// This file is part of Eigen, a lightweight C++ template library // for linear algebra. // // Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com> // Copyright (C) 2008-2011 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/.
// Define the threshold value to fallback from the generic matrix-matrix product // implementation (heavy) to the lightweight coeff-based product one. // See generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct> // in products/GeneralMatrixMatrix.h for more details. // TODO This threshold should also be used in the compile-time selector below. #ifndef EIGEN_GEMM_TO_COEFFBASED_THRESHOLD // This default value has been obtained on a Haswell architecture. #define EIGEN_GEMM_TO_COEFFBASED_THRESHOLD 20 #endif
namespace internal {
template<int Rows, int Cols, int Depth> struct product_type_selector;
// the splitting into different lines of code here, introducing the _select enums and the typedef below, // is to work around an internal compiler error with gcc 4.1 and 4.2. private: enum {
rows_select = product_size_category<Rows,MaxRows>::value,
cols_select = product_size_category<Cols,MaxCols>::value,
depth_select = product_size_category<Depth,MaxDepth>::value
}; typedef product_type_selector<rows_select, cols_select, depth_select> selector;
/* The following allows to select the kind of product at compile time * based on the three dimensions of the product.
* This is a compile time mapping from {1,Small,Large}^3 -> {product types} */ // FIXME I'm not sure the current mapping is the ideal one. template<int M, int N> struct product_type_selector<M,N,1> { enum { ret = OuterProduct }; }; template<int M> struct product_type_selector<M, 1, 1> { enum { ret = LazyCoeffBasedProductMode }; }; template<int N> struct product_type_selector<1, N, 1> { enum { ret = LazyCoeffBasedProductMode }; }; template<int Depth> struct product_type_selector<1, 1, Depth> { enum { ret = InnerProduct }; }; template<> struct product_type_selector<1, 1, 1> { enum { ret = InnerProduct }; }; template<> struct product_type_selector<Small,1, Small> { enum { ret = CoeffBasedProductMode }; }; template<> struct product_type_selector<1, Small,Small> { enum { ret = CoeffBasedProductMode }; }; template<> struct product_type_selector<Small,Small,Small> { enum { ret = CoeffBasedProductMode }; }; template<> struct product_type_selector<Small, Small, 1> { enum { ret = LazyCoeffBasedProductMode }; }; template<> struct product_type_selector<Small, Large, 1> { enum { ret = LazyCoeffBasedProductMode }; }; template<> struct product_type_selector<Large, Small, 1> { enum { ret = LazyCoeffBasedProductMode }; }; template<> struct product_type_selector<1, Large,Small> { enum { ret = CoeffBasedProductMode }; }; template<> struct product_type_selector<1, Large,Large> { enum { ret = GemvProduct }; }; template<> struct product_type_selector<1, Small,Large> { enum { ret = CoeffBasedProductMode }; }; template<> struct product_type_selector<Large,1, Small> { enum { ret = CoeffBasedProductMode }; }; template<> struct product_type_selector<Large,1, Large> { enum { ret = GemvProduct }; }; template<> struct product_type_selector<Small,1, Large> { enum { ret = CoeffBasedProductMode }; }; template<> struct product_type_selector<Small,Small,Large> { enum { ret = GemmProduct }; }; template<> struct product_type_selector<Large,Small,Large> { enum { ret = GemmProduct }; }; template<> struct product_type_selector<Small,Large,Large> { enum { ret = GemmProduct }; }; template<> struct product_type_selector<Large,Large,Large> { enum { ret = GemmProduct }; }; template<> struct product_type_selector<Large,Small,Small> { enum { ret = CoeffBasedProductMode }; }; template<> struct product_type_selector<Small,Large,Small> { enum { ret = CoeffBasedProductMode }; }; template<> struct product_type_selector<Large,Large,Small> { enum { ret = GemmProduct }; };
} // end namespace internal
/*********************************************************************** * Implementation of Inner Vector Vector Product
***********************************************************************/
// FIXME : maybe the "inner product" could return a Scalar // instead of a 1x1 matrix ?? // Pro: more natural for the user // Cons: this could be a problem if in a meta unrolled algorithm a matrix-matrix // product ends up to a row-vector times col-vector product... To tackle this use // case, we could have a specialization for Block<MatrixType,1,1> with: operator=(Scalar x);
/*********************************************************************** * Implementation of Outer Vector Vector Product
***********************************************************************/
/*********************************************************************** * Implementation of General Matrix Vector Product
***********************************************************************/
/* According to the shape/flags of the matrix we have to distinghish 3 different cases: * 1 - the matrix is col-major, BLAS compatible and M is large => call fast BLAS-like colmajor routine * 2 - the matrix is row-major, BLAS compatible and N is large => call fast BLAS-like rowmajor routine * 3 - all other cases are handled using a simple loop along the outer-storage direction. * Therefore we need a lower level meta selector. * Furthermore, if the matrix is the rhs, then the product has to be transposed.
*/ namespace internal {
template<int Side, int StorageOrder, bool BlasCompatible> struct gemv_dense_selector;
// make sure Dest is a compile-time vector type (bug 1166) typedeftypename conditional<Dest::IsVectorAtCompileTime, Dest, typename Dest::ColXpr>::type ActualDest;
enum { // FIXME find a way to allow an inner stride on the result if packet_traits<Scalar>::size==1 // on, the other hand it is good for the cache to pack the vector anyways...
EvalToDestAtCompileTime = (ActualDest::InnerStrideAtCompileTime==1),
ComplexByReal = (NumTraits<LhsScalar>::IsComplex) && (!NumTraits<RhsScalar>::IsComplex),
MightCannotUseDest = ((!EvalToDestAtCompileTime) || ComplexByReal) && (ActualDest::MaxSizeAtCompileTime!=0)
};
if(!MightCannotUseDest)
{ // shortcut if we are sure to be able to use dest directly, // this ease the compiler to generate cleaner and more optimzized code for most common cases
general_matrix_vector_product
<Index,LhsScalar,LhsMapper,ColMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsMapper,RhsBlasTraits::NeedToConjugate>::run(
actualLhs.rows(), actualLhs.cols(),
LhsMapper(actualLhs.data(), actualLhs.outerStride()),
RhsMapper(actualRhs.data(), actualRhs.innerStride()),
dest.data(), 1,
compatibleAlpha);
} else
{
gemv_static_vector_if<ResScalar,ActualDest::SizeAtCompileTime,ActualDest::MaxSizeAtCompileTime,MightCannotUseDest> static_dest;
enum { // FIXME find a way to allow an inner stride on the result if packet_traits<Scalar>::size==1 // on, the other hand it is good for the cache to pack the vector anyways...
DirectlyUseRhs = ActualRhsTypeCleaned::InnerStrideAtCompileTime==1 || ActualRhsTypeCleaned::MaxSizeAtCompileTime==0
};
typedef const_blas_data_mapper<LhsScalar,Index,RowMajor> LhsMapper; typedef const_blas_data_mapper<RhsScalar,Index,ColMajor> RhsMapper;
general_matrix_vector_product
<Index,LhsScalar,LhsMapper,RowMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsMapper,RhsBlasTraits::NeedToConjugate>::run(
actualLhs.rows(), actualLhs.cols(),
LhsMapper(actualLhs.data(), actualLhs.outerStride()),
RhsMapper(actualRhsPtr, 1),
dest.data(), dest.col(0).innerStride(), //NOTE if dest is not a vector at compile-time, then dest.innerStride() might be wrong. (bug 1166)
actualAlpha);
}
};
template<> struct gemv_dense_selector<OnTheRight,ColMajor,false>
{ template<typename Lhs, typename Rhs, typename Dest> staticvoid run(const Lhs &lhs, const Rhs &rhs, Dest& dest, consttypename Dest::Scalar& alpha)
{
EIGEN_STATIC_ASSERT((!nested_eval<Lhs,1>::Evaluate),EIGEN_INTERNAL_COMPILATION_ERROR_OR_YOU_MADE_A_PROGRAMMING_MISTAKE); // TODO if rhs is large enough it might be beneficial to make sure that dest is sequentially stored in memory, otherwise use a temp typename nested_eval<Rhs,1>::type actual_rhs(rhs); const Index size = rhs.rows(); for(Index k=0; k<size; ++k)
dest += (alpha*actual_rhs.coeff(k)) * lhs.col(k);
}
};
/*************************************************************************** * Implementation of matrix base methods
***************************************************************************/
/** \returns the matrix product of \c *this and \a other. * * \note If instead of the matrix product you want the coefficient-wise product, see Cwise::operator*(). * * \sa lazyProduct(), operator*=(const MatrixBase&), Cwise::operator*()
*/ template<typename Derived> template<typename OtherDerived>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Product<Derived, OtherDerived>
MatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const
{ // A note regarding the function declaration: In MSVC, this function will sometimes // not be inlined since DenseStorage is an unwindable object for dynamic // matrices and product types are holding a member to store the result. // Thus it does not help tagging this function with EIGEN_STRONG_INLINE. enum {
ProductIsValid = Derived::ColsAtCompileTime==Dynamic
|| OtherDerived::RowsAtCompileTime==Dynamic
|| int(Derived::ColsAtCompileTime)==int(OtherDerived::RowsAtCompileTime),
AreVectors = Derived::IsVectorAtCompileTime && OtherDerived::IsVectorAtCompileTime,
SameSizes = EIGEN_PREDICATE_SAME_MATRIX_SIZE(Derived,OtherDerived)
}; // note to the lost user: // * for a dot product use: v1.dot(v2) // * for a coeff-wise product use: v1.cwiseProduct(v2)
EIGEN_STATIC_ASSERT(ProductIsValid || !(AreVectors && SameSizes),
INVALID_VECTOR_VECTOR_PRODUCT__IF_YOU_WANTED_A_DOT_OR_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTIONS)
EIGEN_STATIC_ASSERT(ProductIsValid || !(SameSizes && !AreVectors),
INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION)
EIGEN_STATIC_ASSERT(ProductIsValid || SameSizes, INVALID_MATRIX_PRODUCT) #ifdef EIGEN_DEBUG_PRODUCT
internal::product_type<Derived,OtherDerived>::debug(); #endif
/** \returns an expression of the matrix product of \c *this and \a other without implicit evaluation. * * The returned product will behave like any other expressions: the coefficients of the product will be * computed once at a time as requested. This might be useful in some extremely rare cases when only * a small and no coherent fraction of the result's coefficients have to be computed. * * \warning This version of the matrix product can be much much slower. So use it only if you know * what you are doing and that you measured a true speed improvement. * * \sa operator*(const MatrixBase&)
*/ template<typename Derived> template<typename OtherDerived>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Product<Derived,OtherDerived,LazyProduct>
MatrixBase<Derived>::lazyProduct(const MatrixBase<OtherDerived> &other) const
{ enum {
ProductIsValid = Derived::ColsAtCompileTime==Dynamic
|| OtherDerived::RowsAtCompileTime==Dynamic
|| int(Derived::ColsAtCompileTime)==int(OtherDerived::RowsAtCompileTime),
AreVectors = Derived::IsVectorAtCompileTime && OtherDerived::IsVectorAtCompileTime,
SameSizes = EIGEN_PREDICATE_SAME_MATRIX_SIZE(Derived,OtherDerived)
}; // note to the lost user: // * for a dot product use: v1.dot(v2) // * for a coeff-wise product use: v1.cwiseProduct(v2)
EIGEN_STATIC_ASSERT(ProductIsValid || !(AreVectors && SameSizes),
INVALID_VECTOR_VECTOR_PRODUCT__IF_YOU_WANTED_A_DOT_OR_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTIONS)
EIGEN_STATIC_ASSERT(ProductIsValid || !(SameSizes && !AreVectors),
INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION)
EIGEN_STATIC_ASSERT(ProductIsValid || SameSizes, INVALID_MATRIX_PRODUCT)
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.