ruby-eigen 0.0.9 → 0.0.10.pre1

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

@@ -0,0 +1,427 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 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_GENERAL_MATRIX_MATRIX_H
+#define EIGEN_GENERAL_MATRIX_MATRIX_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template<typename _LhsScalar, typename _RhsScalar> class level3_blocking;
+
+/* Specialization for a row-major destination matrix => simple transposition of the product */
+template<
+  typename Index,
+  typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs,
+  typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs>
+struct general_matrix_matrix_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,RowMajor>
+{
+  typedef typename scalar_product_traits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+  static EIGEN_STRONG_INLINE void run(
+    Index rows, Index cols, Index depth,
+    const LhsScalar* lhs, Index lhsStride,
+    const RhsScalar* rhs, Index rhsStride,
+    ResScalar* res, Index resStride,
+    ResScalar alpha,
+    level3_blocking<RhsScalar,LhsScalar>& blocking,
+    GemmParallelInfo<Index>* info = 0)
+  {
+    // transpose the product such that the result is column major
+    general_matrix_matrix_product<Index,
+      RhsScalar, RhsStorageOrder==RowMajor ? ColMajor : RowMajor, ConjugateRhs,
+      LhsScalar, LhsStorageOrder==RowMajor ? ColMajor : RowMajor, ConjugateLhs,
+      ColMajor>
+    ::run(cols,rows,depth,rhs,rhsStride,lhs,lhsStride,res,resStride,alpha,blocking,info);
+  }
+};
+
+/*  Specialization for a col-major destination matrix
+ *    => Blocking algorithm following Goto's paper */
+template<
+  typename Index,
+  typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs,
+  typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs>
+struct general_matrix_matrix_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,ColMajor>
+{
+
+typedef typename scalar_product_traits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+static void run(Index rows, Index cols, Index depth,
+  const LhsScalar* _lhs, Index lhsStride,
+  const RhsScalar* _rhs, Index rhsStride,
+  ResScalar* res, Index resStride,
+  ResScalar alpha,
+  level3_blocking<LhsScalar,RhsScalar>& blocking,
+  GemmParallelInfo<Index>* info = 0)
+{
+  const_blas_data_mapper<LhsScalar, Index, LhsStorageOrder> lhs(_lhs,lhsStride);
+  const_blas_data_mapper<RhsScalar, Index, RhsStorageOrder> rhs(_rhs,rhsStride);
+
+  typedef gebp_traits<LhsScalar,RhsScalar> Traits;
+
+  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
+  //Index nc = blocking.nc(); // cache block size along the N direction
+
+  gemm_pack_lhs<LhsScalar, Index, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+  gemm_pack_rhs<RhsScalar, Index, Traits::nr, RhsStorageOrder> pack_rhs;
+  gebp_kernel<LhsScalar, RhsScalar, Index, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp;
+
+#ifdef EIGEN_HAS_OPENMP
+  if(info)
+  {
+    // this is the parallel version!
+    Index tid = omp_get_thread_num();
+    Index threads = omp_get_num_threads();
+    
+    std::size_t sizeA = kc*mc;
+    std::size_t sizeW = kc*Traits::WorkSpaceFactor;
+    ei_declare_aligned_stack_constructed_variable(LhsScalar, blockA, sizeA, 0);
+    ei_declare_aligned_stack_constructed_variable(RhsScalar, w, sizeW, 0);
+    
+    RhsScalar* blockB = blocking.blockB();
+    eigen_internal_assert(blockB!=0);
+
+    // For each horizontal panel of the rhs, and corresponding vertical panel of the lhs...
+    for(Index k=0; k<depth; k+=kc)
+    {
+      const Index actual_kc = (std::min)(k+kc,depth)-k; // => rows of B', and cols of the A'
+
+      // In order to reduce the chance that a thread has to wait for the other,
+      // let's start by packing A'.
+      pack_lhs(blockA, &lhs(0,k), lhsStride, actual_kc, mc);
+
+      // Pack B_k to B' in a parallel fashion:
+      // each thread packs the sub block B_k,j to B'_j where j is the thread id.
+
+      // However, before copying to B'_j, we have to make sure that no other thread is still using it,
+      // i.e., we test that info[tid].users equals 0.
+      // Then, we set info[tid].users to the number of threads to mark that all other threads are going to use it.
+      while(info[tid].users!=0) {}
+      info[tid].users += threads;
+
+      pack_rhs(blockB+info[tid].rhs_start*actual_kc, &rhs(k,info[tid].rhs_start), rhsStride, actual_kc, info[tid].rhs_length);
+
+      // Notify the other threads that the part B'_j is ready to go.
+      info[tid].sync = k;
+
+      // Computes C_i += A' * B' per B'_j
+      for(Index shift=0; shift<threads; ++shift)
+      {
+        Index j = (tid+shift)%threads;
+
+        // At this point we have to make sure that B'_j has been updated by the thread j,
+        // we use testAndSetOrdered to mimic a volatile access.
+        // However, no need to wait for the B' part which has been updated by the current thread!
+        if(shift>0)
+          while(info[j].sync!=k) {}
+
+        gebp(res+info[j].rhs_start*resStride, resStride, blockA, blockB+info[j].rhs_start*actual_kc, mc, actual_kc, info[j].rhs_length, alpha, -1,-1,0,0, w);
+      }
+
+      // Then keep going as usual with the remaining A'
+      for(Index i=mc; i<rows; i+=mc)
+      {
+        const Index actual_mc = (std::min)(i+mc,rows)-i;
+
+        // pack A_i,k to A'
+        pack_lhs(blockA, &lhs(i,k), lhsStride, actual_kc, actual_mc);
+
+        // C_i += A' * B'
+        gebp(res+i, resStride, blockA, blockB, actual_mc, actual_kc, cols, alpha, -1,-1,0,0, w);
+      }
+
+      // Release all the sub blocks B'_j of B' for the current thread,
+      // i.e., we simply decrement the number of users by 1
+      for(Index j=0; j<threads; ++j)
+        #pragma omp atomic
+        --(info[j].users);
+    }
+  }
+  else
+#endif // EIGEN_HAS_OPENMP
+  {
+    EIGEN_UNUSED_VARIABLE(info);
+
+    // this is the sequential version!
+    std::size_t sizeA = kc*mc;
+    std::size_t sizeB = kc*cols;
+    std::size_t sizeW = kc*Traits::WorkSpaceFactor;
+
+    ei_declare_aligned_stack_constructed_variable(LhsScalar, blockA, sizeA, blocking.blockA());
+    ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, sizeB, blocking.blockB());
+    ei_declare_aligned_stack_constructed_variable(RhsScalar, blockW, sizeW, blocking.blockW());
+
+    // For each horizontal panel of the rhs, and corresponding panel of the lhs...
+    // (==GEMM_VAR1)
+    for(Index k2=0; k2<depth; k2+=kc)
+    {
+      const Index actual_kc = (std::min)(k2+kc,depth)-k2;
+
+      // OK, here we have selected one horizontal panel of rhs and one vertical panel of lhs.
+      // => Pack rhs's panel into a sequential chunk of memory (L2 caching)
+      // Note that this panel will be read as many times as the number of blocks in the lhs's
+      // vertical panel which is, in practice, a very low number.
+      pack_rhs(blockB, &rhs(k2,0), rhsStride, actual_kc, cols);
+
+      // For each mc x kc block of the lhs's vertical panel...
+      // (==GEPP_VAR1)
+      for(Index i2=0; i2<rows; i2+=mc)
+      {
+        const Index actual_mc = (std::min)(i2+mc,rows)-i2;
+
+        // We pack the lhs's block into a sequential chunk of memory (L1 caching)
+        // Note that this block will be read a very high number of times, which is equal to the number of
+        // micro vertical panel of the large rhs's panel (e.g., cols/4 times).
+        pack_lhs(blockA, &lhs(i2,k2), lhsStride, actual_kc, actual_mc);
+
+        // Everything is packed, we can now call the block * panel kernel:
+        gebp(res+i2, resStride, blockA, blockB, actual_mc, actual_kc, cols, alpha, -1, -1, 0, 0, blockW);
+      }
+    }
+  }
+}
+
+};
+
+/*********************************************************************************
+*  Specialization of GeneralProduct<> for "large" GEMM, i.e.,
+*  implementation of the high level wrapper to general_matrix_matrix_product
+**********************************************************************************/
+
+template<typename Lhs, typename Rhs>
+struct traits<GeneralProduct<Lhs,Rhs,GemmProduct> >
+ : traits<ProductBase<GeneralProduct<Lhs,Rhs,GemmProduct>, Lhs, Rhs> >
+{};
+
+template<typename Scalar, typename Index, typename Gemm, typename Lhs, typename Rhs, typename Dest, typename BlockingType>
+struct gemm_functor
+{
+  gemm_functor(const Lhs& lhs, const Rhs& rhs, Dest& dest, const Scalar& actualAlpha,
+                  BlockingType& blocking)
+    : m_lhs(lhs), m_rhs(rhs), m_dest(dest), m_actualAlpha(actualAlpha), m_blocking(blocking)
+  {}
+
+  void initParallelSession() const
+  {
+    m_blocking.allocateB();
+  }
+
+  void operator() (Index row, Index rows, Index col=0, Index cols=-1, GemmParallelInfo<Index>* info=0) const
+  {
+    if(cols==-1)
+      cols = m_rhs.cols();
+
+    Gemm::run(rows, cols, m_lhs.cols(),
+              /*(const Scalar*)*/&m_lhs.coeffRef(row,0), m_lhs.outerStride(),
+              /*(const Scalar*)*/&m_rhs.coeffRef(0,col), m_rhs.outerStride(),
+              (Scalar*)&(m_dest.coeffRef(row,col)), m_dest.outerStride(),
+              m_actualAlpha, m_blocking, info);
+  }
+
+  protected:
+    const Lhs& m_lhs;
+    const Rhs& m_rhs;
+    Dest& m_dest;
+    Scalar m_actualAlpha;
+    BlockingType& m_blocking;
+};
+
+template<int StorageOrder, typename LhsScalar, typename RhsScalar, int MaxRows, int MaxCols, int MaxDepth, int KcFactor=1,
+bool FiniteAtCompileTime = MaxRows!=Dynamic && MaxCols!=Dynamic && MaxDepth != Dynamic> class gemm_blocking_space;
+
+template<typename _LhsScalar, typename _RhsScalar>
+class level3_blocking
+{
+    typedef _LhsScalar LhsScalar;
+    typedef _RhsScalar RhsScalar;
+
+  protected:
+    LhsScalar* m_blockA;
+    RhsScalar* m_blockB;
+    RhsScalar* m_blockW;
+
+    DenseIndex m_mc;
+    DenseIndex m_nc;
+    DenseIndex m_kc;
+
+  public:
+
+    level3_blocking()
+      : m_blockA(0), m_blockB(0), m_blockW(0), m_mc(0), m_nc(0), m_kc(0)
+    {}
+
+    inline DenseIndex mc() const { return m_mc; }
+    inline DenseIndex nc() const { return m_nc; }
+    inline DenseIndex kc() const { return m_kc; }
+
+    inline LhsScalar* blockA() { return m_blockA; }
+    inline RhsScalar* blockB() { return m_blockB; }
+    inline RhsScalar* blockW() { return m_blockW; }
+};
+
+template<int StorageOrder, typename _LhsScalar, typename _RhsScalar, int MaxRows, int MaxCols, int MaxDepth, int KcFactor>
+class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, MaxDepth, KcFactor, true>
+  : public level3_blocking<
+      typename conditional<StorageOrder==RowMajor,_RhsScalar,_LhsScalar>::type,
+      typename conditional<StorageOrder==RowMajor,_LhsScalar,_RhsScalar>::type>
+{
+    enum {
+      Transpose = StorageOrder==RowMajor,
+      ActualRows = Transpose ? MaxCols : MaxRows,
+      ActualCols = Transpose ? MaxRows : MaxCols
+    };
+    typedef typename conditional<Transpose,_RhsScalar,_LhsScalar>::type LhsScalar;
+    typedef typename conditional<Transpose,_LhsScalar,_RhsScalar>::type RhsScalar;
+    typedef gebp_traits<LhsScalar,RhsScalar> Traits;
+    enum {
+      SizeA = ActualRows * MaxDepth,
+      SizeB = ActualCols * MaxDepth,
+      SizeW = MaxDepth * Traits::WorkSpaceFactor
+    };
+
+    EIGEN_ALIGN16 LhsScalar m_staticA[SizeA];
+    EIGEN_ALIGN16 RhsScalar m_staticB[SizeB];
+    EIGEN_ALIGN16 RhsScalar m_staticW[SizeW];
+
+  public:
+
+    gemm_blocking_space(DenseIndex /*rows*/, DenseIndex /*cols*/, DenseIndex /*depth*/)
+    {
+      this->m_mc = ActualRows;
+      this->m_nc = ActualCols;
+      this->m_kc = MaxDepth;
+      this->m_blockA = m_staticA;
+      this->m_blockB = m_staticB;
+      this->m_blockW = m_staticW;
+    }
+
+    inline void allocateA() {}
+    inline void allocateB() {}
+    inline void allocateW() {}
+    inline void allocateAll() {}
+};
+
+template<int StorageOrder, typename _LhsScalar, typename _RhsScalar, int MaxRows, int MaxCols, int MaxDepth, int KcFactor>
+class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, MaxDepth, KcFactor, false>
+  : public level3_blocking<
+      typename conditional<StorageOrder==RowMajor,_RhsScalar,_LhsScalar>::type,
+      typename conditional<StorageOrder==RowMajor,_LhsScalar,_RhsScalar>::type>
+{
+    enum {
+      Transpose = StorageOrder==RowMajor
+    };
+    typedef typename conditional<Transpose,_RhsScalar,_LhsScalar>::type LhsScalar;
+    typedef typename conditional<Transpose,_LhsScalar,_RhsScalar>::type RhsScalar;
+    typedef gebp_traits<LhsScalar,RhsScalar> Traits;
+
+    DenseIndex m_sizeA;
+    DenseIndex m_sizeB;
+    DenseIndex m_sizeW;
+
+  public:
+
+    gemm_blocking_space(DenseIndex rows, DenseIndex cols, DenseIndex depth)
+    {
+      this->m_mc = Transpose ? cols : rows;
+      this->m_nc = Transpose ? rows : cols;
+      this->m_kc = depth;
+
+      computeProductBlockingSizes<LhsScalar,RhsScalar,KcFactor>(this->m_kc, this->m_mc, this->m_nc);
+      m_sizeA = this->m_mc * this->m_kc;
+      m_sizeB = this->m_kc * this->m_nc;
+      m_sizeW = this->m_kc*Traits::WorkSpaceFactor;
+    }
+
+    void allocateA()
+    {
+      if(this->m_blockA==0)
+        this->m_blockA = aligned_new<LhsScalar>(m_sizeA);
+    }
+
+    void allocateB()
+    {
+      if(this->m_blockB==0)
+        this->m_blockB = aligned_new<RhsScalar>(m_sizeB);
+    }
+
+    void allocateW()
+    {
+      if(this->m_blockW==0)
+        this->m_blockW = aligned_new<RhsScalar>(m_sizeW);
+    }
+
+    void allocateAll()
+    {
+      allocateA();
+      allocateB();
+      allocateW();
+    }
+
+    ~gemm_blocking_space()
+    {
+      aligned_delete(this->m_blockA, m_sizeA);
+      aligned_delete(this->m_blockB, m_sizeB);
+      aligned_delete(this->m_blockW, m_sizeW);
+    }
+};
+
+} // end namespace internal
+
+template<typename Lhs, typename Rhs>
+class GeneralProduct<Lhs, Rhs, GemmProduct>
+  : public ProductBase<GeneralProduct<Lhs,Rhs,GemmProduct>, Lhs, Rhs>
+{
+    enum {
+      MaxDepthAtCompileTime = EIGEN_SIZE_MIN_PREFER_FIXED(Lhs::MaxColsAtCompileTime,Rhs::MaxRowsAtCompileTime)
+    };
+  public:
+    EIGEN_PRODUCT_PUBLIC_INTERFACE(GeneralProduct)
+    
+    typedef typename  Lhs::Scalar LhsScalar;
+    typedef typename  Rhs::Scalar RhsScalar;
+    typedef           Scalar      ResScalar;
+
+    GeneralProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs)
+    {
+      typedef internal::scalar_product_op<LhsScalar,RhsScalar> BinOp;
+      EIGEN_CHECK_BINARY_COMPATIBILIY(BinOp,LhsScalar,RhsScalar);
+    }
+
+    template<typename Dest> void scaleAndAddTo(Dest& dst, const Scalar& alpha) const
+    {
+      eigen_assert(dst.rows()==m_lhs.rows() && dst.cols()==m_rhs.cols());
+
+      typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(m_lhs);
+      typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(m_rhs);
+
+      Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(m_lhs)
+                                 * RhsBlasTraits::extractScalarFactor(m_rhs);
+
+      typedef internal::gemm_blocking_space<(Dest::Flags&RowMajorBit) ? RowMajor : ColMajor,LhsScalar,RhsScalar,
+              Dest::MaxRowsAtCompileTime,Dest::MaxColsAtCompileTime,MaxDepthAtCompileTime> BlockingType;
+
+      typedef internal::gemm_functor<
+        Scalar, Index,
+        internal::general_matrix_matrix_product<
+          Index,
+          LhsScalar, (_ActualLhsType::Flags&RowMajorBit) ? RowMajor : ColMajor, bool(LhsBlasTraits::NeedToConjugate),
+          RhsScalar, (_ActualRhsType::Flags&RowMajorBit) ? RowMajor : ColMajor, bool(RhsBlasTraits::NeedToConjugate),
+          (Dest::Flags&RowMajorBit) ? RowMajor : ColMajor>,
+        _ActualLhsType, _ActualRhsType, Dest, BlockingType> GemmFunctor;
+
+      BlockingType blocking(dst.rows(), dst.cols(), lhs.cols());
+
+      internal::parallelize_gemm<(Dest::MaxRowsAtCompileTime>32 || Dest::MaxRowsAtCompileTime==Dynamic)>(GemmFunctor(lhs, rhs, dst, actualAlpha, blocking), this->rows(), this->cols(), Dest::Flags&RowMajorBit);
+    }
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_GENERAL_MATRIX_MATRIX_H

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

@@ -0,0 +1,278 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_H
+#define EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_H
+
+namespace Eigen { 
+
+template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool ConjLhs, bool ConjRhs>
+struct selfadjoint_rank1_update;
+
+namespace internal {
+
+/**********************************************************************
+* This file implements a general A * B product while
+* evaluating only one triangular part of the product.
+* This is more general version of self adjoint product (C += A A^T)
+* as the level 3 SYRK Blas routine.
+**********************************************************************/
+
+// forward declarations (defined at the end of this file)
+template<typename LhsScalar, typename RhsScalar, typename Index, int mr, int nr, bool ConjLhs, bool ConjRhs, int UpLo>
+struct tribb_kernel;
+  
+/* Optimized matrix-matrix product evaluating only one triangular half */
+template <typename Index,
+          typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs,
+          typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs,
+                              int ResStorageOrder, int  UpLo, int Version = Specialized>
+struct general_matrix_matrix_triangular_product;
+
+// as usual if the result is row major => we transpose the product
+template <typename Index, typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs,
+                          typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs, int  UpLo, int Version>
+struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,RowMajor,UpLo,Version>
+{
+  typedef typename scalar_product_traits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+  static EIGEN_STRONG_INLINE void run(Index size, Index depth,const LhsScalar* lhs, Index lhsStride,
+                                      const RhsScalar* rhs, Index rhsStride, ResScalar* res, Index resStride, const ResScalar& alpha)
+  {
+    general_matrix_matrix_triangular_product<Index,
+        RhsScalar, RhsStorageOrder==RowMajor ? ColMajor : RowMajor, ConjugateRhs,
+        LhsScalar, LhsStorageOrder==RowMajor ? ColMajor : RowMajor, ConjugateLhs,
+        ColMajor, UpLo==Lower?Upper:Lower>
+      ::run(size,depth,rhs,rhsStride,lhs,lhsStride,res,resStride,alpha);
+  }
+};
+
+template <typename Index, typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs,
+                          typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs, int  UpLo, int Version>
+struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,ColMajor,UpLo,Version>
+{
+  typedef typename scalar_product_traits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+  static EIGEN_STRONG_INLINE void run(Index size, Index depth,const LhsScalar* _lhs, Index lhsStride,
+                                      const RhsScalar* _rhs, Index rhsStride, ResScalar* res, Index resStride, const ResScalar& alpha)
+  {
+    const_blas_data_mapper<LhsScalar, Index, LhsStorageOrder> lhs(_lhs,lhsStride);
+    const_blas_data_mapper<RhsScalar, Index, RhsStorageOrder> rhs(_rhs,rhsStride);
+
+    typedef gebp_traits<LhsScalar,RhsScalar> Traits;
+
+    Index kc = depth; // cache block size along the K direction
+    Index mc = size;  // cache block size along the M direction
+    Index nc = size;  // cache block size along the N direction
+    computeProductBlockingSizes<LhsScalar,RhsScalar>(kc, mc, nc);
+    // !!! mc must be a multiple of nr:
+    if(mc > Traits::nr)
+      mc = (mc/Traits::nr)*Traits::nr;
+
+    std::size_t sizeW = kc*Traits::WorkSpaceFactor;
+    std::size_t sizeB = sizeW + kc*size;
+    ei_declare_aligned_stack_constructed_variable(LhsScalar, blockA, kc*mc, 0);
+    ei_declare_aligned_stack_constructed_variable(RhsScalar, allocatedBlockB, sizeB, 0);
+    RhsScalar* blockB = allocatedBlockB + sizeW;
+    
+    gemm_pack_lhs<LhsScalar, Index, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+    gemm_pack_rhs<RhsScalar, Index, Traits::nr, RhsStorageOrder> pack_rhs;
+    gebp_kernel <LhsScalar, RhsScalar, Index, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp;
+    tribb_kernel<LhsScalar, RhsScalar, Index, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs, UpLo> sybb;
+
+    for(Index k2=0; k2<depth; k2+=kc)
+    {
+      const Index actual_kc = (std::min)(k2+kc,depth)-k2;
+
+      // note that the actual rhs is the transpose/adjoint of mat
+      pack_rhs(blockB, &rhs(k2,0), rhsStride, actual_kc, size);
+
+      for(Index i2=0; i2<size; i2+=mc)
+      {
+        const Index actual_mc = (std::min)(i2+mc,size)-i2;
+
+        pack_lhs(blockA, &lhs(i2, k2), lhsStride, actual_kc, actual_mc);
+
+        // the selected actual_mc * size panel of res is split into three different part:
+        //  1 - before the diagonal => processed with gebp or skipped
+        //  2 - the actual_mc x actual_mc symmetric block => processed with a special kernel
+        //  3 - after the diagonal => processed with gebp or skipped
+        if (UpLo==Lower)
+          gebp(res+i2, resStride, blockA, blockB, actual_mc, actual_kc, (std::min)(size,i2), alpha,
+               -1, -1, 0, 0, allocatedBlockB);
+
+        sybb(res+resStride*i2 + i2, resStride, blockA, blockB + actual_kc*i2, actual_mc, actual_kc, alpha, allocatedBlockB);
+
+        if (UpLo==Upper)
+        {
+          Index j2 = i2+actual_mc;
+          gebp(res+resStride*j2+i2, resStride, blockA, blockB+actual_kc*j2, actual_mc, actual_kc, (std::max)(Index(0), size-j2), alpha,
+               -1, -1, 0, 0, allocatedBlockB);
+        }
+      }
+    }
+  }
+};
+
+// Optimized packed Block * packed Block product kernel evaluating only one given triangular part
+// This kernel is built on top of the gebp kernel:
+// - the current destination block is processed per panel of actual_mc x BlockSize
+//   where BlockSize is set to the minimal value allowing gebp to be as fast as possible
+// - then, as usual, each panel is split into three parts along the diagonal,
+//   the sub blocks above and below the diagonal are processed as usual,
+//   while the triangular block overlapping the diagonal is evaluated into a
+//   small temporary buffer which is then accumulated into the result using a
+//   triangular traversal.
+template<typename LhsScalar, typename RhsScalar, typename Index, int mr, int nr, bool ConjLhs, bool ConjRhs, int UpLo>
+struct tribb_kernel
+{
+  typedef gebp_traits<LhsScalar,RhsScalar,ConjLhs,ConjRhs> Traits;
+  typedef typename Traits::ResScalar ResScalar;
+  
+  enum {
+    BlockSize  = EIGEN_PLAIN_ENUM_MAX(mr,nr)
+  };
+  void operator()(ResScalar* res, Index resStride, const LhsScalar* blockA, const RhsScalar* blockB, Index size, Index depth, const ResScalar& alpha, RhsScalar* workspace)
+  {
+    gebp_kernel<LhsScalar, RhsScalar, Index, mr, nr, ConjLhs, ConjRhs> gebp_kernel;
+    Matrix<ResScalar,BlockSize,BlockSize,ColMajor> buffer;
+
+    // let's process the block per panel of actual_mc x BlockSize,
+    // again, each is split into three parts, etc.
+    for (Index j=0; j<size; j+=BlockSize)
+    {
+      Index actualBlockSize = std::min<Index>(BlockSize,size - j);
+      const RhsScalar* actual_b = blockB+j*depth;
+
+      if(UpLo==Upper)
+        gebp_kernel(res+j*resStride, resStride, blockA, actual_b, j, depth, actualBlockSize, alpha,
+                    -1, -1, 0, 0, workspace);
+
+      // selfadjoint micro block
+      {
+        Index i = j;
+        buffer.setZero();
+        // 1 - apply the kernel on the temporary buffer
+        gebp_kernel(buffer.data(), BlockSize, blockA+depth*i, actual_b, actualBlockSize, depth, actualBlockSize, alpha,
+                    -1, -1, 0, 0, workspace);
+        // 2 - triangular accumulation
+        for(Index j1=0; j1<actualBlockSize; ++j1)
+        {
+          ResScalar* r = res + (j+j1)*resStride + i;
+          for(Index i1=UpLo==Lower ? j1 : 0;
+              UpLo==Lower ? i1<actualBlockSize : i1<=j1; ++i1)
+            r[i1] += buffer(i1,j1);
+        }
+      }
+
+      if(UpLo==Lower)
+      {
+        Index i = j+actualBlockSize;
+        gebp_kernel(res+j*resStride+i, resStride, blockA+depth*i, actual_b, size-i, depth, actualBlockSize, alpha,
+                    -1, -1, 0, 0, workspace);
+      }
+    }
+  }
+};
+
+} // end namespace internal
+
+// high level API
+
+template<typename MatrixType, typename ProductType, int UpLo, bool IsOuterProduct>
+struct general_product_to_triangular_selector;
+
+
+template<typename MatrixType, typename ProductType, int UpLo>
+struct general_product_to_triangular_selector<MatrixType,ProductType,UpLo,true>
+{
+  static void run(MatrixType& mat, const ProductType& prod, const typename MatrixType::Scalar& alpha)
+  {
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename MatrixType::Index Index;
+    
+    typedef typename internal::remove_all<typename ProductType::LhsNested>::type Lhs;
+    typedef internal::blas_traits<Lhs> LhsBlasTraits;
+    typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhs;
+    typedef typename internal::remove_all<ActualLhs>::type _ActualLhs;
+    typename internal::add_const_on_value_type<ActualLhs>::type actualLhs = LhsBlasTraits::extract(prod.lhs());
+    
+    typedef typename internal::remove_all<typename ProductType::RhsNested>::type Rhs;
+    typedef internal::blas_traits<Rhs> RhsBlasTraits;
+    typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhs;
+    typedef typename internal::remove_all<ActualRhs>::type _ActualRhs;
+    typename internal::add_const_on_value_type<ActualRhs>::type actualRhs = RhsBlasTraits::extract(prod.rhs());
+
+    Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs().derived()) * RhsBlasTraits::extractScalarFactor(prod.rhs().derived());
+
+    enum {
+      StorageOrder = (internal::traits<MatrixType>::Flags&RowMajorBit) ? RowMajor : ColMajor,
+      UseLhsDirectly = _ActualLhs::InnerStrideAtCompileTime==1,
+      UseRhsDirectly = _ActualRhs::InnerStrideAtCompileTime==1
+    };
+    
+    internal::gemv_static_vector_if<Scalar,Lhs::SizeAtCompileTime,Lhs::MaxSizeAtCompileTime,!UseLhsDirectly> static_lhs;
+    ei_declare_aligned_stack_constructed_variable(Scalar, actualLhsPtr, actualLhs.size(),
+      (UseLhsDirectly ? const_cast<Scalar*>(actualLhs.data()) : static_lhs.data()));
+    if(!UseLhsDirectly) Map<typename _ActualLhs::PlainObject>(actualLhsPtr, actualLhs.size()) = actualLhs;
+    
+    internal::gemv_static_vector_if<Scalar,Rhs::SizeAtCompileTime,Rhs::MaxSizeAtCompileTime,!UseRhsDirectly> static_rhs;
+    ei_declare_aligned_stack_constructed_variable(Scalar, actualRhsPtr, actualRhs.size(),
+      (UseRhsDirectly ? const_cast<Scalar*>(actualRhs.data()) : static_rhs.data()));
+    if(!UseRhsDirectly) Map<typename _ActualRhs::PlainObject>(actualRhsPtr, actualRhs.size()) = actualRhs;
+    
+    
+    selfadjoint_rank1_update<Scalar,Index,StorageOrder,UpLo,
+                              LhsBlasTraits::NeedToConjugate && NumTraits<Scalar>::IsComplex,
+                              RhsBlasTraits::NeedToConjugate && NumTraits<Scalar>::IsComplex>
+          ::run(actualLhs.size(), mat.data(), mat.outerStride(), actualLhsPtr, actualRhsPtr, actualAlpha);
+  }
+};
+
+template<typename MatrixType, typename ProductType, int UpLo>
+struct general_product_to_triangular_selector<MatrixType,ProductType,UpLo,false>
+{
+  static void run(MatrixType& mat, const ProductType& prod, const typename MatrixType::Scalar& alpha)
+  {
+    typedef typename MatrixType::Index Index;
+    
+    typedef typename internal::remove_all<typename ProductType::LhsNested>::type Lhs;
+    typedef internal::blas_traits<Lhs> LhsBlasTraits;
+    typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhs;
+    typedef typename internal::remove_all<ActualLhs>::type _ActualLhs;
+    typename internal::add_const_on_value_type<ActualLhs>::type actualLhs = LhsBlasTraits::extract(prod.lhs());
+    
+    typedef typename internal::remove_all<typename ProductType::RhsNested>::type Rhs;
+    typedef internal::blas_traits<Rhs> RhsBlasTraits;
+    typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhs;
+    typedef typename internal::remove_all<ActualRhs>::type _ActualRhs;
+    typename internal::add_const_on_value_type<ActualRhs>::type actualRhs = RhsBlasTraits::extract(prod.rhs());
+
+    typename ProductType::Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs().derived()) * RhsBlasTraits::extractScalarFactor(prod.rhs().derived());
+
+    internal::general_matrix_matrix_triangular_product<Index,
+      typename Lhs::Scalar, _ActualLhs::Flags&RowMajorBit ? RowMajor : ColMajor, LhsBlasTraits::NeedToConjugate,
+      typename Rhs::Scalar, _ActualRhs::Flags&RowMajorBit ? RowMajor : ColMajor, RhsBlasTraits::NeedToConjugate,
+      MatrixType::Flags&RowMajorBit ? RowMajor : ColMajor, UpLo>
+      ::run(mat.cols(), actualLhs.cols(),
+            &actualLhs.coeffRef(0,0), actualLhs.outerStride(), &actualRhs.coeffRef(0,0), actualRhs.outerStride(),
+            mat.data(), mat.outerStride(), actualAlpha);
+  }
+};
+
+template<typename MatrixType, unsigned int UpLo>
+template<typename ProductDerived, typename _Lhs, typename _Rhs>
+TriangularView<MatrixType,UpLo>& TriangularView<MatrixType,UpLo>::assignProduct(const ProductBase<ProductDerived, _Lhs,_Rhs>& prod, const Scalar& alpha)
+{
+  general_product_to_triangular_selector<MatrixType, ProductDerived, UpLo, (_Lhs::ColsAtCompileTime==1) || (_Rhs::RowsAtCompileTime==1)>::run(m_matrix.const_cast_derived(), prod.derived(), alpha);
+  
+  return *this;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_H