tomoto 0.1.0

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

@@ -0,0 +1,136 @@
+/*
+ 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
+ *   General matrix-vector product functionality based on ?GEMV.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_GENERAL_MATRIX_VECTOR_BLAS_H
+#define EIGEN_GENERAL_MATRIX_VECTOR_BLAS_H
+
+namespace Eigen { 
+
+namespace internal {
+
+/**********************************************************************
+* This file implements general matrix-vector multiplication using BLAS
+* gemv function via partial specialization of
+* general_matrix_vector_product::run(..) method for float, double,
+* std::complex<float> and std::complex<double> types
+**********************************************************************/
+
+// gemv specialization
+
+template<typename Index, typename LhsScalar, int StorageOrder, bool ConjugateLhs, typename RhsScalar, bool ConjugateRhs>
+struct general_matrix_vector_product_gemv;
+
+#define EIGEN_BLAS_GEMV_SPECIALIZE(Scalar) \
+template<typename Index, bool ConjugateLhs, bool ConjugateRhs> \
+struct general_matrix_vector_product<Index,Scalar,const_blas_data_mapper<Scalar,Index,ColMajor>,ColMajor,ConjugateLhs,Scalar,const_blas_data_mapper<Scalar,Index,RowMajor>,ConjugateRhs,Specialized> { \
+static void run( \
+  Index rows, Index cols, \
+  const const_blas_data_mapper<Scalar,Index,ColMajor> &lhs, \
+  const const_blas_data_mapper<Scalar,Index,RowMajor> &rhs, \
+  Scalar* res, Index resIncr, Scalar alpha) \
+{ \
+  if (ConjugateLhs) { \
+    general_matrix_vector_product<Index,Scalar,const_blas_data_mapper<Scalar,Index,ColMajor>,ColMajor,ConjugateLhs,Scalar,const_blas_data_mapper<Scalar,Index,RowMajor>,ConjugateRhs,BuiltIn>::run( \
+      rows, cols, lhs, rhs, res, resIncr, alpha); \
+  } else { \
+    general_matrix_vector_product_gemv<Index,Scalar,ColMajor,ConjugateLhs,Scalar,ConjugateRhs>::run( \
+      rows, cols, lhs.data(), lhs.stride(), rhs.data(), rhs.stride(), res, resIncr, alpha); \
+  } \
+} \
+}; \
+template<typename Index, bool ConjugateLhs, bool ConjugateRhs> \
+struct general_matrix_vector_product<Index,Scalar,const_blas_data_mapper<Scalar,Index,RowMajor>,RowMajor,ConjugateLhs,Scalar,const_blas_data_mapper<Scalar,Index,ColMajor>,ConjugateRhs,Specialized> { \
+static void run( \
+  Index rows, Index cols, \
+  const const_blas_data_mapper<Scalar,Index,RowMajor> &lhs, \
+  const const_blas_data_mapper<Scalar,Index,ColMajor> &rhs, \
+  Scalar* res, Index resIncr, Scalar alpha) \
+{ \
+    general_matrix_vector_product_gemv<Index,Scalar,RowMajor,ConjugateLhs,Scalar,ConjugateRhs>::run( \
+      rows, cols, lhs.data(), lhs.stride(), rhs.data(), rhs.stride(), res, resIncr, alpha); \
+} \
+}; \
+
+EIGEN_BLAS_GEMV_SPECIALIZE(double)
+EIGEN_BLAS_GEMV_SPECIALIZE(float)
+EIGEN_BLAS_GEMV_SPECIALIZE(dcomplex)
+EIGEN_BLAS_GEMV_SPECIALIZE(scomplex)
+
+#define EIGEN_BLAS_GEMV_SPECIALIZATION(EIGTYPE,BLASTYPE,BLASFUNC) \
+template<typename Index, int LhsStorageOrder, bool ConjugateLhs, bool ConjugateRhs> \
+struct general_matrix_vector_product_gemv<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,ConjugateRhs> \
+{ \
+typedef Matrix<EIGTYPE,Dynamic,1,ColMajor> GEMVVector;\
+\
+static void run( \
+  Index rows, Index cols, \
+  const EIGTYPE* lhs, Index lhsStride, \
+  const EIGTYPE* rhs, Index rhsIncr, \
+  EIGTYPE* res, Index resIncr, EIGTYPE alpha) \
+{ \
+  BlasIndex m=convert_index<BlasIndex>(rows), n=convert_index<BlasIndex>(cols), \
+            lda=convert_index<BlasIndex>(lhsStride), incx=convert_index<BlasIndex>(rhsIncr), incy=convert_index<BlasIndex>(resIncr); \
+  const EIGTYPE beta(1); \
+  const EIGTYPE *x_ptr; \
+  char trans=(LhsStorageOrder==ColMajor) ? 'N' : (ConjugateLhs) ? 'C' : 'T'; \
+  if (LhsStorageOrder==RowMajor) { \
+    m = convert_index<BlasIndex>(cols); \
+    n = convert_index<BlasIndex>(rows); \
+  }\
+  GEMVVector x_tmp; \
+  if (ConjugateRhs) { \
+    Map<const GEMVVector, 0, InnerStride<> > map_x(rhs,cols,1,InnerStride<>(incx)); \
+    x_tmp=map_x.conjugate(); \
+    x_ptr=x_tmp.data(); \
+    incx=1; \
+  } else x_ptr=rhs; \
+  BLASFUNC(&trans, &m, &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_GEMV_SPECIALIZATION(double,   double, dgemv)
+EIGEN_BLAS_GEMV_SPECIALIZATION(float,    float,  sgemv)
+EIGEN_BLAS_GEMV_SPECIALIZATION(dcomplex, MKL_Complex16, zgemv)
+EIGEN_BLAS_GEMV_SPECIALIZATION(scomplex, MKL_Complex8 , cgemv)
+#else
+EIGEN_BLAS_GEMV_SPECIALIZATION(double,   double, dgemv_)
+EIGEN_BLAS_GEMV_SPECIALIZATION(float,    float,  sgemv_)
+EIGEN_BLAS_GEMV_SPECIALIZATION(dcomplex, double, zgemv_)
+EIGEN_BLAS_GEMV_SPECIALIZATION(scomplex, float,  cgemv_)
+#endif
+
+} // end namespase internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_GENERAL_MATRIX_VECTOR_BLAS_H

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

@@ -0,0 +1,163 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 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_PARALLELIZER_H
+#define EIGEN_PARALLELIZER_H
+
+namespace Eigen {
+
+namespace internal {
+
+/** \internal */
+inline void manage_multi_threading(Action action, int* v)
+{
+  static EIGEN_UNUSED int m_maxThreads = -1;
+
+  if(action==SetAction)
+  {
+    eigen_internal_assert(v!=0);
+    m_maxThreads = *v;
+  }
+  else if(action==GetAction)
+  {
+    eigen_internal_assert(v!=0);
+    #ifdef EIGEN_HAS_OPENMP
+    if(m_maxThreads>0)
+      *v = m_maxThreads;
+    else
+      *v = omp_get_max_threads();
+    #else
+    *v = 1;
+    #endif
+  }
+  else
+  {
+    eigen_internal_assert(false);
+  }
+}
+
+}
+
+/** Must be call first when calling Eigen from multiple threads */
+inline void initParallel()
+{
+  int nbt;
+  internal::manage_multi_threading(GetAction, &nbt);
+  std::ptrdiff_t l1, l2, l3;
+  internal::manage_caching_sizes(GetAction, &l1, &l2, &l3);
+}
+
+/** \returns the max number of threads reserved for Eigen
+  * \sa setNbThreads */
+inline int nbThreads()
+{
+  int ret;
+  internal::manage_multi_threading(GetAction, &ret);
+  return ret;
+}
+
+/** Sets the max number of threads reserved for Eigen
+  * \sa nbThreads */
+inline void setNbThreads(int v)
+{
+  internal::manage_multi_threading(SetAction, &v);
+}
+
+namespace internal {
+
+template<typename Index> struct GemmParallelInfo
+{
+  GemmParallelInfo() : sync(-1), users(0), lhs_start(0), lhs_length(0) {}
+
+  Index volatile sync;
+  int volatile users;
+
+  Index lhs_start;
+  Index lhs_length;
+};
+
+template<bool Condition, typename Functor, typename Index>
+void parallelize_gemm(const Functor& func, Index rows, Index cols, Index depth, bool transpose)
+{
+  // TODO when EIGEN_USE_BLAS is defined,
+  // we should still enable OMP for other scalar types
+#if !(defined (EIGEN_HAS_OPENMP)) || defined (EIGEN_USE_BLAS)
+  // FIXME the transpose variable is only needed to properly split
+  // the matrix product when multithreading is enabled. This is a temporary
+  // fix to support row-major destination matrices. This whole
+  // parallelizer mechanism has to be redisigned anyway.
+  EIGEN_UNUSED_VARIABLE(depth);
+  EIGEN_UNUSED_VARIABLE(transpose);
+  func(0,rows, 0,cols);
+#else
+
+  // Dynamically check whether we should enable or disable OpenMP.
+  // The conditions are:
+  // - the max number of threads we can create is greater than 1
+  // - we are not already in a parallel code
+  // - the sizes are large enough
+
+  // compute the maximal number of threads from the size of the product:
+  // This first heuristic takes into account that the product kernel is fully optimized when working with nr columns at once.
+  Index size = transpose ? rows : cols;
+  Index pb_max_threads = std::max<Index>(1,size / Functor::Traits::nr);
+
+  // compute the maximal number of threads from the total amount of work:
+  double work = static_cast<double>(rows) * static_cast<double>(cols) *
+      static_cast<double>(depth);
+  double kMinTaskSize = 50000;  // FIXME improve this heuristic.
+  pb_max_threads = std::max<Index>(1, std::min<Index>(pb_max_threads, work / kMinTaskSize));
+
+  // compute the number of threads we are going to use
+  Index threads = std::min<Index>(nbThreads(), pb_max_threads);
+
+  // if multi-threading is explicitely disabled, not useful, or if we already are in a parallel session,
+  // then abort multi-threading
+  // FIXME omp_get_num_threads()>1 only works for openmp, what if the user does not use openmp?
+  if((!Condition) || (threads==1) || (omp_get_num_threads()>1))
+    return func(0,rows, 0,cols);
+
+  Eigen::initParallel();
+  func.initParallelSession(threads);
+
+  if(transpose)
+    std::swap(rows,cols);
+
+  ei_declare_aligned_stack_constructed_variable(GemmParallelInfo<Index>,info,threads,0);
+
+  #pragma omp parallel num_threads(threads)
+  {
+    Index i = omp_get_thread_num();
+    // Note that the actual number of threads might be lower than the number of request ones.
+    Index actual_threads = omp_get_num_threads();
+
+    Index blockCols = (cols / actual_threads) & ~Index(0x3);
+    Index blockRows = (rows / actual_threads);
+    blockRows = (blockRows/Functor::Traits::mr)*Functor::Traits::mr;
+
+    Index r0 = i*blockRows;
+    Index actualBlockRows = (i+1==actual_threads) ? rows-r0 : blockRows;
+
+    Index c0 = i*blockCols;
+    Index actualBlockCols = (i+1==actual_threads) ? cols-c0 : blockCols;
+
+    info[i].lhs_start = r0;
+    info[i].lhs_length = actualBlockRows;
+
+    if(transpose) func(c0, actualBlockCols, 0, rows, info);
+    else          func(0, rows, c0, actualBlockCols, info);
+  }
+#endif
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_PARALLELIZER_H

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

@@ -0,0 +1,521 @@
+// 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_MATRIX_MATRIX_H
+#define EIGEN_SELFADJOINT_MATRIX_MATRIX_H
+
+namespace Eigen { 
+
+namespace internal {
+
+// pack a selfadjoint block diagonal for use with the gebp_kernel
+template<typename Scalar, typename Index, int Pack1, int Pack2_dummy, int StorageOrder>
+struct symm_pack_lhs
+{
+  template<int BlockRows> inline
+  void pack(Scalar* blockA, const const_blas_data_mapper<Scalar,Index,StorageOrder>& lhs, Index cols, Index i, Index& count)
+  {
+    // normal copy
+    for(Index k=0; k<i; k++)
+      for(Index w=0; w<BlockRows; w++)
+        blockA[count++] = lhs(i+w,k);           // normal
+    // symmetric copy
+    Index h = 0;
+    for(Index k=i; k<i+BlockRows; k++)
+    {
+      for(Index w=0; w<h; w++)
+        blockA[count++] = numext::conj(lhs(k, i+w)); // transposed
+
+      blockA[count++] = numext::real(lhs(k,k));   // real (diagonal)
+
+      for(Index w=h+1; w<BlockRows; w++)
+        blockA[count++] = lhs(i+w, k);          // normal
+      ++h;
+    }
+    // transposed copy
+    for(Index k=i+BlockRows; k<cols; k++)
+      for(Index w=0; w<BlockRows; w++)
+        blockA[count++] = numext::conj(lhs(k, i+w)); // transposed
+  }
+  void operator()(Scalar* blockA, const Scalar* _lhs, Index lhsStride, Index cols, Index rows)
+  {
+    enum { PacketSize = packet_traits<Scalar>::size };
+    const_blas_data_mapper<Scalar,Index,StorageOrder> lhs(_lhs,lhsStride);
+    Index count = 0;
+    //Index peeled_mc3 = (rows/Pack1)*Pack1;
+    
+    const Index peeled_mc3 = Pack1>=3*PacketSize ? (rows/(3*PacketSize))*(3*PacketSize) : 0;
+    const Index peeled_mc2 = Pack1>=2*PacketSize ? peeled_mc3+((rows-peeled_mc3)/(2*PacketSize))*(2*PacketSize) : 0;
+    const Index peeled_mc1 = Pack1>=1*PacketSize ? (rows/(1*PacketSize))*(1*PacketSize) : 0;
+    
+    if(Pack1>=3*PacketSize)
+      for(Index i=0; i<peeled_mc3; i+=3*PacketSize)
+        pack<3*PacketSize>(blockA, lhs, cols, i, count);
+    
+    if(Pack1>=2*PacketSize)
+      for(Index i=peeled_mc3; i<peeled_mc2; i+=2*PacketSize)
+        pack<2*PacketSize>(blockA, lhs, cols, i, count);
+    
+    if(Pack1>=1*PacketSize)
+      for(Index i=peeled_mc2; i<peeled_mc1; i+=1*PacketSize)
+        pack<1*PacketSize>(blockA, lhs, cols, i, count);
+
+    // do the same with mr==1
+    for(Index i=peeled_mc1; i<rows; i++)
+    {
+      for(Index k=0; k<i; k++)
+        blockA[count++] = lhs(i, k);                   // normal
+
+      blockA[count++] = numext::real(lhs(i, i));       // real (diagonal)
+
+      for(Index k=i+1; k<cols; k++)
+        blockA[count++] = numext::conj(lhs(k, i));     // transposed
+    }
+  }
+};
+
+template<typename Scalar, typename Index, int nr, int StorageOrder>
+struct symm_pack_rhs
+{
+  enum { PacketSize = packet_traits<Scalar>::size };
+  void operator()(Scalar* blockB, const Scalar* _rhs, Index rhsStride, Index rows, Index cols, Index k2)
+  {
+    Index end_k = k2 + rows;
+    Index count = 0;
+    const_blas_data_mapper<Scalar,Index,StorageOrder> rhs(_rhs,rhsStride);
+    Index packet_cols8 = nr>=8 ? (cols/8) * 8 : 0;
+    Index packet_cols4 = nr>=4 ? (cols/4) * 4 : 0;
+
+    // first part: normal case
+    for(Index j2=0; j2<k2; j2+=nr)
+    {
+      for(Index k=k2; k<end_k; k++)
+      {
+        blockB[count+0] = rhs(k,j2+0);
+        blockB[count+1] = rhs(k,j2+1);
+        if (nr>=4)
+        {
+          blockB[count+2] = rhs(k,j2+2);
+          blockB[count+3] = rhs(k,j2+3);
+        }
+        if (nr>=8)
+        {
+          blockB[count+4] = rhs(k,j2+4);
+          blockB[count+5] = rhs(k,j2+5);
+          blockB[count+6] = rhs(k,j2+6);
+          blockB[count+7] = rhs(k,j2+7);
+        }
+        count += nr;
+      }
+    }
+
+    // second part: diagonal block
+    Index end8 = nr>=8 ? (std::min)(k2+rows,packet_cols8) : k2;
+    if(nr>=8)
+    {
+      for(Index j2=k2; j2<end8; j2+=8)
+      {
+        // again we can split vertically in three different parts (transpose, symmetric, normal)
+        // transpose
+        for(Index k=k2; k<j2; k++)
+        {
+          blockB[count+0] = numext::conj(rhs(j2+0,k));
+          blockB[count+1] = numext::conj(rhs(j2+1,k));
+          blockB[count+2] = numext::conj(rhs(j2+2,k));
+          blockB[count+3] = numext::conj(rhs(j2+3,k));
+          blockB[count+4] = numext::conj(rhs(j2+4,k));
+          blockB[count+5] = numext::conj(rhs(j2+5,k));
+          blockB[count+6] = numext::conj(rhs(j2+6,k));
+          blockB[count+7] = numext::conj(rhs(j2+7,k));
+          count += 8;
+        }
+        // symmetric
+        Index h = 0;
+        for(Index k=j2; k<j2+8; k++)
+        {
+          // normal
+          for (Index w=0 ; w<h; ++w)
+            blockB[count+w] = rhs(k,j2+w);
+
+          blockB[count+h] = numext::real(rhs(k,k));
+
+          // transpose
+          for (Index w=h+1 ; w<8; ++w)
+            blockB[count+w] = numext::conj(rhs(j2+w,k));
+          count += 8;
+          ++h;
+        }
+        // normal
+        for(Index k=j2+8; k<end_k; k++)
+        {
+          blockB[count+0] = rhs(k,j2+0);
+          blockB[count+1] = rhs(k,j2+1);
+          blockB[count+2] = rhs(k,j2+2);
+          blockB[count+3] = rhs(k,j2+3);
+          blockB[count+4] = rhs(k,j2+4);
+          blockB[count+5] = rhs(k,j2+5);
+          blockB[count+6] = rhs(k,j2+6);
+          blockB[count+7] = rhs(k,j2+7);
+          count += 8;
+        }
+      }
+    }
+    if(nr>=4)
+    {
+      for(Index j2=end8; j2<(std::min)(k2+rows,packet_cols4); j2+=4)
+      {
+        // again we can split vertically in three different parts (transpose, symmetric, normal)
+        // transpose
+        for(Index k=k2; k<j2; k++)
+        {
+          blockB[count+0] = numext::conj(rhs(j2+0,k));
+          blockB[count+1] = numext::conj(rhs(j2+1,k));
+          blockB[count+2] = numext::conj(rhs(j2+2,k));
+          blockB[count+3] = numext::conj(rhs(j2+3,k));
+          count += 4;
+        }
+        // symmetric
+        Index h = 0;
+        for(Index k=j2; k<j2+4; k++)
+        {
+          // normal
+          for (Index w=0 ; w<h; ++w)
+            blockB[count+w] = rhs(k,j2+w);
+
+          blockB[count+h] = numext::real(rhs(k,k));
+
+          // transpose
+          for (Index w=h+1 ; w<4; ++w)
+            blockB[count+w] = numext::conj(rhs(j2+w,k));
+          count += 4;
+          ++h;
+        }
+        // normal
+        for(Index k=j2+4; k<end_k; k++)
+        {
+          blockB[count+0] = rhs(k,j2+0);
+          blockB[count+1] = rhs(k,j2+1);
+          blockB[count+2] = rhs(k,j2+2);
+          blockB[count+3] = rhs(k,j2+3);
+          count += 4;
+        }
+      }
+    }
+
+    // third part: transposed
+    if(nr>=8)
+    {
+      for(Index j2=k2+rows; j2<packet_cols8; j2+=8)
+      {
+        for(Index k=k2; k<end_k; k++)
+        {
+          blockB[count+0] = numext::conj(rhs(j2+0,k));
+          blockB[count+1] = numext::conj(rhs(j2+1,k));
+          blockB[count+2] = numext::conj(rhs(j2+2,k));
+          blockB[count+3] = numext::conj(rhs(j2+3,k));
+          blockB[count+4] = numext::conj(rhs(j2+4,k));
+          blockB[count+5] = numext::conj(rhs(j2+5,k));
+          blockB[count+6] = numext::conj(rhs(j2+6,k));
+          blockB[count+7] = numext::conj(rhs(j2+7,k));
+          count += 8;
+        }
+      }
+    }
+    if(nr>=4)
+    {
+      for(Index j2=(std::max)(packet_cols8,k2+rows); j2<packet_cols4; j2+=4)
+      {
+        for(Index k=k2; k<end_k; k++)
+        {
+          blockB[count+0] = numext::conj(rhs(j2+0,k));
+          blockB[count+1] = numext::conj(rhs(j2+1,k));
+          blockB[count+2] = numext::conj(rhs(j2+2,k));
+          blockB[count+3] = numext::conj(rhs(j2+3,k));
+          count += 4;
+        }
+      }
+    }
+
+    // copy the remaining columns one at a time (=> the same with nr==1)
+    for(Index j2=packet_cols4; j2<cols; ++j2)
+    {
+      // transpose
+      Index half = (std::min)(end_k,j2);
+      for(Index k=k2; k<half; k++)
+      {
+        blockB[count] = numext::conj(rhs(j2,k));
+        count += 1;
+      }
+
+      if(half==j2 && half<k2+rows)
+      {
+        blockB[count] = numext::real(rhs(j2,j2));
+        count += 1;
+      }
+      else
+        half--;
+
+      // normal
+      for(Index k=half+1; k<k2+rows; k++)
+      {
+        blockB[count] = rhs(k,j2);
+        count += 1;
+      }
+    }
+  }
+};
+
+/* Optimized selfadjoint matrix * matrix (_SYMM) product built on top of
+ * the general matrix matrix product.
+ */
+template <typename Scalar, typename Index,
+          int LhsStorageOrder, bool LhsSelfAdjoint, bool ConjugateLhs,
+          int RhsStorageOrder, bool RhsSelfAdjoint, bool ConjugateRhs,
+          int ResStorageOrder>
+struct product_selfadjoint_matrix;
+
+template <typename Scalar, typename Index,
+          int LhsStorageOrder, bool LhsSelfAdjoint, bool ConjugateLhs,
+          int RhsStorageOrder, bool RhsSelfAdjoint, bool ConjugateRhs>
+struct product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,LhsSelfAdjoint,ConjugateLhs, RhsStorageOrder,RhsSelfAdjoint,ConjugateRhs,RowMajor>
+{
+
+  static EIGEN_STRONG_INLINE void run(
+    Index rows, Index cols,
+    const Scalar* lhs, Index lhsStride,
+    const Scalar* rhs, Index rhsStride,
+    Scalar* res,       Index resStride,
+    const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking)
+  {
+    product_selfadjoint_matrix<Scalar, Index,
+      EIGEN_LOGICAL_XOR(RhsSelfAdjoint,RhsStorageOrder==RowMajor) ? ColMajor : RowMajor,
+      RhsSelfAdjoint, NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(RhsSelfAdjoint,ConjugateRhs),
+      EIGEN_LOGICAL_XOR(LhsSelfAdjoint,LhsStorageOrder==RowMajor) ? ColMajor : RowMajor,
+      LhsSelfAdjoint, NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(LhsSelfAdjoint,ConjugateLhs),
+      ColMajor>
+      ::run(cols, rows,  rhs, rhsStride,  lhs, lhsStride,  res, resStride,  alpha, blocking);
+  }
+};
+
+template <typename Scalar, typename Index,
+          int LhsStorageOrder, bool ConjugateLhs,
+          int RhsStorageOrder, bool ConjugateRhs>
+struct product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,true,ConjugateLhs, RhsStorageOrder,false,ConjugateRhs,ColMajor>
+{
+
+  static EIGEN_DONT_INLINE void run(
+    Index rows, Index cols,
+    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 LhsStorageOrder, bool ConjugateLhs,
+          int RhsStorageOrder, bool ConjugateRhs>
+EIGEN_DONT_INLINE void product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,true,ConjugateLhs, RhsStorageOrder,false,ConjugateRhs,ColMajor>::run(
+    Index rows, Index cols,
+    const Scalar* _lhs, Index lhsStride,
+    const Scalar* _rhs, Index rhsStride,
+    Scalar* _res,        Index resStride,
+    const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking)
+  {
+    Index size = rows;
+
+    typedef gebp_traits<Scalar,Scalar> Traits;
+
+    typedef const_blas_data_mapper<Scalar, Index, LhsStorageOrder> LhsMapper;
+    typedef const_blas_data_mapper<Scalar, Index, (LhsStorageOrder == RowMajor) ? ColMajor : RowMajor> LhsTransposeMapper;
+    typedef const_blas_data_mapper<Scalar, Index, RhsStorageOrder> RhsMapper;
+    typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper;
+    LhsMapper lhs(_lhs,lhsStride);
+    LhsTransposeMapper lhs_transpose(_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
+    // kc must be smaller than mc
+    kc = (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());
+
+    gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
+    symm_pack_lhs<Scalar, Index, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+    gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder> pack_rhs;
+    gemm_pack_lhs<Scalar, Index, LhsTransposeMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder==RowMajor?ColMajor:RowMajor, true> pack_lhs_transposed;
+
+    for(Index k2=0; k2<size; k2+=kc)
+    {
+      const Index actual_kc = (std::min)(k2+kc,size)-k2;
+
+      // we have selected one row panel of rhs and one column panel of lhs
+      // pack rhs's panel into a sequential chunk of memory
+      // and expand each coeff to a constant packet for further reuse
+      pack_rhs(blockB, rhs.getSubMapper(k2,0), actual_kc, cols);
+
+      // the select lhs's panel has to be split in three different parts:
+      //  1 - the transposed panel above the diagonal block => transposed packed copy
+      //  2 - the diagonal block => special packed copy
+      //  3 - the panel below the diagonal block => generic packed copy
+      for(Index i2=0; i2<k2; i2+=mc)
+      {
+        const Index actual_mc = (std::min)(i2+mc,k2)-i2;
+        // transposed packed copy
+        pack_lhs_transposed(blockA, lhs_transpose.getSubMapper(i2, k2), actual_kc, actual_mc);
+
+        gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc, cols, alpha);
+      }
+      // the block diagonal
+      {
+        const Index actual_mc = (std::min)(k2+kc,size)-k2;
+        // symmetric packed copy
+        pack_lhs(blockA, &lhs(k2,k2), lhsStride, actual_kc, actual_mc);
+
+        gebp_kernel(res.getSubMapper(k2, 0), blockA, blockB, actual_mc, actual_kc, cols, alpha);
+      }
+
+      for(Index i2=k2+kc; i2<size; i2+=mc)
+      {
+        const Index actual_mc = (std::min)(i2+mc,size)-i2;
+        gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder,false>()
+          (blockA, lhs.getSubMapper(i2, k2), actual_kc, actual_mc);
+
+        gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc, cols, alpha);
+      }
+    }
+  }
+
+// matrix * selfadjoint product
+template <typename Scalar, typename Index,
+          int LhsStorageOrder, bool ConjugateLhs,
+          int RhsStorageOrder, bool ConjugateRhs>
+struct product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,false,ConjugateLhs, RhsStorageOrder,true,ConjugateRhs,ColMajor>
+{
+
+  static EIGEN_DONT_INLINE void run(
+    Index rows, Index cols,
+    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 LhsStorageOrder, bool ConjugateLhs,
+          int RhsStorageOrder, bool ConjugateRhs>
+EIGEN_DONT_INLINE void product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,false,ConjugateLhs, RhsStorageOrder,true,ConjugateRhs,ColMajor>::run(
+    Index rows, Index cols,
+    const Scalar* _lhs, Index lhsStride,
+    const Scalar* _rhs, Index rhsStride,
+    Scalar* _res,        Index resStride,
+    const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking)
+  {
+    Index size = cols;
+
+    typedef gebp_traits<Scalar,Scalar> Traits;
+
+    typedef const_blas_data_mapper<Scalar, Index, LhsStorageOrder> LhsMapper;
+    typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper;
+    LhsMapper lhs(_lhs,lhsStride);
+    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;
+    ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA());
+    ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
+
+    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;
+    symm_pack_rhs<Scalar, Index, Traits::nr,RhsStorageOrder> pack_rhs;
+
+    for(Index k2=0; k2<size; k2+=kc)
+    {
+      const Index actual_kc = (std::min)(k2+kc,size)-k2;
+
+      pack_rhs(blockB, _rhs, rhsStride, actual_kc, cols, k2);
+
+      // => GEPP
+      for(Index i2=0; i2<rows; i2+=mc)
+      {
+        const Index actual_mc = (std::min)(i2+mc,rows)-i2;
+        pack_lhs(blockA, lhs.getSubMapper(i2, k2), actual_kc, actual_mc);
+
+        gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc, cols, alpha);
+      }
+    }
+  }
+
+} // end namespace internal
+
+/***************************************************************************
+* Wrapper to product_selfadjoint_matrix
+***************************************************************************/
+
+namespace internal {
+  
+template<typename Lhs, int LhsMode, typename Rhs, int RhsMode>
+struct selfadjoint_product_impl<Lhs,LhsMode,false,Rhs,RhsMode,false>
+{
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  
+  typedef internal::blas_traits<Lhs> LhsBlasTraits;
+  typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
+  typedef internal::blas_traits<Rhs> RhsBlasTraits;
+  typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
+  
+  enum {
+    LhsIsUpper = (LhsMode&(Upper|Lower))==Upper,
+    LhsIsSelfAdjoint = (LhsMode&SelfAdjoint)==SelfAdjoint,
+    RhsIsUpper = (RhsMode&(Upper|Lower))==Upper,
+    RhsIsSelfAdjoint = (RhsMode&SelfAdjoint)==SelfAdjoint
+  };
+  
+  template<typename Dest>
+  static void run(Dest &dst, const Lhs &a_lhs, const Rhs &a_rhs, const Scalar& alpha)
+  {
+    eigen_assert(dst.rows()==a_lhs.rows() && dst.cols()==a_rhs.cols());
+
+    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);
+
+    Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(a_lhs)
+                               * RhsBlasTraits::extractScalarFactor(a_rhs);
+
+    typedef internal::gemm_blocking_space<(Dest::Flags&RowMajorBit) ? RowMajor : ColMajor,Scalar,Scalar,
+              Lhs::MaxRowsAtCompileTime, Rhs::MaxColsAtCompileTime, Lhs::MaxColsAtCompileTime,1> BlockingType;
+
+    BlockingType blocking(lhs.rows(), rhs.cols(), lhs.cols(), 1, false);
+
+    internal::product_selfadjoint_matrix<Scalar, Index,
+      EIGEN_LOGICAL_XOR(LhsIsUpper,internal::traits<Lhs>::Flags &RowMajorBit) ? RowMajor : ColMajor, LhsIsSelfAdjoint,
+      NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(LhsIsUpper,bool(LhsBlasTraits::NeedToConjugate)),
+      EIGEN_LOGICAL_XOR(RhsIsUpper,internal::traits<Rhs>::Flags &RowMajorBit) ? RowMajor : ColMajor, RhsIsSelfAdjoint,
+      NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(RhsIsUpper,bool(RhsBlasTraits::NeedToConjugate)),
+      internal::traits<Dest>::Flags&RowMajorBit  ? RowMajor : ColMajor>
+      ::run(
+        lhs.rows(), rhs.cols(),                 // 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                   // alpha
+      );
+  }
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SELFADJOINT_MATRIX_MATRIX_H