@smake/eigen 1.0.2 → 1.1.1

package/eigen/Eigen/src/LU/arch/Inverse_SSE.h

@@ -1,338 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2001 Intel Corporation
-// Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr>
-// Copyright (C) 2009 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/.
-
-// The SSE code for the 4x4 float and double matrix inverse in this file
-// comes from the following Intel's library:
-// http://software.intel.com/en-us/articles/optimized-matrix-library-for-use-with-the-intel-pentiumr-4-processors-sse2-instructions/
-//
-// Here is the respective copyright and license statement:
-//
-//   Copyright (c) 2001 Intel Corporation.
-//
-// Permition is granted to use, copy, distribute and prepare derivative works
-// of this library for any purpose and without fee, provided, that the above
-// copyright notice and this statement appear in all copies.
-// Intel makes no representations about the suitability of this software for
-// any purpose, and specifically disclaims all warranties.
-// See LEGAL.TXT for all the legal information.
-
-#ifndef EIGEN_INVERSE_SSE_H
-#define EIGEN_INVERSE_SSE_H
-
-namespace Eigen { 
-
-namespace internal {
-
-template<typename MatrixType, typename ResultType>
-struct compute_inverse_size4<Architecture::SSE, float, MatrixType, ResultType>
-{
-  enum {
-    MatrixAlignment     = traits<MatrixType>::Alignment,
-    ResultAlignment     = traits<ResultType>::Alignment,
-    StorageOrdersMatch  = (MatrixType::Flags&RowMajorBit) == (ResultType::Flags&RowMajorBit)
-  };
-  typedef typename conditional<(MatrixType::Flags&LinearAccessBit),MatrixType const &,typename MatrixType::PlainObject>::type ActualMatrixType;
-  
-  static void run(const MatrixType& mat, ResultType& result)
-  {
-    ActualMatrixType matrix(mat);
-    const Packet4f p4f_sign_PNNP = _mm_castsi128_ps(_mm_set_epi32(0x00000000, 0x80000000, 0x80000000, 0x00000000));
-
-    // Load the full matrix into registers
-    __m128 _L1 = matrix.template packet<MatrixAlignment>( 0);
-    __m128 _L2 = matrix.template packet<MatrixAlignment>( 4);
-    __m128 _L3 = matrix.template packet<MatrixAlignment>( 8);
-    __m128 _L4 = matrix.template packet<MatrixAlignment>(12);
-
-    // The inverse is calculated using "Divide and Conquer" technique. The
-    // original matrix is divide into four 2x2 sub-matrices. Since each
-    // register holds four matrix element, the smaller matrices are
-    // represented as a registers. Hence we get a better locality of the
-    // calculations.
-
-    __m128 A, B, C, D; // the four sub-matrices
-    if(!StorageOrdersMatch)
-    {
-      A = _mm_unpacklo_ps(_L1, _L2);
-      B = _mm_unpacklo_ps(_L3, _L4);
-      C = _mm_unpackhi_ps(_L1, _L2);
-      D = _mm_unpackhi_ps(_L3, _L4);
-    }
-    else
-    {
-      A = _mm_movelh_ps(_L1, _L2);
-      B = _mm_movehl_ps(_L2, _L1);
-      C = _mm_movelh_ps(_L3, _L4);
-      D = _mm_movehl_ps(_L4, _L3);
-    }
-
-    __m128 iA, iB, iC, iD,                 // partial inverse of the sub-matrices
-            DC, AB;
-    __m128 dA, dB, dC, dD;                 // determinant of the sub-matrices
-    __m128 det, d, d1, d2;
-    __m128 rd;                             // reciprocal of the determinant
-
-    //  AB = A# * B
-    AB = _mm_mul_ps(_mm_shuffle_ps(A,A,0x0F), B);
-    AB = _mm_sub_ps(AB,_mm_mul_ps(_mm_shuffle_ps(A,A,0xA5), _mm_shuffle_ps(B,B,0x4E)));
-    //  DC = D# * C
-    DC = _mm_mul_ps(_mm_shuffle_ps(D,D,0x0F), C);
-    DC = _mm_sub_ps(DC,_mm_mul_ps(_mm_shuffle_ps(D,D,0xA5), _mm_shuffle_ps(C,C,0x4E)));
-
-    //  dA = |A|
-    dA = _mm_mul_ps(_mm_shuffle_ps(A, A, 0x5F),A);
-    dA = _mm_sub_ss(dA, _mm_movehl_ps(dA,dA));
-    //  dB = |B|
-    dB = _mm_mul_ps(_mm_shuffle_ps(B, B, 0x5F),B);
-    dB = _mm_sub_ss(dB, _mm_movehl_ps(dB,dB));
-
-    //  dC = |C|
-    dC = _mm_mul_ps(_mm_shuffle_ps(C, C, 0x5F),C);
-    dC = _mm_sub_ss(dC, _mm_movehl_ps(dC,dC));
-    //  dD = |D|
-    dD = _mm_mul_ps(_mm_shuffle_ps(D, D, 0x5F),D);
-    dD = _mm_sub_ss(dD, _mm_movehl_ps(dD,dD));
-
-    //  d = trace(AB*DC) = trace(A#*B*D#*C)
-    d = _mm_mul_ps(_mm_shuffle_ps(DC,DC,0xD8),AB);
-
-    //  iD = C*A#*B
-    iD = _mm_mul_ps(_mm_shuffle_ps(C,C,0xA0), _mm_movelh_ps(AB,AB));
-    iD = _mm_add_ps(iD,_mm_mul_ps(_mm_shuffle_ps(C,C,0xF5), _mm_movehl_ps(AB,AB)));
-    //  iA = B*D#*C
-    iA = _mm_mul_ps(_mm_shuffle_ps(B,B,0xA0), _mm_movelh_ps(DC,DC));
-    iA = _mm_add_ps(iA,_mm_mul_ps(_mm_shuffle_ps(B,B,0xF5), _mm_movehl_ps(DC,DC)));
-
-    //  d = trace(AB*DC) = trace(A#*B*D#*C) [continue]
-    d  = _mm_add_ps(d, _mm_movehl_ps(d, d));
-    d  = _mm_add_ss(d, _mm_shuffle_ps(d, d, 1));
-    d1 = _mm_mul_ss(dA,dD);
-    d2 = _mm_mul_ss(dB,dC);
-
-    //  iD = D*|A| - C*A#*B
-    iD = _mm_sub_ps(_mm_mul_ps(D,_mm_shuffle_ps(dA,dA,0)), iD);
-
-    //  iA = A*|D| - B*D#*C;
-    iA = _mm_sub_ps(_mm_mul_ps(A,_mm_shuffle_ps(dD,dD,0)), iA);
-
-    //  det = |A|*|D| + |B|*|C| - trace(A#*B*D#*C)
-    det = _mm_sub_ss(_mm_add_ss(d1,d2),d);
-    rd  = _mm_div_ss(_mm_set_ss(1.0f), det);
-
-//     #ifdef ZERO_SINGULAR
-//         rd = _mm_and_ps(_mm_cmpneq_ss(det,_mm_setzero_ps()), rd);
-//     #endif
-
-    //  iB = D * (A#B)# = D*B#*A
-    iB = _mm_mul_ps(D, _mm_shuffle_ps(AB,AB,0x33));
-    iB = _mm_sub_ps(iB, _mm_mul_ps(_mm_shuffle_ps(D,D,0xB1), _mm_shuffle_ps(AB,AB,0x66)));
-    //  iC = A * (D#C)# = A*C#*D
-    iC = _mm_mul_ps(A, _mm_shuffle_ps(DC,DC,0x33));
-    iC = _mm_sub_ps(iC, _mm_mul_ps(_mm_shuffle_ps(A,A,0xB1), _mm_shuffle_ps(DC,DC,0x66)));
-
-    rd = _mm_shuffle_ps(rd,rd,0);
-    rd = _mm_xor_ps(rd, p4f_sign_PNNP);
-
-    //  iB = C*|B| - D*B#*A
-    iB = _mm_sub_ps(_mm_mul_ps(C,_mm_shuffle_ps(dB,dB,0)), iB);
-
-    //  iC = B*|C| - A*C#*D;
-    iC = _mm_sub_ps(_mm_mul_ps(B,_mm_shuffle_ps(dC,dC,0)), iC);
-
-    //  iX = iX / det
-    iA = _mm_mul_ps(rd,iA);
-    iB = _mm_mul_ps(rd,iB);
-    iC = _mm_mul_ps(rd,iC);
-    iD = _mm_mul_ps(rd,iD);
-
-    Index res_stride = result.outerStride();
-    float* res = result.data();
-    pstoret<float, Packet4f, ResultAlignment>(res+0,            _mm_shuffle_ps(iA,iB,0x77));
-    pstoret<float, Packet4f, ResultAlignment>(res+res_stride,   _mm_shuffle_ps(iA,iB,0x22));
-    pstoret<float, Packet4f, ResultAlignment>(res+2*res_stride, _mm_shuffle_ps(iC,iD,0x77));
-    pstoret<float, Packet4f, ResultAlignment>(res+3*res_stride, _mm_shuffle_ps(iC,iD,0x22));
-  }
-
-};
-
-template<typename MatrixType, typename ResultType>
-struct compute_inverse_size4<Architecture::SSE, double, MatrixType, ResultType>
-{
-  enum {
-    MatrixAlignment     = traits<MatrixType>::Alignment,
-    ResultAlignment     = traits<ResultType>::Alignment,
-    StorageOrdersMatch  = (MatrixType::Flags&RowMajorBit) == (ResultType::Flags&RowMajorBit)
-  };
-  typedef typename conditional<(MatrixType::Flags&LinearAccessBit),MatrixType const &,typename MatrixType::PlainObject>::type ActualMatrixType;
-  
-  static void run(const MatrixType& mat, ResultType& result)
-  {
-    ActualMatrixType matrix(mat);
-    const __m128d _Sign_NP = _mm_castsi128_pd(_mm_set_epi32(0x0,0x0,0x80000000,0x0));
-    const __m128d _Sign_PN = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
-
-    // The inverse is calculated using "Divide and Conquer" technique. The
-    // original matrix is divide into four 2x2 sub-matrices. Since each
-    // register of the matrix holds two elements, the smaller matrices are
-    // consisted of two registers. Hence we get a better locality of the
-    // calculations.
-
-    // the four sub-matrices
-    __m128d A1, A2, B1, B2, C1, C2, D1, D2;
-    
-    if(StorageOrdersMatch)
-    {
-      A1 = matrix.template packet<MatrixAlignment>( 0); B1 = matrix.template packet<MatrixAlignment>( 2);
-      A2 = matrix.template packet<MatrixAlignment>( 4); B2 = matrix.template packet<MatrixAlignment>( 6);
-      C1 = matrix.template packet<MatrixAlignment>( 8); D1 = matrix.template packet<MatrixAlignment>(10);
-      C2 = matrix.template packet<MatrixAlignment>(12); D2 = matrix.template packet<MatrixAlignment>(14);
-    }
-    else
-    {
-      __m128d tmp;
-      A1 = matrix.template packet<MatrixAlignment>( 0); C1 = matrix.template packet<MatrixAlignment>( 2);
-      A2 = matrix.template packet<MatrixAlignment>( 4); C2 = matrix.template packet<MatrixAlignment>( 6);
-      tmp = A1;
-      A1 = _mm_unpacklo_pd(A1,A2);
-      A2 = _mm_unpackhi_pd(tmp,A2);
-      tmp = C1;
-      C1 = _mm_unpacklo_pd(C1,C2);
-      C2 = _mm_unpackhi_pd(tmp,C2);
-      
-      B1 = matrix.template packet<MatrixAlignment>( 8); D1 = matrix.template packet<MatrixAlignment>(10);
-      B2 = matrix.template packet<MatrixAlignment>(12); D2 = matrix.template packet<MatrixAlignment>(14);
-      tmp = B1;
-      B1 = _mm_unpacklo_pd(B1,B2);
-      B2 = _mm_unpackhi_pd(tmp,B2);
-      tmp = D1;
-      D1 = _mm_unpacklo_pd(D1,D2);
-      D2 = _mm_unpackhi_pd(tmp,D2);
-    }
-    
-    __m128d iA1, iA2, iB1, iB2, iC1, iC2, iD1, iD2,     // partial invese of the sub-matrices
-            DC1, DC2, AB1, AB2;
-    __m128d dA, dB, dC, dD;     // determinant of the sub-matrices
-    __m128d det, d1, d2, rd;
-
-    //  dA = |A|
-    dA = _mm_shuffle_pd(A2, A2, 1);
-    dA = _mm_mul_pd(A1, dA);
-    dA = _mm_sub_sd(dA, _mm_shuffle_pd(dA,dA,3));
-    //  dB = |B|
-    dB = _mm_shuffle_pd(B2, B2, 1);
-    dB = _mm_mul_pd(B1, dB);
-    dB = _mm_sub_sd(dB, _mm_shuffle_pd(dB,dB,3));
-
-    //  AB = A# * B
-    AB1 = _mm_mul_pd(B1, _mm_shuffle_pd(A2,A2,3));
-    AB2 = _mm_mul_pd(B2, _mm_shuffle_pd(A1,A1,0));
-    AB1 = _mm_sub_pd(AB1, _mm_mul_pd(B2, _mm_shuffle_pd(A1,A1,3)));
-    AB2 = _mm_sub_pd(AB2, _mm_mul_pd(B1, _mm_shuffle_pd(A2,A2,0)));
-
-    //  dC = |C|
-    dC = _mm_shuffle_pd(C2, C2, 1);
-    dC = _mm_mul_pd(C1, dC);
-    dC = _mm_sub_sd(dC, _mm_shuffle_pd(dC,dC,3));
-    //  dD = |D|
-    dD = _mm_shuffle_pd(D2, D2, 1);
-    dD = _mm_mul_pd(D1, dD);
-    dD = _mm_sub_sd(dD, _mm_shuffle_pd(dD,dD,3));
-
-    //  DC = D# * C
-    DC1 = _mm_mul_pd(C1, _mm_shuffle_pd(D2,D2,3));
-    DC2 = _mm_mul_pd(C2, _mm_shuffle_pd(D1,D1,0));
-    DC1 = _mm_sub_pd(DC1, _mm_mul_pd(C2, _mm_shuffle_pd(D1,D1,3)));
-    DC2 = _mm_sub_pd(DC2, _mm_mul_pd(C1, _mm_shuffle_pd(D2,D2,0)));
-
-    //  rd = trace(AB*DC) = trace(A#*B*D#*C)
-    d1 = _mm_mul_pd(AB1, _mm_shuffle_pd(DC1, DC2, 0));
-    d2 = _mm_mul_pd(AB2, _mm_shuffle_pd(DC1, DC2, 3));
-    rd = _mm_add_pd(d1, d2);
-    rd = _mm_add_sd(rd, _mm_shuffle_pd(rd, rd,3));
-
-    //  iD = C*A#*B
-    iD1 = _mm_mul_pd(AB1, _mm_shuffle_pd(C1,C1,0));
-    iD2 = _mm_mul_pd(AB1, _mm_shuffle_pd(C2,C2,0));
-    iD1 = _mm_add_pd(iD1, _mm_mul_pd(AB2, _mm_shuffle_pd(C1,C1,3)));
-    iD2 = _mm_add_pd(iD2, _mm_mul_pd(AB2, _mm_shuffle_pd(C2,C2,3)));
-
-    //  iA = B*D#*C
-    iA1 = _mm_mul_pd(DC1, _mm_shuffle_pd(B1,B1,0));
-    iA2 = _mm_mul_pd(DC1, _mm_shuffle_pd(B2,B2,0));
-    iA1 = _mm_add_pd(iA1, _mm_mul_pd(DC2, _mm_shuffle_pd(B1,B1,3)));
-    iA2 = _mm_add_pd(iA2, _mm_mul_pd(DC2, _mm_shuffle_pd(B2,B2,3)));
-
-    //  iD = D*|A| - C*A#*B
-    dA = _mm_shuffle_pd(dA,dA,0);
-    iD1 = _mm_sub_pd(_mm_mul_pd(D1, dA), iD1);
-    iD2 = _mm_sub_pd(_mm_mul_pd(D2, dA), iD2);
-
-    //  iA = A*|D| - B*D#*C;
-    dD = _mm_shuffle_pd(dD,dD,0);
-    iA1 = _mm_sub_pd(_mm_mul_pd(A1, dD), iA1);
-    iA2 = _mm_sub_pd(_mm_mul_pd(A2, dD), iA2);
-
-    d1 = _mm_mul_sd(dA, dD);
-    d2 = _mm_mul_sd(dB, dC);
-
-    //  iB = D * (A#B)# = D*B#*A
-    iB1 = _mm_mul_pd(D1, _mm_shuffle_pd(AB2,AB1,1));
-    iB2 = _mm_mul_pd(D2, _mm_shuffle_pd(AB2,AB1,1));
-    iB1 = _mm_sub_pd(iB1, _mm_mul_pd(_mm_shuffle_pd(D1,D1,1), _mm_shuffle_pd(AB2,AB1,2)));
-    iB2 = _mm_sub_pd(iB2, _mm_mul_pd(_mm_shuffle_pd(D2,D2,1), _mm_shuffle_pd(AB2,AB1,2)));
-
-    //  det = |A|*|D| + |B|*|C| - trace(A#*B*D#*C)
-    det = _mm_add_sd(d1, d2);
-    det = _mm_sub_sd(det, rd);
-
-    //  iC = A * (D#C)# = A*C#*D
-    iC1 = _mm_mul_pd(A1, _mm_shuffle_pd(DC2,DC1,1));
-    iC2 = _mm_mul_pd(A2, _mm_shuffle_pd(DC2,DC1,1));
-    iC1 = _mm_sub_pd(iC1, _mm_mul_pd(_mm_shuffle_pd(A1,A1,1), _mm_shuffle_pd(DC2,DC1,2)));
-    iC2 = _mm_sub_pd(iC2, _mm_mul_pd(_mm_shuffle_pd(A2,A2,1), _mm_shuffle_pd(DC2,DC1,2)));
-
-    rd = _mm_div_sd(_mm_set_sd(1.0), det);
-//     #ifdef ZERO_SINGULAR
-//         rd = _mm_and_pd(_mm_cmpneq_sd(det,_mm_setzero_pd()), rd);
-//     #endif
-    rd = _mm_shuffle_pd(rd,rd,0);
-
-    //  iB = C*|B| - D*B#*A
-    dB = _mm_shuffle_pd(dB,dB,0);
-    iB1 = _mm_sub_pd(_mm_mul_pd(C1, dB), iB1);
-    iB2 = _mm_sub_pd(_mm_mul_pd(C2, dB), iB2);
-
-    d1 = _mm_xor_pd(rd, _Sign_PN);
-    d2 = _mm_xor_pd(rd, _Sign_NP);
-
-    //  iC = B*|C| - A*C#*D;
-    dC = _mm_shuffle_pd(dC,dC,0);
-    iC1 = _mm_sub_pd(_mm_mul_pd(B1, dC), iC1);
-    iC2 = _mm_sub_pd(_mm_mul_pd(B2, dC), iC2);
-
-    Index res_stride = result.outerStride();
-    double* res = result.data();
-    pstoret<double, Packet2d, ResultAlignment>(res+0,             _mm_mul_pd(_mm_shuffle_pd(iA2, iA1, 3), d1));
-    pstoret<double, Packet2d, ResultAlignment>(res+res_stride,    _mm_mul_pd(_mm_shuffle_pd(iA2, iA1, 0), d2));
-    pstoret<double, Packet2d, ResultAlignment>(res+2,             _mm_mul_pd(_mm_shuffle_pd(iB2, iB1, 3), d1));
-    pstoret<double, Packet2d, ResultAlignment>(res+res_stride+2,  _mm_mul_pd(_mm_shuffle_pd(iB2, iB1, 0), d2));
-    pstoret<double, Packet2d, ResultAlignment>(res+2*res_stride,  _mm_mul_pd(_mm_shuffle_pd(iC2, iC1, 3), d1));
-    pstoret<double, Packet2d, ResultAlignment>(res+3*res_stride,  _mm_mul_pd(_mm_shuffle_pd(iC2, iC1, 0), d2));
-    pstoret<double, Packet2d, ResultAlignment>(res+2*res_stride+2,_mm_mul_pd(_mm_shuffle_pd(iD2, iD1, 3), d1));
-    pstoret<double, Packet2d, ResultAlignment>(res+3*res_stride+2,_mm_mul_pd(_mm_shuffle_pd(iD2, iD1, 0), d2));
-  }
-};
-
-} // end namespace internal
-
-} // end namespace Eigen
-
-#endif // EIGEN_INVERSE_SSE_H

package/eigen/Eigen/src/SparseCore/MappedSparseMatrix.h

@@ -1,67 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// 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
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-#ifndef EIGEN_MAPPED_SPARSEMATRIX_H
-#define EIGEN_MAPPED_SPARSEMATRIX_H
-
-namespace Eigen {
-
-/** \deprecated Use Map<SparseMatrix<> >
-  * \class MappedSparseMatrix
-  *
-  * \brief Sparse matrix
-  *
-  * \param _Scalar the scalar type, i.e. the type of the coefficients
-  *
-  * See http://www.netlib.org/linalg/html_templates/node91.html for details on the storage scheme.
-  *
-  */
-namespace internal {
-template<typename _Scalar, int _Flags, typename _StorageIndex>
-struct traits<MappedSparseMatrix<_Scalar, _Flags, _StorageIndex> > : traits<SparseMatrix<_Scalar, _Flags, _StorageIndex> >
-{};
-} // end namespace internal
-
-template<typename _Scalar, int _Flags, typename _StorageIndex>
-class MappedSparseMatrix
-  : public Map<SparseMatrix<_Scalar, _Flags, _StorageIndex> >
-{
-    typedef Map<SparseMatrix<_Scalar, _Flags, _StorageIndex> > Base;
-
-  public:
-    
-    typedef typename Base::StorageIndex StorageIndex;
-    typedef typename Base::Scalar Scalar;
-
-    inline MappedSparseMatrix(Index rows, Index cols, Index nnz, StorageIndex* outerIndexPtr, StorageIndex* innerIndexPtr, Scalar* valuePtr, StorageIndex* innerNonZeroPtr = 0)
-      : Base(rows, cols, nnz, outerIndexPtr, innerIndexPtr, valuePtr, innerNonZeroPtr)
-    {}
-
-    /** Empty destructor */
-    inline ~MappedSparseMatrix() {}
-};
-
-namespace internal {
-
-template<typename _Scalar, int _Options, typename _StorageIndex>
-struct evaluator<MappedSparseMatrix<_Scalar,_Options,_StorageIndex> >
-  : evaluator<SparseCompressedBase<MappedSparseMatrix<_Scalar,_Options,_StorageIndex> > >
-{
-  typedef MappedSparseMatrix<_Scalar,_Options,_StorageIndex> XprType;
-  typedef evaluator<SparseCompressedBase<XprType> > Base;
-  
-  evaluator() : Base() {}
-  explicit evaluator(const XprType &mat) : Base(mat) {}
-};
-
-}
-
-} // end namespace Eigen
-
-#endif // EIGEN_MAPPED_SPARSEMATRIX_H

package/eigen/Eigen/src/SparseLU/SparseLU_gemm_kernel.h

@@ -1,280 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2012 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_SPARSELU_GEMM_KERNEL_H
-#define EIGEN_SPARSELU_GEMM_KERNEL_H
-
-namespace Eigen {
-
-namespace internal {
-
-
-/** \internal
-  * A general matrix-matrix product kernel optimized for the SparseLU factorization.
-  *  - A, B, and C must be column major
-  *  - lda and ldc must be multiples of the respective packet size
-  *  - C must have the same alignment as A
-  */
-template<typename Scalar>
-EIGEN_DONT_INLINE
-void sparselu_gemm(Index m, Index n, Index d, const Scalar* A, Index lda, const Scalar* B, Index ldb, Scalar* C, Index ldc)
-{
-  using namespace Eigen::internal;
-  
-  typedef typename packet_traits<Scalar>::type Packet;
-  enum {
-    NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS,
-    PacketSize = packet_traits<Scalar>::size,
-    PM = 8,                             // peeling in M
-    RN = 2,                             // register blocking
-    RK = NumberOfRegisters>=16 ? 4 : 2, // register blocking
-    BM = 4096/sizeof(Scalar),           // number of rows of A-C per chunk
-    SM = PM*PacketSize                  // step along M
-  };
-  Index d_end = (d/RK)*RK;    // number of columns of A (rows of B) suitable for full register blocking
-  Index n_end = (n/RN)*RN;    // number of columns of B-C suitable for processing RN columns at once
-  Index i0 = internal::first_default_aligned(A,m);
-  
-  eigen_internal_assert(((lda%PacketSize)==0) && ((ldc%PacketSize)==0) && (i0==internal::first_default_aligned(C,m)));
-  
-  // handle the non aligned rows of A and C without any optimization:
-  for(Index i=0; i<i0; ++i)
-  {
-    for(Index j=0; j<n; ++j)
-    {
-      Scalar c = C[i+j*ldc];
-      for(Index k=0; k<d; ++k)
-        c += B[k+j*ldb] * A[i+k*lda];
-      C[i+j*ldc] = c;
-    }
-  }
-  // process the remaining rows per chunk of BM rows
-  for(Index ib=i0; ib<m; ib+=BM)
-  {
-    Index actual_b = std::min<Index>(BM, m-ib);                 // actual number of rows
-    Index actual_b_end1 = (actual_b/SM)*SM;                   // actual number of rows suitable for peeling
-    Index actual_b_end2 = (actual_b/PacketSize)*PacketSize;   // actual number of rows suitable for vectorization
-    
-    // Let's process two columns of B-C at once
-    for(Index j=0; j<n_end; j+=RN)
-    {
-      const Scalar* Bc0 = B+(j+0)*ldb;
-      const Scalar* Bc1 = B+(j+1)*ldb;
-      
-      for(Index k=0; k<d_end; k+=RK)
-      {
-        
-        // load and expand a RN x RK block of B
-        Packet b00, b10, b20, b30, b01, b11, b21, b31;
-                  { b00 = pset1<Packet>(Bc0[0]); }
-                  { b10 = pset1<Packet>(Bc0[1]); }
-        if(RK==4) { b20 = pset1<Packet>(Bc0[2]); }
-        if(RK==4) { b30 = pset1<Packet>(Bc0[3]); }
-                  { b01 = pset1<Packet>(Bc1[0]); }
-                  { b11 = pset1<Packet>(Bc1[1]); }
-        if(RK==4) { b21 = pset1<Packet>(Bc1[2]); }
-        if(RK==4) { b31 = pset1<Packet>(Bc1[3]); }
-        
-        Packet a0, a1, a2, a3, c0, c1, t0, t1;
-        
-        const Scalar* A0 = A+ib+(k+0)*lda;
-        const Scalar* A1 = A+ib+(k+1)*lda;
-        const Scalar* A2 = A+ib+(k+2)*lda;
-        const Scalar* A3 = A+ib+(k+3)*lda;
-        
-        Scalar* C0 = C+ib+(j+0)*ldc;
-        Scalar* C1 = C+ib+(j+1)*ldc;
-        
-                  a0 = pload<Packet>(A0);
-                  a1 = pload<Packet>(A1);
-        if(RK==4)
-        {
-          a2 = pload<Packet>(A2);
-          a3 = pload<Packet>(A3);
-        }
-        else
-        {
-          // workaround "may be used uninitialized in this function" warning
-          a2 = a3 = a0;
-        }
-        
-#define KMADD(c, a, b, tmp) {tmp = b; tmp = pmul(a,tmp); c = padd(c,tmp);}
-#define WORK(I)  \
-                     c0 = pload<Packet>(C0+i+(I)*PacketSize);    \
-                     c1 = pload<Packet>(C1+i+(I)*PacketSize);    \
-                     KMADD(c0, a0, b00, t0)                      \
-                     KMADD(c1, a0, b01, t1)                      \
-                     a0 = pload<Packet>(A0+i+(I+1)*PacketSize);  \
-                     KMADD(c0, a1, b10, t0)                      \
-                     KMADD(c1, a1, b11, t1)                      \
-                     a1 = pload<Packet>(A1+i+(I+1)*PacketSize);  \
-          if(RK==4){ KMADD(c0, a2, b20, t0)                     }\
-          if(RK==4){ KMADD(c1, a2, b21, t1)                     }\
-          if(RK==4){ a2 = pload<Packet>(A2+i+(I+1)*PacketSize); }\
-          if(RK==4){ KMADD(c0, a3, b30, t0)                     }\
-          if(RK==4){ KMADD(c1, a3, b31, t1)                     }\
-          if(RK==4){ a3 = pload<Packet>(A3+i+(I+1)*PacketSize); }\
-                     pstore(C0+i+(I)*PacketSize, c0);            \
-                     pstore(C1+i+(I)*PacketSize, c1)
-        
-        // process rows of A' - C' with aggressive vectorization and peeling 
-        for(Index i=0; i<actual_b_end1; i+=PacketSize*8)
-        {
-          EIGEN_ASM_COMMENT("SPARSELU_GEMML_KERNEL1");
-                    prefetch((A0+i+(5)*PacketSize));
-                    prefetch((A1+i+(5)*PacketSize));
-          if(RK==4) prefetch((A2+i+(5)*PacketSize));
-          if(RK==4) prefetch((A3+i+(5)*PacketSize));
-
-          WORK(0);
-          WORK(1);
-          WORK(2);
-          WORK(3);
-          WORK(4);
-          WORK(5);
-          WORK(6);
-          WORK(7);
-        }
-        // process the remaining rows with vectorization only
-        for(Index i=actual_b_end1; i<actual_b_end2; i+=PacketSize)
-        {
-          WORK(0);
-        }
-#undef WORK
-        // process the remaining rows without vectorization
-        for(Index i=actual_b_end2; i<actual_b; ++i)
-        {
-          if(RK==4)
-          {
-            C0[i] += A0[i]*Bc0[0]+A1[i]*Bc0[1]+A2[i]*Bc0[2]+A3[i]*Bc0[3];
-            C1[i] += A0[i]*Bc1[0]+A1[i]*Bc1[1]+A2[i]*Bc1[2]+A3[i]*Bc1[3];
-          }
-          else
-          {
-            C0[i] += A0[i]*Bc0[0]+A1[i]*Bc0[1];
-            C1[i] += A0[i]*Bc1[0]+A1[i]*Bc1[1];
-          }
-        }
-        
-        Bc0 += RK;
-        Bc1 += RK;
-      } // peeled loop on k
-    } // peeled loop on the columns j
-    // process the last column (we now perform a matrix-vector product)
-    if((n-n_end)>0)
-    {
-      const Scalar* Bc0 = B+(n-1)*ldb;
-      
-      for(Index k=0; k<d_end; k+=RK)
-      {
-        
-        // load and expand a 1 x RK block of B
-        Packet b00, b10, b20, b30;
-                  b00 = pset1<Packet>(Bc0[0]);
-                  b10 = pset1<Packet>(Bc0[1]);
-        if(RK==4) b20 = pset1<Packet>(Bc0[2]);
-        if(RK==4) b30 = pset1<Packet>(Bc0[3]);
-        
-        Packet a0, a1, a2, a3, c0, t0/*, t1*/;
-        
-        const Scalar* A0 = A+ib+(k+0)*lda;
-        const Scalar* A1 = A+ib+(k+1)*lda;
-        const Scalar* A2 = A+ib+(k+2)*lda;
-        const Scalar* A3 = A+ib+(k+3)*lda;
-        
-        Scalar* C0 = C+ib+(n_end)*ldc;
-        
-                  a0 = pload<Packet>(A0);
-                  a1 = pload<Packet>(A1);
-        if(RK==4)
-        {
-          a2 = pload<Packet>(A2);
-          a3 = pload<Packet>(A3);
-        }
-        else
-        {
-          // workaround "may be used uninitialized in this function" warning
-          a2 = a3 = a0;
-        }
-        
-#define WORK(I) \
-                   c0 = pload<Packet>(C0+i+(I)*PacketSize);     \
-                   KMADD(c0, a0, b00, t0)                       \
-                   a0 = pload<Packet>(A0+i+(I+1)*PacketSize);   \
-                   KMADD(c0, a1, b10, t0)                       \
-                   a1 = pload<Packet>(A1+i+(I+1)*PacketSize);   \
-        if(RK==4){ KMADD(c0, a2, b20, t0)                      }\
-        if(RK==4){ a2 = pload<Packet>(A2+i+(I+1)*PacketSize);  }\
-        if(RK==4){ KMADD(c0, a3, b30, t0)                      }\
-        if(RK==4){ a3 = pload<Packet>(A3+i+(I+1)*PacketSize);  }\
-                   pstore(C0+i+(I)*PacketSize, c0);
-        
-        // agressive vectorization and peeling
-        for(Index i=0; i<actual_b_end1; i+=PacketSize*8)
-        {
-          EIGEN_ASM_COMMENT("SPARSELU_GEMML_KERNEL2");
-          WORK(0);
-          WORK(1);
-          WORK(2);
-          WORK(3);
-          WORK(4);
-          WORK(5);
-          WORK(6);
-          WORK(7);
-        }
-        // vectorization only
-        for(Index i=actual_b_end1; i<actual_b_end2; i+=PacketSize)
-        {
-          WORK(0);
-        }
-        // remaining scalars
-        for(Index i=actual_b_end2; i<actual_b; ++i)
-        {
-          if(RK==4) 
-            C0[i] += A0[i]*Bc0[0]+A1[i]*Bc0[1]+A2[i]*Bc0[2]+A3[i]*Bc0[3];
-          else
-            C0[i] += A0[i]*Bc0[0]+A1[i]*Bc0[1];
-        }
-        
-        Bc0 += RK;
-#undef WORK
-      }
-    }
-    
-    // process the last columns of A, corresponding to the last rows of B
-    Index rd = d-d_end;
-    if(rd>0)
-    {
-      for(Index j=0; j<n; ++j)
-      {
-        enum {
-          Alignment = PacketSize>1 ? Aligned : 0
-        };
-        typedef Map<Matrix<Scalar,Dynamic,1>, Alignment > MapVector;
-        typedef Map<const Matrix<Scalar,Dynamic,1>, Alignment > ConstMapVector;
-        if(rd==1)       MapVector(C+j*ldc+ib,actual_b) += B[0+d_end+j*ldb] * ConstMapVector(A+(d_end+0)*lda+ib, actual_b);
-        
-        else if(rd==2)  MapVector(C+j*ldc+ib,actual_b) += B[0+d_end+j*ldb] * ConstMapVector(A+(d_end+0)*lda+ib, actual_b)
-                                                        + B[1+d_end+j*ldb] * ConstMapVector(A+(d_end+1)*lda+ib, actual_b);
-        
-        else            MapVector(C+j*ldc+ib,actual_b) += B[0+d_end+j*ldb] * ConstMapVector(A+(d_end+0)*lda+ib, actual_b)
-                                                        + B[1+d_end+j*ldb] * ConstMapVector(A+(d_end+1)*lda+ib, actual_b)
-                                                        + B[2+d_end+j*ldb] * ConstMapVector(A+(d_end+2)*lda+ib, actual_b);
-      }
-    }
-  
-  } // blocking on the rows of A and C
-}
-#undef KMADD
-
-} // namespace internal
-
-} // namespace Eigen
-
-#endif // EIGEN_SPARSELU_GEMM_KERNEL_H