tomoto 0.1.0

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

@@ -0,0 +1,118 @@
+/*
+ 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
+ *   Selfadjoint matrix-vector product functionality based on ?SYMV/HEMV.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_SELFADJOINT_MATRIX_VECTOR_BLAS_H
+#define EIGEN_SELFADJOINT_MATRIX_VECTOR_BLAS_H
+
+namespace Eigen { 
+
+namespace internal {
+
+/**********************************************************************
+* This file implements selfadjoint matrix-vector multiplication using BLAS
+**********************************************************************/
+
+// symv/hemv specialization
+
+template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs>
+struct selfadjoint_matrix_vector_product_symv :
+  selfadjoint_matrix_vector_product<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs,BuiltIn> {};
+
+#define EIGEN_BLAS_SYMV_SPECIALIZE(Scalar) \
+template<typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs> \
+struct selfadjoint_matrix_vector_product<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs,Specialized> { \
+static void run( \
+  Index size, const Scalar*  lhs, Index lhsStride, \
+  const Scalar* _rhs, Scalar* res, Scalar alpha) { \
+    enum {\
+      IsColMajor = StorageOrder==ColMajor \
+    }; \
+    if (IsColMajor == ConjugateLhs) {\
+      selfadjoint_matrix_vector_product<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs,BuiltIn>::run( \
+        size, lhs, lhsStride, _rhs, res, alpha);  \
+    } else {\
+      selfadjoint_matrix_vector_product_symv<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs>::run( \
+        size, lhs, lhsStride, _rhs, res, alpha);  \
+    }\
+  } \
+}; \
+
+EIGEN_BLAS_SYMV_SPECIALIZE(double)
+EIGEN_BLAS_SYMV_SPECIALIZE(float)
+EIGEN_BLAS_SYMV_SPECIALIZE(dcomplex)
+EIGEN_BLAS_SYMV_SPECIALIZE(scomplex)
+
+#define EIGEN_BLAS_SYMV_SPECIALIZATION(EIGTYPE,BLASTYPE,BLASFUNC) \
+template<typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs> \
+struct selfadjoint_matrix_vector_product_symv<EIGTYPE,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs> \
+{ \
+typedef Matrix<EIGTYPE,Dynamic,1,ColMajor> SYMVVector;\
+\
+static void run( \
+Index size, const EIGTYPE*  lhs, Index lhsStride, \
+const EIGTYPE* _rhs, EIGTYPE* res, EIGTYPE alpha) \
+{ \
+  enum {\
+    IsRowMajor = StorageOrder==RowMajor ? 1 : 0, \
+    IsLower = UpLo == Lower ? 1 : 0 \
+  }; \
+  BlasIndex n=convert_index<BlasIndex>(size), lda=convert_index<BlasIndex>(lhsStride), incx=1, incy=1; \
+  EIGTYPE beta(1); \
+  const EIGTYPE *x_ptr; \
+  char uplo=(IsRowMajor) ? (IsLower ? 'U' : 'L') : (IsLower ? 'L' : 'U'); \
+  SYMVVector x_tmp; \
+  if (ConjugateRhs) { \
+    Map<const SYMVVector, 0 > map_x(_rhs,size,1); \
+    x_tmp=map_x.conjugate(); \
+    x_ptr=x_tmp.data(); \
+  } else x_ptr=_rhs; \
+  BLASFUNC(&uplo, &n, (const BLASTYPE*)&numext::real_ref(alpha), (const BLASTYPE*)lhs, &lda, (const BLASTYPE*)x_ptr, &incx, (const BLASTYPE*)&numext::real_ref(beta), (BLASTYPE*)res, &incy); \
+}\
+};
+
+#ifdef EIGEN_USE_MKL
+EIGEN_BLAS_SYMV_SPECIALIZATION(double,   double, dsymv)
+EIGEN_BLAS_SYMV_SPECIALIZATION(float,    float,  ssymv)
+EIGEN_BLAS_SYMV_SPECIALIZATION(dcomplex, MKL_Complex16, zhemv)
+EIGEN_BLAS_SYMV_SPECIALIZATION(scomplex, MKL_Complex8,  chemv)
+#else
+EIGEN_BLAS_SYMV_SPECIALIZATION(double,   double, dsymv_)
+EIGEN_BLAS_SYMV_SPECIALIZATION(float,    float,  ssymv_)
+EIGEN_BLAS_SYMV_SPECIALIZATION(dcomplex, double, zhemv_)
+EIGEN_BLAS_SYMV_SPECIALIZATION(scomplex, float,  chemv_)
+#endif
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SELFADJOINT_MATRIX_VECTOR_BLAS_H

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

@@ -0,0 +1,133 @@
+// 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_SELFADJOINT_PRODUCT_H
+#define EIGEN_SELFADJOINT_PRODUCT_H
+
+/**********************************************************************
+* This file implements a self adjoint product: C += A A^T updating only
+* half of the selfadjoint matrix C.
+* It corresponds to the level 3 SYRK and level 2 SYR Blas routines.
+**********************************************************************/
+
+namespace Eigen { 
+
+
+template<typename Scalar, typename Index, int UpLo, bool ConjLhs, bool ConjRhs>
+struct selfadjoint_rank1_update<Scalar,Index,ColMajor,UpLo,ConjLhs,ConjRhs>
+{
+  static void run(Index size, Scalar* mat, Index stride, const Scalar* vecX, const Scalar* vecY, const Scalar& alpha)
+  {
+    internal::conj_if<ConjRhs> cj;
+    typedef Map<const Matrix<Scalar,Dynamic,1> > OtherMap;
+    typedef typename internal::conditional<ConjLhs,typename OtherMap::ConjugateReturnType,const OtherMap&>::type ConjLhsType;
+    for (Index i=0; i<size; ++i)
+    {
+      Map<Matrix<Scalar,Dynamic,1> >(mat+stride*i+(UpLo==Lower ? i : 0), (UpLo==Lower ? size-i : (i+1)))
+          += (alpha * cj(vecY[i])) * ConjLhsType(OtherMap(vecX+(UpLo==Lower ? i : 0),UpLo==Lower ? size-i : (i+1)));
+    }
+  }
+};
+
+template<typename Scalar, typename Index, int UpLo, bool ConjLhs, bool ConjRhs>
+struct selfadjoint_rank1_update<Scalar,Index,RowMajor,UpLo,ConjLhs,ConjRhs>
+{
+  static void run(Index size, Scalar* mat, Index stride, const Scalar* vecX, const Scalar* vecY, const Scalar& alpha)
+  {
+    selfadjoint_rank1_update<Scalar,Index,ColMajor,UpLo==Lower?Upper:Lower,ConjRhs,ConjLhs>::run(size,mat,stride,vecY,vecX,alpha);
+  }
+};
+
+template<typename MatrixType, typename OtherType, int UpLo, bool OtherIsVector = OtherType::IsVectorAtCompileTime>
+struct selfadjoint_product_selector;
+
+template<typename MatrixType, typename OtherType, int UpLo>
+struct selfadjoint_product_selector<MatrixType,OtherType,UpLo,true>
+{
+  static void run(MatrixType& mat, const OtherType& other, const typename MatrixType::Scalar& alpha)
+  {
+    typedef typename MatrixType::Scalar Scalar;
+    typedef internal::blas_traits<OtherType> OtherBlasTraits;
+    typedef typename OtherBlasTraits::DirectLinearAccessType ActualOtherType;
+    typedef typename internal::remove_all<ActualOtherType>::type _ActualOtherType;
+    typename internal::add_const_on_value_type<ActualOtherType>::type actualOther = OtherBlasTraits::extract(other.derived());
+
+    Scalar actualAlpha = alpha * OtherBlasTraits::extractScalarFactor(other.derived());
+
+    enum {
+      StorageOrder = (internal::traits<MatrixType>::Flags&RowMajorBit) ? RowMajor : ColMajor,
+      UseOtherDirectly = _ActualOtherType::InnerStrideAtCompileTime==1
+    };
+    internal::gemv_static_vector_if<Scalar,OtherType::SizeAtCompileTime,OtherType::MaxSizeAtCompileTime,!UseOtherDirectly> static_other;
+
+    ei_declare_aligned_stack_constructed_variable(Scalar, actualOtherPtr, other.size(),
+      (UseOtherDirectly ? const_cast<Scalar*>(actualOther.data()) : static_other.data()));
+      
+    if(!UseOtherDirectly)
+      Map<typename _ActualOtherType::PlainObject>(actualOtherPtr, actualOther.size()) = actualOther;
+    
+    selfadjoint_rank1_update<Scalar,Index,StorageOrder,UpLo,
+                              OtherBlasTraits::NeedToConjugate  && NumTraits<Scalar>::IsComplex,
+                            (!OtherBlasTraits::NeedToConjugate) && NumTraits<Scalar>::IsComplex>
+          ::run(other.size(), mat.data(), mat.outerStride(), actualOtherPtr, actualOtherPtr, actualAlpha);
+  }
+};
+
+template<typename MatrixType, typename OtherType, int UpLo>
+struct selfadjoint_product_selector<MatrixType,OtherType,UpLo,false>
+{
+  static void run(MatrixType& mat, const OtherType& other, const typename MatrixType::Scalar& alpha)
+  {
+    typedef typename MatrixType::Scalar Scalar;
+    typedef internal::blas_traits<OtherType> OtherBlasTraits;
+    typedef typename OtherBlasTraits::DirectLinearAccessType ActualOtherType;
+    typedef typename internal::remove_all<ActualOtherType>::type _ActualOtherType;
+    typename internal::add_const_on_value_type<ActualOtherType>::type actualOther = OtherBlasTraits::extract(other.derived());
+
+    Scalar actualAlpha = alpha * OtherBlasTraits::extractScalarFactor(other.derived());
+
+    enum {
+      IsRowMajor = (internal::traits<MatrixType>::Flags&RowMajorBit) ? 1 : 0,
+      OtherIsRowMajor = _ActualOtherType::Flags&RowMajorBit ? 1 : 0
+    };
+
+    Index size = mat.cols();
+    Index depth = actualOther.cols();
+
+    typedef internal::gemm_blocking_space<IsRowMajor ? RowMajor : ColMajor,Scalar,Scalar,
+              MatrixType::MaxColsAtCompileTime, MatrixType::MaxColsAtCompileTime, _ActualOtherType::MaxColsAtCompileTime> BlockingType;
+
+    BlockingType blocking(size, size, depth, 1, false);
+
+
+    internal::general_matrix_matrix_triangular_product<Index,
+      Scalar, OtherIsRowMajor ? RowMajor : ColMajor,   OtherBlasTraits::NeedToConjugate  && NumTraits<Scalar>::IsComplex,
+      Scalar, OtherIsRowMajor ? ColMajor : RowMajor, (!OtherBlasTraits::NeedToConjugate) && NumTraits<Scalar>::IsComplex,
+      IsRowMajor ? RowMajor : ColMajor, UpLo>
+      ::run(size, depth,
+            &actualOther.coeffRef(0,0), actualOther.outerStride(), &actualOther.coeffRef(0,0), actualOther.outerStride(),
+            mat.data(), mat.outerStride(), actualAlpha, blocking);
+  }
+};
+
+// high level API
+
+template<typename MatrixType, unsigned int UpLo>
+template<typename DerivedU>
+SelfAdjointView<MatrixType,UpLo>& SelfAdjointView<MatrixType,UpLo>
+::rankUpdate(const MatrixBase<DerivedU>& u, const Scalar& alpha)
+{
+  selfadjoint_product_selector<MatrixType,DerivedU,UpLo>::run(_expression().const_cast_derived(), u.derived(), alpha);
+
+  return *this;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SELFADJOINT_PRODUCT_H

data/vendor/eigen/Eigen/src/Core/products/SelfadjointRank2Update.h

@@ -0,0 +1,93 @@
+// 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_SELFADJOINTRANK2UPTADE_H
+#define EIGEN_SELFADJOINTRANK2UPTADE_H
+
+namespace Eigen { 
+
+namespace internal {
+
+/* Optimized selfadjoint matrix += alpha * uv' + conj(alpha)*vu'
+ * It corresponds to the Level2 syr2 BLAS routine
+ */
+
+template<typename Scalar, typename Index, typename UType, typename VType, int UpLo>
+struct selfadjoint_rank2_update_selector;
+
+template<typename Scalar, typename Index, typename UType, typename VType>
+struct selfadjoint_rank2_update_selector<Scalar,Index,UType,VType,Lower>
+{
+  static void run(Scalar* mat, Index stride, const UType& u, const VType& v, const Scalar& alpha)
+  {
+    const Index size = u.size();
+    for (Index i=0; i<size; ++i)
+    {
+      Map<Matrix<Scalar,Dynamic,1> >(mat+stride*i+i, size-i) +=
+                        (numext::conj(alpha) * numext::conj(u.coeff(i))) * v.tail(size-i)
+                      + (alpha * numext::conj(v.coeff(i))) * u.tail(size-i);
+    }
+  }
+};
+
+template<typename Scalar, typename Index, typename UType, typename VType>
+struct selfadjoint_rank2_update_selector<Scalar,Index,UType,VType,Upper>
+{
+  static void run(Scalar* mat, Index stride, const UType& u, const VType& v, const Scalar& alpha)
+  {
+    const Index size = u.size();
+    for (Index i=0; i<size; ++i)
+      Map<Matrix<Scalar,Dynamic,1> >(mat+stride*i, i+1) +=
+                        (numext::conj(alpha)  * numext::conj(u.coeff(i))) * v.head(i+1)
+                      + (alpha * numext::conj(v.coeff(i))) * u.head(i+1);
+  }
+};
+
+template<bool Cond, typename T> struct conj_expr_if
+  : conditional<!Cond, const T&,
+      CwiseUnaryOp<scalar_conjugate_op<typename traits<T>::Scalar>,T> > {};
+
+} // end namespace internal
+
+template<typename MatrixType, unsigned int UpLo>
+template<typename DerivedU, typename DerivedV>
+SelfAdjointView<MatrixType,UpLo>& SelfAdjointView<MatrixType,UpLo>
+::rankUpdate(const MatrixBase<DerivedU>& u, const MatrixBase<DerivedV>& v, const Scalar& alpha)
+{
+  typedef internal::blas_traits<DerivedU> UBlasTraits;
+  typedef typename UBlasTraits::DirectLinearAccessType ActualUType;
+  typedef typename internal::remove_all<ActualUType>::type _ActualUType;
+  typename internal::add_const_on_value_type<ActualUType>::type actualU = UBlasTraits::extract(u.derived());
+
+  typedef internal::blas_traits<DerivedV> VBlasTraits;
+  typedef typename VBlasTraits::DirectLinearAccessType ActualVType;
+  typedef typename internal::remove_all<ActualVType>::type _ActualVType;
+  typename internal::add_const_on_value_type<ActualVType>::type actualV = VBlasTraits::extract(v.derived());
+
+  // If MatrixType is row major, then we use the routine for lower triangular in the upper triangular case and
+  // vice versa, and take the complex conjugate of all coefficients and vector entries.
+
+  enum { IsRowMajor = (internal::traits<MatrixType>::Flags&RowMajorBit) ? 1 : 0 };
+  Scalar actualAlpha = alpha * UBlasTraits::extractScalarFactor(u.derived())
+                             * numext::conj(VBlasTraits::extractScalarFactor(v.derived()));
+  if (IsRowMajor)
+    actualAlpha = numext::conj(actualAlpha);
+
+  typedef typename internal::remove_all<typename internal::conj_expr_if<IsRowMajor ^ UBlasTraits::NeedToConjugate,_ActualUType>::type>::type UType;
+  typedef typename internal::remove_all<typename internal::conj_expr_if<IsRowMajor ^ VBlasTraits::NeedToConjugate,_ActualVType>::type>::type VType;
+  internal::selfadjoint_rank2_update_selector<Scalar, Index, UType, VType,
+    (IsRowMajor ? int(UpLo==Upper ? Lower : Upper) : UpLo)>
+    ::run(_expression().const_cast_derived().data(),_expression().outerStride(),UType(actualU),VType(actualV),actualAlpha);
+
+  return *this;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SELFADJOINTRANK2UPTADE_H

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

@@ -0,0 +1,466 @@
+// 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_TRIANGULAR_MATRIX_MATRIX_H
+#define EIGEN_TRIANGULAR_MATRIX_MATRIX_H
+
+namespace Eigen { 
+
+namespace internal {
+
+// template<typename Scalar, int mr, int StorageOrder, bool Conjugate, int Mode>
+// struct gemm_pack_lhs_triangular
+// {
+//   Matrix<Scalar,mr,mr,
+//   void operator()(Scalar* blockA, const EIGEN_RESTRICT Scalar* _lhs, int lhsStride, int depth, int rows)
+//   {
+//     conj_if<NumTraits<Scalar>::IsComplex && Conjugate> cj;
+//     const_blas_data_mapper<Scalar, StorageOrder> lhs(_lhs,lhsStride);
+//     int count = 0;
+//     const int peeled_mc = (rows/mr)*mr;
+//     for(int i=0; i<peeled_mc; i+=mr)
+//     {
+//       for(int k=0; k<depth; k++)
+//         for(int w=0; w<mr; w++)
+//           blockA[count++] = cj(lhs(i+w, k));
+//     }
+//     for(int i=peeled_mc; i<rows; i++)
+//     {
+//       for(int k=0; k<depth; k++)
+//         blockA[count++] = cj(lhs(i, k));
+//     }
+//   }
+// };
+
+/* Optimized triangular matrix * matrix (_TRMM++) product built on top of
+ * the general matrix matrix product.
+ */
+template <typename Scalar, typename Index,
+          int Mode, bool LhsIsTriangular,
+          int LhsStorageOrder, bool ConjugateLhs,
+          int RhsStorageOrder, bool ConjugateRhs,
+          int ResStorageOrder, int Version = Specialized>
+struct product_triangular_matrix_matrix;
+
+template <typename Scalar, typename Index,
+          int Mode, bool LhsIsTriangular,
+          int LhsStorageOrder, bool ConjugateLhs,
+          int RhsStorageOrder, bool ConjugateRhs, int Version>
+struct product_triangular_matrix_matrix<Scalar,Index,Mode,LhsIsTriangular,
+                                           LhsStorageOrder,ConjugateLhs,
+                                           RhsStorageOrder,ConjugateRhs,RowMajor,Version>
+{
+  static EIGEN_STRONG_INLINE void run(
+    Index rows, Index cols, Index depth,
+    const Scalar* lhs, Index lhsStride,
+    const Scalar* rhs, Index rhsStride,
+    Scalar* res,       Index resStride,
+    const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking)
+  {
+    product_triangular_matrix_matrix<Scalar, Index,
+      (Mode&(UnitDiag|ZeroDiag)) | ((Mode&Upper) ? Lower : Upper),
+      (!LhsIsTriangular),
+      RhsStorageOrder==RowMajor ? ColMajor : RowMajor,
+      ConjugateRhs,
+      LhsStorageOrder==RowMajor ? ColMajor : RowMajor,
+      ConjugateLhs,
+      ColMajor>
+      ::run(cols, rows, depth, rhs, rhsStride, lhs, lhsStride, res, resStride, alpha, blocking);
+  }
+};
+
+// implements col-major += alpha * op(triangular) * op(general)
+template <typename Scalar, typename Index, int Mode,
+          int LhsStorageOrder, bool ConjugateLhs,
+          int RhsStorageOrder, bool ConjugateRhs, int Version>
+struct product_triangular_matrix_matrix<Scalar,Index,Mode,true,
+                                           LhsStorageOrder,ConjugateLhs,
+                                           RhsStorageOrder,ConjugateRhs,ColMajor,Version>
+{
+  
+  typedef gebp_traits<Scalar,Scalar> Traits;
+  enum {
+    SmallPanelWidth   = 2 * EIGEN_PLAIN_ENUM_MAX(Traits::mr,Traits::nr),
+    IsLower = (Mode&Lower) == Lower,
+    SetDiag = (Mode&(ZeroDiag|UnitDiag)) ? 0 : 1
+  };
+
+  static EIGEN_DONT_INLINE void run(
+    Index _rows, Index _cols, Index _depth,
+    const Scalar* _lhs, Index lhsStride,
+    const Scalar* _rhs, Index rhsStride,
+    Scalar* res,        Index resStride,
+    const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking);
+};
+
+template <typename Scalar, typename Index, int Mode,
+          int LhsStorageOrder, bool ConjugateLhs,
+          int RhsStorageOrder, bool ConjugateRhs, int Version>
+EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
+                                                        LhsStorageOrder,ConjugateLhs,
+                                                        RhsStorageOrder,ConjugateRhs,ColMajor,Version>::run(
+    Index _rows, Index _cols, Index _depth,
+    const Scalar* _lhs, Index lhsStride,
+    const Scalar* _rhs, Index rhsStride,
+    Scalar* _res,        Index resStride,
+    const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking)
+  {
+    // strip zeros
+    Index diagSize  = (std::min)(_rows,_depth);
+    Index rows      = IsLower ? _rows : diagSize;
+    Index depth     = IsLower ? diagSize : _depth;
+    Index cols      = _cols;
+    
+    typedef const_blas_data_mapper<Scalar, Index, LhsStorageOrder> LhsMapper;
+    typedef const_blas_data_mapper<Scalar, Index, RhsStorageOrder> RhsMapper;
+    typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper;
+    LhsMapper lhs(_lhs,lhsStride);
+    RhsMapper rhs(_rhs,rhsStride);
+    ResMapper res(_res, resStride);
+
+    Index kc = blocking.kc();                   // cache block size along the K direction
+    Index mc = (std::min)(rows,blocking.mc());  // cache block size along the M direction
+    // The small panel size must not be larger than blocking size.
+    // Usually this should never be the case because SmallPanelWidth^2 is very small
+    // compared to L2 cache size, but let's be safe:
+    Index panelWidth = (std::min)(Index(SmallPanelWidth),(std::min)(kc,mc));
+
+    std::size_t sizeA = kc*mc;
+    std::size_t sizeB = kc*cols;
+
+    ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA());
+    ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
+
+    // To work around an "error: member reference base type 'Matrix<...>
+    // (Eigen::internal::constructor_without_unaligned_array_assert (*)())' is
+    // not a structure or union" compilation error in nvcc (tested V8.0.61),
+    // create a dummy internal::constructor_without_unaligned_array_assert
+    // object to pass to the Matrix constructor.
+    internal::constructor_without_unaligned_array_assert a;
+    Matrix<Scalar,SmallPanelWidth,SmallPanelWidth,LhsStorageOrder> triangularBuffer(a);
+    triangularBuffer.setZero();
+    if((Mode&ZeroDiag)==ZeroDiag)
+      triangularBuffer.diagonal().setZero();
+    else
+      triangularBuffer.diagonal().setOnes();
+
+    gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
+    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+    gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder> pack_rhs;
+
+    for(Index k2=IsLower ? depth : 0;
+        IsLower ? k2>0 : k2<depth;
+        IsLower ? k2-=kc : k2+=kc)
+    {
+      Index actual_kc = (std::min)(IsLower ? k2 : depth-k2, kc);
+      Index actual_k2 = IsLower ? k2-actual_kc : k2;
+
+      // align blocks with the end of the triangular part for trapezoidal lhs
+      if((!IsLower)&&(k2<rows)&&(k2+actual_kc>rows))
+      {
+        actual_kc = rows-k2;
+        k2 = k2+actual_kc-kc;
+      }
+
+      pack_rhs(blockB, rhs.getSubMapper(actual_k2,0), actual_kc, cols);
+
+      // the selected lhs's panel has to be split in three different parts:
+      //  1 - the part which is zero => skip it
+      //  2 - the diagonal block => special kernel
+      //  3 - the dense panel below (lower case) or above (upper case) the diagonal block => GEPP
+
+      // the block diagonal, if any:
+      if(IsLower || actual_k2<rows)
+      {
+        // for each small vertical panels of lhs
+        for (Index k1=0; k1<actual_kc; k1+=panelWidth)
+        {
+          Index actualPanelWidth = std::min<Index>(actual_kc-k1, panelWidth);
+          Index lengthTarget = IsLower ? actual_kc-k1-actualPanelWidth : k1;
+          Index startBlock   = actual_k2+k1;
+          Index blockBOffset = k1;
+
+          // => GEBP with the micro triangular block
+          // The trick is to pack this micro block while filling the opposite triangular part with zeros.
+          // To this end we do an extra triangular copy to a small temporary buffer
+          for (Index k=0;k<actualPanelWidth;++k)
+          {
+            if (SetDiag)
+              triangularBuffer.coeffRef(k,k) = lhs(startBlock+k,startBlock+k);
+            for (Index i=IsLower ? k+1 : 0; IsLower ? i<actualPanelWidth : i<k; ++i)
+              triangularBuffer.coeffRef(i,k) = lhs(startBlock+i,startBlock+k);
+          }
+          pack_lhs(blockA, LhsMapper(triangularBuffer.data(), triangularBuffer.outerStride()), actualPanelWidth, actualPanelWidth);
+
+          gebp_kernel(res.getSubMapper(startBlock, 0), blockA, blockB,
+                      actualPanelWidth, actualPanelWidth, cols, alpha,
+                      actualPanelWidth, actual_kc, 0, blockBOffset);
+
+          // GEBP with remaining micro panel
+          if (lengthTarget>0)
+          {
+            Index startTarget  = IsLower ? actual_k2+k1+actualPanelWidth : actual_k2;
+
+            pack_lhs(blockA, lhs.getSubMapper(startTarget,startBlock), actualPanelWidth, lengthTarget);
+
+            gebp_kernel(res.getSubMapper(startTarget, 0), blockA, blockB,
+                        lengthTarget, actualPanelWidth, cols, alpha,
+                        actualPanelWidth, actual_kc, 0, blockBOffset);
+          }
+        }
+      }
+      // the part below (lower case) or above (upper case) the diagonal => GEPP
+      {
+        Index start = IsLower ? k2 : 0;
+        Index end   = IsLower ? rows : (std::min)(actual_k2,rows);
+        for(Index i2=start; i2<end; i2+=mc)
+        {
+          const Index actual_mc = (std::min)(i2+mc,end)-i2;
+          gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr,Traits::LhsProgress, LhsStorageOrder,false>()
+            (blockA, lhs.getSubMapper(i2, actual_k2), actual_kc, actual_mc);
+
+          gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc,
+                      actual_kc, cols, alpha, -1, -1, 0, 0);
+        }
+      }
+    }
+  }
+
+// implements col-major += alpha * op(general) * op(triangular)
+template <typename Scalar, typename Index, int Mode,
+          int LhsStorageOrder, bool ConjugateLhs,
+          int RhsStorageOrder, bool ConjugateRhs, int Version>
+struct product_triangular_matrix_matrix<Scalar,Index,Mode,false,
+                                        LhsStorageOrder,ConjugateLhs,
+                                        RhsStorageOrder,ConjugateRhs,ColMajor,Version>
+{
+  typedef gebp_traits<Scalar,Scalar> Traits;
+  enum {
+    SmallPanelWidth   = EIGEN_PLAIN_ENUM_MAX(Traits::mr,Traits::nr),
+    IsLower = (Mode&Lower) == Lower,
+    SetDiag = (Mode&(ZeroDiag|UnitDiag)) ? 0 : 1
+  };
+
+  static EIGEN_DONT_INLINE void run(
+    Index _rows, Index _cols, Index _depth,
+    const Scalar* _lhs, Index lhsStride,
+    const Scalar* _rhs, Index rhsStride,
+    Scalar* res,        Index resStride,
+    const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking);
+};
+
+template <typename Scalar, typename Index, int Mode,
+          int LhsStorageOrder, bool ConjugateLhs,
+          int RhsStorageOrder, bool ConjugateRhs, int Version>
+EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
+                                                        LhsStorageOrder,ConjugateLhs,
+                                                        RhsStorageOrder,ConjugateRhs,ColMajor,Version>::run(
+    Index _rows, Index _cols, Index _depth,
+    const Scalar* _lhs, Index lhsStride,
+    const Scalar* _rhs, Index rhsStride,
+    Scalar* _res,        Index resStride,
+    const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking)
+  {
+    const Index PacketBytes = packet_traits<Scalar>::size*sizeof(Scalar);
+    // strip zeros
+    Index diagSize  = (std::min)(_cols,_depth);
+    Index rows      = _rows;
+    Index depth     = IsLower ? _depth : diagSize;
+    Index cols      = IsLower ? diagSize : _cols;
+    
+    typedef const_blas_data_mapper<Scalar, Index, LhsStorageOrder> LhsMapper;
+    typedef const_blas_data_mapper<Scalar, Index, RhsStorageOrder> RhsMapper;
+    typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper;
+    LhsMapper lhs(_lhs,lhsStride);
+    RhsMapper rhs(_rhs,rhsStride);
+    ResMapper res(_res, resStride);
+
+    Index kc = blocking.kc();                   // cache block size along the K direction
+    Index mc = (std::min)(rows,blocking.mc());  // cache block size along the M direction
+
+    std::size_t sizeA = kc*mc;
+    std::size_t sizeB = kc*cols+EIGEN_MAX_ALIGN_BYTES/sizeof(Scalar);
+
+    ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA());
+    ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
+
+    internal::constructor_without_unaligned_array_assert a;
+    Matrix<Scalar,SmallPanelWidth,SmallPanelWidth,RhsStorageOrder> triangularBuffer(a);
+    triangularBuffer.setZero();
+    if((Mode&ZeroDiag)==ZeroDiag)
+      triangularBuffer.diagonal().setZero();
+    else
+      triangularBuffer.diagonal().setOnes();
+
+    gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
+    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+    gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder> pack_rhs;
+    gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder,false,true> pack_rhs_panel;
+
+    for(Index k2=IsLower ? 0 : depth;
+        IsLower ? k2<depth  : k2>0;
+        IsLower ? k2+=kc   : k2-=kc)
+    {
+      Index actual_kc = (std::min)(IsLower ? depth-k2 : k2, kc);
+      Index actual_k2 = IsLower ? k2 : k2-actual_kc;
+
+      // align blocks with the end of the triangular part for trapezoidal rhs
+      if(IsLower && (k2<cols) && (actual_k2+actual_kc>cols))
+      {
+        actual_kc = cols-k2;
+        k2 = actual_k2 + actual_kc - kc;
+      }
+
+      // remaining size
+      Index rs = IsLower ? (std::min)(cols,actual_k2) : cols - k2;
+      // size of the triangular part
+      Index ts = (IsLower && actual_k2>=cols) ? 0 : actual_kc;
+
+      Scalar* geb = blockB+ts*ts;
+      geb = geb + internal::first_aligned<PacketBytes>(geb,PacketBytes/sizeof(Scalar));
+
+      pack_rhs(geb, rhs.getSubMapper(actual_k2,IsLower ? 0 : k2), actual_kc, rs);
+
+      // pack the triangular part of the rhs padding the unrolled blocks with zeros
+      if(ts>0)
+      {
+        for (Index j2=0; j2<actual_kc; j2+=SmallPanelWidth)
+        {
+          Index actualPanelWidth = std::min<Index>(actual_kc-j2, SmallPanelWidth);
+          Index actual_j2 = actual_k2 + j2;
+          Index panelOffset = IsLower ? j2+actualPanelWidth : 0;
+          Index panelLength = IsLower ? actual_kc-j2-actualPanelWidth : j2;
+          // general part
+          pack_rhs_panel(blockB+j2*actual_kc,
+                         rhs.getSubMapper(actual_k2+panelOffset, actual_j2),
+                         panelLength, actualPanelWidth,
+                         actual_kc, panelOffset);
+
+          // append the triangular part via a temporary buffer
+          for (Index j=0;j<actualPanelWidth;++j)
+          {
+            if (SetDiag)
+              triangularBuffer.coeffRef(j,j) = rhs(actual_j2+j,actual_j2+j);
+            for (Index k=IsLower ? j+1 : 0; IsLower ? k<actualPanelWidth : k<j; ++k)
+              triangularBuffer.coeffRef(k,j) = rhs(actual_j2+k,actual_j2+j);
+          }
+
+          pack_rhs_panel(blockB+j2*actual_kc,
+                         RhsMapper(triangularBuffer.data(), triangularBuffer.outerStride()),
+                         actualPanelWidth, actualPanelWidth,
+                         actual_kc, j2);
+        }
+      }
+
+      for (Index i2=0; i2<rows; i2+=mc)
+      {
+        const Index actual_mc = (std::min)(mc,rows-i2);
+        pack_lhs(blockA, lhs.getSubMapper(i2, actual_k2), actual_kc, actual_mc);
+
+        // triangular kernel
+        if(ts>0)
+        {
+          for (Index j2=0; j2<actual_kc; j2+=SmallPanelWidth)
+          {
+            Index actualPanelWidth = std::min<Index>(actual_kc-j2, SmallPanelWidth);
+            Index panelLength = IsLower ? actual_kc-j2 : j2+actualPanelWidth;
+            Index blockOffset = IsLower ? j2 : 0;
+
+            gebp_kernel(res.getSubMapper(i2, actual_k2 + j2),
+                        blockA, blockB+j2*actual_kc,
+                        actual_mc, panelLength, actualPanelWidth,
+                        alpha,
+                        actual_kc, actual_kc,  // strides
+                        blockOffset, blockOffset);// offsets
+          }
+        }
+        gebp_kernel(res.getSubMapper(i2, IsLower ? 0 : k2),
+                    blockA, geb, actual_mc, actual_kc, rs,
+                    alpha,
+                    -1, -1, 0, 0);
+      }
+    }
+  }
+
+/***************************************************************************
+* Wrapper to product_triangular_matrix_matrix
+***************************************************************************/
+
+} // end namespace internal
+
+namespace internal {
+template<int Mode, bool LhsIsTriangular, typename Lhs, typename Rhs>
+struct triangular_product_impl<Mode,LhsIsTriangular,Lhs,false,Rhs,false>
+{
+  template<typename Dest> static void run(Dest& dst, const Lhs &a_lhs, const Rhs &a_rhs, const typename Dest::Scalar& alpha)
+  {
+    typedef typename Lhs::Scalar  LhsScalar;
+    typedef typename Rhs::Scalar  RhsScalar;
+    typedef typename Dest::Scalar Scalar;
+    
+    typedef internal::blas_traits<Lhs> LhsBlasTraits;
+    typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
+    typedef typename internal::remove_all<ActualLhsType>::type ActualLhsTypeCleaned;
+    typedef internal::blas_traits<Rhs> RhsBlasTraits;
+    typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
+    typedef typename internal::remove_all<ActualRhsType>::type ActualRhsTypeCleaned;
+    
+    typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(a_lhs);
+    typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(a_rhs);
+
+    LhsScalar lhs_alpha = LhsBlasTraits::extractScalarFactor(a_lhs);
+    RhsScalar rhs_alpha = RhsBlasTraits::extractScalarFactor(a_rhs);
+    Scalar actualAlpha = alpha * lhs_alpha * rhs_alpha;
+
+    typedef internal::gemm_blocking_space<(Dest::Flags&RowMajorBit) ? RowMajor : ColMajor,Scalar,Scalar,
+              Lhs::MaxRowsAtCompileTime, Rhs::MaxColsAtCompileTime, Lhs::MaxColsAtCompileTime,4> BlockingType;
+
+    enum { IsLower = (Mode&Lower) == Lower };
+    Index stripedRows  = ((!LhsIsTriangular) || (IsLower))  ? lhs.rows() : (std::min)(lhs.rows(),lhs.cols());
+    Index stripedCols  = ((LhsIsTriangular)  || (!IsLower)) ? rhs.cols() : (std::min)(rhs.cols(),rhs.rows());
+    Index stripedDepth = LhsIsTriangular ? ((!IsLower) ? lhs.cols() : (std::min)(lhs.cols(),lhs.rows()))
+                                         : ((IsLower)  ? rhs.rows() : (std::min)(rhs.rows(),rhs.cols()));
+
+    BlockingType blocking(stripedRows, stripedCols, stripedDepth, 1, false);
+
+    internal::product_triangular_matrix_matrix<Scalar, Index,
+      Mode, LhsIsTriangular,
+      (internal::traits<ActualLhsTypeCleaned>::Flags&RowMajorBit) ? RowMajor : ColMajor, LhsBlasTraits::NeedToConjugate,
+      (internal::traits<ActualRhsTypeCleaned>::Flags&RowMajorBit) ? RowMajor : ColMajor, RhsBlasTraits::NeedToConjugate,
+      (internal::traits<Dest          >::Flags&RowMajorBit) ? RowMajor : ColMajor>
+      ::run(
+        stripedRows, stripedCols, stripedDepth,   // sizes
+        &lhs.coeffRef(0,0), lhs.outerStride(),    // lhs info
+        &rhs.coeffRef(0,0), rhs.outerStride(),    // rhs info
+        &dst.coeffRef(0,0), dst.outerStride(),    // result info
+        actualAlpha, blocking
+      );
+
+    // Apply correction if the diagonal is unit and a scalar factor was nested:
+    if ((Mode&UnitDiag)==UnitDiag)
+    {
+      if (LhsIsTriangular && lhs_alpha!=LhsScalar(1))
+      {
+        Index diagSize = (std::min)(lhs.rows(),lhs.cols());
+        dst.topRows(diagSize) -= ((lhs_alpha-LhsScalar(1))*a_rhs).topRows(diagSize);
+      }
+      else if ((!LhsIsTriangular) && rhs_alpha!=RhsScalar(1))
+      {
+        Index diagSize = (std::min)(rhs.rows(),rhs.cols());
+        dst.leftCols(diagSize) -= (rhs_alpha-RhsScalar(1))*a_lhs.leftCols(diagSize);
+      }
+    }
+  }
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TRIANGULAR_MATRIX_MATRIX_H