diff options
Diffstat (limited to 'Eigen/src/Core/products/SelfadjointMatrixVector.h')
| -rw-r--r-- | Eigen/src/Core/products/SelfadjointMatrixVector.h | 127 |
1 files changed, 53 insertions, 74 deletions
diff --git a/Eigen/src/Core/products/SelfadjointMatrixVector.h b/Eigen/src/Core/products/SelfadjointMatrixVector.h index f698f67f9..3fd180e6c 100644 --- a/Eigen/src/Core/products/SelfadjointMatrixVector.h +++ b/Eigen/src/Core/products/SelfadjointMatrixVector.h @@ -30,7 +30,7 @@ struct selfadjoint_matrix_vector_product static EIGEN_DONT_INLINE void run( Index size, const Scalar* lhs, Index lhsStride, - const Scalar* _rhs, Index rhsIncr, + const Scalar* rhs, Scalar* res, Scalar alpha); }; @@ -39,11 +39,12 @@ template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool Conju EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs,Version>::run( Index size, const Scalar* lhs, Index lhsStride, - const Scalar* _rhs, Index rhsIncr, + const Scalar* rhs, Scalar* res, Scalar alpha) { typedef typename packet_traits<Scalar>::type Packet; + typedef typename NumTraits<Scalar>::Real RealScalar; const Index PacketSize = sizeof(Packet)/sizeof(Scalar); enum { @@ -54,23 +55,13 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrd conj_helper<Scalar,Scalar,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs, IsRowMajor), ConjugateRhs> cj0; conj_helper<Scalar,Scalar,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs, !IsRowMajor), ConjugateRhs> cj1; - conj_helper<Scalar,Scalar,NumTraits<Scalar>::IsComplex, ConjugateRhs> cjd; + conj_helper<RealScalar,Scalar,false, ConjugateRhs> cjd; conj_helper<Packet,Packet,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs, IsRowMajor), ConjugateRhs> pcj0; conj_helper<Packet,Packet,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs, !IsRowMajor), ConjugateRhs> pcj1; Scalar cjAlpha = ConjugateRhs ? numext::conj(alpha) : alpha; - // FIXME this copy is now handled outside product_selfadjoint_vector, so it could probably be removed. - // if the rhs is not sequentially stored in memory we copy it to a temporary buffer, - // this is because we need to extract packets - ei_declare_aligned_stack_constructed_variable(Scalar,rhs,size,rhsIncr==1 ? const_cast<Scalar*>(_rhs) : 0); - if (rhsIncr!=1) - { - const Scalar* it = _rhs; - for (Index i=0; i<size; ++i, it+=rhsIncr) - rhs[i] = *it; - } Index bound = (std::max)(Index(0),size-8) & 0xfffffffe; if (FirstTriangular) @@ -92,12 +83,11 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrd Scalar t3(0); Packet ptmp3 = pset1<Packet>(t3); - size_t starti = FirstTriangular ? 0 : j+2; - size_t endi = FirstTriangular ? j : size; - size_t alignedStart = (starti) + internal::first_aligned(&res[starti], endi-starti); - size_t alignedEnd = alignedStart + ((endi-alignedStart)/(PacketSize))*(PacketSize); + Index starti = FirstTriangular ? 0 : j+2; + Index endi = FirstTriangular ? j : size; + Index alignedStart = (starti) + internal::first_default_aligned(&res[starti], endi-starti); + Index alignedEnd = alignedStart + ((endi-alignedStart)/(PacketSize))*(PacketSize); - // TODO make sure this product is a real * complex and that the rhs is properly conjugated if needed res[j] += cjd.pmul(numext::real(A0[j]), t0); res[j+1] += cjd.pmul(numext::real(A1[j+1]), t1); if(FirstTriangular) @@ -111,11 +101,11 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrd t2 += cj1.pmul(A0[j+1], rhs[j+1]); } - for (size_t i=starti; i<alignedStart; ++i) + for (Index i=starti; i<alignedStart; ++i) { - res[i] += t0 * A0[i] + t1 * A1[i]; - t2 += numext::conj(A0[i]) * rhs[i]; - t3 += numext::conj(A1[i]) * rhs[i]; + res[i] += cj0.pmul(A0[i], t0) + cj0.pmul(A1[i],t1); + t2 += cj1.pmul(A0[i], rhs[i]); + t3 += cj1.pmul(A1[i], rhs[i]); } // Yes this an optimization for gcc 4.3 and 4.4 (=> huge speed up) // gcc 4.2 does this optimization automatically. @@ -123,7 +113,7 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrd const Scalar* EIGEN_RESTRICT a1It = A1 + alignedStart; const Scalar* EIGEN_RESTRICT rhsIt = rhs + alignedStart; Scalar* EIGEN_RESTRICT resIt = res + alignedStart; - for (size_t i=alignedStart; i<alignedEnd; i+=PacketSize) + for (Index i=alignedStart; i<alignedEnd; i+=PacketSize) { Packet A0i = ploadu<Packet>(a0It); a0It += PacketSize; Packet A1i = ploadu<Packet>(a1It); a1It += PacketSize; @@ -135,7 +125,7 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrd ptmp3 = pcj1.pmadd(A1i, Bi, ptmp3); pstore(resIt,Xi); resIt += PacketSize; } - for (size_t i=alignedEnd; i<endi; i++) + for (Index i=alignedEnd; i<endi; i++) { res[i] += cj0.pmul(A0[i], t0) + cj0.pmul(A1[i],t1); t2 += cj1.pmul(A0[i], rhs[i]); @@ -151,7 +141,6 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrd Scalar t1 = cjAlpha * rhs[j]; Scalar t2(0); - // TODO make sure this product is a real * complex and that the rhs is properly conjugated if needed res[j] += cjd.pmul(numext::real(A0[j]), t1); for (Index i=FirstTriangular ? 0 : j+1; i<(FirstTriangular ? j : size); i++) { @@ -169,45 +158,44 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrd ***************************************************************************/ namespace internal { -template<typename Lhs, int LhsMode, typename Rhs> -struct traits<SelfadjointProductMatrix<Lhs,LhsMode,false,Rhs,0,true> > - : traits<ProductBase<SelfadjointProductMatrix<Lhs,LhsMode,false,Rhs,0,true>, Lhs, Rhs> > -{}; -} template<typename Lhs, int LhsMode, typename Rhs> -struct SelfadjointProductMatrix<Lhs,LhsMode,false,Rhs,0,true> - : public ProductBase<SelfadjointProductMatrix<Lhs,LhsMode,false,Rhs,0,true>, Lhs, Rhs > +struct selfadjoint_product_impl<Lhs,LhsMode,false,Rhs,0,true> { - EIGEN_PRODUCT_PUBLIC_INTERFACE(SelfadjointProductMatrix) - - enum { - LhsUpLo = LhsMode&(Upper|Lower) - }; - - SelfadjointProductMatrix(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) {} - - template<typename Dest> void scaleAndAddTo(Dest& dest, const Scalar& alpha) const + typedef typename Product<Lhs,Rhs>::Scalar Scalar; + + typedef internal::blas_traits<Lhs> LhsBlasTraits; + typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType; + typedef typename internal::remove_all<ActualLhsType>::type ActualLhsTypeCleaned; + + typedef internal::blas_traits<Rhs> RhsBlasTraits; + typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType; + typedef typename internal::remove_all<ActualRhsType>::type ActualRhsTypeCleaned; + + enum { LhsUpLo = LhsMode&(Upper|Lower) }; + + template<typename Dest> + static void run(Dest& dest, const Lhs &a_lhs, const Rhs &a_rhs, const Scalar& alpha) { typedef typename Dest::Scalar ResScalar; - typedef typename Base::RhsScalar RhsScalar; - typedef Map<Matrix<ResScalar,Dynamic,1>, Aligned> MappedDest; + typedef typename Rhs::Scalar RhsScalar; + typedef Map<Matrix<ResScalar,Dynamic,1>, EIGEN_PLAIN_ENUM_MIN(AlignedMax,internal::packet_traits<ResScalar>::size)> MappedDest; - eigen_assert(dest.rows()==m_lhs.rows() && dest.cols()==m_rhs.cols()); + eigen_assert(dest.rows()==a_lhs.rows() && dest.cols()==a_rhs.cols()); - typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(m_lhs); - typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(m_rhs); + 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(m_lhs) - * RhsBlasTraits::extractScalarFactor(m_rhs); + Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(a_lhs) + * RhsBlasTraits::extractScalarFactor(a_rhs); enum { EvalToDest = (Dest::InnerStrideAtCompileTime==1), - UseRhs = (_ActualRhsType::InnerStrideAtCompileTime==1) + UseRhs = (ActualRhsTypeCleaned::InnerStrideAtCompileTime==1) }; internal::gemv_static_vector_if<ResScalar,Dest::SizeAtCompileTime,Dest::MaxSizeAtCompileTime,!EvalToDest> static_dest; - internal::gemv_static_vector_if<RhsScalar,_ActualRhsType::SizeAtCompileTime,_ActualRhsType::MaxSizeAtCompileTime,!UseRhs> static_rhs; + internal::gemv_static_vector_if<RhsScalar,ActualRhsTypeCleaned::SizeAtCompileTime,ActualRhsTypeCleaned::MaxSizeAtCompileTime,!UseRhs> static_rhs; ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(), EvalToDest ? dest.data() : static_dest.data()); @@ -218,7 +206,7 @@ struct SelfadjointProductMatrix<Lhs,LhsMode,false,Rhs,0,true> if(!EvalToDest) { #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN - int size = dest.size(); + Index size = dest.size(); EIGEN_DENSE_STORAGE_CTOR_PLUGIN #endif MappedDest(actualDestPtr, dest.size()) = dest; @@ -227,18 +215,19 @@ struct SelfadjointProductMatrix<Lhs,LhsMode,false,Rhs,0,true> if(!UseRhs) { #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN - int size = rhs.size(); + Index size = rhs.size(); EIGEN_DENSE_STORAGE_CTOR_PLUGIN #endif - Map<typename _ActualRhsType::PlainObject>(actualRhsPtr, rhs.size()) = rhs; + Map<typename ActualRhsTypeCleaned::PlainObject>(actualRhsPtr, rhs.size()) = rhs; } - internal::selfadjoint_matrix_vector_product<Scalar, Index, (internal::traits<_ActualLhsType>::Flags&RowMajorBit) ? RowMajor : ColMajor, int(LhsUpLo), bool(LhsBlasTraits::NeedToConjugate), bool(RhsBlasTraits::NeedToConjugate)>::run + internal::selfadjoint_matrix_vector_product<Scalar, Index, (internal::traits<ActualLhsTypeCleaned>::Flags&RowMajorBit) ? RowMajor : ColMajor, + int(LhsUpLo), bool(LhsBlasTraits::NeedToConjugate), bool(RhsBlasTraits::NeedToConjugate)>::run ( lhs.rows(), // size &lhs.coeffRef(0,0), lhs.outerStride(), // lhs info - actualRhsPtr, 1, // rhs info + actualRhsPtr, // rhs info actualDestPtr, // result info actualAlpha // scale factor ); @@ -248,34 +237,24 @@ struct SelfadjointProductMatrix<Lhs,LhsMode,false,Rhs,0,true> } }; -namespace internal { -template<typename Lhs, typename Rhs, int RhsMode> -struct traits<SelfadjointProductMatrix<Lhs,0,true,Rhs,RhsMode,false> > - : traits<ProductBase<SelfadjointProductMatrix<Lhs,0,true,Rhs,RhsMode,false>, Lhs, Rhs> > -{}; -} - template<typename Lhs, typename Rhs, int RhsMode> -struct SelfadjointProductMatrix<Lhs,0,true,Rhs,RhsMode,false> - : public ProductBase<SelfadjointProductMatrix<Lhs,0,true,Rhs,RhsMode,false>, Lhs, Rhs > +struct selfadjoint_product_impl<Lhs,0,true,Rhs,RhsMode,false> { - EIGEN_PRODUCT_PUBLIC_INTERFACE(SelfadjointProductMatrix) + typedef typename Product<Lhs,Rhs>::Scalar Scalar; + enum { RhsUpLo = RhsMode&(Upper|Lower) }; - enum { - RhsUpLo = RhsMode&(Upper|Lower) - }; - - SelfadjointProductMatrix(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) {} - - template<typename Dest> void scaleAndAddTo(Dest& dest, const Scalar& alpha) const + template<typename Dest> + static void run(Dest& dest, const Lhs &a_lhs, const Rhs &a_rhs, const Scalar& alpha) { // let's simply transpose the product Transpose<Dest> destT(dest); - SelfadjointProductMatrix<Transpose<const Rhs>, int(RhsUpLo)==Upper ? Lower : Upper, false, - Transpose<const Lhs>, 0, true>(m_rhs.transpose(), m_lhs.transpose()).scaleAndAddTo(destT, alpha); + selfadjoint_product_impl<Transpose<const Rhs>, int(RhsUpLo)==Upper ? Lower : Upper, false, + Transpose<const Lhs>, 0, true>::run(destT, a_rhs.transpose(), a_lhs.transpose(), alpha); } }; +} // end namespace internal + } // end namespace Eigen #endif // EIGEN_SELFADJOINT_MATRIX_VECTOR_H |