update

147 files changed, 29142 insertions, 10582 deletions

diff --git a/eigen/Eigen/src/Core/ArithmeticSequence.h b/eigen/Eigen/src/Core/ArithmeticSequence.h
new file mode 100644
index 0000000..ada1571
--- /dev/null
+++ b/eigen/Eigen/src/Core/ArithmeticSequence.h
@@ -0,0 +1,350 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2017 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ARITHMETIC_SEQUENCE_H
+#define EIGEN_ARITHMETIC_SEQUENCE_H
+
+namespace Eigen {
+
+namespace internal {
+
+#if (!EIGEN_HAS_CXX11) || !((!EIGEN_COMP_GNUC) || EIGEN_COMP_GNUC>=48)
+template<typename T> struct aseq_negate {};
+
+template<> struct aseq_negate<Index> {
+  typedef Index type;
+};
+
+template<int N> struct aseq_negate<FixedInt<N> > {
+  typedef FixedInt<-N> type;
+};
+
+// Compilation error in the following case:
+template<> struct aseq_negate<FixedInt<DynamicIndex> > {};
+
+template<typename FirstType,typename SizeType,typename IncrType,
+         bool FirstIsSymbolic=Symbolic::is_symbolic<FirstType>::value,
+         bool SizeIsSymbolic =Symbolic::is_symbolic<SizeType>::value>
+struct aseq_reverse_first_type {
+  typedef Index type;
+};
+
+template<typename FirstType,typename SizeType,typename IncrType>
+struct aseq_reverse_first_type<FirstType,SizeType,IncrType,true,true> {
+  typedef Symbolic::AddExpr<FirstType,
+                            Symbolic::ProductExpr<Symbolic::AddExpr<SizeType,Symbolic::ValueExpr<FixedInt<-1> > >,
+                                                  Symbolic::ValueExpr<IncrType> >
+                           > type;
+};
+
+template<typename SizeType,typename IncrType,typename EnableIf = void>
+struct aseq_reverse_first_type_aux {
+  typedef Index type;
+};
+
+template<typename SizeType,typename IncrType>
+struct aseq_reverse_first_type_aux<SizeType,IncrType,typename internal::enable_if<bool((SizeType::value+IncrType::value)|0x1)>::type> {
+  typedef FixedInt<(SizeType::value-1)*IncrType::value> type;
+};
+
+template<typename FirstType,typename SizeType,typename IncrType>
+struct aseq_reverse_first_type<FirstType,SizeType,IncrType,true,false> {
+  typedef typename aseq_reverse_first_type_aux<SizeType,IncrType>::type Aux;
+  typedef Symbolic::AddExpr<FirstType,Symbolic::ValueExpr<Aux> > type;
+};
+
+template<typename FirstType,typename SizeType,typename IncrType>
+struct aseq_reverse_first_type<FirstType,SizeType,IncrType,false,true> {
+  typedef Symbolic::AddExpr<Symbolic::ProductExpr<Symbolic::AddExpr<SizeType,Symbolic::ValueExpr<FixedInt<-1> > >,
+                                                  Symbolic::ValueExpr<IncrType> >,
+                            Symbolic::ValueExpr<> > type;
+};
+#endif
+
+// Helper to cleanup the type of the increment:
+template<typename T> struct cleanup_seq_incr {
+  typedef typename cleanup_index_type<T,DynamicIndex>::type type;
+};
+
+}
+
+//--------------------------------------------------------------------------------
+// seq(first,last,incr) and seqN(first,size,incr)
+//--------------------------------------------------------------------------------
+
+template<typename FirstType=Index,typename SizeType=Index,typename IncrType=internal::FixedInt<1> >
+class ArithmeticSequence;
+
+template<typename FirstType,typename SizeType,typename IncrType>
+ArithmeticSequence<typename internal::cleanup_index_type<FirstType>::type,
+                   typename internal::cleanup_index_type<SizeType>::type,
+                   typename internal::cleanup_seq_incr<IncrType>::type >
+seqN(FirstType first, SizeType size, IncrType incr);
+
+/** \class ArithmeticSequence
+  * \ingroup Core_Module
+  *
+  * This class represents an arithmetic progression \f$ a_0, a_1, a_2, ..., a_{n-1}\f$ defined by
+  * its \em first value \f$ a_0 \f$, its \em size (aka length) \em n, and the \em increment (aka stride)
+  * that is equal to \f$ a_{i+1}-a_{i}\f$ for any \em i.
+  *
+  * It is internally used as the return type of the Eigen::seq and Eigen::seqN functions, and as the input arguments
+  * of DenseBase::operator()(const RowIndices&, const ColIndices&), and most of the time this is the
+  * only way it is used.
+  *
+  * \tparam FirstType type of the first element, usually an Index,
+  *                   but internally it can be a symbolic expression
+  * \tparam SizeType type representing the size of the sequence, usually an Index
+  *                  or a compile time integral constant. Internally, it can also be a symbolic expression
+  * \tparam IncrType type of the increment, can be a runtime Index, or a compile time integral constant (default is compile-time 1)
+  *
+  * \sa Eigen::seq, Eigen::seqN, DenseBase::operator()(const RowIndices&, const ColIndices&), class IndexedView
+  */
+template<typename FirstType,typename SizeType,typename IncrType>
+class ArithmeticSequence
+{
+public:
+  ArithmeticSequence(FirstType first, SizeType size) : m_first(first), m_size(size) {}
+  ArithmeticSequence(FirstType first, SizeType size, IncrType incr) : m_first(first), m_size(size), m_incr(incr) {}
+
+  enum {
+    SizeAtCompileTime = internal::get_fixed_value<SizeType>::value,
+    IncrAtCompileTime = internal::get_fixed_value<IncrType,DynamicIndex>::value
+  };
+
+  /** \returns the size, i.e., number of elements, of the sequence */
+  Index size()  const { return m_size; }
+
+  /** \returns the first element \f$ a_0 \f$ in the sequence */
+  Index first()  const { return m_first; }
+
+  /** \returns the value \f$ a_i \f$ at index \a i in the sequence. */
+  Index operator[](Index i) const { return m_first + i * m_incr; }
+
+  const FirstType& firstObject() const { return m_first; }
+  const SizeType&  sizeObject()  const { return m_size; }
+  const IncrType&  incrObject()  const { return m_incr; }
+
+protected:
+  FirstType m_first;
+  SizeType  m_size;
+  IncrType  m_incr;
+
+public:
+
+#if EIGEN_HAS_CXX11 && ((!EIGEN_COMP_GNUC) || EIGEN_COMP_GNUC>=48)
+  auto reverse() const -> decltype(Eigen::seqN(m_first+(m_size+fix<-1>())*m_incr,m_size,-m_incr)) {
+    return seqN(m_first+(m_size+fix<-1>())*m_incr,m_size,-m_incr);
+  }
+#else
+protected:
+  typedef typename internal::aseq_negate<IncrType>::type ReverseIncrType;
+  typedef typename internal::aseq_reverse_first_type<FirstType,SizeType,IncrType>::type ReverseFirstType;
+public:
+  ArithmeticSequence<ReverseFirstType,SizeType,ReverseIncrType>
+  reverse() const {
+    return seqN(m_first+(m_size+fix<-1>())*m_incr,m_size,-m_incr);
+  }
+#endif
+};
+
+/** \returns an ArithmeticSequence starting at \a first, of length \a size, and increment \a incr
+  *
+  * \sa seqN(FirstType,SizeType), seq(FirstType,LastType,IncrType) */
+template<typename FirstType,typename SizeType,typename IncrType>
+ArithmeticSequence<typename internal::cleanup_index_type<FirstType>::type,typename internal::cleanup_index_type<SizeType>::type,typename internal::cleanup_seq_incr<IncrType>::type >
+seqN(FirstType first, SizeType size, IncrType incr)  {
+  return ArithmeticSequence<typename internal::cleanup_index_type<FirstType>::type,typename internal::cleanup_index_type<SizeType>::type,typename internal::cleanup_seq_incr<IncrType>::type>(first,size,incr);
+}
+
+/** \returns an ArithmeticSequence starting at \a first, of length \a size, and unit increment
+  *
+  * \sa seqN(FirstType,SizeType,IncrType), seq(FirstType,LastType) */
+template<typename FirstType,typename SizeType>
+ArithmeticSequence<typename internal::cleanup_index_type<FirstType>::type,typename internal::cleanup_index_type<SizeType>::type >
+seqN(FirstType first, SizeType size)  {
+  return ArithmeticSequence<typename internal::cleanup_index_type<FirstType>::type,typename internal::cleanup_index_type<SizeType>::type>(first,size);
+}
+
+#ifdef EIGEN_PARSED_BY_DOXYGEN
+
+/** \returns an ArithmeticSequence starting at \a f, up (or down) to \a l, and with positive (or negative) increment \a incr
+  *
+  * It is essentially an alias to:
+  * \code
+  * seqN(f, (l-f+incr)/incr, incr);
+  * \endcode
+  *
+  * \sa seqN(FirstType,SizeType,IncrType), seq(FirstType,LastType)
+  */
+template<typename FirstType,typename LastType, typename IncrType>
+auto seq(FirstType f, LastType l, IncrType incr);
+
+/** \returns an ArithmeticSequence starting at \a f, up (or down) to \a l, and unit increment
+  *
+  * It is essentially an alias to:
+  * \code
+  * seqN(f,l-f+1);
+  * \endcode
+  *
+  * \sa seqN(FirstType,SizeType), seq(FirstType,LastType,IncrType)
+  */
+template<typename FirstType,typename LastType>
+auto seq(FirstType f, LastType l);
+
+#else // EIGEN_PARSED_BY_DOXYGEN
+
+#if EIGEN_HAS_CXX11
+template<typename FirstType,typename LastType>
+auto seq(FirstType f, LastType l) -> decltype(seqN(typename internal::cleanup_index_type<FirstType>::type(f),
+                                                   (  typename internal::cleanup_index_type<LastType>::type(l)
+                                                    - typename internal::cleanup_index_type<FirstType>::type(f)+fix<1>())))
+{
+  return seqN(typename internal::cleanup_index_type<FirstType>::type(f),
+              (typename internal::cleanup_index_type<LastType>::type(l)
+               -typename internal::cleanup_index_type<FirstType>::type(f)+fix<1>()));
+}
+
+template<typename FirstType,typename LastType, typename IncrType>
+auto seq(FirstType f, LastType l, IncrType incr)
+  -> decltype(seqN(typename internal::cleanup_index_type<FirstType>::type(f),
+                   (   typename internal::cleanup_index_type<LastType>::type(l)
+                     - typename internal::cleanup_index_type<FirstType>::type(f)+typename internal::cleanup_seq_incr<IncrType>::type(incr)
+                   ) / typename internal::cleanup_seq_incr<IncrType>::type(incr),
+                   typename internal::cleanup_seq_incr<IncrType>::type(incr)))
+{
+  typedef typename internal::cleanup_seq_incr<IncrType>::type CleanedIncrType;
+  return seqN(typename internal::cleanup_index_type<FirstType>::type(f),
+              ( typename internal::cleanup_index_type<LastType>::type(l)
+               -typename internal::cleanup_index_type<FirstType>::type(f)+CleanedIncrType(incr)) / CleanedIncrType(incr),
+              CleanedIncrType(incr));
+}
+#else
+
+template<typename FirstType,typename LastType>
+typename internal::enable_if<!(Symbolic::is_symbolic<FirstType>::value || Symbolic::is_symbolic<LastType>::value),
+                             ArithmeticSequence<typename internal::cleanup_index_type<FirstType>::type,Index> >::type
+seq(FirstType f, LastType l)
+{
+  return seqN(typename internal::cleanup_index_type<FirstType>::type(f),
+              Index((typename internal::cleanup_index_type<LastType>::type(l)-typename internal::cleanup_index_type<FirstType>::type(f)+fix<1>())));
+}
+
+template<typename FirstTypeDerived,typename LastType>
+typename internal::enable_if<!Symbolic::is_symbolic<LastType>::value,
+    ArithmeticSequence<FirstTypeDerived, Symbolic::AddExpr<Symbolic::AddExpr<Symbolic::NegateExpr<FirstTypeDerived>,Symbolic::ValueExpr<> >,
+                                                            Symbolic::ValueExpr<internal::FixedInt<1> > > > >::type
+seq(const Symbolic::BaseExpr<FirstTypeDerived> &f, LastType l)
+{
+  return seqN(f.derived(),(typename internal::cleanup_index_type<LastType>::type(l)-f.derived()+fix<1>()));
+}
+
+template<typename FirstType,typename LastTypeDerived>
+typename internal::enable_if<!Symbolic::is_symbolic<FirstType>::value,
+    ArithmeticSequence<typename internal::cleanup_index_type<FirstType>::type,
+                        Symbolic::AddExpr<Symbolic::AddExpr<LastTypeDerived,Symbolic::ValueExpr<> >,
+                                          Symbolic::ValueExpr<internal::FixedInt<1> > > > >::type
+seq(FirstType f, const Symbolic::BaseExpr<LastTypeDerived> &l)
+{
+  return seqN(typename internal::cleanup_index_type<FirstType>::type(f),(l.derived()-typename internal::cleanup_index_type<FirstType>::type(f)+fix<1>()));
+}
+
+template<typename FirstTypeDerived,typename LastTypeDerived>
+ArithmeticSequence<FirstTypeDerived,
+                    Symbolic::AddExpr<Symbolic::AddExpr<LastTypeDerived,Symbolic::NegateExpr<FirstTypeDerived> >,Symbolic::ValueExpr<internal::FixedInt<1> > > >
+seq(const Symbolic::BaseExpr<FirstTypeDerived> &f, const Symbolic::BaseExpr<LastTypeDerived> &l)
+{
+  return seqN(f.derived(),(l.derived()-f.derived()+fix<1>()));
+}
+
+
+template<typename FirstType,typename LastType, typename IncrType>
+typename internal::enable_if<!(Symbolic::is_symbolic<FirstType>::value || Symbolic::is_symbolic<LastType>::value),
+    ArithmeticSequence<typename internal::cleanup_index_type<FirstType>::type,Index,typename internal::cleanup_seq_incr<IncrType>::type> >::type
+seq(FirstType f, LastType l, IncrType incr)
+{
+  typedef typename internal::cleanup_seq_incr<IncrType>::type CleanedIncrType;
+  return seqN(typename internal::cleanup_index_type<FirstType>::type(f),
+              Index((typename internal::cleanup_index_type<LastType>::type(l)-typename internal::cleanup_index_type<FirstType>::type(f)+CleanedIncrType(incr))/CleanedIncrType(incr)), incr);
+}
+
+template<typename FirstTypeDerived,typename LastType, typename IncrType>
+typename internal::enable_if<!Symbolic::is_symbolic<LastType>::value,
+    ArithmeticSequence<FirstTypeDerived,
+                        Symbolic::QuotientExpr<Symbolic::AddExpr<Symbolic::AddExpr<Symbolic::NegateExpr<FirstTypeDerived>,
+                                                                                   Symbolic::ValueExpr<> >,
+                                                                 Symbolic::ValueExpr<typename internal::cleanup_seq_incr<IncrType>::type> >,
+                                              Symbolic::ValueExpr<typename internal::cleanup_seq_incr<IncrType>::type> >,
+                        typename internal::cleanup_seq_incr<IncrType>::type> >::type
+seq(const Symbolic::BaseExpr<FirstTypeDerived> &f, LastType l, IncrType incr)
+{
+  typedef typename internal::cleanup_seq_incr<IncrType>::type CleanedIncrType;
+  return seqN(f.derived(),(typename internal::cleanup_index_type<LastType>::type(l)-f.derived()+CleanedIncrType(incr))/CleanedIncrType(incr), incr);
+}
+
+template<typename FirstType,typename LastTypeDerived, typename IncrType>
+typename internal::enable_if<!Symbolic::is_symbolic<FirstType>::value,
+    ArithmeticSequence<typename internal::cleanup_index_type<FirstType>::type,
+                        Symbolic::QuotientExpr<Symbolic::AddExpr<Symbolic::AddExpr<LastTypeDerived,Symbolic::ValueExpr<> >,
+                                                                 Symbolic::ValueExpr<typename internal::cleanup_seq_incr<IncrType>::type> >,
+                                               Symbolic::ValueExpr<typename internal::cleanup_seq_incr<IncrType>::type> >,
+                        typename internal::cleanup_seq_incr<IncrType>::type> >::type
+seq(FirstType f, const Symbolic::BaseExpr<LastTypeDerived> &l, IncrType incr)
+{
+  typedef typename internal::cleanup_seq_incr<IncrType>::type CleanedIncrType;
+  return seqN(typename internal::cleanup_index_type<FirstType>::type(f),
+              (l.derived()-typename internal::cleanup_index_type<FirstType>::type(f)+CleanedIncrType(incr))/CleanedIncrType(incr), incr);
+}
+
+template<typename FirstTypeDerived,typename LastTypeDerived, typename IncrType>
+ArithmeticSequence<FirstTypeDerived,
+                    Symbolic::QuotientExpr<Symbolic::AddExpr<Symbolic::AddExpr<LastTypeDerived,
+                                                                               Symbolic::NegateExpr<FirstTypeDerived> >,
+                                                             Symbolic::ValueExpr<typename internal::cleanup_seq_incr<IncrType>::type> >,
+                                          Symbolic::ValueExpr<typename internal::cleanup_seq_incr<IncrType>::type> >,
+                    typename internal::cleanup_seq_incr<IncrType>::type>
+seq(const Symbolic::BaseExpr<FirstTypeDerived> &f, const Symbolic::BaseExpr<LastTypeDerived> &l, IncrType incr)
+{
+  typedef typename internal::cleanup_seq_incr<IncrType>::type CleanedIncrType;
+  return seqN(f.derived(),(l.derived()-f.derived()+CleanedIncrType(incr))/CleanedIncrType(incr), incr);
+}
+#endif
+
+#endif // EIGEN_PARSED_BY_DOXYGEN
+
+namespace internal {
+
+// Convert a symbolic span into a usable one (i.e., remove last/end "keywords")
+template<typename T>
+struct make_size_type {
+  typedef typename internal::conditional<Symbolic::is_symbolic<T>::value, Index, T>::type type;
+};
+
+template<typename FirstType,typename SizeType,typename IncrType,int XprSize>
+struct IndexedViewCompatibleType<ArithmeticSequence<FirstType,SizeType,IncrType>, XprSize> {
+  typedef ArithmeticSequence<Index,typename make_size_type<SizeType>::type,IncrType> type;
+};
+
+template<typename FirstType,typename SizeType,typename IncrType>
+ArithmeticSequence<Index,typename make_size_type<SizeType>::type,IncrType>
+makeIndexedViewCompatible(const ArithmeticSequence<FirstType,SizeType,IncrType>& ids, Index size,SpecializedType) {
+  return ArithmeticSequence<Index,typename make_size_type<SizeType>::type,IncrType>(
+            eval_expr_given_size(ids.firstObject(),size),eval_expr_given_size(ids.sizeObject(),size),ids.incrObject());
+}
+
+template<typename FirstType,typename SizeType,typename IncrType>
+struct get_compile_time_incr<ArithmeticSequence<FirstType,SizeType,IncrType> > {
+  enum { value = get_fixed_value<IncrType,DynamicIndex>::value };
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_ARITHMETIC_SEQUENCE_H
diff --git a/eigen/Eigen/src/Core/Array.h b/eigen/Eigen/src/Core/Array.h
index 0b9c38c..0d34269 100644
--- a/eigen/Eigen/src/Core/Array.h
+++ b/eigen/Eigen/src/Core/Array.h
@@ -12,7 +12,16 @@
 
 namespace Eigen {
 
-/** \class Array 
+namespace internal {
+template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
+struct traits<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> > : traits<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >
+{
+  typedef ArrayXpr XprKind;
+  typedef ArrayBase<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> > XprBase;
+};
+}
+
+/** \class Array
   * \ingroup Core_Module
   *
   * \brief General-purpose arrays with easy API for coefficient-wise operations
@@ -24,20 +33,14 @@ namespace Eigen {
   * API for the %Matrix class provides easy access to linear-algebra
   * operations.
   *
+  * See documentation of class Matrix for detailed information on the template parameters
+  * storage layout.
+  *
   * This class can be extended with the help of the plugin mechanism described on the page
-  * \ref TopicCustomizingEigen by defining the preprocessor symbol \c EIGEN_ARRAY_PLUGIN.
+  * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_ARRAY_PLUGIN.
   *
-  * \sa \ref TutorialArrayClass, \ref TopicClassHierarchy
+  * \sa \blank \ref TutorialArrayClass, \ref TopicClassHierarchy
   */
-namespace internal {
-template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
-struct traits<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> > : traits<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >
-{
-  typedef ArrayXpr XprKind;
-  typedef ArrayBase<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> > XprBase;
-};
-}
-
 template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
 class Array
   : public PlainObjectBase<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >
@@ -69,11 +72,27 @@ class Array
       * the usage of 'using'. This should be done only for operator=.
       */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Array& operator=(const EigenBase<OtherDerived> &other)
     {
       return Base::operator=(other);
     }
 
+    /** Set all the entries to \a value.
+      * \sa DenseBase::setConstant(), DenseBase::fill()
+      */
+    /* This overload is needed because the usage of
+      *   using Base::operator=;
+      * fails on MSVC. Since the code below is working with GCC and MSVC, we skipped
+      * the usage of 'using'. This should be done only for operator=.
+      */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Array& operator=(const Scalar &value)
+    {
+      Base::setConstant(value);
+      return *this;
+    }
+
     /** Copies the value of the expression \a other into \c *this with automatic resizing.
       *
       * *this might be resized to match the dimensions of \a other. If *this was a null matrix (not already initialized),
@@ -84,7 +103,8 @@ class Array
       * remain row-vectors and vectors remain vectors.
       */
     template<typename OtherDerived>
-    EIGEN_STRONG_INLINE Array& operator=(const ArrayBase<OtherDerived>& other)
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Array& operator=(const DenseBase<OtherDerived>& other)
     {
       return Base::_set(other);
     }
@@ -92,11 +112,12 @@ class Array
     /** This is a special case of the templated operator=. Its purpose is to
       * prevent a default operator= from hiding the templated operator=.
       */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Array& operator=(const Array& other)
     {
       return Base::_set(other);
     }
-
+    
     /** Default constructor.
       *
       * For fixed-size matrices, does nothing.
@@ -107,6 +128,7 @@ class Array
       *
       * \sa resize(Index,Index)
       */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Array() : Base()
     {
       Base::_check_template_params();
@@ -116,6 +138,7 @@ class Array
 #ifndef EIGEN_PARSED_BY_DOXYGEN
     // FIXME is it still needed ??
     /** \internal */
+    EIGEN_DEVICE_FUNC
     Array(internal::constructor_without_unaligned_array_assert)
       : Base(internal::constructor_without_unaligned_array_assert())
     {
@@ -124,56 +147,64 @@ class Array
     }
 #endif
 
-#ifdef EIGEN_HAVE_RVALUE_REFERENCES
-    Array(Array&& other)
+#if EIGEN_HAS_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC
+    Array(Array&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_constructible<Scalar>::value)
       : Base(std::move(other))
     {
       Base::_check_template_params();
       if (RowsAtCompileTime!=Dynamic && ColsAtCompileTime!=Dynamic)
         Base::_set_noalias(other);
     }
-    Array& operator=(Array&& other)
+    EIGEN_DEVICE_FUNC
+    Array& operator=(Array&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_assignable<Scalar>::value)
     {
       other.swap(*this);
       return *this;
     }
 #endif
 
-    /** Constructs a vector or row-vector with given dimension. \only_for_vectors
-      *
-      * Note that this is only useful for dynamic-size vectors. For fixed-size vectors,
-      * it is redundant to pass the dimension here, so it makes more sense to use the default
-      * constructor Matrix() instead.
-      */
-    EIGEN_STRONG_INLINE explicit Array(Index dim)
-      : Base(dim, RowsAtCompileTime == 1 ? 1 : dim, ColsAtCompileTime == 1 ? 1 : dim)
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE explicit Array(const T& x)
     {
       Base::_check_template_params();
-      EIGEN_STATIC_ASSERT_VECTOR_ONLY(Array)
-      eigen_assert(dim >= 0);
-      eigen_assert(SizeAtCompileTime == Dynamic || SizeAtCompileTime == dim);
-      EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
+      Base::template _init1<T>(x);
     }
 
-    #ifndef EIGEN_PARSED_BY_DOXYGEN
     template<typename T0, typename T1>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Array(const T0& val0, const T1& val1)
     {
       Base::_check_template_params();
       this->template _init2<T0,T1>(val0, val1);
     }
     #else
-    /** constructs an uninitialized matrix with \a rows rows and \a cols columns.
+    /** \brief Constructs a fixed-sized array initialized with coefficients starting at \a data */
+    EIGEN_DEVICE_FUNC explicit Array(const Scalar *data);
+    /** Constructs a vector or row-vector with given dimension. \only_for_vectors
       *
-      * This is useful for dynamic-size matrices. For fixed-size matrices,
+      * Note that this is only useful for dynamic-size vectors. For fixed-size vectors,
+      * it is redundant to pass the dimension here, so it makes more sense to use the default
+      * constructor Array() instead.
+      */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE explicit Array(Index dim);
+    /** constructs an initialized 1x1 Array with the given coefficient */
+    Array(const Scalar& value);
+    /** constructs an uninitialized array with \a rows rows and \a cols columns.
+      *
+      * This is useful for dynamic-size arrays. For fixed-size arrays,
       * it is redundant to pass these parameters, so one should use the default constructor
-      * Matrix() instead. */
+      * Array() instead. */
     Array(Index rows, Index cols);
     /** constructs an initialized 2D vector with given coefficients */
     Array(const Scalar& val0, const Scalar& val1);
     #endif
 
     /** constructs an initialized 3D vector with given coefficients */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Array(const Scalar& val0, const Scalar& val1, const Scalar& val2)
     {
       Base::_check_template_params();
@@ -183,6 +214,7 @@ class Array
       m_storage.data()[2] = val2;
     }
     /** constructs an initialized 4D vector with given coefficients */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Array(const Scalar& val0, const Scalar& val1, const Scalar& val2, const Scalar& val3)
     {
       Base::_check_template_params();
@@ -193,51 +225,21 @@ class Array
       m_storage.data()[3] = val3;
     }
 
-    explicit Array(const Scalar *data);
-
-    /** Constructor copying the value of the expression \a other */
-    template<typename OtherDerived>
-    EIGEN_STRONG_INLINE Array(const ArrayBase<OtherDerived>& other)
-             : Base(other.rows() * other.cols(), other.rows(), other.cols())
-    {
-      Base::_check_template_params();
-      Base::_set_noalias(other);
-    }
     /** Copy constructor */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Array(const Array& other)
-            : Base(other.rows() * other.cols(), other.rows(), other.cols())
-    {
-      Base::_check_template_params();
-      Base::_set_noalias(other);
-    }
-    /** Copy constructor with in-place evaluation */
-    template<typename OtherDerived>
-    EIGEN_STRONG_INLINE Array(const ReturnByValue<OtherDerived>& other)
-    {
-      Base::_check_template_params();
-      Base::resize(other.rows(), other.cols());
-      other.evalTo(*this);
-    }
+            : Base(other)
+    { }
 
     /** \sa MatrixBase::operator=(const EigenBase<OtherDerived>&) */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Array(const EigenBase<OtherDerived> &other)
-      : Base(other.derived().rows() * other.derived().cols(), other.derived().rows(), other.derived().cols())
-    {
-      Base::_check_template_params();
-      Base::_resize_to_match(other);
-      *this = other;
-    }
-
-    /** Override MatrixBase::swap() since for dynamic-sized matrices of same type it is enough to swap the
-      * data pointers.
-      */
-    template<typename OtherDerived>
-    void swap(ArrayBase<OtherDerived> const & other)
-    { this->_swap(other.derived()); }
+      : Base(other.derived())
+    { }
 
-    inline Index innerStride() const { return 1; }
-    inline Index outerStride() const { return this->innerSize(); }
+    EIGEN_DEVICE_FUNC inline Index innerStride() const { return 1; }
+    EIGEN_DEVICE_FUNC inline Index outerStride() const { return this->innerSize(); }
 
     #ifdef EIGEN_ARRAY_PLUGIN
     #include EIGEN_ARRAY_PLUGIN
diff --git a/eigen/Eigen/src/Core/ArrayBase.h b/eigen/Eigen/src/Core/ArrayBase.h
index 33ff553..9da960f 100644
--- a/eigen/Eigen/src/Core/ArrayBase.h
+++ b/eigen/Eigen/src/Core/ArrayBase.h
@@ -32,7 +32,7 @@ template<typename ExpressionType> class MatrixWrapper;
   * \tparam Derived is the derived type, e.g., an array or an expression type.
   *
   * This class can be extended with the help of the plugin mechanism described on the page
-  * \ref TopicCustomizingEigen by defining the preprocessor symbol \c EIGEN_ARRAYBASE_PLUGIN.
+  * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_ARRAYBASE_PLUGIN.
   *
   * \sa class MatrixBase, \ref TopicClassHierarchy
   */
@@ -47,13 +47,11 @@ template<typename Derived> class ArrayBase
     typedef ArrayBase Eigen_BaseClassForSpecializationOfGlobalMathFuncImpl;
 
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
-    typedef typename internal::traits<Derived>::Index Index;
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef typename internal::packet_traits<Scalar>::type PacketScalar;
     typedef typename NumTraits<Scalar>::Real RealScalar;
 
     typedef DenseBase<Derived> Base;
-    using Base::operator*;
     using Base::RowsAtCompileTime;
     using Base::ColsAtCompileTime;
     using Base::SizeAtCompileTime;
@@ -62,8 +60,7 @@ template<typename Derived> class ArrayBase
     using Base::MaxSizeAtCompileTime;
     using Base::IsVectorAtCompileTime;
     using Base::Flags;
-    using Base::CoeffReadCost;
-
+    
     using Base::derived;
     using Base::const_cast_derived;
     using Base::rows;
@@ -72,6 +69,7 @@ template<typename Derived> class ArrayBase
     using Base::coeff;
     using Base::coeffRef;
     using Base::lazyAssign;
+    using Base::operator-;
     using Base::operator=;
     using Base::operator+=;
     using Base::operator-=;
@@ -83,26 +81,14 @@ template<typename Derived> class ArrayBase
 #endif // not EIGEN_PARSED_BY_DOXYGEN
 
 #ifndef EIGEN_PARSED_BY_DOXYGEN
-    /** \internal the plain matrix type corresponding to this expression. Note that is not necessarily
-      * exactly the return type of eval(): in the case of plain matrices, the return type of eval() is a const
-      * reference to a matrix, not a matrix! It is however guaranteed that the return type of eval() is either
-      * PlainObject or const PlainObject&.
-      */
-    typedef Array<typename internal::traits<Derived>::Scalar,
-                internal::traits<Derived>::RowsAtCompileTime,
-                internal::traits<Derived>::ColsAtCompileTime,
-                AutoAlign | (internal::traits<Derived>::Flags&RowMajorBit ? RowMajor : ColMajor),
-                internal::traits<Derived>::MaxRowsAtCompileTime,
-                internal::traits<Derived>::MaxColsAtCompileTime
-          > PlainObject;
-
+    typedef typename Base::PlainObject PlainObject;
 
     /** \internal Represents a matrix with all coefficients equal to one another*/
-    typedef CwiseNullaryOp<internal::scalar_constant_op<Scalar>,Derived> ConstantReturnType;
+    typedef CwiseNullaryOp<internal::scalar_constant_op<Scalar>,PlainObject> ConstantReturnType;
 #endif // not EIGEN_PARSED_BY_DOXYGEN
 
 #define EIGEN_CURRENT_STORAGE_BASE_CLASS Eigen::ArrayBase
-#   include "../plugins/CommonCwiseUnaryOps.h"
+#define EIGEN_DOC_UNARY_ADDONS(X,Y)
 #   include "../plugins/MatrixCwiseUnaryOps.h"
 #   include "../plugins/ArrayCwiseUnaryOps.h"
 #   include "../plugins/CommonCwiseBinaryOps.h"
@@ -112,44 +98,62 @@ template<typename Derived> class ArrayBase
 #     include EIGEN_ARRAYBASE_PLUGIN
 #   endif
 #undef EIGEN_CURRENT_STORAGE_BASE_CLASS
+#undef EIGEN_DOC_UNARY_ADDONS
 
     /** Special case of the template operator=, in order to prevent the compiler
       * from generating a default operator= (issue hit with g++ 4.1)
       */
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     Derived& operator=(const ArrayBase& other)
     {
-      return internal::assign_selector<Derived,Derived>::run(derived(), other.derived());
+      internal::call_assignment(derived(), other.derived());
+      return derived();
     }
-
-    Derived& operator+=(const Scalar& scalar)
-    { return *this = derived() + scalar; }
-    Derived& operator-=(const Scalar& scalar)
-    { return *this = derived() - scalar; }
+    
+    /** Set all the entries to \a value.
+      * \sa DenseBase::setConstant(), DenseBase::fill() */
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Derived& operator=(const Scalar &value)
+    { Base::setConstant(value); return derived(); }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Derived& operator+=(const Scalar& scalar);
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Derived& operator-=(const Scalar& scalar);
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     Derived& operator+=(const ArrayBase<OtherDerived>& other);
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     Derived& operator-=(const ArrayBase<OtherDerived>& other);
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     Derived& operator*=(const ArrayBase<OtherDerived>& other);
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     Derived& operator/=(const ArrayBase<OtherDerived>& other);
 
   public:
+    EIGEN_DEVICE_FUNC
     ArrayBase<Derived>& array() { return *this; }
+    EIGEN_DEVICE_FUNC
     const ArrayBase<Derived>& array() const { return *this; }
 
     /** \returns an \link Eigen::MatrixBase Matrix \endlink expression of this array
       * \sa MatrixBase::array() */
-    MatrixWrapper<Derived> matrix() { return derived(); }
-    const MatrixWrapper<const Derived> matrix() const { return derived(); }
+    EIGEN_DEVICE_FUNC
+    MatrixWrapper<Derived> matrix() { return MatrixWrapper<Derived>(derived()); }
+    EIGEN_DEVICE_FUNC
+    const MatrixWrapper<const Derived> matrix() const { return MatrixWrapper<const Derived>(derived()); }
 
 //     template<typename Dest>
 //     inline void evalTo(Dest& dst) const { dst = matrix(); }
 
   protected:
+    EIGEN_DEVICE_FUNC
     ArrayBase() : Base() {}
 
   private:
@@ -171,11 +175,10 @@ template<typename Derived> class ArrayBase
   */
 template<typename Derived>
 template<typename OtherDerived>
-EIGEN_STRONG_INLINE Derived &
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived &
 ArrayBase<Derived>::operator-=(const ArrayBase<OtherDerived> &other)
 {
-  SelfCwiseBinaryOp<internal::scalar_difference_op<Scalar>, Derived, OtherDerived> tmp(derived());
-  tmp = other.derived();
+  call_assignment(derived(), other.derived(), internal::sub_assign_op<Scalar,typename OtherDerived::Scalar>());
   return derived();
 }
 
@@ -185,11 +188,10 @@ ArrayBase<Derived>::operator-=(const ArrayBase<OtherDerived> &other)
   */
 template<typename Derived>
 template<typename OtherDerived>
-EIGEN_STRONG_INLINE Derived &
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived &
 ArrayBase<Derived>::operator+=(const ArrayBase<OtherDerived>& other)
 {
-  SelfCwiseBinaryOp<internal::scalar_sum_op<Scalar>, Derived, OtherDerived> tmp(derived());
-  tmp = other.derived();
+  call_assignment(derived(), other.derived(), internal::add_assign_op<Scalar,typename OtherDerived::Scalar>());
   return derived();
 }
 
@@ -199,11 +201,10 @@ ArrayBase<Derived>::operator+=(const ArrayBase<OtherDerived>& other)
   */
 template<typename Derived>
 template<typename OtherDerived>
-EIGEN_STRONG_INLINE Derived &
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived &
 ArrayBase<Derived>::operator*=(const ArrayBase<OtherDerived>& other)
 {
-  SelfCwiseBinaryOp<internal::scalar_product_op<Scalar>, Derived, OtherDerived> tmp(derived());
-  tmp = other.derived();
+  call_assignment(derived(), other.derived(), internal::mul_assign_op<Scalar,typename OtherDerived::Scalar>());
   return derived();
 }
 
@@ -213,11 +214,10 @@ ArrayBase<Derived>::operator*=(const ArrayBase<OtherDerived>& other)
   */
 template<typename Derived>
 template<typename OtherDerived>
-EIGEN_STRONG_INLINE Derived &
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived &
 ArrayBase<Derived>::operator/=(const ArrayBase<OtherDerived>& other)
 {
-  SelfCwiseBinaryOp<internal::scalar_quotient_op<Scalar>, Derived, OtherDerived> tmp(derived());
-  tmp = other.derived();
+  call_assignment(derived(), other.derived(), internal::div_assign_op<Scalar,typename OtherDerived::Scalar>());
   return derived();
 }
 
diff --git a/eigen/Eigen/src/Core/ArrayWrapper.h b/eigen/Eigen/src/Core/ArrayWrapper.h
index b4641e2..a04521a 100644
--- a/eigen/Eigen/src/Core/ArrayWrapper.h
+++ b/eigen/Eigen/src/Core/ArrayWrapper.h
@@ -44,6 +44,7 @@ class ArrayWrapper : public ArrayBase<ArrayWrapper<ExpressionType> >
     typedef ArrayBase<ArrayWrapper> Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(ArrayWrapper)
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(ArrayWrapper)
+    typedef typename internal::remove_all<ExpressionType>::type NestedExpression;
 
     typedef typename internal::conditional<
                        internal::is_lvalue<ExpressionType>::value,
@@ -51,76 +52,45 @@ class ArrayWrapper : public ArrayBase<ArrayWrapper<ExpressionType> >
                        const Scalar
                      >::type ScalarWithConstIfNotLvalue;
 
-    typedef typename internal::nested<ExpressionType>::type NestedExpressionType;
+    typedef typename internal::ref_selector<ExpressionType>::non_const_type NestedExpressionType;
 
-    inline ArrayWrapper(ExpressionType& matrix) : m_expression(matrix) {}
+    using Base::coeffRef;
 
+    EIGEN_DEVICE_FUNC
+    explicit EIGEN_STRONG_INLINE ArrayWrapper(ExpressionType& matrix) : m_expression(matrix) {}
+
+    EIGEN_DEVICE_FUNC
     inline Index rows() const { return m_expression.rows(); }
+    EIGEN_DEVICE_FUNC
     inline Index cols() const { return m_expression.cols(); }
+    EIGEN_DEVICE_FUNC
     inline Index outerStride() const { return m_expression.outerStride(); }
+    EIGEN_DEVICE_FUNC
     inline Index innerStride() const { return m_expression.innerStride(); }
 
-    inline ScalarWithConstIfNotLvalue* data() { return m_expression.const_cast_derived().data(); }
+    EIGEN_DEVICE_FUNC
+    inline ScalarWithConstIfNotLvalue* data() { return m_expression.data(); }
+    EIGEN_DEVICE_FUNC
     inline const Scalar* data() const { return m_expression.data(); }
 
-    inline CoeffReturnType coeff(Index rowId, Index colId) const
-    {
-      return m_expression.coeff(rowId, colId);
-    }
-
-    inline Scalar& coeffRef(Index rowId, Index colId)
-    {
-      return m_expression.const_cast_derived().coeffRef(rowId, colId);
-    }
-
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index rowId, Index colId) const
     {
-      return m_expression.const_cast_derived().coeffRef(rowId, colId);
-    }
-
-    inline CoeffReturnType coeff(Index index) const
-    {
-      return m_expression.coeff(index);
-    }
-
-    inline Scalar& coeffRef(Index index)
-    {
-      return m_expression.const_cast_derived().coeffRef(index);
+      return m_expression.coeffRef(rowId, colId);
     }
 
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index index) const
     {
-      return m_expression.const_cast_derived().coeffRef(index);
-    }
-
-    template<int LoadMode>
-    inline const PacketScalar packet(Index rowId, Index colId) const
-    {
-      return m_expression.template packet<LoadMode>(rowId, colId);
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index rowId, Index colId, const PacketScalar& val)
-    {
-      m_expression.const_cast_derived().template writePacket<LoadMode>(rowId, colId, val);
-    }
-
-    template<int LoadMode>
-    inline const PacketScalar packet(Index index) const
-    {
-      return m_expression.template packet<LoadMode>(index);
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index index, const PacketScalar& val)
-    {
-      m_expression.const_cast_derived().template writePacket<LoadMode>(index, val);
+      return m_expression.coeffRef(index);
     }
 
     template<typename Dest>
+    EIGEN_DEVICE_FUNC
     inline void evalTo(Dest& dst) const { dst = m_expression; }
 
     const typename internal::remove_all<NestedExpressionType>::type& 
+    EIGEN_DEVICE_FUNC
     nestedExpression() const 
     {
       return m_expression;
@@ -128,10 +98,12 @@ class ArrayWrapper : public ArrayBase<ArrayWrapper<ExpressionType> >
 
     /** Forwards the resizing request to the nested expression
       * \sa DenseBase::resize(Index)  */
-    void resize(Index newSize) { m_expression.const_cast_derived().resize(newSize); }
+    EIGEN_DEVICE_FUNC
+    void resize(Index newSize) { m_expression.resize(newSize); }
     /** Forwards the resizing request to the nested expression
       * \sa DenseBase::resize(Index,Index)*/
-    void resize(Index nbRows, Index nbCols) { m_expression.const_cast_derived().resize(nbRows,nbCols); }
+    EIGEN_DEVICE_FUNC
+    void resize(Index rows, Index cols) { m_expression.resize(rows,cols); }
 
   protected:
     NestedExpressionType m_expression;
@@ -169,6 +141,7 @@ class MatrixWrapper : public MatrixBase<MatrixWrapper<ExpressionType> >
     typedef MatrixBase<MatrixWrapper<ExpressionType> > Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(MatrixWrapper)
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(MatrixWrapper)
+    typedef typename internal::remove_all<ExpressionType>::type NestedExpression;
 
     typedef typename internal::conditional<
                        internal::is_lvalue<ExpressionType>::value,
@@ -176,72 +149,40 @@ class MatrixWrapper : public MatrixBase<MatrixWrapper<ExpressionType> >
                        const Scalar
                      >::type ScalarWithConstIfNotLvalue;
 
-    typedef typename internal::nested<ExpressionType>::type NestedExpressionType;
+    typedef typename internal::ref_selector<ExpressionType>::non_const_type NestedExpressionType;
 
-    inline MatrixWrapper(ExpressionType& a_matrix) : m_expression(a_matrix) {}
+    using Base::coeffRef;
 
+    EIGEN_DEVICE_FUNC
+    explicit inline MatrixWrapper(ExpressionType& matrix) : m_expression(matrix) {}
+
+    EIGEN_DEVICE_FUNC
     inline Index rows() const { return m_expression.rows(); }
+    EIGEN_DEVICE_FUNC
     inline Index cols() const { return m_expression.cols(); }
+    EIGEN_DEVICE_FUNC
     inline Index outerStride() const { return m_expression.outerStride(); }
+    EIGEN_DEVICE_FUNC
     inline Index innerStride() const { return m_expression.innerStride(); }
 
-    inline ScalarWithConstIfNotLvalue* data() { return m_expression.const_cast_derived().data(); }
+    EIGEN_DEVICE_FUNC
+    inline ScalarWithConstIfNotLvalue* data() { return m_expression.data(); }
+    EIGEN_DEVICE_FUNC
     inline const Scalar* data() const { return m_expression.data(); }
 
-    inline CoeffReturnType coeff(Index rowId, Index colId) const
-    {
-      return m_expression.coeff(rowId, colId);
-    }
-
-    inline Scalar& coeffRef(Index rowId, Index colId)
-    {
-      return m_expression.const_cast_derived().coeffRef(rowId, colId);
-    }
-
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index rowId, Index colId) const
     {
       return m_expression.derived().coeffRef(rowId, colId);
     }
 
-    inline CoeffReturnType coeff(Index index) const
-    {
-      return m_expression.coeff(index);
-    }
-
-    inline Scalar& coeffRef(Index index)
-    {
-      return m_expression.const_cast_derived().coeffRef(index);
-    }
-
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index index) const
     {
-      return m_expression.const_cast_derived().coeffRef(index);
-    }
-
-    template<int LoadMode>
-    inline const PacketScalar packet(Index rowId, Index colId) const
-    {
-      return m_expression.template packet<LoadMode>(rowId, colId);
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index rowId, Index colId, const PacketScalar& val)
-    {
-      m_expression.const_cast_derived().template writePacket<LoadMode>(rowId, colId, val);
-    }
-
-    template<int LoadMode>
-    inline const PacketScalar packet(Index index) const
-    {
-      return m_expression.template packet<LoadMode>(index);
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index index, const PacketScalar& val)
-    {
-      m_expression.const_cast_derived().template writePacket<LoadMode>(index, val);
+      return m_expression.coeffRef(index);
     }
 
+    EIGEN_DEVICE_FUNC
     const typename internal::remove_all<NestedExpressionType>::type& 
     nestedExpression() const 
     {
@@ -250,10 +191,12 @@ class MatrixWrapper : public MatrixBase<MatrixWrapper<ExpressionType> >
 
     /** Forwards the resizing request to the nested expression
       * \sa DenseBase::resize(Index)  */
-    void resize(Index newSize) { m_expression.const_cast_derived().resize(newSize); }
+    EIGEN_DEVICE_FUNC
+    void resize(Index newSize) { m_expression.resize(newSize); }
     /** Forwards the resizing request to the nested expression
       * \sa DenseBase::resize(Index,Index)*/
-    void resize(Index nbRows, Index nbCols) { m_expression.const_cast_derived().resize(nbRows,nbCols); }
+    EIGEN_DEVICE_FUNC
+    void resize(Index rows, Index cols) { m_expression.resize(rows,cols); }
 
   protected:
     NestedExpressionType m_expression;
diff --git a/eigen/Eigen/src/Core/Assign.h b/eigen/Eigen/src/Core/Assign.h
index f481731..655412e 100644
--- a/eigen/Eigen/src/Core/Assign.h
+++ b/eigen/Eigen/src/Core/Assign.h
@@ -14,481 +14,9 @@
 
 namespace Eigen {
 
-namespace internal {
-
-/***************************************************************************
-* Part 1 : the logic deciding a strategy for traversal and unrolling       *
-***************************************************************************/
-
-template <typename Derived, typename OtherDerived>
-struct assign_traits
-{
-public:
-  enum {
-    DstIsAligned = Derived::Flags & AlignedBit,
-    DstHasDirectAccess = Derived::Flags & DirectAccessBit,
-    SrcIsAligned = OtherDerived::Flags & AlignedBit,
-    JointAlignment = bool(DstIsAligned) && bool(SrcIsAligned) ? Aligned : Unaligned
-  };
-
-private:
-  enum {
-    InnerSize = int(Derived::IsVectorAtCompileTime) ? int(Derived::SizeAtCompileTime)
-              : int(Derived::Flags)&RowMajorBit ? int(Derived::ColsAtCompileTime)
-              : int(Derived::RowsAtCompileTime),
-    InnerMaxSize = int(Derived::IsVectorAtCompileTime) ? int(Derived::MaxSizeAtCompileTime)
-              : int(Derived::Flags)&RowMajorBit ? int(Derived::MaxColsAtCompileTime)
-              : int(Derived::MaxRowsAtCompileTime),
-    MaxSizeAtCompileTime = Derived::SizeAtCompileTime,
-    PacketSize = packet_traits<typename Derived::Scalar>::size
-  };
-
-  enum {
-    StorageOrdersAgree = (int(Derived::IsRowMajor) == int(OtherDerived::IsRowMajor)),
-    MightVectorize = StorageOrdersAgree
-                  && (int(Derived::Flags) & int(OtherDerived::Flags) & ActualPacketAccessBit),
-    MayInnerVectorize  = MightVectorize && int(InnerSize)!=Dynamic && int(InnerSize)%int(PacketSize)==0
-                       && int(DstIsAligned) && int(SrcIsAligned),
-    MayLinearize = StorageOrdersAgree && (int(Derived::Flags) & int(OtherDerived::Flags) & LinearAccessBit),
-    MayLinearVectorize = MightVectorize && MayLinearize && DstHasDirectAccess
-                       && (DstIsAligned || MaxSizeAtCompileTime == Dynamic),
-      /* If the destination isn't aligned, we have to do runtime checks and we don't unroll,
-         so it's only good for large enough sizes. */
-    MaySliceVectorize  = MightVectorize && DstHasDirectAccess
-                       && (int(InnerMaxSize)==Dynamic || int(InnerMaxSize)>=3*PacketSize)
-      /* slice vectorization can be slow, so we only want it if the slices are big, which is
-         indicated by InnerMaxSize rather than InnerSize, think of the case of a dynamic block
-         in a fixed-size matrix */
-  };
-
-public:
-  enum {
-    Traversal = int(MayInnerVectorize)  ? int(InnerVectorizedTraversal)
-              : int(MayLinearVectorize) ? int(LinearVectorizedTraversal)
-              : int(MaySliceVectorize)  ? int(SliceVectorizedTraversal)
-              : int(MayLinearize)       ? int(LinearTraversal)
-                                        : int(DefaultTraversal),
-    Vectorized = int(Traversal) == InnerVectorizedTraversal
-              || int(Traversal) == LinearVectorizedTraversal
-              || int(Traversal) == SliceVectorizedTraversal
-  };
-
-private:
-  enum {
-    UnrollingLimit      = EIGEN_UNROLLING_LIMIT * (Vectorized ? int(PacketSize) : 1),
-    MayUnrollCompletely = int(Derived::SizeAtCompileTime) != Dynamic
-                       && int(OtherDerived::CoeffReadCost) != Dynamic
-                       && int(Derived::SizeAtCompileTime) * int(OtherDerived::CoeffReadCost) <= int(UnrollingLimit),
-    MayUnrollInner      = int(InnerSize) != Dynamic
-                       && int(OtherDerived::CoeffReadCost) != Dynamic
-                       && int(InnerSize) * int(OtherDerived::CoeffReadCost) <= int(UnrollingLimit)
-  };
-
-public:
-  enum {
-    Unrolling = (int(Traversal) == int(InnerVectorizedTraversal) || int(Traversal) == int(DefaultTraversal))
-                ? (
-                    int(MayUnrollCompletely) ? int(CompleteUnrolling)
-                  : int(MayUnrollInner)      ? int(InnerUnrolling)
-                                             : int(NoUnrolling)
-                  )
-              : int(Traversal) == int(LinearVectorizedTraversal)
-                ? ( bool(MayUnrollCompletely) && bool(DstIsAligned) ? int(CompleteUnrolling) : int(NoUnrolling) )
-              : int(Traversal) == int(LinearTraversal)
-                ? ( bool(MayUnrollCompletely) ? int(CompleteUnrolling) : int(NoUnrolling) )
-              : int(NoUnrolling)
-  };
-
-#ifdef EIGEN_DEBUG_ASSIGN
-  static void debug()
-  {
-    EIGEN_DEBUG_VAR(DstIsAligned)
-    EIGEN_DEBUG_VAR(SrcIsAligned)
-    EIGEN_DEBUG_VAR(JointAlignment)
-    EIGEN_DEBUG_VAR(InnerSize)
-    EIGEN_DEBUG_VAR(InnerMaxSize)
-    EIGEN_DEBUG_VAR(PacketSize)
-    EIGEN_DEBUG_VAR(StorageOrdersAgree)
-    EIGEN_DEBUG_VAR(MightVectorize)
-    EIGEN_DEBUG_VAR(MayLinearize)
-    EIGEN_DEBUG_VAR(MayInnerVectorize)
-    EIGEN_DEBUG_VAR(MayLinearVectorize)
-    EIGEN_DEBUG_VAR(MaySliceVectorize)
-    EIGEN_DEBUG_VAR(Traversal)
-    EIGEN_DEBUG_VAR(UnrollingLimit)
-    EIGEN_DEBUG_VAR(MayUnrollCompletely)
-    EIGEN_DEBUG_VAR(MayUnrollInner)
-    EIGEN_DEBUG_VAR(Unrolling)
-  }
-#endif
-};
-
-/***************************************************************************
-* Part 2 : meta-unrollers
-***************************************************************************/
-
-/************************
-*** Default traversal ***
-************************/
-
-template<typename Derived1, typename Derived2, int Index, int Stop>
-struct assign_DefaultTraversal_CompleteUnrolling
-{
-  enum {
-    outer = Index / Derived1::InnerSizeAtCompileTime,
-    inner = Index % Derived1::InnerSizeAtCompileTime
-  };
-
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    dst.copyCoeffByOuterInner(outer, inner, src);
-    assign_DefaultTraversal_CompleteUnrolling<Derived1, Derived2, Index+1, Stop>::run(dst, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Stop>
-struct assign_DefaultTraversal_CompleteUnrolling<Derived1, Derived2, Stop, Stop>
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &) {}
-};
-
-template<typename Derived1, typename Derived2, int Index, int Stop>
-struct assign_DefaultTraversal_InnerUnrolling
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src, typename Derived1::Index outer)
-  {
-    dst.copyCoeffByOuterInner(outer, Index, src);
-    assign_DefaultTraversal_InnerUnrolling<Derived1, Derived2, Index+1, Stop>::run(dst, src, outer);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Stop>
-struct assign_DefaultTraversal_InnerUnrolling<Derived1, Derived2, Stop, Stop>
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &, typename Derived1::Index) {}
-};
-
-/***********************
-*** Linear traversal ***
-***********************/
-
-template<typename Derived1, typename Derived2, int Index, int Stop>
-struct assign_LinearTraversal_CompleteUnrolling
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    dst.copyCoeff(Index, src);
-    assign_LinearTraversal_CompleteUnrolling<Derived1, Derived2, Index+1, Stop>::run(dst, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Stop>
-struct assign_LinearTraversal_CompleteUnrolling<Derived1, Derived2, Stop, Stop>
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &) {}
-};
-
-/**************************
-*** Inner vectorization ***
-**************************/
-
-template<typename Derived1, typename Derived2, int Index, int Stop>
-struct assign_innervec_CompleteUnrolling
-{
-  enum {
-    outer = Index / Derived1::InnerSizeAtCompileTime,
-    inner = Index % Derived1::InnerSizeAtCompileTime,
-    JointAlignment = assign_traits<Derived1,Derived2>::JointAlignment
-  };
-
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    dst.template copyPacketByOuterInner<Derived2, Aligned, JointAlignment>(outer, inner, src);
-    assign_innervec_CompleteUnrolling<Derived1, Derived2,
-      Index+packet_traits<typename Derived1::Scalar>::size, Stop>::run(dst, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Stop>
-struct assign_innervec_CompleteUnrolling<Derived1, Derived2, Stop, Stop>
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &) {}
-};
-
-template<typename Derived1, typename Derived2, int Index, int Stop>
-struct assign_innervec_InnerUnrolling
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src, typename Derived1::Index outer)
-  {
-    dst.template copyPacketByOuterInner<Derived2, Aligned, Aligned>(outer, Index, src);
-    assign_innervec_InnerUnrolling<Derived1, Derived2,
-      Index+packet_traits<typename Derived1::Scalar>::size, Stop>::run(dst, src, outer);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Stop>
-struct assign_innervec_InnerUnrolling<Derived1, Derived2, Stop, Stop>
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &, typename Derived1::Index) {}
-};
-
-/***************************************************************************
-* Part 3 : implementation of all cases
-***************************************************************************/
-
-template<typename Derived1, typename Derived2,
-         int Traversal = assign_traits<Derived1, Derived2>::Traversal,
-         int Unrolling = assign_traits<Derived1, Derived2>::Unrolling,
-         int Version = Specialized>
-struct assign_impl;
-
-/************************
-*** Default traversal ***
-************************/
-
-template<typename Derived1, typename Derived2, int Unrolling, int Version>
-struct assign_impl<Derived1, Derived2, InvalidTraversal, Unrolling, Version>
-{
-  static inline void run(Derived1 &, const Derived2 &) { }
-};
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, DefaultTraversal, NoUnrolling, Version>
-{
-  typedef typename Derived1::Index Index;
-  static inline void run(Derived1 &dst, const Derived2 &src)
-  {
-    const Index innerSize = dst.innerSize();
-    const Index outerSize = dst.outerSize();
-    for(Index outer = 0; outer < outerSize; ++outer)
-      for(Index inner = 0; inner < innerSize; ++inner)
-        dst.copyCoeffByOuterInner(outer, inner, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, DefaultTraversal, CompleteUnrolling, Version>
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    assign_DefaultTraversal_CompleteUnrolling<Derived1, Derived2, 0, Derived1::SizeAtCompileTime>
-      ::run(dst, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, DefaultTraversal, InnerUnrolling, Version>
-{
-  typedef typename Derived1::Index Index;
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    const Index outerSize = dst.outerSize();
-    for(Index outer = 0; outer < outerSize; ++outer)
-      assign_DefaultTraversal_InnerUnrolling<Derived1, Derived2, 0, Derived1::InnerSizeAtCompileTime>
-        ::run(dst, src, outer);
-  }
-};
-
-/***********************
-*** Linear traversal ***
-***********************/
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, LinearTraversal, NoUnrolling, Version>
-{
-  typedef typename Derived1::Index Index;
-  static inline void run(Derived1 &dst, const Derived2 &src)
-  {
-    const Index size = dst.size();
-    for(Index i = 0; i < size; ++i)
-      dst.copyCoeff(i, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, LinearTraversal, CompleteUnrolling, Version>
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    assign_LinearTraversal_CompleteUnrolling<Derived1, Derived2, 0, Derived1::SizeAtCompileTime>
-      ::run(dst, src);
-  }
-};
-
-/**************************
-*** Inner vectorization ***
-**************************/
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, InnerVectorizedTraversal, NoUnrolling, Version>
-{
-  typedef typename Derived1::Index Index;
-  static inline void run(Derived1 &dst, const Derived2 &src)
-  {
-    const Index innerSize = dst.innerSize();
-    const Index outerSize = dst.outerSize();
-    const Index packetSize = packet_traits<typename Derived1::Scalar>::size;
-    for(Index outer = 0; outer < outerSize; ++outer)
-      for(Index inner = 0; inner < innerSize; inner+=packetSize)
-        dst.template copyPacketByOuterInner<Derived2, Aligned, Aligned>(outer, inner, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, InnerVectorizedTraversal, CompleteUnrolling, Version>
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    assign_innervec_CompleteUnrolling<Derived1, Derived2, 0, Derived1::SizeAtCompileTime>
-      ::run(dst, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, InnerVectorizedTraversal, InnerUnrolling, Version>
-{
-  typedef typename Derived1::Index Index;
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    const Index outerSize = dst.outerSize();
-    for(Index outer = 0; outer < outerSize; ++outer)
-      assign_innervec_InnerUnrolling<Derived1, Derived2, 0, Derived1::InnerSizeAtCompileTime>
-        ::run(dst, src, outer);
-  }
-};
-
-/***************************
-*** Linear vectorization ***
-***************************/
-
-template <bool IsAligned = false>
-struct unaligned_assign_impl
-{
-  template <typename Derived, typename OtherDerived>
-  static EIGEN_STRONG_INLINE void run(const Derived&, OtherDerived&, typename Derived::Index, typename Derived::Index) {}
-};
-
-template <>
-struct unaligned_assign_impl<false>
-{
-  // MSVC must not inline this functions. If it does, it fails to optimize the
-  // packet access path.
-#ifdef _MSC_VER
-  template <typename Derived, typename OtherDerived>
-  static EIGEN_DONT_INLINE void run(const Derived& src, OtherDerived& dst, typename Derived::Index start, typename Derived::Index end)
-#else
-  template <typename Derived, typename OtherDerived>
-  static EIGEN_STRONG_INLINE void run(const Derived& src, OtherDerived& dst, typename Derived::Index start, typename Derived::Index end)
-#endif
-  {
-    for (typename Derived::Index index = start; index < end; ++index)
-      dst.copyCoeff(index, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, LinearVectorizedTraversal, NoUnrolling, Version>
-{
-  typedef typename Derived1::Index Index;
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    const Index size = dst.size();
-    typedef packet_traits<typename Derived1::Scalar> PacketTraits;
-    enum {
-      packetSize = PacketTraits::size,
-      dstAlignment = PacketTraits::AlignedOnScalar ? Aligned : int(assign_traits<Derived1,Derived2>::DstIsAligned) ,
-      srcAlignment = assign_traits<Derived1,Derived2>::JointAlignment
-    };
-    const Index alignedStart = assign_traits<Derived1,Derived2>::DstIsAligned ? 0
-                             : internal::first_aligned(&dst.coeffRef(0), size);
-    const Index alignedEnd = alignedStart + ((size-alignedStart)/packetSize)*packetSize;
-
-    unaligned_assign_impl<assign_traits<Derived1,Derived2>::DstIsAligned!=0>::run(src,dst,0,alignedStart);
-
-    for(Index index = alignedStart; index < alignedEnd; index += packetSize)
-    {
-      dst.template copyPacket<Derived2, dstAlignment, srcAlignment>(index, src);
-    }
-
-    unaligned_assign_impl<>::run(src,dst,alignedEnd,size);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, LinearVectorizedTraversal, CompleteUnrolling, Version>
-{
-  typedef typename Derived1::Index Index;
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    enum { size = Derived1::SizeAtCompileTime,
-           packetSize = packet_traits<typename Derived1::Scalar>::size,
-           alignedSize = (size/packetSize)*packetSize };
-
-    assign_innervec_CompleteUnrolling<Derived1, Derived2, 0, alignedSize>::run(dst, src);
-    assign_DefaultTraversal_CompleteUnrolling<Derived1, Derived2, alignedSize, size>::run(dst, src);
-  }
-};
-
-/**************************
-*** Slice vectorization ***
-***************************/
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, SliceVectorizedTraversal, NoUnrolling, Version>
-{
-  typedef typename Derived1::Index Index;
-  static inline void run(Derived1 &dst, const Derived2 &src)
-  {
-    typedef typename Derived1::Scalar Scalar;
-    typedef packet_traits<Scalar> PacketTraits;
-    enum {
-      packetSize = PacketTraits::size,
-      alignable = PacketTraits::AlignedOnScalar,
-      dstIsAligned = assign_traits<Derived1,Derived2>::DstIsAligned,
-      dstAlignment = alignable ? Aligned : int(dstIsAligned),
-      srcAlignment = assign_traits<Derived1,Derived2>::JointAlignment
-    };
-    const Scalar *dst_ptr = &dst.coeffRef(0,0);
-    if((!bool(dstIsAligned)) && (size_t(dst_ptr) % sizeof(Scalar))>0)
-    {
-      // the pointer is not aligend-on scalar, so alignment is not possible
-      return assign_impl<Derived1,Derived2,DefaultTraversal,NoUnrolling>::run(dst, src);
-    }
-    const Index packetAlignedMask = packetSize - 1;
-    const Index innerSize = dst.innerSize();
-    const Index outerSize = dst.outerSize();
-    const Index alignedStep = alignable ? (packetSize - dst.outerStride() % packetSize) & packetAlignedMask : 0;
-    Index alignedStart = ((!alignable) || bool(dstIsAligned)) ? 0 : internal::first_aligned(dst_ptr, innerSize);
-
-    for(Index outer = 0; outer < outerSize; ++outer)
-    {
-      const Index alignedEnd = alignedStart + ((innerSize-alignedStart) & ~packetAlignedMask);
-      // do the non-vectorizable part of the assignment
-      for(Index inner = 0; inner<alignedStart ; ++inner)
-        dst.copyCoeffByOuterInner(outer, inner, src);
-
-      // do the vectorizable part of the assignment
-      for(Index inner = alignedStart; inner<alignedEnd; inner+=packetSize)
-        dst.template copyPacketByOuterInner<Derived2, dstAlignment, Unaligned>(outer, inner, src);
-
-      // do the non-vectorizable part of the assignment
-      for(Index inner = alignedEnd; inner<innerSize ; ++inner)
-        dst.copyCoeffByOuterInner(outer, inner, src);
-
-      alignedStart = std::min<Index>((alignedStart+alignedStep)%packetSize, innerSize);
-    }
-  }
-};
-
-} // end namespace internal
-
-/***************************************************************************
-* Part 4 : implementation of DenseBase methods
-***************************************************************************/
-
 template<typename Derived>
 template<typename OtherDerived>
-EIGEN_STRONG_INLINE Derived& DenseBase<Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>
   ::lazyAssign(const DenseBase<OtherDerived>& other)
 {
   enum{
@@ -499,90 +27,62 @@ EIGEN_STRONG_INLINE Derived& DenseBase<Derived>
   EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Derived,OtherDerived)
   EIGEN_STATIC_ASSERT(SameType,YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
 
-#ifdef EIGEN_DEBUG_ASSIGN
-  internal::assign_traits<Derived, OtherDerived>::debug();
-#endif
   eigen_assert(rows() == other.rows() && cols() == other.cols());
-  internal::assign_impl<Derived, OtherDerived, int(SameType) ? int(internal::assign_traits<Derived, OtherDerived>::Traversal)
-                                                       : int(InvalidTraversal)>::run(derived(),other.derived());
-#ifndef EIGEN_NO_DEBUG
-  checkTransposeAliasing(other.derived());
-#endif
+  internal::call_assignment_no_alias(derived(),other.derived());
+  
   return derived();
 }
 
-namespace internal {
-
-template<typename Derived, typename OtherDerived,
-         bool EvalBeforeAssigning = (int(internal::traits<OtherDerived>::Flags) & EvalBeforeAssigningBit) != 0,
-         bool NeedToTranspose = ((int(Derived::RowsAtCompileTime) == 1 && int(OtherDerived::ColsAtCompileTime) == 1)
-                              |   // FIXME | instead of || to please GCC 4.4.0 stupid warning "suggest parentheses around &&".
-                                  // revert to || as soon as not needed anymore.
-                                  (int(Derived::ColsAtCompileTime) == 1 && int(OtherDerived::RowsAtCompileTime) == 1))
-                              && int(Derived::SizeAtCompileTime) != 1>
-struct assign_selector;
-
-template<typename Derived, typename OtherDerived>
-struct assign_selector<Derived,OtherDerived,false,false> {
-  static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.derived()); }
-  template<typename ActualDerived, typename ActualOtherDerived>
-  static EIGEN_STRONG_INLINE Derived& evalTo(ActualDerived& dst, const ActualOtherDerived& other) { other.evalTo(dst); return dst; }
-};
-template<typename Derived, typename OtherDerived>
-struct assign_selector<Derived,OtherDerived,true,false> {
-  static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.eval()); }
-};
-template<typename Derived, typename OtherDerived>
-struct assign_selector<Derived,OtherDerived,false,true> {
-  static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.transpose()); }
-  template<typename ActualDerived, typename ActualOtherDerived>
-  static EIGEN_STRONG_INLINE Derived& evalTo(ActualDerived& dst, const ActualOtherDerived& other) { Transpose<ActualDerived> dstTrans(dst); other.evalTo(dstTrans); return dst; }
-};
-template<typename Derived, typename OtherDerived>
-struct assign_selector<Derived,OtherDerived,true,true> {
-  static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.transpose().eval()); }
-};
-
-} // end namespace internal
-
 template<typename Derived>
 template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
 EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator=(const DenseBase<OtherDerived>& other)
 {
-  return internal::assign_selector<Derived,OtherDerived>::run(derived(), other.derived());
+  internal::call_assignment(derived(), other.derived());
+  return derived();
 }
 
 template<typename Derived>
+EIGEN_DEVICE_FUNC
 EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator=(const DenseBase& other)
 {
-  return internal::assign_selector<Derived,Derived>::run(derived(), other.derived());
+  internal::call_assignment(derived(), other.derived());
+  return derived();
 }
 
 template<typename Derived>
+EIGEN_DEVICE_FUNC
 EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const MatrixBase& other)
 {
-  return internal::assign_selector<Derived,Derived>::run(derived(), other.derived());
+  internal::call_assignment(derived(), other.derived());
+  return derived();
 }
 
 template<typename Derived>
 template <typename OtherDerived>
+EIGEN_DEVICE_FUNC
 EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const DenseBase<OtherDerived>& other)
 {
-  return internal::assign_selector<Derived,OtherDerived>::run(derived(), other.derived());
+  internal::call_assignment(derived(), other.derived());
+  return derived();
 }
 
 template<typename Derived>
 template <typename OtherDerived>
+EIGEN_DEVICE_FUNC
 EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const EigenBase<OtherDerived>& other)
 {
-  return internal::assign_selector<Derived,OtherDerived,false>::evalTo(derived(), other.derived());
+  internal::call_assignment(derived(), other.derived());
+  return derived();
 }
 
 template<typename Derived>
 template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
 EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const ReturnByValue<OtherDerived>& other)
 {
-  return internal::assign_selector<Derived,OtherDerived,false>::evalTo(derived(), other.derived());
+  other.derived().evalTo(derived());
+  return derived();
 }
 
 } // end namespace Eigen
diff --git a/eigen/Eigen/src/Core/AssignEvaluator.h b/eigen/Eigen/src/Core/AssignEvaluator.h
new file mode 100644
index 0000000..b0ec7b7
--- /dev/null
+++ b/eigen/Eigen/src/Core/AssignEvaluator.h
@@ -0,0 +1,935 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2011 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2011-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2011-2012 Jitse Niesen <jitse@maths.leeds.ac.uk>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ASSIGN_EVALUATOR_H
+#define EIGEN_ASSIGN_EVALUATOR_H
+
+namespace Eigen {
+
+// This implementation is based on Assign.h
+
+namespace internal {
+  
+/***************************************************************************
+* Part 1 : the logic deciding a strategy for traversal and unrolling       *
+***************************************************************************/
+
+// copy_using_evaluator_traits is based on assign_traits
+
+template <typename DstEvaluator, typename SrcEvaluator, typename AssignFunc>
+struct copy_using_evaluator_traits
+{
+  typedef typename DstEvaluator::XprType Dst;
+  typedef typename Dst::Scalar DstScalar;
+  
+  enum {
+    DstFlags = DstEvaluator::Flags,
+    SrcFlags = SrcEvaluator::Flags
+  };
+  
+public:
+  enum {
+    DstAlignment = DstEvaluator::Alignment,
+    SrcAlignment = SrcEvaluator::Alignment,
+    DstHasDirectAccess = DstFlags & DirectAccessBit,
+    JointAlignment = EIGEN_PLAIN_ENUM_MIN(DstAlignment,SrcAlignment)
+  };
+
+private:
+  enum {
+    InnerSize = int(Dst::IsVectorAtCompileTime) ? int(Dst::SizeAtCompileTime)
+              : int(DstFlags)&RowMajorBit ? int(Dst::ColsAtCompileTime)
+              : int(Dst::RowsAtCompileTime),
+    InnerMaxSize = int(Dst::IsVectorAtCompileTime) ? int(Dst::MaxSizeAtCompileTime)
+              : int(DstFlags)&RowMajorBit ? int(Dst::MaxColsAtCompileTime)
+              : int(Dst::MaxRowsAtCompileTime),
+    OuterStride = int(outer_stride_at_compile_time<Dst>::ret),
+    MaxSizeAtCompileTime = Dst::SizeAtCompileTime
+  };
+
+  // TODO distinguish between linear traversal and inner-traversals
+  typedef typename find_best_packet<DstScalar,Dst::SizeAtCompileTime>::type LinearPacketType;
+  typedef typename find_best_packet<DstScalar,InnerSize>::type InnerPacketType;
+
+  enum {
+    LinearPacketSize = unpacket_traits<LinearPacketType>::size,
+    InnerPacketSize = unpacket_traits<InnerPacketType>::size
+  };
+
+public:
+  enum {
+    LinearRequiredAlignment = unpacket_traits<LinearPacketType>::alignment,
+    InnerRequiredAlignment = unpacket_traits<InnerPacketType>::alignment
+  };
+
+private:
+  enum {
+    DstIsRowMajor = DstFlags&RowMajorBit,
+    SrcIsRowMajor = SrcFlags&RowMajorBit,
+    StorageOrdersAgree = (int(DstIsRowMajor) == int(SrcIsRowMajor)),
+    MightVectorize = bool(StorageOrdersAgree)
+                  && (int(DstFlags) & int(SrcFlags) & ActualPacketAccessBit)
+                  && bool(functor_traits<AssignFunc>::PacketAccess),
+    MayInnerVectorize  = MightVectorize
+                       && int(InnerSize)!=Dynamic && int(InnerSize)%int(InnerPacketSize)==0
+                       && int(OuterStride)!=Dynamic && int(OuterStride)%int(InnerPacketSize)==0
+                       && (EIGEN_UNALIGNED_VECTORIZE  || int(JointAlignment)>=int(InnerRequiredAlignment)),
+    MayLinearize = bool(StorageOrdersAgree) && (int(DstFlags) & int(SrcFlags) & LinearAccessBit),
+    MayLinearVectorize = bool(MightVectorize) && MayLinearize && DstHasDirectAccess
+                       && (EIGEN_UNALIGNED_VECTORIZE || (int(DstAlignment)>=int(LinearRequiredAlignment)) || MaxSizeAtCompileTime == Dynamic),
+      /* If the destination isn't aligned, we have to do runtime checks and we don't unroll,
+         so it's only good for large enough sizes. */
+    MaySliceVectorize  = bool(MightVectorize) && bool(DstHasDirectAccess)
+                       && (int(InnerMaxSize)==Dynamic || int(InnerMaxSize)>=(EIGEN_UNALIGNED_VECTORIZE?InnerPacketSize:(3*InnerPacketSize)))
+      /* slice vectorization can be slow, so we only want it if the slices are big, which is
+         indicated by InnerMaxSize rather than InnerSize, think of the case of a dynamic block
+         in a fixed-size matrix
+         However, with EIGEN_UNALIGNED_VECTORIZE and unrolling, slice vectorization is still worth it */
+  };
+
+public:
+  enum {
+    Traversal = int(MayLinearVectorize) && (LinearPacketSize>InnerPacketSize) ? int(LinearVectorizedTraversal)
+              : int(MayInnerVectorize)   ? int(InnerVectorizedTraversal)
+              : int(MayLinearVectorize)  ? int(LinearVectorizedTraversal)
+              : int(MaySliceVectorize)   ? int(SliceVectorizedTraversal)
+              : int(MayLinearize)        ? int(LinearTraversal)
+                                         : int(DefaultTraversal),
+    Vectorized = int(Traversal) == InnerVectorizedTraversal
+              || int(Traversal) == LinearVectorizedTraversal
+              || int(Traversal) == SliceVectorizedTraversal
+  };
+
+  typedef typename conditional<int(Traversal)==LinearVectorizedTraversal, LinearPacketType, InnerPacketType>::type PacketType;
+
+private:
+  enum {
+    ActualPacketSize    = int(Traversal)==LinearVectorizedTraversal ? LinearPacketSize
+                        : Vectorized ? InnerPacketSize
+                        : 1,
+    UnrollingLimit      = EIGEN_UNROLLING_LIMIT * ActualPacketSize,
+    MayUnrollCompletely = int(Dst::SizeAtCompileTime) != Dynamic
+                       && int(Dst::SizeAtCompileTime) * (int(DstEvaluator::CoeffReadCost)+int(SrcEvaluator::CoeffReadCost)) <= int(UnrollingLimit),
+    MayUnrollInner      = int(InnerSize) != Dynamic
+                       && int(InnerSize) * (int(DstEvaluator::CoeffReadCost)+int(SrcEvaluator::CoeffReadCost)) <= int(UnrollingLimit)
+  };
+
+public:
+  enum {
+    Unrolling = (int(Traversal) == int(InnerVectorizedTraversal) || int(Traversal) == int(DefaultTraversal))
+                ? (
+                    int(MayUnrollCompletely) ? int(CompleteUnrolling)
+                  : int(MayUnrollInner)      ? int(InnerUnrolling)
+                                             : int(NoUnrolling)
+                  )
+              : int(Traversal) == int(LinearVectorizedTraversal)
+                ? ( bool(MayUnrollCompletely) && ( EIGEN_UNALIGNED_VECTORIZE || (int(DstAlignment)>=int(LinearRequiredAlignment)))
+                          ? int(CompleteUnrolling)
+                          : int(NoUnrolling) )
+              : int(Traversal) == int(LinearTraversal)
+                ? ( bool(MayUnrollCompletely) ? int(CompleteUnrolling) 
+                                              : int(NoUnrolling) )
+#if EIGEN_UNALIGNED_VECTORIZE
+              : int(Traversal) == int(SliceVectorizedTraversal)
+                ? ( bool(MayUnrollInner) ? int(InnerUnrolling)
+                                         : int(NoUnrolling) )
+#endif
+              : int(NoUnrolling)
+  };
+
+#ifdef EIGEN_DEBUG_ASSIGN
+  static void debug()
+  {
+    std::cerr << "DstXpr: " << typeid(typename DstEvaluator::XprType).name() << std::endl;
+    std::cerr << "SrcXpr: " << typeid(typename SrcEvaluator::XprType).name() << std::endl;
+    std::cerr.setf(std::ios::hex, std::ios::basefield);
+    std::cerr << "DstFlags" << " = " << DstFlags << " (" << demangle_flags(DstFlags) << " )" << std::endl;
+    std::cerr << "SrcFlags" << " = " << SrcFlags << " (" << demangle_flags(SrcFlags) << " )" << std::endl;
+    std::cerr.unsetf(std::ios::hex);
+    EIGEN_DEBUG_VAR(DstAlignment)
+    EIGEN_DEBUG_VAR(SrcAlignment)
+    EIGEN_DEBUG_VAR(LinearRequiredAlignment)
+    EIGEN_DEBUG_VAR(InnerRequiredAlignment)
+    EIGEN_DEBUG_VAR(JointAlignment)
+    EIGEN_DEBUG_VAR(InnerSize)
+    EIGEN_DEBUG_VAR(InnerMaxSize)
+    EIGEN_DEBUG_VAR(LinearPacketSize)
+    EIGEN_DEBUG_VAR(InnerPacketSize)
+    EIGEN_DEBUG_VAR(ActualPacketSize)
+    EIGEN_DEBUG_VAR(StorageOrdersAgree)
+    EIGEN_DEBUG_VAR(MightVectorize)
+    EIGEN_DEBUG_VAR(MayLinearize)
+    EIGEN_DEBUG_VAR(MayInnerVectorize)
+    EIGEN_DEBUG_VAR(MayLinearVectorize)
+    EIGEN_DEBUG_VAR(MaySliceVectorize)
+    std::cerr << "Traversal" << " = " << Traversal << " (" << demangle_traversal(Traversal) << ")" << std::endl;
+    EIGEN_DEBUG_VAR(SrcEvaluator::CoeffReadCost)
+    EIGEN_DEBUG_VAR(UnrollingLimit)
+    EIGEN_DEBUG_VAR(MayUnrollCompletely)
+    EIGEN_DEBUG_VAR(MayUnrollInner)
+    std::cerr << "Unrolling" << " = " << Unrolling << " (" << demangle_unrolling(Unrolling) << ")" << std::endl;
+    std::cerr << std::endl;
+  }
+#endif
+};
+
+/***************************************************************************
+* Part 2 : meta-unrollers
+***************************************************************************/
+
+/************************
+*** Default traversal ***
+************************/
+
+template<typename Kernel, int Index, int Stop>
+struct copy_using_evaluator_DefaultTraversal_CompleteUnrolling
+{
+  // FIXME: this is not very clean, perhaps this information should be provided by the kernel?
+  typedef typename Kernel::DstEvaluatorType DstEvaluatorType;
+  typedef typename DstEvaluatorType::XprType DstXprType;
+  
+  enum {
+    outer = Index / DstXprType::InnerSizeAtCompileTime,
+    inner = Index % DstXprType::InnerSizeAtCompileTime
+  };
+
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    kernel.assignCoeffByOuterInner(outer, inner);
+    copy_using_evaluator_DefaultTraversal_CompleteUnrolling<Kernel, Index+1, Stop>::run(kernel);
+  }
+};
+
+template<typename Kernel, int Stop>
+struct copy_using_evaluator_DefaultTraversal_CompleteUnrolling<Kernel, Stop, Stop>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&) { }
+};
+
+template<typename Kernel, int Index_, int Stop>
+struct copy_using_evaluator_DefaultTraversal_InnerUnrolling
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel, Index outer)
+  {
+    kernel.assignCoeffByOuterInner(outer, Index_);
+    copy_using_evaluator_DefaultTraversal_InnerUnrolling<Kernel, Index_+1, Stop>::run(kernel, outer);
+  }
+};
+
+template<typename Kernel, int Stop>
+struct copy_using_evaluator_DefaultTraversal_InnerUnrolling<Kernel, Stop, Stop>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&, Index) { }
+};
+
+/***********************
+*** Linear traversal ***
+***********************/
+
+template<typename Kernel, int Index, int Stop>
+struct copy_using_evaluator_LinearTraversal_CompleteUnrolling
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel& kernel)
+  {
+    kernel.assignCoeff(Index);
+    copy_using_evaluator_LinearTraversal_CompleteUnrolling<Kernel, Index+1, Stop>::run(kernel);
+  }
+};
+
+template<typename Kernel, int Stop>
+struct copy_using_evaluator_LinearTraversal_CompleteUnrolling<Kernel, Stop, Stop>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&) { }
+};
+
+/**************************
+*** Inner vectorization ***
+**************************/
+
+template<typename Kernel, int Index, int Stop>
+struct copy_using_evaluator_innervec_CompleteUnrolling
+{
+  // FIXME: this is not very clean, perhaps this information should be provided by the kernel?
+  typedef typename Kernel::DstEvaluatorType DstEvaluatorType;
+  typedef typename DstEvaluatorType::XprType DstXprType;
+  typedef typename Kernel::PacketType PacketType;
+  
+  enum {
+    outer = Index / DstXprType::InnerSizeAtCompileTime,
+    inner = Index % DstXprType::InnerSizeAtCompileTime,
+    SrcAlignment = Kernel::AssignmentTraits::SrcAlignment,
+    DstAlignment = Kernel::AssignmentTraits::DstAlignment
+  };
+
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    kernel.template assignPacketByOuterInner<DstAlignment, SrcAlignment, PacketType>(outer, inner);
+    enum { NextIndex = Index + unpacket_traits<PacketType>::size };
+    copy_using_evaluator_innervec_CompleteUnrolling<Kernel, NextIndex, Stop>::run(kernel);
+  }
+};
+
+template<typename Kernel, int Stop>
+struct copy_using_evaluator_innervec_CompleteUnrolling<Kernel, Stop, Stop>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&) { }
+};
+
+template<typename Kernel, int Index_, int Stop, int SrcAlignment, int DstAlignment>
+struct copy_using_evaluator_innervec_InnerUnrolling
+{
+  typedef typename Kernel::PacketType PacketType;
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel, Index outer)
+  {
+    kernel.template assignPacketByOuterInner<DstAlignment, SrcAlignment, PacketType>(outer, Index_);
+    enum { NextIndex = Index_ + unpacket_traits<PacketType>::size };
+    copy_using_evaluator_innervec_InnerUnrolling<Kernel, NextIndex, Stop, SrcAlignment, DstAlignment>::run(kernel, outer);
+  }
+};
+
+template<typename Kernel, int Stop, int SrcAlignment, int DstAlignment>
+struct copy_using_evaluator_innervec_InnerUnrolling<Kernel, Stop, Stop, SrcAlignment, DstAlignment>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &, Index) { }
+};
+
+/***************************************************************************
+* Part 3 : implementation of all cases
+***************************************************************************/
+
+// dense_assignment_loop is based on assign_impl
+
+template<typename Kernel,
+         int Traversal = Kernel::AssignmentTraits::Traversal,
+         int Unrolling = Kernel::AssignmentTraits::Unrolling>
+struct dense_assignment_loop;
+
+/************************
+*** Default traversal ***
+************************/
+
+template<typename Kernel>
+struct dense_assignment_loop<Kernel, DefaultTraversal, NoUnrolling>
+{
+  EIGEN_DEVICE_FUNC static void EIGEN_STRONG_INLINE run(Kernel &kernel)
+  {
+    for(Index outer = 0; outer < kernel.outerSize(); ++outer) {
+      for(Index inner = 0; inner < kernel.innerSize(); ++inner) {
+        kernel.assignCoeffByOuterInner(outer, inner);
+      }
+    }
+  }
+};
+
+template<typename Kernel>
+struct dense_assignment_loop<Kernel, DefaultTraversal, CompleteUnrolling>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
+    copy_using_evaluator_DefaultTraversal_CompleteUnrolling<Kernel, 0, DstXprType::SizeAtCompileTime>::run(kernel);
+  }
+};
+
+template<typename Kernel>
+struct dense_assignment_loop<Kernel, DefaultTraversal, InnerUnrolling>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
+
+    const Index outerSize = kernel.outerSize();
+    for(Index outer = 0; outer < outerSize; ++outer)
+      copy_using_evaluator_DefaultTraversal_InnerUnrolling<Kernel, 0, DstXprType::InnerSizeAtCompileTime>::run(kernel, outer);
+  }
+};
+
+/***************************
+*** Linear vectorization ***
+***************************/
+
+
+// The goal of unaligned_dense_assignment_loop is simply to factorize the handling
+// of the non vectorizable beginning and ending parts
+
+template <bool IsAligned = false>
+struct unaligned_dense_assignment_loop
+{
+  // if IsAligned = true, then do nothing
+  template <typename Kernel>
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&, Index, Index) {}
+};
+
+template <>
+struct unaligned_dense_assignment_loop<false>
+{
+  // MSVC must not inline this functions. If it does, it fails to optimize the
+  // packet access path.
+  // FIXME check which version exhibits this issue
+#if EIGEN_COMP_MSVC
+  template <typename Kernel>
+  static EIGEN_DONT_INLINE void run(Kernel &kernel,
+                                    Index start,
+                                    Index end)
+#else
+  template <typename Kernel>
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel,
+                                      Index start,
+                                      Index end)
+#endif
+  {
+    for (Index index = start; index < end; ++index)
+      kernel.assignCoeff(index);
+  }
+};
+
+template<typename Kernel>
+struct dense_assignment_loop<Kernel, LinearVectorizedTraversal, NoUnrolling>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    const Index size = kernel.size();
+    typedef typename Kernel::Scalar Scalar;
+    typedef typename Kernel::PacketType PacketType;
+    enum {
+      requestedAlignment = Kernel::AssignmentTraits::LinearRequiredAlignment,
+      packetSize = unpacket_traits<PacketType>::size,
+      dstIsAligned = int(Kernel::AssignmentTraits::DstAlignment)>=int(requestedAlignment),
+      dstAlignment = packet_traits<Scalar>::AlignedOnScalar ? int(requestedAlignment)
+                                                            : int(Kernel::AssignmentTraits::DstAlignment),
+      srcAlignment = Kernel::AssignmentTraits::JointAlignment
+    };
+    const Index alignedStart = dstIsAligned ? 0 : internal::first_aligned<requestedAlignment>(kernel.dstDataPtr(), size);
+    const Index alignedEnd = alignedStart + ((size-alignedStart)/packetSize)*packetSize;
+
+    unaligned_dense_assignment_loop<dstIsAligned!=0>::run(kernel, 0, alignedStart);
+
+    for(Index index = alignedStart; index < alignedEnd; index += packetSize)
+      kernel.template assignPacket<dstAlignment, srcAlignment, PacketType>(index);
+
+    unaligned_dense_assignment_loop<>::run(kernel, alignedEnd, size);
+  }
+};
+
+template<typename Kernel>
+struct dense_assignment_loop<Kernel, LinearVectorizedTraversal, CompleteUnrolling>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
+    typedef typename Kernel::PacketType PacketType;
+    
+    enum { size = DstXprType::SizeAtCompileTime,
+           packetSize =unpacket_traits<PacketType>::size,
+           alignedSize = (size/packetSize)*packetSize };
+
+    copy_using_evaluator_innervec_CompleteUnrolling<Kernel, 0, alignedSize>::run(kernel);
+    copy_using_evaluator_DefaultTraversal_CompleteUnrolling<Kernel, alignedSize, size>::run(kernel);
+  }
+};
+
+/**************************
+*** Inner vectorization ***
+**************************/
+
+template<typename Kernel>
+struct dense_assignment_loop<Kernel, InnerVectorizedTraversal, NoUnrolling>
+{
+  typedef typename Kernel::PacketType PacketType;
+  enum {
+    SrcAlignment = Kernel::AssignmentTraits::SrcAlignment,
+    DstAlignment = Kernel::AssignmentTraits::DstAlignment
+  };
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    const Index innerSize = kernel.innerSize();
+    const Index outerSize = kernel.outerSize();
+    const Index packetSize = unpacket_traits<PacketType>::size;
+    for(Index outer = 0; outer < outerSize; ++outer)
+      for(Index inner = 0; inner < innerSize; inner+=packetSize)
+        kernel.template assignPacketByOuterInner<DstAlignment, SrcAlignment, PacketType>(outer, inner);
+  }
+};
+
+template<typename Kernel>
+struct dense_assignment_loop<Kernel, InnerVectorizedTraversal, CompleteUnrolling>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
+    copy_using_evaluator_innervec_CompleteUnrolling<Kernel, 0, DstXprType::SizeAtCompileTime>::run(kernel);
+  }
+};
+
+template<typename Kernel>
+struct dense_assignment_loop<Kernel, InnerVectorizedTraversal, InnerUnrolling>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
+    typedef typename Kernel::AssignmentTraits Traits;
+    const Index outerSize = kernel.outerSize();
+    for(Index outer = 0; outer < outerSize; ++outer)
+      copy_using_evaluator_innervec_InnerUnrolling<Kernel, 0, DstXprType::InnerSizeAtCompileTime,
+                                                   Traits::SrcAlignment, Traits::DstAlignment>::run(kernel, outer);
+  }
+};
+
+/***********************
+*** Linear traversal ***
+***********************/
+
+template<typename Kernel>
+struct dense_assignment_loop<Kernel, LinearTraversal, NoUnrolling>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    const Index size = kernel.size();
+    for(Index i = 0; i < size; ++i)
+      kernel.assignCoeff(i);
+  }
+};
+
+template<typename Kernel>
+struct dense_assignment_loop<Kernel, LinearTraversal, CompleteUnrolling>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
+    copy_using_evaluator_LinearTraversal_CompleteUnrolling<Kernel, 0, DstXprType::SizeAtCompileTime>::run(kernel);
+  }
+};
+
+/**************************
+*** Slice vectorization ***
+***************************/
+
+template<typename Kernel>
+struct dense_assignment_loop<Kernel, SliceVectorizedTraversal, NoUnrolling>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    typedef typename Kernel::Scalar Scalar;
+    typedef typename Kernel::PacketType PacketType;
+    enum {
+      packetSize = unpacket_traits<PacketType>::size,
+      requestedAlignment = int(Kernel::AssignmentTraits::InnerRequiredAlignment),
+      alignable = packet_traits<Scalar>::AlignedOnScalar || int(Kernel::AssignmentTraits::DstAlignment)>=sizeof(Scalar),
+      dstIsAligned = int(Kernel::AssignmentTraits::DstAlignment)>=int(requestedAlignment),
+      dstAlignment = alignable ? int(requestedAlignment)
+                               : int(Kernel::AssignmentTraits::DstAlignment)
+    };
+    const Scalar *dst_ptr = kernel.dstDataPtr();
+    if((!bool(dstIsAligned)) && (UIntPtr(dst_ptr) % sizeof(Scalar))>0)
+    {
+      // the pointer is not aligend-on scalar, so alignment is not possible
+      return dense_assignment_loop<Kernel,DefaultTraversal,NoUnrolling>::run(kernel);
+    }
+    const Index packetAlignedMask = packetSize - 1;
+    const Index innerSize = kernel.innerSize();
+    const Index outerSize = kernel.outerSize();
+    const Index alignedStep = alignable ? (packetSize - kernel.outerStride() % packetSize) & packetAlignedMask : 0;
+    Index alignedStart = ((!alignable) || bool(dstIsAligned)) ? 0 : internal::first_aligned<requestedAlignment>(dst_ptr, innerSize);
+
+    for(Index outer = 0; outer < outerSize; ++outer)
+    {
+      const Index alignedEnd = alignedStart + ((innerSize-alignedStart) & ~packetAlignedMask);
+      // do the non-vectorizable part of the assignment
+      for(Index inner = 0; inner<alignedStart ; ++inner)
+        kernel.assignCoeffByOuterInner(outer, inner);
+
+      // do the vectorizable part of the assignment
+      for(Index inner = alignedStart; inner<alignedEnd; inner+=packetSize)
+        kernel.template assignPacketByOuterInner<dstAlignment, Unaligned, PacketType>(outer, inner);
+
+      // do the non-vectorizable part of the assignment
+      for(Index inner = alignedEnd; inner<innerSize ; ++inner)
+        kernel.assignCoeffByOuterInner(outer, inner);
+
+      alignedStart = numext::mini((alignedStart+alignedStep)%packetSize, innerSize);
+    }
+  }
+};
+
+#if EIGEN_UNALIGNED_VECTORIZE
+template<typename Kernel>
+struct dense_assignment_loop<Kernel, SliceVectorizedTraversal, InnerUnrolling>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
+    typedef typename Kernel::PacketType PacketType;
+
+    enum { size = DstXprType::InnerSizeAtCompileTime,
+           packetSize =unpacket_traits<PacketType>::size,
+           vectorizableSize = (size/packetSize)*packetSize };
+
+    for(Index outer = 0; outer < kernel.outerSize(); ++outer)
+    {
+      copy_using_evaluator_innervec_InnerUnrolling<Kernel, 0, vectorizableSize, 0, 0>::run(kernel, outer);
+      copy_using_evaluator_DefaultTraversal_InnerUnrolling<Kernel, vectorizableSize, size>::run(kernel, outer);
+    }
+  }
+};
+#endif
+
+
+/***************************************************************************
+* Part 4 : Generic dense assignment kernel
+***************************************************************************/
+
+// This class generalize the assignment of a coefficient (or packet) from one dense evaluator
+// to another dense writable evaluator.
+// It is parametrized by the two evaluators, and the actual assignment functor.
+// This abstraction level permits to keep the evaluation loops as simple and as generic as possible.
+// One can customize the assignment using this generic dense_assignment_kernel with different
+// functors, or by completely overloading it, by-passing a functor.
+template<typename DstEvaluatorTypeT, typename SrcEvaluatorTypeT, typename Functor, int Version = Specialized>
+class generic_dense_assignment_kernel
+{
+protected:
+  typedef typename DstEvaluatorTypeT::XprType DstXprType;
+  typedef typename SrcEvaluatorTypeT::XprType SrcXprType;
+public:
+  
+  typedef DstEvaluatorTypeT DstEvaluatorType;
+  typedef SrcEvaluatorTypeT SrcEvaluatorType;
+  typedef typename DstEvaluatorType::Scalar Scalar;
+  typedef copy_using_evaluator_traits<DstEvaluatorTypeT, SrcEvaluatorTypeT, Functor> AssignmentTraits;
+  typedef typename AssignmentTraits::PacketType PacketType;
+  
+  
+  EIGEN_DEVICE_FUNC generic_dense_assignment_kernel(DstEvaluatorType &dst, const SrcEvaluatorType &src, const Functor &func, DstXprType& dstExpr)
+    : m_dst(dst), m_src(src), m_functor(func), m_dstExpr(dstExpr)
+  {
+    #ifdef EIGEN_DEBUG_ASSIGN
+    AssignmentTraits::debug();
+    #endif
+  }
+  
+  EIGEN_DEVICE_FUNC Index size() const        { return m_dstExpr.size(); }
+  EIGEN_DEVICE_FUNC Index innerSize() const   { return m_dstExpr.innerSize(); }
+  EIGEN_DEVICE_FUNC Index outerSize() const   { return m_dstExpr.outerSize(); }
+  EIGEN_DEVICE_FUNC Index rows() const        { return m_dstExpr.rows(); }
+  EIGEN_DEVICE_FUNC Index cols() const        { return m_dstExpr.cols(); }
+  EIGEN_DEVICE_FUNC Index outerStride() const { return m_dstExpr.outerStride(); }
+  
+  EIGEN_DEVICE_FUNC DstEvaluatorType& dstEvaluator() { return m_dst; }
+  EIGEN_DEVICE_FUNC const SrcEvaluatorType& srcEvaluator() const { return m_src; }
+  
+  /// Assign src(row,col) to dst(row,col) through the assignment functor.
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(Index row, Index col)
+  {
+    m_functor.assignCoeff(m_dst.coeffRef(row,col), m_src.coeff(row,col));
+  }
+  
+  /// \sa assignCoeff(Index,Index)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(Index index)
+  {
+    m_functor.assignCoeff(m_dst.coeffRef(index), m_src.coeff(index));
+  }
+  
+  /// \sa assignCoeff(Index,Index)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeffByOuterInner(Index outer, Index inner)
+  {
+    Index row = rowIndexByOuterInner(outer, inner); 
+    Index col = colIndexByOuterInner(outer, inner); 
+    assignCoeff(row, col);
+  }
+  
+  
+  template<int StoreMode, int LoadMode, typename PacketType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignPacket(Index row, Index col)
+  {
+    m_functor.template assignPacket<StoreMode>(&m_dst.coeffRef(row,col), m_src.template packet<LoadMode,PacketType>(row,col));
+  }
+  
+  template<int StoreMode, int LoadMode, typename PacketType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignPacket(Index index)
+  {
+    m_functor.template assignPacket<StoreMode>(&m_dst.coeffRef(index), m_src.template packet<LoadMode,PacketType>(index));
+  }
+  
+  template<int StoreMode, int LoadMode, typename PacketType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignPacketByOuterInner(Index outer, Index inner)
+  {
+    Index row = rowIndexByOuterInner(outer, inner); 
+    Index col = colIndexByOuterInner(outer, inner);
+    assignPacket<StoreMode,LoadMode,PacketType>(row, col);
+  }
+  
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Index rowIndexByOuterInner(Index outer, Index inner)
+  {
+    typedef typename DstEvaluatorType::ExpressionTraits Traits;
+    return int(Traits::RowsAtCompileTime) == 1 ? 0
+      : int(Traits::ColsAtCompileTime) == 1 ? inner
+      : int(DstEvaluatorType::Flags)&RowMajorBit ? outer
+      : inner;
+  }
+
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Index colIndexByOuterInner(Index outer, Index inner)
+  {
+    typedef typename DstEvaluatorType::ExpressionTraits Traits;
+    return int(Traits::ColsAtCompileTime) == 1 ? 0
+      : int(Traits::RowsAtCompileTime) == 1 ? inner
+      : int(DstEvaluatorType::Flags)&RowMajorBit ? inner
+      : outer;
+  }
+
+  EIGEN_DEVICE_FUNC const Scalar* dstDataPtr() const
+  {
+    return m_dstExpr.data();
+  }
+  
+protected:
+  DstEvaluatorType& m_dst;
+  const SrcEvaluatorType& m_src;
+  const Functor &m_functor;
+  // TODO find a way to avoid the needs of the original expression
+  DstXprType& m_dstExpr;
+};
+
+/***************************************************************************
+* Part 5 : Entry point for dense rectangular assignment
+***************************************************************************/
+
+template<typename DstXprType,typename SrcXprType, typename Functor>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void resize_if_allowed(DstXprType &dst, const SrcXprType& src, const Functor &/*func*/)
+{
+  EIGEN_ONLY_USED_FOR_DEBUG(dst);
+  EIGEN_ONLY_USED_FOR_DEBUG(src);
+  eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
+}
+
+template<typename DstXprType,typename SrcXprType, typename T1, typename T2>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void resize_if_allowed(DstXprType &dst, const SrcXprType& src, const internal::assign_op<T1,T2> &/*func*/)
+{
+  Index dstRows = src.rows();
+  Index dstCols = src.cols();
+  if(((dst.rows()!=dstRows) || (dst.cols()!=dstCols)))
+    dst.resize(dstRows, dstCols);
+  eigen_assert(dst.rows() == dstRows && dst.cols() == dstCols);
+}
+
+template<typename DstXprType, typename SrcXprType, typename Functor>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void call_dense_assignment_loop(DstXprType& dst, const SrcXprType& src, const Functor &func)
+{
+  typedef evaluator<DstXprType> DstEvaluatorType;
+  typedef evaluator<SrcXprType> SrcEvaluatorType;
+
+  SrcEvaluatorType srcEvaluator(src);
+
+  // NOTE To properly handle A = (A*A.transpose())/s with A rectangular,
+  // we need to resize the destination after the source evaluator has been created.
+  resize_if_allowed(dst, src, func);
+
+  DstEvaluatorType dstEvaluator(dst);
+    
+  typedef generic_dense_assignment_kernel<DstEvaluatorType,SrcEvaluatorType,Functor> Kernel;
+  Kernel kernel(dstEvaluator, srcEvaluator, func, dst.const_cast_derived());
+
+  dense_assignment_loop<Kernel>::run(kernel);
+}
+
+template<typename DstXprType, typename SrcXprType>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void call_dense_assignment_loop(DstXprType& dst, const SrcXprType& src)
+{
+  call_dense_assignment_loop(dst, src, internal::assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar>());
+}
+
+/***************************************************************************
+* Part 6 : Generic assignment
+***************************************************************************/
+
+// Based on the respective shapes of the destination and source,
+// the class AssignmentKind determine the kind of assignment mechanism.
+// AssignmentKind must define a Kind typedef.
+template<typename DstShape, typename SrcShape> struct AssignmentKind;
+
+// Assignement kind defined in this file:
+struct Dense2Dense {};
+struct EigenBase2EigenBase {};
+
+template<typename,typename> struct AssignmentKind { typedef EigenBase2EigenBase Kind; };
+template<> struct AssignmentKind<DenseShape,DenseShape> { typedef Dense2Dense Kind; };
+    
+// This is the main assignment class
+template< typename DstXprType, typename SrcXprType, typename Functor,
+          typename Kind = typename AssignmentKind< typename evaluator_traits<DstXprType>::Shape , typename evaluator_traits<SrcXprType>::Shape >::Kind,
+          typename EnableIf = void>
+struct Assignment;
+
+
+// The only purpose of this call_assignment() function is to deal with noalias() / "assume-aliasing" and automatic transposition.
+// Indeed, I (Gael) think that this concept of "assume-aliasing" was a mistake, and it makes thing quite complicated.
+// So this intermediate function removes everything related to "assume-aliasing" such that Assignment
+// does not has to bother about these annoying details.
+
+template<typename Dst, typename Src>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void call_assignment(Dst& dst, const Src& src)
+{
+  call_assignment(dst, src, internal::assign_op<typename Dst::Scalar,typename Src::Scalar>());
+}
+template<typename Dst, typename Src>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void call_assignment(const Dst& dst, const Src& src)
+{
+  call_assignment(dst, src, internal::assign_op<typename Dst::Scalar,typename Src::Scalar>());
+}
+                     
+// Deal with "assume-aliasing"
+template<typename Dst, typename Src, typename Func>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable_if< evaluator_assume_aliasing<Src>::value, void*>::type = 0)
+{
+  typename plain_matrix_type<Src>::type tmp(src);
+  call_assignment_no_alias(dst, tmp, func);
+}
+
+template<typename Dst, typename Src, typename Func>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable_if<!evaluator_assume_aliasing<Src>::value, void*>::type = 0)
+{
+  call_assignment_no_alias(dst, src, func);
+}
+
+// by-pass "assume-aliasing"
+// When there is no aliasing, we require that 'dst' has been properly resized
+template<typename Dst, template <typename> class StorageBase, typename Src, typename Func>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void call_assignment(NoAlias<Dst,StorageBase>& dst, const Src& src, const Func& func)
+{
+  call_assignment_no_alias(dst.expression(), src, func);
+}
+
+
+template<typename Dst, typename Src, typename Func>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void call_assignment_no_alias(Dst& dst, const Src& src, const Func& func)
+{
+  enum {
+    NeedToTranspose = (    (int(Dst::RowsAtCompileTime) == 1 && int(Src::ColsAtCompileTime) == 1)
+                        || (int(Dst::ColsAtCompileTime) == 1 && int(Src::RowsAtCompileTime) == 1)
+                      ) && int(Dst::SizeAtCompileTime) != 1
+  };
+
+  typedef typename internal::conditional<NeedToTranspose, Transpose<Dst>, Dst>::type ActualDstTypeCleaned;
+  typedef typename internal::conditional<NeedToTranspose, Transpose<Dst>, Dst&>::type ActualDstType;
+  ActualDstType actualDst(dst);
+  
+  // TODO check whether this is the right place to perform these checks:
+  EIGEN_STATIC_ASSERT_LVALUE(Dst)
+  EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(ActualDstTypeCleaned,Src)
+  EIGEN_CHECK_BINARY_COMPATIBILIY(Func,typename ActualDstTypeCleaned::Scalar,typename Src::Scalar);
+  
+  Assignment<ActualDstTypeCleaned,Src,Func>::run(actualDst, src, func);
+}
+template<typename Dst, typename Src>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void call_assignment_no_alias(Dst& dst, const Src& src)
+{
+  call_assignment_no_alias(dst, src, internal::assign_op<typename Dst::Scalar,typename Src::Scalar>());
+}
+
+template<typename Dst, typename Src, typename Func>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void call_assignment_no_alias_no_transpose(Dst& dst, const Src& src, const Func& func)
+{
+  // TODO check whether this is the right place to perform these checks:
+  EIGEN_STATIC_ASSERT_LVALUE(Dst)
+  EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Dst,Src)
+  EIGEN_CHECK_BINARY_COMPATIBILIY(Func,typename Dst::Scalar,typename Src::Scalar);
+
+  Assignment<Dst,Src,Func>::run(dst, src, func);
+}
+template<typename Dst, typename Src>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void call_assignment_no_alias_no_transpose(Dst& dst, const Src& src)
+{
+  call_assignment_no_alias_no_transpose(dst, src, internal::assign_op<typename Dst::Scalar,typename Src::Scalar>());
+}
+
+// forward declaration
+template<typename Dst, typename Src> void check_for_aliasing(const Dst &dst, const Src &src);
+
+// Generic Dense to Dense assignment
+// Note that the last template argument "Weak" is needed to make it possible to perform
+// both partial specialization+SFINAE without ambiguous specialization
+template< typename DstXprType, typename SrcXprType, typename Functor, typename Weak>
+struct Assignment<DstXprType, SrcXprType, Functor, Dense2Dense, Weak>
+{
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const Functor &func)
+  {
+#ifndef EIGEN_NO_DEBUG
+    internal::check_for_aliasing(dst, src);
+#endif
+    
+    call_dense_assignment_loop(dst, src, func);
+  }
+};
+
+// Generic assignment through evalTo.
+// TODO: not sure we have to keep that one, but it helps porting current code to new evaluator mechanism.
+// Note that the last template argument "Weak" is needed to make it possible to perform
+// both partial specialization+SFINAE without ambiguous specialization
+template< typename DstXprType, typename SrcXprType, typename Functor, typename Weak>
+struct Assignment<DstXprType, SrcXprType, Functor, EigenBase2EigenBase, Weak>
+{
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &/*func*/)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+
+    eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
+    src.evalTo(dst);
+  }
+
+  // NOTE The following two functions are templated to avoid their instanciation if not needed
+  //      This is needed because some expressions supports evalTo only and/or have 'void' as scalar type.
+  template<typename SrcScalarType>
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op<typename DstXprType::Scalar,SrcScalarType> &/*func*/)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+
+    eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
+    src.addTo(dst);
+  }
+
+  template<typename SrcScalarType>
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op<typename DstXprType::Scalar,SrcScalarType> &/*func*/)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+
+    eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
+    src.subTo(dst);
+  }
+};
+
+} // namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_ASSIGN_EVALUATOR_H
diff --git a/eigen/Eigen/src/Core/Assign_MKL.h b/eigen/Eigen/src/Core/Assign_MKL.h
index 7772951..6c2ab92 100644
--- a/eigen/Eigen/src/Core/Assign_MKL.h
+++ b/eigen/Eigen/src/Core/Assign_MKL.h
@@ -1,6 +1,7 @@
 /*
  Copyright (c) 2011, Intel Corporation. All rights reserved.
-
+ Copyright (C) 2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+ 
  Redistribution and use in source and binary forms, with or without modification,
  are permitted provided that the following conditions are met:
 
@@ -37,17 +38,13 @@ namespace Eigen {
 
 namespace internal {
 
-template<typename Op> struct vml_call
-{ enum { IsSupported = 0 }; };
-
-template<typename Dst, typename Src, typename UnaryOp>
+template<typename Dst, typename Src>
 class vml_assign_traits
 {
   private:
     enum {
       DstHasDirectAccess = Dst::Flags & DirectAccessBit,
       SrcHasDirectAccess = Src::Flags & DirectAccessBit,
-
       StorageOrdersAgree = (int(Dst::IsRowMajor) == int(Src::IsRowMajor)),
       InnerSize = int(Dst::IsVectorAtCompileTime) ? int(Dst::SizeAtCompileTime)
                 : int(Dst::Flags)&RowMajorBit ? int(Dst::ColsAtCompileTime)
@@ -57,165 +54,120 @@ class vml_assign_traits
                     : int(Dst::MaxRowsAtCompileTime),
       MaxSizeAtCompileTime = Dst::SizeAtCompileTime,
 
-      MightEnableVml =  vml_call<UnaryOp>::IsSupported && StorageOrdersAgree && DstHasDirectAccess && SrcHasDirectAccess
-                     && Src::InnerStrideAtCompileTime==1 && Dst::InnerStrideAtCompileTime==1,
+      MightEnableVml = StorageOrdersAgree && DstHasDirectAccess && SrcHasDirectAccess && Src::InnerStrideAtCompileTime==1 && Dst::InnerStrideAtCompileTime==1,
       MightLinearize = MightEnableVml && (int(Dst::Flags) & int(Src::Flags) & LinearAccessBit),
       VmlSize = MightLinearize ? MaxSizeAtCompileTime : InnerMaxSize,
-      LargeEnough = VmlSize==Dynamic || VmlSize>=EIGEN_MKL_VML_THRESHOLD,
-      MayEnableVml = MightEnableVml && LargeEnough,
-      MayLinearize = MayEnableVml && MightLinearize
+      LargeEnough = VmlSize==Dynamic || VmlSize>=EIGEN_MKL_VML_THRESHOLD
     };
   public:
     enum {
-      Traversal = MayLinearize ? LinearVectorizedTraversal
-                : MayEnableVml ? InnerVectorizedTraversal
-                : DefaultTraversal
+      EnableVml = MightEnableVml && LargeEnough,
+      Traversal = MightLinearize ? LinearTraversal : DefaultTraversal
     };
 };
 
-template<typename Derived1, typename Derived2, typename UnaryOp, int Traversal, int Unrolling,
-         int VmlTraversal = vml_assign_traits<Derived1, Derived2, UnaryOp>::Traversal >
-struct vml_assign_impl
-  : assign_impl<Derived1, Eigen::CwiseUnaryOp<UnaryOp, Derived2>,Traversal,Unrolling,BuiltIn>
-{
-};
-
-template<typename Derived1, typename Derived2, typename UnaryOp, int Traversal, int Unrolling>
-struct vml_assign_impl<Derived1, Derived2, UnaryOp, Traversal, Unrolling, InnerVectorizedTraversal>
-{
-  typedef typename Derived1::Scalar Scalar;
-  typedef typename Derived1::Index Index;
-  static inline void run(Derived1& dst, const CwiseUnaryOp<UnaryOp, Derived2>& src)
-  {
-    // in case we want to (or have to) skip VML at runtime we can call:
-    // assign_impl<Derived1,Eigen::CwiseUnaryOp<UnaryOp, Derived2>,Traversal,Unrolling,BuiltIn>::run(dst,src);
-    const Index innerSize = dst.innerSize();
-    const Index outerSize = dst.outerSize();
-    for(Index outer = 0; outer < outerSize; ++outer) {
-      const Scalar *src_ptr = src.IsRowMajor ?  &(src.nestedExpression().coeffRef(outer,0)) :
-                                                &(src.nestedExpression().coeffRef(0, outer));
-      Scalar *dst_ptr = dst.IsRowMajor ? &(dst.coeffRef(outer,0)) : &(dst.coeffRef(0, outer));
-      vml_call<UnaryOp>::run(src.functor(), innerSize, src_ptr, dst_ptr );
-    }
-  }
-};
-
-template<typename Derived1, typename Derived2, typename UnaryOp, int Traversal, int Unrolling>
-struct vml_assign_impl<Derived1, Derived2, UnaryOp, Traversal, Unrolling, LinearVectorizedTraversal>
-{
-  static inline void run(Derived1& dst, const CwiseUnaryOp<UnaryOp, Derived2>& src)
-  {
-    // in case we want to (or have to) skip VML at runtime we can call:
-    // assign_impl<Derived1,Eigen::CwiseUnaryOp<UnaryOp, Derived2>,Traversal,Unrolling,BuiltIn>::run(dst,src);
-    vml_call<UnaryOp>::run(src.functor(), dst.size(), src.nestedExpression().data(), dst.data() );
-  }
-};
-
-// Macroses
-
-#define EIGEN_MKL_VML_SPECIALIZE_ASSIGN(TRAVERSAL,UNROLLING) \
-  template<typename Derived1, typename Derived2, typename UnaryOp> \
-  struct assign_impl<Derived1, Eigen::CwiseUnaryOp<UnaryOp, Derived2>, TRAVERSAL, UNROLLING, Specialized>  {  \
-    static inline void run(Derived1 &dst, const Eigen::CwiseUnaryOp<UnaryOp, Derived2> &src) { \
-      vml_assign_impl<Derived1,Derived2,UnaryOp,TRAVERSAL,UNROLLING>::run(dst, src); \
-    } \
-  };
-
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(DefaultTraversal,NoUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(DefaultTraversal,CompleteUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(DefaultTraversal,InnerUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(LinearTraversal,NoUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(LinearTraversal,CompleteUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(InnerVectorizedTraversal,NoUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(InnerVectorizedTraversal,CompleteUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(InnerVectorizedTraversal,InnerUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(LinearVectorizedTraversal,CompleteUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(LinearVectorizedTraversal,NoUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(SliceVectorizedTraversal,NoUnrolling)
-
-
+#define EIGEN_PP_EXPAND(ARG) ARG
 #if !defined (EIGEN_FAST_MATH) || (EIGEN_FAST_MATH != 1)
-#define  EIGEN_MKL_VML_MODE VML_HA
+#define EIGEN_VMLMODE_EXPAND_LA , VML_HA
 #else
-#define  EIGEN_MKL_VML_MODE VML_LA
+#define EIGEN_VMLMODE_EXPAND_LA , VML_LA
 #endif
 
-#define EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE)     \
-  template<> struct vml_call< scalar_##EIGENOP##_op<EIGENTYPE> > {               \
-    enum { IsSupported = 1 };                                                    \
-    static inline void run( const scalar_##EIGENOP##_op<EIGENTYPE>& /*func*/,        \
-                            int size, const EIGENTYPE* src, EIGENTYPE* dst) {    \
-      VMLOP(size, (const VMLTYPE*)src, (VMLTYPE*)dst);                           \
-    }                                                                            \
+#define EIGEN_VMLMODE_EXPAND__ 
+
+#define EIGEN_VMLMODE_PREFIX_LA vm
+#define EIGEN_VMLMODE_PREFIX__  v
+#define EIGEN_VMLMODE_PREFIX(VMLMODE) EIGEN_CAT(EIGEN_VMLMODE_PREFIX_,VMLMODE)
+
+#define EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE, VMLMODE)                                           \
+  template< typename DstXprType, typename SrcXprNested>                                                                         \
+  struct Assignment<DstXprType, CwiseUnaryOp<scalar_##EIGENOP##_op<EIGENTYPE>, SrcXprNested>, assign_op<EIGENTYPE,EIGENTYPE>,   \
+                   Dense2Dense, typename enable_if<vml_assign_traits<DstXprType,SrcXprNested>::EnableVml>::type> {              \
+    typedef CwiseUnaryOp<scalar_##EIGENOP##_op<EIGENTYPE>, SrcXprNested> SrcXprType;                                            \
+    static void run(DstXprType &dst, const SrcXprType &src, const assign_op<EIGENTYPE,EIGENTYPE> &/*func*/) {                   \
+      eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());                                                       \
+      if(vml_assign_traits<DstXprType,SrcXprNested>::Traversal==LinearTraversal) {                                              \
+        VMLOP(dst.size(), (const VMLTYPE*)src.nestedExpression().data(),                                                        \
+              (VMLTYPE*)dst.data() EIGEN_PP_EXPAND(EIGEN_VMLMODE_EXPAND_##VMLMODE) );                                           \
+      } else {                                                                                                                  \
+        const Index outerSize = dst.outerSize();                                                                                \
+        for(Index outer = 0; outer < outerSize; ++outer) {                                                                      \
+          const EIGENTYPE *src_ptr = src.IsRowMajor ? &(src.nestedExpression().coeffRef(outer,0)) :                             \
+                                                      &(src.nestedExpression().coeffRef(0, outer));                             \
+          EIGENTYPE *dst_ptr = dst.IsRowMajor ? &(dst.coeffRef(outer,0)) : &(dst.coeffRef(0, outer));                           \
+          VMLOP( dst.innerSize(), (const VMLTYPE*)src_ptr,                                                                      \
+                (VMLTYPE*)dst_ptr EIGEN_PP_EXPAND(EIGEN_VMLMODE_EXPAND_##VMLMODE));                                             \
+        }                                                                                                                       \
+      }                                                                                                                         \
+    }                                                                                                                           \
+  };                                                                                                                            \
+
+
+#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(EIGENOP, VMLOP, VMLMODE)                                                         \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, EIGEN_CAT(EIGEN_VMLMODE_PREFIX(VMLMODE),s##VMLOP), float, float, VMLMODE)           \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, EIGEN_CAT(EIGEN_VMLMODE_PREFIX(VMLMODE),d##VMLOP), double, double, VMLMODE)
+
+#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_CPLX(EIGENOP, VMLOP, VMLMODE)                                                         \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, EIGEN_CAT(EIGEN_VMLMODE_PREFIX(VMLMODE),c##VMLOP), scomplex, MKL_Complex8, VMLMODE) \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, EIGEN_CAT(EIGEN_VMLMODE_PREFIX(VMLMODE),z##VMLOP), dcomplex, MKL_Complex16, VMLMODE)
+  
+#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS(EIGENOP, VMLOP, VMLMODE)                                                              \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(EIGENOP, VMLOP, VMLMODE)                                                               \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALLS_CPLX(EIGENOP, VMLOP, VMLMODE)
+
+  
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(sin,   Sin,   LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(asin,  Asin,  LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(sinh,  Sinh,  LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(cos,   Cos,   LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(acos,  Acos,  LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(cosh,  Cosh,  LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(tan,   Tan,   LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(atan,  Atan,  LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(tanh,  Tanh,  LA)
+// EIGEN_MKL_VML_DECLARE_UNARY_CALLS(abs,   Abs,    _)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(exp,   Exp,   LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(log,   Ln,    LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(log10, Log10, LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(sqrt,  Sqrt,  _)
+
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(square, Sqr,   _)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_CPLX(arg, Arg,      _)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(round, Round,  _)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(floor, Floor,  _)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(ceil,  Ceil,   _)
+
+#define EIGEN_MKL_VML_DECLARE_POW_CALL(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE, VMLMODE)                                           \
+  template< typename DstXprType, typename SrcXprNested, typename Plain>                                                       \
+  struct Assignment<DstXprType, CwiseBinaryOp<scalar_##EIGENOP##_op<EIGENTYPE,EIGENTYPE>, SrcXprNested,                       \
+                    const CwiseNullaryOp<internal::scalar_constant_op<EIGENTYPE>,Plain> >, assign_op<EIGENTYPE,EIGENTYPE>,    \
+                   Dense2Dense, typename enable_if<vml_assign_traits<DstXprType,SrcXprNested>::EnableVml>::type> {            \
+    typedef CwiseBinaryOp<scalar_##EIGENOP##_op<EIGENTYPE,EIGENTYPE>, SrcXprNested,                                           \
+                    const CwiseNullaryOp<internal::scalar_constant_op<EIGENTYPE>,Plain> > SrcXprType;                         \
+    static void run(DstXprType &dst, const SrcXprType &src, const assign_op<EIGENTYPE,EIGENTYPE> &/*func*/) {                 \
+      eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());                                                     \
+      VMLTYPE exponent = reinterpret_cast<const VMLTYPE&>(src.rhs().functor().m_other);                                       \
+      if(vml_assign_traits<DstXprType,SrcXprNested>::Traversal==LinearTraversal)                                              \
+      {                                                                                                                       \
+        VMLOP( dst.size(), (const VMLTYPE*)src.lhs().data(), exponent,                                                        \
+              (VMLTYPE*)dst.data() EIGEN_PP_EXPAND(EIGEN_VMLMODE_EXPAND_##VMLMODE) );                                         \
+      } else {                                                                                                                \
+        const Index outerSize = dst.outerSize();                                                                              \
+        for(Index outer = 0; outer < outerSize; ++outer) {                                                                    \
+          const EIGENTYPE *src_ptr = src.IsRowMajor ? &(src.lhs().coeffRef(outer,0)) :                                        \
+                                                      &(src.lhs().coeffRef(0, outer));                                        \
+          EIGENTYPE *dst_ptr = dst.IsRowMajor ? &(dst.coeffRef(outer,0)) : &(dst.coeffRef(0, outer));                         \
+          VMLOP( dst.innerSize(), (const VMLTYPE*)src_ptr, exponent,                                                          \
+                 (VMLTYPE*)dst_ptr EIGEN_PP_EXPAND(EIGEN_VMLMODE_EXPAND_##VMLMODE));                                          \
+        }                                                                                                                     \
+      }                                                                                                                       \
+    }                                                                                                                         \
   };
-
-#define EIGEN_MKL_VML_DECLARE_UNARY_CALL_LA(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE)  \
-  template<> struct vml_call< scalar_##EIGENOP##_op<EIGENTYPE> > {               \
-    enum { IsSupported = 1 };                                                    \
-    static inline void run( const scalar_##EIGENOP##_op<EIGENTYPE>& /*func*/,        \
-                            int size, const EIGENTYPE* src, EIGENTYPE* dst) {    \
-      MKL_INT64 vmlMode = EIGEN_MKL_VML_MODE;                                    \
-      VMLOP(size, (const VMLTYPE*)src, (VMLTYPE*)dst, vmlMode);                  \
-    }                                                                            \
-  };
-
-#define EIGEN_MKL_VML_DECLARE_POW_CALL(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE)       \
-  template<> struct vml_call< scalar_##EIGENOP##_op<EIGENTYPE> > {               \
-    enum { IsSupported = 1 };                                                    \
-    static inline void run( const scalar_##EIGENOP##_op<EIGENTYPE>& func,        \
-                          int size, const EIGENTYPE* src, EIGENTYPE* dst) {      \
-      EIGENTYPE exponent = func.m_exponent;                                      \
-      MKL_INT64 vmlMode = EIGEN_MKL_VML_MODE;                                    \
-      VMLOP(&size, (const VMLTYPE*)src, (const VMLTYPE*)&exponent,               \
-                        (VMLTYPE*)dst, &vmlMode);                                \
-    }                                                                            \
-  };
-
-#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(EIGENOP, VMLOP)                   \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, vs##VMLOP, float, float)             \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, vd##VMLOP, double, double)
-
-#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_COMPLEX(EIGENOP, VMLOP)                \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, vc##VMLOP, scomplex, MKL_Complex8)   \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, vz##VMLOP, dcomplex, MKL_Complex16)
-
-#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS(EIGENOP, VMLOP)                        \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(EIGENOP, VMLOP)                         \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALLS_COMPLEX(EIGENOP, VMLOP)
-
-
-#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL_LA(EIGENOP, VMLOP)                \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALL_LA(EIGENOP, vms##VMLOP, float, float)         \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALL_LA(EIGENOP, vmd##VMLOP, double, double)
-
-#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_COMPLEX_LA(EIGENOP, VMLOP)             \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALL_LA(EIGENOP, vmc##VMLOP, scomplex, MKL_Complex8)  \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALL_LA(EIGENOP, vmz##VMLOP, dcomplex, MKL_Complex16)
-
-#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(EIGENOP, VMLOP)                     \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL_LA(EIGENOP, VMLOP)                      \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALLS_COMPLEX_LA(EIGENOP, VMLOP)
-
-
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(sin,  Sin)
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(asin, Asin)
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(cos,  Cos)
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(acos, Acos)
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(tan,  Tan)
-//EIGEN_MKL_VML_DECLARE_UNARY_CALLS(abs,  Abs)
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(exp,  Exp)
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(log,  Ln)
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(sqrt, Sqrt)
-
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(square, Sqr)
-
-// The vm*powx functions are not avaibale in the windows version of MKL.
-#ifndef _WIN32
-EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmspowx_, float, float)
-EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmdpowx_, double, double)
-EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmcpowx_, scomplex, MKL_Complex8)
-EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmzpowx_, dcomplex, MKL_Complex16)
-#endif
+  
+EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmsPowx, float,    float,         LA)
+EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmdPowx, double,   double,        LA)
+EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmcPowx, scomplex, MKL_Complex8,  LA)
+EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmzPowx, dcomplex, MKL_Complex16, LA)
 
 } // end namespace internal
 
diff --git a/eigen/Eigen/src/Core/BandMatrix.h b/eigen/Eigen/src/Core/BandMatrix.h
index ffd7fe8..4978c91 100644
--- a/eigen/Eigen/src/Core/BandMatrix.h
+++ b/eigen/Eigen/src/Core/BandMatrix.h
@@ -32,7 +32,7 @@ class BandMatrixBase : public EigenBase<Derived>
     };
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef Matrix<Scalar,RowsAtCompileTime,ColsAtCompileTime> DenseMatrixType;
-    typedef typename DenseMatrixType::Index Index;
+    typedef typename DenseMatrixType::StorageIndex StorageIndex;
     typedef typename internal::traits<Derived>::CoefficientsType CoefficientsType;
     typedef EigenBase<Derived> Base;
 
@@ -161,15 +161,15 @@ class BandMatrixBase : public EigenBase<Derived>
   *
   * \brief Represents a rectangular matrix with a banded storage
   *
-  * \param _Scalar Numeric type, i.e. float, double, int
-  * \param Rows Number of rows, or \b Dynamic
-  * \param Cols Number of columns, or \b Dynamic
-  * \param Supers Number of super diagonal
-  * \param Subs Number of sub diagonal
-  * \param _Options A combination of either \b #RowMajor or \b #ColMajor, and of \b #SelfAdjoint
-  *                 The former controls \ref TopicStorageOrders "storage order", and defaults to
-  *                 column-major. The latter controls whether the matrix represents a selfadjoint 
-  *                 matrix in which case either Supers of Subs have to be null.
+  * \tparam _Scalar Numeric type, i.e. float, double, int
+  * \tparam _Rows Number of rows, or \b Dynamic
+  * \tparam _Cols Number of columns, or \b Dynamic
+  * \tparam _Supers Number of super diagonal
+  * \tparam _Subs Number of sub diagonal
+  * \tparam _Options A combination of either \b #RowMajor or \b #ColMajor, and of \b #SelfAdjoint
+  *                  The former controls \ref TopicStorageOrders "storage order", and defaults to
+  *                  column-major. The latter controls whether the matrix represents a selfadjoint
+  *                  matrix in which case either Supers of Subs have to be null.
   *
   * \sa class TridiagonalMatrix
   */
@@ -179,7 +179,7 @@ struct traits<BandMatrix<_Scalar,_Rows,_Cols,_Supers,_Subs,_Options> >
 {
   typedef _Scalar Scalar;
   typedef Dense StorageKind;
-  typedef DenseIndex Index;
+  typedef Eigen::Index StorageIndex;
   enum {
     CoeffReadCost = NumTraits<Scalar>::ReadCost,
     RowsAtCompileTime = _Rows,
@@ -201,10 +201,10 @@ class BandMatrix : public BandMatrixBase<BandMatrix<_Scalar,Rows,Cols,Supers,Sub
   public:
 
     typedef typename internal::traits<BandMatrix>::Scalar Scalar;
-    typedef typename internal::traits<BandMatrix>::Index Index;
+    typedef typename internal::traits<BandMatrix>::StorageIndex StorageIndex;
     typedef typename internal::traits<BandMatrix>::CoefficientsType CoefficientsType;
 
-    inline BandMatrix(Index rows=Rows, Index cols=Cols, Index supers=Supers, Index subs=Subs)
+    explicit inline BandMatrix(Index rows=Rows, Index cols=Cols, Index supers=Supers, Index subs=Subs)
       : m_coeffs(1+supers+subs,cols),
         m_rows(rows), m_supers(supers), m_subs(subs)
     {
@@ -241,7 +241,7 @@ struct traits<BandMatrixWrapper<_CoefficientsType,_Rows,_Cols,_Supers,_Subs,_Opt
 {
   typedef typename _CoefficientsType::Scalar Scalar;
   typedef typename _CoefficientsType::StorageKind StorageKind;
-  typedef typename _CoefficientsType::Index Index;
+  typedef typename _CoefficientsType::StorageIndex StorageIndex;
   enum {
     CoeffReadCost = internal::traits<_CoefficientsType>::CoeffReadCost,
     RowsAtCompileTime = _Rows,
@@ -264,9 +264,9 @@ class BandMatrixWrapper : public BandMatrixBase<BandMatrixWrapper<_CoefficientsT
 
     typedef typename internal::traits<BandMatrixWrapper>::Scalar Scalar;
     typedef typename internal::traits<BandMatrixWrapper>::CoefficientsType CoefficientsType;
-    typedef typename internal::traits<BandMatrixWrapper>::Index Index;
+    typedef typename internal::traits<BandMatrixWrapper>::StorageIndex StorageIndex;
 
-    inline BandMatrixWrapper(const CoefficientsType& coeffs, Index rows=_Rows, Index cols=_Cols, Index supers=_Supers, Index subs=_Subs)
+    explicit inline BandMatrixWrapper(const CoefficientsType& coeffs, Index rows=_Rows, Index cols=_Cols, Index supers=_Supers, Index subs=_Subs)
       : m_coeffs(coeffs),
         m_rows(rows), m_supers(supers), m_subs(subs)
     {
@@ -302,9 +302,9 @@ class BandMatrixWrapper : public BandMatrixBase<BandMatrixWrapper<_CoefficientsT
   *
   * \brief Represents a tridiagonal matrix with a compact banded storage
   *
-  * \param _Scalar Numeric type, i.e. float, double, int
-  * \param Size Number of rows and cols, or \b Dynamic
-  * \param _Options Can be 0 or \b SelfAdjoint
+  * \tparam Scalar Numeric type, i.e. float, double, int
+  * \tparam Size Number of rows and cols, or \b Dynamic
+  * \tparam Options Can be 0 or \b SelfAdjoint
   *
   * \sa class BandMatrix
   */
@@ -312,9 +312,9 @@ template<typename Scalar, int Size, int Options>
 class TridiagonalMatrix : public BandMatrix<Scalar,Size,Size,Options&SelfAdjoint?0:1,1,Options|RowMajor>
 {
     typedef BandMatrix<Scalar,Size,Size,Options&SelfAdjoint?0:1,1,Options|RowMajor> Base;
-    typedef typename Base::Index Index;
+    typedef typename Base::StorageIndex StorageIndex;
   public:
-    TridiagonalMatrix(Index size = Size) : Base(size,size,Options&SelfAdjoint?0:1,1) {}
+    explicit TridiagonalMatrix(Index size = Size) : Base(size,size,Options&SelfAdjoint?0:1,1) {}
 
     inline typename Base::template DiagonalIntReturnType<1>::Type super()
     { return Base::template diagonal<1>(); }
@@ -327,6 +327,25 @@ class TridiagonalMatrix : public BandMatrix<Scalar,Size,Size,Options&SelfAdjoint
   protected:
 };
 
+
+struct BandShape {};
+
+template<typename _Scalar, int _Rows, int _Cols, int _Supers, int _Subs, int _Options>
+struct evaluator_traits<BandMatrix<_Scalar,_Rows,_Cols,_Supers,_Subs,_Options> >
+  : public evaluator_traits_base<BandMatrix<_Scalar,_Rows,_Cols,_Supers,_Subs,_Options> >
+{
+  typedef BandShape Shape;
+};
+
+template<typename _CoefficientsType,int _Rows, int _Cols, int _Supers, int _Subs,int _Options>
+struct evaluator_traits<BandMatrixWrapper<_CoefficientsType,_Rows,_Cols,_Supers,_Subs,_Options> >
+  : public evaluator_traits_base<BandMatrixWrapper<_CoefficientsType,_Rows,_Cols,_Supers,_Subs,_Options> >
+{
+  typedef BandShape Shape;
+};
+
+template<> struct AssignmentKind<DenseShape,BandShape> { typedef EigenBase2EigenBase Kind; };
+
 } // end namespace internal
 
 } // end namespace Eigen
diff --git a/eigen/Eigen/src/Core/Block.h b/eigen/Eigen/src/Core/Block.h
index 8278944..11de45c 100644
--- a/eigen/Eigen/src/Core/Block.h
+++ b/eigen/Eigen/src/Core/Block.h
@@ -13,38 +13,6 @@
 
 namespace Eigen { 
 
-/** \class Block
-  * \ingroup Core_Module
-  *
-  * \brief Expression of a fixed-size or dynamic-size block
-  *
-  * \param XprType the type of the expression in which we are taking a block
-  * \param BlockRows the number of rows of the block we are taking at compile time (optional)
-  * \param BlockCols the number of columns of the block we are taking at compile time (optional)
-  *
-  * This class represents an expression of either a fixed-size or dynamic-size block. It is the return
-  * type of DenseBase::block(Index,Index,Index,Index) and DenseBase::block<int,int>(Index,Index) and
-  * most of the time this is the only way it is used.
-  *
-  * However, if you want to directly maniputate block expressions,
-  * for instance if you want to write a function returning such an expression, you
-  * will need to use this class.
-  *
-  * Here is an example illustrating the dynamic case:
-  * \include class_Block.cpp
-  * Output: \verbinclude class_Block.out
-  *
-  * \note Even though this expression has dynamic size, in the case where \a XprType
-  * has fixed size, this expression inherits a fixed maximal size which means that evaluating
-  * it does not cause a dynamic memory allocation.
-  *
-  * Here is an example illustrating the fixed-size case:
-  * \include class_FixedBlock.cpp
-  * Output: \verbinclude class_FixedBlock.out
-  *
-  * \sa DenseBase::block(Index,Index,Index,Index), DenseBase::block(Index,Index), class VectorBlock
-  */
-
 namespace internal {
 template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel>
 struct traits<Block<XprType, BlockRows, BlockCols, InnerPanel> > : traits<XprType>
@@ -52,7 +20,7 @@ struct traits<Block<XprType, BlockRows, BlockCols, InnerPanel> > : traits<XprTyp
   typedef typename traits<XprType>::Scalar Scalar;
   typedef typename traits<XprType>::StorageKind StorageKind;
   typedef typename traits<XprType>::XprKind XprKind;
-  typedef typename nested<XprType>::type XprTypeNested;
+  typedef typename ref_selector<XprType>::type XprTypeNested;
   typedef typename remove_reference<XprTypeNested>::type _XprTypeNested;
   enum{
     MatrixRows = traits<XprType>::RowsAtCompileTime,
@@ -65,10 +33,10 @@ struct traits<Block<XprType, BlockRows, BlockCols, InnerPanel> > : traits<XprTyp
     MaxColsAtCompileTime = BlockCols==0 ? 0
                          : ColsAtCompileTime != Dynamic ? int(ColsAtCompileTime)
                          : int(traits<XprType>::MaxColsAtCompileTime),
+
     XprTypeIsRowMajor = (int(traits<XprType>::Flags)&RowMajorBit) != 0,
-    IsDense = is_same<StorageKind,Dense>::value,
-    IsRowMajor = (IsDense&&MaxRowsAtCompileTime==1&&MaxColsAtCompileTime!=1) ? 1
-               : (IsDense&&MaxColsAtCompileTime==1&&MaxRowsAtCompileTime!=1) ? 0
+    IsRowMajor = (MaxRowsAtCompileTime==1&&MaxColsAtCompileTime!=1) ? 1
+               : (MaxColsAtCompileTime==1&&MaxRowsAtCompileTime!=1) ? 0
                : XprTypeIsRowMajor,
     HasSameStorageOrderAsXprType = (IsRowMajor == XprTypeIsRowMajor),
     InnerSize = IsRowMajor ? int(ColsAtCompileTime) : int(RowsAtCompileTime),
@@ -78,18 +46,16 @@ struct traits<Block<XprType, BlockRows, BlockCols, InnerPanel> > : traits<XprTyp
     OuterStrideAtCompileTime = HasSameStorageOrderAsXprType
                              ? int(outer_stride_at_compile_time<XprType>::ret)
                              : int(inner_stride_at_compile_time<XprType>::ret),
-    MaskPacketAccessBit = (InnerSize == Dynamic || (InnerSize % packet_traits<Scalar>::size) == 0)
-                       && (InnerStrideAtCompileTime == 1)
-                        ? PacketAccessBit : 0,
-    MaskAlignedBit = (InnerPanel && (OuterStrideAtCompileTime!=Dynamic) && (((OuterStrideAtCompileTime * int(sizeof(Scalar))) % 16) == 0)) ? AlignedBit : 0,
-    FlagsLinearAccessBit = (RowsAtCompileTime == 1 || ColsAtCompileTime == 1 || (InnerPanel && (traits<XprType>::Flags&LinearAccessBit))) ? LinearAccessBit : 0,
+
+    // FIXME, this traits is rather specialized for dense object and it needs to be cleaned further
     FlagsLvalueBit = is_lvalue<XprType>::value ? LvalueBit : 0,
     FlagsRowMajorBit = IsRowMajor ? RowMajorBit : 0,
-    Flags0 = traits<XprType>::Flags & ( (HereditaryBits & ~RowMajorBit) |
-                                        DirectAccessBit |
-                                        MaskPacketAccessBit |
-                                        MaskAlignedBit),
-    Flags = Flags0 | FlagsLinearAccessBit | FlagsLvalueBit | FlagsRowMajorBit
+    Flags = (traits<XprType>::Flags & (DirectAccessBit | (InnerPanel?CompressedAccessBit:0))) | FlagsLvalueBit | FlagsRowMajorBit,
+    // FIXME DirectAccessBit should not be handled by expressions
+    // 
+    // Alignment is needed by MapBase's assertions
+    // We can sefely set it to false here. Internal alignment errors will be detected by an eigen_internal_assert in the respective evaluator
+    Alignment = 0
   };
 };
 
@@ -100,6 +66,40 @@ template<typename XprType, int BlockRows=Dynamic, int BlockCols=Dynamic, bool In
 
 template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel, typename StorageKind> class BlockImpl;
 
+/** \class Block
+  * \ingroup Core_Module
+  *
+  * \brief Expression of a fixed-size or dynamic-size block
+  *
+  * \tparam XprType the type of the expression in which we are taking a block
+  * \tparam BlockRows the number of rows of the block we are taking at compile time (optional)
+  * \tparam BlockCols the number of columns of the block we are taking at compile time (optional)
+  * \tparam InnerPanel is true, if the block maps to a set of rows of a row major matrix or
+  *         to set of columns of a column major matrix (optional). The parameter allows to determine
+  *         at compile time whether aligned access is possible on the block expression.
+  *
+  * This class represents an expression of either a fixed-size or dynamic-size block. It is the return
+  * type of DenseBase::block(Index,Index,Index,Index) and DenseBase::block<int,int>(Index,Index) and
+  * most of the time this is the only way it is used.
+  *
+  * However, if you want to directly maniputate block expressions,
+  * for instance if you want to write a function returning such an expression, you
+  * will need to use this class.
+  *
+  * Here is an example illustrating the dynamic case:
+  * \include class_Block.cpp
+  * Output: \verbinclude class_Block.out
+  *
+  * \note Even though this expression has dynamic size, in the case where \a XprType
+  * has fixed size, this expression inherits a fixed maximal size which means that evaluating
+  * it does not cause a dynamic memory allocation.
+  *
+  * Here is an example illustrating the fixed-size case:
+  * \include class_FixedBlock.cpp
+  * Output: \verbinclude class_FixedBlock.out
+  *
+  * \sa DenseBase::block(Index,Index,Index,Index), DenseBase::block(Index,Index), class VectorBlock
+  */
 template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel> class Block
   : public BlockImpl<XprType, BlockRows, BlockCols, InnerPanel, typename internal::traits<XprType>::StorageKind>
 {
@@ -109,9 +109,12 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel> class
     typedef Impl Base;
     EIGEN_GENERIC_PUBLIC_INTERFACE(Block)
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Block)
+    
+    typedef typename internal::remove_all<XprType>::type NestedExpression;
   
     /** Column or Row constructor
       */
+    EIGEN_DEVICE_FUNC
     inline Block(XprType& xpr, Index i) : Impl(xpr,i)
     {
       eigen_assert( (i>=0) && (
@@ -121,25 +124,27 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel> class
 
     /** Fixed-size constructor
       */
-    inline Block(XprType& xpr, Index a_startRow, Index a_startCol)
-      : Impl(xpr, a_startRow, a_startCol)
+    EIGEN_DEVICE_FUNC
+    inline Block(XprType& xpr, Index startRow, Index startCol)
+      : Impl(xpr, startRow, startCol)
     {
       EIGEN_STATIC_ASSERT(RowsAtCompileTime!=Dynamic && ColsAtCompileTime!=Dynamic,THIS_METHOD_IS_ONLY_FOR_FIXED_SIZE)
-      eigen_assert(a_startRow >= 0 && BlockRows >= 1 && a_startRow + BlockRows <= xpr.rows()
-             && a_startCol >= 0 && BlockCols >= 1 && a_startCol + BlockCols <= xpr.cols());
+      eigen_assert(startRow >= 0 && BlockRows >= 0 && startRow + BlockRows <= xpr.rows()
+             && startCol >= 0 && BlockCols >= 0 && startCol + BlockCols <= xpr.cols());
     }
 
     /** Dynamic-size constructor
       */
+    EIGEN_DEVICE_FUNC
     inline Block(XprType& xpr,
-          Index a_startRow, Index a_startCol,
+          Index startRow, Index startCol,
           Index blockRows, Index blockCols)
-      : Impl(xpr, a_startRow, a_startCol, blockRows, blockCols)
+      : Impl(xpr, startRow, startCol, blockRows, blockCols)
     {
       eigen_assert((RowsAtCompileTime==Dynamic || RowsAtCompileTime==blockRows)
           && (ColsAtCompileTime==Dynamic || ColsAtCompileTime==blockCols));
-      eigen_assert(a_startRow >= 0 && blockRows >= 0 && a_startRow  <= xpr.rows() - blockRows
-          && a_startCol >= 0 && blockCols >= 0 && a_startCol <= xpr.cols() - blockCols);
+      eigen_assert(startRow >= 0 && blockRows >= 0 && startRow  <= xpr.rows() - blockRows
+          && startCol >= 0 && blockCols >= 0 && startCol <= xpr.cols() - blockCols);
     }
 };
          
@@ -150,14 +155,15 @@ class BlockImpl<XprType, BlockRows, BlockCols, InnerPanel, Dense>
   : public internal::BlockImpl_dense<XprType, BlockRows, BlockCols, InnerPanel>
 {
     typedef internal::BlockImpl_dense<XprType, BlockRows, BlockCols, InnerPanel> Impl;
-    typedef typename XprType::Index Index;
+    typedef typename XprType::StorageIndex StorageIndex;
   public:
     typedef Impl Base;
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(BlockImpl)
-    inline BlockImpl(XprType& xpr, Index i) : Impl(xpr,i) {}
-    inline BlockImpl(XprType& xpr, Index a_startRow, Index a_startCol) : Impl(xpr, a_startRow, a_startCol) {}
-    inline BlockImpl(XprType& xpr, Index a_startRow, Index a_startCol, Index blockRows, Index blockCols)
-      : Impl(xpr, a_startRow, a_startCol, blockRows, blockCols) {}
+    EIGEN_DEVICE_FUNC inline BlockImpl(XprType& xpr, Index i) : Impl(xpr,i) {}
+    EIGEN_DEVICE_FUNC inline BlockImpl(XprType& xpr, Index startRow, Index startCol) : Impl(xpr, startRow, startCol) {}
+    EIGEN_DEVICE_FUNC
+    inline BlockImpl(XprType& xpr, Index startRow, Index startCol, Index blockRows, Index blockCols)
+      : Impl(xpr, startRow, startCol, blockRows, blockCols) {}
 };
 
 namespace internal {
@@ -167,16 +173,18 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel, bool H
   : public internal::dense_xpr_base<Block<XprType, BlockRows, BlockCols, InnerPanel> >::type
 {
     typedef Block<XprType, BlockRows, BlockCols, InnerPanel> BlockType;
+    typedef typename internal::ref_selector<XprType>::non_const_type XprTypeNested;
   public:
 
     typedef typename internal::dense_xpr_base<BlockType>::type Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(BlockType)
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(BlockImpl_dense)
 
-    class InnerIterator;
+    // class InnerIterator; // FIXME apparently never used
 
     /** Column or Row constructor
       */
+    EIGEN_DEVICE_FUNC
     inline BlockImpl_dense(XprType& xpr, Index i)
       : m_xpr(xpr),
         // It is a row if and only if BlockRows==1 and BlockCols==XprType::ColsAtCompileTime,
@@ -191,75 +199,76 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel, bool H
 
     /** Fixed-size constructor
       */
-    inline BlockImpl_dense(XprType& xpr, Index a_startRow, Index a_startCol)
-      : m_xpr(xpr), m_startRow(a_startRow), m_startCol(a_startCol),
+    EIGEN_DEVICE_FUNC
+    inline BlockImpl_dense(XprType& xpr, Index startRow, Index startCol)
+      : m_xpr(xpr), m_startRow(startRow), m_startCol(startCol),
                     m_blockRows(BlockRows), m_blockCols(BlockCols)
     {}
 
     /** Dynamic-size constructor
       */
+    EIGEN_DEVICE_FUNC
     inline BlockImpl_dense(XprType& xpr,
-          Index a_startRow, Index a_startCol,
+          Index startRow, Index startCol,
           Index blockRows, Index blockCols)
-      : m_xpr(xpr), m_startRow(a_startRow), m_startCol(a_startCol),
+      : m_xpr(xpr), m_startRow(startRow), m_startCol(startCol),
                     m_blockRows(blockRows), m_blockCols(blockCols)
     {}
 
-    inline Index rows() const { return m_blockRows.value(); }
-    inline Index cols() const { return m_blockCols.value(); }
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_blockRows.value(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_blockCols.value(); }
 
+    EIGEN_DEVICE_FUNC
     inline Scalar& coeffRef(Index rowId, Index colId)
     {
       EIGEN_STATIC_ASSERT_LVALUE(XprType)
-      return m_xpr.const_cast_derived()
-               .coeffRef(rowId + m_startRow.value(), colId + m_startCol.value());
+      return m_xpr.coeffRef(rowId + m_startRow.value(), colId + m_startCol.value());
     }
 
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index rowId, Index colId) const
     {
-      return m_xpr.derived()
-               .coeffRef(rowId + m_startRow.value(), colId + m_startCol.value());
+      return m_xpr.derived().coeffRef(rowId + m_startRow.value(), colId + m_startCol.value());
     }
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE const CoeffReturnType coeff(Index rowId, Index colId) const
     {
       return m_xpr.coeff(rowId + m_startRow.value(), colId + m_startCol.value());
     }
 
+    EIGEN_DEVICE_FUNC
     inline Scalar& coeffRef(Index index)
     {
       EIGEN_STATIC_ASSERT_LVALUE(XprType)
-      return m_xpr.const_cast_derived()
-             .coeffRef(m_startRow.value() + (RowsAtCompileTime == 1 ? 0 : index),
-                       m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0));
+      return m_xpr.coeffRef(m_startRow.value() + (RowsAtCompileTime == 1 ? 0 : index),
+                            m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0));
     }
 
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index index) const
     {
-      return m_xpr.const_cast_derived()
-             .coeffRef(m_startRow.value() + (RowsAtCompileTime == 1 ? 0 : index),
-                       m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0));
+      return m_xpr.coeffRef(m_startRow.value() + (RowsAtCompileTime == 1 ? 0 : index),
+                            m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0));
     }
 
+    EIGEN_DEVICE_FUNC
     inline const CoeffReturnType coeff(Index index) const
     {
-      return m_xpr
-             .coeff(m_startRow.value() + (RowsAtCompileTime == 1 ? 0 : index),
-                    m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0));
+      return m_xpr.coeff(m_startRow.value() + (RowsAtCompileTime == 1 ? 0 : index),
+                         m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0));
     }
 
     template<int LoadMode>
     inline PacketScalar packet(Index rowId, Index colId) const
     {
-      return m_xpr.template packet<Unaligned>
-              (rowId + m_startRow.value(), colId + m_startCol.value());
+      return m_xpr.template packet<Unaligned>(rowId + m_startRow.value(), colId + m_startCol.value());
     }
 
     template<int LoadMode>
     inline void writePacket(Index rowId, Index colId, const PacketScalar& val)
     {
-      m_xpr.const_cast_derived().template writePacket<Unaligned>
-              (rowId + m_startRow.value(), colId + m_startCol.value(), val);
+      m_xpr.template writePacket<Unaligned>(rowId + m_startRow.value(), colId + m_startCol.value(), val);
     }
 
     template<int LoadMode>
@@ -273,40 +282,46 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel, bool H
     template<int LoadMode>
     inline void writePacket(Index index, const PacketScalar& val)
     {
-      m_xpr.const_cast_derived().template writePacket<Unaligned>
+      m_xpr.template writePacket<Unaligned>
          (m_startRow.value() + (RowsAtCompileTime == 1 ? 0 : index),
           m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0), val);
     }
 
     #ifdef EIGEN_PARSED_BY_DOXYGEN
     /** \sa MapBase::data() */
-    inline const Scalar* data() const;
-    inline Index innerStride() const;
-    inline Index outerStride() const;
+    EIGEN_DEVICE_FUNC inline const Scalar* data() const;
+    EIGEN_DEVICE_FUNC inline Index innerStride() const;
+    EIGEN_DEVICE_FUNC inline Index outerStride() const;
     #endif
 
-    const typename internal::remove_all<typename XprType::Nested>::type& nestedExpression() const 
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<XprTypeNested>::type& nestedExpression() const
     { 
       return m_xpr; 
     }
+
+    EIGEN_DEVICE_FUNC
+    XprType& nestedExpression() { return m_xpr; }
       
-    Index startRow() const 
+    EIGEN_DEVICE_FUNC
+    StorageIndex startRow() const
     { 
       return m_startRow.value(); 
     }
       
-    Index startCol() const 
+    EIGEN_DEVICE_FUNC
+    StorageIndex startCol() const
     { 
       return m_startCol.value(); 
     }
 
   protected:
 
-    const typename XprType::Nested m_xpr;
-    const internal::variable_if_dynamic<Index, XprType::RowsAtCompileTime == 1 ? 0 : Dynamic> m_startRow;
-    const internal::variable_if_dynamic<Index, XprType::ColsAtCompileTime == 1 ? 0 : Dynamic> m_startCol;
-    const internal::variable_if_dynamic<Index, RowsAtCompileTime> m_blockRows;
-    const internal::variable_if_dynamic<Index, ColsAtCompileTime> m_blockCols;
+    XprTypeNested m_xpr;
+    const internal::variable_if_dynamic<StorageIndex, (XprType::RowsAtCompileTime == 1 && BlockRows==1) ? 0 : Dynamic> m_startRow;
+    const internal::variable_if_dynamic<StorageIndex, (XprType::ColsAtCompileTime == 1 && BlockCols==1) ? 0 : Dynamic> m_startCol;
+    const internal::variable_if_dynamic<StorageIndex, RowsAtCompileTime> m_blockRows;
+    const internal::variable_if_dynamic<StorageIndex, ColsAtCompileTime> m_blockCols;
 };
 
 /** \internal Internal implementation of dense Blocks in the direct access case.*/
@@ -315,6 +330,10 @@ class BlockImpl_dense<XprType,BlockRows,BlockCols, InnerPanel,true>
   : public MapBase<Block<XprType, BlockRows, BlockCols, InnerPanel> >
 {
     typedef Block<XprType, BlockRows, BlockCols, InnerPanel> BlockType;
+    typedef typename internal::ref_selector<XprType>::non_const_type XprTypeNested;
+    enum {
+      XprTypeIsRowMajor = (int(traits<XprType>::Flags)&RowMajorBit) != 0
+    };
   public:
 
     typedef MapBase<BlockType> Base;
@@ -323,42 +342,52 @@ class BlockImpl_dense<XprType,BlockRows,BlockCols, InnerPanel,true>
 
     /** Column or Row constructor
       */
+    EIGEN_DEVICE_FUNC
     inline BlockImpl_dense(XprType& xpr, Index i)
-      : Base(internal::const_cast_ptr(&xpr.coeffRef(
-              (BlockRows==1) && (BlockCols==XprType::ColsAtCompileTime) ? i : 0,
-              (BlockRows==XprType::RowsAtCompileTime) && (BlockCols==1) ? i : 0)),
+      : Base(xpr.data() + i * (    ((BlockRows==1) && (BlockCols==XprType::ColsAtCompileTime) && (!XprTypeIsRowMajor)) 
+                                || ((BlockRows==XprType::RowsAtCompileTime) && (BlockCols==1) && ( XprTypeIsRowMajor)) ? xpr.innerStride() : xpr.outerStride()),
              BlockRows==1 ? 1 : xpr.rows(),
              BlockCols==1 ? 1 : xpr.cols()),
-        m_xpr(xpr)
+        m_xpr(xpr),
+        m_startRow( (BlockRows==1) && (BlockCols==XprType::ColsAtCompileTime) ? i : 0),
+        m_startCol( (BlockRows==XprType::RowsAtCompileTime) && (BlockCols==1) ? i : 0)
     {
       init();
     }
 
     /** Fixed-size constructor
       */
+    EIGEN_DEVICE_FUNC
     inline BlockImpl_dense(XprType& xpr, Index startRow, Index startCol)
-      : Base(internal::const_cast_ptr(&xpr.coeffRef(startRow,startCol))), m_xpr(xpr)
+      : Base(xpr.data()+xpr.innerStride()*(XprTypeIsRowMajor?startCol:startRow) + xpr.outerStride()*(XprTypeIsRowMajor?startRow:startCol)),
+        m_xpr(xpr), m_startRow(startRow), m_startCol(startCol)
     {
       init();
     }
 
     /** Dynamic-size constructor
       */
+    EIGEN_DEVICE_FUNC
     inline BlockImpl_dense(XprType& xpr,
           Index startRow, Index startCol,
           Index blockRows, Index blockCols)
-      : Base(internal::const_cast_ptr(&xpr.coeffRef(startRow,startCol)), blockRows, blockCols),
-        m_xpr(xpr)
+      : Base(xpr.data()+xpr.innerStride()*(XprTypeIsRowMajor?startCol:startRow) + xpr.outerStride()*(XprTypeIsRowMajor?startRow:startCol), blockRows, blockCols),
+        m_xpr(xpr), m_startRow(startRow), m_startCol(startCol)
     {
       init();
     }
 
-    const typename internal::remove_all<typename XprType::Nested>::type& nestedExpression() const 
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<XprTypeNested>::type& nestedExpression() const
     { 
       return m_xpr; 
     }
+
+    EIGEN_DEVICE_FUNC
+    XprType& nestedExpression() { return m_xpr; }
       
     /** \sa MapBase::innerStride() */
+    EIGEN_DEVICE_FUNC
     inline Index innerStride() const
     {
       return internal::traits<BlockType>::HasSameStorageOrderAsXprType
@@ -367,11 +396,24 @@ class BlockImpl_dense<XprType,BlockRows,BlockCols, InnerPanel,true>
     }
 
     /** \sa MapBase::outerStride() */
+    EIGEN_DEVICE_FUNC
     inline Index outerStride() const
     {
       return m_outerStride;
     }
 
+    EIGEN_DEVICE_FUNC
+    StorageIndex startRow() const
+    {
+      return m_startRow.value();
+    }
+
+    EIGEN_DEVICE_FUNC
+    StorageIndex startCol() const
+    {
+      return m_startCol.value();
+    }
+
   #ifndef __SUNPRO_CC
   // FIXME sunstudio is not friendly with the above friend...
   // META-FIXME there is no 'friend' keyword around here. Is this obsolete?
@@ -380,6 +422,7 @@ class BlockImpl_dense<XprType,BlockRows,BlockCols, InnerPanel,true>
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     /** \internal used by allowAligned() */
+    EIGEN_DEVICE_FUNC
     inline BlockImpl_dense(XprType& xpr, const Scalar* data, Index blockRows, Index blockCols)
       : Base(data, blockRows, blockCols), m_xpr(xpr)
     {
@@ -388,6 +431,7 @@ class BlockImpl_dense<XprType,BlockRows,BlockCols, InnerPanel,true>
     #endif
 
   protected:
+    EIGEN_DEVICE_FUNC
     void init()
     {
       m_outerStride = internal::traits<BlockType>::HasSameStorageOrderAsXprType
@@ -395,7 +439,9 @@ class BlockImpl_dense<XprType,BlockRows,BlockCols, InnerPanel,true>
                     : m_xpr.innerStride();
     }
 
-    typename XprType::Nested m_xpr;
+    XprTypeNested m_xpr;
+    const internal::variable_if_dynamic<StorageIndex, (XprType::RowsAtCompileTime == 1 && BlockRows==1) ? 0 : Dynamic> m_startRow;
+    const internal::variable_if_dynamic<StorageIndex, (XprType::ColsAtCompileTime == 1 && BlockCols==1) ? 0 : Dynamic> m_startCol;
     Index m_outerStride;
 };
 
diff --git a/eigen/Eigen/src/Core/BooleanRedux.h b/eigen/Eigen/src/Core/BooleanRedux.h
index be9f48a..ccf5190 100644
--- a/eigen/Eigen/src/Core/BooleanRedux.h
+++ b/eigen/Eigen/src/Core/BooleanRedux.h
@@ -14,54 +14,54 @@ namespace Eigen {
 
 namespace internal {
 
-template<typename Derived, int UnrollCount>
+template<typename Derived, int UnrollCount, int Rows>
 struct all_unroller
 {
   enum {
-    col = (UnrollCount-1) / Derived::RowsAtCompileTime,
-    row = (UnrollCount-1) % Derived::RowsAtCompileTime
+    col = (UnrollCount-1) / Rows,
+    row = (UnrollCount-1) % Rows
   };
 
   static inline bool run(const Derived &mat)
   {
-    return all_unroller<Derived, UnrollCount-1>::run(mat) && mat.coeff(row, col);
+    return all_unroller<Derived, UnrollCount-1, Rows>::run(mat) && mat.coeff(row, col);
   }
 };
 
-template<typename Derived>
-struct all_unroller<Derived, 0>
+template<typename Derived, int Rows>
+struct all_unroller<Derived, 0, Rows>
 {
   static inline bool run(const Derived &/*mat*/) { return true; }
 };
 
-template<typename Derived>
-struct all_unroller<Derived, Dynamic>
+template<typename Derived, int Rows>
+struct all_unroller<Derived, Dynamic, Rows>
 {
   static inline bool run(const Derived &) { return false; }
 };
 
-template<typename Derived, int UnrollCount>
+template<typename Derived, int UnrollCount, int Rows>
 struct any_unroller
 {
   enum {
-    col = (UnrollCount-1) / Derived::RowsAtCompileTime,
-    row = (UnrollCount-1) % Derived::RowsAtCompileTime
+    col = (UnrollCount-1) / Rows,
+    row = (UnrollCount-1) % Rows
   };
-
+  
   static inline bool run(const Derived &mat)
   {
-    return any_unroller<Derived, UnrollCount-1>::run(mat) || mat.coeff(row, col);
+    return any_unroller<Derived, UnrollCount-1, Rows>::run(mat) || mat.coeff(row, col);
   }
 };
 
-template<typename Derived>
-struct any_unroller<Derived, 0>
+template<typename Derived, int Rows>
+struct any_unroller<Derived, 0, Rows>
 {
   static inline bool run(const Derived & /*mat*/) { return false; }
 };
 
-template<typename Derived>
-struct any_unroller<Derived, Dynamic>
+template<typename Derived, int Rows>
+struct any_unroller<Derived, Dynamic, Rows>
 {
   static inline bool run(const Derived &) { return false; }
 };
@@ -76,21 +76,21 @@ struct any_unroller<Derived, Dynamic>
   * \sa any(), Cwise::operator<()
   */
 template<typename Derived>
-inline bool DenseBase<Derived>::all() const
+EIGEN_DEVICE_FUNC inline bool DenseBase<Derived>::all() const
 {
+  typedef internal::evaluator<Derived> Evaluator;
   enum {
     unroll = SizeAtCompileTime != Dynamic
-          && CoeffReadCost != Dynamic
-          && NumTraits<Scalar>::AddCost != Dynamic
-          && SizeAtCompileTime * (CoeffReadCost + NumTraits<Scalar>::AddCost) <= EIGEN_UNROLLING_LIMIT
+          && SizeAtCompileTime * (Evaluator::CoeffReadCost + NumTraits<Scalar>::AddCost) <= EIGEN_UNROLLING_LIMIT
   };
+  Evaluator evaluator(derived());
   if(unroll)
-    return internal::all_unroller<Derived, unroll ? int(SizeAtCompileTime) : Dynamic>::run(derived());
+    return internal::all_unroller<Evaluator, unroll ? int(SizeAtCompileTime) : Dynamic, internal::traits<Derived>::RowsAtCompileTime>::run(evaluator);
   else
   {
     for(Index j = 0; j < cols(); ++j)
       for(Index i = 0; i < rows(); ++i)
-        if (!coeff(i, j)) return false;
+        if (!evaluator.coeff(i, j)) return false;
     return true;
   }
 }
@@ -100,21 +100,21 @@ inline bool DenseBase<Derived>::all() const
   * \sa all()
   */
 template<typename Derived>
-inline bool DenseBase<Derived>::any() const
+EIGEN_DEVICE_FUNC inline bool DenseBase<Derived>::any() const
 {
+  typedef internal::evaluator<Derived> Evaluator;
   enum {
     unroll = SizeAtCompileTime != Dynamic
-          && CoeffReadCost != Dynamic
-          && NumTraits<Scalar>::AddCost != Dynamic
-          && SizeAtCompileTime * (CoeffReadCost + NumTraits<Scalar>::AddCost) <= EIGEN_UNROLLING_LIMIT
+          && SizeAtCompileTime * (Evaluator::CoeffReadCost + NumTraits<Scalar>::AddCost) <= EIGEN_UNROLLING_LIMIT
   };
+  Evaluator evaluator(derived());
   if(unroll)
-    return internal::any_unroller<Derived, unroll ? int(SizeAtCompileTime) : Dynamic>::run(derived());
+    return internal::any_unroller<Evaluator, unroll ? int(SizeAtCompileTime) : Dynamic, internal::traits<Derived>::RowsAtCompileTime>::run(evaluator);
   else
   {
     for(Index j = 0; j < cols(); ++j)
       for(Index i = 0; i < rows(); ++i)
-        if (coeff(i, j)) return true;
+        if (evaluator.coeff(i, j)) return true;
     return false;
   }
 }
@@ -124,7 +124,7 @@ inline bool DenseBase<Derived>::any() const
   * \sa all(), any()
   */
 template<typename Derived>
-inline typename DenseBase<Derived>::Index DenseBase<Derived>::count() const
+EIGEN_DEVICE_FUNC inline Eigen::Index DenseBase<Derived>::count() const
 {
   return derived().template cast<bool>().template cast<Index>().sum();
 }
@@ -136,7 +136,11 @@ inline typename DenseBase<Derived>::Index DenseBase<Derived>::count() const
 template<typename Derived>
 inline bool DenseBase<Derived>::hasNaN() const
 {
+#if EIGEN_COMP_MSVC || (defined __FAST_MATH__)
+  return derived().array().isNaN().any();
+#else
   return !((derived().array()==derived().array()).all());
+#endif
 }
 
 /** \returns true if \c *this contains only finite numbers, i.e., no NaN and no +/-INF values.
@@ -146,7 +150,11 @@ inline bool DenseBase<Derived>::hasNaN() const
 template<typename Derived>
 inline bool DenseBase<Derived>::allFinite() const
 {
+#if EIGEN_COMP_MSVC || (defined __FAST_MATH__)
+  return derived().array().isFinite().all();
+#else
   return !((derived()-derived()).hasNaN());
+#endif
 }
     
 } // end namespace Eigen
diff --git a/eigen/Eigen/src/Core/CMakeLists.txt b/eigen/Eigen/src/Core/CMakeLists.txt
deleted file mode 100644
index 2346fc2..0000000
--- a/eigen/Eigen/src/Core/CMakeLists.txt
+++ /dev/null
@@ -1,10 +0,0 @@
-FILE(GLOB Eigen_Core_SRCS "*.h")
-
-INSTALL(FILES
-  ${Eigen_Core_SRCS}
-  DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core COMPONENT Devel
-  )
-
-ADD_SUBDIRECTORY(products)
-ADD_SUBDIRECTORY(util)
-ADD_SUBDIRECTORY(arch)
diff --git a/eigen/Eigen/src/Core/CommaInitializer.h b/eigen/Eigen/src/Core/CommaInitializer.h
index 56ee38f..35fdbb8 100644
--- a/eigen/Eigen/src/Core/CommaInitializer.h
+++ b/eigen/Eigen/src/Core/CommaInitializer.h
@@ -22,14 +22,14 @@ namespace Eigen {
   * the return type of MatrixBase::operator<<, and most of the time this is the only
   * way it is used.
   *
-  * \sa \ref MatrixBaseCommaInitRef "MatrixBase::operator<<", CommaInitializer::finished()
+  * \sa \blank \ref MatrixBaseCommaInitRef "MatrixBase::operator<<", CommaInitializer::finished()
   */
 template<typename XprType>
 struct CommaInitializer
 {
   typedef typename XprType::Scalar Scalar;
-  typedef typename XprType::Index Index;
 
+  EIGEN_DEVICE_FUNC
   inline CommaInitializer(XprType& xpr, const Scalar& s)
     : m_xpr(xpr), m_row(0), m_col(1), m_currentBlockRows(1)
   {
@@ -37,6 +37,7 @@ struct CommaInitializer
   }
 
   template<typename OtherDerived>
+  EIGEN_DEVICE_FUNC
   inline CommaInitializer(XprType& xpr, const DenseBase<OtherDerived>& other)
     : m_xpr(xpr), m_row(0), m_col(other.cols()), m_currentBlockRows(other.rows())
   {
@@ -46,6 +47,7 @@ struct CommaInitializer
   /* Copy/Move constructor which transfers ownership. This is crucial in 
    * absence of return value optimization to avoid assertions during destruction. */
   // FIXME in C++11 mode this could be replaced by a proper RValue constructor
+  EIGEN_DEVICE_FUNC
   inline CommaInitializer(const CommaInitializer& o)
   : m_xpr(o.m_xpr), m_row(o.m_row), m_col(o.m_col), m_currentBlockRows(o.m_currentBlockRows) {
     // Mark original object as finished. In absence of R-value references we need to const_cast:
@@ -55,6 +57,7 @@ struct CommaInitializer
   }
 
   /* inserts a scalar value in the target matrix */
+  EIGEN_DEVICE_FUNC
   CommaInitializer& operator,(const Scalar& s)
   {
     if (m_col==m_xpr.cols())
@@ -74,14 +77,10 @@ struct CommaInitializer
 
   /* inserts a matrix expression in the target matrix */
   template<typename OtherDerived>
+  EIGEN_DEVICE_FUNC
   CommaInitializer& operator,(const DenseBase<OtherDerived>& other)
   {
-    if(other.rows()==0)
-    {
-      m_col += other.cols();
-      return *this;
-    }
-    if (m_col==m_xpr.cols())
+    if (m_col==m_xpr.cols() && (other.cols()!=0 || other.rows()!=m_currentBlockRows))
     {
       m_row+=m_currentBlockRows;
       m_col = 0;
@@ -89,24 +88,22 @@ struct CommaInitializer
       eigen_assert(m_row+m_currentBlockRows<=m_xpr.rows()
         && "Too many rows passed to comma initializer (operator<<)");
     }
-    eigen_assert((m_col<m_xpr.cols() || (m_xpr.cols()==0 && m_col==0))
+    eigen_assert((m_col + other.cols() <= m_xpr.cols())
       && "Too many coefficients passed to comma initializer (operator<<)");
     eigen_assert(m_currentBlockRows==other.rows());
-    if (OtherDerived::SizeAtCompileTime != Dynamic)
-      m_xpr.template block<OtherDerived::RowsAtCompileTime != Dynamic ? OtherDerived::RowsAtCompileTime : 1,
-                              OtherDerived::ColsAtCompileTime != Dynamic ? OtherDerived::ColsAtCompileTime : 1>
-                    (m_row, m_col) = other;
-    else
-      m_xpr.block(m_row, m_col, other.rows(), other.cols()) = other;
+    m_xpr.template block<OtherDerived::RowsAtCompileTime, OtherDerived::ColsAtCompileTime>
+                    (m_row, m_col, other.rows(), other.cols()) = other;
     m_col += other.cols();
     return *this;
   }
 
+  EIGEN_DEVICE_FUNC
   inline ~CommaInitializer()
+#if defined VERIFY_RAISES_ASSERT && (!defined EIGEN_NO_ASSERTION_CHECKING) && defined EIGEN_EXCEPTIONS
+  EIGEN_EXCEPTION_SPEC(Eigen::eigen_assert_exception)
+#endif
   {
-    eigen_assert((m_row+m_currentBlockRows) == m_xpr.rows()
-         && m_col == m_xpr.cols()
-         && "Too few coefficients passed to comma initializer (operator<<)");
+      finished();
   }
 
   /** \returns the built matrix once all its coefficients have been set.
@@ -116,9 +113,15 @@ struct CommaInitializer
     * quaternion.fromRotationMatrix((Matrix3f() << axis0, axis1, axis2).finished());
     * \endcode
     */
-  inline XprType& finished() { return m_xpr; }
+  EIGEN_DEVICE_FUNC
+  inline XprType& finished() {
+      eigen_assert(((m_row+m_currentBlockRows) == m_xpr.rows() || m_xpr.cols() == 0)
+           && m_col == m_xpr.cols()
+           && "Too few coefficients passed to comma initializer (operator<<)");
+      return m_xpr;
+  }
 
-  XprType& m_xpr;   // target expression
+  XprType& m_xpr;           // target expression
   Index m_row;              // current row id
   Index m_col;              // current col id
   Index m_currentBlockRows; // current block height
@@ -138,7 +141,7 @@ struct CommaInitializer
   * \sa CommaInitializer::finished(), class CommaInitializer
   */
 template<typename Derived>
-inline CommaInitializer<Derived> DenseBase<Derived>::operator<< (const Scalar& s)
+EIGEN_DEVICE_FUNC inline CommaInitializer<Derived> DenseBase<Derived>::operator<< (const Scalar& s)
 {
   return CommaInitializer<Derived>(*static_cast<Derived*>(this), s);
 }
@@ -146,7 +149,7 @@ inline CommaInitializer<Derived> DenseBase<Derived>::operator<< (const Scalar& s
 /** \sa operator<<(const Scalar&) */
 template<typename Derived>
 template<typename OtherDerived>
-inline CommaInitializer<Derived>
+EIGEN_DEVICE_FUNC inline CommaInitializer<Derived>
 DenseBase<Derived>::operator<<(const DenseBase<OtherDerived>& other)
 {
   return CommaInitializer<Derived>(*static_cast<Derived *>(this), other);
diff --git a/eigen/Eigen/src/Core/ConditionEstimator.h b/eigen/Eigen/src/Core/ConditionEstimator.h
new file mode 100644
index 0000000..aa7efdc
--- /dev/null
+++ b/eigen/Eigen/src/Core/ConditionEstimator.h
@@ -0,0 +1,175 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Rasmus Munk Larsen (rmlarsen@google.com)
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CONDITIONESTIMATOR_H
+#define EIGEN_CONDITIONESTIMATOR_H
+
+namespace Eigen {
+
+namespace internal {
+
+template <typename Vector, typename RealVector, bool IsComplex>
+struct rcond_compute_sign {
+  static inline Vector run(const Vector& v) {
+    const RealVector v_abs = v.cwiseAbs();
+    return (v_abs.array() == static_cast<typename Vector::RealScalar>(0))
+            .select(Vector::Ones(v.size()), v.cwiseQuotient(v_abs));
+  }
+};
+
+// Partial specialization to avoid elementwise division for real vectors.
+template <typename Vector>
+struct rcond_compute_sign<Vector, Vector, false> {
+  static inline Vector run(const Vector& v) {
+    return (v.array() < static_cast<typename Vector::RealScalar>(0))
+           .select(-Vector::Ones(v.size()), Vector::Ones(v.size()));
+  }
+};
+
+/**
+  * \returns an estimate of ||inv(matrix)||_1 given a decomposition of
+  * \a matrix that implements .solve() and .adjoint().solve() methods.
+  *
+  * This function implements Algorithms 4.1 and 5.1 from
+  *   http://www.maths.manchester.ac.uk/~higham/narep/narep135.pdf
+  * which also forms the basis for the condition number estimators in
+  * LAPACK. Since at most 10 calls to the solve method of dec are
+  * performed, the total cost is O(dims^2), as opposed to O(dims^3)
+  * needed to compute the inverse matrix explicitly.
+  *
+  * The most common usage is in estimating the condition number
+  * ||matrix||_1 * ||inv(matrix)||_1. The first term ||matrix||_1 can be
+  * computed directly in O(n^2) operations.
+  *
+  * Supports the following decompositions: FullPivLU, PartialPivLU, LDLT, and
+  * LLT.
+  *
+  * \sa FullPivLU, PartialPivLU, LDLT, LLT.
+  */
+template <typename Decomposition>
+typename Decomposition::RealScalar rcond_invmatrix_L1_norm_estimate(const Decomposition& dec)
+{
+  typedef typename Decomposition::MatrixType MatrixType;
+  typedef typename Decomposition::Scalar Scalar;
+  typedef typename Decomposition::RealScalar RealScalar;
+  typedef typename internal::plain_col_type<MatrixType>::type Vector;
+  typedef typename internal::plain_col_type<MatrixType, RealScalar>::type RealVector;
+  const bool is_complex = (NumTraits<Scalar>::IsComplex != 0);
+
+  eigen_assert(dec.rows() == dec.cols());
+  const Index n = dec.rows();
+  if (n == 0)
+    return 0;
+
+  // Disable Index to float conversion warning
+#ifdef __INTEL_COMPILER
+  #pragma warning push
+  #pragma warning ( disable : 2259 )
+#endif
+  Vector v = dec.solve(Vector::Ones(n) / Scalar(n));
+#ifdef __INTEL_COMPILER
+  #pragma warning pop
+#endif
+
+  // lower_bound is a lower bound on
+  //   ||inv(matrix)||_1  = sup_v ||inv(matrix) v||_1 / ||v||_1
+  // and is the objective maximized by the ("super-") gradient ascent
+  // algorithm below.
+  RealScalar lower_bound = v.template lpNorm<1>();
+  if (n == 1)
+    return lower_bound;
+
+  // Gradient ascent algorithm follows: We know that the optimum is achieved at
+  // one of the simplices v = e_i, so in each iteration we follow a
+  // super-gradient to move towards the optimal one.
+  RealScalar old_lower_bound = lower_bound;
+  Vector sign_vector(n);
+  Vector old_sign_vector;
+  Index v_max_abs_index = -1;
+  Index old_v_max_abs_index = v_max_abs_index;
+  for (int k = 0; k < 4; ++k)
+  {
+    sign_vector = internal::rcond_compute_sign<Vector, RealVector, is_complex>::run(v);
+    if (k > 0 && !is_complex && sign_vector == old_sign_vector) {
+      // Break if the solution stagnated.
+      break;
+    }
+    // v_max_abs_index = argmax |real( inv(matrix)^T * sign_vector )|
+    v = dec.adjoint().solve(sign_vector);
+    v.real().cwiseAbs().maxCoeff(&v_max_abs_index);
+    if (v_max_abs_index == old_v_max_abs_index) {
+      // Break if the solution stagnated.
+      break;
+    }
+    // Move to the new simplex e_j, where j = v_max_abs_index.
+    v = dec.solve(Vector::Unit(n, v_max_abs_index));  // v = inv(matrix) * e_j.
+    lower_bound = v.template lpNorm<1>();
+    if (lower_bound <= old_lower_bound) {
+      // Break if the gradient step did not increase the lower_bound.
+      break;
+    }
+    if (!is_complex) {
+      old_sign_vector = sign_vector;
+    }
+    old_v_max_abs_index = v_max_abs_index;
+    old_lower_bound = lower_bound;
+  }
+  // The following calculates an independent estimate of ||matrix||_1 by
+  // multiplying matrix by a vector with entries of slowly increasing
+  // magnitude and alternating sign:
+  //   v_i = (-1)^{i} (1 + (i / (dim-1))), i = 0,...,dim-1.
+  // This improvement to Hager's algorithm above is due to Higham. It was
+  // added to make the algorithm more robust in certain corner cases where
+  // large elements in the matrix might otherwise escape detection due to
+  // exact cancellation (especially when op and op_adjoint correspond to a
+  // sequence of backsubstitutions and permutations), which could cause
+  // Hager's algorithm to vastly underestimate ||matrix||_1.
+  Scalar alternating_sign(RealScalar(1));
+  for (Index i = 0; i < n; ++i) {
+    // The static_cast is needed when Scalar is a complex and RealScalar implements expression templates
+    v[i] = alternating_sign * static_cast<RealScalar>(RealScalar(1) + (RealScalar(i) / (RealScalar(n - 1))));
+    alternating_sign = -alternating_sign;
+  }
+  v = dec.solve(v);
+  const RealScalar alternate_lower_bound = (2 * v.template lpNorm<1>()) / (3 * RealScalar(n));
+  return numext::maxi(lower_bound, alternate_lower_bound);
+}
+
+/** \brief Reciprocal condition number estimator.
+  *
+  * Computing a decomposition of a dense matrix takes O(n^3) operations, while
+  * this method estimates the condition number quickly and reliably in O(n^2)
+  * operations.
+  *
+  * \returns an estimate of the reciprocal condition number
+  * (1 / (||matrix||_1 * ||inv(matrix)||_1)) of matrix, given ||matrix||_1 and
+  * its decomposition. Supports the following decompositions: FullPivLU,
+  * PartialPivLU, LDLT, and LLT.
+  *
+  * \sa FullPivLU, PartialPivLU, LDLT, LLT.
+  */
+template <typename Decomposition>
+typename Decomposition::RealScalar
+rcond_estimate_helper(typename Decomposition::RealScalar matrix_norm, const Decomposition& dec)
+{
+  typedef typename Decomposition::RealScalar RealScalar;
+  eigen_assert(dec.rows() == dec.cols());
+  if (dec.rows() == 0)              return RealScalar(1);
+  if (matrix_norm == RealScalar(0)) return RealScalar(0);
+  if (dec.rows() == 1)              return RealScalar(1);
+  const RealScalar inverse_matrix_norm = rcond_invmatrix_L1_norm_estimate(dec);
+  return (inverse_matrix_norm == RealScalar(0) ? RealScalar(0)
+                                               : (RealScalar(1) / inverse_matrix_norm) / matrix_norm);
+}
+
+}  // namespace internal
+
+}  // namespace Eigen
+
+#endif
diff --git a/eigen/Eigen/src/Core/CoreEvaluators.h b/eigen/Eigen/src/Core/CoreEvaluators.h
new file mode 100644
index 0000000..15b361b
--- /dev/null
+++ b/eigen/Eigen/src/Core/CoreEvaluators.h
@@ -0,0 +1,1728 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2011 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2011-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2011-2012 Jitse Niesen <jitse@maths.leeds.ac.uk>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+
+#ifndef EIGEN_COREEVALUATORS_H
+#define EIGEN_COREEVALUATORS_H
+
+namespace Eigen {
+  
+namespace internal {
+
+// This class returns the evaluator kind from the expression storage kind.
+// Default assumes index based accessors
+template<typename StorageKind>
+struct storage_kind_to_evaluator_kind {
+  typedef IndexBased Kind;
+};
+
+// This class returns the evaluator shape from the expression storage kind.
+// It can be Dense, Sparse, Triangular, Diagonal, SelfAdjoint, Band, etc.
+template<typename StorageKind> struct storage_kind_to_shape;
+
+template<> struct storage_kind_to_shape<Dense>                  { typedef DenseShape Shape;           };
+template<> struct storage_kind_to_shape<SolverStorage>          { typedef SolverShape Shape;           };
+template<> struct storage_kind_to_shape<PermutationStorage>     { typedef PermutationShape Shape;     };
+template<> struct storage_kind_to_shape<TranspositionsStorage>  { typedef TranspositionsShape Shape;  };
+
+// Evaluators have to be specialized with respect to various criteria such as:
+//  - storage/structure/shape
+//  - scalar type
+//  - etc.
+// Therefore, we need specialization of evaluator providing additional template arguments for each kind of evaluators.
+// We currently distinguish the following kind of evaluators:
+// - unary_evaluator    for expressions taking only one arguments (CwiseUnaryOp, CwiseUnaryView, Transpose, MatrixWrapper, ArrayWrapper, Reverse, Replicate)
+// - binary_evaluator   for expression taking two arguments (CwiseBinaryOp)
+// - ternary_evaluator   for expression taking three arguments (CwiseTernaryOp)
+// - product_evaluator  for linear algebra products (Product); special case of binary_evaluator because it requires additional tags for dispatching.
+// - mapbase_evaluator  for Map, Block, Ref
+// - block_evaluator    for Block (special dispatching to a mapbase_evaluator or unary_evaluator)
+
+template< typename T,
+          typename Arg1Kind   = typename evaluator_traits<typename T::Arg1>::Kind,
+          typename Arg2Kind   = typename evaluator_traits<typename T::Arg2>::Kind,
+          typename Arg3Kind   = typename evaluator_traits<typename T::Arg3>::Kind,
+          typename Arg1Scalar = typename traits<typename T::Arg1>::Scalar,
+          typename Arg2Scalar = typename traits<typename T::Arg2>::Scalar,
+          typename Arg3Scalar = typename traits<typename T::Arg3>::Scalar> struct ternary_evaluator;
+
+template< typename T,
+          typename LhsKind   = typename evaluator_traits<typename T::Lhs>::Kind,
+          typename RhsKind   = typename evaluator_traits<typename T::Rhs>::Kind,
+          typename LhsScalar = typename traits<typename T::Lhs>::Scalar,
+          typename RhsScalar = typename traits<typename T::Rhs>::Scalar> struct binary_evaluator;
+
+template< typename T,
+          typename Kind   = typename evaluator_traits<typename T::NestedExpression>::Kind,
+          typename Scalar = typename T::Scalar> struct unary_evaluator;
+          
+// evaluator_traits<T> contains traits for evaluator<T> 
+
+template<typename T>
+struct evaluator_traits_base
+{
+  // by default, get evaluator kind and shape from storage
+  typedef typename storage_kind_to_evaluator_kind<typename traits<T>::StorageKind>::Kind Kind;
+  typedef typename storage_kind_to_shape<typename traits<T>::StorageKind>::Shape Shape;
+};
+
+// Default evaluator traits
+template<typename T>
+struct evaluator_traits : public evaluator_traits_base<T>
+{
+};
+
+template<typename T, typename Shape = typename evaluator_traits<T>::Shape >
+struct evaluator_assume_aliasing {
+  static const bool value = false;
+};
+
+// By default, we assume a unary expression:
+template<typename T>
+struct evaluator : public unary_evaluator<T>
+{
+  typedef unary_evaluator<T> Base;
+  EIGEN_DEVICE_FUNC explicit evaluator(const T& xpr) : Base(xpr) {}
+};
+
+
+// TODO: Think about const-correctness
+template<typename T>
+struct evaluator<const T>
+  : evaluator<T>
+{
+  EIGEN_DEVICE_FUNC
+  explicit evaluator(const T& xpr) : evaluator<T>(xpr) {}
+};
+
+// ---------- base class for all evaluators ----------
+
+template<typename ExpressionType>
+struct evaluator_base
+{
+  // TODO that's not very nice to have to propagate all these traits. They are currently only needed to handle outer,inner indices.
+  typedef traits<ExpressionType> ExpressionTraits;
+  
+  enum {
+    Alignment = 0
+  };
+  // noncopyable:
+  // Don't make this class inherit noncopyable as this kills EBO (Empty Base Optimization)
+  // and make complex evaluator much larger than then should do.
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE evaluator_base() {}
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ~evaluator_base() {}
+private:
+  EIGEN_DEVICE_FUNC evaluator_base(const evaluator_base&);
+  EIGEN_DEVICE_FUNC const evaluator_base& operator=(const evaluator_base&);
+};
+
+// -------------------- Matrix and Array --------------------
+//
+// evaluator<PlainObjectBase> is a common base class for the
+// Matrix and Array evaluators.
+// Here we directly specialize evaluator. This is not really a unary expression, and it is, by definition, dense,
+// so no need for more sophisticated dispatching.
+
+// this helper permits to completely eliminate m_outerStride if it is known at compiletime.
+template<typename Scalar,int OuterStride> class plainobjectbase_evaluator_data {
+public:
+  EIGEN_DEVICE_FUNC plainobjectbase_evaluator_data(const Scalar* ptr, Index outerStride) : data(ptr)
+  {
+    EIGEN_ONLY_USED_FOR_DEBUG(outerStride);
+    eigen_internal_assert(outerStride==OuterStride);
+  }
+  EIGEN_DEVICE_FUNC Index outerStride() const { return OuterStride; }
+  const Scalar *data;
+};
+
+template<typename Scalar> class plainobjectbase_evaluator_data<Scalar,Dynamic> {
+public:
+  EIGEN_DEVICE_FUNC plainobjectbase_evaluator_data(const Scalar* ptr, Index outerStride) : data(ptr), m_outerStride(outerStride) {}
+  EIGEN_DEVICE_FUNC Index outerStride() const { return m_outerStride; }
+  const Scalar *data;
+protected:
+  Index m_outerStride;
+};
+
+template<typename Derived>
+struct evaluator<PlainObjectBase<Derived> >
+  : evaluator_base<Derived>
+{
+  typedef PlainObjectBase<Derived> PlainObjectType;
+  typedef typename PlainObjectType::Scalar Scalar;
+  typedef typename PlainObjectType::CoeffReturnType CoeffReturnType;
+
+  enum {
+    IsRowMajor = PlainObjectType::IsRowMajor,
+    IsVectorAtCompileTime = PlainObjectType::IsVectorAtCompileTime,
+    RowsAtCompileTime = PlainObjectType::RowsAtCompileTime,
+    ColsAtCompileTime = PlainObjectType::ColsAtCompileTime,
+    
+    CoeffReadCost = NumTraits<Scalar>::ReadCost,
+    Flags = traits<Derived>::EvaluatorFlags,
+    Alignment = traits<Derived>::Alignment
+  };
+  enum {
+    // We do not need to know the outer stride for vectors
+    OuterStrideAtCompileTime = IsVectorAtCompileTime  ? 0
+                                                      : int(IsRowMajor) ? ColsAtCompileTime
+                                                                        : RowsAtCompileTime
+  };
+
+  EIGEN_DEVICE_FUNC evaluator()
+    : m_d(0,OuterStrideAtCompileTime)
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+
+  EIGEN_DEVICE_FUNC explicit evaluator(const PlainObjectType& m)
+    : m_d(m.data(),IsVectorAtCompileTime ? 0 : m.outerStride())
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index col) const
+  {
+    if (IsRowMajor)
+      return m_d.data[row * m_d.outerStride() + col];
+    else
+      return m_d.data[row + col * m_d.outerStride()];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  {
+    return m_d.data[index];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index row, Index col)
+  {
+    if (IsRowMajor)
+      return const_cast<Scalar*>(m_d.data)[row * m_d.outerStride() + col];
+    else
+      return const_cast<Scalar*>(m_d.data)[row + col * m_d.outerStride()];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index index)
+  {
+    return const_cast<Scalar*>(m_d.data)[index];
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index row, Index col) const
+  {
+    if (IsRowMajor)
+      return ploadt<PacketType, LoadMode>(m_d.data + row * m_d.outerStride() + col);
+    else
+      return ploadt<PacketType, LoadMode>(m_d.data + row + col * m_d.outerStride());
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index index) const
+  {
+    return ploadt<PacketType, LoadMode>(m_d.data + index);
+  }
+
+  template<int StoreMode,typename PacketType>
+  EIGEN_STRONG_INLINE
+  void writePacket(Index row, Index col, const PacketType& x)
+  {
+    if (IsRowMajor)
+      return pstoret<Scalar, PacketType, StoreMode>
+	            (const_cast<Scalar*>(m_d.data) + row * m_d.outerStride() + col, x);
+    else
+      return pstoret<Scalar, PacketType, StoreMode>
+                    (const_cast<Scalar*>(m_d.data) + row + col * m_d.outerStride(), x);
+  }
+
+  template<int StoreMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  void writePacket(Index index, const PacketType& x)
+  {
+    return pstoret<Scalar, PacketType, StoreMode>(const_cast<Scalar*>(m_d.data) + index, x);
+  }
+
+protected:
+
+  plainobjectbase_evaluator_data<Scalar,OuterStrideAtCompileTime> m_d;
+};
+
+template<typename Scalar, int Rows, int Cols, int Options, int MaxRows, int MaxCols>
+struct evaluator<Matrix<Scalar, Rows, Cols, Options, MaxRows, MaxCols> >
+  : evaluator<PlainObjectBase<Matrix<Scalar, Rows, Cols, Options, MaxRows, MaxCols> > >
+{
+  typedef Matrix<Scalar, Rows, Cols, Options, MaxRows, MaxCols> XprType;
+  
+  EIGEN_DEVICE_FUNC evaluator() {}
+
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& m)
+    : evaluator<PlainObjectBase<XprType> >(m) 
+  { }
+};
+
+template<typename Scalar, int Rows, int Cols, int Options, int MaxRows, int MaxCols>
+struct evaluator<Array<Scalar, Rows, Cols, Options, MaxRows, MaxCols> >
+  : evaluator<PlainObjectBase<Array<Scalar, Rows, Cols, Options, MaxRows, MaxCols> > >
+{
+  typedef Array<Scalar, Rows, Cols, Options, MaxRows, MaxCols> XprType;
+
+  EIGEN_DEVICE_FUNC evaluator() {}
+  
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& m)
+    : evaluator<PlainObjectBase<XprType> >(m) 
+  { }
+};
+
+// -------------------- Transpose --------------------
+
+template<typename ArgType>
+struct unary_evaluator<Transpose<ArgType>, IndexBased>
+  : evaluator_base<Transpose<ArgType> >
+{
+  typedef Transpose<ArgType> XprType;
+  
+  enum {
+    CoeffReadCost = evaluator<ArgType>::CoeffReadCost,    
+    Flags = evaluator<ArgType>::Flags ^ RowMajorBit,
+    Alignment = evaluator<ArgType>::Alignment
+  };
+
+  EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& t) : m_argImpl(t.nestedExpression()) {}
+
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index col) const
+  {
+    return m_argImpl.coeff(col, row);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  {
+    return m_argImpl.coeff(index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index row, Index col)
+  {
+    return m_argImpl.coeffRef(col, row);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  typename XprType::Scalar& coeffRef(Index index)
+  {
+    return m_argImpl.coeffRef(index);
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index row, Index col) const
+  {
+    return m_argImpl.template packet<LoadMode,PacketType>(col, row);
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index index) const
+  {
+    return m_argImpl.template packet<LoadMode,PacketType>(index);
+  }
+
+  template<int StoreMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  void writePacket(Index row, Index col, const PacketType& x)
+  {
+    m_argImpl.template writePacket<StoreMode,PacketType>(col, row, x);
+  }
+
+  template<int StoreMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  void writePacket(Index index, const PacketType& x)
+  {
+    m_argImpl.template writePacket<StoreMode,PacketType>(index, x);
+  }
+
+protected:
+  evaluator<ArgType> m_argImpl;
+};
+
+// -------------------- CwiseNullaryOp --------------------
+// Like Matrix and Array, this is not really a unary expression, so we directly specialize evaluator.
+// Likewise, there is not need to more sophisticated dispatching here.
+
+template<typename Scalar,typename NullaryOp,
+         bool has_nullary = has_nullary_operator<NullaryOp>::value,
+         bool has_unary   = has_unary_operator<NullaryOp>::value,
+         bool has_binary  = has_binary_operator<NullaryOp>::value>
+struct nullary_wrapper
+{
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i, IndexType j) const { return op(i,j); }
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i) const { return op(i); }
+
+  template <typename T, typename IndexType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i, IndexType j) const { return op.template packetOp<T>(i,j); }
+  template <typename T, typename IndexType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i) const { return op.template packetOp<T>(i); }
+};
+
+template<typename Scalar,typename NullaryOp>
+struct nullary_wrapper<Scalar,NullaryOp,true,false,false>
+{
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType=0, IndexType=0) const { return op(); }
+  template <typename T, typename IndexType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType=0, IndexType=0) const { return op.template packetOp<T>(); }
+};
+
+template<typename Scalar,typename NullaryOp>
+struct nullary_wrapper<Scalar,NullaryOp,false,false,true>
+{
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i, IndexType j=0) const { return op(i,j); }
+  template <typename T, typename IndexType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i, IndexType j=0) const { return op.template packetOp<T>(i,j); }
+};
+
+// We need the following specialization for vector-only functors assigned to a runtime vector,
+// for instance, using linspace and assigning a RowVectorXd to a MatrixXd or even a row of a MatrixXd.
+// In this case, i==0 and j is used for the actual iteration.
+template<typename Scalar,typename NullaryOp>
+struct nullary_wrapper<Scalar,NullaryOp,false,true,false>
+{
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i, IndexType j) const {
+    eigen_assert(i==0 || j==0);
+    return op(i+j);
+  }
+  template <typename T, typename IndexType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i, IndexType j) const {
+    eigen_assert(i==0 || j==0);
+    return op.template packetOp<T>(i+j);
+  }
+
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i) const { return op(i); }
+  template <typename T, typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i) const { return op.template packetOp<T>(i); }
+};
+
+template<typename Scalar,typename NullaryOp>
+struct nullary_wrapper<Scalar,NullaryOp,false,false,false> {};
+
+#if 0 && EIGEN_COMP_MSVC>0
+// Disable this ugly workaround. This is now handled in traits<Ref>::match,
+// but this piece of code might still become handly if some other weird compilation
+// erros pop up again.
+
+// MSVC exhibits a weird compilation error when
+// compiling:
+//    Eigen::MatrixXf A = MatrixXf::Random(3,3);
+//    Ref<const MatrixXf> R = 2.f*A;
+// and that has_*ary_operator<scalar_constant_op<float>> have not been instantiated yet.
+// The "problem" is that evaluator<2.f*A> is instantiated by traits<Ref>::match<2.f*A>
+// and at that time has_*ary_operator<T> returns true regardless of T.
+// Then nullary_wrapper is badly instantiated as nullary_wrapper<.,.,true,true,true>.
+// The trick is thus to defer the proper instantiation of nullary_wrapper when coeff(),
+// and packet() are really instantiated as implemented below:
+
+// This is a simple wrapper around Index to enforce the re-instantiation of
+// has_*ary_operator when needed.
+template<typename T> struct nullary_wrapper_workaround_msvc {
+  nullary_wrapper_workaround_msvc(const T&);
+  operator T()const;
+};
+
+template<typename Scalar,typename NullaryOp>
+struct nullary_wrapper<Scalar,NullaryOp,true,true,true>
+{
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i, IndexType j) const {
+    return nullary_wrapper<Scalar,NullaryOp,
+    has_nullary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value,
+    has_unary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value,
+    has_binary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value>().operator()(op,i,j);
+  }
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i) const {
+    return nullary_wrapper<Scalar,NullaryOp,
+    has_nullary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value,
+    has_unary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value,
+    has_binary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value>().operator()(op,i);
+  }
+
+  template <typename T, typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i, IndexType j) const {
+    return nullary_wrapper<Scalar,NullaryOp,
+    has_nullary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value,
+    has_unary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value,
+    has_binary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value>().template packetOp<T>(op,i,j);
+  }
+  template <typename T, typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i) const {
+    return nullary_wrapper<Scalar,NullaryOp,
+    has_nullary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value,
+    has_unary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value,
+    has_binary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value>().template packetOp<T>(op,i);
+  }
+};
+#endif // MSVC workaround
+
+template<typename NullaryOp, typename PlainObjectType>
+struct evaluator<CwiseNullaryOp<NullaryOp,PlainObjectType> >
+  : evaluator_base<CwiseNullaryOp<NullaryOp,PlainObjectType> >
+{
+  typedef CwiseNullaryOp<NullaryOp,PlainObjectType> XprType;
+  typedef typename internal::remove_all<PlainObjectType>::type PlainObjectTypeCleaned;
+  
+  enum {
+    CoeffReadCost = internal::functor_traits<NullaryOp>::Cost,
+    
+    Flags = (evaluator<PlainObjectTypeCleaned>::Flags
+          &  (  HereditaryBits
+              | (functor_has_linear_access<NullaryOp>::ret  ? LinearAccessBit : 0)
+              | (functor_traits<NullaryOp>::PacketAccess    ? PacketAccessBit : 0)))
+          | (functor_traits<NullaryOp>::IsRepeatable ? 0 : EvalBeforeNestingBit),
+    Alignment = AlignedMax
+  };
+
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& n)
+    : m_functor(n.functor()), m_wrapper()
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(IndexType row, IndexType col) const
+  {
+    return m_wrapper(m_functor, row, col);
+  }
+
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(IndexType index) const
+  {
+    return m_wrapper(m_functor,index);
+  }
+
+  template<int LoadMode, typename PacketType, typename IndexType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(IndexType row, IndexType col) const
+  {
+    return m_wrapper.template packetOp<PacketType>(m_functor, row, col);
+  }
+
+  template<int LoadMode, typename PacketType, typename IndexType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(IndexType index) const
+  {
+    return m_wrapper.template packetOp<PacketType>(m_functor, index);
+  }
+
+protected:
+  const NullaryOp m_functor;
+  const internal::nullary_wrapper<CoeffReturnType,NullaryOp> m_wrapper;
+};
+
+// -------------------- CwiseUnaryOp --------------------
+
+template<typename UnaryOp, typename ArgType>
+struct unary_evaluator<CwiseUnaryOp<UnaryOp, ArgType>, IndexBased >
+  : evaluator_base<CwiseUnaryOp<UnaryOp, ArgType> >
+{
+  typedef CwiseUnaryOp<UnaryOp, ArgType> XprType;
+  
+  enum {
+    CoeffReadCost = evaluator<ArgType>::CoeffReadCost + functor_traits<UnaryOp>::Cost,
+    
+    Flags = evaluator<ArgType>::Flags
+          & (HereditaryBits | LinearAccessBit | (functor_traits<UnaryOp>::PacketAccess ? PacketAccessBit : 0)),
+    Alignment = evaluator<ArgType>::Alignment
+  };
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  explicit unary_evaluator(const XprType& op) : m_d(op)
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(functor_traits<UnaryOp>::Cost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index col) const
+  {
+    return m_d.func()(m_d.argImpl.coeff(row, col));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  {
+    return m_d.func()(m_d.argImpl.coeff(index));
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index row, Index col) const
+  {
+    return m_d.func().packetOp(m_d.argImpl.template packet<LoadMode, PacketType>(row, col));
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index index) const
+  {
+    return m_d.func().packetOp(m_d.argImpl.template packet<LoadMode, PacketType>(index));
+  }
+
+protected:
+
+  // this helper permits to completely eliminate the functor if it is empty
+  class Data : private UnaryOp
+  {
+  public:
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Data(const XprType& xpr) : UnaryOp(xpr.functor()), argImpl(xpr.nestedExpression()) {}
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const UnaryOp& func() const { return static_cast<const UnaryOp&>(*this); }
+    evaluator<ArgType> argImpl;
+  };
+
+  Data m_d;
+};
+
+// -------------------- CwiseTernaryOp --------------------
+
+// this is a ternary expression
+template<typename TernaryOp, typename Arg1, typename Arg2, typename Arg3>
+struct evaluator<CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> >
+  : public ternary_evaluator<CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> >
+{
+  typedef CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> XprType;
+  typedef ternary_evaluator<CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> > Base;
+  
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr) : Base(xpr) {}
+};
+
+template<typename TernaryOp, typename Arg1, typename Arg2, typename Arg3>
+struct ternary_evaluator<CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3>, IndexBased, IndexBased>
+  : evaluator_base<CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> >
+{
+  typedef CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> XprType;
+  
+  enum {
+    CoeffReadCost = evaluator<Arg1>::CoeffReadCost + evaluator<Arg2>::CoeffReadCost + evaluator<Arg3>::CoeffReadCost + functor_traits<TernaryOp>::Cost,
+    
+    Arg1Flags = evaluator<Arg1>::Flags,
+    Arg2Flags = evaluator<Arg2>::Flags,
+    Arg3Flags = evaluator<Arg3>::Flags,
+    SameType = is_same<typename Arg1::Scalar,typename Arg2::Scalar>::value && is_same<typename Arg1::Scalar,typename Arg3::Scalar>::value,
+    StorageOrdersAgree = (int(Arg1Flags)&RowMajorBit)==(int(Arg2Flags)&RowMajorBit) && (int(Arg1Flags)&RowMajorBit)==(int(Arg3Flags)&RowMajorBit),
+    Flags0 = (int(Arg1Flags) | int(Arg2Flags) | int(Arg3Flags)) & (
+        HereditaryBits
+        | (int(Arg1Flags) & int(Arg2Flags) & int(Arg3Flags) &
+           ( (StorageOrdersAgree ? LinearAccessBit : 0)
+           | (functor_traits<TernaryOp>::PacketAccess && StorageOrdersAgree && SameType ? PacketAccessBit : 0)
+           )
+        )
+     ),
+    Flags = (Flags0 & ~RowMajorBit) | (Arg1Flags & RowMajorBit),
+    Alignment = EIGEN_PLAIN_ENUM_MIN(
+        EIGEN_PLAIN_ENUM_MIN(evaluator<Arg1>::Alignment, evaluator<Arg2>::Alignment),
+        evaluator<Arg3>::Alignment)
+  };
+
+  EIGEN_DEVICE_FUNC explicit ternary_evaluator(const XprType& xpr) : m_d(xpr)
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(functor_traits<TernaryOp>::Cost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index col) const
+  {
+    return m_d.func()(m_d.arg1Impl.coeff(row, col), m_d.arg2Impl.coeff(row, col), m_d.arg3Impl.coeff(row, col));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  {
+    return m_d.func()(m_d.arg1Impl.coeff(index), m_d.arg2Impl.coeff(index), m_d.arg3Impl.coeff(index));
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index row, Index col) const
+  {
+    return m_d.func().packetOp(m_d.arg1Impl.template packet<LoadMode,PacketType>(row, col),
+                               m_d.arg2Impl.template packet<LoadMode,PacketType>(row, col),
+                               m_d.arg3Impl.template packet<LoadMode,PacketType>(row, col));
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index index) const
+  {
+    return m_d.func().packetOp(m_d.arg1Impl.template packet<LoadMode,PacketType>(index),
+                               m_d.arg2Impl.template packet<LoadMode,PacketType>(index),
+                               m_d.arg3Impl.template packet<LoadMode,PacketType>(index));
+  }
+
+protected:
+  // this helper permits to completely eliminate the functor if it is empty
+  struct Data : private TernaryOp
+  {
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Data(const XprType& xpr) : TernaryOp(xpr.functor()), arg1Impl(xpr.arg1()), arg2Impl(xpr.arg2()), arg3Impl(xpr.arg3()) {}
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TernaryOp& func() const { return static_cast<const TernaryOp&>(*this); }
+    evaluator<Arg1> arg1Impl;
+    evaluator<Arg2> arg2Impl;
+    evaluator<Arg3> arg3Impl;
+  };
+
+  Data m_d;
+};
+
+// -------------------- CwiseBinaryOp --------------------
+
+// this is a binary expression
+template<typename BinaryOp, typename Lhs, typename Rhs>
+struct evaluator<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >
+  : public binary_evaluator<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >
+{
+  typedef CwiseBinaryOp<BinaryOp, Lhs, Rhs> XprType;
+  typedef binary_evaluator<CwiseBinaryOp<BinaryOp, Lhs, Rhs> > Base;
+  
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr) : Base(xpr) {}
+};
+
+template<typename BinaryOp, typename Lhs, typename Rhs>
+struct binary_evaluator<CwiseBinaryOp<BinaryOp, Lhs, Rhs>, IndexBased, IndexBased>
+  : evaluator_base<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >
+{
+  typedef CwiseBinaryOp<BinaryOp, Lhs, Rhs> XprType;
+  
+  enum {
+    CoeffReadCost = evaluator<Lhs>::CoeffReadCost + evaluator<Rhs>::CoeffReadCost + functor_traits<BinaryOp>::Cost,
+    
+    LhsFlags = evaluator<Lhs>::Flags,
+    RhsFlags = evaluator<Rhs>::Flags,
+    SameType = is_same<typename Lhs::Scalar,typename Rhs::Scalar>::value,
+    StorageOrdersAgree = (int(LhsFlags)&RowMajorBit)==(int(RhsFlags)&RowMajorBit),
+    Flags0 = (int(LhsFlags) | int(RhsFlags)) & (
+        HereditaryBits
+      | (int(LhsFlags) & int(RhsFlags) &
+           ( (StorageOrdersAgree ? LinearAccessBit : 0)
+           | (functor_traits<BinaryOp>::PacketAccess && StorageOrdersAgree && SameType ? PacketAccessBit : 0)
+           )
+        )
+     ),
+    Flags = (Flags0 & ~RowMajorBit) | (LhsFlags & RowMajorBit),
+    Alignment = EIGEN_PLAIN_ENUM_MIN(evaluator<Lhs>::Alignment,evaluator<Rhs>::Alignment)
+  };
+
+  EIGEN_DEVICE_FUNC explicit binary_evaluator(const XprType& xpr) : m_d(xpr)
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(functor_traits<BinaryOp>::Cost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index col) const
+  {
+    return m_d.func()(m_d.lhsImpl.coeff(row, col), m_d.rhsImpl.coeff(row, col));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  {
+    return m_d.func()(m_d.lhsImpl.coeff(index), m_d.rhsImpl.coeff(index));
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index row, Index col) const
+  {
+    return m_d.func().packetOp(m_d.lhsImpl.template packet<LoadMode,PacketType>(row, col),
+                               m_d.rhsImpl.template packet<LoadMode,PacketType>(row, col));
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index index) const
+  {
+    return m_d.func().packetOp(m_d.lhsImpl.template packet<LoadMode,PacketType>(index),
+                               m_d.rhsImpl.template packet<LoadMode,PacketType>(index));
+  }
+
+protected:
+
+  // this helper permits to completely eliminate the functor if it is empty
+  struct Data : private BinaryOp
+  {
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Data(const XprType& xpr) : BinaryOp(xpr.functor()), lhsImpl(xpr.lhs()), rhsImpl(xpr.rhs()) {}
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const BinaryOp& func() const { return static_cast<const BinaryOp&>(*this); }
+    evaluator<Lhs> lhsImpl;
+    evaluator<Rhs> rhsImpl;
+  };
+
+  Data m_d;
+};
+
+// -------------------- CwiseUnaryView --------------------
+
+template<typename UnaryOp, typename ArgType>
+struct unary_evaluator<CwiseUnaryView<UnaryOp, ArgType>, IndexBased>
+  : evaluator_base<CwiseUnaryView<UnaryOp, ArgType> >
+{
+  typedef CwiseUnaryView<UnaryOp, ArgType> XprType;
+  
+  enum {
+    CoeffReadCost = evaluator<ArgType>::CoeffReadCost + functor_traits<UnaryOp>::Cost,
+    
+    Flags = (evaluator<ArgType>::Flags & (HereditaryBits | LinearAccessBit | DirectAccessBit)),
+    
+    Alignment = 0 // FIXME it is not very clear why alignment is necessarily lost...
+  };
+
+  EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& op) : m_d(op)
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(functor_traits<UnaryOp>::Cost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index col) const
+  {
+    return m_d.func()(m_d.argImpl.coeff(row, col));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  {
+    return m_d.func()(m_d.argImpl.coeff(index));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index row, Index col)
+  {
+    return m_d.func()(m_d.argImpl.coeffRef(row, col));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index index)
+  {
+    return m_d.func()(m_d.argImpl.coeffRef(index));
+  }
+
+protected:
+
+  // this helper permits to completely eliminate the functor if it is empty
+  struct Data : private UnaryOp
+  {
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Data(const XprType& xpr) : UnaryOp(xpr.functor()), argImpl(xpr.nestedExpression()) {}
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const UnaryOp& func() const { return static_cast<const UnaryOp&>(*this); }
+    evaluator<ArgType> argImpl;
+  };
+
+  Data m_d;
+};
+
+// -------------------- Map --------------------
+
+// FIXME perhaps the PlainObjectType could be provided by Derived::PlainObject ?
+// but that might complicate template specialization
+template<typename Derived, typename PlainObjectType>
+struct mapbase_evaluator;
+
+template<typename Derived, typename PlainObjectType>
+struct mapbase_evaluator : evaluator_base<Derived>
+{
+  typedef Derived  XprType;
+  typedef typename XprType::PointerType PointerType;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  
+  enum {
+    IsRowMajor = XprType::RowsAtCompileTime,
+    ColsAtCompileTime = XprType::ColsAtCompileTime,
+    CoeffReadCost = NumTraits<Scalar>::ReadCost
+  };
+
+  EIGEN_DEVICE_FUNC explicit mapbase_evaluator(const XprType& map)
+    : m_data(const_cast<PointerType>(map.data())),
+      m_innerStride(map.innerStride()),
+      m_outerStride(map.outerStride())
+  {
+    EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(evaluator<Derived>::Flags&PacketAccessBit, internal::inner_stride_at_compile_time<Derived>::ret==1),
+                        PACKET_ACCESS_REQUIRES_TO_HAVE_INNER_STRIDE_FIXED_TO_1);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index col) const
+  {
+    return m_data[col * colStride() + row * rowStride()];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  {
+    return m_data[index * m_innerStride.value()];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index row, Index col)
+  {
+    return m_data[col * colStride() + row * rowStride()];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index index)
+  {
+    return m_data[index * m_innerStride.value()];
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index row, Index col) const
+  {
+    PointerType ptr = m_data + row * rowStride() + col * colStride();
+    return internal::ploadt<PacketType, LoadMode>(ptr);
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index index) const
+  {
+    return internal::ploadt<PacketType, LoadMode>(m_data + index * m_innerStride.value());
+  }
+
+  template<int StoreMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  void writePacket(Index row, Index col, const PacketType& x)
+  {
+    PointerType ptr = m_data + row * rowStride() + col * colStride();
+    return internal::pstoret<Scalar, PacketType, StoreMode>(ptr, x);
+  }
+
+  template<int StoreMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  void writePacket(Index index, const PacketType& x)
+  {
+    internal::pstoret<Scalar, PacketType, StoreMode>(m_data + index * m_innerStride.value(), x);
+  }
+protected:
+  EIGEN_DEVICE_FUNC
+  inline Index rowStride() const { return XprType::IsRowMajor ? m_outerStride.value() : m_innerStride.value(); }
+  EIGEN_DEVICE_FUNC
+  inline Index colStride() const { return XprType::IsRowMajor ? m_innerStride.value() : m_outerStride.value(); }
+
+  PointerType m_data;
+  const internal::variable_if_dynamic<Index, XprType::InnerStrideAtCompileTime> m_innerStride;
+  const internal::variable_if_dynamic<Index, XprType::OuterStrideAtCompileTime> m_outerStride;
+};
+
+template<typename PlainObjectType, int MapOptions, typename StrideType> 
+struct evaluator<Map<PlainObjectType, MapOptions, StrideType> >
+  : public mapbase_evaluator<Map<PlainObjectType, MapOptions, StrideType>, PlainObjectType>
+{
+  typedef Map<PlainObjectType, MapOptions, StrideType> XprType;
+  typedef typename XprType::Scalar Scalar;
+  // TODO: should check for smaller packet types once we can handle multi-sized packet types
+  typedef typename packet_traits<Scalar>::type PacketScalar;
+  
+  enum {
+    InnerStrideAtCompileTime = StrideType::InnerStrideAtCompileTime == 0
+                             ? int(PlainObjectType::InnerStrideAtCompileTime)
+                             : int(StrideType::InnerStrideAtCompileTime),
+    OuterStrideAtCompileTime = StrideType::OuterStrideAtCompileTime == 0
+                             ? int(PlainObjectType::OuterStrideAtCompileTime)
+                             : int(StrideType::OuterStrideAtCompileTime),
+    HasNoInnerStride = InnerStrideAtCompileTime == 1,
+    HasNoOuterStride = StrideType::OuterStrideAtCompileTime == 0,
+    HasNoStride = HasNoInnerStride && HasNoOuterStride,
+    IsDynamicSize = PlainObjectType::SizeAtCompileTime==Dynamic,
+    
+    PacketAccessMask = bool(HasNoInnerStride) ? ~int(0) : ~int(PacketAccessBit),
+    LinearAccessMask = bool(HasNoStride) || bool(PlainObjectType::IsVectorAtCompileTime) ? ~int(0) : ~int(LinearAccessBit),
+    Flags = int( evaluator<PlainObjectType>::Flags) & (LinearAccessMask&PacketAccessMask),
+    
+    Alignment = int(MapOptions)&int(AlignedMask)
+  };
+
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& map)
+    : mapbase_evaluator<XprType, PlainObjectType>(map) 
+  { }
+};
+
+// -------------------- Ref --------------------
+
+template<typename PlainObjectType, int RefOptions, typename StrideType> 
+struct evaluator<Ref<PlainObjectType, RefOptions, StrideType> >
+  : public mapbase_evaluator<Ref<PlainObjectType, RefOptions, StrideType>, PlainObjectType>
+{
+  typedef Ref<PlainObjectType, RefOptions, StrideType> XprType;
+  
+  enum {
+    Flags = evaluator<Map<PlainObjectType, RefOptions, StrideType> >::Flags,
+    Alignment = evaluator<Map<PlainObjectType, RefOptions, StrideType> >::Alignment
+  };
+
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& ref)
+    : mapbase_evaluator<XprType, PlainObjectType>(ref) 
+  { }
+};
+
+// -------------------- Block --------------------
+
+template<typename ArgType, int BlockRows, int BlockCols, bool InnerPanel,
+         bool HasDirectAccess = internal::has_direct_access<ArgType>::ret> struct block_evaluator;
+         
+template<typename ArgType, int BlockRows, int BlockCols, bool InnerPanel> 
+struct evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel> >
+  : block_evaluator<ArgType, BlockRows, BlockCols, InnerPanel>
+{
+  typedef Block<ArgType, BlockRows, BlockCols, InnerPanel> XprType;
+  typedef typename XprType::Scalar Scalar;
+  // TODO: should check for smaller packet types once we can handle multi-sized packet types
+  typedef typename packet_traits<Scalar>::type PacketScalar;
+  
+  enum {
+    CoeffReadCost = evaluator<ArgType>::CoeffReadCost,
+    
+    RowsAtCompileTime = traits<XprType>::RowsAtCompileTime,
+    ColsAtCompileTime = traits<XprType>::ColsAtCompileTime,
+    MaxRowsAtCompileTime = traits<XprType>::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = traits<XprType>::MaxColsAtCompileTime,
+    
+    ArgTypeIsRowMajor = (int(evaluator<ArgType>::Flags)&RowMajorBit) != 0,
+    IsRowMajor = (MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1) ? 1
+               : (MaxColsAtCompileTime==1 && MaxRowsAtCompileTime!=1) ? 0
+               : ArgTypeIsRowMajor,
+    HasSameStorageOrderAsArgType = (IsRowMajor == ArgTypeIsRowMajor),
+    InnerSize = IsRowMajor ? int(ColsAtCompileTime) : int(RowsAtCompileTime),
+    InnerStrideAtCompileTime = HasSameStorageOrderAsArgType
+                             ? int(inner_stride_at_compile_time<ArgType>::ret)
+                             : int(outer_stride_at_compile_time<ArgType>::ret),
+    OuterStrideAtCompileTime = HasSameStorageOrderAsArgType
+                             ? int(outer_stride_at_compile_time<ArgType>::ret)
+                             : int(inner_stride_at_compile_time<ArgType>::ret),
+    MaskPacketAccessBit = (InnerStrideAtCompileTime == 1) ? PacketAccessBit : 0,
+    
+    FlagsLinearAccessBit = (RowsAtCompileTime == 1 || ColsAtCompileTime == 1 || (InnerPanel && (evaluator<ArgType>::Flags&LinearAccessBit))) ? LinearAccessBit : 0,    
+    FlagsRowMajorBit = XprType::Flags&RowMajorBit,
+    Flags0 = evaluator<ArgType>::Flags & ( (HereditaryBits & ~RowMajorBit) |
+                                           DirectAccessBit |
+                                           MaskPacketAccessBit),
+    Flags = Flags0 | FlagsLinearAccessBit | FlagsRowMajorBit,
+    
+    PacketAlignment = unpacket_traits<PacketScalar>::alignment,
+    Alignment0 = (InnerPanel && (OuterStrideAtCompileTime!=Dynamic) && (((OuterStrideAtCompileTime * int(sizeof(Scalar))) % int(PacketAlignment)) == 0)) ? int(PacketAlignment) : 0,
+    Alignment = EIGEN_PLAIN_ENUM_MIN(evaluator<ArgType>::Alignment, Alignment0)
+  };
+  typedef block_evaluator<ArgType, BlockRows, BlockCols, InnerPanel> block_evaluator_type;
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& block) : block_evaluator_type(block)
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+};
+
+// no direct-access => dispatch to a unary evaluator
+template<typename ArgType, int BlockRows, int BlockCols, bool InnerPanel>
+struct block_evaluator<ArgType, BlockRows, BlockCols, InnerPanel, /*HasDirectAccess*/ false>
+  : unary_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel> >
+{
+  typedef Block<ArgType, BlockRows, BlockCols, InnerPanel> XprType;
+
+  EIGEN_DEVICE_FUNC explicit block_evaluator(const XprType& block)
+    : unary_evaluator<XprType>(block) 
+  {}
+};
+
+template<typename ArgType, int BlockRows, int BlockCols, bool InnerPanel>
+struct unary_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel>, IndexBased>
+  : evaluator_base<Block<ArgType, BlockRows, BlockCols, InnerPanel> >
+{
+  typedef Block<ArgType, BlockRows, BlockCols, InnerPanel> XprType;
+
+  EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& block)
+    : m_argImpl(block.nestedExpression()), 
+      m_startRow(block.startRow()), 
+      m_startCol(block.startCol()) 
+  { }
+ 
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  enum {
+    RowsAtCompileTime = XprType::RowsAtCompileTime
+  };
+ 
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index col) const
+  { 
+    return m_argImpl.coeff(m_startRow.value() + row, m_startCol.value() + col); 
+  }
+  
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  { 
+    return coeff(RowsAtCompileTime == 1 ? 0 : index, RowsAtCompileTime == 1 ? index : 0);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index row, Index col)
+  { 
+    return m_argImpl.coeffRef(m_startRow.value() + row, m_startCol.value() + col); 
+  }
+  
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index index)
+  { 
+    return coeffRef(RowsAtCompileTime == 1 ? 0 : index, RowsAtCompileTime == 1 ? index : 0);
+  }
+ 
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index row, Index col) const 
+  { 
+    return m_argImpl.template packet<LoadMode,PacketType>(m_startRow.value() + row, m_startCol.value() + col); 
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index index) const 
+  { 
+    return packet<LoadMode,PacketType>(RowsAtCompileTime == 1 ? 0 : index,
+                                       RowsAtCompileTime == 1 ? index : 0);
+  }
+  
+  template<int StoreMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  void writePacket(Index row, Index col, const PacketType& x) 
+  {
+    return m_argImpl.template writePacket<StoreMode,PacketType>(m_startRow.value() + row, m_startCol.value() + col, x); 
+  }
+  
+  template<int StoreMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  void writePacket(Index index, const PacketType& x) 
+  {
+    return writePacket<StoreMode,PacketType>(RowsAtCompileTime == 1 ? 0 : index,
+                                             RowsAtCompileTime == 1 ? index : 0,
+                                             x);
+  }
+ 
+protected:
+  evaluator<ArgType> m_argImpl;
+  const variable_if_dynamic<Index, (ArgType::RowsAtCompileTime == 1 && BlockRows==1) ? 0 : Dynamic> m_startRow;
+  const variable_if_dynamic<Index, (ArgType::ColsAtCompileTime == 1 && BlockCols==1) ? 0 : Dynamic> m_startCol;
+};
+
+// TODO: This evaluator does not actually use the child evaluator; 
+// all action is via the data() as returned by the Block expression.
+
+template<typename ArgType, int BlockRows, int BlockCols, bool InnerPanel> 
+struct block_evaluator<ArgType, BlockRows, BlockCols, InnerPanel, /* HasDirectAccess */ true>
+  : mapbase_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel>,
+                      typename Block<ArgType, BlockRows, BlockCols, InnerPanel>::PlainObject>
+{
+  typedef Block<ArgType, BlockRows, BlockCols, InnerPanel> XprType;
+  typedef typename XprType::Scalar Scalar;
+
+  EIGEN_DEVICE_FUNC explicit block_evaluator(const XprType& block)
+    : mapbase_evaluator<XprType, typename XprType::PlainObject>(block) 
+  {
+    // TODO: for the 3.3 release, this should be turned to an internal assertion, but let's keep it as is for the beta lifetime
+    eigen_assert(((internal::UIntPtr(block.data()) % EIGEN_PLAIN_ENUM_MAX(1,evaluator<XprType>::Alignment)) == 0) && "data is not aligned");
+  }
+};
+
+
+// -------------------- Select --------------------
+// NOTE shall we introduce a ternary_evaluator?
+
+// TODO enable vectorization for Select
+template<typename ConditionMatrixType, typename ThenMatrixType, typename ElseMatrixType>
+struct evaluator<Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> >
+  : evaluator_base<Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> >
+{
+  typedef Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> XprType;
+  enum {
+    CoeffReadCost = evaluator<ConditionMatrixType>::CoeffReadCost
+                  + EIGEN_PLAIN_ENUM_MAX(evaluator<ThenMatrixType>::CoeffReadCost,
+                                         evaluator<ElseMatrixType>::CoeffReadCost),
+
+    Flags = (unsigned int)evaluator<ThenMatrixType>::Flags & evaluator<ElseMatrixType>::Flags & HereditaryBits,
+    
+    Alignment = EIGEN_PLAIN_ENUM_MIN(evaluator<ThenMatrixType>::Alignment, evaluator<ElseMatrixType>::Alignment)
+  };
+
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& select)
+    : m_conditionImpl(select.conditionMatrix()),
+      m_thenImpl(select.thenMatrix()),
+      m_elseImpl(select.elseMatrix())
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+ 
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index col) const
+  {
+    if (m_conditionImpl.coeff(row, col))
+      return m_thenImpl.coeff(row, col);
+    else
+      return m_elseImpl.coeff(row, col);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  {
+    if (m_conditionImpl.coeff(index))
+      return m_thenImpl.coeff(index);
+    else
+      return m_elseImpl.coeff(index);
+  }
+ 
+protected:
+  evaluator<ConditionMatrixType> m_conditionImpl;
+  evaluator<ThenMatrixType> m_thenImpl;
+  evaluator<ElseMatrixType> m_elseImpl;
+};
+
+
+// -------------------- Replicate --------------------
+
+template<typename ArgType, int RowFactor, int ColFactor> 
+struct unary_evaluator<Replicate<ArgType, RowFactor, ColFactor> >
+  : evaluator_base<Replicate<ArgType, RowFactor, ColFactor> >
+{
+  typedef Replicate<ArgType, RowFactor, ColFactor> XprType;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  enum {
+    Factor = (RowFactor==Dynamic || ColFactor==Dynamic) ? Dynamic : RowFactor*ColFactor
+  };
+  typedef typename internal::nested_eval<ArgType,Factor>::type ArgTypeNested;
+  typedef typename internal::remove_all<ArgTypeNested>::type ArgTypeNestedCleaned;
+  
+  enum {
+    CoeffReadCost = evaluator<ArgTypeNestedCleaned>::CoeffReadCost,
+    LinearAccessMask = XprType::IsVectorAtCompileTime ? LinearAccessBit : 0,
+    Flags = (evaluator<ArgTypeNestedCleaned>::Flags & (HereditaryBits|LinearAccessMask) & ~RowMajorBit) | (traits<XprType>::Flags & RowMajorBit),
+    
+    Alignment = evaluator<ArgTypeNestedCleaned>::Alignment
+  };
+
+  EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& replicate)
+    : m_arg(replicate.nestedExpression()),
+      m_argImpl(m_arg),
+      m_rows(replicate.nestedExpression().rows()),
+      m_cols(replicate.nestedExpression().cols())
+  {}
+ 
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index col) const
+  {
+    // try to avoid using modulo; this is a pure optimization strategy
+    const Index actual_row = internal::traits<XprType>::RowsAtCompileTime==1 ? 0
+                           : RowFactor==1 ? row
+                           : row % m_rows.value();
+    const Index actual_col = internal::traits<XprType>::ColsAtCompileTime==1 ? 0
+                           : ColFactor==1 ? col
+                           : col % m_cols.value();
+    
+    return m_argImpl.coeff(actual_row, actual_col);
+  }
+  
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  {
+    // try to avoid using modulo; this is a pure optimization strategy
+    const Index actual_index = internal::traits<XprType>::RowsAtCompileTime==1
+                                  ? (ColFactor==1 ?  index : index%m_cols.value())
+                                  : (RowFactor==1 ?  index : index%m_rows.value());
+    
+    return m_argImpl.coeff(actual_index);
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index row, Index col) const
+  {
+    const Index actual_row = internal::traits<XprType>::RowsAtCompileTime==1 ? 0
+                           : RowFactor==1 ? row
+                           : row % m_rows.value();
+    const Index actual_col = internal::traits<XprType>::ColsAtCompileTime==1 ? 0
+                           : ColFactor==1 ? col
+                           : col % m_cols.value();
+
+    return m_argImpl.template packet<LoadMode,PacketType>(actual_row, actual_col);
+  }
+  
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index index) const
+  {
+    const Index actual_index = internal::traits<XprType>::RowsAtCompileTime==1
+                                  ? (ColFactor==1 ?  index : index%m_cols.value())
+                                  : (RowFactor==1 ?  index : index%m_rows.value());
+
+    return m_argImpl.template packet<LoadMode,PacketType>(actual_index);
+  }
+ 
+protected:
+  const ArgTypeNested m_arg;
+  evaluator<ArgTypeNestedCleaned> m_argImpl;
+  const variable_if_dynamic<Index, ArgType::RowsAtCompileTime> m_rows;
+  const variable_if_dynamic<Index, ArgType::ColsAtCompileTime> m_cols;
+};
+
+
+// -------------------- PartialReduxExpr --------------------
+
+template< typename ArgType, typename MemberOp, int Direction>
+struct evaluator<PartialReduxExpr<ArgType, MemberOp, Direction> >
+  : evaluator_base<PartialReduxExpr<ArgType, MemberOp, Direction> >
+{
+  typedef PartialReduxExpr<ArgType, MemberOp, Direction> XprType;
+  typedef typename internal::nested_eval<ArgType,1>::type ArgTypeNested;
+  typedef typename internal::remove_all<ArgTypeNested>::type ArgTypeNestedCleaned;
+  typedef typename ArgType::Scalar InputScalar;
+  typedef typename XprType::Scalar Scalar;
+  enum {
+    TraversalSize = Direction==int(Vertical) ? int(ArgType::RowsAtCompileTime) :  int(ArgType::ColsAtCompileTime)
+  };
+  typedef typename MemberOp::template Cost<InputScalar,int(TraversalSize)> CostOpType;
+  enum {
+    CoeffReadCost = TraversalSize==Dynamic ? HugeCost
+                  : TraversalSize * evaluator<ArgType>::CoeffReadCost + int(CostOpType::value),
+    
+    Flags = (traits<XprType>::Flags&RowMajorBit) | (evaluator<ArgType>::Flags&(HereditaryBits&(~RowMajorBit))) | LinearAccessBit,
+    
+    Alignment = 0 // FIXME this will need to be improved once PartialReduxExpr is vectorized
+  };
+
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType xpr)
+    : m_arg(xpr.nestedExpression()), m_functor(xpr.functor())
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(TraversalSize==Dynamic ? HugeCost : int(CostOpType::value));
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  const Scalar coeff(Index i, Index j) const
+  {
+    if (Direction==Vertical)
+      return m_functor(m_arg.col(j));
+    else
+      return m_functor(m_arg.row(i));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  const Scalar coeff(Index index) const
+  {
+    if (Direction==Vertical)
+      return m_functor(m_arg.col(index));
+    else
+      return m_functor(m_arg.row(index));
+  }
+
+protected:
+  typename internal::add_const_on_value_type<ArgTypeNested>::type m_arg;
+  const MemberOp m_functor;
+};
+
+
+// -------------------- MatrixWrapper and ArrayWrapper --------------------
+//
+// evaluator_wrapper_base<T> is a common base class for the
+// MatrixWrapper and ArrayWrapper evaluators.
+
+template<typename XprType>
+struct evaluator_wrapper_base
+  : evaluator_base<XprType>
+{
+  typedef typename remove_all<typename XprType::NestedExpressionType>::type ArgType;
+  enum {
+    CoeffReadCost = evaluator<ArgType>::CoeffReadCost,
+    Flags = evaluator<ArgType>::Flags,
+    Alignment = evaluator<ArgType>::Alignment
+  };
+
+  EIGEN_DEVICE_FUNC explicit evaluator_wrapper_base(const ArgType& arg) : m_argImpl(arg) {}
+
+  typedef typename ArgType::Scalar Scalar;
+  typedef typename ArgType::CoeffReturnType CoeffReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index col) const
+  {
+    return m_argImpl.coeff(row, col);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  {
+    return m_argImpl.coeff(index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index row, Index col)
+  {
+    return m_argImpl.coeffRef(row, col);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index index)
+  {
+    return m_argImpl.coeffRef(index);
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index row, Index col) const
+  {
+    return m_argImpl.template packet<LoadMode,PacketType>(row, col);
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index index) const
+  {
+    return m_argImpl.template packet<LoadMode,PacketType>(index);
+  }
+
+  template<int StoreMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  void writePacket(Index row, Index col, const PacketType& x)
+  {
+    m_argImpl.template writePacket<StoreMode>(row, col, x);
+  }
+
+  template<int StoreMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  void writePacket(Index index, const PacketType& x)
+  {
+    m_argImpl.template writePacket<StoreMode>(index, x);
+  }
+
+protected:
+  evaluator<ArgType> m_argImpl;
+};
+
+template<typename TArgType>
+struct unary_evaluator<MatrixWrapper<TArgType> >
+  : evaluator_wrapper_base<MatrixWrapper<TArgType> >
+{
+  typedef MatrixWrapper<TArgType> XprType;
+
+  EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& wrapper)
+    : evaluator_wrapper_base<MatrixWrapper<TArgType> >(wrapper.nestedExpression())
+  { }
+};
+
+template<typename TArgType>
+struct unary_evaluator<ArrayWrapper<TArgType> >
+  : evaluator_wrapper_base<ArrayWrapper<TArgType> >
+{
+  typedef ArrayWrapper<TArgType> XprType;
+
+  EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& wrapper)
+    : evaluator_wrapper_base<ArrayWrapper<TArgType> >(wrapper.nestedExpression())
+  { }
+};
+
+
+// -------------------- Reverse --------------------
+
+// defined in Reverse.h:
+template<typename PacketType, bool ReversePacket> struct reverse_packet_cond;
+
+template<typename ArgType, int Direction>
+struct unary_evaluator<Reverse<ArgType, Direction> >
+  : evaluator_base<Reverse<ArgType, Direction> >
+{
+  typedef Reverse<ArgType, Direction> XprType;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  enum {
+    IsRowMajor = XprType::IsRowMajor,
+    IsColMajor = !IsRowMajor,
+    ReverseRow = (Direction == Vertical)   || (Direction == BothDirections),
+    ReverseCol = (Direction == Horizontal) || (Direction == BothDirections),
+    ReversePacket = (Direction == BothDirections)
+                    || ((Direction == Vertical)   && IsColMajor)
+                    || ((Direction == Horizontal) && IsRowMajor),
+                    
+    CoeffReadCost = evaluator<ArgType>::CoeffReadCost,
+    
+    // let's enable LinearAccess only with vectorization because of the product overhead
+    // FIXME enable DirectAccess with negative strides?
+    Flags0 = evaluator<ArgType>::Flags,
+    LinearAccess = ( (Direction==BothDirections) && (int(Flags0)&PacketAccessBit) )
+                  || ((ReverseRow && XprType::ColsAtCompileTime==1) || (ReverseCol && XprType::RowsAtCompileTime==1))
+                 ? LinearAccessBit : 0,
+
+    Flags = int(Flags0) & (HereditaryBits | PacketAccessBit | LinearAccess),
+    
+    Alignment = 0 // FIXME in some rare cases, Alignment could be preserved, like a Vector4f.
+  };
+
+  EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& reverse)
+    : m_argImpl(reverse.nestedExpression()),
+      m_rows(ReverseRow ? reverse.nestedExpression().rows() : 1),
+      m_cols(ReverseCol ? reverse.nestedExpression().cols() : 1)
+  { }
+ 
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index col) const
+  {
+    return m_argImpl.coeff(ReverseRow ? m_rows.value() - row - 1 : row,
+                           ReverseCol ? m_cols.value() - col - 1 : col);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  {
+    return m_argImpl.coeff(m_rows.value() * m_cols.value() - index - 1);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index row, Index col)
+  {
+    return m_argImpl.coeffRef(ReverseRow ? m_rows.value() - row - 1 : row,
+                              ReverseCol ? m_cols.value() - col - 1 : col);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index index)
+  {
+    return m_argImpl.coeffRef(m_rows.value() * m_cols.value() - index - 1);
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index row, Index col) const
+  {
+    enum {
+      PacketSize = unpacket_traits<PacketType>::size,
+      OffsetRow  = ReverseRow && IsColMajor ? PacketSize : 1,
+      OffsetCol  = ReverseCol && IsRowMajor ? PacketSize : 1
+    };
+    typedef internal::reverse_packet_cond<PacketType,ReversePacket> reverse_packet;
+    return reverse_packet::run(m_argImpl.template packet<LoadMode,PacketType>(
+                                  ReverseRow ? m_rows.value() - row - OffsetRow : row,
+                                  ReverseCol ? m_cols.value() - col - OffsetCol : col));
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index index) const
+  {
+    enum { PacketSize = unpacket_traits<PacketType>::size };
+    return preverse(m_argImpl.template packet<LoadMode,PacketType>(m_rows.value() * m_cols.value() - index - PacketSize));
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  void writePacket(Index row, Index col, const PacketType& x)
+  {
+    // FIXME we could factorize some code with packet(i,j)
+    enum {
+      PacketSize = unpacket_traits<PacketType>::size,
+      OffsetRow  = ReverseRow && IsColMajor ? PacketSize : 1,
+      OffsetCol  = ReverseCol && IsRowMajor ? PacketSize : 1
+    };
+    typedef internal::reverse_packet_cond<PacketType,ReversePacket> reverse_packet;
+    m_argImpl.template writePacket<LoadMode>(
+                                  ReverseRow ? m_rows.value() - row - OffsetRow : row,
+                                  ReverseCol ? m_cols.value() - col - OffsetCol : col,
+                                  reverse_packet::run(x));
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  void writePacket(Index index, const PacketType& x)
+  {
+    enum { PacketSize = unpacket_traits<PacketType>::size };
+    m_argImpl.template writePacket<LoadMode>
+      (m_rows.value() * m_cols.value() - index - PacketSize, preverse(x));
+  }
+ 
+protected:
+  evaluator<ArgType> m_argImpl;
+
+  // If we do not reverse rows, then we do not need to know the number of rows; same for columns
+  // Nonetheless, in this case it is important to set to 1 such that the coeff(index) method works fine for vectors.
+  const variable_if_dynamic<Index, ReverseRow ? ArgType::RowsAtCompileTime : 1> m_rows;
+  const variable_if_dynamic<Index, ReverseCol ? ArgType::ColsAtCompileTime : 1> m_cols;
+};
+
+
+// -------------------- Diagonal --------------------
+
+template<typename ArgType, int DiagIndex>
+struct evaluator<Diagonal<ArgType, DiagIndex> >
+  : evaluator_base<Diagonal<ArgType, DiagIndex> >
+{
+  typedef Diagonal<ArgType, DiagIndex> XprType;
+  
+  enum {
+    CoeffReadCost = evaluator<ArgType>::CoeffReadCost,
+    
+    Flags = (unsigned int)(evaluator<ArgType>::Flags & (HereditaryBits | DirectAccessBit) & ~RowMajorBit) | LinearAccessBit,
+    
+    Alignment = 0
+  };
+
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& diagonal)
+    : m_argImpl(diagonal.nestedExpression()),
+      m_index(diagonal.index())
+  { }
+ 
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index) const
+  {
+    return m_argImpl.coeff(row + rowOffset(), row + colOffset());
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  {
+    return m_argImpl.coeff(index + rowOffset(), index + colOffset());
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index row, Index)
+  {
+    return m_argImpl.coeffRef(row + rowOffset(), row + colOffset());
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index index)
+  {
+    return m_argImpl.coeffRef(index + rowOffset(), index + colOffset());
+  }
+
+protected:
+  evaluator<ArgType> m_argImpl;
+  const internal::variable_if_dynamicindex<Index, XprType::DiagIndex> m_index;
+
+private:
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index rowOffset() const { return m_index.value() > 0 ? 0 : -m_index.value(); }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index colOffset() const { return m_index.value() > 0 ? m_index.value() : 0; }
+};
+
+
+//----------------------------------------------------------------------
+// deprecated code
+//----------------------------------------------------------------------
+
+// -------------------- EvalToTemp --------------------
+
+// expression class for evaluating nested expression to a temporary
+
+template<typename ArgType> class EvalToTemp;
+
+template<typename ArgType>
+struct traits<EvalToTemp<ArgType> >
+  : public traits<ArgType>
+{ };
+
+template<typename ArgType>
+class EvalToTemp
+  : public dense_xpr_base<EvalToTemp<ArgType> >::type
+{
+ public:
+ 
+  typedef typename dense_xpr_base<EvalToTemp>::type Base;
+  EIGEN_GENERIC_PUBLIC_INTERFACE(EvalToTemp)
+ 
+  explicit EvalToTemp(const ArgType& arg)
+    : m_arg(arg)
+  { }
+ 
+  const ArgType& arg() const
+  {
+    return m_arg;
+  }
+
+  Index rows() const 
+  {
+    return m_arg.rows();
+  }
+
+  Index cols() const 
+  {
+    return m_arg.cols();
+  }
+
+ private:
+  const ArgType& m_arg;
+};
+ 
+template<typename ArgType>
+struct evaluator<EvalToTemp<ArgType> >
+  : public evaluator<typename ArgType::PlainObject>
+{
+  typedef EvalToTemp<ArgType>                   XprType;
+  typedef typename ArgType::PlainObject         PlainObject;
+  typedef evaluator<PlainObject> Base;
+  
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr)
+    : m_result(xpr.arg())
+  {
+    ::new (static_cast<Base*>(this)) Base(m_result);
+  }
+
+  // This constructor is used when nesting an EvalTo evaluator in another evaluator
+  EIGEN_DEVICE_FUNC evaluator(const ArgType& arg)
+    : m_result(arg)
+  {
+    ::new (static_cast<Base*>(this)) Base(m_result);
+  }
+
+protected:
+  PlainObject m_result;
+};
+
+} // namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_COREEVALUATORS_H
diff --git a/eigen/Eigen/src/Core/CoreIterators.h b/eigen/Eigen/src/Core/CoreIterators.h
index 6da4683..b967196 100644
--- a/eigen/Eigen/src/Core/CoreIterators.h
+++ b/eigen/Eigen/src/Core/CoreIterators.h
@@ -1,7 +1,7 @@
 // This file is part of Eigen, a lightweight C++ template library
 // for linear algebra.
 //
-// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2008-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
 //
 // This Source Code Form is subject to the terms of the Mozilla
 // Public License v. 2.0. If a copy of the MPL was not distributed
@@ -15,47 +15,118 @@ namespace Eigen {
 /* This file contains the respective InnerIterator definition of the expressions defined in Eigen/Core
  */
 
-/** \ingroup SparseCore_Module
-  * \class InnerIterator
-  * \brief An InnerIterator allows to loop over the element of a sparse (or dense) matrix or expression
-  *
-  * todo
+namespace internal {
+
+template<typename XprType, typename EvaluatorKind>
+class inner_iterator_selector;
+
+}
+
+/** \class InnerIterator
+  * \brief An InnerIterator allows to loop over the element of any matrix expression.
+  * 
+  * \warning To be used with care because an evaluator is constructed every time an InnerIterator iterator is constructed.
+  * 
+  * TODO: add a usage example
   */
+template<typename XprType>
+class InnerIterator
+{
+protected:
+  typedef internal::inner_iterator_selector<XprType, typename internal::evaluator_traits<XprType>::Kind> IteratorType;
+  typedef internal::evaluator<XprType> EvaluatorType;
+  typedef typename internal::traits<XprType>::Scalar Scalar;
+public:
+  /** Construct an iterator over the \a outerId -th row or column of \a xpr */
+  InnerIterator(const XprType &xpr, const Index &outerId)
+    : m_eval(xpr), m_iter(m_eval, outerId, xpr.innerSize())
+  {}
+  
+  /// \returns the value of the current coefficient.
+  EIGEN_STRONG_INLINE Scalar value() const          { return m_iter.value(); }
+  /** Increment the iterator \c *this to the next non-zero coefficient.
+    * Explicit zeros are not skipped over. To skip explicit zeros, see class SparseView
+    */
+  EIGEN_STRONG_INLINE InnerIterator& operator++()   { m_iter.operator++(); return *this; }
+  EIGEN_STRONG_INLINE InnerIterator& operator+=(Index i) { m_iter.operator+=(i); return *this; }
+  EIGEN_STRONG_INLINE InnerIterator operator+(Index i) 
+  { InnerIterator result(*this); result+=i; return result; }
+    
+
+  /// \returns the column or row index of the current coefficient.
+  EIGEN_STRONG_INLINE Index index() const           { return m_iter.index(); }
+  /// \returns the row index of the current coefficient.
+  EIGEN_STRONG_INLINE Index row() const             { return m_iter.row(); }
+  /// \returns the column index of the current coefficient.
+  EIGEN_STRONG_INLINE Index col() const             { return m_iter.col(); }
+  /// \returns \c true if the iterator \c *this still references a valid coefficient.
+  EIGEN_STRONG_INLINE operator bool() const         { return m_iter; }
+  
+protected:
+  EvaluatorType m_eval;
+  IteratorType m_iter;
+private:
+  // If you get here, then you're not using the right InnerIterator type, e.g.:
+  //   SparseMatrix<double,RowMajor> A;
+  //   SparseMatrix<double>::InnerIterator it(A,0);
+  template<typename T> InnerIterator(const EigenBase<T>&,Index outer);
+};
+
+namespace internal {
+
+// Generic inner iterator implementation for dense objects
+template<typename XprType>
+class inner_iterator_selector<XprType, IndexBased>
+{
+protected:
+  typedef evaluator<XprType> EvaluatorType;
+  typedef typename traits<XprType>::Scalar Scalar;
+  enum { IsRowMajor = (XprType::Flags&RowMajorBit)==RowMajorBit };
+  
+public:
+  EIGEN_STRONG_INLINE inner_iterator_selector(const EvaluatorType &eval, const Index &outerId, const Index &innerSize)
+    : m_eval(eval), m_inner(0), m_outer(outerId), m_end(innerSize)
+  {}
+
+  EIGEN_STRONG_INLINE Scalar value() const
+  {
+    return (IsRowMajor) ? m_eval.coeff(m_outer, m_inner)
+                        : m_eval.coeff(m_inner, m_outer);
+  }
+
+  EIGEN_STRONG_INLINE inner_iterator_selector& operator++() { m_inner++; return *this; }
+
+  EIGEN_STRONG_INLINE Index index() const { return m_inner; }
+  inline Index row() const { return IsRowMajor ? m_outer : index(); }
+  inline Index col() const { return IsRowMajor ? index() : m_outer; }
 
-// generic version for dense matrix and expressions
-template<typename Derived> class DenseBase<Derived>::InnerIterator
+  EIGEN_STRONG_INLINE operator bool() const { return m_inner < m_end && m_inner>=0; }
+
+protected:
+  const EvaluatorType& m_eval;
+  Index m_inner;
+  const Index m_outer;
+  const Index m_end;
+};
+
+// For iterator-based evaluator, inner-iterator is already implemented as
+// evaluator<>::InnerIterator
+template<typename XprType>
+class inner_iterator_selector<XprType, IteratorBased>
+ : public evaluator<XprType>::InnerIterator
 {
-  protected:
-    typedef typename Derived::Scalar Scalar;
-    typedef typename Derived::Index Index;
-
-    enum { IsRowMajor = (Derived::Flags&RowMajorBit)==RowMajorBit };
-  public:
-    EIGEN_STRONG_INLINE InnerIterator(const Derived& expr, Index outer)
-      : m_expression(expr), m_inner(0), m_outer(outer), m_end(expr.innerSize())
-    {}
-
-    EIGEN_STRONG_INLINE Scalar value() const
-    {
-      return (IsRowMajor) ? m_expression.coeff(m_outer, m_inner)
-                          : m_expression.coeff(m_inner, m_outer);
-    }
-
-    EIGEN_STRONG_INLINE InnerIterator& operator++() { m_inner++; return *this; }
-
-    EIGEN_STRONG_INLINE Index index() const { return m_inner; }
-    inline Index row() const { return IsRowMajor ? m_outer : index(); }
-    inline Index col() const { return IsRowMajor ? index() : m_outer; }
-
-    EIGEN_STRONG_INLINE operator bool() const { return m_inner < m_end && m_inner>=0; }
-
-  protected:
-    const Derived& m_expression;
-    Index m_inner;
-    const Index m_outer;
-    const Index m_end;
+protected:
+  typedef typename evaluator<XprType>::InnerIterator Base;
+  typedef evaluator<XprType> EvaluatorType;
+  
+public:
+  EIGEN_STRONG_INLINE inner_iterator_selector(const EvaluatorType &eval, const Index &outerId, const Index &/*innerSize*/)
+    : Base(eval, outerId)
+  {}  
 };
 
+} // end namespace internal
+
 } // end namespace Eigen
 
 #endif // EIGEN_COREITERATORS_H
diff --git a/eigen/Eigen/src/Core/CwiseBinaryOp.h b/eigen/Eigen/src/Core/CwiseBinaryOp.h
index 519a866..bf2632d 100644
--- a/eigen/Eigen/src/Core/CwiseBinaryOp.h
+++ b/eigen/Eigen/src/Core/CwiseBinaryOp.h
@@ -1,7 +1,7 @@
 // This file is part of Eigen, a lightweight C++ template library
 // for linear algebra.
 //
-// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2008-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
 // Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
 //
 // This Source Code Form is subject to the terms of the Mozilla
@@ -13,26 +13,6 @@
 
 namespace Eigen {
 
-/** \class CwiseBinaryOp
-  * \ingroup Core_Module
-  *
-  * \brief Generic expression where a coefficient-wise binary operator is applied to two expressions
-  *
-  * \param BinaryOp template functor implementing the operator
-  * \param Lhs the type of the left-hand side
-  * \param Rhs the type of the right-hand side
-  *
-  * This class represents an expression  where a coefficient-wise binary operator is applied to two expressions.
-  * It is the return type of binary operators, by which we mean only those binary operators where
-  * both the left-hand side and the right-hand side are Eigen expressions.
-  * For example, the return type of matrix1+matrix2 is a CwiseBinaryOp.
-  *
-  * Most of the time, this is the only way that it is used, so you typically don't have to name
-  * CwiseBinaryOp types explicitly.
-  *
-  * \sa MatrixBase::binaryExpr(const MatrixBase<OtherDerived> &,const CustomBinaryOp &) const, class CwiseUnaryOp, class CwiseNullaryOp
-  */
-
 namespace internal {
 template<typename BinaryOp, typename Lhs, typename Rhs>
 struct traits<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >
@@ -52,77 +32,75 @@ struct traits<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >
   // we still want to handle the case when the result type is different.
   typedef typename result_of<
                      BinaryOp(
-                       typename Lhs::Scalar,
-                       typename Rhs::Scalar
+                       const typename Lhs::Scalar&,
+                       const typename Rhs::Scalar&
                      )
                    >::type Scalar;
-  typedef typename promote_storage_type<typename traits<Lhs>::StorageKind,
-                                           typename traits<Rhs>::StorageKind>::ret StorageKind;
-  typedef typename promote_index_type<typename traits<Lhs>::Index,
-                                         typename traits<Rhs>::Index>::type Index;
+  typedef typename cwise_promote_storage_type<typename traits<Lhs>::StorageKind,
+                                              typename traits<Rhs>::StorageKind,
+                                              BinaryOp>::ret StorageKind;
+  typedef typename promote_index_type<typename traits<Lhs>::StorageIndex,
+                                      typename traits<Rhs>::StorageIndex>::type StorageIndex;
   typedef typename Lhs::Nested LhsNested;
   typedef typename Rhs::Nested RhsNested;
   typedef typename remove_reference<LhsNested>::type _LhsNested;
   typedef typename remove_reference<RhsNested>::type _RhsNested;
   enum {
-    LhsCoeffReadCost = _LhsNested::CoeffReadCost,
-    RhsCoeffReadCost = _RhsNested::CoeffReadCost,
-    LhsFlags = _LhsNested::Flags,
-    RhsFlags = _RhsNested::Flags,
-    SameType = is_same<typename _LhsNested::Scalar,typename _RhsNested::Scalar>::value,
-    StorageOrdersAgree = (int(Lhs::Flags)&RowMajorBit)==(int(Rhs::Flags)&RowMajorBit),
-    Flags0 = (int(LhsFlags) | int(RhsFlags)) & (
-        HereditaryBits
-      | (int(LhsFlags) & int(RhsFlags) &
-           ( AlignedBit
-           | (StorageOrdersAgree ? LinearAccessBit : 0)
-           | (functor_traits<BinaryOp>::PacketAccess && StorageOrdersAgree && SameType ? PacketAccessBit : 0)
-           )
-        )
-     ),
-    Flags = (Flags0 & ~RowMajorBit) | (LhsFlags & RowMajorBit),
-    Cost0 = EIGEN_ADD_COST(LhsCoeffReadCost,RhsCoeffReadCost),
-    CoeffReadCost = EIGEN_ADD_COST(Cost0,functor_traits<BinaryOp>::Cost)
+    Flags = cwise_promote_storage_order<typename traits<Lhs>::StorageKind,typename traits<Rhs>::StorageKind,_LhsNested::Flags & RowMajorBit,_RhsNested::Flags & RowMajorBit>::value
   };
 };
 } // end namespace internal
 
-// we require Lhs and Rhs to have the same scalar type. Currently there is no example of a binary functor
-// that would take two operands of different types. If there were such an example, then this check should be
-// moved to the BinaryOp functors, on a per-case basis. This would however require a change in the BinaryOp functors, as
-// currently they take only one typename Scalar template parameter.
-// It is tempting to always allow mixing different types but remember that this is often impossible in the vectorized paths.
-// So allowing mixing different types gives very unexpected errors when enabling vectorization, when the user tries to
-// add together a float matrix and a double matrix.
-#define EIGEN_CHECK_BINARY_COMPATIBILIY(BINOP,LHS,RHS) \
-  EIGEN_STATIC_ASSERT((internal::functor_is_product_like<BINOP>::ret \
-                        ? int(internal::scalar_product_traits<LHS, RHS>::Defined) \
-                        : int(internal::is_same<LHS, RHS>::value)), \
-    YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
-
 template<typename BinaryOp, typename Lhs, typename Rhs, typename StorageKind>
 class CwiseBinaryOpImpl;
 
-template<typename BinaryOp, typename Lhs, typename Rhs>
-class CwiseBinaryOp : internal::no_assignment_operator,
+/** \class CwiseBinaryOp
+  * \ingroup Core_Module
+  *
+  * \brief Generic expression where a coefficient-wise binary operator is applied to two expressions
+  *
+  * \tparam BinaryOp template functor implementing the operator
+  * \tparam LhsType the type of the left-hand side
+  * \tparam RhsType the type of the right-hand side
+  *
+  * This class represents an expression  where a coefficient-wise binary operator is applied to two expressions.
+  * It is the return type of binary operators, by which we mean only those binary operators where
+  * both the left-hand side and the right-hand side are Eigen expressions.
+  * For example, the return type of matrix1+matrix2 is a CwiseBinaryOp.
+  *
+  * Most of the time, this is the only way that it is used, so you typically don't have to name
+  * CwiseBinaryOp types explicitly.
+  *
+  * \sa MatrixBase::binaryExpr(const MatrixBase<OtherDerived> &,const CustomBinaryOp &) const, class CwiseUnaryOp, class CwiseNullaryOp
+  */
+template<typename BinaryOp, typename LhsType, typename RhsType>
+class CwiseBinaryOp : 
   public CwiseBinaryOpImpl<
-          BinaryOp, Lhs, Rhs,
-          typename internal::promote_storage_type<typename internal::traits<Lhs>::StorageKind,
-                                           typename internal::traits<Rhs>::StorageKind>::ret>
+          BinaryOp, LhsType, RhsType,
+          typename internal::cwise_promote_storage_type<typename internal::traits<LhsType>::StorageKind,
+                                                        typename internal::traits<RhsType>::StorageKind,
+                                                        BinaryOp>::ret>,
+  internal::no_assignment_operator
 {
   public:
+    
+    typedef typename internal::remove_all<BinaryOp>::type Functor;
+    typedef typename internal::remove_all<LhsType>::type Lhs;
+    typedef typename internal::remove_all<RhsType>::type Rhs;
 
     typedef typename CwiseBinaryOpImpl<
-        BinaryOp, Lhs, Rhs,
-        typename internal::promote_storage_type<typename internal::traits<Lhs>::StorageKind,
-                                         typename internal::traits<Rhs>::StorageKind>::ret>::Base Base;
+        BinaryOp, LhsType, RhsType,
+        typename internal::cwise_promote_storage_type<typename internal::traits<LhsType>::StorageKind,
+                                                      typename internal::traits<Rhs>::StorageKind,
+                                                      BinaryOp>::ret>::Base Base;
     EIGEN_GENERIC_PUBLIC_INTERFACE(CwiseBinaryOp)
 
-    typedef typename internal::nested<Lhs>::type LhsNested;
-    typedef typename internal::nested<Rhs>::type RhsNested;
+    typedef typename internal::ref_selector<LhsType>::type LhsNested;
+    typedef typename internal::ref_selector<RhsType>::type RhsNested;
     typedef typename internal::remove_reference<LhsNested>::type _LhsNested;
     typedef typename internal::remove_reference<RhsNested>::type _RhsNested;
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE CwiseBinaryOp(const Lhs& aLhs, const Rhs& aRhs, const BinaryOp& func = BinaryOp())
       : m_lhs(aLhs), m_rhs(aRhs), m_functor(func)
     {
@@ -132,6 +110,7 @@ class CwiseBinaryOp : internal::no_assignment_operator,
       eigen_assert(aLhs.rows() == aRhs.rows() && aLhs.cols() == aRhs.cols());
     }
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Index rows() const {
       // return the fixed size type if available to enable compile time optimizations
       if (internal::traits<typename internal::remove_all<LhsNested>::type>::RowsAtCompileTime==Dynamic)
@@ -139,6 +118,7 @@ class CwiseBinaryOp : internal::no_assignment_operator,
       else
         return m_lhs.rows();
     }
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Index cols() const {
       // return the fixed size type if available to enable compile time optimizations
       if (internal::traits<typename internal::remove_all<LhsNested>::type>::ColsAtCompileTime==Dynamic)
@@ -148,10 +128,13 @@ class CwiseBinaryOp : internal::no_assignment_operator,
     }
 
     /** \returns the left hand side nested expression */
+    EIGEN_DEVICE_FUNC
     const _LhsNested& lhs() const { return m_lhs; }
     /** \returns the right hand side nested expression */
+    EIGEN_DEVICE_FUNC
     const _RhsNested& rhs() const { return m_rhs; }
     /** \returns the functor representing the binary operation */
+    EIGEN_DEVICE_FUNC
     const BinaryOp& functor() const { return m_functor; }
 
   protected:
@@ -160,41 +143,13 @@ class CwiseBinaryOp : internal::no_assignment_operator,
     const BinaryOp m_functor;
 };
 
-template<typename BinaryOp, typename Lhs, typename Rhs>
-class CwiseBinaryOpImpl<BinaryOp, Lhs, Rhs, Dense>
-  : public internal::dense_xpr_base<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >::type
+// Generic API dispatcher
+template<typename BinaryOp, typename Lhs, typename Rhs, typename StorageKind>
+class CwiseBinaryOpImpl
+  : public internal::generic_xpr_base<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >::type
 {
-    typedef CwiseBinaryOp<BinaryOp, Lhs, Rhs> Derived;
-  public:
-
-    typedef typename internal::dense_xpr_base<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >::type Base;
-    EIGEN_DENSE_PUBLIC_INTERFACE( Derived )
-
-    EIGEN_STRONG_INLINE const Scalar coeff(Index rowId, Index colId) const
-    {
-      return derived().functor()(derived().lhs().coeff(rowId, colId),
-                                 derived().rhs().coeff(rowId, colId));
-    }
-
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet(Index rowId, Index colId) const
-    {
-      return derived().functor().packetOp(derived().lhs().template packet<LoadMode>(rowId, colId),
-                                          derived().rhs().template packet<LoadMode>(rowId, colId));
-    }
-
-    EIGEN_STRONG_INLINE const Scalar coeff(Index index) const
-    {
-      return derived().functor()(derived().lhs().coeff(index),
-                                 derived().rhs().coeff(index));
-    }
-
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet(Index index) const
-    {
-      return derived().functor().packetOp(derived().lhs().template packet<LoadMode>(index),
-                                          derived().rhs().template packet<LoadMode>(index));
-    }
+public:
+  typedef typename internal::generic_xpr_base<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >::type Base;
 };
 
 /** replaces \c *this by \c *this - \a other.
@@ -203,11 +158,10 @@ class CwiseBinaryOpImpl<BinaryOp, Lhs, Rhs, Dense>
   */
 template<typename Derived>
 template<typename OtherDerived>
-EIGEN_STRONG_INLINE Derived &
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived &
 MatrixBase<Derived>::operator-=(const MatrixBase<OtherDerived> &other)
 {
-  SelfCwiseBinaryOp<internal::scalar_difference_op<Scalar>, Derived, OtherDerived> tmp(derived());
-  tmp = other.derived();
+  call_assignment(derived(), other.derived(), internal::sub_assign_op<Scalar,typename OtherDerived::Scalar>());
   return derived();
 }
 
@@ -217,11 +171,10 @@ MatrixBase<Derived>::operator-=(const MatrixBase<OtherDerived> &other)
   */
 template<typename Derived>
 template<typename OtherDerived>
-EIGEN_STRONG_INLINE Derived &
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived &
 MatrixBase<Derived>::operator+=(const MatrixBase<OtherDerived>& other)
 {
-  SelfCwiseBinaryOp<internal::scalar_sum_op<Scalar>, Derived, OtherDerived> tmp(derived());
-  tmp = other.derived();
+  call_assignment(derived(), other.derived(), internal::add_assign_op<Scalar,typename OtherDerived::Scalar>());
   return derived();
 }
 
diff --git a/eigen/Eigen/src/Core/CwiseNullaryOp.h b/eigen/Eigen/src/Core/CwiseNullaryOp.h
index a93bab2..144608e 100644
--- a/eigen/Eigen/src/Core/CwiseNullaryOp.h
+++ b/eigen/Eigen/src/Core/CwiseNullaryOp.h
@@ -12,13 +12,24 @@
 
 namespace Eigen {
 
+namespace internal {
+template<typename NullaryOp, typename PlainObjectType>
+struct traits<CwiseNullaryOp<NullaryOp, PlainObjectType> > : traits<PlainObjectType>
+{
+  enum {
+    Flags = traits<PlainObjectType>::Flags & RowMajorBit
+  };
+};
+
+} // namespace internal
+
 /** \class CwiseNullaryOp
   * \ingroup Core_Module
   *
   * \brief Generic expression of a matrix where all coefficients are defined by a functor
   *
-  * \param NullaryOp template functor implementing the operator
-  * \param PlainObjectType the underlying plain matrix/array type
+  * \tparam NullaryOp template functor implementing the operator
+  * \tparam PlainObjectType the underlying plain matrix/array type
   *
   * This class represents an expression of a generic nullary operator.
   * It is the return type of the Ones(), Zero(), Constant(), Identity() and Random() methods,
@@ -27,68 +38,49 @@ namespace Eigen {
   * However, if you want to write a function returning such an expression, you
   * will need to use this class.
   *
-  * \sa class CwiseUnaryOp, class CwiseBinaryOp, DenseBase::NullaryExpr()
+  * The functor NullaryOp must expose one of the following method:
+    <table class="manual">
+    <tr            ><td>\c operator()() </td><td>if the procedural generation does not depend on the coefficient entries (e.g., random numbers)</td></tr>
+    <tr class="alt"><td>\c operator()(Index i)</td><td>if the procedural generation makes sense for vectors only and that it depends on the coefficient index \c i (e.g., linspace) </td></tr>
+    <tr            ><td>\c operator()(Index i,Index j)</td><td>if the procedural generation depends on the matrix coordinates \c i, \c j (e.g., to generate a checkerboard with 0 and 1)</td></tr>
+    </table>
+  * It is also possible to expose the last two operators if the generation makes sense for matrices but can be optimized for vectors.
+  *
+  * See DenseBase::NullaryExpr(Index,const CustomNullaryOp&) for an example binding
+  * C++11 random number generators.
+  *
+  * A nullary expression can also be used to implement custom sophisticated matrix manipulations
+  * that cannot be covered by the existing set of natively supported matrix manipulations.
+  * See this \ref TopicCustomizing_NullaryExpr "page" for some examples and additional explanations
+  * on the behavior of CwiseNullaryOp.
+  *
+  * \sa class CwiseUnaryOp, class CwiseBinaryOp, DenseBase::NullaryExpr
   */
-
-namespace internal {
 template<typename NullaryOp, typename PlainObjectType>
-struct traits<CwiseNullaryOp<NullaryOp, PlainObjectType> > : traits<PlainObjectType>
-{
-  enum {
-    Flags = (traits<PlainObjectType>::Flags
-      & (  HereditaryBits
-         | (functor_has_linear_access<NullaryOp>::ret ? LinearAccessBit : 0)
-         | (functor_traits<NullaryOp>::PacketAccess ? PacketAccessBit : 0)))
-      | (functor_traits<NullaryOp>::IsRepeatable ? 0 : EvalBeforeNestingBit),
-    CoeffReadCost = functor_traits<NullaryOp>::Cost
-  };
-};
-}
-
-template<typename NullaryOp, typename PlainObjectType>
-class CwiseNullaryOp : internal::no_assignment_operator,
-  public internal::dense_xpr_base< CwiseNullaryOp<NullaryOp, PlainObjectType> >::type
+class CwiseNullaryOp : public internal::dense_xpr_base< CwiseNullaryOp<NullaryOp, PlainObjectType> >::type, internal::no_assignment_operator
 {
   public:
 
     typedef typename internal::dense_xpr_base<CwiseNullaryOp>::type Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(CwiseNullaryOp)
 
-    CwiseNullaryOp(Index nbRows, Index nbCols, const NullaryOp& func = NullaryOp())
-      : m_rows(nbRows), m_cols(nbCols), m_functor(func)
+    EIGEN_DEVICE_FUNC
+    CwiseNullaryOp(Index rows, Index cols, const NullaryOp& func = NullaryOp())
+      : m_rows(rows), m_cols(cols), m_functor(func)
     {
-      eigen_assert(nbRows >= 0
-            && (RowsAtCompileTime == Dynamic || RowsAtCompileTime == nbRows)
-            &&  nbCols >= 0
-            && (ColsAtCompileTime == Dynamic || ColsAtCompileTime == nbCols));
+      eigen_assert(rows >= 0
+            && (RowsAtCompileTime == Dynamic || RowsAtCompileTime == rows)
+            &&  cols >= 0
+            && (ColsAtCompileTime == Dynamic || ColsAtCompileTime == cols));
     }
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Index rows() const { return m_rows.value(); }
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Index cols() const { return m_cols.value(); }
 
-    EIGEN_STRONG_INLINE const Scalar coeff(Index rowId, Index colId) const
-    {
-      return m_functor(rowId, colId);
-    }
-
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet(Index rowId, Index colId) const
-    {
-      return m_functor.packetOp(rowId, colId);
-    }
-
-    EIGEN_STRONG_INLINE const Scalar coeff(Index index) const
-    {
-      return m_functor(index);
-    }
-
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet(Index index) const
-    {
-      return m_functor.packetOp(index);
-    }
-
     /** \returns the functor representing the nullary operation */
+    EIGEN_DEVICE_FUNC
     const NullaryOp& functor() const { return m_functor; }
 
   protected:
@@ -113,10 +105,10 @@ class CwiseNullaryOp : internal::no_assignment_operator,
   */
 template<typename Derived>
 template<typename CustomNullaryOp>
-EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, typename DenseBase<Derived>::PlainObject>
 DenseBase<Derived>::NullaryExpr(Index rows, Index cols, const CustomNullaryOp& func)
 {
-  return CwiseNullaryOp<CustomNullaryOp, Derived>(rows, cols, func);
+  return CwiseNullaryOp<CustomNullaryOp, PlainObject>(rows, cols, func);
 }
 
 /** \returns an expression of a matrix defined by a custom functor \a func
@@ -132,16 +124,19 @@ DenseBase<Derived>::NullaryExpr(Index rows, Index cols, const CustomNullaryOp& f
   *
   * The template parameter \a CustomNullaryOp is the type of the functor.
   *
+  * Here is an example with C++11 random generators: \include random_cpp11.cpp
+  * Output: \verbinclude random_cpp11.out
+  * 
   * \sa class CwiseNullaryOp
   */
 template<typename Derived>
 template<typename CustomNullaryOp>
-EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, typename DenseBase<Derived>::PlainObject>
 DenseBase<Derived>::NullaryExpr(Index size, const CustomNullaryOp& func)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
-  if(RowsAtCompileTime == 1) return CwiseNullaryOp<CustomNullaryOp, Derived>(1, size, func);
-  else return CwiseNullaryOp<CustomNullaryOp, Derived>(size, 1, func);
+  if(RowsAtCompileTime == 1) return CwiseNullaryOp<CustomNullaryOp, PlainObject>(1, size, func);
+  else return CwiseNullaryOp<CustomNullaryOp, PlainObject>(size, 1, func);
 }
 
 /** \returns an expression of a matrix defined by a custom functor \a func
@@ -155,19 +150,19 @@ DenseBase<Derived>::NullaryExpr(Index size, const CustomNullaryOp& func)
   */
 template<typename Derived>
 template<typename CustomNullaryOp>
-EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, typename DenseBase<Derived>::PlainObject>
 DenseBase<Derived>::NullaryExpr(const CustomNullaryOp& func)
 {
-  return CwiseNullaryOp<CustomNullaryOp, Derived>(RowsAtCompileTime, ColsAtCompileTime, func);
+  return CwiseNullaryOp<CustomNullaryOp, PlainObject>(RowsAtCompileTime, ColsAtCompileTime, func);
 }
 
 /** \returns an expression of a constant matrix of value \a value
   *
-  * The parameters \a nbRows and \a nbCols are the number of rows and of columns of
+  * The parameters \a rows and \a cols are the number of rows and of columns of
   * the returned matrix. Must be compatible with this DenseBase type.
   *
   * This variant is meant to be used for dynamic-size matrix types. For fixed-size types,
-  * it is redundant to pass \a nbRows and \a nbCols as arguments, so Zero() should be used
+  * it is redundant to pass \a rows and \a cols as arguments, so Zero() should be used
   * instead.
   *
   * The template parameter \a CustomNullaryOp is the type of the functor.
@@ -175,10 +170,10 @@ DenseBase<Derived>::NullaryExpr(const CustomNullaryOp& func)
   * \sa class CwiseNullaryOp
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
-DenseBase<Derived>::Constant(Index nbRows, Index nbCols, const Scalar& value)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+DenseBase<Derived>::Constant(Index rows, Index cols, const Scalar& value)
 {
-  return DenseBase<Derived>::NullaryExpr(nbRows, nbCols, internal::scalar_constant_op<Scalar>(value));
+  return DenseBase<Derived>::NullaryExpr(rows, cols, internal::scalar_constant_op<Scalar>(value));
 }
 
 /** \returns an expression of a constant matrix of value \a value
@@ -197,7 +192,7 @@ DenseBase<Derived>::Constant(Index nbRows, Index nbCols, const Scalar& value)
   * \sa class CwiseNullaryOp
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
 DenseBase<Derived>::Constant(Index size, const Scalar& value)
 {
   return DenseBase<Derived>::NullaryExpr(size, internal::scalar_constant_op<Scalar>(value));
@@ -213,53 +208,40 @@ DenseBase<Derived>::Constant(Index size, const Scalar& value)
   * \sa class CwiseNullaryOp
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
 DenseBase<Derived>::Constant(const Scalar& value)
 {
   EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived)
   return DenseBase<Derived>::NullaryExpr(RowsAtCompileTime, ColsAtCompileTime, internal::scalar_constant_op<Scalar>(value));
 }
 
-/**
-  * \brief Sets a linearly space vector.
-  *
-  * The function generates 'size' equally spaced values in the closed interval [low,high].
-  * This particular version of LinSpaced() uses sequential access, i.e. vector access is
-  * assumed to be a(0), a(1), ..., a(size). This assumption allows for better vectorization
-  * and yields faster code than the random access version.
-  *
-  * When size is set to 1, a vector of length 1 containing 'high' is returned.
-  *
-  * \only_for_vectors
-  *
-  * Example: \include DenseBase_LinSpaced_seq.cpp
-  * Output: \verbinclude DenseBase_LinSpaced_seq.out
+/** \deprecated because of accuracy loss. In Eigen 3.3, it is an alias for LinSpaced(Index,const Scalar&,const Scalar&)
   *
-  * \sa setLinSpaced(Index,const Scalar&,const Scalar&), LinSpaced(Index,Scalar,Scalar), CwiseNullaryOp
+  * \sa LinSpaced(Index,Scalar,Scalar), setLinSpaced(Index,const Scalar&,const Scalar&)
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::SequentialLinSpacedReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedReturnType
 DenseBase<Derived>::LinSpaced(Sequential_t, Index size, const Scalar& low, const Scalar& high)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
-  return DenseBase<Derived>::NullaryExpr(size, internal::linspaced_op<Scalar,false>(low,high,size));
+  return DenseBase<Derived>::NullaryExpr(size, internal::linspaced_op<Scalar,PacketScalar>(low,high,size));
 }
 
-/**
-  * \copydoc DenseBase::LinSpaced(Sequential_t, Index, const Scalar&, const Scalar&)
-  * Special version for fixed size types which does not require the size parameter.
+/** \deprecated because of accuracy loss. In Eigen 3.3, it is an alias for LinSpaced(const Scalar&,const Scalar&)
+  *
+  * \sa LinSpaced(Scalar,Scalar)
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::SequentialLinSpacedReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedReturnType
 DenseBase<Derived>::LinSpaced(Sequential_t, const Scalar& low, const Scalar& high)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
   EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived)
-  return DenseBase<Derived>::NullaryExpr(Derived::SizeAtCompileTime, internal::linspaced_op<Scalar,false>(low,high,Derived::SizeAtCompileTime));
+  return DenseBase<Derived>::NullaryExpr(Derived::SizeAtCompileTime, internal::linspaced_op<Scalar,PacketScalar>(low,high,Derived::SizeAtCompileTime));
 }
 
 /**
-  * \brief Sets a linearly space vector.
+  * \brief Sets a linearly spaced vector.
   *
   * The function generates 'size' equally spaced values in the closed interval [low,high].
   * When size is set to 1, a vector of length 1 containing 'high' is returned.
@@ -269,14 +251,24 @@ DenseBase<Derived>::LinSpaced(Sequential_t, const Scalar& low, const Scalar& hig
   * Example: \include DenseBase_LinSpaced.cpp
   * Output: \verbinclude DenseBase_LinSpaced.out
   *
-  * \sa setLinSpaced(Index,const Scalar&,const Scalar&), LinSpaced(Sequential_t,Index,const Scalar&,const Scalar&,Index), CwiseNullaryOp
+  * For integer scalar types, an even spacing is possible if and only if the length of the range,
+  * i.e., \c high-low is a scalar multiple of \c size-1, or if \c size is a scalar multiple of the
+  * number of values \c high-low+1 (meaning each value can be repeated the same number of time).
+  * If one of these two considions is not satisfied, then \c high is lowered to the largest value
+  * satisfying one of this constraint.
+  * Here are some examples:
+  *
+  * Example: \include DenseBase_LinSpacedInt.cpp
+  * Output: \verbinclude DenseBase_LinSpacedInt.out
+  *
+  * \sa setLinSpaced(Index,const Scalar&,const Scalar&), CwiseNullaryOp
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedReturnType
 DenseBase<Derived>::LinSpaced(Index size, const Scalar& low, const Scalar& high)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
-  return DenseBase<Derived>::NullaryExpr(size, internal::linspaced_op<Scalar,true>(low,high,size));
+  return DenseBase<Derived>::NullaryExpr(size, internal::linspaced_op<Scalar,PacketScalar>(low,high,size));
 }
 
 /**
@@ -284,22 +276,23 @@ DenseBase<Derived>::LinSpaced(Index size, const Scalar& low, const Scalar& high)
   * Special version for fixed size types which does not require the size parameter.
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedReturnType
 DenseBase<Derived>::LinSpaced(const Scalar& low, const Scalar& high)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
   EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived)
-  return DenseBase<Derived>::NullaryExpr(Derived::SizeAtCompileTime, internal::linspaced_op<Scalar,true>(low,high,Derived::SizeAtCompileTime));
+  return DenseBase<Derived>::NullaryExpr(Derived::SizeAtCompileTime, internal::linspaced_op<Scalar,PacketScalar>(low,high,Derived::SizeAtCompileTime));
 }
 
 /** \returns true if all coefficients in this matrix are approximately equal to \a val, to within precision \a prec */
 template<typename Derived>
-bool DenseBase<Derived>::isApproxToConstant
+EIGEN_DEVICE_FUNC bool DenseBase<Derived>::isApproxToConstant
 (const Scalar& val, const RealScalar& prec) const
 {
+  typename internal::nested_eval<Derived,1>::type self(derived());
   for(Index j = 0; j < cols(); ++j)
     for(Index i = 0; i < rows(); ++i)
-      if(!internal::isApprox(this->coeff(i, j), val, prec))
+      if(!internal::isApprox(self.coeff(i, j), val, prec))
         return false;
   return true;
 }
@@ -308,7 +301,7 @@ bool DenseBase<Derived>::isApproxToConstant
   *
   * \returns true if all coefficients in this matrix are approximately equal to \a value, to within precision \a prec */
 template<typename Derived>
-bool DenseBase<Derived>::isConstant
+EIGEN_DEVICE_FUNC bool DenseBase<Derived>::isConstant
 (const Scalar& val, const RealScalar& prec) const
 {
   return isApproxToConstant(val, prec);
@@ -319,22 +312,22 @@ bool DenseBase<Derived>::isConstant
   * \sa setConstant(), Constant(), class CwiseNullaryOp
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE void DenseBase<Derived>::fill(const Scalar& val)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void DenseBase<Derived>::fill(const Scalar& val)
 {
   setConstant(val);
 }
 
-/** Sets all coefficients in this expression to \a value.
+/** Sets all coefficients in this expression to value \a val.
   *
   * \sa fill(), setConstant(Index,const Scalar&), setConstant(Index,Index,const Scalar&), setZero(), setOnes(), Constant(), class CwiseNullaryOp, setZero(), setOnes()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setConstant(const Scalar& val)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setConstant(const Scalar& val)
 {
   return derived() = Constant(rows(), cols(), val);
 }
 
-/** Resizes to the given \a size, and sets all coefficients in this expression to the given \a value.
+/** Resizes to the given \a size, and sets all coefficients in this expression to the given value \a val.
   *
   * \only_for_vectors
   *
@@ -344,17 +337,17 @@ EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setConstant(const Scalar& val)
   * \sa MatrixBase::setConstant(const Scalar&), setConstant(Index,Index,const Scalar&), class CwiseNullaryOp, MatrixBase::Constant(const Scalar&)
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived&
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived&
 PlainObjectBase<Derived>::setConstant(Index size, const Scalar& val)
 {
   resize(size);
   return setConstant(val);
 }
 
-/** Resizes to the given size, and sets all coefficients in this expression to the given \a value.
+/** Resizes to the given size, and sets all coefficients in this expression to the given value \a val.
   *
-  * \param nbRows the new number of rows
-  * \param nbCols the new number of columns
+  * \param rows the new number of rows
+  * \param cols the new number of columns
   * \param val the value to which all coefficients are set
   *
   * Example: \include Matrix_setConstant_int_int.cpp
@@ -363,15 +356,15 @@ PlainObjectBase<Derived>::setConstant(Index size, const Scalar& val)
   * \sa MatrixBase::setConstant(const Scalar&), setConstant(Index,const Scalar&), class CwiseNullaryOp, MatrixBase::Constant(const Scalar&)
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived&
-PlainObjectBase<Derived>::setConstant(Index nbRows, Index nbCols, const Scalar& val)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived&
+PlainObjectBase<Derived>::setConstant(Index rows, Index cols, const Scalar& val)
 {
-  resize(nbRows, nbCols);
+  resize(rows, cols);
   return setConstant(val);
 }
 
 /**
-  * \brief Sets a linearly space vector.
+  * \brief Sets a linearly spaced vector.
   *
   * The function generates 'size' equally spaced values in the closed interval [low,high].
   * When size is set to 1, a vector of length 1 containing 'high' is returned.
@@ -381,27 +374,33 @@ PlainObjectBase<Derived>::setConstant(Index nbRows, Index nbCols, const Scalar&
   * Example: \include DenseBase_setLinSpaced.cpp
   * Output: \verbinclude DenseBase_setLinSpaced.out
   *
-  * \sa CwiseNullaryOp
+  * For integer scalar types, do not miss the explanations on the definition
+  * of \link LinSpaced(Index,const Scalar&,const Scalar&) even spacing \endlink.
+  *
+  * \sa LinSpaced(Index,const Scalar&,const Scalar&), CwiseNullaryOp
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setLinSpaced(Index newSize, const Scalar& low, const Scalar& high)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setLinSpaced(Index newSize, const Scalar& low, const Scalar& high)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
-  return derived() = Derived::NullaryExpr(newSize, internal::linspaced_op<Scalar,false>(low,high,newSize));
+  return derived() = Derived::NullaryExpr(newSize, internal::linspaced_op<Scalar,PacketScalar>(low,high,newSize));
 }
 
 /**
-  * \brief Sets a linearly space vector.
+  * \brief Sets a linearly spaced vector.
   *
-  * The function fill *this with equally spaced values in the closed interval [low,high].
+  * The function fills \c *this with equally spaced values in the closed interval [low,high].
   * When size is set to 1, a vector of length 1 containing 'high' is returned.
   *
   * \only_for_vectors
   *
-  * \sa setLinSpaced(Index, const Scalar&, const Scalar&), CwiseNullaryOp
+  * For integer scalar types, do not miss the explanations on the definition
+  * of \link LinSpaced(Index,const Scalar&,const Scalar&) even spacing \endlink.
+  *
+  * \sa LinSpaced(Index,const Scalar&,const Scalar&), setLinSpaced(Index, const Scalar&, const Scalar&), CwiseNullaryOp
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setLinSpaced(const Scalar& low, const Scalar& high)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setLinSpaced(const Scalar& low, const Scalar& high)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
   return setLinSpaced(size(), low, high);
@@ -424,10 +423,10 @@ EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setLinSpaced(const Scalar& low,
   * \sa Zero(), Zero(Index)
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
-DenseBase<Derived>::Zero(Index nbRows, Index nbCols)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+DenseBase<Derived>::Zero(Index rows, Index cols)
 {
-  return Constant(nbRows, nbCols, Scalar(0));
+  return Constant(rows, cols, Scalar(0));
 }
 
 /** \returns an expression of a zero vector.
@@ -447,7 +446,7 @@ DenseBase<Derived>::Zero(Index nbRows, Index nbCols)
   * \sa Zero(), Zero(Index,Index)
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
 DenseBase<Derived>::Zero(Index size)
 {
   return Constant(size, Scalar(0));
@@ -464,7 +463,7 @@ DenseBase<Derived>::Zero(Index size)
   * \sa Zero(Index), Zero(Index,Index)
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
 DenseBase<Derived>::Zero()
 {
   return Constant(Scalar(0));
@@ -479,11 +478,12 @@ DenseBase<Derived>::Zero()
   * \sa class CwiseNullaryOp, Zero()
   */
 template<typename Derived>
-bool DenseBase<Derived>::isZero(const RealScalar& prec) const
+EIGEN_DEVICE_FUNC bool DenseBase<Derived>::isZero(const RealScalar& prec) const
 {
+  typename internal::nested_eval<Derived,1>::type self(derived());
   for(Index j = 0; j < cols(); ++j)
     for(Index i = 0; i < rows(); ++i)
-      if(!internal::isMuchSmallerThan(this->coeff(i, j), static_cast<Scalar>(1), prec))
+      if(!internal::isMuchSmallerThan(self.coeff(i, j), static_cast<Scalar>(1), prec))
         return false;
   return true;
 }
@@ -496,7 +496,7 @@ bool DenseBase<Derived>::isZero(const RealScalar& prec) const
   * \sa class CwiseNullaryOp, Zero()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setZero()
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setZero()
 {
   return setConstant(Scalar(0));
 }
@@ -511,7 +511,7 @@ EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setZero()
   * \sa DenseBase::setZero(), setZero(Index,Index), class CwiseNullaryOp, DenseBase::Zero()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived&
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived&
 PlainObjectBase<Derived>::setZero(Index newSize)
 {
   resize(newSize);
@@ -520,8 +520,8 @@ PlainObjectBase<Derived>::setZero(Index newSize)
 
 /** Resizes to the given size, and sets all coefficients in this expression to zero.
   *
-  * \param nbRows the new number of rows
-  * \param nbCols the new number of columns
+  * \param rows the new number of rows
+  * \param cols the new number of columns
   *
   * Example: \include Matrix_setZero_int_int.cpp
   * Output: \verbinclude Matrix_setZero_int_int.out
@@ -529,10 +529,10 @@ PlainObjectBase<Derived>::setZero(Index newSize)
   * \sa DenseBase::setZero(), setZero(Index), class CwiseNullaryOp, DenseBase::Zero()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived&
-PlainObjectBase<Derived>::setZero(Index nbRows, Index nbCols)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived&
+PlainObjectBase<Derived>::setZero(Index rows, Index cols)
 {
-  resize(nbRows, nbCols);
+  resize(rows, cols);
   return setConstant(Scalar(0));
 }
 
@@ -540,7 +540,7 @@ PlainObjectBase<Derived>::setZero(Index nbRows, Index nbCols)
 
 /** \returns an expression of a matrix where all coefficients equal one.
   *
-  * The parameters \a nbRows and \a nbCols are the number of rows and of columns of
+  * The parameters \a rows and \a cols are the number of rows and of columns of
   * the returned matrix. Must be compatible with this MatrixBase type.
   *
   * This variant is meant to be used for dynamic-size matrix types. For fixed-size types,
@@ -553,10 +553,10 @@ PlainObjectBase<Derived>::setZero(Index nbRows, Index nbCols)
   * \sa Ones(), Ones(Index), isOnes(), class Ones
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
-DenseBase<Derived>::Ones(Index nbRows, Index nbCols)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+DenseBase<Derived>::Ones(Index rows, Index cols)
 {
-  return Constant(nbRows, nbCols, Scalar(1));
+  return Constant(rows, cols, Scalar(1));
 }
 
 /** \returns an expression of a vector where all coefficients equal one.
@@ -576,7 +576,7 @@ DenseBase<Derived>::Ones(Index nbRows, Index nbCols)
   * \sa Ones(), Ones(Index,Index), isOnes(), class Ones
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
 DenseBase<Derived>::Ones(Index newSize)
 {
   return Constant(newSize, Scalar(1));
@@ -593,7 +593,7 @@ DenseBase<Derived>::Ones(Index newSize)
   * \sa Ones(Index), Ones(Index,Index), isOnes(), class Ones
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
 DenseBase<Derived>::Ones()
 {
   return Constant(Scalar(1));
@@ -608,7 +608,7 @@ DenseBase<Derived>::Ones()
   * \sa class CwiseNullaryOp, Ones()
   */
 template<typename Derived>
-bool DenseBase<Derived>::isOnes
+EIGEN_DEVICE_FUNC bool DenseBase<Derived>::isOnes
 (const RealScalar& prec) const
 {
   return isApproxToConstant(Scalar(1), prec);
@@ -622,7 +622,7 @@ bool DenseBase<Derived>::isOnes
   * \sa class CwiseNullaryOp, Ones()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setOnes()
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setOnes()
 {
   return setConstant(Scalar(1));
 }
@@ -637,7 +637,7 @@ EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setOnes()
   * \sa MatrixBase::setOnes(), setOnes(Index,Index), class CwiseNullaryOp, MatrixBase::Ones()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived&
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived&
 PlainObjectBase<Derived>::setOnes(Index newSize)
 {
   resize(newSize);
@@ -646,8 +646,8 @@ PlainObjectBase<Derived>::setOnes(Index newSize)
 
 /** Resizes to the given size, and sets all coefficients in this expression to one.
   *
-  * \param nbRows the new number of rows
-  * \param nbCols the new number of columns
+  * \param rows the new number of rows
+  * \param cols the new number of columns
   *
   * Example: \include Matrix_setOnes_int_int.cpp
   * Output: \verbinclude Matrix_setOnes_int_int.out
@@ -655,10 +655,10 @@ PlainObjectBase<Derived>::setOnes(Index newSize)
   * \sa MatrixBase::setOnes(), setOnes(Index), class CwiseNullaryOp, MatrixBase::Ones()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived&
-PlainObjectBase<Derived>::setOnes(Index nbRows, Index nbCols)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived&
+PlainObjectBase<Derived>::setOnes(Index rows, Index cols)
 {
-  resize(nbRows, nbCols);
+  resize(rows, cols);
   return setConstant(Scalar(1));
 }
 
@@ -666,7 +666,7 @@ PlainObjectBase<Derived>::setOnes(Index nbRows, Index nbCols)
 
 /** \returns an expression of the identity matrix (not necessarily square).
   *
-  * The parameters \a nbRows and \a nbCols are the number of rows and of columns of
+  * The parameters \a rows and \a cols are the number of rows and of columns of
   * the returned matrix. Must be compatible with this MatrixBase type.
   *
   * This variant is meant to be used for dynamic-size matrix types. For fixed-size types,
@@ -679,10 +679,10 @@ PlainObjectBase<Derived>::setOnes(Index nbRows, Index nbCols)
   * \sa Identity(), setIdentity(), isIdentity()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::IdentityReturnType
-MatrixBase<Derived>::Identity(Index nbRows, Index nbCols)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::IdentityReturnType
+MatrixBase<Derived>::Identity(Index rows, Index cols)
 {
-  return DenseBase<Derived>::NullaryExpr(nbRows, nbCols, internal::scalar_identity_op<Scalar>());
+  return DenseBase<Derived>::NullaryExpr(rows, cols, internal::scalar_identity_op<Scalar>());
 }
 
 /** \returns an expression of the identity matrix (not necessarily square).
@@ -696,7 +696,7 @@ MatrixBase<Derived>::Identity(Index nbRows, Index nbCols)
   * \sa Identity(Index,Index), setIdentity(), isIdentity()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::IdentityReturnType
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::IdentityReturnType
 MatrixBase<Derived>::Identity()
 {
   EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived)
@@ -716,18 +716,19 @@ template<typename Derived>
 bool MatrixBase<Derived>::isIdentity
 (const RealScalar& prec) const
 {
+  typename internal::nested_eval<Derived,1>::type self(derived());
   for(Index j = 0; j < cols(); ++j)
   {
     for(Index i = 0; i < rows(); ++i)
     {
       if(i == j)
       {
-        if(!internal::isApprox(this->coeff(i, j), static_cast<Scalar>(1), prec))
+        if(!internal::isApprox(self.coeff(i, j), static_cast<Scalar>(1), prec))
           return false;
       }
       else
       {
-        if(!internal::isMuchSmallerThan(this->coeff(i, j), static_cast<RealScalar>(1), prec))
+        if(!internal::isMuchSmallerThan(self.coeff(i, j), static_cast<RealScalar>(1), prec))
           return false;
       }
     }
@@ -740,6 +741,7 @@ namespace internal {
 template<typename Derived, bool Big = (Derived::SizeAtCompileTime>=16)>
 struct setIdentity_impl
 {
+  EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE Derived& run(Derived& m)
   {
     return m = Derived::Identity(m.rows(), m.cols());
@@ -749,11 +751,11 @@ struct setIdentity_impl
 template<typename Derived>
 struct setIdentity_impl<Derived, true>
 {
-  typedef typename Derived::Index Index;
+  EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE Derived& run(Derived& m)
   {
     m.setZero();
-    const Index size = (std::min)(m.rows(), m.cols());
+    const Index size = numext::mini(m.rows(), m.cols());
     for(Index i = 0; i < size; ++i) m.coeffRef(i,i) = typename Derived::Scalar(1);
     return m;
   }
@@ -769,15 +771,15 @@ struct setIdentity_impl<Derived, true>
   * \sa class CwiseNullaryOp, Identity(), Identity(Index,Index), isIdentity()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setIdentity()
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setIdentity()
 {
   return internal::setIdentity_impl<Derived>::run(derived());
 }
 
 /** \brief Resizes to the given size, and writes the identity expression (not necessarily square) into *this.
   *
-  * \param nbRows the new number of rows
-  * \param nbCols the new number of columns
+  * \param rows the new number of rows
+  * \param cols the new number of columns
   *
   * Example: \include Matrix_setIdentity_int_int.cpp
   * Output: \verbinclude Matrix_setIdentity_int_int.out
@@ -785,9 +787,9 @@ EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setIdentity()
   * \sa MatrixBase::setIdentity(), class CwiseNullaryOp, MatrixBase::Identity()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setIdentity(Index nbRows, Index nbCols)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setIdentity(Index rows, Index cols)
 {
-  derived().resize(nbRows, nbCols);
+  derived().resize(rows, cols);
   return setIdentity();
 }
 
@@ -798,7 +800,7 @@ EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setIdentity(Index nbRows, Inde
   * \sa MatrixBase::Unit(Index), MatrixBase::UnitX(), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::Unit(Index newSize, Index i)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::Unit(Index newSize, Index i)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
   return BasisReturnType(SquareMatrixType::Identity(newSize,newSize), i);
@@ -813,7 +815,7 @@ EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBa
   * \sa MatrixBase::Unit(Index,Index), MatrixBase::UnitX(), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::Unit(Index i)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::Unit(Index i)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
   return BasisReturnType(SquareMatrixType::Identity(),i);
@@ -826,7 +828,7 @@ EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBa
   * \sa MatrixBase::Unit(Index,Index), MatrixBase::Unit(Index), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::UnitX()
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::UnitX()
 { return Derived::Unit(0); }
 
 /** \returns an expression of the Y axis unit vector (0,1{,0}^*)
@@ -836,7 +838,7 @@ EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBa
   * \sa MatrixBase::Unit(Index,Index), MatrixBase::Unit(Index), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::UnitY()
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::UnitY()
 { return Derived::Unit(1); }
 
 /** \returns an expression of the Z axis unit vector (0,0,1{,0}^*)
@@ -846,7 +848,7 @@ EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBa
   * \sa MatrixBase::Unit(Index,Index), MatrixBase::Unit(Index), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::UnitZ()
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::UnitZ()
 { return Derived::Unit(2); }
 
 /** \returns an expression of the W axis unit vector (0,0,0,1)
@@ -856,7 +858,7 @@ EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBa
   * \sa MatrixBase::Unit(Index,Index), MatrixBase::Unit(Index), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::UnitW()
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::UnitW()
 { return Derived::Unit(3); }
 
 } // end namespace Eigen
diff --git a/eigen/Eigen/src/Core/CwiseTernaryOp.h b/eigen/Eigen/src/Core/CwiseTernaryOp.h
new file mode 100644
index 0000000..9f3576f
--- /dev/null
+++ b/eigen/Eigen/src/Core/CwiseTernaryOp.h
@@ -0,0 +1,197 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2016 Eugene Brevdo <ebrevdo@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CWISE_TERNARY_OP_H
+#define EIGEN_CWISE_TERNARY_OP_H
+
+namespace Eigen {
+
+namespace internal {
+template <typename TernaryOp, typename Arg1, typename Arg2, typename Arg3>
+struct traits<CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> > {
+  // we must not inherit from traits<Arg1> since it has
+  // the potential to cause problems with MSVC
+  typedef typename remove_all<Arg1>::type Ancestor;
+  typedef typename traits<Ancestor>::XprKind XprKind;
+  enum {
+    RowsAtCompileTime = traits<Ancestor>::RowsAtCompileTime,
+    ColsAtCompileTime = traits<Ancestor>::ColsAtCompileTime,
+    MaxRowsAtCompileTime = traits<Ancestor>::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = traits<Ancestor>::MaxColsAtCompileTime
+  };
+
+  // even though we require Arg1, Arg2, and Arg3 to have the same scalar type
+  // (see CwiseTernaryOp constructor),
+  // we still want to handle the case when the result type is different.
+  typedef typename result_of<TernaryOp(
+      const typename Arg1::Scalar&, const typename Arg2::Scalar&,
+      const typename Arg3::Scalar&)>::type Scalar;
+
+  typedef typename internal::traits<Arg1>::StorageKind StorageKind;
+  typedef typename internal::traits<Arg1>::StorageIndex StorageIndex;
+
+  typedef typename Arg1::Nested Arg1Nested;
+  typedef typename Arg2::Nested Arg2Nested;
+  typedef typename Arg3::Nested Arg3Nested;
+  typedef typename remove_reference<Arg1Nested>::type _Arg1Nested;
+  typedef typename remove_reference<Arg2Nested>::type _Arg2Nested;
+  typedef typename remove_reference<Arg3Nested>::type _Arg3Nested;
+  enum { Flags = _Arg1Nested::Flags & RowMajorBit };
+};
+}  // end namespace internal
+
+template <typename TernaryOp, typename Arg1, typename Arg2, typename Arg3,
+          typename StorageKind>
+class CwiseTernaryOpImpl;
+
+/** \class CwiseTernaryOp
+  * \ingroup Core_Module
+  *
+  * \brief Generic expression where a coefficient-wise ternary operator is
+ * applied to two expressions
+  *
+  * \tparam TernaryOp template functor implementing the operator
+  * \tparam Arg1Type the type of the first argument
+  * \tparam Arg2Type the type of the second argument
+  * \tparam Arg3Type the type of the third argument
+  *
+  * This class represents an expression where a coefficient-wise ternary
+ * operator is applied to three expressions.
+  * It is the return type of ternary operators, by which we mean only those
+ * ternary operators where
+  * all three arguments are Eigen expressions.
+  * For example, the return type of betainc(matrix1, matrix2, matrix3) is a
+ * CwiseTernaryOp.
+  *
+  * Most of the time, this is the only way that it is used, so you typically
+ * don't have to name
+  * CwiseTernaryOp types explicitly.
+  *
+  * \sa MatrixBase::ternaryExpr(const MatrixBase<Argument2> &, const
+ * MatrixBase<Argument3> &, const CustomTernaryOp &) const, class CwiseBinaryOp,
+ * class CwiseUnaryOp, class CwiseNullaryOp
+  */
+template <typename TernaryOp, typename Arg1Type, typename Arg2Type,
+          typename Arg3Type>
+class CwiseTernaryOp : public CwiseTernaryOpImpl<
+                           TernaryOp, Arg1Type, Arg2Type, Arg3Type,
+                           typename internal::traits<Arg1Type>::StorageKind>,
+                       internal::no_assignment_operator
+{
+ public:
+  typedef typename internal::remove_all<Arg1Type>::type Arg1;
+  typedef typename internal::remove_all<Arg2Type>::type Arg2;
+  typedef typename internal::remove_all<Arg3Type>::type Arg3;
+
+  typedef typename CwiseTernaryOpImpl<
+      TernaryOp, Arg1Type, Arg2Type, Arg3Type,
+      typename internal::traits<Arg1Type>::StorageKind>::Base Base;
+  EIGEN_GENERIC_PUBLIC_INTERFACE(CwiseTernaryOp)
+
+  typedef typename internal::ref_selector<Arg1Type>::type Arg1Nested;
+  typedef typename internal::ref_selector<Arg2Type>::type Arg2Nested;
+  typedef typename internal::ref_selector<Arg3Type>::type Arg3Nested;
+  typedef typename internal::remove_reference<Arg1Nested>::type _Arg1Nested;
+  typedef typename internal::remove_reference<Arg2Nested>::type _Arg2Nested;
+  typedef typename internal::remove_reference<Arg3Nested>::type _Arg3Nested;
+
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE CwiseTernaryOp(const Arg1& a1, const Arg2& a2,
+                                     const Arg3& a3,
+                                     const TernaryOp& func = TernaryOp())
+      : m_arg1(a1), m_arg2(a2), m_arg3(a3), m_functor(func) {
+    // require the sizes to match
+    EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Arg1, Arg2)
+    EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Arg1, Arg3)
+
+    // The index types should match
+    EIGEN_STATIC_ASSERT((internal::is_same<
+                         typename internal::traits<Arg1Type>::StorageKind,
+                         typename internal::traits<Arg2Type>::StorageKind>::value),
+                        STORAGE_KIND_MUST_MATCH)
+    EIGEN_STATIC_ASSERT((internal::is_same<
+                         typename internal::traits<Arg1Type>::StorageKind,
+                         typename internal::traits<Arg3Type>::StorageKind>::value),
+                        STORAGE_KIND_MUST_MATCH)
+
+    eigen_assert(a1.rows() == a2.rows() && a1.cols() == a2.cols() &&
+                 a1.rows() == a3.rows() && a1.cols() == a3.cols());
+  }
+
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE Index rows() const {
+    // return the fixed size type if available to enable compile time
+    // optimizations
+    if (internal::traits<typename internal::remove_all<Arg1Nested>::type>::
+                RowsAtCompileTime == Dynamic &&
+        internal::traits<typename internal::remove_all<Arg2Nested>::type>::
+                RowsAtCompileTime == Dynamic)
+      return m_arg3.rows();
+    else if (internal::traits<typename internal::remove_all<Arg1Nested>::type>::
+                     RowsAtCompileTime == Dynamic &&
+             internal::traits<typename internal::remove_all<Arg3Nested>::type>::
+                     RowsAtCompileTime == Dynamic)
+      return m_arg2.rows();
+    else
+      return m_arg1.rows();
+  }
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE Index cols() const {
+    // return the fixed size type if available to enable compile time
+    // optimizations
+    if (internal::traits<typename internal::remove_all<Arg1Nested>::type>::
+                ColsAtCompileTime == Dynamic &&
+        internal::traits<typename internal::remove_all<Arg2Nested>::type>::
+                ColsAtCompileTime == Dynamic)
+      return m_arg3.cols();
+    else if (internal::traits<typename internal::remove_all<Arg1Nested>::type>::
+                     ColsAtCompileTime == Dynamic &&
+             internal::traits<typename internal::remove_all<Arg3Nested>::type>::
+                     ColsAtCompileTime == Dynamic)
+      return m_arg2.cols();
+    else
+      return m_arg1.cols();
+  }
+
+  /** \returns the first argument nested expression */
+  EIGEN_DEVICE_FUNC
+  const _Arg1Nested& arg1() const { return m_arg1; }
+  /** \returns the first argument nested expression */
+  EIGEN_DEVICE_FUNC
+  const _Arg2Nested& arg2() const { return m_arg2; }
+  /** \returns the third argument nested expression */
+  EIGEN_DEVICE_FUNC
+  const _Arg3Nested& arg3() const { return m_arg3; }
+  /** \returns the functor representing the ternary operation */
+  EIGEN_DEVICE_FUNC
+  const TernaryOp& functor() const { return m_functor; }
+
+ protected:
+  Arg1Nested m_arg1;
+  Arg2Nested m_arg2;
+  Arg3Nested m_arg3;
+  const TernaryOp m_functor;
+};
+
+// Generic API dispatcher
+template <typename TernaryOp, typename Arg1, typename Arg2, typename Arg3,
+          typename StorageKind>
+class CwiseTernaryOpImpl
+    : public internal::generic_xpr_base<
+          CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> >::type {
+ public:
+  typedef typename internal::generic_xpr_base<
+      CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> >::type Base;
+};
+
+}  // end namespace Eigen
+
+#endif  // EIGEN_CWISE_TERNARY_OP_H
diff --git a/eigen/Eigen/src/Core/CwiseUnaryOp.h b/eigen/Eigen/src/Core/CwiseUnaryOp.h
index f7ee60e..1d2dd19 100644
--- a/eigen/Eigen/src/Core/CwiseUnaryOp.h
+++ b/eigen/Eigen/src/Core/CwiseUnaryOp.h
@@ -1,7 +1,7 @@
 // This file is part of Eigen, a lightweight C++ template library
 // for linear algebra.
 //
-// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2008-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
 // Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
 //
 // This Source Code Form is subject to the terms of the Mozilla
@@ -13,41 +13,18 @@
 
 namespace Eigen { 
 
-/** \class CwiseUnaryOp
-  * \ingroup Core_Module
-  *
-  * \brief Generic expression where a coefficient-wise unary operator is applied to an expression
-  *
-  * \param UnaryOp template functor implementing the operator
-  * \param XprType the type of the expression to which we are applying the unary operator
-  *
-  * This class represents an expression where a unary operator is applied to an expression.
-  * It is the return type of all operations taking exactly 1 input expression, regardless of the
-  * presence of other inputs such as scalars. For example, the operator* in the expression 3*matrix
-  * is considered unary, because only the right-hand side is an expression, and its
-  * return type is a specialization of CwiseUnaryOp.
-  *
-  * Most of the time, this is the only way that it is used, so you typically don't have to name
-  * CwiseUnaryOp types explicitly.
-  *
-  * \sa MatrixBase::unaryExpr(const CustomUnaryOp &) const, class CwiseBinaryOp, class CwiseNullaryOp
-  */
-
 namespace internal {
 template<typename UnaryOp, typename XprType>
 struct traits<CwiseUnaryOp<UnaryOp, XprType> >
  : traits<XprType>
 {
   typedef typename result_of<
-                     UnaryOp(typename XprType::Scalar)
+                     UnaryOp(const typename XprType::Scalar&)
                    >::type Scalar;
   typedef typename XprType::Nested XprTypeNested;
   typedef typename remove_reference<XprTypeNested>::type _XprTypeNested;
   enum {
-    Flags = _XprTypeNested::Flags & (
-      HereditaryBits | LinearAccessBit | AlignedBit
-      | (functor_traits<UnaryOp>::PacketAccess ? PacketAccessBit : 0)),
-    CoeffReadCost = EIGEN_ADD_COST(_XprTypeNested::CoeffReadCost, functor_traits<UnaryOp>::Cost)
+    Flags = _XprTypeNested::Flags & RowMajorBit 
   };
 };
 }
@@ -55,70 +32,70 @@ struct traits<CwiseUnaryOp<UnaryOp, XprType> >
 template<typename UnaryOp, typename XprType, typename StorageKind>
 class CwiseUnaryOpImpl;
 
+/** \class CwiseUnaryOp
+  * \ingroup Core_Module
+  *
+  * \brief Generic expression where a coefficient-wise unary operator is applied to an expression
+  *
+  * \tparam UnaryOp template functor implementing the operator
+  * \tparam XprType the type of the expression to which we are applying the unary operator
+  *
+  * This class represents an expression where a unary operator is applied to an expression.
+  * It is the return type of all operations taking exactly 1 input expression, regardless of the
+  * presence of other inputs such as scalars. For example, the operator* in the expression 3*matrix
+  * is considered unary, because only the right-hand side is an expression, and its
+  * return type is a specialization of CwiseUnaryOp.
+  *
+  * Most of the time, this is the only way that it is used, so you typically don't have to name
+  * CwiseUnaryOp types explicitly.
+  *
+  * \sa MatrixBase::unaryExpr(const CustomUnaryOp &) const, class CwiseBinaryOp, class CwiseNullaryOp
+  */
 template<typename UnaryOp, typename XprType>
-class CwiseUnaryOp : internal::no_assignment_operator,
-  public CwiseUnaryOpImpl<UnaryOp, XprType, typename internal::traits<XprType>::StorageKind>
+class CwiseUnaryOp : public CwiseUnaryOpImpl<UnaryOp, XprType, typename internal::traits<XprType>::StorageKind>, internal::no_assignment_operator
 {
   public:
 
     typedef typename CwiseUnaryOpImpl<UnaryOp, XprType,typename internal::traits<XprType>::StorageKind>::Base Base;
     EIGEN_GENERIC_PUBLIC_INTERFACE(CwiseUnaryOp)
+    typedef typename internal::ref_selector<XprType>::type XprTypeNested;
+    typedef typename internal::remove_all<XprType>::type NestedExpression;
 
-    inline CwiseUnaryOp(const XprType& xpr, const UnaryOp& func = UnaryOp())
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    explicit CwiseUnaryOp(const XprType& xpr, const UnaryOp& func = UnaryOp())
       : m_xpr(xpr), m_functor(func) {}
 
-    EIGEN_STRONG_INLINE Index rows() const { return m_xpr.rows(); }
-    EIGEN_STRONG_INLINE Index cols() const { return m_xpr.cols(); }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Index rows() const { return m_xpr.rows(); }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Index cols() const { return m_xpr.cols(); }
 
     /** \returns the functor representing the unary operation */
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     const UnaryOp& functor() const { return m_functor; }
 
     /** \returns the nested expression */
-    const typename internal::remove_all<typename XprType::Nested>::type&
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const typename internal::remove_all<XprTypeNested>::type&
     nestedExpression() const { return m_xpr; }
 
     /** \returns the nested expression */
-    typename internal::remove_all<typename XprType::Nested>::type&
-    nestedExpression() { return m_xpr.const_cast_derived(); }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    typename internal::remove_all<XprTypeNested>::type&
+    nestedExpression() { return m_xpr; }
 
   protected:
-    typename XprType::Nested m_xpr;
+    XprTypeNested m_xpr;
     const UnaryOp m_functor;
 };
 
-// This is the generic implementation for dense storage.
-// It can be used for any expression types implementing the dense concept.
-template<typename UnaryOp, typename XprType>
-class CwiseUnaryOpImpl<UnaryOp,XprType,Dense>
-  : public internal::dense_xpr_base<CwiseUnaryOp<UnaryOp, XprType> >::type
+// Generic API dispatcher
+template<typename UnaryOp, typename XprType, typename StorageKind>
+class CwiseUnaryOpImpl
+  : public internal::generic_xpr_base<CwiseUnaryOp<UnaryOp, XprType> >::type
 {
-  public:
-
-    typedef CwiseUnaryOp<UnaryOp, XprType> Derived;
-    typedef typename internal::dense_xpr_base<CwiseUnaryOp<UnaryOp, XprType> >::type Base;
-    EIGEN_DENSE_PUBLIC_INTERFACE(Derived)
-
-    EIGEN_STRONG_INLINE const Scalar coeff(Index rowId, Index colId) const
-    {
-      return derived().functor()(derived().nestedExpression().coeff(rowId, colId));
-    }
-
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet(Index rowId, Index colId) const
-    {
-      return derived().functor().packetOp(derived().nestedExpression().template packet<LoadMode>(rowId, colId));
-    }
-
-    EIGEN_STRONG_INLINE const Scalar coeff(Index index) const
-    {
-      return derived().functor()(derived().nestedExpression().coeff(index));
-    }
-
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet(Index index) const
-    {
-      return derived().functor().packetOp(derived().nestedExpression().template packet<LoadMode>(index));
-    }
+public:
+  typedef typename internal::generic_xpr_base<CwiseUnaryOp<UnaryOp, XprType> >::type Base;
 };
 
 } // end namespace Eigen
diff --git a/eigen/Eigen/src/Core/CwiseUnaryView.h b/eigen/Eigen/src/Core/CwiseUnaryView.h
index f3b2ffe..2710330 100644
--- a/eigen/Eigen/src/Core/CwiseUnaryView.h
+++ b/eigen/Eigen/src/Core/CwiseUnaryView.h
@@ -12,33 +12,19 @@
 
 namespace Eigen {
 
-/** \class CwiseUnaryView
-  * \ingroup Core_Module
-  *
-  * \brief Generic lvalue expression of a coefficient-wise unary operator of a matrix or a vector
-  *
-  * \param ViewOp template functor implementing the view
-  * \param MatrixType the type of the matrix we are applying the unary operator
-  *
-  * This class represents a lvalue expression of a generic unary view operator of a matrix or a vector.
-  * It is the return type of real() and imag(), and most of the time this is the only way it is used.
-  *
-  * \sa MatrixBase::unaryViewExpr(const CustomUnaryOp &) const, class CwiseUnaryOp
-  */
-
 namespace internal {
 template<typename ViewOp, typename MatrixType>
 struct traits<CwiseUnaryView<ViewOp, MatrixType> >
  : traits<MatrixType>
 {
   typedef typename result_of<
-                     ViewOp(typename traits<MatrixType>::Scalar)
+                     ViewOp(const typename traits<MatrixType>::Scalar&)
                    >::type Scalar;
   typedef typename MatrixType::Nested MatrixTypeNested;
   typedef typename remove_all<MatrixTypeNested>::type _MatrixTypeNested;
   enum {
-    Flags = (traits<_MatrixTypeNested>::Flags & (HereditaryBits | LvalueBit | LinearAccessBit | DirectAccessBit)),
-    CoeffReadCost = EIGEN_ADD_COST(traits<_MatrixTypeNested>::CoeffReadCost, functor_traits<ViewOp>::Cost),
+    FlagsLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0,
+    Flags = traits<_MatrixTypeNested>::Flags & (RowMajorBit | FlagsLvalueBit | DirectAccessBit), // FIXME DirectAccessBit should not be handled by expressions
     MatrixTypeInnerStride =  inner_stride_at_compile_time<MatrixType>::ret,
     // need to cast the sizeof's from size_t to int explicitly, otherwise:
     // "error: no integral type can represent all of the enumerator values
@@ -55,6 +41,19 @@ struct traits<CwiseUnaryView<ViewOp, MatrixType> >
 template<typename ViewOp, typename MatrixType, typename StorageKind>
 class CwiseUnaryViewImpl;
 
+/** \class CwiseUnaryView
+  * \ingroup Core_Module
+  *
+  * \brief Generic lvalue expression of a coefficient-wise unary operator of a matrix or a vector
+  *
+  * \tparam ViewOp template functor implementing the view
+  * \tparam MatrixType the type of the matrix we are applying the unary operator
+  *
+  * This class represents a lvalue expression of a generic unary view operator of a matrix or a vector.
+  * It is the return type of real() and imag(), and most of the time this is the only way it is used.
+  *
+  * \sa MatrixBase::unaryViewExpr(const CustomUnaryOp &) const, class CwiseUnaryOp
+  */
 template<typename ViewOp, typename MatrixType>
 class CwiseUnaryView : public CwiseUnaryViewImpl<ViewOp, MatrixType, typename internal::traits<MatrixType>::StorageKind>
 {
@@ -62,8 +61,10 @@ class CwiseUnaryView : public CwiseUnaryViewImpl<ViewOp, MatrixType, typename in
 
     typedef typename CwiseUnaryViewImpl<ViewOp, MatrixType,typename internal::traits<MatrixType>::StorageKind>::Base Base;
     EIGEN_GENERIC_PUBLIC_INTERFACE(CwiseUnaryView)
+    typedef typename internal::ref_selector<MatrixType>::non_const_type MatrixTypeNested;
+    typedef typename internal::remove_all<MatrixType>::type NestedExpression;
 
-    inline CwiseUnaryView(const MatrixType& mat, const ViewOp& func = ViewOp())
+    explicit inline CwiseUnaryView(MatrixType& mat, const ViewOp& func = ViewOp())
       : m_matrix(mat), m_functor(func) {}
 
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(CwiseUnaryView)
@@ -75,19 +76,27 @@ class CwiseUnaryView : public CwiseUnaryViewImpl<ViewOp, MatrixType, typename in
     const ViewOp& functor() const { return m_functor; }
 
     /** \returns the nested expression */
-    const typename internal::remove_all<typename MatrixType::Nested>::type&
+    const typename internal::remove_all<MatrixTypeNested>::type&
     nestedExpression() const { return m_matrix; }
 
     /** \returns the nested expression */
-    typename internal::remove_all<typename MatrixType::Nested>::type&
+    typename internal::remove_reference<MatrixTypeNested>::type&
     nestedExpression() { return m_matrix.const_cast_derived(); }
 
   protected:
-    // FIXME changed from MatrixType::Nested because of a weird compilation error with sun CC
-    typename internal::nested<MatrixType>::type m_matrix;
+    MatrixTypeNested m_matrix;
     ViewOp m_functor;
 };
 
+// Generic API dispatcher
+template<typename ViewOp, typename XprType, typename StorageKind>
+class CwiseUnaryViewImpl
+  : public internal::generic_xpr_base<CwiseUnaryView<ViewOp, XprType> >::type
+{
+public:
+  typedef typename internal::generic_xpr_base<CwiseUnaryView<ViewOp, XprType> >::type Base;
+};
+
 template<typename ViewOp, typename MatrixType>
 class CwiseUnaryViewImpl<ViewOp,MatrixType,Dense>
   : public internal::dense_xpr_base< CwiseUnaryView<ViewOp, MatrixType> >::type
@@ -100,38 +109,18 @@ class CwiseUnaryViewImpl<ViewOp,MatrixType,Dense>
     EIGEN_DENSE_PUBLIC_INTERFACE(Derived)
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(CwiseUnaryViewImpl)
     
-    inline Scalar* data() { return &coeffRef(0); }
-    inline const Scalar* data() const { return &coeff(0); }
+    EIGEN_DEVICE_FUNC inline Scalar* data() { return &(this->coeffRef(0)); }
+    EIGEN_DEVICE_FUNC inline const Scalar* data() const { return &(this->coeff(0)); }
 
-    inline Index innerStride() const
+    EIGEN_DEVICE_FUNC inline Index innerStride() const
     {
       return derived().nestedExpression().innerStride() * sizeof(typename internal::traits<MatrixType>::Scalar) / sizeof(Scalar);
     }
 
-    inline Index outerStride() const
+    EIGEN_DEVICE_FUNC inline Index outerStride() const
     {
       return derived().nestedExpression().outerStride() * sizeof(typename internal::traits<MatrixType>::Scalar) / sizeof(Scalar);
     }
-
-    EIGEN_STRONG_INLINE CoeffReturnType coeff(Index row, Index col) const
-    {
-      return derived().functor()(derived().nestedExpression().coeff(row, col));
-    }
-
-    EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
-    {
-      return derived().functor()(derived().nestedExpression().coeff(index));
-    }
-
-    EIGEN_STRONG_INLINE Scalar& coeffRef(Index row, Index col)
-    {
-      return derived().functor()(const_cast_derived().nestedExpression().coeffRef(row, col));
-    }
-
-    EIGEN_STRONG_INLINE Scalar& coeffRef(Index index)
-    {
-      return derived().functor()(const_cast_derived().nestedExpression().coeffRef(index));
-    }
 };
 
 } // end namespace Eigen
diff --git a/eigen/Eigen/src/Core/DenseBase.h b/eigen/Eigen/src/Core/DenseBase.h
index 4b371b0..fd933ee 100644
--- a/eigen/Eigen/src/Core/DenseBase.h
+++ b/eigen/Eigen/src/Core/DenseBase.h
@@ -34,37 +34,45 @@ static inline void check_DenseIndex_is_signed() {
   * \tparam Derived is the derived type, e.g., a matrix type or an expression.
   *
   * This class can be extended with the help of the plugin mechanism described on the page
-  * \ref TopicCustomizingEigen by defining the preprocessor symbol \c EIGEN_DENSEBASE_PLUGIN.
+  * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_DENSEBASE_PLUGIN.
   *
-  * \sa \ref TopicClassHierarchy
+  * \sa \blank \ref TopicClassHierarchy
   */
 template<typename Derived> class DenseBase
 #ifndef EIGEN_PARSED_BY_DOXYGEN
-  : public internal::special_scalar_op_base<Derived, typename internal::traits<Derived>::Scalar,
-                                            typename NumTraits<typename internal::traits<Derived>::Scalar>::Real,
-                                            DenseCoeffsBase<Derived> >
-#else
   : public DenseCoeffsBase<Derived>
+#else
+  : public DenseCoeffsBase<Derived,DirectWriteAccessors>
 #endif // not EIGEN_PARSED_BY_DOXYGEN
 {
   public:
 
-    class InnerIterator;
+    /** Inner iterator type to iterate over the coefficients of a row or column.
+      * \sa class InnerIterator
+      */
+    typedef Eigen::InnerIterator<Derived> InnerIterator;
 
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
 
-    /** \brief The type of indices 
-      * \details To change this, \c \#define the preprocessor symbol \c EIGEN_DEFAULT_DENSE_INDEX_TYPE.
-      * \sa \ref TopicPreprocessorDirectives.
-      */
-    typedef typename internal::traits<Derived>::Index Index; 
+    /**
+      * \brief The type used to store indices
+      * \details This typedef is relevant for types that store multiple indices such as
+      *          PermutationMatrix or Transpositions, otherwise it defaults to Eigen::Index
+      * \sa \blank \ref TopicPreprocessorDirectives, Eigen::Index, SparseMatrixBase.
+     */
+    typedef typename internal::traits<Derived>::StorageIndex StorageIndex;
 
+    /** The numeric type of the expression' coefficients, e.g. float, double, int or std::complex<float>, etc. */
     typedef typename internal::traits<Derived>::Scalar Scalar;
-    typedef typename internal::packet_traits<Scalar>::type PacketScalar;
+    
+    /** The numeric type of the expression' coefficients, e.g. float, double, int or std::complex<float>, etc.
+      *
+      * It is an alias for the Scalar type */
+    typedef Scalar value_type;
+    
     typedef typename NumTraits<Scalar>::Real RealScalar;
-    typedef internal::special_scalar_op_base<Derived,Scalar,RealScalar, DenseCoeffsBase<Derived> > Base;
+    typedef DenseCoeffsBase<Derived> Base;
 
-    using Base::operator*;
     using Base::derived;
     using Base::const_cast_derived;
     using Base::rows;
@@ -74,16 +82,6 @@ template<typename Derived> class DenseBase
     using Base::colIndexByOuterInner;
     using Base::coeff;
     using Base::coeffByOuterInner;
-    using Base::packet;
-    using Base::packetByOuterInner;
-    using Base::writePacket;
-    using Base::writePacketByOuterInner;
-    using Base::coeffRef;
-    using Base::coeffRefByOuterInner;
-    using Base::copyCoeff;
-    using Base::copyCoeffByOuterInner;
-    using Base::copyPacket;
-    using Base::copyPacketByOuterInner;
     using Base::operator();
     using Base::operator[];
     using Base::x;
@@ -169,19 +167,46 @@ template<typename Derived> class DenseBase
       InnerSizeAtCompileTime = int(IsVectorAtCompileTime) ? int(SizeAtCompileTime)
                              : int(IsRowMajor) ? int(ColsAtCompileTime) : int(RowsAtCompileTime),
 
-      CoeffReadCost = internal::traits<Derived>::CoeffReadCost,
-        /**< This is a rough measure of how expensive it is to read one coefficient from
-          * this expression.
-          */
-
       InnerStrideAtCompileTime = internal::inner_stride_at_compile_time<Derived>::ret,
       OuterStrideAtCompileTime = internal::outer_stride_at_compile_time<Derived>::ret
     };
+    
+    typedef typename internal::find_best_packet<Scalar,SizeAtCompileTime>::type PacketScalar;
 
-    enum { ThisConstantIsPrivateInPlainObjectBase };
+    enum { IsPlainObjectBase = 0 };
+    
+    /** The plain matrix type corresponding to this expression.
+      * \sa PlainObject */
+    typedef Matrix<typename internal::traits<Derived>::Scalar,
+                internal::traits<Derived>::RowsAtCompileTime,
+                internal::traits<Derived>::ColsAtCompileTime,
+                AutoAlign | (internal::traits<Derived>::Flags&RowMajorBit ? RowMajor : ColMajor),
+                internal::traits<Derived>::MaxRowsAtCompileTime,
+                internal::traits<Derived>::MaxColsAtCompileTime
+          > PlainMatrix;
+    
+    /** The plain array type corresponding to this expression.
+      * \sa PlainObject */
+    typedef Array<typename internal::traits<Derived>::Scalar,
+                internal::traits<Derived>::RowsAtCompileTime,
+                internal::traits<Derived>::ColsAtCompileTime,
+                AutoAlign | (internal::traits<Derived>::Flags&RowMajorBit ? RowMajor : ColMajor),
+                internal::traits<Derived>::MaxRowsAtCompileTime,
+                internal::traits<Derived>::MaxColsAtCompileTime
+          > PlainArray;
+
+    /** \brief The plain matrix or array type corresponding to this expression.
+      *
+      * This is not necessarily exactly the return type of eval(). In the case of plain matrices,
+      * the return type of eval() is a const reference to a matrix, not a matrix! It is however guaranteed
+      * that the return type of eval() is either PlainObject or const PlainObject&.
+      */
+    typedef typename internal::conditional<internal::is_same<typename internal::traits<Derived>::XprKind,MatrixXpr >::value,
+                                 PlainMatrix, PlainArray>::type PlainObject;
 
     /** \returns the number of nonzero coefficients which is in practice the number
       * of stored coefficients. */
+    EIGEN_DEVICE_FUNC
     inline Index nonZeros() const { return size(); }
 
     /** \returns the outer size.
@@ -189,6 +214,7 @@ template<typename Derived> class DenseBase
       * \note For a vector, this returns just 1. For a matrix (non-vector), this is the major dimension
       * with respect to the \ref TopicStorageOrders "storage order", i.e., the number of columns for a
       * column-major matrix, and the number of rows for a row-major matrix. */
+    EIGEN_DEVICE_FUNC
     Index outerSize() const
     {
       return IsVectorAtCompileTime ? 1
@@ -200,6 +226,7 @@ template<typename Derived> class DenseBase
       * \note For a vector, this is just the size. For a matrix (non-vector), this is the minor dimension
       * with respect to the \ref TopicStorageOrders "storage order", i.e., the number of rows for a 
       * column-major matrix, and the number of columns for a row-major matrix. */
+    EIGEN_DEVICE_FUNC
     Index innerSize() const
     {
       return IsVectorAtCompileTime ? this->size()
@@ -210,6 +237,7 @@ template<typename Derived> class DenseBase
       * Matrix::resize() and Array::resize(). The present method only asserts that the new size equals the old size, and does
       * nothing else.
       */
+    EIGEN_DEVICE_FUNC
     void resize(Index newSize)
     {
       EIGEN_ONLY_USED_FOR_DEBUG(newSize);
@@ -220,22 +248,22 @@ template<typename Derived> class DenseBase
       * Matrix::resize() and Array::resize(). The present method only asserts that the new size equals the old size, and does
       * nothing else.
       */
-    void resize(Index nbRows, Index nbCols)
+    EIGEN_DEVICE_FUNC
+    void resize(Index rows, Index cols)
     {
-      EIGEN_ONLY_USED_FOR_DEBUG(nbRows);
-      EIGEN_ONLY_USED_FOR_DEBUG(nbCols);
-      eigen_assert(nbRows == this->rows() && nbCols == this->cols()
+      EIGEN_ONLY_USED_FOR_DEBUG(rows);
+      EIGEN_ONLY_USED_FOR_DEBUG(cols);
+      eigen_assert(rows == this->rows() && cols == this->cols()
                 && "DenseBase::resize() does not actually allow to resize.");
     }
 
 #ifndef EIGEN_PARSED_BY_DOXYGEN
-
     /** \internal Represents a matrix with all coefficients equal to one another*/
-    typedef CwiseNullaryOp<internal::scalar_constant_op<Scalar>,Derived> ConstantReturnType;
-    /** \internal Represents a vector with linearly spaced coefficients that allows sequential access only. */
-    typedef CwiseNullaryOp<internal::linspaced_op<Scalar,false>,Derived> SequentialLinSpacedReturnType;
+    typedef CwiseNullaryOp<internal::scalar_constant_op<Scalar>,PlainObject> ConstantReturnType;
+    /** \internal \deprecated Represents a vector with linearly spaced coefficients that allows sequential access only. */
+    typedef CwiseNullaryOp<internal::linspaced_op<Scalar,PacketScalar>,PlainObject> SequentialLinSpacedReturnType;
     /** \internal Represents a vector with linearly spaced coefficients that allows random access. */
-    typedef CwiseNullaryOp<internal::linspaced_op<Scalar,true>,Derived> RandomAccessLinSpacedReturnType;
+    typedef CwiseNullaryOp<internal::linspaced_op<Scalar,PacketScalar>,PlainObject> RandomAccessLinSpacedReturnType;
     /** \internal the return type of MatrixBase::eigenvalues() */
     typedef Matrix<typename NumTraits<typename internal::traits<Derived>::Scalar>::Real, internal::traits<Derived>::ColsAtCompileTime, 1> EigenvaluesReturnType;
 
@@ -243,120 +271,133 @@ template<typename Derived> class DenseBase
 
     /** Copies \a other into *this. \returns a reference to *this. */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     Derived& operator=(const DenseBase<OtherDerived>& other);
 
     /** Special case of the template operator=, in order to prevent the compiler
       * from generating a default operator= (issue hit with g++ 4.1)
       */
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     Derived& operator=(const DenseBase& other);
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     Derived& operator=(const EigenBase<OtherDerived> &other);
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     Derived& operator+=(const EigenBase<OtherDerived> &other);
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     Derived& operator-=(const EigenBase<OtherDerived> &other);
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     Derived& operator=(const ReturnByValue<OtherDerived>& func);
 
-    /** \internal Copies \a other into *this without evaluating other. \returns a reference to *this. */
+    /** \internal
+      * Copies \a other into *this without evaluating other. \returns a reference to *this.
+      * \deprecated */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     Derived& lazyAssign(const DenseBase<OtherDerived>& other);
 
-    /** \internal Evaluates \a other into *this. \returns a reference to *this. */
-    template<typename OtherDerived>
-    Derived& lazyAssign(const ReturnByValue<OtherDerived>& other);
-
+    EIGEN_DEVICE_FUNC
     CommaInitializer<Derived> operator<< (const Scalar& s);
 
+    /** \deprecated it now returns \c *this */
     template<unsigned int Added,unsigned int Removed>
-    const Flagged<Derived, Added, Removed> flagged() const;
+    EIGEN_DEPRECATED
+    const Derived& flagged() const
+    { return derived(); }
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     CommaInitializer<Derived> operator<< (const DenseBase<OtherDerived>& other);
 
-    Eigen::Transpose<Derived> transpose();
-	typedef typename internal::add_const<Transpose<const Derived> >::type ConstTransposeReturnType;
+    typedef Transpose<Derived> TransposeReturnType;
+    EIGEN_DEVICE_FUNC
+    TransposeReturnType transpose();
+    typedef typename internal::add_const<Transpose<const Derived> >::type ConstTransposeReturnType;
+    EIGEN_DEVICE_FUNC
     ConstTransposeReturnType transpose() const;
+    EIGEN_DEVICE_FUNC
     void transposeInPlace();
-#ifndef EIGEN_NO_DEBUG
-  protected:
-    template<typename OtherDerived>
-    void checkTransposeAliasing(const OtherDerived& other) const;
-  public:
-#endif
-
 
-    static const ConstantReturnType
+    EIGEN_DEVICE_FUNC static const ConstantReturnType
     Constant(Index rows, Index cols, const Scalar& value);
-    static const ConstantReturnType
+    EIGEN_DEVICE_FUNC static const ConstantReturnType
     Constant(Index size, const Scalar& value);
-    static const ConstantReturnType
+    EIGEN_DEVICE_FUNC static const ConstantReturnType
     Constant(const Scalar& value);
 
-    static const SequentialLinSpacedReturnType
+    EIGEN_DEVICE_FUNC static const SequentialLinSpacedReturnType
     LinSpaced(Sequential_t, Index size, const Scalar& low, const Scalar& high);
-    static const RandomAccessLinSpacedReturnType
+    EIGEN_DEVICE_FUNC static const RandomAccessLinSpacedReturnType
     LinSpaced(Index size, const Scalar& low, const Scalar& high);
-    static const SequentialLinSpacedReturnType
+    EIGEN_DEVICE_FUNC static const SequentialLinSpacedReturnType
     LinSpaced(Sequential_t, const Scalar& low, const Scalar& high);
-    static const RandomAccessLinSpacedReturnType
+    EIGEN_DEVICE_FUNC static const RandomAccessLinSpacedReturnType
     LinSpaced(const Scalar& low, const Scalar& high);
 
-    template<typename CustomNullaryOp>
-    static const CwiseNullaryOp<CustomNullaryOp, Derived>
+    template<typename CustomNullaryOp> EIGEN_DEVICE_FUNC
+    static const CwiseNullaryOp<CustomNullaryOp, PlainObject>
     NullaryExpr(Index rows, Index cols, const CustomNullaryOp& func);
-    template<typename CustomNullaryOp>
-    static const CwiseNullaryOp<CustomNullaryOp, Derived>
+    template<typename CustomNullaryOp> EIGEN_DEVICE_FUNC
+    static const CwiseNullaryOp<CustomNullaryOp, PlainObject>
     NullaryExpr(Index size, const CustomNullaryOp& func);
-    template<typename CustomNullaryOp>
-    static const CwiseNullaryOp<CustomNullaryOp, Derived>
+    template<typename CustomNullaryOp> EIGEN_DEVICE_FUNC
+    static const CwiseNullaryOp<CustomNullaryOp, PlainObject>
     NullaryExpr(const CustomNullaryOp& func);
 
-    static const ConstantReturnType Zero(Index rows, Index cols);
-    static const ConstantReturnType Zero(Index size);
-    static const ConstantReturnType Zero();
-    static const ConstantReturnType Ones(Index rows, Index cols);
-    static const ConstantReturnType Ones(Index size);
-    static const ConstantReturnType Ones();
-
-    void fill(const Scalar& value);
-    Derived& setConstant(const Scalar& value);
-    Derived& setLinSpaced(Index size, const Scalar& low, const Scalar& high);
-    Derived& setLinSpaced(const Scalar& low, const Scalar& high);
-    Derived& setZero();
-    Derived& setOnes();
-    Derived& setRandom();
-
-    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC static const ConstantReturnType Zero(Index rows, Index cols);
+    EIGEN_DEVICE_FUNC static const ConstantReturnType Zero(Index size);
+    EIGEN_DEVICE_FUNC static const ConstantReturnType Zero();
+    EIGEN_DEVICE_FUNC static const ConstantReturnType Ones(Index rows, Index cols);
+    EIGEN_DEVICE_FUNC static const ConstantReturnType Ones(Index size);
+    EIGEN_DEVICE_FUNC static const ConstantReturnType Ones();
+
+    EIGEN_DEVICE_FUNC void fill(const Scalar& value);
+    EIGEN_DEVICE_FUNC Derived& setConstant(const Scalar& value);
+    EIGEN_DEVICE_FUNC Derived& setLinSpaced(Index size, const Scalar& low, const Scalar& high);
+    EIGEN_DEVICE_FUNC Derived& setLinSpaced(const Scalar& low, const Scalar& high);
+    EIGEN_DEVICE_FUNC Derived& setZero();
+    EIGEN_DEVICE_FUNC Derived& setOnes();
+    EIGEN_DEVICE_FUNC Derived& setRandom();
+
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC
     bool isApprox(const DenseBase<OtherDerived>& other,
                   const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
+    EIGEN_DEVICE_FUNC 
     bool isMuchSmallerThan(const RealScalar& other,
                            const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
-    template<typename OtherDerived>
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC
     bool isMuchSmallerThan(const DenseBase<OtherDerived>& other,
                            const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
 
-    bool isApproxToConstant(const Scalar& value, const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
-    bool isConstant(const Scalar& value, const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
-    bool isZero(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
-    bool isOnes(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
+    EIGEN_DEVICE_FUNC bool isApproxToConstant(const Scalar& value, const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
+    EIGEN_DEVICE_FUNC bool isConstant(const Scalar& value, const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
+    EIGEN_DEVICE_FUNC bool isZero(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
+    EIGEN_DEVICE_FUNC bool isOnes(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
     
     inline bool hasNaN() const;
     inline bool allFinite() const;
 
-    inline Derived& operator*=(const Scalar& other);
-    inline Derived& operator/=(const Scalar& other);
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Derived& operator*=(const Scalar& other);
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Derived& operator/=(const Scalar& other);
 
     typedef typename internal::add_const_on_value_type<typename internal::eval<Derived>::type>::type EvalReturnType;
     /** \returns the matrix or vector obtained by evaluating this expression.
       *
       * Notice that in the case of a plain matrix or vector (not an expression) this function just returns
       * a const reference, in order to avoid a useless copy.
+      * 
+      * \warning Be carefull with eval() and the auto C++ keyword, as detailed in this \link TopicPitfalls_auto_keyword page \endlink.
       */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE EvalReturnType eval() const
     {
       // Even though MSVC does not honor strong inlining when the return type
@@ -364,61 +405,78 @@ template<typename Derived> class DenseBase
       // size types on MSVC.
       return typename internal::eval<Derived>::type(derived());
     }
-
+    
     /** swaps *this with the expression \a other.
       *
       */
     template<typename OtherDerived>
-    void swap(const DenseBase<OtherDerived>& other,
-              int = OtherDerived::ThisConstantIsPrivateInPlainObjectBase)
+    EIGEN_DEVICE_FUNC
+    void swap(const DenseBase<OtherDerived>& other)
     {
-      SwapWrapper<Derived>(derived()).lazyAssign(other.derived());
+      EIGEN_STATIC_ASSERT(!OtherDerived::IsPlainObjectBase,THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY);
+      eigen_assert(rows()==other.rows() && cols()==other.cols());
+      call_assignment(derived(), other.const_cast_derived(), internal::swap_assign_op<Scalar>());
     }
 
     /** swaps *this with the matrix or array \a other.
       *
       */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     void swap(PlainObjectBase<OtherDerived>& other)
     {
-      SwapWrapper<Derived>(derived()).lazyAssign(other.derived());
+      eigen_assert(rows()==other.rows() && cols()==other.cols());
+      call_assignment(derived(), other.derived(), internal::swap_assign_op<Scalar>());
     }
 
+    EIGEN_DEVICE_FUNC inline const NestByValue<Derived> nestByValue() const;
+    EIGEN_DEVICE_FUNC inline const ForceAlignedAccess<Derived> forceAlignedAccess() const;
+    EIGEN_DEVICE_FUNC inline ForceAlignedAccess<Derived> forceAlignedAccess();
+    template<bool Enable> EIGEN_DEVICE_FUNC
+    inline const typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type forceAlignedAccessIf() const;
+    template<bool Enable> EIGEN_DEVICE_FUNC
+    inline typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type forceAlignedAccessIf();
 
-    inline const NestByValue<Derived> nestByValue() const;
-    inline const ForceAlignedAccess<Derived> forceAlignedAccess() const;
-    inline ForceAlignedAccess<Derived> forceAlignedAccess();
-    template<bool Enable> inline const typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type forceAlignedAccessIf() const;
-    template<bool Enable> inline typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type forceAlignedAccessIf();
-
-    Scalar sum() const;
-    Scalar mean() const;
-    Scalar trace() const;
+    EIGEN_DEVICE_FUNC Scalar sum() const;
+    EIGEN_DEVICE_FUNC Scalar mean() const;
+    EIGEN_DEVICE_FUNC Scalar trace() const;
 
-    Scalar prod() const;
+    EIGEN_DEVICE_FUNC Scalar prod() const;
 
-    typename internal::traits<Derived>::Scalar minCoeff() const;
-    typename internal::traits<Derived>::Scalar maxCoeff() const;
+    EIGEN_DEVICE_FUNC typename internal::traits<Derived>::Scalar minCoeff() const;
+    EIGEN_DEVICE_FUNC typename internal::traits<Derived>::Scalar maxCoeff() const;
 
-    template<typename IndexType>
+    template<typename IndexType> EIGEN_DEVICE_FUNC
     typename internal::traits<Derived>::Scalar minCoeff(IndexType* row, IndexType* col) const;
-    template<typename IndexType>
+    template<typename IndexType> EIGEN_DEVICE_FUNC
     typename internal::traits<Derived>::Scalar maxCoeff(IndexType* row, IndexType* col) const;
-    template<typename IndexType>
+    template<typename IndexType> EIGEN_DEVICE_FUNC
     typename internal::traits<Derived>::Scalar minCoeff(IndexType* index) const;
-    template<typename IndexType>
+    template<typename IndexType> EIGEN_DEVICE_FUNC
     typename internal::traits<Derived>::Scalar maxCoeff(IndexType* index) const;
 
     template<typename BinaryOp>
-    typename internal::result_of<BinaryOp(typename internal::traits<Derived>::Scalar)>::type
-    redux(const BinaryOp& func) const;
+    EIGEN_DEVICE_FUNC
+    Scalar redux(const BinaryOp& func) const;
 
     template<typename Visitor>
+    EIGEN_DEVICE_FUNC
     void visit(Visitor& func) const;
 
-    inline const WithFormat<Derived> format(const IOFormat& fmt) const;
+    /** \returns a WithFormat proxy object allowing to print a matrix the with given
+      * format \a fmt.
+      *
+      * See class IOFormat for some examples.
+      *
+      * \sa class IOFormat, class WithFormat
+      */
+    inline const WithFormat<Derived> format(const IOFormat& fmt) const
+    {
+      return WithFormat<Derived>(derived(), fmt);
+    }
 
     /** \returns the unique coefficient of a 1x1 expression */
+    EIGEN_DEVICE_FUNC
     CoeffReturnType value() const
     {
       EIGEN_STATIC_ASSERT_SIZE_1x1(Derived)
@@ -426,23 +484,44 @@ template<typename Derived> class DenseBase
       return derived().coeff(0,0);
     }
 
-    bool all(void) const;
-    bool any(void) const;
-    Index count() const;
+    EIGEN_DEVICE_FUNC bool all() const;
+    EIGEN_DEVICE_FUNC bool any() const;
+    EIGEN_DEVICE_FUNC Index count() const;
 
     typedef VectorwiseOp<Derived, Horizontal> RowwiseReturnType;
     typedef const VectorwiseOp<const Derived, Horizontal> ConstRowwiseReturnType;
     typedef VectorwiseOp<Derived, Vertical> ColwiseReturnType;
     typedef const VectorwiseOp<const Derived, Vertical> ConstColwiseReturnType;
 
-    ConstRowwiseReturnType rowwise() const;
-    RowwiseReturnType rowwise();
-    ConstColwiseReturnType colwise() const;
-    ColwiseReturnType colwise();
+    /** \returns a VectorwiseOp wrapper of *this providing additional partial reduction operations
+    *
+    * Example: \include MatrixBase_rowwise.cpp
+    * Output: \verbinclude MatrixBase_rowwise.out
+    *
+    * \sa colwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
+    */
+    //Code moved here due to a CUDA compiler bug
+    EIGEN_DEVICE_FUNC inline ConstRowwiseReturnType rowwise() const {
+      return ConstRowwiseReturnType(derived());
+    }
+    EIGEN_DEVICE_FUNC RowwiseReturnType rowwise();
+
+    /** \returns a VectorwiseOp wrapper of *this providing additional partial reduction operations
+    *
+    * Example: \include MatrixBase_colwise.cpp
+    * Output: \verbinclude MatrixBase_colwise.out
+    *
+    * \sa rowwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
+    */
+    EIGEN_DEVICE_FUNC inline ConstColwiseReturnType colwise() const {
+      return ConstColwiseReturnType(derived());
+    }
+    EIGEN_DEVICE_FUNC ColwiseReturnType colwise();
 
-    static const CwiseNullaryOp<internal::scalar_random_op<Scalar>,Derived> Random(Index rows, Index cols);
-    static const CwiseNullaryOp<internal::scalar_random_op<Scalar>,Derived> Random(Index size);
-    static const CwiseNullaryOp<internal::scalar_random_op<Scalar>,Derived> Random();
+    typedef CwiseNullaryOp<internal::scalar_random_op<Scalar>,PlainObject> RandomReturnType;
+    static const RandomReturnType Random(Index rows, Index cols);
+    static const RandomReturnType Random(Index size);
+    static const RandomReturnType Random();
 
     template<typename ThenDerived,typename ElseDerived>
     const Select<Derived,ThenDerived,ElseDerived>
@@ -460,45 +539,60 @@ template<typename Derived> class DenseBase
     template<int p> RealScalar lpNorm() const;
 
     template<int RowFactor, int ColFactor>
-    inline const Replicate<Derived,RowFactor,ColFactor> replicate() const;
-    
-    typedef Replicate<Derived,Dynamic,Dynamic> ReplicateReturnType;
-    inline const ReplicateReturnType replicate(Index rowFacor,Index colFactor) const;
+    EIGEN_DEVICE_FUNC
+    const Replicate<Derived,RowFactor,ColFactor> replicate() const;
+    /**
+    * \return an expression of the replication of \c *this
+    *
+    * Example: \include MatrixBase_replicate_int_int.cpp
+    * Output: \verbinclude MatrixBase_replicate_int_int.out
+    *
+    * \sa VectorwiseOp::replicate(), DenseBase::replicate<int,int>(), class Replicate
+    */
+    //Code moved here due to a CUDA compiler bug
+    EIGEN_DEVICE_FUNC
+    const Replicate<Derived, Dynamic, Dynamic> replicate(Index rowFactor, Index colFactor) const
+    {
+      return Replicate<Derived, Dynamic, Dynamic>(derived(), rowFactor, colFactor);
+    }
 
     typedef Reverse<Derived, BothDirections> ReverseReturnType;
     typedef const Reverse<const Derived, BothDirections> ConstReverseReturnType;
-    ReverseReturnType reverse();
-    ConstReverseReturnType reverse() const;
-    void reverseInPlace();
+    EIGEN_DEVICE_FUNC ReverseReturnType reverse();
+    /** This is the const version of reverse(). */
+    //Code moved here due to a CUDA compiler bug
+    EIGEN_DEVICE_FUNC ConstReverseReturnType reverse() const
+    {
+      return ConstReverseReturnType(derived());
+    }
+    EIGEN_DEVICE_FUNC void reverseInPlace();
 
 #define EIGEN_CURRENT_STORAGE_BASE_CLASS Eigen::DenseBase
+#define EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL
+#define EIGEN_DOC_BLOCK_ADDONS_INNER_PANEL_IF(COND)
+#define EIGEN_DOC_UNARY_ADDONS(X,Y)
+#   include "../plugins/CommonCwiseUnaryOps.h"
 #   include "../plugins/BlockMethods.h"
+#   include "../plugins/IndexedViewMethods.h"
 #   ifdef EIGEN_DENSEBASE_PLUGIN
 #     include EIGEN_DENSEBASE_PLUGIN
 #   endif
 #undef EIGEN_CURRENT_STORAGE_BASE_CLASS
-
-#ifdef EIGEN2_SUPPORT
-
-    Block<Derived> corner(CornerType type, Index cRows, Index cCols);
-    const Block<Derived> corner(CornerType type, Index cRows, Index cCols) const;
-    template<int CRows, int CCols>
-    Block<Derived, CRows, CCols> corner(CornerType type);
-    template<int CRows, int CCols>
-    const Block<Derived, CRows, CCols> corner(CornerType type) const;
-
-#endif // EIGEN2_SUPPORT
-
+#undef EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL
+#undef EIGEN_DOC_BLOCK_ADDONS_INNER_PANEL_IF
+#undef EIGEN_DOC_UNARY_ADDONS
 
     // disable the use of evalTo for dense objects with a nice compilation error
-    template<typename Dest> inline void evalTo(Dest& ) const
+    template<typename Dest>
+    EIGEN_DEVICE_FUNC
+    inline void evalTo(Dest& ) const
     {
       EIGEN_STATIC_ASSERT((internal::is_same<Dest,void>::value),THE_EVAL_EVALTO_FUNCTION_SHOULD_NEVER_BE_CALLED_FOR_DENSE_OBJECTS);
     }
 
   protected:
     /** Default constructor. Do nothing. */
-    DenseBase()
+    EIGEN_DEVICE_FUNC DenseBase()
     {
       /* Just checks for self-consistency of the flags.
        * Only do it when debugging Eigen, as this borders on paranoiac and could slow compilation down
@@ -511,9 +605,9 @@ template<typename Derived> class DenseBase
     }
 
   private:
-    explicit DenseBase(int);
-    DenseBase(int,int);
-    template<typename OtherDerived> explicit DenseBase(const DenseBase<OtherDerived>&);
+    EIGEN_DEVICE_FUNC explicit DenseBase(int);
+    EIGEN_DEVICE_FUNC DenseBase(int,int);
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC explicit DenseBase(const DenseBase<OtherDerived>&);
 };
 
 } // end namespace Eigen
diff --git a/eigen/Eigen/src/Core/DenseCoeffsBase.h b/eigen/Eigen/src/Core/DenseCoeffsBase.h
index 3c890f2..c4af48a 100644
--- a/eigen/Eigen/src/Core/DenseCoeffsBase.h
+++ b/eigen/Eigen/src/Core/DenseCoeffsBase.h
@@ -35,7 +35,6 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
 {
   public:
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
-    typedef typename internal::traits<Derived>::Index Index;
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef typename internal::packet_traits<Scalar>::type PacketScalar;
 
@@ -61,6 +60,7 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
     using Base::size;
     using Base::derived;
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Index rowIndexByOuterInner(Index outer, Index inner) const
     {
       return int(Derived::RowsAtCompileTime) == 1 ? 0
@@ -69,6 +69,7 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
           : inner;
     }
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Index colIndexByOuterInner(Index outer, Index inner) const
     {
       return int(Derived::ColsAtCompileTime) == 1 ? 0
@@ -91,13 +92,15 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
       *
       * \sa operator()(Index,Index) const, coeffRef(Index,Index), coeff(Index) const
       */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE CoeffReturnType coeff(Index row, Index col) const
     {
       eigen_internal_assert(row >= 0 && row < rows()
-                        && col >= 0 && col < cols());
-      return derived().coeff(row, col);
+                         && col >= 0 && col < cols());
+      return internal::evaluator<Derived>(derived()).coeff(row,col);
     }
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE CoeffReturnType coeffByOuterInner(Index outer, Index inner) const
     {
       return coeff(rowIndexByOuterInner(outer, inner),
@@ -108,11 +111,12 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
       *
       * \sa operator()(Index,Index), operator[](Index)
       */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE CoeffReturnType operator()(Index row, Index col) const
     {
       eigen_assert(row >= 0 && row < rows()
           && col >= 0 && col < cols());
-      return derived().coeff(row, col);
+      return coeff(row, col);
     }
 
     /** Short version: don't use this function, use
@@ -130,11 +134,14 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
       * \sa operator[](Index) const, coeffRef(Index), coeff(Index,Index) const
       */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE CoeffReturnType
     coeff(Index index) const
     {
+      EIGEN_STATIC_ASSERT(internal::evaluator<Derived>::Flags & LinearAccessBit,
+                          THIS_COEFFICIENT_ACCESSOR_TAKING_ONE_ACCESS_IS_ONLY_FOR_EXPRESSIONS_ALLOWING_LINEAR_ACCESS)
       eigen_internal_assert(index >= 0 && index < size());
-      return derived().coeff(index);
+      return internal::evaluator<Derived>(derived()).coeff(index);
     }
 
 
@@ -146,15 +153,14 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
       * z() const, w() const
       */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE CoeffReturnType
     operator[](Index index) const
     {
-      #ifndef EIGEN2_SUPPORT
       EIGEN_STATIC_ASSERT(Derived::IsVectorAtCompileTime,
                           THE_BRACKET_OPERATOR_IS_ONLY_FOR_VECTORS__USE_THE_PARENTHESIS_OPERATOR_INSTEAD)
-      #endif
       eigen_assert(index >= 0 && index < size());
-      return derived().coeff(index);
+      return coeff(index);
     }
 
     /** \returns the coefficient at given index.
@@ -167,32 +173,49 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
       * z() const, w() const
       */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE CoeffReturnType
     operator()(Index index) const
     {
       eigen_assert(index >= 0 && index < size());
-      return derived().coeff(index);
+      return coeff(index);
     }
 
     /** equivalent to operator[](0).  */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE CoeffReturnType
     x() const { return (*this)[0]; }
 
     /** equivalent to operator[](1).  */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE CoeffReturnType
-    y() const { return (*this)[1]; }
+    y() const
+    {
+      EIGEN_STATIC_ASSERT(Derived::SizeAtCompileTime==-1 || Derived::SizeAtCompileTime>=2, OUT_OF_RANGE_ACCESS);
+      return (*this)[1];
+    }
 
     /** equivalent to operator[](2).  */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE CoeffReturnType
-    z() const { return (*this)[2]; }
+    z() const
+    {
+      EIGEN_STATIC_ASSERT(Derived::SizeAtCompileTime==-1 || Derived::SizeAtCompileTime>=3, OUT_OF_RANGE_ACCESS);
+      return (*this)[2];
+    }
 
     /** equivalent to operator[](3).  */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE CoeffReturnType
-    w() const { return (*this)[3]; }
+    w() const
+    {
+      EIGEN_STATIC_ASSERT(Derived::SizeAtCompileTime==-1 || Derived::SizeAtCompileTime>=4, OUT_OF_RANGE_ACCESS);
+      return (*this)[3];
+    }
 
     /** \internal
       * \returns the packet of coefficients starting at the given row and column. It is your responsibility
@@ -207,9 +230,9 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
     template<int LoadMode>
     EIGEN_STRONG_INLINE PacketReturnType packet(Index row, Index col) const
     {
-      eigen_internal_assert(row >= 0 && row < rows()
-                      && col >= 0 && col < cols());
-      return derived().template packet<LoadMode>(row,col);
+      typedef typename internal::packet_traits<Scalar>::type DefaultPacketType;
+      eigen_internal_assert(row >= 0 && row < rows() && col >= 0 && col < cols());
+      return internal::evaluator<Derived>(derived()).template packet<LoadMode,DefaultPacketType>(row,col);
     }
 
 
@@ -234,8 +257,11 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
     template<int LoadMode>
     EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
     {
+      EIGEN_STATIC_ASSERT(internal::evaluator<Derived>::Flags & LinearAccessBit,
+                          THIS_COEFFICIENT_ACCESSOR_TAKING_ONE_ACCESS_IS_ONLY_FOR_EXPRESSIONS_ALLOWING_LINEAR_ACCESS)
+      typedef typename internal::packet_traits<Scalar>::type DefaultPacketType;
       eigen_internal_assert(index >= 0 && index < size());
-      return derived().template packet<LoadMode>(index);
+      return internal::evaluator<Derived>(derived()).template packet<LoadMode,DefaultPacketType>(index);
     }
 
   protected:
@@ -278,7 +304,6 @@ class DenseCoeffsBase<Derived, WriteAccessors> : public DenseCoeffsBase<Derived,
     typedef DenseCoeffsBase<Derived, ReadOnlyAccessors> Base;
 
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
-    typedef typename internal::traits<Derived>::Index Index;
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef typename internal::packet_traits<Scalar>::type PacketScalar;
     typedef typename NumTraits<Scalar>::Real RealScalar;
@@ -311,13 +336,15 @@ class DenseCoeffsBase<Derived, WriteAccessors> : public DenseCoeffsBase<Derived,
       *
       * \sa operator()(Index,Index), coeff(Index, Index) const, coeffRef(Index)
       */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Scalar& coeffRef(Index row, Index col)
     {
       eigen_internal_assert(row >= 0 && row < rows()
-                        && col >= 0 && col < cols());
-      return derived().coeffRef(row, col);
+                         && col >= 0 && col < cols());
+      return internal::evaluator<Derived>(derived()).coeffRef(row,col);
     }
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Scalar&
     coeffRefByOuterInner(Index outer, Index inner)
     {
@@ -330,12 +357,13 @@ class DenseCoeffsBase<Derived, WriteAccessors> : public DenseCoeffsBase<Derived,
       * \sa operator[](Index)
       */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Scalar&
     operator()(Index row, Index col)
     {
       eigen_assert(row >= 0 && row < rows()
           && col >= 0 && col < cols());
-      return derived().coeffRef(row, col);
+      return coeffRef(row, col);
     }
 
 
@@ -354,11 +382,14 @@ class DenseCoeffsBase<Derived, WriteAccessors> : public DenseCoeffsBase<Derived,
       * \sa operator[](Index), coeff(Index) const, coeffRef(Index,Index)
       */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Scalar&
     coeffRef(Index index)
     {
+      EIGEN_STATIC_ASSERT(internal::evaluator<Derived>::Flags & LinearAccessBit,
+                          THIS_COEFFICIENT_ACCESSOR_TAKING_ONE_ACCESS_IS_ONLY_FOR_EXPRESSIONS_ALLOWING_LINEAR_ACCESS)
       eigen_internal_assert(index >= 0 && index < size());
-      return derived().coeffRef(index);
+      return internal::evaluator<Derived>(derived()).coeffRef(index);
     }
 
     /** \returns a reference to the coefficient at given index.
@@ -368,15 +399,14 @@ class DenseCoeffsBase<Derived, WriteAccessors> : public DenseCoeffsBase<Derived,
       * \sa operator[](Index) const, operator()(Index,Index), x(), y(), z(), w()
       */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Scalar&
     operator[](Index index)
     {
-      #ifndef EIGEN2_SUPPORT
       EIGEN_STATIC_ASSERT(Derived::IsVectorAtCompileTime,
                           THE_BRACKET_OPERATOR_IS_ONLY_FOR_VECTORS__USE_THE_PARENTHESIS_OPERATOR_INSTEAD)
-      #endif
       eigen_assert(index >= 0 && index < size());
-      return derived().coeffRef(index);
+      return coeffRef(index);
     }
 
     /** \returns a reference to the coefficient at given index.
@@ -388,167 +418,49 @@ class DenseCoeffsBase<Derived, WriteAccessors> : public DenseCoeffsBase<Derived,
       * \sa operator[](Index) const, operator()(Index,Index), x(), y(), z(), w()
       */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Scalar&
     operator()(Index index)
     {
       eigen_assert(index >= 0 && index < size());
-      return derived().coeffRef(index);
+      return coeffRef(index);
     }
 
     /** equivalent to operator[](0).  */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Scalar&
     x() { return (*this)[0]; }
 
     /** equivalent to operator[](1).  */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Scalar&
-    y() { return (*this)[1]; }
-
-    /** equivalent to operator[](2).  */
-
-    EIGEN_STRONG_INLINE Scalar&
-    z() { return (*this)[2]; }
-
-    /** equivalent to operator[](3).  */
-
-    EIGEN_STRONG_INLINE Scalar&
-    w() { return (*this)[3]; }
-
-    /** \internal
-      * Stores the given packet of coefficients, at the given row and column of this expression. It is your responsibility
-      * to ensure that a packet really starts there. This method is only available on expressions having the
-      * PacketAccessBit.
-      *
-      * The \a LoadMode parameter may have the value \a #Aligned or \a #Unaligned. Its effect is to select
-      * the appropriate vectorization instruction. Aligned access is faster, but is only possible for packets
-      * starting at an address which is a multiple of the packet size.
-      */
-
-    template<int StoreMode>
-    EIGEN_STRONG_INLINE void writePacket
-    (Index row, Index col, const typename internal::packet_traits<Scalar>::type& val)
+    y()
     {
-      eigen_internal_assert(row >= 0 && row < rows()
-                        && col >= 0 && col < cols());
-      derived().template writePacket<StoreMode>(row,col,val);
+      EIGEN_STATIC_ASSERT(Derived::SizeAtCompileTime==-1 || Derived::SizeAtCompileTime>=2, OUT_OF_RANGE_ACCESS);
+      return (*this)[1];
     }
 
+    /** equivalent to operator[](2).  */
 
-    /** \internal */
-    template<int StoreMode>
-    EIGEN_STRONG_INLINE void writePacketByOuterInner
-    (Index outer, Index inner, const typename internal::packet_traits<Scalar>::type& val)
-    {
-      writePacket<StoreMode>(rowIndexByOuterInner(outer, inner),
-                            colIndexByOuterInner(outer, inner),
-                            val);
-    }
-
-    /** \internal
-      * Stores the given packet of coefficients, at the given index in this expression. It is your responsibility
-      * to ensure that a packet really starts there. This method is only available on expressions having the
-      * PacketAccessBit and the LinearAccessBit.
-      *
-      * The \a LoadMode parameter may have the value \a Aligned or \a Unaligned. Its effect is to select
-      * the appropriate vectorization instruction. Aligned access is faster, but is only possible for packets
-      * starting at an address which is a multiple of the packet size.
-      */
-    template<int StoreMode>
-    EIGEN_STRONG_INLINE void writePacket
-    (Index index, const typename internal::packet_traits<Scalar>::type& val)
-    {
-      eigen_internal_assert(index >= 0 && index < size());
-      derived().template writePacket<StoreMode>(index,val);
-    }
-
-#ifndef EIGEN_PARSED_BY_DOXYGEN
-
-    /** \internal Copies the coefficient at position (row,col) of other into *this.
-      *
-      * This method is overridden in SwapWrapper, allowing swap() assignments to share 99% of their code
-      * with usual assignments.
-      *
-      * Outside of this internal usage, this method has probably no usefulness. It is hidden in the public API dox.
-      */
-
-    template<typename OtherDerived>
-    EIGEN_STRONG_INLINE void copyCoeff(Index row, Index col, const DenseBase<OtherDerived>& other)
-    {
-      eigen_internal_assert(row >= 0 && row < rows()
-                        && col >= 0 && col < cols());
-      derived().coeffRef(row, col) = other.derived().coeff(row, col);
-    }
-
-    /** \internal Copies the coefficient at the given index of other into *this.
-      *
-      * This method is overridden in SwapWrapper, allowing swap() assignments to share 99% of their code
-      * with usual assignments.
-      *
-      * Outside of this internal usage, this method has probably no usefulness. It is hidden in the public API dox.
-      */
-
-    template<typename OtherDerived>
-    EIGEN_STRONG_INLINE void copyCoeff(Index index, const DenseBase<OtherDerived>& other)
-    {
-      eigen_internal_assert(index >= 0 && index < size());
-      derived().coeffRef(index) = other.derived().coeff(index);
-    }
-
-
-    template<typename OtherDerived>
-    EIGEN_STRONG_INLINE void copyCoeffByOuterInner(Index outer, Index inner, const DenseBase<OtherDerived>& other)
-    {
-      const Index row = rowIndexByOuterInner(outer,inner);
-      const Index col = colIndexByOuterInner(outer,inner);
-      // derived() is important here: copyCoeff() may be reimplemented in Derived!
-      derived().copyCoeff(row, col, other);
-    }
-
-    /** \internal Copies the packet at position (row,col) of other into *this.
-      *
-      * This method is overridden in SwapWrapper, allowing swap() assignments to share 99% of their code
-      * with usual assignments.
-      *
-      * Outside of this internal usage, this method has probably no usefulness. It is hidden in the public API dox.
-      */
-
-    template<typename OtherDerived, int StoreMode, int LoadMode>
-    EIGEN_STRONG_INLINE void copyPacket(Index row, Index col, const DenseBase<OtherDerived>& other)
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar&
+    z()
     {
-      eigen_internal_assert(row >= 0 && row < rows()
-                        && col >= 0 && col < cols());
-      derived().template writePacket<StoreMode>(row, col,
-        other.derived().template packet<LoadMode>(row, col));
+      EIGEN_STATIC_ASSERT(Derived::SizeAtCompileTime==-1 || Derived::SizeAtCompileTime>=3, OUT_OF_RANGE_ACCESS);
+      return (*this)[2];
     }
 
-    /** \internal Copies the packet at the given index of other into *this.
-      *
-      * This method is overridden in SwapWrapper, allowing swap() assignments to share 99% of their code
-      * with usual assignments.
-      *
-      * Outside of this internal usage, this method has probably no usefulness. It is hidden in the public API dox.
-      */
-
-    template<typename OtherDerived, int StoreMode, int LoadMode>
-    EIGEN_STRONG_INLINE void copyPacket(Index index, const DenseBase<OtherDerived>& other)
-    {
-      eigen_internal_assert(index >= 0 && index < size());
-      derived().template writePacket<StoreMode>(index,
-        other.derived().template packet<LoadMode>(index));
-    }
+    /** equivalent to operator[](3).  */
 
-    /** \internal */
-    template<typename OtherDerived, int StoreMode, int LoadMode>
-    EIGEN_STRONG_INLINE void copyPacketByOuterInner(Index outer, Index inner, const DenseBase<OtherDerived>& other)
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar&
+    w()
     {
-      const Index row = rowIndexByOuterInner(outer,inner);
-      const Index col = colIndexByOuterInner(outer,inner);
-      // derived() is important here: copyCoeff() may be reimplemented in Derived!
-      derived().template copyPacket< OtherDerived, StoreMode, LoadMode>(row, col, other);
+      EIGEN_STATIC_ASSERT(Derived::SizeAtCompileTime==-1 || Derived::SizeAtCompileTime>=4, OUT_OF_RANGE_ACCESS);
+      return (*this)[3];
     }
-#endif
-
 };
 
 /** \brief Base class providing direct read-only coefficient access to matrices and arrays.
@@ -560,7 +472,7 @@ class DenseCoeffsBase<Derived, WriteAccessors> : public DenseCoeffsBase<Derived,
   * inherits DenseCoeffsBase<Derived, ReadOnlyAccessors> which defines functions to access entries read-only using
   * \c operator() .
   *
-  * \sa \ref TopicClassHierarchy
+  * \sa \blank \ref TopicClassHierarchy
   */
 template<typename Derived>
 class DenseCoeffsBase<Derived, DirectAccessors> : public DenseCoeffsBase<Derived, ReadOnlyAccessors>
@@ -568,7 +480,6 @@ class DenseCoeffsBase<Derived, DirectAccessors> : public DenseCoeffsBase<Derived
   public:
 
     typedef DenseCoeffsBase<Derived, ReadOnlyAccessors> Base;
-    typedef typename internal::traits<Derived>::Index Index;
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef typename NumTraits<Scalar>::Real RealScalar;
 
@@ -581,6 +492,7 @@ class DenseCoeffsBase<Derived, DirectAccessors> : public DenseCoeffsBase<Derived
       *
       * \sa outerStride(), rowStride(), colStride()
       */
+    EIGEN_DEVICE_FUNC
     inline Index innerStride() const
     {
       return derived().innerStride();
@@ -591,6 +503,7 @@ class DenseCoeffsBase<Derived, DirectAccessors> : public DenseCoeffsBase<Derived
       *
       * \sa innerStride(), rowStride(), colStride()
       */
+    EIGEN_DEVICE_FUNC
     inline Index outerStride() const
     {
       return derived().outerStride();
@@ -606,6 +519,7 @@ class DenseCoeffsBase<Derived, DirectAccessors> : public DenseCoeffsBase<Derived
       *
       * \sa innerStride(), outerStride(), colStride()
       */
+    EIGEN_DEVICE_FUNC
     inline Index rowStride() const
     {
       return Derived::IsRowMajor ? outerStride() : innerStride();
@@ -615,6 +529,7 @@ class DenseCoeffsBase<Derived, DirectAccessors> : public DenseCoeffsBase<Derived
       *
       * \sa innerStride(), outerStride(), rowStride()
       */
+    EIGEN_DEVICE_FUNC
     inline Index colStride() const
     {
       return Derived::IsRowMajor ? innerStride() : outerStride();
@@ -630,7 +545,7 @@ class DenseCoeffsBase<Derived, DirectAccessors> : public DenseCoeffsBase<Derived
   * inherits DenseCoeffsBase<Derived, WriteAccessors> which defines functions to access entries read/write using
   * \c operator().
   *
-  * \sa \ref TopicClassHierarchy
+  * \sa \blank \ref TopicClassHierarchy
   */
 template<typename Derived>
 class DenseCoeffsBase<Derived, DirectWriteAccessors>
@@ -639,7 +554,6 @@ class DenseCoeffsBase<Derived, DirectWriteAccessors>
   public:
 
     typedef DenseCoeffsBase<Derived, WriteAccessors> Base;
-    typedef typename internal::traits<Derived>::Index Index;
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef typename NumTraits<Scalar>::Real RealScalar;
 
@@ -652,6 +566,7 @@ class DenseCoeffsBase<Derived, DirectWriteAccessors>
       *
       * \sa outerStride(), rowStride(), colStride()
       */
+    EIGEN_DEVICE_FUNC
     inline Index innerStride() const
     {
       return derived().innerStride();
@@ -662,6 +577,7 @@ class DenseCoeffsBase<Derived, DirectWriteAccessors>
       *
       * \sa innerStride(), rowStride(), colStride()
       */
+    EIGEN_DEVICE_FUNC
     inline Index outerStride() const
     {
       return derived().outerStride();
@@ -677,6 +593,7 @@ class DenseCoeffsBase<Derived, DirectWriteAccessors>
       *
       * \sa innerStride(), outerStride(), colStride()
       */
+    EIGEN_DEVICE_FUNC
     inline Index rowStride() const
     {
       return Derived::IsRowMajor ? outerStride() : innerStride();
@@ -686,6 +603,7 @@ class DenseCoeffsBase<Derived, DirectWriteAccessors>
       *
       * \sa innerStride(), outerStride(), rowStride()
       */
+    EIGEN_DEVICE_FUNC
     inline Index colStride() const
     {
       return Derived::IsRowMajor ? innerStride() : outerStride();
@@ -694,33 +612,42 @@ class DenseCoeffsBase<Derived, DirectWriteAccessors>
 
 namespace internal {
 
-template<typename Derived, bool JustReturnZero>
+template<int Alignment, typename Derived, bool JustReturnZero>
 struct first_aligned_impl
 {
-  static inline typename Derived::Index run(const Derived&)
+  static inline Index run(const Derived&)
   { return 0; }
 };
 
-template<typename Derived>
-struct first_aligned_impl<Derived, false>
+template<int Alignment, typename Derived>
+struct first_aligned_impl<Alignment, Derived, false>
 {
-  static inline typename Derived::Index run(const Derived& m)
+  static inline Index run(const Derived& m)
   {
-    return internal::first_aligned(&m.const_cast_derived().coeffRef(0,0), m.size());
+    return internal::first_aligned<Alignment>(m.data(), m.size());
   }
 };
 
-/** \internal \returns the index of the first element of the array that is well aligned for vectorization.
+/** \internal \returns the index of the first element of the array stored by \a m that is properly aligned with respect to \a Alignment for vectorization.
+  *
+  * \tparam Alignment requested alignment in Bytes.
   *
   * There is also the variant first_aligned(const Scalar*, Integer) defined in Memory.h. See it for more
   * documentation.
   */
+template<int Alignment, typename Derived>
+static inline Index first_aligned(const DenseBase<Derived>& m)
+{
+  enum { ReturnZero = (int(evaluator<Derived>::Alignment) >= Alignment) || !(Derived::Flags & DirectAccessBit) };
+  return first_aligned_impl<Alignment, Derived, ReturnZero>::run(m.derived());
+}
+
 template<typename Derived>
-static inline typename Derived::Index first_aligned(const Derived& m)
+static inline Index first_default_aligned(const DenseBase<Derived>& m)
 {
-  return first_aligned_impl
-          <Derived, (Derived::Flags & AlignedBit) || !(Derived::Flags & DirectAccessBit)>
-          ::run(m);
+  typedef typename Derived::Scalar Scalar;
+  typedef typename packet_traits<Scalar>::type DefaultPacketType;
+  return internal::first_aligned<int(unpacket_traits<DefaultPacketType>::alignment),Derived>(m);
 }
 
 template<typename Derived, bool HasDirectAccess = has_direct_access<Derived>::ret>
diff --git a/eigen/Eigen/src/Core/DenseStorage.h b/eigen/Eigen/src/Core/DenseStorage.h
index 568493c..7958fee 100644
--- a/eigen/Eigen/src/Core/DenseStorage.h
+++ b/eigen/Eigen/src/Core/DenseStorage.h
@@ -3,7 +3,7 @@
 //
 // Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
 // Copyright (C) 2006-2009 Benoit Jacob <jacob.benoit.1@gmail.com>
-// Copyright (C) 2010 Hauke Heibel <hauke.heibel@gmail.com>
+// Copyright (C) 2010-2013 Hauke Heibel <hauke.heibel@gmail.com>
 //
 // This Source Code Form is subject to the terms of the Mozilla
 // Public License v. 2.0. If a copy of the MPL was not distributed
@@ -13,9 +13,9 @@
 #define EIGEN_MATRIXSTORAGE_H
 
 #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
-  #define EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN EIGEN_DENSE_STORAGE_CTOR_PLUGIN;
+  #define EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(X) X; EIGEN_DENSE_STORAGE_CTOR_PLUGIN;
 #else
-  #define EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN
+  #define EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(X)
 #endif
 
 namespace Eigen {
@@ -24,7 +24,9 @@ namespace internal {
 
 struct constructor_without_unaligned_array_assert {};
 
-template<typename T, int Size> void check_static_allocation_size()
+template<typename T, int Size>
+EIGEN_DEVICE_FUNC
+void check_static_allocation_size()
 {
   // if EIGEN_STACK_ALLOCATION_LIMIT is defined to 0, then no limit
   #if EIGEN_STACK_ALLOCATION_LIMIT
@@ -38,18 +40,19 @@ template<typename T, int Size> void check_static_allocation_size()
   */
 template <typename T, int Size, int MatrixOrArrayOptions,
           int Alignment = (MatrixOrArrayOptions&DontAlign) ? 0
-                        : (((Size*sizeof(T))%16)==0) ? 16
-                        : 0 >
+                        : compute_default_alignment<T,Size>::value >
 struct plain_array
 {
   T array[Size];
 
-  plain_array() 
+  EIGEN_DEVICE_FUNC
+  plain_array()
   { 
     check_static_allocation_size<T,Size>();
   }
 
-  plain_array(constructor_without_unaligned_array_assert) 
+  EIGEN_DEVICE_FUNC
+  plain_array(constructor_without_unaligned_array_assert)
   { 
     check_static_allocation_size<T,Size>();
   }
@@ -64,29 +67,88 @@ struct plain_array
   template<typename PtrType>
   EIGEN_ALWAYS_INLINE PtrType eigen_unaligned_array_assert_workaround_gcc47(PtrType array) { return array; }
   #define EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(sizemask) \
-    eigen_assert((reinterpret_cast<size_t>(eigen_unaligned_array_assert_workaround_gcc47(array)) & sizemask) == 0 \
+    eigen_assert((internal::UIntPtr(eigen_unaligned_array_assert_workaround_gcc47(array)) & (sizemask)) == 0 \
               && "this assertion is explained here: " \
               "http://eigen.tuxfamily.org/dox-devel/group__TopicUnalignedArrayAssert.html" \
               " **** READ THIS WEB PAGE !!! ****");
 #else
   #define EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(sizemask) \
-    eigen_assert((reinterpret_cast<size_t>(array) & sizemask) == 0 \
+    eigen_assert((internal::UIntPtr(array) & (sizemask)) == 0 \
               && "this assertion is explained here: " \
               "http://eigen.tuxfamily.org/dox-devel/group__TopicUnalignedArrayAssert.html" \
               " **** READ THIS WEB PAGE !!! ****");
 #endif
 
 template <typename T, int Size, int MatrixOrArrayOptions>
+struct plain_array<T, Size, MatrixOrArrayOptions, 8>
+{
+  EIGEN_ALIGN_TO_BOUNDARY(8) T array[Size];
+
+  EIGEN_DEVICE_FUNC
+  plain_array() 
+  {
+    EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(7);
+    check_static_allocation_size<T,Size>();
+  }
+
+  EIGEN_DEVICE_FUNC
+  plain_array(constructor_without_unaligned_array_assert) 
+  { 
+    check_static_allocation_size<T,Size>();
+  }
+};
+
+template <typename T, int Size, int MatrixOrArrayOptions>
 struct plain_array<T, Size, MatrixOrArrayOptions, 16>
 {
-  EIGEN_USER_ALIGN16 T array[Size];
+  EIGEN_ALIGN_TO_BOUNDARY(16) T array[Size];
 
+  EIGEN_DEVICE_FUNC
   plain_array() 
   { 
-    EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(0xf);
+    EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(15);
     check_static_allocation_size<T,Size>();
   }
 
+  EIGEN_DEVICE_FUNC
+  plain_array(constructor_without_unaligned_array_assert) 
+  { 
+    check_static_allocation_size<T,Size>();
+  }
+};
+
+template <typename T, int Size, int MatrixOrArrayOptions>
+struct plain_array<T, Size, MatrixOrArrayOptions, 32>
+{
+  EIGEN_ALIGN_TO_BOUNDARY(32) T array[Size];
+
+  EIGEN_DEVICE_FUNC
+  plain_array() 
+  {
+    EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(31);
+    check_static_allocation_size<T,Size>();
+  }
+
+  EIGEN_DEVICE_FUNC
+  plain_array(constructor_without_unaligned_array_assert) 
+  { 
+    check_static_allocation_size<T,Size>();
+  }
+};
+
+template <typename T, int Size, int MatrixOrArrayOptions>
+struct plain_array<T, Size, MatrixOrArrayOptions, 64>
+{
+  EIGEN_ALIGN_TO_BOUNDARY(64) T array[Size];
+
+  EIGEN_DEVICE_FUNC
+  plain_array() 
+  { 
+    EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(63);
+    check_static_allocation_size<T,Size>();
+  }
+
+  EIGEN_DEVICE_FUNC
   plain_array(constructor_without_unaligned_array_assert) 
   { 
     check_static_allocation_size<T,Size>();
@@ -96,9 +158,9 @@ struct plain_array<T, Size, MatrixOrArrayOptions, 16>
 template <typename T, int MatrixOrArrayOptions, int Alignment>
 struct plain_array<T, 0, MatrixOrArrayOptions, Alignment>
 {
-  EIGEN_USER_ALIGN16 T array[1];
-  plain_array() {}
-  plain_array(constructor_without_unaligned_array_assert) {}
+  T array[1];
+  EIGEN_DEVICE_FUNC plain_array() {}
+  EIGEN_DEVICE_FUNC plain_array(constructor_without_unaligned_array_assert) {}
 };
 
 } // end namespace internal
@@ -122,41 +184,54 @@ template<typename T, int Size, int _Rows, int _Cols, int _Options> class DenseSt
 {
     internal::plain_array<T,Size,_Options> m_data;
   public:
-    DenseStorage() {}
-    DenseStorage(internal::constructor_without_unaligned_array_assert)
+    EIGEN_DEVICE_FUNC DenseStorage() {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(Index size = Size)
+    }
+    EIGEN_DEVICE_FUNC
+    explicit DenseStorage(internal::constructor_without_unaligned_array_assert)
       : m_data(internal::constructor_without_unaligned_array_assert()) {}
-    DenseStorage(const DenseStorage& other) : m_data(other.m_data) {}
+    EIGEN_DEVICE_FUNC 
+    DenseStorage(const DenseStorage& other) : m_data(other.m_data) {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(Index size = Size)
+    }
+    EIGEN_DEVICE_FUNC 
     DenseStorage& operator=(const DenseStorage& other)
-    {
+    { 
       if (this != &other) m_data = other.m_data;
-      return *this;
+      return *this; 
     }
-    DenseStorage(DenseIndex,DenseIndex,DenseIndex) {}
-    void swap(DenseStorage& other) { std::swap(m_data,other.m_data); }
-    static DenseIndex rows(void) {return _Rows;}
-    static DenseIndex cols(void) {return _Cols;}
-    void conservativeResize(DenseIndex,DenseIndex,DenseIndex) {}
-    void resize(DenseIndex,DenseIndex,DenseIndex) {}
-    const T *data() const { return m_data.array; }
-    T *data() { return m_data.array; }
+    EIGEN_DEVICE_FUNC DenseStorage(Index size, Index rows, Index cols) {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({})
+      eigen_internal_assert(size==rows*cols && rows==_Rows && cols==_Cols);
+      EIGEN_UNUSED_VARIABLE(size);
+      EIGEN_UNUSED_VARIABLE(rows);
+      EIGEN_UNUSED_VARIABLE(cols);
+    }
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); }
+    EIGEN_DEVICE_FUNC static Index rows(void) {return _Rows;}
+    EIGEN_DEVICE_FUNC static Index cols(void) {return _Cols;}
+    EIGEN_DEVICE_FUNC void conservativeResize(Index,Index,Index) {}
+    EIGEN_DEVICE_FUNC void resize(Index,Index,Index) {}
+    EIGEN_DEVICE_FUNC const T *data() const { return m_data.array; }
+    EIGEN_DEVICE_FUNC T *data() { return m_data.array; }
 };
 
 // null matrix
 template<typename T, int _Rows, int _Cols, int _Options> class DenseStorage<T, 0, _Rows, _Cols, _Options>
 {
   public:
-    DenseStorage() {}
-    DenseStorage(internal::constructor_without_unaligned_array_assert) {}
-    DenseStorage(const DenseStorage&) {}
-    DenseStorage& operator=(const DenseStorage&) { return *this; }
-    DenseStorage(DenseIndex,DenseIndex,DenseIndex) {}
-    void swap(DenseStorage& ) {}
-    static DenseIndex rows(void) {return _Rows;}
-    static DenseIndex cols(void) {return _Cols;}
-    void conservativeResize(DenseIndex,DenseIndex,DenseIndex) {}
-    void resize(DenseIndex,DenseIndex,DenseIndex) {}
-    const T *data() const { return 0; }
-    T *data() { return 0; }
+    EIGEN_DEVICE_FUNC DenseStorage() {}
+    EIGEN_DEVICE_FUNC explicit DenseStorage(internal::constructor_without_unaligned_array_assert) {}
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage&) {}
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage&) { return *this; }
+    EIGEN_DEVICE_FUNC DenseStorage(Index,Index,Index) {}
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& ) {}
+    EIGEN_DEVICE_FUNC static Index rows(void) {return _Rows;}
+    EIGEN_DEVICE_FUNC static Index cols(void) {return _Cols;}
+    EIGEN_DEVICE_FUNC void conservativeResize(Index,Index,Index) {}
+    EIGEN_DEVICE_FUNC void resize(Index,Index,Index) {}
+    EIGEN_DEVICE_FUNC const T *data() const { return 0; }
+    EIGEN_DEVICE_FUNC T *data() { return 0; }
 };
 
 // more specializations for null matrices; these are necessary to resolve ambiguities
@@ -173,74 +248,74 @@ template<typename T, int _Cols, int _Options> class DenseStorage<T, 0, Dynamic,
 template<typename T, int Size, int _Options> class DenseStorage<T, Size, Dynamic, Dynamic, _Options>
 {
     internal::plain_array<T,Size,_Options> m_data;
-    DenseIndex m_rows;
-    DenseIndex m_cols;
+    Index m_rows;
+    Index m_cols;
   public:
-    DenseStorage() : m_rows(0), m_cols(0) {}
-    DenseStorage(internal::constructor_without_unaligned_array_assert)
+    EIGEN_DEVICE_FUNC DenseStorage() : m_rows(0), m_cols(0) {}
+    EIGEN_DEVICE_FUNC explicit DenseStorage(internal::constructor_without_unaligned_array_assert)
       : m_data(internal::constructor_without_unaligned_array_assert()), m_rows(0), m_cols(0) {}
-    DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_rows(other.m_rows), m_cols(other.m_cols) {}
-    DenseStorage& operator=(const DenseStorage& other)
-    {
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_rows(other.m_rows), m_cols(other.m_cols) {}
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other) 
+    { 
       if (this != &other)
       {
         m_data = other.m_data;
         m_rows = other.m_rows;
         m_cols = other.m_cols;
       }
-      return *this;
+      return *this; 
     }
-    DenseStorage(DenseIndex, DenseIndex nbRows, DenseIndex nbCols) : m_rows(nbRows), m_cols(nbCols) {}
-    void swap(DenseStorage& other)
+    EIGEN_DEVICE_FUNC DenseStorage(Index, Index rows, Index cols) : m_rows(rows), m_cols(cols) {}
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other)
     { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); std::swap(m_cols,other.m_cols); }
-    DenseIndex rows() const {return m_rows;}
-    DenseIndex cols() const {return m_cols;}
-    void conservativeResize(DenseIndex, DenseIndex nbRows, DenseIndex nbCols) { m_rows = nbRows; m_cols = nbCols; }
-    void resize(DenseIndex, DenseIndex nbRows, DenseIndex nbCols) { m_rows = nbRows; m_cols = nbCols; }
-    const T *data() const { return m_data.array; }
-    T *data() { return m_data.array; }
+    EIGEN_DEVICE_FUNC Index rows() const {return m_rows;}
+    EIGEN_DEVICE_FUNC Index cols() const {return m_cols;}
+    EIGEN_DEVICE_FUNC void conservativeResize(Index, Index rows, Index cols) { m_rows = rows; m_cols = cols; }
+    EIGEN_DEVICE_FUNC void resize(Index, Index rows, Index cols) { m_rows = rows; m_cols = cols; }
+    EIGEN_DEVICE_FUNC const T *data() const { return m_data.array; }
+    EIGEN_DEVICE_FUNC T *data() { return m_data.array; }
 };
 
 // dynamic-size matrix with fixed-size storage and fixed width
 template<typename T, int Size, int _Cols, int _Options> class DenseStorage<T, Size, Dynamic, _Cols, _Options>
 {
     internal::plain_array<T,Size,_Options> m_data;
-    DenseIndex m_rows;
+    Index m_rows;
   public:
-    DenseStorage() : m_rows(0) {}
-    DenseStorage(internal::constructor_without_unaligned_array_assert)
+    EIGEN_DEVICE_FUNC DenseStorage() : m_rows(0) {}
+    EIGEN_DEVICE_FUNC explicit DenseStorage(internal::constructor_without_unaligned_array_assert)
       : m_data(internal::constructor_without_unaligned_array_assert()), m_rows(0) {}
-    DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_rows(other.m_rows) {}
-    DenseStorage& operator=(const DenseStorage& other)
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_rows(other.m_rows) {}
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other) 
     {
       if (this != &other)
       {
         m_data = other.m_data;
         m_rows = other.m_rows;
       }
-      return *this;
+      return *this; 
     }
-    DenseStorage(DenseIndex, DenseIndex nbRows, DenseIndex) : m_rows(nbRows) {}
-    void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); }
-    DenseIndex rows(void) const {return m_rows;}
-    DenseIndex cols(void) const {return _Cols;}
-    void conservativeResize(DenseIndex, DenseIndex nbRows, DenseIndex) { m_rows = nbRows; }
-    void resize(DenseIndex, DenseIndex nbRows, DenseIndex) { m_rows = nbRows; }
-    const T *data() const { return m_data.array; }
-    T *data() { return m_data.array; }
+    EIGEN_DEVICE_FUNC DenseStorage(Index, Index rows, Index) : m_rows(rows) {}
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); }
+    EIGEN_DEVICE_FUNC Index rows(void) const {return m_rows;}
+    EIGEN_DEVICE_FUNC Index cols(void) const {return _Cols;}
+    EIGEN_DEVICE_FUNC void conservativeResize(Index, Index rows, Index) { m_rows = rows; }
+    EIGEN_DEVICE_FUNC void resize(Index, Index rows, Index) { m_rows = rows; }
+    EIGEN_DEVICE_FUNC const T *data() const { return m_data.array; }
+    EIGEN_DEVICE_FUNC T *data() { return m_data.array; }
 };
 
 // dynamic-size matrix with fixed-size storage and fixed height
 template<typename T, int Size, int _Rows, int _Options> class DenseStorage<T, Size, _Rows, Dynamic, _Options>
 {
     internal::plain_array<T,Size,_Options> m_data;
-    DenseIndex m_cols;
+    Index m_cols;
   public:
-    DenseStorage() : m_cols(0) {}
-    DenseStorage(internal::constructor_without_unaligned_array_assert)
+    EIGEN_DEVICE_FUNC DenseStorage() : m_cols(0) {}
+    EIGEN_DEVICE_FUNC explicit DenseStorage(internal::constructor_without_unaligned_array_assert)
       : m_data(internal::constructor_without_unaligned_array_assert()), m_cols(0) {}
-    DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_cols(other.m_cols) {}
-    DenseStorage& operator=(const DenseStorage& other)
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_cols(other.m_cols) {}
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other)
     {
       if (this != &other)
       {
@@ -249,38 +324,62 @@ template<typename T, int Size, int _Rows, int _Options> class DenseStorage<T, Si
       }
       return *this;
     }
-    DenseStorage(DenseIndex, DenseIndex, DenseIndex nbCols) : m_cols(nbCols) {}
-    void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); }
-    DenseIndex rows(void) const {return _Rows;}
-    DenseIndex cols(void) const {return m_cols;}
-    void conservativeResize(DenseIndex, DenseIndex, DenseIndex nbCols) { m_cols = nbCols; }
-    void resize(DenseIndex, DenseIndex, DenseIndex nbCols) { m_cols = nbCols; }
-    const T *data() const { return m_data.array; }
-    T *data() { return m_data.array; }
+    EIGEN_DEVICE_FUNC DenseStorage(Index, Index, Index cols) : m_cols(cols) {}
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); }
+    EIGEN_DEVICE_FUNC Index rows(void) const {return _Rows;}
+    EIGEN_DEVICE_FUNC Index cols(void) const {return m_cols;}
+    void conservativeResize(Index, Index, Index cols) { m_cols = cols; }
+    void resize(Index, Index, Index cols) { m_cols = cols; }
+    EIGEN_DEVICE_FUNC const T *data() const { return m_data.array; }
+    EIGEN_DEVICE_FUNC T *data() { return m_data.array; }
 };
 
 // purely dynamic matrix.
 template<typename T, int _Options> class DenseStorage<T, Dynamic, Dynamic, Dynamic, _Options>
 {
     T *m_data;
-    DenseIndex m_rows;
-    DenseIndex m_cols;
+    Index m_rows;
+    Index m_cols;
   public:
-    DenseStorage() : m_data(0), m_rows(0), m_cols(0) {}
-    DenseStorage(internal::constructor_without_unaligned_array_assert)
+    EIGEN_DEVICE_FUNC DenseStorage() : m_data(0), m_rows(0), m_cols(0) {}
+    EIGEN_DEVICE_FUNC explicit DenseStorage(internal::constructor_without_unaligned_array_assert)
        : m_data(0), m_rows(0), m_cols(0) {}
-    DenseStorage(DenseIndex size, DenseIndex nbRows, DenseIndex nbCols)
-      : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_rows(nbRows), m_cols(nbCols)
-    { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN }
-#ifdef EIGEN_HAVE_RVALUE_REFERENCES
-    DenseStorage(DenseStorage&& other)
+    EIGEN_DEVICE_FUNC DenseStorage(Index size, Index rows, Index cols)
+      : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_rows(rows), m_cols(cols)
+    {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({})
+      eigen_internal_assert(size==rows*cols && rows>=0 && cols >=0);
+    }
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other)
+      : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(other.m_rows*other.m_cols))
+      , m_rows(other.m_rows)
+      , m_cols(other.m_cols)
+    {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(Index size = m_rows*m_cols)
+      internal::smart_copy(other.m_data, other.m_data+other.m_rows*other.m_cols, m_data);
+    }
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other)
+    {
+      if (this != &other)
+      {
+        DenseStorage tmp(other);
+        this->swap(tmp);
+      }
+      return *this;
+    }
+#if EIGEN_HAS_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC
+    DenseStorage(DenseStorage&& other) EIGEN_NOEXCEPT
       : m_data(std::move(other.m_data))
       , m_rows(std::move(other.m_rows))
       , m_cols(std::move(other.m_cols))
     {
       other.m_data = nullptr;
+      other.m_rows = 0;
+      other.m_cols = 0;
     }
-    DenseStorage& operator=(DenseStorage&& other)
+    EIGEN_DEVICE_FUNC
+    DenseStorage& operator=(DenseStorage&& other) EIGEN_NOEXCEPT
     {
       using std::swap;
       swap(m_data, other.m_data);
@@ -289,18 +388,18 @@ template<typename T, int _Options> class DenseStorage<T, Dynamic, Dynamic, Dynam
       return *this;
     }
 #endif
-    ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, m_rows*m_cols); }
-    void swap(DenseStorage& other)
+    EIGEN_DEVICE_FUNC ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, m_rows*m_cols); }
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other)
     { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); std::swap(m_cols,other.m_cols); }
-    DenseIndex rows(void) const {return m_rows;}
-    DenseIndex cols(void) const {return m_cols;}
-    void conservativeResize(DenseIndex size, DenseIndex nbRows, DenseIndex nbCols)
+    EIGEN_DEVICE_FUNC Index rows(void) const {return m_rows;}
+    EIGEN_DEVICE_FUNC Index cols(void) const {return m_cols;}
+    void conservativeResize(Index size, Index rows, Index cols)
     {
       m_data = internal::conditional_aligned_realloc_new_auto<T,(_Options&DontAlign)==0>(m_data, size, m_rows*m_cols);
-      m_rows = nbRows;
-      m_cols = nbCols;
+      m_rows = rows;
+      m_cols = cols;
     }
-    void resize(DenseIndex size, DenseIndex nbRows, DenseIndex nbCols)
+    EIGEN_DEVICE_FUNC void resize(Index size, Index rows, Index cols)
     {
       if(size != m_rows*m_cols)
       {
@@ -309,36 +408,56 @@ template<typename T, int _Options> class DenseStorage<T, Dynamic, Dynamic, Dynam
           m_data = internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size);
         else
           m_data = 0;
-        EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN
+        EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({})
       }
-      m_rows = nbRows;
-      m_cols = nbCols;
+      m_rows = rows;
+      m_cols = cols;
     }
-    const T *data() const { return m_data; }
-    T *data() { return m_data; }
-  private:
-    DenseStorage(const DenseStorage&);
-    DenseStorage& operator=(const DenseStorage&);
+    EIGEN_DEVICE_FUNC const T *data() const { return m_data; }
+    EIGEN_DEVICE_FUNC T *data() { return m_data; }
 };
 
 // matrix with dynamic width and fixed height (so that matrix has dynamic size).
 template<typename T, int _Rows, int _Options> class DenseStorage<T, Dynamic, _Rows, Dynamic, _Options>
 {
     T *m_data;
-    DenseIndex m_cols;
+    Index m_cols;
   public:
-    DenseStorage() : m_data(0), m_cols(0) {}
-    DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_cols(0) {}
-    DenseStorage(DenseIndex size, DenseIndex, DenseIndex nbCols) : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_cols(nbCols)
-    { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN }
-#ifdef EIGEN_HAVE_RVALUE_REFERENCES
-    DenseStorage(DenseStorage&& other)
+    EIGEN_DEVICE_FUNC DenseStorage() : m_data(0), m_cols(0) {}
+    explicit DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_cols(0) {}
+    EIGEN_DEVICE_FUNC DenseStorage(Index size, Index rows, Index cols) : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_cols(cols)
+    {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({})
+      eigen_internal_assert(size==rows*cols && rows==_Rows && cols >=0);
+      EIGEN_UNUSED_VARIABLE(rows);
+    }
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other)
+      : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(_Rows*other.m_cols))
+      , m_cols(other.m_cols)
+    {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(Index size = m_cols*_Rows)
+      internal::smart_copy(other.m_data, other.m_data+_Rows*m_cols, m_data);
+    }
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other)
+    {
+      if (this != &other)
+      {
+        DenseStorage tmp(other);
+        this->swap(tmp);
+      }
+      return *this;
+    }    
+#if EIGEN_HAS_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC
+    DenseStorage(DenseStorage&& other) EIGEN_NOEXCEPT
       : m_data(std::move(other.m_data))
       , m_cols(std::move(other.m_cols))
     {
       other.m_data = nullptr;
+      other.m_cols = 0;
     }
-    DenseStorage& operator=(DenseStorage&& other)
+    EIGEN_DEVICE_FUNC
+    DenseStorage& operator=(DenseStorage&& other) EIGEN_NOEXCEPT
     {
       using std::swap;
       swap(m_data, other.m_data);
@@ -346,16 +465,16 @@ template<typename T, int _Rows, int _Options> class DenseStorage<T, Dynamic, _Ro
       return *this;
     }
 #endif
-    ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Rows*m_cols); }
-    void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); }
-    static DenseIndex rows(void) {return _Rows;}
-    DenseIndex cols(void) const {return m_cols;}
-    void conservativeResize(DenseIndex size, DenseIndex, DenseIndex nbCols)
+    EIGEN_DEVICE_FUNC ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Rows*m_cols); }
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); }
+    EIGEN_DEVICE_FUNC static Index rows(void) {return _Rows;}
+    EIGEN_DEVICE_FUNC Index cols(void) const {return m_cols;}
+    EIGEN_DEVICE_FUNC void conservativeResize(Index size, Index, Index cols)
     {
       m_data = internal::conditional_aligned_realloc_new_auto<T,(_Options&DontAlign)==0>(m_data, size, _Rows*m_cols);
-      m_cols = nbCols;
+      m_cols = cols;
     }
-    EIGEN_STRONG_INLINE void resize(DenseIndex size, DenseIndex, DenseIndex nbCols)
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void resize(Index size, Index, Index cols)
     {
       if(size != _Rows*m_cols)
       {
@@ -364,35 +483,55 @@ template<typename T, int _Rows, int _Options> class DenseStorage<T, Dynamic, _Ro
           m_data = internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size);
         else
           m_data = 0;
-        EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN
+        EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({})
       }
-      m_cols = nbCols;
+      m_cols = cols;
     }
-    const T *data() const { return m_data; }
-    T *data() { return m_data; }
-  private:
-    DenseStorage(const DenseStorage&);
-    DenseStorage& operator=(const DenseStorage&);
+    EIGEN_DEVICE_FUNC const T *data() const { return m_data; }
+    EIGEN_DEVICE_FUNC T *data() { return m_data; }
 };
 
 // matrix with dynamic height and fixed width (so that matrix has dynamic size).
 template<typename T, int _Cols, int _Options> class DenseStorage<T, Dynamic, Dynamic, _Cols, _Options>
 {
     T *m_data;
-    DenseIndex m_rows;
+    Index m_rows;
   public:
-    DenseStorage() : m_data(0), m_rows(0) {}
-    DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_rows(0) {}
-    DenseStorage(DenseIndex size, DenseIndex nbRows, DenseIndex) : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_rows(nbRows)
-    { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN }
-#ifdef EIGEN_HAVE_RVALUE_REFERENCES
-    DenseStorage(DenseStorage&& other)
+    EIGEN_DEVICE_FUNC DenseStorage() : m_data(0), m_rows(0) {}
+    explicit DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_rows(0) {}
+    EIGEN_DEVICE_FUNC DenseStorage(Index size, Index rows, Index cols) : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_rows(rows)
+    {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({})
+      eigen_internal_assert(size==rows*cols && rows>=0 && cols == _Cols);
+      EIGEN_UNUSED_VARIABLE(cols);
+    }
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other)
+      : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(other.m_rows*_Cols))
+      , m_rows(other.m_rows)
+    {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(Index size = m_rows*_Cols)
+      internal::smart_copy(other.m_data, other.m_data+other.m_rows*_Cols, m_data);
+    }
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other)
+    {
+      if (this != &other)
+      {
+        DenseStorage tmp(other);
+        this->swap(tmp);
+      }
+      return *this;
+    }    
+#if EIGEN_HAS_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC
+    DenseStorage(DenseStorage&& other) EIGEN_NOEXCEPT
       : m_data(std::move(other.m_data))
       , m_rows(std::move(other.m_rows))
     {
       other.m_data = nullptr;
+      other.m_rows = 0;
     }
-    DenseStorage& operator=(DenseStorage&& other)
+    EIGEN_DEVICE_FUNC
+    DenseStorage& operator=(DenseStorage&& other) EIGEN_NOEXCEPT
     {
       using std::swap;
       swap(m_data, other.m_data);
@@ -400,16 +539,16 @@ template<typename T, int _Cols, int _Options> class DenseStorage<T, Dynamic, Dyn
       return *this;
     }
 #endif
-    ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Cols*m_rows); }
-    void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); }
-    DenseIndex rows(void) const {return m_rows;}
-    static DenseIndex cols(void) {return _Cols;}
-    void conservativeResize(DenseIndex size, DenseIndex nbRows, DenseIndex)
+    EIGEN_DEVICE_FUNC ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Cols*m_rows); }
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); }
+    EIGEN_DEVICE_FUNC Index rows(void) const {return m_rows;}
+    EIGEN_DEVICE_FUNC static Index cols(void) {return _Cols;}
+    void conservativeResize(Index size, Index rows, Index)
     {
       m_data = internal::conditional_aligned_realloc_new_auto<T,(_Options&DontAlign)==0>(m_data, size, m_rows*_Cols);
-      m_rows = nbRows;
+      m_rows = rows;
     }
-    EIGEN_STRONG_INLINE void resize(DenseIndex size, DenseIndex nbRows, DenseIndex)
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void resize(Index size, Index rows, Index)
     {
       if(size != m_rows*_Cols)
       {
@@ -418,15 +557,12 @@ template<typename T, int _Cols, int _Options> class DenseStorage<T, Dynamic, Dyn
           m_data = internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size);
         else
           m_data = 0;
-        EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN
+        EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({})
       }
-      m_rows = nbRows;
+      m_rows = rows;
     }
-    const T *data() const { return m_data; }
-    T *data() { return m_data; }
-  private:
-    DenseStorage(const DenseStorage&);
-    DenseStorage& operator=(const DenseStorage&);
+    EIGEN_DEVICE_FUNC const T *data() const { return m_data; }
+    EIGEN_DEVICE_FUNC T *data() { return m_data; }
 };
 
 } // end namespace Eigen
diff --git a/eigen/Eigen/src/Core/Diagonal.h b/eigen/Eigen/src/Core/Diagonal.h
index 68cf6d4..c62f5ff 100644
--- a/eigen/Eigen/src/Core/Diagonal.h
+++ b/eigen/Eigen/src/Core/Diagonal.h
@@ -21,7 +21,7 @@ namespace Eigen {
   * \param MatrixType the type of the object in which we are taking a sub/main/super diagonal
   * \param DiagIndex the index of the sub/super diagonal. The default is 0 and it means the main diagonal.
   *              A positive value means a superdiagonal, a negative value means a subdiagonal.
-  *              You can also use Dynamic so the index can be set at runtime.
+  *              You can also use DynamicIndex so the index can be set at runtime.
   *
   * The matrix is not required to be square.
   *
@@ -37,7 +37,7 @@ template<typename MatrixType, int DiagIndex>
 struct traits<Diagonal<MatrixType,DiagIndex> >
  : traits<MatrixType>
 {
-  typedef typename nested<MatrixType>::type MatrixTypeNested;
+  typedef typename ref_selector<MatrixType>::type MatrixTypeNested;
   typedef typename remove_reference<MatrixTypeNested>::type _MatrixTypeNested;
   typedef typename MatrixType::StorageKind StorageKind;
   enum {
@@ -52,8 +52,7 @@ struct traits<Diagonal<MatrixType,DiagIndex> >
                                                  MatrixType::MaxColsAtCompileTime - EIGEN_PLAIN_ENUM_MAX( DiagIndex, 0))),
     MaxColsAtCompileTime = 1,
     MaskLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0,
-    Flags = (unsigned int)_MatrixTypeNested::Flags & (HereditaryBits | LinearAccessBit | MaskLvalueBit | DirectAccessBit) & ~RowMajorBit,
-    CoeffReadCost = _MatrixTypeNested::CoeffReadCost,
+    Flags = (unsigned int)_MatrixTypeNested::Flags & (RowMajorBit | MaskLvalueBit | DirectAccessBit) & ~RowMajorBit, // FIXME DirectAccessBit should not be handled by expressions
     MatrixTypeOuterStride = outer_stride_at_compile_time<MatrixType>::ret,
     InnerStrideAtCompileTime = MatrixTypeOuterStride == Dynamic ? Dynamic : MatrixTypeOuterStride+1,
     OuterStrideAtCompileTime = 0
@@ -70,20 +69,28 @@ template<typename MatrixType, int _DiagIndex> class Diagonal
     typedef typename internal::dense_xpr_base<Diagonal>::type Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(Diagonal)
 
-    inline Diagonal(MatrixType& matrix, Index a_index = DiagIndex) : m_matrix(matrix), m_index(a_index) {}
+    EIGEN_DEVICE_FUNC
+    explicit inline Diagonal(MatrixType& matrix, Index a_index = DiagIndex) : m_matrix(matrix), m_index(a_index) {}
 
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Diagonal)
 
+    EIGEN_DEVICE_FUNC
     inline Index rows() const
-    { return m_index.value()<0 ? (std::min<Index>)(m_matrix.cols(),m_matrix.rows()+m_index.value()) : (std::min<Index>)(m_matrix.rows(),m_matrix.cols()-m_index.value()); }
+    {
+      return m_index.value()<0 ? numext::mini<Index>(m_matrix.cols(),m_matrix.rows()+m_index.value())
+                               : numext::mini<Index>(m_matrix.rows(),m_matrix.cols()-m_index.value());
+    }
 
+    EIGEN_DEVICE_FUNC
     inline Index cols() const { return 1; }
 
+    EIGEN_DEVICE_FUNC
     inline Index innerStride() const
     {
       return m_matrix.outerStride() + 1;
     }
 
+    EIGEN_DEVICE_FUNC
     inline Index outerStride() const
     {
       return 0;
@@ -95,62 +102,75 @@ template<typename MatrixType, int _DiagIndex> class Diagonal
                        const Scalar
                      >::type ScalarWithConstIfNotLvalue;
 
-    inline ScalarWithConstIfNotLvalue* data() { return &(m_matrix.const_cast_derived().coeffRef(rowOffset(), colOffset())); }
-    inline const Scalar* data() const { return &(m_matrix.const_cast_derived().coeffRef(rowOffset(), colOffset())); }
+    EIGEN_DEVICE_FUNC
+    inline ScalarWithConstIfNotLvalue* data() { return &(m_matrix.coeffRef(rowOffset(), colOffset())); }
+    EIGEN_DEVICE_FUNC
+    inline const Scalar* data() const { return &(m_matrix.coeffRef(rowOffset(), colOffset())); }
 
+    EIGEN_DEVICE_FUNC
     inline Scalar& coeffRef(Index row, Index)
     {
       EIGEN_STATIC_ASSERT_LVALUE(MatrixType)
-      return m_matrix.const_cast_derived().coeffRef(row+rowOffset(), row+colOffset());
+      return m_matrix.coeffRef(row+rowOffset(), row+colOffset());
     }
 
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index row, Index) const
     {
-      return m_matrix.const_cast_derived().coeffRef(row+rowOffset(), row+colOffset());
+      return m_matrix.coeffRef(row+rowOffset(), row+colOffset());
     }
 
+    EIGEN_DEVICE_FUNC
     inline CoeffReturnType coeff(Index row, Index) const
     {
       return m_matrix.coeff(row+rowOffset(), row+colOffset());
     }
 
+    EIGEN_DEVICE_FUNC
     inline Scalar& coeffRef(Index idx)
     {
       EIGEN_STATIC_ASSERT_LVALUE(MatrixType)
-      return m_matrix.const_cast_derived().coeffRef(idx+rowOffset(), idx+colOffset());
+      return m_matrix.coeffRef(idx+rowOffset(), idx+colOffset());
     }
 
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index idx) const
     {
-      return m_matrix.const_cast_derived().coeffRef(idx+rowOffset(), idx+colOffset());
+      return m_matrix.coeffRef(idx+rowOffset(), idx+colOffset());
     }
 
+    EIGEN_DEVICE_FUNC
     inline CoeffReturnType coeff(Index idx) const
     {
       return m_matrix.coeff(idx+rowOffset(), idx+colOffset());
     }
 
-    const typename internal::remove_all<typename MatrixType::Nested>::type& 
+    EIGEN_DEVICE_FUNC
+    inline const typename internal::remove_all<typename MatrixType::Nested>::type& 
     nestedExpression() const 
     {
       return m_matrix;
     }
 
-    int index() const
+    EIGEN_DEVICE_FUNC
+    inline Index index() const
     {
       return m_index.value();
     }
 
   protected:
-    typename MatrixType::Nested m_matrix;
+    typename internal::ref_selector<MatrixType>::non_const_type m_matrix;
     const internal::variable_if_dynamicindex<Index, DiagIndex> m_index;
 
   private:
     // some compilers may fail to optimize std::max etc in case of compile-time constants...
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Index absDiagIndex() const { return m_index.value()>0 ? m_index.value() : -m_index.value(); }
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Index rowOffset() const { return m_index.value()>0 ? 0 : -m_index.value(); }
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Index colOffset() const { return m_index.value()>0 ? m_index.value() : 0; }
-    // triger a compile time error is someone try to call packet
+    // trigger a compile-time error if someone try to call packet
     template<int LoadMode> typename MatrixType::PacketReturnType packet(Index) const;
     template<int LoadMode> typename MatrixType::PacketReturnType packet(Index,Index) const;
 };
@@ -164,15 +184,15 @@ template<typename MatrixType, int _DiagIndex> class Diagonal
   *
   * \sa class Diagonal */
 template<typename Derived>
-inline typename MatrixBase<Derived>::DiagonalReturnType
+EIGEN_DEVICE_FUNC inline typename MatrixBase<Derived>::DiagonalReturnType
 MatrixBase<Derived>::diagonal()
 {
-  return derived();
+  return DiagonalReturnType(derived());
 }
 
 /** This is the const version of diagonal(). */
 template<typename Derived>
-inline typename MatrixBase<Derived>::ConstDiagonalReturnType
+EIGEN_DEVICE_FUNC inline typename MatrixBase<Derived>::ConstDiagonalReturnType
 MatrixBase<Derived>::diagonal() const
 {
   return ConstDiagonalReturnType(derived());
@@ -190,7 +210,7 @@ MatrixBase<Derived>::diagonal() const
   *
   * \sa MatrixBase::diagonal(), class Diagonal */
 template<typename Derived>
-inline typename MatrixBase<Derived>::DiagonalDynamicIndexReturnType
+EIGEN_DEVICE_FUNC inline typename MatrixBase<Derived>::DiagonalDynamicIndexReturnType
 MatrixBase<Derived>::diagonal(Index index)
 {
   return DiagonalDynamicIndexReturnType(derived(), index);
@@ -198,7 +218,7 @@ MatrixBase<Derived>::diagonal(Index index)
 
 /** This is the const version of diagonal(Index). */
 template<typename Derived>
-inline typename MatrixBase<Derived>::ConstDiagonalDynamicIndexReturnType
+EIGEN_DEVICE_FUNC inline typename MatrixBase<Derived>::ConstDiagonalDynamicIndexReturnType
 MatrixBase<Derived>::diagonal(Index index) const
 {
   return ConstDiagonalDynamicIndexReturnType(derived(), index);
@@ -216,20 +236,22 @@ MatrixBase<Derived>::diagonal(Index index) const
   *
   * \sa MatrixBase::diagonal(), class Diagonal */
 template<typename Derived>
-template<int Index>
-inline typename MatrixBase<Derived>::template DiagonalIndexReturnType<Index>::Type
+template<int Index_>
+EIGEN_DEVICE_FUNC
+inline typename MatrixBase<Derived>::template DiagonalIndexReturnType<Index_>::Type
 MatrixBase<Derived>::diagonal()
 {
-  return derived();
+  return typename DiagonalIndexReturnType<Index_>::Type(derived());
 }
 
 /** This is the const version of diagonal<int>(). */
 template<typename Derived>
-template<int Index>
-inline typename MatrixBase<Derived>::template ConstDiagonalIndexReturnType<Index>::Type
+template<int Index_>
+EIGEN_DEVICE_FUNC
+inline typename MatrixBase<Derived>::template ConstDiagonalIndexReturnType<Index_>::Type
 MatrixBase<Derived>::diagonal() const
 {
-  return derived();
+  return typename ConstDiagonalIndexReturnType<Index_>::Type(derived());
 }
 
 } // end namespace Eigen
diff --git a/eigen/Eigen/src/Core/DiagonalMatrix.h b/eigen/Eigen/src/Core/DiagonalMatrix.h
index 53c757b..4e8297e 100644
--- a/eigen/Eigen/src/Core/DiagonalMatrix.h
+++ b/eigen/Eigen/src/Core/DiagonalMatrix.h
@@ -22,7 +22,7 @@ class DiagonalBase : public EigenBase<Derived>
     typedef typename DiagonalVectorType::Scalar Scalar;
     typedef typename DiagonalVectorType::RealScalar RealScalar;
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
-    typedef typename internal::traits<Derived>::Index Index;
+    typedef typename internal::traits<Derived>::StorageIndex StorageIndex;
 
     enum {
       RowsAtCompileTime = DiagonalVectorType::SizeAtCompileTime,
@@ -30,79 +30,61 @@ class DiagonalBase : public EigenBase<Derived>
       MaxRowsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime,
       MaxColsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime,
       IsVectorAtCompileTime = 0,
-      Flags = 0
+      Flags = NoPreferredStorageOrderBit
     };
 
     typedef Matrix<Scalar, RowsAtCompileTime, ColsAtCompileTime, 0, MaxRowsAtCompileTime, MaxColsAtCompileTime> DenseMatrixType;
     typedef DenseMatrixType DenseType;
     typedef DiagonalMatrix<Scalar,DiagonalVectorType::SizeAtCompileTime,DiagonalVectorType::MaxSizeAtCompileTime> PlainObject;
 
+    EIGEN_DEVICE_FUNC
     inline const Derived& derived() const { return *static_cast<const Derived*>(this); }
+    EIGEN_DEVICE_FUNC
     inline Derived& derived() { return *static_cast<Derived*>(this); }
 
+    EIGEN_DEVICE_FUNC
     DenseMatrixType toDenseMatrix() const { return derived(); }
-    template<typename DenseDerived>
-    void evalTo(MatrixBase<DenseDerived> &other) const;
-    template<typename DenseDerived>
-    inline void addTo(MatrixBase<DenseDerived> &other) const
-    { other.diagonal() += diagonal(); }
-    template<typename DenseDerived>
-    inline void subTo(MatrixBase<DenseDerived> &other) const
-    { other.diagonal() -= diagonal(); }
 
+    EIGEN_DEVICE_FUNC
     inline const DiagonalVectorType& diagonal() const { return derived().diagonal(); }
+    EIGEN_DEVICE_FUNC
     inline DiagonalVectorType& diagonal() { return derived().diagonal(); }
 
+    EIGEN_DEVICE_FUNC
     inline Index rows() const { return diagonal().size(); }
+    EIGEN_DEVICE_FUNC
     inline Index cols() const { return diagonal().size(); }
 
-    /** \returns the diagonal matrix product of \c *this by the matrix \a matrix.
-      */
     template<typename MatrixDerived>
-    const DiagonalProduct<MatrixDerived, Derived, OnTheLeft>
+    EIGEN_DEVICE_FUNC
+    const Product<Derived,MatrixDerived,LazyProduct>
     operator*(const MatrixBase<MatrixDerived> &matrix) const
     {
-      return DiagonalProduct<MatrixDerived, Derived, OnTheLeft>(matrix.derived(), derived());
+      return Product<Derived, MatrixDerived, LazyProduct>(derived(),matrix.derived());
     }
 
-    inline const DiagonalWrapper<const CwiseUnaryOp<internal::scalar_inverse_op<Scalar>, const DiagonalVectorType> >
+    typedef DiagonalWrapper<const CwiseUnaryOp<internal::scalar_inverse_op<Scalar>, const DiagonalVectorType> > InverseReturnType;
+    EIGEN_DEVICE_FUNC
+    inline const InverseReturnType
     inverse() const
     {
-      return diagonal().cwiseInverse();
+      return InverseReturnType(diagonal().cwiseInverse());
     }
     
-    inline const DiagonalWrapper<const CwiseUnaryOp<internal::scalar_multiple_op<Scalar>, const DiagonalVectorType> >
+    EIGEN_DEVICE_FUNC
+    inline const DiagonalWrapper<const EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(DiagonalVectorType,Scalar,product) >
     operator*(const Scalar& scalar) const
     {
-      return diagonal() * scalar;
+      return DiagonalWrapper<const EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(DiagonalVectorType,Scalar,product) >(diagonal() * scalar);
     }
-    friend inline const DiagonalWrapper<const CwiseUnaryOp<internal::scalar_multiple_op<Scalar>, const DiagonalVectorType> >
+    EIGEN_DEVICE_FUNC
+    friend inline const DiagonalWrapper<const EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(Scalar,DiagonalVectorType,product) >
     operator*(const Scalar& scalar, const DiagonalBase& other)
     {
-      return other.diagonal() * scalar;
-    }
-    
-    #ifdef EIGEN2_SUPPORT
-    template<typename OtherDerived>
-    bool isApprox(const DiagonalBase<OtherDerived>& other, typename NumTraits<Scalar>::Real precision = NumTraits<Scalar>::dummy_precision()) const
-    {
-      return diagonal().isApprox(other.diagonal(), precision);
-    }
-    template<typename OtherDerived>
-    bool isApprox(const MatrixBase<OtherDerived>& other, typename NumTraits<Scalar>::Real precision = NumTraits<Scalar>::dummy_precision()) const
-    {
-      return toDenseMatrix().isApprox(other, precision);
+      return DiagonalWrapper<const EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(Scalar,DiagonalVectorType,product) >(scalar * other.diagonal());
     }
-    #endif
 };
 
-template<typename Derived>
-template<typename DenseDerived>
-inline void DiagonalBase<Derived>::evalTo(MatrixBase<DenseDerived> &other) const
-{
-  other.setZero();
-  other.diagonal() = diagonal();
-}
 #endif
 
 /** \class DiagonalMatrix
@@ -124,10 +106,9 @@ struct traits<DiagonalMatrix<_Scalar,SizeAtCompileTime,MaxSizeAtCompileTime> >
  : traits<Matrix<_Scalar,SizeAtCompileTime,SizeAtCompileTime,0,MaxSizeAtCompileTime,MaxSizeAtCompileTime> >
 {
   typedef Matrix<_Scalar,SizeAtCompileTime,1,0,MaxSizeAtCompileTime,1> DiagonalVectorType;
-  typedef Dense StorageKind;
-  typedef DenseIndex Index;
+  typedef DiagonalShape StorageKind;
   enum {
-    Flags = LvalueBit
+    Flags = LvalueBit | NoPreferredStorageOrderBit
   };
 };
 }
@@ -141,7 +122,7 @@ class DiagonalMatrix
     typedef const DiagonalMatrix& Nested;
     typedef _Scalar Scalar;
     typedef typename internal::traits<DiagonalMatrix>::StorageKind StorageKind;
-    typedef typename internal::traits<DiagonalMatrix>::Index Index;
+    typedef typename internal::traits<DiagonalMatrix>::StorageIndex StorageIndex;
     #endif
 
   protected:
@@ -151,24 +132,31 @@ class DiagonalMatrix
   public:
 
     /** const version of diagonal(). */
+    EIGEN_DEVICE_FUNC
     inline const DiagonalVectorType& diagonal() const { return m_diagonal; }
     /** \returns a reference to the stored vector of diagonal coefficients. */
+    EIGEN_DEVICE_FUNC
     inline DiagonalVectorType& diagonal() { return m_diagonal; }
 
     /** Default constructor without initialization */
+    EIGEN_DEVICE_FUNC
     inline DiagonalMatrix() {}
 
     /** Constructs a diagonal matrix with given dimension  */
-    inline DiagonalMatrix(Index dim) : m_diagonal(dim) {}
+    EIGEN_DEVICE_FUNC
+    explicit inline DiagonalMatrix(Index dim) : m_diagonal(dim) {}
 
     /** 2D constructor. */
+    EIGEN_DEVICE_FUNC
     inline DiagonalMatrix(const Scalar& x, const Scalar& y) : m_diagonal(x,y) {}
 
     /** 3D constructor. */
+    EIGEN_DEVICE_FUNC
     inline DiagonalMatrix(const Scalar& x, const Scalar& y, const Scalar& z) : m_diagonal(x,y,z) {}
 
     /** Copy constructor. */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     inline DiagonalMatrix(const DiagonalBase<OtherDerived>& other) : m_diagonal(other.diagonal()) {}
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
@@ -178,11 +166,13 @@ class DiagonalMatrix
 
     /** generic constructor from expression of the diagonal coefficients */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     explicit inline DiagonalMatrix(const MatrixBase<OtherDerived>& other) : m_diagonal(other)
     {}
 
     /** Copy operator. */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     DiagonalMatrix& operator=(const DiagonalBase<OtherDerived>& other)
     {
       m_diagonal = other.diagonal();
@@ -193,6 +183,7 @@ class DiagonalMatrix
     /** This is a special case of the templated operator=. Its purpose is to
       * prevent a default operator= from hiding the templated operator=.
       */
+    EIGEN_DEVICE_FUNC
     DiagonalMatrix& operator=(const DiagonalMatrix& other)
     {
       m_diagonal = other.diagonal();
@@ -201,14 +192,19 @@ class DiagonalMatrix
     #endif
 
     /** Resizes to given size. */
+    EIGEN_DEVICE_FUNC
     inline void resize(Index size) { m_diagonal.resize(size); }
     /** Sets all coefficients to zero. */
+    EIGEN_DEVICE_FUNC
     inline void setZero() { m_diagonal.setZero(); }
     /** Resizes and sets all coefficients to zero. */
+    EIGEN_DEVICE_FUNC
     inline void setZero(Index size) { m_diagonal.setZero(size); }
     /** Sets this matrix to be the identity matrix of the current size. */
+    EIGEN_DEVICE_FUNC
     inline void setIdentity() { m_diagonal.setOnes(); }
     /** Sets this matrix to be the identity matrix of the given size. */
+    EIGEN_DEVICE_FUNC
     inline void setIdentity(Index size) { m_diagonal.setOnes(size); }
 };
 
@@ -232,14 +228,15 @@ struct traits<DiagonalWrapper<_DiagonalVectorType> >
 {
   typedef _DiagonalVectorType DiagonalVectorType;
   typedef typename DiagonalVectorType::Scalar Scalar;
-  typedef typename DiagonalVectorType::Index Index;
-  typedef typename DiagonalVectorType::StorageKind StorageKind;
+  typedef typename DiagonalVectorType::StorageIndex StorageIndex;
+  typedef DiagonalShape StorageKind;
+  typedef typename traits<DiagonalVectorType>::XprKind XprKind;
   enum {
     RowsAtCompileTime = DiagonalVectorType::SizeAtCompileTime,
     ColsAtCompileTime = DiagonalVectorType::SizeAtCompileTime,
-    MaxRowsAtCompileTime = DiagonalVectorType::SizeAtCompileTime,
-    MaxColsAtCompileTime = DiagonalVectorType::SizeAtCompileTime,
-    Flags =  traits<DiagonalVectorType>::Flags & LvalueBit
+    MaxRowsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime,
+    MaxColsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime,
+    Flags =  (traits<DiagonalVectorType>::Flags & LvalueBit) | NoPreferredStorageOrderBit
   };
 };
 }
@@ -255,9 +252,11 @@ class DiagonalWrapper
     #endif
 
     /** Constructor from expression of diagonal coefficients to wrap. */
-    inline DiagonalWrapper(DiagonalVectorType& a_diagonal) : m_diagonal(a_diagonal) {}
+    EIGEN_DEVICE_FUNC
+    explicit inline DiagonalWrapper(DiagonalVectorType& a_diagonal) : m_diagonal(a_diagonal) {}
 
     /** \returns a const reference to the wrapped expression of diagonal coefficients. */
+    EIGEN_DEVICE_FUNC
     const DiagonalVectorType& diagonal() const { return m_diagonal; }
 
   protected:
@@ -274,10 +273,10 @@ class DiagonalWrapper
   * \sa class DiagonalWrapper, class DiagonalMatrix, diagonal(), isDiagonal()
   **/
 template<typename Derived>
-inline const DiagonalWrapper<const Derived>
+EIGEN_DEVICE_FUNC inline const DiagonalWrapper<const Derived>
 MatrixBase<Derived>::asDiagonal() const
 {
-  return derived();
+  return DiagonalWrapper<const Derived>(derived());
 }
 
 /** \returns true if *this is approximately equal to a diagonal matrix,
@@ -291,12 +290,11 @@ MatrixBase<Derived>::asDiagonal() const
 template<typename Derived>
 bool MatrixBase<Derived>::isDiagonal(const RealScalar& prec) const
 {
-  using std::abs;
   if(cols() != rows()) return false;
   RealScalar maxAbsOnDiagonal = static_cast<RealScalar>(-1);
   for(Index j = 0; j < cols(); ++j)
   {
-    RealScalar absOnDiagonal = abs(coeff(j,j));
+    RealScalar absOnDiagonal = numext::abs(coeff(j,j));
     if(absOnDiagonal > maxAbsOnDiagonal) maxAbsOnDiagonal = absOnDiagonal;
   }
   for(Index j = 0; j < cols(); ++j)
@@ -308,6 +306,38 @@ bool MatrixBase<Derived>::isDiagonal(const RealScalar& prec) const
   return true;
 }
 
+namespace internal {
+
+template<> struct storage_kind_to_shape<DiagonalShape> { typedef DiagonalShape Shape; };
+
+struct Diagonal2Dense {};
+
+template<> struct AssignmentKind<DenseShape,DiagonalShape> { typedef Diagonal2Dense Kind; };
+
+// Diagonal matrix to Dense assignment
+template< typename DstXprType, typename SrcXprType, typename Functor>
+struct Assignment<DstXprType, SrcXprType, Functor, Diagonal2Dense>
+{
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &/*func*/)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+    
+    dst.setZero();
+    dst.diagonal() = src.diagonal();
+  }
+  
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &/*func*/)
+  { dst.diagonal() += src.diagonal(); }
+  
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &/*func*/)
+  { dst.diagonal() -= src.diagonal(); }
+};
+
+} // namespace internal
+
 } // end namespace Eigen
 
 #endif // EIGEN_DIAGONALMATRIX_H
diff --git a/eigen/Eigen/src/Core/DiagonalProduct.h b/eigen/Eigen/src/Core/DiagonalProduct.h
index cc6b536..7911d1c 100644
--- a/eigen/Eigen/src/Core/DiagonalProduct.h
+++ b/eigen/Eigen/src/Core/DiagonalProduct.h
@@ -13,117 +13,14 @@
 
 namespace Eigen { 
 
-namespace internal {
-template<typename MatrixType, typename DiagonalType, int ProductOrder>
-struct traits<DiagonalProduct<MatrixType, DiagonalType, ProductOrder> >
- : traits<MatrixType>
-{
-  typedef typename scalar_product_traits<typename MatrixType::Scalar, typename DiagonalType::Scalar>::ReturnType Scalar;
-  enum {
-    RowsAtCompileTime = MatrixType::RowsAtCompileTime,
-    ColsAtCompileTime = MatrixType::ColsAtCompileTime,
-    MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
-    MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime,
-
-    _StorageOrder = MatrixType::Flags & RowMajorBit ? RowMajor : ColMajor,
-    _ScalarAccessOnDiag =  !((int(_StorageOrder) == ColMajor && int(ProductOrder) == OnTheLeft)
-                          ||(int(_StorageOrder) == RowMajor && int(ProductOrder) == OnTheRight)),
-    _SameTypes = is_same<typename MatrixType::Scalar, typename DiagonalType::Scalar>::value,
-    // FIXME currently we need same types, but in the future the next rule should be the one
-    //_Vectorizable = bool(int(MatrixType::Flags)&PacketAccessBit) && ((!_PacketOnDiag) || (_SameTypes && bool(int(DiagonalType::DiagonalVectorType::Flags)&PacketAccessBit))),
-    _Vectorizable = bool(int(MatrixType::Flags)&PacketAccessBit) && _SameTypes && (_ScalarAccessOnDiag || (bool(int(DiagonalType::DiagonalVectorType::Flags)&PacketAccessBit))),
-    _LinearAccessMask = (RowsAtCompileTime==1 || ColsAtCompileTime==1) ? LinearAccessBit : 0,
-
-    Flags = ((HereditaryBits|_LinearAccessMask|AlignedBit) & (unsigned int)(MatrixType::Flags)) | (_Vectorizable ? PacketAccessBit : 0),//(int(MatrixType::Flags)&int(DiagonalType::DiagonalVectorType::Flags)&AlignedBit),
-    Cost0 = EIGEN_ADD_COST(NumTraits<Scalar>::MulCost, MatrixType::CoeffReadCost),
-    CoeffReadCost = EIGEN_ADD_COST(Cost0,DiagonalType::DiagonalVectorType::CoeffReadCost)
-  };
-};
-}
-
-template<typename MatrixType, typename DiagonalType, int ProductOrder>
-class DiagonalProduct : internal::no_assignment_operator,
-                        public MatrixBase<DiagonalProduct<MatrixType, DiagonalType, ProductOrder> >
-{
-  public:
-
-    typedef MatrixBase<DiagonalProduct> Base;
-    EIGEN_DENSE_PUBLIC_INTERFACE(DiagonalProduct)
-
-    inline DiagonalProduct(const MatrixType& matrix, const DiagonalType& diagonal)
-      : m_matrix(matrix), m_diagonal(diagonal)
-    {
-      eigen_assert(diagonal.diagonal().size() == (ProductOrder == OnTheLeft ? matrix.rows() : matrix.cols()));
-    }
-
-    EIGEN_STRONG_INLINE Index rows() const { return m_matrix.rows(); }
-    EIGEN_STRONG_INLINE Index cols() const { return m_matrix.cols(); }
-
-    EIGEN_STRONG_INLINE const Scalar coeff(Index row, Index col) const
-    {
-      return m_diagonal.diagonal().coeff(ProductOrder == OnTheLeft ? row : col) * m_matrix.coeff(row, col);
-    }
-    
-    EIGEN_STRONG_INLINE const Scalar coeff(Index idx) const
-    {
-      enum {
-        StorageOrder = int(MatrixType::Flags) & RowMajorBit ? RowMajor : ColMajor
-      };
-      return coeff(int(StorageOrder)==ColMajor?idx:0,int(StorageOrder)==ColMajor?0:idx);
-    }
-
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet(Index row, Index col) const
-    {
-      enum {
-        StorageOrder = Flags & RowMajorBit ? RowMajor : ColMajor
-      };
-      const Index indexInDiagonalVector = ProductOrder == OnTheLeft ? row : col;
-      return packet_impl<LoadMode>(row,col,indexInDiagonalVector,typename internal::conditional<
-        ((int(StorageOrder) == RowMajor && int(ProductOrder) == OnTheLeft)
-       ||(int(StorageOrder) == ColMajor && int(ProductOrder) == OnTheRight)), internal::true_type, internal::false_type>::type());
-    }
-    
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet(Index idx) const
-    {
-      enum {
-        StorageOrder = int(MatrixType::Flags) & RowMajorBit ? RowMajor : ColMajor
-      };
-      return packet<LoadMode>(int(StorageOrder)==ColMajor?idx:0,int(StorageOrder)==ColMajor?0:idx);
-    }
-
-  protected:
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet_impl(Index row, Index col, Index id, internal::true_type) const
-    {
-      return internal::pmul(m_matrix.template packet<LoadMode>(row, col),
-                     internal::pset1<PacketScalar>(m_diagonal.diagonal().coeff(id)));
-    }
-
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet_impl(Index row, Index col, Index id, internal::false_type) const
-    {
-      enum {
-        InnerSize = (MatrixType::Flags & RowMajorBit) ? MatrixType::ColsAtCompileTime : MatrixType::RowsAtCompileTime,
-        DiagonalVectorPacketLoadMode = (LoadMode == Aligned && (((InnerSize%16) == 0) || (int(DiagonalType::DiagonalVectorType::Flags)&AlignedBit)==AlignedBit) ? Aligned : Unaligned)
-      };
-      return internal::pmul(m_matrix.template packet<LoadMode>(row, col),
-                     m_diagonal.diagonal().template packet<DiagonalVectorPacketLoadMode>(id));
-    }
-
-    typename MatrixType::Nested m_matrix;
-    typename DiagonalType::Nested m_diagonal;
-};
-
 /** \returns the diagonal matrix product of \c *this by the diagonal matrix \a diagonal.
   */
 template<typename Derived>
 template<typename DiagonalDerived>
-inline const DiagonalProduct<Derived, DiagonalDerived, OnTheRight>
+EIGEN_DEVICE_FUNC inline const Product<Derived, DiagonalDerived, LazyProduct>
 MatrixBase<Derived>::operator*(const DiagonalBase<DiagonalDerived> &a_diagonal) const
 {
-  return DiagonalProduct<Derived, DiagonalDerived, OnTheRight>(derived(), a_diagonal.derived());
+  return Product<Derived, DiagonalDerived, LazyProduct>(derived(),a_diagonal.derived());
 }
 
 } // end namespace Eigen
diff --git a/eigen/Eigen/src/Core/Dot.h b/eigen/Eigen/src/Core/Dot.h
index 23aab83..bb8e3fe 100644
--- a/eigen/Eigen/src/Core/Dot.h
+++ b/eigen/Eigen/src/Core/Dot.h
@@ -28,26 +28,31 @@ template<typename T, typename U,
 >
 struct dot_nocheck
 {
-  typedef typename scalar_product_traits<typename traits<T>::Scalar,typename traits<U>::Scalar>::ReturnType ResScalar;
+  typedef scalar_conj_product_op<typename traits<T>::Scalar,typename traits<U>::Scalar> conj_prod;
+  typedef typename conj_prod::result_type ResScalar;
+  EIGEN_DEVICE_FUNC
   static inline ResScalar run(const MatrixBase<T>& a, const MatrixBase<U>& b)
   {
-    return a.template binaryExpr<scalar_conj_product_op<typename traits<T>::Scalar,typename traits<U>::Scalar> >(b).sum();
+    return a.template binaryExpr<conj_prod>(b).sum();
   }
 };
 
 template<typename T, typename U>
 struct dot_nocheck<T, U, true>
 {
-  typedef typename scalar_product_traits<typename traits<T>::Scalar,typename traits<U>::Scalar>::ReturnType ResScalar;
+  typedef scalar_conj_product_op<typename traits<T>::Scalar,typename traits<U>::Scalar> conj_prod;
+  typedef typename conj_prod::result_type ResScalar;
+  EIGEN_DEVICE_FUNC
   static inline ResScalar run(const MatrixBase<T>& a, const MatrixBase<U>& b)
   {
-    return a.transpose().template binaryExpr<scalar_conj_product_op<typename traits<T>::Scalar,typename traits<U>::Scalar> >(b).sum();
+    return a.transpose().template binaryExpr<conj_prod>(b).sum();
   }
 };
 
 } // end namespace internal
 
-/** \returns the dot product of *this with other.
+/** \fn MatrixBase::dot
+  * \returns the dot product of *this with other.
   *
   * \only_for_vectors
   *
@@ -59,58 +64,33 @@ struct dot_nocheck<T, U, true>
   */
 template<typename Derived>
 template<typename OtherDerived>
-inline typename internal::scalar_product_traits<typename internal::traits<Derived>::Scalar,typename internal::traits<OtherDerived>::Scalar>::ReturnType
+EIGEN_DEVICE_FUNC
+typename ScalarBinaryOpTraits<typename internal::traits<Derived>::Scalar,typename internal::traits<OtherDerived>::Scalar>::ReturnType
 MatrixBase<Derived>::dot(const MatrixBase<OtherDerived>& other) const
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
   EIGEN_STATIC_ASSERT_SAME_VECTOR_SIZE(Derived,OtherDerived)
+#if !(defined(EIGEN_NO_STATIC_ASSERT) && defined(EIGEN_NO_DEBUG))
   typedef internal::scalar_conj_product_op<Scalar,typename OtherDerived::Scalar> func;
   EIGEN_CHECK_BINARY_COMPATIBILIY(func,Scalar,typename OtherDerived::Scalar);
-
+#endif
+  
   eigen_assert(size() == other.size());
 
   return internal::dot_nocheck<Derived,OtherDerived>::run(*this, other);
 }
 
-#ifdef EIGEN2_SUPPORT
-/** \returns the dot product of *this with other, with the Eigen2 convention that the dot product is linear in the first variable
-  * (conjugating the second variable). Of course this only makes a difference in the complex case.
-  *
-  * This method is only available in EIGEN2_SUPPORT mode.
-  *
-  * \only_for_vectors
-  *
-  * \sa dot()
-  */
-template<typename Derived>
-template<typename OtherDerived>
-typename internal::traits<Derived>::Scalar
-MatrixBase<Derived>::eigen2_dot(const MatrixBase<OtherDerived>& other) const
-{
-  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
-  EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
-  EIGEN_STATIC_ASSERT_SAME_VECTOR_SIZE(Derived,OtherDerived)
-  EIGEN_STATIC_ASSERT((internal::is_same<Scalar, typename OtherDerived::Scalar>::value),
-    YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
-
-  eigen_assert(size() == other.size());
-
-  return internal::dot_nocheck<OtherDerived,Derived>::run(other,*this);
-}
-#endif
-
-
 //---------- implementation of L2 norm and related functions ----------
 
 /** \returns, for vectors, the squared \em l2 norm of \c *this, and for matrices the Frobenius norm.
   * In both cases, it consists in the sum of the square of all the matrix entries.
   * For vectors, this is also equals to the dot product of \c *this with itself.
   *
-  * \sa dot(), norm()
+  * \sa dot(), norm(), lpNorm()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE typename NumTraits<typename internal::traits<Derived>::Scalar>::Real MatrixBase<Derived>::squaredNorm() const
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename NumTraits<typename internal::traits<Derived>::Scalar>::Real MatrixBase<Derived>::squaredNorm() const
 {
   return numext::real((*this).cwiseAbs2().sum());
 }
@@ -119,41 +99,98 @@ EIGEN_STRONG_INLINE typename NumTraits<typename internal::traits<Derived>::Scala
   * In both cases, it consists in the square root of the sum of the square of all the matrix entries.
   * For vectors, this is also equals to the square root of the dot product of \c *this with itself.
   *
-  * \sa dot(), squaredNorm()
+  * \sa lpNorm(), dot(), squaredNorm()
   */
 template<typename Derived>
-inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real MatrixBase<Derived>::norm() const
+EIGEN_DEVICE_FUNC inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real MatrixBase<Derived>::norm() const
 {
-  using std::sqrt;
-  return sqrt(squaredNorm());
+  return numext::sqrt(squaredNorm());
 }
 
-/** \returns an expression of the quotient of *this by its own norm.
+/** \returns an expression of the quotient of \c *this by its own norm.
+  *
+  * \warning If the input vector is too small (i.e., this->norm()==0),
+  *          then this function returns a copy of the input.
   *
   * \only_for_vectors
   *
   * \sa norm(), normalize()
   */
 template<typename Derived>
-inline const typename MatrixBase<Derived>::PlainObject
+EIGEN_DEVICE_FUNC inline const typename MatrixBase<Derived>::PlainObject
 MatrixBase<Derived>::normalized() const
 {
-  typedef typename internal::nested<Derived>::type Nested;
-  typedef typename internal::remove_reference<Nested>::type _Nested;
+  typedef typename internal::nested_eval<Derived,2>::type _Nested;
   _Nested n(derived());
-  return n / n.norm();
+  RealScalar z = n.squaredNorm();
+  // NOTE: after extensive benchmarking, this conditional does not impact performance, at least on recent x86 CPU
+  if(z>RealScalar(0))
+    return n / numext::sqrt(z);
+  else
+    return n;
 }
 
 /** Normalizes the vector, i.e. divides it by its own norm.
   *
   * \only_for_vectors
   *
+  * \warning If the input vector is too small (i.e., this->norm()==0), then \c *this is left unchanged.
+  *
   * \sa norm(), normalized()
   */
 template<typename Derived>
-inline void MatrixBase<Derived>::normalize()
+EIGEN_DEVICE_FUNC inline void MatrixBase<Derived>::normalize()
 {
-  *this /= norm();
+  RealScalar z = squaredNorm();
+  // NOTE: after extensive benchmarking, this conditional does not impact performance, at least on recent x86 CPU
+  if(z>RealScalar(0))
+    derived() /= numext::sqrt(z);
+}
+
+/** \returns an expression of the quotient of \c *this by its own norm while avoiding underflow and overflow.
+  *
+  * \only_for_vectors
+  *
+  * This method is analogue to the normalized() method, but it reduces the risk of
+  * underflow and overflow when computing the norm.
+  *
+  * \warning If the input vector is too small (i.e., this->norm()==0),
+  *          then this function returns a copy of the input.
+  *
+  * \sa stableNorm(), stableNormalize(), normalized()
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC inline const typename MatrixBase<Derived>::PlainObject
+MatrixBase<Derived>::stableNormalized() const
+{
+  typedef typename internal::nested_eval<Derived,3>::type _Nested;
+  _Nested n(derived());
+  RealScalar w = n.cwiseAbs().maxCoeff();
+  RealScalar z = (n/w).squaredNorm();
+  if(z>RealScalar(0))
+    return n / (numext::sqrt(z)*w);
+  else
+    return n;
+}
+
+/** Normalizes the vector while avoid underflow and overflow
+  *
+  * \only_for_vectors
+  *
+  * This method is analogue to the normalize() method, but it reduces the risk of
+  * underflow and overflow when computing the norm.
+  *
+  * \warning If the input vector is too small (i.e., this->norm()==0), then \c *this is left unchanged.
+  *
+  * \sa stableNorm(), stableNormalized(), normalize()
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC inline void MatrixBase<Derived>::stableNormalize()
+{
+  RealScalar w = cwiseAbs().maxCoeff();
+  RealScalar z = (derived()/w).squaredNorm();
+  if(z>RealScalar(0))
+    derived() /= numext::sqrt(z)*w;
 }
 
 //---------- implementation of other norms ----------
@@ -164,9 +201,10 @@ template<typename Derived, int p>
 struct lpNorm_selector
 {
   typedef typename NumTraits<typename traits<Derived>::Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
   static inline RealScalar run(const MatrixBase<Derived>& m)
   {
-    using std::pow;
+    EIGEN_USING_STD_MATH(pow)
     return pow(m.cwiseAbs().array().pow(p).sum(), RealScalar(1)/p);
   }
 };
@@ -174,6 +212,7 @@ struct lpNorm_selector
 template<typename Derived>
 struct lpNorm_selector<Derived, 1>
 {
+  EIGEN_DEVICE_FUNC
   static inline typename NumTraits<typename traits<Derived>::Scalar>::Real run(const MatrixBase<Derived>& m)
   {
     return m.cwiseAbs().sum();
@@ -183,6 +222,7 @@ struct lpNorm_selector<Derived, 1>
 template<typename Derived>
 struct lpNorm_selector<Derived, 2>
 {
+  EIGEN_DEVICE_FUNC
   static inline typename NumTraits<typename traits<Derived>::Scalar>::Real run(const MatrixBase<Derived>& m)
   {
     return m.norm();
@@ -192,23 +232,35 @@ struct lpNorm_selector<Derived, 2>
 template<typename Derived>
 struct lpNorm_selector<Derived, Infinity>
 {
-  static inline typename NumTraits<typename traits<Derived>::Scalar>::Real run(const MatrixBase<Derived>& m)
+  typedef typename NumTraits<typename traits<Derived>::Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
+  static inline RealScalar run(const MatrixBase<Derived>& m)
   {
+    if(Derived::SizeAtCompileTime==0 || (Derived::SizeAtCompileTime==Dynamic && m.size()==0))
+      return RealScalar(0);
     return m.cwiseAbs().maxCoeff();
   }
 };
 
 } // end namespace internal
 
-/** \returns the \f$ \ell^p \f$ norm of *this, that is, returns the p-th root of the sum of the p-th powers of the absolute values
-  *          of the coefficients of *this. If \a p is the special value \a Eigen::Infinity, this function returns the \f$ \ell^\infty \f$
-  *          norm, that is the maximum of the absolute values of the coefficients of *this.
+/** \returns the \b coefficient-wise \f$ \ell^p \f$ norm of \c *this, that is, returns the p-th root of the sum of the p-th powers of the absolute values
+  *          of the coefficients of \c *this. If \a p is the special value \a Eigen::Infinity, this function returns the \f$ \ell^\infty \f$
+  *          norm, that is the maximum of the absolute values of the coefficients of \c *this.
+  *
+  * In all cases, if \c *this is empty, then the value 0 is returned.
+  *
+  * \note For matrices, this function does not compute the <a href="https://en.wikipedia.org/wiki/Operator_norm">operator-norm</a>. That is, if \c *this is a matrix, then its coefficients are interpreted as a 1D vector. Nonetheless, you can easily compute the 1-norm and \f$\infty\f$-norm matrix operator norms using \link TutorialReductionsVisitorsBroadcastingReductionsNorm partial reductions \endlink.
   *
   * \sa norm()
   */
 template<typename Derived>
 template<int p>
-inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+EIGEN_DEVICE_FUNC inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real
+#else
+EIGEN_DEVICE_FUNC MatrixBase<Derived>::RealScalar
+#endif
 MatrixBase<Derived>::lpNorm() const
 {
   return internal::lpNorm_selector<Derived, p>::run(*this);
@@ -227,8 +279,8 @@ template<typename OtherDerived>
 bool MatrixBase<Derived>::isOrthogonal
 (const MatrixBase<OtherDerived>& other, const RealScalar& prec) const
 {
-  typename internal::nested<Derived,2>::type nested(derived());
-  typename internal::nested<OtherDerived,2>::type otherNested(other.derived());
+  typename internal::nested_eval<Derived,2>::type nested(derived());
+  typename internal::nested_eval<OtherDerived,2>::type otherNested(other.derived());
   return numext::abs2(nested.dot(otherNested)) <= prec * prec * nested.squaredNorm() * otherNested.squaredNorm();
 }
 
@@ -246,13 +298,13 @@ bool MatrixBase<Derived>::isOrthogonal
 template<typename Derived>
 bool MatrixBase<Derived>::isUnitary(const RealScalar& prec) const
 {
-  typename Derived::Nested nested(derived());
+  typename internal::nested_eval<Derived,1>::type self(derived());
   for(Index i = 0; i < cols(); ++i)
   {
-    if(!internal::isApprox(nested.col(i).squaredNorm(), static_cast<RealScalar>(1), prec))
+    if(!internal::isApprox(self.col(i).squaredNorm(), static_cast<RealScalar>(1), prec))
       return false;
     for(Index j = 0; j < i; ++j)
-      if(!internal::isMuchSmallerThan(nested.col(i).dot(nested.col(j)), static_cast<Scalar>(1), prec))
+      if(!internal::isMuchSmallerThan(self.col(i).dot(self.col(j)), static_cast<Scalar>(1), prec))
         return false;
   }
   return true;
diff --git a/eigen/Eigen/src/Core/EigenBase.h b/eigen/Eigen/src/Core/EigenBase.h
index fadb458..ccc122c 100644
--- a/eigen/Eigen/src/Core/EigenBase.h
+++ b/eigen/Eigen/src/Core/EigenBase.h
@@ -13,7 +13,9 @@
 
 namespace Eigen {
 
-/** Common base class for all classes T such that MatrixBase has an operator=(T) and a constructor MatrixBase(T).
+/** \class EigenBase
+  * 
+  * Common base class for all classes T such that MatrixBase has an operator=(T) and a constructor MatrixBase(T).
   *
   * In other words, an EigenBase object is an object that can be copied into a MatrixBase.
   *
@@ -21,39 +23,57 @@ namespace Eigen {
   *
   * Notice that this class is trivial, it is only used to disambiguate overloaded functions.
   *
-  * \sa \ref TopicClassHierarchy
+  * \sa \blank \ref TopicClassHierarchy
   */
 template<typename Derived> struct EigenBase
 {
 //   typedef typename internal::plain_matrix_type<Derived>::type PlainObject;
-
+  
+  /** \brief The interface type of indices
+    * \details To change this, \c \#define the preprocessor symbol \c EIGEN_DEFAULT_DENSE_INDEX_TYPE.
+    * \deprecated Since Eigen 3.3, its usage is deprecated. Use Eigen::Index instead.
+    * \sa StorageIndex, \ref TopicPreprocessorDirectives.
+    */
+  typedef Eigen::Index Index;
+
+  // FIXME is it needed?
   typedef typename internal::traits<Derived>::StorageKind StorageKind;
-  typedef typename internal::traits<Derived>::Index Index;
 
   /** \returns a reference to the derived object */
+  EIGEN_DEVICE_FUNC
   Derived& derived() { return *static_cast<Derived*>(this); }
   /** \returns a const reference to the derived object */
+  EIGEN_DEVICE_FUNC
   const Derived& derived() const { return *static_cast<const Derived*>(this); }
 
+  EIGEN_DEVICE_FUNC
   inline Derived& const_cast_derived() const
   { return *static_cast<Derived*>(const_cast<EigenBase*>(this)); }
+  EIGEN_DEVICE_FUNC
   inline const Derived& const_derived() const
   { return *static_cast<const Derived*>(this); }
 
   /** \returns the number of rows. \sa cols(), RowsAtCompileTime */
+  EIGEN_DEVICE_FUNC
   inline Index rows() const { return derived().rows(); }
   /** \returns the number of columns. \sa rows(), ColsAtCompileTime*/
+  EIGEN_DEVICE_FUNC
   inline Index cols() const { return derived().cols(); }
   /** \returns the number of coefficients, which is rows()*cols().
     * \sa rows(), cols(), SizeAtCompileTime. */
+  EIGEN_DEVICE_FUNC
   inline Index size() const { return rows() * cols(); }
 
   /** \internal Don't use it, but do the equivalent: \code dst = *this; \endcode */
-  template<typename Dest> inline void evalTo(Dest& dst) const
+  template<typename Dest>
+  EIGEN_DEVICE_FUNC
+  inline void evalTo(Dest& dst) const
   { derived().evalTo(dst); }
 
   /** \internal Don't use it, but do the equivalent: \code dst += *this; \endcode */
-  template<typename Dest> inline void addTo(Dest& dst) const
+  template<typename Dest>
+  EIGEN_DEVICE_FUNC
+  inline void addTo(Dest& dst) const
   {
     // This is the default implementation,
     // derived class can reimplement it in a more optimized way.
@@ -63,7 +83,9 @@ template<typename Derived> struct EigenBase
   }
 
   /** \internal Don't use it, but do the equivalent: \code dst -= *this; \endcode */
-  template<typename Dest> inline void subTo(Dest& dst) const
+  template<typename Dest>
+  EIGEN_DEVICE_FUNC
+  inline void subTo(Dest& dst) const
   {
     // This is the default implementation,
     // derived class can reimplement it in a more optimized way.
@@ -73,7 +95,8 @@ template<typename Derived> struct EigenBase
   }
 
   /** \internal Don't use it, but do the equivalent: \code dst.applyOnTheRight(*this); \endcode */
-  template<typename Dest> inline void applyThisOnTheRight(Dest& dst) const
+  template<typename Dest>
+  EIGEN_DEVICE_FUNC inline void applyThisOnTheRight(Dest& dst) const
   {
     // This is the default implementation,
     // derived class can reimplement it in a more optimized way.
@@ -81,7 +104,8 @@ template<typename Derived> struct EigenBase
   }
 
   /** \internal Don't use it, but do the equivalent: \code dst.applyOnTheLeft(*this); \endcode */
-  template<typename Dest> inline void applyThisOnTheLeft(Dest& dst) const
+  template<typename Dest>
+  EIGEN_DEVICE_FUNC inline void applyThisOnTheLeft(Dest& dst) const
   {
     // This is the default implementation,
     // derived class can reimplement it in a more optimized way.
@@ -104,25 +128,28 @@ template<typename Derived> struct EigenBase
   */
 template<typename Derived>
 template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
 Derived& DenseBase<Derived>::operator=(const EigenBase<OtherDerived> &other)
 {
-  other.derived().evalTo(derived());
+  call_assignment(derived(), other.derived());
   return derived();
 }
 
 template<typename Derived>
 template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
 Derived& DenseBase<Derived>::operator+=(const EigenBase<OtherDerived> &other)
 {
-  other.derived().addTo(derived());
+  call_assignment(derived(), other.derived(), internal::add_assign_op<Scalar,typename OtherDerived::Scalar>());
   return derived();
 }
 
 template<typename Derived>
 template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
 Derived& DenseBase<Derived>::operator-=(const EigenBase<OtherDerived> &other)
 {
-  other.derived().subTo(derived());
+  call_assignment(derived(), other.derived(), internal::sub_assign_op<Scalar,typename OtherDerived::Scalar>());
   return derived();
 }
 
diff --git a/eigen/Eigen/src/Core/Flagged.h b/eigen/Eigen/src/Core/Flagged.h
deleted file mode 100644
index 1f2955f..0000000
--- a/eigen/Eigen/src/Core/Flagged.h
+++ /dev/null
@@ -1,140 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2008 Benoit Jacob <jacob.benoit.1@gmail.com>
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-#ifndef EIGEN_FLAGGED_H
-#define EIGEN_FLAGGED_H
-
-namespace Eigen { 
-
-/** \class Flagged
-  * \ingroup Core_Module
-  *
-  * \brief Expression with modified flags
-  *
-  * \param ExpressionType the type of the object of which we are modifying the flags
-  * \param Added the flags added to the expression
-  * \param Removed the flags removed from the expression (has priority over Added).
-  *
-  * This class represents an expression whose flags have been modified.
-  * It is the return type of MatrixBase::flagged()
-  * and most of the time this is the only way it is used.
-  *
-  * \sa MatrixBase::flagged()
-  */
-
-namespace internal {
-template<typename ExpressionType, unsigned int Added, unsigned int Removed>
-struct traits<Flagged<ExpressionType, Added, Removed> > : traits<ExpressionType>
-{
-  enum { Flags = (ExpressionType::Flags | Added) & ~Removed };
-};
-}
-
-template<typename ExpressionType, unsigned int Added, unsigned int Removed> class Flagged
-  : public MatrixBase<Flagged<ExpressionType, Added, Removed> >
-{
-  public:
-
-    typedef MatrixBase<Flagged> Base;
-    
-    EIGEN_DENSE_PUBLIC_INTERFACE(Flagged)
-    typedef typename internal::conditional<internal::must_nest_by_value<ExpressionType>::ret,
-        ExpressionType, const ExpressionType&>::type ExpressionTypeNested;
-    typedef typename ExpressionType::InnerIterator InnerIterator;
-
-    inline Flagged(const ExpressionType& matrix) : m_matrix(matrix) {}
-
-    inline Index rows() const { return m_matrix.rows(); }
-    inline Index cols() const { return m_matrix.cols(); }
-    inline Index outerStride() const { return m_matrix.outerStride(); }
-    inline Index innerStride() const { return m_matrix.innerStride(); }
-
-    inline CoeffReturnType coeff(Index row, Index col) const
-    {
-      return m_matrix.coeff(row, col);
-    }
-
-    inline CoeffReturnType coeff(Index index) const
-    {
-      return m_matrix.coeff(index);
-    }
-    
-    inline const Scalar& coeffRef(Index row, Index col) const
-    {
-      return m_matrix.const_cast_derived().coeffRef(row, col);
-    }
-
-    inline const Scalar& coeffRef(Index index) const
-    {
-      return m_matrix.const_cast_derived().coeffRef(index);
-    }
-
-    inline Scalar& coeffRef(Index row, Index col)
-    {
-      return m_matrix.const_cast_derived().coeffRef(row, col);
-    }
-
-    inline Scalar& coeffRef(Index index)
-    {
-      return m_matrix.const_cast_derived().coeffRef(index);
-    }
-
-    template<int LoadMode>
-    inline const PacketScalar packet(Index row, Index col) const
-    {
-      return m_matrix.template packet<LoadMode>(row, col);
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index row, Index col, const PacketScalar& x)
-    {
-      m_matrix.const_cast_derived().template writePacket<LoadMode>(row, col, x);
-    }
-
-    template<int LoadMode>
-    inline const PacketScalar packet(Index index) const
-    {
-      return m_matrix.template packet<LoadMode>(index);
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index index, const PacketScalar& x)
-    {
-      m_matrix.const_cast_derived().template writePacket<LoadMode>(index, x);
-    }
-
-    const ExpressionType& _expression() const { return m_matrix; }
-
-    template<typename OtherDerived>
-    typename ExpressionType::PlainObject solveTriangular(const MatrixBase<OtherDerived>& other) const;
-
-    template<typename OtherDerived>
-    void solveTriangularInPlace(const MatrixBase<OtherDerived>& other) const;
-
-  protected:
-    ExpressionTypeNested m_matrix;
-};
-
-/** \returns an expression of *this with added and removed flags
-  *
-  * This is mostly for internal use.
-  *
-  * \sa class Flagged
-  */
-template<typename Derived>
-template<unsigned int Added,unsigned int Removed>
-inline const Flagged<Derived, Added, Removed>
-DenseBase<Derived>::flagged() const
-{
-  return derived();
-}
-
-} // end namespace Eigen
-
-#endif // EIGEN_FLAGGED_H
diff --git a/eigen/Eigen/src/Core/ForceAlignedAccess.h b/eigen/Eigen/src/Core/ForceAlignedAccess.h
index 807c7a2..7b08b45 100644
--- a/eigen/Eigen/src/Core/ForceAlignedAccess.h
+++ b/eigen/Eigen/src/Core/ForceAlignedAccess.h
@@ -39,29 +39,29 @@ template<typename ExpressionType> class ForceAlignedAccess
     typedef typename internal::dense_xpr_base<ForceAlignedAccess>::type Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(ForceAlignedAccess)
 
-    inline ForceAlignedAccess(const ExpressionType& matrix) : m_expression(matrix) {}
+    EIGEN_DEVICE_FUNC explicit inline ForceAlignedAccess(const ExpressionType& matrix) : m_expression(matrix) {}
 
-    inline Index rows() const { return m_expression.rows(); }
-    inline Index cols() const { return m_expression.cols(); }
-    inline Index outerStride() const { return m_expression.outerStride(); }
-    inline Index innerStride() const { return m_expression.innerStride(); }
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_expression.rows(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_expression.cols(); }
+    EIGEN_DEVICE_FUNC inline Index outerStride() const { return m_expression.outerStride(); }
+    EIGEN_DEVICE_FUNC inline Index innerStride() const { return m_expression.innerStride(); }
 
-    inline const CoeffReturnType coeff(Index row, Index col) const
+    EIGEN_DEVICE_FUNC inline const CoeffReturnType coeff(Index row, Index col) const
     {
       return m_expression.coeff(row, col);
     }
 
-    inline Scalar& coeffRef(Index row, Index col)
+    EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index row, Index col)
     {
       return m_expression.const_cast_derived().coeffRef(row, col);
     }
 
-    inline const CoeffReturnType coeff(Index index) const
+    EIGEN_DEVICE_FUNC inline const CoeffReturnType coeff(Index index) const
     {
       return m_expression.coeff(index);
     }
 
-    inline Scalar& coeffRef(Index index)
+    EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index index)
     {
       return m_expression.const_cast_derived().coeffRef(index);
     }
@@ -90,7 +90,7 @@ template<typename ExpressionType> class ForceAlignedAccess
       m_expression.const_cast_derived().template writePacket<Aligned>(index, x);
     }
 
-    operator const ExpressionType&() const { return m_expression; }
+    EIGEN_DEVICE_FUNC operator const ExpressionType&() const { return m_expression; }
 
   protected:
     const ExpressionType& m_expression;
@@ -127,7 +127,7 @@ template<bool Enable>
 inline typename internal::add_const_on_value_type<typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type>::type
 MatrixBase<Derived>::forceAlignedAccessIf() const
 {
-  return derived();
+  return derived();  // FIXME This should not work but apparently is never used
 }
 
 /** \returns an expression of *this with forced aligned access if \a Enable is true.
@@ -138,7 +138,7 @@ template<bool Enable>
 inline typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type
 MatrixBase<Derived>::forceAlignedAccessIf()
 {
-  return derived();
+  return derived();  // FIXME This should not work but apparently is never used
 }
 
 } // end namespace Eigen
diff --git a/eigen/Eigen/src/Core/Functors.h b/eigen/Eigen/src/Core/Functors.h
deleted file mode 100644
index 5f14c65..0000000
--- a/eigen/Eigen/src/Core/Functors.h
+++ /dev/null
@@ -1,1026 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-#ifndef EIGEN_FUNCTORS_H
-#define EIGEN_FUNCTORS_H
-
-namespace Eigen {
-
-namespace internal {
-
-// associative functors:
-
-/** \internal
-  * \brief Template functor to compute the sum of two scalars
-  *
-  * \sa class CwiseBinaryOp, MatrixBase::operator+, class VectorwiseOp, MatrixBase::sum()
-  */
-template<typename Scalar> struct scalar_sum_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_sum_op)
-  EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { return a + b; }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
-  { return internal::padd(a,b); }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Scalar predux(const Packet& a) const
-  { return internal::predux(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_sum_op<Scalar> > {
-  enum {
-    Cost = NumTraits<Scalar>::AddCost,
-    PacketAccess = packet_traits<Scalar>::HasAdd
-  };
-};
-
-/** \internal
-  * \brief Template functor to compute the product of two scalars
-  *
-  * \sa class CwiseBinaryOp, Cwise::operator*(), class VectorwiseOp, MatrixBase::redux()
-  */
-template<typename LhsScalar,typename RhsScalar> struct scalar_product_op {
-  enum {
-    // TODO vectorize mixed product
-    Vectorizable = is_same<LhsScalar,RhsScalar>::value && packet_traits<LhsScalar>::HasMul && packet_traits<RhsScalar>::HasMul
-  };
-  typedef typename scalar_product_traits<LhsScalar,RhsScalar>::ReturnType result_type;
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_product_op)
-  EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a * b; }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
-  { return internal::pmul(a,b); }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const result_type predux(const Packet& a) const
-  { return internal::predux_mul(a); }
-};
-template<typename LhsScalar,typename RhsScalar>
-struct functor_traits<scalar_product_op<LhsScalar,RhsScalar> > {
-  enum {
-    Cost = (NumTraits<LhsScalar>::MulCost + NumTraits<RhsScalar>::MulCost)/2, // rough estimate!
-    PacketAccess = scalar_product_op<LhsScalar,RhsScalar>::Vectorizable
-  };
-};
-
-/** \internal
-  * \brief Template functor to compute the conjugate product of two scalars
-  *
-  * This is a short cut for conj(x) * y which is needed for optimization purpose; in Eigen2 support mode, this becomes x * conj(y)
-  */
-template<typename LhsScalar,typename RhsScalar> struct scalar_conj_product_op {
-
-  enum {
-    Conj = NumTraits<LhsScalar>::IsComplex
-  };
-  
-  typedef typename scalar_product_traits<LhsScalar,RhsScalar>::ReturnType result_type;
-  
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_conj_product_op)
-  EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const
-  { return conj_helper<LhsScalar,RhsScalar,Conj,false>().pmul(a,b); }
-  
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
-  { return conj_helper<Packet,Packet,Conj,false>().pmul(a,b); }
-};
-template<typename LhsScalar,typename RhsScalar>
-struct functor_traits<scalar_conj_product_op<LhsScalar,RhsScalar> > {
-  enum {
-    Cost = NumTraits<LhsScalar>::MulCost,
-    PacketAccess = internal::is_same<LhsScalar, RhsScalar>::value && packet_traits<LhsScalar>::HasMul
-  };
-};
-
-/** \internal
-  * \brief Template functor to compute the min of two scalars
-  *
-  * \sa class CwiseBinaryOp, MatrixBase::cwiseMin, class VectorwiseOp, MatrixBase::minCoeff()
-  */
-template<typename Scalar> struct scalar_min_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_min_op)
-  EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { using std::min; return (min)(a, b); }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
-  { return internal::pmin(a,b); }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Scalar predux(const Packet& a) const
-  { return internal::predux_min(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_min_op<Scalar> > {
-  enum {
-    Cost = NumTraits<Scalar>::AddCost,
-    PacketAccess = packet_traits<Scalar>::HasMin
-  };
-};
-
-/** \internal
-  * \brief Template functor to compute the max of two scalars
-  *
-  * \sa class CwiseBinaryOp, MatrixBase::cwiseMax, class VectorwiseOp, MatrixBase::maxCoeff()
-  */
-template<typename Scalar> struct scalar_max_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_max_op)
-  EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { using std::max; return (max)(a, b); }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
-  { return internal::pmax(a,b); }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Scalar predux(const Packet& a) const
-  { return internal::predux_max(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_max_op<Scalar> > {
-  enum {
-    Cost = NumTraits<Scalar>::AddCost,
-    PacketAccess = packet_traits<Scalar>::HasMax
-  };
-};
-
-/** \internal
-  * \brief Template functor to compute the hypot of two scalars
-  *
-  * \sa MatrixBase::stableNorm(), class Redux
-  */
-template<typename Scalar> struct scalar_hypot_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_hypot_op)
-//   typedef typename NumTraits<Scalar>::Real result_type;
-  EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& _x, const Scalar& _y) const
-  {
-    using std::max;
-    using std::min;
-    using std::sqrt;
-    Scalar p = (max)(_x, _y);
-    Scalar q = (min)(_x, _y);
-    Scalar qp = q/p;
-    return p * sqrt(Scalar(1) + qp*qp);
-  }
-};
-template<typename Scalar>
-struct functor_traits<scalar_hypot_op<Scalar> > {
-  enum { Cost = 5 * NumTraits<Scalar>::MulCost, PacketAccess=0 };
-};
-
-/** \internal
-  * \brief Template functor to compute the pow of two scalars
-  */
-template<typename Scalar, typename OtherScalar> struct scalar_binary_pow_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_binary_pow_op)
-  inline Scalar operator() (const Scalar& a, const OtherScalar& b) const { return numext::pow(a, b); }
-};
-template<typename Scalar, typename OtherScalar>
-struct functor_traits<scalar_binary_pow_op<Scalar,OtherScalar> > {
-  enum { Cost = 5 * NumTraits<Scalar>::MulCost, PacketAccess = false };
-};
-
-// other binary functors:
-
-/** \internal
-  * \brief Template functor to compute the difference of two scalars
-  *
-  * \sa class CwiseBinaryOp, MatrixBase::operator-
-  */
-template<typename Scalar> struct scalar_difference_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_difference_op)
-  EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { return a - b; }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
-  { return internal::psub(a,b); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_difference_op<Scalar> > {
-  enum {
-    Cost = NumTraits<Scalar>::AddCost,
-    PacketAccess = packet_traits<Scalar>::HasSub
-  };
-};
-
-/** \internal
-  * \brief Template functor to compute the quotient of two scalars
-  *
-  * \sa class CwiseBinaryOp, Cwise::operator/()
-  */
-template<typename LhsScalar,typename RhsScalar> struct scalar_quotient_op {
-  enum {
-    // TODO vectorize mixed product
-    Vectorizable = is_same<LhsScalar,RhsScalar>::value && packet_traits<LhsScalar>::HasDiv && packet_traits<RhsScalar>::HasDiv
-  };
-  typedef typename scalar_product_traits<LhsScalar,RhsScalar>::ReturnType result_type;
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_quotient_op)
-  EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a / b; }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
-  { return internal::pdiv(a,b); }
-};
-template<typename LhsScalar,typename RhsScalar>
-struct functor_traits<scalar_quotient_op<LhsScalar,RhsScalar> > {
-  enum {
-    Cost = (NumTraits<LhsScalar>::MulCost + NumTraits<RhsScalar>::MulCost), // rough estimate!
-    PacketAccess = scalar_quotient_op<LhsScalar,RhsScalar>::Vectorizable
-  };
-};
-
-
-
-/** \internal
-  * \brief Template functor to compute the and of two booleans
-  *
-  * \sa class CwiseBinaryOp, ArrayBase::operator&&
-  */
-struct scalar_boolean_and_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_boolean_and_op)
-  EIGEN_STRONG_INLINE bool operator() (const bool& a, const bool& b) const { return a && b; }
-};
-template<> struct functor_traits<scalar_boolean_and_op> {
-  enum {
-    Cost = NumTraits<bool>::AddCost,
-    PacketAccess = false
-  };
-};
-
-/** \internal
-  * \brief Template functor to compute the or of two booleans
-  *
-  * \sa class CwiseBinaryOp, ArrayBase::operator||
-  */
-struct scalar_boolean_or_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_boolean_or_op)
-  EIGEN_STRONG_INLINE bool operator() (const bool& a, const bool& b) const { return a || b; }
-};
-template<> struct functor_traits<scalar_boolean_or_op> {
-  enum {
-    Cost = NumTraits<bool>::AddCost,
-    PacketAccess = false
-  };
-};
-
-/** \internal
-  * \brief Template functors for comparison of two scalars
-  * \todo Implement packet-comparisons
-  */
-template<typename Scalar, ComparisonName cmp> struct scalar_cmp_op;
-
-template<typename Scalar, ComparisonName cmp>
-struct functor_traits<scalar_cmp_op<Scalar, cmp> > {
-  enum {
-    Cost = NumTraits<Scalar>::AddCost,
-    PacketAccess = false
-  };
-};
-
-template<ComparisonName Cmp, typename Scalar>
-struct result_of<scalar_cmp_op<Scalar, Cmp>(Scalar,Scalar)> {
-  typedef bool type;
-};
-
-
-template<typename Scalar> struct scalar_cmp_op<Scalar, cmp_EQ> {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
-  EIGEN_STRONG_INLINE bool operator()(const Scalar& a, const Scalar& b) const {return a==b;}
-};
-template<typename Scalar> struct scalar_cmp_op<Scalar, cmp_LT> {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
-  EIGEN_STRONG_INLINE bool operator()(const Scalar& a, const Scalar& b) const {return a<b;}
-};
-template<typename Scalar> struct scalar_cmp_op<Scalar, cmp_LE> {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
-  EIGEN_STRONG_INLINE bool operator()(const Scalar& a, const Scalar& b) const {return a<=b;}
-};
-template<typename Scalar> struct scalar_cmp_op<Scalar, cmp_UNORD> {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
-  EIGEN_STRONG_INLINE bool operator()(const Scalar& a, const Scalar& b) const {return !(a<=b || b<=a);}
-};
-template<typename Scalar> struct scalar_cmp_op<Scalar, cmp_NEQ> {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
-  EIGEN_STRONG_INLINE bool operator()(const Scalar& a, const Scalar& b) const {return a!=b;}
-};
-
-// unary functors:
-
-/** \internal
-  * \brief Template functor to compute the opposite of a scalar
-  *
-  * \sa class CwiseUnaryOp, MatrixBase::operator-
-  */
-template<typename Scalar> struct scalar_opposite_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_opposite_op)
-  EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { return -a; }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
-  { return internal::pnegate(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_opposite_op<Scalar> >
-{ enum {
-    Cost = NumTraits<Scalar>::AddCost,
-    PacketAccess = packet_traits<Scalar>::HasNegate };
-};
-
-/** \internal
-  * \brief Template functor to compute the absolute value of a scalar
-  *
-  * \sa class CwiseUnaryOp, Cwise::abs
-  */
-template<typename Scalar> struct scalar_abs_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_abs_op)
-  typedef typename NumTraits<Scalar>::Real result_type;
-  EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a) const { using std::abs; return abs(a); }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
-  { return internal::pabs(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_abs_op<Scalar> >
-{
-  enum {
-    Cost = NumTraits<Scalar>::AddCost,
-    PacketAccess = packet_traits<Scalar>::HasAbs
-  };
-};
-
-/** \internal
-  * \brief Template functor to compute the squared absolute value of a scalar
-  *
-  * \sa class CwiseUnaryOp, Cwise::abs2
-  */
-template<typename Scalar> struct scalar_abs2_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_abs2_op)
-  typedef typename NumTraits<Scalar>::Real result_type;
-  EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a) const { return numext::abs2(a); }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
-  { return internal::pmul(a,a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_abs2_op<Scalar> >
-{ enum { Cost = NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasAbs2 }; };
-
-/** \internal
-  * \brief Template functor to compute the conjugate of a complex value
-  *
-  * \sa class CwiseUnaryOp, MatrixBase::conjugate()
-  */
-template<typename Scalar> struct scalar_conjugate_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_conjugate_op)
-  EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { using numext::conj; return conj(a); }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const { return internal::pconj(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_conjugate_op<Scalar> >
-{
-  enum {
-    Cost = NumTraits<Scalar>::IsComplex ? NumTraits<Scalar>::AddCost : 0,
-    PacketAccess = packet_traits<Scalar>::HasConj
-  };
-};
-
-/** \internal
-  * \brief Template functor to cast a scalar to another type
-  *
-  * \sa class CwiseUnaryOp, MatrixBase::cast()
-  */
-template<typename Scalar, typename NewType>
-struct scalar_cast_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
-  typedef NewType result_type;
-  EIGEN_STRONG_INLINE const NewType operator() (const Scalar& a) const { return cast<Scalar, NewType>(a); }
-};
-template<typename Scalar, typename NewType>
-struct functor_traits<scalar_cast_op<Scalar,NewType> >
-{ enum { Cost = is_same<Scalar, NewType>::value ? 0 : NumTraits<NewType>::AddCost, PacketAccess = false }; };
-
-/** \internal
-  * \brief Template functor to extract the real part of a complex
-  *
-  * \sa class CwiseUnaryOp, MatrixBase::real()
-  */
-template<typename Scalar>
-struct scalar_real_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_real_op)
-  typedef typename NumTraits<Scalar>::Real result_type;
-  EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const { return numext::real(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_real_op<Scalar> >
-{ enum { Cost = 0, PacketAccess = false }; };
-
-/** \internal
-  * \brief Template functor to extract the imaginary part of a complex
-  *
-  * \sa class CwiseUnaryOp, MatrixBase::imag()
-  */
-template<typename Scalar>
-struct scalar_imag_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_imag_op)
-  typedef typename NumTraits<Scalar>::Real result_type;
-  EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const { return numext::imag(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_imag_op<Scalar> >
-{ enum { Cost = 0, PacketAccess = false }; };
-
-/** \internal
-  * \brief Template functor to extract the real part of a complex as a reference
-  *
-  * \sa class CwiseUnaryOp, MatrixBase::real()
-  */
-template<typename Scalar>
-struct scalar_real_ref_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_real_ref_op)
-  typedef typename NumTraits<Scalar>::Real result_type;
-  EIGEN_STRONG_INLINE result_type& operator() (const Scalar& a) const { return numext::real_ref(*const_cast<Scalar*>(&a)); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_real_ref_op<Scalar> >
-{ enum { Cost = 0, PacketAccess = false }; };
-
-/** \internal
-  * \brief Template functor to extract the imaginary part of a complex as a reference
-  *
-  * \sa class CwiseUnaryOp, MatrixBase::imag()
-  */
-template<typename Scalar>
-struct scalar_imag_ref_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_imag_ref_op)
-  typedef typename NumTraits<Scalar>::Real result_type;
-  EIGEN_STRONG_INLINE result_type& operator() (const Scalar& a) const { return numext::imag_ref(*const_cast<Scalar*>(&a)); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_imag_ref_op<Scalar> >
-{ enum { Cost = 0, PacketAccess = false }; };
-
-/** \internal
-  *
-  * \brief Template functor to compute the exponential of a scalar
-  *
-  * \sa class CwiseUnaryOp, Cwise::exp()
-  */
-template<typename Scalar> struct scalar_exp_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_exp_op)
-  inline const Scalar operator() (const Scalar& a) const { using std::exp; return exp(a); }
-  typedef typename packet_traits<Scalar>::type Packet;
-  inline Packet packetOp(const Packet& a) const { return internal::pexp(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_exp_op<Scalar> >
-{ enum { Cost = 5 * NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasExp }; };
-
-/** \internal
-  *
-  * \brief Template functor to compute the logarithm of a scalar
-  *
-  * \sa class CwiseUnaryOp, Cwise::log()
-  */
-template<typename Scalar> struct scalar_log_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_log_op)
-  inline const Scalar operator() (const Scalar& a) const { using std::log; return log(a); }
-  typedef typename packet_traits<Scalar>::type Packet;
-  inline Packet packetOp(const Packet& a) const { return internal::plog(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_log_op<Scalar> >
-{ enum { Cost = 5 * NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasLog }; };
-
-/** \internal
-  * \brief Template functor to multiply a scalar by a fixed other one
-  *
-  * \sa class CwiseUnaryOp, MatrixBase::operator*, MatrixBase::operator/
-  */
-/* NOTE why doing the pset1() in packetOp *is* an optimization ?
- * indeed it seems better to declare m_other as a Packet and do the pset1() once
- * in the constructor. However, in practice:
- *  - GCC does not like m_other as a Packet and generate a load every time it needs it
- *  - on the other hand GCC is able to moves the pset1() outside the loop :)
- *  - simpler code ;)
- * (ICC and gcc 4.4 seems to perform well in both cases, the issue is visible with y = a*x + b*y)
- */
-template<typename Scalar>
-struct scalar_multiple_op {
-  typedef typename packet_traits<Scalar>::type Packet;
-  // FIXME default copy constructors seems bugged with std::complex<>
-  EIGEN_STRONG_INLINE scalar_multiple_op(const scalar_multiple_op& other) : m_other(other.m_other) { }
-  EIGEN_STRONG_INLINE scalar_multiple_op(const Scalar& other) : m_other(other) { }
-  EIGEN_STRONG_INLINE Scalar operator() (const Scalar& a) const { return a * m_other; }
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
-  { return internal::pmul(a, pset1<Packet>(m_other)); }
-  typename add_const_on_value_type<typename NumTraits<Scalar>::Nested>::type m_other;
-};
-template<typename Scalar>
-struct functor_traits<scalar_multiple_op<Scalar> >
-{ enum { Cost = NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasMul }; };
-
-template<typename Scalar1, typename Scalar2>
-struct scalar_multiple2_op {
-  typedef typename scalar_product_traits<Scalar1,Scalar2>::ReturnType result_type;
-  EIGEN_STRONG_INLINE scalar_multiple2_op(const scalar_multiple2_op& other) : m_other(other.m_other) { }
-  EIGEN_STRONG_INLINE scalar_multiple2_op(const Scalar2& other) : m_other(other) { }
-  EIGEN_STRONG_INLINE result_type operator() (const Scalar1& a) const { return a * m_other; }
-  typename add_const_on_value_type<typename NumTraits<Scalar2>::Nested>::type m_other;
-};
-template<typename Scalar1,typename Scalar2>
-struct functor_traits<scalar_multiple2_op<Scalar1,Scalar2> >
-{ enum { Cost = NumTraits<Scalar1>::MulCost, PacketAccess = false }; };
-
-/** \internal
-  * \brief Template functor to divide a scalar by a fixed other one
-  *
-  * This functor is used to implement the quotient of a matrix by
-  * a scalar where the scalar type is not necessarily a floating point type.
-  *
-  * \sa class CwiseUnaryOp, MatrixBase::operator/
-  */
-template<typename Scalar>
-struct scalar_quotient1_op {
-  typedef typename packet_traits<Scalar>::type Packet;
-  // FIXME default copy constructors seems bugged with std::complex<>
-  EIGEN_STRONG_INLINE scalar_quotient1_op(const scalar_quotient1_op& other) : m_other(other.m_other) { }
-  EIGEN_STRONG_INLINE scalar_quotient1_op(const Scalar& other) : m_other(other) {}
-  EIGEN_STRONG_INLINE Scalar operator() (const Scalar& a) const { return a / m_other; }
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
-  { return internal::pdiv(a, pset1<Packet>(m_other)); }
-  typename add_const_on_value_type<typename NumTraits<Scalar>::Nested>::type m_other;
-};
-template<typename Scalar>
-struct functor_traits<scalar_quotient1_op<Scalar> >
-{ enum { Cost = 2 * NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasDiv }; };
-
-// nullary functors
-
-template<typename Scalar>
-struct scalar_constant_op {
-  typedef typename packet_traits<Scalar>::type Packet;
-  EIGEN_STRONG_INLINE scalar_constant_op(const scalar_constant_op& other) : m_other(other.m_other) { }
-  EIGEN_STRONG_INLINE scalar_constant_op(const Scalar& other) : m_other(other) { }
-  template<typename Index>
-  EIGEN_STRONG_INLINE const Scalar operator() (Index, Index = 0) const { return m_other; }
-  template<typename Index>
-  EIGEN_STRONG_INLINE const Packet packetOp(Index, Index = 0) const { return internal::pset1<Packet>(m_other); }
-  const Scalar m_other;
-};
-template<typename Scalar>
-struct functor_traits<scalar_constant_op<Scalar> >
-// FIXME replace this packet test by a safe one
-{ enum { Cost = 1, PacketAccess = packet_traits<Scalar>::Vectorizable, IsRepeatable = true }; };
-
-template<typename Scalar> struct scalar_identity_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_identity_op)
-  template<typename Index>
-  EIGEN_STRONG_INLINE const Scalar operator() (Index row, Index col) const { return row==col ? Scalar(1) : Scalar(0); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_identity_op<Scalar> >
-{ enum { Cost = NumTraits<Scalar>::AddCost, PacketAccess = false, IsRepeatable = true }; };
-
-template <typename Scalar, bool RandomAccess> struct linspaced_op_impl;
-
-// linear access for packet ops:
-// 1) initialization
-//   base = [low, ..., low] + ([step, ..., step] * [-size, ..., 0])
-// 2) each step (where size is 1 for coeff access or PacketSize for packet access)
-//   base += [size*step, ..., size*step]
-//
-// TODO: Perhaps it's better to initialize lazily (so not in the constructor but in packetOp)
-//       in order to avoid the padd() in operator() ?
-template <typename Scalar>
-struct linspaced_op_impl<Scalar,false>
-{
-  typedef typename packet_traits<Scalar>::type Packet;
-
-  linspaced_op_impl(const Scalar& low, const Scalar& step) :
-  m_low(low), m_step(step),
-  m_packetStep(pset1<Packet>(packet_traits<Scalar>::size*step)),
-  m_base(padd(pset1<Packet>(low), pmul(pset1<Packet>(step),plset<Scalar>(-packet_traits<Scalar>::size)))) {}
-
-  template<typename Index>
-  EIGEN_STRONG_INLINE const Scalar operator() (Index i) const 
-  { 
-    m_base = padd(m_base, pset1<Packet>(m_step));
-    return m_low+Scalar(i)*m_step; 
-  }
-
-  template<typename Index>
-  EIGEN_STRONG_INLINE const Packet packetOp(Index) const { return m_base = padd(m_base,m_packetStep); }
-
-  const Scalar m_low;
-  const Scalar m_step;
-  const Packet m_packetStep;
-  mutable Packet m_base;
-};
-
-// random access for packet ops:
-// 1) each step
-//   [low, ..., low] + ( [step, ..., step] * ( [i, ..., i] + [0, ..., size] ) )
-template <typename Scalar>
-struct linspaced_op_impl<Scalar,true>
-{
-  typedef typename packet_traits<Scalar>::type Packet;
-
-  linspaced_op_impl(const Scalar& low, const Scalar& step) :
-  m_low(low), m_step(step),
-  m_lowPacket(pset1<Packet>(m_low)), m_stepPacket(pset1<Packet>(m_step)), m_interPacket(plset<Scalar>(0)) {}
-
-  template<typename Index>
-  EIGEN_STRONG_INLINE const Scalar operator() (Index i) const { return m_low+i*m_step; }
-
-  template<typename Index>
-  EIGEN_STRONG_INLINE const Packet packetOp(Index i) const
-  { return internal::padd(m_lowPacket, pmul(m_stepPacket, padd(pset1<Packet>(Scalar(i)),m_interPacket))); }
-
-  const Scalar m_low;
-  const Scalar m_step;
-  const Packet m_lowPacket;
-  const Packet m_stepPacket;
-  const Packet m_interPacket;
-};
-
-// ----- Linspace functor ----------------------------------------------------------------
-
-// Forward declaration (we default to random access which does not really give
-// us a speed gain when using packet access but it allows to use the functor in
-// nested expressions).
-template <typename Scalar, bool RandomAccess = true> struct linspaced_op;
-template <typename Scalar, bool RandomAccess> struct functor_traits< linspaced_op<Scalar,RandomAccess> >
-{ enum { Cost = 1, PacketAccess = packet_traits<Scalar>::HasSetLinear, IsRepeatable = true }; };
-template <typename Scalar, bool RandomAccess> struct linspaced_op
-{
-  typedef typename packet_traits<Scalar>::type Packet;
-  linspaced_op(const Scalar& low, const Scalar& high, DenseIndex num_steps) : impl((num_steps==1 ? high : low), (num_steps==1 ? Scalar() : (high-low)/Scalar(num_steps-1))) {}
-
-  template<typename Index>
-  EIGEN_STRONG_INLINE const Scalar operator() (Index i) const { return impl(i); }
-
-  // We need this function when assigning e.g. a RowVectorXd to a MatrixXd since
-  // there row==0 and col is used for the actual iteration.
-  template<typename Index>
-  EIGEN_STRONG_INLINE const Scalar operator() (Index row, Index col) const 
-  {
-    eigen_assert(col==0 || row==0);
-    return impl(col + row);
-  }
-
-  template<typename Index>
-  EIGEN_STRONG_INLINE const Packet packetOp(Index i) const { return impl.packetOp(i); }
-
-  // We need this function when assigning e.g. a RowVectorXd to a MatrixXd since
-  // there row==0 and col is used for the actual iteration.
-  template<typename Index>
-  EIGEN_STRONG_INLINE const Packet packetOp(Index row, Index col) const
-  {
-    eigen_assert(col==0 || row==0);
-    return impl.packetOp(col + row);
-  }
-
-  // This proxy object handles the actual required temporaries, the different
-  // implementations (random vs. sequential access) as well as the
-  // correct piping to size 2/4 packet operations.
-  const linspaced_op_impl<Scalar,RandomAccess> impl;
-};
-
-// all functors allow linear access, except scalar_identity_op. So we fix here a quick meta
-// to indicate whether a functor allows linear access, just always answering 'yes' except for
-// scalar_identity_op.
-// FIXME move this to functor_traits adding a functor_default
-template<typename Functor> struct functor_has_linear_access { enum { ret = 1 }; };
-template<typename Scalar> struct functor_has_linear_access<scalar_identity_op<Scalar> > { enum { ret = 0 }; };
-
-// In Eigen, any binary op (Product, CwiseBinaryOp) require the Lhs and Rhs to have the same scalar type, except for multiplication
-// where the mixing of different types is handled by scalar_product_traits
-// In particular, real * complex<real> is allowed.
-// FIXME move this to functor_traits adding a functor_default
-template<typename Functor> struct functor_is_product_like { enum { ret = 0 }; };
-template<typename LhsScalar,typename RhsScalar> struct functor_is_product_like<scalar_product_op<LhsScalar,RhsScalar> > { enum { ret = 1 }; };
-template<typename LhsScalar,typename RhsScalar> struct functor_is_product_like<scalar_conj_product_op<LhsScalar,RhsScalar> > { enum { ret = 1 }; };
-template<typename LhsScalar,typename RhsScalar> struct functor_is_product_like<scalar_quotient_op<LhsScalar,RhsScalar> > { enum { ret = 1 }; };
-
-
-/** \internal
-  * \brief Template functor to add a scalar to a fixed other one
-  * \sa class CwiseUnaryOp, Array::operator+
-  */
-/* If you wonder why doing the pset1() in packetOp() is an optimization check scalar_multiple_op */
-template<typename Scalar>
-struct scalar_add_op {
-  typedef typename packet_traits<Scalar>::type Packet;
-  // FIXME default copy constructors seems bugged with std::complex<>
-  inline scalar_add_op(const scalar_add_op& other) : m_other(other.m_other) { }
-  inline scalar_add_op(const Scalar& other) : m_other(other) { }
-  inline Scalar operator() (const Scalar& a) const { return a + m_other; }
-  inline const Packet packetOp(const Packet& a) const
-  { return internal::padd(a, pset1<Packet>(m_other)); }
-  const Scalar m_other;
-};
-template<typename Scalar>
-struct functor_traits<scalar_add_op<Scalar> >
-{ enum { Cost = NumTraits<Scalar>::AddCost, PacketAccess = packet_traits<Scalar>::HasAdd }; };
-
-/** \internal
-  * \brief Template functor to compute the square root of a scalar
-  * \sa class CwiseUnaryOp, Cwise::sqrt()
-  */
-template<typename Scalar> struct scalar_sqrt_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_sqrt_op)
-  inline const Scalar operator() (const Scalar& a) const { using std::sqrt; return sqrt(a); }
-  typedef typename packet_traits<Scalar>::type Packet;
-  inline Packet packetOp(const Packet& a) const { return internal::psqrt(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_sqrt_op<Scalar> >
-{ enum {
-    Cost = 5 * NumTraits<Scalar>::MulCost,
-    PacketAccess = packet_traits<Scalar>::HasSqrt
-  };
-};
-
-/** \internal
-  * \brief Template functor to compute the cosine of a scalar
-  * \sa class CwiseUnaryOp, ArrayBase::cos()
-  */
-template<typename Scalar> struct scalar_cos_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_cos_op)
-  inline Scalar operator() (const Scalar& a) const { using std::cos; return cos(a); }
-  typedef typename packet_traits<Scalar>::type Packet;
-  inline Packet packetOp(const Packet& a) const { return internal::pcos(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_cos_op<Scalar> >
-{
-  enum {
-    Cost = 5 * NumTraits<Scalar>::MulCost,
-    PacketAccess = packet_traits<Scalar>::HasCos
-  };
-};
-
-/** \internal
-  * \brief Template functor to compute the sine of a scalar
-  * \sa class CwiseUnaryOp, ArrayBase::sin()
-  */
-template<typename Scalar> struct scalar_sin_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_sin_op)
-  inline const Scalar operator() (const Scalar& a) const { using std::sin; return sin(a); }
-  typedef typename packet_traits<Scalar>::type Packet;
-  inline Packet packetOp(const Packet& a) const { return internal::psin(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_sin_op<Scalar> >
-{
-  enum {
-    Cost = 5 * NumTraits<Scalar>::MulCost,
-    PacketAccess = packet_traits<Scalar>::HasSin
-  };
-};
-
-
-/** \internal
-  * \brief Template functor to compute the tan of a scalar
-  * \sa class CwiseUnaryOp, ArrayBase::tan()
-  */
-template<typename Scalar> struct scalar_tan_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_tan_op)
-  inline const Scalar operator() (const Scalar& a) const { using std::tan; return tan(a); }
-  typedef typename packet_traits<Scalar>::type Packet;
-  inline Packet packetOp(const Packet& a) const { return internal::ptan(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_tan_op<Scalar> >
-{
-  enum {
-    Cost = 5 * NumTraits<Scalar>::MulCost,
-    PacketAccess = packet_traits<Scalar>::HasTan
-  };
-};
-
-/** \internal
-  * \brief Template functor to compute the arc cosine of a scalar
-  * \sa class CwiseUnaryOp, ArrayBase::acos()
-  */
-template<typename Scalar> struct scalar_acos_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_acos_op)
-  inline const Scalar operator() (const Scalar& a) const { using std::acos; return acos(a); }
-  typedef typename packet_traits<Scalar>::type Packet;
-  inline Packet packetOp(const Packet& a) const { return internal::pacos(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_acos_op<Scalar> >
-{
-  enum {
-    Cost = 5 * NumTraits<Scalar>::MulCost,
-    PacketAccess = packet_traits<Scalar>::HasACos
-  };
-};
-
-/** \internal
-  * \brief Template functor to compute the arc sine of a scalar
-  * \sa class CwiseUnaryOp, ArrayBase::asin()
-  */
-template<typename Scalar> struct scalar_asin_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_asin_op)
-  inline const Scalar operator() (const Scalar& a) const { using std::asin; return asin(a); }
-  typedef typename packet_traits<Scalar>::type Packet;
-  inline Packet packetOp(const Packet& a) const { return internal::pasin(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_asin_op<Scalar> >
-{
-  enum {
-    Cost = 5 * NumTraits<Scalar>::MulCost,
-    PacketAccess = packet_traits<Scalar>::HasASin
-  };
-};
-
-/** \internal
-  * \brief Template functor to raise a scalar to a power
-  * \sa class CwiseUnaryOp, Cwise::pow
-  */
-template<typename Scalar>
-struct scalar_pow_op {
-  // FIXME default copy constructors seems bugged with std::complex<>
-  inline scalar_pow_op(const scalar_pow_op& other) : m_exponent(other.m_exponent) { }
-  inline scalar_pow_op(const Scalar& exponent) : m_exponent(exponent) {}
-  inline Scalar operator() (const Scalar& a) const { return numext::pow(a, m_exponent); }
-  const Scalar m_exponent;
-};
-template<typename Scalar>
-struct functor_traits<scalar_pow_op<Scalar> >
-{ enum { Cost = 5 * NumTraits<Scalar>::MulCost, PacketAccess = false }; };
-
-/** \internal
-  * \brief Template functor to compute the quotient between a scalar and array entries.
-  * \sa class CwiseUnaryOp, Cwise::inverse()
-  */
-template<typename Scalar>
-struct scalar_inverse_mult_op {
-  scalar_inverse_mult_op(const Scalar& other) : m_other(other) {}
-  inline Scalar operator() (const Scalar& a) const { return m_other / a; }
-  template<typename Packet>
-  inline const Packet packetOp(const Packet& a) const
-  { return internal::pdiv(pset1<Packet>(m_other),a); }
-  Scalar m_other;
-};
-
-/** \internal
-  * \brief Template functor to compute the inverse of a scalar
-  * \sa class CwiseUnaryOp, Cwise::inverse()
-  */
-template<typename Scalar>
-struct scalar_inverse_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_inverse_op)
-  inline Scalar operator() (const Scalar& a) const { return Scalar(1)/a; }
-  template<typename Packet>
-  inline const Packet packetOp(const Packet& a) const
-  { return internal::pdiv(pset1<Packet>(Scalar(1)),a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_inverse_op<Scalar> >
-{ enum { Cost = NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasDiv }; };
-
-/** \internal
-  * \brief Template functor to compute the square of a scalar
-  * \sa class CwiseUnaryOp, Cwise::square()
-  */
-template<typename Scalar>
-struct scalar_square_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_square_op)
-  inline Scalar operator() (const Scalar& a) const { return a*a; }
-  template<typename Packet>
-  inline const Packet packetOp(const Packet& a) const
-  { return internal::pmul(a,a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_square_op<Scalar> >
-{ enum { Cost = NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasMul }; };
-
-/** \internal
-  * \brief Template functor to compute the cube of a scalar
-  * \sa class CwiseUnaryOp, Cwise::cube()
-  */
-template<typename Scalar>
-struct scalar_cube_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_cube_op)
-  inline Scalar operator() (const Scalar& a) const { return a*a*a; }
-  template<typename Packet>
-  inline const Packet packetOp(const Packet& a) const
-  { return internal::pmul(a,pmul(a,a)); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_cube_op<Scalar> >
-{ enum { Cost = 2*NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasMul }; };
-
-// default functor traits for STL functors:
-
-template<typename T>
-struct functor_traits<std::multiplies<T> >
-{ enum { Cost = NumTraits<T>::MulCost, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::divides<T> >
-{ enum { Cost = NumTraits<T>::MulCost, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::plus<T> >
-{ enum { Cost = NumTraits<T>::AddCost, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::minus<T> >
-{ enum { Cost = NumTraits<T>::AddCost, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::negate<T> >
-{ enum { Cost = NumTraits<T>::AddCost, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::logical_or<T> >
-{ enum { Cost = 1, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::logical_and<T> >
-{ enum { Cost = 1, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::logical_not<T> >
-{ enum { Cost = 1, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::greater<T> >
-{ enum { Cost = 1, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::less<T> >
-{ enum { Cost = 1, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::greater_equal<T> >
-{ enum { Cost = 1, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::less_equal<T> >
-{ enum { Cost = 1, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::equal_to<T> >
-{ enum { Cost = 1, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::not_equal_to<T> >
-{ enum { Cost = 1, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::binder2nd<T> >
-{ enum { Cost = functor_traits<T>::Cost, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::binder1st<T> >
-{ enum { Cost = functor_traits<T>::Cost, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::unary_negate<T> >
-{ enum { Cost = 1 + functor_traits<T>::Cost, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::binary_negate<T> >
-{ enum { Cost = 1 + functor_traits<T>::Cost, PacketAccess = false }; };
-
-#ifdef EIGEN_STDEXT_SUPPORT
-
-template<typename T0,typename T1>
-struct functor_traits<std::project1st<T0,T1> >
-{ enum { Cost = 0, PacketAccess = false }; };
-
-template<typename T0,typename T1>
-struct functor_traits<std::project2nd<T0,T1> >
-{ enum { Cost = 0, PacketAccess = false }; };
-
-template<typename T0,typename T1>
-struct functor_traits<std::select2nd<std::pair<T0,T1> > >
-{ enum { Cost = 0, PacketAccess = false }; };
-
-template<typename T0,typename T1>
-struct functor_traits<std::select1st<std::pair<T0,T1> > >
-{ enum { Cost = 0, PacketAccess = false }; };
-
-template<typename T0,typename T1>
-struct functor_traits<std::unary_compose<T0,T1> >
-{ enum { Cost = functor_traits<T0>::Cost + functor_traits<T1>::Cost, PacketAccess = false }; };
-
-template<typename T0,typename T1,typename T2>
-struct functor_traits<std::binary_compose<T0,T1,T2> >
-{ enum { Cost = functor_traits<T0>::Cost + functor_traits<T1>::Cost + functor_traits<T2>::Cost, PacketAccess = false }; };
-
-#endif // EIGEN_STDEXT_SUPPORT
-
-// allow to add new functors and specializations of functor_traits from outside Eigen.
-// this macro is really needed because functor_traits must be specialized after it is declared but before it is used...
-#ifdef EIGEN_FUNCTORS_PLUGIN
-#include EIGEN_FUNCTORS_PLUGIN
-#endif
-
-} // end namespace internal
-
-} // end namespace Eigen
-
-#endif // EIGEN_FUNCTORS_H
diff --git a/eigen/Eigen/src/Core/Fuzzy.h b/eigen/Eigen/src/Core/Fuzzy.h
index fe63bd2..43aa49b 100644
--- a/eigen/Eigen/src/Core/Fuzzy.h
+++ b/eigen/Eigen/src/Core/Fuzzy.h
@@ -19,18 +19,19 @@ namespace internal
 template<typename Derived, typename OtherDerived, bool is_integer = NumTraits<typename Derived::Scalar>::IsInteger>
 struct isApprox_selector
 {
+  EIGEN_DEVICE_FUNC
   static bool run(const Derived& x, const OtherDerived& y, const typename Derived::RealScalar& prec)
   {
-    using std::min;
-    typename internal::nested<Derived,2>::type nested(x);
-    typename internal::nested<OtherDerived,2>::type otherNested(y);
-    return (nested - otherNested).cwiseAbs2().sum() <= prec * prec * (min)(nested.cwiseAbs2().sum(), otherNested.cwiseAbs2().sum());
+    typename internal::nested_eval<Derived,2>::type nested(x);
+    typename internal::nested_eval<OtherDerived,2>::type otherNested(y);
+    return (nested - otherNested).cwiseAbs2().sum() <= prec * prec * numext::mini(nested.cwiseAbs2().sum(), otherNested.cwiseAbs2().sum());
   }
 };
 
 template<typename Derived, typename OtherDerived>
 struct isApprox_selector<Derived, OtherDerived, true>
 {
+  EIGEN_DEVICE_FUNC
   static bool run(const Derived& x, const OtherDerived& y, const typename Derived::RealScalar&)
   {
     return x.matrix() == y.matrix();
@@ -40,6 +41,7 @@ struct isApprox_selector<Derived, OtherDerived, true>
 template<typename Derived, typename OtherDerived, bool is_integer = NumTraits<typename Derived::Scalar>::IsInteger>
 struct isMuchSmallerThan_object_selector
 {
+  EIGEN_DEVICE_FUNC
   static bool run(const Derived& x, const OtherDerived& y, const typename Derived::RealScalar& prec)
   {
     return x.cwiseAbs2().sum() <= numext::abs2(prec) * y.cwiseAbs2().sum();
@@ -49,6 +51,7 @@ struct isMuchSmallerThan_object_selector
 template<typename Derived, typename OtherDerived>
 struct isMuchSmallerThan_object_selector<Derived, OtherDerived, true>
 {
+  EIGEN_DEVICE_FUNC
   static bool run(const Derived& x, const OtherDerived&, const typename Derived::RealScalar&)
   {
     return x.matrix() == Derived::Zero(x.rows(), x.cols()).matrix();
@@ -58,6 +61,7 @@ struct isMuchSmallerThan_object_selector<Derived, OtherDerived, true>
 template<typename Derived, bool is_integer = NumTraits<typename Derived::Scalar>::IsInteger>
 struct isMuchSmallerThan_scalar_selector
 {
+  EIGEN_DEVICE_FUNC
   static bool run(const Derived& x, const typename Derived::RealScalar& y, const typename Derived::RealScalar& prec)
   {
     return x.cwiseAbs2().sum() <= numext::abs2(prec * y);
@@ -67,6 +71,7 @@ struct isMuchSmallerThan_scalar_selector
 template<typename Derived>
 struct isMuchSmallerThan_scalar_selector<Derived, true>
 {
+  EIGEN_DEVICE_FUNC
   static bool run(const Derived& x, const typename Derived::RealScalar&, const typename Derived::RealScalar&)
   {
     return x.matrix() == Derived::Zero(x.rows(), x.cols()).matrix();
@@ -95,7 +100,7 @@ struct isMuchSmallerThan_scalar_selector<Derived, true>
   */
 template<typename Derived>
 template<typename OtherDerived>
-bool DenseBase<Derived>::isApprox(
+EIGEN_DEVICE_FUNC bool DenseBase<Derived>::isApprox(
   const DenseBase<OtherDerived>& other,
   const RealScalar& prec
 ) const
@@ -117,7 +122,7 @@ bool DenseBase<Derived>::isApprox(
   * \sa isApprox(), isMuchSmallerThan(const DenseBase<OtherDerived>&, RealScalar) const
   */
 template<typename Derived>
-bool DenseBase<Derived>::isMuchSmallerThan(
+EIGEN_DEVICE_FUNC bool DenseBase<Derived>::isMuchSmallerThan(
   const typename NumTraits<Scalar>::Real& other,
   const RealScalar& prec
 ) const
@@ -137,7 +142,7 @@ bool DenseBase<Derived>::isMuchSmallerThan(
   */
 template<typename Derived>
 template<typename OtherDerived>
-bool DenseBase<Derived>::isMuchSmallerThan(
+EIGEN_DEVICE_FUNC bool DenseBase<Derived>::isMuchSmallerThan(
   const DenseBase<OtherDerived>& other,
   const RealScalar& prec
 ) const
diff --git a/eigen/Eigen/src/Core/GeneralProduct.h b/eigen/Eigen/src/Core/GeneralProduct.h
index 5744eb7..b206b0a 100644
--- a/eigen/Eigen/src/Core/GeneralProduct.h
+++ b/eigen/Eigen/src/Core/GeneralProduct.h
@@ -11,29 +11,7 @@
 #ifndef EIGEN_GENERAL_PRODUCT_H
 #define EIGEN_GENERAL_PRODUCT_H
 
-namespace Eigen { 
-
-/** \class GeneralProduct
-  * \ingroup Core_Module
-  *
-  * \brief Expression of the product of two general matrices or vectors
-  *
-  * \param LhsNested the type used to store the left-hand side
-  * \param RhsNested the type used to store the right-hand side
-  * \param ProductMode the type of the product
-  *
-  * This class represents an expression of the product of two general matrices.
-  * We call a general matrix, a dense matrix with full storage. For instance,
-  * This excludes triangular, selfadjoint, and sparse matrices.
-  * It is the return type of the operator* between general matrices. Its template
-  * arguments are determined automatically by ProductReturnType. Therefore,
-  * GeneralProduct should never be used direclty. To determine the result type of a
-  * function which involves a matrix product, use ProductReturnType::Type.
-  *
-  * \sa ProductReturnType, MatrixBase::operator*(const MatrixBase<OtherDerived>&)
-  */
-template<typename Lhs, typename Rhs, int ProductType = internal::product_type<Lhs,Rhs>::value>
-class GeneralProduct;
+namespace Eigen {
 
 enum {
   Large = 2,
@@ -47,7 +25,8 @@ template<int Rows, int Cols, int Depth> struct product_type_selector;
 template<int Size, int MaxSize> struct product_size_category
 {
   enum { is_large = MaxSize == Dynamic ||
-                    Size >= EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD,
+                    Size >= EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD ||
+                    (Size==Dynamic && MaxSize>=EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD),
          value = is_large  ? Large
                : Size == 1 ? 1
                            : Small
@@ -59,15 +38,14 @@ template<typename Lhs, typename Rhs> struct product_type
   typedef typename remove_all<Lhs>::type _Lhs;
   typedef typename remove_all<Rhs>::type _Rhs;
   enum {
-    MaxRows  = _Lhs::MaxRowsAtCompileTime,
-    Rows  = _Lhs::RowsAtCompileTime,
-    MaxCols  = _Rhs::MaxColsAtCompileTime,
-    Cols  = _Rhs::ColsAtCompileTime,
-    MaxDepth = EIGEN_SIZE_MIN_PREFER_FIXED(_Lhs::MaxColsAtCompileTime,
-                                           _Rhs::MaxRowsAtCompileTime),
-    Depth = EIGEN_SIZE_MIN_PREFER_FIXED(_Lhs::ColsAtCompileTime,
-                                        _Rhs::RowsAtCompileTime),
-    LargeThreshold = EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
+    MaxRows = traits<_Lhs>::MaxRowsAtCompileTime,
+    Rows    = traits<_Lhs>::RowsAtCompileTime,
+    MaxCols = traits<_Rhs>::MaxColsAtCompileTime,
+    Cols    = traits<_Rhs>::ColsAtCompileTime,
+    MaxDepth = EIGEN_SIZE_MIN_PREFER_FIXED(traits<_Lhs>::MaxColsAtCompileTime,
+                                           traits<_Rhs>::MaxRowsAtCompileTime),
+    Depth = EIGEN_SIZE_MIN_PREFER_FIXED(traits<_Lhs>::ColsAtCompileTime,
+                                        traits<_Rhs>::RowsAtCompileTime)
   };
 
   // the splitting into different lines of code here, introducing the _select enums and the typedef below,
@@ -82,7 +60,8 @@ private:
 
 public:
   enum {
-    value = selector::ret
+    value = selector::ret,
+    ret = selector::ret
   };
 #ifdef EIGEN_DEBUG_PRODUCT
   static void debug()
@@ -98,12 +77,13 @@ public:
 #endif
 };
 
-
 /* The following allows to select the kind of product at compile time
  * based on the three dimensions of the product.
  * This is a compile time mapping from {1,Small,Large}^3 -> {product types} */
 // FIXME I'm not sure the current mapping is the ideal one.
 template<int M, int N>  struct product_type_selector<M,N,1>              { enum { ret = OuterProduct }; };
+template<int M>         struct product_type_selector<M, 1, 1>            { enum { ret = LazyCoeffBasedProductMode }; };
+template<int N>         struct product_type_selector<1, N, 1>            { enum { ret = LazyCoeffBasedProductMode }; };
 template<int Depth>     struct product_type_selector<1,    1,    Depth>  { enum { ret = InnerProduct }; };
 template<>              struct product_type_selector<1,    1,    1>      { enum { ret = InnerProduct }; };
 template<>              struct product_type_selector<Small,1,    Small>  { enum { ret = CoeffBasedProductMode }; };
@@ -122,60 +102,12 @@ template<>              struct product_type_selector<Small,Small,Large>  { enum
 template<>              struct product_type_selector<Large,Small,Large>  { enum { ret = GemmProduct }; };
 template<>              struct product_type_selector<Small,Large,Large>  { enum { ret = GemmProduct }; };
 template<>              struct product_type_selector<Large,Large,Large>  { enum { ret = GemmProduct }; };
-template<>              struct product_type_selector<Large,Small,Small>  { enum { ret = GemmProduct }; };
-template<>              struct product_type_selector<Small,Large,Small>  { enum { ret = GemmProduct }; };
+template<>              struct product_type_selector<Large,Small,Small>  { enum { ret = CoeffBasedProductMode }; };
+template<>              struct product_type_selector<Small,Large,Small>  { enum { ret = CoeffBasedProductMode }; };
 template<>              struct product_type_selector<Large,Large,Small>  { enum { ret = GemmProduct }; };
 
 } // end namespace internal
 
-/** \class ProductReturnType
-  * \ingroup Core_Module
-  *
-  * \brief Helper class to get the correct and optimized returned type of operator*
-  *
-  * \param Lhs the type of the left-hand side
-  * \param Rhs the type of the right-hand side
-  * \param ProductMode the type of the product (determined automatically by internal::product_mode)
-  *
-  * This class defines the typename Type representing the optimized product expression
-  * between two matrix expressions. In practice, using ProductReturnType<Lhs,Rhs>::Type
-  * is the recommended way to define the result type of a function returning an expression
-  * which involve a matrix product. The class Product should never be
-  * used directly.
-  *
-  * \sa class Product, MatrixBase::operator*(const MatrixBase<OtherDerived>&)
-  */
-template<typename Lhs, typename Rhs, int ProductType>
-struct ProductReturnType
-{
-  // TODO use the nested type to reduce instanciations ????
-//   typedef typename internal::nested<Lhs,Rhs::ColsAtCompileTime>::type LhsNested;
-//   typedef typename internal::nested<Rhs,Lhs::RowsAtCompileTime>::type RhsNested;
-
-  typedef GeneralProduct<Lhs/*Nested*/, Rhs/*Nested*/, ProductType> Type;
-};
-
-template<typename Lhs, typename Rhs>
-struct ProductReturnType<Lhs,Rhs,CoeffBasedProductMode>
-{
-  typedef typename internal::nested<Lhs, Rhs::ColsAtCompileTime, typename internal::plain_matrix_type<Lhs>::type >::type LhsNested;
-  typedef typename internal::nested<Rhs, Lhs::RowsAtCompileTime, typename internal::plain_matrix_type<Rhs>::type >::type RhsNested;
-  typedef CoeffBasedProduct<LhsNested, RhsNested, EvalBeforeAssigningBit | EvalBeforeNestingBit> Type;
-};
-
-template<typename Lhs, typename Rhs>
-struct ProductReturnType<Lhs,Rhs,LazyCoeffBasedProductMode>
-{
-  typedef typename internal::nested<Lhs, Rhs::ColsAtCompileTime, typename internal::plain_matrix_type<Lhs>::type >::type LhsNested;
-  typedef typename internal::nested<Rhs, Lhs::RowsAtCompileTime, typename internal::plain_matrix_type<Rhs>::type >::type RhsNested;
-  typedef CoeffBasedProduct<LhsNested, RhsNested, NestByRefBit> Type;
-};
-
-// this is a workaround for sun CC
-template<typename Lhs, typename Rhs>
-struct LazyProductReturnType : public ProductReturnType<Lhs,Rhs,LazyCoeffBasedProductMode>
-{};
-
 /***********************************************************************
 *  Implementation of Inner Vector Vector Product
 ***********************************************************************/
@@ -187,114 +119,10 @@ struct LazyProductReturnType : public ProductReturnType<Lhs,Rhs,LazyCoeffBasedPr
 // product ends up to a row-vector times col-vector product... To tackle this use
 // case, we could have a specialization for Block<MatrixType,1,1> with: operator=(Scalar x);
 
-namespace internal {
-
-template<typename Lhs, typename Rhs>
-struct traits<GeneralProduct<Lhs,Rhs,InnerProduct> >
- : traits<Matrix<typename scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType,1,1> >
-{};
-
-}
-
-template<typename Lhs, typename Rhs>
-class GeneralProduct<Lhs, Rhs, InnerProduct>
-  : internal::no_assignment_operator,
-    public Matrix<typename internal::scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType,1,1>
-{
-    typedef Matrix<typename internal::scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType,1,1> Base;
-  public:
-    GeneralProduct(const Lhs& lhs, const Rhs& rhs)
-    {
-      Base::coeffRef(0,0) = (lhs.transpose().cwiseProduct(rhs)).sum();
-    }
-
-    /** Convertion to scalar */
-    operator const typename Base::Scalar() const {
-      return Base::coeff(0,0);
-    }
-};
-
 /***********************************************************************
 *  Implementation of Outer Vector Vector Product
 ***********************************************************************/
 
-namespace internal {
-
-// Column major
-template<typename ProductType, typename Dest, typename Func>
-EIGEN_DONT_INLINE void outer_product_selector_run(const ProductType& prod, Dest& dest, const Func& func, const false_type&)
-{
-  typedef typename Dest::Index Index;
-  // FIXME make sure lhs is sequentially stored
-  // FIXME not very good if rhs is real and lhs complex while alpha is real too
-  const Index cols = dest.cols();
-  for (Index j=0; j<cols; ++j)
-    func(dest.col(j), prod.rhs().coeff(0,j) * prod.lhs());
-}
-
-// Row major
-template<typename ProductType, typename Dest, typename Func>
-EIGEN_DONT_INLINE void outer_product_selector_run(const ProductType& prod, Dest& dest, const Func& func, const true_type&) {
-  typedef typename Dest::Index Index;
-  // FIXME make sure rhs is sequentially stored
-  // FIXME not very good if lhs is real and rhs complex while alpha is real too
-  const Index rows = dest.rows();
-  for (Index i=0; i<rows; ++i)
-    func(dest.row(i), prod.lhs().coeff(i,0) * prod.rhs());
-}
-
-template<typename Lhs, typename Rhs>
-struct traits<GeneralProduct<Lhs,Rhs,OuterProduct> >
- : traits<ProductBase<GeneralProduct<Lhs,Rhs,OuterProduct>, Lhs, Rhs> >
-{};
-
-}
-
-template<typename Lhs, typename Rhs>
-class GeneralProduct<Lhs, Rhs, OuterProduct>
-  : public ProductBase<GeneralProduct<Lhs,Rhs,OuterProduct>, Lhs, Rhs>
-{
-    template<typename T> struct is_row_major : internal::conditional<(int(T::Flags)&RowMajorBit), internal::true_type, internal::false_type>::type {};
-    
-  public:
-    EIGEN_PRODUCT_PUBLIC_INTERFACE(GeneralProduct)
-
-    GeneralProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs)
-    {
-    }
-    
-    struct set  { template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived()  = src; } };
-    struct add  { template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived() += src; } };
-    struct sub  { template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived() -= src; } };
-    struct adds {
-      Scalar m_scale;
-      adds(const Scalar& s) : m_scale(s) {}
-      template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const {
-        dst.const_cast_derived() += m_scale * src;
-      }
-    };
-    
-    template<typename Dest>
-    inline void evalTo(Dest& dest) const {
-      internal::outer_product_selector_run(*this, dest, set(), is_row_major<Dest>());
-    }
-    
-    template<typename Dest>
-    inline void addTo(Dest& dest) const {
-      internal::outer_product_selector_run(*this, dest, add(), is_row_major<Dest>());
-    }
-
-    template<typename Dest>
-    inline void subTo(Dest& dest) const {
-      internal::outer_product_selector_run(*this, dest, sub(), is_row_major<Dest>());
-    }
-
-    template<typename Dest> void scaleAndAddTo(Dest& dest, const Scalar& alpha) const
-    {
-      internal::outer_product_selector_run(*this, dest, adds(alpha), is_row_major<Dest>());
-    }
-};
-
 /***********************************************************************
 *  Implementation of General Matrix Vector Product
 ***********************************************************************/
@@ -308,60 +136,13 @@ class GeneralProduct<Lhs, Rhs, OuterProduct>
  */
 namespace internal {
 
-template<typename Lhs, typename Rhs>
-struct traits<GeneralProduct<Lhs,Rhs,GemvProduct> >
- : traits<ProductBase<GeneralProduct<Lhs,Rhs,GemvProduct>, Lhs, Rhs> >
-{};
-
 template<int Side, int StorageOrder, bool BlasCompatible>
-struct gemv_selector;
+struct gemv_dense_selector;
 
 } // end namespace internal
 
-template<typename Lhs, typename Rhs>
-class GeneralProduct<Lhs, Rhs, GemvProduct>
-  : public ProductBase<GeneralProduct<Lhs,Rhs,GemvProduct>, Lhs, Rhs>
-{
-  public:
-    EIGEN_PRODUCT_PUBLIC_INTERFACE(GeneralProduct)
-
-    typedef typename Lhs::Scalar LhsScalar;
-    typedef typename Rhs::Scalar RhsScalar;
-
-    GeneralProduct(const Lhs& a_lhs, const Rhs& a_rhs) : Base(a_lhs,a_rhs)
-    {
-//       EIGEN_STATIC_ASSERT((internal::is_same<typename Lhs::Scalar, typename Rhs::Scalar>::value),
-//         YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
-    }
-
-    enum { Side = Lhs::IsVectorAtCompileTime ? OnTheLeft : OnTheRight };
-    typedef typename internal::conditional<int(Side)==OnTheRight,_LhsNested,_RhsNested>::type MatrixType;
-
-    template<typename Dest> void scaleAndAddTo(Dest& dst, const Scalar& alpha) const
-    {
-      eigen_assert(m_lhs.rows() == dst.rows() && m_rhs.cols() == dst.cols());
-      internal::gemv_selector<Side,(int(MatrixType::Flags)&RowMajorBit) ? RowMajor : ColMajor,
-                       bool(internal::blas_traits<MatrixType>::HasUsableDirectAccess)>::run(*this, dst, alpha);
-    }
-};
-
 namespace internal {
 
-// The vector is on the left => transposition
-template<int StorageOrder, bool BlasCompatible>
-struct gemv_selector<OnTheLeft,StorageOrder,BlasCompatible>
-{
-  template<typename ProductType, typename Dest>
-  static void run(const ProductType& prod, Dest& dest, const typename ProductType::Scalar& alpha)
-  {
-    Transpose<Dest> destT(dest);
-    enum { OtherStorageOrder = StorageOrder == RowMajor ? ColMajor : RowMajor };
-    gemv_selector<OnTheRight,OtherStorageOrder,BlasCompatible>
-      ::run(GeneralProduct<Transpose<const typename ProductType::_RhsNested>,Transpose<const typename ProductType::_LhsNested>, GemvProduct>
-        (prod.rhs().transpose(), prod.lhs().transpose()), destT, alpha);
-  }
-};
-
 template<typename Scalar,int Size,int MaxSize,bool Cond> struct gemv_static_vector_if;
 
 template<typename Scalar,int Size,int MaxSize>
@@ -379,46 +160,61 @@ struct gemv_static_vector_if<Scalar,Size,Dynamic,true>
 template<typename Scalar,int Size,int MaxSize>
 struct gemv_static_vector_if<Scalar,Size,MaxSize,true>
 {
-  #if EIGEN_ALIGN_STATICALLY
-  internal::plain_array<Scalar,EIGEN_SIZE_MIN_PREFER_FIXED(Size,MaxSize),0> m_data;
-  EIGEN_STRONG_INLINE Scalar* data() { return m_data.array; }
-  #else
-  // Some architectures cannot align on the stack,
-  // => let's manually enforce alignment by allocating more data and return the address of the first aligned element.
   enum {
     ForceAlignment  = internal::packet_traits<Scalar>::Vectorizable,
     PacketSize      = internal::packet_traits<Scalar>::size
   };
-  internal::plain_array<Scalar,EIGEN_SIZE_MIN_PREFER_FIXED(Size,MaxSize)+(ForceAlignment?PacketSize:0),0> m_data;
+  #if EIGEN_MAX_STATIC_ALIGN_BYTES!=0
+  internal::plain_array<Scalar,EIGEN_SIZE_MIN_PREFER_FIXED(Size,MaxSize),0,EIGEN_PLAIN_ENUM_MIN(AlignedMax,PacketSize)> m_data;
+  EIGEN_STRONG_INLINE Scalar* data() { return m_data.array; }
+  #else
+  // Some architectures cannot align on the stack,
+  // => let's manually enforce alignment by allocating more data and return the address of the first aligned element.
+  internal::plain_array<Scalar,EIGEN_SIZE_MIN_PREFER_FIXED(Size,MaxSize)+(ForceAlignment?EIGEN_MAX_ALIGN_BYTES:0),0> m_data;
   EIGEN_STRONG_INLINE Scalar* data() {
     return ForceAlignment
-            ? reinterpret_cast<Scalar*>((reinterpret_cast<size_t>(m_data.array) & ~(size_t(15))) + 16)
+            ? reinterpret_cast<Scalar*>((internal::UIntPtr(m_data.array) & ~(std::size_t(EIGEN_MAX_ALIGN_BYTES-1))) + EIGEN_MAX_ALIGN_BYTES)
             : m_data.array;
   }
   #endif
 };
 
-template<> struct gemv_selector<OnTheRight,ColMajor,true>
+// The vector is on the left => transposition
+template<int StorageOrder, bool BlasCompatible>
+struct gemv_dense_selector<OnTheLeft,StorageOrder,BlasCompatible>
+{
+  template<typename Lhs, typename Rhs, typename Dest>
+  static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
+  {
+    Transpose<Dest> destT(dest);
+    enum { OtherStorageOrder = StorageOrder == RowMajor ? ColMajor : RowMajor };
+    gemv_dense_selector<OnTheRight,OtherStorageOrder,BlasCompatible>
+      ::run(rhs.transpose(), lhs.transpose(), destT, alpha);
+  }
+};
+
+template<> struct gemv_dense_selector<OnTheRight,ColMajor,true>
 {
-  template<typename ProductType, typename Dest>
-  static inline void run(const ProductType& prod, Dest& dest, const typename ProductType::Scalar& alpha)
+  template<typename Lhs, typename Rhs, typename Dest>
+  static inline void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
   {
-    typedef typename ProductType::Index Index;
-    typedef typename ProductType::LhsScalar   LhsScalar;
-    typedef typename ProductType::RhsScalar   RhsScalar;
-    typedef typename ProductType::Scalar      ResScalar;
-    typedef typename ProductType::RealScalar  RealScalar;
-    typedef typename ProductType::ActualLhsType ActualLhsType;
-    typedef typename ProductType::ActualRhsType ActualRhsType;
-    typedef typename ProductType::LhsBlasTraits LhsBlasTraits;
-    typedef typename ProductType::RhsBlasTraits RhsBlasTraits;
-    typedef Map<Matrix<ResScalar,Dynamic,1>, Aligned> MappedDest;
-
-    ActualLhsType actualLhs = LhsBlasTraits::extract(prod.lhs());
-    ActualRhsType actualRhs = RhsBlasTraits::extract(prod.rhs());
-
-    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs())
-                                  * RhsBlasTraits::extractScalarFactor(prod.rhs());
+    typedef typename Lhs::Scalar   LhsScalar;
+    typedef typename Rhs::Scalar   RhsScalar;
+    typedef typename Dest::Scalar  ResScalar;
+    typedef typename Dest::RealScalar  RealScalar;
+    
+    typedef internal::blas_traits<Lhs> LhsBlasTraits;
+    typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
+    typedef internal::blas_traits<Rhs> RhsBlasTraits;
+    typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
+  
+    typedef Map<Matrix<ResScalar,Dynamic,1>, EIGEN_PLAIN_ENUM_MIN(AlignedMax,internal::packet_traits<ResScalar>::size)> MappedDest;
+
+    ActualLhsType actualLhs = LhsBlasTraits::extract(lhs);
+    ActualRhsType actualRhs = RhsBlasTraits::extract(rhs);
+
+    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(lhs)
+                                  * RhsBlasTraits::extractScalarFactor(rhs);
 
     // make sure Dest is a compile-time vector type (bug 1166)
     typedef typename conditional<Dest::IsVectorAtCompileTime, Dest, typename Dest::ColXpr>::type ActualDest;
@@ -428,80 +224,97 @@ template<> struct gemv_selector<OnTheRight,ColMajor,true>
       // on, the other hand it is good for the cache to pack the vector anyways...
       EvalToDestAtCompileTime = (ActualDest::InnerStrideAtCompileTime==1),
       ComplexByReal = (NumTraits<LhsScalar>::IsComplex) && (!NumTraits<RhsScalar>::IsComplex),
-      MightCannotUseDest = (ActualDest::InnerStrideAtCompileTime!=1) || ComplexByReal
+      MightCannotUseDest = (!EvalToDestAtCompileTime) || ComplexByReal
     };
 
-    gemv_static_vector_if<ResScalar,ActualDest::SizeAtCompileTime,ActualDest::MaxSizeAtCompileTime,MightCannotUseDest> static_dest;
-
-    bool alphaIsCompatible = (!ComplexByReal) || (numext::imag(actualAlpha)==RealScalar(0));
-    bool evalToDest = EvalToDestAtCompileTime && alphaIsCompatible;
-    
+    typedef const_blas_data_mapper<LhsScalar,Index,ColMajor> LhsMapper;
+    typedef const_blas_data_mapper<RhsScalar,Index,RowMajor> RhsMapper;
     RhsScalar compatibleAlpha = get_factor<ResScalar,RhsScalar>::run(actualAlpha);
 
-    ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(),
-                                                  evalToDest ? dest.data() : static_dest.data());
-    
-    if(!evalToDest)
+    if(!MightCannotUseDest)
     {
-      #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
-      int size = dest.size();
-      EIGEN_DENSE_STORAGE_CTOR_PLUGIN
-      #endif
-      if(!alphaIsCompatible)
+      // shortcut if we are sure to be able to use dest directly,
+      // this ease the compiler to generate cleaner and more optimzized code for most common cases
+      general_matrix_vector_product
+          <Index,LhsScalar,LhsMapper,ColMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsMapper,RhsBlasTraits::NeedToConjugate>::run(
+          actualLhs.rows(), actualLhs.cols(),
+          LhsMapper(actualLhs.data(), actualLhs.outerStride()),
+          RhsMapper(actualRhs.data(), actualRhs.innerStride()),
+          dest.data(), 1,
+          compatibleAlpha);
+    }
+    else
+    {
+      gemv_static_vector_if<ResScalar,ActualDest::SizeAtCompileTime,ActualDest::MaxSizeAtCompileTime,MightCannotUseDest> static_dest;
+
+      const bool alphaIsCompatible = (!ComplexByReal) || (numext::imag(actualAlpha)==RealScalar(0));
+      const bool evalToDest = EvalToDestAtCompileTime && alphaIsCompatible;
+
+      ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(),
+                                                    evalToDest ? dest.data() : static_dest.data());
+
+      if(!evalToDest)
       {
-        MappedDest(actualDestPtr, dest.size()).setZero();
-        compatibleAlpha = RhsScalar(1);
+        #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
+        Index size = dest.size();
+        EIGEN_DENSE_STORAGE_CTOR_PLUGIN
+        #endif
+        if(!alphaIsCompatible)
+        {
+          MappedDest(actualDestPtr, dest.size()).setZero();
+          compatibleAlpha = RhsScalar(1);
+        }
+        else
+          MappedDest(actualDestPtr, dest.size()) = dest;
       }
-      else
-        MappedDest(actualDestPtr, dest.size()) = dest;
-    }
 
-    general_matrix_vector_product
-      <Index,LhsScalar,ColMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsBlasTraits::NeedToConjugate>::run(
-        actualLhs.rows(), actualLhs.cols(),
-        actualLhs.data(), actualLhs.outerStride(),
-        actualRhs.data(), actualRhs.innerStride(),
-        actualDestPtr, 1,
-        compatibleAlpha);
+      general_matrix_vector_product
+          <Index,LhsScalar,LhsMapper,ColMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsMapper,RhsBlasTraits::NeedToConjugate>::run(
+          actualLhs.rows(), actualLhs.cols(),
+          LhsMapper(actualLhs.data(), actualLhs.outerStride()),
+          RhsMapper(actualRhs.data(), actualRhs.innerStride()),
+          actualDestPtr, 1,
+          compatibleAlpha);
 
-    if (!evalToDest)
-    {
-      if(!alphaIsCompatible)
-        dest += actualAlpha * MappedDest(actualDestPtr, dest.size());
-      else
-        dest = MappedDest(actualDestPtr, dest.size());
+      if (!evalToDest)
+      {
+        if(!alphaIsCompatible)
+          dest.matrix() += actualAlpha * MappedDest(actualDestPtr, dest.size());
+        else
+          dest = MappedDest(actualDestPtr, dest.size());
+      }
     }
   }
 };
 
-template<> struct gemv_selector<OnTheRight,RowMajor,true>
+template<> struct gemv_dense_selector<OnTheRight,RowMajor,true>
 {
-  template<typename ProductType, typename Dest>
-  static void run(const ProductType& prod, Dest& dest, const typename ProductType::Scalar& alpha)
+  template<typename Lhs, typename Rhs, typename Dest>
+  static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
   {
-    typedef typename ProductType::LhsScalar LhsScalar;
-    typedef typename ProductType::RhsScalar RhsScalar;
-    typedef typename ProductType::Scalar    ResScalar;
-    typedef typename ProductType::Index Index;
-    typedef typename ProductType::ActualLhsType ActualLhsType;
-    typedef typename ProductType::ActualRhsType ActualRhsType;
-    typedef typename ProductType::_ActualRhsType _ActualRhsType;
-    typedef typename ProductType::LhsBlasTraits LhsBlasTraits;
-    typedef typename ProductType::RhsBlasTraits RhsBlasTraits;
-
-    typename add_const<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(prod.lhs());
-    typename add_const<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(prod.rhs());
-
-    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs())
-                                  * RhsBlasTraits::extractScalarFactor(prod.rhs());
+    typedef typename Lhs::Scalar   LhsScalar;
+    typedef typename Rhs::Scalar   RhsScalar;
+    typedef typename Dest::Scalar  ResScalar;
+    
+    typedef internal::blas_traits<Lhs> LhsBlasTraits;
+    typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
+    typedef internal::blas_traits<Rhs> RhsBlasTraits;
+    typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
+    typedef typename internal::remove_all<ActualRhsType>::type ActualRhsTypeCleaned;
+
+    typename add_const<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(lhs);
+    typename add_const<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(rhs);
+
+    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(lhs)
+                                  * RhsBlasTraits::extractScalarFactor(rhs);
 
     enum {
       // FIXME find a way to allow an inner stride on the result if packet_traits<Scalar>::size==1
       // on, the other hand it is good for the cache to pack the vector anyways...
-      DirectlyUseRhs = _ActualRhsType::InnerStrideAtCompileTime==1
+      DirectlyUseRhs = ActualRhsTypeCleaned::InnerStrideAtCompileTime==1
     };
 
-    gemv_static_vector_if<RhsScalar,_ActualRhsType::SizeAtCompileTime,_ActualRhsType::MaxSizeAtCompileTime,!DirectlyUseRhs> static_rhs;
+    gemv_static_vector_if<RhsScalar,ActualRhsTypeCleaned::SizeAtCompileTime,ActualRhsTypeCleaned::MaxSizeAtCompileTime,!DirectlyUseRhs> static_rhs;
 
     ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhsPtr,actualRhs.size(),
         DirectlyUseRhs ? const_cast<RhsScalar*>(actualRhs.data()) : static_rhs.data());
@@ -509,45 +322,48 @@ template<> struct gemv_selector<OnTheRight,RowMajor,true>
     if(!DirectlyUseRhs)
     {
       #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
-      int size = actualRhs.size();
+      Index size = actualRhs.size();
       EIGEN_DENSE_STORAGE_CTOR_PLUGIN
       #endif
-      Map<typename _ActualRhsType::PlainObject>(actualRhsPtr, actualRhs.size()) = actualRhs;
+      Map<typename ActualRhsTypeCleaned::PlainObject>(actualRhsPtr, actualRhs.size()) = actualRhs;
     }
 
+    typedef const_blas_data_mapper<LhsScalar,Index,RowMajor> LhsMapper;
+    typedef const_blas_data_mapper<RhsScalar,Index,ColMajor> RhsMapper;
     general_matrix_vector_product
-      <Index,LhsScalar,RowMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsBlasTraits::NeedToConjugate>::run(
+        <Index,LhsScalar,LhsMapper,RowMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsMapper,RhsBlasTraits::NeedToConjugate>::run(
         actualLhs.rows(), actualLhs.cols(),
-        actualLhs.data(), actualLhs.outerStride(),
-        actualRhsPtr, 1,
+        LhsMapper(actualLhs.data(), actualLhs.outerStride()),
+        RhsMapper(actualRhsPtr, 1),
         dest.data(), dest.col(0).innerStride(), //NOTE  if dest is not a vector at compile-time, then dest.innerStride() might be wrong. (bug 1166)
         actualAlpha);
   }
 };
 
-template<> struct gemv_selector<OnTheRight,ColMajor,false>
+template<> struct gemv_dense_selector<OnTheRight,ColMajor,false>
 {
-  template<typename ProductType, typename Dest>
-  static void run(const ProductType& prod, Dest& dest, const typename ProductType::Scalar& alpha)
+  template<typename Lhs, typename Rhs, typename Dest>
+  static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
   {
-    typedef typename Dest::Index Index;
-    // TODO makes sure dest is sequentially stored in memory, otherwise use a temp
-    const Index size = prod.rhs().rows();
+    EIGEN_STATIC_ASSERT((!nested_eval<Lhs,1>::Evaluate),EIGEN_INTERNAL_COMPILATION_ERROR_OR_YOU_MADE_A_PROGRAMMING_MISTAKE);
+    // TODO if rhs is large enough it might be beneficial to make sure that dest is sequentially stored in memory, otherwise use a temp
+    typename nested_eval<Rhs,1>::type actual_rhs(rhs);
+    const Index size = rhs.rows();
     for(Index k=0; k<size; ++k)
-      dest += (alpha*prod.rhs().coeff(k)) * prod.lhs().col(k);
+      dest += (alpha*actual_rhs.coeff(k)) * lhs.col(k);
   }
 };
 
-template<> struct gemv_selector<OnTheRight,RowMajor,false>
+template<> struct gemv_dense_selector<OnTheRight,RowMajor,false>
 {
-  template<typename ProductType, typename Dest>
-  static void run(const ProductType& prod, Dest& dest, const typename ProductType::Scalar& alpha)
+  template<typename Lhs, typename Rhs, typename Dest>
+  static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
   {
-    typedef typename Dest::Index Index;
-    // TODO makes sure rhs is sequentially stored in memory, otherwise use a temp
-    const Index rows = prod.rows();
+    EIGEN_STATIC_ASSERT((!nested_eval<Lhs,1>::Evaluate),EIGEN_INTERNAL_COMPILATION_ERROR_OR_YOU_MADE_A_PROGRAMMING_MISTAKE);
+    typename nested_eval<Rhs,Lhs::RowsAtCompileTime>::type actual_rhs(rhs);
+    const Index rows = dest.rows();
     for(Index i=0; i<rows; ++i)
-      dest.coeffRef(i) += alpha * (prod.lhs().row(i).cwiseProduct(prod.rhs().transpose())).sum();
+      dest.coeffRef(i) += alpha * (lhs.row(i).cwiseProduct(actual_rhs.transpose())).sum();
   }
 };
 
@@ -563,9 +379,11 @@ template<> struct gemv_selector<OnTheRight,RowMajor,false>
   *
   * \sa lazyProduct(), operator*=(const MatrixBase&), Cwise::operator*()
   */
+#ifndef __CUDACC__
+
 template<typename Derived>
 template<typename OtherDerived>
-inline const typename ProductReturnType<Derived, OtherDerived>::Type
+inline const Product<Derived, OtherDerived>
 MatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const
 {
   // A note regarding the function declaration: In MSVC, this function will sometimes
@@ -590,9 +408,12 @@ MatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const
 #ifdef EIGEN_DEBUG_PRODUCT
   internal::product_type<Derived,OtherDerived>::debug();
 #endif
-  return typename ProductReturnType<Derived,OtherDerived>::Type(derived(), other.derived());
+
+  return Product<Derived, OtherDerived>(derived(), other.derived());
 }
 
+#endif // __CUDACC__
+
 /** \returns an expression of the matrix product of \c *this and \a other without implicit evaluation.
   *
   * The returned product will behave like any other expressions: the coefficients of the product will be
@@ -606,8 +427,8 @@ MatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const
   */
 template<typename Derived>
 template<typename OtherDerived>
-const typename LazyProductReturnType<Derived,OtherDerived>::Type
-MatrixBase<Derived>::lazyProduct(const MatrixBase<OtherDerived> &other) const
+const Product<Derived,OtherDerived,LazyProduct>
+EIGEN_DEVICE_FUNC MatrixBase<Derived>::lazyProduct(const MatrixBase<OtherDerived> &other) const
 {
   enum {
     ProductIsValid =  Derived::ColsAtCompileTime==Dynamic
@@ -625,7 +446,7 @@ MatrixBase<Derived>::lazyProduct(const MatrixBase<OtherDerived> &other) const
     INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION)
   EIGEN_STATIC_ASSERT(ProductIsValid || SameSizes, INVALID_MATRIX_PRODUCT)
 
-  return typename LazyProductReturnType<Derived,OtherDerived>::Type(derived(), other.derived());
+  return Product<Derived,OtherDerived,LazyProduct>(derived(), other.derived());
 }
 
 } // end namespace Eigen
diff --git a/eigen/Eigen/src/Core/GenericPacketMath.h b/eigen/Eigen/src/Core/GenericPacketMath.h
index 5f783eb..d19d5bb 100644
--- a/eigen/Eigen/src/Core/GenericPacketMath.h
+++ b/eigen/Eigen/src/Core/GenericPacketMath.h
@@ -42,21 +42,29 @@ namespace internal {
 struct default_packet_traits
 {
   enum {
+    HasHalfPacket = 0,
+
     HasAdd    = 1,
     HasSub    = 1,
     HasMul    = 1,
     HasNegate = 1,
     HasAbs    = 1,
+    HasArg    = 0,
     HasAbs2   = 1,
     HasMin    = 1,
     HasMax    = 1,
     HasConj   = 1,
     HasSetLinear = 1,
+    HasBlend  = 0,
 
     HasDiv    = 0,
     HasSqrt   = 0,
+    HasRsqrt  = 0,
     HasExp    = 0,
+    HasExpm1  = 0,
     HasLog    = 0,
+    HasLog1p  = 0,
+    HasLog10  = 0,
     HasPow    = 0,
 
     HasSin    = 0,
@@ -64,17 +72,37 @@ struct default_packet_traits
     HasTan    = 0,
     HasASin   = 0,
     HasACos   = 0,
-    HasATan   = 0
+    HasATan   = 0,
+    HasSinh   = 0,
+    HasCosh   = 0,
+    HasTanh   = 0,
+    HasLGamma = 0,
+    HasDiGamma = 0,
+    HasZeta = 0,
+    HasPolygamma = 0,
+    HasErf = 0,
+    HasErfc = 0,
+    HasIGamma = 0,
+    HasIGammac = 0,
+    HasBetaInc = 0,
+
+    HasRound  = 0,
+    HasFloor  = 0,
+    HasCeil   = 0,
+
+    HasSign   = 0
   };
 };
 
 template<typename T> struct packet_traits : default_packet_traits
 {
   typedef T type;
+  typedef T half;
   enum {
     Vectorizable = 0,
     size = 1,
-    AlignedOnScalar = 0
+    AlignedOnScalar = 0,
+    HasHalfPacket = 0
   };
   enum {
     HasAdd    = 0,
@@ -90,135 +118,239 @@ template<typename T> struct packet_traits : default_packet_traits
   };
 };
 
+template<typename T> struct packet_traits<const T> : packet_traits<T> { };
+
+template <typename Src, typename Tgt> struct type_casting_traits {
+  enum {
+    VectorizedCast = 0,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+
+/** \internal \returns static_cast<TgtType>(a) (coeff-wise) */
+template <typename SrcPacket, typename TgtPacket>
+EIGEN_DEVICE_FUNC inline TgtPacket
+pcast(const SrcPacket& a) {
+  return static_cast<TgtPacket>(a);
+}
+template <typename SrcPacket, typename TgtPacket>
+EIGEN_DEVICE_FUNC inline TgtPacket
+pcast(const SrcPacket& a, const SrcPacket& /*b*/) {
+  return static_cast<TgtPacket>(a);
+}
+
+template <typename SrcPacket, typename TgtPacket>
+EIGEN_DEVICE_FUNC inline TgtPacket
+pcast(const SrcPacket& a, const SrcPacket& /*b*/, const SrcPacket& /*c*/, const SrcPacket& /*d*/) {
+  return static_cast<TgtPacket>(a);
+}
+
 /** \internal \returns a + b (coeff-wise) */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 padd(const Packet& a,
         const Packet& b) { return a+b; }
 
 /** \internal \returns a - b (coeff-wise) */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 psub(const Packet& a,
         const Packet& b) { return a-b; }
 
 /** \internal \returns -a (coeff-wise) */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pnegate(const Packet& a) { return -a; }
 
 /** \internal \returns conj(a) (coeff-wise) */
-template<typename Packet> inline Packet
+
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pconj(const Packet& a) { return numext::conj(a); }
 
 /** \internal \returns a * b (coeff-wise) */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pmul(const Packet& a,
         const Packet& b) { return a*b; }
 
 /** \internal \returns a / b (coeff-wise) */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pdiv(const Packet& a,
         const Packet& b) { return a/b; }
 
 /** \internal \returns the min of \a a and \a b  (coeff-wise) */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pmin(const Packet& a,
-        const Packet& b) { using std::min; return (min)(a, b); }
+        const Packet& b) { return numext::mini(a, b); }
 
 /** \internal \returns the max of \a a and \a b  (coeff-wise) */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pmax(const Packet& a,
-        const Packet& b) { using std::max; return (max)(a, b); }
+        const Packet& b) { return numext::maxi(a, b); }
 
 /** \internal \returns the absolute value of \a a */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pabs(const Packet& a) { using std::abs; return abs(a); }
 
+/** \internal \returns the phase angle of \a a */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+parg(const Packet& a) { using numext::arg; return arg(a); }
+
 /** \internal \returns the bitwise and of \a a and \a b */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pand(const Packet& a, const Packet& b) { return a & b; }
 
 /** \internal \returns the bitwise or of \a a and \a b */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 por(const Packet& a, const Packet& b) { return a | b; }
 
 /** \internal \returns the bitwise xor of \a a and \a b */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pxor(const Packet& a, const Packet& b) { return a ^ b; }
 
 /** \internal \returns the bitwise andnot of \a a and \a b */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pandnot(const Packet& a, const Packet& b) { return a & (!b); }
 
 /** \internal \returns a packet version of \a *from, from must be 16 bytes aligned */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pload(const typename unpacket_traits<Packet>::type* from) { return *from; }
 
 /** \internal \returns a packet version of \a *from, (un-aligned load) */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 ploadu(const typename unpacket_traits<Packet>::type* from) { return *from; }
 
+/** \internal \returns a packet with constant coefficients \a a, e.g.: (a,a,a,a) */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pset1(const typename unpacket_traits<Packet>::type& a) { return a; }
+
+/** \internal \returns a packet with constant coefficients \a a[0], e.g.: (a[0],a[0],a[0],a[0]) */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pload1(const typename unpacket_traits<Packet>::type  *a) { return pset1<Packet>(*a); }
+
 /** \internal \returns a packet with elements of \a *from duplicated.
-  * For instance, for a packet of 8 elements, 4 scalar will be read from \a *from and
-  * duplicated to form: {from[0],from[0],from[1],from[1],,from[2],from[2],,from[3],from[3]}
+  * For instance, for a packet of 8 elements, 4 scalars will be read from \a *from and
+  * duplicated to form: {from[0],from[0],from[1],from[1],from[2],from[2],from[3],from[3]}
   * Currently, this function is only used for scalar * complex products.
- */
-template<typename Packet> inline Packet
+  */
+template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet
 ploaddup(const typename unpacket_traits<Packet>::type* from) { return *from; }
 
-/** \internal \returns a packet with constant coefficients \a a, e.g.: (a,a,a,a) */
-template<typename Packet> inline Packet
-pset1(const typename unpacket_traits<Packet>::type& a) { return a; }
+/** \internal \returns a packet with elements of \a *from quadrupled.
+  * For instance, for a packet of 8 elements, 2 scalars will be read from \a *from and
+  * replicated to form: {from[0],from[0],from[0],from[0],from[1],from[1],from[1],from[1]}
+  * Currently, this function is only used in matrix products.
+  * For packet-size smaller or equal to 4, this function is equivalent to pload1 
+  */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+ploadquad(const typename unpacket_traits<Packet>::type* from)
+{ return pload1<Packet>(from); }
+
+/** \internal equivalent to
+  * \code
+  * a0 = pload1(a+0);
+  * a1 = pload1(a+1);
+  * a2 = pload1(a+2);
+  * a3 = pload1(a+3);
+  * \endcode
+  * \sa pset1, pload1, ploaddup, pbroadcast2
+  */
+template<typename Packet> EIGEN_DEVICE_FUNC
+inline void pbroadcast4(const typename unpacket_traits<Packet>::type *a,
+                        Packet& a0, Packet& a1, Packet& a2, Packet& a3)
+{
+  a0 = pload1<Packet>(a+0);
+  a1 = pload1<Packet>(a+1);
+  a2 = pload1<Packet>(a+2);
+  a3 = pload1<Packet>(a+3);
+}
+
+/** \internal equivalent to
+  * \code
+  * a0 = pload1(a+0);
+  * a1 = pload1(a+1);
+  * \endcode
+  * \sa pset1, pload1, ploaddup, pbroadcast4
+  */
+template<typename Packet> EIGEN_DEVICE_FUNC
+inline void pbroadcast2(const typename unpacket_traits<Packet>::type *a,
+                        Packet& a0, Packet& a1)
+{
+  a0 = pload1<Packet>(a+0);
+  a1 = pload1<Packet>(a+1);
+}
 
 /** \internal \brief Returns a packet with coefficients (a,a+1,...,a+packet_size-1). */
-template<typename Scalar> inline typename packet_traits<Scalar>::type
-plset(const Scalar& a) { return a; }
+template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet
+plset(const typename unpacket_traits<Packet>::type& a) { return a; }
 
 /** \internal copy the packet \a from to \a *to, \a to must be 16 bytes aligned */
-template<typename Scalar, typename Packet> inline void pstore(Scalar* to, const Packet& from)
+template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline void pstore(Scalar* to, const Packet& from)
 { (*to) = from; }
 
 /** \internal copy the packet \a from to \a *to, (un-aligned store) */
-template<typename Scalar, typename Packet> inline void pstoreu(Scalar* to, const Packet& from)
-{ (*to) = from; }
+template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline void pstoreu(Scalar* to, const Packet& from)
+{  (*to) = from; }
+
+ template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline Packet pgather(const Scalar* from, Index /*stride*/)
+ { return ploadu<Packet>(from); }
+
+ template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline void pscatter(Scalar* to, const Packet& from, Index /*stride*/)
+ { pstore(to, from); }
 
 /** \internal tries to do cache prefetching of \a addr */
-template<typename Scalar> inline void prefetch(const Scalar* addr)
+template<typename Scalar> EIGEN_DEVICE_FUNC inline void prefetch(const Scalar* addr)
 {
-#if !defined(_MSC_VER)
-__builtin_prefetch(addr);
+#ifdef __CUDA_ARCH__
+#if defined(__LP64__)
+  // 64-bit pointer operand constraint for inlined asm
+  asm(" prefetch.L1 [ %1 ];" : "=l"(addr) : "l"(addr));
+#else
+  // 32-bit pointer operand constraint for inlined asm
+  asm(" prefetch.L1 [ %1 ];" : "=r"(addr) : "r"(addr));
+#endif
+#elif (!EIGEN_COMP_MSVC) && (EIGEN_COMP_GNUC || EIGEN_COMP_CLANG || EIGEN_COMP_ICC)
+  __builtin_prefetch(addr);
 #endif
 }
 
 /** \internal \returns the first element of a packet */
-template<typename Packet> inline typename unpacket_traits<Packet>::type pfirst(const Packet& a)
+template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type pfirst(const Packet& a)
 { return a; }
 
 /** \internal \returns a packet where the element i contains the sum of the packet of \a vec[i] */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 preduxp(const Packet* vecs) { return vecs[0]; }
 
 /** \internal \returns the sum of the elements of \a a*/
-template<typename Packet> inline typename unpacket_traits<Packet>::type predux(const Packet& a)
+template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type predux(const Packet& a)
+{ return a; }
+
+/** \internal \returns the sum of the elements of \a a by block of 4 elements.
+  * For a packet {a0, a1, a2, a3, a4, a5, a6, a7}, it returns a half packet {a0+a4, a1+a5, a2+a6, a3+a7}
+  * For packet-size smaller or equal to 4, this boils down to a noop.
+  */
+template<typename Packet> EIGEN_DEVICE_FUNC inline
+typename conditional<(unpacket_traits<Packet>::size%8)==0,typename unpacket_traits<Packet>::half,Packet>::type
+predux_downto4(const Packet& a)
 { return a; }
 
 /** \internal \returns the product of the elements of \a a*/
-template<typename Packet> inline typename unpacket_traits<Packet>::type predux_mul(const Packet& a)
+template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type predux_mul(const Packet& a)
 { return a; }
 
 /** \internal \returns the min of the elements of \a a*/
-template<typename Packet> inline typename unpacket_traits<Packet>::type predux_min(const Packet& a)
+template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type predux_min(const Packet& a)
 { return a; }
 
 /** \internal \returns the max of the elements of \a a*/
-template<typename Packet> inline typename unpacket_traits<Packet>::type predux_max(const Packet& a)
+template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type predux_max(const Packet& a)
 { return a; }
 
 /** \internal \returns the reversed elements of \a a*/
-template<typename Packet> inline Packet preverse(const Packet& a)
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet preverse(const Packet& a)
 { return a; }
 
-
 /** \internal \returns \a a with real and imaginary part flipped (for complex type only) */
-template<typename Packet> inline Packet pcplxflip(const Packet& a)
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet pcplxflip(const Packet& a)
 {
   // FIXME: uncomment the following in case we drop the internal imag and real functions.
 //   using std::imag;
@@ -250,18 +382,64 @@ Packet pasin(const Packet& a) { using std::asin; return asin(a); }
 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
 Packet pacos(const Packet& a) { using std::acos; return acos(a); }
 
+/** \internal \returns the arc tangent of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet patan(const Packet& a) { using std::atan; return atan(a); }
+
+/** \internal \returns the hyperbolic sine of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet psinh(const Packet& a) { using std::sinh; return sinh(a); }
+
+/** \internal \returns the hyperbolic cosine of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pcosh(const Packet& a) { using std::cosh; return cosh(a); }
+
+/** \internal \returns the hyperbolic tan of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet ptanh(const Packet& a) { using std::tanh; return tanh(a); }
+
 /** \internal \returns the exp of \a a (coeff-wise) */
 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
 Packet pexp(const Packet& a) { using std::exp; return exp(a); }
 
+/** \internal \returns the expm1 of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pexpm1(const Packet& a) { return numext::expm1(a); }
+
 /** \internal \returns the log of \a a (coeff-wise) */
 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
 Packet plog(const Packet& a) { using std::log; return log(a); }
 
+/** \internal \returns the log1p of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet plog1p(const Packet& a) { return numext::log1p(a); }
+
+/** \internal \returns the log10 of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet plog10(const Packet& a) { using std::log10; return log10(a); }
+
 /** \internal \returns the square-root of \a a (coeff-wise) */
 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
 Packet psqrt(const Packet& a) { using std::sqrt; return sqrt(a); }
 
+/** \internal \returns the reciprocal square-root of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet prsqrt(const Packet& a) {
+  return pdiv(pset1<Packet>(1), psqrt(a));
+}
+
+/** \internal \returns the rounded value of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pround(const Packet& a) { using numext::round; return round(a); }
+
+/** \internal \returns the floor of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pfloor(const Packet& a) { using numext::floor; return floor(a); }
+
+/** \internal \returns the ceil of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pceil(const Packet& a) { using numext::ceil; return ceil(a); }
+
 /***************************************************************************
 * The following functions might not have to be overwritten for vectorized types
 ***************************************************************************/
@@ -275,34 +453,45 @@ inline void pstore1(typename unpacket_traits<Packet>::type* to, const typename u
 }
 
 /** \internal \returns a * b + c (coeff-wise) */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pmadd(const Packet&  a,
          const Packet&  b,
          const Packet&  c)
 { return padd(pmul(a, b),c); }
 
 /** \internal \returns a packet version of \a *from.
-  * If LoadMode equals #Aligned, \a from must be 16 bytes aligned */
-template<typename Packet, int LoadMode>
-inline Packet ploadt(const typename unpacket_traits<Packet>::type* from)
+  * The pointer \a from must be aligned on a \a Alignment bytes boundary. */
+template<typename Packet, int Alignment>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet ploadt(const typename unpacket_traits<Packet>::type* from)
 {
-  if(LoadMode == Aligned)
+  if(Alignment >= unpacket_traits<Packet>::alignment)
     return pload<Packet>(from);
   else
     return ploadu<Packet>(from);
 }
 
 /** \internal copy the packet \a from to \a *to.
-  * If StoreMode equals #Aligned, \a to must be 16 bytes aligned */
-template<typename Scalar, typename Packet, int LoadMode>
-inline void pstoret(Scalar* to, const Packet& from)
+  * The pointer \a from must be aligned on a \a Alignment bytes boundary. */
+template<typename Scalar, typename Packet, int Alignment>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void pstoret(Scalar* to, const Packet& from)
 {
-  if(LoadMode == Aligned)
+  if(Alignment >= unpacket_traits<Packet>::alignment)
     pstore(to, from);
   else
     pstoreu(to, from);
 }
 
+/** \internal \returns a packet version of \a *from.
+  * Unlike ploadt, ploadt_ro takes advantage of the read-only memory path on the
+  * hardware if available to speedup the loading of data that won't be modified
+  * by the current computation.
+  */
+template<typename Packet, int LoadMode>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet ploadt_ro(const typename unpacket_traits<Packet>::type* from)
+{
+  return ploadt<Packet, LoadMode>(from);
+}
+
 /** \internal default implementation of palign() allowing partial specialization */
 template<int Offset,typename PacketType>
 struct palign_impl
@@ -336,15 +525,74 @@ inline void palign(PacketType& first, const PacketType& second)
 * Fast complex products (GCC generates a function call which is very slow)
 ***************************************************************************/
 
+// Eigen+CUDA does not support complexes.
+#ifndef __CUDACC__
+
 template<> inline std::complex<float> pmul(const std::complex<float>& a, const std::complex<float>& b)
 { return std::complex<float>(real(a)*real(b) - imag(a)*imag(b), imag(a)*real(b) + real(a)*imag(b)); }
 
 template<> inline std::complex<double> pmul(const std::complex<double>& a, const std::complex<double>& b)
 { return std::complex<double>(real(a)*real(b) - imag(a)*imag(b), imag(a)*real(b) + real(a)*imag(b)); }
 
+#endif
+
+
+/***************************************************************************
+ * PacketBlock, that is a collection of N packets where the number of words
+ * in the packet is a multiple of N.
+***************************************************************************/
+template <typename Packet,int N=unpacket_traits<Packet>::size> struct PacketBlock {
+  Packet packet[N];
+};
+
+template<typename Packet> EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet,1>& /*kernel*/) {
+  // Nothing to do in the scalar case, i.e. a 1x1 matrix.
+}
+
+/***************************************************************************
+ * Selector, i.e. vector of N boolean values used to select (i.e. blend)
+ * words from 2 packets.
+***************************************************************************/
+template <size_t N> struct Selector {
+  bool select[N];
+};
+
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pblend(const Selector<unpacket_traits<Packet>::size>& ifPacket, const Packet& thenPacket, const Packet& elsePacket) {
+  return ifPacket.select[0] ? thenPacket : elsePacket;
+}
+
+/** \internal \returns \a a with the first coefficient replaced by the scalar b */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pinsertfirst(const Packet& a, typename unpacket_traits<Packet>::type b)
+{
+  // Default implementation based on pblend.
+  // It must be specialized for higher performance.
+  Selector<unpacket_traits<Packet>::size> mask;
+  mask.select[0] = true;
+  // This for loop should be optimized away by the compiler.
+  for(Index i=1; i<unpacket_traits<Packet>::size; ++i)
+    mask.select[i] = false;
+  return pblend(mask, pset1<Packet>(b), a);
+}
+
+/** \internal \returns \a a with the last coefficient replaced by the scalar b */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pinsertlast(const Packet& a, typename unpacket_traits<Packet>::type b)
+{
+  // Default implementation based on pblend.
+  // It must be specialized for higher performance.
+  Selector<unpacket_traits<Packet>::size> mask;
+  // This for loop should be optimized away by the compiler.
+  for(Index i=0; i<unpacket_traits<Packet>::size-1; ++i)
+    mask.select[i] = false;
+  mask.select[unpacket_traits<Packet>::size-1] = true;
+  return pblend(mask, pset1<Packet>(b), a);
+}
+
 } // end namespace internal
 
 } // end namespace Eigen
 
 #endif // EIGEN_GENERIC_PACKET_MATH_H
-
diff --git a/eigen/Eigen/src/Core/GlobalFunctions.h b/eigen/Eigen/src/Core/GlobalFunctions.h
index 2acf977..12828a7 100644
--- a/eigen/Eigen/src/Core/GlobalFunctions.h
+++ b/eigen/Eigen/src/Core/GlobalFunctions.h
@@ -1,7 +1,7 @@
 // This file is part of Eigen, a lightweight C++ template library
 // for linear algebra.
 //
-// Copyright (C) 2010-2012 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2010-2016 Gael Guennebaud <gael.guennebaud@inria.fr>
 // Copyright (C) 2010 Benoit Jacob <jacob.benoit.1@gmail.com>
 //
 // This Source Code Form is subject to the terms of the Mozilla
@@ -11,13 +11,30 @@
 #ifndef EIGEN_GLOBAL_FUNCTIONS_H
 #define EIGEN_GLOBAL_FUNCTIONS_H
 
-#define EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(NAME,FUNCTOR) \
+#ifdef EIGEN_PARSED_BY_DOXYGEN
+
+#define EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(NAME,FUNCTOR,DOC_OP,DOC_DETAILS) \
+  /** \returns an expression of the coefficient-wise DOC_OP of \a x
+
+    DOC_DETAILS
+
+    \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_##NAME">Math functions</a>, class CwiseUnaryOp
+    */ \
+  template<typename Derived> \
+  inline const Eigen::CwiseUnaryOp<Eigen::internal::FUNCTOR<typename Derived::Scalar>, const Derived> \
+  NAME(const Eigen::ArrayBase<Derived>& x);
+
+#else
+
+#define EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(NAME,FUNCTOR,DOC_OP,DOC_DETAILS) \
   template<typename Derived> \
   inline const Eigen::CwiseUnaryOp<Eigen::internal::FUNCTOR<typename Derived::Scalar>, const Derived> \
-  NAME(const Eigen::ArrayBase<Derived>& x) { \
-    return x.derived(); \
+  (NAME)(const Eigen::ArrayBase<Derived>& x) { \
+    return Eigen::CwiseUnaryOp<Eigen::internal::FUNCTOR<typename Derived::Scalar>, const Derived>(x.derived()); \
   }
 
+#endif // EIGEN_PARSED_BY_DOXYGEN
+
 #define EIGEN_ARRAY_DECLARE_GLOBAL_EIGEN_UNARY(NAME,FUNCTOR) \
   \
   template<typename Derived> \
@@ -30,55 +47,134 @@
   { \
     static inline typename NAME##_retval<ArrayBase<Derived> >::type run(const Eigen::ArrayBase<Derived>& x) \
     { \
-      return x.derived(); \
+      return typename NAME##_retval<ArrayBase<Derived> >::type(x.derived()); \
     } \
   };
 
-
 namespace Eigen
 {
-  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(real,scalar_real_op)
-  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(imag,scalar_imag_op)
-  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(conj,scalar_conjugate_op)
-  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sin,scalar_sin_op)
-  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(cos,scalar_cos_op)
-  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(asin,scalar_asin_op)
-  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(acos,scalar_acos_op)
-  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(tan,scalar_tan_op)
-  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(exp,scalar_exp_op)
-  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(log,scalar_log_op)
-  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(abs,scalar_abs_op)
-  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sqrt,scalar_sqrt_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(real,scalar_real_op,real part,\sa ArrayBase::real)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(imag,scalar_imag_op,imaginary part,\sa ArrayBase::imag)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(conj,scalar_conjugate_op,complex conjugate,\sa ArrayBase::conjugate)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(inverse,scalar_inverse_op,inverse,\sa ArrayBase::inverse)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sin,scalar_sin_op,sine,\sa ArrayBase::sin)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(cos,scalar_cos_op,cosine,\sa ArrayBase::cos)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(tan,scalar_tan_op,tangent,\sa ArrayBase::tan)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(atan,scalar_atan_op,arc-tangent,\sa ArrayBase::atan)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(asin,scalar_asin_op,arc-sine,\sa ArrayBase::asin)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(acos,scalar_acos_op,arc-consine,\sa ArrayBase::acos)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sinh,scalar_sinh_op,hyperbolic sine,\sa ArrayBase::sinh)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(cosh,scalar_cosh_op,hyperbolic cosine,\sa ArrayBase::cosh)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(tanh,scalar_tanh_op,hyperbolic tangent,\sa ArrayBase::tanh)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(lgamma,scalar_lgamma_op,natural logarithm of the gamma function,\sa ArrayBase::lgamma)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(digamma,scalar_digamma_op,derivative of lgamma,\sa ArrayBase::digamma)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(erf,scalar_erf_op,error function,\sa ArrayBase::erf)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(erfc,scalar_erfc_op,complement error function,\sa ArrayBase::erfc)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(exp,scalar_exp_op,exponential,\sa ArrayBase::exp)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(expm1,scalar_expm1_op,exponential of a value minus 1,\sa ArrayBase::expm1)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(log,scalar_log_op,natural logarithm,\sa Eigen::log10 DOXCOMMA ArrayBase::log)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(log1p,scalar_log1p_op,natural logarithm of 1 plus the value,\sa ArrayBase::log1p)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(log10,scalar_log10_op,base 10 logarithm,\sa Eigen::log DOXCOMMA ArrayBase::log)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(abs,scalar_abs_op,absolute value,\sa ArrayBase::abs DOXCOMMA MatrixBase::cwiseAbs)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(abs2,scalar_abs2_op,squared absolute value,\sa ArrayBase::abs2 DOXCOMMA MatrixBase::cwiseAbs2)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(arg,scalar_arg_op,complex argument,\sa ArrayBase::arg)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sqrt,scalar_sqrt_op,square root,\sa ArrayBase::sqrt DOXCOMMA MatrixBase::cwiseSqrt)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(rsqrt,scalar_rsqrt_op,reciprocal square root,\sa ArrayBase::rsqrt)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(square,scalar_square_op,square (power 2),\sa Eigen::abs2 DOXCOMMA Eigen::pow DOXCOMMA ArrayBase::square)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(cube,scalar_cube_op,cube (power 3),\sa Eigen::pow DOXCOMMA ArrayBase::cube)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(round,scalar_round_op,nearest integer,\sa Eigen::floor DOXCOMMA Eigen::ceil DOXCOMMA ArrayBase::round)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(floor,scalar_floor_op,nearest integer not greater than the giben value,\sa Eigen::ceil DOXCOMMA ArrayBase::floor)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(ceil,scalar_ceil_op,nearest integer not less than the giben value,\sa Eigen::floor DOXCOMMA ArrayBase::ceil)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(isnan,scalar_isnan_op,not-a-number test,\sa Eigen::isinf DOXCOMMA Eigen::isfinite DOXCOMMA ArrayBase::isnan)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(isinf,scalar_isinf_op,infinite value test,\sa Eigen::isnan DOXCOMMA Eigen::isfinite DOXCOMMA ArrayBase::isinf)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(isfinite,scalar_isfinite_op,finite value test,\sa Eigen::isinf DOXCOMMA Eigen::isnan DOXCOMMA ArrayBase::isfinite)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sign,scalar_sign_op,sign (or 0),\sa ArrayBase::sign)
   
+  /** \returns an expression of the coefficient-wise power of \a x to the given constant \a exponent.
+    *
+    * \tparam ScalarExponent is the scalar type of \a exponent. It must be compatible with the scalar type of the given expression (\c Derived::Scalar).
+    *
+    * \sa ArrayBase::pow()
+    *
+    * \relates ArrayBase
+    */
+#ifdef EIGEN_PARSED_BY_DOXYGEN
+  template<typename Derived,typename ScalarExponent>
+  inline const CwiseBinaryOp<internal::scalar_pow_op<Derived::Scalar,ScalarExponent>,Derived,Constant<ScalarExponent> >
+  pow(const Eigen::ArrayBase<Derived>& x, const ScalarExponent& exponent);
+#else
+  template<typename Derived,typename ScalarExponent>
+  inline typename internal::enable_if<   !(internal::is_same<typename Derived::Scalar,ScalarExponent>::value) && EIGEN_SCALAR_BINARY_SUPPORTED(pow,typename Derived::Scalar,ScalarExponent),
+          const EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(Derived,ScalarExponent,pow) >::type
+  pow(const Eigen::ArrayBase<Derived>& x, const ScalarExponent& exponent) {
+    return x.derived().pow(exponent);
+  }
+
   template<typename Derived>
-  inline const Eigen::CwiseUnaryOp<Eigen::internal::scalar_pow_op<typename Derived::Scalar>, const Derived>
+  inline const EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(Derived,typename Derived::Scalar,pow)
   pow(const Eigen::ArrayBase<Derived>& x, const typename Derived::Scalar& exponent) {
     return x.derived().pow(exponent);
   }
+#endif
 
-  template<typename Derived>
-  inline const Eigen::CwiseBinaryOp<Eigen::internal::scalar_binary_pow_op<typename Derived::Scalar, typename Derived::Scalar>, const Derived, const Derived>
-  pow(const Eigen::ArrayBase<Derived>& x, const Eigen::ArrayBase<Derived>& exponents) 
+  /** \returns an expression of the coefficient-wise power of \a x to the given array of \a exponents.
+    *
+    * This function computes the coefficient-wise power.
+    *
+    * Example: \include Cwise_array_power_array.cpp
+    * Output: \verbinclude Cwise_array_power_array.out
+    * 
+    * \sa ArrayBase::pow()
+    *
+    * \relates ArrayBase
+    */
+  template<typename Derived,typename ExponentDerived>
+  inline const Eigen::CwiseBinaryOp<Eigen::internal::scalar_pow_op<typename Derived::Scalar, typename ExponentDerived::Scalar>, const Derived, const ExponentDerived>
+  pow(const Eigen::ArrayBase<Derived>& x, const Eigen::ArrayBase<ExponentDerived>& exponents) 
   {
-    return Eigen::CwiseBinaryOp<Eigen::internal::scalar_binary_pow_op<typename Derived::Scalar, typename Derived::Scalar>, const Derived, const Derived>(
+    return Eigen::CwiseBinaryOp<Eigen::internal::scalar_pow_op<typename Derived::Scalar, typename ExponentDerived::Scalar>, const Derived, const ExponentDerived>(
       x.derived(),
       exponents.derived()
     );
   }
   
-  /**
-  * \brief Component-wise division of a scalar by array elements.
-  **/
-  template <typename Derived>
-  inline const Eigen::CwiseUnaryOp<Eigen::internal::scalar_inverse_mult_op<typename Derived::Scalar>, const Derived>
-    operator/(const typename Derived::Scalar& s, const Eigen::ArrayBase<Derived>& a)
+  /** \returns an expression of the coefficient-wise power of the scalar \a x to the given array of \a exponents.
+    *
+    * This function computes the coefficient-wise power between a scalar and an array of exponents.
+    *
+    * \tparam Scalar is the scalar type of \a x. It must be compatible with the scalar type of the given array expression (\c Derived::Scalar).
+    *
+    * Example: \include Cwise_scalar_power_array.cpp
+    * Output: \verbinclude Cwise_scalar_power_array.out
+    * 
+    * \sa ArrayBase::pow()
+    *
+    * \relates ArrayBase
+    */
+#ifdef EIGEN_PARSED_BY_DOXYGEN
+  template<typename Scalar,typename Derived>
+  inline const CwiseBinaryOp<internal::scalar_pow_op<Scalar,Derived::Scalar>,Constant<Scalar>,Derived>
+  pow(const Scalar& x,const Eigen::ArrayBase<Derived>& x);
+#else
+  template<typename Scalar, typename Derived>
+  inline typename internal::enable_if<   !(internal::is_same<typename Derived::Scalar,Scalar>::value) && EIGEN_SCALAR_BINARY_SUPPORTED(pow,Scalar,typename Derived::Scalar),
+          const EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(Scalar,Derived,pow) >::type
+  pow(const Scalar& x, const Eigen::ArrayBase<Derived>& exponents)
   {
-    return Eigen::CwiseUnaryOp<Eigen::internal::scalar_inverse_mult_op<typename Derived::Scalar>, const Derived>(
-      a.derived(),
-      Eigen::internal::scalar_inverse_mult_op<typename Derived::Scalar>(s)  
-    );
+    return EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(Scalar,Derived,pow)(
+            typename internal::plain_constant_type<Derived,Scalar>::type(exponents.rows(), exponents.cols(), x), exponents.derived() );
   }
 
+  template<typename Derived>
+  inline const EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(typename Derived::Scalar,Derived,pow)
+  pow(const typename Derived::Scalar& x, const Eigen::ArrayBase<Derived>& exponents)
+  {
+    return EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(typename Derived::Scalar,Derived,pow)(
+      typename internal::plain_constant_type<Derived,typename Derived::Scalar>::type(exponents.rows(), exponents.cols(), x), exponents.derived() );
+  }
+#endif
+
+
   namespace internal
   {
     EIGEN_ARRAY_DECLARE_GLOBAL_EIGEN_UNARY(real,scalar_real_op)
diff --git a/eigen/Eigen/src/Core/IO.h b/eigen/Eigen/src/Core/IO.h
index 8d4bc59..da7fd6c 100644
--- a/eigen/Eigen/src/Core/IO.h
+++ b/eigen/Eigen/src/Core/IO.h
@@ -49,7 +49,7 @@ std::ostream & print_matrix(std::ostream & s, const Derived& _m, const IOFormat&
   */
 struct IOFormat
 {
-  /** Default contructor, see class IOFormat for the meaning of the parameters */
+  /** Default constructor, see class IOFormat for the meaning of the parameters */
   IOFormat(int _precision = StreamPrecision, int _flags = 0,
     const std::string& _coeffSeparator = " ",
     const std::string& _rowSeparator = "\n", const std::string& _rowPrefix="", const std::string& _rowSuffix="",
@@ -57,6 +57,10 @@ struct IOFormat
   : matPrefix(_matPrefix), matSuffix(_matSuffix), rowPrefix(_rowPrefix), rowSuffix(_rowSuffix), rowSeparator(_rowSeparator),
     rowSpacer(""), coeffSeparator(_coeffSeparator), precision(_precision), flags(_flags)
   {
+    // TODO check if rowPrefix, rowSuffix or rowSeparator contains a newline
+    // don't add rowSpacer if columns are not to be aligned
+    if((flags & DontAlignCols))
+      return;
     int i = int(matSuffix.length())-1;
     while (i>=0 && matSuffix[i]!='\n')
     {
@@ -76,7 +80,7 @@ struct IOFormat
   *
   * \brief Pseudo expression providing matrix output with given format
   *
-  * \param ExpressionType the type of the object on which IO stream operations are performed
+  * \tparam ExpressionType the type of the object on which IO stream operations are performed
   *
   * This class represents an expression with stream operators controlled by a given IOFormat.
   * It is the return type of DenseBase::format()
@@ -101,52 +105,24 @@ class WithFormat
     }
 
   protected:
-    const typename ExpressionType::Nested m_matrix;
+    typename ExpressionType::Nested m_matrix;
     IOFormat m_format;
 };
 
-/** \returns a WithFormat proxy object allowing to print a matrix the with given
-  * format \a fmt.
-  *
-  * See class IOFormat for some examples.
-  *
-  * \sa class IOFormat, class WithFormat
-  */
-template<typename Derived>
-inline const WithFormat<Derived>
-DenseBase<Derived>::format(const IOFormat& fmt) const
-{
-  return WithFormat<Derived>(derived(), fmt);
-}
-
 namespace internal {
 
-template<typename Scalar, bool IsInteger>
-struct significant_decimals_default_impl
-{
-  typedef typename NumTraits<Scalar>::Real RealScalar;
-  static inline int run()
-  {
-    using std::ceil;
-    using std::log;
-    return cast<RealScalar,int>(ceil(-log(NumTraits<RealScalar>::epsilon())/log(RealScalar(10))));
-  }
-};
-
+// NOTE: This helper is kept for backward compatibility with previous code specializing
+//       this internal::significant_decimals_impl structure. In the future we should directly
+//       call digits10() which has been introduced in July 2016 in 3.3.
 template<typename Scalar>
-struct significant_decimals_default_impl<Scalar, true>
+struct significant_decimals_impl
 {
   static inline int run()
   {
-    return 0;
+    return NumTraits<Scalar>::digits10();
   }
 };
 
-template<typename Scalar>
-struct significant_decimals_impl
-  : significant_decimals_default_impl<Scalar, NumTraits<Scalar>::IsInteger>
-{};
-
 /** \internal
   * print the matrix \a _m to the output stream \a s using the output format \a fmt */
 template<typename Derived>
@@ -160,7 +136,6 @@ std::ostream & print_matrix(std::ostream & s, const Derived& _m, const IOFormat&
   
   typename Derived::Nested m = _m;
   typedef typename Derived::Scalar Scalar;
-  typedef typename Derived::Index Index;
 
   Index width = 0;
 
diff --git a/eigen/Eigen/src/Core/IndexedView.h b/eigen/Eigen/src/Core/IndexedView.h
new file mode 100644
index 0000000..8c57a27
--- /dev/null
+++ b/eigen/Eigen/src/Core/IndexedView.h
@@ -0,0 +1,207 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2017 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_INDEXED_VIEW_H
+#define EIGEN_INDEXED_VIEW_H
+
+namespace Eigen {
+
+namespace internal {
+
+template<typename XprType, typename RowIndices, typename ColIndices>
+struct traits<IndexedView<XprType, RowIndices, ColIndices> >
+ : traits<XprType>
+{
+  enum {
+    RowsAtCompileTime = int(array_size<RowIndices>::value),
+    ColsAtCompileTime = int(array_size<ColIndices>::value),
+    MaxRowsAtCompileTime = RowsAtCompileTime != Dynamic ? int(RowsAtCompileTime) : int(traits<XprType>::MaxRowsAtCompileTime),
+    MaxColsAtCompileTime = ColsAtCompileTime != Dynamic ? int(ColsAtCompileTime) : int(traits<XprType>::MaxColsAtCompileTime),
+
+    XprTypeIsRowMajor = (int(traits<XprType>::Flags)&RowMajorBit) != 0,
+    IsRowMajor = (MaxRowsAtCompileTime==1&&MaxColsAtCompileTime!=1) ? 1
+               : (MaxColsAtCompileTime==1&&MaxRowsAtCompileTime!=1) ? 0
+               : XprTypeIsRowMajor,
+
+    RowIncr = int(get_compile_time_incr<RowIndices>::value),
+    ColIncr = int(get_compile_time_incr<ColIndices>::value),
+    InnerIncr = IsRowMajor ? ColIncr : RowIncr,
+    OuterIncr = IsRowMajor ? RowIncr : ColIncr,
+
+    HasSameStorageOrderAsXprType = (IsRowMajor == XprTypeIsRowMajor),
+    XprInnerStride = HasSameStorageOrderAsXprType ? int(inner_stride_at_compile_time<XprType>::ret) : int(outer_stride_at_compile_time<XprType>::ret),
+    XprOuterstride = HasSameStorageOrderAsXprType ? int(outer_stride_at_compile_time<XprType>::ret) : int(inner_stride_at_compile_time<XprType>::ret),
+
+    InnerSize = XprTypeIsRowMajor ? ColsAtCompileTime : RowsAtCompileTime,
+    IsBlockAlike = InnerIncr==1 && OuterIncr==1,
+    IsInnerPannel = HasSameStorageOrderAsXprType && is_same<AllRange<InnerSize>,typename conditional<XprTypeIsRowMajor,ColIndices,RowIndices>::type>::value,
+
+    InnerStrideAtCompileTime = InnerIncr<0 || InnerIncr==DynamicIndex || XprInnerStride==Dynamic ? Dynamic : XprInnerStride * InnerIncr,
+    OuterStrideAtCompileTime = OuterIncr<0 || OuterIncr==DynamicIndex || XprOuterstride==Dynamic ? Dynamic : XprOuterstride * OuterIncr,
+
+    ReturnAsScalar = is_same<RowIndices,SingleRange>::value && is_same<ColIndices,SingleRange>::value,
+    ReturnAsBlock = (!ReturnAsScalar) && IsBlockAlike,
+    ReturnAsIndexedView = (!ReturnAsScalar) && (!ReturnAsBlock),
+
+    // FIXME we deal with compile-time strides if and only if we have DirectAccessBit flag,
+    // but this is too strict regarding negative strides...
+    DirectAccessMask = (int(InnerIncr)!=UndefinedIncr && int(OuterIncr)!=UndefinedIncr && InnerIncr>=0 && OuterIncr>=0) ? DirectAccessBit : 0,
+    FlagsRowMajorBit = IsRowMajor ? RowMajorBit : 0,
+    FlagsLvalueBit = is_lvalue<XprType>::value ? LvalueBit : 0,
+    Flags = (traits<XprType>::Flags & (HereditaryBits | DirectAccessMask)) | FlagsLvalueBit | FlagsRowMajorBit
+  };
+
+  typedef Block<XprType,RowsAtCompileTime,ColsAtCompileTime,IsInnerPannel> BlockType;
+};
+
+}
+
+template<typename XprType, typename RowIndices, typename ColIndices, typename StorageKind>
+class IndexedViewImpl;
+
+
+/** \class IndexedView
+  * \ingroup Core_Module
+  *
+  * \brief Expression of a non-sequential sub-matrix defined by arbitrary sequences of row and column indices
+  *
+  * \tparam XprType the type of the expression in which we are taking the intersections of sub-rows and sub-columns
+  * \tparam RowIndices the type of the object defining the sequence of row indices
+  * \tparam ColIndices the type of the object defining the sequence of column indices
+  *
+  * This class represents an expression of a sub-matrix (or sub-vector) defined as the intersection
+  * of sub-sets of rows and columns, that are themself defined by generic sequences of row indices \f$ \{r_0,r_1,..r_{m-1}\} \f$
+  * and column indices \f$ \{c_0,c_1,..c_{n-1} \}\f$. Let \f$ A \f$  be the nested matrix, then the resulting matrix \f$ B \f$ has \c m
+  * rows and \c n columns, and its entries are given by: \f$ B(i,j) = A(r_i,c_j) \f$.
+  *
+  * The \c RowIndices and \c ColIndices types must be compatible with the following API:
+  * \code
+  * <integral type> operator[](Index) const;
+  * Index size() const;
+  * \endcode
+  *
+  * Typical supported types thus include:
+  *  - std::vector<int>
+  *  - std::valarray<int>
+  *  - std::array<int>
+  *  - Plain C arrays: int[N]
+  *  - Eigen::ArrayXi
+  *  - decltype(ArrayXi::LinSpaced(...))
+  *  - Any view/expressions of the previous types
+  *  - Eigen::ArithmeticSequence
+  *  - Eigen::internal::AllRange      (helper for Eigen::all)
+  *  - Eigen::internal::SingleRange  (helper for single index)
+  *  - etc.
+  *
+  * In typical usages of %Eigen, this class should never be used directly. It is the return type of
+  * DenseBase::operator()(const RowIndices&, const ColIndices&).
+  *
+  * \sa class Block
+  */
+template<typename XprType, typename RowIndices, typename ColIndices>
+class IndexedView : public IndexedViewImpl<XprType, RowIndices, ColIndices, typename internal::traits<XprType>::StorageKind>
+{
+public:
+  typedef typename IndexedViewImpl<XprType, RowIndices, ColIndices, typename internal::traits<XprType>::StorageKind>::Base Base;
+  EIGEN_GENERIC_PUBLIC_INTERFACE(IndexedView)
+  EIGEN_INHERIT_ASSIGNMENT_OPERATORS(IndexedView)
+
+  typedef typename internal::ref_selector<XprType>::non_const_type MatrixTypeNested;
+  typedef typename internal::remove_all<XprType>::type NestedExpression;
+
+  template<typename T0, typename T1>
+  IndexedView(XprType& xpr, const T0& rowIndices, const T1& colIndices)
+    : m_xpr(xpr), m_rowIndices(rowIndices), m_colIndices(colIndices)
+  {}
+
+  /** \returns number of rows */
+  Index rows() const { return internal::size(m_rowIndices); }
+
+  /** \returns number of columns */
+  Index cols() const { return internal::size(m_colIndices); }
+
+  /** \returns the nested expression */
+  const typename internal::remove_all<XprType>::type&
+  nestedExpression() const { return m_xpr; }
+
+  /** \returns the nested expression */
+  typename internal::remove_reference<XprType>::type&
+  nestedExpression() { return m_xpr.const_cast_derived(); }
+
+  /** \returns a const reference to the object storing/generating the row indices */
+  const RowIndices& rowIndices() const { return m_rowIndices; }
+
+  /** \returns a const reference to the object storing/generating the column indices */
+  const ColIndices& colIndices() const { return m_colIndices; }
+
+protected:
+  MatrixTypeNested m_xpr;
+  RowIndices m_rowIndices;
+  ColIndices m_colIndices;
+};
+
+
+// Generic API dispatcher
+template<typename XprType, typename RowIndices, typename ColIndices, typename StorageKind>
+class IndexedViewImpl
+  : public internal::generic_xpr_base<IndexedView<XprType, RowIndices, ColIndices> >::type
+{
+public:
+  typedef typename internal::generic_xpr_base<IndexedView<XprType, RowIndices, ColIndices> >::type Base;
+};
+
+namespace internal {
+
+
+template<typename ArgType, typename RowIndices, typename ColIndices>
+struct unary_evaluator<IndexedView<ArgType, RowIndices, ColIndices>, IndexBased>
+  : evaluator_base<IndexedView<ArgType, RowIndices, ColIndices> >
+{
+  typedef IndexedView<ArgType, RowIndices, ColIndices> XprType;
+
+  enum {
+    CoeffReadCost = evaluator<ArgType>::CoeffReadCost /* TODO + cost of row/col index */,
+
+    Flags = (evaluator<ArgType>::Flags & (HereditaryBits /*| LinearAccessBit | DirectAccessBit*/)),
+
+    Alignment = 0
+  };
+
+  EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& xpr) : m_argImpl(xpr.nestedExpression()), m_xpr(xpr)
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index col) const
+  {
+    return m_argImpl.coeff(m_xpr.rowIndices()[row], m_xpr.colIndices()[col]);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index row, Index col)
+  {
+    return m_argImpl.coeffRef(m_xpr.rowIndices()[row], m_xpr.colIndices()[col]);
+  }
+
+protected:
+
+  evaluator<ArgType> m_argImpl;
+  const XprType& m_xpr;
+
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_INDEXED_VIEW_H
diff --git a/eigen/Eigen/src/Core/Inverse.h b/eigen/Eigen/src/Core/Inverse.h
new file mode 100644
index 0000000..b76f043
--- /dev/null
+++ b/eigen/Eigen/src/Core/Inverse.h
@@ -0,0 +1,118 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_INVERSE_H
+#define EIGEN_INVERSE_H
+
+namespace Eigen { 
+
+template<typename XprType,typename StorageKind> class InverseImpl;
+
+namespace internal {
+
+template<typename XprType>
+struct traits<Inverse<XprType> >
+  : traits<typename XprType::PlainObject>
+{
+  typedef typename XprType::PlainObject PlainObject;
+  typedef traits<PlainObject> BaseTraits;
+  enum {
+    Flags = BaseTraits::Flags & RowMajorBit
+  };
+};
+
+} // end namespace internal
+
+/** \class Inverse
+  *
+  * \brief Expression of the inverse of another expression
+  *
+  * \tparam XprType the type of the expression we are taking the inverse
+  *
+  * This class represents an abstract expression of A.inverse()
+  * and most of the time this is the only way it is used.
+  *
+  */
+template<typename XprType>
+class Inverse : public InverseImpl<XprType,typename internal::traits<XprType>::StorageKind>
+{
+public:
+  typedef typename XprType::StorageIndex StorageIndex;
+  typedef typename XprType::PlainObject                       PlainObject;
+  typedef typename XprType::Scalar                            Scalar;
+  typedef typename internal::ref_selector<XprType>::type      XprTypeNested;
+  typedef typename internal::remove_all<XprTypeNested>::type  XprTypeNestedCleaned;
+  typedef typename internal::ref_selector<Inverse>::type Nested;
+  typedef typename internal::remove_all<XprType>::type NestedExpression;
+  
+  explicit EIGEN_DEVICE_FUNC Inverse(const XprType &xpr)
+    : m_xpr(xpr)
+  {}
+
+  EIGEN_DEVICE_FUNC Index rows() const { return m_xpr.rows(); }
+  EIGEN_DEVICE_FUNC Index cols() const { return m_xpr.cols(); }
+
+  EIGEN_DEVICE_FUNC const XprTypeNestedCleaned& nestedExpression() const { return m_xpr; }
+
+protected:
+  XprTypeNested m_xpr;
+};
+
+// Generic API dispatcher
+template<typename XprType, typename StorageKind>
+class InverseImpl
+  : public internal::generic_xpr_base<Inverse<XprType> >::type
+{
+public:
+  typedef typename internal::generic_xpr_base<Inverse<XprType> >::type Base;
+  typedef typename XprType::Scalar Scalar;
+private:
+
+  Scalar coeff(Index row, Index col) const;
+  Scalar coeff(Index i) const;
+};
+
+namespace internal {
+
+/** \internal
+  * \brief Default evaluator for Inverse expression.
+  * 
+  * This default evaluator for Inverse expression simply evaluate the inverse into a temporary
+  * by a call to internal::call_assignment_no_alias.
+  * Therefore, inverse implementers only have to specialize Assignment<Dst,Inverse<...>, ...> for
+  * there own nested expression.
+  *
+  * \sa class Inverse
+  */
+template<typename ArgType>
+struct unary_evaluator<Inverse<ArgType> >
+  : public evaluator<typename Inverse<ArgType>::PlainObject>
+{
+  typedef Inverse<ArgType> InverseType;
+  typedef typename InverseType::PlainObject PlainObject;
+  typedef evaluator<PlainObject> Base;
+  
+  enum { Flags = Base::Flags | EvalBeforeNestingBit };
+
+  unary_evaluator(const InverseType& inv_xpr)
+    : m_result(inv_xpr.rows(), inv_xpr.cols())
+  {
+    ::new (static_cast<Base*>(this)) Base(m_result);
+    internal::call_assignment_no_alias(m_result, inv_xpr);
+  }
+  
+protected:
+  PlainObject m_result;
+};
+  
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_INVERSE_H
diff --git a/eigen/Eigen/src/Core/Map.h b/eigen/Eigen/src/Core/Map.h
index f804c89..06d1967 100644
--- a/eigen/Eigen/src/Core/Map.h
+++ b/eigen/Eigen/src/Core/Map.h
@@ -13,13 +13,35 @@
 
 namespace Eigen { 
 
+namespace internal {
+template<typename PlainObjectType, int MapOptions, typename StrideType>
+struct traits<Map<PlainObjectType, MapOptions, StrideType> >
+  : public traits<PlainObjectType>
+{
+  typedef traits<PlainObjectType> TraitsBase;
+  enum {
+    InnerStrideAtCompileTime = StrideType::InnerStrideAtCompileTime == 0
+                             ? int(PlainObjectType::InnerStrideAtCompileTime)
+                             : int(StrideType::InnerStrideAtCompileTime),
+    OuterStrideAtCompileTime = StrideType::OuterStrideAtCompileTime == 0
+                             ? int(PlainObjectType::OuterStrideAtCompileTime)
+                             : int(StrideType::OuterStrideAtCompileTime),
+    Alignment = int(MapOptions)&int(AlignedMask),
+    Flags0 = TraitsBase::Flags & (~NestByRefBit),
+    Flags = is_lvalue<PlainObjectType>::value ? int(Flags0) : (int(Flags0) & ~LvalueBit)
+  };
+private:
+  enum { Options }; // Expressions don't have Options
+};
+}
+
 /** \class Map
   * \ingroup Core_Module
   *
   * \brief A matrix or vector expression mapping an existing array of data.
   *
   * \tparam PlainObjectType the equivalent matrix type of the mapped data
-  * \tparam MapOptions specifies whether the pointer is \c #Aligned, or \c #Unaligned.
+  * \tparam MapOptions specifies the pointer alignment in bytes. It can be: \c #Aligned128, , \c #Aligned64, \c #Aligned32, \c #Aligned16, \c #Aligned8 or \c #Unaligned.
   *                The default is \c #Unaligned.
   * \tparam StrideType optionally specifies strides. By default, Map assumes the memory layout
   *                   of an ordinary, contiguous array. This can be overridden by specifying strides.
@@ -63,44 +85,6 @@ namespace Eigen {
   *
   * \sa PlainObjectBase::Map(), \ref TopicStorageOrders
   */
-
-namespace internal {
-template<typename PlainObjectType, int MapOptions, typename StrideType>
-struct traits<Map<PlainObjectType, MapOptions, StrideType> >
-  : public traits<PlainObjectType>
-{
-  typedef traits<PlainObjectType> TraitsBase;
-  typedef typename PlainObjectType::Index Index;
-  typedef typename PlainObjectType::Scalar Scalar;
-  enum {
-    InnerStrideAtCompileTime = StrideType::InnerStrideAtCompileTime == 0
-                             ? int(PlainObjectType::InnerStrideAtCompileTime)
-                             : int(StrideType::InnerStrideAtCompileTime),
-    OuterStrideAtCompileTime = StrideType::OuterStrideAtCompileTime == 0
-                             ? int(PlainObjectType::OuterStrideAtCompileTime)
-                             : int(StrideType::OuterStrideAtCompileTime),
-    HasNoInnerStride = InnerStrideAtCompileTime == 1,
-    HasNoOuterStride = StrideType::OuterStrideAtCompileTime == 0,
-    HasNoStride = HasNoInnerStride && HasNoOuterStride,
-    IsAligned = bool(EIGEN_ALIGN) && ((int(MapOptions)&Aligned)==Aligned),
-    IsDynamicSize = PlainObjectType::SizeAtCompileTime==Dynamic,
-    KeepsPacketAccess = bool(HasNoInnerStride)
-                        && ( bool(IsDynamicSize)
-                           || HasNoOuterStride
-                           || ( OuterStrideAtCompileTime!=Dynamic
-                           && ((static_cast<int>(sizeof(Scalar))*OuterStrideAtCompileTime)%16)==0 ) ),
-    Flags0 = TraitsBase::Flags & (~NestByRefBit),
-    Flags1 = IsAligned ? (int(Flags0) | AlignedBit) : (int(Flags0) & ~AlignedBit),
-    Flags2 = (bool(HasNoStride) || bool(PlainObjectType::IsVectorAtCompileTime))
-           ? int(Flags1) : int(Flags1 & ~LinearAccessBit),
-    Flags3 = is_lvalue<PlainObjectType>::value ? int(Flags2) : (int(Flags2) & ~LvalueBit),
-    Flags = KeepsPacketAccess ? int(Flags3) : (int(Flags3) & ~PacketAccessBit)
-  };
-private:
-  enum { Options }; // Expressions don't have Options
-};
-}
-
 template<typename PlainObjectType, int MapOptions, typename StrideType> class Map
   : public MapBase<Map<PlainObjectType, MapOptions, StrideType> >
 {
@@ -110,19 +94,17 @@ template<typename PlainObjectType, int MapOptions, typename StrideType> class Ma
     EIGEN_DENSE_PUBLIC_INTERFACE(Map)
 
     typedef typename Base::PointerType PointerType;
-#if EIGEN2_SUPPORT_STAGE <= STAGE30_FULL_EIGEN3_API
-    typedef const Scalar* PointerArgType;
-    inline PointerType cast_to_pointer_type(PointerArgType ptr) { return const_cast<PointerType>(ptr); }
-#else
     typedef PointerType PointerArgType;
+    EIGEN_DEVICE_FUNC
     inline PointerType cast_to_pointer_type(PointerArgType ptr) { return ptr; }
-#endif
 
+    EIGEN_DEVICE_FUNC
     inline Index innerStride() const
     {
       return StrideType::InnerStrideAtCompileTime != 0 ? m_stride.inner() : 1;
     }
 
+    EIGEN_DEVICE_FUNC
     inline Index outerStride() const
     {
       return StrideType::OuterStrideAtCompileTime != 0 ? m_stride.outer()
@@ -134,10 +116,11 @@ template<typename PlainObjectType, int MapOptions, typename StrideType> class Ma
     /** Constructor in the fixed-size case.
       *
       * \param dataPtr pointer to the array to map
-      * \param a_stride optional Stride object, passing the strides.
+      * \param stride optional Stride object, passing the strides.
       */
-    inline Map(PointerArgType dataPtr, const StrideType& a_stride = StrideType())
-      : Base(cast_to_pointer_type(dataPtr)), m_stride(a_stride)
+    EIGEN_DEVICE_FUNC
+    explicit inline Map(PointerArgType dataPtr, const StrideType& stride = StrideType())
+      : Base(cast_to_pointer_type(dataPtr)), m_stride(stride)
     {
       PlainObjectType::Base::_check_template_params();
     }
@@ -145,11 +128,12 @@ template<typename PlainObjectType, int MapOptions, typename StrideType> class Ma
     /** Constructor in the dynamic-size vector case.
       *
       * \param dataPtr pointer to the array to map
-      * \param a_size the size of the vector expression
-      * \param a_stride optional Stride object, passing the strides.
+      * \param size the size of the vector expression
+      * \param stride optional Stride object, passing the strides.
       */
-    inline Map(PointerArgType dataPtr, Index a_size, const StrideType& a_stride = StrideType())
-      : Base(cast_to_pointer_type(dataPtr), a_size), m_stride(a_stride)
+    EIGEN_DEVICE_FUNC
+    inline Map(PointerArgType dataPtr, Index size, const StrideType& stride = StrideType())
+      : Base(cast_to_pointer_type(dataPtr), size), m_stride(stride)
     {
       PlainObjectType::Base::_check_template_params();
     }
@@ -157,12 +141,13 @@ template<typename PlainObjectType, int MapOptions, typename StrideType> class Ma
     /** Constructor in the dynamic-size matrix case.
       *
       * \param dataPtr pointer to the array to map
-      * \param nbRows the number of rows of the matrix expression
-      * \param nbCols the number of columns of the matrix expression
-      * \param a_stride optional Stride object, passing the strides.
+      * \param rows the number of rows of the matrix expression
+      * \param cols the number of columns of the matrix expression
+      * \param stride optional Stride object, passing the strides.
       */
-    inline Map(PointerArgType dataPtr, Index nbRows, Index nbCols, const StrideType& a_stride = StrideType())
-      : Base(cast_to_pointer_type(dataPtr), nbRows, nbCols), m_stride(a_stride)
+    EIGEN_DEVICE_FUNC
+    inline Map(PointerArgType dataPtr, Index rows, Index cols, const StrideType& stride = StrideType())
+      : Base(cast_to_pointer_type(dataPtr), rows, cols), m_stride(stride)
     {
       PlainObjectType::Base::_check_template_params();
     }
@@ -173,19 +158,6 @@ template<typename PlainObjectType, int MapOptions, typename StrideType> class Ma
     StrideType m_stride;
 };
 
-template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
-inline Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>
-  ::Array(const Scalar *data)
-{
-  this->_set_noalias(Eigen::Map<const Array>(data));
-}
-
-template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
-inline Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>
-  ::Matrix(const Scalar *data)
-{
-  this->_set_noalias(Eigen::Map<const Matrix>(data));
-}
 
 } // end namespace Eigen
 
diff --git a/eigen/Eigen/src/Core/MapBase.h b/eigen/Eigen/src/Core/MapBase.h
index 81efc4a..020f939 100644
--- a/eigen/Eigen/src/Core/MapBase.h
+++ b/eigen/Eigen/src/Core/MapBase.h
@@ -12,15 +12,25 @@
 #define EIGEN_MAPBASE_H
 
 #define EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived) \
-      EIGEN_STATIC_ASSERT((int(internal::traits<Derived>::Flags) & LinearAccessBit) || Derived::IsVectorAtCompileTime, \
+      EIGEN_STATIC_ASSERT((int(internal::evaluator<Derived>::Flags) & LinearAccessBit) || Derived::IsVectorAtCompileTime, \
                           YOU_ARE_TRYING_TO_USE_AN_INDEX_BASED_ACCESSOR_ON_AN_EXPRESSION_THAT_DOES_NOT_SUPPORT_THAT)
 
 namespace Eigen { 
 
-/** \class MapBase
-  * \ingroup Core_Module
+/** \ingroup Core_Module
   *
-  * \brief Base class for Map and Block expression with direct access
+  * \brief Base class for dense Map and Block expression with direct access
+  *
+  * This base class provides the const low-level accessors (e.g. coeff, coeffRef) of dense
+  * Map and Block objects with direct access.
+  * Typical users do not have to directly deal with this class.
+  *
+  * This class can be extended by through the macro plugin \c EIGEN_MAPBASE_PLUGIN.
+  * See \link TopicCustomizing_Plugins customizing Eigen \endlink for details.
+  *
+  * The \c Derived class has to provide the following two methods describing the memory layout:
+  *  \code Index innerStride() const; \endcode
+  *  \code Index outerStride() const; \endcode
   *
   * \sa class Map, class Block
   */
@@ -37,7 +47,6 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
     };
 
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
-    typedef typename internal::traits<Derived>::Index Index;
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef typename internal::packet_traits<Scalar>::type PacketScalar;
     typedef typename NumTraits<Scalar>::Real RealScalar;
@@ -76,8 +85,10 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
 
     typedef typename Base::CoeffReturnType CoeffReturnType;
 
-    inline Index rows() const { return m_rows.value(); }
-    inline Index cols() const { return m_cols.value(); }
+    /** \copydoc DenseBase::rows() */
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_rows.value(); }
+    /** \copydoc DenseBase::cols() */
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_cols.value(); }
 
     /** Returns a pointer to the first coefficient of the matrix or vector.
       *
@@ -85,30 +96,39 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
       *
       * \sa innerStride(), outerStride()
       */
-    inline const Scalar* data() const { return m_data; }
+    EIGEN_DEVICE_FUNC inline const Scalar* data() const { return m_data; }
 
+    /** \copydoc PlainObjectBase::coeff(Index,Index) const */
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeff(Index rowId, Index colId) const
     {
       return m_data[colId * colStride() + rowId * rowStride()];
     }
 
+    /** \copydoc PlainObjectBase::coeff(Index) const */
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeff(Index index) const
     {
       EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived)
       return m_data[index * innerStride()];
     }
 
+    /** \copydoc PlainObjectBase::coeffRef(Index,Index) const */
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index rowId, Index colId) const
     {
       return this->m_data[colId * colStride() + rowId * rowStride()];
     }
 
+    /** \copydoc PlainObjectBase::coeffRef(Index) const */
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index index) const
     {
       EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived)
       return this->m_data[index * innerStride()];
     }
 
+    /** \internal */
     template<int LoadMode>
     inline PacketScalar packet(Index rowId, Index colId) const
     {
@@ -116,6 +136,7 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
                (m_data + (colId * colStride() + rowId * rowStride()));
     }
 
+    /** \internal */
     template<int LoadMode>
     inline PacketScalar packet(Index index) const
     {
@@ -123,12 +144,16 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
       return internal::ploadt<PacketScalar, LoadMode>(m_data + index * innerStride());
     }
 
+    /** \internal Constructor for fixed size matrices or vectors */
+    EIGEN_DEVICE_FUNC
     explicit inline MapBase(PointerType dataPtr) : m_data(dataPtr), m_rows(RowsAtCompileTime), m_cols(ColsAtCompileTime)
     {
       EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived)
-      checkSanity();
+      checkSanity<Derived>();
     }
 
+    /** \internal Constructor for dynamically sized vectors */
+    EIGEN_DEVICE_FUNC
     inline MapBase(PointerType dataPtr, Index vecSize)
             : m_data(dataPtr),
               m_rows(RowsAtCompileTime == Dynamic ? vecSize : Index(RowsAtCompileTime)),
@@ -137,16 +162,18 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
       EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
       eigen_assert(vecSize >= 0);
       eigen_assert(dataPtr == 0 || SizeAtCompileTime == Dynamic || SizeAtCompileTime == vecSize);
-      checkSanity();
+      checkSanity<Derived>();
     }
 
-    inline MapBase(PointerType dataPtr, Index nbRows, Index nbCols)
-            : m_data(dataPtr), m_rows(nbRows), m_cols(nbCols)
+    /** \internal Constructor for dynamically sized matrices */
+    EIGEN_DEVICE_FUNC
+    inline MapBase(PointerType dataPtr, Index rows, Index cols)
+            : m_data(dataPtr), m_rows(rows), m_cols(cols)
     {
       eigen_assert( (dataPtr == 0)
-              || (   nbRows >= 0 && (RowsAtCompileTime == Dynamic || RowsAtCompileTime == nbRows)
-                  && nbCols >= 0 && (ColsAtCompileTime == Dynamic || ColsAtCompileTime == nbCols)));
-      checkSanity();
+              || (   rows >= 0 && (RowsAtCompileTime == Dynamic || RowsAtCompileTime == rows)
+                  && cols >= 0 && (ColsAtCompileTime == Dynamic || ColsAtCompileTime == cols)));
+      checkSanity<Derived>();
     }
 
     #ifdef EIGEN_MAPBASE_PLUGIN
@@ -155,20 +182,36 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
 
   protected:
 
-    void checkSanity() const
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    void checkSanity(typename internal::enable_if<(internal::traits<T>::Alignment>0),void*>::type = 0) const
     {
-      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(internal::traits<Derived>::Flags&PacketAccessBit,
-                                        internal::inner_stride_at_compile_time<Derived>::ret==1),
-                          PACKET_ACCESS_REQUIRES_TO_HAVE_INNER_STRIDE_FIXED_TO_1);
-      eigen_assert(EIGEN_IMPLIES(internal::traits<Derived>::Flags&AlignedBit, (size_t(m_data) % 16) == 0)
-                   && "input pointer is not aligned on a 16 byte boundary");
+#if EIGEN_MAX_ALIGN_BYTES>0
+      eigen_assert((   ((internal::UIntPtr(m_data) % internal::traits<Derived>::Alignment) == 0)
+                    || (cols() * rows() * innerStride() * sizeof(Scalar)) < internal::traits<Derived>::Alignment ) && "data is not aligned");
+#endif
     }
 
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    void checkSanity(typename internal::enable_if<internal::traits<T>::Alignment==0,void*>::type = 0) const
+    {}
+
     PointerType m_data;
     const internal::variable_if_dynamic<Index, RowsAtCompileTime> m_rows;
     const internal::variable_if_dynamic<Index, ColsAtCompileTime> m_cols;
 };
 
+/** \ingroup Core_Module
+  *
+  * \brief Base class for non-const dense Map and Block expression with direct access
+  *
+  * This base class provides the non-const low-level accessors (e.g. coeff and coeffRef) of
+  * dense Map and Block objects with direct access.
+  * It inherits MapBase<Derived, ReadOnlyAccessors> which defines the const variant for reading specific entries.
+  *
+  * \sa class Map, class Block
+  */
 template<typename Derived> class MapBase<Derived, WriteAccessors>
   : public MapBase<Derived, ReadOnlyAccessors>
 {
@@ -179,7 +222,7 @@ template<typename Derived> class MapBase<Derived, WriteAccessors>
 
     typedef typename Base::Scalar Scalar;
     typedef typename Base::PacketScalar PacketScalar;
-    typedef typename Base::Index Index;
+    typedef typename Base::StorageIndex StorageIndex;
     typedef typename Base::PointerType PointerType;
 
     using Base::derived;
@@ -200,14 +243,18 @@ template<typename Derived> class MapBase<Derived, WriteAccessors>
                     const Scalar
                   >::type ScalarWithConstIfNotLvalue;
 
+    EIGEN_DEVICE_FUNC
     inline const Scalar* data() const { return this->m_data; }
+    EIGEN_DEVICE_FUNC
     inline ScalarWithConstIfNotLvalue* data() { return this->m_data; } // no const-cast here so non-const-correct code will give a compile error
 
+    EIGEN_DEVICE_FUNC
     inline ScalarWithConstIfNotLvalue& coeffRef(Index row, Index col)
     {
       return this->m_data[col * colStride() + row * rowStride()];
     }
 
+    EIGEN_DEVICE_FUNC
     inline ScalarWithConstIfNotLvalue& coeffRef(Index index)
     {
       EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived)
@@ -229,10 +276,11 @@ template<typename Derived> class MapBase<Derived, WriteAccessors>
                 (this->m_data + index * innerStride(), val);
     }
 
-    explicit inline MapBase(PointerType dataPtr) : Base(dataPtr) {}
-    inline MapBase(PointerType dataPtr, Index vecSize) : Base(dataPtr, vecSize) {}
-    inline MapBase(PointerType dataPtr, Index nbRows, Index nbCols) : Base(dataPtr, nbRows, nbCols) {}
+    EIGEN_DEVICE_FUNC explicit inline MapBase(PointerType dataPtr) : Base(dataPtr) {}
+    EIGEN_DEVICE_FUNC inline MapBase(PointerType dataPtr, Index vecSize) : Base(dataPtr, vecSize) {}
+    EIGEN_DEVICE_FUNC inline MapBase(PointerType dataPtr, Index rows, Index cols) : Base(dataPtr, rows, cols) {}
 
+    EIGEN_DEVICE_FUNC
     Derived& operator=(const MapBase& other)
     {
       ReadOnlyMapBase::Base::operator=(other);
diff --git a/eigen/Eigen/src/Core/MathFunctions.h b/eigen/Eigen/src/Core/MathFunctions.h
index dc2c698..5ec6c39 100644
--- a/eigen/Eigen/src/Core/MathFunctions.h
+++ b/eigen/Eigen/src/Core/MathFunctions.h
@@ -10,11 +10,24 @@
 #ifndef EIGEN_MATHFUNCTIONS_H
 #define EIGEN_MATHFUNCTIONS_H
 
+// source: http://www.geom.uiuc.edu/~huberty/math5337/groupe/digits.html
+// TODO this should better be moved to NumTraits
+#define EIGEN_PI 3.141592653589793238462643383279502884197169399375105820974944592307816406L
+
 namespace Eigen {
 
+// On WINCE, std::abs is defined for int only, so let's defined our own overloads:
+// This issue has been confirmed with MSVC 2008 only, but the issue might exist for more recent versions too.
+#if EIGEN_OS_WINCE && EIGEN_COMP_MSVC && EIGEN_COMP_MSVC<=1500
+long        abs(long        x) { return (labs(x));  }
+double      abs(double      x) { return (fabs(x));  }
+float       abs(float       x) { return (fabsf(x)); }
+long double abs(long double x) { return (fabsl(x)); }
+#endif
+
 namespace internal {
 
-/** \internal \struct global_math_functions_filtering_base
+/** \internal \class global_math_functions_filtering_base
   *
   * What it does:
   * Defines a typedef 'type' as follows:
@@ -62,6 +75,7 @@ template<typename Scalar, bool IsComplex = NumTraits<Scalar>::IsComplex>
 struct real_default_impl
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
   static inline RealScalar run(const Scalar& x)
   {
     return x;
@@ -72,6 +86,7 @@ template<typename Scalar>
 struct real_default_impl<Scalar,true>
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
   static inline RealScalar run(const Scalar& x)
   {
     using std::real;
@@ -81,13 +96,25 @@ struct real_default_impl<Scalar,true>
 
 template<typename Scalar> struct real_impl : real_default_impl<Scalar> {};
 
+#ifdef __CUDA_ARCH__
+template<typename T>
+struct real_impl<std::complex<T> >
+{
+  typedef T RealScalar;
+  EIGEN_DEVICE_FUNC
+  static inline T run(const std::complex<T>& x)
+  {
+    return x.real();
+  }
+};
+#endif
+
 template<typename Scalar>
 struct real_retval
 {
   typedef typename NumTraits<Scalar>::Real type;
 };
 
-
 /****************************************************************************
 * Implementation of imag                                                 *
 ****************************************************************************/
@@ -96,6 +123,7 @@ template<typename Scalar, bool IsComplex = NumTraits<Scalar>::IsComplex>
 struct imag_default_impl
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
   static inline RealScalar run(const Scalar&)
   {
     return RealScalar(0);
@@ -106,6 +134,7 @@ template<typename Scalar>
 struct imag_default_impl<Scalar,true>
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
   static inline RealScalar run(const Scalar& x)
   {
     using std::imag;
@@ -115,6 +144,19 @@ struct imag_default_impl<Scalar,true>
 
 template<typename Scalar> struct imag_impl : imag_default_impl<Scalar> {};
 
+#ifdef __CUDA_ARCH__
+template<typename T>
+struct imag_impl<std::complex<T> >
+{
+  typedef T RealScalar;
+  EIGEN_DEVICE_FUNC
+  static inline T run(const std::complex<T>& x)
+  {
+    return x.imag();
+  }
+};
+#endif
+
 template<typename Scalar>
 struct imag_retval
 {
@@ -129,10 +171,12 @@ template<typename Scalar>
 struct real_ref_impl
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
   static inline RealScalar& run(Scalar& x)
   {
     return reinterpret_cast<RealScalar*>(&x)[0];
   }
+  EIGEN_DEVICE_FUNC
   static inline const RealScalar& run(const Scalar& x)
   {
     return reinterpret_cast<const RealScalar*>(&x)[0];
@@ -153,10 +197,12 @@ template<typename Scalar, bool IsComplex>
 struct imag_ref_default_impl
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
   static inline RealScalar& run(Scalar& x)
   {
     return reinterpret_cast<RealScalar*>(&x)[1];
   }
+  EIGEN_DEVICE_FUNC
   static inline const RealScalar& run(const Scalar& x)
   {
     return reinterpret_cast<RealScalar*>(&x)[1];
@@ -166,10 +212,12 @@ struct imag_ref_default_impl
 template<typename Scalar>
 struct imag_ref_default_impl<Scalar, false>
 {
+  EIGEN_DEVICE_FUNC
   static inline Scalar run(Scalar&)
   {
     return Scalar(0);
   }
+  EIGEN_DEVICE_FUNC
   static inline const Scalar run(const Scalar&)
   {
     return Scalar(0);
@@ -192,6 +240,7 @@ struct imag_ref_retval
 template<typename Scalar, bool IsComplex = NumTraits<Scalar>::IsComplex>
 struct conj_impl
 {
+  EIGEN_DEVICE_FUNC
   static inline Scalar run(const Scalar& x)
   {
     return x;
@@ -201,6 +250,7 @@ struct conj_impl
 template<typename Scalar>
 struct conj_impl<Scalar,true>
 {
+  EIGEN_DEVICE_FUNC
   static inline Scalar run(const Scalar& x)
   {
     using std::conj;
@@ -222,6 +272,7 @@ template<typename Scalar,bool IsComplex>
 struct abs2_impl_default
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
   static inline RealScalar run(const Scalar& x)
   {
     return x*x;
@@ -232,6 +283,7 @@ template<typename Scalar>
 struct abs2_impl_default<Scalar, true> // IsComplex
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
   static inline RealScalar run(const Scalar& x)
   {
     return real(x)*real(x) + imag(x)*imag(x);
@@ -242,6 +294,7 @@ template<typename Scalar>
 struct abs2_impl
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
   static inline RealScalar run(const Scalar& x)
   {
     return abs2_impl_default<Scalar,NumTraits<Scalar>::IsComplex>::run(x);
@@ -262,9 +315,10 @@ template<typename Scalar, bool IsComplex>
 struct norm1_default_impl
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
   static inline RealScalar run(const Scalar& x)
   {
-    using std::abs;
+    EIGEN_USING_STD_MATH(abs);
     return abs(real(x)) + abs(imag(x));
   }
 };
@@ -272,9 +326,10 @@ struct norm1_default_impl
 template<typename Scalar>
 struct norm1_default_impl<Scalar, false>
 {
+  EIGEN_DEVICE_FUNC
   static inline Scalar run(const Scalar& x)
   {
-    using std::abs;
+    EIGEN_USING_STD_MATH(abs);
     return abs(x);
   }
 };
@@ -298,16 +353,22 @@ struct hypot_impl
   typedef typename NumTraits<Scalar>::Real RealScalar;
   static inline RealScalar run(const Scalar& x, const Scalar& y)
   {
-    using std::max;
-    using std::min;
-    using std::abs;
-    using std::sqrt;
+    EIGEN_USING_STD_MATH(abs);
+    EIGEN_USING_STD_MATH(sqrt);
     RealScalar _x = abs(x);
     RealScalar _y = abs(y);
-    RealScalar p = (max)(_x, _y);
+    Scalar p, qp;
+    if(_x>_y)
+    {
+      p = _x;
+      qp = _y / p;
+    }
+    else
+    {
+      p = _y;
+      qp = _x / p;
+    }
     if(p==RealScalar(0)) return RealScalar(0);
-    RealScalar q = (min)(_x, _y);
-    RealScalar qp = q/p;
     return p * sqrt(RealScalar(1) + qp*qp);
   }
 };
@@ -325,6 +386,7 @@ struct hypot_retval
 template<typename OldType, typename NewType>
 struct cast_impl
 {
+  EIGEN_DEVICE_FUNC
   static inline NewType run(const OldType& x)
   {
     return static_cast<NewType>(x);
@@ -334,48 +396,173 @@ struct cast_impl
 // here, for once, we're plainly returning NewType: we don't want cast to do weird things.
 
 template<typename OldType, typename NewType>
+EIGEN_DEVICE_FUNC
 inline NewType cast(const OldType& x)
 {
   return cast_impl<OldType, NewType>::run(x);
 }
 
 /****************************************************************************
-* Implementation of atanh2                                                *
+* Implementation of round                                                   *
 ****************************************************************************/
 
-template<typename Scalar, bool IsInteger>
-struct atanh2_default_impl
+#if EIGEN_HAS_CXX11_MATH
+  template<typename Scalar>
+  struct round_impl {
+    static inline Scalar run(const Scalar& x)
+    {
+      EIGEN_STATIC_ASSERT((!NumTraits<Scalar>::IsComplex), NUMERIC_TYPE_MUST_BE_REAL)
+      EIGEN_USING_STD_MATH(round);
+      return round(x);
+    }
+  };
+#else
+  template<typename Scalar>
+  struct round_impl
+  {
+    static inline Scalar run(const Scalar& x)
+    {
+      EIGEN_STATIC_ASSERT((!NumTraits<Scalar>::IsComplex), NUMERIC_TYPE_MUST_BE_REAL)
+      EIGEN_USING_STD_MATH(floor);
+      EIGEN_USING_STD_MATH(ceil);
+      return (x > Scalar(0)) ? floor(x + Scalar(0.5)) : ceil(x - Scalar(0.5));
+    }
+  };
+#endif
+
+template<typename Scalar>
+struct round_retval
 {
-  typedef Scalar retval;
-  typedef typename NumTraits<Scalar>::Real RealScalar;
-  static inline Scalar run(const Scalar& x, const Scalar& y)
+  typedef Scalar type;
+};
+
+/****************************************************************************
+* Implementation of arg                                                     *
+****************************************************************************/
+
+#if EIGEN_HAS_CXX11_MATH
+  template<typename Scalar>
+  struct arg_impl {
+    static inline Scalar run(const Scalar& x)
+    {
+      EIGEN_USING_STD_MATH(arg);
+      return arg(x);
+    }
+  };
+#else
+  template<typename Scalar, bool IsComplex = NumTraits<Scalar>::IsComplex>
+  struct arg_default_impl
   {
-    using std::abs;
-    using std::log;
-    using std::sqrt;
-    Scalar z = x / y;
-    if (y == Scalar(0) || abs(z) > sqrt(NumTraits<RealScalar>::epsilon()))
-      return RealScalar(0.5) * log((y + x) / (y - x));
-    else
-      return z + z*z*z / RealScalar(3);
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    EIGEN_DEVICE_FUNC
+    static inline RealScalar run(const Scalar& x)
+    {
+      return (x < Scalar(0)) ? Scalar(EIGEN_PI) : Scalar(0); }
+  };
+
+  template<typename Scalar>
+  struct arg_default_impl<Scalar,true>
+  {
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    EIGEN_DEVICE_FUNC
+    static inline RealScalar run(const Scalar& x)
+    {
+      EIGEN_USING_STD_MATH(arg);
+      return arg(x);
+    }
+  };
+
+  template<typename Scalar> struct arg_impl : arg_default_impl<Scalar> {};
+#endif
+
+template<typename Scalar>
+struct arg_retval
+{
+  typedef typename NumTraits<Scalar>::Real type;
+};
+
+/****************************************************************************
+* Implementation of expm1                                                   *
+****************************************************************************/
+
+// This implementation is based on GSL Math's expm1.
+namespace std_fallback {
+  // fallback expm1 implementation in case there is no expm1(Scalar) function in namespace of Scalar,
+  // or that there is no suitable std::expm1 function available. Implementation
+  // attributed to Kahan. See: http://www.plunk.org/~hatch/rightway.php.
+  template<typename Scalar>
+  EIGEN_DEVICE_FUNC inline Scalar expm1(const Scalar& x) {
+    EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar)
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+
+    EIGEN_USING_STD_MATH(exp);
+    Scalar u = exp(x);
+    if (u == Scalar(1)) {
+      return x;
+    }
+    Scalar um1 = u - RealScalar(1);
+    if (um1 == Scalar(-1)) {
+      return RealScalar(-1);
+    }
+
+    EIGEN_USING_STD_MATH(log);
+    return (u - RealScalar(1)) * x / log(u);
+  }
+}
+
+template<typename Scalar>
+struct expm1_impl {
+  EIGEN_DEVICE_FUNC static inline Scalar run(const Scalar& x)
+  {
+    EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar)
+    #if EIGEN_HAS_CXX11_MATH
+    using std::expm1;
+    #endif
+    using std_fallback::expm1;
+    return expm1(x);
   }
 };
 
+
 template<typename Scalar>
-struct atanh2_default_impl<Scalar, true>
+struct expm1_retval
 {
-  static inline Scalar run(const Scalar&, const Scalar&)
+  typedef Scalar type;
+};
+
+/****************************************************************************
+* Implementation of log1p                                                   *
+****************************************************************************/
+
+namespace std_fallback {
+  // fallback log1p implementation in case there is no log1p(Scalar) function in namespace of Scalar,
+  // or that there is no suitable std::log1p function available
+  template<typename Scalar>
+  EIGEN_DEVICE_FUNC inline Scalar log1p(const Scalar& x) {
+    EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar)
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    EIGEN_USING_STD_MATH(log);
+    Scalar x1p = RealScalar(1) + x;
+    return ( x1p == Scalar(1) ) ? x : x * ( log(x1p) / (x1p - RealScalar(1)) );
+  }
+}
+
+template<typename Scalar>
+struct log1p_impl {
+  EIGEN_DEVICE_FUNC static inline Scalar run(const Scalar& x)
   {
     EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar)
-    return Scalar(0);
+    #if EIGEN_HAS_CXX11_MATH
+    using std::log1p;
+    #endif
+    using std_fallback::log1p;
+    return log1p(x);
   }
 };
 
-template<typename Scalar>
-struct atanh2_impl : atanh2_default_impl<Scalar, NumTraits<Scalar>::IsInteger> {};
 
 template<typename Scalar>
-struct atanh2_retval
+struct log1p_retval
 {
   typedef Scalar type;
 };
@@ -384,24 +571,26 @@ struct atanh2_retval
 * Implementation of pow                                                  *
 ****************************************************************************/
 
-template<typename Scalar, bool IsInteger>
-struct pow_default_impl
+template<typename ScalarX,typename ScalarY, bool IsInteger = NumTraits<ScalarX>::IsInteger&&NumTraits<ScalarY>::IsInteger>
+struct pow_impl
 {
-  typedef Scalar retval;
-  static inline Scalar run(const Scalar& x, const Scalar& y)
+  //typedef Scalar retval;
+  typedef typename ScalarBinaryOpTraits<ScalarX,ScalarY,internal::scalar_pow_op<ScalarX,ScalarY> >::ReturnType result_type;
+  static EIGEN_DEVICE_FUNC inline result_type run(const ScalarX& x, const ScalarY& y)
   {
-    using std::pow;
+    EIGEN_USING_STD_MATH(pow);
     return pow(x, y);
   }
 };
 
-template<typename Scalar>
-struct pow_default_impl<Scalar, true>
+template<typename ScalarX,typename ScalarY>
+struct pow_impl<ScalarX,ScalarY, true>
 {
-  static inline Scalar run(Scalar x, Scalar y)
+  typedef ScalarX result_type;
+  static EIGEN_DEVICE_FUNC inline ScalarX run(ScalarX x, ScalarY y)
   {
-    Scalar res(1);
-    eigen_assert(!NumTraits<Scalar>::IsSigned || y >= 0);
+    ScalarX res(1);
+    eigen_assert(!NumTraits<ScalarY>::IsSigned || y >= 0);
     if(y & 1) res *= x;
     y >>= 1;
     while(y)
@@ -414,15 +603,6 @@ struct pow_default_impl<Scalar, true>
   }
 };
 
-template<typename Scalar>
-struct pow_impl : pow_default_impl<Scalar, NumTraits<Scalar>::IsInteger> {};
-
-template<typename Scalar>
-struct pow_retval
-{
-  typedef Scalar type;
-};
-
 /****************************************************************************
 * Implementation of random                                               *
 ****************************************************************************/
@@ -458,48 +638,48 @@ struct random_default_impl<Scalar, false, false>
 };
 
 enum {
-  floor_log2_terminate,
-  floor_log2_move_up,
-  floor_log2_move_down,
-  floor_log2_bogus
+  meta_floor_log2_terminate,
+  meta_floor_log2_move_up,
+  meta_floor_log2_move_down,
+  meta_floor_log2_bogus
 };
 
-template<unsigned int n, int lower, int upper> struct floor_log2_selector
+template<unsigned int n, int lower, int upper> struct meta_floor_log2_selector
 {
   enum { middle = (lower + upper) / 2,
-         value = (upper <= lower + 1) ? int(floor_log2_terminate)
-               : (n < (1 << middle)) ? int(floor_log2_move_down)
-               : (n==0) ? int(floor_log2_bogus)
-               : int(floor_log2_move_up)
+         value = (upper <= lower + 1) ? int(meta_floor_log2_terminate)
+               : (n < (1 << middle)) ? int(meta_floor_log2_move_down)
+               : (n==0) ? int(meta_floor_log2_bogus)
+               : int(meta_floor_log2_move_up)
   };
 };
 
 template<unsigned int n,
          int lower = 0,
          int upper = sizeof(unsigned int) * CHAR_BIT - 1,
-         int selector = floor_log2_selector<n, lower, upper>::value>
-struct floor_log2 {};
+         int selector = meta_floor_log2_selector<n, lower, upper>::value>
+struct meta_floor_log2 {};
 
 template<unsigned int n, int lower, int upper>
-struct floor_log2<n, lower, upper, floor_log2_move_down>
+struct meta_floor_log2<n, lower, upper, meta_floor_log2_move_down>
 {
-  enum { value = floor_log2<n, lower, floor_log2_selector<n, lower, upper>::middle>::value };
+  enum { value = meta_floor_log2<n, lower, meta_floor_log2_selector<n, lower, upper>::middle>::value };
 };
 
 template<unsigned int n, int lower, int upper>
-struct floor_log2<n, lower, upper, floor_log2_move_up>
+struct meta_floor_log2<n, lower, upper, meta_floor_log2_move_up>
 {
-  enum { value = floor_log2<n, floor_log2_selector<n, lower, upper>::middle, upper>::value };
+  enum { value = meta_floor_log2<n, meta_floor_log2_selector<n, lower, upper>::middle, upper>::value };
 };
 
 template<unsigned int n, int lower, int upper>
-struct floor_log2<n, lower, upper, floor_log2_terminate>
+struct meta_floor_log2<n, lower, upper, meta_floor_log2_terminate>
 {
   enum { value = (n >= ((unsigned int)(1) << (lower+1))) ? lower+1 : lower };
 };
 
 template<unsigned int n, int lower, int upper>
-struct floor_log2<n, lower, upper, floor_log2_bogus>
+struct meta_floor_log2<n, lower, upper, meta_floor_log2_bogus>
 {
   // no value, error at compile time
 };
@@ -507,11 +687,24 @@ struct floor_log2<n, lower, upper, floor_log2_bogus>
 template<typename Scalar>
 struct random_default_impl<Scalar, false, true>
 {
-  typedef typename NumTraits<Scalar>::NonInteger NonInteger;
-
   static inline Scalar run(const Scalar& x, const Scalar& y)
   {
-    return x + Scalar((NonInteger(y)-x+1) * std::rand() / (RAND_MAX + NonInteger(1)));
+    typedef typename conditional<NumTraits<Scalar>::IsSigned,std::ptrdiff_t,std::size_t>::type ScalarX;
+    if(y<x)
+      return x;
+    // the following difference might overflow on a 32 bits system,
+    // but since y>=x the result converted to an unsigned long is still correct.
+    std::size_t range = ScalarX(y)-ScalarX(x);
+    std::size_t offset = 0;
+    // rejection sampling
+    std::size_t divisor = 1;
+    std::size_t multiplier = 1;
+    if(range<RAND_MAX) divisor = (std::size_t(RAND_MAX)+1)/(range+1);
+    else               multiplier = 1 + range/(std::size_t(RAND_MAX)+1);
+    do {
+      offset = (std::size_t(std::rand()) * multiplier) / divisor;
+    } while (offset > range);
+    return Scalar(ScalarX(x) + offset);
   }
 
   static inline Scalar run()
@@ -519,7 +712,7 @@ struct random_default_impl<Scalar, false, true>
 #ifdef EIGEN_MAKING_DOCS
     return run(Scalar(NumTraits<Scalar>::IsSigned ? -10 : 0), Scalar(10));
 #else
-    enum { rand_bits = floor_log2<(unsigned int)(RAND_MAX)+1>::value,
+    enum { rand_bits = meta_floor_log2<(unsigned int)(RAND_MAX)+1>::value,
            scalar_bits = sizeof(Scalar) * CHAR_BIT,
            shift = EIGEN_PLAIN_ENUM_MAX(0, int(rand_bits) - int(scalar_bits)),
            offset = NumTraits<Scalar>::IsSigned ? (1 << (EIGEN_PLAIN_ENUM_MIN(rand_bits,scalar_bits)-1)) : 0
@@ -556,97 +749,789 @@ inline EIGEN_MATHFUNC_RETVAL(random, Scalar) random()
   return EIGEN_MATHFUNC_IMPL(random, Scalar)::run();
 }
 
+// Implementatin of is* functions
+
+// std::is* do not work with fast-math and gcc, std::is* are available on MSVC 2013 and newer, as well as in clang.
+#if (EIGEN_HAS_CXX11_MATH && !(EIGEN_COMP_GNUC_STRICT && __FINITE_MATH_ONLY__)) || (EIGEN_COMP_MSVC>=1800) || (EIGEN_COMP_CLANG)
+#define EIGEN_USE_STD_FPCLASSIFY 1
+#else
+#define EIGEN_USE_STD_FPCLASSIFY 0
+#endif
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+typename internal::enable_if<internal::is_integral<T>::value,bool>::type
+isnan_impl(const T&) { return false; }
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+typename internal::enable_if<internal::is_integral<T>::value,bool>::type
+isinf_impl(const T&) { return false; }
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+typename internal::enable_if<internal::is_integral<T>::value,bool>::type
+isfinite_impl(const T&) { return true; }
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+typename internal::enable_if<(!internal::is_integral<T>::value)&&(!NumTraits<T>::IsComplex),bool>::type
+isfinite_impl(const T& x)
+{
+  #ifdef __CUDA_ARCH__
+    return (::isfinite)(x);
+  #elif EIGEN_USE_STD_FPCLASSIFY
+    using std::isfinite;
+    return isfinite EIGEN_NOT_A_MACRO (x);
+  #else
+    return x<=NumTraits<T>::highest() && x>=NumTraits<T>::lowest();
+  #endif
+}
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+typename internal::enable_if<(!internal::is_integral<T>::value)&&(!NumTraits<T>::IsComplex),bool>::type
+isinf_impl(const T& x)
+{
+  #ifdef __CUDA_ARCH__
+    return (::isinf)(x);
+  #elif EIGEN_USE_STD_FPCLASSIFY
+    using std::isinf;
+    return isinf EIGEN_NOT_A_MACRO (x);
+  #else
+    return x>NumTraits<T>::highest() || x<NumTraits<T>::lowest();
+  #endif
+}
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+typename internal::enable_if<(!internal::is_integral<T>::value)&&(!NumTraits<T>::IsComplex),bool>::type
+isnan_impl(const T& x)
+{
+  #ifdef __CUDA_ARCH__
+    return (::isnan)(x);
+  #elif EIGEN_USE_STD_FPCLASSIFY
+    using std::isnan;
+    return isnan EIGEN_NOT_A_MACRO (x);
+  #else
+    return x != x;
+  #endif
+}
+
+#if (!EIGEN_USE_STD_FPCLASSIFY)
+
+#if EIGEN_COMP_MSVC
+
+template<typename T> EIGEN_DEVICE_FUNC bool isinf_msvc_helper(T x)
+{
+  return _fpclass(x)==_FPCLASS_NINF || _fpclass(x)==_FPCLASS_PINF;
+}
+
+//MSVC defines a _isnan builtin function, but for double only
+EIGEN_DEVICE_FUNC inline bool isnan_impl(const long double& x) { return _isnan(x)!=0; }
+EIGEN_DEVICE_FUNC inline bool isnan_impl(const double& x)      { return _isnan(x)!=0; }
+EIGEN_DEVICE_FUNC inline bool isnan_impl(const float& x)       { return _isnan(x)!=0; }
+
+EIGEN_DEVICE_FUNC inline bool isinf_impl(const long double& x) { return isinf_msvc_helper(x); }
+EIGEN_DEVICE_FUNC inline bool isinf_impl(const double& x)      { return isinf_msvc_helper(x); }
+EIGEN_DEVICE_FUNC inline bool isinf_impl(const float& x)       { return isinf_msvc_helper(x); }
+
+#elif (defined __FINITE_MATH_ONLY__ && __FINITE_MATH_ONLY__ && EIGEN_COMP_GNUC)
+
+#if EIGEN_GNUC_AT_LEAST(5,0)
+  #define EIGEN_TMP_NOOPT_ATTRIB EIGEN_DEVICE_FUNC inline __attribute__((optimize("no-finite-math-only")))
+#else
+  // NOTE the inline qualifier and noinline attribute are both needed: the former is to avoid linking issue (duplicate symbol),
+  //      while the second prevent too aggressive optimizations in fast-math mode:
+  #define EIGEN_TMP_NOOPT_ATTRIB EIGEN_DEVICE_FUNC inline __attribute__((noinline,optimize("no-finite-math-only")))
+#endif
+
+template<> EIGEN_TMP_NOOPT_ATTRIB bool isnan_impl(const long double& x) { return __builtin_isnan(x); }
+template<> EIGEN_TMP_NOOPT_ATTRIB bool isnan_impl(const double& x)      { return __builtin_isnan(x); }
+template<> EIGEN_TMP_NOOPT_ATTRIB bool isnan_impl(const float& x)       { return __builtin_isnan(x); }
+template<> EIGEN_TMP_NOOPT_ATTRIB bool isinf_impl(const double& x)      { return __builtin_isinf(x); }
+template<> EIGEN_TMP_NOOPT_ATTRIB bool isinf_impl(const float& x)       { return __builtin_isinf(x); }
+template<> EIGEN_TMP_NOOPT_ATTRIB bool isinf_impl(const long double& x) { return __builtin_isinf(x); }
+
+#undef EIGEN_TMP_NOOPT_ATTRIB
+
+#endif
+
+#endif
+
+// The following overload are defined at the end of this file
+template<typename T> EIGEN_DEVICE_FUNC bool isfinite_impl(const std::complex<T>& x);
+template<typename T> EIGEN_DEVICE_FUNC bool isnan_impl(const std::complex<T>& x);
+template<typename T> EIGEN_DEVICE_FUNC bool isinf_impl(const std::complex<T>& x);
+
+template<typename T> T generic_fast_tanh_float(const T& a_x);
+
 } // end namespace internal
 
 /****************************************************************************
-* Generic math function                                                    *
+* Generic math functions                                                    *
 ****************************************************************************/
 
 namespace numext {
 
+#if !defined(__CUDA_ARCH__) && !defined(__SYCL_DEVICE_ONLY__)
+template<typename T>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE T mini(const T& x, const T& y)
+{
+  EIGEN_USING_STD_MATH(min);
+  return min EIGEN_NOT_A_MACRO (x,y);
+}
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE T maxi(const T& x, const T& y)
+{
+  EIGEN_USING_STD_MATH(max);
+  return max EIGEN_NOT_A_MACRO (x,y);
+}
+
+
+#elif defined(__SYCL_DEVICE_ONLY__)
+template<typename T>
+EIGEN_ALWAYS_INLINE T mini(const T& x, const T& y)
+{
+
+  return y < x ? y : x;
+}
+
+template<typename T>
+EIGEN_ALWAYS_INLINE T maxi(const T& x, const T& y)
+{
+
+  return x < y ? y : x;
+}
+
+EIGEN_ALWAYS_INLINE int mini(const int& x, const int& y)
+{
+  return cl::sycl::min(x,y);
+}
+
+EIGEN_ALWAYS_INLINE int maxi(const int& x, const int& y)
+{
+  return cl::sycl::max(x,y);
+}
+
+EIGEN_ALWAYS_INLINE unsigned int mini(const unsigned int& x, const unsigned int& y)
+{
+  return cl::sycl::min(x,y);
+}
+
+EIGEN_ALWAYS_INLINE unsigned int maxi(const unsigned int& x, const unsigned int& y)
+{
+  return cl::sycl::max(x,y);
+}
+
+EIGEN_ALWAYS_INLINE  long mini(const long & x, const long & y)
+{
+  return cl::sycl::min(x,y);
+}
+
+EIGEN_ALWAYS_INLINE  long maxi(const long & x, const long & y)
+{
+  return cl::sycl::max(x,y);
+}
+
+EIGEN_ALWAYS_INLINE unsigned long mini(const unsigned long& x, const unsigned long& y)
+{
+  return cl::sycl::min(x,y);
+}
+
+EIGEN_ALWAYS_INLINE unsigned long maxi(const unsigned long& x, const unsigned long& y)
+{
+  return cl::sycl::max(x,y);
+}
+
+
+EIGEN_ALWAYS_INLINE float mini(const float& x, const float& y)
+{
+  return cl::sycl::fmin(x,y);
+}
+
+EIGEN_ALWAYS_INLINE float maxi(const float& x, const float& y)
+{
+  return cl::sycl::fmax(x,y);
+}
+
+EIGEN_ALWAYS_INLINE double mini(const double& x, const double& y)
+{
+  return cl::sycl::fmin(x,y);
+}
+
+EIGEN_ALWAYS_INLINE double maxi(const double& x, const double& y)
+{
+  return cl::sycl::fmax(x,y);
+}
+
+#else
+template<typename T>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE T mini(const T& x, const T& y)
+{
+  return y < x ? y : x;
+}
+template<>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE float mini(const float& x, const float& y)
+{
+  return fminf(x, y);
+}
+template<typename T>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE T maxi(const T& x, const T& y)
+{
+  return x < y ? y : x;
+}
+template<>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE float maxi(const float& x, const float& y)
+{
+  return fmaxf(x, y);
+}
+#endif
+
+
 template<typename Scalar>
+EIGEN_DEVICE_FUNC
 inline EIGEN_MATHFUNC_RETVAL(real, Scalar) real(const Scalar& x)
 {
   return EIGEN_MATHFUNC_IMPL(real, Scalar)::run(x);
-}  
+}
 
 template<typename Scalar>
+EIGEN_DEVICE_FUNC
 inline typename internal::add_const_on_value_type< EIGEN_MATHFUNC_RETVAL(real_ref, Scalar) >::type real_ref(const Scalar& x)
 {
   return internal::real_ref_impl<Scalar>::run(x);
 }
 
 template<typename Scalar>
+EIGEN_DEVICE_FUNC
 inline EIGEN_MATHFUNC_RETVAL(real_ref, Scalar) real_ref(Scalar& x)
 {
   return EIGEN_MATHFUNC_IMPL(real_ref, Scalar)::run(x);
 }
 
 template<typename Scalar>
+EIGEN_DEVICE_FUNC
 inline EIGEN_MATHFUNC_RETVAL(imag, Scalar) imag(const Scalar& x)
 {
   return EIGEN_MATHFUNC_IMPL(imag, Scalar)::run(x);
 }
 
 template<typename Scalar>
+EIGEN_DEVICE_FUNC
+inline EIGEN_MATHFUNC_RETVAL(arg, Scalar) arg(const Scalar& x)
+{
+  return EIGEN_MATHFUNC_IMPL(arg, Scalar)::run(x);
+}
+
+template<typename Scalar>
+EIGEN_DEVICE_FUNC
 inline typename internal::add_const_on_value_type< EIGEN_MATHFUNC_RETVAL(imag_ref, Scalar) >::type imag_ref(const Scalar& x)
 {
   return internal::imag_ref_impl<Scalar>::run(x);
 }
 
 template<typename Scalar>
+EIGEN_DEVICE_FUNC
 inline EIGEN_MATHFUNC_RETVAL(imag_ref, Scalar) imag_ref(Scalar& x)
 {
   return EIGEN_MATHFUNC_IMPL(imag_ref, Scalar)::run(x);
 }
 
 template<typename Scalar>
+EIGEN_DEVICE_FUNC
 inline EIGEN_MATHFUNC_RETVAL(conj, Scalar) conj(const Scalar& x)
 {
   return EIGEN_MATHFUNC_IMPL(conj, Scalar)::run(x);
 }
 
 template<typename Scalar>
+EIGEN_DEVICE_FUNC
 inline EIGEN_MATHFUNC_RETVAL(abs2, Scalar) abs2(const Scalar& x)
 {
   return EIGEN_MATHFUNC_IMPL(abs2, Scalar)::run(x);
 }
 
 template<typename Scalar>
+EIGEN_DEVICE_FUNC
 inline EIGEN_MATHFUNC_RETVAL(norm1, Scalar) norm1(const Scalar& x)
 {
   return EIGEN_MATHFUNC_IMPL(norm1, Scalar)::run(x);
 }
 
 template<typename Scalar>
+EIGEN_DEVICE_FUNC
 inline EIGEN_MATHFUNC_RETVAL(hypot, Scalar) hypot(const Scalar& x, const Scalar& y)
 {
   return EIGEN_MATHFUNC_IMPL(hypot, Scalar)::run(x, y);
 }
 
 template<typename Scalar>
-inline EIGEN_MATHFUNC_RETVAL(atanh2, Scalar) atanh2(const Scalar& x, const Scalar& y)
+EIGEN_DEVICE_FUNC
+inline EIGEN_MATHFUNC_RETVAL(log1p, Scalar) log1p(const Scalar& x)
+{
+  return EIGEN_MATHFUNC_IMPL(log1p, Scalar)::run(x);
+}
+
+#if defined(__SYCL_DEVICE_ONLY__)
+EIGEN_ALWAYS_INLINE float   log1p(float x) { return cl::sycl::log1p(x); }
+EIGEN_ALWAYS_INLINE double  log1p(double x) { return cl::sycl::log1p(x); }
+#endif // defined(__SYCL_DEVICE_ONLY__)
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float log1p(const float &x) { return ::log1pf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double log1p(const double &x) { return ::log1p(x); }
+#endif
+
+template<typename ScalarX,typename ScalarY>
+EIGEN_DEVICE_FUNC
+inline typename internal::pow_impl<ScalarX,ScalarY>::result_type pow(const ScalarX& x, const ScalarY& y)
 {
-  return EIGEN_MATHFUNC_IMPL(atanh2, Scalar)::run(x, y);
+  return internal::pow_impl<ScalarX,ScalarY>::run(x, y);
 }
 
+#if defined(__SYCL_DEVICE_ONLY__)
+EIGEN_ALWAYS_INLINE float   pow(float x, float y) { return cl::sycl::pow(x, y); }
+EIGEN_ALWAYS_INLINE double  pow(double x, double y) { return cl::sycl::pow(x, y); }
+#endif // defined(__SYCL_DEVICE_ONLY__)
+
+template<typename T> EIGEN_DEVICE_FUNC bool (isnan)   (const T &x) { return internal::isnan_impl(x); }
+template<typename T> EIGEN_DEVICE_FUNC bool (isinf)   (const T &x) { return internal::isinf_impl(x); }
+template<typename T> EIGEN_DEVICE_FUNC bool (isfinite)(const T &x) { return internal::isfinite_impl(x); }
+
+#if defined(__SYCL_DEVICE_ONLY__)
+EIGEN_ALWAYS_INLINE float   isnan(float x) { return cl::sycl::isnan(x); }
+EIGEN_ALWAYS_INLINE double  isnan(double x) { return cl::sycl::isnan(x); }
+EIGEN_ALWAYS_INLINE float   isinf(float x) { return cl::sycl::isinf(x); }
+EIGEN_ALWAYS_INLINE double  isinf(double x) { return cl::sycl::isinf(x); }
+EIGEN_ALWAYS_INLINE float   isfinite(float x) { return cl::sycl::isfinite(x); }
+EIGEN_ALWAYS_INLINE double  isfinite(double x) { return cl::sycl::isfinite(x); }
+#endif // defined(__SYCL_DEVICE_ONLY__)
+
 template<typename Scalar>
-inline EIGEN_MATHFUNC_RETVAL(pow, Scalar) pow(const Scalar& x, const Scalar& y)
+EIGEN_DEVICE_FUNC
+inline EIGEN_MATHFUNC_RETVAL(round, Scalar) round(const Scalar& x)
+{
+  return EIGEN_MATHFUNC_IMPL(round, Scalar)::run(x);
+}
+
+#if defined(__SYCL_DEVICE_ONLY__)
+EIGEN_ALWAYS_INLINE float   round(float x) { return cl::sycl::round(x); }
+EIGEN_ALWAYS_INLINE double  round(double x) { return cl::sycl::round(x); }
+#endif // defined(__SYCL_DEVICE_ONLY__)
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+T (floor)(const T& x)
 {
-  return EIGEN_MATHFUNC_IMPL(pow, Scalar)::run(x, y);
+  EIGEN_USING_STD_MATH(floor);
+  return floor(x);
 }
 
-// std::isfinite is non standard, so let's define our own version,
-// even though it is not very efficient.
-template<typename T> bool (isfinite)(const T& x)
+#if defined(__SYCL_DEVICE_ONLY__)
+EIGEN_ALWAYS_INLINE float   floor(float x) { return cl::sycl::floor(x); }
+EIGEN_ALWAYS_INLINE double  floor(double x) { return cl::sycl::floor(x); }
+#endif // defined(__SYCL_DEVICE_ONLY__)
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float floor(const float &x) { return ::floorf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double floor(const double &x) { return ::floor(x); }
+#endif
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+T (ceil)(const T& x)
 {
-  return x<NumTraits<T>::highest() && x>NumTraits<T>::lowest();
+  EIGEN_USING_STD_MATH(ceil);
+  return ceil(x);
 }
 
+#if defined(__SYCL_DEVICE_ONLY__)
+EIGEN_ALWAYS_INLINE float   ceil(float x) { return cl::sycl::ceil(x); }
+EIGEN_ALWAYS_INLINE double  ceil(double x) { return cl::sycl::ceil(x); }
+#endif // defined(__SYCL_DEVICE_ONLY__)
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float ceil(const float &x) { return ::ceilf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double ceil(const double &x) { return ::ceil(x); }
+#endif
+
+
+/** Log base 2 for 32 bits positive integers.
+  * Conveniently returns 0 for x==0. */
+inline int log2(int x)
+{
+  eigen_assert(x>=0);
+  unsigned int v(x);
+  static const int table[32] = { 0, 9, 1, 10, 13, 21, 2, 29, 11, 14, 16, 18, 22, 25, 3, 30, 8, 12, 20, 28, 15, 17, 24, 7, 19, 27, 23, 6, 26, 5, 4, 31 };
+  v |= v >> 1;
+  v |= v >> 2;
+  v |= v >> 4;
+  v |= v >> 8;
+  v |= v >> 16;
+  return table[(v * 0x07C4ACDDU) >> 27];
+}
+
+/** \returns the square root of \a x.
+  *
+  * It is essentially equivalent to \code using std::sqrt; return sqrt(x); \endcode,
+  * but slightly faster for float/double and some compilers (e.g., gcc), thanks to
+  * specializations when SSE is enabled.
+  *
+  * It's usage is justified in performance critical functions, like norm/normalize.
+  */
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T sqrt(const T &x)
+{
+  EIGEN_USING_STD_MATH(sqrt);
+  return sqrt(x);
+}
+
+#if defined(__SYCL_DEVICE_ONLY__)
+EIGEN_ALWAYS_INLINE float   sqrt(float x) { return cl::sycl::sqrt(x); }
+EIGEN_ALWAYS_INLINE double  sqrt(double x) { return cl::sycl::sqrt(x); }
+#endif // defined(__SYCL_DEVICE_ONLY__)
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T log(const T &x) {
+  EIGEN_USING_STD_MATH(log);
+  return log(x);
+}
+
+#if defined(__SYCL_DEVICE_ONLY__)
+EIGEN_ALWAYS_INLINE float   log(float x) { return cl::sycl::log(x); }
+EIGEN_ALWAYS_INLINE double  log(double x) { return cl::sycl::log(x); }
+#endif // defined(__SYCL_DEVICE_ONLY__)
+
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float log(const float &x) { return ::logf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double log(const double &x) { return ::log(x); }
+#endif
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+typename NumTraits<T>::Real abs(const T &x) {
+  EIGEN_USING_STD_MATH(abs);
+  return abs(x);
+}
+
+#if defined(__SYCL_DEVICE_ONLY__)
+EIGEN_ALWAYS_INLINE float   abs(float x) { return cl::sycl::fabs(x); }
+EIGEN_ALWAYS_INLINE double  abs(double x) { return cl::sycl::fabs(x); }
+#endif // defined(__SYCL_DEVICE_ONLY__)
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float abs(const float &x) { return ::fabsf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double abs(const double &x) { return ::fabs(x); }
+
+template <> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float abs(const std::complex<float>& x) {
+  return ::hypotf(x.real(), x.imag());
+}
+
+template <> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double abs(const std::complex<double>& x) {
+  return ::hypot(x.real(), x.imag());
+}
+#endif
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T exp(const T &x) {
+  EIGEN_USING_STD_MATH(exp);
+  return exp(x);
+}
+
+#if defined(__SYCL_DEVICE_ONLY__)
+EIGEN_ALWAYS_INLINE float   exp(float x) { return cl::sycl::exp(x); }
+EIGEN_ALWAYS_INLINE double  exp(double x) { return cl::sycl::exp(x); }
+#endif // defined(__SYCL_DEVICE_ONLY__)
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float exp(const float &x) { return ::expf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double exp(const double &x) { return ::exp(x); }
+#endif
+
+template<typename Scalar>
+EIGEN_DEVICE_FUNC
+inline EIGEN_MATHFUNC_RETVAL(expm1, Scalar) expm1(const Scalar& x)
+{
+  return EIGEN_MATHFUNC_IMPL(expm1, Scalar)::run(x);
+}
+
+#if defined(__SYCL_DEVICE_ONLY__)
+EIGEN_ALWAYS_INLINE float   expm1(float x) { return cl::sycl::expm1(x); }
+EIGEN_ALWAYS_INLINE double  expm1(double x) { return cl::sycl::expm1(x); }
+#endif // defined(__SYCL_DEVICE_ONLY__)
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float expm1(const float &x) { return ::expm1f(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double expm1(const double &x) { return ::expm1(x); }
+#endif
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T cos(const T &x) {
+  EIGEN_USING_STD_MATH(cos);
+  return cos(x);
+}
+
+#if defined(__SYCL_DEVICE_ONLY__)
+EIGEN_ALWAYS_INLINE float   cos(float x) { return cl::sycl::cos(x); }
+EIGEN_ALWAYS_INLINE double  cos(double x) { return cl::sycl::cos(x); }
+#endif // defined(__SYCL_DEVICE_ONLY__)
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float cos(const float &x) { return ::cosf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double cos(const double &x) { return ::cos(x); }
+#endif
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T sin(const T &x) {
+  EIGEN_USING_STD_MATH(sin);
+  return sin(x);
+}
+
+#if defined(__SYCL_DEVICE_ONLY__)
+EIGEN_ALWAYS_INLINE float   sin(float x) { return cl::sycl::sin(x); }
+EIGEN_ALWAYS_INLINE double  sin(double x) { return cl::sycl::sin(x); }
+#endif // defined(__SYCL_DEVICE_ONLY__)
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float sin(const float &x) { return ::sinf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double sin(const double &x) { return ::sin(x); }
+#endif
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T tan(const T &x) {
+  EIGEN_USING_STD_MATH(tan);
+  return tan(x);
+}
+
+#if defined(__SYCL_DEVICE_ONLY__)
+EIGEN_ALWAYS_INLINE float   tan(float x) { return cl::sycl::tan(x); }
+EIGEN_ALWAYS_INLINE double  tan(double x) { return cl::sycl::tan(x); }
+#endif // defined(__SYCL_DEVICE_ONLY__)
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float tan(const float &x) { return ::tanf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double tan(const double &x) { return ::tan(x); }
+#endif
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T acos(const T &x) {
+  EIGEN_USING_STD_MATH(acos);
+  return acos(x);
+}
+
+#if defined(__SYCL_DEVICE_ONLY__)
+EIGEN_ALWAYS_INLINE float   acos(float x) { return cl::sycl::acos(x); }
+EIGEN_ALWAYS_INLINE double  acos(double x) { return cl::sycl::acos(x); }
+#endif // defined(__SYCL_DEVICE_ONLY__)
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float acos(const float &x) { return ::acosf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double acos(const double &x) { return ::acos(x); }
+#endif
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T asin(const T &x) {
+  EIGEN_USING_STD_MATH(asin);
+  return asin(x);
+}
+
+#if defined(__SYCL_DEVICE_ONLY__)
+EIGEN_ALWAYS_INLINE float   asin(float x) { return cl::sycl::asin(x); }
+EIGEN_ALWAYS_INLINE double  asin(double x) { return cl::sycl::asin(x); }
+#endif // defined(__SYCL_DEVICE_ONLY__)
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float asin(const float &x) { return ::asinf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double asin(const double &x) { return ::asin(x); }
+#endif
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T atan(const T &x) {
+  EIGEN_USING_STD_MATH(atan);
+  return atan(x);
+}
+
+#if defined(__SYCL_DEVICE_ONLY__)
+EIGEN_ALWAYS_INLINE float   atan(float x) { return cl::sycl::atan(x); }
+EIGEN_ALWAYS_INLINE double  atan(double x) { return cl::sycl::atan(x); }
+#endif // defined(__SYCL_DEVICE_ONLY__)
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float atan(const float &x) { return ::atanf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double atan(const double &x) { return ::atan(x); }
+#endif
+
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T cosh(const T &x) {
+  EIGEN_USING_STD_MATH(cosh);
+  return cosh(x);
+}
+
+#if defined(__SYCL_DEVICE_ONLY__)
+EIGEN_ALWAYS_INLINE float   cosh(float x) { return cl::sycl::cosh(x); }
+EIGEN_ALWAYS_INLINE double  cosh(double x) { return cl::sycl::cosh(x); }
+#endif // defined(__SYCL_DEVICE_ONLY__)
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float cosh(const float &x) { return ::coshf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double cosh(const double &x) { return ::cosh(x); }
+#endif
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T sinh(const T &x) {
+  EIGEN_USING_STD_MATH(sinh);
+  return sinh(x);
+}
+
+#if defined(__SYCL_DEVICE_ONLY__)
+EIGEN_ALWAYS_INLINE float   sinh(float x) { return cl::sycl::sinh(x); }
+EIGEN_ALWAYS_INLINE double  sinh(double x) { return cl::sycl::sinh(x); }
+#endif // defined(__SYCL_DEVICE_ONLY__)
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float sinh(const float &x) { return ::sinhf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double sinh(const double &x) { return ::sinh(x); }
+#endif
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T tanh(const T &x) {
+  EIGEN_USING_STD_MATH(tanh);
+  return tanh(x);
+}
+
+#if defined(__SYCL_DEVICE_ONLY__)
+EIGEN_ALWAYS_INLINE float   tanh(float x) { return cl::sycl::tanh(x); }
+EIGEN_ALWAYS_INLINE double  tanh(double x) { return cl::sycl::tanh(x); }
+#elif (!defined(__CUDACC__)) && EIGEN_FAST_MATH
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float tanh(float x) { return internal::generic_fast_tanh_float(x); }
+#endif
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float tanh(const float &x) { return ::tanhf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double tanh(const double &x) { return ::tanh(x); }
+#endif
+
+template <typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T fmod(const T& a, const T& b) {
+  EIGEN_USING_STD_MATH(fmod);
+  return fmod(a, b);
+}
+
+#if defined(__SYCL_DEVICE_ONLY__)
+EIGEN_ALWAYS_INLINE float   fmod(float x, float y) { return cl::sycl::fmod(x, y); }
+EIGEN_ALWAYS_INLINE double  fmod(double x, double y) { return cl::sycl::fmod(x, y); }
+#endif // defined(__SYCL_DEVICE_ONLY__)
+
+#ifdef __CUDACC__
+template <>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float fmod(const float& a, const float& b) {
+  return ::fmodf(a, b);
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double fmod(const double& a, const double& b) {
+  return ::fmod(a, b);
+}
+#endif
+
 } // end namespace numext
 
 namespace internal {
 
+template<typename T>
+EIGEN_DEVICE_FUNC bool isfinite_impl(const std::complex<T>& x)
+{
+  return (numext::isfinite)(numext::real(x)) && (numext::isfinite)(numext::imag(x));
+}
+
+template<typename T>
+EIGEN_DEVICE_FUNC bool isnan_impl(const std::complex<T>& x)
+{
+  return (numext::isnan)(numext::real(x)) || (numext::isnan)(numext::imag(x));
+}
+
+template<typename T>
+EIGEN_DEVICE_FUNC bool isinf_impl(const std::complex<T>& x)
+{
+  return ((numext::isinf)(numext::real(x)) || (numext::isinf)(numext::imag(x))) && (!(numext::isnan)(x));
+}
+
 /****************************************************************************
 * Implementation of fuzzy comparisons                                       *
 ****************************************************************************/
@@ -660,18 +1545,17 @@ template<typename Scalar>
 struct scalar_fuzzy_default_impl<Scalar, false, false>
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
-  template<typename OtherScalar>
+  template<typename OtherScalar> EIGEN_DEVICE_FUNC
   static inline bool isMuchSmallerThan(const Scalar& x, const OtherScalar& y, const RealScalar& prec)
   {
-    using std::abs;
-    return abs(x) <= abs(y) * prec;
+    return numext::abs(x) <= numext::abs(y) * prec;
   }
+  EIGEN_DEVICE_FUNC
   static inline bool isApprox(const Scalar& x, const Scalar& y, const RealScalar& prec)
   {
-    using std::min;
-    using std::abs;
-    return abs(x - y) <= (min)(abs(x), abs(y)) * prec;
+    return numext::abs(x - y) <= numext::mini(numext::abs(x), numext::abs(y)) * prec;
   }
+  EIGEN_DEVICE_FUNC
   static inline bool isApproxOrLessThan(const Scalar& x, const Scalar& y, const RealScalar& prec)
   {
     return x <= y || isApprox(x, y, prec);
@@ -682,15 +1566,17 @@ template<typename Scalar>
 struct scalar_fuzzy_default_impl<Scalar, false, true>
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
-  template<typename OtherScalar>
+  template<typename OtherScalar> EIGEN_DEVICE_FUNC
   static inline bool isMuchSmallerThan(const Scalar& x, const Scalar&, const RealScalar&)
   {
     return x == Scalar(0);
   }
+  EIGEN_DEVICE_FUNC
   static inline bool isApprox(const Scalar& x, const Scalar& y, const RealScalar&)
   {
     return x == y;
   }
+  EIGEN_DEVICE_FUNC
   static inline bool isApproxOrLessThan(const Scalar& x, const Scalar& y, const RealScalar&)
   {
     return x <= y;
@@ -701,36 +1587,36 @@ template<typename Scalar>
 struct scalar_fuzzy_default_impl<Scalar, true, false>
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
-  template<typename OtherScalar>
+  template<typename OtherScalar> EIGEN_DEVICE_FUNC
   static inline bool isMuchSmallerThan(const Scalar& x, const OtherScalar& y, const RealScalar& prec)
   {
     return numext::abs2(x) <= numext::abs2(y) * prec * prec;
   }
+  EIGEN_DEVICE_FUNC
   static inline bool isApprox(const Scalar& x, const Scalar& y, const RealScalar& prec)
   {
-    using std::min;
-    return numext::abs2(x - y) <= (min)(numext::abs2(x), numext::abs2(y)) * prec * prec;
+    return numext::abs2(x - y) <= numext::mini(numext::abs2(x), numext::abs2(y)) * prec * prec;
   }
 };
 
 template<typename Scalar>
 struct scalar_fuzzy_impl : scalar_fuzzy_default_impl<Scalar, NumTraits<Scalar>::IsComplex, NumTraits<Scalar>::IsInteger> {};
 
-template<typename Scalar, typename OtherScalar>
+template<typename Scalar, typename OtherScalar> EIGEN_DEVICE_FUNC
 inline bool isMuchSmallerThan(const Scalar& x, const OtherScalar& y,
                               const typename NumTraits<Scalar>::Real &precision = NumTraits<Scalar>::dummy_precision())
 {
   return scalar_fuzzy_impl<Scalar>::template isMuchSmallerThan<OtherScalar>(x, y, precision);
 }
 
-template<typename Scalar>
+template<typename Scalar> EIGEN_DEVICE_FUNC
 inline bool isApprox(const Scalar& x, const Scalar& y,
                      const typename NumTraits<Scalar>::Real &precision = NumTraits<Scalar>::dummy_precision())
 {
   return scalar_fuzzy_impl<Scalar>::isApprox(x, y, precision);
 }
 
-template<typename Scalar>
+template<typename Scalar> EIGEN_DEVICE_FUNC
 inline bool isApproxOrLessThan(const Scalar& x, const Scalar& y,
                                const typename NumTraits<Scalar>::Real &precision = NumTraits<Scalar>::dummy_precision())
 {
@@ -752,26 +1638,28 @@ template<> struct random_impl<bool>
 template<> struct scalar_fuzzy_impl<bool>
 {
   typedef bool RealScalar;
-  
-  template<typename OtherScalar>
+
+  template<typename OtherScalar> EIGEN_DEVICE_FUNC
   static inline bool isMuchSmallerThan(const bool& x, const bool&, const bool&)
   {
     return !x;
   }
-  
+
+  EIGEN_DEVICE_FUNC
   static inline bool isApprox(bool x, bool y, bool)
   {
     return x == y;
   }
 
+  EIGEN_DEVICE_FUNC
   static inline bool isApproxOrLessThan(const bool& x, const bool& y, const bool&)
   {
     return (!x) || y;
   }
-  
+
 };
 
-  
+
 } // end namespace internal
 
 } // end namespace Eigen
diff --git a/eigen/Eigen/src/Core/MathFunctionsImpl.h b/eigen/Eigen/src/Core/MathFunctionsImpl.h
new file mode 100644
index 0000000..ae1386b
--- /dev/null
+++ b/eigen/Eigen/src/Core/MathFunctionsImpl.h
@@ -0,0 +1,73 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Pedro Gonnet (pedro.gonnet@gmail.com)
+// Copyright (C) 2016 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MATHFUNCTIONSIMPL_H
+#define EIGEN_MATHFUNCTIONSIMPL_H
+
+namespace Eigen {
+
+namespace internal {
+
+/** \internal \returns the hyperbolic tan of \a a (coeff-wise)
+    Doesn't do anything fancy, just a 13/6-degree rational interpolant which
+    is accurate up to a couple of ulp in the range [-9, 9], outside of which
+    the tanh(x) = +/-1.
+
+    This implementation works on both scalars and packets.
+*/
+template<typename T>
+T generic_fast_tanh_float(const T& a_x)
+{
+  // Clamp the inputs to the range [-9, 9] since anything outside
+  // this range is +/-1.0f in single-precision.
+  const T plus_9 = pset1<T>(9.f);
+  const T minus_9 = pset1<T>(-9.f);
+  const T x = pmax(pmin(a_x, plus_9), minus_9);
+  // The monomial coefficients of the numerator polynomial (odd).
+  const T alpha_1 = pset1<T>(4.89352455891786e-03f);
+  const T alpha_3 = pset1<T>(6.37261928875436e-04f);
+  const T alpha_5 = pset1<T>(1.48572235717979e-05f);
+  const T alpha_7 = pset1<T>(5.12229709037114e-08f);
+  const T alpha_9 = pset1<T>(-8.60467152213735e-11f);
+  const T alpha_11 = pset1<T>(2.00018790482477e-13f);
+  const T alpha_13 = pset1<T>(-2.76076847742355e-16f);
+
+  // The monomial coefficients of the denominator polynomial (even).
+  const T beta_0 = pset1<T>(4.89352518554385e-03f);
+  const T beta_2 = pset1<T>(2.26843463243900e-03f);
+  const T beta_4 = pset1<T>(1.18534705686654e-04f);
+  const T beta_6 = pset1<T>(1.19825839466702e-06f);
+
+  // Since the polynomials are odd/even, we need x^2.
+  const T x2 = pmul(x, x);
+
+  // Evaluate the numerator polynomial p.
+  T p = pmadd(x2, alpha_13, alpha_11);
+  p = pmadd(x2, p, alpha_9);
+  p = pmadd(x2, p, alpha_7);
+  p = pmadd(x2, p, alpha_5);
+  p = pmadd(x2, p, alpha_3);
+  p = pmadd(x2, p, alpha_1);
+  p = pmul(x, p);
+
+  // Evaluate the denominator polynomial p.
+  T q = pmadd(x2, beta_6, beta_4);
+  q = pmadd(x2, q, beta_2);
+  q = pmadd(x2, q, beta_0);
+
+  // Divide the numerator by the denominator.
+  return pdiv(p, q);
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_MATHFUNCTIONSIMPL_H
diff --git a/eigen/Eigen/src/Core/Matrix.h b/eigen/Eigen/src/Core/Matrix.h
index 02be142..90c336d 100644
--- a/eigen/Eigen/src/Core/Matrix.h
+++ b/eigen/Eigen/src/Core/Matrix.h
@@ -13,6 +13,45 @@
 
 namespace Eigen {
 
+namespace internal {
+template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
+struct traits<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >
+{
+private:
+  enum { size = internal::size_at_compile_time<_Rows,_Cols>::ret };
+  typedef typename find_best_packet<_Scalar,size>::type PacketScalar;
+  enum {
+      row_major_bit = _Options&RowMajor ? RowMajorBit : 0,
+      is_dynamic_size_storage = _MaxRows==Dynamic || _MaxCols==Dynamic,
+      max_size = is_dynamic_size_storage ? Dynamic : _MaxRows*_MaxCols,
+      default_alignment = compute_default_alignment<_Scalar,max_size>::value,
+      actual_alignment = ((_Options&DontAlign)==0) ? default_alignment : 0,
+      required_alignment = unpacket_traits<PacketScalar>::alignment,
+      packet_access_bit = (packet_traits<_Scalar>::Vectorizable && (EIGEN_UNALIGNED_VECTORIZE || (actual_alignment>=required_alignment))) ? PacketAccessBit : 0
+    };
+    
+public:
+  typedef _Scalar Scalar;
+  typedef Dense StorageKind;
+  typedef Eigen::Index StorageIndex;
+  typedef MatrixXpr XprKind;
+  enum {
+    RowsAtCompileTime = _Rows,
+    ColsAtCompileTime = _Cols,
+    MaxRowsAtCompileTime = _MaxRows,
+    MaxColsAtCompileTime = _MaxCols,
+    Flags = compute_matrix_flags<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>::ret,
+    Options = _Options,
+    InnerStrideAtCompileTime = 1,
+    OuterStrideAtCompileTime = (Options&RowMajor) ? ColsAtCompileTime : RowsAtCompileTime,
+    
+    // FIXME, the following flag in only used to define NeedsToAlign in PlainObjectBase
+    EvaluatorFlags = LinearAccessBit | DirectAccessBit | packet_access_bit | row_major_bit,
+    Alignment = actual_alignment
+  };
+};
+}
+
 /** \class Matrix
   * \ingroup Core_Module
   *
@@ -24,13 +63,13 @@ namespace Eigen {
   * The %Matrix class encompasses \em both fixed-size and dynamic-size objects (\ref fixedsize "note").
   *
   * The first three template parameters are required:
-  * \tparam _Scalar \anchor matrix_tparam_scalar Numeric type, e.g. float, double, int or std::complex<float>.
-  *                 User defined sclar types are supported as well (see \ref user_defined_scalars "here").
+  * \tparam _Scalar Numeric type, e.g. float, double, int or std::complex<float>.
+  *                 User defined scalar types are supported as well (see \ref user_defined_scalars "here").
   * \tparam _Rows Number of rows, or \b Dynamic
   * \tparam _Cols Number of columns, or \b Dynamic
   *
   * The remaining template parameters are optional -- in most cases you don't have to worry about them.
-  * \tparam _Options \anchor matrix_tparam_options A combination of either \b #RowMajor or \b #ColMajor, and of either
+  * \tparam _Options A combination of either \b #RowMajor or \b #ColMajor, and of either
   *                 \b #AutoAlign or \b #DontAlign.
   *                 The former controls \ref TopicStorageOrders "storage order", and defaults to column-major. The latter controls alignment, which is required
   *                 for vectorization. It defaults to aligning matrices except for fixed sizes that aren't a multiple of the packet size.
@@ -67,7 +106,7 @@ namespace Eigen {
   * \endcode
   *
   * This class can be extended with the help of the plugin mechanism described on the page
-  * \ref TopicCustomizingEigen by defining the preprocessor symbol \c EIGEN_MATRIX_PLUGIN.
+  * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_MATRIX_PLUGIN.
   *
   * <i><b>Some notes:</b></i>
   *
@@ -97,32 +136,44 @@ namespace Eigen {
   * are the dimensions of the original matrix, while _Rows and _Cols are Dynamic.</dd>
   * </dl>
   *
-  * \see MatrixBase for the majority of the API methods for matrices, \ref TopicClassHierarchy, 
-  * \ref TopicStorageOrders 
+  * <i><b>ABI and storage layout</b></i>
+  *
+  * The table below summarizes the ABI of some possible Matrix instances which is fixed thorough the lifetime of Eigen 3.
+  * <table  class="manual">
+  * <tr><th>Matrix type</th><th>Equivalent C structure</th></tr>
+  * <tr><td>\code Matrix<T,Dynamic,Dynamic> \endcode</td><td>\code
+  * struct {
+  *   T *data;                  // with (size_t(data)%EIGEN_MAX_ALIGN_BYTES)==0
+  *   Eigen::Index rows, cols;
+  *  };
+  * \endcode</td></tr>
+  * <tr class="alt"><td>\code
+  * Matrix<T,Dynamic,1>
+  * Matrix<T,1,Dynamic> \endcode</td><td>\code
+  * struct {
+  *   T *data;                  // with (size_t(data)%EIGEN_MAX_ALIGN_BYTES)==0
+  *   Eigen::Index size;
+  *  };
+  * \endcode</td></tr>
+  * <tr><td>\code Matrix<T,Rows,Cols> \endcode</td><td>\code
+  * struct {
+  *   T data[Rows*Cols];        // with (size_t(data)%A(Rows*Cols*sizeof(T)))==0
+  *  };
+  * \endcode</td></tr>
+  * <tr class="alt"><td>\code Matrix<T,Dynamic,Dynamic,0,MaxRows,MaxCols> \endcode</td><td>\code
+  * struct {
+  *   T data[MaxRows*MaxCols];  // with (size_t(data)%A(MaxRows*MaxCols*sizeof(T)))==0
+  *   Eigen::Index rows, cols;
+  *  };
+  * \endcode</td></tr>
+  * </table>
+  * Note that in this table Rows, Cols, MaxRows and MaxCols are all positive integers. A(S) is defined to the largest possible power-of-two
+  * smaller to EIGEN_MAX_STATIC_ALIGN_BYTES.
+  *
+  * \see MatrixBase for the majority of the API methods for matrices, \ref TopicClassHierarchy,
+  * \ref TopicStorageOrders
   */
 
-namespace internal {
-template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
-struct traits<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >
-{
-  typedef _Scalar Scalar;
-  typedef Dense StorageKind;
-  typedef DenseIndex Index;
-  typedef MatrixXpr XprKind;
-  enum {
-    RowsAtCompileTime = _Rows,
-    ColsAtCompileTime = _Cols,
-    MaxRowsAtCompileTime = _MaxRows,
-    MaxColsAtCompileTime = _MaxCols,
-    Flags = compute_matrix_flags<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>::ret,
-    CoeffReadCost = NumTraits<Scalar>::ReadCost,
-    Options = _Options,
-    InnerStrideAtCompileTime = 1,
-    OuterStrideAtCompileTime = (Options&RowMajor) ? ColsAtCompileTime : RowsAtCompileTime
-  };
-};
-}
-
 template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
 class Matrix
   : public PlainObjectBase<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >
@@ -151,6 +202,7 @@ class Matrix
       *
       * \callgraph
       */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Matrix& operator=(const Matrix& other)
     {
       return Base::_set(other);
@@ -167,7 +219,8 @@ class Matrix
       * remain row-vectors and vectors remain vectors.
       */
     template<typename OtherDerived>
-    EIGEN_STRONG_INLINE Matrix& operator=(const MatrixBase<OtherDerived>& other)
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Matrix& operator=(const DenseBase<OtherDerived>& other)
     {
       return Base::_set(other);
     }
@@ -179,12 +232,14 @@ class Matrix
       * \copydetails DenseBase::operator=(const EigenBase<OtherDerived> &other)
       */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Matrix& operator=(const EigenBase<OtherDerived> &other)
     {
       return Base::operator=(other);
     }
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Matrix& operator=(const ReturnByValue<OtherDerived>& func)
     {
       return Base::operator=(func);
@@ -200,6 +255,7 @@ class Matrix
       *
       * \sa resize(Index,Index)
       */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Matrix() : Base()
     {
       Base::_check_template_params();
@@ -207,60 +263,87 @@ class Matrix
     }
 
     // FIXME is it still needed
-    Matrix(internal::constructor_without_unaligned_array_assert)
+    EIGEN_DEVICE_FUNC
+    explicit Matrix(internal::constructor_without_unaligned_array_assert)
       : Base(internal::constructor_without_unaligned_array_assert())
     { Base::_check_template_params(); EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED }
 
-#ifdef EIGEN_HAVE_RVALUE_REFERENCES
-    Matrix(Matrix&& other)
+#if EIGEN_HAS_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC
+    Matrix(Matrix&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_constructible<Scalar>::value)
       : Base(std::move(other))
     {
       Base::_check_template_params();
       if (RowsAtCompileTime!=Dynamic && ColsAtCompileTime!=Dynamic)
         Base::_set_noalias(other);
     }
-    Matrix& operator=(Matrix&& other)
+    EIGEN_DEVICE_FUNC
+    Matrix& operator=(Matrix&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_assignable<Scalar>::value)
     {
       other.swap(*this);
       return *this;
     }
 #endif
 
-    /** \brief Constructs a vector or row-vector with given dimension. \only_for_vectors
-      *
-      * Note that this is only useful for dynamic-size vectors. For fixed-size vectors,
-      * it is redundant to pass the dimension here, so it makes more sense to use the default
-      * constructor Matrix() instead.
-      */
-    EIGEN_STRONG_INLINE explicit Matrix(Index dim)
-      : Base(dim, RowsAtCompileTime == 1 ? 1 : dim, ColsAtCompileTime == 1 ? 1 : dim)
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+
+    // This constructor is for both 1x1 matrices and dynamic vectors
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE explicit Matrix(const T& x)
     {
       Base::_check_template_params();
-      EIGEN_STATIC_ASSERT_VECTOR_ONLY(Matrix)
-      eigen_assert(dim >= 0);
-      eigen_assert(SizeAtCompileTime == Dynamic || SizeAtCompileTime == dim);
-      EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
+      Base::template _init1<T>(x);
     }
 
-    #ifndef EIGEN_PARSED_BY_DOXYGEN
     template<typename T0, typename T1>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Matrix(const T0& x, const T1& y)
     {
       Base::_check_template_params();
       Base::template _init2<T0,T1>(x, y);
     }
     #else
+    /** \brief Constructs a fixed-sized matrix initialized with coefficients starting at \a data */
+    EIGEN_DEVICE_FUNC
+    explicit Matrix(const Scalar *data);
+
+    /** \brief Constructs a vector or row-vector with given dimension. \only_for_vectors
+      *
+      * This is useful for dynamic-size vectors. For fixed-size vectors,
+      * it is redundant to pass these parameters, so one should use the default constructor
+      * Matrix() instead.
+      * 
+      * \warning This constructor is disabled for fixed-size \c 1x1 matrices. For instance,
+      * calling Matrix<double,1,1>(1) will call the initialization constructor: Matrix(const Scalar&).
+      * For fixed-size \c 1x1 matrices it is therefore recommended to use the default
+      * constructor Matrix() instead, especially when using one of the non standard
+      * \c EIGEN_INITIALIZE_MATRICES_BY_{ZERO,\c NAN} macros (see \ref TopicPreprocessorDirectives).
+      */
+    EIGEN_STRONG_INLINE explicit Matrix(Index dim);
+    /** \brief Constructs an initialized 1x1 matrix with the given coefficient */
+    Matrix(const Scalar& x);
     /** \brief Constructs an uninitialized matrix with \a rows rows and \a cols columns.
       *
       * This is useful for dynamic-size matrices. For fixed-size matrices,
       * it is redundant to pass these parameters, so one should use the default constructor
-      * Matrix() instead. */
+      * Matrix() instead.
+      * 
+      * \warning This constructor is disabled for fixed-size \c 1x2 and \c 2x1 vectors. For instance,
+      * calling Matrix2f(2,1) will call the initialization constructor: Matrix(const Scalar& x, const Scalar& y).
+      * For fixed-size \c 1x2 or \c 2x1 vectors it is therefore recommended to use the default
+      * constructor Matrix() instead, especially when using one of the non standard
+      * \c EIGEN_INITIALIZE_MATRICES_BY_{ZERO,\c NAN} macros (see \ref TopicPreprocessorDirectives).
+      */
+    EIGEN_DEVICE_FUNC
     Matrix(Index rows, Index cols);
+    
     /** \brief Constructs an initialized 2D vector with given coefficients */
     Matrix(const Scalar& x, const Scalar& y);
     #endif
 
     /** \brief Constructs an initialized 3D vector with given coefficients */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Matrix(const Scalar& x, const Scalar& y, const Scalar& z)
     {
       Base::_check_template_params();
@@ -270,6 +353,7 @@ class Matrix
       m_storage.data()[2] = z;
     }
     /** \brief Constructs an initialized 4D vector with given coefficients */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Matrix(const Scalar& x, const Scalar& y, const Scalar& z, const Scalar& w)
     {
       Base::_check_template_params();
@@ -280,76 +364,33 @@ class Matrix
       m_storage.data()[3] = w;
     }
 
-    explicit Matrix(const Scalar *data);
 
-    /** \brief Constructor copying the value of the expression \a other */
-    template<typename OtherDerived>
-    EIGEN_STRONG_INLINE Matrix(const MatrixBase<OtherDerived>& other)
-             : Base(other.rows() * other.cols(), other.rows(), other.cols())
-    {
-      // This test resides here, to bring the error messages closer to the user. Normally, these checks
-      // are performed deeply within the library, thus causing long and scary error traces.
-      EIGEN_STATIC_ASSERT((internal::is_same<Scalar, typename OtherDerived::Scalar>::value),
-        YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
-
-      Base::_check_template_params();
-      Base::_set_noalias(other);
-    }
     /** \brief Copy constructor */
-    EIGEN_STRONG_INLINE Matrix(const Matrix& other)
-            : Base(other.rows() * other.cols(), other.rows(), other.cols())
-    {
-      Base::_check_template_params();
-      Base::_set_noalias(other);
-    }
-    /** \brief Copy constructor with in-place evaluation */
-    template<typename OtherDerived>
-    EIGEN_STRONG_INLINE Matrix(const ReturnByValue<OtherDerived>& other)
-    {
-      Base::_check_template_params();
-      Base::resize(other.rows(), other.cols());
-      other.evalTo(*this);
-    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Matrix(const Matrix& other) : Base(other)
+    { }
 
     /** \brief Copy constructor for generic expressions.
       * \sa MatrixBase::operator=(const EigenBase<OtherDerived>&)
       */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Matrix(const EigenBase<OtherDerived> &other)
-      : Base(other.derived().rows() * other.derived().cols(), other.derived().rows(), other.derived().cols())
-    {
-      Base::_check_template_params();
-      Base::_resize_to_match(other);
-      // FIXME/CHECK: isn't *this = other.derived() more efficient. it allows to
-      //              go for pure _set() implementations, right?
-      *this = other;
-    }
-
-    /** \internal
-      * \brief Override MatrixBase::swap() since for dynamic-sized matrices
-      * of same type it is enough to swap the data pointers.
-      */
-    template<typename OtherDerived>
-    void swap(MatrixBase<OtherDerived> const & other)
-    { this->_swap(other.derived()); }
+      : Base(other.derived())
+    { }
 
-    inline Index innerStride() const { return 1; }
-    inline Index outerStride() const { return this->innerSize(); }
+    EIGEN_DEVICE_FUNC inline Index innerStride() const { return 1; }
+    EIGEN_DEVICE_FUNC inline Index outerStride() const { return this->innerSize(); }
 
     /////////// Geometry module ///////////
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     explicit Matrix(const RotationBase<OtherDerived,ColsAtCompileTime>& r);
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     Matrix& operator=(const RotationBase<OtherDerived,ColsAtCompileTime>& r);
 
-    #ifdef EIGEN2_SUPPORT
-    template<typename OtherDerived>
-    explicit Matrix(const eigen2_RotationBase<OtherDerived,ColsAtCompileTime>& r);
-    template<typename OtherDerived>
-    Matrix& operator=(const eigen2_RotationBase<OtherDerived,ColsAtCompileTime>& r);
-    #endif
-
     // allow to extend Matrix outside Eigen
     #ifdef EIGEN_MATRIX_PLUGIN
     #include EIGEN_MATRIX_PLUGIN
diff --git a/eigen/Eigen/src/Core/MatrixBase.h b/eigen/Eigen/src/Core/MatrixBase.h
index e83ef4d..200e577 100644
--- a/eigen/Eigen/src/Core/MatrixBase.h
+++ b/eigen/Eigen/src/Core/MatrixBase.h
@@ -41,9 +41,9 @@ namespace Eigen {
   * \endcode
   *
   * This class can be extended with the help of the plugin mechanism described on the page
-  * \ref TopicCustomizingEigen by defining the preprocessor symbol \c EIGEN_MATRIXBASE_PLUGIN.
+  * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_MATRIXBASE_PLUGIN.
   *
-  * \sa \ref TopicClassHierarchy
+  * \sa \blank \ref TopicClassHierarchy
   */
 template<typename Derived> class MatrixBase
   : public DenseBase<Derived>
@@ -52,7 +52,7 @@ template<typename Derived> class MatrixBase
 #ifndef EIGEN_PARSED_BY_DOXYGEN
     typedef MatrixBase StorageBaseType;
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
-    typedef typename internal::traits<Derived>::Index Index;
+    typedef typename internal::traits<Derived>::StorageIndex StorageIndex;
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef typename internal::packet_traits<Scalar>::type PacketScalar;
     typedef typename NumTraits<Scalar>::Real RealScalar;
@@ -66,7 +66,6 @@ template<typename Derived> class MatrixBase
     using Base::MaxSizeAtCompileTime;
     using Base::IsVectorAtCompileTime;
     using Base::Flags;
-    using Base::CoeffReadCost;
 
     using Base::derived;
     using Base::const_cast_derived;
@@ -77,6 +76,7 @@ template<typename Derived> class MatrixBase
     using Base::coeffRef;
     using Base::lazyAssign;
     using Base::eval;
+    using Base::operator-;
     using Base::operator+=;
     using Base::operator-=;
     using Base::operator*=;
@@ -98,25 +98,14 @@ template<typename Derived> class MatrixBase
 
     /** \returns the size of the main diagonal, which is min(rows(),cols()).
       * \sa rows(), cols(), SizeAtCompileTime. */
-    inline Index diagonalSize() const { return (std::min)(rows(),cols()); }
+    EIGEN_DEVICE_FUNC
+    inline Index diagonalSize() const { return (numext::mini)(rows(),cols()); }
 
-    /** \brief The plain matrix type corresponding to this expression.
-      *
-      * This is not necessarily exactly the return type of eval(). In the case of plain matrices,
-      * the return type of eval() is a const reference to a matrix, not a matrix! It is however guaranteed
-      * that the return type of eval() is either PlainObject or const PlainObject&.
-      */
-    typedef Matrix<typename internal::traits<Derived>::Scalar,
-                internal::traits<Derived>::RowsAtCompileTime,
-                internal::traits<Derived>::ColsAtCompileTime,
-                AutoAlign | (internal::traits<Derived>::Flags&RowMajorBit ? RowMajor : ColMajor),
-                internal::traits<Derived>::MaxRowsAtCompileTime,
-                internal::traits<Derived>::MaxColsAtCompileTime
-          > PlainObject;
+    typedef typename Base::PlainObject PlainObject;
 
 #ifndef EIGEN_PARSED_BY_DOXYGEN
     /** \internal Represents a matrix with all coefficients equal to one another*/
-    typedef CwiseNullaryOp<internal::scalar_constant_op<Scalar>,Derived> ConstantReturnType;
+    typedef CwiseNullaryOp<internal::scalar_constant_op<Scalar>,PlainObject> ConstantReturnType;
     /** \internal the return type of MatrixBase::adjoint() */
     typedef typename internal::conditional<NumTraits<Scalar>::IsComplex,
                         CwiseUnaryOp<internal::scalar_conjugate_op<Scalar>, ConstTransposeReturnType>,
@@ -125,7 +114,7 @@ template<typename Derived> class MatrixBase
     /** \internal Return type of eigenvalues() */
     typedef Matrix<std::complex<RealScalar>, internal::traits<Derived>::ColsAtCompileTime, 1, ColMajor> EigenvaluesReturnType;
     /** \internal the return type of identity */
-    typedef CwiseNullaryOp<internal::scalar_identity_op<Scalar>,Derived> IdentityReturnType;
+    typedef CwiseNullaryOp<internal::scalar_identity_op<Scalar>,PlainObject> IdentityReturnType;
     /** \internal the return type of unit vectors */
     typedef Block<const CwiseNullaryOp<internal::scalar_identity_op<Scalar>, SquareMatrixType>,
                   internal::traits<Derived>::RowsAtCompileTime,
@@ -133,7 +122,7 @@ template<typename Derived> class MatrixBase
 #endif // not EIGEN_PARSED_BY_DOXYGEN
 
 #define EIGEN_CURRENT_STORAGE_BASE_CLASS Eigen::MatrixBase
-#   include "../plugins/CommonCwiseUnaryOps.h"
+#define EIGEN_DOC_UNARY_ADDONS(X,Y)
 #   include "../plugins/CommonCwiseBinaryOps.h"
 #   include "../plugins/MatrixCwiseUnaryOps.h"
 #   include "../plugins/MatrixCwiseBinaryOps.h"
@@ -141,41 +130,53 @@ template<typename Derived> class MatrixBase
 #     include EIGEN_MATRIXBASE_PLUGIN
 #   endif
 #undef EIGEN_CURRENT_STORAGE_BASE_CLASS
+#undef EIGEN_DOC_UNARY_ADDONS
 
     /** Special case of the template operator=, in order to prevent the compiler
       * from generating a default operator= (issue hit with g++ 4.1)
       */
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     Derived& operator=(const MatrixBase& other);
 
     // We cannot inherit here via Base::operator= since it is causing
     // trouble with MSVC.
 
     template <typename OtherDerived>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     Derived& operator=(const DenseBase<OtherDerived>& other);
 
     template <typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     Derived& operator=(const EigenBase<OtherDerived>& other);
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     Derived& operator=(const ReturnByValue<OtherDerived>& other);
 
-    template<typename ProductDerived, typename Lhs, typename Rhs>
-    Derived& lazyAssign(const ProductBase<ProductDerived, Lhs,Rhs>& other);
-
-    template<typename MatrixPower, typename Lhs, typename Rhs>
-    Derived& lazyAssign(const MatrixPowerProduct<MatrixPower, Lhs,Rhs>& other);
-
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     Derived& operator+=(const MatrixBase<OtherDerived>& other);
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     Derived& operator-=(const MatrixBase<OtherDerived>& other);
 
+#ifdef __CUDACC__
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    const Product<Derived,OtherDerived,LazyProduct>
+    operator*(const MatrixBase<OtherDerived> &other) const
+    { return this->lazyProduct(other); }
+#else
+
     template<typename OtherDerived>
-    const typename ProductReturnType<Derived,OtherDerived>::Type
+    const Product<Derived,OtherDerived>
     operator*(const MatrixBase<OtherDerived> &other) const;
 
+#endif
+
     template<typename OtherDerived>
-    const typename LazyProductReturnType<Derived,OtherDerived>::Type
+    EIGEN_DEVICE_FUNC
+    const Product<Derived,OtherDerived,LazyProduct>
     lazyProduct(const MatrixBase<OtherDerived> &other) const;
 
     template<typename OtherDerived>
@@ -188,84 +189,93 @@ template<typename Derived> class MatrixBase
     void applyOnTheRight(const EigenBase<OtherDerived>& other);
 
     template<typename DiagonalDerived>
-    const DiagonalProduct<Derived, DiagonalDerived, OnTheRight>
+    EIGEN_DEVICE_FUNC
+    const Product<Derived, DiagonalDerived, LazyProduct>
     operator*(const DiagonalBase<DiagonalDerived> &diagonal) const;
 
     template<typename OtherDerived>
-    typename internal::scalar_product_traits<typename internal::traits<Derived>::Scalar,typename internal::traits<OtherDerived>::Scalar>::ReturnType
+    EIGEN_DEVICE_FUNC
+    typename ScalarBinaryOpTraits<typename internal::traits<Derived>::Scalar,typename internal::traits<OtherDerived>::Scalar>::ReturnType
     dot(const MatrixBase<OtherDerived>& other) const;
 
-    #ifdef EIGEN2_SUPPORT
-      template<typename OtherDerived>
-      Scalar eigen2_dot(const MatrixBase<OtherDerived>& other) const;
-    #endif
-
-    RealScalar squaredNorm() const;
-    RealScalar norm() const;
+    EIGEN_DEVICE_FUNC RealScalar squaredNorm() const;
+    EIGEN_DEVICE_FUNC RealScalar norm() const;
     RealScalar stableNorm() const;
     RealScalar blueNorm() const;
     RealScalar hypotNorm() const;
-    const PlainObject normalized() const;
-    void normalize();
+    EIGEN_DEVICE_FUNC const PlainObject normalized() const;
+    EIGEN_DEVICE_FUNC const PlainObject stableNormalized() const;
+    EIGEN_DEVICE_FUNC void normalize();
+    EIGEN_DEVICE_FUNC void stableNormalize();
 
-    const AdjointReturnType adjoint() const;
-    void adjointInPlace();
+    EIGEN_DEVICE_FUNC const AdjointReturnType adjoint() const;
+    EIGEN_DEVICE_FUNC void adjointInPlace();
 
     typedef Diagonal<Derived> DiagonalReturnType;
+    EIGEN_DEVICE_FUNC
     DiagonalReturnType diagonal();
+
     typedef typename internal::add_const<Diagonal<const Derived> >::type ConstDiagonalReturnType;
+    EIGEN_DEVICE_FUNC
     ConstDiagonalReturnType diagonal() const;
 
     template<int Index> struct DiagonalIndexReturnType { typedef Diagonal<Derived,Index> Type; };
     template<int Index> struct ConstDiagonalIndexReturnType { typedef const Diagonal<const Derived,Index> Type; };
 
-    template<int Index> typename DiagonalIndexReturnType<Index>::Type diagonal();
-    template<int Index> typename ConstDiagonalIndexReturnType<Index>::Type diagonal() const;
-    
+    template<int Index>
+    EIGEN_DEVICE_FUNC
+    typename DiagonalIndexReturnType<Index>::Type diagonal();
+
+    template<int Index>
+    EIGEN_DEVICE_FUNC
+    typename ConstDiagonalIndexReturnType<Index>::Type diagonal() const;
+
     typedef Diagonal<Derived,DynamicIndex> DiagonalDynamicIndexReturnType;
     typedef typename internal::add_const<Diagonal<const Derived,DynamicIndex> >::type ConstDiagonalDynamicIndexReturnType;
 
+    EIGEN_DEVICE_FUNC
     DiagonalDynamicIndexReturnType diagonal(Index index);
+    EIGEN_DEVICE_FUNC
     ConstDiagonalDynamicIndexReturnType diagonal(Index index) const;
 
-    #ifdef EIGEN2_SUPPORT
-    template<unsigned int Mode> typename internal::eigen2_part_return_type<Derived, Mode>::type part();
-    template<unsigned int Mode> const typename internal::eigen2_part_return_type<Derived, Mode>::type part() const;
-    
-    // huuuge hack. make Eigen2's matrix.part<Diagonal>() work in eigen3. Problem: Diagonal is now a class template instead
-    // of an integer constant. Solution: overload the part() method template wrt template parameters list.
-    template<template<typename T, int N> class U>
-    const DiagonalWrapper<ConstDiagonalReturnType> part() const
-    { return diagonal().asDiagonal(); }
-    #endif // EIGEN2_SUPPORT
-
     template<unsigned int Mode> struct TriangularViewReturnType { typedef TriangularView<Derived, Mode> Type; };
     template<unsigned int Mode> struct ConstTriangularViewReturnType { typedef const TriangularView<const Derived, Mode> Type; };
 
-    template<unsigned int Mode> typename TriangularViewReturnType<Mode>::Type triangularView();
-    template<unsigned int Mode> typename ConstTriangularViewReturnType<Mode>::Type triangularView() const;
+    template<unsigned int Mode>
+    EIGEN_DEVICE_FUNC
+    typename TriangularViewReturnType<Mode>::Type triangularView();
+    template<unsigned int Mode>
+    EIGEN_DEVICE_FUNC
+    typename ConstTriangularViewReturnType<Mode>::Type triangularView() const;
 
     template<unsigned int UpLo> struct SelfAdjointViewReturnType { typedef SelfAdjointView<Derived, UpLo> Type; };
     template<unsigned int UpLo> struct ConstSelfAdjointViewReturnType { typedef const SelfAdjointView<const Derived, UpLo> Type; };
 
-    template<unsigned int UpLo> typename SelfAdjointViewReturnType<UpLo>::Type selfadjointView();
-    template<unsigned int UpLo> typename ConstSelfAdjointViewReturnType<UpLo>::Type selfadjointView() const;
+    template<unsigned int UpLo>
+    EIGEN_DEVICE_FUNC
+    typename SelfAdjointViewReturnType<UpLo>::Type selfadjointView();
+    template<unsigned int UpLo>
+    EIGEN_DEVICE_FUNC
+    typename ConstSelfAdjointViewReturnType<UpLo>::Type selfadjointView() const;
 
     const SparseView<Derived> sparseView(const Scalar& m_reference = Scalar(0),
                                          const typename NumTraits<Scalar>::Real& m_epsilon = NumTraits<Scalar>::dummy_precision()) const;
-    static const IdentityReturnType Identity();
-    static const IdentityReturnType Identity(Index rows, Index cols);
-    static const BasisReturnType Unit(Index size, Index i);
-    static const BasisReturnType Unit(Index i);
-    static const BasisReturnType UnitX();
-    static const BasisReturnType UnitY();
-    static const BasisReturnType UnitZ();
-    static const BasisReturnType UnitW();
-
+    EIGEN_DEVICE_FUNC static const IdentityReturnType Identity();
+    EIGEN_DEVICE_FUNC static const IdentityReturnType Identity(Index rows, Index cols);
+    EIGEN_DEVICE_FUNC static const BasisReturnType Unit(Index size, Index i);
+    EIGEN_DEVICE_FUNC static const BasisReturnType Unit(Index i);
+    EIGEN_DEVICE_FUNC static const BasisReturnType UnitX();
+    EIGEN_DEVICE_FUNC static const BasisReturnType UnitY();
+    EIGEN_DEVICE_FUNC static const BasisReturnType UnitZ();
+    EIGEN_DEVICE_FUNC static const BasisReturnType UnitW();
+
+    EIGEN_DEVICE_FUNC
     const DiagonalWrapper<const Derived> asDiagonal() const;
     const PermutationWrapper<const Derived> asPermutation() const;
 
+    EIGEN_DEVICE_FUNC
     Derived& setIdentity();
+    EIGEN_DEVICE_FUNC
     Derived& setIdentity(Index rows, Index cols);
 
     bool isIdentity(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
@@ -284,7 +294,7 @@ template<typename Derived> class MatrixBase
       *          fuzzy comparison such as isApprox()
       * \sa isApprox(), operator!= */
     template<typename OtherDerived>
-    inline bool operator==(const MatrixBase<OtherDerived>& other) const
+    EIGEN_DEVICE_FUNC inline bool operator==(const MatrixBase<OtherDerived>& other) const
     { return cwiseEqual(other).all(); }
 
     /** \returns true if at least one pair of coefficients of \c *this and \a other are not exactly equal to each other.
@@ -292,64 +302,50 @@ template<typename Derived> class MatrixBase
       *          fuzzy comparison such as isApprox()
       * \sa isApprox(), operator== */
     template<typename OtherDerived>
-    inline bool operator!=(const MatrixBase<OtherDerived>& other) const
+    EIGEN_DEVICE_FUNC inline bool operator!=(const MatrixBase<OtherDerived>& other) const
     { return cwiseNotEqual(other).any(); }
 
     NoAlias<Derived,Eigen::MatrixBase > noalias();
 
-    inline const ForceAlignedAccess<Derived> forceAlignedAccess() const;
-    inline ForceAlignedAccess<Derived> forceAlignedAccess();
-    template<bool Enable> inline typename internal::add_const_on_value_type<typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type>::type forceAlignedAccessIf() const;
-    template<bool Enable> inline typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type forceAlignedAccessIf();
+    // TODO forceAlignedAccess is temporarily disabled
+    // Need to find a nicer workaround.
+    inline const Derived& forceAlignedAccess() const { return derived(); }
+    inline Derived& forceAlignedAccess() { return derived(); }
+    template<bool Enable> inline const Derived& forceAlignedAccessIf() const { return derived(); }
+    template<bool Enable> inline Derived& forceAlignedAccessIf() { return derived(); }
 
-    Scalar trace() const;
+    EIGEN_DEVICE_FUNC Scalar trace() const;
 
-/////////// Array module ///////////
+    template<int p> EIGEN_DEVICE_FUNC RealScalar lpNorm() const;
 
-    template<int p> RealScalar lpNorm() const;
-
-    MatrixBase<Derived>& matrix() { return *this; }
-    const MatrixBase<Derived>& matrix() const { return *this; }
+    EIGEN_DEVICE_FUNC MatrixBase<Derived>& matrix() { return *this; }
+    EIGEN_DEVICE_FUNC const MatrixBase<Derived>& matrix() const { return *this; }
 
     /** \returns an \link Eigen::ArrayBase Array \endlink expression of this matrix
       * \sa ArrayBase::matrix() */
-    ArrayWrapper<Derived> array() { return derived(); }
-    const ArrayWrapper<const Derived> array() const { return derived(); }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ArrayWrapper<Derived> array() { return ArrayWrapper<Derived>(derived()); }
+    /** \returns a const \link Eigen::ArrayBase Array \endlink expression of this matrix
+      * \sa ArrayBase::matrix() */
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const ArrayWrapper<const Derived> array() const { return ArrayWrapper<const Derived>(derived()); }
 
 /////////// LU module ///////////
 
-    const FullPivLU<PlainObject> fullPivLu() const;
-    const PartialPivLU<PlainObject> partialPivLu() const;
+    inline const FullPivLU<PlainObject> fullPivLu() const;
+    inline const PartialPivLU<PlainObject> partialPivLu() const;
 
-    #if EIGEN2_SUPPORT_STAGE < STAGE20_RESOLVE_API_CONFLICTS
-    const LU<PlainObject> lu() const;
-    #endif
+    inline const PartialPivLU<PlainObject> lu() const;
 
-    #ifdef EIGEN2_SUPPORT
-    const LU<PlainObject> eigen2_lu() const;
-    #endif
+    inline const Inverse<Derived> inverse() const;
 
-    #if EIGEN2_SUPPORT_STAGE > STAGE20_RESOLVE_API_CONFLICTS
-    const PartialPivLU<PlainObject> lu() const;
-    #endif
-    
-    #ifdef EIGEN2_SUPPORT
     template<typename ResultType>
-    void computeInverse(MatrixBase<ResultType> *result) const {
-      *result = this->inverse();
-    }
-    #endif
-
-    const internal::inverse_impl<Derived> inverse() const;
-    template<typename ResultType>
-    void computeInverseAndDetWithCheck(
+    inline void computeInverseAndDetWithCheck(
       ResultType& inverse,
       typename ResultType::Scalar& determinant,
       bool& invertible,
       const RealScalar& absDeterminantThreshold = NumTraits<Scalar>::dummy_precision()
     ) const;
     template<typename ResultType>
-    void computeInverseWithCheck(
+    inline void computeInverseWithCheck(
       ResultType& inverse,
       bool& invertible,
       const RealScalar& absDeterminantThreshold = NumTraits<Scalar>::dummy_precision()
@@ -358,65 +354,70 @@ template<typename Derived> class MatrixBase
 
 /////////// Cholesky module ///////////
 
-    const LLT<PlainObject>  llt() const;
-    const LDLT<PlainObject> ldlt() const;
+    inline const LLT<PlainObject>  llt() const;
+    inline const LDLT<PlainObject> ldlt() const;
 
 /////////// QR module ///////////
 
-    const HouseholderQR<PlainObject> householderQr() const;
-    const ColPivHouseholderQR<PlainObject> colPivHouseholderQr() const;
-    const FullPivHouseholderQR<PlainObject> fullPivHouseholderQr() const;
-    
-    #ifdef EIGEN2_SUPPORT
-    const QR<PlainObject> qr() const;
-    #endif
+    inline const HouseholderQR<PlainObject> householderQr() const;
+    inline const ColPivHouseholderQR<PlainObject> colPivHouseholderQr() const;
+    inline const FullPivHouseholderQR<PlainObject> fullPivHouseholderQr() const;
+    inline const CompleteOrthogonalDecomposition<PlainObject> completeOrthogonalDecomposition() const;
 
-    EigenvaluesReturnType eigenvalues() const;
-    RealScalar operatorNorm() const;
+/////////// Eigenvalues module ///////////
 
-/////////// SVD module ///////////
+    inline EigenvaluesReturnType eigenvalues() const;
+    inline RealScalar operatorNorm() const;
 
-    JacobiSVD<PlainObject> jacobiSvd(unsigned int computationOptions = 0) const;
+/////////// SVD module ///////////
 
-    #ifdef EIGEN2_SUPPORT
-    SVD<PlainObject> svd() const;
-    #endif
+    inline JacobiSVD<PlainObject> jacobiSvd(unsigned int computationOptions = 0) const;
+    inline BDCSVD<PlainObject>    bdcSvd(unsigned int computationOptions = 0) const;
 
 /////////// Geometry module ///////////
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     /// \internal helper struct to form the return type of the cross product
     template<typename OtherDerived> struct cross_product_return_type {
-      typedef typename internal::scalar_product_traits<typename internal::traits<Derived>::Scalar,typename internal::traits<OtherDerived>::Scalar>::ReturnType Scalar;
+      typedef typename ScalarBinaryOpTraits<typename internal::traits<Derived>::Scalar,typename internal::traits<OtherDerived>::Scalar>::ReturnType Scalar;
       typedef Matrix<Scalar,MatrixBase::RowsAtCompileTime,MatrixBase::ColsAtCompileTime> type;
     };
     #endif // EIGEN_PARSED_BY_DOXYGEN
     template<typename OtherDerived>
-    typename cross_product_return_type<OtherDerived>::type
+    EIGEN_DEVICE_FUNC
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+    inline typename cross_product_return_type<OtherDerived>::type
+#else
+    inline PlainObject
+#endif
     cross(const MatrixBase<OtherDerived>& other) const;
+
     template<typename OtherDerived>
-    PlainObject cross3(const MatrixBase<OtherDerived>& other) const;
-    PlainObject unitOrthogonal(void) const;
-    Matrix<Scalar,3,1> eulerAngles(Index a0, Index a1, Index a2) const;
-    
-    #if EIGEN2_SUPPORT_STAGE > STAGE20_RESOLVE_API_CONFLICTS
-    ScalarMultipleReturnType operator*(const UniformScaling<Scalar>& s) const;
+    EIGEN_DEVICE_FUNC
+    inline PlainObject cross3(const MatrixBase<OtherDerived>& other) const;
+
+    EIGEN_DEVICE_FUNC
+    inline PlainObject unitOrthogonal(void) const;
+
+    EIGEN_DEVICE_FUNC
+    inline Matrix<Scalar,3,1> eulerAngles(Index a0, Index a1, Index a2) const;
+
     // put this as separate enum value to work around possible GCC 4.3 bug (?)
-    enum { HomogeneousReturnTypeDirection = ColsAtCompileTime==1?Vertical:Horizontal };
+    enum { HomogeneousReturnTypeDirection = ColsAtCompileTime==1&&RowsAtCompileTime==1 ? ((internal::traits<Derived>::Flags&RowMajorBit)==RowMajorBit ? Horizontal : Vertical)
+                                          : ColsAtCompileTime==1 ? Vertical : Horizontal };
     typedef Homogeneous<Derived, HomogeneousReturnTypeDirection> HomogeneousReturnType;
-    HomogeneousReturnType homogeneous() const;
-    #endif
-    
+    EIGEN_DEVICE_FUNC
+    inline HomogeneousReturnType homogeneous() const;
+
     enum {
       SizeMinusOne = SizeAtCompileTime==Dynamic ? Dynamic : SizeAtCompileTime-1
     };
     typedef Block<const Derived,
                   internal::traits<Derived>::ColsAtCompileTime==1 ? SizeMinusOne : 1,
                   internal::traits<Derived>::ColsAtCompileTime==1 ? 1 : SizeMinusOne> ConstStartMinusOne;
-    typedef CwiseUnaryOp<internal::scalar_quotient1_op<typename internal::traits<Derived>::Scalar>,
-                const ConstStartMinusOne > HNormalizedReturnType;
-
-    const HNormalizedReturnType hnormalized() const;
+    typedef EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(ConstStartMinusOne,Scalar,quotient) HNormalizedReturnType;
+    EIGEN_DEVICE_FUNC
+    inline const HNormalizedReturnType hnormalized() const;
 
 ////////// Householder module ///////////
 
@@ -461,49 +462,15 @@ template<typename Derived> class MatrixBase
     const MatrixSquareRootReturnValue<Derived> sqrt() const;
     const MatrixLogarithmReturnValue<Derived> log() const;
     const MatrixPowerReturnValue<Derived> pow(const RealScalar& p) const;
-
-#ifdef EIGEN2_SUPPORT
-    template<typename ProductDerived, typename Lhs, typename Rhs>
-    Derived& operator+=(const Flagged<ProductBase<ProductDerived, Lhs,Rhs>, 0,
-                                      EvalBeforeAssigningBit>& other);
-
-    template<typename ProductDerived, typename Lhs, typename Rhs>
-    Derived& operator-=(const Flagged<ProductBase<ProductDerived, Lhs,Rhs>, 0,
-                                      EvalBeforeAssigningBit>& other);
-
-    /** \deprecated because .lazy() is deprecated
-      * Overloaded for cache friendly product evaluation */
-    template<typename OtherDerived>
-    Derived& lazyAssign(const Flagged<OtherDerived, 0, EvalBeforeAssigningBit>& other)
-    { return lazyAssign(other._expression()); }
-
-    template<unsigned int Added>
-    const Flagged<Derived, Added, 0> marked() const;
-    const Flagged<Derived, 0, EvalBeforeAssigningBit> lazy() const;
-
-    inline const Cwise<Derived> cwise() const;
-    inline Cwise<Derived> cwise();
-
-    VectorBlock<Derived> start(Index size);
-    const VectorBlock<const Derived> start(Index size) const;
-    VectorBlock<Derived> end(Index size);
-    const VectorBlock<const Derived> end(Index size) const;
-    template<int Size> VectorBlock<Derived,Size> start();
-    template<int Size> const VectorBlock<const Derived,Size> start() const;
-    template<int Size> VectorBlock<Derived,Size> end();
-    template<int Size> const VectorBlock<const Derived,Size> end() const;
-
-    Minor<Derived> minor(Index row, Index col);
-    const Minor<Derived> minor(Index row, Index col) const;
-#endif
+    const MatrixComplexPowerReturnValue<Derived> pow(const std::complex<RealScalar>& p) const;
 
   protected:
-    MatrixBase() : Base() {}
+    EIGEN_DEVICE_FUNC MatrixBase() : Base() {}
 
   private:
-    explicit MatrixBase(int);
-    MatrixBase(int,int);
-    template<typename OtherDerived> explicit MatrixBase(const MatrixBase<OtherDerived>&);
+    EIGEN_DEVICE_FUNC explicit MatrixBase(int);
+    EIGEN_DEVICE_FUNC MatrixBase(int,int);
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC explicit MatrixBase(const MatrixBase<OtherDerived>&);
   protected:
     // mixing arrays and matrices is not legal
     template<typename OtherDerived> Derived& operator+=(const ArrayBase<OtherDerived>& )
diff --git a/eigen/Eigen/src/Core/NestByValue.h b/eigen/Eigen/src/Core/NestByValue.h
index a893b17..01cf192 100644
--- a/eigen/Eigen/src/Core/NestByValue.h
+++ b/eigen/Eigen/src/Core/NestByValue.h
@@ -13,25 +13,24 @@
 
 namespace Eigen {
 
+namespace internal {
+template<typename ExpressionType>
+struct traits<NestByValue<ExpressionType> > : public traits<ExpressionType>
+{};
+}
+
 /** \class NestByValue
   * \ingroup Core_Module
   *
   * \brief Expression which must be nested by value
   *
-  * \param ExpressionType the type of the object of which we are requiring nesting-by-value
+  * \tparam ExpressionType the type of the object of which we are requiring nesting-by-value
   *
   * This class is the return type of MatrixBase::nestByValue()
   * and most of the time this is the only way it is used.
   *
   * \sa MatrixBase::nestByValue()
   */
-
-namespace internal {
-template<typename ExpressionType>
-struct traits<NestByValue<ExpressionType> > : public traits<ExpressionType>
-{};
-}
-
 template<typename ExpressionType> class NestByValue
   : public internal::dense_xpr_base< NestByValue<ExpressionType> >::type
 {
@@ -40,58 +39,58 @@ template<typename ExpressionType> class NestByValue
     typedef typename internal::dense_xpr_base<NestByValue>::type Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(NestByValue)
 
-    inline NestByValue(const ExpressionType& matrix) : m_expression(matrix) {}
+    EIGEN_DEVICE_FUNC explicit inline NestByValue(const ExpressionType& matrix) : m_expression(matrix) {}
 
-    inline Index rows() const { return m_expression.rows(); }
-    inline Index cols() const { return m_expression.cols(); }
-    inline Index outerStride() const { return m_expression.outerStride(); }
-    inline Index innerStride() const { return m_expression.innerStride(); }
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_expression.rows(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_expression.cols(); }
+    EIGEN_DEVICE_FUNC inline Index outerStride() const { return m_expression.outerStride(); }
+    EIGEN_DEVICE_FUNC inline Index innerStride() const { return m_expression.innerStride(); }
 
-    inline const CoeffReturnType coeff(Index row, Index col) const
+    EIGEN_DEVICE_FUNC inline const CoeffReturnType coeff(Index row, Index col) const
     {
       return m_expression.coeff(row, col);
     }
 
-    inline Scalar& coeffRef(Index row, Index col)
+    EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index row, Index col)
     {
       return m_expression.const_cast_derived().coeffRef(row, col);
     }
 
-    inline const CoeffReturnType coeff(Index index) const
+    EIGEN_DEVICE_FUNC inline const CoeffReturnType coeff(Index index) const
     {
       return m_expression.coeff(index);
     }
 
-    inline Scalar& coeffRef(Index index)
+    EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index index)
     {
       return m_expression.const_cast_derived().coeffRef(index);
     }
 
     template<int LoadMode>
-    inline const PacketScalar packet(Index row, Index col) const
+    EIGEN_DEVICE_FUNC inline const PacketScalar packet(Index row, Index col) const
     {
       return m_expression.template packet<LoadMode>(row, col);
     }
 
     template<int LoadMode>
-    inline void writePacket(Index row, Index col, const PacketScalar& x)
+    EIGEN_DEVICE_FUNC inline void writePacket(Index row, Index col, const PacketScalar& x)
     {
       m_expression.const_cast_derived().template writePacket<LoadMode>(row, col, x);
     }
 
     template<int LoadMode>
-    inline const PacketScalar packet(Index index) const
+    EIGEN_DEVICE_FUNC inline const PacketScalar packet(Index index) const
     {
       return m_expression.template packet<LoadMode>(index);
     }
 
     template<int LoadMode>
-    inline void writePacket(Index index, const PacketScalar& x)
+    EIGEN_DEVICE_FUNC inline void writePacket(Index index, const PacketScalar& x)
     {
       m_expression.const_cast_derived().template writePacket<LoadMode>(index, x);
     }
 
-    operator const ExpressionType&() const { return m_expression; }
+    EIGEN_DEVICE_FUNC operator const ExpressionType&() const { return m_expression; }
 
   protected:
     const ExpressionType m_expression;
@@ -100,7 +99,7 @@ template<typename ExpressionType> class NestByValue
 /** \returns an expression of the temporary version of *this.
   */
 template<typename Derived>
-inline const NestByValue<Derived>
+EIGEN_DEVICE_FUNC inline const NestByValue<Derived>
 DenseBase<Derived>::nestByValue() const
 {
   return NestByValue<Derived>(derived());
diff --git a/eigen/Eigen/src/Core/NoAlias.h b/eigen/Eigen/src/Core/NoAlias.h
index 768bfb1..3390801 100644
--- a/eigen/Eigen/src/Core/NoAlias.h
+++ b/eigen/Eigen/src/Core/NoAlias.h
@@ -17,7 +17,7 @@ namespace Eigen {
   *
   * \brief Pseudo expression providing an operator = assuming no aliasing
   *
-  * \param ExpressionType the type of the object on which to do the lazy assignment
+  * \tparam ExpressionType the type of the object on which to do the lazy assignment
   *
   * This class represents an expression with special assignment operators
   * assuming no aliasing between the target expression and the source expression.
@@ -30,62 +30,36 @@ namespace Eigen {
 template<typename ExpressionType, template <typename> class StorageBase>
 class NoAlias
 {
-    typedef typename ExpressionType::Scalar Scalar;
   public:
-    NoAlias(ExpressionType& expression) : m_expression(expression) {}
-
-    /** Behaves like MatrixBase::lazyAssign(other)
-      * \sa MatrixBase::lazyAssign() */
+    typedef typename ExpressionType::Scalar Scalar;
+    
+    explicit NoAlias(ExpressionType& expression) : m_expression(expression) {}
+    
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE ExpressionType& operator=(const StorageBase<OtherDerived>& other)
-    { return internal::assign_selector<ExpressionType,OtherDerived,false>::run(m_expression,other.derived()); }
-
-    /** \sa MatrixBase::operator+= */
+    {
+      call_assignment_no_alias(m_expression, other.derived(), internal::assign_op<Scalar,typename OtherDerived::Scalar>());
+      return m_expression;
+    }
+    
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE ExpressionType& operator+=(const StorageBase<OtherDerived>& other)
     {
-      typedef SelfCwiseBinaryOp<internal::scalar_sum_op<Scalar>, ExpressionType, OtherDerived> SelfAdder;
-      SelfAdder tmp(m_expression);
-      typedef typename internal::nested<OtherDerived>::type OtherDerivedNested;
-      typedef typename internal::remove_all<OtherDerivedNested>::type _OtherDerivedNested;
-      internal::assign_selector<SelfAdder,_OtherDerivedNested,false>::run(tmp,OtherDerivedNested(other.derived()));
+      call_assignment_no_alias(m_expression, other.derived(), internal::add_assign_op<Scalar,typename OtherDerived::Scalar>());
       return m_expression;
     }
-
-    /** \sa MatrixBase::operator-= */
+    
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE ExpressionType& operator-=(const StorageBase<OtherDerived>& other)
     {
-      typedef SelfCwiseBinaryOp<internal::scalar_difference_op<Scalar>, ExpressionType, OtherDerived> SelfAdder;
-      SelfAdder tmp(m_expression);
-      typedef typename internal::nested<OtherDerived>::type OtherDerivedNested;
-      typedef typename internal::remove_all<OtherDerivedNested>::type _OtherDerivedNested;
-      internal::assign_selector<SelfAdder,_OtherDerivedNested,false>::run(tmp,OtherDerivedNested(other.derived()));
+      call_assignment_no_alias(m_expression, other.derived(), internal::sub_assign_op<Scalar,typename OtherDerived::Scalar>());
       return m_expression;
     }
 
-#ifndef EIGEN_PARSED_BY_DOXYGEN
-    template<typename ProductDerived, typename Lhs, typename Rhs>
-    EIGEN_STRONG_INLINE ExpressionType& operator+=(const ProductBase<ProductDerived, Lhs,Rhs>& other)
-    { other.derived().addTo(m_expression); return m_expression; }
-
-    template<typename ProductDerived, typename Lhs, typename Rhs>
-    EIGEN_STRONG_INLINE ExpressionType& operator-=(const ProductBase<ProductDerived, Lhs,Rhs>& other)
-    { other.derived().subTo(m_expression); return m_expression; }
-
-    template<typename Lhs, typename Rhs, int NestingFlags>
-    EIGEN_STRONG_INLINE ExpressionType& operator+=(const CoeffBasedProduct<Lhs,Rhs,NestingFlags>& other)
-    { return m_expression.derived() += CoeffBasedProduct<Lhs,Rhs,NestByRefBit>(other.lhs(), other.rhs()); }
-
-    template<typename Lhs, typename Rhs, int NestingFlags>
-    EIGEN_STRONG_INLINE ExpressionType& operator-=(const CoeffBasedProduct<Lhs,Rhs,NestingFlags>& other)
-    { return m_expression.derived() -= CoeffBasedProduct<Lhs,Rhs,NestByRefBit>(other.lhs(), other.rhs()); }
-    
-    template<typename OtherDerived>
-    ExpressionType& operator=(const ReturnByValue<OtherDerived>& func)
-    { return m_expression = func; }
-#endif
-
+    EIGEN_DEVICE_FUNC
     ExpressionType& expression() const
     {
       return m_expression;
@@ -126,7 +100,7 @@ class NoAlias
 template<typename Derived>
 NoAlias<Derived,MatrixBase> MatrixBase<Derived>::noalias()
 {
-  return derived();
+  return NoAlias<Derived, Eigen::MatrixBase >(derived());
 }
 
 } // end namespace Eigen
diff --git a/eigen/Eigen/src/Core/NumTraits.h b/eigen/Eigen/src/Core/NumTraits.h
index bac9e50..aebc0c2 100644
--- a/eigen/Eigen/src/Core/NumTraits.h
+++ b/eigen/Eigen/src/Core/NumTraits.h
@@ -12,24 +12,57 @@
 
 namespace Eigen {
 
+namespace internal {
+
+// default implementation of digits10(), based on numeric_limits if specialized,
+// 0 for integer types, and log10(epsilon()) otherwise.
+template< typename T,
+          bool use_numeric_limits = std::numeric_limits<T>::is_specialized,
+          bool is_integer = NumTraits<T>::IsInteger>
+struct default_digits10_impl
+{
+  static int run() { return std::numeric_limits<T>::digits10; }
+};
+
+template<typename T>
+struct default_digits10_impl<T,false,false> // Floating point
+{
+  static int run() {
+    using std::log10;
+    using std::ceil;
+    typedef typename NumTraits<T>::Real Real;
+    return int(ceil(-log10(NumTraits<Real>::epsilon())));
+  }
+};
+
+template<typename T>
+struct default_digits10_impl<T,false,true> // Integer
+{
+  static int run() { return 0; }
+};
+
+} // end namespace internal
+
 /** \class NumTraits
   * \ingroup Core_Module
   *
   * \brief Holds information about the various numeric (i.e. scalar) types allowed by Eigen.
   *
-  * \param T the numeric type at hand
+  * \tparam T the numeric type at hand
   *
   * This class stores enums, typedefs and static methods giving information about a numeric type.
   *
   * The provided data consists of:
-  * \li A typedef \a Real, giving the "real part" type of \a T. If \a T is already real,
-  *     then \a Real is just a typedef to \a T. If \a T is \c std::complex<U> then \a Real
+  * \li A typedef \c Real, giving the "real part" type of \a T. If \a T is already real,
+  *     then \c Real is just a typedef to \a T. If \a T is \c std::complex<U> then \c Real
   *     is a typedef to \a U.
-  * \li A typedef \a NonInteger, giving the type that should be used for operations producing non-integral values,
+  * \li A typedef \c NonInteger, giving the type that should be used for operations producing non-integral values,
   *     such as quotients, square roots, etc. If \a T is a floating-point type, then this typedef just gives
   *     \a T again. Note however that many Eigen functions such as internal::sqrt simply refuse to
   *     take integers. Outside of a few cases, Eigen doesn't do automatic type promotion. Thus, this typedef is
   *     only intended as a helper for code that needs to explicitly promote types.
+  * \li A typedef \c Literal giving the type to use for numeric literals such as "2" or "0.5". For instance, for \c std::complex<U>, Literal is defined as \c U.
+  *     Of course, this type must be fully compatible with \a T. In doubt, just use \a T here.
   * \li A typedef \a Nested giving the type to use to nest a value inside of the expression tree. If you don't know what
   *     this means, just use \a T here.
   * \li An enum value \a IsComplex. It is equal to 1 if \a T is a \c std::complex
@@ -38,14 +71,18 @@ namespace Eigen {
   *     and to \c 0 otherwise.
   * \li Enum values ReadCost, AddCost and MulCost representing a rough estimate of the number of CPU cycles needed
   *     to by move / add / mul instructions respectively, assuming the data is already stored in CPU registers.
-  *     Stay vague here. No need to do architecture-specific stuff.
+  *     Stay vague here. No need to do architecture-specific stuff. If you don't know what this means, just use \c Eigen::HugeCost.
   * \li An enum value \a IsSigned. It is equal to \c 1 if \a T is a signed type and to 0 if \a T is unsigned.
   * \li An enum value \a RequireInitialization. It is equal to \c 1 if the constructor of the numeric type \a T must
   *     be called, and to 0 if it is safe not to call it. Default is 0 if \a T is an arithmetic type, and 1 otherwise.
-  * \li An epsilon() function which, unlike std::numeric_limits::epsilon(), returns a \a Real instead of a \a T.
+  * \li An epsilon() function which, unlike <a href="http://en.cppreference.com/w/cpp/types/numeric_limits/epsilon">std::numeric_limits::epsilon()</a>,
+  *     it returns a \a Real instead of a \a T.
   * \li A dummy_precision() function returning a weak epsilon value. It is mainly used as a default
   *     value by the fuzzy comparison operators.
   * \li highest() and lowest() functions returning the highest and lowest possible values respectively.
+  * \li digits10() function returning the number of decimal digits that can be represented without change. This is
+  *     the analogue of <a href="http://en.cppreference.com/w/cpp/types/numeric_limits/digits10">std::numeric_limits<T>::digits10</a>
+  *     which is used as the default implementation if specialized.
   */
 
 template<typename T> struct GenericNumTraits
@@ -67,22 +104,47 @@ template<typename T> struct GenericNumTraits
                      T
                    >::type NonInteger;
   typedef T Nested;
+  typedef T Literal;
+
+  EIGEN_DEVICE_FUNC
+  static inline Real epsilon()
+  {
+    return numext::numeric_limits<T>::epsilon();
+  }
 
-  static inline Real epsilon() { return std::numeric_limits<T>::epsilon(); }
+  EIGEN_DEVICE_FUNC
+  static inline int digits10()
+  {
+    return internal::default_digits10_impl<T>::run();
+  }
+
+  EIGEN_DEVICE_FUNC
   static inline Real dummy_precision()
   {
     // make sure to override this for floating-point types
     return Real(0);
   }
-  static inline T highest() { return (std::numeric_limits<T>::max)(); }
-  static inline T lowest()  { return IsInteger ? (std::numeric_limits<T>::min)() : (-(std::numeric_limits<T>::max)()); }
-  
-#ifdef EIGEN2_SUPPORT
-  enum {
-    HasFloatingPoint = !IsInteger
-  };
-  typedef NonInteger FloatingPoint;
-#endif
+
+
+  EIGEN_DEVICE_FUNC
+  static inline T highest() {
+    return (numext::numeric_limits<T>::max)();
+  }
+
+  EIGEN_DEVICE_FUNC
+  static inline T lowest()  {
+    return IsInteger ? (numext::numeric_limits<T>::min)() : (-(numext::numeric_limits<T>::max)());
+  }
+
+  EIGEN_DEVICE_FUNC
+  static inline T infinity() {
+    return numext::numeric_limits<T>::infinity();
+  }
+
+  EIGEN_DEVICE_FUNC
+  static inline T quiet_NaN() {
+    return numext::numeric_limits<T>::quiet_NaN();
+  }
 };
 
 template<typename T> struct NumTraits : GenericNumTraits<T>
@@ -91,11 +153,13 @@ template<typename T> struct NumTraits : GenericNumTraits<T>
 template<> struct NumTraits<float>
   : GenericNumTraits<float>
 {
+  EIGEN_DEVICE_FUNC
   static inline float dummy_precision() { return 1e-5f; }
 };
 
 template<> struct NumTraits<double> : GenericNumTraits<double>
 {
+  EIGEN_DEVICE_FUNC
   static inline double dummy_precision() { return 1e-12; }
 };
 
@@ -109,6 +173,7 @@ template<typename _Real> struct NumTraits<std::complex<_Real> >
   : GenericNumTraits<std::complex<_Real> >
 {
   typedef _Real Real;
+  typedef typename NumTraits<_Real>::Literal Literal;
   enum {
     IsComplex = 1,
     RequireInitialization = NumTraits<_Real>::RequireInitialization,
@@ -117,8 +182,12 @@ template<typename _Real> struct NumTraits<std::complex<_Real> >
     MulCost = 4 * NumTraits<Real>::MulCost + 2 * NumTraits<Real>::AddCost
   };
 
+  EIGEN_DEVICE_FUNC
   static inline Real epsilon() { return NumTraits<Real>::epsilon(); }
+  EIGEN_DEVICE_FUNC
   static inline Real dummy_precision() { return NumTraits<Real>::dummy_precision(); }
+  EIGEN_DEVICE_FUNC
+  static inline int digits10() { return NumTraits<Real>::digits10(); }
 };
 
 template<typename Scalar, int Rows, int Cols, int Options, int MaxRows, int MaxCols>
@@ -130,21 +199,48 @@ struct NumTraits<Array<Scalar, Rows, Cols, Options, MaxRows, MaxCols> >
   typedef typename NumTraits<Scalar>::NonInteger NonIntegerScalar;
   typedef Array<NonIntegerScalar, Rows, Cols, Options, MaxRows, MaxCols> NonInteger;
   typedef ArrayType & Nested;
-  
+  typedef typename NumTraits<Scalar>::Literal Literal;
+
   enum {
     IsComplex = NumTraits<Scalar>::IsComplex,
     IsInteger = NumTraits<Scalar>::IsInteger,
     IsSigned  = NumTraits<Scalar>::IsSigned,
     RequireInitialization = 1,
-    ReadCost = ArrayType::SizeAtCompileTime==Dynamic ? Dynamic : ArrayType::SizeAtCompileTime * NumTraits<Scalar>::ReadCost,
-    AddCost  = ArrayType::SizeAtCompileTime==Dynamic ? Dynamic : ArrayType::SizeAtCompileTime * NumTraits<Scalar>::AddCost,
-    MulCost  = ArrayType::SizeAtCompileTime==Dynamic ? Dynamic : ArrayType::SizeAtCompileTime * NumTraits<Scalar>::MulCost
+    ReadCost = ArrayType::SizeAtCompileTime==Dynamic ? HugeCost : ArrayType::SizeAtCompileTime * NumTraits<Scalar>::ReadCost,
+    AddCost  = ArrayType::SizeAtCompileTime==Dynamic ? HugeCost : ArrayType::SizeAtCompileTime * NumTraits<Scalar>::AddCost,
+    MulCost  = ArrayType::SizeAtCompileTime==Dynamic ? HugeCost : ArrayType::SizeAtCompileTime * NumTraits<Scalar>::MulCost
   };
-  
+
+  EIGEN_DEVICE_FUNC
   static inline RealScalar epsilon() { return NumTraits<RealScalar>::epsilon(); }
+  EIGEN_DEVICE_FUNC
   static inline RealScalar dummy_precision() { return NumTraits<RealScalar>::dummy_precision(); }
 };
 
+template<> struct NumTraits<std::string>
+  : GenericNumTraits<std::string>
+{
+  enum {
+    RequireInitialization = 1,
+    ReadCost = HugeCost,
+    AddCost  = HugeCost,
+    MulCost  = HugeCost
+  };
+
+  static inline int digits10() { return 0; }
+
+private:
+  static inline std::string epsilon();
+  static inline std::string dummy_precision();
+  static inline std::string lowest();
+  static inline std::string highest();
+  static inline std::string infinity();
+  static inline std::string quiet_NaN();
+};
+
+// Empty specialization for void to allow template specialization based on NumTraits<T>::Real with T==void and SFINAE.
+template<> struct NumTraits<void> {};
+
 } // end namespace Eigen
 
 #endif // EIGEN_NUMTRAITS_H
diff --git a/eigen/Eigen/src/Core/PermutationMatrix.h b/eigen/Eigen/src/Core/PermutationMatrix.h
index bda79fa..b1fb455 100644
--- a/eigen/Eigen/src/Core/PermutationMatrix.h
+++ b/eigen/Eigen/src/Core/PermutationMatrix.h
@@ -2,7 +2,7 @@
 // for linear algebra.
 //
 // Copyright (C) 2009 Benoit Jacob <jacob.benoit.1@gmail.com>
-// Copyright (C) 2009-2011 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2009-2015 Gael Guennebaud <gael.guennebaud@inria.fr>
 //
 // This Source Code Form is subject to the terms of the Mozilla
 // Public License v. 2.0. If a copy of the MPL was not distributed
@@ -13,14 +13,18 @@
 
 namespace Eigen { 
 
-template<int RowCol,typename IndicesType,typename MatrixType, typename StorageKind> class PermutedImpl;
+namespace internal {
+
+enum PermPermProduct_t {PermPermProduct};
+
+} // end namespace internal
 
 /** \class PermutationBase
   * \ingroup Core_Module
   *
   * \brief Base class for permutations
   *
-  * \param Derived the derived class
+  * \tparam Derived the derived class
   *
   * This class is the base class for all expressions representing a permutation matrix,
   * internally stored as a vector of integers.
@@ -38,17 +42,6 @@ template<int RowCol,typename IndicesType,typename MatrixType, typename StorageKi
   *
   * \sa class PermutationMatrix, class PermutationWrapper
   */
-
-namespace internal {
-
-template<typename PermutationType, typename MatrixType, int Side, bool Transposed=false>
-struct permut_matrix_product_retval;
-template<typename PermutationType, typename MatrixType, int Side, bool Transposed=false>
-struct permut_sparsematrix_product_retval;
-enum PermPermProduct_t {PermPermProduct};
-
-} // end namespace internal
-
 template<typename Derived>
 class PermutationBase : public EigenBase<Derived>
 {
@@ -60,19 +53,20 @@ class PermutationBase : public EigenBase<Derived>
     typedef typename Traits::IndicesType IndicesType;
     enum {
       Flags = Traits::Flags,
-      CoeffReadCost = Traits::CoeffReadCost,
       RowsAtCompileTime = Traits::RowsAtCompileTime,
       ColsAtCompileTime = Traits::ColsAtCompileTime,
       MaxRowsAtCompileTime = Traits::MaxRowsAtCompileTime,
       MaxColsAtCompileTime = Traits::MaxColsAtCompileTime
     };
-    typedef typename Traits::Scalar Scalar;
-    typedef typename Traits::Index Index;
-    typedef Matrix<Scalar,RowsAtCompileTime,ColsAtCompileTime,0,MaxRowsAtCompileTime,MaxColsAtCompileTime>
+    typedef typename Traits::StorageIndex StorageIndex;
+    typedef Matrix<StorageIndex,RowsAtCompileTime,ColsAtCompileTime,0,MaxRowsAtCompileTime,MaxColsAtCompileTime>
             DenseMatrixType;
-    typedef PermutationMatrix<IndicesType::SizeAtCompileTime,IndicesType::MaxSizeAtCompileTime,Index>
+    typedef PermutationMatrix<IndicesType::SizeAtCompileTime,IndicesType::MaxSizeAtCompileTime,StorageIndex>
             PlainPermutationType;
+    typedef PlainPermutationType PlainObject;
     using Base::derived;
+    typedef Inverse<Derived> InverseReturnType;
+    typedef void Scalar;
     #endif
 
     /** Copies the other permutation into *this */
@@ -118,7 +112,7 @@ class PermutationBase : public EigenBase<Derived>
     void evalTo(MatrixBase<DenseDerived>& other) const
     {
       other.setZero();
-      for (int i=0; i<rows();++i)
+      for (Index i=0; i<rows(); ++i)
         other.coeffRef(indices().coeff(i),i) = typename DenseDerived::Scalar(1);
     }
     #endif
@@ -147,7 +141,8 @@ class PermutationBase : public EigenBase<Derived>
     /** Sets *this to be the identity permutation matrix */
     void setIdentity()
     {
-      for(Index i = 0; i < size(); ++i)
+      StorageIndex n = StorageIndex(size());
+      for(StorageIndex i = 0; i < n; ++i)
         indices().coeffRef(i) = i;
     }
 
@@ -163,18 +158,18 @@ class PermutationBase : public EigenBase<Derived>
       *
       * \returns a reference to *this.
       *
-      * \warning This is much slower than applyTranspositionOnTheRight(int,int):
+      * \warning This is much slower than applyTranspositionOnTheRight(Index,Index):
       * this has linear complexity and requires a lot of branching.
       *
-      * \sa applyTranspositionOnTheRight(int,int)
+      * \sa applyTranspositionOnTheRight(Index,Index)
       */
     Derived& applyTranspositionOnTheLeft(Index i, Index j)
     {
       eigen_assert(i>=0 && j>=0 && i<size() && j<size());
       for(Index k = 0; k < size(); ++k)
       {
-        if(indices().coeff(k) == i) indices().coeffRef(k) = j;
-        else if(indices().coeff(k) == j) indices().coeffRef(k) = i;
+        if(indices().coeff(k) == i) indices().coeffRef(k) = StorageIndex(j);
+        else if(indices().coeff(k) == j) indices().coeffRef(k) = StorageIndex(i);
       }
       return derived();
     }
@@ -185,7 +180,7 @@ class PermutationBase : public EigenBase<Derived>
       *
       * This is a fast operation, it only consists in swapping two indices.
       *
-      * \sa applyTranspositionOnTheLeft(int,int)
+      * \sa applyTranspositionOnTheLeft(Index,Index)
       */
     Derived& applyTranspositionOnTheRight(Index i, Index j)
     {
@@ -196,16 +191,16 @@ class PermutationBase : public EigenBase<Derived>
 
     /** \returns the inverse permutation matrix.
       *
-      * \note \note_try_to_help_rvo
+      * \note \blank \note_try_to_help_rvo
       */
-    inline Transpose<PermutationBase> inverse() const
-    { return derived(); }
+    inline InverseReturnType inverse() const
+    { return InverseReturnType(derived()); }
     /** \returns the tranpose permutation matrix.
       *
-      * \note \note_try_to_help_rvo
+      * \note \blank \note_try_to_help_rvo
       */
-    inline Transpose<PermutationBase> transpose() const
-    { return derived(); }
+    inline InverseReturnType transpose() const
+    { return InverseReturnType(derived()); }
 
     /**** multiplication helpers to hopefully get RVO ****/
 
@@ -215,13 +210,13 @@ class PermutationBase : public EigenBase<Derived>
     template<typename OtherDerived>
     void assignTranspose(const PermutationBase<OtherDerived>& other)
     {
-      for (int i=0; i<rows();++i) indices().coeffRef(other.indices().coeff(i)) = i;
+      for (Index i=0; i<rows();++i) indices().coeffRef(other.indices().coeff(i)) = i;
     }
     template<typename Lhs,typename Rhs>
     void assignProduct(const Lhs& lhs, const Rhs& rhs)
     {
       eigen_assert(lhs.cols() == rhs.rows());
-      for (int i=0; i<rows();++i) indices().coeffRef(i) = lhs.indices().coeff(rhs.indices().coeff(i));
+      for (Index i=0; i<rows();++i) indices().coeffRef(i) = lhs.indices().coeff(rhs.indices().coeff(i));
     }
 #endif
 
@@ -229,7 +224,7 @@ class PermutationBase : public EigenBase<Derived>
 
     /** \returns the product permutation matrix.
       *
-      * \note \note_try_to_help_rvo
+      * \note \blank \note_try_to_help_rvo
       */
     template<typename Other>
     inline PlainPermutationType operator*(const PermutationBase<Other>& other) const
@@ -237,18 +232,18 @@ class PermutationBase : public EigenBase<Derived>
 
     /** \returns the product of a permutation with another inverse permutation.
       *
-      * \note \note_try_to_help_rvo
+      * \note \blank \note_try_to_help_rvo
       */
     template<typename Other>
-    inline PlainPermutationType operator*(const Transpose<PermutationBase<Other> >& other) const
+    inline PlainPermutationType operator*(const InverseImpl<Other,PermutationStorage>& other) const
     { return PlainPermutationType(internal::PermPermProduct, *this, other.eval()); }
 
     /** \returns the product of an inverse permutation with another permutation.
       *
-      * \note \note_try_to_help_rvo
+      * \note \blank \note_try_to_help_rvo
       */
     template<typename Other> friend
-    inline PlainPermutationType operator*(const Transpose<PermutationBase<Other> >& other, const PermutationBase& perm)
+    inline PlainPermutationType operator*(const InverseImpl<Other, PermutationStorage>& other, const PermutationBase& perm)
     { return PlainPermutationType(internal::PermPermProduct, other.eval(), perm); }
     
     /** \returns the determinant of the permutation matrix, which is either 1 or -1 depending on the parity of the permutation.
@@ -284,39 +279,43 @@ class PermutationBase : public EigenBase<Derived>
 
 };
 
+namespace internal {
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex>
+struct traits<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndex> >
+ : traits<Matrix<_StorageIndex,SizeAtCompileTime,SizeAtCompileTime,0,MaxSizeAtCompileTime,MaxSizeAtCompileTime> >
+{
+  typedef PermutationStorage StorageKind;
+  typedef Matrix<_StorageIndex, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1> IndicesType;
+  typedef _StorageIndex StorageIndex;
+  typedef void Scalar;
+};
+}
+
 /** \class PermutationMatrix
   * \ingroup Core_Module
   *
   * \brief Permutation matrix
   *
-  * \param SizeAtCompileTime the number of rows/cols, or Dynamic
-  * \param MaxSizeAtCompileTime the maximum number of rows/cols, or Dynamic. This optional parameter defaults to SizeAtCompileTime. Most of the time, you should not have to specify it.
-  * \param IndexType the interger type of the indices
+  * \tparam SizeAtCompileTime the number of rows/cols, or Dynamic
+  * \tparam MaxSizeAtCompileTime the maximum number of rows/cols, or Dynamic. This optional parameter defaults to SizeAtCompileTime. Most of the time, you should not have to specify it.
+  * \tparam _StorageIndex the integer type of the indices
   *
   * This class represents a permutation matrix, internally stored as a vector of integers.
   *
   * \sa class PermutationBase, class PermutationWrapper, class DiagonalMatrix
   */
-
-namespace internal {
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename IndexType>
-struct traits<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, IndexType> >
- : traits<Matrix<IndexType,SizeAtCompileTime,SizeAtCompileTime,0,MaxSizeAtCompileTime,MaxSizeAtCompileTime> >
-{
-  typedef IndexType Index;
-  typedef Matrix<IndexType, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1> IndicesType;
-};
-}
-
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename IndexType>
-class PermutationMatrix : public PermutationBase<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, IndexType> >
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex>
+class PermutationMatrix : public PermutationBase<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndex> >
 {
     typedef PermutationBase<PermutationMatrix> Base;
     typedef internal::traits<PermutationMatrix> Traits;
   public:
 
+    typedef const PermutationMatrix& Nested;
+
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     typedef typename Traits::IndicesType IndicesType;
+    typedef typename Traits::StorageIndex StorageIndex;
     #endif
 
     inline PermutationMatrix()
@@ -324,8 +323,10 @@ class PermutationMatrix : public PermutationBase<PermutationMatrix<SizeAtCompile
 
     /** Constructs an uninitialized permutation matrix of given size.
       */
-    inline PermutationMatrix(int size) : m_indices(size)
-    {}
+    explicit inline PermutationMatrix(Index size) : m_indices(size)
+    {
+      eigen_internal_assert(size <= NumTraits<StorageIndex>::highest());
+    }
 
     /** Copy constructor. */
     template<typename OtherDerived>
@@ -346,7 +347,7 @@ class PermutationMatrix : public PermutationBase<PermutationMatrix<SizeAtCompile
       * array's size.
       */
     template<typename Other>
-    explicit inline PermutationMatrix(const MatrixBase<Other>& a_indices) : m_indices(a_indices)
+    explicit inline PermutationMatrix(const MatrixBase<Other>& indices) : m_indices(indices)
     {}
 
     /** Convert the Transpositions \a tr to a permutation matrix */
@@ -393,10 +394,13 @@ class PermutationMatrix : public PermutationBase<PermutationMatrix<SizeAtCompile
 
 #ifndef EIGEN_PARSED_BY_DOXYGEN
     template<typename Other>
-    PermutationMatrix(const Transpose<PermutationBase<Other> >& other)
-      : m_indices(other.nestedPermutation().size())
+    PermutationMatrix(const InverseImpl<Other,PermutationStorage>& other)
+      : m_indices(other.derived().nestedExpression().size())
     {
-      for (int i=0; i<m_indices.size();++i) m_indices.coeffRef(other.nestedPermutation().indices().coeff(i)) = i;
+      eigen_internal_assert(m_indices.size() <= NumTraits<StorageIndex>::highest());
+      StorageIndex end = StorageIndex(m_indices.size());
+      for (StorageIndex i=0; i<end;++i)
+        m_indices.coeffRef(other.derived().nestedExpression().indices().coeff(i)) = i;
     }
     template<typename Lhs,typename Rhs>
     PermutationMatrix(internal::PermPermProduct_t, const Lhs& lhs, const Rhs& rhs)
@@ -413,18 +417,20 @@ class PermutationMatrix : public PermutationBase<PermutationMatrix<SizeAtCompile
 
 
 namespace internal {
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename IndexType, int _PacketAccess>
-struct traits<Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, IndexType>,_PacketAccess> >
- : traits<Matrix<IndexType,SizeAtCompileTime,SizeAtCompileTime,0,MaxSizeAtCompileTime,MaxSizeAtCompileTime> >
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex, int _PacketAccess>
+struct traits<Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndex>,_PacketAccess> >
+ : traits<Matrix<_StorageIndex,SizeAtCompileTime,SizeAtCompileTime,0,MaxSizeAtCompileTime,MaxSizeAtCompileTime> >
 {
-  typedef IndexType Index;
-  typedef Map<const Matrix<IndexType, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1>, _PacketAccess> IndicesType;
+  typedef PermutationStorage StorageKind;
+  typedef Map<const Matrix<_StorageIndex, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1>, _PacketAccess> IndicesType;
+  typedef _StorageIndex StorageIndex;
+  typedef void Scalar;
 };
 }
 
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename IndexType, int _PacketAccess>
-class Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, IndexType>,_PacketAccess>
-  : public PermutationBase<Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, IndexType>,_PacketAccess> >
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex, int _PacketAccess>
+class Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndex>,_PacketAccess>
+  : public PermutationBase<Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndex>,_PacketAccess> >
 {
     typedef PermutationBase<Map> Base;
     typedef internal::traits<Map> Traits;
@@ -432,14 +438,14 @@ class Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, IndexType>,
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     typedef typename Traits::IndicesType IndicesType;
-    typedef typename IndicesType::Scalar Index;
+    typedef typename IndicesType::Scalar StorageIndex;
     #endif
 
-    inline Map(const Index* indicesPtr)
+    inline Map(const StorageIndex* indicesPtr)
       : m_indices(indicesPtr)
     {}
 
-    inline Map(const Index* indicesPtr, Index size)
+    inline Map(const StorageIndex* indicesPtr, Index size)
       : m_indices(indicesPtr,size)
     {}
 
@@ -474,40 +480,36 @@ class Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, IndexType>,
     IndicesType m_indices;
 };
 
-/** \class PermutationWrapper
-  * \ingroup Core_Module
-  *
-  * \brief Class to view a vector of integers as a permutation matrix
-  *
-  * \param _IndicesType the type of the vector of integer (can be any compatible expression)
-  *
-  * This class allows to view any vector expression of integers as a permutation matrix.
-  *
-  * \sa class PermutationBase, class PermutationMatrix
-  */
-
-struct PermutationStorage {};
-
 template<typename _IndicesType> class TranspositionsWrapper;
 namespace internal {
 template<typename _IndicesType>
 struct traits<PermutationWrapper<_IndicesType> >
 {
   typedef PermutationStorage StorageKind;
-  typedef typename _IndicesType::Scalar Scalar;
-  typedef typename _IndicesType::Scalar Index;
+  typedef void Scalar;
+  typedef typename _IndicesType::Scalar StorageIndex;
   typedef _IndicesType IndicesType;
   enum {
     RowsAtCompileTime = _IndicesType::SizeAtCompileTime,
     ColsAtCompileTime = _IndicesType::SizeAtCompileTime,
-    MaxRowsAtCompileTime = IndicesType::MaxRowsAtCompileTime,
-    MaxColsAtCompileTime = IndicesType::MaxColsAtCompileTime,
-    Flags = 0,
-    CoeffReadCost = _IndicesType::CoeffReadCost
+    MaxRowsAtCompileTime = IndicesType::MaxSizeAtCompileTime,
+    MaxColsAtCompileTime = IndicesType::MaxSizeAtCompileTime,
+    Flags = 0
   };
 };
 }
 
+/** \class PermutationWrapper
+  * \ingroup Core_Module
+  *
+  * \brief Class to view a vector of integers as a permutation matrix
+  *
+  * \tparam _IndicesType the type of the vector of integer (can be any compatible expression)
+  *
+  * This class allows to view any vector expression of integers as a permutation matrix.
+  *
+  * \sa class PermutationBase, class PermutationMatrix
+  */
 template<typename _IndicesType>
 class PermutationWrapper : public PermutationBase<PermutationWrapper<_IndicesType> >
 {
@@ -519,8 +521,8 @@ class PermutationWrapper : public PermutationBase<PermutationWrapper<_IndicesTyp
     typedef typename Traits::IndicesType IndicesType;
     #endif
 
-    inline PermutationWrapper(const IndicesType& a_indices)
-      : m_indices(a_indices)
+    inline PermutationWrapper(const IndicesType& indices)
+      : m_indices(indices)
     {}
 
     /** const version of indices(). */
@@ -532,183 +534,86 @@ class PermutationWrapper : public PermutationBase<PermutationWrapper<_IndicesTyp
     typename IndicesType::Nested m_indices;
 };
 
+
 /** \returns the matrix with the permutation applied to the columns.
   */
-template<typename Derived, typename PermutationDerived>
-inline const internal::permut_matrix_product_retval<PermutationDerived, Derived, OnTheRight>
-operator*(const MatrixBase<Derived>& matrix,
-          const PermutationBase<PermutationDerived> &permutation)
+template<typename MatrixDerived, typename PermutationDerived>
+EIGEN_DEVICE_FUNC
+const Product<MatrixDerived, PermutationDerived, AliasFreeProduct>
+operator*(const MatrixBase<MatrixDerived> &matrix,
+          const PermutationBase<PermutationDerived>& permutation)
 {
-  return internal::permut_matrix_product_retval
-           <PermutationDerived, Derived, OnTheRight>
-           (permutation.derived(), matrix.derived());
+  return Product<MatrixDerived, PermutationDerived, AliasFreeProduct>
+            (matrix.derived(), permutation.derived());
 }
 
 /** \returns the matrix with the permutation applied to the rows.
   */
-template<typename Derived, typename PermutationDerived>
-inline const internal::permut_matrix_product_retval
-               <PermutationDerived, Derived, OnTheLeft>
+template<typename PermutationDerived, typename MatrixDerived>
+EIGEN_DEVICE_FUNC
+const Product<PermutationDerived, MatrixDerived, AliasFreeProduct>
 operator*(const PermutationBase<PermutationDerived> &permutation,
-          const MatrixBase<Derived>& matrix)
+          const MatrixBase<MatrixDerived>& matrix)
 {
-  return internal::permut_matrix_product_retval
-           <PermutationDerived, Derived, OnTheLeft>
-           (permutation.derived(), matrix.derived());
+  return Product<PermutationDerived, MatrixDerived, AliasFreeProduct>
+            (permutation.derived(), matrix.derived());
 }
 
-namespace internal {
 
-template<typename PermutationType, typename MatrixType, int Side, bool Transposed>
-struct traits<permut_matrix_product_retval<PermutationType, MatrixType, Side, Transposed> >
+template<typename PermutationType>
+class InverseImpl<PermutationType, PermutationStorage>
+  : public EigenBase<Inverse<PermutationType> >
 {
-  typedef typename MatrixType::PlainObject ReturnType;
-};
-
-template<typename PermutationType, typename MatrixType, int Side, bool Transposed>
-struct permut_matrix_product_retval
- : public ReturnByValue<permut_matrix_product_retval<PermutationType, MatrixType, Side, Transposed> >
-{
-    typedef typename remove_all<typename MatrixType::Nested>::type MatrixTypeNestedCleaned;
-    typedef typename MatrixType::Index Index;
-
-    permut_matrix_product_retval(const PermutationType& perm, const MatrixType& matrix)
-      : m_permutation(perm), m_matrix(matrix)
-    {}
-
-    inline Index rows() const { return m_matrix.rows(); }
-    inline Index cols() const { return m_matrix.cols(); }
-
-    template<typename Dest> inline void evalTo(Dest& dst) const
-    {
-      const Index n = Side==OnTheLeft ? rows() : cols();
-      // FIXME we need an is_same for expression that is not sensitive to constness. For instance
-      // is_same_xpr<Block<const Matrix>, Block<Matrix> >::value should be true.
-      const typename Dest::Scalar *dst_data = internal::extract_data(dst);
-      if(    is_same<MatrixTypeNestedCleaned,Dest>::value
-          && blas_traits<MatrixTypeNestedCleaned>::HasUsableDirectAccess
-          && blas_traits<Dest>::HasUsableDirectAccess
-          && dst_data!=0 && dst_data == extract_data(m_matrix))
-      {
-        // apply the permutation inplace
-        Matrix<bool,PermutationType::RowsAtCompileTime,1,0,PermutationType::MaxRowsAtCompileTime> mask(m_permutation.size());
-        mask.fill(false);
-        Index r = 0;
-        while(r < m_permutation.size())
-        {
-          // search for the next seed
-          while(r<m_permutation.size() && mask[r]) r++;
-          if(r>=m_permutation.size())
-            break;
-          // we got one, let's follow it until we are back to the seed
-          Index k0 = r++;
-          Index kPrev = k0;
-          mask.coeffRef(k0) = true;
-          for(Index k=m_permutation.indices().coeff(k0); k!=k0; k=m_permutation.indices().coeff(k))
-          {
-                  Block<Dest, Side==OnTheLeft ? 1 : Dest::RowsAtCompileTime, Side==OnTheRight ? 1 : Dest::ColsAtCompileTime>(dst, k)
-            .swap(Block<Dest, Side==OnTheLeft ? 1 : Dest::RowsAtCompileTime, Side==OnTheRight ? 1 : Dest::ColsAtCompileTime>
-                       (dst,((Side==OnTheLeft) ^ Transposed) ? k0 : kPrev));
-
-            mask.coeffRef(k) = true;
-            kPrev = k;
-          }
-        }
-      }
-      else
-      {
-        for(int i = 0; i < n; ++i)
-        {
-          Block<Dest, Side==OnTheLeft ? 1 : Dest::RowsAtCompileTime, Side==OnTheRight ? 1 : Dest::ColsAtCompileTime>
-               (dst, ((Side==OnTheLeft) ^ Transposed) ? m_permutation.indices().coeff(i) : i)
-
-          =
-
-          Block<const MatrixTypeNestedCleaned,Side==OnTheLeft ? 1 : MatrixType::RowsAtCompileTime,Side==OnTheRight ? 1 : MatrixType::ColsAtCompileTime>
-               (m_matrix, ((Side==OnTheRight) ^ Transposed) ? m_permutation.indices().coeff(i) : i);
-        }
-      }
-    }
-
-  protected:
-    const PermutationType& m_permutation;
-    typename MatrixType::Nested m_matrix;
-};
-
-/* Template partial specialization for transposed/inverse permutations */
-
-template<typename Derived>
-struct traits<Transpose<PermutationBase<Derived> > >
- : traits<Derived>
-{};
-
-} // end namespace internal
-
-template<typename Derived>
-class Transpose<PermutationBase<Derived> >
-  : public EigenBase<Transpose<PermutationBase<Derived> > >
-{
-    typedef Derived PermutationType;
-    typedef typename PermutationType::IndicesType IndicesType;
     typedef typename PermutationType::PlainPermutationType PlainPermutationType;
+    typedef internal::traits<PermutationType> PermTraits;
+  protected:
+    InverseImpl() {}
   public:
+    typedef Inverse<PermutationType> InverseType;
+    using EigenBase<Inverse<PermutationType> >::derived;
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
-    typedef internal::traits<PermutationType> Traits;
-    typedef typename Derived::DenseMatrixType DenseMatrixType;
+    typedef typename PermutationType::DenseMatrixType DenseMatrixType;
     enum {
-      Flags = Traits::Flags,
-      CoeffReadCost = Traits::CoeffReadCost,
-      RowsAtCompileTime = Traits::RowsAtCompileTime,
-      ColsAtCompileTime = Traits::ColsAtCompileTime,
-      MaxRowsAtCompileTime = Traits::MaxRowsAtCompileTime,
-      MaxColsAtCompileTime = Traits::MaxColsAtCompileTime
+      RowsAtCompileTime = PermTraits::RowsAtCompileTime,
+      ColsAtCompileTime = PermTraits::ColsAtCompileTime,
+      MaxRowsAtCompileTime = PermTraits::MaxRowsAtCompileTime,
+      MaxColsAtCompileTime = PermTraits::MaxColsAtCompileTime
     };
-    typedef typename Traits::Scalar Scalar;
     #endif
 
-    Transpose(const PermutationType& p) : m_permutation(p) {}
-
-    inline int rows() const { return m_permutation.rows(); }
-    inline int cols() const { return m_permutation.cols(); }
-
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     template<typename DenseDerived>
     void evalTo(MatrixBase<DenseDerived>& other) const
     {
       other.setZero();
-      for (int i=0; i<rows();++i)
-        other.coeffRef(i, m_permutation.indices().coeff(i)) = typename DenseDerived::Scalar(1);
+      for (Index i=0; i<derived().rows();++i)
+        other.coeffRef(i, derived().nestedExpression().indices().coeff(i)) = typename DenseDerived::Scalar(1);
     }
     #endif
 
     /** \return the equivalent permutation matrix */
-    PlainPermutationType eval() const { return *this; }
+    PlainPermutationType eval() const { return derived(); }
 
-    DenseMatrixType toDenseMatrix() const { return *this; }
+    DenseMatrixType toDenseMatrix() const { return derived(); }
 
     /** \returns the matrix with the inverse permutation applied to the columns.
       */
     template<typename OtherDerived> friend
-    inline const internal::permut_matrix_product_retval<PermutationType, OtherDerived, OnTheRight, true>
-    operator*(const MatrixBase<OtherDerived>& matrix, const Transpose& trPerm)
+    const Product<OtherDerived, InverseType, AliasFreeProduct>
+    operator*(const MatrixBase<OtherDerived>& matrix, const InverseType& trPerm)
     {
-      return internal::permut_matrix_product_retval<PermutationType, OtherDerived, OnTheRight, true>(trPerm.m_permutation, matrix.derived());
+      return Product<OtherDerived, InverseType, AliasFreeProduct>(matrix.derived(), trPerm.derived());
     }
 
     /** \returns the matrix with the inverse permutation applied to the rows.
       */
     template<typename OtherDerived>
-    inline const internal::permut_matrix_product_retval<PermutationType, OtherDerived, OnTheLeft, true>
+    const Product<InverseType, OtherDerived, AliasFreeProduct>
     operator*(const MatrixBase<OtherDerived>& matrix) const
     {
-      return internal::permut_matrix_product_retval<PermutationType, OtherDerived, OnTheLeft, true>(m_permutation, matrix.derived());
+      return Product<InverseType, OtherDerived, AliasFreeProduct>(derived(), matrix.derived());
     }
-
-    const PermutationType& nestedPermutation() const { return m_permutation; }
-
-  protected:
-    const PermutationType& m_permutation;
 };
 
 template<typename Derived>
@@ -717,6 +622,12 @@ const PermutationWrapper<const Derived> MatrixBase<Derived>::asPermutation() con
   return derived();
 }
 
+namespace internal {
+
+template<> struct AssignmentKind<DenseShape,PermutationShape> { typedef EigenBase2EigenBase Kind; };
+
+} // end namespace internal
+
 } // end namespace Eigen
 
 #endif // EIGEN_PERMUTATIONMATRIX_H
diff --git a/eigen/Eigen/src/Core/PlainObjectBase.h b/eigen/Eigen/src/Core/PlainObjectBase.h
index 9f71956..77f4f60 100644
--- a/eigen/Eigen/src/Core/PlainObjectBase.h
+++ b/eigen/Eigen/src/Core/PlainObjectBase.h
@@ -28,6 +28,7 @@ namespace internal {
 
 template<int MaxSizeAtCompileTime> struct check_rows_cols_for_overflow {
   template<typename Index>
+  EIGEN_DEVICE_FUNC
   static EIGEN_ALWAYS_INLINE void run(Index, Index)
   {
   }
@@ -35,11 +36,12 @@ template<int MaxSizeAtCompileTime> struct check_rows_cols_for_overflow {
 
 template<> struct check_rows_cols_for_overflow<Dynamic> {
   template<typename Index>
+  EIGEN_DEVICE_FUNC
   static EIGEN_ALWAYS_INLINE void run(Index rows, Index cols)
   {
     // http://hg.mozilla.org/mozilla-central/file/6c8a909977d3/xpcom/ds/CheckedInt.h#l242
     // we assume Index is signed
-    Index max_index = (size_t(1) << (8 * sizeof(Index) - 1)) - 1; // assume Index is signed
+    Index max_index = (std::size_t(1) << (8 * sizeof(Index) - 1)) - 1; // assume Index is signed
     bool error = (rows == 0 || cols == 0) ? false
                : (rows > max_index / cols);
     if (error)
@@ -56,33 +58,41 @@ template<typename MatrixTypeA, typename MatrixTypeB, bool SwapPointers> struct m
 
 } // end namespace internal
 
-/** \class PlainObjectBase
-  * \brief %Dense storage base class for matrices and arrays.
-  *
-  * This class can be extended with the help of the plugin mechanism described on the page
-  * \ref TopicCustomizingEigen by defining the preprocessor symbol \c EIGEN_PLAINOBJECTBASE_PLUGIN.
-  *
-  * \sa \ref TopicClassHierarchy
-  */
 #ifdef EIGEN_PARSED_BY_DOXYGEN
-namespace internal {
+namespace doxygen {
 
-// this is a warkaround to doxygen not being able to understand the inheritence logic
+// This is a workaround to doxygen not being able to understand the inheritance logic
 // when it is hidden by the dense_xpr_base helper struct.
-template<typename Derived> struct dense_xpr_base_dispatcher_for_doxygen;// : public MatrixBase<Derived> {};
+// Moreover, doxygen fails to include members that are not documented in the declaration body of
+// MatrixBase if we inherits MatrixBase<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >,
+// this is why we simply inherits MatrixBase, though this does not make sense.
+
+/** This class is just a workaround for Doxygen and it does not not actually exist. */
+template<typename Derived> struct dense_xpr_base_dispatcher;
 /** This class is just a workaround for Doxygen and it does not not actually exist. */
 template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
-struct dense_xpr_base_dispatcher_for_doxygen<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >
-    : public MatrixBase<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> > {};
+struct dense_xpr_base_dispatcher<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >
+    : public MatrixBase {};
 /** This class is just a workaround for Doxygen and it does not not actually exist. */
 template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
-struct dense_xpr_base_dispatcher_for_doxygen<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >
-    : public ArrayBase<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> > {};
+struct dense_xpr_base_dispatcher<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >
+    : public ArrayBase {};
 
-} // namespace internal
+} // namespace doxygen
 
+/** \class PlainObjectBase
+  * \ingroup Core_Module
+  * \brief %Dense storage base class for matrices and arrays.
+  *
+  * This class can be extended with the help of the plugin mechanism described on the page
+  * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_PLAINOBJECTBASE_PLUGIN.
+  *
+  * \tparam Derived is the derived type, e.g., a Matrix or Array
+  *
+  * \sa \ref TopicClassHierarchy
+  */
 template<typename Derived>
-class PlainObjectBase : public internal::dense_xpr_base_dispatcher_for_doxygen<Derived>
+class PlainObjectBase : public doxygen::dense_xpr_base_dispatcher<Derived>
 #else
 template<typename Derived>
 class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
@@ -93,8 +103,8 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     typedef typename internal::dense_xpr_base<Derived>::type Base;
 
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
-    typedef typename internal::traits<Derived>::Index Index;
     typedef typename internal::traits<Derived>::Scalar Scalar;
+    
     typedef typename internal::packet_traits<Scalar>::type PacketScalar;
     typedef typename NumTraits<Scalar>::Real RealScalar;
     typedef Derived DenseType;
@@ -113,28 +123,40 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     typedef Eigen::Map<Derived, Unaligned>  MapType;
     friend  class Eigen::Map<const Derived, Unaligned>;
     typedef const Eigen::Map<const Derived, Unaligned> ConstMapType;
-    friend  class Eigen::Map<Derived, Aligned>;
-    typedef Eigen::Map<Derived, Aligned> AlignedMapType;
-    friend  class Eigen::Map<const Derived, Aligned>;
-    typedef const Eigen::Map<const Derived, Aligned> ConstAlignedMapType;
+#if EIGEN_MAX_ALIGN_BYTES>0
+    // for EIGEN_MAX_ALIGN_BYTES==0, AlignedMax==Unaligned, and many compilers generate warnings for friend-ing a class twice.
+    friend  class Eigen::Map<Derived, AlignedMax>;
+    friend  class Eigen::Map<const Derived, AlignedMax>;
+#endif
+    typedef Eigen::Map<Derived, AlignedMax> AlignedMapType;
+    typedef const Eigen::Map<const Derived, AlignedMax> ConstAlignedMapType;
     template<typename StrideType> struct StridedMapType { typedef Eigen::Map<Derived, Unaligned, StrideType> type; };
     template<typename StrideType> struct StridedConstMapType { typedef Eigen::Map<const Derived, Unaligned, StrideType> type; };
-    template<typename StrideType> struct StridedAlignedMapType { typedef Eigen::Map<Derived, Aligned, StrideType> type; };
-    template<typename StrideType> struct StridedConstAlignedMapType { typedef Eigen::Map<const Derived, Aligned, StrideType> type; };
+    template<typename StrideType> struct StridedAlignedMapType { typedef Eigen::Map<Derived, AlignedMax, StrideType> type; };
+    template<typename StrideType> struct StridedConstAlignedMapType { typedef Eigen::Map<const Derived, AlignedMax, StrideType> type; };
 
   protected:
     DenseStorage<Scalar, Base::MaxSizeAtCompileTime, Base::RowsAtCompileTime, Base::ColsAtCompileTime, Options> m_storage;
 
   public:
-    enum { NeedsToAlign = SizeAtCompileTime != Dynamic && (internal::traits<Derived>::Flags & AlignedBit) != 0 };
+    enum { NeedsToAlign = (SizeAtCompileTime != Dynamic) && (internal::traits<Derived>::Alignment>0) };
     EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)
 
+    EIGEN_DEVICE_FUNC
     Base& base() { return *static_cast<Base*>(this); }
+    EIGEN_DEVICE_FUNC
     const Base& base() const { return *static_cast<const Base*>(this); }
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Index rows() const { return m_storage.rows(); }
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Index cols() const { return m_storage.cols(); }
 
+    /** This is an overloaded version of DenseCoeffsBase<Derived,ReadOnlyAccessors>::coeff(Index,Index) const
+      * provided to by-pass the creation of an evaluator of the expression, thus saving compilation efforts.
+      *
+      * See DenseCoeffsBase<Derived,ReadOnlyAccessors>::coeff(Index) const for details. */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE const Scalar& coeff(Index rowId, Index colId) const
     {
       if(Flags & RowMajorBit)
@@ -143,11 +165,21 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
         return m_storage.data()[rowId + colId * m_storage.rows()];
     }
 
+    /** This is an overloaded version of DenseCoeffsBase<Derived,ReadOnlyAccessors>::coeff(Index) const
+      * provided to by-pass the creation of an evaluator of the expression, thus saving compilation efforts.
+      *
+      * See DenseCoeffsBase<Derived,ReadOnlyAccessors>::coeff(Index) const for details. */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE const Scalar& coeff(Index index) const
     {
       return m_storage.data()[index];
     }
 
+    /** This is an overloaded version of DenseCoeffsBase<Derived,WriteAccessors>::coeffRef(Index,Index) const
+      * provided to by-pass the creation of an evaluator of the expression, thus saving compilation efforts.
+      *
+      * See DenseCoeffsBase<Derived,WriteAccessors>::coeffRef(Index,Index) const for details. */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Scalar& coeffRef(Index rowId, Index colId)
     {
       if(Flags & RowMajorBit)
@@ -156,11 +188,19 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
         return m_storage.data()[rowId + colId * m_storage.rows()];
     }
 
+    /** This is an overloaded version of DenseCoeffsBase<Derived,WriteAccessors>::coeffRef(Index) const
+      * provided to by-pass the creation of an evaluator of the expression, thus saving compilation efforts.
+      *
+      * See DenseCoeffsBase<Derived,WriteAccessors>::coeffRef(Index) const for details. */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Scalar& coeffRef(Index index)
     {
       return m_storage.data()[index];
     }
 
+    /** This is the const version of coeffRef(Index,Index) which is thus synonym of coeff(Index,Index).
+      * It is provided for convenience. */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE const Scalar& coeffRef(Index rowId, Index colId) const
     {
       if(Flags & RowMajorBit)
@@ -169,6 +209,9 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
         return m_storage.data()[rowId + colId * m_storage.rows()];
     }
 
+    /** This is the const version of coeffRef(Index) which is thus synonym of coeff(Index).
+      * It is provided for convenience. */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE const Scalar& coeffRef(Index index) const
     {
       return m_storage.data()[index];
@@ -209,11 +252,11 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     }
 
     /** \returns a const pointer to the data array of this matrix */
-    EIGEN_STRONG_INLINE const Scalar *data() const
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar *data() const
     { return m_storage.data(); }
 
     /** \returns a pointer to the data array of this matrix */
-    EIGEN_STRONG_INLINE Scalar *data()
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar *data()
     { return m_storage.data(); }
 
     /** Resizes \c *this to a \a rows x \a cols matrix.
@@ -232,22 +275,22 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       *
       * \sa resize(Index) for vectors, resize(NoChange_t, Index), resize(Index, NoChange_t)
       */
-    EIGEN_STRONG_INLINE void resize(Index nbRows, Index nbCols)
-    {
-      eigen_assert(   EIGEN_IMPLIES(RowsAtCompileTime!=Dynamic,nbRows==RowsAtCompileTime)
-                   && EIGEN_IMPLIES(ColsAtCompileTime!=Dynamic,nbCols==ColsAtCompileTime)
-                   && EIGEN_IMPLIES(RowsAtCompileTime==Dynamic && MaxRowsAtCompileTime!=Dynamic,nbRows<=MaxRowsAtCompileTime)
-                   && EIGEN_IMPLIES(ColsAtCompileTime==Dynamic && MaxColsAtCompileTime!=Dynamic,nbCols<=MaxColsAtCompileTime)
-                   && nbRows>=0 && nbCols>=0 && "Invalid sizes when resizing a matrix or array.");
-      internal::check_rows_cols_for_overflow<MaxSizeAtCompileTime>::run(nbRows, nbCols);
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void resize(Index rows, Index cols)
+    {
+      eigen_assert(   EIGEN_IMPLIES(RowsAtCompileTime!=Dynamic,rows==RowsAtCompileTime)
+                   && EIGEN_IMPLIES(ColsAtCompileTime!=Dynamic,cols==ColsAtCompileTime)
+                   && EIGEN_IMPLIES(RowsAtCompileTime==Dynamic && MaxRowsAtCompileTime!=Dynamic,rows<=MaxRowsAtCompileTime)
+                   && EIGEN_IMPLIES(ColsAtCompileTime==Dynamic && MaxColsAtCompileTime!=Dynamic,cols<=MaxColsAtCompileTime)
+                   && rows>=0 && cols>=0 && "Invalid sizes when resizing a matrix or array.");
+      internal::check_rows_cols_for_overflow<MaxSizeAtCompileTime>::run(rows, cols);
       #ifdef EIGEN_INITIALIZE_COEFFS
-        Index size = nbRows*nbCols;
+        Index size = rows*cols;
         bool size_changed = size != this->size();
-        m_storage.resize(size, nbRows, nbCols);
+        m_storage.resize(size, rows, cols);
         if(size_changed) EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
       #else
-        internal::check_rows_cols_for_overflow<MaxSizeAtCompileTime>::run(nbRows, nbCols);
-        m_storage.resize(nbRows*nbCols, nbRows, nbCols);
+        m_storage.resize(rows*cols, rows, cols);
       #endif
     }
 
@@ -262,6 +305,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       *
       * \sa resize(Index,Index), resize(NoChange_t, Index), resize(Index, NoChange_t)
       */
+    EIGEN_DEVICE_FUNC
     inline void resize(Index size)
     {
       EIGEN_STATIC_ASSERT_VECTOR_ONLY(PlainObjectBase)
@@ -286,9 +330,10 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       *
       * \sa resize(Index,Index)
       */
-    inline void resize(NoChange_t, Index nbCols)
+    EIGEN_DEVICE_FUNC
+    inline void resize(NoChange_t, Index cols)
     {
-      resize(rows(), nbCols);
+      resize(rows(), cols);
     }
 
     /** Resizes the matrix, changing only the number of rows. For the parameter of type NoChange_t, just pass the special value \c NoChange
@@ -299,9 +344,10 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       *
       * \sa resize(Index,Index)
       */
-    inline void resize(Index nbRows, NoChange_t)
+    EIGEN_DEVICE_FUNC
+    inline void resize(Index rows, NoChange_t)
     {
-      resize(nbRows, cols());
+      resize(rows, cols());
     }
 
     /** Resizes \c *this to have the same dimensions as \a other.
@@ -312,11 +358,12 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       * remain row-vectors and vectors remain vectors.
       */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC 
     EIGEN_STRONG_INLINE void resizeLike(const EigenBase<OtherDerived>& _other)
     {
       const OtherDerived& other = _other.derived();
-      internal::check_rows_cols_for_overflow<MaxSizeAtCompileTime>::run(Index(other.rows()), Index(other.cols()));
-      const Index othersize = Index(other.rows())*Index(other.cols());
+      internal::check_rows_cols_for_overflow<MaxSizeAtCompileTime>::run(other.rows(), other.cols());
+      const Index othersize = other.rows()*other.cols();
       if(RowsAtCompileTime == 1)
       {
         eigen_assert(other.rows() == 1 || other.cols() == 1);
@@ -339,9 +386,10 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       * Matrices are resized relative to the top-left element. In case values need to be 
       * appended to the matrix they will be uninitialized.
       */
-    EIGEN_STRONG_INLINE void conservativeResize(Index nbRows, Index nbCols)
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void conservativeResize(Index rows, Index cols)
     {
-      internal::conservative_resize_like_impl<Derived>::run(*this, nbRows, nbCols);
+      internal::conservative_resize_like_impl<Derived>::run(*this, rows, cols);
     }
 
     /** Resizes the matrix to \a rows x \a cols while leaving old values untouched.
@@ -351,10 +399,11 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       *
       * In case the matrix is growing, new rows will be uninitialized.
       */
-    EIGEN_STRONG_INLINE void conservativeResize(Index nbRows, NoChange_t)
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void conservativeResize(Index rows, NoChange_t)
     {
       // Note: see the comment in conservativeResize(Index,Index)
-      conservativeResize(nbRows, cols());
+      conservativeResize(rows, cols());
     }
 
     /** Resizes the matrix to \a rows x \a cols while leaving old values untouched.
@@ -364,10 +413,11 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       *
       * In case the matrix is growing, new columns will be uninitialized.
       */
-    EIGEN_STRONG_INLINE void conservativeResize(NoChange_t, Index nbCols)
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void conservativeResize(NoChange_t, Index cols)
     {
       // Note: see the comment in conservativeResize(Index,Index)
-      conservativeResize(rows(), nbCols);
+      conservativeResize(rows(), cols);
     }
 
     /** Resizes the vector to \a size while retaining old values.
@@ -378,6 +428,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       *
       * When values are appended, they will be uninitialized.
       */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE void conservativeResize(Index size)
     {
       internal::conservative_resize_like_impl<Derived>::run(*this, size);
@@ -393,6 +444,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       * appended to the matrix they will copied from \c other.
       */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE void conservativeResizeLike(const DenseBase<OtherDerived>& other)
     {
       internal::conservative_resize_like_impl<Derived,OtherDerived>::run(*this, other);
@@ -401,6 +453,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     /** This is a special case of the templated operator=. Its purpose is to
       * prevent a default operator= from hiding the templated operator=.
       */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Derived& operator=(const PlainObjectBase& other)
     {
       return _set(other);
@@ -408,6 +461,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
 
     /** \sa MatrixBase::lazyAssign() */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Derived& lazyAssign(const DenseBase<OtherDerived>& other)
     {
       _resize_to_match(other);
@@ -415,12 +469,18 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     }
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Derived& operator=(const ReturnByValue<OtherDerived>& func)
     {
       resize(func.rows(), func.cols());
       return Base::operator=(func);
     }
 
+    // Prevent user from trying to instantiate PlainObjectBase objects
+    // by making all its constructor protected. See bug 1074.
+  protected:
+
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE PlainObjectBase() : m_storage()
     {
 //       _check_template_params();
@@ -430,20 +490,23 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
 #ifndef EIGEN_PARSED_BY_DOXYGEN
     // FIXME is it still needed ?
     /** \internal */
-    PlainObjectBase(internal::constructor_without_unaligned_array_assert)
+    EIGEN_DEVICE_FUNC
+    explicit PlainObjectBase(internal::constructor_without_unaligned_array_assert)
       : m_storage(internal::constructor_without_unaligned_array_assert())
     {
 //       _check_template_params(); EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
     }
 #endif
 
-#ifdef EIGEN_HAVE_RVALUE_REFERENCES
-    PlainObjectBase(PlainObjectBase&& other)
+#if EIGEN_HAS_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC
+    PlainObjectBase(PlainObjectBase&& other) EIGEN_NOEXCEPT
       : m_storage( std::move(other.m_storage) )
     {
     }
 
-    PlainObjectBase& operator=(PlainObjectBase&& other)
+    EIGEN_DEVICE_FUNC
+    PlainObjectBase& operator=(PlainObjectBase&& other) EIGEN_NOEXCEPT
     {
       using std::swap;
       swap(m_storage, other.m_storage);
@@ -452,31 +515,56 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
 #endif
 
     /** Copy constructor */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE PlainObjectBase(const PlainObjectBase& other)
-      : m_storage()
+      : Base(), m_storage(other.m_storage) { }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE PlainObjectBase(Index size, Index rows, Index cols)
+      : m_storage(size, rows, cols)
     {
-      _check_template_params();
-      lazyAssign(other);
+//       _check_template_params();
+//       EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
     }
 
+    /** \sa PlainObjectBase::operator=(const EigenBase<OtherDerived>&) */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE PlainObjectBase(const DenseBase<OtherDerived> &other)
       : m_storage()
     {
       _check_template_params();
-      lazyAssign(other);
+      resizeLike(other);
+      _set_noalias(other);
     }
 
-    EIGEN_STRONG_INLINE PlainObjectBase(Index a_size, Index nbRows, Index nbCols)
-      : m_storage(a_size, nbRows, nbCols)
+    /** \sa PlainObjectBase::operator=(const EigenBase<OtherDerived>&) */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE PlainObjectBase(const EigenBase<OtherDerived> &other)
+      : m_storage()
     {
-//       _check_template_params();
-//       EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
+      _check_template_params();
+      resizeLike(other);
+      *this = other.derived();
+    }
+    /** \brief Copy constructor with in-place evaluation */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE PlainObjectBase(const ReturnByValue<OtherDerived>& other)
+    {
+      _check_template_params();
+      // FIXME this does not automatically transpose vectors if necessary
+      resize(other.rows(), other.cols());
+      other.evalTo(this->derived());
     }
 
-    /** \copydoc MatrixBase::operator=(const EigenBase<OtherDerived>&)
+  public:
+
+    /** \brief Copies the generic expression \a other into *this.
+      * \copydetails DenseBase::operator=(const EigenBase<OtherDerived> &other)
       */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC 
     EIGEN_STRONG_INLINE Derived& operator=(const EigenBase<OtherDerived> &other)
     {
       _resize_to_match(other);
@@ -484,16 +572,6 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       return this->derived();
     }
 
-    /** \sa MatrixBase::operator=(const EigenBase<OtherDerived>&) */
-    template<typename OtherDerived>
-    EIGEN_STRONG_INLINE PlainObjectBase(const EigenBase<OtherDerived> &other)
-      : m_storage(Index(other.derived().rows()) * Index(other.derived().cols()), other.derived().rows(), other.derived().cols())
-    {
-      _check_template_params();
-      internal::check_rows_cols_for_overflow<MaxSizeAtCompileTime>::run(other.derived().rows(), other.derived().cols());
-      Base::operator=(other.derived());
-    }
-
     /** \name Map
       * These are convenience functions returning Map objects. The Map() static functions return unaligned Map objects,
       * while the AlignedMap() functions return aligned Map objects and thus should be called only with 16-byte-aligned
@@ -568,16 +646,16 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     //@}
 
     using Base::setConstant;
-    Derived& setConstant(Index size, const Scalar& value);
-    Derived& setConstant(Index rows, Index cols, const Scalar& value);
+    EIGEN_DEVICE_FUNC Derived& setConstant(Index size, const Scalar& val);
+    EIGEN_DEVICE_FUNC Derived& setConstant(Index rows, Index cols, const Scalar& val);
 
     using Base::setZero;
-    Derived& setZero(Index size);
-    Derived& setZero(Index rows, Index cols);
+    EIGEN_DEVICE_FUNC Derived& setZero(Index size);
+    EIGEN_DEVICE_FUNC Derived& setZero(Index rows, Index cols);
 
     using Base::setOnes;
-    Derived& setOnes(Index size);
-    Derived& setOnes(Index rows, Index cols);
+    EIGEN_DEVICE_FUNC Derived& setOnes(Index size);
+    EIGEN_DEVICE_FUNC Derived& setOnes(Index rows, Index cols);
 
     using Base::setRandom;
     Derived& setRandom(Index size);
@@ -596,6 +674,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       * remain row-vectors and vectors remain vectors.
       */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC 
     EIGEN_STRONG_INLINE void _resize_to_match(const EigenBase<OtherDerived>& other)
     {
       #ifdef EIGEN_NO_AUTOMATIC_RESIZING
@@ -603,8 +682,6 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
                  : (rows() == other.rows() && cols() == other.cols())))
         && "Size mismatch. Automatic resizing is disabled because EIGEN_NO_AUTOMATIC_RESIZING is defined");
       EIGEN_ONLY_USED_FOR_DEBUG(other);
-      if(this->size()==0)
-        resizeLike(other);
       #else
       resizeLike(other);
       #endif
@@ -624,25 +701,23 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       *
       * \internal
       */
+    // aliasing is dealt once in internall::call_assignment
+    // so at this stage we have to assume aliasing... and resising has to be done later.
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC 
     EIGEN_STRONG_INLINE Derived& _set(const DenseBase<OtherDerived>& other)
     {
-      _set_selector(other.derived(), typename internal::conditional<static_cast<bool>(int(OtherDerived::Flags) & EvalBeforeAssigningBit), internal::true_type, internal::false_type>::type());
+      internal::call_assignment(this->derived(), other.derived());
       return this->derived();
     }
 
-    template<typename OtherDerived>
-    EIGEN_STRONG_INLINE void _set_selector(const OtherDerived& other, const internal::true_type&) { _set_noalias(other.eval()); }
-
-    template<typename OtherDerived>
-    EIGEN_STRONG_INLINE void _set_selector(const OtherDerived& other, const internal::false_type&) { _set_noalias(other); }
-
     /** \internal Like _set() but additionally makes the assumption that no aliasing effect can happen (which
       * is the case when creating a new matrix) so one can enforce lazy evaluation.
       *
       * \sa operator=(const MatrixBase<OtherDerived>&), _set()
       */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC 
     EIGEN_STRONG_INLINE Derived& _set_noalias(const DenseBase<OtherDerived>& other)
     {
       // I don't think we need this resize call since the lazyAssign will anyways resize
@@ -650,40 +725,175 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       //_resize_to_match(other);
       // the 'false' below means to enforce lazy evaluation. We don't use lazyAssign() because
       // it wouldn't allow to copy a row-vector into a column-vector.
-      return internal::assign_selector<Derived,OtherDerived,false>::run(this->derived(), other.derived());
+      internal::call_assignment_no_alias(this->derived(), other.derived(), internal::assign_op<Scalar,typename OtherDerived::Scalar>());
+      return this->derived();
     }
 
     template<typename T0, typename T1>
-    EIGEN_STRONG_INLINE void _init2(Index nbRows, Index nbCols, typename internal::enable_if<Base::SizeAtCompileTime!=2,T0>::type* = 0)
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init2(Index rows, Index cols, typename internal::enable_if<Base::SizeAtCompileTime!=2,T0>::type* = 0)
     {
       EIGEN_STATIC_ASSERT(bool(NumTraits<T0>::IsInteger) &&
                           bool(NumTraits<T1>::IsInteger),
                           FLOATING_POINT_ARGUMENT_PASSED__INTEGER_WAS_EXPECTED)
-      resize(nbRows,nbCols);
+      resize(rows,cols);
+    }
+    
+    template<typename T0, typename T1>
+    EIGEN_DEVICE_FUNC 
+    EIGEN_STRONG_INLINE void _init2(const T0& val0, const T1& val1, typename internal::enable_if<Base::SizeAtCompileTime==2,T0>::type* = 0)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, 2)
+      m_storage.data()[0] = Scalar(val0);
+      m_storage.data()[1] = Scalar(val1);
     }
+    
     template<typename T0, typename T1>
-    EIGEN_STRONG_INLINE void _init2(const Scalar& val0, const Scalar& val1, typename internal::enable_if<Base::SizeAtCompileTime==2,T0>::type* = 0)
+    EIGEN_DEVICE_FUNC 
+    EIGEN_STRONG_INLINE void _init2(const Index& val0, const Index& val1,
+                                    typename internal::enable_if<    (!internal::is_same<Index,Scalar>::value)
+                                                                  && (internal::is_same<T0,Index>::value)
+                                                                  && (internal::is_same<T1,Index>::value)
+                                                                  && Base::SizeAtCompileTime==2,T1>::type* = 0)
     {
       EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, 2)
+      m_storage.data()[0] = Scalar(val0);
+      m_storage.data()[1] = Scalar(val1);
+    }
+
+    // The argument is convertible to the Index type and we either have a non 1x1 Matrix, or a dynamic-sized Array,
+    // then the argument is meant to be the size of the object.
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(Index size, typename internal::enable_if<    (Base::SizeAtCompileTime!=1 || !internal::is_convertible<T, Scalar>::value)
+                                                                              && ((!internal::is_same<typename internal::traits<Derived>::XprKind,ArrayXpr>::value || Base::SizeAtCompileTime==Dynamic)),T>::type* = 0)
+    {
+      // NOTE MSVC 2008 complains if we directly put bool(NumTraits<T>::IsInteger) as the EIGEN_STATIC_ASSERT argument.
+      const bool is_integer = NumTraits<T>::IsInteger;
+      EIGEN_UNUSED_VARIABLE(is_integer);
+      EIGEN_STATIC_ASSERT(is_integer,
+                          FLOATING_POINT_ARGUMENT_PASSED__INTEGER_WAS_EXPECTED)
+      resize(size);
+    }
+    
+    // We have a 1x1 matrix/array => the argument is interpreted as the value of the unique coefficient (case where scalar type can be implicitely converted)
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const Scalar& val0, typename internal::enable_if<Base::SizeAtCompileTime==1 && internal::is_convertible<T, Scalar>::value,T>::type* = 0)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, 1)
       m_storage.data()[0] = val0;
-      m_storage.data()[1] = val1;
+    }
+    
+    // We have a 1x1 matrix/array => the argument is interpreted as the value of the unique coefficient (case where scalar type match the index type)
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const Index& val0,
+                                    typename internal::enable_if<    (!internal::is_same<Index,Scalar>::value)
+                                                                  && (internal::is_same<Index,T>::value)
+                                                                  && Base::SizeAtCompileTime==1
+                                                                  && internal::is_convertible<T, Scalar>::value,T*>::type* = 0)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, 1)
+      m_storage.data()[0] = Scalar(val0);
     }
 
+    // Initialize a fixed size matrix from a pointer to raw data
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const Scalar* data){
+      this->_set_noalias(ConstMapType(data));
+    }
+
+    // Initialize an arbitrary matrix from a dense expression
+    template<typename T, typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const DenseBase<OtherDerived>& other){
+      this->_set_noalias(other);
+    }
+
+    // Initialize an arbitrary matrix from an object convertible to the Derived type.
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const Derived& other){
+      this->_set_noalias(other);
+    }
+
+    // Initialize an arbitrary matrix from a generic Eigen expression
+    template<typename T, typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const EigenBase<OtherDerived>& other){
+      this->derived() = other;
+    }
+
+    template<typename T, typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const ReturnByValue<OtherDerived>& other)
+    {
+      resize(other.rows(), other.cols());
+      other.evalTo(this->derived());
+    }
+
+    template<typename T, typename OtherDerived, int ColsAtCompileTime>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const RotationBase<OtherDerived,ColsAtCompileTime>& r)
+    {
+      this->derived() = r;
+    }
+    
+    // For fixed-size Array<Scalar,...>
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const Scalar& val0,
+                                    typename internal::enable_if<    Base::SizeAtCompileTime!=Dynamic
+                                                                  && Base::SizeAtCompileTime!=1
+                                                                  && internal::is_convertible<T, Scalar>::value
+                                                                  && internal::is_same<typename internal::traits<Derived>::XprKind,ArrayXpr>::value,T>::type* = 0)
+    {
+      Base::setConstant(val0);
+    }
+    
+    // For fixed-size Array<Index,...>
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const Index& val0,
+                                    typename internal::enable_if<    (!internal::is_same<Index,Scalar>::value)
+                                                                  && (internal::is_same<Index,T>::value)
+                                                                  && Base::SizeAtCompileTime!=Dynamic
+                                                                  && Base::SizeAtCompileTime!=1
+                                                                  && internal::is_convertible<T, Scalar>::value
+                                                                  && internal::is_same<typename internal::traits<Derived>::XprKind,ArrayXpr>::value,T*>::type* = 0)
+    {
+      Base::setConstant(val0);
+    }
+    
     template<typename MatrixTypeA, typename MatrixTypeB, bool SwapPointers>
     friend struct internal::matrix_swap_impl;
 
-    /** \internal generic implementation of swap for dense storage since for dynamic-sized matrices of same type it is enough to swap the
-      * data pointers.
+  public:
+    
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** \internal
+      * \brief Override DenseBase::swap() since for dynamic-sized matrices
+      * of same type it is enough to swap the data pointers.
       */
     template<typename OtherDerived>
-    void _swap(DenseBase<OtherDerived> const & other)
+    EIGEN_DEVICE_FUNC
+    void swap(DenseBase<OtherDerived> & other)
     {
       enum { SwapPointers = internal::is_same<Derived, OtherDerived>::value && Base::SizeAtCompileTime==Dynamic };
-      internal::matrix_swap_impl<Derived, OtherDerived, bool(SwapPointers)>::run(this->derived(), other.const_cast_derived());
+      internal::matrix_swap_impl<Derived, OtherDerived, bool(SwapPointers)>::run(this->derived(), other.derived());
     }
-
-  public:
-#ifndef EIGEN_PARSED_BY_DOXYGEN
+    
+    /** \internal
+      * \brief const version forwarded to DenseBase::swap
+      */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    void swap(DenseBase<OtherDerived> const & other)
+    { Base::swap(other.derived()); }
+    
+    EIGEN_DEVICE_FUNC 
     static EIGEN_STRONG_INLINE void _check_template_params()
     {
       EIGEN_STATIC_ASSERT((EIGEN_IMPLIES(MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1, (Options&RowMajor)==RowMajor)
@@ -697,10 +907,9 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
                         && (Options & (DontAlign|RowMajor)) == Options),
         INVALID_MATRIX_TEMPLATE_PARAMETERS)
     }
-#endif
 
-private:
-    enum { ThisConstantIsPrivateInPlainObjectBase };
+    enum { IsPlainObjectBase = 1 };
+#endif
 };
 
 namespace internal {
@@ -708,7 +917,6 @@ namespace internal {
 template <typename Derived, typename OtherDerived, bool IsVector>
 struct conservative_resize_like_impl
 {
-  typedef typename Derived::Index Index;
   static void run(DenseBase<Derived>& _this, Index rows, Index cols)
   {
     if (_this.rows() == rows && _this.cols() == cols) return;
@@ -724,8 +932,8 @@ struct conservative_resize_like_impl
     {
       // The storage order does not allow us to use reallocation.
       typename Derived::PlainObject tmp(rows,cols);
-      const Index common_rows = (std::min)(rows, _this.rows());
-      const Index common_cols = (std::min)(cols, _this.cols());
+      const Index common_rows = numext::mini(rows, _this.rows());
+      const Index common_cols = numext::mini(cols, _this.cols());
       tmp.block(0,0,common_rows,common_cols) = _this.block(0,0,common_rows,common_cols);
       _this.derived().swap(tmp);
     }
@@ -758,8 +966,8 @@ struct conservative_resize_like_impl
     {
       // The storage order does not allow us to use reallocation.
       typename Derived::PlainObject tmp(other);
-      const Index common_rows = (std::min)(tmp.rows(), _this.rows());
-      const Index common_cols = (std::min)(tmp.cols(), _this.cols());
+      const Index common_rows = numext::mini(tmp.rows(), _this.rows());
+      const Index common_cols = numext::mini(tmp.cols(), _this.cols());
       tmp.block(0,0,common_rows,common_cols) = _this.block(0,0,common_rows,common_cols);
       _this.derived().swap(tmp);
     }
@@ -774,7 +982,6 @@ struct conservative_resize_like_impl<Derived,OtherDerived,true>
 {
   using conservative_resize_like_impl<Derived,OtherDerived,false>::run;
   
-  typedef typename Derived::Index Index;
   static void run(DenseBase<Derived>& _this, Index size)
   {
     const Index new_rows = Derived::RowsAtCompileTime==1 ? 1 : size;
@@ -800,6 +1007,7 @@ struct conservative_resize_like_impl<Derived,OtherDerived,true>
 template<typename MatrixTypeA, typename MatrixTypeB, bool SwapPointers>
 struct matrix_swap_impl
 {
+  EIGEN_DEVICE_FUNC
   static inline void run(MatrixTypeA& a, MatrixTypeB& b)
   {
     a.base().swap(b);
@@ -809,6 +1017,7 @@ struct matrix_swap_impl
 template<typename MatrixTypeA, typename MatrixTypeB>
 struct matrix_swap_impl<MatrixTypeA, MatrixTypeB, true>
 {
+  EIGEN_DEVICE_FUNC
   static inline void run(MatrixTypeA& a, MatrixTypeB& b)
   {
     static_cast<typename MatrixTypeA::Base&>(a).m_storage.swap(static_cast<typename MatrixTypeB::Base&>(b).m_storage);
diff --git a/eigen/Eigen/src/Core/Product.h b/eigen/Eigen/src/Core/Product.h
new file mode 100644
index 0000000..ae0c94b
--- /dev/null
+++ b/eigen/Eigen/src/Core/Product.h
@@ -0,0 +1,186 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2011 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PRODUCT_H
+#define EIGEN_PRODUCT_H
+
+namespace Eigen {
+
+template<typename Lhs, typename Rhs, int Option, typename StorageKind> class ProductImpl;
+
+namespace internal {
+
+template<typename Lhs, typename Rhs, int Option>
+struct traits<Product<Lhs, Rhs, Option> >
+{
+  typedef typename remove_all<Lhs>::type LhsCleaned;
+  typedef typename remove_all<Rhs>::type RhsCleaned;
+  typedef traits<LhsCleaned> LhsTraits;
+  typedef traits<RhsCleaned> RhsTraits;
+  
+  typedef MatrixXpr XprKind;
+  
+  typedef typename ScalarBinaryOpTraits<typename traits<LhsCleaned>::Scalar, typename traits<RhsCleaned>::Scalar>::ReturnType Scalar;
+  typedef typename product_promote_storage_type<typename LhsTraits::StorageKind,
+                                                typename RhsTraits::StorageKind,
+                                                internal::product_type<Lhs,Rhs>::ret>::ret StorageKind;
+  typedef typename promote_index_type<typename LhsTraits::StorageIndex,
+                                      typename RhsTraits::StorageIndex>::type StorageIndex;
+  
+  enum {
+    RowsAtCompileTime    = LhsTraits::RowsAtCompileTime,
+    ColsAtCompileTime    = RhsTraits::ColsAtCompileTime,
+    MaxRowsAtCompileTime = LhsTraits::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = RhsTraits::MaxColsAtCompileTime,
+    
+    // FIXME: only needed by GeneralMatrixMatrixTriangular
+    InnerSize = EIGEN_SIZE_MIN_PREFER_FIXED(LhsTraits::ColsAtCompileTime, RhsTraits::RowsAtCompileTime),
+    
+    // The storage order is somewhat arbitrary here. The correct one will be determined through the evaluator.
+    Flags = (MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1) ? RowMajorBit
+          : (MaxColsAtCompileTime==1 && MaxRowsAtCompileTime!=1) ? 0
+          : (   ((LhsTraits::Flags&NoPreferredStorageOrderBit) && (RhsTraits::Flags&RowMajorBit))
+             || ((RhsTraits::Flags&NoPreferredStorageOrderBit) && (LhsTraits::Flags&RowMajorBit)) ) ? RowMajorBit
+          : NoPreferredStorageOrderBit
+  };
+};
+
+} // end namespace internal
+
+/** \class Product
+  * \ingroup Core_Module
+  *
+  * \brief Expression of the product of two arbitrary matrices or vectors
+  *
+  * \tparam _Lhs the type of the left-hand side expression
+  * \tparam _Rhs the type of the right-hand side expression
+  *
+  * This class represents an expression of the product of two arbitrary matrices.
+  *
+  * The other template parameters are:
+  * \tparam Option     can be DefaultProduct, AliasFreeProduct, or LazyProduct
+  *
+  */
+template<typename _Lhs, typename _Rhs, int Option>
+class Product : public ProductImpl<_Lhs,_Rhs,Option,
+                                   typename internal::product_promote_storage_type<typename internal::traits<_Lhs>::StorageKind,
+                                                                                   typename internal::traits<_Rhs>::StorageKind,
+                                                                                   internal::product_type<_Lhs,_Rhs>::ret>::ret>
+{
+  public:
+    
+    typedef _Lhs Lhs;
+    typedef _Rhs Rhs;
+    
+    typedef typename ProductImpl<
+        Lhs, Rhs, Option,
+        typename internal::product_promote_storage_type<typename internal::traits<Lhs>::StorageKind,
+                                                        typename internal::traits<Rhs>::StorageKind,
+                                                        internal::product_type<Lhs,Rhs>::ret>::ret>::Base Base;
+    EIGEN_GENERIC_PUBLIC_INTERFACE(Product)
+
+    typedef typename internal::ref_selector<Lhs>::type LhsNested;
+    typedef typename internal::ref_selector<Rhs>::type RhsNested;
+    typedef typename internal::remove_all<LhsNested>::type LhsNestedCleaned;
+    typedef typename internal::remove_all<RhsNested>::type RhsNestedCleaned;
+
+    EIGEN_DEVICE_FUNC Product(const Lhs& lhs, const Rhs& rhs) : m_lhs(lhs), m_rhs(rhs)
+    {
+      eigen_assert(lhs.cols() == rhs.rows()
+        && "invalid matrix product"
+        && "if you wanted a coeff-wise or a dot product use the respective explicit functions");
+    }
+
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_lhs.rows(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_rhs.cols(); }
+
+    EIGEN_DEVICE_FUNC const LhsNestedCleaned& lhs() const { return m_lhs; }
+    EIGEN_DEVICE_FUNC const RhsNestedCleaned& rhs() const { return m_rhs; }
+
+  protected:
+
+    LhsNested m_lhs;
+    RhsNested m_rhs;
+};
+
+namespace internal {
+  
+template<typename Lhs, typename Rhs, int Option, int ProductTag = internal::product_type<Lhs,Rhs>::ret>
+class dense_product_base
+ : public internal::dense_xpr_base<Product<Lhs,Rhs,Option> >::type
+{};
+
+/** Convertion to scalar for inner-products */
+template<typename Lhs, typename Rhs, int Option>
+class dense_product_base<Lhs, Rhs, Option, InnerProduct>
+ : public internal::dense_xpr_base<Product<Lhs,Rhs,Option> >::type
+{
+  typedef Product<Lhs,Rhs,Option> ProductXpr;
+  typedef typename internal::dense_xpr_base<ProductXpr>::type Base;
+public:
+  using Base::derived;
+  typedef typename Base::Scalar Scalar;
+  
+  operator const Scalar() const
+  {
+    return internal::evaluator<ProductXpr>(derived()).coeff(0,0);
+  }
+};
+
+} // namespace internal
+
+// Generic API dispatcher
+template<typename Lhs, typename Rhs, int Option, typename StorageKind>
+class ProductImpl : public internal::generic_xpr_base<Product<Lhs,Rhs,Option>, MatrixXpr, StorageKind>::type
+{
+  public:
+    typedef typename internal::generic_xpr_base<Product<Lhs,Rhs,Option>, MatrixXpr, StorageKind>::type Base;
+};
+
+template<typename Lhs, typename Rhs, int Option>
+class ProductImpl<Lhs,Rhs,Option,Dense>
+  : public internal::dense_product_base<Lhs,Rhs,Option>
+{
+    typedef Product<Lhs, Rhs, Option> Derived;
+    
+  public:
+    
+    typedef typename internal::dense_product_base<Lhs, Rhs, Option> Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(Derived)
+  protected:
+    enum {
+      IsOneByOne = (RowsAtCompileTime == 1 || RowsAtCompileTime == Dynamic) && 
+                   (ColsAtCompileTime == 1 || ColsAtCompileTime == Dynamic),
+      EnableCoeff = IsOneByOne || Option==LazyProduct
+    };
+    
+  public:
+  
+    EIGEN_DEVICE_FUNC Scalar coeff(Index row, Index col) const
+    {
+      EIGEN_STATIC_ASSERT(EnableCoeff, THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS);
+      eigen_assert( (Option==LazyProduct) || (this->rows() == 1 && this->cols() == 1) );
+      
+      return internal::evaluator<Derived>(derived()).coeff(row,col);
+    }
+
+    EIGEN_DEVICE_FUNC Scalar coeff(Index i) const
+    {
+      EIGEN_STATIC_ASSERT(EnableCoeff, THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS);
+      eigen_assert( (Option==LazyProduct) || (this->rows() == 1 && this->cols() == 1) );
+      
+      return internal::evaluator<Derived>(derived()).coeff(i);
+    }
+    
+  
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_PRODUCT_H
diff --git a/eigen/Eigen/src/Core/ProductBase.h b/eigen/Eigen/src/Core/ProductBase.h
deleted file mode 100644
index cf74470..0000000
--- a/eigen/Eigen/src/Core/ProductBase.h
+++ /dev/null
@@ -1,290 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2009-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-#ifndef EIGEN_PRODUCTBASE_H
-#define EIGEN_PRODUCTBASE_H
-
-namespace Eigen { 
-
-/** \class ProductBase
-  * \ingroup Core_Module
-  *
-  */
-
-namespace internal {
-template<typename Derived, typename _Lhs, typename _Rhs>
-struct traits<ProductBase<Derived,_Lhs,_Rhs> >
-{
-  typedef MatrixXpr XprKind;
-  typedef typename remove_all<_Lhs>::type Lhs;
-  typedef typename remove_all<_Rhs>::type Rhs;
-  typedef typename scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType Scalar;
-  typedef typename promote_storage_type<typename traits<Lhs>::StorageKind,
-                                           typename traits<Rhs>::StorageKind>::ret StorageKind;
-  typedef typename promote_index_type<typename traits<Lhs>::Index,
-                                         typename traits<Rhs>::Index>::type Index;
-  enum {
-    RowsAtCompileTime = traits<Lhs>::RowsAtCompileTime,
-    ColsAtCompileTime = traits<Rhs>::ColsAtCompileTime,
-    MaxRowsAtCompileTime = traits<Lhs>::MaxRowsAtCompileTime,
-    MaxColsAtCompileTime = traits<Rhs>::MaxColsAtCompileTime,
-    Flags = (MaxRowsAtCompileTime==1 ? RowMajorBit : 0)
-          | EvalBeforeNestingBit | EvalBeforeAssigningBit | NestByRefBit,
-                  // Note that EvalBeforeNestingBit and NestByRefBit
-                  // are not used in practice because nested is overloaded for products
-    CoeffReadCost = 0 // FIXME why is it needed ?
-  };
-};
-}
-
-#define EIGEN_PRODUCT_PUBLIC_INTERFACE(Derived) \
-  typedef ProductBase<Derived, Lhs, Rhs > Base; \
-  EIGEN_DENSE_PUBLIC_INTERFACE(Derived) \
-  typedef typename Base::LhsNested LhsNested; \
-  typedef typename Base::_LhsNested _LhsNested; \
-  typedef typename Base::LhsBlasTraits LhsBlasTraits; \
-  typedef typename Base::ActualLhsType ActualLhsType; \
-  typedef typename Base::_ActualLhsType _ActualLhsType; \
-  typedef typename Base::RhsNested RhsNested; \
-  typedef typename Base::_RhsNested _RhsNested; \
-  typedef typename Base::RhsBlasTraits RhsBlasTraits; \
-  typedef typename Base::ActualRhsType ActualRhsType; \
-  typedef typename Base::_ActualRhsType _ActualRhsType; \
-  using Base::m_lhs; \
-  using Base::m_rhs;
-
-template<typename Derived, typename Lhs, typename Rhs>
-class ProductBase : public MatrixBase<Derived>
-{
-  public:
-    typedef MatrixBase<Derived> Base;
-    EIGEN_DENSE_PUBLIC_INTERFACE(ProductBase)
-    
-    typedef typename Lhs::Nested LhsNested;
-    typedef typename internal::remove_all<LhsNested>::type _LhsNested;
-    typedef internal::blas_traits<_LhsNested> LhsBlasTraits;
-    typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
-    typedef typename internal::remove_all<ActualLhsType>::type _ActualLhsType;
-    typedef typename internal::traits<Lhs>::Scalar LhsScalar;
-
-    typedef typename Rhs::Nested RhsNested;
-    typedef typename internal::remove_all<RhsNested>::type _RhsNested;
-    typedef internal::blas_traits<_RhsNested> RhsBlasTraits;
-    typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
-    typedef typename internal::remove_all<ActualRhsType>::type _ActualRhsType;
-    typedef typename internal::traits<Rhs>::Scalar RhsScalar;
-
-    // Diagonal of a product: no need to evaluate the arguments because they are going to be evaluated only once
-    typedef CoeffBasedProduct<LhsNested, RhsNested, 0> FullyLazyCoeffBaseProductType;
-
-  public:
-
-#ifndef EIGEN_NO_MALLOC
-    typedef typename Base::PlainObject BasePlainObject;
-    typedef Matrix<Scalar,RowsAtCompileTime==1?1:Dynamic,ColsAtCompileTime==1?1:Dynamic,BasePlainObject::Options> DynPlainObject;
-    typedef typename internal::conditional<(BasePlainObject::SizeAtCompileTime==Dynamic) || (BasePlainObject::SizeAtCompileTime*int(sizeof(Scalar)) < int(EIGEN_STACK_ALLOCATION_LIMIT)),
-                                           BasePlainObject, DynPlainObject>::type PlainObject;
-#else
-    typedef typename Base::PlainObject PlainObject;
-#endif
-
-    ProductBase(const Lhs& a_lhs, const Rhs& a_rhs)
-      : m_lhs(a_lhs), m_rhs(a_rhs)
-    {
-      eigen_assert(a_lhs.cols() == a_rhs.rows()
-        && "invalid matrix product"
-        && "if you wanted a coeff-wise or a dot product use the respective explicit functions");
-    }
-
-    inline Index rows() const { return m_lhs.rows(); }
-    inline Index cols() const { return m_rhs.cols(); }
-
-    template<typename Dest>
-    inline void evalTo(Dest& dst) const { dst.setZero(); scaleAndAddTo(dst,Scalar(1)); }
-
-    template<typename Dest>
-    inline void addTo(Dest& dst) const { scaleAndAddTo(dst,Scalar(1)); }
-
-    template<typename Dest>
-    inline void subTo(Dest& dst) const { scaleAndAddTo(dst,Scalar(-1)); }
-
-    template<typename Dest>
-    inline void scaleAndAddTo(Dest& dst, const Scalar& alpha) const { derived().scaleAndAddTo(dst,alpha); }
-
-    const _LhsNested& lhs() const { return m_lhs; }
-    const _RhsNested& rhs() const { return m_rhs; }
-
-    // Implicit conversion to the nested type (trigger the evaluation of the product)
-    operator const PlainObject& () const
-    {
-      m_result.resize(m_lhs.rows(), m_rhs.cols());
-      derived().evalTo(m_result);
-      return m_result;
-    }
-
-    const Diagonal<const FullyLazyCoeffBaseProductType,0> diagonal() const
-    { return FullyLazyCoeffBaseProductType(m_lhs, m_rhs); }
-
-    template<int Index>
-    const Diagonal<FullyLazyCoeffBaseProductType,Index> diagonal() const
-    { return FullyLazyCoeffBaseProductType(m_lhs, m_rhs); }
-
-    const Diagonal<FullyLazyCoeffBaseProductType,Dynamic> diagonal(Index index) const
-    { return FullyLazyCoeffBaseProductType(m_lhs, m_rhs).diagonal(index); }
-
-    // restrict coeff accessors to 1x1 expressions. No need to care about mutators here since this isnt a Lvalue expression
-    typename Base::CoeffReturnType coeff(Index row, Index col) const
-    {
-#ifdef EIGEN2_SUPPORT
-      return lhs().row(row).cwiseProduct(rhs().col(col).transpose()).sum();
-#else
-      EIGEN_STATIC_ASSERT_SIZE_1x1(Derived)
-      eigen_assert(this->rows() == 1 && this->cols() == 1);
-      Matrix<Scalar,1,1> result = *this;
-      return result.coeff(row,col);
-#endif
-    }
-
-    typename Base::CoeffReturnType coeff(Index i) const
-    {
-      EIGEN_STATIC_ASSERT_SIZE_1x1(Derived)
-      eigen_assert(this->rows() == 1 && this->cols() == 1);
-      Matrix<Scalar,1,1> result = *this;
-      return result.coeff(i);
-    }
-
-    const Scalar& coeffRef(Index row, Index col) const
-    {
-      EIGEN_STATIC_ASSERT_SIZE_1x1(Derived)
-      eigen_assert(this->rows() == 1 && this->cols() == 1);
-      return derived().coeffRef(row,col);
-    }
-
-    const Scalar& coeffRef(Index i) const
-    {
-      EIGEN_STATIC_ASSERT_SIZE_1x1(Derived)
-      eigen_assert(this->rows() == 1 && this->cols() == 1);
-      return derived().coeffRef(i);
-    }
-
-  protected:
-
-    LhsNested m_lhs;
-    RhsNested m_rhs;
-
-    mutable PlainObject m_result;
-};
-
-// here we need to overload the nested rule for products
-// such that the nested type is a const reference to a plain matrix
-namespace internal {
-template<typename Lhs, typename Rhs, int Mode, int N, typename PlainObject>
-struct nested<GeneralProduct<Lhs,Rhs,Mode>, N, PlainObject>
-{
-  typedef typename GeneralProduct<Lhs,Rhs,Mode>::PlainObject const& type;
-};
-template<typename Lhs, typename Rhs, int Mode, int N, typename PlainObject>
-struct nested<const GeneralProduct<Lhs,Rhs,Mode>, N, PlainObject>
-{
-  typedef typename GeneralProduct<Lhs,Rhs,Mode>::PlainObject const& type;
-};
-}
-
-template<typename NestedProduct>
-class ScaledProduct;
-
-// Note that these two operator* functions are not defined as member
-// functions of ProductBase, because, otherwise we would have to
-// define all overloads defined in MatrixBase. Furthermore, Using
-// "using Base::operator*" would not work with MSVC.
-//
-// Also note that here we accept any compatible scalar types
-template<typename Derived,typename Lhs,typename Rhs>
-const ScaledProduct<Derived>
-operator*(const ProductBase<Derived,Lhs,Rhs>& prod, const typename Derived::Scalar& x)
-{ return ScaledProduct<Derived>(prod.derived(), x); }
-
-template<typename Derived,typename Lhs,typename Rhs>
-typename internal::enable_if<!internal::is_same<typename Derived::Scalar,typename Derived::RealScalar>::value,
-                      const ScaledProduct<Derived> >::type
-operator*(const ProductBase<Derived,Lhs,Rhs>& prod, const typename Derived::RealScalar& x)
-{ return ScaledProduct<Derived>(prod.derived(), x); }
-
-
-template<typename Derived,typename Lhs,typename Rhs>
-const ScaledProduct<Derived>
-operator*(const typename Derived::Scalar& x,const ProductBase<Derived,Lhs,Rhs>& prod)
-{ return ScaledProduct<Derived>(prod.derived(), x); }
-
-template<typename Derived,typename Lhs,typename Rhs>
-typename internal::enable_if<!internal::is_same<typename Derived::Scalar,typename Derived::RealScalar>::value,
-                      const ScaledProduct<Derived> >::type
-operator*(const typename Derived::RealScalar& x,const ProductBase<Derived,Lhs,Rhs>& prod)
-{ return ScaledProduct<Derived>(prod.derived(), x); }
-
-namespace internal {
-template<typename NestedProduct>
-struct traits<ScaledProduct<NestedProduct> >
- : traits<ProductBase<ScaledProduct<NestedProduct>,
-                         typename NestedProduct::_LhsNested,
-                         typename NestedProduct::_RhsNested> >
-{
-  typedef typename traits<NestedProduct>::StorageKind StorageKind;
-};
-}
-
-template<typename NestedProduct>
-class ScaledProduct
-  : public ProductBase<ScaledProduct<NestedProduct>,
-                       typename NestedProduct::_LhsNested,
-                       typename NestedProduct::_RhsNested>
-{
-  public:
-    typedef ProductBase<ScaledProduct<NestedProduct>,
-                       typename NestedProduct::_LhsNested,
-                       typename NestedProduct::_RhsNested> Base;
-    typedef typename Base::Scalar Scalar;
-    typedef typename Base::PlainObject PlainObject;
-//     EIGEN_PRODUCT_PUBLIC_INTERFACE(ScaledProduct)
-
-    ScaledProduct(const NestedProduct& prod, const Scalar& x)
-    : Base(prod.lhs(),prod.rhs()), m_prod(prod), m_alpha(x) {}
-
-    template<typename Dest>
-    inline void evalTo(Dest& dst) const { dst.setZero(); scaleAndAddTo(dst, Scalar(1)); }
-
-    template<typename Dest>
-    inline void addTo(Dest& dst) const { scaleAndAddTo(dst, Scalar(1)); }
-
-    template<typename Dest>
-    inline void subTo(Dest& dst) const { scaleAndAddTo(dst, Scalar(-1)); }
-
-    template<typename Dest>
-    inline void scaleAndAddTo(Dest& dst, const Scalar& a_alpha) const { m_prod.derived().scaleAndAddTo(dst,a_alpha * m_alpha); }
-
-    const Scalar& alpha() const { return m_alpha; }
-    
-  protected:
-    const NestedProduct& m_prod;
-    Scalar m_alpha;
-};
-
-/** \internal
-  * Overloaded to perform an efficient C = (A*B).lazy() */
-template<typename Derived>
-template<typename ProductDerived, typename Lhs, typename Rhs>
-Derived& MatrixBase<Derived>::lazyAssign(const ProductBase<ProductDerived, Lhs,Rhs>& other)
-{
-  other.derived().evalTo(derived());
-  return derived();
-}
-
-} // end namespace Eigen
-
-#endif // EIGEN_PRODUCTBASE_H
diff --git a/eigen/Eigen/src/Core/ProductEvaluators.h b/eigen/Eigen/src/Core/ProductEvaluators.h
new file mode 100644
index 0000000..583b7f5
--- /dev/null
+++ b/eigen/Eigen/src/Core/ProductEvaluators.h
@@ -0,0 +1,1099 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2011 Jitse Niesen <jitse@maths.leeds.ac.uk>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+
+#ifndef EIGEN_PRODUCTEVALUATORS_H
+#define EIGEN_PRODUCTEVALUATORS_H
+
+namespace Eigen {
+  
+namespace internal {
+
+/** \internal
+  * Evaluator of a product expression.
+  * Since products require special treatments to handle all possible cases,
+  * we simply deffer the evaluation logic to a product_evaluator class
+  * which offers more partial specialization possibilities.
+  * 
+  * \sa class product_evaluator
+  */
+template<typename Lhs, typename Rhs, int Options>
+struct evaluator<Product<Lhs, Rhs, Options> > 
+ : public product_evaluator<Product<Lhs, Rhs, Options> >
+{
+  typedef Product<Lhs, Rhs, Options> XprType;
+  typedef product_evaluator<XprType> Base;
+  
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr) : Base(xpr) {}
+};
+ 
+// Catch "scalar * ( A * B )" and transform it to "(A*scalar) * B"
+// TODO we should apply that rule only if that's really helpful
+template<typename Lhs, typename Rhs, typename Scalar1, typename Scalar2, typename Plain1>
+struct evaluator_assume_aliasing<CwiseBinaryOp<internal::scalar_product_op<Scalar1,Scalar2>,
+                                               const CwiseNullaryOp<internal::scalar_constant_op<Scalar1>, Plain1>,
+                                               const Product<Lhs, Rhs, DefaultProduct> > >
+{
+  static const bool value = true;
+};
+template<typename Lhs, typename Rhs, typename Scalar1, typename Scalar2, typename Plain1>
+struct evaluator<CwiseBinaryOp<internal::scalar_product_op<Scalar1,Scalar2>,
+                               const CwiseNullaryOp<internal::scalar_constant_op<Scalar1>, Plain1>,
+                               const Product<Lhs, Rhs, DefaultProduct> > >
+ : public evaluator<Product<EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(Scalar1,Lhs,product), Rhs, DefaultProduct> >
+{
+  typedef CwiseBinaryOp<internal::scalar_product_op<Scalar1,Scalar2>,
+                               const CwiseNullaryOp<internal::scalar_constant_op<Scalar1>, Plain1>,
+                               const Product<Lhs, Rhs, DefaultProduct> > XprType;
+  typedef evaluator<Product<EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(Scalar1,Lhs,product), Rhs, DefaultProduct> > Base;
+
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr)
+    : Base(xpr.lhs().functor().m_other * xpr.rhs().lhs() * xpr.rhs().rhs())
+  {}
+};
+
+
+template<typename Lhs, typename Rhs, int DiagIndex>
+struct evaluator<Diagonal<const Product<Lhs, Rhs, DefaultProduct>, DiagIndex> > 
+ : public evaluator<Diagonal<const Product<Lhs, Rhs, LazyProduct>, DiagIndex> >
+{
+  typedef Diagonal<const Product<Lhs, Rhs, DefaultProduct>, DiagIndex> XprType;
+  typedef evaluator<Diagonal<const Product<Lhs, Rhs, LazyProduct>, DiagIndex> > Base;
+  
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr)
+    : Base(Diagonal<const Product<Lhs, Rhs, LazyProduct>, DiagIndex>(
+        Product<Lhs, Rhs, LazyProduct>(xpr.nestedExpression().lhs(), xpr.nestedExpression().rhs()),
+        xpr.index() ))
+  {}
+};
+
+
+// Helper class to perform a matrix product with the destination at hand.
+// Depending on the sizes of the factors, there are different evaluation strategies
+// as controlled by internal::product_type.
+template< typename Lhs, typename Rhs,
+          typename LhsShape = typename evaluator_traits<Lhs>::Shape,
+          typename RhsShape = typename evaluator_traits<Rhs>::Shape,
+          int ProductType = internal::product_type<Lhs,Rhs>::value>
+struct generic_product_impl;
+
+template<typename Lhs, typename Rhs>
+struct evaluator_assume_aliasing<Product<Lhs, Rhs, DefaultProduct> > {
+  static const bool value = true;
+};
+
+// This is the default evaluator implementation for products:
+// It creates a temporary and call generic_product_impl
+template<typename Lhs, typename Rhs, int Options, int ProductTag, typename LhsShape, typename RhsShape>
+struct product_evaluator<Product<Lhs, Rhs, Options>, ProductTag, LhsShape, RhsShape>
+  : public evaluator<typename Product<Lhs, Rhs, Options>::PlainObject>
+{
+  typedef Product<Lhs, Rhs, Options> XprType;
+  typedef typename XprType::PlainObject PlainObject;
+  typedef evaluator<PlainObject> Base;
+  enum {
+    Flags = Base::Flags | EvalBeforeNestingBit
+  };
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  explicit product_evaluator(const XprType& xpr)
+    : m_result(xpr.rows(), xpr.cols())
+  {
+    ::new (static_cast<Base*>(this)) Base(m_result);
+    
+// FIXME shall we handle nested_eval here?,
+// if so, then we must take care at removing the call to nested_eval in the specializations (e.g., in permutation_matrix_product, transposition_matrix_product, etc.)
+//     typedef typename internal::nested_eval<Lhs,Rhs::ColsAtCompileTime>::type LhsNested;
+//     typedef typename internal::nested_eval<Rhs,Lhs::RowsAtCompileTime>::type RhsNested;
+//     typedef typename internal::remove_all<LhsNested>::type LhsNestedCleaned;
+//     typedef typename internal::remove_all<RhsNested>::type RhsNestedCleaned;
+//     
+//     const LhsNested lhs(xpr.lhs());
+//     const RhsNested rhs(xpr.rhs());
+//   
+//     generic_product_impl<LhsNestedCleaned, RhsNestedCleaned>::evalTo(m_result, lhs, rhs);
+
+    generic_product_impl<Lhs, Rhs, LhsShape, RhsShape, ProductTag>::evalTo(m_result, xpr.lhs(), xpr.rhs());
+  }
+  
+protected:  
+  PlainObject m_result;
+};
+
+// The following three shortcuts are enabled only if the scalar types match excatly.
+// TODO: we could enable them for different scalar types when the product is not vectorized.
+
+// Dense = Product
+template< typename DstXprType, typename Lhs, typename Rhs, int Options, typename Scalar>
+struct Assignment<DstXprType, Product<Lhs,Rhs,Options>, internal::assign_op<Scalar,Scalar>, Dense2Dense,
+  typename enable_if<(Options==DefaultProduct || Options==AliasFreeProduct)>::type>
+{
+  typedef Product<Lhs,Rhs,Options> SrcXprType;
+  static EIGEN_STRONG_INLINE
+  void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,Scalar> &)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+    // FIXME shall we handle nested_eval here?
+    generic_product_impl<Lhs, Rhs>::evalTo(dst, src.lhs(), src.rhs());
+  }
+};
+
+// Dense += Product
+template< typename DstXprType, typename Lhs, typename Rhs, int Options, typename Scalar>
+struct Assignment<DstXprType, Product<Lhs,Rhs,Options>, internal::add_assign_op<Scalar,Scalar>, Dense2Dense,
+  typename enable_if<(Options==DefaultProduct || Options==AliasFreeProduct)>::type>
+{
+  typedef Product<Lhs,Rhs,Options> SrcXprType;
+  static EIGEN_STRONG_INLINE
+  void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op<Scalar,Scalar> &)
+  {
+    eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
+    // FIXME shall we handle nested_eval here?
+    generic_product_impl<Lhs, Rhs>::addTo(dst, src.lhs(), src.rhs());
+  }
+};
+
+// Dense -= Product
+template< typename DstXprType, typename Lhs, typename Rhs, int Options, typename Scalar>
+struct Assignment<DstXprType, Product<Lhs,Rhs,Options>, internal::sub_assign_op<Scalar,Scalar>, Dense2Dense,
+  typename enable_if<(Options==DefaultProduct || Options==AliasFreeProduct)>::type>
+{
+  typedef Product<Lhs,Rhs,Options> SrcXprType;
+  static EIGEN_STRONG_INLINE
+  void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op<Scalar,Scalar> &)
+  {
+    eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
+    // FIXME shall we handle nested_eval here?
+    generic_product_impl<Lhs, Rhs>::subTo(dst, src.lhs(), src.rhs());
+  }
+};
+
+
+// Dense ?= scalar * Product
+// TODO we should apply that rule if that's really helpful
+// for instance, this is not good for inner products
+template< typename DstXprType, typename Lhs, typename Rhs, typename AssignFunc, typename Scalar, typename ScalarBis, typename Plain>
+struct Assignment<DstXprType, CwiseBinaryOp<internal::scalar_product_op<ScalarBis,Scalar>, const CwiseNullaryOp<internal::scalar_constant_op<ScalarBis>,Plain>,
+                                           const Product<Lhs,Rhs,DefaultProduct> >, AssignFunc, Dense2Dense>
+{
+  typedef CwiseBinaryOp<internal::scalar_product_op<ScalarBis,Scalar>,
+                        const CwiseNullaryOp<internal::scalar_constant_op<ScalarBis>,Plain>,
+                        const Product<Lhs,Rhs,DefaultProduct> > SrcXprType;
+  static EIGEN_STRONG_INLINE
+  void run(DstXprType &dst, const SrcXprType &src, const AssignFunc& func)
+  {
+    call_assignment_no_alias(dst, (src.lhs().functor().m_other * src.rhs().lhs())*src.rhs().rhs(), func);
+  }
+};
+
+//----------------------------------------
+// Catch "Dense ?= xpr + Product<>" expression to save one temporary
+// FIXME we could probably enable these rules for any product, i.e., not only Dense and DefaultProduct
+
+template<typename OtherXpr, typename Lhs, typename Rhs>
+struct evaluator_assume_aliasing<CwiseBinaryOp<internal::scalar_sum_op<typename OtherXpr::Scalar,typename Product<Lhs,Rhs,DefaultProduct>::Scalar>, const OtherXpr,
+                                               const Product<Lhs,Rhs,DefaultProduct> >, DenseShape > {
+  static const bool value = true;
+};
+
+template<typename DstXprType, typename OtherXpr, typename ProductType, typename Func1, typename Func2>
+struct assignment_from_xpr_op_product
+{
+  template<typename SrcXprType, typename InitialFunc>
+  static EIGEN_STRONG_INLINE
+  void run(DstXprType &dst, const SrcXprType &src, const InitialFunc& /*func*/)
+  {
+    call_assignment_no_alias(dst, src.lhs(), Func1());
+    call_assignment_no_alias(dst, src.rhs(), Func2());
+  }
+};
+
+#define EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(ASSIGN_OP,BINOP,ASSIGN_OP2) \
+  template< typename DstXprType, typename OtherXpr, typename Lhs, typename Rhs, typename DstScalar, typename SrcScalar, typename OtherScalar,typename ProdScalar> \
+  struct Assignment<DstXprType, CwiseBinaryOp<internal::BINOP<OtherScalar,ProdScalar>, const OtherXpr, \
+                                            const Product<Lhs,Rhs,DefaultProduct> >, internal::ASSIGN_OP<DstScalar,SrcScalar>, Dense2Dense> \
+    : assignment_from_xpr_op_product<DstXprType, OtherXpr, Product<Lhs,Rhs,DefaultProduct>, internal::ASSIGN_OP<DstScalar,OtherScalar>, internal::ASSIGN_OP2<DstScalar,ProdScalar> > \
+  {}
+
+EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(assign_op,    scalar_sum_op,add_assign_op);
+EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(add_assign_op,scalar_sum_op,add_assign_op);
+EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(sub_assign_op,scalar_sum_op,sub_assign_op);
+
+EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(assign_op,    scalar_difference_op,sub_assign_op);
+EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(add_assign_op,scalar_difference_op,sub_assign_op);
+EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(sub_assign_op,scalar_difference_op,add_assign_op);
+
+//----------------------------------------
+
+template<typename Lhs, typename Rhs>
+struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,InnerProduct>
+{
+  template<typename Dst>
+  static inline void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    dst.coeffRef(0,0) = (lhs.transpose().cwiseProduct(rhs)).sum();
+  }
+  
+  template<typename Dst>
+  static inline void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    dst.coeffRef(0,0) += (lhs.transpose().cwiseProduct(rhs)).sum();
+  }
+  
+  template<typename Dst>
+  static void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  { dst.coeffRef(0,0) -= (lhs.transpose().cwiseProduct(rhs)).sum(); }
+};
+
+
+/***********************************************************************
+*  Implementation of outer dense * dense vector product
+***********************************************************************/
+
+// Column major result
+template<typename Dst, typename Lhs, typename Rhs, typename Func>
+void outer_product_selector_run(Dst& dst, const Lhs &lhs, const Rhs &rhs, const Func& func, const false_type&)
+{
+  evaluator<Rhs> rhsEval(rhs);
+  typename nested_eval<Lhs,Rhs::SizeAtCompileTime>::type actual_lhs(lhs);
+  // FIXME if cols is large enough, then it might be useful to make sure that lhs is sequentially stored
+  // FIXME not very good if rhs is real and lhs complex while alpha is real too
+  const Index cols = dst.cols();
+  for (Index j=0; j<cols; ++j)
+    func(dst.col(j), rhsEval.coeff(Index(0),j) * actual_lhs);
+}
+
+// Row major result
+template<typename Dst, typename Lhs, typename Rhs, typename Func>
+void outer_product_selector_run(Dst& dst, const Lhs &lhs, const Rhs &rhs, const Func& func, const true_type&)
+{
+  evaluator<Lhs> lhsEval(lhs);
+  typename nested_eval<Rhs,Lhs::SizeAtCompileTime>::type actual_rhs(rhs);
+  // FIXME if rows is large enough, then it might be useful to make sure that rhs is sequentially stored
+  // FIXME not very good if lhs is real and rhs complex while alpha is real too
+  const Index rows = dst.rows();
+  for (Index i=0; i<rows; ++i)
+    func(dst.row(i), lhsEval.coeff(i,Index(0)) * actual_rhs);
+}
+
+template<typename Lhs, typename Rhs>
+struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,OuterProduct>
+{
+  template<typename T> struct is_row_major : internal::conditional<(int(T::Flags)&RowMajorBit), internal::true_type, internal::false_type>::type {};
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  
+  // TODO it would be nice to be able to exploit our *_assign_op functors for that purpose
+  struct set  { template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived()  = src; } };
+  struct add  { template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived() += src; } };
+  struct sub  { template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived() -= src; } };
+  struct adds {
+    Scalar m_scale;
+    explicit adds(const Scalar& s) : m_scale(s) {}
+    template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const {
+      dst.const_cast_derived() += m_scale * src;
+    }
+  };
+  
+  template<typename Dst>
+  static inline void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    internal::outer_product_selector_run(dst, lhs, rhs, set(), is_row_major<Dst>());
+  }
+  
+  template<typename Dst>
+  static inline void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    internal::outer_product_selector_run(dst, lhs, rhs, add(), is_row_major<Dst>());
+  }
+  
+  template<typename Dst>
+  static inline void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    internal::outer_product_selector_run(dst, lhs, rhs, sub(), is_row_major<Dst>());
+  }
+  
+  template<typename Dst>
+  static inline void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
+  {
+    internal::outer_product_selector_run(dst, lhs, rhs, adds(alpha), is_row_major<Dst>());
+  }
+  
+};
+
+
+// This base class provides default implementations for evalTo, addTo, subTo, in terms of scaleAndAddTo
+template<typename Lhs, typename Rhs, typename Derived>
+struct generic_product_impl_base
+{
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  
+  template<typename Dst>
+  static EIGEN_STRONG_INLINE void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  { dst.setZero(); scaleAndAddTo(dst, lhs, rhs, Scalar(1)); }
+
+  template<typename Dst>
+  static EIGEN_STRONG_INLINE void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  { scaleAndAddTo(dst,lhs, rhs, Scalar(1)); }
+
+  template<typename Dst>
+  static EIGEN_STRONG_INLINE void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  { scaleAndAddTo(dst, lhs, rhs, Scalar(-1)); }
+  
+  template<typename Dst>
+  static EIGEN_STRONG_INLINE void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
+  { Derived::scaleAndAddTo(dst,lhs,rhs,alpha); }
+
+};
+
+template<typename Lhs, typename Rhs>
+struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemvProduct>
+  : generic_product_impl_base<Lhs,Rhs,generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemvProduct> >
+{
+  typedef typename nested_eval<Lhs,1>::type LhsNested;
+  typedef typename nested_eval<Rhs,1>::type RhsNested;
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  enum { Side = Lhs::IsVectorAtCompileTime ? OnTheLeft : OnTheRight };
+  typedef typename internal::remove_all<typename internal::conditional<int(Side)==OnTheRight,LhsNested,RhsNested>::type>::type MatrixType;
+
+  template<typename Dest>
+  static EIGEN_STRONG_INLINE void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
+  {
+    LhsNested actual_lhs(lhs);
+    RhsNested actual_rhs(rhs);
+    internal::gemv_dense_selector<Side,
+                            (int(MatrixType::Flags)&RowMajorBit) ? RowMajor : ColMajor,
+                            bool(internal::blas_traits<MatrixType>::HasUsableDirectAccess)
+                           >::run(actual_lhs, actual_rhs, dst, alpha);
+  }
+};
+
+template<typename Lhs, typename Rhs>
+struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,CoeffBasedProductMode> 
+{
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  
+  template<typename Dst>
+  static EIGEN_STRONG_INLINE void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    // Same as: dst.noalias() = lhs.lazyProduct(rhs);
+    // but easier on the compiler side
+    call_assignment_no_alias(dst, lhs.lazyProduct(rhs), internal::assign_op<typename Dst::Scalar,Scalar>());
+  }
+  
+  template<typename Dst>
+  static EIGEN_STRONG_INLINE void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    // dst.noalias() += lhs.lazyProduct(rhs);
+    call_assignment_no_alias(dst, lhs.lazyProduct(rhs), internal::add_assign_op<typename Dst::Scalar,Scalar>());
+  }
+  
+  template<typename Dst>
+  static EIGEN_STRONG_INLINE void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    // dst.noalias() -= lhs.lazyProduct(rhs);
+    call_assignment_no_alias(dst, lhs.lazyProduct(rhs), internal::sub_assign_op<typename Dst::Scalar,Scalar>());
+  }
+  
+//   template<typename Dst>
+//   static inline void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
+//   { dst.noalias() += alpha * lhs.lazyProduct(rhs); }
+};
+
+// This specialization enforces the use of a coefficient-based evaluation strategy
+template<typename Lhs, typename Rhs>
+struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,LazyCoeffBasedProductMode>
+  : generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,CoeffBasedProductMode> {};
+
+// Case 2: Evaluate coeff by coeff
+//
+// This is mostly taken from CoeffBasedProduct.h
+// The main difference is that we add an extra argument to the etor_product_*_impl::run() function
+// for the inner dimension of the product, because evaluator object do not know their size.
+
+template<int Traversal, int UnrollingIndex, typename Lhs, typename Rhs, typename RetScalar>
+struct etor_product_coeff_impl;
+
+template<int StorageOrder, int UnrollingIndex, typename Lhs, typename Rhs, typename Packet, int LoadMode>
+struct etor_product_packet_impl;
+
+template<typename Lhs, typename Rhs, int ProductTag>
+struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape, DenseShape>
+    : evaluator_base<Product<Lhs, Rhs, LazyProduct> >
+{
+  typedef Product<Lhs, Rhs, LazyProduct> XprType;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  explicit product_evaluator(const XprType& xpr)
+    : m_lhs(xpr.lhs()),
+      m_rhs(xpr.rhs()),
+      m_lhsImpl(m_lhs),     // FIXME the creation of the evaluator objects should result in a no-op, but check that!
+      m_rhsImpl(m_rhs),     //       Moreover, they are only useful for the packet path, so we could completely disable them when not needed,
+                            //       or perhaps declare them on the fly on the packet method... We have experiment to check what's best.
+      m_innerDim(xpr.lhs().cols())
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(NumTraits<Scalar>::MulCost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(NumTraits<Scalar>::AddCost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+#if 0
+    std::cerr << "LhsOuterStrideBytes=  " << LhsOuterStrideBytes << "\n";
+    std::cerr << "RhsOuterStrideBytes=  " << RhsOuterStrideBytes << "\n";
+    std::cerr << "LhsAlignment=         " << LhsAlignment << "\n";
+    std::cerr << "RhsAlignment=         " << RhsAlignment << "\n";
+    std::cerr << "CanVectorizeLhs=      " << CanVectorizeLhs << "\n";
+    std::cerr << "CanVectorizeRhs=      " << CanVectorizeRhs << "\n";
+    std::cerr << "CanVectorizeInner=    " << CanVectorizeInner << "\n";
+    std::cerr << "EvalToRowMajor=       " << EvalToRowMajor << "\n";
+    std::cerr << "Alignment=            " << Alignment << "\n";
+    std::cerr << "Flags=                " << Flags << "\n";
+#endif
+  }
+
+  // Everything below here is taken from CoeffBasedProduct.h
+
+  typedef typename internal::nested_eval<Lhs,Rhs::ColsAtCompileTime>::type LhsNested;
+  typedef typename internal::nested_eval<Rhs,Lhs::RowsAtCompileTime>::type RhsNested;
+  
+  typedef typename internal::remove_all<LhsNested>::type LhsNestedCleaned;
+  typedef typename internal::remove_all<RhsNested>::type RhsNestedCleaned;
+
+  typedef evaluator<LhsNestedCleaned> LhsEtorType;
+  typedef evaluator<RhsNestedCleaned> RhsEtorType;
+
+  enum {
+    RowsAtCompileTime = LhsNestedCleaned::RowsAtCompileTime,
+    ColsAtCompileTime = RhsNestedCleaned::ColsAtCompileTime,
+    InnerSize = EIGEN_SIZE_MIN_PREFER_FIXED(LhsNestedCleaned::ColsAtCompileTime, RhsNestedCleaned::RowsAtCompileTime),
+    MaxRowsAtCompileTime = LhsNestedCleaned::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = RhsNestedCleaned::MaxColsAtCompileTime
+  };
+
+  typedef typename find_best_packet<Scalar,RowsAtCompileTime>::type LhsVecPacketType;
+  typedef typename find_best_packet<Scalar,ColsAtCompileTime>::type RhsVecPacketType;
+
+  enum {
+      
+    LhsCoeffReadCost = LhsEtorType::CoeffReadCost,
+    RhsCoeffReadCost = RhsEtorType::CoeffReadCost,
+    CoeffReadCost = InnerSize==0 ? NumTraits<Scalar>::ReadCost
+                  : InnerSize == Dynamic ? HugeCost
+                  : InnerSize * (NumTraits<Scalar>::MulCost + LhsCoeffReadCost + RhsCoeffReadCost)
+                    + (InnerSize - 1) * NumTraits<Scalar>::AddCost,
+
+    Unroll = CoeffReadCost <= EIGEN_UNROLLING_LIMIT,
+    
+    LhsFlags = LhsEtorType::Flags,
+    RhsFlags = RhsEtorType::Flags,
+    
+    LhsRowMajor = LhsFlags & RowMajorBit,
+    RhsRowMajor = RhsFlags & RowMajorBit,
+
+    LhsVecPacketSize = unpacket_traits<LhsVecPacketType>::size,
+    RhsVecPacketSize = unpacket_traits<RhsVecPacketType>::size,
+
+    // Here, we don't care about alignment larger than the usable packet size.
+    LhsAlignment = EIGEN_PLAIN_ENUM_MIN(LhsEtorType::Alignment,LhsVecPacketSize*int(sizeof(typename LhsNestedCleaned::Scalar))),
+    RhsAlignment = EIGEN_PLAIN_ENUM_MIN(RhsEtorType::Alignment,RhsVecPacketSize*int(sizeof(typename RhsNestedCleaned::Scalar))),
+      
+    SameType = is_same<typename LhsNestedCleaned::Scalar,typename RhsNestedCleaned::Scalar>::value,
+
+    CanVectorizeRhs = bool(RhsRowMajor) && (RhsFlags & PacketAccessBit) && (ColsAtCompileTime!=1),
+    CanVectorizeLhs = (!LhsRowMajor) && (LhsFlags & PacketAccessBit) && (RowsAtCompileTime!=1),
+
+    EvalToRowMajor = (MaxRowsAtCompileTime==1&&MaxColsAtCompileTime!=1) ? 1
+                    : (MaxColsAtCompileTime==1&&MaxRowsAtCompileTime!=1) ? 0
+                    : (bool(RhsRowMajor) && !CanVectorizeLhs),
+
+    Flags = ((unsigned int)(LhsFlags | RhsFlags) & HereditaryBits & ~RowMajorBit)
+          | (EvalToRowMajor ? RowMajorBit : 0)
+          // TODO enable vectorization for mixed types
+          | (SameType && (CanVectorizeLhs || CanVectorizeRhs) ? PacketAccessBit : 0)
+          | (XprType::IsVectorAtCompileTime ? LinearAccessBit : 0),
+          
+    LhsOuterStrideBytes = int(LhsNestedCleaned::OuterStrideAtCompileTime) * int(sizeof(typename LhsNestedCleaned::Scalar)),
+    RhsOuterStrideBytes = int(RhsNestedCleaned::OuterStrideAtCompileTime) * int(sizeof(typename RhsNestedCleaned::Scalar)),
+
+    Alignment = bool(CanVectorizeLhs) ? (LhsOuterStrideBytes<=0 || (int(LhsOuterStrideBytes) % EIGEN_PLAIN_ENUM_MAX(1,LhsAlignment))!=0 ? 0 : LhsAlignment)
+              : bool(CanVectorizeRhs) ? (RhsOuterStrideBytes<=0 || (int(RhsOuterStrideBytes) % EIGEN_PLAIN_ENUM_MAX(1,RhsAlignment))!=0 ? 0 : RhsAlignment)
+              : 0,
+
+    /* CanVectorizeInner deserves special explanation. It does not affect the product flags. It is not used outside
+     * of Product. If the Product itself is not a packet-access expression, there is still a chance that the inner
+     * loop of the product might be vectorized. This is the meaning of CanVectorizeInner. Since it doesn't affect
+     * the Flags, it is safe to make this value depend on ActualPacketAccessBit, that doesn't affect the ABI.
+     */
+    CanVectorizeInner =    SameType
+                        && LhsRowMajor
+                        && (!RhsRowMajor)
+                        && (LhsFlags & RhsFlags & ActualPacketAccessBit)
+                        && (InnerSize % packet_traits<Scalar>::size == 0)
+  };
+  
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CoeffReturnType coeff(Index row, Index col) const
+  {
+    return (m_lhs.row(row).transpose().cwiseProduct( m_rhs.col(col) )).sum();
+  }
+
+  /* Allow index-based non-packet access. It is impossible though to allow index-based packed access,
+   * which is why we don't set the LinearAccessBit.
+   * TODO: this seems possible when the result is a vector
+   */
+  EIGEN_DEVICE_FUNC const CoeffReturnType coeff(Index index) const
+  {
+    const Index row = (RowsAtCompileTime == 1 || MaxRowsAtCompileTime==1) ? 0 : index;
+    const Index col = (RowsAtCompileTime == 1 || MaxRowsAtCompileTime==1) ? index : 0;
+    return (m_lhs.row(row).transpose().cwiseProduct( m_rhs.col(col) )).sum();
+  }
+
+  template<int LoadMode, typename PacketType>
+  const PacketType packet(Index row, Index col) const
+  {
+    PacketType res;
+    typedef etor_product_packet_impl<bool(int(Flags)&RowMajorBit) ? RowMajor : ColMajor,
+                                     Unroll ? int(InnerSize) : Dynamic,
+                                     LhsEtorType, RhsEtorType, PacketType, LoadMode> PacketImpl;
+    PacketImpl::run(row, col, m_lhsImpl, m_rhsImpl, m_innerDim, res);
+    return res;
+  }
+
+  template<int LoadMode, typename PacketType>
+  const PacketType packet(Index index) const
+  {
+    const Index row = (RowsAtCompileTime == 1 || MaxRowsAtCompileTime==1) ? 0 : index;
+    const Index col = (RowsAtCompileTime == 1 || MaxRowsAtCompileTime==1) ? index : 0;
+    return packet<LoadMode,PacketType>(row,col);
+  }
+
+protected:
+  typename internal::add_const_on_value_type<LhsNested>::type m_lhs;
+  typename internal::add_const_on_value_type<RhsNested>::type m_rhs;
+  
+  LhsEtorType m_lhsImpl;
+  RhsEtorType m_rhsImpl;
+
+  // TODO: Get rid of m_innerDim if known at compile time
+  Index m_innerDim;
+};
+
+template<typename Lhs, typename Rhs>
+struct product_evaluator<Product<Lhs, Rhs, DefaultProduct>, LazyCoeffBasedProductMode, DenseShape, DenseShape>
+  : product_evaluator<Product<Lhs, Rhs, LazyProduct>, CoeffBasedProductMode, DenseShape, DenseShape>
+{
+  typedef Product<Lhs, Rhs, DefaultProduct> XprType;
+  typedef Product<Lhs, Rhs, LazyProduct> BaseProduct;
+  typedef product_evaluator<BaseProduct, CoeffBasedProductMode, DenseShape, DenseShape> Base;
+  enum {
+    Flags = Base::Flags | EvalBeforeNestingBit
+  };
+  EIGEN_DEVICE_FUNC explicit product_evaluator(const XprType& xpr)
+    : Base(BaseProduct(xpr.lhs(),xpr.rhs()))
+  {}
+};
+
+/****************************************
+*** Coeff based product, Packet path  ***
+****************************************/
+
+template<int UnrollingIndex, typename Lhs, typename Rhs, typename Packet, int LoadMode>
+struct etor_product_packet_impl<RowMajor, UnrollingIndex, Lhs, Rhs, Packet, LoadMode>
+{
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, Packet &res)
+  {
+    etor_product_packet_impl<RowMajor, UnrollingIndex-1, Lhs, Rhs, Packet, LoadMode>::run(row, col, lhs, rhs, innerDim, res);
+    res =  pmadd(pset1<Packet>(lhs.coeff(row, Index(UnrollingIndex-1))), rhs.template packet<LoadMode,Packet>(Index(UnrollingIndex-1), col), res);
+  }
+};
+
+template<int UnrollingIndex, typename Lhs, typename Rhs, typename Packet, int LoadMode>
+struct etor_product_packet_impl<ColMajor, UnrollingIndex, Lhs, Rhs, Packet, LoadMode>
+{
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, Packet &res)
+  {
+    etor_product_packet_impl<ColMajor, UnrollingIndex-1, Lhs, Rhs, Packet, LoadMode>::run(row, col, lhs, rhs, innerDim, res);
+    res =  pmadd(lhs.template packet<LoadMode,Packet>(row, Index(UnrollingIndex-1)), pset1<Packet>(rhs.coeff(Index(UnrollingIndex-1), col)), res);
+  }
+};
+
+template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
+struct etor_product_packet_impl<RowMajor, 1, Lhs, Rhs, Packet, LoadMode>
+{
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index /*innerDim*/, Packet &res)
+  {
+    res = pmul(pset1<Packet>(lhs.coeff(row, Index(0))),rhs.template packet<LoadMode,Packet>(Index(0), col));
+  }
+};
+
+template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
+struct etor_product_packet_impl<ColMajor, 1, Lhs, Rhs, Packet, LoadMode>
+{
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index /*innerDim*/, Packet &res)
+  {
+    res = pmul(lhs.template packet<LoadMode,Packet>(row, Index(0)), pset1<Packet>(rhs.coeff(Index(0), col)));
+  }
+};
+
+template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
+struct etor_product_packet_impl<RowMajor, 0, Lhs, Rhs, Packet, LoadMode>
+{
+  static EIGEN_STRONG_INLINE void run(Index /*row*/, Index /*col*/, const Lhs& /*lhs*/, const Rhs& /*rhs*/, Index /*innerDim*/, Packet &res)
+  {
+    res = pset1<Packet>(typename unpacket_traits<Packet>::type(0));
+  }
+};
+
+template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
+struct etor_product_packet_impl<ColMajor, 0, Lhs, Rhs, Packet, LoadMode>
+{
+  static EIGEN_STRONG_INLINE void run(Index /*row*/, Index /*col*/, const Lhs& /*lhs*/, const Rhs& /*rhs*/, Index /*innerDim*/, Packet &res)
+  {
+    res = pset1<Packet>(typename unpacket_traits<Packet>::type(0));
+  }
+};
+
+template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
+struct etor_product_packet_impl<RowMajor, Dynamic, Lhs, Rhs, Packet, LoadMode>
+{
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, Packet& res)
+  {
+    res = pset1<Packet>(typename unpacket_traits<Packet>::type(0));
+    for(Index i = 0; i < innerDim; ++i)
+      res =  pmadd(pset1<Packet>(lhs.coeff(row, i)), rhs.template packet<LoadMode,Packet>(i, col), res);
+  }
+};
+
+template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
+struct etor_product_packet_impl<ColMajor, Dynamic, Lhs, Rhs, Packet, LoadMode>
+{
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, Packet& res)
+  {
+    res = pset1<Packet>(typename unpacket_traits<Packet>::type(0));
+    for(Index i = 0; i < innerDim; ++i)
+      res =  pmadd(lhs.template packet<LoadMode,Packet>(row, i), pset1<Packet>(rhs.coeff(i, col)), res);
+  }
+};
+
+
+/***************************************************************************
+* Triangular products
+***************************************************************************/
+template<int Mode, bool LhsIsTriangular,
+         typename Lhs, bool LhsIsVector,
+         typename Rhs, bool RhsIsVector>
+struct triangular_product_impl;
+
+template<typename Lhs, typename Rhs, int ProductTag>
+struct generic_product_impl<Lhs,Rhs,TriangularShape,DenseShape,ProductTag>
+  : generic_product_impl_base<Lhs,Rhs,generic_product_impl<Lhs,Rhs,TriangularShape,DenseShape,ProductTag> >
+{
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  
+  template<typename Dest>
+  static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
+  {
+    triangular_product_impl<Lhs::Mode,true,typename Lhs::MatrixType,false,Rhs, Rhs::ColsAtCompileTime==1>
+        ::run(dst, lhs.nestedExpression(), rhs, alpha);
+  }
+};
+
+template<typename Lhs, typename Rhs, int ProductTag>
+struct generic_product_impl<Lhs,Rhs,DenseShape,TriangularShape,ProductTag>
+: generic_product_impl_base<Lhs,Rhs,generic_product_impl<Lhs,Rhs,DenseShape,TriangularShape,ProductTag> >
+{
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  
+  template<typename Dest>
+  static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
+  {
+    triangular_product_impl<Rhs::Mode,false,Lhs,Lhs::RowsAtCompileTime==1, typename Rhs::MatrixType, false>::run(dst, lhs, rhs.nestedExpression(), alpha);
+  }
+};
+
+
+/***************************************************************************
+* SelfAdjoint products
+***************************************************************************/
+template <typename Lhs, int LhsMode, bool LhsIsVector,
+          typename Rhs, int RhsMode, bool RhsIsVector>
+struct selfadjoint_product_impl;
+
+template<typename Lhs, typename Rhs, int ProductTag>
+struct generic_product_impl<Lhs,Rhs,SelfAdjointShape,DenseShape,ProductTag>
+  : generic_product_impl_base<Lhs,Rhs,generic_product_impl<Lhs,Rhs,SelfAdjointShape,DenseShape,ProductTag> >
+{
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  
+  template<typename Dest>
+  static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
+  {
+    selfadjoint_product_impl<typename Lhs::MatrixType,Lhs::Mode,false,Rhs,0,Rhs::IsVectorAtCompileTime>::run(dst, lhs.nestedExpression(), rhs, alpha);
+  }
+};
+
+template<typename Lhs, typename Rhs, int ProductTag>
+struct generic_product_impl<Lhs,Rhs,DenseShape,SelfAdjointShape,ProductTag>
+: generic_product_impl_base<Lhs,Rhs,generic_product_impl<Lhs,Rhs,DenseShape,SelfAdjointShape,ProductTag> >
+{
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  
+  template<typename Dest>
+  static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
+  {
+    selfadjoint_product_impl<Lhs,0,Lhs::IsVectorAtCompileTime,typename Rhs::MatrixType,Rhs::Mode,false>::run(dst, lhs, rhs.nestedExpression(), alpha);
+  }
+};
+
+
+/***************************************************************************
+* Diagonal products
+***************************************************************************/
+  
+template<typename MatrixType, typename DiagonalType, typename Derived, int ProductOrder>
+struct diagonal_product_evaluator_base
+  : evaluator_base<Derived>
+{
+   typedef typename ScalarBinaryOpTraits<typename MatrixType::Scalar, typename DiagonalType::Scalar>::ReturnType Scalar;
+public:
+  enum {
+    CoeffReadCost = NumTraits<Scalar>::MulCost + evaluator<MatrixType>::CoeffReadCost + evaluator<DiagonalType>::CoeffReadCost,
+    
+    MatrixFlags = evaluator<MatrixType>::Flags,
+    DiagFlags = evaluator<DiagonalType>::Flags,
+    _StorageOrder = MatrixFlags & RowMajorBit ? RowMajor : ColMajor,
+    _ScalarAccessOnDiag =  !((int(_StorageOrder) == ColMajor && int(ProductOrder) == OnTheLeft)
+                           ||(int(_StorageOrder) == RowMajor && int(ProductOrder) == OnTheRight)),
+    _SameTypes = is_same<typename MatrixType::Scalar, typename DiagonalType::Scalar>::value,
+    // FIXME currently we need same types, but in the future the next rule should be the one
+    //_Vectorizable = bool(int(MatrixFlags)&PacketAccessBit) && ((!_PacketOnDiag) || (_SameTypes && bool(int(DiagFlags)&PacketAccessBit))),
+    _Vectorizable = bool(int(MatrixFlags)&PacketAccessBit) && _SameTypes && (_ScalarAccessOnDiag || (bool(int(DiagFlags)&PacketAccessBit))),
+    _LinearAccessMask = (MatrixType::RowsAtCompileTime==1 || MatrixType::ColsAtCompileTime==1) ? LinearAccessBit : 0,
+    Flags = ((HereditaryBits|_LinearAccessMask) & (unsigned int)(MatrixFlags)) | (_Vectorizable ? PacketAccessBit : 0),
+    Alignment = evaluator<MatrixType>::Alignment
+  };
+  
+  diagonal_product_evaluator_base(const MatrixType &mat, const DiagonalType &diag)
+    : m_diagImpl(diag), m_matImpl(mat)
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(NumTraits<Scalar>::MulCost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+  
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index idx) const
+  {
+    return m_diagImpl.coeff(idx) * m_matImpl.coeff(idx);
+  }
+  
+protected:
+  template<int LoadMode,typename PacketType>
+  EIGEN_STRONG_INLINE PacketType packet_impl(Index row, Index col, Index id, internal::true_type) const
+  {
+    return internal::pmul(m_matImpl.template packet<LoadMode,PacketType>(row, col),
+                          internal::pset1<PacketType>(m_diagImpl.coeff(id)));
+  }
+  
+  template<int LoadMode,typename PacketType>
+  EIGEN_STRONG_INLINE PacketType packet_impl(Index row, Index col, Index id, internal::false_type) const
+  {
+    enum {
+      InnerSize = (MatrixType::Flags & RowMajorBit) ? MatrixType::ColsAtCompileTime : MatrixType::RowsAtCompileTime,
+      DiagonalPacketLoadMode = EIGEN_PLAIN_ENUM_MIN(LoadMode,((InnerSize%16) == 0) ? int(Aligned16) : int(evaluator<DiagonalType>::Alignment)) // FIXME hardcoded 16!!
+    };
+    return internal::pmul(m_matImpl.template packet<LoadMode,PacketType>(row, col),
+                          m_diagImpl.template packet<DiagonalPacketLoadMode,PacketType>(id));
+  }
+  
+  evaluator<DiagonalType> m_diagImpl;
+  evaluator<MatrixType>   m_matImpl;
+};
+
+// diagonal * dense
+template<typename Lhs, typename Rhs, int ProductKind, int ProductTag>
+struct product_evaluator<Product<Lhs, Rhs, ProductKind>, ProductTag, DiagonalShape, DenseShape>
+  : diagonal_product_evaluator_base<Rhs, typename Lhs::DiagonalVectorType, Product<Lhs, Rhs, LazyProduct>, OnTheLeft>
+{
+  typedef diagonal_product_evaluator_base<Rhs, typename Lhs::DiagonalVectorType, Product<Lhs, Rhs, LazyProduct>, OnTheLeft> Base;
+  using Base::m_diagImpl;
+  using Base::m_matImpl;
+  using Base::coeff;
+  typedef typename Base::Scalar Scalar;
+  
+  typedef Product<Lhs, Rhs, ProductKind> XprType;
+  typedef typename XprType::PlainObject PlainObject;
+  
+  enum {
+    StorageOrder = int(Rhs::Flags) & RowMajorBit ? RowMajor : ColMajor
+  };
+
+  EIGEN_DEVICE_FUNC explicit product_evaluator(const XprType& xpr)
+    : Base(xpr.rhs(), xpr.lhs().diagonal())
+  {
+  }
+  
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index row, Index col) const
+  {
+    return m_diagImpl.coeff(row) * m_matImpl.coeff(row, col);
+  }
+  
+#ifndef __CUDACC__
+  template<int LoadMode,typename PacketType>
+  EIGEN_STRONG_INLINE PacketType packet(Index row, Index col) const
+  {
+    // FIXME: NVCC used to complain about the template keyword, but we have to check whether this is still the case.
+    // See also similar calls below.
+    return this->template packet_impl<LoadMode,PacketType>(row,col, row,
+                                 typename internal::conditional<int(StorageOrder)==RowMajor, internal::true_type, internal::false_type>::type());
+  }
+  
+  template<int LoadMode,typename PacketType>
+  EIGEN_STRONG_INLINE PacketType packet(Index idx) const
+  {
+    return packet<LoadMode,PacketType>(int(StorageOrder)==ColMajor?idx:0,int(StorageOrder)==ColMajor?0:idx);
+  }
+#endif
+};
+
+// dense * diagonal
+template<typename Lhs, typename Rhs, int ProductKind, int ProductTag>
+struct product_evaluator<Product<Lhs, Rhs, ProductKind>, ProductTag, DenseShape, DiagonalShape>
+  : diagonal_product_evaluator_base<Lhs, typename Rhs::DiagonalVectorType, Product<Lhs, Rhs, LazyProduct>, OnTheRight>
+{
+  typedef diagonal_product_evaluator_base<Lhs, typename Rhs::DiagonalVectorType, Product<Lhs, Rhs, LazyProduct>, OnTheRight> Base;
+  using Base::m_diagImpl;
+  using Base::m_matImpl;
+  using Base::coeff;
+  typedef typename Base::Scalar Scalar;
+  
+  typedef Product<Lhs, Rhs, ProductKind> XprType;
+  typedef typename XprType::PlainObject PlainObject;
+  
+  enum { StorageOrder = int(Lhs::Flags) & RowMajorBit ? RowMajor : ColMajor };
+
+  EIGEN_DEVICE_FUNC explicit product_evaluator(const XprType& xpr)
+    : Base(xpr.lhs(), xpr.rhs().diagonal())
+  {
+  }
+  
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index row, Index col) const
+  {
+    return m_matImpl.coeff(row, col) * m_diagImpl.coeff(col);
+  }
+  
+#ifndef __CUDACC__
+  template<int LoadMode,typename PacketType>
+  EIGEN_STRONG_INLINE PacketType packet(Index row, Index col) const
+  {
+    return this->template packet_impl<LoadMode,PacketType>(row,col, col,
+                                 typename internal::conditional<int(StorageOrder)==ColMajor, internal::true_type, internal::false_type>::type());
+  }
+  
+  template<int LoadMode,typename PacketType>
+  EIGEN_STRONG_INLINE PacketType packet(Index idx) const
+  {
+    return packet<LoadMode,PacketType>(int(StorageOrder)==ColMajor?idx:0,int(StorageOrder)==ColMajor?0:idx);
+  }
+#endif
+};
+
+/***************************************************************************
+* Products with permutation matrices
+***************************************************************************/
+
+/** \internal
+  * \class permutation_matrix_product
+  * Internal helper class implementing the product between a permutation matrix and a matrix.
+  * This class is specialized for DenseShape below and for SparseShape in SparseCore/SparsePermutation.h
+  */
+template<typename ExpressionType, int Side, bool Transposed, typename ExpressionShape>
+struct permutation_matrix_product;
+
+template<typename ExpressionType, int Side, bool Transposed>
+struct permutation_matrix_product<ExpressionType, Side, Transposed, DenseShape>
+{
+    typedef typename nested_eval<ExpressionType, 1>::type MatrixType;
+    typedef typename remove_all<MatrixType>::type MatrixTypeCleaned;
+
+    template<typename Dest, typename PermutationType>
+    static inline void run(Dest& dst, const PermutationType& perm, const ExpressionType& xpr)
+    {
+      MatrixType mat(xpr);
+      const Index n = Side==OnTheLeft ? mat.rows() : mat.cols();
+      // FIXME we need an is_same for expression that is not sensitive to constness. For instance
+      // is_same_xpr<Block<const Matrix>, Block<Matrix> >::value should be true.
+      //if(is_same<MatrixTypeCleaned,Dest>::value && extract_data(dst) == extract_data(mat))
+      if(is_same_dense(dst, mat))
+      {
+        // apply the permutation inplace
+        Matrix<bool,PermutationType::RowsAtCompileTime,1,0,PermutationType::MaxRowsAtCompileTime> mask(perm.size());
+        mask.fill(false);
+        Index r = 0;
+        while(r < perm.size())
+        {
+          // search for the next seed
+          while(r<perm.size() && mask[r]) r++;
+          if(r>=perm.size())
+            break;
+          // we got one, let's follow it until we are back to the seed
+          Index k0 = r++;
+          Index kPrev = k0;
+          mask.coeffRef(k0) = true;
+          for(Index k=perm.indices().coeff(k0); k!=k0; k=perm.indices().coeff(k))
+          {
+                  Block<Dest, Side==OnTheLeft ? 1 : Dest::RowsAtCompileTime, Side==OnTheRight ? 1 : Dest::ColsAtCompileTime>(dst, k)
+            .swap(Block<Dest, Side==OnTheLeft ? 1 : Dest::RowsAtCompileTime, Side==OnTheRight ? 1 : Dest::ColsAtCompileTime>
+                       (dst,((Side==OnTheLeft) ^ Transposed) ? k0 : kPrev));
+
+            mask.coeffRef(k) = true;
+            kPrev = k;
+          }
+        }
+      }
+      else
+      {
+        for(Index i = 0; i < n; ++i)
+        {
+          Block<Dest, Side==OnTheLeft ? 1 : Dest::RowsAtCompileTime, Side==OnTheRight ? 1 : Dest::ColsAtCompileTime>
+               (dst, ((Side==OnTheLeft) ^ Transposed) ? perm.indices().coeff(i) : i)
+
+          =
+
+          Block<const MatrixTypeCleaned,Side==OnTheLeft ? 1 : MatrixTypeCleaned::RowsAtCompileTime,Side==OnTheRight ? 1 : MatrixTypeCleaned::ColsAtCompileTime>
+               (mat, ((Side==OnTheRight) ^ Transposed) ? perm.indices().coeff(i) : i);
+        }
+      }
+    }
+};
+
+template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape>
+struct generic_product_impl<Lhs, Rhs, PermutationShape, MatrixShape, ProductTag>
+{
+  template<typename Dest>
+  static void evalTo(Dest& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    permutation_matrix_product<Rhs, OnTheLeft, false, MatrixShape>::run(dst, lhs, rhs);
+  }
+};
+
+template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape>
+struct generic_product_impl<Lhs, Rhs, MatrixShape, PermutationShape, ProductTag>
+{
+  template<typename Dest>
+  static void evalTo(Dest& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    permutation_matrix_product<Lhs, OnTheRight, false, MatrixShape>::run(dst, rhs, lhs);
+  }
+};
+
+template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape>
+struct generic_product_impl<Inverse<Lhs>, Rhs, PermutationShape, MatrixShape, ProductTag>
+{
+  template<typename Dest>
+  static void evalTo(Dest& dst, const Inverse<Lhs>& lhs, const Rhs& rhs)
+  {
+    permutation_matrix_product<Rhs, OnTheLeft, true, MatrixShape>::run(dst, lhs.nestedExpression(), rhs);
+  }
+};
+
+template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape>
+struct generic_product_impl<Lhs, Inverse<Rhs>, MatrixShape, PermutationShape, ProductTag>
+{
+  template<typename Dest>
+  static void evalTo(Dest& dst, const Lhs& lhs, const Inverse<Rhs>& rhs)
+  {
+    permutation_matrix_product<Lhs, OnTheRight, true, MatrixShape>::run(dst, rhs.nestedExpression(), lhs);
+  }
+};
+
+
+/***************************************************************************
+* Products with transpositions matrices
+***************************************************************************/
+
+// FIXME could we unify Transpositions and Permutation into a single "shape"??
+
+/** \internal
+  * \class transposition_matrix_product
+  * Internal helper class implementing the product between a permutation matrix and a matrix.
+  */
+template<typename ExpressionType, int Side, bool Transposed, typename ExpressionShape>
+struct transposition_matrix_product
+{
+  typedef typename nested_eval<ExpressionType, 1>::type MatrixType;
+  typedef typename remove_all<MatrixType>::type MatrixTypeCleaned;
+  
+  template<typename Dest, typename TranspositionType>
+  static inline void run(Dest& dst, const TranspositionType& tr, const ExpressionType& xpr)
+  {
+    MatrixType mat(xpr);
+    typedef typename TranspositionType::StorageIndex StorageIndex;
+    const Index size = tr.size();
+    StorageIndex j = 0;
+
+    if(!is_same_dense(dst,mat))
+      dst = mat;
+
+    for(Index k=(Transposed?size-1:0) ; Transposed?k>=0:k<size ; Transposed?--k:++k)
+      if(Index(j=tr.coeff(k))!=k)
+      {
+        if(Side==OnTheLeft)        dst.row(k).swap(dst.row(j));
+        else if(Side==OnTheRight)  dst.col(k).swap(dst.col(j));
+      }
+  }
+};
+
+template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape>
+struct generic_product_impl<Lhs, Rhs, TranspositionsShape, MatrixShape, ProductTag>
+{
+  template<typename Dest>
+  static void evalTo(Dest& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    transposition_matrix_product<Rhs, OnTheLeft, false, MatrixShape>::run(dst, lhs, rhs);
+  }
+};
+
+template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape>
+struct generic_product_impl<Lhs, Rhs, MatrixShape, TranspositionsShape, ProductTag>
+{
+  template<typename Dest>
+  static void evalTo(Dest& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    transposition_matrix_product<Lhs, OnTheRight, false, MatrixShape>::run(dst, rhs, lhs);
+  }
+};
+
+
+template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape>
+struct generic_product_impl<Transpose<Lhs>, Rhs, TranspositionsShape, MatrixShape, ProductTag>
+{
+  template<typename Dest>
+  static void evalTo(Dest& dst, const Transpose<Lhs>& lhs, const Rhs& rhs)
+  {
+    transposition_matrix_product<Rhs, OnTheLeft, true, MatrixShape>::run(dst, lhs.nestedExpression(), rhs);
+  }
+};
+
+template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape>
+struct generic_product_impl<Lhs, Transpose<Rhs>, MatrixShape, TranspositionsShape, ProductTag>
+{
+  template<typename Dest>
+  static void evalTo(Dest& dst, const Lhs& lhs, const Transpose<Rhs>& rhs)
+  {
+    transposition_matrix_product<Lhs, OnTheRight, true, MatrixShape>::run(dst, rhs.nestedExpression(), lhs);
+  }
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_PRODUCT_EVALUATORS_H
diff --git a/eigen/Eigen/src/Core/Random.h b/eigen/Eigen/src/Core/Random.h
index 480fea4..486e9ed 100644
--- a/eigen/Eigen/src/Core/Random.h
+++ b/eigen/Eigen/src/Core/Random.h
@@ -16,8 +16,7 @@ namespace internal {
 
 template<typename Scalar> struct scalar_random_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_random_op)
-  template<typename Index>
-  inline const Scalar operator() (Index, Index = 0) const { return random<Scalar>(); }
+  inline const Scalar operator() () const { return random<Scalar>(); }
 };
 
 template<typename Scalar>
@@ -28,12 +27,18 @@ struct functor_traits<scalar_random_op<Scalar> >
 
 /** \returns a random matrix expression
   *
+  * Numbers are uniformly spread through their whole definition range for integer types,
+  * and in the [-1:1] range for floating point scalar types.
+  * 
   * The parameters \a rows and \a cols are the number of rows and of columns of
   * the returned matrix. Must be compatible with this MatrixBase type.
   *
+  * \not_reentrant
+  * 
   * This variant is meant to be used for dynamic-size matrix types. For fixed-size types,
   * it is redundant to pass \a rows and \a cols as arguments, so Random() should be used
   * instead.
+  * 
   *
   * Example: \include MatrixBase_random_int_int.cpp
   * Output: \verbinclude MatrixBase_random_int_int.out
@@ -41,11 +46,13 @@ struct functor_traits<scalar_random_op<Scalar> >
   * This expression has the "evaluate before nesting" flag so that it will be evaluated into
   * a temporary matrix whenever it is nested in a larger expression. This prevents unexpected
   * behavior with expressions involving random matrices.
+  * 
+  * See DenseBase::NullaryExpr(Index, const CustomNullaryOp&) for an example using C++11 random generators.
   *
-  * \sa MatrixBase::setRandom(), MatrixBase::Random(Index), MatrixBase::Random()
+  * \sa DenseBase::setRandom(), DenseBase::Random(Index), DenseBase::Random()
   */
 template<typename Derived>
-inline const CwiseNullaryOp<internal::scalar_random_op<typename internal::traits<Derived>::Scalar>, Derived>
+inline const typename DenseBase<Derived>::RandomReturnType
 DenseBase<Derived>::Random(Index rows, Index cols)
 {
   return NullaryExpr(rows, cols, internal::scalar_random_op<Scalar>());
@@ -53,10 +60,14 @@ DenseBase<Derived>::Random(Index rows, Index cols)
 
 /** \returns a random vector expression
   *
+  * Numbers are uniformly spread through their whole definition range for integer types,
+  * and in the [-1:1] range for floating point scalar types.
+  *
   * The parameter \a size is the size of the returned vector.
   * Must be compatible with this MatrixBase type.
   *
   * \only_for_vectors
+  * \not_reentrant
   *
   * This variant is meant to be used for dynamic-size vector types. For fixed-size types,
   * it is redundant to pass \a size as argument, so Random() should be used
@@ -69,10 +80,10 @@ DenseBase<Derived>::Random(Index rows, Index cols)
   * a temporary vector whenever it is nested in a larger expression. This prevents unexpected
   * behavior with expressions involving random matrices.
   *
-  * \sa MatrixBase::setRandom(), MatrixBase::Random(Index,Index), MatrixBase::Random()
+  * \sa DenseBase::setRandom(), DenseBase::Random(Index,Index), DenseBase::Random()
   */
 template<typename Derived>
-inline const CwiseNullaryOp<internal::scalar_random_op<typename internal::traits<Derived>::Scalar>, Derived>
+inline const typename DenseBase<Derived>::RandomReturnType
 DenseBase<Derived>::Random(Index size)
 {
   return NullaryExpr(size, internal::scalar_random_op<Scalar>());
@@ -80,6 +91,9 @@ DenseBase<Derived>::Random(Index size)
 
 /** \returns a fixed-size random matrix or vector expression
   *
+  * Numbers are uniformly spread through their whole definition range for integer types,
+  * and in the [-1:1] range for floating point scalar types.
+  * 
   * This variant is only for fixed-size MatrixBase types. For dynamic-size types, you
   * need to use the variants taking size arguments.
   *
@@ -89,11 +103,13 @@ DenseBase<Derived>::Random(Index size)
   * This expression has the "evaluate before nesting" flag so that it will be evaluated into
   * a temporary matrix whenever it is nested in a larger expression. This prevents unexpected
   * behavior with expressions involving random matrices.
+  * 
+  * \not_reentrant
   *
-  * \sa MatrixBase::setRandom(), MatrixBase::Random(Index,Index), MatrixBase::Random(Index)
+  * \sa DenseBase::setRandom(), DenseBase::Random(Index,Index), DenseBase::Random(Index)
   */
 template<typename Derived>
-inline const CwiseNullaryOp<internal::scalar_random_op<typename internal::traits<Derived>::Scalar>, Derived>
+inline const typename DenseBase<Derived>::RandomReturnType
 DenseBase<Derived>::Random()
 {
   return NullaryExpr(RowsAtCompileTime, ColsAtCompileTime, internal::scalar_random_op<Scalar>());
@@ -101,25 +117,34 @@ DenseBase<Derived>::Random()
 
 /** Sets all coefficients in this expression to random values.
   *
+  * Numbers are uniformly spread through their whole definition range for integer types,
+  * and in the [-1:1] range for floating point scalar types.
+  * 
+  * \not_reentrant
+  * 
   * Example: \include MatrixBase_setRandom.cpp
   * Output: \verbinclude MatrixBase_setRandom.out
   *
   * \sa class CwiseNullaryOp, setRandom(Index), setRandom(Index,Index)
   */
 template<typename Derived>
-inline Derived& DenseBase<Derived>::setRandom()
+EIGEN_DEVICE_FUNC inline Derived& DenseBase<Derived>::setRandom()
 {
   return *this = Random(rows(), cols());
 }
 
 /** Resizes to the given \a newSize, and sets all coefficients in this expression to random values.
   *
+  * Numbers are uniformly spread through their whole definition range for integer types,
+  * and in the [-1:1] range for floating point scalar types.
+  * 
   * \only_for_vectors
+  * \not_reentrant
   *
   * Example: \include Matrix_setRandom_int.cpp
   * Output: \verbinclude Matrix_setRandom_int.out
   *
-  * \sa MatrixBase::setRandom(), setRandom(Index,Index), class CwiseNullaryOp, MatrixBase::Random()
+  * \sa DenseBase::setRandom(), setRandom(Index,Index), class CwiseNullaryOp, DenseBase::Random()
   */
 template<typename Derived>
 EIGEN_STRONG_INLINE Derived&
@@ -131,19 +156,24 @@ PlainObjectBase<Derived>::setRandom(Index newSize)
 
 /** Resizes to the given size, and sets all coefficients in this expression to random values.
   *
-  * \param nbRows the new number of rows
-  * \param nbCols the new number of columns
+  * Numbers are uniformly spread through their whole definition range for integer types,
+  * and in the [-1:1] range for floating point scalar types.
+  *
+  * \not_reentrant
+  * 
+  * \param rows the new number of rows
+  * \param cols the new number of columns
   *
   * Example: \include Matrix_setRandom_int_int.cpp
   * Output: \verbinclude Matrix_setRandom_int_int.out
   *
-  * \sa MatrixBase::setRandom(), setRandom(Index), class CwiseNullaryOp, MatrixBase::Random()
+  * \sa DenseBase::setRandom(), setRandom(Index), class CwiseNullaryOp, DenseBase::Random()
   */
 template<typename Derived>
 EIGEN_STRONG_INLINE Derived&
-PlainObjectBase<Derived>::setRandom(Index nbRows, Index nbCols)
+PlainObjectBase<Derived>::setRandom(Index rows, Index cols)
 {
-  resize(nbRows, nbCols);
+  resize(rows, cols);
   return setRandom();
 }
 
diff --git a/eigen/Eigen/src/Core/Redux.h b/eigen/Eigen/src/Core/Redux.h
index 9b8662a..2b5b73b 100644
--- a/eigen/Eigen/src/Core/Redux.h
+++ b/eigen/Eigen/src/Core/Redux.h
@@ -27,8 +27,9 @@ template<typename Func, typename Derived>
 struct redux_traits
 {
 public:
+    typedef typename find_best_packet<typename Derived::Scalar,Derived::SizeAtCompileTime>::type PacketType;
   enum {
-    PacketSize = packet_traits<typename Derived::Scalar>::size,
+    PacketSize = unpacket_traits<PacketType>::size,
     InnerMaxSize = int(Derived::IsRowMajor)
                  ? Derived::MaxColsAtCompileTime
                  : Derived::MaxRowsAtCompileTime
@@ -37,8 +38,8 @@ public:
   enum {
     MightVectorize = (int(Derived::Flags)&ActualPacketAccessBit)
                   && (functor_traits<Func>::PacketAccess),
-    MayLinearVectorize = MightVectorize && (int(Derived::Flags)&LinearAccessBit),
-    MaySliceVectorize  = MightVectorize && int(InnerMaxSize)>=3*PacketSize
+    MayLinearVectorize = bool(MightVectorize) && (int(Derived::Flags)&LinearAccessBit),
+    MaySliceVectorize  = bool(MightVectorize) && int(InnerMaxSize)>=3*PacketSize
   };
 
 public:
@@ -50,21 +51,34 @@ public:
 
 public:
   enum {
-    Cost = (  Derived::SizeAtCompileTime == Dynamic
-           || Derived::CoeffReadCost == Dynamic
-           || (Derived::SizeAtCompileTime!=1 && functor_traits<Func>::Cost == Dynamic)
-           ) ? Dynamic
-           : Derived::SizeAtCompileTime * Derived::CoeffReadCost
-               + (Derived::SizeAtCompileTime-1) * functor_traits<Func>::Cost,
+    Cost = Derived::SizeAtCompileTime == Dynamic ? HugeCost
+         : Derived::SizeAtCompileTime * Derived::CoeffReadCost + (Derived::SizeAtCompileTime-1) * functor_traits<Func>::Cost,
     UnrollingLimit = EIGEN_UNROLLING_LIMIT * (int(Traversal) == int(DefaultTraversal) ? 1 : int(PacketSize))
   };
 
 public:
   enum {
-    Unrolling = Cost != Dynamic && Cost <= UnrollingLimit
-              ? CompleteUnrolling
-              : NoUnrolling
+    Unrolling = Cost <= UnrollingLimit ? CompleteUnrolling : NoUnrolling
   };
+  
+#ifdef EIGEN_DEBUG_ASSIGN
+  static void debug()
+  {
+    std::cerr << "Xpr: " << typeid(typename Derived::XprType).name() << std::endl;
+    std::cerr.setf(std::ios::hex, std::ios::basefield);
+    EIGEN_DEBUG_VAR(Derived::Flags)
+    std::cerr.unsetf(std::ios::hex);
+    EIGEN_DEBUG_VAR(InnerMaxSize)
+    EIGEN_DEBUG_VAR(PacketSize)
+    EIGEN_DEBUG_VAR(MightVectorize)
+    EIGEN_DEBUG_VAR(MayLinearVectorize)
+    EIGEN_DEBUG_VAR(MaySliceVectorize)
+    EIGEN_DEBUG_VAR(Traversal)
+    EIGEN_DEBUG_VAR(UnrollingLimit)
+    EIGEN_DEBUG_VAR(Unrolling)
+    std::cerr << std::endl;
+  }
+#endif
 };
 
 /***************************************************************************
@@ -82,6 +96,7 @@ struct redux_novec_unroller
 
   typedef typename Derived::Scalar Scalar;
 
+  EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func)
   {
     return func(redux_novec_unroller<Func, Derived, Start, HalfLength>::run(mat,func),
@@ -99,6 +114,7 @@ struct redux_novec_unroller<Func, Derived, Start, 1>
 
   typedef typename Derived::Scalar Scalar;
 
+  EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func&)
   {
     return mat.coeffByOuterInner(outer, inner);
@@ -112,6 +128,7 @@ template<typename Func, typename Derived, int Start>
 struct redux_novec_unroller<Func, Derived, Start, 0>
 {
   typedef typename Derived::Scalar Scalar;
+  EIGEN_DEVICE_FUNC 
   static EIGEN_STRONG_INLINE Scalar run(const Derived&, const Func&) { return Scalar(); }
 };
 
@@ -121,12 +138,12 @@ template<typename Func, typename Derived, int Start, int Length>
 struct redux_vec_unroller
 {
   enum {
-    PacketSize = packet_traits<typename Derived::Scalar>::size,
+    PacketSize = redux_traits<Func, Derived>::PacketSize,
     HalfLength = Length/2
   };
 
   typedef typename Derived::Scalar Scalar;
-  typedef typename packet_traits<Scalar>::type PacketScalar;
+  typedef typename redux_traits<Func, Derived>::PacketType PacketScalar;
 
   static EIGEN_STRONG_INLINE PacketScalar run(const Derived &mat, const Func& func)
   {
@@ -140,18 +157,18 @@ template<typename Func, typename Derived, int Start>
 struct redux_vec_unroller<Func, Derived, Start, 1>
 {
   enum {
-    index = Start * packet_traits<typename Derived::Scalar>::size,
+    index = Start * redux_traits<Func, Derived>::PacketSize,
     outer = index / int(Derived::InnerSizeAtCompileTime),
     inner = index % int(Derived::InnerSizeAtCompileTime),
-    alignment = (Derived::Flags & AlignedBit) ? Aligned : Unaligned
+    alignment = Derived::Alignment
   };
 
   typedef typename Derived::Scalar Scalar;
-  typedef typename packet_traits<Scalar>::type PacketScalar;
+  typedef typename redux_traits<Func, Derived>::PacketType PacketScalar;
 
   static EIGEN_STRONG_INLINE PacketScalar run(const Derived &mat, const Func&)
   {
-    return mat.template packetByOuterInner<alignment>(outer, inner);
+    return mat.template packetByOuterInner<alignment,PacketScalar>(outer, inner);
   }
 };
 
@@ -169,8 +186,8 @@ template<typename Func, typename Derived>
 struct redux_impl<Func, Derived, DefaultTraversal, NoUnrolling>
 {
   typedef typename Derived::Scalar Scalar;
-  typedef typename Derived::Index Index;
-  static EIGEN_STRONG_INLINE Scalar run(const Derived& mat, const Func& func)
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func)
   {
     eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix");
     Scalar res;
@@ -193,19 +210,19 @@ template<typename Func, typename Derived>
 struct redux_impl<Func, Derived, LinearVectorizedTraversal, NoUnrolling>
 {
   typedef typename Derived::Scalar Scalar;
-  typedef typename packet_traits<Scalar>::type PacketScalar;
-  typedef typename Derived::Index Index;
+  typedef typename redux_traits<Func, Derived>::PacketType PacketScalar;
 
-  static Scalar run(const Derived& mat, const Func& func)
+  static Scalar run(const Derived &mat, const Func& func)
   {
     const Index size = mat.size();
-    eigen_assert(size && "you are using an empty matrix");
-    const Index packetSize = packet_traits<Scalar>::size;
-    const Index alignedStart = internal::first_aligned(mat);
+    
+    const Index packetSize = redux_traits<Func, Derived>::PacketSize;
+    const int packetAlignment = unpacket_traits<PacketScalar>::alignment;
     enum {
-      alignment = bool(Derived::Flags & DirectAccessBit) || bool(Derived::Flags & AlignedBit)
-                ? Aligned : Unaligned
+      alignment0 = (bool(Derived::Flags & DirectAccessBit) && bool(packet_traits<Scalar>::AlignedOnScalar)) ? int(packetAlignment) : int(Unaligned),
+      alignment = EIGEN_PLAIN_ENUM_MAX(alignment0, Derived::Alignment)
     };
+    const Index alignedStart = internal::first_default_aligned(mat.nestedExpression());
     const Index alignedSize2 = ((size-alignedStart)/(2*packetSize))*(2*packetSize);
     const Index alignedSize = ((size-alignedStart)/(packetSize))*(packetSize);
     const Index alignedEnd2 = alignedStart + alignedSize2;
@@ -213,19 +230,19 @@ struct redux_impl<Func, Derived, LinearVectorizedTraversal, NoUnrolling>
     Scalar res;
     if(alignedSize)
     {
-      PacketScalar packet_res0 = mat.template packet<alignment>(alignedStart);
+      PacketScalar packet_res0 = mat.template packet<alignment,PacketScalar>(alignedStart);
       if(alignedSize>packetSize) // we have at least two packets to partly unroll the loop
       {
-        PacketScalar packet_res1 = mat.template packet<alignment>(alignedStart+packetSize);
+        PacketScalar packet_res1 = mat.template packet<alignment,PacketScalar>(alignedStart+packetSize);
         for(Index index = alignedStart + 2*packetSize; index < alignedEnd2; index += 2*packetSize)
         {
-          packet_res0 = func.packetOp(packet_res0, mat.template packet<alignment>(index));
-          packet_res1 = func.packetOp(packet_res1, mat.template packet<alignment>(index+packetSize));
+          packet_res0 = func.packetOp(packet_res0, mat.template packet<alignment,PacketScalar>(index));
+          packet_res1 = func.packetOp(packet_res1, mat.template packet<alignment,PacketScalar>(index+packetSize));
         }
 
         packet_res0 = func.packetOp(packet_res0,packet_res1);
         if(alignedEnd>alignedEnd2)
-          packet_res0 = func.packetOp(packet_res0, mat.template packet<alignment>(alignedEnd2));
+          packet_res0 = func.packetOp(packet_res0, mat.template packet<alignment,PacketScalar>(alignedEnd2));
       }
       res = func.predux(packet_res0);
 
@@ -252,25 +269,24 @@ template<typename Func, typename Derived, int Unrolling>
 struct redux_impl<Func, Derived, SliceVectorizedTraversal, Unrolling>
 {
   typedef typename Derived::Scalar Scalar;
-  typedef typename packet_traits<Scalar>::type PacketScalar;
-  typedef typename Derived::Index Index;
+  typedef typename redux_traits<Func, Derived>::PacketType PacketType;
 
-  static Scalar run(const Derived& mat, const Func& func)
+  EIGEN_DEVICE_FUNC static Scalar run(const Derived &mat, const Func& func)
   {
     eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix");
     const Index innerSize = mat.innerSize();
     const Index outerSize = mat.outerSize();
     enum {
-      packetSize = packet_traits<Scalar>::size
+      packetSize = redux_traits<Func, Derived>::PacketSize
     };
     const Index packetedInnerSize = ((innerSize)/packetSize)*packetSize;
     Scalar res;
     if(packetedInnerSize)
     {
-      PacketScalar packet_res = mat.template packet<Unaligned>(0,0);
+      PacketType packet_res = mat.template packet<Unaligned,PacketType>(0,0);
       for(Index j=0; j<outerSize; ++j)
         for(Index i=(j==0?packetSize:0); i<packetedInnerSize; i+=Index(packetSize))
-          packet_res = func.packetOp(packet_res, mat.template packetByOuterInner<Unaligned>(j,i));
+          packet_res = func.packetOp(packet_res, mat.template packetByOuterInner<Unaligned,PacketType>(j,i));
 
       res = func.predux(packet_res);
       for(Index j=0; j<outerSize; ++j)
@@ -291,22 +307,90 @@ template<typename Func, typename Derived>
 struct redux_impl<Func, Derived, LinearVectorizedTraversal, CompleteUnrolling>
 {
   typedef typename Derived::Scalar Scalar;
-  typedef typename packet_traits<Scalar>::type PacketScalar;
+
+  typedef typename redux_traits<Func, Derived>::PacketType PacketScalar;
   enum {
-    PacketSize = packet_traits<Scalar>::size,
+    PacketSize = redux_traits<Func, Derived>::PacketSize,
     Size = Derived::SizeAtCompileTime,
     VectorizedSize = (Size / PacketSize) * PacketSize
   };
-  static EIGEN_STRONG_INLINE Scalar run(const Derived& mat, const Func& func)
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func)
   {
     eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix");
-    Scalar res = func.predux(redux_vec_unroller<Func, Derived, 0, Size / PacketSize>::run(mat,func));
-    if (VectorizedSize != Size)
-      res = func(res,redux_novec_unroller<Func, Derived, VectorizedSize, Size-VectorizedSize>::run(mat,func));
-    return res;
+    if (VectorizedSize > 0) {
+      Scalar res = func.predux(redux_vec_unroller<Func, Derived, 0, Size / PacketSize>::run(mat,func));
+      if (VectorizedSize != Size)
+        res = func(res,redux_novec_unroller<Func, Derived, VectorizedSize, Size-VectorizedSize>::run(mat,func));
+      return res;
+    }
+    else {
+      return redux_novec_unroller<Func, Derived, 0, Size>::run(mat,func);
+    }
   }
 };
 
+// evaluator adaptor
+template<typename _XprType>
+class redux_evaluator
+{
+public:
+  typedef _XprType XprType;
+  EIGEN_DEVICE_FUNC explicit redux_evaluator(const XprType &xpr) : m_evaluator(xpr), m_xpr(xpr) {}
+  
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename XprType::PacketScalar PacketScalar;
+  typedef typename XprType::PacketReturnType PacketReturnType;
+  
+  enum {
+    MaxRowsAtCompileTime = XprType::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = XprType::MaxColsAtCompileTime,
+    // TODO we should not remove DirectAccessBit and rather find an elegant way to query the alignment offset at runtime from the evaluator
+    Flags = evaluator<XprType>::Flags & ~DirectAccessBit,
+    IsRowMajor = XprType::IsRowMajor,
+    SizeAtCompileTime = XprType::SizeAtCompileTime,
+    InnerSizeAtCompileTime = XprType::InnerSizeAtCompileTime,
+    CoeffReadCost = evaluator<XprType>::CoeffReadCost,
+    Alignment = evaluator<XprType>::Alignment
+  };
+  
+  EIGEN_DEVICE_FUNC Index rows() const { return m_xpr.rows(); }
+  EIGEN_DEVICE_FUNC Index cols() const { return m_xpr.cols(); }
+  EIGEN_DEVICE_FUNC Index size() const { return m_xpr.size(); }
+  EIGEN_DEVICE_FUNC Index innerSize() const { return m_xpr.innerSize(); }
+  EIGEN_DEVICE_FUNC Index outerSize() const { return m_xpr.outerSize(); }
+
+  EIGEN_DEVICE_FUNC
+  CoeffReturnType coeff(Index row, Index col) const
+  { return m_evaluator.coeff(row, col); }
+
+  EIGEN_DEVICE_FUNC
+  CoeffReturnType coeff(Index index) const
+  { return m_evaluator.coeff(index); }
+
+  template<int LoadMode, typename PacketType>
+  PacketType packet(Index row, Index col) const
+  { return m_evaluator.template packet<LoadMode,PacketType>(row, col); }
+
+  template<int LoadMode, typename PacketType>
+  PacketType packet(Index index) const
+  { return m_evaluator.template packet<LoadMode,PacketType>(index); }
+  
+  EIGEN_DEVICE_FUNC
+  CoeffReturnType coeffByOuterInner(Index outer, Index inner) const
+  { return m_evaluator.coeff(IsRowMajor ? outer : inner, IsRowMajor ? inner : outer); }
+  
+  template<int LoadMode, typename PacketType>
+  PacketType packetByOuterInner(Index outer, Index inner) const
+  { return m_evaluator.template packet<LoadMode,PacketType>(IsRowMajor ? outer : inner, IsRowMajor ? inner : outer); }
+  
+  const XprType & nestedExpression() const { return m_xpr; }
+  
+protected:
+  internal::evaluator<XprType> m_evaluator;
+  const XprType &m_xpr;
+};
+
 } // end namespace internal
 
 /***************************************************************************
@@ -317,51 +401,56 @@ struct redux_impl<Func, Derived, LinearVectorizedTraversal, CompleteUnrolling>
 /** \returns the result of a full redux operation on the whole matrix or vector using \a func
   *
   * The template parameter \a BinaryOp is the type of the functor \a func which must be
-  * an associative operator. Both current STL and TR1 functor styles are handled.
+  * an associative operator. Both current C++98 and C++11 functor styles are handled.
   *
   * \sa DenseBase::sum(), DenseBase::minCoeff(), DenseBase::maxCoeff(), MatrixBase::colwise(), MatrixBase::rowwise()
   */
 template<typename Derived>
 template<typename Func>
-EIGEN_STRONG_INLINE typename internal::result_of<Func(typename internal::traits<Derived>::Scalar)>::type
+EIGEN_DEVICE_FUNC typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::redux(const Func& func) const
 {
-  typedef typename internal::remove_all<typename Derived::Nested>::type ThisNested;
-  return internal::redux_impl<Func, ThisNested>
-            ::run(derived(), func);
+  eigen_assert(this->rows()>0 && this->cols()>0 && "you are using an empty matrix");
+
+  typedef typename internal::redux_evaluator<Derived> ThisEvaluator;
+  ThisEvaluator thisEval(derived());
+  
+  return internal::redux_impl<Func, ThisEvaluator>::run(thisEval, func);
 }
 
 /** \returns the minimum of all coefficients of \c *this.
   * \warning the result is undefined if \c *this contains NaN.
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::minCoeff() const
 {
-  return this->redux(Eigen::internal::scalar_min_op<Scalar>());
+  return derived().redux(Eigen::internal::scalar_min_op<Scalar,Scalar>());
 }
 
 /** \returns the maximum of all coefficients of \c *this.
   * \warning the result is undefined if \c *this contains NaN.
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::maxCoeff() const
 {
-  return this->redux(Eigen::internal::scalar_max_op<Scalar>());
+  return derived().redux(Eigen::internal::scalar_max_op<Scalar,Scalar>());
 }
 
-/** \returns the sum of all coefficients of *this
+/** \returns the sum of all coefficients of \c *this
+  *
+  * If \c *this is empty, then the value 0 is returned.
   *
   * \sa trace(), prod(), mean()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::sum() const
 {
   if(SizeAtCompileTime==0 || (SizeAtCompileTime==Dynamic && size()==0))
     return Scalar(0);
-  return this->redux(Eigen::internal::scalar_sum_op<Scalar>());
+  return derived().redux(Eigen::internal::scalar_sum_op<Scalar,Scalar>());
 }
 
 /** \returns the mean of all coefficients of *this
@@ -369,10 +458,17 @@ DenseBase<Derived>::sum() const
 * \sa trace(), prod(), sum()
 */
 template<typename Derived>
-EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::mean() const
 {
-  return Scalar(this->redux(Eigen::internal::scalar_sum_op<Scalar>())) / Scalar(this->size());
+#ifdef __INTEL_COMPILER
+  #pragma warning push
+  #pragma warning ( disable : 2259 )
+#endif
+  return Scalar(derived().redux(Eigen::internal::scalar_sum_op<Scalar,Scalar>())) / Scalar(this->size());
+#ifdef __INTEL_COMPILER
+  #pragma warning pop
+#endif
 }
 
 /** \returns the product of all coefficients of *this
@@ -383,12 +479,12 @@ DenseBase<Derived>::mean() const
   * \sa sum(), mean(), trace()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::prod() const
 {
   if(SizeAtCompileTime==0 || (SizeAtCompileTime==Dynamic && size()==0))
     return Scalar(1);
-  return this->redux(Eigen::internal::scalar_product_op<Scalar>());
+  return derived().redux(Eigen::internal::scalar_product_op<Scalar>());
 }
 
 /** \returns the trace of \c *this, i.e. the sum of the coefficients on the main diagonal.
@@ -398,7 +494,7 @@ DenseBase<Derived>::prod() const
   * \sa diagonal(), sum()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
 MatrixBase<Derived>::trace() const
 {
   return derived().diagonal().sum();
diff --git a/eigen/Eigen/src/Core/Ref.h b/eigen/Eigen/src/Core/Ref.h
index 7a3beca..abb1e51 100644
--- a/eigen/Eigen/src/Core/Ref.h
+++ b/eigen/Eigen/src/Core/Ref.h
@@ -12,79 +12,6 @@
 
 namespace Eigen { 
 
-template<typename Derived> class RefBase;
-template<typename PlainObjectType, int Options = 0,
-         typename StrideType = typename internal::conditional<PlainObjectType::IsVectorAtCompileTime,InnerStride<1>,OuterStride<> >::type > class Ref;
-
-/** \class Ref
-  * \ingroup Core_Module
-  *
-  * \brief A matrix or vector expression mapping an existing expressions
-  *
-  * \tparam PlainObjectType the equivalent matrix type of the mapped data
-  * \tparam Options specifies whether the pointer is \c #Aligned, or \c #Unaligned.
-  *                The default is \c #Unaligned.
-  * \tparam StrideType optionally specifies strides. By default, Ref implies a contiguous storage along the inner dimension (inner stride==1),
-  *                   but accept a variable outer stride (leading dimension).
-  *                   This can be overridden by specifying strides.
-  *                   The type passed here must be a specialization of the Stride template, see examples below.
-  *
-  * This class permits to write non template functions taking Eigen's object as parameters while limiting the number of copies.
-  * A Ref<> object can represent either a const expression or a l-value:
-  * \code
-  * // in-out argument:
-  * void foo1(Ref<VectorXf> x);
-  *
-  * // read-only const argument:
-  * void foo2(const Ref<const VectorXf>& x);
-  * \endcode
-  *
-  * In the in-out case, the input argument must satisfies the constraints of the actual Ref<> type, otherwise a compilation issue will be triggered.
-  * By default, a Ref<VectorXf> can reference any dense vector expression of float having a contiguous memory layout.
-  * Likewise, a Ref<MatrixXf> can reference any column major dense matrix expression of float whose column's elements are contiguously stored with
-  * the possibility to have a constant space inbetween each column, i.e.: the inner stride mmust be equal to 1, but the outer-stride (or leading dimension),
-  * can be greater than the number of rows.
-  *
-  * In the const case, if the input expression does not match the above requirement, then it is evaluated into a temporary before being passed to the function.
-  * Here are some examples:
-  * \code
-  * MatrixXf A;
-  * VectorXf a;
-  * foo1(a.head());             // OK
-  * foo1(A.col());              // OK
-  * foo1(A.row());              // compilation error because here innerstride!=1
-  * foo2(A.row());              // The row is copied into a contiguous temporary
-  * foo2(2*a);                  // The expression is evaluated into a temporary
-  * foo2(A.col().segment(2,4)); // No temporary
-  * \endcode
-  *
-  * The range of inputs that can be referenced without temporary can be enlarged using the last two template parameter.
-  * Here is an example accepting an innerstride!=1:
-  * \code
-  * // in-out argument:
-  * void foo3(Ref<VectorXf,0,InnerStride<> > x);
-  * foo3(A.row());              // OK
-  * \endcode
-  * The downside here is that the function foo3 might be significantly slower than foo1 because it won't be able to exploit vectorization, and will involved more
-  * expensive address computations even if the input is contiguously stored in memory. To overcome this issue, one might propose to overloads internally calling a
-  * template function, e.g.:
-  * \code
-  * // in the .h:
-  * void foo(const Ref<MatrixXf>& A);
-  * void foo(const Ref<MatrixXf,0,Stride<> >& A);
-  *
-  * // in the .cpp:
-  * template<typename TypeOfA> void foo_impl(const TypeOfA& A) {
-  *     ... // crazy code goes here
-  * }
-  * void foo(const Ref<MatrixXf>& A) { foo_impl(A); }
-  * void foo(const Ref<MatrixXf,0,Stride<> >& A) { foo_impl(A); }
-  * \endcode
-  *
-  *
-  * \sa PlainObjectBase::Map(), \ref TopicStorageOrders
-  */
-
 namespace internal {
 
 template<typename _PlainObjectType, int _Options, typename _StrideType>
@@ -95,7 +22,8 @@ struct traits<Ref<_PlainObjectType, _Options, _StrideType> >
   typedef _StrideType StrideType;
   enum {
     Options = _Options,
-    Flags = traits<Map<_PlainObjectType, _Options, _StrideType> >::Flags | NestByRefBit
+    Flags = traits<Map<_PlainObjectType, _Options, _StrideType> >::Flags | NestByRefBit,
+    Alignment = traits<Map<_PlainObjectType, _Options, _StrideType> >::Alignment
   };
 
   template<typename Derived> struct match {
@@ -107,7 +35,13 @@ struct traits<Ref<_PlainObjectType, _Options, _StrideType> >
                       || (int(StrideType::InnerStrideAtCompileTime)==0 && int(Derived::InnerStrideAtCompileTime)==1),
       OuterStrideMatch = Derived::IsVectorAtCompileTime
                       || int(StrideType::OuterStrideAtCompileTime)==int(Dynamic) || int(StrideType::OuterStrideAtCompileTime)==int(Derived::OuterStrideAtCompileTime),
-      AlignmentMatch = (_Options!=Aligned) || ((PlainObjectType::Flags&AlignedBit)==0) || ((traits<Derived>::Flags&AlignedBit)==AlignedBit),
+      // NOTE, this indirection of evaluator<Derived>::Alignment is needed
+      // to workaround a very strange bug in MSVC related to the instantiation
+      // of has_*ary_operator in evaluator<CwiseNullaryOp>.
+      // This line is surprisingly very sensitive. For instance, simply adding parenthesis
+      // as "DerivedAlignment = (int(evaluator<Derived>::Alignment))," will make MSVC fail...
+      DerivedAlignment = int(evaluator<Derived>::Alignment),
+      AlignmentMatch = (int(traits<PlainObjectType>::Alignment)==int(Unaligned)) || (DerivedAlignment >= int(Alignment)), // FIXME the first condition is not very clear, it should be replaced by the required alignment
       ScalarTypeMatch = internal::is_same<typename PlainObjectType::Scalar, typename Derived::Scalar>::value,
       MatchAtCompileTime = HasDirectAccess && StorageOrderMatch && InnerStrideMatch && OuterStrideMatch && AlignmentMatch && ScalarTypeMatch
     };
@@ -132,12 +66,12 @@ public:
   typedef MapBase<Derived> Base;
   EIGEN_DENSE_PUBLIC_INTERFACE(RefBase)
 
-  inline Index innerStride() const
+  EIGEN_DEVICE_FUNC inline Index innerStride() const
   {
     return StrideType::InnerStrideAtCompileTime != 0 ? m_stride.inner() : 1;
   }
 
-  inline Index outerStride() const
+  EIGEN_DEVICE_FUNC inline Index outerStride() const
   {
     return StrideType::OuterStrideAtCompileTime != 0 ? m_stride.outer()
          : IsVectorAtCompileTime ? this->size()
@@ -145,7 +79,7 @@ public:
          : this->rows();
   }
 
-  RefBase()
+  EIGEN_DEVICE_FUNC RefBase()
     : Base(0,RowsAtCompileTime==Dynamic?0:RowsAtCompileTime,ColsAtCompileTime==Dynamic?0:ColsAtCompileTime),
       // Stride<> does not allow default ctor for Dynamic strides, so let' initialize it with dummy values:
       m_stride(StrideType::OuterStrideAtCompileTime==Dynamic?0:StrideType::OuterStrideAtCompileTime,
@@ -159,7 +93,7 @@ protected:
   typedef Stride<StrideType::OuterStrideAtCompileTime,StrideType::InnerStrideAtCompileTime> StrideBase;
 
   template<typename Expression>
-  void construct(Expression& expr)
+  EIGEN_DEVICE_FUNC void construct(Expression& expr)
   {
     if(PlainObjectType::RowsAtCompileTime==1)
     {
@@ -184,15 +118,85 @@ protected:
   StrideBase m_stride;
 };
 
-
+/** \class Ref
+  * \ingroup Core_Module
+  *
+  * \brief A matrix or vector expression mapping an existing expression
+  *
+  * \tparam PlainObjectType the equivalent matrix type of the mapped data
+  * \tparam Options specifies the pointer alignment in bytes. It can be: \c #Aligned128, , \c #Aligned64, \c #Aligned32, \c #Aligned16, \c #Aligned8 or \c #Unaligned.
+  *                 The default is \c #Unaligned.
+  * \tparam StrideType optionally specifies strides. By default, Ref implies a contiguous storage along the inner dimension (inner stride==1),
+  *                   but accepts a variable outer stride (leading dimension).
+  *                   This can be overridden by specifying strides.
+  *                   The type passed here must be a specialization of the Stride template, see examples below.
+  *
+  * This class provides a way to write non-template functions taking Eigen objects as parameters while limiting the number of copies.
+  * A Ref<> object can represent either a const expression or a l-value:
+  * \code
+  * // in-out argument:
+  * void foo1(Ref<VectorXf> x);
+  *
+  * // read-only const argument:
+  * void foo2(const Ref<const VectorXf>& x);
+  * \endcode
+  *
+  * In the in-out case, the input argument must satisfy the constraints of the actual Ref<> type, otherwise a compilation issue will be triggered.
+  * By default, a Ref<VectorXf> can reference any dense vector expression of float having a contiguous memory layout.
+  * Likewise, a Ref<MatrixXf> can reference any column-major dense matrix expression of float whose column's elements are contiguously stored with
+  * the possibility to have a constant space in-between each column, i.e. the inner stride must be equal to 1, but the outer stride (or leading dimension)
+  * can be greater than the number of rows.
+  *
+  * In the const case, if the input expression does not match the above requirement, then it is evaluated into a temporary before being passed to the function.
+  * Here are some examples:
+  * \code
+  * MatrixXf A;
+  * VectorXf a;
+  * foo1(a.head());             // OK
+  * foo1(A.col());              // OK
+  * foo1(A.row());              // Compilation error because here innerstride!=1
+  * foo2(A.row());              // Compilation error because A.row() is a 1xN object while foo2 is expecting a Nx1 object
+  * foo2(A.row().transpose());  // The row is copied into a contiguous temporary
+  * foo2(2*a);                  // The expression is evaluated into a temporary
+  * foo2(A.col().segment(2,4)); // No temporary
+  * \endcode
+  *
+  * The range of inputs that can be referenced without temporary can be enlarged using the last two template parameters.
+  * Here is an example accepting an innerstride!=1:
+  * \code
+  * // in-out argument:
+  * void foo3(Ref<VectorXf,0,InnerStride<> > x);
+  * foo3(A.row());              // OK
+  * \endcode
+  * The downside here is that the function foo3 might be significantly slower than foo1 because it won't be able to exploit vectorization, and will involve more
+  * expensive address computations even if the input is contiguously stored in memory. To overcome this issue, one might propose to overload internally calling a
+  * template function, e.g.:
+  * \code
+  * // in the .h:
+  * void foo(const Ref<MatrixXf>& A);
+  * void foo(const Ref<MatrixXf,0,Stride<> >& A);
+  *
+  * // in the .cpp:
+  * template<typename TypeOfA> void foo_impl(const TypeOfA& A) {
+  *     ... // crazy code goes here
+  * }
+  * void foo(const Ref<MatrixXf>& A) { foo_impl(A); }
+  * void foo(const Ref<MatrixXf,0,Stride<> >& A) { foo_impl(A); }
+  * \endcode
+  *
+  * See also the following stackoverflow questions for further references:
+  *  - <a href="http://stackoverflow.com/questions/21132538/correct-usage-of-the-eigenref-class">Correct usage of the Eigen::Ref<> class</a>
+  *
+  * \sa PlainObjectBase::Map(), \ref TopicStorageOrders
+  */
 template<typename PlainObjectType, int Options, typename StrideType> class Ref
   : public RefBase<Ref<PlainObjectType, Options, StrideType> >
 {
   private:
     typedef internal::traits<Ref> Traits;
     template<typename Derived>
-    inline Ref(const PlainObjectBase<Derived>& expr,
-               typename internal::enable_if<bool(Traits::template match<Derived>::MatchAtCompileTime),Derived>::type* = 0);
+    EIGEN_DEVICE_FUNC inline Ref(const PlainObjectBase<Derived>& expr,
+                                 typename internal::enable_if<bool(Traits::template match<Derived>::MatchAtCompileTime),Derived>::type* = 0);
   public:
 
     typedef RefBase<Ref> Base;
@@ -201,23 +205,24 @@ template<typename PlainObjectType, int Options, typename StrideType> class Ref
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     template<typename Derived>
-    inline Ref(PlainObjectBase<Derived>& expr,
-               typename internal::enable_if<bool(Traits::template match<Derived>::MatchAtCompileTime),Derived>::type* = 0)
+    EIGEN_DEVICE_FUNC inline Ref(PlainObjectBase<Derived>& expr,
+                                 typename internal::enable_if<bool(Traits::template match<Derived>::MatchAtCompileTime),Derived>::type* = 0)
     {
-      EIGEN_STATIC_ASSERT(static_cast<bool>(Traits::template match<Derived>::MatchAtCompileTime), STORAGE_LAYOUT_DOES_NOT_MATCH);
+      EIGEN_STATIC_ASSERT(bool(Traits::template match<Derived>::MatchAtCompileTime), STORAGE_LAYOUT_DOES_NOT_MATCH);
       Base::construct(expr.derived());
     }
     template<typename Derived>
-    inline Ref(const DenseBase<Derived>& expr,
-               typename internal::enable_if<bool(Traits::template match<Derived>::MatchAtCompileTime),Derived>::type* = 0)
+    EIGEN_DEVICE_FUNC inline Ref(const DenseBase<Derived>& expr,
+                                 typename internal::enable_if<bool(Traits::template match<Derived>::MatchAtCompileTime),Derived>::type* = 0)
     #else
+    /** Implicit constructor from any dense expression */
     template<typename Derived>
     inline Ref(DenseBase<Derived>& expr)
     #endif
     {
-      EIGEN_STATIC_ASSERT(static_cast<bool>(internal::is_lvalue<Derived>::value), THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY);
-      EIGEN_STATIC_ASSERT(static_cast<bool>(Traits::template match<Derived>::MatchAtCompileTime), STORAGE_LAYOUT_DOES_NOT_MATCH);
-      enum { THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY = Derived::ThisConstantIsPrivateInPlainObjectBase};
+      EIGEN_STATIC_ASSERT(bool(internal::is_lvalue<Derived>::value), THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY);
+      EIGEN_STATIC_ASSERT(bool(Traits::template match<Derived>::MatchAtCompileTime), STORAGE_LAYOUT_DOES_NOT_MATCH);
+      EIGEN_STATIC_ASSERT(!Derived::IsPlainObjectBase,THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY);
       Base::construct(expr.const_cast_derived());
     }
 
@@ -236,36 +241,36 @@ template<typename TPlainObjectType, int Options, typename StrideType> class Ref<
     EIGEN_DENSE_PUBLIC_INTERFACE(Ref)
 
     template<typename Derived>
-    inline Ref(const DenseBase<Derived>& expr,
-               typename internal::enable_if<bool(Traits::template match<Derived>::ScalarTypeMatch),Derived>::type* = 0)
+    EIGEN_DEVICE_FUNC inline Ref(const DenseBase<Derived>& expr,
+                                 typename internal::enable_if<bool(Traits::template match<Derived>::ScalarTypeMatch),Derived>::type* = 0)
     {
 //      std::cout << match_helper<Derived>::HasDirectAccess << "," << match_helper<Derived>::OuterStrideMatch << "," << match_helper<Derived>::InnerStrideMatch << "\n";
 //      std::cout << int(StrideType::OuterStrideAtCompileTime) << " - " << int(Derived::OuterStrideAtCompileTime) << "\n";
 //      std::cout << int(StrideType::InnerStrideAtCompileTime) << " - " << int(Derived::InnerStrideAtCompileTime) << "\n";
       construct(expr.derived(), typename Traits::template match<Derived>::type());
     }
-    
-    inline Ref(const Ref& other) : Base(other) {
+
+    EIGEN_DEVICE_FUNC inline Ref(const Ref& other) : Base(other) {
       // copy constructor shall not copy the m_object, to avoid unnecessary malloc and copy
     }
 
     template<typename OtherRef>
-    inline Ref(const RefBase<OtherRef>& other) {
+    EIGEN_DEVICE_FUNC inline Ref(const RefBase<OtherRef>& other) {
       construct(other.derived(), typename Traits::template match<OtherRef>::type());
     }
 
   protected:
 
     template<typename Expression>
-    void construct(const Expression& expr,internal::true_type)
+    EIGEN_DEVICE_FUNC void construct(const Expression& expr,internal::true_type)
     {
       Base::construct(expr);
     }
 
     template<typename Expression>
-    void construct(const Expression& expr, internal::false_type)
+    EIGEN_DEVICE_FUNC void construct(const Expression& expr, internal::false_type)
     {
-      m_object.lazyAssign(expr);
+      internal::call_assignment_no_alias(m_object,expr,internal::assign_op<Scalar,Scalar>());
       Base::construct(m_object);
     }
 
diff --git a/eigen/Eigen/src/Core/Replicate.h b/eigen/Eigen/src/Core/Replicate.h
index ac4537c..0b2d6d7 100644
--- a/eigen/Eigen/src/Core/Replicate.h
+++ b/eigen/Eigen/src/Core/Replicate.h
@@ -12,21 +12,6 @@
 
 namespace Eigen { 
 
-/**
-  * \class Replicate
-  * \ingroup Core_Module
-  *
-  * \brief Expression of the multiple replication of a matrix or vector
-  *
-  * \param MatrixType the type of the object we are replicating
-  *
-  * This class represents an expression of the multiple replication of a matrix or vector.
-  * It is the return type of DenseBase::replicate() and most of the time
-  * this is the only way it is used.
-  *
-  * \sa DenseBase::replicate()
-  */
-
 namespace internal {
 template<typename MatrixType,int RowFactor,int ColFactor>
 struct traits<Replicate<MatrixType,RowFactor,ColFactor> >
@@ -35,10 +20,7 @@ struct traits<Replicate<MatrixType,RowFactor,ColFactor> >
   typedef typename MatrixType::Scalar Scalar;
   typedef typename traits<MatrixType>::StorageKind StorageKind;
   typedef typename traits<MatrixType>::XprKind XprKind;
-  enum {
-    Factor = (RowFactor==Dynamic || ColFactor==Dynamic) ? Dynamic : RowFactor*ColFactor
-  };
-  typedef typename nested<MatrixType,Factor>::type MatrixTypeNested;
+  typedef typename ref_selector<MatrixType>::type MatrixTypeNested;
   typedef typename remove_reference<MatrixTypeNested>::type _MatrixTypeNested;
   enum {
     RowsAtCompileTime = RowFactor==Dynamic || int(MatrixType::RowsAtCompileTime)==Dynamic
@@ -53,12 +35,29 @@ struct traits<Replicate<MatrixType,RowFactor,ColFactor> >
     IsRowMajor = MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1 ? 1
                : MaxColsAtCompileTime==1 && MaxRowsAtCompileTime!=1 ? 0
                : (MatrixType::Flags & RowMajorBit) ? 1 : 0,
-    Flags = (_MatrixTypeNested::Flags & HereditaryBits & ~RowMajorBit) | (IsRowMajor ? RowMajorBit : 0),
-    CoeffReadCost = _MatrixTypeNested::CoeffReadCost
+    
+    // FIXME enable DirectAccess with negative strides?
+    Flags = IsRowMajor ? RowMajorBit : 0
   };
 };
 }
 
+/**
+  * \class Replicate
+  * \ingroup Core_Module
+  *
+  * \brief Expression of the multiple replication of a matrix or vector
+  *
+  * \tparam MatrixType the type of the object we are replicating
+  * \tparam RowFactor number of repetitions at compile time along the vertical direction, can be Dynamic.
+  * \tparam ColFactor number of repetitions at compile time along the horizontal direction, can be Dynamic.
+  *
+  * This class represents an expression of the multiple replication of a matrix or vector.
+  * It is the return type of DenseBase::replicate() and most of the time
+  * this is the only way it is used.
+  *
+  * \sa DenseBase::replicate()
+  */
 template<typename MatrixType,int RowFactor,int ColFactor> class Replicate
   : public internal::dense_xpr_base< Replicate<MatrixType,RowFactor,ColFactor> >::type
 {
@@ -68,10 +67,12 @@ template<typename MatrixType,int RowFactor,int ColFactor> class Replicate
 
     typedef typename internal::dense_xpr_base<Replicate>::type Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(Replicate)
+    typedef typename internal::remove_all<MatrixType>::type NestedExpression;
 
     template<typename OriginalMatrixType>
-    inline explicit Replicate(const OriginalMatrixType& a_matrix)
-      : m_matrix(a_matrix), m_rowFactor(RowFactor), m_colFactor(ColFactor)
+    EIGEN_DEVICE_FUNC
+    inline explicit Replicate(const OriginalMatrixType& matrix)
+      : m_matrix(matrix), m_rowFactor(RowFactor), m_colFactor(ColFactor)
     {
       EIGEN_STATIC_ASSERT((internal::is_same<typename internal::remove_const<MatrixType>::type,OriginalMatrixType>::value),
                           THE_MATRIX_OR_EXPRESSION_THAT_YOU_PASSED_DOES_NOT_HAVE_THE_EXPECTED_TYPE)
@@ -79,41 +80,20 @@ template<typename MatrixType,int RowFactor,int ColFactor> class Replicate
     }
 
     template<typename OriginalMatrixType>
-    inline Replicate(const OriginalMatrixType& a_matrix, Index rowFactor, Index colFactor)
-      : m_matrix(a_matrix), m_rowFactor(rowFactor), m_colFactor(colFactor)
+    EIGEN_DEVICE_FUNC
+    inline Replicate(const OriginalMatrixType& matrix, Index rowFactor, Index colFactor)
+      : m_matrix(matrix), m_rowFactor(rowFactor), m_colFactor(colFactor)
     {
       EIGEN_STATIC_ASSERT((internal::is_same<typename internal::remove_const<MatrixType>::type,OriginalMatrixType>::value),
                           THE_MATRIX_OR_EXPRESSION_THAT_YOU_PASSED_DOES_NOT_HAVE_THE_EXPECTED_TYPE)
     }
 
+    EIGEN_DEVICE_FUNC
     inline Index rows() const { return m_matrix.rows() * m_rowFactor.value(); }
+    EIGEN_DEVICE_FUNC
     inline Index cols() const { return m_matrix.cols() * m_colFactor.value(); }
 
-    inline Scalar coeff(Index rowId, Index colId) const
-    {
-      // try to avoid using modulo; this is a pure optimization strategy
-      const Index actual_row  = internal::traits<MatrixType>::RowsAtCompileTime==1 ? 0
-                            : RowFactor==1 ? rowId
-                            : rowId%m_matrix.rows();
-      const Index actual_col  = internal::traits<MatrixType>::ColsAtCompileTime==1 ? 0
-                            : ColFactor==1 ? colId
-                            : colId%m_matrix.cols();
-
-      return m_matrix.coeff(actual_row, actual_col);
-    }
-    template<int LoadMode>
-    inline PacketScalar packet(Index rowId, Index colId) const
-    {
-      const Index actual_row  = internal::traits<MatrixType>::RowsAtCompileTime==1 ? 0
-                            : RowFactor==1 ? rowId
-                            : rowId%m_matrix.rows();
-      const Index actual_col  = internal::traits<MatrixType>::ColsAtCompileTime==1 ? 0
-                            : ColFactor==1 ? colId
-                            : colId%m_matrix.cols();
-
-      return m_matrix.template packet<LoadMode>(actual_row, actual_col);
-    }
-
+    EIGEN_DEVICE_FUNC
     const _MatrixTypeNested& nestedExpression() const
     { 
       return m_matrix; 
@@ -135,28 +115,13 @@ template<typename MatrixType,int RowFactor,int ColFactor> class Replicate
   */
 template<typename Derived>
 template<int RowFactor, int ColFactor>
-const Replicate<Derived,RowFactor,ColFactor>
+EIGEN_DEVICE_FUNC const Replicate<Derived,RowFactor,ColFactor>
 DenseBase<Derived>::replicate() const
 {
   return Replicate<Derived,RowFactor,ColFactor>(derived());
 }
 
 /**
-  * \return an expression of the replication of \c *this
-  *
-  * Example: \include MatrixBase_replicate_int_int.cpp
-  * Output: \verbinclude MatrixBase_replicate_int_int.out
-  *
-  * \sa VectorwiseOp::replicate(), DenseBase::replicate<int,int>(), class Replicate
-  */
-template<typename Derived>
-const typename DenseBase<Derived>::ReplicateReturnType
-DenseBase<Derived>::replicate(Index rowFactor,Index colFactor) const
-{
-  return Replicate<Derived,Dynamic,Dynamic>(derived(),rowFactor,colFactor);
-}
-
-/**
   * \return an expression of the replication of each column (or row) of \c *this
   *
   * Example: \include DirectionWise_replicate_int.cpp
@@ -165,7 +130,7 @@ DenseBase<Derived>::replicate(Index rowFactor,Index colFactor) const
   * \sa VectorwiseOp::replicate(), DenseBase::replicate(), class Replicate
   */
 template<typename ExpressionType, int Direction>
-const typename VectorwiseOp<ExpressionType,Direction>::ReplicateReturnType
+EIGEN_DEVICE_FUNC const typename VectorwiseOp<ExpressionType,Direction>::ReplicateReturnType
 VectorwiseOp<ExpressionType,Direction>::replicate(Index factor) const
 {
   return typename VectorwiseOp<ExpressionType,Direction>::ReplicateReturnType
diff --git a/eigen/Eigen/src/Core/ReturnByValue.h b/eigen/Eigen/src/Core/ReturnByValue.h
index f635598..11dc86d 100644
--- a/eigen/Eigen/src/Core/ReturnByValue.h
+++ b/eigen/Eigen/src/Core/ReturnByValue.h
@@ -13,11 +13,6 @@
 
 namespace Eigen {
 
-/** \class ReturnByValue
-  * \ingroup Core_Module
-  *
-  */
-
 namespace internal {
 
 template<typename Derived>
@@ -38,17 +33,22 @@ struct traits<ReturnByValue<Derived> >
  * So internal::nested always gives the plain return matrix type.
  *
  * FIXME: I don't understand why we need this specialization: isn't this taken care of by the EvalBeforeNestingBit ??
+ * Answer: EvalBeforeNestingBit should be deprecated since we have the evaluators
  */
 template<typename Derived,int n,typename PlainObject>
-struct nested<ReturnByValue<Derived>, n, PlainObject>
+struct nested_eval<ReturnByValue<Derived>, n, PlainObject>
 {
   typedef typename traits<Derived>::ReturnType type;
 };
 
 } // end namespace internal
 
+/** \class ReturnByValue
+  * \ingroup Core_Module
+  *
+  */
 template<typename Derived> class ReturnByValue
-  : internal::no_assignment_operator, public internal::dense_xpr_base< ReturnByValue<Derived> >::type
+  : public internal::dense_xpr_base< ReturnByValue<Derived> >::type, internal::no_assignment_operator
 {
   public:
     typedef typename internal::traits<Derived>::ReturnType ReturnType;
@@ -57,10 +57,11 @@ template<typename Derived> class ReturnByValue
     EIGEN_DENSE_PUBLIC_INTERFACE(ReturnByValue)
 
     template<typename Dest>
+    EIGEN_DEVICE_FUNC
     inline void evalTo(Dest& dst) const
     { static_cast<const Derived*>(this)->evalTo(dst); }
-    inline Index rows() const { return static_cast<const Derived*>(this)->rows(); }
-    inline Index cols() const { return static_cast<const Derived*>(this)->cols(); }
+    EIGEN_DEVICE_FUNC inline Index rows() const { return static_cast<const Derived*>(this)->rows(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return static_cast<const Derived*>(this)->cols(); }
 
 #ifndef EIGEN_PARSED_BY_DOXYGEN
 #define Unusable YOU_ARE_TRYING_TO_ACCESS_A_SINGLE_COEFFICIENT_IN_A_SPECIAL_EXPRESSION_WHERE_THAT_IS_NOT_ALLOWED_BECAUSE_THAT_WOULD_BE_INEFFICIENT
@@ -72,27 +73,44 @@ template<typename Derived> class ReturnByValue
     const Unusable& coeff(Index,Index) const { return *reinterpret_cast<const Unusable*>(this); }
     Unusable& coeffRef(Index) { return *reinterpret_cast<Unusable*>(this); }
     Unusable& coeffRef(Index,Index) { return *reinterpret_cast<Unusable*>(this); }
-    template<int LoadMode>  Unusable& packet(Index) const;
-    template<int LoadMode>  Unusable& packet(Index, Index) const;
+#undef Unusable
 #endif
 };
 
 template<typename Derived>
 template<typename OtherDerived>
-Derived& DenseBase<Derived>::operator=(const ReturnByValue<OtherDerived>& other)
+EIGEN_DEVICE_FUNC Derived& DenseBase<Derived>::operator=(const ReturnByValue<OtherDerived>& other)
 {
   other.evalTo(derived());
   return derived();
 }
 
+namespace internal {
+
+// Expression is evaluated in a temporary; default implementation of Assignment is bypassed so that
+// when a ReturnByValue expression is assigned, the evaluator is not constructed.
+// TODO: Finalize port to new regime; ReturnByValue should not exist in the expression world
+  
 template<typename Derived>
-template<typename OtherDerived>
-Derived& DenseBase<Derived>::lazyAssign(const ReturnByValue<OtherDerived>& other)
+struct evaluator<ReturnByValue<Derived> >
+  : public evaluator<typename internal::traits<Derived>::ReturnType>
 {
-  other.evalTo(derived());
-  return derived();
-}
+  typedef ReturnByValue<Derived> XprType;
+  typedef typename internal::traits<Derived>::ReturnType PlainObject;
+  typedef evaluator<PlainObject> Base;
+  
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr)
+    : m_result(xpr.rows(), xpr.cols())
+  {
+    ::new (static_cast<Base*>(this)) Base(m_result);
+    xpr.evalTo(m_result);
+  }
+
+protected:
+  PlainObject m_result;
+};
 
+} // end namespace internal
 
 } // end namespace Eigen
 
diff --git a/eigen/Eigen/src/Core/Reverse.h b/eigen/Eigen/src/Core/Reverse.h
index e30ae3d..8b6b3ab 100644
--- a/eigen/Eigen/src/Core/Reverse.h
+++ b/eigen/Eigen/src/Core/Reverse.h
@@ -14,20 +14,6 @@
 
 namespace Eigen { 
 
-/** \class Reverse
-  * \ingroup Core_Module
-  *
-  * \brief Expression of the reverse of a vector or matrix
-  *
-  * \param MatrixType the type of the object of which we are taking the reverse
-  *
-  * This class represents an expression of the reverse of a vector.
-  * It is the return type of MatrixBase::reverse() and VectorwiseOp::reverse()
-  * and most of the time this is the only way it is used.
-  *
-  * \sa MatrixBase::reverse(), VectorwiseOp::reverse()
-  */
-
 namespace internal {
 
 template<typename MatrixType, int Direction>
@@ -37,36 +23,43 @@ struct traits<Reverse<MatrixType, Direction> >
   typedef typename MatrixType::Scalar Scalar;
   typedef typename traits<MatrixType>::StorageKind StorageKind;
   typedef typename traits<MatrixType>::XprKind XprKind;
-  typedef typename nested<MatrixType>::type MatrixTypeNested;
+  typedef typename ref_selector<MatrixType>::type MatrixTypeNested;
   typedef typename remove_reference<MatrixTypeNested>::type _MatrixTypeNested;
   enum {
     RowsAtCompileTime = MatrixType::RowsAtCompileTime,
     ColsAtCompileTime = MatrixType::ColsAtCompileTime,
     MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
     MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime,
-
-    // let's enable LinearAccess only with vectorization because of the product overhead
-    LinearAccess = ( (Direction==BothDirections) && (int(_MatrixTypeNested::Flags)&PacketAccessBit) )
-                 ? LinearAccessBit : 0,
-
-    Flags = int(_MatrixTypeNested::Flags) & (HereditaryBits | LvalueBit | PacketAccessBit | LinearAccess),
-
-    CoeffReadCost = _MatrixTypeNested::CoeffReadCost
+    Flags = _MatrixTypeNested::Flags & (RowMajorBit | LvalueBit)
   };
 };
 
-template<typename PacketScalar, bool ReversePacket> struct reverse_packet_cond
+template<typename PacketType, bool ReversePacket> struct reverse_packet_cond
 {
-  static inline PacketScalar run(const PacketScalar& x) { return preverse(x); }
+  static inline PacketType run(const PacketType& x) { return preverse(x); }
 };
 
-template<typename PacketScalar> struct reverse_packet_cond<PacketScalar,false>
+template<typename PacketType> struct reverse_packet_cond<PacketType,false>
 {
-  static inline PacketScalar run(const PacketScalar& x) { return x; }
+  static inline PacketType run(const PacketType& x) { return x; }
 };
 
 } // end namespace internal 
 
+/** \class Reverse
+  * \ingroup Core_Module
+  *
+  * \brief Expression of the reverse of a vector or matrix
+  *
+  * \tparam MatrixType the type of the object of which we are taking the reverse
+  * \tparam Direction defines the direction of the reverse operation, can be Vertical, Horizontal, or BothDirections
+  *
+  * This class represents an expression of the reverse of a vector.
+  * It is the return type of MatrixBase::reverse() and VectorwiseOp::reverse()
+  * and most of the time this is the only way it is used.
+  *
+  * \sa MatrixBase::reverse(), VectorwiseOp::reverse()
+  */
 template<typename MatrixType, int Direction> class Reverse
   : public internal::dense_xpr_base< Reverse<MatrixType, Direction> >::type
 {
@@ -74,12 +67,9 @@ template<typename MatrixType, int Direction> class Reverse
 
     typedef typename internal::dense_xpr_base<Reverse>::type Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(Reverse)
+    typedef typename internal::remove_all<MatrixType>::type NestedExpression;
     using Base::IsRowMajor;
 
-    // next line is necessary because otherwise const version of operator()
-    // is hidden by non-const version defined in this file
-    using Base::operator(); 
-
   protected:
     enum {
       PacketSize = internal::packet_traits<Scalar>::size,
@@ -95,82 +85,19 @@ template<typename MatrixType, int Direction> class Reverse
     typedef internal::reverse_packet_cond<PacketScalar,ReversePacket> reverse_packet;
   public:
 
-    inline Reverse(const MatrixType& matrix) : m_matrix(matrix) { }
+    EIGEN_DEVICE_FUNC explicit inline Reverse(const MatrixType& matrix) : m_matrix(matrix) { }
 
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Reverse)
 
-    inline Index rows() const { return m_matrix.rows(); }
-    inline Index cols() const { return m_matrix.cols(); }
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_matrix.rows(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_matrix.cols(); }
 
-    inline Index innerStride() const
+    EIGEN_DEVICE_FUNC inline Index innerStride() const
     {
       return -m_matrix.innerStride();
     }
 
-    inline Scalar& operator()(Index row, Index col)
-    {
-      eigen_assert(row >= 0 && row < rows() && col >= 0 && col < cols());
-      return coeffRef(row, col);
-    }
-
-    inline Scalar& coeffRef(Index row, Index col)
-    {
-      return m_matrix.const_cast_derived().coeffRef(ReverseRow ? m_matrix.rows() - row - 1 : row,
-                                                    ReverseCol ? m_matrix.cols() - col - 1 : col);
-    }
-
-    inline CoeffReturnType coeff(Index row, Index col) const
-    {
-      return m_matrix.coeff(ReverseRow ? m_matrix.rows() - row - 1 : row,
-                            ReverseCol ? m_matrix.cols() - col - 1 : col);
-    }
-
-    inline CoeffReturnType coeff(Index index) const
-    {
-      return m_matrix.coeff(m_matrix.size() - index - 1);
-    }
-
-    inline Scalar& coeffRef(Index index)
-    {
-      return m_matrix.const_cast_derived().coeffRef(m_matrix.size() - index - 1);
-    }
-
-    inline Scalar& operator()(Index index)
-    {
-      eigen_assert(index >= 0 && index < m_matrix.size());
-      return coeffRef(index);
-    }
-
-    template<int LoadMode>
-    inline const PacketScalar packet(Index row, Index col) const
-    {
-      return reverse_packet::run(m_matrix.template packet<LoadMode>(
-                                    ReverseRow ? m_matrix.rows() - row - OffsetRow : row,
-                                    ReverseCol ? m_matrix.cols() - col - OffsetCol : col));
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index row, Index col, const PacketScalar& x)
-    {
-      m_matrix.const_cast_derived().template writePacket<LoadMode>(
-                                      ReverseRow ? m_matrix.rows() - row - OffsetRow : row,
-                                      ReverseCol ? m_matrix.cols() - col - OffsetCol : col,
-                                      reverse_packet::run(x));
-    }
-
-    template<int LoadMode>
-    inline const PacketScalar packet(Index index) const
-    {
-      return internal::preverse(m_matrix.template packet<LoadMode>( m_matrix.size() - index - PacketSize ));
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index index, const PacketScalar& x)
-    {
-      m_matrix.const_cast_derived().template writePacket<LoadMode>(m_matrix.size() - index - PacketSize, internal::preverse(x));
-    }
-
-    const typename internal::remove_all<typename MatrixType::Nested>::type& 
+    EIGEN_DEVICE_FUNC const typename internal::remove_all<typename MatrixType::Nested>::type&
     nestedExpression() const 
     {
       return m_matrix;
@@ -187,36 +114,96 @@ template<typename MatrixType, int Direction> class Reverse
   *
   */
 template<typename Derived>
-inline typename DenseBase<Derived>::ReverseReturnType
+EIGEN_DEVICE_FUNC inline typename DenseBase<Derived>::ReverseReturnType
 DenseBase<Derived>::reverse()
 {
-  return derived();
+  return ReverseReturnType(derived());
 }
 
-/** This is the const version of reverse(). */
-template<typename Derived>
-inline const typename DenseBase<Derived>::ConstReverseReturnType
-DenseBase<Derived>::reverse() const
-{
-  return derived();
-}
+
+//reverse const overload moved DenseBase.h due to a CUDA compiler bug
 
 /** This is the "in place" version of reverse: it reverses \c *this.
   *
   * In most cases it is probably better to simply use the reversed expression
   * of a matrix. However, when reversing the matrix data itself is really needed,
   * then this "in-place" version is probably the right choice because it provides
-  * the following additional features:
+  * the following additional benefits:
   *  - less error prone: doing the same operation with .reverse() requires special care:
   *    \code m = m.reverse().eval(); \endcode
-  *  - this API allows to avoid creating a temporary (the current implementation creates a temporary, but that could be avoided using swap)
+  *  - this API enables reverse operations without the need for a temporary
   *  - it allows future optimizations (cache friendliness, etc.)
   *
-  * \sa reverse() */
+  * \sa VectorwiseOp::reverseInPlace(), reverse() */
 template<typename Derived>
-inline void DenseBase<Derived>::reverseInPlace()
+EIGEN_DEVICE_FUNC inline void DenseBase<Derived>::reverseInPlace()
+{
+  if(cols()>rows())
+  {
+    Index half = cols()/2;
+    leftCols(half).swap(rightCols(half).reverse());
+    if((cols()%2)==1)
+    {
+      Index half2 = rows()/2;
+      col(half).head(half2).swap(col(half).tail(half2).reverse());
+    }
+  }
+  else
+  {
+    Index half = rows()/2;
+    topRows(half).swap(bottomRows(half).reverse());
+    if((rows()%2)==1)
+    {
+      Index half2 = cols()/2;
+      row(half).head(half2).swap(row(half).tail(half2).reverse());
+    }
+  }
+}
+
+namespace internal {
+  
+template<int Direction>
+struct vectorwise_reverse_inplace_impl;
+
+template<>
+struct vectorwise_reverse_inplace_impl<Vertical>
+{
+  template<typename ExpressionType>
+  static void run(ExpressionType &xpr)
+  {
+    Index half = xpr.rows()/2;
+    xpr.topRows(half).swap(xpr.bottomRows(half).colwise().reverse());
+  }
+};
+
+template<>
+struct vectorwise_reverse_inplace_impl<Horizontal>
+{
+  template<typename ExpressionType>
+  static void run(ExpressionType &xpr)
+  {
+    Index half = xpr.cols()/2;
+    xpr.leftCols(half).swap(xpr.rightCols(half).rowwise().reverse());
+  }
+};
+
+} // end namespace internal
+
+/** This is the "in place" version of VectorwiseOp::reverse: it reverses each column or row of \c *this.
+  *
+  * In most cases it is probably better to simply use the reversed expression
+  * of a matrix. However, when reversing the matrix data itself is really needed,
+  * then this "in-place" version is probably the right choice because it provides
+  * the following additional benefits:
+  *  - less error prone: doing the same operation with .reverse() requires special care:
+  *    \code m = m.reverse().eval(); \endcode
+  *  - this API enables reverse operations without the need for a temporary
+  *
+  * \sa DenseBase::reverseInPlace(), reverse() */
+template<typename ExpressionType, int Direction>
+EIGEN_DEVICE_FUNC void VectorwiseOp<ExpressionType,Direction>::reverseInPlace()
 {
-  derived() = derived().reverse().eval();
+  internal::vectorwise_reverse_inplace_impl<Direction>::run(_expression().const_cast_derived());
 }
 
 } // end namespace Eigen
diff --git a/eigen/Eigen/src/Core/Select.h b/eigen/Eigen/src/Core/Select.h
index 87993bb..79eec1b 100644
--- a/eigen/Eigen/src/Core/Select.h
+++ b/eigen/Eigen/src/Core/Select.h
@@ -43,23 +43,21 @@ struct traits<Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> >
     ColsAtCompileTime = ConditionMatrixType::ColsAtCompileTime,
     MaxRowsAtCompileTime = ConditionMatrixType::MaxRowsAtCompileTime,
     MaxColsAtCompileTime = ConditionMatrixType::MaxColsAtCompileTime,
-    Flags = (unsigned int)ThenMatrixType::Flags & ElseMatrixType::Flags & HereditaryBits,
-    CoeffReadCost = traits<typename remove_all<ConditionMatrixNested>::type>::CoeffReadCost
-                  + EIGEN_SIZE_MAX(traits<typename remove_all<ThenMatrixNested>::type>::CoeffReadCost,
-                                   traits<typename remove_all<ElseMatrixNested>::type>::CoeffReadCost)
+    Flags = (unsigned int)ThenMatrixType::Flags & ElseMatrixType::Flags & RowMajorBit
   };
 };
 }
 
 template<typename ConditionMatrixType, typename ThenMatrixType, typename ElseMatrixType>
-class Select : internal::no_assignment_operator,
-  public internal::dense_xpr_base< Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> >::type
+class Select : public internal::dense_xpr_base< Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> >::type,
+               internal::no_assignment_operator
 {
   public:
 
     typedef typename internal::dense_xpr_base<Select>::type Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(Select)
 
+    inline EIGEN_DEVICE_FUNC
     Select(const ConditionMatrixType& a_conditionMatrix,
            const ThenMatrixType& a_thenMatrix,
            const ElseMatrixType& a_elseMatrix)
@@ -69,9 +67,10 @@ class Select : internal::no_assignment_operator,
       eigen_assert(m_condition.cols() == m_then.cols() && m_condition.cols() == m_else.cols());
     }
 
-    Index rows() const { return m_condition.rows(); }
-    Index cols() const { return m_condition.cols(); }
+    inline EIGEN_DEVICE_FUNC Index rows() const { return m_condition.rows(); }
+    inline EIGEN_DEVICE_FUNC Index cols() const { return m_condition.cols(); }
 
+    inline EIGEN_DEVICE_FUNC
     const Scalar coeff(Index i, Index j) const
     {
       if (m_condition.coeff(i,j))
@@ -80,6 +79,7 @@ class Select : internal::no_assignment_operator,
         return m_else.coeff(i,j);
     }
 
+    inline EIGEN_DEVICE_FUNC
     const Scalar coeff(Index i) const
     {
       if (m_condition.coeff(i))
@@ -88,17 +88,17 @@ class Select : internal::no_assignment_operator,
         return m_else.coeff(i);
     }
 
-    const ConditionMatrixType& conditionMatrix() const
+    inline EIGEN_DEVICE_FUNC const ConditionMatrixType& conditionMatrix() const
     {
       return m_condition;
     }
 
-    const ThenMatrixType& thenMatrix() const
+    inline EIGEN_DEVICE_FUNC const ThenMatrixType& thenMatrix() const
     {
       return m_then;
     }
 
-    const ElseMatrixType& elseMatrix() const
+    inline EIGEN_DEVICE_FUNC const ElseMatrixType& elseMatrix() const
     {
       return m_else;
     }
diff --git a/eigen/Eigen/src/Core/SelfAdjointView.h b/eigen/Eigen/src/Core/SelfAdjointView.h
index 6fa7cd1..7e71fe3 100644
--- a/eigen/Eigen/src/Core/SelfAdjointView.h
+++ b/eigen/Eigen/src/Core/SelfAdjointView.h
@@ -32,54 +32,60 @@ namespace internal {
 template<typename MatrixType, unsigned int UpLo>
 struct traits<SelfAdjointView<MatrixType, UpLo> > : traits<MatrixType>
 {
-  typedef typename nested<MatrixType>::type MatrixTypeNested;
+  typedef typename ref_selector<MatrixType>::non_const_type MatrixTypeNested;
   typedef typename remove_all<MatrixTypeNested>::type MatrixTypeNestedCleaned;
   typedef MatrixType ExpressionType;
-  typedef typename MatrixType::PlainObject DenseMatrixType;
+  typedef typename MatrixType::PlainObject FullMatrixType;
   enum {
     Mode = UpLo | SelfAdjoint,
-    Flags =  MatrixTypeNestedCleaned::Flags & (HereditaryBits)
-           & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit)), // FIXME these flags should be preserved
-    CoeffReadCost = MatrixTypeNestedCleaned::CoeffReadCost
+    FlagsLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0,
+    Flags =  MatrixTypeNestedCleaned::Flags & (HereditaryBits|FlagsLvalueBit)
+           & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit)) // FIXME these flags should be preserved
   };
 };
 }
 
-template <typename Lhs, int LhsMode, bool LhsIsVector,
-          typename Rhs, int RhsMode, bool RhsIsVector>
-struct SelfadjointProductMatrix;
 
-// FIXME could also be called SelfAdjointWrapper to be consistent with DiagonalWrapper ??
-template<typename MatrixType, unsigned int UpLo> class SelfAdjointView
-  : public TriangularBase<SelfAdjointView<MatrixType, UpLo> >
+template<typename _MatrixType, unsigned int UpLo> class SelfAdjointView
+  : public TriangularBase<SelfAdjointView<_MatrixType, UpLo> >
 {
   public:
 
+    typedef _MatrixType MatrixType;
     typedef TriangularBase<SelfAdjointView> Base;
     typedef typename internal::traits<SelfAdjointView>::MatrixTypeNested MatrixTypeNested;
     typedef typename internal::traits<SelfAdjointView>::MatrixTypeNestedCleaned MatrixTypeNestedCleaned;
+    typedef MatrixTypeNestedCleaned NestedExpression;
 
     /** \brief The type of coefficients in this matrix */
     typedef typename internal::traits<SelfAdjointView>::Scalar Scalar; 
-
-    typedef typename MatrixType::Index Index;
+    typedef typename MatrixType::StorageIndex StorageIndex;
+    typedef typename internal::remove_all<typename MatrixType::ConjugateReturnType>::type MatrixConjugateReturnType;
 
     enum {
-      Mode = internal::traits<SelfAdjointView>::Mode
+      Mode = internal::traits<SelfAdjointView>::Mode,
+      Flags = internal::traits<SelfAdjointView>::Flags,
+      TransposeMode = ((Mode & Upper) ? Lower : 0) | ((Mode & Lower) ? Upper : 0)
     };
     typedef typename MatrixType::PlainObject PlainObject;
 
-    inline SelfAdjointView(MatrixType& matrix) : m_matrix(matrix)
+    EIGEN_DEVICE_FUNC
+    explicit inline SelfAdjointView(MatrixType& matrix) : m_matrix(matrix)
     {}
 
+    EIGEN_DEVICE_FUNC
     inline Index rows() const { return m_matrix.rows(); }
+    EIGEN_DEVICE_FUNC
     inline Index cols() const { return m_matrix.cols(); }
+    EIGEN_DEVICE_FUNC
     inline Index outerStride() const { return m_matrix.outerStride(); }
+    EIGEN_DEVICE_FUNC
     inline Index innerStride() const { return m_matrix.innerStride(); }
 
     /** \sa MatrixBase::coeff()
       * \warning the coordinates must fit into the referenced triangular part
       */
+    EIGEN_DEVICE_FUNC
     inline Scalar coeff(Index row, Index col) const
     {
       Base::check_coordinates_internal(row, col);
@@ -89,36 +95,46 @@ template<typename MatrixType, unsigned int UpLo> class SelfAdjointView
     /** \sa MatrixBase::coeffRef()
       * \warning the coordinates must fit into the referenced triangular part
       */
+    EIGEN_DEVICE_FUNC
     inline Scalar& coeffRef(Index row, Index col)
     {
+      EIGEN_STATIC_ASSERT_LVALUE(SelfAdjointView);
       Base::check_coordinates_internal(row, col);
-      return m_matrix.const_cast_derived().coeffRef(row, col);
+      return m_matrix.coeffRef(row, col);
     }
 
     /** \internal */
+    EIGEN_DEVICE_FUNC
     const MatrixTypeNestedCleaned& _expression() const { return m_matrix; }
 
+    EIGEN_DEVICE_FUNC
     const MatrixTypeNestedCleaned& nestedExpression() const { return m_matrix; }
-    MatrixTypeNestedCleaned& nestedExpression() { return *const_cast<MatrixTypeNestedCleaned*>(&m_matrix); }
+    EIGEN_DEVICE_FUNC
+    MatrixTypeNestedCleaned& nestedExpression() { return m_matrix; }
 
-    /** Efficient self-adjoint matrix times vector/matrix product */
+    /** Efficient triangular matrix times vector/matrix product */
     template<typename OtherDerived>
-    SelfadjointProductMatrix<MatrixType,Mode,false,OtherDerived,0,OtherDerived::IsVectorAtCompileTime>
+    EIGEN_DEVICE_FUNC
+    const Product<SelfAdjointView,OtherDerived>
     operator*(const MatrixBase<OtherDerived>& rhs) const
     {
-      return SelfadjointProductMatrix
-              <MatrixType,Mode,false,OtherDerived,0,OtherDerived::IsVectorAtCompileTime>
-              (m_matrix, rhs.derived());
+      return Product<SelfAdjointView,OtherDerived>(*this, rhs.derived());
     }
 
-    /** Efficient vector/matrix times self-adjoint matrix product */
+    /** Efficient vector/matrix times triangular matrix product */
     template<typename OtherDerived> friend
-    SelfadjointProductMatrix<OtherDerived,0,OtherDerived::IsVectorAtCompileTime,MatrixType,Mode,false>
+    EIGEN_DEVICE_FUNC
+    const Product<OtherDerived,SelfAdjointView>
     operator*(const MatrixBase<OtherDerived>& lhs, const SelfAdjointView& rhs)
     {
-      return SelfadjointProductMatrix
-              <OtherDerived,0,OtherDerived::IsVectorAtCompileTime,MatrixType,Mode,false>
-              (lhs.derived(),rhs.m_matrix);
+      return Product<OtherDerived,SelfAdjointView>(lhs.derived(),rhs);
+    }
+    
+    friend EIGEN_DEVICE_FUNC
+    const SelfAdjointView<const EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(Scalar,MatrixType,product),UpLo>
+    operator*(const Scalar& s, const SelfAdjointView& mat)
+    {
+      return (s*mat.nestedExpression()).template selfadjointView<UpLo>();
     }
 
     /** Perform a symmetric rank 2 update of the selfadjoint matrix \c *this:
@@ -132,6 +148,7 @@ template<typename MatrixType, unsigned int UpLo> class SelfAdjointView
       * \sa rankUpdate(const MatrixBase<DerivedU>&, Scalar)
       */
     template<typename DerivedU, typename DerivedV>
+    EIGEN_DEVICE_FUNC
     SelfAdjointView& rankUpdate(const MatrixBase<DerivedU>& u, const MatrixBase<DerivedV>& v, const Scalar& alpha = Scalar(1));
 
     /** Perform a symmetric rank K update of the selfadjoint matrix \c *this:
@@ -145,8 +162,74 @@ template<typename MatrixType, unsigned int UpLo> class SelfAdjointView
       * \sa rankUpdate(const MatrixBase<DerivedU>&, const MatrixBase<DerivedV>&, Scalar)
       */
     template<typename DerivedU>
+    EIGEN_DEVICE_FUNC
     SelfAdjointView& rankUpdate(const MatrixBase<DerivedU>& u, const Scalar& alpha = Scalar(1));
 
+    /** \returns an expression of a triangular view extracted from the current selfadjoint view of a given triangular part
+      *
+      * The parameter \a TriMode can have the following values: \c #Upper, \c #StrictlyUpper, \c #UnitUpper,
+      * \c #Lower, \c #StrictlyLower, \c #UnitLower.
+      *
+      * If \c TriMode references the same triangular part than \c *this, then this method simply return a \c TriangularView of the nested expression,
+      * otherwise, the nested expression is first transposed, thus returning a \c TriangularView<Transpose<MatrixType>> object.
+      *
+      * \sa MatrixBase::triangularView(), class TriangularView
+      */
+    template<unsigned int TriMode>
+    EIGEN_DEVICE_FUNC
+    typename internal::conditional<(TriMode&(Upper|Lower))==(UpLo&(Upper|Lower)),
+                                   TriangularView<MatrixType,TriMode>,
+                                   TriangularView<typename MatrixType::AdjointReturnType,TriMode> >::type
+    triangularView() const
+    {
+      typename internal::conditional<(TriMode&(Upper|Lower))==(UpLo&(Upper|Lower)), MatrixType&, typename MatrixType::ConstTransposeReturnType>::type tmp1(m_matrix);
+      typename internal::conditional<(TriMode&(Upper|Lower))==(UpLo&(Upper|Lower)), MatrixType&, typename MatrixType::AdjointReturnType>::type tmp2(tmp1);
+      return typename internal::conditional<(TriMode&(Upper|Lower))==(UpLo&(Upper|Lower)),
+                                   TriangularView<MatrixType,TriMode>,
+                                   TriangularView<typename MatrixType::AdjointReturnType,TriMode> >::type(tmp2);
+    }
+
+    typedef SelfAdjointView<const MatrixConjugateReturnType,Mode> ConjugateReturnType;
+    /** \sa MatrixBase::conjugate() const */
+    EIGEN_DEVICE_FUNC
+    inline const ConjugateReturnType conjugate() const
+    { return ConjugateReturnType(m_matrix.conjugate()); }
+
+    typedef SelfAdjointView<const typename MatrixType::AdjointReturnType,TransposeMode> AdjointReturnType;
+    /** \sa MatrixBase::adjoint() const */
+    EIGEN_DEVICE_FUNC
+    inline const AdjointReturnType adjoint() const
+    { return AdjointReturnType(m_matrix.adjoint()); }
+
+    typedef SelfAdjointView<typename MatrixType::TransposeReturnType,TransposeMode> TransposeReturnType;
+     /** \sa MatrixBase::transpose() */
+    EIGEN_DEVICE_FUNC
+    inline TransposeReturnType transpose()
+    {
+      EIGEN_STATIC_ASSERT_LVALUE(MatrixType)
+      typename MatrixType::TransposeReturnType tmp(m_matrix);
+      return TransposeReturnType(tmp);
+    }
+
+    typedef SelfAdjointView<const typename MatrixType::ConstTransposeReturnType,TransposeMode> ConstTransposeReturnType;
+    /** \sa MatrixBase::transpose() const */
+    EIGEN_DEVICE_FUNC
+    inline const ConstTransposeReturnType transpose() const
+    {
+      return ConstTransposeReturnType(m_matrix.transpose());
+    }
+
+    /** \returns a const expression of the main diagonal of the matrix \c *this
+      *
+      * This method simply returns the diagonal of the nested expression, thus by-passing the SelfAdjointView decorator.
+      *
+      * \sa MatrixBase::diagonal(), class Diagonal */
+    EIGEN_DEVICE_FUNC
+    typename MatrixType::ConstDiagonalReturnType diagonal() const
+    {
+      return typename MatrixType::ConstDiagonalReturnType(m_matrix);
+    }
+
 /////////// Cholesky module ///////////
 
     const LLT<PlainObject, UpLo> llt() const;
@@ -159,31 +242,10 @@ template<typename MatrixType, unsigned int UpLo> class SelfAdjointView
     /** Return type of eigenvalues() */
     typedef Matrix<RealScalar, internal::traits<MatrixType>::ColsAtCompileTime, 1> EigenvaluesReturnType;
 
+    EIGEN_DEVICE_FUNC
     EigenvaluesReturnType eigenvalues() const;
+    EIGEN_DEVICE_FUNC
     RealScalar operatorNorm() const;
-    
-    #ifdef EIGEN2_SUPPORT
-    template<typename OtherDerived>
-    SelfAdjointView& operator=(const MatrixBase<OtherDerived>& other)
-    {
-      enum {
-        OtherPart = UpLo == Upper ? StrictlyLower : StrictlyUpper
-      };
-      m_matrix.const_cast_derived().template triangularView<UpLo>() = other;
-      m_matrix.const_cast_derived().template triangularView<OtherPart>() = other.adjoint();
-      return *this;
-    }
-    template<typename OtherMatrixType, unsigned int OtherMode>
-    SelfAdjointView& operator=(const TriangularView<OtherMatrixType, OtherMode>& other)
-    {
-      enum {
-        OtherPart = UpLo == Upper ? StrictlyLower : StrictlyUpper
-      };
-      m_matrix.const_cast_derived().template triangularView<UpLo>() = other.toDenseMatrix();
-      m_matrix.const_cast_derived().template triangularView<OtherPart>() = other.toDenseMatrix().adjoint();
-      return *this;
-    }
-    #endif
 
   protected:
     MatrixTypeNested m_matrix;
@@ -201,90 +263,54 @@ template<typename MatrixType, unsigned int UpLo> class SelfAdjointView
 
 namespace internal {
 
-template<typename Derived1, typename Derived2, int UnrollCount, bool ClearOpposite>
-struct triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Upper), UnrollCount, ClearOpposite>
-{
-  enum {
-    col = (UnrollCount-1) / Derived1::RowsAtCompileTime,
-    row = (UnrollCount-1) % Derived1::RowsAtCompileTime
-  };
-
-  static inline void run(Derived1 &dst, const Derived2 &src)
-  {
-    triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Upper), UnrollCount-1, ClearOpposite>::run(dst, src);
-
-    if(row == col)
-      dst.coeffRef(row, col) = numext::real(src.coeff(row, col));
-    else if(row < col)
-      dst.coeffRef(col, row) = numext::conj(dst.coeffRef(row, col) = src.coeff(row, col));
-  }
-};
-
-template<typename Derived1, typename Derived2, bool ClearOpposite>
-struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Upper, 0, ClearOpposite>
-{
-  static inline void run(Derived1 &, const Derived2 &) {}
-};
-
-template<typename Derived1, typename Derived2, int UnrollCount, bool ClearOpposite>
-struct triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Lower), UnrollCount, ClearOpposite>
-{
-  enum {
-    col = (UnrollCount-1) / Derived1::RowsAtCompileTime,
-    row = (UnrollCount-1) % Derived1::RowsAtCompileTime
-  };
-
-  static inline void run(Derived1 &dst, const Derived2 &src)
-  {
-    triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Lower), UnrollCount-1, ClearOpposite>::run(dst, src);
-
-    if(row == col)
-      dst.coeffRef(row, col) = numext::real(src.coeff(row, col));
-    else if(row > col)
-      dst.coeffRef(col, row) = numext::conj(dst.coeffRef(row, col) = src.coeff(row, col));
-  }
-};
-
-template<typename Derived1, typename Derived2, bool ClearOpposite>
-struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Lower, 0, ClearOpposite>
+// TODO currently a selfadjoint expression has the form SelfAdjointView<.,.>
+//      in the future selfadjoint-ness should be defined by the expression traits
+//      such that Transpose<SelfAdjointView<.,.> > is valid. (currently TriangularBase::transpose() is overloaded to make it work)
+template<typename MatrixType, unsigned int Mode>
+struct evaluator_traits<SelfAdjointView<MatrixType,Mode> >
 {
-  static inline void run(Derived1 &, const Derived2 &) {}
+  typedef typename storage_kind_to_evaluator_kind<typename MatrixType::StorageKind>::Kind Kind;
+  typedef SelfAdjointShape Shape;
 };
 
-template<typename Derived1, typename Derived2, bool ClearOpposite>
-struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Upper, Dynamic, ClearOpposite>
+template<int UpLo, int SetOpposite, typename DstEvaluatorTypeT, typename SrcEvaluatorTypeT, typename Functor, int Version>
+class triangular_dense_assignment_kernel<UpLo,SelfAdjoint,SetOpposite,DstEvaluatorTypeT,SrcEvaluatorTypeT,Functor,Version>
+  : public generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, Functor, Version>
 {
-  typedef typename Derived1::Index Index;
-  static inline void run(Derived1 &dst, const Derived2 &src)
+protected:
+  typedef generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, Functor, Version> Base;
+  typedef typename Base::DstXprType DstXprType;
+  typedef typename Base::SrcXprType SrcXprType;
+  using Base::m_dst;
+  using Base::m_src;
+  using Base::m_functor;
+public:
+  
+  typedef typename Base::DstEvaluatorType DstEvaluatorType;
+  typedef typename Base::SrcEvaluatorType SrcEvaluatorType;
+  typedef typename Base::Scalar Scalar;
+  typedef typename Base::AssignmentTraits AssignmentTraits;
+  
+  
+  EIGEN_DEVICE_FUNC triangular_dense_assignment_kernel(DstEvaluatorType &dst, const SrcEvaluatorType &src, const Functor &func, DstXprType& dstExpr)
+    : Base(dst, src, func, dstExpr)
+  {}
+  
+  EIGEN_DEVICE_FUNC void assignCoeff(Index row, Index col)
   {
-    for(Index j = 0; j < dst.cols(); ++j)
-    {
-      for(Index i = 0; i < j; ++i)
-      {
-        dst.copyCoeff(i, j, src);
-        dst.coeffRef(j,i) = numext::conj(dst.coeff(i,j));
-      }
-      dst.copyCoeff(j, j, src);
-    }
+    eigen_internal_assert(row!=col);
+    Scalar tmp = m_src.coeff(row,col);
+    m_functor.assignCoeff(m_dst.coeffRef(row,col), tmp);
+    m_functor.assignCoeff(m_dst.coeffRef(col,row), numext::conj(tmp));
   }
-};
-
-template<typename Derived1, typename Derived2, bool ClearOpposite>
-struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Lower, Dynamic, ClearOpposite>
-{
-  static inline void run(Derived1 &dst, const Derived2 &src)
+  
+  EIGEN_DEVICE_FUNC void assignDiagonalCoeff(Index id)
   {
-  typedef typename Derived1::Index Index;
-    for(Index i = 0; i < dst.rows(); ++i)
-    {
-      for(Index j = 0; j < i; ++j)
-      {
-        dst.copyCoeff(i, j, src);
-        dst.coeffRef(j,i) = numext::conj(dst.coeff(i,j));
-      }
-      dst.copyCoeff(i, i, src);
-    }
+    Base::assignCoeff(id,id);
   }
+  
+  EIGEN_DEVICE_FUNC void assignOppositeCoeff(Index, Index)
+  { eigen_internal_assert(false && "should never be called"); }
 };
 
 } // end namespace internal
@@ -293,20 +319,30 @@ struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Lower, Dyn
 * Implementation of MatrixBase methods
 ***************************************************************************/
 
+/** This is the const version of MatrixBase::selfadjointView() */
 template<typename Derived>
 template<unsigned int UpLo>
-typename MatrixBase<Derived>::template ConstSelfAdjointViewReturnType<UpLo>::Type
+EIGEN_DEVICE_FUNC typename MatrixBase<Derived>::template ConstSelfAdjointViewReturnType<UpLo>::Type
 MatrixBase<Derived>::selfadjointView() const
 {
-  return derived();
+  return typename ConstSelfAdjointViewReturnType<UpLo>::Type(derived());
 }
 
+/** \returns an expression of a symmetric/self-adjoint view extracted from the upper or lower triangular part of the current matrix
+  *
+  * The parameter \a UpLo can be either \c #Upper or \c #Lower
+  *
+  * Example: \include MatrixBase_selfadjointView.cpp
+  * Output: \verbinclude MatrixBase_selfadjointView.out
+  *
+  * \sa class SelfAdjointView
+  */
 template<typename Derived>
 template<unsigned int UpLo>
-typename MatrixBase<Derived>::template SelfAdjointViewReturnType<UpLo>::Type
+EIGEN_DEVICE_FUNC typename MatrixBase<Derived>::template SelfAdjointViewReturnType<UpLo>::Type
 MatrixBase<Derived>::selfadjointView()
 {
-  return derived();
+  return typename SelfAdjointViewReturnType<UpLo>::Type(derived());
 }
 
 } // end namespace Eigen
diff --git a/eigen/Eigen/src/Core/SelfCwiseBinaryOp.h b/eigen/Eigen/src/Core/SelfCwiseBinaryOp.h
index 0956475..50099df 100644
--- a/eigen/Eigen/src/Core/SelfCwiseBinaryOp.h
+++ b/eigen/Eigen/src/Core/SelfCwiseBinaryOp.h
@@ -12,177 +12,37 @@
 
 namespace Eigen { 
 
-/** \class SelfCwiseBinaryOp
-  * \ingroup Core_Module
-  *
-  * \internal
-  *
-  * \brief Internal helper class for optimizing operators like +=, -=
-  *
-  * This is a pseudo expression class re-implementing the copyCoeff/copyPacket
-  * method to directly performs a +=/-= operations in an optimal way. In particular,
-  * this allows to make sure that the input/output data are loaded only once using
-  * aligned packet loads.
-  *
-  * \sa class SwapWrapper for a similar trick.
-  */
+// TODO generalize the scalar type of 'other'
 
-namespace internal {
-template<typename BinaryOp, typename Lhs, typename Rhs>
-struct traits<SelfCwiseBinaryOp<BinaryOp,Lhs,Rhs> >
-  : traits<CwiseBinaryOp<BinaryOp,Lhs,Rhs> >
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator*=(const Scalar& other)
 {
-  enum {
-    // Note that it is still a good idea to preserve the DirectAccessBit
-    // so that assign can correctly align the data.
-    Flags = traits<CwiseBinaryOp<BinaryOp,Lhs,Rhs> >::Flags | (Lhs::Flags&DirectAccessBit) | (Lhs::Flags&LvalueBit),
-    OuterStrideAtCompileTime = Lhs::OuterStrideAtCompileTime,
-    InnerStrideAtCompileTime = Lhs::InnerStrideAtCompileTime
-  };
-};
+  typedef typename Derived::PlainObject PlainObject;
+  internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::mul_assign_op<Scalar,Scalar>());
+  return derived();
 }
 
-template<typename BinaryOp, typename Lhs, typename Rhs> class SelfCwiseBinaryOp
-  : public internal::dense_xpr_base< SelfCwiseBinaryOp<BinaryOp, Lhs, Rhs> >::type
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& ArrayBase<Derived>::operator+=(const Scalar& other)
 {
-  public:
-
-    typedef typename internal::dense_xpr_base<SelfCwiseBinaryOp>::type Base;
-    EIGEN_DENSE_PUBLIC_INTERFACE(SelfCwiseBinaryOp)
-
-    typedef typename internal::packet_traits<Scalar>::type Packet;
-
-    inline SelfCwiseBinaryOp(Lhs& xpr, const BinaryOp& func = BinaryOp()) : m_matrix(xpr), m_functor(func) {}
-
-    inline Index rows() const { return m_matrix.rows(); }
-    inline Index cols() const { return m_matrix.cols(); }
-    inline Index outerStride() const { return m_matrix.outerStride(); }
-    inline Index innerStride() const { return m_matrix.innerStride(); }
-    inline const Scalar* data() const { return m_matrix.data(); }
-
-    // note that this function is needed by assign to correctly align loads/stores
-    // TODO make Assign use .data()
-    inline Scalar& coeffRef(Index row, Index col)
-    {
-      EIGEN_STATIC_ASSERT_LVALUE(Lhs)
-      return m_matrix.const_cast_derived().coeffRef(row, col);
-    }
-    inline const Scalar& coeffRef(Index row, Index col) const
-    {
-      return m_matrix.coeffRef(row, col);
-    }
-
-    // note that this function is needed by assign to correctly align loads/stores
-    // TODO make Assign use .data()
-    inline Scalar& coeffRef(Index index)
-    {
-      EIGEN_STATIC_ASSERT_LVALUE(Lhs)
-      return m_matrix.const_cast_derived().coeffRef(index);
-    }
-    inline const Scalar& coeffRef(Index index) const
-    {
-      return m_matrix.const_cast_derived().coeffRef(index);
-    }
-
-    template<typename OtherDerived>
-    void copyCoeff(Index row, Index col, const DenseBase<OtherDerived>& other)
-    {
-      OtherDerived& _other = other.const_cast_derived();
-      eigen_internal_assert(row >= 0 && row < rows()
-                         && col >= 0 && col < cols());
-      Scalar& tmp = m_matrix.coeffRef(row,col);
-      tmp = m_functor(tmp, _other.coeff(row,col));
-    }
-
-    template<typename OtherDerived>
-    void copyCoeff(Index index, const DenseBase<OtherDerived>& other)
-    {
-      OtherDerived& _other = other.const_cast_derived();
-      eigen_internal_assert(index >= 0 && index < m_matrix.size());
-      Scalar& tmp = m_matrix.coeffRef(index);
-      tmp = m_functor(tmp, _other.coeff(index));
-    }
-
-    template<typename OtherDerived, int StoreMode, int LoadMode>
-    void copyPacket(Index row, Index col, const DenseBase<OtherDerived>& other)
-    {
-      OtherDerived& _other = other.const_cast_derived();
-      eigen_internal_assert(row >= 0 && row < rows()
-                        && col >= 0 && col < cols());
-      m_matrix.template writePacket<StoreMode>(row, col,
-        m_functor.packetOp(m_matrix.template packet<StoreMode>(row, col),_other.template packet<LoadMode>(row, col)) );
-    }
-
-    template<typename OtherDerived, int StoreMode, int LoadMode>
-    void copyPacket(Index index, const DenseBase<OtherDerived>& other)
-    {
-      OtherDerived& _other = other.const_cast_derived();
-      eigen_internal_assert(index >= 0 && index < m_matrix.size());
-      m_matrix.template writePacket<StoreMode>(index,
-        m_functor.packetOp(m_matrix.template packet<StoreMode>(index),_other.template packet<LoadMode>(index)) );
-    }
-
-    // reimplement lazyAssign to handle complex *= real
-    // see CwiseBinaryOp ctor for details
-    template<typename RhsDerived>
-    EIGEN_STRONG_INLINE SelfCwiseBinaryOp& lazyAssign(const DenseBase<RhsDerived>& rhs)
-    {
-      EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Lhs,RhsDerived)
-      EIGEN_CHECK_BINARY_COMPATIBILIY(BinaryOp,typename Lhs::Scalar,typename RhsDerived::Scalar);
-      
-    #ifdef EIGEN_DEBUG_ASSIGN
-      internal::assign_traits<SelfCwiseBinaryOp, RhsDerived>::debug();
-    #endif
-      eigen_assert(rows() == rhs.rows() && cols() == rhs.cols());
-      internal::assign_impl<SelfCwiseBinaryOp, RhsDerived>::run(*this,rhs.derived());
-    #ifndef EIGEN_NO_DEBUG
-      this->checkTransposeAliasing(rhs.derived());
-    #endif
-      return *this;
-    }
-    
-    // overloaded to honor evaluation of special matrices
-    // maybe another solution would be to not use SelfCwiseBinaryOp
-    // at first...
-    SelfCwiseBinaryOp& operator=(const Rhs& _rhs)
-    {
-      typename internal::nested<Rhs>::type rhs(_rhs);
-      return Base::operator=(rhs);
-    }
-
-    Lhs& expression() const 
-    { 
-      return m_matrix;
-    }
-
-    const BinaryOp& functor() const 
-    { 
-      return m_functor;
-    }
-
-  protected:
-    Lhs& m_matrix;
-    const BinaryOp& m_functor;
-
-  private:
-    SelfCwiseBinaryOp& operator=(const SelfCwiseBinaryOp&);
-};
+  typedef typename Derived::PlainObject PlainObject;
+  internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::add_assign_op<Scalar,Scalar>());
+  return derived();
+}
 
 template<typename Derived>
-inline Derived& DenseBase<Derived>::operator*=(const Scalar& other)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& ArrayBase<Derived>::operator-=(const Scalar& other)
 {
   typedef typename Derived::PlainObject PlainObject;
-  SelfCwiseBinaryOp<internal::scalar_product_op<Scalar>, Derived, typename PlainObject::ConstantReturnType> tmp(derived());
-  tmp = PlainObject::Constant(rows(),cols(),other);
+  internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::sub_assign_op<Scalar,Scalar>());
   return derived();
 }
 
 template<typename Derived>
-inline Derived& DenseBase<Derived>::operator/=(const Scalar& other)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator/=(const Scalar& other)
 {
   typedef typename Derived::PlainObject PlainObject;
-  SelfCwiseBinaryOp<internal::scalar_quotient_op<Scalar>, Derived, typename PlainObject::ConstantReturnType> tmp(derived());
-  tmp = PlainObject::Constant(rows(),cols(), other);
+  internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::div_assign_op<Scalar,Scalar>());
   return derived();
 }
 
diff --git a/eigen/Eigen/src/Core/Solve.h b/eigen/Eigen/src/Core/Solve.h
new file mode 100644
index 0000000..960a585
--- /dev/null
+++ b/eigen/Eigen/src/Core/Solve.h
@@ -0,0 +1,188 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SOLVE_H
+#define EIGEN_SOLVE_H
+
+namespace Eigen {
+
+template<typename Decomposition, typename RhsType, typename StorageKind> class SolveImpl;
+  
+/** \class Solve
+  * \ingroup Core_Module
+  *
+  * \brief Pseudo expression representing a solving operation
+  *
+  * \tparam Decomposition the type of the matrix or decomposion object
+  * \tparam Rhstype the type of the right-hand side
+  *
+  * This class represents an expression of A.solve(B)
+  * and most of the time this is the only way it is used.
+  *
+  */
+namespace internal {
+
+// this solve_traits class permits to determine the evaluation type with respect to storage kind (Dense vs Sparse)
+template<typename Decomposition, typename RhsType,typename StorageKind> struct solve_traits;
+
+template<typename Decomposition, typename RhsType>
+struct solve_traits<Decomposition,RhsType,Dense>
+{
+  typedef Matrix<typename RhsType::Scalar,
+                 Decomposition::ColsAtCompileTime,
+                 RhsType::ColsAtCompileTime,
+                 RhsType::PlainObject::Options,
+                 Decomposition::MaxColsAtCompileTime,
+                 RhsType::MaxColsAtCompileTime> PlainObject;  
+};
+
+template<typename Decomposition, typename RhsType>
+struct traits<Solve<Decomposition, RhsType> >
+  : traits<typename solve_traits<Decomposition,RhsType,typename internal::traits<RhsType>::StorageKind>::PlainObject>
+{
+  typedef typename solve_traits<Decomposition,RhsType,typename internal::traits<RhsType>::StorageKind>::PlainObject PlainObject;
+  typedef typename promote_index_type<typename Decomposition::StorageIndex, typename RhsType::StorageIndex>::type StorageIndex;
+  typedef traits<PlainObject> BaseTraits;
+  enum {
+    Flags = BaseTraits::Flags & RowMajorBit,
+    CoeffReadCost = HugeCost
+  };
+};
+
+}
+
+
+template<typename Decomposition, typename RhsType>
+class Solve : public SolveImpl<Decomposition,RhsType,typename internal::traits<RhsType>::StorageKind>
+{
+public:
+  typedef typename internal::traits<Solve>::PlainObject PlainObject;
+  typedef typename internal::traits<Solve>::StorageIndex StorageIndex;
+  
+  Solve(const Decomposition &dec, const RhsType &rhs)
+    : m_dec(dec), m_rhs(rhs)
+  {}
+  
+  EIGEN_DEVICE_FUNC Index rows() const { return m_dec.cols(); }
+  EIGEN_DEVICE_FUNC Index cols() const { return m_rhs.cols(); }
+
+  EIGEN_DEVICE_FUNC const Decomposition& dec() const { return m_dec; }
+  EIGEN_DEVICE_FUNC const RhsType&       rhs() const { return m_rhs; }
+
+protected:
+  const Decomposition &m_dec;
+  const RhsType       &m_rhs;
+};
+
+
+// Specialization of the Solve expression for dense results
+template<typename Decomposition, typename RhsType>
+class SolveImpl<Decomposition,RhsType,Dense>
+  : public MatrixBase<Solve<Decomposition,RhsType> >
+{
+  typedef Solve<Decomposition,RhsType> Derived;
+  
+public:
+  
+  typedef MatrixBase<Solve<Decomposition,RhsType> > Base;
+  EIGEN_DENSE_PUBLIC_INTERFACE(Derived)
+
+private:
+  
+  Scalar coeff(Index row, Index col) const;
+  Scalar coeff(Index i) const;
+};
+
+// Generic API dispatcher
+template<typename Decomposition, typename RhsType, typename StorageKind>
+class SolveImpl : public internal::generic_xpr_base<Solve<Decomposition,RhsType>, MatrixXpr, StorageKind>::type
+{
+  public:
+    typedef typename internal::generic_xpr_base<Solve<Decomposition,RhsType>, MatrixXpr, StorageKind>::type Base;
+};
+
+namespace internal {
+
+// Evaluator of Solve -> eval into a temporary
+template<typename Decomposition, typename RhsType>
+struct evaluator<Solve<Decomposition,RhsType> >
+  : public evaluator<typename Solve<Decomposition,RhsType>::PlainObject>
+{
+  typedef Solve<Decomposition,RhsType> SolveType;
+  typedef typename SolveType::PlainObject PlainObject;
+  typedef evaluator<PlainObject> Base;
+
+  enum { Flags = Base::Flags | EvalBeforeNestingBit };
+  
+  EIGEN_DEVICE_FUNC explicit evaluator(const SolveType& solve)
+    : m_result(solve.rows(), solve.cols())
+  {
+    ::new (static_cast<Base*>(this)) Base(m_result);
+    solve.dec()._solve_impl(solve.rhs(), m_result);
+  }
+  
+protected:  
+  PlainObject m_result;
+};
+
+// Specialization for "dst = dec.solve(rhs)"
+// NOTE we need to specialize it for Dense2Dense to avoid ambiguous specialization error and a Sparse2Sparse specialization must exist somewhere
+template<typename DstXprType, typename DecType, typename RhsType, typename Scalar>
+struct Assignment<DstXprType, Solve<DecType,RhsType>, internal::assign_op<Scalar,Scalar>, Dense2Dense>
+{
+  typedef Solve<DecType,RhsType> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,Scalar> &)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+
+    src.dec()._solve_impl(src.rhs(), dst);
+  }
+};
+
+// Specialization for "dst = dec.transpose().solve(rhs)"
+template<typename DstXprType, typename DecType, typename RhsType, typename Scalar>
+struct Assignment<DstXprType, Solve<Transpose<const DecType>,RhsType>, internal::assign_op<Scalar,Scalar>, Dense2Dense>
+{
+  typedef Solve<Transpose<const DecType>,RhsType> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,Scalar> &)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+
+    src.dec().nestedExpression().template _solve_impl_transposed<false>(src.rhs(), dst);
+  }
+};
+
+// Specialization for "dst = dec.adjoint().solve(rhs)"
+template<typename DstXprType, typename DecType, typename RhsType, typename Scalar>
+struct Assignment<DstXprType, Solve<CwiseUnaryOp<internal::scalar_conjugate_op<typename DecType::Scalar>, const Transpose<const DecType> >,RhsType>,
+                  internal::assign_op<Scalar,Scalar>, Dense2Dense>
+{
+  typedef Solve<CwiseUnaryOp<internal::scalar_conjugate_op<typename DecType::Scalar>, const Transpose<const DecType> >,RhsType> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,Scalar> &)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+    
+    src.dec().nestedExpression().nestedExpression().template _solve_impl_transposed<true>(src.rhs(), dst);
+  }
+};
+
+} // end namepsace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SOLVE_H
diff --git a/eigen/Eigen/src/Core/SolveTriangular.h b/eigen/Eigen/src/Core/SolveTriangular.h
index 30c9c38..a0011d4 100644
--- a/eigen/Eigen/src/Core/SolveTriangular.h
+++ b/eigen/Eigen/src/Core/SolveTriangular.h
@@ -68,7 +68,7 @@ struct triangular_solver_selector<Lhs,Rhs,Side,Mode,NoUnrolling,1>
     if(!useRhsDirectly)
       MappedRhs(actualRhs,rhs.size()) = rhs;
 
-    triangular_solve_vector<LhsScalar, RhsScalar, typename Lhs::Index, Side, Mode, LhsProductTraits::NeedToConjugate,
+    triangular_solve_vector<LhsScalar, RhsScalar, Index, Side, Mode, LhsProductTraits::NeedToConjugate,
                             (int(Lhs::Flags) & RowMajorBit) ? RowMajor : ColMajor>
       ::run(actualLhs.cols(), actualLhs.data(), actualLhs.outerStride(), actualRhs);
 
@@ -82,7 +82,6 @@ template<typename Lhs, typename Rhs, int Side, int Mode>
 struct triangular_solver_selector<Lhs,Rhs,Side,Mode,NoUnrolling,Dynamic>
 {
   typedef typename Rhs::Scalar Scalar;
-  typedef typename Rhs::Index Index;
   typedef blas_traits<Lhs> LhsProductTraits;
   typedef typename LhsProductTraits::DirectLinearAccessType ActualLhsType;
 
@@ -96,7 +95,7 @@ struct triangular_solver_selector<Lhs,Rhs,Side,Mode,NoUnrolling,Dynamic>
     typedef internal::gemm_blocking_space<(Rhs::Flags&RowMajorBit) ? RowMajor : ColMajor,Scalar,Scalar,
               Rhs::MaxRowsAtCompileTime, Rhs::MaxColsAtCompileTime, Lhs::MaxRowsAtCompileTime,4> BlockingType;
 
-    BlockingType blocking(rhs.rows(), rhs.cols(), size);
+    BlockingType blocking(rhs.rows(), rhs.cols(), size, 1, false);
 
     triangular_solve_matrix<Scalar,Index,Side,Mode,LhsProductTraits::NeedToConjugate,(int(Lhs::Flags) & RowMajorBit) ? RowMajor : ColMajor,
                                (Rhs::Flags&RowMajorBit) ? RowMajor : ColMajor>
@@ -108,32 +107,32 @@ struct triangular_solver_selector<Lhs,Rhs,Side,Mode,NoUnrolling,Dynamic>
 * meta-unrolling implementation
 ***************************************************************************/
 
-template<typename Lhs, typename Rhs, int Mode, int Index, int Size,
-         bool Stop = Index==Size>
+template<typename Lhs, typename Rhs, int Mode, int LoopIndex, int Size,
+         bool Stop = LoopIndex==Size>
 struct triangular_solver_unroller;
 
-template<typename Lhs, typename Rhs, int Mode, int Index, int Size>
-struct triangular_solver_unroller<Lhs,Rhs,Mode,Index,Size,false> {
+template<typename Lhs, typename Rhs, int Mode, int LoopIndex, int Size>
+struct triangular_solver_unroller<Lhs,Rhs,Mode,LoopIndex,Size,false> {
   enum {
     IsLower = ((Mode&Lower)==Lower),
-    RowIndex = IsLower ? Index : Size - Index - 1,
-    S = IsLower ? 0     : RowIndex+1
+    DiagIndex  = IsLower ? LoopIndex : Size - LoopIndex - 1,
+    StartIndex = IsLower ? 0         : DiagIndex+1
   };
   static void run(const Lhs& lhs, Rhs& rhs)
   {
-    if (Index>0)
-      rhs.coeffRef(RowIndex) -= lhs.row(RowIndex).template segment<Index>(S).transpose()
-                         .cwiseProduct(rhs.template segment<Index>(S)).sum();
+    if (LoopIndex>0)
+      rhs.coeffRef(DiagIndex) -= lhs.row(DiagIndex).template segment<LoopIndex>(StartIndex).transpose()
+                                .cwiseProduct(rhs.template segment<LoopIndex>(StartIndex)).sum();
 
     if(!(Mode & UnitDiag))
-      rhs.coeffRef(RowIndex) /= lhs.coeff(RowIndex,RowIndex);
+      rhs.coeffRef(DiagIndex) /= lhs.coeff(DiagIndex,DiagIndex);
 
-    triangular_solver_unroller<Lhs,Rhs,Mode,Index+1,Size>::run(lhs,rhs);
+    triangular_solver_unroller<Lhs,Rhs,Mode,LoopIndex+1,Size>::run(lhs,rhs);
   }
 };
 
-template<typename Lhs, typename Rhs, int Mode, int Index, int Size>
-struct triangular_solver_unroller<Lhs,Rhs,Mode,Index,Size,true> {
+template<typename Lhs, typename Rhs, int Mode, int LoopIndex, int Size>
+struct triangular_solver_unroller<Lhs,Rhs,Mode,LoopIndex,Size,true> {
   static void run(const Lhs&, Rhs&) {}
 };
 
@@ -162,61 +161,35 @@ struct triangular_solver_selector<Lhs,Rhs,OnTheRight,Mode,CompleteUnrolling,1> {
 * TriangularView methods
 ***************************************************************************/
 
-/** "in-place" version of TriangularView::solve() where the result is written in \a other
-  *
-  * \warning The parameter is only marked 'const' to make the C++ compiler accept a temporary expression here.
-  * This function will const_cast it, so constness isn't honored here.
-  *
-  * See TriangularView:solve() for the details.
-  */
+#ifndef EIGEN_PARSED_BY_DOXYGEN
 template<typename MatrixType, unsigned int Mode>
 template<int Side, typename OtherDerived>
-void TriangularView<MatrixType,Mode>::solveInPlace(const MatrixBase<OtherDerived>& _other) const
+EIGEN_DEVICE_FUNC void TriangularViewImpl<MatrixType,Mode,Dense>::solveInPlace(const MatrixBase<OtherDerived>& _other) const
 {
   OtherDerived& other = _other.const_cast_derived();
-  eigen_assert( cols() == rows() && ((Side==OnTheLeft && cols() == other.rows()) || (Side==OnTheRight && cols() == other.cols())) );
+  eigen_assert( derived().cols() == derived().rows() && ((Side==OnTheLeft && derived().cols() == other.rows()) || (Side==OnTheRight && derived().cols() == other.cols())) );
   eigen_assert((!(Mode & ZeroDiag)) && bool(Mode & (Upper|Lower)));
 
-  enum { copy = internal::traits<OtherDerived>::Flags & RowMajorBit  && OtherDerived::IsVectorAtCompileTime };
+  enum { copy = (internal::traits<OtherDerived>::Flags & RowMajorBit)  && OtherDerived::IsVectorAtCompileTime && OtherDerived::SizeAtCompileTime!=1};
   typedef typename internal::conditional<copy,
     typename internal::plain_matrix_type_column_major<OtherDerived>::type, OtherDerived&>::type OtherCopy;
   OtherCopy otherCopy(other);
 
   internal::triangular_solver_selector<MatrixType, typename internal::remove_reference<OtherCopy>::type,
-    Side, Mode>::run(nestedExpression(), otherCopy);
+    Side, Mode>::run(derived().nestedExpression(), otherCopy);
 
   if (copy)
     other = otherCopy;
 }
 
-/** \returns the product of the inverse of \c *this with \a other, \a *this being triangular.
-  *
-  * This function computes the inverse-matrix matrix product inverse(\c *this) * \a other if
-  * \a Side==OnTheLeft (the default), or the right-inverse-multiply  \a other * inverse(\c *this) if
-  * \a Side==OnTheRight.
-  *
-  * The matrix \c *this must be triangular and invertible (i.e., all the coefficients of the
-  * diagonal must be non zero). It works as a forward (resp. backward) substitution if \c *this
-  * is an upper (resp. lower) triangular matrix.
-  *
-  * Example: \include MatrixBase_marked.cpp
-  * Output: \verbinclude MatrixBase_marked.out
-  *
-  * This function returns an expression of the inverse-multiply and can works in-place if it is assigned
-  * to the same matrix or vector \a other.
-  *
-  * For users coming from BLAS, this function (and more specifically solveInPlace()) offer
-  * all the operations supported by the \c *TRSV and \c *TRSM BLAS routines.
-  *
-  * \sa TriangularView::solveInPlace()
-  */
 template<typename Derived, unsigned int Mode>
 template<int Side, typename Other>
 const internal::triangular_solve_retval<Side,TriangularView<Derived,Mode>,Other>
-TriangularView<Derived,Mode>::solve(const MatrixBase<Other>& other) const
+TriangularViewImpl<Derived,Mode,Dense>::solve(const MatrixBase<Other>& other) const
 {
-  return internal::triangular_solve_retval<Side,TriangularView,Other>(*this, other.derived());
+  return internal::triangular_solve_retval<Side,TriangularViewType,Other>(derived(), other.derived());
 }
+#endif
 
 namespace internal {
 
@@ -232,7 +205,6 @@ template<int Side, typename TriangularType, typename Rhs> struct triangular_solv
 {
   typedef typename remove_all<typename Rhs::Nested>::type RhsNestedCleaned;
   typedef ReturnByValue<triangular_solve_retval> Base;
-  typedef typename Base::Index Index;
 
   triangular_solve_retval(const TriangularType& tri, const Rhs& rhs)
     : m_triangularMatrix(tri), m_rhs(rhs)
@@ -243,8 +215,7 @@ template<int Side, typename TriangularType, typename Rhs> struct triangular_solv
 
   template<typename Dest> inline void evalTo(Dest& dst) const
   {
-    const typename Dest::Scalar *dst_data = internal::extract_data(dst);
-    if(!(is_same<RhsNestedCleaned,Dest>::value && dst_data!=0 && extract_data(dst) == extract_data(m_rhs)))
+    if(!is_same_dense(dst,m_rhs))
       dst = m_rhs;
     m_triangularMatrix.template solveInPlace<Side>(dst);
   }
diff --git a/eigen/Eigen/src/Core/SolverBase.h b/eigen/Eigen/src/Core/SolverBase.h
new file mode 100644
index 0000000..8a4adc2
--- /dev/null
+++ b/eigen/Eigen/src/Core/SolverBase.h
@@ -0,0 +1,130 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SOLVERBASE_H
+#define EIGEN_SOLVERBASE_H
+
+namespace Eigen {
+
+namespace internal {
+
+
+
+} // end namespace internal
+
+/** \class SolverBase
+  * \brief A base class for matrix decomposition and solvers
+  *
+  * \tparam Derived the actual type of the decomposition/solver.
+  *
+  * Any matrix decomposition inheriting this base class provide the following API:
+  *
+  * \code
+  * MatrixType A, b, x;
+  * DecompositionType dec(A);
+  * x = dec.solve(b);             // solve A   * x = b
+  * x = dec.transpose().solve(b); // solve A^T * x = b
+  * x = dec.adjoint().solve(b);   // solve A'  * x = b
+  * \endcode
+  *
+  * \warning Currently, any other usage of transpose() and adjoint() are not supported and will produce compilation errors.
+  *
+  * \sa class PartialPivLU, class FullPivLU
+  */
+template<typename Derived>
+class SolverBase : public EigenBase<Derived>
+{
+  public:
+
+    typedef EigenBase<Derived> Base;
+    typedef typename internal::traits<Derived>::Scalar Scalar;
+    typedef Scalar CoeffReturnType;
+
+    enum {
+      RowsAtCompileTime = internal::traits<Derived>::RowsAtCompileTime,
+      ColsAtCompileTime = internal::traits<Derived>::ColsAtCompileTime,
+      SizeAtCompileTime = (internal::size_at_compile_time<internal::traits<Derived>::RowsAtCompileTime,
+                                                          internal::traits<Derived>::ColsAtCompileTime>::ret),
+      MaxRowsAtCompileTime = internal::traits<Derived>::MaxRowsAtCompileTime,
+      MaxColsAtCompileTime = internal::traits<Derived>::MaxColsAtCompileTime,
+      MaxSizeAtCompileTime = (internal::size_at_compile_time<internal::traits<Derived>::MaxRowsAtCompileTime,
+                                                             internal::traits<Derived>::MaxColsAtCompileTime>::ret),
+      IsVectorAtCompileTime = internal::traits<Derived>::MaxRowsAtCompileTime == 1
+                           || internal::traits<Derived>::MaxColsAtCompileTime == 1
+    };
+
+    /** Default constructor */
+    SolverBase()
+    {}
+
+    ~SolverBase()
+    {}
+
+    using Base::derived;
+
+    /** \returns an expression of the solution x of \f$ A x = b \f$ using the current decomposition of A.
+      */
+    template<typename Rhs>
+    inline const Solve<Derived, Rhs>
+    solve(const MatrixBase<Rhs>& b) const
+    {
+      eigen_assert(derived().rows()==b.rows() && "solve(): invalid number of rows of the right hand side matrix b");
+      return Solve<Derived, Rhs>(derived(), b.derived());
+    }
+
+    /** \internal the return type of transpose() */
+    typedef typename internal::add_const<Transpose<const Derived> >::type ConstTransposeReturnType;
+    /** \returns an expression of the transposed of the factored matrix.
+      *
+      * A typical usage is to solve for the transposed problem A^T x = b:
+      * \code x = dec.transpose().solve(b); \endcode
+      *
+      * \sa adjoint(), solve()
+      */
+    inline ConstTransposeReturnType transpose() const
+    {
+      return ConstTransposeReturnType(derived());
+    }
+
+    /** \internal the return type of adjoint() */
+    typedef typename internal::conditional<NumTraits<Scalar>::IsComplex,
+                        CwiseUnaryOp<internal::scalar_conjugate_op<Scalar>, ConstTransposeReturnType>,
+                        ConstTransposeReturnType
+                     >::type AdjointReturnType;
+    /** \returns an expression of the adjoint of the factored matrix
+      *
+      * A typical usage is to solve for the adjoint problem A' x = b:
+      * \code x = dec.adjoint().solve(b); \endcode
+      *
+      * For real scalar types, this function is equivalent to transpose().
+      *
+      * \sa transpose(), solve()
+      */
+    inline AdjointReturnType adjoint() const
+    {
+      return AdjointReturnType(derived().transpose());
+    }
+
+  protected:
+};
+
+namespace internal {
+
+template<typename Derived>
+struct generic_xpr_base<Derived, MatrixXpr, SolverStorage>
+{
+  typedef SolverBase<Derived> type;
+
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SOLVERBASE_H
diff --git a/eigen/Eigen/src/Core/StableNorm.h b/eigen/Eigen/src/Core/StableNorm.h
index 389d942..be04ed4 100644
--- a/eigen/Eigen/src/Core/StableNorm.h
+++ b/eigen/Eigen/src/Core/StableNorm.h
@@ -17,10 +17,9 @@ namespace internal {
 template<typename ExpressionType, typename Scalar>
 inline void stable_norm_kernel(const ExpressionType& bl, Scalar& ssq, Scalar& scale, Scalar& invScale)
 {
-  using std::max;
   Scalar maxCoeff = bl.cwiseAbs().maxCoeff();
   
-  if (maxCoeff>scale)
+  if(maxCoeff>scale)
   {
     ssq = ssq * numext::abs2(scale/maxCoeff);
     Scalar tmp = Scalar(1)/maxCoeff;
@@ -29,12 +28,21 @@ inline void stable_norm_kernel(const ExpressionType& bl, Scalar& ssq, Scalar& sc
       invScale = NumTraits<Scalar>::highest();
       scale = Scalar(1)/invScale;
     }
+    else if(maxCoeff>NumTraits<Scalar>::highest()) // we got a INF
+    {
+      invScale = Scalar(1);
+      scale = maxCoeff;
+    }
     else
     {
       scale = maxCoeff;
       invScale = tmp;
     }
   }
+  else if(maxCoeff!=maxCoeff) // we got a NaN
+  {
+    scale = maxCoeff;
+  }
   
   // TODO if the maxCoeff is much much smaller than the current scale,
   // then we can neglect this sub vector
@@ -47,15 +55,12 @@ inline typename NumTraits<typename traits<Derived>::Scalar>::Real
 blueNorm_impl(const EigenBase<Derived>& _vec)
 {
   typedef typename Derived::RealScalar RealScalar;  
-  typedef typename Derived::Index Index;
   using std::pow;
-  using std::min;
-  using std::max;
   using std::sqrt;
   using std::abs;
   const Derived& vec(_vec.derived());
   static bool initialized = false;
-  static RealScalar b1, b2, s1m, s2m, overfl, rbig, relerr;
+  static RealScalar b1, b2, s1m, s2m, rbig, relerr;
   if(!initialized)
   {
     int ibeta, it, iemin, iemax, iexp;
@@ -84,7 +89,6 @@ blueNorm_impl(const EigenBase<Derived>& _vec)
     iexp  = - ((iemax+it)/2);
     s2m   = RealScalar(pow(RealScalar(ibeta),RealScalar(iexp)));    // scaling factor for upper range
 
-    overfl  = rbig*s2m;                                             // overflow boundary for abig
     eps     = RealScalar(pow(double(ibeta), 1-it));
     relerr  = sqrt(eps);                                            // tolerance for neglecting asml
     initialized = true;
@@ -101,13 +105,13 @@ blueNorm_impl(const EigenBase<Derived>& _vec)
     else if(ax < b1) asml += numext::abs2(ax*s1m);
     else             amed += numext::abs2(ax);
   }
+  if(amed!=amed)
+    return amed;  // we got a NaN
   if(abig > RealScalar(0))
   {
     abig = sqrt(abig);
-    if(abig > overfl)
-    {
-      return rbig;
-    }
+    if(abig > rbig) // overflow, or *this contains INF values
+      return abig;  // return INF
     if(amed > RealScalar(0))
     {
       abig = abig/s2m;
@@ -128,8 +132,8 @@ blueNorm_impl(const EigenBase<Derived>& _vec)
   }
   else
     return sqrt(amed);
-  asml = (min)(abig, amed);
-  abig = (max)(abig, amed);
+  asml = numext::mini(abig, amed);
+  abig = numext::maxi(abig, amed);
   if(asml <= abig*relerr)
     return abig;
   else
@@ -152,21 +156,35 @@ template<typename Derived>
 inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real
 MatrixBase<Derived>::stableNorm() const
 {
-  using std::min;
   using std::sqrt;
+  using std::abs;
   const Index blockSize = 4096;
   RealScalar scale(0);
   RealScalar invScale(1);
   RealScalar ssq(0); // sum of square
+  
+  typedef typename internal::nested_eval<Derived,2>::type DerivedCopy;
+  typedef typename internal::remove_all<DerivedCopy>::type DerivedCopyClean;
+  DerivedCopy copy(derived());
+  
   enum {
-    Alignment = (int(Flags)&DirectAccessBit) || (int(Flags)&AlignedBit) ? 1 : 0
+    CanAlign = (   (int(DerivedCopyClean::Flags)&DirectAccessBit)
+                || (int(internal::evaluator<DerivedCopyClean>::Alignment)>0) // FIXME Alignment)>0 might not be enough
+               ) && (blockSize*sizeof(Scalar)*2<EIGEN_STACK_ALLOCATION_LIMIT)
+                 && (EIGEN_MAX_STATIC_ALIGN_BYTES>0) // if we cannot allocate on the stack, then let's not bother about this optimization
   };
+  typedef typename internal::conditional<CanAlign, Ref<const Matrix<Scalar,Dynamic,1,0,blockSize,1>, internal::evaluator<DerivedCopyClean>::Alignment>,
+                                                   typename DerivedCopyClean::ConstSegmentReturnType>::type SegmentWrapper;
   Index n = size();
-  Index bi = internal::first_aligned(derived());
+  
+  if(n==1)
+    return abs(this->coeff(0));
+  
+  Index bi = internal::first_default_aligned(copy);
   if (bi>0)
-    internal::stable_norm_kernel(this->head(bi), ssq, scale, invScale);
+    internal::stable_norm_kernel(copy.head(bi), ssq, scale, invScale);
   for (; bi<n; bi+=blockSize)
-    internal::stable_norm_kernel(this->segment(bi,(min)(blockSize, n - bi)).template forceAlignedAccessIf<Alignment>(), ssq, scale, invScale);
+    internal::stable_norm_kernel(SegmentWrapper(copy.segment(bi,numext::mini(blockSize, n - bi))), ssq, scale, invScale);
   return scale * sqrt(ssq);
 }
 
diff --git a/eigen/Eigen/src/Core/Stride.h b/eigen/Eigen/src/Core/Stride.h
index 1e3f5fe..513742f 100644
--- a/eigen/Eigen/src/Core/Stride.h
+++ b/eigen/Eigen/src/Core/Stride.h
@@ -31,8 +31,8 @@ namespace Eigen {
   * arguments to the constructor.
   *
   * Indeed, this class takes two template parameters:
-  *  \param _OuterStrideAtCompileTime the outer stride, or Dynamic if you want to specify it at runtime.
-  *  \param _InnerStrideAtCompileTime the inner stride, or Dynamic if you want to specify it at runtime.
+  *  \tparam _OuterStrideAtCompileTime the outer stride, or Dynamic if you want to specify it at runtime.
+  *  \tparam _InnerStrideAtCompileTime the inner stride, or Dynamic if you want to specify it at runtime.
   *
   * Here is an example:
   * \include Map_general_stride.cpp
@@ -44,13 +44,14 @@ template<int _OuterStrideAtCompileTime, int _InnerStrideAtCompileTime>
 class Stride
 {
   public:
-    typedef DenseIndex Index;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
     enum {
       InnerStrideAtCompileTime = _InnerStrideAtCompileTime,
       OuterStrideAtCompileTime = _OuterStrideAtCompileTime
     };
 
     /** Default constructor, for use when strides are fixed at compile time */
+    EIGEN_DEVICE_FUNC
     Stride()
       : m_outer(OuterStrideAtCompileTime), m_inner(InnerStrideAtCompileTime)
     {
@@ -58,6 +59,7 @@ class Stride
     }
 
     /** Constructor allowing to pass the strides at runtime */
+    EIGEN_DEVICE_FUNC
     Stride(Index outerStride, Index innerStride)
       : m_outer(outerStride), m_inner(innerStride)
     {
@@ -65,13 +67,16 @@ class Stride
     }
 
     /** Copy constructor */
+    EIGEN_DEVICE_FUNC
     Stride(const Stride& other)
       : m_outer(other.outer()), m_inner(other.inner())
     {}
 
     /** \returns the outer stride */
+    EIGEN_DEVICE_FUNC
     inline Index outer() const { return m_outer.value(); }
     /** \returns the inner stride */
+    EIGEN_DEVICE_FUNC
     inline Index inner() const { return m_inner.value(); }
 
   protected:
@@ -81,26 +86,24 @@ class Stride
 
 /** \brief Convenience specialization of Stride to specify only an inner stride
   * See class Map for some examples */
-template<int Value = Dynamic>
+template<int Value>
 class InnerStride : public Stride<0, Value>
 {
     typedef Stride<0, Value> Base;
   public:
-    typedef DenseIndex Index;
-    InnerStride() : Base() {}
-    InnerStride(Index v) : Base(0, v) {}
+    EIGEN_DEVICE_FUNC InnerStride() : Base() {}
+    EIGEN_DEVICE_FUNC InnerStride(Index v) : Base(0, v) {} // FIXME making this explicit could break valid code
 };
 
 /** \brief Convenience specialization of Stride to specify only an outer stride
   * See class Map for some examples */
-template<int Value = Dynamic>
+template<int Value>
 class OuterStride : public Stride<Value, 0>
 {
     typedef Stride<Value, 0> Base;
   public:
-    typedef DenseIndex Index;
-    OuterStride() : Base() {}
-    OuterStride(Index v) : Base(v,0) {}
+    EIGEN_DEVICE_FUNC OuterStride() : Base() {}
+    EIGEN_DEVICE_FUNC OuterStride(Index v) : Base(v,0) {} // FIXME making this explicit could break valid code
 };
 
 } // end namespace Eigen
diff --git a/eigen/Eigen/src/Core/Swap.h b/eigen/Eigen/src/Core/Swap.h
index bf58bd5..d702009 100644
--- a/eigen/Eigen/src/Core/Swap.h
+++ b/eigen/Eigen/src/Core/Swap.h
@@ -12,115 +12,56 @@
 
 namespace Eigen { 
 
-/** \class SwapWrapper
-  * \ingroup Core_Module
-  *
-  * \internal
-  *
-  * \brief Internal helper class for swapping two expressions
-  */
 namespace internal {
-template<typename ExpressionType>
-struct traits<SwapWrapper<ExpressionType> > : traits<ExpressionType> {};
-}
 
-template<typename ExpressionType> class SwapWrapper
-  : public internal::dense_xpr_base<SwapWrapper<ExpressionType> >::type
+// Overload default assignPacket behavior for swapping them
+template<typename DstEvaluatorTypeT, typename SrcEvaluatorTypeT>
+class generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, swap_assign_op<typename DstEvaluatorTypeT::Scalar>, Specialized>
+ : public generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, swap_assign_op<typename DstEvaluatorTypeT::Scalar>, BuiltIn>
 {
-  public:
-
-    typedef typename internal::dense_xpr_base<SwapWrapper>::type Base;
-    EIGEN_DENSE_PUBLIC_INTERFACE(SwapWrapper)
-    typedef typename internal::packet_traits<Scalar>::type Packet;
-
-    inline SwapWrapper(ExpressionType& xpr) : m_expression(xpr) {}
-
-    inline Index rows() const { return m_expression.rows(); }
-    inline Index cols() const { return m_expression.cols(); }
-    inline Index outerStride() const { return m_expression.outerStride(); }
-    inline Index innerStride() const { return m_expression.innerStride(); }
-    
-    typedef typename internal::conditional<
-                       internal::is_lvalue<ExpressionType>::value,
-                       Scalar,
-                       const Scalar
-                     >::type ScalarWithConstIfNotLvalue;
-                     
-    inline ScalarWithConstIfNotLvalue* data() { return m_expression.data(); }
-    inline const Scalar* data() const { return m_expression.data(); }
-
-    inline Scalar& coeffRef(Index rowId, Index colId)
-    {
-      return m_expression.const_cast_derived().coeffRef(rowId, colId);
-    }
-
-    inline Scalar& coeffRef(Index index)
-    {
-      return m_expression.const_cast_derived().coeffRef(index);
-    }
-
-    inline Scalar& coeffRef(Index rowId, Index colId) const
-    {
-      return m_expression.coeffRef(rowId, colId);
-    }
-
-    inline Scalar& coeffRef(Index index) const
-    {
-      return m_expression.coeffRef(index);
-    }
-
-    template<typename OtherDerived>
-    void copyCoeff(Index rowId, Index colId, const DenseBase<OtherDerived>& other)
-    {
-      OtherDerived& _other = other.const_cast_derived();
-      eigen_internal_assert(rowId >= 0 && rowId < rows()
-                         && colId >= 0 && colId < cols());
-      Scalar tmp = m_expression.coeff(rowId, colId);
-      m_expression.coeffRef(rowId, colId) = _other.coeff(rowId, colId);
-      _other.coeffRef(rowId, colId) = tmp;
-    }
-
-    template<typename OtherDerived>
-    void copyCoeff(Index index, const DenseBase<OtherDerived>& other)
-    {
-      OtherDerived& _other = other.const_cast_derived();
-      eigen_internal_assert(index >= 0 && index < m_expression.size());
-      Scalar tmp = m_expression.coeff(index);
-      m_expression.coeffRef(index) = _other.coeff(index);
-      _other.coeffRef(index) = tmp;
-    }
-
-    template<typename OtherDerived, int StoreMode, int LoadMode>
-    void copyPacket(Index rowId, Index colId, const DenseBase<OtherDerived>& other)
-    {
-      OtherDerived& _other = other.const_cast_derived();
-      eigen_internal_assert(rowId >= 0 && rowId < rows()
-                        && colId >= 0 && colId < cols());
-      Packet tmp = m_expression.template packet<StoreMode>(rowId, colId);
-      m_expression.template writePacket<StoreMode>(rowId, colId,
-        _other.template packet<LoadMode>(rowId, colId)
-      );
-      _other.template writePacket<LoadMode>(rowId, colId, tmp);
-    }
-
-    template<typename OtherDerived, int StoreMode, int LoadMode>
-    void copyPacket(Index index, const DenseBase<OtherDerived>& other)
-    {
-      OtherDerived& _other = other.const_cast_derived();
-      eigen_internal_assert(index >= 0 && index < m_expression.size());
-      Packet tmp = m_expression.template packet<StoreMode>(index);
-      m_expression.template writePacket<StoreMode>(index,
-        _other.template packet<LoadMode>(index)
-      );
-      _other.template writePacket<LoadMode>(index, tmp);
-    }
-
-    ExpressionType& expression() const { return m_expression; }
-
-  protected:
-    ExpressionType& m_expression;
+protected:
+  typedef generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, swap_assign_op<typename DstEvaluatorTypeT::Scalar>, BuiltIn> Base;
+  using Base::m_dst;
+  using Base::m_src;
+  using Base::m_functor;
+  
+public:
+  typedef typename Base::Scalar Scalar;
+  typedef typename Base::DstXprType DstXprType;
+  typedef swap_assign_op<Scalar> Functor;
+  
+  EIGEN_DEVICE_FUNC generic_dense_assignment_kernel(DstEvaluatorTypeT &dst, const SrcEvaluatorTypeT &src, const Functor &func, DstXprType& dstExpr)
+    : Base(dst, src, func, dstExpr)
+  {}
+  
+  template<int StoreMode, int LoadMode, typename PacketType>
+  void assignPacket(Index row, Index col)
+  {
+    PacketType tmp = m_src.template packet<LoadMode,PacketType>(row,col);
+    const_cast<SrcEvaluatorTypeT&>(m_src).template writePacket<LoadMode>(row,col, m_dst.template packet<StoreMode,PacketType>(row,col));
+    m_dst.template writePacket<StoreMode>(row,col,tmp);
+  }
+  
+  template<int StoreMode, int LoadMode, typename PacketType>
+  void assignPacket(Index index)
+  {
+    PacketType tmp = m_src.template packet<LoadMode,PacketType>(index);
+    const_cast<SrcEvaluatorTypeT&>(m_src).template writePacket<LoadMode>(index, m_dst.template packet<StoreMode,PacketType>(index));
+    m_dst.template writePacket<StoreMode>(index,tmp);
+  }
+  
+  // TODO find a simple way not to have to copy/paste this function from generic_dense_assignment_kernel, by simple I mean no CRTP (Gael)
+  template<int StoreMode, int LoadMode, typename PacketType>
+  void assignPacketByOuterInner(Index outer, Index inner)
+  {
+    Index row = Base::rowIndexByOuterInner(outer, inner); 
+    Index col = Base::colIndexByOuterInner(outer, inner);
+    assignPacket<StoreMode,LoadMode,PacketType>(row, col);
+  }
 };
 
+} // namespace internal
+
 } // end namespace Eigen
 
 #endif // EIGEN_SWAP_H
diff --git a/eigen/Eigen/src/Core/Transpose.h b/eigen/Eigen/src/Core/Transpose.h
index 2abce3c..ba7d6e6 100644
--- a/eigen/Eigen/src/Core/Transpose.h
+++ b/eigen/Eigen/src/Core/Transpose.h
@@ -2,7 +2,7 @@
 // for linear algebra.
 //
 // Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
-// Copyright (C) 2009-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2009-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
 //
 // This Source Code Form is subject to the terms of the Mozilla
 // Public License v. 2.0. If a copy of the MPL was not distributed
@@ -13,39 +13,21 @@
 
 namespace Eigen { 
 
-/** \class Transpose
-  * \ingroup Core_Module
-  *
-  * \brief Expression of the transpose of a matrix
-  *
-  * \param MatrixType the type of the object of which we are taking the transpose
-  *
-  * This class represents an expression of the transpose of a matrix.
-  * It is the return type of MatrixBase::transpose() and MatrixBase::adjoint()
-  * and most of the time this is the only way it is used.
-  *
-  * \sa MatrixBase::transpose(), MatrixBase::adjoint()
-  */
-
 namespace internal {
 template<typename MatrixType>
-struct traits<Transpose<MatrixType> > : traits<MatrixType>
+struct traits<Transpose<MatrixType> > : public traits<MatrixType>
 {
-  typedef typename MatrixType::Scalar Scalar;
-  typedef typename nested<MatrixType>::type MatrixTypeNested;
+  typedef typename ref_selector<MatrixType>::type MatrixTypeNested;
   typedef typename remove_reference<MatrixTypeNested>::type MatrixTypeNestedPlain;
-  typedef typename traits<MatrixType>::StorageKind StorageKind;
-  typedef typename traits<MatrixType>::XprKind XprKind;
   enum {
     RowsAtCompileTime = MatrixType::ColsAtCompileTime,
     ColsAtCompileTime = MatrixType::RowsAtCompileTime,
     MaxRowsAtCompileTime = MatrixType::MaxColsAtCompileTime,
     MaxColsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
     FlagsLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0,
-    Flags0 = MatrixTypeNestedPlain::Flags & ~(LvalueBit | NestByRefBit),
+    Flags0 = traits<MatrixTypeNestedPlain>::Flags & ~(LvalueBit | NestByRefBit),
     Flags1 = Flags0 | FlagsLvalueBit,
     Flags = Flags1 ^ RowMajorBit,
-    CoeffReadCost = MatrixTypeNestedPlain::CoeffReadCost,
     InnerStrideAtCompileTime = inner_stride_at_compile_time<MatrixType>::ret,
     OuterStrideAtCompileTime = outer_stride_at_compile_time<MatrixType>::ret
   };
@@ -54,31 +36,55 @@ struct traits<Transpose<MatrixType> > : traits<MatrixType>
 
 template<typename MatrixType, typename StorageKind> class TransposeImpl;
 
+/** \class Transpose
+  * \ingroup Core_Module
+  *
+  * \brief Expression of the transpose of a matrix
+  *
+  * \tparam MatrixType the type of the object of which we are taking the transpose
+  *
+  * This class represents an expression of the transpose of a matrix.
+  * It is the return type of MatrixBase::transpose() and MatrixBase::adjoint()
+  * and most of the time this is the only way it is used.
+  *
+  * \sa MatrixBase::transpose(), MatrixBase::adjoint()
+  */
 template<typename MatrixType> class Transpose
   : public TransposeImpl<MatrixType,typename internal::traits<MatrixType>::StorageKind>
 {
   public:
 
+    typedef typename internal::ref_selector<MatrixType>::non_const_type MatrixTypeNested;
+
     typedef typename TransposeImpl<MatrixType,typename internal::traits<MatrixType>::StorageKind>::Base Base;
     EIGEN_GENERIC_PUBLIC_INTERFACE(Transpose)
+    typedef typename internal::remove_all<MatrixType>::type NestedExpression;
 
-    inline Transpose(MatrixType& a_matrix) : m_matrix(a_matrix) {}
+    EIGEN_DEVICE_FUNC
+    explicit inline Transpose(MatrixType& matrix) : m_matrix(matrix) {}
 
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Transpose)
 
-    inline Index rows() const { return m_matrix.cols(); }
-    inline Index cols() const { return m_matrix.rows(); }
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_matrix.cols(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_matrix.rows(); }
 
     /** \returns the nested expression */
-    const typename internal::remove_all<typename MatrixType::Nested>::type&
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<MatrixTypeNested>::type&
     nestedExpression() const { return m_matrix; }
 
     /** \returns the nested expression */
-    typename internal::remove_all<typename MatrixType::Nested>::type&
-    nestedExpression() { return m_matrix.const_cast_derived(); }
+    EIGEN_DEVICE_FUNC
+    typename internal::remove_reference<MatrixTypeNested>::type&
+    nestedExpression() { return m_matrix; }
+
+    /** \internal */
+    void resize(Index nrows, Index ncols) {
+      m_matrix.resize(ncols,nrows);
+    }
 
   protected:
-    typename MatrixType::Nested m_matrix;
+    typename internal::ref_selector<MatrixType>::non_const_type m_matrix;
 };
 
 namespace internal {
@@ -97,17 +103,27 @@ struct TransposeImpl_base<MatrixType, false>
 
 } // end namespace internal
 
+// Generic API dispatcher
+template<typename XprType, typename StorageKind>
+class TransposeImpl
+  : public internal::generic_xpr_base<Transpose<XprType> >::type
+{
+public:
+  typedef typename internal::generic_xpr_base<Transpose<XprType> >::type Base;
+};
+
 template<typename MatrixType> class TransposeImpl<MatrixType,Dense>
   : public internal::TransposeImpl_base<MatrixType>::type
 {
   public:
 
     typedef typename internal::TransposeImpl_base<MatrixType>::type Base;
+    using Base::coeffRef;
     EIGEN_DENSE_PUBLIC_INTERFACE(Transpose<MatrixType>)
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(TransposeImpl)
 
-    inline Index innerStride() const { return derived().nestedExpression().innerStride(); }
-    inline Index outerStride() const { return derived().nestedExpression().outerStride(); }
+    EIGEN_DEVICE_FUNC inline Index innerStride() const { return derived().nestedExpression().innerStride(); }
+    EIGEN_DEVICE_FUNC inline Index outerStride() const { return derived().nestedExpression().outerStride(); }
 
     typedef typename internal::conditional<
                        internal::is_lvalue<MatrixType>::value,
@@ -115,64 +131,21 @@ template<typename MatrixType> class TransposeImpl<MatrixType,Dense>
                        const Scalar
                      >::type ScalarWithConstIfNotLvalue;
 
-    inline ScalarWithConstIfNotLvalue* data() { return derived().nestedExpression().data(); }
-    inline const Scalar* data() const { return derived().nestedExpression().data(); }
-
-    inline ScalarWithConstIfNotLvalue& coeffRef(Index rowId, Index colId)
-    {
-      EIGEN_STATIC_ASSERT_LVALUE(MatrixType)
-      return derived().nestedExpression().const_cast_derived().coeffRef(colId, rowId);
-    }
-
-    inline ScalarWithConstIfNotLvalue& coeffRef(Index index)
-    {
-      EIGEN_STATIC_ASSERT_LVALUE(MatrixType)
-      return derived().nestedExpression().const_cast_derived().coeffRef(index);
-    }
+    EIGEN_DEVICE_FUNC inline ScalarWithConstIfNotLvalue* data() { return derived().nestedExpression().data(); }
+    EIGEN_DEVICE_FUNC inline const Scalar* data() const { return derived().nestedExpression().data(); }
 
+    // FIXME: shall we keep the const version of coeffRef?
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index rowId, Index colId) const
     {
       return derived().nestedExpression().coeffRef(colId, rowId);
     }
 
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index index) const
     {
       return derived().nestedExpression().coeffRef(index);
     }
-
-    inline CoeffReturnType coeff(Index rowId, Index colId) const
-    {
-      return derived().nestedExpression().coeff(colId, rowId);
-    }
-
-    inline CoeffReturnType coeff(Index index) const
-    {
-      return derived().nestedExpression().coeff(index);
-    }
-
-    template<int LoadMode>
-    inline const PacketScalar packet(Index rowId, Index colId) const
-    {
-      return derived().nestedExpression().template packet<LoadMode>(colId, rowId);
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index rowId, Index colId, const PacketScalar& x)
-    {
-      derived().nestedExpression().const_cast_derived().template writePacket<LoadMode>(colId, rowId, x);
-    }
-
-    template<int LoadMode>
-    inline const PacketScalar packet(Index index) const
-    {
-      return derived().nestedExpression().template packet<LoadMode>(index);
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index index, const PacketScalar& x)
-    {
-      derived().nestedExpression().const_cast_derived().template writePacket<LoadMode>(index, x);
-    }
 };
 
 /** \returns an expression of the transpose of *this.
@@ -195,10 +168,10 @@ template<typename MatrixType> class TransposeImpl<MatrixType,Dense>
   *
   * \sa transposeInPlace(), adjoint() */
 template<typename Derived>
-inline Transpose<Derived>
+EIGEN_DEVICE_FUNC inline Transpose<Derived>
 DenseBase<Derived>::transpose()
 {
-  return derived();
+  return TransposeReturnType(derived());
 }
 
 /** This is the const version of transpose().
@@ -207,7 +180,7 @@ DenseBase<Derived>::transpose()
   *
   * \sa transposeInPlace(), adjoint() */
 template<typename Derived>
-inline typename DenseBase<Derived>::ConstTransposeReturnType
+EIGEN_DEVICE_FUNC inline typename DenseBase<Derived>::ConstTransposeReturnType
 DenseBase<Derived>::transpose() const
 {
   return ConstTransposeReturnType(derived());
@@ -233,11 +206,10 @@ DenseBase<Derived>::transpose() const
   *
   * \sa adjointInPlace(), transpose(), conjugate(), class Transpose, class internal::scalar_conjugate_op */
 template<typename Derived>
-inline const typename MatrixBase<Derived>::AdjointReturnType
+EIGEN_DEVICE_FUNC inline const typename MatrixBase<Derived>::AdjointReturnType
 MatrixBase<Derived>::adjoint() const
 {
-  return this->transpose(); // in the complex case, the .conjugate() is be implicit here
-                            // due to implicit conversion to return type
+  return AdjointReturnType(this->transpose());
 }
 
 /***************************************************************************
@@ -247,18 +219,38 @@ MatrixBase<Derived>::adjoint() const
 namespace internal {
 
 template<typename MatrixType,
-  bool IsSquare = (MatrixType::RowsAtCompileTime == MatrixType::ColsAtCompileTime) && MatrixType::RowsAtCompileTime!=Dynamic>
+  bool IsSquare = (MatrixType::RowsAtCompileTime == MatrixType::ColsAtCompileTime) && MatrixType::RowsAtCompileTime!=Dynamic,
+  bool MatchPacketSize =
+        (int(MatrixType::RowsAtCompileTime) == int(internal::packet_traits<typename MatrixType::Scalar>::size))
+    &&  (internal::evaluator<MatrixType>::Flags&PacketAccessBit) >
 struct inplace_transpose_selector;
 
 template<typename MatrixType>
-struct inplace_transpose_selector<MatrixType,true> { // square matrix
+struct inplace_transpose_selector<MatrixType,true,false> { // square matrix
   static void run(MatrixType& m) {
     m.matrix().template triangularView<StrictlyUpper>().swap(m.matrix().transpose());
   }
 };
 
+// TODO: vectorized path is currently limited to LargestPacketSize x LargestPacketSize cases only.
 template<typename MatrixType>
-struct inplace_transpose_selector<MatrixType,false> { // non square matrix
+struct inplace_transpose_selector<MatrixType,true,true> { // PacketSize x PacketSize
+  static void run(MatrixType& m) {
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename internal::packet_traits<typename MatrixType::Scalar>::type Packet;
+    const Index PacketSize = internal::packet_traits<Scalar>::size;
+    const Index Alignment = internal::evaluator<MatrixType>::Alignment;
+    PacketBlock<Packet> A;
+    for (Index i=0; i<PacketSize; ++i)
+      A.packet[i] = m.template packetByOuterInner<Alignment>(i,0);
+    internal::ptranspose(A);
+    for (Index i=0; i<PacketSize; ++i)
+      m.template writePacket<Alignment>(m.rowIndexByOuterInner(i,0), m.colIndexByOuterInner(i,0), A.packet[i]);
+  }
+};
+
+template<typename MatrixType,bool MatchPacketSize>
+struct inplace_transpose_selector<MatrixType,false,MatchPacketSize> { // non square matrix
   static void run(MatrixType& m) {
     if (m.rows()==m.cols())
       m.matrix().template triangularView<StrictlyUpper>().swap(m.matrix().transpose());
@@ -289,7 +281,7 @@ struct inplace_transpose_selector<MatrixType,false> { // non square matrix
   *
   * \sa transpose(), adjoint(), adjointInPlace() */
 template<typename Derived>
-inline void DenseBase<Derived>::transposeInPlace()
+EIGEN_DEVICE_FUNC inline void DenseBase<Derived>::transposeInPlace()
 {
   eigen_assert((rows() == cols() || (RowsAtCompileTime == Dynamic && ColsAtCompileTime == Dynamic))
                && "transposeInPlace() called on a non-square non-resizable matrix");
@@ -320,7 +312,7 @@ inline void DenseBase<Derived>::transposeInPlace()
   *
   * \sa transpose(), adjoint(), transposeInPlace() */
 template<typename Derived>
-inline void MatrixBase<Derived>::adjointInPlace()
+EIGEN_DEVICE_FUNC inline void MatrixBase<Derived>::adjointInPlace()
 {
   derived() = adjoint().eval();
 }
@@ -331,14 +323,6 @@ inline void MatrixBase<Derived>::adjointInPlace()
 
 namespace internal {
 
-template<typename BinOp,typename Xpr,typename Rhs>
-struct blas_traits<SelfCwiseBinaryOp<BinOp,Xpr,Rhs> >
- : blas_traits<typename internal::remove_all<typename Xpr::Nested>::type>
-{
-  typedef SelfCwiseBinaryOp<BinOp,Xpr,Rhs> XprType;
-  static inline const XprType extract(const XprType& x) { return x; }
-};
-
 template<bool DestIsTransposed, typename OtherDerived>
 struct check_transpose_aliasing_compile_time_selector
 {
@@ -392,6 +376,7 @@ struct checkTransposeAliasing_impl
                       ::run(extract_data(dst), other))
           && "aliasing detected during transposition, use transposeInPlace() "
              "or evaluate the rhs into a temporary using .eval()");
+
     }
 };
 
@@ -403,15 +388,15 @@ struct checkTransposeAliasing_impl<Derived, OtherDerived, false>
     }
 };
 
-} // end namespace internal
-
-template<typename Derived>
-template<typename OtherDerived>
-void DenseBase<Derived>::checkTransposeAliasing(const OtherDerived& other) const
+template<typename Dst, typename Src>
+void check_for_aliasing(const Dst &dst, const Src &src)
 {
-    internal::checkTransposeAliasing_impl<Derived, OtherDerived>::run(derived(), other);
+  internal::checkTransposeAliasing_impl<Dst, Src>::run(dst, src);
 }
-#endif
+
+} // end namespace internal
+
+#endif // EIGEN_NO_DEBUG
 
 } // end namespace Eigen
 
diff --git a/eigen/Eigen/src/Core/Transpositions.h b/eigen/Eigen/src/Core/Transpositions.h
index 16bc1ce..19c17bb 100644
--- a/eigen/Eigen/src/Core/Transpositions.h
+++ b/eigen/Eigen/src/Core/Transpositions.h
@@ -12,39 +12,6 @@
 
 namespace Eigen { 
 
-/** \class Transpositions
-  * \ingroup Core_Module
-  *
-  * \brief Represents a sequence of transpositions (row/column interchange)
-  *
-  * \param SizeAtCompileTime the number of transpositions, or Dynamic
-  * \param MaxSizeAtCompileTime the maximum number of transpositions, or Dynamic. This optional parameter defaults to SizeAtCompileTime. Most of the time, you should not have to specify it.
-  *
-  * This class represents a permutation transformation as a sequence of \em n transpositions
-  * \f$[T_{n-1} \ldots T_{i} \ldots T_{0}]\f$. It is internally stored as a vector of integers \c indices.
-  * Each transposition \f$ T_{i} \f$ applied on the left of a matrix (\f$ T_{i} M\f$) interchanges
-  * the rows \c i and \c indices[i] of the matrix \c M.
-  * A transposition applied on the right (e.g., \f$ M T_{i}\f$) yields a column interchange.
-  *
-  * Compared to the class PermutationMatrix, such a sequence of transpositions is what is
-  * computed during a decomposition with pivoting, and it is faster when applying the permutation in-place.
-  * 
-  * To apply a sequence of transpositions to a matrix, simply use the operator * as in the following example:
-  * \code
-  * Transpositions tr;
-  * MatrixXf mat;
-  * mat = tr * mat;
-  * \endcode
-  * In this example, we detect that the matrix appears on both side, and so the transpositions
-  * are applied in-place without any temporary or extra copy.
-  *
-  * \sa class PermutationMatrix
-  */
-
-namespace internal {
-template<typename TranspositionType, typename MatrixType, int Side, bool Transposed=false> struct transposition_matrix_product_retval;
-}
-
 template<typename Derived>
 class TranspositionsBase
 {
@@ -53,7 +20,8 @@ class TranspositionsBase
   public:
 
     typedef typename Traits::IndicesType IndicesType;
-    typedef typename IndicesType::Scalar Index;
+    typedef typename IndicesType::Scalar StorageIndex;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
 
     Derived& derived() { return *static_cast<Derived*>(this); }
     const Derived& derived() const { return *static_cast<const Derived*>(this); }
@@ -65,7 +33,7 @@ class TranspositionsBase
       indices() = other.indices();
       return derived();
     }
-
+    
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     /** This is a special case of the templated operator=. Its purpose is to
       * prevent a default operator= from hiding the templated operator=.
@@ -78,20 +46,24 @@ class TranspositionsBase
     #endif
 
     /** \returns the number of transpositions */
-    inline Index size() const { return indices().size(); }
+    Index size() const { return indices().size(); }
+    /** \returns the number of rows of the equivalent permutation matrix */
+    Index rows() const { return indices().size(); }
+    /** \returns the number of columns of the equivalent permutation matrix */
+    Index cols() const { return indices().size(); }
 
     /** Direct access to the underlying index vector */
-    inline const Index& coeff(Index i) const { return indices().coeff(i); }
+    inline const StorageIndex& coeff(Index i) const { return indices().coeff(i); }
     /** Direct access to the underlying index vector */
-    inline Index& coeffRef(Index i) { return indices().coeffRef(i); }
+    inline StorageIndex& coeffRef(Index i) { return indices().coeffRef(i); }
     /** Direct access to the underlying index vector */
-    inline const Index& operator()(Index i) const { return indices()(i); }
+    inline const StorageIndex& operator()(Index i) const { return indices()(i); }
     /** Direct access to the underlying index vector */
-    inline Index& operator()(Index i) { return indices()(i); }
+    inline StorageIndex& operator()(Index i) { return indices()(i); }
     /** Direct access to the underlying index vector */
-    inline const Index& operator[](Index i) const { return indices()(i); }
+    inline const StorageIndex& operator[](Index i) const { return indices()(i); }
     /** Direct access to the underlying index vector */
-    inline Index& operator[](Index i) { return indices()(i); }
+    inline StorageIndex& operator[](Index i) { return indices()(i); }
 
     /** const version of indices(). */
     const IndicesType& indices() const { return derived().indices(); }
@@ -99,7 +71,7 @@ class TranspositionsBase
     IndicesType& indices() { return derived().indices(); }
 
     /** Resizes to given size. */
-    inline void resize(int newSize)
+    inline void resize(Index newSize)
     {
       indices().resize(newSize);
     }
@@ -107,7 +79,7 @@ class TranspositionsBase
     /** Sets \c *this to represents an identity transformation */
     void setIdentity()
     {
-      for(int i = 0; i < indices().size(); ++i)
+      for(StorageIndex i = 0; i < indices().size(); ++i)
         coeffRef(i) = i;
     }
 
@@ -144,23 +116,53 @@ class TranspositionsBase
 };
 
 namespace internal {
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename IndexType>
-struct traits<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,IndexType> >
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex>
+struct traits<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex> >
+ : traits<PermutationMatrix<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex> >
 {
-  typedef IndexType Index;
-  typedef Matrix<Index, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1> IndicesType;
+  typedef Matrix<_StorageIndex, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1> IndicesType;
+  typedef TranspositionsStorage StorageKind;
 };
 }
 
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename IndexType>
-class Transpositions : public TranspositionsBase<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,IndexType> >
+/** \class Transpositions
+  * \ingroup Core_Module
+  *
+  * \brief Represents a sequence of transpositions (row/column interchange)
+  *
+  * \tparam SizeAtCompileTime the number of transpositions, or Dynamic
+  * \tparam MaxSizeAtCompileTime the maximum number of transpositions, or Dynamic. This optional parameter defaults to SizeAtCompileTime. Most of the time, you should not have to specify it.
+  *
+  * This class represents a permutation transformation as a sequence of \em n transpositions
+  * \f$[T_{n-1} \ldots T_{i} \ldots T_{0}]\f$. It is internally stored as a vector of integers \c indices.
+  * Each transposition \f$ T_{i} \f$ applied on the left of a matrix (\f$ T_{i} M\f$) interchanges
+  * the rows \c i and \c indices[i] of the matrix \c M.
+  * A transposition applied on the right (e.g., \f$ M T_{i}\f$) yields a column interchange.
+  *
+  * Compared to the class PermutationMatrix, such a sequence of transpositions is what is
+  * computed during a decomposition with pivoting, and it is faster when applying the permutation in-place.
+  *
+  * To apply a sequence of transpositions to a matrix, simply use the operator * as in the following example:
+  * \code
+  * Transpositions tr;
+  * MatrixXf mat;
+  * mat = tr * mat;
+  * \endcode
+  * In this example, we detect that the matrix appears on both side, and so the transpositions
+  * are applied in-place without any temporary or extra copy.
+  *
+  * \sa class PermutationMatrix
+  */
+
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex>
+class Transpositions : public TranspositionsBase<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex> >
 {
     typedef internal::traits<Transpositions> Traits;
   public:
 
     typedef TranspositionsBase<Transpositions> Base;
     typedef typename Traits::IndicesType IndicesType;
-    typedef typename IndicesType::Scalar Index;
+    typedef typename IndicesType::Scalar StorageIndex;
 
     inline Transpositions() {}
 
@@ -177,7 +179,7 @@ class Transpositions : public TranspositionsBase<Transpositions<SizeAtCompileTim
 
     /** Generic constructor from expression of the transposition indices. */
     template<typename Other>
-    explicit inline Transpositions(const MatrixBase<Other>& a_indices) : m_indices(a_indices)
+    explicit inline Transpositions(const MatrixBase<Other>& indices) : m_indices(indices)
     {}
 
     /** Copies the \a other transpositions into \c *this */
@@ -215,30 +217,32 @@ class Transpositions : public TranspositionsBase<Transpositions<SizeAtCompileTim
 
 
 namespace internal {
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename IndexType, int _PacketAccess>
-struct traits<Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,IndexType>,_PacketAccess> >
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex, int _PacketAccess>
+struct traits<Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex>,_PacketAccess> >
+ : traits<PermutationMatrix<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex> >
 {
-  typedef IndexType Index;
-  typedef Map<const Matrix<Index,SizeAtCompileTime,1,0,MaxSizeAtCompileTime,1>, _PacketAccess> IndicesType;
+  typedef Map<const Matrix<_StorageIndex,SizeAtCompileTime,1,0,MaxSizeAtCompileTime,1>, _PacketAccess> IndicesType;
+  typedef _StorageIndex StorageIndex;
+  typedef TranspositionsStorage StorageKind;
 };
 }
 
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename IndexType, int PacketAccess>
-class Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,IndexType>,PacketAccess>
- : public TranspositionsBase<Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,IndexType>,PacketAccess> >
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex, int PacketAccess>
+class Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex>,PacketAccess>
+ : public TranspositionsBase<Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex>,PacketAccess> >
 {
     typedef internal::traits<Map> Traits;
   public:
 
     typedef TranspositionsBase<Map> Base;
     typedef typename Traits::IndicesType IndicesType;
-    typedef typename IndicesType::Scalar Index;
+    typedef typename IndicesType::Scalar StorageIndex;
 
-    inline Map(const Index* indicesPtr)
+    explicit inline Map(const StorageIndex* indicesPtr)
       : m_indices(indicesPtr)
     {}
 
-    inline Map(const Index* indicesPtr, Index size)
+    inline Map(const StorageIndex* indicesPtr, Index size)
       : m_indices(indicesPtr,size)
     {}
 
@@ -274,9 +278,9 @@ class Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,IndexType>,Packe
 namespace internal {
 template<typename _IndicesType>
 struct traits<TranspositionsWrapper<_IndicesType> >
+ : traits<PermutationWrapper<_IndicesType> >
 {
-  typedef typename _IndicesType::Scalar Index;
-  typedef _IndicesType IndicesType;
+  typedef TranspositionsStorage StorageKind;
 };
 }
 
@@ -289,10 +293,10 @@ class TranspositionsWrapper
 
     typedef TranspositionsBase<TranspositionsWrapper> Base;
     typedef typename Traits::IndicesType IndicesType;
-    typedef typename IndicesType::Scalar Index;
+    typedef typename IndicesType::Scalar StorageIndex;
 
-    inline TranspositionsWrapper(IndicesType& a_indices)
-      : m_indices(a_indices)
+    explicit inline TranspositionsWrapper(IndicesType& indices)
+      : m_indices(indices)
     {}
 
     /** Copies the \a other transpositions into \c *this */
@@ -321,84 +325,46 @@ class TranspositionsWrapper
 
   protected:
 
-    const typename IndicesType::Nested m_indices;
+    typename IndicesType::Nested m_indices;
 };
 
+
+
 /** \returns the \a matrix with the \a transpositions applied to the columns.
   */
-template<typename Derived, typename TranspositionsDerived>
-inline const internal::transposition_matrix_product_retval<TranspositionsDerived, Derived, OnTheRight>
-operator*(const MatrixBase<Derived>& matrix,
-          const TranspositionsBase<TranspositionsDerived> &transpositions)
+template<typename MatrixDerived, typename TranspositionsDerived>
+EIGEN_DEVICE_FUNC
+const Product<MatrixDerived, TranspositionsDerived, AliasFreeProduct>
+operator*(const MatrixBase<MatrixDerived> &matrix,
+          const TranspositionsBase<TranspositionsDerived>& transpositions)
 {
-  return internal::transposition_matrix_product_retval
-           <TranspositionsDerived, Derived, OnTheRight>
-           (transpositions.derived(), matrix.derived());
+  return Product<MatrixDerived, TranspositionsDerived, AliasFreeProduct>
+            (matrix.derived(), transpositions.derived());
 }
 
 /** \returns the \a matrix with the \a transpositions applied to the rows.
   */
-template<typename Derived, typename TranspositionDerived>
-inline const internal::transposition_matrix_product_retval
-               <TranspositionDerived, Derived, OnTheLeft>
-operator*(const TranspositionsBase<TranspositionDerived> &transpositions,
-          const MatrixBase<Derived>& matrix)
+template<typename TranspositionsDerived, typename MatrixDerived>
+EIGEN_DEVICE_FUNC
+const Product<TranspositionsDerived, MatrixDerived, AliasFreeProduct>
+operator*(const TranspositionsBase<TranspositionsDerived> &transpositions,
+          const MatrixBase<MatrixDerived>& matrix)
 {
-  return internal::transposition_matrix_product_retval
-           <TranspositionDerived, Derived, OnTheLeft>
-           (transpositions.derived(), matrix.derived());
+  return Product<TranspositionsDerived, MatrixDerived, AliasFreeProduct>
+            (transpositions.derived(), matrix.derived());
 }
 
-namespace internal {
-
-template<typename TranspositionType, typename MatrixType, int Side, bool Transposed>
-struct traits<transposition_matrix_product_retval<TranspositionType, MatrixType, Side, Transposed> >
-{
-  typedef typename MatrixType::PlainObject ReturnType;
-};
-
-template<typename TranspositionType, typename MatrixType, int Side, bool Transposed>
-struct transposition_matrix_product_retval
- : public ReturnByValue<transposition_matrix_product_retval<TranspositionType, MatrixType, Side, Transposed> >
-{
-    typedef typename remove_all<typename MatrixType::Nested>::type MatrixTypeNestedCleaned;
-    typedef typename TranspositionType::Index Index;
-
-    transposition_matrix_product_retval(const TranspositionType& tr, const MatrixType& matrix)
-      : m_transpositions(tr), m_matrix(matrix)
-    {}
+// Template partial specialization for transposed/inverse transpositions
 
-    inline int rows() const { return m_matrix.rows(); }
-    inline int cols() const { return m_matrix.cols(); }
-
-    template<typename Dest> inline void evalTo(Dest& dst) const
-    {
-      const int size = m_transpositions.size();
-      Index j = 0;
-
-      const typename Dest::Scalar *dst_data = internal::extract_data(dst);
-      if(!(is_same<MatrixTypeNestedCleaned,Dest>::value && dst_data!=0 && dst_data == extract_data(m_matrix)))
-        dst = m_matrix;
-
-      for(int k=(Transposed?size-1:0) ; Transposed?k>=0:k<size ; Transposed?--k:++k)
-        if((j=m_transpositions.coeff(k))!=k)
-        {
-          if(Side==OnTheLeft)
-            dst.row(k).swap(dst.row(j));
-          else if(Side==OnTheRight)
-            dst.col(k).swap(dst.col(j));
-        }
-    }
+namespace internal {
 
-  protected:
-    const TranspositionType& m_transpositions;
-    typename MatrixType::Nested m_matrix;
-};
+template<typename Derived>
+struct traits<Transpose<TranspositionsBase<Derived> > >
+ : traits<Derived>
+{};
 
 } // end namespace internal
 
-/* Template partial specialization for transposed/inverse transpositions */
-
 template<typename TranspositionsDerived>
 class Transpose<TranspositionsBase<TranspositionsDerived> >
 {
@@ -406,27 +372,31 @@ class Transpose<TranspositionsBase<TranspositionsDerived> >
     typedef typename TranspositionType::IndicesType IndicesType;
   public:
 
-    Transpose(const TranspositionType& t) : m_transpositions(t) {}
+    explicit Transpose(const TranspositionType& t) : m_transpositions(t) {}
 
-    inline int size() const { return m_transpositions.size(); }
+    Index size() const { return m_transpositions.size(); }
+    Index rows() const { return m_transpositions.size(); }
+    Index cols() const { return m_transpositions.size(); }
 
     /** \returns the \a matrix with the inverse transpositions applied to the columns.
       */
-    template<typename Derived> friend
-    inline const internal::transposition_matrix_product_retval<TranspositionType, Derived, OnTheRight, true>
-    operator*(const MatrixBase<Derived>& matrix, const Transpose& trt)
+    template<typename OtherDerived> friend
+    const Product<OtherDerived, Transpose, AliasFreeProduct>
+    operator*(const MatrixBase<OtherDerived>& matrix, const Transpose& trt)
     {
-      return internal::transposition_matrix_product_retval<TranspositionType, Derived, OnTheRight, true>(trt.m_transpositions, matrix.derived());
+      return Product<OtherDerived, Transpose, AliasFreeProduct>(matrix.derived(), trt.derived());
     }
 
     /** \returns the \a matrix with the inverse transpositions applied to the rows.
       */
-    template<typename Derived>
-    inline const internal::transposition_matrix_product_retval<TranspositionType, Derived, OnTheLeft, true>
-    operator*(const MatrixBase<Derived>& matrix) const
+    template<typename OtherDerived>
+    const Product<Transpose, OtherDerived, AliasFreeProduct>
+    operator*(const MatrixBase<OtherDerived>& matrix) const
     {
-      return internal::transposition_matrix_product_retval<TranspositionType, Derived, OnTheLeft, true>(m_transpositions, matrix.derived());
+      return Product<Transpose, OtherDerived, AliasFreeProduct>(*this, matrix.derived());
     }
+    
+    const TranspositionType& nestedExpression() const { return m_transpositions; }
 
   protected:
     const TranspositionType& m_transpositions;
diff --git a/eigen/Eigen/src/Core/TriangularMatrix.h b/eigen/Eigen/src/Core/TriangularMatrix.h
index 4d65392..ed80da3 100644
--- a/eigen/Eigen/src/Core/TriangularMatrix.h
+++ b/eigen/Eigen/src/Core/TriangularMatrix.h
@@ -19,9 +19,7 @@ template<int Side, typename TriangularType, typename Rhs> struct triangular_solv
   
 }
 
-/** \internal
-  *
-  * \class TriangularBase
+/** \class TriangularBase
   * \ingroup Core_Module
   *
   * \brief Base class for triangular part in a matrix
@@ -32,41 +30,69 @@ template<typename Derived> class TriangularBase : public EigenBase<Derived>
 
     enum {
       Mode = internal::traits<Derived>::Mode,
-      CoeffReadCost = internal::traits<Derived>::CoeffReadCost,
       RowsAtCompileTime = internal::traits<Derived>::RowsAtCompileTime,
       ColsAtCompileTime = internal::traits<Derived>::ColsAtCompileTime,
       MaxRowsAtCompileTime = internal::traits<Derived>::MaxRowsAtCompileTime,
-      MaxColsAtCompileTime = internal::traits<Derived>::MaxColsAtCompileTime
+      MaxColsAtCompileTime = internal::traits<Derived>::MaxColsAtCompileTime,
+      
+      SizeAtCompileTime = (internal::size_at_compile_time<internal::traits<Derived>::RowsAtCompileTime,
+                                                   internal::traits<Derived>::ColsAtCompileTime>::ret),
+      /**< This is equal to the number of coefficients, i.e. the number of
+          * rows times the number of columns, or to \a Dynamic if this is not
+          * known at compile-time. \sa RowsAtCompileTime, ColsAtCompileTime */
+      
+      MaxSizeAtCompileTime = (internal::size_at_compile_time<internal::traits<Derived>::MaxRowsAtCompileTime,
+                                                   internal::traits<Derived>::MaxColsAtCompileTime>::ret)
+        
     };
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
-    typedef typename internal::traits<Derived>::Index Index;
-    typedef typename internal::traits<Derived>::DenseMatrixType DenseMatrixType;
+    typedef typename internal::traits<Derived>::StorageIndex StorageIndex;
+    typedef typename internal::traits<Derived>::FullMatrixType DenseMatrixType;
     typedef DenseMatrixType DenseType;
+    typedef Derived const& Nested;
 
+    EIGEN_DEVICE_FUNC
     inline TriangularBase() { eigen_assert(!((Mode&UnitDiag) && (Mode&ZeroDiag))); }
 
+    EIGEN_DEVICE_FUNC
     inline Index rows() const { return derived().rows(); }
+    EIGEN_DEVICE_FUNC
     inline Index cols() const { return derived().cols(); }
+    EIGEN_DEVICE_FUNC
     inline Index outerStride() const { return derived().outerStride(); }
+    EIGEN_DEVICE_FUNC
     inline Index innerStride() const { return derived().innerStride(); }
+    
+    // dummy resize function
+    void resize(Index rows, Index cols)
+    {
+      EIGEN_UNUSED_VARIABLE(rows);
+      EIGEN_UNUSED_VARIABLE(cols);
+      eigen_assert(rows==this->rows() && cols==this->cols());
+    }
 
+    EIGEN_DEVICE_FUNC
     inline Scalar coeff(Index row, Index col) const  { return derived().coeff(row,col); }
+    EIGEN_DEVICE_FUNC
     inline Scalar& coeffRef(Index row, Index col) { return derived().coeffRef(row,col); }
 
     /** \see MatrixBase::copyCoeff(row,col)
       */
     template<typename Other>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE void copyCoeff(Index row, Index col, Other& other)
     {
       derived().coeffRef(row, col) = other.coeff(row, col);
     }
 
+    EIGEN_DEVICE_FUNC
     inline Scalar operator()(Index row, Index col) const
     {
       check_coordinates(row, col);
       return coeff(row,col);
     }
+    EIGEN_DEVICE_FUNC
     inline Scalar& operator()(Index row, Index col)
     {
       check_coordinates(row, col);
@@ -74,15 +100,20 @@ template<typename Derived> class TriangularBase : public EigenBase<Derived>
     }
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
+    EIGEN_DEVICE_FUNC
     inline const Derived& derived() const { return *static_cast<const Derived*>(this); }
+    EIGEN_DEVICE_FUNC
     inline Derived& derived() { return *static_cast<Derived*>(this); }
     #endif // not EIGEN_PARSED_BY_DOXYGEN
 
     template<typename DenseDerived>
+    EIGEN_DEVICE_FUNC
     void evalTo(MatrixBase<DenseDerived> &other) const;
     template<typename DenseDerived>
+    EIGEN_DEVICE_FUNC
     void evalToLazy(MatrixBase<DenseDerived> &other) const;
 
+    EIGEN_DEVICE_FUNC
     DenseMatrixType toDenseMatrix() const
     {
       DenseMatrixType res(rows(), cols());
@@ -119,17 +150,17 @@ template<typename Derived> class TriangularBase : public EigenBase<Derived>
 /** \class TriangularView
   * \ingroup Core_Module
   *
-  * \brief Base class for triangular part in a matrix
+  * \brief Expression of a triangular part in a matrix
   *
   * \param MatrixType the type of the object in which we are taking the triangular part
   * \param Mode the kind of triangular matrix expression to construct. Can be #Upper,
   *             #Lower, #UnitUpper, #UnitLower, #StrictlyUpper, or #StrictlyLower.
   *             This is in fact a bit field; it must have either #Upper or #Lower, 
-  *             and additionnaly it may have #UnitDiag or #ZeroDiag or neither.
+  *             and additionally it may have #UnitDiag or #ZeroDiag or neither.
   *
   * This class represents a triangular part of a matrix, not necessarily square. Strictly speaking, for rectangular
   * matrices one should speak of "trapezoid" parts. This class is the return type
-  * of MatrixBase::triangularView() and most of the time this is the only way it is used.
+  * of MatrixBase::triangularView() and SparseMatrixBase::triangularView(), and most of the time this is the only way it is used.
   *
   * \sa MatrixBase::triangularView()
   */
@@ -137,538 +168,426 @@ namespace internal {
 template<typename MatrixType, unsigned int _Mode>
 struct traits<TriangularView<MatrixType, _Mode> > : traits<MatrixType>
 {
-  typedef typename nested<MatrixType>::type MatrixTypeNested;
+  typedef typename ref_selector<MatrixType>::non_const_type MatrixTypeNested;
   typedef typename remove_reference<MatrixTypeNested>::type MatrixTypeNestedNonRef;
   typedef typename remove_all<MatrixTypeNested>::type MatrixTypeNestedCleaned;
+  typedef typename MatrixType::PlainObject FullMatrixType;
   typedef MatrixType ExpressionType;
-  typedef typename MatrixType::PlainObject DenseMatrixType;
   enum {
     Mode = _Mode,
-    Flags = (MatrixTypeNestedCleaned::Flags & (HereditaryBits) & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit))) | Mode,
-    CoeffReadCost = MatrixTypeNestedCleaned::CoeffReadCost
+    FlagsLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0,
+    Flags = (MatrixTypeNestedCleaned::Flags & (HereditaryBits | FlagsLvalueBit) & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit)))
   };
 };
 }
 
-template<int Mode, bool LhsIsTriangular,
-         typename Lhs, bool LhsIsVector,
-         typename Rhs, bool RhsIsVector>
-struct TriangularProduct;
+template<typename _MatrixType, unsigned int _Mode, typename StorageKind> class TriangularViewImpl;
 
 template<typename _MatrixType, unsigned int _Mode> class TriangularView
-  : public TriangularBase<TriangularView<_MatrixType, _Mode> >
+  : public TriangularViewImpl<_MatrixType, _Mode, typename internal::traits<_MatrixType>::StorageKind >
 {
   public:
 
-    typedef TriangularBase<TriangularView> Base;
+    typedef TriangularViewImpl<_MatrixType, _Mode, typename internal::traits<_MatrixType>::StorageKind > Base;
     typedef typename internal::traits<TriangularView>::Scalar Scalar;
-
     typedef _MatrixType MatrixType;
-    typedef typename internal::traits<TriangularView>::DenseMatrixType DenseMatrixType;
-    typedef DenseMatrixType PlainObject;
 
   protected:
     typedef typename internal::traits<TriangularView>::MatrixTypeNested MatrixTypeNested;
     typedef typename internal::traits<TriangularView>::MatrixTypeNestedNonRef MatrixTypeNestedNonRef;
-    typedef typename internal::traits<TriangularView>::MatrixTypeNestedCleaned MatrixTypeNestedCleaned;
 
     typedef typename internal::remove_all<typename MatrixType::ConjugateReturnType>::type MatrixConjugateReturnType;
     
   public:
-    using Base::evalToLazy;
-  
 
     typedef typename internal::traits<TriangularView>::StorageKind StorageKind;
-    typedef typename internal::traits<TriangularView>::Index Index;
+    typedef typename internal::traits<TriangularView>::MatrixTypeNestedCleaned NestedExpression;
 
     enum {
       Mode = _Mode,
+      Flags = internal::traits<TriangularView>::Flags,
       TransposeMode = (Mode & Upper ? Lower : 0)
                     | (Mode & Lower ? Upper : 0)
                     | (Mode & (UnitDiag))
-                    | (Mode & (ZeroDiag))
+                    | (Mode & (ZeroDiag)),
+      IsVectorAtCompileTime = false
     };
 
-    inline TriangularView(const MatrixType& matrix) : m_matrix(matrix)
+    EIGEN_DEVICE_FUNC
+    explicit inline TriangularView(MatrixType& matrix) : m_matrix(matrix)
     {}
+    
+    using Base::operator=;
+    TriangularView& operator=(const TriangularView &other)
+    { return Base::operator=(other); }
 
+    /** \copydoc EigenBase::rows() */
+    EIGEN_DEVICE_FUNC
     inline Index rows() const { return m_matrix.rows(); }
+    /** \copydoc EigenBase::cols() */
+    EIGEN_DEVICE_FUNC
     inline Index cols() const { return m_matrix.cols(); }
-    inline Index outerStride() const { return m_matrix.outerStride(); }
-    inline Index innerStride() const { return m_matrix.innerStride(); }
+
+    /** \returns a const reference to the nested expression */
+    EIGEN_DEVICE_FUNC
+    const NestedExpression& nestedExpression() const { return m_matrix; }
+
+    /** \returns a reference to the nested expression */
+    EIGEN_DEVICE_FUNC
+    NestedExpression& nestedExpression() { return m_matrix; }
+    
+    typedef TriangularView<const MatrixConjugateReturnType,Mode> ConjugateReturnType;
+    /** \sa MatrixBase::conjugate() const */
+    EIGEN_DEVICE_FUNC
+    inline const ConjugateReturnType conjugate() const
+    { return ConjugateReturnType(m_matrix.conjugate()); }
+
+    typedef TriangularView<const typename MatrixType::AdjointReturnType,TransposeMode> AdjointReturnType;
+    /** \sa MatrixBase::adjoint() const */
+    EIGEN_DEVICE_FUNC
+    inline const AdjointReturnType adjoint() const
+    { return AdjointReturnType(m_matrix.adjoint()); }
+
+    typedef TriangularView<typename MatrixType::TransposeReturnType,TransposeMode> TransposeReturnType;
+     /** \sa MatrixBase::transpose() */
+    EIGEN_DEVICE_FUNC
+    inline TransposeReturnType transpose()
+    {
+      EIGEN_STATIC_ASSERT_LVALUE(MatrixType)
+      typename MatrixType::TransposeReturnType tmp(m_matrix);
+      return TransposeReturnType(tmp);
+    }
+    
+    typedef TriangularView<const typename MatrixType::ConstTransposeReturnType,TransposeMode> ConstTransposeReturnType;
+    /** \sa MatrixBase::transpose() const */
+    EIGEN_DEVICE_FUNC
+    inline const ConstTransposeReturnType transpose() const
+    {
+      return ConstTransposeReturnType(m_matrix.transpose());
+    }
+
+    template<typename Other>
+    EIGEN_DEVICE_FUNC
+    inline const Solve<TriangularView, Other> 
+    solve(const MatrixBase<Other>& other) const
+    { return Solve<TriangularView, Other>(*this, other.derived()); }
+    
+  // workaround MSVC ICE
+  #if EIGEN_COMP_MSVC
+    template<int Side, typename Other>
+    EIGEN_DEVICE_FUNC
+    inline const internal::triangular_solve_retval<Side,TriangularView, Other>
+    solve(const MatrixBase<Other>& other) const
+    { return Base::template solve<Side>(other); }
+  #else
+    using Base::solve;
+  #endif
+
+    /** \returns a selfadjoint view of the referenced triangular part which must be either \c #Upper or \c #Lower.
+      *
+      * This is a shortcut for \code this->nestedExpression().selfadjointView<(*this)::Mode>() \endcode
+      * \sa MatrixBase::selfadjointView() */
+    EIGEN_DEVICE_FUNC
+    SelfAdjointView<MatrixTypeNestedNonRef,Mode> selfadjointView()
+    {
+      EIGEN_STATIC_ASSERT((Mode&(UnitDiag|ZeroDiag))==0,PROGRAMMING_ERROR);
+      return SelfAdjointView<MatrixTypeNestedNonRef,Mode>(m_matrix);
+    }
+
+    /** This is the const version of selfadjointView() */
+    EIGEN_DEVICE_FUNC
+    const SelfAdjointView<MatrixTypeNestedNonRef,Mode> selfadjointView() const
+    {
+      EIGEN_STATIC_ASSERT((Mode&(UnitDiag|ZeroDiag))==0,PROGRAMMING_ERROR);
+      return SelfAdjointView<MatrixTypeNestedNonRef,Mode>(m_matrix);
+    }
+
+
+    /** \returns the determinant of the triangular matrix
+      * \sa MatrixBase::determinant() */
+    EIGEN_DEVICE_FUNC
+    Scalar determinant() const
+    {
+      if (Mode & UnitDiag)
+        return 1;
+      else if (Mode & ZeroDiag)
+        return 0;
+      else
+        return m_matrix.diagonal().prod();
+    }
+      
+  protected:
+
+    MatrixTypeNested m_matrix;
+};
+
+/** \ingroup Core_Module
+  *
+  * \brief Base class for a triangular part in a \b dense matrix
+  *
+  * This class is an abstract base class of class TriangularView, and objects of type TriangularViewImpl cannot be instantiated.
+  * It extends class TriangularView with additional methods which available for dense expressions only.
+  *
+  * \sa class TriangularView, MatrixBase::triangularView()
+  */
+template<typename _MatrixType, unsigned int _Mode> class TriangularViewImpl<_MatrixType,_Mode,Dense>
+  : public TriangularBase<TriangularView<_MatrixType, _Mode> >
+{
+  public:
+
+    typedef TriangularView<_MatrixType, _Mode> TriangularViewType;
+    typedef TriangularBase<TriangularViewType> Base;
+    typedef typename internal::traits<TriangularViewType>::Scalar Scalar;
+
+    typedef _MatrixType MatrixType;
+    typedef typename MatrixType::PlainObject DenseMatrixType;
+    typedef DenseMatrixType PlainObject;
+
+  public:
+    using Base::evalToLazy;
+    using Base::derived;
+
+    typedef typename internal::traits<TriangularViewType>::StorageKind StorageKind;
+
+    enum {
+      Mode = _Mode,
+      Flags = internal::traits<TriangularViewType>::Flags
+    };
+
+    /** \returns the outer-stride of the underlying dense matrix
+      * \sa DenseCoeffsBase::outerStride() */
+    EIGEN_DEVICE_FUNC
+    inline Index outerStride() const { return derived().nestedExpression().outerStride(); }
+    /** \returns the inner-stride of the underlying dense matrix
+      * \sa DenseCoeffsBase::innerStride() */
+    EIGEN_DEVICE_FUNC
+    inline Index innerStride() const { return derived().nestedExpression().innerStride(); }
 
     /** \sa MatrixBase::operator+=() */
-    template<typename Other> TriangularView&  operator+=(const DenseBase<Other>& other) { return *this = m_matrix + other.derived(); }
+    template<typename Other>
+    EIGEN_DEVICE_FUNC
+    TriangularViewType&  operator+=(const DenseBase<Other>& other) {
+      internal::call_assignment_no_alias(derived(), other.derived(), internal::add_assign_op<Scalar,typename Other::Scalar>());
+      return derived();
+    }
     /** \sa MatrixBase::operator-=() */
-    template<typename Other> TriangularView&  operator-=(const DenseBase<Other>& other) { return *this = m_matrix - other.derived(); }
+    template<typename Other>
+    EIGEN_DEVICE_FUNC
+    TriangularViewType&  operator-=(const DenseBase<Other>& other) {
+      internal::call_assignment_no_alias(derived(), other.derived(), internal::sub_assign_op<Scalar,typename Other::Scalar>());
+      return derived();
+    }
+    
     /** \sa MatrixBase::operator*=() */
-    TriangularView&  operator*=(const typename internal::traits<MatrixType>::Scalar& other) { return *this = m_matrix * other; }
-    /** \sa MatrixBase::operator/=() */
-    TriangularView&  operator/=(const typename internal::traits<MatrixType>::Scalar& other) { return *this = m_matrix / other; }
+    EIGEN_DEVICE_FUNC
+    TriangularViewType&  operator*=(const typename internal::traits<MatrixType>::Scalar& other) { return *this = derived().nestedExpression() * other; }
+    /** \sa DenseBase::operator/=() */
+    EIGEN_DEVICE_FUNC
+    TriangularViewType&  operator/=(const typename internal::traits<MatrixType>::Scalar& other) { return *this = derived().nestedExpression() / other; }
 
     /** \sa MatrixBase::fill() */
+    EIGEN_DEVICE_FUNC
     void fill(const Scalar& value) { setConstant(value); }
     /** \sa MatrixBase::setConstant() */
-    TriangularView& setConstant(const Scalar& value)
-    { return *this = MatrixType::Constant(rows(), cols(), value); }
+    EIGEN_DEVICE_FUNC
+    TriangularViewType& setConstant(const Scalar& value)
+    { return *this = MatrixType::Constant(derived().rows(), derived().cols(), value); }
     /** \sa MatrixBase::setZero() */
-    TriangularView& setZero() { return setConstant(Scalar(0)); }
+    EIGEN_DEVICE_FUNC
+    TriangularViewType& setZero() { return setConstant(Scalar(0)); }
     /** \sa MatrixBase::setOnes() */
-    TriangularView& setOnes() { return setConstant(Scalar(1)); }
+    EIGEN_DEVICE_FUNC
+    TriangularViewType& setOnes() { return setConstant(Scalar(1)); }
 
     /** \sa MatrixBase::coeff()
       * \warning the coordinates must fit into the referenced triangular part
       */
+    EIGEN_DEVICE_FUNC
     inline Scalar coeff(Index row, Index col) const
     {
       Base::check_coordinates_internal(row, col);
-      return m_matrix.coeff(row, col);
+      return derived().nestedExpression().coeff(row, col);
     }
 
     /** \sa MatrixBase::coeffRef()
       * \warning the coordinates must fit into the referenced triangular part
       */
+    EIGEN_DEVICE_FUNC
     inline Scalar& coeffRef(Index row, Index col)
     {
+      EIGEN_STATIC_ASSERT_LVALUE(TriangularViewType);
       Base::check_coordinates_internal(row, col);
-      return m_matrix.const_cast_derived().coeffRef(row, col);
+      return derived().nestedExpression().coeffRef(row, col);
     }
 
-    const MatrixTypeNestedCleaned& nestedExpression() const { return m_matrix; }
-    MatrixTypeNestedCleaned& nestedExpression() { return *const_cast<MatrixTypeNestedCleaned*>(&m_matrix); }
-
     /** Assigns a triangular matrix to a triangular part of a dense matrix */
     template<typename OtherDerived>
-    TriangularView& operator=(const TriangularBase<OtherDerived>& other);
+    EIGEN_DEVICE_FUNC
+    TriangularViewType& operator=(const TriangularBase<OtherDerived>& other);
 
+    /** Shortcut for\code *this = other.other.triangularView<(*this)::Mode>() \endcode */
     template<typename OtherDerived>
-    TriangularView& operator=(const MatrixBase<OtherDerived>& other);
+    EIGEN_DEVICE_FUNC
+    TriangularViewType& operator=(const MatrixBase<OtherDerived>& other);
 
-    TriangularView& operator=(const TriangularView& other)
-    { return *this = other.nestedExpression(); }
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+    EIGEN_DEVICE_FUNC
+    TriangularViewType& operator=(const TriangularViewImpl& other)
+    { return *this = other.derived().nestedExpression(); }
 
+    /** \deprecated */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     void lazyAssign(const TriangularBase<OtherDerived>& other);
 
+    /** \deprecated */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     void lazyAssign(const MatrixBase<OtherDerived>& other);
-
-    /** \sa MatrixBase::conjugate() */
-    inline TriangularView<MatrixConjugateReturnType,Mode> conjugate()
-    { return m_matrix.conjugate(); }
-    /** \sa MatrixBase::conjugate() const */
-    inline const TriangularView<MatrixConjugateReturnType,Mode> conjugate() const
-    { return m_matrix.conjugate(); }
-
-    /** \sa MatrixBase::adjoint() const */
-    inline const TriangularView<const typename MatrixType::AdjointReturnType,TransposeMode> adjoint() const
-    { return m_matrix.adjoint(); }
-
-    /** \sa MatrixBase::transpose() */
-    inline TriangularView<Transpose<MatrixType>,TransposeMode> transpose()
-    {
-      EIGEN_STATIC_ASSERT_LVALUE(MatrixType)
-      return m_matrix.const_cast_derived().transpose();
-    }
-    /** \sa MatrixBase::transpose() const */
-    inline const TriangularView<Transpose<MatrixType>,TransposeMode> transpose() const
-    {
-      return m_matrix.transpose();
-    }
+#endif
 
     /** Efficient triangular matrix times vector/matrix product */
     template<typename OtherDerived>
-    TriangularProduct<Mode, true, MatrixType, false, OtherDerived, OtherDerived::ColsAtCompileTime==1>
+    EIGEN_DEVICE_FUNC
+    const Product<TriangularViewType,OtherDerived>
     operator*(const MatrixBase<OtherDerived>& rhs) const
     {
-      return TriangularProduct
-              <Mode, true, MatrixType, false, OtherDerived, OtherDerived::ColsAtCompileTime==1>
-              (m_matrix, rhs.derived());
+      return Product<TriangularViewType,OtherDerived>(derived(), rhs.derived());
     }
 
     /** Efficient vector/matrix times triangular matrix product */
     template<typename OtherDerived> friend
-    TriangularProduct<Mode, false, OtherDerived, OtherDerived::RowsAtCompileTime==1, MatrixType, false>
-    operator*(const MatrixBase<OtherDerived>& lhs, const TriangularView& rhs)
-    {
-      return TriangularProduct
-              <Mode, false, OtherDerived, OtherDerived::RowsAtCompileTime==1, MatrixType, false>
-              (lhs.derived(),rhs.m_matrix);
-    }
-
-    #ifdef EIGEN2_SUPPORT
-    template<typename OtherDerived>
-    struct eigen2_product_return_type
-    {
-      typedef typename TriangularView<MatrixType,Mode>::DenseMatrixType DenseMatrixType;
-      typedef typename OtherDerived::PlainObject::DenseType OtherPlainObject;
-      typedef typename ProductReturnType<DenseMatrixType, OtherPlainObject>::Type ProdRetType;
-      typedef typename ProdRetType::PlainObject type;
-    };
-    template<typename OtherDerived>
-    const typename eigen2_product_return_type<OtherDerived>::type
-    operator*(const EigenBase<OtherDerived>& rhs) const
-    {
-      typename OtherDerived::PlainObject::DenseType rhsPlainObject;
-      rhs.evalTo(rhsPlainObject);
-      return this->toDenseMatrix() * rhsPlainObject;
-    }
-    template<typename OtherMatrixType>
-    bool isApprox(const TriangularView<OtherMatrixType, Mode>& other, typename NumTraits<Scalar>::Real precision = NumTraits<Scalar>::dummy_precision()) const
-    {
-      return this->toDenseMatrix().isApprox(other.toDenseMatrix(), precision);
-    }
-    template<typename OtherDerived>
-    bool isApprox(const MatrixBase<OtherDerived>& other, typename NumTraits<Scalar>::Real precision = NumTraits<Scalar>::dummy_precision()) const
-    {
-      return this->toDenseMatrix().isApprox(other, precision);
-    }
-    #endif // EIGEN2_SUPPORT
-
+    EIGEN_DEVICE_FUNC
+    const Product<OtherDerived,TriangularViewType>
+    operator*(const MatrixBase<OtherDerived>& lhs, const TriangularViewImpl& rhs)
+    {
+      return Product<OtherDerived,TriangularViewType>(lhs.derived(),rhs.derived());
+    }
+
+    /** \returns the product of the inverse of \c *this with \a other, \a *this being triangular.
+      *
+      * This function computes the inverse-matrix matrix product inverse(\c *this) * \a other if
+      * \a Side==OnTheLeft (the default), or the right-inverse-multiply  \a other * inverse(\c *this) if
+      * \a Side==OnTheRight.
+      *
+      * Note that the template parameter \c Side can be ommitted, in which case \c Side==OnTheLeft
+      *
+      * The matrix \c *this must be triangular and invertible (i.e., all the coefficients of the
+      * diagonal must be non zero). It works as a forward (resp. backward) substitution if \c *this
+      * is an upper (resp. lower) triangular matrix.
+      *
+      * Example: \include Triangular_solve.cpp
+      * Output: \verbinclude Triangular_solve.out
+      *
+      * This function returns an expression of the inverse-multiply and can works in-place if it is assigned
+      * to the same matrix or vector \a other.
+      *
+      * For users coming from BLAS, this function (and more specifically solveInPlace()) offer
+      * all the operations supported by the \c *TRSV and \c *TRSM BLAS routines.
+      *
+      * \sa TriangularView::solveInPlace()
+      */
     template<int Side, typename Other>
-    inline const internal::triangular_solve_retval<Side,TriangularView, Other>
+    inline const internal::triangular_solve_retval<Side,TriangularViewType, Other>
     solve(const MatrixBase<Other>& other) const;
 
+    /** "in-place" version of TriangularView::solve() where the result is written in \a other
+      *
+      * \warning The parameter is only marked 'const' to make the C++ compiler accept a temporary expression here.
+      * This function will const_cast it, so constness isn't honored here.
+      *
+      * Note that the template parameter \c Side can be ommitted, in which case \c Side==OnTheLeft
+      *
+      * See TriangularView:solve() for the details.
+      */
     template<int Side, typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     void solveInPlace(const MatrixBase<OtherDerived>& other) const;
 
-    template<typename Other>
-    inline const internal::triangular_solve_retval<OnTheLeft,TriangularView, Other> 
-    solve(const MatrixBase<Other>& other) const
-    { return solve<OnTheLeft>(other); }
-
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     void solveInPlace(const MatrixBase<OtherDerived>& other) const
     { return solveInPlace<OnTheLeft>(other); }
 
-    const SelfAdjointView<MatrixTypeNestedNonRef,Mode> selfadjointView() const
-    {
-      EIGEN_STATIC_ASSERT((Mode&UnitDiag)==0,PROGRAMMING_ERROR);
-      return SelfAdjointView<MatrixTypeNestedNonRef,Mode>(m_matrix);
-    }
-    SelfAdjointView<MatrixTypeNestedNonRef,Mode> selfadjointView()
-    {
-      EIGEN_STATIC_ASSERT((Mode&UnitDiag)==0,PROGRAMMING_ERROR);
-      return SelfAdjointView<MatrixTypeNestedNonRef,Mode>(m_matrix);
-    }
-
+    /** Swaps the coefficients of the common triangular parts of two matrices */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+#ifdef EIGEN_PARSED_BY_DOXYGEN
+    void swap(TriangularBase<OtherDerived> &other)
+#else
     void swap(TriangularBase<OtherDerived> const & other)
+#endif
     {
-      TriangularView<SwapWrapper<MatrixType>,Mode>(const_cast<MatrixType&>(m_matrix)).lazyAssign(other.derived());
+      EIGEN_STATIC_ASSERT_LVALUE(OtherDerived);
+      call_assignment(derived(), other.const_cast_derived(), internal::swap_assign_op<Scalar>());
     }
 
+    /** \deprecated
+      * Shortcut for \code (*this).swap(other.triangularView<(*this)::Mode>()) \endcode */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     void swap(MatrixBase<OtherDerived> const & other)
     {
-      SwapWrapper<MatrixType> swaper(const_cast<MatrixType&>(m_matrix));
-      TriangularView<SwapWrapper<MatrixType>,Mode>(swaper).lazyAssign(other.derived());
+      EIGEN_STATIC_ASSERT_LVALUE(OtherDerived);
+      call_assignment(derived(), other.const_cast_derived(), internal::swap_assign_op<Scalar>());
     }
 
-    Scalar determinant() const
-    {
-      if (Mode & UnitDiag)
-        return 1;
-      else if (Mode & ZeroDiag)
-        return 0;
-      else
-        return m_matrix.diagonal().prod();
-    }
-    
-    // TODO simplify the following:
-    template<typename ProductDerived, typename Lhs, typename Rhs>
-    EIGEN_STRONG_INLINE TriangularView& operator=(const ProductBase<ProductDerived, Lhs,Rhs>& other)
-    {
-      setZero();
-      return assignProduct(other.derived(),1);
-    }
-    
-    template<typename ProductDerived, typename Lhs, typename Rhs>
-    EIGEN_STRONG_INLINE TriangularView& operator+=(const ProductBase<ProductDerived, Lhs,Rhs>& other)
-    {
-      return assignProduct(other.derived(),1);
-    }
-    
-    template<typename ProductDerived, typename Lhs, typename Rhs>
-    EIGEN_STRONG_INLINE TriangularView& operator-=(const ProductBase<ProductDerived, Lhs,Rhs>& other)
-    {
-      return assignProduct(other.derived(),-1);
-    }
-    
-    
-    template<typename ProductDerived>
-    EIGEN_STRONG_INLINE TriangularView& operator=(const ScaledProduct<ProductDerived>& other)
-    {
-      setZero();
-      return assignProduct(other.derived(),other.alpha());
-    }
-    
-    template<typename ProductDerived>
-    EIGEN_STRONG_INLINE TriangularView& operator+=(const ScaledProduct<ProductDerived>& other)
-    {
-      return assignProduct(other.derived(),other.alpha());
-    }
-    
-    template<typename ProductDerived>
-    EIGEN_STRONG_INLINE TriangularView& operator-=(const ScaledProduct<ProductDerived>& other)
-    {
-      return assignProduct(other.derived(),-other.alpha());
-    }
-    
-  protected:
-    
-    template<typename ProductDerived, typename Lhs, typename Rhs>
-    EIGEN_STRONG_INLINE TriangularView& assignProduct(const ProductBase<ProductDerived, Lhs,Rhs>& prod, const Scalar& alpha);
-    
-    template<int Mode, bool LhsIsTriangular,
-         typename Lhs, bool LhsIsVector,
-         typename Rhs, bool RhsIsVector>
-    EIGEN_STRONG_INLINE TriangularView& assignProduct(const TriangularProduct<Mode, LhsIsTriangular, Lhs, LhsIsVector, Rhs, RhsIsVector>& prod, const Scalar& alpha)
-    {
-      lazyAssign(alpha*prod.eval());
-      return *this;
+    template<typename RhsType, typename DstType>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _solve_impl(const RhsType &rhs, DstType &dst) const {
+      if(!internal::is_same_dense(dst,rhs))
+        dst = rhs;
+      this->solveInPlace(dst);
     }
 
-    MatrixTypeNested m_matrix;
+    template<typename ProductType>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE TriangularViewType& _assignProduct(const ProductType& prod, const Scalar& alpha, bool beta);
 };
 
 /***************************************************************************
 * Implementation of triangular evaluation/assignment
 ***************************************************************************/
 
-namespace internal {
-
-template<typename Derived1, typename Derived2, unsigned int Mode, int UnrollCount, bool ClearOpposite>
-struct triangular_assignment_selector
-{
-  enum {
-    col = (UnrollCount-1) / Derived1::RowsAtCompileTime,
-    row = (UnrollCount-1) % Derived1::RowsAtCompileTime
-  };
-  
-  typedef typename Derived1::Scalar Scalar;
-
-  static inline void run(Derived1 &dst, const Derived2 &src)
-  {
-    triangular_assignment_selector<Derived1, Derived2, Mode, UnrollCount-1, ClearOpposite>::run(dst, src);
-
-    eigen_assert( Mode == Upper || Mode == Lower
-            || Mode == StrictlyUpper || Mode == StrictlyLower
-            || Mode == UnitUpper || Mode == UnitLower);
-    if((Mode == Upper && row <= col)
-    || (Mode == Lower && row >= col)
-    || (Mode == StrictlyUpper && row < col)
-    || (Mode == StrictlyLower && row > col)
-    || (Mode == UnitUpper && row < col)
-    || (Mode == UnitLower && row > col))
-      dst.copyCoeff(row, col, src);
-    else if(ClearOpposite)
-    {
-      if (Mode&UnitDiag && row==col)
-        dst.coeffRef(row, col) = Scalar(1);
-      else
-        dst.coeffRef(row, col) = Scalar(0);
-    }
-  }
-};
-
-// prevent buggy user code from causing an infinite recursion
-template<typename Derived1, typename Derived2, unsigned int Mode, bool ClearOpposite>
-struct triangular_assignment_selector<Derived1, Derived2, Mode, 0, ClearOpposite>
-{
-  static inline void run(Derived1 &, const Derived2 &) {}
-};
-
-template<typename Derived1, typename Derived2, bool ClearOpposite>
-struct triangular_assignment_selector<Derived1, Derived2, Upper, Dynamic, ClearOpposite>
-{
-  typedef typename Derived1::Index Index;
-  typedef typename Derived1::Scalar Scalar;
-  static inline void run(Derived1 &dst, const Derived2 &src)
-  {
-    for(Index j = 0; j < dst.cols(); ++j)
-    {
-      Index maxi = (std::min)(j, dst.rows()-1);
-      for(Index i = 0; i <= maxi; ++i)
-        dst.copyCoeff(i, j, src);
-      if (ClearOpposite)
-        for(Index i = maxi+1; i < dst.rows(); ++i)
-          dst.coeffRef(i, j) = Scalar(0);
-    }
-  }
-};
-
-template<typename Derived1, typename Derived2, bool ClearOpposite>
-struct triangular_assignment_selector<Derived1, Derived2, Lower, Dynamic, ClearOpposite>
-{
-  typedef typename Derived1::Index Index;
-  static inline void run(Derived1 &dst, const Derived2 &src)
-  {
-    for(Index j = 0; j < dst.cols(); ++j)
-    {
-      for(Index i = j; i < dst.rows(); ++i)
-        dst.copyCoeff(i, j, src);
-      Index maxi = (std::min)(j, dst.rows());
-      if (ClearOpposite)
-        for(Index i = 0; i < maxi; ++i)
-          dst.coeffRef(i, j) = static_cast<typename Derived1::Scalar>(0);
-    }
-  }
-};
-
-template<typename Derived1, typename Derived2, bool ClearOpposite>
-struct triangular_assignment_selector<Derived1, Derived2, StrictlyUpper, Dynamic, ClearOpposite>
-{
-  typedef typename Derived1::Index Index;
-  typedef typename Derived1::Scalar Scalar;
-  static inline void run(Derived1 &dst, const Derived2 &src)
-  {
-    for(Index j = 0; j < dst.cols(); ++j)
-    {
-      Index maxi = (std::min)(j, dst.rows());
-      for(Index i = 0; i < maxi; ++i)
-        dst.copyCoeff(i, j, src);
-      if (ClearOpposite)
-        for(Index i = maxi; i < dst.rows(); ++i)
-          dst.coeffRef(i, j) = Scalar(0);
-    }
-  }
-};
-
-template<typename Derived1, typename Derived2, bool ClearOpposite>
-struct triangular_assignment_selector<Derived1, Derived2, StrictlyLower, Dynamic, ClearOpposite>
-{
-  typedef typename Derived1::Index Index;
-  static inline void run(Derived1 &dst, const Derived2 &src)
-  {
-    for(Index j = 0; j < dst.cols(); ++j)
-    {
-      for(Index i = j+1; i < dst.rows(); ++i)
-        dst.copyCoeff(i, j, src);
-      Index maxi = (std::min)(j, dst.rows()-1);
-      if (ClearOpposite)
-        for(Index i = 0; i <= maxi; ++i)
-          dst.coeffRef(i, j) = static_cast<typename Derived1::Scalar>(0);
-    }
-  }
-};
-
-template<typename Derived1, typename Derived2, bool ClearOpposite>
-struct triangular_assignment_selector<Derived1, Derived2, UnitUpper, Dynamic, ClearOpposite>
-{
-  typedef typename Derived1::Index Index;
-  static inline void run(Derived1 &dst, const Derived2 &src)
-  {
-    for(Index j = 0; j < dst.cols(); ++j)
-    {
-      Index maxi = (std::min)(j, dst.rows());
-      for(Index i = 0; i < maxi; ++i)
-        dst.copyCoeff(i, j, src);
-      if (ClearOpposite)
-      {
-        for(Index i = maxi+1; i < dst.rows(); ++i)
-          dst.coeffRef(i, j) = 0;
-      }
-    }
-    dst.diagonal().setOnes();
-  }
-};
-template<typename Derived1, typename Derived2, bool ClearOpposite>
-struct triangular_assignment_selector<Derived1, Derived2, UnitLower, Dynamic, ClearOpposite>
-{
-  typedef typename Derived1::Index Index;
-  static inline void run(Derived1 &dst, const Derived2 &src)
-  {
-    for(Index j = 0; j < dst.cols(); ++j)
-    {
-      Index maxi = (std::min)(j, dst.rows());
-      for(Index i = maxi+1; i < dst.rows(); ++i)
-        dst.copyCoeff(i, j, src);
-      if (ClearOpposite)
-      {
-        for(Index i = 0; i < maxi; ++i)
-          dst.coeffRef(i, j) = 0;
-      }
-    }
-    dst.diagonal().setOnes();
-  }
-};
-
-} // end namespace internal
-
+#ifndef EIGEN_PARSED_BY_DOXYGEN
 // FIXME should we keep that possibility
 template<typename MatrixType, unsigned int Mode>
 template<typename OtherDerived>
-inline TriangularView<MatrixType, Mode>&
-TriangularView<MatrixType, Mode>::operator=(const MatrixBase<OtherDerived>& other)
+EIGEN_DEVICE_FUNC inline TriangularView<MatrixType, Mode>&
+TriangularViewImpl<MatrixType, Mode, Dense>::operator=(const MatrixBase<OtherDerived>& other)
 {
-  if(OtherDerived::Flags & EvalBeforeAssigningBit)
-  {
-    typename internal::plain_matrix_type<OtherDerived>::type other_evaluated(other.rows(), other.cols());
-    other_evaluated.template triangularView<Mode>().lazyAssign(other.derived());
-    lazyAssign(other_evaluated);
-  }
-  else
-    lazyAssign(other.derived());
-  return *this;
+  internal::call_assignment_no_alias(derived(), other.derived(), internal::assign_op<Scalar,typename OtherDerived::Scalar>());
+  return derived();
 }
 
 // FIXME should we keep that possibility
 template<typename MatrixType, unsigned int Mode>
 template<typename OtherDerived>
-void TriangularView<MatrixType, Mode>::lazyAssign(const MatrixBase<OtherDerived>& other)
+EIGEN_DEVICE_FUNC void TriangularViewImpl<MatrixType, Mode, Dense>::lazyAssign(const MatrixBase<OtherDerived>& other)
 {
-  enum {
-    unroll = MatrixType::SizeAtCompileTime != Dynamic
-          && internal::traits<OtherDerived>::CoeffReadCost != Dynamic
-          && MatrixType::SizeAtCompileTime*internal::traits<OtherDerived>::CoeffReadCost/2 <= EIGEN_UNROLLING_LIMIT
-  };
-  eigen_assert(m_matrix.rows() == other.rows() && m_matrix.cols() == other.cols());
-
-  internal::triangular_assignment_selector
-    <MatrixType, OtherDerived, int(Mode),
-    unroll ? int(MatrixType::SizeAtCompileTime) : Dynamic,
-    false // do not change the opposite triangular part
-    >::run(m_matrix.const_cast_derived(), other.derived());
+  internal::call_assignment_no_alias(derived(), other.template triangularView<Mode>());
 }
 
 
 
 template<typename MatrixType, unsigned int Mode>
 template<typename OtherDerived>
-inline TriangularView<MatrixType, Mode>&
-TriangularView<MatrixType, Mode>::operator=(const TriangularBase<OtherDerived>& other)
+EIGEN_DEVICE_FUNC inline TriangularView<MatrixType, Mode>&
+TriangularViewImpl<MatrixType, Mode, Dense>::operator=(const TriangularBase<OtherDerived>& other)
 {
   eigen_assert(Mode == int(OtherDerived::Mode));
-  if(internal::traits<OtherDerived>::Flags & EvalBeforeAssigningBit)
-  {
-    typename OtherDerived::DenseMatrixType other_evaluated(other.rows(), other.cols());
-    other_evaluated.template triangularView<Mode>().lazyAssign(other.derived().nestedExpression());
-    lazyAssign(other_evaluated);
-  }
-  else
-    lazyAssign(other.derived().nestedExpression());
-  return *this;
+  internal::call_assignment(derived(), other.derived());
+  return derived();
 }
 
 template<typename MatrixType, unsigned int Mode>
 template<typename OtherDerived>
-void TriangularView<MatrixType, Mode>::lazyAssign(const TriangularBase<OtherDerived>& other)
+EIGEN_DEVICE_FUNC void TriangularViewImpl<MatrixType, Mode, Dense>::lazyAssign(const TriangularBase<OtherDerived>& other)
 {
-  enum {
-    unroll = MatrixType::SizeAtCompileTime != Dynamic
-                   && internal::traits<OtherDerived>::CoeffReadCost != Dynamic
-                   && MatrixType::SizeAtCompileTime * internal::traits<OtherDerived>::CoeffReadCost / 2
-                        <= EIGEN_UNROLLING_LIMIT
-  };
-  eigen_assert(m_matrix.rows() == other.rows() && m_matrix.cols() == other.cols());
-
-  internal::triangular_assignment_selector
-    <MatrixType, OtherDerived, int(Mode),
-    unroll ? int(MatrixType::SizeAtCompileTime) : Dynamic,
-    false // preserve the opposite triangular part
-    >::run(m_matrix.const_cast_derived(), other.derived().nestedExpression());
+  eigen_assert(Mode == int(OtherDerived::Mode));
+  internal::call_assignment_no_alias(derived(), other.derived());
 }
+#endif
 
 /***************************************************************************
 * Implementation of TriangularBase methods
@@ -678,37 +597,9 @@ void TriangularView<MatrixType, Mode>::lazyAssign(const TriangularBase<OtherDeri
   * If the matrix is triangular, the opposite part is set to zero. */
 template<typename Derived>
 template<typename DenseDerived>
-void TriangularBase<Derived>::evalTo(MatrixBase<DenseDerived> &other) const
+EIGEN_DEVICE_FUNC void TriangularBase<Derived>::evalTo(MatrixBase<DenseDerived> &other) const
 {
-  if(internal::traits<Derived>::Flags & EvalBeforeAssigningBit)
-  {
-    typename internal::plain_matrix_type<Derived>::type other_evaluated(rows(), cols());
-    evalToLazy(other_evaluated);
-    other.derived().swap(other_evaluated);
-  }
-  else
-    evalToLazy(other.derived());
-}
-
-/** Assigns a triangular or selfadjoint matrix to a dense matrix.
-  * If the matrix is triangular, the opposite part is set to zero. */
-template<typename Derived>
-template<typename DenseDerived>
-void TriangularBase<Derived>::evalToLazy(MatrixBase<DenseDerived> &other) const
-{
-  enum {
-    unroll = DenseDerived::SizeAtCompileTime != Dynamic
-                   && internal::traits<Derived>::CoeffReadCost != Dynamic
-                   && DenseDerived::SizeAtCompileTime * internal::traits<Derived>::CoeffReadCost / 2
-                        <= EIGEN_UNROLLING_LIMIT
-  };
-  other.derived().resize(this->rows(), this->cols());
-
-  internal::triangular_assignment_selector
-    <DenseDerived, typename internal::traits<Derived>::MatrixTypeNestedCleaned, Derived::Mode,
-    unroll ? int(DenseDerived::SizeAtCompileTime) : Dynamic,
-    true // clear the opposite triangular part
-    >::run(other.derived(), derived().nestedExpression());
+  evalToLazy(other.derived());
 }
 
 /***************************************************************************
@@ -719,67 +610,34 @@ void TriangularBase<Derived>::evalToLazy(MatrixBase<DenseDerived> &other) const
 * Implementation of MatrixBase methods
 ***************************************************************************/
 
-#ifdef EIGEN2_SUPPORT
-
-// implementation of part<>(), including the SelfAdjoint case.
-
-namespace internal {
-template<typename MatrixType, unsigned int Mode>
-struct eigen2_part_return_type
-{
-  typedef TriangularView<MatrixType, Mode> type;
-};
-
-template<typename MatrixType>
-struct eigen2_part_return_type<MatrixType, SelfAdjoint>
-{
-  typedef SelfAdjointView<MatrixType, Upper> type;
-};
-}
-
-/** \deprecated use MatrixBase::triangularView() */
-template<typename Derived>
-template<unsigned int Mode>
-const typename internal::eigen2_part_return_type<Derived, Mode>::type MatrixBase<Derived>::part() const
-{
-  return derived();
-}
-
-/** \deprecated use MatrixBase::triangularView() */
-template<typename Derived>
-template<unsigned int Mode>
-typename internal::eigen2_part_return_type<Derived, Mode>::type MatrixBase<Derived>::part()
-{
-  return derived();
-}
-#endif
-
 /**
   * \returns an expression of a triangular view extracted from the current matrix
   *
   * The parameter \a Mode can have the following values: \c #Upper, \c #StrictlyUpper, \c #UnitUpper,
   * \c #Lower, \c #StrictlyLower, \c #UnitLower.
   *
-  * Example: \include MatrixBase_extract.cpp
-  * Output: \verbinclude MatrixBase_extract.out
+  * Example: \include MatrixBase_triangularView.cpp
+  * Output: \verbinclude MatrixBase_triangularView.out
   *
   * \sa class TriangularView
   */
 template<typename Derived>
 template<unsigned int Mode>
+EIGEN_DEVICE_FUNC
 typename MatrixBase<Derived>::template TriangularViewReturnType<Mode>::Type
 MatrixBase<Derived>::triangularView()
 {
-  return derived();
+  return typename TriangularViewReturnType<Mode>::Type(derived());
 }
 
 /** This is the const version of MatrixBase::triangularView() */
 template<typename Derived>
 template<unsigned int Mode>
+EIGEN_DEVICE_FUNC
 typename MatrixBase<Derived>::template ConstTriangularViewReturnType<Mode>::Type
 MatrixBase<Derived>::triangularView() const
 {
-  return derived();
+  return typename ConstTriangularViewReturnType<Mode>::Type(derived());
 }
 
 /** \returns true if *this is approximately equal to an upper triangular matrix,
@@ -790,21 +648,20 @@ MatrixBase<Derived>::triangularView() const
 template<typename Derived>
 bool MatrixBase<Derived>::isUpperTriangular(const RealScalar& prec) const
 {
-  using std::abs;
   RealScalar maxAbsOnUpperPart = static_cast<RealScalar>(-1);
   for(Index j = 0; j < cols(); ++j)
   {
-    Index maxi = (std::min)(j, rows()-1);
+    Index maxi = numext::mini(j, rows()-1);
     for(Index i = 0; i <= maxi; ++i)
     {
-      RealScalar absValue = abs(coeff(i,j));
+      RealScalar absValue = numext::abs(coeff(i,j));
       if(absValue > maxAbsOnUpperPart) maxAbsOnUpperPart = absValue;
     }
   }
   RealScalar threshold = maxAbsOnUpperPart * prec;
   for(Index j = 0; j < cols(); ++j)
     for(Index i = j+1; i < rows(); ++i)
-      if(abs(coeff(i, j)) > threshold) return false;
+      if(numext::abs(coeff(i, j)) > threshold) return false;
   return true;
 }
 
@@ -816,24 +673,312 @@ bool MatrixBase<Derived>::isUpperTriangular(const RealScalar& prec) const
 template<typename Derived>
 bool MatrixBase<Derived>::isLowerTriangular(const RealScalar& prec) const
 {
-  using std::abs;
   RealScalar maxAbsOnLowerPart = static_cast<RealScalar>(-1);
   for(Index j = 0; j < cols(); ++j)
     for(Index i = j; i < rows(); ++i)
     {
-      RealScalar absValue = abs(coeff(i,j));
+      RealScalar absValue = numext::abs(coeff(i,j));
       if(absValue > maxAbsOnLowerPart) maxAbsOnLowerPart = absValue;
     }
   RealScalar threshold = maxAbsOnLowerPart * prec;
   for(Index j = 1; j < cols(); ++j)
   {
-    Index maxi = (std::min)(j, rows()-1);
+    Index maxi = numext::mini(j, rows()-1);
     for(Index i = 0; i < maxi; ++i)
-      if(abs(coeff(i, j)) > threshold) return false;
+      if(numext::abs(coeff(i, j)) > threshold) return false;
   }
   return true;
 }
 
+
+/***************************************************************************
+****************************************************************************
+* Evaluators and Assignment of triangular expressions
+***************************************************************************
+***************************************************************************/
+
+namespace internal {
+
+  
+// TODO currently a triangular expression has the form TriangularView<.,.>
+//      in the future triangular-ness should be defined by the expression traits
+//      such that Transpose<TriangularView<.,.> > is valid. (currently TriangularBase::transpose() is overloaded to make it work)
+template<typename MatrixType, unsigned int Mode>
+struct evaluator_traits<TriangularView<MatrixType,Mode> >
+{
+  typedef typename storage_kind_to_evaluator_kind<typename MatrixType::StorageKind>::Kind Kind;
+  typedef typename glue_shapes<typename evaluator_traits<MatrixType>::Shape, TriangularShape>::type Shape;
+};
+
+template<typename MatrixType, unsigned int Mode>
+struct unary_evaluator<TriangularView<MatrixType,Mode>, IndexBased>
+ : evaluator<typename internal::remove_all<MatrixType>::type>
+{
+  typedef TriangularView<MatrixType,Mode> XprType;
+  typedef evaluator<typename internal::remove_all<MatrixType>::type> Base;
+  unary_evaluator(const XprType &xpr) : Base(xpr.nestedExpression()) {}
+};
+
+// Additional assignment kinds:
+struct Triangular2Triangular    {};
+struct Triangular2Dense         {};
+struct Dense2Triangular         {};
+
+
+template<typename Kernel, unsigned int Mode, int UnrollCount, bool ClearOpposite> struct triangular_assignment_loop;
+
+ 
+/** \internal Specialization of the dense assignment kernel for triangular matrices.
+  * The main difference is that the triangular, diagonal, and opposite parts are processed through three different functions.
+  * \tparam UpLo must be either Lower or Upper
+  * \tparam Mode must be either 0, UnitDiag, ZeroDiag, or SelfAdjoint
+  */
+template<int UpLo, int Mode, int SetOpposite, typename DstEvaluatorTypeT, typename SrcEvaluatorTypeT, typename Functor, int Version = Specialized>
+class triangular_dense_assignment_kernel : public generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, Functor, Version>
+{
+protected:
+  typedef generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, Functor, Version> Base;
+  typedef typename Base::DstXprType DstXprType;
+  typedef typename Base::SrcXprType SrcXprType;
+  using Base::m_dst;
+  using Base::m_src;
+  using Base::m_functor;
+public:
+  
+  typedef typename Base::DstEvaluatorType DstEvaluatorType;
+  typedef typename Base::SrcEvaluatorType SrcEvaluatorType;
+  typedef typename Base::Scalar Scalar;
+  typedef typename Base::AssignmentTraits AssignmentTraits;
+  
+  
+  EIGEN_DEVICE_FUNC triangular_dense_assignment_kernel(DstEvaluatorType &dst, const SrcEvaluatorType &src, const Functor &func, DstXprType& dstExpr)
+    : Base(dst, src, func, dstExpr)
+  {}
+  
+#ifdef EIGEN_INTERNAL_DEBUGGING
+  EIGEN_DEVICE_FUNC void assignCoeff(Index row, Index col)
+  {
+    eigen_internal_assert(row!=col);
+    Base::assignCoeff(row,col);
+  }
+#else
+  using Base::assignCoeff;
+#endif
+  
+  EIGEN_DEVICE_FUNC void assignDiagonalCoeff(Index id)
+  {
+         if(Mode==UnitDiag && SetOpposite) m_functor.assignCoeff(m_dst.coeffRef(id,id), Scalar(1));
+    else if(Mode==ZeroDiag && SetOpposite) m_functor.assignCoeff(m_dst.coeffRef(id,id), Scalar(0));
+    else if(Mode==0)                       Base::assignCoeff(id,id);
+  }
+  
+  EIGEN_DEVICE_FUNC void assignOppositeCoeff(Index row, Index col)
+  { 
+    eigen_internal_assert(row!=col);
+    if(SetOpposite)
+      m_functor.assignCoeff(m_dst.coeffRef(row,col), Scalar(0));
+  }
+};
+
+template<int Mode, bool SetOpposite, typename DstXprType, typename SrcXprType, typename Functor>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void call_triangular_assignment_loop(DstXprType& dst, const SrcXprType& src, const Functor &func)
+{
+  typedef evaluator<DstXprType> DstEvaluatorType;
+  typedef evaluator<SrcXprType> SrcEvaluatorType;
+
+  SrcEvaluatorType srcEvaluator(src);
+
+  Index dstRows = src.rows();
+  Index dstCols = src.cols();
+  if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+    dst.resize(dstRows, dstCols);
+  DstEvaluatorType dstEvaluator(dst);
+    
+  typedef triangular_dense_assignment_kernel< Mode&(Lower|Upper),Mode&(UnitDiag|ZeroDiag|SelfAdjoint),SetOpposite,
+                                              DstEvaluatorType,SrcEvaluatorType,Functor> Kernel;
+  Kernel kernel(dstEvaluator, srcEvaluator, func, dst.const_cast_derived());
+  
+  enum {
+      unroll = DstXprType::SizeAtCompileTime != Dynamic
+            && SrcEvaluatorType::CoeffReadCost < HugeCost
+            && DstXprType::SizeAtCompileTime * (DstEvaluatorType::CoeffReadCost+SrcEvaluatorType::CoeffReadCost) / 2 <= EIGEN_UNROLLING_LIMIT
+    };
+  
+  triangular_assignment_loop<Kernel, Mode, unroll ? int(DstXprType::SizeAtCompileTime) : Dynamic, SetOpposite>::run(kernel);
+}
+
+template<int Mode, bool SetOpposite, typename DstXprType, typename SrcXprType>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void call_triangular_assignment_loop(DstXprType& dst, const SrcXprType& src)
+{
+  call_triangular_assignment_loop<Mode,SetOpposite>(dst, src, internal::assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar>());
+}
+
+template<> struct AssignmentKind<TriangularShape,TriangularShape> { typedef Triangular2Triangular Kind; };
+template<> struct AssignmentKind<DenseShape,TriangularShape>      { typedef Triangular2Dense      Kind; };
+template<> struct AssignmentKind<TriangularShape,DenseShape>      { typedef Dense2Triangular      Kind; };
+
+
+template< typename DstXprType, typename SrcXprType, typename Functor>
+struct Assignment<DstXprType, SrcXprType, Functor, Triangular2Triangular>
+{
+  EIGEN_DEVICE_FUNC static void run(DstXprType &dst, const SrcXprType &src, const Functor &func)
+  {
+    eigen_assert(int(DstXprType::Mode) == int(SrcXprType::Mode));
+    
+    call_triangular_assignment_loop<DstXprType::Mode, false>(dst, src, func);  
+  }
+};
+
+template< typename DstXprType, typename SrcXprType, typename Functor>
+struct Assignment<DstXprType, SrcXprType, Functor, Triangular2Dense>
+{
+  EIGEN_DEVICE_FUNC static void run(DstXprType &dst, const SrcXprType &src, const Functor &func)
+  {
+    call_triangular_assignment_loop<SrcXprType::Mode, (SrcXprType::Mode&SelfAdjoint)==0>(dst, src, func);  
+  }
+};
+
+template< typename DstXprType, typename SrcXprType, typename Functor>
+struct Assignment<DstXprType, SrcXprType, Functor, Dense2Triangular>
+{
+  EIGEN_DEVICE_FUNC static void run(DstXprType &dst, const SrcXprType &src, const Functor &func)
+  {
+    call_triangular_assignment_loop<DstXprType::Mode, false>(dst, src, func);  
+  }
+};
+
+
+template<typename Kernel, unsigned int Mode, int UnrollCount, bool SetOpposite>
+struct triangular_assignment_loop
+{
+  // FIXME: this is not very clean, perhaps this information should be provided by the kernel?
+  typedef typename Kernel::DstEvaluatorType DstEvaluatorType;
+  typedef typename DstEvaluatorType::XprType DstXprType;
+  
+  enum {
+    col = (UnrollCount-1) / DstXprType::RowsAtCompileTime,
+    row = (UnrollCount-1) % DstXprType::RowsAtCompileTime
+  };
+  
+  typedef typename Kernel::Scalar Scalar;
+
+  EIGEN_DEVICE_FUNC
+  static inline void run(Kernel &kernel)
+  {
+    triangular_assignment_loop<Kernel, Mode, UnrollCount-1, SetOpposite>::run(kernel);
+    
+    if(row==col)
+      kernel.assignDiagonalCoeff(row);
+    else if( ((Mode&Lower) && row>col) || ((Mode&Upper) && row<col) )
+      kernel.assignCoeff(row,col);
+    else if(SetOpposite)
+      kernel.assignOppositeCoeff(row,col);
+  }
+};
+
+// prevent buggy user code from causing an infinite recursion
+template<typename Kernel, unsigned int Mode, bool SetOpposite>
+struct triangular_assignment_loop<Kernel, Mode, 0, SetOpposite>
+{
+  EIGEN_DEVICE_FUNC
+  static inline void run(Kernel &) {}
+};
+
+
+
+// TODO: experiment with a recursive assignment procedure splitting the current
+//       triangular part into one rectangular and two triangular parts.
+
+
+template<typename Kernel, unsigned int Mode, bool SetOpposite>
+struct triangular_assignment_loop<Kernel, Mode, Dynamic, SetOpposite>
+{
+  typedef typename Kernel::Scalar Scalar;
+  EIGEN_DEVICE_FUNC
+  static inline void run(Kernel &kernel)
+  {
+    for(Index j = 0; j < kernel.cols(); ++j)
+    {
+      Index maxi = numext::mini(j, kernel.rows());
+      Index i = 0;
+      if (((Mode&Lower) && SetOpposite) || (Mode&Upper))
+      {
+        for(; i < maxi; ++i)
+          if(Mode&Upper) kernel.assignCoeff(i, j);
+          else           kernel.assignOppositeCoeff(i, j);
+      }
+      else
+        i = maxi;
+      
+      if(i<kernel.rows()) // then i==j
+        kernel.assignDiagonalCoeff(i++);
+      
+      if (((Mode&Upper) && SetOpposite) || (Mode&Lower))
+      {
+        for(; i < kernel.rows(); ++i)
+          if(Mode&Lower) kernel.assignCoeff(i, j);
+          else           kernel.assignOppositeCoeff(i, j);
+      }
+    }
+  }
+};
+
+} // end namespace internal
+
+/** Assigns a triangular or selfadjoint matrix to a dense matrix.
+  * If the matrix is triangular, the opposite part is set to zero. */
+template<typename Derived>
+template<typename DenseDerived>
+EIGEN_DEVICE_FUNC void TriangularBase<Derived>::evalToLazy(MatrixBase<DenseDerived> &other) const
+{
+  other.derived().resize(this->rows(), this->cols());
+  internal::call_triangular_assignment_loop<Derived::Mode,(Derived::Mode&SelfAdjoint)==0 /* SetOpposite */>(other.derived(), derived().nestedExpression());
+}
+
+namespace internal {
+  
+// Triangular = Product
+template< typename DstXprType, typename Lhs, typename Rhs, typename Scalar>
+struct Assignment<DstXprType, Product<Lhs,Rhs,DefaultProduct>, internal::assign_op<Scalar,typename Product<Lhs,Rhs,DefaultProduct>::Scalar>, Dense2Triangular>
+{
+  typedef Product<Lhs,Rhs,DefaultProduct> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,typename SrcXprType::Scalar> &)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+
+    dst._assignProduct(src, 1, 0);
+  }
+};
+
+// Triangular += Product
+template< typename DstXprType, typename Lhs, typename Rhs, typename Scalar>
+struct Assignment<DstXprType, Product<Lhs,Rhs,DefaultProduct>, internal::add_assign_op<Scalar,typename Product<Lhs,Rhs,DefaultProduct>::Scalar>, Dense2Triangular>
+{
+  typedef Product<Lhs,Rhs,DefaultProduct> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op<Scalar,typename SrcXprType::Scalar> &)
+  {
+    dst._assignProduct(src, 1, 1);
+  }
+};
+
+// Triangular -= Product
+template< typename DstXprType, typename Lhs, typename Rhs, typename Scalar>
+struct Assignment<DstXprType, Product<Lhs,Rhs,DefaultProduct>, internal::sub_assign_op<Scalar,typename Product<Lhs,Rhs,DefaultProduct>::Scalar>, Dense2Triangular>
+{
+  typedef Product<Lhs,Rhs,DefaultProduct> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op<Scalar,typename SrcXprType::Scalar> &)
+  {
+    dst._assignProduct(src, -1, 1);
+  }
+};
+
+} // end namespace internal
+
 } // end namespace Eigen
 
 #endif // EIGEN_TRIANGULARMATRIX_H
diff --git a/eigen/Eigen/src/Core/VectorBlock.h b/eigen/Eigen/src/Core/VectorBlock.h
index 1a7330f..d72fbf7 100644
--- a/eigen/Eigen/src/Core/VectorBlock.h
+++ b/eigen/Eigen/src/Core/VectorBlock.h
@@ -13,13 +13,23 @@
 
 namespace Eigen { 
 
+namespace internal {
+template<typename VectorType, int Size>
+struct traits<VectorBlock<VectorType, Size> >
+  : public traits<Block<VectorType,
+                     traits<VectorType>::Flags & RowMajorBit ? 1 : Size,
+                     traits<VectorType>::Flags & RowMajorBit ? Size : 1> >
+{
+};
+}
+
 /** \class VectorBlock
   * \ingroup Core_Module
   *
   * \brief Expression of a fixed-size or dynamic-size sub-vector
   *
-  * \param VectorType the type of the object in which we are taking a sub-vector
-  * \param Size size of the sub-vector we are taking at compile time (optional)
+  * \tparam VectorType the type of the object in which we are taking a sub-vector
+  * \tparam Size size of the sub-vector we are taking at compile time (optional)
   *
   * This class represents an expression of either a fixed-size or dynamic-size sub-vector.
   * It is the return type of DenseBase::segment(Index,Index) and DenseBase::segment<int>(Index) and
@@ -43,17 +53,6 @@ namespace Eigen {
   *
   * \sa class Block, DenseBase::segment(Index,Index,Index,Index), DenseBase::segment(Index,Index)
   */
-
-namespace internal {
-template<typename VectorType, int Size>
-struct traits<VectorBlock<VectorType, Size> >
-  : public traits<Block<VectorType,
-                     traits<VectorType>::Flags & RowMajorBit ? 1 : Size,
-                     traits<VectorType>::Flags & RowMajorBit ? Size : 1> >
-{
-};
-}
-
 template<typename VectorType, int Size> class VectorBlock
   : public Block<VectorType,
                      internal::traits<VectorType>::Flags & RowMajorBit ? 1 : Size,
@@ -72,6 +71,7 @@ template<typename VectorType, int Size> class VectorBlock
 
     /** Dynamic-size constructor
       */
+    EIGEN_DEVICE_FUNC
     inline VectorBlock(VectorType& vector, Index start, Index size)
       : Base(vector,
              IsColVector ? start : 0, IsColVector ? 0 : start,
@@ -82,6 +82,7 @@ template<typename VectorType, int Size> class VectorBlock
 
     /** Fixed-size constructor
       */
+    EIGEN_DEVICE_FUNC
     inline VectorBlock(VectorType& vector, Index start)
       : Base(vector, IsColVector ? start : 0, IsColVector ? 0 : start)
     {
diff --git a/eigen/Eigen/src/Core/VectorwiseOp.h b/eigen/Eigen/src/Core/VectorwiseOp.h
index d5ab036..893bc79 100644
--- a/eigen/Eigen/src/Core/VectorwiseOp.h
+++ b/eigen/Eigen/src/Core/VectorwiseOp.h
@@ -11,7 +11,7 @@
 #ifndef EIGEN_PARTIAL_REDUX_H
 #define EIGEN_PARTIAL_REDUX_H
 
-namespace Eigen { 
+namespace Eigen {
 
 /** \class PartialReduxExpr
   * \ingroup Core_Module
@@ -41,64 +41,43 @@ struct traits<PartialReduxExpr<MatrixType, MemberOp, Direction> >
   typedef typename traits<MatrixType>::StorageKind StorageKind;
   typedef typename traits<MatrixType>::XprKind XprKind;
   typedef typename MatrixType::Scalar InputScalar;
-  typedef typename nested<MatrixType>::type MatrixTypeNested;
-  typedef typename remove_all<MatrixTypeNested>::type _MatrixTypeNested;
   enum {
     RowsAtCompileTime = Direction==Vertical   ? 1 : MatrixType::RowsAtCompileTime,
     ColsAtCompileTime = Direction==Horizontal ? 1 : MatrixType::ColsAtCompileTime,
     MaxRowsAtCompileTime = Direction==Vertical   ? 1 : MatrixType::MaxRowsAtCompileTime,
     MaxColsAtCompileTime = Direction==Horizontal ? 1 : MatrixType::MaxColsAtCompileTime,
-    Flags0 = (unsigned int)_MatrixTypeNested::Flags & HereditaryBits,
-    Flags = (Flags0 & ~RowMajorBit) | (RowsAtCompileTime == 1 ? RowMajorBit : 0),
+    Flags = RowsAtCompileTime == 1 ? RowMajorBit : 0,
     TraversalSize = Direction==Vertical ? MatrixType::RowsAtCompileTime :  MatrixType::ColsAtCompileTime
   };
-  #if EIGEN_GNUC_AT_LEAST(3,4)
-  typedef typename MemberOp::template Cost<InputScalar,int(TraversalSize)> CostOpType;
-  #else
-  typedef typename MemberOp::template Cost<InputScalar,TraversalSize> CostOpType;
-  #endif
-  enum {
-    CoeffReadCost = TraversalSize==Dynamic ? Dynamic
-                  : TraversalSize * traits<_MatrixTypeNested>::CoeffReadCost + int(CostOpType::value)
-  };
 };
 }
 
 template< typename MatrixType, typename MemberOp, int Direction>
-class PartialReduxExpr : internal::no_assignment_operator,
-  public internal::dense_xpr_base< PartialReduxExpr<MatrixType, MemberOp, Direction> >::type
+class PartialReduxExpr : public internal::dense_xpr_base< PartialReduxExpr<MatrixType, MemberOp, Direction> >::type,
+                         internal::no_assignment_operator
 {
   public:
 
     typedef typename internal::dense_xpr_base<PartialReduxExpr>::type Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(PartialReduxExpr)
-    typedef typename internal::traits<PartialReduxExpr>::MatrixTypeNested MatrixTypeNested;
-    typedef typename internal::traits<PartialReduxExpr>::_MatrixTypeNested _MatrixTypeNested;
 
-    PartialReduxExpr(const MatrixType& mat, const MemberOp& func = MemberOp())
+    EIGEN_DEVICE_FUNC
+    explicit PartialReduxExpr(const MatrixType& mat, const MemberOp& func = MemberOp())
       : m_matrix(mat), m_functor(func) {}
 
+    EIGEN_DEVICE_FUNC
     Index rows() const { return (Direction==Vertical   ? 1 : m_matrix.rows()); }
+    EIGEN_DEVICE_FUNC
     Index cols() const { return (Direction==Horizontal ? 1 : m_matrix.cols()); }
 
-    EIGEN_STRONG_INLINE const Scalar coeff(Index i, Index j) const
-    {
-      if (Direction==Vertical)
-        return m_functor(m_matrix.col(j));
-      else
-        return m_functor(m_matrix.row(i));
-    }
+    EIGEN_DEVICE_FUNC
+    typename MatrixType::Nested nestedExpression() const { return m_matrix; }
 
-    const Scalar coeff(Index index) const
-    {
-      if (Direction==Vertical)
-        return m_functor(m_matrix.col(index));
-      else
-        return m_functor(m_matrix.row(index));
-    }
+    EIGEN_DEVICE_FUNC
+    const MemberOp& functor() const { return m_functor; }
 
   protected:
-    MatrixTypeNested m_matrix;
+    typename MatrixType::Nested m_matrix;
     const MemberOp m_functor;
 };
 
@@ -110,7 +89,8 @@ class PartialReduxExpr : internal::no_assignment_operator,
     template<typename Scalar, int Size> struct Cost                     \
     { enum { value = COST }; };                                         \
     template<typename XprType>                                          \
-    EIGEN_STRONG_INLINE ResultType operator()(const XprType& mat) const \
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE                               \
+    ResultType operator()(const XprType& mat) const                     \
     { return mat.MEMBER(); } \
   }
 
@@ -130,17 +110,27 @@ EIGEN_MEMBER_FUNCTOR(any, (Size-1)*NumTraits<Scalar>::AddCost);
 EIGEN_MEMBER_FUNCTOR(count, (Size-1)*NumTraits<Scalar>::AddCost);
 EIGEN_MEMBER_FUNCTOR(prod, (Size-1)*NumTraits<Scalar>::MulCost);
 
+template <int p, typename ResultType>
+struct member_lpnorm {
+  typedef ResultType result_type;
+  template<typename Scalar, int Size> struct Cost
+  { enum { value = (Size+5) * NumTraits<Scalar>::MulCost + (Size-1)*NumTraits<Scalar>::AddCost }; };
+  EIGEN_DEVICE_FUNC member_lpnorm() {}
+  template<typename XprType>
+  EIGEN_DEVICE_FUNC inline ResultType operator()(const XprType& mat) const
+  { return mat.template lpNorm<p>(); }
+};
 
 template <typename BinaryOp, typename Scalar>
 struct member_redux {
   typedef typename result_of<
-                     BinaryOp(Scalar)
+                     BinaryOp(const Scalar&,const Scalar&)
                    >::type  result_type;
   template<typename _Scalar, int Size> struct Cost
   { enum { value = (Size-1) * functor_traits<BinaryOp>::Cost }; };
-  member_redux(const BinaryOp func) : m_functor(func) {}
+  EIGEN_DEVICE_FUNC explicit member_redux(const BinaryOp func) : m_functor(func) {}
   template<typename Derived>
-  inline result_type operator()(const DenseBase<Derived>& mat) const
+  EIGEN_DEVICE_FUNC inline result_type operator()(const DenseBase<Derived>& mat) const
   { return mat.redux(m_functor); }
   const BinaryOp m_functor;
 };
@@ -151,8 +141,8 @@ struct member_redux {
   *
   * \brief Pseudo expression providing partial reduction operations
   *
-  * \param ExpressionType the type of the object on which to do partial reductions
-  * \param Direction indicates the direction of the redux (#Vertical or #Horizontal)
+  * \tparam ExpressionType the type of the object on which to do partial reductions
+  * \tparam Direction indicates the direction of the redux (#Vertical or #Horizontal)
   *
   * This class represents a pseudo expression with partial reduction features.
   * It is the return type of DenseBase::colwise() and DenseBase::rowwise()
@@ -169,16 +159,15 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 
     typedef typename ExpressionType::Scalar Scalar;
     typedef typename ExpressionType::RealScalar RealScalar;
-    typedef typename ExpressionType::Index Index;
-    typedef typename internal::conditional<internal::must_nest_by_value<ExpressionType>::ret,
-        ExpressionType, ExpressionType&>::type ExpressionTypeNested;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+    typedef typename internal::ref_selector<ExpressionType>::non_const_type ExpressionTypeNested;
     typedef typename internal::remove_all<ExpressionTypeNested>::type ExpressionTypeNestedCleaned;
 
     template<template<typename _Scalar> class Functor,
-                      typename Scalar=typename internal::traits<ExpressionType>::Scalar> struct ReturnType
+                      typename Scalar_=Scalar> struct ReturnType
     {
       typedef PartialReduxExpr<ExpressionType,
-                               Functor<Scalar>,
+                               Functor<Scalar_>,
                                Direction
                               > Type;
     };
@@ -186,23 +175,24 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
     template<typename BinaryOp> struct ReduxReturnType
     {
       typedef PartialReduxExpr<ExpressionType,
-                               internal::member_redux<BinaryOp,typename internal::traits<ExpressionType>::Scalar>,
+                               internal::member_redux<BinaryOp,Scalar>,
                                Direction
                               > Type;
     };
 
     enum {
-      IsVertical   = (Direction==Vertical) ? 1 : 0,
-      IsHorizontal = (Direction==Horizontal) ? 1 : 0
+      isVertical   = (Direction==Vertical) ? 1 : 0,
+      isHorizontal = (Direction==Horizontal) ? 1 : 0
     };
 
   protected:
 
-    /** \internal
-      * \returns the i-th subvector according to the \c Direction */
-    typedef typename internal::conditional<Direction==Vertical,
+    typedef typename internal::conditional<isVertical,
                                typename ExpressionType::ColXpr,
                                typename ExpressionType::RowXpr>::type SubVector;
+    /** \internal
+      * \returns the i-th subvector according to the \c Direction */
+    EIGEN_DEVICE_FUNC
     SubVector subVector(Index i)
     {
       return SubVector(m_matrix.derived(),i);
@@ -210,58 +200,62 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 
     /** \internal
       * \returns the number of subvectors in the direction \c Direction */
+    EIGEN_DEVICE_FUNC
     Index subVectors() const
-    { return Direction==Vertical?m_matrix.cols():m_matrix.rows(); }
+    { return isVertical?m_matrix.cols():m_matrix.rows(); }
 
     template<typename OtherDerived> struct ExtendedType {
       typedef Replicate<OtherDerived,
-                        Direction==Vertical   ? 1 : ExpressionType::RowsAtCompileTime,
-                        Direction==Horizontal ? 1 : ExpressionType::ColsAtCompileTime> Type;
+                        isVertical   ? 1 : ExpressionType::RowsAtCompileTime,
+                        isHorizontal ? 1 : ExpressionType::ColsAtCompileTime> Type;
     };
 
     /** \internal
       * Replicates a vector to match the size of \c *this */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     typename ExtendedType<OtherDerived>::Type
     extendedTo(const DenseBase<OtherDerived>& other) const
     {
-      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(Direction==Vertical, OtherDerived::MaxColsAtCompileTime==1),
+      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(isVertical, OtherDerived::MaxColsAtCompileTime==1),
                           YOU_PASSED_A_ROW_VECTOR_BUT_A_COLUMN_VECTOR_WAS_EXPECTED)
-      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(Direction==Horizontal, OtherDerived::MaxRowsAtCompileTime==1),
+      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(isHorizontal, OtherDerived::MaxRowsAtCompileTime==1),
                           YOU_PASSED_A_COLUMN_VECTOR_BUT_A_ROW_VECTOR_WAS_EXPECTED)
       return typename ExtendedType<OtherDerived>::Type
                       (other.derived(),
-                       Direction==Vertical   ? 1 : m_matrix.rows(),
-                       Direction==Horizontal ? 1 : m_matrix.cols());
+                       isVertical   ? 1 : m_matrix.rows(),
+                       isHorizontal ? 1 : m_matrix.cols());
     }
-    
+
     template<typename OtherDerived> struct OppositeExtendedType {
       typedef Replicate<OtherDerived,
-                        Direction==Horizontal ? 1 : ExpressionType::RowsAtCompileTime,
-                        Direction==Vertical   ? 1 : ExpressionType::ColsAtCompileTime> Type;
+                        isHorizontal ? 1 : ExpressionType::RowsAtCompileTime,
+                        isVertical   ? 1 : ExpressionType::ColsAtCompileTime> Type;
     };
 
     /** \internal
       * Replicates a vector in the opposite direction to match the size of \c *this */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     typename OppositeExtendedType<OtherDerived>::Type
     extendedToOpposite(const DenseBase<OtherDerived>& other) const
     {
-      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(Direction==Horizontal, OtherDerived::MaxColsAtCompileTime==1),
+      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(isHorizontal, OtherDerived::MaxColsAtCompileTime==1),
                           YOU_PASSED_A_ROW_VECTOR_BUT_A_COLUMN_VECTOR_WAS_EXPECTED)
-      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(Direction==Vertical, OtherDerived::MaxRowsAtCompileTime==1),
+      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(isVertical, OtherDerived::MaxRowsAtCompileTime==1),
                           YOU_PASSED_A_COLUMN_VECTOR_BUT_A_ROW_VECTOR_WAS_EXPECTED)
       return typename OppositeExtendedType<OtherDerived>::Type
                       (other.derived(),
-                       Direction==Horizontal  ? 1 : m_matrix.rows(),
-                       Direction==Vertical    ? 1 : m_matrix.cols());
+                       isHorizontal  ? 1 : m_matrix.rows(),
+                       isVertical    ? 1 : m_matrix.cols());
     }
 
   public:
-
-    inline VectorwiseOp(ExpressionType& matrix) : m_matrix(matrix) {}
+    EIGEN_DEVICE_FUNC
+    explicit inline VectorwiseOp(ExpressionType& matrix) : m_matrix(matrix) {}
 
     /** \internal */
+    EIGEN_DEVICE_FUNC
     inline const ExpressionType& _expression() const { return m_matrix; }
 
     /** \returns a row or column vector expression of \c *this reduxed by \a func
@@ -272,80 +266,126 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * \sa class VectorwiseOp, DenseBase::colwise(), DenseBase::rowwise()
       */
     template<typename BinaryOp>
+    EIGEN_DEVICE_FUNC
     const typename ReduxReturnType<BinaryOp>::Type
     redux(const BinaryOp& func = BinaryOp()) const
-    { return typename ReduxReturnType<BinaryOp>::Type(_expression(), func); }
+    { return typename ReduxReturnType<BinaryOp>::Type(_expression(), internal::member_redux<BinaryOp,Scalar>(func)); }
+
+    typedef typename ReturnType<internal::member_minCoeff>::Type MinCoeffReturnType;
+    typedef typename ReturnType<internal::member_maxCoeff>::Type MaxCoeffReturnType;
+    typedef typename ReturnType<internal::member_squaredNorm,RealScalar>::Type SquaredNormReturnType;
+    typedef typename ReturnType<internal::member_norm,RealScalar>::Type NormReturnType;
+    typedef typename ReturnType<internal::member_blueNorm,RealScalar>::Type BlueNormReturnType;
+    typedef typename ReturnType<internal::member_stableNorm,RealScalar>::Type StableNormReturnType;
+    typedef typename ReturnType<internal::member_hypotNorm,RealScalar>::Type HypotNormReturnType;
+    typedef typename ReturnType<internal::member_sum>::Type SumReturnType;
+    typedef typename ReturnType<internal::member_mean>::Type MeanReturnType;
+    typedef typename ReturnType<internal::member_all>::Type AllReturnType;
+    typedef typename ReturnType<internal::member_any>::Type AnyReturnType;
+    typedef PartialReduxExpr<ExpressionType, internal::member_count<Index>, Direction> CountReturnType;
+    typedef typename ReturnType<internal::member_prod>::Type ProdReturnType;
+    typedef Reverse<const ExpressionType, Direction> ConstReverseReturnType;
+    typedef Reverse<ExpressionType, Direction> ReverseReturnType;
+
+    template<int p> struct LpNormReturnType {
+      typedef PartialReduxExpr<ExpressionType, internal::member_lpnorm<p,RealScalar>,Direction> Type;
+    };
 
     /** \returns a row (or column) vector expression of the smallest coefficient
       * of each column (or row) of the referenced expression.
-      * 
+      *
       * \warning the result is undefined if \c *this contains NaN.
       *
       * Example: \include PartialRedux_minCoeff.cpp
       * Output: \verbinclude PartialRedux_minCoeff.out
       *
       * \sa DenseBase::minCoeff() */
-    const typename ReturnType<internal::member_minCoeff>::Type minCoeff() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const MinCoeffReturnType minCoeff() const
+    { return MinCoeffReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression of the largest coefficient
       * of each column (or row) of the referenced expression.
-      * 
+      *
       * \warning the result is undefined if \c *this contains NaN.
       *
       * Example: \include PartialRedux_maxCoeff.cpp
       * Output: \verbinclude PartialRedux_maxCoeff.out
       *
       * \sa DenseBase::maxCoeff() */
-    const typename ReturnType<internal::member_maxCoeff>::Type maxCoeff() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const MaxCoeffReturnType maxCoeff() const
+    { return MaxCoeffReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression of the squared norm
       * of each column (or row) of the referenced expression.
+      * This is a vector with real entries, even if the original matrix has complex entries.
       *
       * Example: \include PartialRedux_squaredNorm.cpp
       * Output: \verbinclude PartialRedux_squaredNorm.out
       *
       * \sa DenseBase::squaredNorm() */
-    const typename ReturnType<internal::member_squaredNorm,RealScalar>::Type squaredNorm() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const SquaredNormReturnType squaredNorm() const
+    { return SquaredNormReturnType(_expression()); }
+
+    /** \returns a row (or column) vector expression of the norm
+      * of each column (or row) of the referenced expression.
+      * This is a vector with real entries, even if the original matrix has complex entries.
+      *
+      * Example: \include PartialRedux_norm.cpp
+      * Output: \verbinclude PartialRedux_norm.out
+      *
+      * \sa DenseBase::norm() */
+    EIGEN_DEVICE_FUNC
+    const NormReturnType norm() const
+    { return NormReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression of the norm
       * of each column (or row) of the referenced expression.
+      * This is a vector with real entries, even if the original matrix has complex entries.
       *
       * Example: \include PartialRedux_norm.cpp
       * Output: \verbinclude PartialRedux_norm.out
       *
       * \sa DenseBase::norm() */
-    const typename ReturnType<internal::member_norm,RealScalar>::Type norm() const
-    { return _expression(); }
+    template<int p>
+    EIGEN_DEVICE_FUNC
+    const typename LpNormReturnType<p>::Type lpNorm() const
+    { return typename LpNormReturnType<p>::Type(_expression()); }
 
 
     /** \returns a row (or column) vector expression of the norm
       * of each column (or row) of the referenced expression, using
-      * blue's algorithm.
+      * Blue's algorithm.
+      * This is a vector with real entries, even if the original matrix has complex entries.
       *
       * \sa DenseBase::blueNorm() */
-    const typename ReturnType<internal::member_blueNorm,RealScalar>::Type blueNorm() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const BlueNormReturnType blueNorm() const
+    { return BlueNormReturnType(_expression()); }
 
 
     /** \returns a row (or column) vector expression of the norm
       * of each column (or row) of the referenced expression, avoiding
       * underflow and overflow.
+      * This is a vector with real entries, even if the original matrix has complex entries.
       *
       * \sa DenseBase::stableNorm() */
-    const typename ReturnType<internal::member_stableNorm,RealScalar>::Type stableNorm() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const StableNormReturnType stableNorm() const
+    { return StableNormReturnType(_expression()); }
 
 
     /** \returns a row (or column) vector expression of the norm
       * of each column (or row) of the referenced expression, avoiding
       * underflow and overflow using a concatenation of hypot() calls.
+      * This is a vector with real entries, even if the original matrix has complex entries.
       *
       * \sa DenseBase::hypotNorm() */
-    const typename ReturnType<internal::member_hypotNorm,RealScalar>::Type hypotNorm() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const HypotNormReturnType hypotNorm() const
+    { return HypotNormReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression of the sum
       * of each column (or row) of the referenced expression.
@@ -354,39 +394,48 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * Output: \verbinclude PartialRedux_sum.out
       *
       * \sa DenseBase::sum() */
-    const typename ReturnType<internal::member_sum>::Type sum() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const SumReturnType sum() const
+    { return SumReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression of the mean
     * of each column (or row) of the referenced expression.
     *
     * \sa DenseBase::mean() */
-    const typename ReturnType<internal::member_mean>::Type mean() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const MeanReturnType mean() const
+    { return MeanReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression representing
       * whether \b all coefficients of each respective column (or row) are \c true.
+      * This expression can be assigned to a vector with entries of type \c bool.
       *
       * \sa DenseBase::all() */
-    const typename ReturnType<internal::member_all>::Type all() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const AllReturnType all() const
+    { return AllReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression representing
       * whether \b at \b least one coefficient of each respective column (or row) is \c true.
+      * This expression can be assigned to a vector with entries of type \c bool.
       *
       * \sa DenseBase::any() */
-    const typename ReturnType<internal::member_any>::Type any() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const AnyReturnType any() const
+    { return AnyReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression representing
       * the number of \c true coefficients of each respective column (or row).
+      * This expression can be assigned to a vector whose entries have the same type as is used to
+      * index entries of the original matrix; for dense matrices, this is \c std::ptrdiff_t .
       *
       * Example: \include PartialRedux_count.cpp
       * Output: \verbinclude PartialRedux_count.out
       *
       * \sa DenseBase::count() */
-    const PartialReduxExpr<ExpressionType, internal::member_count<Index>, Direction> count() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const CountReturnType count() const
+    { return CountReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression of the product
       * of each column (or row) of the referenced expression.
@@ -395,8 +444,9 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * Output: \verbinclude PartialRedux_prod.out
       *
       * \sa DenseBase::prod() */
-    const typename ReturnType<internal::member_prod>::Type prod() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const ProdReturnType prod() const
+    { return ProdReturnType(_expression()); }
 
 
     /** \returns a matrix expression
@@ -406,10 +456,20 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * Output: \verbinclude Vectorwise_reverse.out
       *
       * \sa DenseBase::reverse() */
-    const Reverse<ExpressionType, Direction> reverse() const
-    { return Reverse<ExpressionType, Direction>( _expression() ); }
+    EIGEN_DEVICE_FUNC
+    const ConstReverseReturnType reverse() const
+    { return ConstReverseReturnType( _expression() ); }
 
-    typedef Replicate<ExpressionType,Direction==Vertical?Dynamic:1,Direction==Horizontal?Dynamic:1> ReplicateReturnType;
+    /** \returns a writable matrix expression
+      * where each column (or row) are reversed.
+      *
+      * \sa reverse() const */
+    EIGEN_DEVICE_FUNC
+    ReverseReturnType reverse()
+    { return ReverseReturnType( _expression() ); }
+
+    typedef Replicate<ExpressionType,(isVertical?Dynamic:1),(isHorizontal?Dynamic:1)> ReplicateReturnType;
+    EIGEN_DEVICE_FUNC
     const ReplicateReturnType replicate(Index factor) const;
 
     /**
@@ -421,17 +481,20 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * \sa VectorwiseOp::replicate(Index), DenseBase::replicate(), class Replicate
       */
     // NOTE implemented here because of sunstudio's compilation errors
-    template<int Factor> const Replicate<ExpressionType,(IsVertical?Factor:1),(IsHorizontal?Factor:1)>
+    // isVertical*Factor+isHorizontal instead of (isVertical?Factor:1) to handle CUDA bug with ternary operator
+    template<int Factor> const Replicate<ExpressionType,isVertical*Factor+isHorizontal,isHorizontal*Factor+isVertical>
+    EIGEN_DEVICE_FUNC
     replicate(Index factor = Factor) const
     {
-      return Replicate<ExpressionType,Direction==Vertical?Factor:1,Direction==Horizontal?Factor:1>
-          (_expression(),Direction==Vertical?factor:1,Direction==Horizontal?factor:1);
+      return Replicate<ExpressionType,(isVertical?Factor:1),(isHorizontal?Factor:1)>
+          (_expression(),isVertical?factor:1,isHorizontal?factor:1);
     }
 
 /////////// Artithmetic operators ///////////
 
     /** Copies the vector \a other to each subvector of \c *this */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     ExpressionType& operator=(const DenseBase<OtherDerived>& other)
     {
       EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
@@ -442,6 +505,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 
     /** Adds the vector \a other to each subvector of \c *this */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     ExpressionType& operator+=(const DenseBase<OtherDerived>& other)
     {
       EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
@@ -451,6 +515,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 
     /** Substracts the vector \a other to each subvector of \c *this */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     ExpressionType& operator-=(const DenseBase<OtherDerived>& other)
     {
       EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
@@ -460,6 +525,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 
     /** Multiples each subvector of \c *this by the vector \a other */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     ExpressionType& operator*=(const DenseBase<OtherDerived>& other)
     {
       EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
@@ -471,6 +537,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 
     /** Divides each subvector of \c *this by the vector \a other */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     ExpressionType& operator/=(const DenseBase<OtherDerived>& other)
     {
       EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
@@ -481,8 +548,8 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
     }
 
     /** Returns the expression of the sum of the vector \a other to each subvector of \c *this */
-    template<typename OtherDerived> EIGEN_STRONG_INLINE
-    CwiseBinaryOp<internal::scalar_sum_op<Scalar>,
+    template<typename OtherDerived> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
+    CwiseBinaryOp<internal::scalar_sum_op<Scalar,typename OtherDerived::Scalar>,
                   const ExpressionTypeNestedCleaned,
                   const typename ExtendedType<OtherDerived>::Type>
     operator+(const DenseBase<OtherDerived>& other) const
@@ -494,7 +561,8 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 
     /** Returns the expression of the difference between each subvector of \c *this and the vector \a other */
     template<typename OtherDerived>
-    CwiseBinaryOp<internal::scalar_difference_op<Scalar>,
+    EIGEN_DEVICE_FUNC
+    CwiseBinaryOp<internal::scalar_difference_op<Scalar,typename OtherDerived::Scalar>,
                   const ExpressionTypeNestedCleaned,
                   const typename ExtendedType<OtherDerived>::Type>
     operator-(const DenseBase<OtherDerived>& other) const
@@ -506,10 +574,11 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 
     /** Returns the expression where each subvector is the product of the vector \a other
       * by the corresponding subvector of \c *this */
-    template<typename OtherDerived> EIGEN_STRONG_INLINE
+    template<typename OtherDerived> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
     CwiseBinaryOp<internal::scalar_product_op<Scalar>,
                   const ExpressionTypeNestedCleaned,
                   const typename ExtendedType<OtherDerived>::Type>
+    EIGEN_DEVICE_FUNC
     operator*(const DenseBase<OtherDerived>& other) const
     {
       EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
@@ -521,6 +590,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
     /** Returns the expression where each subvector is the quotient of the corresponding
       * subvector of \c *this by the vector \a other */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     CwiseBinaryOp<internal::scalar_quotient_op<Scalar>,
                   const ExpressionTypeNestedCleaned,
                   const typename ExtendedType<OtherDerived>::Type>
@@ -531,32 +601,36 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
       return m_matrix / extendedTo(other.derived());
     }
-    
-    /** \returns an expression where each column of row of the referenced matrix are normalized.
+
+    /** \returns an expression where each column (or row) of the referenced matrix are normalized.
       * The referenced matrix is \b not modified.
       * \sa MatrixBase::normalized(), normalize()
       */
+    EIGEN_DEVICE_FUNC
     CwiseBinaryOp<internal::scalar_quotient_op<Scalar>,
                   const ExpressionTypeNestedCleaned,
                   const typename OppositeExtendedType<typename ReturnType<internal::member_norm,RealScalar>::Type>::Type>
     normalized() const { return m_matrix.cwiseQuotient(extendedToOpposite(this->norm())); }
-    
-    
+
+
     /** Normalize in-place each row or columns of the referenced matrix.
       * \sa MatrixBase::normalize(), normalized()
       */
-    void normalize() {
+    EIGEN_DEVICE_FUNC void normalize() {
       m_matrix = this->normalized();
     }
 
+    EIGEN_DEVICE_FUNC inline void reverseInPlace();
+
 /////////// Geometry module ///////////
 
-    #if EIGEN2_SUPPORT_STAGE > STAGE20_RESOLVE_API_CONFLICTS
-    Homogeneous<ExpressionType,Direction> homogeneous() const;
-    #endif
+    typedef Homogeneous<ExpressionType,Direction> HomogeneousReturnType;
+    EIGEN_DEVICE_FUNC
+    HomogeneousReturnType homogeneous() const;
 
     typedef typename ExpressionType::PlainObject CrossReturnType;
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     const CrossReturnType cross(const MatrixBase<OtherDerived>& other) const;
 
     enum {
@@ -581,60 +655,39 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
                   Direction==Horizontal ? HNormalized_SizeMinusOne : 1> >
             HNormalizedReturnType;
 
+    EIGEN_DEVICE_FUNC
     const HNormalizedReturnType hnormalized() const;
 
   protected:
     ExpressionTypeNested m_matrix;
 };
 
-/** \returns a VectorwiseOp wrapper of *this providing additional partial reduction operations
-  *
-  * Example: \include MatrixBase_colwise.cpp
-  * Output: \verbinclude MatrixBase_colwise.out
-  *
-  * \sa rowwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
-  */
-template<typename Derived>
-inline const typename DenseBase<Derived>::ConstColwiseReturnType
-DenseBase<Derived>::colwise() const
-{
-  return derived();
-}
+//const colwise moved to DenseBase.h due to CUDA compiler bug
+
 
 /** \returns a writable VectorwiseOp wrapper of *this providing additional partial reduction operations
   *
   * \sa rowwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
   */
 template<typename Derived>
-inline typename DenseBase<Derived>::ColwiseReturnType
+EIGEN_DEVICE_FUNC inline typename DenseBase<Derived>::ColwiseReturnType
 DenseBase<Derived>::colwise()
 {
-  return derived();
+  return ColwiseReturnType(derived());
 }
 
-/** \returns a VectorwiseOp wrapper of *this providing additional partial reduction operations
-  *
-  * Example: \include MatrixBase_rowwise.cpp
-  * Output: \verbinclude MatrixBase_rowwise.out
-  *
-  * \sa colwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
-  */
-template<typename Derived>
-inline const typename DenseBase<Derived>::ConstRowwiseReturnType
-DenseBase<Derived>::rowwise() const
-{
-  return derived();
-}
+//const rowwise moved to DenseBase.h due to CUDA compiler bug
+
 
 /** \returns a writable VectorwiseOp wrapper of *this providing additional partial reduction operations
   *
   * \sa colwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
   */
 template<typename Derived>
-inline typename DenseBase<Derived>::RowwiseReturnType
+EIGEN_DEVICE_FUNC inline typename DenseBase<Derived>::RowwiseReturnType
 DenseBase<Derived>::rowwise()
 {
-  return derived();
+  return RowwiseReturnType(derived());
 }
 
 } // end namespace Eigen
diff --git a/eigen/Eigen/src/Core/Visitor.h b/eigen/Eigen/src/Core/Visitor.h
index dd94eb8..54c1883 100644
--- a/eigen/Eigen/src/Core/Visitor.h
+++ b/eigen/Eigen/src/Core/Visitor.h
@@ -22,6 +22,7 @@ struct visitor_impl
     row = (UnrollCount-1) % Derived::RowsAtCompileTime
   };
 
+  EIGEN_DEVICE_FUNC
   static inline void run(const Derived &mat, Visitor& visitor)
   {
     visitor_impl<Visitor, Derived, UnrollCount-1>::run(mat, visitor);
@@ -32,6 +33,7 @@ struct visitor_impl
 template<typename Visitor, typename Derived>
 struct visitor_impl<Visitor, Derived, 1>
 {
+  EIGEN_DEVICE_FUNC
   static inline void run(const Derived &mat, Visitor& visitor)
   {
     return visitor.init(mat.coeff(0, 0), 0, 0);
@@ -41,7 +43,7 @@ struct visitor_impl<Visitor, Derived, 1>
 template<typename Visitor, typename Derived>
 struct visitor_impl<Visitor, Derived, Dynamic>
 {
-  typedef typename Derived::Index Index;
+  EIGEN_DEVICE_FUNC
   static inline void run(const Derived& mat, Visitor& visitor)
   {
     visitor.init(mat.coeff(0,0), 0, 0);
@@ -53,6 +55,33 @@ struct visitor_impl<Visitor, Derived, Dynamic>
   }
 };
 
+// evaluator adaptor
+template<typename XprType>
+class visitor_evaluator
+{
+public:
+  EIGEN_DEVICE_FUNC
+  explicit visitor_evaluator(const XprType &xpr) : m_evaluator(xpr), m_xpr(xpr) {}
+  
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  
+  enum {
+    RowsAtCompileTime = XprType::RowsAtCompileTime,
+    CoeffReadCost = internal::evaluator<XprType>::CoeffReadCost
+  };
+  
+  EIGEN_DEVICE_FUNC Index rows() const { return m_xpr.rows(); }
+  EIGEN_DEVICE_FUNC Index cols() const { return m_xpr.cols(); }
+  EIGEN_DEVICE_FUNC Index size() const { return m_xpr.size(); }
+
+  EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index col) const
+  { return m_evaluator.coeff(row, col); }
+  
+protected:
+  internal::evaluator<XprType> m_evaluator;
+  const XprType &m_xpr;
+};
 } // end namespace internal
 
 /** Applies the visitor \a visitor to the whole coefficients of the matrix or vector.
@@ -74,19 +103,17 @@ struct visitor_impl<Visitor, Derived, Dynamic>
   */
 template<typename Derived>
 template<typename Visitor>
+EIGEN_DEVICE_FUNC
 void DenseBase<Derived>::visit(Visitor& visitor) const
 {
-  typedef typename internal::remove_all<typename Derived::Nested>::type ThisNested;
-  typename Derived::Nested thisNested(derived());
-
-  enum { unroll = SizeAtCompileTime != Dynamic
-                   && CoeffReadCost != Dynamic
-                   && (SizeAtCompileTime == 1 || internal::functor_traits<Visitor>::Cost != Dynamic)
-                   && SizeAtCompileTime * CoeffReadCost + (SizeAtCompileTime-1) * internal::functor_traits<Visitor>::Cost
-                      <= EIGEN_UNROLLING_LIMIT };
-  return internal::visitor_impl<Visitor, ThisNested,
-      unroll ? int(SizeAtCompileTime) : Dynamic
-    >::run(thisNested, visitor);
+  typedef typename internal::visitor_evaluator<Derived> ThisEvaluator;
+  ThisEvaluator thisEval(derived());
+  
+  enum {
+    unroll =  SizeAtCompileTime != Dynamic
+           && SizeAtCompileTime * ThisEvaluator::CoeffReadCost + (SizeAtCompileTime-1) * internal::functor_traits<Visitor>::Cost <= EIGEN_UNROLLING_LIMIT
+  };
+  return internal::visitor_impl<Visitor, ThisEvaluator, unroll ? int(SizeAtCompileTime) : Dynamic>::run(thisEval, visitor);
 }
 
 namespace internal {
@@ -97,10 +124,10 @@ namespace internal {
 template <typename Derived>
 struct coeff_visitor
 {
-  typedef typename Derived::Index Index;
   typedef typename Derived::Scalar Scalar;
   Index row, col;
   Scalar res;
+  EIGEN_DEVICE_FUNC
   inline void init(const Scalar& value, Index i, Index j)
   {
     res = value;
@@ -117,8 +144,8 @@ struct coeff_visitor
 template <typename Derived>
 struct min_coeff_visitor : coeff_visitor<Derived>
 {
-  typedef typename Derived::Index Index;
   typedef typename Derived::Scalar Scalar;
+  EIGEN_DEVICE_FUNC
   void operator() (const Scalar& value, Index i, Index j)
   {
     if(value < this->res)
@@ -145,8 +172,8 @@ struct functor_traits<min_coeff_visitor<Scalar> > {
 template <typename Derived>
 struct max_coeff_visitor : coeff_visitor<Derived>
 {
-  typedef typename Derived::Index Index;
-  typedef typename Derived::Scalar Scalar;
+  typedef typename Derived::Scalar Scalar; 
+  EIGEN_DEVICE_FUNC
   void operator() (const Scalar& value, Index i, Index j)
   {
     if(value > this->res)
@@ -167,13 +194,15 @@ struct functor_traits<max_coeff_visitor<Scalar> > {
 
 } // end namespace internal
 
-/** \returns the minimum of all coefficients of *this and puts in *row and *col its location.
+/** \fn DenseBase<Derived>::minCoeff(IndexType* rowId, IndexType* colId) const
+  * \returns the minimum of all coefficients of *this and puts in *row and *col its location.
   * \warning the result is undefined if \c *this contains NaN.
   *
-  * \sa DenseBase::minCoeff(Index*), DenseBase::maxCoeff(Index*,Index*), DenseBase::visitor(), DenseBase::minCoeff()
+  * \sa DenseBase::minCoeff(Index*), DenseBase::maxCoeff(Index*,Index*), DenseBase::visit(), DenseBase::minCoeff()
   */
 template<typename Derived>
 template<typename IndexType>
+EIGEN_DEVICE_FUNC
 typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::minCoeff(IndexType* rowId, IndexType* colId) const
 {
@@ -187,27 +216,30 @@ DenseBase<Derived>::minCoeff(IndexType* rowId, IndexType* colId) const
 /** \returns the minimum of all coefficients of *this and puts in *index its location.
   * \warning the result is undefined if \c *this contains NaN. 
   *
-  * \sa DenseBase::minCoeff(IndexType*,IndexType*), DenseBase::maxCoeff(IndexType*,IndexType*), DenseBase::visitor(), DenseBase::minCoeff()
+  * \sa DenseBase::minCoeff(IndexType*,IndexType*), DenseBase::maxCoeff(IndexType*,IndexType*), DenseBase::visit(), DenseBase::minCoeff()
   */
 template<typename Derived>
 template<typename IndexType>
+EIGEN_DEVICE_FUNC
 typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::minCoeff(IndexType* index) const
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
   internal::min_coeff_visitor<Derived> minVisitor;
   this->visit(minVisitor);
-  *index = (RowsAtCompileTime==1) ? minVisitor.col : minVisitor.row;
+  *index = IndexType((RowsAtCompileTime==1) ? minVisitor.col : minVisitor.row);
   return minVisitor.res;
 }
 
-/** \returns the maximum of all coefficients of *this and puts in *row and *col its location.
+/** \fn DenseBase<Derived>::maxCoeff(IndexType* rowId, IndexType* colId) const
+  * \returns the maximum of all coefficients of *this and puts in *row and *col its location.
   * \warning the result is undefined if \c *this contains NaN. 
   *
-  * \sa DenseBase::minCoeff(IndexType*,IndexType*), DenseBase::visitor(), DenseBase::maxCoeff()
+  * \sa DenseBase::minCoeff(IndexType*,IndexType*), DenseBase::visit(), DenseBase::maxCoeff()
   */
 template<typename Derived>
 template<typename IndexType>
+EIGEN_DEVICE_FUNC
 typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::maxCoeff(IndexType* rowPtr, IndexType* colPtr) const
 {
@@ -225,6 +257,7 @@ DenseBase<Derived>::maxCoeff(IndexType* rowPtr, IndexType* colPtr) const
   */
 template<typename Derived>
 template<typename IndexType>
+EIGEN_DEVICE_FUNC
 typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::maxCoeff(IndexType* index) const
 {
diff --git a/eigen/Eigen/src/Core/arch/AVX/Complex.h b/eigen/Eigen/src/Core/arch/AVX/Complex.h
new file mode 100644
index 0000000..99439c8
--- /dev/null
+++ b/eigen/Eigen/src/Core/arch/AVX/Complex.h
@@ -0,0 +1,483 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner (benoit.steiner.goog@gmail.com)
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_COMPLEX_AVX_H
+#define EIGEN_COMPLEX_AVX_H
+
+namespace Eigen {
+
+namespace internal {
+
+//---------- float ----------
+struct Packet4cf
+{
+  EIGEN_STRONG_INLINE Packet4cf() {}
+  EIGEN_STRONG_INLINE explicit Packet4cf(const __m256& a) : v(a) {}
+  __m256  v;
+};
+
+template<> struct packet_traits<std::complex<float> >  : default_packet_traits
+{
+  typedef Packet4cf type;
+  typedef Packet2cf half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 4,
+    HasHalfPacket = 1,
+
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasDiv    = 1,
+    HasNegate = 1,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasSetLinear = 0
+  };
+};
+
+template<> struct unpacket_traits<Packet4cf> { typedef std::complex<float> type; enum {size=4, alignment=Aligned32}; typedef Packet2cf half; };
+
+template<> EIGEN_STRONG_INLINE Packet4cf padd<Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_add_ps(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet4cf psub<Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_sub_ps(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet4cf pnegate(const Packet4cf& a)
+{
+  return Packet4cf(pnegate(a.v));
+}
+template<> EIGEN_STRONG_INLINE Packet4cf pconj(const Packet4cf& a)
+{
+  const __m256 mask = _mm256_castsi256_ps(_mm256_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000,0x00000000,0x80000000,0x00000000,0x80000000));
+  return Packet4cf(_mm256_xor_ps(a.v,mask));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4cf pmul<Packet4cf>(const Packet4cf& a, const Packet4cf& b)
+{
+  __m256 tmp1 = _mm256_mul_ps(_mm256_moveldup_ps(a.v), b.v);
+  __m256 tmp2 = _mm256_mul_ps(_mm256_movehdup_ps(a.v), _mm256_permute_ps(b.v, _MM_SHUFFLE(2,3,0,1)));
+  __m256 result = _mm256_addsub_ps(tmp1, tmp2);
+  return Packet4cf(result);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4cf pand   <Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_and_ps(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet4cf por    <Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_or_ps(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet4cf pxor   <Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_xor_ps(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet4cf pandnot<Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_andnot_ps(a.v,b.v)); }
+
+template<> EIGEN_STRONG_INLINE Packet4cf pload <Packet4cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet4cf(pload<Packet8f>(&numext::real_ref(*from))); }
+template<> EIGEN_STRONG_INLINE Packet4cf ploadu<Packet4cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet4cf(ploadu<Packet8f>(&numext::real_ref(*from))); }
+
+
+template<> EIGEN_STRONG_INLINE Packet4cf pset1<Packet4cf>(const std::complex<float>& from)
+{
+  return Packet4cf(_mm256_castpd_ps(_mm256_broadcast_sd((const double*)(const void*)&from)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4cf ploaddup<Packet4cf>(const std::complex<float>* from)
+{
+  // FIXME The following might be optimized using _mm256_movedup_pd
+  Packet2cf a = ploaddup<Packet2cf>(from);
+  Packet2cf b = ploaddup<Packet2cf>(from+1);
+  return  Packet4cf(_mm256_insertf128_ps(_mm256_castps128_ps256(a.v), b.v, 1));
+}
+
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float>* to, const Packet4cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore(&numext::real_ref(*to), from.v); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float>* to, const Packet4cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&numext::real_ref(*to), from.v); }
+
+template<> EIGEN_DEVICE_FUNC inline Packet4cf pgather<std::complex<float>, Packet4cf>(const std::complex<float>* from, Index stride)
+{
+  return Packet4cf(_mm256_set_ps(std::imag(from[3*stride]), std::real(from[3*stride]),
+                                 std::imag(from[2*stride]), std::real(from[2*stride]),
+                                 std::imag(from[1*stride]), std::real(from[1*stride]),
+                                 std::imag(from[0*stride]), std::real(from[0*stride])));
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet4cf>(std::complex<float>* to, const Packet4cf& from, Index stride)
+{
+  __m128 low = _mm256_extractf128_ps(from.v, 0);
+  to[stride*0] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(low, low, 0)),
+                                     _mm_cvtss_f32(_mm_shuffle_ps(low, low, 1)));
+  to[stride*1] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(low, low, 2)),
+                                     _mm_cvtss_f32(_mm_shuffle_ps(low, low, 3)));
+
+  __m128 high = _mm256_extractf128_ps(from.v, 1);
+  to[stride*2] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(high, high, 0)),
+                                     _mm_cvtss_f32(_mm_shuffle_ps(high, high, 1)));
+  to[stride*3] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(high, high, 2)),
+                                     _mm_cvtss_f32(_mm_shuffle_ps(high, high, 3)));
+
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<float>  pfirst<Packet4cf>(const Packet4cf& a)
+{
+  return pfirst(Packet2cf(_mm256_castps256_ps128(a.v)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4cf preverse(const Packet4cf& a) {
+  __m128 low  = _mm256_extractf128_ps(a.v, 0);
+  __m128 high = _mm256_extractf128_ps(a.v, 1);
+  __m128d lowd  = _mm_castps_pd(low);
+  __m128d highd = _mm_castps_pd(high);
+  low  = _mm_castpd_ps(_mm_shuffle_pd(lowd,lowd,0x1));
+  high = _mm_castpd_ps(_mm_shuffle_pd(highd,highd,0x1));
+  __m256 result = _mm256_setzero_ps();
+  result = _mm256_insertf128_ps(result, low, 1);
+  result = _mm256_insertf128_ps(result, high, 0);
+  return Packet4cf(result);
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet4cf>(const Packet4cf& a)
+{
+  return predux(padd(Packet2cf(_mm256_extractf128_ps(a.v,0)),
+                     Packet2cf(_mm256_extractf128_ps(a.v,1))));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4cf preduxp<Packet4cf>(const Packet4cf* vecs)
+{
+  Packet8f t0 = _mm256_shuffle_ps(vecs[0].v, vecs[0].v, _MM_SHUFFLE(3, 1, 2 ,0));
+  Packet8f t1 = _mm256_shuffle_ps(vecs[1].v, vecs[1].v, _MM_SHUFFLE(3, 1, 2 ,0));
+  t0 = _mm256_hadd_ps(t0,t1);
+  Packet8f t2 = _mm256_shuffle_ps(vecs[2].v, vecs[2].v, _MM_SHUFFLE(3, 1, 2 ,0));
+  Packet8f t3 = _mm256_shuffle_ps(vecs[3].v, vecs[3].v, _MM_SHUFFLE(3, 1, 2 ,0));
+  t2 = _mm256_hadd_ps(t2,t3);
+  
+  t1 = _mm256_permute2f128_ps(t0,t2, 0 + (2<<4));
+  t3 = _mm256_permute2f128_ps(t0,t2, 1 + (3<<4));
+
+  return Packet4cf(_mm256_add_ps(t1,t3));
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet4cf>(const Packet4cf& a)
+{
+  return predux_mul(pmul(Packet2cf(_mm256_extractf128_ps(a.v, 0)),
+                         Packet2cf(_mm256_extractf128_ps(a.v, 1))));
+}
+
+template<int Offset>
+struct palign_impl<Offset,Packet4cf>
+{
+  static EIGEN_STRONG_INLINE void run(Packet4cf& first, const Packet4cf& second)
+  {
+    if (Offset==0) return;
+    palign_impl<Offset*2,Packet8f>::run(first.v, second.v);
+  }
+};
+
+template<> struct conj_helper<Packet4cf, Packet4cf, false,true>
+{
+  EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet4cf& x, const Packet4cf& y, const Packet4cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet4cf pmul(const Packet4cf& a, const Packet4cf& b) const
+  {
+    return internal::pmul(a, pconj(b));
+  }
+};
+
+template<> struct conj_helper<Packet4cf, Packet4cf, true,false>
+{
+  EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet4cf& x, const Packet4cf& y, const Packet4cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet4cf pmul(const Packet4cf& a, const Packet4cf& b) const
+  {
+    return internal::pmul(pconj(a), b);
+  }
+};
+
+template<> struct conj_helper<Packet4cf, Packet4cf, true,true>
+{
+  EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet4cf& x, const Packet4cf& y, const Packet4cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet4cf pmul(const Packet4cf& a, const Packet4cf& b) const
+  {
+    return pconj(internal::pmul(a, b));
+  }
+};
+
+template<> struct conj_helper<Packet8f, Packet4cf, false,false>
+{
+  EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet8f& x, const Packet4cf& y, const Packet4cf& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Packet4cf pmul(const Packet8f& x, const Packet4cf& y) const
+  { return Packet4cf(Eigen::internal::pmul(x, y.v)); }
+};
+
+template<> struct conj_helper<Packet4cf, Packet8f, false,false>
+{
+  EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet4cf& x, const Packet8f& y, const Packet4cf& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Packet4cf pmul(const Packet4cf& x, const Packet8f& y) const
+  { return Packet4cf(Eigen::internal::pmul(x.v, y)); }
+};
+
+template<> EIGEN_STRONG_INLINE Packet4cf pdiv<Packet4cf>(const Packet4cf& a, const Packet4cf& b)
+{
+  Packet4cf num = pmul(a, pconj(b));
+  __m256 tmp = _mm256_mul_ps(b.v, b.v);
+  __m256 tmp2    = _mm256_shuffle_ps(tmp,tmp,0xB1);
+  __m256 denom = _mm256_add_ps(tmp, tmp2);
+  return Packet4cf(_mm256_div_ps(num.v, denom));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4cf pcplxflip<Packet4cf>(const Packet4cf& x)
+{
+  return Packet4cf(_mm256_shuffle_ps(x.v, x.v, _MM_SHUFFLE(2, 3, 0 ,1)));
+}
+
+//---------- double ----------
+struct Packet2cd
+{
+  EIGEN_STRONG_INLINE Packet2cd() {}
+  EIGEN_STRONG_INLINE explicit Packet2cd(const __m256d& a) : v(a) {}
+  __m256d  v;
+};
+
+template<> struct packet_traits<std::complex<double> >  : default_packet_traits
+{
+  typedef Packet2cd type;
+  typedef Packet1cd half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 0,
+    size = 2,
+    HasHalfPacket = 1,
+
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasDiv    = 1,
+    HasNegate = 1,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasSetLinear = 0
+  };
+};
+
+template<> struct unpacket_traits<Packet2cd> { typedef std::complex<double> type; enum {size=2, alignment=Aligned32}; typedef Packet1cd half; };
+
+template<> EIGEN_STRONG_INLINE Packet2cd padd<Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_add_pd(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cd psub<Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_sub_pd(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cd pnegate(const Packet2cd& a) { return Packet2cd(pnegate(a.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cd pconj(const Packet2cd& a)
+{
+  const __m256d mask = _mm256_castsi256_pd(_mm256_set_epi32(0x80000000,0x0,0x0,0x0,0x80000000,0x0,0x0,0x0));
+  return Packet2cd(_mm256_xor_pd(a.v,mask));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cd pmul<Packet2cd>(const Packet2cd& a, const Packet2cd& b)
+{
+  __m256d tmp1 = _mm256_shuffle_pd(a.v,a.v,0x0);
+  __m256d even = _mm256_mul_pd(tmp1, b.v);
+  __m256d tmp2 = _mm256_shuffle_pd(a.v,a.v,0xF);
+  __m256d tmp3 = _mm256_shuffle_pd(b.v,b.v,0x5);
+  __m256d odd  = _mm256_mul_pd(tmp2, tmp3);
+  return Packet2cd(_mm256_addsub_pd(even, odd));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cd pand   <Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_and_pd(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cd por    <Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_or_pd(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cd pxor   <Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_xor_pd(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cd pandnot<Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_andnot_pd(a.v,b.v)); }
+
+template<> EIGEN_STRONG_INLINE Packet2cd pload <Packet2cd>(const std::complex<double>* from)
+{ EIGEN_DEBUG_ALIGNED_LOAD return Packet2cd(pload<Packet4d>((const double*)from)); }
+template<> EIGEN_STRONG_INLINE Packet2cd ploadu<Packet2cd>(const std::complex<double>* from)
+{ EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cd(ploadu<Packet4d>((const double*)from)); }
+
+template<> EIGEN_STRONG_INLINE Packet2cd pset1<Packet2cd>(const std::complex<double>& from)
+{
+  // in case casting to a __m128d* is really not safe, then we can still fallback to this version: (much slower though)
+//   return Packet2cd(_mm256_loadu2_m128d((const double*)&from,(const double*)&from));
+    return Packet2cd(_mm256_broadcast_pd((const __m128d*)(const void*)&from));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cd ploaddup<Packet2cd>(const std::complex<double>* from) { return pset1<Packet2cd>(*from); }
+
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> *   to, const Packet2cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> *   to, const Packet2cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); }
+
+template<> EIGEN_DEVICE_FUNC inline Packet2cd pgather<std::complex<double>, Packet2cd>(const std::complex<double>* from, Index stride)
+{
+  return Packet2cd(_mm256_set_pd(std::imag(from[1*stride]), std::real(from[1*stride]),
+				 std::imag(from[0*stride]), std::real(from[0*stride])));
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet2cd>(std::complex<double>* to, const Packet2cd& from, Index stride)
+{
+  __m128d low = _mm256_extractf128_pd(from.v, 0);
+  to[stride*0] = std::complex<double>(_mm_cvtsd_f64(low), _mm_cvtsd_f64(_mm_shuffle_pd(low, low, 1)));
+  __m128d high = _mm256_extractf128_pd(from.v, 1);
+  to[stride*1] = std::complex<double>(_mm_cvtsd_f64(high), _mm_cvtsd_f64(_mm_shuffle_pd(high, high, 1)));
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet2cd>(const Packet2cd& a)
+{
+  __m128d low = _mm256_extractf128_pd(a.v, 0);
+  EIGEN_ALIGN16 double res[2];
+  _mm_store_pd(res, low);
+  return std::complex<double>(res[0],res[1]);
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cd preverse(const Packet2cd& a) {
+  __m256d result = _mm256_permute2f128_pd(a.v, a.v, 1);
+  return Packet2cd(result);
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet2cd>(const Packet2cd& a)
+{
+  return predux(padd(Packet1cd(_mm256_extractf128_pd(a.v,0)),
+                     Packet1cd(_mm256_extractf128_pd(a.v,1))));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cd preduxp<Packet2cd>(const Packet2cd* vecs)
+{
+  Packet4d t0 = _mm256_permute2f128_pd(vecs[0].v,vecs[1].v, 0 + (2<<4));
+  Packet4d t1 = _mm256_permute2f128_pd(vecs[0].v,vecs[1].v, 1 + (3<<4));
+
+  return Packet2cd(_mm256_add_pd(t0,t1));
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet2cd>(const Packet2cd& a)
+{
+  return predux(pmul(Packet1cd(_mm256_extractf128_pd(a.v,0)),
+                     Packet1cd(_mm256_extractf128_pd(a.v,1))));
+}
+
+template<int Offset>
+struct palign_impl<Offset,Packet2cd>
+{
+  static EIGEN_STRONG_INLINE void run(Packet2cd& first, const Packet2cd& second)
+  {
+    if (Offset==0) return;
+    palign_impl<Offset*2,Packet4d>::run(first.v, second.v);
+  }
+};
+
+template<> struct conj_helper<Packet2cd, Packet2cd, false,true>
+{
+  EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet2cd& x, const Packet2cd& y, const Packet2cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cd pmul(const Packet2cd& a, const Packet2cd& b) const
+  {
+    return internal::pmul(a, pconj(b));
+  }
+};
+
+template<> struct conj_helper<Packet2cd, Packet2cd, true,false>
+{
+  EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet2cd& x, const Packet2cd& y, const Packet2cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cd pmul(const Packet2cd& a, const Packet2cd& b) const
+  {
+    return internal::pmul(pconj(a), b);
+  }
+};
+
+template<> struct conj_helper<Packet2cd, Packet2cd, true,true>
+{
+  EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet2cd& x, const Packet2cd& y, const Packet2cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cd pmul(const Packet2cd& a, const Packet2cd& b) const
+  {
+    return pconj(internal::pmul(a, b));
+  }
+};
+
+template<> struct conj_helper<Packet4d, Packet2cd, false,false>
+{
+  EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet4d& x, const Packet2cd& y, const Packet2cd& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Packet2cd pmul(const Packet4d& x, const Packet2cd& y) const
+  { return Packet2cd(Eigen::internal::pmul(x, y.v)); }
+};
+
+template<> struct conj_helper<Packet2cd, Packet4d, false,false>
+{
+  EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet2cd& x, const Packet4d& y, const Packet2cd& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Packet2cd pmul(const Packet2cd& x, const Packet4d& y) const
+  { return Packet2cd(Eigen::internal::pmul(x.v, y)); }
+};
+
+template<> EIGEN_STRONG_INLINE Packet2cd pdiv<Packet2cd>(const Packet2cd& a, const Packet2cd& b)
+{
+  Packet2cd num = pmul(a, pconj(b));
+  __m256d tmp = _mm256_mul_pd(b.v, b.v);
+  __m256d denom = _mm256_hadd_pd(tmp, tmp);
+  return Packet2cd(_mm256_div_pd(num.v, denom));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cd pcplxflip<Packet2cd>(const Packet2cd& x)
+{
+  return Packet2cd(_mm256_shuffle_pd(x.v, x.v, 0x5));
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet4cf,4>& kernel) {
+  __m256d P0 = _mm256_castps_pd(kernel.packet[0].v);
+  __m256d P1 = _mm256_castps_pd(kernel.packet[1].v);
+  __m256d P2 = _mm256_castps_pd(kernel.packet[2].v);
+  __m256d P3 = _mm256_castps_pd(kernel.packet[3].v);
+
+  __m256d T0 = _mm256_shuffle_pd(P0, P1, 15);
+  __m256d T1 = _mm256_shuffle_pd(P0, P1, 0);
+  __m256d T2 = _mm256_shuffle_pd(P2, P3, 15);
+  __m256d T3 = _mm256_shuffle_pd(P2, P3, 0);
+
+  kernel.packet[1].v = _mm256_castpd_ps(_mm256_permute2f128_pd(T0, T2, 32));
+  kernel.packet[3].v = _mm256_castpd_ps(_mm256_permute2f128_pd(T0, T2, 49));
+  kernel.packet[0].v = _mm256_castpd_ps(_mm256_permute2f128_pd(T1, T3, 32));
+  kernel.packet[2].v = _mm256_castpd_ps(_mm256_permute2f128_pd(T1, T3, 49));
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet2cd,2>& kernel) {
+  __m256d tmp = _mm256_permute2f128_pd(kernel.packet[0].v, kernel.packet[1].v, 0+(2<<4));
+  kernel.packet[1].v = _mm256_permute2f128_pd(kernel.packet[0].v, kernel.packet[1].v, 1+(3<<4));
+ kernel.packet[0].v = tmp;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4cf pinsertfirst(const Packet4cf& a, std::complex<float> b)
+{
+  return Packet4cf(_mm256_blend_ps(a.v,pset1<Packet4cf>(b).v,1|2));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cd pinsertfirst(const Packet2cd& a, std::complex<double> b)
+{
+  return Packet2cd(_mm256_blend_pd(a.v,pset1<Packet2cd>(b).v,1|2));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4cf pinsertlast(const Packet4cf& a, std::complex<float> b)
+{
+  return Packet4cf(_mm256_blend_ps(a.v,pset1<Packet4cf>(b).v,(1<<7)|(1<<6)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cd pinsertlast(const Packet2cd& a, std::complex<double> b)
+{
+  return Packet2cd(_mm256_blend_pd(a.v,pset1<Packet2cd>(b).v,(1<<3)|(1<<2)));
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_COMPLEX_AVX_H
diff --git a/eigen/Eigen/src/Core/arch/AVX/MathFunctions.h b/eigen/Eigen/src/Core/arch/AVX/MathFunctions.h
new file mode 100644
index 0000000..6af67ce
--- /dev/null
+++ b/eigen/Eigen/src/Core/arch/AVX/MathFunctions.h
@@ -0,0 +1,439 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Pedro Gonnet (pedro.gonnet@gmail.com)
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MATH_FUNCTIONS_AVX_H
+#define EIGEN_MATH_FUNCTIONS_AVX_H
+
+/* The sin, cos, exp, and log functions of this file are loosely derived from
+ * Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/
+ */
+
+namespace Eigen {
+
+namespace internal {
+
+inline Packet8i pshiftleft(Packet8i v, int n)
+{
+#ifdef EIGEN_VECTORIZE_AVX2
+  return _mm256_slli_epi32(v, n);
+#else
+  __m128i lo = _mm_slli_epi32(_mm256_extractf128_si256(v, 0), n);
+  __m128i hi = _mm_slli_epi32(_mm256_extractf128_si256(v, 1), n);
+  return _mm256_insertf128_si256(_mm256_castsi128_si256(lo), (hi), 1);
+#endif
+}
+
+inline Packet8f pshiftright(Packet8f v, int n)
+{
+#ifdef EIGEN_VECTORIZE_AVX2
+  return _mm256_cvtepi32_ps(_mm256_srli_epi32(_mm256_castps_si256(v), n));
+#else
+  __m128i lo = _mm_srli_epi32(_mm256_extractf128_si256(_mm256_castps_si256(v), 0), n);
+  __m128i hi = _mm_srli_epi32(_mm256_extractf128_si256(_mm256_castps_si256(v), 1), n);
+  return _mm256_cvtepi32_ps(_mm256_insertf128_si256(_mm256_castsi128_si256(lo), (hi), 1));
+#endif
+}
+
+// Sine function
+// Computes sin(x) by wrapping x to the interval [-Pi/4,3*Pi/4] and
+// evaluating interpolants in [-Pi/4,Pi/4] or [Pi/4,3*Pi/4]. The interpolants
+// are (anti-)symmetric and thus have only odd/even coefficients
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f
+psin<Packet8f>(const Packet8f& _x) {
+  Packet8f x = _x;
+
+  // Some useful values.
+  _EIGEN_DECLARE_CONST_Packet8i(one, 1);
+  _EIGEN_DECLARE_CONST_Packet8f(one, 1.0f);
+  _EIGEN_DECLARE_CONST_Packet8f(two, 2.0f);
+  _EIGEN_DECLARE_CONST_Packet8f(one_over_four, 0.25f);
+  _EIGEN_DECLARE_CONST_Packet8f(one_over_pi, 3.183098861837907e-01f);
+  _EIGEN_DECLARE_CONST_Packet8f(neg_pi_first, -3.140625000000000e+00f);
+  _EIGEN_DECLARE_CONST_Packet8f(neg_pi_second, -9.670257568359375e-04f);
+  _EIGEN_DECLARE_CONST_Packet8f(neg_pi_third, -6.278329571784980e-07f);
+  _EIGEN_DECLARE_CONST_Packet8f(four_over_pi, 1.273239544735163e+00f);
+
+  // Map x from [-Pi/4,3*Pi/4] to z in [-1,3] and subtract the shifted period.
+  Packet8f z = pmul(x, p8f_one_over_pi);
+  Packet8f shift = _mm256_floor_ps(padd(z, p8f_one_over_four));
+  x = pmadd(shift, p8f_neg_pi_first, x);
+  x = pmadd(shift, p8f_neg_pi_second, x);
+  x = pmadd(shift, p8f_neg_pi_third, x);
+  z = pmul(x, p8f_four_over_pi);
+
+  // Make a mask for the entries that need flipping, i.e. wherever the shift
+  // is odd.
+  Packet8i shift_ints = _mm256_cvtps_epi32(shift);
+  Packet8i shift_isodd = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(shift_ints), _mm256_castsi256_ps(p8i_one)));
+  Packet8i sign_flip_mask = pshiftleft(shift_isodd, 31);
+
+  // Create a mask for which interpolant to use, i.e. if z > 1, then the mask
+  // is set to ones for that entry.
+  Packet8f ival_mask = _mm256_cmp_ps(z, p8f_one, _CMP_GT_OQ);
+
+  // Evaluate the polynomial for the interval [1,3] in z.
+  _EIGEN_DECLARE_CONST_Packet8f(coeff_right_0, 9.999999724233232e-01f);
+  _EIGEN_DECLARE_CONST_Packet8f(coeff_right_2, -3.084242535619928e-01f);
+  _EIGEN_DECLARE_CONST_Packet8f(coeff_right_4, 1.584991525700324e-02f);
+  _EIGEN_DECLARE_CONST_Packet8f(coeff_right_6, -3.188805084631342e-04f);
+  Packet8f z_minus_two = psub(z, p8f_two);
+  Packet8f z_minus_two2 = pmul(z_minus_two, z_minus_two);
+  Packet8f right = pmadd(p8f_coeff_right_6, z_minus_two2, p8f_coeff_right_4);
+  right = pmadd(right, z_minus_two2, p8f_coeff_right_2);
+  right = pmadd(right, z_minus_two2, p8f_coeff_right_0);
+
+  // Evaluate the polynomial for the interval [-1,1] in z.
+  _EIGEN_DECLARE_CONST_Packet8f(coeff_left_1, 7.853981525427295e-01f);
+  _EIGEN_DECLARE_CONST_Packet8f(coeff_left_3, -8.074536727092352e-02f);
+  _EIGEN_DECLARE_CONST_Packet8f(coeff_left_5, 2.489871967827018e-03f);
+  _EIGEN_DECLARE_CONST_Packet8f(coeff_left_7, -3.587725841214251e-05f);
+  Packet8f z2 = pmul(z, z);
+  Packet8f left = pmadd(p8f_coeff_left_7, z2, p8f_coeff_left_5);
+  left = pmadd(left, z2, p8f_coeff_left_3);
+  left = pmadd(left, z2, p8f_coeff_left_1);
+  left = pmul(left, z);
+
+  // Assemble the results, i.e. select the left and right polynomials.
+  left = _mm256_andnot_ps(ival_mask, left);
+  right = _mm256_and_ps(ival_mask, right);
+  Packet8f res = _mm256_or_ps(left, right);
+
+  // Flip the sign on the odd intervals and return the result.
+  res = _mm256_xor_ps(res, _mm256_castsi256_ps(sign_flip_mask));
+  return res;
+}
+
+// Natural logarithm
+// Computes log(x) as log(2^e * m) = C*e + log(m), where the constant C =log(2)
+// and m is in the range [sqrt(1/2),sqrt(2)). In this range, the logarithm can
+// be easily approximated by a polynomial centered on m=1 for stability.
+// TODO(gonnet): Further reduce the interval allowing for lower-degree
+//               polynomial interpolants -> ... -> profit!
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f
+plog<Packet8f>(const Packet8f& _x) {
+  Packet8f x = _x;
+  _EIGEN_DECLARE_CONST_Packet8f(1, 1.0f);
+  _EIGEN_DECLARE_CONST_Packet8f(half, 0.5f);
+  _EIGEN_DECLARE_CONST_Packet8f(126f, 126.0f);
+
+  _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(inv_mant_mask, ~0x7f800000);
+
+  // The smallest non denormalized float number.
+  _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(min_norm_pos, 0x00800000);
+  _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(minus_inf, 0xff800000);
+
+  // Polynomial coefficients.
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_SQRTHF, 0.707106781186547524f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p0, 7.0376836292E-2f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p1, -1.1514610310E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p2, 1.1676998740E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p3, -1.2420140846E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p4, +1.4249322787E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p5, -1.6668057665E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p6, +2.0000714765E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p7, -2.4999993993E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p8, +3.3333331174E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_q1, -2.12194440e-4f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_q2, 0.693359375f);
+
+  Packet8f invalid_mask = _mm256_cmp_ps(x, _mm256_setzero_ps(), _CMP_NGE_UQ); // not greater equal is true if x is NaN
+  Packet8f iszero_mask = _mm256_cmp_ps(x, _mm256_setzero_ps(), _CMP_EQ_OQ);
+
+  // Truncate input values to the minimum positive normal.
+  x = pmax(x, p8f_min_norm_pos);
+
+  Packet8f emm0 = pshiftright(x,23);
+  Packet8f e = _mm256_sub_ps(emm0, p8f_126f);
+
+  // Set the exponents to -1, i.e. x are in the range [0.5,1).
+  x = _mm256_and_ps(x, p8f_inv_mant_mask);
+  x = _mm256_or_ps(x, p8f_half);
+
+  // part2: Shift the inputs from the range [0.5,1) to [sqrt(1/2),sqrt(2))
+  // and shift by -1. The values are then centered around 0, which improves
+  // the stability of the polynomial evaluation.
+  //   if( x < SQRTHF ) {
+  //     e -= 1;
+  //     x = x + x - 1.0;
+  //   } else { x = x - 1.0; }
+  Packet8f mask = _mm256_cmp_ps(x, p8f_cephes_SQRTHF, _CMP_LT_OQ);
+  Packet8f tmp = _mm256_and_ps(x, mask);
+  x = psub(x, p8f_1);
+  e = psub(e, _mm256_and_ps(p8f_1, mask));
+  x = padd(x, tmp);
+
+  Packet8f x2 = pmul(x, x);
+  Packet8f x3 = pmul(x2, x);
+
+  // Evaluate the polynomial approximant of degree 8 in three parts, probably
+  // to improve instruction-level parallelism.
+  Packet8f y, y1, y2;
+  y = pmadd(p8f_cephes_log_p0, x, p8f_cephes_log_p1);
+  y1 = pmadd(p8f_cephes_log_p3, x, p8f_cephes_log_p4);
+  y2 = pmadd(p8f_cephes_log_p6, x, p8f_cephes_log_p7);
+  y = pmadd(y, x, p8f_cephes_log_p2);
+  y1 = pmadd(y1, x, p8f_cephes_log_p5);
+  y2 = pmadd(y2, x, p8f_cephes_log_p8);
+  y = pmadd(y, x3, y1);
+  y = pmadd(y, x3, y2);
+  y = pmul(y, x3);
+
+  // Add the logarithm of the exponent back to the result of the interpolation.
+  y1 = pmul(e, p8f_cephes_log_q1);
+  tmp = pmul(x2, p8f_half);
+  y = padd(y, y1);
+  x = psub(x, tmp);
+  y2 = pmul(e, p8f_cephes_log_q2);
+  x = padd(x, y);
+  x = padd(x, y2);
+
+  // Filter out invalid inputs, i.e. negative arg will be NAN, 0 will be -INF.
+  return _mm256_or_ps(
+      _mm256_andnot_ps(iszero_mask, _mm256_or_ps(x, invalid_mask)),
+      _mm256_and_ps(iszero_mask, p8f_minus_inf));
+}
+
+// Exponential function. Works by writing "x = m*log(2) + r" where
+// "m = floor(x/log(2)+1/2)" and "r" is the remainder. The result is then
+// "exp(x) = 2^m*exp(r)" where exp(r) is in the range [-1,1).
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f
+pexp<Packet8f>(const Packet8f& _x) {
+  _EIGEN_DECLARE_CONST_Packet8f(1, 1.0f);
+  _EIGEN_DECLARE_CONST_Packet8f(half, 0.5f);
+  _EIGEN_DECLARE_CONST_Packet8f(127, 127.0f);
+
+  _EIGEN_DECLARE_CONST_Packet8f(exp_hi, 88.3762626647950f);
+  _EIGEN_DECLARE_CONST_Packet8f(exp_lo, -88.3762626647949f);
+
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_LOG2EF, 1.44269504088896341f);
+
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p0, 1.9875691500E-4f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p1, 1.3981999507E-3f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p2, 8.3334519073E-3f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p3, 4.1665795894E-2f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p4, 1.6666665459E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p5, 5.0000001201E-1f);
+
+  // Clamp x.
+  Packet8f x = pmax(pmin(_x, p8f_exp_hi), p8f_exp_lo);
+
+  // Express exp(x) as exp(m*ln(2) + r), start by extracting
+  // m = floor(x/ln(2) + 0.5).
+  Packet8f m = _mm256_floor_ps(pmadd(x, p8f_cephes_LOG2EF, p8f_half));
+
+// Get r = x - m*ln(2). If no FMA instructions are available, m*ln(2) is
+// subtracted out in two parts, m*C1+m*C2 = m*ln(2), to avoid accumulating
+// truncation errors. Note that we don't use the "pmadd" function here to
+// ensure that a precision-preserving FMA instruction is used.
+#ifdef EIGEN_VECTORIZE_FMA
+  _EIGEN_DECLARE_CONST_Packet8f(nln2, -0.6931471805599453f);
+  Packet8f r = _mm256_fmadd_ps(m, p8f_nln2, x);
+#else
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_C1, 0.693359375f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_C2, -2.12194440e-4f);
+  Packet8f r = psub(x, pmul(m, p8f_cephes_exp_C1));
+  r = psub(r, pmul(m, p8f_cephes_exp_C2));
+#endif
+
+  Packet8f r2 = pmul(r, r);
+
+  // TODO(gonnet): Split into odd/even polynomials and try to exploit
+  //               instruction-level parallelism.
+  Packet8f y = p8f_cephes_exp_p0;
+  y = pmadd(y, r, p8f_cephes_exp_p1);
+  y = pmadd(y, r, p8f_cephes_exp_p2);
+  y = pmadd(y, r, p8f_cephes_exp_p3);
+  y = pmadd(y, r, p8f_cephes_exp_p4);
+  y = pmadd(y, r, p8f_cephes_exp_p5);
+  y = pmadd(y, r2, r);
+  y = padd(y, p8f_1);
+
+  // Build emm0 = 2^m.
+  Packet8i emm0 = _mm256_cvttps_epi32(padd(m, p8f_127));
+  emm0 = pshiftleft(emm0, 23);
+
+  // Return 2^m * exp(r).
+  return pmax(pmul(y, _mm256_castsi256_ps(emm0)), _x);
+}
+
+// Hyperbolic Tangent function.
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f
+ptanh<Packet8f>(const Packet8f& x) {
+  return internal::generic_fast_tanh_float(x);
+}
+
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4d
+pexp<Packet4d>(const Packet4d& _x) {
+  Packet4d x = _x;
+
+  _EIGEN_DECLARE_CONST_Packet4d(1, 1.0);
+  _EIGEN_DECLARE_CONST_Packet4d(2, 2.0);
+  _EIGEN_DECLARE_CONST_Packet4d(half, 0.5);
+
+  _EIGEN_DECLARE_CONST_Packet4d(exp_hi, 709.437);
+  _EIGEN_DECLARE_CONST_Packet4d(exp_lo, -709.436139303);
+
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_LOG2EF, 1.4426950408889634073599);
+
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_p0, 1.26177193074810590878e-4);
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_p1, 3.02994407707441961300e-2);
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_p2, 9.99999999999999999910e-1);
+
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_q0, 3.00198505138664455042e-6);
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_q1, 2.52448340349684104192e-3);
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_q2, 2.27265548208155028766e-1);
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_q3, 2.00000000000000000009e0);
+
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_C1, 0.693145751953125);
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_C2, 1.42860682030941723212e-6);
+  _EIGEN_DECLARE_CONST_Packet4i(1023, 1023);
+
+  Packet4d tmp, fx;
+
+  // clamp x
+  x = pmax(pmin(x, p4d_exp_hi), p4d_exp_lo);
+  // Express exp(x) as exp(g + n*log(2)).
+  fx = pmadd(p4d_cephes_LOG2EF, x, p4d_half);
+
+  // Get the integer modulus of log(2), i.e. the "n" described above.
+  fx = _mm256_floor_pd(fx);
+
+  // Get the remainder modulo log(2), i.e. the "g" described above. Subtract
+  // n*log(2) out in two steps, i.e. n*C1 + n*C2, C1+C2=log2 to get the last
+  // digits right.
+  tmp = pmul(fx, p4d_cephes_exp_C1);
+  Packet4d z = pmul(fx, p4d_cephes_exp_C2);
+  x = psub(x, tmp);
+  x = psub(x, z);
+
+  Packet4d x2 = pmul(x, x);
+
+  // Evaluate the numerator polynomial of the rational interpolant.
+  Packet4d px = p4d_cephes_exp_p0;
+  px = pmadd(px, x2, p4d_cephes_exp_p1);
+  px = pmadd(px, x2, p4d_cephes_exp_p2);
+  px = pmul(px, x);
+
+  // Evaluate the denominator polynomial of the rational interpolant.
+  Packet4d qx = p4d_cephes_exp_q0;
+  qx = pmadd(qx, x2, p4d_cephes_exp_q1);
+  qx = pmadd(qx, x2, p4d_cephes_exp_q2);
+  qx = pmadd(qx, x2, p4d_cephes_exp_q3);
+
+  // I don't really get this bit, copied from the SSE2 routines, so...
+  // TODO(gonnet): Figure out what is going on here, perhaps find a better
+  // rational interpolant?
+  x = _mm256_div_pd(px, psub(qx, px));
+  x = pmadd(p4d_2, x, p4d_1);
+
+  // Build e=2^n by constructing the exponents in a 128-bit vector and
+  // shifting them to where they belong in double-precision values.
+  __m128i emm0 = _mm256_cvtpd_epi32(fx);
+  emm0 = _mm_add_epi32(emm0, p4i_1023);
+  emm0 = _mm_shuffle_epi32(emm0, _MM_SHUFFLE(3, 1, 2, 0));
+  __m128i lo = _mm_slli_epi64(emm0, 52);
+  __m128i hi = _mm_slli_epi64(_mm_srli_epi64(emm0, 32), 52);
+  __m256i e = _mm256_insertf128_si256(_mm256_setzero_si256(), lo, 0);
+  e = _mm256_insertf128_si256(e, hi, 1);
+
+  // Construct the result 2^n * exp(g) = e * x. The max is used to catch
+  // non-finite values in the input.
+  return pmax(pmul(x, _mm256_castsi256_pd(e)), _x);
+}
+
+// Functions for sqrt.
+// The EIGEN_FAST_MATH version uses the _mm_rsqrt_ps approximation and one step
+// of Newton's method, at a cost of 1-2 bits of precision as opposed to the
+// exact solution. It does not handle +inf, or denormalized numbers correctly.
+// The main advantage of this approach is not just speed, but also the fact that
+// it can be inlined and pipelined with other computations, further reducing its
+// effective latency. This is similar to Quake3's fast inverse square root.
+// For detail see here: http://www.beyond3d.com/content/articles/8/
+#if EIGEN_FAST_MATH
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f
+psqrt<Packet8f>(const Packet8f& _x) {
+  Packet8f half = pmul(_x, pset1<Packet8f>(.5f));
+  Packet8f denormal_mask = _mm256_and_ps(
+      _mm256_cmp_ps(_x, pset1<Packet8f>((std::numeric_limits<float>::min)()),
+                    _CMP_LT_OQ),
+      _mm256_cmp_ps(_x, _mm256_setzero_ps(), _CMP_GE_OQ));
+
+  // Compute approximate reciprocal sqrt.
+  Packet8f x = _mm256_rsqrt_ps(_x);
+  // Do a single step of Newton's iteration.
+  x = pmul(x, psub(pset1<Packet8f>(1.5f), pmul(half, pmul(x,x))));
+  // Flush results for denormals to zero.
+  return _mm256_andnot_ps(denormal_mask, pmul(_x,x));
+}
+#else
+template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet8f psqrt<Packet8f>(const Packet8f& x) {
+  return _mm256_sqrt_ps(x);
+}
+#endif
+template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4d psqrt<Packet4d>(const Packet4d& x) {
+  return _mm256_sqrt_pd(x);
+}
+#if EIGEN_FAST_MATH
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet8f prsqrt<Packet8f>(const Packet8f& _x) {
+  _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(inf, 0x7f800000);
+  _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(nan, 0x7fc00000);
+  _EIGEN_DECLARE_CONST_Packet8f(one_point_five, 1.5f);
+  _EIGEN_DECLARE_CONST_Packet8f(minus_half, -0.5f);
+  _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(flt_min, 0x00800000);
+
+  Packet8f neg_half = pmul(_x, p8f_minus_half);
+
+  // select only the inverse sqrt of positive normal inputs (denormals are
+  // flushed to zero and cause infs as well).
+  Packet8f le_zero_mask = _mm256_cmp_ps(_x, p8f_flt_min, _CMP_LT_OQ);
+  Packet8f x = _mm256_andnot_ps(le_zero_mask, _mm256_rsqrt_ps(_x));
+
+  // Fill in NaNs and Infs for the negative/zero entries.
+  Packet8f neg_mask = _mm256_cmp_ps(_x, _mm256_setzero_ps(), _CMP_LT_OQ);
+  Packet8f zero_mask = _mm256_andnot_ps(neg_mask, le_zero_mask);
+  Packet8f infs_and_nans = _mm256_or_ps(_mm256_and_ps(neg_mask, p8f_nan),
+                                        _mm256_and_ps(zero_mask, p8f_inf));
+
+  // Do a single step of Newton's iteration.
+  x = pmul(x, pmadd(neg_half, pmul(x, x), p8f_one_point_five));
+
+  // Insert NaNs and Infs in all the right places.
+  return _mm256_or_ps(x, infs_and_nans);
+}
+
+#else
+template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet8f prsqrt<Packet8f>(const Packet8f& x) {
+  _EIGEN_DECLARE_CONST_Packet8f(one, 1.0f);
+  return _mm256_div_ps(p8f_one, _mm256_sqrt_ps(x));
+}
+#endif
+
+template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4d prsqrt<Packet4d>(const Packet4d& x) {
+  _EIGEN_DECLARE_CONST_Packet4d(one, 1.0);
+  return _mm256_div_pd(p4d_one, _mm256_sqrt_pd(x));
+}
+
+
+}  // end namespace internal
+
+}  // end namespace Eigen
+
+#endif  // EIGEN_MATH_FUNCTIONS_AVX_H
diff --git a/eigen/Eigen/src/Core/arch/AVX/PacketMath.h b/eigen/Eigen/src/Core/arch/AVX/PacketMath.h
new file mode 100644
index 0000000..6362309
--- /dev/null
+++ b/eigen/Eigen/src/Core/arch/AVX/PacketMath.h
@@ -0,0 +1,643 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner (benoit.steiner.goog@gmail.com)
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PACKET_MATH_AVX_H
+#define EIGEN_PACKET_MATH_AVX_H
+
+namespace Eigen {
+
+namespace internal {
+
+#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
+#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8
+#endif
+
+#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
+#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*))
+#endif
+
+#ifdef __FMA__
+#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#endif
+#endif
+
+typedef __m256  Packet8f;
+typedef __m256i Packet8i;
+typedef __m256d Packet4d;
+
+template<> struct is_arithmetic<__m256>  { enum { value = true }; };
+template<> struct is_arithmetic<__m256i> { enum { value = true }; };
+template<> struct is_arithmetic<__m256d> { enum { value = true }; };
+
+#define _EIGEN_DECLARE_CONST_Packet8f(NAME,X) \
+  const Packet8f p8f_##NAME = pset1<Packet8f>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet4d(NAME,X) \
+  const Packet4d p4d_##NAME = pset1<Packet4d>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(NAME,X) \
+  const Packet8f p8f_##NAME = _mm256_castsi256_ps(pset1<Packet8i>(X))
+
+#define _EIGEN_DECLARE_CONST_Packet8i(NAME,X) \
+  const Packet8i p8i_##NAME = pset1<Packet8i>(X)
+
+// Use the packet_traits defined in AVX512/PacketMath.h instead if we're going
+// to leverage AVX512 instructions.
+#ifndef EIGEN_VECTORIZE_AVX512
+template<> struct packet_traits<float>  : default_packet_traits
+{
+  typedef Packet8f type;
+  typedef Packet4f half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=8,
+    HasHalfPacket = 1,
+
+    HasDiv  = 1,
+    HasSin  = EIGEN_FAST_MATH,
+    HasCos  = 0,
+    HasLog  = 1,
+    HasExp  = 1,
+    HasSqrt = 1,
+    HasRsqrt = 1,
+    HasTanh  = EIGEN_FAST_MATH,
+    HasBlend = 1,
+    HasRound = 1,
+    HasFloor = 1,
+    HasCeil = 1
+  };
+};
+template<> struct packet_traits<double> : default_packet_traits
+{
+  typedef Packet4d type;
+  typedef Packet2d half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=4,
+    HasHalfPacket = 1,
+
+    HasDiv  = 1,
+    HasExp  = 1,
+    HasSqrt = 1,
+    HasRsqrt = 1,
+    HasBlend = 1,
+    HasRound = 1,
+    HasFloor = 1,
+    HasCeil = 1
+  };
+};
+#endif
+
+template<> struct scalar_div_cost<float,true> { enum { value = 14 }; };
+template<> struct scalar_div_cost<double,true> { enum { value = 16 }; };
+
+/* Proper support for integers is only provided by AVX2. In the meantime, we'll
+   use SSE instructions and packets to deal with integers.
+template<> struct packet_traits<int>    : default_packet_traits
+{
+  typedef Packet8i type;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=8
+  };
+};
+*/
+
+template<> struct unpacket_traits<Packet8f> { typedef float  type; typedef Packet4f half; enum {size=8, alignment=Aligned32}; };
+template<> struct unpacket_traits<Packet4d> { typedef double type; typedef Packet2d half; enum {size=4, alignment=Aligned32}; };
+template<> struct unpacket_traits<Packet8i> { typedef int    type; typedef Packet4i half; enum {size=8, alignment=Aligned32}; };
+
+template<> EIGEN_STRONG_INLINE Packet8f pset1<Packet8f>(const float&  from) { return _mm256_set1_ps(from); }
+template<> EIGEN_STRONG_INLINE Packet4d pset1<Packet4d>(const double& from) { return _mm256_set1_pd(from); }
+template<> EIGEN_STRONG_INLINE Packet8i pset1<Packet8i>(const int&    from) { return _mm256_set1_epi32(from); }
+
+template<> EIGEN_STRONG_INLINE Packet8f pload1<Packet8f>(const float*  from) { return _mm256_broadcast_ss(from); }
+template<> EIGEN_STRONG_INLINE Packet4d pload1<Packet4d>(const double* from) { return _mm256_broadcast_sd(from); }
+
+template<> EIGEN_STRONG_INLINE Packet8f plset<Packet8f>(const float& a) { return _mm256_add_ps(_mm256_set1_ps(a), _mm256_set_ps(7.0,6.0,5.0,4.0,3.0,2.0,1.0,0.0)); }
+template<> EIGEN_STRONG_INLINE Packet4d plset<Packet4d>(const double& a) { return _mm256_add_pd(_mm256_set1_pd(a), _mm256_set_pd(3.0,2.0,1.0,0.0)); }
+
+template<> EIGEN_STRONG_INLINE Packet8f padd<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_add_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4d padd<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_add_pd(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet8f psub<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_sub_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4d psub<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_sub_pd(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet8f pnegate(const Packet8f& a)
+{
+  return _mm256_sub_ps(_mm256_set1_ps(0.0),a);
+}
+template<> EIGEN_STRONG_INLINE Packet4d pnegate(const Packet4d& a)
+{
+  return _mm256_sub_pd(_mm256_set1_pd(0.0),a);
+}
+
+template<> EIGEN_STRONG_INLINE Packet8f pconj(const Packet8f& a) { return a; }
+template<> EIGEN_STRONG_INLINE Packet4d pconj(const Packet4d& a) { return a; }
+template<> EIGEN_STRONG_INLINE Packet8i pconj(const Packet8i& a) { return a; }
+
+template<> EIGEN_STRONG_INLINE Packet8f pmul<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_mul_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4d pmul<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_mul_pd(a,b); }
+
+
+template<> EIGEN_STRONG_INLINE Packet8f pdiv<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_div_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4d pdiv<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_div_pd(a,b); }
+template<> EIGEN_STRONG_INLINE Packet8i pdiv<Packet8i>(const Packet8i& /*a*/, const Packet8i& /*b*/)
+{ eigen_assert(false && "packet integer division are not supported by AVX");
+  return pset1<Packet8i>(0);
+}
+
+#ifdef __FMA__
+template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f& b, const Packet8f& c) {
+#if ( EIGEN_COMP_GNUC_STRICT || (EIGEN_COMP_CLANG && (EIGEN_COMP_CLANG<308)) )
+  // clang stupidly generates a vfmadd213ps instruction plus some vmovaps on registers,
+  // and gcc stupidly generates a vfmadd132ps instruction,
+  // so let's enforce it to generate a vfmadd231ps instruction since the most common use case is to accumulate
+  // the result of the product.
+  Packet8f res = c;
+  __asm__("vfmadd231ps %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b));
+  return res;
+#else
+  return _mm256_fmadd_ps(a,b,c);
+#endif
+}
+template<> EIGEN_STRONG_INLINE Packet4d pmadd(const Packet4d& a, const Packet4d& b, const Packet4d& c) {
+#if ( EIGEN_COMP_GNUC_STRICT || (EIGEN_COMP_CLANG && (EIGEN_COMP_CLANG<308)) )
+  // see above
+  Packet4d res = c;
+  __asm__("vfmadd231pd %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b));
+  return res;
+#else
+  return _mm256_fmadd_pd(a,b,c);
+#endif
+}
+#endif
+
+template<> EIGEN_STRONG_INLINE Packet8f pmin<Packet8f>(const Packet8f& a, const Packet8f& b) {
+  // Arguments are swapped to match NaN propagation behavior of std::min.
+  return _mm256_min_ps(b,a);
+}
+template<> EIGEN_STRONG_INLINE Packet4d pmin<Packet4d>(const Packet4d& a, const Packet4d& b) {
+  // Arguments are swapped to match NaN propagation behavior of std::min.
+  return _mm256_min_pd(b,a);
+}
+template<> EIGEN_STRONG_INLINE Packet8f pmax<Packet8f>(const Packet8f& a, const Packet8f& b) {
+  // Arguments are swapped to match NaN propagation behavior of std::max.
+  return _mm256_max_ps(b,a);
+}
+template<> EIGEN_STRONG_INLINE Packet4d pmax<Packet4d>(const Packet4d& a, const Packet4d& b) {
+  // Arguments are swapped to match NaN propagation behavior of std::max.
+  return _mm256_max_pd(b,a);
+}
+template<> EIGEN_STRONG_INLINE Packet8f pround<Packet8f>(const Packet8f& a) { return _mm256_round_ps(a, _MM_FROUND_CUR_DIRECTION); }
+template<> EIGEN_STRONG_INLINE Packet4d pround<Packet4d>(const Packet4d& a) { return _mm256_round_pd(a, _MM_FROUND_CUR_DIRECTION); }
+
+template<> EIGEN_STRONG_INLINE Packet8f pceil<Packet8f>(const Packet8f& a) { return _mm256_ceil_ps(a); }
+template<> EIGEN_STRONG_INLINE Packet4d pceil<Packet4d>(const Packet4d& a) { return _mm256_ceil_pd(a); }
+
+template<> EIGEN_STRONG_INLINE Packet8f pfloor<Packet8f>(const Packet8f& a) { return _mm256_floor_ps(a); }
+template<> EIGEN_STRONG_INLINE Packet4d pfloor<Packet4d>(const Packet4d& a) { return _mm256_floor_pd(a); }
+
+template<> EIGEN_STRONG_INLINE Packet8f pand<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_and_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4d pand<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_and_pd(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet8f por<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_or_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4d por<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_or_pd(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet8f pxor<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_xor_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4d pxor<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_xor_pd(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet8f pandnot<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_andnot_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4d pandnot<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_andnot_pd(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet8f pload<Packet8f>(const float*   from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm256_load_ps(from); }
+template<> EIGEN_STRONG_INLINE Packet4d pload<Packet4d>(const double*  from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm256_load_pd(from); }
+template<> EIGEN_STRONG_INLINE Packet8i pload<Packet8i>(const int*     from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm256_load_si256(reinterpret_cast<const __m256i*>(from)); }
+
+template<> EIGEN_STRONG_INLINE Packet8f ploadu<Packet8f>(const float* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm256_loadu_ps(from); }
+template<> EIGEN_STRONG_INLINE Packet4d ploadu<Packet4d>(const double* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm256_loadu_pd(from); }
+template<> EIGEN_STRONG_INLINE Packet8i ploadu<Packet8i>(const int* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm256_loadu_si256(reinterpret_cast<const __m256i*>(from)); }
+
+// Loads 4 floats from memory a returns the packet {a0, a0  a1, a1, a2, a2, a3, a3}
+template<> EIGEN_STRONG_INLINE Packet8f ploaddup<Packet8f>(const float* from)
+{
+  // TODO try to find a way to avoid the need of a temporary register
+//   Packet8f tmp  = _mm256_castps128_ps256(_mm_loadu_ps(from));
+//   tmp = _mm256_insertf128_ps(tmp, _mm_movehl_ps(_mm256_castps256_ps128(tmp),_mm256_castps256_ps128(tmp)), 1);
+//   return _mm256_unpacklo_ps(tmp,tmp);
+
+  // _mm256_insertf128_ps is very slow on Haswell, thus:
+  Packet8f tmp = _mm256_broadcast_ps((const __m128*)(const void*)from);
+  // mimic an "inplace" permutation of the lower 128bits using a blend
+  tmp = _mm256_blend_ps(tmp,_mm256_castps128_ps256(_mm_permute_ps( _mm256_castps256_ps128(tmp), _MM_SHUFFLE(1,0,1,0))), 15);
+  // then we can perform a consistent permutation on the global register to get everything in shape:
+  return  _mm256_permute_ps(tmp, _MM_SHUFFLE(3,3,2,2));
+}
+// Loads 2 doubles from memory a returns the packet {a0, a0  a1, a1}
+template<> EIGEN_STRONG_INLINE Packet4d ploaddup<Packet4d>(const double* from)
+{
+  Packet4d tmp = _mm256_broadcast_pd((const __m128d*)(const void*)from);
+  return  _mm256_permute_pd(tmp, 3<<2);
+}
+
+// Loads 2 floats from memory a returns the packet {a0, a0  a0, a0, a1, a1, a1, a1}
+template<> EIGEN_STRONG_INLINE Packet8f ploadquad<Packet8f>(const float* from)
+{
+  Packet8f tmp = _mm256_castps128_ps256(_mm_broadcast_ss(from));
+  return _mm256_insertf128_ps(tmp, _mm_broadcast_ss(from+1), 1);
+}
+
+template<> EIGEN_STRONG_INLINE void pstore<float>(float*   to, const Packet8f& from) { EIGEN_DEBUG_ALIGNED_STORE _mm256_store_ps(to, from); }
+template<> EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet4d& from) { EIGEN_DEBUG_ALIGNED_STORE _mm256_store_pd(to, from); }
+template<> EIGEN_STRONG_INLINE void pstore<int>(int*       to, const Packet8i& from) { EIGEN_DEBUG_ALIGNED_STORE _mm256_storeu_si256(reinterpret_cast<__m256i*>(to), from); }
+
+template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*   to, const Packet8f& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_ps(to, from); }
+template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet4d& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_pd(to, from); }
+template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*       to, const Packet8i& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_si256(reinterpret_cast<__m256i*>(to), from); }
+
+// NOTE: leverage _mm256_i32gather_ps and _mm256_i32gather_pd if AVX2 instructions are available
+// NOTE: for the record the following seems to be slower: return _mm256_i32gather_ps(from, _mm256_set1_epi32(stride), 4);
+template<> EIGEN_DEVICE_FUNC inline Packet8f pgather<float, Packet8f>(const float* from, Index stride)
+{
+  return _mm256_set_ps(from[7*stride], from[6*stride], from[5*stride], from[4*stride],
+                       from[3*stride], from[2*stride], from[1*stride], from[0*stride]);
+}
+template<> EIGEN_DEVICE_FUNC inline Packet4d pgather<double, Packet4d>(const double* from, Index stride)
+{
+  return _mm256_set_pd(from[3*stride], from[2*stride], from[1*stride], from[0*stride]);
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet8f>(float* to, const Packet8f& from, Index stride)
+{
+  __m128 low = _mm256_extractf128_ps(from, 0);
+  to[stride*0] = _mm_cvtss_f32(low);
+  to[stride*1] = _mm_cvtss_f32(_mm_shuffle_ps(low, low, 1));
+  to[stride*2] = _mm_cvtss_f32(_mm_shuffle_ps(low, low, 2));
+  to[stride*3] = _mm_cvtss_f32(_mm_shuffle_ps(low, low, 3));
+
+  __m128 high = _mm256_extractf128_ps(from, 1);
+  to[stride*4] = _mm_cvtss_f32(high);
+  to[stride*5] = _mm_cvtss_f32(_mm_shuffle_ps(high, high, 1));
+  to[stride*6] = _mm_cvtss_f32(_mm_shuffle_ps(high, high, 2));
+  to[stride*7] = _mm_cvtss_f32(_mm_shuffle_ps(high, high, 3));
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet4d>(double* to, const Packet4d& from, Index stride)
+{
+  __m128d low = _mm256_extractf128_pd(from, 0);
+  to[stride*0] = _mm_cvtsd_f64(low);
+  to[stride*1] = _mm_cvtsd_f64(_mm_shuffle_pd(low, low, 1));
+  __m128d high = _mm256_extractf128_pd(from, 1);
+  to[stride*2] = _mm_cvtsd_f64(high);
+  to[stride*3] = _mm_cvtsd_f64(_mm_shuffle_pd(high, high, 1));
+}
+
+template<> EIGEN_STRONG_INLINE void pstore1<Packet8f>(float* to, const float& a)
+{
+  Packet8f pa = pset1<Packet8f>(a);
+  pstore(to, pa);
+}
+template<> EIGEN_STRONG_INLINE void pstore1<Packet4d>(double* to, const double& a)
+{
+  Packet4d pa = pset1<Packet4d>(a);
+  pstore(to, pa);
+}
+template<> EIGEN_STRONG_INLINE void pstore1<Packet8i>(int* to, const int& a)
+{
+  Packet8i pa = pset1<Packet8i>(a);
+  pstore(to, pa);
+}
+
+#ifndef EIGEN_VECTORIZE_AVX512
+template<> EIGEN_STRONG_INLINE void prefetch<float>(const float*   addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*       addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+#endif
+
+template<> EIGEN_STRONG_INLINE float  pfirst<Packet8f>(const Packet8f& a) {
+  return _mm_cvtss_f32(_mm256_castps256_ps128(a));
+}
+template<> EIGEN_STRONG_INLINE double pfirst<Packet4d>(const Packet4d& a) {
+  return _mm_cvtsd_f64(_mm256_castpd256_pd128(a));
+}
+template<> EIGEN_STRONG_INLINE int    pfirst<Packet8i>(const Packet8i& a) {
+  return _mm_cvtsi128_si32(_mm256_castsi256_si128(a));
+}
+
+
+template<> EIGEN_STRONG_INLINE Packet8f preverse(const Packet8f& a)
+{
+  __m256 tmp = _mm256_shuffle_ps(a,a,0x1b);
+  return _mm256_permute2f128_ps(tmp, tmp, 1);
+}
+template<> EIGEN_STRONG_INLINE Packet4d preverse(const Packet4d& a)
+{
+   __m256d tmp = _mm256_shuffle_pd(a,a,5);
+  return _mm256_permute2f128_pd(tmp, tmp, 1);
+
+  __m256d swap_halves = _mm256_permute2f128_pd(a,a,1);
+    return _mm256_permute_pd(swap_halves,5);
+}
+
+// pabs should be ok
+template<> EIGEN_STRONG_INLINE Packet8f pabs(const Packet8f& a)
+{
+  const Packet8f mask = _mm256_castsi256_ps(_mm256_setr_epi32(0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF));
+  return _mm256_and_ps(a,mask);
+}
+template<> EIGEN_STRONG_INLINE Packet4d pabs(const Packet4d& a)
+{
+  const Packet4d mask = _mm256_castsi256_pd(_mm256_setr_epi32(0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF));
+  return _mm256_and_pd(a,mask);
+}
+
+// preduxp should be ok
+// FIXME: why is this ok? why isn't the simply implementation working as expected?
+template<> EIGEN_STRONG_INLINE Packet8f preduxp<Packet8f>(const Packet8f* vecs)
+{
+    __m256 hsum1 = _mm256_hadd_ps(vecs[0], vecs[1]);
+    __m256 hsum2 = _mm256_hadd_ps(vecs[2], vecs[3]);
+    __m256 hsum3 = _mm256_hadd_ps(vecs[4], vecs[5]);
+    __m256 hsum4 = _mm256_hadd_ps(vecs[6], vecs[7]);
+
+    __m256 hsum5 = _mm256_hadd_ps(hsum1, hsum1);
+    __m256 hsum6 = _mm256_hadd_ps(hsum2, hsum2);
+    __m256 hsum7 = _mm256_hadd_ps(hsum3, hsum3);
+    __m256 hsum8 = _mm256_hadd_ps(hsum4, hsum4);
+
+    __m256 perm1 =  _mm256_permute2f128_ps(hsum5, hsum5, 0x23);
+    __m256 perm2 =  _mm256_permute2f128_ps(hsum6, hsum6, 0x23);
+    __m256 perm3 =  _mm256_permute2f128_ps(hsum7, hsum7, 0x23);
+    __m256 perm4 =  _mm256_permute2f128_ps(hsum8, hsum8, 0x23);
+
+    __m256 sum1 = _mm256_add_ps(perm1, hsum5);
+    __m256 sum2 = _mm256_add_ps(perm2, hsum6);
+    __m256 sum3 = _mm256_add_ps(perm3, hsum7);
+    __m256 sum4 = _mm256_add_ps(perm4, hsum8);
+
+    __m256 blend1 = _mm256_blend_ps(sum1, sum2, 0xcc);
+    __m256 blend2 = _mm256_blend_ps(sum3, sum4, 0xcc);
+
+    __m256 final = _mm256_blend_ps(blend1, blend2, 0xf0);
+    return final;
+}
+template<> EIGEN_STRONG_INLINE Packet4d preduxp<Packet4d>(const Packet4d* vecs)
+{
+ Packet4d tmp0, tmp1;
+
+  tmp0 = _mm256_hadd_pd(vecs[0], vecs[1]);
+  tmp0 = _mm256_add_pd(tmp0, _mm256_permute2f128_pd(tmp0, tmp0, 1));
+
+  tmp1 = _mm256_hadd_pd(vecs[2], vecs[3]);
+  tmp1 = _mm256_add_pd(tmp1, _mm256_permute2f128_pd(tmp1, tmp1, 1));
+
+  return _mm256_blend_pd(tmp0, tmp1, 0xC);
+}
+
+template<> EIGEN_STRONG_INLINE float predux<Packet8f>(const Packet8f& a)
+{
+  return predux(Packet4f(_mm_add_ps(_mm256_castps256_ps128(a),_mm256_extractf128_ps(a,1))));
+}
+template<> EIGEN_STRONG_INLINE double predux<Packet4d>(const Packet4d& a)
+{
+  return predux(Packet2d(_mm_add_pd(_mm256_castpd256_pd128(a),_mm256_extractf128_pd(a,1))));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f predux_downto4<Packet8f>(const Packet8f& a)
+{
+  return _mm_add_ps(_mm256_castps256_ps128(a),_mm256_extractf128_ps(a,1));
+}
+
+template<> EIGEN_STRONG_INLINE float predux_mul<Packet8f>(const Packet8f& a)
+{
+  Packet8f tmp;
+  tmp = _mm256_mul_ps(a, _mm256_permute2f128_ps(a,a,1));
+  tmp = _mm256_mul_ps(tmp, _mm256_shuffle_ps(tmp,tmp,_MM_SHUFFLE(1,0,3,2)));
+  return pfirst(_mm256_mul_ps(tmp, _mm256_shuffle_ps(tmp,tmp,1)));
+}
+template<> EIGEN_STRONG_INLINE double predux_mul<Packet4d>(const Packet4d& a)
+{
+  Packet4d tmp;
+  tmp = _mm256_mul_pd(a, _mm256_permute2f128_pd(a,a,1));
+  return pfirst(_mm256_mul_pd(tmp, _mm256_shuffle_pd(tmp,tmp,1)));
+}
+
+template<> EIGEN_STRONG_INLINE float predux_min<Packet8f>(const Packet8f& a)
+{
+  Packet8f tmp = _mm256_min_ps(a, _mm256_permute2f128_ps(a,a,1));
+  tmp = _mm256_min_ps(tmp, _mm256_shuffle_ps(tmp,tmp,_MM_SHUFFLE(1,0,3,2)));
+  return pfirst(_mm256_min_ps(tmp, _mm256_shuffle_ps(tmp,tmp,1)));
+}
+template<> EIGEN_STRONG_INLINE double predux_min<Packet4d>(const Packet4d& a)
+{
+  Packet4d tmp = _mm256_min_pd(a, _mm256_permute2f128_pd(a,a,1));
+  return pfirst(_mm256_min_pd(tmp, _mm256_shuffle_pd(tmp, tmp, 1)));
+}
+
+template<> EIGEN_STRONG_INLINE float predux_max<Packet8f>(const Packet8f& a)
+{
+  Packet8f tmp = _mm256_max_ps(a, _mm256_permute2f128_ps(a,a,1));
+  tmp = _mm256_max_ps(tmp, _mm256_shuffle_ps(tmp,tmp,_MM_SHUFFLE(1,0,3,2)));
+  return pfirst(_mm256_max_ps(tmp, _mm256_shuffle_ps(tmp,tmp,1)));
+}
+
+template<> EIGEN_STRONG_INLINE double predux_max<Packet4d>(const Packet4d& a)
+{
+  Packet4d tmp = _mm256_max_pd(a, _mm256_permute2f128_pd(a,a,1));
+  return pfirst(_mm256_max_pd(tmp, _mm256_shuffle_pd(tmp, tmp, 1)));
+}
+
+
+template<int Offset>
+struct palign_impl<Offset,Packet8f>
+{
+  static EIGEN_STRONG_INLINE void run(Packet8f& first, const Packet8f& second)
+  {
+    if (Offset==1)
+    {
+      first = _mm256_blend_ps(first, second, 1);
+      Packet8f tmp1 = _mm256_permute_ps (first, _MM_SHUFFLE(0,3,2,1));
+      Packet8f tmp2 = _mm256_permute2f128_ps (tmp1, tmp1, 1);
+      first = _mm256_blend_ps(tmp1, tmp2, 0x88);
+    }
+    else if (Offset==2)
+    {
+      first = _mm256_blend_ps(first, second, 3);
+      Packet8f tmp1 = _mm256_permute_ps (first, _MM_SHUFFLE(1,0,3,2));
+      Packet8f tmp2 = _mm256_permute2f128_ps (tmp1, tmp1, 1);
+      first = _mm256_blend_ps(tmp1, tmp2, 0xcc);
+    }
+    else if (Offset==3)
+    {
+      first = _mm256_blend_ps(first, second, 7);
+      Packet8f tmp1 = _mm256_permute_ps (first, _MM_SHUFFLE(2,1,0,3));
+      Packet8f tmp2 = _mm256_permute2f128_ps (tmp1, tmp1, 1);
+      first = _mm256_blend_ps(tmp1, tmp2, 0xee);
+    }
+    else if (Offset==4)
+    {
+      first = _mm256_blend_ps(first, second, 15);
+      Packet8f tmp1 = _mm256_permute_ps (first, _MM_SHUFFLE(3,2,1,0));
+      Packet8f tmp2 = _mm256_permute2f128_ps (tmp1, tmp1, 1);
+      first = _mm256_permute_ps(tmp2, _MM_SHUFFLE(3,2,1,0));
+    }
+    else if (Offset==5)
+    {
+      first = _mm256_blend_ps(first, second, 31);
+      first = _mm256_permute2f128_ps(first, first, 1);
+      Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(0,3,2,1));
+      first = _mm256_permute2f128_ps(tmp, tmp, 1);
+      first = _mm256_blend_ps(tmp, first, 0x88);
+    }
+    else if (Offset==6)
+    {
+      first = _mm256_blend_ps(first, second, 63);
+      first = _mm256_permute2f128_ps(first, first, 1);
+      Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(1,0,3,2));
+      first = _mm256_permute2f128_ps(tmp, tmp, 1);
+      first = _mm256_blend_ps(tmp, first, 0xcc);
+    }
+    else if (Offset==7)
+    {
+      first = _mm256_blend_ps(first, second, 127);
+      first = _mm256_permute2f128_ps(first, first, 1);
+      Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(2,1,0,3));
+      first = _mm256_permute2f128_ps(tmp, tmp, 1);
+      first = _mm256_blend_ps(tmp, first, 0xee);
+    }
+  }
+};
+
+template<int Offset>
+struct palign_impl<Offset,Packet4d>
+{
+  static EIGEN_STRONG_INLINE void run(Packet4d& first, const Packet4d& second)
+  {
+    if (Offset==1)
+    {
+      first = _mm256_blend_pd(first, second, 1);
+      __m256d tmp = _mm256_permute_pd(first, 5);
+      first = _mm256_permute2f128_pd(tmp, tmp, 1);
+      first = _mm256_blend_pd(tmp, first, 0xA);
+    }
+    else if (Offset==2)
+    {
+      first = _mm256_blend_pd(first, second, 3);
+      first = _mm256_permute2f128_pd(first, first, 1);
+    }
+    else if (Offset==3)
+    {
+      first = _mm256_blend_pd(first, second, 7);
+      __m256d tmp = _mm256_permute_pd(first, 5);
+      first = _mm256_permute2f128_pd(tmp, tmp, 1);
+      first = _mm256_blend_pd(tmp, first, 5);
+    }
+  }
+};
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet8f,8>& kernel) {
+  __m256 T0 = _mm256_unpacklo_ps(kernel.packet[0], kernel.packet[1]);
+  __m256 T1 = _mm256_unpackhi_ps(kernel.packet[0], kernel.packet[1]);
+  __m256 T2 = _mm256_unpacklo_ps(kernel.packet[2], kernel.packet[3]);
+  __m256 T3 = _mm256_unpackhi_ps(kernel.packet[2], kernel.packet[3]);
+  __m256 T4 = _mm256_unpacklo_ps(kernel.packet[4], kernel.packet[5]);
+  __m256 T5 = _mm256_unpackhi_ps(kernel.packet[4], kernel.packet[5]);
+  __m256 T6 = _mm256_unpacklo_ps(kernel.packet[6], kernel.packet[7]);
+  __m256 T7 = _mm256_unpackhi_ps(kernel.packet[6], kernel.packet[7]);
+  __m256 S0 = _mm256_shuffle_ps(T0,T2,_MM_SHUFFLE(1,0,1,0));
+  __m256 S1 = _mm256_shuffle_ps(T0,T2,_MM_SHUFFLE(3,2,3,2));
+  __m256 S2 = _mm256_shuffle_ps(T1,T3,_MM_SHUFFLE(1,0,1,0));
+  __m256 S3 = _mm256_shuffle_ps(T1,T3,_MM_SHUFFLE(3,2,3,2));
+  __m256 S4 = _mm256_shuffle_ps(T4,T6,_MM_SHUFFLE(1,0,1,0));
+  __m256 S5 = _mm256_shuffle_ps(T4,T6,_MM_SHUFFLE(3,2,3,2));
+  __m256 S6 = _mm256_shuffle_ps(T5,T7,_MM_SHUFFLE(1,0,1,0));
+  __m256 S7 = _mm256_shuffle_ps(T5,T7,_MM_SHUFFLE(3,2,3,2));
+  kernel.packet[0] = _mm256_permute2f128_ps(S0, S4, 0x20);
+  kernel.packet[1] = _mm256_permute2f128_ps(S1, S5, 0x20);
+  kernel.packet[2] = _mm256_permute2f128_ps(S2, S6, 0x20);
+  kernel.packet[3] = _mm256_permute2f128_ps(S3, S7, 0x20);
+  kernel.packet[4] = _mm256_permute2f128_ps(S0, S4, 0x31);
+  kernel.packet[5] = _mm256_permute2f128_ps(S1, S5, 0x31);
+  kernel.packet[6] = _mm256_permute2f128_ps(S2, S6, 0x31);
+  kernel.packet[7] = _mm256_permute2f128_ps(S3, S7, 0x31);
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet8f,4>& kernel) {
+  __m256 T0 = _mm256_unpacklo_ps(kernel.packet[0], kernel.packet[1]);
+  __m256 T1 = _mm256_unpackhi_ps(kernel.packet[0], kernel.packet[1]);
+  __m256 T2 = _mm256_unpacklo_ps(kernel.packet[2], kernel.packet[3]);
+  __m256 T3 = _mm256_unpackhi_ps(kernel.packet[2], kernel.packet[3]);
+
+  __m256 S0 = _mm256_shuffle_ps(T0,T2,_MM_SHUFFLE(1,0,1,0));
+  __m256 S1 = _mm256_shuffle_ps(T0,T2,_MM_SHUFFLE(3,2,3,2));
+  __m256 S2 = _mm256_shuffle_ps(T1,T3,_MM_SHUFFLE(1,0,1,0));
+  __m256 S3 = _mm256_shuffle_ps(T1,T3,_MM_SHUFFLE(3,2,3,2));
+
+  kernel.packet[0] = _mm256_permute2f128_ps(S0, S1, 0x20);
+  kernel.packet[1] = _mm256_permute2f128_ps(S2, S3, 0x20);
+  kernel.packet[2] = _mm256_permute2f128_ps(S0, S1, 0x31);
+  kernel.packet[3] = _mm256_permute2f128_ps(S2, S3, 0x31);
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet4d,4>& kernel) {
+  __m256d T0 = _mm256_shuffle_pd(kernel.packet[0], kernel.packet[1], 15);
+  __m256d T1 = _mm256_shuffle_pd(kernel.packet[0], kernel.packet[1], 0);
+  __m256d T2 = _mm256_shuffle_pd(kernel.packet[2], kernel.packet[3], 15);
+  __m256d T3 = _mm256_shuffle_pd(kernel.packet[2], kernel.packet[3], 0);
+
+  kernel.packet[1] = _mm256_permute2f128_pd(T0, T2, 32);
+  kernel.packet[3] = _mm256_permute2f128_pd(T0, T2, 49);
+  kernel.packet[0] = _mm256_permute2f128_pd(T1, T3, 32);
+  kernel.packet[2] = _mm256_permute2f128_pd(T1, T3, 49);
+}
+
+template<> EIGEN_STRONG_INLINE Packet8f pblend(const Selector<8>& ifPacket, const Packet8f& thenPacket, const Packet8f& elsePacket) {
+  const __m256 zero = _mm256_setzero_ps();
+  const __m256 select = _mm256_set_ps(ifPacket.select[7], ifPacket.select[6], ifPacket.select[5], ifPacket.select[4], ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]);
+  __m256 false_mask = _mm256_cmp_ps(select, zero, _CMP_EQ_UQ);
+  return _mm256_blendv_ps(thenPacket, elsePacket, false_mask);
+}
+template<> EIGEN_STRONG_INLINE Packet4d pblend(const Selector<4>& ifPacket, const Packet4d& thenPacket, const Packet4d& elsePacket) {
+  const __m256d zero = _mm256_setzero_pd();
+  const __m256d select = _mm256_set_pd(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]);
+  __m256d false_mask = _mm256_cmp_pd(select, zero, _CMP_EQ_UQ);
+  return _mm256_blendv_pd(thenPacket, elsePacket, false_mask);
+}
+
+template<> EIGEN_STRONG_INLINE Packet8f pinsertfirst(const Packet8f& a, float b)
+{
+  return _mm256_blend_ps(a,pset1<Packet8f>(b),1);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4d pinsertfirst(const Packet4d& a, double b)
+{
+  return _mm256_blend_pd(a,pset1<Packet4d>(b),1);
+}
+
+template<> EIGEN_STRONG_INLINE Packet8f pinsertlast(const Packet8f& a, float b)
+{
+  return _mm256_blend_ps(a,pset1<Packet8f>(b),(1<<7));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4d pinsertlast(const Packet4d& a, double b)
+{
+  return _mm256_blend_pd(a,pset1<Packet4d>(b),(1<<3));
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_PACKET_MATH_AVX_H
diff --git a/eigen/Eigen/src/Core/arch/AVX/TypeCasting.h b/eigen/Eigen/src/Core/arch/AVX/TypeCasting.h
new file mode 100644
index 0000000..83bfdc6
--- /dev/null
+++ b/eigen/Eigen/src/Core/arch/AVX/TypeCasting.h
@@ -0,0 +1,51 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_TYPE_CASTING_AVX_H
+#define EIGEN_TYPE_CASTING_AVX_H
+
+namespace Eigen {
+
+namespace internal {
+
+// For now we use SSE to handle integers, so we can't use AVX instructions to cast
+// from int to float
+template <>
+struct type_casting_traits<float, int> {
+  enum {
+    VectorizedCast = 0,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template <>
+struct type_casting_traits<int, float> {
+  enum {
+    VectorizedCast = 0,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+
+
+template<> EIGEN_STRONG_INLINE Packet8i pcast<Packet8f, Packet8i>(const Packet8f& a) {
+  return _mm256_cvtps_epi32(a);
+}
+
+template<> EIGEN_STRONG_INLINE Packet8f pcast<Packet8i, Packet8f>(const Packet8i& a) {
+  return _mm256_cvtepi32_ps(a);
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TYPE_CASTING_AVX_H
diff --git a/eigen/Eigen/src/Core/arch/AVX512/MathFunctions.h b/eigen/Eigen/src/Core/arch/AVX512/MathFunctions.h
new file mode 100644
index 0000000..399be0e
--- /dev/null
+++ b/eigen/Eigen/src/Core/arch/AVX512/MathFunctions.h
@@ -0,0 +1,396 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Pedro Gonnet (pedro.gonnet@gmail.com)
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef THIRD_PARTY_EIGEN3_EIGEN_SRC_CORE_ARCH_AVX512_MATHFUNCTIONS_H_
+#define THIRD_PARTY_EIGEN3_EIGEN_SRC_CORE_ARCH_AVX512_MATHFUNCTIONS_H_
+
+namespace Eigen {
+
+namespace internal {
+
+// Disable the code for older versions of gcc that don't support many of the required avx512 instrinsics.
+#if EIGEN_GNUC_AT_LEAST(5, 3)
+
+#define _EIGEN_DECLARE_CONST_Packet16f(NAME, X) \
+  const Packet16f p16f_##NAME = pset1<Packet16f>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(NAME, X) \
+  const Packet16f p16f_##NAME = (__m512)pset1<Packet16i>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet8d(NAME, X) \
+  const Packet8d p8d_##NAME = pset1<Packet8d>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet8d_FROM_INT64(NAME, X) \
+  const Packet8d p8d_##NAME = _mm512_castsi512_pd(_mm512_set1_epi64(X))
+
+// Natural logarithm
+// Computes log(x) as log(2^e * m) = C*e + log(m), where the constant C =log(2)
+// and m is in the range [sqrt(1/2),sqrt(2)). In this range, the logarithm can
+// be easily approximated by a polynomial centered on m=1 for stability.
+#if defined(EIGEN_VECTORIZE_AVX512DQ)
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet16f
+plog<Packet16f>(const Packet16f& _x) {
+  Packet16f x = _x;
+  _EIGEN_DECLARE_CONST_Packet16f(1, 1.0f);
+  _EIGEN_DECLARE_CONST_Packet16f(half, 0.5f);
+  _EIGEN_DECLARE_CONST_Packet16f(126f, 126.0f);
+
+  _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(inv_mant_mask, ~0x7f800000);
+
+  // The smallest non denormalized float number.
+  _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(min_norm_pos, 0x00800000);
+  _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(minus_inf, 0xff800000);
+  _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(nan, 0x7fc00000);
+
+  // Polynomial coefficients.
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_SQRTHF, 0.707106781186547524f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p0, 7.0376836292E-2f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p1, -1.1514610310E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p2, 1.1676998740E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p3, -1.2420140846E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p4, +1.4249322787E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p5, -1.6668057665E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p6, +2.0000714765E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p7, -2.4999993993E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p8, +3.3333331174E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_q1, -2.12194440e-4f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_q2, 0.693359375f);
+
+  // invalid_mask is set to true when x is NaN
+  __mmask16 invalid_mask =
+      _mm512_cmp_ps_mask(x, _mm512_setzero_ps(), _CMP_NGE_UQ);
+  __mmask16 iszero_mask =
+      _mm512_cmp_ps_mask(x, _mm512_setzero_ps(), _CMP_EQ_UQ);
+
+  // Truncate input values to the minimum positive normal.
+  x = pmax(x, p16f_min_norm_pos);
+
+  // Extract the shifted exponents.
+  Packet16f emm0 = _mm512_cvtepi32_ps(_mm512_srli_epi32((__m512i)x, 23));
+  Packet16f e = _mm512_sub_ps(emm0, p16f_126f);
+
+  // Set the exponents to -1, i.e. x are in the range [0.5,1).
+  x = _mm512_and_ps(x, p16f_inv_mant_mask);
+  x = _mm512_or_ps(x, p16f_half);
+
+  // part2: Shift the inputs from the range [0.5,1) to [sqrt(1/2),sqrt(2))
+  // and shift by -1. The values are then centered around 0, which improves
+  // the stability of the polynomial evaluation.
+  //   if( x < SQRTHF ) {
+  //     e -= 1;
+  //     x = x + x - 1.0;
+  //   } else { x = x - 1.0; }
+  __mmask16 mask = _mm512_cmp_ps_mask(x, p16f_cephes_SQRTHF, _CMP_LT_OQ);
+  Packet16f tmp = _mm512_mask_blend_ps(mask, x, _mm512_setzero_ps());
+  x = psub(x, p16f_1);
+  e = psub(e, _mm512_mask_blend_ps(mask, p16f_1, _mm512_setzero_ps()));
+  x = padd(x, tmp);
+
+  Packet16f x2 = pmul(x, x);
+  Packet16f x3 = pmul(x2, x);
+
+  // Evaluate the polynomial approximant of degree 8 in three parts, probably
+  // to improve instruction-level parallelism.
+  Packet16f y, y1, y2;
+  y = pmadd(p16f_cephes_log_p0, x, p16f_cephes_log_p1);
+  y1 = pmadd(p16f_cephes_log_p3, x, p16f_cephes_log_p4);
+  y2 = pmadd(p16f_cephes_log_p6, x, p16f_cephes_log_p7);
+  y = pmadd(y, x, p16f_cephes_log_p2);
+  y1 = pmadd(y1, x, p16f_cephes_log_p5);
+  y2 = pmadd(y2, x, p16f_cephes_log_p8);
+  y = pmadd(y, x3, y1);
+  y = pmadd(y, x3, y2);
+  y = pmul(y, x3);
+
+  // Add the logarithm of the exponent back to the result of the interpolation.
+  y1 = pmul(e, p16f_cephes_log_q1);
+  tmp = pmul(x2, p16f_half);
+  y = padd(y, y1);
+  x = psub(x, tmp);
+  y2 = pmul(e, p16f_cephes_log_q2);
+  x = padd(x, y);
+  x = padd(x, y2);
+
+  // Filter out invalid inputs, i.e. negative arg will be NAN, 0 will be -INF.
+  return _mm512_mask_blend_ps(iszero_mask, p16f_minus_inf,
+                              _mm512_mask_blend_ps(invalid_mask, p16f_nan, x));
+}
+#endif
+
+// Exponential function. Works by writing "x = m*log(2) + r" where
+// "m = floor(x/log(2)+1/2)" and "r" is the remainder. The result is then
+// "exp(x) = 2^m*exp(r)" where exp(r) is in the range [-1,1).
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet16f
+pexp<Packet16f>(const Packet16f& _x) {
+  _EIGEN_DECLARE_CONST_Packet16f(1, 1.0f);
+  _EIGEN_DECLARE_CONST_Packet16f(half, 0.5f);
+  _EIGEN_DECLARE_CONST_Packet16f(127, 127.0f);
+
+  _EIGEN_DECLARE_CONST_Packet16f(exp_hi, 88.3762626647950f);
+  _EIGEN_DECLARE_CONST_Packet16f(exp_lo, -88.3762626647949f);
+
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_LOG2EF, 1.44269504088896341f);
+
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p0, 1.9875691500E-4f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p1, 1.3981999507E-3f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p2, 8.3334519073E-3f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p3, 4.1665795894E-2f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p4, 1.6666665459E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p5, 5.0000001201E-1f);
+
+  // Clamp x.
+  Packet16f x = pmax(pmin(_x, p16f_exp_hi), p16f_exp_lo);
+
+  // Express exp(x) as exp(m*ln(2) + r), start by extracting
+  // m = floor(x/ln(2) + 0.5).
+  Packet16f m = _mm512_floor_ps(pmadd(x, p16f_cephes_LOG2EF, p16f_half));
+
+  // Get r = x - m*ln(2). Note that we can do this without losing more than one
+  // ulp precision due to the FMA instruction.
+  _EIGEN_DECLARE_CONST_Packet16f(nln2, -0.6931471805599453f);
+  Packet16f r = _mm512_fmadd_ps(m, p16f_nln2, x);
+  Packet16f r2 = pmul(r, r);
+
+  // TODO(gonnet): Split into odd/even polynomials and try to exploit
+  //               instruction-level parallelism.
+  Packet16f y = p16f_cephes_exp_p0;
+  y = pmadd(y, r, p16f_cephes_exp_p1);
+  y = pmadd(y, r, p16f_cephes_exp_p2);
+  y = pmadd(y, r, p16f_cephes_exp_p3);
+  y = pmadd(y, r, p16f_cephes_exp_p4);
+  y = pmadd(y, r, p16f_cephes_exp_p5);
+  y = pmadd(y, r2, r);
+  y = padd(y, p16f_1);
+
+  // Build emm0 = 2^m.
+  Packet16i emm0 = _mm512_cvttps_epi32(padd(m, p16f_127));
+  emm0 = _mm512_slli_epi32(emm0, 23);
+
+  // Return 2^m * exp(r).
+  return pmax(pmul(y, _mm512_castsi512_ps(emm0)), _x);
+}
+
+/*template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8d
+pexp<Packet8d>(const Packet8d& _x) {
+  Packet8d x = _x;
+
+  _EIGEN_DECLARE_CONST_Packet8d(1, 1.0);
+  _EIGEN_DECLARE_CONST_Packet8d(2, 2.0);
+
+  _EIGEN_DECLARE_CONST_Packet8d(exp_hi, 709.437);
+  _EIGEN_DECLARE_CONST_Packet8d(exp_lo, -709.436139303);
+
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_LOG2EF, 1.4426950408889634073599);
+
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_p0, 1.26177193074810590878e-4);
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_p1, 3.02994407707441961300e-2);
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_p2, 9.99999999999999999910e-1);
+
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_q0, 3.00198505138664455042e-6);
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_q1, 2.52448340349684104192e-3);
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_q2, 2.27265548208155028766e-1);
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_q3, 2.00000000000000000009e0);
+
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_C1, 0.693145751953125);
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_C2, 1.42860682030941723212e-6);
+
+  // clamp x
+  x = pmax(pmin(x, p8d_exp_hi), p8d_exp_lo);
+
+  // Express exp(x) as exp(g + n*log(2)).
+  const Packet8d n =
+      _mm512_mul_round_pd(p8d_cephes_LOG2EF, x, _MM_FROUND_TO_NEAREST_INT);
+
+  // Get the remainder modulo log(2), i.e. the "g" described above. Subtract
+  // n*log(2) out in two steps, i.e. n*C1 + n*C2, C1+C2=log2 to get the last
+  // digits right.
+  const Packet8d nC1 = pmul(n, p8d_cephes_exp_C1);
+  const Packet8d nC2 = pmul(n, p8d_cephes_exp_C2);
+  x = psub(x, nC1);
+  x = psub(x, nC2);
+
+  const Packet8d x2 = pmul(x, x);
+
+  // Evaluate the numerator polynomial of the rational interpolant.
+  Packet8d px = p8d_cephes_exp_p0;
+  px = pmadd(px, x2, p8d_cephes_exp_p1);
+  px = pmadd(px, x2, p8d_cephes_exp_p2);
+  px = pmul(px, x);
+
+  // Evaluate the denominator polynomial of the rational interpolant.
+  Packet8d qx = p8d_cephes_exp_q0;
+  qx = pmadd(qx, x2, p8d_cephes_exp_q1);
+  qx = pmadd(qx, x2, p8d_cephes_exp_q2);
+  qx = pmadd(qx, x2, p8d_cephes_exp_q3);
+
+  // I don't really get this bit, copied from the SSE2 routines, so...
+  // TODO(gonnet): Figure out what is going on here, perhaps find a better
+  // rational interpolant?
+  x = _mm512_div_pd(px, psub(qx, px));
+  x = pmadd(p8d_2, x, p8d_1);
+
+  // Build e=2^n.
+  const Packet8d e = _mm512_castsi512_pd(_mm512_slli_epi64(
+      _mm512_add_epi64(_mm512_cvtpd_epi64(n), _mm512_set1_epi64(1023)), 52));
+
+  // Construct the result 2^n * exp(g) = e * x. The max is used to catch
+  // non-finite values in the input.
+  return pmax(pmul(x, e), _x);
+  }*/
+
+// Functions for sqrt.
+// The EIGEN_FAST_MATH version uses the _mm_rsqrt_ps approximation and one step
+// of Newton's method, at a cost of 1-2 bits of precision as opposed to the
+// exact solution. The main advantage of this approach is not just speed, but
+// also the fact that it can be inlined and pipelined with other computations,
+// further reducing its effective latency.
+#if EIGEN_FAST_MATH
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet16f
+psqrt<Packet16f>(const Packet16f& _x) {
+  _EIGEN_DECLARE_CONST_Packet16f(one_point_five, 1.5f);
+  _EIGEN_DECLARE_CONST_Packet16f(minus_half, -0.5f);
+  _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(flt_min, 0x00800000);
+
+  Packet16f neg_half = pmul(_x, p16f_minus_half);
+
+  // select only the inverse sqrt of positive normal inputs (denormals are
+  // flushed to zero and cause infs as well).
+  __mmask16 non_zero_mask = _mm512_cmp_ps_mask(_x, p16f_flt_min, _CMP_GE_OQ);
+  Packet16f x = _mm512_mask_blend_ps(non_zero_mask, _mm512_rsqrt14_ps(_x),
+                                     _mm512_setzero_ps());
+
+  // Do a single step of Newton's iteration.
+  x = pmul(x, pmadd(neg_half, pmul(x, x), p16f_one_point_five));
+
+  // Multiply the original _x by it's reciprocal square root to extract the
+  // square root.
+  return pmul(_x, x);
+}
+
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8d
+psqrt<Packet8d>(const Packet8d& _x) {
+  _EIGEN_DECLARE_CONST_Packet8d(one_point_five, 1.5);
+  _EIGEN_DECLARE_CONST_Packet8d(minus_half, -0.5);
+  _EIGEN_DECLARE_CONST_Packet8d_FROM_INT64(dbl_min, 0x0010000000000000LL);
+
+  Packet8d neg_half = pmul(_x, p8d_minus_half);
+
+  // select only the inverse sqrt of positive normal inputs (denormals are
+  // flushed to zero and cause infs as well).
+  __mmask8 non_zero_mask = _mm512_cmp_pd_mask(_x, p8d_dbl_min, _CMP_GE_OQ);
+  Packet8d x = _mm512_mask_blend_pd(non_zero_mask, _mm512_rsqrt14_pd(_x),
+                                    _mm512_setzero_pd());
+
+  // Do a first step of Newton's iteration.
+  x = pmul(x, pmadd(neg_half, pmul(x, x), p8d_one_point_five));
+
+  // Do a second step of Newton's iteration.
+  x = pmul(x, pmadd(neg_half, pmul(x, x), p8d_one_point_five));
+
+  // Multiply the original _x by it's reciprocal square root to extract the
+  // square root.
+  return pmul(_x, x);
+}
+#else
+template <>
+EIGEN_STRONG_INLINE Packet16f psqrt<Packet16f>(const Packet16f& x) {
+  return _mm512_sqrt_ps(x);
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d psqrt<Packet8d>(const Packet8d& x) {
+  return _mm512_sqrt_pd(x);
+}
+#endif
+
+// Functions for rsqrt.
+// Almost identical to the sqrt routine, just leave out the last multiplication
+// and fill in NaN/Inf where needed. Note that this function only exists as an
+// iterative version for doubles since there is no instruction for diretly
+// computing the reciprocal square root in AVX-512.
+#ifdef EIGEN_FAST_MATH
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet16f
+prsqrt<Packet16f>(const Packet16f& _x) {
+  _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(inf, 0x7f800000);
+  _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(nan, 0x7fc00000);
+  _EIGEN_DECLARE_CONST_Packet16f(one_point_five, 1.5f);
+  _EIGEN_DECLARE_CONST_Packet16f(minus_half, -0.5f);
+  _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(flt_min, 0x00800000);
+
+  Packet16f neg_half = pmul(_x, p16f_minus_half);
+
+  // select only the inverse sqrt of positive normal inputs (denormals are
+  // flushed to zero and cause infs as well).
+  __mmask16 le_zero_mask = _mm512_cmp_ps_mask(_x, p16f_flt_min, _CMP_LT_OQ);
+  Packet16f x = _mm512_mask_blend_ps(le_zero_mask, _mm512_setzero_ps(),
+                                     _mm512_rsqrt14_ps(_x));
+
+  // Fill in NaNs and Infs for the negative/zero entries.
+  __mmask16 neg_mask = _mm512_cmp_ps_mask(_x, _mm512_setzero_ps(), _CMP_LT_OQ);
+  Packet16f infs_and_nans = _mm512_mask_blend_ps(
+      neg_mask, p16f_nan,
+      _mm512_mask_blend_ps(le_zero_mask, p16f_inf, _mm512_setzero_ps()));
+
+  // Do a single step of Newton's iteration.
+  x = pmul(x, pmadd(neg_half, pmul(x, x), p16f_one_point_five));
+
+  // Insert NaNs and Infs in all the right places.
+  return _mm512_mask_blend_ps(le_zero_mask, infs_and_nans, x);
+}
+
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8d
+prsqrt<Packet8d>(const Packet8d& _x) {
+  _EIGEN_DECLARE_CONST_Packet8d_FROM_INT64(inf, 0x7ff0000000000000LL);
+  _EIGEN_DECLARE_CONST_Packet8d_FROM_INT64(nan, 0x7ff1000000000000LL);
+  _EIGEN_DECLARE_CONST_Packet8d(one_point_five, 1.5);
+  _EIGEN_DECLARE_CONST_Packet8d(minus_half, -0.5);
+  _EIGEN_DECLARE_CONST_Packet8d_FROM_INT64(dbl_min, 0x0010000000000000LL);
+
+  Packet8d neg_half = pmul(_x, p8d_minus_half);
+
+  // select only the inverse sqrt of positive normal inputs (denormals are
+  // flushed to zero and cause infs as well).
+  __mmask8 le_zero_mask = _mm512_cmp_pd_mask(_x, p8d_dbl_min, _CMP_LT_OQ);
+  Packet8d x = _mm512_mask_blend_pd(le_zero_mask, _mm512_setzero_pd(),
+                                    _mm512_rsqrt14_pd(_x));
+
+  // Fill in NaNs and Infs for the negative/zero entries.
+  __mmask8 neg_mask = _mm512_cmp_pd_mask(_x, _mm512_setzero_pd(), _CMP_LT_OQ);
+  Packet8d infs_and_nans = _mm512_mask_blend_pd(
+      neg_mask, p8d_nan,
+      _mm512_mask_blend_pd(le_zero_mask, p8d_inf, _mm512_setzero_pd()));
+
+  // Do a first step of Newton's iteration.
+  x = pmul(x, pmadd(neg_half, pmul(x, x), p8d_one_point_five));
+
+  // Do a second step of Newton's iteration.
+  x = pmul(x, pmadd(neg_half, pmul(x, x), p8d_one_point_five));
+
+  // Insert NaNs and Infs in all the right places.
+  return _mm512_mask_blend_pd(le_zero_mask, infs_and_nans, x);
+}
+#else
+template <>
+EIGEN_STRONG_INLINE Packet16f prsqrt<Packet16f>(const Packet16f& x) {
+  return _mm512_rsqrt28_ps(x);
+}
+#endif
+#endif
+
+}  // end namespace internal
+
+}  // end namespace Eigen
+
+#endif  // THIRD_PARTY_EIGEN3_EIGEN_SRC_CORE_ARCH_AVX512_MATHFUNCTIONS_H_
diff --git a/eigen/Eigen/src/Core/arch/AVX512/PacketMath.h b/eigen/Eigen/src/Core/arch/AVX512/PacketMath.h
new file mode 100644
index 0000000..12b8975
--- /dev/null
+++ b/eigen/Eigen/src/Core/arch/AVX512/PacketMath.h
@@ -0,0 +1,1286 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Benoit Steiner (benoit.steiner.goog@gmail.com)
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PACKET_MATH_AVX512_H
+#define EIGEN_PACKET_MATH_AVX512_H
+
+namespace Eigen {
+
+namespace internal {
+
+#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
+#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8
+#endif
+
+#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
+#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*))
+#endif
+
+#ifdef __FMA__
+#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#endif
+#endif
+
+typedef __m512 Packet16f;
+typedef __m512i Packet16i;
+typedef __m512d Packet8d;
+
+template <>
+struct is_arithmetic<__m512> {
+  enum { value = true };
+};
+template <>
+struct is_arithmetic<__m512i> {
+  enum { value = true };
+};
+template <>
+struct is_arithmetic<__m512d> {
+  enum { value = true };
+};
+
+template<> struct packet_traits<float>  : default_packet_traits
+{
+  typedef Packet16f type;
+  typedef Packet8f half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 16,
+    HasHalfPacket = 1,
+#if EIGEN_GNUC_AT_LEAST(5, 3)
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+    HasLog = 1,
+#endif
+    HasExp = 1,
+    HasSqrt = EIGEN_FAST_MATH,
+    HasRsqrt = EIGEN_FAST_MATH,
+#endif
+    HasDiv = 1
+  };
+ };
+template<> struct packet_traits<double> : default_packet_traits
+{
+  typedef Packet8d type;
+  typedef Packet4d half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 8,
+    HasHalfPacket = 1,
+#if EIGEN_GNUC_AT_LEAST(5, 3)
+    HasSqrt = EIGEN_FAST_MATH,
+    HasRsqrt = EIGEN_FAST_MATH,
+#endif
+    HasDiv = 1
+  };
+};
+
+/* TODO Implement AVX512 for integers
+template<> struct packet_traits<int>    : default_packet_traits
+{
+  typedef Packet16i type;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=8
+  };
+};
+*/
+
+template <>
+struct unpacket_traits<Packet16f> {
+  typedef float type;
+  typedef Packet8f half;
+  enum { size = 16, alignment=Aligned64 };
+};
+template <>
+struct unpacket_traits<Packet8d> {
+  typedef double type;
+  typedef Packet4d half;
+  enum { size = 8, alignment=Aligned64 };
+};
+template <>
+struct unpacket_traits<Packet16i> {
+  typedef int type;
+  typedef Packet8i half;
+  enum { size = 16, alignment=Aligned64 };
+};
+
+template <>
+EIGEN_STRONG_INLINE Packet16f pset1<Packet16f>(const float& from) {
+  return _mm512_set1_ps(from);
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pset1<Packet8d>(const double& from) {
+  return _mm512_set1_pd(from);
+}
+template <>
+EIGEN_STRONG_INLINE Packet16i pset1<Packet16i>(const int& from) {
+  return _mm512_set1_epi32(from);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f pload1<Packet16f>(const float* from) {
+  return _mm512_broadcastss_ps(_mm_load_ps1(from));
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pload1<Packet8d>(const double* from) {
+  return _mm512_broadcastsd_pd(_mm_load_pd1(from));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f plset<Packet16f>(const float& a) {
+  return _mm512_add_ps(
+      _mm512_set1_ps(a),
+      _mm512_set_ps(15.0f, 14.0f, 13.0f, 12.0f, 11.0f, 10.0f, 9.0f, 8.0f, 7.0f, 6.0f, 5.0f,
+                    4.0f, 3.0f, 2.0f, 1.0f, 0.0f));
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d plset<Packet8d>(const double& a) {
+  return _mm512_add_pd(_mm512_set1_pd(a),
+                       _mm512_set_pd(7.0, 6.0, 5.0, 4.0, 3.0, 2.0, 1.0, 0.0));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f padd<Packet16f>(const Packet16f& a,
+                                              const Packet16f& b) {
+  return _mm512_add_ps(a, b);
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d padd<Packet8d>(const Packet8d& a,
+                                            const Packet8d& b) {
+  return _mm512_add_pd(a, b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f psub<Packet16f>(const Packet16f& a,
+                                              const Packet16f& b) {
+  return _mm512_sub_ps(a, b);
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d psub<Packet8d>(const Packet8d& a,
+                                            const Packet8d& b) {
+  return _mm512_sub_pd(a, b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f pnegate(const Packet16f& a) {
+  return _mm512_sub_ps(_mm512_set1_ps(0.0), a);
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pnegate(const Packet8d& a) {
+  return _mm512_sub_pd(_mm512_set1_pd(0.0), a);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f pconj(const Packet16f& a) {
+  return a;
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pconj(const Packet8d& a) {
+  return a;
+}
+template <>
+EIGEN_STRONG_INLINE Packet16i pconj(const Packet16i& a) {
+  return a;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f pmul<Packet16f>(const Packet16f& a,
+                                              const Packet16f& b) {
+  return _mm512_mul_ps(a, b);
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pmul<Packet8d>(const Packet8d& a,
+                                            const Packet8d& b) {
+  return _mm512_mul_pd(a, b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f pdiv<Packet16f>(const Packet16f& a,
+                                              const Packet16f& b) {
+  return _mm512_div_ps(a, b);
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pdiv<Packet8d>(const Packet8d& a,
+                                            const Packet8d& b) {
+  return _mm512_div_pd(a, b);
+}
+
+#ifdef __FMA__
+template <>
+EIGEN_STRONG_INLINE Packet16f pmadd(const Packet16f& a, const Packet16f& b,
+                                    const Packet16f& c) {
+  return _mm512_fmadd_ps(a, b, c);
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pmadd(const Packet8d& a, const Packet8d& b,
+                                   const Packet8d& c) {
+  return _mm512_fmadd_pd(a, b, c);
+}
+#endif
+
+template <>
+EIGEN_STRONG_INLINE Packet16f pmin<Packet16f>(const Packet16f& a,
+                                              const Packet16f& b) {
+  // Arguments are reversed to match NaN propagation behavior of std::min.
+  return _mm512_min_ps(b, a);
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pmin<Packet8d>(const Packet8d& a,
+                                            const Packet8d& b) {
+  // Arguments are reversed to match NaN propagation behavior of std::min.
+  return _mm512_min_pd(b, a);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f pmax<Packet16f>(const Packet16f& a,
+                                              const Packet16f& b) {
+  // Arguments are reversed to match NaN propagation behavior of std::max.
+  return _mm512_max_ps(b, a);
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pmax<Packet8d>(const Packet8d& a,
+                                            const Packet8d& b) {
+  // Arguments are reversed to match NaN propagation behavior of std::max.
+  return _mm512_max_pd(b, a);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f pand<Packet16f>(const Packet16f& a,
+                                              const Packet16f& b) {
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+  return _mm512_and_ps(a, b);
+#else
+  Packet16f res = _mm512_undefined_ps();
+  Packet4f lane0_a = _mm512_extractf32x4_ps(a, 0);
+  Packet4f lane0_b = _mm512_extractf32x4_ps(b, 0);
+  res = _mm512_insertf32x4(res, _mm_and_ps(lane0_a, lane0_b), 0);
+
+  Packet4f lane1_a = _mm512_extractf32x4_ps(a, 1);
+  Packet4f lane1_b = _mm512_extractf32x4_ps(b, 1);
+  res = _mm512_insertf32x4(res, _mm_and_ps(lane1_a, lane1_b), 1);
+
+  Packet4f lane2_a = _mm512_extractf32x4_ps(a, 2);
+  Packet4f lane2_b = _mm512_extractf32x4_ps(b, 2);
+  res = _mm512_insertf32x4(res, _mm_and_ps(lane2_a, lane2_b), 2);
+
+  Packet4f lane3_a = _mm512_extractf32x4_ps(a, 3);
+  Packet4f lane3_b = _mm512_extractf32x4_ps(b, 3);
+  res = _mm512_insertf32x4(res, _mm_and_ps(lane3_a, lane3_b), 3);
+
+  return res;
+#endif
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pand<Packet8d>(const Packet8d& a,
+                                            const Packet8d& b) {
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+  return _mm512_and_pd(a, b);
+#else
+  Packet8d res = _mm512_undefined_pd();
+  Packet4d lane0_a = _mm512_extractf64x4_pd(a, 0);
+  Packet4d lane0_b = _mm512_extractf64x4_pd(b, 0);
+  res = _mm512_insertf64x4(res, _mm256_and_pd(lane0_a, lane0_b), 0);
+
+  Packet4d lane1_a = _mm512_extractf64x4_pd(a, 1);
+  Packet4d lane1_b = _mm512_extractf64x4_pd(b, 1);
+  res = _mm512_insertf64x4(res, _mm256_and_pd(lane1_a, lane1_b), 1);
+
+  return res;
+#endif
+}
+template <>
+EIGEN_STRONG_INLINE Packet16f por<Packet16f>(const Packet16f& a,
+                                             const Packet16f& b) {
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+  return _mm512_or_ps(a, b);
+#else
+  Packet16f res = _mm512_undefined_ps();
+  Packet4f lane0_a = _mm512_extractf32x4_ps(a, 0);
+  Packet4f lane0_b = _mm512_extractf32x4_ps(b, 0);
+  res = _mm512_insertf32x4(res, _mm_or_ps(lane0_a, lane0_b), 0);
+
+  Packet4f lane1_a = _mm512_extractf32x4_ps(a, 1);
+  Packet4f lane1_b = _mm512_extractf32x4_ps(b, 1);
+  res = _mm512_insertf32x4(res, _mm_or_ps(lane1_a, lane1_b), 1);
+
+  Packet4f lane2_a = _mm512_extractf32x4_ps(a, 2);
+  Packet4f lane2_b = _mm512_extractf32x4_ps(b, 2);
+  res = _mm512_insertf32x4(res, _mm_or_ps(lane2_a, lane2_b), 2);
+
+  Packet4f lane3_a = _mm512_extractf32x4_ps(a, 3);
+  Packet4f lane3_b = _mm512_extractf32x4_ps(b, 3);
+  res = _mm512_insertf32x4(res, _mm_or_ps(lane3_a, lane3_b), 3);
+
+  return res;
+#endif
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet8d por<Packet8d>(const Packet8d& a,
+                                           const Packet8d& b) {
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+  return _mm512_or_pd(a, b);
+#else
+  Packet8d res = _mm512_undefined_pd();
+  Packet4d lane0_a = _mm512_extractf64x4_pd(a, 0);
+  Packet4d lane0_b = _mm512_extractf64x4_pd(b, 0);
+  res = _mm512_insertf64x4(res, _mm256_or_pd(lane0_a, lane0_b), 0);
+
+  Packet4d lane1_a = _mm512_extractf64x4_pd(a, 1);
+  Packet4d lane1_b = _mm512_extractf64x4_pd(b, 1);
+  res = _mm512_insertf64x4(res, _mm256_or_pd(lane1_a, lane1_b), 1);
+
+  return res;
+#endif
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f pxor<Packet16f>(const Packet16f& a,
+                                              const Packet16f& b) {
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+  return _mm512_xor_ps(a, b);
+#else
+  Packet16f res = _mm512_undefined_ps();
+  Packet4f lane0_a = _mm512_extractf32x4_ps(a, 0);
+  Packet4f lane0_b = _mm512_extractf32x4_ps(b, 0);
+  res = _mm512_insertf32x4(res, _mm_xor_ps(lane0_a, lane0_b), 0);
+
+  Packet4f lane1_a = _mm512_extractf32x4_ps(a, 1);
+  Packet4f lane1_b = _mm512_extractf32x4_ps(b, 1);
+  res = _mm512_insertf32x4(res, _mm_xor_ps(lane1_a, lane1_b), 1);
+
+  Packet4f lane2_a = _mm512_extractf32x4_ps(a, 2);
+  Packet4f lane2_b = _mm512_extractf32x4_ps(b, 2);
+  res = _mm512_insertf32x4(res, _mm_xor_ps(lane2_a, lane2_b), 2);
+
+  Packet4f lane3_a = _mm512_extractf32x4_ps(a, 3);
+  Packet4f lane3_b = _mm512_extractf32x4_ps(b, 3);
+  res = _mm512_insertf32x4(res, _mm_xor_ps(lane3_a, lane3_b), 3);
+
+  return res;
+#endif
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pxor<Packet8d>(const Packet8d& a,
+                                            const Packet8d& b) {
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+  return _mm512_xor_pd(a, b);
+#else
+  Packet8d res = _mm512_undefined_pd();
+  Packet4d lane0_a = _mm512_extractf64x4_pd(a, 0);
+  Packet4d lane0_b = _mm512_extractf64x4_pd(b, 0);
+  res = _mm512_insertf64x4(res, _mm256_xor_pd(lane0_a, lane0_b), 0);
+
+  Packet4d lane1_a = _mm512_extractf64x4_pd(a, 1);
+  Packet4d lane1_b = _mm512_extractf64x4_pd(b, 1);
+  res = _mm512_insertf64x4(res, _mm256_xor_pd(lane1_a, lane1_b), 1);
+
+  return res;
+#endif
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f pandnot<Packet16f>(const Packet16f& a,
+                                                 const Packet16f& b) {
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+  return _mm512_andnot_ps(a, b);
+#else
+  Packet16f res = _mm512_undefined_ps();
+  Packet4f lane0_a = _mm512_extractf32x4_ps(a, 0);
+  Packet4f lane0_b = _mm512_extractf32x4_ps(b, 0);
+  res = _mm512_insertf32x4(res, _mm_andnot_ps(lane0_a, lane0_b), 0);
+
+  Packet4f lane1_a = _mm512_extractf32x4_ps(a, 1);
+  Packet4f lane1_b = _mm512_extractf32x4_ps(b, 1);
+  res = _mm512_insertf32x4(res, _mm_andnot_ps(lane1_a, lane1_b), 1);
+
+  Packet4f lane2_a = _mm512_extractf32x4_ps(a, 2);
+  Packet4f lane2_b = _mm512_extractf32x4_ps(b, 2);
+  res = _mm512_insertf32x4(res, _mm_andnot_ps(lane2_a, lane2_b), 2);
+
+  Packet4f lane3_a = _mm512_extractf32x4_ps(a, 3);
+  Packet4f lane3_b = _mm512_extractf32x4_ps(b, 3);
+  res = _mm512_insertf32x4(res, _mm_andnot_ps(lane3_a, lane3_b), 3);
+
+  return res;
+#endif
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pandnot<Packet8d>(const Packet8d& a,
+                                               const Packet8d& b) {
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+  return _mm512_andnot_pd(a, b);
+#else
+  Packet8d res = _mm512_undefined_pd();
+  Packet4d lane0_a = _mm512_extractf64x4_pd(a, 0);
+  Packet4d lane0_b = _mm512_extractf64x4_pd(b, 0);
+  res = _mm512_insertf64x4(res, _mm256_andnot_pd(lane0_a, lane0_b), 0);
+
+  Packet4d lane1_a = _mm512_extractf64x4_pd(a, 1);
+  Packet4d lane1_b = _mm512_extractf64x4_pd(b, 1);
+  res = _mm512_insertf64x4(res, _mm256_andnot_pd(lane1_a, lane1_b), 1);
+
+  return res;
+#endif
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f pload<Packet16f>(const float* from) {
+  EIGEN_DEBUG_ALIGNED_LOAD return _mm512_load_ps(from);
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pload<Packet8d>(const double* from) {
+  EIGEN_DEBUG_ALIGNED_LOAD return _mm512_load_pd(from);
+}
+template <>
+EIGEN_STRONG_INLINE Packet16i pload<Packet16i>(const int* from) {
+  EIGEN_DEBUG_ALIGNED_LOAD return _mm512_load_si512(
+      reinterpret_cast<const __m512i*>(from));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f ploadu<Packet16f>(const float* from) {
+  EIGEN_DEBUG_UNALIGNED_LOAD return _mm512_loadu_ps(from);
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d ploadu<Packet8d>(const double* from) {
+  EIGEN_DEBUG_UNALIGNED_LOAD return _mm512_loadu_pd(from);
+}
+template <>
+EIGEN_STRONG_INLINE Packet16i ploadu<Packet16i>(const int* from) {
+  EIGEN_DEBUG_UNALIGNED_LOAD return _mm512_loadu_si512(
+      reinterpret_cast<const __m512i*>(from));
+}
+
+// Loads 8 floats from memory a returns the packet
+// {a0, a0  a1, a1, a2, a2, a3, a3, a4, a4, a5, a5, a6, a6, a7, a7}
+template <>
+EIGEN_STRONG_INLINE Packet16f ploaddup<Packet16f>(const float* from) {
+  __m256i low_half = _mm256_load_si256(reinterpret_cast<const __m256i*>(from));
+  __m512 even_elements = _mm512_castsi512_ps(_mm512_cvtepu32_epi64(low_half));
+  __m512 pairs = _mm512_permute_ps(even_elements, _MM_SHUFFLE(2, 2, 0, 0));
+  return pairs;
+}
+// Loads 4 doubles from memory a returns the packet {a0, a0  a1, a1, a2, a2, a3,
+// a3}
+template <>
+EIGEN_STRONG_INLINE Packet8d ploaddup<Packet8d>(const double* from) {
+ __m512d x = _mm512_setzero_pd();
+  x = _mm512_insertf64x2(x, _mm_loaddup_pd(&from[0]), 0);
+  x = _mm512_insertf64x2(x, _mm_loaddup_pd(&from[1]), 1);
+  x = _mm512_insertf64x2(x, _mm_loaddup_pd(&from[2]), 2);
+  x = _mm512_insertf64x2(x, _mm_loaddup_pd(&from[3]), 3);
+  return x;
+}
+
+// Loads 4 floats from memory a returns the packet
+// {a0, a0  a0, a0, a1, a1, a1, a1, a2, a2, a2, a2, a3, a3, a3, a3}
+template <>
+EIGEN_STRONG_INLINE Packet16f ploadquad<Packet16f>(const float* from) {
+  Packet16f tmp = _mm512_undefined_ps();
+  tmp = _mm512_insertf32x4(tmp, _mm_load_ps1(from), 0);
+  tmp = _mm512_insertf32x4(tmp, _mm_load_ps1(from + 1), 1);
+  tmp = _mm512_insertf32x4(tmp, _mm_load_ps1(from + 2), 2);
+  tmp = _mm512_insertf32x4(tmp, _mm_load_ps1(from + 3), 3);
+  return tmp;
+}
+// Loads 2 doubles from memory a returns the packet
+// {a0, a0  a0, a0, a1, a1, a1, a1}
+template <>
+EIGEN_STRONG_INLINE Packet8d ploadquad<Packet8d>(const double* from) {
+  __m128d tmp0 = _mm_load_pd1(from);
+  __m256d lane0 = _mm256_broadcastsd_pd(tmp0);
+  __m128d tmp1 = _mm_load_pd1(from + 1);
+  __m256d lane1 = _mm256_broadcastsd_pd(tmp1);
+  __m512d tmp = _mm512_undefined_pd();
+  tmp = _mm512_insertf64x4(tmp, lane0, 0);
+  return _mm512_insertf64x4(tmp, lane1, 1);
+}
+
+template <>
+EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet16f& from) {
+  EIGEN_DEBUG_ALIGNED_STORE _mm512_store_ps(to, from);
+}
+template <>
+EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet8d& from) {
+  EIGEN_DEBUG_ALIGNED_STORE _mm512_store_pd(to, from);
+}
+template <>
+EIGEN_STRONG_INLINE void pstore<int>(int* to, const Packet16i& from) {
+  EIGEN_DEBUG_ALIGNED_STORE _mm512_storeu_si512(reinterpret_cast<__m512i*>(to),
+                                                from);
+}
+
+template <>
+EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet16f& from) {
+  EIGEN_DEBUG_UNALIGNED_STORE _mm512_storeu_ps(to, from);
+}
+template <>
+EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet8d& from) {
+  EIGEN_DEBUG_UNALIGNED_STORE _mm512_storeu_pd(to, from);
+}
+template <>
+EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet16i& from) {
+  EIGEN_DEBUG_UNALIGNED_STORE _mm512_storeu_si512(
+      reinterpret_cast<__m512i*>(to), from);
+}
+
+template <>
+EIGEN_DEVICE_FUNC inline Packet16f pgather<float, Packet16f>(const float* from,
+                                                             Index stride) {
+  Packet16i stride_vector = _mm512_set1_epi32(stride);
+  Packet16i stride_multiplier =
+      _mm512_set_epi32(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
+  Packet16i indices = _mm512_mullo_epi32(stride_vector, stride_multiplier);
+
+  return _mm512_i32gather_ps(indices, from, 4);
+}
+template <>
+EIGEN_DEVICE_FUNC inline Packet8d pgather<double, Packet8d>(const double* from,
+                                                            Index stride) {
+  Packet8i stride_vector = _mm256_set1_epi32(stride);
+  Packet8i stride_multiplier = _mm256_set_epi32(7, 6, 5, 4, 3, 2, 1, 0);
+  Packet8i indices = _mm256_mullo_epi32(stride_vector, stride_multiplier);
+
+  return _mm512_i32gather_pd(indices, from, 8);
+}
+
+template <>
+EIGEN_DEVICE_FUNC inline void pscatter<float, Packet16f>(float* to,
+                                                         const Packet16f& from,
+                                                         Index stride) {
+  Packet16i stride_vector = _mm512_set1_epi32(stride);
+  Packet16i stride_multiplier =
+      _mm512_set_epi32(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
+  Packet16i indices = _mm512_mullo_epi32(stride_vector, stride_multiplier);
+  _mm512_i32scatter_ps(to, indices, from, 4);
+}
+template <>
+EIGEN_DEVICE_FUNC inline void pscatter<double, Packet8d>(double* to,
+                                                         const Packet8d& from,
+                                                         Index stride) {
+  Packet8i stride_vector = _mm256_set1_epi32(stride);
+  Packet8i stride_multiplier = _mm256_set_epi32(7, 6, 5, 4, 3, 2, 1, 0);
+  Packet8i indices = _mm256_mullo_epi32(stride_vector, stride_multiplier);
+  _mm512_i32scatter_pd(to, indices, from, 8);
+}
+
+template <>
+EIGEN_STRONG_INLINE void pstore1<Packet16f>(float* to, const float& a) {
+  Packet16f pa = pset1<Packet16f>(a);
+  pstore(to, pa);
+}
+template <>
+EIGEN_STRONG_INLINE void pstore1<Packet8d>(double* to, const double& a) {
+  Packet8d pa = pset1<Packet8d>(a);
+  pstore(to, pa);
+}
+template <>
+EIGEN_STRONG_INLINE void pstore1<Packet16i>(int* to, const int& a) {
+  Packet16i pa = pset1<Packet16i>(a);
+  pstore(to, pa);
+}
+
+template<> EIGEN_STRONG_INLINE void prefetch<float>(const float*   addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*       addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+
+template <>
+EIGEN_STRONG_INLINE float pfirst<Packet16f>(const Packet16f& a) {
+  return _mm_cvtss_f32(_mm512_extractf32x4_ps(a, 0));
+}
+template <>
+EIGEN_STRONG_INLINE double pfirst<Packet8d>(const Packet8d& a) {
+  return _mm_cvtsd_f64(_mm256_extractf128_pd(_mm512_extractf64x4_pd(a, 0), 0));
+}
+template <>
+EIGEN_STRONG_INLINE int pfirst<Packet16i>(const Packet16i& a) {
+  return _mm_extract_epi32(_mm512_extracti32x4_epi32(a, 0), 0);
+}
+
+template<> EIGEN_STRONG_INLINE Packet16f preverse(const Packet16f& a)
+{
+  return _mm512_permutexvar_ps(_mm512_set_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15), a);
+}
+
+template<> EIGEN_STRONG_INLINE Packet8d preverse(const Packet8d& a)
+{
+  return _mm512_permutexvar_pd(_mm512_set_epi32(0, 0, 0, 1, 0, 2, 0, 3, 0, 4, 0, 5, 0, 6, 0, 7), a);
+}
+
+template<> EIGEN_STRONG_INLINE Packet16f pabs(const Packet16f& a)
+{
+  // _mm512_abs_ps intrinsic not found, so hack around it
+  return (__m512)_mm512_and_si512((__m512i)a, _mm512_set1_epi32(0x7fffffff));
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pabs(const Packet8d& a) {
+  // _mm512_abs_ps intrinsic not found, so hack around it
+  return (__m512d)_mm512_and_si512((__m512i)a,
+                                   _mm512_set1_epi64(0x7fffffffffffffff));
+}
+
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+// AVX512F does not define _mm512_extractf32x8_ps to extract _m256 from _m512
+#define EIGEN_EXTRACT_8f_FROM_16f(INPUT, OUTPUT)                           \
+  __m256 OUTPUT##_0 = _mm512_extractf32x8_ps(INPUT, 0);                    \
+  __m256 OUTPUT##_1 = _mm512_extractf32x8_ps(INPUT, 1)
+#else
+#define EIGEN_EXTRACT_8f_FROM_16f(INPUT, OUTPUT)                \
+  __m256 OUTPUT##_0 = _mm256_insertf128_ps(                     \
+      _mm256_castps128_ps256(_mm512_extractf32x4_ps(INPUT, 0)), \
+      _mm512_extractf32x4_ps(INPUT, 1), 1);                     \
+  __m256 OUTPUT##_1 = _mm256_insertf128_ps(                     \
+      _mm256_castps128_ps256(_mm512_extractf32x4_ps(INPUT, 2)), \
+      _mm512_extractf32x4_ps(INPUT, 3), 1);
+#endif
+
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+#define EIGEN_INSERT_8f_INTO_16f(OUTPUT, INPUTA, INPUTB) \
+  OUTPUT = _mm512_insertf32x8(OUTPUT, INPUTA, 0);        \
+  OUTPUT = _mm512_insertf32x8(OUTPUT, INPUTB, 1);
+#else
+#define EIGEN_INSERT_8f_INTO_16f(OUTPUT, INPUTA, INPUTB)                    \
+  OUTPUT = _mm512_insertf32x4(OUTPUT, _mm256_extractf128_ps(INPUTA, 0), 0); \
+  OUTPUT = _mm512_insertf32x4(OUTPUT, _mm256_extractf128_ps(INPUTA, 1), 1); \
+  OUTPUT = _mm512_insertf32x4(OUTPUT, _mm256_extractf128_ps(INPUTB, 0), 2); \
+  OUTPUT = _mm512_insertf32x4(OUTPUT, _mm256_extractf128_ps(INPUTB, 1), 3);
+#endif
+template<> EIGEN_STRONG_INLINE Packet16f preduxp<Packet16f>(const Packet16f*
+vecs)
+{
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[0], vecs0);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[1], vecs1);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[2], vecs2);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[3], vecs3);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[4], vecs4);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[5], vecs5);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[6], vecs6);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[7], vecs7);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[8], vecs8);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[9], vecs9);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[10], vecs10);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[11], vecs11);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[12], vecs12);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[13], vecs13);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[14], vecs14);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[15], vecs15);
+
+  __m256 hsum1 = _mm256_hadd_ps(vecs0_0, vecs1_0);
+  __m256 hsum2 = _mm256_hadd_ps(vecs2_0, vecs3_0);
+  __m256 hsum3 = _mm256_hadd_ps(vecs4_0, vecs5_0);
+  __m256 hsum4 = _mm256_hadd_ps(vecs6_0, vecs7_0);
+
+  __m256 hsum5 = _mm256_hadd_ps(hsum1, hsum1);
+  __m256 hsum6 = _mm256_hadd_ps(hsum2, hsum2);
+  __m256 hsum7 = _mm256_hadd_ps(hsum3, hsum3);
+  __m256 hsum8 = _mm256_hadd_ps(hsum4, hsum4);
+
+  __m256 perm1 = _mm256_permute2f128_ps(hsum5, hsum5, 0x23);
+  __m256 perm2 = _mm256_permute2f128_ps(hsum6, hsum6, 0x23);
+  __m256 perm3 = _mm256_permute2f128_ps(hsum7, hsum7, 0x23);
+  __m256 perm4 = _mm256_permute2f128_ps(hsum8, hsum8, 0x23);
+
+  __m256 sum1 = _mm256_add_ps(perm1, hsum5);
+  __m256 sum2 = _mm256_add_ps(perm2, hsum6);
+  __m256 sum3 = _mm256_add_ps(perm3, hsum7);
+  __m256 sum4 = _mm256_add_ps(perm4, hsum8);
+
+  __m256 blend1 = _mm256_blend_ps(sum1, sum2, 0xcc);
+  __m256 blend2 = _mm256_blend_ps(sum3, sum4, 0xcc);
+
+  __m256 final = _mm256_blend_ps(blend1, blend2, 0xf0);
+
+  hsum1 = _mm256_hadd_ps(vecs0_1, vecs1_1);
+  hsum2 = _mm256_hadd_ps(vecs2_1, vecs3_1);
+  hsum3 = _mm256_hadd_ps(vecs4_1, vecs5_1);
+  hsum4 = _mm256_hadd_ps(vecs6_1, vecs7_1);
+
+  hsum5 = _mm256_hadd_ps(hsum1, hsum1);
+  hsum6 = _mm256_hadd_ps(hsum2, hsum2);
+  hsum7 = _mm256_hadd_ps(hsum3, hsum3);
+  hsum8 = _mm256_hadd_ps(hsum4, hsum4);
+
+  perm1 = _mm256_permute2f128_ps(hsum5, hsum5, 0x23);
+  perm2 = _mm256_permute2f128_ps(hsum6, hsum6, 0x23);
+  perm3 = _mm256_permute2f128_ps(hsum7, hsum7, 0x23);
+  perm4 = _mm256_permute2f128_ps(hsum8, hsum8, 0x23);
+
+  sum1 = _mm256_add_ps(perm1, hsum5);
+  sum2 = _mm256_add_ps(perm2, hsum6);
+  sum3 = _mm256_add_ps(perm3, hsum7);
+  sum4 = _mm256_add_ps(perm4, hsum8);
+
+  blend1 = _mm256_blend_ps(sum1, sum2, 0xcc);
+  blend2 = _mm256_blend_ps(sum3, sum4, 0xcc);
+
+  final = _mm256_add_ps(final, _mm256_blend_ps(blend1, blend2, 0xf0));
+
+  hsum1 = _mm256_hadd_ps(vecs8_0, vecs9_0);
+  hsum2 = _mm256_hadd_ps(vecs10_0, vecs11_0);
+  hsum3 = _mm256_hadd_ps(vecs12_0, vecs13_0);
+  hsum4 = _mm256_hadd_ps(vecs14_0, vecs15_0);
+
+  hsum5 = _mm256_hadd_ps(hsum1, hsum1);
+  hsum6 = _mm256_hadd_ps(hsum2, hsum2);
+  hsum7 = _mm256_hadd_ps(hsum3, hsum3);
+  hsum8 = _mm256_hadd_ps(hsum4, hsum4);
+
+  perm1 = _mm256_permute2f128_ps(hsum5, hsum5, 0x23);
+  perm2 = _mm256_permute2f128_ps(hsum6, hsum6, 0x23);
+  perm3 = _mm256_permute2f128_ps(hsum7, hsum7, 0x23);
+  perm4 = _mm256_permute2f128_ps(hsum8, hsum8, 0x23);
+
+  sum1 = _mm256_add_ps(perm1, hsum5);
+  sum2 = _mm256_add_ps(perm2, hsum6);
+  sum3 = _mm256_add_ps(perm3, hsum7);
+  sum4 = _mm256_add_ps(perm4, hsum8);
+
+  blend1 = _mm256_blend_ps(sum1, sum2, 0xcc);
+  blend2 = _mm256_blend_ps(sum3, sum4, 0xcc);
+
+  __m256 final_1 = _mm256_blend_ps(blend1, blend2, 0xf0);
+
+  hsum1 = _mm256_hadd_ps(vecs8_1, vecs9_1);
+  hsum2 = _mm256_hadd_ps(vecs10_1, vecs11_1);
+  hsum3 = _mm256_hadd_ps(vecs12_1, vecs13_1);
+  hsum4 = _mm256_hadd_ps(vecs14_1, vecs15_1);
+
+  hsum5 = _mm256_hadd_ps(hsum1, hsum1);
+  hsum6 = _mm256_hadd_ps(hsum2, hsum2);
+  hsum7 = _mm256_hadd_ps(hsum3, hsum3);
+  hsum8 = _mm256_hadd_ps(hsum4, hsum4);
+
+  perm1 = _mm256_permute2f128_ps(hsum5, hsum5, 0x23);
+  perm2 = _mm256_permute2f128_ps(hsum6, hsum6, 0x23);
+  perm3 = _mm256_permute2f128_ps(hsum7, hsum7, 0x23);
+  perm4 = _mm256_permute2f128_ps(hsum8, hsum8, 0x23);
+
+  sum1 = _mm256_add_ps(perm1, hsum5);
+  sum2 = _mm256_add_ps(perm2, hsum6);
+  sum3 = _mm256_add_ps(perm3, hsum7);
+  sum4 = _mm256_add_ps(perm4, hsum8);
+
+  blend1 = _mm256_blend_ps(sum1, sum2, 0xcc);
+  blend2 = _mm256_blend_ps(sum3, sum4, 0xcc);
+
+  final_1 = _mm256_add_ps(final_1, _mm256_blend_ps(blend1, blend2, 0xf0));
+
+  __m512 final_output;
+
+  EIGEN_INSERT_8f_INTO_16f(final_output, final, final_1);
+  return final_output;
+}
+
+template<> EIGEN_STRONG_INLINE Packet8d preduxp<Packet8d>(const Packet8d* vecs)
+{
+  Packet4d vecs0_0 = _mm512_extractf64x4_pd(vecs[0], 0);
+  Packet4d vecs0_1 = _mm512_extractf64x4_pd(vecs[0], 1);
+
+  Packet4d vecs1_0 = _mm512_extractf64x4_pd(vecs[1], 0);
+  Packet4d vecs1_1 = _mm512_extractf64x4_pd(vecs[1], 1);
+
+  Packet4d vecs2_0 = _mm512_extractf64x4_pd(vecs[2], 0);
+  Packet4d vecs2_1 = _mm512_extractf64x4_pd(vecs[2], 1);
+
+  Packet4d vecs3_0 = _mm512_extractf64x4_pd(vecs[3], 0);
+  Packet4d vecs3_1 = _mm512_extractf64x4_pd(vecs[3], 1);
+
+  Packet4d vecs4_0 = _mm512_extractf64x4_pd(vecs[4], 0);
+  Packet4d vecs4_1 = _mm512_extractf64x4_pd(vecs[4], 1);
+
+  Packet4d vecs5_0 = _mm512_extractf64x4_pd(vecs[5], 0);
+  Packet4d vecs5_1 = _mm512_extractf64x4_pd(vecs[5], 1);
+
+  Packet4d vecs6_0 = _mm512_extractf64x4_pd(vecs[6], 0);
+  Packet4d vecs6_1 = _mm512_extractf64x4_pd(vecs[6], 1);
+
+  Packet4d vecs7_0 = _mm512_extractf64x4_pd(vecs[7], 0);
+  Packet4d vecs7_1 = _mm512_extractf64x4_pd(vecs[7], 1);
+
+  Packet4d tmp0, tmp1;
+
+  tmp0 = _mm256_hadd_pd(vecs0_0, vecs1_0);
+  tmp0 = _mm256_add_pd(tmp0, _mm256_permute2f128_pd(tmp0, tmp0, 1));
+
+  tmp1 = _mm256_hadd_pd(vecs2_0, vecs3_0);
+  tmp1 = _mm256_add_pd(tmp1, _mm256_permute2f128_pd(tmp1, tmp1, 1));
+
+  __m256d final_0 = _mm256_blend_pd(tmp0, tmp1, 0xC);
+
+  tmp0 = _mm256_hadd_pd(vecs0_1, vecs1_1);
+  tmp0 = _mm256_add_pd(tmp0, _mm256_permute2f128_pd(tmp0, tmp0, 1));
+
+  tmp1 = _mm256_hadd_pd(vecs2_1, vecs3_1);
+  tmp1 = _mm256_add_pd(tmp1, _mm256_permute2f128_pd(tmp1, tmp1, 1));
+
+  final_0 = _mm256_add_pd(final_0, _mm256_blend_pd(tmp0, tmp1, 0xC));
+
+  tmp0 = _mm256_hadd_pd(vecs4_0, vecs5_0);
+  tmp0 = _mm256_add_pd(tmp0, _mm256_permute2f128_pd(tmp0, tmp0, 1));
+
+  tmp1 = _mm256_hadd_pd(vecs6_0, vecs7_0);
+  tmp1 = _mm256_add_pd(tmp1, _mm256_permute2f128_pd(tmp1, tmp1, 1));
+
+  __m256d final_1 = _mm256_blend_pd(tmp0, tmp1, 0xC);
+
+  tmp0 = _mm256_hadd_pd(vecs4_1, vecs5_1);
+  tmp0 = _mm256_add_pd(tmp0, _mm256_permute2f128_pd(tmp0, tmp0, 1));
+
+  tmp1 = _mm256_hadd_pd(vecs6_1, vecs7_1);
+  tmp1 = _mm256_add_pd(tmp1, _mm256_permute2f128_pd(tmp1, tmp1, 1));
+
+  final_1 = _mm256_add_pd(final_1, _mm256_blend_pd(tmp0, tmp1, 0xC));
+
+  __m512d final_output = _mm512_insertf64x4(final_output, final_0, 0);
+
+  return _mm512_insertf64x4(final_output, final_1, 1);
+}
+
+template <>
+EIGEN_STRONG_INLINE float predux<Packet16f>(const Packet16f& a) {
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+  __m256 lane0 = _mm512_extractf32x8_ps(a, 0);
+  __m256 lane1 = _mm512_extractf32x8_ps(a, 1);
+  Packet8f x = _mm256_add_ps(lane0, lane1);
+  return predux<Packet8f>(x);
+#else
+  __m128 lane0 = _mm512_extractf32x4_ps(a, 0);
+  __m128 lane1 = _mm512_extractf32x4_ps(a, 1);
+  __m128 lane2 = _mm512_extractf32x4_ps(a, 2);
+  __m128 lane3 = _mm512_extractf32x4_ps(a, 3);
+  __m128 sum = _mm_add_ps(_mm_add_ps(lane0, lane1), _mm_add_ps(lane2, lane3));
+  sum = _mm_hadd_ps(sum, sum);
+  sum = _mm_hadd_ps(sum, _mm_permute_ps(sum, 1));
+  return _mm_cvtss_f32(sum);
+#endif
+}
+template <>
+EIGEN_STRONG_INLINE double predux<Packet8d>(const Packet8d& a) {
+  __m256d lane0 = _mm512_extractf64x4_pd(a, 0);
+  __m256d lane1 = _mm512_extractf64x4_pd(a, 1);
+  __m256d sum = _mm256_add_pd(lane0, lane1);
+  __m256d tmp0 = _mm256_hadd_pd(sum, _mm256_permute2f128_pd(sum, sum, 1));
+  return _mm_cvtsd_f64(_mm256_castpd256_pd128(_mm256_hadd_pd(tmp0, tmp0)));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet8f predux_downto4<Packet16f>(const Packet16f& a) {
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+  __m256 lane0 = _mm512_extractf32x8_ps(a, 0);
+  __m256 lane1 = _mm512_extractf32x8_ps(a, 1);
+  return _mm256_add_ps(lane0, lane1);
+#else
+  __m128 lane0 = _mm512_extractf32x4_ps(a, 0);
+  __m128 lane1 = _mm512_extractf32x4_ps(a, 1);
+  __m128 lane2 = _mm512_extractf32x4_ps(a, 2);
+  __m128 lane3 = _mm512_extractf32x4_ps(a, 3);
+  __m128 sum0 = _mm_add_ps(lane0, lane2);
+  __m128 sum1 = _mm_add_ps(lane1, lane3);
+  return _mm256_insertf128_ps(_mm256_castps128_ps256(sum0), sum1, 1);
+#endif
+}
+template <>
+EIGEN_STRONG_INLINE Packet4d predux_downto4<Packet8d>(const Packet8d& a) {
+  __m256d lane0 = _mm512_extractf64x4_pd(a, 0);
+  __m256d lane1 = _mm512_extractf64x4_pd(a, 1);
+  __m256d res = _mm256_add_pd(lane0, lane1);
+  return res;
+}
+
+template <>
+EIGEN_STRONG_INLINE float predux_mul<Packet16f>(const Packet16f& a) {
+//#ifdef EIGEN_VECTORIZE_AVX512DQ
+#if 0
+  Packet8f lane0 = _mm512_extractf32x8_ps(a, 0);
+  Packet8f lane1 = _mm512_extractf32x8_ps(a, 1);
+  Packet8f res = pmul(lane0, lane1);
+  res = pmul(res, _mm256_permute2f128_ps(res, res, 1));
+  res = pmul(res, _mm_permute_ps(res, _MM_SHUFFLE(0, 0, 3, 2)));
+  return pfirst(pmul(res, _mm_permute_ps(res, _MM_SHUFFLE(0, 0, 0, 1))));
+#else
+  __m128 lane0 = _mm512_extractf32x4_ps(a, 0);
+  __m128 lane1 = _mm512_extractf32x4_ps(a, 1);
+  __m128 lane2 = _mm512_extractf32x4_ps(a, 2);
+  __m128 lane3 = _mm512_extractf32x4_ps(a, 3);
+  __m128 res = pmul(pmul(lane0, lane1), pmul(lane2, lane3));
+  res = pmul(res, _mm_permute_ps(res, _MM_SHUFFLE(0, 0, 3, 2)));
+  return pfirst(pmul(res, _mm_permute_ps(res, _MM_SHUFFLE(0, 0, 0, 1))));
+#endif
+}
+template <>
+EIGEN_STRONG_INLINE double predux_mul<Packet8d>(const Packet8d& a) {
+  __m256d lane0 = _mm512_extractf64x4_pd(a, 0);
+  __m256d lane1 = _mm512_extractf64x4_pd(a, 1);
+  __m256d res = pmul(lane0, lane1);
+  res = pmul(res, _mm256_permute2f128_pd(res, res, 1));
+  return pfirst(pmul(res, _mm256_shuffle_pd(res, res, 1)));
+}
+
+template <>
+EIGEN_STRONG_INLINE float predux_min<Packet16f>(const Packet16f& a) {
+  __m128 lane0 = _mm512_extractf32x4_ps(a, 0);
+  __m128 lane1 = _mm512_extractf32x4_ps(a, 1);
+  __m128 lane2 = _mm512_extractf32x4_ps(a, 2);
+  __m128 lane3 = _mm512_extractf32x4_ps(a, 3);
+  __m128 res = _mm_min_ps(_mm_min_ps(lane0, lane1), _mm_min_ps(lane2, lane3));
+  res = _mm_min_ps(res, _mm_permute_ps(res, _MM_SHUFFLE(0, 0, 3, 2)));
+  return pfirst(_mm_min_ps(res, _mm_permute_ps(res, _MM_SHUFFLE(0, 0, 0, 1))));
+}
+template <>
+EIGEN_STRONG_INLINE double predux_min<Packet8d>(const Packet8d& a) {
+  __m256d lane0 = _mm512_extractf64x4_pd(a, 0);
+  __m256d lane1 = _mm512_extractf64x4_pd(a, 1);
+  __m256d res = _mm256_min_pd(lane0, lane1);
+  res = _mm256_min_pd(res, _mm256_permute2f128_pd(res, res, 1));
+  return pfirst(_mm256_min_pd(res, _mm256_shuffle_pd(res, res, 1)));
+}
+
+template <>
+EIGEN_STRONG_INLINE float predux_max<Packet16f>(const Packet16f& a) {
+  __m128 lane0 = _mm512_extractf32x4_ps(a, 0);
+  __m128 lane1 = _mm512_extractf32x4_ps(a, 1);
+  __m128 lane2 = _mm512_extractf32x4_ps(a, 2);
+  __m128 lane3 = _mm512_extractf32x4_ps(a, 3);
+  __m128 res = _mm_max_ps(_mm_max_ps(lane0, lane1), _mm_max_ps(lane2, lane3));
+  res = _mm_max_ps(res, _mm_permute_ps(res, _MM_SHUFFLE(0, 0, 3, 2)));
+  return pfirst(_mm_max_ps(res, _mm_permute_ps(res, _MM_SHUFFLE(0, 0, 0, 1))));
+}
+
+template <>
+EIGEN_STRONG_INLINE double predux_max<Packet8d>(const Packet8d& a) {
+  __m256d lane0 = _mm512_extractf64x4_pd(a, 0);
+  __m256d lane1 = _mm512_extractf64x4_pd(a, 1);
+  __m256d res = _mm256_max_pd(lane0, lane1);
+  res = _mm256_max_pd(res, _mm256_permute2f128_pd(res, res, 1));
+  return pfirst(_mm256_max_pd(res, _mm256_shuffle_pd(res, res, 1)));
+}
+
+template <int Offset>
+struct palign_impl<Offset, Packet16f> {
+  static EIGEN_STRONG_INLINE void run(Packet16f& first,
+                                      const Packet16f& second) {
+    if (Offset != 0) {
+      __m512i first_idx = _mm512_set_epi32(
+          Offset + 15, Offset + 14, Offset + 13, Offset + 12, Offset + 11,
+          Offset + 10, Offset + 9, Offset + 8, Offset + 7, Offset + 6,
+          Offset + 5, Offset + 4, Offset + 3, Offset + 2, Offset + 1, Offset);
+
+      __m512i second_idx =
+          _mm512_set_epi32(Offset - 1, Offset - 2, Offset - 3, Offset - 4,
+                           Offset - 5, Offset - 6, Offset - 7, Offset - 8,
+                           Offset - 9, Offset - 10, Offset - 11, Offset - 12,
+                           Offset - 13, Offset - 14, Offset - 15, Offset - 16);
+
+      unsigned short mask = 0xFFFF;
+      mask <<= (16 - Offset);
+
+      first = _mm512_permutexvar_ps(first_idx, first);
+      Packet16f tmp = _mm512_permutexvar_ps(second_idx, second);
+      first = _mm512_mask_blend_ps(mask, first, tmp);
+    }
+  }
+};
+template <int Offset>
+struct palign_impl<Offset, Packet8d> {
+  static EIGEN_STRONG_INLINE void run(Packet8d& first, const Packet8d& second) {
+    if (Offset != 0) {
+      __m512i first_idx = _mm512_set_epi32(
+          0, Offset + 7, 0, Offset + 6, 0, Offset + 5, 0, Offset + 4, 0,
+          Offset + 3, 0, Offset + 2, 0, Offset + 1, 0, Offset);
+
+      __m512i second_idx = _mm512_set_epi32(
+          0, Offset - 1, 0, Offset - 2, 0, Offset - 3, 0, Offset - 4, 0,
+          Offset - 5, 0, Offset - 6, 0, Offset - 7, 0, Offset - 8);
+
+      unsigned char mask = 0xFF;
+      mask <<= (8 - Offset);
+
+      first = _mm512_permutexvar_pd(first_idx, first);
+      Packet8d tmp = _mm512_permutexvar_pd(second_idx, second);
+      first = _mm512_mask_blend_pd(mask, first, tmp);
+    }
+  }
+};
+
+
+#define PACK_OUTPUT(OUTPUT, INPUT, INDEX, STRIDE) \
+  EIGEN_INSERT_8f_INTO_16f(OUTPUT[INDEX], INPUT[INDEX], INPUT[INDEX + STRIDE]);
+
+EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet16f, 16>& kernel) {
+  __m512 T0 = _mm512_unpacklo_ps(kernel.packet[0], kernel.packet[1]);
+  __m512 T1 = _mm512_unpackhi_ps(kernel.packet[0], kernel.packet[1]);
+  __m512 T2 = _mm512_unpacklo_ps(kernel.packet[2], kernel.packet[3]);
+  __m512 T3 = _mm512_unpackhi_ps(kernel.packet[2], kernel.packet[3]);
+  __m512 T4 = _mm512_unpacklo_ps(kernel.packet[4], kernel.packet[5]);
+  __m512 T5 = _mm512_unpackhi_ps(kernel.packet[4], kernel.packet[5]);
+  __m512 T6 = _mm512_unpacklo_ps(kernel.packet[6], kernel.packet[7]);
+  __m512 T7 = _mm512_unpackhi_ps(kernel.packet[6], kernel.packet[7]);
+  __m512 T8 = _mm512_unpacklo_ps(kernel.packet[8], kernel.packet[9]);
+  __m512 T9 = _mm512_unpackhi_ps(kernel.packet[8], kernel.packet[9]);
+  __m512 T10 = _mm512_unpacklo_ps(kernel.packet[10], kernel.packet[11]);
+  __m512 T11 = _mm512_unpackhi_ps(kernel.packet[10], kernel.packet[11]);
+  __m512 T12 = _mm512_unpacklo_ps(kernel.packet[12], kernel.packet[13]);
+  __m512 T13 = _mm512_unpackhi_ps(kernel.packet[12], kernel.packet[13]);
+  __m512 T14 = _mm512_unpacklo_ps(kernel.packet[14], kernel.packet[15]);
+  __m512 T15 = _mm512_unpackhi_ps(kernel.packet[14], kernel.packet[15]);
+  __m512 S0 = _mm512_shuffle_ps(T0, T2, _MM_SHUFFLE(1, 0, 1, 0));
+  __m512 S1 = _mm512_shuffle_ps(T0, T2, _MM_SHUFFLE(3, 2, 3, 2));
+  __m512 S2 = _mm512_shuffle_ps(T1, T3, _MM_SHUFFLE(1, 0, 1, 0));
+  __m512 S3 = _mm512_shuffle_ps(T1, T3, _MM_SHUFFLE(3, 2, 3, 2));
+  __m512 S4 = _mm512_shuffle_ps(T4, T6, _MM_SHUFFLE(1, 0, 1, 0));
+  __m512 S5 = _mm512_shuffle_ps(T4, T6, _MM_SHUFFLE(3, 2, 3, 2));
+  __m512 S6 = _mm512_shuffle_ps(T5, T7, _MM_SHUFFLE(1, 0, 1, 0));
+  __m512 S7 = _mm512_shuffle_ps(T5, T7, _MM_SHUFFLE(3, 2, 3, 2));
+  __m512 S8 = _mm512_shuffle_ps(T8, T10, _MM_SHUFFLE(1, 0, 1, 0));
+  __m512 S9 = _mm512_shuffle_ps(T8, T10, _MM_SHUFFLE(3, 2, 3, 2));
+  __m512 S10 = _mm512_shuffle_ps(T9, T11, _MM_SHUFFLE(1, 0, 1, 0));
+  __m512 S11 = _mm512_shuffle_ps(T9, T11, _MM_SHUFFLE(3, 2, 3, 2));
+  __m512 S12 = _mm512_shuffle_ps(T12, T14, _MM_SHUFFLE(1, 0, 1, 0));
+  __m512 S13 = _mm512_shuffle_ps(T12, T14, _MM_SHUFFLE(3, 2, 3, 2));
+  __m512 S14 = _mm512_shuffle_ps(T13, T15, _MM_SHUFFLE(1, 0, 1, 0));
+  __m512 S15 = _mm512_shuffle_ps(T13, T15, _MM_SHUFFLE(3, 2, 3, 2));
+
+  EIGEN_EXTRACT_8f_FROM_16f(S0, S0);
+  EIGEN_EXTRACT_8f_FROM_16f(S1, S1);
+  EIGEN_EXTRACT_8f_FROM_16f(S2, S2);
+  EIGEN_EXTRACT_8f_FROM_16f(S3, S3);
+  EIGEN_EXTRACT_8f_FROM_16f(S4, S4);
+  EIGEN_EXTRACT_8f_FROM_16f(S5, S5);
+  EIGEN_EXTRACT_8f_FROM_16f(S6, S6);
+  EIGEN_EXTRACT_8f_FROM_16f(S7, S7);
+  EIGEN_EXTRACT_8f_FROM_16f(S8, S8);
+  EIGEN_EXTRACT_8f_FROM_16f(S9, S9);
+  EIGEN_EXTRACT_8f_FROM_16f(S10, S10);
+  EIGEN_EXTRACT_8f_FROM_16f(S11, S11);
+  EIGEN_EXTRACT_8f_FROM_16f(S12, S12);
+  EIGEN_EXTRACT_8f_FROM_16f(S13, S13);
+  EIGEN_EXTRACT_8f_FROM_16f(S14, S14);
+  EIGEN_EXTRACT_8f_FROM_16f(S15, S15);
+
+  PacketBlock<Packet8f, 32> tmp;
+
+  tmp.packet[0] = _mm256_permute2f128_ps(S0_0, S4_0, 0x20);
+  tmp.packet[1] = _mm256_permute2f128_ps(S1_0, S5_0, 0x20);
+  tmp.packet[2] = _mm256_permute2f128_ps(S2_0, S6_0, 0x20);
+  tmp.packet[3] = _mm256_permute2f128_ps(S3_0, S7_0, 0x20);
+  tmp.packet[4] = _mm256_permute2f128_ps(S0_0, S4_0, 0x31);
+  tmp.packet[5] = _mm256_permute2f128_ps(S1_0, S5_0, 0x31);
+  tmp.packet[6] = _mm256_permute2f128_ps(S2_0, S6_0, 0x31);
+  tmp.packet[7] = _mm256_permute2f128_ps(S3_0, S7_0, 0x31);
+
+  tmp.packet[8] = _mm256_permute2f128_ps(S0_1, S4_1, 0x20);
+  tmp.packet[9] = _mm256_permute2f128_ps(S1_1, S5_1, 0x20);
+  tmp.packet[10] = _mm256_permute2f128_ps(S2_1, S6_1, 0x20);
+  tmp.packet[11] = _mm256_permute2f128_ps(S3_1, S7_1, 0x20);
+  tmp.packet[12] = _mm256_permute2f128_ps(S0_1, S4_1, 0x31);
+  tmp.packet[13] = _mm256_permute2f128_ps(S1_1, S5_1, 0x31);
+  tmp.packet[14] = _mm256_permute2f128_ps(S2_1, S6_1, 0x31);
+  tmp.packet[15] = _mm256_permute2f128_ps(S3_1, S7_1, 0x31);
+
+  // Second set of _m256 outputs
+  tmp.packet[16] = _mm256_permute2f128_ps(S8_0, S12_0, 0x20);
+  tmp.packet[17] = _mm256_permute2f128_ps(S9_0, S13_0, 0x20);
+  tmp.packet[18] = _mm256_permute2f128_ps(S10_0, S14_0, 0x20);
+  tmp.packet[19] = _mm256_permute2f128_ps(S11_0, S15_0, 0x20);
+  tmp.packet[20] = _mm256_permute2f128_ps(S8_0, S12_0, 0x31);
+  tmp.packet[21] = _mm256_permute2f128_ps(S9_0, S13_0, 0x31);
+  tmp.packet[22] = _mm256_permute2f128_ps(S10_0, S14_0, 0x31);
+  tmp.packet[23] = _mm256_permute2f128_ps(S11_0, S15_0, 0x31);
+
+  tmp.packet[24] = _mm256_permute2f128_ps(S8_1, S12_1, 0x20);
+  tmp.packet[25] = _mm256_permute2f128_ps(S9_1, S13_1, 0x20);
+  tmp.packet[26] = _mm256_permute2f128_ps(S10_1, S14_1, 0x20);
+  tmp.packet[27] = _mm256_permute2f128_ps(S11_1, S15_1, 0x20);
+  tmp.packet[28] = _mm256_permute2f128_ps(S8_1, S12_1, 0x31);
+  tmp.packet[29] = _mm256_permute2f128_ps(S9_1, S13_1, 0x31);
+  tmp.packet[30] = _mm256_permute2f128_ps(S10_1, S14_1, 0x31);
+  tmp.packet[31] = _mm256_permute2f128_ps(S11_1, S15_1, 0x31);
+
+  // Pack them into the output
+  PACK_OUTPUT(kernel.packet, tmp.packet, 0, 16);
+  PACK_OUTPUT(kernel.packet, tmp.packet, 1, 16);
+  PACK_OUTPUT(kernel.packet, tmp.packet, 2, 16);
+  PACK_OUTPUT(kernel.packet, tmp.packet, 3, 16);
+
+  PACK_OUTPUT(kernel.packet, tmp.packet, 4, 16);
+  PACK_OUTPUT(kernel.packet, tmp.packet, 5, 16);
+  PACK_OUTPUT(kernel.packet, tmp.packet, 6, 16);
+  PACK_OUTPUT(kernel.packet, tmp.packet, 7, 16);
+
+  PACK_OUTPUT(kernel.packet, tmp.packet, 8, 16);
+  PACK_OUTPUT(kernel.packet, tmp.packet, 9, 16);
+  PACK_OUTPUT(kernel.packet, tmp.packet, 10, 16);
+  PACK_OUTPUT(kernel.packet, tmp.packet, 11, 16);
+
+  PACK_OUTPUT(kernel.packet, tmp.packet, 12, 16);
+  PACK_OUTPUT(kernel.packet, tmp.packet, 13, 16);
+  PACK_OUTPUT(kernel.packet, tmp.packet, 14, 16);
+  PACK_OUTPUT(kernel.packet, tmp.packet, 15, 16);
+}
+#define PACK_OUTPUT_2(OUTPUT, INPUT, INDEX, STRIDE)         \
+  EIGEN_INSERT_8f_INTO_16f(OUTPUT[INDEX], INPUT[2 * INDEX], \
+                           INPUT[2 * INDEX + STRIDE]);
+
+EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet16f, 4>& kernel) {
+  __m512 T0 = _mm512_unpacklo_ps(kernel.packet[0], kernel.packet[1]);
+  __m512 T1 = _mm512_unpackhi_ps(kernel.packet[0], kernel.packet[1]);
+  __m512 T2 = _mm512_unpacklo_ps(kernel.packet[2], kernel.packet[3]);
+  __m512 T3 = _mm512_unpackhi_ps(kernel.packet[2], kernel.packet[3]);
+
+  __m512 S0 = _mm512_shuffle_ps(T0, T2, _MM_SHUFFLE(1, 0, 1, 0));
+  __m512 S1 = _mm512_shuffle_ps(T0, T2, _MM_SHUFFLE(3, 2, 3, 2));
+  __m512 S2 = _mm512_shuffle_ps(T1, T3, _MM_SHUFFLE(1, 0, 1, 0));
+  __m512 S3 = _mm512_shuffle_ps(T1, T3, _MM_SHUFFLE(3, 2, 3, 2));
+
+  EIGEN_EXTRACT_8f_FROM_16f(S0, S0);
+  EIGEN_EXTRACT_8f_FROM_16f(S1, S1);
+  EIGEN_EXTRACT_8f_FROM_16f(S2, S2);
+  EIGEN_EXTRACT_8f_FROM_16f(S3, S3);
+
+  PacketBlock<Packet8f, 8> tmp;
+
+  tmp.packet[0] = _mm256_permute2f128_ps(S0_0, S1_0, 0x20);
+  tmp.packet[1] = _mm256_permute2f128_ps(S2_0, S3_0, 0x20);
+  tmp.packet[2] = _mm256_permute2f128_ps(S0_0, S1_0, 0x31);
+  tmp.packet[3] = _mm256_permute2f128_ps(S2_0, S3_0, 0x31);
+
+  tmp.packet[4] = _mm256_permute2f128_ps(S0_1, S1_1, 0x20);
+  tmp.packet[5] = _mm256_permute2f128_ps(S2_1, S3_1, 0x20);
+  tmp.packet[6] = _mm256_permute2f128_ps(S0_1, S1_1, 0x31);
+  tmp.packet[7] = _mm256_permute2f128_ps(S2_1, S3_1, 0x31);
+
+  PACK_OUTPUT_2(kernel.packet, tmp.packet, 0, 1);
+  PACK_OUTPUT_2(kernel.packet, tmp.packet, 1, 1);
+  PACK_OUTPUT_2(kernel.packet, tmp.packet, 2, 1);
+  PACK_OUTPUT_2(kernel.packet, tmp.packet, 3, 1);
+}
+
+#define PACK_OUTPUT_SQ_D(OUTPUT, INPUT, INDEX, STRIDE)                \
+  OUTPUT[INDEX] = _mm512_insertf64x4(OUTPUT[INDEX], INPUT[INDEX], 0); \
+  OUTPUT[INDEX] = _mm512_insertf64x4(OUTPUT[INDEX], INPUT[INDEX + STRIDE], 1);
+
+#define PACK_OUTPUT_D(OUTPUT, INPUT, INDEX, STRIDE)                         \
+  OUTPUT[INDEX] = _mm512_insertf64x4(OUTPUT[INDEX], INPUT[(2 * INDEX)], 0); \
+  OUTPUT[INDEX] =                                                           \
+      _mm512_insertf64x4(OUTPUT[INDEX], INPUT[(2 * INDEX) + STRIDE], 1);
+
+EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8d, 4>& kernel) {
+  __m512d T0 = _mm512_shuffle_pd(kernel.packet[0], kernel.packet[1], 0);
+  __m512d T1 = _mm512_shuffle_pd(kernel.packet[0], kernel.packet[1], 0xff);
+  __m512d T2 = _mm512_shuffle_pd(kernel.packet[2], kernel.packet[3], 0);
+  __m512d T3 = _mm512_shuffle_pd(kernel.packet[2], kernel.packet[3], 0xff);
+
+  PacketBlock<Packet4d, 8> tmp;
+
+  tmp.packet[0] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T0, 0),
+                                         _mm512_extractf64x4_pd(T2, 0), 0x20);
+  tmp.packet[1] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T1, 0),
+                                         _mm512_extractf64x4_pd(T3, 0), 0x20);
+  tmp.packet[2] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T0, 0),
+                                         _mm512_extractf64x4_pd(T2, 0), 0x31);
+  tmp.packet[3] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T1, 0),
+                                         _mm512_extractf64x4_pd(T3, 0), 0x31);
+
+  tmp.packet[4] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T0, 1),
+                                         _mm512_extractf64x4_pd(T2, 1), 0x20);
+  tmp.packet[5] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T1, 1),
+                                         _mm512_extractf64x4_pd(T3, 1), 0x20);
+  tmp.packet[6] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T0, 1),
+                                         _mm512_extractf64x4_pd(T2, 1), 0x31);
+  tmp.packet[7] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T1, 1),
+                                         _mm512_extractf64x4_pd(T3, 1), 0x31);
+
+  PACK_OUTPUT_D(kernel.packet, tmp.packet, 0, 1);
+  PACK_OUTPUT_D(kernel.packet, tmp.packet, 1, 1);
+  PACK_OUTPUT_D(kernel.packet, tmp.packet, 2, 1);
+  PACK_OUTPUT_D(kernel.packet, tmp.packet, 3, 1);
+}
+
+EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8d, 8>& kernel) {
+  __m512d T0 = _mm512_unpacklo_pd(kernel.packet[0], kernel.packet[1]);
+  __m512d T1 = _mm512_unpackhi_pd(kernel.packet[0], kernel.packet[1]);
+  __m512d T2 = _mm512_unpacklo_pd(kernel.packet[2], kernel.packet[3]);
+  __m512d T3 = _mm512_unpackhi_pd(kernel.packet[2], kernel.packet[3]);
+  __m512d T4 = _mm512_unpacklo_pd(kernel.packet[4], kernel.packet[5]);
+  __m512d T5 = _mm512_unpackhi_pd(kernel.packet[4], kernel.packet[5]);
+  __m512d T6 = _mm512_unpacklo_pd(kernel.packet[6], kernel.packet[7]);
+  __m512d T7 = _mm512_unpackhi_pd(kernel.packet[6], kernel.packet[7]);
+
+  PacketBlock<Packet4d, 16> tmp;
+
+  tmp.packet[0] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T0, 0),
+                                         _mm512_extractf64x4_pd(T2, 0), 0x20);
+  tmp.packet[1] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T1, 0),
+                                         _mm512_extractf64x4_pd(T3, 0), 0x20);
+  tmp.packet[2] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T0, 0),
+                                         _mm512_extractf64x4_pd(T2, 0), 0x31);
+  tmp.packet[3] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T1, 0),
+                                         _mm512_extractf64x4_pd(T3, 0), 0x31);
+
+  tmp.packet[4] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T0, 1),
+                                         _mm512_extractf64x4_pd(T2, 1), 0x20);
+  tmp.packet[5] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T1, 1),
+                                         _mm512_extractf64x4_pd(T3, 1), 0x20);
+  tmp.packet[6] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T0, 1),
+                                         _mm512_extractf64x4_pd(T2, 1), 0x31);
+  tmp.packet[7] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T1, 1),
+                                         _mm512_extractf64x4_pd(T3, 1), 0x31);
+
+  tmp.packet[8] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T4, 0),
+                                         _mm512_extractf64x4_pd(T6, 0), 0x20);
+  tmp.packet[9] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T5, 0),
+                                         _mm512_extractf64x4_pd(T7, 0), 0x20);
+  tmp.packet[10] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T4, 0),
+                                          _mm512_extractf64x4_pd(T6, 0), 0x31);
+  tmp.packet[11] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T5, 0),
+                                          _mm512_extractf64x4_pd(T7, 0), 0x31);
+
+  tmp.packet[12] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T4, 1),
+                                          _mm512_extractf64x4_pd(T6, 1), 0x20);
+  tmp.packet[13] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T5, 1),
+                                          _mm512_extractf64x4_pd(T7, 1), 0x20);
+  tmp.packet[14] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T4, 1),
+                                          _mm512_extractf64x4_pd(T6, 1), 0x31);
+  tmp.packet[15] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T5, 1),
+                                          _mm512_extractf64x4_pd(T7, 1), 0x31);
+
+  PACK_OUTPUT_SQ_D(kernel.packet, tmp.packet, 0, 8);
+  PACK_OUTPUT_SQ_D(kernel.packet, tmp.packet, 1, 8);
+  PACK_OUTPUT_SQ_D(kernel.packet, tmp.packet, 2, 8);
+  PACK_OUTPUT_SQ_D(kernel.packet, tmp.packet, 3, 8);
+
+  PACK_OUTPUT_SQ_D(kernel.packet, tmp.packet, 4, 8);
+  PACK_OUTPUT_SQ_D(kernel.packet, tmp.packet, 5, 8);
+  PACK_OUTPUT_SQ_D(kernel.packet, tmp.packet, 6, 8);
+  PACK_OUTPUT_SQ_D(kernel.packet, tmp.packet, 7, 8);
+}
+template <>
+EIGEN_STRONG_INLINE Packet16f pblend(const Selector<16>& /*ifPacket*/,
+                                     const Packet16f& /*thenPacket*/,
+                                     const Packet16f& /*elsePacket*/) {
+  assert(false && "To be implemented");
+  return Packet16f();
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pblend(const Selector<8>& /*ifPacket*/,
+                                    const Packet8d& /*thenPacket*/,
+                                    const Packet8d& /*elsePacket*/) {
+  assert(false && "To be implemented");
+  return Packet8d();
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_PACKET_MATH_AVX512_H
diff --git a/eigen/Eigen/src/Core/arch/AltiVec/CMakeLists.txt b/eigen/Eigen/src/Core/arch/AltiVec/CMakeLists.txt
deleted file mode 100644
index 9f8d2e9..0000000
--- a/eigen/Eigen/src/Core/arch/AltiVec/CMakeLists.txt
+++ /dev/null
@@ -1,6 +0,0 @@
-FILE(GLOB Eigen_Core_arch_AltiVec_SRCS "*.h")
-
-INSTALL(FILES
-  ${Eigen_Core_arch_AltiVec_SRCS}
-  DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/arch/AltiVec COMPONENT Devel
-)
diff --git a/eigen/Eigen/src/Core/arch/AltiVec/Complex.h b/eigen/Eigen/src/Core/arch/AltiVec/Complex.h
index 68d9a2b..67db2f8 100644
--- a/eigen/Eigen/src/Core/arch/AltiVec/Complex.h
+++ b/eigen/Eigen/src/Core/arch/AltiVec/Complex.h
@@ -2,30 +2,34 @@
 // for linear algebra.
 //
 // Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2010-2016 Konstantinos Margaritis <markos@freevec.org>
 //
 // This Source Code Form is subject to the terms of the Mozilla
 // Public License v. 2.0. If a copy of the MPL was not distributed
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
-#ifndef EIGEN_COMPLEX_ALTIVEC_H
-#define EIGEN_COMPLEX_ALTIVEC_H
+#ifndef EIGEN_COMPLEX32_ALTIVEC_H
+#define EIGEN_COMPLEX32_ALTIVEC_H
 
 namespace Eigen {
 
 namespace internal {
 
-static Packet4ui  p4ui_CONJ_XOR = vec_mergeh((Packet4ui)p4i_ZERO, (Packet4ui)p4f_ZERO_);//{ 0x00000000, 0x80000000, 0x00000000, 0x80000000 };
-static Packet16uc p16uc_COMPLEX_RE   = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 };
-static Packet16uc p16uc_COMPLEX_IM   = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 1), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 };
-static Packet16uc p16uc_COMPLEX_REV  = vec_sld(p16uc_REVERSE, p16uc_REVERSE, 8);//{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 };
-static Packet16uc p16uc_COMPLEX_REV2 = vec_sld(p16uc_FORWARD, p16uc_FORWARD, 8);//{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
-static Packet16uc p16uc_PSET_HI = (Packet16uc) vec_mergeh((Packet4ui) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet4ui) vec_splat((Packet4ui)p16uc_FORWARD, 1));//{ 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 };
-static Packet16uc p16uc_PSET_LO = (Packet16uc) vec_mergeh((Packet4ui) vec_splat((Packet4ui)p16uc_FORWARD, 2), (Packet4ui) vec_splat((Packet4ui)p16uc_FORWARD, 3));//{ 8,9,10,11, 12,13,14,15, 8,9,10,11, 12,13,14,15 };
+static Packet4ui  p4ui_CONJ_XOR = vec_mergeh((Packet4ui)p4i_ZERO, (Packet4ui)p4f_MZERO);//{ 0x00000000, 0x80000000, 0x00000000, 0x80000000 };
+#ifdef __VSX__
+#if defined(_BIG_ENDIAN)
+static Packet2ul  p2ul_CONJ_XOR1 = (Packet2ul) vec_sld((Packet4ui) p2d_MZERO, (Packet4ui) p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 };
+static Packet2ul  p2ul_CONJ_XOR2 = (Packet2ul) vec_sld((Packet4ui) p2l_ZERO,  (Packet4ui) p2d_MZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 };
+#else
+static Packet2ul  p2ul_CONJ_XOR1 = (Packet2ul) vec_sld((Packet4ui) p2l_ZERO,  (Packet4ui) p2d_MZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 };
+static Packet2ul  p2ul_CONJ_XOR2 = (Packet2ul) vec_sld((Packet4ui) p2d_MZERO, (Packet4ui) p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 };
+#endif
+#endif
 
 //---------- float ----------
 struct Packet2cf
 {
-  EIGEN_STRONG_INLINE Packet2cf() {}
+  EIGEN_STRONG_INLINE explicit Packet2cf() : v(p4f_ZERO) {}
   EIGEN_STRONG_INLINE explicit Packet2cf(const Packet4f& a) : v(a) {}
   Packet4f  v;
 };
@@ -33,10 +37,12 @@ struct Packet2cf
 template<> struct packet_traits<std::complex<float> >  : default_packet_traits
 {
   typedef Packet2cf type;
+  typedef Packet2cf half;
   enum {
     Vectorizable = 1,
     AlignedOnScalar = 1,
     size = 2,
+    HasHalfPacket = 0,
 
     HasAdd    = 1,
     HasSub    = 1,
@@ -47,65 +53,78 @@ template<> struct packet_traits<std::complex<float> >  : default_packet_traits
     HasAbs2   = 0,
     HasMin    = 0,
     HasMax    = 0,
+#ifdef __VSX__
+    HasBlend  = 1,
+#endif
     HasSetLinear = 0
   };
 };
 
-template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2}; };
+template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16}; typedef Packet2cf half; };
 
 template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>&  from)
 {
   Packet2cf res;
-  /* On AltiVec we cannot load 64-bit registers, so wa have to take care of alignment */
-  if((ptrdiff_t(&from) % 16) == 0)
+  if((std::ptrdiff_t(&from) % 16) == 0)
     res.v = pload<Packet4f>((const float *)&from);
   else
     res.v = ploadu<Packet4f>((const float *)&from);
-  res.v = vec_perm(res.v, res.v, p16uc_PSET_HI);
+  res.v = vec_perm(res.v, res.v, p16uc_PSET64_HI);
   return res;
 }
 
-template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_add(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_sub(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pload<Packet2cf>(const std::complex<float>*        from) { return Packet2cf(pload<Packet4f>((const float *) from)); }
+template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>*       from) { return Packet2cf(ploadu<Packet4f>((const float*) from)); }
+template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>*     from) { return pset1<Packet2cf>(*from); }
+
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> *   to, const Packet2cf& from) { pstore((float*)to, from.v); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> *   to, const Packet2cf& from) { pstoreu((float*)to, from.v); }
+
+template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from, Index stride)
+{
+  std::complex<float> EIGEN_ALIGN16 af[2];
+  af[0] = from[0*stride];
+  af[1] = from[1*stride];
+  return pload<Packet2cf>(af);
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, Index stride)
+{
+  std::complex<float> EIGEN_ALIGN16 af[2];
+  pstore<std::complex<float> >((std::complex<float> *) af, from);
+  to[0*stride] = af[0];
+  to[1*stride] = af[1];
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(a.v + b.v); }
+template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(a.v - b.v); }
 template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) { return Packet2cf(pnegate(a.v)); }
-template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a) { return Packet2cf((Packet4f)vec_xor((Packet4ui)a.v, p4ui_CONJ_XOR)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a) { return Packet2cf(pxor<Packet4f>(a.v, reinterpret_cast<Packet4f>(p4ui_CONJ_XOR))); }
 
 template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
 {
   Packet4f v1, v2;
 
   // Permute and multiply the real parts of a and b
-  v1 = vec_perm(a.v, a.v, p16uc_COMPLEX_RE);
+  v1 = vec_perm(a.v, a.v, p16uc_PSET32_WODD);
   // Get the imaginary parts of a
-  v2 = vec_perm(a.v, a.v, p16uc_COMPLEX_IM);
+  v2 = vec_perm(a.v, a.v, p16uc_PSET32_WEVEN);
   // multiply a_re * b 
   v1 = vec_madd(v1, b.v, p4f_ZERO);
   // multiply a_im * b and get the conjugate result
   v2 = vec_madd(v2, b.v, p4f_ZERO);
-  v2 = (Packet4f) vec_xor((Packet4ui)v2, p4ui_CONJ_XOR);
+  v2 = reinterpret_cast<Packet4f>(pxor(v2, reinterpret_cast<Packet4f>(p4ui_CONJ_XOR)));
   // permute back to a proper order
-  v2 = vec_perm(v2, v2, p16uc_COMPLEX_REV);
+  v2 = vec_perm(v2, v2, p16uc_COMPLEX32_REV);
   
-  return Packet2cf(vec_add(v1, v2));
+  return Packet2cf(padd<Packet4f>(v1, v2));
 }
 
-template<> EIGEN_STRONG_INLINE Packet2cf pand   <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_and(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet2cf por    <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_or(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet2cf pxor   <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_xor(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_and(a.v, vec_nor(b.v,b.v))); }
+template<> EIGEN_STRONG_INLINE Packet2cf pand   <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pand<Packet4f>(a.v, b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf por    <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(por<Packet4f>(a.v, b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pxor   <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pxor<Packet4f>(a.v, b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pandnot<Packet4f>(a.v, b.v)); }
 
-template<> EIGEN_STRONG_INLINE Packet2cf pload <Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>((const float*)from)); }
-template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>((const float*)from)); }
-
-template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>*     from)
-{
-  return pset1<Packet2cf>(*from);
-}
-
-template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> *   to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((float*)to, from.v); }
-template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> *   to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((float*)to, from.v); }
-
-template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> *   addr) { vec_dstt((float *)addr, DST_CTRL(2,2,32), DST_CHAN); }
+template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> * addr)    { EIGEN_PPC_PREFETCH(addr); }
 
 template<> EIGEN_STRONG_INLINE std::complex<float>  pfirst<Packet2cf>(const Packet2cf& a)
 {
@@ -118,26 +137,30 @@ template<> EIGEN_STRONG_INLINE std::complex<float>  pfirst<Packet2cf>(const Pack
 template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a)
 {
   Packet4f rev_a;
-  rev_a = vec_perm(a.v, a.v, p16uc_COMPLEX_REV2);
+  rev_a = vec_perm(a.v, a.v, p16uc_COMPLEX32_REV2);
   return Packet2cf(rev_a);
 }
 
 template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a)
 {
   Packet4f b;
-  b = (Packet4f) vec_sld(a.v, a.v, 8);
-  b = padd(a.v, b);
-  return pfirst(Packet2cf(b));
+  b = vec_sld(a.v, a.v, 8);
+  b = padd<Packet4f>(a.v, b);
+  return pfirst<Packet2cf>(Packet2cf(b));
 }
 
 template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs)
 {
   Packet4f b1, b2;
-  
-  b1 = (Packet4f) vec_sld(vecs[0].v, vecs[1].v, 8);
-  b2 = (Packet4f) vec_sld(vecs[1].v, vecs[0].v, 8);
-  b2 = (Packet4f) vec_sld(b2, b2, 8);
-  b2 = padd(b1, b2);
+#ifdef _BIG_ENDIAN  
+  b1 = vec_sld(vecs[0].v, vecs[1].v, 8);
+  b2 = vec_sld(vecs[1].v, vecs[0].v, 8);
+#else
+  b1 = vec_sld(vecs[1].v, vecs[0].v, 8);
+  b2 = vec_sld(vecs[0].v, vecs[1].v, 8);
+#endif
+  b2 = vec_sld(b2, b2, 8);
+  b2 = padd<Packet4f>(b1, b2);
 
   return Packet2cf(b2);
 }
@@ -146,10 +169,10 @@ template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const P
 {
   Packet4f b;
   Packet2cf prod;
-  b = (Packet4f) vec_sld(a.v, a.v, 8);
-  prod = pmul(a, Packet2cf(b));
+  b = vec_sld(a.v, a.v, 8);
+  prod = pmul<Packet2cf>(a, Packet2cf(b));
 
-  return pfirst(prod);
+  return pfirst<Packet2cf>(prod);
 }
 
 template<int Offset>
@@ -159,7 +182,11 @@ struct palign_impl<Offset,Packet2cf>
   {
     if (Offset==1)
     {
+#ifdef _BIG_ENDIAN
       first.v = vec_sld(first.v, second.v, 8);
+#else
+      first.v = vec_sld(second.v, first.v, 8);
+#endif
     }
   }
 };
@@ -197,21 +224,238 @@ template<> struct conj_helper<Packet2cf, Packet2cf, true,true>
   }
 };
 
+template<> struct conj_helper<Packet4f, Packet2cf, false,false>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet4f& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet4f& x, const Packet2cf& y) const
+  { return Packet2cf(internal::pmul<Packet4f>(x, y.v)); }
+};
+
+template<> struct conj_helper<Packet2cf, Packet4f, false,false>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet4f& y, const Packet2cf& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& x, const Packet4f& y) const
+  { return Packet2cf(internal::pmul<Packet4f>(x.v, y)); }
+};
+
 template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
 {
   // TODO optimize it for AltiVec
-  Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a,b);
-  Packet4f s = vec_madd(b.v, b.v, p4f_ZERO);
-  return Packet2cf(pdiv(res.v, vec_add(s,vec_perm(s, s, p16uc_COMPLEX_REV))));
+  Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a, b);
+  Packet4f s = pmul<Packet4f>(b.v, b.v);
+  return Packet2cf(pdiv(res.v, padd<Packet4f>(s, vec_perm(s, s, p16uc_COMPLEX32_REV))));
 }
 
 template<> EIGEN_STRONG_INLINE Packet2cf pcplxflip<Packet2cf>(const Packet2cf& x)
 {
-  return Packet2cf(vec_perm(x.v, x.v, p16uc_COMPLEX_REV));
+  return Packet2cf(vec_perm(x.v, x.v, p16uc_COMPLEX32_REV));
+}
+
+EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2cf,2>& kernel)
+{
+  Packet4f tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_HI);
+  kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO);
+  kernel.packet[0].v = tmp;
+}
+
+#ifdef __VSX__
+template<> EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, const Packet2cf& thenPacket, const Packet2cf& elsePacket) {
+  Packet2cf result;
+  result.v = reinterpret_cast<Packet4f>(pblend<Packet2d>(ifPacket, reinterpret_cast<Packet2d>(thenPacket.v), reinterpret_cast<Packet2d>(elsePacket.v)));
+  return result;
+}
+#endif
+
+//---------- double ----------
+#ifdef __VSX__
+struct Packet1cd
+{
+  EIGEN_STRONG_INLINE Packet1cd() {}
+  EIGEN_STRONG_INLINE explicit Packet1cd(const Packet2d& a) : v(a) {}
+  Packet2d v;
+};
+
+template<> struct packet_traits<std::complex<double> >  : default_packet_traits
+{
+  typedef Packet1cd type;
+  typedef Packet1cd half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 0,
+    size = 1,
+    HasHalfPacket = 0,
+
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasDiv    = 1,
+    HasNegate = 1,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasSetLinear = 0
+  };
+};
+
+template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16}; typedef Packet1cd half; };
+
+template<> EIGEN_STRONG_INLINE Packet1cd pload <Packet1cd>(const std::complex<double>* from) { return Packet1cd(pload<Packet2d>((const double*)from)); }
+template<> EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from) { return Packet1cd(ploadu<Packet2d>((const double*)from)); }
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { pstore((double*)to, from.v); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { pstoreu((double*)to, from.v); }
+
+template<> EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<double>&  from)
+{ /* here we really have to use unaligned loads :( */ return ploadu<Packet1cd>(&from); }
+
+template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather<std::complex<double>, Packet1cd>(const std::complex<double>* from, Index stride)
+{
+  std::complex<double> EIGEN_ALIGN16 af[2];
+  af[0] = from[0*stride];
+  af[1] = from[1*stride];
+  return pload<Packet1cd>(af);
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet1cd>(std::complex<double>* to, const Packet1cd& from, Index stride)
+{
+  std::complex<double> EIGEN_ALIGN16 af[2];
+  pstore<std::complex<double> >(af, from);
+  to[0*stride] = af[0];
+  to[1*stride] = af[1];
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(a.v + b.v); }
+template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(a.v - b.v); }
+template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(Packet2d(a.v))); }
+template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) { return Packet1cd(pxor(a.v, reinterpret_cast<Packet2d>(p2ul_CONJ_XOR2))); }
+
+template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
+{
+  Packet2d a_re, a_im, v1, v2;
+
+  // Permute and multiply the real parts of a and b
+  a_re = vec_perm(a.v, a.v, p16uc_PSET64_HI);
+  // Get the imaginary parts of a
+  a_im = vec_perm(a.v, a.v, p16uc_PSET64_LO);
+  // multiply a_re * b
+  v1 = vec_madd(a_re, b.v, p2d_ZERO);
+  // multiply a_im * b and get the conjugate result
+  v2 = vec_madd(a_im, b.v, p2d_ZERO);
+  v2 = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(v2), reinterpret_cast<Packet4ui>(v2), 8));
+  v2 = pxor(v2, reinterpret_cast<Packet2d>(p2ul_CONJ_XOR1));
+
+  return Packet1cd(padd<Packet2d>(v1, v2));
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd pand   <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(pand(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd por    <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(por(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pxor   <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(pxor(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(pandnot(a.v, b.v)); }
+
+template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>*     from)  { return pset1<Packet1cd>(*from); }
+
+template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> * addr)    { EIGEN_PPC_PREFETCH(addr); }
+
+template<> EIGEN_STRONG_INLINE std::complex<double>  pfirst<Packet1cd>(const Packet1cd& a)
+{
+  std::complex<double> EIGEN_ALIGN16 res[2];
+  pstore<std::complex<double> >(res, a);
+
+  return res[0];
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; }
+
+template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a) { return pfirst(a); }
+template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs)        { return vecs[0]; }
+
+template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a) { return pfirst(a); }
+
+template<int Offset>
+struct palign_impl<Offset,Packet1cd>
+{
+  static EIGEN_STRONG_INLINE void run(Packet1cd& /*first*/, const Packet1cd& /*second*/)
+  {
+    // FIXME is it sure we never have to align a Packet1cd?
+    // Even though a std::complex<double> has 16 bytes, it is not necessarily aligned on a 16 bytes boundary...
+  }
+};
+
+template<> struct conj_helper<Packet1cd, Packet1cd, false,true>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
+  {
+    return internal::pmul(a, pconj(b));
+  }
+};
+
+template<> struct conj_helper<Packet1cd, Packet1cd, true,false>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
+  {
+    return internal::pmul(pconj(a), b);
+  }
+};
+
+template<> struct conj_helper<Packet1cd, Packet1cd, true,true>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
+  {
+    return pconj(internal::pmul(a, b));
+  }
+};
+template<> struct conj_helper<Packet2d, Packet1cd, false,false>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet2d& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet2d& x, const Packet1cd& y) const
+  { return Packet1cd(internal::pmul<Packet2d>(x, y.v)); }
+};
+
+template<> struct conj_helper<Packet1cd, Packet2d, false,false>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet2d& y, const Packet1cd& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& x, const Packet2d& y) const
+  { return Packet1cd(internal::pmul<Packet2d>(x.v, y)); }
+};
+
+template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
+{
+  // TODO optimize it for AltiVec
+  Packet1cd res = conj_helper<Packet1cd,Packet1cd,false,true>().pmul(a,b);
+  Packet2d s = pmul<Packet2d>(b.v, b.v);
+  return Packet1cd(pdiv(res.v, padd<Packet2d>(s, vec_perm(s, s, p16uc_REVERSE64))));
 }
 
+EIGEN_STRONG_INLINE Packet1cd pcplxflip/*<Packet1cd>*/(const Packet1cd& x)
+{
+  return Packet1cd(preverse(Packet2d(x.v)));
+}
+
+EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet1cd,2>& kernel)
+{
+  Packet2d tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_HI);
+  kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO);
+  kernel.packet[0].v = tmp;
+}
+#endif // __VSX__
 } // end namespace internal
 
 } // end namespace Eigen
 
-#endif // EIGEN_COMPLEX_ALTIVEC_H
+#endif // EIGEN_COMPLEX32_ALTIVEC_H
diff --git a/eigen/Eigen/src/Core/arch/AltiVec/MathFunctions.h b/eigen/Eigen/src/Core/arch/AltiVec/MathFunctions.h
new file mode 100644
index 0000000..c5e4bed
--- /dev/null
+++ b/eigen/Eigen/src/Core/arch/AltiVec/MathFunctions.h
@@ -0,0 +1,322 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2007 Julien Pommier
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2016 Konstantinos Margaritis <markos@freevec.org>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/* The sin, cos, exp, and log functions of this file come from
+ * Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/
+ */
+
+#ifndef EIGEN_MATH_FUNCTIONS_ALTIVEC_H
+#define EIGEN_MATH_FUNCTIONS_ALTIVEC_H
+
+namespace Eigen {
+
+namespace internal {
+
+static _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f);
+static _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
+static _EIGEN_DECLARE_CONST_Packet4i(0x7f, 0x7f);
+static _EIGEN_DECLARE_CONST_Packet4i(23, 23);
+
+static _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(inv_mant_mask, ~0x7f800000);
+
+/* the smallest non denormalized float number */
+static _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(min_norm_pos,  0x00800000);
+static _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(minus_inf,     0xff800000); // -1.f/0.f
+static _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(minus_nan,     0xffffffff);
+  
+/* natural logarithm computed for 4 simultaneous float
+  return NaN for x <= 0
+*/
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_SQRTHF, 0.707106781186547524f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p0, 7.0376836292E-2f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p1, - 1.1514610310E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p2, 1.1676998740E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p3, - 1.2420140846E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p4, + 1.4249322787E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p5, - 1.6668057665E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p6, + 2.0000714765E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p7, - 2.4999993993E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p8, + 3.3333331174E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q1, -2.12194440e-4f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q2, 0.693359375f);
+
+static _EIGEN_DECLARE_CONST_Packet4f(exp_hi,  88.3762626647950f);
+static _EIGEN_DECLARE_CONST_Packet4f(exp_lo, -88.3762626647949f);
+
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_LOG2EF, 1.44269504088896341f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C1, 0.693359375f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C2, -2.12194440e-4f);
+
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p0, 1.9875691500E-4f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p1, 1.3981999507E-3f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p2, 8.3334519073E-3f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p3, 4.1665795894E-2f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201E-1f);
+
+#ifdef __VSX__
+static _EIGEN_DECLARE_CONST_Packet2d(1 , 1.0);
+static _EIGEN_DECLARE_CONST_Packet2d(2 , 2.0);
+static _EIGEN_DECLARE_CONST_Packet2d(half, 0.5);
+
+static _EIGEN_DECLARE_CONST_Packet2d(exp_hi,  709.437);
+static _EIGEN_DECLARE_CONST_Packet2d(exp_lo, -709.436139303);
+
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_LOG2EF, 1.4426950408889634073599);
+
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p0, 1.26177193074810590878e-4);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p1, 3.02994407707441961300e-2);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p2, 9.99999999999999999910e-1);
+
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q0, 3.00198505138664455042e-6);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q1, 2.52448340349684104192e-3);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q2, 2.27265548208155028766e-1);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q3, 2.00000000000000000009e0);
+
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C1, 0.693145751953125);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6);
+
+#ifdef __POWER8_VECTOR__
+static Packet2l p2l_1023 = { 1023, 1023 };
+static Packet2ul p2ul_52 = { 52, 52 };
+#endif
+
+#endif
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f plog<Packet4f>(const Packet4f& _x)
+{
+  Packet4f x = _x;
+
+  Packet4i emm0;
+
+  /* isvalid_mask is 0 if x < 0 or x is NaN. */
+  Packet4ui isvalid_mask = reinterpret_cast<Packet4ui>(vec_cmpge(x, p4f_ZERO));
+  Packet4ui iszero_mask = reinterpret_cast<Packet4ui>(vec_cmpeq(x, p4f_ZERO));
+
+  x = pmax(x, p4f_min_norm_pos);  /* cut off denormalized stuff */
+  emm0 = vec_sr(reinterpret_cast<Packet4i>(x),
+                reinterpret_cast<Packet4ui>(p4i_23));
+
+  /* keep only the fractional part */
+  x = pand(x, p4f_inv_mant_mask);
+  x = por(x, p4f_half);
+
+  emm0 = psub(emm0, p4i_0x7f);
+  Packet4f e = padd(vec_ctf(emm0, 0), p4f_1);
+
+  /* part2:
+     if( x < SQRTHF ) {
+       e -= 1;
+       x = x + x - 1.0;
+     } else { x = x - 1.0; }
+  */
+  Packet4f mask = reinterpret_cast<Packet4f>(vec_cmplt(x, p4f_cephes_SQRTHF));
+  Packet4f tmp = pand(x, mask);
+  x = psub(x, p4f_1);
+  e = psub(e, pand(p4f_1, mask));
+  x = padd(x, tmp);
+
+  Packet4f x2 = pmul(x,x);
+  Packet4f x3 = pmul(x2,x);
+
+  Packet4f y, y1, y2;
+  y  = pmadd(p4f_cephes_log_p0, x, p4f_cephes_log_p1);
+  y1 = pmadd(p4f_cephes_log_p3, x, p4f_cephes_log_p4);
+  y2 = pmadd(p4f_cephes_log_p6, x, p4f_cephes_log_p7);
+  y  = pmadd(y , x, p4f_cephes_log_p2);
+  y1 = pmadd(y1, x, p4f_cephes_log_p5);
+  y2 = pmadd(y2, x, p4f_cephes_log_p8);
+  y = pmadd(y, x3, y1);
+  y = pmadd(y, x3, y2);
+  y = pmul(y, x3);
+
+  y1 = pmul(e, p4f_cephes_log_q1);
+  tmp = pmul(x2, p4f_half);
+  y = padd(y, y1);
+  x = psub(x, tmp);
+  y2 = pmul(e, p4f_cephes_log_q2);
+  x = padd(x, y);
+  x = padd(x, y2);
+  // negative arg will be NAN, 0 will be -INF
+  x = vec_sel(x, p4f_minus_inf, iszero_mask);
+  x = vec_sel(p4f_minus_nan, x, isvalid_mask);
+  return x;
+}
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f pexp<Packet4f>(const Packet4f& _x)
+{
+  Packet4f x = _x;
+
+  Packet4f tmp, fx;
+  Packet4i emm0;
+
+  // clamp x
+  x = pmax(pmin(x, p4f_exp_hi), p4f_exp_lo);
+
+  // express exp(x) as exp(g + n*log(2))
+  fx = pmadd(x, p4f_cephes_LOG2EF, p4f_half);
+
+  fx = pfloor(fx);
+
+  tmp = pmul(fx, p4f_cephes_exp_C1);
+  Packet4f z = pmul(fx, p4f_cephes_exp_C2);
+  x = psub(x, tmp);
+  x = psub(x, z);
+
+  z = pmul(x,x);
+
+  Packet4f y = p4f_cephes_exp_p0;
+  y = pmadd(y, x, p4f_cephes_exp_p1);
+  y = pmadd(y, x, p4f_cephes_exp_p2);
+  y = pmadd(y, x, p4f_cephes_exp_p3);
+  y = pmadd(y, x, p4f_cephes_exp_p4);
+  y = pmadd(y, x, p4f_cephes_exp_p5);
+  y = pmadd(y, z, x);
+  y = padd(y, p4f_1);
+
+  // build 2^n
+  emm0 = vec_cts(fx, 0);
+  emm0 = vec_add(emm0, p4i_0x7f);
+  emm0 = vec_sl(emm0, reinterpret_cast<Packet4ui>(p4i_23));
+
+  // Altivec's max & min operators just drop silent NaNs. Check NaNs in 
+  // inputs and return them unmodified.
+  Packet4ui isnumber_mask = reinterpret_cast<Packet4ui>(vec_cmpeq(_x, _x));
+  return vec_sel(_x, pmax(pmul(y, reinterpret_cast<Packet4f>(emm0)), _x),
+                 isnumber_mask);
+}
+
+#ifndef EIGEN_COMP_CLANG
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f prsqrt<Packet4f>(const Packet4f& x)
+{
+  return  vec_rsqrt(x);
+}
+#endif
+
+#ifdef __VSX__
+#ifndef EIGEN_COMP_CLANG
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet2d prsqrt<Packet2d>(const Packet2d& x)
+{
+  return  vec_rsqrt(x);
+}
+#endif
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f psqrt<Packet4f>(const Packet4f& x)
+{
+  return  vec_sqrt(x);
+}
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet2d psqrt<Packet2d>(const Packet2d& x)
+{
+  return  vec_sqrt(x);
+}
+
+// VSX support varies between different compilers and even different
+// versions of the same compiler.  For gcc version >= 4.9.3, we can use
+// vec_cts to efficiently convert Packet2d to Packet2l.  Otherwise, use
+// a slow version that works with older compilers. 
+// Update: apparently vec_cts/vec_ctf intrinsics for 64-bit doubles
+// are buggy, https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70963
+static inline Packet2l ConvertToPacket2l(const Packet2d& x) {
+#if EIGEN_GNUC_AT_LEAST(5, 4) || \
+    (EIGEN_GNUC_AT(6, 1) && __GNUC_PATCHLEVEL__ >= 1)
+  return vec_cts(x, 0);    // TODO: check clang version.
+#else
+  double tmp[2];
+  memcpy(tmp, &x, sizeof(tmp));
+  Packet2l l = { static_cast<long long>(tmp[0]),
+                 static_cast<long long>(tmp[1]) };
+  return l;
+#endif
+}
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet2d pexp<Packet2d>(const Packet2d& _x)
+{
+  Packet2d x = _x;
+
+  Packet2d tmp, fx;
+  Packet2l emm0;
+
+  // clamp x
+  x = pmax(pmin(x, p2d_exp_hi), p2d_exp_lo);
+
+  /* express exp(x) as exp(g + n*log(2)) */
+  fx = pmadd(x, p2d_cephes_LOG2EF, p2d_half);
+
+  fx = pfloor(fx);
+
+  tmp = pmul(fx, p2d_cephes_exp_C1);
+  Packet2d z = pmul(fx, p2d_cephes_exp_C2);
+  x = psub(x, tmp);
+  x = psub(x, z);
+
+  Packet2d x2 = pmul(x,x);
+
+  Packet2d px = p2d_cephes_exp_p0;
+  px = pmadd(px, x2, p2d_cephes_exp_p1);
+  px = pmadd(px, x2, p2d_cephes_exp_p2);
+  px = pmul (px, x);
+
+  Packet2d qx = p2d_cephes_exp_q0;
+  qx = pmadd(qx, x2, p2d_cephes_exp_q1);
+  qx = pmadd(qx, x2, p2d_cephes_exp_q2);
+  qx = pmadd(qx, x2, p2d_cephes_exp_q3);
+
+  x = pdiv(px,psub(qx,px));
+  x = pmadd(p2d_2,x,p2d_1);
+
+  // build 2^n
+  emm0 = ConvertToPacket2l(fx);
+
+#ifdef __POWER8_VECTOR__ 
+  emm0 = vec_add(emm0, p2l_1023);
+  emm0 = vec_sl(emm0, p2ul_52);
+#else
+  // Code is a bit complex for POWER7.  There is actually a
+  // vec_xxsldi intrinsic but it is not supported by some gcc versions.
+  // So we shift (52-32) bits and do a word swap with zeros.
+  _EIGEN_DECLARE_CONST_Packet4i(1023, 1023);
+  _EIGEN_DECLARE_CONST_Packet4i(20, 20);    // 52 - 32
+
+  Packet4i emm04i = reinterpret_cast<Packet4i>(emm0);
+  emm04i = vec_add(emm04i, p4i_1023);
+  emm04i = vec_sl(emm04i, reinterpret_cast<Packet4ui>(p4i_20));
+  static const Packet16uc perm = {
+    0x14, 0x15, 0x16, 0x17, 0x00, 0x01, 0x02, 0x03, 
+    0x1c, 0x1d, 0x1e, 0x1f, 0x08, 0x09, 0x0a, 0x0b };
+#ifdef  _BIG_ENDIAN
+  emm0 = reinterpret_cast<Packet2l>(vec_perm(p4i_ZERO, emm04i, perm));
+#else
+  emm0 = reinterpret_cast<Packet2l>(vec_perm(emm04i, p4i_ZERO, perm));
+#endif
+
+#endif
+
+  // Altivec's max & min operators just drop silent NaNs. Check NaNs in 
+  // inputs and return them unmodified.
+  Packet2ul isnumber_mask = reinterpret_cast<Packet2ul>(vec_cmpeq(_x, _x));
+  return vec_sel(_x, pmax(pmul(x, reinterpret_cast<Packet2d>(emm0)), _x),
+                 isnumber_mask);
+}
+#endif
+
+}  // end namespace internal
+
+}  // end namespace Eigen
+
+#endif  // EIGEN_MATH_FUNCTIONS_ALTIVEC_H
diff --git a/eigen/Eigen/src/Core/arch/AltiVec/PacketMath.h b/eigen/Eigen/src/Core/arch/AltiVec/PacketMath.h
index e408996..b3f1ea1 100644
--- a/eigen/Eigen/src/Core/arch/AltiVec/PacketMath.h
+++ b/eigen/Eigen/src/Core/arch/AltiVec/PacketMath.h
@@ -1,7 +1,7 @@
 // This file is part of Eigen, a lightweight C++ template library
 // for linear algebra.
 //
-// Copyright (C) 2008 Konstantinos Margaritis <markos@codex.gr>
+// Copyright (C) 2008-2016 Konstantinos Margaritis <markos@freevec.org>
 //
 // This Source Code Form is subject to the terms of the Mozilla
 // Public License v. 2.0. If a copy of the MPL was not distributed
@@ -18,13 +18,17 @@ namespace internal {
 #define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 4
 #endif
 
-#ifndef EIGEN_HAS_FUSE_CJMADD
-#define EIGEN_HAS_FUSE_CJMADD 1
+#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#endif
+
+#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD
+#define EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD
 #endif
 
 // NOTE Altivec has 32 registers, but Eigen only accepts a value of 8 or 16
 #ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
-#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 16
+#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS  32
 #endif
 
 typedef __vector float          Packet4f;
@@ -38,7 +42,7 @@ typedef __vector unsigned char  Packet16uc;
 // and it doesn't really work to declare them global, so we define macros instead
 
 #define _EIGEN_DECLARE_CONST_FAST_Packet4f(NAME,X) \
-  Packet4f p4f_##NAME = (Packet4f) vec_splat_s32(X)
+  Packet4f p4f_##NAME = reinterpret_cast<Packet4f>(vec_splat_s32(X))
 
 #define _EIGEN_DECLARE_CONST_FAST_Packet4i(NAME,X) \
   Packet4i p4i_##NAME = vec_splat_s32(X)
@@ -46,60 +50,158 @@ typedef __vector unsigned char  Packet16uc;
 #define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \
   Packet4f p4f_##NAME = pset1<Packet4f>(X)
 
-#define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
-  Packet4f p4f_##NAME = vreinterpretq_f32_u32(pset1<int>(X))
-
 #define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
   Packet4i p4i_##NAME = pset1<Packet4i>(X)
 
+#define _EIGEN_DECLARE_CONST_Packet2d(NAME,X) \
+  Packet2d p2d_##NAME = pset1<Packet2d>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet2l(NAME,X) \
+  Packet2l p2l_##NAME = pset1<Packet2l>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
+  const Packet4f p4f_##NAME = reinterpret_cast<Packet4f>(pset1<Packet4i>(X))
+
 #define DST_CHAN 1
 #define DST_CTRL(size, count, stride) (((size) << 24) | ((count) << 16) | (stride))
 
+
+// These constants are endian-agnostic
+static _EIGEN_DECLARE_CONST_FAST_Packet4f(ZERO, 0); //{ 0.0, 0.0, 0.0, 0.0}
+static _EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0); //{ 0, 0, 0, 0,}
+static _EIGEN_DECLARE_CONST_FAST_Packet4i(ONE,1); //{ 1, 1, 1, 1}
+static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS16,-16); //{ -16, -16, -16, -16}
+static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS1,-1); //{ -1, -1, -1, -1}
+static Packet4f p4f_MZERO = (Packet4f) vec_sl((Packet4ui)p4i_MINUS1, (Packet4ui)p4i_MINUS1); //{ 0x80000000, 0x80000000, 0x80000000, 0x80000000}
+#ifndef __VSX__
+static Packet4f p4f_ONE = vec_ctf(p4i_ONE, 0); //{ 1.0, 1.0, 1.0, 1.0}
+#endif
+
+static Packet4f p4f_COUNTDOWN = { 0.0, 1.0, 2.0, 3.0 };
+static Packet4i p4i_COUNTDOWN = { 0, 1, 2, 3 };
+
+static Packet16uc p16uc_REVERSE32 = { 12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3 };
+static Packet16uc p16uc_DUPLICATE32_HI = { 0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7 };
+
+// Mask alignment
+#ifdef __PPC64__
+#define _EIGEN_MASK_ALIGNMENT	0xfffffffffffffff0
+#else
+#define _EIGEN_MASK_ALIGNMENT	0xfffffff0
+#endif
+
+#define _EIGEN_ALIGNED_PTR(x)	((std::ptrdiff_t)(x) & _EIGEN_MASK_ALIGNMENT)
+
+// Handle endianness properly while loading constants
 // Define global static constants:
-static Packet4f p4f_COUNTDOWN = { 3.0, 2.0, 1.0, 0.0 };
-static Packet4i p4i_COUNTDOWN = { 3, 2, 1, 0 };
-static Packet16uc p16uc_REVERSE = {12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3};
+#ifdef _BIG_ENDIAN
 static Packet16uc p16uc_FORWARD = vec_lvsl(0, (float*)0);
-static Packet16uc p16uc_DUPLICATE = {0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7};
-
-static _EIGEN_DECLARE_CONST_FAST_Packet4f(ZERO, 0);
-static _EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0);
-static _EIGEN_DECLARE_CONST_FAST_Packet4i(ONE,1);
-static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS16,-16);
-static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS1,-1);
-static Packet4f p4f_ONE = vec_ctf(p4i_ONE, 0);
-static Packet4f p4f_ZERO_ = (Packet4f) vec_sl((Packet4ui)p4i_MINUS1, (Packet4ui)p4i_MINUS1);
+#ifdef __VSX__
+static Packet16uc p16uc_REVERSE64 = { 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
+#endif
+static Packet16uc p16uc_PSET32_WODD   = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 };
+static Packet16uc p16uc_PSET32_WEVEN  = vec_sld(p16uc_DUPLICATE32_HI, (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 };
+static Packet16uc p16uc_HALF64_0_16 = vec_sld((Packet16uc)p4i_ZERO, vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 3), 8);      //{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16};
+#else
+static Packet16uc p16uc_FORWARD = p16uc_REVERSE32; 
+static Packet16uc p16uc_REVERSE64 = { 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
+static Packet16uc p16uc_PSET32_WODD = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 1), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 };
+static Packet16uc p16uc_PSET32_WEVEN = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 };
+static Packet16uc p16uc_HALF64_0_16 = vec_sld(vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 0), (Packet16uc)p4i_ZERO, 8);      //{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16};
+#endif // _BIG_ENDIAN
+
+static Packet16uc p16uc_PSET64_HI = (Packet16uc) vec_mergeh((Packet4ui)p16uc_PSET32_WODD, (Packet4ui)p16uc_PSET32_WEVEN);     //{ 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 };
+static Packet16uc p16uc_PSET64_LO = (Packet16uc) vec_mergel((Packet4ui)p16uc_PSET32_WODD, (Packet4ui)p16uc_PSET32_WEVEN);     //{ 8,9,10,11, 12,13,14,15, 8,9,10,11, 12,13,14,15 };
+static Packet16uc p16uc_TRANSPOSE64_HI = p16uc_PSET64_HI + p16uc_HALF64_0_16;                                         //{ 0,1,2,3, 4,5,6,7, 16,17,18,19, 20,21,22,23};
+static Packet16uc p16uc_TRANSPOSE64_LO = p16uc_PSET64_LO + p16uc_HALF64_0_16;                                         //{ 8,9,10,11, 12,13,14,15, 24,25,26,27, 28,29,30,31};
+
+static Packet16uc p16uc_COMPLEX32_REV = vec_sld(p16uc_REVERSE32, p16uc_REVERSE32, 8);                                         //{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 };
+
+#ifdef _BIG_ENDIAN
+static Packet16uc p16uc_COMPLEX32_REV2 = vec_sld(p16uc_FORWARD, p16uc_FORWARD, 8);                                            //{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
+#else
+static Packet16uc p16uc_COMPLEX32_REV2 = vec_sld(p16uc_PSET64_HI, p16uc_PSET64_LO, 8);                                            //{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
+#endif // _BIG_ENDIAN
+
+#if EIGEN_HAS_BUILTIN(__builtin_prefetch) || EIGEN_COMP_GNUC
+  #define EIGEN_PPC_PREFETCH(ADDR) __builtin_prefetch(ADDR);
+#else
+  #define EIGEN_PPC_PREFETCH(ADDR) asm( "   dcbt [%[addr]]\n" :: [addr] "r" (ADDR) : "cc" );
+#endif
 
 template<> struct packet_traits<float>  : default_packet_traits
 {
   typedef Packet4f type;
+  typedef Packet4f half;
   enum {
     Vectorizable = 1,
     AlignedOnScalar = 1,
     size=4,
-
-    // FIXME check the Has*
+    HasHalfPacket = 1,
+
+    HasAdd  = 1,
+    HasSub  = 1,
+    HasMul  = 1,
+    HasDiv  = 1,
+    HasMin  = 1,
+    HasMax  = 1,
+    HasAbs  = 1,
     HasSin  = 0,
     HasCos  = 0,
     HasLog  = 0,
-    HasExp  = 0,
-    HasSqrt = 0
+    HasExp  = 1,
+#ifdef __VSX__
+    HasSqrt = 1,
+#if !EIGEN_COMP_CLANG
+    HasRsqrt = 1,
+#else
+    HasRsqrt = 0,
+#endif
+#else
+    HasSqrt = 0,
+    HasRsqrt = 0,
+#endif
+    HasRound = 1,
+    HasFloor = 1,
+    HasCeil = 1,
+    HasNegate = 1,
+    HasBlend = 1
   };
 };
 template<> struct packet_traits<int>    : default_packet_traits
 {
   typedef Packet4i type;
+  typedef Packet4i half;
   enum {
-    // FIXME check the Has*
     Vectorizable = 1,
     AlignedOnScalar = 1,
-    size=4
+    size = 4,
+    HasHalfPacket = 0,
+
+    HasAdd  = 1,
+    HasSub  = 1,
+    HasMul  = 1,
+    HasDiv  = 0,
+    HasBlend = 1
   };
 };
 
-template<> struct unpacket_traits<Packet4f> { typedef float  type; enum {size=4}; };
-template<> struct unpacket_traits<Packet4i> { typedef int    type; enum {size=4}; };
-/*
+
+template<> struct unpacket_traits<Packet4f> { typedef float  type; enum {size=4, alignment=Aligned16}; typedef Packet4f half; };
+template<> struct unpacket_traits<Packet4i> { typedef int    type; enum {size=4, alignment=Aligned16}; typedef Packet4i half; };
+
+inline std::ostream & operator <<(std::ostream & s, const Packet16uc & v)
+{
+  union {
+    Packet16uc   v;
+    unsigned char n[16];
+  } vt;
+  vt.v = v;
+  for (int i=0; i< 16; i++)
+    s << (int)vt.n[i] << ", ";
+  return s;
+}
+
 inline std::ostream & operator <<(std::ostream & s, const Packet4f & v)
 {
   union {
@@ -133,89 +235,136 @@ inline std::ostream & operator <<(std::ostream & s, const Packet4ui & v)
   return s;
 }
 
-inline std::ostream & operator <<(std::ostream & s, const Packetbi & v)
+// Need to define them first or we get specialization after instantiation errors
+template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from)
 {
-  union {
-    Packet4bi v;
-    unsigned int n[4];
-  } vt;
-  vt.v = v;
-  s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3];
-  return s;
-}
-*/
-template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float&  from) {
-  // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
-  float EIGEN_ALIGN16 af[4];
-  af[0] = from;
-  Packet4f vc = vec_ld(0, af);
-  vc = vec_splat(vc, 0);
-  return vc;
+  EIGEN_DEBUG_ALIGNED_LOAD
+#ifdef __VSX__
+  return vec_vsx_ld(0, from);
+#else
+  return vec_ld(0, from);
+#endif
 }
 
-template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int&    from)   {
-  int EIGEN_ALIGN16 ai[4];
-  ai[0] = from;
-  Packet4i vc = vec_ld(0, ai);
-  vc = vec_splat(vc, 0);
-  return vc;
+template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int*     from)
+{
+  EIGEN_DEBUG_ALIGNED_LOAD
+#ifdef __VSX__
+  return vec_vsx_ld(0, from);
+#else
+  return vec_ld(0, from);
+#endif
 }
 
-template<> EIGEN_STRONG_INLINE Packet4f plset<float>(const float& a) { return vec_add(pset1<Packet4f>(a), p4f_COUNTDOWN); }
-template<> EIGEN_STRONG_INLINE Packet4i plset<int>(const int& a)     { return vec_add(pset1<Packet4i>(a), p4i_COUNTDOWN); }
-
-template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_add(a,b); }
-template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_add(a,b); }
+template<> EIGEN_STRONG_INLINE void pstore<float>(float*   to, const Packet4f& from)
+{
+  EIGEN_DEBUG_ALIGNED_STORE
+#ifdef __VSX__
+  vec_vsx_st(from, 0, to);
+#else
+  vec_st(from, 0, to);
+#endif
+}
 
-template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_sub(a,b); }
-template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_sub(a,b); }
+template<> EIGEN_STRONG_INLINE void pstore<int>(int*       to, const Packet4i& from)
+{
+  EIGEN_DEBUG_ALIGNED_STORE
+#ifdef __VSX__
+  vec_vsx_st(from, 0, to);
+#else
+  vec_st(from, 0, to);
+#endif
+}
 
-template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a) { return psub<Packet4f>(p4f_ZERO, a); }
-template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return psub<Packet4i>(p4i_ZERO, a); }
+template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float&  from) {
+  Packet4f v = {from, from, from, from};
+  return v;
+}
 
-template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; }
-template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; }
+template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int&    from)   {
+  Packet4i v = {from, from, from, from};
+  return v;
+}
+template<> EIGEN_STRONG_INLINE void
+pbroadcast4<Packet4f>(const float *a,
+                      Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3)
+{
+  a3 = pload<Packet4f>(a);
+  a0 = vec_splat(a3, 0);
+  a1 = vec_splat(a3, 1);
+  a2 = vec_splat(a3, 2);
+  a3 = vec_splat(a3, 3);
+}
+template<> EIGEN_STRONG_INLINE void
+pbroadcast4<Packet4i>(const int *a,
+                      Packet4i& a0, Packet4i& a1, Packet4i& a2, Packet4i& a3)
+{
+  a3 = pload<Packet4i>(a);
+  a0 = vec_splat(a3, 0);
+  a1 = vec_splat(a3, 1);
+  a2 = vec_splat(a3, 2);
+  a3 = vec_splat(a3, 3);
+}
 
-template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_madd(a,b,p4f_ZERO); }
-/* Commented out: it's actually slower than processing it scalar
- *
-template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b)
+template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride)
+{
+  float EIGEN_ALIGN16 af[4];
+  af[0] = from[0*stride];
+  af[1] = from[1*stride];
+  af[2] = from[2*stride];
+  af[3] = from[3*stride];
+ return pload<Packet4f>(af);
+}
+template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* from, Index stride)
+{
+  int EIGEN_ALIGN16 ai[4];
+  ai[0] = from[0*stride];
+  ai[1] = from[1*stride];
+  ai[2] = from[2*stride];
+  ai[3] = from[3*stride];
+ return pload<Packet4i>(ai);
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride)
 {
-  // Detailed in: http://freevec.org/content/32bit_signed_integer_multiplication_altivec
-  //Set up constants, variables
-  Packet4i a1, b1, bswap, low_prod, high_prod, prod, prod_, v1sel;
+  float EIGEN_ALIGN16 af[4];
+  pstore<float>(af, from);
+  to[0*stride] = af[0];
+  to[1*stride] = af[1];
+  to[2*stride] = af[2];
+  to[3*stride] = af[3];
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const Packet4i& from, Index stride)
+{
+  int EIGEN_ALIGN16 ai[4];
+  pstore<int>((int *)ai, from);
+  to[0*stride] = ai[0];
+  to[1*stride] = ai[1];
+  to[2*stride] = ai[2];
+  to[3*stride] = ai[3];
+}
 
-  // Get the absolute values
-  a1  = vec_abs(a);
-  b1  = vec_abs(b);
+template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a) { return pset1<Packet4f>(a) + p4f_COUNTDOWN; }
+template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int& a)   { return pset1<Packet4i>(a) + p4i_COUNTDOWN; }
 
-  // Get the signs using xor
-  Packet4bi sgn = (Packet4bi) vec_cmplt(vec_xor(a, b), p4i_ZERO);
+template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return a + b; }
+template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return a + b; }
 
-  // Do the multiplication for the asbolute values.
-  bswap = (Packet4i) vec_rl((Packet4ui) b1, (Packet4ui) p4i_MINUS16 );
-  low_prod = vec_mulo((Packet8i) a1, (Packet8i)b1);
-  high_prod = vec_msum((Packet8i) a1, (Packet8i) bswap, p4i_ZERO);
-  high_prod = (Packet4i) vec_sl((Packet4ui) high_prod, (Packet4ui) p4i_MINUS16);
-  prod = vec_add( low_prod, high_prod );
+template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { return a - b; }
+template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return a - b; }
 
-  // NOR the product and select only the negative elements according to the sign mask
-  prod_ = vec_nor(prod, prod);
-  prod_ = vec_sel(p4i_ZERO, prod_, sgn);
+template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a) { return p4f_ZERO - a; }
+template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return p4i_ZERO - a; }
 
-  // Add 1 to the result to get the negative numbers
-  v1sel = vec_sel(p4i_ZERO, p4i_ONE, sgn);
-  prod_ = vec_add(prod_, v1sel);
+template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; }
+template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; }
 
-  // Merge the results back to the final vector.
-  prod = vec_sel(prod, prod_, sgn);
+template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_madd(a,b, p4f_MZERO); }
+template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) { return a * b; }
 
-  return prod;
-}
-*/
 template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b)
 {
-  Packet4f t, y_0, y_1, res;
+#ifndef __VSX__  // VSX actually provides a div instruction
+  Packet4f t, y_0, y_1;
 
   // Altivec does not offer a divide instruction, we have to do a reciprocal approximation
   y_0 = vec_re(b);
@@ -224,8 +373,10 @@ template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const
   t   = vec_nmsub(y_0, b, p4f_ONE);
   y_1 = vec_madd(y_0, t, y_0);
 
-  res = vec_madd(a, y_1, p4f_ZERO);
-  return res;
+  return vec_madd(a, y_1, p4f_MZERO);
+#else
+  return vec_div(a, b);
+#endif
 }
 
 template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, const Packet4i& /*b*/)
@@ -234,8 +385,8 @@ template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, co
 }
 
 // for some weird raisons, it has to be overloaded for packet of integers
-template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vec_madd(a, b, c); }
-template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd(pmul(a,b), c); }
+template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vec_madd(a,b,c); }
+template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return a*b + c; }
 
 template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_min(a, b); }
 template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_min(a, b); }
@@ -243,7 +394,6 @@ template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const
 template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_max(a, b); }
 template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_max(a, b); }
 
-// Logical Operations are not supported for float, so we have to reinterpret casts using NEON intrinsics
 template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_and(a, b); }
 template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, b); }
 
@@ -256,13 +406,14 @@ template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const
 template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_and(a, vec_nor(b, b)); }
 template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, vec_nor(b, b)); }
 
-template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return vec_ld(0, from); }
-template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int*     from) { EIGEN_DEBUG_ALIGNED_LOAD return vec_ld(0, from); }
+template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a) { return vec_round(a); }
+template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const  Packet4f& a) { return vec_ceil(a); }
+template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) { return vec_floor(a); }
 
+#ifdef _BIG_ENDIAN
 template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from)
 {
   EIGEN_DEBUG_ALIGNED_LOAD
-  // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
   Packet16uc MSQ, LSQ;
   Packet16uc mask;
   MSQ = vec_ld(0, (unsigned char *)from);          // most significant quadword
@@ -282,25 +433,36 @@ template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)
   mask = vec_lvsl(0, from);                        // create the permute mask
   return (Packet4i) vec_perm(MSQ, LSQ, mask);    // align the data
 }
+#else
+// We also need ot redefine little endian loading of Packet4i/Packet4f using VSX
+template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)
+{
+  EIGEN_DEBUG_UNALIGNED_LOAD
+  return (Packet4i) vec_vsx_ld((long)from & 15, (const int*) _EIGEN_ALIGNED_PTR(from));
+}
+template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from)
+{
+  EIGEN_DEBUG_UNALIGNED_LOAD
+  return (Packet4f) vec_vsx_ld((long)from & 15, (const float*) _EIGEN_ALIGNED_PTR(from));
+}
+#endif
 
 template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float*   from)
 {
   Packet4f p;
-  if((ptrdiff_t(&from) % 16) == 0)  p = pload<Packet4f>(from);
-  else                              p = ploadu<Packet4f>(from);
-  return vec_perm(p, p, p16uc_DUPLICATE);
+  if((std::ptrdiff_t(from) % 16) == 0)  p = pload<Packet4f>(from);
+  else                                  p = ploadu<Packet4f>(from);
+  return vec_perm(p, p, p16uc_DUPLICATE32_HI);
 }
 template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int*     from)
 {
   Packet4i p;
-  if((ptrdiff_t(&from) % 16) == 0)  p = pload<Packet4i>(from);
-  else                              p = ploadu<Packet4i>(from);
-  return vec_perm(p, p, p16uc_DUPLICATE);
+  if((std::ptrdiff_t(from) % 16) == 0)  p = pload<Packet4i>(from);
+  else                                  p = ploadu<Packet4i>(from);
+  return vec_perm(p, p, p16uc_DUPLICATE32_HI);
 }
 
-template<> EIGEN_STRONG_INLINE void pstore<float>(float*   to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE vec_st(from, 0, to); }
-template<> EIGEN_STRONG_INLINE void pstore<int>(int*       to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE vec_st(from, 0, to); }
-
+#ifdef _BIG_ENDIAN
 template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*  to, const Packet4f& from)
 {
   EIGEN_DEBUG_UNALIGNED_STORE
@@ -337,15 +499,33 @@ template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*      to, const Packet4i& f
   vec_st( LSQ, 15, (unsigned char *)to );                   // Store the LSQ part first
   vec_st( MSQ, 0, (unsigned char *)to );                    // Store the MSQ part
 }
+#else
+// We also need ot redefine little endian loading of Packet4i/Packet4f using VSX
+template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*       to, const Packet4i& from)
+{
+  EIGEN_DEBUG_ALIGNED_STORE
+  vec_vsx_st(from, (long)to & 15, (int*) _EIGEN_ALIGNED_PTR(to));
+}
+template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*   to, const Packet4f& from)
+{
+  EIGEN_DEBUG_ALIGNED_STORE
+  vec_vsx_st(from, (long)to & 15, (float*) _EIGEN_ALIGNED_PTR(to));
+}
+#endif
 
-template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); }
-template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*     addr) { vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); }
+template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr)    { EIGEN_PPC_PREFETCH(addr); }
+template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*     addr)    { EIGEN_PPC_PREFETCH(addr); }
 
-template<> EIGEN_STRONG_INLINE float  pfirst<Packet4f>(const Packet4f& a) { float EIGEN_ALIGN16 x[4]; vec_st(a, 0, x); return x[0]; }
-template<> EIGEN_STRONG_INLINE int    pfirst<Packet4i>(const Packet4i& a) { int   EIGEN_ALIGN16 x[4]; vec_st(a, 0, x); return x[0]; }
+template<> EIGEN_STRONG_INLINE float  pfirst<Packet4f>(const Packet4f& a) { float EIGEN_ALIGN16 x; vec_ste(a, 0, &x); return x; }
+template<> EIGEN_STRONG_INLINE int    pfirst<Packet4i>(const Packet4i& a) { int   EIGEN_ALIGN16 x; vec_ste(a, 0, &x); return x; }
 
-template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) { return (Packet4f)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE); }
-template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) { return (Packet4i)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE); }
+template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a)
+{
+  return reinterpret_cast<Packet4f>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE32));
+}
+template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a)
+{
+  return reinterpret_cast<Packet4i>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE32)); }
 
 template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a) { return vec_abs(a); }
 template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) { return vec_abs(a); }
@@ -353,10 +533,10 @@ template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) { return vec_abs
 template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
 {
   Packet4f b, sum;
-  b   = (Packet4f) vec_sld(a, a, 8);
-  sum = vec_add(a, b);
-  b   = (Packet4f) vec_sld(sum, sum, 4);
-  sum = vec_add(sum, b);
+  b   = vec_sld(a, a, 8);
+  sum = a + b;
+  b   = vec_sld(sum, sum, 4);
+  sum += b;
   return pfirst(sum);
 }
 
@@ -379,11 +559,11 @@ template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
 
   // Now do the summation:
   // Lines 0+1
-  sum[0] = vec_add(sum[0], sum[1]);
+  sum[0] = sum[0] + sum[1];
   // Lines 2+3
-  sum[1] = vec_add(sum[2], sum[3]);
+  sum[1] = sum[2] + sum[3];
   // Add the results
-  sum[0] = vec_add(sum[0], sum[1]);
+  sum[0] = sum[0] + sum[1];
 
   return sum[0];
 }
@@ -392,7 +572,11 @@ template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
 {
   Packet4i sum;
   sum = vec_sums(a, p4i_ZERO);
+#ifdef _BIG_ENDIAN
   sum = vec_sld(sum, p4i_ZERO, 12);
+#else
+  sum = vec_sld(p4i_ZERO, sum, 4);
+#endif
   return pfirst(sum);
 }
 
@@ -415,11 +599,11 @@ template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
 
   // Now do the summation:
   // Lines 0+1
-  sum[0] = vec_add(sum[0], sum[1]);
+  sum[0] = sum[0] + sum[1];
   // Lines 2+3
-  sum[1] = vec_add(sum[2], sum[3]);
+  sum[1] = sum[2] + sum[3];
   // Add the results
-  sum[0] = vec_add(sum[0], sum[1]);
+  sum[0] = sum[0] + sum[1];
 
   return sum[0];
 }
@@ -429,8 +613,8 @@ template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
 template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
 {
   Packet4f prod;
-  prod = pmul(a, (Packet4f)vec_sld(a, a, 8));
-  return pfirst(pmul(prod, (Packet4f)vec_sld(prod, prod, 4)));
+  prod = pmul(a, vec_sld(a, a, 8));
+  return pfirst(pmul(prod, vec_sld(prod, prod, 4)));
 }
 
 template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a)
@@ -479,8 +663,25 @@ struct palign_impl<Offset,Packet4f>
 {
   static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second)
   {
-    if (Offset!=0)
-      first = vec_sld(first, second, Offset*4);
+#ifdef _BIG_ENDIAN
+    switch (Offset % 4) {
+    case 1:
+      first = vec_sld(first, second, 4); break;
+    case 2:
+      first = vec_sld(first, second, 8); break;
+    case 3:
+      first = vec_sld(first, second, 12); break;
+    }
+#else
+    switch (Offset % 4) {
+    case 1:
+      first = vec_sld(second, first, 12); break;
+    case 2:
+      first = vec_sld(second, first, 8); break;
+    case 3:
+      first = vec_sld(second, first, 4); break;
+    }
+#endif
   }
 };
 
@@ -489,11 +690,342 @@ struct palign_impl<Offset,Packet4i>
 {
   static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second)
   {
-    if (Offset!=0)
-      first = vec_sld(first, second, Offset*4);
+#ifdef _BIG_ENDIAN
+    switch (Offset % 4) {
+    case 1:
+      first = vec_sld(first, second, 4); break;
+    case 2:
+      first = vec_sld(first, second, 8); break;
+    case 3:
+      first = vec_sld(first, second, 12); break;
+    }
+#else
+    switch (Offset % 4) {
+    case 1:
+      first = vec_sld(second, first, 12); break;
+    case 2:
+      first = vec_sld(second, first, 8); break;
+    case 3:
+      first = vec_sld(second, first, 4); break;
+    }
+#endif
   }
 };
 
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet4f,4>& kernel) {
+  Packet4f t0, t1, t2, t3;
+  t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]);
+  t1 = vec_mergel(kernel.packet[0], kernel.packet[2]);
+  t2 = vec_mergeh(kernel.packet[1], kernel.packet[3]);
+  t3 = vec_mergel(kernel.packet[1], kernel.packet[3]);
+  kernel.packet[0] = vec_mergeh(t0, t2);
+  kernel.packet[1] = vec_mergel(t0, t2);
+  kernel.packet[2] = vec_mergeh(t1, t3);
+  kernel.packet[3] = vec_mergel(t1, t3);
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet4i,4>& kernel) {
+  Packet4i t0, t1, t2, t3;
+  t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]);
+  t1 = vec_mergel(kernel.packet[0], kernel.packet[2]);
+  t2 = vec_mergeh(kernel.packet[1], kernel.packet[3]);
+  t3 = vec_mergel(kernel.packet[1], kernel.packet[3]);
+  kernel.packet[0] = vec_mergeh(t0, t2);
+  kernel.packet[1] = vec_mergel(t0, t2);
+  kernel.packet[2] = vec_mergeh(t1, t3);
+  kernel.packet[3] = vec_mergel(t1, t3);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pblend(const Selector<4>& ifPacket, const Packet4i& thenPacket, const Packet4i& elsePacket) {
+  Packet4ui select = { ifPacket.select[0], ifPacket.select[1], ifPacket.select[2], ifPacket.select[3] };
+  Packet4ui mask = reinterpret_cast<Packet4ui>(vec_cmpeq(reinterpret_cast<Packet4ui>(select), reinterpret_cast<Packet4ui>(p4i_ONE)));
+  return vec_sel(elsePacket, thenPacket, mask);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket, const Packet4f& elsePacket) {
+  Packet4ui select = { ifPacket.select[0], ifPacket.select[1], ifPacket.select[2], ifPacket.select[3] };
+  Packet4ui mask = reinterpret_cast<Packet4ui>(vec_cmpeq(reinterpret_cast<Packet4ui>(select), reinterpret_cast<Packet4ui>(p4i_ONE)));
+  return vec_sel(elsePacket, thenPacket, mask);
+}
+
+
+//---------- double ----------
+#ifdef __VSX__
+typedef __vector double              Packet2d;
+typedef __vector unsigned long long  Packet2ul;
+typedef __vector long long           Packet2l;
+#if EIGEN_COMP_CLANG
+typedef Packet2ul                    Packet2bl;
+#else
+typedef __vector __bool long         Packet2bl;
+#endif
+
+static Packet2l  p2l_ONE  = { 1, 1 };
+static Packet2l  p2l_ZERO = reinterpret_cast<Packet2l>(p4i_ZERO);
+static Packet2d  p2d_ONE  = { 1.0, 1.0 }; 
+static Packet2d  p2d_ZERO = reinterpret_cast<Packet2d>(p4f_ZERO);
+static Packet2d  p2d_MZERO = { -0.0, -0.0 };
+
+#ifdef _BIG_ENDIAN
+static Packet2d p2d_COUNTDOWN = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(p2d_ZERO), reinterpret_cast<Packet4f>(p2d_ONE), 8));
+#else
+static Packet2d p2d_COUNTDOWN = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(p2d_ONE), reinterpret_cast<Packet4f>(p2d_ZERO), 8));
+#endif
+
+template<int index> Packet2d vec_splat_dbl(Packet2d& a);
+
+template<> EIGEN_STRONG_INLINE Packet2d vec_splat_dbl<0>(Packet2d& a)
+{
+  return reinterpret_cast<Packet2d>(vec_perm(a, a, p16uc_PSET64_HI));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d vec_splat_dbl<1>(Packet2d& a)
+{
+  return reinterpret_cast<Packet2d>(vec_perm(a, a, p16uc_PSET64_LO));
+}
+
+template<> struct packet_traits<double> : default_packet_traits
+{
+  typedef Packet2d type;
+  typedef Packet2d half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=2,
+    HasHalfPacket = 1,
+
+    HasAdd  = 1,
+    HasSub  = 1,
+    HasMul  = 1,
+    HasDiv  = 1,
+    HasMin  = 1,
+    HasMax  = 1,
+    HasAbs  = 1,
+    HasSin  = 0,
+    HasCos  = 0,
+    HasLog  = 0,
+    HasExp  = 1,
+    HasSqrt = 1,
+    HasRsqrt = 1,
+    HasRound = 1,
+    HasFloor = 1,
+    HasCeil = 1,
+    HasNegate = 1,
+    HasBlend = 1
+  };
+};
+
+template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16}; typedef Packet2d half; };
+
+inline std::ostream & operator <<(std::ostream & s, const Packet2l & v)
+{
+  union {
+    Packet2l   v;
+    int64_t n[2];
+  } vt;
+  vt.v = v;
+  s << vt.n[0] << ", " << vt.n[1];
+  return s;
+}
+
+inline std::ostream & operator <<(std::ostream & s, const Packet2d & v)
+{
+  union {
+    Packet2d   v;
+    double n[2];
+  } vt;
+  vt.v = v;
+  s << vt.n[0] << ", " << vt.n[1];
+  return s;
+}
+
+// Need to define them first or we get specialization after instantiation errors
+template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from)
+{
+  EIGEN_DEBUG_ALIGNED_LOAD
+#ifdef __VSX__
+  return vec_vsx_ld(0, from);
+#else
+  return vec_ld(0, from);
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE void pstore<double>(double*   to, const Packet2d& from)
+{
+  EIGEN_DEBUG_ALIGNED_STORE
+#ifdef __VSX__
+  vec_vsx_st(from, 0, to);
+#else
+  vec_st(from, 0, to);
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double&  from) {
+  Packet2d v = {from, from};
+  return v;
+}
+
+template<> EIGEN_STRONG_INLINE void
+pbroadcast4<Packet2d>(const double *a,
+                      Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3)
+{
+  a1 = pload<Packet2d>(a);
+  a0 = vec_splat_dbl<0>(a1);
+  a1 = vec_splat_dbl<1>(a1);
+  a3 = pload<Packet2d>(a+2);
+  a2 = vec_splat_dbl<0>(a3);
+  a3 = vec_splat_dbl<1>(a3);
+}
+
+template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride)
+{
+  double EIGEN_ALIGN16 af[2];
+  af[0] = from[0*stride];
+  af[1] = from[1*stride];
+ return pload<Packet2d>(af);
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride)
+{
+  double EIGEN_ALIGN16 af[2];
+  pstore<double>(af, from);
+  to[0*stride] = af[0];
+  to[1*stride] = af[1];
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a) { return pset1<Packet2d>(a) + p2d_COUNTDOWN; }
+
+template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return a + b; }
+
+template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return a - b; }
+
+template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) { return p2d_ZERO - a; }
+
+template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; }
+
+template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_madd(a,b,p2d_MZERO); }
+template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_div(a,b); }
+
+// for some weird raisons, it has to be overloaded for packet of integers
+template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vec_madd(a, b, c); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_min(a, b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_max(a, b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_or(a, b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_xor(a, b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, vec_nor(b, b)); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a) { return vec_round(a); }
+template<> EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const  Packet2d& a) { return vec_ceil(a); }
+template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) { return vec_floor(a); }
+
+template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from)
+{
+  EIGEN_DEBUG_ALIGNED_LOAD
+  return (Packet2d) vec_vsx_ld((long)from & 15, (const double*) _EIGEN_ALIGNED_PTR(from));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double*   from)
+{
+  Packet2d p;
+  if((std::ptrdiff_t(from) % 16) == 0)  p = pload<Packet2d>(from);
+  else                                  p = ploadu<Packet2d>(from);
+  return vec_splat_dbl<0>(p);
+}
+
+template<> EIGEN_STRONG_INLINE void pstoreu<double>(double*  to, const Packet2d& from)
+{
+  EIGEN_DEBUG_ALIGNED_STORE
+  vec_vsx_st((Packet4f)from, (long)to & 15, (float*) _EIGEN_ALIGNED_PTR(to));
+}
+
+template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { EIGEN_PPC_PREFETCH(addr); }
+
+template<> EIGEN_STRONG_INLINE double  pfirst<Packet2d>(const Packet2d& a) { double EIGEN_ALIGN16 x[2]; pstore<double>(x, a); return x[0]; }
+
+template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a)
+{
+  return reinterpret_cast<Packet2d>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE64));
+}
+template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a) { return vec_abs(a); }
+
+template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a)
+{
+  Packet2d b, sum;
+  b   = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(a), reinterpret_cast<Packet4f>(a), 8));
+  sum = a + b;
+  return pfirst<Packet2d>(sum);
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
+{
+  Packet2d v[2], sum;
+  v[0] = vecs[0] + reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(vecs[0]), reinterpret_cast<Packet4f>(vecs[0]), 8));
+  v[1] = vecs[1] + reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(vecs[1]), reinterpret_cast<Packet4f>(vecs[1]), 8));
+ 
+#ifdef _BIG_ENDIAN
+  sum = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(v[0]), reinterpret_cast<Packet4f>(v[1]), 8));
+#else
+  sum = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(v[1]), reinterpret_cast<Packet4f>(v[0]), 8));
+#endif
+
+  return sum;
+}
+// Other reduction functions:
+// mul
+template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a)
+{
+  return pfirst(pmul(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(a), reinterpret_cast<Packet4ui>(a), 8))));
+}
+
+// min
+template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a)
+{
+  return pfirst(pmin(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(a), reinterpret_cast<Packet4ui>(a), 8))));
+}
+
+// max
+template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a)
+{
+  return pfirst(pmax(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(a), reinterpret_cast<Packet4ui>(a), 8))));
+}
+
+template<int Offset>
+struct palign_impl<Offset,Packet2d>
+{
+  static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second)
+  {
+    if (Offset == 1)
+#ifdef _BIG_ENDIAN
+      first = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(first), reinterpret_cast<Packet4ui>(second), 8));
+#else
+      first = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(second), reinterpret_cast<Packet4ui>(first), 8));
+#endif
+  }
+};
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet2d,2>& kernel) {
+  Packet2d t0, t1;
+  t0 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_HI);
+  t1 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_LO);
+  kernel.packet[0] = t0;
+  kernel.packet[1] = t1;
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket, const Packet2d& elsePacket) {
+  Packet2l select = { ifPacket.select[0], ifPacket.select[1] };
+  Packet2bl mask = vec_cmpeq(reinterpret_cast<Packet2d>(select), reinterpret_cast<Packet2d>(p2l_ONE));
+  return vec_sel(elsePacket, thenPacket, mask);
+}
+#endif // __VSX__
 } // end namespace internal
 
 } // end namespace Eigen
diff --git a/eigen/Eigen/src/Core/arch/CMakeLists.txt b/eigen/Eigen/src/Core/arch/CMakeLists.txt
deleted file mode 100644
index 8456dec..0000000
--- a/eigen/Eigen/src/Core/arch/CMakeLists.txt
+++ /dev/null
@@ -1,4 +0,0 @@
-ADD_SUBDIRECTORY(SSE)
-ADD_SUBDIRECTORY(AltiVec)
-ADD_SUBDIRECTORY(NEON)
-ADD_SUBDIRECTORY(Default)
diff --git a/eigen/Eigen/src/Core/arch/CUDA/Complex.h b/eigen/Eigen/src/Core/arch/CUDA/Complex.h
new file mode 100644
index 0000000..9c25365
--- /dev/null
+++ b/eigen/Eigen/src/Core/arch/CUDA/Complex.h
@@ -0,0 +1,103 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_COMPLEX_CUDA_H
+#define EIGEN_COMPLEX_CUDA_H
+
+// clang-format off
+
+namespace Eigen {
+
+namespace internal {
+
+#if defined(__CUDACC__) && defined(EIGEN_USE_GPU)
+
+// Many std::complex methods such as operator+, operator-, operator* and
+// operator/ are not constexpr. Due to this, clang does not treat them as device
+// functions and thus Eigen functors making use of these operators fail to
+// compile. Here, we manually specialize these functors for complex types when
+// building for CUDA to avoid non-constexpr methods.
+
+// Sum
+template<typename T> struct scalar_sum_op<const std::complex<T>, const std::complex<T> > : binary_op_base<const std::complex<T>, const std::complex<T> > {
+  typedef typename std::complex<T> result_type;
+
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_sum_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::complex<T> operator() (const std::complex<T>& a, const std::complex<T>& b) const {
+    return std::complex<T>(numext::real(a) + numext::real(b),
+                           numext::imag(a) + numext::imag(b));
+  }
+};
+
+template<typename T> struct scalar_sum_op<std::complex<T>, std::complex<T> > : scalar_sum_op<const std::complex<T>, const std::complex<T> > {};
+
+
+// Difference
+template<typename T> struct scalar_difference_op<const std::complex<T>, const std::complex<T> >  : binary_op_base<const std::complex<T>, const std::complex<T> > {
+  typedef typename std::complex<T> result_type;
+
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_difference_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::complex<T> operator() (const std::complex<T>& a, const std::complex<T>& b) const {
+    return std::complex<T>(numext::real(a) - numext::real(b),
+                           numext::imag(a) - numext::imag(b));
+  }
+};
+
+template<typename T> struct scalar_difference_op<std::complex<T>, std::complex<T> > : scalar_difference_op<const std::complex<T>, const std::complex<T> > {};
+
+
+// Product
+template<typename T> struct scalar_product_op<const std::complex<T>, const std::complex<T> >  : binary_op_base<const std::complex<T>, const std::complex<T> > {
+  enum {
+    Vectorizable = packet_traits<std::complex<T>>::HasMul
+  };
+  typedef typename std::complex<T> result_type;
+
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_product_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::complex<T> operator() (const std::complex<T>& a, const std::complex<T>& b) const {
+    const T a_real = numext::real(a);
+    const T a_imag = numext::imag(a);
+    const T b_real = numext::real(b);
+    const T b_imag = numext::imag(b);
+    return std::complex<T>(a_real * b_real - a_imag * b_imag,
+                           a_real * b_imag + a_imag * b_real);
+  }
+};
+
+template<typename T> struct scalar_product_op<std::complex<T>, std::complex<T> > : scalar_product_op<const std::complex<T>, const std::complex<T> > {};
+
+
+// Quotient
+template<typename T> struct scalar_quotient_op<const std::complex<T>, const std::complex<T> > : binary_op_base<const std::complex<T>, const std::complex<T> > {
+  enum {
+    Vectorizable = packet_traits<std::complex<T>>::HasDiv
+  };
+  typedef typename std::complex<T> result_type;
+
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_quotient_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::complex<T> operator() (const std::complex<T>& a, const std::complex<T>& b) const {
+    const T a_real = numext::real(a);
+    const T a_imag = numext::imag(a);
+    const T b_real = numext::real(b);
+    const T b_imag = numext::imag(b);
+    const T norm = T(1) / (b_real * b_real + b_imag * b_imag);
+    return std::complex<T>((a_real * b_real + a_imag * b_imag) * norm,
+                           (a_imag * b_real - a_real * b_imag) * norm);
+  }
+};
+
+template<typename T> struct scalar_quotient_op<std::complex<T>, std::complex<T> > : scalar_quotient_op<const std::complex<T>, const std::complex<T> > {};
+
+#endif
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_COMPLEX_CUDA_H
diff --git a/eigen/Eigen/src/Core/arch/CUDA/Half.h b/eigen/Eigen/src/Core/arch/CUDA/Half.h
new file mode 100644
index 0000000..67518da
--- /dev/null
+++ b/eigen/Eigen/src/Core/arch/CUDA/Half.h
@@ -0,0 +1,604 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+//
+// The conversion routines are Copyright (c) Fabian Giesen, 2016.
+// The original license follows:
+//
+// Copyright (c) Fabian Giesen, 2016
+// All rights reserved.
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted.
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+
+// Standard 16-bit float type, mostly useful for GPUs. Defines a new
+// type Eigen::half (inheriting from CUDA's __half struct) with
+// operator overloads such that it behaves basically as an arithmetic
+// type. It will be quite slow on CPUs (so it is recommended to stay
+// in fp32 for CPUs, except for simple parameter conversions, I/O
+// to disk and the likes), but fast on GPUs.
+
+
+#ifndef EIGEN_HALF_CUDA_H
+#define EIGEN_HALF_CUDA_H
+
+#if __cplusplus > 199711L
+#define EIGEN_EXPLICIT_CAST(tgt_type) explicit operator tgt_type()
+#else
+#define EIGEN_EXPLICIT_CAST(tgt_type) operator tgt_type()
+#endif
+
+
+namespace Eigen {
+
+struct half;
+
+namespace half_impl {
+
+#if !defined(EIGEN_HAS_CUDA_FP16)
+
+// Make our own __half definition that is similar to CUDA's.
+struct __half {
+  EIGEN_DEVICE_FUNC __half() : x(0) {}
+  explicit EIGEN_DEVICE_FUNC __half(unsigned short raw) : x(raw) {}
+  unsigned short x;
+};
+
+#endif
+
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half raw_uint16_to_half(unsigned short x);
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half float_to_half_rtne(float ff);
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half h);
+
+struct half_base : public __half {
+  EIGEN_DEVICE_FUNC half_base() {}
+  EIGEN_DEVICE_FUNC half_base(const half_base& h) : __half(h) {}
+  EIGEN_DEVICE_FUNC half_base(const __half& h) : __half(h) {}
+};
+
+} // namespace half_impl
+
+// Class definition.
+struct half : public half_impl::half_base {
+  #if !defined(EIGEN_HAS_CUDA_FP16)
+    typedef half_impl::__half __half;
+  #endif
+
+  EIGEN_DEVICE_FUNC half() {}
+
+  EIGEN_DEVICE_FUNC half(const __half& h) : half_impl::half_base(h) {}
+  EIGEN_DEVICE_FUNC half(const half& h) : half_impl::half_base(h) {}
+
+  explicit EIGEN_DEVICE_FUNC half(bool b)
+      : half_impl::half_base(half_impl::raw_uint16_to_half(b ? 0x3c00 : 0)) {}
+  template<class T>
+  explicit EIGEN_DEVICE_FUNC half(const T& val)
+      : half_impl::half_base(half_impl::float_to_half_rtne(static_cast<float>(val))) {}
+  explicit EIGEN_DEVICE_FUNC half(float f)
+      : half_impl::half_base(half_impl::float_to_half_rtne(f)) {}
+
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(bool) const {
+    // +0.0 and -0.0 become false, everything else becomes true.
+    return (x & 0x7fff) != 0;
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(signed char) const {
+    return static_cast<signed char>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned char) const {
+    return static_cast<unsigned char>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(short) const {
+    return static_cast<short>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned short) const {
+    return static_cast<unsigned short>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(int) const {
+    return static_cast<int>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned int) const {
+    return static_cast<unsigned int>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(long) const {
+    return static_cast<long>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned long) const {
+    return static_cast<unsigned long>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(long long) const {
+    return static_cast<long long>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned long long) const {
+    return static_cast<unsigned long long>(half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(float) const {
+    return half_impl::half_to_float(*this);
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(double) const {
+    return static_cast<double>(half_impl::half_to_float(*this));
+  }
+
+  EIGEN_DEVICE_FUNC half& operator=(const half& other) {
+    x = other.x;
+    return *this;
+  }
+};
+
+namespace half_impl {
+
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+
+// Intrinsics for native fp16 support. Note that on current hardware,
+// these are no faster than fp32 arithmetic (you need to use the half2
+// versions to get the ALU speed increased), but you do save the
+// conversion steps back and forth.
+
+__device__ half operator + (const half& a, const half& b) {
+  return __hadd(a, b);
+}
+__device__ half operator * (const half& a, const half& b) {
+  return __hmul(a, b);
+}
+__device__ half operator - (const half& a, const half& b) {
+  return __hsub(a, b);
+}
+__device__ half operator / (const half& a, const half& b) {
+  float num = __half2float(a);
+  float denom = __half2float(b);
+  return __float2half(num / denom);
+}
+__device__ half operator - (const half& a) {
+  return __hneg(a);
+}
+__device__ half& operator += (half& a, const half& b) {
+  a = a + b;
+  return a;
+}
+__device__ half& operator *= (half& a, const half& b) {
+  a = a * b;
+  return a;
+}
+__device__ half& operator -= (half& a, const half& b) {
+  a = a - b;
+  return a;
+}
+__device__ half& operator /= (half& a, const half& b) {
+  a = a / b;
+  return a;
+}
+__device__ bool operator == (const half& a, const half& b) {
+  return __heq(a, b);
+}
+__device__ bool operator != (const half& a, const half& b) {
+  return __hne(a, b);
+}
+__device__ bool operator < (const half& a, const half& b) {
+  return __hlt(a, b);
+}
+__device__ bool operator <= (const half& a, const half& b) {
+  return __hle(a, b);
+}
+__device__ bool operator > (const half& a, const half& b) {
+  return __hgt(a, b);
+}
+__device__ bool operator >= (const half& a, const half& b) {
+  return __hge(a, b);
+}
+
+#else  // Emulate support for half floats
+
+// Definitions for CPUs and older CUDA, mostly working through conversion
+// to/from fp32.
+
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator + (const half& a, const half& b) {
+  return half(float(a) + float(b));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator * (const half& a, const half& b) {
+  return half(float(a) * float(b));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator - (const half& a, const half& b) {
+  return half(float(a) - float(b));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator / (const half& a, const half& b) {
+  return half(float(a) / float(b));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator - (const half& a) {
+  half result;
+  result.x = a.x ^ 0x8000;
+  return result;
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator += (half& a, const half& b) {
+  a = half(float(a) + float(b));
+  return a;
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator *= (half& a, const half& b) {
+  a = half(float(a) * float(b));
+  return a;
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator -= (half& a, const half& b) {
+  a = half(float(a) - float(b));
+  return a;
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator /= (half& a, const half& b) {
+  a = half(float(a) / float(b));
+  return a;
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator == (const half& a, const half& b) {
+  return float(a) == float(b);
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator != (const half& a, const half& b) {
+  return float(a) != float(b);
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator < (const half& a, const half& b) {
+  return float(a) < float(b);
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator <= (const half& a, const half& b) {
+  return float(a) <= float(b);
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator > (const half& a, const half& b) {
+  return float(a) > float(b);
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator >= (const half& a, const half& b) {
+  return float(a) >= float(b);
+}
+
+#endif  // Emulate support for half floats
+
+// Division by an index. Do it in full float precision to avoid accuracy
+// issues in converting the denominator to half.
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator / (const half& a, Index b) {
+  return half(static_cast<float>(a) / static_cast<float>(b));
+}
+
+// Conversion routines, including fallbacks for the host or older CUDA.
+// Note that newer Intel CPUs (Haswell or newer) have vectorized versions of
+// these in hardware. If we need more performance on older/other CPUs, they are
+// also possible to vectorize directly.
+
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half raw_uint16_to_half(unsigned short x) {
+  __half h;
+  h.x = x;
+  return h;
+}
+
+union FP32 {
+  unsigned int u;
+  float f;
+};
+
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half float_to_half_rtne(float ff) {
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+  return __float2half(ff);
+
+#elif defined(EIGEN_HAS_FP16_C)
+  __half h;
+  h.x = _cvtss_sh(ff, 0);
+  return h;
+
+#else
+  FP32 f; f.f = ff;
+
+  const FP32 f32infty = { 255 << 23 };
+  const FP32 f16max = { (127 + 16) << 23 };
+  const FP32 denorm_magic = { ((127 - 15) + (23 - 10) + 1) << 23 };
+  unsigned int sign_mask = 0x80000000u;
+  __half o;
+  o.x = static_cast<unsigned short>(0x0u);
+
+  unsigned int sign = f.u & sign_mask;
+  f.u ^= sign;
+
+  // NOTE all the integer compares in this function can be safely
+  // compiled into signed compares since all operands are below
+  // 0x80000000. Important if you want fast straight SSE2 code
+  // (since there's no unsigned PCMPGTD).
+
+  if (f.u >= f16max.u) {  // result is Inf or NaN (all exponent bits set)
+    o.x = (f.u > f32infty.u) ? 0x7e00 : 0x7c00; // NaN->qNaN and Inf->Inf
+  } else {  // (De)normalized number or zero
+    if (f.u < (113 << 23)) {  // resulting FP16 is subnormal or zero
+      // use a magic value to align our 10 mantissa bits at the bottom of
+      // the float. as long as FP addition is round-to-nearest-even this
+      // just works.
+      f.f += denorm_magic.f;
+
+      // and one integer subtract of the bias later, we have our final float!
+      o.x = static_cast<unsigned short>(f.u - denorm_magic.u);
+    } else {
+      unsigned int mant_odd = (f.u >> 13) & 1; // resulting mantissa is odd
+
+      // update exponent, rounding bias part 1
+      f.u += ((unsigned int)(15 - 127) << 23) + 0xfff;
+      // rounding bias part 2
+      f.u += mant_odd;
+      // take the bits!
+      o.x = static_cast<unsigned short>(f.u >> 13);
+    }
+  }
+
+  o.x |= static_cast<unsigned short>(sign >> 16);
+  return o;
+#endif
+}
+
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half h) {
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+  return __half2float(h);
+
+#elif defined(EIGEN_HAS_FP16_C)
+  return _cvtsh_ss(h.x);
+
+#else
+  const FP32 magic = { 113 << 23 };
+  const unsigned int shifted_exp = 0x7c00 << 13; // exponent mask after shift
+  FP32 o;
+
+  o.u = (h.x & 0x7fff) << 13;             // exponent/mantissa bits
+  unsigned int exp = shifted_exp & o.u;   // just the exponent
+  o.u += (127 - 15) << 23;                // exponent adjust
+
+  // handle exponent special cases
+  if (exp == shifted_exp) {     // Inf/NaN?
+    o.u += (128 - 16) << 23;    // extra exp adjust
+  } else if (exp == 0) {        // Zero/Denormal?
+    o.u += 1 << 23;             // extra exp adjust
+    o.f -= magic.f;             // renormalize
+  }
+
+  o.u |= (h.x & 0x8000) << 16;    // sign bit
+  return o.f;
+#endif
+}
+
+// --- standard functions ---
+
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isinf)(const half& a) {
+  return (a.x & 0x7fff) == 0x7c00;
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isnan)(const half& a) {
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+  return __hisnan(a);
+#else
+  return (a.x & 0x7fff) > 0x7c00;
+#endif
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isfinite)(const half& a) {
+  return !(isinf EIGEN_NOT_A_MACRO (a)) && !(isnan EIGEN_NOT_A_MACRO (a));
+}
+
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half abs(const half& a) {
+  half result;
+  result.x = a.x & 0x7FFF;
+  return result;
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half exp(const half& a) {
+#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 530
+  return half(hexp(a));
+#else
+   return half(::expf(float(a)));
+#endif
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half expm1(const half& a) {
+  return half(numext::expm1(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log(const half& a) {
+#if defined(EIGEN_HAS_CUDA_FP16) && defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+  return half(::hlog(a));
+#else
+  return half(::logf(float(a)));
+#endif
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log1p(const half& a) {
+  return half(numext::log1p(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log10(const half& a) {
+  return half(::log10f(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half sqrt(const half& a) {
+#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 530
+  return half(hsqrt(a));
+#else
+    return half(::sqrtf(float(a)));
+#endif
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half pow(const half& a, const half& b) {
+  return half(::powf(float(a), float(b)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half sin(const half& a) {
+  return half(::sinf(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half cos(const half& a) {
+  return half(::cosf(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half tan(const half& a) {
+  return half(::tanf(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half tanh(const half& a) {
+  return half(::tanhf(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half floor(const half& a) {
+#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 300
+  return half(hfloor(a));
+#else
+  return half(::floorf(float(a)));
+#endif
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half ceil(const half& a) {
+#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 300
+  return half(hceil(a));
+#else
+  return half(::ceilf(float(a)));
+#endif
+}
+
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half (min)(const half& a, const half& b) {
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+  return __hlt(b, a) ? b : a;
+#else
+  const float f1 = static_cast<float>(a);
+  const float f2 = static_cast<float>(b);
+  return f2 < f1 ? b : a;
+#endif
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half (max)(const half& a, const half& b) {
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+  return __hlt(a, b) ? b : a;
+#else
+  const float f1 = static_cast<float>(a);
+  const float f2 = static_cast<float>(b);
+  return f1 < f2 ? b : a;
+#endif
+}
+
+EIGEN_ALWAYS_INLINE std::ostream& operator << (std::ostream& os, const half& v) {
+  os << static_cast<float>(v);
+  return os;
+}
+
+} // end namespace half_impl
+
+// import Eigen::half_impl::half into Eigen namespace
+// using half_impl::half;
+
+namespace internal {
+
+template<>
+struct random_default_impl<half, false, false>
+{
+  static inline half run(const half& x, const half& y)
+  {
+    return x + (y-x) * half(float(std::rand()) / float(RAND_MAX));
+  }
+  static inline half run()
+  {
+    return run(half(-1.f), half(1.f));
+  }
+};
+
+template<> struct is_arithmetic<half> { enum { value = true }; };
+
+} // end namespace internal
+
+template<> struct NumTraits<Eigen::half>
+    : GenericNumTraits<Eigen::half>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half epsilon() {
+    return half_impl::raw_uint16_to_half(0x0800);
+  }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half dummy_precision() { return Eigen::half(1e-2f); }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half highest() {
+    return half_impl::raw_uint16_to_half(0x7bff);
+  }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half lowest() {
+    return half_impl::raw_uint16_to_half(0xfbff);
+  }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half infinity() {
+    return half_impl::raw_uint16_to_half(0x7c00);
+  }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half quiet_NaN() {
+    return half_impl::raw_uint16_to_half(0x7c01);
+  }
+};
+
+} // end namespace Eigen
+
+// C-like standard mathematical functions and trancendentals.
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half fabsh(const Eigen::half& a) {
+  Eigen::half result;
+  result.x = a.x & 0x7FFF;
+  return result;
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half exph(const Eigen::half& a) {
+  return Eigen::half(::expf(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half logh(const Eigen::half& a) {
+#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+  return Eigen::half(::hlog(a));
+#else
+  return Eigen::half(::logf(float(a)));
+#endif
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half sqrth(const Eigen::half& a) {
+  return Eigen::half(::sqrtf(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half powh(const Eigen::half& a, const Eigen::half& b) {
+  return Eigen::half(::powf(float(a), float(b)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half floorh(const Eigen::half& a) {
+  return Eigen::half(::floorf(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half ceilh(const Eigen::half& a) {
+  return Eigen::half(::ceilf(float(a)));
+}
+
+namespace std {
+
+#if __cplusplus > 199711L
+template <>
+struct hash<Eigen::half> {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::size_t operator()(const Eigen::half& a) const {
+    return static_cast<std::size_t>(a.x);
+  }
+};
+#endif
+
+} // end namespace std
+
+
+// Add the missing shfl_xor intrinsic
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+__device__ EIGEN_STRONG_INLINE Eigen::half __shfl_xor(Eigen::half var, int laneMask, int width=warpSize) {
+  return static_cast<Eigen::half>(__shfl_xor(static_cast<float>(var), laneMask, width));
+}
+#endif
+
+// ldg() has an overload for __half, but we also need one for Eigen::half.
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half __ldg(const Eigen::half* ptr) {
+  return Eigen::half_impl::raw_uint16_to_half(
+      __ldg(reinterpret_cast<const unsigned short*>(ptr)));
+}
+#endif
+
+
+#if defined(__CUDA_ARCH__)
+namespace Eigen {
+namespace numext {
+
+template<>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+bool (isnan)(const Eigen::half& h) {
+  return (half_impl::isnan)(h);
+}
+
+template<>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+bool (isinf)(const Eigen::half& h) {
+  return (half_impl::isinf)(h);
+}
+
+template<>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+bool (isfinite)(const Eigen::half& h) {
+  return (half_impl::isfinite)(h);
+}
+
+} // namespace Eigen
+}  // namespace numext
+#endif
+
+#endif // EIGEN_HALF_CUDA_H
diff --git a/eigen/Eigen/src/Core/arch/CUDA/MathFunctions.h b/eigen/Eigen/src/Core/arch/CUDA/MathFunctions.h
new file mode 100644
index 0000000..987a529
--- /dev/null
+++ b/eigen/Eigen/src/Core/arch/CUDA/MathFunctions.h
@@ -0,0 +1,103 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MATH_FUNCTIONS_CUDA_H
+#define EIGEN_MATH_FUNCTIONS_CUDA_H
+
+namespace Eigen {
+
+namespace internal {
+
+// Make sure this is only available when targeting a GPU: we don't want to
+// introduce conflicts between these packet_traits definitions and the ones
+// we'll use on the host side (SSE, AVX, ...)
+#if defined(__CUDACC__) && defined(EIGEN_USE_GPU)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 plog<float4>(const float4& a)
+{
+  return make_float4(logf(a.x), logf(a.y), logf(a.z), logf(a.w));
+}
+
+template<>  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 plog<double2>(const double2& a)
+{
+  using ::log;
+  return make_double2(log(a.x), log(a.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 plog1p<float4>(const float4& a)
+{
+  return make_float4(log1pf(a.x), log1pf(a.y), log1pf(a.z), log1pf(a.w));
+}
+
+template<>  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 plog1p<double2>(const double2& a)
+{
+  return make_double2(log1p(a.x), log1p(a.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 pexp<float4>(const float4& a)
+{
+  return make_float4(expf(a.x), expf(a.y), expf(a.z), expf(a.w));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 pexp<double2>(const double2& a)
+{
+  using ::exp;
+  return make_double2(exp(a.x), exp(a.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 pexpm1<float4>(const float4& a)
+{
+  return make_float4(expm1f(a.x), expm1f(a.y), expm1f(a.z), expm1f(a.w));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 pexpm1<double2>(const double2& a)
+{
+  return make_double2(expm1(a.x), expm1(a.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 psqrt<float4>(const float4& a)
+{
+  return make_float4(sqrtf(a.x), sqrtf(a.y), sqrtf(a.z), sqrtf(a.w));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 psqrt<double2>(const double2& a)
+{
+  using ::sqrt;
+  return make_double2(sqrt(a.x), sqrt(a.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 prsqrt<float4>(const float4& a)
+{
+  return make_float4(rsqrtf(a.x), rsqrtf(a.y), rsqrtf(a.z), rsqrtf(a.w));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 prsqrt<double2>(const double2& a)
+{
+  return make_double2(rsqrt(a.x), rsqrt(a.y));
+}
+
+
+#endif
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_MATH_FUNCTIONS_CUDA_H
diff --git a/eigen/Eigen/src/Core/arch/CUDA/PacketMath.h b/eigen/Eigen/src/Core/arch/CUDA/PacketMath.h
new file mode 100644
index 0000000..8c46af0
--- /dev/null
+++ b/eigen/Eigen/src/Core/arch/CUDA/PacketMath.h
@@ -0,0 +1,333 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PACKET_MATH_CUDA_H
+#define EIGEN_PACKET_MATH_CUDA_H
+
+namespace Eigen {
+
+namespace internal {
+
+// Make sure this is only available when targeting a GPU: we don't want to
+// introduce conflicts between these packet_traits definitions and the ones
+// we'll use on the host side (SSE, AVX, ...)
+#if defined(__CUDACC__) && defined(EIGEN_USE_GPU)
+template<> struct is_arithmetic<float4>  { enum { value = true }; };
+template<> struct is_arithmetic<double2> { enum { value = true }; };
+
+template<> struct packet_traits<float> : default_packet_traits
+{
+  typedef float4 type;
+  typedef float4 half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=4,
+    HasHalfPacket = 0,
+
+    HasDiv  = 1,
+    HasSin  = 0,
+    HasCos  = 0,
+    HasLog  = 1,
+    HasExp  = 1,
+    HasSqrt = 1,
+    HasRsqrt = 1,
+    HasLGamma = 1,
+    HasDiGamma = 1,
+    HasZeta = 1,
+    HasPolygamma = 1,
+    HasErf = 1,
+    HasErfc = 1,
+    HasIGamma = 1,
+    HasIGammac = 1,
+    HasBetaInc = 1,
+
+    HasBlend = 0,
+  };
+};
+
+template<> struct packet_traits<double> : default_packet_traits
+{
+  typedef double2 type;
+  typedef double2 half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=2,
+    HasHalfPacket = 0,
+
+    HasDiv  = 1,
+    HasLog  = 1,
+    HasExp  = 1,
+    HasSqrt = 1,
+    HasRsqrt = 1,
+    HasLGamma = 1,
+    HasDiGamma = 1,
+    HasZeta = 1,
+    HasPolygamma = 1,
+    HasErf = 1,
+    HasErfc = 1,
+    HasIGamma = 1,
+    HasIGammac = 1,
+    HasBetaInc = 1,
+
+    HasBlend = 0,
+  };
+};
+
+
+template<> struct unpacket_traits<float4>  { typedef float  type; enum {size=4, alignment=Aligned16}; typedef float4 half; };
+template<> struct unpacket_traits<double2> { typedef double type; enum {size=2, alignment=Aligned16}; typedef double2 half; };
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pset1<float4>(const float&  from) {
+  return make_float4(from, from, from, from);
+}
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pset1<double2>(const double& from) {
+  return make_double2(from, from);
+}
+
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 plset<float4>(const float& a) {
+  return make_float4(a, a+1, a+2, a+3);
+}
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 plset<double2>(const double& a) {
+  return make_double2(a, a+1);
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 padd<float4>(const float4& a, const float4& b) {
+  return make_float4(a.x+b.x, a.y+b.y, a.z+b.z, a.w+b.w);
+}
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 padd<double2>(const double2& a, const double2& b) {
+  return make_double2(a.x+b.x, a.y+b.y);
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 psub<float4>(const float4& a, const float4& b) {
+  return make_float4(a.x-b.x, a.y-b.y, a.z-b.z, a.w-b.w);
+}
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 psub<double2>(const double2& a, const double2& b) {
+  return make_double2(a.x-b.x, a.y-b.y);
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pnegate(const float4& a) {
+  return make_float4(-a.x, -a.y, -a.z, -a.w);
+}
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pnegate(const double2& a) {
+  return make_double2(-a.x, -a.y);
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pconj(const float4& a) { return a; }
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pconj(const double2& a) { return a; }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pmul<float4>(const float4& a, const float4& b) {
+  return make_float4(a.x*b.x, a.y*b.y, a.z*b.z, a.w*b.w);
+}
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pmul<double2>(const double2& a, const double2& b) {
+  return make_double2(a.x*b.x, a.y*b.y);
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pdiv<float4>(const float4& a, const float4& b) {
+  return make_float4(a.x/b.x, a.y/b.y, a.z/b.z, a.w/b.w);
+}
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pdiv<double2>(const double2& a, const double2& b) {
+  return make_double2(a.x/b.x, a.y/b.y);
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pmin<float4>(const float4& a, const float4& b) {
+  return make_float4(fminf(a.x, b.x), fminf(a.y, b.y), fminf(a.z, b.z), fminf(a.w, b.w));
+}
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pmin<double2>(const double2& a, const double2& b) {
+  return make_double2(fmin(a.x, b.x), fmin(a.y, b.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pmax<float4>(const float4& a, const float4& b) {
+  return make_float4(fmaxf(a.x, b.x), fmaxf(a.y, b.y), fmaxf(a.z, b.z), fmaxf(a.w, b.w));
+}
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pmax<double2>(const double2& a, const double2& b) {
+  return make_double2(fmax(a.x, b.x), fmax(a.y, b.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pload<float4>(const float* from) {
+  return *reinterpret_cast<const float4*>(from);
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pload<double2>(const double* from) {
+  return *reinterpret_cast<const double2*>(from);
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 ploadu<float4>(const float* from) {
+  return make_float4(from[0], from[1], from[2], from[3]);
+}
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 ploadu<double2>(const double* from) {
+  return make_double2(from[0], from[1]);
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 ploaddup<float4>(const float*   from) {
+  return make_float4(from[0], from[0], from[1], from[1]);
+}
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 ploaddup<double2>(const double*  from) {
+  return make_double2(from[0], from[0]);
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstore<float>(float*   to, const float4& from) {
+  *reinterpret_cast<float4*>(to) = from;
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstore<double>(double* to, const double2& from) {
+  *reinterpret_cast<double2*>(to) = from;
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstoreu<float>(float*  to, const float4& from) {
+  to[0] = from.x;
+  to[1] = from.y;
+  to[2] = from.z;
+  to[3] = from.w;
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const double2& from) {
+  to[0] = from.x;
+  to[1] = from.y;
+}
+
+template<>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE float4 ploadt_ro<float4, Aligned>(const float* from) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350
+  return __ldg((const float4*)from);
+#else
+  return make_float4(from[0], from[1], from[2], from[3]);
+#endif
+}
+template<>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE double2 ploadt_ro<double2, Aligned>(const double* from) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350
+  return __ldg((const double2*)from);
+#else
+  return make_double2(from[0], from[1]);
+#endif
+}
+
+template<>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE float4 ploadt_ro<float4, Unaligned>(const float* from) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350
+  return make_float4(__ldg(from+0), __ldg(from+1), __ldg(from+2), __ldg(from+3));
+#else
+  return make_float4(from[0], from[1], from[2], from[3]);
+#endif
+}
+template<>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE double2 ploadt_ro<double2, Unaligned>(const double* from) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350
+  return make_double2(__ldg(from+0), __ldg(from+1));
+#else
+  return make_double2(from[0], from[1]);
+#endif
+}
+
+template<> EIGEN_DEVICE_FUNC inline float4 pgather<float, float4>(const float* from, Index stride) {
+  return make_float4(from[0*stride], from[1*stride], from[2*stride], from[3*stride]);
+}
+
+template<> EIGEN_DEVICE_FUNC inline double2 pgather<double, double2>(const double* from, Index stride) {
+  return make_double2(from[0*stride], from[1*stride]);
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<float, float4>(float* to, const float4& from, Index stride) {
+  to[stride*0] = from.x;
+  to[stride*1] = from.y;
+  to[stride*2] = from.z;
+  to[stride*3] = from.w;
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<double, double2>(double* to, const double2& from, Index stride) {
+  to[stride*0] = from.x;
+  to[stride*1] = from.y;
+}
+
+template<> EIGEN_DEVICE_FUNC inline float  pfirst<float4>(const float4& a) {
+  return a.x;
+}
+template<> EIGEN_DEVICE_FUNC inline double pfirst<double2>(const double2& a) {
+  return a.x;
+}
+
+template<> EIGEN_DEVICE_FUNC inline float  predux<float4>(const float4& a) {
+  return a.x + a.y + a.z + a.w;
+}
+template<> EIGEN_DEVICE_FUNC inline double predux<double2>(const double2& a) {
+  return a.x + a.y;
+}
+
+template<> EIGEN_DEVICE_FUNC inline float  predux_max<float4>(const float4& a) {
+  return fmaxf(fmaxf(a.x, a.y), fmaxf(a.z, a.w));
+}
+template<> EIGEN_DEVICE_FUNC inline double predux_max<double2>(const double2& a) {
+  return fmax(a.x, a.y);
+}
+
+template<> EIGEN_DEVICE_FUNC inline float  predux_min<float4>(const float4& a) {
+  return fminf(fminf(a.x, a.y), fminf(a.z, a.w));
+}
+template<> EIGEN_DEVICE_FUNC inline double predux_min<double2>(const double2& a) {
+  return fmin(a.x, a.y);
+}
+
+template<> EIGEN_DEVICE_FUNC inline float  predux_mul<float4>(const float4& a) {
+  return a.x * a.y * a.z * a.w;
+}
+template<> EIGEN_DEVICE_FUNC inline double predux_mul<double2>(const double2& a) {
+  return a.x * a.y;
+}
+
+template<> EIGEN_DEVICE_FUNC inline float4  pabs<float4>(const float4& a) {
+  return make_float4(fabsf(a.x), fabsf(a.y), fabsf(a.z), fabsf(a.w));
+}
+template<> EIGEN_DEVICE_FUNC inline double2 pabs<double2>(const double2& a) {
+  return make_double2(fabs(a.x), fabs(a.y));
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<float4,4>& kernel) {
+  float tmp = kernel.packet[0].y;
+  kernel.packet[0].y = kernel.packet[1].x;
+  kernel.packet[1].x = tmp;
+
+  tmp = kernel.packet[0].z;
+  kernel.packet[0].z = kernel.packet[2].x;
+  kernel.packet[2].x = tmp;
+
+  tmp = kernel.packet[0].w;
+  kernel.packet[0].w = kernel.packet[3].x;
+  kernel.packet[3].x = tmp;
+
+  tmp = kernel.packet[1].z;
+  kernel.packet[1].z = kernel.packet[2].y;
+  kernel.packet[2].y = tmp;
+
+  tmp = kernel.packet[1].w;
+  kernel.packet[1].w = kernel.packet[3].y;
+  kernel.packet[3].y = tmp;
+
+  tmp = kernel.packet[2].w;
+  kernel.packet[2].w = kernel.packet[3].z;
+  kernel.packet[3].z = tmp;
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<double2,2>& kernel) {
+  double tmp = kernel.packet[0].y;
+  kernel.packet[0].y = kernel.packet[1].x;
+  kernel.packet[1].x = tmp;
+}
+
+#endif
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+
+#endif // EIGEN_PACKET_MATH_CUDA_H
diff --git a/eigen/Eigen/src/Core/arch/CUDA/PacketMathHalf.h b/eigen/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
new file mode 100644
index 0000000..b9a125b
--- /dev/null
+++ b/eigen/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
@@ -0,0 +1,1132 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PACKET_MATH_HALF_CUDA_H
+#define EIGEN_PACKET_MATH_HALF_CUDA_H
+
+
+namespace Eigen {
+namespace internal {
+
+// Most of the following operations require arch >= 3.0
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDACC__) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+
+template<> struct is_arithmetic<half2> { enum { value = true }; };
+
+template<> struct packet_traits<Eigen::half> : default_packet_traits
+{
+  typedef half2 type;
+  typedef half2 half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=2,
+    HasHalfPacket = 0,
+    HasAdd    = 1,
+    HasMul    = 1,
+    HasDiv    = 1,
+    HasSqrt   = 1,
+    HasRsqrt  = 1,
+    HasExp    = 1,
+    HasExpm1  = 1,
+    HasLog    = 1,
+    HasLog1p  = 1
+  };
+};
+
+template<> struct unpacket_traits<half2> { typedef Eigen::half type; enum {size=2, alignment=Aligned16}; typedef half2 half; };
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pset1<half2>(const Eigen::half& from) {
+  return __half2half2(from);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pload<half2>(const Eigen::half* from) {
+  return *reinterpret_cast<const half2*>(from);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 ploadu<half2>(const Eigen::half* from) {
+  return __halves2half2(from[0], from[1]);
+}
+
+template<> EIGEN_STRONG_INLINE half2 ploaddup<half2>(const Eigen::half*  from) {
+  return __halves2half2(from[0], from[0]);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE void pstore<Eigen::half>(Eigen::half* to, const half2& from) {
+  *reinterpret_cast<half2*>(to) = from;
+}
+
+template<> __device__ EIGEN_STRONG_INLINE void pstoreu<Eigen::half>(Eigen::half* to, const half2& from) {
+  to[0] = __low2half(from);
+  to[1] = __high2half(from);
+}
+
+template<>
+ __device__ EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Aligned>(const Eigen::half* from) {
+#if __CUDA_ARCH__ >= 350
+   return __ldg((const half2*)from);
+#else
+  return __halves2half2(*(from+0), *(from+1));
+#endif
+}
+
+template<>
+__device__ EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Unaligned>(const Eigen::half* from) {
+#if __CUDA_ARCH__ >= 350
+   return __halves2half2(__ldg(from+0), __ldg(from+1));
+#else
+  return __halves2half2(*(from+0), *(from+1));
+#endif
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pgather<Eigen::half, half2>(const Eigen::half* from, Index stride) {
+  return __halves2half2(from[0*stride], from[1*stride]);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE void pscatter<Eigen::half, half2>(Eigen::half* to, const half2& from, Index stride) {
+  to[stride*0] = __low2half(from);
+  to[stride*1] = __high2half(from);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE Eigen::half pfirst<half2>(const half2& a) {
+  return __low2half(a);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pabs<half2>(const half2& a) {
+  half2 result;
+  result.x = a.x & 0x7FFF7FFF;
+  return result;
+}
+
+
+__device__ EIGEN_STRONG_INLINE void
+ptranspose(PacketBlock<half2,2>& kernel) {
+  __half a1 = __low2half(kernel.packet[0]);
+  __half a2 = __high2half(kernel.packet[0]);
+  __half b1 = __low2half(kernel.packet[1]);
+  __half b2 = __high2half(kernel.packet[1]);
+  kernel.packet[0] = __halves2half2(a1, b1);
+  kernel.packet[1] = __halves2half2(a2, b2);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 plset<half2>(const Eigen::half& a) {
+#if __CUDA_ARCH__ >= 530
+  return __halves2half2(a, __hadd(a, __float2half(1.0f)));
+#else
+  float f = __half2float(a) + 1.0f;
+  return __halves2half2(a, __float2half(f));
+#endif
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 padd<half2>(const half2& a, const half2& b) {
+#if __CUDA_ARCH__ >= 530
+  return __hadd2(a, b);
+#else
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float b1 = __low2float(b);
+  float b2 = __high2float(b);
+  float r1 = a1 + b1;
+  float r2 = a2 + b2;
+  return __floats2half2_rn(r1, r2);
+#endif
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 psub<half2>(const half2& a, const half2& b) {
+#if __CUDA_ARCH__ >= 530
+  return __hsub2(a, b);
+#else
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float b1 = __low2float(b);
+  float b2 = __high2float(b);
+  float r1 = a1 - b1;
+  float r2 = a2 - b2;
+  return __floats2half2_rn(r1, r2);
+#endif
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pnegate(const half2& a) {
+#if __CUDA_ARCH__ >= 530
+  return __hneg2(a);
+#else
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  return __floats2half2_rn(-a1, -a2);
+#endif
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pconj(const half2& a) { return a; }
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pmul<half2>(const half2& a, const half2& b) {
+#if __CUDA_ARCH__ >= 530
+  return __hmul2(a, b);
+#else
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float b1 = __low2float(b);
+  float b2 = __high2float(b);
+  float r1 = a1 * b1;
+  float r2 = a2 * b2;
+  return __floats2half2_rn(r1, r2);
+#endif
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pmadd<half2>(const half2& a, const half2& b, const half2& c) {
+#if __CUDA_ARCH__ >= 530
+   return __hfma2(a, b, c);
+#else
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float b1 = __low2float(b);
+  float b2 = __high2float(b);
+  float c1 = __low2float(c);
+  float c2 = __high2float(c);
+  float r1 = a1 * b1 + c1;
+  float r2 = a2 * b2 + c2;
+  return __floats2half2_rn(r1, r2);
+#endif
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pdiv<half2>(const half2& a, const half2& b) {
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float b1 = __low2float(b);
+  float b2 = __high2float(b);
+  float r1 = a1 / b1;
+  float r2 = a2 / b2;
+  return __floats2half2_rn(r1, r2);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pmin<half2>(const half2& a, const half2& b) {
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float b1 = __low2float(b);
+  float b2 = __high2float(b);
+  __half r1 = a1 < b1 ? __low2half(a) : __low2half(b);
+  __half r2 = a2 < b2 ? __high2half(a) : __high2half(b);
+  return __halves2half2(r1, r2);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pmax<half2>(const half2& a, const half2& b) {
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float b1 = __low2float(b);
+  float b2 = __high2float(b);
+  __half r1 = a1 > b1 ? __low2half(a) : __low2half(b);
+  __half r2 = a2 > b2 ? __high2half(a) : __high2half(b);
+  return __halves2half2(r1, r2);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux<half2>(const half2& a) {
+#if __CUDA_ARCH__ >= 530
+  return __hadd(__low2half(a), __high2half(a));
+#else
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  return Eigen::half(half_impl::raw_uint16_to_half(__float2half_rn(a1 + a2)));
+#endif
+}
+
+template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_max<half2>(const half2& a) {
+#if __CUDA_ARCH__ >= 530
+  __half first = __low2half(a);
+  __half second = __high2half(a);
+  return __hgt(first, second) ? first : second;
+#else
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  return a1 > a2 ? __low2half(a) : __high2half(a);
+#endif
+}
+
+template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_min<half2>(const half2& a) {
+#if __CUDA_ARCH__ >= 530
+  __half first = __low2half(a);
+  __half second = __high2half(a);
+  return __hlt(first, second) ? first : second;
+#else
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  return a1 < a2 ? __low2half(a) : __high2half(a);
+#endif
+}
+
+template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_mul<half2>(const half2& a) {
+#if __CUDA_ARCH__ >= 530
+  return __hmul(__low2half(a), __high2half(a));
+#else
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  return Eigen::half(half_impl::raw_uint16_to_half(__float2half_rn(a1 * a2)));
+#endif
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 plog1p<half2>(const half2& a) {
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float r1 = log1pf(a1);
+  float r2 = log1pf(a2);
+  return __floats2half2_rn(r1, r2);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pexpm1<half2>(const half2& a) {
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float r1 = expm1f(a1);
+  float r2 = expm1f(a2);
+  return __floats2half2_rn(r1, r2);
+}
+
+#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 530
+
+template<>  __device__ EIGEN_STRONG_INLINE
+half2 plog<half2>(const half2& a) {
+  return h2log(a);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE
+half2 pexp<half2>(const half2& a) {
+  return h2exp(a);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE
+half2 psqrt<half2>(const half2& a) {
+  return h2sqrt(a);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE
+half2 prsqrt<half2>(const half2& a) {
+  return h2rsqrt(a);
+}
+
+#else
+
+template<> __device__ EIGEN_STRONG_INLINE half2 plog<half2>(const half2& a) {
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float r1 = logf(a1);
+  float r2 = logf(a2);
+  return __floats2half2_rn(r1, r2);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pexp<half2>(const half2& a) {
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float r1 = expf(a1);
+  float r2 = expf(a2);
+  return __floats2half2_rn(r1, r2);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 psqrt<half2>(const half2& a) {
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float r1 = sqrtf(a1);
+  float r2 = sqrtf(a2);
+  return __floats2half2_rn(r1, r2);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 prsqrt<half2>(const half2& a) {
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float r1 = rsqrtf(a1);
+  float r2 = rsqrtf(a2);
+  return __floats2half2_rn(r1, r2);
+}
+
+#endif
+
+#elif defined EIGEN_VECTORIZE_AVX512
+
+typedef struct {
+  __m256i x;
+} Packet16h;
+
+
+template<> struct is_arithmetic<Packet16h> { enum { value = true }; };
+
+template <>
+struct packet_traits<half> : default_packet_traits {
+  typedef Packet16h type;
+  // There is no half-size packet for Packet16h.
+  typedef Packet16h half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 16,
+    HasHalfPacket = 0,
+    HasAdd    = 0,
+    HasSub    = 0,
+    HasMul    = 0,
+    HasNegate = 0,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasConj   = 0,
+    HasSetLinear = 0,
+    HasDiv = 0,
+    HasSqrt = 0,
+    HasRsqrt = 0,
+    HasExp = 0,
+    HasLog = 0,
+    HasBlend = 0
+  };
+};
+
+
+template<> struct unpacket_traits<Packet16h> { typedef Eigen::half type; enum {size=16, alignment=Aligned32}; typedef Packet16h half; };
+
+template<> EIGEN_STRONG_INLINE Packet16h pset1<Packet16h>(const Eigen::half& from) {
+  Packet16h result;
+  result.x = _mm256_set1_epi16(from.x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Eigen::half pfirst<Packet16h>(const Packet16h& from) {
+  return half_impl::raw_uint16_to_half(static_cast<unsigned short>(_mm256_extract_epi16(from.x, 0)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h pload<Packet16h>(const Eigen::half* from) {
+  Packet16h result;
+  result.x = _mm256_load_si256(reinterpret_cast<const __m256i*>(from));
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h ploadu<Packet16h>(const Eigen::half* from) {
+  Packet16h result;
+  result.x = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(from));
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE void pstore<half>(Eigen::half* to, const Packet16h& from) {
+  _mm256_store_si256((__m256i*)to, from.x);
+}
+
+template<> EIGEN_STRONG_INLINE void pstoreu<half>(Eigen::half* to, const Packet16h& from) {
+  _mm256_storeu_si256((__m256i*)to, from.x);
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h
+ploadquad(const Eigen::half* from) {
+  Packet16h result;
+  unsigned short a = from[0].x;
+  unsigned short b = from[1].x;
+  unsigned short c = from[2].x;
+  unsigned short d = from[3].x;
+  result.x = _mm256_set_epi16(d, d, d, d, c, c, c, c, b, b, b, b, a, a, a, a);
+  return result;
+}
+
+EIGEN_STRONG_INLINE Packet16f half2float(const Packet16h& a) {
+#ifdef EIGEN_HAS_FP16_C
+  return _mm512_cvtph_ps(a.x);
+#else
+  EIGEN_ALIGN64 half aux[16];
+  pstore(aux, a);
+  float f0(aux[0]);
+  float f1(aux[1]);
+  float f2(aux[2]);
+  float f3(aux[3]);
+  float f4(aux[4]);
+  float f5(aux[5]);
+  float f6(aux[6]);
+  float f7(aux[7]);
+  float f8(aux[8]);
+  float f9(aux[9]);
+  float fa(aux[10]);
+  float fb(aux[11]);
+  float fc(aux[12]);
+  float fd(aux[13]);
+  float fe(aux[14]);
+  float ff(aux[15]);
+
+  return _mm512_set_ps(
+      ff, fe, fd, fc, fb, fa, f9, f8, f7, f6, f5, f4, f3, f2, f1, f0);
+#endif
+}
+
+EIGEN_STRONG_INLINE Packet16h float2half(const Packet16f& a) {
+#ifdef EIGEN_HAS_FP16_C
+  Packet16h result;
+  result.x = _mm512_cvtps_ph(a, _MM_FROUND_TO_NEAREST_INT|_MM_FROUND_NO_EXC);
+  return result;
+#else
+  EIGEN_ALIGN64 float aux[16];
+  pstore(aux, a);
+  half h0(aux[0]);
+  half h1(aux[1]);
+  half h2(aux[2]);
+  half h3(aux[3]);
+  half h4(aux[4]);
+  half h5(aux[5]);
+  half h6(aux[6]);
+  half h7(aux[7]);
+  half h8(aux[8]);
+  half h9(aux[9]);
+  half ha(aux[10]);
+  half hb(aux[11]);
+  half hc(aux[12]);
+  half hd(aux[13]);
+  half he(aux[14]);
+  half hf(aux[15]);
+
+  Packet16h result;
+  result.x = _mm256_set_epi16(
+      hf.x, he.x, hd.x, hc.x, hb.x, ha.x, h9.x, h8.x,
+      h7.x, h6.x, h5.x, h4.x, h3.x, h2.x, h1.x, h0.x);
+  return result;
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h padd<Packet16h>(const Packet16h& a, const Packet16h& b) {
+  Packet16f af = half2float(a);
+  Packet16f bf = half2float(b);
+  Packet16f rf = padd(af, bf);
+  return float2half(rf);
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h pmul<Packet16h>(const Packet16h& a, const Packet16h& b) {
+  Packet16f af = half2float(a);
+  Packet16f bf = half2float(b);
+  Packet16f rf = pmul(af, bf);
+  return float2half(rf);
+}
+
+template<> EIGEN_STRONG_INLINE half predux<Packet16h>(const Packet16h& from) {
+  Packet16f from_float = half2float(from);
+  return half(predux(from_float));
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h pgather<Eigen::half, Packet16h>(const Eigen::half* from, Index stride)
+{
+  Packet16h result;
+  result.x = _mm256_set_epi16(
+      from[15*stride].x, from[14*stride].x, from[13*stride].x, from[12*stride].x,
+      from[11*stride].x, from[10*stride].x, from[9*stride].x, from[8*stride].x,
+      from[7*stride].x, from[6*stride].x, from[5*stride].x, from[4*stride].x,
+      from[3*stride].x, from[2*stride].x, from[1*stride].x, from[0*stride].x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE void pscatter<half, Packet16h>(half* to, const Packet16h& from, Index stride)
+{
+  EIGEN_ALIGN64 half aux[16];
+  pstore(aux, from);
+  to[stride*0].x = aux[0].x;
+  to[stride*1].x = aux[1].x;
+  to[stride*2].x = aux[2].x;
+  to[stride*3].x = aux[3].x;
+  to[stride*4].x = aux[4].x;
+  to[stride*5].x = aux[5].x;
+  to[stride*6].x = aux[6].x;
+  to[stride*7].x = aux[7].x;
+  to[stride*8].x = aux[8].x;
+  to[stride*9].x = aux[9].x;
+  to[stride*10].x = aux[10].x;
+  to[stride*11].x = aux[11].x;
+  to[stride*12].x = aux[12].x;
+  to[stride*13].x = aux[13].x;
+  to[stride*14].x = aux[14].x;
+  to[stride*15].x = aux[15].x;
+}
+
+EIGEN_STRONG_INLINE void
+ptranspose(PacketBlock<Packet16h,16>& kernel) {
+  __m256i a = kernel.packet[0].x;
+  __m256i b = kernel.packet[1].x;
+  __m256i c = kernel.packet[2].x;
+  __m256i d = kernel.packet[3].x;
+  __m256i e = kernel.packet[4].x;
+  __m256i f = kernel.packet[5].x;
+  __m256i g = kernel.packet[6].x;
+  __m256i h = kernel.packet[7].x;
+  __m256i i = kernel.packet[8].x;
+  __m256i j = kernel.packet[9].x;
+  __m256i k = kernel.packet[10].x;
+  __m256i l = kernel.packet[11].x;
+  __m256i m = kernel.packet[12].x;
+  __m256i n = kernel.packet[13].x;
+  __m256i o = kernel.packet[14].x;
+  __m256i p = kernel.packet[15].x;
+
+  __m256i ab_07 = _mm256_unpacklo_epi16(a, b);
+  __m256i cd_07 = _mm256_unpacklo_epi16(c, d);
+  __m256i ef_07 = _mm256_unpacklo_epi16(e, f);
+  __m256i gh_07 = _mm256_unpacklo_epi16(g, h);
+  __m256i ij_07 = _mm256_unpacklo_epi16(i, j);
+  __m256i kl_07 = _mm256_unpacklo_epi16(k, l);
+  __m256i mn_07 = _mm256_unpacklo_epi16(m, n);
+  __m256i op_07 = _mm256_unpacklo_epi16(o, p);
+
+  __m256i ab_8f = _mm256_unpackhi_epi16(a, b);
+  __m256i cd_8f = _mm256_unpackhi_epi16(c, d);
+  __m256i ef_8f = _mm256_unpackhi_epi16(e, f);
+  __m256i gh_8f = _mm256_unpackhi_epi16(g, h);
+  __m256i ij_8f = _mm256_unpackhi_epi16(i, j);
+  __m256i kl_8f = _mm256_unpackhi_epi16(k, l);
+  __m256i mn_8f = _mm256_unpackhi_epi16(m, n);
+  __m256i op_8f = _mm256_unpackhi_epi16(o, p);
+
+  __m256i abcd_03 = _mm256_unpacklo_epi32(ab_07, cd_07);
+  __m256i abcd_47 = _mm256_unpackhi_epi32(ab_07, cd_07);
+  __m256i efgh_03 = _mm256_unpacklo_epi32(ef_07, gh_07);
+  __m256i efgh_47 = _mm256_unpackhi_epi32(ef_07, gh_07);
+  __m256i ijkl_03 = _mm256_unpacklo_epi32(ij_07, kl_07);
+  __m256i ijkl_47 = _mm256_unpackhi_epi32(ij_07, kl_07);
+  __m256i mnop_03 = _mm256_unpacklo_epi32(mn_07, op_07);
+  __m256i mnop_47 = _mm256_unpackhi_epi32(mn_07, op_07);
+
+  __m256i abcd_8b = _mm256_unpacklo_epi32(ab_8f, cd_8f);
+  __m256i abcd_cf = _mm256_unpackhi_epi32(ab_8f, cd_8f);
+  __m256i efgh_8b = _mm256_unpacklo_epi32(ef_8f, gh_8f);
+  __m256i efgh_cf = _mm256_unpackhi_epi32(ef_8f, gh_8f);
+  __m256i ijkl_8b = _mm256_unpacklo_epi32(ij_8f, kl_8f);
+  __m256i ijkl_cf = _mm256_unpackhi_epi32(ij_8f, kl_8f);
+  __m256i mnop_8b = _mm256_unpacklo_epi32(mn_8f, op_8f);
+  __m256i mnop_cf = _mm256_unpackhi_epi32(mn_8f, op_8f);
+
+  __m256i abcdefgh_01 = _mm256_unpacklo_epi64(abcd_03, efgh_03);
+  __m256i abcdefgh_23 = _mm256_unpackhi_epi64(abcd_03, efgh_03);
+  __m256i ijklmnop_01 = _mm256_unpacklo_epi64(ijkl_03, mnop_03);
+  __m256i ijklmnop_23 = _mm256_unpackhi_epi64(ijkl_03, mnop_03);
+  __m256i abcdefgh_45 = _mm256_unpacklo_epi64(abcd_47, efgh_47);
+  __m256i abcdefgh_67 = _mm256_unpackhi_epi64(abcd_47, efgh_47);
+  __m256i ijklmnop_45 = _mm256_unpacklo_epi64(ijkl_47, mnop_47);
+  __m256i ijklmnop_67 = _mm256_unpackhi_epi64(ijkl_47, mnop_47);
+  __m256i abcdefgh_89 = _mm256_unpacklo_epi64(abcd_8b, efgh_8b);
+  __m256i abcdefgh_ab = _mm256_unpackhi_epi64(abcd_8b, efgh_8b);
+  __m256i ijklmnop_89 = _mm256_unpacklo_epi64(ijkl_8b, mnop_8b);
+  __m256i ijklmnop_ab = _mm256_unpackhi_epi64(ijkl_8b, mnop_8b);
+  __m256i abcdefgh_cd = _mm256_unpacklo_epi64(abcd_cf, efgh_cf);
+  __m256i abcdefgh_ef = _mm256_unpackhi_epi64(abcd_cf, efgh_cf);
+  __m256i ijklmnop_cd = _mm256_unpacklo_epi64(ijkl_cf, mnop_cf);
+  __m256i ijklmnop_ef = _mm256_unpackhi_epi64(ijkl_cf, mnop_cf);
+
+  // NOTE: no unpacklo/hi instr in this case, so using permute instr.
+  __m256i a_p_0 = _mm256_permute2x128_si256(abcdefgh_01, ijklmnop_01, 0x20);
+  __m256i a_p_1 = _mm256_permute2x128_si256(abcdefgh_01, ijklmnop_01, 0x31);
+  __m256i a_p_2 = _mm256_permute2x128_si256(abcdefgh_23, ijklmnop_23, 0x20);
+  __m256i a_p_3 = _mm256_permute2x128_si256(abcdefgh_23, ijklmnop_23, 0x31);
+  __m256i a_p_4 = _mm256_permute2x128_si256(abcdefgh_45, ijklmnop_45, 0x20);
+  __m256i a_p_5 = _mm256_permute2x128_si256(abcdefgh_45, ijklmnop_45, 0x31);
+  __m256i a_p_6 = _mm256_permute2x128_si256(abcdefgh_67, ijklmnop_67, 0x20);
+  __m256i a_p_7 = _mm256_permute2x128_si256(abcdefgh_67, ijklmnop_67, 0x31);
+  __m256i a_p_8 = _mm256_permute2x128_si256(abcdefgh_89, ijklmnop_89, 0x20);
+  __m256i a_p_9 = _mm256_permute2x128_si256(abcdefgh_89, ijklmnop_89, 0x31);
+  __m256i a_p_a = _mm256_permute2x128_si256(abcdefgh_ab, ijklmnop_ab, 0x20);
+  __m256i a_p_b = _mm256_permute2x128_si256(abcdefgh_ab, ijklmnop_ab, 0x31);
+  __m256i a_p_c = _mm256_permute2x128_si256(abcdefgh_cd, ijklmnop_cd, 0x20);
+  __m256i a_p_d = _mm256_permute2x128_si256(abcdefgh_cd, ijklmnop_cd, 0x31);
+  __m256i a_p_e = _mm256_permute2x128_si256(abcdefgh_ef, ijklmnop_ef, 0x20);
+  __m256i a_p_f = _mm256_permute2x128_si256(abcdefgh_ef, ijklmnop_ef, 0x31);
+
+  kernel.packet[0].x = a_p_0;
+  kernel.packet[1].x = a_p_1;
+  kernel.packet[2].x = a_p_2;
+  kernel.packet[3].x = a_p_3;
+  kernel.packet[4].x = a_p_4;
+  kernel.packet[5].x = a_p_5;
+  kernel.packet[6].x = a_p_6;
+  kernel.packet[7].x = a_p_7;
+  kernel.packet[8].x = a_p_8;
+  kernel.packet[9].x = a_p_9;
+  kernel.packet[10].x = a_p_a;
+  kernel.packet[11].x = a_p_b;
+  kernel.packet[12].x = a_p_c;
+  kernel.packet[13].x = a_p_d;
+  kernel.packet[14].x = a_p_e;
+  kernel.packet[15].x = a_p_f;
+}
+
+EIGEN_STRONG_INLINE void
+ptranspose(PacketBlock<Packet16h,8>& kernel) {
+  EIGEN_ALIGN64 half in[8][16];
+  pstore<half>(in[0], kernel.packet[0]);
+  pstore<half>(in[1], kernel.packet[1]);
+  pstore<half>(in[2], kernel.packet[2]);
+  pstore<half>(in[3], kernel.packet[3]);
+  pstore<half>(in[4], kernel.packet[4]);
+  pstore<half>(in[5], kernel.packet[5]);
+  pstore<half>(in[6], kernel.packet[6]);
+  pstore<half>(in[7], kernel.packet[7]);
+
+  EIGEN_ALIGN64 half out[8][16];
+
+  for (int i = 0; i < 8; ++i) {
+    for (int j = 0; j < 8; ++j) {
+      out[i][j] = in[j][2*i];
+    }
+    for (int j = 0; j < 8; ++j) {
+      out[i][j+8] = in[j][2*i+1];
+    }
+  }
+
+  kernel.packet[0] = pload<Packet16h>(out[0]);
+  kernel.packet[1] = pload<Packet16h>(out[1]);
+  kernel.packet[2] = pload<Packet16h>(out[2]);
+  kernel.packet[3] = pload<Packet16h>(out[3]);
+  kernel.packet[4] = pload<Packet16h>(out[4]);
+  kernel.packet[5] = pload<Packet16h>(out[5]);
+  kernel.packet[6] = pload<Packet16h>(out[6]);
+  kernel.packet[7] = pload<Packet16h>(out[7]);
+}
+
+EIGEN_STRONG_INLINE void
+ptranspose(PacketBlock<Packet16h,4>& kernel) {
+  EIGEN_ALIGN64 half in[4][16];
+  pstore<half>(in[0], kernel.packet[0]);
+  pstore<half>(in[1], kernel.packet[1]);
+  pstore<half>(in[2], kernel.packet[2]);
+  pstore<half>(in[3], kernel.packet[3]);
+
+  EIGEN_ALIGN64 half out[4][16];
+
+  for (int i = 0; i < 4; ++i) {
+    for (int j = 0; j < 4; ++j) {
+      out[i][j] = in[j][4*i];
+    }
+    for (int j = 0; j < 4; ++j) {
+      out[i][j+4] = in[j][4*i+1];
+    }
+    for (int j = 0; j < 4; ++j) {
+      out[i][j+8] = in[j][4*i+2];
+    }
+    for (int j = 0; j < 4; ++j) {
+      out[i][j+12] = in[j][4*i+3];
+    }
+  }
+
+  kernel.packet[0] = pload<Packet16h>(out[0]);
+  kernel.packet[1] = pload<Packet16h>(out[1]);
+  kernel.packet[2] = pload<Packet16h>(out[2]);
+  kernel.packet[3] = pload<Packet16h>(out[3]);
+}
+
+
+#elif defined EIGEN_VECTORIZE_AVX
+
+typedef struct {
+  __m128i x;
+} Packet8h;
+
+
+template<> struct is_arithmetic<Packet8h> { enum { value = true }; };
+
+template <>
+struct packet_traits<Eigen::half> : default_packet_traits {
+  typedef Packet8h type;
+  // There is no half-size packet for Packet8h.
+  typedef Packet8h half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 8,
+    HasHalfPacket = 0,
+    HasAdd    = 0,
+    HasSub    = 0,
+    HasMul    = 0,
+    HasNegate = 0,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasConj   = 0,
+    HasSetLinear = 0,
+    HasDiv = 0,
+    HasSqrt = 0,
+    HasRsqrt = 0,
+    HasExp = 0,
+    HasLog = 0,
+    HasBlend = 0
+  };
+};
+
+
+template<> struct unpacket_traits<Packet8h> { typedef Eigen::half type; enum {size=8, alignment=Aligned16}; typedef Packet8h half; };
+
+template<> EIGEN_STRONG_INLINE Packet8h pset1<Packet8h>(const Eigen::half& from) {
+  Packet8h result;
+  result.x = _mm_set1_epi16(from.x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Eigen::half pfirst<Packet8h>(const Packet8h& from) {
+  return half_impl::raw_uint16_to_half(static_cast<unsigned short>(_mm_extract_epi16(from.x, 0)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet8h pload<Packet8h>(const Eigen::half* from) {
+  Packet8h result;
+  result.x = _mm_load_si128(reinterpret_cast<const __m128i*>(from));
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Packet8h ploadu<Packet8h>(const Eigen::half* from) {
+  Packet8h result;
+  result.x = _mm_loadu_si128(reinterpret_cast<const __m128i*>(from));
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE void pstore<Eigen::half>(Eigen::half* to, const Packet8h& from) {
+  _mm_store_si128(reinterpret_cast<__m128i*>(to), from.x);
+}
+
+template<> EIGEN_STRONG_INLINE void pstoreu<Eigen::half>(Eigen::half* to, const Packet8h& from) {
+  _mm_storeu_si128(reinterpret_cast<__m128i*>(to), from.x);
+}
+
+template<> EIGEN_STRONG_INLINE Packet8h
+ploadquad<Packet8h>(const Eigen::half* from) {
+  Packet8h result;
+  unsigned short a = from[0].x;
+  unsigned short b = from[1].x;
+  result.x = _mm_set_epi16(b, b, b, b, a, a, a, a);
+  return result;
+}
+
+EIGEN_STRONG_INLINE Packet8f half2float(const Packet8h& a) {
+#ifdef EIGEN_HAS_FP16_C
+  return _mm256_cvtph_ps(a.x);
+#else
+  EIGEN_ALIGN32 Eigen::half aux[8];
+  pstore(aux, a);
+  float f0(aux[0]);
+  float f1(aux[1]);
+  float f2(aux[2]);
+  float f3(aux[3]);
+  float f4(aux[4]);
+  float f5(aux[5]);
+  float f6(aux[6]);
+  float f7(aux[7]);
+
+  return _mm256_set_ps(f7, f6, f5, f4, f3, f2, f1, f0);
+#endif
+}
+
+EIGEN_STRONG_INLINE Packet8h float2half(const Packet8f& a) {
+#ifdef EIGEN_HAS_FP16_C
+  Packet8h result;
+  result.x = _mm256_cvtps_ph(a, _MM_FROUND_TO_NEAREST_INT|_MM_FROUND_NO_EXC);
+  return result;
+#else
+  EIGEN_ALIGN32 float aux[8];
+  pstore(aux, a);
+  Eigen::half h0(aux[0]);
+  Eigen::half h1(aux[1]);
+  Eigen::half h2(aux[2]);
+  Eigen::half h3(aux[3]);
+  Eigen::half h4(aux[4]);
+  Eigen::half h5(aux[5]);
+  Eigen::half h6(aux[6]);
+  Eigen::half h7(aux[7]);
+
+  Packet8h result;
+  result.x = _mm_set_epi16(h7.x, h6.x, h5.x, h4.x, h3.x, h2.x, h1.x, h0.x);
+  return result;
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet8h pconj(const Packet8h& a) { return a; }
+
+template<> EIGEN_STRONG_INLINE Packet8h padd<Packet8h>(const Packet8h& a, const Packet8h& b) {
+  Packet8f af = half2float(a);
+  Packet8f bf = half2float(b);
+  Packet8f rf = padd(af, bf);
+  return float2half(rf);
+}
+
+template<> EIGEN_STRONG_INLINE Packet8h pmul<Packet8h>(const Packet8h& a, const Packet8h& b) {
+  Packet8f af = half2float(a);
+  Packet8f bf = half2float(b);
+  Packet8f rf = pmul(af, bf);
+  return float2half(rf);
+}
+
+template<> EIGEN_STRONG_INLINE Packet8h pgather<Eigen::half, Packet8h>(const Eigen::half* from, Index stride)
+{
+  Packet8h result;
+  result.x = _mm_set_epi16(from[7*stride].x, from[6*stride].x, from[5*stride].x, from[4*stride].x, from[3*stride].x, from[2*stride].x, from[1*stride].x, from[0*stride].x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE void pscatter<Eigen::half, Packet8h>(Eigen::half* to, const Packet8h& from, Index stride)
+{
+  EIGEN_ALIGN32 Eigen::half aux[8];
+  pstore(aux, from);
+  to[stride*0].x = aux[0].x;
+  to[stride*1].x = aux[1].x;
+  to[stride*2].x = aux[2].x;
+  to[stride*3].x = aux[3].x;
+  to[stride*4].x = aux[4].x;
+  to[stride*5].x = aux[5].x;
+  to[stride*6].x = aux[6].x;
+  to[stride*7].x = aux[7].x;
+}
+
+template<> EIGEN_STRONG_INLINE Eigen::half predux<Packet8h>(const Packet8h& a) {
+  Packet8f af = half2float(a);
+  float reduced = predux<Packet8f>(af);
+  return Eigen::half(reduced);
+}
+
+template<> EIGEN_STRONG_INLINE Eigen::half predux_max<Packet8h>(const Packet8h& a) {
+  Packet8f af = half2float(a);
+  float reduced = predux_max<Packet8f>(af);
+  return Eigen::half(reduced);
+}
+
+template<> EIGEN_STRONG_INLINE Eigen::half predux_min<Packet8h>(const Packet8h& a) {
+  Packet8f af = half2float(a);
+  float reduced = predux_min<Packet8f>(af);
+  return Eigen::half(reduced);
+}
+
+template<> EIGEN_STRONG_INLINE Eigen::half predux_mul<Packet8h>(const Packet8h& a) {
+  Packet8f af = half2float(a);
+  float reduced = predux_mul<Packet8f>(af);
+  return Eigen::half(reduced);
+}
+
+EIGEN_STRONG_INLINE void
+ptranspose(PacketBlock<Packet8h,8>& kernel) {
+  __m128i a = kernel.packet[0].x;
+  __m128i b = kernel.packet[1].x;
+  __m128i c = kernel.packet[2].x;
+  __m128i d = kernel.packet[3].x;
+  __m128i e = kernel.packet[4].x;
+  __m128i f = kernel.packet[5].x;
+  __m128i g = kernel.packet[6].x;
+  __m128i h = kernel.packet[7].x;
+
+  __m128i a03b03 = _mm_unpacklo_epi16(a, b);
+  __m128i c03d03 = _mm_unpacklo_epi16(c, d);
+  __m128i e03f03 = _mm_unpacklo_epi16(e, f);
+  __m128i g03h03 = _mm_unpacklo_epi16(g, h);
+  __m128i a47b47 = _mm_unpackhi_epi16(a, b);
+  __m128i c47d47 = _mm_unpackhi_epi16(c, d);
+  __m128i e47f47 = _mm_unpackhi_epi16(e, f);
+  __m128i g47h47 = _mm_unpackhi_epi16(g, h);
+
+  __m128i a01b01c01d01 = _mm_unpacklo_epi32(a03b03, c03d03);
+  __m128i a23b23c23d23 = _mm_unpackhi_epi32(a03b03, c03d03);
+  __m128i e01f01g01h01 = _mm_unpacklo_epi32(e03f03, g03h03);
+  __m128i e23f23g23h23 = _mm_unpackhi_epi32(e03f03, g03h03);
+  __m128i a45b45c45d45 = _mm_unpacklo_epi32(a47b47, c47d47);
+  __m128i a67b67c67d67 = _mm_unpackhi_epi32(a47b47, c47d47);
+  __m128i e45f45g45h45 = _mm_unpacklo_epi32(e47f47, g47h47);
+  __m128i e67f67g67h67 = _mm_unpackhi_epi32(e47f47, g47h47);
+
+  __m128i a0b0c0d0e0f0g0h0 = _mm_unpacklo_epi64(a01b01c01d01, e01f01g01h01);
+  __m128i a1b1c1d1e1f1g1h1 = _mm_unpackhi_epi64(a01b01c01d01, e01f01g01h01);
+  __m128i a2b2c2d2e2f2g2h2 = _mm_unpacklo_epi64(a23b23c23d23, e23f23g23h23);
+  __m128i a3b3c3d3e3f3g3h3 = _mm_unpackhi_epi64(a23b23c23d23, e23f23g23h23);
+  __m128i a4b4c4d4e4f4g4h4 = _mm_unpacklo_epi64(a45b45c45d45, e45f45g45h45);
+  __m128i a5b5c5d5e5f5g5h5 = _mm_unpackhi_epi64(a45b45c45d45, e45f45g45h45);
+  __m128i a6b6c6d6e6f6g6h6 = _mm_unpacklo_epi64(a67b67c67d67, e67f67g67h67);
+  __m128i a7b7c7d7e7f7g7h7 = _mm_unpackhi_epi64(a67b67c67d67, e67f67g67h67);
+
+  kernel.packet[0].x = a0b0c0d0e0f0g0h0;
+  kernel.packet[1].x = a1b1c1d1e1f1g1h1;
+  kernel.packet[2].x = a2b2c2d2e2f2g2h2;
+  kernel.packet[3].x = a3b3c3d3e3f3g3h3;
+  kernel.packet[4].x = a4b4c4d4e4f4g4h4;
+  kernel.packet[5].x = a5b5c5d5e5f5g5h5;
+  kernel.packet[6].x = a6b6c6d6e6f6g6h6;
+  kernel.packet[7].x = a7b7c7d7e7f7g7h7;
+}
+
+EIGEN_STRONG_INLINE void
+ptranspose(PacketBlock<Packet8h,4>& kernel) {
+  EIGEN_ALIGN32 Eigen::half in[4][8];
+  pstore<Eigen::half>(in[0], kernel.packet[0]);
+  pstore<Eigen::half>(in[1], kernel.packet[1]);
+  pstore<Eigen::half>(in[2], kernel.packet[2]);
+  pstore<Eigen::half>(in[3], kernel.packet[3]);
+
+  EIGEN_ALIGN32 Eigen::half out[4][8];
+
+  for (int i = 0; i < 4; ++i) {
+    for (int j = 0; j < 4; ++j) {
+      out[i][j] = in[j][2*i];
+    }
+    for (int j = 0; j < 4; ++j) {
+      out[i][j+4] = in[j][2*i+1];
+    }
+  }
+
+  kernel.packet[0] = pload<Packet8h>(out[0]);
+  kernel.packet[1] = pload<Packet8h>(out[1]);
+  kernel.packet[2] = pload<Packet8h>(out[2]);
+  kernel.packet[3] = pload<Packet8h>(out[3]);
+}
+
+
+// Disable the following code since it's broken on too many platforms / compilers.
+//#elif defined(EIGEN_VECTORIZE_SSE) && (!EIGEN_ARCH_x86_64) && (!EIGEN_COMP_MSVC)
+#elif 0
+
+typedef struct {
+  __m64 x;
+} Packet4h;
+
+
+template<> struct is_arithmetic<Packet4h> { enum { value = true }; };
+
+template <>
+struct packet_traits<Eigen::half> : default_packet_traits {
+  typedef Packet4h type;
+  // There is no half-size packet for Packet4h.
+  typedef Packet4h half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 4,
+    HasHalfPacket = 0,
+    HasAdd    = 0,
+    HasSub    = 0,
+    HasMul    = 0,
+    HasNegate = 0,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasConj   = 0,
+    HasSetLinear = 0,
+    HasDiv = 0,
+    HasSqrt = 0,
+    HasRsqrt = 0,
+    HasExp = 0,
+    HasLog = 0,
+    HasBlend = 0
+  };
+};
+
+
+template<> struct unpacket_traits<Packet4h> { typedef Eigen::half type; enum {size=4, alignment=Aligned16}; typedef Packet4h half; };
+
+template<> EIGEN_STRONG_INLINE Packet4h pset1<Packet4h>(const Eigen::half& from) {
+  Packet4h result;
+  result.x = _mm_set1_pi16(from.x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Eigen::half pfirst<Packet4h>(const Packet4h& from) {
+  return half_impl::raw_uint16_to_half(static_cast<unsigned short>(_mm_cvtsi64_si32(from.x)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4h pconj(const Packet4h& a) { return a; }
+
+template<> EIGEN_STRONG_INLINE Packet4h padd<Packet4h>(const Packet4h& a, const Packet4h& b) {
+  __int64_t a64 = _mm_cvtm64_si64(a.x);
+  __int64_t b64 = _mm_cvtm64_si64(b.x);
+
+  Eigen::half h[4];
+
+  Eigen::half ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64));
+  Eigen::half hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64));
+  h[0] = ha + hb;
+  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 16));
+  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 16));
+  h[1] = ha + hb;
+  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 32));
+  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 32));
+  h[2] = ha + hb;
+  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 48));
+  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 48));
+  h[3] = ha + hb;
+  Packet4h result;
+  result.x = _mm_set_pi16(h[3].x, h[2].x, h[1].x, h[0].x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4h pmul<Packet4h>(const Packet4h& a, const Packet4h& b) {
+  __int64_t a64 = _mm_cvtm64_si64(a.x);
+  __int64_t b64 = _mm_cvtm64_si64(b.x);
+
+  Eigen::half h[4];
+
+  Eigen::half ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64));
+  Eigen::half hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64));
+  h[0] = ha * hb;
+  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 16));
+  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 16));
+  h[1] = ha * hb;
+  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 32));
+  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 32));
+  h[2] = ha * hb;
+  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 48));
+  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 48));
+  h[3] = ha * hb;
+  Packet4h result;
+  result.x = _mm_set_pi16(h[3].x, h[2].x, h[1].x, h[0].x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4h pload<Packet4h>(const Eigen::half* from) {
+  Packet4h result;
+  result.x = _mm_cvtsi64_m64(*reinterpret_cast<const __int64_t*>(from));
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4h ploadu<Packet4h>(const Eigen::half* from) {
+  Packet4h result;
+  result.x = _mm_cvtsi64_m64(*reinterpret_cast<const __int64_t*>(from));
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE void pstore<Eigen::half>(Eigen::half* to, const Packet4h& from) {
+  __int64_t r = _mm_cvtm64_si64(from.x);
+  *(reinterpret_cast<__int64_t*>(to)) = r;
+}
+
+template<> EIGEN_STRONG_INLINE void pstoreu<Eigen::half>(Eigen::half* to, const Packet4h& from) {
+  __int64_t r = _mm_cvtm64_si64(from.x);
+  *(reinterpret_cast<__int64_t*>(to)) = r;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4h
+ploadquad<Packet4h>(const Eigen::half* from) {
+  return pset1<Packet4h>(*from);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4h pgather<Eigen::half, Packet4h>(const Eigen::half* from, Index stride)
+{
+  Packet4h result;
+  result.x = _mm_set_pi16(from[3*stride].x, from[2*stride].x, from[1*stride].x, from[0*stride].x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE void pscatter<Eigen::half, Packet4h>(Eigen::half* to, const Packet4h& from, Index stride)
+{
+  __int64_t a = _mm_cvtm64_si64(from.x);
+  to[stride*0].x = static_cast<unsigned short>(a);
+  to[stride*1].x = static_cast<unsigned short>(a >> 16);
+  to[stride*2].x = static_cast<unsigned short>(a >> 32);
+  to[stride*3].x = static_cast<unsigned short>(a >> 48);
+}
+
+EIGEN_STRONG_INLINE void
+ptranspose(PacketBlock<Packet4h,4>& kernel) {
+  __m64 T0 = _mm_unpacklo_pi16(kernel.packet[0].x, kernel.packet[1].x);
+  __m64 T1 = _mm_unpacklo_pi16(kernel.packet[2].x, kernel.packet[3].x);
+  __m64 T2 = _mm_unpackhi_pi16(kernel.packet[0].x, kernel.packet[1].x);
+  __m64 T3 = _mm_unpackhi_pi16(kernel.packet[2].x, kernel.packet[3].x);
+
+  kernel.packet[0].x = _mm_unpacklo_pi32(T0, T1);
+  kernel.packet[1].x = _mm_unpackhi_pi32(T0, T1);
+  kernel.packet[2].x = _mm_unpacklo_pi32(T2, T3);
+  kernel.packet[3].x = _mm_unpackhi_pi32(T2, T3);
+}
+
+#endif
+
+}
+}
+
+#endif // EIGEN_PACKET_MATH_HALF_CUDA_H
diff --git a/eigen/Eigen/src/Core/arch/CUDA/TypeCasting.h b/eigen/Eigen/src/Core/arch/CUDA/TypeCasting.h
new file mode 100644
index 0000000..aa5fbce
--- /dev/null
+++ b/eigen/Eigen/src/Core/arch/CUDA/TypeCasting.h
@@ -0,0 +1,212 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_TYPE_CASTING_CUDA_H
+#define EIGEN_TYPE_CASTING_CUDA_H
+
+namespace Eigen {
+
+namespace internal {
+
+template<>
+struct scalar_cast_op<float, Eigen::half> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
+  typedef Eigen::half result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half operator() (const float& a) const {
+    #if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+      return __float2half(a);
+    #else
+      return Eigen::half(a);
+    #endif
+  }
+};
+
+template<>
+struct functor_traits<scalar_cast_op<float, Eigen::half> >
+{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; };
+
+
+template<>
+struct scalar_cast_op<int, Eigen::half> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
+  typedef Eigen::half result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half operator() (const int& a) const {
+    #if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+      return __float2half(static_cast<float>(a));
+    #else
+      return Eigen::half(static_cast<float>(a));
+    #endif
+  }
+};
+
+template<>
+struct functor_traits<scalar_cast_op<int, Eigen::half> >
+{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; };
+
+
+template<>
+struct scalar_cast_op<Eigen::half, float> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
+  typedef float result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float operator() (const Eigen::half& a) const {
+    #if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+      return __half2float(a);
+    #else
+      return static_cast<float>(a);
+    #endif
+  }
+};
+
+template<>
+struct functor_traits<scalar_cast_op<Eigen::half, float> >
+{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; };
+
+
+
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+
+template <>
+struct type_casting_traits<Eigen::half, float> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 2,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pcast<half2, float4>(const half2& a, const half2& b) {
+  float2 r1 = __half22float2(a);
+  float2 r2 = __half22float2(b);
+  return make_float4(r1.x, r1.y, r2.x, r2.y);
+}
+
+template <>
+struct type_casting_traits<float, Eigen::half> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 2
+  };
+};
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pcast<float4, half2>(const float4& a) {
+  // Simply discard the second half of the input
+  return __floats2half2_rn(a.x, a.y);
+}
+
+#elif defined EIGEN_VECTORIZE_AVX512
+template <>
+struct type_casting_traits<half, float> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet16f pcast<Packet16h, Packet16f>(const Packet16h& a) {
+  return half2float(a);
+}
+
+template <>
+struct type_casting_traits<float, half> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet16h pcast<Packet16f, Packet16h>(const Packet16f& a) {
+  return float2half(a);
+}
+
+#elif defined EIGEN_VECTORIZE_AVX
+
+template <>
+struct type_casting_traits<Eigen::half, float> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet8f pcast<Packet8h, Packet8f>(const Packet8h& a) {
+  return half2float(a);
+}
+
+template <>
+struct type_casting_traits<float, Eigen::half> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet8h pcast<Packet8f, Packet8h>(const Packet8f& a) {
+  return float2half(a);
+}
+
+// Disable the following code since it's broken on too many platforms / compilers.
+//#elif defined(EIGEN_VECTORIZE_SSE) && (!EIGEN_ARCH_x86_64) && (!EIGEN_COMP_MSVC)
+#elif 0
+
+template <>
+struct type_casting_traits<Eigen::half, float> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet4f pcast<Packet4h, Packet4f>(const Packet4h& a) {
+  __int64_t a64 = _mm_cvtm64_si64(a.x);
+  Eigen::half h = raw_uint16_to_half(static_cast<unsigned short>(a64));
+  float f1 = static_cast<float>(h);
+  h = raw_uint16_to_half(static_cast<unsigned short>(a64 >> 16));
+  float f2 = static_cast<float>(h);
+  h = raw_uint16_to_half(static_cast<unsigned short>(a64 >> 32));
+  float f3 = static_cast<float>(h);
+  h = raw_uint16_to_half(static_cast<unsigned short>(a64 >> 48));
+  float f4 = static_cast<float>(h);
+  return _mm_set_ps(f4, f3, f2, f1);
+}
+
+template <>
+struct type_casting_traits<float, Eigen::half> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet4h pcast<Packet4f, Packet4h>(const Packet4f& a) {
+  EIGEN_ALIGN16 float aux[4];
+  pstore(aux, a);
+  Eigen::half h0(aux[0]);
+  Eigen::half h1(aux[1]);
+  Eigen::half h2(aux[2]);
+  Eigen::half h3(aux[3]);
+
+  Packet4h result;
+  result.x = _mm_set_pi16(h3.x, h2.x, h1.x, h0.x);
+  return result;
+}
+
+#endif
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TYPE_CASTING_CUDA_H
diff --git a/eigen/Eigen/src/Core/arch/Default/CMakeLists.txt b/eigen/Eigen/src/Core/arch/Default/CMakeLists.txt
deleted file mode 100644
index 339c091..0000000
--- a/eigen/Eigen/src/Core/arch/Default/CMakeLists.txt
+++ /dev/null
@@ -1,6 +0,0 @@
-FILE(GLOB Eigen_Core_arch_Default_SRCS "*.h")
-
-INSTALL(FILES
-  ${Eigen_Core_arch_Default_SRCS}
-  DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/arch/Default COMPONENT Devel
-)
diff --git a/eigen/Eigen/src/Core/arch/NEON/CMakeLists.txt b/eigen/Eigen/src/Core/arch/NEON/CMakeLists.txt
deleted file mode 100644
index fd4d4af..0000000
--- a/eigen/Eigen/src/Core/arch/NEON/CMakeLists.txt
+++ /dev/null
@@ -1,6 +0,0 @@
-FILE(GLOB Eigen_Core_arch_NEON_SRCS "*.h")
-
-INSTALL(FILES
-  ${Eigen_Core_arch_NEON_SRCS}
-  DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/arch/NEON COMPONENT Devel
-)
diff --git a/eigen/Eigen/src/Core/arch/NEON/Complex.h b/eigen/Eigen/src/Core/arch/NEON/Complex.h
index 8d9255e..57e9b43 100644
--- a/eigen/Eigen/src/Core/arch/NEON/Complex.h
+++ b/eigen/Eigen/src/Core/arch/NEON/Complex.h
@@ -2,6 +2,7 @@
 // for linear algebra.
 //
 // Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2010 Konstantinos Margaritis <markos@freevec.org>
 //
 // This Source Code Form is subject to the terms of the Mozilla
 // Public License v. 2.0. If a copy of the MPL was not distributed
@@ -14,8 +15,21 @@ namespace Eigen {
 
 namespace internal {
 
-static uint32x4_t p4ui_CONJ_XOR = EIGEN_INIT_NEON_PACKET4(0x00000000, 0x80000000, 0x00000000, 0x80000000);
-static uint32x2_t p2ui_CONJ_XOR = EIGEN_INIT_NEON_PACKET2(0x00000000, 0x80000000);
+inline uint32x4_t p4ui_CONJ_XOR() {
+// See bug 1325, clang fails to call vld1q_u64.
+#if EIGEN_COMP_CLANG
+  uint32x4_t ret = { 0x00000000, 0x80000000, 0x00000000, 0x80000000 };
+  return ret;
+#else
+  static const uint32_t conj_XOR_DATA[] = { 0x00000000, 0x80000000, 0x00000000, 0x80000000 };
+  return vld1q_u32( conj_XOR_DATA );
+#endif
+}
+
+inline uint32x2_t p2ui_CONJ_XOR() {
+  static const uint32_t conj_XOR_DATA[] = { 0x00000000, 0x80000000 };
+  return vld1_u32( conj_XOR_DATA );
+}
 
 //---------- float ----------
 struct Packet2cf
@@ -28,10 +42,12 @@ struct Packet2cf
 template<> struct packet_traits<std::complex<float> >  : default_packet_traits
 {
   typedef Packet2cf type;
+  typedef Packet2cf half;
   enum {
     Vectorizable = 1,
     AlignedOnScalar = 1,
     size = 2,
+    HasHalfPacket = 0,
 
     HasAdd    = 1,
     HasSub    = 1,
@@ -46,7 +62,7 @@ template<> struct packet_traits<std::complex<float> >  : default_packet_traits
   };
 };
 
-template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2}; };
+template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16}; typedef Packet2cf half; };
 
 template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>&  from)
 {
@@ -62,7 +78,7 @@ template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) { return Pa
 template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a)
 {
   Packet4ui b = vreinterpretq_u32_f32(a.v);
-  return Packet2cf(vreinterpretq_f32_u32(veorq_u32(b, p4ui_CONJ_XOR)));
+  return Packet2cf(vreinterpretq_f32_u32(veorq_u32(b, p4ui_CONJ_XOR())));
 }
 
 template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
@@ -71,14 +87,14 @@ template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, con
 
   // Get the real values of a | a1_re | a1_re | a2_re | a2_re |
   v1 = vcombine_f32(vdup_lane_f32(vget_low_f32(a.v), 0), vdup_lane_f32(vget_high_f32(a.v), 0));
-  // Get the real values of a | a1_im | a1_im | a2_im | a2_im |
+  // Get the imag values of a | a1_im | a1_im | a2_im | a2_im |
   v2 = vcombine_f32(vdup_lane_f32(vget_low_f32(a.v), 1), vdup_lane_f32(vget_high_f32(a.v), 1));
   // Multiply the real a with b
   v1 = vmulq_f32(v1, b.v);
   // Multiply the imag a with b
   v2 = vmulq_f32(v2, b.v);
   // Conjugate v2 
-  v2 = vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(v2), p4ui_CONJ_XOR));
+  v2 = vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(v2), p4ui_CONJ_XOR()));
   // Swap real/imag elements in v2.
   v2 = vrev64q_f32(v2);
   // Add and return the result
@@ -87,7 +103,7 @@ template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, con
 
 template<> EIGEN_STRONG_INLINE Packet2cf pand   <Packet2cf>(const Packet2cf& a, const Packet2cf& b)
 {
-  return Packet2cf(vreinterpretq_f32_u32(vorrq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v))));
+  return Packet2cf(vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v))));
 }
 template<> EIGEN_STRONG_INLINE Packet2cf por    <Packet2cf>(const Packet2cf& a, const Packet2cf& b)
 {
@@ -110,6 +126,22 @@ template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<
 template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> *   to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((float*)to, from.v); }
 template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> *   to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((float*)to, from.v); }
 
+template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from, Index stride)
+{
+  Packet4f res = pset1<Packet4f>(0.f);
+  res = vsetq_lane_f32(std::real(from[0*stride]), res, 0);
+  res = vsetq_lane_f32(std::imag(from[0*stride]), res, 1);
+  res = vsetq_lane_f32(std::real(from[1*stride]), res, 2);
+  res = vsetq_lane_f32(std::imag(from[1*stride]), res, 3);
+  return Packet2cf(res);
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, Index stride)
+{
+  to[stride*0] = std::complex<float>(vgetq_lane_f32(from.v, 0), vgetq_lane_f32(from.v, 1));
+  to[stride*1] = std::complex<float>(vgetq_lane_f32(from.v, 2), vgetq_lane_f32(from.v, 3));
+}
+
 template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> *   addr) { EIGEN_ARM_PREFETCH((float *)addr); }
 
 template<> EIGEN_STRONG_INLINE std::complex<float>  pfirst<Packet2cf>(const Packet2cf& a)
@@ -177,7 +209,7 @@ template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const P
   // Multiply the imag a with b
   v2 = vmul_f32(v2, a2);
   // Conjugate v2 
-  v2 = vreinterpret_f32_u32(veor_u32(vreinterpret_u32_f32(v2), p2ui_CONJ_XOR));
+  v2 = vreinterpret_f32_u32(veor_u32(vreinterpret_u32_f32(v2), p2ui_CONJ_XOR()));
   // Swap real/imag elements in v2.
   v2 = vrev64_f32(v2);
   // Add v1, v2
@@ -235,7 +267,7 @@ template<> struct conj_helper<Packet2cf, Packet2cf, true,true>
 
 template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
 {
-  // TODO optimize it for AltiVec
+  // TODO optimize it for NEON
   Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a,b);
   Packet4f s, rev_s;
 
@@ -246,6 +278,207 @@ template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, con
   return Packet2cf(pdiv(res.v, vaddq_f32(s,rev_s)));
 }
 
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet2cf,2>& kernel) {
+  Packet4f tmp = vcombine_f32(vget_high_f32(kernel.packet[0].v), vget_high_f32(kernel.packet[1].v));
+  kernel.packet[0].v = vcombine_f32(vget_low_f32(kernel.packet[0].v), vget_low_f32(kernel.packet[1].v));
+  kernel.packet[1].v = tmp;
+}
+
+//---------- double ----------
+#if EIGEN_ARCH_ARM64 && !EIGEN_APPLE_DOUBLE_NEON_BUG
+
+// See bug 1325, clang fails to call vld1q_u64.
+#if EIGEN_COMP_CLANG
+  static uint64x2_t p2ul_CONJ_XOR = {0x0, 0x8000000000000000};
+#else
+  const uint64_t  p2ul_conj_XOR_DATA[] = { 0x0, 0x8000000000000000 };
+  static uint64x2_t p2ul_CONJ_XOR = vld1q_u64( p2ul_conj_XOR_DATA );
+#endif
+
+struct Packet1cd
+{
+  EIGEN_STRONG_INLINE Packet1cd() {}
+  EIGEN_STRONG_INLINE explicit Packet1cd(const Packet2d& a) : v(a) {}
+  Packet2d v;
+};
+
+template<> struct packet_traits<std::complex<double> >  : default_packet_traits
+{
+  typedef Packet1cd type;
+  typedef Packet1cd half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 0,
+    size = 1,
+    HasHalfPacket = 0,
+
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasDiv    = 1,
+    HasNegate = 1,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasSetLinear = 0
+  };
+};
+
+template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16}; typedef Packet1cd half; };
+
+template<> EIGEN_STRONG_INLINE Packet1cd pload<Packet1cd>(const std::complex<double>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(pload<Packet2d>((const double*)from)); }
+template<> EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu<Packet2d>((const double*)from)); }
+
+template<> EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<double>&  from)
+{ /* here we really have to use unaligned loads :( */ return ploadu<Packet1cd>(&from); }
+
+template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(padd<Packet2d>(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(psub<Packet2d>(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate<Packet2d>(a.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) { return Packet1cd(vreinterpretq_f64_u64(veorq_u64(vreinterpretq_u64_f64(a.v), p2ul_CONJ_XOR))); }
+
+template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
+{
+  Packet2d v1, v2;
+
+  // Get the real values of a 
+  v1 = vdupq_lane_f64(vget_low_f64(a.v), 0);
+  // Get the imag values of a
+  v2 = vdupq_lane_f64(vget_high_f64(a.v), 0);
+  // Multiply the real a with b
+  v1 = vmulq_f64(v1, b.v);
+  // Multiply the imag a with b
+  v2 = vmulq_f64(v2, b.v);
+  // Conjugate v2 
+  v2 = vreinterpretq_f64_u64(veorq_u64(vreinterpretq_u64_f64(v2), p2ul_CONJ_XOR));
+  // Swap real/imag elements in v2.
+  v2 = preverse<Packet2d>(v2);
+  // Add and return the result
+  return Packet1cd(vaddq_f64(v1, v2));
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd pand   <Packet1cd>(const Packet1cd& a, const Packet1cd& b)
+{
+  return Packet1cd(vreinterpretq_f64_u64(vandq_u64(vreinterpretq_u64_f64(a.v),vreinterpretq_u64_f64(b.v))));
+}
+template<> EIGEN_STRONG_INLINE Packet1cd por    <Packet1cd>(const Packet1cd& a, const Packet1cd& b)
+{
+  return Packet1cd(vreinterpretq_f64_u64(vorrq_u64(vreinterpretq_u64_f64(a.v),vreinterpretq_u64_f64(b.v))));
+}
+template<> EIGEN_STRONG_INLINE Packet1cd pxor   <Packet1cd>(const Packet1cd& a, const Packet1cd& b)
+{
+  return Packet1cd(vreinterpretq_f64_u64(veorq_u64(vreinterpretq_u64_f64(a.v),vreinterpretq_u64_f64(b.v))));
+}
+template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
+{
+  return Packet1cd(vreinterpretq_f64_u64(vbicq_u64(vreinterpretq_u64_f64(a.v),vreinterpretq_u64_f64(b.v))));
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>* from) { return pset1<Packet1cd>(*from); }
+
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); }
+
+template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> *   addr) { EIGEN_ARM_PREFETCH((double *)addr); }
+
+template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather<std::complex<double>, Packet1cd>(const std::complex<double>* from, Index stride)
+{
+  Packet2d res = pset1<Packet2d>(0.0);
+  res = vsetq_lane_f64(std::real(from[0*stride]), res, 0);
+  res = vsetq_lane_f64(std::imag(from[0*stride]), res, 1);
+  return Packet1cd(res);
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet1cd>(std::complex<double>* to, const Packet1cd& from, Index stride)
+{
+  to[stride*0] = std::complex<double>(vgetq_lane_f64(from.v, 0), vgetq_lane_f64(from.v, 1));
+}
+
+
+template<> EIGEN_STRONG_INLINE std::complex<double>  pfirst<Packet1cd>(const Packet1cd& a)
+{
+  std::complex<double> EIGEN_ALIGN16 res;
+  pstore<std::complex<double> >(&res, a);
+
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; }
+
+template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a) { return pfirst(a); }
+
+template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs) { return vecs[0]; }
+
+template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a) { return pfirst(a); }
+
+template<int Offset>
+struct palign_impl<Offset,Packet1cd>
+{
+  static EIGEN_STRONG_INLINE void run(Packet1cd& /*first*/, const Packet1cd& /*second*/)
+  {
+    // FIXME is it sure we never have to align a Packet1cd?
+    // Even though a std::complex<double> has 16 bytes, it is not necessarily aligned on a 16 bytes boundary...
+  }
+};
+
+template<> struct conj_helper<Packet1cd, Packet1cd, false,true>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
+  {
+    return internal::pmul(a, pconj(b));
+  }
+};
+
+template<> struct conj_helper<Packet1cd, Packet1cd, true,false>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
+  {
+    return internal::pmul(pconj(a), b);
+  }
+};
+
+template<> struct conj_helper<Packet1cd, Packet1cd, true,true>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
+  {
+    return pconj(internal::pmul(a, b));
+  }
+};
+
+template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
+{
+  // TODO optimize it for NEON
+  Packet1cd res = conj_helper<Packet1cd,Packet1cd,false,true>().pmul(a,b);
+  Packet2d s = pmul<Packet2d>(b.v, b.v);
+  Packet2d rev_s = preverse<Packet2d>(s);
+
+  return Packet1cd(pdiv(res.v, padd<Packet2d>(s,rev_s)));
+}
+
+EIGEN_STRONG_INLINE Packet1cd pcplxflip/*<Packet1cd>*/(const Packet1cd& x)
+{
+  return Packet1cd(preverse(Packet2d(x.v)));
+}
+
+EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet1cd,2>& kernel)
+{
+  Packet2d tmp = vcombine_f64(vget_high_f64(kernel.packet[0].v), vget_high_f64(kernel.packet[1].v));
+  kernel.packet[0].v = vcombine_f64(vget_low_f64(kernel.packet[0].v), vget_low_f64(kernel.packet[1].v));
+  kernel.packet[1].v = tmp;
+}
+#endif // EIGEN_ARCH_ARM64
+
 } // end namespace internal
 
 } // end namespace Eigen
diff --git a/eigen/Eigen/src/Core/arch/NEON/MathFunctions.h b/eigen/Eigen/src/Core/arch/NEON/MathFunctions.h
new file mode 100644
index 0000000..6bb05bb
--- /dev/null
+++ b/eigen/Eigen/src/Core/arch/NEON/MathFunctions.h
@@ -0,0 +1,91 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/* The sin, cos, exp, and log functions of this file come from
+ * Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/
+ */
+
+#ifndef EIGEN_MATH_FUNCTIONS_NEON_H
+#define EIGEN_MATH_FUNCTIONS_NEON_H
+
+namespace Eigen {
+
+namespace internal {
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f pexp<Packet4f>(const Packet4f& _x)
+{
+  Packet4f x = _x;
+  Packet4f tmp, fx;
+
+  _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f);
+  _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
+  _EIGEN_DECLARE_CONST_Packet4i(0x7f, 0x7f);
+  _EIGEN_DECLARE_CONST_Packet4f(exp_hi,  88.3762626647950f);
+  _EIGEN_DECLARE_CONST_Packet4f(exp_lo, -88.3762626647949f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_LOG2EF, 1.44269504088896341f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C1, 0.693359375f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C2, -2.12194440e-4f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p0, 1.9875691500E-4f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p1, 1.3981999507E-3f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p2, 8.3334519073E-3f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p3, 4.1665795894E-2f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201E-1f);
+
+  x = vminq_f32(x, p4f_exp_hi);
+  x = vmaxq_f32(x, p4f_exp_lo);
+
+  /* express exp(x) as exp(g + n*log(2)) */
+  fx = vmlaq_f32(p4f_half, x, p4f_cephes_LOG2EF);
+
+  /* perform a floorf */
+  tmp = vcvtq_f32_s32(vcvtq_s32_f32(fx));
+
+  /* if greater, substract 1 */
+  Packet4ui mask = vcgtq_f32(tmp, fx);
+  mask = vandq_u32(mask, vreinterpretq_u32_f32(p4f_1));
+
+  fx = vsubq_f32(tmp, vreinterpretq_f32_u32(mask));
+
+  tmp = vmulq_f32(fx, p4f_cephes_exp_C1);
+  Packet4f z = vmulq_f32(fx, p4f_cephes_exp_C2);
+  x = vsubq_f32(x, tmp);
+  x = vsubq_f32(x, z);
+
+  Packet4f y = vmulq_f32(p4f_cephes_exp_p0, x);
+  z = vmulq_f32(x, x);
+  y = vaddq_f32(y, p4f_cephes_exp_p1);
+  y = vmulq_f32(y, x);
+  y = vaddq_f32(y, p4f_cephes_exp_p2);
+  y = vmulq_f32(y, x);
+  y = vaddq_f32(y, p4f_cephes_exp_p3);
+  y = vmulq_f32(y, x);
+  y = vaddq_f32(y, p4f_cephes_exp_p4);
+  y = vmulq_f32(y, x);
+  y = vaddq_f32(y, p4f_cephes_exp_p5);
+
+  y = vmulq_f32(y, z);
+  y = vaddq_f32(y, x);
+  y = vaddq_f32(y, p4f_1);
+
+  /* build 2^n */
+  int32x4_t mm;
+  mm = vcvtq_s32_f32(fx);
+  mm = vaddq_s32(mm, p4i_0x7f);
+  mm = vshlq_n_s32(mm, 23);
+  Packet4f pow2n = vreinterpretq_f32_s32(mm);
+
+  y = vmulq_f32(y, pow2n);
+  return y;
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_MATH_FUNCTIONS_NEON_H
diff --git a/eigen/Eigen/src/Core/arch/NEON/PacketMath.h b/eigen/Eigen/src/Core/arch/NEON/PacketMath.h
index d49670e..aede4a6 100644
--- a/eigen/Eigen/src/Core/arch/NEON/PacketMath.h
+++ b/eigen/Eigen/src/Core/arch/NEON/PacketMath.h
@@ -2,7 +2,7 @@
 // for linear algebra.
 //
 // Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
-// Copyright (C) 2010 Konstantinos Margaritis <markos@codex.gr>
+// Copyright (C) 2010 Konstantinos Margaritis <markos@freevec.org>
 // Heavily based on Gael's SSE version.
 //
 // This Source Code Form is subject to the terms of the Mozilla
@@ -20,43 +20,48 @@ namespace internal {
 #define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8
 #endif
 
-// FIXME NEON has 16 quad registers, but since the current register allocator
-// is so bad, it is much better to reduce it to 8
+#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#endif
+
+#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD
+#define EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD
+#endif
+
 #ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
-#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 8
+#if EIGEN_ARCH_ARM64
+#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 32
+#else
+#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 16 
+#endif
 #endif
 
+typedef float32x2_t Packet2f;
 typedef float32x4_t Packet4f;
 typedef int32x4_t   Packet4i;
+typedef int32x2_t   Packet2i;
 typedef uint32x4_t  Packet4ui;
 
 #define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \
   const Packet4f p4f_##NAME = pset1<Packet4f>(X)
 
 #define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
-  const Packet4f p4f_##NAME = vreinterpretq_f32_u32(pset1<int>(X))
+  const Packet4f p4f_##NAME = vreinterpretq_f32_u32(pset1<int32_t>(X))
 
 #define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
   const Packet4i p4i_##NAME = pset1<Packet4i>(X)
 
-#if defined(__llvm__) && !defined(__clang__)
-  //Special treatment for Apple's llvm-gcc, its NEON packet types are unions
-  #define EIGEN_INIT_NEON_PACKET2(X, Y)       {{X, Y}}
-  #define EIGEN_INIT_NEON_PACKET4(X, Y, Z, W) {{X, Y, Z, W}}
-#else
-  //Default initializer for packets
-  #define EIGEN_INIT_NEON_PACKET2(X, Y)       {X, Y}
-  #define EIGEN_INIT_NEON_PACKET4(X, Y, Z, W) {X, Y, Z, W}
-#endif
-
-// arm64 does have the pld instruction. If available, let's trust the __builtin_prefetch built-in function
-// which available on LLVM and GCC (at least)
-#if EIGEN_HAS_BUILTIN(__builtin_prefetch) || defined(__GNUC__)
+#if EIGEN_ARCH_ARM64
+  // __builtin_prefetch tends to do nothing on ARM64 compilers because the
+  // prefetch instructions there are too detailed for __builtin_prefetch to map
+  // meaningfully to them.
+  #define EIGEN_ARM_PREFETCH(ADDR)  __asm__ __volatile__("prfm pldl1keep, [%[addr]]\n" ::[addr] "r"(ADDR) : );
+#elif EIGEN_HAS_BUILTIN(__builtin_prefetch) || EIGEN_COMP_GNUC
   #define EIGEN_ARM_PREFETCH(ADDR) __builtin_prefetch(ADDR);
 #elif defined __pld
   #define EIGEN_ARM_PREFETCH(ADDR) __pld(ADDR)
-#elif !defined(__aarch64__)
-  #define EIGEN_ARM_PREFETCH(ADDR) __asm__ __volatile__ ( "   pld [%[addr]]\n" :: [addr] "r" (ADDR) : "cc" );
+#elif EIGEN_ARCH_ARM32
+  #define EIGEN_ARM_PREFETCH(ADDR) __asm__ __volatile__ ("pld [%[addr]]\n" :: [addr] "r" (ADDR) : );
 #else
   // by default no explicit prefetching
   #define EIGEN_ARM_PREFETCH(ADDR)
@@ -65,53 +70,60 @@ typedef uint32x4_t  Packet4ui;
 template<> struct packet_traits<float>  : default_packet_traits
 {
   typedef Packet4f type;
+  typedef Packet4f half; // Packet2f intrinsics not implemented yet
   enum {
     Vectorizable = 1,
     AlignedOnScalar = 1,
     size = 4,
+    HasHalfPacket=0, // Packet2f intrinsics not implemented yet
    
     HasDiv  = 1,
     // FIXME check the Has*
     HasSin  = 0,
     HasCos  = 0,
     HasLog  = 0,
-    HasExp  = 0,
+    HasExp  = 1,
     HasSqrt = 0
   };
 };
-template<> struct packet_traits<int>    : default_packet_traits
+template<> struct packet_traits<int32_t>    : default_packet_traits
 {
   typedef Packet4i type;
+  typedef Packet4i half; // Packet2i intrinsics not implemented yet
   enum {
     Vectorizable = 1,
     AlignedOnScalar = 1,
-    size=4
+    size=4,
+    HasHalfPacket=0 // Packet2i intrinsics not implemented yet
     // FIXME check the Has*
   };
 };
 
-#if EIGEN_GNUC_AT_MOST(4,4) && !defined(__llvm__)
+#if EIGEN_GNUC_AT_MOST(4,4) && !EIGEN_COMP_LLVM
 // workaround gcc 4.2, 4.3 and 4.4 compilatin issue
 EIGEN_STRONG_INLINE float32x4_t vld1q_f32(const float* x) { return ::vld1q_f32((const float32_t*)x); }
 EIGEN_STRONG_INLINE float32x2_t vld1_f32 (const float* x) { return ::vld1_f32 ((const float32_t*)x); }
+EIGEN_STRONG_INLINE float32x2_t vld1_dup_f32 (const float* x) { return ::vld1_dup_f32 ((const float32_t*)x); }
 EIGEN_STRONG_INLINE void        vst1q_f32(float* to, float32x4_t from) { ::vst1q_f32((float32_t*)to,from); }
 EIGEN_STRONG_INLINE void        vst1_f32 (float* to, float32x2_t from) { ::vst1_f32 ((float32_t*)to,from); }
 #endif
 
-template<> struct unpacket_traits<Packet4f> { typedef float  type; enum {size=4}; };
-template<> struct unpacket_traits<Packet4i> { typedef int    type; enum {size=4}; };
+template<> struct unpacket_traits<Packet4f> { typedef float   type; enum {size=4, alignment=Aligned16}; typedef Packet4f half; };
+template<> struct unpacket_traits<Packet4i> { typedef int32_t type; enum {size=4, alignment=Aligned16}; typedef Packet4i half; };
 
 template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float&  from) { return vdupq_n_f32(from); }
-template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int&    from)   { return vdupq_n_s32(from); }
+template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int32_t&    from)   { return vdupq_n_s32(from); }
 
-template<> EIGEN_STRONG_INLINE Packet4f plset<float>(const float& a)
+template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a)
 {
-  Packet4f countdown = EIGEN_INIT_NEON_PACKET4(0, 1, 2, 3);
+  const float32_t f[] = {0, 1, 2, 3};
+  Packet4f countdown = vld1q_f32(f);
   return vaddq_f32(pset1<Packet4f>(a), countdown);
 }
-template<> EIGEN_STRONG_INLINE Packet4i plset<int>(const int& a)
+template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int32_t& a)
 {
-  Packet4i countdown = EIGEN_INIT_NEON_PACKET4(0, 1, 2, 3);
+  const int32_t i[] = {0, 1, 2, 3};
+  Packet4i countdown = vld1q_s32(i);
   return vaddq_s32(pset1<Packet4i>(a), countdown);
 }
 
@@ -132,6 +144,9 @@ template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const
 
 template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b)
 {
+#if EIGEN_ARCH_ARM64
+  return vdivq_f32(a,b);
+#else
   Packet4f inv, restep, div;
 
   // NEON does not offer a divide instruction, we have to do a reciprocal approximation
@@ -150,14 +165,51 @@ template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const
   div = vmulq_f32(a, inv);
 
   return div;
+#endif
 }
+
 template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, const Packet4i& /*b*/)
 { eigen_assert(false && "packet integer division are not supported by NEON");
   return pset1<Packet4i>(0);
 }
 
-// for some weird raisons, it has to be overloaded for packet of integers
-template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vmlaq_f32(c,a,b); }
+// Clang/ARM wrongly advertises __ARM_FEATURE_FMA even when it's not available,
+// then implements a slow software scalar fallback calling fmaf()!
+// Filed LLVM bug:
+//     https://llvm.org/bugs/show_bug.cgi?id=27216
+#if (defined __ARM_FEATURE_FMA) && !(EIGEN_COMP_CLANG && EIGEN_ARCH_ARM)
+// See bug 936.
+// FMA is available on VFPv4 i.e. when compiling with -mfpu=neon-vfpv4.
+// FMA is a true fused multiply-add i.e. only 1 rounding at the end, no intermediate rounding.
+// MLA is not fused i.e. does 2 roundings.
+// In addition to giving better accuracy, FMA also gives better performance here on a Krait (Nexus 4):
+// MLA: 10 GFlop/s ; FMA: 12 GFlops/s.
+template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vfmaq_f32(c,a,b); }
+#else
+template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) {
+#if EIGEN_COMP_CLANG && EIGEN_ARCH_ARM
+  // Clang/ARM will replace VMLA by VMUL+VADD at least for some values of -mcpu,
+  // at least -mcpu=cortex-a8 and -mcpu=cortex-a7. Since the former is the default on
+  // -march=armv7-a, that is a very common case.
+  // See e.g. this thread:
+  //     http://lists.llvm.org/pipermail/llvm-dev/2013-December/068806.html
+  // Filed LLVM bug:
+  //     https://llvm.org/bugs/show_bug.cgi?id=27219
+  Packet4f r = c;
+  asm volatile(
+    "vmla.f32 %q[r], %q[a], %q[b]"
+    : [r] "+w" (r)
+    : [a] "w" (a),
+      [b] "w" (b)
+    : );
+  return r;
+#else
+  return vmlaq_f32(c,a,b);
+#endif
+}
+#endif
+
+// No FMA instruction for int, so use MLA unconditionally.
 template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return vmlaq_s32(c,a,b); }
 
 template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { return vminq_f32(a,b); }
@@ -191,20 +243,20 @@ template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, con
 }
 template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return vbicq_s32(a,b); }
 
-template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return vld1q_f32(from); }
-template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int*   from) { EIGEN_DEBUG_ALIGNED_LOAD return vld1q_s32(from); }
+template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float*    from) { EIGEN_DEBUG_ALIGNED_LOAD return vld1q_f32(from); }
+template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int32_t*  from) { EIGEN_DEBUG_ALIGNED_LOAD return vld1q_s32(from); }
 
-template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) { EIGEN_DEBUG_UNALIGNED_LOAD return vld1q_f32(from); }
-template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)   { EIGEN_DEBUG_UNALIGNED_LOAD return vld1q_s32(from); }
+template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float*   from) { EIGEN_DEBUG_UNALIGNED_LOAD return vld1q_f32(from); }
+template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int32_t* from) { EIGEN_DEBUG_UNALIGNED_LOAD return vld1q_s32(from); }
 
-template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float*   from)
+template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from)
 {
   float32x2_t lo, hi;
   lo = vld1_dup_f32(from);
   hi = vld1_dup_f32(from+1);
   return vcombine_f32(lo, hi);
 }
-template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int*     from)
+template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int32_t* from)
 {
   int32x2_t lo, hi;
   lo = vld1_dup_s32(from);
@@ -212,18 +264,52 @@ template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int*     from)
   return vcombine_s32(lo, hi);
 }
 
-template<> EIGEN_STRONG_INLINE void pstore<float>(float*   to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_f32(to, from); }
-template<> EIGEN_STRONG_INLINE void pstore<int>(int*       to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_s32(to, from); }
+template<> EIGEN_STRONG_INLINE void pstore<float>  (float*    to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_f32(to, from); }
+template<> EIGEN_STRONG_INLINE void pstore<int32_t>(int32_t*  to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_s32(to, from); }
+
+template<> EIGEN_STRONG_INLINE void pstoreu<float>  (float*   to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_f32(to, from); }
+template<> EIGEN_STRONG_INLINE void pstoreu<int32_t>(int32_t* to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_s32(to, from); }
 
-template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*  to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_f32(to, from); }
-template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*      to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_s32(to, from); }
+template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride)
+{
+  Packet4f res = pset1<Packet4f>(0.f);
+  res = vsetq_lane_f32(from[0*stride], res, 0);
+  res = vsetq_lane_f32(from[1*stride], res, 1);
+  res = vsetq_lane_f32(from[2*stride], res, 2);
+  res = vsetq_lane_f32(from[3*stride], res, 3);
+  return res;
+}
+template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int32_t, Packet4i>(const int32_t* from, Index stride)
+{
+  Packet4i res = pset1<Packet4i>(0);
+  res = vsetq_lane_s32(from[0*stride], res, 0);
+  res = vsetq_lane_s32(from[1*stride], res, 1);
+  res = vsetq_lane_s32(from[2*stride], res, 2);
+  res = vsetq_lane_s32(from[3*stride], res, 3);
+  return res;
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride)
+{
+  to[stride*0] = vgetq_lane_f32(from, 0);
+  to[stride*1] = vgetq_lane_f32(from, 1);
+  to[stride*2] = vgetq_lane_f32(from, 2);
+  to[stride*3] = vgetq_lane_f32(from, 3);
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<int32_t, Packet4i>(int32_t* to, const Packet4i& from, Index stride)
+{
+  to[stride*0] = vgetq_lane_s32(from, 0);
+  to[stride*1] = vgetq_lane_s32(from, 1);
+  to[stride*2] = vgetq_lane_s32(from, 2);
+  to[stride*3] = vgetq_lane_s32(from, 3);
+}
 
-template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { EIGEN_ARM_PREFETCH(addr); }
-template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*     addr) { EIGEN_ARM_PREFETCH(addr); }
+template<> EIGEN_STRONG_INLINE void prefetch<float>  (const float*    addr) { EIGEN_ARM_PREFETCH(addr); }
+template<> EIGEN_STRONG_INLINE void prefetch<int32_t>(const int32_t*  addr) { EIGEN_ARM_PREFETCH(addr); }
 
 // FIXME only store the 2 first elements ?
-template<> EIGEN_STRONG_INLINE float  pfirst<Packet4f>(const Packet4f& a) { float EIGEN_ALIGN16 x[4]; vst1q_f32(x, a); return x[0]; }
-template<> EIGEN_STRONG_INLINE int    pfirst<Packet4i>(const Packet4i& a) { int   EIGEN_ALIGN16 x[4]; vst1q_s32(x, a); return x[0]; }
+template<> EIGEN_STRONG_INLINE float   pfirst<Packet4f>(const Packet4f& a) { float   EIGEN_ALIGN16 x[4]; vst1q_f32(x, a); return x[0]; }
+template<> EIGEN_STRONG_INLINE int32_t pfirst<Packet4i>(const Packet4i& a) { int32_t EIGEN_ALIGN16 x[4]; vst1q_s32(x, a); return x[0]; }
 
 template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) {
   float32x2_t a_lo, a_hi;
@@ -243,6 +329,7 @@ template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) {
   a_hi = vget_high_s32(a_r64);
   return vcombine_s32(a_hi, a_lo);
 }
+
 template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a) { return vabsq_f32(a); }
 template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) { return vabsq_s32(a); }
 
@@ -277,7 +364,7 @@ template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
   return sum;
 }
 
-template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
+template<> EIGEN_STRONG_INLINE int32_t predux<Packet4i>(const Packet4i& a)
 {
   int32x2_t a_lo, a_hi, sum;
 
@@ -324,7 +411,7 @@ template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
 
   return vget_lane_f32(prod, 0);
 }
-template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a)
+template<> EIGEN_STRONG_INLINE int32_t predux_mul<Packet4i>(const Packet4i& a)
 {
   int32x2_t a_lo, a_hi, prod;
 
@@ -352,7 +439,7 @@ template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a)
   return vget_lane_f32(min, 0);
 }
 
-template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a)
+template<> EIGEN_STRONG_INLINE int32_t predux_min<Packet4i>(const Packet4i& a)
 {
   int32x2_t a_lo, a_hi, min;
 
@@ -377,7 +464,7 @@ template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a)
   return vget_lane_f32(max, 0);
 }
 
-template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a)
+template<> EIGEN_STRONG_INLINE int32_t predux_max<Packet4i>(const Packet4i& a)
 {
   int32x2_t a_lo, a_hi, max;
 
@@ -410,9 +497,231 @@ PALIGN_NEON(0,Packet4i,vextq_s32)
 PALIGN_NEON(1,Packet4i,vextq_s32)
 PALIGN_NEON(2,Packet4i,vextq_s32)
 PALIGN_NEON(3,Packet4i,vextq_s32)
-    
+
 #undef PALIGN_NEON
 
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet4f,4>& kernel) {
+  float32x4x2_t tmp1 = vzipq_f32(kernel.packet[0], kernel.packet[1]);
+  float32x4x2_t tmp2 = vzipq_f32(kernel.packet[2], kernel.packet[3]);
+
+  kernel.packet[0] = vcombine_f32(vget_low_f32(tmp1.val[0]), vget_low_f32(tmp2.val[0]));
+  kernel.packet[1] = vcombine_f32(vget_high_f32(tmp1.val[0]), vget_high_f32(tmp2.val[0]));
+  kernel.packet[2] = vcombine_f32(vget_low_f32(tmp1.val[1]), vget_low_f32(tmp2.val[1]));
+  kernel.packet[3] = vcombine_f32(vget_high_f32(tmp1.val[1]), vget_high_f32(tmp2.val[1]));
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet4i,4>& kernel) {
+  int32x4x2_t tmp1 = vzipq_s32(kernel.packet[0], kernel.packet[1]);
+  int32x4x2_t tmp2 = vzipq_s32(kernel.packet[2], kernel.packet[3]);
+  kernel.packet[0] = vcombine_s32(vget_low_s32(tmp1.val[0]), vget_low_s32(tmp2.val[0]));
+  kernel.packet[1] = vcombine_s32(vget_high_s32(tmp1.val[0]), vget_high_s32(tmp2.val[0]));
+  kernel.packet[2] = vcombine_s32(vget_low_s32(tmp1.val[1]), vget_low_s32(tmp2.val[1]));
+  kernel.packet[3] = vcombine_s32(vget_high_s32(tmp1.val[1]), vget_high_s32(tmp2.val[1]));
+}
+
+//---------- double ----------
+
+// Clang 3.5 in the iOS toolchain has an ICE triggered by NEON intrisics for double.
+// Confirmed at least with __apple_build_version__ = 6000054.
+#ifdef __apple_build_version__
+// Let's hope that by the time __apple_build_version__ hits the 601* range, the bug will be fixed.
+// https://gist.github.com/yamaya/2924292 suggests that the 3 first digits are only updated with
+// major toolchain updates.
+#define EIGEN_APPLE_DOUBLE_NEON_BUG (__apple_build_version__ < 6010000)
+#else
+#define EIGEN_APPLE_DOUBLE_NEON_BUG 0
+#endif
+
+#if EIGEN_ARCH_ARM64 && !EIGEN_APPLE_DOUBLE_NEON_BUG
+
+// Bug 907: workaround missing declarations of the following two functions in the ADK
+// Defining these functions as templates ensures that if these intrinsics are
+// already defined in arm_neon.h, then our workaround doesn't cause a conflict
+// and has lower priority in overload resolution.
+template <typename T>
+uint64x2_t vreinterpretq_u64_f64(T a)
+{
+  return (uint64x2_t) a;
+}
+
+template <typename T>
+float64x2_t vreinterpretq_f64_u64(T a)
+{
+  return (float64x2_t) a;
+}
+
+typedef float64x2_t Packet2d;
+typedef float64x1_t Packet1d;
+
+template<> struct packet_traits<double>  : default_packet_traits
+{
+  typedef Packet2d type;
+  typedef Packet2d half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 2,
+    HasHalfPacket=0,
+   
+    HasDiv  = 1,
+    // FIXME check the Has*
+    HasSin  = 0,
+    HasCos  = 0,
+    HasLog  = 0,
+    HasExp  = 0,
+    HasSqrt = 0
+  };
+};
+
+template<> struct unpacket_traits<Packet2d> { typedef double  type; enum {size=2, alignment=Aligned16}; typedef Packet2d half; };
+
+template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double&  from) { return vdupq_n_f64(from); }
+
+template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a)
+{
+  const double countdown_raw[] = {0.0,1.0};
+  const Packet2d countdown = vld1q_f64(countdown_raw);
+  return vaddq_f64(pset1<Packet2d>(a), countdown);
+}
+template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return vaddq_f64(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return vsubq_f64(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) { return vnegq_f64(a); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; }
+
+template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return vmulq_f64(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return vdivq_f64(a,b); }
+
+#ifdef __ARM_FEATURE_FMA
+// See bug 936. See above comment about FMA for float.
+template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vfmaq_f64(c,a,b); }
+#else
+template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vmlaq_f64(c,a,b); }
+#endif
+
+template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return vminq_f64(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return vmaxq_f64(a,b); }
+
+// Logical Operations are not supported for float, so we have to reinterpret casts using NEON intrinsics
+template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b)
+{
+  return vreinterpretq_f64_u64(vandq_u64(vreinterpretq_u64_f64(a),vreinterpretq_u64_f64(b)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b)
+{
+  return vreinterpretq_f64_u64(vorrq_u64(vreinterpretq_u64_f64(a),vreinterpretq_u64_f64(b)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b)
+{
+  return vreinterpretq_f64_u64(veorq_u64(vreinterpretq_u64_f64(a),vreinterpretq_u64_f64(b)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b)
+{
+  return vreinterpretq_f64_u64(vbicq_u64(vreinterpretq_u64_f64(a),vreinterpretq_u64_f64(b)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from) { EIGEN_DEBUG_ALIGNED_LOAD return vld1q_f64(from); }
+
+template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from) { EIGEN_DEBUG_UNALIGNED_LOAD return vld1q_f64(from); }
+
+template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double*   from)
+{
+  return vld1q_dup_f64(from);
+}
+template<> EIGEN_STRONG_INLINE void pstore<double>(double*   to, const Packet2d& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_f64(to, from); }
+
+template<> EIGEN_STRONG_INLINE void pstoreu<double>(double*  to, const Packet2d& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_f64(to, from); }
+
+template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride)
+{
+  Packet2d res = pset1<Packet2d>(0.0);
+  res = vsetq_lane_f64(from[0*stride], res, 0);
+  res = vsetq_lane_f64(from[1*stride], res, 1);
+  return res;
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride)
+{
+  to[stride*0] = vgetq_lane_f64(from, 0);
+  to[stride*1] = vgetq_lane_f64(from, 1);
+}
+template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { EIGEN_ARM_PREFETCH(addr); }
+
+// FIXME only store the 2 first elements ?
+template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { return vgetq_lane_f64(a, 0); }
+
+template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a) { return vcombine_f64(vget_high_f64(a), vget_low_f64(a)); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a) { return vabsq_f64(a); }
+
+#if EIGEN_COMP_CLANG && defined(__apple_build_version__)
+// workaround ICE, see bug 907
+template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) { return (vget_low_f64(a) + vget_high_f64(a))[0]; }
+#else
+template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) { return vget_lane_f64(vget_low_f64(a) + vget_high_f64(a), 0); }
+#endif
+
+template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
+{
+  float64x2_t trn1, trn2;
+
+  // NEON zip performs interleaving of the supplied vectors.
+  // We perform two interleaves in a row to acquire the transposed vector
+  trn1 = vzip1q_f64(vecs[0], vecs[1]);
+  trn2 = vzip2q_f64(vecs[0], vecs[1]);
+
+  // Do the addition of the resulting vectors
+  return vaddq_f64(trn1, trn2);
+}
+// Other reduction functions:
+// mul
+#if EIGEN_COMP_CLANG && defined(__apple_build_version__)
+template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a) { return (vget_low_f64(a) * vget_high_f64(a))[0]; }
+#else
+template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a) { return vget_lane_f64(vget_low_f64(a) * vget_high_f64(a), 0); }
+#endif
+
+// min
+template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a) { return vgetq_lane_f64(vpminq_f64(a, a), 0); }
+
+// max
+template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a) { return vgetq_lane_f64(vpmaxq_f64(a, a), 0); }
+
+// this PALIGN_NEON business is to work around a bug in LLVM Clang 3.0 causing incorrect compilation errors,
+// see bug 347 and this LLVM bug: http://llvm.org/bugs/show_bug.cgi?id=11074
+#define PALIGN_NEON(Offset,Type,Command) \
+template<>\
+struct palign_impl<Offset,Type>\
+{\
+    EIGEN_STRONG_INLINE static void run(Type& first, const Type& second)\
+    {\
+        if (Offset!=0)\
+            first = Command(first, second, Offset);\
+    }\
+};\
+
+PALIGN_NEON(0,Packet2d,vextq_f64)
+PALIGN_NEON(1,Packet2d,vextq_f64)
+#undef PALIGN_NEON
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet2d,2>& kernel) {
+  float64x2_t trn1 = vzip1q_f64(kernel.packet[0], kernel.packet[1]);
+  float64x2_t trn2 = vzip2q_f64(kernel.packet[0], kernel.packet[1]);
+
+  kernel.packet[0] = trn1;
+  kernel.packet[1] = trn2;
+}
+#endif // EIGEN_ARCH_ARM64 
+
 } // end namespace internal
 
 } // end namespace Eigen
diff --git a/eigen/Eigen/src/Core/arch/SSE/CMakeLists.txt b/eigen/Eigen/src/Core/arch/SSE/CMakeLists.txt
deleted file mode 100644
index 46ea7cc..0000000
--- a/eigen/Eigen/src/Core/arch/SSE/CMakeLists.txt
+++ /dev/null
@@ -1,6 +0,0 @@
-FILE(GLOB Eigen_Core_arch_SSE_SRCS "*.h")
-
-INSTALL(FILES
-  ${Eigen_Core_arch_SSE_SRCS}
-  DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/arch/SSE COMPONENT Devel
-)
diff --git a/eigen/Eigen/src/Core/arch/SSE/Complex.h b/eigen/Eigen/src/Core/arch/SSE/Complex.h
index 91bba5e..5607fe0 100644
--- a/eigen/Eigen/src/Core/arch/SSE/Complex.h
+++ b/eigen/Eigen/src/Core/arch/SSE/Complex.h
@@ -22,13 +22,18 @@ struct Packet2cf
   __m128  v;
 };
 
+// Use the packet_traits defined in AVX/PacketMath.h instead if we're going
+// to leverage AVX instructions.
+#ifndef EIGEN_VECTORIZE_AVX
 template<> struct packet_traits<std::complex<float> >  : default_packet_traits
 {
   typedef Packet2cf type;
+  typedef Packet2cf half;
   enum {
     Vectorizable = 1,
     AlignedOnScalar = 1,
     size = 2,
+    HasHalfPacket = 0,
 
     HasAdd    = 1,
     HasSub    = 1,
@@ -39,11 +44,13 @@ template<> struct packet_traits<std::complex<float> >  : default_packet_traits
     HasAbs2   = 0,
     HasMin    = 0,
     HasMax    = 0,
-    HasSetLinear = 0
+    HasSetLinear = 0,
+    HasBlend = 1
   };
 };
+#endif
 
-template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2}; };
+template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16}; typedef Packet2cf half; };
 
 template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_add_ps(a.v,b.v)); }
 template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_sub_ps(a.v,b.v)); }
@@ -60,7 +67,6 @@ template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a)
 
 template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
 {
-  // TODO optimize it for SSE3 and 4
   #ifdef EIGEN_VECTORIZE_SSE3
   return Packet2cf(_mm_addsub_ps(_mm_mul_ps(_mm_moveldup_ps(a.v), b.v),
                                  _mm_mul_ps(_mm_movehdup_ps(a.v),
@@ -104,8 +110,23 @@ template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<flo
 
 template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from) { return pset1<Packet2cf>(*from); }
 
-template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> *   to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore(&numext::real_ref(*to), from.v); }
-template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> *   to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&numext::real_ref(*to), from.v); }
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> *   to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore(&numext::real_ref(*to), Packet4f(from.v)); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> *   to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&numext::real_ref(*to), Packet4f(from.v)); }
+
+
+template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from, Index stride)
+{
+  return Packet2cf(_mm_set_ps(std::imag(from[1*stride]), std::real(from[1*stride]),
+                              std::imag(from[0*stride]), std::real(from[0*stride])));
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, Index stride)
+{
+  to[stride*0] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 0)),
+                                     _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 1)));
+  to[stride*1] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 2)),
+                                     _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 3)));
+}
 
 template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> *   addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
 
@@ -124,7 +145,7 @@ template<> EIGEN_STRONG_INLINE std::complex<float>  pfirst<Packet2cf>(const Pack
   #endif
 }
 
-template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) { return Packet2cf(_mm_castpd_ps(preverse(_mm_castps_pd(a.v)))); }
+template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) { return Packet2cf(_mm_castpd_ps(preverse(Packet2d(_mm_castps_pd(a.v))))); }
 
 template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a)
 {
@@ -214,7 +235,7 @@ template<> struct conj_helper<Packet4f, Packet2cf, false,false>
   { return padd(c, pmul(x,y)); }
 
   EIGEN_STRONG_INLINE Packet2cf pmul(const Packet4f& x, const Packet2cf& y) const
-  { return Packet2cf(Eigen::internal::pmul(x, y.v)); }
+  { return Packet2cf(Eigen::internal::pmul<Packet4f>(x, y.v)); }
 };
 
 template<> struct conj_helper<Packet2cf, Packet4f, false,false>
@@ -223,7 +244,7 @@ template<> struct conj_helper<Packet2cf, Packet4f, false,false>
   { return padd(c, pmul(x,y)); }
 
   EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& x, const Packet4f& y) const
-  { return Packet2cf(Eigen::internal::pmul(x.v, y)); }
+  { return Packet2cf(Eigen::internal::pmul<Packet4f>(x.v, y)); }
 };
 
 template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
@@ -234,7 +255,7 @@ template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, con
   return Packet2cf(_mm_div_ps(res.v,_mm_add_ps(s,_mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(s), 0xb1)))));
 }
 
-EIGEN_STRONG_INLINE Packet2cf pcplxflip/*<Packet2cf>*/(const Packet2cf& x)
+EIGEN_STRONG_INLINE Packet2cf pcplxflip/* <Packet2cf> */(const Packet2cf& x)
 {
   return Packet2cf(vec4f_swizzle1(x.v, 1, 0, 3, 2));
 }
@@ -248,13 +269,18 @@ struct Packet1cd
   __m128d  v;
 };
 
+// Use the packet_traits defined in AVX/PacketMath.h instead if we're going
+// to leverage AVX instructions.
+#ifndef EIGEN_VECTORIZE_AVX
 template<> struct packet_traits<std::complex<double> >  : default_packet_traits
 {
   typedef Packet1cd type;
+  typedef Packet1cd half;
   enum {
     Vectorizable = 1,
     AlignedOnScalar = 0,
     size = 1,
+    HasHalfPacket = 0,
 
     HasAdd    = 1,
     HasSub    = 1,
@@ -268,12 +294,13 @@ template<> struct packet_traits<std::complex<double> >  : default_packet_traits
     HasSetLinear = 0
   };
 };
+#endif
 
-template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1}; };
+template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16}; typedef Packet1cd half; };
 
 template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_add_pd(a.v,b.v)); }
 template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_sub_pd(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(a.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(Packet2d(a.v))); }
 template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a)
 {
   const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
@@ -282,9 +309,8 @@ template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a)
 
 template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
 {
-  // TODO optimize it for SSE3 and 4
   #ifdef EIGEN_VECTORIZE_SSE3
-  return Packet1cd(_mm_addsub_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v),
+  return Packet1cd(_mm_addsub_pd(_mm_mul_pd(_mm_movedup_pd(a.v), b.v),
                                  _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
                                             vec2d_swizzle1(b.v, 1, 0))));
   #else
@@ -311,8 +337,8 @@ template<> EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<dou
 template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>* from) { return pset1<Packet1cd>(*from); }
 
 // FIXME force unaligned store, this is a temporary fix
-template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); }
-template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); }
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, Packet2d(from.v)); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, Packet2d(from.v)); }
 
 template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> *   addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
 
@@ -410,7 +436,7 @@ template<> struct conj_helper<Packet2d, Packet1cd, false,false>
   { return padd(c, pmul(x,y)); }
 
   EIGEN_STRONG_INLINE Packet1cd pmul(const Packet2d& x, const Packet1cd& y) const
-  { return Packet1cd(Eigen::internal::pmul(x, y.v)); }
+  { return Packet1cd(Eigen::internal::pmul<Packet2d>(x, y.v)); }
 };
 
 template<> struct conj_helper<Packet1cd, Packet2d, false,false>
@@ -419,7 +445,7 @@ template<> struct conj_helper<Packet1cd, Packet2d, false,false>
   { return padd(c, pmul(x,y)); }
 
   EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& x, const Packet2d& y) const
-  { return Packet1cd(Eigen::internal::pmul(x.v, y)); }
+  { return Packet1cd(Eigen::internal::pmul<Packet2d>(x.v, y)); }
 };
 
 template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
@@ -430,9 +456,44 @@ template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, con
   return Packet1cd(_mm_div_pd(res.v, _mm_add_pd(s,_mm_shuffle_pd(s, s, 0x1))));
 }
 
-EIGEN_STRONG_INLINE Packet1cd pcplxflip/*<Packet1cd>*/(const Packet1cd& x)
+EIGEN_STRONG_INLINE Packet1cd pcplxflip/* <Packet1cd> */(const Packet1cd& x)
+{
+  return Packet1cd(preverse(Packet2d(x.v)));
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet2cf,2>& kernel) {
+  __m128d w1 = _mm_castps_pd(kernel.packet[0].v);
+  __m128d w2 = _mm_castps_pd(kernel.packet[1].v);
+
+  __m128 tmp = _mm_castpd_ps(_mm_unpackhi_pd(w1, w2));
+  kernel.packet[0].v = _mm_castpd_ps(_mm_unpacklo_pd(w1, w2));
+  kernel.packet[1].v = tmp;
+}
+
+template<>  EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, const Packet2cf& thenPacket, const Packet2cf& elsePacket) {
+  __m128d result = pblend<Packet2d>(ifPacket, _mm_castps_pd(thenPacket.v), _mm_castps_pd(elsePacket.v));
+  return Packet2cf(_mm_castpd_ps(result));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf pinsertfirst(const Packet2cf& a, std::complex<float> b)
+{
+  return Packet2cf(_mm_loadl_pi(a.v, reinterpret_cast<const __m64*>(&b)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd pinsertfirst(const Packet1cd&, std::complex<double> b)
+{
+  return pset1<Packet1cd>(b);
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf pinsertlast(const Packet2cf& a, std::complex<float> b)
+{
+  return Packet2cf(_mm_loadh_pi(a.v, reinterpret_cast<const __m64*>(&b)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd pinsertlast(const Packet1cd&, std::complex<double> b)
 {
-  return Packet1cd(preverse(x.v));
+  return pset1<Packet1cd>(b);
 }
 
 } // end namespace internal
diff --git a/eigen/Eigen/src/Core/arch/SSE/MathFunctions.h b/eigen/Eigen/src/Core/arch/SSE/MathFunctions.h
index 2b07168..7b5f948 100644
--- a/eigen/Eigen/src/Core/arch/SSE/MathFunctions.h
+++ b/eigen/Eigen/src/Core/arch/SSE/MathFunctions.h
@@ -32,7 +32,7 @@ Packet4f plog<Packet4f>(const Packet4f& _x)
   /* the smallest non denormalized float number */
   _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(min_norm_pos,  0x00800000);
   _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(minus_inf,     0xff800000);//-1.f/0.f);
-  
+
   /* natural logarithm computed for 4 simultaneous float
     return NaN for x <= 0
   */
@@ -63,7 +63,7 @@ Packet4f plog<Packet4f>(const Packet4f& _x)
   x = _mm_or_ps(x, p4f_half);
 
   emm0 = _mm_sub_epi32(emm0, p4i_0x7f);
-  Packet4f e = padd(_mm_cvtepi32_ps(emm0), p4f_1);
+  Packet4f e = padd(Packet4f(_mm_cvtepi32_ps(emm0)), p4f_1);
 
   /* part2:
      if( x < SQRTHF ) {
@@ -72,9 +72,9 @@ Packet4f plog<Packet4f>(const Packet4f& _x)
      } else { x = x - 1.0; }
   */
   Packet4f mask = _mm_cmplt_ps(x, p4f_cephes_SQRTHF);
-  Packet4f tmp = _mm_and_ps(x, mask);
+  Packet4f tmp = pand(x, mask);
   x = psub(x, p4f_1);
-  e = psub(e, _mm_and_ps(p4f_1, mask));
+  e = psub(e, pand(p4f_1, mask));
   x = padd(x, tmp);
 
   Packet4f x2 = pmul(x,x);
@@ -444,32 +444,119 @@ Packet4f pcos<Packet4f>(const Packet4f& _x)
 
 #if EIGEN_FAST_MATH
 
-// This is based on Quake3's fast inverse square root.
+// Functions for sqrt.
+// The EIGEN_FAST_MATH version uses the _mm_rsqrt_ps approximation and one step
+// of Newton's method, at a cost of 1-2 bits of precision as opposed to the
+// exact solution. It does not handle +inf, or denormalized numbers correctly.
+// The main advantage of this approach is not just speed, but also the fact that
+// it can be inlined and pipelined with other computations, further reducing its
+// effective latency. This is similar to Quake3's fast inverse square root.
 // For detail see here: http://www.beyond3d.com/content/articles/8/
-// It lacks 1 (or 2 bits in some rare cases) of precision, and does not handle negative, +inf, or denormalized numbers correctly.
 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
 Packet4f psqrt<Packet4f>(const Packet4f& _x)
 {
   Packet4f half = pmul(_x, pset1<Packet4f>(.5f));
+  Packet4f denormal_mask = _mm_and_ps(
+      _mm_cmpge_ps(_x, _mm_setzero_ps()),
+      _mm_cmplt_ps(_x, pset1<Packet4f>((std::numeric_limits<float>::min)())));
 
-  /* select only the inverse sqrt of non-zero inputs */
-  Packet4f non_zero_mask = _mm_cmpge_ps(_x, pset1<Packet4f>((std::numeric_limits<float>::min)()));
-  Packet4f x = _mm_and_ps(non_zero_mask, _mm_rsqrt_ps(_x));
-
+  // Compute approximate reciprocal sqrt.
+  Packet4f x = _mm_rsqrt_ps(_x);
+  // Do a single step of Newton's iteration.
   x = pmul(x, psub(pset1<Packet4f>(1.5f), pmul(half, pmul(x,x))));
-  return pmul(_x,x);
+  // Flush results for denormals to zero.
+  return _mm_andnot_ps(denormal_mask, pmul(_x,x));
 }
 
 #else
 
-template<> EIGEN_STRONG_INLINE Packet4f psqrt<Packet4f>(const Packet4f& x) { return _mm_sqrt_ps(x); }
+template<>EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f psqrt<Packet4f>(const Packet4f& x) { return _mm_sqrt_ps(x); }
+
+#endif
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet2d psqrt<Packet2d>(const Packet2d& x) { return _mm_sqrt_pd(x); }
+
+#if EIGEN_FAST_MATH
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f prsqrt<Packet4f>(const Packet4f& _x) {
+  _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(inf, 0x7f800000);
+  _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(nan, 0x7fc00000);
+  _EIGEN_DECLARE_CONST_Packet4f(one_point_five, 1.5f);
+  _EIGEN_DECLARE_CONST_Packet4f(minus_half, -0.5f);
+  _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(flt_min, 0x00800000);
+
+  Packet4f neg_half = pmul(_x, p4f_minus_half);
+
+  // select only the inverse sqrt of positive normal inputs (denormals are
+  // flushed to zero and cause infs as well).
+  Packet4f le_zero_mask = _mm_cmple_ps(_x, p4f_flt_min);
+  Packet4f x = _mm_andnot_ps(le_zero_mask, _mm_rsqrt_ps(_x));
+
+  // Fill in NaNs and Infs for the negative/zero entries.
+  Packet4f neg_mask = _mm_cmplt_ps(_x, _mm_setzero_ps());
+  Packet4f zero_mask = _mm_andnot_ps(neg_mask, le_zero_mask);
+  Packet4f infs_and_nans = _mm_or_ps(_mm_and_ps(neg_mask, p4f_nan),
+                                     _mm_and_ps(zero_mask, p4f_inf));
+
+  // Do a single step of Newton's iteration.
+  x = pmul(x, pmadd(neg_half, pmul(x, x), p4f_one_point_five));
+
+  // Insert NaNs and Infs in all the right places.
+  return _mm_or_ps(x, infs_and_nans);
+}
+
+#else
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f prsqrt<Packet4f>(const Packet4f& x) {
+  // Unfortunately we can't use the much faster mm_rqsrt_ps since it only provides an approximation.
+  return _mm_div_ps(pset1<Packet4f>(1.0f), _mm_sqrt_ps(x));
+}
 
 #endif
 
-template<> EIGEN_STRONG_INLINE Packet2d psqrt<Packet2d>(const Packet2d& x) { return _mm_sqrt_pd(x); }
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet2d prsqrt<Packet2d>(const Packet2d& x) {
+  // Unfortunately we can't use the much faster mm_rqsrt_pd since it only provides an approximation.
+  return _mm_div_pd(pset1<Packet2d>(1.0), _mm_sqrt_pd(x));
+}
+
+// Hyperbolic Tangent function.
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f
+ptanh<Packet4f>(const Packet4f& x) {
+  return internal::generic_fast_tanh_float(x);
+}
 
 } // end namespace internal
 
+namespace numext {
+
+template<>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float sqrt(const float &x)
+{
+  return internal::pfirst(internal::Packet4f(_mm_sqrt_ss(_mm_set_ss(x))));
+}
+
+template<>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double sqrt(const double &x)
+{
+#if EIGEN_COMP_GNUC_STRICT
+  // This works around a GCC bug generating poor code for _mm_sqrt_pd
+  // See https://bitbucket.org/eigen/eigen/commits/14f468dba4d350d7c19c9b93072e19f7b3df563b
+  return internal::pfirst(internal::Packet2d(__builtin_ia32_sqrtsd(_mm_set_sd(x))));
+#else
+  return internal::pfirst(internal::Packet2d(_mm_sqrt_pd(_mm_set_sd(x))));
+#endif
+}
+
+} // end namespace numex
+
 } // end namespace Eigen
 
 #endif // EIGEN_MATH_FUNCTIONS_SSE_H
diff --git a/eigen/Eigen/src/Core/arch/SSE/PacketMath.h b/eigen/Eigen/src/Core/arch/SSE/PacketMath.h
index bef898b..03c8a2c 100644
--- a/eigen/Eigen/src/Core/arch/SSE/PacketMath.h
+++ b/eigen/Eigen/src/Core/arch/SSE/PacketMath.h
@@ -22,9 +22,40 @@ namespace internal {
 #define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*))
 #endif
 
+#ifdef __FMA__
+#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD 1
+#endif
+#endif
+
+#if (defined EIGEN_VECTORIZE_AVX) && (EIGEN_COMP_GNUC_STRICT || EIGEN_COMP_MINGW) && (__GXX_ABI_VERSION < 1004)
+// With GCC's default ABI version, a __m128 or __m256 are the same types and therefore we cannot
+// have overloads for both types without linking error.
+// One solution is to increase ABI version using -fabi-version=4 (or greater).
+// Otherwise, we workaround this inconvenience by wrapping 128bit types into the following helper
+// structure:
+template<typename T>
+struct eigen_packet_wrapper
+{
+  EIGEN_ALWAYS_INLINE operator T&() { return m_val; }
+  EIGEN_ALWAYS_INLINE operator const T&() const { return m_val; }
+  EIGEN_ALWAYS_INLINE eigen_packet_wrapper() {}
+  EIGEN_ALWAYS_INLINE eigen_packet_wrapper(const T &v) : m_val(v) {}
+  EIGEN_ALWAYS_INLINE eigen_packet_wrapper& operator=(const T &v) {
+    m_val = v;
+    return *this;
+  }
+
+  T m_val;
+};
+typedef eigen_packet_wrapper<__m128>  Packet4f;
+typedef eigen_packet_wrapper<__m128i> Packet4i;
+typedef eigen_packet_wrapper<__m128d> Packet2d;
+#else
 typedef __m128  Packet4f;
 typedef __m128i Packet4i;
 typedef __m128d Packet2d;
+#endif
 
 template<> struct is_arithmetic<__m128>  { enum { value = true }; };
 template<> struct is_arithmetic<__m128i> { enum { value = true }; };
@@ -38,7 +69,7 @@ template<> struct is_arithmetic<__m128d> { enum { value = true }; };
 
 #define vec2d_swizzle1(v,p,q) \
   (_mm_castsi128_pd(_mm_shuffle_epi32( _mm_castpd_si128(v), ((q*2+1)<<6|(q*2)<<4|(p*2+1)<<2|(p*2)))))
-  
+
 #define vec4f_swizzle2(a,b,p,q,r,s) \
   (_mm_shuffle_ps( (a), (b), ((s)<<6|(r)<<4|(q)<<2|(p))))
 
@@ -58,51 +89,85 @@ template<> struct is_arithmetic<__m128d> { enum { value = true }; };
   const Packet4i p4i_##NAME = pset1<Packet4i>(X)
 
 
+// Use the packet_traits defined in AVX/PacketMath.h instead if we're going
+// to leverage AVX instructions.
+#ifndef EIGEN_VECTORIZE_AVX
 template<> struct packet_traits<float>  : default_packet_traits
 {
   typedef Packet4f type;
+  typedef Packet4f half;
   enum {
     Vectorizable = 1,
     AlignedOnScalar = 1,
     size=4,
+    HasHalfPacket = 0,
 
     HasDiv  = 1,
     HasSin  = EIGEN_FAST_MATH,
     HasCos  = EIGEN_FAST_MATH,
     HasLog  = 1,
     HasExp  = 1,
-    HasSqrt = 1
+    HasSqrt = 1,
+    HasRsqrt = 1,
+    HasTanh  = EIGEN_FAST_MATH,
+    HasBlend = 1
+
+#ifdef EIGEN_VECTORIZE_SSE4_1
+    ,
+    HasRound = 1,
+    HasFloor = 1,
+    HasCeil = 1
+#endif
   };
 };
 template<> struct packet_traits<double> : default_packet_traits
 {
   typedef Packet2d type;
+  typedef Packet2d half;
   enum {
     Vectorizable = 1,
     AlignedOnScalar = 1,
     size=2,
+    HasHalfPacket = 0,
 
     HasDiv  = 1,
     HasExp  = 1,
-    HasSqrt = 1
+    HasSqrt = 1,
+    HasRsqrt = 1,
+    HasBlend = 1
+
+#ifdef EIGEN_VECTORIZE_SSE4_1
+    ,
+    HasRound = 1,
+    HasFloor = 1,
+    HasCeil = 1
+#endif
   };
 };
+#endif
 template<> struct packet_traits<int>    : default_packet_traits
 {
   typedef Packet4i type;
+  typedef Packet4i half;
   enum {
-    // FIXME check the Has*
     Vectorizable = 1,
     AlignedOnScalar = 1,
-    size=4
+    size=4,
+
+    HasBlend = 1
   };
 };
 
-template<> struct unpacket_traits<Packet4f> { typedef float  type; enum {size=4}; };
-template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2}; };
-template<> struct unpacket_traits<Packet4i> { typedef int    type; enum {size=4}; };
+template<> struct unpacket_traits<Packet4f> { typedef float  type; enum {size=4, alignment=Aligned16}; typedef Packet4f half; };
+template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16}; typedef Packet2d half; };
+template<> struct unpacket_traits<Packet4i> { typedef int    type; enum {size=4, alignment=Aligned16}; typedef Packet4i half; };
+
+#ifndef EIGEN_VECTORIZE_AVX
+template<> struct scalar_div_cost<float,true> { enum { value = 7 }; };
+template<> struct scalar_div_cost<double,true> { enum { value = 8 }; };
+#endif
 
-#if defined(_MSC_VER) && (_MSC_VER==1500)
+#if EIGEN_COMP_MSVC==1500
 // Workaround MSVC 9 internal compiler error.
 // TODO: It has been detected with win64 builds (amd64), so let's check whether it also happens in 32bits+SSE mode
 // TODO: let's check whether there does not exist a better fix, like adding a pset0() function. (it crashed on pset1(0)).
@@ -110,14 +175,25 @@ template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float&  from) { re
 template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { return _mm_set_pd(from,from); }
 template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int&    from) { return _mm_set_epi32(from,from,from,from); }
 #else
-template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float&  from) { return _mm_set1_ps(from); }
+template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float&  from) { return _mm_set_ps1(from); }
 template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { return _mm_set1_pd(from); }
 template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int&    from) { return _mm_set1_epi32(from); }
 #endif
 
-template<> EIGEN_STRONG_INLINE Packet4f plset<float>(const float& a) { return _mm_add_ps(pset1<Packet4f>(a), _mm_set_ps(3,2,1,0)); }
-template<> EIGEN_STRONG_INLINE Packet2d plset<double>(const double& a) { return _mm_add_pd(pset1<Packet2d>(a),_mm_set_pd(1,0)); }
-template<> EIGEN_STRONG_INLINE Packet4i plset<int>(const int& a) { return _mm_add_epi32(pset1<Packet4i>(a),_mm_set_epi32(3,2,1,0)); }
+// GCC generates a shufps instruction for _mm_set1_ps/_mm_load1_ps instead of the more efficient pshufd instruction.
+// However, using inrinsics for pset1 makes gcc to generate crappy code in some cases (see bug 203)
+// Using inline assembly is also not an option because then gcc fails to reorder properly the instructions.
+// Therefore, we introduced the pload1 functions to be used in product kernels for which bug 203 does not apply.
+// Also note that with AVX, we want it to generate a vbroadcastss.
+#if EIGEN_COMP_GNUC_STRICT && (!defined __AVX__)
+template<> EIGEN_STRONG_INLINE Packet4f pload1<Packet4f>(const float *from) {
+  return vec4f_swizzle1(_mm_load_ss(from),0,0,0,0);
+}
+#endif
+
+template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a) { return _mm_add_ps(pset1<Packet4f>(a), _mm_set_ps(3,2,1,0)); }
+template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a) { return _mm_add_pd(pset1<Packet2d>(a),_mm_set_pd(1,0)); }
+template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int& a) { return _mm_add_epi32(pset1<Packet4i>(a),_mm_set_epi32(3,2,1,0)); }
 
 template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_add_ps(a,b); }
 template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_add_pd(a,b); }
@@ -139,7 +215,7 @@ template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a)
 }
 template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a)
 {
-  return psub(_mm_setr_epi32(0,0,0,0), a);
+  return psub(Packet4i(_mm_setr_epi32(0,0,0,0)), a);
 }
 
 template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; }
@@ -166,16 +242,42 @@ template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const
 
 template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_div_ps(a,b); }
 template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_div_pd(a,b); }
-template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, const Packet4i& /*b*/)
-{ eigen_assert(false && "packet integer division are not supported by SSE");
-  return pset1<Packet4i>(0);
-}
 
 // for some weird raisons, it has to be overloaded for packet of integers
 template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd(pmul(a,b), c); }
+#ifdef __FMA__
+template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return _mm_fmadd_ps(a,b,c); }
+template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return _mm_fmadd_pd(a,b,c); }
+#endif
 
-template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_min_ps(a,b); }
-template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_min_pd(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) {
+#if EIGEN_COMP_GNUC
+  // There appears to be a bug in GCC, by which the optimizer may
+  // flip the argument order in calls to _mm_min_ps, so we have to
+  // resort to inline ASM here. This is supposed to be fixed in gcc6.3,
+  // see also: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72867
+  Packet4f res = b;
+  asm("minps %[a], %[res]" : [res] "+x" (res) : [a] "x" (a));
+  return res;
+#else
+  // Arguments are reversed to match NaN propagation behavior of std::min.
+  return _mm_min_ps(b, a);
+#endif
+}
+template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) {
+#if EIGEN_COMP_GNUC
+  // There appears to be a bug in GCC, by which the optimizer may
+  // flip the argument order in calls to _mm_min_pd, so we have to
+  // resort to inline ASM here. This is supposed to be fixed in gcc6.3,
+  // see also: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72867
+  Packet2d res = b;
+  asm("minpd %[a], %[res]" : [res] "+x" (res) : [a] "x" (a));
+  return res;
+#else
+  // Arguments are reversed to match NaN propagation behavior of std::min.
+  return _mm_min_pd(b, a);
+#endif
+}
 template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b)
 {
 #ifdef EIGEN_VECTORIZE_SSE4_1
@@ -187,8 +289,34 @@ template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const
 #endif
 }
 
-template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_max_ps(a,b); }
-template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_max_pd(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) {
+#if EIGEN_COMP_GNUC
+  // There appears to be a bug in GCC, by which the optimizer may
+  // flip the argument order in calls to _mm_max_ps, so we have to
+  // resort to inline ASM here. This is supposed to be fixed in gcc6.3,
+  // see also: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72867
+  Packet4f res = b;
+  asm("maxps %[a], %[res]" : [res] "+x" (res) : [a] "x" (a));
+  return res;
+#else
+  // Arguments are reversed to match NaN propagation behavior of std::max.
+  return _mm_max_ps(b, a);
+#endif
+}
+template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) {
+#if EIGEN_COMP_GNUC
+  // There appears to be a bug in GCC, by which the optimizer may
+  // flip the argument order in calls to _mm_max_pd, so we have to
+  // resort to inline ASM here. This is supposed to be fixed in gcc6.3,
+  // see also: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72867
+  Packet2d res = b;
+  asm("maxpd %[a], %[res]" : [res] "+x" (res) : [a] "x" (a));
+  return res;
+#else
+  // Arguments are reversed to match NaN propagation behavior of std::max.
+  return _mm_max_pd(b, a);
+#endif
+}
 template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b)
 {
 #ifdef EIGEN_VECTORIZE_SSE4_1
@@ -200,6 +328,17 @@ template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const
 #endif
 }
 
+#ifdef EIGEN_VECTORIZE_SSE4_1
+template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a) { return _mm_round_ps(a, 0); }
+template<> EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a) { return _mm_round_pd(a, 0); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const Packet4f& a) { return _mm_ceil_ps(a); }
+template<> EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const Packet2d& a) { return _mm_ceil_pd(a); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) { return _mm_floor_ps(a); }
+template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) { return _mm_floor_pd(a); }
+#endif
+
 template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_and_ps(a,b); }
 template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_and_pd(a,b); }
 template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_and_si128(a,b); }
@@ -218,16 +357,14 @@ template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, con
 
 template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float*   from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_ps(from); }
 template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double*  from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_pd(from); }
-template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int*     from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_si128(reinterpret_cast<const Packet4i*>(from)); }
+template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int*     from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_si128(reinterpret_cast<const __m128i*>(from)); }
 
-#if defined(_MSC_VER)
+#if EIGEN_COMP_MSVC
   template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float*  from) {
     EIGEN_DEBUG_UNALIGNED_LOAD
-    #if (_MSC_VER==1600)
+    #if (EIGEN_COMP_MSVC==1600)
     // NOTE Some version of MSVC10 generates bad code when using _mm_loadu_ps
     // (i.e., it does not generate an unaligned load!!
-    // TODO On most architectures this version should also be faster than a single _mm_loadu_ps
-    // so we could also enable it for MSVC08 but first we have to make this later does not generate crap when doing so...
     __m128 res = _mm_loadl_pi(_mm_set1_ps(0.0f), (const __m64*)(from));
     res = _mm_loadh_pi(res, (const __m64*)(from+2));
     return res;
@@ -266,46 +403,77 @@ template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double*  from)
 template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int*     from)
 {
   Packet4i tmp;
-  tmp = _mm_loadl_epi64(reinterpret_cast<const Packet4i*>(from));
+  tmp = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(from));
   return vec4i_swizzle1(tmp, 0, 0, 1, 1);
 }
 
 template<> EIGEN_STRONG_INLINE void pstore<float>(float*   to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_ps(to, from); }
 template<> EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet2d& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_pd(to, from); }
-template<> EIGEN_STRONG_INLINE void pstore<int>(int*       to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_si128(reinterpret_cast<Packet4i*>(to), from); }
+template<> EIGEN_STRONG_INLINE void pstore<int>(int*       to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_si128(reinterpret_cast<__m128i*>(to), from); }
+
+template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_pd(to, from); }
+template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*   to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_ps(to, from); }
+template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*       to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_si128(reinterpret_cast<__m128i*>(to), from); }
+
+template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride)
+{
+ return _mm_set_ps(from[3*stride], from[2*stride], from[1*stride], from[0*stride]);
+}
+template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride)
+{
+ return _mm_set_pd(from[1*stride], from[0*stride]);
+}
+template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* from, Index stride)
+{
+ return _mm_set_epi32(from[3*stride], from[2*stride], from[1*stride], from[0*stride]);
+ }
 
-template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d& from) {
-  EIGEN_DEBUG_UNALIGNED_STORE
-  _mm_storel_pd((to), from);
-  _mm_storeh_pd((to+1), from);
+template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride)
+{
+  to[stride*0] = _mm_cvtss_f32(from);
+  to[stride*1] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 1));
+  to[stride*2] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 2));
+  to[stride*3] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 3));
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride)
+{
+  to[stride*0] = _mm_cvtsd_f64(from);
+  to[stride*1] = _mm_cvtsd_f64(_mm_shuffle_pd(from, from, 1));
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const Packet4i& from, Index stride)
+{
+  to[stride*0] = _mm_cvtsi128_si32(from);
+  to[stride*1] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 1));
+  to[stride*2] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 2));
+  to[stride*3] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 3));
 }
-template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*  to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast<double*>(to), _mm_castps_pd(from)); }
-template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*      to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast<double*>(to), _mm_castsi128_pd(from)); }
 
 // some compilers might be tempted to perform multiple moves instead of using a vector path.
 template<> EIGEN_STRONG_INLINE void pstore1<Packet4f>(float* to, const float& a)
 {
   Packet4f pa = _mm_set_ss(a);
-  pstore(to, vec4f_swizzle1(pa,0,0,0,0));
+  pstore(to, Packet4f(vec4f_swizzle1(pa,0,0,0,0)));
 }
 // some compilers might be tempted to perform multiple moves instead of using a vector path.
 template<> EIGEN_STRONG_INLINE void pstore1<Packet2d>(double* to, const double& a)
 {
   Packet2d pa = _mm_set_sd(a);
-  pstore(to, vec2d_swizzle1(pa,0,0));
+  pstore(to, Packet2d(vec2d_swizzle1(pa,0,0)));
 }
 
+#ifndef EIGEN_VECTORIZE_AVX
 template<> EIGEN_STRONG_INLINE void prefetch<float>(const float*   addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
 template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
 template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*       addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+#endif
 
-#if defined(_MSC_VER) && defined(_WIN64) && !defined(__INTEL_COMPILER)
+#if EIGEN_COMP_MSVC_STRICT && EIGEN_OS_WIN64
 // The temporary variable fixes an internal compilation error in vs <= 2008 and a wrong-result bug in vs 2010
 // Direct of the struct members fixed bug #62.
 template<> EIGEN_STRONG_INLINE float  pfirst<Packet4f>(const Packet4f& a) { return a.m128_f32[0]; }
 template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { return a.m128d_f64[0]; }
 template<> EIGEN_STRONG_INLINE int    pfirst<Packet4i>(const Packet4i& a) { int x = _mm_cvtsi128_si32(a); return x; }
-#elif defined(_MSC_VER) && !defined(__INTEL_COMPILER)
+#elif EIGEN_COMP_MSVC_STRICT
 // The temporary variable fixes an internal compilation error in vs <= 2008 and a wrong-result bug in vs 2010
 template<> EIGEN_STRONG_INLINE float  pfirst<Packet4f>(const Packet4f& a) { float x = _mm_cvtss_f32(a); return x; }
 template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { double x = _mm_cvtsd_f64(a); return x; }
@@ -323,7 +491,6 @@ template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a)
 template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a)
 { return _mm_shuffle_epi32(a,0x1B); }
 
-
 template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a)
 {
   const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF));
@@ -344,6 +511,38 @@ template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a)
   #endif
 }
 
+// with AVX, the default implementations based on pload1 are faster
+#ifndef __AVX__
+template<> EIGEN_STRONG_INLINE void
+pbroadcast4<Packet4f>(const float *a,
+                      Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3)
+{
+  a3 = pload<Packet4f>(a);
+  a0 = vec4f_swizzle1(a3, 0,0,0,0);
+  a1 = vec4f_swizzle1(a3, 1,1,1,1);
+  a2 = vec4f_swizzle1(a3, 2,2,2,2);
+  a3 = vec4f_swizzle1(a3, 3,3,3,3);
+}
+template<> EIGEN_STRONG_INLINE void
+pbroadcast4<Packet2d>(const double *a,
+                      Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3)
+{
+#ifdef EIGEN_VECTORIZE_SSE3
+  a0 = _mm_loaddup_pd(a+0);
+  a1 = _mm_loaddup_pd(a+1);
+  a2 = _mm_loaddup_pd(a+2);
+  a3 = _mm_loaddup_pd(a+3);
+#else
+  a1 = pload<Packet2d>(a);
+  a0 = vec2d_swizzle1(a1, 0,0);
+  a1 = vec2d_swizzle1(a1, 1,1);
+  a3 = pload<Packet2d>(a+2);
+  a2 = vec2d_swizzle1(a3, 0,0);
+  a3 = vec2d_swizzle1(a3, 1,1);
+#endif
+}
+#endif
+
 EIGEN_STRONG_INLINE void punpackp(Packet4f* vecs)
 {
   vecs[1] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0x55));
@@ -353,47 +552,17 @@ EIGEN_STRONG_INLINE void punpackp(Packet4f* vecs)
 }
 
 #ifdef EIGEN_VECTORIZE_SSE3
-// TODO implement SSE2 versions as well as integer versions
 template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
 {
   return _mm_hadd_ps(_mm_hadd_ps(vecs[0], vecs[1]),_mm_hadd_ps(vecs[2], vecs[3]));
 }
+
 template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
 {
   return _mm_hadd_pd(vecs[0], vecs[1]);
 }
-// SSSE3 version:
-// EIGEN_STRONG_INLINE Packet4i preduxp(const Packet4i* vecs)
-// {
-//   return _mm_hadd_epi32(_mm_hadd_epi32(vecs[0], vecs[1]),_mm_hadd_epi32(vecs[2], vecs[3]));
-// }
 
-template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
-{
-  Packet4f tmp0 = _mm_hadd_ps(a,a);
-  return pfirst(_mm_hadd_ps(tmp0, tmp0));
-}
-
-template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) { return pfirst(_mm_hadd_pd(a, a)); }
-
-// SSSE3 version:
-// EIGEN_STRONG_INLINE float predux(const Packet4i& a)
-// {
-//   Packet4i tmp0 = _mm_hadd_epi32(a,a);
-//   return pfirst(_mm_hadd_epi32(tmp0, tmp0));
-// }
 #else
-// SSE2 versions
-template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
-{
-  Packet4f tmp = _mm_add_ps(a, _mm_movehl_ps(a,a));
-  return pfirst(_mm_add_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
-}
-template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a)
-{
-  return pfirst(_mm_add_sd(a, _mm_unpackhi_pd(a,a)));
-}
-
 template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
 {
   Packet4f tmp0, tmp1, tmp2;
@@ -414,10 +583,45 @@ template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
 }
 #endif  // SSE3
 
+template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
+{
+  // Disable SSE3 _mm_hadd_pd that is extremely slow on all existing Intel's architectures
+  // (from Nehalem to Haswell)
+// #ifdef EIGEN_VECTORIZE_SSE3
+//   Packet4f tmp = _mm_add_ps(a, vec4f_swizzle1(a,2,3,2,3));
+//   return pfirst<Packet4f>(_mm_hadd_ps(tmp, tmp));
+// #else
+  Packet4f tmp = _mm_add_ps(a, _mm_movehl_ps(a,a));
+  return pfirst<Packet4f>(_mm_add_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
+// #endif
+}
+
+template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a)
+{
+  // Disable SSE3 _mm_hadd_pd that is extremely slow on all existing Intel's architectures
+  // (from Nehalem to Haswell)
+// #ifdef EIGEN_VECTORIZE_SSE3
+//   return pfirst<Packet2d>(_mm_hadd_pd(a, a));
+// #else
+  return pfirst<Packet2d>(_mm_add_sd(a, _mm_unpackhi_pd(a,a)));
+// #endif
+}
+
+#ifdef EIGEN_VECTORIZE_SSSE3
+template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
+{
+  return _mm_hadd_epi32(_mm_hadd_epi32(vecs[0], vecs[1]),_mm_hadd_epi32(vecs[2], vecs[3]));
+}
+template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
+{
+  Packet4i tmp0 = _mm_hadd_epi32(a,a);
+  return pfirst<Packet4i>(_mm_hadd_epi32(tmp0,tmp0));
+}
+#else
 template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
 {
   Packet4i tmp = _mm_add_epi32(a, _mm_unpackhi_epi64(a,a));
-  return pfirst(tmp) + pfirst(_mm_shuffle_epi32(tmp, 1));
+  return pfirst(tmp) + pfirst<Packet4i>(_mm_shuffle_epi32(tmp, 1));
 }
 
 template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
@@ -433,18 +637,18 @@ template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
   tmp0 = _mm_unpackhi_epi64(tmp0, tmp1);
   return _mm_add_epi32(tmp0, tmp2);
 }
-
+#endif
 // Other reduction functions:
 
 // mul
 template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
 {
   Packet4f tmp = _mm_mul_ps(a, _mm_movehl_ps(a,a));
-  return pfirst(_mm_mul_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
+  return pfirst<Packet4f>(_mm_mul_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
 }
 template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a)
 {
-  return pfirst(_mm_mul_sd(a, _mm_unpackhi_pd(a,a)));
+  return pfirst<Packet2d>(_mm_mul_sd(a, _mm_unpackhi_pd(a,a)));
 }
 template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a)
 {
@@ -460,14 +664,18 @@ template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a)
 template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a)
 {
   Packet4f tmp = _mm_min_ps(a, _mm_movehl_ps(a,a));
-  return pfirst(_mm_min_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
+  return pfirst<Packet4f>(_mm_min_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
 }
 template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a)
 {
-  return pfirst(_mm_min_sd(a, _mm_unpackhi_pd(a,a)));
+  return pfirst<Packet2d>(_mm_min_sd(a, _mm_unpackhi_pd(a,a)));
 }
 template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a)
 {
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  Packet4i tmp = _mm_min_epi32(a, _mm_shuffle_epi32(a, _MM_SHUFFLE(0,0,3,2)));
+  return pfirst<Packet4i>(_mm_min_epi32(tmp,_mm_shuffle_epi32(tmp, 1)));
+#else
   // after some experiments, it is seems this is the fastest way to implement it
   // for GCC (eg., it does not like using std::min after the pstore !!)
   EIGEN_ALIGN16 int aux[4];
@@ -475,20 +683,25 @@ template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a)
   int aux0 = aux[0]<aux[1] ? aux[0] : aux[1];
   int aux2 = aux[2]<aux[3] ? aux[2] : aux[3];
   return aux0<aux2 ? aux0 : aux2;
+#endif // EIGEN_VECTORIZE_SSE4_1
 }
 
 // max
 template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a)
 {
   Packet4f tmp = _mm_max_ps(a, _mm_movehl_ps(a,a));
-  return pfirst(_mm_max_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
+  return pfirst<Packet4f>(_mm_max_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
 }
 template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a)
 {
-  return pfirst(_mm_max_sd(a, _mm_unpackhi_pd(a,a)));
+  return pfirst<Packet2d>(_mm_max_sd(a, _mm_unpackhi_pd(a,a)));
 }
 template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a)
 {
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  Packet4i tmp = _mm_max_epi32(a, _mm_shuffle_epi32(a, _MM_SHUFFLE(0,0,3,2)));
+  return pfirst<Packet4i>(_mm_max_epi32(tmp,_mm_shuffle_epi32(tmp, 1)));
+#else
   // after some experiments, it is seems this is the fastest way to implement it
   // for GCC (eg., it does not like using std::min after the pstore !!)
   EIGEN_ALIGN16 int aux[4];
@@ -496,9 +709,10 @@ template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a)
   int aux0 = aux[0]>aux[1] ? aux[0] : aux[1];
   int aux2 = aux[2]>aux[3] ? aux[2] : aux[3];
   return aux0>aux2 ? aux0 : aux2;
+#endif // EIGEN_VECTORIZE_SSE4_1
 }
 
-#if (defined __GNUC__)
+#if EIGEN_COMP_GNUC
 // template <> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f&  a, const Packet4f&  b, const Packet4f&  c)
 // {
 //   Packet4f res = b;
@@ -606,6 +820,110 @@ struct palign_impl<Offset,Packet2d>
 };
 #endif
 
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet4f,4>& kernel) {
+  _MM_TRANSPOSE4_PS(kernel.packet[0], kernel.packet[1], kernel.packet[2], kernel.packet[3]);
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet2d,2>& kernel) {
+  __m128d tmp = _mm_unpackhi_pd(kernel.packet[0], kernel.packet[1]);
+  kernel.packet[0] = _mm_unpacklo_pd(kernel.packet[0], kernel.packet[1]);
+  kernel.packet[1] = tmp;
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet4i,4>& kernel) {
+  __m128i T0 = _mm_unpacklo_epi32(kernel.packet[0], kernel.packet[1]);
+  __m128i T1 = _mm_unpacklo_epi32(kernel.packet[2], kernel.packet[3]);
+  __m128i T2 = _mm_unpackhi_epi32(kernel.packet[0], kernel.packet[1]);
+  __m128i T3 = _mm_unpackhi_epi32(kernel.packet[2], kernel.packet[3]);
+
+  kernel.packet[0] = _mm_unpacklo_epi64(T0, T1);
+  kernel.packet[1] = _mm_unpackhi_epi64(T0, T1);
+  kernel.packet[2] = _mm_unpacklo_epi64(T2, T3);
+  kernel.packet[3] = _mm_unpackhi_epi64(T2, T3);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pblend(const Selector<4>& ifPacket, const Packet4i& thenPacket, const Packet4i& elsePacket) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i select = _mm_set_epi32(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]);
+  __m128i false_mask = _mm_cmpeq_epi32(select, zero);
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  return _mm_blendv_epi8(thenPacket, elsePacket, false_mask);
+#else
+  return _mm_or_si128(_mm_andnot_si128(false_mask, thenPacket), _mm_and_si128(false_mask, elsePacket));
+#endif
+}
+template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket, const Packet4f& elsePacket) {
+  const __m128 zero = _mm_setzero_ps();
+  const __m128 select = _mm_set_ps(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]);
+  __m128 false_mask = _mm_cmpeq_ps(select, zero);
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  return _mm_blendv_ps(thenPacket, elsePacket, false_mask);
+#else
+  return _mm_or_ps(_mm_andnot_ps(false_mask, thenPacket), _mm_and_ps(false_mask, elsePacket));
+#endif
+}
+template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket, const Packet2d& elsePacket) {
+  const __m128d zero = _mm_setzero_pd();
+  const __m128d select = _mm_set_pd(ifPacket.select[1], ifPacket.select[0]);
+  __m128d false_mask = _mm_cmpeq_pd(select, zero);
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  return _mm_blendv_pd(thenPacket, elsePacket, false_mask);
+#else
+  return _mm_or_pd(_mm_andnot_pd(false_mask, thenPacket), _mm_and_pd(false_mask, elsePacket));
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pinsertfirst(const Packet4f& a, float b)
+{
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  return _mm_blend_ps(a,pset1<Packet4f>(b),1);
+#else
+  return _mm_move_ss(a, _mm_load_ss(&b));
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d pinsertfirst(const Packet2d& a, double b)
+{
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  return _mm_blend_pd(a,pset1<Packet2d>(b),1);
+#else
+  return _mm_move_sd(a, _mm_load_sd(&b));
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pinsertlast(const Packet4f& a, float b)
+{
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  return _mm_blend_ps(a,pset1<Packet4f>(b),(1<<3));
+#else
+  const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x0,0x0,0x0,0xFFFFFFFF));
+  return _mm_or_ps(_mm_andnot_ps(mask, a), _mm_and_ps(mask, pset1<Packet4f>(b)));
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d pinsertlast(const Packet2d& a, double b)
+{
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  return _mm_blend_pd(a,pset1<Packet2d>(b),(1<<1));
+#else
+  const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0x0,0x0,0xFFFFFFFF,0xFFFFFFFF));
+  return _mm_or_pd(_mm_andnot_pd(mask, a), _mm_and_pd(mask, pset1<Packet2d>(b)));
+#endif
+}
+
+// Scalar path for pmadd with FMA to ensure consistency with vectorized path.
+#ifdef __FMA__
+template<> EIGEN_STRONG_INLINE float pmadd(const float& a, const float& b, const float& c) {
+  return ::fmaf(a,b,c);
+}
+template<> EIGEN_STRONG_INLINE double pmadd(const double& a, const double& b, const double& c) {
+  return ::fma(a,b,c);
+}
+#endif
+
 } // end namespace internal
 
 } // end namespace Eigen
diff --git a/eigen/Eigen/src/Core/arch/SSE/TypeCasting.h b/eigen/Eigen/src/Core/arch/SSE/TypeCasting.h
new file mode 100644
index 0000000..c848932
--- /dev/null
+++ b/eigen/Eigen/src/Core/arch/SSE/TypeCasting.h
@@ -0,0 +1,77 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_TYPE_CASTING_SSE_H
+#define EIGEN_TYPE_CASTING_SSE_H
+
+namespace Eigen {
+
+namespace internal {
+
+template <>
+struct type_casting_traits<float, int> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet4i pcast<Packet4f, Packet4i>(const Packet4f& a) {
+  return _mm_cvttps_epi32(a);
+}
+
+
+template <>
+struct type_casting_traits<int, float> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet4f pcast<Packet4i, Packet4f>(const Packet4i& a) {
+  return _mm_cvtepi32_ps(a);
+}
+
+
+template <>
+struct type_casting_traits<double, float> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 2,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet4f pcast<Packet2d, Packet4f>(const Packet2d& a, const Packet2d& b) {
+  return _mm_shuffle_ps(_mm_cvtpd_ps(a), _mm_cvtpd_ps(b), (1 << 2) | (1 << 6));
+}
+
+template <>
+struct type_casting_traits<float, double> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 2
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet2d pcast<Packet4f, Packet2d>(const Packet4f& a) {
+  // Simply discard the second half of the input
+  return _mm_cvtps_pd(a);
+}
+
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TYPE_CASTING_SSE_H
diff --git a/eigen/Eigen/src/Core/arch/ZVector/Complex.h b/eigen/Eigen/src/Core/arch/ZVector/Complex.h
new file mode 100644
index 0000000..d39d2d1
--- /dev/null
+++ b/eigen/Eigen/src/Core/arch/ZVector/Complex.h
@@ -0,0 +1,394 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2016 Konstantinos Margaritis <markos@freevec.org>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_COMPLEX32_ALTIVEC_H
+#define EIGEN_COMPLEX32_ALTIVEC_H
+
+namespace Eigen {
+
+namespace internal {
+
+static Packet2ul  p2ul_CONJ_XOR1 = (Packet2ul) vec_sld((Packet4ui) p2d_ZERO_, (Packet4ui) p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 };
+static Packet2ul  p2ul_CONJ_XOR2 = (Packet2ul) vec_sld((Packet4ui) p2l_ZERO,  (Packet4ui) p2d_ZERO_, 8);//{ 0x8000000000000000, 0x0000000000000000 };
+
+struct Packet1cd
+{
+  EIGEN_STRONG_INLINE Packet1cd() {}
+  EIGEN_STRONG_INLINE explicit Packet1cd(const Packet2d& a) : v(a) {}
+  Packet2d v;
+};
+
+struct Packet2cf
+{
+  EIGEN_STRONG_INLINE Packet2cf() {}
+  EIGEN_STRONG_INLINE explicit Packet2cf(const Packet4f& a) : v(a) {}
+  union {
+    Packet4f v;
+    Packet1cd cd[2];
+  };
+};
+
+template<> struct packet_traits<std::complex<float> >  : default_packet_traits
+{
+  typedef Packet2cf type;
+  typedef Packet2cf half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 2,
+    HasHalfPacket = 0,
+
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasDiv    = 1,
+    HasNegate = 1,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasBlend  = 1,
+    HasSetLinear = 0
+  };
+};
+
+
+template<> struct packet_traits<std::complex<double> >  : default_packet_traits
+{
+  typedef Packet1cd type;
+  typedef Packet1cd half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 1,
+    HasHalfPacket = 0,
+
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasDiv    = 1,
+    HasNegate = 1,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasSetLinear = 0
+  };
+};
+
+template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float>  type; enum {size=2, alignment=Aligned16}; typedef Packet2cf half; };
+template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16}; typedef Packet1cd half; };
+
+/* Forward declaration */
+EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2cf,2>& kernel);
+
+template<> EIGEN_STRONG_INLINE Packet2cf pload <Packet2cf>(const std::complex<float>* from)  { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>((const float*)from)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pload <Packet1cd>(const std::complex<double>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(pload<Packet2d>((const double*)from)); }
+template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from)  { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>((const float*)from)); }
+template<> EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu<Packet2d>((const double*)from)); }
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> *     to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((float*)to, from.v); }
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> *     to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((float*)to, from.v); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); }
+
+template<> EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<double>&  from)
+{ /* here we really have to use unaligned loads :( */ return ploadu<Packet1cd>(&from); }
+
+template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>&  from)
+{
+  Packet2cf res;
+  res.cd[0] = Packet1cd(vec_ld2f((const float *)&from));
+  res.cd[1] = res.cd[0];
+  return res;
+}
+template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from, Index stride)
+{
+  std::complex<float> EIGEN_ALIGN16 af[2];
+  af[0] = from[0*stride];
+  af[1] = from[1*stride];
+  return pload<Packet2cf>(af);
+}
+template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather<std::complex<double>, Packet1cd>(const std::complex<double>* from, Index stride EIGEN_UNUSED)
+{
+  return pload<Packet1cd>(from);
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, Index stride)
+{
+  std::complex<float> EIGEN_ALIGN16 af[2];
+  pstore<std::complex<float> >((std::complex<float> *) af, from);
+  to[0*stride] = af[0];
+  to[1*stride] = af[1];
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet1cd>(std::complex<double>* to, const Packet1cd& from, Index stride EIGEN_UNUSED)
+{
+  pstore<std::complex<double> >(to, from);
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(padd<Packet4f>(a.v, b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(a.v + b.v); }
+template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(psub<Packet4f>(a.v, b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(a.v - b.v); }
+template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(Packet2d(a.v))); }
+template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) { return Packet2cf(pnegate(Packet4f(a.v))); }
+template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) { return Packet1cd((Packet2d)vec_xor((Packet2d)a.v, (Packet2d)p2ul_CONJ_XOR2)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a)
+{
+  Packet2cf res;
+  res.v.v4f[0] = pconj(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[0]))).v;
+  res.v.v4f[1] = pconj(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[1]))).v;
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
+{
+  Packet2d a_re, a_im, v1, v2;
+
+  // Permute and multiply the real parts of a and b
+  a_re = vec_perm(a.v, a.v, p16uc_PSET64_HI);
+  // Get the imaginary parts of a
+  a_im = vec_perm(a.v, a.v, p16uc_PSET64_LO);
+  // multiply a_re * b
+  v1 = vec_madd(a_re, b.v, p2d_ZERO);
+  // multiply a_im * b and get the conjugate result
+  v2 = vec_madd(a_im, b.v, p2d_ZERO);
+  v2 = (Packet2d) vec_sld((Packet4ui)v2, (Packet4ui)v2, 8);
+  v2 = (Packet2d) vec_xor((Packet2d)v2, (Packet2d) p2ul_CONJ_XOR1);
+
+  return Packet1cd(v1 + v2);
+}
+template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
+{
+  Packet2cf res;
+  res.v.v4f[0] = pmul(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[0])), Packet1cd(reinterpret_cast<Packet2d>(b.v.v4f[0]))).v;
+  res.v.v4f[1] = pmul(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[1])), Packet1cd(reinterpret_cast<Packet2d>(b.v.v4f[1]))).v;
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd pand   <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pand   <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pand<Packet4f>(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd por    <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_or(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf por    <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(por<Packet4f>(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pxor   <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_xor(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pxor   <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pxor<Packet4f>(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v, vec_nor(b.v,b.v))); }
+template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pandnot<Packet4f>(a.v,b.v)); }
+
+template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>*     from) {  return pset1<Packet1cd>(*from); }
+template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>*      from) {  return pset1<Packet2cf>(*from); }
+
+template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> *     addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
+template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> *   addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
+
+template<> EIGEN_STRONG_INLINE std::complex<double>  pfirst<Packet1cd>(const Packet1cd& a)
+{
+  std::complex<double> EIGEN_ALIGN16 res;
+  pstore<std::complex<double> >(&res, a);
+
+  return res;
+}
+template<> EIGEN_STRONG_INLINE std::complex<float>  pfirst<Packet2cf>(const Packet2cf& a)
+{
+  std::complex<float> EIGEN_ALIGN16 res[2];
+  pstore<std::complex<float> >(res, a);
+
+  return res[0];
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; }
+template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a)
+{
+  Packet2cf res;
+  res.cd[0] = a.cd[1];
+  res.cd[1] = a.cd[0];
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a)
+{
+  return pfirst(a);
+}
+template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a)
+{
+  std::complex<float> res;
+  Packet1cd b = padd<Packet1cd>(a.cd[0], a.cd[1]);
+  vec_st2f(b.v, (float*)&res);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs)
+{
+  return vecs[0];
+}
+template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs)
+{
+  PacketBlock<Packet2cf,2> transpose;
+  transpose.packet[0] = vecs[0];
+  transpose.packet[1] = vecs[1];
+  ptranspose(transpose);
+
+  return padd<Packet2cf>(transpose.packet[0], transpose.packet[1]);
+} 
+
+template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a)
+{
+  return pfirst(a);
+}
+template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a)
+{
+  std::complex<float> res;
+  Packet1cd b = pmul<Packet1cd>(a.cd[0], a.cd[1]);
+  vec_st2f(b.v, (float*)&res);
+  return res;
+}
+
+template<int Offset>
+struct palign_impl<Offset,Packet1cd>
+{
+  static EIGEN_STRONG_INLINE void run(Packet1cd& /*first*/, const Packet1cd& /*second*/)
+  {
+    // FIXME is it sure we never have to align a Packet1cd?
+    // Even though a std::complex<double> has 16 bytes, it is not necessarily aligned on a 16 bytes boundary...
+  }
+};
+
+template<int Offset>
+struct palign_impl<Offset,Packet2cf>
+{
+  static EIGEN_STRONG_INLINE void run(Packet2cf& first, const Packet2cf& second)
+  {
+    if (Offset == 1) {
+      first.cd[0] = first.cd[1];
+      first.cd[1] = second.cd[0];
+    }
+  }
+};
+
+template<> struct conj_helper<Packet1cd, Packet1cd, false,true>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
+  {
+    return internal::pmul(a, pconj(b));
+  }
+};
+
+template<> struct conj_helper<Packet1cd, Packet1cd, true,false>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
+  {
+    return internal::pmul(pconj(a), b);
+  }
+};
+
+template<> struct conj_helper<Packet1cd, Packet1cd, true,true>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
+  {
+    return pconj(internal::pmul(a, b));
+  }
+};
+
+template<> struct conj_helper<Packet2cf, Packet2cf, false,true>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
+  {
+    return internal::pmul(a, pconj(b));
+  }
+};
+
+template<> struct conj_helper<Packet2cf, Packet2cf, true,false>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
+  {
+    return internal::pmul(pconj(a), b);
+  }
+};
+
+template<> struct conj_helper<Packet2cf, Packet2cf, true,true>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
+  {
+    return pconj(internal::pmul(a, b));
+  }
+};
+
+template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
+{
+  // TODO optimize it for AltiVec
+  Packet1cd res = conj_helper<Packet1cd,Packet1cd,false,true>().pmul(a,b);
+  Packet2d s = vec_madd(b.v, b.v, p2d_ZERO_);
+  return Packet1cd(pdiv(res.v, s + vec_perm(s, s, p16uc_REVERSE64)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
+{
+  // TODO optimize it for AltiVec
+  Packet2cf res;
+  res.cd[0] = pdiv<Packet1cd>(a.cd[0], b.cd[0]);
+  res.cd[1] = pdiv<Packet1cd>(a.cd[1], b.cd[1]);
+  return res;
+}
+
+EIGEN_STRONG_INLINE Packet1cd pcplxflip/*<Packet1cd>*/(const Packet1cd& x)
+{
+  return Packet1cd(preverse(Packet2d(x.v)));
+}
+
+EIGEN_STRONG_INLINE Packet2cf pcplxflip/*<Packet2cf>*/(const Packet2cf& x)
+{
+  Packet2cf res;
+  res.cd[0] = pcplxflip(x.cd[0]);
+  res.cd[1] = pcplxflip(x.cd[1]);
+  return res;
+}
+
+EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet1cd,2>& kernel)
+{
+  Packet2d tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_HI);
+  kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO);
+  kernel.packet[0].v = tmp;
+}
+
+EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2cf,2>& kernel)
+{
+  Packet1cd tmp = kernel.packet[0].cd[1];
+  kernel.packet[0].cd[1] = kernel.packet[1].cd[0];
+  kernel.packet[1].cd[0] = tmp;
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, const Packet2cf& thenPacket, const Packet2cf& elsePacket) {
+  Packet2cf result;
+  const Selector<4> ifPacket4 = { ifPacket.select[0], ifPacket.select[0], ifPacket.select[1], ifPacket.select[1] };
+  result.v = pblend<Packet4f>(ifPacket4, thenPacket.v, elsePacket.v);
+  return result;
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_COMPLEX32_ALTIVEC_H
diff --git a/eigen/Eigen/src/Core/arch/ZVector/MathFunctions.h b/eigen/Eigen/src/Core/arch/ZVector/MathFunctions.h
new file mode 100644
index 0000000..5c7aa72
--- /dev/null
+++ b/eigen/Eigen/src/Core/arch/ZVector/MathFunctions.h
@@ -0,0 +1,137 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2007 Julien Pommier
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2016 Konstantinos Margaritis <markos@freevec.org>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/* The sin, cos, exp, and log functions of this file come from
+ * Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/
+ */
+
+#ifndef EIGEN_MATH_FUNCTIONS_ALTIVEC_H
+#define EIGEN_MATH_FUNCTIONS_ALTIVEC_H
+
+namespace Eigen {
+
+namespace internal {
+
+static _EIGEN_DECLARE_CONST_Packet2d(1 , 1.0);
+static _EIGEN_DECLARE_CONST_Packet2d(2 , 2.0);
+static _EIGEN_DECLARE_CONST_Packet2d(half, 0.5);
+
+static _EIGEN_DECLARE_CONST_Packet2d(exp_hi,  709.437);
+static _EIGEN_DECLARE_CONST_Packet2d(exp_lo, -709.436139303);
+
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_LOG2EF, 1.4426950408889634073599);
+
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p0, 1.26177193074810590878e-4);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p1, 3.02994407707441961300e-2);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p2, 9.99999999999999999910e-1);
+
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q0, 3.00198505138664455042e-6);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q1, 2.52448340349684104192e-3);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q2, 2.27265548208155028766e-1);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q3, 2.00000000000000000009e0);
+
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C1, 0.693145751953125);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6);
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet2d pexp<Packet2d>(const Packet2d& _x)
+{
+  Packet2d x = _x;
+
+  Packet2d tmp, fx;
+  Packet2l emm0;
+
+  // clamp x
+  x = pmax(pmin(x, p2d_exp_hi), p2d_exp_lo);
+  /* express exp(x) as exp(g + n*log(2)) */
+  fx = pmadd(p2d_cephes_LOG2EF, x, p2d_half);
+
+  fx = vec_floor(fx);
+
+  tmp = pmul(fx, p2d_cephes_exp_C1);
+  Packet2d z = pmul(fx, p2d_cephes_exp_C2);
+  x = psub(x, tmp);
+  x = psub(x, z);
+
+  Packet2d x2 = pmul(x,x);
+
+  Packet2d px = p2d_cephes_exp_p0;
+  px = pmadd(px, x2, p2d_cephes_exp_p1);
+  px = pmadd(px, x2, p2d_cephes_exp_p2);
+  px = pmul (px, x);
+
+  Packet2d qx = p2d_cephes_exp_q0;
+  qx = pmadd(qx, x2, p2d_cephes_exp_q1);
+  qx = pmadd(qx, x2, p2d_cephes_exp_q2);
+  qx = pmadd(qx, x2, p2d_cephes_exp_q3);
+
+  x = pdiv(px,psub(qx,px));
+  x = pmadd(p2d_2,x,p2d_1);
+
+  // build 2^n
+  emm0 = vec_ctsl(fx, 0);
+
+  static const Packet2l p2l_1023 = { 1023, 1023 };
+  static const Packet2ul p2ul_52 = { 52, 52 };
+
+  emm0 = emm0 + p2l_1023;
+  emm0 = emm0 << reinterpret_cast<Packet2l>(p2ul_52);
+
+  // Altivec's max & min operators just drop silent NaNs. Check NaNs in 
+  // inputs and return them unmodified.
+  Packet2ul isnumber_mask = reinterpret_cast<Packet2ul>(vec_cmpeq(_x, _x));
+  return vec_sel(_x, pmax(pmul(x, reinterpret_cast<Packet2d>(emm0)), _x),
+                 isnumber_mask);
+}
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f pexp<Packet4f>(const Packet4f& x)
+{
+  Packet4f res;
+  res.v4f[0] = pexp<Packet2d>(x.v4f[0]);
+  res.v4f[1] = pexp<Packet2d>(x.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet2d psqrt<Packet2d>(const Packet2d& x)
+{
+  return  __builtin_s390_vfsqdb(x);
+}
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f psqrt<Packet4f>(const Packet4f& x)
+{
+  Packet4f res;
+  res.v4f[0] = psqrt<Packet2d>(x.v4f[0]);
+  res.v4f[1] = psqrt<Packet2d>(x.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet2d prsqrt<Packet2d>(const Packet2d& x) {
+  // Unfortunately we can't use the much faster mm_rqsrt_pd since it only provides an approximation.
+  return pset1<Packet2d>(1.0) / psqrt<Packet2d>(x);
+}
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f prsqrt<Packet4f>(const Packet4f& x) {
+  Packet4f res;
+  res.v4f[0] = prsqrt<Packet2d>(x.v4f[0]);
+  res.v4f[1] = prsqrt<Packet2d>(x.v4f[1]);
+  return res;
+}
+
+}  // end namespace internal
+
+}  // end namespace Eigen
+
+#endif  // EIGEN_MATH_FUNCTIONS_ALTIVEC_H
diff --git a/eigen/Eigen/src/Core/arch/ZVector/PacketMath.h b/eigen/Eigen/src/Core/arch/ZVector/PacketMath.h
new file mode 100644
index 0000000..57b01fc
--- /dev/null
+++ b/eigen/Eigen/src/Core/arch/ZVector/PacketMath.h
@@ -0,0 +1,945 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Konstantinos Margaritis <markos@freevec.org>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PACKET_MATH_ZVECTOR_H
+#define EIGEN_PACKET_MATH_ZVECTOR_H
+
+#include <stdint.h>
+
+namespace Eigen {
+
+namespace internal {
+
+#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
+#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 4
+#endif
+
+#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#endif
+
+#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD
+#define EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD
+#endif
+
+#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
+#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS  16
+#endif
+
+typedef __vector int                 Packet4i;
+typedef __vector unsigned int        Packet4ui;
+typedef __vector __bool int          Packet4bi;
+typedef __vector short int           Packet8i;
+typedef __vector unsigned char       Packet16uc;
+typedef __vector double              Packet2d;
+typedef __vector unsigned long long  Packet2ul;
+typedef __vector long long           Packet2l;
+
+typedef struct {
+	Packet2d  v4f[2];
+} Packet4f;
+
+typedef union {
+  int32_t   i[4];
+  uint32_t ui[4];
+  int64_t   l[2];
+  uint64_t ul[2];
+  double    d[2];
+  Packet4i  v4i;
+  Packet4ui v4ui;
+  Packet2l  v2l;
+  Packet2ul v2ul;
+  Packet2d  v2d;
+} Packet;
+
+// We don't want to write the same code all the time, but we need to reuse the constants
+// and it doesn't really work to declare them global, so we define macros instead
+
+#define _EIGEN_DECLARE_CONST_FAST_Packet4i(NAME,X) \
+  Packet4i p4i_##NAME = reinterpret_cast<Packet4i>(vec_splat_s32(X))
+
+#define _EIGEN_DECLARE_CONST_FAST_Packet2d(NAME,X) \
+  Packet2d p2d_##NAME = reinterpret_cast<Packet2d>(vec_splat_s64(X))
+
+#define _EIGEN_DECLARE_CONST_FAST_Packet2l(NAME,X) \
+  Packet2l p2l_##NAME = reinterpret_cast<Packet2l>(vec_splat_s64(X))
+
+#define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
+  Packet4i p4i_##NAME = pset1<Packet4i>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet2d(NAME,X) \
+  Packet2d p2d_##NAME = pset1<Packet2d>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet2l(NAME,X) \
+  Packet2l p2l_##NAME = pset1<Packet2l>(X)
+
+// These constants are endian-agnostic
+//static _EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0); //{ 0, 0, 0, 0,}
+static _EIGEN_DECLARE_CONST_FAST_Packet4i(ONE, 1); //{ 1, 1, 1, 1}
+
+static _EIGEN_DECLARE_CONST_FAST_Packet2d(ZERO, 0);
+static _EIGEN_DECLARE_CONST_FAST_Packet2l(ZERO, 0);
+static _EIGEN_DECLARE_CONST_FAST_Packet2l(ONE, 1);
+
+static Packet2d p2d_ONE = { 1.0, 1.0 }; 
+static Packet2d p2d_ZERO_ = { -0.0, -0.0 };
+
+static Packet4i p4i_COUNTDOWN = { 0, 1, 2, 3 };
+static Packet4f p4f_COUNTDOWN = { 0.0, 1.0, 2.0, 3.0 };
+static Packet2d p2d_COUNTDOWN = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet16uc>(p2d_ZERO), reinterpret_cast<Packet16uc>(p2d_ONE), 8));
+
+static Packet16uc p16uc_PSET64_HI = { 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 };
+static Packet16uc p16uc_DUPLICATE32_HI = { 0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7 };
+
+// Mask alignment
+#define _EIGEN_MASK_ALIGNMENT	0xfffffffffffffff0
+
+#define _EIGEN_ALIGNED_PTR(x)	((std::ptrdiff_t)(x) & _EIGEN_MASK_ALIGNMENT)
+
+// Handle endianness properly while loading constants
+// Define global static constants:
+
+static Packet16uc p16uc_FORWARD =   { 0,1,2,3, 4,5,6,7, 8,9,10,11, 12,13,14,15 };
+static Packet16uc p16uc_REVERSE32 = { 12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3 };
+static Packet16uc p16uc_REVERSE64 = { 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
+
+static Packet16uc p16uc_PSET32_WODD   = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 };
+static Packet16uc p16uc_PSET32_WEVEN  = vec_sld(p16uc_DUPLICATE32_HI, (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 };
+/*static Packet16uc p16uc_HALF64_0_16 = vec_sld((Packet16uc)p4i_ZERO, vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 3), 8);      //{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16};
+
+static Packet16uc p16uc_PSET64_HI = (Packet16uc) vec_mergeh((Packet4ui)p16uc_PSET32_WODD, (Packet4ui)p16uc_PSET32_WEVEN);     //{ 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 };*/
+static Packet16uc p16uc_PSET64_LO = (Packet16uc) vec_mergel((Packet4ui)p16uc_PSET32_WODD, (Packet4ui)p16uc_PSET32_WEVEN);     //{ 8,9,10,11, 12,13,14,15, 8,9,10,11, 12,13,14,15 };
+/*static Packet16uc p16uc_TRANSPOSE64_HI = vec_add(p16uc_PSET64_HI, p16uc_HALF64_0_16);                                         //{ 0,1,2,3, 4,5,6,7, 16,17,18,19, 20,21,22,23};
+static Packet16uc p16uc_TRANSPOSE64_LO = vec_add(p16uc_PSET64_LO, p16uc_HALF64_0_16);                                         //{ 8,9,10,11, 12,13,14,15, 24,25,26,27, 28,29,30,31};*/
+static Packet16uc p16uc_TRANSPOSE64_HI = { 0,1,2,3, 4,5,6,7, 16,17,18,19, 20,21,22,23};
+static Packet16uc p16uc_TRANSPOSE64_LO = { 8,9,10,11, 12,13,14,15, 24,25,26,27, 28,29,30,31};
+
+//static Packet16uc p16uc_COMPLEX32_REV = vec_sld(p16uc_REVERSE32, p16uc_REVERSE32, 8);                                         //{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 };
+
+//static Packet16uc p16uc_COMPLEX32_REV2 = vec_sld(p16uc_FORWARD, p16uc_FORWARD, 8);                                            //{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
+
+
+#if EIGEN_HAS_BUILTIN(__builtin_prefetch) || EIGEN_COMP_GNUC
+  #define EIGEN_ZVECTOR_PREFETCH(ADDR) __builtin_prefetch(ADDR);
+#else
+  #define EIGEN_ZVECTOR_PREFETCH(ADDR) asm( "   pfd [%[addr]]\n" :: [addr] "r" (ADDR) : "cc" );
+#endif
+
+template<> struct packet_traits<int>    : default_packet_traits
+{
+  typedef Packet4i type;
+  typedef Packet4i half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 4,
+    HasHalfPacket = 0,
+
+    HasAdd  = 1,
+    HasSub  = 1,
+    HasMul  = 1,
+    HasDiv  = 1,
+    HasBlend = 1
+  };
+};
+
+template<> struct packet_traits<float> : default_packet_traits
+{
+  typedef Packet4f type;
+  typedef Packet4f half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=4,
+    HasHalfPacket = 0,
+
+    HasAdd  = 1,
+    HasSub  = 1,
+    HasMul  = 1,
+    HasDiv  = 1,
+    HasMin  = 1,
+    HasMax  = 1,
+    HasAbs  = 1,
+    HasSin  = 0,
+    HasCos  = 0,
+    HasLog  = 0,
+    HasExp  = 1,
+    HasSqrt = 1,
+    HasRsqrt = 1,
+    HasRound = 1,
+    HasFloor = 1,
+    HasCeil = 1,
+    HasNegate = 1,
+    HasBlend = 1
+  };
+};
+
+template<> struct packet_traits<double> : default_packet_traits
+{
+  typedef Packet2d type;
+  typedef Packet2d half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=2,
+    HasHalfPacket = 1,
+
+    HasAdd  = 1,
+    HasSub  = 1,
+    HasMul  = 1,
+    HasDiv  = 1,
+    HasMin  = 1,
+    HasMax  = 1,
+    HasAbs  = 1,
+    HasSin  = 0,
+    HasCos  = 0,
+    HasLog  = 0,
+    HasExp  = 1,
+    HasSqrt = 1,
+    HasRsqrt = 1,
+    HasRound = 1,
+    HasFloor = 1,
+    HasCeil = 1,
+    HasNegate = 1,
+    HasBlend = 1
+  };
+};
+
+template<> struct unpacket_traits<Packet4i> { typedef int    type; enum {size=4, alignment=Aligned16}; typedef Packet4i half; };
+template<> struct unpacket_traits<Packet4f> { typedef float  type; enum {size=4, alignment=Aligned16}; typedef Packet4f half; };
+template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16}; typedef Packet2d half; };
+
+/* Forward declaration */
+EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4f,4>& kernel);
+ 
+inline std::ostream & operator <<(std::ostream & s, const Packet4i & v)
+{
+  Packet vt;
+  vt.v4i = v;
+  s << vt.i[0] << ", " << vt.i[1] << ", " << vt.i[2] << ", " << vt.i[3];
+  return s;
+}
+
+inline std::ostream & operator <<(std::ostream & s, const Packet4ui & v)
+{
+  Packet vt;
+  vt.v4ui = v;
+  s << vt.ui[0] << ", " << vt.ui[1] << ", " << vt.ui[2] << ", " << vt.ui[3];
+  return s;
+}
+
+inline std::ostream & operator <<(std::ostream & s, const Packet2l & v)
+{
+  Packet vt;
+  vt.v2l = v;
+  s << vt.l[0] << ", " << vt.l[1];
+  return s;
+}
+
+inline std::ostream & operator <<(std::ostream & s, const Packet2ul & v)
+{
+  Packet vt;
+  vt.v2ul = v;
+  s << vt.ul[0] << ", " << vt.ul[1] ;
+  return s;
+}
+
+inline std::ostream & operator <<(std::ostream & s, const Packet2d & v)
+{
+  Packet vt;
+  vt.v2d = v;
+  s << vt.d[0] << ", " << vt.d[1];
+  return s;
+}
+
+/* Helper function to simulate a vec_splat_packet4f
+ */
+template<int element> EIGEN_STRONG_INLINE Packet4f vec_splat_packet4f(const Packet4f&   from)
+{
+  Packet4f splat;
+  switch (element) {
+  case 0:
+    splat.v4f[0] = vec_splat(from.v4f[0], 0);
+    splat.v4f[1] = splat.v4f[0];
+    break;
+  case 1:
+    splat.v4f[0] = vec_splat(from.v4f[0], 1);
+    splat.v4f[1] = splat.v4f[0];
+    break;
+  case 2:
+    splat.v4f[0] = vec_splat(from.v4f[1], 0);
+    splat.v4f[1] = splat.v4f[0];
+    break;
+  case 3:
+    splat.v4f[0] = vec_splat(from.v4f[1], 1);
+    splat.v4f[1] = splat.v4f[0];
+    break;
+  }
+  return splat;
+}
+
+template<int Offset>
+struct palign_impl<Offset,Packet4i>
+{
+  static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second)
+  {
+    switch (Offset % 4) {
+    case 1:
+      first = vec_sld(first, second, 4); break;
+    case 2:
+      first = vec_sld(first, second, 8); break;
+    case 3:
+      first = vec_sld(first, second, 12); break;
+    }
+  }
+};
+
+/* This is a tricky one, we have to translate float alignment to vector elements of sizeof double
+ */
+template<int Offset>
+struct palign_impl<Offset,Packet4f>
+{
+  static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second)
+  {
+    switch (Offset % 4) {
+    case 1:
+      first.v4f[0] = vec_sld(first.v4f[0], first.v4f[1], 8);
+      first.v4f[1] = vec_sld(first.v4f[1], second.v4f[0], 8);
+      break;
+    case 2:
+      first.v4f[0] = first.v4f[1];
+      first.v4f[1] = second.v4f[0];
+      break;
+    case 3:
+      first.v4f[0] = vec_sld(first.v4f[1],  second.v4f[0], 8);
+      first.v4f[1] = vec_sld(second.v4f[0], second.v4f[1], 8);
+      break;
+    }
+  }
+};
+
+
+template<int Offset>
+struct palign_impl<Offset,Packet2d>
+{
+  static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second)
+  {
+    if (Offset == 1)
+      first = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(first), reinterpret_cast<Packet4i>(second), 8));
+  }
+};
+
+template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int*     from)
+{
+  // FIXME: No intrinsic yet
+  EIGEN_DEBUG_ALIGNED_LOAD
+  Packet *vfrom;
+  vfrom = (Packet *) from;
+  return vfrom->v4i;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float*   from)
+{
+  // FIXME: No intrinsic yet
+  EIGEN_DEBUG_ALIGNED_LOAD
+  Packet4f vfrom;
+  vfrom.v4f[0] = vec_ld2f(&from[0]);
+  vfrom.v4f[1] = vec_ld2f(&from[2]);
+  return vfrom;
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from)
+{
+  // FIXME: No intrinsic yet
+  EIGEN_DEBUG_ALIGNED_LOAD
+  Packet *vfrom;
+  vfrom = (Packet *) from;
+  return vfrom->v2d;
+}
+
+template<> EIGEN_STRONG_INLINE void pstore<int>(int*       to, const Packet4i& from)
+{
+  // FIXME: No intrinsic yet
+  EIGEN_DEBUG_ALIGNED_STORE
+  Packet *vto;
+  vto = (Packet *) to;
+  vto->v4i = from;
+}
+
+template<> EIGEN_STRONG_INLINE void pstore<float>(float*   to, const Packet4f& from)
+{
+  // FIXME: No intrinsic yet
+  EIGEN_DEBUG_ALIGNED_STORE
+  vec_st2f(from.v4f[0], &to[0]);
+  vec_st2f(from.v4f[1], &to[2]);
+}
+
+
+template<> EIGEN_STRONG_INLINE void pstore<double>(double*   to, const Packet2d& from)
+{
+  // FIXME: No intrinsic yet
+  EIGEN_DEBUG_ALIGNED_STORE
+  Packet *vto;
+  vto = (Packet *) to;
+  vto->v2d = from;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int&    from)
+{
+  return vec_splats(from);
+}
+template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) {
+  return vec_splats(from);
+}
+template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float&    from)
+{
+  Packet4f to;
+  to.v4f[0] = pset1<Packet2d>(static_cast<const double&>(from));
+  to.v4f[1] = to.v4f[0];
+  return to;
+}
+
+template<> EIGEN_STRONG_INLINE void
+pbroadcast4<Packet4i>(const int *a,
+                      Packet4i& a0, Packet4i& a1, Packet4i& a2, Packet4i& a3)
+{
+  a3 = pload<Packet4i>(a);
+  a0 = vec_splat(a3, 0);
+  a1 = vec_splat(a3, 1);
+  a2 = vec_splat(a3, 2);
+  a3 = vec_splat(a3, 3);
+}
+
+template<> EIGEN_STRONG_INLINE void
+pbroadcast4<Packet4f>(const float *a,
+                      Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3)
+{
+  a3 = pload<Packet4f>(a);
+  a0 = vec_splat_packet4f<0>(a3);
+  a1 = vec_splat_packet4f<1>(a3);
+  a2 = vec_splat_packet4f<2>(a3);
+  a3 = vec_splat_packet4f<3>(a3);
+}
+
+template<> EIGEN_STRONG_INLINE void
+pbroadcast4<Packet2d>(const double *a,
+                      Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3)
+{
+  a1 = pload<Packet2d>(a);
+  a0 = vec_splat(a1, 0);
+  a1 = vec_splat(a1, 1);
+  a3 = pload<Packet2d>(a+2);
+  a2 = vec_splat(a3, 0);
+  a3 = vec_splat(a3, 1);
+}
+
+template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* from, Index stride)
+{
+  int EIGEN_ALIGN16 ai[4];
+  ai[0] = from[0*stride];
+  ai[1] = from[1*stride];
+  ai[2] = from[2*stride];
+  ai[3] = from[3*stride];
+ return pload<Packet4i>(ai);
+}
+
+template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride)
+{
+  float EIGEN_ALIGN16 ai[4];
+  ai[0] = from[0*stride];
+  ai[1] = from[1*stride];
+  ai[2] = from[2*stride];
+  ai[3] = from[3*stride];
+ return pload<Packet4f>(ai);
+}
+
+template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride)
+{
+  double EIGEN_ALIGN16 af[2];
+  af[0] = from[0*stride];
+  af[1] = from[1*stride];
+ return pload<Packet2d>(af);
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const Packet4i& from, Index stride)
+{
+  int EIGEN_ALIGN16 ai[4];
+  pstore<int>((int *)ai, from);
+  to[0*stride] = ai[0];
+  to[1*stride] = ai[1];
+  to[2*stride] = ai[2];
+  to[3*stride] = ai[3];
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride)
+{
+  float EIGEN_ALIGN16 ai[4];
+  pstore<float>((float *)ai, from);
+  to[0*stride] = ai[0];
+  to[1*stride] = ai[1];
+  to[2*stride] = ai[2];
+  to[3*stride] = ai[3];
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride)
+{
+  double EIGEN_ALIGN16 af[2];
+  pstore<double>(af, from);
+  to[0*stride] = af[0];
+  to[1*stride] = af[1];
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a + b); }
+template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f c;
+  c.v4f[0] = a.v4f[0] + b.v4f[0];
+  c.v4f[1] = a.v4f[1] + b.v4f[1];
+  return c;
+}
+template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a + b); }
+
+template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a - b); }
+template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f c;
+  c.v4f[0] = a.v4f[0] - b.v4f[0];
+  c.v4f[1] = a.v4f[1] - b.v4f[1];
+  return c;
+}
+template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a - b); }
+
+template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a * b); }
+template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f c;
+  c.v4f[0] = a.v4f[0] * b.v4f[0];
+  c.v4f[1] = a.v4f[1] * b.v4f[1];
+  return c;
+}
+template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a * b); }
+
+template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a / b); }
+template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f c;
+  c.v4f[0] = a.v4f[0] / b.v4f[0];
+  c.v4f[1] = a.v4f[1] / b.v4f[1];
+  return c;
+}
+template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a / b); }
+
+template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return (-a); }
+template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a)
+{
+  Packet4f c;
+  c.v4f[0] = -a.v4f[0];
+  c.v4f[1] = -a.v4f[1];
+  return c;
+}
+template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) { return (-a); }
+
+template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; }
+template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; }
+template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; }
+
+template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd<Packet4i>(pmul<Packet4i>(a, b), c); }
+template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c)
+{
+  Packet4f res;
+  res.v4f[0] = vec_madd(a.v4f[0], b.v4f[0], c.v4f[0]);
+  res.v4f[1] = vec_madd(a.v4f[1], b.v4f[1], c.v4f[1]);
+  return res;
+}
+template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vec_madd(a, b, c); }
+
+template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int& a)    { return padd<Packet4i>(pset1<Packet4i>(a), p4i_COUNTDOWN); }
+template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a)  { return padd<Packet4f>(pset1<Packet4f>(a), p4f_COUNTDOWN); }
+template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a) { return padd<Packet2d>(pset1<Packet2d>(a), p2d_COUNTDOWN); }
+
+template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_min(a, b); }
+template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_min(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pmin(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pmin(a.v4f[1], b.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_max(a, b); }
+template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_max(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pmax(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pmax(a.v4f[1], b.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, b); }
+template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pand(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pand(a.v4f[1], b.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_or(a, b); }
+template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_or(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pand(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pand(a.v4f[1], b.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_xor(a, b); }
+template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_xor(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pand(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pand(a.v4f[1], b.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return pand<Packet4i>(a, vec_nor(b, b)); }
+template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, vec_nor(b, b)); }
+template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pandnot(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pandnot(a.v4f[1], b.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a)
+{
+  Packet4f res;
+  res.v4f[0] = vec_round(a.v4f[0]);
+  res.v4f[1] = vec_round(a.v4f[1]);
+  return res;
+}
+template<> EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a) { return vec_round(a); }
+template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const  Packet4f& a)
+{
+  Packet4f res;
+  res.v4f[0] = vec_ceil(a.v4f[0]);
+  res.v4f[1] = vec_ceil(a.v4f[1]);
+  return res;
+}
+template<> EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const  Packet2d& a) { return vec_ceil(a); }
+template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a)
+{
+  Packet4f res;
+  res.v4f[0] = vec_floor(a.v4f[0]);
+  res.v4f[1] = vec_floor(a.v4f[1]);
+  return res;
+}
+template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) { return vec_floor(a); }
+
+template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int*       from) { return pload<Packet4i>(from); }
+template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float*     from) { return pload<Packet4f>(from); }
+template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double*    from) { return pload<Packet2d>(from); }
+
+
+template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int*     from)
+{
+  Packet4i p = pload<Packet4i>(from);
+  return vec_perm(p, p, p16uc_DUPLICATE32_HI);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float*    from)
+{
+  Packet4f p = pload<Packet4f>(from);
+  p.v4f[1] = vec_splat(p.v4f[0], 1);
+  p.v4f[0] = vec_splat(p.v4f[0], 0);
+  return p;
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double*   from)
+{
+  Packet2d p = pload<Packet2d>(from);
+  return vec_perm(p, p, p16uc_PSET64_HI);
+}
+
+template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*        to, const Packet4i& from) { pstore<int>(to, from); }
+template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*    to, const Packet4f& from) { pstore<float>(to, from); }
+template<> EIGEN_STRONG_INLINE void pstoreu<double>(double*  to, const Packet2d& from) { pstore<double>(to, from); }
+
+template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*       addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
+template<> EIGEN_STRONG_INLINE void prefetch<float>(const float*   addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
+template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
+
+template<> EIGEN_STRONG_INLINE int    pfirst<Packet4i>(const Packet4i& a) { int    EIGEN_ALIGN16 x[4]; pstore(x, a); return x[0]; }
+template<> EIGEN_STRONG_INLINE float  pfirst<Packet4f>(const Packet4f& a) { float  EIGEN_ALIGN16 x[2]; vec_st2f(a.v4f[0], &x[0]); return x[0]; }
+template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { double EIGEN_ALIGN16 x[2]; pstore(x, a); return x[0]; }
+
+template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a)
+{
+  return reinterpret_cast<Packet4i>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE32));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a)
+{
+  return reinterpret_cast<Packet2d>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE64));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a)
+{
+  Packet4f rev;
+  rev.v4f[0] = preverse<Packet2d>(a.v4f[1]);
+  rev.v4f[1] = preverse<Packet2d>(a.v4f[0]);
+  return rev;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pabs<Packet4i>(const Packet4i& a) { return vec_abs(a); }
+template<> EIGEN_STRONG_INLINE Packet2d pabs<Packet2d>(const Packet2d& a) { return vec_abs(a); }
+template<> EIGEN_STRONG_INLINE Packet4f pabs<Packet4f>(const Packet4f& a)
+{
+  Packet4f res;
+  res.v4f[0] = pabs(a.v4f[0]);
+  res.v4f[1] = pabs(a.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
+{
+  Packet4i b, sum;
+  b   = vec_sld(a, a, 8);
+  sum = padd<Packet4i>(a, b);
+  b   = vec_sld(sum, sum, 4);
+  sum = padd<Packet4i>(sum, b);
+  return pfirst(sum);
+}
+
+template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a)
+{
+  Packet2d b, sum;
+  b   = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8));
+  sum = padd<Packet2d>(a, b);
+  return pfirst(sum);
+}
+template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
+{
+  Packet2d sum;
+  sum = padd<Packet2d>(a.v4f[0], a.v4f[1]);
+  double first = predux<Packet2d>(sum);
+  return static_cast<float>(first);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
+{
+  Packet4i v[4], sum[4];
+
+  // It's easier and faster to transpose then add as columns
+  // Check: http://www.freevec.org/function/matrix_4x4_transpose_floats for explanation
+  // Do the transpose, first set of moves
+  v[0] = vec_mergeh(vecs[0], vecs[2]);
+  v[1] = vec_mergel(vecs[0], vecs[2]);
+  v[2] = vec_mergeh(vecs[1], vecs[3]);
+  v[3] = vec_mergel(vecs[1], vecs[3]);
+  // Get the resulting vectors
+  sum[0] = vec_mergeh(v[0], v[2]);
+  sum[1] = vec_mergel(v[0], v[2]);
+  sum[2] = vec_mergeh(v[1], v[3]);
+  sum[3] = vec_mergel(v[1], v[3]);
+
+  // Now do the summation:
+  // Lines 0+1
+  sum[0] = padd<Packet4i>(sum[0], sum[1]);
+  // Lines 2+3
+  sum[1] = padd<Packet4i>(sum[2], sum[3]);
+  // Add the results
+  sum[0] = padd<Packet4i>(sum[0], sum[1]);
+
+  return sum[0];
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
+{
+  Packet2d v[2], sum;
+  v[0] = padd<Packet2d>(vecs[0], reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(vecs[0]), reinterpret_cast<Packet4ui>(vecs[0]), 8)));
+  v[1] = padd<Packet2d>(vecs[1], reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(vecs[1]), reinterpret_cast<Packet4ui>(vecs[1]), 8)));
+ 
+  sum = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(v[0]), reinterpret_cast<Packet4ui>(v[1]), 8));
+
+  return sum;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
+{
+  PacketBlock<Packet4f,4> transpose;
+  transpose.packet[0] = vecs[0];
+  transpose.packet[1] = vecs[1];
+  transpose.packet[2] = vecs[2];
+  transpose.packet[3] = vecs[3];
+  ptranspose(transpose);
+
+  Packet4f sum = padd(transpose.packet[0], transpose.packet[1]);
+  sum = padd(sum, transpose.packet[2]);
+  sum = padd(sum, transpose.packet[3]);
+  return sum;
+}
+
+// Other reduction functions:
+// mul
+template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a)
+{
+  EIGEN_ALIGN16 int aux[4];
+  pstore(aux, a);
+  return aux[0] * aux[1] * aux[2] * aux[3];
+}
+
+template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a)
+{
+  return pfirst(pmul(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8))));
+}
+
+template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
+{
+  // Return predux_mul<Packet2d> of the subvectors product
+  return static_cast<float>(pfirst(predux_mul(pmul(a.v4f[0], a.v4f[1]))));
+}
+
+// min
+template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a)
+{
+  Packet4i b, res;
+  b   = pmin<Packet4i>(a, vec_sld(a, a, 8));
+  res = pmin<Packet4i>(b, vec_sld(b, b, 4));
+  return pfirst(res);
+}
+
+template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a)
+{
+  return pfirst(pmin<Packet2d>(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8))));
+}
+
+template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a)
+{
+  Packet2d b, res;
+  b   = pmin<Packet2d>(a.v4f[0], a.v4f[1]);
+  res = pmin<Packet2d>(b, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(b), reinterpret_cast<Packet4i>(b), 8)));
+  return static_cast<float>(pfirst(res));
+}
+
+// max
+template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a)
+{
+  Packet4i b, res;
+  b = pmax<Packet4i>(a, vec_sld(a, a, 8));
+  res = pmax<Packet4i>(b, vec_sld(b, b, 4));
+  return pfirst(res);
+}
+
+// max
+template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a)
+{
+  return pfirst(pmax<Packet2d>(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8))));
+}
+
+template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a)
+{
+  Packet2d b, res;
+  b   = pmax<Packet2d>(a.v4f[0], a.v4f[1]);
+  res = pmax<Packet2d>(b, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(b), reinterpret_cast<Packet4i>(b), 8)));
+  return static_cast<float>(pfirst(res));
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet4i,4>& kernel) {
+  Packet4i t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]);
+  Packet4i t1 = vec_mergel(kernel.packet[0], kernel.packet[2]);
+  Packet4i t2 = vec_mergeh(kernel.packet[1], kernel.packet[3]);
+  Packet4i t3 = vec_mergel(kernel.packet[1], kernel.packet[3]);
+  kernel.packet[0] = vec_mergeh(t0, t2);
+  kernel.packet[1] = vec_mergel(t0, t2);
+  kernel.packet[2] = vec_mergeh(t1, t3);
+  kernel.packet[3] = vec_mergel(t1, t3);
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet2d,2>& kernel) {
+  Packet2d t0 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_HI);
+  Packet2d t1 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_LO);
+  kernel.packet[0] = t0;
+  kernel.packet[1] = t1;
+}
+
+/* Split the Packet4f PacketBlock into 4 Packet2d PacketBlocks and transpose each one
+ */
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet4f,4>& kernel) {
+  PacketBlock<Packet2d,2> t0,t1,t2,t3;
+  // copy top-left 2x2 Packet2d block
+  t0.packet[0] = kernel.packet[0].v4f[0];
+  t0.packet[1] = kernel.packet[1].v4f[0];
+
+  // copy top-right 2x2 Packet2d block
+  t1.packet[0] = kernel.packet[0].v4f[1];
+  t1.packet[1] = kernel.packet[1].v4f[1];
+
+  // copy bottom-left 2x2 Packet2d block
+  t2.packet[0] = kernel.packet[2].v4f[0];
+  t2.packet[1] = kernel.packet[3].v4f[0];
+
+  // copy bottom-right 2x2 Packet2d block
+  t3.packet[0] = kernel.packet[2].v4f[1];
+  t3.packet[1] = kernel.packet[3].v4f[1];
+
+  // Transpose all 2x2 blocks
+  ptranspose(t0);
+  ptranspose(t1);
+  ptranspose(t2);
+  ptranspose(t3);
+
+  // Copy back transposed blocks, but exchange t1 and t2 due to transposition
+  kernel.packet[0].v4f[0] = t0.packet[0];
+  kernel.packet[0].v4f[1] = t2.packet[0];
+  kernel.packet[1].v4f[0] = t0.packet[1];
+  kernel.packet[1].v4f[1] = t2.packet[1];
+  kernel.packet[2].v4f[0] = t1.packet[0];
+  kernel.packet[2].v4f[1] = t3.packet[0];
+  kernel.packet[3].v4f[0] = t1.packet[1];
+  kernel.packet[3].v4f[1] = t3.packet[1];
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pblend(const Selector<4>& ifPacket, const Packet4i& thenPacket, const Packet4i& elsePacket) {
+  Packet4ui select = { ifPacket.select[0], ifPacket.select[1], ifPacket.select[2], ifPacket.select[3] };
+  Packet4ui mask = vec_cmpeq(select, reinterpret_cast<Packet4ui>(p4i_ONE));
+  return vec_sel(elsePacket, thenPacket, mask);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket, const Packet4f& elsePacket) {
+  Packet2ul select_hi = { ifPacket.select[0], ifPacket.select[1] };
+  Packet2ul select_lo = { ifPacket.select[2], ifPacket.select[3] };
+  Packet2ul mask_hi = vec_cmpeq(select_hi, reinterpret_cast<Packet2ul>(p2l_ONE));
+  Packet2ul mask_lo = vec_cmpeq(select_lo, reinterpret_cast<Packet2ul>(p2l_ONE));
+  Packet4f result;
+  result.v4f[0] = vec_sel(elsePacket.v4f[0], thenPacket.v4f[0], mask_hi);
+  result.v4f[1] = vec_sel(elsePacket.v4f[1], thenPacket.v4f[1], mask_lo);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket, const Packet2d& elsePacket) {
+  Packet2ul select = { ifPacket.select[0], ifPacket.select[1] };
+  Packet2ul mask = vec_cmpeq(select, reinterpret_cast<Packet2ul>(p2l_ONE));
+  return vec_sel(elsePacket, thenPacket, mask);
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_PACKET_MATH_ZVECTOR_H
diff --git a/eigen/Eigen/src/Core/functors/AssignmentFunctors.h b/eigen/Eigen/src/Core/functors/AssignmentFunctors.h
new file mode 100644
index 0000000..4153b87
--- /dev/null
+++ b/eigen/Eigen/src/Core/functors/AssignmentFunctors.h
@@ -0,0 +1,168 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ASSIGNMENT_FUNCTORS_H
+#define EIGEN_ASSIGNMENT_FUNCTORS_H
+
+namespace Eigen {
+
+namespace internal {
+  
+/** \internal
+  * \brief Template functor for scalar/packet assignment
+  *
+  */
+template<typename DstScalar,typename SrcScalar> struct assign_op {
+
+  EIGEN_EMPTY_STRUCT_CTOR(assign_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(DstScalar& a, const SrcScalar& b) const { a = b; }
+  
+  template<int Alignment, typename Packet>
+  EIGEN_STRONG_INLINE void assignPacket(DstScalar* a, const Packet& b) const
+  { internal::pstoret<DstScalar,Packet,Alignment>(a,b); }
+};
+
+// Empty overload for void type (used by PermutationMatrix)
+template<typename DstScalar> struct assign_op<DstScalar,void> {};
+
+template<typename DstScalar,typename SrcScalar>
+struct functor_traits<assign_op<DstScalar,SrcScalar> > {
+  enum {
+    Cost = NumTraits<DstScalar>::ReadCost,
+    PacketAccess = is_same<DstScalar,SrcScalar>::value && packet_traits<DstScalar>::Vectorizable && packet_traits<SrcScalar>::Vectorizable
+  };
+};
+
+/** \internal
+  * \brief Template functor for scalar/packet assignment with addition
+  *
+  */
+template<typename DstScalar,typename SrcScalar> struct add_assign_op {
+
+  EIGEN_EMPTY_STRUCT_CTOR(add_assign_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(DstScalar& a, const SrcScalar& b) const { a += b; }
+  
+  template<int Alignment, typename Packet>
+  EIGEN_STRONG_INLINE void assignPacket(DstScalar* a, const Packet& b) const
+  { internal::pstoret<DstScalar,Packet,Alignment>(a,internal::padd(internal::ploadt<Packet,Alignment>(a),b)); }
+};
+template<typename DstScalar,typename SrcScalar>
+struct functor_traits<add_assign_op<DstScalar,SrcScalar> > {
+  enum {
+    Cost = NumTraits<DstScalar>::ReadCost + NumTraits<DstScalar>::AddCost,
+    PacketAccess = is_same<DstScalar,SrcScalar>::value && packet_traits<DstScalar>::HasAdd
+  };
+};
+
+/** \internal
+  * \brief Template functor for scalar/packet assignment with subtraction
+  *
+  */
+template<typename DstScalar,typename SrcScalar> struct sub_assign_op {
+
+  EIGEN_EMPTY_STRUCT_CTOR(sub_assign_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(DstScalar& a, const SrcScalar& b) const { a -= b; }
+  
+  template<int Alignment, typename Packet>
+  EIGEN_STRONG_INLINE void assignPacket(DstScalar* a, const Packet& b) const
+  { internal::pstoret<DstScalar,Packet,Alignment>(a,internal::psub(internal::ploadt<Packet,Alignment>(a),b)); }
+};
+template<typename DstScalar,typename SrcScalar>
+struct functor_traits<sub_assign_op<DstScalar,SrcScalar> > {
+  enum {
+    Cost = NumTraits<DstScalar>::ReadCost + NumTraits<DstScalar>::AddCost,
+    PacketAccess = is_same<DstScalar,SrcScalar>::value && packet_traits<DstScalar>::HasSub
+  };
+};
+
+/** \internal
+  * \brief Template functor for scalar/packet assignment with multiplication
+  *
+  */
+template<typename DstScalar, typename SrcScalar=DstScalar>
+struct mul_assign_op {
+
+  EIGEN_EMPTY_STRUCT_CTOR(mul_assign_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(DstScalar& a, const SrcScalar& b) const { a *= b; }
+  
+  template<int Alignment, typename Packet>
+  EIGEN_STRONG_INLINE void assignPacket(DstScalar* a, const Packet& b) const
+  { internal::pstoret<DstScalar,Packet,Alignment>(a,internal::pmul(internal::ploadt<Packet,Alignment>(a),b)); }
+};
+template<typename DstScalar, typename SrcScalar>
+struct functor_traits<mul_assign_op<DstScalar,SrcScalar> > {
+  enum {
+    Cost = NumTraits<DstScalar>::ReadCost + NumTraits<DstScalar>::MulCost,
+    PacketAccess = is_same<DstScalar,SrcScalar>::value && packet_traits<DstScalar>::HasMul
+  };
+};
+
+/** \internal
+  * \brief Template functor for scalar/packet assignment with diviving
+  *
+  */
+template<typename DstScalar, typename SrcScalar=DstScalar> struct div_assign_op {
+
+  EIGEN_EMPTY_STRUCT_CTOR(div_assign_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(DstScalar& a, const SrcScalar& b) const { a /= b; }
+  
+  template<int Alignment, typename Packet>
+  EIGEN_STRONG_INLINE void assignPacket(DstScalar* a, const Packet& b) const
+  { internal::pstoret<DstScalar,Packet,Alignment>(a,internal::pdiv(internal::ploadt<Packet,Alignment>(a),b)); }
+};
+template<typename DstScalar, typename SrcScalar>
+struct functor_traits<div_assign_op<DstScalar,SrcScalar> > {
+  enum {
+    Cost = NumTraits<DstScalar>::ReadCost + NumTraits<DstScalar>::MulCost,
+    PacketAccess = is_same<DstScalar,SrcScalar>::value && packet_traits<DstScalar>::HasDiv
+  };
+};
+
+/** \internal
+  * \brief Template functor for scalar/packet assignment with swapping
+  *
+  * It works as follow. For a non-vectorized evaluation loop, we have:
+  *   for(i) func(A.coeffRef(i), B.coeff(i));
+  * where B is a SwapWrapper expression. The trick is to make SwapWrapper::coeff behaves like a non-const coeffRef.
+  * Actually, SwapWrapper might not even be needed since even if B is a plain expression, since it has to be writable
+  * B.coeff already returns a const reference to the underlying scalar value.
+  * 
+  * The case of a vectorized loop is more tricky:
+  *   for(i,j) func.assignPacket<A_Align>(&A.coeffRef(i,j), B.packet<B_Align>(i,j));
+  * Here, B must be a SwapWrapper whose packet function actually returns a proxy object holding a Scalar*,
+  * the actual alignment and Packet type.
+  *
+  */
+template<typename Scalar> struct swap_assign_op {
+
+  EIGEN_EMPTY_STRUCT_CTOR(swap_assign_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(Scalar& a, const Scalar& b) const
+  {
+#ifdef __CUDACC__
+    // FIXME is there some kind of cuda::swap?
+    Scalar t=b; const_cast<Scalar&>(b)=a; a=t;
+#else
+    using std::swap;
+    swap(a,const_cast<Scalar&>(b));
+#endif
+  }
+};
+template<typename Scalar>
+struct functor_traits<swap_assign_op<Scalar> > {
+  enum {
+    Cost = 3 * NumTraits<Scalar>::ReadCost,
+    PacketAccess = packet_traits<Scalar>::Vectorizable
+  };
+};
+
+} // namespace internal
+
+} // namespace Eigen
+
+#endif // EIGEN_ASSIGNMENT_FUNCTORS_H
diff --git a/eigen/Eigen/src/Core/functors/BinaryFunctors.h b/eigen/Eigen/src/Core/functors/BinaryFunctors.h
new file mode 100644
index 0000000..96747ba
--- /dev/null
+++ b/eigen/Eigen/src/Core/functors/BinaryFunctors.h
@@ -0,0 +1,482 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_BINARY_FUNCTORS_H
+#define EIGEN_BINARY_FUNCTORS_H
+
+namespace Eigen {
+
+namespace internal {
+
+//---------- associative binary functors ----------
+
+template<typename Arg1, typename Arg2>
+struct binary_op_base
+{
+  typedef Arg1 first_argument_type;
+  typedef Arg2 second_argument_type;
+};
+
+/** \internal
+  * \brief Template functor to compute the sum of two scalars
+  *
+  * \sa class CwiseBinaryOp, MatrixBase::operator+, class VectorwiseOp, DenseBase::sum()
+  */
+template<typename LhsScalar,typename RhsScalar>
+struct scalar_sum_op : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar,scalar_sum_op>::ReturnType result_type;
+#ifndef EIGEN_SCALAR_BINARY_OP_PLUGIN
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_sum_op)
+#else
+  scalar_sum_op() {
+    EIGEN_SCALAR_BINARY_OP_PLUGIN
+  }
+#endif
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a + b; }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
+  { return internal::padd(a,b); }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type predux(const Packet& a) const
+  { return internal::predux(a); }
+};
+template<typename LhsScalar,typename RhsScalar>
+struct functor_traits<scalar_sum_op<LhsScalar,RhsScalar> > {
+  enum {
+    Cost = (NumTraits<LhsScalar>::AddCost+NumTraits<RhsScalar>::AddCost)/2, // rough estimate!
+    PacketAccess = is_same<LhsScalar,RhsScalar>::value && packet_traits<LhsScalar>::HasAdd && packet_traits<RhsScalar>::HasAdd
+    // TODO vectorize mixed sum
+  };
+};
+
+/** \internal
+  * \brief Template specialization to deprecate the summation of boolean expressions.
+  * This is required to solve Bug 426.
+  * \sa DenseBase::count(), DenseBase::any(), ArrayBase::cast(), MatrixBase::cast()
+  */
+template<> struct scalar_sum_op<bool,bool> : scalar_sum_op<int,int> {
+  EIGEN_DEPRECATED
+  scalar_sum_op() {}
+};
+
+
+/** \internal
+  * \brief Template functor to compute the product of two scalars
+  *
+  * \sa class CwiseBinaryOp, Cwise::operator*(), class VectorwiseOp, MatrixBase::redux()
+  */
+template<typename LhsScalar,typename RhsScalar>
+struct scalar_product_op  : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar,scalar_product_op>::ReturnType result_type;
+#ifndef EIGEN_SCALAR_BINARY_OP_PLUGIN
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_product_op)
+#else
+  scalar_product_op() {
+    EIGEN_SCALAR_BINARY_OP_PLUGIN
+  }
+#endif
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a * b; }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
+  { return internal::pmul(a,b); }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type predux(const Packet& a) const
+  { return internal::predux_mul(a); }
+};
+template<typename LhsScalar,typename RhsScalar>
+struct functor_traits<scalar_product_op<LhsScalar,RhsScalar> > {
+  enum {
+    Cost = (NumTraits<LhsScalar>::MulCost + NumTraits<RhsScalar>::MulCost)/2, // rough estimate!
+    PacketAccess = is_same<LhsScalar,RhsScalar>::value && packet_traits<LhsScalar>::HasMul && packet_traits<RhsScalar>::HasMul
+    // TODO vectorize mixed product
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the conjugate product of two scalars
+  *
+  * This is a short cut for conj(x) * y which is needed for optimization purpose; in Eigen2 support mode, this becomes x * conj(y)
+  */
+template<typename LhsScalar,typename RhsScalar>
+struct scalar_conj_product_op  : binary_op_base<LhsScalar,RhsScalar>
+{
+
+  enum {
+    Conj = NumTraits<LhsScalar>::IsComplex
+  };
+  
+  typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar,scalar_conj_product_op>::ReturnType result_type;
+  
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_conj_product_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const
+  { return conj_helper<LhsScalar,RhsScalar,Conj,false>().pmul(a,b); }
+  
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
+  { return conj_helper<Packet,Packet,Conj,false>().pmul(a,b); }
+};
+template<typename LhsScalar,typename RhsScalar>
+struct functor_traits<scalar_conj_product_op<LhsScalar,RhsScalar> > {
+  enum {
+    Cost = NumTraits<LhsScalar>::MulCost,
+    PacketAccess = internal::is_same<LhsScalar, RhsScalar>::value && packet_traits<LhsScalar>::HasMul
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the min of two scalars
+  *
+  * \sa class CwiseBinaryOp, MatrixBase::cwiseMin, class VectorwiseOp, MatrixBase::minCoeff()
+  */
+template<typename LhsScalar,typename RhsScalar>
+struct scalar_min_op : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar,scalar_min_op>::ReturnType result_type;
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_min_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return numext::mini(a, b); }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
+  { return internal::pmin(a,b); }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type predux(const Packet& a) const
+  { return internal::predux_min(a); }
+};
+template<typename LhsScalar,typename RhsScalar>
+struct functor_traits<scalar_min_op<LhsScalar,RhsScalar> > {
+  enum {
+    Cost = (NumTraits<LhsScalar>::AddCost+NumTraits<RhsScalar>::AddCost)/2,
+    PacketAccess = internal::is_same<LhsScalar, RhsScalar>::value && packet_traits<LhsScalar>::HasMin
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the max of two scalars
+  *
+  * \sa class CwiseBinaryOp, MatrixBase::cwiseMax, class VectorwiseOp, MatrixBase::maxCoeff()
+  */
+template<typename LhsScalar,typename RhsScalar>
+struct scalar_max_op  : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar,scalar_max_op>::ReturnType result_type;
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_max_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return numext::maxi(a, b); }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
+  { return internal::pmax(a,b); }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type predux(const Packet& a) const
+  { return internal::predux_max(a); }
+};
+template<typename LhsScalar,typename RhsScalar>
+struct functor_traits<scalar_max_op<LhsScalar,RhsScalar> > {
+  enum {
+    Cost = (NumTraits<LhsScalar>::AddCost+NumTraits<RhsScalar>::AddCost)/2,
+    PacketAccess = internal::is_same<LhsScalar, RhsScalar>::value && packet_traits<LhsScalar>::HasMax
+  };
+};
+
+/** \internal
+  * \brief Template functors for comparison of two scalars
+  * \todo Implement packet-comparisons
+  */
+template<typename LhsScalar, typename RhsScalar, ComparisonName cmp> struct scalar_cmp_op;
+
+template<typename LhsScalar, typename RhsScalar, ComparisonName cmp>
+struct functor_traits<scalar_cmp_op<LhsScalar,RhsScalar, cmp> > {
+  enum {
+    Cost = (NumTraits<LhsScalar>::AddCost+NumTraits<RhsScalar>::AddCost)/2,
+    PacketAccess = false
+  };
+};
+
+template<ComparisonName Cmp, typename LhsScalar, typename RhsScalar>
+struct result_of<scalar_cmp_op<LhsScalar, RhsScalar, Cmp>(LhsScalar,RhsScalar)> {
+  typedef bool type;
+};
+
+
+template<typename LhsScalar, typename RhsScalar>
+struct scalar_cmp_op<LhsScalar,RhsScalar, cmp_EQ> : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef bool result_type;
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator()(const LhsScalar& a, const RhsScalar& b) const {return a==b;}
+};
+template<typename LhsScalar, typename RhsScalar>
+struct scalar_cmp_op<LhsScalar,RhsScalar, cmp_LT> : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef bool result_type;
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator()(const LhsScalar& a, const RhsScalar& b) const {return a<b;}
+};
+template<typename LhsScalar, typename RhsScalar>
+struct scalar_cmp_op<LhsScalar,RhsScalar, cmp_LE> : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef bool result_type;
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator()(const LhsScalar& a, const RhsScalar& b) const {return a<=b;}
+};
+template<typename LhsScalar, typename RhsScalar>
+struct scalar_cmp_op<LhsScalar,RhsScalar, cmp_GT> : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef bool result_type;
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator()(const LhsScalar& a, const RhsScalar& b) const {return a>b;}
+};
+template<typename LhsScalar, typename RhsScalar>
+struct scalar_cmp_op<LhsScalar,RhsScalar, cmp_GE> : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef bool result_type;
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator()(const LhsScalar& a, const RhsScalar& b) const {return a>=b;}
+};
+template<typename LhsScalar, typename RhsScalar>
+struct scalar_cmp_op<LhsScalar,RhsScalar, cmp_UNORD> : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef bool result_type;
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator()(const LhsScalar& a, const RhsScalar& b) const {return !(a<=b || b<=a);}
+};
+template<typename LhsScalar, typename RhsScalar>
+struct scalar_cmp_op<LhsScalar,RhsScalar, cmp_NEQ> : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef bool result_type;
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator()(const LhsScalar& a, const RhsScalar& b) const {return a!=b;}
+};
+
+
+/** \internal
+  * \brief Template functor to compute the hypot of two scalars
+  *
+  * \sa MatrixBase::stableNorm(), class Redux
+  */
+template<typename Scalar>
+struct scalar_hypot_op<Scalar,Scalar> : binary_op_base<Scalar,Scalar>
+{
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_hypot_op)
+//   typedef typename NumTraits<Scalar>::Real result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& _x, const Scalar& _y) const
+  {
+    EIGEN_USING_STD_MATH(sqrt)
+    Scalar p, qp;
+    if(_x>_y)
+    {
+      p = _x;
+      qp = _y / p;
+    }
+    else
+    {
+      p = _y;
+      qp = _x / p;
+    }
+    return p * sqrt(Scalar(1) + qp*qp);
+  }
+};
+template<typename Scalar>
+struct functor_traits<scalar_hypot_op<Scalar,Scalar> > {
+  enum
+  {
+    Cost = 3 * NumTraits<Scalar>::AddCost +
+           2 * NumTraits<Scalar>::MulCost +
+           2 * scalar_div_cost<Scalar,false>::value,
+    PacketAccess = false
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the pow of two scalars
+  */
+template<typename Scalar, typename Exponent>
+struct scalar_pow_op  : binary_op_base<Scalar,Exponent>
+{
+  typedef typename ScalarBinaryOpTraits<Scalar,Exponent,scalar_pow_op>::ReturnType result_type;
+#ifndef EIGEN_SCALAR_BINARY_OP_PLUGIN
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_pow_op)
+#else
+  scalar_pow_op() {
+    typedef Scalar LhsScalar;
+    typedef Exponent RhsScalar;
+    EIGEN_SCALAR_BINARY_OP_PLUGIN
+  }
+#endif
+  EIGEN_DEVICE_FUNC
+  inline result_type operator() (const Scalar& a, const Exponent& b) const { return numext::pow(a, b); }
+};
+template<typename Scalar, typename Exponent>
+struct functor_traits<scalar_pow_op<Scalar,Exponent> > {
+  enum { Cost = 5 * NumTraits<Scalar>::MulCost, PacketAccess = false };
+};
+
+
+
+//---------- non associative binary functors ----------
+
+/** \internal
+  * \brief Template functor to compute the difference of two scalars
+  *
+  * \sa class CwiseBinaryOp, MatrixBase::operator-
+  */
+template<typename LhsScalar,typename RhsScalar>
+struct scalar_difference_op : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar,scalar_difference_op>::ReturnType result_type;
+#ifndef EIGEN_SCALAR_BINARY_OP_PLUGIN
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_difference_op)
+#else
+  scalar_difference_op() {
+    EIGEN_SCALAR_BINARY_OP_PLUGIN
+  }
+#endif
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a - b; }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
+  { return internal::psub(a,b); }
+};
+template<typename LhsScalar,typename RhsScalar>
+struct functor_traits<scalar_difference_op<LhsScalar,RhsScalar> > {
+  enum {
+    Cost = (NumTraits<LhsScalar>::AddCost+NumTraits<RhsScalar>::AddCost)/2,
+    PacketAccess = is_same<LhsScalar,RhsScalar>::value && packet_traits<LhsScalar>::HasSub && packet_traits<RhsScalar>::HasSub
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the quotient of two scalars
+  *
+  * \sa class CwiseBinaryOp, Cwise::operator/()
+  */
+template<typename LhsScalar,typename RhsScalar>
+struct scalar_quotient_op  : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar,scalar_quotient_op>::ReturnType result_type;
+#ifndef EIGEN_SCALAR_BINARY_OP_PLUGIN
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_quotient_op)
+#else
+  scalar_quotient_op() {
+    EIGEN_SCALAR_BINARY_OP_PLUGIN
+  }
+#endif
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a / b; }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
+  { return internal::pdiv(a,b); }
+};
+template<typename LhsScalar,typename RhsScalar>
+struct functor_traits<scalar_quotient_op<LhsScalar,RhsScalar> > {
+  typedef typename scalar_quotient_op<LhsScalar,RhsScalar>::result_type result_type;
+  enum {
+    PacketAccess = is_same<LhsScalar,RhsScalar>::value && packet_traits<LhsScalar>::HasDiv && packet_traits<RhsScalar>::HasDiv,
+    Cost = scalar_div_cost<result_type,PacketAccess>::value
+  };
+};
+
+
+
+/** \internal
+  * \brief Template functor to compute the and of two booleans
+  *
+  * \sa class CwiseBinaryOp, ArrayBase::operator&&
+  */
+struct scalar_boolean_and_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_boolean_and_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator() (const bool& a, const bool& b) const { return a && b; }
+};
+template<> struct functor_traits<scalar_boolean_and_op> {
+  enum {
+    Cost = NumTraits<bool>::AddCost,
+    PacketAccess = false
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the or of two booleans
+  *
+  * \sa class CwiseBinaryOp, ArrayBase::operator||
+  */
+struct scalar_boolean_or_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_boolean_or_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator() (const bool& a, const bool& b) const { return a || b; }
+};
+template<> struct functor_traits<scalar_boolean_or_op> {
+  enum {
+    Cost = NumTraits<bool>::AddCost,
+    PacketAccess = false
+  };
+};
+
+/** \internal
+ * \brief Template functor to compute the xor of two booleans
+ *
+ * \sa class CwiseBinaryOp, ArrayBase::operator^
+ */
+struct scalar_boolean_xor_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_boolean_xor_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator() (const bool& a, const bool& b) const { return a ^ b; }
+};
+template<> struct functor_traits<scalar_boolean_xor_op> {
+  enum {
+    Cost = NumTraits<bool>::AddCost,
+    PacketAccess = false
+  };
+};
+
+
+
+//---------- binary functors bound to a constant, thus appearing as a unary functor ----------
+
+// The following two classes permits to turn any binary functor into a unary one with one argument bound to a constant value.
+// They are analogues to std::binder1st/binder2nd but with the following differences:
+//  - they are compatible with packetOp
+//  - they are portable across C++ versions (the std::binder* are deprecated in C++11)
+template<typename BinaryOp> struct bind1st_op : BinaryOp {
+
+  typedef typename BinaryOp::first_argument_type  first_argument_type;
+  typedef typename BinaryOp::second_argument_type second_argument_type;
+  typedef typename BinaryOp::result_type          result_type;
+
+  bind1st_op(const first_argument_type &val) : m_value(val) {}
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const second_argument_type& b) const { return BinaryOp::operator()(m_value,b); }
+
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& b) const
+  { return BinaryOp::packetOp(internal::pset1<Packet>(m_value), b); }
+
+  first_argument_type m_value;
+};
+template<typename BinaryOp> struct functor_traits<bind1st_op<BinaryOp> > : functor_traits<BinaryOp> {};
+
+
+template<typename BinaryOp> struct bind2nd_op : BinaryOp {
+
+  typedef typename BinaryOp::first_argument_type  first_argument_type;
+  typedef typename BinaryOp::second_argument_type second_argument_type;
+  typedef typename BinaryOp::result_type          result_type;
+
+  bind2nd_op(const second_argument_type &val) : m_value(val) {}
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const first_argument_type& a) const { return BinaryOp::operator()(a,m_value); }
+
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
+  { return BinaryOp::packetOp(a,internal::pset1<Packet>(m_value)); }
+
+  second_argument_type m_value;
+};
+template<typename BinaryOp> struct functor_traits<bind2nd_op<BinaryOp> > : functor_traits<BinaryOp> {};
+
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_BINARY_FUNCTORS_H
diff --git a/eigen/Eigen/src/Core/functors/NullaryFunctors.h b/eigen/Eigen/src/Core/functors/NullaryFunctors.h
new file mode 100644
index 0000000..6a30466
--- /dev/null
+++ b/eigen/Eigen/src/Core/functors/NullaryFunctors.h
@@ -0,0 +1,189 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2016 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_NULLARY_FUNCTORS_H
+#define EIGEN_NULLARY_FUNCTORS_H
+
+namespace Eigen {
+
+namespace internal {
+
+template<typename Scalar>
+struct scalar_constant_op {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_constant_op(const scalar_constant_op& other) : m_other(other.m_other) { }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_constant_op(const Scalar& other) : m_other(other) { }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() () const { return m_other; }
+  template<typename PacketType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const PacketType packetOp() const { return internal::pset1<PacketType>(m_other); }
+  const Scalar m_other;
+};
+template<typename Scalar>
+struct functor_traits<scalar_constant_op<Scalar> >
+{ enum { Cost = 0 /* as the constant value should be loaded in register only once for the whole expression */,
+         PacketAccess = packet_traits<Scalar>::Vectorizable, IsRepeatable = true }; };
+
+template<typename Scalar> struct scalar_identity_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_identity_op)
+  template<typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (IndexType row, IndexType col) const { return row==col ? Scalar(1) : Scalar(0); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_identity_op<Scalar> >
+{ enum { Cost = NumTraits<Scalar>::AddCost, PacketAccess = false, IsRepeatable = true }; };
+
+template <typename Scalar, typename Packet, bool IsInteger> struct linspaced_op_impl;
+
+template <typename Scalar, typename Packet>
+struct linspaced_op_impl<Scalar,Packet,/*IsInteger*/false>
+{
+  linspaced_op_impl(const Scalar& low, const Scalar& high, Index num_steps) :
+    m_low(low), m_high(high), m_size1(num_steps==1 ? 1 : num_steps-1), m_step(num_steps==1 ? Scalar() : (high-low)/Scalar(num_steps-1)),
+    m_interPacket(plset<Packet>(0)),
+    m_flip(numext::abs(high)<numext::abs(low))
+  {}
+
+  template<typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (IndexType i) const {
+    if(m_flip)
+      return (i==0)? m_low : (m_high - (m_size1-i)*m_step);
+    else
+      return (i==m_size1)? m_high : (m_low + i*m_step);
+  }
+
+  template<typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(IndexType i) const
+  {
+    // Principle:
+    // [low, ..., low] + ( [step, ..., step] * ( [i, ..., i] + [0, ..., size] ) )
+    if(m_flip)
+    {
+      Packet pi = padd(pset1<Packet>(Scalar(i-m_size1)),m_interPacket);
+      Packet res = padd(pset1<Packet>(m_high), pmul(pset1<Packet>(m_step), pi));
+      if(i==0)
+        res = pinsertfirst(res, m_low);
+      return res;
+    }
+    else
+    {
+      Packet pi = padd(pset1<Packet>(Scalar(i)),m_interPacket);
+      Packet res = padd(pset1<Packet>(m_low), pmul(pset1<Packet>(m_step), pi));
+      if(i==m_size1-unpacket_traits<Packet>::size+1)
+        res = pinsertlast(res, m_high);
+      return res;
+    }
+  }
+
+  const Scalar m_low;
+  const Scalar m_high;
+  const Index m_size1;
+  const Scalar m_step;
+  const Packet m_interPacket;
+  const bool m_flip;
+};
+
+template <typename Scalar, typename Packet>
+struct linspaced_op_impl<Scalar,Packet,/*IsInteger*/true>
+{
+  linspaced_op_impl(const Scalar& low, const Scalar& high, Index num_steps) :
+    m_low(low),
+    m_multiplier((high-low)/convert_index<Scalar>(num_steps<=1 ? 1 : num_steps-1)),
+    m_divisor(convert_index<Scalar>((high>=low?num_steps:-num_steps)+(high-low))/((numext::abs(high-low)+1)==0?1:(numext::abs(high-low)+1))),
+    m_use_divisor(num_steps>1 && (numext::abs(high-low)+1)<num_steps)
+  {}
+
+  template<typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  const Scalar operator() (IndexType i) const
+  {
+    if(m_use_divisor) return m_low + convert_index<Scalar>(i)/m_divisor;
+    else              return m_low + convert_index<Scalar>(i)*m_multiplier;
+  }
+
+  const Scalar m_low;
+  const Scalar m_multiplier;
+  const Scalar m_divisor;
+  const bool m_use_divisor;
+};
+
+// ----- Linspace functor ----------------------------------------------------------------
+
+// Forward declaration (we default to random access which does not really give
+// us a speed gain when using packet access but it allows to use the functor in
+// nested expressions).
+template <typename Scalar, typename PacketType> struct linspaced_op;
+template <typename Scalar, typename PacketType> struct functor_traits< linspaced_op<Scalar,PacketType> >
+{
+  enum
+  {
+    Cost = 1,
+    PacketAccess =   (!NumTraits<Scalar>::IsInteger) && packet_traits<Scalar>::HasSetLinear && packet_traits<Scalar>::HasBlend,
+                  /*&& ((!NumTraits<Scalar>::IsInteger) || packet_traits<Scalar>::HasDiv),*/ // <- vectorization for integer is currently disabled
+    IsRepeatable = true
+  };
+};
+template <typename Scalar, typename PacketType> struct linspaced_op
+{
+  linspaced_op(const Scalar& low, const Scalar& high, Index num_steps)
+    : impl((num_steps==1 ? high : low),high,num_steps)
+  {}
+
+  template<typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (IndexType i) const { return impl(i); }
+
+  template<typename Packet,typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(IndexType i) const { return impl.packetOp(i); }
+
+  // This proxy object handles the actual required temporaries and the different
+  // implementations (integer vs. floating point).
+  const linspaced_op_impl<Scalar,PacketType,NumTraits<Scalar>::IsInteger> impl;
+};
+
+// Linear access is automatically determined from the operator() prototypes available for the given functor.
+// If it exposes an operator()(i,j), then we assume the i and j coefficients are required independently
+// and linear access is not possible. In all other cases, linear access is enabled.
+// Users should not have to deal with this structure.
+template<typename Functor> struct functor_has_linear_access { enum { ret = !has_binary_operator<Functor>::value }; };
+
+// For unreliable compilers, let's specialize the has_*ary_operator
+// helpers so that at least built-in nullary functors work fine.
+#if !( (EIGEN_COMP_MSVC>1600) || (EIGEN_GNUC_AT_LEAST(4,8)) || (EIGEN_COMP_ICC>=1600))
+template<typename Scalar,typename IndexType>
+struct has_nullary_operator<scalar_constant_op<Scalar>,IndexType> { enum { value = 1}; };
+template<typename Scalar,typename IndexType>
+struct has_unary_operator<scalar_constant_op<Scalar>,IndexType> { enum { value = 0}; };
+template<typename Scalar,typename IndexType>
+struct has_binary_operator<scalar_constant_op<Scalar>,IndexType> { enum { value = 0}; };
+
+template<typename Scalar,typename IndexType>
+struct has_nullary_operator<scalar_identity_op<Scalar>,IndexType> { enum { value = 0}; };
+template<typename Scalar,typename IndexType>
+struct has_unary_operator<scalar_identity_op<Scalar>,IndexType> { enum { value = 0}; };
+template<typename Scalar,typename IndexType>
+struct has_binary_operator<scalar_identity_op<Scalar>,IndexType> { enum { value = 1}; };
+
+template<typename Scalar, typename PacketType,typename IndexType>
+struct has_nullary_operator<linspaced_op<Scalar,PacketType>,IndexType> { enum { value = 0}; };
+template<typename Scalar, typename PacketType,typename IndexType>
+struct has_unary_operator<linspaced_op<Scalar,PacketType>,IndexType> { enum { value = 1}; };
+template<typename Scalar, typename PacketType,typename IndexType>
+struct has_binary_operator<linspaced_op<Scalar,PacketType>,IndexType> { enum { value = 0}; };
+
+template<typename Scalar,typename IndexType>
+struct has_nullary_operator<scalar_random_op<Scalar>,IndexType> { enum { value = 1}; };
+template<typename Scalar,typename IndexType>
+struct has_unary_operator<scalar_random_op<Scalar>,IndexType> { enum { value = 0}; };
+template<typename Scalar,typename IndexType>
+struct has_binary_operator<scalar_random_op<Scalar>,IndexType> { enum { value = 0}; };
+#endif
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_NULLARY_FUNCTORS_H
diff --git a/eigen/Eigen/src/Core/functors/StlFunctors.h b/eigen/Eigen/src/Core/functors/StlFunctors.h
new file mode 100644
index 0000000..6df3fa5
--- /dev/null
+++ b/eigen/Eigen/src/Core/functors/StlFunctors.h
@@ -0,0 +1,132 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_STL_FUNCTORS_H
+#define EIGEN_STL_FUNCTORS_H
+
+namespace Eigen {
+
+namespace internal {
+
+// default functor traits for STL functors:
+
+template<typename T>
+struct functor_traits<std::multiplies<T> >
+{ enum { Cost = NumTraits<T>::MulCost, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::divides<T> >
+{ enum { Cost = NumTraits<T>::MulCost, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::plus<T> >
+{ enum { Cost = NumTraits<T>::AddCost, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::minus<T> >
+{ enum { Cost = NumTraits<T>::AddCost, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::negate<T> >
+{ enum { Cost = NumTraits<T>::AddCost, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::logical_or<T> >
+{ enum { Cost = 1, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::logical_and<T> >
+{ enum { Cost = 1, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::logical_not<T> >
+{ enum { Cost = 1, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::greater<T> >
+{ enum { Cost = 1, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::less<T> >
+{ enum { Cost = 1, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::greater_equal<T> >
+{ enum { Cost = 1, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::less_equal<T> >
+{ enum { Cost = 1, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::equal_to<T> >
+{ enum { Cost = 1, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::not_equal_to<T> >
+{ enum { Cost = 1, PacketAccess = false }; };
+
+#if (__cplusplus < 201103L) && (EIGEN_COMP_MSVC <= 1900)
+// std::binder* are deprecated since c++11 and will be removed in c++17
+template<typename T>
+struct functor_traits<std::binder2nd<T> >
+{ enum { Cost = functor_traits<T>::Cost, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::binder1st<T> >
+{ enum { Cost = functor_traits<T>::Cost, PacketAccess = false }; };
+#endif
+
+template<typename T>
+struct functor_traits<std::unary_negate<T> >
+{ enum { Cost = 1 + functor_traits<T>::Cost, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::binary_negate<T> >
+{ enum { Cost = 1 + functor_traits<T>::Cost, PacketAccess = false }; };
+
+#ifdef EIGEN_STDEXT_SUPPORT
+
+template<typename T0,typename T1>
+struct functor_traits<std::project1st<T0,T1> >
+{ enum { Cost = 0, PacketAccess = false }; };
+
+template<typename T0,typename T1>
+struct functor_traits<std::project2nd<T0,T1> >
+{ enum { Cost = 0, PacketAccess = false }; };
+
+template<typename T0,typename T1>
+struct functor_traits<std::select2nd<std::pair<T0,T1> > >
+{ enum { Cost = 0, PacketAccess = false }; };
+
+template<typename T0,typename T1>
+struct functor_traits<std::select1st<std::pair<T0,T1> > >
+{ enum { Cost = 0, PacketAccess = false }; };
+
+template<typename T0,typename T1>
+struct functor_traits<std::unary_compose<T0,T1> >
+{ enum { Cost = functor_traits<T0>::Cost + functor_traits<T1>::Cost, PacketAccess = false }; };
+
+template<typename T0,typename T1,typename T2>
+struct functor_traits<std::binary_compose<T0,T1,T2> >
+{ enum { Cost = functor_traits<T0>::Cost + functor_traits<T1>::Cost + functor_traits<T2>::Cost, PacketAccess = false }; };
+
+#endif // EIGEN_STDEXT_SUPPORT
+
+// allow to add new functors and specializations of functor_traits from outside Eigen.
+// this macro is really needed because functor_traits must be specialized after it is declared but before it is used...
+#ifdef EIGEN_FUNCTORS_PLUGIN
+#include EIGEN_FUNCTORS_PLUGIN
+#endif
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_STL_FUNCTORS_H
diff --git a/eigen/Eigen/src/Core/functors/TernaryFunctors.h b/eigen/Eigen/src/Core/functors/TernaryFunctors.h
new file mode 100644
index 0000000..b254e96
--- /dev/null
+++ b/eigen/Eigen/src/Core/functors/TernaryFunctors.h
@@ -0,0 +1,25 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Eugene Brevdo <ebrevdo@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_TERNARY_FUNCTORS_H
+#define EIGEN_TERNARY_FUNCTORS_H
+
+namespace Eigen {
+
+namespace internal {
+
+//---------- associative ternary functors ----------
+
+
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TERNARY_FUNCTORS_H
diff --git a/eigen/Eigen/src/Core/functors/UnaryFunctors.h b/eigen/Eigen/src/Core/functors/UnaryFunctors.h
new file mode 100644
index 0000000..bfc0465
--- /dev/null
+++ b/eigen/Eigen/src/Core/functors/UnaryFunctors.h
@@ -0,0 +1,830 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2016 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_UNARY_FUNCTORS_H
+#define EIGEN_UNARY_FUNCTORS_H
+
+namespace Eigen {
+
+namespace internal {
+
+/** \internal
+  * \brief Template functor to compute the opposite of a scalar
+  *
+  * \sa class CwiseUnaryOp, MatrixBase::operator-
+  */
+template<typename Scalar> struct scalar_opposite_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_opposite_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { return -a; }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
+  { return internal::pnegate(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_opposite_op<Scalar> >
+{ enum {
+    Cost = NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasNegate };
+};
+
+/** \internal
+  * \brief Template functor to compute the absolute value of a scalar
+  *
+  * \sa class CwiseUnaryOp, Cwise::abs
+  */
+template<typename Scalar> struct scalar_abs_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_abs_op)
+  typedef typename NumTraits<Scalar>::Real result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a) const { return numext::abs(a); }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
+  { return internal::pabs(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_abs_op<Scalar> >
+{
+  enum {
+    Cost = NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasAbs
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the score of a scalar, to chose a pivot
+  *
+  * \sa class CwiseUnaryOp
+  */
+template<typename Scalar> struct scalar_score_coeff_op : scalar_abs_op<Scalar>
+{
+  typedef void Score_is_abs;
+};
+template<typename Scalar>
+struct functor_traits<scalar_score_coeff_op<Scalar> > : functor_traits<scalar_abs_op<Scalar> > {};
+
+/* Avoid recomputing abs when we know the score and they are the same. Not a true Eigen functor.  */
+template<typename Scalar, typename=void> struct abs_knowing_score
+{
+  EIGEN_EMPTY_STRUCT_CTOR(abs_knowing_score)
+  typedef typename NumTraits<Scalar>::Real result_type;
+  template<typename Score>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a, const Score&) const { return numext::abs(a); }
+};
+template<typename Scalar> struct abs_knowing_score<Scalar, typename scalar_score_coeff_op<Scalar>::Score_is_abs>
+{
+  EIGEN_EMPTY_STRUCT_CTOR(abs_knowing_score)
+  typedef typename NumTraits<Scalar>::Real result_type;
+  template<typename Scal>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const Scal&, const result_type& a) const { return a; }
+};
+
+/** \internal
+  * \brief Template functor to compute the squared absolute value of a scalar
+  *
+  * \sa class CwiseUnaryOp, Cwise::abs2
+  */
+template<typename Scalar> struct scalar_abs2_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_abs2_op)
+  typedef typename NumTraits<Scalar>::Real result_type;
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a) const { return numext::abs2(a); }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
+  { return internal::pmul(a,a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_abs2_op<Scalar> >
+{ enum { Cost = NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasAbs2 }; };
+
+/** \internal
+  * \brief Template functor to compute the conjugate of a complex value
+  *
+  * \sa class CwiseUnaryOp, MatrixBase::conjugate()
+  */
+template<typename Scalar> struct scalar_conjugate_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_conjugate_op)
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { using numext::conj; return conj(a); }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const { return internal::pconj(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_conjugate_op<Scalar> >
+{
+  enum {
+    Cost = NumTraits<Scalar>::IsComplex ? NumTraits<Scalar>::AddCost : 0,
+    PacketAccess = packet_traits<Scalar>::HasConj
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the phase angle of a complex
+  *
+  * \sa class CwiseUnaryOp, Cwise::arg
+  */
+template<typename Scalar> struct scalar_arg_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_arg_op)
+  typedef typename NumTraits<Scalar>::Real result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a) const { using numext::arg; return arg(a); }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
+  { return internal::parg(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_arg_op<Scalar> >
+{
+  enum {
+    Cost = NumTraits<Scalar>::IsComplex ? 5 * NumTraits<Scalar>::MulCost : NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasArg
+  };
+};
+/** \internal
+  * \brief Template functor to cast a scalar to another type
+  *
+  * \sa class CwiseUnaryOp, MatrixBase::cast()
+  */
+template<typename Scalar, typename NewType>
+struct scalar_cast_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
+  typedef NewType result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const NewType operator() (const Scalar& a) const { return cast<Scalar, NewType>(a); }
+};
+template<typename Scalar, typename NewType>
+struct functor_traits<scalar_cast_op<Scalar,NewType> >
+{ enum { Cost = is_same<Scalar, NewType>::value ? 0 : NumTraits<NewType>::AddCost, PacketAccess = false }; };
+
+/** \internal
+  * \brief Template functor to extract the real part of a complex
+  *
+  * \sa class CwiseUnaryOp, MatrixBase::real()
+  */
+template<typename Scalar>
+struct scalar_real_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_real_op)
+  typedef typename NumTraits<Scalar>::Real result_type;
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const { return numext::real(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_real_op<Scalar> >
+{ enum { Cost = 0, PacketAccess = false }; };
+
+/** \internal
+  * \brief Template functor to extract the imaginary part of a complex
+  *
+  * \sa class CwiseUnaryOp, MatrixBase::imag()
+  */
+template<typename Scalar>
+struct scalar_imag_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_imag_op)
+  typedef typename NumTraits<Scalar>::Real result_type;
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const { return numext::imag(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_imag_op<Scalar> >
+{ enum { Cost = 0, PacketAccess = false }; };
+
+/** \internal
+  * \brief Template functor to extract the real part of a complex as a reference
+  *
+  * \sa class CwiseUnaryOp, MatrixBase::real()
+  */
+template<typename Scalar>
+struct scalar_real_ref_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_real_ref_op)
+  typedef typename NumTraits<Scalar>::Real result_type;
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE result_type& operator() (const Scalar& a) const { return numext::real_ref(*const_cast<Scalar*>(&a)); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_real_ref_op<Scalar> >
+{ enum { Cost = 0, PacketAccess = false }; };
+
+/** \internal
+  * \brief Template functor to extract the imaginary part of a complex as a reference
+  *
+  * \sa class CwiseUnaryOp, MatrixBase::imag()
+  */
+template<typename Scalar>
+struct scalar_imag_ref_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_imag_ref_op)
+  typedef typename NumTraits<Scalar>::Real result_type;
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE result_type& operator() (const Scalar& a) const { return numext::imag_ref(*const_cast<Scalar*>(&a)); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_imag_ref_op<Scalar> >
+{ enum { Cost = 0, PacketAccess = false }; };
+
+/** \internal
+  *
+  * \brief Template functor to compute the exponential of a scalar
+  *
+  * \sa class CwiseUnaryOp, Cwise::exp()
+  */
+template<typename Scalar> struct scalar_exp_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_exp_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::exp(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pexp(a); }
+};
+template <typename Scalar>
+struct functor_traits<scalar_exp_op<Scalar> > {
+  enum {
+    PacketAccess = packet_traits<Scalar>::HasExp,
+    // The following numbers are based on the AVX implementation.
+#ifdef EIGEN_VECTORIZE_FMA
+    // Haswell can issue 2 add/mul/madd per cycle.
+    Cost =
+    (sizeof(Scalar) == 4
+     // float: 8 pmadd, 4 pmul, 2 padd/psub, 6 other
+     ? (8 * NumTraits<Scalar>::AddCost + 6 * NumTraits<Scalar>::MulCost)
+     // double: 7 pmadd, 5 pmul, 3 padd/psub, 1 div,  13 other
+     : (14 * NumTraits<Scalar>::AddCost +
+        6 * NumTraits<Scalar>::MulCost +
+        scalar_div_cost<Scalar,packet_traits<Scalar>::HasDiv>::value))
+#else
+    Cost =
+    (sizeof(Scalar) == 4
+     // float: 7 pmadd, 6 pmul, 4 padd/psub, 10 other
+     ? (21 * NumTraits<Scalar>::AddCost + 13 * NumTraits<Scalar>::MulCost)
+     // double: 7 pmadd, 5 pmul, 3 padd/psub, 1 div,  13 other
+     : (23 * NumTraits<Scalar>::AddCost +
+        12 * NumTraits<Scalar>::MulCost +
+        scalar_div_cost<Scalar,packet_traits<Scalar>::HasDiv>::value))
+#endif
+  };
+};
+
+/** \internal
+  *
+  * \brief Template functor to compute the exponential of a scalar - 1.
+  *
+  * \sa class CwiseUnaryOp, ArrayBase::expm1()
+  */
+template<typename Scalar> struct scalar_expm1_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_expm1_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::expm1(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pexpm1(a); }
+};
+template <typename Scalar>
+struct functor_traits<scalar_expm1_op<Scalar> > {
+  enum {
+    PacketAccess = packet_traits<Scalar>::HasExpm1,
+    Cost = functor_traits<scalar_exp_op<Scalar> >::Cost // TODO measure cost of expm1
+  };
+};
+
+/** \internal
+  *
+  * \brief Template functor to compute the logarithm of a scalar
+  *
+  * \sa class CwiseUnaryOp, ArrayBase::log()
+  */
+template<typename Scalar> struct scalar_log_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_log_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::log(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::plog(a); }
+};
+template <typename Scalar>
+struct functor_traits<scalar_log_op<Scalar> > {
+  enum {
+    PacketAccess = packet_traits<Scalar>::HasLog,
+    Cost =
+    (PacketAccess
+     // The following numbers are based on the AVX implementation.
+#ifdef EIGEN_VECTORIZE_FMA
+     // 8 pmadd, 6 pmul, 8 padd/psub, 16 other, can issue 2 add/mul/madd per cycle.
+     ? (20 * NumTraits<Scalar>::AddCost + 7 * NumTraits<Scalar>::MulCost)
+#else
+     // 8 pmadd, 6 pmul, 8 padd/psub, 20 other
+     ? (36 * NumTraits<Scalar>::AddCost + 14 * NumTraits<Scalar>::MulCost)
+#endif
+     // Measured cost of std::log.
+     : sizeof(Scalar)==4 ? 40 : 85)
+  };
+};
+
+/** \internal
+  *
+  * \brief Template functor to compute the logarithm of 1 plus a scalar value
+  *
+  * \sa class CwiseUnaryOp, ArrayBase::log1p()
+  */
+template<typename Scalar> struct scalar_log1p_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_log1p_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::log1p(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::plog1p(a); }
+};
+template <typename Scalar>
+struct functor_traits<scalar_log1p_op<Scalar> > {
+  enum {
+    PacketAccess = packet_traits<Scalar>::HasLog1p,
+    Cost = functor_traits<scalar_log_op<Scalar> >::Cost // TODO measure cost of log1p
+  };
+};
+
+/** \internal
+  *
+  * \brief Template functor to compute the base-10 logarithm of a scalar
+  *
+  * \sa class CwiseUnaryOp, Cwise::log10()
+  */
+template<typename Scalar> struct scalar_log10_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_log10_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { EIGEN_USING_STD_MATH(log10) return log10(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::plog10(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_log10_op<Scalar> >
+{ enum { Cost = 5 * NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasLog10 }; };
+
+/** \internal
+  * \brief Template functor to compute the square root of a scalar
+  * \sa class CwiseUnaryOp, Cwise::sqrt()
+  */
+template<typename Scalar> struct scalar_sqrt_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_sqrt_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::sqrt(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::psqrt(a); }
+};
+template <typename Scalar>
+struct functor_traits<scalar_sqrt_op<Scalar> > {
+  enum {
+#if EIGEN_FAST_MATH
+    // The following numbers are based on the AVX implementation.
+    Cost = (sizeof(Scalar) == 8 ? 28
+                                // 4 pmul, 1 pmadd, 3 other
+                                : (3 * NumTraits<Scalar>::AddCost +
+                                   5 * NumTraits<Scalar>::MulCost)),
+#else
+    // The following numbers are based on min VSQRT throughput on Haswell.
+    Cost = (sizeof(Scalar) == 8 ? 28 : 14),
+#endif
+    PacketAccess = packet_traits<Scalar>::HasSqrt
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the reciprocal square root of a scalar
+  * \sa class CwiseUnaryOp, Cwise::rsqrt()
+  */
+template<typename Scalar> struct scalar_rsqrt_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_rsqrt_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return Scalar(1)/numext::sqrt(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::prsqrt(a); }
+};
+
+template<typename Scalar>
+struct functor_traits<scalar_rsqrt_op<Scalar> >
+{ enum {
+    Cost = 5 * NumTraits<Scalar>::MulCost,
+    PacketAccess = packet_traits<Scalar>::HasRsqrt
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the cosine of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::cos()
+  */
+template<typename Scalar> struct scalar_cos_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cos_op)
+  EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return numext::cos(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pcos(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_cos_op<Scalar> >
+{
+  enum {
+    Cost = 5 * NumTraits<Scalar>::MulCost,
+    PacketAccess = packet_traits<Scalar>::HasCos
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the sine of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::sin()
+  */
+template<typename Scalar> struct scalar_sin_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_sin_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::sin(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::psin(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_sin_op<Scalar> >
+{
+  enum {
+    Cost = 5 * NumTraits<Scalar>::MulCost,
+    PacketAccess = packet_traits<Scalar>::HasSin
+  };
+};
+
+
+/** \internal
+  * \brief Template functor to compute the tan of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::tan()
+  */
+template<typename Scalar> struct scalar_tan_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_tan_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::tan(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::ptan(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_tan_op<Scalar> >
+{
+  enum {
+    Cost = 5 * NumTraits<Scalar>::MulCost,
+    PacketAccess = packet_traits<Scalar>::HasTan
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the arc cosine of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::acos()
+  */
+template<typename Scalar> struct scalar_acos_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_acos_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::acos(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pacos(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_acos_op<Scalar> >
+{
+  enum {
+    Cost = 5 * NumTraits<Scalar>::MulCost,
+    PacketAccess = packet_traits<Scalar>::HasACos
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the arc sine of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::asin()
+  */
+template<typename Scalar> struct scalar_asin_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_asin_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::asin(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pasin(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_asin_op<Scalar> >
+{
+  enum {
+    Cost = 5 * NumTraits<Scalar>::MulCost,
+    PacketAccess = packet_traits<Scalar>::HasASin
+  };
+};
+
+
+/** \internal
+  * \brief Template functor to compute the atan of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::atan()
+  */
+template<typename Scalar> struct scalar_atan_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_atan_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::atan(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::patan(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_atan_op<Scalar> >
+{
+  enum {
+    Cost = 5 * NumTraits<Scalar>::MulCost,
+    PacketAccess = packet_traits<Scalar>::HasATan
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the tanh of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::tanh()
+  */
+template <typename Scalar>
+struct scalar_tanh_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_tanh_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator()(const Scalar& a) const { return numext::tanh(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& x) const { return ptanh(x); }
+};
+
+template <typename Scalar>
+struct functor_traits<scalar_tanh_op<Scalar> > {
+  enum {
+    PacketAccess = packet_traits<Scalar>::HasTanh,
+    Cost = ( (EIGEN_FAST_MATH && is_same<Scalar,float>::value)
+// The following numbers are based on the AVX implementation,
+#ifdef EIGEN_VECTORIZE_FMA
+                // Haswell can issue 2 add/mul/madd per cycle.
+                // 9 pmadd, 2 pmul, 1 div, 2 other
+                ? (2 * NumTraits<Scalar>::AddCost +
+                   6 * NumTraits<Scalar>::MulCost +
+                   scalar_div_cost<Scalar,packet_traits<Scalar>::HasDiv>::value)
+#else
+                ? (11 * NumTraits<Scalar>::AddCost +
+                   11 * NumTraits<Scalar>::MulCost +
+                   scalar_div_cost<Scalar,packet_traits<Scalar>::HasDiv>::value)
+#endif
+                // This number assumes a naive implementation of tanh
+                : (6 * NumTraits<Scalar>::AddCost +
+                   3 * NumTraits<Scalar>::MulCost +
+                   2 * scalar_div_cost<Scalar,packet_traits<Scalar>::HasDiv>::value +
+                   functor_traits<scalar_exp_op<Scalar> >::Cost))
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the sinh of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::sinh()
+  */
+template<typename Scalar> struct scalar_sinh_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_sinh_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::sinh(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::psinh(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_sinh_op<Scalar> >
+{
+  enum {
+    Cost = 5 * NumTraits<Scalar>::MulCost,
+    PacketAccess = packet_traits<Scalar>::HasSinh
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the cosh of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::cosh()
+  */
+template<typename Scalar> struct scalar_cosh_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cosh_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::cosh(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pcosh(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_cosh_op<Scalar> >
+{
+  enum {
+    Cost = 5 * NumTraits<Scalar>::MulCost,
+    PacketAccess = packet_traits<Scalar>::HasCosh
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the inverse of a scalar
+  * \sa class CwiseUnaryOp, Cwise::inverse()
+  */
+template<typename Scalar>
+struct scalar_inverse_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_inverse_op)
+  EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return Scalar(1)/a; }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC inline const Packet packetOp(const Packet& a) const
+  { return internal::pdiv(pset1<Packet>(Scalar(1)),a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_inverse_op<Scalar> >
+{ enum { Cost = NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasDiv }; };
+
+/** \internal
+  * \brief Template functor to compute the square of a scalar
+  * \sa class CwiseUnaryOp, Cwise::square()
+  */
+template<typename Scalar>
+struct scalar_square_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_square_op)
+  EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return a*a; }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC inline const Packet packetOp(const Packet& a) const
+  { return internal::pmul(a,a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_square_op<Scalar> >
+{ enum { Cost = NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasMul }; };
+
+/** \internal
+  * \brief Template functor to compute the cube of a scalar
+  * \sa class CwiseUnaryOp, Cwise::cube()
+  */
+template<typename Scalar>
+struct scalar_cube_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cube_op)
+  EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return a*a*a; }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC inline const Packet packetOp(const Packet& a) const
+  { return internal::pmul(a,pmul(a,a)); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_cube_op<Scalar> >
+{ enum { Cost = 2*NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasMul }; };
+
+/** \internal
+  * \brief Template functor to compute the rounded value of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::round()
+  */
+template<typename Scalar> struct scalar_round_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_round_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { return numext::round(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pround(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_round_op<Scalar> >
+{
+  enum {
+    Cost = NumTraits<Scalar>::MulCost,
+    PacketAccess = packet_traits<Scalar>::HasRound
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the floor of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::floor()
+  */
+template<typename Scalar> struct scalar_floor_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_floor_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { return numext::floor(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pfloor(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_floor_op<Scalar> >
+{
+  enum {
+    Cost = NumTraits<Scalar>::MulCost,
+    PacketAccess = packet_traits<Scalar>::HasFloor
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the ceil of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::ceil()
+  */
+template<typename Scalar> struct scalar_ceil_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_ceil_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { return numext::ceil(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pceil(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_ceil_op<Scalar> >
+{
+  enum {
+    Cost = NumTraits<Scalar>::MulCost,
+    PacketAccess = packet_traits<Scalar>::HasCeil
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute whether a scalar is NaN
+  * \sa class CwiseUnaryOp, ArrayBase::isnan()
+  */
+template<typename Scalar> struct scalar_isnan_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_isnan_op)
+  typedef bool result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const {
+#if defined(__SYCL_DEVICE_ONLY__)
+    return numext::isnan(a);
+#else  
+    return (numext::isnan)(a);
+#endif
+  }
+};
+template<typename Scalar>
+struct functor_traits<scalar_isnan_op<Scalar> >
+{
+  enum {
+    Cost = NumTraits<Scalar>::MulCost,
+    PacketAccess = false
+  };
+};
+
+/** \internal
+  * \brief Template functor to check whether a scalar is +/-inf
+  * \sa class CwiseUnaryOp, ArrayBase::isinf()
+  */
+template<typename Scalar> struct scalar_isinf_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_isinf_op)
+  typedef bool result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const {
+#if defined(__SYCL_DEVICE_ONLY__)
+    return numext::isinf(a);
+#else
+    return (numext::isinf)(a);
+#endif
+  }
+};
+template<typename Scalar>
+struct functor_traits<scalar_isinf_op<Scalar> >
+{
+  enum {
+    Cost = NumTraits<Scalar>::MulCost,
+    PacketAccess = false
+  };
+};
+
+/** \internal
+  * \brief Template functor to check whether a scalar has a finite value
+  * \sa class CwiseUnaryOp, ArrayBase::isfinite()
+  */
+template<typename Scalar> struct scalar_isfinite_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_isfinite_op)
+  typedef bool result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const {
+#if defined(__SYCL_DEVICE_ONLY__)
+    return numext::isfinite(a);
+#else
+    return (numext::isfinite)(a);
+#endif
+  }
+};
+template<typename Scalar>
+struct functor_traits<scalar_isfinite_op<Scalar> >
+{
+  enum {
+    Cost = NumTraits<Scalar>::MulCost,
+    PacketAccess = false
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the logical not of a boolean
+  *
+  * \sa class CwiseUnaryOp, ArrayBase::operator!
+  */
+template<typename Scalar> struct scalar_boolean_not_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_boolean_not_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator() (const bool& a) const { return !a; }
+};
+template<typename Scalar>
+struct functor_traits<scalar_boolean_not_op<Scalar> > {
+  enum {
+    Cost = NumTraits<bool>::AddCost,
+    PacketAccess = false
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the signum of a scalar
+  * \sa class CwiseUnaryOp, Cwise::sign()
+  */
+template<typename Scalar,bool iscpx=(NumTraits<Scalar>::IsComplex!=0) > struct scalar_sign_op;
+template<typename Scalar>
+struct scalar_sign_op<Scalar,false> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_sign_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const
+  {
+      return Scalar( (a>Scalar(0)) - (a<Scalar(0)) );
+  }
+  //TODO
+  //template <typename Packet>
+  //EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::psign(a); }
+};
+template<typename Scalar>
+struct scalar_sign_op<Scalar,true> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_sign_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const
+  {
+    typedef typename NumTraits<Scalar>::Real real_type;
+    real_type aa = numext::abs(a);
+    if (aa==real_type(0))
+      return Scalar(0);
+    aa = real_type(1)/aa;
+    return Scalar(real(a)*aa, imag(a)*aa );
+  }
+  //TODO
+  //template <typename Packet>
+  //EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::psign(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_sign_op<Scalar> >
+{ enum {
+    Cost = 
+        NumTraits<Scalar>::IsComplex
+        ? ( 8*NumTraits<Scalar>::MulCost  ) // roughly
+        : ( 3*NumTraits<Scalar>::AddCost),
+    PacketAccess = packet_traits<Scalar>::HasSign
+  };
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_FUNCTORS_H
diff --git a/eigen/Eigen/src/Core/products/CMakeLists.txt b/eigen/Eigen/src/Core/products/CMakeLists.txt
deleted file mode 100644
index 21fc94a..0000000
--- a/eigen/Eigen/src/Core/products/CMakeLists.txt
+++ /dev/null
@@ -1,6 +0,0 @@
-FILE(GLOB Eigen_Core_Product_SRCS "*.h")
-
-INSTALL(FILES
-  ${Eigen_Core_Product_SRCS}
-  DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/products COMPONENT Devel
-  )
diff --git a/eigen/Eigen/src/Core/products/CoeffBasedProduct.h b/eigen/Eigen/src/Core/products/CoeffBasedProduct.h
deleted file mode 100644
index 2a9d65b..0000000
--- a/eigen/Eigen/src/Core/products/CoeffBasedProduct.h
+++ /dev/null
@@ -1,476 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
-// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-#ifndef EIGEN_COEFFBASED_PRODUCT_H
-#define EIGEN_COEFFBASED_PRODUCT_H
-
-namespace Eigen { 
-
-namespace internal {
-
-/*********************************************************************************
-*  Coefficient based product implementation.
-*  It is designed for the following use cases:
-*  - small fixed sizes
-*  - lazy products
-*********************************************************************************/
-
-/* Since the all the dimensions of the product are small, here we can rely
- * on the generic Assign mechanism to evaluate the product per coeff (or packet).
- *
- * Note that here the inner-loops should always be unrolled.
- */
-
-template<int Traversal, int UnrollingIndex, typename Lhs, typename Rhs, typename RetScalar>
-struct product_coeff_impl;
-
-template<int StorageOrder, int UnrollingIndex, typename Lhs, typename Rhs, typename Packet, int LoadMode>
-struct product_packet_impl;
-
-template<typename LhsNested, typename RhsNested, int NestingFlags>
-struct traits<CoeffBasedProduct<LhsNested,RhsNested,NestingFlags> >
-{
-  typedef MatrixXpr XprKind;
-  typedef typename remove_all<LhsNested>::type _LhsNested;
-  typedef typename remove_all<RhsNested>::type _RhsNested;
-  typedef typename scalar_product_traits<typename _LhsNested::Scalar, typename _RhsNested::Scalar>::ReturnType Scalar;
-  typedef typename promote_storage_type<typename traits<_LhsNested>::StorageKind,
-                                           typename traits<_RhsNested>::StorageKind>::ret StorageKind;
-  typedef typename promote_index_type<typename traits<_LhsNested>::Index,
-                                         typename traits<_RhsNested>::Index>::type Index;
-
-  enum {
-      LhsCoeffReadCost = _LhsNested::CoeffReadCost,
-      RhsCoeffReadCost = _RhsNested::CoeffReadCost,
-      LhsFlags = _LhsNested::Flags,
-      RhsFlags = _RhsNested::Flags,
-
-      RowsAtCompileTime = _LhsNested::RowsAtCompileTime,
-      ColsAtCompileTime = _RhsNested::ColsAtCompileTime,
-      InnerSize = EIGEN_SIZE_MIN_PREFER_FIXED(_LhsNested::ColsAtCompileTime, _RhsNested::RowsAtCompileTime),
-
-      MaxRowsAtCompileTime = _LhsNested::MaxRowsAtCompileTime,
-      MaxColsAtCompileTime = _RhsNested::MaxColsAtCompileTime,
-
-      LhsRowMajor = LhsFlags & RowMajorBit,
-      RhsRowMajor = RhsFlags & RowMajorBit,
-
-      SameType = is_same<typename _LhsNested::Scalar,typename _RhsNested::Scalar>::value,
-
-      CanVectorizeRhs = RhsRowMajor && (RhsFlags & PacketAccessBit)
-                      && (ColsAtCompileTime == Dynamic
-                          || ( (ColsAtCompileTime % packet_traits<Scalar>::size) == 0
-                              && (RhsFlags&AlignedBit)
-                             )
-                         ),
-
-      CanVectorizeLhs = (!LhsRowMajor) && (LhsFlags & PacketAccessBit)
-                      && (RowsAtCompileTime == Dynamic
-                          || ( (RowsAtCompileTime % packet_traits<Scalar>::size) == 0
-                              && (LhsFlags&AlignedBit)
-                             )
-                         ),
-
-      EvalToRowMajor = (MaxRowsAtCompileTime==1&&MaxColsAtCompileTime!=1) ? 1
-                     : (MaxColsAtCompileTime==1&&MaxRowsAtCompileTime!=1) ? 0
-                     : (RhsRowMajor && !CanVectorizeLhs),
-
-      Flags = ((unsigned int)(LhsFlags | RhsFlags) & HereditaryBits & ~RowMajorBit)
-            | (EvalToRowMajor ? RowMajorBit : 0)
-            | NestingFlags
-            | (LhsFlags & RhsFlags & AlignedBit)
-            // TODO enable vectorization for mixed types
-            | (SameType && (CanVectorizeLhs || CanVectorizeRhs) ? PacketAccessBit : 0),
-
-      CoeffReadCost = InnerSize == Dynamic ? Dynamic
-                    : InnerSize == 0 ? 0
-                    : InnerSize * (NumTraits<Scalar>::MulCost + LhsCoeffReadCost + RhsCoeffReadCost)
-                      + (InnerSize - 1) * NumTraits<Scalar>::AddCost,
-
-      /* CanVectorizeInner deserves special explanation. It does not affect the product flags. It is not used outside
-      * of Product. If the Product itself is not a packet-access expression, there is still a chance that the inner
-      * loop of the product might be vectorized. This is the meaning of CanVectorizeInner. Since it doesn't affect
-      * the Flags, it is safe to make this value depend on ActualPacketAccessBit, that doesn't affect the ABI.
-      */
-      CanVectorizeInner =    SameType
-                          && LhsRowMajor
-                          && (!RhsRowMajor)
-                          && (LhsFlags & RhsFlags & ActualPacketAccessBit)
-                          && (LhsFlags & RhsFlags & AlignedBit)
-                          && (InnerSize % packet_traits<Scalar>::size == 0)
-    };
-};
-
-} // end namespace internal
-
-template<typename LhsNested, typename RhsNested, int NestingFlags>
-class CoeffBasedProduct
-  : internal::no_assignment_operator,
-    public MatrixBase<CoeffBasedProduct<LhsNested, RhsNested, NestingFlags> >
-{
-  public:
-
-    typedef MatrixBase<CoeffBasedProduct> Base;
-    EIGEN_DENSE_PUBLIC_INTERFACE(CoeffBasedProduct)
-    typedef typename Base::PlainObject PlainObject;
-
-  private:
-
-    typedef typename internal::traits<CoeffBasedProduct>::_LhsNested _LhsNested;
-    typedef typename internal::traits<CoeffBasedProduct>::_RhsNested _RhsNested;
-
-    enum {
-      PacketSize = internal::packet_traits<Scalar>::size,
-      InnerSize  = internal::traits<CoeffBasedProduct>::InnerSize,
-      Unroll = CoeffReadCost != Dynamic && CoeffReadCost <= EIGEN_UNROLLING_LIMIT,
-      CanVectorizeInner = internal::traits<CoeffBasedProduct>::CanVectorizeInner
-    };
-
-    typedef internal::product_coeff_impl<CanVectorizeInner ? InnerVectorizedTraversal : DefaultTraversal,
-                                   Unroll ? InnerSize : Dynamic,
-                                   _LhsNested, _RhsNested, Scalar> ScalarCoeffImpl;
-
-    typedef CoeffBasedProduct<LhsNested,RhsNested,NestByRefBit> LazyCoeffBasedProductType;
-
-  public:
-
-    inline CoeffBasedProduct(const CoeffBasedProduct& other)
-      : Base(), m_lhs(other.m_lhs), m_rhs(other.m_rhs)
-    {}
-
-    template<typename Lhs, typename Rhs>
-    inline CoeffBasedProduct(const Lhs& lhs, const Rhs& rhs)
-      : m_lhs(lhs), m_rhs(rhs)
-    {
-      // we don't allow taking products of matrices of different real types, as that wouldn't be vectorizable.
-      // We still allow to mix T and complex<T>.
-      EIGEN_STATIC_ASSERT((internal::scalar_product_traits<typename Lhs::RealScalar, typename Rhs::RealScalar>::Defined),
-        YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
-      eigen_assert(lhs.cols() == rhs.rows()
-        && "invalid matrix product"
-        && "if you wanted a coeff-wise or a dot product use the respective explicit functions");
-    }
-
-    EIGEN_STRONG_INLINE Index rows() const { return m_lhs.rows(); }
-    EIGEN_STRONG_INLINE Index cols() const { return m_rhs.cols(); }
-
-    EIGEN_STRONG_INLINE const Scalar coeff(Index row, Index col) const
-    {
-      Scalar res;
-      ScalarCoeffImpl::run(row, col, m_lhs, m_rhs, res);
-      return res;
-    }
-
-    /* Allow index-based non-packet access. It is impossible though to allow index-based packed access,
-     * which is why we don't set the LinearAccessBit.
-     */
-    EIGEN_STRONG_INLINE const Scalar coeff(Index index) const
-    {
-      Scalar res;
-      const Index row = RowsAtCompileTime == 1 ? 0 : index;
-      const Index col = RowsAtCompileTime == 1 ? index : 0;
-      ScalarCoeffImpl::run(row, col, m_lhs, m_rhs, res);
-      return res;
-    }
-
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE const PacketScalar packet(Index row, Index col) const
-    {
-      PacketScalar res;
-      internal::product_packet_impl<Flags&RowMajorBit ? RowMajor : ColMajor,
-                              Unroll ? InnerSize : Dynamic,
-                              _LhsNested, _RhsNested, PacketScalar, LoadMode>
-        ::run(row, col, m_lhs, m_rhs, res);
-      return res;
-    }
-
-    // Implicit conversion to the nested type (trigger the evaluation of the product)
-    EIGEN_STRONG_INLINE operator const PlainObject& () const
-    {
-      m_result.lazyAssign(*this);
-      return m_result;
-    }
-
-    const _LhsNested& lhs() const { return m_lhs; }
-    const _RhsNested& rhs() const { return m_rhs; }
-
-    const Diagonal<const LazyCoeffBasedProductType,0> diagonal() const
-    { return reinterpret_cast<const LazyCoeffBasedProductType&>(*this); }
-
-    template<int DiagonalIndex>
-    const Diagonal<const LazyCoeffBasedProductType,DiagonalIndex> diagonal() const
-    { return reinterpret_cast<const LazyCoeffBasedProductType&>(*this); }
-
-    const Diagonal<const LazyCoeffBasedProductType,Dynamic> diagonal(Index index) const
-    { return reinterpret_cast<const LazyCoeffBasedProductType&>(*this).diagonal(index); }
-
-  protected:
-    typename internal::add_const_on_value_type<LhsNested>::type m_lhs;
-    typename internal::add_const_on_value_type<RhsNested>::type m_rhs;
-
-    mutable PlainObject m_result;
-};
-
-namespace internal {
-
-// here we need to overload the nested rule for products
-// such that the nested type is a const reference to a plain matrix
-template<typename Lhs, typename Rhs, int N, typename PlainObject>
-struct nested<CoeffBasedProduct<Lhs,Rhs,EvalBeforeNestingBit|EvalBeforeAssigningBit>, N, PlainObject>
-{
-  typedef PlainObject const& type;
-};
-
-/***************************************************************************
-* Normal product .coeff() implementation (with meta-unrolling)
-***************************************************************************/
-
-/**************************************
-*** Scalar path  - no vectorization ***
-**************************************/
-
-template<int UnrollingIndex, typename Lhs, typename Rhs, typename RetScalar>
-struct product_coeff_impl<DefaultTraversal, UnrollingIndex, Lhs, Rhs, RetScalar>
-{
-  typedef typename Lhs::Index Index;
-  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, RetScalar &res)
-  {
-    product_coeff_impl<DefaultTraversal, UnrollingIndex-1, Lhs, Rhs, RetScalar>::run(row, col, lhs, rhs, res);
-    res += lhs.coeff(row, UnrollingIndex-1) * rhs.coeff(UnrollingIndex-1, col);
-  }
-};
-
-template<typename Lhs, typename Rhs, typename RetScalar>
-struct product_coeff_impl<DefaultTraversal, 1, Lhs, Rhs, RetScalar>
-{
-  typedef typename Lhs::Index Index;
-  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, RetScalar &res)
-  {
-    res = lhs.coeff(row, 0) * rhs.coeff(0, col);
-  }
-};
-
-template<typename Lhs, typename Rhs, typename RetScalar>
-struct product_coeff_impl<DefaultTraversal, 0, Lhs, Rhs, RetScalar>
-{
-  typedef typename Lhs::Index Index;
-  static EIGEN_STRONG_INLINE void run(Index /*row*/, Index /*col*/, const Lhs& /*lhs*/, const Rhs& /*rhs*/, RetScalar &res)
-  {
-    res = RetScalar(0);
-  }
-};
-
-template<typename Lhs, typename Rhs, typename RetScalar>
-struct product_coeff_impl<DefaultTraversal, Dynamic, Lhs, Rhs, RetScalar>
-{
-  typedef typename Lhs::Index Index;
-  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, RetScalar& res)
-  {
-    res = (lhs.row(row).transpose().cwiseProduct( rhs.col(col) )).sum();
-  }
-};
-
-/*******************************************
-*** Scalar path with inner vectorization ***
-*******************************************/
-
-template<int UnrollingIndex, typename Lhs, typename Rhs, typename Packet>
-struct product_coeff_vectorized_unroller
-{
-  typedef typename Lhs::Index Index;
-  enum { PacketSize = packet_traits<typename Lhs::Scalar>::size };
-  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, typename Lhs::PacketScalar &pres)
-  {
-    product_coeff_vectorized_unroller<UnrollingIndex-PacketSize, Lhs, Rhs, Packet>::run(row, col, lhs, rhs, pres);
-    pres = padd(pres, pmul( lhs.template packet<Aligned>(row, UnrollingIndex) , rhs.template packet<Aligned>(UnrollingIndex, col) ));
-  }
-};
-
-template<typename Lhs, typename Rhs, typename Packet>
-struct product_coeff_vectorized_unroller<0, Lhs, Rhs, Packet>
-{
-  typedef typename Lhs::Index Index;
-  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, typename Lhs::PacketScalar &pres)
-  {
-    pres = pmul(lhs.template packet<Aligned>(row, 0) , rhs.template packet<Aligned>(0, col));
-  }
-};
-
-template<typename Lhs, typename Rhs, typename RetScalar>
-struct product_coeff_impl<InnerVectorizedTraversal, 0, Lhs, Rhs, RetScalar>
-{
-  typedef typename Lhs::Index Index;
-  static EIGEN_STRONG_INLINE void run(Index /*row*/, Index /*col*/, const Lhs& /*lhs*/, const Rhs& /*rhs*/, RetScalar &res)
-  {
-    res = 0;
-  }
-};
-
-template<int UnrollingIndex, typename Lhs, typename Rhs, typename RetScalar>
-struct product_coeff_impl<InnerVectorizedTraversal, UnrollingIndex, Lhs, Rhs, RetScalar>
-{
-  typedef typename Lhs::PacketScalar Packet;
-  typedef typename Lhs::Index Index;
-  enum { PacketSize = packet_traits<typename Lhs::Scalar>::size };
-  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, RetScalar &res)
-  {
-    Packet pres;
-    product_coeff_vectorized_unroller<UnrollingIndex-PacketSize, Lhs, Rhs, Packet>::run(row, col, lhs, rhs, pres);
-    res = predux(pres);
-  }
-};
-
-template<typename Lhs, typename Rhs, int LhsRows = Lhs::RowsAtCompileTime, int RhsCols = Rhs::ColsAtCompileTime>
-struct product_coeff_vectorized_dyn_selector
-{
-  typedef typename Lhs::Index Index;
-  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, typename Lhs::Scalar &res)
-  {
-    res = lhs.row(row).transpose().cwiseProduct(rhs.col(col)).sum();
-  }
-};
-
-// NOTE the 3 following specializations are because taking .col(0) on a vector is a bit slower
-// NOTE maybe they are now useless since we have a specialization for Block<Matrix>
-template<typename Lhs, typename Rhs, int RhsCols>
-struct product_coeff_vectorized_dyn_selector<Lhs,Rhs,1,RhsCols>
-{
-  typedef typename Lhs::Index Index;
-  static EIGEN_STRONG_INLINE void run(Index /*row*/, Index col, const Lhs& lhs, const Rhs& rhs, typename Lhs::Scalar &res)
-  {
-    res = lhs.transpose().cwiseProduct(rhs.col(col)).sum();
-  }
-};
-
-template<typename Lhs, typename Rhs, int LhsRows>
-struct product_coeff_vectorized_dyn_selector<Lhs,Rhs,LhsRows,1>
-{
-  typedef typename Lhs::Index Index;
-  static EIGEN_STRONG_INLINE void run(Index row, Index /*col*/, const Lhs& lhs, const Rhs& rhs, typename Lhs::Scalar &res)
-  {
-    res = lhs.row(row).transpose().cwiseProduct(rhs).sum();
-  }
-};
-
-template<typename Lhs, typename Rhs>
-struct product_coeff_vectorized_dyn_selector<Lhs,Rhs,1,1>
-{
-  typedef typename Lhs::Index Index;
-  static EIGEN_STRONG_INLINE void run(Index /*row*/, Index /*col*/, const Lhs& lhs, const Rhs& rhs, typename Lhs::Scalar &res)
-  {
-    res = lhs.transpose().cwiseProduct(rhs).sum();
-  }
-};
-
-template<typename Lhs, typename Rhs, typename RetScalar>
-struct product_coeff_impl<InnerVectorizedTraversal, Dynamic, Lhs, Rhs, RetScalar>
-{
-  typedef typename Lhs::Index Index;
-  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, typename Lhs::Scalar &res)
-  {
-    product_coeff_vectorized_dyn_selector<Lhs,Rhs>::run(row, col, lhs, rhs, res);
-  }
-};
-
-/*******************
-*** Packet path  ***
-*******************/
-
-template<int UnrollingIndex, typename Lhs, typename Rhs, typename Packet, int LoadMode>
-struct product_packet_impl<RowMajor, UnrollingIndex, Lhs, Rhs, Packet, LoadMode>
-{
-  typedef typename Lhs::Index Index;
-  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet &res)
-  {
-    product_packet_impl<RowMajor, UnrollingIndex-1, Lhs, Rhs, Packet, LoadMode>::run(row, col, lhs, rhs, res);
-    res =  pmadd(pset1<Packet>(lhs.coeff(row, UnrollingIndex-1)), rhs.template packet<LoadMode>(UnrollingIndex-1, col), res);
-  }
-};
-
-template<int UnrollingIndex, typename Lhs, typename Rhs, typename Packet, int LoadMode>
-struct product_packet_impl<ColMajor, UnrollingIndex, Lhs, Rhs, Packet, LoadMode>
-{
-  typedef typename Lhs::Index Index;
-  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet &res)
-  {
-    product_packet_impl<ColMajor, UnrollingIndex-1, Lhs, Rhs, Packet, LoadMode>::run(row, col, lhs, rhs, res);
-    res =  pmadd(lhs.template packet<LoadMode>(row, UnrollingIndex-1), pset1<Packet>(rhs.coeff(UnrollingIndex-1, col)), res);
-  }
-};
-
-template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
-struct product_packet_impl<RowMajor, 1, Lhs, Rhs, Packet, LoadMode>
-{
-  typedef typename Lhs::Index Index;
-  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet &res)
-  {
-    res = pmul(pset1<Packet>(lhs.coeff(row, 0)),rhs.template packet<LoadMode>(0, col));
-  }
-};
-
-template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
-struct product_packet_impl<ColMajor, 1, Lhs, Rhs, Packet, LoadMode>
-{
-  typedef typename Lhs::Index Index;
-  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet &res)
-  {
-    res = pmul(lhs.template packet<LoadMode>(row, 0), pset1<Packet>(rhs.coeff(0, col)));
-  }
-};
-
-template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
-struct product_packet_impl<RowMajor, 0, Lhs, Rhs, Packet, LoadMode>
-{
-  typedef typename Lhs::Index Index;
-  static EIGEN_STRONG_INLINE void run(Index /*row*/, Index /*col*/, const Lhs& /*lhs*/, const Rhs& /*rhs*/, Packet &res)
-  {
-    res = pset1<Packet>(0);
-  }
-};
-
-template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
-struct product_packet_impl<ColMajor, 0, Lhs, Rhs, Packet, LoadMode>
-{
-  typedef typename Lhs::Index Index;
-  static EIGEN_STRONG_INLINE void run(Index /*row*/, Index /*col*/, const Lhs& /*lhs*/, const Rhs& /*rhs*/, Packet &res)
-  {
-    res = pset1<Packet>(0);
-  }
-};
-
-template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
-struct product_packet_impl<RowMajor, Dynamic, Lhs, Rhs, Packet, LoadMode>
-{
-  typedef typename Lhs::Index Index;
-  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet& res)
-  {
-    res = pset1<Packet>(0);
-    for(Index i = 0; i < lhs.cols(); ++i)
-      res =  pmadd(pset1<Packet>(lhs.coeff(row, i)), rhs.template packet<LoadMode>(i, col), res);
-  }
-};
-
-template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
-struct product_packet_impl<ColMajor, Dynamic, Lhs, Rhs, Packet, LoadMode>
-{
-  typedef typename Lhs::Index Index;
-  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet& res)
-  {
-    res = pset1<Packet>(0);
-    for(Index i = 0; i < lhs.cols(); ++i)
-      res =  pmadd(lhs.template packet<LoadMode>(row, i), pset1<Packet>(rhs.coeff(i, col)), res);
-  }
-};
-
-} // end namespace internal
-
-} // end namespace Eigen
-
-#endif // EIGEN_COEFFBASED_PRODUCT_H
diff --git a/eigen/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/eigen/Eigen/src/Core/products/GeneralBlockPanelKernel.h
index bcdca5b..45230bc 100644
--- a/eigen/Eigen/src/Core/products/GeneralBlockPanelKernel.h
+++ b/eigen/Eigen/src/Core/products/GeneralBlockPanelKernel.h
@@ -10,8 +10,9 @@
 #ifndef EIGEN_GENERAL_BLOCK_PANEL_H
 #define EIGEN_GENERAL_BLOCK_PANEL_H
 
-namespace Eigen { 
-  
+
+namespace Eigen {
+
 namespace internal {
 
 template<typename _LhsScalar, typename _RhsScalar, bool _ConjLhs=false, bool _ConjRhs=false>
@@ -24,29 +25,51 @@ inline std::ptrdiff_t manage_caching_sizes_helper(std::ptrdiff_t a, std::ptrdiff
   return a<=0 ? b : a;
 }
 
+#if EIGEN_ARCH_i386_OR_x86_64
+const std::ptrdiff_t defaultL1CacheSize = 32*1024;
+const std::ptrdiff_t defaultL2CacheSize = 256*1024;
+const std::ptrdiff_t defaultL3CacheSize = 2*1024*1024;
+#else
+const std::ptrdiff_t defaultL1CacheSize = 16*1024;
+const std::ptrdiff_t defaultL2CacheSize = 512*1024;
+const std::ptrdiff_t defaultL3CacheSize = 512*1024;
+#endif
+
 /** \internal */
-inline void manage_caching_sizes(Action action, std::ptrdiff_t* l1=0, std::ptrdiff_t* l2=0)
-{
-  static std::ptrdiff_t m_l1CacheSize = 0;
-  static std::ptrdiff_t m_l2CacheSize = 0;
-  if(m_l2CacheSize==0)
-  {
-    m_l1CacheSize = manage_caching_sizes_helper(queryL1CacheSize(),8 * 1024);
-    m_l2CacheSize = manage_caching_sizes_helper(queryTopLevelCacheSize(),1*1024*1024);
+struct CacheSizes {
+  CacheSizes(): m_l1(-1),m_l2(-1),m_l3(-1) {
+    int l1CacheSize, l2CacheSize, l3CacheSize;
+    queryCacheSizes(l1CacheSize, l2CacheSize, l3CacheSize);
+    m_l1 = manage_caching_sizes_helper(l1CacheSize, defaultL1CacheSize);
+    m_l2 = manage_caching_sizes_helper(l2CacheSize, defaultL2CacheSize);
+    m_l3 = manage_caching_sizes_helper(l3CacheSize, defaultL3CacheSize);
   }
-  
+
+  std::ptrdiff_t m_l1;
+  std::ptrdiff_t m_l2;
+  std::ptrdiff_t m_l3;
+};
+
+
+/** \internal */
+inline void manage_caching_sizes(Action action, std::ptrdiff_t* l1, std::ptrdiff_t* l2, std::ptrdiff_t* l3)
+{
+  static CacheSizes m_cacheSizes;
+
   if(action==SetAction)
   {
     // set the cpu cache size and cache all block sizes from a global cache size in byte
     eigen_internal_assert(l1!=0 && l2!=0);
-    m_l1CacheSize = *l1;
-    m_l2CacheSize = *l2;
+    m_cacheSizes.m_l1 = *l1;
+    m_cacheSizes.m_l2 = *l2;
+    m_cacheSizes.m_l3 = *l3;
   }
   else if(action==GetAction)
   {
     eigen_internal_assert(l1!=0 && l2!=0);
-    *l1 = m_l1CacheSize;
-    *l2 = m_l2CacheSize;
+    *l1 = m_cacheSizes.m_l1;
+    *l2 = m_cacheSizes.m_l2;
+    *l3 = m_cacheSizes.m_l3;
   }
   else
   {
@@ -54,6 +77,206 @@ inline void manage_caching_sizes(Action action, std::ptrdiff_t* l1=0, std::ptrdi
   }
 }
 
+/* Helper for computeProductBlockingSizes.
+ *
+ * Given a m x k times k x n matrix product of scalar types \c LhsScalar and \c RhsScalar,
+ * this function computes the blocking size parameters along the respective dimensions
+ * for matrix products and related algorithms. The blocking sizes depends on various
+ * parameters:
+ * - the L1 and L2 cache sizes,
+ * - the register level blocking sizes defined by gebp_traits,
+ * - the number of scalars that fit into a packet (when vectorization is enabled).
+ *
+ * \sa setCpuCacheSizes */
+
+template<typename LhsScalar, typename RhsScalar, int KcFactor, typename Index>
+void evaluateProductBlockingSizesHeuristic(Index& k, Index& m, Index& n, Index num_threads = 1)
+{
+  typedef gebp_traits<LhsScalar,RhsScalar> Traits;
+
+  // Explanations:
+  // Let's recall that the product algorithms form mc x kc vertical panels A' on the lhs and
+  // kc x nc blocks B' on the rhs. B' has to fit into L2/L3 cache. Moreover, A' is processed
+  // per mr x kc horizontal small panels where mr is the blocking size along the m dimension
+  // at the register level. This small horizontal panel has to stay within L1 cache.
+  std::ptrdiff_t l1, l2, l3;
+  manage_caching_sizes(GetAction, &l1, &l2, &l3);
+
+  if (num_threads > 1) {
+    typedef typename Traits::ResScalar ResScalar;
+    enum {
+      kdiv = KcFactor * (Traits::mr * sizeof(LhsScalar) + Traits::nr * sizeof(RhsScalar)),
+      ksub = Traits::mr * Traits::nr * sizeof(ResScalar),
+      kr = 8,
+      mr = Traits::mr,
+      nr = Traits::nr
+    };
+    // Increasing k gives us more time to prefetch the content of the "C"
+    // registers. However once the latency is hidden there is no point in
+    // increasing the value of k, so we'll cap it at 320 (value determined
+    // experimentally).
+    const Index k_cache = (numext::mini<Index>)((l1-ksub)/kdiv, 320);
+    if (k_cache < k) {
+      k = k_cache - (k_cache % kr);
+      eigen_internal_assert(k > 0);
+    }
+
+    const Index n_cache = (l2-l1) / (nr * sizeof(RhsScalar) * k);
+    const Index n_per_thread = numext::div_ceil(n, num_threads);
+    if (n_cache <= n_per_thread) {
+      // Don't exceed the capacity of the l2 cache.
+      eigen_internal_assert(n_cache >= static_cast<Index>(nr));
+      n = n_cache - (n_cache % nr);
+      eigen_internal_assert(n > 0);
+    } else {
+      n = (numext::mini<Index>)(n, (n_per_thread + nr - 1) - ((n_per_thread + nr - 1) % nr));
+    }
+
+    if (l3 > l2) {
+      // l3 is shared between all cores, so we'll give each thread its own chunk of l3.
+      const Index m_cache = (l3-l2) / (sizeof(LhsScalar) * k * num_threads);
+      const Index m_per_thread = numext::div_ceil(m, num_threads);
+      if(m_cache < m_per_thread && m_cache >= static_cast<Index>(mr)) {
+        m = m_cache - (m_cache % mr);
+        eigen_internal_assert(m > 0);
+      } else {
+        m = (numext::mini<Index>)(m, (m_per_thread + mr - 1) - ((m_per_thread + mr - 1) % mr));
+      }
+    }
+  }
+  else {
+    // In unit tests we do not want to use extra large matrices,
+    // so we reduce the cache size to check the blocking strategy is not flawed
+#ifdef EIGEN_DEBUG_SMALL_PRODUCT_BLOCKS
+    l1 = 9*1024;
+    l2 = 32*1024;
+    l3 = 512*1024;
+#endif
+
+    // Early return for small problems because the computation below are time consuming for small problems.
+    // Perhaps it would make more sense to consider k*n*m??
+    // Note that for very tiny problem, this function should be bypassed anyway
+    // because we use the coefficient-based implementation for them.
+    if((numext::maxi)(k,(numext::maxi)(m,n))<48)
+      return;
+
+    typedef typename Traits::ResScalar ResScalar;
+    enum {
+      k_peeling = 8,
+      k_div = KcFactor * (Traits::mr * sizeof(LhsScalar) + Traits::nr * sizeof(RhsScalar)),
+      k_sub = Traits::mr * Traits::nr * sizeof(ResScalar)
+    };
+
+    // ---- 1st level of blocking on L1, yields kc ----
+
+    // Blocking on the third dimension (i.e., k) is chosen so that an horizontal panel
+    // of size mr x kc of the lhs plus a vertical panel of kc x nr of the rhs both fits within L1 cache.
+    // We also include a register-level block of the result (mx x nr).
+    // (In an ideal world only the lhs panel would stay in L1)
+    // Moreover, kc has to be a multiple of 8 to be compatible with loop peeling, leading to a maximum blocking size of:
+    const Index max_kc = numext::maxi<Index>(((l1-k_sub)/k_div) & (~(k_peeling-1)),1);
+    const Index old_k = k;
+    if(k>max_kc)
+    {
+      // We are really blocking on the third dimension:
+      // -> reduce blocking size to make sure the last block is as large as possible
+      //    while keeping the same number of sweeps over the result.
+      k = (k%max_kc)==0 ? max_kc
+                        : max_kc - k_peeling * ((max_kc-1-(k%max_kc))/(k_peeling*(k/max_kc+1)));
+
+      eigen_internal_assert(((old_k/k) == (old_k/max_kc)) && "the number of sweeps has to remain the same");
+    }
+
+    // ---- 2nd level of blocking on max(L2,L3), yields nc ----
+
+    // TODO find a reliable way to get the actual amount of cache per core to use for 2nd level blocking, that is:
+    //      actual_l2 = max(l2, l3/nb_core_sharing_l3)
+    // The number below is quite conservative: it is better to underestimate the cache size rather than overestimating it)
+    // For instance, it corresponds to 6MB of L3 shared among 4 cores.
+    #ifdef EIGEN_DEBUG_SMALL_PRODUCT_BLOCKS
+    const Index actual_l2 = l3;
+    #else
+    const Index actual_l2 = 1572864; // == 1.5 MB
+    #endif
+
+    // Here, nc is chosen such that a block of kc x nc of the rhs fit within half of L2.
+    // The second half is implicitly reserved to access the result and lhs coefficients.
+    // When k<max_kc, then nc can arbitrarily growth. In practice, it seems to be fruitful
+    // to limit this growth: we bound nc to growth by a factor x1.5.
+    // However, if the entire lhs block fit within L1, then we are not going to block on the rows at all,
+    // and it becomes fruitful to keep the packed rhs blocks in L1 if there is enough remaining space.
+    Index max_nc;
+    const Index lhs_bytes = m * k * sizeof(LhsScalar);
+    const Index remaining_l1 = l1- k_sub - lhs_bytes;
+    if(remaining_l1 >= Index(Traits::nr*sizeof(RhsScalar))*k)
+    {
+      // L1 blocking
+      max_nc = remaining_l1 / (k*sizeof(RhsScalar));
+    }
+    else
+    {
+      // L2 blocking
+      max_nc = (3*actual_l2)/(2*2*max_kc*sizeof(RhsScalar));
+    }
+    // WARNING Below, we assume that Traits::nr is a power of two.
+    Index nc = numext::mini<Index>(actual_l2/(2*k*sizeof(RhsScalar)), max_nc) & (~(Traits::nr-1));
+    if(n>nc)
+    {
+      // We are really blocking over the columns:
+      // -> reduce blocking size to make sure the last block is as large as possible
+      //    while keeping the same number of sweeps over the packed lhs.
+      //    Here we allow one more sweep if this gives us a perfect match, thus the commented "-1"
+      n = (n%nc)==0 ? nc
+                    : (nc - Traits::nr * ((nc/*-1*/-(n%nc))/(Traits::nr*(n/nc+1))));
+    }
+    else if(old_k==k)
+    {
+      // So far, no blocking at all, i.e., kc==k, and nc==n.
+      // In this case, let's perform a blocking over the rows such that the packed lhs data is kept in cache L1/L2
+      // TODO: part of this blocking strategy is now implemented within the kernel itself, so the L1-based heuristic here should be obsolete.
+      Index problem_size = k*n*sizeof(LhsScalar);
+      Index actual_lm = actual_l2;
+      Index max_mc = m;
+      if(problem_size<=1024)
+      {
+        // problem is small enough to keep in L1
+        // Let's choose m such that lhs's block fit in 1/3 of L1
+        actual_lm = l1;
+      }
+      else if(l3!=0 && problem_size<=32768)
+      {
+        // we have both L2 and L3, and problem is small enough to be kept in L2
+        // Let's choose m such that lhs's block fit in 1/3 of L2
+        actual_lm = l2;
+        max_mc = (numext::mini<Index>)(576,max_mc);
+      }
+      Index mc = (numext::mini<Index>)(actual_lm/(3*k*sizeof(LhsScalar)), max_mc);
+      if (mc > Traits::mr) mc -= mc % Traits::mr;
+      else if (mc==0) return;
+      m = (m%mc)==0 ? mc
+                    : (mc - Traits::mr * ((mc/*-1*/-(m%mc))/(Traits::mr*(m/mc+1))));
+    }
+  }
+}
+
+template <typename Index>
+inline bool useSpecificBlockingSizes(Index& k, Index& m, Index& n)
+{
+#ifdef EIGEN_TEST_SPECIFIC_BLOCKING_SIZES
+  if (EIGEN_TEST_SPECIFIC_BLOCKING_SIZES) {
+    k = numext::mini<Index>(k, EIGEN_TEST_SPECIFIC_BLOCKING_SIZE_K);
+    m = numext::mini<Index>(m, EIGEN_TEST_SPECIFIC_BLOCKING_SIZE_M);
+    n = numext::mini<Index>(n, EIGEN_TEST_SPECIFIC_BLOCKING_SIZE_N);
+    return true;
+  }
+#else
+  EIGEN_UNUSED_VARIABLE(k)
+  EIGEN_UNUSED_VARIABLE(m)
+  EIGEN_UNUSED_VARIABLE(n)
+#endif
+  return false;
+}
+
 /** \brief Computes the blocking parameters for a m x k times k x n matrix product
   *
   * \param[in,out] k Input: the third dimension of the product. Output: the blocking size along the same dimension.
@@ -62,48 +285,30 @@ inline void manage_caching_sizes(Action action, std::ptrdiff_t* l1=0, std::ptrdi
   *
   * Given a m x k times k x n matrix product of scalar types \c LhsScalar and \c RhsScalar,
   * this function computes the blocking size parameters along the respective dimensions
-  * for matrix products and related algorithms. The blocking sizes depends on various
-  * parameters:
-  * - the L1 and L2 cache sizes,
-  * - the register level blocking sizes defined by gebp_traits,
-  * - the number of scalars that fit into a packet (when vectorization is enabled).
+  * for matrix products and related algorithms.
+  *
+  * The blocking size parameters may be evaluated:
+  *   - either by a heuristic based on cache sizes;
+  *   - or using fixed prescribed values (for testing purposes).
   *
   * \sa setCpuCacheSizes */
-template<typename LhsScalar, typename RhsScalar, int KcFactor, typename SizeType>
-void computeProductBlockingSizes(SizeType& k, SizeType& m, SizeType& n)
-{
-  EIGEN_UNUSED_VARIABLE(n);
-  // Explanations:
-  // Let's recall the product algorithms form kc x nc horizontal panels B' on the rhs and
-  // mc x kc blocks A' on the lhs. A' has to fit into L2 cache. Moreover, B' is processed
-  // per kc x nr vertical small panels where nr is the blocking size along the n dimension
-  // at the register level. For vectorization purpose, these small vertical panels are unpacked,
-  // e.g., each coefficient is replicated to fit a packet. This small vertical panel has to
-  // stay in L1 cache.
-  std::ptrdiff_t l1, l2;
 
-  typedef gebp_traits<LhsScalar,RhsScalar> Traits;
-  enum {
-    kdiv = KcFactor * 2 * Traits::nr
-         * Traits::RhsProgress * sizeof(RhsScalar),
-    mr = gebp_traits<LhsScalar,RhsScalar>::mr,
-    mr_mask = (0xffffffff/mr)*mr
-  };
-
-  manage_caching_sizes(GetAction, &l1, &l2);
-  k = std::min<SizeType>(k, l1/kdiv);
-  SizeType _m = k>0 ? l2/(4 * sizeof(LhsScalar) * k) : 0;
-  if(_m<m) m = _m & mr_mask;
+template<typename LhsScalar, typename RhsScalar, int KcFactor, typename Index>
+void computeProductBlockingSizes(Index& k, Index& m, Index& n, Index num_threads = 1)
+{
+  if (!useSpecificBlockingSizes(k, m, n)) {
+    evaluateProductBlockingSizesHeuristic<LhsScalar, RhsScalar, KcFactor, Index>(k, m, n, num_threads);
+  }
 }
 
-template<typename LhsScalar, typename RhsScalar, typename SizeType>
-inline void computeProductBlockingSizes(SizeType& k, SizeType& m, SizeType& n)
+template<typename LhsScalar, typename RhsScalar, typename Index>
+inline void computeProductBlockingSizes(Index& k, Index& m, Index& n, Index num_threads = 1)
 {
-  computeProductBlockingSizes<LhsScalar,RhsScalar,1>(k, m, n);
+  computeProductBlockingSizes<LhsScalar,RhsScalar,1,Index>(k, m, n, num_threads);
 }
 
-#ifdef EIGEN_HAS_FUSE_CJMADD
-  #define MADD(CJ,A,B,C,T)  C = CJ.pmadd(A,B,C);
+#ifdef EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD
+  #define CJMADD(CJ,A,B,C,T)  C = CJ.pmadd(A,B,C);
 #else
 
   // FIXME (a bit overkill maybe ?)
@@ -128,8 +333,8 @@ inline void computeProductBlockingSizes(SizeType& k, SizeType& m, SizeType& n)
     gebp_madd_selector<CJ,A,B,C,T>::run(cj,a,b,c,t);
   }
 
-  #define MADD(CJ,A,B,C,T)  gebp_madd(CJ,A,B,C,T);
-//   #define MADD(CJ,A,B,C,T)  T = B; T = CJ.pmul(A,T); C = padd(C,T);
+  #define CJMADD(CJ,A,B,C,T)  gebp_madd(CJ,A,B,C,T);
+//   #define CJMADD(CJ,A,B,C,T)  T = B; T = CJ.pmul(A,T); C = padd(C,T);
 #endif
 
 /* Vectorization logic
@@ -148,7 +353,7 @@ class gebp_traits
 public:
   typedef _LhsScalar LhsScalar;
   typedef _RhsScalar RhsScalar;
-  typedef typename scalar_product_traits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+  typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
 
   enum {
     ConjLhs = _ConjLhs,
@@ -160,16 +365,22 @@ public:
     
     NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS,
 
-    // register block size along the N direction (must be either 2 or 4)
-    nr = NumberOfRegisters/4,
+    // register block size along the N direction must be 1 or 4
+    nr = 4,
 
     // register block size along the M direction (currently, this one cannot be modified)
-    mr = 2 * LhsPacketSize,
+    default_mr = (EIGEN_PLAIN_ENUM_MIN(16,NumberOfRegisters)/2/nr)*LhsPacketSize,
+#if defined(EIGEN_HAS_SINGLE_INSTRUCTION_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) && !defined(EIGEN_VECTORIZE_VSX)
+    // we assume 16 registers
+    // See bug 992, if the scalar type is not vectorizable but that EIGEN_HAS_SINGLE_INSTRUCTION_MADD is defined,
+    // then using 3*LhsPacketSize triggers non-implemented paths in syrk.
+    mr = Vectorizable ? 3*LhsPacketSize : default_mr,
+#else
+    mr = default_mr,
+#endif
     
-    WorkSpaceFactor = nr * RhsPacketSize,
-
     LhsProgress = LhsPacketSize,
-    RhsProgress = RhsPacketSize
+    RhsProgress = 1
   };
 
   typedef typename packet_traits<LhsScalar>::type  _LhsPacket;
@@ -186,36 +397,67 @@ public:
   {
     p = pset1<ResPacket>(ResScalar(0));
   }
-
-  EIGEN_STRONG_INLINE void unpackRhs(DenseIndex n, const RhsScalar* rhs, RhsScalar* b)
+  
+  EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3)
+  {
+    pbroadcast4(b, b0, b1, b2, b3);
+  }
+  
+//   EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1)
+//   {
+//     pbroadcast2(b, b0, b1);
+//   }
+  
+  template<typename RhsPacketType>
+  EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacketType& dest) const
+  {
+    dest = pset1<RhsPacketType>(*b);
+  }
+  
+  EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, RhsPacket& dest) const
   {
-    for(DenseIndex k=0; k<n; k++)
-      pstore1<RhsPacket>(&b[k*RhsPacketSize], rhs[k]);
+    dest = ploadquad<RhsPacket>(b);
   }
 
-  EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacket& dest) const
+  template<typename LhsPacketType>
+  EIGEN_STRONG_INLINE void loadLhs(const LhsScalar* a, LhsPacketType& dest) const
   {
-    dest = pload<RhsPacket>(b);
+    dest = pload<LhsPacketType>(a);
   }
 
-  EIGEN_STRONG_INLINE void loadLhs(const LhsScalar* a, LhsPacket& dest) const
+  template<typename LhsPacketType>
+  EIGEN_STRONG_INLINE void loadLhsUnaligned(const LhsScalar* a, LhsPacketType& dest) const
   {
-    dest = pload<LhsPacket>(a);
+    dest = ploadu<LhsPacketType>(a);
   }
 
-  EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, AccPacket& c, AccPacket& tmp) const
+  template<typename LhsPacketType, typename RhsPacketType, typename AccPacketType>
+  EIGEN_STRONG_INLINE void madd(const LhsPacketType& a, const RhsPacketType& b, AccPacketType& c, AccPacketType& tmp) const
   {
-    tmp = b; tmp = pmul(a,tmp); c = padd(c,tmp);
+    conj_helper<LhsPacketType,RhsPacketType,ConjLhs,ConjRhs> cj;
+    // It would be a lot cleaner to call pmadd all the time. Unfortunately if we
+    // let gcc allocate the register in which to store the result of the pmul
+    // (in the case where there is no FMA) gcc fails to figure out how to avoid
+    // spilling register.
+#ifdef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+    EIGEN_UNUSED_VARIABLE(tmp);
+    c = cj.pmadd(a,b,c);
+#else
+    tmp = b; tmp = cj.pmul(a,tmp); c = padd(c,tmp);
+#endif
   }
 
   EIGEN_STRONG_INLINE void acc(const AccPacket& c, const ResPacket& alpha, ResPacket& r) const
   {
     r = pmadd(c,alpha,r);
   }
+  
+  template<typename ResPacketHalf>
+  EIGEN_STRONG_INLINE void acc(const ResPacketHalf& c, const ResPacketHalf& alpha, ResPacketHalf& r) const
+  {
+    r = pmadd(c,alpha,r);
+  }
 
-protected:
-//   conj_helper<LhsScalar,RhsScalar,ConjLhs,ConjRhs> cj;
-//   conj_helper<LhsPacket,RhsPacket,ConjLhs,ConjRhs> pcj;
 };
 
 template<typename RealScalar, bool _ConjLhs>
@@ -224,7 +466,7 @@ class gebp_traits<std::complex<RealScalar>, RealScalar, _ConjLhs, false>
 public:
   typedef std::complex<RealScalar> LhsScalar;
   typedef RealScalar RhsScalar;
-  typedef typename scalar_product_traits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+  typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
 
   enum {
     ConjLhs = _ConjLhs,
@@ -235,12 +477,16 @@ public:
     ResPacketSize = Vectorizable ? packet_traits<ResScalar>::size : 1,
     
     NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS,
-    nr = NumberOfRegisters/4,
-    mr = 2 * LhsPacketSize,
-    WorkSpaceFactor = nr*RhsPacketSize,
+    nr = 4,
+#if defined(EIGEN_HAS_SINGLE_INSTRUCTION_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) && !defined(EIGEN_VECTORIZE_VSX)
+    // we assume 16 registers
+    mr = 3*LhsPacketSize,
+#else
+    mr = (EIGEN_PLAIN_ENUM_MIN(16,NumberOfRegisters)/2/nr)*LhsPacketSize,
+#endif
 
     LhsProgress = LhsPacketSize,
-    RhsProgress = RhsPacketSize
+    RhsProgress = 1
   };
 
   typedef typename packet_traits<LhsScalar>::type  _LhsPacket;
@@ -258,15 +504,14 @@ public:
     p = pset1<ResPacket>(ResScalar(0));
   }
 
-  EIGEN_STRONG_INLINE void unpackRhs(DenseIndex n, const RhsScalar* rhs, RhsScalar* b)
+  EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacket& dest) const
   {
-    for(DenseIndex k=0; k<n; k++)
-      pstore1<RhsPacket>(&b[k*RhsPacketSize], rhs[k]);
+    dest = pset1<RhsPacket>(*b);
   }
-
-  EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacket& dest) const
+  
+  EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, RhsPacket& dest) const
   {
-    dest = pload<RhsPacket>(b);
+    dest = pset1<RhsPacket>(*b);
   }
 
   EIGEN_STRONG_INLINE void loadLhs(const LhsScalar* a, LhsPacket& dest) const
@@ -274,6 +519,21 @@ public:
     dest = pload<LhsPacket>(a);
   }
 
+  EIGEN_STRONG_INLINE void loadLhsUnaligned(const LhsScalar* a, LhsPacket& dest) const
+  {
+    dest = ploadu<LhsPacket>(a);
+  }
+
+  EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3)
+  {
+    pbroadcast4(b, b0, b1, b2, b3);
+  }
+  
+//   EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1)
+//   {
+//     pbroadcast2(b, b0, b1);
+//   }
+
   EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp) const
   {
     madd_impl(a, b, c, tmp, typename conditional<Vectorizable,true_type,false_type>::type());
@@ -281,7 +541,12 @@ public:
 
   EIGEN_STRONG_INLINE void madd_impl(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp, const true_type&) const
   {
+#ifdef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+    EIGEN_UNUSED_VARIABLE(tmp);
+    c.v = pmadd(a.v,b,c.v);
+#else
     tmp = b; tmp = pmul(a.v,tmp); c.v = padd(c.v,tmp);
+#endif
   }
 
   EIGEN_STRONG_INLINE void madd_impl(const LhsScalar& a, const RhsScalar& b, ResScalar& c, RhsScalar& /*tmp*/, const false_type&) const
@@ -298,6 +563,38 @@ protected:
   conj_helper<ResPacket,ResPacket,ConjLhs,false> cj;
 };
 
+template<typename Packet>
+struct DoublePacket
+{
+  Packet first;
+  Packet second;
+};
+
+template<typename Packet>
+DoublePacket<Packet> padd(const DoublePacket<Packet> &a, const DoublePacket<Packet> &b)
+{
+  DoublePacket<Packet> res;
+  res.first  = padd(a.first, b.first);
+  res.second = padd(a.second,b.second);
+  return res;
+}
+
+template<typename Packet>
+const DoublePacket<Packet>& predux_downto4(const DoublePacket<Packet> &a)
+{
+  return a;
+}
+
+template<typename Packet> struct unpacket_traits<DoublePacket<Packet> > { typedef DoublePacket<Packet> half; };
+// template<typename Packet>
+// DoublePacket<Packet> pmadd(const DoublePacket<Packet> &a, const DoublePacket<Packet> &b)
+// {
+//   DoublePacket<Packet> res;
+//   res.first  = padd(a.first, b.first);
+//   res.second = padd(a.second,b.second);
+//   return res;
+// }
+
 template<typename RealScalar, bool _ConjLhs, bool _ConjRhs>
 class gebp_traits<std::complex<RealScalar>, std::complex<RealScalar>, _ConjLhs, _ConjRhs >
 {
@@ -314,60 +611,80 @@ public:
                 && packet_traits<Scalar>::Vectorizable,
     RealPacketSize  = Vectorizable ? packet_traits<RealScalar>::size : 1,
     ResPacketSize   = Vectorizable ? packet_traits<ResScalar>::size : 1,
-    
-    nr = 2,
-    mr = 2 * ResPacketSize,
-    WorkSpaceFactor = Vectorizable ? 2*nr*RealPacketSize : nr,
+    LhsPacketSize = Vectorizable ? packet_traits<LhsScalar>::size : 1,
+    RhsPacketSize = Vectorizable ? packet_traits<RhsScalar>::size : 1,
+
+    // FIXME: should depend on NumberOfRegisters
+    nr = 4,
+    mr = ResPacketSize,
 
     LhsProgress = ResPacketSize,
-    RhsProgress = Vectorizable ? 2*ResPacketSize : 1
+    RhsProgress = 1
   };
   
   typedef typename packet_traits<RealScalar>::type RealPacket;
   typedef typename packet_traits<Scalar>::type     ScalarPacket;
-  struct DoublePacket
-  {
-    RealPacket first;
-    RealPacket second;
-  };
+  typedef DoublePacket<RealPacket> DoublePacketType;
 
   typedef typename conditional<Vectorizable,RealPacket,  Scalar>::type LhsPacket;
-  typedef typename conditional<Vectorizable,DoublePacket,Scalar>::type RhsPacket;
+  typedef typename conditional<Vectorizable,DoublePacketType,Scalar>::type RhsPacket;
   typedef typename conditional<Vectorizable,ScalarPacket,Scalar>::type ResPacket;
-  typedef typename conditional<Vectorizable,DoublePacket,Scalar>::type AccPacket;
+  typedef typename conditional<Vectorizable,DoublePacketType,Scalar>::type AccPacket;
   
   EIGEN_STRONG_INLINE void initAcc(Scalar& p) { p = Scalar(0); }
 
-  EIGEN_STRONG_INLINE void initAcc(DoublePacket& p)
+  EIGEN_STRONG_INLINE void initAcc(DoublePacketType& p)
   {
     p.first   = pset1<RealPacket>(RealScalar(0));
     p.second  = pset1<RealPacket>(RealScalar(0));
   }
 
-  /* Unpack the rhs coeff such that each complex coefficient is spread into
-   * two packects containing respectively the real and imaginary coefficient
-   * duplicated as many time as needed: (x+iy) => [x, ..., x] [y, ..., y]
-   */
-  EIGEN_STRONG_INLINE void unpackRhs(DenseIndex n, const Scalar* rhs, Scalar* b)
+  // Scalar path
+  EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, ResPacket& dest) const
   {
-    for(DenseIndex k=0; k<n; k++)
-    {
-      if(Vectorizable)
-      {
-        pstore1<RealPacket>((RealScalar*)&b[k*ResPacketSize*2+0],             real(rhs[k]));
-        pstore1<RealPacket>((RealScalar*)&b[k*ResPacketSize*2+ResPacketSize], imag(rhs[k]));
-      }
-      else
-        b[k] = rhs[k];
-    }
+    dest = pset1<ResPacket>(*b);
   }
 
-  EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, ResPacket& dest) const { dest = *b; }
-
-  EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, DoublePacket& dest) const
+  // Vectorized path
+  EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, DoublePacketType& dest) const
+  {
+    dest.first  = pset1<RealPacket>(real(*b));
+    dest.second = pset1<RealPacket>(imag(*b));
+  }
+  
+  EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, ResPacket& dest) const
+  {
+    loadRhs(b,dest);
+  }
+  EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, DoublePacketType& dest) const
+  {
+    eigen_internal_assert(unpacket_traits<ScalarPacket>::size<=4);
+    loadRhs(b,dest);
+  }
+  
+  EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3)
   {
-    dest.first  = pload<RealPacket>((const RealScalar*)b);
-    dest.second = pload<RealPacket>((const RealScalar*)(b+ResPacketSize));
+    // FIXME not sure that's the best way to implement it!
+    loadRhs(b+0, b0);
+    loadRhs(b+1, b1);
+    loadRhs(b+2, b2);
+    loadRhs(b+3, b3);
+  }
+  
+  // Vectorized path
+  EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, DoublePacketType& b0, DoublePacketType& b1)
+  {
+    // FIXME not sure that's the best way to implement it!
+    loadRhs(b+0, b0);
+    loadRhs(b+1, b1);
+  }
+  
+  // Scalar path
+  EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsScalar& b0, RhsScalar& b1)
+  {
+    // FIXME not sure that's the best way to implement it!
+    loadRhs(b+0, b0);
+    loadRhs(b+1, b1);
   }
 
   // nothing special here
@@ -376,7 +693,12 @@ public:
     dest = pload<LhsPacket>((const typename unpacket_traits<LhsPacket>::type*)(a));
   }
 
-  EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, DoublePacket& c, RhsPacket& /*tmp*/) const
+  EIGEN_STRONG_INLINE void loadLhsUnaligned(const LhsScalar* a, LhsPacket& dest) const
+  {
+    dest = ploadu<LhsPacket>((const typename unpacket_traits<LhsPacket>::type*)(a));
+  }
+
+  EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, DoublePacketType& c, RhsPacket& /*tmp*/) const
   {
     c.first   = padd(pmul(a,b.first), c.first);
     c.second  = padd(pmul(a,b.second),c.second);
@@ -389,7 +711,7 @@ public:
   
   EIGEN_STRONG_INLINE void acc(const Scalar& c, const Scalar& alpha, Scalar& r) const { r += alpha * c; }
   
-  EIGEN_STRONG_INLINE void acc(const DoublePacket& c, const ResPacket& alpha, ResPacket& r) const
+  EIGEN_STRONG_INLINE void acc(const DoublePacketType& c, const ResPacket& alpha, ResPacket& r) const
   {
     // assemble c
     ResPacket tmp;
@@ -440,12 +762,12 @@ public:
     ResPacketSize = Vectorizable ? packet_traits<ResScalar>::size : 1,
     
     NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS,
+    // FIXME: should depend on NumberOfRegisters
     nr = 4,
-    mr = 2*ResPacketSize,
-    WorkSpaceFactor = nr*RhsPacketSize,
+    mr = (EIGEN_PLAIN_ENUM_MIN(16,NumberOfRegisters)/2/nr)*ResPacketSize,
 
     LhsProgress = ResPacketSize,
-    RhsProgress = ResPacketSize
+    RhsProgress = 1
   };
 
   typedef typename packet_traits<LhsScalar>::type  _LhsPacket;
@@ -463,21 +785,38 @@ public:
     p = pset1<ResPacket>(ResScalar(0));
   }
 
-  EIGEN_STRONG_INLINE void unpackRhs(DenseIndex n, const RhsScalar* rhs, RhsScalar* b)
+  EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacket& dest) const
   {
-    for(DenseIndex k=0; k<n; k++)
-      pstore1<RhsPacket>(&b[k*RhsPacketSize], rhs[k]);
+    dest = pset1<RhsPacket>(*b);
   }
-
-  EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacket& dest) const
+  
+  void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3)
   {
-    dest = pload<RhsPacket>(b);
+    pbroadcast4(b, b0, b1, b2, b3);
   }
+  
+//   EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1)
+//   {
+//     // FIXME not sure that's the best way to implement it!
+//     b0 = pload1<RhsPacket>(b+0);
+//     b1 = pload1<RhsPacket>(b+1);
+//   }
 
   EIGEN_STRONG_INLINE void loadLhs(const LhsScalar* a, LhsPacket& dest) const
   {
     dest = ploaddup<LhsPacket>(a);
   }
+  
+  EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, RhsPacket& dest) const
+  {
+    eigen_internal_assert(unpacket_traits<RhsPacket>::size<=4);
+    loadRhs(b,dest);
+  }
+
+  EIGEN_STRONG_INLINE void loadLhsUnaligned(const LhsScalar* a, LhsPacket& dest) const
+  {
+    dest = ploaddup<LhsPacket>(a);
+  }
 
   EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp) const
   {
@@ -486,7 +825,13 @@ public:
 
   EIGEN_STRONG_INLINE void madd_impl(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp, const true_type&) const
   {
+#ifdef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+    EIGEN_UNUSED_VARIABLE(tmp);
+    c.v = pmadd(a,b.v,c.v);
+#else
     tmp = b; tmp.v = pmul(a,tmp.v); c = padd(c,tmp);
+#endif
+    
   }
 
   EIGEN_STRONG_INLINE void madd_impl(const LhsScalar& a, const RhsScalar& b, ResScalar& c, RhsScalar& /*tmp*/, const false_type&) const
@@ -510,7 +855,7 @@ protected:
  *  |real |cplx | no vectorization yet, would require to pack A with duplication
  *  |cplx |real | easy vectorization
  */
-template<typename LhsScalar, typename RhsScalar, typename Index, int mr, int nr, bool ConjugateLhs, bool ConjugateRhs>
+template<typename LhsScalar, typename RhsScalar, typename Index, typename DataMapper, int mr, int nr, bool ConjugateLhs, bool ConjugateRhs>
 struct gebp_kernel
 {
   typedef gebp_traits<LhsScalar,RhsScalar,ConjugateLhs,ConjugateRhs> Traits;
@@ -520,6 +865,15 @@ struct gebp_kernel
   typedef typename Traits::ResPacket ResPacket;
   typedef typename Traits::AccPacket AccPacket;
 
+  typedef gebp_traits<RhsScalar,LhsScalar,ConjugateRhs,ConjugateLhs> SwappedTraits;
+  typedef typename SwappedTraits::ResScalar SResScalar;
+  typedef typename SwappedTraits::LhsPacket SLhsPacket;
+  typedef typename SwappedTraits::RhsPacket SRhsPacket;
+  typedef typename SwappedTraits::ResPacket SResPacket;
+  typedef typename SwappedTraits::AccPacket SAccPacket;
+
+  typedef typename DataMapper::LinearMapper LinearMapper;
+
   enum {
     Vectorizable  = Traits::Vectorizable,
     LhsProgress   = Traits::LhsProgress,
@@ -528,571 +882,788 @@ struct gebp_kernel
   };
 
   EIGEN_DONT_INLINE
-  void operator()(ResScalar* res, Index resStride, const LhsScalar* blockA, const RhsScalar* blockB, Index rows, Index depth, Index cols, ResScalar alpha,
-                  Index strideA=-1, Index strideB=-1, Index offsetA=0, Index offsetB=0, RhsScalar* unpackedB=0);
+  void operator()(const DataMapper& res, const LhsScalar* blockA, const RhsScalar* blockB,
+                  Index rows, Index depth, Index cols, ResScalar alpha,
+                  Index strideA=-1, Index strideB=-1, Index offsetA=0, Index offsetB=0);
 };
 
-template<typename LhsScalar, typename RhsScalar, typename Index, int mr, int nr, bool ConjugateLhs, bool ConjugateRhs>
+template<typename LhsScalar, typename RhsScalar, typename Index, typename DataMapper, int mr, int nr, bool ConjugateLhs, bool ConjugateRhs>
 EIGEN_DONT_INLINE
-void gebp_kernel<LhsScalar,RhsScalar,Index,mr,nr,ConjugateLhs,ConjugateRhs>
-  ::operator()(ResScalar* res, Index resStride, const LhsScalar* blockA, const RhsScalar* blockB, Index rows, Index depth, Index cols, ResScalar alpha,
-               Index strideA, Index strideB, Index offsetA, Index offsetB, RhsScalar* unpackedB)
+void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,ConjugateRhs>
+  ::operator()(const DataMapper& res, const LhsScalar* blockA, const RhsScalar* blockB,
+               Index rows, Index depth, Index cols, ResScalar alpha,
+               Index strideA, Index strideB, Index offsetA, Index offsetB)
   {
     Traits traits;
+    SwappedTraits straits;
     
     if(strideA==-1) strideA = depth;
     if(strideB==-1) strideB = depth;
     conj_helper<LhsScalar,RhsScalar,ConjugateLhs,ConjugateRhs> cj;
-//     conj_helper<LhsPacket,RhsPacket,ConjugateLhs,ConjugateRhs> pcj;
-    Index packet_cols = (cols/nr) * nr;
-    const Index peeled_mc = (rows/mr)*mr;
-    // FIXME:
-    const Index peeled_mc2 = peeled_mc + (rows-peeled_mc >= LhsProgress ? LhsProgress : 0);
-    const Index peeled_kc = (depth/4)*4;
-
-    if(unpackedB==0)
-      unpackedB = const_cast<RhsScalar*>(blockB - strideB * nr * RhsProgress);
-
-    // loops on each micro vertical panel of rhs (depth x nr)
-    for(Index j2=0; j2<packet_cols; j2+=nr)
+    Index packet_cols4 = nr>=4 ? (cols/4) * 4 : 0;
+    const Index peeled_mc3 = mr>=3*Traits::LhsProgress ? (rows/(3*LhsProgress))*(3*LhsProgress) : 0;
+    const Index peeled_mc2 = mr>=2*Traits::LhsProgress ? peeled_mc3+((rows-peeled_mc3)/(2*LhsProgress))*(2*LhsProgress) : 0;
+    const Index peeled_mc1 = mr>=1*Traits::LhsProgress ? (rows/(1*LhsProgress))*(1*LhsProgress) : 0;
+    enum { pk = 8 }; // NOTE Such a large peeling factor is important for large matrices (~ +5% when >1000 on Haswell)
+    const Index peeled_kc  = depth & ~(pk-1);
+    const Index prefetch_res_offset = 32/sizeof(ResScalar);    
+//     const Index depth2     = depth & ~1;
+
+    //---------- Process 3 * LhsProgress rows at once ----------
+    // This corresponds to 3*LhsProgress x nr register blocks.
+    // Usually, make sense only with FMA
+    if(mr>=3*Traits::LhsProgress)
     {
-      traits.unpackRhs(depth*nr,&blockB[j2*strideB+offsetB*nr],unpackedB); 
-
-      // loops on each largest micro horizontal panel of lhs (mr x depth)
-      // => we select a mr x nr micro block of res which is entirely
-      //    stored into mr/packet_size x nr registers.
-      for(Index i=0; i<peeled_mc; i+=mr)
+      // Here, the general idea is to loop on each largest micro horizontal panel of the lhs (3*Traits::LhsProgress x depth)
+      // and on each largest micro vertical panel of the rhs (depth * nr).
+      // Blocking sizes, i.e., 'depth' has been computed so that the micro horizontal panel of the lhs fit in L1.
+      // However, if depth is too small, we can extend the number of rows of these horizontal panels.
+      // This actual number of rows is computed as follow:
+      const Index l1 = defaultL1CacheSize; // in Bytes, TODO, l1 should be passed to this function.
+      // The max(1, ...) here is needed because we may be using blocking params larger than what our known l1 cache size
+      // suggests we should be using: either because our known l1 cache size is inaccurate (e.g. on Android, we can only guess),
+      // or because we are testing specific blocking sizes.
+      const Index actual_panel_rows = (3*LhsProgress) * std::max<Index>(1,( (l1 - sizeof(ResScalar)*mr*nr - depth*nr*sizeof(RhsScalar)) / (depth * sizeof(LhsScalar) * 3*LhsProgress) ));
+      for(Index i1=0; i1<peeled_mc3; i1+=actual_panel_rows)
       {
-        const LhsScalar* blA = &blockA[i*strideA+offsetA*mr];
-        prefetch(&blA[0]);
-
-        // gets res block as register
-        AccPacket C0, C1, C2, C3, C4, C5, C6, C7;
-                  traits.initAcc(C0);
-                  traits.initAcc(C1);
-        if(nr==4) traits.initAcc(C2);
-        if(nr==4) traits.initAcc(C3);
-                  traits.initAcc(C4);
-                  traits.initAcc(C5);
-        if(nr==4) traits.initAcc(C6);
-        if(nr==4) traits.initAcc(C7);
-
-        ResScalar* r0 = &res[(j2+0)*resStride + i];
-        ResScalar* r1 = r0 + resStride;
-        ResScalar* r2 = r1 + resStride;
-        ResScalar* r3 = r2 + resStride;
-
-        prefetch(r0+16);
-        prefetch(r1+16);
-        prefetch(r2+16);
-        prefetch(r3+16);
-
-        // performs "inner" product
-        // TODO let's check wether the folowing peeled loop could not be
-        //      optimized via optimal prefetching from one loop to the other
-        const RhsScalar* blB = unpackedB;
-        for(Index k=0; k<peeled_kc; k+=4)
+        const Index actual_panel_end = (std::min)(i1+actual_panel_rows, peeled_mc3);
+        for(Index j2=0; j2<packet_cols4; j2+=nr)
         {
-          if(nr==2)
-          {
-            LhsPacket A0, A1;
-            RhsPacket B_0;
-            RhsPacket T0;
-            
-EIGEN_ASM_COMMENT("mybegin2");
-            traits.loadLhs(&blA[0*LhsProgress], A0);
-            traits.loadLhs(&blA[1*LhsProgress], A1);
-            traits.loadRhs(&blB[0*RhsProgress], B_0);
-            traits.madd(A0,B_0,C0,T0);
-            traits.madd(A1,B_0,C4,B_0);
-            traits.loadRhs(&blB[1*RhsProgress], B_0);
-            traits.madd(A0,B_0,C1,T0);
-            traits.madd(A1,B_0,C5,B_0);
-
-            traits.loadLhs(&blA[2*LhsProgress], A0);
-            traits.loadLhs(&blA[3*LhsProgress], A1);
-            traits.loadRhs(&blB[2*RhsProgress], B_0);
-            traits.madd(A0,B_0,C0,T0);
-            traits.madd(A1,B_0,C4,B_0);
-            traits.loadRhs(&blB[3*RhsProgress], B_0);
-            traits.madd(A0,B_0,C1,T0);
-            traits.madd(A1,B_0,C5,B_0);
-
-            traits.loadLhs(&blA[4*LhsProgress], A0);
-            traits.loadLhs(&blA[5*LhsProgress], A1);
-            traits.loadRhs(&blB[4*RhsProgress], B_0);
-            traits.madd(A0,B_0,C0,T0);
-            traits.madd(A1,B_0,C4,B_0);
-            traits.loadRhs(&blB[5*RhsProgress], B_0);
-            traits.madd(A0,B_0,C1,T0);
-            traits.madd(A1,B_0,C5,B_0);
-
-            traits.loadLhs(&blA[6*LhsProgress], A0);
-            traits.loadLhs(&blA[7*LhsProgress], A1);
-            traits.loadRhs(&blB[6*RhsProgress], B_0);
-            traits.madd(A0,B_0,C0,T0);
-            traits.madd(A1,B_0,C4,B_0);
-            traits.loadRhs(&blB[7*RhsProgress], B_0);
-            traits.madd(A0,B_0,C1,T0);
-            traits.madd(A1,B_0,C5,B_0);
-EIGEN_ASM_COMMENT("myend");
-          }
-          else
+          for(Index i=i1; i<actual_panel_end; i+=3*LhsProgress)
           {
-EIGEN_ASM_COMMENT("mybegin4");
-            LhsPacket A0, A1;
-            RhsPacket B_0, B1, B2, B3;
-            RhsPacket T0;
-            
-            traits.loadLhs(&blA[0*LhsProgress], A0);
-            traits.loadLhs(&blA[1*LhsProgress], A1);
-            traits.loadRhs(&blB[0*RhsProgress], B_0);
-            traits.loadRhs(&blB[1*RhsProgress], B1);
-
-            traits.madd(A0,B_0,C0,T0);
-            traits.loadRhs(&blB[2*RhsProgress], B2);
-            traits.madd(A1,B_0,C4,B_0);
-            traits.loadRhs(&blB[3*RhsProgress], B3);
-            traits.loadRhs(&blB[4*RhsProgress], B_0);
-            traits.madd(A0,B1,C1,T0);
-            traits.madd(A1,B1,C5,B1);
-            traits.loadRhs(&blB[5*RhsProgress], B1);
-            traits.madd(A0,B2,C2,T0);
-            traits.madd(A1,B2,C6,B2);
-            traits.loadRhs(&blB[6*RhsProgress], B2);
-            traits.madd(A0,B3,C3,T0);
-            traits.loadLhs(&blA[2*LhsProgress], A0);
-            traits.madd(A1,B3,C7,B3);
-            traits.loadLhs(&blA[3*LhsProgress], A1);
-            traits.loadRhs(&blB[7*RhsProgress], B3);
-            traits.madd(A0,B_0,C0,T0);
-            traits.madd(A1,B_0,C4,B_0);
-            traits.loadRhs(&blB[8*RhsProgress], B_0);
-            traits.madd(A0,B1,C1,T0);
-            traits.madd(A1,B1,C5,B1);
-            traits.loadRhs(&blB[9*RhsProgress], B1);
-            traits.madd(A0,B2,C2,T0);
-            traits.madd(A1,B2,C6,B2);
-            traits.loadRhs(&blB[10*RhsProgress], B2);
-            traits.madd(A0,B3,C3,T0);
-            traits.loadLhs(&blA[4*LhsProgress], A0);
-            traits.madd(A1,B3,C7,B3);
-            traits.loadLhs(&blA[5*LhsProgress], A1);
-            traits.loadRhs(&blB[11*RhsProgress], B3);
-
-            traits.madd(A0,B_0,C0,T0);
-            traits.madd(A1,B_0,C4,B_0);
-            traits.loadRhs(&blB[12*RhsProgress], B_0);
-            traits.madd(A0,B1,C1,T0);
-            traits.madd(A1,B1,C5,B1);
-            traits.loadRhs(&blB[13*RhsProgress], B1);
-            traits.madd(A0,B2,C2,T0);
-            traits.madd(A1,B2,C6,B2);
-            traits.loadRhs(&blB[14*RhsProgress], B2);
-            traits.madd(A0,B3,C3,T0);
-            traits.loadLhs(&blA[6*LhsProgress], A0);
-            traits.madd(A1,B3,C7,B3);
-            traits.loadLhs(&blA[7*LhsProgress], A1);
-            traits.loadRhs(&blB[15*RhsProgress], B3);
-            traits.madd(A0,B_0,C0,T0);
-            traits.madd(A1,B_0,C4,B_0);
-            traits.madd(A0,B1,C1,T0);
-            traits.madd(A1,B1,C5,B1);
-            traits.madd(A0,B2,C2,T0);
-            traits.madd(A1,B2,C6,B2);
-            traits.madd(A0,B3,C3,T0);
-            traits.madd(A1,B3,C7,B3);
-          }
+          
+          // We selected a 3*Traits::LhsProgress x nr micro block of res which is entirely
+          // stored into 3 x nr registers.
+          
+          const LhsScalar* blA = &blockA[i*strideA+offsetA*(3*LhsProgress)];
+          prefetch(&blA[0]);
+
+          // gets res block as register
+          AccPacket C0, C1, C2,  C3,
+                    C4, C5, C6,  C7,
+                    C8, C9, C10, C11;
+          traits.initAcc(C0);  traits.initAcc(C1);  traits.initAcc(C2);  traits.initAcc(C3);
+          traits.initAcc(C4);  traits.initAcc(C5);  traits.initAcc(C6);  traits.initAcc(C7);
+          traits.initAcc(C8);  traits.initAcc(C9);  traits.initAcc(C10); traits.initAcc(C11);
+
+          LinearMapper r0 = res.getLinearMapper(i, j2 + 0);
+          LinearMapper r1 = res.getLinearMapper(i, j2 + 1);
+          LinearMapper r2 = res.getLinearMapper(i, j2 + 2);
+          LinearMapper r3 = res.getLinearMapper(i, j2 + 3);
+
+          r0.prefetch(0);
+          r1.prefetch(0);
+          r2.prefetch(0);
+          r3.prefetch(0);
+
+          // performs "inner" products
+          const RhsScalar* blB = &blockB[j2*strideB+offsetB*nr];
+          prefetch(&blB[0]);
+          LhsPacket A0, A1;
 
-          blB += 4*nr*RhsProgress;
-          blA += 4*mr;
-        }
-        // process remaining peeled loop
-        for(Index k=peeled_kc; k<depth; k++)
-        {
-          if(nr==2)
+          for(Index k=0; k<peeled_kc; k+=pk)
           {
-            LhsPacket A0, A1;
-            RhsPacket B_0;
-            RhsPacket T0;
-
-            traits.loadLhs(&blA[0*LhsProgress], A0);
-            traits.loadLhs(&blA[1*LhsProgress], A1);
-            traits.loadRhs(&blB[0*RhsProgress], B_0);
-            traits.madd(A0,B_0,C0,T0);
-            traits.madd(A1,B_0,C4,B_0);
-            traits.loadRhs(&blB[1*RhsProgress], B_0);
-            traits.madd(A0,B_0,C1,T0);
-            traits.madd(A1,B_0,C5,B_0);
+            EIGEN_ASM_COMMENT("begin gebp micro kernel 3pX4");
+            RhsPacket B_0, T0;
+            LhsPacket A2;
+
+#define EIGEN_GEBP_ONESTEP(K) \
+            do { \
+              EIGEN_ASM_COMMENT("begin step of gebp micro kernel 3pX4"); \
+              EIGEN_ASM_COMMENT("Note: these asm comments work around bug 935!"); \
+              internal::prefetch(blA+(3*K+16)*LhsProgress); \
+              if (EIGEN_ARCH_ARM) { internal::prefetch(blB+(4*K+16)*RhsProgress); } /* Bug 953 */ \
+              traits.loadLhs(&blA[(0+3*K)*LhsProgress], A0);  \
+              traits.loadLhs(&blA[(1+3*K)*LhsProgress], A1);  \
+              traits.loadLhs(&blA[(2+3*K)*LhsProgress], A2);  \
+              traits.loadRhs(blB + (0+4*K)*Traits::RhsProgress, B_0); \
+              traits.madd(A0, B_0, C0, T0); \
+              traits.madd(A1, B_0, C4, T0); \
+              traits.madd(A2, B_0, C8, B_0); \
+              traits.loadRhs(blB + (1+4*K)*Traits::RhsProgress, B_0); \
+              traits.madd(A0, B_0, C1, T0); \
+              traits.madd(A1, B_0, C5, T0); \
+              traits.madd(A2, B_0, C9, B_0); \
+              traits.loadRhs(blB + (2+4*K)*Traits::RhsProgress, B_0); \
+              traits.madd(A0, B_0, C2,  T0); \
+              traits.madd(A1, B_0, C6,  T0); \
+              traits.madd(A2, B_0, C10, B_0); \
+              traits.loadRhs(blB + (3+4*K)*Traits::RhsProgress, B_0); \
+              traits.madd(A0, B_0, C3 , T0); \
+              traits.madd(A1, B_0, C7,  T0); \
+              traits.madd(A2, B_0, C11, B_0); \
+              EIGEN_ASM_COMMENT("end step of gebp micro kernel 3pX4"); \
+            } while(false)
+
+            internal::prefetch(blB);
+            EIGEN_GEBP_ONESTEP(0);
+            EIGEN_GEBP_ONESTEP(1);
+            EIGEN_GEBP_ONESTEP(2);
+            EIGEN_GEBP_ONESTEP(3);
+            EIGEN_GEBP_ONESTEP(4);
+            EIGEN_GEBP_ONESTEP(5);
+            EIGEN_GEBP_ONESTEP(6);
+            EIGEN_GEBP_ONESTEP(7);
+
+            blB += pk*4*RhsProgress;
+            blA += pk*3*Traits::LhsProgress;
+
+            EIGEN_ASM_COMMENT("end gebp micro kernel 3pX4");
           }
-          else
+          // process remaining peeled loop
+          for(Index k=peeled_kc; k<depth; k++)
           {
-            LhsPacket A0, A1;
-            RhsPacket B_0, B1, B2, B3;
-            RhsPacket T0;
-
-            traits.loadLhs(&blA[0*LhsProgress], A0);
-            traits.loadLhs(&blA[1*LhsProgress], A1);
-            traits.loadRhs(&blB[0*RhsProgress], B_0);
-            traits.loadRhs(&blB[1*RhsProgress], B1);
-
-            traits.madd(A0,B_0,C0,T0);
-            traits.loadRhs(&blB[2*RhsProgress], B2);
-            traits.madd(A1,B_0,C4,B_0);
-            traits.loadRhs(&blB[3*RhsProgress], B3);
-            traits.madd(A0,B1,C1,T0);
-            traits.madd(A1,B1,C5,B1);
-            traits.madd(A0,B2,C2,T0);
-            traits.madd(A1,B2,C6,B2);
-            traits.madd(A0,B3,C3,T0);
-            traits.madd(A1,B3,C7,B3);
+            RhsPacket B_0, T0;
+            LhsPacket A2;
+            EIGEN_GEBP_ONESTEP(0);
+            blB += 4*RhsProgress;
+            blA += 3*Traits::LhsProgress;
           }
 
-          blB += nr*RhsProgress;
-          blA += mr;
-        }
+#undef EIGEN_GEBP_ONESTEP
 
-        if(nr==4)
-        {
-          ResPacket R0, R1, R2, R3, R4, R5, R6;
+          ResPacket R0, R1, R2;
           ResPacket alphav = pset1<ResPacket>(alpha);
 
-          R0 = ploadu<ResPacket>(r0);
-          R1 = ploadu<ResPacket>(r1);
-          R2 = ploadu<ResPacket>(r2);
-          R3 = ploadu<ResPacket>(r3);
-          R4 = ploadu<ResPacket>(r0 + ResPacketSize);
-          R5 = ploadu<ResPacket>(r1 + ResPacketSize);
-          R6 = ploadu<ResPacket>(r2 + ResPacketSize);
+          R0 = r0.loadPacket(0 * Traits::ResPacketSize);
+          R1 = r0.loadPacket(1 * Traits::ResPacketSize);
+          R2 = r0.loadPacket(2 * Traits::ResPacketSize);
           traits.acc(C0, alphav, R0);
-          pstoreu(r0, R0);
-          R0 = ploadu<ResPacket>(r3 + ResPacketSize);
-
-          traits.acc(C1, alphav, R1);
-          traits.acc(C2, alphav, R2);
-          traits.acc(C3, alphav, R3);
-          traits.acc(C4, alphav, R4);
-          traits.acc(C5, alphav, R5);
-          traits.acc(C6, alphav, R6);
-          traits.acc(C7, alphav, R0);
-          
-          pstoreu(r1, R1);
-          pstoreu(r2, R2);
-          pstoreu(r3, R3);
-          pstoreu(r0 + ResPacketSize, R4);
-          pstoreu(r1 + ResPacketSize, R5);
-          pstoreu(r2 + ResPacketSize, R6);
-          pstoreu(r3 + ResPacketSize, R0);
+          traits.acc(C4, alphav, R1);
+          traits.acc(C8, alphav, R2);
+          r0.storePacket(0 * Traits::ResPacketSize, R0);
+          r0.storePacket(1 * Traits::ResPacketSize, R1);
+          r0.storePacket(2 * Traits::ResPacketSize, R2);
+
+          R0 = r1.loadPacket(0 * Traits::ResPacketSize);
+          R1 = r1.loadPacket(1 * Traits::ResPacketSize);
+          R2 = r1.loadPacket(2 * Traits::ResPacketSize);
+          traits.acc(C1, alphav, R0);
+          traits.acc(C5, alphav, R1);
+          traits.acc(C9, alphav, R2);
+          r1.storePacket(0 * Traits::ResPacketSize, R0);
+          r1.storePacket(1 * Traits::ResPacketSize, R1);
+          r1.storePacket(2 * Traits::ResPacketSize, R2);
+
+          R0 = r2.loadPacket(0 * Traits::ResPacketSize);
+          R1 = r2.loadPacket(1 * Traits::ResPacketSize);
+          R2 = r2.loadPacket(2 * Traits::ResPacketSize);
+          traits.acc(C2, alphav, R0);
+          traits.acc(C6, alphav, R1);
+          traits.acc(C10, alphav, R2);
+          r2.storePacket(0 * Traits::ResPacketSize, R0);
+          r2.storePacket(1 * Traits::ResPacketSize, R1);
+          r2.storePacket(2 * Traits::ResPacketSize, R2);
+
+          R0 = r3.loadPacket(0 * Traits::ResPacketSize);
+          R1 = r3.loadPacket(1 * Traits::ResPacketSize);
+          R2 = r3.loadPacket(2 * Traits::ResPacketSize);
+          traits.acc(C3, alphav, R0);
+          traits.acc(C7, alphav, R1);
+          traits.acc(C11, alphav, R2);
+          r3.storePacket(0 * Traits::ResPacketSize, R0);
+          r3.storePacket(1 * Traits::ResPacketSize, R1);
+          r3.storePacket(2 * Traits::ResPacketSize, R2);          
+          }
         }
-        else
+
+        // Deal with remaining columns of the rhs
+        for(Index j2=packet_cols4; j2<cols; j2++)
         {
-          ResPacket R0, R1, R4;
+          for(Index i=i1; i<actual_panel_end; i+=3*LhsProgress)
+          {
+          // One column at a time
+          const LhsScalar* blA = &blockA[i*strideA+offsetA*(3*Traits::LhsProgress)];
+          prefetch(&blA[0]);
+
+          // gets res block as register
+          AccPacket C0, C4, C8;
+          traits.initAcc(C0);
+          traits.initAcc(C4);
+          traits.initAcc(C8);
+
+          LinearMapper r0 = res.getLinearMapper(i, j2);
+          r0.prefetch(0);
+
+          // performs "inner" products
+          const RhsScalar* blB = &blockB[j2*strideB+offsetB];
+          LhsPacket A0, A1, A2;
+          
+          for(Index k=0; k<peeled_kc; k+=pk)
+          {
+            EIGEN_ASM_COMMENT("begin gebp micro kernel 3pX1");
+            RhsPacket B_0;
+#define EIGEN_GEBGP_ONESTEP(K) \
+            do { \
+              EIGEN_ASM_COMMENT("begin step of gebp micro kernel 3pX1"); \
+              EIGEN_ASM_COMMENT("Note: these asm comments work around bug 935!"); \
+              traits.loadLhs(&blA[(0+3*K)*LhsProgress], A0);  \
+              traits.loadLhs(&blA[(1+3*K)*LhsProgress], A1);  \
+              traits.loadLhs(&blA[(2+3*K)*LhsProgress], A2);  \
+              traits.loadRhs(&blB[(0+K)*RhsProgress], B_0);   \
+              traits.madd(A0, B_0, C0, B_0); \
+              traits.madd(A1, B_0, C4, B_0); \
+              traits.madd(A2, B_0, C8, B_0); \
+              EIGEN_ASM_COMMENT("end step of gebp micro kernel 3pX1"); \
+            } while(false)
+        
+            EIGEN_GEBGP_ONESTEP(0);
+            EIGEN_GEBGP_ONESTEP(1);
+            EIGEN_GEBGP_ONESTEP(2);
+            EIGEN_GEBGP_ONESTEP(3);
+            EIGEN_GEBGP_ONESTEP(4);
+            EIGEN_GEBGP_ONESTEP(5);
+            EIGEN_GEBGP_ONESTEP(6);
+            EIGEN_GEBGP_ONESTEP(7);
+
+            blB += pk*RhsProgress;
+            blA += pk*3*Traits::LhsProgress;
+
+            EIGEN_ASM_COMMENT("end gebp micro kernel 3pX1");
+          }
+
+          // process remaining peeled loop
+          for(Index k=peeled_kc; k<depth; k++)
+          {
+            RhsPacket B_0;
+            EIGEN_GEBGP_ONESTEP(0);
+            blB += RhsProgress;
+            blA += 3*Traits::LhsProgress;
+          }
+#undef EIGEN_GEBGP_ONESTEP
+          ResPacket R0, R1, R2;
           ResPacket alphav = pset1<ResPacket>(alpha);
 
-          R0 = ploadu<ResPacket>(r0);
-          R1 = ploadu<ResPacket>(r1);
-          R4 = ploadu<ResPacket>(r0 + ResPacketSize);
+          R0 = r0.loadPacket(0 * Traits::ResPacketSize);
+          R1 = r0.loadPacket(1 * Traits::ResPacketSize);
+          R2 = r0.loadPacket(2 * Traits::ResPacketSize);
           traits.acc(C0, alphav, R0);
-          pstoreu(r0, R0);
-          R0 = ploadu<ResPacket>(r1 + ResPacketSize);
-          traits.acc(C1, alphav, R1);
-          traits.acc(C4, alphav, R4);
-          traits.acc(C5, alphav, R0);
-          pstoreu(r1, R1);
-          pstoreu(r0 + ResPacketSize, R4);
-          pstoreu(r1 + ResPacketSize, R0);
+          traits.acc(C4, alphav, R1);
+          traits.acc(C8, alphav, R2);
+          r0.storePacket(0 * Traits::ResPacketSize, R0);
+          r0.storePacket(1 * Traits::ResPacketSize, R1);
+          r0.storePacket(2 * Traits::ResPacketSize, R2);          
+          }
         }
-        
       }
-      
-      if(rows-peeled_mc>=LhsProgress)
+    }
+
+    //---------- Process 2 * LhsProgress rows at once ----------
+    if(mr>=2*Traits::LhsProgress)
+    {
+      const Index l1 = defaultL1CacheSize; // in Bytes, TODO, l1 should be passed to this function.
+      // The max(1, ...) here is needed because we may be using blocking params larger than what our known l1 cache size
+      // suggests we should be using: either because our known l1 cache size is inaccurate (e.g. on Android, we can only guess),
+      // or because we are testing specific blocking sizes.
+      Index actual_panel_rows = (2*LhsProgress) * std::max<Index>(1,( (l1 - sizeof(ResScalar)*mr*nr - depth*nr*sizeof(RhsScalar)) / (depth * sizeof(LhsScalar) * 2*LhsProgress) ));
+
+      for(Index i1=peeled_mc3; i1<peeled_mc2; i1+=actual_panel_rows)
       {
-        Index i = peeled_mc;
-        const LhsScalar* blA = &blockA[i*strideA+offsetA*LhsProgress];
-        prefetch(&blA[0]);
-
-        // gets res block as register
-        AccPacket C0, C1, C2, C3;
-                  traits.initAcc(C0);
-                  traits.initAcc(C1);
-        if(nr==4) traits.initAcc(C2);
-        if(nr==4) traits.initAcc(C3);
-
-        // performs "inner" product
-        const RhsScalar* blB = unpackedB;
-        for(Index k=0; k<peeled_kc; k+=4)
+        Index actual_panel_end = (std::min)(i1+actual_panel_rows, peeled_mc2);
+        for(Index j2=0; j2<packet_cols4; j2+=nr)
         {
-          if(nr==2)
+          for(Index i=i1; i<actual_panel_end; i+=2*LhsProgress)
           {
-            LhsPacket A0;
-            RhsPacket B_0, B1;
+          
+          // We selected a 2*Traits::LhsProgress x nr micro block of res which is entirely
+          // stored into 2 x nr registers.
+          
+          const LhsScalar* blA = &blockA[i*strideA+offsetA*(2*Traits::LhsProgress)];
+          prefetch(&blA[0]);
+
+          // gets res block as register
+          AccPacket C0, C1, C2, C3,
+                    C4, C5, C6, C7;
+          traits.initAcc(C0); traits.initAcc(C1); traits.initAcc(C2); traits.initAcc(C3);
+          traits.initAcc(C4); traits.initAcc(C5); traits.initAcc(C6); traits.initAcc(C7);
+
+          LinearMapper r0 = res.getLinearMapper(i, j2 + 0);
+          LinearMapper r1 = res.getLinearMapper(i, j2 + 1);
+          LinearMapper r2 = res.getLinearMapper(i, j2 + 2);
+          LinearMapper r3 = res.getLinearMapper(i, j2 + 3);
+
+          r0.prefetch(prefetch_res_offset);
+          r1.prefetch(prefetch_res_offset);
+          r2.prefetch(prefetch_res_offset);
+          r3.prefetch(prefetch_res_offset);
+
+          // performs "inner" products
+          const RhsScalar* blB = &blockB[j2*strideB+offsetB*nr];
+          prefetch(&blB[0]);
+          LhsPacket A0, A1;
 
-            traits.loadLhs(&blA[0*LhsProgress], A0);
-            traits.loadRhs(&blB[0*RhsProgress], B_0);
-            traits.loadRhs(&blB[1*RhsProgress], B1);
-            traits.madd(A0,B_0,C0,B_0);
-            traits.loadRhs(&blB[2*RhsProgress], B_0);
-            traits.madd(A0,B1,C1,B1);
-            traits.loadLhs(&blA[1*LhsProgress], A0);
-            traits.loadRhs(&blB[3*RhsProgress], B1);
-            traits.madd(A0,B_0,C0,B_0);
-            traits.loadRhs(&blB[4*RhsProgress], B_0);
-            traits.madd(A0,B1,C1,B1);
-            traits.loadLhs(&blA[2*LhsProgress], A0);
-            traits.loadRhs(&blB[5*RhsProgress], B1);
-            traits.madd(A0,B_0,C0,B_0);
-            traits.loadRhs(&blB[6*RhsProgress], B_0);
-            traits.madd(A0,B1,C1,B1);
-            traits.loadLhs(&blA[3*LhsProgress], A0);
-            traits.loadRhs(&blB[7*RhsProgress], B1);
-            traits.madd(A0,B_0,C0,B_0);
-            traits.madd(A0,B1,C1,B1);
-          }
-          else
+          for(Index k=0; k<peeled_kc; k+=pk)
           {
-            LhsPacket A0;
-            RhsPacket B_0, B1, B2, B3;
-
-            traits.loadLhs(&blA[0*LhsProgress], A0);
-            traits.loadRhs(&blB[0*RhsProgress], B_0);
-            traits.loadRhs(&blB[1*RhsProgress], B1);
-
-            traits.madd(A0,B_0,C0,B_0);
-            traits.loadRhs(&blB[2*RhsProgress], B2);
-            traits.loadRhs(&blB[3*RhsProgress], B3);
-            traits.loadRhs(&blB[4*RhsProgress], B_0);
-            traits.madd(A0,B1,C1,B1);
-            traits.loadRhs(&blB[5*RhsProgress], B1);
-            traits.madd(A0,B2,C2,B2);
-            traits.loadRhs(&blB[6*RhsProgress], B2);
-            traits.madd(A0,B3,C3,B3);
-            traits.loadLhs(&blA[1*LhsProgress], A0);
-            traits.loadRhs(&blB[7*RhsProgress], B3);
-            traits.madd(A0,B_0,C0,B_0);
-            traits.loadRhs(&blB[8*RhsProgress], B_0);
-            traits.madd(A0,B1,C1,B1);
-            traits.loadRhs(&blB[9*RhsProgress], B1);
-            traits.madd(A0,B2,C2,B2);
-            traits.loadRhs(&blB[10*RhsProgress], B2);
-            traits.madd(A0,B3,C3,B3);
-            traits.loadLhs(&blA[2*LhsProgress], A0);
-            traits.loadRhs(&blB[11*RhsProgress], B3);
-
-            traits.madd(A0,B_0,C0,B_0);
-            traits.loadRhs(&blB[12*RhsProgress], B_0);
-            traits.madd(A0,B1,C1,B1);
-            traits.loadRhs(&blB[13*RhsProgress], B1);
-            traits.madd(A0,B2,C2,B2);
-            traits.loadRhs(&blB[14*RhsProgress], B2);
-            traits.madd(A0,B3,C3,B3);
-
-            traits.loadLhs(&blA[3*LhsProgress], A0);
-            traits.loadRhs(&blB[15*RhsProgress], B3);
-            traits.madd(A0,B_0,C0,B_0);
-            traits.madd(A0,B1,C1,B1);
-            traits.madd(A0,B2,C2,B2);
-            traits.madd(A0,B3,C3,B3);
+            EIGEN_ASM_COMMENT("begin gebp micro kernel 2pX4");
+            RhsPacket B_0, B1, B2, B3, T0;
+
+   #define EIGEN_GEBGP_ONESTEP(K) \
+            do {                                                                \
+              EIGEN_ASM_COMMENT("begin step of gebp micro kernel 2pX4");        \
+              EIGEN_ASM_COMMENT("Note: these asm comments work around bug 935!"); \
+              traits.loadLhs(&blA[(0+2*K)*LhsProgress], A0);                    \
+              traits.loadLhs(&blA[(1+2*K)*LhsProgress], A1);                    \
+              traits.broadcastRhs(&blB[(0+4*K)*RhsProgress], B_0, B1, B2, B3);  \
+              traits.madd(A0, B_0, C0, T0);                                     \
+              traits.madd(A1, B_0, C4, B_0);                                    \
+              traits.madd(A0, B1,  C1, T0);                                     \
+              traits.madd(A1, B1,  C5, B1);                                     \
+              traits.madd(A0, B2,  C2, T0);                                     \
+              traits.madd(A1, B2,  C6, B2);                                     \
+              traits.madd(A0, B3,  C3, T0);                                     \
+              traits.madd(A1, B3,  C7, B3);                                     \
+              EIGEN_ASM_COMMENT("end step of gebp micro kernel 2pX4");          \
+            } while(false)
+            
+            internal::prefetch(blB+(48+0));
+            EIGEN_GEBGP_ONESTEP(0);
+            EIGEN_GEBGP_ONESTEP(1);
+            EIGEN_GEBGP_ONESTEP(2);
+            EIGEN_GEBGP_ONESTEP(3);
+            internal::prefetch(blB+(48+16));
+            EIGEN_GEBGP_ONESTEP(4);
+            EIGEN_GEBGP_ONESTEP(5);
+            EIGEN_GEBGP_ONESTEP(6);
+            EIGEN_GEBGP_ONESTEP(7);
+
+            blB += pk*4*RhsProgress;
+            blA += pk*(2*Traits::LhsProgress);
+
+            EIGEN_ASM_COMMENT("end gebp micro kernel 2pX4");
           }
-
-          blB += nr*4*RhsProgress;
-          blA += 4*LhsProgress;
-        }
-        // process remaining peeled loop
-        for(Index k=peeled_kc; k<depth; k++)
-        {
-          if(nr==2)
+          // process remaining peeled loop
+          for(Index k=peeled_kc; k<depth; k++)
           {
-            LhsPacket A0;
-            RhsPacket B_0, B1;
-
-            traits.loadLhs(&blA[0*LhsProgress], A0);
-            traits.loadRhs(&blB[0*RhsProgress], B_0);
-            traits.loadRhs(&blB[1*RhsProgress], B1);
-            traits.madd(A0,B_0,C0,B_0);
-            traits.madd(A0,B1,C1,B1);
+            RhsPacket B_0, B1, B2, B3, T0;
+            EIGEN_GEBGP_ONESTEP(0);
+            blB += 4*RhsProgress;
+            blA += 2*Traits::LhsProgress;
           }
-          else
-          {
-            LhsPacket A0;
-            RhsPacket B_0, B1, B2, B3;
+#undef EIGEN_GEBGP_ONESTEP
 
-            traits.loadLhs(&blA[0*LhsProgress], A0);
-            traits.loadRhs(&blB[0*RhsProgress], B_0);
-            traits.loadRhs(&blB[1*RhsProgress], B1);
-            traits.loadRhs(&blB[2*RhsProgress], B2);
-            traits.loadRhs(&blB[3*RhsProgress], B3);
+          ResPacket R0, R1, R2, R3;
+          ResPacket alphav = pset1<ResPacket>(alpha);
 
-            traits.madd(A0,B_0,C0,B_0);
-            traits.madd(A0,B1,C1,B1);
-            traits.madd(A0,B2,C2,B2);
-            traits.madd(A0,B3,C3,B3);
+          R0 = r0.loadPacket(0 * Traits::ResPacketSize);
+          R1 = r0.loadPacket(1 * Traits::ResPacketSize);
+          R2 = r1.loadPacket(0 * Traits::ResPacketSize);
+          R3 = r1.loadPacket(1 * Traits::ResPacketSize);
+          traits.acc(C0, alphav, R0);
+          traits.acc(C4, alphav, R1);
+          traits.acc(C1, alphav, R2);
+          traits.acc(C5, alphav, R3);
+          r0.storePacket(0 * Traits::ResPacketSize, R0);
+          r0.storePacket(1 * Traits::ResPacketSize, R1);
+          r1.storePacket(0 * Traits::ResPacketSize, R2);
+          r1.storePacket(1 * Traits::ResPacketSize, R3);
+
+          R0 = r2.loadPacket(0 * Traits::ResPacketSize);
+          R1 = r2.loadPacket(1 * Traits::ResPacketSize);
+          R2 = r3.loadPacket(0 * Traits::ResPacketSize);
+          R3 = r3.loadPacket(1 * Traits::ResPacketSize);
+          traits.acc(C2,  alphav, R0);
+          traits.acc(C6,  alphav, R1);
+          traits.acc(C3,  alphav, R2);
+          traits.acc(C7,  alphav, R3);
+          r2.storePacket(0 * Traits::ResPacketSize, R0);
+          r2.storePacket(1 * Traits::ResPacketSize, R1);
+          r3.storePacket(0 * Traits::ResPacketSize, R2);
+          r3.storePacket(1 * Traits::ResPacketSize, R3);
           }
-
-          blB += nr*RhsProgress;
-          blA += LhsProgress;
         }
+      
+        // Deal with remaining columns of the rhs
+        for(Index j2=packet_cols4; j2<cols; j2++)
+        {
+          for(Index i=i1; i<actual_panel_end; i+=2*LhsProgress)
+          {
+          // One column at a time
+          const LhsScalar* blA = &blockA[i*strideA+offsetA*(2*Traits::LhsProgress)];
+          prefetch(&blA[0]);
 
-        ResPacket R0, R1, R2, R3;
-        ResPacket alphav = pset1<ResPacket>(alpha);
-
-        ResScalar* r0 = &res[(j2+0)*resStride + i];
-        ResScalar* r1 = r0 + resStride;
-        ResScalar* r2 = r1 + resStride;
-        ResScalar* r3 = r2 + resStride;
+          // gets res block as register
+          AccPacket C0, C4;
+          traits.initAcc(C0);
+          traits.initAcc(C4);
 
-                  R0 = ploadu<ResPacket>(r0);
-                  R1 = ploadu<ResPacket>(r1);
-        if(nr==4) R2 = ploadu<ResPacket>(r2);
-        if(nr==4) R3 = ploadu<ResPacket>(r3);
+          LinearMapper r0 = res.getLinearMapper(i, j2);
+          r0.prefetch(prefetch_res_offset);
 
-                  traits.acc(C0, alphav, R0);
-                  traits.acc(C1, alphav, R1);
-        if(nr==4) traits.acc(C2, alphav, R2);
-        if(nr==4) traits.acc(C3, alphav, R3);
+          // performs "inner" products
+          const RhsScalar* blB = &blockB[j2*strideB+offsetB];
+          LhsPacket A0, A1;
 
-                  pstoreu(r0, R0);
-                  pstoreu(r1, R1);
-        if(nr==4) pstoreu(r2, R2);
-        if(nr==4) pstoreu(r3, R3);
-      }
-      for(Index i=peeled_mc2; i<rows; i++)
-      {
-        const LhsScalar* blA = &blockA[i*strideA+offsetA];
-        prefetch(&blA[0]);
-
-        // gets a 1 x nr res block as registers
-        ResScalar C0(0), C1(0), C2(0), C3(0);
-        // TODO directly use blockB ???
-        const RhsScalar* blB = &blockB[j2*strideB+offsetB*nr];
-        for(Index k=0; k<depth; k++)
-        {
-          if(nr==2)
+          for(Index k=0; k<peeled_kc; k+=pk)
           {
-            LhsScalar A0;
-            RhsScalar B_0, B1;
-
-            A0 = blA[k];
-            B_0 = blB[0];
-            B1 = blB[1];
-            MADD(cj,A0,B_0,C0,B_0);
-            MADD(cj,A0,B1,C1,B1);
+            EIGEN_ASM_COMMENT("begin gebp micro kernel 2pX1");
+            RhsPacket B_0, B1;
+        
+#define EIGEN_GEBGP_ONESTEP(K) \
+            do {                                                                  \
+              EIGEN_ASM_COMMENT("begin step of gebp micro kernel 2pX1");          \
+              EIGEN_ASM_COMMENT("Note: these asm comments work around bug 935!"); \
+              traits.loadLhs(&blA[(0+2*K)*LhsProgress], A0);                      \
+              traits.loadLhs(&blA[(1+2*K)*LhsProgress], A1);                      \
+              traits.loadRhs(&blB[(0+K)*RhsProgress], B_0);                       \
+              traits.madd(A0, B_0, C0, B1);                                       \
+              traits.madd(A1, B_0, C4, B_0);                                      \
+              EIGEN_ASM_COMMENT("end step of gebp micro kernel 2pX1");            \
+            } while(false)
+        
+            EIGEN_GEBGP_ONESTEP(0);
+            EIGEN_GEBGP_ONESTEP(1);
+            EIGEN_GEBGP_ONESTEP(2);
+            EIGEN_GEBGP_ONESTEP(3);
+            EIGEN_GEBGP_ONESTEP(4);
+            EIGEN_GEBGP_ONESTEP(5);
+            EIGEN_GEBGP_ONESTEP(6);
+            EIGEN_GEBGP_ONESTEP(7);
+
+            blB += pk*RhsProgress;
+            blA += pk*2*Traits::LhsProgress;
+
+            EIGEN_ASM_COMMENT("end gebp micro kernel 2pX1");
           }
-          else
+
+          // process remaining peeled loop
+          for(Index k=peeled_kc; k<depth; k++)
           {
-            LhsScalar A0;
-            RhsScalar B_0, B1, B2, B3;
-
-            A0 = blA[k];
-            B_0 = blB[0];
-            B1 = blB[1];
-            B2 = blB[2];
-            B3 = blB[3];
-
-            MADD(cj,A0,B_0,C0,B_0);
-            MADD(cj,A0,B1,C1,B1);
-            MADD(cj,A0,B2,C2,B2);
-            MADD(cj,A0,B3,C3,B3);
+            RhsPacket B_0, B1;
+            EIGEN_GEBGP_ONESTEP(0);
+            blB += RhsProgress;
+            blA += 2*Traits::LhsProgress;
           }
+#undef EIGEN_GEBGP_ONESTEP
+          ResPacket R0, R1;
+          ResPacket alphav = pset1<ResPacket>(alpha);
 
-          blB += nr;
+          R0 = r0.loadPacket(0 * Traits::ResPacketSize);
+          R1 = r0.loadPacket(1 * Traits::ResPacketSize);
+          traits.acc(C0, alphav, R0);
+          traits.acc(C4, alphav, R1);
+          r0.storePacket(0 * Traits::ResPacketSize, R0);
+          r0.storePacket(1 * Traits::ResPacketSize, R1);
+          }
         }
-                  res[(j2+0)*resStride + i] += alpha*C0;
-                  res[(j2+1)*resStride + i] += alpha*C1;
-        if(nr==4) res[(j2+2)*resStride + i] += alpha*C2;
-        if(nr==4) res[(j2+3)*resStride + i] += alpha*C3;
       }
     }
-    // process remaining rhs/res columns one at a time
-    // => do the same but with nr==1
-    for(Index j2=packet_cols; j2<cols; j2++)
+    //---------- Process 1 * LhsProgress rows at once ----------
+    if(mr>=1*Traits::LhsProgress)
     {
-      // unpack B
-      traits.unpackRhs(depth, &blockB[j2*strideB+offsetB], unpackedB);
-
-      for(Index i=0; i<peeled_mc; i+=mr)
+      // loops on each largest micro horizontal panel of lhs (1*LhsProgress x depth)
+      for(Index i=peeled_mc2; i<peeled_mc1; i+=1*LhsProgress)
       {
-        const LhsScalar* blA = &blockA[i*strideA+offsetA*mr];
-        prefetch(&blA[0]);
+        // loops on each largest micro vertical panel of rhs (depth * nr)
+        for(Index j2=0; j2<packet_cols4; j2+=nr)
+        {
+          // We select a 1*Traits::LhsProgress x nr micro block of res which is entirely
+          // stored into 1 x nr registers.
+          
+          const LhsScalar* blA = &blockA[i*strideA+offsetA*(1*Traits::LhsProgress)];
+          prefetch(&blA[0]);
+
+          // gets res block as register
+          AccPacket C0, C1, C2, C3;
+          traits.initAcc(C0);
+          traits.initAcc(C1);
+          traits.initAcc(C2);
+          traits.initAcc(C3);
+
+          LinearMapper r0 = res.getLinearMapper(i, j2 + 0);
+          LinearMapper r1 = res.getLinearMapper(i, j2 + 1);
+          LinearMapper r2 = res.getLinearMapper(i, j2 + 2);
+          LinearMapper r3 = res.getLinearMapper(i, j2 + 3);
+
+          r0.prefetch(prefetch_res_offset);
+          r1.prefetch(prefetch_res_offset);
+          r2.prefetch(prefetch_res_offset);
+          r3.prefetch(prefetch_res_offset);
+
+          // performs "inner" products
+          const RhsScalar* blB = &blockB[j2*strideB+offsetB*nr];
+          prefetch(&blB[0]);
+          LhsPacket A0;
+
+          for(Index k=0; k<peeled_kc; k+=pk)
+          {
+            EIGEN_ASM_COMMENT("begin gebp micro kernel 1pX4");
+            RhsPacket B_0, B1, B2, B3;
+               
+#define EIGEN_GEBGP_ONESTEP(K) \
+            do {                                                                \
+              EIGEN_ASM_COMMENT("begin step of gebp micro kernel 1pX4");        \
+              EIGEN_ASM_COMMENT("Note: these asm comments work around bug 935!"); \
+              traits.loadLhs(&blA[(0+1*K)*LhsProgress], A0);                    \
+              traits.broadcastRhs(&blB[(0+4*K)*RhsProgress], B_0, B1, B2, B3);  \
+              traits.madd(A0, B_0, C0, B_0);                                    \
+              traits.madd(A0, B1,  C1, B1);                                     \
+              traits.madd(A0, B2,  C2, B2);                                     \
+              traits.madd(A0, B3,  C3, B3);                                     \
+              EIGEN_ASM_COMMENT("end step of gebp micro kernel 1pX4");          \
+            } while(false)
+            
+            internal::prefetch(blB+(48+0));
+            EIGEN_GEBGP_ONESTEP(0);
+            EIGEN_GEBGP_ONESTEP(1);
+            EIGEN_GEBGP_ONESTEP(2);
+            EIGEN_GEBGP_ONESTEP(3);
+            internal::prefetch(blB+(48+16));
+            EIGEN_GEBGP_ONESTEP(4);
+            EIGEN_GEBGP_ONESTEP(5);
+            EIGEN_GEBGP_ONESTEP(6);
+            EIGEN_GEBGP_ONESTEP(7);
+
+            blB += pk*4*RhsProgress;
+            blA += pk*1*LhsProgress;
+
+            EIGEN_ASM_COMMENT("end gebp micro kernel 1pX4");
+          }
+          // process remaining peeled loop
+          for(Index k=peeled_kc; k<depth; k++)
+          {
+            RhsPacket B_0, B1, B2, B3;
+            EIGEN_GEBGP_ONESTEP(0);
+            blB += 4*RhsProgress;
+            blA += 1*LhsProgress;
+          }
+#undef EIGEN_GEBGP_ONESTEP
 
-        // TODO move the res loads to the stores
+          ResPacket R0, R1;
+          ResPacket alphav = pset1<ResPacket>(alpha);
 
-        // get res block as registers
-        AccPacket C0, C4;
-        traits.initAcc(C0);
-        traits.initAcc(C4);
+          R0 = r0.loadPacket(0 * Traits::ResPacketSize);
+          R1 = r1.loadPacket(0 * Traits::ResPacketSize);
+          traits.acc(C0, alphav, R0);
+          traits.acc(C1,  alphav, R1);
+          r0.storePacket(0 * Traits::ResPacketSize, R0);
+          r1.storePacket(0 * Traits::ResPacketSize, R1);
+
+          R0 = r2.loadPacket(0 * Traits::ResPacketSize);
+          R1 = r3.loadPacket(0 * Traits::ResPacketSize);
+          traits.acc(C2,  alphav, R0);
+          traits.acc(C3,  alphav, R1);
+          r2.storePacket(0 * Traits::ResPacketSize, R0);
+          r3.storePacket(0 * Traits::ResPacketSize, R1);
+        }
 
-        const RhsScalar* blB = unpackedB;
-        for(Index k=0; k<depth; k++)
+        // Deal with remaining columns of the rhs
+        for(Index j2=packet_cols4; j2<cols; j2++)
         {
-          LhsPacket A0, A1;
-          RhsPacket B_0;
-          RhsPacket T0;
-
-          traits.loadLhs(&blA[0*LhsProgress], A0);
-          traits.loadLhs(&blA[1*LhsProgress], A1);
-          traits.loadRhs(&blB[0*RhsProgress], B_0);
-          traits.madd(A0,B_0,C0,T0);
-          traits.madd(A1,B_0,C4,B_0);
+          // One column at a time
+          const LhsScalar* blA = &blockA[i*strideA+offsetA*(1*Traits::LhsProgress)];
+          prefetch(&blA[0]);
 
-          blB += RhsProgress;
-          blA += 2*LhsProgress;
-        }
-        ResPacket R0, R4;
-        ResPacket alphav = pset1<ResPacket>(alpha);
+          // gets res block as register
+          AccPacket C0;
+          traits.initAcc(C0);
 
-        ResScalar* r0 = &res[(j2+0)*resStride + i];
+          LinearMapper r0 = res.getLinearMapper(i, j2);
 
-        R0 = ploadu<ResPacket>(r0);
-        R4 = ploadu<ResPacket>(r0+ResPacketSize);
+          // performs "inner" products
+          const RhsScalar* blB = &blockB[j2*strideB+offsetB];
+          LhsPacket A0;
 
-        traits.acc(C0, alphav, R0);
-        traits.acc(C4, alphav, R4);
+          for(Index k=0; k<peeled_kc; k+=pk)
+          {
+            EIGEN_ASM_COMMENT("begin gebp micro kernel 1pX1");
+            RhsPacket B_0;
+        
+#define EIGEN_GEBGP_ONESTEP(K) \
+            do {                                                                \
+              EIGEN_ASM_COMMENT("begin step of gebp micro kernel 1pX1");        \
+              EIGEN_ASM_COMMENT("Note: these asm comments work around bug 935!"); \
+              traits.loadLhs(&blA[(0+1*K)*LhsProgress], A0);                    \
+              traits.loadRhs(&blB[(0+K)*RhsProgress], B_0);                     \
+              traits.madd(A0, B_0, C0, B_0);                                    \
+              EIGEN_ASM_COMMENT("end step of gebp micro kernel 1pX1");          \
+            } while(false);
+
+            EIGEN_GEBGP_ONESTEP(0);
+            EIGEN_GEBGP_ONESTEP(1);
+            EIGEN_GEBGP_ONESTEP(2);
+            EIGEN_GEBGP_ONESTEP(3);
+            EIGEN_GEBGP_ONESTEP(4);
+            EIGEN_GEBGP_ONESTEP(5);
+            EIGEN_GEBGP_ONESTEP(6);
+            EIGEN_GEBGP_ONESTEP(7);
+
+            blB += pk*RhsProgress;
+            blA += pk*1*Traits::LhsProgress;
+
+            EIGEN_ASM_COMMENT("end gebp micro kernel 1pX1");
+          }
 
-        pstoreu(r0,               R0);
-        pstoreu(r0+ResPacketSize, R4);
+          // process remaining peeled loop
+          for(Index k=peeled_kc; k<depth; k++)
+          {
+            RhsPacket B_0;
+            EIGEN_GEBGP_ONESTEP(0);
+            blB += RhsProgress;
+            blA += 1*Traits::LhsProgress;
+          }
+#undef EIGEN_GEBGP_ONESTEP
+          ResPacket R0;
+          ResPacket alphav = pset1<ResPacket>(alpha);
+          R0 = r0.loadPacket(0 * Traits::ResPacketSize);
+          traits.acc(C0, alphav, R0);
+          r0.storePacket(0 * Traits::ResPacketSize, R0);
+        }
       }
-      if(rows-peeled_mc>=LhsProgress)
+    }
+    //---------- Process remaining rows, 1 at once ----------
+    if(peeled_mc1<rows)
+    {
+      // loop on each panel of the rhs
+      for(Index j2=0; j2<packet_cols4; j2+=nr)
       {
-        Index i = peeled_mc;
-        const LhsScalar* blA = &blockA[i*strideA+offsetA*LhsProgress];
-        prefetch(&blA[0]);
-
-        AccPacket C0;
-        traits.initAcc(C0);
-
-        const RhsScalar* blB = unpackedB;
-        for(Index k=0; k<depth; k++)
+        // loop on each row of the lhs (1*LhsProgress x depth)
+        for(Index i=peeled_mc1; i<rows; i+=1)
         {
-          LhsPacket A0;
-          RhsPacket B_0;
-          traits.loadLhs(blA, A0);
-          traits.loadRhs(blB, B_0);
-          traits.madd(A0, B_0, C0, B_0);
-          blB += RhsProgress;
-          blA += LhsProgress;
+          const LhsScalar* blA = &blockA[i*strideA+offsetA];
+          prefetch(&blA[0]);
+          const RhsScalar* blB = &blockB[j2*strideB+offsetB*nr];
+
+          // The following piece of code wont work for 512 bit registers
+          // Moreover, if LhsProgress==8 it assumes that there is a half packet of the same size
+          // as nr (which is currently 4) for the return type.
+          typedef typename unpacket_traits<SResPacket>::half SResPacketHalf;
+          if ((SwappedTraits::LhsProgress % 4) == 0 &&
+              (SwappedTraits::LhsProgress <= 8) &&
+              (SwappedTraits::LhsProgress!=8 || unpacket_traits<SResPacketHalf>::size==nr))
+          {
+            SAccPacket C0, C1, C2, C3;
+            straits.initAcc(C0);
+            straits.initAcc(C1);
+            straits.initAcc(C2);
+            straits.initAcc(C3);
+
+            const Index spk   = (std::max)(1,SwappedTraits::LhsProgress/4);
+            const Index endk  = (depth/spk)*spk;
+            const Index endk4 = (depth/(spk*4))*(spk*4);
+
+            Index k=0;
+            for(; k<endk4; k+=4*spk)
+            {
+              SLhsPacket A0,A1;
+              SRhsPacket B_0,B_1;
+
+              straits.loadLhsUnaligned(blB+0*SwappedTraits::LhsProgress, A0);
+              straits.loadLhsUnaligned(blB+1*SwappedTraits::LhsProgress, A1);
+
+              straits.loadRhsQuad(blA+0*spk, B_0);
+              straits.loadRhsQuad(blA+1*spk, B_1);
+              straits.madd(A0,B_0,C0,B_0);
+              straits.madd(A1,B_1,C1,B_1);
+
+              straits.loadLhsUnaligned(blB+2*SwappedTraits::LhsProgress, A0);
+              straits.loadLhsUnaligned(blB+3*SwappedTraits::LhsProgress, A1);
+              straits.loadRhsQuad(blA+2*spk, B_0);
+              straits.loadRhsQuad(blA+3*spk, B_1);
+              straits.madd(A0,B_0,C2,B_0);
+              straits.madd(A1,B_1,C3,B_1);
+
+              blB += 4*SwappedTraits::LhsProgress;
+              blA += 4*spk;
+            }
+            C0 = padd(padd(C0,C1),padd(C2,C3));
+            for(; k<endk; k+=spk)
+            {
+              SLhsPacket A0;
+              SRhsPacket B_0;
+
+              straits.loadLhsUnaligned(blB, A0);
+              straits.loadRhsQuad(blA, B_0);
+              straits.madd(A0,B_0,C0,B_0);
+
+              blB += SwappedTraits::LhsProgress;
+              blA += spk;
+            }
+            if(SwappedTraits::LhsProgress==8)
+            {
+              // Special case where we have to first reduce the accumulation register C0
+              typedef typename conditional<SwappedTraits::LhsProgress>=8,typename unpacket_traits<SResPacket>::half,SResPacket>::type SResPacketHalf;
+              typedef typename conditional<SwappedTraits::LhsProgress>=8,typename unpacket_traits<SLhsPacket>::half,SLhsPacket>::type SLhsPacketHalf;
+              typedef typename conditional<SwappedTraits::LhsProgress>=8,typename unpacket_traits<SLhsPacket>::half,SRhsPacket>::type SRhsPacketHalf;
+              typedef typename conditional<SwappedTraits::LhsProgress>=8,typename unpacket_traits<SAccPacket>::half,SAccPacket>::type SAccPacketHalf;
+
+              SResPacketHalf R = res.template gatherPacket<SResPacketHalf>(i, j2);
+              SResPacketHalf alphav = pset1<SResPacketHalf>(alpha);
+
+              if(depth-endk>0)
+              {
+                // We have to handle the last row of the rhs which corresponds to a half-packet
+                SLhsPacketHalf a0;
+                SRhsPacketHalf b0;
+                straits.loadLhsUnaligned(blB, a0);
+                straits.loadRhs(blA, b0);
+                SAccPacketHalf c0 = predux_downto4(C0);
+                straits.madd(a0,b0,c0,b0);
+                straits.acc(c0, alphav, R);
+              }
+              else
+              {
+                straits.acc(predux_downto4(C0), alphav, R);
+              }
+              res.scatterPacket(i, j2, R);
+            }
+            else
+            {
+              SResPacket R = res.template gatherPacket<SResPacket>(i, j2);
+              SResPacket alphav = pset1<SResPacket>(alpha);
+              straits.acc(C0, alphav, R);
+              res.scatterPacket(i, j2, R);
+            }
+          }
+          else // scalar path
+          {
+            // get a 1 x 4 res block as registers
+            ResScalar C0(0), C1(0), C2(0), C3(0);
+
+            for(Index k=0; k<depth; k++)
+            {
+              LhsScalar A0;
+              RhsScalar B_0, B_1;
+
+              A0 = blA[k];
+
+              B_0 = blB[0];
+              B_1 = blB[1];
+              CJMADD(cj,A0,B_0,C0,  B_0);
+              CJMADD(cj,A0,B_1,C1,  B_1);
+              
+              B_0 = blB[2];
+              B_1 = blB[3];
+              CJMADD(cj,A0,B_0,C2,  B_0);
+              CJMADD(cj,A0,B_1,C3,  B_1);
+              
+              blB += 4;
+            }
+            res(i, j2 + 0) += alpha * C0;
+            res(i, j2 + 1) += alpha * C1;
+            res(i, j2 + 2) += alpha * C2;
+            res(i, j2 + 3) += alpha * C3;
+          }
         }
-
-        ResPacket alphav = pset1<ResPacket>(alpha);
-        ResPacket R0 = ploadu<ResPacket>(&res[(j2+0)*resStride + i]);
-        traits.acc(C0, alphav, R0);
-        pstoreu(&res[(j2+0)*resStride + i], R0);
       }
-      for(Index i=peeled_mc2; i<rows; i++)
+      // remaining columns
+      for(Index j2=packet_cols4; j2<cols; j2++)
       {
-        const LhsScalar* blA = &blockA[i*strideA+offsetA];
-        prefetch(&blA[0]);
-
-        // gets a 1 x 1 res block as registers
-        ResScalar C0(0);
-        // FIXME directly use blockB ??
-        const RhsScalar* blB = &blockB[j2*strideB+offsetB];
-        for(Index k=0; k<depth; k++)
+        // loop on each row of the lhs (1*LhsProgress x depth)
+        for(Index i=peeled_mc1; i<rows; i+=1)
         {
-          LhsScalar A0 = blA[k];
-          RhsScalar B_0 = blB[k];
-          MADD(cj, A0, B_0, C0, B_0);
+          const LhsScalar* blA = &blockA[i*strideA+offsetA];
+          prefetch(&blA[0]);
+          // gets a 1 x 1 res block as registers
+          ResScalar C0(0);
+          const RhsScalar* blB = &blockB[j2*strideB+offsetB];
+          for(Index k=0; k<depth; k++)
+          {
+            LhsScalar A0 = blA[k];
+            RhsScalar B_0 = blB[k];
+            CJMADD(cj, A0, B_0, C0, B_0);
+          }
+          res(i, j2) += alpha * C0;
         }
-        res[(j2+0)*resStride + i] += alpha*C0;
       }
     }
   }
@@ -1114,81 +1685,193 @@ EIGEN_ASM_COMMENT("mybegin4");
 //
 //  32 33 34 35 ...
 //  36 36 38 39 ...
-template<typename Scalar, typename Index, int Pack1, int Pack2, int StorageOrder, bool Conjugate, bool PanelMode>
-struct gemm_pack_lhs
+template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, bool Conjugate, bool PanelMode>
+struct gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, ColMajor, Conjugate, PanelMode>
 {
-  EIGEN_DONT_INLINE void operator()(Scalar* blockA, const Scalar* EIGEN_RESTRICT _lhs, Index lhsStride, Index depth, Index rows, Index stride=0, Index offset=0);
+  typedef typename DataMapper::LinearMapper LinearMapper;
+  EIGEN_DONT_INLINE void operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride=0, Index offset=0);
 };
 
-template<typename Scalar, typename Index, int Pack1, int Pack2, int StorageOrder, bool Conjugate, bool PanelMode>
-EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, Pack1, Pack2, StorageOrder, Conjugate, PanelMode>
-  ::operator()(Scalar* blockA, const Scalar* EIGEN_RESTRICT _lhs, Index lhsStride, Index depth, Index rows, Index stride, Index offset)
+template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, bool Conjugate, bool PanelMode>
+EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, ColMajor, Conjugate, PanelMode>
+  ::operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride, Index offset)
 {
   typedef typename packet_traits<Scalar>::type Packet;
   enum { PacketSize = packet_traits<Scalar>::size };
 
   EIGEN_ASM_COMMENT("EIGEN PRODUCT PACK LHS");
-  EIGEN_UNUSED_VARIABLE(stride)
-  EIGEN_UNUSED_VARIABLE(offset)
+  EIGEN_UNUSED_VARIABLE(stride);
+  EIGEN_UNUSED_VARIABLE(offset);
   eigen_assert(((!PanelMode) && stride==0 && offset==0) || (PanelMode && stride>=depth && offset<=stride));
-  eigen_assert( (StorageOrder==RowMajor) || ((Pack1%PacketSize)==0 && Pack1<=4*PacketSize) );
+  eigen_assert( ((Pack1%PacketSize)==0 && Pack1<=4*PacketSize) || (Pack1<=4) );
   conj_if<NumTraits<Scalar>::IsComplex && Conjugate> cj;
-  const_blas_data_mapper<Scalar, Index, StorageOrder> lhs(_lhs,lhsStride);
   Index count = 0;
-  Index peeled_mc = (rows/Pack1)*Pack1;
-  for(Index i=0; i<peeled_mc; i+=Pack1)
+
+  const Index peeled_mc3 = Pack1>=3*PacketSize ? (rows/(3*PacketSize))*(3*PacketSize) : 0;
+  const Index peeled_mc2 = Pack1>=2*PacketSize ? peeled_mc3+((rows-peeled_mc3)/(2*PacketSize))*(2*PacketSize) : 0;
+  const Index peeled_mc1 = Pack1>=1*PacketSize ? (rows/(1*PacketSize))*(1*PacketSize) : 0;
+  const Index peeled_mc0 = Pack2>=1*PacketSize ? peeled_mc1
+                         : Pack2>1             ? (rows/Pack2)*Pack2 : 0;
+
+  Index i=0;
+
+  // Pack 3 packets
+  if(Pack1>=3*PacketSize)
   {
-    if(PanelMode) count += Pack1 * offset;
+    for(; i<peeled_mc3; i+=3*PacketSize)
+    {
+      if(PanelMode) count += (3*PacketSize) * offset;
 
-    if(StorageOrder==ColMajor)
+      for(Index k=0; k<depth; k++)
+      {
+        Packet A, B, C;
+        A = lhs.loadPacket(i+0*PacketSize, k);
+        B = lhs.loadPacket(i+1*PacketSize, k);
+        C = lhs.loadPacket(i+2*PacketSize, k);
+        pstore(blockA+count, cj.pconj(A)); count+=PacketSize;
+        pstore(blockA+count, cj.pconj(B)); count+=PacketSize;
+        pstore(blockA+count, cj.pconj(C)); count+=PacketSize;
+      }
+      if(PanelMode) count += (3*PacketSize) * (stride-offset-depth);
+    }
+  }
+  // Pack 2 packets
+  if(Pack1>=2*PacketSize)
+  {
+    for(; i<peeled_mc2; i+=2*PacketSize)
     {
+      if(PanelMode) count += (2*PacketSize) * offset;
+
       for(Index k=0; k<depth; k++)
       {
-        Packet A, B, C, D;
-        if(Pack1>=1*PacketSize) A = ploadu<Packet>(&lhs(i+0*PacketSize, k));
-        if(Pack1>=2*PacketSize) B = ploadu<Packet>(&lhs(i+1*PacketSize, k));
-        if(Pack1>=3*PacketSize) C = ploadu<Packet>(&lhs(i+2*PacketSize, k));
-        if(Pack1>=4*PacketSize) D = ploadu<Packet>(&lhs(i+3*PacketSize, k));
-        if(Pack1>=1*PacketSize) { pstore(blockA+count, cj.pconj(A)); count+=PacketSize; }
-        if(Pack1>=2*PacketSize) { pstore(blockA+count, cj.pconj(B)); count+=PacketSize; }
-        if(Pack1>=3*PacketSize) { pstore(blockA+count, cj.pconj(C)); count+=PacketSize; }
-        if(Pack1>=4*PacketSize) { pstore(blockA+count, cj.pconj(D)); count+=PacketSize; }
+        Packet A, B;
+        A = lhs.loadPacket(i+0*PacketSize, k);
+        B = lhs.loadPacket(i+1*PacketSize, k);
+        pstore(blockA+count, cj.pconj(A)); count+=PacketSize;
+        pstore(blockA+count, cj.pconj(B)); count+=PacketSize;
       }
+      if(PanelMode) count += (2*PacketSize) * (stride-offset-depth);
     }
-    else
+  }
+  // Pack 1 packets
+  if(Pack1>=1*PacketSize)
+  {
+    for(; i<peeled_mc1; i+=1*PacketSize)
     {
+      if(PanelMode) count += (1*PacketSize) * offset;
+
+      for(Index k=0; k<depth; k++)
+      {
+        Packet A;
+        A = lhs.loadPacket(i+0*PacketSize, k);
+        pstore(blockA+count, cj.pconj(A));
+        count+=PacketSize;
+      }
+      if(PanelMode) count += (1*PacketSize) * (stride-offset-depth);
+    }
+  }
+  // Pack scalars
+  if(Pack2<PacketSize && Pack2>1)
+  {
+    for(; i<peeled_mc0; i+=Pack2)
+    {
+      if(PanelMode) count += Pack2 * offset;
+
       for(Index k=0; k<depth; k++)
+        for(Index w=0; w<Pack2; w++)
+          blockA[count++] = cj(lhs(i+w, k));
+
+      if(PanelMode) count += Pack2 * (stride-offset-depth);
+    }
+  }
+  for(; i<rows; i++)
+  {
+    if(PanelMode) count += offset;
+    for(Index k=0; k<depth; k++)
+      blockA[count++] = cj(lhs(i, k));
+    if(PanelMode) count += (stride-offset-depth);
+  }
+}
+
+template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, bool Conjugate, bool PanelMode>
+struct gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, RowMajor, Conjugate, PanelMode>
+{
+  typedef typename DataMapper::LinearMapper LinearMapper;
+  EIGEN_DONT_INLINE void operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride=0, Index offset=0);
+};
+
+template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, bool Conjugate, bool PanelMode>
+EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, RowMajor, Conjugate, PanelMode>
+  ::operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride, Index offset)
+{
+  typedef typename packet_traits<Scalar>::type Packet;
+  enum { PacketSize = packet_traits<Scalar>::size };
+
+  EIGEN_ASM_COMMENT("EIGEN PRODUCT PACK LHS");
+  EIGEN_UNUSED_VARIABLE(stride);
+  EIGEN_UNUSED_VARIABLE(offset);
+  eigen_assert(((!PanelMode) && stride==0 && offset==0) || (PanelMode && stride>=depth && offset<=stride));
+  conj_if<NumTraits<Scalar>::IsComplex && Conjugate> cj;
+  Index count = 0;
+
+//   const Index peeled_mc3 = Pack1>=3*PacketSize ? (rows/(3*PacketSize))*(3*PacketSize) : 0;
+//   const Index peeled_mc2 = Pack1>=2*PacketSize ? peeled_mc3+((rows-peeled_mc3)/(2*PacketSize))*(2*PacketSize) : 0;
+//   const Index peeled_mc1 = Pack1>=1*PacketSize ? (rows/(1*PacketSize))*(1*PacketSize) : 0;
+
+  int pack = Pack1;
+  Index i = 0;
+  while(pack>0)
+  {
+    Index remaining_rows = rows-i;
+    Index peeled_mc = i+(remaining_rows/pack)*pack;
+    for(; i<peeled_mc; i+=pack)
+    {
+      if(PanelMode) count += pack * offset;
+
+      const Index peeled_k = (depth/PacketSize)*PacketSize;
+      Index k=0;
+      if(pack>=PacketSize)
+      {
+        for(; k<peeled_k; k+=PacketSize)
+        {
+          for (Index m = 0; m < pack; m += PacketSize)
+          {
+            PacketBlock<Packet> kernel;
+            for (int p = 0; p < PacketSize; ++p) kernel.packet[p] = lhs.loadPacket(i+p+m, k);
+            ptranspose(kernel);
+            for (int p = 0; p < PacketSize; ++p) pstore(blockA+count+m+(pack)*p, cj.pconj(kernel.packet[p]));
+          }
+          count += PacketSize*pack;
+        }
+      }
+      for(; k<depth; k++)
       {
-        // TODO add a vectorized transpose here
         Index w=0;
-        for(; w<Pack1-3; w+=4)
+        for(; w<pack-3; w+=4)
         {
           Scalar a(cj(lhs(i+w+0, k))),
-                  b(cj(lhs(i+w+1, k))),
-                  c(cj(lhs(i+w+2, k))),
-                  d(cj(lhs(i+w+3, k)));
+                 b(cj(lhs(i+w+1, k))),
+                 c(cj(lhs(i+w+2, k))),
+                 d(cj(lhs(i+w+3, k)));
           blockA[count++] = a;
           blockA[count++] = b;
           blockA[count++] = c;
           blockA[count++] = d;
         }
-        if(Pack1%4)
-          for(;w<Pack1;++w)
+        if(pack%4)
+          for(;w<pack;++w)
             blockA[count++] = cj(lhs(i+w, k));
       }
+
+      if(PanelMode) count += pack * (stride-offset-depth);
     }
-    if(PanelMode) count += Pack1 * (stride-offset-depth);
-  }
-  if(rows-peeled_mc>=Pack2)
-  {
-    if(PanelMode) count += Pack2*offset;
-    for(Index k=0; k<depth; k++)
-      for(Index w=0; w<Pack2; w++)
-        blockA[count++] = cj(lhs(peeled_mc+w, k));
-    if(PanelMode) count += Pack2 * (stride-offset-depth);
-    peeled_mc += Pack2;
+
+    pack -= PacketSize;
+    if(pack<Pack2 && (pack+PacketSize)!=Pack2)
+      pack = Pack2;
   }
-  for(Index i=peeled_mc; i<rows; i++)
+
+  for(; i<rows; i++)
   {
     if(PanelMode) count += offset;
     for(Index k=0; k<depth; k++)
@@ -1204,53 +1887,123 @@ EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, Pack1, Pack2, StorageOrder,
 //  4  5  6  7   16 17 18 19   25 28
 //  8  9 10 11   20 21 22 23   26 29
 //  .  .  .  .    .  .  .  .    .  .
-template<typename Scalar, typename Index, int nr, bool Conjugate, bool PanelMode>
-struct gemm_pack_rhs<Scalar, Index, nr, ColMajor, Conjugate, PanelMode>
+template<typename Scalar, typename Index, typename DataMapper, int nr, bool Conjugate, bool PanelMode>
+struct gemm_pack_rhs<Scalar, Index, DataMapper, nr, ColMajor, Conjugate, PanelMode>
 {
   typedef typename packet_traits<Scalar>::type Packet;
+  typedef typename DataMapper::LinearMapper LinearMapper;
   enum { PacketSize = packet_traits<Scalar>::size };
-  EIGEN_DONT_INLINE void operator()(Scalar* blockB, const Scalar* rhs, Index rhsStride, Index depth, Index cols, Index stride=0, Index offset=0);
+  EIGEN_DONT_INLINE void operator()(Scalar* blockB, const DataMapper& rhs, Index depth, Index cols, Index stride=0, Index offset=0);
 };
 
-template<typename Scalar, typename Index, int nr, bool Conjugate, bool PanelMode>
-EIGEN_DONT_INLINE void gemm_pack_rhs<Scalar, Index, nr, ColMajor, Conjugate, PanelMode>
-  ::operator()(Scalar* blockB, const Scalar* rhs, Index rhsStride, Index depth, Index cols, Index stride, Index offset)
+template<typename Scalar, typename Index, typename DataMapper, int nr, bool Conjugate, bool PanelMode>
+EIGEN_DONT_INLINE void gemm_pack_rhs<Scalar, Index, DataMapper, nr, ColMajor, Conjugate, PanelMode>
+  ::operator()(Scalar* blockB, const DataMapper& rhs, Index depth, Index cols, Index stride, Index offset)
 {
   EIGEN_ASM_COMMENT("EIGEN PRODUCT PACK RHS COLMAJOR");
-  EIGEN_UNUSED_VARIABLE(stride)
-  EIGEN_UNUSED_VARIABLE(offset)
+  EIGEN_UNUSED_VARIABLE(stride);
+  EIGEN_UNUSED_VARIABLE(offset);
   eigen_assert(((!PanelMode) && stride==0 && offset==0) || (PanelMode && stride>=depth && offset<=stride));
   conj_if<NumTraits<Scalar>::IsComplex && Conjugate> cj;
-  Index packet_cols = (cols/nr) * nr;
+  Index packet_cols8 = nr>=8 ? (cols/8) * 8 : 0;
+  Index packet_cols4 = nr>=4 ? (cols/4) * 4 : 0;
   Index count = 0;
-  for(Index j2=0; j2<packet_cols; j2+=nr)
+  const Index peeled_k = (depth/PacketSize)*PacketSize;
+//   if(nr>=8)
+//   {
+//     for(Index j2=0; j2<packet_cols8; j2+=8)
+//     {
+//       // skip what we have before
+//       if(PanelMode) count += 8 * offset;
+//       const Scalar* b0 = &rhs[(j2+0)*rhsStride];
+//       const Scalar* b1 = &rhs[(j2+1)*rhsStride];
+//       const Scalar* b2 = &rhs[(j2+2)*rhsStride];
+//       const Scalar* b3 = &rhs[(j2+3)*rhsStride];
+//       const Scalar* b4 = &rhs[(j2+4)*rhsStride];
+//       const Scalar* b5 = &rhs[(j2+5)*rhsStride];
+//       const Scalar* b6 = &rhs[(j2+6)*rhsStride];
+//       const Scalar* b7 = &rhs[(j2+7)*rhsStride];
+//       Index k=0;
+//       if(PacketSize==8) // TODO enbale vectorized transposition for PacketSize==4
+//       {
+//         for(; k<peeled_k; k+=PacketSize) {
+//           PacketBlock<Packet> kernel;
+//           for (int p = 0; p < PacketSize; ++p) {
+//             kernel.packet[p] = ploadu<Packet>(&rhs[(j2+p)*rhsStride+k]);
+//           }
+//           ptranspose(kernel);
+//           for (int p = 0; p < PacketSize; ++p) {
+//             pstoreu(blockB+count, cj.pconj(kernel.packet[p]));
+//             count+=PacketSize;
+//           }
+//         }
+//       }
+//       for(; k<depth; k++)
+//       {
+//         blockB[count+0] = cj(b0[k]);
+//         blockB[count+1] = cj(b1[k]);
+//         blockB[count+2] = cj(b2[k]);
+//         blockB[count+3] = cj(b3[k]);
+//         blockB[count+4] = cj(b4[k]);
+//         blockB[count+5] = cj(b5[k]);
+//         blockB[count+6] = cj(b6[k]);
+//         blockB[count+7] = cj(b7[k]);
+//         count += 8;
+//       }
+//       // skip what we have after
+//       if(PanelMode) count += 8 * (stride-offset-depth);
+//     }
+//   }
+
+  if(nr>=4)
   {
-    // skip what we have before
-    if(PanelMode) count += nr * offset;
-    const Scalar* b0 = &rhs[(j2+0)*rhsStride];
-    const Scalar* b1 = &rhs[(j2+1)*rhsStride];
-    const Scalar* b2 = &rhs[(j2+2)*rhsStride];
-    const Scalar* b3 = &rhs[(j2+3)*rhsStride];
-    for(Index k=0; k<depth; k++)
+    for(Index j2=packet_cols8; j2<packet_cols4; j2+=4)
     {
-                blockB[count+0] = cj(b0[k]);
-                blockB[count+1] = cj(b1[k]);
-      if(nr==4) blockB[count+2] = cj(b2[k]);
-      if(nr==4) blockB[count+3] = cj(b3[k]);
-      count += nr;
+      // skip what we have before
+      if(PanelMode) count += 4 * offset;
+      const LinearMapper dm0 = rhs.getLinearMapper(0, j2 + 0);
+      const LinearMapper dm1 = rhs.getLinearMapper(0, j2 + 1);
+      const LinearMapper dm2 = rhs.getLinearMapper(0, j2 + 2);
+      const LinearMapper dm3 = rhs.getLinearMapper(0, j2 + 3);
+
+      Index k=0;
+      if((PacketSize%4)==0) // TODO enable vectorized transposition for PacketSize==2 ??
+      {
+        for(; k<peeled_k; k+=PacketSize) {
+          PacketBlock<Packet,(PacketSize%4)==0?4:PacketSize> kernel;
+          kernel.packet[0] = dm0.loadPacket(k);
+          kernel.packet[1%PacketSize] = dm1.loadPacket(k);
+          kernel.packet[2%PacketSize] = dm2.loadPacket(k);
+          kernel.packet[3%PacketSize] = dm3.loadPacket(k);
+          ptranspose(kernel);
+          pstoreu(blockB+count+0*PacketSize, cj.pconj(kernel.packet[0]));
+          pstoreu(blockB+count+1*PacketSize, cj.pconj(kernel.packet[1%PacketSize]));
+          pstoreu(blockB+count+2*PacketSize, cj.pconj(kernel.packet[2%PacketSize]));
+          pstoreu(blockB+count+3*PacketSize, cj.pconj(kernel.packet[3%PacketSize]));
+          count+=4*PacketSize;
+        }
+      }
+      for(; k<depth; k++)
+      {
+        blockB[count+0] = cj(dm0(k));
+        blockB[count+1] = cj(dm1(k));
+        blockB[count+2] = cj(dm2(k));
+        blockB[count+3] = cj(dm3(k));
+        count += 4;
+      }
+      // skip what we have after
+      if(PanelMode) count += 4 * (stride-offset-depth);
     }
-    // skip what we have after
-    if(PanelMode) count += nr * (stride-offset-depth);
   }
 
   // copy the remaining columns one at a time (nr==1)
-  for(Index j2=packet_cols; j2<cols; ++j2)
+  for(Index j2=packet_cols4; j2<cols; ++j2)
   {
     if(PanelMode) count += offset;
-    const Scalar* b0 = &rhs[(j2+0)*rhsStride];
+    const LinearMapper dm0 = rhs.getLinearMapper(0, j2);
     for(Index k=0; k<depth; k++)
     {
-      blockB[count] = cj(b0[k]);
+      blockB[count] = cj(dm0(k));
       count += 1;
     }
     if(PanelMode) count += (stride-offset-depth);
@@ -1258,48 +2011,93 @@ EIGEN_DONT_INLINE void gemm_pack_rhs<Scalar, Index, nr, ColMajor, Conjugate, Pan
 }
 
 // this version is optimized for row major matrices
-template<typename Scalar, typename Index, int nr, bool Conjugate, bool PanelMode>
-struct gemm_pack_rhs<Scalar, Index, nr, RowMajor, Conjugate, PanelMode>
+template<typename Scalar, typename Index, typename DataMapper, int nr, bool Conjugate, bool PanelMode>
+struct gemm_pack_rhs<Scalar, Index, DataMapper, nr, RowMajor, Conjugate, PanelMode>
 {
+  typedef typename packet_traits<Scalar>::type Packet;
+  typedef typename DataMapper::LinearMapper LinearMapper;
   enum { PacketSize = packet_traits<Scalar>::size };
-  EIGEN_DONT_INLINE void operator()(Scalar* blockB, const Scalar* rhs, Index rhsStride, Index depth, Index cols, Index stride=0, Index offset=0);
+  EIGEN_DONT_INLINE void operator()(Scalar* blockB, const DataMapper& rhs, Index depth, Index cols, Index stride=0, Index offset=0);
 };
 
-template<typename Scalar, typename Index, int nr, bool Conjugate, bool PanelMode>
-EIGEN_DONT_INLINE void gemm_pack_rhs<Scalar, Index, nr, RowMajor, Conjugate, PanelMode>
-  ::operator()(Scalar* blockB, const Scalar* rhs, Index rhsStride, Index depth, Index cols, Index stride, Index offset)
+template<typename Scalar, typename Index, typename DataMapper, int nr, bool Conjugate, bool PanelMode>
+EIGEN_DONT_INLINE void gemm_pack_rhs<Scalar, Index, DataMapper, nr, RowMajor, Conjugate, PanelMode>
+  ::operator()(Scalar* blockB, const DataMapper& rhs, Index depth, Index cols, Index stride, Index offset)
 {
   EIGEN_ASM_COMMENT("EIGEN PRODUCT PACK RHS ROWMAJOR");
-  EIGEN_UNUSED_VARIABLE(stride)
-  EIGEN_UNUSED_VARIABLE(offset)
+  EIGEN_UNUSED_VARIABLE(stride);
+  EIGEN_UNUSED_VARIABLE(offset);
   eigen_assert(((!PanelMode) && stride==0 && offset==0) || (PanelMode && stride>=depth && offset<=stride));
   conj_if<NumTraits<Scalar>::IsComplex && Conjugate> cj;
-  Index packet_cols = (cols/nr) * nr;
+  Index packet_cols8 = nr>=8 ? (cols/8) * 8 : 0;
+  Index packet_cols4 = nr>=4 ? (cols/4) * 4 : 0;
   Index count = 0;
-  for(Index j2=0; j2<packet_cols; j2+=nr)
+
+//   if(nr>=8)
+//   {
+//     for(Index j2=0; j2<packet_cols8; j2+=8)
+//     {
+//       // skip what we have before
+//       if(PanelMode) count += 8 * offset;
+//       for(Index k=0; k<depth; k++)
+//       {
+//         if (PacketSize==8) {
+//           Packet A = ploadu<Packet>(&rhs[k*rhsStride + j2]);
+//           pstoreu(blockB+count, cj.pconj(A));
+//         } else if (PacketSize==4) {
+//           Packet A = ploadu<Packet>(&rhs[k*rhsStride + j2]);
+//           Packet B = ploadu<Packet>(&rhs[k*rhsStride + j2 + PacketSize]);
+//           pstoreu(blockB+count, cj.pconj(A));
+//           pstoreu(blockB+count+PacketSize, cj.pconj(B));
+//         } else {
+//           const Scalar* b0 = &rhs[k*rhsStride + j2];
+//           blockB[count+0] = cj(b0[0]);
+//           blockB[count+1] = cj(b0[1]);
+//           blockB[count+2] = cj(b0[2]);
+//           blockB[count+3] = cj(b0[3]);
+//           blockB[count+4] = cj(b0[4]);
+//           blockB[count+5] = cj(b0[5]);
+//           blockB[count+6] = cj(b0[6]);
+//           blockB[count+7] = cj(b0[7]);
+//         }
+//         count += 8;
+//       }
+//       // skip what we have after
+//       if(PanelMode) count += 8 * (stride-offset-depth);
+//     }
+//   }
+  if(nr>=4)
   {
-    // skip what we have before
-    if(PanelMode) count += nr * offset;
-    for(Index k=0; k<depth; k++)
+    for(Index j2=packet_cols8; j2<packet_cols4; j2+=4)
     {
-      const Scalar* b0 = &rhs[k*rhsStride + j2];
-                blockB[count+0] = cj(b0[0]);
-                blockB[count+1] = cj(b0[1]);
-      if(nr==4) blockB[count+2] = cj(b0[2]);
-      if(nr==4) blockB[count+3] = cj(b0[3]);
-      count += nr;
+      // skip what we have before
+      if(PanelMode) count += 4 * offset;
+      for(Index k=0; k<depth; k++)
+      {
+        if (PacketSize==4) {
+          Packet A = rhs.loadPacket(k, j2);
+          pstoreu(blockB+count, cj.pconj(A));
+          count += PacketSize;
+        } else {
+          const LinearMapper dm0 = rhs.getLinearMapper(k, j2);
+          blockB[count+0] = cj(dm0(0));
+          blockB[count+1] = cj(dm0(1));
+          blockB[count+2] = cj(dm0(2));
+          blockB[count+3] = cj(dm0(3));
+          count += 4;
+        }
+      }
+      // skip what we have after
+      if(PanelMode) count += 4 * (stride-offset-depth);
     }
-    // skip what we have after
-    if(PanelMode) count += nr * (stride-offset-depth);
   }
   // copy the remaining columns one at a time (nr==1)
-  for(Index j2=packet_cols; j2<cols; ++j2)
+  for(Index j2=packet_cols4; j2<cols; ++j2)
   {
     if(PanelMode) count += offset;
-    const Scalar* b0 = &rhs[j2];
     for(Index k=0; k<depth; k++)
     {
-      blockB[count] = cj(b0[k*rhsStride]);
+      blockB[count] = cj(rhs(k, j2));
       count += 1;
     }
     if(PanelMode) count += stride-offset-depth;
@@ -1312,8 +2110,8 @@ EIGEN_DONT_INLINE void gemm_pack_rhs<Scalar, Index, nr, RowMajor, Conjugate, Pan
   * \sa setCpuCacheSize */
 inline std::ptrdiff_t l1CacheSize()
 {
-  std::ptrdiff_t l1, l2;
-  internal::manage_caching_sizes(GetAction, &l1, &l2);
+  std::ptrdiff_t l1, l2, l3;
+  internal::manage_caching_sizes(GetAction, &l1, &l2, &l3);
   return l1;
 }
 
@@ -1321,19 +2119,29 @@ inline std::ptrdiff_t l1CacheSize()
   * \sa setCpuCacheSize */
 inline std::ptrdiff_t l2CacheSize()
 {
-  std::ptrdiff_t l1, l2;
-  internal::manage_caching_sizes(GetAction, &l1, &l2);
+  std::ptrdiff_t l1, l2, l3;
+  internal::manage_caching_sizes(GetAction, &l1, &l2, &l3);
   return l2;
 }
 
+/** \returns the currently set level 3 cpu cache size (in bytes) used to estimate the ideal blocking size paramete\
+rs.                                                                                                                
+* \sa setCpuCacheSize */
+inline std::ptrdiff_t l3CacheSize()
+{
+  std::ptrdiff_t l1, l2, l3;
+  internal::manage_caching_sizes(GetAction, &l1, &l2, &l3);
+  return l3;
+}
+
 /** Set the cpu L1 and L2 cache sizes (in bytes).
   * These values are use to adjust the size of the blocks
   * for the algorithms working per blocks.
   *
   * \sa computeProductBlockingSizes */
-inline void setCpuCacheSizes(std::ptrdiff_t l1, std::ptrdiff_t l2)
+inline void setCpuCacheSizes(std::ptrdiff_t l1, std::ptrdiff_t l2, std::ptrdiff_t l3)
 {
-  internal::manage_caching_sizes(SetAction, &l1, &l2);
+  internal::manage_caching_sizes(SetAction, &l1, &l2, &l3);
 }
 
 } // end namespace Eigen
diff --git a/eigen/Eigen/src/Core/products/GeneralMatrixMatrix.h b/eigen/Eigen/src/Core/products/GeneralMatrixMatrix.h
index cfd2f00..6440e1d 100644
--- a/eigen/Eigen/src/Core/products/GeneralMatrixMatrix.h
+++ b/eigen/Eigen/src/Core/products/GeneralMatrixMatrix.h
@@ -10,7 +10,7 @@
 #ifndef EIGEN_GENERAL_MATRIX_MATRIX_H
 #define EIGEN_GENERAL_MATRIX_MATRIX_H
 
-namespace Eigen { 
+namespace Eigen {
 
 namespace internal {
 
@@ -23,7 +23,9 @@ template<
   typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs>
 struct general_matrix_matrix_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,RowMajor>
 {
-  typedef typename scalar_product_traits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+  typedef gebp_traits<RhsScalar,LhsScalar> Traits;
+
+  typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
   static EIGEN_STRONG_INLINE void run(
     Index rows, Index cols, Index depth,
     const LhsScalar* lhs, Index lhsStride,
@@ -51,42 +53,44 @@ template<
 struct general_matrix_matrix_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,ColMajor>
 {
 
-typedef typename scalar_product_traits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+typedef gebp_traits<LhsScalar,RhsScalar> Traits;
+
+typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
 static void run(Index rows, Index cols, Index depth,
   const LhsScalar* _lhs, Index lhsStride,
   const RhsScalar* _rhs, Index rhsStride,
-  ResScalar* res, Index resStride,
+  ResScalar* _res, Index resStride,
   ResScalar alpha,
   level3_blocking<LhsScalar,RhsScalar>& blocking,
   GemmParallelInfo<Index>* info = 0)
 {
-  const_blas_data_mapper<LhsScalar, Index, LhsStorageOrder> lhs(_lhs,lhsStride);
-  const_blas_data_mapper<RhsScalar, Index, RhsStorageOrder> rhs(_rhs,rhsStride);
-
-  typedef gebp_traits<LhsScalar,RhsScalar> Traits;
+  typedef const_blas_data_mapper<LhsScalar, Index, LhsStorageOrder> LhsMapper;
+  typedef const_blas_data_mapper<RhsScalar, Index, RhsStorageOrder> RhsMapper;
+  typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper;
+  LhsMapper lhs(_lhs,lhsStride);
+  RhsMapper rhs(_rhs,rhsStride);
+  ResMapper res(_res, resStride);
 
   Index kc = blocking.kc();                   // cache block size along the K direction
   Index mc = (std::min)(rows,blocking.mc());  // cache block size along the M direction
-  //Index nc = blocking.nc(); // cache block size along the N direction
+  Index nc = (std::min)(cols,blocking.nc());  // cache block size along the N direction
 
-  gemm_pack_lhs<LhsScalar, Index, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
-  gemm_pack_rhs<RhsScalar, Index, Traits::nr, RhsStorageOrder> pack_rhs;
-  gebp_kernel<LhsScalar, RhsScalar, Index, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp;
+  gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+  gemm_pack_rhs<RhsScalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs;
+  gebp_kernel<LhsScalar, RhsScalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp;
 
 #ifdef EIGEN_HAS_OPENMP
   if(info)
   {
     // this is the parallel version!
-    Index tid = omp_get_thread_num();
-    Index threads = omp_get_num_threads();
-    
-    std::size_t sizeA = kc*mc;
-    std::size_t sizeW = kc*Traits::WorkSpaceFactor;
-    ei_declare_aligned_stack_constructed_variable(LhsScalar, blockA, sizeA, 0);
-    ei_declare_aligned_stack_constructed_variable(RhsScalar, w, sizeW, 0);
-    
-    RhsScalar* blockB = blocking.blockB();
-    eigen_internal_assert(blockB!=0);
+    int tid = omp_get_thread_num();
+    int threads = omp_get_num_threads();
+
+    LhsScalar* blockA = blocking.blockA();
+    eigen_internal_assert(blockA!=0);
+
+    std::size_t sizeB = kc*nc;
+    ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, sizeB, 0);
 
     // For each horizontal panel of the rhs, and corresponding vertical panel of the lhs...
     for(Index k=0; k<depth; k+=kc)
@@ -94,56 +98,56 @@ static void run(Index rows, Index cols, Index depth,
       const Index actual_kc = (std::min)(k+kc,depth)-k; // => rows of B', and cols of the A'
 
       // In order to reduce the chance that a thread has to wait for the other,
-      // let's start by packing A'.
-      pack_lhs(blockA, &lhs(0,k), lhsStride, actual_kc, mc);
+      // let's start by packing B'.
+      pack_rhs(blockB, rhs.getSubMapper(k,0), actual_kc, nc);
 
-      // Pack B_k to B' in a parallel fashion:
-      // each thread packs the sub block B_k,j to B'_j where j is the thread id.
+      // Pack A_k to A' in a parallel fashion:
+      // each thread packs the sub block A_k,i to A'_i where i is the thread id.
 
-      // However, before copying to B'_j, we have to make sure that no other thread is still using it,
+      // However, before copying to A'_i, we have to make sure that no other thread is still using it,
       // i.e., we test that info[tid].users equals 0.
       // Then, we set info[tid].users to the number of threads to mark that all other threads are going to use it.
       while(info[tid].users!=0) {}
       info[tid].users += threads;
 
-      pack_rhs(blockB+info[tid].rhs_start*actual_kc, &rhs(k,info[tid].rhs_start), rhsStride, actual_kc, info[tid].rhs_length);
+      pack_lhs(blockA+info[tid].lhs_start*actual_kc, lhs.getSubMapper(info[tid].lhs_start,k), actual_kc, info[tid].lhs_length);
 
-      // Notify the other threads that the part B'_j is ready to go.
+      // Notify the other threads that the part A'_i is ready to go.
       info[tid].sync = k;
 
-      // Computes C_i += A' * B' per B'_j
-      for(Index shift=0; shift<threads; ++shift)
+      // Computes C_i += A' * B' per A'_i
+      for(int shift=0; shift<threads; ++shift)
       {
-        Index j = (tid+shift)%threads;
+        int i = (tid+shift)%threads;
 
-        // At this point we have to make sure that B'_j has been updated by the thread j,
+        // At this point we have to make sure that A'_i has been updated by the thread i,
         // we use testAndSetOrdered to mimic a volatile access.
         // However, no need to wait for the B' part which has been updated by the current thread!
-        if(shift>0)
-          while(info[j].sync!=k) {}
+        if (shift>0) {
+          while(info[i].sync!=k) {
+          }
+        }
 
-        gebp(res+info[j].rhs_start*resStride, resStride, blockA, blockB+info[j].rhs_start*actual_kc, mc, actual_kc, info[j].rhs_length, alpha, -1,-1,0,0, w);
+        gebp(res.getSubMapper(info[i].lhs_start, 0), blockA+info[i].lhs_start*actual_kc, blockB, info[i].lhs_length, actual_kc, nc, alpha);
       }
 
-      // Then keep going as usual with the remaining A'
-      for(Index i=mc; i<rows; i+=mc)
+      // Then keep going as usual with the remaining B'
+      for(Index j=nc; j<cols; j+=nc)
       {
-        const Index actual_mc = (std::min)(i+mc,rows)-i;
+        const Index actual_nc = (std::min)(j+nc,cols)-j;
 
-        // pack A_i,k to A'
-        pack_lhs(blockA, &lhs(i,k), lhsStride, actual_kc, actual_mc);
+        // pack B_k,j to B'
+        pack_rhs(blockB, rhs.getSubMapper(k,j), actual_kc, actual_nc);
 
-        // C_i += A' * B'
-        gebp(res+i, resStride, blockA, blockB, actual_mc, actual_kc, cols, alpha, -1,-1,0,0, w);
+        // C_j += A' * B'
+        gebp(res.getSubMapper(0, j), blockA, blockB, rows, actual_kc, actual_nc, alpha);
       }
 
-      // Release all the sub blocks B'_j of B' for the current thread,
+      // Release all the sub blocks A'_i of A' for the current thread,
       // i.e., we simply decrement the number of users by 1
-      for(Index j=0; j<threads; ++j)
-      {
+      for(Index i=0; i<threads; ++i)
         #pragma omp atomic
-        info[j].users -= 1;
-      }
+        info[i].users -= 1;
     }
   }
   else
@@ -153,38 +157,42 @@ static void run(Index rows, Index cols, Index depth,
 
     // this is the sequential version!
     std::size_t sizeA = kc*mc;
-    std::size_t sizeB = kc*cols;
-    std::size_t sizeW = kc*Traits::WorkSpaceFactor;
+    std::size_t sizeB = kc*nc;
 
     ei_declare_aligned_stack_constructed_variable(LhsScalar, blockA, sizeA, blocking.blockA());
     ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, sizeB, blocking.blockB());
-    ei_declare_aligned_stack_constructed_variable(RhsScalar, blockW, sizeW, blocking.blockW());
+
+    const bool pack_rhs_once = mc!=rows && kc==depth && nc==cols;
 
     // For each horizontal panel of the rhs, and corresponding panel of the lhs...
-    // (==GEMM_VAR1)
-    for(Index k2=0; k2<depth; k2+=kc)
+    for(Index i2=0; i2<rows; i2+=mc)
     {
-      const Index actual_kc = (std::min)(k2+kc,depth)-k2;
-
-      // OK, here we have selected one horizontal panel of rhs and one vertical panel of lhs.
-      // => Pack rhs's panel into a sequential chunk of memory (L2 caching)
-      // Note that this panel will be read as many times as the number of blocks in the lhs's
-      // vertical panel which is, in practice, a very low number.
-      pack_rhs(blockB, &rhs(k2,0), rhsStride, actual_kc, cols);
+      const Index actual_mc = (std::min)(i2+mc,rows)-i2;
 
-      // For each mc x kc block of the lhs's vertical panel...
-      // (==GEPP_VAR1)
-      for(Index i2=0; i2<rows; i2+=mc)
+      for(Index k2=0; k2<depth; k2+=kc)
       {
-        const Index actual_mc = (std::min)(i2+mc,rows)-i2;
-
-        // We pack the lhs's block into a sequential chunk of memory (L1 caching)
-        // Note that this block will be read a very high number of times, which is equal to the number of
-        // micro vertical panel of the large rhs's panel (e.g., cols/4 times).
-        pack_lhs(blockA, &lhs(i2,k2), lhsStride, actual_kc, actual_mc);
-
-        // Everything is packed, we can now call the block * panel kernel:
-        gebp(res+i2, resStride, blockA, blockB, actual_mc, actual_kc, cols, alpha, -1, -1, 0, 0, blockW);
+        const Index actual_kc = (std::min)(k2+kc,depth)-k2;
+
+        // OK, here we have selected one horizontal panel of rhs and one vertical panel of lhs.
+        // => Pack lhs's panel into a sequential chunk of memory (L2/L3 caching)
+        // Note that this panel will be read as many times as the number of blocks in the rhs's
+        // horizontal panel which is, in practice, a very low number.
+        pack_lhs(blockA, lhs.getSubMapper(i2,k2), actual_kc, actual_mc);
+
+        // For each kc x nc block of the rhs's horizontal panel...
+        for(Index j2=0; j2<cols; j2+=nc)
+        {
+          const Index actual_nc = (std::min)(j2+nc,cols)-j2;
+
+          // We pack the rhs's block into a sequential chunk of memory (L2 caching)
+          // Note that this block will be read a very high number of times, which is equal to the number of
+          // micro horizontal panel of the large rhs's panel (e.g., rows/12 times).
+          if((!pack_rhs_once) || i2==0)
+            pack_rhs(blockB, rhs.getSubMapper(k2,j2), actual_kc, actual_nc);
+
+          // Everything is packed, we can now call the panel * block kernel:
+          gebp(res.getSubMapper(i2, j2), blockA, blockB, actual_mc, actual_kc, actual_nc, alpha);
+        }
       }
     }
   }
@@ -193,26 +201,21 @@ static void run(Index rows, Index cols, Index depth,
 };
 
 /*********************************************************************************
-*  Specialization of GeneralProduct<> for "large" GEMM, i.e.,
+*  Specialization of generic_product_impl for "large" GEMM, i.e.,
 *  implementation of the high level wrapper to general_matrix_matrix_product
 **********************************************************************************/
 
-template<typename Lhs, typename Rhs>
-struct traits<GeneralProduct<Lhs,Rhs,GemmProduct> >
- : traits<ProductBase<GeneralProduct<Lhs,Rhs,GemmProduct>, Lhs, Rhs> >
-{};
-
 template<typename Scalar, typename Index, typename Gemm, typename Lhs, typename Rhs, typename Dest, typename BlockingType>
 struct gemm_functor
 {
-  gemm_functor(const Lhs& lhs, const Rhs& rhs, Dest& dest, const Scalar& actualAlpha,
-                  BlockingType& blocking)
+  gemm_functor(const Lhs& lhs, const Rhs& rhs, Dest& dest, const Scalar& actualAlpha, BlockingType& blocking)
     : m_lhs(lhs), m_rhs(rhs), m_dest(dest), m_actualAlpha(actualAlpha), m_blocking(blocking)
   {}
 
-  void initParallelSession() const
+  void initParallelSession(Index num_threads) const
   {
-    m_blocking.allocateB();
+    m_blocking.initParallel(m_lhs.rows(), m_rhs.cols(), m_lhs.cols(), num_threads);
+    m_blocking.allocateA();
   }
 
   void operator() (Index row, Index rows, Index col=0, Index cols=-1, GemmParallelInfo<Index>* info=0) const
@@ -221,12 +224,14 @@ struct gemm_functor
       cols = m_rhs.cols();
 
     Gemm::run(rows, cols, m_lhs.cols(),
-              /*(const Scalar*)*/&m_lhs.coeffRef(row,0), m_lhs.outerStride(),
-              /*(const Scalar*)*/&m_rhs.coeffRef(0,col), m_rhs.outerStride(),
+              &m_lhs.coeffRef(row,0), m_lhs.outerStride(),
+              &m_rhs.coeffRef(0,col), m_rhs.outerStride(),
               (Scalar*)&(m_dest.coeffRef(row,col)), m_dest.outerStride(),
               m_actualAlpha, m_blocking, info);
   }
 
+  typedef typename Gemm::Traits Traits;
+
   protected:
     const Lhs& m_lhs;
     const Rhs& m_rhs;
@@ -247,29 +252,27 @@ class level3_blocking
   protected:
     LhsScalar* m_blockA;
     RhsScalar* m_blockB;
-    RhsScalar* m_blockW;
 
-    DenseIndex m_mc;
-    DenseIndex m_nc;
-    DenseIndex m_kc;
+    Index m_mc;
+    Index m_nc;
+    Index m_kc;
 
   public:
 
     level3_blocking()
-      : m_blockA(0), m_blockB(0), m_blockW(0), m_mc(0), m_nc(0), m_kc(0)
+      : m_blockA(0), m_blockB(0), m_mc(0), m_nc(0), m_kc(0)
     {}
 
-    inline DenseIndex mc() const { return m_mc; }
-    inline DenseIndex nc() const { return m_nc; }
-    inline DenseIndex kc() const { return m_kc; }
+    inline Index mc() const { return m_mc; }
+    inline Index nc() const { return m_nc; }
+    inline Index kc() const { return m_kc; }
 
     inline LhsScalar* blockA() { return m_blockA; }
     inline RhsScalar* blockB() { return m_blockB; }
-    inline RhsScalar* blockW() { return m_blockW; }
 };
 
 template<int StorageOrder, typename _LhsScalar, typename _RhsScalar, int MaxRows, int MaxCols, int MaxDepth, int KcFactor>
-class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, MaxDepth, KcFactor, true>
+class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, MaxDepth, KcFactor, true /* == FiniteAtCompileTime */>
   : public level3_blocking<
       typename conditional<StorageOrder==RowMajor,_RhsScalar,_LhsScalar>::type,
       typename conditional<StorageOrder==RowMajor,_LhsScalar,_RhsScalar>::type>
@@ -284,29 +287,38 @@ class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, M
     typedef gebp_traits<LhsScalar,RhsScalar> Traits;
     enum {
       SizeA = ActualRows * MaxDepth,
-      SizeB = ActualCols * MaxDepth,
-      SizeW = MaxDepth * Traits::WorkSpaceFactor
+      SizeB = ActualCols * MaxDepth
     };
 
-    EIGEN_ALIGN16 LhsScalar m_staticA[SizeA];
-    EIGEN_ALIGN16 RhsScalar m_staticB[SizeB];
-    EIGEN_ALIGN16 RhsScalar m_staticW[SizeW];
+#if EIGEN_MAX_STATIC_ALIGN_BYTES >= EIGEN_DEFAULT_ALIGN_BYTES
+    EIGEN_ALIGN_MAX LhsScalar m_staticA[SizeA];
+    EIGEN_ALIGN_MAX RhsScalar m_staticB[SizeB];
+#else
+    EIGEN_ALIGN_MAX char m_staticA[SizeA * sizeof(LhsScalar) + EIGEN_DEFAULT_ALIGN_BYTES-1];
+    EIGEN_ALIGN_MAX char m_staticB[SizeB * sizeof(RhsScalar) + EIGEN_DEFAULT_ALIGN_BYTES-1];
+#endif
 
   public:
 
-    gemm_blocking_space(DenseIndex /*rows*/, DenseIndex /*cols*/, DenseIndex /*depth*/)
+    gemm_blocking_space(Index /*rows*/, Index /*cols*/, Index /*depth*/, Index /*num_threads*/, bool /*full_rows = false*/)
     {
       this->m_mc = ActualRows;
       this->m_nc = ActualCols;
       this->m_kc = MaxDepth;
+#if EIGEN_MAX_STATIC_ALIGN_BYTES >= EIGEN_DEFAULT_ALIGN_BYTES
       this->m_blockA = m_staticA;
       this->m_blockB = m_staticB;
-      this->m_blockW = m_staticW;
+#else
+      this->m_blockA = reinterpret_cast<LhsScalar*>((internal::UIntPtr(m_staticA) + (EIGEN_DEFAULT_ALIGN_BYTES-1)) & ~std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1));
+      this->m_blockB = reinterpret_cast<RhsScalar*>((internal::UIntPtr(m_staticB) + (EIGEN_DEFAULT_ALIGN_BYTES-1)) & ~std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1));
+#endif
     }
 
+    void initParallel(Index, Index, Index, Index)
+    {}
+
     inline void allocateA() {}
     inline void allocateB() {}
-    inline void allocateW() {}
     inline void allocateAll() {}
 };
 
@@ -323,22 +335,42 @@ class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, M
     typedef typename conditional<Transpose,_LhsScalar,_RhsScalar>::type RhsScalar;
     typedef gebp_traits<LhsScalar,RhsScalar> Traits;
 
-    DenseIndex m_sizeA;
-    DenseIndex m_sizeB;
-    DenseIndex m_sizeW;
+    Index m_sizeA;
+    Index m_sizeB;
 
   public:
 
-    gemm_blocking_space(DenseIndex rows, DenseIndex cols, DenseIndex depth)
+    gemm_blocking_space(Index rows, Index cols, Index depth, Index num_threads, bool l3_blocking)
+    {
+      this->m_mc = Transpose ? cols : rows;
+      this->m_nc = Transpose ? rows : cols;
+      this->m_kc = depth;
+
+      if(l3_blocking)
+      {
+        computeProductBlockingSizes<LhsScalar,RhsScalar,KcFactor>(this->m_kc, this->m_mc, this->m_nc, num_threads);
+      }
+      else  // no l3 blocking
+      {
+        Index n = this->m_nc;
+        computeProductBlockingSizes<LhsScalar,RhsScalar,KcFactor>(this->m_kc, this->m_mc, n, num_threads);
+      }
+
+      m_sizeA = this->m_mc * this->m_kc;
+      m_sizeB = this->m_kc * this->m_nc;
+    }
+
+    void initParallel(Index rows, Index cols, Index depth, Index num_threads)
     {
       this->m_mc = Transpose ? cols : rows;
       this->m_nc = Transpose ? rows : cols;
       this->m_kc = depth;
 
-      computeProductBlockingSizes<LhsScalar,RhsScalar,KcFactor>(this->m_kc, this->m_mc, this->m_nc);
+      eigen_internal_assert(this->m_blockA==0 && this->m_blockB==0);
+      Index m = this->m_mc;
+      computeProductBlockingSizes<LhsScalar,RhsScalar,KcFactor>(this->m_kc, m, this->m_nc, num_threads);
       m_sizeA = this->m_mc * this->m_kc;
       m_sizeB = this->m_kc * this->m_nc;
-      m_sizeW = this->m_kc*Traits::WorkSpaceFactor;
     }
 
     void allocateA()
@@ -353,81 +385,108 @@ class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, M
         this->m_blockB = aligned_new<RhsScalar>(m_sizeB);
     }
 
-    void allocateW()
-    {
-      if(this->m_blockW==0)
-        this->m_blockW = aligned_new<RhsScalar>(m_sizeW);
-    }
-
     void allocateAll()
     {
       allocateA();
       allocateB();
-      allocateW();
     }
 
     ~gemm_blocking_space()
     {
       aligned_delete(this->m_blockA, m_sizeA);
       aligned_delete(this->m_blockB, m_sizeB);
-      aligned_delete(this->m_blockW, m_sizeW);
     }
 };
 
 } // end namespace internal
 
+namespace internal {
+
 template<typename Lhs, typename Rhs>
-class GeneralProduct<Lhs, Rhs, GemmProduct>
-  : public ProductBase<GeneralProduct<Lhs,Rhs,GemmProduct>, Lhs, Rhs>
+struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
+  : generic_product_impl_base<Lhs,Rhs,generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct> >
 {
-    enum {
-      MaxDepthAtCompileTime = EIGEN_SIZE_MIN_PREFER_FIXED(Lhs::MaxColsAtCompileTime,Rhs::MaxRowsAtCompileTime)
-    };
-  public:
-    EIGEN_PRODUCT_PUBLIC_INTERFACE(GeneralProduct)
-    
-    typedef typename  Lhs::Scalar LhsScalar;
-    typedef typename  Rhs::Scalar RhsScalar;
-    typedef           Scalar      ResScalar;
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  typedef typename Lhs::Scalar LhsScalar;
+  typedef typename Rhs::Scalar RhsScalar;
 
-    GeneralProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs)
-    {
-#if !(defined(EIGEN_NO_STATIC_ASSERT) && defined(EIGEN_NO_DEBUG))
-      typedef internal::scalar_product_op<LhsScalar,RhsScalar> BinOp;
-      EIGEN_CHECK_BINARY_COMPATIBILIY(BinOp,LhsScalar,RhsScalar);
-#endif
-    }
+  typedef internal::blas_traits<Lhs> LhsBlasTraits;
+  typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
+  typedef typename internal::remove_all<ActualLhsType>::type ActualLhsTypeCleaned;
 
-    template<typename Dest> void scaleAndAddTo(Dest& dst, const Scalar& alpha) const
-    {
-      eigen_assert(dst.rows()==m_lhs.rows() && dst.cols()==m_rhs.cols());
-      if(m_lhs.cols()==0 || m_lhs.rows()==0 || m_rhs.cols()==0)
-        return;
+  typedef internal::blas_traits<Rhs> RhsBlasTraits;
+  typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
+  typedef typename internal::remove_all<ActualRhsType>::type ActualRhsTypeCleaned;
 
-      typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(m_lhs);
-      typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(m_rhs);
+  enum {
+    MaxDepthAtCompileTime = EIGEN_SIZE_MIN_PREFER_FIXED(Lhs::MaxColsAtCompileTime,Rhs::MaxRowsAtCompileTime)
+  };
 
-      Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(m_lhs)
-                                 * RhsBlasTraits::extractScalarFactor(m_rhs);
+  typedef generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,CoeffBasedProductMode> lazyproduct;
 
-      typedef internal::gemm_blocking_space<(Dest::Flags&RowMajorBit) ? RowMajor : ColMajor,LhsScalar,RhsScalar,
-              Dest::MaxRowsAtCompileTime,Dest::MaxColsAtCompileTime,MaxDepthAtCompileTime> BlockingType;
+  template<typename Dst>
+  static void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0)
+      lazyproduct::evalTo(dst, lhs, rhs);
+    else
+    {
+      dst.setZero();
+      scaleAndAddTo(dst, lhs, rhs, Scalar(1));
+    }
+  }
 
-      typedef internal::gemm_functor<
-        Scalar, Index,
-        internal::general_matrix_matrix_product<
-          Index,
-          LhsScalar, (_ActualLhsType::Flags&RowMajorBit) ? RowMajor : ColMajor, bool(LhsBlasTraits::NeedToConjugate),
-          RhsScalar, (_ActualRhsType::Flags&RowMajorBit) ? RowMajor : ColMajor, bool(RhsBlasTraits::NeedToConjugate),
-          (Dest::Flags&RowMajorBit) ? RowMajor : ColMajor>,
-        _ActualLhsType, _ActualRhsType, Dest, BlockingType> GemmFunctor;
+  template<typename Dst>
+  static void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0)
+      lazyproduct::addTo(dst, lhs, rhs);
+    else
+      scaleAndAddTo(dst,lhs, rhs, Scalar(1));
+  }
 
-      BlockingType blocking(dst.rows(), dst.cols(), lhs.cols());
+  template<typename Dst>
+  static void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0)
+      lazyproduct::subTo(dst, lhs, rhs);
+    else
+      scaleAndAddTo(dst, lhs, rhs, Scalar(-1));
+  }
 
-      internal::parallelize_gemm<(Dest::MaxRowsAtCompileTime>32 || Dest::MaxRowsAtCompileTime==Dynamic)>(GemmFunctor(lhs, rhs, dst, actualAlpha, blocking), this->rows(), this->cols(), Dest::Flags&RowMajorBit);
-    }
+  template<typename Dest>
+  static void scaleAndAddTo(Dest& dst, const Lhs& a_lhs, const Rhs& a_rhs, const Scalar& alpha)
+  {
+    eigen_assert(dst.rows()==a_lhs.rows() && dst.cols()==a_rhs.cols());
+    if(a_lhs.cols()==0 || a_lhs.rows()==0 || a_rhs.cols()==0)
+      return;
+
+    typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(a_lhs);
+    typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(a_rhs);
+
+    Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(a_lhs)
+                               * RhsBlasTraits::extractScalarFactor(a_rhs);
+
+    typedef internal::gemm_blocking_space<(Dest::Flags&RowMajorBit) ? RowMajor : ColMajor,LhsScalar,RhsScalar,
+            Dest::MaxRowsAtCompileTime,Dest::MaxColsAtCompileTime,MaxDepthAtCompileTime> BlockingType;
+
+    typedef internal::gemm_functor<
+      Scalar, Index,
+      internal::general_matrix_matrix_product<
+        Index,
+        LhsScalar, (ActualLhsTypeCleaned::Flags&RowMajorBit) ? RowMajor : ColMajor, bool(LhsBlasTraits::NeedToConjugate),
+        RhsScalar, (ActualRhsTypeCleaned::Flags&RowMajorBit) ? RowMajor : ColMajor, bool(RhsBlasTraits::NeedToConjugate),
+        (Dest::Flags&RowMajorBit) ? RowMajor : ColMajor>,
+      ActualLhsTypeCleaned, ActualRhsTypeCleaned, Dest, BlockingType> GemmFunctor;
+
+    BlockingType blocking(dst.rows(), dst.cols(), lhs.cols(), 1, true);
+    internal::parallelize_gemm<(Dest::MaxRowsAtCompileTime>32 || Dest::MaxRowsAtCompileTime==Dynamic)>
+        (GemmFunctor(lhs, rhs, dst, actualAlpha, blocking), a_lhs.rows(), a_rhs.cols(), a_lhs.cols(), Dest::Flags&RowMajorBit);
+  }
 };
 
+} // end namespace internal
+
 } // end namespace Eigen
 
 #endif // EIGEN_GENERAL_MATRIX_MATRIX_H
diff --git a/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h b/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h
index 5c37639..ad38bcf 100644
--- a/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h
+++ b/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h
@@ -20,7 +20,7 @@ namespace internal {
 /**********************************************************************
 * This file implements a general A * B product while
 * evaluating only one triangular part of the product.
-* This is more general version of self adjoint product (C += A A^T)
+* This is a more general version of self adjoint product (C += A A^T)
 * as the level 3 SYRK Blas routine.
 **********************************************************************/
 
@@ -40,15 +40,16 @@ template <typename Index, typename LhsScalar, int LhsStorageOrder, bool Conjugat
                           typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs, int  UpLo, int Version>
 struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,RowMajor,UpLo,Version>
 {
-  typedef typename scalar_product_traits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+  typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
   static EIGEN_STRONG_INLINE void run(Index size, Index depth,const LhsScalar* lhs, Index lhsStride,
-                                      const RhsScalar* rhs, Index rhsStride, ResScalar* res, Index resStride, const ResScalar& alpha)
+                                      const RhsScalar* rhs, Index rhsStride, ResScalar* res, Index resStride,
+                                      const ResScalar& alpha, level3_blocking<RhsScalar,LhsScalar>& blocking)
   {
     general_matrix_matrix_triangular_product<Index,
         RhsScalar, RhsStorageOrder==RowMajor ? ColMajor : RowMajor, ConjugateRhs,
         LhsScalar, LhsStorageOrder==RowMajor ? ColMajor : RowMajor, ConjugateLhs,
         ColMajor, UpLo==Lower?Upper:Lower>
-      ::run(size,depth,rhs,rhsStride,lhs,lhsStride,res,resStride,alpha);
+      ::run(size,depth,rhs,rhsStride,lhs,lhsStride,res,resStride,alpha,blocking);
   }
 };
 
@@ -56,32 +57,36 @@ template <typename Index, typename LhsScalar, int LhsStorageOrder, bool Conjugat
                           typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs, int  UpLo, int Version>
 struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,ColMajor,UpLo,Version>
 {
-  typedef typename scalar_product_traits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+  typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
   static EIGEN_STRONG_INLINE void run(Index size, Index depth,const LhsScalar* _lhs, Index lhsStride,
-                                      const RhsScalar* _rhs, Index rhsStride, ResScalar* res, Index resStride, const ResScalar& alpha)
+                                      const RhsScalar* _rhs, Index rhsStride, ResScalar* _res, Index resStride,
+                                      const ResScalar& alpha, level3_blocking<LhsScalar,RhsScalar>& blocking)
   {
-    const_blas_data_mapper<LhsScalar, Index, LhsStorageOrder> lhs(_lhs,lhsStride);
-    const_blas_data_mapper<RhsScalar, Index, RhsStorageOrder> rhs(_rhs,rhsStride);
-
     typedef gebp_traits<LhsScalar,RhsScalar> Traits;
 
-    Index kc = depth; // cache block size along the K direction
-    Index mc = size;  // cache block size along the M direction
-    Index nc = size;  // cache block size along the N direction
-    computeProductBlockingSizes<LhsScalar,RhsScalar>(kc, mc, nc);
+    typedef const_blas_data_mapper<LhsScalar, Index, LhsStorageOrder> LhsMapper;
+    typedef const_blas_data_mapper<RhsScalar, Index, RhsStorageOrder> RhsMapper;
+    typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper;
+    LhsMapper lhs(_lhs,lhsStride);
+    RhsMapper rhs(_rhs,rhsStride);
+    ResMapper res(_res, resStride);
+
+    Index kc = blocking.kc();
+    Index mc = (std::min)(size,blocking.mc());
+
     // !!! mc must be a multiple of nr:
     if(mc > Traits::nr)
       mc = (mc/Traits::nr)*Traits::nr;
 
-    std::size_t sizeW = kc*Traits::WorkSpaceFactor;
-    std::size_t sizeB = sizeW + kc*size;
-    ei_declare_aligned_stack_constructed_variable(LhsScalar, blockA, kc*mc, 0);
-    ei_declare_aligned_stack_constructed_variable(RhsScalar, allocatedBlockB, sizeB, 0);
-    RhsScalar* blockB = allocatedBlockB + sizeW;
-    
-    gemm_pack_lhs<LhsScalar, Index, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
-    gemm_pack_rhs<RhsScalar, Index, Traits::nr, RhsStorageOrder> pack_rhs;
-    gebp_kernel <LhsScalar, RhsScalar, Index, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp;
+    std::size_t sizeA = kc*mc;
+    std::size_t sizeB = kc*size;
+
+    ei_declare_aligned_stack_constructed_variable(LhsScalar, blockA, sizeA, blocking.blockA());
+    ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, sizeB, blocking.blockB());
+
+    gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+    gemm_pack_rhs<RhsScalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs;
+    gebp_kernel<LhsScalar, RhsScalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp;
     tribb_kernel<LhsScalar, RhsScalar, Index, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs, UpLo> sybb;
 
     for(Index k2=0; k2<depth; k2+=kc)
@@ -89,29 +94,30 @@ struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,
       const Index actual_kc = (std::min)(k2+kc,depth)-k2;
 
       // note that the actual rhs is the transpose/adjoint of mat
-      pack_rhs(blockB, &rhs(k2,0), rhsStride, actual_kc, size);
+      pack_rhs(blockB, rhs.getSubMapper(k2,0), actual_kc, size);
 
       for(Index i2=0; i2<size; i2+=mc)
       {
         const Index actual_mc = (std::min)(i2+mc,size)-i2;
 
-        pack_lhs(blockA, &lhs(i2, k2), lhsStride, actual_kc, actual_mc);
+        pack_lhs(blockA, lhs.getSubMapper(i2, k2), actual_kc, actual_mc);
 
         // the selected actual_mc * size panel of res is split into three different part:
         //  1 - before the diagonal => processed with gebp or skipped
         //  2 - the actual_mc x actual_mc symmetric block => processed with a special kernel
         //  3 - after the diagonal => processed with gebp or skipped
         if (UpLo==Lower)
-          gebp(res+i2, resStride, blockA, blockB, actual_mc, actual_kc, (std::min)(size,i2), alpha,
-               -1, -1, 0, 0, allocatedBlockB);
+          gebp(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc,
+               (std::min)(size,i2), alpha, -1, -1, 0, 0);
 
-        sybb(res+resStride*i2 + i2, resStride, blockA, blockB + actual_kc*i2, actual_mc, actual_kc, alpha, allocatedBlockB);
+
+        sybb(_res+resStride*i2 + i2, resStride, blockA, blockB + actual_kc*i2, actual_mc, actual_kc, alpha);
 
         if (UpLo==Upper)
         {
           Index j2 = i2+actual_mc;
-          gebp(res+resStride*j2+i2, resStride, blockA, blockB+actual_kc*j2, actual_mc, actual_kc, (std::max)(Index(0), size-j2), alpha,
-               -1, -1, 0, 0, allocatedBlockB);
+          gebp(res.getSubMapper(i2, j2), blockA, blockB+actual_kc*j2, actual_mc,
+               actual_kc, (std::max)(Index(0), size-j2), alpha, -1, -1, 0, 0);
         }
       }
     }
@@ -132,14 +138,17 @@ struct tribb_kernel
 {
   typedef gebp_traits<LhsScalar,RhsScalar,ConjLhs,ConjRhs> Traits;
   typedef typename Traits::ResScalar ResScalar;
-  
+
   enum {
-    BlockSize  = EIGEN_PLAIN_ENUM_MAX(mr,nr)
+    BlockSize  = meta_least_common_multiple<EIGEN_PLAIN_ENUM_MAX(mr,nr),EIGEN_PLAIN_ENUM_MIN(mr,nr)>::ret
   };
-  void operator()(ResScalar* res, Index resStride, const LhsScalar* blockA, const RhsScalar* blockB, Index size, Index depth, const ResScalar& alpha, RhsScalar* workspace)
+  void operator()(ResScalar* _res, Index resStride, const LhsScalar* blockA, const RhsScalar* blockB, Index size, Index depth, const ResScalar& alpha)
   {
-    gebp_kernel<LhsScalar, RhsScalar, Index, mr, nr, ConjLhs, ConjRhs> gebp_kernel;
-    Matrix<ResScalar,BlockSize,BlockSize,ColMajor> buffer;
+    typedef blas_data_mapper<ResScalar, Index, ColMajor> ResMapper;
+    ResMapper res(_res, resStride);
+    gebp_kernel<LhsScalar, RhsScalar, Index, ResMapper, mr, nr, ConjLhs, ConjRhs> gebp_kernel;
+
+    Matrix<ResScalar,BlockSize,BlockSize,ColMajor> buffer((internal::constructor_without_unaligned_array_assert()));
 
     // let's process the block per panel of actual_mc x BlockSize,
     // again, each is split into three parts, etc.
@@ -149,20 +158,20 @@ struct tribb_kernel
       const RhsScalar* actual_b = blockB+j*depth;
 
       if(UpLo==Upper)
-        gebp_kernel(res+j*resStride, resStride, blockA, actual_b, j, depth, actualBlockSize, alpha,
-                    -1, -1, 0, 0, workspace);
+        gebp_kernel(res.getSubMapper(0, j), blockA, actual_b, j, depth, actualBlockSize, alpha,
+                    -1, -1, 0, 0);
 
       // selfadjoint micro block
       {
         Index i = j;
         buffer.setZero();
         // 1 - apply the kernel on the temporary buffer
-        gebp_kernel(buffer.data(), BlockSize, blockA+depth*i, actual_b, actualBlockSize, depth, actualBlockSize, alpha,
-                    -1, -1, 0, 0, workspace);
+        gebp_kernel(ResMapper(buffer.data(), BlockSize), blockA+depth*i, actual_b, actualBlockSize, depth, actualBlockSize, alpha,
+                    -1, -1, 0, 0);
         // 2 - triangular accumulation
         for(Index j1=0; j1<actualBlockSize; ++j1)
         {
-          ResScalar* r = res + (j+j1)*resStride + i;
+          ResScalar* r = &res(i, j + j1);
           for(Index i1=UpLo==Lower ? j1 : 0;
               UpLo==Lower ? i1<actualBlockSize : i1<=j1; ++i1)
             r[i1] += buffer(i1,j1);
@@ -172,8 +181,8 @@ struct tribb_kernel
       if(UpLo==Lower)
       {
         Index i = j+actualBlockSize;
-        gebp_kernel(res+j*resStride+i, resStride, blockA+depth*i, actual_b, size-i, depth, actualBlockSize, alpha,
-                    -1, -1, 0, 0, workspace);
+        gebp_kernel(res.getSubMapper(i, j), blockA+depth*i, actual_b, size-i, 
+                    depth, actualBlockSize, alpha, -1, -1, 0, 0);
       }
     }
   }
@@ -190,10 +199,9 @@ struct general_product_to_triangular_selector;
 template<typename MatrixType, typename ProductType, int UpLo>
 struct general_product_to_triangular_selector<MatrixType,ProductType,UpLo,true>
 {
-  static void run(MatrixType& mat, const ProductType& prod, const typename MatrixType::Scalar& alpha)
+  static void run(MatrixType& mat, const ProductType& prod, const typename MatrixType::Scalar& alpha, bool beta)
   {
     typedef typename MatrixType::Scalar Scalar;
-    typedef typename MatrixType::Index Index;
     
     typedef typename internal::remove_all<typename ProductType::LhsNested>::type Lhs;
     typedef internal::blas_traits<Lhs> LhsBlasTraits;
@@ -209,6 +217,9 @@ struct general_product_to_triangular_selector<MatrixType,ProductType,UpLo,true>
 
     Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs().derived()) * RhsBlasTraits::extractScalarFactor(prod.rhs().derived());
 
+    if(!beta)
+      mat.template triangularView<UpLo>().setZero();
+
     enum {
       StorageOrder = (internal::traits<MatrixType>::Flags&RowMajorBit) ? RowMajor : ColMajor,
       UseLhsDirectly = _ActualLhs::InnerStrideAtCompileTime==1,
@@ -236,10 +247,8 @@ struct general_product_to_triangular_selector<MatrixType,ProductType,UpLo,true>
 template<typename MatrixType, typename ProductType, int UpLo>
 struct general_product_to_triangular_selector<MatrixType,ProductType,UpLo,false>
 {
-  static void run(MatrixType& mat, const ProductType& prod, const typename MatrixType::Scalar& alpha)
+  static void run(MatrixType& mat, const ProductType& prod, const typename MatrixType::Scalar& alpha, bool beta)
   {
-    typedef typename MatrixType::Index Index;
-    
     typedef typename internal::remove_all<typename ProductType::LhsNested>::type Lhs;
     typedef internal::blas_traits<Lhs> LhsBlasTraits;
     typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhs;
@@ -254,23 +263,42 @@ struct general_product_to_triangular_selector<MatrixType,ProductType,UpLo,false>
 
     typename ProductType::Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs().derived()) * RhsBlasTraits::extractScalarFactor(prod.rhs().derived());
 
+    if(!beta)
+      mat.template triangularView<UpLo>().setZero();
+
+    enum {
+      IsRowMajor = (internal::traits<MatrixType>::Flags&RowMajorBit) ? 1 : 0,
+      LhsIsRowMajor = _ActualLhs::Flags&RowMajorBit ? 1 : 0,
+      RhsIsRowMajor = _ActualRhs::Flags&RowMajorBit ? 1 : 0
+    };
+
+    Index size = mat.cols();
+    Index depth = actualLhs.cols();
+
+    typedef internal::gemm_blocking_space<IsRowMajor ? RowMajor : ColMajor,typename Lhs::Scalar,typename Rhs::Scalar,
+          MatrixType::MaxColsAtCompileTime, MatrixType::MaxColsAtCompileTime, _ActualRhs::MaxColsAtCompileTime> BlockingType;
+
+    BlockingType blocking(size, size, depth, 1, false);
+
     internal::general_matrix_matrix_triangular_product<Index,
-      typename Lhs::Scalar, _ActualLhs::Flags&RowMajorBit ? RowMajor : ColMajor, LhsBlasTraits::NeedToConjugate,
-      typename Rhs::Scalar, _ActualRhs::Flags&RowMajorBit ? RowMajor : ColMajor, RhsBlasTraits::NeedToConjugate,
-      MatrixType::Flags&RowMajorBit ? RowMajor : ColMajor, UpLo>
-      ::run(mat.cols(), actualLhs.cols(),
+      typename Lhs::Scalar, LhsIsRowMajor ? RowMajor : ColMajor, LhsBlasTraits::NeedToConjugate,
+      typename Rhs::Scalar, RhsIsRowMajor ? RowMajor : ColMajor, RhsBlasTraits::NeedToConjugate,
+      IsRowMajor ? RowMajor : ColMajor, UpLo>
+      ::run(size, depth,
             &actualLhs.coeffRef(0,0), actualLhs.outerStride(), &actualRhs.coeffRef(0,0), actualRhs.outerStride(),
-            mat.data(), mat.outerStride(), actualAlpha);
+            mat.data(), mat.outerStride(), actualAlpha, blocking);
   }
 };
 
 template<typename MatrixType, unsigned int UpLo>
-template<typename ProductDerived, typename _Lhs, typename _Rhs>
-TriangularView<MatrixType,UpLo>& TriangularView<MatrixType,UpLo>::assignProduct(const ProductBase<ProductDerived, _Lhs,_Rhs>& prod, const Scalar& alpha)
+template<typename ProductType>
+EIGEN_DEVICE_FUNC TriangularView<MatrixType,UpLo>& TriangularViewImpl<MatrixType,UpLo,Dense>::_assignProduct(const ProductType& prod, const Scalar& alpha, bool beta)
 {
-  general_product_to_triangular_selector<MatrixType, ProductDerived, UpLo, (_Lhs::ColsAtCompileTime==1) || (_Rhs::RowsAtCompileTime==1)>::run(m_matrix.const_cast_derived(), prod.derived(), alpha);
-  
-  return *this;
+  eigen_assert(derived().nestedExpression().rows() == prod.rows() && derived().cols() == prod.cols());
+
+  general_product_to_triangular_selector<MatrixType, ProductType, UpLo, internal::traits<ProductType>::InnerSize==1>::run(derived().nestedExpression().const_cast_derived(), prod, alpha, beta);
+
+  return derived();
 }
 
 } // end namespace Eigen
diff --git a/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_MKL.h b/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h
index 3deed06..5b7c15c 100644
--- a/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_MKL.h
+++ b/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h
@@ -25,15 +25,15 @@
  SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
  ********************************************************************************
- *   Content : Eigen bindings to Intel(R) MKL
+ *   Content : Eigen bindings to BLAS F77
  *   Level 3 BLAS SYRK/HERK implementation.
  ********************************************************************************
 */
 
-#ifndef EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_MKL_H
-#define EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_MKL_H
+#ifndef EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_BLAS_H
+#define EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_BLAS_H
 
-namespace Eigen { 
+namespace Eigen {
 
 namespace internal {
 
@@ -44,34 +44,35 @@ struct general_matrix_matrix_rankupdate :
 
 
 // try to go to BLAS specialization
-#define EIGEN_MKL_RANKUPDATE_SPECIALIZE(Scalar) \
+#define EIGEN_BLAS_RANKUPDATE_SPECIALIZE(Scalar) \
 template <typename Index, int LhsStorageOrder, bool ConjugateLhs, \
                           int RhsStorageOrder, bool ConjugateRhs, int  UpLo> \
 struct general_matrix_matrix_triangular_product<Index,Scalar,LhsStorageOrder,ConjugateLhs, \
                Scalar,RhsStorageOrder,ConjugateRhs,ColMajor,UpLo,Specialized> { \
   static EIGEN_STRONG_INLINE void run(Index size, Index depth,const Scalar* lhs, Index lhsStride, \
-                          const Scalar* rhs, Index rhsStride, Scalar* res, Index resStride, Scalar alpha) \
+                          const Scalar* rhs, Index rhsStride, Scalar* res, Index resStride, Scalar alpha, level3_blocking<Scalar, Scalar>& blocking) \
   { \
     if (lhs==rhs) { \
       general_matrix_matrix_rankupdate<Index,Scalar,LhsStorageOrder,ConjugateLhs,ColMajor,UpLo> \
-      ::run(size,depth,lhs,lhsStride,rhs,rhsStride,res,resStride,alpha); \
+      ::run(size,depth,lhs,lhsStride,rhs,rhsStride,res,resStride,alpha,blocking); \
     } else { \
       general_matrix_matrix_triangular_product<Index, \
         Scalar, LhsStorageOrder, ConjugateLhs, \
         Scalar, RhsStorageOrder, ConjugateRhs, \
         ColMajor, UpLo, BuiltIn> \
-      ::run(size,depth,lhs,lhsStride,rhs,rhsStride,res,resStride,alpha); \
+      ::run(size,depth,lhs,lhsStride,rhs,rhsStride,res,resStride,alpha,blocking); \
     } \
   } \
 };
 
-EIGEN_MKL_RANKUPDATE_SPECIALIZE(double)
-//EIGEN_MKL_RANKUPDATE_SPECIALIZE(dcomplex)
-EIGEN_MKL_RANKUPDATE_SPECIALIZE(float)
-//EIGEN_MKL_RANKUPDATE_SPECIALIZE(scomplex)
+EIGEN_BLAS_RANKUPDATE_SPECIALIZE(double)
+EIGEN_BLAS_RANKUPDATE_SPECIALIZE(float)
+// TODO handle complex cases
+// EIGEN_BLAS_RANKUPDATE_SPECIALIZE(dcomplex)
+// EIGEN_BLAS_RANKUPDATE_SPECIALIZE(scomplex)
 
 // SYRK for float/double
-#define EIGEN_MKL_RANKUPDATE_R(EIGTYPE, MKLTYPE, MKLFUNC) \
+#define EIGEN_BLAS_RANKUPDATE_R(EIGTYPE, BLASTYPE, BLASFUNC) \
 template <typename Index, int AStorageOrder, bool ConjugateA, int  UpLo> \
 struct general_matrix_matrix_rankupdate<Index,EIGTYPE,AStorageOrder,ConjugateA,ColMajor,UpLo> { \
   enum { \
@@ -80,23 +81,19 @@ struct general_matrix_matrix_rankupdate<Index,EIGTYPE,AStorageOrder,ConjugateA,C
     conjA = ((AStorageOrder==ColMajor) && ConjugateA) ? 1 : 0 \
   }; \
   static EIGEN_STRONG_INLINE void run(Index size, Index depth,const EIGTYPE* lhs, Index lhsStride, \
-                          const EIGTYPE* rhs, Index rhsStride, EIGTYPE* res, Index resStride, EIGTYPE alpha) \
+                          const EIGTYPE* /*rhs*/, Index /*rhsStride*/, EIGTYPE* res, Index resStride, EIGTYPE alpha, level3_blocking<EIGTYPE, EIGTYPE>& /*blocking*/) \
   { \
   /* typedef Matrix<EIGTYPE, Dynamic, Dynamic, RhsStorageOrder> MatrixRhs;*/ \
 \
-   MKL_INT lda=lhsStride, ldc=resStride, n=size, k=depth; \
-   char uplo=(IsLower) ? 'L' : 'U', trans=(AStorageOrder==RowMajor) ? 'T':'N'; \
-   MKLTYPE alpha_, beta_; \
-\
-/* Set alpha_ & beta_ */ \
-   assign_scalar_eig2mkl<MKLTYPE, EIGTYPE>(alpha_, alpha); \
-   assign_scalar_eig2mkl<MKLTYPE, EIGTYPE>(beta_, EIGTYPE(1)); \
-   MKLFUNC(&uplo, &trans, &n, &k, &alpha_, lhs, &lda, &beta_, res, &ldc); \
+   BlasIndex lda=convert_index<BlasIndex>(lhsStride), ldc=convert_index<BlasIndex>(resStride), n=convert_index<BlasIndex>(size), k=convert_index<BlasIndex>(depth); \
+   char uplo=((IsLower) ? 'L' : 'U'), trans=((AStorageOrder==RowMajor) ? 'T':'N'); \
+   EIGTYPE beta(1); \
+   BLASFUNC(&uplo, &trans, &n, &k, &numext::real_ref(alpha), lhs, &lda, &numext::real_ref(beta), res, &ldc); \
   } \
 };
 
 // HERK for complex data
-#define EIGEN_MKL_RANKUPDATE_C(EIGTYPE, MKLTYPE, RTYPE, MKLFUNC) \
+#define EIGEN_BLAS_RANKUPDATE_C(EIGTYPE, BLASTYPE, RTYPE, BLASFUNC) \
 template <typename Index, int AStorageOrder, bool ConjugateA, int  UpLo> \
 struct general_matrix_matrix_rankupdate<Index,EIGTYPE,AStorageOrder,ConjugateA,ColMajor,UpLo> { \
   enum { \
@@ -105,18 +102,15 @@ struct general_matrix_matrix_rankupdate<Index,EIGTYPE,AStorageOrder,ConjugateA,C
     conjA = (((AStorageOrder==ColMajor) && ConjugateA) || ((AStorageOrder==RowMajor) && !ConjugateA)) ? 1 : 0 \
   }; \
   static EIGEN_STRONG_INLINE void run(Index size, Index depth,const EIGTYPE* lhs, Index lhsStride, \
-                          const EIGTYPE* rhs, Index rhsStride, EIGTYPE* res, Index resStride, EIGTYPE alpha) \
+                          const EIGTYPE* /*rhs*/, Index /*rhsStride*/, EIGTYPE* res, Index resStride, EIGTYPE alpha, level3_blocking<EIGTYPE, EIGTYPE>& /*blocking*/) \
   { \
    typedef Matrix<EIGTYPE, Dynamic, Dynamic, AStorageOrder> MatrixType; \
 \
-   MKL_INT lda=lhsStride, ldc=resStride, n=size, k=depth; \
-   char uplo=(IsLower) ? 'L' : 'U', trans=(AStorageOrder==RowMajor) ? 'C':'N'; \
+   BlasIndex lda=convert_index<BlasIndex>(lhsStride), ldc=convert_index<BlasIndex>(resStride), n=convert_index<BlasIndex>(size), k=convert_index<BlasIndex>(depth); \
+   char uplo=((IsLower) ? 'L' : 'U'), trans=((AStorageOrder==RowMajor) ? 'C':'N'); \
    RTYPE alpha_, beta_; \
    const EIGTYPE* a_ptr; \
 \
-/* Set alpha_ & beta_ */ \
-/*   assign_scalar_eig2mkl<MKLTYPE, EIGTYPE>(alpha_, alpha); */\
-/*   assign_scalar_eig2mkl<MKLTYPE, EIGTYPE>(beta_, EIGTYPE(1));*/ \
    alpha_ = alpha.real(); \
    beta_ = 1.0; \
 /* Copy with conjugation in some cases*/ \
@@ -127,20 +121,21 @@ struct general_matrix_matrix_rankupdate<Index,EIGTYPE,AStorageOrder,ConjugateA,C
      lda = a.outerStride(); \
      a_ptr = a.data(); \
    } else a_ptr=lhs; \
-   MKLFUNC(&uplo, &trans, &n, &k, &alpha_, (MKLTYPE*)a_ptr, &lda, &beta_, (MKLTYPE*)res, &ldc); \
+   BLASFUNC(&uplo, &trans, &n, &k, &alpha_, (BLASTYPE*)a_ptr, &lda, &beta_, (BLASTYPE*)res, &ldc); \
   } \
 };
 
 
-EIGEN_MKL_RANKUPDATE_R(double, double, dsyrk)
-EIGEN_MKL_RANKUPDATE_R(float,  float,  ssyrk)
+EIGEN_BLAS_RANKUPDATE_R(double, double, dsyrk_)
+EIGEN_BLAS_RANKUPDATE_R(float,  float,  ssyrk_)
 
-//EIGEN_MKL_RANKUPDATE_C(dcomplex, MKL_Complex16, double, zherk)
-//EIGEN_MKL_RANKUPDATE_C(scomplex, MKL_Complex8,  double, cherk)
+// TODO hanlde complex cases
+// EIGEN_BLAS_RANKUPDATE_C(dcomplex, double, double, zherk_)
+// EIGEN_BLAS_RANKUPDATE_C(scomplex, float,  float, cherk_)
 
 
 } // end namespace internal
 
 } // end namespace Eigen
 
-#endif // EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_MKL_H
+#endif // EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_BLAS_H
diff --git a/eigen/Eigen/src/Core/products/GeneralMatrixMatrix_MKL.h b/eigen/Eigen/src/Core/products/GeneralMatrixMatrix_BLAS.h
index 060af32..7a3bdbf 100644
--- a/eigen/Eigen/src/Core/products/GeneralMatrixMatrix_MKL.h
+++ b/eigen/Eigen/src/Core/products/GeneralMatrixMatrix_BLAS.h
@@ -25,13 +25,13 @@
  SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
  ********************************************************************************
- *   Content : Eigen bindings to Intel(R) MKL
+ *   Content : Eigen bindings to BLAS F77
  *   General matrix-matrix product functionality based on ?GEMM.
  ********************************************************************************
 */
 
-#ifndef EIGEN_GENERAL_MATRIX_MATRIX_MKL_H
-#define EIGEN_GENERAL_MATRIX_MATRIX_MKL_H
+#ifndef EIGEN_GENERAL_MATRIX_MATRIX_BLAS_H
+#define EIGEN_GENERAL_MATRIX_MATRIX_BLAS_H
 
 namespace Eigen { 
 
@@ -46,13 +46,15 @@ namespace internal {
 
 // gemm specialization
 
-#define GEMM_SPECIALIZATION(EIGTYPE, EIGPREFIX, MKLTYPE, MKLPREFIX) \
+#define GEMM_SPECIALIZATION(EIGTYPE, EIGPREFIX, BLASTYPE, BLASPREFIX) \
 template< \
   typename Index, \
   int LhsStorageOrder, bool ConjugateLhs, \
   int RhsStorageOrder, bool ConjugateRhs> \
 struct general_matrix_matrix_product<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,RhsStorageOrder,ConjugateRhs,ColMajor> \
 { \
+typedef gebp_traits<EIGTYPE,EIGTYPE> Traits; \
+\
 static void run(Index rows, Index cols, Index depth, \
   const EIGTYPE* _lhs, Index lhsStride, \
   const EIGTYPE* _rhs, Index rhsStride, \
@@ -64,55 +66,50 @@ static void run(Index rows, Index cols, Index depth, \
   using std::conj; \
 \
   char transa, transb; \
-  MKL_INT m, n, k, lda, ldb, ldc; \
+  BlasIndex m, n, k, lda, ldb, ldc; \
   const EIGTYPE *a, *b; \
-  MKLTYPE alpha_, beta_; \
+  EIGTYPE beta(1); \
   MatrixX##EIGPREFIX a_tmp, b_tmp; \
-  EIGTYPE myone(1);\
 \
 /* Set transpose options */ \
   transa = (LhsStorageOrder==RowMajor) ? ((ConjugateLhs) ? 'C' : 'T') : 'N'; \
   transb = (RhsStorageOrder==RowMajor) ? ((ConjugateRhs) ? 'C' : 'T') : 'N'; \
 \
 /* Set m, n, k */ \
-  m = (MKL_INT)rows;  \
-  n = (MKL_INT)cols;  \
-  k = (MKL_INT)depth; \
-\
-/* Set alpha_ & beta_ */ \
-  assign_scalar_eig2mkl(alpha_, alpha); \
-  assign_scalar_eig2mkl(beta_, myone); \
+  m = convert_index<BlasIndex>(rows);  \
+  n = convert_index<BlasIndex>(cols);  \
+  k = convert_index<BlasIndex>(depth); \
 \
 /* Set lda, ldb, ldc */ \
-  lda = (MKL_INT)lhsStride; \
-  ldb = (MKL_INT)rhsStride; \
-  ldc = (MKL_INT)resStride; \
+  lda = convert_index<BlasIndex>(lhsStride); \
+  ldb = convert_index<BlasIndex>(rhsStride); \
+  ldc = convert_index<BlasIndex>(resStride); \
 \
 /* Set a, b, c */ \
   if ((LhsStorageOrder==ColMajor) && (ConjugateLhs)) { \
     Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > lhs(_lhs,m,k,OuterStride<>(lhsStride)); \
     a_tmp = lhs.conjugate(); \
     a = a_tmp.data(); \
-    lda = a_tmp.outerStride(); \
+    lda = convert_index<BlasIndex>(a_tmp.outerStride()); \
   } else a = _lhs; \
 \
   if ((RhsStorageOrder==ColMajor) && (ConjugateRhs)) { \
     Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > rhs(_rhs,k,n,OuterStride<>(rhsStride)); \
     b_tmp = rhs.conjugate(); \
     b = b_tmp.data(); \
-    ldb = b_tmp.outerStride(); \
+    ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \
   } else b = _rhs; \
 \
-  MKLPREFIX##gemm(&transa, &transb, &m, &n, &k, &alpha_, (const MKLTYPE*)a, &lda, (const MKLTYPE*)b, &ldb, &beta_, (MKLTYPE*)res, &ldc); \
+  BLASPREFIX##gemm_(&transa, &transb, &m, &n, &k, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, &numext::real_ref(beta), (BLASTYPE*)res, &ldc); \
 }};
 
-GEMM_SPECIALIZATION(double,   d,  double,        d)
-GEMM_SPECIALIZATION(float,    f,  float,         s)
-GEMM_SPECIALIZATION(dcomplex, cd, MKL_Complex16, z)
-GEMM_SPECIALIZATION(scomplex, cf, MKL_Complex8,  c)
+GEMM_SPECIALIZATION(double,   d,  double, d)
+GEMM_SPECIALIZATION(float,    f,  float,  s)
+GEMM_SPECIALIZATION(dcomplex, cd, double, z)
+GEMM_SPECIALIZATION(scomplex, cf, float,  c)
 
 } // end namespase internal
 
 } // end namespace Eigen
 
-#endif // EIGEN_GENERAL_MATRIX_MATRIX_MKL_H
+#endif // EIGEN_GENERAL_MATRIX_MATRIX_BLAS_H
diff --git a/eigen/Eigen/src/Core/products/GeneralMatrixVector.h b/eigen/Eigen/src/Core/products/GeneralMatrixVector.h
index 0938770..41d8242 100644
--- a/eigen/Eigen/src/Core/products/GeneralMatrixVector.h
+++ b/eigen/Eigen/src/Core/products/GeneralMatrixVector.h
@@ -1,7 +1,7 @@
 // This file is part of Eigen, a lightweight C++ template library
 // for linear algebra.
 //
-// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2008-2016 Gael Guennebaud <gael.guennebaud@inria.fr>
 //
 // This Source Code Form is subject to the terms of the Mozilla
 // Public License v. 2.0. If a copy of the MPL was not distributed
@@ -10,15 +10,13 @@
 #ifndef EIGEN_GENERAL_MATRIX_VECTOR_H
 #define EIGEN_GENERAL_MATRIX_VECTOR_H
 
-namespace Eigen { 
+namespace Eigen {
 
 namespace internal {
 
 /* Optimized col-major matrix * vector product:
- * This algorithm processes 4 columns at onces that allows to both reduce
- * the number of load/stores of the result by a factor 4 and to reduce
- * the instruction dependency. Moreover, we know that all bands have the
- * same alignment pattern.
+ * This algorithm processes the matrix per vertical panels,
+ * which are then processed horizontaly per chunck of 8*PacketSize x 1 vertical segments.
  *
  * Mixing type logic: C += alpha * A * B
  *  |  A  |  B  |alpha| comments
@@ -26,11 +24,13 @@ namespace internal {
  *  |real |cplx |real | alpha is converted to a cplx when calling the run function, no vectorization
  *  |cplx |real |cplx | invalid, the caller has to do tmp: = A * B; C += alpha*tmp
  *  |cplx |real |real | optimal case, vectorization possible via real-cplx mul
+ *
+ * The same reasoning apply for the transposed case.
  */
-template<typename Index, typename LhsScalar, bool ConjugateLhs, typename RhsScalar, bool ConjugateRhs, int Version>
-struct general_matrix_vector_product<Index,LhsScalar,ColMajor,ConjugateLhs,RhsScalar,ConjugateRhs,Version>
+template<typename Index, typename LhsScalar, typename LhsMapper, bool ConjugateLhs, typename RhsScalar, typename RhsMapper, bool ConjugateRhs, int Version>
+struct general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,ConjugateLhs,RhsScalar,RhsMapper,ConjugateRhs,Version>
 {
-typedef typename scalar_product_traits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+  typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
 
 enum {
   Vectorizable = packet_traits<LhsScalar>::Vectorizable && packet_traits<RhsScalar>::Vectorizable
@@ -50,234 +50,154 @@ typedef typename conditional<Vectorizable,_ResPacket,ResScalar>::type ResPacket;
 
 EIGEN_DONT_INLINE static void run(
   Index rows, Index cols,
-  const LhsScalar* lhs, Index lhsStride,
-  const RhsScalar* rhs, Index rhsIncr,
-  ResScalar* res, Index resIncr, RhsScalar alpha);
+  const LhsMapper& lhs,
+  const RhsMapper& rhs,
+        ResScalar* res, Index resIncr,
+  RhsScalar alpha);
 };
 
-template<typename Index, typename LhsScalar, bool ConjugateLhs, typename RhsScalar, bool ConjugateRhs, int Version>
-EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,ColMajor,ConjugateLhs,RhsScalar,ConjugateRhs,Version>::run(
+template<typename Index, typename LhsScalar, typename LhsMapper, bool ConjugateLhs, typename RhsScalar, typename RhsMapper, bool ConjugateRhs, int Version>
+EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,ConjugateLhs,RhsScalar,RhsMapper,ConjugateRhs,Version>::run(
   Index rows, Index cols,
-  const LhsScalar* lhs, Index lhsStride,
-  const RhsScalar* rhs, Index rhsIncr,
-  ResScalar* res, Index resIncr, RhsScalar alpha)
+  const LhsMapper& alhs,
+  const RhsMapper& rhs,
+        ResScalar* res, Index resIncr,
+  RhsScalar alpha)
 {
-  EIGEN_UNUSED_VARIABLE(resIncr)
+  EIGEN_UNUSED_VARIABLE(resIncr);
   eigen_internal_assert(resIncr==1);
-  #ifdef _EIGEN_ACCUMULATE_PACKETS
-  #error _EIGEN_ACCUMULATE_PACKETS has already been defined
-  #endif
-  #define _EIGEN_ACCUMULATE_PACKETS(A0,A13,A2) \
-    pstore(&res[j], \
-      padd(pload<ResPacket>(&res[j]), \
-        padd( \
-          padd(pcj.pmul(EIGEN_CAT(ploa , A0)<LhsPacket>(&lhs0[j]),    ptmp0), \
-                  pcj.pmul(EIGEN_CAT(ploa , A13)<LhsPacket>(&lhs1[j]),   ptmp1)), \
-          padd(pcj.pmul(EIGEN_CAT(ploa , A2)<LhsPacket>(&lhs2[j]),    ptmp2), \
-                  pcj.pmul(EIGEN_CAT(ploa , A13)<LhsPacket>(&lhs3[j]),   ptmp3)) )))
+
+  // The following copy tells the compiler that lhs's attributes are not modified outside this function
+  // This helps GCC to generate propoer code.
+  LhsMapper lhs(alhs);
 
   conj_helper<LhsScalar,RhsScalar,ConjugateLhs,ConjugateRhs> cj;
   conj_helper<LhsPacket,RhsPacket,ConjugateLhs,ConjugateRhs> pcj;
-  if(ConjugateRhs)
-    alpha = numext::conj(alpha);
-
-  enum { AllAligned = 0, EvenAligned, FirstAligned, NoneAligned };
-  const Index columnsAtOnce = 4;
-  const Index peels = 2;
-  const Index LhsPacketAlignedMask = LhsPacketSize-1;
-  const Index ResPacketAlignedMask = ResPacketSize-1;
-//  const Index PeelAlignedMask = ResPacketSize*peels-1;
-  const Index size = rows;
-  
-  // How many coeffs of the result do we have to skip to be aligned.
-  // Here we assume data are at least aligned on the base scalar type.
-  Index alignedStart = internal::first_aligned(res,size);
-  Index alignedSize = ResPacketSize>1 ? alignedStart + ((size-alignedStart) & ~ResPacketAlignedMask) : 0;
-  const Index peeledSize = alignedSize - RhsPacketSize*peels - RhsPacketSize + 1;
-
-  const Index alignmentStep = LhsPacketSize>1 ? (LhsPacketSize - lhsStride % LhsPacketSize) & LhsPacketAlignedMask : 0;
-  Index alignmentPattern = alignmentStep==0 ? AllAligned
-                       : alignmentStep==(LhsPacketSize/2) ? EvenAligned
-                       : FirstAligned;
-
-  // we cannot assume the first element is aligned because of sub-matrices
-  const Index lhsAlignmentOffset = internal::first_aligned(lhs,size);
-
-  // find how many columns do we have to skip to be aligned with the result (if possible)
-  Index skipColumns = 0;
-  // if the data cannot be aligned (TODO add some compile time tests when possible, e.g. for floats)
-  if( (size_t(lhs)%sizeof(LhsScalar)) || (size_t(res)%sizeof(ResScalar)) )
-  {
-    alignedSize = 0;
-    alignedStart = 0;
-  }
-  else if (LhsPacketSize>1)
-  {
-    eigen_internal_assert(size_t(lhs+lhsAlignmentOffset)%sizeof(LhsPacket)==0 || size<LhsPacketSize);
+  const Index lhsStride = lhs.stride();
+  // TODO: for padded aligned inputs, we could enable aligned reads
+  enum { LhsAlignment = Unaligned };
 
-    while (skipColumns<LhsPacketSize &&
-          alignedStart != ((lhsAlignmentOffset + alignmentStep*skipColumns)%LhsPacketSize))
-      ++skipColumns;
-    if (skipColumns==LhsPacketSize)
-    {
-      // nothing can be aligned, no need to skip any column
-      alignmentPattern = NoneAligned;
-      skipColumns = 0;
-    }
-    else
-    {
-      skipColumns = (std::min)(skipColumns,cols);
-      // note that the skiped columns are processed later.
-    }
-
-    eigen_internal_assert(  (alignmentPattern==NoneAligned)
-                      || (skipColumns + columnsAtOnce >= cols)
-                      || LhsPacketSize > size
-                      || (size_t(lhs+alignedStart+lhsStride*skipColumns)%sizeof(LhsPacket))==0);
-  }
-  else if(Vectorizable)
-  {
-    alignedStart = 0;
-    alignedSize = size;
-    alignmentPattern = AllAligned;
-  }
+  const Index n8 = rows-8*ResPacketSize+1;
+  const Index n4 = rows-4*ResPacketSize+1;
+  const Index n3 = rows-3*ResPacketSize+1;
+  const Index n2 = rows-2*ResPacketSize+1;
+  const Index n1 = rows-1*ResPacketSize+1;
 
-  Index offset1 = (FirstAligned && alignmentStep==1?3:1);
-  Index offset3 = (FirstAligned && alignmentStep==1?1:3);
+  // TODO: improve the following heuristic:
+  const Index block_cols = cols<128 ? cols : (lhsStride*sizeof(LhsScalar)<32000?16:4);
+  ResPacket palpha = pset1<ResPacket>(alpha);
 
-  Index columnBound = ((cols-skipColumns)/columnsAtOnce)*columnsAtOnce + skipColumns;
-  for (Index i=skipColumns; i<columnBound; i+=columnsAtOnce)
+  for(Index j2=0; j2<cols; j2+=block_cols)
   {
-    RhsPacket ptmp0 = pset1<RhsPacket>(alpha*rhs[i*rhsIncr]),
-              ptmp1 = pset1<RhsPacket>(alpha*rhs[(i+offset1)*rhsIncr]),
-              ptmp2 = pset1<RhsPacket>(alpha*rhs[(i+2)*rhsIncr]),
-              ptmp3 = pset1<RhsPacket>(alpha*rhs[(i+offset3)*rhsIncr]);
-
-    // this helps a lot generating better binary code
-    const LhsScalar *lhs0 = lhs + i*lhsStride,     *lhs1 = lhs + (i+offset1)*lhsStride,
-                    *lhs2 = lhs + (i+2)*lhsStride, *lhs3 = lhs + (i+offset3)*lhsStride;
-
-    if (Vectorizable)
+    Index jend = numext::mini(j2+block_cols,cols);
+    Index i=0;
+    for(; i<n8; i+=ResPacketSize*8)
     {
-      /* explicit vectorization */
-      // process initial unaligned coeffs
-      for (Index j=0; j<alignedStart; ++j)
+      ResPacket c0 = pset1<ResPacket>(ResScalar(0)),
+                c1 = pset1<ResPacket>(ResScalar(0)),
+                c2 = pset1<ResPacket>(ResScalar(0)),
+                c3 = pset1<ResPacket>(ResScalar(0)),
+                c4 = pset1<ResPacket>(ResScalar(0)),
+                c5 = pset1<ResPacket>(ResScalar(0)),
+                c6 = pset1<ResPacket>(ResScalar(0)),
+                c7 = pset1<ResPacket>(ResScalar(0));
+
+      for(Index j=j2; j<jend; j+=1)
       {
-        res[j] = cj.pmadd(lhs0[j], pfirst(ptmp0), res[j]);
-        res[j] = cj.pmadd(lhs1[j], pfirst(ptmp1), res[j]);
-        res[j] = cj.pmadd(lhs2[j], pfirst(ptmp2), res[j]);
-        res[j] = cj.pmadd(lhs3[j], pfirst(ptmp3), res[j]);
+        RhsPacket b0 = pset1<RhsPacket>(rhs(j,0));
+        c0 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+LhsPacketSize*0,j),b0,c0);
+        c1 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+LhsPacketSize*1,j),b0,c1);
+        c2 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+LhsPacketSize*2,j),b0,c2);
+        c3 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+LhsPacketSize*3,j),b0,c3);
+        c4 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+LhsPacketSize*4,j),b0,c4);
+        c5 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+LhsPacketSize*5,j),b0,c5);
+        c6 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+LhsPacketSize*6,j),b0,c6);
+        c7 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+LhsPacketSize*7,j),b0,c7);
       }
+      pstoreu(res+i+ResPacketSize*0, pmadd(c0,palpha,ploadu<ResPacket>(res+i+ResPacketSize*0)));
+      pstoreu(res+i+ResPacketSize*1, pmadd(c1,palpha,ploadu<ResPacket>(res+i+ResPacketSize*1)));
+      pstoreu(res+i+ResPacketSize*2, pmadd(c2,palpha,ploadu<ResPacket>(res+i+ResPacketSize*2)));
+      pstoreu(res+i+ResPacketSize*3, pmadd(c3,palpha,ploadu<ResPacket>(res+i+ResPacketSize*3)));
+      pstoreu(res+i+ResPacketSize*4, pmadd(c4,palpha,ploadu<ResPacket>(res+i+ResPacketSize*4)));
+      pstoreu(res+i+ResPacketSize*5, pmadd(c5,palpha,ploadu<ResPacket>(res+i+ResPacketSize*5)));
+      pstoreu(res+i+ResPacketSize*6, pmadd(c6,palpha,ploadu<ResPacket>(res+i+ResPacketSize*6)));
+      pstoreu(res+i+ResPacketSize*7, pmadd(c7,palpha,ploadu<ResPacket>(res+i+ResPacketSize*7)));
+    }
+    if(i<n4)
+    {
+      ResPacket c0 = pset1<ResPacket>(ResScalar(0)),
+                c1 = pset1<ResPacket>(ResScalar(0)),
+                c2 = pset1<ResPacket>(ResScalar(0)),
+                c3 = pset1<ResPacket>(ResScalar(0));
 
-      if (alignedSize>alignedStart)
+      for(Index j=j2; j<jend; j+=1)
       {
-        switch(alignmentPattern)
-        {
-          case AllAligned:
-            for (Index j = alignedStart; j<alignedSize; j+=ResPacketSize)
-              _EIGEN_ACCUMULATE_PACKETS(d,d,d);
-            break;
-          case EvenAligned:
-            for (Index j = alignedStart; j<alignedSize; j+=ResPacketSize)
-              _EIGEN_ACCUMULATE_PACKETS(d,du,d);
-            break;
-          case FirstAligned:
-          {
-            Index j = alignedStart;
-            if(peels>1)
-            {
-              LhsPacket A00, A01, A02, A03, A10, A11, A12, A13;
-              ResPacket T0, T1;
-
-              A01 = pload<LhsPacket>(&lhs1[alignedStart-1]);
-              A02 = pload<LhsPacket>(&lhs2[alignedStart-2]);
-              A03 = pload<LhsPacket>(&lhs3[alignedStart-3]);
-
-              for (; j<peeledSize; j+=peels*ResPacketSize)
-              {
-                A11 = pload<LhsPacket>(&lhs1[j-1+LhsPacketSize]);  palign<1>(A01,A11);
-                A12 = pload<LhsPacket>(&lhs2[j-2+LhsPacketSize]);  palign<2>(A02,A12);
-                A13 = pload<LhsPacket>(&lhs3[j-3+LhsPacketSize]);  palign<3>(A03,A13);
-
-                A00 = pload<LhsPacket>(&lhs0[j]);
-                A10 = pload<LhsPacket>(&lhs0[j+LhsPacketSize]);
-                T0  = pcj.pmadd(A00, ptmp0, pload<ResPacket>(&res[j]));
-                T1  = pcj.pmadd(A10, ptmp0, pload<ResPacket>(&res[j+ResPacketSize]));
-
-                T0  = pcj.pmadd(A01, ptmp1, T0);
-                A01 = pload<LhsPacket>(&lhs1[j-1+2*LhsPacketSize]);  palign<1>(A11,A01);
-                T0  = pcj.pmadd(A02, ptmp2, T0);
-                A02 = pload<LhsPacket>(&lhs2[j-2+2*LhsPacketSize]);  palign<2>(A12,A02);
-                T0  = pcj.pmadd(A03, ptmp3, T0);
-                pstore(&res[j],T0);
-                A03 = pload<LhsPacket>(&lhs3[j-3+2*LhsPacketSize]);  palign<3>(A13,A03);
-                T1  = pcj.pmadd(A11, ptmp1, T1);
-                T1  = pcj.pmadd(A12, ptmp2, T1);
-                T1  = pcj.pmadd(A13, ptmp3, T1);
-                pstore(&res[j+ResPacketSize],T1);
-              }
-            }
-            for (; j<alignedSize; j+=ResPacketSize)
-              _EIGEN_ACCUMULATE_PACKETS(d,du,du);
-            break;
-          }
-          default:
-            for (Index j = alignedStart; j<alignedSize; j+=ResPacketSize)
-              _EIGEN_ACCUMULATE_PACKETS(du,du,du);
-            break;
-        }
+        RhsPacket b0 = pset1<RhsPacket>(rhs(j,0));
+        c0 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+LhsPacketSize*0,j),b0,c0);
+        c1 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+LhsPacketSize*1,j),b0,c1);
+        c2 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+LhsPacketSize*2,j),b0,c2);
+        c3 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+LhsPacketSize*3,j),b0,c3);
       }
-    } // end explicit vectorization
+      pstoreu(res+i+ResPacketSize*0, pmadd(c0,palpha,ploadu<ResPacket>(res+i+ResPacketSize*0)));
+      pstoreu(res+i+ResPacketSize*1, pmadd(c1,palpha,ploadu<ResPacket>(res+i+ResPacketSize*1)));
+      pstoreu(res+i+ResPacketSize*2, pmadd(c2,palpha,ploadu<ResPacket>(res+i+ResPacketSize*2)));
+      pstoreu(res+i+ResPacketSize*3, pmadd(c3,palpha,ploadu<ResPacket>(res+i+ResPacketSize*3)));
 
-    /* process remaining coeffs (or all if there is no explicit vectorization) */
-    for (Index j=alignedSize; j<size; ++j)
-    {
-      res[j] = cj.pmadd(lhs0[j], pfirst(ptmp0), res[j]);
-      res[j] = cj.pmadd(lhs1[j], pfirst(ptmp1), res[j]);
-      res[j] = cj.pmadd(lhs2[j], pfirst(ptmp2), res[j]);
-      res[j] = cj.pmadd(lhs3[j], pfirst(ptmp3), res[j]);
+      i+=ResPacketSize*4;
     }
-  }
-
-  // process remaining first and last columns (at most columnsAtOnce-1)
-  Index end = cols;
-  Index start = columnBound;
-  do
-  {
-    for (Index k=start; k<end; ++k)
+    if(i<n3)
     {
-      RhsPacket ptmp0 = pset1<RhsPacket>(alpha*rhs[k*rhsIncr]);
-      const LhsScalar* lhs0 = lhs + k*lhsStride;
+      ResPacket c0 = pset1<ResPacket>(ResScalar(0)),
+                c1 = pset1<ResPacket>(ResScalar(0)),
+                c2 = pset1<ResPacket>(ResScalar(0));
 
-      if (Vectorizable)
+      for(Index j=j2; j<jend; j+=1)
       {
-        /* explicit vectorization */
-        // process first unaligned result's coeffs
-        for (Index j=0; j<alignedStart; ++j)
-          res[j] += cj.pmul(lhs0[j], pfirst(ptmp0));
-        // process aligned result's coeffs
-        if ((size_t(lhs0+alignedStart)%sizeof(LhsPacket))==0)
-          for (Index i = alignedStart;i<alignedSize;i+=ResPacketSize)
-            pstore(&res[i], pcj.pmadd(pload<LhsPacket>(&lhs0[i]), ptmp0, pload<ResPacket>(&res[i])));
-        else
-          for (Index i = alignedStart;i<alignedSize;i+=ResPacketSize)
-            pstore(&res[i], pcj.pmadd(ploadu<LhsPacket>(&lhs0[i]), ptmp0, pload<ResPacket>(&res[i])));
+        RhsPacket b0 = pset1<RhsPacket>(rhs(j,0));
+        c0 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+LhsPacketSize*0,j),b0,c0);
+        c1 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+LhsPacketSize*1,j),b0,c1);
+        c2 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+LhsPacketSize*2,j),b0,c2);
       }
+      pstoreu(res+i+ResPacketSize*0, pmadd(c0,palpha,ploadu<ResPacket>(res+i+ResPacketSize*0)));
+      pstoreu(res+i+ResPacketSize*1, pmadd(c1,palpha,ploadu<ResPacket>(res+i+ResPacketSize*1)));
+      pstoreu(res+i+ResPacketSize*2, pmadd(c2,palpha,ploadu<ResPacket>(res+i+ResPacketSize*2)));
 
-      // process remaining scalars (or all if no explicit vectorization)
-      for (Index i=alignedSize; i<size; ++i)
-        res[i] += cj.pmul(lhs0[i], pfirst(ptmp0));
+      i+=ResPacketSize*3;
     }
-    if (skipColumns)
+    if(i<n2)
     {
-      start = 0;
-      end = skipColumns;
-      skipColumns = 0;
+      ResPacket c0 = pset1<ResPacket>(ResScalar(0)),
+                c1 = pset1<ResPacket>(ResScalar(0));
+
+      for(Index j=j2; j<jend; j+=1)
+      {
+        RhsPacket b0 = pset1<RhsPacket>(rhs(j,0));
+        c0 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+LhsPacketSize*0,j),b0,c0);
+        c1 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+LhsPacketSize*1,j),b0,c1);
+      }
+      pstoreu(res+i+ResPacketSize*0, pmadd(c0,palpha,ploadu<ResPacket>(res+i+ResPacketSize*0)));
+      pstoreu(res+i+ResPacketSize*1, pmadd(c1,palpha,ploadu<ResPacket>(res+i+ResPacketSize*1)));
+      i+=ResPacketSize*2;
+    }
+    if(i<n1)
+    {
+      ResPacket c0 = pset1<ResPacket>(ResScalar(0));
+      for(Index j=j2; j<jend; j+=1)
+      {
+        RhsPacket b0 = pset1<RhsPacket>(rhs(j,0));
+        c0 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+0,j),b0,c0);
+      }
+      pstoreu(res+i+ResPacketSize*0, pmadd(c0,palpha,ploadu<ResPacket>(res+i+ResPacketSize*0)));
+      i+=ResPacketSize;
     }
-    else
-      break;
-  } while(Vectorizable);
-  #undef _EIGEN_ACCUMULATE_PACKETS
+    for(;i<rows;++i)
+    {
+      ResScalar c0(0);
+      for(Index j=j2; j<jend; j+=1)
+        c0 += cj.pmul(lhs(i,j), rhs(j,0));
+      res[i] += alpha*c0;
+    }
+  }
 }
 
 /* Optimized row-major matrix * vector product:
@@ -290,10 +210,10 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,ColMajor,Co
  *  - alpha is always a complex (or converted to a complex)
  *  - no vectorization
  */
-template<typename Index, typename LhsScalar, bool ConjugateLhs, typename RhsScalar, bool ConjugateRhs, int Version>
-struct general_matrix_vector_product<Index,LhsScalar,RowMajor,ConjugateLhs,RhsScalar,ConjugateRhs,Version>
+template<typename Index, typename LhsScalar, typename LhsMapper, bool ConjugateLhs, typename RhsScalar, typename RhsMapper, bool ConjugateRhs, int Version>
+struct general_matrix_vector_product<Index,LhsScalar,LhsMapper,RowMajor,ConjugateLhs,RhsScalar,RhsMapper,ConjugateRhs,Version>
 {
-typedef typename scalar_product_traits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
 
 enum {
   Vectorizable = packet_traits<LhsScalar>::Vectorizable && packet_traits<RhsScalar>::Vectorizable
@@ -310,253 +230,172 @@ typedef typename packet_traits<ResScalar>::type  _ResPacket;
 typedef typename conditional<Vectorizable,_LhsPacket,LhsScalar>::type LhsPacket;
 typedef typename conditional<Vectorizable,_RhsPacket,RhsScalar>::type RhsPacket;
 typedef typename conditional<Vectorizable,_ResPacket,ResScalar>::type ResPacket;
-  
+
 EIGEN_DONT_INLINE static void run(
   Index rows, Index cols,
-  const LhsScalar* lhs, Index lhsStride,
-  const RhsScalar* rhs, Index rhsIncr,
-  ResScalar* res, Index resIncr,
+  const LhsMapper& lhs,
+  const RhsMapper& rhs,
+        ResScalar* res, Index resIncr,
   ResScalar alpha);
 };
 
-template<typename Index, typename LhsScalar, bool ConjugateLhs, typename RhsScalar, bool ConjugateRhs, int Version>
-EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,RowMajor,ConjugateLhs,RhsScalar,ConjugateRhs,Version>::run(
+template<typename Index, typename LhsScalar, typename LhsMapper, bool ConjugateLhs, typename RhsScalar, typename RhsMapper, bool ConjugateRhs, int Version>
+EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,LhsMapper,RowMajor,ConjugateLhs,RhsScalar,RhsMapper,ConjugateRhs,Version>::run(
   Index rows, Index cols,
-  const LhsScalar* lhs, Index lhsStride,
-  const RhsScalar* rhs, Index rhsIncr,
+  const LhsMapper& alhs,
+  const RhsMapper& rhs,
   ResScalar* res, Index resIncr,
   ResScalar alpha)
 {
-  EIGEN_UNUSED_VARIABLE(rhsIncr);
-  eigen_internal_assert(rhsIncr==1);
-  #ifdef _EIGEN_ACCUMULATE_PACKETS
-  #error _EIGEN_ACCUMULATE_PACKETS has already been defined
-  #endif
-
-  #define _EIGEN_ACCUMULATE_PACKETS(A0,A13,A2) {\
-    RhsPacket b = pload<RhsPacket>(&rhs[j]); \
-    ptmp0 = pcj.pmadd(EIGEN_CAT(ploa,A0) <LhsPacket>(&lhs0[j]), b, ptmp0); \
-    ptmp1 = pcj.pmadd(EIGEN_CAT(ploa,A13)<LhsPacket>(&lhs1[j]), b, ptmp1); \
-    ptmp2 = pcj.pmadd(EIGEN_CAT(ploa,A2) <LhsPacket>(&lhs2[j]), b, ptmp2); \
-    ptmp3 = pcj.pmadd(EIGEN_CAT(ploa,A13)<LhsPacket>(&lhs3[j]), b, ptmp3); }
+  // The following copy tells the compiler that lhs's attributes are not modified outside this function
+  // This helps GCC to generate propoer code.
+  LhsMapper lhs(alhs);
 
+  eigen_internal_assert(rhs.stride()==1);
   conj_helper<LhsScalar,RhsScalar,ConjugateLhs,ConjugateRhs> cj;
   conj_helper<LhsPacket,RhsPacket,ConjugateLhs,ConjugateRhs> pcj;
 
-  enum { AllAligned=0, EvenAligned=1, FirstAligned=2, NoneAligned=3 };
-  const Index rowsAtOnce = 4;
-  const Index peels = 2;
-  const Index RhsPacketAlignedMask = RhsPacketSize-1;
-  const Index LhsPacketAlignedMask = LhsPacketSize-1;
-//   const Index PeelAlignedMask = RhsPacketSize*peels-1;
-  const Index depth = cols;
-
-  // How many coeffs of the result do we have to skip to be aligned.
-  // Here we assume data are at least aligned on the base scalar type
-  // if that's not the case then vectorization is discarded, see below.
-  Index alignedStart = internal::first_aligned(rhs, depth);
-  Index alignedSize = RhsPacketSize>1 ? alignedStart + ((depth-alignedStart) & ~RhsPacketAlignedMask) : 0;
-  const Index peeledSize = alignedSize - RhsPacketSize*peels - RhsPacketSize + 1;
-
-  const Index alignmentStep = LhsPacketSize>1 ? (LhsPacketSize - lhsStride % LhsPacketSize) & LhsPacketAlignedMask : 0;
-  Index alignmentPattern = alignmentStep==0 ? AllAligned
-                         : alignmentStep==(LhsPacketSize/2) ? EvenAligned
-                         : FirstAligned;
-
-  // we cannot assume the first element is aligned because of sub-matrices
-  const Index lhsAlignmentOffset = internal::first_aligned(lhs,depth);
-
-  // find how many rows do we have to skip to be aligned with rhs (if possible)
-  Index skipRows = 0;
-  // if the data cannot be aligned (TODO add some compile time tests when possible, e.g. for floats)
-  if( (sizeof(LhsScalar)!=sizeof(RhsScalar)) || (size_t(lhs)%sizeof(LhsScalar)) || (size_t(rhs)%sizeof(RhsScalar)) )
-  {
-    alignedSize = 0;
-    alignedStart = 0;
-  }
-  else if (LhsPacketSize>1)
-  {
-    eigen_internal_assert(size_t(lhs+lhsAlignmentOffset)%sizeof(LhsPacket)==0  || depth<LhsPacketSize);
+  // TODO: fine tune the following heuristic. The rationale is that if the matrix is very large,
+  //       processing 8 rows at once might be counter productive wrt cache.
+  const Index n8 = lhs.stride()*sizeof(LhsScalar)>32000 ? 0 : rows-7;
+  const Index n4 = rows-3;
+  const Index n2 = rows-1;
+
+  // TODO: for padded aligned inputs, we could enable aligned reads
+  enum { LhsAlignment = Unaligned };
 
-    while (skipRows<LhsPacketSize &&
-           alignedStart != ((lhsAlignmentOffset + alignmentStep*skipRows)%LhsPacketSize))
-      ++skipRows;
-    if (skipRows==LhsPacketSize)
+  Index i=0;
+  for(; i<n8; i+=8)
+  {
+    ResPacket c0 = pset1<ResPacket>(ResScalar(0)),
+              c1 = pset1<ResPacket>(ResScalar(0)),
+              c2 = pset1<ResPacket>(ResScalar(0)),
+              c3 = pset1<ResPacket>(ResScalar(0)),
+              c4 = pset1<ResPacket>(ResScalar(0)),
+              c5 = pset1<ResPacket>(ResScalar(0)),
+              c6 = pset1<ResPacket>(ResScalar(0)),
+              c7 = pset1<ResPacket>(ResScalar(0));
+
+    Index j=0;
+    for(; j+LhsPacketSize<=cols; j+=LhsPacketSize)
     {
-      // nothing can be aligned, no need to skip any column
-      alignmentPattern = NoneAligned;
-      skipRows = 0;
+      RhsPacket b0 = rhs.template load<RhsPacket, Unaligned>(j,0);
+
+      c0 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+0,j),b0,c0);
+      c1 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+1,j),b0,c1);
+      c2 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+2,j),b0,c2);
+      c3 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+3,j),b0,c3);
+      c4 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+4,j),b0,c4);
+      c5 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+5,j),b0,c5);
+      c6 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+6,j),b0,c6);
+      c7 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+7,j),b0,c7);
     }
-    else
+    ResScalar cc0 = predux(c0);
+    ResScalar cc1 = predux(c1);
+    ResScalar cc2 = predux(c2);
+    ResScalar cc3 = predux(c3);
+    ResScalar cc4 = predux(c4);
+    ResScalar cc5 = predux(c5);
+    ResScalar cc6 = predux(c6);
+    ResScalar cc7 = predux(c7);
+    for(; j<cols; ++j)
     {
-      skipRows = (std::min)(skipRows,Index(rows));
-      // note that the skiped columns are processed later.
+      RhsScalar b0 = rhs(j,0);
+
+      cc0 += cj.pmul(lhs(i+0,j), b0);
+      cc1 += cj.pmul(lhs(i+1,j), b0);
+      cc2 += cj.pmul(lhs(i+2,j), b0);
+      cc3 += cj.pmul(lhs(i+3,j), b0);
+      cc4 += cj.pmul(lhs(i+4,j), b0);
+      cc5 += cj.pmul(lhs(i+5,j), b0);
+      cc6 += cj.pmul(lhs(i+6,j), b0);
+      cc7 += cj.pmul(lhs(i+7,j), b0);
     }
-    eigen_internal_assert(  alignmentPattern==NoneAligned
-                      || LhsPacketSize==1
-                      || (skipRows + rowsAtOnce >= rows)
-                      || LhsPacketSize > depth
-                      || (size_t(lhs+alignedStart+lhsStride*skipRows)%sizeof(LhsPacket))==0);
-  }
-  else if(Vectorizable)
-  {
-    alignedStart = 0;
-    alignedSize = depth;
-    alignmentPattern = AllAligned;
+    res[(i+0)*resIncr] += alpha*cc0;
+    res[(i+1)*resIncr] += alpha*cc1;
+    res[(i+2)*resIncr] += alpha*cc2;
+    res[(i+3)*resIncr] += alpha*cc3;
+    res[(i+4)*resIncr] += alpha*cc4;
+    res[(i+5)*resIncr] += alpha*cc5;
+    res[(i+6)*resIncr] += alpha*cc6;
+    res[(i+7)*resIncr] += alpha*cc7;
   }
-
-  Index offset1 = (FirstAligned && alignmentStep==1?3:1);
-  Index offset3 = (FirstAligned && alignmentStep==1?1:3);
-
-  Index rowBound = ((rows-skipRows)/rowsAtOnce)*rowsAtOnce + skipRows;
-  for (Index i=skipRows; i<rowBound; i+=rowsAtOnce)
+  for(; i<n4; i+=4)
   {
-    EIGEN_ALIGN16 ResScalar tmp0 = ResScalar(0);
-    ResScalar tmp1 = ResScalar(0), tmp2 = ResScalar(0), tmp3 = ResScalar(0);
+    ResPacket c0 = pset1<ResPacket>(ResScalar(0)),
+              c1 = pset1<ResPacket>(ResScalar(0)),
+              c2 = pset1<ResPacket>(ResScalar(0)),
+              c3 = pset1<ResPacket>(ResScalar(0));
 
-    // this helps the compiler generating good binary code
-    const LhsScalar *lhs0 = lhs + i*lhsStride,     *lhs1 = lhs + (i+offset1)*lhsStride,
-                    *lhs2 = lhs + (i+2)*lhsStride, *lhs3 = lhs + (i+offset3)*lhsStride;
+    Index j=0;
+    for(; j+LhsPacketSize<=cols; j+=LhsPacketSize)
+    {
+      RhsPacket b0 = rhs.template load<RhsPacket, Unaligned>(j,0);
 
-    if (Vectorizable)
+      c0 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+0,j),b0,c0);
+      c1 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+1,j),b0,c1);
+      c2 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+2,j),b0,c2);
+      c3 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+3,j),b0,c3);
+    }
+    ResScalar cc0 = predux(c0);
+    ResScalar cc1 = predux(c1);
+    ResScalar cc2 = predux(c2);
+    ResScalar cc3 = predux(c3);
+    for(; j<cols; ++j)
     {
-      /* explicit vectorization */
-      ResPacket ptmp0 = pset1<ResPacket>(ResScalar(0)), ptmp1 = pset1<ResPacket>(ResScalar(0)),
-                ptmp2 = pset1<ResPacket>(ResScalar(0)), ptmp3 = pset1<ResPacket>(ResScalar(0));
+      RhsScalar b0 = rhs(j,0);
 
-      // process initial unaligned coeffs
-      // FIXME this loop get vectorized by the compiler !
-      for (Index j=0; j<alignedStart; ++j)
-      {
-        RhsScalar b = rhs[j];
-        tmp0 += cj.pmul(lhs0[j],b); tmp1 += cj.pmul(lhs1[j],b);
-        tmp2 += cj.pmul(lhs2[j],b); tmp3 += cj.pmul(lhs3[j],b);
-      }
+      cc0 += cj.pmul(lhs(i+0,j), b0);
+      cc1 += cj.pmul(lhs(i+1,j), b0);
+      cc2 += cj.pmul(lhs(i+2,j), b0);
+      cc3 += cj.pmul(lhs(i+3,j), b0);
+    }
+    res[(i+0)*resIncr] += alpha*cc0;
+    res[(i+1)*resIncr] += alpha*cc1;
+    res[(i+2)*resIncr] += alpha*cc2;
+    res[(i+3)*resIncr] += alpha*cc3;
+  }
+  for(; i<n2; i+=2)
+  {
+    ResPacket c0 = pset1<ResPacket>(ResScalar(0)),
+              c1 = pset1<ResPacket>(ResScalar(0));
 
-      if (alignedSize>alignedStart)
-      {
-        switch(alignmentPattern)
-        {
-          case AllAligned:
-            for (Index j = alignedStart; j<alignedSize; j+=RhsPacketSize)
-              _EIGEN_ACCUMULATE_PACKETS(d,d,d);
-            break;
-          case EvenAligned:
-            for (Index j = alignedStart; j<alignedSize; j+=RhsPacketSize)
-              _EIGEN_ACCUMULATE_PACKETS(d,du,d);
-            break;
-          case FirstAligned:
-          {
-            Index j = alignedStart;
-            if (peels>1)
-            {
-              /* Here we proccess 4 rows with with two peeled iterations to hide
-               * the overhead of unaligned loads. Moreover unaligned loads are handled
-               * using special shift/move operations between the two aligned packets
-               * overlaping the desired unaligned packet. This is *much* more efficient
-               * than basic unaligned loads.
-               */
-              LhsPacket A01, A02, A03, A11, A12, A13;
-              A01 = pload<LhsPacket>(&lhs1[alignedStart-1]);
-              A02 = pload<LhsPacket>(&lhs2[alignedStart-2]);
-              A03 = pload<LhsPacket>(&lhs3[alignedStart-3]);
-
-              for (; j<peeledSize; j+=peels*RhsPacketSize)
-              {
-                RhsPacket b = pload<RhsPacket>(&rhs[j]);
-                A11 = pload<LhsPacket>(&lhs1[j-1+LhsPacketSize]);  palign<1>(A01,A11);
-                A12 = pload<LhsPacket>(&lhs2[j-2+LhsPacketSize]);  palign<2>(A02,A12);
-                A13 = pload<LhsPacket>(&lhs3[j-3+LhsPacketSize]);  palign<3>(A03,A13);
-
-                ptmp0 = pcj.pmadd(pload<LhsPacket>(&lhs0[j]), b, ptmp0);
-                ptmp1 = pcj.pmadd(A01, b, ptmp1);
-                A01 = pload<LhsPacket>(&lhs1[j-1+2*LhsPacketSize]);  palign<1>(A11,A01);
-                ptmp2 = pcj.pmadd(A02, b, ptmp2);
-                A02 = pload<LhsPacket>(&lhs2[j-2+2*LhsPacketSize]);  palign<2>(A12,A02);
-                ptmp3 = pcj.pmadd(A03, b, ptmp3);
-                A03 = pload<LhsPacket>(&lhs3[j-3+2*LhsPacketSize]);  palign<3>(A13,A03);
-
-                b = pload<RhsPacket>(&rhs[j+RhsPacketSize]);
-                ptmp0 = pcj.pmadd(pload<LhsPacket>(&lhs0[j+LhsPacketSize]), b, ptmp0);
-                ptmp1 = pcj.pmadd(A11, b, ptmp1);
-                ptmp2 = pcj.pmadd(A12, b, ptmp2);
-                ptmp3 = pcj.pmadd(A13, b, ptmp3);
-              }
-            }
-            for (; j<alignedSize; j+=RhsPacketSize)
-              _EIGEN_ACCUMULATE_PACKETS(d,du,du);
-            break;
-          }
-          default:
-            for (Index j = alignedStart; j<alignedSize; j+=RhsPacketSize)
-              _EIGEN_ACCUMULATE_PACKETS(du,du,du);
-            break;
-        }
-        tmp0 += predux(ptmp0);
-        tmp1 += predux(ptmp1);
-        tmp2 += predux(ptmp2);
-        tmp3 += predux(ptmp3);
-      }
-    } // end explicit vectorization
+    Index j=0;
+    for(; j+LhsPacketSize<=cols; j+=LhsPacketSize)
+    {
+      RhsPacket b0 = rhs.template load<RhsPacket, Unaligned>(j,0);
 
-    // process remaining coeffs (or all if no explicit vectorization)
-    // FIXME this loop get vectorized by the compiler !
-    for (Index j=alignedSize; j<depth; ++j)
+      c0 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+0,j),b0,c0);
+      c1 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i+1,j),b0,c1);
+    }
+    ResScalar cc0 = predux(c0);
+    ResScalar cc1 = predux(c1);
+    for(; j<cols; ++j)
     {
-      RhsScalar b = rhs[j];
-      tmp0 += cj.pmul(lhs0[j],b); tmp1 += cj.pmul(lhs1[j],b);
-      tmp2 += cj.pmul(lhs2[j],b); tmp3 += cj.pmul(lhs3[j],b);
+      RhsScalar b0 = rhs(j,0);
+
+      cc0 += cj.pmul(lhs(i+0,j), b0);
+      cc1 += cj.pmul(lhs(i+1,j), b0);
     }
-    res[i*resIncr]            += alpha*tmp0;
-    res[(i+offset1)*resIncr]  += alpha*tmp1;
-    res[(i+2)*resIncr]        += alpha*tmp2;
-    res[(i+offset3)*resIncr]  += alpha*tmp3;
+    res[(i+0)*resIncr] += alpha*cc0;
+    res[(i+1)*resIncr] += alpha*cc1;
   }
-
-  // process remaining first and last rows (at most columnsAtOnce-1)
-  Index end = rows;
-  Index start = rowBound;
-  do
+  for(; i<rows; ++i)
   {
-    for (Index i=start; i<end; ++i)
+    ResPacket c0 = pset1<ResPacket>(ResScalar(0));
+    Index j=0;
+    for(; j+LhsPacketSize<=cols; j+=LhsPacketSize)
     {
-      EIGEN_ALIGN16 ResScalar tmp0 = ResScalar(0);
-      ResPacket ptmp0 = pset1<ResPacket>(tmp0);
-      const LhsScalar* lhs0 = lhs + i*lhsStride;
-      // process first unaligned result's coeffs
-      // FIXME this loop get vectorized by the compiler !
-      for (Index j=0; j<alignedStart; ++j)
-        tmp0 += cj.pmul(lhs0[j], rhs[j]);
-
-      if (alignedSize>alignedStart)
-      {
-        // process aligned rhs coeffs
-        if ((size_t(lhs0+alignedStart)%sizeof(LhsPacket))==0)
-          for (Index j = alignedStart;j<alignedSize;j+=RhsPacketSize)
-            ptmp0 = pcj.pmadd(pload<LhsPacket>(&lhs0[j]), pload<RhsPacket>(&rhs[j]), ptmp0);
-        else
-          for (Index j = alignedStart;j<alignedSize;j+=RhsPacketSize)
-            ptmp0 = pcj.pmadd(ploadu<LhsPacket>(&lhs0[j]), pload<RhsPacket>(&rhs[j]), ptmp0);
-        tmp0 += predux(ptmp0);
-      }
-
-      // process remaining scalars
-      // FIXME this loop get vectorized by the compiler !
-      for (Index j=alignedSize; j<depth; ++j)
-        tmp0 += cj.pmul(lhs0[j], rhs[j]);
-      res[i*resIncr] += alpha*tmp0;
+      RhsPacket b0 = rhs.template load<RhsPacket,Unaligned>(j,0);
+      c0 = pcj.pmadd(lhs.template load<LhsPacket,LhsAlignment>(i,j),b0,c0);
     }
-    if (skipRows)
+    ResScalar cc0 = predux(c0);
+    for(; j<cols; ++j)
     {
-      start = 0;
-      end = skipRows;
-      skipRows = 0;
+      cc0 += cj.pmul(lhs(i,j), rhs(j,0));
     }
-    else
-      break;
-  } while(Vectorizable);
-
-  #undef _EIGEN_ACCUMULATE_PACKETS
+    res[i*resIncr] += alpha*cc0;
+  }
 }
 
 } // end namespace internal
diff --git a/eigen/Eigen/src/Core/products/GeneralMatrixVector_MKL.h b/eigen/Eigen/src/Core/products/GeneralMatrixVector_BLAS.h
index 1cb9fe6..e3a5d58 100644
--- a/eigen/Eigen/src/Core/products/GeneralMatrixVector_MKL.h
+++ b/eigen/Eigen/src/Core/products/GeneralMatrixVector_BLAS.h
@@ -25,13 +25,13 @@
  SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
  ********************************************************************************
- *   Content : Eigen bindings to Intel(R) MKL
+ *   Content : Eigen bindings to BLAS F77
  *   General matrix-vector product functionality based on ?GEMV.
  ********************************************************************************
 */
 
-#ifndef EIGEN_GENERAL_MATRIX_VECTOR_MKL_H
-#define EIGEN_GENERAL_MATRIX_VECTOR_MKL_H
+#ifndef EIGEN_GENERAL_MATRIX_VECTOR_BLAS_H
+#define EIGEN_GENERAL_MATRIX_VECTOR_BLAS_H
 
 namespace Eigen { 
 
@@ -46,47 +46,46 @@ namespace internal {
 
 // gemv specialization
 
-template<typename Index, typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs, typename RhsScalar, bool ConjugateRhs>
-struct general_matrix_vector_product_gemv :
-  general_matrix_vector_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,ConjugateRhs,BuiltIn> {};
+template<typename Index, typename LhsScalar, int StorageOrder, bool ConjugateLhs, typename RhsScalar, bool ConjugateRhs>
+struct general_matrix_vector_product_gemv;
 
-#define EIGEN_MKL_GEMV_SPECIALIZE(Scalar) \
+#define EIGEN_BLAS_GEMV_SPECIALIZE(Scalar) \
 template<typename Index, bool ConjugateLhs, bool ConjugateRhs> \
-struct general_matrix_vector_product<Index,Scalar,ColMajor,ConjugateLhs,Scalar,ConjugateRhs,Specialized> { \
+struct general_matrix_vector_product<Index,Scalar,const_blas_data_mapper<Scalar,Index,ColMajor>,ColMajor,ConjugateLhs,Scalar,const_blas_data_mapper<Scalar,Index,RowMajor>,ConjugateRhs,Specialized> { \
 static void run( \
   Index rows, Index cols, \
-  const Scalar* lhs, Index lhsStride, \
-  const Scalar* rhs, Index rhsIncr, \
+  const const_blas_data_mapper<Scalar,Index,ColMajor> &lhs, \
+  const const_blas_data_mapper<Scalar,Index,RowMajor> &rhs, \
   Scalar* res, Index resIncr, Scalar alpha) \
 { \
   if (ConjugateLhs) { \
-    general_matrix_vector_product<Index,Scalar,ColMajor,ConjugateLhs,Scalar,ConjugateRhs,BuiltIn>::run( \
-      rows, cols, lhs, lhsStride, rhs, rhsIncr, res, resIncr, alpha); \
+    general_matrix_vector_product<Index,Scalar,const_blas_data_mapper<Scalar,Index,ColMajor>,ColMajor,ConjugateLhs,Scalar,const_blas_data_mapper<Scalar,Index,RowMajor>,ConjugateRhs,BuiltIn>::run( \
+      rows, cols, lhs, rhs, res, resIncr, alpha); \
   } else { \
     general_matrix_vector_product_gemv<Index,Scalar,ColMajor,ConjugateLhs,Scalar,ConjugateRhs>::run( \
-      rows, cols, lhs, lhsStride, rhs, rhsIncr, res, resIncr, alpha); \
+      rows, cols, lhs.data(), lhs.stride(), rhs.data(), rhs.stride(), res, resIncr, alpha); \
   } \
 } \
 }; \
 template<typename Index, bool ConjugateLhs, bool ConjugateRhs> \
-struct general_matrix_vector_product<Index,Scalar,RowMajor,ConjugateLhs,Scalar,ConjugateRhs,Specialized> { \
+struct general_matrix_vector_product<Index,Scalar,const_blas_data_mapper<Scalar,Index,RowMajor>,RowMajor,ConjugateLhs,Scalar,const_blas_data_mapper<Scalar,Index,ColMajor>,ConjugateRhs,Specialized> { \
 static void run( \
   Index rows, Index cols, \
-  const Scalar* lhs, Index lhsStride, \
-  const Scalar* rhs, Index rhsIncr, \
+  const const_blas_data_mapper<Scalar,Index,RowMajor> &lhs, \
+  const const_blas_data_mapper<Scalar,Index,ColMajor> &rhs, \
   Scalar* res, Index resIncr, Scalar alpha) \
 { \
     general_matrix_vector_product_gemv<Index,Scalar,RowMajor,ConjugateLhs,Scalar,ConjugateRhs>::run( \
-      rows, cols, lhs, lhsStride, rhs, rhsIncr, res, resIncr, alpha); \
+      rows, cols, lhs.data(), lhs.stride(), rhs.data(), rhs.stride(), res, resIncr, alpha); \
 } \
 }; \
 
-EIGEN_MKL_GEMV_SPECIALIZE(double)
-EIGEN_MKL_GEMV_SPECIALIZE(float)
-EIGEN_MKL_GEMV_SPECIALIZE(dcomplex)
-EIGEN_MKL_GEMV_SPECIALIZE(scomplex)
+EIGEN_BLAS_GEMV_SPECIALIZE(double)
+EIGEN_BLAS_GEMV_SPECIALIZE(float)
+EIGEN_BLAS_GEMV_SPECIALIZE(dcomplex)
+EIGEN_BLAS_GEMV_SPECIALIZE(scomplex)
 
-#define EIGEN_MKL_GEMV_SPECIALIZATION(EIGTYPE,MKLTYPE,MKLPREFIX) \
+#define EIGEN_BLAS_GEMV_SPECIALIZATION(EIGTYPE,BLASTYPE,BLASPREFIX) \
 template<typename Index, int LhsStorageOrder, bool ConjugateLhs, bool ConjugateRhs> \
 struct general_matrix_vector_product_gemv<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,ConjugateRhs> \
 { \
@@ -98,16 +97,15 @@ static void run( \
   const EIGTYPE* rhs, Index rhsIncr, \
   EIGTYPE* res, Index resIncr, EIGTYPE alpha) \
 { \
-  MKL_INT m=rows, n=cols, lda=lhsStride, incx=rhsIncr, incy=resIncr; \
-  MKLTYPE alpha_, beta_; \
-  const EIGTYPE *x_ptr, myone(1); \
+  BlasIndex m=convert_index<BlasIndex>(rows), n=convert_index<BlasIndex>(cols), \
+            lda=convert_index<BlasIndex>(lhsStride), incx=convert_index<BlasIndex>(rhsIncr), incy=convert_index<BlasIndex>(resIncr); \
+  const EIGTYPE beta(1); \
+  const EIGTYPE *x_ptr; \
   char trans=(LhsStorageOrder==ColMajor) ? 'N' : (ConjugateLhs) ? 'C' : 'T'; \
   if (LhsStorageOrder==RowMajor) { \
-    m=cols; \
-    n=rows; \
+    m = convert_index<BlasIndex>(cols); \
+    n = convert_index<BlasIndex>(rows); \
   }\
-  assign_scalar_eig2mkl(alpha_, alpha); \
-  assign_scalar_eig2mkl(beta_, myone); \
   GEMVVector x_tmp; \
   if (ConjugateRhs) { \
     Map<const GEMVVector, 0, InnerStride<> > map_x(rhs,cols,1,InnerStride<>(incx)); \
@@ -115,17 +113,17 @@ static void run( \
     x_ptr=x_tmp.data(); \
     incx=1; \
   } else x_ptr=rhs; \
-  MKLPREFIX##gemv(&trans, &m, &n, &alpha_, (const MKLTYPE*)lhs, &lda, (const MKLTYPE*)x_ptr, &incx, &beta_, (MKLTYPE*)res, &incy); \
+  BLASPREFIX##gemv_(&trans, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)lhs, &lda, (const BLASTYPE*)x_ptr, &incx, &numext::real_ref(beta), (BLASTYPE*)res, &incy); \
 }\
 };
 
-EIGEN_MKL_GEMV_SPECIALIZATION(double,   double,        d)
-EIGEN_MKL_GEMV_SPECIALIZATION(float,    float,         s)
-EIGEN_MKL_GEMV_SPECIALIZATION(dcomplex, MKL_Complex16, z)
-EIGEN_MKL_GEMV_SPECIALIZATION(scomplex, MKL_Complex8,  c)
+EIGEN_BLAS_GEMV_SPECIALIZATION(double,   double, d)
+EIGEN_BLAS_GEMV_SPECIALIZATION(float,    float,  s)
+EIGEN_BLAS_GEMV_SPECIALIZATION(dcomplex, double, z)
+EIGEN_BLAS_GEMV_SPECIALIZATION(scomplex, float,  c)
 
 } // end namespase internal
 
 } // end namespace Eigen
 
-#endif // EIGEN_GENERAL_MATRIX_VECTOR_MKL_H
+#endif // EIGEN_GENERAL_MATRIX_VECTOR_BLAS_H
diff --git a/eigen/Eigen/src/Core/products/Parallelizer.h b/eigen/Eigen/src/Core/products/Parallelizer.h
index 6937ee3..c2f084c 100644
--- a/eigen/Eigen/src/Core/products/Parallelizer.h
+++ b/eigen/Eigen/src/Core/products/Parallelizer.h
@@ -10,7 +10,7 @@
 #ifndef EIGEN_PARALLELIZER_H
 #define EIGEN_PARALLELIZER_H
 
-namespace Eigen { 
+namespace Eigen {
 
 namespace internal {
 
@@ -49,8 +49,8 @@ inline void initParallel()
 {
   int nbt;
   internal::manage_multi_threading(GetAction, &nbt);
-  std::ptrdiff_t l1, l2;
-  internal::manage_caching_sizes(GetAction, &l1, &l2);
+  std::ptrdiff_t l1, l2, l3;
+  internal::manage_caching_sizes(GetAction, &l1, &l2, &l3);
 }
 
 /** \returns the max number of threads reserved for Eigen
@@ -73,17 +73,17 @@ namespace internal {
 
 template<typename Index> struct GemmParallelInfo
 {
-  GemmParallelInfo() : sync(-1), users(0), rhs_start(0), rhs_length(0) {}
+  GemmParallelInfo() : sync(-1), users(0), lhs_start(0), lhs_length(0) {}
 
-  int volatile sync;
+  Index volatile sync;
   int volatile users;
 
-  Index rhs_start;
-  Index rhs_length;
+  Index lhs_start;
+  Index lhs_length;
 };
 
 template<bool Condition, typename Functor, typename Index>
-void parallelize_gemm(const Functor& func, Index rows, Index cols, bool transpose)
+void parallelize_gemm(const Functor& func, Index rows, Index cols, Index depth, bool transpose)
 {
   // TODO when EIGEN_USE_BLAS is defined,
   // we should still enable OMP for other scalar types
@@ -92,6 +92,7 @@ void parallelize_gemm(const Functor& func, Index rows, Index cols, bool transpos
   // the matrix product when multithreading is enabled. This is a temporary
   // fix to support row-major destination matrices. This whole
   // parallelizer mechanism has to be redisigned anyway.
+  EIGEN_UNUSED_VARIABLE(depth);
   EIGEN_UNUSED_VARIABLE(transpose);
   func(0,rows, 0,cols);
 #else
@@ -102,56 +103,56 @@ void parallelize_gemm(const Functor& func, Index rows, Index cols, bool transpos
   // - we are not already in a parallel code
   // - the sizes are large enough
 
-  // 1- are we already in a parallel session?
-  // FIXME omp_get_num_threads()>1 only works for openmp, what if the user does not use openmp?
-  if((!Condition) || (omp_get_num_threads()>1))
-    return func(0,rows, 0,cols);
+  // compute the maximal number of threads from the size of the product:
+  // This first heuristic takes into account that the product kernel is fully optimized when working with nr columns at once.
+  Index size = transpose ? rows : cols;
+  Index pb_max_threads = std::max<Index>(1,size / Functor::Traits::nr);
 
-  Index size = transpose ? cols : rows;
+  // compute the maximal number of threads from the total amount of work:
+  double work = static_cast<double>(rows) * static_cast<double>(cols) *
+      static_cast<double>(depth);
+  double kMinTaskSize = 50000;  // FIXME improve this heuristic.
+  pb_max_threads = std::max<Index>(1, std::min<Index>(pb_max_threads, work / kMinTaskSize));
 
-  // 2- compute the maximal number of threads from the size of the product:
-  // FIXME this has to be fine tuned
-  Index max_threads = std::max<Index>(1,size / 32);
+  // compute the number of threads we are going to use
+  Index threads = std::min<Index>(nbThreads(), pb_max_threads);
 
-  // 3 - compute the number of threads we are going to use
-  Index threads = std::min<Index>(nbThreads(), max_threads);
-
-  if(threads==1)
+  // if multi-threading is explicitely disabled, not useful, or if we already are in a parallel session,
+  // then abort multi-threading
+  // FIXME omp_get_num_threads()>1 only works for openmp, what if the user does not use openmp?
+  if((!Condition) || (threads==1) || (omp_get_num_threads()>1))
     return func(0,rows, 0,cols);
 
   Eigen::initParallel();
-  func.initParallelSession();
+  func.initParallelSession(threads);
 
   if(transpose)
     std::swap(rows,cols);
 
-  GemmParallelInfo<Index>* info = new GemmParallelInfo<Index>[threads];
+  ei_declare_aligned_stack_constructed_variable(GemmParallelInfo<Index>,info,threads,0);
 
   #pragma omp parallel num_threads(threads)
   {
     Index i = omp_get_thread_num();
     // Note that the actual number of threads might be lower than the number of request ones.
     Index actual_threads = omp_get_num_threads();
-    
+
     Index blockCols = (cols / actual_threads) & ~Index(0x3);
-    Index blockRows = (rows / actual_threads) & ~Index(0x7);
-    
+    Index blockRows = (rows / actual_threads);
+    blockRows = (blockRows/Functor::Traits::mr)*Functor::Traits::mr;
+
     Index r0 = i*blockRows;
     Index actualBlockRows = (i+1==actual_threads) ? rows-r0 : blockRows;
 
     Index c0 = i*blockCols;
     Index actualBlockCols = (i+1==actual_threads) ? cols-c0 : blockCols;
 
-    info[i].rhs_start = c0;
-    info[i].rhs_length = actualBlockCols;
+    info[i].lhs_start = r0;
+    info[i].lhs_length = actualBlockRows;
 
-    if(transpose)
-      func(0, cols, r0, actualBlockRows, info);
-    else
-      func(r0, actualBlockRows, 0,cols, info);
+    if(transpose) func(c0, actualBlockCols, 0, rows, info);
+    else          func(0, rows, c0, actualBlockCols, info);
   }
-
-  delete[] info;
 #endif
 }
 
diff --git a/eigen/Eigen/src/Core/products/SelfadjointMatrixMatrix.h b/eigen/Eigen/src/Core/products/SelfadjointMatrixMatrix.h
index 99cf9e0..da6f82a 100644
--- a/eigen/Eigen/src/Core/products/SelfadjointMatrixMatrix.h
+++ b/eigen/Eigen/src/Core/products/SelfadjointMatrixMatrix.h
@@ -15,7 +15,7 @@ namespace Eigen {
 namespace internal {
 
 // pack a selfadjoint block diagonal for use with the gebp_kernel
-template<typename Scalar, typename Index, int Pack1, int Pack2, int StorageOrder>
+template<typename Scalar, typename Index, int Pack1, int Pack2_dummy, int StorageOrder>
 struct symm_pack_lhs
 {
   template<int BlockRows> inline
@@ -45,25 +45,32 @@ struct symm_pack_lhs
   }
   void operator()(Scalar* blockA, const Scalar* _lhs, Index lhsStride, Index cols, Index rows)
   {
+    enum { PacketSize = packet_traits<Scalar>::size };
     const_blas_data_mapper<Scalar,Index,StorageOrder> lhs(_lhs,lhsStride);
     Index count = 0;
-    Index peeled_mc = (rows/Pack1)*Pack1;
-    for(Index i=0; i<peeled_mc; i+=Pack1)
-    {
-      pack<Pack1>(blockA, lhs, cols, i, count);
-    }
-
-    if(rows-peeled_mc>=Pack2)
-    {
-      pack<Pack2>(blockA, lhs, cols, peeled_mc, count);
-      peeled_mc += Pack2;
-    }
+    //Index peeled_mc3 = (rows/Pack1)*Pack1;
+    
+    const Index peeled_mc3 = Pack1>=3*PacketSize ? (rows/(3*PacketSize))*(3*PacketSize) : 0;
+    const Index peeled_mc2 = Pack1>=2*PacketSize ? peeled_mc3+((rows-peeled_mc3)/(2*PacketSize))*(2*PacketSize) : 0;
+    const Index peeled_mc1 = Pack1>=1*PacketSize ? (rows/(1*PacketSize))*(1*PacketSize) : 0;
+    
+    if(Pack1>=3*PacketSize)
+      for(Index i=0; i<peeled_mc3; i+=3*PacketSize)
+        pack<3*PacketSize>(blockA, lhs, cols, i, count);
+    
+    if(Pack1>=2*PacketSize)
+      for(Index i=peeled_mc3; i<peeled_mc2; i+=2*PacketSize)
+        pack<2*PacketSize>(blockA, lhs, cols, i, count);
+    
+    if(Pack1>=1*PacketSize)
+      for(Index i=peeled_mc2; i<peeled_mc1; i+=1*PacketSize)
+        pack<1*PacketSize>(blockA, lhs, cols, i, count);
 
     // do the same with mr==1
-    for(Index i=peeled_mc; i<rows; i++)
+    for(Index i=peeled_mc1; i<rows; i++)
     {
       for(Index k=0; k<i; k++)
-        blockA[count++] = lhs(i, k);              // normal
+        blockA[count++] = lhs(i, k);                   // normal
 
       blockA[count++] = numext::real(lhs(i, i));       // real (diagonal)
 
@@ -82,7 +89,8 @@ struct symm_pack_rhs
     Index end_k = k2 + rows;
     Index count = 0;
     const_blas_data_mapper<Scalar,Index,StorageOrder> rhs(_rhs,rhsStride);
-    Index packet_cols = (cols/nr)*nr;
+    Index packet_cols8 = nr>=8 ? (cols/8) * 8 : 0;
+    Index packet_cols4 = nr>=4 ? (cols/4) * 4 : 0;
 
     // first part: normal case
     for(Index j2=0; j2<k2; j2+=nr)
@@ -91,79 +99,151 @@ struct symm_pack_rhs
       {
         blockB[count+0] = rhs(k,j2+0);
         blockB[count+1] = rhs(k,j2+1);
-        if (nr==4)
+        if (nr>=4)
         {
           blockB[count+2] = rhs(k,j2+2);
           blockB[count+3] = rhs(k,j2+3);
         }
+        if (nr>=8)
+        {
+          blockB[count+4] = rhs(k,j2+4);
+          blockB[count+5] = rhs(k,j2+5);
+          blockB[count+6] = rhs(k,j2+6);
+          blockB[count+7] = rhs(k,j2+7);
+        }
         count += nr;
       }
     }
 
     // second part: diagonal block
-    for(Index j2=k2; j2<(std::min)(k2+rows,packet_cols); j2+=nr)
+    Index end8 = nr>=8 ? (std::min)(k2+rows,packet_cols8) : k2;
+    if(nr>=8)
     {
-      // again we can split vertically in three different parts (transpose, symmetric, normal)
-      // transpose
-      for(Index k=k2; k<j2; k++)
+      for(Index j2=k2; j2<end8; j2+=8)
       {
-        blockB[count+0] = numext::conj(rhs(j2+0,k));
-        blockB[count+1] = numext::conj(rhs(j2+1,k));
-        if (nr==4)
+        // again we can split vertically in three different parts (transpose, symmetric, normal)
+        // transpose
+        for(Index k=k2; k<j2; k++)
         {
+          blockB[count+0] = numext::conj(rhs(j2+0,k));
+          blockB[count+1] = numext::conj(rhs(j2+1,k));
           blockB[count+2] = numext::conj(rhs(j2+2,k));
           blockB[count+3] = numext::conj(rhs(j2+3,k));
+          blockB[count+4] = numext::conj(rhs(j2+4,k));
+          blockB[count+5] = numext::conj(rhs(j2+5,k));
+          blockB[count+6] = numext::conj(rhs(j2+6,k));
+          blockB[count+7] = numext::conj(rhs(j2+7,k));
+          count += 8;
         }
-        count += nr;
-      }
-      // symmetric
-      Index h = 0;
-      for(Index k=j2; k<j2+nr; k++)
-      {
-        // normal
-        for (Index w=0 ; w<h; ++w)
-          blockB[count+w] = rhs(k,j2+w);
+        // symmetric
+        Index h = 0;
+        for(Index k=j2; k<j2+8; k++)
+        {
+          // normal
+          for (Index w=0 ; w<h; ++w)
+            blockB[count+w] = rhs(k,j2+w);
 
-        blockB[count+h] = numext::real(rhs(k,k));
+          blockB[count+h] = numext::real(rhs(k,k));
 
-        // transpose
-        for (Index w=h+1 ; w<nr; ++w)
-          blockB[count+w] = numext::conj(rhs(j2+w,k));
-        count += nr;
-        ++h;
+          // transpose
+          for (Index w=h+1 ; w<8; ++w)
+            blockB[count+w] = numext::conj(rhs(j2+w,k));
+          count += 8;
+          ++h;
+        }
+        // normal
+        for(Index k=j2+8; k<end_k; k++)
+        {
+          blockB[count+0] = rhs(k,j2+0);
+          blockB[count+1] = rhs(k,j2+1);
+          blockB[count+2] = rhs(k,j2+2);
+          blockB[count+3] = rhs(k,j2+3);
+          blockB[count+4] = rhs(k,j2+4);
+          blockB[count+5] = rhs(k,j2+5);
+          blockB[count+6] = rhs(k,j2+6);
+          blockB[count+7] = rhs(k,j2+7);
+          count += 8;
+        }
       }
-      // normal
-      for(Index k=j2+nr; k<end_k; k++)
+    }
+    if(nr>=4)
+    {
+      for(Index j2=end8; j2<(std::min)(k2+rows,packet_cols4); j2+=4)
       {
-        blockB[count+0] = rhs(k,j2+0);
-        blockB[count+1] = rhs(k,j2+1);
-        if (nr==4)
+        // again we can split vertically in three different parts (transpose, symmetric, normal)
+        // transpose
+        for(Index k=k2; k<j2; k++)
+        {
+          blockB[count+0] = numext::conj(rhs(j2+0,k));
+          blockB[count+1] = numext::conj(rhs(j2+1,k));
+          blockB[count+2] = numext::conj(rhs(j2+2,k));
+          blockB[count+3] = numext::conj(rhs(j2+3,k));
+          count += 4;
+        }
+        // symmetric
+        Index h = 0;
+        for(Index k=j2; k<j2+4; k++)
         {
+          // normal
+          for (Index w=0 ; w<h; ++w)
+            blockB[count+w] = rhs(k,j2+w);
+
+          blockB[count+h] = numext::real(rhs(k,k));
+
+          // transpose
+          for (Index w=h+1 ; w<4; ++w)
+            blockB[count+w] = numext::conj(rhs(j2+w,k));
+          count += 4;
+          ++h;
+        }
+        // normal
+        for(Index k=j2+4; k<end_k; k++)
+        {
+          blockB[count+0] = rhs(k,j2+0);
+          blockB[count+1] = rhs(k,j2+1);
           blockB[count+2] = rhs(k,j2+2);
           blockB[count+3] = rhs(k,j2+3);
+          count += 4;
         }
-        count += nr;
       }
     }
 
     // third part: transposed
-    for(Index j2=k2+rows; j2<packet_cols; j2+=nr)
+    if(nr>=8)
     {
-      for(Index k=k2; k<end_k; k++)
+      for(Index j2=k2+rows; j2<packet_cols8; j2+=8)
       {
-        blockB[count+0] = numext::conj(rhs(j2+0,k));
-        blockB[count+1] = numext::conj(rhs(j2+1,k));
-        if (nr==4)
+        for(Index k=k2; k<end_k; k++)
         {
+          blockB[count+0] = numext::conj(rhs(j2+0,k));
+          blockB[count+1] = numext::conj(rhs(j2+1,k));
           blockB[count+2] = numext::conj(rhs(j2+2,k));
           blockB[count+3] = numext::conj(rhs(j2+3,k));
+          blockB[count+4] = numext::conj(rhs(j2+4,k));
+          blockB[count+5] = numext::conj(rhs(j2+5,k));
+          blockB[count+6] = numext::conj(rhs(j2+6,k));
+          blockB[count+7] = numext::conj(rhs(j2+7,k));
+          count += 8;
+        }
+      }
+    }
+    if(nr>=4)
+    {
+      for(Index j2=(std::max)(packet_cols8,k2+rows); j2<packet_cols4; j2+=4)
+      {
+        for(Index k=k2; k<end_k; k++)
+        {
+          blockB[count+0] = numext::conj(rhs(j2+0,k));
+          blockB[count+1] = numext::conj(rhs(j2+1,k));
+          blockB[count+2] = numext::conj(rhs(j2+2,k));
+          blockB[count+3] = numext::conj(rhs(j2+3,k));
+          count += 4;
         }
-        count += nr;
       }
     }
 
     // copy the remaining columns one at a time (=> the same with nr==1)
-    for(Index j2=packet_cols; j2<cols; ++j2)
+    for(Index j2=packet_cols4; j2<cols; ++j2)
     {
       // transpose
       Index half = (std::min)(end_k,j2);
@@ -211,7 +291,7 @@ struct product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,LhsSelfAdjoint,Co
     const Scalar* lhs, Index lhsStride,
     const Scalar* rhs, Index rhsStride,
     Scalar* res,       Index resStride,
-    const Scalar& alpha)
+    const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking)
   {
     product_selfadjoint_matrix<Scalar, Index,
       EIGEN_LOGICAL_XOR(RhsSelfAdjoint,RhsStorageOrder==RowMajor) ? ColMajor : RowMajor,
@@ -219,7 +299,7 @@ struct product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,LhsSelfAdjoint,Co
       EIGEN_LOGICAL_XOR(LhsSelfAdjoint,LhsStorageOrder==RowMajor) ? ColMajor : RowMajor,
       LhsSelfAdjoint, NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(LhsSelfAdjoint,ConjugateLhs),
       ColMajor>
-      ::run(cols, rows,  rhs, rhsStride,  lhs, lhsStride,  res, resStride,  alpha);
+      ::run(cols, rows,  rhs, rhsStride,  lhs, lhsStride,  res, resStride,  alpha, blocking);
   }
 };
 
@@ -234,7 +314,7 @@ struct product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,true,ConjugateLhs
     const Scalar* _lhs, Index lhsStride,
     const Scalar* _rhs, Index rhsStride,
     Scalar* res,        Index resStride,
-    const Scalar& alpha);
+    const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking);
 };
 
 template <typename Scalar, typename Index,
@@ -244,33 +324,35 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,t
     Index rows, Index cols,
     const Scalar* _lhs, Index lhsStride,
     const Scalar* _rhs, Index rhsStride,
-    Scalar* res,        Index resStride,
-    const Scalar& alpha)
+    Scalar* _res,        Index resStride,
+    const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking)
   {
     Index size = rows;
 
-    const_blas_data_mapper<Scalar, Index, LhsStorageOrder> lhs(_lhs,lhsStride);
-    const_blas_data_mapper<Scalar, Index, RhsStorageOrder> rhs(_rhs,rhsStride);
-
     typedef gebp_traits<Scalar,Scalar> Traits;
 
-    Index kc = size;  // cache block size along the K direction
-    Index mc = rows;  // cache block size along the M direction
-    Index nc = cols;  // cache block size along the N direction
-    computeProductBlockingSizes<Scalar,Scalar>(kc, mc, nc);
-    // kc must smaller than mc
+    typedef const_blas_data_mapper<Scalar, Index, LhsStorageOrder> LhsMapper;
+    typedef const_blas_data_mapper<Scalar, Index, (LhsStorageOrder == RowMajor) ? ColMajor : RowMajor> LhsTransposeMapper;
+    typedef const_blas_data_mapper<Scalar, Index, RhsStorageOrder> RhsMapper;
+    typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper;
+    LhsMapper lhs(_lhs,lhsStride);
+    LhsTransposeMapper lhs_transpose(_lhs,lhsStride);
+    RhsMapper rhs(_rhs,rhsStride);
+    ResMapper res(_res, resStride);
+
+    Index kc = blocking.kc();                   // cache block size along the K direction
+    Index mc = (std::min)(rows,blocking.mc());  // cache block size along the M direction
+    // kc must be smaller than mc
     kc = (std::min)(kc,mc);
+    std::size_t sizeA = kc*mc;
+    std::size_t sizeB = kc*cols;
+    ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA());
+    ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
 
-    std::size_t sizeW = kc*Traits::WorkSpaceFactor;
-    std::size_t sizeB = sizeW + kc*cols;
-    ei_declare_aligned_stack_constructed_variable(Scalar, blockA, kc*mc, 0);
-    ei_declare_aligned_stack_constructed_variable(Scalar, allocatedBlockB, sizeB, 0);
-    Scalar* blockB = allocatedBlockB + sizeW;
-
-    gebp_kernel<Scalar, Scalar, Index, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
+    gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
     symm_pack_lhs<Scalar, Index, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
-    gemm_pack_rhs<Scalar, Index, Traits::nr,RhsStorageOrder> pack_rhs;
-    gemm_pack_lhs<Scalar, Index, Traits::mr, Traits::LhsProgress, LhsStorageOrder==RowMajor?ColMajor:RowMajor, true> pack_lhs_transposed;
+    gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder> pack_rhs;
+    gemm_pack_lhs<Scalar, Index, LhsTransposeMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder==RowMajor?ColMajor:RowMajor, true> pack_lhs_transposed;
 
     for(Index k2=0; k2<size; k2+=kc)
     {
@@ -279,7 +361,7 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,t
       // we have selected one row panel of rhs and one column panel of lhs
       // pack rhs's panel into a sequential chunk of memory
       // and expand each coeff to a constant packet for further reuse
-      pack_rhs(blockB, &rhs(k2,0), rhsStride, actual_kc, cols);
+      pack_rhs(blockB, rhs.getSubMapper(k2,0), actual_kc, cols);
 
       // the select lhs's panel has to be split in three different parts:
       //  1 - the transposed panel above the diagonal block => transposed packed copy
@@ -289,9 +371,9 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,t
       {
         const Index actual_mc = (std::min)(i2+mc,k2)-i2;
         // transposed packed copy
-        pack_lhs_transposed(blockA, &lhs(k2, i2), lhsStride, actual_kc, actual_mc);
+        pack_lhs_transposed(blockA, lhs_transpose.getSubMapper(i2, k2), actual_kc, actual_mc);
 
-        gebp_kernel(res+i2, resStride, blockA, blockB, actual_mc, actual_kc, cols, alpha);
+        gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc, cols, alpha);
       }
       // the block diagonal
       {
@@ -299,16 +381,16 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,t
         // symmetric packed copy
         pack_lhs(blockA, &lhs(k2,k2), lhsStride, actual_kc, actual_mc);
 
-        gebp_kernel(res+k2, resStride, blockA, blockB, actual_mc, actual_kc, cols, alpha);
+        gebp_kernel(res.getSubMapper(k2, 0), blockA, blockB, actual_mc, actual_kc, cols, alpha);
       }
 
       for(Index i2=k2+kc; i2<size; i2+=mc)
       {
         const Index actual_mc = (std::min)(i2+mc,size)-i2;
-        gemm_pack_lhs<Scalar, Index, Traits::mr, Traits::LhsProgress, LhsStorageOrder,false>()
-          (blockA, &lhs(i2, k2), lhsStride, actual_kc, actual_mc);
+        gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder,false>()
+          (blockA, lhs.getSubMapper(i2, k2), actual_kc, actual_mc);
 
-        gebp_kernel(res+i2, resStride, blockA, blockB, actual_mc, actual_kc, cols, alpha);
+        gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc, cols, alpha);
       }
     }
   }
@@ -325,7 +407,7 @@ struct product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,false,ConjugateLh
     const Scalar* _lhs, Index lhsStride,
     const Scalar* _rhs, Index rhsStride,
     Scalar* res,        Index resStride,
-    const Scalar& alpha);
+    const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking);
 };
 
 template <typename Scalar, typename Index,
@@ -335,27 +417,27 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,f
     Index rows, Index cols,
     const Scalar* _lhs, Index lhsStride,
     const Scalar* _rhs, Index rhsStride,
-    Scalar* res,        Index resStride,
-    const Scalar& alpha)
+    Scalar* _res,        Index resStride,
+    const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking)
   {
     Index size = cols;
 
-    const_blas_data_mapper<Scalar, Index, LhsStorageOrder> lhs(_lhs,lhsStride);
-
     typedef gebp_traits<Scalar,Scalar> Traits;
 
-    Index kc = size; // cache block size along the K direction
-    Index mc = rows;  // cache block size along the M direction
-    Index nc = cols;  // cache block size along the N direction
-    computeProductBlockingSizes<Scalar,Scalar>(kc, mc, nc);
-    std::size_t sizeW = kc*Traits::WorkSpaceFactor;
-    std::size_t sizeB = sizeW + kc*cols;
-    ei_declare_aligned_stack_constructed_variable(Scalar, blockA, kc*mc, 0);
-    ei_declare_aligned_stack_constructed_variable(Scalar, allocatedBlockB, sizeB, 0);
-    Scalar* blockB = allocatedBlockB + sizeW;
-
-    gebp_kernel<Scalar, Scalar, Index, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
-    gemm_pack_lhs<Scalar, Index, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+    typedef const_blas_data_mapper<Scalar, Index, LhsStorageOrder> LhsMapper;
+    typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper;
+    LhsMapper lhs(_lhs,lhsStride);
+    ResMapper res(_res,resStride);
+
+    Index kc = blocking.kc();                   // cache block size along the K direction
+    Index mc = (std::min)(rows,blocking.mc());  // cache block size along the M direction
+    std::size_t sizeA = kc*mc;
+    std::size_t sizeB = kc*cols;
+    ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA());
+    ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
+
+    gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
+    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
     symm_pack_rhs<Scalar, Index, Traits::nr,RhsStorageOrder> pack_rhs;
 
     for(Index k2=0; k2<size; k2+=kc)
@@ -368,9 +450,9 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,f
       for(Index i2=0; i2<rows; i2+=mc)
       {
         const Index actual_mc = (std::min)(i2+mc,rows)-i2;
-        pack_lhs(blockA, &lhs(i2, k2), lhsStride, actual_kc, actual_mc);
+        pack_lhs(blockA, lhs.getSubMapper(i2, k2), actual_kc, actual_mc);
 
-        gebp_kernel(res+i2, resStride, blockA, blockB, actual_mc, actual_kc, cols, alpha);
+        gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc, cols, alpha);
       }
     }
   }
@@ -382,55 +464,58 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,f
 ***************************************************************************/
 
 namespace internal {
+  
 template<typename Lhs, int LhsMode, typename Rhs, int RhsMode>
-struct traits<SelfadjointProductMatrix<Lhs,LhsMode,false,Rhs,RhsMode,false> >
-  : traits<ProductBase<SelfadjointProductMatrix<Lhs,LhsMode,false,Rhs,RhsMode,false>, Lhs, Rhs> >
-{};
-}
-
-template<typename Lhs, int LhsMode, typename Rhs, int RhsMode>
-struct SelfadjointProductMatrix<Lhs,LhsMode,false,Rhs,RhsMode,false>
-  : public ProductBase<SelfadjointProductMatrix<Lhs,LhsMode,false,Rhs,RhsMode,false>, Lhs, Rhs >
+struct selfadjoint_product_impl<Lhs,LhsMode,false,Rhs,RhsMode,false>
 {
-  EIGEN_PRODUCT_PUBLIC_INTERFACE(SelfadjointProductMatrix)
-
-  SelfadjointProductMatrix(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) {}
-
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  
+  typedef internal::blas_traits<Lhs> LhsBlasTraits;
+  typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
+  typedef internal::blas_traits<Rhs> RhsBlasTraits;
+  typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
+  
   enum {
     LhsIsUpper = (LhsMode&(Upper|Lower))==Upper,
     LhsIsSelfAdjoint = (LhsMode&SelfAdjoint)==SelfAdjoint,
     RhsIsUpper = (RhsMode&(Upper|Lower))==Upper,
     RhsIsSelfAdjoint = (RhsMode&SelfAdjoint)==SelfAdjoint
   };
-
-  template<typename Dest> void scaleAndAddTo(Dest& dst, const Scalar& alpha) const
+  
+  template<typename Dest>
+  static void run(Dest &dst, const Lhs &a_lhs, const Rhs &a_rhs, const Scalar& alpha)
   {
-    eigen_assert(dst.rows()==m_lhs.rows() && dst.cols()==m_rhs.cols());
+    eigen_assert(dst.rows()==a_lhs.rows() && dst.cols()==a_rhs.cols());
 
-    typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(m_lhs);
-    typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(m_rhs);
+    typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(a_lhs);
+    typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(a_rhs);
 
-    Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(m_lhs)
-                               * RhsBlasTraits::extractScalarFactor(m_rhs);
+    Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(a_lhs)
+                               * RhsBlasTraits::extractScalarFactor(a_rhs);
+
+    typedef internal::gemm_blocking_space<(Dest::Flags&RowMajorBit) ? RowMajor : ColMajor,Scalar,Scalar,
+              Lhs::MaxRowsAtCompileTime, Rhs::MaxColsAtCompileTime, Lhs::MaxColsAtCompileTime,1> BlockingType;
+
+    BlockingType blocking(lhs.rows(), rhs.cols(), lhs.cols(), 1, false);
 
     internal::product_selfadjoint_matrix<Scalar, Index,
-      EIGEN_LOGICAL_XOR(LhsIsUpper,
-                        internal::traits<Lhs>::Flags &RowMajorBit) ? RowMajor : ColMajor, LhsIsSelfAdjoint,
+      EIGEN_LOGICAL_XOR(LhsIsUpper,internal::traits<Lhs>::Flags &RowMajorBit) ? RowMajor : ColMajor, LhsIsSelfAdjoint,
       NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(LhsIsUpper,bool(LhsBlasTraits::NeedToConjugate)),
-      EIGEN_LOGICAL_XOR(RhsIsUpper,
-                        internal::traits<Rhs>::Flags &RowMajorBit) ? RowMajor : ColMajor, RhsIsSelfAdjoint,
+      EIGEN_LOGICAL_XOR(RhsIsUpper,internal::traits<Rhs>::Flags &RowMajorBit) ? RowMajor : ColMajor, RhsIsSelfAdjoint,
       NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(RhsIsUpper,bool(RhsBlasTraits::NeedToConjugate)),
       internal::traits<Dest>::Flags&RowMajorBit  ? RowMajor : ColMajor>
       ::run(
         lhs.rows(), rhs.cols(),                 // sizes
-        &lhs.coeffRef(0,0),    lhs.outerStride(),  // lhs info
-        &rhs.coeffRef(0,0),    rhs.outerStride(),  // rhs info
+        &lhs.coeffRef(0,0), lhs.outerStride(),  // lhs info
+        &rhs.coeffRef(0,0), rhs.outerStride(),  // rhs info
         &dst.coeffRef(0,0), dst.outerStride(),  // result info
-        actualAlpha                             // alpha
+        actualAlpha, blocking                   // alpha
       );
   }
 };
 
+} // end namespace internal
+
 } // end namespace Eigen
 
 #endif // EIGEN_SELFADJOINT_MATRIX_MATRIX_H
diff --git a/eigen/Eigen/src/Core/products/SelfadjointMatrixMatrix_MKL.h b/eigen/Eigen/src/Core/products/SelfadjointMatrixMatrix_BLAS.h
index dfa687f..a45238d 100644
--- a/eigen/Eigen/src/Core/products/SelfadjointMatrixMatrix_MKL.h
+++ b/eigen/Eigen/src/Core/products/SelfadjointMatrixMatrix_BLAS.h
@@ -25,13 +25,13 @@
  SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 //
  ********************************************************************************
- *   Content : Eigen bindings to Intel(R) MKL
+ *   Content : Eigen bindings to BLAS F77
  *   Self adjoint matrix * matrix product functionality based on ?SYMM/?HEMM.
  ********************************************************************************
 */
 
-#ifndef EIGEN_SELFADJOINT_MATRIX_MATRIX_MKL_H
-#define EIGEN_SELFADJOINT_MATRIX_MATRIX_MKL_H
+#ifndef EIGEN_SELFADJOINT_MATRIX_MATRIX_BLAS_H
+#define EIGEN_SELFADJOINT_MATRIX_MATRIX_BLAS_H
 
 namespace Eigen { 
 
@@ -40,7 +40,7 @@ namespace internal {
 
 /* Optimized selfadjoint matrix * matrix (?SYMM/?HEMM) product */
 
-#define EIGEN_MKL_SYMM_L(EIGTYPE, MKLTYPE, EIGPREFIX, MKLPREFIX) \
+#define EIGEN_BLAS_SYMM_L(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \
 template <typename Index, \
           int LhsStorageOrder, bool ConjugateLhs, \
           int RhsStorageOrder, bool ConjugateRhs> \
@@ -52,28 +52,23 @@ struct product_selfadjoint_matrix<EIGTYPE,Index,LhsStorageOrder,true,ConjugateLh
     const EIGTYPE* _lhs, Index lhsStride, \
     const EIGTYPE* _rhs, Index rhsStride, \
     EIGTYPE* res,        Index resStride, \
-    EIGTYPE alpha) \
+    EIGTYPE alpha, level3_blocking<EIGTYPE, EIGTYPE>& /*blocking*/) \
   { \
     char side='L', uplo='L'; \
-    MKL_INT m, n, lda, ldb, ldc; \
+    BlasIndex m, n, lda, ldb, ldc; \
     const EIGTYPE *a, *b; \
-    MKLTYPE alpha_, beta_; \
+    EIGTYPE beta(1); \
     MatrixX##EIGPREFIX b_tmp; \
-    EIGTYPE myone(1);\
 \
 /* Set transpose options */ \
 /* Set m, n, k */ \
-    m = (MKL_INT)rows;  \
-    n = (MKL_INT)cols;  \
-\
-/* Set alpha_ & beta_ */ \
-    assign_scalar_eig2mkl(alpha_, alpha); \
-    assign_scalar_eig2mkl(beta_, myone); \
+    m = convert_index<BlasIndex>(rows);  \
+    n = convert_index<BlasIndex>(cols);  \
 \
 /* Set lda, ldb, ldc */ \
-    lda = (MKL_INT)lhsStride; \
-    ldb = (MKL_INT)rhsStride; \
-    ldc = (MKL_INT)resStride; \
+    lda = convert_index<BlasIndex>(lhsStride); \
+    ldb = convert_index<BlasIndex>(rhsStride); \
+    ldc = convert_index<BlasIndex>(resStride); \
 \
 /* Set a, b, c */ \
     if (LhsStorageOrder==RowMajor) uplo='U'; \
@@ -83,16 +78,16 @@ struct product_selfadjoint_matrix<EIGTYPE,Index,LhsStorageOrder,true,ConjugateLh
       Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > rhs(_rhs,n,m,OuterStride<>(rhsStride)); \
       b_tmp = rhs.adjoint(); \
       b = b_tmp.data(); \
-      ldb = b_tmp.outerStride(); \
+      ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \
     } else b = _rhs; \
 \
-    MKLPREFIX##symm(&side, &uplo, &m, &n, &alpha_, (const MKLTYPE*)a, &lda, (const MKLTYPE*)b, &ldb, &beta_, (MKLTYPE*)res, &ldc); \
+    BLASPREFIX##symm_(&side, &uplo, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, &numext::real_ref(beta), (BLASTYPE*)res, &ldc); \
 \
   } \
 };
 
 
-#define EIGEN_MKL_HEMM_L(EIGTYPE, MKLTYPE, EIGPREFIX, MKLPREFIX) \
+#define EIGEN_BLAS_HEMM_L(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \
 template <typename Index, \
           int LhsStorageOrder, bool ConjugateLhs, \
           int RhsStorageOrder, bool ConjugateRhs> \
@@ -103,36 +98,31 @@ struct product_selfadjoint_matrix<EIGTYPE,Index,LhsStorageOrder,true,ConjugateLh
     const EIGTYPE* _lhs, Index lhsStride, \
     const EIGTYPE* _rhs, Index rhsStride, \
     EIGTYPE* res,        Index resStride, \
-    EIGTYPE alpha) \
+    EIGTYPE alpha, level3_blocking<EIGTYPE, EIGTYPE>& /*blocking*/) \
   { \
     char side='L', uplo='L'; \
-    MKL_INT m, n, lda, ldb, ldc; \
+    BlasIndex m, n, lda, ldb, ldc; \
     const EIGTYPE *a, *b; \
-    MKLTYPE alpha_, beta_; \
+    EIGTYPE beta(1); \
     MatrixX##EIGPREFIX b_tmp; \
     Matrix<EIGTYPE, Dynamic, Dynamic, LhsStorageOrder> a_tmp; \
-    EIGTYPE myone(1); \
 \
 /* Set transpose options */ \
 /* Set m, n, k */ \
-    m = (MKL_INT)rows; \
-    n = (MKL_INT)cols; \
-\
-/* Set alpha_ & beta_ */ \
-    assign_scalar_eig2mkl(alpha_, alpha); \
-    assign_scalar_eig2mkl(beta_, myone); \
+    m = convert_index<BlasIndex>(rows); \
+    n = convert_index<BlasIndex>(cols); \
 \
 /* Set lda, ldb, ldc */ \
-    lda = (MKL_INT)lhsStride; \
-    ldb = (MKL_INT)rhsStride; \
-    ldc = (MKL_INT)resStride; \
+    lda = convert_index<BlasIndex>(lhsStride); \
+    ldb = convert_index<BlasIndex>(rhsStride); \
+    ldc = convert_index<BlasIndex>(resStride); \
 \
 /* Set a, b, c */ \
     if (((LhsStorageOrder==ColMajor) && ConjugateLhs) || ((LhsStorageOrder==RowMajor) && (!ConjugateLhs))) { \
       Map<const Matrix<EIGTYPE, Dynamic, Dynamic, LhsStorageOrder>, 0, OuterStride<> > lhs(_lhs,m,m,OuterStride<>(lhsStride)); \
       a_tmp = lhs.conjugate(); \
       a = a_tmp.data(); \
-      lda = a_tmp.outerStride(); \
+      lda = convert_index<BlasIndex>(a_tmp.outerStride()); \
     } else a = _lhs; \
     if (LhsStorageOrder==RowMajor) uplo='U'; \
 \
@@ -151,23 +141,23 @@ struct product_selfadjoint_matrix<EIGTYPE,Index,LhsStorageOrder,true,ConjugateLh
         b_tmp = rhs.transpose(); \
       } \
       b = b_tmp.data(); \
-      ldb = b_tmp.outerStride(); \
+      ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \
     } \
 \
-    MKLPREFIX##hemm(&side, &uplo, &m, &n, &alpha_, (const MKLTYPE*)a, &lda, (const MKLTYPE*)b, &ldb, &beta_, (MKLTYPE*)res, &ldc); \
+    BLASPREFIX##hemm_(&side, &uplo, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, &numext::real_ref(beta), (BLASTYPE*)res, &ldc); \
 \
   } \
 };
 
-EIGEN_MKL_SYMM_L(double, double, d, d)
-EIGEN_MKL_SYMM_L(float, float, f, s)
-EIGEN_MKL_HEMM_L(dcomplex, MKL_Complex16, cd, z)
-EIGEN_MKL_HEMM_L(scomplex, MKL_Complex8, cf, c)
+EIGEN_BLAS_SYMM_L(double, double, d, d)
+EIGEN_BLAS_SYMM_L(float, float, f, s)
+EIGEN_BLAS_HEMM_L(dcomplex, double, cd, z)
+EIGEN_BLAS_HEMM_L(scomplex, float, cf, c)
 
 
 /* Optimized matrix * selfadjoint matrix (?SYMM/?HEMM) product */
 
-#define EIGEN_MKL_SYMM_R(EIGTYPE, MKLTYPE, EIGPREFIX, MKLPREFIX) \
+#define EIGEN_BLAS_SYMM_R(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \
 template <typename Index, \
           int LhsStorageOrder, bool ConjugateLhs, \
           int RhsStorageOrder, bool ConjugateRhs> \
@@ -179,27 +169,22 @@ struct product_selfadjoint_matrix<EIGTYPE,Index,LhsStorageOrder,false,ConjugateL
     const EIGTYPE* _lhs, Index lhsStride, \
     const EIGTYPE* _rhs, Index rhsStride, \
     EIGTYPE* res,        Index resStride, \
-    EIGTYPE alpha) \
+    EIGTYPE alpha, level3_blocking<EIGTYPE, EIGTYPE>& /*blocking*/) \
   { \
     char side='R', uplo='L'; \
-    MKL_INT m, n, lda, ldb, ldc; \
+    BlasIndex m, n, lda, ldb, ldc; \
     const EIGTYPE *a, *b; \
-    MKLTYPE alpha_, beta_; \
+    EIGTYPE beta(1); \
     MatrixX##EIGPREFIX b_tmp; \
-    EIGTYPE myone(1);\
 \
 /* Set m, n, k */ \
-    m = (MKL_INT)rows;  \
-    n = (MKL_INT)cols;  \
-\
-/* Set alpha_ & beta_ */ \
-    assign_scalar_eig2mkl(alpha_, alpha); \
-    assign_scalar_eig2mkl(beta_, myone); \
+    m = convert_index<BlasIndex>(rows);  \
+    n = convert_index<BlasIndex>(cols);  \
 \
 /* Set lda, ldb, ldc */ \
-    lda = (MKL_INT)rhsStride; \
-    ldb = (MKL_INT)lhsStride; \
-    ldc = (MKL_INT)resStride; \
+    lda = convert_index<BlasIndex>(rhsStride); \
+    ldb = convert_index<BlasIndex>(lhsStride); \
+    ldc = convert_index<BlasIndex>(resStride); \
 \
 /* Set a, b, c */ \
     if (RhsStorageOrder==RowMajor) uplo='U'; \
@@ -209,16 +194,16 @@ struct product_selfadjoint_matrix<EIGTYPE,Index,LhsStorageOrder,false,ConjugateL
       Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > lhs(_lhs,n,m,OuterStride<>(rhsStride)); \
       b_tmp = lhs.adjoint(); \
       b = b_tmp.data(); \
-      ldb = b_tmp.outerStride(); \
+      ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \
     } else b = _lhs; \
 \
-    MKLPREFIX##symm(&side, &uplo, &m, &n, &alpha_, (const MKLTYPE*)a, &lda, (const MKLTYPE*)b, &ldb, &beta_, (MKLTYPE*)res, &ldc); \
+    BLASPREFIX##symm_(&side, &uplo, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, &numext::real_ref(beta), (BLASTYPE*)res, &ldc); \
 \
   } \
 };
 
 
-#define EIGEN_MKL_HEMM_R(EIGTYPE, MKLTYPE, EIGPREFIX, MKLPREFIX) \
+#define EIGEN_BLAS_HEMM_R(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \
 template <typename Index, \
           int LhsStorageOrder, bool ConjugateLhs, \
           int RhsStorageOrder, bool ConjugateRhs> \
@@ -229,35 +214,30 @@ struct product_selfadjoint_matrix<EIGTYPE,Index,LhsStorageOrder,false,ConjugateL
     const EIGTYPE* _lhs, Index lhsStride, \
     const EIGTYPE* _rhs, Index rhsStride, \
     EIGTYPE* res,        Index resStride, \
-    EIGTYPE alpha) \
+    EIGTYPE alpha, level3_blocking<EIGTYPE, EIGTYPE>& /*blocking*/) \
   { \
     char side='R', uplo='L'; \
-    MKL_INT m, n, lda, ldb, ldc; \
+    BlasIndex m, n, lda, ldb, ldc; \
     const EIGTYPE *a, *b; \
-    MKLTYPE alpha_, beta_; \
+    EIGTYPE beta(1); \
     MatrixX##EIGPREFIX b_tmp; \
     Matrix<EIGTYPE, Dynamic, Dynamic, RhsStorageOrder> a_tmp; \
-    EIGTYPE myone(1); \
 \
 /* Set m, n, k */ \
-    m = (MKL_INT)rows; \
-    n = (MKL_INT)cols; \
-\
-/* Set alpha_ & beta_ */ \
-    assign_scalar_eig2mkl(alpha_, alpha); \
-    assign_scalar_eig2mkl(beta_, myone); \
+    m = convert_index<BlasIndex>(rows); \
+    n = convert_index<BlasIndex>(cols); \
 \
 /* Set lda, ldb, ldc */ \
-    lda = (MKL_INT)rhsStride; \
-    ldb = (MKL_INT)lhsStride; \
-    ldc = (MKL_INT)resStride; \
+    lda = convert_index<BlasIndex>(rhsStride); \
+    ldb = convert_index<BlasIndex>(lhsStride); \
+    ldc = convert_index<BlasIndex>(resStride); \
 \
 /* Set a, b, c */ \
     if (((RhsStorageOrder==ColMajor) && ConjugateRhs) || ((RhsStorageOrder==RowMajor) && (!ConjugateRhs))) { \
       Map<const Matrix<EIGTYPE, Dynamic, Dynamic, RhsStorageOrder>, 0, OuterStride<> > rhs(_rhs,n,n,OuterStride<>(rhsStride)); \
       a_tmp = rhs.conjugate(); \
       a = a_tmp.data(); \
-      lda = a_tmp.outerStride(); \
+      lda = convert_index<BlasIndex>(a_tmp.outerStride()); \
     } else a = _rhs; \
     if (RhsStorageOrder==RowMajor) uplo='U'; \
 \
@@ -276,20 +256,20 @@ struct product_selfadjoint_matrix<EIGTYPE,Index,LhsStorageOrder,false,ConjugateL
         b_tmp = lhs.transpose(); \
       } \
       b = b_tmp.data(); \
-      ldb = b_tmp.outerStride(); \
+      ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \
     } \
 \
-    MKLPREFIX##hemm(&side, &uplo, &m, &n, &alpha_, (const MKLTYPE*)a, &lda, (const MKLTYPE*)b, &ldb, &beta_, (MKLTYPE*)res, &ldc); \
+    BLASPREFIX##hemm_(&side, &uplo, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, &numext::real_ref(beta), (BLASTYPE*)res, &ldc); \
   } \
 };
 
-EIGEN_MKL_SYMM_R(double, double, d, d)
-EIGEN_MKL_SYMM_R(float, float, f, s)
-EIGEN_MKL_HEMM_R(dcomplex, MKL_Complex16, cd, z)
-EIGEN_MKL_HEMM_R(scomplex, MKL_Complex8, cf, c)
+EIGEN_BLAS_SYMM_R(double, double, d, d)
+EIGEN_BLAS_SYMM_R(float, float, f, s)
+EIGEN_BLAS_HEMM_R(dcomplex, double, cd, z)
+EIGEN_BLAS_HEMM_R(scomplex, float, cf, c)
 
 } // end namespace internal
 
 } // end namespace Eigen
 
-#endif // EIGEN_SELFADJOINT_MATRIX_MATRIX_MKL_H
+#endif // EIGEN_SELFADJOINT_MATRIX_MATRIX_BLAS_H
diff --git a/eigen/Eigen/src/Core/products/SelfadjointMatrixVector.h b/eigen/Eigen/src/Core/products/SelfadjointMatrixVector.h
index f698f67..3fd180e 100644
--- a/eigen/Eigen/src/Core/products/SelfadjointMatrixVector.h
+++ b/eigen/Eigen/src/Core/products/SelfadjointMatrixVector.h
@@ -30,7 +30,7 @@ struct selfadjoint_matrix_vector_product
 static EIGEN_DONT_INLINE void run(
   Index size,
   const Scalar*  lhs, Index lhsStride,
-  const Scalar* _rhs, Index rhsIncr,
+  const Scalar*  rhs,
   Scalar* res,
   Scalar alpha);
 };
@@ -39,11 +39,12 @@ template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool Conju
 EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs,Version>::run(
   Index size,
   const Scalar*  lhs, Index lhsStride,
-  const Scalar* _rhs, Index rhsIncr,
+  const Scalar*  rhs,
   Scalar* res,
   Scalar alpha)
 {
   typedef typename packet_traits<Scalar>::type Packet;
+  typedef typename NumTraits<Scalar>::Real RealScalar;
   const Index PacketSize = sizeof(Packet)/sizeof(Scalar);
 
   enum {
@@ -54,23 +55,13 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrd
 
   conj_helper<Scalar,Scalar,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs,  IsRowMajor), ConjugateRhs> cj0;
   conj_helper<Scalar,Scalar,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs, !IsRowMajor), ConjugateRhs> cj1;
-  conj_helper<Scalar,Scalar,NumTraits<Scalar>::IsComplex, ConjugateRhs> cjd;
+  conj_helper<RealScalar,Scalar,false, ConjugateRhs> cjd;
 
   conj_helper<Packet,Packet,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs,  IsRowMajor), ConjugateRhs> pcj0;
   conj_helper<Packet,Packet,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs, !IsRowMajor), ConjugateRhs> pcj1;
 
   Scalar cjAlpha = ConjugateRhs ? numext::conj(alpha) : alpha;
 
-  // FIXME this copy is now handled outside product_selfadjoint_vector, so it could probably be removed.
-  // if the rhs is not sequentially stored in memory we copy it to a temporary buffer,
-  // this is because we need to extract packets
-  ei_declare_aligned_stack_constructed_variable(Scalar,rhs,size,rhsIncr==1 ? const_cast<Scalar*>(_rhs) : 0);  
-  if (rhsIncr!=1)
-  {
-    const Scalar* it = _rhs;
-    for (Index i=0; i<size; ++i, it+=rhsIncr)
-      rhs[i] = *it;
-  }
 
   Index bound = (std::max)(Index(0),size-8) & 0xfffffffe;
   if (FirstTriangular)
@@ -92,12 +83,11 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrd
     Scalar t3(0);
     Packet ptmp3 = pset1<Packet>(t3);
 
-    size_t starti = FirstTriangular ? 0 : j+2;
-    size_t endi   = FirstTriangular ? j : size;
-    size_t alignedStart = (starti) + internal::first_aligned(&res[starti], endi-starti);
-    size_t alignedEnd = alignedStart + ((endi-alignedStart)/(PacketSize))*(PacketSize);
+    Index starti = FirstTriangular ? 0 : j+2;
+    Index endi   = FirstTriangular ? j : size;
+    Index alignedStart = (starti) + internal::first_default_aligned(&res[starti], endi-starti);
+    Index alignedEnd = alignedStart + ((endi-alignedStart)/(PacketSize))*(PacketSize);
 
-    // TODO make sure this product is a real * complex and that the rhs is properly conjugated if needed
     res[j]   += cjd.pmul(numext::real(A0[j]), t0);
     res[j+1] += cjd.pmul(numext::real(A1[j+1]), t1);
     if(FirstTriangular)
@@ -111,11 +101,11 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrd
       t2 += cj1.pmul(A0[j+1], rhs[j+1]);
     }
 
-    for (size_t i=starti; i<alignedStart; ++i)
+    for (Index i=starti; i<alignedStart; ++i)
     {
-      res[i] += t0 * A0[i] + t1 * A1[i];
-      t2 += numext::conj(A0[i]) * rhs[i];
-      t3 += numext::conj(A1[i]) * rhs[i];
+      res[i] += cj0.pmul(A0[i], t0) + cj0.pmul(A1[i],t1);
+      t2 += cj1.pmul(A0[i], rhs[i]);
+      t3 += cj1.pmul(A1[i], rhs[i]);
     }
     // Yes this an optimization for gcc 4.3 and 4.4 (=> huge speed up)
     // gcc 4.2 does this optimization automatically.
@@ -123,7 +113,7 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrd
     const Scalar* EIGEN_RESTRICT a1It  = A1  + alignedStart;
     const Scalar* EIGEN_RESTRICT rhsIt = rhs + alignedStart;
           Scalar* EIGEN_RESTRICT resIt = res + alignedStart;
-    for (size_t i=alignedStart; i<alignedEnd; i+=PacketSize)
+    for (Index i=alignedStart; i<alignedEnd; i+=PacketSize)
     {
       Packet A0i = ploadu<Packet>(a0It);  a0It  += PacketSize;
       Packet A1i = ploadu<Packet>(a1It);  a1It  += PacketSize;
@@ -135,7 +125,7 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrd
       ptmp3 = pcj1.pmadd(A1i,  Bi, ptmp3);
       pstore(resIt,Xi); resIt += PacketSize;
     }
-    for (size_t i=alignedEnd; i<endi; i++)
+    for (Index i=alignedEnd; i<endi; i++)
     {
       res[i] += cj0.pmul(A0[i], t0) + cj0.pmul(A1[i],t1);
       t2 += cj1.pmul(A0[i], rhs[i]);
@@ -151,7 +141,6 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrd
 
     Scalar t1 = cjAlpha * rhs[j];
     Scalar t2(0);
-    // TODO make sure this product is a real * complex and that the rhs is properly conjugated if needed
     res[j] += cjd.pmul(numext::real(A0[j]), t1);
     for (Index i=FirstTriangular ? 0 : j+1; i<(FirstTriangular ? j : size); i++)
     {
@@ -169,45 +158,44 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrd
 ***************************************************************************/
 
 namespace internal {
-template<typename Lhs, int LhsMode, typename Rhs>
-struct traits<SelfadjointProductMatrix<Lhs,LhsMode,false,Rhs,0,true> >
-  : traits<ProductBase<SelfadjointProductMatrix<Lhs,LhsMode,false,Rhs,0,true>, Lhs, Rhs> >
-{};
-}
 
 template<typename Lhs, int LhsMode, typename Rhs>
-struct SelfadjointProductMatrix<Lhs,LhsMode,false,Rhs,0,true>
-  : public ProductBase<SelfadjointProductMatrix<Lhs,LhsMode,false,Rhs,0,true>, Lhs, Rhs >
+struct selfadjoint_product_impl<Lhs,LhsMode,false,Rhs,0,true>
 {
-  EIGEN_PRODUCT_PUBLIC_INTERFACE(SelfadjointProductMatrix)
-
-  enum {
-    LhsUpLo = LhsMode&(Upper|Lower)
-  };
-
-  SelfadjointProductMatrix(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) {}
-
-  template<typename Dest> void scaleAndAddTo(Dest& dest, const Scalar& alpha) const
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  
+  typedef internal::blas_traits<Lhs> LhsBlasTraits;
+  typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
+  typedef typename internal::remove_all<ActualLhsType>::type ActualLhsTypeCleaned;
+  
+  typedef internal::blas_traits<Rhs> RhsBlasTraits;
+  typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
+  typedef typename internal::remove_all<ActualRhsType>::type ActualRhsTypeCleaned;
+
+  enum { LhsUpLo = LhsMode&(Upper|Lower) };
+
+  template<typename Dest>
+  static void run(Dest& dest, const Lhs &a_lhs, const Rhs &a_rhs, const Scalar& alpha)
   {
     typedef typename Dest::Scalar ResScalar;
-    typedef typename Base::RhsScalar RhsScalar;
-    typedef Map<Matrix<ResScalar,Dynamic,1>, Aligned> MappedDest;
+    typedef typename Rhs::Scalar RhsScalar;
+    typedef Map<Matrix<ResScalar,Dynamic,1>, EIGEN_PLAIN_ENUM_MIN(AlignedMax,internal::packet_traits<ResScalar>::size)> MappedDest;
     
-    eigen_assert(dest.rows()==m_lhs.rows() && dest.cols()==m_rhs.cols());
+    eigen_assert(dest.rows()==a_lhs.rows() && dest.cols()==a_rhs.cols());
 
-    typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(m_lhs);
-    typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(m_rhs);
+    typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(a_lhs);
+    typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(a_rhs);
 
-    Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(m_lhs)
-                               * RhsBlasTraits::extractScalarFactor(m_rhs);
+    Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(a_lhs)
+                               * RhsBlasTraits::extractScalarFactor(a_rhs);
 
     enum {
       EvalToDest = (Dest::InnerStrideAtCompileTime==1),
-      UseRhs = (_ActualRhsType::InnerStrideAtCompileTime==1)
+      UseRhs = (ActualRhsTypeCleaned::InnerStrideAtCompileTime==1)
     };
     
     internal::gemv_static_vector_if<ResScalar,Dest::SizeAtCompileTime,Dest::MaxSizeAtCompileTime,!EvalToDest> static_dest;
-    internal::gemv_static_vector_if<RhsScalar,_ActualRhsType::SizeAtCompileTime,_ActualRhsType::MaxSizeAtCompileTime,!UseRhs> static_rhs;
+    internal::gemv_static_vector_if<RhsScalar,ActualRhsTypeCleaned::SizeAtCompileTime,ActualRhsTypeCleaned::MaxSizeAtCompileTime,!UseRhs> static_rhs;
 
     ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(),
                                                   EvalToDest ? dest.data() : static_dest.data());
@@ -218,7 +206,7 @@ struct SelfadjointProductMatrix<Lhs,LhsMode,false,Rhs,0,true>
     if(!EvalToDest)
     {
       #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
-      int size = dest.size();
+      Index size = dest.size();
       EIGEN_DENSE_STORAGE_CTOR_PLUGIN
       #endif
       MappedDest(actualDestPtr, dest.size()) = dest;
@@ -227,18 +215,19 @@ struct SelfadjointProductMatrix<Lhs,LhsMode,false,Rhs,0,true>
     if(!UseRhs)
     {
       #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
-      int size = rhs.size();
+      Index size = rhs.size();
       EIGEN_DENSE_STORAGE_CTOR_PLUGIN
       #endif
-      Map<typename _ActualRhsType::PlainObject>(actualRhsPtr, rhs.size()) = rhs;
+      Map<typename ActualRhsTypeCleaned::PlainObject>(actualRhsPtr, rhs.size()) = rhs;
     }
       
       
-    internal::selfadjoint_matrix_vector_product<Scalar, Index, (internal::traits<_ActualLhsType>::Flags&RowMajorBit) ? RowMajor : ColMajor, int(LhsUpLo), bool(LhsBlasTraits::NeedToConjugate), bool(RhsBlasTraits::NeedToConjugate)>::run
+    internal::selfadjoint_matrix_vector_product<Scalar, Index, (internal::traits<ActualLhsTypeCleaned>::Flags&RowMajorBit) ? RowMajor : ColMajor,
+                                                int(LhsUpLo), bool(LhsBlasTraits::NeedToConjugate), bool(RhsBlasTraits::NeedToConjugate)>::run
       (
         lhs.rows(),                             // size
         &lhs.coeffRef(0,0),  lhs.outerStride(), // lhs info
-        actualRhsPtr, 1,                        // rhs info
+        actualRhsPtr,                           // rhs info
         actualDestPtr,                          // result info
         actualAlpha                             // scale factor
       );
@@ -248,34 +237,24 @@ struct SelfadjointProductMatrix<Lhs,LhsMode,false,Rhs,0,true>
   }
 };
 
-namespace internal {
-template<typename Lhs, typename Rhs, int RhsMode>
-struct traits<SelfadjointProductMatrix<Lhs,0,true,Rhs,RhsMode,false> >
-  : traits<ProductBase<SelfadjointProductMatrix<Lhs,0,true,Rhs,RhsMode,false>, Lhs, Rhs> >
-{};
-}
-
 template<typename Lhs, typename Rhs, int RhsMode>
-struct SelfadjointProductMatrix<Lhs,0,true,Rhs,RhsMode,false>
-  : public ProductBase<SelfadjointProductMatrix<Lhs,0,true,Rhs,RhsMode,false>, Lhs, Rhs >
+struct selfadjoint_product_impl<Lhs,0,true,Rhs,RhsMode,false>
 {
-  EIGEN_PRODUCT_PUBLIC_INTERFACE(SelfadjointProductMatrix)
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  enum { RhsUpLo = RhsMode&(Upper|Lower)  };
 
-  enum {
-    RhsUpLo = RhsMode&(Upper|Lower)
-  };
-
-  SelfadjointProductMatrix(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) {}
-
-  template<typename Dest> void scaleAndAddTo(Dest& dest, const Scalar& alpha) const
+  template<typename Dest>
+  static void run(Dest& dest, const Lhs &a_lhs, const Rhs &a_rhs, const Scalar& alpha)
   {
     // let's simply transpose the product
     Transpose<Dest> destT(dest);
-    SelfadjointProductMatrix<Transpose<const Rhs>, int(RhsUpLo)==Upper ? Lower : Upper, false,
-                             Transpose<const Lhs>, 0, true>(m_rhs.transpose(), m_lhs.transpose()).scaleAndAddTo(destT, alpha);
+    selfadjoint_product_impl<Transpose<const Rhs>, int(RhsUpLo)==Upper ? Lower : Upper, false,
+                             Transpose<const Lhs>, 0, true>::run(destT, a_rhs.transpose(), a_lhs.transpose(), alpha);
   }
 };
 
+} // end namespace internal
+
 } // end namespace Eigen
 
 #endif // EIGEN_SELFADJOINT_MATRIX_VECTOR_H
diff --git a/eigen/Eigen/src/Core/products/SelfadjointMatrixVector_MKL.h b/eigen/Eigen/src/Core/products/SelfadjointMatrixVector_BLAS.h
index 86684b6..38f23ac 100644
--- a/eigen/Eigen/src/Core/products/SelfadjointMatrixVector_MKL.h
+++ b/eigen/Eigen/src/Core/products/SelfadjointMatrixVector_BLAS.h
@@ -25,13 +25,13 @@
  SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
  ********************************************************************************
- *   Content : Eigen bindings to Intel(R) MKL
+ *   Content : Eigen bindings to BLAS F77
  *   Selfadjoint matrix-vector product functionality based on ?SYMV/HEMV.
  ********************************************************************************
 */
 
-#ifndef EIGEN_SELFADJOINT_MATRIX_VECTOR_MKL_H
-#define EIGEN_SELFADJOINT_MATRIX_VECTOR_MKL_H
+#ifndef EIGEN_SELFADJOINT_MATRIX_VECTOR_BLAS_H
+#define EIGEN_SELFADJOINT_MATRIX_VECTOR_BLAS_H
 
 namespace Eigen { 
 
@@ -47,31 +47,31 @@ template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool Conju
 struct selfadjoint_matrix_vector_product_symv :
   selfadjoint_matrix_vector_product<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs,BuiltIn> {};
 
-#define EIGEN_MKL_SYMV_SPECIALIZE(Scalar) \
+#define EIGEN_BLAS_SYMV_SPECIALIZE(Scalar) \
 template<typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs> \
 struct selfadjoint_matrix_vector_product<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs,Specialized> { \
 static void run( \
   Index size, const Scalar*  lhs, Index lhsStride, \
-  const Scalar* _rhs, Index rhsIncr, Scalar* res, Scalar alpha) { \
+  const Scalar* _rhs, Scalar* res, Scalar alpha) { \
     enum {\
       IsColMajor = StorageOrder==ColMajor \
     }; \
     if (IsColMajor == ConjugateLhs) {\
       selfadjoint_matrix_vector_product<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs,BuiltIn>::run( \
-        size, lhs, lhsStride, _rhs, rhsIncr, res, alpha);  \
+        size, lhs, lhsStride, _rhs, res, alpha);  \
     } else {\
       selfadjoint_matrix_vector_product_symv<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs>::run( \
-        size, lhs, lhsStride, _rhs, rhsIncr, res, alpha);  \
+        size, lhs, lhsStride, _rhs, res, alpha);  \
     }\
   } \
 }; \
 
-EIGEN_MKL_SYMV_SPECIALIZE(double)
-EIGEN_MKL_SYMV_SPECIALIZE(float)
-EIGEN_MKL_SYMV_SPECIALIZE(dcomplex)
-EIGEN_MKL_SYMV_SPECIALIZE(scomplex)
+EIGEN_BLAS_SYMV_SPECIALIZE(double)
+EIGEN_BLAS_SYMV_SPECIALIZE(float)
+EIGEN_BLAS_SYMV_SPECIALIZE(dcomplex)
+EIGEN_BLAS_SYMV_SPECIALIZE(scomplex)
 
-#define EIGEN_MKL_SYMV_SPECIALIZATION(EIGTYPE,MKLTYPE,MKLFUNC) \
+#define EIGEN_BLAS_SYMV_SPECIALIZATION(EIGTYPE,BLASTYPE,BLASFUNC) \
 template<typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs> \
 struct selfadjoint_matrix_vector_product_symv<EIGTYPE,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs> \
 { \
@@ -79,36 +79,33 @@ typedef Matrix<EIGTYPE,Dynamic,1,ColMajor> SYMVVector;\
 \
 static void run( \
 Index size, const EIGTYPE*  lhs, Index lhsStride, \
-const EIGTYPE* _rhs, Index rhsIncr, EIGTYPE* res, EIGTYPE alpha) \
+const EIGTYPE* _rhs, EIGTYPE* res, EIGTYPE alpha) \
 { \
   enum {\
     IsRowMajor = StorageOrder==RowMajor ? 1 : 0, \
     IsLower = UpLo == Lower ? 1 : 0 \
   }; \
-  MKL_INT n=size, lda=lhsStride, incx=rhsIncr, incy=1; \
-  MKLTYPE alpha_, beta_; \
-  const EIGTYPE *x_ptr, myone(1); \
+  BlasIndex n=convert_index<BlasIndex>(size), lda=convert_index<BlasIndex>(lhsStride), incx=1, incy=1; \
+  EIGTYPE beta(1); \
+  const EIGTYPE *x_ptr; \
   char uplo=(IsRowMajor) ? (IsLower ? 'U' : 'L') : (IsLower ? 'L' : 'U'); \
-  assign_scalar_eig2mkl(alpha_, alpha); \
-  assign_scalar_eig2mkl(beta_, myone); \
   SYMVVector x_tmp; \
   if (ConjugateRhs) { \
-    Map<const SYMVVector, 0, InnerStride<> > map_x(_rhs,size,1,InnerStride<>(incx)); \
+    Map<const SYMVVector, 0 > map_x(_rhs,size,1); \
     x_tmp=map_x.conjugate(); \
     x_ptr=x_tmp.data(); \
-    incx=1; \
   } else x_ptr=_rhs; \
-  MKLFUNC(&uplo, &n, &alpha_, (const MKLTYPE*)lhs, &lda, (const MKLTYPE*)x_ptr, &incx, &beta_, (MKLTYPE*)res, &incy); \
+  BLASFUNC(&uplo, &n, &numext::real_ref(alpha), (const BLASTYPE*)lhs, &lda, (const BLASTYPE*)x_ptr, &incx, &numext::real_ref(beta), (BLASTYPE*)res, &incy); \
 }\
 };
 
-EIGEN_MKL_SYMV_SPECIALIZATION(double,   double,        dsymv)
-EIGEN_MKL_SYMV_SPECIALIZATION(float,    float,         ssymv)
-EIGEN_MKL_SYMV_SPECIALIZATION(dcomplex, MKL_Complex16, zhemv)
-EIGEN_MKL_SYMV_SPECIALIZATION(scomplex, MKL_Complex8,  chemv)
+EIGEN_BLAS_SYMV_SPECIALIZATION(double,   double, dsymv_)
+EIGEN_BLAS_SYMV_SPECIALIZATION(float,    float,  ssymv_)
+EIGEN_BLAS_SYMV_SPECIALIZATION(dcomplex, double, zhemv_)
+EIGEN_BLAS_SYMV_SPECIALIZATION(scomplex, float,  chemv_)
 
 } // end namespace internal
 
 } // end namespace Eigen
 
-#endif // EIGEN_SELFADJOINT_MATRIX_VECTOR_MKL_H
+#endif // EIGEN_SELFADJOINT_MATRIX_VECTOR_BLAS_H
diff --git a/eigen/Eigen/src/Core/products/SelfadjointProduct.h b/eigen/Eigen/src/Core/products/SelfadjointProduct.h
index 6ca4ae6..39c5b59 100644
--- a/eigen/Eigen/src/Core/products/SelfadjointProduct.h
+++ b/eigen/Eigen/src/Core/products/SelfadjointProduct.h
@@ -53,7 +53,6 @@ struct selfadjoint_product_selector<MatrixType,OtherType,UpLo,true>
   static void run(MatrixType& mat, const OtherType& other, const typename MatrixType::Scalar& alpha)
   {
     typedef typename MatrixType::Scalar Scalar;
-    typedef typename MatrixType::Index Index;
     typedef internal::blas_traits<OtherType> OtherBlasTraits;
     typedef typename OtherBlasTraits::DirectLinearAccessType ActualOtherType;
     typedef typename internal::remove_all<ActualOtherType>::type _ActualOtherType;
@@ -86,7 +85,6 @@ struct selfadjoint_product_selector<MatrixType,OtherType,UpLo,false>
   static void run(MatrixType& mat, const OtherType& other, const typename MatrixType::Scalar& alpha)
   {
     typedef typename MatrixType::Scalar Scalar;
-    typedef typename MatrixType::Index Index;
     typedef internal::blas_traits<OtherType> OtherBlasTraits;
     typedef typename OtherBlasTraits::DirectLinearAccessType ActualOtherType;
     typedef typename internal::remove_all<ActualOtherType>::type _ActualOtherType;
@@ -94,15 +92,27 @@ struct selfadjoint_product_selector<MatrixType,OtherType,UpLo,false>
 
     Scalar actualAlpha = alpha * OtherBlasTraits::extractScalarFactor(other.derived());
 
-    enum { IsRowMajor = (internal::traits<MatrixType>::Flags&RowMajorBit) ? 1 : 0 };
+    enum {
+      IsRowMajor = (internal::traits<MatrixType>::Flags&RowMajorBit) ? 1 : 0,
+      OtherIsRowMajor = _ActualOtherType::Flags&RowMajorBit ? 1 : 0
+    };
+
+    Index size = mat.cols();
+    Index depth = actualOther.cols();
+
+    typedef internal::gemm_blocking_space<IsRowMajor ? RowMajor : ColMajor,Scalar,Scalar,
+              MatrixType::MaxColsAtCompileTime, MatrixType::MaxColsAtCompileTime, _ActualOtherType::MaxColsAtCompileTime> BlockingType;
+
+    BlockingType blocking(size, size, depth, 1, false);
+
 
     internal::general_matrix_matrix_triangular_product<Index,
-      Scalar, _ActualOtherType::Flags&RowMajorBit ? RowMajor : ColMajor,   OtherBlasTraits::NeedToConjugate  && NumTraits<Scalar>::IsComplex,
-      Scalar, _ActualOtherType::Flags&RowMajorBit ? ColMajor : RowMajor, (!OtherBlasTraits::NeedToConjugate) && NumTraits<Scalar>::IsComplex,
-      MatrixType::Flags&RowMajorBit ? RowMajor : ColMajor, UpLo>
-      ::run(mat.cols(), actualOther.cols(),
+      Scalar, OtherIsRowMajor ? RowMajor : ColMajor,   OtherBlasTraits::NeedToConjugate  && NumTraits<Scalar>::IsComplex,
+      Scalar, OtherIsRowMajor ? ColMajor : RowMajor, (!OtherBlasTraits::NeedToConjugate) && NumTraits<Scalar>::IsComplex,
+      IsRowMajor ? RowMajor : ColMajor, UpLo>
+      ::run(size, depth,
             &actualOther.coeffRef(0,0), actualOther.outerStride(), &actualOther.coeffRef(0,0), actualOther.outerStride(),
-            mat.data(), mat.outerStride(), actualAlpha);
+            mat.data(), mat.outerStride(), actualAlpha, blocking);
   }
 };
 
@@ -110,7 +120,7 @@ struct selfadjoint_product_selector<MatrixType,OtherType,UpLo,false>
 
 template<typename MatrixType, unsigned int UpLo>
 template<typename DerivedU>
-SelfAdjointView<MatrixType,UpLo>& SelfAdjointView<MatrixType,UpLo>
+EIGEN_DEVICE_FUNC SelfAdjointView<MatrixType,UpLo>& SelfAdjointView<MatrixType,UpLo>
 ::rankUpdate(const MatrixBase<DerivedU>& u, const Scalar& alpha)
 {
   selfadjoint_product_selector<MatrixType,DerivedU,UpLo>::run(_expression().const_cast_derived(), u.derived(), alpha);
diff --git a/eigen/Eigen/src/Core/products/SelfadjointRank2Update.h b/eigen/Eigen/src/Core/products/SelfadjointRank2Update.h
index 8594a97..d395888 100644
--- a/eigen/Eigen/src/Core/products/SelfadjointRank2Update.h
+++ b/eigen/Eigen/src/Core/products/SelfadjointRank2Update.h
@@ -57,7 +57,7 @@ template<bool Cond, typename T> struct conj_expr_if
 
 template<typename MatrixType, unsigned int UpLo>
 template<typename DerivedU, typename DerivedV>
-SelfAdjointView<MatrixType,UpLo>& SelfAdjointView<MatrixType,UpLo>
+EIGEN_DEVICE_FUNC SelfAdjointView<MatrixType,UpLo>& SelfAdjointView<MatrixType,UpLo>
 ::rankUpdate(const MatrixBase<DerivedU>& u, const MatrixBase<DerivedV>& v, const Scalar& alpha)
 {
   typedef internal::blas_traits<DerivedU> UBlasTraits;
@@ -79,11 +79,11 @@ SelfAdjointView<MatrixType,UpLo>& SelfAdjointView<MatrixType,UpLo>
   if (IsRowMajor)
     actualAlpha = numext::conj(actualAlpha);
 
-  internal::selfadjoint_rank2_update_selector<Scalar, Index,
-    typename internal::remove_all<typename internal::conj_expr_if<IsRowMajor ^ UBlasTraits::NeedToConjugate,_ActualUType>::type>::type,
-    typename internal::remove_all<typename internal::conj_expr_if<IsRowMajor ^ VBlasTraits::NeedToConjugate,_ActualVType>::type>::type,
+  typedef typename internal::remove_all<typename internal::conj_expr_if<IsRowMajor ^ UBlasTraits::NeedToConjugate,_ActualUType>::type>::type UType;
+  typedef typename internal::remove_all<typename internal::conj_expr_if<IsRowMajor ^ VBlasTraits::NeedToConjugate,_ActualVType>::type>::type VType;
+  internal::selfadjoint_rank2_update_selector<Scalar, Index, UType, VType,
     (IsRowMajor ? int(UpLo==Upper ? Lower : Upper) : UpLo)>
-    ::run(_expression().const_cast_derived().data(),_expression().outerStride(),actualU,actualV,actualAlpha);
+    ::run(_expression().const_cast_derived().data(),_expression().outerStride(),UType(actualU),VType(actualV),actualAlpha);
 
   return *this;
 }
diff --git a/eigen/Eigen/src/Core/products/TriangularMatrixMatrix.h b/eigen/Eigen/src/Core/products/TriangularMatrixMatrix.h
index 8110507..6ec5a8a 100644
--- a/eigen/Eigen/src/Core/products/TriangularMatrixMatrix.h
+++ b/eigen/Eigen/src/Core/products/TriangularMatrixMatrix.h
@@ -108,7 +108,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
     Index _rows, Index _cols, Index _depth,
     const Scalar* _lhs, Index lhsStride,
     const Scalar* _rhs, Index rhsStride,
-    Scalar* res,        Index resStride,
+    Scalar* _res,        Index resStride,
     const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking)
   {
     // strip zeros
@@ -117,30 +117,36 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
     Index depth     = IsLower ? diagSize : _depth;
     Index cols      = _cols;
     
-    const_blas_data_mapper<Scalar, Index, LhsStorageOrder> lhs(_lhs,lhsStride);
-    const_blas_data_mapper<Scalar, Index, RhsStorageOrder> rhs(_rhs,rhsStride);
+    typedef const_blas_data_mapper<Scalar, Index, LhsStorageOrder> LhsMapper;
+    typedef const_blas_data_mapper<Scalar, Index, RhsStorageOrder> RhsMapper;
+    typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper;
+    LhsMapper lhs(_lhs,lhsStride);
+    RhsMapper rhs(_rhs,rhsStride);
+    ResMapper res(_res, resStride);
 
     Index kc = blocking.kc();                   // cache block size along the K direction
     Index mc = (std::min)(rows,blocking.mc());  // cache block size along the M direction
+    // The small panel size must not be larger than blocking size.
+    // Usually this should never be the case because SmallPanelWidth^2 is very small
+    // compared to L2 cache size, but let's be safe:
+    Index panelWidth = (std::min)(Index(SmallPanelWidth),(std::min)(kc,mc));
 
     std::size_t sizeA = kc*mc;
     std::size_t sizeB = kc*cols;
-    std::size_t sizeW = kc*Traits::WorkSpaceFactor;
 
     ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA());
     ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
-    ei_declare_aligned_stack_constructed_variable(Scalar, blockW, sizeW, blocking.blockW());
 
-    Matrix<Scalar,SmallPanelWidth,SmallPanelWidth,LhsStorageOrder> triangularBuffer;
+    Matrix<Scalar,SmallPanelWidth,SmallPanelWidth,LhsStorageOrder> triangularBuffer((internal::constructor_without_unaligned_array_assert()));
     triangularBuffer.setZero();
     if((Mode&ZeroDiag)==ZeroDiag)
       triangularBuffer.diagonal().setZero();
     else
       triangularBuffer.diagonal().setOnes();
 
-    gebp_kernel<Scalar, Scalar, Index, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
-    gemm_pack_lhs<Scalar, Index, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
-    gemm_pack_rhs<Scalar, Index, Traits::nr,RhsStorageOrder> pack_rhs;
+    gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
+    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+    gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder> pack_rhs;
 
     for(Index k2=IsLower ? depth : 0;
         IsLower ? k2>0 : k2<depth;
@@ -156,7 +162,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
         k2 = k2+actual_kc-kc;
       }
 
-      pack_rhs(blockB, &rhs(actual_k2,0), rhsStride, actual_kc, cols);
+      pack_rhs(blockB, rhs.getSubMapper(actual_k2,0), actual_kc, cols);
 
       // the selected lhs's panel has to be split in three different parts:
       //  1 - the part which is zero => skip it
@@ -167,9 +173,9 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
       if(IsLower || actual_k2<rows)
       {
         // for each small vertical panels of lhs
-        for (Index k1=0; k1<actual_kc; k1+=SmallPanelWidth)
+        for (Index k1=0; k1<actual_kc; k1+=panelWidth)
         {
-          Index actualPanelWidth = std::min<Index>(actual_kc-k1, SmallPanelWidth);
+          Index actualPanelWidth = std::min<Index>(actual_kc-k1, panelWidth);
           Index lengthTarget = IsLower ? actual_kc-k1-actualPanelWidth : k1;
           Index startBlock   = actual_k2+k1;
           Index blockBOffset = k1;
@@ -184,20 +190,22 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
             for (Index i=IsLower ? k+1 : 0; IsLower ? i<actualPanelWidth : i<k; ++i)
               triangularBuffer.coeffRef(i,k) = lhs(startBlock+i,startBlock+k);
           }
-          pack_lhs(blockA, triangularBuffer.data(), triangularBuffer.outerStride(), actualPanelWidth, actualPanelWidth);
+          pack_lhs(blockA, LhsMapper(triangularBuffer.data(), triangularBuffer.outerStride()), actualPanelWidth, actualPanelWidth);
 
-          gebp_kernel(res+startBlock, resStride, blockA, blockB, actualPanelWidth, actualPanelWidth, cols, alpha,
-                      actualPanelWidth, actual_kc, 0, blockBOffset, blockW);
+          gebp_kernel(res.getSubMapper(startBlock, 0), blockA, blockB,
+                      actualPanelWidth, actualPanelWidth, cols, alpha,
+                      actualPanelWidth, actual_kc, 0, blockBOffset);
 
           // GEBP with remaining micro panel
           if (lengthTarget>0)
           {
             Index startTarget  = IsLower ? actual_k2+k1+actualPanelWidth : actual_k2;
 
-            pack_lhs(blockA, &lhs(startTarget,startBlock), lhsStride, actualPanelWidth, lengthTarget);
+            pack_lhs(blockA, lhs.getSubMapper(startTarget,startBlock), actualPanelWidth, lengthTarget);
 
-            gebp_kernel(res+startTarget, resStride, blockA, blockB, lengthTarget, actualPanelWidth, cols, alpha,
-                        actualPanelWidth, actual_kc, 0, blockBOffset, blockW);
+            gebp_kernel(res.getSubMapper(startTarget, 0), blockA, blockB,
+                        lengthTarget, actualPanelWidth, cols, alpha,
+                        actualPanelWidth, actual_kc, 0, blockBOffset);
           }
         }
       }
@@ -208,10 +216,11 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
         for(Index i2=start; i2<end; i2+=mc)
         {
           const Index actual_mc = (std::min)(i2+mc,end)-i2;
-          gemm_pack_lhs<Scalar, Index, Traits::mr,Traits::LhsProgress, LhsStorageOrder,false>()
-            (blockA, &lhs(i2, actual_k2), lhsStride, actual_kc, actual_mc);
+          gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr,Traits::LhsProgress, LhsStorageOrder,false>()
+            (blockA, lhs.getSubMapper(i2, actual_k2), actual_kc, actual_mc);
 
-          gebp_kernel(res+i2, resStride, blockA, blockB, actual_mc, actual_kc, cols, alpha, -1, -1, 0, 0, blockW);
+          gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc,
+                      actual_kc, cols, alpha, -1, -1, 0, 0);
         }
       }
     }
@@ -249,40 +258,43 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
     Index _rows, Index _cols, Index _depth,
     const Scalar* _lhs, Index lhsStride,
     const Scalar* _rhs, Index rhsStride,
-    Scalar* res,        Index resStride,
+    Scalar* _res,        Index resStride,
     const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking)
   {
+    const Index PacketBytes = packet_traits<Scalar>::size*sizeof(Scalar);
     // strip zeros
     Index diagSize  = (std::min)(_cols,_depth);
     Index rows      = _rows;
     Index depth     = IsLower ? _depth : diagSize;
     Index cols      = IsLower ? diagSize : _cols;
     
-    const_blas_data_mapper<Scalar, Index, LhsStorageOrder> lhs(_lhs,lhsStride);
-    const_blas_data_mapper<Scalar, Index, RhsStorageOrder> rhs(_rhs,rhsStride);
+    typedef const_blas_data_mapper<Scalar, Index, LhsStorageOrder> LhsMapper;
+    typedef const_blas_data_mapper<Scalar, Index, RhsStorageOrder> RhsMapper;
+    typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper;
+    LhsMapper lhs(_lhs,lhsStride);
+    RhsMapper rhs(_rhs,rhsStride);
+    ResMapper res(_res, resStride);
 
     Index kc = blocking.kc();                   // cache block size along the K direction
     Index mc = (std::min)(rows,blocking.mc());  // cache block size along the M direction
 
     std::size_t sizeA = kc*mc;
-    std::size_t sizeB = kc*cols;
-    std::size_t sizeW = kc*Traits::WorkSpaceFactor;
+    std::size_t sizeB = kc*cols+EIGEN_MAX_ALIGN_BYTES/sizeof(Scalar);
 
     ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA());
     ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
-    ei_declare_aligned_stack_constructed_variable(Scalar, blockW, sizeW, blocking.blockW());
 
-    Matrix<Scalar,SmallPanelWidth,SmallPanelWidth,RhsStorageOrder> triangularBuffer;
+    Matrix<Scalar,SmallPanelWidth,SmallPanelWidth,RhsStorageOrder> triangularBuffer((internal::constructor_without_unaligned_array_assert()));
     triangularBuffer.setZero();
     if((Mode&ZeroDiag)==ZeroDiag)
       triangularBuffer.diagonal().setZero();
     else
       triangularBuffer.diagonal().setOnes();
 
-    gebp_kernel<Scalar, Scalar, Index, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
-    gemm_pack_lhs<Scalar, Index, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
-    gemm_pack_rhs<Scalar, Index, Traits::nr,RhsStorageOrder> pack_rhs;
-    gemm_pack_rhs<Scalar, Index, Traits::nr,RhsStorageOrder,false,true> pack_rhs_panel;
+    gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
+    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+    gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder> pack_rhs;
+    gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder,false,true> pack_rhs_panel;
 
     for(Index k2=IsLower ? 0 : depth;
         IsLower ? k2<depth  : k2>0;
@@ -304,8 +316,9 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
       Index ts = (IsLower && actual_k2>=cols) ? 0 : actual_kc;
 
       Scalar* geb = blockB+ts*ts;
+      geb = geb + internal::first_aligned<PacketBytes>(geb,PacketBytes/sizeof(Scalar));
 
-      pack_rhs(geb, &rhs(actual_k2,IsLower ? 0 : k2), rhsStride, actual_kc, rs);
+      pack_rhs(geb, rhs.getSubMapper(actual_k2,IsLower ? 0 : k2), actual_kc, rs);
 
       // pack the triangular part of the rhs padding the unrolled blocks with zeros
       if(ts>0)
@@ -318,7 +331,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
           Index panelLength = IsLower ? actual_kc-j2-actualPanelWidth : j2;
           // general part
           pack_rhs_panel(blockB+j2*actual_kc,
-                         &rhs(actual_k2+panelOffset, actual_j2), rhsStride,
+                         rhs.getSubMapper(actual_k2+panelOffset, actual_j2),
                          panelLength, actualPanelWidth,
                          actual_kc, panelOffset);
 
@@ -332,7 +345,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
           }
 
           pack_rhs_panel(blockB+j2*actual_kc,
-                         triangularBuffer.data(), triangularBuffer.outerStride(),
+                         RhsMapper(triangularBuffer.data(), triangularBuffer.outerStride()),
                          actualPanelWidth, actualPanelWidth,
                          actual_kc, j2);
         }
@@ -341,7 +354,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
       for (Index i2=0; i2<rows; i2+=mc)
       {
         const Index actual_mc = (std::min)(mc,rows-i2);
-        pack_lhs(blockA, &lhs(i2, actual_k2), lhsStride, actual_kc, actual_mc);
+        pack_lhs(blockA, lhs.getSubMapper(i2, actual_k2), actual_kc, actual_mc);
 
         // triangular kernel
         if(ts>0)
@@ -352,19 +365,18 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
             Index panelLength = IsLower ? actual_kc-j2 : j2+actualPanelWidth;
             Index blockOffset = IsLower ? j2 : 0;
 
-            gebp_kernel(res+i2+(actual_k2+j2)*resStride, resStride,
+            gebp_kernel(res.getSubMapper(i2, actual_k2 + j2),
                         blockA, blockB+j2*actual_kc,
                         actual_mc, panelLength, actualPanelWidth,
                         alpha,
                         actual_kc, actual_kc,  // strides
-                        blockOffset, blockOffset,// offsets
-                        blockW); // workspace
+                        blockOffset, blockOffset);// offsets
           }
         }
-        gebp_kernel(res+i2+(IsLower ? 0 : k2)*resStride, resStride,
+        gebp_kernel(res.getSubMapper(i2, IsLower ? 0 : k2),
                     blockA, geb, actual_mc, actual_kc, rs,
                     alpha,
-                    -1, -1, 0, 0, blockW);
+                    -1, -1, 0, 0);
       }
     }
   }
@@ -373,28 +385,28 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
 * Wrapper to product_triangular_matrix_matrix
 ***************************************************************************/
 
-template<int Mode, bool LhsIsTriangular, typename Lhs, typename Rhs>
-struct traits<TriangularProduct<Mode,LhsIsTriangular,Lhs,false,Rhs,false> >
-  : traits<ProductBase<TriangularProduct<Mode,LhsIsTriangular,Lhs,false,Rhs,false>, Lhs, Rhs> >
-{};
-
 } // end namespace internal
 
+namespace internal {
 template<int Mode, bool LhsIsTriangular, typename Lhs, typename Rhs>
-struct TriangularProduct<Mode,LhsIsTriangular,Lhs,false,Rhs,false>
-  : public ProductBase<TriangularProduct<Mode,LhsIsTriangular,Lhs,false,Rhs,false>, Lhs, Rhs >
+struct triangular_product_impl<Mode,LhsIsTriangular,Lhs,false,Rhs,false>
 {
-  EIGEN_PRODUCT_PUBLIC_INTERFACE(TriangularProduct)
-
-  TriangularProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) {}
-
-  template<typename Dest> void scaleAndAddTo(Dest& dst, const Scalar& alpha) const
+  template<typename Dest> static void run(Dest& dst, const Lhs &a_lhs, const Rhs &a_rhs, const typename Dest::Scalar& alpha)
   {
-    typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(m_lhs);
-    typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(m_rhs);
+    typedef typename Dest::Scalar     Scalar;
+    
+    typedef internal::blas_traits<Lhs> LhsBlasTraits;
+    typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
+    typedef typename internal::remove_all<ActualLhsType>::type ActualLhsTypeCleaned;
+    typedef internal::blas_traits<Rhs> RhsBlasTraits;
+    typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
+    typedef typename internal::remove_all<ActualRhsType>::type ActualRhsTypeCleaned;
+    
+    typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(a_lhs);
+    typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(a_rhs);
 
-    Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(m_lhs)
-                               * RhsBlasTraits::extractScalarFactor(m_rhs);
+    Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(a_lhs)
+                               * RhsBlasTraits::extractScalarFactor(a_rhs);
 
     typedef internal::gemm_blocking_space<(Dest::Flags&RowMajorBit) ? RowMajor : ColMajor,Scalar,Scalar,
               Lhs::MaxRowsAtCompileTime, Rhs::MaxColsAtCompileTime, Lhs::MaxColsAtCompileTime,4> BlockingType;
@@ -405,23 +417,25 @@ struct TriangularProduct<Mode,LhsIsTriangular,Lhs,false,Rhs,false>
     Index stripedDepth = LhsIsTriangular ? ((!IsLower) ? lhs.cols() : (std::min)(lhs.cols(),lhs.rows()))
                                          : ((IsLower)  ? rhs.rows() : (std::min)(rhs.rows(),rhs.cols()));
 
-    BlockingType blocking(stripedRows, stripedCols, stripedDepth);
+    BlockingType blocking(stripedRows, stripedCols, stripedDepth, 1, false);
 
     internal::product_triangular_matrix_matrix<Scalar, Index,
       Mode, LhsIsTriangular,
-      (internal::traits<_ActualLhsType>::Flags&RowMajorBit) ? RowMajor : ColMajor, LhsBlasTraits::NeedToConjugate,
-      (internal::traits<_ActualRhsType>::Flags&RowMajorBit) ? RowMajor : ColMajor, RhsBlasTraits::NeedToConjugate,
+      (internal::traits<ActualLhsTypeCleaned>::Flags&RowMajorBit) ? RowMajor : ColMajor, LhsBlasTraits::NeedToConjugate,
+      (internal::traits<ActualRhsTypeCleaned>::Flags&RowMajorBit) ? RowMajor : ColMajor, RhsBlasTraits::NeedToConjugate,
       (internal::traits<Dest          >::Flags&RowMajorBit) ? RowMajor : ColMajor>
       ::run(
         stripedRows, stripedCols, stripedDepth,   // sizes
-        &lhs.coeffRef(0,0),    lhs.outerStride(), // lhs info
-        &rhs.coeffRef(0,0),    rhs.outerStride(), // rhs info
+        &lhs.coeffRef(0,0), lhs.outerStride(),    // lhs info
+        &rhs.coeffRef(0,0), rhs.outerStride(),    // rhs info
         &dst.coeffRef(0,0), dst.outerStride(),    // result info
         actualAlpha, blocking
       );
   }
 };
 
+} // end namespace internal
+
 } // end namespace Eigen
 
 #endif // EIGEN_TRIANGULAR_MATRIX_MATRIX_H
diff --git a/eigen/Eigen/src/Core/products/TriangularMatrixMatrix_MKL.h b/eigen/Eigen/src/Core/products/TriangularMatrixMatrix_BLAS.h
index 4cc56a4..aecded6 100644
--- a/eigen/Eigen/src/Core/products/TriangularMatrixMatrix_MKL.h
+++ b/eigen/Eigen/src/Core/products/TriangularMatrixMatrix_BLAS.h
@@ -25,13 +25,13 @@
  SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
  ********************************************************************************
- *   Content : Eigen bindings to Intel(R) MKL
+ *   Content : Eigen bindings to BLAS F77
  *   Triangular matrix * matrix product functionality based on ?TRMM.
  ********************************************************************************
 */
 
-#ifndef EIGEN_TRIANGULAR_MATRIX_MATRIX_MKL_H
-#define EIGEN_TRIANGULAR_MATRIX_MATRIX_MKL_H
+#ifndef EIGEN_TRIANGULAR_MATRIX_MATRIX_BLAS_H
+#define EIGEN_TRIANGULAR_MATRIX_MATRIX_BLAS_H
 
 namespace Eigen { 
 
@@ -50,7 +50,7 @@ struct product_triangular_matrix_matrix_trmm :
 
 
 // try to go to BLAS specialization
-#define EIGEN_MKL_TRMM_SPECIALIZE(Scalar, LhsIsTriangular) \
+#define EIGEN_BLAS_TRMM_SPECIALIZE(Scalar, LhsIsTriangular) \
 template <typename Index, int Mode, \
           int LhsStorageOrder, bool ConjugateLhs, \
           int RhsStorageOrder, bool ConjugateRhs> \
@@ -65,17 +65,17 @@ struct product_triangular_matrix_matrix<Scalar,Index, Mode, LhsIsTriangular, \
   } \
 };
 
-EIGEN_MKL_TRMM_SPECIALIZE(double, true)
-EIGEN_MKL_TRMM_SPECIALIZE(double, false)
-EIGEN_MKL_TRMM_SPECIALIZE(dcomplex, true)
-EIGEN_MKL_TRMM_SPECIALIZE(dcomplex, false)
-EIGEN_MKL_TRMM_SPECIALIZE(float, true)
-EIGEN_MKL_TRMM_SPECIALIZE(float, false)
-EIGEN_MKL_TRMM_SPECIALIZE(scomplex, true)
-EIGEN_MKL_TRMM_SPECIALIZE(scomplex, false)
+EIGEN_BLAS_TRMM_SPECIALIZE(double, true)
+EIGEN_BLAS_TRMM_SPECIALIZE(double, false)
+EIGEN_BLAS_TRMM_SPECIALIZE(dcomplex, true)
+EIGEN_BLAS_TRMM_SPECIALIZE(dcomplex, false)
+EIGEN_BLAS_TRMM_SPECIALIZE(float, true)
+EIGEN_BLAS_TRMM_SPECIALIZE(float, false)
+EIGEN_BLAS_TRMM_SPECIALIZE(scomplex, true)
+EIGEN_BLAS_TRMM_SPECIALIZE(scomplex, false)
 
 // implements col-major += alpha * op(triangular) * op(general)
-#define EIGEN_MKL_TRMM_L(EIGTYPE, MKLTYPE, EIGPREFIX, MKLPREFIX) \
+#define EIGEN_BLAS_TRMM_L(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \
 template <typename Index, int Mode, \
           int LhsStorageOrder, bool ConjugateLhs, \
           int RhsStorageOrder, bool ConjugateRhs> \
@@ -106,13 +106,14 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,true, \
    typedef Matrix<EIGTYPE, Dynamic, Dynamic, LhsStorageOrder> MatrixLhs; \
    typedef Matrix<EIGTYPE, Dynamic, Dynamic, RhsStorageOrder> MatrixRhs; \
 \
-/* Non-square case - doesn't fit to MKL ?TRMM. Fall to default triangular product or call MKL ?GEMM*/ \
+/* Non-square case - doesn't fit to BLAS ?TRMM. Fall to default triangular product or call BLAS ?GEMM*/ \
    if (rows != depth) { \
 \
-     int nthr = mkl_domain_get_max_threads(EIGEN_MKL_DOMAIN_BLAS); \
+     /* FIXME handle mkl_domain_get_max_threads */ \
+     /*int nthr = mkl_domain_get_max_threads(EIGEN_BLAS_DOMAIN_BLAS);*/ int nthr = 1;\
 \
      if (((nthr==1) && (((std::max)(rows,depth)-diagSize)/(double)diagSize < 0.5))) { \
-     /* Most likely no benefit to call TRMM or GEMM from MKL*/ \
+     /* Most likely no benefit to call TRMM or GEMM from BLAS */ \
        product_triangular_matrix_matrix<EIGTYPE,Index,Mode,true, \
        LhsStorageOrder,ConjugateLhs, RhsStorageOrder, ConjugateRhs, ColMajor, BuiltIn>::run( \
            _rows, _cols, _depth, _lhs, lhsStride, _rhs, rhsStride, res, resStride, alpha, blocking); \
@@ -121,27 +122,23 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,true, \
      /* Make sense to call GEMM */ \
        Map<const MatrixLhs, 0, OuterStride<> > lhsMap(_lhs,rows,depth,OuterStride<>(lhsStride)); \
        MatrixLhs aa_tmp=lhsMap.template triangularView<Mode>(); \
-       MKL_INT aStride = aa_tmp.outerStride(); \
-       gemm_blocking_space<ColMajor,EIGTYPE,EIGTYPE,Dynamic,Dynamic,Dynamic> gemm_blocking(_rows,_cols,_depth); \
+       BlasIndex aStride = convert_index<BlasIndex>(aa_tmp.outerStride()); \
+       gemm_blocking_space<ColMajor,EIGTYPE,EIGTYPE,Dynamic,Dynamic,Dynamic> gemm_blocking(_rows,_cols,_depth, 1, true); \
        general_matrix_matrix_product<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,RhsStorageOrder,ConjugateRhs,ColMajor>::run( \
        rows, cols, depth, aa_tmp.data(), aStride, _rhs, rhsStride, res, resStride, alpha, gemm_blocking, 0); \
 \
-     /*std::cout << "TRMM_L: A is not square! Go to MKL GEMM implementation! " << nthr<<" \n";*/ \
+     /*std::cout << "TRMM_L: A is not square! Go to BLAS GEMM implementation! " << nthr<<" \n";*/ \
      } \
      return; \
    } \
    char side = 'L', transa, uplo, diag = 'N'; \
    EIGTYPE *b; \
    const EIGTYPE *a; \
-   MKL_INT m, n, lda, ldb; \
-   MKLTYPE alpha_; \
-\
-/* Set alpha_*/ \
-   assign_scalar_eig2mkl<MKLTYPE, EIGTYPE>(alpha_, alpha); \
+   BlasIndex m, n, lda, ldb; \
 \
 /* Set m, n */ \
-   m = (MKL_INT)diagSize; \
-   n = (MKL_INT)cols; \
+   m = convert_index<BlasIndex>(diagSize); \
+   n = convert_index<BlasIndex>(cols); \
 \
 /* Set trans */ \
    transa = (LhsStorageOrder==RowMajor) ? ((ConjugateLhs) ? 'C' : 'T') : 'N'; \
@@ -152,7 +149,7 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,true, \
 \
    if (ConjugateRhs) b_tmp = rhs.conjugate(); else b_tmp = rhs; \
    b = b_tmp.data(); \
-   ldb = b_tmp.outerStride(); \
+   ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \
 \
 /* Set uplo */ \
    uplo = IsLower ? 'L' : 'U'; \
@@ -168,14 +165,14 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,true, \
      else if (IsUnitDiag) \
        a_tmp.diagonal().setOnes();\
      a = a_tmp.data(); \
-     lda = a_tmp.outerStride(); \
+     lda = convert_index<BlasIndex>(a_tmp.outerStride()); \
    } else { \
      a = _lhs; \
-     lda = lhsStride; \
+     lda = convert_index<BlasIndex>(lhsStride); \
    } \
-   /*std::cout << "TRMM_L: A is square! Go to MKL TRMM implementation! \n";*/ \
+   /*std::cout << "TRMM_L: A is square! Go to BLAS TRMM implementation! \n";*/ \
 /* call ?trmm*/ \
-   MKLPREFIX##trmm(&side, &uplo, &transa, &diag, &m, &n, &alpha_, (const MKLTYPE*)a, &lda, (MKLTYPE*)b, &ldb); \
+   BLASPREFIX##trmm_(&side, &uplo, &transa, &diag, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (BLASTYPE*)b, &ldb); \
 \
 /* Add op(a_triangular)*b into res*/ \
    Map<MatrixX##EIGPREFIX, 0, OuterStride<> > res_tmp(res,rows,cols,OuterStride<>(resStride)); \
@@ -183,13 +180,13 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,true, \
   } \
 };
 
-EIGEN_MKL_TRMM_L(double, double, d, d)
-EIGEN_MKL_TRMM_L(dcomplex, MKL_Complex16, cd, z)
-EIGEN_MKL_TRMM_L(float, float, f, s)
-EIGEN_MKL_TRMM_L(scomplex, MKL_Complex8, cf, c)
+EIGEN_BLAS_TRMM_L(double, double, d, d)
+EIGEN_BLAS_TRMM_L(dcomplex, double, cd, z)
+EIGEN_BLAS_TRMM_L(float, float, f, s)
+EIGEN_BLAS_TRMM_L(scomplex, float, cf, c)
 
 // implements col-major += alpha * op(general) * op(triangular)
-#define EIGEN_MKL_TRMM_R(EIGTYPE, MKLTYPE, EIGPREFIX, MKLPREFIX) \
+#define EIGEN_BLAS_TRMM_R(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \
 template <typename Index, int Mode, \
           int LhsStorageOrder, bool ConjugateLhs, \
           int RhsStorageOrder, bool ConjugateRhs> \
@@ -220,13 +217,13 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,false, \
    typedef Matrix<EIGTYPE, Dynamic, Dynamic, LhsStorageOrder> MatrixLhs; \
    typedef Matrix<EIGTYPE, Dynamic, Dynamic, RhsStorageOrder> MatrixRhs; \
 \
-/* Non-square case - doesn't fit to MKL ?TRMM. Fall to default triangular product or call MKL ?GEMM*/ \
+/* Non-square case - doesn't fit to BLAS ?TRMM. Fall to default triangular product or call BLAS ?GEMM*/ \
    if (cols != depth) { \
 \
-     int nthr = mkl_domain_get_max_threads(EIGEN_MKL_DOMAIN_BLAS); \
+     int nthr = 1 /*mkl_domain_get_max_threads(EIGEN_BLAS_DOMAIN_BLAS)*/; \
 \
      if ((nthr==1) && (((std::max)(cols,depth)-diagSize)/(double)diagSize < 0.5)) { \
-     /* Most likely no benefit to call TRMM or GEMM from MKL*/ \
+     /* Most likely no benefit to call TRMM or GEMM from BLAS*/ \
        product_triangular_matrix_matrix<EIGTYPE,Index,Mode,false, \
        LhsStorageOrder,ConjugateLhs, RhsStorageOrder, ConjugateRhs, ColMajor, BuiltIn>::run( \
            _rows, _cols, _depth, _lhs, lhsStride, _rhs, rhsStride, res, resStride, alpha, blocking); \
@@ -235,27 +232,23 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,false, \
      /* Make sense to call GEMM */ \
        Map<const MatrixRhs, 0, OuterStride<> > rhsMap(_rhs,depth,cols, OuterStride<>(rhsStride)); \
        MatrixRhs aa_tmp=rhsMap.template triangularView<Mode>(); \
-       MKL_INT aStride = aa_tmp.outerStride(); \
-       gemm_blocking_space<ColMajor,EIGTYPE,EIGTYPE,Dynamic,Dynamic,Dynamic> gemm_blocking(_rows,_cols,_depth); \
+       BlasIndex aStride = convert_index<BlasIndex>(aa_tmp.outerStride()); \
+       gemm_blocking_space<ColMajor,EIGTYPE,EIGTYPE,Dynamic,Dynamic,Dynamic> gemm_blocking(_rows,_cols,_depth, 1, true); \
        general_matrix_matrix_product<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,RhsStorageOrder,ConjugateRhs,ColMajor>::run( \
        rows, cols, depth, _lhs, lhsStride, aa_tmp.data(), aStride, res, resStride, alpha, gemm_blocking, 0); \
 \
-     /*std::cout << "TRMM_R: A is not square! Go to MKL GEMM implementation! " << nthr<<" \n";*/ \
+     /*std::cout << "TRMM_R: A is not square! Go to BLAS GEMM implementation! " << nthr<<" \n";*/ \
      } \
      return; \
    } \
    char side = 'R', transa, uplo, diag = 'N'; \
    EIGTYPE *b; \
    const EIGTYPE *a; \
-   MKL_INT m, n, lda, ldb; \
-   MKLTYPE alpha_; \
-\
-/* Set alpha_*/ \
-   assign_scalar_eig2mkl<MKLTYPE, EIGTYPE>(alpha_, alpha); \
+   BlasIndex m, n, lda, ldb; \
 \
 /* Set m, n */ \
-   m = (MKL_INT)rows; \
-   n = (MKL_INT)diagSize; \
+   m = convert_index<BlasIndex>(rows); \
+   n = convert_index<BlasIndex>(diagSize); \
 \
 /* Set trans */ \
    transa = (RhsStorageOrder==RowMajor) ? ((ConjugateRhs) ? 'C' : 'T') : 'N'; \
@@ -266,7 +259,7 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,false, \
 \
    if (ConjugateLhs) b_tmp = lhs.conjugate(); else b_tmp = lhs; \
    b = b_tmp.data(); \
-   ldb = b_tmp.outerStride(); \
+   ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \
 \
 /* Set uplo */ \
    uplo = IsLower ? 'L' : 'U'; \
@@ -282,14 +275,14 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,false, \
      else if (IsUnitDiag) \
        a_tmp.diagonal().setOnes();\
      a = a_tmp.data(); \
-     lda = a_tmp.outerStride(); \
+     lda = convert_index<BlasIndex>(a_tmp.outerStride()); \
    } else { \
      a = _rhs; \
-     lda = rhsStride; \
+     lda = convert_index<BlasIndex>(rhsStride); \
    } \
-   /*std::cout << "TRMM_R: A is square! Go to MKL TRMM implementation! \n";*/ \
+   /*std::cout << "TRMM_R: A is square! Go to BLAS TRMM implementation! \n";*/ \
 /* call ?trmm*/ \
-   MKLPREFIX##trmm(&side, &uplo, &transa, &diag, &m, &n, &alpha_, (const MKLTYPE*)a, &lda, (MKLTYPE*)b, &ldb); \
+   BLASPREFIX##trmm_(&side, &uplo, &transa, &diag, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (BLASTYPE*)b, &ldb); \
 \
 /* Add op(a_triangular)*b into res*/ \
    Map<MatrixX##EIGPREFIX, 0, OuterStride<> > res_tmp(res,rows,cols,OuterStride<>(resStride)); \
@@ -297,13 +290,13 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,false, \
   } \
 };
 
-EIGEN_MKL_TRMM_R(double, double, d, d)
-EIGEN_MKL_TRMM_R(dcomplex, MKL_Complex16, cd, z)
-EIGEN_MKL_TRMM_R(float, float, f, s)
-EIGEN_MKL_TRMM_R(scomplex, MKL_Complex8, cf, c)
+EIGEN_BLAS_TRMM_R(double, double, d, d)
+EIGEN_BLAS_TRMM_R(dcomplex, double, cd, z)
+EIGEN_BLAS_TRMM_R(float, float, f, s)
+EIGEN_BLAS_TRMM_R(scomplex, float, cf, c)
 
 } // end namespace internal
 
 } // end namespace Eigen
 
-#endif // EIGEN_TRIANGULAR_MATRIX_MATRIX_MKL_H
+#endif // EIGEN_TRIANGULAR_MATRIX_MATRIX_BLAS_H
diff --git a/eigen/Eigen/src/Core/products/TriangularMatrixVector.h b/eigen/Eigen/src/Core/products/TriangularMatrixVector.h
index 6117d5a..4b292e7 100644
--- a/eigen/Eigen/src/Core/products/TriangularMatrixVector.h
+++ b/eigen/Eigen/src/Core/products/TriangularMatrixVector.h
@@ -10,7 +10,7 @@
 #ifndef EIGEN_TRIANGULARMATRIXVECTOR_H
 #define EIGEN_TRIANGULARMATRIXVECTOR_H
 
-namespace Eigen { 
+namespace Eigen {
 
 namespace internal {
 
@@ -20,20 +20,20 @@ struct triangular_matrix_vector_product;
 template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs, int Version>
 struct triangular_matrix_vector_product<Index,Mode,LhsScalar,ConjLhs,RhsScalar,ConjRhs,ColMajor,Version>
 {
-  typedef typename scalar_product_traits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+  typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
   enum {
     IsLower = ((Mode&Lower)==Lower),
     HasUnitDiag = (Mode & UnitDiag)==UnitDiag,
     HasZeroDiag = (Mode & ZeroDiag)==ZeroDiag
   };
   static EIGEN_DONT_INLINE  void run(Index _rows, Index _cols, const LhsScalar* _lhs, Index lhsStride,
-                                     const RhsScalar* _rhs, Index rhsIncr, ResScalar* _res, Index resIncr, const ResScalar& alpha);
+                                     const RhsScalar* _rhs, Index rhsIncr, ResScalar* _res, Index resIncr, const RhsScalar& alpha);
 };
 
 template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs, int Version>
 EIGEN_DONT_INLINE void triangular_matrix_vector_product<Index,Mode,LhsScalar,ConjLhs,RhsScalar,ConjRhs,ColMajor,Version>
   ::run(Index _rows, Index _cols, const LhsScalar* _lhs, Index lhsStride,
-        const RhsScalar* _rhs, Index rhsIncr, ResScalar* _res, Index resIncr, const ResScalar& alpha)
+        const RhsScalar* _rhs, Index rhsIncr, ResScalar* _res, Index resIncr, const RhsScalar& alpha)
   {
     static const Index PanelWidth = EIGEN_TUNE_TRIANGULAR_PANEL_WIDTH;
     Index size = (std::min)(_rows,_cols);
@@ -43,7 +43,7 @@ EIGEN_DONT_INLINE void triangular_matrix_vector_product<Index,Mode,LhsScalar,Con
     typedef Map<const Matrix<LhsScalar,Dynamic,Dynamic,ColMajor>, 0, OuterStride<> > LhsMap;
     const LhsMap lhs(_lhs,rows,cols,OuterStride<>(lhsStride));
     typename conj_expr_if<ConjLhs,LhsMap>::type cjLhs(lhs);
-    
+
     typedef Map<const Matrix<RhsScalar,Dynamic,1>, 0, InnerStride<> > RhsMap;
     const RhsMap rhs(_rhs,cols,InnerStride<>(rhsIncr));
     typename conj_expr_if<ConjRhs,RhsMap>::type cjRhs(rhs);
@@ -51,6 +51,9 @@ EIGEN_DONT_INLINE void triangular_matrix_vector_product<Index,Mode,LhsScalar,Con
     typedef Map<Matrix<ResScalar,Dynamic,1> > ResMap;
     ResMap res(_res,rows);
 
+    typedef const_blas_data_mapper<LhsScalar,Index,ColMajor> LhsMapper;
+    typedef const_blas_data_mapper<RhsScalar,Index,RowMajor> RhsMapper;
+
     for (Index pi=0; pi<size; pi+=PanelWidth)
     {
       Index actualPanelWidth = (std::min)(PanelWidth, size-pi);
@@ -68,19 +71,19 @@ EIGEN_DONT_INLINE void triangular_matrix_vector_product<Index,Mode,LhsScalar,Con
       if (r>0)
       {
         Index s = IsLower ? pi+actualPanelWidth : 0;
-        general_matrix_vector_product<Index,LhsScalar,ColMajor,ConjLhs,RhsScalar,ConjRhs,BuiltIn>::run(
+        general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,ConjLhs,RhsScalar,RhsMapper,ConjRhs,BuiltIn>::run(
             r, actualPanelWidth,
-            &lhs.coeffRef(s,pi), lhsStride,
-            &rhs.coeffRef(pi), rhsIncr,
+            LhsMapper(&lhs.coeffRef(s,pi), lhsStride),
+            RhsMapper(&rhs.coeffRef(pi), rhsIncr),
             &res.coeffRef(s), resIncr, alpha);
       }
     }
     if((!IsLower) && cols>size)
     {
-      general_matrix_vector_product<Index,LhsScalar,ColMajor,ConjLhs,RhsScalar,ConjRhs>::run(
+      general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,ConjLhs,RhsScalar,RhsMapper,ConjRhs>::run(
           rows, cols-size,
-          &lhs.coeffRef(0,size), lhsStride,
-          &rhs.coeffRef(size), rhsIncr,
+          LhsMapper(&lhs.coeffRef(0,size), lhsStride),
+          RhsMapper(&rhs.coeffRef(size), rhsIncr),
           _res, resIncr, alpha);
     }
   }
@@ -88,7 +91,7 @@ EIGEN_DONT_INLINE void triangular_matrix_vector_product<Index,Mode,LhsScalar,Con
 template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs,int Version>
 struct triangular_matrix_vector_product<Index,Mode,LhsScalar,ConjLhs,RhsScalar,ConjRhs,RowMajor,Version>
 {
-  typedef typename scalar_product_traits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+  typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
   enum {
     IsLower = ((Mode&Lower)==Lower),
     HasUnitDiag = (Mode & UnitDiag)==UnitDiag,
@@ -118,7 +121,10 @@ EIGEN_DONT_INLINE void triangular_matrix_vector_product<Index,Mode,LhsScalar,Con
 
     typedef Map<Matrix<ResScalar,Dynamic,1>, 0, InnerStride<> > ResMap;
     ResMap res(_res,rows,InnerStride<>(resIncr));
-    
+
+    typedef const_blas_data_mapper<LhsScalar,Index,RowMajor> LhsMapper;
+    typedef const_blas_data_mapper<RhsScalar,Index,RowMajor> RhsMapper;
+
     for (Index pi=0; pi<diagSize; pi+=PanelWidth)
     {
       Index actualPanelWidth = (std::min)(PanelWidth, diagSize-pi);
@@ -136,19 +142,19 @@ EIGEN_DONT_INLINE void triangular_matrix_vector_product<Index,Mode,LhsScalar,Con
       if (r>0)
       {
         Index s = IsLower ? 0 : pi + actualPanelWidth;
-        general_matrix_vector_product<Index,LhsScalar,RowMajor,ConjLhs,RhsScalar,ConjRhs,BuiltIn>::run(
+        general_matrix_vector_product<Index,LhsScalar,LhsMapper,RowMajor,ConjLhs,RhsScalar,RhsMapper,ConjRhs,BuiltIn>::run(
             actualPanelWidth, r,
-            &lhs.coeffRef(pi,s), lhsStride,
-            &rhs.coeffRef(s), rhsIncr,
+            LhsMapper(&lhs.coeffRef(pi,s), lhsStride),
+            RhsMapper(&rhs.coeffRef(s), rhsIncr),
             &res.coeffRef(pi), resIncr, alpha);
       }
     }
     if(IsLower && rows>diagSize)
     {
-      general_matrix_vector_product<Index,LhsScalar,RowMajor,ConjLhs,RhsScalar,ConjRhs>::run(
+      general_matrix_vector_product<Index,LhsScalar,LhsMapper,RowMajor,ConjLhs,RhsScalar,RhsMapper,ConjRhs>::run(
             rows-diagSize, cols,
-            &lhs.coeffRef(diagSize,0), lhsStride,
-            &rhs.coeffRef(0), rhsIncr,
+            LhsMapper(&lhs.coeffRef(diagSize,0), lhsStride),
+            RhsMapper(&rhs.coeffRef(0), rhsIncr),
             &res.coeffRef(diagSize), resIncr, alpha);
     }
   }
@@ -157,83 +163,66 @@ EIGEN_DONT_INLINE void triangular_matrix_vector_product<Index,Mode,LhsScalar,Con
 * Wrapper to product_triangular_vector
 ***************************************************************************/
 
-template<int Mode, bool LhsIsTriangular, typename Lhs, typename Rhs>
-struct traits<TriangularProduct<Mode,LhsIsTriangular,Lhs,false,Rhs,true> >
- : traits<ProductBase<TriangularProduct<Mode,LhsIsTriangular,Lhs,false,Rhs,true>, Lhs, Rhs> >
-{};
-
-template<int Mode, bool LhsIsTriangular, typename Lhs, typename Rhs>
-struct traits<TriangularProduct<Mode,LhsIsTriangular,Lhs,true,Rhs,false> >
- : traits<ProductBase<TriangularProduct<Mode,LhsIsTriangular,Lhs,true,Rhs,false>, Lhs, Rhs> >
-{};
-
-
-template<int StorageOrder>
+template<int Mode,int StorageOrder>
 struct trmv_selector;
 
 } // end namespace internal
 
+namespace internal {
+
 template<int Mode, typename Lhs, typename Rhs>
-struct TriangularProduct<Mode,true,Lhs,false,Rhs,true>
-  : public ProductBase<TriangularProduct<Mode,true,Lhs,false,Rhs,true>, Lhs, Rhs >
+struct triangular_product_impl<Mode,true,Lhs,false,Rhs,true>
 {
-  EIGEN_PRODUCT_PUBLIC_INTERFACE(TriangularProduct)
-
-  TriangularProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) {}
-
-  template<typename Dest> void scaleAndAddTo(Dest& dst, const Scalar& alpha) const
+  template<typename Dest> static void run(Dest& dst, const Lhs &lhs, const Rhs &rhs, const typename Dest::Scalar& alpha)
   {
-    eigen_assert(dst.rows()==m_lhs.rows() && dst.cols()==m_rhs.cols());
+    eigen_assert(dst.rows()==lhs.rows() && dst.cols()==rhs.cols());
   
-    internal::trmv_selector<(int(internal::traits<Lhs>::Flags)&RowMajorBit) ? RowMajor : ColMajor>::run(*this, dst, alpha);
+    internal::trmv_selector<Mode,(int(internal::traits<Lhs>::Flags)&RowMajorBit) ? RowMajor : ColMajor>::run(lhs, rhs, dst, alpha);
   }
 };
 
 template<int Mode, typename Lhs, typename Rhs>
-struct TriangularProduct<Mode,false,Lhs,true,Rhs,false>
-  : public ProductBase<TriangularProduct<Mode,false,Lhs,true,Rhs,false>, Lhs, Rhs >
+struct triangular_product_impl<Mode,false,Lhs,true,Rhs,false>
 {
-  EIGEN_PRODUCT_PUBLIC_INTERFACE(TriangularProduct)
-
-  TriangularProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) {}
-
-  template<typename Dest> void scaleAndAddTo(Dest& dst, const Scalar& alpha) const
+  template<typename Dest> static void run(Dest& dst, const Lhs &lhs, const Rhs &rhs, const typename Dest::Scalar& alpha)
   {
-    eigen_assert(dst.rows()==m_lhs.rows() && dst.cols()==m_rhs.cols());
+    eigen_assert(dst.rows()==lhs.rows() && dst.cols()==rhs.cols());
 
-    typedef TriangularProduct<(Mode & (UnitDiag|ZeroDiag)) | ((Mode & Lower) ? Upper : Lower),true,Transpose<const Rhs>,false,Transpose<const Lhs>,true> TriangularProductTranspose;
     Transpose<Dest> dstT(dst);
-    internal::trmv_selector<(int(internal::traits<Rhs>::Flags)&RowMajorBit) ? ColMajor : RowMajor>::run(
-      TriangularProductTranspose(m_rhs.transpose(),m_lhs.transpose()), dstT, alpha);
+    internal::trmv_selector<(Mode & (UnitDiag|ZeroDiag)) | ((Mode & Lower) ? Upper : Lower),
+                            (int(internal::traits<Rhs>::Flags)&RowMajorBit) ? ColMajor : RowMajor>
+            ::run(rhs.transpose(),lhs.transpose(), dstT, alpha);
   }
 };
 
+} // end namespace internal
+
 namespace internal {
 
 // TODO: find a way to factorize this piece of code with gemv_selector since the logic is exactly the same.
   
-template<> struct trmv_selector<ColMajor>
+template<int Mode> struct trmv_selector<Mode,ColMajor>
 {
-  template<int Mode, typename Lhs, typename Rhs, typename Dest>
-  static void run(const TriangularProduct<Mode,true,Lhs,false,Rhs,true>& prod, Dest& dest, const typename TriangularProduct<Mode,true,Lhs,false,Rhs,true>::Scalar& alpha)
+  template<typename Lhs, typename Rhs, typename Dest>
+  static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
   {
-    typedef TriangularProduct<Mode,true,Lhs,false,Rhs,true> ProductType;
-    typedef typename ProductType::Index Index;
-    typedef typename ProductType::LhsScalar   LhsScalar;
-    typedef typename ProductType::RhsScalar   RhsScalar;
-    typedef typename ProductType::Scalar      ResScalar;
-    typedef typename ProductType::RealScalar  RealScalar;
-    typedef typename ProductType::ActualLhsType ActualLhsType;
-    typedef typename ProductType::ActualRhsType ActualRhsType;
-    typedef typename ProductType::LhsBlasTraits LhsBlasTraits;
-    typedef typename ProductType::RhsBlasTraits RhsBlasTraits;
-    typedef Map<Matrix<ResScalar,Dynamic,1>, Aligned> MappedDest;
-
-    typename internal::add_const_on_value_type<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(prod.lhs());
-    typename internal::add_const_on_value_type<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(prod.rhs());
-
-    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs())
-                                  * RhsBlasTraits::extractScalarFactor(prod.rhs());
+    typedef typename Lhs::Scalar      LhsScalar;
+    typedef typename Rhs::Scalar      RhsScalar;
+    typedef typename Dest::Scalar     ResScalar;
+    typedef typename Dest::RealScalar RealScalar;
+    
+    typedef internal::blas_traits<Lhs> LhsBlasTraits;
+    typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
+    typedef internal::blas_traits<Rhs> RhsBlasTraits;
+    typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
+    
+    typedef Map<Matrix<ResScalar,Dynamic,1>, EIGEN_PLAIN_ENUM_MIN(AlignedMax,internal::packet_traits<ResScalar>::size)> MappedDest;
+
+    typename internal::add_const_on_value_type<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(lhs);
+    typename internal::add_const_on_value_type<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(rhs);
+
+    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(lhs)
+                                  * RhsBlasTraits::extractScalarFactor(rhs);
 
     enum {
       // FIXME find a way to allow an inner stride on the result if packet_traits<Scalar>::size==1
@@ -247,7 +236,7 @@ template<> struct trmv_selector<ColMajor>
 
     bool alphaIsCompatible = (!ComplexByReal) || (numext::imag(actualAlpha)==RealScalar(0));
     bool evalToDest = EvalToDestAtCompileTime && alphaIsCompatible;
-    
+
     RhsScalar compatibleAlpha = get_factor<ResScalar,RhsScalar>::run(actualAlpha);
 
     ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(),
@@ -267,7 +256,7 @@ template<> struct trmv_selector<ColMajor>
       else
         MappedDest(actualDestPtr, dest.size()) = dest;
     }
-    
+
     internal::triangular_matrix_vector_product
       <Index,Mode,
        LhsScalar, LhsBlasTraits::NeedToConjugate,
@@ -288,33 +277,32 @@ template<> struct trmv_selector<ColMajor>
   }
 };
 
-template<> struct trmv_selector<RowMajor>
+template<int Mode> struct trmv_selector<Mode,RowMajor>
 {
-  template<int Mode, typename Lhs, typename Rhs, typename Dest>
-  static void run(const TriangularProduct<Mode,true,Lhs,false,Rhs,true>& prod, Dest& dest, const typename TriangularProduct<Mode,true,Lhs,false,Rhs,true>::Scalar& alpha)
+  template<typename Lhs, typename Rhs, typename Dest>
+  static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
   {
-    typedef TriangularProduct<Mode,true,Lhs,false,Rhs,true> ProductType;
-    typedef typename ProductType::LhsScalar LhsScalar;
-    typedef typename ProductType::RhsScalar RhsScalar;
-    typedef typename ProductType::Scalar    ResScalar;
-    typedef typename ProductType::Index Index;
-    typedef typename ProductType::ActualLhsType ActualLhsType;
-    typedef typename ProductType::ActualRhsType ActualRhsType;
-    typedef typename ProductType::_ActualRhsType _ActualRhsType;
-    typedef typename ProductType::LhsBlasTraits LhsBlasTraits;
-    typedef typename ProductType::RhsBlasTraits RhsBlasTraits;
-
-    typename add_const<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(prod.lhs());
-    typename add_const<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(prod.rhs());
-
-    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs())
-                                  * RhsBlasTraits::extractScalarFactor(prod.rhs());
+    typedef typename Lhs::Scalar      LhsScalar;
+    typedef typename Rhs::Scalar      RhsScalar;
+    typedef typename Dest::Scalar     ResScalar;
+    
+    typedef internal::blas_traits<Lhs> LhsBlasTraits;
+    typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
+    typedef internal::blas_traits<Rhs> RhsBlasTraits;
+    typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
+    typedef typename internal::remove_all<ActualRhsType>::type ActualRhsTypeCleaned;
+
+    typename add_const<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(lhs);
+    typename add_const<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(rhs);
+
+    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(lhs)
+                                  * RhsBlasTraits::extractScalarFactor(rhs);
 
     enum {
-      DirectlyUseRhs = _ActualRhsType::InnerStrideAtCompileTime==1
+      DirectlyUseRhs = ActualRhsTypeCleaned::InnerStrideAtCompileTime==1
     };
 
-    gemv_static_vector_if<RhsScalar,_ActualRhsType::SizeAtCompileTime,_ActualRhsType::MaxSizeAtCompileTime,!DirectlyUseRhs> static_rhs;
+    gemv_static_vector_if<RhsScalar,ActualRhsTypeCleaned::SizeAtCompileTime,ActualRhsTypeCleaned::MaxSizeAtCompileTime,!DirectlyUseRhs> static_rhs;
 
     ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhsPtr,actualRhs.size(),
         DirectlyUseRhs ? const_cast<RhsScalar*>(actualRhs.data()) : static_rhs.data());
@@ -322,12 +310,12 @@ template<> struct trmv_selector<RowMajor>
     if(!DirectlyUseRhs)
     {
       #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
-      int size = actualRhs.size();
+      Index size = actualRhs.size();
       EIGEN_DENSE_STORAGE_CTOR_PLUGIN
       #endif
-      Map<typename _ActualRhsType::PlainObject>(actualRhsPtr, actualRhs.size()) = actualRhs;
+      Map<typename ActualRhsTypeCleaned::PlainObject>(actualRhsPtr, actualRhs.size()) = actualRhs;
     }
-    
+
     internal::triangular_matrix_vector_product
       <Index,Mode,
        LhsScalar, LhsBlasTraits::NeedToConjugate,
diff --git a/eigen/Eigen/src/Core/products/TriangularMatrixVector_MKL.h b/eigen/Eigen/src/Core/products/TriangularMatrixVector_BLAS.h
index 09f110d..07bf26c 100644
--- a/eigen/Eigen/src/Core/products/TriangularMatrixVector_MKL.h
+++ b/eigen/Eigen/src/Core/products/TriangularMatrixVector_BLAS.h
@@ -25,13 +25,13 @@
  SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
  ********************************************************************************
- *   Content : Eigen bindings to Intel(R) MKL
+ *   Content : Eigen bindings to BLAS F77
  *   Triangular matrix-vector product functionality based on ?TRMV.
  ********************************************************************************
 */
 
-#ifndef EIGEN_TRIANGULAR_MATRIX_VECTOR_MKL_H
-#define EIGEN_TRIANGULAR_MATRIX_VECTOR_MKL_H
+#ifndef EIGEN_TRIANGULAR_MATRIX_VECTOR_BLAS_H
+#define EIGEN_TRIANGULAR_MATRIX_VECTOR_BLAS_H
 
 namespace Eigen { 
 
@@ -47,7 +47,7 @@ template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename Rh
 struct triangular_matrix_vector_product_trmv :
   triangular_matrix_vector_product<Index,Mode,LhsScalar,ConjLhs,RhsScalar,ConjRhs,StorageOrder,BuiltIn> {};
 
-#define EIGEN_MKL_TRMV_SPECIALIZE(Scalar) \
+#define EIGEN_BLAS_TRMV_SPECIALIZE(Scalar) \
 template<typename Index, int Mode, bool ConjLhs, bool ConjRhs> \
 struct triangular_matrix_vector_product<Index,Mode,Scalar,ConjLhs,Scalar,ConjRhs,ColMajor,Specialized> { \
  static void run(Index _rows, Index _cols, const Scalar* _lhs, Index lhsStride, \
@@ -65,13 +65,13 @@ struct triangular_matrix_vector_product<Index,Mode,Scalar,ConjLhs,Scalar,ConjRhs
   } \
 };
 
-EIGEN_MKL_TRMV_SPECIALIZE(double)
-EIGEN_MKL_TRMV_SPECIALIZE(float)
-EIGEN_MKL_TRMV_SPECIALIZE(dcomplex)
-EIGEN_MKL_TRMV_SPECIALIZE(scomplex)
+EIGEN_BLAS_TRMV_SPECIALIZE(double)
+EIGEN_BLAS_TRMV_SPECIALIZE(float)
+EIGEN_BLAS_TRMV_SPECIALIZE(dcomplex)
+EIGEN_BLAS_TRMV_SPECIALIZE(scomplex)
 
 // implements col-major: res += alpha * op(triangular) * vector
-#define EIGEN_MKL_TRMV_CM(EIGTYPE, MKLTYPE, EIGPREFIX, MKLPREFIX) \
+#define EIGEN_BLAS_TRMV_CM(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \
 template<typename Index, int Mode, bool ConjLhs, bool ConjRhs> \
 struct triangular_matrix_vector_product_trmv<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,ConjRhs,ColMajor> { \
   enum { \
@@ -105,17 +105,15 @@ struct triangular_matrix_vector_product_trmv<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,
 /* Square part handling */\
 \
    char trans, uplo, diag; \
-   MKL_INT m, n, lda, incx, incy; \
+   BlasIndex m, n, lda, incx, incy; \
    EIGTYPE const *a; \
-   MKLTYPE alpha_, beta_; \
-   assign_scalar_eig2mkl<MKLTYPE, EIGTYPE>(alpha_, alpha); \
-   assign_scalar_eig2mkl<MKLTYPE, EIGTYPE>(beta_, EIGTYPE(1)); \
+   EIGTYPE beta(1); \
 \
 /* Set m, n */ \
-   n = (MKL_INT)size; \
-   lda = lhsStride; \
+   n = convert_index<BlasIndex>(size); \
+   lda = convert_index<BlasIndex>(lhsStride); \
    incx = 1; \
-   incy = resIncr; \
+   incy = convert_index<BlasIndex>(resIncr); \
 \
 /* Set uplo, trans and diag*/ \
    trans = 'N'; \
@@ -123,40 +121,39 @@ struct triangular_matrix_vector_product_trmv<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,
    diag = IsUnitDiag ? 'U' : 'N'; \
 \
 /* call ?TRMV*/ \
-   MKLPREFIX##trmv(&uplo, &trans, &diag, &n, (const MKLTYPE*)_lhs, &lda, (MKLTYPE*)x, &incx); \
+   BLASPREFIX##trmv_(&uplo, &trans, &diag, &n, (const BLASTYPE*)_lhs, &lda, (BLASTYPE*)x, &incx); \
 \
 /* Add op(a_tr)rhs into res*/ \
-   MKLPREFIX##axpy(&n, &alpha_,(const MKLTYPE*)x, &incx, (MKLTYPE*)_res, &incy); \
-/* Non-square case - doesn't fit to MKL ?TRMV. Fall to default triangular product*/ \
+   BLASPREFIX##axpy_(&n, &numext::real_ref(alpha),(const BLASTYPE*)x, &incx, (BLASTYPE*)_res, &incy); \
+/* Non-square case - doesn't fit to BLAS ?TRMV. Fall to default triangular product*/ \
    if (size<(std::max)(rows,cols)) { \
-     typedef Matrix<EIGTYPE, Dynamic, Dynamic> MatrixLhs; \
      if (ConjRhs) x_tmp = rhs.conjugate(); else x_tmp = rhs; \
      x = x_tmp.data(); \
      if (size<rows) { \
        y = _res + size*resIncr; \
        a = _lhs + size; \
-       m = rows-size; \
-       n = size; \
+       m = convert_index<BlasIndex>(rows-size); \
+       n = convert_index<BlasIndex>(size); \
      } \
      else { \
        x += size; \
        y = _res; \
        a = _lhs + size*lda; \
-       m = size; \
-       n = cols-size; \
+       m = convert_index<BlasIndex>(size); \
+       n = convert_index<BlasIndex>(cols-size); \
      } \
-     MKLPREFIX##gemv(&trans, &m, &n, &alpha_, (const MKLTYPE*)a, &lda, (const MKLTYPE*)x, &incx, &beta_, (MKLTYPE*)y, &incy); \
+     BLASPREFIX##gemv_(&trans, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)x, &incx, &numext::real_ref(beta), (BLASTYPE*)y, &incy); \
    } \
   } \
 };
 
-EIGEN_MKL_TRMV_CM(double, double, d, d)
-EIGEN_MKL_TRMV_CM(dcomplex, MKL_Complex16, cd, z)
-EIGEN_MKL_TRMV_CM(float, float, f, s)
-EIGEN_MKL_TRMV_CM(scomplex, MKL_Complex8, cf, c)
+EIGEN_BLAS_TRMV_CM(double,   double, d,  d)
+EIGEN_BLAS_TRMV_CM(dcomplex, double, cd, z)
+EIGEN_BLAS_TRMV_CM(float,    float,  f,  s)
+EIGEN_BLAS_TRMV_CM(scomplex, float,  cf, c)
 
 // implements row-major: res += alpha * op(triangular) * vector
-#define EIGEN_MKL_TRMV_RM(EIGTYPE, MKLTYPE, EIGPREFIX, MKLPREFIX) \
+#define EIGEN_BLAS_TRMV_RM(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \
 template<typename Index, int Mode, bool ConjLhs, bool ConjRhs> \
 struct triangular_matrix_vector_product_trmv<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,ConjRhs,RowMajor> { \
   enum { \
@@ -190,17 +187,15 @@ struct triangular_matrix_vector_product_trmv<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,
 /* Square part handling */\
 \
    char trans, uplo, diag; \
-   MKL_INT m, n, lda, incx, incy; \
+   BlasIndex m, n, lda, incx, incy; \
    EIGTYPE const *a; \
-   MKLTYPE alpha_, beta_; \
-   assign_scalar_eig2mkl<MKLTYPE, EIGTYPE>(alpha_, alpha); \
-   assign_scalar_eig2mkl<MKLTYPE, EIGTYPE>(beta_, EIGTYPE(1)); \
+   EIGTYPE beta(1); \
 \
 /* Set m, n */ \
-   n = (MKL_INT)size; \
-   lda = lhsStride; \
+   n = convert_index<BlasIndex>(size); \
+   lda = convert_index<BlasIndex>(lhsStride); \
    incx = 1; \
-   incy = resIncr; \
+   incy = convert_index<BlasIndex>(resIncr); \
 \
 /* Set uplo, trans and diag*/ \
    trans = ConjLhs ? 'C' : 'T'; \
@@ -208,40 +203,39 @@ struct triangular_matrix_vector_product_trmv<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,
    diag = IsUnitDiag ? 'U' : 'N'; \
 \
 /* call ?TRMV*/ \
-   MKLPREFIX##trmv(&uplo, &trans, &diag, &n, (const MKLTYPE*)_lhs, &lda, (MKLTYPE*)x, &incx); \
+   BLASPREFIX##trmv_(&uplo, &trans, &diag, &n, (const BLASTYPE*)_lhs, &lda, (BLASTYPE*)x, &incx); \
 \
 /* Add op(a_tr)rhs into res*/ \
-   MKLPREFIX##axpy(&n, &alpha_,(const MKLTYPE*)x, &incx, (MKLTYPE*)_res, &incy); \
-/* Non-square case - doesn't fit to MKL ?TRMV. Fall to default triangular product*/ \
+   BLASPREFIX##axpy_(&n, &numext::real_ref(alpha),(const BLASTYPE*)x, &incx, (BLASTYPE*)_res, &incy); \
+/* Non-square case - doesn't fit to BLAS ?TRMV. Fall to default triangular product*/ \
    if (size<(std::max)(rows,cols)) { \
-     typedef Matrix<EIGTYPE, Dynamic, Dynamic> MatrixLhs; \
      if (ConjRhs) x_tmp = rhs.conjugate(); else x_tmp = rhs; \
      x = x_tmp.data(); \
      if (size<rows) { \
        y = _res + size*resIncr; \
        a = _lhs + size*lda; \
-       m = rows-size; \
-       n = size; \
+       m = convert_index<BlasIndex>(rows-size); \
+       n = convert_index<BlasIndex>(size); \
      } \
      else { \
        x += size; \
        y = _res; \
        a = _lhs + size; \
-       m = size; \
-       n = cols-size; \
+       m = convert_index<BlasIndex>(size); \
+       n = convert_index<BlasIndex>(cols-size); \
      } \
-     MKLPREFIX##gemv(&trans, &n, &m, &alpha_, (const MKLTYPE*)a, &lda, (const MKLTYPE*)x, &incx, &beta_, (MKLTYPE*)y, &incy); \
+     BLASPREFIX##gemv_(&trans, &n, &m, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)x, &incx, &numext::real_ref(beta), (BLASTYPE*)y, &incy); \
    } \
   } \
 };
 
-EIGEN_MKL_TRMV_RM(double, double, d, d)
-EIGEN_MKL_TRMV_RM(dcomplex, MKL_Complex16, cd, z)
-EIGEN_MKL_TRMV_RM(float, float, f, s)
-EIGEN_MKL_TRMV_RM(scomplex, MKL_Complex8, cf, c)
+EIGEN_BLAS_TRMV_RM(double,   double, d,  d)
+EIGEN_BLAS_TRMV_RM(dcomplex, double, cd, z)
+EIGEN_BLAS_TRMV_RM(float,    float,  f,  s)
+EIGEN_BLAS_TRMV_RM(scomplex, float,  cf, c)
 
 } // end namespase internal
 
 } // end namespace Eigen
 
-#endif // EIGEN_TRIANGULAR_MATRIX_VECTOR_MKL_H
+#endif // EIGEN_TRIANGULAR_MATRIX_VECTOR_BLAS_H
diff --git a/eigen/Eigen/src/Core/products/TriangularSolverMatrix.h b/eigen/Eigen/src/Core/products/TriangularSolverMatrix.h
index 3984b80..223c38b 100644
--- a/eigen/Eigen/src/Core/products/TriangularSolverMatrix.h
+++ b/eigen/Eigen/src/Core/products/TriangularSolverMatrix.h
@@ -52,10 +52,14 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conju
     level3_blocking<Scalar,Scalar>& blocking)
   {
     Index cols = otherSize;
-    const_blas_data_mapper<Scalar, Index, TriStorageOrder> tri(_tri,triStride);
-    blas_data_mapper<Scalar, Index, ColMajor> other(_other,otherStride);
+
+    typedef const_blas_data_mapper<Scalar, Index, TriStorageOrder> TriMapper;
+    typedef blas_data_mapper<Scalar, Index, ColMajor> OtherMapper;
+    TriMapper tri(_tri, triStride);
+    OtherMapper other(_other, otherStride);
 
     typedef gebp_traits<Scalar,Scalar> Traits;
+
     enum {
       SmallPanelWidth   = EIGEN_PLAIN_ENUM_MAX(Traits::mr,Traits::nr),
       IsLower = (Mode&Lower) == Lower
@@ -66,22 +70,20 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conju
 
     std::size_t sizeA = kc*mc;
     std::size_t sizeB = kc*cols;
-    std::size_t sizeW = kc*Traits::WorkSpaceFactor;
 
     ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA());
     ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
-    ei_declare_aligned_stack_constructed_variable(Scalar, blockW, sizeW, blocking.blockW());
 
     conj_if<Conjugate> conj;
-    gebp_kernel<Scalar, Scalar, Index, Traits::mr, Traits::nr, Conjugate, false> gebp_kernel;
-    gemm_pack_lhs<Scalar, Index, Traits::mr, Traits::LhsProgress, TriStorageOrder> pack_lhs;
-    gemm_pack_rhs<Scalar, Index, Traits::nr, ColMajor, false, true> pack_rhs;
+    gebp_kernel<Scalar, Scalar, Index, OtherMapper, Traits::mr, Traits::nr, Conjugate, false> gebp_kernel;
+    gemm_pack_lhs<Scalar, Index, TriMapper, Traits::mr, Traits::LhsProgress, TriStorageOrder> pack_lhs;
+    gemm_pack_rhs<Scalar, Index, OtherMapper, Traits::nr, ColMajor, false, true> pack_rhs;
 
     // the goal here is to subdivise the Rhs panels such that we keep some cache
     // coherence when accessing the rhs elements
-    std::ptrdiff_t l1, l2;
-    manage_caching_sizes(GetAction, &l1, &l2);
-    Index subcols = cols>0 ? l2/(4 * sizeof(Scalar) *  std::max<Index>(otherStride,size)) : 0;
+    std::ptrdiff_t l1, l2, l3;
+    manage_caching_sizes(GetAction, &l1, &l2, &l3);
+    Index subcols = cols>0 ? l2/(4 * sizeof(Scalar) * std::max<Index>(otherStride,size)) : 0;
     subcols = std::max<Index>((subcols/Traits::nr)*Traits::nr, Traits::nr);
 
     for(Index k2=IsLower ? 0 : size;
@@ -148,17 +150,17 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conju
           Index blockBOffset = IsLower ? k1 : lengthTarget;
 
           // update the respective rows of B from other
-          pack_rhs(blockB+actual_kc*j2, &other(startBlock,j2), otherStride, actualPanelWidth, actual_cols, actual_kc, blockBOffset);
+          pack_rhs(blockB+actual_kc*j2, other.getSubMapper(startBlock,j2), actualPanelWidth, actual_cols, actual_kc, blockBOffset);
 
           // GEBP
           if (lengthTarget>0)
           {
             Index startTarget  = IsLower ? k2+k1+actualPanelWidth : k2-actual_kc;
 
-            pack_lhs(blockA, &tri(startTarget,startBlock), triStride, actualPanelWidth, lengthTarget);
+            pack_lhs(blockA, tri.getSubMapper(startTarget,startBlock), actualPanelWidth, lengthTarget);
 
-            gebp_kernel(&other(startTarget,j2), otherStride, blockA, blockB+actual_kc*j2, lengthTarget, actualPanelWidth, actual_cols, Scalar(-1),
-                        actualPanelWidth, actual_kc, 0, blockBOffset, blockW);
+            gebp_kernel(other.getSubMapper(startTarget,j2), blockA, blockB+actual_kc*j2, lengthTarget, actualPanelWidth, actual_cols, Scalar(-1),
+                        actualPanelWidth, actual_kc, 0, blockBOffset);
           }
         }
       }
@@ -172,16 +174,16 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conju
           const Index actual_mc = (std::min)(mc,end-i2);
           if (actual_mc>0)
           {
-            pack_lhs(blockA, &tri(i2, IsLower ? k2 : k2-kc), triStride, actual_kc, actual_mc);
+            pack_lhs(blockA, tri.getSubMapper(i2, IsLower ? k2 : k2-kc), actual_kc, actual_mc);
 
-            gebp_kernel(_other+i2, otherStride, blockA, blockB, actual_mc, actual_kc, cols, Scalar(-1), -1, -1, 0, 0, blockW);
+            gebp_kernel(other.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc, cols, Scalar(-1), -1, -1, 0, 0);
           }
         }
       }
     }
   }
 
-/* Optimized triangular solver with multiple left hand sides and the trinagular matrix on the right
+/* Optimized triangular solver with multiple left hand sides and the triangular matrix on the right
  */
 template <typename Scalar, typename Index, int Mode, bool Conjugate, int TriStorageOrder>
 struct triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conjugate,TriStorageOrder,ColMajor>
@@ -200,8 +202,12 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conj
     level3_blocking<Scalar,Scalar>& blocking)
   {
     Index rows = otherSize;
-    const_blas_data_mapper<Scalar, Index, TriStorageOrder> rhs(_tri,triStride);
-    blas_data_mapper<Scalar, Index, ColMajor> lhs(_other,otherStride);
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+
+    typedef blas_data_mapper<Scalar, Index, ColMajor> LhsMapper;
+    typedef const_blas_data_mapper<Scalar, Index, TriStorageOrder> RhsMapper;
+    LhsMapper lhs(_other, otherStride);
+    RhsMapper rhs(_tri, triStride);
 
     typedef gebp_traits<Scalar,Scalar> Traits;
     enum {
@@ -215,17 +221,15 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conj
 
     std::size_t sizeA = kc*mc;
     std::size_t sizeB = kc*size;
-    std::size_t sizeW = kc*Traits::WorkSpaceFactor;
 
     ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA());
     ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
-    ei_declare_aligned_stack_constructed_variable(Scalar, blockW, sizeW, blocking.blockW());
 
     conj_if<Conjugate> conj;
-    gebp_kernel<Scalar,Scalar, Index, Traits::mr, Traits::nr, false, Conjugate> gebp_kernel;
-    gemm_pack_rhs<Scalar, Index, Traits::nr,RhsStorageOrder> pack_rhs;
-    gemm_pack_rhs<Scalar, Index, Traits::nr,RhsStorageOrder,false,true> pack_rhs_panel;
-    gemm_pack_lhs<Scalar, Index, Traits::mr, Traits::LhsProgress, ColMajor, false, true> pack_lhs_panel;
+    gebp_kernel<Scalar, Scalar, Index, LhsMapper, Traits::mr, Traits::nr, false, Conjugate> gebp_kernel;
+    gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs;
+    gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr, RhsStorageOrder,false,true> pack_rhs_panel;
+    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, ColMajor, false, true> pack_lhs_panel;
 
     for(Index k2=IsLower ? size : 0;
         IsLower ? k2>0 : k2<size;
@@ -238,7 +242,7 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conj
       Index rs = IsLower ? actual_k2 : size - actual_k2 - actual_kc;
       Scalar* geb = blockB+actual_kc*actual_kc;
 
-      if (rs>0) pack_rhs(geb, &rhs(actual_k2,startPanel), triStride, actual_kc, rs);
+      if (rs>0) pack_rhs(geb, rhs.getSubMapper(actual_k2,startPanel), actual_kc, rs);
 
       // triangular packing (we only pack the panels off the diagonal,
       // neglecting the blocks overlapping the diagonal
@@ -252,7 +256,7 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conj
 
           if (panelLength>0)
           pack_rhs_panel(blockB+j2*actual_kc,
-                         &rhs(actual_k2+panelOffset, actual_j2), triStride,
+                         rhs.getSubMapper(actual_k2+panelOffset, actual_j2),
                          panelLength, actualPanelWidth,
                          actual_kc, panelOffset);
         }
@@ -280,13 +284,12 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conj
             // GEBP
             if(panelLength>0)
             {
-              gebp_kernel(&lhs(i2,absolute_j2), otherStride,
+              gebp_kernel(lhs.getSubMapper(i2,absolute_j2),
                           blockA, blockB+j2*actual_kc,
                           actual_mc, panelLength, actualPanelWidth,
                           Scalar(-1),
                           actual_kc, actual_kc, // strides
-                          panelOffset, panelOffset, // offsets
-                          blockW);  // workspace
+                          panelOffset, panelOffset); // offsets
             }
 
             // unblocked triangular solve
@@ -304,23 +307,23 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conj
               }
               if((Mode & UnitDiag)==0)
               {
-                Scalar b = conj(rhs(j,j));
+                Scalar inv_rjj = RealScalar(1)/conj(rhs(j,j));
                 for (Index i=0; i<actual_mc; ++i)
-                  r[i] /= b;
+                  r[i] *= inv_rjj;
               }
             }
 
             // pack the just computed part of lhs to A
-            pack_lhs_panel(blockA, _other+absolute_j2*otherStride+i2, otherStride,
+            pack_lhs_panel(blockA, LhsMapper(_other+absolute_j2*otherStride+i2, otherStride),
                            actualPanelWidth, actual_mc,
                            actual_kc, j2);
           }
         }
 
         if (rs>0)
-          gebp_kernel(_other+i2+startPanel*otherStride, otherStride, blockA, geb,
+          gebp_kernel(lhs.getSubMapper(i2, startPanel), blockA, geb,
                       actual_mc, actual_kc, rs, Scalar(-1),
-                      -1, -1, 0, 0, blockW);
+                      -1, -1, 0, 0);
       }
     }
   }
diff --git a/eigen/Eigen/src/Core/products/TriangularSolverMatrix_MKL.h b/eigen/Eigen/src/Core/products/TriangularSolverMatrix_BLAS.h
index 6a0bb83..88c0fb7 100644
--- a/eigen/Eigen/src/Core/products/TriangularSolverMatrix_MKL.h
+++ b/eigen/Eigen/src/Core/products/TriangularSolverMatrix_BLAS.h
@@ -25,20 +25,20 @@
  SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
  ********************************************************************************
- *   Content : Eigen bindings to Intel(R) MKL
+ *   Content : Eigen bindings to BLAS F77
  *   Triangular matrix * matrix product functionality based on ?TRMM.
  ********************************************************************************
 */
 
-#ifndef EIGEN_TRIANGULAR_SOLVER_MATRIX_MKL_H
-#define EIGEN_TRIANGULAR_SOLVER_MATRIX_MKL_H
+#ifndef EIGEN_TRIANGULAR_SOLVER_MATRIX_BLAS_H
+#define EIGEN_TRIANGULAR_SOLVER_MATRIX_BLAS_H
 
 namespace Eigen {
 
 namespace internal {
 
 // implements LeftSide op(triangular)^-1 * general
-#define EIGEN_MKL_TRSM_L(EIGTYPE, MKLTYPE, MKLPREFIX) \
+#define EIGEN_BLAS_TRSM_L(EIGTYPE, BLASTYPE, BLASPREFIX) \
 template <typename Index, int Mode, bool Conjugate, int TriStorageOrder> \
 struct triangular_solve_matrix<EIGTYPE,Index,OnTheLeft,Mode,Conjugate,TriStorageOrder,ColMajor> \
 { \
@@ -53,13 +53,11 @@ struct triangular_solve_matrix<EIGTYPE,Index,OnTheLeft,Mode,Conjugate,TriStorage
       const EIGTYPE* _tri, Index triStride, \
       EIGTYPE* _other, Index otherStride, level3_blocking<EIGTYPE,EIGTYPE>& /*blocking*/) \
   { \
-   MKL_INT m = size, n = otherSize, lda, ldb; \
+   BlasIndex m = convert_index<BlasIndex>(size), n = convert_index<BlasIndex>(otherSize), lda, ldb; \
    char side = 'L', uplo, diag='N', transa; \
    /* Set alpha_ */ \
-   MKLTYPE alpha; \
-   EIGTYPE myone(1); \
-   assign_scalar_eig2mkl(alpha, myone); \
-   ldb = otherStride;\
+   EIGTYPE alpha(1); \
+   ldb = convert_index<BlasIndex>(otherStride);\
 \
    const EIGTYPE *a; \
 /* Set trans */ \
@@ -75,25 +73,25 @@ struct triangular_solve_matrix<EIGTYPE,Index,OnTheLeft,Mode,Conjugate,TriStorage
    if (conjA) { \
      a_tmp = tri.conjugate(); \
      a = a_tmp.data(); \
-     lda = a_tmp.outerStride(); \
+     lda = convert_index<BlasIndex>(a_tmp.outerStride()); \
    } else { \
      a = _tri; \
-     lda = triStride; \
+     lda = convert_index<BlasIndex>(triStride); \
    } \
    if (IsUnitDiag) diag='U'; \
 /* call ?trsm*/ \
-   MKLPREFIX##trsm(&side, &uplo, &transa, &diag, &m, &n, &alpha, (const MKLTYPE*)a, &lda, (MKLTYPE*)_other, &ldb); \
+   BLASPREFIX##trsm_(&side, &uplo, &transa, &diag, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (BLASTYPE*)_other, &ldb); \
  } \
 };
 
-EIGEN_MKL_TRSM_L(double, double, d)
-EIGEN_MKL_TRSM_L(dcomplex, MKL_Complex16, z)
-EIGEN_MKL_TRSM_L(float, float, s)
-EIGEN_MKL_TRSM_L(scomplex, MKL_Complex8, c)
+EIGEN_BLAS_TRSM_L(double,   double, d)
+EIGEN_BLAS_TRSM_L(dcomplex, double, z)
+EIGEN_BLAS_TRSM_L(float,    float,  s)
+EIGEN_BLAS_TRSM_L(scomplex, float,  c)
 
 
 // implements RightSide general * op(triangular)^-1
-#define EIGEN_MKL_TRSM_R(EIGTYPE, MKLTYPE, MKLPREFIX) \
+#define EIGEN_BLAS_TRSM_R(EIGTYPE, BLASTYPE, BLASPREFIX) \
 template <typename Index, int Mode, bool Conjugate, int TriStorageOrder> \
 struct triangular_solve_matrix<EIGTYPE,Index,OnTheRight,Mode,Conjugate,TriStorageOrder,ColMajor> \
 { \
@@ -108,13 +106,11 @@ struct triangular_solve_matrix<EIGTYPE,Index,OnTheRight,Mode,Conjugate,TriStorag
       const EIGTYPE* _tri, Index triStride, \
       EIGTYPE* _other, Index otherStride, level3_blocking<EIGTYPE,EIGTYPE>& /*blocking*/) \
   { \
-   MKL_INT m = otherSize, n = size, lda, ldb; \
+   BlasIndex m = convert_index<BlasIndex>(otherSize), n = convert_index<BlasIndex>(size), lda, ldb; \
    char side = 'R', uplo, diag='N', transa; \
    /* Set alpha_ */ \
-   MKLTYPE alpha; \
-   EIGTYPE myone(1); \
-   assign_scalar_eig2mkl(alpha, myone); \
-   ldb = otherStride;\
+   EIGTYPE alpha(1); \
+   ldb = convert_index<BlasIndex>(otherStride);\
 \
    const EIGTYPE *a; \
 /* Set trans */ \
@@ -130,26 +126,26 @@ struct triangular_solve_matrix<EIGTYPE,Index,OnTheRight,Mode,Conjugate,TriStorag
    if (conjA) { \
      a_tmp = tri.conjugate(); \
      a = a_tmp.data(); \
-     lda = a_tmp.outerStride(); \
+     lda = convert_index<BlasIndex>(a_tmp.outerStride()); \
    } else { \
      a = _tri; \
-     lda = triStride; \
+     lda = convert_index<BlasIndex>(triStride); \
    } \
    if (IsUnitDiag) diag='U'; \
 /* call ?trsm*/ \
-   MKLPREFIX##trsm(&side, &uplo, &transa, &diag, &m, &n, &alpha, (const MKLTYPE*)a, &lda, (MKLTYPE*)_other, &ldb); \
+   BLASPREFIX##trsm_(&side, &uplo, &transa, &diag, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (BLASTYPE*)_other, &ldb); \
    /*std::cout << "TRMS_L specialization!\n";*/ \
  } \
 };
 
-EIGEN_MKL_TRSM_R(double, double, d)
-EIGEN_MKL_TRSM_R(dcomplex, MKL_Complex16, z)
-EIGEN_MKL_TRSM_R(float, float, s)
-EIGEN_MKL_TRSM_R(scomplex, MKL_Complex8, c)
+EIGEN_BLAS_TRSM_R(double,   double, d)
+EIGEN_BLAS_TRSM_R(dcomplex, double, z)
+EIGEN_BLAS_TRSM_R(float,    float,  s)
+EIGEN_BLAS_TRSM_R(scomplex, float,  c)
 
 
 } // end namespace internal
 
 } // end namespace Eigen
 
-#endif // EIGEN_TRIANGULAR_SOLVER_MATRIX_MKL_H
+#endif // EIGEN_TRIANGULAR_SOLVER_MATRIX_BLAS_H
diff --git a/eigen/Eigen/src/Core/products/TriangularSolverVector.h b/eigen/Eigen/src/Core/products/TriangularSolverVector.h
index ce4d100..b994759 100644
--- a/eigen/Eigen/src/Core/products/TriangularSolverVector.h
+++ b/eigen/Eigen/src/Core/products/TriangularSolverVector.h
@@ -10,7 +10,7 @@
 #ifndef EIGEN_TRIANGULAR_SOLVER_VECTOR_H
 #define EIGEN_TRIANGULAR_SOLVER_VECTOR_H
 
-namespace Eigen { 
+namespace Eigen {
 
 namespace internal {
 
@@ -25,7 +25,7 @@ struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheRight, Mode, Co
       >::run(size, _lhs, lhsStride, rhs);
   }
 };
-    
+
 // forward and backward substitution, row-major, rhs is a vector
 template<typename LhsScalar, typename RhsScalar, typename Index, int Mode, bool Conjugate>
 struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheLeft, Mode, Conjugate, RowMajor>
@@ -37,6 +37,10 @@ struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheLeft, Mode, Con
   {
     typedef Map<const Matrix<LhsScalar,Dynamic,Dynamic,RowMajor>, 0, OuterStride<> > LhsMap;
     const LhsMap lhs(_lhs,size,size,OuterStride<>(lhsStride));
+
+    typedef const_blas_data_mapper<LhsScalar,Index,RowMajor> LhsMapper;
+    typedef const_blas_data_mapper<RhsScalar,Index,ColMajor> RhsMapper;
+
     typename internal::conditional<
                           Conjugate,
                           const CwiseUnaryOp<typename internal::scalar_conjugate_op<LhsScalar>,LhsMap>,
@@ -58,10 +62,10 @@ struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheLeft, Mode, Con
         Index startRow = IsLower ? pi : pi-actualPanelWidth;
         Index startCol = IsLower ? 0 : pi;
 
-        general_matrix_vector_product<Index,LhsScalar,RowMajor,Conjugate,RhsScalar,false>::run(
+        general_matrix_vector_product<Index,LhsScalar,LhsMapper,RowMajor,Conjugate,RhsScalar,RhsMapper,false>::run(
           actualPanelWidth, r,
-          &lhs.coeffRef(startRow,startCol), lhsStride,
-          rhs + startCol, 1,
+          LhsMapper(&lhs.coeffRef(startRow,startCol), lhsStride),
+          RhsMapper(rhs + startCol, 1),
           rhs + startRow, 1,
           RhsScalar(-1));
       }
@@ -72,7 +76,7 @@ struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheLeft, Mode, Con
         Index s = IsLower ? pi   : i+1;
         if (k>0)
           rhs[i] -= (cjLhs.row(i).segment(s,k).transpose().cwiseProduct(Map<const Matrix<RhsScalar,Dynamic,1> >(rhs+s,k))).sum();
-        
+
         if(!(Mode & UnitDiag))
           rhs[i] /= cjLhs(i,i);
       }
@@ -91,6 +95,8 @@ struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheLeft, Mode, Con
   {
     typedef Map<const Matrix<LhsScalar,Dynamic,Dynamic,ColMajor>, 0, OuterStride<> > LhsMap;
     const LhsMap lhs(_lhs,size,size,OuterStride<>(lhsStride));
+    typedef const_blas_data_mapper<LhsScalar,Index,ColMajor> LhsMapper;
+    typedef const_blas_data_mapper<RhsScalar,Index,ColMajor> RhsMapper;
     typename internal::conditional<Conjugate,
                                    const CwiseUnaryOp<typename internal::scalar_conjugate_op<LhsScalar>,LhsMap>,
                                    const LhsMap&
@@ -122,10 +128,10 @@ struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheLeft, Mode, Con
         // let's directly call the low level product function because:
         // 1 - it is faster to compile
         // 2 - it is slighlty faster at runtime
-        general_matrix_vector_product<Index,LhsScalar,ColMajor,Conjugate,RhsScalar,false>::run(
+        general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,Conjugate,RhsScalar,RhsMapper,false>::run(
             r, actualPanelWidth,
-            &lhs.coeffRef(endBlock,startBlock), lhsStride,
-            rhs+startBlock, 1,
+            LhsMapper(&lhs.coeffRef(endBlock,startBlock), lhsStride),
+            RhsMapper(rhs+startBlock, 1),
             rhs+endBlock, 1, RhsScalar(-1));
       }
     }
diff --git a/eigen/Eigen/src/Core/util/BlasUtil.h b/eigen/Eigen/src/Core/util/BlasUtil.h
index 9d03af3..b1791fb 100644
--- a/eigen/Eigen/src/Core/util/BlasUtil.h
+++ b/eigen/Eigen/src/Core/util/BlasUtil.h
@@ -18,13 +18,13 @@ namespace Eigen {
 namespace internal {
 
 // forward declarations
-template<typename LhsScalar, typename RhsScalar, typename Index, int mr, int nr, bool ConjugateLhs=false, bool ConjugateRhs=false>
+template<typename LhsScalar, typename RhsScalar, typename Index, typename DataMapper, int mr, int nr, bool ConjugateLhs=false, bool ConjugateRhs=false>
 struct gebp_kernel;
 
-template<typename Scalar, typename Index, int nr, int StorageOrder, bool Conjugate = false, bool PanelMode=false>
+template<typename Scalar, typename Index, typename DataMapper, int nr, int StorageOrder, bool Conjugate = false, bool PanelMode=false>
 struct gemm_pack_rhs;
 
-template<typename Scalar, typename Index, int Pack1, int Pack2, int StorageOrder, bool Conjugate = false, bool PanelMode = false>
+template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, int StorageOrder, bool Conjugate = false, bool PanelMode = false>
 struct gemm_pack_lhs;
 
 template<
@@ -34,7 +34,9 @@ template<
   int ResStorageOrder>
 struct general_matrix_matrix_product;
 
-template<typename Index, typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs, typename RhsScalar, bool ConjugateRhs, int Version=Specialized>
+template<typename Index,
+         typename LhsScalar, typename LhsMapper, int LhsStorageOrder, bool ConjugateLhs,
+         typename RhsScalar, typename RhsMapper, bool ConjugateRhs, int Version=Specialized>
 struct general_matrix_vector_product;
 
 
@@ -42,22 +44,35 @@ template<bool Conjugate> struct conj_if;
 
 template<> struct conj_if<true> {
   template<typename T>
-  inline T operator()(const T& x) { return numext::conj(x); }
+  inline T operator()(const T& x) const { return numext::conj(x); }
   template<typename T>
-  inline T pconj(const T& x) { return internal::pconj(x); }
+  inline T pconj(const T& x) const { return internal::pconj(x); }
 };
 
 template<> struct conj_if<false> {
   template<typename T>
-  inline const T& operator()(const T& x) { return x; }
+  inline const T& operator()(const T& x) const { return x; }
   template<typename T>
-  inline const T& pconj(const T& x) { return x; }
+  inline const T& pconj(const T& x) const { return x; }
+};
+
+// Generic implementation for custom complex types.
+template<typename LhsScalar, typename RhsScalar, bool ConjLhs, bool ConjRhs>
+struct conj_helper
+{
+  typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar>::ReturnType Scalar;
+
+  EIGEN_STRONG_INLINE Scalar pmadd(const LhsScalar& x, const RhsScalar& y, const Scalar& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Scalar pmul(const LhsScalar& x, const RhsScalar& y) const
+  { return conj_if<ConjLhs>()(x) *  conj_if<ConjRhs>()(y); }
 };
 
 template<typename Scalar> struct conj_helper<Scalar,Scalar,false,false>
 {
-  EIGEN_STRONG_INLINE Scalar pmadd(const Scalar& x, const Scalar& y, const Scalar& c) const { return internal::pmadd(x,y,c); }
-  EIGEN_STRONG_INLINE Scalar pmul(const Scalar& x, const Scalar& y) const { return internal::pmul(x,y); }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar pmadd(const Scalar& x, const Scalar& y, const Scalar& c) const { return internal::pmadd(x,y,c); }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar pmul(const Scalar& x, const Scalar& y) const { return internal::pmul(x,y); }
 };
 
 template<typename RealScalar> struct conj_helper<std::complex<RealScalar>, std::complex<RealScalar>, false,true>
@@ -109,39 +124,147 @@ template<typename RealScalar,bool Conj> struct conj_helper<RealScalar, std::comp
 };
 
 template<typename From,typename To> struct get_factor {
-  static EIGEN_STRONG_INLINE To run(const From& x) { return x; }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE To run(const From& x) { return To(x); }
 };
 
 template<typename Scalar> struct get_factor<Scalar,typename NumTraits<Scalar>::Real> {
+  EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE typename NumTraits<Scalar>::Real run(const Scalar& x) { return numext::real(x); }
 };
 
+
+template<typename Scalar, typename Index>
+class BlasVectorMapper {
+  public:
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE BlasVectorMapper(Scalar *data) : m_data(data) {}
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Scalar operator()(Index i) const {
+    return m_data[i];
+  }
+  template <typename Packet, int AlignmentType>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet load(Index i) const {
+    return ploadt<Packet, AlignmentType>(m_data + i);
+  }
+
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC bool aligned(Index i) const {
+    return (UIntPtr(m_data+i)%sizeof(Packet))==0;
+  }
+
+  protected:
+  Scalar* m_data;
+};
+
+template<typename Scalar, typename Index, int AlignmentType>
+class BlasLinearMapper {
+  public:
+  typedef typename packet_traits<Scalar>::type Packet;
+  typedef typename packet_traits<Scalar>::half HalfPacket;
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE BlasLinearMapper(Scalar *data) : m_data(data) {}
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void prefetch(int i) const {
+    internal::prefetch(&operator()(i));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Scalar& operator()(Index i) const {
+    return m_data[i];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet loadPacket(Index i) const {
+    return ploadt<Packet, AlignmentType>(m_data + i);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE HalfPacket loadHalfPacket(Index i) const {
+    return ploadt<HalfPacket, AlignmentType>(m_data + i);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void storePacket(Index i, const Packet &p) const {
+    pstoret<Scalar, Packet, AlignmentType>(m_data + i, p);
+  }
+
+  protected:
+  Scalar *m_data;
+};
+
 // Lightweight helper class to access matrix coefficients.
-// Yes, this is somehow redundant with Map<>, but this version is much much lighter,
-// and so I hope better compilation performance (time and code quality).
-template<typename Scalar, typename Index, int StorageOrder>
-class blas_data_mapper
-{
+template<typename Scalar, typename Index, int StorageOrder, int AlignmentType = Unaligned>
+class blas_data_mapper {
   public:
-    blas_data_mapper(Scalar* data, Index stride) : m_data(data), m_stride(stride) {}
-    EIGEN_STRONG_INLINE Scalar& operator()(Index i, Index j)
-    { return m_data[StorageOrder==RowMajor ? j + i*m_stride : i + j*m_stride]; }
+  typedef typename packet_traits<Scalar>::type Packet;
+  typedef typename packet_traits<Scalar>::half HalfPacket;
+
+  typedef BlasLinearMapper<Scalar, Index, AlignmentType> LinearMapper;
+  typedef BlasVectorMapper<Scalar, Index> VectorMapper;
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE blas_data_mapper(Scalar* data, Index stride) : m_data(data), m_stride(stride) {}
+
+  EIGEN_DEVICE_FUNC  EIGEN_ALWAYS_INLINE blas_data_mapper<Scalar, Index, StorageOrder, AlignmentType>
+  getSubMapper(Index i, Index j) const {
+    return blas_data_mapper<Scalar, Index, StorageOrder, AlignmentType>(&operator()(i, j), m_stride);
+  }
+
+  EIGEN_DEVICE_FUNC  EIGEN_ALWAYS_INLINE LinearMapper getLinearMapper(Index i, Index j) const {
+    return LinearMapper(&operator()(i, j));
+  }
+
+  EIGEN_DEVICE_FUNC  EIGEN_ALWAYS_INLINE VectorMapper getVectorMapper(Index i, Index j) const {
+    return VectorMapper(&operator()(i, j));
+  }
+
+
+  EIGEN_DEVICE_FUNC
+  EIGEN_ALWAYS_INLINE Scalar& operator()(Index i, Index j) const {
+    return m_data[StorageOrder==RowMajor ? j + i*m_stride : i + j*m_stride];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet loadPacket(Index i, Index j) const {
+    return ploadt<Packet, AlignmentType>(&operator()(i, j));
+  }
+
+  template <typename PacketT, int AlignmentT>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE PacketT load(Index i, Index j) const {
+    return ploadt<PacketT, AlignmentT>(&operator()(i, j));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE HalfPacket loadHalfPacket(Index i, Index j) const {
+    return ploadt<HalfPacket, AlignmentType>(&operator()(i, j));
+  }
+
+  template<typename SubPacket>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void scatterPacket(Index i, Index j, const SubPacket &p) const {
+    pscatter<Scalar, SubPacket>(&operator()(i, j), p, m_stride);
+  }
+
+  template<typename SubPacket>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE SubPacket gatherPacket(Index i, Index j) const {
+    return pgather<Scalar, SubPacket>(&operator()(i, j), m_stride);
+  }
+
+  EIGEN_DEVICE_FUNC const Index stride() const { return m_stride; }
+  EIGEN_DEVICE_FUNC const Scalar* data() const { return m_data; }
+
+  EIGEN_DEVICE_FUNC Index firstAligned(Index size) const {
+    if (UIntPtr(m_data)%sizeof(Scalar)) {
+      return -1;
+    }
+    return internal::first_default_aligned(m_data, size);
+  }
+
   protected:
-    Scalar* EIGEN_RESTRICT m_data;
-    Index m_stride;
+  Scalar* EIGEN_RESTRICT m_data;
+  const Index m_stride;
 };
 
 // lightweight helper class to access matrix coefficients (const version)
 template<typename Scalar, typename Index, int StorageOrder>
-class const_blas_data_mapper
-{
+class const_blas_data_mapper : public blas_data_mapper<const Scalar, Index, StorageOrder> {
   public:
-    const_blas_data_mapper(const Scalar* data, Index stride) : m_data(data), m_stride(stride) {}
-    EIGEN_STRONG_INLINE const Scalar& operator()(Index i, Index j) const
-    { return m_data[StorageOrder==RowMajor ? j + i*m_stride : i + j*m_stride]; }
-  protected:
-    const Scalar* EIGEN_RESTRICT m_data;
-    Index m_stride;
+  EIGEN_ALWAYS_INLINE const_blas_data_mapper(const Scalar *data, Index stride) : blas_data_mapper<const Scalar, Index, StorageOrder>(data, stride) {}
+
+  EIGEN_ALWAYS_INLINE const_blas_data_mapper<Scalar, Index, StorageOrder> getSubMapper(Index i, Index j) const {
+    return const_blas_data_mapper<Scalar, Index, StorageOrder>(&(this->operator()(i, j)), this->m_stride);
+  }
 };
 
 
@@ -171,13 +294,12 @@ template<typename XprType> struct blas_traits
 };
 
 // pop conjugate
-template<typename Scalar, typename Xpr>
-struct blas_traits<CwiseUnaryOp<scalar_conjugate_op<Scalar>, Xpr> >
- : blas_traits<typename internal::remove_all<typename Xpr::Nested>::type>
+template<typename Scalar, typename NestedXpr>
+struct blas_traits<CwiseUnaryOp<scalar_conjugate_op<Scalar>, NestedXpr> >
+ : blas_traits<NestedXpr>
 {
-  typedef typename internal::remove_all<typename Xpr::Nested>::type NestedXpr;
   typedef blas_traits<NestedXpr> Base;
-  typedef CwiseUnaryOp<scalar_conjugate_op<Scalar>, Xpr> XprType;
+  typedef CwiseUnaryOp<scalar_conjugate_op<Scalar>, NestedXpr> XprType;
   typedef typename Base::ExtractType ExtractType;
 
   enum {
@@ -189,27 +311,41 @@ struct blas_traits<CwiseUnaryOp<scalar_conjugate_op<Scalar>, Xpr> >
 };
 
 // pop scalar multiple
-template<typename Scalar, typename Xpr>
-struct blas_traits<CwiseUnaryOp<scalar_multiple_op<Scalar>, Xpr> >
- : blas_traits<typename internal::remove_all<typename Xpr::Nested>::type>
+template<typename Scalar, typename NestedXpr, typename Plain>
+struct blas_traits<CwiseBinaryOp<scalar_product_op<Scalar>, const CwiseNullaryOp<scalar_constant_op<Scalar>,Plain>, NestedXpr> >
+ : blas_traits<NestedXpr>
 {
-  typedef typename internal::remove_all<typename Xpr::Nested>::type NestedXpr;
   typedef blas_traits<NestedXpr> Base;
-  typedef CwiseUnaryOp<scalar_multiple_op<Scalar>, Xpr> XprType;
+  typedef CwiseBinaryOp<scalar_product_op<Scalar>, const CwiseNullaryOp<scalar_constant_op<Scalar>,Plain>, NestedXpr> XprType;
   typedef typename Base::ExtractType ExtractType;
-  static inline ExtractType extract(const XprType& x) { return Base::extract(x.nestedExpression()); }
+  static inline ExtractType extract(const XprType& x) { return Base::extract(x.rhs()); }
   static inline Scalar extractScalarFactor(const XprType& x)
-  { return x.functor().m_other * Base::extractScalarFactor(x.nestedExpression()); }
+  { return x.lhs().functor().m_other * Base::extractScalarFactor(x.rhs()); }
 };
+template<typename Scalar, typename NestedXpr, typename Plain>
+struct blas_traits<CwiseBinaryOp<scalar_product_op<Scalar>, NestedXpr, const CwiseNullaryOp<scalar_constant_op<Scalar>,Plain> > >
+ : blas_traits<NestedXpr>
+{
+  typedef blas_traits<NestedXpr> Base;
+  typedef CwiseBinaryOp<scalar_product_op<Scalar>, NestedXpr, const CwiseNullaryOp<scalar_constant_op<Scalar>,Plain> > XprType;
+  typedef typename Base::ExtractType ExtractType;
+  static inline ExtractType extract(const XprType& x) { return Base::extract(x.lhs()); }
+  static inline Scalar extractScalarFactor(const XprType& x)
+  { return Base::extractScalarFactor(x.lhs()) * x.rhs().functor().m_other; }
+};
+template<typename Scalar, typename Plain1, typename Plain2>
+struct blas_traits<CwiseBinaryOp<scalar_product_op<Scalar>, const CwiseNullaryOp<scalar_constant_op<Scalar>,Plain1>,
+                                                            const CwiseNullaryOp<scalar_constant_op<Scalar>,Plain2> > >
+ : blas_traits<CwiseNullaryOp<scalar_constant_op<Scalar>,Plain1> >
+{};
 
 // pop opposite
-template<typename Scalar, typename Xpr>
-struct blas_traits<CwiseUnaryOp<scalar_opposite_op<Scalar>, Xpr> >
- : blas_traits<typename internal::remove_all<typename Xpr::Nested>::type>
+template<typename Scalar, typename NestedXpr>
+struct blas_traits<CwiseUnaryOp<scalar_opposite_op<Scalar>, NestedXpr> >
+ : blas_traits<NestedXpr>
 {
-  typedef typename internal::remove_all<typename Xpr::Nested>::type NestedXpr;
   typedef blas_traits<NestedXpr> Base;
-  typedef CwiseUnaryOp<scalar_opposite_op<Scalar>, Xpr> XprType;
+  typedef CwiseUnaryOp<scalar_opposite_op<Scalar>, NestedXpr> XprType;
   typedef typename Base::ExtractType ExtractType;
   static inline ExtractType extract(const XprType& x) { return Base::extract(x.nestedExpression()); }
   static inline Scalar extractScalarFactor(const XprType& x)
@@ -217,14 +353,13 @@ struct blas_traits<CwiseUnaryOp<scalar_opposite_op<Scalar>, Xpr> >
 };
 
 // pop/push transpose
-template<typename Xpr>
-struct blas_traits<Transpose<Xpr> >
- : blas_traits<typename internal::remove_all<typename Xpr::Nested>::type>
+template<typename NestedXpr>
+struct blas_traits<Transpose<NestedXpr> >
+ : blas_traits<NestedXpr>
 {
-  typedef typename internal::remove_all<typename Xpr::Nested>::type NestedXpr;
   typedef typename NestedXpr::Scalar Scalar;
   typedef blas_traits<NestedXpr> Base;
-  typedef Transpose<Xpr> XprType;
+  typedef Transpose<NestedXpr> XprType;
   typedef Transpose<const typename Base::_ExtractType>  ExtractType; // const to get rid of a compile error; anyway blas traits are only used on the RHS
   typedef Transpose<const typename Base::_ExtractType> _ExtractType;
   typedef typename conditional<bool(Base::HasUsableDirectAccess),
@@ -234,7 +369,7 @@ struct blas_traits<Transpose<Xpr> >
   enum {
     IsTransposed = Base::IsTransposed ? 0 : 1
   };
-  static inline ExtractType extract(const XprType& x) { return Base::extract(x.nestedExpression()); }
+  static inline ExtractType extract(const XprType& x) { return ExtractType(Base::extract(x.nestedExpression())); }
   static inline Scalar extractScalarFactor(const XprType& x) { return Base::extractScalarFactor(x.nestedExpression()); }
 };
 
diff --git a/eigen/Eigen/src/Core/util/CMakeLists.txt b/eigen/Eigen/src/Core/util/CMakeLists.txt
deleted file mode 100644
index a1e2e52..0000000
--- a/eigen/Eigen/src/Core/util/CMakeLists.txt
+++ /dev/null
@@ -1,6 +0,0 @@
-FILE(GLOB Eigen_Core_util_SRCS "*.h")
-
-INSTALL(FILES 
-  ${Eigen_Core_util_SRCS}
-  DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/util COMPONENT Devel
-  )
diff --git a/eigen/Eigen/src/Core/util/Constants.h b/eigen/Eigen/src/Core/util/Constants.h
index 78499ff..5d37e5d 100644
--- a/eigen/Eigen/src/Core/util/Constants.h
+++ b/eigen/Eigen/src/Core/util/Constants.h
@@ -1,7 +1,7 @@
 // This file is part of Eigen, a lightweight C++ template library
 // for linear algebra.
 //
-// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2008-2015 Gael Guennebaud <gael.guennebaud@inria.fr>
 // Copyright (C) 2007-2009 Benoit Jacob <jacob.benoit.1@gmail.com>
 //
 // This Source Code Form is subject to the terms of the Mozilla
@@ -25,11 +25,23 @@ const int Dynamic = -1;
   */
 const int DynamicIndex = 0xffffff;
 
+/** This value means that the increment to go from one value to another in a sequence is not constant for each step.
+  */
+const int UndefinedIncr = 0xfffffe;
+
 /** This value means +Infinity; it is currently used only as the p parameter to MatrixBase::lpNorm<int>().
   * The value Infinity there means the L-infinity norm.
   */
 const int Infinity = -1;
 
+/** This value means that the cost to evaluate an expression coefficient is either very expensive or
+  * cannot be known at compile time.
+  *
+  * This value has to be positive to (1) simplify cost computation, and (2) allow to distinguish between a very expensive and very very expensive expressions.
+  * It thus must also be large enough to make sure unrolling won't happen and that sub expressions will be evaluated, but not too large to avoid overflow.
+  */
+const int HugeCost = 10000;
+
 /** \defgroup flags Flags
   * \ingroup Core_Module
   *
@@ -48,19 +60,19 @@ const int Infinity = -1;
   * for a matrix, this means that the storage order is row-major.
   * If this bit is not set, the storage order is column-major.
   * For an expression, this determines the storage order of
-  * the matrix created by evaluation of that expression. 
-  * \sa \ref TopicStorageOrders */
+  * the matrix created by evaluation of that expression.
+  * \sa \blank  \ref TopicStorageOrders */
 const unsigned int RowMajorBit = 0x1;
 
 /** \ingroup flags
-  *
   * means the expression should be evaluated by the calling expression */
 const unsigned int EvalBeforeNestingBit = 0x2;
 
 /** \ingroup flags
-  *
+  * \deprecated
   * means the expression should be evaluated before any assignment */
-const unsigned int EvalBeforeAssigningBit = 0x4;
+EIGEN_DEPRECATED
+const unsigned int EvalBeforeAssigningBit = 0x4; // FIXME deprecated
 
 /** \ingroup flags
   *
@@ -141,17 +153,46 @@ const unsigned int LvalueBit = 0x20;
   */
 const unsigned int DirectAccessBit = 0x40;
 
-/** \ingroup flags
+/** \deprecated \ingroup flags
   *
-  * means the first coefficient packet is guaranteed to be aligned */
-const unsigned int AlignedBit = 0x80;
+  * means the first coefficient packet is guaranteed to be aligned.
+  * An expression cannot has the AlignedBit without the PacketAccessBit flag.
+  * In other words, this means we are allow to perform an aligned packet access to the first element regardless
+  * of the expression kind:
+  * \code
+  * expression.packet<Aligned>(0);
+  * \endcode
+  */
+EIGEN_DEPRECATED const unsigned int AlignedBit = 0x80;
 
 const unsigned int NestByRefBit = 0x100;
 
+/** \ingroup flags
+  *
+  * for an expression, this means that the storage order
+  * can be either row-major or column-major.
+  * The precise choice will be decided at evaluation time or when
+  * combined with other expressions.
+  * \sa \blank  \ref RowMajorBit, \ref TopicStorageOrders */
+const unsigned int NoPreferredStorageOrderBit = 0x200;
+
+/** \ingroup flags
+  *
+  * Means that the underlying coefficients can be accessed through pointers to the sparse (un)compressed storage format,
+  * that is, the expression provides:
+  * \code
+    inline const Scalar* valuePtr() const;
+    inline const Index* innerIndexPtr() const;
+    inline const Index* outerIndexPtr() const;
+    inline const Index* innerNonZeroPtr() const;
+    \endcode
+  */
+const unsigned int CompressedAccessBit = 0x400;
+
+
 // list of flags that are inherited by default
 const unsigned int HereditaryBits = RowMajorBit
-                                  | EvalBeforeNestingBit
-                                  | EvalBeforeAssigningBit;
+                                  | EvalBeforeNestingBit;
 
 /** \defgroup enums Enumerations
   * \ingroup Core_Module
@@ -160,8 +201,8 @@ const unsigned int HereditaryBits = RowMajorBit
   */
 
 /** \ingroup enums
-  * Enum containing possible values for the \p Mode parameter of 
-  * MatrixBase::selfadjointView() and MatrixBase::triangularView(). */
+  * Enum containing possible values for the \c Mode or \c UpLo parameter of
+  * MatrixBase::selfadjointView() and MatrixBase::triangularView(), and selfadjoint solvers. */
 enum UpLoType {
   /** View matrix as a lower triangular matrix. */
   Lower=0x1,                      
@@ -186,12 +227,31 @@ enum UpLoType {
 };
 
 /** \ingroup enums
-  * Enum for indicating whether an object is aligned or not. */
+  * Enum for indicating whether a buffer is aligned or not. */
 enum AlignmentType {
-  /** Object is not correctly aligned for vectorization. */
-  Unaligned=0, 
-  /** Object is aligned for vectorization. */
-  Aligned=1 
+  Unaligned=0,        /**< Data pointer has no specific alignment. */
+  Aligned8=8,         /**< Data pointer is aligned on a 8 bytes boundary. */
+  Aligned16=16,       /**< Data pointer is aligned on a 16 bytes boundary. */
+  Aligned32=32,       /**< Data pointer is aligned on a 32 bytes boundary. */
+  Aligned64=64,       /**< Data pointer is aligned on a 64 bytes boundary. */
+  Aligned128=128,     /**< Data pointer is aligned on a 128 bytes boundary. */
+  AlignedMask=255,
+  Aligned=16,         /**< \deprecated Synonym for Aligned16. */
+#if EIGEN_MAX_ALIGN_BYTES==128
+  AlignedMax = Aligned128
+#elif EIGEN_MAX_ALIGN_BYTES==64
+  AlignedMax = Aligned64
+#elif EIGEN_MAX_ALIGN_BYTES==32
+  AlignedMax = Aligned32
+#elif EIGEN_MAX_ALIGN_BYTES==16
+  AlignedMax = Aligned16
+#elif EIGEN_MAX_ALIGN_BYTES==8
+  AlignedMax = Aligned8
+#elif EIGEN_MAX_ALIGN_BYTES==0
+  AlignedMax = Unaligned
+#else
+#error Invalid value for EIGEN_MAX_ALIGN_BYTES
+#endif
 };
 
 /** \ingroup enums
@@ -297,7 +357,7 @@ enum Default_t    { Default };
 
 /** \internal \ingroup enums
   * Used in AmbiVector. */
-enum {
+enum AmbiVectorMode {
   IsDense         = 0,
   IsSparse
 };
@@ -406,10 +466,16 @@ namespace Architecture
     Generic = 0x0,
     SSE = 0x1,
     AltiVec = 0x2,
+    VSX = 0x3,
+    NEON = 0x4,
 #if defined EIGEN_VECTORIZE_SSE
     Target = SSE
 #elif defined EIGEN_VECTORIZE_ALTIVEC
     Target = AltiVec
+#elif defined EIGEN_VECTORIZE_VSX
+    Target = VSX
+#elif defined EIGEN_VECTORIZE_NEON
+    Target = NEON
 #else
     Target = Generic
 #endif
@@ -417,8 +483,9 @@ namespace Architecture
 }
 
 /** \internal \ingroup enums
-  * Enum used as template parameter in GeneralProduct. */
-enum ProductImplType { CoeffBasedProductMode, LazyCoeffBasedProductMode, OuterProduct, InnerProduct, GemvProduct, GemmProduct };
+  * Enum used as template parameter in Product and product evaluators. */
+enum ProductImplType
+{ DefaultProduct=0, LazyProduct, AliasFreeProduct, CoeffBasedProductMode, LazyCoeffBasedProductMode, OuterProduct, InnerProduct, GemvProduct, GemmProduct };
 
 /** \internal \ingroup enums
   * Enum used in experimental parallel implementation. */
@@ -427,24 +494,57 @@ enum Action {GetAction, SetAction};
 /** The type used to identify a dense storage. */
 struct Dense {};
 
+/** The type used to identify a general sparse storage. */
+struct Sparse {};
+
+/** The type used to identify a general solver (factored) storage. */
+struct SolverStorage {};
+
+/** The type used to identify a permutation storage. */
+struct PermutationStorage {};
+
+/** The type used to identify a permutation storage. */
+struct TranspositionsStorage {};
+
 /** The type used to identify a matrix expression */
 struct MatrixXpr {};
 
 /** The type used to identify an array expression */
 struct ArrayXpr {};
 
+// An evaluator must define its shape. By default, it can be one of the following:
+struct DenseShape             { static std::string debugName() { return "DenseShape"; } };
+struct SolverShape            { static std::string debugName() { return "SolverShape"; } };
+struct HomogeneousShape       { static std::string debugName() { return "HomogeneousShape"; } };
+struct DiagonalShape          { static std::string debugName() { return "DiagonalShape"; } };
+struct BandShape              { static std::string debugName() { return "BandShape"; } };
+struct TriangularShape        { static std::string debugName() { return "TriangularShape"; } };
+struct SelfAdjointShape       { static std::string debugName() { return "SelfAdjointShape"; } };
+struct PermutationShape       { static std::string debugName() { return "PermutationShape"; } };
+struct TranspositionsShape    { static std::string debugName() { return "TranspositionsShape"; } };
+struct SparseShape            { static std::string debugName() { return "SparseShape"; } };
+
 namespace internal {
-  /** \internal
-  * Constants for comparison functors
-  */
-  enum ComparisonName {
-    cmp_EQ = 0,
-    cmp_LT = 1,
-    cmp_LE = 2,
-    cmp_UNORD = 3,
-    cmp_NEQ = 4
-  };
-}
+
+  // random access iterators based on coeff*() accessors.
+struct IndexBased {};
+
+// evaluator based on iterators to access coefficients. 
+struct IteratorBased {};
+
+/** \internal
+ * Constants for comparison functors
+ */
+enum ComparisonName {
+  cmp_EQ = 0,
+  cmp_LT = 1,
+  cmp_LE = 2,
+  cmp_UNORD = 3,
+  cmp_NEQ = 4,
+  cmp_GT = 5,
+  cmp_GE = 6
+};
+} // end namespace internal
 
 } // end namespace Eigen
 
diff --git a/eigen/Eigen/src/Core/util/DisableStupidWarnings.h b/eigen/Eigen/src/Core/util/DisableStupidWarnings.h
index c53457a..4431f2f 100644
--- a/eigen/Eigen/src/Core/util/DisableStupidWarnings.h
+++ b/eigen/Eigen/src/Core/util/DisableStupidWarnings.h
@@ -4,30 +4,36 @@
 #ifdef _MSC_VER
   // 4100 - unreferenced formal parameter (occurred e.g. in aligned_allocator::destroy(pointer p))
   // 4101 - unreferenced local variable
-  // 4127 - conditional expression is constant
   // 4181 - qualifier applied to reference type ignored
   // 4211 - nonstandard extension used : redefined extern to static
   // 4244 - 'argument' : conversion from 'type1' to 'type2', possible loss of data
   // 4273 - QtAlignedMalloc, inconsistent DLL linkage
   // 4324 - structure was padded due to declspec(align())
+  // 4503 - decorated name length exceeded, name was truncated
   // 4512 - assignment operator could not be generated
   // 4522 - 'class' : multiple assignment operators specified
   // 4700 - uninitialized local variable 'xyz' used
+  // 4714 - function marked as __forceinline not inlined
   // 4717 - 'function' : recursive on all control paths, function will cause runtime stack overflow
+  // 4800 - 'type' : forcing value to bool 'true' or 'false' (performance warning)
   #ifndef EIGEN_PERMANENTLY_DISABLE_STUPID_WARNINGS
     #pragma warning( push )
   #endif
-  #pragma warning( disable : 4100 4101 4127 4181 4211 4244 4273 4324 4512 4522 4700 4717 )
+  #pragma warning( disable : 4100 4101 4181 4211 4244 4273 4324 4503 4512 4522 4700 4714 4717 4800)
+
 #elif defined __INTEL_COMPILER
   // 2196 - routine is both "inline" and "noinline" ("noinline" assumed)
   //        ICC 12 generates this warning even without any inline keyword, when defining class methods 'inline' i.e. inside of class body
   //        typedef that may be a reference type.
   // 279  - controlling expression is constant
   //        ICC 12 generates this warning on assert(constant_expression_depending_on_template_params) and frankly this is a legitimate use case.
+  // 1684 - conversion from pointer to same-sized integral type (potential portability problem)
+  // 2259 - non-pointer conversion from "Eigen::Index={ptrdiff_t={long}}" to "int" may lose significant bits
   #ifndef EIGEN_PERMANENTLY_DISABLE_STUPID_WARNINGS
     #pragma warning push
   #endif
-  #pragma warning disable 2196 279
+  #pragma warning disable 2196 279 1684 2259
+
 #elif defined __clang__
   // -Wconstant-logical-operand - warning: use of logical && with constant operand; switch to bitwise & or remove constant
   //     this is really a stupid warning as it warns on compile-time expressions involving enums
@@ -35,6 +41,9 @@
     #pragma clang diagnostic push
   #endif
   #pragma clang diagnostic ignored "-Wconstant-logical-operand"
+  #if __clang_major__ >= 3 && __clang_minor__ >= 5
+    #pragma clang diagnostic ignored "-Wabsolute-value"
+  #endif
 
 #elif defined __GNUC__ && __GNUC__>=6
 
@@ -45,4 +54,24 @@
 
 #endif
 
+#if defined __NVCC__
+  // Disable the "statement is unreachable" message
+  #pragma diag_suppress code_is_unreachable
+  // Disable the "dynamic initialization in unreachable code" message
+  #pragma diag_suppress initialization_not_reachable
+  // Disable the "invalid error number" message that we get with older versions of nvcc
+  #pragma diag_suppress 1222
+  // Disable the "calling a __host__ function from a __host__ __device__ function is not allowed" messages (yes, there are many of them and they seem to change with every version of the compiler)
+  #pragma diag_suppress 2527
+  #pragma diag_suppress 2529
+  #pragma diag_suppress 2651
+  #pragma diag_suppress 2653
+  #pragma diag_suppress 2668
+  #pragma diag_suppress 2669
+  #pragma diag_suppress 2670
+  #pragma diag_suppress 2671
+  #pragma diag_suppress 2735
+  #pragma diag_suppress 2737
+#endif
+
 #endif // not EIGEN_WARNINGS_DISABLED
diff --git a/eigen/Eigen/src/Core/util/ForwardDeclarations.h b/eigen/Eigen/src/Core/util/ForwardDeclarations.h
index f277720..1a48cff 100644
--- a/eigen/Eigen/src/Core/util/ForwardDeclarations.h
+++ b/eigen/Eigen/src/Core/util/ForwardDeclarations.h
@@ -36,6 +36,10 @@ template<typename Derived> struct accessors_level
   };
 };
 
+template<typename T> struct evaluator_traits;
+
+template< typename T> struct evaluator;
+
 } // end namespace internal
 
 template<typename T> struct NumTraits;
@@ -51,18 +55,18 @@ class DenseCoeffsBase;
 
 template<typename _Scalar, int _Rows, int _Cols,
          int _Options = AutoAlign |
-#if defined(__GNUC__) && __GNUC__==3 && __GNUC_MINOR__==4
+#if EIGEN_GNUC_AT(3,4)
     // workaround a bug in at least gcc 3.4.6
     // the innermost ?: ternary operator is misparsed. We write it slightly
     // differently and this makes gcc 3.4.6 happy, but it's ugly.
     // The error would only show up with EIGEN_DEFAULT_TO_ROW_MAJOR is defined
     // (when EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION is RowMajor)
-                          ( (_Rows==1 && _Cols!=1) ? RowMajor
+                          ( (_Rows==1 && _Cols!=1) ? Eigen::RowMajor
                           : !(_Cols==1 && _Rows!=1) ?  EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION
-                          : ColMajor ),
+                          : Eigen::ColMajor ),
 #else
-                          ( (_Rows==1 && _Cols!=1) ? RowMajor
-                          : (_Cols==1 && _Rows!=1) ? ColMajor
+                          ( (_Rows==1 && _Cols!=1) ? Eigen::RowMajor
+                          : (_Cols==1 && _Rows!=1) ? Eigen::ColMajor
                           : EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION ),
 #endif
          int _MaxRows = _Rows,
@@ -79,6 +83,7 @@ template<typename ExpressionType> class ForceAlignedAccess;
 template<typename ExpressionType> class SwapWrapper;
 
 template<typename XprType, int BlockRows=Dynamic, int BlockCols=Dynamic, bool InnerPanel = false> class Block;
+template<typename XprType, typename RowIndices, typename ColIndices> class IndexedView;
 
 template<typename MatrixType, int Size=Dynamic> class VectorBlock;
 template<typename MatrixType> class Transpose;
@@ -87,10 +92,11 @@ template<typename NullaryOp, typename MatrixType>         class CwiseNullaryOp;
 template<typename UnaryOp,   typename MatrixType>         class CwiseUnaryOp;
 template<typename ViewOp,    typename MatrixType>         class CwiseUnaryView;
 template<typename BinaryOp,  typename Lhs, typename Rhs>  class CwiseBinaryOp;
-template<typename BinOp,     typename Lhs, typename Rhs>  class SelfCwiseBinaryOp;
-template<typename Derived,   typename Lhs, typename Rhs>  class ProductBase;
-template<typename Lhs, typename Rhs, int Mode>            class GeneralProduct;
-template<typename Lhs, typename Rhs, int NestingFlags>    class CoeffBasedProduct;
+template<typename TernaryOp, typename Arg1, typename Arg2, typename Arg3>  class CwiseTernaryOp;
+template<typename Decomposition, typename Rhstype>        class Solve;
+template<typename XprType>                                class Inverse;
+
+template<typename Lhs, typename Rhs, int Option = DefaultProduct> class Product;
 
 template<typename Derived> class DiagonalBase;
 template<typename _DiagonalVectorType> class DiagonalWrapper;
@@ -108,7 +114,12 @@ template<typename Derived,
          int Level = internal::accessors_level<Derived>::has_write_access ? WriteAccessors : ReadOnlyAccessors
 > class MapBase;
 template<int InnerStrideAtCompileTime, int OuterStrideAtCompileTime> class Stride;
+template<int Value = Dynamic> class InnerStride;
+template<int Value = Dynamic> class OuterStride;
 template<typename MatrixType, int MapOptions=Unaligned, typename StrideType = Stride<0,0> > class Map;
+template<typename Derived> class RefBase;
+template<typename PlainObjectType, int Options = 0,
+         typename StrideType = typename internal::conditional<PlainObjectType::IsVectorAtCompileTime,InnerStride<1>,OuterStride<> >::type > class Ref;
 
 template<typename Derived> class TriangularBase;
 template<typename MatrixType, unsigned int Mode> class TriangularView;
@@ -119,10 +130,10 @@ template<typename MatrixType> struct CommaInitializer;
 template<typename Derived> class ReturnByValue;
 template<typename ExpressionType> class ArrayWrapper;
 template<typename ExpressionType> class MatrixWrapper;
+template<typename Derived> class SolverBase;
+template<typename XprType> class InnerIterator;
 
 namespace internal {
-template<typename DecompositionType, typename Rhs> struct solve_retval_base;
-template<typename DecompositionType, typename Rhs> struct solve_retval;
 template<typename DecompositionType> struct kernel_retval_base;
 template<typename DecompositionType> struct kernel_retval;
 template<typename DecompositionType> struct image_retval_base;
@@ -135,6 +146,21 @@ template<typename _Scalar, int Rows=Dynamic, int Cols=Dynamic, int Supers=Dynami
 
 namespace internal {
 template<typename Lhs, typename Rhs> struct product_type;
+
+template<bool> struct EnableIf;
+
+/** \internal
+  * \class product_evaluator
+  * Products need their own evaluator with more template arguments allowing for
+  * easier partial template specializations.
+  */
+template< typename T,
+          int ProductTag = internal::product_type<typename T::Lhs,typename T::Rhs>::ret,
+          typename LhsShape = typename evaluator_traits<typename T::Lhs>::Shape,
+          typename RhsShape = typename evaluator_traits<typename T::Rhs>::Shape,
+          typename LhsScalar = typename traits<typename T::Lhs>::Scalar,
+          typename RhsScalar = typename traits<typename T::Rhs>::Scalar
+        > struct product_evaluator;
 }
 
 template<typename Lhs, typename Rhs,
@@ -150,9 +176,11 @@ namespace internal {
 // with optional conjugation of the arguments.
 template<typename LhsScalar, typename RhsScalar, bool ConjLhs=false, bool ConjRhs=false> struct conj_helper;
 
-template<typename Scalar> struct scalar_sum_op;
-template<typename Scalar> struct scalar_difference_op;
-template<typename LhsScalar,typename RhsScalar> struct scalar_conj_product_op;
+template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_sum_op;
+template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_difference_op;
+template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_conj_product_op;
+template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_min_op;
+template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_max_op;
 template<typename Scalar> struct scalar_opposite_op;
 template<typename Scalar> struct scalar_conjugate_op;
 template<typename Scalar> struct scalar_real_op;
@@ -160,6 +188,7 @@ template<typename Scalar> struct scalar_imag_op;
 template<typename Scalar> struct scalar_abs_op;
 template<typename Scalar> struct scalar_abs2_op;
 template<typename Scalar> struct scalar_sqrt_op;
+template<typename Scalar> struct scalar_rsqrt_op;
 template<typename Scalar> struct scalar_exp_op;
 template<typename Scalar> struct scalar_log_op;
 template<typename Scalar> struct scalar_cos_op;
@@ -167,24 +196,29 @@ template<typename Scalar> struct scalar_sin_op;
 template<typename Scalar> struct scalar_acos_op;
 template<typename Scalar> struct scalar_asin_op;
 template<typename Scalar> struct scalar_tan_op;
-template<typename Scalar> struct scalar_pow_op;
 template<typename Scalar> struct scalar_inverse_op;
 template<typename Scalar> struct scalar_square_op;
 template<typename Scalar> struct scalar_cube_op;
 template<typename Scalar, typename NewType> struct scalar_cast_op;
-template<typename Scalar> struct scalar_multiple_op;
-template<typename Scalar> struct scalar_quotient1_op;
-template<typename Scalar> struct scalar_min_op;
-template<typename Scalar> struct scalar_max_op;
 template<typename Scalar> struct scalar_random_op;
-template<typename Scalar> struct scalar_add_op;
 template<typename Scalar> struct scalar_constant_op;
 template<typename Scalar> struct scalar_identity_op;
-
+template<typename Scalar,bool iscpx> struct scalar_sign_op;
+template<typename Scalar,typename ScalarExponent> struct scalar_pow_op;
+template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_hypot_op;
 template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_product_op;
-template<typename LhsScalar,typename RhsScalar> struct scalar_multiple2_op;
 template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_quotient_op;
 
+// SpecialFunctions module
+template<typename Scalar> struct scalar_lgamma_op;
+template<typename Scalar> struct scalar_digamma_op;
+template<typename Scalar> struct scalar_erf_op;
+template<typename Scalar> struct scalar_erfc_op;
+template<typename Scalar> struct scalar_igamma_op;
+template<typename Scalar> struct scalar_igammac_op;
+template<typename Scalar> struct scalar_zeta_op;
+template<typename Scalar> struct scalar_betainc_op;
+
 } // end namespace internal
 
 struct IOFormat;
@@ -192,18 +226,18 @@ struct IOFormat;
 // Array module
 template<typename _Scalar, int _Rows, int _Cols,
          int _Options = AutoAlign |
-#if defined(__GNUC__) && __GNUC__==3 && __GNUC_MINOR__==4
+#if EIGEN_GNUC_AT(3,4)
     // workaround a bug in at least gcc 3.4.6
     // the innermost ?: ternary operator is misparsed. We write it slightly
     // differently and this makes gcc 3.4.6 happy, but it's ugly.
     // The error would only show up with EIGEN_DEFAULT_TO_ROW_MAJOR is defined
     // (when EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION is RowMajor)
-                          ( (_Rows==1 && _Cols!=1) ? RowMajor
+                          ( (_Rows==1 && _Cols!=1) ? Eigen::RowMajor
                           : !(_Cols==1 && _Rows!=1) ?  EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION
-                          : ColMajor ),
+                          : Eigen::ColMajor ),
 #else
-                          ( (_Rows==1 && _Cols!=1) ? RowMajor
-                          : (_Cols==1 && _Rows!=1) ? ColMajor
+                          ( (_Rows==1 && _Cols!=1) ? Eigen::RowMajor
+                          : (_Cols==1 && _Rows!=1) ? Eigen::ColMajor
                           : EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION ),
 #endif
          int _MaxRows = _Rows, int _MaxCols = _Cols> class Array;
@@ -221,7 +255,9 @@ template<typename MatrixType> struct inverse_impl;
 template<typename MatrixType> class HouseholderQR;
 template<typename MatrixType> class ColPivHouseholderQR;
 template<typename MatrixType> class FullPivHouseholderQR;
+template<typename MatrixType> class CompleteOrthogonalDecomposition;
 template<typename MatrixType, int QRPreconditioner = ColPivHouseholderQRPreconditioner> class JacobiSVD;
+template<typename MatrixType> class BDCSVD;
 template<typename MatrixType, int UpLo = Lower> class LLT;
 template<typename MatrixType, int UpLo = Lower> class LDLT;
 template<typename VectorsType, typename CoeffsType, int Side=OnTheLeft> class HouseholderSequence;
@@ -234,39 +270,16 @@ template<typename Derived> class QuaternionBase;
 template<typename Scalar> class Rotation2D;
 template<typename Scalar> class AngleAxis;
 template<typename Scalar,int Dim> class Translation;
-
-// Sparse module:
-template<typename Derived> class SparseMatrixBase;
-
-#ifdef EIGEN2_SUPPORT
-template<typename Derived, int _Dim> class eigen2_RotationBase;
-template<typename Lhs, typename Rhs> class eigen2_Cross;
-template<typename Scalar> class eigen2_Quaternion;
-template<typename Scalar> class eigen2_Rotation2D;
-template<typename Scalar> class eigen2_AngleAxis;
-template<typename Scalar,int Dim> class eigen2_Transform;
-template <typename _Scalar, int _AmbientDim> class eigen2_ParametrizedLine;
-template <typename _Scalar, int _AmbientDim> class eigen2_Hyperplane;
-template<typename Scalar,int Dim> class eigen2_Translation;
-template<typename Scalar,int Dim> class eigen2_Scaling;
-#endif
-
-#if EIGEN2_SUPPORT_STAGE < STAGE20_RESOLVE_API_CONFLICTS
-template<typename Scalar> class Quaternion;
-template<typename Scalar,int Dim> class Transform;
-template <typename _Scalar, int _AmbientDim> class ParametrizedLine;
-template <typename _Scalar, int _AmbientDim> class Hyperplane;
-template<typename Scalar,int Dim> class Scaling;
-#endif
-
-#if EIGEN2_SUPPORT_STAGE > STAGE20_RESOLVE_API_CONFLICTS
+template<typename Scalar,int Dim> class AlignedBox;
 template<typename Scalar, int Options = AutoAlign> class Quaternion;
 template<typename Scalar,int Dim,int Mode,int _Options=AutoAlign> class Transform;
 template <typename _Scalar, int _AmbientDim, int Options=AutoAlign> class ParametrizedLine;
 template <typename _Scalar, int _AmbientDim, int Options=AutoAlign> class Hyperplane;
 template<typename Scalar> class UniformScaling;
 template<typename MatrixType,int Direction> class Homogeneous;
-#endif
+
+// Sparse module:
+template<typename Derived> class SparseMatrixBase;
 
 // MatrixFunctions module
 template<typename Derived> struct MatrixExponentialReturnValue;
@@ -274,7 +287,7 @@ template<typename Derived> class MatrixFunctionReturnValue;
 template<typename Derived> class MatrixSquareRootReturnValue;
 template<typename Derived> class MatrixLogarithmReturnValue;
 template<typename Derived> class MatrixPowerReturnValue;
-template<typename Derived, typename Lhs, typename Rhs> class MatrixPowerProduct;
+template<typename Derived> class MatrixComplexPowerReturnValue;
 
 namespace internal {
 template <typename Scalar>
@@ -285,18 +298,6 @@ struct stem_function
 };
 }
 
-
-#ifdef EIGEN2_SUPPORT
-template<typename ExpressionType> class Cwise;
-template<typename MatrixType> class Minor;
-template<typename MatrixType> class LU;
-template<typename MatrixType> class QR;
-template<typename MatrixType> class SVD;
-namespace internal {
-template<typename MatrixType, unsigned int Mode> struct eigen2_part_return_type;
-}
-#endif
-
 } // end namespace Eigen
 
 #endif // EIGEN_FORWARDDECLARATIONS_H
diff --git a/eigen/Eigen/src/Core/util/IndexedViewHelper.h b/eigen/Eigen/src/Core/util/IndexedViewHelper.h
new file mode 100644
index 0000000..ab01c85
--- /dev/null
+++ b/eigen/Eigen/src/Core/util/IndexedViewHelper.h
@@ -0,0 +1,187 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2017 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+
+#ifndef EIGEN_INDEXED_VIEW_HELPER_H
+#define EIGEN_INDEXED_VIEW_HELPER_H
+
+namespace Eigen {
+
+/** \namespace Eigen::placeholders
+  * \ingroup Core_Module
+  *
+  * Namespace containing symbolic placeholder and identifiers
+  */
+namespace placeholders {
+
+namespace internal {
+struct symbolic_last_tag {};
+}
+
+/** \var last
+  * \ingroup Core_Module
+  *
+  * Can be used as a parameter to Eigen::seq and Eigen::seqN functions to symbolically reference the last element/row/columns
+  * of the underlying vector or matrix once passed to DenseBase::operator()(const RowIndices&, const ColIndices&).
+  *
+  * This symbolic placeholder support standard arithmetic operation.
+  *
+  * A typical usage example would be:
+  * \code
+  * using namespace Eigen;
+  * using Eigen::placeholders::last;
+  * VectorXd v(n);
+  * v(seq(2,last-2)).setOnes();
+  * \endcode
+  *
+  * \sa end
+  */
+static const Symbolic::SymbolExpr<internal::symbolic_last_tag> last;
+
+/** \var end
+  * \ingroup Core_Module
+  *
+  * Can be used as a parameter to Eigen::seq and Eigen::seqN functions to symbolically reference the last+1 element/row/columns
+  * of the underlying vector or matrix once passed to DenseBase::operator()(const RowIndices&, const ColIndices&).
+  *
+  * This symbolic placeholder support standard arithmetic operation.
+  * It is essentially an alias to last+1
+  *
+  * \sa last
+  */
+#ifdef EIGEN_PARSED_BY_DOXYGEN
+static const auto end = last+1;
+#else
+// Using a FixedExpr<1> expression is important here to make sure the compiler
+// can fully optimize the computation starting indices with zero overhead.
+static const Symbolic::AddExpr<Symbolic::SymbolExpr<internal::symbolic_last_tag>,Symbolic::ValueExpr<Eigen::internal::FixedInt<1> > > end(last+fix<1>());
+#endif
+
+} // end namespace placeholders
+
+namespace internal {
+
+ // Replace symbolic last/end "keywords" by their true runtime value
+inline Index eval_expr_given_size(Index x, Index /* size */)   { return x; }
+
+template<int N>
+FixedInt<N> eval_expr_given_size(FixedInt<N> x, Index /*size*/)   { return x; }
+
+template<typename Derived>
+Index eval_expr_given_size(const Symbolic::BaseExpr<Derived> &x, Index size)
+{
+  return x.derived().eval(placeholders::last=size-1);
+}
+
+// Extract increment/step at compile time
+template<typename T, typename EnableIf = void> struct get_compile_time_incr {
+  enum { value = UndefinedIncr };
+};
+
+// Analogue of std::get<0>(x), but tailored for our needs.
+template<typename T>
+Index first(const T& x) { return x.first(); }
+
+// IndexedViewCompatibleType/makeIndexedViewCompatible turn an arbitrary object of type T into something usable by MatrixSlice
+// The generic implementation is a no-op
+template<typename T,int XprSize,typename EnableIf=void>
+struct IndexedViewCompatibleType {
+  typedef T type;
+};
+
+template<typename T,typename Q>
+const T& makeIndexedViewCompatible(const T& x, Index /*size*/, Q) { return x; }
+
+//--------------------------------------------------------------------------------
+// Handling of a single Index
+//--------------------------------------------------------------------------------
+
+struct SingleRange {
+  enum {
+    SizeAtCompileTime = 1
+  };
+  SingleRange(Index val) : m_value(val) {}
+  Index operator[](Index) const { return m_value; }
+  Index size() const { return 1; }
+  Index first() const { return m_value; }
+  Index m_value;
+};
+
+template<> struct get_compile_time_incr<SingleRange> {
+  enum { value = 1 }; // 1 or 0 ??
+};
+
+// Turn a single index into something that looks like an array (i.e., that exposes a .size(), and operatro[](int) methods)
+template<typename T, int XprSize>
+struct IndexedViewCompatibleType<T,XprSize,typename internal::enable_if<internal::is_integral<T>::value>::type> {
+  // Here we could simply use Array, but maybe it's less work for the compiler to use
+  // a simpler wrapper as SingleRange
+  //typedef Eigen::Array<Index,1,1> type;
+  typedef SingleRange type;
+};
+
+template<typename T, int XprSize>
+struct IndexedViewCompatibleType<T, XprSize, typename enable_if<Symbolic::is_symbolic<T>::value>::type> {
+  typedef SingleRange type;
+};
+
+
+template<typename T>
+typename enable_if<Symbolic::is_symbolic<T>::value,SingleRange>::type
+makeIndexedViewCompatible(const T& id, Index size, SpecializedType) {
+  return eval_expr_given_size(id,size);
+}
+
+//--------------------------------------------------------------------------------
+// Handling of all
+//--------------------------------------------------------------------------------
+
+struct all_t { all_t() {} };
+
+// Convert a symbolic 'all' into a usable range type
+template<int XprSize>
+struct AllRange {
+  enum { SizeAtCompileTime = XprSize };
+  AllRange(Index size = XprSize) : m_size(size) {}
+  Index operator[](Index i) const { return i; }
+  Index size() const { return m_size.value(); }
+  Index first() const { return 0; }
+  variable_if_dynamic<Index,XprSize> m_size;
+};
+
+template<int XprSize>
+struct IndexedViewCompatibleType<all_t,XprSize> {
+  typedef AllRange<XprSize> type;
+};
+
+template<typename XprSizeType>
+inline AllRange<get_fixed_value<XprSizeType>::value> makeIndexedViewCompatible(all_t , XprSizeType size, SpecializedType) {
+  return AllRange<get_fixed_value<XprSizeType>::value>(size);
+}
+
+template<int Size> struct get_compile_time_incr<AllRange<Size> > {
+  enum { value = 1 };
+};
+
+} // end namespace internal
+
+
+namespace placeholders {
+
+/** \var all
+  * \ingroup Core_Module
+  * Can be used as a parameter to DenseBase::operator()(const RowIndices&, const ColIndices&) to index all rows or columns
+  */
+static const Eigen::internal::all_t all;
+
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_INDEXED_VIEW_HELPER_H
diff --git a/eigen/Eigen/src/Core/util/IntegralConstant.h b/eigen/Eigen/src/Core/util/IntegralConstant.h
new file mode 100644
index 0000000..78a4705
--- /dev/null
+++ b/eigen/Eigen/src/Core/util/IntegralConstant.h
@@ -0,0 +1,270 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2017 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+
+#ifndef EIGEN_INTEGRAL_CONSTANT_H
+#define EIGEN_INTEGRAL_CONSTANT_H
+
+namespace Eigen {
+
+namespace internal {
+
+template<int N> class FixedInt;
+template<int N> class VariableAndFixedInt;
+
+/** \internal
+  * \class FixedInt
+  *
+  * This class embeds a compile-time integer \c N.
+  *
+  * It is similar to c++11 std::integral_constant<int,N> but with some additional features
+  * such as:
+  *  - implicit conversion to int
+  *  - arithmetic and some bitwise operators: -, +, *, /, %, &, |
+  *  - c++98/14 compatibility with fix<N> and fix<N>() syntax to define integral constants.
+  *
+  * It is strongly discouraged to directly deal with this class FixedInt. Instances are expcected to
+  * be created by the user using Eigen::fix<N> or Eigen::fix<N>(). In C++98-11, the former syntax does
+  * not create a FixedInt<N> instance but rather a point to function that needs to be \em cleaned-up
+  * using the generic helper:
+  * \code
+  * internal::cleanup_index_type<T>::type
+  * internal::cleanup_index_type<T,DynamicKey>::type
+  * \endcode
+  * where T can a FixedInt<N>, a pointer to function FixedInt<N> (*)(), or numerous other integer-like representations.
+  * \c DynamicKey is either Dynamic (default) or DynamicIndex and used to identify true compile-time values.
+  *
+  * For convenience, you can extract the compile-time value \c N in a generic way using the following helper:
+  * \code
+  * internal::get_fixed_value<T,DefaultVal>::value
+  * \endcode
+  * that will give you \c N if T equals FixedInt<N> or FixedInt<N> (*)(), and \c DefaultVal if T does not embed any compile-time value (e.g., T==int).
+  *
+  * \sa fix<N>, class VariableAndFixedInt
+  */
+template<int N> class FixedInt
+{
+public:
+  static const int value = N;
+  operator int() const { return value; }
+  FixedInt() {}
+  FixedInt( VariableAndFixedInt<N> other) {
+    EIGEN_ONLY_USED_FOR_DEBUG(other);
+    eigen_internal_assert(int(other)==N);
+  }
+
+  FixedInt<-N> operator-() const { return FixedInt<-N>(); }
+  template<int M>
+  FixedInt<N+M> operator+( FixedInt<M>) const { return FixedInt<N+M>(); }
+  template<int M>
+  FixedInt<N-M> operator-( FixedInt<M>) const { return FixedInt<N-M>(); }
+  template<int M>
+  FixedInt<N*M> operator*( FixedInt<M>) const { return FixedInt<N*M>(); }
+  template<int M>
+  FixedInt<N/M> operator/( FixedInt<M>) const { return FixedInt<N/M>(); }
+  template<int M>
+  FixedInt<N%M> operator%( FixedInt<M>) const { return FixedInt<N%M>(); }
+  template<int M>
+  FixedInt<N|M> operator|( FixedInt<M>) const { return FixedInt<N|M>(); }
+  template<int M>
+  FixedInt<N&M> operator&( FixedInt<M>) const { return FixedInt<N&M>(); }
+
+#if EIGEN_HAS_CXX14
+  // Needed in C++14 to allow fix<N>():
+  FixedInt operator() () const { return *this; }
+
+  VariableAndFixedInt<N> operator() (int val) const { return VariableAndFixedInt<N>(val); }
+#else
+  FixedInt ( FixedInt<N> (*)() ) {}
+#endif
+
+#if EIGEN_HAS_CXX11
+  FixedInt(std::integral_constant<int,N>) {}
+#endif
+};
+
+/** \internal
+  * \class VariableAndFixedInt
+  *
+  * This class embeds both a compile-time integer \c N and a runtime integer.
+  * Both values are supposed to be equal unless the compile-time value \c N has a special
+  * value meaning that the runtime-value should be used. Depending on the context, this special
+  * value can be either Eigen::Dynamic (for positive quantities) or Eigen::DynamicIndex (for
+  * quantities that can be negative).
+  *
+  * It is the return-type of the function Eigen::fix<N>(int), and most of the time this is the only
+  * way it is used. It is strongly discouraged to directly deal with instances of VariableAndFixedInt.
+  * Indeed, in order to write generic code, it is the responsibility of the callee to properly convert
+  * it to either a true compile-time quantity (i.e. a FixedInt<N>), or to a runtime quantity (e.g., an Index)
+  * using the following generic helper:
+  * \code
+  * internal::cleanup_index_type<T>::type
+  * internal::cleanup_index_type<T,DynamicKey>::type
+  * \endcode
+  * where T can be a template instantiation of VariableAndFixedInt or numerous other integer-like representations.
+  * \c DynamicKey is either Dynamic (default) or DynamicIndex and used to identify true compile-time values.
+  *
+  * For convenience, you can also extract the compile-time value \c N using the following helper:
+  * \code
+  * internal::get_fixed_value<T,DefaultVal>::value
+  * \endcode
+  * that will give you \c N if T equals VariableAndFixedInt<N>, and \c DefaultVal if T does not embed any compile-time value (e.g., T==int).
+  *
+  * \sa fix<N>(int), class FixedInt
+  */
+template<int N> class VariableAndFixedInt
+{
+public:
+  static const int value = N;
+  operator int() const { return m_value; }
+  VariableAndFixedInt(int val) { m_value = val; }
+protected:
+  int m_value;
+};
+
+template<typename T, int Default=Dynamic> struct get_fixed_value {
+  static const int value = Default;
+};
+
+template<int N,int Default> struct get_fixed_value<FixedInt<N>,Default> {
+  static const int value = N;
+};
+
+#if !EIGEN_HAS_CXX14
+template<int N,int Default> struct get_fixed_value<FixedInt<N> (*)(),Default> {
+  static const int value = N;
+};
+#endif
+
+template<int N,int Default> struct get_fixed_value<VariableAndFixedInt<N>,Default> {
+  static const int value = N ;
+};
+
+template<typename T, int N, int Default>
+struct get_fixed_value<variable_if_dynamic<T,N>,Default> {
+  static const int value = N;
+};
+
+template<typename T> EIGEN_DEVICE_FUNC Index get_runtime_value(const T &x) { return x; }
+#if !EIGEN_HAS_CXX14
+template<int N> EIGEN_DEVICE_FUNC Index get_runtime_value(FixedInt<N> (*)()) { return N; }
+#endif
+
+// Cleanup integer/FixedInt/VariableAndFixedInt/etc types:
+
+// By default, no cleanup:
+template<typename T, int DynamicKey=Dynamic, typename EnableIf=void> struct cleanup_index_type { typedef T type; };
+
+// Convert any integral type (e.g., short, int, unsigned int, etc.) to Eigen::Index
+template<typename T, int DynamicKey> struct cleanup_index_type<T,DynamicKey,typename internal::enable_if<internal::is_integral<T>::value>::type> { typedef Index type; };
+
+#if !EIGEN_HAS_CXX14
+// In c++98/c++11, fix<N> is a pointer to function that we better cleanup to a true FixedInt<N>:
+template<int N, int DynamicKey> struct cleanup_index_type<FixedInt<N> (*)(), DynamicKey> { typedef FixedInt<N> type; };
+#endif
+
+// If VariableAndFixedInt does not match DynamicKey, then we turn it to a pure compile-time value:
+template<int N, int DynamicKey> struct cleanup_index_type<VariableAndFixedInt<N>, DynamicKey> { typedef FixedInt<N> type; };
+// If VariableAndFixedInt matches DynamicKey, then we turn it to a pure runtime-value (aka Index):
+template<int DynamicKey> struct cleanup_index_type<VariableAndFixedInt<DynamicKey>, DynamicKey> { typedef Index type; };
+
+#if EIGEN_HAS_CXX11
+template<int N, int DynamicKey> struct cleanup_index_type<std::integral_constant<int,N>, DynamicKey> { typedef FixedInt<N> type; };
+#endif
+
+} // end namespace internal
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+
+#if EIGEN_HAS_CXX14
+template<int N>
+static const internal::FixedInt<N> fix{};
+#else
+template<int N>
+inline internal::FixedInt<N> fix() { return internal::FixedInt<N>(); }
+
+// The generic typename T is mandatory. Otherwise, a code like fix<N> could refer to either the function above or this next overload.
+// This way a code like fix<N> can only refer to the previous function.
+template<int N,typename T>
+inline internal::VariableAndFixedInt<N> fix(T val) { return internal::VariableAndFixedInt<N>(val); }
+#endif
+
+#else // EIGEN_PARSED_BY_DOXYGEN
+
+/** \var fix<N>()
+  * \ingroup Core_Module
+  *
+  * This \em identifier permits to construct an object embedding a compile-time integer \c N.
+  *
+  * \tparam N the compile-time integer value
+  *
+  * It is typically used in conjunction with the Eigen::seq and Eigen::seqN functions to pass compile-time values to them:
+  * \code
+  * seqN(10,fix<4>,fix<-3>)   // <=> [10 7 4 1]
+  * \endcode
+  *
+  * See also the function fix(int) to pass both a compile-time and runtime value.
+  *
+  * In c++14, it is implemented as:
+  * \code
+  * template<int N> static const internal::FixedInt<N> fix{};
+  * \endcode
+  * where internal::FixedInt<N> is an internal template class similar to
+  * <a href="http://en.cppreference.com/w/cpp/types/integral_constant">\c std::integral_constant </a><tt> <int,N> </tt>
+  * Here, \c fix<N> is thus an object of type \c internal::FixedInt<N>.
+  *
+  * In c++98/11, it is implemented as a function:
+  * \code
+  * template<int N> inline internal::FixedInt<N> fix();
+  * \endcode
+  * Here internal::FixedInt<N> is thus a pointer to function.
+  *
+  * If for some reason you want a true object in c++98 then you can write: \code fix<N>() \endcode which is also valid in c++14.
+  *
+  * \sa fix<N>(int), seq, seqN
+  */
+template<int N>
+static const auto fix();
+
+/** \fn fix<N>(int)
+  * \ingroup Core_Module
+  *
+  * This function returns an object embedding both a compile-time integer \c N, and a fallback runtime value \a val.
+  *
+  * \tparam N the compile-time integer value
+  * \param  val the fallback runtime integer value
+  *
+  * This function is a more general version of the \ref fix identifier/function that can be used in template code
+  * where the compile-time value could turn out to actually mean "undefined at compile-time". For positive integers
+  * such as a size or a dimension, this case is identified by Eigen::Dynamic, whereas runtime signed integers
+  * (e.g., an increment/stride) are identified as Eigen::DynamicIndex. In such a case, the runtime value \a val
+  * will be used as a fallback.
+  *
+  * A typical use case would be:
+  * \code
+  * template<typename Derived> void foo(const MatrixBase<Derived> &mat) {
+  *   const int N = Derived::RowsAtCompileTime==Dynamic ? Dynamic : Derived::RowsAtCompileTime/2;
+  *   const int n = mat.rows()/2;
+  *   ... mat( seqN(0,fix<N>(n) ) ...;
+  * }
+  * \endcode
+  * In this example, the function Eigen::seqN knows that the second argument is expected to be a size.
+  * If the passed compile-time value N equals Eigen::Dynamic, then the proxy object returned by fix will be dissmissed, and converted to an Eigen::Index of value \c n.
+  * Otherwise, the runtime-value \c n will be dissmissed, and the returned ArithmeticSequence will be of the exact same type as <tt> seqN(0,fix<N>) </tt>.
+  *
+  * \sa fix, seqN, class ArithmeticSequence
+  */
+template<int N>
+static const auto fix(int val);
+
+#endif // EIGEN_PARSED_BY_DOXYGEN
+
+} // end namespace Eigen
+
+#endif // EIGEN_INTEGRAL_CONSTANT_H
diff --git a/eigen/Eigen/src/Core/util/MKL_support.h b/eigen/Eigen/src/Core/util/MKL_support.h
index 1ef3b61..26b5966 100644
--- a/eigen/Eigen/src/Core/util/MKL_support.h
+++ b/eigen/Eigen/src/Core/util/MKL_support.h
@@ -49,7 +49,7 @@
   #define EIGEN_USE_LAPACKE
 #endif
 
-#if defined(EIGEN_USE_BLAS) || defined(EIGEN_USE_LAPACKE) || defined(EIGEN_USE_MKL_VML)
+#if defined(EIGEN_USE_MKL_VML)
   #define EIGEN_USE_MKL
 #endif
 
@@ -64,7 +64,6 @@
 #   ifndef EIGEN_USE_MKL
     /*If the MKL version is too old, undef everything*/
 #       undef   EIGEN_USE_MKL_ALL
-#       undef   EIGEN_USE_BLAS
 #       undef   EIGEN_USE_LAPACKE
 #       undef   EIGEN_USE_MKL_VML
 #       undef   EIGEN_USE_LAPACKE_STRICT
@@ -73,7 +72,7 @@
 #endif
 
 #if defined EIGEN_USE_MKL
-#include <mkl_lapacke.h>
+
 #define EIGEN_MKL_VML_THRESHOLD 128
 
 /* MKL_DOMAIN_BLAS, etc are defined only in 10.3 update 7 */
@@ -107,52 +106,23 @@
 #else
 #define EIGEN_MKL_DOMAIN_PARDISO MKL_PARDISO
 #endif
+#endif
 
 namespace Eigen {
 
 typedef std::complex<double> dcomplex;
 typedef std::complex<float>  scomplex;
 
-namespace internal {
-
-template<typename MKLType, typename EigenType>
-static inline void assign_scalar_eig2mkl(MKLType& mklScalar, const EigenType& eigenScalar) {
-  mklScalar=eigenScalar;
-}
-
-template<typename MKLType, typename EigenType>
-static inline void assign_conj_scalar_eig2mkl(MKLType& mklScalar, const EigenType& eigenScalar) {
-  mklScalar=eigenScalar;
-}
-
-template <>
-inline void assign_scalar_eig2mkl<MKL_Complex16,dcomplex>(MKL_Complex16& mklScalar, const dcomplex& eigenScalar) {
-  mklScalar.real=eigenScalar.real();
-  mklScalar.imag=eigenScalar.imag();
-}
-
-template <>
-inline void assign_scalar_eig2mkl<MKL_Complex8,scomplex>(MKL_Complex8& mklScalar, const scomplex& eigenScalar) {
-  mklScalar.real=eigenScalar.real();
-  mklScalar.imag=eigenScalar.imag();
-}
-
-template <>
-inline void assign_conj_scalar_eig2mkl<MKL_Complex16,dcomplex>(MKL_Complex16& mklScalar, const dcomplex& eigenScalar) {
-  mklScalar.real=eigenScalar.real();
-  mklScalar.imag=-eigenScalar.imag();
-}
-
-template <>
-inline void assign_conj_scalar_eig2mkl<MKL_Complex8,scomplex>(MKL_Complex8& mklScalar, const scomplex& eigenScalar) {
-  mklScalar.real=eigenScalar.real();
-  mklScalar.imag=-eigenScalar.imag();
-}
-
-} // end namespace internal
+#if defined(EIGEN_USE_MKL)
+typedef MKL_INT BlasIndex;
+#else
+typedef int BlasIndex;
+#endif
 
 } // end namespace Eigen
 
+#if defined(EIGEN_USE_BLAS)
+#include "../../misc/blas.h"
 #endif
 
 #endif // EIGEN_MKL_SUPPORT_H
diff --git a/eigen/Eigen/src/Core/util/Macros.h b/eigen/Eigen/src/Core/util/Macros.h
index 16c248c..14ec87d 100644
--- a/eigen/Eigen/src/Core/util/Macros.h
+++ b/eigen/Eigen/src/Core/util/Macros.h
@@ -1,7 +1,7 @@
 // This file is part of Eigen, a lightweight C++ template library
 // for linear algebra.
 //
-// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2008-2015 Gael Guennebaud <gael.guennebaud@inria.fr>
 // Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
 //
 // This Source Code Form is subject to the terms of the Mozilla
@@ -12,26 +12,25 @@
 #define EIGEN_MACROS_H
 
 #define EIGEN_WORLD_VERSION 3
-#define EIGEN_MAJOR_VERSION 2
-#define EIGEN_MINOR_VERSION 9
+#define EIGEN_MAJOR_VERSION 3
+#define EIGEN_MINOR_VERSION 90
 
 #define EIGEN_VERSION_AT_LEAST(x,y,z) (EIGEN_WORLD_VERSION>x || (EIGEN_WORLD_VERSION>=x && \
                                       (EIGEN_MAJOR_VERSION>y || (EIGEN_MAJOR_VERSION>=y && \
                                                                  EIGEN_MINOR_VERSION>=z))))
 
-
 // Compiler identification, EIGEN_COMP_*
 
 /// \internal EIGEN_COMP_GNUC set to 1 for all compilers compatible with GCC
 #ifdef __GNUC__
-  #define EIGEN_COMP_GNUC 1
+  #define EIGEN_COMP_GNUC (__GNUC__*10+__GNUC_MINOR__)
 #else
   #define EIGEN_COMP_GNUC 0
 #endif
 
-/// \internal EIGEN_COMP_CLANG set to 1 if the compiler is clang (alias for __clang__)
+/// \internal EIGEN_COMP_CLANG set to major+minor version (e.g., 307 for clang 3.7) if the compiler is clang
 #if defined(__clang__)
-  #define EIGEN_COMP_CLANG 1
+  #define EIGEN_COMP_CLANG (__clang_major__*100+__clang_minor__)
 #else
   #define EIGEN_COMP_CLANG 0
 #endif
@@ -72,8 +71,17 @@
   #define EIGEN_COMP_MSVC 0
 #endif
 
-/// \internal EIGEN_COMP_MSVC_STRICT set to 1 if the compiler is really Microsoft Visual C++ and not ,e.g., ICC
-#if EIGEN_COMP_MSVC && !(EIGEN_COMP_ICC)
+// For the record, here is a table summarizing the possible values for EIGEN_COMP_MSVC:
+//  name  ver   MSC_VER
+//  2008    9      1500
+//  2010   10      1600
+//  2012   11      1700
+//  2013   12      1800
+//  2015   14      1900
+//  "15"   15      1900
+
+/// \internal EIGEN_COMP_MSVC_STRICT set to 1 if the compiler is really Microsoft Visual C++ and not ,e.g., ICC or clang-cl
+#if EIGEN_COMP_MSVC && !(EIGEN_COMP_ICC || EIGEN_COMP_LLVM || EIGEN_COMP_CLANG)
   #define EIGEN_COMP_MSVC_STRICT _MSC_VER
 #else
   #define EIGEN_COMP_MSVC_STRICT 0
@@ -100,9 +108,16 @@
   #define EIGEN_COMP_ARM 0
 #endif
 
+/// \internal EIGEN_COMP_ARM set to 1 if the compiler is ARM Compiler
+#if defined(__EMSCRIPTEN__)
+  #define EIGEN_COMP_EMSCRIPTEN 1
+#else
+  #define EIGEN_COMP_EMSCRIPTEN 0
+#endif
+
 
 /// \internal EIGEN_GNUC_STRICT set to 1 if the compiler is really GCC and not a compatible compiler (e.g., ICC, clang, mingw, etc.)
-#if EIGEN_COMP_GNUC && !(EIGEN_COMP_CLANG || EIGEN_COMP_ICC || EIGEN_COMP_MINGW || EIGEN_COMP_PGI || EIGEN_COMP_IBM || EIGEN_COMP_ARM )
+#if EIGEN_COMP_GNUC && !(EIGEN_COMP_CLANG || EIGEN_COMP_ICC || EIGEN_COMP_MINGW || EIGEN_COMP_PGI || EIGEN_COMP_IBM || EIGEN_COMP_ARM || EIGEN_COMP_EMSCRIPTEN)
   #define EIGEN_COMP_GNUC_STRICT 1
 #else
   #define EIGEN_COMP_GNUC_STRICT 0
@@ -299,91 +314,176 @@
 #endif
 
 
-#if EIGEN_GNUC_AT_MOST(4,3) && !defined(__clang__)
+
+#if EIGEN_GNUC_AT_MOST(4,3) && !EIGEN_COMP_CLANG
   // see bug 89
   #define EIGEN_SAFE_TO_USE_STANDARD_ASSERT_MACRO 0
 #else
   #define EIGEN_SAFE_TO_USE_STANDARD_ASSERT_MACRO 1
 #endif
 
-// 16 byte alignment is only useful for vectorization. Since it affects the ABI, we need to enable
-// 16 byte alignment on all platforms where vectorization might be enabled. In theory we could always
-// enable alignment, but it can be a cause of problems on some platforms, so we just disable it in
-// certain common platform (compiler+architecture combinations) to avoid these problems.
-// Only static alignment is really problematic (relies on nonstandard compiler extensions that don't
-// work everywhere, for example don't work on GCC/ARM), try to keep heap alignment even
-// when we have to disable static alignment.
-#if defined(__GNUC__) && !(defined(__i386__) || defined(__x86_64__) || defined(__powerpc__) || defined(__ppc__) || defined(__ia64__))
-#define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 1
+// This macro can be used to prevent from macro expansion, e.g.:
+//   std::max EIGEN_NOT_A_MACRO(a,b)
+#define EIGEN_NOT_A_MACRO
+
+#ifdef EIGEN_DEFAULT_TO_ROW_MAJOR
+#define EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION Eigen::RowMajor
 #else
-#define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 0
+#define EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION Eigen::ColMajor
 #endif
 
-// static alignment is completely disabled with GCC 3, Sun Studio, and QCC/QNX
-#if !EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT \
- && !EIGEN_GCC3_OR_OLDER \
- && !defined(__SUNPRO_CC) \
- && !defined(__QNXNTO__)
-  #define EIGEN_ARCH_WANTS_STACK_ALIGNMENT 1
-#else
-  #define EIGEN_ARCH_WANTS_STACK_ALIGNMENT 0
+#ifndef EIGEN_DEFAULT_DENSE_INDEX_TYPE
+#define EIGEN_DEFAULT_DENSE_INDEX_TYPE std::ptrdiff_t
 #endif
 
-#ifdef EIGEN_DONT_ALIGN
-  #ifndef EIGEN_DONT_ALIGN_STATICALLY
-    #define EIGEN_DONT_ALIGN_STATICALLY
-  #endif
-  #define EIGEN_ALIGN 0
+// Cross compiler wrapper around LLVM's __has_builtin
+#ifdef __has_builtin
+#  define EIGEN_HAS_BUILTIN(x) __has_builtin(x)
 #else
-  #define EIGEN_ALIGN 1
+#  define EIGEN_HAS_BUILTIN(x) 0
 #endif
 
-// EIGEN_ALIGN_STATICALLY is the true test whether we want to align arrays on the stack or not. It takes into account both the user choice to explicitly disable
-// alignment (EIGEN_DONT_ALIGN_STATICALLY) and the architecture config (EIGEN_ARCH_WANTS_STACK_ALIGNMENT). Henceforth, only EIGEN_ALIGN_STATICALLY should be used.
-#if EIGEN_ARCH_WANTS_STACK_ALIGNMENT && !defined(EIGEN_DONT_ALIGN_STATICALLY)
-  #define EIGEN_ALIGN_STATICALLY 1
-#else
-  #define EIGEN_ALIGN_STATICALLY 0
-  #ifndef EIGEN_DISABLE_UNALIGNED_ARRAY_ASSERT
-    #define EIGEN_DISABLE_UNALIGNED_ARRAY_ASSERT
-  #endif
+// A Clang feature extension to determine compiler features.
+// We use it to determine 'cxx_rvalue_references'
+#ifndef __has_feature
+# define __has_feature(x) 0
 #endif
 
-#ifdef EIGEN_DEFAULT_TO_ROW_MAJOR
-#define EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION RowMajor
+// Some old compilers do not support template specializations like:
+// template<typename T,int N> void foo(const T x[N]);
+#if !( EIGEN_COMP_CLANG && ((EIGEN_COMP_CLANG<309) || defined(__apple_build_version__)) || EIGEN_COMP_GNUC_STRICT && EIGEN_COMP_GNUC<49)
+#define EIGEN_HAS_STATIC_ARRAY_TEMPLATE 1
 #else
-#define EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION ColMajor
+#define EIGEN_HAS_STATIC_ARRAY_TEMPLATE 0
 #endif
 
-#ifndef EIGEN_DEFAULT_DENSE_INDEX_TYPE
-#define EIGEN_DEFAULT_DENSE_INDEX_TYPE std::ptrdiff_t
+// Upperbound on the C++ version to use.
+// Expected values are 03, 11, 14, 17, etc.
+// By default, let's use an arbitrarily large C++ version.
+#ifndef EIGEN_MAX_CPP_VER
+#define EIGEN_MAX_CPP_VER 99
 #endif
 
-// A Clang feature extension to determine compiler features.
-// We use it to determine 'cxx_rvalue_references'
-#ifndef __has_feature
-# define __has_feature(x) 0
+#if EIGEN_MAX_CPP_VER>=11 && (defined(__cplusplus) && (__cplusplus >= 201103L) || EIGEN_COMP_MSVC >= 1900)
+#define EIGEN_HAS_CXX11 1
+#else
+#define EIGEN_HAS_CXX11 0
+#endif
+
+#if EIGEN_MAX_CPP_VER>=14 && (defined(__cplusplus) && (__cplusplus > 201103L) || EIGEN_COMP_MSVC >= 1910)
+#define EIGEN_HAS_CXX14 1
+#else
+#define EIGEN_HAS_CXX14 0
 #endif
 
 // Do we support r-value references?
-#if (__has_feature(cxx_rvalue_references) || \
-     (defined(__cplusplus) && __cplusplus >= 201103L) || \
-     (defined(_MSC_VER) && _MSC_VER >= 1600))
-  #define EIGEN_HAVE_RVALUE_REFERENCES
+#ifndef EIGEN_HAS_RVALUE_REFERENCES
+#if EIGEN_MAX_CPP_VER>=11 && \
+    (__has_feature(cxx_rvalue_references) || \
+    (defined(__cplusplus) && __cplusplus >= 201103L) || \
+    (EIGEN_COMP_MSVC >= 1600))
+  #define EIGEN_HAS_RVALUE_REFERENCES 1
+#else
+  #define EIGEN_HAS_RVALUE_REFERENCES 0
+#endif
 #endif
 
+// Does the compiler support C99?
+#ifndef EIGEN_HAS_C99_MATH
+#if EIGEN_MAX_CPP_VER>=11 && \
+    ((defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901))       \
+  || (defined(__GNUC__) && defined(_GLIBCXX_USE_C99)) \
+  || (defined(_LIBCPP_VERSION) && !defined(_MSC_VER)) \
+  || (EIGEN_COMP_MSVC >= 1900) || defined(__SYCL_DEVICE_ONLY__))
+  #define EIGEN_HAS_C99_MATH 1
+#else
+  #define EIGEN_HAS_C99_MATH 0
+#endif
+#endif
 
-// Cross compiler wrapper around LLVM's __has_builtin
-#ifdef __has_builtin
-#  define EIGEN_HAS_BUILTIN(x) __has_builtin(x)
+// Does the compiler support result_of?
+#ifndef EIGEN_HAS_STD_RESULT_OF
+#if EIGEN_MAX_CPP_VER>=11 && ((__has_feature(cxx_lambdas) || (defined(__cplusplus) && __cplusplus >= 201103L)))
+#define EIGEN_HAS_STD_RESULT_OF 1
 #else
-#  define EIGEN_HAS_BUILTIN(x) 0
+#define EIGEN_HAS_STD_RESULT_OF 0
+#endif
+#endif
+
+// Does the compiler support variadic templates?
+#ifndef EIGEN_HAS_VARIADIC_TEMPLATES
+#if EIGEN_MAX_CPP_VER>=11 && (__cplusplus > 199711L || EIGEN_COMP_MSVC >= 1900) \
+  && (!defined(__NVCC__) || !EIGEN_ARCH_ARM_OR_ARM64 || (defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000) )
+    // ^^ Disable the use of variadic templates when compiling with versions of nvcc older than 8.0 on ARM devices:
+    //    this prevents nvcc from crashing when compiling Eigen on Tegra X1
+#define EIGEN_HAS_VARIADIC_TEMPLATES 1
+#elif  EIGEN_MAX_CPP_VER>=11 && (__cplusplus > 199711L || EIGEN_COMP_MSVC >= 1900) && defined(__SYCL_DEVICE_ONLY__)
+#define EIGEN_HAS_VARIADIC_TEMPLATES 1
+#else
+#define EIGEN_HAS_VARIADIC_TEMPLATES 0
+#endif
+#endif
+
+// Does the compiler fully support const expressions? (as in c++14)
+#ifndef EIGEN_HAS_CONSTEXPR
+
+#if defined(__CUDACC__)
+// Const expressions are supported provided that c++11 is enabled and we're using either clang or nvcc 7.5 or above
+#if EIGEN_MAX_CPP_VER>=14 && (__cplusplus > 199711L && defined(__CUDACC_VER__) && (EIGEN_COMP_CLANG || __CUDACC_VER__ >= 70500))
+  #define EIGEN_HAS_CONSTEXPR 1
+#endif
+#elif EIGEN_MAX_CPP_VER>=14 && (__has_feature(cxx_relaxed_constexpr) || (defined(__cplusplus) && __cplusplus >= 201402L) || \
+  (EIGEN_GNUC_AT_LEAST(4,8) && (__cplusplus > 199711L)) || \
+  (EIGEN_COMP_CLANG >= 306 && (__cplusplus > 199711L)))
+#define EIGEN_HAS_CONSTEXPR 1
+#endif
+
+#ifndef EIGEN_HAS_CONSTEXPR
+#define EIGEN_HAS_CONSTEXPR 0
+#endif
+
+#endif
+
+// Does the compiler support C++11 math?
+// Let's be conservative and enable the default C++11 implementation only if we are sure it exists
+#ifndef EIGEN_HAS_CXX11_MATH
+  #if EIGEN_MAX_CPP_VER>=11 && ((__cplusplus > 201103L) || (__cplusplus >= 201103L) && (EIGEN_COMP_GNUC_STRICT || EIGEN_COMP_CLANG || EIGEN_COMP_MSVC || EIGEN_COMP_ICC)  \
+      && (EIGEN_ARCH_i386_OR_x86_64) && (EIGEN_OS_GNULINUX || EIGEN_OS_WIN_STRICT || EIGEN_OS_MAC))
+    #define EIGEN_HAS_CXX11_MATH 1
+  #else
+    #define EIGEN_HAS_CXX11_MATH 0
+  #endif
+#endif
+
+// Does the compiler support proper C++11 containers?
+#ifndef EIGEN_HAS_CXX11_CONTAINERS
+  #if    EIGEN_MAX_CPP_VER>=11 && \
+         ((__cplusplus > 201103L) \
+      || ((__cplusplus >= 201103L) && (EIGEN_COMP_GNUC_STRICT || EIGEN_COMP_CLANG || EIGEN_COMP_ICC>=1400)) \
+      || EIGEN_COMP_MSVC >= 1900)
+    #define EIGEN_HAS_CXX11_CONTAINERS 1
+  #else
+    #define EIGEN_HAS_CXX11_CONTAINERS 0
+  #endif
+#endif
+
+// Does the compiler support C++11 noexcept?
+#ifndef EIGEN_HAS_CXX11_NOEXCEPT
+  #if    EIGEN_MAX_CPP_VER>=11 && \
+         (__has_feature(cxx_noexcept) \
+      || (__cplusplus > 201103L) \
+      || ((__cplusplus >= 201103L) && (EIGEN_COMP_GNUC_STRICT || EIGEN_COMP_CLANG || EIGEN_COMP_ICC>=1400)) \
+      || EIGEN_COMP_MSVC >= 1900)
+    #define EIGEN_HAS_CXX11_NOEXCEPT 1
+  #else
+    #define EIGEN_HAS_CXX11_NOEXCEPT 0
+  #endif
 #endif
 
 /** Allows to disable some optimizations which might affect the accuracy of the result.
   * Such optimization are enabled by default, and set EIGEN_FAST_MATH to 0 to disable them.
   * They currently include:
-  *   - single precision Cwise::sin() and Cwise::cos() when SSE vectorization is enabled.
+  *   - single precision ArrayBase::sin() and ArrayBase::cos() for SSE and AVX vectorization.
   */
 #ifndef EIGEN_FAST_MATH
 #define EIGEN_FAST_MATH 1
@@ -395,6 +495,8 @@
 #define EIGEN_CAT2(a,b) a ## b
 #define EIGEN_CAT(a,b) EIGEN_CAT2(a,b)
 
+#define EIGEN_COMMA ,
+
 // convert a token to a string
 #define EIGEN_MAKESTRING2(a) #a
 #define EIGEN_MAKESTRING(a) EIGEN_MAKESTRING2(a)
@@ -402,7 +504,7 @@
 // EIGEN_STRONG_INLINE is a stronger version of the inline, using __forceinline on MSVC,
 // but it still doesn't use GCC's always_inline. This is useful in (common) situations where MSVC needs forceinline
 // but GCC is still doing fine with just inline.
-#if (defined _MSC_VER) || (defined __INTEL_COMPILER)
+#if EIGEN_COMP_MSVC || EIGEN_COMP_ICC
 #define EIGEN_STRONG_INLINE __forceinline
 #else
 #define EIGEN_STRONG_INLINE inline
@@ -412,24 +514,25 @@
 // attribute to maximize inlining. This should only be used when really necessary: in particular,
 // it uses __attribute__((always_inline)) on GCC, which most of the time is useless and can severely harm compile times.
 // FIXME with the always_inline attribute,
-// gcc 3.4.x reports the following compilation error:
+// gcc 3.4.x and 4.1 reports the following compilation error:
 //   Eval.h:91: sorry, unimplemented: inlining failed in call to 'const Eigen::Eval<Derived> Eigen::MatrixBase<Scalar, Derived>::eval() const'
 //    : function body not available
-#if EIGEN_GNUC_AT_LEAST(4,0)
+//   See also bug 1367
+#if EIGEN_GNUC_AT_LEAST(4,2)
 #define EIGEN_ALWAYS_INLINE __attribute__((always_inline)) inline
 #else
 #define EIGEN_ALWAYS_INLINE EIGEN_STRONG_INLINE
 #endif
 
-#if (defined __GNUC__)
+#if EIGEN_COMP_GNUC
 #define EIGEN_DONT_INLINE __attribute__((noinline))
-#elif (defined _MSC_VER)
+#elif EIGEN_COMP_MSVC
 #define EIGEN_DONT_INLINE __declspec(noinline)
 #else
 #define EIGEN_DONT_INLINE
 #endif
 
-#if (defined __GNUC__)
+#if EIGEN_COMP_GNUC
 #define EIGEN_PERMISSIVE_EXPR __extension__
 #else
 #define EIGEN_PERMISSIVE_EXPR
@@ -439,8 +542,8 @@
 //  - static is not very good because it prevents definitions from different object files to be merged.
 //           So static causes the resulting linked executable to be bloated with multiple copies of the same function.
 //  - inline is not perfect either as it unwantedly hints the compiler toward inlining the function.
-#define EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
-#define EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS inline
+#define EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_DEVICE_FUNC
+#define EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_DEVICE_FUNC inline
 
 #ifdef NDEBUG
 # ifndef EIGEN_NO_DEBUG
@@ -498,15 +601,15 @@
 #endif
 
 #ifdef EIGEN_NO_DEBUG
-#define EIGEN_ONLY_USED_FOR_DEBUG(x) (void)x
+#define EIGEN_ONLY_USED_FOR_DEBUG(x) EIGEN_UNUSED_VARIABLE(x)
 #else
 #define EIGEN_ONLY_USED_FOR_DEBUG(x)
 #endif
 
 #ifndef EIGEN_NO_DEPRECATED_WARNING
-  #if (defined __GNUC__)
+  #if EIGEN_COMP_GNUC
     #define EIGEN_DEPRECATED __attribute__((deprecated))
-  #elif (defined _MSC_VER)
+  #elif EIGEN_COMP_MSVC
     #define EIGEN_DEPRECATED __declspec(deprecated)
   #else
     #define EIGEN_DEPRECATED
@@ -515,7 +618,7 @@
   #define EIGEN_DEPRECATED
 #endif
 
-#if (defined __GNUC__)
+#if EIGEN_COMP_GNUC
 #define EIGEN_UNUSED __attribute__((unused))
 #else
 #define EIGEN_UNUSED
@@ -524,19 +627,41 @@
 // Suppresses 'unused variable' warnings.
 namespace Eigen {
   namespace internal {
-    template<typename T> void ignore_unused_variable(const T&) {}
+    template<typename T> EIGEN_DEVICE_FUNC void ignore_unused_variable(const T&) {}
   }
 }
 #define EIGEN_UNUSED_VARIABLE(var) Eigen::internal::ignore_unused_variable(var);
 
 #if !defined(EIGEN_ASM_COMMENT)
-  #if (defined __GNUC__) && ( defined(__i386__) || defined(__x86_64__) )
+  #if EIGEN_COMP_GNUC && (EIGEN_ARCH_i386_OR_x86_64 || EIGEN_ARCH_ARM_OR_ARM64)
     #define EIGEN_ASM_COMMENT(X)  __asm__("#" X)
   #else
     #define EIGEN_ASM_COMMENT(X)
   #endif
 #endif
 
+
+#if EIGEN_COMP_MSVC
+  // NOTE MSVC often gives C4127 warnings with compiletime if statements. See bug 1362.
+  // This workaround is ugly, but it does the job.
+#  define EIGEN_CONST_CONDITIONAL(cond)  (void)0, cond
+#else
+#  define EIGEN_CONST_CONDITIONAL(cond)  cond
+#endif
+
+//------------------------------------------------------------------------------------------
+// Static and dynamic alignment control
+//
+// The main purpose of this section is to define EIGEN_MAX_ALIGN_BYTES and EIGEN_MAX_STATIC_ALIGN_BYTES
+// as the maximal boundary in bytes on which dynamically and statically allocated data may be alignment respectively.
+// The values of EIGEN_MAX_ALIGN_BYTES and EIGEN_MAX_STATIC_ALIGN_BYTES can be specified by the user. If not,
+// a default value is automatically computed based on architecture, compiler, and OS.
+//
+// This section also defines macros EIGEN_ALIGN_TO_BOUNDARY(N) and the shortcuts EIGEN_ALIGN{8,16,32,_MAX}
+// to be used to declare statically aligned buffers.
+//------------------------------------------------------------------------------------------
+
+
 /* EIGEN_ALIGN_TO_BOUNDARY(n) forces data to be n-byte aligned. This is used to satisfy SIMD requirements.
  * However, we do that EVEN if vectorization (EIGEN_VECTORIZE) is disabled,
  * so that vectorization doesn't affect binary compatibility.
@@ -544,28 +669,149 @@ namespace Eigen {
  * If we made alignment depend on whether or not EIGEN_VECTORIZE is defined, it would be impossible to link
  * vectorized and non-vectorized code.
  */
-#if (defined __GNUC__) || (defined __PGI) || (defined __IBMCPP__) || (defined __ARMCC_VERSION)
+#if (defined __CUDACC__)
+  #define EIGEN_ALIGN_TO_BOUNDARY(n) __align__(n)
+#elif EIGEN_COMP_GNUC || EIGEN_COMP_PGI || EIGEN_COMP_IBM || EIGEN_COMP_ARM
   #define EIGEN_ALIGN_TO_BOUNDARY(n) __attribute__((aligned(n)))
-#elif (defined _MSC_VER)
+#elif EIGEN_COMP_MSVC
   #define EIGEN_ALIGN_TO_BOUNDARY(n) __declspec(align(n))
-#elif (defined __SUNPRO_CC)
+#elif EIGEN_COMP_SUNCC
   // FIXME not sure about this one:
   #define EIGEN_ALIGN_TO_BOUNDARY(n) __attribute__((aligned(n)))
 #else
   #error Please tell me what is the equivalent of __attribute__((aligned(n))) for your compiler
 #endif
 
+// If the user explicitly disable vectorization, then we also disable alignment
+#if defined(EIGEN_DONT_VECTORIZE)
+  #define EIGEN_IDEAL_MAX_ALIGN_BYTES 0
+#elif defined(EIGEN_VECTORIZE_AVX512)
+  // 64 bytes static alignmeent is preferred only if really required
+  #define EIGEN_IDEAL_MAX_ALIGN_BYTES 64
+#elif defined(__AVX__)
+  // 32 bytes static alignmeent is preferred only if really required
+  #define EIGEN_IDEAL_MAX_ALIGN_BYTES 32
+#else
+  #define EIGEN_IDEAL_MAX_ALIGN_BYTES 16
+#endif
+
+
+// EIGEN_MIN_ALIGN_BYTES defines the minimal value for which the notion of explicit alignment makes sense
+#define EIGEN_MIN_ALIGN_BYTES 16
+
+// Defined the boundary (in bytes) on which the data needs to be aligned. Note
+// that unless EIGEN_ALIGN is defined and not equal to 0, the data may not be
+// aligned at all regardless of the value of this #define.
+
+#if (defined(EIGEN_DONT_ALIGN_STATICALLY) || defined(EIGEN_DONT_ALIGN))  && defined(EIGEN_MAX_STATIC_ALIGN_BYTES) && EIGEN_MAX_STATIC_ALIGN_BYTES>0
+#error EIGEN_MAX_STATIC_ALIGN_BYTES and EIGEN_DONT_ALIGN[_STATICALLY] are both defined with EIGEN_MAX_STATIC_ALIGN_BYTES!=0. Use EIGEN_MAX_STATIC_ALIGN_BYTES=0 as a synonym of EIGEN_DONT_ALIGN_STATICALLY.
+#endif
+
+// EIGEN_DONT_ALIGN_STATICALLY and EIGEN_DONT_ALIGN are deprectated
+// They imply EIGEN_MAX_STATIC_ALIGN_BYTES=0
+#if defined(EIGEN_DONT_ALIGN_STATICALLY) || defined(EIGEN_DONT_ALIGN)
+  #ifdef EIGEN_MAX_STATIC_ALIGN_BYTES
+    #undef EIGEN_MAX_STATIC_ALIGN_BYTES
+  #endif
+  #define EIGEN_MAX_STATIC_ALIGN_BYTES 0
+#endif
+
+#ifndef EIGEN_MAX_STATIC_ALIGN_BYTES
+
+  // Try to automatically guess what is the best default value for EIGEN_MAX_STATIC_ALIGN_BYTES
+
+  // 16 byte alignment is only useful for vectorization. Since it affects the ABI, we need to enable
+  // 16 byte alignment on all platforms where vectorization might be enabled. In theory we could always
+  // enable alignment, but it can be a cause of problems on some platforms, so we just disable it in
+  // certain common platform (compiler+architecture combinations) to avoid these problems.
+  // Only static alignment is really problematic (relies on nonstandard compiler extensions),
+  // try to keep heap alignment even when we have to disable static alignment.
+  #if EIGEN_COMP_GNUC && !(EIGEN_ARCH_i386_OR_x86_64 || EIGEN_ARCH_ARM_OR_ARM64 || EIGEN_ARCH_PPC || EIGEN_ARCH_IA64)
+  #define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 1
+  #elif EIGEN_ARCH_ARM_OR_ARM64 && EIGEN_COMP_GNUC_STRICT && EIGEN_GNUC_AT_MOST(4, 6)
+  // Old versions of GCC on ARM, at least 4.4, were once seen to have buggy static alignment support.
+  // Not sure which version fixed it, hopefully it doesn't affect 4.7, which is still somewhat in use.
+  // 4.8 and newer seem definitely unaffected.
+  #define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 1
+  #else
+  #define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 0
+  #endif
+
+  // static alignment is completely disabled with GCC 3, Sun Studio, and QCC/QNX
+  #if !EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT \
+  && !EIGEN_GCC3_OR_OLDER \
+  && !EIGEN_COMP_SUNCC \
+  && !EIGEN_OS_QNX
+    #define EIGEN_ARCH_WANTS_STACK_ALIGNMENT 1
+  #else
+    #define EIGEN_ARCH_WANTS_STACK_ALIGNMENT 0
+  #endif
+
+  #if EIGEN_ARCH_WANTS_STACK_ALIGNMENT
+    #define EIGEN_MAX_STATIC_ALIGN_BYTES EIGEN_IDEAL_MAX_ALIGN_BYTES
+  #else
+    #define EIGEN_MAX_STATIC_ALIGN_BYTES 0
+  #endif
+
+#endif
+
+// If EIGEN_MAX_ALIGN_BYTES is defined, then it is considered as an upper bound for EIGEN_MAX_ALIGN_BYTES
+#if defined(EIGEN_MAX_ALIGN_BYTES) && EIGEN_MAX_ALIGN_BYTES<EIGEN_MAX_STATIC_ALIGN_BYTES
+#undef EIGEN_MAX_STATIC_ALIGN_BYTES
+#define EIGEN_MAX_STATIC_ALIGN_BYTES EIGEN_MAX_ALIGN_BYTES
+#endif
+
+#if EIGEN_MAX_STATIC_ALIGN_BYTES==0 && !defined(EIGEN_DISABLE_UNALIGNED_ARRAY_ASSERT)
+  #define EIGEN_DISABLE_UNALIGNED_ARRAY_ASSERT
+#endif
+
+// At this stage, EIGEN_MAX_STATIC_ALIGN_BYTES>0 is the true test whether we want to align arrays on the stack or not.
+// It takes into account both the user choice to explicitly enable/disable alignment (by settting EIGEN_MAX_STATIC_ALIGN_BYTES)
+// and the architecture config (EIGEN_ARCH_WANTS_STACK_ALIGNMENT).
+// Henceforth, only EIGEN_MAX_STATIC_ALIGN_BYTES should be used.
+
+
+// Shortcuts to EIGEN_ALIGN_TO_BOUNDARY
 #define EIGEN_ALIGN8  EIGEN_ALIGN_TO_BOUNDARY(8)
 #define EIGEN_ALIGN16 EIGEN_ALIGN_TO_BOUNDARY(16)
+#define EIGEN_ALIGN32 EIGEN_ALIGN_TO_BOUNDARY(32)
+#define EIGEN_ALIGN64 EIGEN_ALIGN_TO_BOUNDARY(64)
+#if EIGEN_MAX_STATIC_ALIGN_BYTES>0
+#define EIGEN_ALIGN_MAX EIGEN_ALIGN_TO_BOUNDARY(EIGEN_MAX_STATIC_ALIGN_BYTES)
+#else
+#define EIGEN_ALIGN_MAX
+#endif
+
+
+// Dynamic alignment control
+
+#if defined(EIGEN_DONT_ALIGN) && defined(EIGEN_MAX_ALIGN_BYTES) && EIGEN_MAX_ALIGN_BYTES>0
+#error EIGEN_MAX_ALIGN_BYTES and EIGEN_DONT_ALIGN are both defined with EIGEN_MAX_ALIGN_BYTES!=0. Use EIGEN_MAX_ALIGN_BYTES=0 as a synonym of EIGEN_DONT_ALIGN.
+#endif
 
-#if EIGEN_ALIGN_STATICALLY
-#define EIGEN_USER_ALIGN_TO_BOUNDARY(n) EIGEN_ALIGN_TO_BOUNDARY(n)
-#define EIGEN_USER_ALIGN16 EIGEN_ALIGN16
+#ifdef EIGEN_DONT_ALIGN
+  #ifdef EIGEN_MAX_ALIGN_BYTES
+    #undef EIGEN_MAX_ALIGN_BYTES
+  #endif
+  #define EIGEN_MAX_ALIGN_BYTES 0
+#elif !defined(EIGEN_MAX_ALIGN_BYTES)
+  #define EIGEN_MAX_ALIGN_BYTES EIGEN_IDEAL_MAX_ALIGN_BYTES
+#endif
+
+#if EIGEN_IDEAL_MAX_ALIGN_BYTES > EIGEN_MAX_ALIGN_BYTES
+#define EIGEN_DEFAULT_ALIGN_BYTES EIGEN_IDEAL_MAX_ALIGN_BYTES
 #else
-#define EIGEN_USER_ALIGN_TO_BOUNDARY(n)
-#define EIGEN_USER_ALIGN16
+#define EIGEN_DEFAULT_ALIGN_BYTES EIGEN_MAX_ALIGN_BYTES
 #endif
 
+
+#ifndef EIGEN_UNALIGNED_VECTORIZE
+#define EIGEN_UNALIGNED_VECTORIZE 1
+#endif
+
+//----------------------------------------------------------------------
+
+
 #ifdef EIGEN_DONT_USE_RESTRICT_KEYWORD
   #define EIGEN_RESTRICT
 #endif
@@ -591,25 +837,26 @@ namespace Eigen {
 // just an empty macro !
 #define EIGEN_EMPTY
 
-#if defined(_MSC_VER) && (_MSC_VER < 1900) && (!defined(__INTEL_COMPILER))
-#define EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived) \
-  using Base::operator =;
-#elif defined(__clang__) // workaround clang bug (see http://forum.kde.org/viewtopic.php?f=74&t=102653)
-#define EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived) \
-  using Base::operator =; \
-  EIGEN_STRONG_INLINE Derived& operator=(const Derived& other) { Base::operator=(other); return *this; } \
-  template <typename OtherDerived> \
-  EIGEN_STRONG_INLINE Derived& operator=(const DenseBase<OtherDerived>& other) { Base::operator=(other.derived()); return *this; }
-#else
-#define EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived) \
-  using Base::operator =; \
-  EIGEN_STRONG_INLINE Derived& operator=(const Derived& other) \
-  { \
-    Base::operator=(other); \
-    return *this; \
-  }
+#if EIGEN_COMP_MSVC_STRICT && (EIGEN_COMP_MSVC < 1900 ||  defined(__CUDACC_VER__)) // for older MSVC versions, as well as 1900 && CUDA 8, using the base operator is sufficient (cf Bugs 1000, 1324)
+  #define EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived) \
+    using Base::operator =;
+#elif EIGEN_COMP_CLANG // workaround clang bug (see http://forum.kde.org/viewtopic.php?f=74&t=102653)
+  #define EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived) \
+    using Base::operator =; \
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& operator=(const Derived& other) { Base::operator=(other); return *this; } \
+    template <typename OtherDerived> \
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& operator=(const DenseBase<OtherDerived>& other) { Base::operator=(other.derived()); return *this; }
+#else
+  #define EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived) \
+    using Base::operator =; \
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& operator=(const Derived& other) \
+    { \
+      Base::operator=(other); \
+      return *this; \
+    }
 #endif
 
+
 /** \internal
  * \brief Macro to manually inherit assignment operators.
  * This is necessary, because the implicitly defined assignment operator gets deleted when a custom operator= is defined.
@@ -628,32 +875,13 @@ namespace Eigen {
   typedef typename Eigen::internal::traits<Derived>::Scalar Scalar; /*!< \brief Numeric type, e.g. float, double, int or std::complex<float>. */ \
   typedef typename Eigen::NumTraits<Scalar>::Real RealScalar; /*!< \brief The underlying numeric type for composed scalar types. \details In cases where Scalar is e.g. std::complex<T>, T were corresponding to RealScalar. */ \
   typedef typename Base::CoeffReturnType CoeffReturnType; /*!< \brief The return type for coefficient access. \details Depending on whether the object allows direct coefficient access (e.g. for a MatrixXd), this type is either 'const Scalar&' or simply 'Scalar' for objects that do not allow direct coefficient access. */ \
-  typedef typename Eigen::internal::nested<Derived>::type Nested; \
+  typedef typename Eigen::internal::ref_selector<Derived>::type Nested; \
   typedef typename Eigen::internal::traits<Derived>::StorageKind StorageKind; \
-  typedef typename Eigen::internal::traits<Derived>::Index Index; \
-  enum { RowsAtCompileTime = Eigen::internal::traits<Derived>::RowsAtCompileTime, \
+  typedef typename Eigen::internal::traits<Derived>::StorageIndex StorageIndex; \
+  enum CompileTimeTraits \
+      { RowsAtCompileTime = Eigen::internal::traits<Derived>::RowsAtCompileTime, \
         ColsAtCompileTime = Eigen::internal::traits<Derived>::ColsAtCompileTime, \
         Flags = Eigen::internal::traits<Derived>::Flags, \
-        CoeffReadCost = Eigen::internal::traits<Derived>::CoeffReadCost, \
-        SizeAtCompileTime = Base::SizeAtCompileTime, \
-        MaxSizeAtCompileTime = Base::MaxSizeAtCompileTime, \
-        IsVectorAtCompileTime = Base::IsVectorAtCompileTime };
-
-
-#define EIGEN_DENSE_PUBLIC_INTERFACE(Derived) \
-  typedef typename Eigen::internal::traits<Derived>::Scalar Scalar; /*!< \brief Numeric type, e.g. float, double, int or std::complex<float>. */ \
-  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar; /*!< \brief The underlying numeric type for composed scalar types. \details In cases where Scalar is e.g. std::complex<T>, T were corresponding to RealScalar. */ \
-  typedef typename Base::PacketScalar PacketScalar; \
-  typedef typename Base::CoeffReturnType CoeffReturnType; /*!< \brief The return type for coefficient access. \details Depending on whether the object allows direct coefficient access (e.g. for a MatrixXd), this type is either 'const Scalar&' or simply 'Scalar' for objects that do not allow direct coefficient access. */ \
-  typedef typename Eigen::internal::nested<Derived>::type Nested; \
-  typedef typename Eigen::internal::traits<Derived>::StorageKind StorageKind; \
-  typedef typename Eigen::internal::traits<Derived>::Index Index; \
-  enum { RowsAtCompileTime = Eigen::internal::traits<Derived>::RowsAtCompileTime, \
-        ColsAtCompileTime = Eigen::internal::traits<Derived>::ColsAtCompileTime, \
-        MaxRowsAtCompileTime = Eigen::internal::traits<Derived>::MaxRowsAtCompileTime, \
-        MaxColsAtCompileTime = Eigen::internal::traits<Derived>::MaxColsAtCompileTime, \
-        Flags = Eigen::internal::traits<Derived>::Flags, \
-        CoeffReadCost = Eigen::internal::traits<Derived>::CoeffReadCost, \
         SizeAtCompileTime = Base::SizeAtCompileTime, \
         MaxSizeAtCompileTime = Base::MaxSizeAtCompileTime, \
         IsVectorAtCompileTime = Base::IsVectorAtCompileTime }; \
@@ -661,6 +889,12 @@ namespace Eigen {
   using Base::const_cast_derived;
 
 
+// FIXME Maybe the EIGEN_DENSE_PUBLIC_INTERFACE could be removed as importing PacketScalar is rarely needed
+#define EIGEN_DENSE_PUBLIC_INTERFACE(Derived) \
+  EIGEN_GENERIC_PUBLIC_INTERFACE(Derived) \
+  typedef typename Base::PacketScalar PacketScalar;
+
+
 #define EIGEN_PLAIN_ENUM_MIN(a,b) (((int)a <= (int)b) ? (int)a : (int)b)
 #define EIGEN_PLAIN_ENUM_MAX(a,b) (((int)a >= (int)b) ? (int)a : (int)b)
 
@@ -686,24 +920,14 @@ namespace Eigen {
 #define EIGEN_SIZE_MAX(a,b) (((int)a == Dynamic || (int)b == Dynamic) ? Dynamic \
                            : ((int)a >= (int)b) ? (int)a : (int)b)
 
-#define EIGEN_ADD_COST(a,b) int(a)==Dynamic || int(b)==Dynamic ? Dynamic : int(a)+int(b)
-
 #define EIGEN_LOGICAL_XOR(a,b) (((a) || (b)) && !((a) && (b)))
 
 #define EIGEN_IMPLIES(a,b) (!(a) || (b))
 
-#define EIGEN_MAKE_CWISE_BINARY_OP(METHOD,FUNCTOR) \
-  template<typename OtherDerived> \
-  EIGEN_STRONG_INLINE const CwiseBinaryOp<FUNCTOR<Scalar>, const Derived, const OtherDerived> \
-  (METHOD)(const EIGEN_CURRENT_STORAGE_BASE_CLASS<OtherDerived> &other) const \
-  { \
-    return CwiseBinaryOp<FUNCTOR<Scalar>, const Derived, const OtherDerived>(derived(), other.derived()); \
-  }
-
-// the expression type of a cwise product
-#define EIGEN_CWISE_PRODUCT_RETURN_TYPE(LHS,RHS) \
+// the expression type of a standard coefficient wise binary operation
+#define EIGEN_CWISE_BINARY_RETURN_TYPE(LHS,RHS,OPNAME) \
     CwiseBinaryOp< \
-      internal::scalar_product_op< \
+      EIGEN_CAT(EIGEN_CAT(internal::scalar_,OPNAME),_op)< \
           typename internal::traits<LHS>::Scalar, \
           typename internal::traits<RHS>::Scalar \
       >, \
@@ -711,4 +935,84 @@ namespace Eigen {
       const RHS \
     >
 
+#define EIGEN_MAKE_CWISE_BINARY_OP(METHOD,OPNAME) \
+  template<typename OtherDerived> \
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const EIGEN_CWISE_BINARY_RETURN_TYPE(Derived,OtherDerived,OPNAME) \
+  (METHOD)(const EIGEN_CURRENT_STORAGE_BASE_CLASS<OtherDerived> &other) const \
+  { \
+    return EIGEN_CWISE_BINARY_RETURN_TYPE(Derived,OtherDerived,OPNAME)(derived(), other.derived()); \
+  }
+
+#define EIGEN_SCALAR_BINARY_SUPPORTED(OPNAME,TYPEA,TYPEB) \
+  (Eigen::internal::has_ReturnType<Eigen::ScalarBinaryOpTraits<TYPEA,TYPEB,EIGEN_CAT(EIGEN_CAT(Eigen::internal::scalar_,OPNAME),_op)<TYPEA,TYPEB>  > >::value)
+
+#define EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(EXPR,SCALAR,OPNAME) \
+  CwiseBinaryOp<EIGEN_CAT(EIGEN_CAT(internal::scalar_,OPNAME),_op)<typename internal::traits<EXPR>::Scalar,SCALAR>, const EXPR, \
+                const typename internal::plain_constant_type<EXPR,SCALAR>::type>
+
+#define EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(SCALAR,EXPR,OPNAME) \
+  CwiseBinaryOp<EIGEN_CAT(EIGEN_CAT(internal::scalar_,OPNAME),_op)<SCALAR,typename internal::traits<EXPR>::Scalar>, \
+                const typename internal::plain_constant_type<EXPR,SCALAR>::type, const EXPR>
+
+// Workaround for MSVC 2010 (see ML thread "patch with compile for for MSVC 2010")
+#if EIGEN_COMP_MSVC_STRICT<=1600
+#define EIGEN_MSVC10_WORKAROUND_BINARYOP_RETURN_TYPE(X) typename internal::enable_if<true,X>::type
+#else
+#define EIGEN_MSVC10_WORKAROUND_BINARYOP_RETURN_TYPE(X) X
+#endif
+
+#define EIGEN_MAKE_SCALAR_BINARY_OP_ONTHERIGHT(METHOD,OPNAME) \
+  template <typename T> EIGEN_DEVICE_FUNC inline \
+  EIGEN_MSVC10_WORKAROUND_BINARYOP_RETURN_TYPE(const EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(Derived,typename internal::promote_scalar_arg<Scalar EIGEN_COMMA T EIGEN_COMMA EIGEN_SCALAR_BINARY_SUPPORTED(OPNAME,Scalar,T)>::type,OPNAME))\
+  (METHOD)(const T& scalar) const { \
+    typedef typename internal::promote_scalar_arg<Scalar,T,EIGEN_SCALAR_BINARY_SUPPORTED(OPNAME,Scalar,T)>::type PromotedT; \
+    return EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(Derived,PromotedT,OPNAME)(derived(), \
+           typename internal::plain_constant_type<Derived,PromotedT>::type(derived().rows(), derived().cols(), internal::scalar_constant_op<PromotedT>(scalar))); \
+  }
+
+#define EIGEN_MAKE_SCALAR_BINARY_OP_ONTHELEFT(METHOD,OPNAME) \
+  template <typename T> EIGEN_DEVICE_FUNC inline friend \
+  EIGEN_MSVC10_WORKAROUND_BINARYOP_RETURN_TYPE(const EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(typename internal::promote_scalar_arg<Scalar EIGEN_COMMA T EIGEN_COMMA EIGEN_SCALAR_BINARY_SUPPORTED(OPNAME,T,Scalar)>::type,Derived,OPNAME)) \
+  (METHOD)(const T& scalar, const StorageBaseType& matrix) { \
+    typedef typename internal::promote_scalar_arg<Scalar,T,EIGEN_SCALAR_BINARY_SUPPORTED(OPNAME,T,Scalar)>::type PromotedT; \
+    return EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(PromotedT,Derived,OPNAME)( \
+           typename internal::plain_constant_type<Derived,PromotedT>::type(matrix.derived().rows(), matrix.derived().cols(), internal::scalar_constant_op<PromotedT>(scalar)), matrix.derived()); \
+  }
+
+#define EIGEN_MAKE_SCALAR_BINARY_OP(METHOD,OPNAME) \
+  EIGEN_MAKE_SCALAR_BINARY_OP_ONTHELEFT(METHOD,OPNAME) \
+  EIGEN_MAKE_SCALAR_BINARY_OP_ONTHERIGHT(METHOD,OPNAME)
+
+
+#ifdef EIGEN_EXCEPTIONS
+#  define EIGEN_THROW_X(X) throw X
+#  define EIGEN_THROW throw
+#  define EIGEN_TRY try
+#  define EIGEN_CATCH(X) catch (X)
+#else
+#  ifdef __CUDA_ARCH__
+#    define EIGEN_THROW_X(X) asm("trap;")
+#    define EIGEN_THROW asm("trap;")
+#  else
+#    define EIGEN_THROW_X(X) std::abort()
+#    define EIGEN_THROW std::abort()
+#  endif
+#  define EIGEN_TRY if (true)
+#  define EIGEN_CATCH(X) else
+#endif
+
+
+#if EIGEN_HAS_CXX11_NOEXCEPT
+#   define EIGEN_INCLUDE_TYPE_TRAITS
+#   define EIGEN_NOEXCEPT noexcept
+#   define EIGEN_NOEXCEPT_IF(x) noexcept(x)
+#   define EIGEN_NO_THROW noexcept(true)
+#   define EIGEN_EXCEPTION_SPEC(X) noexcept(false)
+#else
+#   define EIGEN_NOEXCEPT
+#   define EIGEN_NOEXCEPT_IF(x)
+#   define EIGEN_NO_THROW throw()
+#   define EIGEN_EXCEPTION_SPEC(X) throw(X)
+#endif
+
 #endif // EIGEN_MACROS_H
diff --git a/eigen/Eigen/src/Core/util/Memory.h b/eigen/Eigen/src/Core/util/Memory.h
index 74e37a4..7d90534 100644
--- a/eigen/Eigen/src/Core/util/Memory.h
+++ b/eigen/Eigen/src/Core/util/Memory.h
@@ -1,11 +1,12 @@
 // This file is part of Eigen, a lightweight C++ template library
 // for linear algebra.
 //
-// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2008-2015 Gael Guennebaud <gael.guennebaud@inria.fr>
 // Copyright (C) 2008-2009 Benoit Jacob <jacob.benoit.1@gmail.com>
 // Copyright (C) 2009 Kenneth Riddile <kfriddile@yahoo.com>
 // Copyright (C) 2010 Hauke Heibel <hauke.heibel@gmail.com>
 // Copyright (C) 2010 Thomas Capricelli <orzel@freehackers.org>
+// Copyright (C) 2013 Pavel Holoborodko <pavel@holoborodko.com>
 //
 // This Source Code Form is subject to the terms of the Mozilla
 // Public License v. 2.0. If a copy of the MPL was not distributed
@@ -31,7 +32,7 @@
 // page 114, "[The] LP64 model [...] is used by all 64-bit UNIX ports" so it's indeed
 // quite safe, at least within the context of glibc, to equate 64-bit with LP64.
 #if defined(__GLIBC__) && ((__GLIBC__>=2 && __GLIBC_MINOR__ >= 8) || __GLIBC__>2) \
- && defined(__LP64__) && ! defined( __SANITIZE_ADDRESS__ )
+ && defined(__LP64__) && ! defined( __SANITIZE_ADDRESS__ ) && (EIGEN_DEFAULT_ALIGN_BYTES == 16)
   #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 1
 #else
   #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 0
@@ -41,15 +42,15 @@
 //   See http://svn.freebsd.org/viewvc/base/stable/6/lib/libc/stdlib/malloc.c?view=markup
 // FreeBSD 7 seems to have 16-byte aligned malloc except on ARM and MIPS architectures
 //   See http://svn.freebsd.org/viewvc/base/stable/7/lib/libc/stdlib/malloc.c?view=markup
-#if defined(__FreeBSD__) && !defined(__arm__) && !defined(__mips__)
+#if defined(__FreeBSD__) && !(EIGEN_ARCH_ARM || EIGEN_ARCH_MIPS) && (EIGEN_DEFAULT_ALIGN_BYTES == 16)
   #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 1
 #else
   #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 0
 #endif
 
-#if defined(__APPLE__) \
- || defined(_WIN64) \
- || EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED \
+#if (EIGEN_OS_MAC && (EIGEN_DEFAULT_ALIGN_BYTES == 16))     \
+ || (EIGEN_OS_WIN64 && (EIGEN_DEFAULT_ALIGN_BYTES == 16))   \
+ || EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED              \
  || EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED
   #define EIGEN_MALLOC_ALREADY_ALIGNED 1
 #else
@@ -58,36 +59,17 @@
 
 #endif
 
-// See bug 554 (http://eigen.tuxfamily.org/bz/show_bug.cgi?id=554)
-// It seems to be unsafe to check _POSIX_ADVISORY_INFO without including unistd.h first.
-// Currently, let's include it only on unix systems:
-#if defined(__unix__) || defined(__unix)
-  #include <unistd.h>
-  #if ((defined __QNXNTO__) || (defined _GNU_SOURCE) || (defined __PGI) || ((defined _XOPEN_SOURCE) && (_XOPEN_SOURCE >= 600))) && (defined _POSIX_ADVISORY_INFO) && (_POSIX_ADVISORY_INFO > 0)
-    #define EIGEN_HAS_POSIX_MEMALIGN 1
-  #endif
-#endif
-
-#ifndef EIGEN_HAS_POSIX_MEMALIGN
-  #define EIGEN_HAS_POSIX_MEMALIGN 0
-#endif
-
-#ifdef EIGEN_VECTORIZE_SSE
-  #define EIGEN_HAS_MM_MALLOC 1
-#else
-  #define EIGEN_HAS_MM_MALLOC 0
-#endif
-
 namespace Eigen {
 
 namespace internal {
 
+EIGEN_DEVICE_FUNC
 inline void throw_std_bad_alloc()
 {
   #ifdef EIGEN_EXCEPTIONS
     throw std::bad_alloc();
   #else
-    std::size_t huge = -1;
+    std::size_t huge = static_cast<std::size_t>(-1);
     new int[huge];
   #endif
 }
@@ -103,9 +85,9 @@ inline void throw_std_bad_alloc()
   */
 inline void* handmade_aligned_malloc(std::size_t size)
 {
-  void *original = std::malloc(size+16);
+  void *original = std::malloc(size+EIGEN_DEFAULT_ALIGN_BYTES);
   if (original == 0) return 0;
-  void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(15))) + 16);
+  void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1))) + EIGEN_DEFAULT_ALIGN_BYTES);
   *(reinterpret_cast<void**>(aligned) - 1) = original;
   return aligned;
 }
@@ -118,7 +100,7 @@ inline void handmade_aligned_free(void *ptr)
 
 /** \internal
   * \brief Reallocates aligned memory.
-  * Since we know that our handmade version is based on std::realloc
+  * Since we know that our handmade version is based on std::malloc
   * we can use std::realloc to implement efficient reallocation.
   */
 inline void* handmade_aligned_realloc(void* ptr, std::size_t size, std::size_t = 0)
@@ -126,9 +108,9 @@ inline void* handmade_aligned_realloc(void* ptr, std::size_t size, std::size_t =
   if (ptr == 0) return handmade_aligned_malloc(size);
   void *original = *(reinterpret_cast<void**>(ptr) - 1);
   std::ptrdiff_t previous_offset = static_cast<char *>(ptr)-static_cast<char *>(original);
-  original = std::realloc(original,size+16);
+  original = std::realloc(original,size+EIGEN_DEFAULT_ALIGN_BYTES);
   if (original == 0) return 0;
-  void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(15))) + 16);
+  void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1))) + EIGEN_DEFAULT_ALIGN_BYTES);
   void *previous_aligned = static_cast<char *>(original)+previous_offset;
   if(aligned!=previous_aligned)
     std::memmove(aligned, previous_aligned, size);
@@ -138,92 +120,46 @@ inline void* handmade_aligned_realloc(void* ptr, std::size_t size, std::size_t =
 }
 
 /*****************************************************************************
-*** Implementation of generic aligned realloc (when no realloc can be used)***
-*****************************************************************************/
-
-void* aligned_malloc(std::size_t size);
-void  aligned_free(void *ptr);
-
-/** \internal
-  * \brief Reallocates aligned memory.
-  * Allows reallocation with aligned ptr types. This implementation will
-  * always create a new memory chunk and copy the old data.
-  */
-inline void* generic_aligned_realloc(void* ptr, size_t size, size_t old_size)
-{
-  if (ptr==0)
-    return aligned_malloc(size);
-
-  if (size==0)
-  {
-    aligned_free(ptr);
-    return 0;
-  }
-
-  void* newptr = aligned_malloc(size);
-  if (newptr == 0)
-  {
-    #ifdef EIGEN_HAS_ERRNO
-    errno = ENOMEM; // according to the standard
-    #endif
-    return 0;
-  }
-
-  if (ptr != 0)
-  {
-    std::memcpy(newptr, ptr, (std::min)(size,old_size));
-    aligned_free(ptr);
-  }
-
-  return newptr;
-}
-
-/*****************************************************************************
 *** Implementation of portable aligned versions of malloc/free/realloc     ***
 *****************************************************************************/
 
 #ifdef EIGEN_NO_MALLOC
-inline void check_that_malloc_is_allowed()
+EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed()
 {
   eigen_assert(false && "heap allocation is forbidden (EIGEN_NO_MALLOC is defined)");
 }
 #elif defined EIGEN_RUNTIME_NO_MALLOC
-inline bool is_malloc_allowed_impl(bool update, bool new_value = false)
+EIGEN_DEVICE_FUNC inline bool is_malloc_allowed_impl(bool update, bool new_value = false)
 {
   static bool value = true;
   if (update == 1)
     value = new_value;
   return value;
 }
-inline bool is_malloc_allowed() { return is_malloc_allowed_impl(false); }
-inline bool set_is_malloc_allowed(bool new_value) { return is_malloc_allowed_impl(true, new_value); }
-inline void check_that_malloc_is_allowed()
+EIGEN_DEVICE_FUNC inline bool is_malloc_allowed() { return is_malloc_allowed_impl(false); }
+EIGEN_DEVICE_FUNC inline bool set_is_malloc_allowed(bool new_value) { return is_malloc_allowed_impl(true, new_value); }
+EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed()
 {
   eigen_assert(is_malloc_allowed() && "heap allocation is forbidden (EIGEN_RUNTIME_NO_MALLOC is defined and g_is_malloc_allowed is false)");
 }
 #else
-inline void check_that_malloc_is_allowed()
+EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed()
 {}
 #endif
 
-/** \internal Allocates \a size bytes. The returned pointer is guaranteed to have 16 bytes alignment.
+/** \internal Allocates \a size bytes. The returned pointer is guaranteed to have 16 or 32 bytes alignment depending on the requirements.
   * On allocation error, the returned pointer is null, and std::bad_alloc is thrown.
   */
-inline void* aligned_malloc(size_t size)
+EIGEN_DEVICE_FUNC inline void* aligned_malloc(std::size_t size)
 {
   check_that_malloc_is_allowed();
 
   void *result;
-  #if !EIGEN_ALIGN
-    result = std::malloc(size);
-  #elif EIGEN_MALLOC_ALREADY_ALIGNED
+  #if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED
     result = std::malloc(size);
-  #elif EIGEN_HAS_POSIX_MEMALIGN
-    if(posix_memalign(&result, 16, size)) result = 0;
-  #elif EIGEN_HAS_MM_MALLOC
-    result = _mm_malloc(size, 16);
-  #elif defined(_MSC_VER) && (!defined(_WIN32_WCE))
-    result = _aligned_malloc(size, 16);
+    #if EIGEN_DEFAULT_ALIGN_BYTES==16
+    eigen_assert((size<16 || (std::size_t(result)%16)==0) && "System's malloc returned an unaligned pointer. Compile with EIGEN_MALLOC_ALREADY_ALIGNED=0 to fallback to handmade alignd memory allocator.");
+    #endif
   #else
     result = handmade_aligned_malloc(size);
   #endif
@@ -235,50 +171,27 @@ inline void* aligned_malloc(size_t size)
 }
 
 /** \internal Frees memory allocated with aligned_malloc. */
-inline void aligned_free(void *ptr)
+EIGEN_DEVICE_FUNC inline void aligned_free(void *ptr)
 {
-  #if !EIGEN_ALIGN
-    std::free(ptr);
-  #elif EIGEN_MALLOC_ALREADY_ALIGNED
+  #if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED
     std::free(ptr);
-  #elif EIGEN_HAS_POSIX_MEMALIGN
-    std::free(ptr);
-  #elif EIGEN_HAS_MM_MALLOC
-    _mm_free(ptr);
-  #elif defined(_MSC_VER) && (!defined(_WIN32_WCE))
-    _aligned_free(ptr);
   #else
     handmade_aligned_free(ptr);
   #endif
 }
 
 /**
-* \internal
-* \brief Reallocates an aligned block of memory.
-* \throws std::bad_alloc on allocation failure
-**/
-inline void* aligned_realloc(void *ptr, size_t new_size, size_t old_size)
+  * \internal
+  * \brief Reallocates an aligned block of memory.
+  * \throws std::bad_alloc on allocation failure
+  */
+inline void* aligned_realloc(void *ptr, std::size_t new_size, std::size_t old_size)
 {
   EIGEN_UNUSED_VARIABLE(old_size);
 
   void *result;
-#if !EIGEN_ALIGN
-  result = std::realloc(ptr,new_size);
-#elif EIGEN_MALLOC_ALREADY_ALIGNED
+#if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED
   result = std::realloc(ptr,new_size);
-#elif EIGEN_HAS_POSIX_MEMALIGN
-  result = generic_aligned_realloc(ptr,new_size,old_size);
-#elif EIGEN_HAS_MM_MALLOC
-  // The defined(_mm_free) is just here to verify that this MSVC version
-  // implements _mm_malloc/_mm_free based on the corresponding _aligned_
-  // functions. This may not always be the case and we just try to be safe.
-  #if defined(_MSC_VER) && (!defined(_WIN32_WCE)) && defined(_mm_free)
-    result = _aligned_realloc(ptr,new_size,16);
-  #else
-    result = generic_aligned_realloc(ptr,new_size,old_size);
-  #endif
-#elif defined(_MSC_VER) && (!defined(_WIN32_WCE))
-  result = _aligned_realloc(ptr,new_size,16);
 #else
   result = handmade_aligned_realloc(ptr,new_size,old_size);
 #endif
@@ -296,12 +209,12 @@ inline void* aligned_realloc(void *ptr, size_t new_size, size_t old_size)
 /** \internal Allocates \a size bytes. If Align is true, then the returned ptr is 16-byte-aligned.
   * On allocation error, the returned pointer is null, and a std::bad_alloc is thrown.
   */
-template<bool Align> inline void* conditional_aligned_malloc(size_t size)
+template<bool Align> EIGEN_DEVICE_FUNC inline void* conditional_aligned_malloc(std::size_t size)
 {
   return aligned_malloc(size);
 }
 
-template<> inline void* conditional_aligned_malloc<false>(size_t size)
+template<> EIGEN_DEVICE_FUNC inline void* conditional_aligned_malloc<false>(std::size_t size)
 {
   check_that_malloc_is_allowed();
 
@@ -312,22 +225,22 @@ template<> inline void* conditional_aligned_malloc<false>(size_t size)
 }
 
 /** \internal Frees memory allocated with conditional_aligned_malloc */
-template<bool Align> inline void conditional_aligned_free(void *ptr)
+template<bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_free(void *ptr)
 {
   aligned_free(ptr);
 }
 
-template<> inline void conditional_aligned_free<false>(void *ptr)
+template<> EIGEN_DEVICE_FUNC inline void conditional_aligned_free<false>(void *ptr)
 {
   std::free(ptr);
 }
 
-template<bool Align> inline void* conditional_aligned_realloc(void* ptr, size_t new_size, size_t old_size)
+template<bool Align> inline void* conditional_aligned_realloc(void* ptr, std::size_t new_size, std::size_t old_size)
 {
   return aligned_realloc(ptr, new_size, old_size);
 }
 
-template<> inline void* conditional_aligned_realloc<false>(void* ptr, size_t new_size, size_t)
+template<> inline void* conditional_aligned_realloc<false>(void* ptr, std::size_t new_size, std::size_t)
 {
   return std::realloc(ptr, new_size);
 }
@@ -336,33 +249,43 @@ template<> inline void* conditional_aligned_realloc<false>(void* ptr, size_t new
 *** Construction/destruction of array elements                             ***
 *****************************************************************************/
 
-/** \internal Constructs the elements of an array.
-  * The \a size parameter tells on how many objects to call the constructor of T.
-  */
-template<typename T> inline T* construct_elements_of_array(T *ptr, size_t size)
-{
-  for (size_t i=0; i < size; ++i) ::new (ptr + i) T;
-  return ptr;
-}
-
 /** \internal Destructs the elements of an array.
   * The \a size parameters tells on how many objects to call the destructor of T.
   */
-template<typename T> inline void destruct_elements_of_array(T *ptr, size_t size)
+template<typename T> EIGEN_DEVICE_FUNC inline void destruct_elements_of_array(T *ptr, std::size_t size)
 {
   // always destruct an array starting from the end.
   if(ptr)
     while(size) ptr[--size].~T();
 }
 
+/** \internal Constructs the elements of an array.
+  * The \a size parameter tells on how many objects to call the constructor of T.
+  */
+template<typename T> EIGEN_DEVICE_FUNC inline T* construct_elements_of_array(T *ptr, std::size_t size)
+{
+  std::size_t i;
+  EIGEN_TRY
+  {
+      for (i = 0; i < size; ++i) ::new (ptr + i) T;
+      return ptr;
+  }
+  EIGEN_CATCH(...)
+  {
+    destruct_elements_of_array(ptr, i);
+    EIGEN_THROW;
+  }
+  return NULL;
+}
+
 /*****************************************************************************
 *** Implementation of aligned new/delete-like functions                    ***
 *****************************************************************************/
 
 template<typename T>
-EIGEN_ALWAYS_INLINE void check_size_for_overflow(size_t size)
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void check_size_for_overflow(std::size_t size)
 {
-  if(size > size_t(-1) / sizeof(T))
+  if(size > std::size_t(-1) / sizeof(T))
     throw_std_bad_alloc();
 }
 
@@ -370,24 +293,42 @@ EIGEN_ALWAYS_INLINE void check_size_for_overflow(size_t size)
   * On allocation error, the returned pointer is undefined, but a std::bad_alloc is thrown.
   * The default constructor of T is called.
   */
-template<typename T> inline T* aligned_new(size_t size)
+template<typename T> EIGEN_DEVICE_FUNC inline T* aligned_new(std::size_t size)
 {
   check_size_for_overflow<T>(size);
   T *result = reinterpret_cast<T*>(aligned_malloc(sizeof(T)*size));
-  return construct_elements_of_array(result, size);
+  EIGEN_TRY
+  {
+    return construct_elements_of_array(result, size);
+  }
+  EIGEN_CATCH(...)
+  {
+    aligned_free(result);
+    EIGEN_THROW;
+  }
+  return result;
 }
 
-template<typename T, bool Align> inline T* conditional_aligned_new(size_t size)
+template<typename T, bool Align> EIGEN_DEVICE_FUNC inline T* conditional_aligned_new(std::size_t size)
 {
   check_size_for_overflow<T>(size);
   T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
-  return construct_elements_of_array(result, size);
+  EIGEN_TRY
+  {
+    return construct_elements_of_array(result, size);
+  }
+  EIGEN_CATCH(...)
+  {
+    conditional_aligned_free<Align>(result);
+    EIGEN_THROW;
+  }
+  return result;
 }
 
 /** \internal Deletes objects constructed with aligned_new
   * The \a size parameters tells on how many objects to call the destructor of T.
   */
-template<typename T> inline void aligned_delete(T *ptr, size_t size)
+template<typename T> EIGEN_DEVICE_FUNC inline void aligned_delete(T *ptr, std::size_t size)
 {
   destruct_elements_of_array<T>(ptr, size);
   aligned_free(ptr);
@@ -396,13 +337,13 @@ template<typename T> inline void aligned_delete(T *ptr, size_t size)
 /** \internal Deletes objects constructed with conditional_aligned_new
   * The \a size parameters tells on how many objects to call the destructor of T.
   */
-template<typename T, bool Align> inline void conditional_aligned_delete(T *ptr, size_t size)
+template<typename T, bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_delete(T *ptr, std::size_t size)
 {
   destruct_elements_of_array<T>(ptr, size);
   conditional_aligned_free<Align>(ptr);
 }
 
-template<typename T, bool Align> inline T* conditional_aligned_realloc_new(T* pts, size_t new_size, size_t old_size)
+template<typename T, bool Align> EIGEN_DEVICE_FUNC inline T* conditional_aligned_realloc_new(T* pts, std::size_t new_size, std::size_t old_size)
 {
   check_size_for_overflow<T>(new_size);
   check_size_for_overflow<T>(old_size);
@@ -410,23 +351,43 @@ template<typename T, bool Align> inline T* conditional_aligned_realloc_new(T* pt
     destruct_elements_of_array(pts+new_size, old_size-new_size);
   T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
   if(new_size > old_size)
-    construct_elements_of_array(result+old_size, new_size-old_size);
+  {
+    EIGEN_TRY
+    {
+      construct_elements_of_array(result+old_size, new_size-old_size);
+    }
+    EIGEN_CATCH(...)
+    {
+      conditional_aligned_free<Align>(result);
+      EIGEN_THROW;
+    }
+  }
   return result;
 }
 
 
-template<typename T, bool Align> inline T* conditional_aligned_new_auto(size_t size)
+template<typename T, bool Align> EIGEN_DEVICE_FUNC inline T* conditional_aligned_new_auto(std::size_t size)
 {
   if(size==0)
     return 0; // short-cut. Also fixes Bug 884
   check_size_for_overflow<T>(size);
   T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
   if(NumTraits<T>::RequireInitialization)
-    construct_elements_of_array(result, size);
+  {
+    EIGEN_TRY
+    {
+      construct_elements_of_array(result, size);
+    }
+    EIGEN_CATCH(...)
+    {
+      conditional_aligned_free<Align>(result);
+      EIGEN_THROW;
+    }
+  }
   return result;
 }
 
-template<typename T, bool Align> inline T* conditional_aligned_realloc_new_auto(T* pts, size_t new_size, size_t old_size)
+template<typename T, bool Align> inline T* conditional_aligned_realloc_new_auto(T* pts, std::size_t new_size, std::size_t old_size)
 {
   check_size_for_overflow<T>(new_size);
   check_size_for_overflow<T>(old_size);
@@ -434,11 +395,21 @@ template<typename T, bool Align> inline T* conditional_aligned_realloc_new_auto(
     destruct_elements_of_array(pts+new_size, old_size-new_size);
   T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
   if(NumTraits<T>::RequireInitialization && (new_size > old_size))
-    construct_elements_of_array(result+old_size, new_size-old_size);
+  {
+    EIGEN_TRY
+    {
+      construct_elements_of_array(result+old_size, new_size-old_size);
+    }
+    EIGEN_CATCH(...)
+    {
+      conditional_aligned_free<Align>(result);
+      EIGEN_THROW;
+    }
+  }
   return result;
 }
 
-template<typename T, bool Align> inline void conditional_aligned_delete_auto(T *ptr, size_t size)
+template<typename T, bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_delete_auto(T *ptr, std::size_t size)
 {
   if(NumTraits<T>::RequireInitialization)
     destruct_elements_of_array<T>(ptr, size);
@@ -447,51 +418,62 @@ template<typename T, bool Align> inline void conditional_aligned_delete_auto(T *
 
 /****************************************************************************/
 
-/** \internal Returns the index of the first element of the array that is well aligned for vectorization.
+/** \internal Returns the index of the first element of the array that is well aligned with respect to the requested \a Alignment.
   *
+  * \tparam Alignment requested alignment in Bytes.
   * \param array the address of the start of the array
   * \param size the size of the array
   *
-  * \note If no element of the array is well aligned, the size of the array is returned. Typically,
-  * for example with SSE, "well aligned" means 16-byte-aligned. If vectorization is disabled or if the
+  * \note If no element of the array is well aligned or the requested alignment is not a multiple of a scalar,
+  * the size of the array is returned. For example with SSE, the requested alignment is typically 16-bytes. If
   * packet size for the given scalar type is 1, then everything is considered well-aligned.
   *
-  * \note If the scalar type is vectorizable, we rely on the following assumptions: sizeof(Scalar) is a
-  * power of 2, the packet size in bytes is also a power of 2, and is a multiple of sizeof(Scalar). On the
-  * other hand, we do not assume that the array address is a multiple of sizeof(Scalar), as that fails for
+  * \note Otherwise, if the Alignment is larger that the scalar size, we rely on the assumptions that sizeof(Scalar) is a
+  * power of 2. On the other hand, we do not assume that the array address is a multiple of sizeof(Scalar), as that fails for
   * example with Scalar=double on certain 32-bit platforms, see bug #79.
   *
   * There is also the variant first_aligned(const MatrixBase&) defined in DenseCoeffsBase.h.
+  * \sa first_default_aligned()
   */
-template<typename Scalar, typename Index>
-static inline Index first_aligned(const Scalar* array, Index size)
+template<int Alignment, typename Scalar, typename Index>
+EIGEN_DEVICE_FUNC inline Index first_aligned(const Scalar* array, Index size)
 {
-  static const Index PacketSize = packet_traits<Scalar>::size;
-  static const Index PacketAlignedMask = PacketSize-1;
+  const Index ScalarSize = sizeof(Scalar);
+  const Index AlignmentSize = Alignment / ScalarSize;
+  const Index AlignmentMask = AlignmentSize-1;
 
-  if(PacketSize==1)
+  if(AlignmentSize<=1)
   {
-    // Either there is no vectorization, or a packet consists of exactly 1 scalar so that all elements
-    // of the array have the same alignment.
+    // Either the requested alignment if smaller than a scalar, or it exactly match a 1 scalar
+    // so that all elements of the array have the same alignment.
     return 0;
   }
-  else if(size_t(array) & (sizeof(Scalar)-1))
+  else if( (UIntPtr(array) & (sizeof(Scalar)-1)) || (Alignment%ScalarSize)!=0)
   {
-    // There is vectorization for this scalar type, but the array is not aligned to the size of a single scalar.
+    // The array is not aligned to the size of a single scalar, or the requested alignment is not a multiple of the scalar size.
     // Consequently, no element of the array is well aligned.
     return size;
   }
   else
   {
-    return std::min<Index>( (PacketSize - (Index((size_t(array)/sizeof(Scalar))) & PacketAlignedMask))
-                           & PacketAlignedMask, size);
+    Index first = (AlignmentSize - (Index((UIntPtr(array)/sizeof(Scalar))) & AlignmentMask)) & AlignmentMask;
+    return (first < size) ? first : size;
   }
 }
 
+/** \internal Returns the index of the first element of the array that is well aligned with respect the largest packet requirement.
+   * \sa first_aligned(Scalar*,Index) and first_default_aligned(DenseBase<Derived>) */
+template<typename Scalar, typename Index>
+EIGEN_DEVICE_FUNC inline Index first_default_aligned(const Scalar* array, Index size)
+{
+  typedef typename packet_traits<Scalar>::type DefaultPacketType;
+  return first_aligned<unpacket_traits<DefaultPacketType>::alignment>(array, size);
+}
+
 /** \internal Returns the smallest integer multiple of \a base and greater or equal to \a size
   */
 template<typename Index>
-inline static Index first_multiple(Index size, Index base)
+inline Index first_multiple(Index size, Index base)
 {
   return ((size+base-1)/base)*base;
 }
@@ -500,15 +482,15 @@ inline static Index first_multiple(Index size, Index base)
 // use memcpy on trivial types, i.e., on types that does not require an initialization ctor.
 template<typename T, bool UseMemcpy> struct smart_copy_helper;
 
-template<typename T> void smart_copy(const T* start, const T* end, T* target)
+template<typename T> EIGEN_DEVICE_FUNC void smart_copy(const T* start, const T* end, T* target)
 {
   smart_copy_helper<T,!NumTraits<T>::RequireInitialization>::run(start, end, target);
 }
 
 template<typename T> struct smart_copy_helper<T,true> {
-  static inline void run(const T* start, const T* end, T* target)
+  EIGEN_DEVICE_FUNC static inline void run(const T* start, const T* end, T* target)
   {
-    std::ptrdiff_t size = std::ptrdiff_t(end)-std::ptrdiff_t(start);
+    IntPtr size = IntPtr(end)-IntPtr(start);
     if(size==0) return;
     eigen_internal_assert(start!=0 && end!=0 && target!=0);
     memcpy(target, start, size);
@@ -516,10 +498,44 @@ template<typename T> struct smart_copy_helper<T,true> {
 };
 
 template<typename T> struct smart_copy_helper<T,false> {
-  static inline void run(const T* start, const T* end, T* target)
+  EIGEN_DEVICE_FUNC static inline void run(const T* start, const T* end, T* target)
   { std::copy(start, end, target); }
 };
 
+// intelligent memmove. falls back to std::memmove for POD types, uses std::copy otherwise.
+template<typename T, bool UseMemmove> struct smart_memmove_helper;
+
+template<typename T> void smart_memmove(const T* start, const T* end, T* target)
+{
+  smart_memmove_helper<T,!NumTraits<T>::RequireInitialization>::run(start, end, target);
+}
+
+template<typename T> struct smart_memmove_helper<T,true> {
+  static inline void run(const T* start, const T* end, T* target)
+  {
+    IntPtr size = IntPtr(end)-IntPtr(start);
+    if(size==0) return;
+    eigen_internal_assert(start!=0 && end!=0 && target!=0);
+    std::memmove(target, start, size);
+  }
+};
+
+template<typename T> struct smart_memmove_helper<T,false> {
+  static inline void run(const T* start, const T* end, T* target)
+  {
+    if (UIntPtr(target) < UIntPtr(start))
+    {
+      std::copy(start, end, target);
+    }
+    else
+    {
+      std::ptrdiff_t count = (std::ptrdiff_t(end)-std::ptrdiff_t(start)) / sizeof(T);
+      std::copy_backward(start, end, target + count);
+    }
+  }
+};
+
+
 /*****************************************************************************
 *** Implementation of runtime stack allocation (falling back to malloc)    ***
 *****************************************************************************/
@@ -527,16 +543,16 @@ template<typename T> struct smart_copy_helper<T,false> {
 // you can overwrite Eigen's default behavior regarding alloca by defining EIGEN_ALLOCA
 // to the appropriate stack allocation function
 #ifndef EIGEN_ALLOCA
-  #if (defined __linux__) || (defined __APPLE__) || (defined alloca)
+  #if EIGEN_OS_LINUX || EIGEN_OS_MAC || (defined alloca)
     #define EIGEN_ALLOCA alloca
-  #elif defined(_MSC_VER)
+  #elif EIGEN_COMP_MSVC
     #define EIGEN_ALLOCA _alloca
   #endif
 #endif
 
 // This helper class construct the allocated memory, and takes care of destructing and freeing the handled data
 // at destruction time. In practice this helper class is mainly useful to avoid memory leak in case of exceptions.
-template<typename T> class aligned_stack_memory_handler
+template<typename T> class aligned_stack_memory_handler : noncopyable
 {
   public:
     /* Creates a stack_memory_handler responsible for the buffer \a ptr of size \a size.
@@ -545,7 +561,7 @@ template<typename T> class aligned_stack_memory_handler
      * In this case, the buffer elements will also be destructed when this handler will be destructed.
      * Finally, if \a dealloc is true, then the pointer \a ptr is freed.
      **/
-    aligned_stack_memory_handler(T* ptr, size_t size, bool dealloc)
+    aligned_stack_memory_handler(T* ptr, std::size_t size, bool dealloc)
       : m_ptr(ptr), m_size(size), m_deallocate(dealloc)
     {
       if(NumTraits<T>::RequireInitialization && m_ptr)
@@ -560,10 +576,34 @@ template<typename T> class aligned_stack_memory_handler
     }
   protected:
     T* m_ptr;
-    size_t m_size;
+    std::size_t m_size;
     bool m_deallocate;
 };
 
+template<typename T> class scoped_array : noncopyable
+{
+  T* m_ptr;
+public:
+  explicit scoped_array(std::ptrdiff_t size)
+  {
+    m_ptr = new T[size];
+  }
+  ~scoped_array()
+  {
+    delete[] m_ptr;
+  }
+  T& operator[](std::ptrdiff_t i) { return m_ptr[i]; }
+  const T& operator[](std::ptrdiff_t i) const { return m_ptr[i]; }
+  T* &ptr() { return m_ptr; }
+  const T* ptr() const { return m_ptr; }
+  operator const T*() const { return m_ptr; }
+};
+
+template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b)
+{
+  std::swap(a.ptr(),b.ptr());
+}
+
 } // end namespace internal
 
 /** \internal
@@ -583,10 +623,12 @@ template<typename T> class aligned_stack_memory_handler
   */
 #ifdef EIGEN_ALLOCA
 
-  #if defined(__arm__) || defined(_WIN32)
-    #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast<void*>((reinterpret_cast<size_t>(EIGEN_ALLOCA(SIZE+16)) & ~(size_t(15))) + 16)
+  #if EIGEN_DEFAULT_ALIGN_BYTES>0
+    // We always manually re-align the result of EIGEN_ALLOCA.
+    // If alloca is already aligned, the compiler should be smart enough to optimize away the re-alignment.
+    #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast<void*>((internal::UIntPtr(EIGEN_ALLOCA(SIZE+EIGEN_DEFAULT_ALIGN_BYTES-1)) + EIGEN_DEFAULT_ALIGN_BYTES-1) & ~(std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1)))
   #else
-    #define EIGEN_ALIGNED_ALLOCA EIGEN_ALLOCA
+    #define EIGEN_ALIGNED_ALLOCA(SIZE) EIGEN_ALLOCA(SIZE)
   #endif
 
   #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
@@ -611,41 +653,33 @@ template<typename T> class aligned_stack_memory_handler
 *** Implementation of EIGEN_MAKE_ALIGNED_OPERATOR_NEW [_IF]                ***
 *****************************************************************************/
 
-#if EIGEN_ALIGN
-  #ifdef EIGEN_EXCEPTIONS
-    #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
-      void* operator new(size_t size, const std::nothrow_t&) noexcept { \
-        try { return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); } \
-        catch (...) { return 0; } \
+#if EIGEN_MAX_ALIGN_BYTES!=0
+  #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
+      void* operator new(std::size_t size, const std::nothrow_t&) EIGEN_NO_THROW { \
+        EIGEN_TRY { return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); } \
+        EIGEN_CATCH (...) { return 0; } \
       }
-  #else
-    #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
-      void* operator new(size_t size, const std::nothrow_t&) noexcept { \
-        return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
-      }
-  #endif
-
   #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign) \
-      void *operator new(size_t size) { \
+      void *operator new(std::size_t size) { \
         return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
       } \
-      void *operator new[](size_t size) { \
+      void *operator new[](std::size_t size) { \
         return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
       } \
-      void operator delete(void * ptr) noexcept { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
-      void operator delete[](void * ptr) noexcept { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
-      void operator delete(void * ptr, std::size_t /* sz */) noexcept { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
-      void operator delete[](void * ptr, std::size_t /* sz */) noexcept { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
+      void operator delete(void * ptr) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
+      void operator delete[](void * ptr) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
+      void operator delete(void * ptr, std::size_t /* sz */) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
+      void operator delete[](void * ptr, std::size_t /* sz */) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
       /* in-place new and delete. since (at least afaik) there is no actual   */ \
       /* memory allocated we can safely let the default implementation handle */ \
       /* this particular case. */ \
-      static void *operator new(size_t size, void *ptr) { return ::operator new(size,ptr); } \
-      static void *operator new[](size_t size, void* ptr) { return ::operator new[](size,ptr); } \
-      void operator delete(void * memory, void *ptr) noexcept { return ::operator delete(memory,ptr); } \
-      void operator delete[](void * memory, void *ptr) noexcept { return ::operator delete[](memory,ptr); } \
+      static void *operator new(std::size_t size, void *ptr) { return ::operator new(size,ptr); } \
+      static void *operator new[](std::size_t size, void* ptr) { return ::operator new[](size,ptr); } \
+      void operator delete(void * memory, void *ptr) EIGEN_NO_THROW { return ::operator delete(memory,ptr); } \
+      void operator delete[](void * memory, void *ptr) EIGEN_NO_THROW { return ::operator delete[](memory,ptr); } \
       /* nothrow-new (returns zero instead of std::bad_alloc) */ \
       EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
-      void operator delete(void *ptr, const std::nothrow_t&) noexcept { \
+      void operator delete(void *ptr, const std::nothrow_t&) EIGEN_NO_THROW { \
         Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); \
       } \
       typedef void eigen_aligned_operator_new_marker_type;
@@ -655,32 +689,31 @@ template<typename T> class aligned_stack_memory_handler
 
 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(true)
 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size) \
-  EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool(((Size)!=Eigen::Dynamic) && ((sizeof(Scalar)*(Size))%16==0)))
+  EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool(((Size)!=Eigen::Dynamic) && ((sizeof(Scalar)*(Size))%EIGEN_MAX_ALIGN_BYTES==0)))
 
 /****************************************************************************/
 
-
 /** \class aligned_allocator
-  * \ingroup Core_Module
-  *
-  * \brief STL compatible allocator to use with with 16 byte aligned types
-  *
-  * Example:
-  * \code
-  * // Matrix4f requires 16 bytes alignment:
-  * std::map< int, Matrix4f, std::less<int>,
-  *           aligned_allocator<std::pair<const int, Matrix4f> > > my_map_mat4;
-  * // Vector3f does not require 16 bytes alignment, no need to use Eigen's allocator:
-  * std::map< int, Vector3f > my_map_vec3;
-  * \endcode
-  *
-  * \sa \blank \ref TopicStlContainers.
-  */
+* \ingroup Core_Module
+*
+* \brief STL compatible allocator to use with with 16 byte aligned types
+*
+* Example:
+* \code
+* // Matrix4f requires 16 bytes alignment:
+* std::map< int, Matrix4f, std::less<int>,
+*           aligned_allocator<std::pair<const int, Matrix4f> > > my_map_mat4;
+* // Vector3f does not require 16 bytes alignment, no need to use Eigen's allocator:
+* std::map< int, Vector3f > my_map_vec3;
+* \endcode
+*
+* \sa \blank \ref TopicStlContainers.
+*/
 template<class T>
 class aligned_allocator : public std::allocator<T>
 {
 public:
-  typedef size_t          size_type;
+  typedef std::size_t     size_type;
   typedef std::ptrdiff_t  difference_type;
   typedef T*              pointer;
   typedef const T*        const_pointer;
@@ -718,12 +751,12 @@ public:
 //---------- Cache sizes ----------
 
 #if !defined(EIGEN_NO_CPUID)
-#  if defined(__GNUC__) && ( defined(__i386__) || defined(__x86_64__) )
-#    if defined(__PIC__) && defined(__i386__)
+#  if EIGEN_COMP_GNUC && EIGEN_ARCH_i386_OR_x86_64
+#    if defined(__PIC__) && EIGEN_ARCH_i386
        // Case for x86 with PIC
 #      define EIGEN_CPUID(abcd,func,id) \
          __asm__ __volatile__ ("xchgl %%ebx, %k1;cpuid; xchgl %%ebx,%k1": "=a" (abcd[0]), "=&r" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "a" (func), "c" (id));
-#    elif defined(__PIC__) && defined(__x86_64__)
+#    elif defined(__PIC__) && EIGEN_ARCH_x86_64
        // Case for x64 with PIC. In theory this is only a problem with recent gcc and with medium or large code model, not with the default small code model.
        // However, we cannot detect which code model is used, and the xchg overhead is negligible anyway.
 #      define EIGEN_CPUID(abcd,func,id) \
@@ -733,8 +766,8 @@ public:
 #      define EIGEN_CPUID(abcd,func,id) \
          __asm__ __volatile__ ("cpuid": "=a" (abcd[0]), "=b" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "0" (func), "2" (id) );
 #    endif
-#  elif defined(_MSC_VER)
-#    if (_MSC_VER > 1500) && ( defined(_M_IX86) || defined(_M_X64) )
+#  elif EIGEN_COMP_MSVC
+#    if (EIGEN_COMP_MSVC > 1500) && EIGEN_ARCH_i386_OR_x86_64
 #      define EIGEN_CPUID(abcd,func,id) __cpuidex((int*)abcd,func,id)
 #    endif
 #  endif
diff --git a/eigen/Eigen/src/Core/util/Meta.h b/eigen/Eigen/src/Core/util/Meta.h
index 71d5871..8de6055 100644
--- a/eigen/Eigen/src/Core/util/Meta.h
+++ b/eigen/Eigen/src/Core/util/Meta.h
@@ -1,7 +1,7 @@
 // This file is part of Eigen, a lightweight C++ template library
 // for linear algebra.
 //
-// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2008-2015 Gael Guennebaud <gael.guennebaud@inria.fr>
 // Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
 //
 // This Source Code Form is subject to the terms of the Mozilla
@@ -11,8 +11,27 @@
 #ifndef EIGEN_META_H
 #define EIGEN_META_H
 
+#if defined(__CUDA_ARCH__)
+#include <cfloat>
+#include <math_constants.h>
+#endif
+
+#if EIGEN_COMP_ICC>=1600 &&  __cplusplus >= 201103L
+#include <cstdint>
+#endif
+
 namespace Eigen {
 
+typedef EIGEN_DEFAULT_DENSE_INDEX_TYPE DenseIndex;
+
+/**
+ * \brief The Index type as used for the API.
+ * \details To change this, \c \#define the preprocessor symbol \c EIGEN_DEFAULT_DENSE_INDEX_TYPE.
+ * \sa \blank \ref TopicPreprocessorDirectives, StorageIndex.
+ */
+
+typedef EIGEN_DEFAULT_DENSE_INDEX_TYPE Index;
+
 namespace internal {
 
 /** \internal
@@ -22,6 +41,16 @@ namespace internal {
   * we however don't want to add a dependency to Boost.
   */
 
+// Only recent versions of ICC complain about using ptrdiff_t to hold pointers,
+// and older versions do not provide *intptr_t types.
+#if EIGEN_COMP_ICC>=1600 &&  __cplusplus >= 201103L
+typedef std::intptr_t  IntPtr;
+typedef std::uintptr_t UIntPtr;
+#else
+typedef std::ptrdiff_t IntPtr;
+typedef std::size_t UIntPtr;
+#endif
+
 struct true_type {  enum { value = 1 }; };
 struct false_type { enum { value = 0 }; };
 
@@ -68,6 +97,23 @@ template<> struct is_arithmetic<unsigned int>  { enum { value = true }; };
 template<> struct is_arithmetic<signed long>   { enum { value = true }; };
 template<> struct is_arithmetic<unsigned long> { enum { value = true }; };
 
+#if EIGEN_HAS_CXX11
+using std::is_integral;
+#else
+template<typename T> struct is_integral               { enum { value = false }; };
+template<> struct is_integral<bool>                   { enum { value = true }; };
+template<> struct is_integral<char>                   { enum { value = true }; };
+template<> struct is_integral<signed char>            { enum { value = true }; };
+template<> struct is_integral<unsigned char>          { enum { value = true }; };
+template<> struct is_integral<signed short>           { enum { value = true }; };
+template<> struct is_integral<unsigned short>         { enum { value = true }; };
+template<> struct is_integral<signed int>             { enum { value = true }; };
+template<> struct is_integral<unsigned int>           { enum { value = true }; };
+template<> struct is_integral<signed long>            { enum { value = true }; };
+template<> struct is_integral<unsigned long>          { enum { value = true }; };
+#endif
+
+
 template <typename T> struct add_const { typedef const T type; };
 template <typename T> struct add_const<T&> { typedef T& type; };
 
@@ -80,28 +126,215 @@ template<typename T> struct add_const_on_value_type<T*>        { typedef T const
 template<typename T> struct add_const_on_value_type<T* const>  { typedef T const* const type; };
 template<typename T> struct add_const_on_value_type<T const* const>  { typedef T const* const type; };
 
+
+template<typename From, typename To>
+struct is_convertible_impl
+{
+private:
+  struct any_conversion
+  {
+    template <typename T> any_conversion(const volatile T&);
+    template <typename T> any_conversion(T&);
+  };
+  struct yes {int a[1];};
+  struct no  {int a[2];};
+
+  static yes test(const To&, int);
+  static no  test(any_conversion, ...);
+
+public:
+  static From ms_from;
+#ifdef __INTEL_COMPILER
+  #pragma warning push
+  #pragma warning ( disable : 2259 )
+#endif
+  enum { value = sizeof(test(ms_from, 0))==sizeof(yes) };
+#ifdef __INTEL_COMPILER
+  #pragma warning pop
+#endif
+};
+
+template<typename From, typename To>
+struct is_convertible
+{
+  enum { value = is_convertible_impl<typename remove_all<From>::type,
+                                     typename remove_all<To  >::type>::value };
+};
+
 /** \internal Allows to enable/disable an overload
   * according to a compile time condition.
   */
-template<bool Condition, typename T> struct enable_if;
+template<bool Condition, typename T=void> struct enable_if;
 
 template<typename T> struct enable_if<true,T>
 { typedef T type; };
 
+#if defined(__CUDA_ARCH__)
+#if !defined(__FLT_EPSILON__)
+#define __FLT_EPSILON__ FLT_EPSILON
+#define __DBL_EPSILON__ DBL_EPSILON
+#endif
 
+namespace device {
+
+template<typename T> struct numeric_limits
+{
+  EIGEN_DEVICE_FUNC
+  static T epsilon() { return 0; }
+  static T (max)() { assert(false && "Highest not supported for this type"); }
+  static T (min)() { assert(false && "Lowest not supported for this type"); }
+  static T infinity() { assert(false && "Infinity not supported for this type"); }
+  static T quiet_NaN() { assert(false && "quiet_NaN not supported for this type"); }
+};
+template<> struct numeric_limits<float>
+{
+  EIGEN_DEVICE_FUNC
+  static float epsilon() { return __FLT_EPSILON__; }
+  EIGEN_DEVICE_FUNC
+  static float (max)() { return CUDART_MAX_NORMAL_F; }
+  EIGEN_DEVICE_FUNC
+  static float (min)() { return FLT_MIN; }
+  EIGEN_DEVICE_FUNC
+  static float infinity() { return CUDART_INF_F; }
+  EIGEN_DEVICE_FUNC
+  static float quiet_NaN() { return CUDART_NAN_F; }
+};
+template<> struct numeric_limits<double>
+{
+  EIGEN_DEVICE_FUNC
+  static double epsilon() { return __DBL_EPSILON__; }
+  EIGEN_DEVICE_FUNC
+  static double (max)() { return DBL_MAX; }
+  EIGEN_DEVICE_FUNC
+  static double (min)() { return DBL_MIN; }
+  EIGEN_DEVICE_FUNC
+  static double infinity() { return CUDART_INF; }
+  EIGEN_DEVICE_FUNC
+  static double quiet_NaN() { return CUDART_NAN; }
+};
+template<> struct numeric_limits<int>
+{
+  EIGEN_DEVICE_FUNC
+  static int epsilon() { return 0; }
+  EIGEN_DEVICE_FUNC
+  static int (max)() { return INT_MAX; }
+  EIGEN_DEVICE_FUNC
+  static int (min)() { return INT_MIN; }
+};
+template<> struct numeric_limits<unsigned int>
+{
+  EIGEN_DEVICE_FUNC
+  static unsigned int epsilon() { return 0; }
+  EIGEN_DEVICE_FUNC
+  static unsigned int (max)() { return UINT_MAX; }
+  EIGEN_DEVICE_FUNC
+  static unsigned int (min)() { return 0; }
+};
+template<> struct numeric_limits<long>
+{
+  EIGEN_DEVICE_FUNC
+  static long epsilon() { return 0; }
+  EIGEN_DEVICE_FUNC
+  static long (max)() { return LONG_MAX; }
+  EIGEN_DEVICE_FUNC
+  static long (min)() { return LONG_MIN; }
+};
+template<> struct numeric_limits<unsigned long>
+{
+  EIGEN_DEVICE_FUNC
+  static unsigned long epsilon() { return 0; }
+  EIGEN_DEVICE_FUNC
+  static unsigned long (max)() { return ULONG_MAX; }
+  EIGEN_DEVICE_FUNC
+  static unsigned long (min)() { return 0; }
+};
+template<> struct numeric_limits<long long>
+{
+  EIGEN_DEVICE_FUNC
+  static long long epsilon() { return 0; }
+  EIGEN_DEVICE_FUNC
+  static long long (max)() { return LLONG_MAX; }
+  EIGEN_DEVICE_FUNC
+  static long long (min)() { return LLONG_MIN; }
+};
+template<> struct numeric_limits<unsigned long long>
+{
+  EIGEN_DEVICE_FUNC
+  static unsigned long long epsilon() { return 0; }
+  EIGEN_DEVICE_FUNC
+  static unsigned long long (max)() { return ULLONG_MAX; }
+  EIGEN_DEVICE_FUNC
+  static unsigned long long (min)() { return 0; }
+};
+
+}
+
+#endif
 
 /** \internal
   * A base class do disable default copy ctor and copy assignement operator.
   */
 class noncopyable
 {
-  noncopyable(const noncopyable&);
-  const noncopyable& operator=(const noncopyable&);
+  EIGEN_DEVICE_FUNC noncopyable(const noncopyable&);
+  EIGEN_DEVICE_FUNC const noncopyable& operator=(const noncopyable&);
 protected:
-  noncopyable() {}
-  ~noncopyable() {}
+  EIGEN_DEVICE_FUNC noncopyable() {}
+  EIGEN_DEVICE_FUNC ~noncopyable() {}
+};
+
+/** \internal
+  * Provides access to the number of elements in the object of as a compile-time constant expression.
+  * It "returns" Eigen::Dynamic if the size cannot be resolved at compile-time (default).
+  *
+  * Similar to std::tuple_size, but more general.
+  *
+  * It currently supports:
+  *  - any types T defining T::SizeAtCompileTime
+  *  - plain C arrays as T[N]
+  *  - std::array (c++11)
+  *  - some internal types such as SingleRange and AllRange
+  *
+  * The second template parameter eases SFINAE-based specializations.
+  */
+template<typename T, typename EnableIf = void> struct array_size {
+  enum { value = Dynamic };
 };
 
+template<typename T> struct array_size<T,typename internal::enable_if<((T::SizeAtCompileTime&0)==0)>::type> {
+  enum { value = T::SizeAtCompileTime };
+};
+
+template<typename T, int N> struct array_size<const T (&)[N]> {
+  enum { value = N };
+};
+template<typename T, int N> struct array_size<T (&)[N]> {
+  enum { value = N };
+};
+
+#if EIGEN_HAS_CXX11
+template<typename T, std::size_t N> struct array_size<const std::array<T,N> > {
+  enum { value = N };
+};
+template<typename T, std::size_t N> struct array_size<std::array<T,N> > {
+  enum { value = N };
+};
+#endif
+
+/** \internal
+  * Analogue of the std::size free function.
+  * It returns the size of the container or view \a x of type \c T
+  *
+  * It currently supports:
+  *  - any types T defining a member T::size() const
+  *  - plain C arrays as T[N]
+  *
+  */
+template<typename T>
+Index size(const T& x) { return x.size(); }
+
+template<typename T,std::size_t N>
+Index size(const T (&) [N]) { return N; }
 
 /** \internal
   * Convenient struct to get the result type of a unary or binary functor.
@@ -110,14 +343,20 @@ protected:
   * upcoming next STL generation (using a templated result member).
   * If none of these members is provided, then the type of the first argument is returned. FIXME, that behavior is a pretty bad hack.
   */
-template<typename T> struct result_of {};
+#if EIGEN_HAS_STD_RESULT_OF
+template<typename T> struct result_of {
+  typedef typename std::result_of<T>::type type1;
+  typedef typename remove_all<type1>::type type;
+};
+#else
+template<typename T> struct result_of { };
 
 struct has_none {int a[1];};
 struct has_std_result_type {int a[2];};
 struct has_tr1_result {int a[3];};
 
 template<typename Func, typename ArgType, int SizeOf=sizeof(has_none)>
-struct unary_result_of_select {typedef ArgType type;};
+struct unary_result_of_select {typedef typename internal::remove_all<ArgType>::type type;};
 
 template<typename Func, typename ArgType>
 struct unary_result_of_select<Func, ArgType, sizeof(has_std_result_type)> {typedef typename Func::result_type type;};
@@ -128,10 +367,10 @@ struct unary_result_of_select<Func, ArgType, sizeof(has_tr1_result)> {typedef ty
 template<typename Func, typename ArgType>
 struct result_of<Func(ArgType)> {
     template<typename T>
-    static has_std_result_type testFunctor(T const *, typename T::result_type const * = 0);
+    static has_std_result_type    testFunctor(T const *, typename T::result_type const * = 0);
     template<typename T>
-    static has_tr1_result      testFunctor(T const *, typename T::template result<T(ArgType)>::type const * = 0);
-    static has_none            testFunctor(...);
+    static has_tr1_result         testFunctor(T const *, typename T::template result<T(ArgType)>::type const * = 0);
+    static has_none               testFunctor(...);
 
     // note that the following indirection is needed for gcc-3.3
     enum {FunctorType = sizeof(testFunctor(static_cast<Func*>(0)))};
@@ -139,7 +378,7 @@ struct result_of<Func(ArgType)> {
 };
 
 template<typename Func, typename ArgType0, typename ArgType1, int SizeOf=sizeof(has_none)>
-struct binary_result_of_select {typedef ArgType0 type;};
+struct binary_result_of_select {typedef typename internal::remove_all<ArgType0>::type type;};
 
 template<typename Func, typename ArgType0, typename ArgType1>
 struct binary_result_of_select<Func, ArgType0, ArgType1, sizeof(has_std_result_type)>
@@ -152,16 +391,83 @@ struct binary_result_of_select<Func, ArgType0, ArgType1, sizeof(has_tr1_result)>
 template<typename Func, typename ArgType0, typename ArgType1>
 struct result_of<Func(ArgType0,ArgType1)> {
     template<typename T>
-    static has_std_result_type testFunctor(T const *, typename T::result_type const * = 0);
+    static has_std_result_type    testFunctor(T const *, typename T::result_type const * = 0);
     template<typename T>
-    static has_tr1_result      testFunctor(T const *, typename T::template result<T(ArgType0,ArgType1)>::type const * = 0);
-    static has_none            testFunctor(...);
+    static has_tr1_result         testFunctor(T const *, typename T::template result<T(ArgType0,ArgType1)>::type const * = 0);
+    static has_none               testFunctor(...);
 
     // note that the following indirection is needed for gcc-3.3
     enum {FunctorType = sizeof(testFunctor(static_cast<Func*>(0)))};
     typedef typename binary_result_of_select<Func, ArgType0, ArgType1, FunctorType>::type type;
 };
 
+template<typename Func, typename ArgType0, typename ArgType1, typename ArgType2, int SizeOf=sizeof(has_none)>
+struct ternary_result_of_select {typedef typename internal::remove_all<ArgType0>::type type;};
+
+template<typename Func, typename ArgType0, typename ArgType1, typename ArgType2>
+struct ternary_result_of_select<Func, ArgType0, ArgType1, ArgType2, sizeof(has_std_result_type)>
+{typedef typename Func::result_type type;};
+
+template<typename Func, typename ArgType0, typename ArgType1, typename ArgType2>
+struct ternary_result_of_select<Func, ArgType0, ArgType1, ArgType2, sizeof(has_tr1_result)>
+{typedef typename Func::template result<Func(ArgType0,ArgType1,ArgType2)>::type type;};
+
+template<typename Func, typename ArgType0, typename ArgType1, typename ArgType2>
+struct result_of<Func(ArgType0,ArgType1,ArgType2)> {
+    template<typename T>
+    static has_std_result_type    testFunctor(T const *, typename T::result_type const * = 0);
+    template<typename T>
+    static has_tr1_result         testFunctor(T const *, typename T::template result<T(ArgType0,ArgType1,ArgType2)>::type const * = 0);
+    static has_none               testFunctor(...);
+
+    // note that the following indirection is needed for gcc-3.3
+    enum {FunctorType = sizeof(testFunctor(static_cast<Func*>(0)))};
+    typedef typename ternary_result_of_select<Func, ArgType0, ArgType1, ArgType2, FunctorType>::type type;
+};
+#endif
+
+struct meta_yes { char a[1]; };
+struct meta_no  { char a[2]; };
+
+// Check whether T::ReturnType does exist
+template <typename T>
+struct has_ReturnType
+{
+  template <typename C> static meta_yes testFunctor(C const *, typename C::ReturnType const * = 0);
+  template <typename C> static meta_no  testFunctor(...);
+
+  enum { value = sizeof(testFunctor<T>(static_cast<T*>(0))) == sizeof(meta_yes) };
+};
+
+template<typename T> const T* return_ptr();
+
+template <typename T, typename IndexType=Index>
+struct has_nullary_operator
+{
+  template <typename C> static meta_yes testFunctor(C const *,typename enable_if<(sizeof(return_ptr<C>()->operator()())>0)>::type * = 0);
+  static meta_no testFunctor(...);
+
+  enum { value = sizeof(testFunctor(static_cast<T*>(0))) == sizeof(meta_yes) };
+};
+
+template <typename T, typename IndexType=Index>
+struct has_unary_operator
+{
+  template <typename C> static meta_yes testFunctor(C const *,typename enable_if<(sizeof(return_ptr<C>()->operator()(IndexType(0)))>0)>::type * = 0);
+  static meta_no testFunctor(...);
+
+  enum { value = sizeof(testFunctor(static_cast<T*>(0))) == sizeof(meta_yes) };
+};
+
+template <typename T, typename IndexType=Index>
+struct has_binary_operator
+{
+  template <typename C> static meta_yes testFunctor(C const *,typename enable_if<(sizeof(return_ptr<C>()->operator()(IndexType(0),IndexType(0)))>0)>::type * = 0);
+  static meta_no testFunctor(...);
+
+  enum { value = sizeof(testFunctor(static_cast<T*>(0))) == sizeof(meta_yes) };
+};
+
 /** \internal In short, it computes int(sqrt(\a Y)) with \a Y an integer.
   * Usage example: \code meta_sqrt<1023>::ret \endcode
   */
@@ -185,37 +491,26 @@ class meta_sqrt
 template<int Y, int InfX, int SupX>
 class meta_sqrt<Y, InfX, SupX, true> { public:  enum { ret = (SupX*SupX <= Y) ? SupX : InfX }; };
 
-/** \internal determines whether the product of two numeric types is allowed and what the return type is */
-template<typename T, typename U> struct scalar_product_traits
-{
-  enum { Defined = 0 };
-};
 
-template<typename T> struct scalar_product_traits<T,T>
+/** \internal Computes the least common multiple of two positive integer A and B
+  * at compile-time. It implements a naive algorithm testing all multiples of A.
+  * It thus works better if A>=B.
+  */
+template<int A, int B, int K=1, bool Done = ((A*K)%B)==0>
+struct meta_least_common_multiple
 {
-  enum {
-    // Cost = NumTraits<T>::MulCost,
-    Defined = 1
-  };
-  typedef T ReturnType;
+  enum { ret = meta_least_common_multiple<A,B,K+1>::ret };
 };
-
-template<typename T> struct scalar_product_traits<T,std::complex<T> >
+template<int A, int B, int K>
+struct meta_least_common_multiple<A,B,K,true>
 {
-  enum {
-    // Cost = 2*NumTraits<T>::MulCost,
-    Defined = 1
-  };
-  typedef std::complex<T> ReturnType;
+  enum { ret = A*K };
 };
 
-template<typename T> struct scalar_product_traits<std::complex<T>, T>
+/** \internal determines whether the product of two numeric types is allowed and what the return type is */
+template<typename T, typename U> struct scalar_product_traits
 {
-  enum {
-    // Cost = 2*NumTraits<T>::MulCost,
-    Defined = 1
-  };
-  typedef std::complex<T> ReturnType;
+  enum { Defined = 0 };
 };
 
 // FIXME quick workaround around current limitation of result_of
@@ -224,19 +519,31 @@ template<typename T> struct scalar_product_traits<std::complex<T>, T>
 // typedef typename scalar_product_traits<typename remove_all<ArgType0>::type, typename remove_all<ArgType1>::type>::ReturnType type;
 // };
 
-template<typename T> struct is_diagonal
-{ enum { ret = false }; };
-
-template<typename T> struct is_diagonal<DiagonalBase<T> >
-{ enum { ret = true }; };
-
-template<typename T> struct is_diagonal<DiagonalWrapper<T> >
-{ enum { ret = true }; };
+} // end namespace internal
 
-template<typename T, int S> struct is_diagonal<DiagonalMatrix<T,S> >
-{ enum { ret = true }; };
+namespace numext {
+  
+#if defined(__CUDA_ARCH__)
+template<typename T> EIGEN_DEVICE_FUNC   void swap(T &a, T &b) { T tmp = b; b = a; a = tmp; }
+#else
+template<typename T> EIGEN_STRONG_INLINE void swap(T &a, T &b) { std::swap(a,b); }
+#endif
+
+#if defined(__CUDA_ARCH__)
+using internal::device::numeric_limits;
+#else
+using std::numeric_limits;
+#endif
+
+// Integer division with rounding up.
+// T is assumed to be an integer type with a>=0, and b>0
+template<typename T>
+T div_ceil(const T &a, const T &b)
+{
+  return (a+b-1) / b;
+}
 
-} // end namespace internal
+} // end namespace numext
 
 } // end namespace Eigen
 
diff --git a/eigen/Eigen/src/Core/util/ReenableStupidWarnings.h b/eigen/Eigen/src/Core/util/ReenableStupidWarnings.h
index d573bbd..86b60f5 100644
--- a/eigen/Eigen/src/Core/util/ReenableStupidWarnings.h
+++ b/eigen/Eigen/src/Core/util/ReenableStupidWarnings.h
@@ -12,6 +12,16 @@
     #pragma GCC diagnostic pop
   #endif
 
+  #if defined __NVCC__
+//    Don't reenable the diagnostic messages, as it turns out these messages need
+//    to be disabled at the point of the template instantiation (i.e the user code)
+//    otherwise they'll be triggered by nvcc.
+//    #pragma diag_default code_is_unreachable
+//    #pragma diag_default initialization_not_reachable
+//    #pragma diag_default 2651
+//    #pragma diag_default 2653
+  #endif
+
 #endif
 
 #endif // EIGEN_WARNINGS_DISABLED
diff --git a/eigen/Eigen/src/Core/util/StaticAssert.h b/eigen/Eigen/src/Core/util/StaticAssert.h
index e53d2b8..983361a 100644
--- a/eigen/Eigen/src/Core/util/StaticAssert.h
+++ b/eigen/Eigen/src/Core/util/StaticAssert.h
@@ -26,7 +26,7 @@
 
 #ifndef EIGEN_NO_STATIC_ASSERT
 
-  #if __has_feature(cxx_static_assert) || (defined(__cplusplus) && __cplusplus >= 201103L) || (EIGEN_COMP_MSVC >= 1600)
+  #if EIGEN_MAX_CPP_VER>=11 && (__has_feature(cxx_static_assert) || (defined(__cplusplus) && __cplusplus >= 201103L) || (EIGEN_COMP_MSVC >= 1600))
 
     // if native static_assert is enabled, let's use it
     #define EIGEN_STATIC_ASSERT(X,MSG) static_assert(X,#MSG);
@@ -50,6 +50,7 @@
         THIS_METHOD_IS_ONLY_FOR_VECTORS_OF_A_SPECIFIC_SIZE,
         THIS_METHOD_IS_ONLY_FOR_MATRICES_OF_A_SPECIFIC_SIZE,
         THIS_METHOD_IS_ONLY_FOR_OBJECTS_OF_A_SPECIFIC_SIZE,
+        OUT_OF_RANGE_ACCESS,
         YOU_MADE_A_PROGRAMMING_MISTAKE,
         EIGEN_INTERNAL_ERROR_PLEASE_FILE_A_BUG_REPORT,
         EIGEN_INTERNAL_COMPILATION_ERROR_OR_YOU_MADE_A_PROGRAMMING_MISTAKE,
@@ -84,6 +85,7 @@
         THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY,
         YOU_ARE_TRYING_TO_USE_AN_INDEX_BASED_ACCESSOR_ON_AN_EXPRESSION_THAT_DOES_NOT_SUPPORT_THAT,
         THIS_METHOD_IS_ONLY_FOR_1x1_EXPRESSIONS,
+        THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS,
         THIS_METHOD_IS_ONLY_FOR_EXPRESSIONS_OF_BOOL,
         THIS_METHOD_IS_ONLY_FOR_ARRAYS_NOT_MATRICES,
         YOU_PASSED_A_ROW_VECTOR_BUT_A_COLUMN_VECTOR_WAS_EXPECTED,
@@ -92,7 +94,14 @@
         THE_STORAGE_ORDER_OF_BOTH_SIDES_MUST_MATCH,
         OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG,
         IMPLICIT_CONVERSION_TO_SCALAR_IS_FOR_INNER_PRODUCT_ONLY,
-        STORAGE_LAYOUT_DOES_NOT_MATCH
+        STORAGE_LAYOUT_DOES_NOT_MATCH,
+        EIGEN_INTERNAL_ERROR_PLEASE_FILE_A_BUG_REPORT__INVALID_COST_VALUE,
+        THIS_COEFFICIENT_ACCESSOR_TAKING_ONE_ACCESS_IS_ONLY_FOR_EXPRESSIONS_ALLOWING_LINEAR_ACCESS,
+        MATRIX_FREE_CONJUGATE_GRADIENT_IS_COMPATIBLE_WITH_UPPER_UNION_LOWER_MODE_ONLY,
+        THIS_TYPE_IS_NOT_SUPPORTED,
+        STORAGE_KIND_MUST_MATCH,
+        STORAGE_INDEX_MUST_MATCH,
+        CHOLMOD_SUPPORTS_DOUBLE_PRECISION_ONLY
       };
     };
 
@@ -103,15 +112,15 @@
     // Specialized implementation for MSVC to avoid "conditional
     // expression is constant" warnings.  This implementation doesn't
     // appear to work under GCC, hence the multiple implementations.
-    #ifdef _MSC_VER
+    #if EIGEN_COMP_MSVC
 
       #define EIGEN_STATIC_ASSERT(CONDITION,MSG) \
         {Eigen::internal::static_assertion<bool(CONDITION)>::MSG;}
 
     #else
-
+      // In some cases clang interprets bool(CONDITION) as function declaration
       #define EIGEN_STATIC_ASSERT(CONDITION,MSG) \
-        if (Eigen::internal::static_assertion<bool(CONDITION)>::MSG) {}
+        if (Eigen::internal::static_assertion<static_cast<bool>(CONDITION)>::MSG) {}
 
     #endif
 
@@ -159,7 +168,7 @@
 
 #define EIGEN_PREDICATE_SAME_MATRIX_SIZE(TYPE0,TYPE1) \
      ( \
-        (int(TYPE0::SizeAtCompileTime)==0 && int(TYPE1::SizeAtCompileTime)==0) \
+        (int(Eigen::internal::size_of_xpr_at_compile_time<TYPE0>::ret)==0 && int(Eigen::internal::size_of_xpr_at_compile_time<TYPE1>::ret)==0) \
     || (\
           (int(TYPE0::RowsAtCompileTime)==Eigen::Dynamic \
         || int(TYPE1::RowsAtCompileTime)==Eigen::Dynamic \
@@ -170,13 +179,8 @@
        ) \
      )
 
-#ifdef EIGEN2_SUPPORT
-  #define EIGEN_STATIC_ASSERT_NON_INTEGER(TYPE) \
-    eigen_assert(!NumTraits<Scalar>::IsInteger);
-#else
-  #define EIGEN_STATIC_ASSERT_NON_INTEGER(TYPE) \
+#define EIGEN_STATIC_ASSERT_NON_INTEGER(TYPE) \
     EIGEN_STATIC_ASSERT(!NumTraits<TYPE>::IsInteger, THIS_FUNCTION_IS_NOT_FOR_INTEGER_NUMERIC_TYPES)
-#endif
 
 
 // static assertion failing if it is guaranteed at compile-time that the two matrix expression types have different sizes
@@ -191,18 +195,22 @@
                           THIS_METHOD_IS_ONLY_FOR_1x1_EXPRESSIONS)
 
 #define EIGEN_STATIC_ASSERT_LVALUE(Derived) \
-      EIGEN_STATIC_ASSERT(internal::is_lvalue<Derived>::value, \
+      EIGEN_STATIC_ASSERT(Eigen::internal::is_lvalue<Derived>::value, \
                           THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY)
 
 #define EIGEN_STATIC_ASSERT_ARRAYXPR(Derived) \
-      EIGEN_STATIC_ASSERT((internal::is_same<typename internal::traits<Derived>::XprKind, ArrayXpr>::value), \
+      EIGEN_STATIC_ASSERT((Eigen::internal::is_same<typename Eigen::internal::traits<Derived>::XprKind, ArrayXpr>::value), \
                           THIS_METHOD_IS_ONLY_FOR_ARRAYS_NOT_MATRICES)
 
 #define EIGEN_STATIC_ASSERT_SAME_XPR_KIND(Derived1, Derived2) \
-      EIGEN_STATIC_ASSERT((internal::is_same<typename internal::traits<Derived1>::XprKind, \
-                                             typename internal::traits<Derived2>::XprKind \
+      EIGEN_STATIC_ASSERT((Eigen::internal::is_same<typename Eigen::internal::traits<Derived1>::XprKind, \
+                                             typename Eigen::internal::traits<Derived2>::XprKind \
                                             >::value), \
                           YOU_CANNOT_MIX_ARRAYS_AND_MATRICES)
 
+// Check that a cost value is positive, and that is stay within a reasonable range
+// TODO this check could be enabled for internal debugging only
+#define EIGEN_INTERNAL_CHECK_COST_VALUE(C) \
+      EIGEN_STATIC_ASSERT((C)>=0 && (C)<=HugeCost*HugeCost, EIGEN_INTERNAL_ERROR_PLEASE_FILE_A_BUG_REPORT__INVALID_COST_VALUE);
 
 #endif // EIGEN_STATIC_ASSERT_H
diff --git a/eigen/Eigen/src/Core/util/SymbolicIndex.h b/eigen/Eigen/src/Core/util/SymbolicIndex.h
new file mode 100644
index 0000000..bb6349e
--- /dev/null
+++ b/eigen/Eigen/src/Core/util/SymbolicIndex.h
@@ -0,0 +1,300 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2017 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SYMBOLIC_INDEX_H
+#define EIGEN_SYMBOLIC_INDEX_H
+
+namespace Eigen {
+
+/** \namespace Eigen::Symbolic
+  * \ingroup Core_Module
+  *
+  * This namespace defines a set of classes and functions to build and evaluate symbolic expressions of scalar type Index.
+  * Here is a simple example:
+  *
+  * \code
+  * // First step, defines symbols:
+  * struct x_tag {};  static const Symbolic::SymbolExpr<x_tag> x;
+  * struct y_tag {};  static const Symbolic::SymbolExpr<y_tag> y;
+  * struct z_tag {};  static const Symbolic::SymbolExpr<z_tag> z;
+  *
+  * // Defines an expression:
+  * auto expr = (x+3)/y+z;
+  *
+  * // And evaluate it: (c++14)
+  * std::cout << expr.eval(x=6,y=3,z=-13) << "\n";
+  *
+  * // In c++98/11, only one symbol per expression is supported for now:
+  * auto expr98 = (3-x)/2;
+  * std::cout << expr98.eval(x=6) << "\n";
+  * \endcode
+  *
+  * It is currently only used internally to define and minipulate the placeholders::last and placeholders::end symbols in Eigen::seq and Eigen::seqN.
+  *
+  */
+namespace Symbolic {
+
+template<typename Tag> class Symbol;
+template<typename Arg0> class NegateExpr;
+template<typename Arg1,typename Arg2> class AddExpr;
+template<typename Arg1,typename Arg2> class ProductExpr;
+template<typename Arg1,typename Arg2> class QuotientExpr;
+
+// A simple wrapper around an integral value to provide the eval method.
+// We could also use a free-function symbolic_eval...
+template<typename IndexType=Index>
+class ValueExpr {
+public:
+  ValueExpr(IndexType val) : m_value(val) {}
+  template<typename T>
+  IndexType eval_impl(const T&) const { return m_value; }
+protected:
+  IndexType m_value;
+};
+
+// Specialization for compile-time value,
+// It is similar to ValueExpr(N) but this version helps the compiler to generate better code.
+template<int N>
+class ValueExpr<internal::FixedInt<N> > {
+public:
+  ValueExpr() {}
+  template<typename T>
+  Index eval_impl(const T&) const { return N; }
+};
+
+
+/** \class BaseExpr
+  * \ingroup Core_Module
+  * Common base class of any symbolic expressions
+  */
+template<typename Derived>
+class BaseExpr
+{
+public:
+  const Derived& derived() const { return *static_cast<const Derived*>(this); }
+
+  /** Evaluate the expression given the \a values of the symbols.
+    *
+    * \param values defines the values of the symbols, it can either be a SymbolValue or a std::tuple of SymbolValue
+    *               as constructed by SymbolExpr::operator= operator.
+    *
+    */
+  template<typename T>
+  Index eval(const T& values) const { return derived().eval_impl(values); }
+
+#if EIGEN_HAS_CXX14
+  template<typename... Types>
+  Index eval(Types&&... values) const { return derived().eval_impl(std::make_tuple(values...)); }
+#endif
+
+  NegateExpr<Derived> operator-() const { return NegateExpr<Derived>(derived()); }
+
+  AddExpr<Derived,ValueExpr<> > operator+(Index b) const
+  { return AddExpr<Derived,ValueExpr<> >(derived(),  b); }
+  AddExpr<Derived,ValueExpr<> > operator-(Index a) const
+  { return AddExpr<Derived,ValueExpr<> >(derived(), -a); }
+  ProductExpr<Derived,ValueExpr<> > operator*(Index a) const
+  { return ProductExpr<Derived,ValueExpr<> >(derived(),a); }
+  QuotientExpr<Derived,ValueExpr<> > operator/(Index a) const
+  { return QuotientExpr<Derived,ValueExpr<> >(derived(),a); }
+
+  friend AddExpr<Derived,ValueExpr<> > operator+(Index a, const BaseExpr& b)
+  { return AddExpr<Derived,ValueExpr<> >(b.derived(), a); }
+  friend AddExpr<NegateExpr<Derived>,ValueExpr<> > operator-(Index a, const BaseExpr& b)
+  { return AddExpr<NegateExpr<Derived>,ValueExpr<> >(-b.derived(), a); }
+  friend ProductExpr<ValueExpr<>,Derived> operator*(Index a, const BaseExpr& b)
+  { return ProductExpr<ValueExpr<>,Derived>(a,b.derived()); }
+  friend QuotientExpr<ValueExpr<>,Derived> operator/(Index a, const BaseExpr& b)
+  { return QuotientExpr<ValueExpr<>,Derived>(a,b.derived()); }
+
+  template<int N>
+  AddExpr<Derived,ValueExpr<internal::FixedInt<N> > > operator+(internal::FixedInt<N>) const
+  { return AddExpr<Derived,ValueExpr<internal::FixedInt<N> > >(derived(), ValueExpr<internal::FixedInt<N> >()); }
+  template<int N>
+  AddExpr<Derived,ValueExpr<internal::FixedInt<-N> > > operator-(internal::FixedInt<N>) const
+  { return AddExpr<Derived,ValueExpr<internal::FixedInt<-N> > >(derived(), ValueExpr<internal::FixedInt<-N> >()); }
+  template<int N>
+  ProductExpr<Derived,ValueExpr<internal::FixedInt<N> > > operator*(internal::FixedInt<N>) const
+  { return ProductExpr<Derived,ValueExpr<internal::FixedInt<N> > >(derived(),ValueExpr<internal::FixedInt<N> >()); }
+  template<int N>
+  QuotientExpr<Derived,ValueExpr<internal::FixedInt<N> > > operator/(internal::FixedInt<N>) const
+  { return QuotientExpr<Derived,ValueExpr<internal::FixedInt<N> > >(derived(),ValueExpr<internal::FixedInt<N> >()); }
+
+  template<int N>
+  friend AddExpr<Derived,ValueExpr<internal::FixedInt<N> > > operator+(internal::FixedInt<N>, const BaseExpr& b)
+  { return AddExpr<Derived,ValueExpr<internal::FixedInt<N> > >(b.derived(), ValueExpr<internal::FixedInt<N> >()); }
+  template<int N>
+  friend AddExpr<NegateExpr<Derived>,ValueExpr<internal::FixedInt<N> > > operator-(internal::FixedInt<N>, const BaseExpr& b)
+  { return AddExpr<NegateExpr<Derived>,ValueExpr<internal::FixedInt<N> > >(-b.derived(), ValueExpr<internal::FixedInt<N> >()); }
+  template<int N>
+  friend ProductExpr<ValueExpr<internal::FixedInt<N> >,Derived> operator*(internal::FixedInt<N>, const BaseExpr& b)
+  { return ProductExpr<ValueExpr<internal::FixedInt<N> >,Derived>(ValueExpr<internal::FixedInt<N> >(),b.derived()); }
+  template<int N>
+  friend QuotientExpr<ValueExpr<internal::FixedInt<N> >,Derived> operator/(internal::FixedInt<N>, const BaseExpr& b)
+  { return QuotientExpr<ValueExpr<internal::FixedInt<N> > ,Derived>(ValueExpr<internal::FixedInt<N> >(),b.derived()); }
+
+#if (!EIGEN_HAS_CXX14)
+  template<int N>
+  AddExpr<Derived,ValueExpr<internal::FixedInt<N> > > operator+(internal::FixedInt<N> (*)()) const
+  { return AddExpr<Derived,ValueExpr<internal::FixedInt<N> > >(derived(), ValueExpr<internal::FixedInt<N> >()); }
+  template<int N>
+  AddExpr<Derived,ValueExpr<internal::FixedInt<-N> > > operator-(internal::FixedInt<N> (*)()) const
+  { return AddExpr<Derived,ValueExpr<internal::FixedInt<-N> > >(derived(), ValueExpr<internal::FixedInt<-N> >()); }
+  template<int N>
+  ProductExpr<Derived,ValueExpr<internal::FixedInt<N> > > operator*(internal::FixedInt<N> (*)()) const
+  { return ProductExpr<Derived,ValueExpr<internal::FixedInt<N> > >(derived(),ValueExpr<internal::FixedInt<N> >()); }
+  template<int N>
+  QuotientExpr<Derived,ValueExpr<internal::FixedInt<N> > > operator/(internal::FixedInt<N> (*)()) const
+  { return QuotientExpr<Derived,ValueExpr<internal::FixedInt<N> > >(derived(),ValueExpr<internal::FixedInt<N> >()); }
+
+  template<int N>
+  friend AddExpr<Derived,ValueExpr<internal::FixedInt<N> > > operator+(internal::FixedInt<N> (*)(), const BaseExpr& b)
+  { return AddExpr<Derived,ValueExpr<internal::FixedInt<N> > >(b.derived(), ValueExpr<internal::FixedInt<N> >()); }
+  template<int N>
+  friend AddExpr<NegateExpr<Derived>,ValueExpr<internal::FixedInt<N> > > operator-(internal::FixedInt<N> (*)(), const BaseExpr& b)
+  { return AddExpr<NegateExpr<Derived>,ValueExpr<internal::FixedInt<N> > >(-b.derived(), ValueExpr<internal::FixedInt<N> >()); }
+  template<int N>
+  friend ProductExpr<ValueExpr<internal::FixedInt<N> >,Derived> operator*(internal::FixedInt<N> (*)(), const BaseExpr& b)
+  { return ProductExpr<ValueExpr<internal::FixedInt<N> >,Derived>(ValueExpr<internal::FixedInt<N> >(),b.derived()); }
+  template<int N>
+  friend QuotientExpr<ValueExpr<internal::FixedInt<N> >,Derived> operator/(internal::FixedInt<N> (*)(), const BaseExpr& b)
+  { return QuotientExpr<ValueExpr<internal::FixedInt<N> > ,Derived>(ValueExpr<internal::FixedInt<N> >(),b.derived()); }
+#endif
+
+
+  template<typename OtherDerived>
+  AddExpr<Derived,OtherDerived> operator+(const BaseExpr<OtherDerived> &b) const
+  { return AddExpr<Derived,OtherDerived>(derived(),  b.derived()); }
+
+  template<typename OtherDerived>
+  AddExpr<Derived,NegateExpr<OtherDerived> > operator-(const BaseExpr<OtherDerived> &b) const
+  { return AddExpr<Derived,NegateExpr<OtherDerived> >(derived(), -b.derived()); }
+
+  template<typename OtherDerived>
+  ProductExpr<Derived,OtherDerived> operator*(const BaseExpr<OtherDerived> &b) const
+  { return ProductExpr<Derived,OtherDerived>(derived(), b.derived()); }
+
+  template<typename OtherDerived>
+  QuotientExpr<Derived,OtherDerived> operator/(const BaseExpr<OtherDerived> &b) const
+  { return QuotientExpr<Derived,OtherDerived>(derived(), b.derived()); }
+};
+
+template<typename T>
+struct is_symbolic {
+  // BaseExpr has no conversion ctor, so we only have to check whether T can be staticaly cast to its base class BaseExpr<T>.
+  enum { value = internal::is_convertible<T,BaseExpr<T> >::value };
+};
+
+// Specialization for functions, because is_convertible fails in this case.
+// Useful in c++98/11 mode when testing is_symbolic<decltype(fix<N>)>
+template<typename T>
+struct is_symbolic<T (*)()> {
+  enum { value = false };
+};
+
+/** Represents the actual value of a symbol identified by its tag
+  *
+  * It is the return type of SymbolValue::operator=, and most of the time this is only way it is used.
+  */
+template<typename Tag>
+class SymbolValue
+{
+public:
+  /** Default constructor from the value \a val */
+  SymbolValue(Index val) : m_value(val) {}
+
+  /** \returns the stored value of the symbol */
+  Index value() const { return m_value; }
+protected:
+  Index m_value;
+};
+
+/** Expression of a symbol uniquely identified by the template parameter type \c tag */
+template<typename tag>
+class SymbolExpr : public BaseExpr<SymbolExpr<tag> >
+{
+public:
+  /** Alias to the template parameter \c tag */
+  typedef tag Tag;
+
+  SymbolExpr() {}
+
+  /** Associate the value \a val to the given symbol \c *this, uniquely identified by its \c Tag.
+    *
+    * The returned object should be passed to ExprBase::eval() to evaluate a given expression with this specified runtime-time value.
+    */
+  SymbolValue<Tag> operator=(Index val) const {
+    return SymbolValue<Tag>(val);
+  }
+
+  Index eval_impl(const SymbolValue<Tag> &values) const { return values.value(); }
+
+#if EIGEN_HAS_CXX14
+  // C++14 versions suitable for multiple symbols
+  template<typename... Types>
+  Index eval_impl(const std::tuple<Types...>& values) const { return std::get<SymbolValue<Tag> >(values).value(); }
+#endif
+};
+
+template<typename Arg0>
+class NegateExpr : public BaseExpr<NegateExpr<Arg0> >
+{
+public:
+  NegateExpr(const Arg0& arg0) : m_arg0(arg0) {}
+
+  template<typename T>
+  Index eval_impl(const T& values) const { return -m_arg0.eval_impl(values); }
+protected:
+  Arg0 m_arg0;
+};
+
+template<typename Arg0, typename Arg1>
+class AddExpr : public BaseExpr<AddExpr<Arg0,Arg1> >
+{
+public:
+  AddExpr(const Arg0& arg0, const Arg1& arg1) : m_arg0(arg0), m_arg1(arg1) {}
+
+  template<typename T>
+  Index eval_impl(const T& values) const { return m_arg0.eval_impl(values) + m_arg1.eval_impl(values); }
+protected:
+  Arg0 m_arg0;
+  Arg1 m_arg1;
+};
+
+template<typename Arg0, typename Arg1>
+class ProductExpr : public BaseExpr<ProductExpr<Arg0,Arg1> >
+{
+public:
+  ProductExpr(const Arg0& arg0, const Arg1& arg1) : m_arg0(arg0), m_arg1(arg1) {}
+
+  template<typename T>
+  Index eval_impl(const T& values) const { return m_arg0.eval_impl(values) * m_arg1.eval_impl(values); }
+protected:
+  Arg0 m_arg0;
+  Arg1 m_arg1;
+};
+
+template<typename Arg0, typename Arg1>
+class QuotientExpr : public BaseExpr<QuotientExpr<Arg0,Arg1> >
+{
+public:
+  QuotientExpr(const Arg0& arg0, const Arg1& arg1) : m_arg0(arg0), m_arg1(arg1) {}
+
+  template<typename T>
+  Index eval_impl(const T& values) const { return m_arg0.eval_impl(values) / m_arg1.eval_impl(values); }
+protected:
+  Arg0 m_arg0;
+  Arg1 m_arg1;
+};
+
+} // end namespace Symbolic
+
+} // end namespace Eigen
+
+#endif // EIGEN_SYMBOLIC_INDEX_H
diff --git a/eigen/Eigen/src/Core/util/XprHelper.h b/eigen/Eigen/src/Core/util/XprHelper.h
index d05f8e5..4b337f2 100644
--- a/eigen/Eigen/src/Core/util/XprHelper.h
+++ b/eigen/Eigen/src/Core/util/XprHelper.h
@@ -14,20 +14,77 @@
 // just a workaround because GCC seems to not really like empty structs
 // FIXME: gcc 4.3 generates bad code when strict-aliasing is enabled
 // so currently we simply disable this optimization for gcc 4.3
-#if (defined __GNUG__) && !((__GNUC__==4) && (__GNUC_MINOR__==3))
+#if EIGEN_COMP_GNUC && !EIGEN_GNUC_AT(4,3)
   #define EIGEN_EMPTY_STRUCT_CTOR(X) \
-    EIGEN_STRONG_INLINE X() {} \
-    EIGEN_STRONG_INLINE X(const X& ) {}
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE X() {} \
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE X(const X& ) {}
 #else
   #define EIGEN_EMPTY_STRUCT_CTOR(X)
 #endif
 
 namespace Eigen {
 
-typedef EIGEN_DEFAULT_DENSE_INDEX_TYPE DenseIndex;
-
 namespace internal {
 
+template<typename IndexDest, typename IndexSrc>
+EIGEN_DEVICE_FUNC
+inline IndexDest convert_index(const IndexSrc& idx) {
+  // for sizeof(IndexDest)>=sizeof(IndexSrc) compilers should be able to optimize this away:
+  eigen_internal_assert(idx <= NumTraits<IndexDest>::highest() && "Index value to big for target type");
+  return IndexDest(idx);
+}
+
+
+// promote_scalar_arg is an helper used in operation between an expression and a scalar, like:
+//    expression * scalar
+// Its role is to determine how the type T of the scalar operand should be promoted given the scalar type ExprScalar of the given expression.
+// The IsSupported template parameter must be provided by the caller as: internal::has_ReturnType<ScalarBinaryOpTraits<ExprScalar,T,op> >::value using the proper order for ExprScalar and T.
+// Then the logic is as follows:
+//  - if the operation is natively supported as defined by IsSupported, then the scalar type is not promoted, and T is returned.
+//  - otherwise, NumTraits<ExprScalar>::Literal is returned if T is implicitly convertible to NumTraits<ExprScalar>::Literal AND that this does not imply a float to integer conversion.
+//  - otherwise, ExprScalar is returned if T is implicitly convertible to ExprScalar AND that this does not imply a float to integer conversion.
+//  - In all other cases, the promoted type is not defined, and the respective operation is thus invalid and not available (SFINAE).
+template<typename ExprScalar,typename T, bool IsSupported>
+struct promote_scalar_arg;
+
+template<typename S,typename T>
+struct promote_scalar_arg<S,T,true>
+{
+  typedef T type;
+};
+
+// Recursively check safe conversion to PromotedType, and then ExprScalar if they are different.
+template<typename ExprScalar,typename T,typename PromotedType,
+  bool ConvertibleToLiteral = internal::is_convertible<T,PromotedType>::value,
+  bool IsSafe = NumTraits<T>::IsInteger || !NumTraits<PromotedType>::IsInteger>
+struct promote_scalar_arg_unsupported;
+
+// Start recursion with NumTraits<ExprScalar>::Literal
+template<typename S,typename T>
+struct promote_scalar_arg<S,T,false> : promote_scalar_arg_unsupported<S,T,typename NumTraits<S>::Literal> {};
+
+// We found a match!
+template<typename S,typename T, typename PromotedType>
+struct promote_scalar_arg_unsupported<S,T,PromotedType,true,true>
+{
+  typedef PromotedType type;
+};
+
+// No match, but no real-to-integer issues, and ExprScalar and current PromotedType are different,
+// so let's try to promote to ExprScalar
+template<typename ExprScalar,typename T, typename PromotedType>
+struct promote_scalar_arg_unsupported<ExprScalar,T,PromotedType,false,true>
+   : promote_scalar_arg_unsupported<ExprScalar,T,ExprScalar>
+{};
+
+// Unsafe real-to-integer, let's stop.
+template<typename S,typename T, typename PromotedType, bool ConvertibleToLiteral>
+struct promote_scalar_arg_unsupported<S,T,PromotedType,ConvertibleToLiteral,false> {};
+
+// T is not even convertible to ExprScalar, let's stop.
+template<typename S,typename T>
+struct promote_scalar_arg_unsupported<S,T,S,false,true> {};
+
 //classes inheriting no_assignment_operator don't generate a default operator=.
 class no_assignment_operator
 {
@@ -50,19 +107,21 @@ template<typename T, int Value> class variable_if_dynamic
 {
   public:
     EIGEN_EMPTY_STRUCT_CTOR(variable_if_dynamic)
-    explicit variable_if_dynamic(T v) { EIGEN_ONLY_USED_FOR_DEBUG(v); assert(v == T(Value)); }
-    static T value() { return T(Value); }
-    void setValue(T) {}
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit variable_if_dynamic(T v) { EIGEN_ONLY_USED_FOR_DEBUG(v); eigen_assert(v == T(Value)); }
+    EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE T value() { return T(Value); }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE operator T() const { return T(Value); }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void setValue(T) {}
 };
 
 template<typename T> class variable_if_dynamic<T, Dynamic>
 {
     T m_value;
-    variable_if_dynamic() { assert(false); }
+    EIGEN_DEVICE_FUNC variable_if_dynamic() { eigen_assert(false); }
   public:
-    explicit variable_if_dynamic(T value) : m_value(value) {}
-    T value() const { return m_value; }
-    void setValue(T value) { m_value = value; }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit variable_if_dynamic(T value) : m_value(value) {}
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T value() const { return m_value; }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE operator T() const { return m_value; }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void setValue(T value) { m_value = value; }
 };
 
 /** \internal like variable_if_dynamic but for DynamicIndex
@@ -71,19 +130,19 @@ template<typename T, int Value> class variable_if_dynamicindex
 {
   public:
     EIGEN_EMPTY_STRUCT_CTOR(variable_if_dynamicindex)
-    explicit variable_if_dynamicindex(T v) { EIGEN_ONLY_USED_FOR_DEBUG(v); assert(v == T(Value)); }
-    static T value() { return T(Value); }
-    void setValue(T) {}
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit variable_if_dynamicindex(T v) { EIGEN_ONLY_USED_FOR_DEBUG(v); eigen_assert(v == T(Value)); }
+    EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE T value() { return T(Value); }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void setValue(T) {}
 };
 
 template<typename T> class variable_if_dynamicindex<T, DynamicIndex>
 {
     T m_value;
-    variable_if_dynamicindex() { assert(false); }
+    EIGEN_DEVICE_FUNC variable_if_dynamicindex() { eigen_assert(false); }
   public:
-    explicit variable_if_dynamicindex(T value) : m_value(value) {}
-    T value() const { return m_value; }
-    void setValue(T value) { m_value = value; }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit variable_if_dynamicindex(T value) : m_value(value) {}
+    EIGEN_DEVICE_FUNC T EIGEN_STRONG_INLINE value() const { return m_value; }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void setValue(T value) { m_value = value; }
 };
 
 template<typename T> struct functor_traits
@@ -101,7 +160,73 @@ template<typename T> struct packet_traits;
 template<typename T> struct unpacket_traits
 {
   typedef T type;
-  enum {size=1};
+  typedef T half;
+  enum
+  {
+    size = 1,
+    alignment = 1
+  };
+};
+
+template<int Size, typename PacketType,
+         bool Stop = Size==Dynamic || (Size%unpacket_traits<PacketType>::size)==0 || is_same<PacketType,typename unpacket_traits<PacketType>::half>::value>
+struct find_best_packet_helper;
+
+template< int Size, typename PacketType>
+struct find_best_packet_helper<Size,PacketType,true>
+{
+  typedef PacketType type;
+};
+
+template<int Size, typename PacketType>
+struct find_best_packet_helper<Size,PacketType,false>
+{
+  typedef typename find_best_packet_helper<Size,typename unpacket_traits<PacketType>::half>::type type;
+};
+
+template<typename T, int Size>
+struct find_best_packet
+{
+  typedef typename find_best_packet_helper<Size,typename packet_traits<T>::type>::type type;
+};
+
+#if EIGEN_MAX_STATIC_ALIGN_BYTES>0
+template<int ArrayBytes, int AlignmentBytes,
+         bool Match     =  bool((ArrayBytes%AlignmentBytes)==0),
+         bool TryHalf   =  bool(EIGEN_MIN_ALIGN_BYTES<AlignmentBytes) >
+struct compute_default_alignment_helper
+{
+  enum { value = 0 };
+};
+
+template<int ArrayBytes, int AlignmentBytes, bool TryHalf>
+struct compute_default_alignment_helper<ArrayBytes, AlignmentBytes, true, TryHalf> // Match
+{
+  enum { value = AlignmentBytes };
+};
+
+template<int ArrayBytes, int AlignmentBytes>
+struct compute_default_alignment_helper<ArrayBytes, AlignmentBytes, false, true> // Try-half
+{
+  // current packet too large, try with an half-packet
+  enum { value = compute_default_alignment_helper<ArrayBytes, AlignmentBytes/2>::value };
+};
+#else
+// If static alignment is disabled, no need to bother.
+// This also avoids a division by zero in "bool Match =  bool((ArrayBytes%AlignmentBytes)==0)"
+template<int ArrayBytes, int AlignmentBytes>
+struct compute_default_alignment_helper
+{
+  enum { value = 0 };
+};
+#endif
+
+template<typename T, int Size> struct compute_default_alignment {
+  enum { value = compute_default_alignment_helper<Size*sizeof(T),EIGEN_MAX_STATIC_ALIGN_BYTES>::value };
+};
+
+template<typename T> struct compute_default_alignment<T,Dynamic> {
+  enum { value = EIGEN_MAX_ALIGN_BYTES };
 };
 
 template<typename _Scalar, int _Rows, int _Cols,
@@ -127,35 +252,12 @@ template<typename _Scalar, int _Rows, int _Cols,
 template<typename Scalar, int Rows, int Cols, int Options, int MaxRows, int MaxCols>
 class compute_matrix_flags
 {
-    enum {
-      row_major_bit = Options&RowMajor ? RowMajorBit : 0,
-      is_dynamic_size_storage = MaxRows==Dynamic || MaxCols==Dynamic,
-
-      aligned_bit =
-      (
-            ((Options&DontAlign)==0)
-        && (
-#if EIGEN_ALIGN_STATICALLY
-             ((!is_dynamic_size_storage) && (((MaxCols*MaxRows*int(sizeof(Scalar))) % 16) == 0))
-#else
-             0
-#endif
-
-          ||
-
-#if EIGEN_ALIGN
-             is_dynamic_size_storage
-#else
-             0
-#endif
-
-          )
-      ) ? AlignedBit : 0,
-      packet_access_bit = packet_traits<Scalar>::Vectorizable && aligned_bit ? PacketAccessBit : 0
-    };
-
+    enum { row_major_bit = Options&RowMajor ? RowMajorBit : 0 };
   public:
-    enum { ret = LinearAccessBit | LvalueBit | DirectAccessBit | NestByRefBit | packet_access_bit | row_major_bit | aligned_bit };
+    // FIXME currently we still have to handle DirectAccessBit at the expression level to handle DenseCoeffsBase<>
+    // and then propagate this information to the evaluator's flags.
+    // However, I (Gael) think that DirectAccessBit should only matter at the evaluation stage.
+    enum { ret = DirectAccessBit | LvalueBit | NestByRefBit | row_major_bit };
 };
 
 template<int _Rows, int _Cols> struct size_at_compile_time
@@ -163,34 +265,43 @@ template<int _Rows, int _Cols> struct size_at_compile_time
   enum { ret = (_Rows==Dynamic || _Cols==Dynamic) ? Dynamic : _Rows * _Cols };
 };
 
+template<typename XprType> struct size_of_xpr_at_compile_time
+{
+  enum { ret = size_at_compile_time<traits<XprType>::RowsAtCompileTime,traits<XprType>::ColsAtCompileTime>::ret };
+};
+
 /* plain_matrix_type : the difference from eval is that plain_matrix_type is always a plain matrix type,
  * whereas eval is a const reference in the case of a matrix
  */
 
 template<typename T, typename StorageKind = typename traits<T>::StorageKind> struct plain_matrix_type;
-template<typename T, typename BaseClassType> struct plain_matrix_type_dense;
+template<typename T, typename BaseClassType, int Flags> struct plain_matrix_type_dense;
 template<typename T> struct plain_matrix_type<T,Dense>
 {
-  typedef typename plain_matrix_type_dense<T,typename traits<T>::XprKind>::type type;
+  typedef typename plain_matrix_type_dense<T,typename traits<T>::XprKind, traits<T>::Flags>::type type;
+};
+template<typename T> struct plain_matrix_type<T,DiagonalShape>
+{
+  typedef typename T::PlainObject type;
 };
 
-template<typename T> struct plain_matrix_type_dense<T,MatrixXpr>
+template<typename T, int Flags> struct plain_matrix_type_dense<T,MatrixXpr,Flags>
 {
   typedef Matrix<typename traits<T>::Scalar,
                 traits<T>::RowsAtCompileTime,
                 traits<T>::ColsAtCompileTime,
-                AutoAlign | (traits<T>::Flags&RowMajorBit ? RowMajor : ColMajor),
+                AutoAlign | (Flags&RowMajorBit ? RowMajor : ColMajor),
                 traits<T>::MaxRowsAtCompileTime,
                 traits<T>::MaxColsAtCompileTime
           > type;
 };
 
-template<typename T> struct plain_matrix_type_dense<T,ArrayXpr>
+template<typename T, int Flags> struct plain_matrix_type_dense<T,ArrayXpr,Flags>
 {
   typedef Array<typename traits<T>::Scalar,
                 traits<T>::RowsAtCompileTime,
                 traits<T>::ColsAtCompileTime,
-                AutoAlign | (traits<T>::Flags&RowMajorBit ? RowMajor : ColMajor),
+                AutoAlign | (Flags&RowMajorBit ? RowMajor : ColMajor),
                 traits<T>::MaxRowsAtCompileTime,
                 traits<T>::MaxColsAtCompileTime
           > type;
@@ -215,6 +326,11 @@ template<typename T> struct eval<T,Dense>
 //           > type;
 };
 
+template<typename T> struct eval<T,DiagonalShape>
+{
+  typedef typename plain_matrix_type<T>::type type;
+};
+
 // for matrices, no need to evaluate, just use a const reference to avoid a useless copy
 template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
 struct eval<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>, Dense>
@@ -229,6 +345,15 @@ struct eval<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>, Dense>
 };
 
 
+/* similar to plain_matrix_type, but using the evaluator's Flags */
+template<typename T, typename StorageKind = typename traits<T>::StorageKind> struct plain_object_eval;
+
+template<typename T>
+struct plain_object_eval<T,Dense>
+{
+  typedef typename plain_matrix_type_dense<T,typename traits<T>::XprKind, evaluator<T>::Flags>::type type;
+};
+
 
 /* plain_matrix_type_column_major : same as plain_matrix_type but guaranteed to be column-major
  */
@@ -266,9 +391,6 @@ template<typename T> struct plain_matrix_type_row_major
           > type;
 };
 
-// we should be able to get rid of this one too
-template<typename T> struct must_nest_by_value { enum { ret = false }; };
-
 /** \internal The reference selector for template expressions. The idea is that we don't
   * need to use references for expressions since they are light weight proxy
   * objects which should generate no copying overhead. */
@@ -280,6 +402,12 @@ struct ref_selector
     T const&,
     const T
   >::type type;
+  
+  typedef typename conditional<
+    bool(traits<T>::Flags & NestByRefBit),
+    T &,
+    T
+  >::type non_const_type;
 };
 
 /** \internal Adds the const qualifier on the value-type of T2 if and only if T1 is a const type */
@@ -293,54 +421,41 @@ struct transfer_constness
   >::type type;
 };
 
-/** \internal Determines how a given expression should be nested into another one.
+
+// However, we still need a mechanism to detect whether an expression which is evaluated multiple time
+// has to be evaluated into a temporary.
+// That's the purpose of this new nested_eval helper:
+/** \internal Determines how a given expression should be nested when evaluated multiple times.
   * For example, when you do a * (b+c), Eigen will determine how the expression b+c should be
-  * nested into the bigger product expression. The choice is between nesting the expression b+c as-is, or
+  * evaluated into the bigger product expression. The choice is between nesting the expression b+c as-is, or
   * evaluating that expression b+c into a temporary variable d, and nest d so that the resulting expression is
   * a*d. Evaluating can be beneficial for example if every coefficient access in the resulting expression causes
   * many coefficient accesses in the nested expressions -- as is the case with matrix product for example.
   *
-  * \param T the type of the expression being nested
-  * \param n the number of coefficient accesses in the nested expression for each coefficient access in the bigger expression.
-  *
-  * Note that if no evaluation occur, then the constness of T is preserved.
-  *
-  * Example. Suppose that a, b, and c are of type Matrix3d. The user forms the expression a*(b+c).
-  * b+c is an expression "sum of matrices", which we will denote by S. In order to determine how to nest it,
-  * the Product expression uses: nested<S, 3>::ret, which turns out to be Matrix3d because the internal logic of
-  * nested determined that in this case it was better to evaluate the expression b+c into a temporary. On the other hand,
-  * since a is of type Matrix3d, the Product expression nests it as nested<Matrix3d, 3>::ret, which turns out to be
-  * const Matrix3d&, because the internal logic of nested determined that since a was already a matrix, there was no point
-  * in copying it into another matrix.
+  * \tparam T the type of the expression being nested.
+  * \tparam n the number of coefficient accesses in the nested expression for each coefficient access in the bigger expression.
+  * \tparam PlainObject the type of the temporary if needed.
   */
-template<typename T, int n=1, typename PlainObject = typename eval<T>::type> struct nested
+template<typename T, int n, typename PlainObject = typename plain_object_eval<T>::type> struct nested_eval
 {
   enum {
-    // for the purpose of this test, to keep it reasonably simple, we arbitrarily choose a value of Dynamic values.
-    // the choice of 10000 makes it larger than any practical fixed value and even most dynamic values.
-    // in extreme cases where these assumptions would be wrong, we would still at worst suffer performance issues
-    // (poor choice of temporaries).
-    // it's important that this value can still be squared without integer overflowing.
-    DynamicAsInteger = 10000,
     ScalarReadCost = NumTraits<typename traits<T>::Scalar>::ReadCost,
-    ScalarReadCostAsInteger = ScalarReadCost == Dynamic ? int(DynamicAsInteger) : int(ScalarReadCost),
-    CoeffReadCost = traits<T>::CoeffReadCost,
-    CoeffReadCostAsInteger = CoeffReadCost == Dynamic ? int(DynamicAsInteger) : int(CoeffReadCost),
-    NAsInteger = n == Dynamic ? int(DynamicAsInteger) : n,
-    CostEvalAsInteger   = (NAsInteger+1) * ScalarReadCostAsInteger + CoeffReadCostAsInteger,
-    CostNoEvalAsInteger = NAsInteger * CoeffReadCostAsInteger
+    CoeffReadCost = evaluator<T>::CoeffReadCost,  // NOTE What if an evaluator evaluate itself into a tempory?
+                                                  //      Then CoeffReadCost will be small (e.g., 1) but we still have to evaluate, especially if n>1.
+                                                  //      This situation is already taken care by the EvalBeforeNestingBit flag, which is turned ON
+                                                  //      for all evaluator creating a temporary. This flag is then propagated by the parent evaluators.
+                                                  //      Another solution could be to count the number of temps?
+    NAsInteger = n == Dynamic ? HugeCost : n,
+    CostEval   = (NAsInteger+1) * ScalarReadCost + CoeffReadCost,
+    CostNoEval = NAsInteger * CoeffReadCost,
+    Evaluate = (int(evaluator<T>::Flags) & EvalBeforeNestingBit) || (int(CostEval) < int(CostNoEval))
   };
 
-  typedef typename conditional<
-      ( (int(traits<T>::Flags) & EvalBeforeNestingBit) ||
-        int(CostEvalAsInteger) < int(CostNoEvalAsInteger)
-      ),
-      PlainObject,
-      typename ref_selector<T>::type
-  >::type type;
+  typedef typename conditional<Evaluate, PlainObject, typename ref_selector<T>::type>::type type;
 };
 
 template<typename T>
+EIGEN_DEVICE_FUNC
 inline T* const_cast_ptr(const T* ptr)
 {
   return const_cast<T*>(ptr);
@@ -364,30 +479,13 @@ struct dense_xpr_base<Derived, ArrayXpr>
   typedef ArrayBase<Derived> type;
 };
 
-/** \internal Helper base class to add a scalar multiple operator
-  * overloads for complex types */
-template<typename Derived, typename Scalar, typename OtherScalar, typename BaseType,
-         bool EnableIt = !is_same<Scalar,OtherScalar>::value >
-struct special_scalar_op_base : public BaseType
-{
-  // dummy operator* so that the
-  // "using special_scalar_op_base::operator*" compiles
-  void operator*() const;
-};
+template<typename Derived, typename XprKind = typename traits<Derived>::XprKind, typename StorageKind = typename traits<Derived>::StorageKind>
+struct generic_xpr_base;
 
-template<typename Derived,typename Scalar,typename OtherScalar, typename BaseType>
-struct special_scalar_op_base<Derived,Scalar,OtherScalar,BaseType,true>  : public BaseType
+template<typename Derived, typename XprKind>
+struct generic_xpr_base<Derived, XprKind, Dense>
 {
-  const CwiseUnaryOp<scalar_multiple2_op<Scalar,OtherScalar>, Derived>
-  operator*(const OtherScalar& scalar) const
-  {
-    return CwiseUnaryOp<scalar_multiple2_op<Scalar,OtherScalar>, Derived>
-      (*static_cast<const Derived*>(this), scalar_multiple2_op<Scalar,OtherScalar>(scalar));
-  }
-
-  inline friend const CwiseUnaryOp<scalar_multiple2_op<Scalar,OtherScalar>, Derived>
-  operator*(const OtherScalar& scalar, const Derived& matrix)
-  { return static_cast<const special_scalar_op_base&>(matrix).operator*(scalar); }
+  typedef typename dense_xpr_base<Derived,XprKind>::type type;
 };
 
 template<typename XprType, typename CastType> struct cast_return_type
@@ -405,9 +503,79 @@ template <typename A> struct promote_storage_type<A,A>
 {
   typedef A ret;
 };
+template <typename A> struct promote_storage_type<A, const A>
+{
+  typedef A ret;
+};
+template <typename A> struct promote_storage_type<const A, A>
+{
+  typedef A ret;
+};
+
+/** \internal Specify the "storage kind" of applying a coefficient-wise
+  * binary operations between two expressions of kinds A and B respectively.
+  * The template parameter Functor permits to specialize the resulting storage kind wrt to
+  * the functor.
+  * The default rules are as follows:
+  * \code
+  * A      op A      -> A
+  * A      op dense  -> dense
+  * dense  op B      -> dense
+  * sparse op dense  -> sparse
+  * dense  op sparse -> sparse
+  * \endcode
+  */
+template <typename A, typename B, typename Functor> struct cwise_promote_storage_type;
+
+template <typename A, typename Functor>                   struct cwise_promote_storage_type<A,A,Functor>                                      { typedef A      ret; };
+template <typename Functor>                               struct cwise_promote_storage_type<Dense,Dense,Functor>                              { typedef Dense  ret; };
+template <typename A, typename Functor>                   struct cwise_promote_storage_type<A,Dense,Functor>                                  { typedef Dense  ret; };
+template <typename B, typename Functor>                   struct cwise_promote_storage_type<Dense,B,Functor>                                  { typedef Dense  ret; };
+template <typename Functor>                               struct cwise_promote_storage_type<Sparse,Dense,Functor>                             { typedef Sparse ret; };
+template <typename Functor>                               struct cwise_promote_storage_type<Dense,Sparse,Functor>                             { typedef Sparse ret; };
+
+template <typename LhsKind, typename RhsKind, int LhsOrder, int RhsOrder> struct cwise_promote_storage_order {
+  enum { value = LhsOrder };
+};
+
+template <typename LhsKind, int LhsOrder, int RhsOrder>   struct cwise_promote_storage_order<LhsKind,Sparse,LhsOrder,RhsOrder>                { enum { value = RhsOrder }; };
+template <typename RhsKind, int LhsOrder, int RhsOrder>   struct cwise_promote_storage_order<Sparse,RhsKind,LhsOrder,RhsOrder>                { enum { value = LhsOrder }; };
+template <int Order>                                      struct cwise_promote_storage_order<Sparse,Sparse,Order,Order>                       { enum { value = Order }; };
+
+
+/** \internal Specify the "storage kind" of multiplying an expression of kind A with kind B.
+  * The template parameter ProductTag permits to specialize the resulting storage kind wrt to
+  * some compile-time properties of the product: GemmProduct, GemvProduct, OuterProduct, InnerProduct.
+  * The default rules are as follows:
+  * \code
+  *  K * K            -> K
+  *  dense * K        -> dense
+  *  K * dense        -> dense
+  *  diag * K         -> K
+  *  K * diag         -> K
+  *  Perm * K         -> K
+  * K * Perm          -> K
+  * \endcode
+  */
+template <typename A, typename B, int ProductTag> struct product_promote_storage_type;
+
+template <typename A, int ProductTag> struct product_promote_storage_type<A,                  A,                  ProductTag> { typedef A     ret;};
+template <int ProductTag>             struct product_promote_storage_type<Dense,              Dense,              ProductTag> { typedef Dense ret;};
+template <typename A, int ProductTag> struct product_promote_storage_type<A,                  Dense,              ProductTag> { typedef Dense ret; };
+template <typename B, int ProductTag> struct product_promote_storage_type<Dense,              B,                  ProductTag> { typedef Dense ret; };
+
+template <typename A, int ProductTag> struct product_promote_storage_type<A,                  DiagonalShape,      ProductTag> { typedef A ret; };
+template <typename B, int ProductTag> struct product_promote_storage_type<DiagonalShape,      B,                  ProductTag> { typedef B ret; };
+template <int ProductTag>             struct product_promote_storage_type<Dense,              DiagonalShape,      ProductTag> { typedef Dense ret; };
+template <int ProductTag>             struct product_promote_storage_type<DiagonalShape,      Dense,              ProductTag> { typedef Dense ret; };
+
+template <typename A, int ProductTag> struct product_promote_storage_type<A,                  PermutationStorage, ProductTag> { typedef A ret; };
+template <typename B, int ProductTag> struct product_promote_storage_type<PermutationStorage, B,                  ProductTag> { typedef B ret; };
+template <int ProductTag>             struct product_promote_storage_type<Dense,              PermutationStorage, ProductTag> { typedef Dense ret; };
+template <int ProductTag>             struct product_promote_storage_type<PermutationStorage, Dense,              ProductTag> { typedef Dense ret; };
 
 /** \internal gives the plain matrix or array type to store a row/column/diagonal of a matrix type.
-  * \param Scalar optional parameter allowing to pass a different scalar type than the one of the MatrixType.
+  * \tparam Scalar optional parameter allowing to pass a different scalar type than the one of the MatrixType.
   */
 template<typename ExpressionType, typename Scalar = typename ExpressionType::Scalar>
 struct plain_row_type
@@ -455,15 +623,201 @@ struct plain_diag_type
   >::type type;
 };
 
+template<typename Expr,typename Scalar = typename Expr::Scalar>
+struct plain_constant_type
+{
+  enum { Options = (traits<Expr>::Flags&RowMajorBit)?RowMajor:0 };
+
+  typedef Array<Scalar,  traits<Expr>::RowsAtCompileTime,   traits<Expr>::ColsAtCompileTime,
+                Options, traits<Expr>::MaxRowsAtCompileTime,traits<Expr>::MaxColsAtCompileTime> array_type;
+
+  typedef Matrix<Scalar,  traits<Expr>::RowsAtCompileTime,   traits<Expr>::ColsAtCompileTime,
+                 Options, traits<Expr>::MaxRowsAtCompileTime,traits<Expr>::MaxColsAtCompileTime> matrix_type;
+
+  typedef CwiseNullaryOp<scalar_constant_op<Scalar>, const typename conditional<is_same< typename traits<Expr>::XprKind, MatrixXpr >::value, matrix_type, array_type>::type > type;
+};
+
 template<typename ExpressionType>
 struct is_lvalue
 {
-  enum { value = !bool(is_const<ExpressionType>::value) &&
+  enum { value = (!bool(is_const<ExpressionType>::value)) &&
                  bool(traits<ExpressionType>::Flags & LvalueBit) };
 };
 
+template<typename T> struct is_diagonal
+{ enum { ret = false }; };
+
+template<typename T> struct is_diagonal<DiagonalBase<T> >
+{ enum { ret = true }; };
+
+template<typename T> struct is_diagonal<DiagonalWrapper<T> >
+{ enum { ret = true }; };
+
+template<typename T, int S> struct is_diagonal<DiagonalMatrix<T,S> >
+{ enum { ret = true }; };
+
+template<typename S1, typename S2> struct glue_shapes;
+template<> struct glue_shapes<DenseShape,TriangularShape> { typedef TriangularShape type;  };
+
+template<typename T1, typename T2>
+bool is_same_dense(const T1 &mat1, const T2 &mat2, typename enable_if<has_direct_access<T1>::ret&&has_direct_access<T2>::ret, T1>::type * = 0)
+{
+  return (mat1.data()==mat2.data()) && (mat1.innerStride()==mat2.innerStride()) && (mat1.outerStride()==mat2.outerStride());
+}
+
+template<typename T1, typename T2>
+bool is_same_dense(const T1 &, const T2 &, typename enable_if<!(has_direct_access<T1>::ret&&has_direct_access<T2>::ret), T1>::type * = 0)
+{
+  return false;
+}
+
+// Internal helper defining the cost of a scalar division for the type T.
+// The default heuristic can be specialized for each scalar type and architecture.
+template<typename T,bool Vectorized=false,typename EnableIf = void>
+struct scalar_div_cost {
+  enum { value = 8*NumTraits<T>::MulCost };
+};
+
+template<typename T,bool Vectorized>
+struct scalar_div_cost<std::complex<T>, Vectorized> {
+  enum { value = 2*scalar_div_cost<T>::value
+               + 6*NumTraits<T>::MulCost
+               + 3*NumTraits<T>::AddCost
+  };
+};
+
+
+template<bool Vectorized>
+struct scalar_div_cost<signed long,Vectorized,typename conditional<sizeof(long)==8,void,false_type>::type> { enum { value = 24 }; };
+template<bool Vectorized>
+struct scalar_div_cost<unsigned long,Vectorized,typename conditional<sizeof(long)==8,void,false_type>::type> { enum { value = 21 }; };
+
+
+#ifdef EIGEN_DEBUG_ASSIGN
+std::string demangle_traversal(int t)
+{
+  if(t==DefaultTraversal) return "DefaultTraversal";
+  if(t==LinearTraversal) return "LinearTraversal";
+  if(t==InnerVectorizedTraversal) return "InnerVectorizedTraversal";
+  if(t==LinearVectorizedTraversal) return "LinearVectorizedTraversal";
+  if(t==SliceVectorizedTraversal) return "SliceVectorizedTraversal";
+  return "?";
+}
+std::string demangle_unrolling(int t)
+{
+  if(t==NoUnrolling) return "NoUnrolling";
+  if(t==InnerUnrolling) return "InnerUnrolling";
+  if(t==CompleteUnrolling) return "CompleteUnrolling";
+  return "?";
+}
+std::string demangle_flags(int f)
+{
+  std::string res;
+  if(f&RowMajorBit)                 res += " | RowMajor";
+  if(f&PacketAccessBit)             res += " | Packet";
+  if(f&LinearAccessBit)             res += " | Linear";
+  if(f&LvalueBit)                   res += " | Lvalue";
+  if(f&DirectAccessBit)             res += " | Direct";
+  if(f&NestByRefBit)                res += " | NestByRef";
+  if(f&NoPreferredStorageOrderBit)  res += " | NoPreferredStorageOrderBit";
+  
+  return res;
+}
+#endif
+
 } // end namespace internal
 
+
+/** \class ScalarBinaryOpTraits
+  * \ingroup Core_Module
+  *
+  * \brief Determines whether the given binary operation of two numeric types is allowed and what the scalar return type is.
+  *
+  * This class permits to control the scalar return type of any binary operation performed on two different scalar types through (partial) template specializations.
+  *
+  * For instance, let \c U1, \c U2 and \c U3 be three user defined scalar types for which most operations between instances of \c U1 and \c U2 returns an \c U3.
+  * You can let %Eigen knows that by defining:
+    \code
+    template<typename BinaryOp>
+    struct ScalarBinaryOpTraits<U1,U2,BinaryOp> { typedef U3 ReturnType;  };
+    template<typename BinaryOp>
+    struct ScalarBinaryOpTraits<U2,U1,BinaryOp> { typedef U3 ReturnType;  };
+    \endcode
+  * You can then explicitly disable some particular operations to get more explicit error messages:
+    \code
+    template<>
+    struct ScalarBinaryOpTraits<U1,U2,internal::scalar_max_op<U1,U2> > {};
+    \endcode
+  * Or customize the return type for individual operation:
+    \code
+    template<>
+    struct ScalarBinaryOpTraits<U1,U2,internal::scalar_sum_op<U1,U2> > { typedef U1 ReturnType; };
+    \endcode
+  *
+  * By default, the following generic combinations are supported:
+  <table class="manual">
+  <tr><th>ScalarA</th><th>ScalarB</th><th>BinaryOp</th><th>ReturnType</th><th>Note</th></tr>
+  <tr            ><td>\c T </td><td>\c T </td><td>\c * </td><td>\c T </td><td></td></tr>
+  <tr class="alt"><td>\c NumTraits<T>::Real </td><td>\c T </td><td>\c * </td><td>\c T </td><td>Only if \c NumTraits<T>::IsComplex </td></tr>
+  <tr            ><td>\c T </td><td>\c NumTraits<T>::Real </td><td>\c * </td><td>\c T </td><td>Only if \c NumTraits<T>::IsComplex </td></tr>
+  </table>
+  *
+  * \sa CwiseBinaryOp
+  */
+template<typename ScalarA, typename ScalarB, typename BinaryOp=internal::scalar_product_op<ScalarA,ScalarB> >
+struct ScalarBinaryOpTraits
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+  // for backward compatibility, use the hints given by the (deprecated) internal::scalar_product_traits class.
+  : internal::scalar_product_traits<ScalarA,ScalarB>
+#endif // EIGEN_PARSED_BY_DOXYGEN
+{};
+
+template<typename T, typename BinaryOp>
+struct ScalarBinaryOpTraits<T,T,BinaryOp>
+{
+  typedef T ReturnType;
+};
+
+template <typename T, typename BinaryOp>
+struct ScalarBinaryOpTraits<T, typename NumTraits<typename internal::enable_if<NumTraits<T>::IsComplex,T>::type>::Real, BinaryOp>
+{
+  typedef T ReturnType;
+};
+template <typename T, typename BinaryOp>
+struct ScalarBinaryOpTraits<typename NumTraits<typename internal::enable_if<NumTraits<T>::IsComplex,T>::type>::Real, T, BinaryOp>
+{
+  typedef T ReturnType;
+};
+
+// For Matrix * Permutation
+template<typename T, typename BinaryOp>
+struct ScalarBinaryOpTraits<T,void,BinaryOp>
+{
+  typedef T ReturnType;
+};
+
+// For Permutation * Matrix
+template<typename T, typename BinaryOp>
+struct ScalarBinaryOpTraits<void,T,BinaryOp>
+{
+  typedef T ReturnType;
+};
+
+// for Permutation*Permutation
+template<typename BinaryOp>
+struct ScalarBinaryOpTraits<void,void,BinaryOp>
+{
+  typedef void ReturnType;
+};
+
+// We require Lhs and Rhs to have "compatible" scalar types.
+// It is tempting to always allow mixing different types but remember that this is often impossible in the vectorized paths.
+// So allowing mixing different types gives very unexpected errors when enabling vectorization, when the user tries to
+// add together a float matrix and a double matrix.
+#define EIGEN_CHECK_BINARY_COMPATIBILIY(BINOP,LHS,RHS) \
+  EIGEN_STATIC_ASSERT((Eigen::internal::has_ReturnType<ScalarBinaryOpTraits<LHS, RHS,BINOP> >::value), \
+    YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
+    
 } // end namespace Eigen
 
 #endif // EIGEN_XPRHELPER_H