tomoto 0.1.0

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

@@ -0,0 +1,315 @@
+/*
+ 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_MATRIX_MATRIX_BLAS_H
+#define EIGEN_TRIANGULAR_MATRIX_MATRIX_BLAS_H
+
+namespace Eigen { 
+
+namespace internal {
+
+
+template <typename Scalar, typename Index,
+          int Mode, bool LhsIsTriangular,
+          int LhsStorageOrder, bool ConjugateLhs,
+          int RhsStorageOrder, bool ConjugateRhs,
+          int ResStorageOrder>
+struct product_triangular_matrix_matrix_trmm :
+       product_triangular_matrix_matrix<Scalar,Index,Mode,
+          LhsIsTriangular,LhsStorageOrder,ConjugateLhs,
+          RhsStorageOrder, ConjugateRhs, ResStorageOrder, BuiltIn> {};
+
+
+// try to go to BLAS specialization
+#define EIGEN_BLAS_TRMM_SPECIALIZE(Scalar, LhsIsTriangular) \
+template <typename Index, int Mode, \
+          int LhsStorageOrder, bool ConjugateLhs, \
+          int RhsStorageOrder, bool ConjugateRhs> \
+struct product_triangular_matrix_matrix<Scalar,Index, Mode, LhsIsTriangular, \
+           LhsStorageOrder,ConjugateLhs, RhsStorageOrder,ConjugateRhs,ColMajor,Specialized> { \
+  static inline void run(Index _rows, Index _cols, Index _depth, const Scalar* _lhs, Index lhsStride,\
+    const Scalar* _rhs, Index rhsStride, Scalar* res, Index resStride, Scalar alpha, level3_blocking<Scalar,Scalar>& blocking) { \
+      product_triangular_matrix_matrix_trmm<Scalar,Index,Mode, \
+        LhsIsTriangular,LhsStorageOrder,ConjugateLhs, \
+        RhsStorageOrder, ConjugateRhs, ColMajor>::run( \
+        _rows, _cols, _depth, _lhs, lhsStride, _rhs, rhsStride, res, resStride, alpha, blocking); \
+  } \
+};
+
+EIGEN_BLAS_TRMM_SPECIALIZE(double, true)
+EIGEN_BLAS_TRMM_SPECIALIZE(double, false)
+EIGEN_BLAS_TRMM_SPECIALIZE(dcomplex, true)
+EIGEN_BLAS_TRMM_SPECIALIZE(dcomplex, false)
+EIGEN_BLAS_TRMM_SPECIALIZE(float, true)
+EIGEN_BLAS_TRMM_SPECIALIZE(float, false)
+EIGEN_BLAS_TRMM_SPECIALIZE(scomplex, true)
+EIGEN_BLAS_TRMM_SPECIALIZE(scomplex, false)
+
+// implements col-major += alpha * op(triangular) * op(general)
+#define EIGEN_BLAS_TRMM_L(EIGTYPE, BLASTYPE, EIGPREFIX, BLASFUNC) \
+template <typename Index, int Mode, \
+          int LhsStorageOrder, bool ConjugateLhs, \
+          int RhsStorageOrder, bool ConjugateRhs> \
+struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,true, \
+         LhsStorageOrder,ConjugateLhs,RhsStorageOrder,ConjugateRhs,ColMajor> \
+{ \
+  enum { \
+    IsLower = (Mode&Lower) == Lower, \
+    SetDiag = (Mode&(ZeroDiag|UnitDiag)) ? 0 : 1, \
+    IsUnitDiag  = (Mode&UnitDiag) ? 1 : 0, \
+    IsZeroDiag  = (Mode&ZeroDiag) ? 1 : 0, \
+    LowUp = IsLower ? Lower : Upper, \
+    conjA = ((LhsStorageOrder==ColMajor) && ConjugateLhs) ? 1 : 0 \
+  }; \
+\
+  static void run( \
+    Index _rows, Index _cols, Index _depth, \
+    const EIGTYPE* _lhs, Index lhsStride, \
+    const EIGTYPE* _rhs, Index rhsStride, \
+    EIGTYPE* res,        Index resStride, \
+    EIGTYPE alpha, level3_blocking<EIGTYPE,EIGTYPE>& blocking) \
+  { \
+   Index diagSize  = (std::min)(_rows,_depth); \
+   Index rows      = IsLower ? _rows : diagSize; \
+   Index depth     = IsLower ? diagSize : _depth; \
+   Index cols      = _cols; \
+\
+   typedef Matrix<EIGTYPE, Dynamic, Dynamic, LhsStorageOrder> MatrixLhs; \
+   typedef Matrix<EIGTYPE, Dynamic, Dynamic, RhsStorageOrder> MatrixRhs; \
+\
+/* Non-square case - doesn't fit to BLAS ?TRMM. Fall to default triangular product or call BLAS ?GEMM*/ \
+   if (rows != depth) { \
+\
+     /* FIXME handle mkl_domain_get_max_threads */ \
+     /*int nthr = mkl_domain_get_max_threads(EIGEN_BLAS_DOMAIN_BLAS);*/ int nthr = 1;\
+\
+     if (((nthr==1) && (((std::max)(rows,depth)-diagSize)/(double)diagSize < 0.5))) { \
+     /* Most likely no benefit to call TRMM or GEMM from BLAS */ \
+       product_triangular_matrix_matrix<EIGTYPE,Index,Mode,true, \
+       LhsStorageOrder,ConjugateLhs, RhsStorageOrder, ConjugateRhs, ColMajor, BuiltIn>::run( \
+           _rows, _cols, _depth, _lhs, lhsStride, _rhs, rhsStride, res, resStride, alpha, blocking); \
+     /*std::cout << "TRMM_L: A is not square! Go to Eigen TRMM implementation!\n";*/ \
+     } else { \
+     /* Make sense to call GEMM */ \
+       Map<const MatrixLhs, 0, OuterStride<> > lhsMap(_lhs,rows,depth,OuterStride<>(lhsStride)); \
+       MatrixLhs aa_tmp=lhsMap.template triangularView<Mode>(); \
+       BlasIndex aStride = convert_index<BlasIndex>(aa_tmp.outerStride()); \
+       gemm_blocking_space<ColMajor,EIGTYPE,EIGTYPE,Dynamic,Dynamic,Dynamic> gemm_blocking(_rows,_cols,_depth, 1, true); \
+       general_matrix_matrix_product<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,RhsStorageOrder,ConjugateRhs,ColMajor>::run( \
+       rows, cols, depth, aa_tmp.data(), aStride, _rhs, rhsStride, res, resStride, alpha, gemm_blocking, 0); \
+\
+     /*std::cout << "TRMM_L: A is not square! Go to BLAS GEMM implementation! " << nthr<<" \n";*/ \
+     } \
+     return; \
+   } \
+   char side = 'L', transa, uplo, diag = 'N'; \
+   EIGTYPE *b; \
+   const EIGTYPE *a; \
+   BlasIndex m, n, lda, ldb; \
+\
+/* Set m, n */ \
+   m = convert_index<BlasIndex>(diagSize); \
+   n = convert_index<BlasIndex>(cols); \
+\
+/* Set trans */ \
+   transa = (LhsStorageOrder==RowMajor) ? ((ConjugateLhs) ? 'C' : 'T') : 'N'; \
+\
+/* Set b, ldb */ \
+   Map<const MatrixRhs, 0, OuterStride<> > rhs(_rhs,depth,cols,OuterStride<>(rhsStride)); \
+   MatrixX##EIGPREFIX b_tmp; \
+\
+   if (ConjugateRhs) b_tmp = rhs.conjugate(); else b_tmp = rhs; \
+   b = b_tmp.data(); \
+   ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \
+\
+/* Set uplo */ \
+   uplo = IsLower ? 'L' : 'U'; \
+   if (LhsStorageOrder==RowMajor) uplo = (uplo == 'L') ? 'U' : 'L'; \
+/* Set a, lda */ \
+   Map<const MatrixLhs, 0, OuterStride<> > lhs(_lhs,rows,depth,OuterStride<>(lhsStride)); \
+   MatrixLhs a_tmp; \
+\
+   if ((conjA!=0) || (SetDiag==0)) { \
+     if (conjA) a_tmp = lhs.conjugate(); else a_tmp = lhs; \
+     if (IsZeroDiag) \
+       a_tmp.diagonal().setZero(); \
+     else if (IsUnitDiag) \
+       a_tmp.diagonal().setOnes();\
+     a = a_tmp.data(); \
+     lda = convert_index<BlasIndex>(a_tmp.outerStride()); \
+   } else { \
+     a = _lhs; \
+     lda = convert_index<BlasIndex>(lhsStride); \
+   } \
+   /*std::cout << "TRMM_L: A is square! Go to BLAS TRMM implementation! \n";*/ \
+/* call ?trmm*/ \
+   BLASFUNC(&side, &uplo, &transa, &diag, &m, &n, (const BLASTYPE*)&numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (BLASTYPE*)b, &ldb); \
+\
+/* Add op(a_triangular)*b into res*/ \
+   Map<MatrixX##EIGPREFIX, 0, OuterStride<> > res_tmp(res,rows,cols,OuterStride<>(resStride)); \
+   res_tmp=res_tmp+b_tmp; \
+  } \
+};
+
+#ifdef EIGEN_USE_MKL
+EIGEN_BLAS_TRMM_L(double, double, d, dtrmm)
+EIGEN_BLAS_TRMM_L(dcomplex, MKL_Complex16, cd, ztrmm)
+EIGEN_BLAS_TRMM_L(float, float, f, strmm)
+EIGEN_BLAS_TRMM_L(scomplex, MKL_Complex8, cf, ctrmm)
+#else
+EIGEN_BLAS_TRMM_L(double, double, d, dtrmm_)
+EIGEN_BLAS_TRMM_L(dcomplex, double, cd, ztrmm_)
+EIGEN_BLAS_TRMM_L(float, float, f, strmm_)
+EIGEN_BLAS_TRMM_L(scomplex, float, cf, ctrmm_)
+#endif
+
+// implements col-major += alpha * op(general) * op(triangular)
+#define EIGEN_BLAS_TRMM_R(EIGTYPE, BLASTYPE, EIGPREFIX, BLASFUNC) \
+template <typename Index, int Mode, \
+          int LhsStorageOrder, bool ConjugateLhs, \
+          int RhsStorageOrder, bool ConjugateRhs> \
+struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,false, \
+         LhsStorageOrder,ConjugateLhs,RhsStorageOrder,ConjugateRhs,ColMajor> \
+{ \
+  enum { \
+    IsLower = (Mode&Lower) == Lower, \
+    SetDiag = (Mode&(ZeroDiag|UnitDiag)) ? 0 : 1, \
+    IsUnitDiag  = (Mode&UnitDiag) ? 1 : 0, \
+    IsZeroDiag  = (Mode&ZeroDiag) ? 1 : 0, \
+    LowUp = IsLower ? Lower : Upper, \
+    conjA = ((RhsStorageOrder==ColMajor) && ConjugateRhs) ? 1 : 0 \
+  }; \
+\
+  static void run( \
+    Index _rows, Index _cols, Index _depth, \
+    const EIGTYPE* _lhs, Index lhsStride, \
+    const EIGTYPE* _rhs, Index rhsStride, \
+    EIGTYPE* res,        Index resStride, \
+    EIGTYPE alpha, level3_blocking<EIGTYPE,EIGTYPE>& blocking) \
+  { \
+   Index diagSize  = (std::min)(_cols,_depth); \
+   Index rows      = _rows; \
+   Index depth     = IsLower ? _depth : diagSize; \
+   Index cols      = IsLower ? diagSize : _cols; \
+\
+   typedef Matrix<EIGTYPE, Dynamic, Dynamic, LhsStorageOrder> MatrixLhs; \
+   typedef Matrix<EIGTYPE, Dynamic, Dynamic, RhsStorageOrder> MatrixRhs; \
+\
+/* Non-square case - doesn't fit to BLAS ?TRMM. Fall to default triangular product or call BLAS ?GEMM*/ \
+   if (cols != depth) { \
+\
+     int nthr = 1 /*mkl_domain_get_max_threads(EIGEN_BLAS_DOMAIN_BLAS)*/; \
+\
+     if ((nthr==1) && (((std::max)(cols,depth)-diagSize)/(double)diagSize < 0.5)) { \
+     /* Most likely no benefit to call TRMM or GEMM from BLAS*/ \
+       product_triangular_matrix_matrix<EIGTYPE,Index,Mode,false, \
+       LhsStorageOrder,ConjugateLhs, RhsStorageOrder, ConjugateRhs, ColMajor, BuiltIn>::run( \
+           _rows, _cols, _depth, _lhs, lhsStride, _rhs, rhsStride, res, resStride, alpha, blocking); \
+       /*std::cout << "TRMM_R: A is not square! Go to Eigen TRMM implementation!\n";*/ \
+     } else { \
+     /* Make sense to call GEMM */ \
+       Map<const MatrixRhs, 0, OuterStride<> > rhsMap(_rhs,depth,cols, OuterStride<>(rhsStride)); \
+       MatrixRhs aa_tmp=rhsMap.template triangularView<Mode>(); \
+       BlasIndex aStride = convert_index<BlasIndex>(aa_tmp.outerStride()); \
+       gemm_blocking_space<ColMajor,EIGTYPE,EIGTYPE,Dynamic,Dynamic,Dynamic> gemm_blocking(_rows,_cols,_depth, 1, true); \
+       general_matrix_matrix_product<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,RhsStorageOrder,ConjugateRhs,ColMajor>::run( \
+       rows, cols, depth, _lhs, lhsStride, aa_tmp.data(), aStride, res, resStride, alpha, gemm_blocking, 0); \
+\
+     /*std::cout << "TRMM_R: A is not square! Go to BLAS GEMM implementation! " << nthr<<" \n";*/ \
+     } \
+     return; \
+   } \
+   char side = 'R', transa, uplo, diag = 'N'; \
+   EIGTYPE *b; \
+   const EIGTYPE *a; \
+   BlasIndex m, n, lda, ldb; \
+\
+/* Set m, n */ \
+   m = convert_index<BlasIndex>(rows); \
+   n = convert_index<BlasIndex>(diagSize); \
+\
+/* Set trans */ \
+   transa = (RhsStorageOrder==RowMajor) ? ((ConjugateRhs) ? 'C' : 'T') : 'N'; \
+\
+/* Set b, ldb */ \
+   Map<const MatrixLhs, 0, OuterStride<> > lhs(_lhs,rows,depth,OuterStride<>(lhsStride)); \
+   MatrixX##EIGPREFIX b_tmp; \
+\
+   if (ConjugateLhs) b_tmp = lhs.conjugate(); else b_tmp = lhs; \
+   b = b_tmp.data(); \
+   ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \
+\
+/* Set uplo */ \
+   uplo = IsLower ? 'L' : 'U'; \
+   if (RhsStorageOrder==RowMajor) uplo = (uplo == 'L') ? 'U' : 'L'; \
+/* Set a, lda */ \
+   Map<const MatrixRhs, 0, OuterStride<> > rhs(_rhs,depth,cols, OuterStride<>(rhsStride)); \
+   MatrixRhs a_tmp; \
+\
+   if ((conjA!=0) || (SetDiag==0)) { \
+     if (conjA) a_tmp = rhs.conjugate(); else a_tmp = rhs; \
+     if (IsZeroDiag) \
+       a_tmp.diagonal().setZero(); \
+     else if (IsUnitDiag) \
+       a_tmp.diagonal().setOnes();\
+     a = a_tmp.data(); \
+     lda = convert_index<BlasIndex>(a_tmp.outerStride()); \
+   } else { \
+     a = _rhs; \
+     lda = convert_index<BlasIndex>(rhsStride); \
+   } \
+   /*std::cout << "TRMM_R: A is square! Go to BLAS TRMM implementation! \n";*/ \
+/* call ?trmm*/ \
+   BLASFUNC(&side, &uplo, &transa, &diag, &m, &n, (const BLASTYPE*)&numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (BLASTYPE*)b, &ldb); \
+\
+/* Add op(a_triangular)*b into res*/ \
+   Map<MatrixX##EIGPREFIX, 0, OuterStride<> > res_tmp(res,rows,cols,OuterStride<>(resStride)); \
+   res_tmp=res_tmp+b_tmp; \
+  } \
+};
+
+#ifdef EIGEN_USE_MKL
+EIGEN_BLAS_TRMM_R(double, double, d, dtrmm)
+EIGEN_BLAS_TRMM_R(dcomplex, MKL_Complex16, cd, ztrmm)
+EIGEN_BLAS_TRMM_R(float, float, f, strmm)
+EIGEN_BLAS_TRMM_R(scomplex, MKL_Complex8, cf, ctrmm)
+#else
+EIGEN_BLAS_TRMM_R(double, double, d, dtrmm_)
+EIGEN_BLAS_TRMM_R(dcomplex, double, cd, ztrmm_)
+EIGEN_BLAS_TRMM_R(float, float, f, strmm_)
+EIGEN_BLAS_TRMM_R(scomplex, float, cf, ctrmm_)
+#endif
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TRIANGULAR_MATRIX_MATRIX_BLAS_H

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

@@ -0,0 +1,350 @@
+// 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_TRIANGULARMATRIXVECTOR_H
+#define EIGEN_TRIANGULARMATRIXVECTOR_H
+
+namespace Eigen {
+
+namespace internal {
+
+template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs, int StorageOrder, int Version=Specialized>
+struct triangular_matrix_vector_product;
+
+template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs, int Version>
+struct triangular_matrix_vector_product<Index,Mode,LhsScalar,ConjLhs,RhsScalar,ConjRhs,ColMajor,Version>
+{
+  typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+  enum {
+    IsLower = ((Mode&Lower)==Lower),
+    HasUnitDiag = (Mode & UnitDiag)==UnitDiag,
+    HasZeroDiag = (Mode & ZeroDiag)==ZeroDiag
+  };
+  static EIGEN_DONT_INLINE  void run(Index _rows, Index _cols, const LhsScalar* _lhs, Index lhsStride,
+                                     const RhsScalar* _rhs, Index rhsIncr, ResScalar* _res, Index resIncr, const RhsScalar& alpha);
+};
+
+template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs, int Version>
+EIGEN_DONT_INLINE void triangular_matrix_vector_product<Index,Mode,LhsScalar,ConjLhs,RhsScalar,ConjRhs,ColMajor,Version>
+  ::run(Index _rows, Index _cols, const LhsScalar* _lhs, Index lhsStride,
+        const RhsScalar* _rhs, Index rhsIncr, ResScalar* _res, Index resIncr, const RhsScalar& alpha)
+  {
+    static const Index PanelWidth = EIGEN_TUNE_TRIANGULAR_PANEL_WIDTH;
+    Index size = (std::min)(_rows,_cols);
+    Index rows = IsLower ? _rows : (std::min)(_rows,_cols);
+    Index cols = IsLower ? (std::min)(_rows,_cols) : _cols;
+
+    typedef Map<const Matrix<LhsScalar,Dynamic,Dynamic,ColMajor>, 0, OuterStride<> > LhsMap;
+    const LhsMap lhs(_lhs,rows,cols,OuterStride<>(lhsStride));
+    typename conj_expr_if<ConjLhs,LhsMap>::type cjLhs(lhs);
+
+    typedef Map<const Matrix<RhsScalar,Dynamic,1>, 0, InnerStride<> > RhsMap;
+    const RhsMap rhs(_rhs,cols,InnerStride<>(rhsIncr));
+    typename conj_expr_if<ConjRhs,RhsMap>::type cjRhs(rhs);
+
+    typedef Map<Matrix<ResScalar,Dynamic,1> > ResMap;
+    ResMap res(_res,rows);
+
+    typedef const_blas_data_mapper<LhsScalar,Index,ColMajor> LhsMapper;
+    typedef const_blas_data_mapper<RhsScalar,Index,RowMajor> RhsMapper;
+
+    for (Index pi=0; pi<size; pi+=PanelWidth)
+    {
+      Index actualPanelWidth = (std::min)(PanelWidth, size-pi);
+      for (Index k=0; k<actualPanelWidth; ++k)
+      {
+        Index i = pi + k;
+        Index s = IsLower ? ((HasUnitDiag||HasZeroDiag) ? i+1 : i ) : pi;
+        Index r = IsLower ? actualPanelWidth-k : k+1;
+        if ((!(HasUnitDiag||HasZeroDiag)) || (--r)>0)
+          res.segment(s,r) += (alpha * cjRhs.coeff(i)) * cjLhs.col(i).segment(s,r);
+        if (HasUnitDiag)
+          res.coeffRef(i) += alpha * cjRhs.coeff(i);
+      }
+      Index r = IsLower ? rows - pi - actualPanelWidth : pi;
+      if (r>0)
+      {
+        Index s = IsLower ? pi+actualPanelWidth : 0;
+        general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,ConjLhs,RhsScalar,RhsMapper,ConjRhs,BuiltIn>::run(
+            r, actualPanelWidth,
+            LhsMapper(&lhs.coeffRef(s,pi), lhsStride),
+            RhsMapper(&rhs.coeffRef(pi), rhsIncr),
+            &res.coeffRef(s), resIncr, alpha);
+      }
+    }
+    if((!IsLower) && cols>size)
+    {
+      general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,ConjLhs,RhsScalar,RhsMapper,ConjRhs>::run(
+          rows, cols-size,
+          LhsMapper(&lhs.coeffRef(0,size), lhsStride),
+          RhsMapper(&rhs.coeffRef(size), rhsIncr),
+          _res, resIncr, alpha);
+    }
+  }
+
+template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs,int Version>
+struct triangular_matrix_vector_product<Index,Mode,LhsScalar,ConjLhs,RhsScalar,ConjRhs,RowMajor,Version>
+{
+  typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+  enum {
+    IsLower = ((Mode&Lower)==Lower),
+    HasUnitDiag = (Mode & UnitDiag)==UnitDiag,
+    HasZeroDiag = (Mode & ZeroDiag)==ZeroDiag
+  };
+  static EIGEN_DONT_INLINE void run(Index _rows, Index _cols, const LhsScalar* _lhs, Index lhsStride,
+                                    const RhsScalar* _rhs, Index rhsIncr, ResScalar* _res, Index resIncr, const ResScalar& alpha);
+};
+
+template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs,int Version>
+EIGEN_DONT_INLINE void triangular_matrix_vector_product<Index,Mode,LhsScalar,ConjLhs,RhsScalar,ConjRhs,RowMajor,Version>
+  ::run(Index _rows, Index _cols, const LhsScalar* _lhs, Index lhsStride,
+        const RhsScalar* _rhs, Index rhsIncr, ResScalar* _res, Index resIncr, const ResScalar& alpha)
+  {
+    static const Index PanelWidth = EIGEN_TUNE_TRIANGULAR_PANEL_WIDTH;
+    Index diagSize = (std::min)(_rows,_cols);
+    Index rows = IsLower ? _rows : diagSize;
+    Index cols = IsLower ? diagSize : _cols;
+
+    typedef Map<const Matrix<LhsScalar,Dynamic,Dynamic,RowMajor>, 0, OuterStride<> > LhsMap;
+    const LhsMap lhs(_lhs,rows,cols,OuterStride<>(lhsStride));
+    typename conj_expr_if<ConjLhs,LhsMap>::type cjLhs(lhs);
+
+    typedef Map<const Matrix<RhsScalar,Dynamic,1> > RhsMap;
+    const RhsMap rhs(_rhs,cols);
+    typename conj_expr_if<ConjRhs,RhsMap>::type cjRhs(rhs);
+
+    typedef Map<Matrix<ResScalar,Dynamic,1>, 0, InnerStride<> > ResMap;
+    ResMap res(_res,rows,InnerStride<>(resIncr));
+
+    typedef const_blas_data_mapper<LhsScalar,Index,RowMajor> LhsMapper;
+    typedef const_blas_data_mapper<RhsScalar,Index,RowMajor> RhsMapper;
+
+    for (Index pi=0; pi<diagSize; pi+=PanelWidth)
+    {
+      Index actualPanelWidth = (std::min)(PanelWidth, diagSize-pi);
+      for (Index k=0; k<actualPanelWidth; ++k)
+      {
+        Index i = pi + k;
+        Index s = IsLower ? pi  : ((HasUnitDiag||HasZeroDiag) ? i+1 : i);
+        Index r = IsLower ? k+1 : actualPanelWidth-k;
+        if ((!(HasUnitDiag||HasZeroDiag)) || (--r)>0)
+          res.coeffRef(i) += alpha * (cjLhs.row(i).segment(s,r).cwiseProduct(cjRhs.segment(s,r).transpose())).sum();
+        if (HasUnitDiag)
+          res.coeffRef(i) += alpha * cjRhs.coeff(i);
+      }
+      Index r = IsLower ? pi : cols - pi - actualPanelWidth;
+      if (r>0)
+      {
+        Index s = IsLower ? 0 : pi + actualPanelWidth;
+        general_matrix_vector_product<Index,LhsScalar,LhsMapper,RowMajor,ConjLhs,RhsScalar,RhsMapper,ConjRhs,BuiltIn>::run(
+            actualPanelWidth, r,
+            LhsMapper(&lhs.coeffRef(pi,s), lhsStride),
+            RhsMapper(&rhs.coeffRef(s), rhsIncr),
+            &res.coeffRef(pi), resIncr, alpha);
+      }
+    }
+    if(IsLower && rows>diagSize)
+    {
+      general_matrix_vector_product<Index,LhsScalar,LhsMapper,RowMajor,ConjLhs,RhsScalar,RhsMapper,ConjRhs>::run(
+            rows-diagSize, cols,
+            LhsMapper(&lhs.coeffRef(diagSize,0), lhsStride),
+            RhsMapper(&rhs.coeffRef(0), rhsIncr),
+            &res.coeffRef(diagSize), resIncr, alpha);
+    }
+  }
+
+/***************************************************************************
+* Wrapper to product_triangular_vector
+***************************************************************************/
+
+template<int Mode,int StorageOrder>
+struct trmv_selector;
+
+} // end namespace internal
+
+namespace internal {
+
+template<int Mode, typename Lhs, typename Rhs>
+struct triangular_product_impl<Mode,true,Lhs,false,Rhs,true>
+{
+  template<typename Dest> static void run(Dest& dst, const Lhs &lhs, const Rhs &rhs, const typename Dest::Scalar& alpha)
+  {
+    eigen_assert(dst.rows()==lhs.rows() && dst.cols()==rhs.cols());
+  
+    internal::trmv_selector<Mode,(int(internal::traits<Lhs>::Flags)&RowMajorBit) ? RowMajor : ColMajor>::run(lhs, rhs, dst, alpha);
+  }
+};
+
+template<int Mode, typename Lhs, typename Rhs>
+struct triangular_product_impl<Mode,false,Lhs,true,Rhs,false>
+{
+  template<typename Dest> static void run(Dest& dst, const Lhs &lhs, const Rhs &rhs, const typename Dest::Scalar& alpha)
+  {
+    eigen_assert(dst.rows()==lhs.rows() && dst.cols()==rhs.cols());
+
+    Transpose<Dest> dstT(dst);
+    internal::trmv_selector<(Mode & (UnitDiag|ZeroDiag)) | ((Mode & Lower) ? Upper : Lower),
+                            (int(internal::traits<Rhs>::Flags)&RowMajorBit) ? ColMajor : RowMajor>
+            ::run(rhs.transpose(),lhs.transpose(), dstT, alpha);
+  }
+};
+
+} // end namespace internal
+
+namespace internal {
+
+// TODO: find a way to factorize this piece of code with gemv_selector since the logic is exactly the same.
+  
+template<int Mode> struct trmv_selector<Mode,ColMajor>
+{
+  template<typename Lhs, typename Rhs, typename Dest>
+  static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
+  {
+    typedef typename Lhs::Scalar      LhsScalar;
+    typedef typename Rhs::Scalar      RhsScalar;
+    typedef typename Dest::Scalar     ResScalar;
+    typedef typename Dest::RealScalar RealScalar;
+    
+    typedef internal::blas_traits<Lhs> LhsBlasTraits;
+    typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
+    typedef internal::blas_traits<Rhs> RhsBlasTraits;
+    typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
+    
+    typedef Map<Matrix<ResScalar,Dynamic,1>, EIGEN_PLAIN_ENUM_MIN(AlignedMax,internal::packet_traits<ResScalar>::size)> MappedDest;
+
+    typename internal::add_const_on_value_type<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(lhs);
+    typename internal::add_const_on_value_type<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(rhs);
+
+    LhsScalar lhs_alpha = LhsBlasTraits::extractScalarFactor(lhs);
+    RhsScalar rhs_alpha = RhsBlasTraits::extractScalarFactor(rhs);
+    ResScalar actualAlpha = alpha * lhs_alpha * rhs_alpha;
+
+    enum {
+      // FIXME find a way to allow an inner stride on the result if packet_traits<Scalar>::size==1
+      // on, the other hand it is good for the cache to pack the vector anyways...
+      EvalToDestAtCompileTime = Dest::InnerStrideAtCompileTime==1,
+      ComplexByReal = (NumTraits<LhsScalar>::IsComplex) && (!NumTraits<RhsScalar>::IsComplex),
+      MightCannotUseDest = (Dest::InnerStrideAtCompileTime!=1) || ComplexByReal
+    };
+
+    gemv_static_vector_if<ResScalar,Dest::SizeAtCompileTime,Dest::MaxSizeAtCompileTime,MightCannotUseDest> static_dest;
+
+    bool alphaIsCompatible = (!ComplexByReal) || (numext::imag(actualAlpha)==RealScalar(0));
+    bool evalToDest = EvalToDestAtCompileTime && alphaIsCompatible;
+
+    RhsScalar compatibleAlpha = get_factor<ResScalar,RhsScalar>::run(actualAlpha);
+
+    ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(),
+                                                  evalToDest ? dest.data() : static_dest.data());
+
+    if(!evalToDest)
+    {
+      #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
+      Index size = dest.size();
+      EIGEN_DENSE_STORAGE_CTOR_PLUGIN
+      #endif
+      if(!alphaIsCompatible)
+      {
+        MappedDest(actualDestPtr, dest.size()).setZero();
+        compatibleAlpha = RhsScalar(1);
+      }
+      else
+        MappedDest(actualDestPtr, dest.size()) = dest;
+    }
+
+    internal::triangular_matrix_vector_product
+      <Index,Mode,
+       LhsScalar, LhsBlasTraits::NeedToConjugate,
+       RhsScalar, RhsBlasTraits::NeedToConjugate,
+       ColMajor>
+      ::run(actualLhs.rows(),actualLhs.cols(),
+            actualLhs.data(),actualLhs.outerStride(),
+            actualRhs.data(),actualRhs.innerStride(),
+            actualDestPtr,1,compatibleAlpha);
+
+    if (!evalToDest)
+    {
+      if(!alphaIsCompatible)
+        dest += actualAlpha * MappedDest(actualDestPtr, dest.size());
+      else
+        dest = MappedDest(actualDestPtr, dest.size());
+    }
+
+    if ( ((Mode&UnitDiag)==UnitDiag) && (lhs_alpha!=LhsScalar(1)) )
+    {
+      Index diagSize = (std::min)(lhs.rows(),lhs.cols());
+      dest.head(diagSize) -= (lhs_alpha-LhsScalar(1))*rhs.head(diagSize);
+    }
+  }
+};
+
+template<int Mode> struct trmv_selector<Mode,RowMajor>
+{
+  template<typename Lhs, typename Rhs, typename Dest>
+  static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
+  {
+    typedef typename Lhs::Scalar      LhsScalar;
+    typedef typename Rhs::Scalar      RhsScalar;
+    typedef typename Dest::Scalar     ResScalar;
+    
+    typedef internal::blas_traits<Lhs> LhsBlasTraits;
+    typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
+    typedef internal::blas_traits<Rhs> RhsBlasTraits;
+    typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
+    typedef typename internal::remove_all<ActualRhsType>::type ActualRhsTypeCleaned;
+
+    typename add_const<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(lhs);
+    typename add_const<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(rhs);
+
+    LhsScalar lhs_alpha = LhsBlasTraits::extractScalarFactor(lhs);
+    RhsScalar rhs_alpha = RhsBlasTraits::extractScalarFactor(rhs);
+    ResScalar actualAlpha = alpha * lhs_alpha * rhs_alpha;
+
+    enum {
+      DirectlyUseRhs = ActualRhsTypeCleaned::InnerStrideAtCompileTime==1
+    };
+
+    gemv_static_vector_if<RhsScalar,ActualRhsTypeCleaned::SizeAtCompileTime,ActualRhsTypeCleaned::MaxSizeAtCompileTime,!DirectlyUseRhs> static_rhs;
+
+    ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhsPtr,actualRhs.size(),
+        DirectlyUseRhs ? const_cast<RhsScalar*>(actualRhs.data()) : static_rhs.data());
+
+    if(!DirectlyUseRhs)
+    {
+      #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
+      Index size = actualRhs.size();
+      EIGEN_DENSE_STORAGE_CTOR_PLUGIN
+      #endif
+      Map<typename ActualRhsTypeCleaned::PlainObject>(actualRhsPtr, actualRhs.size()) = actualRhs;
+    }
+
+    internal::triangular_matrix_vector_product
+      <Index,Mode,
+       LhsScalar, LhsBlasTraits::NeedToConjugate,
+       RhsScalar, RhsBlasTraits::NeedToConjugate,
+       RowMajor>
+      ::run(actualLhs.rows(),actualLhs.cols(),
+            actualLhs.data(),actualLhs.outerStride(),
+            actualRhsPtr,1,
+            dest.data(),dest.innerStride(),
+            actualAlpha);
+
+    if ( ((Mode&UnitDiag)==UnitDiag) && (lhs_alpha!=LhsScalar(1)) )
+    {
+      Index diagSize = (std::min)(lhs.rows(),lhs.cols());
+      dest.head(diagSize) -= (lhs_alpha-LhsScalar(1))*rhs.head(diagSize);
+    }
+  }
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TRIANGULARMATRIXVECTOR_H