tomoto 0.1.0

data/vendor/eigen/Eigen/src/Core/products/TriangularSolverMatrix_BLAS.h

@@ -0,0 +1,163 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ 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 OWNER 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.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to BLAS F77
+ *   Triangular matrix * matrix product functionality based on ?TRMM.
+ ********************************************************************************
+*/
+
+#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_BLAS_TRSM_L(EIGTYPE, BLASTYPE, BLASFUNC) \
+template <typename Index, int Mode, bool Conjugate, int TriStorageOrder> \
+struct triangular_solve_matrix<EIGTYPE,Index,OnTheLeft,Mode,Conjugate,TriStorageOrder,ColMajor> \
+{ \
+  enum { \
+    IsLower = (Mode&Lower) == Lower, \
+    IsUnitDiag  = (Mode&UnitDiag) ? 1 : 0, \
+    IsZeroDiag  = (Mode&ZeroDiag) ? 1 : 0, \
+    conjA = ((TriStorageOrder==ColMajor) && Conjugate) ? 1 : 0 \
+  }; \
+  static void run( \
+      Index size, Index otherSize, \
+      const EIGTYPE* _tri, Index triStride, \
+      EIGTYPE* _other, Index otherStride, level3_blocking<EIGTYPE,EIGTYPE>& /*blocking*/) \
+  { \
+   BlasIndex m = convert_index<BlasIndex>(size), n = convert_index<BlasIndex>(otherSize), lda, ldb; \
+   char side = 'L', uplo, diag='N', transa; \
+   /* Set alpha_ */ \
+   EIGTYPE alpha(1); \
+   ldb = convert_index<BlasIndex>(otherStride);\
+\
+   const EIGTYPE *a; \
+/* Set trans */ \
+   transa = (TriStorageOrder==RowMajor) ? ((Conjugate) ? 'C' : 'T') : 'N'; \
+/* Set uplo */ \
+   uplo = IsLower ? 'L' : 'U'; \
+   if (TriStorageOrder==RowMajor) uplo = (uplo == 'L') ? 'U' : 'L'; \
+/* Set a, lda */ \
+   typedef Matrix<EIGTYPE, Dynamic, Dynamic, TriStorageOrder> MatrixTri; \
+   Map<const MatrixTri, 0, OuterStride<> > tri(_tri,size,size,OuterStride<>(triStride)); \
+   MatrixTri a_tmp; \
+\
+   if (conjA) { \
+     a_tmp = tri.conjugate(); \
+     a = a_tmp.data(); \
+     lda = convert_index<BlasIndex>(a_tmp.outerStride()); \
+   } else { \
+     a = _tri; \
+     lda = convert_index<BlasIndex>(triStride); \
+   } \
+   if (IsUnitDiag) diag='U'; \
+/* call ?trsm*/ \
+   BLASFUNC(&side, &uplo, &transa, &diag, &m, &n, (const BLASTYPE*)&numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (BLASTYPE*)_other, &ldb); \
+ } \
+};
+
+#ifdef EIGEN_USE_MKL
+EIGEN_BLAS_TRSM_L(double,   double, dtrsm)
+EIGEN_BLAS_TRSM_L(dcomplex, MKL_Complex16, ztrsm)
+EIGEN_BLAS_TRSM_L(float,    float,  strsm)
+EIGEN_BLAS_TRSM_L(scomplex, MKL_Complex8, ctrsm)
+#else
+EIGEN_BLAS_TRSM_L(double,   double, dtrsm_)
+EIGEN_BLAS_TRSM_L(dcomplex, double, ztrsm_)
+EIGEN_BLAS_TRSM_L(float,    float,  strsm_)
+EIGEN_BLAS_TRSM_L(scomplex, float,  ctrsm_)
+#endif
+
+// implements RightSide general * op(triangular)^-1
+#define EIGEN_BLAS_TRSM_R(EIGTYPE, BLASTYPE, BLASFUNC) \
+template <typename Index, int Mode, bool Conjugate, int TriStorageOrder> \
+struct triangular_solve_matrix<EIGTYPE,Index,OnTheRight,Mode,Conjugate,TriStorageOrder,ColMajor> \
+{ \
+  enum { \
+    IsLower = (Mode&Lower) == Lower, \
+    IsUnitDiag  = (Mode&UnitDiag) ? 1 : 0, \
+    IsZeroDiag  = (Mode&ZeroDiag) ? 1 : 0, \
+    conjA = ((TriStorageOrder==ColMajor) && Conjugate) ? 1 : 0 \
+  }; \
+  static void run( \
+      Index size, Index otherSize, \
+      const EIGTYPE* _tri, Index triStride, \
+      EIGTYPE* _other, Index otherStride, level3_blocking<EIGTYPE,EIGTYPE>& /*blocking*/) \
+  { \
+   BlasIndex m = convert_index<BlasIndex>(otherSize), n = convert_index<BlasIndex>(size), lda, ldb; \
+   char side = 'R', uplo, diag='N', transa; \
+   /* Set alpha_ */ \
+   EIGTYPE alpha(1); \
+   ldb = convert_index<BlasIndex>(otherStride);\
+\
+   const EIGTYPE *a; \
+/* Set trans */ \
+   transa = (TriStorageOrder==RowMajor) ? ((Conjugate) ? 'C' : 'T') : 'N'; \
+/* Set uplo */ \
+   uplo = IsLower ? 'L' : 'U'; \
+   if (TriStorageOrder==RowMajor) uplo = (uplo == 'L') ? 'U' : 'L'; \
+/* Set a, lda */ \
+   typedef Matrix<EIGTYPE, Dynamic, Dynamic, TriStorageOrder> MatrixTri; \
+   Map<const MatrixTri, 0, OuterStride<> > tri(_tri,size,size,OuterStride<>(triStride)); \
+   MatrixTri a_tmp; \
+\
+   if (conjA) { \
+     a_tmp = tri.conjugate(); \
+     a = a_tmp.data(); \
+     lda = convert_index<BlasIndex>(a_tmp.outerStride()); \
+   } else { \
+     a = _tri; \
+     lda = convert_index<BlasIndex>(triStride); \
+   } \
+   if (IsUnitDiag) diag='U'; \
+/* call ?trsm*/ \
+   BLASFUNC(&side, &uplo, &transa, &diag, &m, &n, (const BLASTYPE*)&numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (BLASTYPE*)_other, &ldb); \
+   /*std::cout << "TRMS_L specialization!\n";*/ \
+ } \
+};
+
+#ifdef EIGEN_USE_MKL
+EIGEN_BLAS_TRSM_R(double,   double, dtrsm)
+EIGEN_BLAS_TRSM_R(dcomplex, MKL_Complex16, ztrsm)
+EIGEN_BLAS_TRSM_R(float,    float,  strsm)
+EIGEN_BLAS_TRSM_R(scomplex, MKL_Complex8,  ctrsm)
+#else
+EIGEN_BLAS_TRSM_R(double,   double, dtrsm_)
+EIGEN_BLAS_TRSM_R(dcomplex, double, ztrsm_)
+EIGEN_BLAS_TRSM_R(float,    float,  strsm_)
+EIGEN_BLAS_TRSM_R(scomplex, float,  ctrsm_)
+#endif
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TRIANGULAR_SOLVER_MATRIX_BLAS_H

data/vendor/eigen/Eigen/src/Core/products/TriangularSolverVector.h

@@ -0,0 +1,145 @@
+// 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_TRIANGULAR_SOLVER_VECTOR_H
+#define EIGEN_TRIANGULAR_SOLVER_VECTOR_H
+
+namespace Eigen {
+
+namespace internal {
+
+template<typename LhsScalar, typename RhsScalar, typename Index, int Mode, bool Conjugate, int StorageOrder>
+struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheRight, Mode, Conjugate, StorageOrder>
+{
+  static void run(Index size, const LhsScalar* _lhs, Index lhsStride, RhsScalar* rhs)
+  {
+    triangular_solve_vector<LhsScalar,RhsScalar,Index,OnTheLeft,
+        ((Mode&Upper)==Upper ? Lower : Upper) | (Mode&UnitDiag),
+        Conjugate,StorageOrder==RowMajor?ColMajor:RowMajor
+      >::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>
+{
+  enum {
+    IsLower = ((Mode&Lower)==Lower)
+  };
+  static void run(Index size, const LhsScalar* _lhs, Index lhsStride, RhsScalar* rhs)
+  {
+    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>,
+                          const LhsMap&>
+                        ::type cjLhs(lhs);
+    static const Index PanelWidth = EIGEN_TUNE_TRIANGULAR_PANEL_WIDTH;
+    for(Index pi=IsLower ? 0 : size;
+        IsLower ? pi<size : pi>0;
+        IsLower ? pi+=PanelWidth : pi-=PanelWidth)
+    {
+      Index actualPanelWidth = (std::min)(IsLower ? size - pi : pi, PanelWidth);
+
+      Index r = IsLower ? pi : size - pi; // remaining size
+      if (r > 0)
+      {
+        // let's directly call the low level product function because:
+        // 1 - it is faster to compile
+        // 2 - it is slighlty faster at runtime
+        Index startRow = IsLower ? pi : pi-actualPanelWidth;
+        Index startCol = IsLower ? 0 : pi;
+
+        general_matrix_vector_product<Index,LhsScalar,LhsMapper,RowMajor,Conjugate,RhsScalar,RhsMapper,false>::run(
+          actualPanelWidth, r,
+          LhsMapper(&lhs.coeffRef(startRow,startCol), lhsStride),
+          RhsMapper(rhs + startCol, 1),
+          rhs + startRow, 1,
+          RhsScalar(-1));
+      }
+
+      for(Index k=0; k<actualPanelWidth; ++k)
+      {
+        Index i = IsLower ? pi+k : pi-k-1;
+        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);
+      }
+    }
+  }
+};
+
+// forward and backward substitution, column-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, ColMajor>
+{
+  enum {
+    IsLower = ((Mode&Lower)==Lower)
+  };
+  static void run(Index size, const LhsScalar* _lhs, Index lhsStride, RhsScalar* rhs)
+  {
+    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&
+                                  >::type cjLhs(lhs);
+    static const Index PanelWidth = EIGEN_TUNE_TRIANGULAR_PANEL_WIDTH;
+
+    for(Index pi=IsLower ? 0 : size;
+        IsLower ? pi<size : pi>0;
+        IsLower ? pi+=PanelWidth : pi-=PanelWidth)
+    {
+      Index actualPanelWidth = (std::min)(IsLower ? size - pi : pi, PanelWidth);
+      Index startBlock = IsLower ? pi : pi-actualPanelWidth;
+      Index endBlock = IsLower ? pi + actualPanelWidth : 0;
+
+      for(Index k=0; k<actualPanelWidth; ++k)
+      {
+        Index i = IsLower ? pi+k : pi-k-1;
+        if(!(Mode & UnitDiag))
+          rhs[i] /= cjLhs.coeff(i,i);
+
+        Index r = actualPanelWidth - k - 1; // remaining size
+        Index s = IsLower ? i+1 : i-r;
+        if (r>0)
+          Map<Matrix<RhsScalar,Dynamic,1> >(rhs+s,r) -= rhs[i] * cjLhs.col(i).segment(s,r);
+      }
+      Index r = IsLower ? size - endBlock : startBlock; // remaining size
+      if (r > 0)
+      {
+        // 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,LhsMapper,ColMajor,Conjugate,RhsScalar,RhsMapper,false>::run(
+            r, actualPanelWidth,
+            LhsMapper(&lhs.coeffRef(endBlock,startBlock), lhsStride),
+            RhsMapper(rhs+startBlock, 1),
+            rhs+endBlock, 1, RhsScalar(-1));
+      }
+    }
+  }
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TRIANGULAR_SOLVER_VECTOR_H

data/vendor/eigen/Eigen/src/Core/util/BlasUtil.h

@@ -0,0 +1,398 @@
+// 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_BLASUTIL_H
+#define EIGEN_BLASUTIL_H
+
+// This file contains many lightweight helper classes used to
+// implement and control fast level 2 and level 3 BLAS-like routines.
+
+namespace Eigen {
+
+namespace internal {
+
+// forward declarations
+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, typename DataMapper, int nr, int StorageOrder, bool Conjugate = false, bool PanelMode=false>
+struct gemm_pack_rhs;
+
+template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, int StorageOrder, bool Conjugate = false, bool PanelMode = false>
+struct gemm_pack_lhs;
+
+template<
+  typename Index,
+  typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs,
+  typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs,
+  int ResStorageOrder>
+struct general_matrix_matrix_product;
+
+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;
+
+
+template<bool Conjugate> struct conj_if;
+
+template<> struct conj_if<true> {
+  template<typename T>
+  inline T operator()(const T& x) const { return numext::conj(x); }
+  template<typename T>
+  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) const { return x; }
+  template<typename T>
+  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_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>
+{
+  typedef std::complex<RealScalar> Scalar;
+  EIGEN_STRONG_INLINE Scalar pmadd(const Scalar& x, const Scalar& y, const Scalar& c) const
+  { return c + pmul(x,y); }
+
+  EIGEN_STRONG_INLINE Scalar pmul(const Scalar& x, const Scalar& y) const
+  { return Scalar(numext::real(x)*numext::real(y) + numext::imag(x)*numext::imag(y), numext::imag(x)*numext::real(y) - numext::real(x)*numext::imag(y)); }
+};
+
+template<typename RealScalar> struct conj_helper<std::complex<RealScalar>, std::complex<RealScalar>, true,false>
+{
+  typedef std::complex<RealScalar> Scalar;
+  EIGEN_STRONG_INLINE Scalar pmadd(const Scalar& x, const Scalar& y, const Scalar& c) const
+  { return c + pmul(x,y); }
+
+  EIGEN_STRONG_INLINE Scalar pmul(const Scalar& x, const Scalar& y) const
+  { return Scalar(numext::real(x)*numext::real(y) + numext::imag(x)*numext::imag(y), numext::real(x)*numext::imag(y) - numext::imag(x)*numext::real(y)); }
+};
+
+template<typename RealScalar> struct conj_helper<std::complex<RealScalar>, std::complex<RealScalar>, true,true>
+{
+  typedef std::complex<RealScalar> Scalar;
+  EIGEN_STRONG_INLINE Scalar pmadd(const Scalar& x, const Scalar& y, const Scalar& c) const
+  { return c + pmul(x,y); }
+
+  EIGEN_STRONG_INLINE Scalar pmul(const Scalar& x, const Scalar& y) const
+  { return Scalar(numext::real(x)*numext::real(y) - numext::imag(x)*numext::imag(y), - numext::real(x)*numext::imag(y) - numext::imag(x)*numext::real(y)); }
+};
+
+template<typename RealScalar,bool Conj> struct conj_helper<std::complex<RealScalar>, RealScalar, Conj,false>
+{
+  typedef std::complex<RealScalar> Scalar;
+  EIGEN_STRONG_INLINE Scalar pmadd(const Scalar& x, const RealScalar& y, const Scalar& c) const
+  { return padd(c, pmul(x,y)); }
+  EIGEN_STRONG_INLINE Scalar pmul(const Scalar& x, const RealScalar& y) const
+  { return conj_if<Conj>()(x)*y; }
+};
+
+template<typename RealScalar,bool Conj> struct conj_helper<RealScalar, std::complex<RealScalar>, false,Conj>
+{
+  typedef std::complex<RealScalar> Scalar;
+  EIGEN_STRONG_INLINE Scalar pmadd(const RealScalar& x, const Scalar& y, const Scalar& c) const
+  { return padd(c, pmul(x,y)); }
+  EIGEN_STRONG_INLINE Scalar pmul(const RealScalar& x, const Scalar& y) const
+  { return x*conj_if<Conj>()(y); }
+};
+
+template<typename From,typename To> struct get_factor {
+  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.
+template<typename Scalar, typename Index, int StorageOrder, int AlignmentType = Unaligned>
+class blas_data_mapper {
+  public:
+  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));
+  }
+
+  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;
+  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 : public blas_data_mapper<const Scalar, Index, StorageOrder> {
+  public:
+  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);
+  }
+};
+
+
+/* Helper class to analyze the factors of a Product expression.
+ * In particular it allows to pop out operator-, scalar multiples,
+ * and conjugate */
+template<typename XprType> struct blas_traits
+{
+  typedef typename traits<XprType>::Scalar Scalar;
+  typedef const XprType& ExtractType;
+  typedef XprType _ExtractType;
+  enum {
+    IsComplex = NumTraits<Scalar>::IsComplex,
+    IsTransposed = false,
+    NeedToConjugate = false,
+    HasUsableDirectAccess = (    (int(XprType::Flags)&DirectAccessBit)
+                              && (   bool(XprType::IsVectorAtCompileTime)
+                                  || int(inner_stride_at_compile_time<XprType>::ret) == 1)
+                             ) ?  1 : 0
+  };
+  typedef typename conditional<bool(HasUsableDirectAccess),
+    ExtractType,
+    typename _ExtractType::PlainObject
+    >::type DirectLinearAccessType;
+  static inline ExtractType extract(const XprType& x) { return x; }
+  static inline const Scalar extractScalarFactor(const XprType&) { return Scalar(1); }
+};
+
+// pop conjugate
+template<typename Scalar, typename NestedXpr>
+struct blas_traits<CwiseUnaryOp<scalar_conjugate_op<Scalar>, NestedXpr> >
+ : blas_traits<NestedXpr>
+{
+  typedef blas_traits<NestedXpr> Base;
+  typedef CwiseUnaryOp<scalar_conjugate_op<Scalar>, NestedXpr> XprType;
+  typedef typename Base::ExtractType ExtractType;
+
+  enum {
+    IsComplex = NumTraits<Scalar>::IsComplex,
+    NeedToConjugate = Base::NeedToConjugate ? 0 : IsComplex
+  };
+  static inline ExtractType extract(const XprType& x) { return Base::extract(x.nestedExpression()); }
+  static inline Scalar extractScalarFactor(const XprType& x) { return conj(Base::extractScalarFactor(x.nestedExpression())); }
+};
+
+// pop scalar multiple
+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 blas_traits<NestedXpr> Base;
+  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.rhs()); }
+  static inline Scalar extractScalarFactor(const XprType& x)
+  { 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 NestedXpr>
+struct blas_traits<CwiseUnaryOp<scalar_opposite_op<Scalar>, NestedXpr> >
+ : blas_traits<NestedXpr>
+{
+  typedef blas_traits<NestedXpr> Base;
+  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)
+  { return - Base::extractScalarFactor(x.nestedExpression()); }
+};
+
+// pop/push transpose
+template<typename NestedXpr>
+struct blas_traits<Transpose<NestedXpr> >
+ : blas_traits<NestedXpr>
+{
+  typedef typename NestedXpr::Scalar Scalar;
+  typedef blas_traits<NestedXpr> Base;
+  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),
+    ExtractType,
+    typename ExtractType::PlainObject
+    >::type DirectLinearAccessType;
+  enum {
+    IsTransposed = Base::IsTransposed ? 0 : 1
+  };
+  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()); }
+};
+
+template<typename T>
+struct blas_traits<const T>
+     : blas_traits<T>
+{};
+
+template<typename T, bool HasUsableDirectAccess=blas_traits<T>::HasUsableDirectAccess>
+struct extract_data_selector {
+  static const typename T::Scalar* run(const T& m)
+  {
+    return blas_traits<T>::extract(m).data();
+  }
+};
+
+template<typename T>
+struct extract_data_selector<T,false> {
+  static typename T::Scalar* run(const T&) { return 0; }
+};
+
+template<typename T> const typename T::Scalar* extract_data(const T& m)
+{
+  return extract_data_selector<T>::run(m);
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_BLASUTIL_H