ruby-eigen 0.0.9 → 0.0.10.pre1

data/ext/eigen/eigen3/Eigen/src/Core/products/TriangularSolverMatrix_MKL.h

@@ -0,0 +1,155 @@
+/*
+ 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 Intel(R) MKL
+ *   Triangular matrix * matrix product functionality based on ?TRMM.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_TRIANGULAR_SOLVER_MATRIX_MKL_H
+#define EIGEN_TRIANGULAR_SOLVER_MATRIX_MKL_H
+
+namespace Eigen {
+
+namespace internal {
+
+// implements LeftSide op(triangular)^-1 * general
+#define EIGEN_MKL_TRSM_L(EIGTYPE, MKLTYPE, MKLPREFIX) \
+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*/) \
+  { \
+   MKL_INT m = size, n = otherSize, lda, ldb; \
+   char side = 'L', uplo, diag='N', transa; \
+   /* Set alpha_ */ \
+   MKLTYPE alpha; \
+   EIGTYPE myone(1); \
+   assign_scalar_eig2mkl(alpha, myone); \
+   ldb = 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 = a_tmp.outerStride(); \
+   } else { \
+     a = _tri; \
+     lda = triStride; \
+   } \
+   if (IsUnitDiag) diag='U'; \
+/* call ?trsm*/ \
+   MKLPREFIX##trsm(&side, &uplo, &transa, &diag, &m, &n, &alpha, (const MKLTYPE*)a, &lda, (MKLTYPE*)_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)
+
+
+// implements RightSide general * op(triangular)^-1
+#define EIGEN_MKL_TRSM_R(EIGTYPE, MKLTYPE, MKLPREFIX) \
+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*/) \
+  { \
+   MKL_INT m = otherSize, n = size, lda, ldb; \
+   char side = 'R', uplo, diag='N', transa; \
+   /* Set alpha_ */ \
+   MKLTYPE alpha; \
+   EIGTYPE myone(1); \
+   assign_scalar_eig2mkl(alpha, myone); \
+   ldb = 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 = a_tmp.outerStride(); \
+   } else { \
+     a = _tri; \
+     lda = triStride; \
+   } \
+   if (IsUnitDiag) diag='U'; \
+/* call ?trsm*/ \
+   MKLPREFIX##trsm(&side, &uplo, &transa, &diag, &m, &n, &alpha, (const MKLTYPE*)a, &lda, (MKLTYPE*)_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)
+
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TRIANGULAR_SOLVER_MATRIX_MKL_H

data/ext/eigen/eigen3/Eigen/src/Core/products/TriangularSolverVector.h

@@ -0,0 +1,139 @@
+// 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));
+    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,RowMajor,Conjugate,RhsScalar,false>::run(
+          actualPanelWidth, r,
+          &lhs.coeffRef(startRow,startCol), lhsStride,
+          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));
+    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,ColMajor,Conjugate,RhsScalar,false>::run(
+            r, actualPanelWidth,
+            &lhs.coeffRef(endBlock,startBlock), lhsStride,
+            rhs+startBlock, 1,
+            rhs+endBlock, 1, RhsScalar(-1));
+      }
+    }
+  }
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TRIANGULAR_SOLVER_VECTOR_H

data/ext/eigen/eigen3/Eigen/src/Core/util/BlasUtil.h

@@ -0,0 +1,264 @@
+// 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, 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>
+struct gemm_pack_rhs;
+
+template<typename Scalar, typename Index, 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, int LhsStorageOrder, bool ConjugateLhs, typename RhsScalar, 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) { return numext::conj(x); }
+  template<typename T>
+  inline T pconj(const T& x) { return internal::pconj(x); }
+};
+
+template<> struct conj_if<false> {
+  template<typename T>
+  inline const T& operator()(const T& x) { return x; }
+  template<typename T>
+  inline const T& pconj(const T& x) { return x; }
+};
+
+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); }
+};
+
+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 {
+  static EIGEN_STRONG_INLINE To run(const From& x) { return x; }
+};
+
+template<typename Scalar> struct get_factor<Scalar,typename NumTraits<Scalar>::Real> {
+  static EIGEN_STRONG_INLINE typename NumTraits<Scalar>::Real run(const Scalar& x) { return numext::real(x); }
+};
+
+// 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
+{
+  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]; }
+  protected:
+    Scalar* EIGEN_RESTRICT m_data;
+    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:
+    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;
+};
+
+
+/* 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>
+struct blas_traits<CwiseUnaryOp<scalar_multiple_op<Scalar>, NestedXpr> >
+ : blas_traits<NestedXpr>
+{
+  typedef blas_traits<NestedXpr> Base;
+  typedef CwiseUnaryOp<scalar_multiple_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 x.functor().m_other * Base::extractScalarFactor(x.nestedExpression()); }
+};
+
+// 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 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