diff options
author | Gael Guennebaud <g.gael@free.fr> | 2011-03-23 16:12:21 +0100 |
---|---|---|
committer | Gael Guennebaud <g.gael@free.fr> | 2011-03-23 16:12:21 +0100 |
commit | 816541d82c3b853c0c73b42f9e978f069afc43e1 (patch) | |
tree | cb572e4024a9af5adf47359bb10de7296e021bb9 /Eigen | |
parent | cfd5c2d74eaa4355eef8b9a9bda59e1cd6babf84 (diff) |
add a stupid Product<A,B> expression produced by prod(a,b), and implement a first version of its evaluator
Diffstat (limited to 'Eigen')
-rw-r--r-- | Eigen/Core | 3 | ||||
-rw-r--r-- | Eigen/src/Core/GeneralProduct.h | 609 | ||||
-rw-r--r-- | Eigen/src/Core/Product.h | 602 |
3 files changed, 671 insertions, 543 deletions
diff --git a/Eigen/Core b/Eigen/Core index aeaefbed6..ad86be3dd 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -321,7 +321,7 @@ using std::size_t; #include "src/Core/CommaInitializer.h" #include "src/Core/Flagged.h" #include "src/Core/ProductBase.h" -#include "src/Core/Product.h" +#include "src/Core/GeneralProduct.h" #include "src/Core/TriangularMatrix.h" #include "src/Core/SelfAdjointView.h" #include "src/Core/SolveTriangular.h" @@ -351,6 +351,7 @@ using std::size_t; #include "src/Core/ArrayWrapper.h" #ifdef EIGEN_ENABLE_EVALUATORS +#include "src/Core/Product.h" #include "src/Core/CoreEvaluators.h" #include "src/Core/AssignEvaluator.h" #endif diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h new file mode 100644 index 000000000..2d63120cf --- /dev/null +++ b/Eigen/src/Core/GeneralProduct.h @@ -0,0 +1,609 @@ +// 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> +// +// Eigen is free software; you can redistribute it and/or +// modify it under the terms of the GNU Lesser General Public +// License as published by the Free Software Foundation; either +// version 3 of the License, or (at your option) any later version. +// +// Alternatively, you can redistribute it and/or +// modify it under the terms of the GNU General Public License as +// published by the Free Software Foundation; either version 2 of +// the License, or (at your option) any later version. +// +// Eigen is distributed in the hope that it will be useful, but WITHOUT ANY +// WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS +// FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License or the +// GNU General Public License for more details. +// +// You should have received a copy of the GNU Lesser General Public +// License and a copy of the GNU General Public License along with +// Eigen. If not, see <http://www.gnu.org/licenses/>. + +#ifndef EIGEN_GENERAL_PRODUCT_H +#define EIGEN_GENERAL_PRODUCT_H + +/** \class GeneralProduct + * \ingroup Core_Module + * + * \brief Expression of the product of two general matrices or vectors + * + * \param LhsNested the type used to store the left-hand side + * \param RhsNested the type used to store the right-hand side + * \param ProductMode the type of the product + * + * This class represents an expression of the product of two general matrices. + * We call a general matrix, a dense matrix with full storage. For instance, + * This excludes triangular, selfadjoint, and sparse matrices. + * It is the return type of the operator* between general matrices. Its template + * arguments are determined automatically by ProductReturnType. Therefore, + * GeneralProduct should never be used direclty. To determine the result type of a + * function which involves a matrix product, use ProductReturnType::Type. + * + * \sa ProductReturnType, MatrixBase::operator*(const MatrixBase<OtherDerived>&) + */ +template<typename Lhs, typename Rhs, int ProductType = internal::product_type<Lhs,Rhs>::value> +class GeneralProduct; + +enum { + Large = 2, + Small = 3 +}; + +namespace internal { + +template<int Rows, int Cols, int Depth> struct product_type_selector; + +template<int Size, int MaxSize> struct product_size_category +{ + enum { is_large = MaxSize == Dynamic || + Size >= EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD, + value = is_large ? Large + : Size == 1 ? 1 + : Small + }; +}; + +template<typename Lhs, typename Rhs> struct product_type +{ + typedef typename remove_all<Lhs>::type _Lhs; + typedef typename remove_all<Rhs>::type _Rhs; + enum { + MaxRows = _Lhs::MaxRowsAtCompileTime, + Rows = _Lhs::RowsAtCompileTime, + MaxCols = _Rhs::MaxColsAtCompileTime, + Cols = _Rhs::ColsAtCompileTime, + MaxDepth = EIGEN_SIZE_MIN_PREFER_FIXED(_Lhs::MaxColsAtCompileTime, + _Rhs::MaxRowsAtCompileTime), + Depth = EIGEN_SIZE_MIN_PREFER_FIXED(_Lhs::ColsAtCompileTime, + _Rhs::RowsAtCompileTime), + LargeThreshold = EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD + }; + + // 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; + +public: + enum { + value = selector::ret + }; +#ifdef EIGEN_DEBUG_PRODUCT + static void debug() + { + EIGEN_DEBUG_VAR(Rows); + EIGEN_DEBUG_VAR(Cols); + EIGEN_DEBUG_VAR(Depth); + EIGEN_DEBUG_VAR(rows_select); + EIGEN_DEBUG_VAR(cols_select); + EIGEN_DEBUG_VAR(depth_select); + EIGEN_DEBUG_VAR(value); + } +#endif +}; + + +/* 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 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 = GemmProduct }; }; +template<> struct product_type_selector<Small,Large,Small> { enum { ret = GemmProduct }; }; +template<> struct product_type_selector<Large,Large,Small> { enum { ret = GemmProduct }; }; + +} // end namespace internal + +/** \class ProductReturnType + * \ingroup Core_Module + * + * \brief Helper class to get the correct and optimized returned type of operator* + * + * \param Lhs the type of the left-hand side + * \param Rhs the type of the right-hand side + * \param ProductMode the type of the product (determined automatically by internal::product_mode) + * + * This class defines the typename Type representing the optimized product expression + * between two matrix expressions. In practice, using ProductReturnType<Lhs,Rhs>::Type + * is the recommended way to define the result type of a function returning an expression + * which involve a matrix product. The class Product should never be + * used directly. + * + * \sa class Product, MatrixBase::operator*(const MatrixBase<OtherDerived>&) + */ +template<typename Lhs, typename Rhs, int ProductType> +struct ProductReturnType +{ + // TODO use the nested type to reduce instanciations ???? +// typedef typename internal::nested<Lhs,Rhs::ColsAtCompileTime>::type LhsNested; +// typedef typename internal::nested<Rhs,Lhs::RowsAtCompileTime>::type RhsNested; + + typedef GeneralProduct<Lhs/*Nested*/, Rhs/*Nested*/, ProductType> Type; +}; + +template<typename Lhs, typename Rhs> +struct ProductReturnType<Lhs,Rhs,CoeffBasedProductMode> +{ + typedef typename internal::nested<Lhs, Rhs::ColsAtCompileTime, typename internal::plain_matrix_type<Lhs>::type >::type LhsNested; + typedef typename internal::nested<Rhs, Lhs::RowsAtCompileTime, typename internal::plain_matrix_type<Rhs>::type >::type RhsNested; + typedef CoeffBasedProduct<LhsNested, RhsNested, EvalBeforeAssigningBit | EvalBeforeNestingBit> Type; +}; + +template<typename Lhs, typename Rhs> +struct ProductReturnType<Lhs,Rhs,LazyCoeffBasedProductMode> +{ + typedef typename internal::nested<Lhs, Rhs::ColsAtCompileTime, typename internal::plain_matrix_type<Lhs>::type >::type LhsNested; + typedef typename internal::nested<Rhs, Lhs::RowsAtCompileTime, typename internal::plain_matrix_type<Rhs>::type >::type RhsNested; + typedef CoeffBasedProduct<LhsNested, RhsNested, NestByRefBit> Type; +}; + +// this is a workaround for sun CC +template<typename Lhs, typename Rhs> +struct LazyProductReturnType : public ProductReturnType<Lhs,Rhs,LazyCoeffBasedProductMode> +{}; + +/*********************************************************************** +* 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); + +namespace internal { + +template<typename Lhs, typename Rhs> +struct traits<GeneralProduct<Lhs,Rhs,InnerProduct> > + : traits<Matrix<typename scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType,1,1> > +{}; + +} + +template<typename Lhs, typename Rhs> +class GeneralProduct<Lhs, Rhs, InnerProduct> + : internal::no_assignment_operator, + public Matrix<typename internal::scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType,1,1> +{ + typedef Matrix<typename internal::scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType,1,1> Base; + public: + GeneralProduct(const Lhs& lhs, const Rhs& rhs) + { + EIGEN_STATIC_ASSERT((internal::is_same<typename Lhs::RealScalar, typename Rhs::RealScalar>::value), + YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) + + Base::coeffRef(0,0) = (lhs.transpose().cwiseProduct(rhs)).sum(); + } + + /** Convertion to scalar */ + operator const typename Base::Scalar() const { + return Base::coeff(0,0); + } +}; + +/*********************************************************************** +* Implementation of Outer Vector Vector Product +***********************************************************************/ + +namespace internal { +template<int StorageOrder> struct outer_product_selector; + +template<typename Lhs, typename Rhs> +struct traits<GeneralProduct<Lhs,Rhs,OuterProduct> > + : traits<ProductBase<GeneralProduct<Lhs,Rhs,OuterProduct>, Lhs, Rhs> > +{}; + +} + +template<typename Lhs, typename Rhs> +class GeneralProduct<Lhs, Rhs, OuterProduct> + : public ProductBase<GeneralProduct<Lhs,Rhs,OuterProduct>, Lhs, Rhs> +{ + public: + EIGEN_PRODUCT_PUBLIC_INTERFACE(GeneralProduct) + + GeneralProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) + { + EIGEN_STATIC_ASSERT((internal::is_same<typename Lhs::RealScalar, typename Rhs::RealScalar>::value), + YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) + } + + template<typename Dest> void scaleAndAddTo(Dest& dest, Scalar alpha) const + { + internal::outer_product_selector<(int(Dest::Flags)&RowMajorBit) ? RowMajor : ColMajor>::run(*this, dest, alpha); + } +}; + +namespace internal { + +template<> struct outer_product_selector<ColMajor> { + template<typename ProductType, typename Dest> + static EIGEN_DONT_INLINE void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha) { + typedef typename Dest::Index Index; + // FIXME make sure lhs is sequentially stored + // FIXME not very good if rhs is real and lhs complex while alpha is real too + const Index cols = dest.cols(); + for (Index j=0; j<cols; ++j) + dest.col(j) += (alpha * prod.rhs().coeff(j)) * prod.lhs(); + } +}; + +template<> struct outer_product_selector<RowMajor> { + template<typename ProductType, typename Dest> + static EIGEN_DONT_INLINE void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha) { + typedef typename Dest::Index Index; + // FIXME make sure rhs is sequentially stored + // FIXME not very good if lhs is real and rhs complex while alpha is real too + const Index rows = dest.rows(); + for (Index i=0; i<rows; ++i) + dest.row(i) += (alpha * prod.lhs().coeff(i)) * prod.rhs(); + } +}; + +} // end namespace internal + +/*********************************************************************** +* 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<typename Lhs, typename Rhs> +struct traits<GeneralProduct<Lhs,Rhs,GemvProduct> > + : traits<ProductBase<GeneralProduct<Lhs,Rhs,GemvProduct>, Lhs, Rhs> > +{}; + +template<int Side, int StorageOrder, bool BlasCompatible> +struct gemv_selector; + +} // end namespace internal + +template<typename Lhs, typename Rhs> +class GeneralProduct<Lhs, Rhs, GemvProduct> + : public ProductBase<GeneralProduct<Lhs,Rhs,GemvProduct>, Lhs, Rhs> +{ + public: + EIGEN_PRODUCT_PUBLIC_INTERFACE(GeneralProduct) + + typedef typename Lhs::Scalar LhsScalar; + typedef typename Rhs::Scalar RhsScalar; + + GeneralProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) + { +// EIGEN_STATIC_ASSERT((internal::is_same<typename Lhs::Scalar, typename Rhs::Scalar>::value), +// YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) + } + + enum { Side = Lhs::IsVectorAtCompileTime ? OnTheLeft : OnTheRight }; + typedef typename internal::conditional<int(Side)==OnTheRight,_LhsNested,_RhsNested>::type MatrixType; + + template<typename Dest> void scaleAndAddTo(Dest& dst, Scalar alpha) const + { + eigen_assert(m_lhs.rows() == dst.rows() && m_rhs.cols() == dst.cols()); + internal::gemv_selector<Side,(int(MatrixType::Flags)&RowMajorBit) ? RowMajor : ColMajor, + bool(internal::blas_traits<MatrixType>::HasUsableDirectAccess)>::run(*this, dst, alpha); + } +}; + +namespace internal { + +// The vector is on the left => transposition +template<int StorageOrder, bool BlasCompatible> +struct gemv_selector<OnTheLeft,StorageOrder,BlasCompatible> +{ + template<typename ProductType, typename Dest> + static void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha) + { + Transpose<Dest> destT(dest); + enum { OtherStorageOrder = StorageOrder == RowMajor ? ColMajor : RowMajor }; + gemv_selector<OnTheRight,OtherStorageOrder,BlasCompatible> + ::run(GeneralProduct<Transpose<const typename ProductType::_RhsNested>,Transpose<const typename ProductType::_LhsNested>, GemvProduct> + (prod.rhs().transpose(), prod.lhs().transpose()), destT, alpha); + } +}; + +template<typename Scalar,int Size,int MaxSize,bool Cond> struct gemv_static_vector_if; + +template<typename Scalar,int Size,int MaxSize> +struct gemv_static_vector_if<Scalar,Size,MaxSize,false> +{ + EIGEN_STRONG_INLINE Scalar* data() { eigen_internal_assert(false && "should never be called"); return 0; } +}; + +template<typename Scalar,int Size> +struct gemv_static_vector_if<Scalar,Size,Dynamic,true> +{ + EIGEN_STRONG_INLINE Scalar* data() { return 0; } +}; + +template<typename Scalar,int Size,int MaxSize> +struct gemv_static_vector_if<Scalar,Size,MaxSize,true> +{ + internal::plain_array<Scalar,EIGEN_SIZE_MIN_PREFER_FIXED(Size,MaxSize),0> m_data; + EIGEN_STRONG_INLINE Scalar* data() { return m_data.array; } +}; + +template<> struct gemv_selector<OnTheRight,ColMajor,true> +{ + template<typename ProductType, typename Dest> + static inline void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha) + { + typedef typename ProductType::Index Index; + typedef typename ProductType::LhsScalar LhsScalar; + typedef typename ProductType::RhsScalar RhsScalar; + typedef typename ProductType::Scalar ResScalar; + typedef typename ProductType::RealScalar RealScalar; + typedef typename ProductType::ActualLhsType ActualLhsType; + typedef typename ProductType::ActualRhsType ActualRhsType; + typedef typename ProductType::LhsBlasTraits LhsBlasTraits; + typedef typename ProductType::RhsBlasTraits RhsBlasTraits; + typedef Map<Matrix<ResScalar,Dynamic,1>, Aligned> MappedDest; + + const ActualLhsType actualLhs = LhsBlasTraits::extract(prod.lhs()); + const ActualRhsType actualRhs = RhsBlasTraits::extract(prod.rhs()); + + ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs()) + * RhsBlasTraits::extractScalarFactor(prod.rhs()); + + 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 = Dest::InnerStrideAtCompileTime==1, + ComplexByReal = (NumTraits<LhsScalar>::IsComplex) && (!NumTraits<RhsScalar>::IsComplex), + MightCannotUseDest = (Dest::InnerStrideAtCompileTime!=1) || ComplexByReal + }; + + gemv_static_vector_if<ResScalar,Dest::SizeAtCompileTime,Dest::MaxSizeAtCompileTime,MightCannotUseDest> static_dest; + + bool alphaIsCompatible = (!ComplexByReal) || (imag(actualAlpha)==RealScalar(0)); + bool evalToDest = EvalToDestAtCompileTime && alphaIsCompatible; + + RhsScalar compatibleAlpha = get_factor<ResScalar,RhsScalar>::run(actualAlpha); + + ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(), + evalToDest ? dest.data() : static_dest.data()); + + if(!evalToDest) + { + #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN + int size = dest.size(); + EIGEN_DENSE_STORAGE_CTOR_PLUGIN + #endif + if(!alphaIsCompatible) + { + MappedDest(actualDestPtr, dest.size()).setZero(); + compatibleAlpha = RhsScalar(1); + } + else + MappedDest(actualDestPtr, dest.size()) = dest; + } + + general_matrix_vector_product + <Index,LhsScalar,ColMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsBlasTraits::NeedToConjugate>::run( + actualLhs.rows(), actualLhs.cols(), + &actualLhs.coeffRef(0,0), actualLhs.outerStride(), + actualRhs.data(), actualRhs.innerStride(), + actualDestPtr, 1, + compatibleAlpha); + + if (!evalToDest) + { + if(!alphaIsCompatible) + dest += actualAlpha * MappedDest(actualDestPtr, dest.size()); + else + dest = MappedDest(actualDestPtr, dest.size()); + } + } +}; + +template<> struct gemv_selector<OnTheRight,RowMajor,true> +{ + template<typename ProductType, typename Dest> + static void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha) + { + typedef typename ProductType::LhsScalar LhsScalar; + typedef typename ProductType::RhsScalar RhsScalar; + typedef typename ProductType::Scalar ResScalar; + typedef typename ProductType::Index Index; + typedef typename ProductType::ActualLhsType ActualLhsType; + typedef typename ProductType::ActualRhsType ActualRhsType; + typedef typename ProductType::_ActualRhsType _ActualRhsType; + typedef typename ProductType::LhsBlasTraits LhsBlasTraits; + typedef typename ProductType::RhsBlasTraits RhsBlasTraits; + + typename add_const<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(prod.lhs()); + typename add_const<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(prod.rhs()); + + ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs()) + * RhsBlasTraits::extractScalarFactor(prod.rhs()); + + 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 = _ActualRhsType::InnerStrideAtCompileTime==1 + }; + + gemv_static_vector_if<RhsScalar,_ActualRhsType::SizeAtCompileTime,_ActualRhsType::MaxSizeAtCompileTime,!DirectlyUseRhs> static_rhs; + + ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhsPtr,actualRhs.size(), + DirectlyUseRhs ? const_cast<RhsScalar*>(actualRhs.data()) : static_rhs.data()); + + if(!DirectlyUseRhs) + { + #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN + int size = actualRhs.size(); + EIGEN_DENSE_STORAGE_CTOR_PLUGIN + #endif + Map<typename _ActualRhsType::PlainObject>(actualRhsPtr, actualRhs.size()) = actualRhs; + } + + general_matrix_vector_product + <Index,LhsScalar,RowMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsBlasTraits::NeedToConjugate>::run( + actualLhs.rows(), actualLhs.cols(), + &actualLhs.coeffRef(0,0), actualLhs.outerStride(), + actualRhsPtr, 1, + &dest.coeffRef(0,0), dest.innerStride(), + actualAlpha); + } +}; + +template<> struct gemv_selector<OnTheRight,ColMajor,false> +{ + template<typename ProductType, typename Dest> + static void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha) + { + typedef typename Dest::Index Index; + // TODO makes sure dest is sequentially stored in memory, otherwise use a temp + const Index size = prod.rhs().rows(); + for(Index k=0; k<size; ++k) + dest += (alpha*prod.rhs().coeff(k)) * prod.lhs().col(k); + } +}; + +template<> struct gemv_selector<OnTheRight,RowMajor,false> +{ + template<typename ProductType, typename Dest> + static void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha) + { + typedef typename Dest::Index Index; + // TODO makes sure rhs is sequentially stored in memory, otherwise use a temp + const Index rows = prod.rows(); + for(Index i=0; i<rows; ++i) + dest.coeffRef(i) += alpha * (prod.lhs().row(i).cwiseProduct(prod.rhs().transpose())).sum(); + } +}; + +} // end namespace internal + +/*************************************************************************** +* 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> +inline const typename ProductReturnType<Derived,OtherDerived>::Type +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 + return typename ProductReturnType<Derived,OtherDerived>::Type(derived(), other.derived()); +} + +/** \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> +const typename LazyProductReturnType<Derived,OtherDerived>::Type +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) + + return typename LazyProductReturnType<Derived,OtherDerived>::Type(derived(), other.derived()); +} + +#endif // EIGEN_PRODUCT_H diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h index bde25375d..9bea26886 100644 --- a/Eigen/src/Core/Product.h +++ b/Eigen/src/Core/Product.h @@ -1,8 +1,7 @@ // 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 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2008-2011 Gael Guennebaud <gael.guennebaud@inria.fr> // // Eigen is free software; you can redistribute it and/or // modify it under the terms of the GNU Lesser General Public @@ -26,584 +25,103 @@ #ifndef EIGEN_PRODUCT_H #define EIGEN_PRODUCT_H -/** \class GeneralProduct +template<typename Lhs, typename Rhs> class Product; +template<typename Lhs, typename Rhs, typename StorageKind> class ProductImpl; + +/** \class Product * \ingroup Core_Module * - * \brief Expression of the product of two general matrices or vectors + * \brief Expression of the product of two arbitrary matrices or vectors * - * \param LhsNested the type used to store the left-hand side - * \param RhsNested the type used to store the right-hand side - * \param ProductMode the type of the product + * \param Lhs the type of the left-hand side expression + * \param Rhs the type of the right-hand side expression * - * This class represents an expression of the product of two general matrices. - * We call a general matrix, a dense matrix with full storage. For instance, - * This excludes triangular, selfadjoint, and sparse matrices. - * It is the return type of the operator* between general matrices. Its template - * arguments are determined automatically by ProductReturnType. Therefore, - * GeneralProduct should never be used direclty. To determine the result type of a - * function which involves a matrix product, use ProductReturnType::Type. + * This class represents an expression of the product of two arbitrary matrices. * - * \sa ProductReturnType, MatrixBase::operator*(const MatrixBase<OtherDerived>&) */ -template<typename Lhs, typename Rhs, int ProductType = internal::product_type<Lhs,Rhs>::value> -class GeneralProduct; - -enum { - Large = 2, - Small = 3 -}; namespace internal { - -template<int Rows, int Cols, int Depth> struct product_type_selector; - -template<int Size, int MaxSize> struct product_size_category -{ - enum { is_large = MaxSize == Dynamic || - Size >= EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD, - value = is_large ? Large - : Size == 1 ? 1 - : Small - }; -}; - -template<typename Lhs, typename Rhs> struct product_type -{ - typedef typename remove_all<Lhs>::type _Lhs; - typedef typename remove_all<Rhs>::type _Rhs; - enum { - MaxRows = _Lhs::MaxRowsAtCompileTime, - Rows = _Lhs::RowsAtCompileTime, - MaxCols = _Rhs::MaxColsAtCompileTime, - Cols = _Rhs::ColsAtCompileTime, - MaxDepth = EIGEN_SIZE_MIN_PREFER_FIXED(_Lhs::MaxColsAtCompileTime, - _Rhs::MaxRowsAtCompileTime), - Depth = EIGEN_SIZE_MIN_PREFER_FIXED(_Lhs::ColsAtCompileTime, - _Rhs::RowsAtCompileTime), - LargeThreshold = EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD - }; - - // 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; - -public: +template<typename Lhs, typename Rhs> +struct traits<Product<Lhs, Rhs> > +{ + typedef MatrixXpr XprKind; + typedef typename remove_all<Lhs>::type LhsCleaned; + typedef typename remove_all<Rhs>::type RhsCleaned; + typedef typename scalar_product_traits<typename traits<LhsCleaned>::Scalar, typename traits<RhsCleaned>::Scalar>::ReturnType Scalar; + typedef typename promote_storage_type<typename traits<LhsCleaned>::StorageKind, + typename traits<RhsCleaned>::StorageKind>::ret StorageKind; + typedef typename promote_index_type<typename traits<LhsCleaned>::Index, + typename traits<RhsCleaned>::Index>::type Index; enum { - value = selector::ret + RowsAtCompileTime = LhsCleaned::RowsAtCompileTime, + ColsAtCompileTime = RhsCleaned::ColsAtCompileTime, + MaxRowsAtCompileTime = LhsCleaned::MaxRowsAtCompileTime, + MaxColsAtCompileTime = RhsCleaned::MaxColsAtCompileTime, + Flags = (MaxRowsAtCompileTime==1 ? RowMajorBit : 0), // TODO should be no storage order + CoeffReadCost = 0 // TODO CoeffReadCost should not be part of the expression traits }; -#ifdef EIGEN_DEBUG_PRODUCT - static void debug() - { - EIGEN_DEBUG_VAR(Rows); - EIGEN_DEBUG_VAR(Cols); - EIGEN_DEBUG_VAR(Depth); - EIGEN_DEBUG_VAR(rows_select); - EIGEN_DEBUG_VAR(cols_select); - EIGEN_DEBUG_VAR(depth_select); - EIGEN_DEBUG_VAR(value); - } -#endif }; - - -/* 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 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 = GemmProduct }; }; -template<> struct product_type_selector<Small,Large,Small> { enum { ret = GemmProduct }; }; -template<> struct product_type_selector<Large,Large,Small> { enum { ret = GemmProduct }; }; - } // end namespace internal -/** \class ProductReturnType - * \ingroup Core_Module - * - * \brief Helper class to get the correct and optimized returned type of operator* - * - * \param Lhs the type of the left-hand side - * \param Rhs the type of the right-hand side - * \param ProductMode the type of the product (determined automatically by internal::product_mode) - * - * This class defines the typename Type representing the optimized product expression - * between two matrix expressions. In practice, using ProductReturnType<Lhs,Rhs>::Type - * is the recommended way to define the result type of a function returning an expression - * which involve a matrix product. The class Product should never be - * used directly. - * - * \sa class Product, MatrixBase::operator*(const MatrixBase<OtherDerived>&) - */ -template<typename Lhs, typename Rhs, int ProductType> -struct ProductReturnType -{ - // TODO use the nested type to reduce instanciations ???? -// typedef typename internal::nested<Lhs,Rhs::ColsAtCompileTime>::type LhsNested; -// typedef typename internal::nested<Rhs,Lhs::RowsAtCompileTime>::type RhsNested; - - typedef GeneralProduct<Lhs/*Nested*/, Rhs/*Nested*/, ProductType> Type; -}; template<typename Lhs, typename Rhs> -struct ProductReturnType<Lhs,Rhs,CoeffBasedProductMode> -{ - typedef typename internal::nested<Lhs, Rhs::ColsAtCompileTime, typename internal::plain_matrix_type<Lhs>::type >::type LhsNested; - typedef typename internal::nested<Rhs, Lhs::RowsAtCompileTime, typename internal::plain_matrix_type<Rhs>::type >::type RhsNested; - typedef CoeffBasedProduct<LhsNested, RhsNested, EvalBeforeAssigningBit | EvalBeforeNestingBit> Type; -}; - -template<typename Lhs, typename Rhs> -struct ProductReturnType<Lhs,Rhs,LazyCoeffBasedProductMode> -{ - typedef typename internal::nested<Lhs, Rhs::ColsAtCompileTime, typename internal::plain_matrix_type<Lhs>::type >::type LhsNested; - typedef typename internal::nested<Rhs, Lhs::RowsAtCompileTime, typename internal::plain_matrix_type<Rhs>::type >::type RhsNested; - typedef CoeffBasedProduct<LhsNested, RhsNested, NestByRefBit> Type; -}; - -// this is a workaround for sun CC -template<typename Lhs, typename Rhs> -struct LazyProductReturnType : public ProductReturnType<Lhs,Rhs,LazyCoeffBasedProductMode> -{}; - -/*********************************************************************** -* 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); - -namespace internal { - -template<typename Lhs, typename Rhs> -struct traits<GeneralProduct<Lhs,Rhs,InnerProduct> > - : traits<Matrix<typename scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType,1,1> > -{}; - -} - -template<typename Lhs, typename Rhs> -class GeneralProduct<Lhs, Rhs, InnerProduct> - : internal::no_assignment_operator, - public Matrix<typename internal::scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType,1,1> -{ - typedef Matrix<typename internal::scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType,1,1> Base; - public: - GeneralProduct(const Lhs& lhs, const Rhs& rhs) - { - EIGEN_STATIC_ASSERT((internal::is_same<typename Lhs::RealScalar, typename Rhs::RealScalar>::value), - YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) - - Base::coeffRef(0,0) = (lhs.transpose().cwiseProduct(rhs)).sum(); - } - - /** Convertion to scalar */ - operator const typename Base::Scalar() const { - return Base::coeff(0,0); - } -}; - -/*********************************************************************** -* Implementation of Outer Vector Vector Product -***********************************************************************/ - -namespace internal { -template<int StorageOrder> struct outer_product_selector; - -template<typename Lhs, typename Rhs> -struct traits<GeneralProduct<Lhs,Rhs,OuterProduct> > - : traits<ProductBase<GeneralProduct<Lhs,Rhs,OuterProduct>, Lhs, Rhs> > -{}; - -} - -template<typename Lhs, typename Rhs> -class GeneralProduct<Lhs, Rhs, OuterProduct> - : public ProductBase<GeneralProduct<Lhs,Rhs,OuterProduct>, Lhs, Rhs> +class Product : public ProductImpl<Lhs,Rhs,typename internal::promote_storage_type<typename internal::traits<Lhs>::StorageKind, + typename internal::traits<Rhs>::StorageKind>::ret> { public: - EIGEN_PRODUCT_PUBLIC_INTERFACE(GeneralProduct) - - GeneralProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) + + typedef typename ProductImpl< + Lhs, Rhs, + typename internal::promote_storage_type<typename Lhs::StorageKind, + typename Rhs::StorageKind>::ret>::Base Base; + EIGEN_GENERIC_PUBLIC_INTERFACE(Product) + + typedef typename Lhs::Nested LhsNested; + typedef typename Rhs::Nested RhsNested; + typedef typename internal::remove_all<LhsNested>::type LhsNestedCleaned; + typedef typename internal::remove_all<RhsNested>::type RhsNestedCleaned; + + Product(const Lhs& lhs, const Rhs& rhs) : m_lhs(lhs), m_rhs(rhs) { - EIGEN_STATIC_ASSERT((internal::is_same<typename Lhs::RealScalar, typename Rhs::RealScalar>::value), - YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) + eigen_assert(lhs.cols() == rhs.rows() + && "invalid matrix product" + && "if you wanted a coeff-wise or a dot product use the respective explicit functions"); } - template<typename Dest> void scaleAndAddTo(Dest& dest, Scalar alpha) const - { - internal::outer_product_selector<(int(Dest::Flags)&RowMajorBit) ? RowMajor : ColMajor>::run(*this, dest, alpha); - } -}; + inline Index rows() const { return m_lhs.rows(); } + inline Index cols() const { return m_rhs.cols(); } -namespace internal { + const LhsNestedCleaned& lhs() const { return m_lhs; } + const RhsNestedCleaned& rhs() const { return m_rhs; } -template<> struct outer_product_selector<ColMajor> { - template<typename ProductType, typename Dest> - static EIGEN_DONT_INLINE void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha) { - typedef typename Dest::Index Index; - // FIXME make sure lhs is sequentially stored - // FIXME not very good if rhs is real and lhs complex while alpha is real too - const Index cols = dest.cols(); - for (Index j=0; j<cols; ++j) - dest.col(j) += (alpha * prod.rhs().coeff(j)) * prod.lhs(); - } -}; + protected: -template<> struct outer_product_selector<RowMajor> { - template<typename ProductType, typename Dest> - static EIGEN_DONT_INLINE void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha) { - typedef typename Dest::Index Index; - // FIXME make sure rhs is sequentially stored - // FIXME not very good if lhs is real and rhs complex while alpha is real too - const Index rows = dest.rows(); - for (Index i=0; i<rows; ++i) - dest.row(i) += (alpha * prod.lhs().coeff(i)) * prod.rhs(); - } + const LhsNested m_lhs; + const RhsNested m_rhs; }; -} // end namespace internal - -/*********************************************************************** -* 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<typename Lhs, typename Rhs> -struct traits<GeneralProduct<Lhs,Rhs,GemvProduct> > - : traits<ProductBase<GeneralProduct<Lhs,Rhs,GemvProduct>, Lhs, Rhs> > -{}; - -template<int Side, int StorageOrder, bool BlasCompatible> -struct gemv_selector; - -} // end namespace internal - -template<typename Lhs, typename Rhs> -class GeneralProduct<Lhs, Rhs, GemvProduct> - : public ProductBase<GeneralProduct<Lhs,Rhs,GemvProduct>, Lhs, Rhs> +class ProductImpl<Lhs,Rhs,Dense> : public internal::dense_xpr_base<Product<Lhs,Rhs> >::type { + typedef Product<Lhs, Rhs> Derived; public: - EIGEN_PRODUCT_PUBLIC_INTERFACE(GeneralProduct) - - typedef typename Lhs::Scalar LhsScalar; - typedef typename Rhs::Scalar RhsScalar; - - GeneralProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) - { -// EIGEN_STATIC_ASSERT((internal::is_same<typename Lhs::Scalar, typename Rhs::Scalar>::value), -// YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) - } - - enum { Side = Lhs::IsVectorAtCompileTime ? OnTheLeft : OnTheRight }; - typedef typename internal::conditional<int(Side)==OnTheRight,_LhsNested,_RhsNested>::type MatrixType; - - template<typename Dest> void scaleAndAddTo(Dest& dst, Scalar alpha) const - { - eigen_assert(m_lhs.rows() == dst.rows() && m_rhs.cols() == dst.cols()); - internal::gemv_selector<Side,(int(MatrixType::Flags)&RowMajorBit) ? RowMajor : ColMajor, - bool(internal::blas_traits<MatrixType>::HasUsableDirectAccess)>::run(*this, dst, alpha); - } -}; - -namespace internal { - -// The vector is on the left => transposition -template<int StorageOrder, bool BlasCompatible> -struct gemv_selector<OnTheLeft,StorageOrder,BlasCompatible> -{ - template<typename ProductType, typename Dest> - static void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha) - { - Transpose<Dest> destT(dest); - enum { OtherStorageOrder = StorageOrder == RowMajor ? ColMajor : RowMajor }; - gemv_selector<OnTheRight,OtherStorageOrder,BlasCompatible> - ::run(GeneralProduct<Transpose<const typename ProductType::_RhsNested>,Transpose<const typename ProductType::_LhsNested>, GemvProduct> - (prod.rhs().transpose(), prod.lhs().transpose()), destT, alpha); - } -}; - -template<typename Scalar,int Size,int MaxSize,bool Cond> struct gemv_static_vector_if; - -template<typename Scalar,int Size,int MaxSize> -struct gemv_static_vector_if<Scalar,Size,MaxSize,false> -{ - EIGEN_STRONG_INLINE Scalar* data() { eigen_internal_assert(false && "should never be called"); return 0; } -}; - -template<typename Scalar,int Size> -struct gemv_static_vector_if<Scalar,Size,Dynamic,true> -{ - EIGEN_STRONG_INLINE Scalar* data() { return 0; } -}; - -template<typename Scalar,int Size,int MaxSize> -struct gemv_static_vector_if<Scalar,Size,MaxSize,true> -{ - internal::plain_array<Scalar,EIGEN_SIZE_MIN_PREFER_FIXED(Size,MaxSize),0> m_data; - EIGEN_STRONG_INLINE Scalar* data() { return m_data.array; } -}; - -template<> struct gemv_selector<OnTheRight,ColMajor,true> -{ - template<typename ProductType, typename Dest> - static inline void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha) - { - typedef typename ProductType::Index Index; - typedef typename ProductType::LhsScalar LhsScalar; - typedef typename ProductType::RhsScalar RhsScalar; - typedef typename ProductType::Scalar ResScalar; - typedef typename ProductType::RealScalar RealScalar; - typedef typename ProductType::ActualLhsType ActualLhsType; - typedef typename ProductType::ActualRhsType ActualRhsType; - typedef typename ProductType::LhsBlasTraits LhsBlasTraits; - typedef typename ProductType::RhsBlasTraits RhsBlasTraits; - typedef Map<Matrix<ResScalar,Dynamic,1>, Aligned> MappedDest; - - const ActualLhsType actualLhs = LhsBlasTraits::extract(prod.lhs()); - const ActualRhsType actualRhs = RhsBlasTraits::extract(prod.rhs()); - - ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs()) - * RhsBlasTraits::extractScalarFactor(prod.rhs()); - - 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 = Dest::InnerStrideAtCompileTime==1, - ComplexByReal = (NumTraits<LhsScalar>::IsComplex) && (!NumTraits<RhsScalar>::IsComplex), - MightCannotUseDest = (Dest::InnerStrideAtCompileTime!=1) || ComplexByReal - }; - - gemv_static_vector_if<ResScalar,Dest::SizeAtCompileTime,Dest::MaxSizeAtCompileTime,MightCannotUseDest> static_dest; - - bool alphaIsCompatible = (!ComplexByReal) || (imag(actualAlpha)==RealScalar(0)); - bool evalToDest = EvalToDestAtCompileTime && alphaIsCompatible; - - RhsScalar compatibleAlpha = get_factor<ResScalar,RhsScalar>::run(actualAlpha); - - ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(), - evalToDest ? dest.data() : static_dest.data()); - - if(!evalToDest) - { - #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN - int size = dest.size(); - EIGEN_DENSE_STORAGE_CTOR_PLUGIN - #endif - if(!alphaIsCompatible) - { - MappedDest(actualDestPtr, dest.size()).setZero(); - compatibleAlpha = RhsScalar(1); - } - else - MappedDest(actualDestPtr, dest.size()) = dest; - } - - general_matrix_vector_product - <Index,LhsScalar,ColMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsBlasTraits::NeedToConjugate>::run( - actualLhs.rows(), actualLhs.cols(), - &actualLhs.coeffRef(0,0), actualLhs.outerStride(), - actualRhs.data(), actualRhs.innerStride(), - actualDestPtr, 1, - compatibleAlpha); - - if (!evalToDest) - { - if(!alphaIsCompatible) - dest += actualAlpha * MappedDest(actualDestPtr, dest.size()); - else - dest = MappedDest(actualDestPtr, dest.size()); - } - } -}; - -template<> struct gemv_selector<OnTheRight,RowMajor,true> -{ - template<typename ProductType, typename Dest> - static void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha) - { - typedef typename ProductType::LhsScalar LhsScalar; - typedef typename ProductType::RhsScalar RhsScalar; - typedef typename ProductType::Scalar ResScalar; - typedef typename ProductType::Index Index; - typedef typename ProductType::ActualLhsType ActualLhsType; - typedef typename ProductType::ActualRhsType ActualRhsType; - typedef typename ProductType::_ActualRhsType _ActualRhsType; - typedef typename ProductType::LhsBlasTraits LhsBlasTraits; - typedef typename ProductType::RhsBlasTraits RhsBlasTraits; - - typename add_const<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(prod.lhs()); - typename add_const<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(prod.rhs()); - - ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs()) - * RhsBlasTraits::extractScalarFactor(prod.rhs()); - - 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 = _ActualRhsType::InnerStrideAtCompileTime==1 - }; - - gemv_static_vector_if<RhsScalar,_ActualRhsType::SizeAtCompileTime,_ActualRhsType::MaxSizeAtCompileTime,!DirectlyUseRhs> static_rhs; - - ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhsPtr,actualRhs.size(), - DirectlyUseRhs ? const_cast<RhsScalar*>(actualRhs.data()) : static_rhs.data()); - - if(!DirectlyUseRhs) - { - #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN - int size = actualRhs.size(); - EIGEN_DENSE_STORAGE_CTOR_PLUGIN - #endif - Map<typename _ActualRhsType::PlainObject>(actualRhsPtr, actualRhs.size()) = actualRhs; - } - - general_matrix_vector_product - <Index,LhsScalar,RowMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsBlasTraits::NeedToConjugate>::run( - actualLhs.rows(), actualLhs.cols(), - &actualLhs.coeffRef(0,0), actualLhs.outerStride(), - actualRhsPtr, 1, - &dest.coeffRef(0,0), dest.innerStride(), - actualAlpha); - } -}; -template<> struct gemv_selector<OnTheRight,ColMajor,false> -{ - template<typename ProductType, typename Dest> - static void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha) - { - typedef typename Dest::Index Index; - // TODO makes sure dest is sequentially stored in memory, otherwise use a temp - const Index size = prod.rhs().rows(); - for(Index k=0; k<size; ++k) - dest += (alpha*prod.rhs().coeff(k)) * prod.lhs().col(k); - } + typedef typename internal::dense_xpr_base<Product<Lhs, Rhs> >::type Base; + EIGEN_DENSE_PUBLIC_INTERFACE(Derived) }; -template<> struct gemv_selector<OnTheRight,RowMajor,false> -{ - template<typename ProductType, typename Dest> - static void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha) - { - typedef typename Dest::Index Index; - // TODO makes sure rhs is sequentially stored in memory, otherwise use a temp - const Index rows = prod.rows(); - for(Index i=0; i<rows; ++i) - dest.coeffRef(i) += alpha * (prod.lhs().row(i).cwiseProduct(prod.rhs().transpose())).sum(); - } -}; - -} // end namespace internal - /*************************************************************************** * 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> -inline const typename ProductReturnType<Derived,OtherDerived>::Type -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 - return typename ProductReturnType<Derived,OtherDerived>::Type(derived(), other.derived()); -} -/** \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&) +/** \internal used to test the evaluator only */ -template<typename Derived> -template<typename OtherDerived> -const typename LazyProductReturnType<Derived,OtherDerived>::Type -MatrixBase<Derived>::lazyProduct(const MatrixBase<OtherDerived> &other) const +template<typename Lhs,typename Rhs> +const Product<Lhs,Rhs> +prod(const Lhs& lhs, const Rhs& rhs) { - 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) - - return typename LazyProductReturnType<Derived,OtherDerived>::Type(derived(), other.derived()); + return Product<Lhs,Rhs>(lhs,rhs); } #endif // EIGEN_PRODUCT_H |