@smake/eigen 1.0.2 → 1.1.0

package/eigen/Eigen/src/Core/products/GeneralMatrixMatrix.h

@@ -77,7 +77,7 @@ static void run(Index rows, Index cols, Index depth,
   Index mc = (std::min)(rows,blocking.mc());  // cache block size along the M direction
   Index nc = (std::min)(cols,blocking.nc());  // cache block size along the N direction
 
-  gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+  gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, LhsStorageOrder> pack_lhs;
   gemm_pack_rhs<RhsScalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs;
   gebp_kernel<LhsScalar, RhsScalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp;
 
@@ -110,7 +110,7 @@ static void run(Index rows, Index cols, Index depth,
       // 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;
+      info[tid].users = threads;
 
       pack_lhs(blockA+info[tid].lhs_start*actual_kc, lhs.getSubMapper(info[tid].lhs_start,k), actual_kc, info[tid].lhs_length);
 
@@ -148,7 +148,9 @@ static void run(Index rows, Index cols, Index depth,
       // Release all the sub blocks A'_i of A' for the current thread,
       // i.e., we simply decrement the number of users by 1
       for(Index i=0; i<threads; ++i)
+#if !EIGEN_HAS_CXX11_ATOMIC
         #pragma omp atomic
+#endif
         info[i].users -= 1;
     }
   }
@@ -429,7 +431,13 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
   template<typename Dst>
   static void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
-    if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0)
+    // See http://eigen.tuxfamily.org/bz/show_bug.cgi?id=404 for a discussion and helper program
+    // to determine the following heuristic.
+    // EIGEN_GEMM_TO_COEFFBASED_THRESHOLD is typically defined to 20 in GeneralProduct.h,
+    // unless it has been specialized by the user or for a given architecture.
+    // Note that the condition rhs.rows()>0 was required because lazy product is (was?) not happy with empty inputs.
+    // I'm not sure it is still required.
+    if((rhs.rows()+dst.rows()+dst.cols())<EIGEN_GEMM_TO_COEFFBASED_THRESHOLD && rhs.rows()>0)
       lazyproduct::eval_dynamic(dst, lhs, rhs, internal::assign_op<typename Dst::Scalar,Scalar>());
     else
     {
@@ -441,7 +449,7 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
   template<typename Dst>
   static void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
-    if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0)
+    if((rhs.rows()+dst.rows()+dst.cols())<EIGEN_GEMM_TO_COEFFBASED_THRESHOLD && rhs.rows()>0)
       lazyproduct::eval_dynamic(dst, lhs, rhs, internal::add_assign_op<typename Dst::Scalar,Scalar>());
     else
       scaleAndAddTo(dst,lhs, rhs, Scalar(1));
@@ -450,7 +458,7 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
   template<typename Dst>
   static void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
-    if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0)
+    if((rhs.rows()+dst.rows()+dst.cols())<EIGEN_GEMM_TO_COEFFBASED_THRESHOLD && rhs.rows()>0)
       lazyproduct::eval_dynamic(dst, lhs, rhs, internal::sub_assign_op<typename Dst::Scalar,Scalar>());
     else
       scaleAndAddTo(dst, lhs, rhs, Scalar(-1));
@@ -463,11 +471,25 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
     if(a_lhs.cols()==0 || a_lhs.rows()==0 || a_rhs.cols()==0)
       return;
 
+    if (dst.cols() == 1)
+    {
+      // Fallback to GEMV if either the lhs or rhs is a runtime vector
+      typename Dest::ColXpr dst_vec(dst.col(0));
+      return internal::generic_product_impl<Lhs,typename Rhs::ConstColXpr,DenseShape,DenseShape,GemvProduct>
+        ::scaleAndAddTo(dst_vec, a_lhs, a_rhs.col(0), alpha);
+    }
+    else if (dst.rows() == 1)
+    {
+      // Fallback to GEMV if either the lhs or rhs is a runtime vector
+      typename Dest::RowXpr dst_vec(dst.row(0));
+      return internal::generic_product_impl<typename Lhs::ConstRowXpr,Rhs,DenseShape,DenseShape,GemvProduct>
+        ::scaleAndAddTo(dst_vec, a_lhs.row(0), a_rhs, alpha);
+    }
+
     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);
+    Scalar actualAlpha = combine_scalar_factors(alpha, a_lhs, a_rhs);
 
     typedef internal::gemm_blocking_space<(Dest::Flags&RowMajorBit) ? RowMajor : ColMajor,LhsScalar,RhsScalar,
             Dest::MaxRowsAtCompileTime,Dest::MaxColsAtCompileTime,MaxDepthAtCompileTime> BlockingType;

package/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h

@@ -87,7 +87,7 @@ struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,
     ei_declare_aligned_stack_constructed_variable(LhsScalar, blockA, sizeA, blocking.blockA());
     ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, sizeB, blocking.blockB());
 
-    gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+    gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, LhsStorageOrder> pack_lhs;
     gemm_pack_rhs<RhsScalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs;
     gebp_kernel<LhsScalar, RhsScalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp;
     tribb_kernel<LhsScalar, RhsScalar, Index, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs, ResInnerStride, UpLo> sybb;
@@ -302,13 +302,13 @@ struct general_product_to_triangular_selector<MatrixType,ProductType,UpLo,false>
 
 template<typename MatrixType, unsigned int UpLo>
 template<typename ProductType>
-TriangularView<MatrixType,UpLo>& TriangularViewImpl<MatrixType,UpLo,Dense>::_assignProduct(const ProductType& prod, const Scalar& alpha, bool beta)
+EIGEN_DEVICE_FUNC TriangularView<MatrixType,UpLo>& TriangularViewImpl<MatrixType,UpLo,Dense>::_assignProduct(const ProductType& prod, const Scalar& alpha, bool beta)
 {
   EIGEN_STATIC_ASSERT((UpLo&UnitDiag)==0, WRITING_TO_TRIANGULAR_PART_WITH_UNIT_DIAGONAL_IS_NOT_SUPPORTED);
   eigen_assert(derived().nestedExpression().rows() == prod.rows() && derived().cols() == prod.cols());
-  
+
   general_product_to_triangular_selector<MatrixType, ProductType, UpLo, internal::traits<ProductType>::InnerSize==1>::run(derived().nestedExpression().const_cast_derived(), prod, alpha, beta);
-  
+
   return derived();
 }
 

package/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h

@@ -37,7 +37,7 @@ namespace Eigen {
 
 namespace internal {
 
-template <typename Index, typename Scalar, int AStorageOrder, bool ConjugateA, int ResStorageOrder, int  UpLo>
+template <typename Index, typename Scalar, int AStorageOrder, bool ConjugateA, int ResStorageOrder, int UpLo>
 struct general_matrix_matrix_rankupdate :
        general_matrix_matrix_triangular_product<
          Index,Scalar,AStorageOrder,ConjugateA,Scalar,AStorageOrder,ConjugateA,ResStorageOrder,1,UpLo,BuiltIn> {};