umappp 0.1.2

data/vendor/Eigen/src/Core/NoAlias.h

@@ -0,0 +1,109 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 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_NOALIAS_H
+#define EIGEN_NOALIAS_H
+
+namespace Eigen {
+
+/** \class NoAlias
+  * \ingroup Core_Module
+  *
+  * \brief Pseudo expression providing an operator = assuming no aliasing
+  *
+  * \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.
+  * More precisely it alloas to bypass the EvalBeforeAssignBit flag of the source expression.
+  * It is the return type of MatrixBase::noalias()
+  * and most of the time this is the only way it is used.
+  *
+  * \sa MatrixBase::noalias()
+  */
+template<typename ExpressionType, template <typename> class StorageBase>
+class NoAlias
+{
+  public:
+    typedef typename ExpressionType::Scalar Scalar;
+    
+    EIGEN_DEVICE_FUNC
+    explicit NoAlias(ExpressionType& expression) : m_expression(expression) {}
+    
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE ExpressionType& operator=(const StorageBase<OtherDerived>& other)
+    {
+      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)
+    {
+      call_assignment_no_alias(m_expression, other.derived(), internal::add_assign_op<Scalar,typename OtherDerived::Scalar>());
+      return m_expression;
+    }
+    
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE ExpressionType& operator-=(const StorageBase<OtherDerived>& other)
+    {
+      call_assignment_no_alias(m_expression, other.derived(), internal::sub_assign_op<Scalar,typename OtherDerived::Scalar>());
+      return m_expression;
+    }
+
+    EIGEN_DEVICE_FUNC
+    ExpressionType& expression() const
+    {
+      return m_expression;
+    }
+
+  protected:
+    ExpressionType& m_expression;
+};
+
+/** \returns a pseudo expression of \c *this with an operator= assuming
+  * no aliasing between \c *this and the source expression.
+  *
+  * More precisely, noalias() allows to bypass the EvalBeforeAssignBit flag.
+  * Currently, even though several expressions may alias, only product
+  * expressions have this flag. Therefore, noalias() is only useful when
+  * the source expression contains a matrix product.
+  *
+  * Here are some examples where noalias is useful:
+  * \code
+  * D.noalias()  = A * B;
+  * D.noalias() += A.transpose() * B;
+  * D.noalias() -= 2 * A * B.adjoint();
+  * \endcode
+  *
+  * On the other hand the following example will lead to a \b wrong result:
+  * \code
+  * A.noalias() = A * B;
+  * \endcode
+  * because the result matrix A is also an operand of the matrix product. Therefore,
+  * there is no alternative than evaluating A * B in a temporary, that is the default
+  * behavior when you write:
+  * \code
+  * A = A * B;
+  * \endcode
+  *
+  * \sa class NoAlias
+  */
+template<typename Derived>
+NoAlias<Derived,MatrixBase> EIGEN_DEVICE_FUNC MatrixBase<Derived>::noalias()
+{
+  return NoAlias<Derived, Eigen::MatrixBase >(derived());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_NOALIAS_H

data/vendor/Eigen/src/Core/NumTraits.h

@@ -0,0 +1,335 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2006-2010 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_NUMTRAITS_H
+#define EIGEN_NUMTRAITS_H
+
+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
+{
+  EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
+  static int run() { return std::numeric_limits<T>::digits10; }
+};
+
+template<typename T>
+struct default_digits10_impl<T,false,false> // Floating point
+{
+  EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
+  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
+{
+  EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
+  static int run() { return 0; }
+};
+
+
+// default implementation of digits(), based on numeric_limits if specialized,
+// 0 for integer types, and log2(epsilon()) otherwise.
+template< typename T,
+          bool use_numeric_limits = std::numeric_limits<T>::is_specialized,
+          bool is_integer = NumTraits<T>::IsInteger>
+struct default_digits_impl
+{
+  EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
+  static int run() { return std::numeric_limits<T>::digits; }
+};
+
+template<typename T>
+struct default_digits_impl<T,false,false> // Floating point
+{
+  EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
+  static int run() {
+    using std::log;
+    using std::ceil;
+    typedef typename NumTraits<T>::Real Real;
+    return int(ceil(-log(NumTraits<Real>::epsilon())/log(static_cast<Real>(2))));
+  }
+};
+
+template<typename T>
+struct default_digits_impl<T,false,true> // Integer
+{
+  EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
+  static int run() { return 0; }
+};
+
+} // end namespace internal
+
+namespace numext {
+/** \internal bit-wise cast without changing the underlying bit representation. */
+
+// TODO: Replace by std::bit_cast (available in C++20)
+template <typename Tgt, typename Src>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Tgt bit_cast(const Src& src) {
+#if EIGEN_HAS_TYPE_TRAITS
+  // The behaviour of memcpy is not specified for non-trivially copyable types
+  EIGEN_STATIC_ASSERT(std::is_trivially_copyable<Src>::value, THIS_TYPE_IS_NOT_SUPPORTED);
+  EIGEN_STATIC_ASSERT(std::is_trivially_copyable<Tgt>::value && std::is_default_constructible<Tgt>::value,
+                      THIS_TYPE_IS_NOT_SUPPORTED);
+#endif
+
+  EIGEN_STATIC_ASSERT(sizeof(Src) == sizeof(Tgt), THIS_TYPE_IS_NOT_SUPPORTED);
+  Tgt tgt;
+  EIGEN_USING_STD(memcpy)
+  memcpy(&tgt, &src, sizeof(Tgt));
+  return tgt;
+}
+}  // namespace numext
+
+/** \class NumTraits
+  * \ingroup Core_Module
+  *
+  * \brief Holds information about the various numeric (i.e. scalar) types allowed by Eigen.
+  *
+  * \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 \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 \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
+  *     type, and to 0 otherwise.
+  * \li An enum value \a IsInteger. It is equal to \c 1 if \a T is an integer type such as \c int,
+  *     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. 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 <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 digits() function returning the number of radix digits (non-sign digits for integers, mantissa for floating-point). This is
+  *     the analogue of <a href="http://en.cppreference.com/w/cpp/types/numeric_limits/digits">std::numeric_limits<T>::digits</a>
+  *     which is used as the default implementation if specialized.
+  * \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.
+  * \li min_exponent() and max_exponent() functions returning the highest and lowest possible values, respectively,
+  *     such that the radix raised to the power exponent-1 is a normalized floating-point number.  These are equivalent to
+  *     <a href="http://en.cppreference.com/w/cpp/types/numeric_limits/min_exponent">std::numeric_limits<T>::min_exponent</a>/
+  *     <a href="http://en.cppreference.com/w/cpp/types/numeric_limits/max_exponent">std::numeric_limits<T>::max_exponent</a>.
+  * \li infinity() function returning a representation of positive infinity, if available.
+  * \li quiet_NaN function returning a non-signaling "not-a-number", if available.
+  */
+
+template<typename T> struct GenericNumTraits
+{
+  enum {
+    IsInteger = std::numeric_limits<T>::is_integer,
+    IsSigned = std::numeric_limits<T>::is_signed,
+    IsComplex = 0,
+    RequireInitialization = internal::is_arithmetic<T>::value ? 0 : 1,
+    ReadCost = 1,
+    AddCost = 1,
+    MulCost = 1
+  };
+
+  typedef T Real;
+  typedef typename internal::conditional<
+                     IsInteger,
+                     typename internal::conditional<sizeof(T)<=2, float, double>::type,
+                     T
+                   >::type NonInteger;
+  typedef T Nested;
+  typedef T Literal;
+
+  EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
+  static inline Real epsilon()
+  {
+    return numext::numeric_limits<T>::epsilon();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
+  static inline int digits10()
+  {
+    return internal::default_digits10_impl<T>::run();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
+  static inline int digits()
+  {
+    return internal::default_digits_impl<T>::run();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
+  static inline int min_exponent()
+  {
+    return numext::numeric_limits<T>::min_exponent;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
+  static inline int max_exponent()
+  {
+    return numext::numeric_limits<T>::max_exponent;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
+  static inline Real dummy_precision()
+  {
+    // make sure to override this for floating-point types
+    return Real(0);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
+  static inline T highest() {
+    return (numext::numeric_limits<T>::max)();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
+  static inline T lowest()  {
+    return IsInteger ? (numext::numeric_limits<T>::min)()
+                     : static_cast<T>(-(numext::numeric_limits<T>::max)());
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
+  static inline T infinity() {
+    return numext::numeric_limits<T>::infinity();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
+  static inline T quiet_NaN() {
+    return numext::numeric_limits<T>::quiet_NaN();
+  }
+};
+
+template<typename T> struct NumTraits : GenericNumTraits<T>
+{};
+
+template<> struct NumTraits<float>
+  : GenericNumTraits<float>
+{
+  EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
+  static inline float dummy_precision() { return 1e-5f; }
+};
+
+template<> struct NumTraits<double> : GenericNumTraits<double>
+{
+  EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
+  static inline double dummy_precision() { return 1e-12; }
+};
+
+template<> struct NumTraits<long double>
+  : GenericNumTraits<long double>
+{
+  EIGEN_CONSTEXPR
+  static inline long double dummy_precision() { return 1e-15l; }
+};
+
+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,
+    ReadCost = 2 * NumTraits<_Real>::ReadCost,
+    AddCost = 2 * NumTraits<Real>::AddCost,
+    MulCost = 4 * NumTraits<Real>::MulCost + 2 * NumTraits<Real>::AddCost
+  };
+
+  EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
+  static inline Real epsilon() { return NumTraits<Real>::epsilon(); }
+  EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
+  static inline Real dummy_precision() { return NumTraits<Real>::dummy_precision(); }
+  EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
+  static inline int digits10() { return NumTraits<Real>::digits10(); }
+};
+
+template<typename Scalar, int Rows, int Cols, int Options, int MaxRows, int MaxCols>
+struct NumTraits<Array<Scalar, Rows, Cols, Options, MaxRows, MaxCols> >
+{
+  typedef Array<Scalar, Rows, Cols, Options, MaxRows, MaxCols> ArrayType;
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  typedef Array<RealScalar, Rows, Cols, Options, MaxRows, MaxCols> Real;
+  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 ? HugeCost : ArrayType::SizeAtCompileTime * int(NumTraits<Scalar>::ReadCost),
+    AddCost  = ArrayType::SizeAtCompileTime==Dynamic ? HugeCost : ArrayType::SizeAtCompileTime * int(NumTraits<Scalar>::AddCost),
+    MulCost  = ArrayType::SizeAtCompileTime==Dynamic ? HugeCost : ArrayType::SizeAtCompileTime * int(NumTraits<Scalar>::MulCost)
+  };
+
+  EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
+  static inline RealScalar epsilon() { return NumTraits<RealScalar>::epsilon(); }
+  EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
+  static inline RealScalar dummy_precision() { return NumTraits<RealScalar>::dummy_precision(); }
+
+  EIGEN_CONSTEXPR
+  static inline int digits10() { return NumTraits<Scalar>::digits10(); }
+};
+
+template<> struct NumTraits<std::string>
+  : GenericNumTraits<std::string>
+{
+  enum {
+    RequireInitialization = 1,
+    ReadCost = HugeCost,
+    AddCost  = HugeCost,
+    MulCost  = HugeCost
+  };
+
+  EIGEN_CONSTEXPR
+  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> {};
+
+template<> struct NumTraits<bool> : GenericNumTraits<bool> {};
+
+} // end namespace Eigen
+
+#endif // EIGEN_NUMTRAITS_H

data/vendor/Eigen/src/Core/PartialReduxEvaluator.h

@@ -0,0 +1,232 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2011-2018 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_PARTIALREDUX_H
+#define EIGEN_PARTIALREDUX_H
+
+namespace Eigen { 
+
+namespace internal {
+
+
+/***************************************************************************
+*
+* This file provides evaluators for partial reductions.
+* There are two modes:
+*
+*  - scalar path: simply calls the respective function on the column or row.
+*    -> nothing special here, all the tricky part is handled by the return
+*       types of VectorwiseOp's members. They embed the functor calling the
+*       respective DenseBase's member function.
+*
+*  - vectorized path: implements a packet-wise reductions followed by
+*    some (optional) processing of the outcome, e.g., division by n for mean.
+*
+* For the vectorized path let's observe that the packet-size and outer-unrolling
+* are both decided by the assignement logic. So all we have to do is to decide
+* on the inner unrolling.
+*
+* For the unrolling, we can reuse "internal::redux_vec_unroller" from Redux.h,
+* but be need to be careful to specify correct increment.
+*
+***************************************************************************/
+
+
+/* logic deciding a strategy for unrolling of vectorized paths */
+template<typename Func, typename Evaluator>
+struct packetwise_redux_traits
+{
+  enum {
+    OuterSize = int(Evaluator::IsRowMajor) ? Evaluator::RowsAtCompileTime : Evaluator::ColsAtCompileTime,
+    Cost = OuterSize == Dynamic ? HugeCost
+         : OuterSize * Evaluator::CoeffReadCost + (OuterSize-1) * functor_traits<Func>::Cost,
+    Unrolling = Cost <= EIGEN_UNROLLING_LIMIT ? CompleteUnrolling : NoUnrolling
+  };
+
+};
+
+/* Value to be returned when size==0 , by default let's return 0 */
+template<typename PacketType,typename Func>
+EIGEN_DEVICE_FUNC
+PacketType packetwise_redux_empty_value(const Func& ) { return pset1<PacketType>(0); }
+
+/* For products the default is 1 */
+template<typename PacketType,typename Scalar>
+EIGEN_DEVICE_FUNC
+PacketType packetwise_redux_empty_value(const scalar_product_op<Scalar,Scalar>& ) { return pset1<PacketType>(1); }
+
+/* Perform the actual reduction */
+template<typename Func, typename Evaluator,
+         int Unrolling = packetwise_redux_traits<Func, Evaluator>::Unrolling
+>
+struct packetwise_redux_impl;
+
+/* Perform the actual reduction with unrolling */
+template<typename Func, typename Evaluator>
+struct packetwise_redux_impl<Func, Evaluator, CompleteUnrolling>
+{
+  typedef redux_novec_unroller<Func,Evaluator, 0, Evaluator::SizeAtCompileTime> Base;
+  typedef typename Evaluator::Scalar Scalar;
+
+  template<typename PacketType>
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE
+  PacketType run(const Evaluator &eval, const Func& func, Index /*size*/)
+  {
+    return redux_vec_unroller<Func, Evaluator, 0, packetwise_redux_traits<Func, Evaluator>::OuterSize>::template run<PacketType>(eval,func);
+  }
+};
+
+/* Add a specialization of redux_vec_unroller for size==0 at compiletime.
+ * This specialization is not required for general reductions, which is
+ * why it is defined here.
+ */
+template<typename Func, typename Evaluator, int Start>
+struct redux_vec_unroller<Func, Evaluator, Start, 0>
+{
+  template<typename PacketType>
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE PacketType run(const Evaluator &, const Func& f)
+  {
+    return packetwise_redux_empty_value<PacketType>(f);
+  }
+};
+
+/* Perform the actual reduction for dynamic sizes */
+template<typename Func, typename Evaluator>
+struct packetwise_redux_impl<Func, Evaluator, NoUnrolling>
+{
+  typedef typename Evaluator::Scalar Scalar;
+  typedef typename redux_traits<Func, Evaluator>::PacketType PacketScalar;
+
+  template<typename PacketType>
+  EIGEN_DEVICE_FUNC
+  static PacketType run(const Evaluator &eval, const Func& func, Index size)
+  {
+    if(size==0)
+      return packetwise_redux_empty_value<PacketType>(func);
+    
+    const Index size4 = (size-1)&(~3);
+    PacketType p = eval.template packetByOuterInner<Unaligned,PacketType>(0,0);
+    Index i = 1;
+    // This loop is optimized for instruction pipelining:
+    // - each iteration generates two independent instructions
+    // - thanks to branch prediction and out-of-order execution we have independent instructions across loops
+    for(; i<size4; i+=4)
+      p = func.packetOp(p,
+            func.packetOp(
+              func.packetOp(eval.template packetByOuterInner<Unaligned,PacketType>(i+0,0),eval.template packetByOuterInner<Unaligned,PacketType>(i+1,0)),
+              func.packetOp(eval.template packetByOuterInner<Unaligned,PacketType>(i+2,0),eval.template packetByOuterInner<Unaligned,PacketType>(i+3,0))));
+    for(; i<size; ++i)
+      p = func.packetOp(p, eval.template packetByOuterInner<Unaligned,PacketType>(i,0));
+    return p;
+  }
+};
+
+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::add_const_on_value_type<ArgTypeNested>::type ConstArgTypeNested;
+  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<int(TraversalSize)> CostOpType;
+  enum {
+    CoeffReadCost = TraversalSize==Dynamic ? HugeCost
+                  : TraversalSize==0 ? 1
+                  : int(TraversalSize) * int(evaluator<ArgType>::CoeffReadCost) + int(CostOpType::value),
+    
+    _ArgFlags = evaluator<ArgType>::Flags,
+
+    _Vectorizable =  bool(int(_ArgFlags)&PacketAccessBit)
+                  && bool(MemberOp::Vectorizable)
+                  && (Direction==int(Vertical) ? bool(_ArgFlags&RowMajorBit) : (_ArgFlags&RowMajorBit)==0)
+                  && (TraversalSize!=0),
+                  
+    Flags = (traits<XprType>::Flags&RowMajorBit)
+          | (evaluator<ArgType>::Flags&(HereditaryBits&(~RowMajorBit)))
+          | (_Vectorizable ? PacketAccessBit : 0)
+          | 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 : (TraversalSize==0 ? 1 : 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
+  {
+    return coeff(Direction==Vertical ? j : i);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  const Scalar coeff(Index index) const
+  {
+    return m_functor(m_arg.template subVector<DirectionType(Direction)>(index));
+  }
+
+  template<int LoadMode,typename PacketType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  PacketType packet(Index i, Index j) const
+  {
+    return packet<LoadMode,PacketType>(Direction==Vertical ? j : i);
+  }
+  
+  template<int LoadMode,typename PacketType>
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
+  PacketType packet(Index idx) const
+  {
+    enum { PacketSize = internal::unpacket_traits<PacketType>::size };
+    typedef Block<const ArgTypeNestedCleaned,
+                  Direction==Vertical ? int(ArgType::RowsAtCompileTime) : int(PacketSize),
+                  Direction==Vertical ? int(PacketSize) : int(ArgType::ColsAtCompileTime),
+                  true /* InnerPanel */> PanelType;
+    
+    PanelType panel(m_arg,
+                    Direction==Vertical ? 0 : idx,
+                    Direction==Vertical ? idx : 0,
+                    Direction==Vertical ? m_arg.rows() : Index(PacketSize),
+                    Direction==Vertical ? Index(PacketSize) : m_arg.cols());
+
+    // FIXME
+    // See bug 1612, currently if PacketSize==1 (i.e. complex<double> with 128bits registers) then the storage-order of panel get reversed
+    // and methods like packetByOuterInner do not make sense anymore in this context.
+    // So let's just by pass "vectorization" in this case:
+    if(PacketSize==1)
+      return internal::pset1<PacketType>(coeff(idx));
+    
+    typedef typename internal::redux_evaluator<PanelType> PanelEvaluator;
+    PanelEvaluator panel_eval(panel);
+    typedef typename MemberOp::BinaryOp BinaryOp;
+    PacketType p = internal::packetwise_redux_impl<BinaryOp,PanelEvaluator>::template run<PacketType>(panel_eval,m_functor.binaryFunc(),m_arg.outerSize());
+    return p;
+  }
+
+protected:
+  ConstArgTypeNested m_arg;
+  const MemberOp m_functor;
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_PARTIALREDUX_H