diff --git a/xs/src/eigen/Eigen/Cholesky b/xs/src/eigen/Eigen/Cholesky index 369d1f5ec..1332b540d 100644 --- a/xs/src/eigen/Eigen/Cholesky +++ b/xs/src/eigen/Eigen/Cholesky @@ -9,6 +9,7 @@ #define EIGEN_CHOLESKY_MODULE_H #include "Core" +#include "Jacobi" #include "src/Core/util/DisableStupidWarnings.h" @@ -31,7 +32,11 @@ #include "src/Cholesky/LLT.h" #include "src/Cholesky/LDLT.h" #ifdef EIGEN_USE_LAPACKE +#ifdef EIGEN_USE_MKL +#include "mkl_lapacke.h" +#else #include "src/misc/lapacke.h" +#endif #include "src/Cholesky/LLT_LAPACKE.h" #endif diff --git a/xs/src/eigen/Eigen/Core b/xs/src/eigen/Eigen/Core index 0f7fa630d..4d4901e03 100644 --- a/xs/src/eigen/Eigen/Core +++ b/xs/src/eigen/Eigen/Core @@ -14,6 +14,22 @@ // first thing Eigen does: stop the compiler from committing suicide #include "src/Core/util/DisableStupidWarnings.h" +#if defined(__CUDACC__) && !defined(EIGEN_NO_CUDA) + #define EIGEN_CUDACC __CUDACC__ +#endif + +#if defined(__CUDA_ARCH__) && !defined(EIGEN_NO_CUDA) + #define EIGEN_CUDA_ARCH __CUDA_ARCH__ +#endif + +#if defined(__CUDACC_VER_MAJOR__) && (__CUDACC_VER_MAJOR__ >= 9) +#define EIGEN_CUDACC_VER ((__CUDACC_VER_MAJOR__ * 10000) + (__CUDACC_VER_MINOR__ * 100)) +#elif defined(__CUDACC_VER__) +#define EIGEN_CUDACC_VER __CUDACC_VER__ +#else +#define EIGEN_CUDACC_VER 0 +#endif + // Handle NVCC/CUDA/SYCL #if defined(__CUDACC__) || defined(__SYCL_DEVICE_ONLY__) // Do not try asserts on CUDA and SYCL! @@ -155,6 +171,9 @@ #ifdef __AVX512DQ__ #define EIGEN_VECTORIZE_AVX512DQ #endif + #ifdef __AVX512ER__ + #define EIGEN_VECTORIZE_AVX512ER + #endif #endif // include files @@ -229,7 +248,7 @@ #if defined __CUDACC__ #define EIGEN_VECTORIZE_CUDA #include - #if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500 + #if EIGEN_CUDACC_VER >= 70500 #define EIGEN_HAS_CUDA_FP16 #endif #endif @@ -352,6 +371,7 @@ using std::ptrdiff_t; #include "src/Core/MathFunctions.h" #include "src/Core/GenericPacketMath.h" #include "src/Core/MathFunctionsImpl.h" +#include "src/Core/arch/Default/ConjHelper.h" #if defined EIGEN_VECTORIZE_AVX512 #include "src/Core/arch/SSE/PacketMath.h" @@ -367,6 +387,7 @@ using std::ptrdiff_t; #include "src/Core/arch/AVX/MathFunctions.h" #include "src/Core/arch/AVX/Complex.h" #include "src/Core/arch/AVX/TypeCasting.h" + #include "src/Core/arch/SSE/TypeCasting.h" #elif defined EIGEN_VECTORIZE_SSE #include "src/Core/arch/SSE/PacketMath.h" #include "src/Core/arch/SSE/MathFunctions.h" diff --git a/xs/src/eigen/Eigen/Eigenvalues b/xs/src/eigen/Eigen/Eigenvalues index 009e529e1..f3f661b07 100644 --- a/xs/src/eigen/Eigen/Eigenvalues +++ b/xs/src/eigen/Eigen/Eigenvalues @@ -45,7 +45,11 @@ #include "src/Eigenvalues/GeneralizedEigenSolver.h" #include "src/Eigenvalues/MatrixBaseEigenvalues.h" #ifdef EIGEN_USE_LAPACKE +#ifdef EIGEN_USE_MKL +#include "mkl_lapacke.h" +#else #include "src/misc/lapacke.h" +#endif #include "src/Eigenvalues/RealSchur_LAPACKE.h" #include "src/Eigenvalues/ComplexSchur_LAPACKE.h" #include "src/Eigenvalues/SelfAdjointEigenSolver_LAPACKE.h" diff --git a/xs/src/eigen/Eigen/LU b/xs/src/eigen/Eigen/LU index 6f6c55629..6418a86e1 100644 --- a/xs/src/eigen/Eigen/LU +++ b/xs/src/eigen/Eigen/LU @@ -28,7 +28,11 @@ #include "src/LU/FullPivLU.h" #include "src/LU/PartialPivLU.h" #ifdef EIGEN_USE_LAPACKE +#ifdef EIGEN_USE_MKL +#include "mkl_lapacke.h" +#else #include "src/misc/lapacke.h" +#endif #include "src/LU/PartialPivLU_LAPACKE.h" #endif #include "src/LU/Determinant.h" diff --git a/xs/src/eigen/Eigen/QR b/xs/src/eigen/Eigen/QR index 80838e3bd..c7e914469 100644 --- a/xs/src/eigen/Eigen/QR +++ b/xs/src/eigen/Eigen/QR @@ -36,7 +36,11 @@ #include "src/QR/ColPivHouseholderQR.h" #include "src/QR/CompleteOrthogonalDecomposition.h" #ifdef EIGEN_USE_LAPACKE +#ifdef EIGEN_USE_MKL +#include "mkl_lapacke.h" +#else #include "src/misc/lapacke.h" +#endif #include "src/QR/HouseholderQR_LAPACKE.h" #include "src/QR/ColPivHouseholderQR_LAPACKE.h" #endif diff --git a/xs/src/eigen/Eigen/QtAlignedMalloc b/xs/src/eigen/Eigen/QtAlignedMalloc index c6571f129..4f07df02a 100644 --- a/xs/src/eigen/Eigen/QtAlignedMalloc +++ b/xs/src/eigen/Eigen/QtAlignedMalloc @@ -27,7 +27,7 @@ void qFree(void *ptr) void *qRealloc(void *ptr, std::size_t size) { void* newPtr = Eigen::internal::aligned_malloc(size); - memcpy(newPtr, ptr, size); + std::memcpy(newPtr, ptr, size); Eigen::internal::aligned_free(ptr); return newPtr; } diff --git a/xs/src/eigen/Eigen/SVD b/xs/src/eigen/Eigen/SVD index 86143c23d..5d0e75f7f 100644 --- a/xs/src/eigen/Eigen/SVD +++ b/xs/src/eigen/Eigen/SVD @@ -37,7 +37,11 @@ #include "src/SVD/JacobiSVD.h" #include "src/SVD/BDCSVD.h" #if defined(EIGEN_USE_LAPACKE) && !defined(EIGEN_USE_LAPACKE_STRICT) +#ifdef EIGEN_USE_MKL +#include "mkl_lapacke.h" +#else #include "src/misc/lapacke.h" +#endif #include "src/SVD/JacobiSVD_LAPACKE.h" #endif diff --git a/xs/src/eigen/Eigen/src/Cholesky/LDLT.h b/xs/src/eigen/Eigen/src/Cholesky/LDLT.h index fcee7b2e3..0313a54bf 100644 --- a/xs/src/eigen/Eigen/src/Cholesky/LDLT.h +++ b/xs/src/eigen/Eigen/src/Cholesky/LDLT.h @@ -248,7 +248,7 @@ template class LDLT /** \brief Reports whether previous computation was successful. * * \returns \c Success if computation was succesful, - * \c NumericalIssue if the matrix.appears to be negative. + * \c NumericalIssue if the factorization failed because of a zero pivot. */ ComputationInfo info() const { @@ -376,6 +376,8 @@ template<> struct ldlt_inplace if((rs>0) && pivot_is_valid) A21 /= realAkk; + else if(rs>0) + ret = ret && (A21.array()==Scalar(0)).all(); if(found_zero_pivot && pivot_is_valid) ret = false; // factorization failed else if(!pivot_is_valid) found_zero_pivot = true; @@ -568,13 +570,14 @@ void LDLT<_MatrixType,_UpLo>::_solve_impl(const RhsType &rhs, DstType &dst) cons // more precisely, use pseudo-inverse of D (see bug 241) using std::abs; const typename Diagonal::RealReturnType vecD(vectorD()); - // In some previous versions, tolerance was set to the max of 1/highest and the maximal diagonal entry * epsilon - // as motivated by LAPACK's xGELSS: + // In some previous versions, tolerance was set to the max of 1/highest (or rather numeric_limits::min()) + // and the maximal diagonal entry * epsilon as motivated by LAPACK's xGELSS: // RealScalar tolerance = numext::maxi(vecD.array().abs().maxCoeff() * NumTraits::epsilon(),RealScalar(1) / NumTraits::highest()); // However, LDLT is not rank revealing, and so adjusting the tolerance wrt to the highest // diagonal element is not well justified and leads to numerical issues in some cases. // Moreover, Lapack's xSYTRS routines use 0 for the tolerance. - RealScalar tolerance = RealScalar(1) / NumTraits::highest(); + // Using numeric_limits::min() gives us more robustness to denormals. + RealScalar tolerance = (std::numeric_limits::min)(); for (Index i = 0; i < vecD.size(); ++i) { diff --git a/xs/src/eigen/Eigen/src/Cholesky/LLT.h b/xs/src/eigen/Eigen/src/Cholesky/LLT.h index 87ca8d423..e1624d21b 100644 --- a/xs/src/eigen/Eigen/src/Cholesky/LLT.h +++ b/xs/src/eigen/Eigen/src/Cholesky/LLT.h @@ -24,7 +24,7 @@ template struct LLT_Traits; * * \tparam _MatrixType the type of the matrix of which we are computing the LL^T Cholesky decomposition * \tparam _UpLo the triangular part that will be used for the decompositon: Lower (default) or Upper. - * The other triangular part won't be read. + * The other triangular part won't be read. * * This class performs a LL^T Cholesky decomposition of a symmetric, positive definite * matrix A such that A = LL^* = U^*U, where L is lower triangular. @@ -41,14 +41,18 @@ template struct LLT_Traits; * Example: \include LLT_example.cpp * Output: \verbinclude LLT_example.out * + * \b Performance: for best performance, it is recommended to use a column-major storage format + * with the Lower triangular part (the default), or, equivalently, a row-major storage format + * with the Upper triangular part. Otherwise, you might get a 20% slowdown for the full factorization + * step, and rank-updates can be up to 3 times slower. + * * This class supports the \link InplaceDecomposition inplace decomposition \endlink mechanism. * + * Note that during the decomposition, only the lower (or upper, as defined by _UpLo) triangular part of A is considered. + * Therefore, the strict lower part does not have to store correct values. + * * \sa MatrixBase::llt(), SelfAdjointView::llt(), class LDLT */ - /* HEY THIS DOX IS DISABLED BECAUSE THERE's A BUG EITHER HERE OR IN LDLT ABOUT THAT (OR BOTH) - * Note that during the decomposition, only the upper triangular part of A is considered. Therefore, - * the strict lower part does not have to store correct values. - */ template class LLT { public: @@ -146,7 +150,7 @@ template class LLT } template - void solveInPlace(MatrixBase &bAndX) const; + void solveInPlace(const MatrixBase &bAndX) const; template LLT& compute(const EigenBase& matrix); @@ -177,7 +181,7 @@ template class LLT /** \brief Reports whether previous computation was successful. * * \returns \c Success if computation was succesful, - * \c NumericalIssue if the matrix.appears to be negative. + * \c NumericalIssue if the matrix.appears not to be positive definite. */ ComputationInfo info() const { @@ -425,7 +429,8 @@ LLT& LLT::compute(const EigenBase eigen_assert(a.rows()==a.cols()); const Index size = a.rows(); m_matrix.resize(size, size); - m_matrix = a.derived(); + if (!internal::is_same_dense(m_matrix, a.derived())) + m_matrix = a.derived(); // Compute matrix L1 norm = max abs column sum. m_l1_norm = RealScalar(0); @@ -485,11 +490,14 @@ void LLT<_MatrixType,_UpLo>::_solve_impl(const RhsType &rhs, DstType &dst) const * * This version avoids a copy when the right hand side matrix b is not needed anymore. * + * \warning The parameter is only marked 'const' to make the C++ compiler accept a temporary expression here. + * This function will const_cast it, so constness isn't honored here. + * * \sa LLT::solve(), MatrixBase::llt() */ template template -void LLT::solveInPlace(MatrixBase &bAndX) const +void LLT::solveInPlace(const MatrixBase &bAndX) const { eigen_assert(m_isInitialized && "LLT is not initialized."); eigen_assert(m_matrix.rows()==bAndX.rows()); diff --git a/xs/src/eigen/Eigen/src/Core/Array.h b/xs/src/eigen/Eigen/src/Core/Array.h index 0d34269fd..e10020d4f 100644 --- a/xs/src/eigen/Eigen/src/Core/Array.h +++ b/xs/src/eigen/Eigen/src/Core/Array.h @@ -231,10 +231,16 @@ class Array : Base(other) { } + private: + struct PrivateType {}; + public: + /** \sa MatrixBase::operator=(const EigenBase&) */ template EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE Array(const EigenBase &other) + EIGEN_STRONG_INLINE Array(const EigenBase &other, + typename internal::enable_if::value, + PrivateType>::type = PrivateType()) : Base(other.derived()) { } diff --git a/xs/src/eigen/Eigen/src/Core/ArrayBase.h b/xs/src/eigen/Eigen/src/Core/ArrayBase.h index f0232f65e..3dbc7084c 100644 --- a/xs/src/eigen/Eigen/src/Core/ArrayBase.h +++ b/xs/src/eigen/Eigen/src/Core/ArrayBase.h @@ -175,7 +175,7 @@ template class ArrayBase */ template template -EIGEN_STRONG_INLINE Derived & +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived & ArrayBase::operator-=(const ArrayBase &other) { call_assignment(derived(), other.derived(), internal::sub_assign_op()); @@ -188,7 +188,7 @@ ArrayBase::operator-=(const ArrayBase &other) */ template template -EIGEN_STRONG_INLINE Derived & +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived & ArrayBase::operator+=(const ArrayBase& other) { call_assignment(derived(), other.derived(), internal::add_assign_op()); @@ -201,7 +201,7 @@ ArrayBase::operator+=(const ArrayBase& other) */ template template -EIGEN_STRONG_INLINE Derived & +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived & ArrayBase::operator*=(const ArrayBase& other) { call_assignment(derived(), other.derived(), internal::mul_assign_op()); @@ -214,7 +214,7 @@ ArrayBase::operator*=(const ArrayBase& other) */ template template -EIGEN_STRONG_INLINE Derived & +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived & ArrayBase::operator/=(const ArrayBase& other) { call_assignment(derived(), other.derived(), internal::div_assign_op()); diff --git a/xs/src/eigen/Eigen/src/Core/ArrayWrapper.h b/xs/src/eigen/Eigen/src/Core/ArrayWrapper.h index a04521a16..688aadd62 100644 --- a/xs/src/eigen/Eigen/src/Core/ArrayWrapper.h +++ b/xs/src/eigen/Eigen/src/Core/ArrayWrapper.h @@ -32,7 +32,8 @@ struct traits > // Let's remove NestByRefBit enum { Flags0 = traits::type >::Flags, - Flags = Flags0 & ~NestByRefBit + LvalueBitFlag = is_lvalue::value ? LvalueBit : 0, + Flags = (Flags0 & ~(NestByRefBit | LvalueBit)) | LvalueBitFlag }; }; } @@ -129,7 +130,8 @@ struct traits > // Let's remove NestByRefBit enum { Flags0 = traits::type >::Flags, - Flags = Flags0 & ~NestByRefBit + LvalueBitFlag = is_lvalue::value ? LvalueBit : 0, + Flags = (Flags0 & ~(NestByRefBit | LvalueBit)) | LvalueBitFlag }; }; } diff --git a/xs/src/eigen/Eigen/src/Core/AssignEvaluator.h b/xs/src/eigen/Eigen/src/Core/AssignEvaluator.h index b0ec7b7ca..dbe435d86 100644 --- a/xs/src/eigen/Eigen/src/Core/AssignEvaluator.h +++ b/xs/src/eigen/Eigen/src/Core/AssignEvaluator.h @@ -39,7 +39,7 @@ public: enum { DstAlignment = DstEvaluator::Alignment, SrcAlignment = SrcEvaluator::Alignment, - DstHasDirectAccess = DstFlags & DirectAccessBit, + DstHasDirectAccess = (DstFlags & DirectAccessBit) == DirectAccessBit, JointAlignment = EIGEN_PLAIN_ENUM_MIN(DstAlignment,SrcAlignment) }; @@ -83,7 +83,7 @@ private: && int(OuterStride)!=Dynamic && int(OuterStride)%int(InnerPacketSize)==0 && (EIGEN_UNALIGNED_VECTORIZE || int(JointAlignment)>=int(InnerRequiredAlignment)), MayLinearize = bool(StorageOrdersAgree) && (int(DstFlags) & int(SrcFlags) & LinearAccessBit), - MayLinearVectorize = bool(MightVectorize) && MayLinearize && DstHasDirectAccess + MayLinearVectorize = bool(MightVectorize) && bool(MayLinearize) && bool(DstHasDirectAccess) && (EIGEN_UNALIGNED_VECTORIZE || (int(DstAlignment)>=int(LinearRequiredAlignment)) || MaxSizeAtCompileTime == Dynamic), /* If the destination isn't aligned, we have to do runtime checks and we don't unroll, so it's only good for large enough sizes. */ diff --git a/xs/src/eigen/Eigen/src/Core/Assign_MKL.h b/xs/src/eigen/Eigen/src/Core/Assign_MKL.h index 6c2ab9264..6866095bf 100644 --- a/xs/src/eigen/Eigen/src/Core/Assign_MKL.h +++ b/xs/src/eigen/Eigen/src/Core/Assign_MKL.h @@ -84,7 +84,8 @@ class vml_assign_traits struct Assignment, SrcXprNested>, assign_op, \ Dense2Dense, typename enable_if::EnableVml>::type> { \ typedef CwiseUnaryOp, SrcXprNested> SrcXprType; \ - static void run(DstXprType &dst, const SrcXprType &src, const assign_op &/*func*/) { \ + static void run(DstXprType &dst, const SrcXprType &src, const assign_op &func) { \ + resize_if_allowed(dst, src, func); \ eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); \ if(vml_assign_traits::Traversal==LinearTraversal) { \ VMLOP(dst.size(), (const VMLTYPE*)src.nestedExpression().data(), \ @@ -144,7 +145,8 @@ EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(ceil, Ceil, _) Dense2Dense, typename enable_if::EnableVml>::type> { \ typedef CwiseBinaryOp, SrcXprNested, \ const CwiseNullaryOp,Plain> > SrcXprType; \ - static void run(DstXprType &dst, const SrcXprType &src, const assign_op &/*func*/) { \ + static void run(DstXprType &dst, const SrcXprType &src, const assign_op &func) { \ + resize_if_allowed(dst, src, func); \ eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); \ VMLTYPE exponent = reinterpret_cast(src.rhs().functor().m_other); \ if(vml_assign_traits::Traversal==LinearTraversal) \ diff --git a/xs/src/eigen/Eigen/src/Core/CoreEvaluators.h b/xs/src/eigen/Eigen/src/Core/CoreEvaluators.h index f7c1effca..910889efa 100644 --- a/xs/src/eigen/Eigen/src/Core/CoreEvaluators.h +++ b/xs/src/eigen/Eigen/src/Core/CoreEvaluators.h @@ -977,7 +977,7 @@ struct evaluator > OuterStrideAtCompileTime = HasSameStorageOrderAsArgType ? int(outer_stride_at_compile_time::ret) : int(inner_stride_at_compile_time::ret), - MaskPacketAccessBit = (InnerStrideAtCompileTime == 1) ? PacketAccessBit : 0, + MaskPacketAccessBit = (InnerStrideAtCompileTime == 1 || HasSameStorageOrderAsArgType) ? PacketAccessBit : 0, FlagsLinearAccessBit = (RowsAtCompileTime == 1 || ColsAtCompileTime == 1 || (InnerPanel && (evaluator::Flags&LinearAccessBit))) ? LinearAccessBit : 0, FlagsRowMajorBit = XprType::Flags&RowMajorBit, @@ -987,7 +987,9 @@ struct evaluator > Flags = Flags0 | FlagsLinearAccessBit | FlagsRowMajorBit, PacketAlignment = unpacket_traits::alignment, - Alignment0 = (InnerPanel && (OuterStrideAtCompileTime!=Dynamic) && (((OuterStrideAtCompileTime * int(sizeof(Scalar))) % int(PacketAlignment)) == 0)) ? int(PacketAlignment) : 0, + Alignment0 = (InnerPanel && (OuterStrideAtCompileTime!=Dynamic) + && (OuterStrideAtCompileTime!=0) + && (((OuterStrideAtCompileTime * int(sizeof(Scalar))) % int(PacketAlignment)) == 0)) ? int(PacketAlignment) : 0, Alignment = EIGEN_PLAIN_ENUM_MIN(evaluator::Alignment, Alignment0) }; typedef block_evaluator block_evaluator_type; @@ -1018,14 +1020,16 @@ struct unary_evaluator, IndexBa EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& block) : m_argImpl(block.nestedExpression()), m_startRow(block.startRow()), - m_startCol(block.startCol()) + m_startCol(block.startCol()), + m_linear_offset(InnerPanel?(XprType::IsRowMajor ? block.startRow()*block.cols() : block.startCol()*block.rows()):0) { } typedef typename XprType::Scalar Scalar; typedef typename XprType::CoeffReturnType CoeffReturnType; enum { - RowsAtCompileTime = XprType::RowsAtCompileTime + RowsAtCompileTime = XprType::RowsAtCompileTime, + ForwardLinearAccess = InnerPanel && bool(evaluator::Flags&LinearAccessBit) }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE @@ -1037,7 +1041,10 @@ struct unary_evaluator, IndexBa EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const { - return coeff(RowsAtCompileTime == 1 ? 0 : index, RowsAtCompileTime == 1 ? index : 0); + if (ForwardLinearAccess) + return m_argImpl.coeff(m_linear_offset.value() + index); + else + return coeff(RowsAtCompileTime == 1 ? 0 : index, RowsAtCompileTime == 1 ? index : 0); } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE @@ -1049,7 +1056,10 @@ struct unary_evaluator, IndexBa EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& coeffRef(Index index) { - return coeffRef(RowsAtCompileTime == 1 ? 0 : index, RowsAtCompileTime == 1 ? index : 0); + if (ForwardLinearAccess) + return m_argImpl.coeffRef(m_linear_offset.value() + index); + else + return coeffRef(RowsAtCompileTime == 1 ? 0 : index, RowsAtCompileTime == 1 ? index : 0); } template @@ -1063,8 +1073,11 @@ struct unary_evaluator, IndexBa EIGEN_STRONG_INLINE PacketType packet(Index index) const { - return packet(RowsAtCompileTime == 1 ? 0 : index, - RowsAtCompileTime == 1 ? index : 0); + if (ForwardLinearAccess) + return m_argImpl.template packet(m_linear_offset.value() + index); + else + return packet(RowsAtCompileTime == 1 ? 0 : index, + RowsAtCompileTime == 1 ? index : 0); } template @@ -1078,15 +1091,19 @@ struct unary_evaluator, IndexBa EIGEN_STRONG_INLINE void writePacket(Index index, const PacketType& x) { - return writePacket(RowsAtCompileTime == 1 ? 0 : index, - RowsAtCompileTime == 1 ? index : 0, - x); + if (ForwardLinearAccess) + return m_argImpl.template writePacket(m_linear_offset.value() + index, x); + else + return writePacket(RowsAtCompileTime == 1 ? 0 : index, + RowsAtCompileTime == 1 ? index : 0, + x); } protected: evaluator m_argImpl; const variable_if_dynamic m_startRow; const variable_if_dynamic m_startCol; + const variable_if_dynamic m_linear_offset; }; // TODO: This evaluator does not actually use the child evaluator; diff --git a/xs/src/eigen/Eigen/src/Core/CwiseNullaryOp.h b/xs/src/eigen/Eigen/src/Core/CwiseNullaryOp.h index dd498f758..ddd607e38 100644 --- a/xs/src/eigen/Eigen/src/Core/CwiseNullaryOp.h +++ b/xs/src/eigen/Eigen/src/Core/CwiseNullaryOp.h @@ -105,7 +105,7 @@ class CwiseNullaryOp : public internal::dense_xpr_base< CwiseNullaryOp template -EIGEN_STRONG_INLINE const CwiseNullaryOp::PlainObject> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseNullaryOp::PlainObject> DenseBase::NullaryExpr(Index rows, Index cols, const CustomNullaryOp& func) { return CwiseNullaryOp(rows, cols, func); @@ -150,7 +150,7 @@ DenseBase::NullaryExpr(Index size, const CustomNullaryOp& func) */ template template -EIGEN_STRONG_INLINE const CwiseNullaryOp::PlainObject> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseNullaryOp::PlainObject> DenseBase::NullaryExpr(const CustomNullaryOp& func) { return CwiseNullaryOp(RowsAtCompileTime, ColsAtCompileTime, func); @@ -192,7 +192,7 @@ DenseBase::Constant(Index rows, Index cols, const Scalar& value) * \sa class CwiseNullaryOp */ template -EIGEN_STRONG_INLINE const typename DenseBase::ConstantReturnType +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase::ConstantReturnType DenseBase::Constant(Index size, const Scalar& value) { return DenseBase::NullaryExpr(size, internal::scalar_constant_op(value)); @@ -208,7 +208,7 @@ DenseBase::Constant(Index size, const Scalar& value) * \sa class CwiseNullaryOp */ template -EIGEN_STRONG_INLINE const typename DenseBase::ConstantReturnType +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase::ConstantReturnType DenseBase::Constant(const Scalar& value) { EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived) @@ -220,7 +220,7 @@ DenseBase::Constant(const Scalar& value) * \sa LinSpaced(Index,Scalar,Scalar), setLinSpaced(Index,const Scalar&,const Scalar&) */ template -EIGEN_STRONG_INLINE const typename DenseBase::RandomAccessLinSpacedReturnType +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase::RandomAccessLinSpacedReturnType DenseBase::LinSpaced(Sequential_t, Index size, const Scalar& low, const Scalar& high) { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) @@ -232,7 +232,7 @@ DenseBase::LinSpaced(Sequential_t, Index size, const Scalar& low, const * \sa LinSpaced(Scalar,Scalar) */ template -EIGEN_STRONG_INLINE const typename DenseBase::RandomAccessLinSpacedReturnType +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase::RandomAccessLinSpacedReturnType DenseBase::LinSpaced(Sequential_t, const Scalar& low, const Scalar& high) { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) @@ -264,7 +264,7 @@ DenseBase::LinSpaced(Sequential_t, const Scalar& low, const Scalar& hig * \sa setLinSpaced(Index,const Scalar&,const Scalar&), CwiseNullaryOp */ template -EIGEN_STRONG_INLINE const typename DenseBase::RandomAccessLinSpacedReturnType +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase::RandomAccessLinSpacedReturnType DenseBase::LinSpaced(Index size, const Scalar& low, const Scalar& high) { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) @@ -276,7 +276,7 @@ DenseBase::LinSpaced(Index size, const Scalar& low, const Scalar& high) * Special version for fixed size types which does not require the size parameter. */ template -EIGEN_STRONG_INLINE const typename DenseBase::RandomAccessLinSpacedReturnType +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase::RandomAccessLinSpacedReturnType DenseBase::LinSpaced(const Scalar& low, const Scalar& high) { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) @@ -286,7 +286,7 @@ DenseBase::LinSpaced(const Scalar& low, const Scalar& high) /** \returns true if all coefficients in this matrix are approximately equal to \a val, to within precision \a prec */ template -bool DenseBase::isApproxToConstant +EIGEN_DEVICE_FUNC bool DenseBase::isApproxToConstant (const Scalar& val, const RealScalar& prec) const { typename internal::nested_eval::type self(derived()); @@ -301,7 +301,7 @@ bool DenseBase::isApproxToConstant * * \returns true if all coefficients in this matrix are approximately equal to \a value, to within precision \a prec */ template -bool DenseBase::isConstant +EIGEN_DEVICE_FUNC bool DenseBase::isConstant (const Scalar& val, const RealScalar& prec) const { return isApproxToConstant(val, prec); @@ -312,7 +312,7 @@ bool DenseBase::isConstant * \sa setConstant(), Constant(), class CwiseNullaryOp */ template -EIGEN_STRONG_INLINE void DenseBase::fill(const Scalar& val) +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void DenseBase::fill(const Scalar& val) { setConstant(val); } @@ -322,7 +322,7 @@ EIGEN_STRONG_INLINE void DenseBase::fill(const Scalar& val) * \sa fill(), setConstant(Index,const Scalar&), setConstant(Index,Index,const Scalar&), setZero(), setOnes(), Constant(), class CwiseNullaryOp, setZero(), setOnes() */ template -EIGEN_STRONG_INLINE Derived& DenseBase::setConstant(const Scalar& val) +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase::setConstant(const Scalar& val) { return derived() = Constant(rows(), cols(), val); } @@ -337,7 +337,7 @@ EIGEN_STRONG_INLINE Derived& DenseBase::setConstant(const Scalar& val) * \sa MatrixBase::setConstant(const Scalar&), setConstant(Index,Index,const Scalar&), class CwiseNullaryOp, MatrixBase::Constant(const Scalar&) */ template -EIGEN_STRONG_INLINE Derived& +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& PlainObjectBase::setConstant(Index size, const Scalar& val) { resize(size); @@ -356,7 +356,7 @@ PlainObjectBase::setConstant(Index size, const Scalar& val) * \sa MatrixBase::setConstant(const Scalar&), setConstant(Index,const Scalar&), class CwiseNullaryOp, MatrixBase::Constant(const Scalar&) */ template -EIGEN_STRONG_INLINE Derived& +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& PlainObjectBase::setConstant(Index rows, Index cols, const Scalar& val) { resize(rows, cols); @@ -380,7 +380,7 @@ PlainObjectBase::setConstant(Index rows, Index cols, const Scalar& val) * \sa LinSpaced(Index,const Scalar&,const Scalar&), CwiseNullaryOp */ template -EIGEN_STRONG_INLINE Derived& DenseBase::setLinSpaced(Index newSize, const Scalar& low, const Scalar& high) +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase::setLinSpaced(Index newSize, const Scalar& low, const Scalar& high) { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) return derived() = Derived::NullaryExpr(newSize, internal::linspaced_op(low,high,newSize)); @@ -400,7 +400,7 @@ EIGEN_STRONG_INLINE Derived& DenseBase::setLinSpaced(Index newSize, con * \sa LinSpaced(Index,const Scalar&,const Scalar&), setLinSpaced(Index, const Scalar&, const Scalar&), CwiseNullaryOp */ template -EIGEN_STRONG_INLINE Derived& DenseBase::setLinSpaced(const Scalar& low, const Scalar& high) +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase::setLinSpaced(const Scalar& low, const Scalar& high) { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) return setLinSpaced(size(), low, high); @@ -423,7 +423,7 @@ EIGEN_STRONG_INLINE Derived& DenseBase::setLinSpaced(const Scalar& low, * \sa Zero(), Zero(Index) */ template -EIGEN_STRONG_INLINE const typename DenseBase::ConstantReturnType +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase::ConstantReturnType DenseBase::Zero(Index rows, Index cols) { return Constant(rows, cols, Scalar(0)); @@ -446,7 +446,7 @@ DenseBase::Zero(Index rows, Index cols) * \sa Zero(), Zero(Index,Index) */ template -EIGEN_STRONG_INLINE const typename DenseBase::ConstantReturnType +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase::ConstantReturnType DenseBase::Zero(Index size) { return Constant(size, Scalar(0)); @@ -463,7 +463,7 @@ DenseBase::Zero(Index size) * \sa Zero(Index), Zero(Index,Index) */ template -EIGEN_STRONG_INLINE const typename DenseBase::ConstantReturnType +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase::ConstantReturnType DenseBase::Zero() { return Constant(Scalar(0)); @@ -478,7 +478,7 @@ DenseBase::Zero() * \sa class CwiseNullaryOp, Zero() */ template -bool DenseBase::isZero(const RealScalar& prec) const +EIGEN_DEVICE_FUNC bool DenseBase::isZero(const RealScalar& prec) const { typename internal::nested_eval::type self(derived()); for(Index j = 0; j < cols(); ++j) @@ -496,7 +496,7 @@ bool DenseBase::isZero(const RealScalar& prec) const * \sa class CwiseNullaryOp, Zero() */ template -EIGEN_STRONG_INLINE Derived& DenseBase::setZero() +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase::setZero() { return setConstant(Scalar(0)); } @@ -511,7 +511,7 @@ EIGEN_STRONG_INLINE Derived& DenseBase::setZero() * \sa DenseBase::setZero(), setZero(Index,Index), class CwiseNullaryOp, DenseBase::Zero() */ template -EIGEN_STRONG_INLINE Derived& +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& PlainObjectBase::setZero(Index newSize) { resize(newSize); @@ -529,7 +529,7 @@ PlainObjectBase::setZero(Index newSize) * \sa DenseBase::setZero(), setZero(Index), class CwiseNullaryOp, DenseBase::Zero() */ template -EIGEN_STRONG_INLINE Derived& +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& PlainObjectBase::setZero(Index rows, Index cols) { resize(rows, cols); @@ -553,7 +553,7 @@ PlainObjectBase::setZero(Index rows, Index cols) * \sa Ones(), Ones(Index), isOnes(), class Ones */ template -EIGEN_STRONG_INLINE const typename DenseBase::ConstantReturnType +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase::ConstantReturnType DenseBase::Ones(Index rows, Index cols) { return Constant(rows, cols, Scalar(1)); @@ -576,7 +576,7 @@ DenseBase::Ones(Index rows, Index cols) * \sa Ones(), Ones(Index,Index), isOnes(), class Ones */ template -EIGEN_STRONG_INLINE const typename DenseBase::ConstantReturnType +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase::ConstantReturnType DenseBase::Ones(Index newSize) { return Constant(newSize, Scalar(1)); @@ -593,7 +593,7 @@ DenseBase::Ones(Index newSize) * \sa Ones(Index), Ones(Index,Index), isOnes(), class Ones */ template -EIGEN_STRONG_INLINE const typename DenseBase::ConstantReturnType +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase::ConstantReturnType DenseBase::Ones() { return Constant(Scalar(1)); @@ -608,7 +608,7 @@ DenseBase::Ones() * \sa class CwiseNullaryOp, Ones() */ template -bool DenseBase::isOnes +EIGEN_DEVICE_FUNC bool DenseBase::isOnes (const RealScalar& prec) const { return isApproxToConstant(Scalar(1), prec); @@ -622,7 +622,7 @@ bool DenseBase::isOnes * \sa class CwiseNullaryOp, Ones() */ template -EIGEN_STRONG_INLINE Derived& DenseBase::setOnes() +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase::setOnes() { return setConstant(Scalar(1)); } @@ -637,7 +637,7 @@ EIGEN_STRONG_INLINE Derived& DenseBase::setOnes() * \sa MatrixBase::setOnes(), setOnes(Index,Index), class CwiseNullaryOp, MatrixBase::Ones() */ template -EIGEN_STRONG_INLINE Derived& +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& PlainObjectBase::setOnes(Index newSize) { resize(newSize); @@ -655,7 +655,7 @@ PlainObjectBase::setOnes(Index newSize) * \sa MatrixBase::setOnes(), setOnes(Index), class CwiseNullaryOp, MatrixBase::Ones() */ template -EIGEN_STRONG_INLINE Derived& +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& PlainObjectBase::setOnes(Index rows, Index cols) { resize(rows, cols); @@ -679,7 +679,7 @@ PlainObjectBase::setOnes(Index rows, Index cols) * \sa Identity(), setIdentity(), isIdentity() */ template -EIGEN_STRONG_INLINE const typename MatrixBase::IdentityReturnType +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase::IdentityReturnType MatrixBase::Identity(Index rows, Index cols) { return DenseBase::NullaryExpr(rows, cols, internal::scalar_identity_op()); @@ -696,7 +696,7 @@ MatrixBase::Identity(Index rows, Index cols) * \sa Identity(Index,Index), setIdentity(), isIdentity() */ template -EIGEN_STRONG_INLINE const typename MatrixBase::IdentityReturnType +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase::IdentityReturnType MatrixBase::Identity() { EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived) @@ -771,7 +771,7 @@ struct setIdentity_impl * \sa class CwiseNullaryOp, Identity(), Identity(Index,Index), isIdentity() */ template -EIGEN_STRONG_INLINE Derived& MatrixBase::setIdentity() +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& MatrixBase::setIdentity() { return internal::setIdentity_impl::run(derived()); } @@ -787,7 +787,7 @@ EIGEN_STRONG_INLINE Derived& MatrixBase::setIdentity() * \sa MatrixBase::setIdentity(), class CwiseNullaryOp, MatrixBase::Identity() */ template -EIGEN_STRONG_INLINE Derived& MatrixBase::setIdentity(Index rows, Index cols) +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& MatrixBase::setIdentity(Index rows, Index cols) { derived().resize(rows, cols); return setIdentity(); @@ -800,7 +800,7 @@ EIGEN_STRONG_INLINE Derived& MatrixBase::setIdentity(Index rows, Index * \sa MatrixBase::Unit(Index), MatrixBase::UnitX(), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW() */ template -EIGEN_STRONG_INLINE const typename MatrixBase::BasisReturnType MatrixBase::Unit(Index newSize, Index i) +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase::BasisReturnType MatrixBase::Unit(Index newSize, Index i) { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) return BasisReturnType(SquareMatrixType::Identity(newSize,newSize), i); @@ -815,7 +815,7 @@ EIGEN_STRONG_INLINE const typename MatrixBase::BasisReturnType MatrixBa * \sa MatrixBase::Unit(Index,Index), MatrixBase::UnitX(), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW() */ template -EIGEN_STRONG_INLINE const typename MatrixBase::BasisReturnType MatrixBase::Unit(Index i) +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase::BasisReturnType MatrixBase::Unit(Index i) { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) return BasisReturnType(SquareMatrixType::Identity(),i); @@ -828,7 +828,7 @@ EIGEN_STRONG_INLINE const typename MatrixBase::BasisReturnType MatrixBa * \sa MatrixBase::Unit(Index,Index), MatrixBase::Unit(Index), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW() */ template -EIGEN_STRONG_INLINE const typename MatrixBase::BasisReturnType MatrixBase::UnitX() +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase::BasisReturnType MatrixBase::UnitX() { return Derived::Unit(0); } /** \returns an expression of the Y axis unit vector (0,1{,0}^*) @@ -838,7 +838,7 @@ EIGEN_STRONG_INLINE const typename MatrixBase::BasisReturnType MatrixBa * \sa MatrixBase::Unit(Index,Index), MatrixBase::Unit(Index), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW() */ template -EIGEN_STRONG_INLINE const typename MatrixBase::BasisReturnType MatrixBase::UnitY() +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase::BasisReturnType MatrixBase::UnitY() { return Derived::Unit(1); } /** \returns an expression of the Z axis unit vector (0,0,1{,0}^*) @@ -848,7 +848,7 @@ EIGEN_STRONG_INLINE const typename MatrixBase::BasisReturnType MatrixBa * \sa MatrixBase::Unit(Index,Index), MatrixBase::Unit(Index), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW() */ template -EIGEN_STRONG_INLINE const typename MatrixBase::BasisReturnType MatrixBase::UnitZ() +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase::BasisReturnType MatrixBase::UnitZ() { return Derived::Unit(2); } /** \returns an expression of the W axis unit vector (0,0,0,1) @@ -858,7 +858,7 @@ EIGEN_STRONG_INLINE const typename MatrixBase::BasisReturnType MatrixBa * \sa MatrixBase::Unit(Index,Index), MatrixBase::Unit(Index), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW() */ template -EIGEN_STRONG_INLINE const typename MatrixBase::BasisReturnType MatrixBase::UnitW() +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase::BasisReturnType MatrixBase::UnitW() { return Derived::Unit(3); } } // end namespace Eigen diff --git a/xs/src/eigen/Eigen/src/Core/DenseBase.h b/xs/src/eigen/Eigen/src/Core/DenseBase.h index 46fe5193c..90066ae73 100644 --- a/xs/src/eigen/Eigen/src/Core/DenseBase.h +++ b/xs/src/eigen/Eigen/src/Core/DenseBase.h @@ -296,7 +296,7 @@ template class DenseBase EIGEN_DEVICE_FUNC Derived& operator=(const ReturnByValue& func); - /** \ínternal + /** \internal * Copies \a other into *this without evaluating other. \returns a reference to *this. * \deprecated */ template @@ -484,9 +484,9 @@ template class DenseBase return derived().coeff(0,0); } - bool all() const; - bool any() const; - Index count() const; + EIGEN_DEVICE_FUNC bool all() const; + EIGEN_DEVICE_FUNC bool any() const; + EIGEN_DEVICE_FUNC Index count() const; typedef VectorwiseOp RowwiseReturnType; typedef const VectorwiseOp ConstRowwiseReturnType; diff --git a/xs/src/eigen/Eigen/src/Core/Diagonal.h b/xs/src/eigen/Eigen/src/Core/Diagonal.h index 49e711257..afcaf3575 100644 --- a/xs/src/eigen/Eigen/src/Core/Diagonal.h +++ b/xs/src/eigen/Eigen/src/Core/Diagonal.h @@ -70,7 +70,10 @@ template class Diagonal EIGEN_DENSE_PUBLIC_INTERFACE(Diagonal) EIGEN_DEVICE_FUNC - explicit inline Diagonal(MatrixType& matrix, Index a_index = DiagIndex) : m_matrix(matrix), m_index(a_index) {} + explicit inline Diagonal(MatrixType& matrix, Index a_index = DiagIndex) : m_matrix(matrix), m_index(a_index) + { + eigen_assert( a_index <= m_matrix.cols() && -a_index <= m_matrix.rows() ); + } EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Diagonal) diff --git a/xs/src/eigen/Eigen/src/Core/Dot.h b/xs/src/eigen/Eigen/src/Core/Dot.h index 06ef18b8b..1fe7a84a4 100644 --- a/xs/src/eigen/Eigen/src/Core/Dot.h +++ b/xs/src/eigen/Eigen/src/Core/Dot.h @@ -31,7 +31,8 @@ struct dot_nocheck typedef scalar_conj_product_op::Scalar,typename traits::Scalar> conj_prod; typedef typename conj_prod::result_type ResScalar; EIGEN_DEVICE_FUNC - static inline ResScalar run(const MatrixBase& a, const MatrixBase& b) + EIGEN_STRONG_INLINE + static ResScalar run(const MatrixBase& a, const MatrixBase& b) { return a.template binaryExpr(b).sum(); } @@ -43,7 +44,8 @@ struct dot_nocheck typedef scalar_conj_product_op::Scalar,typename traits::Scalar> conj_prod; typedef typename conj_prod::result_type ResScalar; EIGEN_DEVICE_FUNC - static inline ResScalar run(const MatrixBase& a, const MatrixBase& b) + EIGEN_STRONG_INLINE + static ResScalar run(const MatrixBase& a, const MatrixBase& b) { return a.transpose().template binaryExpr(b).sum(); } @@ -65,6 +67,7 @@ struct dot_nocheck template template EIGEN_DEVICE_FUNC +EIGEN_STRONG_INLINE typename ScalarBinaryOpTraits::Scalar,typename internal::traits::Scalar>::ReturnType MatrixBase::dot(const MatrixBase& other) const { @@ -102,7 +105,7 @@ EIGEN_STRONG_INLINE typename NumTraits::Scala * \sa lpNorm(), dot(), squaredNorm() */ template -inline typename NumTraits::Scalar>::Real MatrixBase::norm() const +EIGEN_STRONG_INLINE typename NumTraits::Scalar>::Real MatrixBase::norm() const { return numext::sqrt(squaredNorm()); } @@ -117,7 +120,7 @@ inline typename NumTraits::Scalar>::Real Matr * \sa norm(), normalize() */ template -inline const typename MatrixBase::PlainObject +EIGEN_STRONG_INLINE const typename MatrixBase::PlainObject MatrixBase::normalized() const { typedef typename internal::nested_eval::type _Nested; @@ -139,7 +142,7 @@ MatrixBase::normalized() const * \sa norm(), normalized() */ template -inline void MatrixBase::normalize() +EIGEN_STRONG_INLINE void MatrixBase::normalize() { RealScalar z = squaredNorm(); // NOTE: after extensive benchmarking, this conditional does not impact performance, at least on recent x86 CPU @@ -160,7 +163,7 @@ inline void MatrixBase::normalize() * \sa stableNorm(), stableNormalize(), normalized() */ template -inline const typename MatrixBase::PlainObject +EIGEN_STRONG_INLINE const typename MatrixBase::PlainObject MatrixBase::stableNormalized() const { typedef typename internal::nested_eval::type _Nested; @@ -185,7 +188,7 @@ MatrixBase::stableNormalized() const * \sa stableNorm(), stableNormalized(), normalize() */ template -inline void MatrixBase::stableNormalize() +EIGEN_STRONG_INLINE void MatrixBase::stableNormalize() { RealScalar w = cwiseAbs().maxCoeff(); RealScalar z = (derived()/w).squaredNorm(); diff --git a/xs/src/eigen/Eigen/src/Core/EigenBase.h b/xs/src/eigen/Eigen/src/Core/EigenBase.h index f76995af9..b195506a9 100644 --- a/xs/src/eigen/Eigen/src/Core/EigenBase.h +++ b/xs/src/eigen/Eigen/src/Core/EigenBase.h @@ -14,6 +14,7 @@ namespace Eigen { /** \class EigenBase + * \ingroup Core_Module * * Common base class for all classes T such that MatrixBase has an operator=(T) and a constructor MatrixBase(T). * @@ -128,6 +129,7 @@ template struct EigenBase */ template template +EIGEN_DEVICE_FUNC Derived& DenseBase::operator=(const EigenBase &other) { call_assignment(derived(), other.derived()); @@ -136,6 +138,7 @@ Derived& DenseBase::operator=(const EigenBase &other) template template +EIGEN_DEVICE_FUNC Derived& DenseBase::operator+=(const EigenBase &other) { call_assignment(derived(), other.derived(), internal::add_assign_op()); @@ -144,6 +147,7 @@ Derived& DenseBase::operator+=(const EigenBase &other) template template +EIGEN_DEVICE_FUNC Derived& DenseBase::operator-=(const EigenBase &other) { call_assignment(derived(), other.derived(), internal::sub_assign_op()); diff --git a/xs/src/eigen/Eigen/src/Core/GeneralProduct.h b/xs/src/eigen/Eigen/src/Core/GeneralProduct.h index 0f16cd8e3..6f0cc80e9 100644 --- a/xs/src/eigen/Eigen/src/Core/GeneralProduct.h +++ b/xs/src/eigen/Eigen/src/Core/GeneralProduct.h @@ -24,12 +24,17 @@ template struct product_type_selector; template struct product_size_category { - enum { is_large = MaxSize == Dynamic || - Size >= EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD || - (Size==Dynamic && MaxSize>=EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD), - value = is_large ? Large - : Size == 1 ? 1 - : Small + enum { + #ifndef EIGEN_CUDA_ARCH + is_large = MaxSize == Dynamic || + Size >= EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD || + (Size==Dynamic && MaxSize>=EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD), + #else + is_large = 0, + #endif + value = is_large ? Large + : Size == 1 ? 1 + : Small }; }; @@ -379,8 +384,6 @@ template<> struct gemv_dense_selector * * \sa lazyProduct(), operator*=(const MatrixBase&), Cwise::operator*() */ -#ifndef __CUDACC__ - template template inline const Product @@ -412,8 +415,6 @@ MatrixBase::operator*(const MatrixBase &other) const return Product(derived(), other.derived()); } -#endif // __CUDACC__ - /** \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 diff --git a/xs/src/eigen/Eigen/src/Core/GenericPacketMath.h b/xs/src/eigen/Eigen/src/Core/GenericPacketMath.h index 27033a2dd..029f8ac36 100644 --- a/xs/src/eigen/Eigen/src/Core/GenericPacketMath.h +++ b/xs/src/eigen/Eigen/src/Core/GenericPacketMath.h @@ -230,7 +230,7 @@ pload1(const typename unpacket_traits::type *a) { return pset1( * duplicated to form: {from[0],from[0],from[1],from[1],from[2],from[2],from[3],from[3]} * Currently, this function is only used for scalar * complex products. */ -template EIGEN_DEVICE_FUNC inline Packet +template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet ploaddup(const typename unpacket_traits::type* from) { return *from; } /** \internal \returns a packet with elements of \a *from quadrupled. @@ -278,7 +278,7 @@ inline void pbroadcast2(const typename unpacket_traits::type *a, } /** \internal \brief Returns a packet with coefficients (a,a+1,...,a+packet_size-1). */ -template inline Packet +template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet plset(const typename unpacket_traits::type& a) { return a; } /** \internal copy the packet \a from to \a *to, \a to must be 16 bytes aligned */ @@ -482,7 +482,7 @@ EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void pstoret(Scalar* to, const Packet& fro * by the current computation. */ template -inline Packet ploadt_ro(const typename unpacket_traits::type* from) +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet ploadt_ro(const typename unpacket_traits::type* from) { return ploadt(from); } diff --git a/xs/src/eigen/Eigen/src/Core/Map.h b/xs/src/eigen/Eigen/src/Core/Map.h index 06d196702..548bf9a2d 100644 --- a/xs/src/eigen/Eigen/src/Core/Map.h +++ b/xs/src/eigen/Eigen/src/Core/Map.h @@ -20,11 +20,17 @@ struct traits > { typedef traits TraitsBase; enum { + PlainObjectTypeInnerSize = ((traits::Flags&RowMajorBit)==RowMajorBit) + ? PlainObjectType::ColsAtCompileTime + : PlainObjectType::RowsAtCompileTime, + InnerStrideAtCompileTime = StrideType::InnerStrideAtCompileTime == 0 ? int(PlainObjectType::InnerStrideAtCompileTime) : int(StrideType::InnerStrideAtCompileTime), OuterStrideAtCompileTime = StrideType::OuterStrideAtCompileTime == 0 - ? int(PlainObjectType::OuterStrideAtCompileTime) + ? (InnerStrideAtCompileTime==Dynamic || PlainObjectTypeInnerSize==Dynamic + ? Dynamic + : int(InnerStrideAtCompileTime) * int(PlainObjectTypeInnerSize)) : int(StrideType::OuterStrideAtCompileTime), Alignment = int(MapOptions)&int(AlignedMask), Flags0 = TraitsBase::Flags & (~NestByRefBit), @@ -107,10 +113,11 @@ template class Ma EIGEN_DEVICE_FUNC inline Index outerStride() const { - return StrideType::OuterStrideAtCompileTime != 0 ? m_stride.outer() - : IsVectorAtCompileTime ? this->size() - : int(Flags)&RowMajorBit ? this->cols() - : this->rows(); + return int(StrideType::OuterStrideAtCompileTime) != 0 ? m_stride.outer() + : int(internal::traits::OuterStrideAtCompileTime) != Dynamic ? Index(internal::traits::OuterStrideAtCompileTime) + : IsVectorAtCompileTime ? (this->size() * innerStride()) + : (int(Flags)&RowMajorBit) ? (this->cols() * innerStride()) + : (this->rows() * innerStride()); } /** Constructor in the fixed-size case. diff --git a/xs/src/eigen/Eigen/src/Core/MathFunctions.h b/xs/src/eigen/Eigen/src/Core/MathFunctions.h index 8d47fb8a4..6eb974d41 100644 --- a/xs/src/eigen/Eigen/src/Core/MathFunctions.h +++ b/xs/src/eigen/Eigen/src/Core/MathFunctions.h @@ -348,31 +348,7 @@ struct norm1_retval * Implementation of hypot * ****************************************************************************/ -template -struct hypot_impl -{ - typedef typename NumTraits::Real RealScalar; - static inline RealScalar run(const Scalar& x, const Scalar& y) - { - EIGEN_USING_STD_MATH(abs); - EIGEN_USING_STD_MATH(sqrt); - RealScalar _x = abs(x); - RealScalar _y = abs(y); - Scalar p, qp; - if(_x>_y) - { - p = _x; - qp = _y / p; - } - else - { - p = _y; - qp = _x / p; - } - if(p==RealScalar(0)) return RealScalar(0); - return p * sqrt(RealScalar(1) + qp*qp); - } -}; +template struct hypot_impl; template struct hypot_retval @@ -495,7 +471,7 @@ namespace std_fallback { typedef typename NumTraits::Real RealScalar; EIGEN_USING_STD_MATH(log); Scalar x1p = RealScalar(1) + x; - return ( x1p == Scalar(1) ) ? x : x * ( log(x1p) / (x1p - RealScalar(1)) ); + return numext::equal_strict(x1p, Scalar(1)) ? x : x * ( log(x1p) / (x1p - RealScalar(1)) ); } } @@ -1061,11 +1037,24 @@ double log(const double &x) { return ::log(x); } template EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE -typename NumTraits::Real abs(const T &x) { +typename internal::enable_if::IsSigned || NumTraits::IsComplex,typename NumTraits::Real>::type +abs(const T &x) { EIGEN_USING_STD_MATH(abs); return abs(x); } +template +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +typename internal::enable_if::IsSigned || NumTraits::IsComplex),typename NumTraits::Real>::type +abs(const T &x) { + return x; +} + +#if defined(__SYCL_DEVICE_ONLY__) +EIGEN_ALWAYS_INLINE float abs(float x) { return cl::sycl::fabs(x); } +EIGEN_ALWAYS_INLINE double abs(double x) { return cl::sycl::fabs(x); } +#endif // defined(__SYCL_DEVICE_ONLY__) + #ifdef __CUDACC__ template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE float abs(const float &x) { return ::fabsf(x); } diff --git a/xs/src/eigen/Eigen/src/Core/MathFunctionsImpl.h b/xs/src/eigen/Eigen/src/Core/MathFunctionsImpl.h index 3c9ef22fa..9c1ceb0eb 100644 --- a/xs/src/eigen/Eigen/src/Core/MathFunctionsImpl.h +++ b/xs/src/eigen/Eigen/src/Core/MathFunctionsImpl.h @@ -71,6 +71,29 @@ T generic_fast_tanh_float(const T& a_x) return pdiv(p, q); } +template +EIGEN_STRONG_INLINE +RealScalar positive_real_hypot(const RealScalar& x, const RealScalar& y) +{ + EIGEN_USING_STD_MATH(sqrt); + RealScalar p, qp; + p = numext::maxi(x,y); + if(p==RealScalar(0)) return RealScalar(0); + qp = numext::mini(y,x) / p; + return p * sqrt(RealScalar(1) + qp*qp); +} + +template +struct hypot_impl +{ + typedef typename NumTraits::Real RealScalar; + static inline RealScalar run(const Scalar& x, const Scalar& y) + { + EIGEN_USING_STD_MATH(abs); + return positive_real_hypot(abs(x), abs(y)); + } +}; + } // end namespace internal } // end namespace Eigen diff --git a/xs/src/eigen/Eigen/src/Core/MatrixBase.h b/xs/src/eigen/Eigen/src/Core/MatrixBase.h index f7cf04cde..05db48813 100644 --- a/xs/src/eigen/Eigen/src/Core/MatrixBase.h +++ b/xs/src/eigen/Eigen/src/Core/MatrixBase.h @@ -160,20 +160,11 @@ template class MatrixBase EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& operator-=(const MatrixBase& other); -#ifdef __CUDACC__ template EIGEN_DEVICE_FUNC - const Product - operator*(const MatrixBase &other) const - { return this->lazyProduct(other); } -#else - - template const Product operator*(const MatrixBase &other) const; -#endif - template EIGEN_DEVICE_FUNC const Product @@ -294,7 +285,7 @@ template class MatrixBase * fuzzy comparison such as isApprox() * \sa isApprox(), operator!= */ template - inline bool operator==(const MatrixBase& other) const + EIGEN_DEVICE_FUNC inline bool operator==(const MatrixBase& other) const { return cwiseEqual(other).all(); } /** \returns true if at least one pair of coefficients of \c *this and \a other are not exactly equal to each other. @@ -302,7 +293,7 @@ template class MatrixBase * fuzzy comparison such as isApprox() * \sa isApprox(), operator== */ template - inline bool operator!=(const MatrixBase& other) const + EIGEN_DEVICE_FUNC inline bool operator!=(const MatrixBase& other) const { return cwiseNotEqual(other).any(); } NoAlias noalias(); diff --git a/xs/src/eigen/Eigen/src/Core/NumTraits.h b/xs/src/eigen/Eigen/src/Core/NumTraits.h index dd61195bc..daf489878 100644 --- a/xs/src/eigen/Eigen/src/Core/NumTraits.h +++ b/xs/src/eigen/Eigen/src/Core/NumTraits.h @@ -215,6 +215,8 @@ struct NumTraits > static inline RealScalar epsilon() { return NumTraits::epsilon(); } EIGEN_DEVICE_FUNC static inline RealScalar dummy_precision() { return NumTraits::dummy_precision(); } + + static inline int digits10() { return NumTraits::digits10(); } }; template<> struct NumTraits diff --git a/xs/src/eigen/Eigen/src/Core/PlainObjectBase.h b/xs/src/eigen/Eigen/src/Core/PlainObjectBase.h index 77f4f6066..1dc7e223a 100644 --- a/xs/src/eigen/Eigen/src/Core/PlainObjectBase.h +++ b/xs/src/eigen/Eigen/src/Core/PlainObjectBase.h @@ -577,6 +577,10 @@ class PlainObjectBase : public internal::dense_xpr_base::type * while the AlignedMap() functions return aligned Map objects and thus should be called only with 16-byte-aligned * \a data pointers. * + * Here is an example using strides: + * \include Matrix_Map_stride.cpp + * Output: \verbinclude Matrix_Map_stride.out + * * \see class Map */ //@{ diff --git a/xs/src/eigen/Eigen/src/Core/Product.h b/xs/src/eigen/Eigen/src/Core/Product.h index ae0c94b38..676c48027 100644 --- a/xs/src/eigen/Eigen/src/Core/Product.h +++ b/xs/src/eigen/Eigen/src/Core/Product.h @@ -97,8 +97,8 @@ class Product : public ProductImpl<_Lhs,_Rhs,Option, && "if you wanted a coeff-wise or a dot product use the respective explicit functions"); } - EIGEN_DEVICE_FUNC inline Index rows() const { return m_lhs.rows(); } - EIGEN_DEVICE_FUNC inline Index cols() const { return m_rhs.cols(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index rows() const { return m_lhs.rows(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index cols() const { return m_rhs.cols(); } EIGEN_DEVICE_FUNC const LhsNestedCleaned& lhs() const { return m_lhs; } EIGEN_DEVICE_FUNC const RhsNestedCleaned& rhs() const { return m_rhs; } @@ -127,7 +127,7 @@ public: using Base::derived; typedef typename Base::Scalar Scalar; - operator const Scalar() const + EIGEN_STRONG_INLINE operator const Scalar() const { return internal::evaluator(derived()).coeff(0,0); } @@ -162,7 +162,7 @@ class ProductImpl public: - EIGEN_DEVICE_FUNC Scalar coeff(Index row, Index col) const + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar coeff(Index row, Index col) const { EIGEN_STATIC_ASSERT(EnableCoeff, THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS); eigen_assert( (Option==LazyProduct) || (this->rows() == 1 && this->cols() == 1) ); @@ -170,7 +170,7 @@ class ProductImpl return internal::evaluator(derived()).coeff(row,col); } - EIGEN_DEVICE_FUNC Scalar coeff(Index i) const + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar coeff(Index i) const { EIGEN_STATIC_ASSERT(EnableCoeff, THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS); eigen_assert( (Option==LazyProduct) || (this->rows() == 1 && this->cols() == 1) ); diff --git a/xs/src/eigen/Eigen/src/Core/ProductEvaluators.h b/xs/src/eigen/Eigen/src/Core/ProductEvaluators.h index 583b7f59e..9b99bd769 100644 --- a/xs/src/eigen/Eigen/src/Core/ProductEvaluators.h +++ b/xs/src/eigen/Eigen/src/Core/ProductEvaluators.h @@ -32,7 +32,7 @@ struct evaluator > typedef Product XprType; typedef product_evaluator Base; - EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr) : Base(xpr) {} + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit evaluator(const XprType& xpr) : Base(xpr) {} }; // Catch "scalar * ( A * B )" and transform it to "(A*scalar) * B" @@ -55,7 +55,7 @@ struct evaluator, const Product > XprType; typedef evaluator > Base; - EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit evaluator(const XprType& xpr) : Base(xpr.lhs().functor().m_other * xpr.rhs().lhs() * xpr.rhs().rhs()) {} }; @@ -68,7 +68,7 @@ struct evaluator, DiagIndex> > typedef Diagonal, DiagIndex> XprType; typedef evaluator, DiagIndex> > Base; - EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit evaluator(const XprType& xpr) : Base(Diagonal, DiagIndex>( Product(xpr.nestedExpression().lhs(), xpr.nestedExpression().rhs()), xpr.index() )) @@ -207,6 +207,12 @@ struct evaluator_assume_aliasing +struct evaluator_assume_aliasing::Scalar>, const OtherXpr, + const Product >, DenseShape > { + static const bool value = true; +}; + template struct assignment_from_xpr_op_product { @@ -240,19 +246,19 @@ template struct generic_product_impl { template - static inline void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + static EIGEN_STRONG_INLINE void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) { dst.coeffRef(0,0) = (lhs.transpose().cwiseProduct(rhs)).sum(); } template - static inline void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + static EIGEN_STRONG_INLINE void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) { dst.coeffRef(0,0) += (lhs.transpose().cwiseProduct(rhs)).sum(); } template - static void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + static EIGEN_STRONG_INLINE void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) { dst.coeffRef(0,0) -= (lhs.transpose().cwiseProduct(rhs)).sum(); } }; @@ -306,25 +312,25 @@ struct generic_product_impl }; template - static inline void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + static EIGEN_STRONG_INLINE void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) { internal::outer_product_selector_run(dst, lhs, rhs, set(), is_row_major()); } template - static inline void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + static EIGEN_STRONG_INLINE void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) { internal::outer_product_selector_run(dst, lhs, rhs, add(), is_row_major()); } template - static inline void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + static EIGEN_STRONG_INLINE void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) { internal::outer_product_selector_run(dst, lhs, rhs, sub(), is_row_major()); } template - static inline void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) + static EIGEN_STRONG_INLINE void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) { internal::outer_product_selector_run(dst, lhs, rhs, adds(alpha), is_row_major()); } @@ -779,7 +785,11 @@ public: _Vectorizable = bool(int(MatrixFlags)&PacketAccessBit) && _SameTypes && (_ScalarAccessOnDiag || (bool(int(DiagFlags)&PacketAccessBit))), _LinearAccessMask = (MatrixType::RowsAtCompileTime==1 || MatrixType::ColsAtCompileTime==1) ? LinearAccessBit : 0, Flags = ((HereditaryBits|_LinearAccessMask) & (unsigned int)(MatrixFlags)) | (_Vectorizable ? PacketAccessBit : 0), - Alignment = evaluator::Alignment + Alignment = evaluator::Alignment, + + AsScalarProduct = (DiagonalType::SizeAtCompileTime==1) + || (DiagonalType::SizeAtCompileTime==Dynamic && MatrixType::RowsAtCompileTime==1 && ProductOrder==OnTheLeft) + || (DiagonalType::SizeAtCompileTime==Dynamic && MatrixType::ColsAtCompileTime==1 && ProductOrder==OnTheRight) }; diagonal_product_evaluator_base(const MatrixType &mat, const DiagonalType &diag) @@ -791,7 +801,10 @@ public: EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index idx) const { - return m_diagImpl.coeff(idx) * m_matImpl.coeff(idx); + if(AsScalarProduct) + return m_diagImpl.coeff(0) * m_matImpl.coeff(idx); + else + return m_diagImpl.coeff(idx) * m_matImpl.coeff(idx); } protected: diff --git a/xs/src/eigen/Eigen/src/Core/Redux.h b/xs/src/eigen/Eigen/src/Core/Redux.h index b6e8f8887..760e9f861 100644 --- a/xs/src/eigen/Eigen/src/Core/Redux.h +++ b/xs/src/eigen/Eigen/src/Core/Redux.h @@ -407,7 +407,7 @@ protected: */ template template -typename internal::traits::Scalar +EIGEN_STRONG_INLINE typename internal::traits::Scalar DenseBase::redux(const Func& func) const { eigen_assert(this->rows()>0 && this->cols()>0 && "you are using an empty matrix"); diff --git a/xs/src/eigen/Eigen/src/Core/Ref.h b/xs/src/eigen/Eigen/src/Core/Ref.h index bdf24f52a..9c6e3c5d9 100644 --- a/xs/src/eigen/Eigen/src/Core/Ref.h +++ b/xs/src/eigen/Eigen/src/Core/Ref.h @@ -95,6 +95,8 @@ protected: template EIGEN_DEVICE_FUNC void construct(Expression& expr) { + EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(PlainObjectType,Expression); + if(PlainObjectType::RowsAtCompileTime==1) { eigen_assert(expr.rows()==1 || expr.cols()==1); diff --git a/xs/src/eigen/Eigen/src/Core/SelfAdjointView.h b/xs/src/eigen/Eigen/src/Core/SelfAdjointView.h index 504c98f0e..b2e51f37a 100644 --- a/xs/src/eigen/Eigen/src/Core/SelfAdjointView.h +++ b/xs/src/eigen/Eigen/src/Core/SelfAdjointView.h @@ -71,7 +71,9 @@ template class SelfAdjointView EIGEN_DEVICE_FUNC explicit inline SelfAdjointView(MatrixType& matrix) : m_matrix(matrix) - {} + { + EIGEN_STATIC_ASSERT(UpLo==Lower || UpLo==Upper,SELFADJOINTVIEW_ACCEPTS_UPPER_AND_LOWER_MODE_ONLY); + } EIGEN_DEVICE_FUNC inline Index rows() const { return m_matrix.rows(); } @@ -189,7 +191,7 @@ template class SelfAdjointView TriangularView >::type(tmp2); } - typedef SelfAdjointView ConjugateReturnType; + typedef SelfAdjointView ConjugateReturnType; /** \sa MatrixBase::conjugate() const */ EIGEN_DEVICE_FUNC inline const ConjugateReturnType conjugate() const diff --git a/xs/src/eigen/Eigen/src/Core/SelfCwiseBinaryOp.h b/xs/src/eigen/Eigen/src/Core/SelfCwiseBinaryOp.h index 719ed72a5..7c89c2e23 100644 --- a/xs/src/eigen/Eigen/src/Core/SelfCwiseBinaryOp.h +++ b/xs/src/eigen/Eigen/src/Core/SelfCwiseBinaryOp.h @@ -15,33 +15,29 @@ namespace Eigen { // TODO generalize the scalar type of 'other' template -EIGEN_STRONG_INLINE Derived& DenseBase::operator*=(const Scalar& other) +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase::operator*=(const Scalar& other) { - typedef typename Derived::PlainObject PlainObject; internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::mul_assign_op()); return derived(); } template -EIGEN_STRONG_INLINE Derived& ArrayBase::operator+=(const Scalar& other) +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& ArrayBase::operator+=(const Scalar& other) { - typedef typename Derived::PlainObject PlainObject; internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::add_assign_op()); return derived(); } template -EIGEN_STRONG_INLINE Derived& ArrayBase::operator-=(const Scalar& other) +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& ArrayBase::operator-=(const Scalar& other) { - typedef typename Derived::PlainObject PlainObject; internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::sub_assign_op()); return derived(); } template -EIGEN_STRONG_INLINE Derived& DenseBase::operator/=(const Scalar& other) +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase::operator/=(const Scalar& other) { - typedef typename Derived::PlainObject PlainObject; internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::div_assign_op()); return derived(); } diff --git a/xs/src/eigen/Eigen/src/Core/Solve.h b/xs/src/eigen/Eigen/src/Core/Solve.h index 960a58597..a8daea511 100644 --- a/xs/src/eigen/Eigen/src/Core/Solve.h +++ b/xs/src/eigen/Eigen/src/Core/Solve.h @@ -34,12 +34,12 @@ template struct s template struct solve_traits { - typedef Matrix PlainObject; + RhsType::MaxColsAtCompileTime>::type PlainObject; }; template diff --git a/xs/src/eigen/Eigen/src/Core/StableNorm.h b/xs/src/eigen/Eigen/src/Core/StableNorm.h index d2fe1e199..88c8d9890 100644 --- a/xs/src/eigen/Eigen/src/Core/StableNorm.h +++ b/xs/src/eigen/Eigen/src/Core/StableNorm.h @@ -165,12 +165,13 @@ MatrixBase::stableNorm() const typedef typename internal::nested_eval::type DerivedCopy; typedef typename internal::remove_all::type DerivedCopyClean; - DerivedCopy copy(derived()); + const DerivedCopy copy(derived()); enum { CanAlign = ( (int(DerivedCopyClean::Flags)&DirectAccessBit) || (int(internal::evaluator::Alignment)>0) // FIXME Alignment)>0 might not be enough - ) && (blockSize*sizeof(Scalar)*20) // if we cannot allocate on the stack, then let's not bother about this optimization }; typedef typename internal::conditional, internal::evaluator::Alignment>, typename DerivedCopyClean::ConstSegmentReturnType>::type SegmentWrapper; diff --git a/xs/src/eigen/Eigen/src/Core/Transpositions.h b/xs/src/eigen/Eigen/src/Core/Transpositions.h index 19c17bb4a..86da5af59 100644 --- a/xs/src/eigen/Eigen/src/Core/Transpositions.h +++ b/xs/src/eigen/Eigen/src/Core/Transpositions.h @@ -384,7 +384,7 @@ class Transpose > const Product operator*(const MatrixBase& matrix, const Transpose& trt) { - return Product(matrix.derived(), trt.derived()); + return Product(matrix.derived(), trt); } /** \returns the \a matrix with the inverse transpositions applied to the rows. diff --git a/xs/src/eigen/Eigen/src/Core/arch/AVX/Complex.h b/xs/src/eigen/Eigen/src/Core/arch/AVX/Complex.h index 99439c8aa..7fa61969d 100644 --- a/xs/src/eigen/Eigen/src/Core/arch/AVX/Complex.h +++ b/xs/src/eigen/Eigen/src/Core/arch/AVX/Complex.h @@ -204,23 +204,7 @@ template<> struct conj_helper } }; -template<> struct conj_helper -{ - EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet8f& x, const Packet4cf& y, const Packet4cf& c) const - { return padd(c, pmul(x,y)); } - - EIGEN_STRONG_INLINE Packet4cf pmul(const Packet8f& x, const Packet4cf& y) const - { return Packet4cf(Eigen::internal::pmul(x, y.v)); } -}; - -template<> struct conj_helper -{ - EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet4cf& x, const Packet8f& y, const Packet4cf& c) const - { return padd(c, pmul(x,y)); } - - EIGEN_STRONG_INLINE Packet4cf pmul(const Packet4cf& x, const Packet8f& y) const - { return Packet4cf(Eigen::internal::pmul(x.v, y)); } -}; +EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet4cf,Packet8f) template<> EIGEN_STRONG_INLINE Packet4cf pdiv(const Packet4cf& a, const Packet4cf& b) { @@ -400,23 +384,7 @@ template<> struct conj_helper } }; -template<> struct conj_helper -{ - EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet4d& x, const Packet2cd& y, const Packet2cd& c) const - { return padd(c, pmul(x,y)); } - - EIGEN_STRONG_INLINE Packet2cd pmul(const Packet4d& x, const Packet2cd& y) const - { return Packet2cd(Eigen::internal::pmul(x, y.v)); } -}; - -template<> struct conj_helper -{ - EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet2cd& x, const Packet4d& y, const Packet2cd& c) const - { return padd(c, pmul(x,y)); } - - EIGEN_STRONG_INLINE Packet2cd pmul(const Packet2cd& x, const Packet4d& y) const - { return Packet2cd(Eigen::internal::pmul(x.v, y)); } -}; +EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet2cd,Packet4d) template<> EIGEN_STRONG_INLINE Packet2cd pdiv(const Packet2cd& a, const Packet2cd& b) { diff --git a/xs/src/eigen/Eigen/src/Core/arch/AVX/PacketMath.h b/xs/src/eigen/Eigen/src/Core/arch/AVX/PacketMath.h index 195d40fb4..61c3dfcab 100644 --- a/xs/src/eigen/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/xs/src/eigen/Eigen/src/Core/arch/AVX/PacketMath.h @@ -308,9 +308,9 @@ template<> EIGEN_STRONG_INLINE void pstore1(int* to, const int& a) } #ifndef EIGEN_VECTORIZE_AVX512 -template<> EIGEN_STRONG_INLINE void prefetch(const float* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } -template<> EIGEN_STRONG_INLINE void prefetch(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } -template<> EIGEN_STRONG_INLINE void prefetch(const int* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } +template<> EIGEN_STRONG_INLINE void prefetch(const float* addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); } +template<> EIGEN_STRONG_INLINE void prefetch(const double* addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); } +template<> EIGEN_STRONG_INLINE void prefetch(const int* addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); } #endif template<> EIGEN_STRONG_INLINE float pfirst(const Packet8f& a) { @@ -333,9 +333,12 @@ template<> EIGEN_STRONG_INLINE Packet4d preverse(const Packet4d& a) { __m256d tmp = _mm256_shuffle_pd(a,a,5); return _mm256_permute2f128_pd(tmp, tmp, 1); - + #if 0 + // This version is unlikely to be faster as _mm256_shuffle_ps and _mm256_permute_pd + // exhibit the same latency/throughput, but it is here for future reference/benchmarking... __m256d swap_halves = _mm256_permute2f128_pd(a,a,1); return _mm256_permute_pd(swap_halves,5); + #endif } // pabs should be ok diff --git a/xs/src/eigen/Eigen/src/Core/arch/AVX512/MathFunctions.h b/xs/src/eigen/Eigen/src/Core/arch/AVX512/MathFunctions.h index 399be0ee4..9c1717f76 100644 --- a/xs/src/eigen/Eigen/src/Core/arch/AVX512/MathFunctions.h +++ b/xs/src/eigen/Eigen/src/Core/arch/AVX512/MathFunctions.h @@ -88,9 +88,9 @@ plog(const Packet16f& _x) { // x = x + x - 1.0; // } else { x = x - 1.0; } __mmask16 mask = _mm512_cmp_ps_mask(x, p16f_cephes_SQRTHF, _CMP_LT_OQ); - Packet16f tmp = _mm512_mask_blend_ps(mask, x, _mm512_setzero_ps()); + Packet16f tmp = _mm512_mask_blend_ps(mask, _mm512_setzero_ps(), x); x = psub(x, p16f_1); - e = psub(e, _mm512_mask_blend_ps(mask, p16f_1, _mm512_setzero_ps())); + e = psub(e, _mm512_mask_blend_ps(mask, _mm512_setzero_ps(), p16f_1)); x = padd(x, tmp); Packet16f x2 = pmul(x, x); @@ -119,8 +119,9 @@ plog(const Packet16f& _x) { x = padd(x, y2); // Filter out invalid inputs, i.e. negative arg will be NAN, 0 will be -INF. - return _mm512_mask_blend_ps(iszero_mask, p16f_minus_inf, - _mm512_mask_blend_ps(invalid_mask, p16f_nan, x)); + return _mm512_mask_blend_ps(iszero_mask, + _mm512_mask_blend_ps(invalid_mask, x, p16f_nan), + p16f_minus_inf); } #endif @@ -266,8 +267,7 @@ psqrt(const Packet16f& _x) { // select only the inverse sqrt of positive normal inputs (denormals are // flushed to zero and cause infs as well). __mmask16 non_zero_mask = _mm512_cmp_ps_mask(_x, p16f_flt_min, _CMP_GE_OQ); - Packet16f x = _mm512_mask_blend_ps(non_zero_mask, _mm512_rsqrt14_ps(_x), - _mm512_setzero_ps()); + Packet16f x = _mm512_mask_blend_ps(non_zero_mask, _mm512_setzero_ps(), _mm512_rsqrt14_ps(_x)); // Do a single step of Newton's iteration. x = pmul(x, pmadd(neg_half, pmul(x, x), p16f_one_point_five)); @@ -289,8 +289,7 @@ psqrt(const Packet8d& _x) { // select only the inverse sqrt of positive normal inputs (denormals are // flushed to zero and cause infs as well). __mmask8 non_zero_mask = _mm512_cmp_pd_mask(_x, p8d_dbl_min, _CMP_GE_OQ); - Packet8d x = _mm512_mask_blend_pd(non_zero_mask, _mm512_rsqrt14_pd(_x), - _mm512_setzero_pd()); + Packet8d x = _mm512_mask_blend_pd(non_zero_mask, _mm512_setzero_pd(), _mm512_rsqrt14_pd(_x)); // Do a first step of Newton's iteration. x = pmul(x, pmadd(neg_half, pmul(x, x), p8d_one_point_five)); @@ -333,20 +332,18 @@ prsqrt(const Packet16f& _x) { // select only the inverse sqrt of positive normal inputs (denormals are // flushed to zero and cause infs as well). __mmask16 le_zero_mask = _mm512_cmp_ps_mask(_x, p16f_flt_min, _CMP_LT_OQ); - Packet16f x = _mm512_mask_blend_ps(le_zero_mask, _mm512_setzero_ps(), - _mm512_rsqrt14_ps(_x)); + Packet16f x = _mm512_mask_blend_ps(le_zero_mask, _mm512_rsqrt14_ps(_x), _mm512_setzero_ps()); // Fill in NaNs and Infs for the negative/zero entries. __mmask16 neg_mask = _mm512_cmp_ps_mask(_x, _mm512_setzero_ps(), _CMP_LT_OQ); Packet16f infs_and_nans = _mm512_mask_blend_ps( - neg_mask, p16f_nan, - _mm512_mask_blend_ps(le_zero_mask, p16f_inf, _mm512_setzero_ps())); + neg_mask, _mm512_mask_blend_ps(le_zero_mask, _mm512_setzero_ps(), p16f_inf), p16f_nan); // Do a single step of Newton's iteration. x = pmul(x, pmadd(neg_half, pmul(x, x), p16f_one_point_five)); // Insert NaNs and Infs in all the right places. - return _mm512_mask_blend_ps(le_zero_mask, infs_and_nans, x); + return _mm512_mask_blend_ps(le_zero_mask, x, infs_and_nans); } template <> @@ -363,14 +360,12 @@ prsqrt(const Packet8d& _x) { // select only the inverse sqrt of positive normal inputs (denormals are // flushed to zero and cause infs as well). __mmask8 le_zero_mask = _mm512_cmp_pd_mask(_x, p8d_dbl_min, _CMP_LT_OQ); - Packet8d x = _mm512_mask_blend_pd(le_zero_mask, _mm512_setzero_pd(), - _mm512_rsqrt14_pd(_x)); + Packet8d x = _mm512_mask_blend_pd(le_zero_mask, _mm512_rsqrt14_pd(_x), _mm512_setzero_pd()); // Fill in NaNs and Infs for the negative/zero entries. __mmask8 neg_mask = _mm512_cmp_pd_mask(_x, _mm512_setzero_pd(), _CMP_LT_OQ); Packet8d infs_and_nans = _mm512_mask_blend_pd( - neg_mask, p8d_nan, - _mm512_mask_blend_pd(le_zero_mask, p8d_inf, _mm512_setzero_pd())); + neg_mask, _mm512_mask_blend_pd(le_zero_mask, _mm512_setzero_pd(), p8d_inf), p8d_nan); // Do a first step of Newton's iteration. x = pmul(x, pmadd(neg_half, pmul(x, x), p8d_one_point_five)); @@ -379,9 +374,9 @@ prsqrt(const Packet8d& _x) { x = pmul(x, pmadd(neg_half, pmul(x, x), p8d_one_point_five)); // Insert NaNs and Infs in all the right places. - return _mm512_mask_blend_pd(le_zero_mask, infs_and_nans, x); + return _mm512_mask_blend_pd(le_zero_mask, x, infs_and_nans); } -#else +#elif defined(EIGEN_VECTORIZE_AVX512ER) template <> EIGEN_STRONG_INLINE Packet16f prsqrt(const Packet16f& x) { return _mm512_rsqrt28_ps(x); diff --git a/xs/src/eigen/Eigen/src/Core/arch/AVX512/PacketMath.h b/xs/src/eigen/Eigen/src/Core/arch/AVX512/PacketMath.h index f6500a16e..89705248a 100644 --- a/xs/src/eigen/Eigen/src/Core/arch/AVX512/PacketMath.h +++ b/xs/src/eigen/Eigen/src/Core/arch/AVX512/PacketMath.h @@ -618,9 +618,9 @@ EIGEN_STRONG_INLINE void pstore1(int* to, const int& a) { pstore(to, pa); } -template<> EIGEN_STRONG_INLINE void prefetch(const float* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } -template<> EIGEN_STRONG_INLINE void prefetch(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } -template<> EIGEN_STRONG_INLINE void prefetch(const int* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } +template<> EIGEN_STRONG_INLINE void prefetch(const float* addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); } +template<> EIGEN_STRONG_INLINE void prefetch(const double* addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); } +template<> EIGEN_STRONG_INLINE void prefetch(const int* addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); } template <> EIGEN_STRONG_INLINE float pfirst(const Packet16f& a) { diff --git a/xs/src/eigen/Eigen/src/Core/arch/AltiVec/Complex.h b/xs/src/eigen/Eigen/src/Core/arch/AltiVec/Complex.h index 67db2f8ee..3e665730c 100644 --- a/xs/src/eigen/Eigen/src/Core/arch/AltiVec/Complex.h +++ b/xs/src/eigen/Eigen/src/Core/arch/AltiVec/Complex.h @@ -224,23 +224,7 @@ template<> struct conj_helper } }; -template<> struct conj_helper -{ - EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet4f& x, const Packet2cf& y, const Packet2cf& c) const - { return padd(c, pmul(x,y)); } - - EIGEN_STRONG_INLINE Packet2cf pmul(const Packet4f& x, const Packet2cf& y) const - { return Packet2cf(internal::pmul(x, y.v)); } -}; - -template<> struct conj_helper -{ - EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet4f& y, const Packet2cf& c) const - { return padd(c, pmul(x,y)); } - - EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& x, const Packet4f& y) const - { return Packet2cf(internal::pmul(x.v, y)); } -}; +EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet2cf,Packet4f) template<> EIGEN_STRONG_INLINE Packet2cf pdiv(const Packet2cf& a, const Packet2cf& b) { @@ -416,23 +400,8 @@ template<> struct conj_helper return pconj(internal::pmul(a, b)); } }; -template<> struct conj_helper -{ - EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet2d& x, const Packet1cd& y, const Packet1cd& c) const - { return padd(c, pmul(x,y)); } - EIGEN_STRONG_INLINE Packet1cd pmul(const Packet2d& x, const Packet1cd& y) const - { return Packet1cd(internal::pmul(x, y.v)); } -}; - -template<> struct conj_helper -{ - EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet2d& y, const Packet1cd& c) const - { return padd(c, pmul(x,y)); } - - EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& x, const Packet2d& y) const - { return Packet1cd(internal::pmul(x.v, y)); } -}; +EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet1cd,Packet2d) template<> EIGEN_STRONG_INLINE Packet1cd pdiv(const Packet1cd& a, const Packet1cd& b) { diff --git a/xs/src/eigen/Eigen/src/Core/arch/AltiVec/PacketMath.h b/xs/src/eigen/Eigen/src/Core/arch/AltiVec/PacketMath.h index b3f1ea199..08a27d153 100644 --- a/xs/src/eigen/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/xs/src/eigen/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -103,7 +103,7 @@ static Packet16uc p16uc_PSET32_WODD = vec_sld((Packet16uc) vec_splat((Packet4u static Packet16uc p16uc_PSET32_WEVEN = vec_sld(p16uc_DUPLICATE32_HI, (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 }; static Packet16uc p16uc_HALF64_0_16 = vec_sld((Packet16uc)p4i_ZERO, vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 3), 8); //{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16}; #else -static Packet16uc p16uc_FORWARD = p16uc_REVERSE32; +static Packet16uc p16uc_FORWARD = p16uc_REVERSE32; static Packet16uc p16uc_REVERSE64 = { 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 }; static Packet16uc p16uc_PSET32_WODD = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 1), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 }; static Packet16uc p16uc_PSET32_WEVEN = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 }; @@ -388,10 +388,28 @@ template<> EIGEN_STRONG_INLINE Packet4i pdiv(const Packet4i& /*a*/, co template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vec_madd(a,b,c); } template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return a*b + c; } -template<> EIGEN_STRONG_INLINE Packet4f pmin(const Packet4f& a, const Packet4f& b) { return vec_min(a, b); } +template<> EIGEN_STRONG_INLINE Packet4f pmin(const Packet4f& a, const Packet4f& b) +{ + #ifdef __VSX__ + Packet4f ret; + __asm__ ("xvcmpgesp %x0,%x1,%x2\n\txxsel %x0,%x1,%x2,%x0" : "=&wa" (ret) : "wa" (a), "wa" (b)); + return ret; + #else + return vec_min(a, b); + #endif +} template<> EIGEN_STRONG_INLINE Packet4i pmin(const Packet4i& a, const Packet4i& b) { return vec_min(a, b); } -template<> EIGEN_STRONG_INLINE Packet4f pmax(const Packet4f& a, const Packet4f& b) { return vec_max(a, b); } +template<> EIGEN_STRONG_INLINE Packet4f pmax(const Packet4f& a, const Packet4f& b) +{ + #ifdef __VSX__ + Packet4f ret; + __asm__ ("xvcmpgtsp %x0,%x2,%x1\n\txxsel %x0,%x1,%x2,%x0" : "=&wa" (ret) : "wa" (a), "wa" (b)); + return ret; + #else + return vec_max(a, b); + #endif +} template<> EIGEN_STRONG_INLINE Packet4i pmax(const Packet4i& a, const Packet4i& b) { return vec_max(a, b); } template<> EIGEN_STRONG_INLINE Packet4f pand(const Packet4f& a, const Packet4f& b) { return vec_and(a, b); } @@ -764,7 +782,7 @@ typedef __vector __bool long Packet2bl; static Packet2l p2l_ONE = { 1, 1 }; static Packet2l p2l_ZERO = reinterpret_cast(p4i_ZERO); -static Packet2d p2d_ONE = { 1.0, 1.0 }; +static Packet2d p2d_ONE = { 1.0, 1.0 }; static Packet2d p2d_ZERO = reinterpret_cast(p4f_ZERO); static Packet2d p2d_MZERO = { -0.0, -0.0 }; @@ -910,9 +928,19 @@ template<> EIGEN_STRONG_INLINE Packet2d pdiv(const Packet2d& a, const // for some weird raisons, it has to be overloaded for packet of integers template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vec_madd(a, b, c); } -template<> EIGEN_STRONG_INLINE Packet2d pmin(const Packet2d& a, const Packet2d& b) { return vec_min(a, b); } +template<> EIGEN_STRONG_INLINE Packet2d pmin(const Packet2d& a, const Packet2d& b) +{ + Packet2d ret; + __asm__ ("xvcmpgedp %x0,%x1,%x2\n\txxsel %x0,%x1,%x2,%x0" : "=&wa" (ret) : "wa" (a), "wa" (b)); + return ret; + } -template<> EIGEN_STRONG_INLINE Packet2d pmax(const Packet2d& a, const Packet2d& b) { return vec_max(a, b); } +template<> EIGEN_STRONG_INLINE Packet2d pmax(const Packet2d& a, const Packet2d& b) +{ + Packet2d ret; + __asm__ ("xvcmpgtdp %x0,%x2,%x1\n\txxsel %x0,%x1,%x2,%x0" : "=&wa" (ret) : "wa" (a), "wa" (b)); + return ret; +} template<> EIGEN_STRONG_INLINE Packet2d pand(const Packet2d& a, const Packet2d& b) { return vec_and(a, b); } @@ -969,7 +997,7 @@ template<> EIGEN_STRONG_INLINE Packet2d preduxp(const Packet2d* vecs) Packet2d v[2], sum; v[0] = vecs[0] + reinterpret_cast(vec_sld(reinterpret_cast(vecs[0]), reinterpret_cast(vecs[0]), 8)); v[1] = vecs[1] + reinterpret_cast(vec_sld(reinterpret_cast(vecs[1]), reinterpret_cast(vecs[1]), 8)); - + #ifdef _BIG_ENDIAN sum = reinterpret_cast(vec_sld(reinterpret_cast(v[0]), reinterpret_cast(v[1]), 8)); #else @@ -1022,7 +1050,7 @@ ptranspose(PacketBlock& kernel) { template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket, const Packet2d& elsePacket) { Packet2l select = { ifPacket.select[0], ifPacket.select[1] }; - Packet2bl mask = vec_cmpeq(reinterpret_cast(select), reinterpret_cast(p2l_ONE)); + Packet2bl mask = reinterpret_cast( vec_cmpeq(reinterpret_cast(select), reinterpret_cast(p2l_ONE)) ); return vec_sel(elsePacket, thenPacket, mask); } #endif // __VSX__ diff --git a/xs/src/eigen/Eigen/src/Core/arch/CUDA/Half.h b/xs/src/eigen/Eigen/src/Core/arch/CUDA/Half.h index 52892db38..02ac0c23a 100644 --- a/xs/src/eigen/Eigen/src/Core/arch/CUDA/Half.h +++ b/xs/src/eigen/Eigen/src/Core/arch/CUDA/Half.h @@ -13,7 +13,7 @@ // Redistribution and use in source and binary forms, with or without // modification, are permitted. // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS -// “AS IS” AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// "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 // HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, @@ -147,55 +147,55 @@ namespace half_impl { // versions to get the ALU speed increased), but you do save the // conversion steps back and forth. -__device__ half operator + (const half& a, const half& b) { +EIGEN_STRONG_INLINE __device__ half operator + (const half& a, const half& b) { return __hadd(a, b); } -__device__ half operator * (const half& a, const half& b) { +EIGEN_STRONG_INLINE __device__ half operator * (const half& a, const half& b) { return __hmul(a, b); } -__device__ half operator - (const half& a, const half& b) { +EIGEN_STRONG_INLINE __device__ half operator - (const half& a, const half& b) { return __hsub(a, b); } -__device__ half operator / (const half& a, const half& b) { +EIGEN_STRONG_INLINE __device__ half operator / (const half& a, const half& b) { float num = __half2float(a); float denom = __half2float(b); return __float2half(num / denom); } -__device__ half operator - (const half& a) { +EIGEN_STRONG_INLINE __device__ half operator - (const half& a) { return __hneg(a); } -__device__ half& operator += (half& a, const half& b) { +EIGEN_STRONG_INLINE __device__ half& operator += (half& a, const half& b) { a = a + b; return a; } -__device__ half& operator *= (half& a, const half& b) { +EIGEN_STRONG_INLINE __device__ half& operator *= (half& a, const half& b) { a = a * b; return a; } -__device__ half& operator -= (half& a, const half& b) { +EIGEN_STRONG_INLINE __device__ half& operator -= (half& a, const half& b) { a = a - b; return a; } -__device__ half& operator /= (half& a, const half& b) { +EIGEN_STRONG_INLINE __device__ half& operator /= (half& a, const half& b) { a = a / b; return a; } -__device__ bool operator == (const half& a, const half& b) { +EIGEN_STRONG_INLINE __device__ bool operator == (const half& a, const half& b) { return __heq(a, b); } -__device__ bool operator != (const half& a, const half& b) { +EIGEN_STRONG_INLINE __device__ bool operator != (const half& a, const half& b) { return __hne(a, b); } -__device__ bool operator < (const half& a, const half& b) { +EIGEN_STRONG_INLINE __device__ bool operator < (const half& a, const half& b) { return __hlt(a, b); } -__device__ bool operator <= (const half& a, const half& b) { +EIGEN_STRONG_INLINE __device__ bool operator <= (const half& a, const half& b) { return __hle(a, b); } -__device__ bool operator > (const half& a, const half& b) { +EIGEN_STRONG_INLINE __device__ bool operator > (const half& a, const half& b) { return __hgt(a, b); } -__device__ bool operator >= (const half& a, const half& b) { +EIGEN_STRONG_INLINE __device__ bool operator >= (const half& a, const half& b) { return __hge(a, b); } @@ -238,10 +238,10 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator /= (half& a, const half& b) return a; } EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator == (const half& a, const half& b) { - return float(a) == float(b); + return numext::equal_strict(float(a),float(b)); } EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator != (const half& a, const half& b) { - return float(a) != float(b); + return numext::not_equal_strict(float(a), float(b)); } EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator < (const half& a, const half& b) { return float(a) < float(b); @@ -386,11 +386,15 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half abs(const half& a) { return result; } EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half exp(const half& a) { - return half(::expf(float(a))); +#if EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530 + return half(hexp(a)); +#else + return half(::expf(float(a))); +#endif } EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log(const half& a) { -#if defined(EIGEN_HAS_CUDA_FP16) && defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530 - return Eigen::half(::hlog(a)); +#if defined(EIGEN_HAS_CUDA_FP16) && EIGEN_CUDACC_VER >= 80000 && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530 + return half(::hlog(a)); #else return half(::logf(float(a))); #endif @@ -402,7 +406,11 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log10(const half& a) { return half(::log10f(float(a))); } EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half sqrt(const half& a) { - return half(::sqrtf(float(a))); +#if EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530 + return half(hsqrt(a)); +#else + return half(::sqrtf(float(a))); +#endif } EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half pow(const half& a, const half& b) { return half(::powf(float(a), float(b))); @@ -420,10 +428,18 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half tanh(const half& a) { return half(::tanhf(float(a))); } EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half floor(const half& a) { +#if EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 300 + return half(hfloor(a)); +#else return half(::floorf(float(a))); +#endif } EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half ceil(const half& a) { +#if EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 300 + return half(hceil(a)); +#else return half(::ceilf(float(a))); +#endif } EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half (min)(const half& a, const half& b) { @@ -474,9 +490,59 @@ template<> struct is_arithmetic { enum { value = true }; }; } // end namespace internal +} // end namespace Eigen + +namespace std { +template<> +struct numeric_limits { + static const bool is_specialized = true; + static const bool is_signed = true; + static const bool is_integer = false; + static const bool is_exact = false; + static const bool has_infinity = true; + static const bool has_quiet_NaN = true; + static const bool has_signaling_NaN = true; + static const float_denorm_style has_denorm = denorm_present; + static const bool has_denorm_loss = false; + static const std::float_round_style round_style = std::round_to_nearest; + static const bool is_iec559 = false; + static const bool is_bounded = false; + static const bool is_modulo = false; + static const int digits = 11; + static const int digits10 = 3; // according to http://half.sourceforge.net/structstd_1_1numeric__limits_3_01half__float_1_1half_01_4.html + static const int max_digits10 = 5; // according to http://half.sourceforge.net/structstd_1_1numeric__limits_3_01half__float_1_1half_01_4.html + static const int radix = 2; + static const int min_exponent = -13; + static const int min_exponent10 = -4; + static const int max_exponent = 16; + static const int max_exponent10 = 4; + static const bool traps = true; + static const bool tinyness_before = false; + + static Eigen::half (min)() { return Eigen::half_impl::raw_uint16_to_half(0x400); } + static Eigen::half lowest() { return Eigen::half_impl::raw_uint16_to_half(0xfbff); } + static Eigen::half (max)() { return Eigen::half_impl::raw_uint16_to_half(0x7bff); } + static Eigen::half epsilon() { return Eigen::half_impl::raw_uint16_to_half(0x0800); } + static Eigen::half round_error() { return Eigen::half(0.5); } + static Eigen::half infinity() { return Eigen::half_impl::raw_uint16_to_half(0x7c00); } + static Eigen::half quiet_NaN() { return Eigen::half_impl::raw_uint16_to_half(0x7e00); } + static Eigen::half signaling_NaN() { return Eigen::half_impl::raw_uint16_to_half(0x7e00); } + static Eigen::half denorm_min() { return Eigen::half_impl::raw_uint16_to_half(0x1); } +}; +} + +namespace Eigen { + template<> struct NumTraits : GenericNumTraits { + enum { + IsSigned = true, + IsInteger = false, + IsComplex = false, + RequireInitialization = false + }; + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half epsilon() { return half_impl::raw_uint16_to_half(0x0800); } @@ -507,7 +573,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half exph(const Eigen::half& a) { return Eigen::half(::expf(float(a))); } EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half logh(const Eigen::half& a) { -#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530 +#if EIGEN_CUDACC_VER >= 80000 && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530 return Eigen::half(::hlog(a)); #else return Eigen::half(::logf(float(a))); diff --git a/xs/src/eigen/Eigen/src/Core/arch/CUDA/PacketMath.h b/xs/src/eigen/Eigen/src/Core/arch/CUDA/PacketMath.h index ad66399e0..4dda63188 100644 --- a/xs/src/eigen/Eigen/src/Core/arch/CUDA/PacketMath.h +++ b/xs/src/eigen/Eigen/src/Core/arch/CUDA/PacketMath.h @@ -291,7 +291,7 @@ template<> EIGEN_DEVICE_FUNC inline double2 pabs(const double2& a) { EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock& kernel) { - double tmp = kernel.packet[0].y; + float tmp = kernel.packet[0].y; kernel.packet[0].y = kernel.packet[1].x; kernel.packet[1].x = tmp; diff --git a/xs/src/eigen/Eigen/src/Core/arch/CUDA/PacketMathHalf.h b/xs/src/eigen/Eigen/src/Core/arch/CUDA/PacketMathHalf.h index ae54225f8..943e0b06d 100644 --- a/xs/src/eigen/Eigen/src/Core/arch/CUDA/PacketMathHalf.h +++ b/xs/src/eigen/Eigen/src/Core/arch/CUDA/PacketMathHalf.h @@ -275,7 +275,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 plog1p(const half2& a) { return __floats2half2_rn(r1, r2); } -#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 530 +#if EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530 template<> __device__ EIGEN_STRONG_INLINE half2 plog(const half2& a) { diff --git a/xs/src/eigen/Eigen/src/Core/arch/Default/ConjHelper.h b/xs/src/eigen/Eigen/src/Core/arch/Default/ConjHelper.h new file mode 100644 index 000000000..4cfe34e05 --- /dev/null +++ b/xs/src/eigen/Eigen/src/Core/arch/Default/ConjHelper.h @@ -0,0 +1,29 @@ + +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2017 Gael Guennebaud +// +// 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_ARCH_CONJ_HELPER_H +#define EIGEN_ARCH_CONJ_HELPER_H + +#define EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(PACKET_CPLX, PACKET_REAL) \ + template<> struct conj_helper { \ + EIGEN_STRONG_INLINE PACKET_CPLX pmadd(const PACKET_REAL& x, const PACKET_CPLX& y, const PACKET_CPLX& c) const \ + { return padd(c, pmul(x,y)); } \ + EIGEN_STRONG_INLINE PACKET_CPLX pmul(const PACKET_REAL& x, const PACKET_CPLX& y) const \ + { return PACKET_CPLX(Eigen::internal::pmul(x, y.v)); } \ + }; \ + \ + template<> struct conj_helper { \ + EIGEN_STRONG_INLINE PACKET_CPLX pmadd(const PACKET_CPLX& x, const PACKET_REAL& y, const PACKET_CPLX& c) const \ + { return padd(c, pmul(x,y)); } \ + EIGEN_STRONG_INLINE PACKET_CPLX pmul(const PACKET_CPLX& x, const PACKET_REAL& y) const \ + { return PACKET_CPLX(Eigen::internal::pmul(x.v, y)); } \ + }; + +#endif // EIGEN_ARCH_CONJ_HELPER_H diff --git a/xs/src/eigen/Eigen/src/Core/arch/NEON/Complex.h b/xs/src/eigen/Eigen/src/Core/arch/NEON/Complex.h index 57e9b431f..306a309be 100644 --- a/xs/src/eigen/Eigen/src/Core/arch/NEON/Complex.h +++ b/xs/src/eigen/Eigen/src/Core/arch/NEON/Complex.h @@ -67,7 +67,7 @@ template<> struct unpacket_traits { typedef std::complex type; template<> EIGEN_STRONG_INLINE Packet2cf pset1(const std::complex& from) { float32x2_t r64; - r64 = vld1_f32((float *)&from); + r64 = vld1_f32((const float *)&from); return Packet2cf(vcombine_f32(r64, r64)); } @@ -142,7 +142,7 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet2cf to[stride*1] = std::complex(vgetq_lane_f32(from.v, 2), vgetq_lane_f32(from.v, 3)); } -template<> EIGEN_STRONG_INLINE void prefetch >(const std::complex * addr) { EIGEN_ARM_PREFETCH((float *)addr); } +template<> EIGEN_STRONG_INLINE void prefetch >(const std::complex * addr) { EIGEN_ARM_PREFETCH((const float *)addr); } template<> EIGEN_STRONG_INLINE std::complex pfirst(const Packet2cf& a) { @@ -265,6 +265,8 @@ template<> struct conj_helper } }; +EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet2cf,Packet4f) + template<> EIGEN_STRONG_INLINE Packet2cf pdiv(const Packet2cf& a, const Packet2cf& b) { // TODO optimize it for NEON @@ -275,7 +277,7 @@ template<> EIGEN_STRONG_INLINE Packet2cf pdiv(const Packet2cf& a, con s = vmulq_f32(b.v, b.v); rev_s = vrev64q_f32(s); - return Packet2cf(pdiv(res.v, vaddq_f32(s,rev_s))); + return Packet2cf(pdiv(res.v, vaddq_f32(s,rev_s))); } EIGEN_DEVICE_FUNC inline void @@ -381,7 +383,7 @@ template<> EIGEN_STRONG_INLINE Packet1cd ploaddup(const std::complex< template<> EIGEN_STRONG_INLINE void pstore >(std::complex * to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); } template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex * to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); } -template<> EIGEN_STRONG_INLINE void prefetch >(const std::complex * addr) { EIGEN_ARM_PREFETCH((double *)addr); } +template<> EIGEN_STRONG_INLINE void prefetch >(const std::complex * addr) { EIGEN_ARM_PREFETCH((const double *)addr); } template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather, Packet1cd>(const std::complex* from, Index stride) { @@ -456,6 +458,8 @@ template<> struct conj_helper } }; +EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet1cd,Packet2d) + template<> EIGEN_STRONG_INLINE Packet1cd pdiv(const Packet1cd& a, const Packet1cd& b) { // TODO optimize it for NEON diff --git a/xs/src/eigen/Eigen/src/Core/arch/NEON/PacketMath.h b/xs/src/eigen/Eigen/src/Core/arch/NEON/PacketMath.h index 84a56bdcc..3d5ed0d24 100644 --- a/xs/src/eigen/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/xs/src/eigen/Eigen/src/Core/arch/NEON/PacketMath.h @@ -36,12 +36,43 @@ namespace internal { #endif #endif +#if EIGEN_COMP_MSVC + +// In MSVC's arm_neon.h header file, all NEON vector types +// are aliases to the same underlying type __n128. +// We thus have to wrap them to make them different C++ types. +// (See also bug 1428) + +template +struct eigen_packet_wrapper +{ + operator T&() { return m_val; } + operator const T&() const { return m_val; } + eigen_packet_wrapper() {} + eigen_packet_wrapper(const T &v) : m_val(v) {} + eigen_packet_wrapper& operator=(const T &v) { + m_val = v; + return *this; + } + + T m_val; +}; +typedef eigen_packet_wrapper Packet2f; +typedef eigen_packet_wrapper Packet4f; +typedef eigen_packet_wrapper Packet4i; +typedef eigen_packet_wrapper Packet2i; +typedef eigen_packet_wrapper Packet4ui; + +#else + typedef float32x2_t Packet2f; typedef float32x4_t Packet4f; typedef int32x4_t Packet4i; typedef int32x2_t Packet2i; typedef uint32x4_t Packet4ui; +#endif // EIGEN_COMP_MSVC + #define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \ const Packet4f p4f_##NAME = pset1(X) @@ -51,14 +82,17 @@ typedef uint32x4_t Packet4ui; #define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \ const Packet4i p4i_##NAME = pset1(X) -// arm64 does have the pld instruction. If available, let's trust the __builtin_prefetch built-in function -// which available on LLVM and GCC (at least) -#if EIGEN_HAS_BUILTIN(__builtin_prefetch) || EIGEN_COMP_GNUC +#if EIGEN_ARCH_ARM64 + // __builtin_prefetch tends to do nothing on ARM64 compilers because the + // prefetch instructions there are too detailed for __builtin_prefetch to map + // meaningfully to them. + #define EIGEN_ARM_PREFETCH(ADDR) __asm__ __volatile__("prfm pldl1keep, [%[addr]]\n" ::[addr] "r"(ADDR) : ); +#elif EIGEN_HAS_BUILTIN(__builtin_prefetch) || EIGEN_COMP_GNUC #define EIGEN_ARM_PREFETCH(ADDR) __builtin_prefetch(ADDR); #elif defined __pld #define EIGEN_ARM_PREFETCH(ADDR) __pld(ADDR) -#elif !EIGEN_ARCH_ARM64 - #define EIGEN_ARM_PREFETCH(ADDR) __asm__ __volatile__ ( " pld [%[addr]]\n" :: [addr] "r" (ADDR) : "cc" ); +#elif EIGEN_ARCH_ARM32 + #define EIGEN_ARM_PREFETCH(ADDR) __asm__ __volatile__ ("pld [%[addr]]\n" :: [addr] "r" (ADDR) : ); #else // by default no explicit prefetching #define EIGEN_ARM_PREFETCH(ADDR) @@ -113,7 +147,7 @@ template<> EIGEN_STRONG_INLINE Packet4i pset1(const int32_t& from) template<> EIGEN_STRONG_INLINE Packet4f plset(const float& a) { - const float32_t f[] = {0, 1, 2, 3}; + const float f[] = {0, 1, 2, 3}; Packet4f countdown = vld1q_f32(f); return vaddq_f32(pset1(a), countdown); } diff --git a/xs/src/eigen/Eigen/src/Core/arch/SSE/Complex.h b/xs/src/eigen/Eigen/src/Core/arch/SSE/Complex.h index 5607fe0ab..d075043ce 100644 --- a/xs/src/eigen/Eigen/src/Core/arch/SSE/Complex.h +++ b/xs/src/eigen/Eigen/src/Core/arch/SSE/Complex.h @@ -128,7 +128,7 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet2cf _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 3))); } -template<> EIGEN_STRONG_INLINE void prefetch >(const std::complex * addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } +template<> EIGEN_STRONG_INLINE void prefetch >(const std::complex * addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); } template<> EIGEN_STRONG_INLINE std::complex pfirst(const Packet2cf& a) { @@ -229,23 +229,7 @@ template<> struct conj_helper } }; -template<> struct conj_helper -{ - EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet4f& x, const Packet2cf& y, const Packet2cf& c) const - { return padd(c, pmul(x,y)); } - - EIGEN_STRONG_INLINE Packet2cf pmul(const Packet4f& x, const Packet2cf& y) const - { return Packet2cf(Eigen::internal::pmul(x, y.v)); } -}; - -template<> struct conj_helper -{ - EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet4f& y, const Packet2cf& c) const - { return padd(c, pmul(x,y)); } - - EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& x, const Packet4f& y) const - { return Packet2cf(Eigen::internal::pmul(x.v, y)); } -}; +EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet2cf,Packet4f) template<> EIGEN_STRONG_INLINE Packet2cf pdiv(const Packet2cf& a, const Packet2cf& b) { @@ -340,7 +324,7 @@ template<> EIGEN_STRONG_INLINE Packet1cd ploaddup(const std::complex< template<> EIGEN_STRONG_INLINE void pstore >(std::complex * to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, Packet2d(from.v)); } template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex * to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, Packet2d(from.v)); } -template<> EIGEN_STRONG_INLINE void prefetch >(const std::complex * addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } +template<> EIGEN_STRONG_INLINE void prefetch >(const std::complex * addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); } template<> EIGEN_STRONG_INLINE std::complex pfirst(const Packet1cd& a) { @@ -430,23 +414,7 @@ template<> struct conj_helper } }; -template<> struct conj_helper -{ - EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet2d& x, const Packet1cd& y, const Packet1cd& c) const - { return padd(c, pmul(x,y)); } - - EIGEN_STRONG_INLINE Packet1cd pmul(const Packet2d& x, const Packet1cd& y) const - { return Packet1cd(Eigen::internal::pmul(x, y.v)); } -}; - -template<> struct conj_helper -{ - EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet2d& y, const Packet1cd& c) const - { return padd(c, pmul(x,y)); } - - EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& x, const Packet2d& y) const - { return Packet1cd(Eigen::internal::pmul(x.v, y)); } -}; +EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet1cd,Packet2d) template<> EIGEN_STRONG_INLINE Packet1cd pdiv(const Packet1cd& a, const Packet1cd& b) { diff --git a/xs/src/eigen/Eigen/src/Core/arch/SSE/PacketMath.h b/xs/src/eigen/Eigen/src/Core/arch/SSE/PacketMath.h index 3832de147..5e652cc79 100644 --- a/xs/src/eigen/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/xs/src/eigen/Eigen/src/Core/arch/SSE/PacketMath.h @@ -409,10 +409,16 @@ template<> EIGEN_STRONG_INLINE void pstore1(double* to, const double& pstore(to, Packet2d(vec2d_swizzle1(pa,0,0))); } +#if EIGEN_COMP_PGI +typedef const void * SsePrefetchPtrType; +#else +typedef const char * SsePrefetchPtrType; +#endif + #ifndef EIGEN_VECTORIZE_AVX -template<> EIGEN_STRONG_INLINE void prefetch(const float* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } -template<> EIGEN_STRONG_INLINE void prefetch(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } -template<> EIGEN_STRONG_INLINE void prefetch(const int* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } +template<> EIGEN_STRONG_INLINE void prefetch(const float* addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); } +template<> EIGEN_STRONG_INLINE void prefetch(const double* addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); } +template<> EIGEN_STRONG_INLINE void prefetch(const int* addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); } #endif #if EIGEN_COMP_MSVC_STRICT && EIGEN_OS_WIN64 @@ -876,4 +882,14 @@ template<> EIGEN_STRONG_INLINE double pmadd(const double& a, const double& b, co } // end namespace Eigen +#if EIGEN_COMP_PGI +// PGI++ does not define the following intrinsics in C++ mode. +static inline __m128 _mm_castpd_ps (__m128d x) { return reinterpret_cast<__m128&>(x); } +static inline __m128i _mm_castpd_si128(__m128d x) { return reinterpret_cast<__m128i&>(x); } +static inline __m128d _mm_castps_pd (__m128 x) { return reinterpret_cast<__m128d&>(x); } +static inline __m128i _mm_castps_si128(__m128 x) { return reinterpret_cast<__m128i&>(x); } +static inline __m128 _mm_castsi128_ps(__m128i x) { return reinterpret_cast<__m128&>(x); } +static inline __m128d _mm_castsi128_pd(__m128i x) { return reinterpret_cast<__m128d&>(x); } +#endif + #endif // EIGEN_PACKET_MATH_SSE_H diff --git a/xs/src/eigen/Eigen/src/Core/arch/SSE/TypeCasting.h b/xs/src/eigen/Eigen/src/Core/arch/SSE/TypeCasting.h index c84893230..c6ca8c716 100644 --- a/xs/src/eigen/Eigen/src/Core/arch/SSE/TypeCasting.h +++ b/xs/src/eigen/Eigen/src/Core/arch/SSE/TypeCasting.h @@ -14,6 +14,7 @@ namespace Eigen { namespace internal { +#ifndef EIGEN_VECTORIZE_AVX template <> struct type_casting_traits { enum { @@ -23,11 +24,6 @@ struct type_casting_traits { }; }; -template<> EIGEN_STRONG_INLINE Packet4i pcast(const Packet4f& a) { - return _mm_cvttps_epi32(a); -} - - template <> struct type_casting_traits { enum { @@ -37,11 +33,6 @@ struct type_casting_traits { }; }; -template<> EIGEN_STRONG_INLINE Packet4f pcast(const Packet4i& a) { - return _mm_cvtepi32_ps(a); -} - - template <> struct type_casting_traits { enum { @@ -51,10 +42,6 @@ struct type_casting_traits { }; }; -template<> EIGEN_STRONG_INLINE Packet4f pcast(const Packet2d& a, const Packet2d& b) { - return _mm_shuffle_ps(_mm_cvtpd_ps(a), _mm_cvtpd_ps(b), (1 << 2) | (1 << 6)); -} - template <> struct type_casting_traits { enum { @@ -63,6 +50,19 @@ struct type_casting_traits { TgtCoeffRatio = 2 }; }; +#endif + +template<> EIGEN_STRONG_INLINE Packet4i pcast(const Packet4f& a) { + return _mm_cvttps_epi32(a); +} + +template<> EIGEN_STRONG_INLINE Packet4f pcast(const Packet4i& a) { + return _mm_cvtepi32_ps(a); +} + +template<> EIGEN_STRONG_INLINE Packet4f pcast(const Packet2d& a, const Packet2d& b) { + return _mm_shuffle_ps(_mm_cvtpd_ps(a), _mm_cvtpd_ps(b), (1 << 2) | (1 << 6)); +} template<> EIGEN_STRONG_INLINE Packet2d pcast(const Packet4f& a) { // Simply discard the second half of the input diff --git a/xs/src/eigen/Eigen/src/Core/arch/ZVector/Complex.h b/xs/src/eigen/Eigen/src/Core/arch/ZVector/Complex.h index d39d2d105..1bfb73397 100644 --- a/xs/src/eigen/Eigen/src/Core/arch/ZVector/Complex.h +++ b/xs/src/eigen/Eigen/src/Core/arch/ZVector/Complex.h @@ -336,6 +336,9 @@ template<> struct conj_helper } }; +EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet2cf,Packet4f) +EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet1cd,Packet2d) + template<> EIGEN_STRONG_INLINE Packet1cd pdiv(const Packet1cd& a, const Packet1cd& b) { // TODO optimize it for AltiVec diff --git a/xs/src/eigen/Eigen/src/Core/functors/BinaryFunctors.h b/xs/src/eigen/Eigen/src/Core/functors/BinaryFunctors.h index 96747bac7..3eae6b8ca 100644 --- a/xs/src/eigen/Eigen/src/Core/functors/BinaryFunctors.h +++ b/xs/src/eigen/Eigen/src/Core/functors/BinaryFunctors.h @@ -255,7 +255,7 @@ struct scalar_cmp_op : binary_op_base struct scalar_hypot_op : binary_op_base { EIGEN_EMPTY_STRUCT_CTOR(scalar_hypot_op) -// typedef typename NumTraits::Real result_type; - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& _x, const Scalar& _y) const + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar &x, const Scalar &y) const { - EIGEN_USING_STD_MATH(sqrt) - Scalar p, qp; - if(_x>_y) - { - p = _x; - qp = _y / p; - } - else - { - p = _y; - qp = _x / p; - } - return p * sqrt(Scalar(1) + qp*qp); + // This functor is used by hypotNorm only for which it is faster to first apply abs + // on all coefficients prior to reduction through hypot. + // This way we avoid calling abs on positive and real entries, and this also permits + // to seamlessly handle complexes. Otherwise we would have to handle both real and complexes + // through the same functor... + return internal::positive_real_hypot(x,y); } }; template diff --git a/xs/src/eigen/Eigen/src/Core/functors/NullaryFunctors.h b/xs/src/eigen/Eigen/src/Core/functors/NullaryFunctors.h index 6a30466fb..b03be0269 100644 --- a/xs/src/eigen/Eigen/src/Core/functors/NullaryFunctors.h +++ b/xs/src/eigen/Eigen/src/Core/functors/NullaryFunctors.h @@ -44,16 +44,16 @@ struct linspaced_op_impl { linspaced_op_impl(const Scalar& low, const Scalar& high, Index num_steps) : m_low(low), m_high(high), m_size1(num_steps==1 ? 1 : num_steps-1), m_step(num_steps==1 ? Scalar() : (high-low)/Scalar(num_steps-1)), - m_interPacket(plset(0)), m_flip(numext::abs(high) EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (IndexType i) const { + typedef typename NumTraits::Real RealScalar; if(m_flip) - return (i==0)? m_low : (m_high - (m_size1-i)*m_step); + return (i==0)? m_low : (m_high - RealScalar(m_size1-i)*m_step); else - return (i==m_size1)? m_high : (m_low + i*m_step); + return (i==m_size1)? m_high : (m_low + RealScalar(i)*m_step); } template @@ -63,7 +63,7 @@ struct linspaced_op_impl // [low, ..., low] + ( [step, ..., step] * ( [i, ..., i] + [0, ..., size] ) ) if(m_flip) { - Packet pi = padd(pset1(Scalar(i-m_size1)),m_interPacket); + Packet pi = plset(Scalar(i-m_size1)); Packet res = padd(pset1(m_high), pmul(pset1(m_step), pi)); if(i==0) res = pinsertfirst(res, m_low); @@ -71,7 +71,7 @@ struct linspaced_op_impl } else { - Packet pi = padd(pset1(Scalar(i)),m_interPacket); + Packet pi = plset(Scalar(i)); Packet res = padd(pset1(m_low), pmul(pset1(m_step), pi)); if(i==m_size1-unpacket_traits::size+1) res = pinsertlast(res, m_high); @@ -83,7 +83,6 @@ struct linspaced_op_impl const Scalar m_high; const Index m_size1; const Scalar m_step; - const Packet m_interPacket; const bool m_flip; }; diff --git a/xs/src/eigen/Eigen/src/Core/functors/StlFunctors.h b/xs/src/eigen/Eigen/src/Core/functors/StlFunctors.h index 6df3fa501..9c1d75850 100644 --- a/xs/src/eigen/Eigen/src/Core/functors/StlFunctors.h +++ b/xs/src/eigen/Eigen/src/Core/functors/StlFunctors.h @@ -83,13 +83,17 @@ struct functor_traits > { enum { Cost = functor_traits::Cost, PacketAccess = false }; }; #endif +#if (__cplusplus < 201703L) && (EIGEN_COMP_MSVC < 1910) +// std::unary_negate is deprecated since c++17 and will be removed in c++20 template struct functor_traits > { enum { Cost = 1 + functor_traits::Cost, PacketAccess = false }; }; +// std::binary_negate is deprecated since c++17 and will be removed in c++20 template struct functor_traits > { enum { Cost = 1 + functor_traits::Cost, PacketAccess = false }; }; +#endif #ifdef EIGEN_STDEXT_SUPPORT diff --git a/xs/src/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h b/xs/src/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h index 7122efa60..e844e37d1 100644 --- a/xs/src/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h +++ b/xs/src/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h @@ -269,10 +269,13 @@ struct general_product_to_triangular_selector enum { IsRowMajor = (internal::traits::Flags&RowMajorBit) ? 1 : 0, LhsIsRowMajor = _ActualLhs::Flags&RowMajorBit ? 1 : 0, - RhsIsRowMajor = _ActualRhs::Flags&RowMajorBit ? 1 : 0 + RhsIsRowMajor = _ActualRhs::Flags&RowMajorBit ? 1 : 0, + SkipDiag = (UpLo&(UnitDiag|ZeroDiag))!=0 }; Index size = mat.cols(); + if(SkipDiag) + size--; Index depth = actualLhs.cols(); typedef internal::gemm_blocking_space internal::general_matrix_matrix_triangular_product + IsRowMajor ? RowMajor : ColMajor, UpLo&(Lower|Upper)> ::run(size, depth, - &actualLhs.coeffRef(0,0), actualLhs.outerStride(), &actualRhs.coeffRef(0,0), actualRhs.outerStride(), - mat.data(), mat.outerStride(), actualAlpha, blocking); + &actualLhs.coeffRef(SkipDiag&&(UpLo&Lower)==Lower ? 1 : 0,0), actualLhs.outerStride(), + &actualRhs.coeffRef(0,SkipDiag&&(UpLo&Upper)==Upper ? 1 : 0), actualRhs.outerStride(), + mat.data() + (SkipDiag ? (bool(IsRowMajor) != ((UpLo&Lower)==Lower) ? 1 : mat.outerStride() ) : 0), mat.outerStride(), actualAlpha, blocking); } }; @@ -294,6 +298,7 @@ template template TriangularView& TriangularViewImpl::_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::InnerSize==1>::run(derived().nestedExpression().const_cast_derived(), prod, alpha, beta); diff --git a/xs/src/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h b/xs/src/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h index 5b7c15cca..9176a1382 100644 --- a/xs/src/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h +++ b/xs/src/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h @@ -52,7 +52,7 @@ struct general_matrix_matrix_triangular_product& blocking) \ { \ - if (lhs==rhs) { \ + if ( lhs==rhs && ((UpLo&(Lower|Upper)==UpLo)) ) { \ general_matrix_matrix_rankupdate \ ::run(size,depth,lhs,lhsStride,rhs,rhsStride,res,resStride,alpha,blocking); \ } else { \ @@ -88,7 +88,7 @@ struct general_matrix_matrix_rankupdate(lhsStride), ldc=convert_index(resStride), n=convert_index(size), k=convert_index(depth); \ char uplo=((IsLower) ? 'L' : 'U'), trans=((AStorageOrder==RowMajor) ? 'T':'N'); \ EIGTYPE beta(1); \ - BLASFUNC(&uplo, &trans, &n, &k, &numext::real_ref(alpha), lhs, &lda, &numext::real_ref(beta), res, &ldc); \ + BLASFUNC(&uplo, &trans, &n, &k, (const BLASTYPE*)&numext::real_ref(alpha), lhs, &lda, (const BLASTYPE*)&numext::real_ref(beta), res, &ldc); \ } \ }; @@ -125,9 +125,13 @@ struct general_matrix_matrix_rankupdate(b_tmp.outerStride()); \ } else b = _rhs; \ \ - BLASPREFIX##gemm_(&transa, &transb, &m, &n, &k, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, &numext::real_ref(beta), (BLASTYPE*)res, &ldc); \ + BLASFUNC(&transa, &transb, &m, &n, &k, (const BLASTYPE*)&numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, (const BLASTYPE*)&numext::real_ref(beta), (BLASTYPE*)res, &ldc); \ }}; -GEMM_SPECIALIZATION(double, d, double, d) -GEMM_SPECIALIZATION(float, f, float, s) -GEMM_SPECIALIZATION(dcomplex, cd, double, z) -GEMM_SPECIALIZATION(scomplex, cf, float, c) +#ifdef EIGEN_USE_MKL +GEMM_SPECIALIZATION(double, d, double, dgemm) +GEMM_SPECIALIZATION(float, f, float, sgemm) +GEMM_SPECIALIZATION(dcomplex, cd, MKL_Complex16, zgemm) +GEMM_SPECIALIZATION(scomplex, cf, MKL_Complex8, cgemm) +#else +GEMM_SPECIALIZATION(double, d, double, dgemm_) +GEMM_SPECIALIZATION(float, f, float, sgemm_) +GEMM_SPECIALIZATION(dcomplex, cd, double, zgemm_) +GEMM_SPECIALIZATION(scomplex, cf, float, cgemm_) +#endif } // end namespase internal diff --git a/xs/src/eigen/Eigen/src/Core/products/GeneralMatrixVector.h b/xs/src/eigen/Eigen/src/Core/products/GeneralMatrixVector.h index 3c1a7fc40..a597c1f4e 100644 --- a/xs/src/eigen/Eigen/src/Core/products/GeneralMatrixVector.h +++ b/xs/src/eigen/Eigen/src/Core/products/GeneralMatrixVector.h @@ -183,8 +183,8 @@ EIGEN_DONT_INLINE void general_matrix_vector_product \ struct general_matrix_vector_product_gemv \ { \ @@ -113,14 +113,21 @@ static void run( \ x_ptr=x_tmp.data(); \ incx=1; \ } else x_ptr=rhs; \ - BLASPREFIX##gemv_(&trans, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)lhs, &lda, (const BLASTYPE*)x_ptr, &incx, &numext::real_ref(beta), (BLASTYPE*)res, &incy); \ + 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); \ }\ }; -EIGEN_BLAS_GEMV_SPECIALIZATION(double, double, d) -EIGEN_BLAS_GEMV_SPECIALIZATION(float, float, s) -EIGEN_BLAS_GEMV_SPECIALIZATION(dcomplex, double, z) -EIGEN_BLAS_GEMV_SPECIALIZATION(scomplex, float, c) +#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 diff --git a/xs/src/eigen/Eigen/src/Core/products/SelfadjointMatrixMatrix_BLAS.h b/xs/src/eigen/Eigen/src/Core/products/SelfadjointMatrixMatrix_BLAS.h index a45238d69..9a5318507 100644 --- a/xs/src/eigen/Eigen/src/Core/products/SelfadjointMatrixMatrix_BLAS.h +++ b/xs/src/eigen/Eigen/src/Core/products/SelfadjointMatrixMatrix_BLAS.h @@ -40,7 +40,7 @@ namespace internal { /* Optimized selfadjoint matrix * matrix (?SYMM/?HEMM) product */ -#define EIGEN_BLAS_SYMM_L(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \ +#define EIGEN_BLAS_SYMM_L(EIGTYPE, BLASTYPE, EIGPREFIX, BLASFUNC) \ template \ @@ -81,13 +81,13 @@ struct product_selfadjoint_matrix(b_tmp.outerStride()); \ } else b = _rhs; \ \ - BLASPREFIX##symm_(&side, &uplo, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, &numext::real_ref(beta), (BLASTYPE*)res, &ldc); \ + BLASFUNC(&side, &uplo, &m, &n, (const BLASTYPE*)&numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, (const BLASTYPE*)&numext::real_ref(beta), (BLASTYPE*)res, &ldc); \ \ } \ }; -#define EIGEN_BLAS_HEMM_L(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \ +#define EIGEN_BLAS_HEMM_L(EIGTYPE, BLASTYPE, EIGPREFIX, BLASFUNC) \ template \ @@ -144,20 +144,26 @@ struct product_selfadjoint_matrix(b_tmp.outerStride()); \ } \ \ - BLASPREFIX##hemm_(&side, &uplo, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, &numext::real_ref(beta), (BLASTYPE*)res, &ldc); \ + BLASFUNC(&side, &uplo, &m, &n, (const BLASTYPE*)&numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, (const BLASTYPE*)&numext::real_ref(beta), (BLASTYPE*)res, &ldc); \ \ } \ }; -EIGEN_BLAS_SYMM_L(double, double, d, d) -EIGEN_BLAS_SYMM_L(float, float, f, s) -EIGEN_BLAS_HEMM_L(dcomplex, double, cd, z) -EIGEN_BLAS_HEMM_L(scomplex, float, cf, c) - +#ifdef EIGEN_USE_MKL +EIGEN_BLAS_SYMM_L(double, double, d, dsymm) +EIGEN_BLAS_SYMM_L(float, float, f, ssymm) +EIGEN_BLAS_HEMM_L(dcomplex, MKL_Complex16, cd, zhemm) +EIGEN_BLAS_HEMM_L(scomplex, MKL_Complex8, cf, chemm) +#else +EIGEN_BLAS_SYMM_L(double, double, d, dsymm_) +EIGEN_BLAS_SYMM_L(float, float, f, ssymm_) +EIGEN_BLAS_HEMM_L(dcomplex, double, cd, zhemm_) +EIGEN_BLAS_HEMM_L(scomplex, float, cf, chemm_) +#endif /* Optimized matrix * selfadjoint matrix (?SYMM/?HEMM) product */ -#define EIGEN_BLAS_SYMM_R(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \ +#define EIGEN_BLAS_SYMM_R(EIGTYPE, BLASTYPE, EIGPREFIX, BLASFUNC) \ template \ @@ -197,13 +203,13 @@ struct product_selfadjoint_matrix(b_tmp.outerStride()); \ } else b = _lhs; \ \ - BLASPREFIX##symm_(&side, &uplo, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, &numext::real_ref(beta), (BLASTYPE*)res, &ldc); \ + BLASFUNC(&side, &uplo, &m, &n, (const BLASTYPE*)&numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, (const BLASTYPE*)&numext::real_ref(beta), (BLASTYPE*)res, &ldc); \ \ } \ }; -#define EIGEN_BLAS_HEMM_R(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \ +#define EIGEN_BLAS_HEMM_R(EIGTYPE, BLASTYPE, EIGPREFIX, BLASFUNC) \ template \ @@ -259,15 +265,21 @@ struct product_selfadjoint_matrix(b_tmp.outerStride()); \ } \ \ - BLASPREFIX##hemm_(&side, &uplo, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, &numext::real_ref(beta), (BLASTYPE*)res, &ldc); \ + BLASFUNC(&side, &uplo, &m, &n, (const BLASTYPE*)&numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, (const BLASTYPE*)&numext::real_ref(beta), (BLASTYPE*)res, &ldc); \ } \ }; -EIGEN_BLAS_SYMM_R(double, double, d, d) -EIGEN_BLAS_SYMM_R(float, float, f, s) -EIGEN_BLAS_HEMM_R(dcomplex, double, cd, z) -EIGEN_BLAS_HEMM_R(scomplex, float, cf, c) - +#ifdef EIGEN_USE_MKL +EIGEN_BLAS_SYMM_R(double, double, d, dsymm) +EIGEN_BLAS_SYMM_R(float, float, f, ssymm) +EIGEN_BLAS_HEMM_R(dcomplex, MKL_Complex16, cd, zhemm) +EIGEN_BLAS_HEMM_R(scomplex, MKL_Complex8, cf, chemm) +#else +EIGEN_BLAS_SYMM_R(double, double, d, dsymm_) +EIGEN_BLAS_SYMM_R(float, float, f, ssymm_) +EIGEN_BLAS_HEMM_R(dcomplex, double, cd, zhemm_) +EIGEN_BLAS_HEMM_R(scomplex, float, cf, chemm_) +#endif } // end namespace internal } // end namespace Eigen diff --git a/xs/src/eigen/Eigen/src/Core/products/SelfadjointMatrixVector_BLAS.h b/xs/src/eigen/Eigen/src/Core/products/SelfadjointMatrixVector_BLAS.h index 38f23accf..1238345e3 100644 --- a/xs/src/eigen/Eigen/src/Core/products/SelfadjointMatrixVector_BLAS.h +++ b/xs/src/eigen/Eigen/src/Core/products/SelfadjointMatrixVector_BLAS.h @@ -95,14 +95,21 @@ const EIGTYPE* _rhs, EIGTYPE* res, EIGTYPE alpha) \ x_tmp=map_x.conjugate(); \ x_ptr=x_tmp.data(); \ } else x_ptr=_rhs; \ - BLASFUNC(&uplo, &n, &numext::real_ref(alpha), (const BLASTYPE*)lhs, &lda, (const BLASTYPE*)x_ptr, &incx, &numext::real_ref(beta), (BLASTYPE*)res, &incy); \ + BLASFUNC(&uplo, &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_SYMV_SPECIALIZATION(double, double, dsymv) +EIGEN_BLAS_SYMV_SPECIALIZATION(float, float, ssymv) +EIGEN_BLAS_SYMV_SPECIALIZATION(dcomplex, MKL_Complex16, zhemv) +EIGEN_BLAS_SYMV_SPECIALIZATION(scomplex, MKL_Complex8, chemv) +#else EIGEN_BLAS_SYMV_SPECIALIZATION(double, double, dsymv_) EIGEN_BLAS_SYMV_SPECIALIZATION(float, float, ssymv_) EIGEN_BLAS_SYMV_SPECIALIZATION(dcomplex, double, zhemv_) EIGEN_BLAS_SYMV_SPECIALIZATION(scomplex, float, chemv_) +#endif } // end namespace internal diff --git a/xs/src/eigen/Eigen/src/Core/products/TriangularMatrixMatrix.h b/xs/src/eigen/Eigen/src/Core/products/TriangularMatrixMatrix.h index 6ec5a8a0b..f784507e7 100644 --- a/xs/src/eigen/Eigen/src/Core/products/TriangularMatrixMatrix.h +++ b/xs/src/eigen/Eigen/src/Core/products/TriangularMatrixMatrix.h @@ -137,7 +137,13 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix triangularBuffer((internal::constructor_without_unaligned_array_assert())); + // To work around an "error: member reference base type 'Matrix<...> + // (Eigen::internal::constructor_without_unaligned_array_assert (*)())' is + // not a structure or union" compilation error in nvcc (tested V8.0.61), + // create a dummy internal::constructor_without_unaligned_array_assert + // object to pass to the Matrix constructor. + internal::constructor_without_unaligned_array_assert a; + Matrix triangularBuffer(a); triangularBuffer.setZero(); if((Mode&ZeroDiag)==ZeroDiag) triangularBuffer.diagonal().setZero(); @@ -284,7 +290,8 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix triangularBuffer((internal::constructor_without_unaligned_array_assert())); + internal::constructor_without_unaligned_array_assert a; + Matrix triangularBuffer(a); triangularBuffer.setZero(); if((Mode&ZeroDiag)==ZeroDiag) triangularBuffer.diagonal().setZero(); @@ -393,7 +400,9 @@ struct triangular_product_impl { template static void run(Dest& dst, const Lhs &a_lhs, const Rhs &a_rhs, const typename Dest::Scalar& alpha) { - typedef typename Dest::Scalar Scalar; + typedef typename Lhs::Scalar LhsScalar; + typedef typename Rhs::Scalar RhsScalar; + typedef typename Dest::Scalar Scalar; typedef internal::blas_traits LhsBlasTraits; typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType; @@ -405,8 +414,9 @@ struct triangular_product_impl typename internal::add_const_on_value_type::type lhs = LhsBlasTraits::extract(a_lhs); typename internal::add_const_on_value_type::type rhs = RhsBlasTraits::extract(a_rhs); - Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(a_lhs) - * RhsBlasTraits::extractScalarFactor(a_rhs); + LhsScalar lhs_alpha = LhsBlasTraits::extractScalarFactor(a_lhs); + RhsScalar rhs_alpha = RhsBlasTraits::extractScalarFactor(a_rhs); + Scalar actualAlpha = alpha * lhs_alpha * rhs_alpha; typedef internal::gemm_blocking_space<(Dest::Flags&RowMajorBit) ? RowMajor : ColMajor,Scalar,Scalar, Lhs::MaxRowsAtCompileTime, Rhs::MaxColsAtCompileTime, Lhs::MaxColsAtCompileTime,4> BlockingType; @@ -431,6 +441,21 @@ struct triangular_product_impl &dst.coeffRef(0,0), dst.outerStride(), // result info actualAlpha, blocking ); + + // Apply correction if the diagonal is unit and a scalar factor was nested: + if ((Mode&UnitDiag)==UnitDiag) + { + if (LhsIsTriangular && lhs_alpha!=LhsScalar(1)) + { + Index diagSize = (std::min)(lhs.rows(),lhs.cols()); + dst.topRows(diagSize) -= ((lhs_alpha-LhsScalar(1))*a_rhs).topRows(diagSize); + } + else if ((!LhsIsTriangular) && rhs_alpha!=RhsScalar(1)) + { + Index diagSize = (std::min)(rhs.rows(),rhs.cols()); + dst.leftCols(diagSize) -= (rhs_alpha-RhsScalar(1))*a_lhs.leftCols(diagSize); + } + } } }; diff --git a/xs/src/eigen/Eigen/src/Core/products/TriangularMatrixMatrix_BLAS.h b/xs/src/eigen/Eigen/src/Core/products/TriangularMatrixMatrix_BLAS.h index aecded6bb..a25197ab0 100644 --- a/xs/src/eigen/Eigen/src/Core/products/TriangularMatrixMatrix_BLAS.h +++ b/xs/src/eigen/Eigen/src/Core/products/TriangularMatrixMatrix_BLAS.h @@ -75,7 +75,7 @@ EIGEN_BLAS_TRMM_SPECIALIZE(scomplex, true) EIGEN_BLAS_TRMM_SPECIALIZE(scomplex, false) // implements col-major += alpha * op(triangular) * op(general) -#define EIGEN_BLAS_TRMM_L(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \ +#define EIGEN_BLAS_TRMM_L(EIGTYPE, BLASTYPE, EIGPREFIX, BLASFUNC) \ template \ @@ -172,7 +172,7 @@ struct product_triangular_matrix_matrix_trmm > res_tmp(res,rows,cols,OuterStride<>(resStride)); \ @@ -180,13 +180,20 @@ struct product_triangular_matrix_matrix_trmm \ @@ -282,7 +289,7 @@ struct product_triangular_matrix_matrix_trmm > res_tmp(res,rows,cols,OuterStride<>(resStride)); \ @@ -290,11 +297,17 @@ struct product_triangular_matrix_matrix_trmm struct trmv_selector typename internal::add_const_on_value_type::type actualLhs = LhsBlasTraits::extract(lhs); typename internal::add_const_on_value_type::type actualRhs = RhsBlasTraits::extract(rhs); - ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(lhs) - * RhsBlasTraits::extractScalarFactor(rhs); + LhsScalar lhs_alpha = LhsBlasTraits::extractScalarFactor(lhs); + RhsScalar rhs_alpha = RhsBlasTraits::extractScalarFactor(rhs); + ResScalar actualAlpha = alpha * lhs_alpha * rhs_alpha; enum { // FIXME find a way to allow an inner stride on the result if packet_traits::size==1 @@ -274,6 +275,12 @@ template struct trmv_selector else dest = MappedDest(actualDestPtr, dest.size()); } + + if ( ((Mode&UnitDiag)==UnitDiag) && (lhs_alpha!=LhsScalar(1)) ) + { + Index diagSize = (std::min)(lhs.rows(),lhs.cols()); + dest.head(diagSize) -= (lhs_alpha-LhsScalar(1))*rhs.head(diagSize); + } } }; @@ -295,8 +302,9 @@ template struct trmv_selector typename add_const::type actualLhs = LhsBlasTraits::extract(lhs); typename add_const::type actualRhs = RhsBlasTraits::extract(rhs); - ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(lhs) - * RhsBlasTraits::extractScalarFactor(rhs); + LhsScalar lhs_alpha = LhsBlasTraits::extractScalarFactor(lhs); + RhsScalar rhs_alpha = RhsBlasTraits::extractScalarFactor(rhs); + ResScalar actualAlpha = alpha * lhs_alpha * rhs_alpha; enum { DirectlyUseRhs = ActualRhsTypeCleaned::InnerStrideAtCompileTime==1 @@ -326,6 +334,12 @@ template struct trmv_selector actualRhsPtr,1, dest.data(),dest.innerStride(), actualAlpha); + + if ( ((Mode&UnitDiag)==UnitDiag) && (lhs_alpha!=LhsScalar(1)) ) + { + Index diagSize = (std::min)(lhs.rows(),lhs.cols()); + dest.head(diagSize) -= (lhs_alpha-LhsScalar(1))*rhs.head(diagSize); + } } }; diff --git a/xs/src/eigen/Eigen/src/Core/products/TriangularMatrixVector_BLAS.h b/xs/src/eigen/Eigen/src/Core/products/TriangularMatrixVector_BLAS.h index 07bf26ce5..3d47a2b94 100644 --- a/xs/src/eigen/Eigen/src/Core/products/TriangularMatrixVector_BLAS.h +++ b/xs/src/eigen/Eigen/src/Core/products/TriangularMatrixVector_BLAS.h @@ -71,7 +71,7 @@ EIGEN_BLAS_TRMV_SPECIALIZE(dcomplex) EIGEN_BLAS_TRMV_SPECIALIZE(scomplex) // implements col-major: res += alpha * op(triangular) * vector -#define EIGEN_BLAS_TRMV_CM(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \ +#define EIGEN_BLAS_TRMV_CM(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX, BLASPOSTFIX) \ template \ struct triangular_matrix_vector_product_trmv { \ enum { \ @@ -121,10 +121,10 @@ struct triangular_matrix_vector_product_trmv(size); \ n = convert_index(cols-size); \ } \ - BLASPREFIX##gemv_(&trans, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)x, &incx, &numext::real_ref(beta), (BLASTYPE*)y, &incy); \ + BLASPREFIX##gemv##BLASPOSTFIX(&trans, &m, &n, (const BLASTYPE*)&numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)x, &incx, (const BLASTYPE*)&numext::real_ref(beta), (BLASTYPE*)y, &incy); \ } \ } \ }; -EIGEN_BLAS_TRMV_CM(double, double, d, d) -EIGEN_BLAS_TRMV_CM(dcomplex, double, cd, z) -EIGEN_BLAS_TRMV_CM(float, float, f, s) -EIGEN_BLAS_TRMV_CM(scomplex, float, cf, c) +#ifdef EIGEN_USE_MKL +EIGEN_BLAS_TRMV_CM(double, double, d, d,) +EIGEN_BLAS_TRMV_CM(dcomplex, MKL_Complex16, cd, z,) +EIGEN_BLAS_TRMV_CM(float, float, f, s,) +EIGEN_BLAS_TRMV_CM(scomplex, MKL_Complex8, cf, c,) +#else +EIGEN_BLAS_TRMV_CM(double, double, d, d, _) +EIGEN_BLAS_TRMV_CM(dcomplex, double, cd, z, _) +EIGEN_BLAS_TRMV_CM(float, float, f, s, _) +EIGEN_BLAS_TRMV_CM(scomplex, float, cf, c, _) +#endif // implements row-major: res += alpha * op(triangular) * vector -#define EIGEN_BLAS_TRMV_RM(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \ +#define EIGEN_BLAS_TRMV_RM(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX, BLASPOSTFIX) \ template \ struct triangular_matrix_vector_product_trmv { \ enum { \ @@ -203,10 +210,10 @@ struct triangular_matrix_vector_product_trmv(size); \ n = convert_index(cols-size); \ } \ - BLASPREFIX##gemv_(&trans, &n, &m, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)x, &incx, &numext::real_ref(beta), (BLASTYPE*)y, &incy); \ + BLASPREFIX##gemv##BLASPOSTFIX(&trans, &n, &m, (const BLASTYPE*)&numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)x, &incx, (const BLASTYPE*)&numext::real_ref(beta), (BLASTYPE*)y, &incy); \ } \ } \ }; -EIGEN_BLAS_TRMV_RM(double, double, d, d) -EIGEN_BLAS_TRMV_RM(dcomplex, double, cd, z) -EIGEN_BLAS_TRMV_RM(float, float, f, s) -EIGEN_BLAS_TRMV_RM(scomplex, float, cf, c) +#ifdef EIGEN_USE_MKL +EIGEN_BLAS_TRMV_RM(double, double, d, d,) +EIGEN_BLAS_TRMV_RM(dcomplex, MKL_Complex16, cd, z,) +EIGEN_BLAS_TRMV_RM(float, float, f, s,) +EIGEN_BLAS_TRMV_RM(scomplex, MKL_Complex8, cf, c,) +#else +EIGEN_BLAS_TRMV_RM(double, double, d, d,_) +EIGEN_BLAS_TRMV_RM(dcomplex, double, cd, z,_) +EIGEN_BLAS_TRMV_RM(float, float, f, s,_) +EIGEN_BLAS_TRMV_RM(scomplex, float, cf, c,_) +#endif } // end namespase internal diff --git a/xs/src/eigen/Eigen/src/Core/products/TriangularSolverMatrix_BLAS.h b/xs/src/eigen/Eigen/src/Core/products/TriangularSolverMatrix_BLAS.h index 88c0fb794..f0775116a 100644 --- a/xs/src/eigen/Eigen/src/Core/products/TriangularSolverMatrix_BLAS.h +++ b/xs/src/eigen/Eigen/src/Core/products/TriangularSolverMatrix_BLAS.h @@ -38,7 +38,7 @@ namespace Eigen { namespace internal { // implements LeftSide op(triangular)^-1 * general -#define EIGEN_BLAS_TRSM_L(EIGTYPE, BLASTYPE, BLASPREFIX) \ +#define EIGEN_BLAS_TRSM_L(EIGTYPE, BLASTYPE, BLASFUNC) \ template \ struct triangular_solve_matrix \ { \ @@ -80,18 +80,24 @@ struct triangular_solve_matrix \ struct triangular_solve_matrix \ { \ @@ -133,16 +139,22 @@ struct triangular_solve_matrix /*Check IMKL version for compatibility: < 10.3 is not usable with Eigen*/ @@ -108,6 +109,10 @@ #endif #endif +#if defined(EIGEN_USE_BLAS) && !defined(EIGEN_USE_MKL) +#include "../../misc/blas.h" +#endif + namespace Eigen { typedef std::complex dcomplex; @@ -121,8 +126,5 @@ typedef int BlasIndex; } // end namespace Eigen -#if defined(EIGEN_USE_BLAS) -#include "../../misc/blas.h" -#endif #endif // EIGEN_MKL_SUPPORT_H diff --git a/xs/src/eigen/Eigen/src/Core/util/Macros.h b/xs/src/eigen/Eigen/src/Core/util/Macros.h index 427d3cd6b..02d21d2cd 100644 --- a/xs/src/eigen/Eigen/src/Core/util/Macros.h +++ b/xs/src/eigen/Eigen/src/Core/util/Macros.h @@ -13,7 +13,7 @@ #define EIGEN_WORLD_VERSION 3 #define EIGEN_MAJOR_VERSION 3 -#define EIGEN_MINOR_VERSION 3 +#define EIGEN_MINOR_VERSION 5 #define EIGEN_VERSION_AT_LEAST(x,y,z) (EIGEN_WORLD_VERSION>x || (EIGEN_WORLD_VERSION>=x && \ (EIGEN_MAJOR_VERSION>y || (EIGEN_MAJOR_VERSION>=y && \ @@ -399,7 +399,7 @@ // Does the compiler support variadic templates? #ifndef EIGEN_HAS_VARIADIC_TEMPLATES #if EIGEN_MAX_CPP_VER>=11 && (__cplusplus > 199711L || EIGEN_COMP_MSVC >= 1900) \ - && ( !defined(__NVCC__) || !EIGEN_ARCH_ARM_OR_ARM64 || (defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000) ) + && (!defined(__NVCC__) || !EIGEN_ARCH_ARM_OR_ARM64 || (EIGEN_CUDACC_VER >= 80000) ) // ^^ Disable the use of variadic templates when compiling with versions of nvcc older than 8.0 on ARM devices: // this prevents nvcc from crashing when compiling Eigen on Tegra X1 #define EIGEN_HAS_VARIADIC_TEMPLATES 1 @@ -413,7 +413,7 @@ #ifdef __CUDACC__ // Const expressions are supported provided that c++11 is enabled and we're using either clang or nvcc 7.5 or above -#if EIGEN_MAX_CPP_VER>=14 && (__cplusplus > 199711L && defined(__CUDACC_VER__) && (EIGEN_COMP_CLANG || __CUDACC_VER__ >= 70500)) +#if EIGEN_MAX_CPP_VER>=14 && (__cplusplus > 199711L && (EIGEN_COMP_CLANG || EIGEN_CUDACC_VER >= 70500)) #define EIGEN_HAS_CONSTEXPR 1 #endif #elif EIGEN_MAX_CPP_VER>=14 && (__has_feature(cxx_relaxed_constexpr) || (defined(__cplusplus) && __cplusplus >= 201402L) || \ @@ -487,11 +487,13 @@ // EIGEN_STRONG_INLINE is a stronger version of the inline, using __forceinline on MSVC, // but it still doesn't use GCC's always_inline. This is useful in (common) situations where MSVC needs forceinline // but GCC is still doing fine with just inline. +#ifndef EIGEN_STRONG_INLINE #if EIGEN_COMP_MSVC || EIGEN_COMP_ICC #define EIGEN_STRONG_INLINE __forceinline #else #define EIGEN_STRONG_INLINE inline #endif +#endif // EIGEN_ALWAYS_INLINE is the stronget, it has the effect of making the function inline and adding every possible // attribute to maximize inlining. This should only be used when really necessary: in particular, @@ -812,7 +814,8 @@ namespace Eigen { // just an empty macro ! #define EIGEN_EMPTY -#if EIGEN_COMP_MSVC_STRICT && (EIGEN_COMP_MSVC < 1900 || defined(__CUDACC_VER__)) // for older MSVC versions, as well as 1900 && CUDA 8, using the base operator is sufficient (cf Bugs 1000, 1324) +#if EIGEN_COMP_MSVC_STRICT && (EIGEN_COMP_MSVC < 1900 || EIGEN_CUDACC_VER>0) + // for older MSVC versions, as well as 1900 && CUDA 8, using the base operator is sufficient (cf Bugs 1000, 1324) #define EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived) \ using Base::operator =; #elif EIGEN_COMP_CLANG // workaround clang bug (see http://forum.kde.org/viewtopic.php?f=74&t=102653) @@ -986,7 +989,13 @@ namespace Eigen { # define EIGEN_NOEXCEPT # define EIGEN_NOEXCEPT_IF(x) # define EIGEN_NO_THROW throw() -# define EIGEN_EXCEPTION_SPEC(X) throw(X) +# if EIGEN_COMP_MSVC + // MSVC does not support exception specifications (warning C4290), + // and they are deprecated in c++11 anyway. +# define EIGEN_EXCEPTION_SPEC(X) throw() +# else +# define EIGEN_EXCEPTION_SPEC(X) throw(X) +# endif #endif #endif // EIGEN_MACROS_H diff --git a/xs/src/eigen/Eigen/src/Core/util/Memory.h b/xs/src/eigen/Eigen/src/Core/util/Memory.h index c634d7ea0..66cdbd8dd 100644 --- a/xs/src/eigen/Eigen/src/Core/util/Memory.h +++ b/xs/src/eigen/Eigen/src/Core/util/Memory.h @@ -70,7 +70,7 @@ inline void throw_std_bad_alloc() throw std::bad_alloc(); #else std::size_t huge = static_cast(-1); - new int[huge]; + ::operator new(huge); #endif } @@ -493,7 +493,7 @@ template struct smart_copy_helper { IntPtr size = IntPtr(end)-IntPtr(start); if(size==0) return; eigen_internal_assert(start!=0 && end!=0 && target!=0); - memcpy(target, start, size); + std::memcpy(target, start, size); } }; @@ -696,7 +696,15 @@ template void swap(scoped_array &a,scoped_array &b) /** \class aligned_allocator * \ingroup Core_Module * -* \brief STL compatible allocator to use with with 16 byte aligned types +* \brief STL compatible allocator to use with types requiring a non standrad alignment. +* +* The memory is aligned as for dynamically aligned matrix/array types such as MatrixXd. +* By default, it will thus provide at least 16 bytes alignment and more in following cases: +* - 32 bytes alignment if AVX is enabled. +* - 64 bytes alignment if AVX512 is enabled. +* +* This can be controled using the \c EIGEN_MAX_ALIGN_BYTES macro as documented +* \link TopicPreprocessorDirectivesPerformance there \endlink. * * Example: * \code diff --git a/xs/src/eigen/Eigen/src/Core/util/Meta.h b/xs/src/eigen/Eigen/src/Core/util/Meta.h index 7f6370755..1d73f05d6 100644 --- a/xs/src/eigen/Eigen/src/Core/util/Meta.h +++ b/xs/src/eigen/Eigen/src/Core/util/Meta.h @@ -485,6 +485,26 @@ T div_ceil(const T &a, const T &b) return (a+b-1) / b; } +// The aim of the following functions is to bypass -Wfloat-equal warnings +// when we really want a strict equality comparison on floating points. +template EIGEN_STRONG_INLINE +bool equal_strict(const X& x,const Y& y) { return x == y; } + +template<> EIGEN_STRONG_INLINE +bool equal_strict(const float& x,const float& y) { return std::equal_to()(x,y); } + +template<> EIGEN_STRONG_INLINE +bool equal_strict(const double& x,const double& y) { return std::equal_to()(x,y); } + +template EIGEN_STRONG_INLINE +bool not_equal_strict(const X& x,const Y& y) { return x != y; } + +template<> EIGEN_STRONG_INLINE +bool not_equal_strict(const float& x,const float& y) { return std::not_equal_to()(x,y); } + +template<> EIGEN_STRONG_INLINE +bool not_equal_strict(const double& x,const double& y) { return std::not_equal_to()(x,y); } + } // end namespace numext } // end namespace Eigen diff --git a/xs/src/eigen/Eigen/src/Core/util/StaticAssert.h b/xs/src/eigen/Eigen/src/Core/util/StaticAssert.h index 983361a45..500e47792 100644 --- a/xs/src/eigen/Eigen/src/Core/util/StaticAssert.h +++ b/xs/src/eigen/Eigen/src/Core/util/StaticAssert.h @@ -24,6 +24,7 @@ * */ +#ifndef EIGEN_STATIC_ASSERT #ifndef EIGEN_NO_STATIC_ASSERT #if EIGEN_MAX_CPP_VER>=11 && (__has_feature(cxx_static_assert) || (defined(__cplusplus) && __cplusplus >= 201103L) || (EIGEN_COMP_MSVC >= 1600)) @@ -44,64 +45,65 @@ struct static_assertion { enum { - YOU_TRIED_CALLING_A_VECTOR_METHOD_ON_A_MATRIX, - YOU_MIXED_VECTORS_OF_DIFFERENT_SIZES, - YOU_MIXED_MATRICES_OF_DIFFERENT_SIZES, - THIS_METHOD_IS_ONLY_FOR_VECTORS_OF_A_SPECIFIC_SIZE, - THIS_METHOD_IS_ONLY_FOR_MATRICES_OF_A_SPECIFIC_SIZE, - THIS_METHOD_IS_ONLY_FOR_OBJECTS_OF_A_SPECIFIC_SIZE, - OUT_OF_RANGE_ACCESS, - YOU_MADE_A_PROGRAMMING_MISTAKE, - EIGEN_INTERNAL_ERROR_PLEASE_FILE_A_BUG_REPORT, - EIGEN_INTERNAL_COMPILATION_ERROR_OR_YOU_MADE_A_PROGRAMMING_MISTAKE, - YOU_CALLED_A_FIXED_SIZE_METHOD_ON_A_DYNAMIC_SIZE_MATRIX_OR_VECTOR, - YOU_CALLED_A_DYNAMIC_SIZE_METHOD_ON_A_FIXED_SIZE_MATRIX_OR_VECTOR, - UNALIGNED_LOAD_AND_STORE_OPERATIONS_UNIMPLEMENTED_ON_ALTIVEC, - THIS_FUNCTION_IS_NOT_FOR_INTEGER_NUMERIC_TYPES, - FLOATING_POINT_ARGUMENT_PASSED__INTEGER_WAS_EXPECTED, - NUMERIC_TYPE_MUST_BE_REAL, - COEFFICIENT_WRITE_ACCESS_TO_SELFADJOINT_NOT_SUPPORTED, - WRITING_TO_TRIANGULAR_PART_WITH_UNIT_DIAGONAL_IS_NOT_SUPPORTED, - THIS_METHOD_IS_ONLY_FOR_FIXED_SIZE, - INVALID_MATRIX_PRODUCT, - INVALID_VECTOR_VECTOR_PRODUCT__IF_YOU_WANTED_A_DOT_OR_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTIONS, - INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION, - YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY, - THIS_METHOD_IS_ONLY_FOR_COLUMN_MAJOR_MATRICES, - THIS_METHOD_IS_ONLY_FOR_ROW_MAJOR_MATRICES, - INVALID_MATRIX_TEMPLATE_PARAMETERS, - INVALID_MATRIXBASE_TEMPLATE_PARAMETERS, - BOTH_MATRICES_MUST_HAVE_THE_SAME_STORAGE_ORDER, - THIS_METHOD_IS_ONLY_FOR_DIAGONAL_MATRIX, - THE_MATRIX_OR_EXPRESSION_THAT_YOU_PASSED_DOES_NOT_HAVE_THE_EXPECTED_TYPE, - THIS_METHOD_IS_ONLY_FOR_EXPRESSIONS_WITH_DIRECT_MEMORY_ACCESS_SUCH_AS_MAP_OR_PLAIN_MATRICES, - YOU_ALREADY_SPECIFIED_THIS_STRIDE, - INVALID_STORAGE_ORDER_FOR_THIS_VECTOR_EXPRESSION, - THE_BRACKET_OPERATOR_IS_ONLY_FOR_VECTORS__USE_THE_PARENTHESIS_OPERATOR_INSTEAD, - PACKET_ACCESS_REQUIRES_TO_HAVE_INNER_STRIDE_FIXED_TO_1, - THIS_METHOD_IS_ONLY_FOR_SPECIFIC_TRANSFORMATIONS, - YOU_CANNOT_MIX_ARRAYS_AND_MATRICES, - YOU_PERFORMED_AN_INVALID_TRANSFORMATION_CONVERSION, - THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY, - YOU_ARE_TRYING_TO_USE_AN_INDEX_BASED_ACCESSOR_ON_AN_EXPRESSION_THAT_DOES_NOT_SUPPORT_THAT, - THIS_METHOD_IS_ONLY_FOR_1x1_EXPRESSIONS, - THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS, - THIS_METHOD_IS_ONLY_FOR_EXPRESSIONS_OF_BOOL, - THIS_METHOD_IS_ONLY_FOR_ARRAYS_NOT_MATRICES, - YOU_PASSED_A_ROW_VECTOR_BUT_A_COLUMN_VECTOR_WAS_EXPECTED, - YOU_PASSED_A_COLUMN_VECTOR_BUT_A_ROW_VECTOR_WAS_EXPECTED, - THE_INDEX_TYPE_MUST_BE_A_SIGNED_TYPE, - THE_STORAGE_ORDER_OF_BOTH_SIDES_MUST_MATCH, - OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG, - IMPLICIT_CONVERSION_TO_SCALAR_IS_FOR_INNER_PRODUCT_ONLY, - STORAGE_LAYOUT_DOES_NOT_MATCH, - EIGEN_INTERNAL_ERROR_PLEASE_FILE_A_BUG_REPORT__INVALID_COST_VALUE, - THIS_COEFFICIENT_ACCESSOR_TAKING_ONE_ACCESS_IS_ONLY_FOR_EXPRESSIONS_ALLOWING_LINEAR_ACCESS, - MATRIX_FREE_CONJUGATE_GRADIENT_IS_COMPATIBLE_WITH_UPPER_UNION_LOWER_MODE_ONLY, - THIS_TYPE_IS_NOT_SUPPORTED, - STORAGE_KIND_MUST_MATCH, - STORAGE_INDEX_MUST_MATCH, - CHOLMOD_SUPPORTS_DOUBLE_PRECISION_ONLY + YOU_TRIED_CALLING_A_VECTOR_METHOD_ON_A_MATRIX=1, + YOU_MIXED_VECTORS_OF_DIFFERENT_SIZES=1, + YOU_MIXED_MATRICES_OF_DIFFERENT_SIZES=1, + THIS_METHOD_IS_ONLY_FOR_VECTORS_OF_A_SPECIFIC_SIZE=1, + THIS_METHOD_IS_ONLY_FOR_MATRICES_OF_A_SPECIFIC_SIZE=1, + THIS_METHOD_IS_ONLY_FOR_OBJECTS_OF_A_SPECIFIC_SIZE=1, + OUT_OF_RANGE_ACCESS=1, + YOU_MADE_A_PROGRAMMING_MISTAKE=1, + EIGEN_INTERNAL_ERROR_PLEASE_FILE_A_BUG_REPORT=1, + EIGEN_INTERNAL_COMPILATION_ERROR_OR_YOU_MADE_A_PROGRAMMING_MISTAKE=1, + YOU_CALLED_A_FIXED_SIZE_METHOD_ON_A_DYNAMIC_SIZE_MATRIX_OR_VECTOR=1, + YOU_CALLED_A_DYNAMIC_SIZE_METHOD_ON_A_FIXED_SIZE_MATRIX_OR_VECTOR=1, + UNALIGNED_LOAD_AND_STORE_OPERATIONS_UNIMPLEMENTED_ON_ALTIVEC=1, + THIS_FUNCTION_IS_NOT_FOR_INTEGER_NUMERIC_TYPES=1, + FLOATING_POINT_ARGUMENT_PASSED__INTEGER_WAS_EXPECTED=1, + NUMERIC_TYPE_MUST_BE_REAL=1, + COEFFICIENT_WRITE_ACCESS_TO_SELFADJOINT_NOT_SUPPORTED=1, + WRITING_TO_TRIANGULAR_PART_WITH_UNIT_DIAGONAL_IS_NOT_SUPPORTED=1, + THIS_METHOD_IS_ONLY_FOR_FIXED_SIZE=1, + INVALID_MATRIX_PRODUCT=1, + INVALID_VECTOR_VECTOR_PRODUCT__IF_YOU_WANTED_A_DOT_OR_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTIONS=1, + INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION=1, + YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY=1, + THIS_METHOD_IS_ONLY_FOR_COLUMN_MAJOR_MATRICES=1, + THIS_METHOD_IS_ONLY_FOR_ROW_MAJOR_MATRICES=1, + INVALID_MATRIX_TEMPLATE_PARAMETERS=1, + INVALID_MATRIXBASE_TEMPLATE_PARAMETERS=1, + BOTH_MATRICES_MUST_HAVE_THE_SAME_STORAGE_ORDER=1, + THIS_METHOD_IS_ONLY_FOR_DIAGONAL_MATRIX=1, + THE_MATRIX_OR_EXPRESSION_THAT_YOU_PASSED_DOES_NOT_HAVE_THE_EXPECTED_TYPE=1, + THIS_METHOD_IS_ONLY_FOR_EXPRESSIONS_WITH_DIRECT_MEMORY_ACCESS_SUCH_AS_MAP_OR_PLAIN_MATRICES=1, + YOU_ALREADY_SPECIFIED_THIS_STRIDE=1, + INVALID_STORAGE_ORDER_FOR_THIS_VECTOR_EXPRESSION=1, + THE_BRACKET_OPERATOR_IS_ONLY_FOR_VECTORS__USE_THE_PARENTHESIS_OPERATOR_INSTEAD=1, + PACKET_ACCESS_REQUIRES_TO_HAVE_INNER_STRIDE_FIXED_TO_1=1, + THIS_METHOD_IS_ONLY_FOR_SPECIFIC_TRANSFORMATIONS=1, + YOU_CANNOT_MIX_ARRAYS_AND_MATRICES=1, + YOU_PERFORMED_AN_INVALID_TRANSFORMATION_CONVERSION=1, + THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY=1, + YOU_ARE_TRYING_TO_USE_AN_INDEX_BASED_ACCESSOR_ON_AN_EXPRESSION_THAT_DOES_NOT_SUPPORT_THAT=1, + THIS_METHOD_IS_ONLY_FOR_1x1_EXPRESSIONS=1, + THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS=1, + THIS_METHOD_IS_ONLY_FOR_EXPRESSIONS_OF_BOOL=1, + THIS_METHOD_IS_ONLY_FOR_ARRAYS_NOT_MATRICES=1, + YOU_PASSED_A_ROW_VECTOR_BUT_A_COLUMN_VECTOR_WAS_EXPECTED=1, + YOU_PASSED_A_COLUMN_VECTOR_BUT_A_ROW_VECTOR_WAS_EXPECTED=1, + THE_INDEX_TYPE_MUST_BE_A_SIGNED_TYPE=1, + THE_STORAGE_ORDER_OF_BOTH_SIDES_MUST_MATCH=1, + OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG=1, + IMPLICIT_CONVERSION_TO_SCALAR_IS_FOR_INNER_PRODUCT_ONLY=1, + STORAGE_LAYOUT_DOES_NOT_MATCH=1, + EIGEN_INTERNAL_ERROR_PLEASE_FILE_A_BUG_REPORT__INVALID_COST_VALUE=1, + THIS_COEFFICIENT_ACCESSOR_TAKING_ONE_ACCESS_IS_ONLY_FOR_EXPRESSIONS_ALLOWING_LINEAR_ACCESS=1, + MATRIX_FREE_CONJUGATE_GRADIENT_IS_COMPATIBLE_WITH_UPPER_UNION_LOWER_MODE_ONLY=1, + THIS_TYPE_IS_NOT_SUPPORTED=1, + STORAGE_KIND_MUST_MATCH=1, + STORAGE_INDEX_MUST_MATCH=1, + CHOLMOD_SUPPORTS_DOUBLE_PRECISION_ONLY=1, + SELFADJOINTVIEW_ACCEPTS_UPPER_AND_LOWER_MODE_ONLY=1 }; }; @@ -131,7 +133,7 @@ #define EIGEN_STATIC_ASSERT(CONDITION,MSG) eigen_assert((CONDITION) && #MSG); #endif // EIGEN_NO_STATIC_ASSERT - +#endif // EIGEN_STATIC_ASSERT // static assertion failing if the type \a TYPE is not a vector type #define EIGEN_STATIC_ASSERT_VECTOR_ONLY(TYPE) \ diff --git a/xs/src/eigen/Eigen/src/Eigenvalues/GeneralizedEigenSolver.h b/xs/src/eigen/Eigen/src/Eigenvalues/GeneralizedEigenSolver.h index 36a91dffc..87d789b3f 100644 --- a/xs/src/eigen/Eigen/src/Eigenvalues/GeneralizedEigenSolver.h +++ b/xs/src/eigen/Eigen/src/Eigenvalues/GeneralizedEigenSolver.h @@ -311,7 +311,6 @@ GeneralizedEigenSolver::compute(const MatrixType& A, const MatrixTyp // Aliases: Map v(reinterpret_cast(m_tmp.data()), size); ComplexVectorType &cv = m_tmp; - const MatrixType &mZ = m_realQZ.matrixZ(); const MatrixType &mS = m_realQZ.matrixS(); const MatrixType &mT = m_realQZ.matrixT(); @@ -351,7 +350,7 @@ GeneralizedEigenSolver::compute(const MatrixType& A, const MatrixTyp } } } - m_eivec.col(i).real().noalias() = mZ.transpose() * v; + m_eivec.col(i).real().noalias() = m_realQZ.matrixZ().transpose() * v; m_eivec.col(i).real().normalize(); m_eivec.col(i).imag().setConstant(0); } @@ -400,7 +399,7 @@ GeneralizedEigenSolver::compute(const MatrixType& A, const MatrixTyp / (alpha*mT.coeffRef(j,j) - static_cast(beta*mS.coeffRef(j,j))); } } - m_eivec.col(i+1).noalias() = (mZ.transpose() * cv); + m_eivec.col(i+1).noalias() = (m_realQZ.matrixZ().transpose() * cv); m_eivec.col(i+1).normalize(); m_eivec.col(i) = m_eivec.col(i+1).conjugate(); } diff --git a/xs/src/eigen/Eigen/src/Eigenvalues/RealSchur.h b/xs/src/eigen/Eigen/src/Eigenvalues/RealSchur.h index f5c86041d..17ea903f5 100644 --- a/xs/src/eigen/Eigen/src/Eigenvalues/RealSchur.h +++ b/xs/src/eigen/Eigen/src/Eigenvalues/RealSchur.h @@ -303,7 +303,7 @@ RealSchur& RealSchur::computeFromHessenberg(const HessMa Scalar exshift(0); // sum of exceptional shifts Scalar norm = computeNormOfT(); - if(norm!=0) + if(norm!=Scalar(0)) { while (iu >= 0) { @@ -327,7 +327,7 @@ RealSchur& RealSchur::computeFromHessenberg(const HessMa else // No convergence yet { // The firstHouseholderVector vector has to be initialized to something to get rid of a silly GCC warning (-O1 -Wall -DNDEBUG ) - Vector3s firstHouseholderVector(0,0,0), shiftInfo; + Vector3s firstHouseholderVector = Vector3s::Zero(), shiftInfo; computeShift(iu, iter, exshift, shiftInfo); iter = iter + 1; totalIter = totalIter + 1; diff --git a/xs/src/eigen/Eigen/src/Eigenvalues/SelfAdjointEigenSolver_LAPACKE.h b/xs/src/eigen/Eigen/src/Eigenvalues/SelfAdjointEigenSolver_LAPACKE.h index 3891cf883..b0c947dc0 100644 --- a/xs/src/eigen/Eigen/src/Eigenvalues/SelfAdjointEigenSolver_LAPACKE.h +++ b/xs/src/eigen/Eigen/src/Eigenvalues/SelfAdjointEigenSolver_LAPACKE.h @@ -37,7 +37,7 @@ namespace Eigen { /** \internal Specialization for the data types supported by LAPACKe */ -#define EIGEN_LAPACKE_EIG_SELFADJ(EIGTYPE, LAPACKE_TYPE, LAPACKE_RTYPE, LAPACKE_NAME, EIGCOLROW, LAPACKE_COLROW ) \ +#define EIGEN_LAPACKE_EIG_SELFADJ_2(EIGTYPE, LAPACKE_TYPE, LAPACKE_RTYPE, LAPACKE_NAME, EIGCOLROW ) \ template<> template inline \ SelfAdjointEigenSolver >& \ SelfAdjointEigenSolver >::compute(const EigenBase& matrix, int options) \ @@ -47,7 +47,7 @@ SelfAdjointEigenSolver >::compute(c && (options&EigVecMask)!=EigVecMask \ && "invalid option parameter"); \ bool computeEigenvectors = (options&ComputeEigenvectors)==ComputeEigenvectors; \ - lapack_int n = internal::convert_index(matrix.cols()), lda, matrix_order, info; \ + lapack_int n = internal::convert_index(matrix.cols()), lda, info; \ m_eivalues.resize(n,1); \ m_subdiag.resize(n-1); \ m_eivec = matrix; \ @@ -63,27 +63,24 @@ SelfAdjointEigenSolver >::compute(c } \ \ lda = internal::convert_index(m_eivec.outerStride()); \ - matrix_order=LAPACKE_COLROW; \ char jobz, uplo='L'/*, range='A'*/; \ jobz = computeEigenvectors ? 'V' : 'N'; \ \ - info = LAPACKE_##LAPACKE_NAME( matrix_order, jobz, uplo, n, (LAPACKE_TYPE*)m_eivec.data(), lda, (LAPACKE_RTYPE*)m_eivalues.data() ); \ + info = LAPACKE_##LAPACKE_NAME( LAPACK_COL_MAJOR, jobz, uplo, n, (LAPACKE_TYPE*)m_eivec.data(), lda, (LAPACKE_RTYPE*)m_eivalues.data() ); \ m_info = (info==0) ? Success : NoConvergence; \ m_isInitialized = true; \ m_eigenvectorsOk = computeEigenvectors; \ return *this; \ } +#define EIGEN_LAPACKE_EIG_SELFADJ(EIGTYPE, LAPACKE_TYPE, LAPACKE_RTYPE, LAPACKE_NAME ) \ + EIGEN_LAPACKE_EIG_SELFADJ_2(EIGTYPE, LAPACKE_TYPE, LAPACKE_RTYPE, LAPACKE_NAME, ColMajor ) \ + EIGEN_LAPACKE_EIG_SELFADJ_2(EIGTYPE, LAPACKE_TYPE, LAPACKE_RTYPE, LAPACKE_NAME, RowMajor ) -EIGEN_LAPACKE_EIG_SELFADJ(double, double, double, dsyev, ColMajor, LAPACK_COL_MAJOR) -EIGEN_LAPACKE_EIG_SELFADJ(float, float, float, ssyev, ColMajor, LAPACK_COL_MAJOR) -EIGEN_LAPACKE_EIG_SELFADJ(dcomplex, lapack_complex_double, double, zheev, ColMajor, LAPACK_COL_MAJOR) -EIGEN_LAPACKE_EIG_SELFADJ(scomplex, lapack_complex_float, float, cheev, ColMajor, LAPACK_COL_MAJOR) - -EIGEN_LAPACKE_EIG_SELFADJ(double, double, double, dsyev, RowMajor, LAPACK_ROW_MAJOR) -EIGEN_LAPACKE_EIG_SELFADJ(float, float, float, ssyev, RowMajor, LAPACK_ROW_MAJOR) -EIGEN_LAPACKE_EIG_SELFADJ(dcomplex, lapack_complex_double, double, zheev, RowMajor, LAPACK_ROW_MAJOR) -EIGEN_LAPACKE_EIG_SELFADJ(scomplex, lapack_complex_float, float, cheev, RowMajor, LAPACK_ROW_MAJOR) +EIGEN_LAPACKE_EIG_SELFADJ(double, double, double, dsyev) +EIGEN_LAPACKE_EIG_SELFADJ(float, float, float, ssyev) +EIGEN_LAPACKE_EIG_SELFADJ(dcomplex, lapack_complex_double, double, zheev) +EIGEN_LAPACKE_EIG_SELFADJ(scomplex, lapack_complex_float, float, cheev) } // end namespace Eigen diff --git a/xs/src/eigen/Eigen/src/Geometry/AngleAxis.h b/xs/src/eigen/Eigen/src/Geometry/AngleAxis.h index 0af3c1b08..83ee1be46 100644 --- a/xs/src/eigen/Eigen/src/Geometry/AngleAxis.h +++ b/xs/src/eigen/Eigen/src/Geometry/AngleAxis.h @@ -178,7 +178,7 @@ EIGEN_DEVICE_FUNC AngleAxis& AngleAxis::operator=(const Quaterni if (n != Scalar(0)) { m_angle = Scalar(2)*atan2(n, abs(q.w())); - if(q.w() < 0) + if(q.w() < Scalar(0)) n = -n; m_axis = q.vec() / n; } diff --git a/xs/src/eigen/Eigen/src/Geometry/Quaternion.h b/xs/src/eigen/Eigen/src/Geometry/Quaternion.h index f6ef1bcf6..c3fd8c3e0 100644 --- a/xs/src/eigen/Eigen/src/Geometry/Quaternion.h +++ b/xs/src/eigen/Eigen/src/Geometry/Quaternion.h @@ -43,6 +43,11 @@ class QuaternionBase : public RotationBase typedef typename internal::traits::Scalar Scalar; typedef typename NumTraits::Real RealScalar; typedef typename internal::traits::Coefficients Coefficients; + typedef typename Coefficients::CoeffReturnType CoeffReturnType; + typedef typename internal::conditional::Flags&LvalueBit), + Scalar&, CoeffReturnType>::type NonConstCoeffReturnType; + + enum { Flags = Eigen::internal::traits::Flags }; @@ -58,22 +63,22 @@ class QuaternionBase : public RotationBase /** \returns the \c x coefficient */ - EIGEN_DEVICE_FUNC inline Scalar x() const { return this->derived().coeffs().coeff(0); } + EIGEN_DEVICE_FUNC inline CoeffReturnType x() const { return this->derived().coeffs().coeff(0); } /** \returns the \c y coefficient */ - EIGEN_DEVICE_FUNC inline Scalar y() const { return this->derived().coeffs().coeff(1); } + EIGEN_DEVICE_FUNC inline CoeffReturnType y() const { return this->derived().coeffs().coeff(1); } /** \returns the \c z coefficient */ - EIGEN_DEVICE_FUNC inline Scalar z() const { return this->derived().coeffs().coeff(2); } + EIGEN_DEVICE_FUNC inline CoeffReturnType z() const { return this->derived().coeffs().coeff(2); } /** \returns the \c w coefficient */ - EIGEN_DEVICE_FUNC inline Scalar w() const { return this->derived().coeffs().coeff(3); } + EIGEN_DEVICE_FUNC inline CoeffReturnType w() const { return this->derived().coeffs().coeff(3); } - /** \returns a reference to the \c x coefficient */ - EIGEN_DEVICE_FUNC inline Scalar& x() { return this->derived().coeffs().coeffRef(0); } - /** \returns a reference to the \c y coefficient */ - EIGEN_DEVICE_FUNC inline Scalar& y() { return this->derived().coeffs().coeffRef(1); } - /** \returns a reference to the \c z coefficient */ - EIGEN_DEVICE_FUNC inline Scalar& z() { return this->derived().coeffs().coeffRef(2); } - /** \returns a reference to the \c w coefficient */ - EIGEN_DEVICE_FUNC inline Scalar& w() { return this->derived().coeffs().coeffRef(3); } + /** \returns a reference to the \c x coefficient (if Derived is a non-const lvalue) */ + EIGEN_DEVICE_FUNC inline NonConstCoeffReturnType x() { return this->derived().coeffs().x(); } + /** \returns a reference to the \c y coefficient (if Derived is a non-const lvalue) */ + EIGEN_DEVICE_FUNC inline NonConstCoeffReturnType y() { return this->derived().coeffs().y(); } + /** \returns a reference to the \c z coefficient (if Derived is a non-const lvalue) */ + EIGEN_DEVICE_FUNC inline NonConstCoeffReturnType z() { return this->derived().coeffs().z(); } + /** \returns a reference to the \c w coefficient (if Derived is a non-const lvalue) */ + EIGEN_DEVICE_FUNC inline NonConstCoeffReturnType w() { return this->derived().coeffs().w(); } /** \returns a read-only vector expression of the imaginary part (x,y,z) */ EIGEN_DEVICE_FUNC inline const VectorBlock vec() const { return coeffs().template head<3>(); } @@ -423,7 +428,7 @@ typedef Map, Aligned> QuaternionMapAlignedd; // Generic Quaternion * Quaternion product // This product can be specialized for a given architecture via the Arch template argument. namespace internal { -template struct quat_product +template struct quat_product { EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Quaternion run(const QuaternionBase& a, const QuaternionBase& b){ return Quaternion @@ -446,8 +451,7 @@ QuaternionBase::operator* (const QuaternionBase& other) c EIGEN_STATIC_ASSERT((internal::is_same::value), YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) return internal::quat_product::Scalar, - EIGEN_PLAIN_ENUM_MIN(internal::traits::Alignment, internal::traits::Alignment)>::run(*this, other); + typename internal::traits::Scalar>::run(*this, other); } /** \sa operator*(Quaternion) */ @@ -672,7 +676,7 @@ EIGEN_DEVICE_FUNC inline Quaternion::Scalar> // Generic conjugate of a Quaternion namespace internal { -template struct quat_conj +template struct quat_conj { EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Quaternion run(const QuaternionBase& q){ return Quaternion(q.w(),-q.x(),-q.y(),-q.z()); @@ -691,8 +695,7 @@ EIGEN_DEVICE_FUNC inline Quaternion::Scalar> QuaternionBase::conjugate() const { return internal::quat_conj::Scalar, - internal::traits::Alignment>::run(*this); + typename internal::traits::Scalar>::run(*this); } diff --git a/xs/src/eigen/Eigen/src/Geometry/arch/Geometry_SSE.h b/xs/src/eigen/Eigen/src/Geometry/arch/Geometry_SSE.h index 1a86ff837..f68cab583 100644 --- a/xs/src/eigen/Eigen/src/Geometry/arch/Geometry_SSE.h +++ b/xs/src/eigen/Eigen/src/Geometry/arch/Geometry_SSE.h @@ -16,17 +16,23 @@ namespace Eigen { namespace internal { template -struct quat_product +struct quat_product { + enum { + AAlignment = traits::Alignment, + BAlignment = traits::Alignment, + ResAlignment = traits >::Alignment + }; static inline Quaternion run(const QuaternionBase& _a, const QuaternionBase& _b) { Quaternion res; const __m128 mask = _mm_setr_ps(0.f,0.f,0.f,-0.f); - __m128 a = _a.coeffs().template packet(0); - __m128 b = _b.coeffs().template packet(0); + __m128 a = _a.coeffs().template packet(0); + __m128 b = _b.coeffs().template packet(0); __m128 s1 = _mm_mul_ps(vec4f_swizzle1(a,1,2,0,2),vec4f_swizzle1(b,2,0,1,2)); __m128 s2 = _mm_mul_ps(vec4f_swizzle1(a,3,3,3,1),vec4f_swizzle1(b,0,1,2,1)); - pstore(&res.x(), + pstoret( + &res.x(), _mm_add_ps(_mm_sub_ps(_mm_mul_ps(a,vec4f_swizzle1(b,3,3,3,3)), _mm_mul_ps(vec4f_swizzle1(a,2,0,1,0), vec4f_swizzle1(b,1,2,0,0))), @@ -36,14 +42,17 @@ struct quat_product } }; -template -struct quat_conj +template +struct quat_conj { + enum { + ResAlignment = traits >::Alignment + }; static inline Quaternion run(const QuaternionBase& q) { Quaternion res; const __m128 mask = _mm_setr_ps(-0.f,-0.f,-0.f,0.f); - pstore(&res.x(), _mm_xor_ps(mask, q.coeffs().template packet(0))); + pstoret(&res.x(), _mm_xor_ps(mask, q.coeffs().template packet::Alignment>(0))); return res; } }; @@ -52,6 +61,9 @@ struct quat_conj template struct cross3_impl { + enum { + ResAlignment = traits::type>::Alignment + }; static inline typename plain_matrix_type::type run(const VectorLhs& lhs, const VectorRhs& rhs) { @@ -60,7 +72,7 @@ struct cross3_impl __m128 mul1=_mm_mul_ps(vec4f_swizzle1(a,1,2,0,3),vec4f_swizzle1(b,2,0,1,3)); __m128 mul2=_mm_mul_ps(vec4f_swizzle1(a,2,0,1,3),vec4f_swizzle1(b,1,2,0,3)); typename plain_matrix_type::type res; - pstore(&res.x(),_mm_sub_ps(mul1,mul2)); + pstoret(&res.x(),_mm_sub_ps(mul1,mul2)); return res; } }; @@ -68,9 +80,14 @@ struct cross3_impl -template -struct quat_product +template +struct quat_product { + enum { + BAlignment = traits::Alignment, + ResAlignment = traits >::Alignment + }; + static inline Quaternion run(const QuaternionBase& _a, const QuaternionBase& _b) { const Packet2d mask = _mm_castsi128_pd(_mm_set_epi32(0x0,0x0,0x80000000,0x0)); @@ -78,8 +95,8 @@ struct quat_product Quaternion res; const double* a = _a.coeffs().data(); - Packet2d b_xy = _b.coeffs().template packet(0); - Packet2d b_zw = _b.coeffs().template packet(2); + Packet2d b_xy = _b.coeffs().template packet(0); + Packet2d b_zw = _b.coeffs().template packet(2); Packet2d a_xx = pset1(a[0]); Packet2d a_yy = pset1(a[1]); Packet2d a_zz = pset1(a[2]); @@ -97,9 +114,9 @@ struct quat_product t2 = psub(pmul(a_zz, b_xy), pmul(a_xx, b_zw)); #ifdef EIGEN_VECTORIZE_SSE3 EIGEN_UNUSED_VARIABLE(mask) - pstore(&res.x(), _mm_addsub_pd(t1, preverse(t2))); + pstoret(&res.x(), _mm_addsub_pd(t1, preverse(t2))); #else - pstore(&res.x(), padd(t1, pxor(mask,preverse(t2)))); + pstoret(&res.x(), padd(t1, pxor(mask,preverse(t2)))); #endif /* @@ -111,25 +128,28 @@ struct quat_product t2 = padd(pmul(a_zz, b_zw), pmul(a_xx, b_xy)); #ifdef EIGEN_VECTORIZE_SSE3 EIGEN_UNUSED_VARIABLE(mask) - pstore(&res.z(), preverse(_mm_addsub_pd(preverse(t1), t2))); + pstoret(&res.z(), preverse(_mm_addsub_pd(preverse(t1), t2))); #else - pstore(&res.z(), psub(t1, pxor(mask,preverse(t2)))); + pstoret(&res.z(), psub(t1, pxor(mask,preverse(t2)))); #endif return res; } }; -template -struct quat_conj +template +struct quat_conj { + enum { + ResAlignment = traits >::Alignment + }; static inline Quaternion run(const QuaternionBase& q) { Quaternion res; const __m128d mask0 = _mm_setr_pd(-0.,-0.); const __m128d mask2 = _mm_setr_pd(-0.,0.); - pstore(&res.x(), _mm_xor_pd(mask0, q.coeffs().template packet(0))); - pstore(&res.z(), _mm_xor_pd(mask2, q.coeffs().template packet(2))); + pstoret(&res.x(), _mm_xor_pd(mask0, q.coeffs().template packet::Alignment>(0))); + pstoret(&res.z(), _mm_xor_pd(mask2, q.coeffs().template packet::Alignment>(2))); return res; } }; diff --git a/xs/src/eigen/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h b/xs/src/eigen/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h index 358444aff..f66c846ef 100644 --- a/xs/src/eigen/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h +++ b/xs/src/eigen/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h @@ -152,13 +152,28 @@ class LeastSquareDiagonalPreconditioner : public DiagonalPreconditioner<_Scalar> { // Compute the inverse squared-norm of each column of mat m_invdiag.resize(mat.cols()); - for(Index j=0; j0) - m_invdiag(j) = RealScalar(1)/sum; - else - m_invdiag(j) = RealScalar(1); + m_invdiag.setZero(); + for(Index j=0; jRealScalar(0)) + m_invdiag(j) = RealScalar(1)/numext::real(m_invdiag(j)); + } + else + { + for(Index j=0; jRealScalar(0)) + m_invdiag(j) = RealScalar(1)/sum; + else + m_invdiag(j) = RealScalar(1); + } } Base::m_isInitialized = true; return *this; diff --git a/xs/src/eigen/Eigen/src/Jacobi/Jacobi.h b/xs/src/eigen/Eigen/src/Jacobi/Jacobi.h index d25af8e90..437e666a3 100644 --- a/xs/src/eigen/Eigen/src/Jacobi/Jacobi.h +++ b/xs/src/eigen/Eigen/src/Jacobi/Jacobi.h @@ -298,12 +298,144 @@ inline void MatrixBase::applyOnTheRight(Index p, Index q, const JacobiR } namespace internal { + +template +struct apply_rotation_in_the_plane_selector +{ + static inline void run(Scalar *x, Index incrx, Scalar *y, Index incry, Index size, OtherScalar c, OtherScalar s) + { + for(Index i=0; i +struct apply_rotation_in_the_plane_selector +{ + static inline void run(Scalar *x, Index incrx, Scalar *y, Index incry, Index size, OtherScalar c, OtherScalar s) + { + enum { + PacketSize = packet_traits::size, + OtherPacketSize = packet_traits::size + }; + typedef typename packet_traits::type Packet; + typedef typename packet_traits::type OtherPacket; + + /*** dynamic-size vectorized paths ***/ + if(SizeAtCompileTime == Dynamic && ((incrx==1 && incry==1) || PacketSize == 1)) + { + // both vectors are sequentially stored in memory => vectorization + enum { Peeling = 2 }; + + Index alignedStart = internal::first_default_aligned(y, size); + Index alignedEnd = alignedStart + ((size-alignedStart)/PacketSize)*PacketSize; + + const OtherPacket pc = pset1(c); + const OtherPacket ps = pset1(s); + conj_helper::IsComplex,false> pcj; + conj_helper pm; + + for(Index i=0; i(px); + Packet yi = pload(py); + pstore(px, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi))); + pstore(py, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi))); + px += PacketSize; + py += PacketSize; + } + } + else + { + Index peelingEnd = alignedStart + ((size-alignedStart)/(Peeling*PacketSize))*(Peeling*PacketSize); + for(Index i=alignedStart; i(px); + Packet xi1 = ploadu(px+PacketSize); + Packet yi = pload (py); + Packet yi1 = pload (py+PacketSize); + pstoreu(px, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi))); + pstoreu(px+PacketSize, padd(pm.pmul(pc,xi1),pcj.pmul(ps,yi1))); + pstore (py, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi))); + pstore (py+PacketSize, psub(pcj.pmul(pc,yi1),pm.pmul(ps,xi1))); + px += Peeling*PacketSize; + py += Peeling*PacketSize; + } + if(alignedEnd!=peelingEnd) + { + Packet xi = ploadu(x+peelingEnd); + Packet yi = pload (y+peelingEnd); + pstoreu(x+peelingEnd, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi))); + pstore (y+peelingEnd, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi))); + } + } + + for(Index i=alignedEnd; i0) // FIXME should be compared to the required alignment + { + const OtherPacket pc = pset1(c); + const OtherPacket ps = pset1(s); + conj_helper::IsComplex,false> pcj; + conj_helper pm; + Scalar* EIGEN_RESTRICT px = x; + Scalar* EIGEN_RESTRICT py = y; + for(Index i=0; i(px); + Packet yi = pload(py); + pstore(px, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi))); + pstore(py, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi))); + px += PacketSize; + py += PacketSize; + } + } + + /*** non-vectorized path ***/ + else + { + apply_rotation_in_the_plane_selector::run(x,incrx,y,incry,size,c,s); + } + } +}; + template void /*EIGEN_DONT_INLINE*/ apply_rotation_in_the_plane(DenseBase& xpr_x, DenseBase& xpr_y, const JacobiRotation& j) { typedef typename VectorX::Scalar Scalar; - enum { PacketSize = packet_traits::size }; - typedef typename packet_traits::type Packet; + const bool Vectorizable = (VectorX::Flags & VectorY::Flags & PacketAccessBit) + && (int(packet_traits::size) == int(packet_traits::size)); + eigen_assert(xpr_x.size() == xpr_y.size()); Index size = xpr_x.size(); Index incrx = xpr_x.derived().innerStride(); @@ -317,113 +449,11 @@ void /*EIGEN_DONT_INLINE*/ apply_rotation_in_the_plane(DenseBase& xpr_x if (c==OtherScalar(1) && s==OtherScalar(0)) return; - /*** dynamic-size vectorized paths ***/ - - if(VectorX::SizeAtCompileTime == Dynamic && - (VectorX::Flags & VectorY::Flags & PacketAccessBit) && - ((incrx==1 && incry==1) || PacketSize == 1)) - { - // both vectors are sequentially stored in memory => vectorization - enum { Peeling = 2 }; - - Index alignedStart = internal::first_default_aligned(y, size); - Index alignedEnd = alignedStart + ((size-alignedStart)/PacketSize)*PacketSize; - - const Packet pc = pset1(c); - const Packet ps = pset1(s); - conj_helper::IsComplex,false> pcj; - - for(Index i=0; i(px); - Packet yi = pload(py); - pstore(px, padd(pmul(pc,xi),pcj.pmul(ps,yi))); - pstore(py, psub(pcj.pmul(pc,yi),pmul(ps,xi))); - px += PacketSize; - py += PacketSize; - } - } - else - { - Index peelingEnd = alignedStart + ((size-alignedStart)/(Peeling*PacketSize))*(Peeling*PacketSize); - for(Index i=alignedStart; i(px); - Packet xi1 = ploadu(px+PacketSize); - Packet yi = pload (py); - Packet yi1 = pload (py+PacketSize); - pstoreu(px, padd(pmul(pc,xi),pcj.pmul(ps,yi))); - pstoreu(px+PacketSize, padd(pmul(pc,xi1),pcj.pmul(ps,yi1))); - pstore (py, psub(pcj.pmul(pc,yi),pmul(ps,xi))); - pstore (py+PacketSize, psub(pcj.pmul(pc,yi1),pmul(ps,xi1))); - px += Peeling*PacketSize; - py += Peeling*PacketSize; - } - if(alignedEnd!=peelingEnd) - { - Packet xi = ploadu(x+peelingEnd); - Packet yi = pload (y+peelingEnd); - pstoreu(x+peelingEnd, padd(pmul(pc,xi),pcj.pmul(ps,yi))); - pstore (y+peelingEnd, psub(pcj.pmul(pc,yi),pmul(ps,xi))); - } - } - - for(Index i=alignedEnd; i::Alignment, evaluator::Alignment)>0)) // FIXME should be compared to the required alignment - { - const Packet pc = pset1(c); - const Packet ps = pset1(s); - conj_helper::IsComplex,false> pcj; - Scalar* EIGEN_RESTRICT px = x; - Scalar* EIGEN_RESTRICT py = y; - for(Index i=0; i(px); - Packet yi = pload(py); - pstore(px, padd(pmul(pc,xi),pcj.pmul(ps,yi))); - pstore(py, psub(pcj.pmul(pc,yi),pmul(ps,xi))); - px += PacketSize; - py += PacketSize; - } - } - - /*** non-vectorized path ***/ - else - { - for(Index i=0; i::Alignment, evaluator::Alignment), + Vectorizable>::run(x,incrx,y,incry,size,c,s); } } // end namespace internal diff --git a/xs/src/eigen/Eigen/src/LU/InverseImpl.h b/xs/src/eigen/Eigen/src/LU/InverseImpl.h index 018f99b58..f49f23360 100644 --- a/xs/src/eigen/Eigen/src/LU/InverseImpl.h +++ b/xs/src/eigen/Eigen/src/LU/InverseImpl.h @@ -404,7 +404,7 @@ inline void MatrixBase::computeInverseWithCheck( const RealScalar& absDeterminantThreshold ) const { - RealScalar determinant; + Scalar determinant; // i'd love to put some static assertions there, but SFINAE means that they have no effect... eigen_assert(rows() == cols()); computeInverseAndDetWithCheck(inverse,determinant,invertible,absDeterminantThreshold); diff --git a/xs/src/eigen/Eigen/src/OrderingMethods/Eigen_Colamd.h b/xs/src/eigen/Eigen/src/OrderingMethods/Eigen_Colamd.h index 933cd564b..da85b4d6e 100644 --- a/xs/src/eigen/Eigen/src/OrderingMethods/Eigen_Colamd.h +++ b/xs/src/eigen/Eigen/src/OrderingMethods/Eigen_Colamd.h @@ -1004,7 +1004,7 @@ static IndexType find_ordering /* return the number of garbage collections */ COLAMD_ASSERT (head [min_score] >= COLAMD_EMPTY) ; /* get pivot column from head of minimum degree list */ - while (head [min_score] == COLAMD_EMPTY && min_score < n_col) + while (min_score < n_col && head [min_score] == COLAMD_EMPTY) { min_score++ ; } diff --git a/xs/src/eigen/Eigen/src/PaStiXSupport/PaStiXSupport.h b/xs/src/eigen/Eigen/src/PaStiXSupport/PaStiXSupport.h index d2ebfd7bb..160d8a523 100644 --- a/xs/src/eigen/Eigen/src/PaStiXSupport/PaStiXSupport.h +++ b/xs/src/eigen/Eigen/src/PaStiXSupport/PaStiXSupport.h @@ -64,28 +64,28 @@ namespace internal typedef typename _MatrixType::StorageIndex StorageIndex; }; - void eigen_pastix(pastix_data_t **pastix_data, int pastix_comm, int n, int *ptr, int *idx, float *vals, int *perm, int * invp, float *x, int nbrhs, int *iparm, double *dparm) + inline void eigen_pastix(pastix_data_t **pastix_data, int pastix_comm, int n, int *ptr, int *idx, float *vals, int *perm, int * invp, float *x, int nbrhs, int *iparm, double *dparm) { if (n == 0) { ptr = NULL; idx = NULL; vals = NULL; } if (nbrhs == 0) {x = NULL; nbrhs=1;} s_pastix(pastix_data, pastix_comm, n, ptr, idx, vals, perm, invp, x, nbrhs, iparm, dparm); } - void eigen_pastix(pastix_data_t **pastix_data, int pastix_comm, int n, int *ptr, int *idx, double *vals, int *perm, int * invp, double *x, int nbrhs, int *iparm, double *dparm) + inline void eigen_pastix(pastix_data_t **pastix_data, int pastix_comm, int n, int *ptr, int *idx, double *vals, int *perm, int * invp, double *x, int nbrhs, int *iparm, double *dparm) { if (n == 0) { ptr = NULL; idx = NULL; vals = NULL; } if (nbrhs == 0) {x = NULL; nbrhs=1;} d_pastix(pastix_data, pastix_comm, n, ptr, idx, vals, perm, invp, x, nbrhs, iparm, dparm); } - void eigen_pastix(pastix_data_t **pastix_data, int pastix_comm, int n, int *ptr, int *idx, std::complex *vals, int *perm, int * invp, std::complex *x, int nbrhs, int *iparm, double *dparm) + inline void eigen_pastix(pastix_data_t **pastix_data, int pastix_comm, int n, int *ptr, int *idx, std::complex *vals, int *perm, int * invp, std::complex *x, int nbrhs, int *iparm, double *dparm) { if (n == 0) { ptr = NULL; idx = NULL; vals = NULL; } if (nbrhs == 0) {x = NULL; nbrhs=1;} c_pastix(pastix_data, pastix_comm, n, ptr, idx, reinterpret_cast(vals), perm, invp, reinterpret_cast(x), nbrhs, iparm, dparm); } - void eigen_pastix(pastix_data_t **pastix_data, int pastix_comm, int n, int *ptr, int *idx, std::complex *vals, int *perm, int * invp, std::complex *x, int nbrhs, int *iparm, double *dparm) + inline void eigen_pastix(pastix_data_t **pastix_data, int pastix_comm, int n, int *ptr, int *idx, std::complex *vals, int *perm, int * invp, std::complex *x, int nbrhs, int *iparm, double *dparm) { if (n == 0) { ptr = NULL; idx = NULL; vals = NULL; } if (nbrhs == 0) {x = NULL; nbrhs=1;} diff --git a/xs/src/eigen/Eigen/src/QR/ColPivHouseholderQR.h b/xs/src/eigen/Eigen/src/QR/ColPivHouseholderQR.h index 0e47c8332..a7b47d55d 100644 --- a/xs/src/eigen/Eigen/src/QR/ColPivHouseholderQR.h +++ b/xs/src/eigen/Eigen/src/QR/ColPivHouseholderQR.h @@ -506,8 +506,8 @@ void ColPivHouseholderQR::computeInPlace() m_colNormsUpdated.coeffRef(k) = m_colNormsDirect.coeffRef(k); } - RealScalar threshold_helper = numext::abs2(m_colNormsUpdated.maxCoeff() * NumTraits::epsilon()) / RealScalar(rows); - RealScalar norm_downdate_threshold = numext::sqrt(NumTraits::epsilon()); + RealScalar threshold_helper = numext::abs2(m_colNormsUpdated.maxCoeff() * NumTraits::epsilon()) / RealScalar(rows); + RealScalar norm_downdate_threshold = numext::sqrt(NumTraits::epsilon()); m_nonzero_pivots = size; // the generic case is that in which all pivots are nonzero (invertible case) m_maxpivot = RealScalar(0); @@ -553,12 +553,12 @@ void ColPivHouseholderQR::computeInPlace() // http://www.netlib.org/lapack/lawnspdf/lawn176.pdf // and used in LAPACK routines xGEQPF and xGEQP3. // See lines 278-297 in http://www.netlib.org/lapack/explore-html/dc/df4/sgeqpf_8f_source.html - if (m_colNormsUpdated.coeffRef(j) != 0) { + if (m_colNormsUpdated.coeffRef(j) != RealScalar(0)) { RealScalar temp = abs(m_qr.coeffRef(k, j)) / m_colNormsUpdated.coeffRef(j); temp = (RealScalar(1) + temp) * (RealScalar(1) - temp); - temp = temp < 0 ? 0 : temp; - RealScalar temp2 = temp * numext::abs2(m_colNormsUpdated.coeffRef(j) / - m_colNormsDirect.coeffRef(j)); + temp = temp < RealScalar(0) ? RealScalar(0) : temp; + RealScalar temp2 = temp * numext::abs2(m_colNormsUpdated.coeffRef(j) / + m_colNormsDirect.coeffRef(j)); if (temp2 <= norm_downdate_threshold) { // The updated norm has become too inaccurate so re-compute the column // norm directly. diff --git a/xs/src/eigen/Eigen/src/SVD/BDCSVD.h b/xs/src/eigen/Eigen/src/SVD/BDCSVD.h index 25fca6f4d..1134d66e7 100644 --- a/xs/src/eigen/Eigen/src/SVD/BDCSVD.h +++ b/xs/src/eigen/Eigen/src/SVD/BDCSVD.h @@ -11,7 +11,7 @@ // Copyright (C) 2013 Jean Ceccato // Copyright (C) 2013 Pierre Zoppitelli // Copyright (C) 2013 Jitse Niesen -// Copyright (C) 2014-2016 Gael Guennebaud +// Copyright (C) 2014-2017 Gael Guennebaud // // 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 @@ -77,6 +77,7 @@ public: typedef _MatrixType MatrixType; typedef typename MatrixType::Scalar Scalar; typedef typename NumTraits::Real RealScalar; + typedef typename NumTraits::Literal Literal; enum { RowsAtCompileTime = MatrixType::RowsAtCompileTime, ColsAtCompileTime = MatrixType::ColsAtCompileTime, @@ -259,7 +260,7 @@ BDCSVD& BDCSVD::compute(const MatrixType& matrix, unsign //**** step 0 - Copy the input matrix and apply scaling to reduce over/under-flows RealScalar scale = matrix.cwiseAbs().maxCoeff(); - if(scale==RealScalar(0)) scale = RealScalar(1); + if(scale==Literal(0)) scale = Literal(1); MatrixX copy; if (m_isTranspose) copy = matrix.adjoint()/scale; else copy = matrix/scale; @@ -351,13 +352,13 @@ void BDCSVD::structured_update(Block A, co Index k1=0, k2=0; for(Index j=0; j::divide (Index firstCol, Index lastCol, Index firstRowW, l = m_naiveU.row(1).segment(firstCol, k); f = m_naiveU.row(0).segment(firstCol + k + 1, n - k - 1); } - if (m_compV) m_naiveV(firstRowW+k, firstColW) = 1; + if (m_compV) m_naiveV(firstRowW+k, firstColW) = Literal(1); if (r0::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec ArrayRef col0 = m_computed.col(firstCol).segment(firstCol, n); m_workspace.head(n) = m_computed.block(firstCol, firstCol, n, n).diagonal(); ArrayRef diag = m_workspace.head(n); - diag(0) = 0; + diag(0) = Literal(0); // Allocate space for singular values and vectors singVals.resize(n); @@ -590,7 +591,7 @@ void BDCSVD::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec // but others are interleaved and we must ignore them at this stage. // To this end, let's compute a permutation skipping them: Index actual_n = n; - while(actual_n>1 && diag(actual_n-1)==0) --actual_n; + while(actual_n>1 && diag(actual_n-1)==Literal(0)) --actual_n; Index m = 0; // size of the deflated problem for(Index k=0;kconsiderZero) @@ -691,11 +692,13 @@ template typename BDCSVD::RealScalar BDCSVD::secularEq(RealScalar mu, const ArrayRef& col0, const ArrayRef& diag, const IndicesRef &perm, const ArrayRef& diagShifted, RealScalar shift) { Index m = perm.size(); - RealScalar res = 1; + RealScalar res = Literal(1); for(Index i=0; i::computeSingVals(const ArrayRef& col0, const ArrayRef& d { using std::abs; using std::swap; + using std::sqrt; Index n = col0.size(); Index actual_n = n; - while(actual_n>1 && col0(actual_n-1)==0) --actual_n; + // Note that here actual_n is computed based on col0(i)==0 instead of diag(i)==0 as above + // because 1) we have diag(i)==0 => col0(i)==0 and 2) if col0(i)==0, then diag(i) is already a singular value. + while(actual_n>1 && col0(actual_n-1)==Literal(0)) --actual_n; for (Index k = 0; k < n; ++k) { - if (col0(k) == 0 || actual_n==1) + if (col0(k) == Literal(0) || actual_n==1) { // if col0(k) == 0, then entry is deflated, so singular value is on diagonal // if actual_n==1, then the deflated problem is already diagonalized singVals(k) = k==0 ? col0(0) : diag(k); - mus(k) = 0; + mus(k) = Literal(0); shifts(k) = k==0 ? col0(0) : diag(k); continue; } @@ -731,15 +737,17 @@ void BDCSVD::computeSingVals(const ArrayRef& col0, const ArrayRef& d right = (diag(actual_n-1) + col0.matrix().norm()); else { - // Skip deflated singular values + // Skip deflated singular values, + // recall that at this stage we assume that z[j]!=0 and all entries for which z[j]==0 have been put aside. + // This should be equivalent to using perm[] Index l = k+1; - while(col0(l)==0) { ++l; eigen_internal_assert(l::computeSingVals(const ArrayRef& col0, const ArrayRef& d << " " << secularEq(0.8*(left+right), col0, diag, perm, diag, 0) << " " << secularEq(0.9*(left+right), col0, diag, perm, diag, 0) << "\n"; #endif - RealScalar shift = (k == actual_n-1 || fMid > 0) ? left : right; + RealScalar shift = (k == actual_n-1 || fMid > Literal(0)) ? left : right; // measure everything relative to shift Map diagShifted(m_workspace.data()+4*n, n); @@ -785,13 +793,13 @@ void BDCSVD::computeSingVals(const ArrayRef& col0, const ArrayRef& d // rational interpolation: fit a function of the form a / mu + b through the two previous // iterates and use its zero to compute the next iterate - bool useBisection = fPrev*fCur>0; - while (fCur!=0 && abs(muCur - muPrev) > 8 * NumTraits::epsilon() * numext::maxi(abs(muCur), abs(muPrev)) && abs(fCur - fPrev)>NumTraits::epsilon() && !useBisection) + bool useBisection = fPrev*fCur>Literal(0); + while (fCur!=Literal(0) && abs(muCur - muPrev) > Literal(8) * NumTraits::epsilon() * numext::maxi(abs(muCur), abs(muPrev)) && abs(fCur - fPrev)>NumTraits::epsilon() && !useBisection) { ++m_numIters; // Find a and b such that the function f(mu) = a / mu + b matches the current and previous samples. - RealScalar a = (fCur - fPrev) / (1/muCur - 1/muPrev); + RealScalar a = (fCur - fPrev) / (Literal(1)/muCur - Literal(1)/muPrev); RealScalar b = fCur - a / muCur; // And find mu such that f(mu)==0: RealScalar muZero = -a/b; @@ -803,8 +811,8 @@ void BDCSVD::computeSingVals(const ArrayRef& col0, const ArrayRef& d fCur = fZero; - if (shift == left && (muCur < 0 || muCur > right - left)) useBisection = true; - if (shift == right && (muCur < -(right - left) || muCur > 0)) useBisection = true; + if (shift == left && (muCur < Literal(0) || muCur > right - left)) useBisection = true; + if (shift == right && (muCur < -(right - left) || muCur > Literal(0))) useBisection = true; if (abs(fCur)>abs(fPrev)) useBisection = true; } @@ -817,15 +825,23 @@ void BDCSVD::computeSingVals(const ArrayRef& col0, const ArrayRef& d RealScalar leftShifted, rightShifted; if (shift == left) { - leftShifted = (std::numeric_limits::min)(); + // to avoid overflow, we must have mu > max(real_min, |z(k)|/sqrt(real_max)), + // the factor 2 is to be more conservative + leftShifted = numext::maxi( (std::numeric_limits::min)(), Literal(2) * abs(col0(k)) / sqrt((std::numeric_limits::max)()) ); + + // check that we did it right: + eigen_internal_assert( (numext::isfinite)( (col0(k)/leftShifted)*(col0(k)/(diag(k)+shift+leftShifted)) ) ); // I don't understand why the case k==0 would be special there: - // if (k == 0) rightShifted = right - left; else - rightShifted = (k==actual_n-1) ? right : ((right - left) * RealScalar(0.6)); // theoretically we can take 0.5, but let's be safe + // if (k == 0) rightShifted = right - left; else + rightShifted = (k==actual_n-1) ? right : ((right - left) * RealScalar(0.51)); // theoretically we can take 0.5, but let's be safe } else { - leftShifted = -(right - left) * RealScalar(0.6); - rightShifted = -(std::numeric_limits::min)(); + leftShifted = -(right - left) * RealScalar(0.51); + if(k+1( (std::numeric_limits::min)(), abs(col0(k+1)) / sqrt((std::numeric_limits::max)()) ); + else + rightShifted = -(std::numeric_limits::min)(); } RealScalar fLeft = secularEq(leftShifted, col0, diag, perm, diagShifted, shift); @@ -841,13 +857,13 @@ void BDCSVD::computeSingVals(const ArrayRef& col0, const ArrayRef& d std::cout << k << " : " << fLeft << " * " << fRight << " == " << fLeft * fRight << " ; " << left << " - " << right << " -> " << leftShifted << " " << rightShifted << " shift=" << shift << "\n"; } #endif - eigen_internal_assert(fLeft * fRight < 0); + eigen_internal_assert(fLeft * fRight < Literal(0)); - while (rightShifted - leftShifted > 2 * NumTraits::epsilon() * numext::maxi(abs(leftShifted), abs(rightShifted))) + while (rightShifted - leftShifted > Literal(2) * NumTraits::epsilon() * numext::maxi(abs(leftShifted), abs(rightShifted))) { - RealScalar midShifted = (leftShifted + rightShifted) / 2; + RealScalar midShifted = (leftShifted + rightShifted) / Literal(2); fMid = secularEq(midShifted, col0, diag, perm, diagShifted, shift); - if (fLeft * fMid < 0) + if (fLeft * fMid < Literal(0)) { rightShifted = midShifted; } @@ -858,7 +874,7 @@ void BDCSVD::computeSingVals(const ArrayRef& col0, const ArrayRef& d } } - muCur = (leftShifted + rightShifted) / 2; + muCur = (leftShifted + rightShifted) / Literal(2); } singVals[k] = shift + muCur; @@ -892,8 +908,8 @@ void BDCSVD::perturbCol0 // The offset permits to skip deflated entries while computing zhat for (Index k = 0; k < n; ++k) { - if (col0(k) == 0) // deflated - zhat(k) = 0; + if (col0(k) == Literal(0)) // deflated + zhat(k) = Literal(0); else { // see equation (3.6) @@ -918,7 +934,7 @@ void BDCSVD::perturbCol0 std::cout << "zhat(" << k << ") = sqrt( " << prod << ") ; " << (singVals(last) + dk) << " * " << mus(last) + shifts(last) << " - " << dk << "\n"; #endif RealScalar tmp = sqrt(prod); - zhat(k) = col0(k) > 0 ? tmp : -tmp; + zhat(k) = col0(k) > Literal(0) ? tmp : -tmp; } } } @@ -934,7 +950,7 @@ void BDCSVD::computeSingVecs for (Index k = 0; k < n; ++k) { - if (zhat(k) == 0) + if (zhat(k) == Literal(0)) { U.col(k) = VectorType::Unit(n+1, k); if (m_compV) V.col(k) = VectorType::Unit(n, k); @@ -947,7 +963,7 @@ void BDCSVD::computeSingVecs Index i = perm(l); U(i,k) = zhat(i)/(((diag(i) - shifts(k)) - mus(k)) )/( (diag(i) + singVals[k])); } - U(n,k) = 0; + U(n,k) = Literal(0); U.col(k).normalize(); if (m_compV) @@ -958,7 +974,7 @@ void BDCSVD::computeSingVecs Index i = perm(l); V(i,k) = diag(i) * zhat(i) / (((diag(i) - shifts(k)) - mus(k)) )/( (diag(i) + singVals[k])); } - V(0,k) = -1; + V(0,k) = Literal(-1); V.col(k).normalize(); } } @@ -979,15 +995,15 @@ void BDCSVD::deflation43(Index firstCol, Index shift, Index i, Index Index start = firstCol + shift; RealScalar c = m_computed(start, start); RealScalar s = m_computed(start+i, start); - RealScalar r = sqrt(numext::abs2(c) + numext::abs2(s)); - if (r == 0) + RealScalar r = numext::hypot(c,s); + if (r == Literal(0)) { - m_computed(start+i, start+i) = 0; + m_computed(start+i, start+i) = Literal(0); return; } m_computed(start,start) = r; - m_computed(start+i, start) = 0; - m_computed(start+i, start+i) = 0; + m_computed(start+i, start) = Literal(0); + m_computed(start+i, start+i) = Literal(0); JacobiRotation J(c/r,-s/r); if (m_compU) m_naiveU.middleRows(firstCol, size+1).applyOnTheRight(firstCol, firstCol+i, J); @@ -1020,7 +1036,7 @@ void BDCSVD::deflation44(Index firstColu , Index firstColm, Index fi << m_computed(firstColm + i+1, firstColm+i+1) << " " << m_computed(firstColm + i+2, firstColm+i+2) << "\n"; #endif - if (r==0) + if (r==Literal(0)) { m_computed(firstColm + i, firstColm + i) = m_computed(firstColm + j, firstColm + j); return; @@ -1029,7 +1045,7 @@ void BDCSVD::deflation44(Index firstColu , Index firstColm, Index fi s/=r; m_computed(firstColm + i, firstColm) = r; m_computed(firstColm + j, firstColm + j) = m_computed(firstColm + i, firstColm + i); - m_computed(firstColm + j, firstColm) = 0; + m_computed(firstColm + j, firstColm) = Literal(0); JacobiRotation J(c,-s); if (m_compU) m_naiveU.middleRows(firstColu, size+1).applyOnTheRight(firstColu + i, firstColu + j, J); @@ -1053,7 +1069,7 @@ void BDCSVD::deflation(Index firstCol, Index lastCol, Index k, Index const RealScalar considerZero = (std::numeric_limits::min)(); RealScalar maxDiag = diag.tail((std::max)(Index(1),length-1)).cwiseAbs().maxCoeff(); RealScalar epsilon_strict = numext::maxi(considerZero,NumTraits::epsilon() * maxDiag); - RealScalar epsilon_coarse = 8 * NumTraits::epsilon() * numext::maxi(col0.cwiseAbs().maxCoeff(), maxDiag); + RealScalar epsilon_coarse = Literal(8) * NumTraits::epsilon() * numext::maxi(col0.cwiseAbs().maxCoeff(), maxDiag); #ifdef EIGEN_BDCSVD_SANITY_CHECKS assert(m_naiveU.allFinite()); @@ -1081,7 +1097,7 @@ void BDCSVD::deflation(Index firstCol, Index lastCol, Index k, Index #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE std::cout << "deflation 4.2, set z(" << i << ") to zero because " << abs(col0(i)) << " < " << epsilon_strict << " (diag(" << i << ")=" << diag(i) << ")\n"; #endif - col0(i) = 0; + col0(i) = Literal(0); } //condition 4.3 diff --git a/xs/src/eigen/Eigen/src/SVD/JacobiSVD_LAPACKE.h b/xs/src/eigen/Eigen/src/SVD/JacobiSVD_LAPACKE.h index 50272154f..ff0516f61 100644 --- a/xs/src/eigen/Eigen/src/SVD/JacobiSVD_LAPACKE.h +++ b/xs/src/eigen/Eigen/src/SVD/JacobiSVD_LAPACKE.h @@ -61,9 +61,10 @@ JacobiSVD, ColPiv u = (LAPACKE_TYPE*)m_matrixU.data(); \ } else { ldu=1; u=&dummy; }\ MatrixType localV; \ - ldvt = (m_computeFullV) ? internal::convert_index(m_cols) : (m_computeThinV) ? internal::convert_index(m_diagSize) : 1; \ + lapack_int vt_rows = (m_computeFullV) ? internal::convert_index(m_cols) : (m_computeThinV) ? internal::convert_index(m_diagSize) : 1; \ if (computeV()) { \ - localV.resize(ldvt, m_cols); \ + localV.resize(vt_rows, m_cols); \ + ldvt = internal::convert_index(localV.outerStride()); \ vt = (LAPACKE_TYPE*)localV.data(); \ } else { ldvt=1; vt=&dummy; }\ Matrix superb; superb.resize(m_diagSize, 1); \ diff --git a/xs/src/eigen/Eigen/src/SVD/UpperBidiagonalization.h b/xs/src/eigen/Eigen/src/SVD/UpperBidiagonalization.h index 0b1460894..11ac847e1 100644 --- a/xs/src/eigen/Eigen/src/SVD/UpperBidiagonalization.h +++ b/xs/src/eigen/Eigen/src/SVD/UpperBidiagonalization.h @@ -159,6 +159,8 @@ void upperbidiagonalization_blocked_helper(MatrixType& A, traits::Flags & RowMajorBit> > Y) { typedef typename MatrixType::Scalar Scalar; + typedef typename MatrixType::RealScalar RealScalar; + typedef typename NumTraits::Literal Literal; enum { StorageOrder = traits::Flags & RowMajorBit }; typedef InnerStride ColInnerStride; typedef InnerStride RowInnerStride; @@ -263,7 +265,7 @@ void upperbidiagonalization_blocked_helper(MatrixType& A, SubMatType A10( A.block(bs,0, brows-bs,bs) ); SubMatType A01( A.block(0,bs, bs,bcols-bs) ); Scalar tmp = A01(bs-1,0); - A01(bs-1,0) = 1; + A01(bs-1,0) = Literal(1); A11.noalias() -= A10 * Y.topLeftCorner(bcols,bs).bottomRows(bcols-bs).adjoint(); A11.noalias() -= X.topLeftCorner(brows,bs).bottomRows(brows-bs) * A01; A01(bs-1,0) = tmp; diff --git a/xs/src/eigen/Eigen/src/SparseCore/AmbiVector.h b/xs/src/eigen/Eigen/src/SparseCore/AmbiVector.h index 8a5cc91f2..e0295f2af 100644 --- a/xs/src/eigen/Eigen/src/SparseCore/AmbiVector.h +++ b/xs/src/eigen/Eigen/src/SparseCore/AmbiVector.h @@ -94,7 +94,7 @@ class AmbiVector Index allocSize = m_allocatedElements * sizeof(ListEl); allocSize = (allocSize + sizeof(Scalar) - 1)/sizeof(Scalar); Scalar* newBuffer = new Scalar[allocSize]; - memcpy(newBuffer, m_buffer, copyElements * sizeof(ListEl)); + std::memcpy(newBuffer, m_buffer, copyElements * sizeof(ListEl)); delete[] m_buffer; m_buffer = newBuffer; } diff --git a/xs/src/eigen/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h b/xs/src/eigen/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h index 492eb0a29..9db119b67 100644 --- a/xs/src/eigen/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h +++ b/xs/src/eigen/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h @@ -17,7 +17,9 @@ namespace internal { template static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& rhs, ResultType& res, bool sortedInsertion = false) { - typedef typename remove_all::type::Scalar Scalar; + typedef typename remove_all::type::Scalar LhsScalar; + typedef typename remove_all::type::Scalar RhsScalar; + typedef typename remove_all::type::Scalar ResScalar; // make sure to call innerSize/outerSize since we fake the storage order. Index rows = lhs.innerSize(); @@ -25,7 +27,7 @@ static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& r eigen_assert(lhs.outerSize() == rhs.innerSize()); ei_declare_aligned_stack_constructed_variable(bool, mask, rows, 0); - ei_declare_aligned_stack_constructed_variable(Scalar, values, rows, 0); + ei_declare_aligned_stack_constructed_variable(ResScalar, values, rows, 0); ei_declare_aligned_stack_constructed_variable(Index, indices, rows, 0); std::memset(mask,0,sizeof(bool)*rows); @@ -51,12 +53,12 @@ static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& r Index nnz = 0; for (typename evaluator::InnerIterator rhsIt(rhsEval, j); rhsIt; ++rhsIt) { - Scalar y = rhsIt.value(); + RhsScalar y = rhsIt.value(); Index k = rhsIt.index(); for (typename evaluator::InnerIterator lhsIt(lhsEval, k); lhsIt; ++lhsIt) { Index i = lhsIt.index(); - Scalar x = lhsIt.value(); + LhsScalar x = lhsIt.value(); if(!mask[i]) { mask[i] = true; @@ -166,11 +168,12 @@ struct conservative_sparse_sparse_product_selector RowMajorMatrix; - RowMajorMatrix rhsRow = rhs; - RowMajorMatrix resRow(lhs.rows(), rhs.cols()); - internal::conservative_sparse_sparse_product_impl(rhsRow, lhs, resRow); - res = resRow; + typedef SparseMatrix RowMajorRhs; + typedef SparseMatrix RowMajorRes; + RowMajorRhs rhsRow = rhs; + RowMajorRes resRow(lhs.rows(), rhs.cols()); + internal::conservative_sparse_sparse_product_impl(rhsRow, lhs, resRow); + res = resRow; } }; @@ -179,10 +182,11 @@ struct conservative_sparse_sparse_product_selector RowMajorMatrix; - RowMajorMatrix lhsRow = lhs; - RowMajorMatrix resRow(lhs.rows(), rhs.cols()); - internal::conservative_sparse_sparse_product_impl(rhs, lhsRow, resRow); + typedef SparseMatrix RowMajorLhs; + typedef SparseMatrix RowMajorRes; + RowMajorLhs lhsRow = lhs; + RowMajorRes resRow(lhs.rows(), rhs.cols()); + internal::conservative_sparse_sparse_product_impl(rhs, lhsRow, resRow); res = resRow; } }; @@ -219,10 +223,11 @@ struct conservative_sparse_sparse_product_selector ColMajorMatrix; - ColMajorMatrix lhsCol = lhs; - ColMajorMatrix resCol(lhs.rows(), rhs.cols()); - internal::conservative_sparse_sparse_product_impl(lhsCol, rhs, resCol); + typedef SparseMatrix ColMajorLhs; + typedef SparseMatrix ColMajorRes; + ColMajorLhs lhsCol = lhs; + ColMajorRes resCol(lhs.rows(), rhs.cols()); + internal::conservative_sparse_sparse_product_impl(lhsCol, rhs, resCol); res = resCol; } }; @@ -232,10 +237,11 @@ struct conservative_sparse_sparse_product_selector ColMajorMatrix; - ColMajorMatrix rhsCol = rhs; - ColMajorMatrix resCol(lhs.rows(), rhs.cols()); - internal::conservative_sparse_sparse_product_impl(lhs, rhsCol, resCol); + typedef SparseMatrix ColMajorRhs; + typedef SparseMatrix ColMajorRes; + ColMajorRhs rhsCol = rhs; + ColMajorRes resCol(lhs.rows(), rhs.cols()); + internal::conservative_sparse_sparse_product_impl(lhs, rhsCol, resCol); res = resCol; } }; @@ -263,7 +269,8 @@ namespace internal { template static void sparse_sparse_to_dense_product_impl(const Lhs& lhs, const Rhs& rhs, ResultType& res) { - typedef typename remove_all::type::Scalar Scalar; + typedef typename remove_all::type::Scalar LhsScalar; + typedef typename remove_all::type::Scalar RhsScalar; Index cols = rhs.outerSize(); eigen_assert(lhs.outerSize() == rhs.innerSize()); @@ -274,12 +281,12 @@ static void sparse_sparse_to_dense_product_impl(const Lhs& lhs, const Rhs& rhs, { for (typename evaluator::InnerIterator rhsIt(rhsEval, j); rhsIt; ++rhsIt) { - Scalar y = rhsIt.value(); + RhsScalar y = rhsIt.value(); Index k = rhsIt.index(); for (typename evaluator::InnerIterator lhsIt(lhsEval, k); lhsIt; ++lhsIt) { Index i = lhsIt.index(); - Scalar x = lhsIt.value(); + LhsScalar x = lhsIt.value(); res.coeffRef(i,j) += x * y; } } @@ -310,9 +317,9 @@ struct sparse_sparse_to_dense_product_selector ColMajorMatrix; - ColMajorMatrix lhsCol(lhs); - internal::sparse_sparse_to_dense_product_impl(lhsCol, rhs, res); + typedef SparseMatrix ColMajorLhs; + ColMajorLhs lhsCol(lhs); + internal::sparse_sparse_to_dense_product_impl(lhsCol, rhs, res); } }; @@ -321,9 +328,9 @@ struct sparse_sparse_to_dense_product_selector ColMajorMatrix; - ColMajorMatrix rhsCol(rhs); - internal::sparse_sparse_to_dense_product_impl(lhs, rhsCol, res); + typedef SparseMatrix ColMajorRhs; + ColMajorRhs rhsCol(rhs); + internal::sparse_sparse_to_dense_product_impl(lhs, rhsCol, res); } }; diff --git a/xs/src/eigen/Eigen/src/SparseCore/SparseSelfAdjointView.h b/xs/src/eigen/Eigen/src/SparseCore/SparseSelfAdjointView.h index 9e39be738..65611b3d4 100644 --- a/xs/src/eigen/Eigen/src/SparseCore/SparseSelfAdjointView.h +++ b/xs/src/eigen/Eigen/src/SparseCore/SparseSelfAdjointView.h @@ -47,6 +47,7 @@ template class SparseSelfAdjointView enum { Mode = _Mode, + TransposeMode = ((Mode & Upper) ? Lower : 0) | ((Mode & Lower) ? Upper : 0), RowsAtCompileTime = internal::traits::RowsAtCompileTime, ColsAtCompileTime = internal::traits::ColsAtCompileTime }; @@ -310,7 +311,7 @@ inline void sparse_selfadjoint_time_dense_product(const SparseLhsType& lhs, cons while (i && i.index() dstT(dst); - internal::sparse_selfadjoint_time_dense_product(rhsNested.transpose(), lhsNested.transpose(), dstT, alpha); + internal::sparse_selfadjoint_time_dense_product(rhsNested.transpose(), lhsNested.transpose(), dstT, alpha); } }; diff --git a/xs/src/eigen/Eigen/src/SparseCore/SparseSparseProductWithPruning.h b/xs/src/eigen/Eigen/src/SparseCore/SparseSparseProductWithPruning.h index 21c419002..88820a48f 100644 --- a/xs/src/eigen/Eigen/src/SparseCore/SparseSparseProductWithPruning.h +++ b/xs/src/eigen/Eigen/src/SparseCore/SparseSparseProductWithPruning.h @@ -21,7 +21,8 @@ static void sparse_sparse_product_with_pruning_impl(const Lhs& lhs, const Rhs& r { // return sparse_sparse_product_with_pruning_impl2(lhs,rhs,res); - typedef typename remove_all::type::Scalar Scalar; + typedef typename remove_all::type::Scalar RhsScalar; + typedef typename remove_all::type::Scalar ResScalar; typedef typename remove_all::type::StorageIndex StorageIndex; // make sure to call innerSize/outerSize since we fake the storage order. @@ -31,7 +32,7 @@ static void sparse_sparse_product_with_pruning_impl(const Lhs& lhs, const Rhs& r eigen_assert(lhs.outerSize() == rhs.innerSize()); // allocate a temporary buffer - AmbiVector tempVector(rows); + AmbiVector tempVector(rows); // mimics a resizeByInnerOuter: if(ResultType::IsRowMajor) @@ -63,14 +64,14 @@ static void sparse_sparse_product_with_pruning_impl(const Lhs& lhs, const Rhs& r { // FIXME should be written like this: tmp += rhsIt.value() * lhs.col(rhsIt.index()) tempVector.restart(); - Scalar x = rhsIt.value(); + RhsScalar x = rhsIt.value(); for (typename evaluator::InnerIterator lhsIt(lhsEval, rhsIt.index()); lhsIt; ++lhsIt) { tempVector.coeffRef(lhsIt.index()) += lhsIt.value() * x; } } res.startVec(j); - for (typename AmbiVector::Iterator it(tempVector,tolerance); it; ++it) + for (typename AmbiVector::Iterator it(tempVector,tolerance); it; ++it) res.insertBackByOuterInner(j,it.index()) = it.value(); } res.finalize(); @@ -85,7 +86,6 @@ struct sparse_sparse_product_with_pruning_selector; template struct sparse_sparse_product_with_pruning_selector { - typedef typename traits::type>::Scalar Scalar; typedef typename ResultType::RealScalar RealScalar; static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res, const RealScalar& tolerance) @@ -129,8 +129,8 @@ struct sparse_sparse_product_with_pruning_selector ColMajorMatrixLhs; - typedef SparseMatrix ColMajorMatrixRhs; + typedef SparseMatrix ColMajorMatrixLhs; + typedef SparseMatrix ColMajorMatrixRhs; ColMajorMatrixLhs colLhs(lhs); ColMajorMatrixRhs colRhs(rhs); internal::sparse_sparse_product_with_pruning_impl(colLhs, colRhs, res, tolerance); @@ -149,7 +149,7 @@ struct sparse_sparse_product_with_pruning_selector RowMajorMatrixLhs; + typedef SparseMatrix RowMajorMatrixLhs; RowMajorMatrixLhs rowLhs(lhs); sparse_sparse_product_with_pruning_selector(rowLhs,rhs,res,tolerance); } @@ -161,7 +161,7 @@ struct sparse_sparse_product_with_pruning_selector RowMajorMatrixRhs; + typedef SparseMatrix RowMajorMatrixRhs; RowMajorMatrixRhs rowRhs(rhs); sparse_sparse_product_with_pruning_selector(lhs,rowRhs,res,tolerance); } @@ -173,7 +173,7 @@ struct sparse_sparse_product_with_pruning_selector ColMajorMatrixRhs; + typedef SparseMatrix ColMajorMatrixRhs; ColMajorMatrixRhs colRhs(rhs); internal::sparse_sparse_product_with_pruning_impl(lhs, colRhs, res, tolerance); } @@ -185,7 +185,7 @@ struct sparse_sparse_product_with_pruning_selector ColMajorMatrixLhs; + typedef SparseMatrix ColMajorMatrixLhs; ColMajorMatrixLhs colLhs(lhs); internal::sparse_sparse_product_with_pruning_impl(colLhs, rhs, res, tolerance); } diff --git a/xs/src/eigen/Eigen/src/SparseQR/SparseQR.h b/xs/src/eigen/Eigen/src/SparseQR/SparseQR.h index 2d4498b03..7409fcae9 100644 --- a/xs/src/eigen/Eigen/src/SparseQR/SparseQR.h +++ b/xs/src/eigen/Eigen/src/SparseQR/SparseQR.h @@ -52,7 +52,7 @@ namespace internal { * rank-revealing permutations. Use colsPermutation() to get it. * * Q is the orthogonal matrix represented as products of Householder reflectors. - * Use matrixQ() to get an expression and matrixQ().transpose() to get the transpose. + * Use matrixQ() to get an expression and matrixQ().adjoint() to get the adjoint. * You can then apply it to a vector. * * R is the sparse triangular or trapezoidal matrix. The later occurs when A is rank-deficient. @@ -65,6 +65,7 @@ namespace internal { * \implsparsesolverconcept * * \warning The input sparse matrix A must be in compressed mode (see SparseMatrix::makeCompressed()). + * \warning For complex matrices matrixQ().transpose() will actually return the adjoint matrix. * */ template @@ -196,9 +197,9 @@ class SparseQR : public SparseSolverBase > Index rank = this->rank(); - // Compute Q^T * b; + // Compute Q^* * b; typename Dest::PlainObject y, b; - y = this->matrixQ().transpose() * B; + y = this->matrixQ().adjoint() * B; b = y; // Solve with the triangular matrix R @@ -604,7 +605,7 @@ struct SparseQR_QProduct : ReturnByValue @@ -668,13 +672,14 @@ struct SparseQRMatrixQReturnType : public EigenBase(m_qr,other.derived(),false); } + // To use for operations with the adjoint of Q SparseQRMatrixQTransposeReturnType adjoint() const { return SparseQRMatrixQTransposeReturnType(m_qr); } inline Index rows() const { return m_qr.rows(); } - inline Index cols() const { return (std::min)(m_qr.rows(),m_qr.cols()); } - // To use for operations with the transpose of Q + inline Index cols() const { return m_qr.rows(); } + // To use for operations with the transpose of Q FIXME this is the same as adjoint at the moment SparseQRMatrixQTransposeReturnType transpose() const { return SparseQRMatrixQTransposeReturnType(m_qr); @@ -682,6 +687,7 @@ struct SparseQRMatrixQReturnType : public EigenBase struct SparseQRMatrixQTransposeReturnType { @@ -712,7 +718,7 @@ struct Assignment, internal: typedef typename DstXprType::StorageIndex StorageIndex; static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &/*func*/) { - typename DstXprType::PlainObject idMat(src.m_qr.rows(), src.m_qr.rows()); + typename DstXprType::PlainObject idMat(src.rows(), src.cols()); idMat.setIdentity(); // Sort the sparse householder reflectors if needed const_cast(&src.m_qr)->_sort_matrix_Q(); diff --git a/xs/src/eigen/Eigen/src/UmfPackSupport/UmfPackSupport.h b/xs/src/eigen/Eigen/src/UmfPackSupport/UmfPackSupport.h index dc74de935..91c09ab13 100644 --- a/xs/src/eigen/Eigen/src/UmfPackSupport/UmfPackSupport.h +++ b/xs/src/eigen/Eigen/src/UmfPackSupport/UmfPackSupport.h @@ -10,19 +10,37 @@ #ifndef EIGEN_UMFPACKSUPPORT_H #define EIGEN_UMFPACKSUPPORT_H -namespace Eigen { +namespace Eigen { /* TODO extract L, extract U, compute det, etc... */ // generic double/complex wrapper functions: -inline void umfpack_defaults(double control[UMFPACK_CONTROL], double) +inline void umfpack_defaults(double control[UMFPACK_CONTROL], double) { umfpack_di_defaults(control); } -inline void umfpack_defaults(double control[UMFPACK_CONTROL], std::complex) +inline void umfpack_defaults(double control[UMFPACK_CONTROL], std::complex) { umfpack_zi_defaults(control); } +inline void umfpack_report_info(double control[UMFPACK_CONTROL], double info[UMFPACK_INFO], double) +{ umfpack_di_report_info(control, info);} + +inline void umfpack_report_info(double control[UMFPACK_CONTROL], double info[UMFPACK_INFO], std::complex) +{ umfpack_zi_report_info(control, info);} + +inline void umfpack_report_status(double control[UMFPACK_CONTROL], int status, double) +{ umfpack_di_report_status(control, status);} + +inline void umfpack_report_status(double control[UMFPACK_CONTROL], int status, std::complex) +{ umfpack_zi_report_status(control, status);} + +inline void umfpack_report_control(double control[UMFPACK_CONTROL], double) +{ umfpack_di_report_control(control);} + +inline void umfpack_report_control(double control[UMFPACK_CONTROL], std::complex) +{ umfpack_zi_report_control(control);} + inline void umfpack_free_numeric(void **Numeric, double) { umfpack_di_free_numeric(Numeric); *Numeric = 0; } @@ -156,6 +174,7 @@ class UmfPackLU : public SparseSolverBase > public: typedef Array UmfpackControl; + typedef Array UmfpackInfo; UmfPackLU() : m_dummy(0,0), mp_matrix(m_dummy) @@ -215,7 +234,7 @@ class UmfPackLU : public SparseSolverBase > return m_q; } - /** Computes the sparse Cholesky decomposition of \a matrix + /** Computes the sparse Cholesky decomposition of \a matrix * Note that the matrix should be column-major, and in compressed format for best performance. * \sa SparseMatrix::makeCompressed(). */ @@ -240,7 +259,7 @@ class UmfPackLU : public SparseSolverBase > { if(m_symbolic) umfpack_free_symbolic(&m_symbolic,Scalar()); if(m_numeric) umfpack_free_numeric(&m_numeric,Scalar()); - + grab(matrix.derived()); analyzePattern_impl(); @@ -267,7 +286,7 @@ class UmfPackLU : public SparseSolverBase > { return m_control; } - + /** Provides access to the control settings array used by UmfPack. * * If this array contains NaN's, the default values are used. @@ -278,7 +297,7 @@ class UmfPackLU : public SparseSolverBase > { return m_control; } - + /** Performs a numeric decomposition of \a matrix * * The given matrix must has the same sparcity than the matrix on which the pattern anylysis has been performed. @@ -293,10 +312,38 @@ class UmfPackLU : public SparseSolverBase > umfpack_free_numeric(&m_numeric,Scalar()); grab(matrix.derived()); - + factorize_impl(); } + /** Prints the current UmfPack control settings. + * + * \sa umfpackControl() + */ + void umfpackReportControl() + { + umfpack_report_control(m_control.data(), Scalar()); + } + + /** Prints statistics collected by UmfPack. + * + * \sa analyzePattern(), compute() + */ + void umfpackReportInfo() + { + eigen_assert(m_analysisIsOk && "UmfPackLU: you must first call analyzePattern()"); + umfpack_report_info(m_control.data(), m_umfpackInfo.data(), Scalar()); + } + + /** Prints the status of the previous factorization operation performed by UmfPack (symbolic or numerical factorization). + * + * \sa analyzePattern(), compute() + */ + void umfpackReportStatus() { + eigen_assert(m_analysisIsOk && "UmfPackLU: you must first call analyzePattern()"); + umfpack_report_status(m_control.data(), m_fact_errorCode, Scalar()); + } + /** \internal */ template bool _solve_impl(const MatrixBase &b, MatrixBase &x) const; @@ -314,41 +361,42 @@ class UmfPackLU : public SparseSolverBase > m_numeric = 0; m_symbolic = 0; m_extractedDataAreDirty = true; + + umfpack_defaults(m_control.data(), Scalar()); } - + void analyzePattern_impl() { - umfpack_defaults(m_control.data(), Scalar()); - int errorCode = 0; - errorCode = umfpack_symbolic(internal::convert_index(mp_matrix.rows()), - internal::convert_index(mp_matrix.cols()), - mp_matrix.outerIndexPtr(), mp_matrix.innerIndexPtr(), mp_matrix.valuePtr(), - &m_symbolic, m_control.data(), 0); + m_fact_errorCode = umfpack_symbolic(internal::convert_index(mp_matrix.rows()), + internal::convert_index(mp_matrix.cols()), + mp_matrix.outerIndexPtr(), mp_matrix.innerIndexPtr(), mp_matrix.valuePtr(), + &m_symbolic, m_control.data(), m_umfpackInfo.data()); m_isInitialized = true; - m_info = errorCode ? InvalidInput : Success; + m_info = m_fact_errorCode ? InvalidInput : Success; m_analysisIsOk = true; m_factorizationIsOk = false; m_extractedDataAreDirty = true; } - + void factorize_impl() { + m_fact_errorCode = umfpack_numeric(mp_matrix.outerIndexPtr(), mp_matrix.innerIndexPtr(), mp_matrix.valuePtr(), - m_symbolic, &m_numeric, m_control.data(), 0); + m_symbolic, &m_numeric, m_control.data(), m_umfpackInfo.data()); m_info = m_fact_errorCode == UMFPACK_OK ? Success : NumericalIssue; m_factorizationIsOk = true; m_extractedDataAreDirty = true; } - + template void grab(const EigenBase &A) { mp_matrix.~UmfpackMatrixRef(); ::new (&mp_matrix) UmfpackMatrixRef(A.derived()); } - + void grab(const UmfpackMatrixRef &A) { if(&(A.derived()) != &mp_matrix) @@ -357,19 +405,20 @@ class UmfPackLU : public SparseSolverBase > ::new (&mp_matrix) UmfpackMatrixRef(A); } } - + // cached data to reduce reallocation, etc. mutable LUMatrixType m_l; int m_fact_errorCode; UmfpackControl m_control; - + mutable UmfpackInfo m_umfpackInfo; + mutable LUMatrixType m_u; mutable IntColVectorType m_p; mutable IntRowVectorType m_q; UmfpackMatrixType m_dummy; UmfpackMatrixRef mp_matrix; - + void* m_numeric; void* m_symbolic; @@ -377,7 +426,7 @@ class UmfPackLU : public SparseSolverBase > int m_factorizationIsOk; int m_analysisIsOk; mutable bool m_extractedDataAreDirty; - + private: UmfPackLU(const UmfPackLU& ) { } }; @@ -427,7 +476,7 @@ bool UmfPackLU::_solve_impl(const MatrixBase &b, MatrixBas eigen_assert((BDerived::Flags&RowMajorBit)==0 && "UmfPackLU backend does not support non col-major rhs yet"); eigen_assert((XDerived::Flags&RowMajorBit)==0 && "UmfPackLU backend does not support non col-major result yet"); eigen_assert(b.derived().data() != x.derived().data() && " Umfpack does not support inplace solve"); - + int errorCode; Scalar* x_ptr = 0; Matrix x_tmp; @@ -442,7 +491,7 @@ bool UmfPackLU::_solve_impl(const MatrixBase &b, MatrixBas x_ptr = &x.col(j).coeffRef(0); errorCode = umfpack_solve(UMFPACK_A, mp_matrix.outerIndexPtr(), mp_matrix.innerIndexPtr(), mp_matrix.valuePtr(), - x_ptr, &b.const_cast_derived().col(j).coeffRef(0), m_numeric, m_control.data(), 0); + x_ptr, &b.const_cast_derived().col(j).coeffRef(0), m_numeric, m_control.data(), m_umfpackInfo.data()); if(x.innerStride()!=1) x.col(j) = x_tmp; if (errorCode!=0) diff --git a/xs/src/eigen/README.md b/xs/src/eigen/README.md index f56262cca..39892c00b 100644 --- a/xs/src/eigen/README.md +++ b/xs/src/eigen/README.md @@ -1,5 +1,5 @@ THIS IS NOT THE COMPLETE EIGEN DISTRIBUTION. ONLY FILES NEEDED FOR COMPILING EIGEN INTO SLIC3R WERE PUT INTO THE SLIC3R SOURCE DISTRIBUTION. -THIS DIRECTORY CONTAINS PIECES OF THE EIGEN 3.3.3 SOURCE DISTRIBUTION. +THIS DIRECTORY CONTAINS PIECES OF THE EIGEN 3.3.5 b3f3d4950030 SOURCE DISTRIBUTION. **Eigen is a C++ template library for linear algebra: matrices, vectors, numerical solvers, and related algorithms.**