diff options
author | Deven Desai <deven.desai.amd@gmail.com> | 2019-03-19 16:52:38 -0400 |
---|---|---|
committer | Deven Desai <deven.desai.amd@gmail.com> | 2019-03-19 16:52:38 -0400 |
commit | 2dbea5510fe5cb64dbfdef9042c04a3a92b87f76 (patch) | |
tree | c187e7ec5e90a191e19466ff6084dd8f053dba7e /Eigen/src/Core | |
parent | e7e6809e6b38a5928efc0b5ca9520258e4d1fb3a (diff) | |
parent | 5c93b38c5fca514a08084e32feb8a8fb27bf3665 (diff) |
Merged eigen/eigen into default
Diffstat (limited to 'Eigen/src/Core')
79 files changed, 4360 insertions, 2137 deletions
diff --git a/Eigen/src/Core/Array.h b/Eigen/src/Core/Array.h index e10020d4f..ee12d96fc 100644 --- a/Eigen/src/Core/Array.h +++ b/Eigen/src/Core/Array.h @@ -153,8 +153,6 @@ class Array : Base(std::move(other)) { Base::_check_template_params(); - if (RowsAtCompileTime!=Dynamic && ColsAtCompileTime!=Dynamic) - Base::_set_noalias(other); } EIGEN_DEVICE_FUNC Array& operator=(Array&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_assignable<Scalar>::value) @@ -180,6 +178,46 @@ class Array Base::_check_template_params(); this->template _init2<T0,T1>(val0, val1); } + + #if EIGEN_HAS_CXX11 + /** \copydoc PlainObjectBase(const Scalar& a0, const Scalar& a1, const Scalar& a2, const Scalar& a3, const ArgTypes&... args) + * + * Example: \include Array_variadic_ctor_cxx11.cpp + * Output: \verbinclude Array_variadic_ctor_cxx11.out + * + * \sa Array(const std::initializer_list<std::initializer_list<Scalar>>&) + * \sa Array(Scalar), Array(Scalar,Scalar) + */ + template <typename... ArgTypes> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Array(const Scalar& a0, const Scalar& a1, const Scalar& a2, const Scalar& a3, const ArgTypes&... args) + : Base(a0, a1, a2, a3, args...) {} + + /** \brief Constructs an array and initializes it from the coefficients given as initializer-lists grouped by row. \cpp11 + * + * In the general case, the constructor takes a list of rows, each row being represented as a list of coefficients: + * + * Example: \include Array_initializer_list_23_cxx11.cpp + * Output: \verbinclude Array_initializer_list_23_cxx11.out + * + * Each of the inner initializer lists must contain the exact same number of elements, otherwise an assertion is triggered. + * + * In the case of a compile-time column 1D array, implicit transposition from a single row is allowed. + * Therefore <code> Array<int,Dynamic,1>{{1,2,3,4,5}}</code> is legal and the more verbose syntax + * <code>Array<int,Dynamic,1>{{1},{2},{3},{4},{5}}</code> can be avoided: + * + * Example: \include Array_initializer_list_vector_cxx11.cpp + * Output: \verbinclude Array_initializer_list_vector_cxx11.out + * + * In the case of fixed-sized arrays, the initializer list sizes must exactly match the array sizes, + * and implicit transposition is allowed for compile-time 1D arrays only. + * + * \sa Array(const Scalar& a0, const Scalar& a1, const Scalar& a2, const Scalar& a3, const ArgTypes&... args) + */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Array(const std::initializer_list<std::initializer_list<Scalar> >& list) : Base(list) {} + #endif // end EIGEN_HAS_CXX11 + #else /** \brief Constructs a fixed-sized array initialized with coefficients starting at \a data */ EIGEN_DEVICE_FUNC explicit Array(const Scalar *data); @@ -191,7 +229,8 @@ class Array */ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit Array(Index dim); - /** constructs an initialized 1x1 Array with the given coefficient */ + /** constructs an initialized 1x1 Array with the given coefficient + * \sa const Scalar& a0, const Scalar& a1, const Scalar& a2, const Scalar& a3, const ArgTypes&... args */ Array(const Scalar& value); /** constructs an uninitialized array with \a rows rows and \a cols columns. * @@ -199,11 +238,14 @@ class Array * it is redundant to pass these parameters, so one should use the default constructor * Array() instead. */ Array(Index rows, Index cols); - /** constructs an initialized 2D vector with given coefficients */ + /** constructs an initialized 2D vector with given coefficients + * \sa Array(const Scalar& a0, const Scalar& a1, const Scalar& a2, const Scalar& a3, const ArgTypes&... args) */ Array(const Scalar& val0, const Scalar& val1); - #endif + #endif // end EIGEN_PARSED_BY_DOXYGEN - /** constructs an initialized 3D vector with given coefficients */ + /** constructs an initialized 3D vector with given coefficients + * \sa Array(const Scalar& a0, const Scalar& a1, const Scalar& a2, const Scalar& a3, const ArgTypes&... args) + */ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Array(const Scalar& val0, const Scalar& val1, const Scalar& val2) { @@ -213,7 +255,9 @@ class Array m_storage.data()[1] = val1; m_storage.data()[2] = val2; } - /** constructs an initialized 4D vector with given coefficients */ + /** constructs an initialized 4D vector with given coefficients + * \sa Array(const Scalar& a0, const Scalar& a1, const Scalar& a2, const Scalar& a3, const ArgTypes&... args) + */ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Array(const Scalar& val0, const Scalar& val1, const Scalar& val2, const Scalar& val3) { @@ -260,7 +304,7 @@ class Array /** \defgroup arraytypedefs Global array typedefs * \ingroup Core_Module * - * Eigen defines several typedef shortcuts for most common 1D and 2D array types. + * %Eigen defines several typedef shortcuts for most common 1D and 2D array types. * * The general patterns are the following: * @@ -273,6 +317,12 @@ class Array * There are also \c ArraySizeType which are self-explanatory. For example, \c Array4cf is * a fixed-size 1D array of 4 complex floats. * + * With \cpp11, template alias are also defined for common sizes. + * They follow the same pattern as above except that the scalar type suffix is replaced by a + * template parameter, i.e.: + * - `ArrayRowsCols<Type>` where `Rows` and `Cols` can be \c 2,\c 3,\c 4, or \c X for fixed or dynamic size. + * - `ArraySize<Type>` where `Size` can be \c 2,\c 3,\c 4 or \c X for fixed or dynamic size 1D arrays. + * * \sa class Array */ @@ -305,9 +355,43 @@ EIGEN_MAKE_ARRAY_TYPEDEFS_ALL_SIZES(std::complex<double>, cd) #undef EIGEN_MAKE_ARRAY_TYPEDEFS_ALL_SIZES #undef EIGEN_MAKE_ARRAY_TYPEDEFS +#undef EIGEN_MAKE_ARRAY_FIXED_TYPEDEFS + +#if EIGEN_HAS_CXX11 + +#define EIGEN_MAKE_ARRAY_TYPEDEFS(Size, SizeSuffix) \ +/** \ingroup arraytypedefs */ \ +/** \brief \cpp11 */ \ +template <typename Type> \ +using Array##SizeSuffix##SizeSuffix = Array<Type, Size, Size>; \ +/** \ingroup arraytypedefs */ \ +/** \brief \cpp11 */ \ +template <typename Type> \ +using Array##SizeSuffix = Array<Type, Size, 1>; + +#define EIGEN_MAKE_ARRAY_FIXED_TYPEDEFS(Size) \ +/** \ingroup arraytypedefs */ \ +/** \brief \cpp11 */ \ +template <typename Type> \ +using Array##Size##X = Array<Type, Size, Dynamic>; \ +/** \ingroup arraytypedefs */ \ +/** \brief \cpp11 */ \ +template <typename Type> \ +using Array##X##Size = Array<Type, Dynamic, Size>; + +EIGEN_MAKE_ARRAY_TYPEDEFS(2, 2) +EIGEN_MAKE_ARRAY_TYPEDEFS(3, 3) +EIGEN_MAKE_ARRAY_TYPEDEFS(4, 4) +EIGEN_MAKE_ARRAY_TYPEDEFS(Dynamic, X) +EIGEN_MAKE_ARRAY_FIXED_TYPEDEFS(2) +EIGEN_MAKE_ARRAY_FIXED_TYPEDEFS(3) +EIGEN_MAKE_ARRAY_FIXED_TYPEDEFS(4) -#undef EIGEN_MAKE_ARRAY_TYPEDEFS_LARGE +#undef EIGEN_MAKE_ARRAY_TYPEDEFS +#undef EIGEN_MAKE_ARRAY_FIXED_TYPEDEFS +#endif // EIGEN_HAS_CXX11 + #define EIGEN_USING_ARRAY_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, SizeSuffix) \ using Eigen::Matrix##SizeSuffix##TypeSuffix; \ using Eigen::Vector##SizeSuffix##TypeSuffix; \ diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 79575e1b4..229e25854 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -611,7 +611,8 @@ public: typedef typename AssignmentTraits::PacketType PacketType; - EIGEN_DEVICE_FUNC generic_dense_assignment_kernel(DstEvaluatorType &dst, const SrcEvaluatorType &src, const Functor &func, DstXprType& dstExpr) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + generic_dense_assignment_kernel(DstEvaluatorType &dst, const SrcEvaluatorType &src, const Functor &func, DstXprType& dstExpr) : m_dst(dst), m_src(src), m_functor(func), m_dstExpr(dstExpr) { #ifdef EIGEN_DEBUG_ASSIGN diff --git a/Eigen/src/Core/Assign_MKL.h b/Eigen/src/Core/Assign_MKL.h index 6866095bf..c6140d185 100755 --- a/Eigen/src/Core/Assign_MKL.h +++ b/Eigen/src/Core/Assign_MKL.h @@ -68,16 +68,16 @@ class vml_assign_traits #define EIGEN_PP_EXPAND(ARG) ARG #if !defined (EIGEN_FAST_MATH) || (EIGEN_FAST_MATH != 1) -#define EIGEN_VMLMODE_EXPAND_LA , VML_HA +#define EIGEN_VMLMODE_EXPAND_xLA , VML_HA #else -#define EIGEN_VMLMODE_EXPAND_LA , VML_LA +#define EIGEN_VMLMODE_EXPAND_xLA , VML_LA #endif -#define EIGEN_VMLMODE_EXPAND__ +#define EIGEN_VMLMODE_EXPAND_x_ -#define EIGEN_VMLMODE_PREFIX_LA vm -#define EIGEN_VMLMODE_PREFIX__ v -#define EIGEN_VMLMODE_PREFIX(VMLMODE) EIGEN_CAT(EIGEN_VMLMODE_PREFIX_,VMLMODE) +#define EIGEN_VMLMODE_PREFIX_xLA vm +#define EIGEN_VMLMODE_PREFIX_x_ v +#define EIGEN_VMLMODE_PREFIX(VMLMODE) EIGEN_CAT(EIGEN_VMLMODE_PREFIX_x,VMLMODE) #define EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE, VMLMODE) \ template< typename DstXprType, typename SrcXprNested> \ @@ -89,7 +89,7 @@ class vml_assign_traits eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); \ if(vml_assign_traits<DstXprType,SrcXprNested>::Traversal==LinearTraversal) { \ VMLOP(dst.size(), (const VMLTYPE*)src.nestedExpression().data(), \ - (VMLTYPE*)dst.data() EIGEN_PP_EXPAND(EIGEN_VMLMODE_EXPAND_##VMLMODE) ); \ + (VMLTYPE*)dst.data() EIGEN_PP_EXPAND(EIGEN_VMLMODE_EXPAND_x##VMLMODE) ); \ } else { \ const Index outerSize = dst.outerSize(); \ for(Index outer = 0; outer < outerSize; ++outer) { \ @@ -97,7 +97,7 @@ class vml_assign_traits &(src.nestedExpression().coeffRef(0, outer)); \ EIGENTYPE *dst_ptr = dst.IsRowMajor ? &(dst.coeffRef(outer,0)) : &(dst.coeffRef(0, outer)); \ VMLOP( dst.innerSize(), (const VMLTYPE*)src_ptr, \ - (VMLTYPE*)dst_ptr EIGEN_PP_EXPAND(EIGEN_VMLMODE_EXPAND_##VMLMODE)); \ + (VMLTYPE*)dst_ptr EIGEN_PP_EXPAND(EIGEN_VMLMODE_EXPAND_x##VMLMODE)); \ } \ } \ } \ @@ -152,7 +152,7 @@ EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(ceil, Ceil, _) if(vml_assign_traits<DstXprType,SrcXprNested>::Traversal==LinearTraversal) \ { \ VMLOP( dst.size(), (const VMLTYPE*)src.lhs().data(), exponent, \ - (VMLTYPE*)dst.data() EIGEN_PP_EXPAND(EIGEN_VMLMODE_EXPAND_##VMLMODE) ); \ + (VMLTYPE*)dst.data() EIGEN_PP_EXPAND(EIGEN_VMLMODE_EXPAND_x##VMLMODE) ); \ } else { \ const Index outerSize = dst.outerSize(); \ for(Index outer = 0; outer < outerSize; ++outer) { \ @@ -160,7 +160,7 @@ EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(ceil, Ceil, _) &(src.lhs().coeffRef(0, outer)); \ EIGENTYPE *dst_ptr = dst.IsRowMajor ? &(dst.coeffRef(outer,0)) : &(dst.coeffRef(0, outer)); \ VMLOP( dst.innerSize(), (const VMLTYPE*)src_ptr, exponent, \ - (VMLTYPE*)dst_ptr EIGEN_PP_EXPAND(EIGEN_VMLMODE_EXPAND_##VMLMODE)); \ + (VMLTYPE*)dst_ptr EIGEN_PP_EXPAND(EIGEN_VMLMODE_EXPAND_x##VMLMODE)); \ } \ } \ } \ diff --git a/Eigen/src/Core/Block.h b/Eigen/src/Core/Block.h index 11de45c2e..6e938ea58 100644 --- a/Eigen/src/Core/Block.h +++ b/Eigen/src/Core/Block.h @@ -114,8 +114,8 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel> class /** Column or Row constructor */ - EIGEN_DEVICE_FUNC - inline Block(XprType& xpr, Index i) : Impl(xpr,i) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Block(XprType& xpr, Index i) : Impl(xpr,i) { eigen_assert( (i>=0) && ( ((BlockRows==1) && (BlockCols==XprType::ColsAtCompileTime) && i<xpr.rows()) @@ -124,8 +124,8 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel> class /** Fixed-size constructor */ - EIGEN_DEVICE_FUNC - inline Block(XprType& xpr, Index startRow, Index startCol) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Block(XprType& xpr, Index startRow, Index startCol) : Impl(xpr, startRow, startCol) { EIGEN_STATIC_ASSERT(RowsAtCompileTime!=Dynamic && ColsAtCompileTime!=Dynamic,THIS_METHOD_IS_ONLY_FOR_FIXED_SIZE) @@ -135,8 +135,8 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel> class /** Dynamic-size constructor */ - EIGEN_DEVICE_FUNC - inline Block(XprType& xpr, + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Block(XprType& xpr, Index startRow, Index startCol, Index blockRows, Index blockCols) : Impl(xpr, startRow, startCol, blockRows, blockCols) @@ -159,10 +159,10 @@ class BlockImpl<XprType, BlockRows, BlockCols, InnerPanel, Dense> public: typedef Impl Base; EIGEN_INHERIT_ASSIGNMENT_OPERATORS(BlockImpl) - EIGEN_DEVICE_FUNC inline BlockImpl(XprType& xpr, Index i) : Impl(xpr,i) {} - EIGEN_DEVICE_FUNC inline BlockImpl(XprType& xpr, Index startRow, Index startCol) : Impl(xpr, startRow, startCol) {} + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE BlockImpl(XprType& xpr, Index i) : Impl(xpr,i) {} + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE BlockImpl(XprType& xpr, Index startRow, Index startCol) : Impl(xpr, startRow, startCol) {} EIGEN_DEVICE_FUNC - inline BlockImpl(XprType& xpr, Index startRow, Index startCol, Index blockRows, Index blockCols) + EIGEN_STRONG_INLINE BlockImpl(XprType& xpr, Index startRow, Index startCol, Index blockRows, Index blockCols) : Impl(xpr, startRow, startCol, blockRows, blockCols) {} }; @@ -294,22 +294,22 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel, bool H EIGEN_DEVICE_FUNC inline Index outerStride() const; #endif - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename internal::remove_all<XprTypeNested>::type& nestedExpression() const { return m_xpr; } - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE XprType& nestedExpression() { return m_xpr; } - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE StorageIndex startRow() const { return m_startRow.value(); } - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE StorageIndex startCol() const { return m_startCol.value(); @@ -342,8 +342,8 @@ class BlockImpl_dense<XprType,BlockRows,BlockCols, InnerPanel,true> /** Column or Row constructor */ - EIGEN_DEVICE_FUNC - inline BlockImpl_dense(XprType& xpr, Index i) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + BlockImpl_dense(XprType& xpr, Index i) : Base(xpr.data() + i * ( ((BlockRows==1) && (BlockCols==XprType::ColsAtCompileTime) && (!XprTypeIsRowMajor)) || ((BlockRows==XprType::RowsAtCompileTime) && (BlockCols==1) && ( XprTypeIsRowMajor)) ? xpr.innerStride() : xpr.outerStride()), BlockRows==1 ? 1 : xpr.rows(), @@ -357,8 +357,8 @@ class BlockImpl_dense<XprType,BlockRows,BlockCols, InnerPanel,true> /** Fixed-size constructor */ - EIGEN_DEVICE_FUNC - inline BlockImpl_dense(XprType& xpr, Index startRow, Index startCol) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + BlockImpl_dense(XprType& xpr, Index startRow, Index startCol) : Base(xpr.data()+xpr.innerStride()*(XprTypeIsRowMajor?startCol:startRow) + xpr.outerStride()*(XprTypeIsRowMajor?startRow:startCol)), m_xpr(xpr), m_startRow(startRow), m_startCol(startCol) { @@ -367,8 +367,8 @@ class BlockImpl_dense<XprType,BlockRows,BlockCols, InnerPanel,true> /** Dynamic-size constructor */ - EIGEN_DEVICE_FUNC - inline BlockImpl_dense(XprType& xpr, + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + BlockImpl_dense(XprType& xpr, Index startRow, Index startCol, Index blockRows, Index blockCols) : Base(xpr.data()+xpr.innerStride()*(XprTypeIsRowMajor?startCol:startRow) + xpr.outerStride()*(XprTypeIsRowMajor?startRow:startCol), blockRows, blockCols), @@ -377,18 +377,18 @@ class BlockImpl_dense<XprType,BlockRows,BlockCols, InnerPanel,true> init(); } - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename internal::remove_all<XprTypeNested>::type& nestedExpression() const { return m_xpr; } - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE XprType& nestedExpression() { return m_xpr; } /** \sa MapBase::innerStride() */ - EIGEN_DEVICE_FUNC - inline Index innerStride() const + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Index innerStride() const { return internal::traits<BlockType>::HasSameStorageOrderAsXprType ? m_xpr.innerStride() @@ -396,19 +396,19 @@ class BlockImpl_dense<XprType,BlockRows,BlockCols, InnerPanel,true> } /** \sa MapBase::outerStride() */ - EIGEN_DEVICE_FUNC - inline Index outerStride() const + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Index outerStride() const { return m_outerStride; } - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE StorageIndex startRow() const { return m_startRow.value(); } - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE StorageIndex startCol() const { return m_startCol.value(); @@ -422,8 +422,8 @@ class BlockImpl_dense<XprType,BlockRows,BlockCols, InnerPanel,true> #ifndef EIGEN_PARSED_BY_DOXYGEN /** \internal used by allowAligned() */ - EIGEN_DEVICE_FUNC - inline BlockImpl_dense(XprType& xpr, const Scalar* data, Index blockRows, Index blockCols) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + BlockImpl_dense(XprType& xpr, const Scalar* data, Index blockRows, Index blockCols) : Base(data, blockRows, blockCols), m_xpr(xpr) { init(); @@ -431,7 +431,7 @@ class BlockImpl_dense<XprType,BlockRows,BlockCols, InnerPanel,true> #endif protected: - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void init() { m_outerStride = internal::traits<BlockType>::HasSameStorageOrderAsXprType diff --git a/Eigen/src/Core/ConditionEstimator.h b/Eigen/src/Core/ConditionEstimator.h index aa7efdc76..51a2e5f1b 100644 --- a/Eigen/src/Core/ConditionEstimator.h +++ b/Eigen/src/Core/ConditionEstimator.h @@ -160,7 +160,7 @@ rcond_estimate_helper(typename Decomposition::RealScalar matrix_norm, const Deco { typedef typename Decomposition::RealScalar RealScalar; eigen_assert(dec.rows() == dec.cols()); - if (dec.rows() == 0) return RealScalar(1); + if (dec.rows() == 0) return NumTraits<RealScalar>::infinity(); if (matrix_norm == RealScalar(0)) return RealScalar(0); if (dec.rows() == 1) return RealScalar(1); const RealScalar inverse_matrix_norm = rcond_invmatrix_L1_norm_estimate(dec); diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index d5da5cdec..670fa77b5 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -90,7 +90,8 @@ template<typename T> struct evaluator : public unary_evaluator<T> { typedef unary_evaluator<T> Base; - EIGEN_DEVICE_FUNC explicit evaluator(const T& xpr) : Base(xpr) {} + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit evaluator(const T& xpr) : Base(xpr) {} }; @@ -99,7 +100,7 @@ template<typename T> struct evaluator<const T> : evaluator<T> { - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit evaluator(const T& xpr) : evaluator<T>(xpr) {} }; @@ -134,21 +135,25 @@ private: // this helper permits to completely eliminate m_outerStride if it is known at compiletime. template<typename Scalar,int OuterStride> class plainobjectbase_evaluator_data { public: - EIGEN_DEVICE_FUNC plainobjectbase_evaluator_data(const Scalar* ptr, Index outerStride) : data(ptr) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + plainobjectbase_evaluator_data(const Scalar* ptr, Index outerStride) : data(ptr) { #ifndef EIGEN_INTERNAL_DEBUGGING EIGEN_UNUSED_VARIABLE(outerStride); #endif eigen_internal_assert(outerStride==OuterStride); } - EIGEN_DEVICE_FUNC Index outerStride() const { return OuterStride; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Index outerStride() const { return OuterStride; } const Scalar *data; }; template<typename Scalar> class plainobjectbase_evaluator_data<Scalar,Dynamic> { public: - EIGEN_DEVICE_FUNC plainobjectbase_evaluator_data(const Scalar* ptr, Index outerStride) : data(ptr), m_outerStride(outerStride) {} - EIGEN_DEVICE_FUNC Index outerStride() const { return m_outerStride; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + plainobjectbase_evaluator_data(const Scalar* ptr, Index outerStride) : data(ptr), m_outerStride(outerStride) {} + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Index outerStride() const { return m_outerStride; } const Scalar *data; protected: Index m_outerStride; @@ -179,13 +184,15 @@ struct evaluator<PlainObjectBase<Derived> > : RowsAtCompileTime }; - EIGEN_DEVICE_FUNC evaluator() + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + evaluator() : m_d(0,OuterStrideAtCompileTime) { EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost); } - EIGEN_DEVICE_FUNC explicit evaluator(const PlainObjectType& m) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit evaluator(const PlainObjectType& m) : m_d(m.data(),IsVectorAtCompileTime ? 0 : m.outerStride()) { EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost); @@ -268,9 +275,11 @@ struct evaluator<Matrix<Scalar, Rows, Cols, Options, MaxRows, MaxCols> > { typedef Matrix<Scalar, Rows, Cols, Options, MaxRows, MaxCols> XprType; - EIGEN_DEVICE_FUNC evaluator() {} + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + evaluator() {} - EIGEN_DEVICE_FUNC explicit evaluator(const XprType& m) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit evaluator(const XprType& m) : evaluator<PlainObjectBase<XprType> >(m) { } }; @@ -281,9 +290,11 @@ struct evaluator<Array<Scalar, Rows, Cols, Options, MaxRows, MaxCols> > { typedef Array<Scalar, Rows, Cols, Options, MaxRows, MaxCols> XprType; - EIGEN_DEVICE_FUNC evaluator() {} + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + evaluator() {} - EIGEN_DEVICE_FUNC explicit evaluator(const XprType& m) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit evaluator(const XprType& m) : evaluator<PlainObjectBase<XprType> >(m) { } }; @@ -302,7 +313,8 @@ struct unary_evaluator<Transpose<ArgType>, IndexBased> Alignment = evaluator<ArgType>::Alignment }; - EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& t) : m_argImpl(t.nestedExpression()) {} + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit unary_evaluator(const XprType& t) : m_argImpl(t.nestedExpression()) {} typedef typename XprType::Scalar Scalar; typedef typename XprType::CoeffReturnType CoeffReturnType; @@ -712,7 +724,8 @@ struct evaluator<CwiseBinaryOp<BinaryOp, Lhs, Rhs> > typedef CwiseBinaryOp<BinaryOp, Lhs, Rhs> XprType; typedef binary_evaluator<CwiseBinaryOp<BinaryOp, Lhs, Rhs> > Base; - EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr) : Base(xpr) {} + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit evaluator(const XprType& xpr) : Base(xpr) {} }; template<typename BinaryOp, typename Lhs, typename Rhs> @@ -740,7 +753,8 @@ struct binary_evaluator<CwiseBinaryOp<BinaryOp, Lhs, Rhs>, IndexBased, IndexBase Alignment = EIGEN_PLAIN_ENUM_MIN(evaluator<Lhs>::Alignment,evaluator<Rhs>::Alignment) }; - EIGEN_DEVICE_FUNC explicit binary_evaluator(const XprType& xpr) : m_d(xpr) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit binary_evaluator(const XprType& xpr) : m_d(xpr) { EIGEN_INTERNAL_CHECK_COST_VALUE(functor_traits<BinaryOp>::Cost); EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost); @@ -877,7 +891,8 @@ struct mapbase_evaluator : evaluator_base<Derived> CoeffReadCost = NumTraits<Scalar>::ReadCost }; - EIGEN_DEVICE_FUNC explicit mapbase_evaluator(const XprType& map) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit mapbase_evaluator(const XprType& map) : m_data(const_cast<PointerType>(map.data())), m_innerStride(map.innerStride()), m_outerStride(map.outerStride()) @@ -941,10 +956,10 @@ struct mapbase_evaluator : evaluator_base<Derived> internal::pstoret<Scalar, PacketType, StoreMode>(m_data + index * m_innerStride.value(), x); } protected: - EIGEN_DEVICE_FUNC - inline Index rowStride() const { return XprType::IsRowMajor ? m_outerStride.value() : m_innerStride.value(); } - EIGEN_DEVICE_FUNC - inline Index colStride() const { return XprType::IsRowMajor ? m_innerStride.value() : m_outerStride.value(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Index rowStride() const { return XprType::IsRowMajor ? m_outerStride.value() : m_innerStride.value(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Index colStride() const { return XprType::IsRowMajor ? m_innerStride.value() : m_outerStride.value(); } PointerType m_data; const internal::variable_if_dynamic<Index, XprType::InnerStrideAtCompileTime> m_innerStride; @@ -997,7 +1012,8 @@ struct evaluator<Ref<PlainObjectType, RefOptions, StrideType> > Alignment = evaluator<Map<PlainObjectType, RefOptions, StrideType> >::Alignment }; - EIGEN_DEVICE_FUNC explicit evaluator(const XprType& ref) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit evaluator(const XprType& ref) : mapbase_evaluator<XprType, PlainObjectType>(ref) { } }; @@ -1052,7 +1068,8 @@ struct evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel> > Alignment = EIGEN_PLAIN_ENUM_MIN(evaluator<ArgType>::Alignment, Alignment0) }; typedef block_evaluator<ArgType, BlockRows, BlockCols, InnerPanel> block_evaluator_type; - EIGEN_DEVICE_FUNC explicit evaluator(const XprType& block) : block_evaluator_type(block) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit evaluator(const XprType& block) : block_evaluator_type(block) { EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost); } @@ -1065,7 +1082,8 @@ struct block_evaluator<ArgType, BlockRows, BlockCols, InnerPanel, /*HasDirectAcc { typedef Block<ArgType, BlockRows, BlockCols, InnerPanel> XprType; - EIGEN_DEVICE_FUNC explicit block_evaluator(const XprType& block) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit block_evaluator(const XprType& block) : unary_evaluator<XprType>(block) {} }; @@ -1076,7 +1094,8 @@ struct unary_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel>, IndexBa { typedef Block<ArgType, BlockRows, BlockCols, InnerPanel> XprType; - EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& block) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit unary_evaluator(const XprType& block) : m_argImpl(block.nestedExpression()), m_startRow(block.startRow()), m_startCol(block.startCol()), @@ -1176,7 +1195,8 @@ struct block_evaluator<ArgType, BlockRows, BlockCols, InnerPanel, /* HasDirectAc typedef Block<ArgType, BlockRows, BlockCols, InnerPanel> XprType; typedef typename XprType::Scalar Scalar; - EIGEN_DEVICE_FUNC explicit block_evaluator(const XprType& block) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit block_evaluator(const XprType& block) : mapbase_evaluator<XprType, typename XprType::PlainObject>(block) { // TODO: for the 3.3 release, this should be turned to an internal assertion, but let's keep it as is for the beta lifetime @@ -1204,7 +1224,8 @@ struct evaluator<Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> > Alignment = EIGEN_PLAIN_ENUM_MIN(evaluator<ThenMatrixType>::Alignment, evaluator<ElseMatrixType>::Alignment) }; - EIGEN_DEVICE_FUNC explicit evaluator(const XprType& select) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit evaluator(const XprType& select) : m_conditionImpl(select.conditionMatrix()), m_thenImpl(select.thenMatrix()), m_elseImpl(select.elseMatrix()) @@ -1261,7 +1282,8 @@ struct unary_evaluator<Replicate<ArgType, RowFactor, ColFactor> > Alignment = evaluator<ArgTypeNestedCleaned>::Alignment }; - EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& replicate) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit unary_evaluator(const XprType& replicate) : m_arg(replicate.nestedExpression()), m_argImpl(m_arg), m_rows(replicate.nestedExpression().rows()), @@ -1341,7 +1363,8 @@ struct evaluator_wrapper_base Alignment = evaluator<ArgType>::Alignment }; - EIGEN_DEVICE_FUNC explicit evaluator_wrapper_base(const ArgType& arg) : m_argImpl(arg) {} + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit evaluator_wrapper_base(const ArgType& arg) : m_argImpl(arg) {} typedef typename ArgType::Scalar Scalar; typedef typename ArgType::CoeffReturnType CoeffReturnType; @@ -1408,7 +1431,8 @@ struct unary_evaluator<MatrixWrapper<TArgType> > { typedef MatrixWrapper<TArgType> XprType; - EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& wrapper) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit unary_evaluator(const XprType& wrapper) : evaluator_wrapper_base<MatrixWrapper<TArgType> >(wrapper.nestedExpression()) { } }; @@ -1419,7 +1443,8 @@ struct unary_evaluator<ArrayWrapper<TArgType> > { typedef ArrayWrapper<TArgType> XprType; - EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& wrapper) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit unary_evaluator(const XprType& wrapper) : evaluator_wrapper_base<ArrayWrapper<TArgType> >(wrapper.nestedExpression()) { } }; @@ -1461,7 +1486,8 @@ struct unary_evaluator<Reverse<ArgType, Direction> > Alignment = 0 // FIXME in some rare cases, Alignment could be preserved, like a Vector4f. }; - EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& reverse) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit unary_evaluator(const XprType& reverse) : m_argImpl(reverse.nestedExpression()), m_rows(ReverseRow ? reverse.nestedExpression().rows() : 1), m_cols(ReverseCol ? reverse.nestedExpression().cols() : 1) @@ -1568,7 +1594,8 @@ struct evaluator<Diagonal<ArgType, DiagIndex> > Alignment = 0 }; - EIGEN_DEVICE_FUNC explicit evaluator(const XprType& diagonal) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit evaluator(const XprType& diagonal) : m_argImpl(diagonal.nestedExpression()), m_index(diagonal.index()) { } diff --git a/Eigen/src/Core/CwiseBinaryOp.h b/Eigen/src/Core/CwiseBinaryOp.h index bf2632d9e..8b8de8382 100644 --- a/Eigen/src/Core/CwiseBinaryOp.h +++ b/Eigen/src/Core/CwiseBinaryOp.h @@ -100,8 +100,14 @@ class CwiseBinaryOp : typedef typename internal::remove_reference<LhsNested>::type _LhsNested; typedef typename internal::remove_reference<RhsNested>::type _RhsNested; - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE CwiseBinaryOp(const Lhs& aLhs, const Rhs& aRhs, const BinaryOp& func = BinaryOp()) +#if EIGEN_COMP_MSVC && EIGEN_HAS_CXX11 + //Required for Visual Studio or the Copy constructor will probably not get inlined! + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CwiseBinaryOp(const CwiseBinaryOp<BinaryOp,LhsType,RhsType>&) = default; +#endif + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CwiseBinaryOp(const Lhs& aLhs, const Rhs& aRhs, const BinaryOp& func = BinaryOp()) : m_lhs(aLhs), m_rhs(aRhs), m_functor(func) { EIGEN_CHECK_BINARY_COMPATIBILIY(BinaryOp,typename Lhs::Scalar,typename Rhs::Scalar); @@ -110,16 +116,16 @@ class CwiseBinaryOp : eigen_assert(aLhs.rows() == aRhs.rows() && aLhs.cols() == aRhs.cols()); } - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE Index rows() const { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Index rows() const { // return the fixed size type if available to enable compile time optimizations if (internal::traits<typename internal::remove_all<LhsNested>::type>::RowsAtCompileTime==Dynamic) return m_rhs.rows(); else return m_lhs.rows(); } - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE Index cols() const { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Index cols() const { // return the fixed size type if available to enable compile time optimizations if (internal::traits<typename internal::remove_all<LhsNested>::type>::ColsAtCompileTime==Dynamic) return m_rhs.cols(); @@ -128,13 +134,13 @@ class CwiseBinaryOp : } /** \returns the left hand side nested expression */ - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const _LhsNested& lhs() const { return m_lhs; } /** \returns the right hand side nested expression */ - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const _RhsNested& rhs() const { return m_rhs; } /** \returns the functor representing the binary operation */ - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const BinaryOp& functor() const { return m_functor; } protected: diff --git a/Eigen/src/Core/CwiseNullaryOp.h b/Eigen/src/Core/CwiseNullaryOp.h index d149abe93..ef708197b 100644 --- a/Eigen/src/Core/CwiseNullaryOp.h +++ b/Eigen/src/Core/CwiseNullaryOp.h @@ -239,7 +239,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomA DenseBase<Derived>::LinSpaced(Sequential_t, Index size, const Scalar& low, const Scalar& high) { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) - return DenseBase<Derived>::NullaryExpr(size, internal::linspaced_op<Scalar,PacketScalar>(low,high,size)); + return DenseBase<Derived>::NullaryExpr(size, internal::linspaced_op<Scalar>(low,high,size)); } /** \deprecated because of accuracy loss. In Eigen 3.3, it is an alias for LinSpaced(const Scalar&,const Scalar&) @@ -252,7 +252,7 @@ DenseBase<Derived>::LinSpaced(Sequential_t, const Scalar& low, const Scalar& hig { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived) - return DenseBase<Derived>::NullaryExpr(Derived::SizeAtCompileTime, internal::linspaced_op<Scalar,PacketScalar>(low,high,Derived::SizeAtCompileTime)); + return DenseBase<Derived>::NullaryExpr(Derived::SizeAtCompileTime, internal::linspaced_op<Scalar>(low,high,Derived::SizeAtCompileTime)); } /** @@ -283,7 +283,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomA DenseBase<Derived>::LinSpaced(Index size, const Scalar& low, const Scalar& high) { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) - return DenseBase<Derived>::NullaryExpr(size, internal::linspaced_op<Scalar,PacketScalar>(low,high,size)); + return DenseBase<Derived>::NullaryExpr(size, internal::linspaced_op<Scalar>(low,high,size)); } /** @@ -296,7 +296,7 @@ DenseBase<Derived>::LinSpaced(const Scalar& low, const Scalar& high) { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived) - return DenseBase<Derived>::NullaryExpr(Derived::SizeAtCompileTime, internal::linspaced_op<Scalar,PacketScalar>(low,high,Derived::SizeAtCompileTime)); + return DenseBase<Derived>::NullaryExpr(Derived::SizeAtCompileTime, internal::linspaced_op<Scalar>(low,high,Derived::SizeAtCompileTime)); } /** \returns true if all coefficients in this matrix are approximately equal to \a val, to within precision \a prec */ @@ -398,7 +398,7 @@ template<typename Derived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setLinSpaced(Index newSize, const Scalar& low, const Scalar& high) { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) - return derived() = Derived::NullaryExpr(newSize, internal::linspaced_op<Scalar,PacketScalar>(low,high,newSize)); + return derived() = Derived::NullaryExpr(newSize, internal::linspaced_op<Scalar>(low,high,newSize)); } /** diff --git a/Eigen/src/Core/CwiseUnaryView.h b/Eigen/src/Core/CwiseUnaryView.h index 271033056..21cf5ea9e 100644 --- a/Eigen/src/Core/CwiseUnaryView.h +++ b/Eigen/src/Core/CwiseUnaryView.h @@ -81,7 +81,7 @@ class CwiseUnaryView : public CwiseUnaryViewImpl<ViewOp, MatrixType, typename in /** \returns the nested expression */ typename internal::remove_reference<MatrixTypeNested>::type& - nestedExpression() { return m_matrix.const_cast_derived(); } + nestedExpression() { return m_matrix; } protected: MatrixTypeNested m_matrix; diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h index 2a0927317..2289fe41f 100644 --- a/Eigen/src/Core/DenseBase.h +++ b/Eigen/src/Core/DenseBase.h @@ -40,7 +40,7 @@ static inline void check_DenseIndex_is_signed() { */ template<typename Derived> class DenseBase #ifndef EIGEN_PARSED_BY_DOXYGEN - : public DenseCoeffsBase<Derived> + : public DenseCoeffsBase<Derived, internal::accessors_level<Derived>::value> #else : public DenseCoeffsBase<Derived,DirectWriteAccessors> #endif // not EIGEN_PARSED_BY_DOXYGEN @@ -71,7 +71,7 @@ template<typename Derived> class DenseBase typedef Scalar value_type; typedef typename NumTraits<Scalar>::Real RealScalar; - typedef DenseCoeffsBase<Derived> Base; + typedef DenseCoeffsBase<Derived, internal::accessors_level<Derived>::value> Base; using Base::derived; using Base::const_cast_derived; @@ -150,8 +150,8 @@ template<typename Derived> class DenseBase * \sa SizeAtCompileTime, MaxRowsAtCompileTime, MaxColsAtCompileTime */ - IsVectorAtCompileTime = internal::traits<Derived>::MaxRowsAtCompileTime == 1 - || internal::traits<Derived>::MaxColsAtCompileTime == 1, + IsVectorAtCompileTime = internal::traits<Derived>::RowsAtCompileTime == 1 + || internal::traits<Derived>::ColsAtCompileTime == 1, /**< This is set to true if either the number of rows or the number of * columns is known at compile-time to be equal to 1. Indeed, in that case, * we are dealing with a column-vector (if there is only one column) or with @@ -266,9 +266,9 @@ template<typename Derived> class DenseBase /** \internal Represents a matrix with all coefficients equal to one another*/ typedef CwiseNullaryOp<internal::scalar_constant_op<Scalar>,PlainObject> ConstantReturnType; /** \internal \deprecated Represents a vector with linearly spaced coefficients that allows sequential access only. */ - typedef CwiseNullaryOp<internal::linspaced_op<Scalar,PacketScalar>,PlainObject> SequentialLinSpacedReturnType; + typedef CwiseNullaryOp<internal::linspaced_op<Scalar>,PlainObject> SequentialLinSpacedReturnType; /** \internal Represents a vector with linearly spaced coefficients that allows random access. */ - typedef CwiseNullaryOp<internal::linspaced_op<Scalar,PacketScalar>,PlainObject> RandomAccessLinSpacedReturnType; + typedef CwiseNullaryOp<internal::linspaced_op<Scalar>,PlainObject> RandomAccessLinSpacedReturnType; /** \internal the return type of MatrixBase::eigenvalues() */ typedef Matrix<typename NumTraits<typename internal::traits<Derived>::Scalar>::Real, internal::traits<Derived>::ColsAtCompileTime, 1> EigenvaluesReturnType; @@ -415,7 +415,7 @@ template<typename Derived> class DenseBase * */ template<typename OtherDerived> - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void swap(const DenseBase<OtherDerived>& other) { EIGEN_STATIC_ASSERT(!OtherDerived::IsPlainObjectBase,THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY); @@ -427,7 +427,7 @@ template<typename Derived> class DenseBase * */ template<typename OtherDerived> - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void swap(PlainObjectBase<OtherDerived>& other) { eigen_assert(rows()==other.rows() && cols()==other.cols()); diff --git a/Eigen/src/Core/DiagonalMatrix.h b/Eigen/src/Core/DiagonalMatrix.h index afab2f1b6..542685c65 100644 --- a/Eigen/src/Core/DiagonalMatrix.h +++ b/Eigen/src/Core/DiagonalMatrix.h @@ -178,6 +178,30 @@ class DiagonalMatrix EIGEN_DEVICE_FUNC inline DiagonalMatrix(const Scalar& x, const Scalar& y, const Scalar& z) : m_diagonal(x,y,z) {} + #if EIGEN_HAS_CXX11 + /** \brief Construct a diagonal matrix with fixed size from an arbitrary number of coefficients. \cpp11 + * + * There exists C++98 anologue constructors for fixed-size diagonal matrices having 2 or 3 coefficients. + * + * \warning To construct a diagonal matrix of fixed size, the number of values passed to this + * constructor must match the fixed dimension of \c *this. + * + * \sa DiagonalMatrix(const Scalar&, const Scalar&) + * \sa DiagonalMatrix(const Scalar&, const Scalar&, const Scalar&) + */ + template <typename... ArgTypes> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + DiagonalMatrix(const Scalar& a0, const Scalar& a1, const Scalar& a2, const ArgTypes&... args) + : m_diagonal(a0, a1, a2, args...) {} + + /** \brief Constructs a DiagonalMatrix and initializes it by elements given by an initializer list of initializer + * lists \cpp11 + */ + EIGEN_DEVICE_FUNC + explicit EIGEN_STRONG_INLINE DiagonalMatrix(const std::initializer_list<std::initializer_list<Scalar>>& list) + : m_diagonal(list) {} + #endif // EIGEN_HAS_CXX11 + /** Copy constructor. */ template<typename OtherDerived> EIGEN_DEVICE_FUNC diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h index 43f3b84c8..bf7ef54b5 100644 --- a/Eigen/src/Core/GeneralProduct.h +++ b/Eigen/src/Core/GeneralProduct.h @@ -239,7 +239,7 @@ template<> struct gemv_dense_selector<OnTheRight,ColMajor,true> // on, the other hand it is good for the cache to pack the vector anyways... EvalToDestAtCompileTime = (ActualDest::InnerStrideAtCompileTime==1), ComplexByReal = (NumTraits<LhsScalar>::IsComplex) && (!NumTraits<RhsScalar>::IsComplex), - MightCannotUseDest = (!EvalToDestAtCompileTime) || ComplexByReal + MightCannotUseDest = ((!EvalToDestAtCompileTime) || ComplexByReal) && (ActualDest::MaxSizeAtCompileTime!=0) }; typedef const_blas_data_mapper<LhsScalar,Index,ColMajor> LhsMapper; @@ -326,7 +326,7 @@ template<> struct gemv_dense_selector<OnTheRight,RowMajor,true> enum { // FIXME find a way to allow an inner stride on the result if packet_traits<Scalar>::size==1 // on, the other hand it is good for the cache to pack the vector anyways... - DirectlyUseRhs = ActualRhsTypeCleaned::InnerStrideAtCompileTime==1 + DirectlyUseRhs = ActualRhsTypeCleaned::InnerStrideAtCompileTime==1 || ActualRhsTypeCleaned::MaxSizeAtCompileTime==0 }; gemv_static_vector_if<RhsScalar,ActualRhsTypeCleaned::SizeAtCompileTime,ActualRhsTypeCleaned::MaxSizeAtCompileTime,!DirectlyUseRhs> static_rhs; @@ -396,8 +396,8 @@ template<> struct gemv_dense_selector<OnTheRight,RowMajor,false> */ template<typename Derived> template<typename OtherDerived> -EIGEN_DEVICE_FUNC -inline const Product<Derived, OtherDerived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +const Product<Derived, OtherDerived> MatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const { // A note regarding the function declaration: In MSVC, this function will sometimes @@ -439,8 +439,9 @@ MatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const */ template<typename Derived> template<typename OtherDerived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Product<Derived,OtherDerived,LazyProduct> -EIGEN_DEVICE_FUNC MatrixBase<Derived>::lazyProduct(const MatrixBase<OtherDerived> &other) const +MatrixBase<Derived>::lazyProduct(const MatrixBase<OtherDerived> &other) const { enum { ProductIsValid = Derived::ColsAtCompileTime==Dynamic diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h index da1350f1b..04a321b9f 100644 --- a/Eigen/src/Core/GenericPacketMath.h +++ b/Eigen/src/Core/GenericPacketMath.h @@ -56,6 +56,7 @@ struct default_packet_traits HasConj = 1, HasSetLinear = 1, HasBlend = 0, + HasReduxp = 1, HasDiv = 0, HasSqrt = 0, @@ -151,15 +152,18 @@ pcast(const SrcPacket& a, const SrcPacket& /*b*/, const SrcPacket& /*c*/, const return static_cast<TgtPacket>(a); } +/** \internal \returns reinterpret_cast<Target>(a) */ +template <typename Target, typename Packet> +EIGEN_DEVICE_FUNC inline Target +preinterpret(const Packet& a); /* { return reinterpret_cast<const Target&>(a); } */ + /** \internal \returns a + b (coeff-wise) */ template<typename Packet> EIGEN_DEVICE_FUNC inline Packet -padd(const Packet& a, - const Packet& b) { return a+b; } +padd(const Packet& a, const Packet& b) { return a+b; } /** \internal \returns a - b (coeff-wise) */ template<typename Packet> EIGEN_DEVICE_FUNC inline Packet -psub(const Packet& a, - const Packet& b) { return a-b; } +psub(const Packet& a, const Packet& b) { return a-b; } /** \internal \returns -a (coeff-wise) */ template<typename Packet> EIGEN_DEVICE_FUNC inline Packet @@ -172,23 +176,19 @@ pconj(const Packet& a) { return numext::conj(a); } /** \internal \returns a * b (coeff-wise) */ template<typename Packet> EIGEN_DEVICE_FUNC inline Packet -pmul(const Packet& a, - const Packet& b) { return a*b; } +pmul(const Packet& a, const Packet& b) { return a*b; } /** \internal \returns a / b (coeff-wise) */ template<typename Packet> EIGEN_DEVICE_FUNC inline Packet -pdiv(const Packet& a, - const Packet& b) { return a/b; } +pdiv(const Packet& a, const Packet& b) { return a/b; } /** \internal \returns the min of \a a and \a b (coeff-wise) */ template<typename Packet> EIGEN_DEVICE_FUNC inline Packet -pmin(const Packet& a, - const Packet& b) { return numext::mini(a, b); } +pmin(const Packet& a, const Packet& b) { return numext::mini(a, b); } /** \internal \returns the max of \a a and \a b (coeff-wise) */ template<typename Packet> EIGEN_DEVICE_FUNC inline Packet -pmax(const Packet& a, - const Packet& b) { return numext::maxi(a, b); } +pmax(const Packet& a, const Packet& b) { return numext::maxi(a, b); } /** \internal \returns the absolute value of \a a */ template<typename Packet> EIGEN_DEVICE_FUNC inline Packet @@ -212,7 +212,72 @@ pxor(const Packet& a, const Packet& b) { return a ^ b; } /** \internal \returns the bitwise andnot of \a a and \a b */ template<typename Packet> EIGEN_DEVICE_FUNC inline Packet -pandnot(const Packet& a, const Packet& b) { return a & (!b); } +pandnot(const Packet& a, const Packet& b) { return a & (~b); } + +/** \internal \returns ones */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +ptrue(const Packet& /*a*/) { Packet b; memset((void*)&b, 0xff, sizeof(b)); return b;} + +template <typename RealScalar> +EIGEN_DEVICE_FUNC inline std::complex<RealScalar> ptrue(const std::complex<RealScalar>& /*a*/) { + RealScalar b; + b = ptrue(b); + return std::complex<RealScalar>(b, b); +} + +/** \internal \returns the bitwise not of \a a */ +template <typename Packet> EIGEN_DEVICE_FUNC inline Packet +pnot(const Packet& a) { return pxor(ptrue(a), a);} + +/** \internal \returns \a a shifted by N bits to the right */ +template<int N> EIGEN_DEVICE_FUNC inline int +pshiftright(const int& a) { return a >> N; } +template<int N> EIGEN_DEVICE_FUNC inline long int +pshiftright(const long int& a) { return a >> N; } + +/** \internal \returns \a a shifted by N bits to the left */ +template<int N> EIGEN_DEVICE_FUNC inline int +pshiftleft(const int& a) { return a << N; } +template<int N> EIGEN_DEVICE_FUNC inline long int +pshiftleft(const long int& a) { return a << N; } + +/** \internal \returns the significant and exponent of the underlying floating point numbers + * See https://en.cppreference.com/w/cpp/numeric/math/frexp + */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +pfrexp(const Packet &a, Packet &exponent) { return std::frexp(a,&exponent); } + +/** \internal \returns a * 2^exponent + * See https://en.cppreference.com/w/cpp/numeric/math/ldexp + */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +pldexp(const Packet &a, const Packet &exponent) { return std::ldexp(a,exponent); } + +/** \internal \returns zeros */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +pzero(const Packet& a) { return pxor(a,a); } + +/** \internal \returns bits of \a or \b according to the input bit mask \a mask */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +pselect(const Packet& mask, const Packet& a, const Packet& b) { + return por(pand(a,mask),pandnot(b,mask)); +} + +/** \internal \returns a <= b as a bit mask */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +pcmp_le(const Packet& a, const Packet& b) { return a<=b ? ptrue(a) : pzero(a); } + +/** \internal \returns a < b as a bit mask */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +pcmp_lt(const Packet& a, const Packet& b) { return a<b ? ptrue(a) : pzero(a); } + +/** \internal \returns a == b as a bit mask */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +pcmp_eq(const Packet& a, const Packet& b) { return a==b ? ptrue(a) : pzero(a); } + +/** \internal \returns a < b or a==NaN or b==NaN as a bit mask */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +pcmp_lt_or_nan(const Packet& a, const Packet& b) { return pnot(pcmp_le(b,a)); } /** \internal \returns a packet version of \a *from, from must be 16 bytes aligned */ template<typename Packet> EIGEN_DEVICE_FUNC inline Packet @@ -226,6 +291,10 @@ ploadu(const typename unpacket_traits<Packet>::type* from) { return *from; } template<typename Packet> EIGEN_DEVICE_FUNC inline Packet pset1(const typename unpacket_traits<Packet>::type& a) { return a; } +/** \internal \returns a packet with constant coefficients set from bits */ +template<typename Packet,typename BitsType> EIGEN_DEVICE_FUNC inline Packet +pset1frombits(BitsType a); + /** \internal \returns a packet with constant coefficients \a a[0], e.g.: (a[0],a[0],a[0],a[0]) */ template<typename Packet> EIGEN_DEVICE_FUNC inline Packet pload1(const typename unpacket_traits<Packet>::type *a) { return pset1<Packet>(*a); } @@ -339,18 +408,39 @@ typename conditional<(unpacket_traits<Packet>::size%8)==0,typename unpacket_trai predux_half_dowto4(const Packet& a) { return a; } -/** \internal \returns the product of the elements of \a a*/ +/** \internal \returns the product of the elements of \a a */ template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type predux_mul(const Packet& a) { return a; } -/** \internal \returns the min of the elements of \a a*/ +/** \internal \returns the min of the elements of \a a */ template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type predux_min(const Packet& a) { return a; } -/** \internal \returns the max of the elements of \a a*/ +/** \internal \returns the max of the elements of \a a */ template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type predux_max(const Packet& a) { return a; } +/** \internal \returns true if all coeffs of \a a means "true" + * It is supposed to be called on values returned by pcmp_*. + */ +// not needed yet +// template<typename Packet> EIGEN_DEVICE_FUNC inline bool predux_all(const Packet& a) +// { return bool(a); } + +/** \internal \returns true if any coeffs of \a a means "true" + * It is supposed to be called on values returned by pcmp_*. + */ +template<typename Packet> EIGEN_DEVICE_FUNC inline bool predux_any(const Packet& a) +{ + // Dirty but generic implementation where "true" is assumed to be non 0 and all the sames. + // It is expected that "true" is either: + // - Scalar(1) + // - bits full of ones (NaN for floats), + // - or first bit equals to 1 (1 for ints, smallest denormal for floats). + // For all these cases, taking the sum is just fine, and this boils down to a no-op for scalars. + return bool(predux(a)); +} + /** \internal \returns the reversed elements of \a a*/ template<typename Packet> EIGEN_DEVICE_FUNC inline Packet preverse(const Packet& a) { return a; } @@ -597,6 +687,22 @@ pinsertlast(const Packet& a, typename unpacket_traits<Packet>::type b) return pblend(mask, pset1<Packet>(b), a); } +/*************************************************************************** + * Some generic implementations to be used by implementors +***************************************************************************/ + +/** Default implementation of pfrexp for float. + * It is expected to be called by implementers of template<> pfrexp. + */ +template<typename Packet> EIGEN_STRONG_INLINE Packet +pfrexp_float(const Packet& a, Packet& exponent); + +/** Default implementation of pldexp for float. + * It is expected to be called by implementers of template<> pldexp. + */ +template<typename Packet> EIGEN_STRONG_INLINE Packet +pldexp_float(Packet a, Packet exponent); + } // end namespace internal } // end namespace Eigen diff --git a/Eigen/src/Core/GlobalFunctions.h b/Eigen/src/Core/GlobalFunctions.h index 563df6e84..71377cee5 100644 --- a/Eigen/src/Core/GlobalFunctions.h +++ b/Eigen/src/Core/GlobalFunctions.h @@ -66,6 +66,11 @@ namespace Eigen EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sinh,scalar_sinh_op,hyperbolic sine,\sa ArrayBase::sinh) EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(cosh,scalar_cosh_op,hyperbolic cosine,\sa ArrayBase::cosh) EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(tanh,scalar_tanh_op,hyperbolic tangent,\sa ArrayBase::tanh) +#if EIGEN_HAS_CXX11_MATH + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(asinh,scalar_asinh_op,inverse hyperbolic sine,\sa ArrayBase::asinh) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(acosh,scalar_acosh_op,inverse hyperbolic cosine,\sa ArrayBase::acosh) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(atanh,scalar_atanh_op,inverse hyperbolic tangent,\sa ArrayBase::atanh) +#endif EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(logistic,scalar_logistic_op,logistic function,\sa ArrayBase::logistic) EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(lgamma,scalar_lgamma_op,natural logarithm of the gamma function,\sa ArrayBase::lgamma) EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(digamma,scalar_digamma_op,derivative of lgamma,\sa ArrayBase::digamma) diff --git a/Eigen/src/Core/IO.h b/Eigen/src/Core/IO.h index da7fd6cce..063511f24 100644 --- a/Eigen/src/Core/IO.h +++ b/Eigen/src/Core/IO.h @@ -41,6 +41,7 @@ std::ostream & print_matrix(std::ostream & s, const Derived& _m, const IOFormat& * - \b rowSuffix string printed at the end of each row * - \b matPrefix string printed at the beginning of the matrix * - \b matSuffix string printed at the end of the matrix + * - \b fill character printed to fill the empty space in aligned columns * * Example: \include IOFormat.cpp * Output: \verbinclude IOFormat.out @@ -53,9 +54,9 @@ struct IOFormat IOFormat(int _precision = StreamPrecision, int _flags = 0, const std::string& _coeffSeparator = " ", const std::string& _rowSeparator = "\n", const std::string& _rowPrefix="", const std::string& _rowSuffix="", - const std::string& _matPrefix="", const std::string& _matSuffix="") + const std::string& _matPrefix="", const std::string& _matSuffix="", const char _fill=' ') : matPrefix(_matPrefix), matSuffix(_matSuffix), rowPrefix(_rowPrefix), rowSuffix(_rowSuffix), rowSeparator(_rowSeparator), - rowSpacer(""), coeffSeparator(_coeffSeparator), precision(_precision), flags(_flags) + rowSpacer(""), coeffSeparator(_coeffSeparator), fill(_fill), precision(_precision), flags(_flags) { // TODO check if rowPrefix, rowSuffix or rowSeparator contains a newline // don't add rowSpacer if columns are not to be aligned @@ -71,6 +72,7 @@ struct IOFormat std::string matPrefix, matSuffix; std::string rowPrefix, rowSuffix, rowSeparator, rowSpacer; std::string coeffSeparator; + char fill; int precision; int flags; }; @@ -176,18 +178,26 @@ std::ostream & print_matrix(std::ostream & s, const Derived& _m, const IOFormat& width = std::max<Index>(width, Index(sstr.str().length())); } } + std::streamsize old_width = s.width(); + char old_fill_character = s.fill(); s << fmt.matPrefix; for(Index i = 0; i < m.rows(); ++i) { if (i) s << fmt.rowSpacer; s << fmt.rowPrefix; - if(width) s.width(width); + if(width) { + s.fill(fmt.fill); + s.width(width); + } s << m.coeff(i, 0); for(Index j = 1; j < m.cols(); ++j) { s << fmt.coeffSeparator; - if (width) s.width(width); + if(width) { + s.fill(fmt.fill); + s.width(width); + } s << m.coeff(i, j); } s << fmt.rowSuffix; @@ -196,6 +206,10 @@ std::ostream & print_matrix(std::ostream & s, const Derived& _m, const IOFormat& } s << fmt.matSuffix; if(explicit_precision) s.precision(old_precision); + if(width) { + s.fill(old_fill_character); + s.width(old_width); + } return s; } diff --git a/Eigen/src/Core/IndexedView.h b/Eigen/src/Core/IndexedView.h index 3485d8f46..377f8a5cc 100644 --- a/Eigen/src/Core/IndexedView.h +++ b/Eigen/src/Core/IndexedView.h @@ -132,7 +132,7 @@ public: /** \returns the nested expression */ typename internal::remove_reference<XprType>::type& - nestedExpression() { return m_xpr.const_cast_derived(); } + nestedExpression() { return m_xpr; } /** \returns a const reference to the object storing/generating the row indices */ const RowIndices& rowIndices() const { return m_rowIndices; } diff --git a/Eigen/src/Core/Matrix.h b/Eigen/src/Core/Matrix.h index 90c336d8c..4b714328c 100644 --- a/Eigen/src/Core/Matrix.h +++ b/Eigen/src/Core/Matrix.h @@ -255,29 +255,27 @@ class Matrix * * \sa resize(Index,Index) */ - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE Matrix() : Base() + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Matrix() : Base() { Base::_check_template_params(); EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED } // FIXME is it still needed - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit Matrix(internal::constructor_without_unaligned_array_assert) : Base(internal::constructor_without_unaligned_array_assert()) { Base::_check_template_params(); EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED } #if EIGEN_HAS_RVALUE_REFERENCES - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Matrix(Matrix&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_constructible<Scalar>::value) : Base(std::move(other)) { Base::_check_template_params(); - if (RowsAtCompileTime!=Dynamic && ColsAtCompileTime!=Dynamic) - Base::_set_noalias(other); } - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Matrix& operator=(Matrix&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_assignable<Scalar>::value) { other.swap(*this); @@ -289,20 +287,59 @@ class Matrix // This constructor is for both 1x1 matrices and dynamic vectors template<typename T> - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE explicit Matrix(const T& x) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit Matrix(const T& x) { Base::_check_template_params(); Base::template _init1<T>(x); } template<typename T0, typename T1> - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE Matrix(const T0& x, const T1& y) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Matrix(const T0& x, const T1& y) { Base::_check_template_params(); Base::template _init2<T0,T1>(x, y); } + + #if EIGEN_HAS_CXX11 + /** \copydoc PlainObjectBase(const Scalar&, const Scalar&, const Scalar&, const Scalar&, const ArgTypes&...) + * + * Example: \include Matrix_variadic_ctor_cxx11.cpp + * Output: \verbinclude Matrix_variadic_ctor_cxx11.out + * + * \sa Matrix(const std::initializer_list<std::initializer_list<Scalar>>&) + */ + template <typename... ArgTypes> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Matrix(const Scalar& a0, const Scalar& a1, const Scalar& a2, const Scalar& a3, const ArgTypes&... args) + : Base(a0, a1, a2, a3, args...) {} + + /** \brief Constructs a Matrix and initializes it from the coefficients given as initializer-lists grouped by row. \cpp11 + * + * In the general case, the constructor takes a list of rows, each row being represented as a list of coefficients: + * + * Example: \include Matrix_initializer_list_23_cxx11.cpp + * Output: \verbinclude Matrix_initializer_list_23_cxx11.out + * + * Each of the inner initializer lists must contain the exact same number of elements, otherwise an assertion is triggered. + * + * In the case of a compile-time column vector, implicit transposition from a single row is allowed. + * Therefore <code>VectorXd{{1,2,3,4,5}}</code> is legal and the more verbose syntax + * <code>RowVectorXd{{1},{2},{3},{4},{5}}</code> can be avoided: + * + * Example: \include Matrix_initializer_list_vector_cxx11.cpp + * Output: \verbinclude Matrix_initializer_list_vector_cxx11.out + * + * In the case of fixed-sized matrices, the initializer list sizes must exactly match the matrix sizes, + * and implicit transposition is allowed for compile-time vectors only. + * + * \sa Matrix(const Scalar&, const Scalar&, const Scalar&, const Scalar&, const ArgTypes&...) + */ + EIGEN_DEVICE_FUNC + explicit EIGEN_STRONG_INLINE Matrix(const std::initializer_list<std::initializer_list<Scalar>>& list) : Base(list) {} + #endif // end EIGEN_HAS_CXX11 + #else /** \brief Constructs a fixed-sized matrix initialized with coefficients starting at \a data */ EIGEN_DEVICE_FUNC @@ -321,7 +358,8 @@ class Matrix * \c EIGEN_INITIALIZE_MATRICES_BY_{ZERO,\c NAN} macros (see \ref TopicPreprocessorDirectives). */ EIGEN_STRONG_INLINE explicit Matrix(Index dim); - /** \brief Constructs an initialized 1x1 matrix with the given coefficient */ + /** \brief Constructs an initialized 1x1 matrix with the given coefficient + * \sa Matrix(const Scalar&, const Scalar&, const Scalar&, const Scalar&, const ArgTypes&...) */ Matrix(const Scalar& x); /** \brief Constructs an uninitialized matrix with \a rows rows and \a cols columns. * @@ -338,11 +376,14 @@ class Matrix EIGEN_DEVICE_FUNC Matrix(Index rows, Index cols); - /** \brief Constructs an initialized 2D vector with given coefficients */ + /** \brief Constructs an initialized 2D vector with given coefficients + * \sa Matrix(const Scalar&, const Scalar&, const Scalar&, const Scalar&, const ArgTypes&...) */ Matrix(const Scalar& x, const Scalar& y); - #endif + #endif // end EIGEN_PARSED_BY_DOXYGEN - /** \brief Constructs an initialized 3D vector with given coefficients */ + /** \brief Constructs an initialized 3D vector with given coefficients + * \sa Matrix(const Scalar&, const Scalar&, const Scalar&, const Scalar&, const ArgTypes&...) + */ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Matrix(const Scalar& x, const Scalar& y, const Scalar& z) { @@ -352,7 +393,9 @@ class Matrix m_storage.data()[1] = y; m_storage.data()[2] = z; } - /** \brief Constructs an initialized 4D vector with given coefficients */ + /** \brief Constructs an initialized 4D vector with given coefficients + * \sa Matrix(const Scalar&, const Scalar&, const Scalar&, const Scalar&, const ArgTypes&...) + */ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Matrix(const Scalar& x, const Scalar& y, const Scalar& z, const Scalar& w) { @@ -407,7 +450,7 @@ class Matrix * * \ingroup Core_Module * - * Eigen defines several typedef shortcuts for most common matrix and vector types. + * %Eigen defines several typedef shortcuts for most common matrix and vector types. * * The general patterns are the following: * @@ -419,6 +462,15 @@ class Matrix * * There are also \c VectorSizeType and \c RowVectorSizeType which are self-explanatory. For example, \c Vector4cf is * a fixed-size vector of 4 complex floats. + * + * With \cpp11, template alias are also defined for common sizes. + * They follow the same pattern as above except that the scalar type suffix is replaced by a + * template parameter, i.e.: + * - `MatrixSize<Type>` where `Size` can be \c 2,\c 3,\c 4 for fixed size square matrices or \c X for dynamic size. + * - `MatrixXSize<Type>` and `MatrixSizeX<Type>` where `Size` can be \c 2,\c 3,\c 4 for hybrid dynamic/fixed matrices. + * - `VectorSize<Type>` and `RowVectorSize<Type>` for column and row vectors. + * + * With \cpp11, you can also use fully generic column and row vector types: `Vector<Type,Size>` and `RowVector<Type,Size>`. * * \sa class Matrix */ @@ -456,6 +508,55 @@ EIGEN_MAKE_TYPEDEFS_ALL_SIZES(std::complex<double>, cd) #undef EIGEN_MAKE_TYPEDEFS #undef EIGEN_MAKE_FIXED_TYPEDEFS +#if EIGEN_HAS_CXX11 + +#define EIGEN_MAKE_TYPEDEFS(Size, SizeSuffix) \ +/** \ingroup matrixtypedefs */ \ +/** \brief \cpp11 */ \ +template <typename Type> \ +using Matrix##SizeSuffix = Matrix<Type, Size, Size>; \ +/** \ingroup matrixtypedefs */ \ +/** \brief \cpp11 */ \ +template <typename Type> \ +using Vector##SizeSuffix = Matrix<Type, Size, 1>; \ +/** \ingroup matrixtypedefs */ \ +/** \brief \cpp11 */ \ +template <typename Type> \ +using RowVector##SizeSuffix = Matrix<Type, 1, Size>; + +#define EIGEN_MAKE_FIXED_TYPEDEFS(Size) \ +/** \ingroup matrixtypedefs */ \ +/** \brief \cpp11 */ \ +template <typename Type> \ +using Matrix##Size##X = Matrix<Type, Size, Dynamic>; \ +/** \ingroup matrixtypedefs */ \ +/** \brief \cpp11 */ \ +template <typename Type> \ +using Matrix##X##Size = Matrix<Type, Dynamic, Size>; + +EIGEN_MAKE_TYPEDEFS(2, 2) +EIGEN_MAKE_TYPEDEFS(3, 3) +EIGEN_MAKE_TYPEDEFS(4, 4) +EIGEN_MAKE_TYPEDEFS(Dynamic, X) +EIGEN_MAKE_FIXED_TYPEDEFS(2) +EIGEN_MAKE_FIXED_TYPEDEFS(3) +EIGEN_MAKE_FIXED_TYPEDEFS(4) + +/** \ingroup matrixtypedefs + * \brief \cpp11 */ +template <typename Type, int Size> +using Vector = Matrix<Type, Size, 1>; + +/** \ingroup matrixtypedefs + * \brief \cpp11 */ +template <typename Type, int Size> +using RowVector = Matrix<Type, 1, Size>; + +#undef EIGEN_MAKE_TYPEDEFS +#undef EIGEN_MAKE_FIXED_TYPEDEFS + +#endif // EIGEN_HAS_CXX11 + } // end namespace Eigen #endif // EIGEN_MATRIX_H diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h index 596cdd133..4744e5cc4 100644 --- a/Eigen/src/Core/MatrixBase.h +++ b/Eigen/src/Core/MatrixBase.h @@ -468,6 +468,11 @@ template<typename Derived> class MatrixBase const MatrixFunctionReturnValue<Derived> matrixFunction(StemFunction f) const; EIGEN_MATRIX_FUNCTION(MatrixFunctionReturnValue, cosh, hyperbolic cosine) EIGEN_MATRIX_FUNCTION(MatrixFunctionReturnValue, sinh, hyperbolic sine) +#if EIGEN_HAS_CXX11_MATH + EIGEN_MATRIX_FUNCTION(MatrixFunctionReturnValue, atanh, inverse hyperbolic cosine) + EIGEN_MATRIX_FUNCTION(MatrixFunctionReturnValue, acosh, inverse hyperbolic cosine) + EIGEN_MATRIX_FUNCTION(MatrixFunctionReturnValue, asinh, inverse hyperbolic sine) +#endif EIGEN_MATRIX_FUNCTION(MatrixFunctionReturnValue, cos, cosine) EIGEN_MATRIX_FUNCTION(MatrixFunctionReturnValue, sin, sine) EIGEN_MATRIX_FUNCTION(MatrixSquareRootReturnValue, sqrt, square root) diff --git a/Eigen/src/Core/NestByValue.h b/Eigen/src/Core/NestByValue.h index 01cf192e9..239bbba63 100644 --- a/Eigen/src/Core/NestByValue.h +++ b/Eigen/src/Core/NestByValue.h @@ -16,7 +16,11 @@ namespace Eigen { namespace internal { template<typename ExpressionType> struct traits<NestByValue<ExpressionType> > : public traits<ExpressionType> -{}; +{ + enum { + Flags = traits<ExpressionType>::Flags & ~NestByRefBit + }; +}; } /** \class NestByValue @@ -43,55 +47,11 @@ template<typename ExpressionType> class NestByValue EIGEN_DEVICE_FUNC inline Index rows() const { return m_expression.rows(); } EIGEN_DEVICE_FUNC inline Index cols() const { return m_expression.cols(); } - EIGEN_DEVICE_FUNC inline Index outerStride() const { return m_expression.outerStride(); } - EIGEN_DEVICE_FUNC inline Index innerStride() const { return m_expression.innerStride(); } - - EIGEN_DEVICE_FUNC inline const CoeffReturnType coeff(Index row, Index col) const - { - return m_expression.coeff(row, col); - } - - EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index row, Index col) - { - return m_expression.const_cast_derived().coeffRef(row, col); - } - - EIGEN_DEVICE_FUNC inline const CoeffReturnType coeff(Index index) const - { - return m_expression.coeff(index); - } - - EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index index) - { - return m_expression.const_cast_derived().coeffRef(index); - } - - template<int LoadMode> - EIGEN_DEVICE_FUNC inline const PacketScalar packet(Index row, Index col) const - { - return m_expression.template packet<LoadMode>(row, col); - } - - template<int LoadMode> - EIGEN_DEVICE_FUNC inline void writePacket(Index row, Index col, const PacketScalar& x) - { - m_expression.const_cast_derived().template writePacket<LoadMode>(row, col, x); - } - - template<int LoadMode> - EIGEN_DEVICE_FUNC inline const PacketScalar packet(Index index) const - { - return m_expression.template packet<LoadMode>(index); - } - - template<int LoadMode> - EIGEN_DEVICE_FUNC inline void writePacket(Index index, const PacketScalar& x) - { - m_expression.const_cast_derived().template writePacket<LoadMode>(index, x); - } EIGEN_DEVICE_FUNC operator const ExpressionType&() const { return m_expression; } + EIGEN_DEVICE_FUNC const ExpressionType& nestedExpression() const { return m_expression; } + protected: const ExpressionType m_expression; }; @@ -105,6 +65,21 @@ DenseBase<Derived>::nestByValue() const return NestByValue<Derived>(derived()); } +namespace internal { + +// Evaluator of Solve -> eval into a temporary +template<typename ArgType> +struct evaluator<NestByValue<ArgType> > + : public evaluator<ArgType> +{ + typedef evaluator<ArgType> Base; + + EIGEN_DEVICE_FUNC explicit evaluator(const NestByValue<ArgType>& xpr) + : Base(xpr.nestedExpression()) + {} +}; +} + } // end namespace Eigen #endif // EIGEN_NESTBYVALUE_H diff --git a/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h index f551dabb0..6de78fd2f 100644 --- a/Eigen/src/Core/PlainObjectBase.h +++ b/Eigen/src/Core/PlainObjectBase.h @@ -526,6 +526,71 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type // EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED } + #if EIGEN_HAS_CXX11 + /** \brief Construct a row of column vector with fixed size from an arbitrary number of coefficients. \cpp11 + * + * \only_for_vectors + * + * This constructor is for 1D array or vectors with more than 4 coefficients. + * There exists C++98 anologue constructors for fixed-size array/vector having 1, 2, 3, or 4 coefficients. + * + * \warning To construct a column (resp. row) vector of fixed length, the number of values passed to this + * constructor must match the the fixed number of rows (resp. columns) of \c *this. + */ + template <typename... ArgTypes> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + PlainObjectBase(const Scalar& a0, const Scalar& a1, const Scalar& a2, const Scalar& a3, const ArgTypes&... args) + : m_storage() + { + _check_template_params(); + EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, sizeof...(args) + 4); + m_storage.data()[0] = a0; + m_storage.data()[1] = a1; + m_storage.data()[2] = a2; + m_storage.data()[3] = a3; + int i = 4; + auto x = {(m_storage.data()[i++] = args, 0)...}; + static_cast<void>(x); + } + + /** \brief Constructs a Matrix or Array and initializes it by elements given by an initializer list of initializer + * lists \cpp11 + */ + EIGEN_DEVICE_FUNC + explicit EIGEN_STRONG_INLINE PlainObjectBase(const std::initializer_list<std::initializer_list<Scalar>>& list) + : m_storage() + { + _check_template_params(); + + size_t list_size = 0; + if (list.begin() != list.end()) { + list_size = list.begin()->size(); + } + + // This is to allow syntax like VectorXi {{1, 2, 3, 4}} + if (ColsAtCompileTime == 1 && list.size() == 1) { + eigen_assert(list_size == static_cast<size_t>(RowsAtCompileTime) || RowsAtCompileTime == Dynamic); + resize(list_size, ColsAtCompileTime); + std::copy(list.begin()->begin(), list.begin()->end(), m_storage.data()); + } else { + eigen_assert(list.size() == static_cast<size_t>(RowsAtCompileTime) || RowsAtCompileTime == Dynamic); + eigen_assert(list_size == static_cast<size_t>(ColsAtCompileTime) || ColsAtCompileTime == Dynamic); + resize(list.size(), list_size); + + Index row_index = 0; + for (const std::initializer_list<Scalar>& row : list) { + eigen_assert(list_size == row.size()); + Index col_index = 0; + for (const Scalar& e : row) { + coeffRef(row_index, col_index) = e; + ++col_index; + } + ++row_index; + } + } + } + #endif // end EIGEN_HAS_CXX11 + /** \sa PlainObjectBase::operator=(const EigenBase<OtherDerived>&) */ template<typename OtherDerived> EIGEN_DEVICE_FUNC @@ -737,8 +802,10 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void _init2(Index rows, Index cols, typename internal::enable_if<Base::SizeAtCompileTime!=2,T0>::type* = 0) { - EIGEN_STATIC_ASSERT(bool(NumTraits<T0>::IsInteger) && - bool(NumTraits<T1>::IsInteger), + const bool t0_is_integer_alike = internal::is_valid_index_type<T0>::value; + const bool t1_is_integer_alike = internal::is_valid_index_type<T1>::value; + EIGEN_STATIC_ASSERT(t0_is_integer_alike && + t1_is_integer_alike, FLOATING_POINT_ARGUMENT_PASSED__INTEGER_WAS_EXPECTED) resize(rows,cols); } @@ -773,9 +840,9 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type && ((!internal::is_same<typename internal::traits<Derived>::XprKind,ArrayXpr>::value || Base::SizeAtCompileTime==Dynamic)),T>::type* = 0) { // NOTE MSVC 2008 complains if we directly put bool(NumTraits<T>::IsInteger) as the EIGEN_STATIC_ASSERT argument. - const bool is_integer = NumTraits<T>::IsInteger; - EIGEN_UNUSED_VARIABLE(is_integer); - EIGEN_STATIC_ASSERT(is_integer, + const bool is_integer_alike = internal::is_valid_index_type<T>::value; + EIGEN_UNUSED_VARIABLE(is_integer_alike); + EIGEN_STATIC_ASSERT(is_integer_alike, FLOATING_POINT_ARGUMENT_PASSED__INTEGER_WAS_EXPECTED) resize(size); } @@ -882,7 +949,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type * of same type it is enough to swap the data pointers. */ template<typename OtherDerived> - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void swap(DenseBase<OtherDerived> & other) { enum { SwapPointers = internal::is_same<Derived, OtherDerived>::value && Base::SizeAtCompileTime==Dynamic }; @@ -893,7 +960,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type * \brief const version forwarded to DenseBase::swap */ template<typename OtherDerived> - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void swap(DenseBase<OtherDerived> const & other) { Base::swap(other.derived()); } @@ -1027,7 +1094,7 @@ template<typename MatrixTypeA, typename MatrixTypeB, bool SwapPointers> struct matrix_swap_impl { EIGEN_DEVICE_FUNC - static inline void run(MatrixTypeA& a, MatrixTypeB& b) + static EIGEN_STRONG_INLINE void run(MatrixTypeA& a, MatrixTypeB& b) { a.base().swap(b); } diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h index 70790dbd4..13d5662df 100644 --- a/Eigen/src/Core/Product.h +++ b/Eigen/src/Core/Product.h @@ -90,18 +90,23 @@ class Product : public ProductImpl<_Lhs,_Rhs,Option, typedef typename internal::remove_all<LhsNested>::type LhsNestedCleaned; typedef typename internal::remove_all<RhsNested>::type RhsNestedCleaned; - EIGEN_DEVICE_FUNC Product(const Lhs& lhs, const Rhs& rhs) : m_lhs(lhs), m_rhs(rhs) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Product(const Lhs& lhs, const Rhs& rhs) : m_lhs(lhs), m_rhs(rhs) { eigen_assert(lhs.cols() == rhs.rows() && "invalid matrix product" && "if you wanted a coeff-wise or a dot product use the respective explicit functions"); } - 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 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; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const LhsNestedCleaned& lhs() const { return m_lhs; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const RhsNestedCleaned& rhs() const { return m_rhs; } protected: diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index 246bca3e5..d53dc30a3 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -411,35 +411,58 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,CoeffBasedProductMode> call_assignment_no_alias(dst, lhs.lazyProduct(rhs), internal::sub_assign_op<typename Dst::Scalar,Scalar>()); } - // Catch "dst {,+,-}= (s*A)*B" and evaluate it lazily by moving out the scalar factor: - // dst {,+,-}= s * (A.lazyProduct(B)) - // This is a huge benefit for heap-allocated matrix types as it save one costly allocation. - // For them, this strategy is also faster than simply by-passing the heap allocation through - // stack allocation. - // For fixed sizes matrices, this is less obvious, it is sometimes x2 faster, but sometimes x3 slower, - // and the behavior depends also a lot on the compiler... so let's be conservative and enable them for dynamic-size only, - // that is when coming from generic_product_impl<...,GemmProduct> in file GeneralMatrixMatrix.h - template<typename Dst, typename Scalar1, typename Scalar2, typename Plain1, typename Xpr2, typename Func> + // This is a special evaluation path called from generic_product_impl<...,GemmProduct> in file GeneralMatrixMatrix.h + // This variant tries to extract scalar multiples from both the LHS and RHS and factor them out. For instance: + // dst {,+,-}= (s1*A)*(B*s2) + // will be rewritten as: + // dst {,+,-}= (s1*s2) * (A.lazyProduct(B)) + // There are at least four benefits of doing so: + // 1 - huge performance gain for heap-allocated matrix types as it save costly allocations. + // 2 - it is faster than simply by-passing the heap allocation through stack allocation. + // 3 - it makes this fallback consistent with the heavy GEMM routine. + // 4 - it fully by-passes huge stack allocation attempts when multiplying huge fixed-size matrices. + // (see https://stackoverflow.com/questions/54738495) + // For small fixed sizes matrices, howver, the gains are less obvious, it is sometimes x2 faster, but sometimes x3 slower, + // and the behavior depends also a lot on the compiler... This is why this re-writting strategy is currently + // enabled only when falling back from the main GEMM. + template<typename Dst, typename Func> static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE - void eval_dynamic(Dst& dst, const CwiseBinaryOp<internal::scalar_product_op<Scalar1,Scalar2>, - const CwiseNullaryOp<internal::scalar_constant_op<Scalar1>, Plain1>, Xpr2>& lhs, const Rhs& rhs, const Func &func) + void eval_dynamic(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Func &func) { - call_restricted_packet_assignment_no_alias(dst, lhs.lhs().functor().m_other * lhs.rhs().lazyProduct(rhs), func); + enum { + HasScalarFactor = blas_traits<Lhs>::HasScalarFactor || blas_traits<Rhs>::HasScalarFactor, + ConjLhs = blas_traits<Lhs>::NeedToConjugate, + ConjRhs = blas_traits<Rhs>::NeedToConjugate + }; + // FIXME: in c++11 this should be auto, and extractScalarFactor should also return auto + // this is important for real*complex_mat + Scalar actualAlpha = blas_traits<Lhs>::extractScalarFactor(lhs) + * blas_traits<Rhs>::extractScalarFactor(rhs); + eval_dynamic_impl(dst, + blas_traits<Lhs>::extract(lhs).template conjugateIf<ConjLhs>(), + blas_traits<Rhs>::extract(rhs).template conjugateIf<ConjRhs>(), + func, + actualAlpha, + typename conditional<HasScalarFactor,true_type,false_type>::type()); } - // Here, we we always have LhsT==Lhs, but we need to make it a template type to make the above - // overload more specialized. - template<typename Dst, typename LhsT, typename Func> +protected: + + template<typename Dst, typename LhsT, typename RhsT, typename Func, typename Scalar> static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE - void eval_dynamic(Dst& dst, const LhsT& lhs, const Rhs& rhs, const Func &func) + void eval_dynamic_impl(Dst& dst, const LhsT& lhs, const RhsT& rhs, const Func &func, const Scalar& s /* == 1 */, false_type) { + EIGEN_UNUSED_VARIABLE(s); + eigen_internal_assert(s==Scalar(1)); call_restricted_packet_assignment_no_alias(dst, lhs.lazyProduct(rhs), func); } - - -// template<typename Dst> -// static inline void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) -// { dst.noalias() += alpha * lhs.lazyProduct(rhs); } + + template<typename Dst, typename LhsT, typename RhsT, typename Func, typename Scalar> + static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + void eval_dynamic_impl(Dst& dst, const LhsT& lhs, const RhsT& rhs, const Func &func, const Scalar& s, true_type) + { + call_restricted_packet_assignment_no_alias(dst, s * lhs.lazyProduct(rhs), func); + } }; // This specialization enforces the use of a coefficient-based evaluation strategy @@ -582,7 +605,8 @@ struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape, * which is why we don't set the LinearAccessBit. * TODO: this seems possible when the result is a vector */ - EIGEN_DEVICE_FUNC const CoeffReturnType coeff(Index index) const + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const CoeffReturnType coeff(Index index) const { const Index row = (RowsAtCompileTime == 1 || MaxRowsAtCompileTime==1) ? 0 : index; const Index col = (RowsAtCompileTime == 1 || MaxRowsAtCompileTime==1) ? index : 0; @@ -590,6 +614,7 @@ struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape, } template<int LoadMode, typename PacketType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const PacketType packet(Index row, Index col) const { PacketType res; @@ -601,6 +626,7 @@ struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape, } template<int LoadMode, typename PacketType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const PacketType packet(Index index) const { const Index row = (RowsAtCompileTime == 1 || MaxRowsAtCompileTime==1) ? 0 : index; @@ -629,7 +655,8 @@ struct product_evaluator<Product<Lhs, Rhs, DefaultProduct>, LazyCoeffBasedProduc enum { Flags = Base::Flags | EvalBeforeNestingBit }; - EIGEN_DEVICE_FUNC explicit product_evaluator(const XprType& xpr) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit product_evaluator(const XprType& xpr) : Base(BaseProduct(xpr.lhs(),xpr.rhs())) {} }; diff --git a/Eigen/src/Core/Redux.h b/Eigen/src/Core/Redux.h index 720b6030c..2eef5abc5 100644 --- a/Eigen/src/Core/Redux.h +++ b/Eigen/src/Core/Redux.h @@ -359,7 +359,8 @@ class redux_evaluator : public internal::evaluator<_XprType> typedef internal::evaluator<_XprType> Base; public: typedef _XprType XprType; - EIGEN_DEVICE_FUNC explicit redux_evaluator(const XprType &xpr) : Base(xpr) {} + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit redux_evaluator(const XprType &xpr) : Base(xpr) {} typedef typename XprType::Scalar Scalar; typedef typename XprType::CoeffReturnType CoeffReturnType; @@ -375,11 +376,12 @@ public: InnerSizeAtCompileTime = XprType::InnerSizeAtCompileTime }; - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeffByOuterInner(Index outer, Index inner) const { return Base::coeff(IsRowMajor ? outer : inner, IsRowMajor ? inner : outer); } template<int LoadMode, typename PacketType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketType packetByOuterInner(Index outer, Index inner) const { return Base::template packet<LoadMode,PacketType>(IsRowMajor ? outer : inner, IsRowMajor ? inner : outer); } @@ -397,6 +399,8 @@ public: * The template parameter \a BinaryOp is the type of the functor \a func which must be * an associative operator. Both current C++98 and C++11 functor styles are handled. * + * \warning the matrix must be not empty, otherwise an assertion is triggered. + * * \sa DenseBase::sum(), DenseBase::minCoeff(), DenseBase::maxCoeff(), MatrixBase::colwise(), MatrixBase::rowwise() */ template<typename Derived> @@ -415,6 +419,7 @@ DenseBase<Derived>::redux(const Func& func) const } /** \returns the minimum of all coefficients of \c *this. + * \warning the matrix must be not empty, otherwise an assertion is triggered. * \warning the result is undefined if \c *this contains NaN. */ template<typename Derived> @@ -425,6 +430,7 @@ DenseBase<Derived>::minCoeff() const } /** \returns the maximum of all coefficients of \c *this. + * \warning the matrix must be not empty, otherwise an assertion is triggered. * \warning the result is undefined if \c *this contains NaN. */ template<typename Derived> diff --git a/Eigen/src/Core/Ref.h b/Eigen/src/Core/Ref.h index ac9502bc4..172c8ffb6 100644 --- a/Eigen/src/Core/Ref.h +++ b/Eigen/src/Core/Ref.h @@ -28,12 +28,13 @@ struct traits<Ref<_PlainObjectType, _Options, _StrideType> > template<typename Derived> struct match { enum { + IsVectorAtCompileTime = PlainObjectType::IsVectorAtCompileTime || Derived::IsVectorAtCompileTime, HasDirectAccess = internal::has_direct_access<Derived>::ret, - StorageOrderMatch = PlainObjectType::IsVectorAtCompileTime || Derived::IsVectorAtCompileTime || ((PlainObjectType::Flags&RowMajorBit)==(Derived::Flags&RowMajorBit)), + StorageOrderMatch = IsVectorAtCompileTime || ((PlainObjectType::Flags&RowMajorBit)==(Derived::Flags&RowMajorBit)), InnerStrideMatch = int(StrideType::InnerStrideAtCompileTime)==int(Dynamic) || int(StrideType::InnerStrideAtCompileTime)==int(Derived::InnerStrideAtCompileTime) || (int(StrideType::InnerStrideAtCompileTime)==0 && int(Derived::InnerStrideAtCompileTime)==1), - OuterStrideMatch = Derived::IsVectorAtCompileTime + OuterStrideMatch = IsVectorAtCompileTime || int(StrideType::OuterStrideAtCompileTime)==int(Dynamic) || int(StrideType::OuterStrideAtCompileTime)==int(Derived::OuterStrideAtCompileTime), // NOTE, this indirection of evaluator<Derived>::Alignment is needed // to workaround a very strange bug in MSVC related to the instantiation diff --git a/Eigen/src/Core/Reshaped.h b/Eigen/src/Core/Reshaped.h index b7bd1b292..c955815e6 100644 --- a/Eigen/src/Core/Reshaped.h +++ b/Eigen/src/Core/Reshaped.h @@ -191,7 +191,7 @@ class ReshapedImpl_dense<XprType,Rows,Cols,Order,false> /** \returns the nested expression */ EIGEN_DEVICE_FUNC typename internal::remove_reference<XprType>::type& - nestedExpression() { return m_xpr.const_cast_derived(); } + nestedExpression() { return m_xpr; } protected: diff --git a/Eigen/src/Core/Reverse.h b/Eigen/src/Core/Reverse.h index 8b6b3ab03..853093923 100644 --- a/Eigen/src/Core/Reverse.h +++ b/Eigen/src/Core/Reverse.h @@ -171,8 +171,10 @@ struct vectorwise_reverse_inplace_impl<Vertical> template<typename ExpressionType> static void run(ExpressionType &xpr) { + const int HalfAtCompileTime = ExpressionType::RowsAtCompileTime==Dynamic?Dynamic:ExpressionType::RowsAtCompileTime/2; Index half = xpr.rows()/2; - xpr.topRows(half).swap(xpr.bottomRows(half).colwise().reverse()); + xpr.topRows(fix<HalfAtCompileTime>(half)) + .swap(xpr.bottomRows(fix<HalfAtCompileTime>(half)).colwise().reverse()); } }; @@ -182,8 +184,10 @@ struct vectorwise_reverse_inplace_impl<Horizontal> template<typename ExpressionType> static void run(ExpressionType &xpr) { + const int HalfAtCompileTime = ExpressionType::ColsAtCompileTime==Dynamic?Dynamic:ExpressionType::ColsAtCompileTime/2; Index half = xpr.cols()/2; - xpr.leftCols(half).swap(xpr.rightCols(half).rowwise().reverse()); + xpr.leftCols(fix<HalfAtCompileTime>(half)) + .swap(xpr.rightCols(fix<HalfAtCompileTime>(half)).rowwise().reverse()); } }; @@ -203,7 +207,7 @@ struct vectorwise_reverse_inplace_impl<Horizontal> template<typename ExpressionType, int Direction> EIGEN_DEVICE_FUNC void VectorwiseOp<ExpressionType,Direction>::reverseInPlace() { - internal::vectorwise_reverse_inplace_impl<Direction>::run(_expression().const_cast_derived()); + internal::vectorwise_reverse_inplace_impl<Direction>::run(m_matrix); } } // end namespace Eigen diff --git a/Eigen/src/Core/SelfAdjointView.h b/Eigen/src/Core/SelfAdjointView.h index 2cf3fa1ef..2173799d9 100644 --- a/Eigen/src/Core/SelfAdjointView.h +++ b/Eigen/src/Core/SelfAdjointView.h @@ -61,6 +61,7 @@ template<typename _MatrixType, unsigned int UpLo> class SelfAdjointView typedef typename internal::traits<SelfAdjointView>::Scalar Scalar; typedef typename MatrixType::StorageIndex StorageIndex; typedef typename internal::remove_all<typename MatrixType::ConjugateReturnType>::type MatrixConjugateReturnType; + typedef SelfAdjointView<typename internal::add_const<MatrixType>::type, UpLo> ConstSelfAdjointView; enum { Mode = internal::traits<SelfAdjointView>::Mode, @@ -197,6 +198,18 @@ template<typename _MatrixType, unsigned int UpLo> class SelfAdjointView inline const ConjugateReturnType conjugate() const { return ConjugateReturnType(m_matrix.conjugate()); } + /** \returns an expression of the complex conjugate of \c *this if Cond==true, + * returns \c *this otherwise. + */ + template<bool Cond> + EIGEN_DEVICE_FUNC + inline typename internal::conditional<Cond,ConjugateReturnType,ConstSelfAdjointView>::type + conjugateIf() const + { + typedef typename internal::conditional<Cond,ConjugateReturnType,ConstSelfAdjointView>::type ReturnType; + return ReturnType(m_matrix.template conjugateIf<Cond>()); + } + typedef SelfAdjointView<const typename MatrixType::AdjointReturnType,TransposeMode> AdjointReturnType; /** \sa MatrixBase::adjoint() const */ EIGEN_DEVICE_FUNC diff --git a/Eigen/src/Core/Solve.h b/Eigen/src/Core/Solve.h index 2bf940a26..ec4b4a987 100644 --- a/Eigen/src/Core/Solve.h +++ b/Eigen/src/Core/Solve.h @@ -19,7 +19,7 @@ template<typename Decomposition, typename RhsType, typename StorageKind> class S * * \brief Pseudo expression representing a solving operation * - * \tparam Decomposition the type of the matrix or decomposion object + * \tparam Decomposition the type of the matrix or decomposition object * \tparam Rhstype the type of the right-hand side * * This class represents an expression of A.solve(B) diff --git a/Eigen/src/Core/SolverBase.h b/Eigen/src/Core/SolverBase.h index 702a5485c..501461042 100644 --- a/Eigen/src/Core/SolverBase.h +++ b/Eigen/src/Core/SolverBase.h @@ -14,8 +14,35 @@ namespace Eigen { namespace internal { +template<typename Derived> +struct solve_assertion { + template<bool Transpose_, typename Rhs> + static void run(const Derived& solver, const Rhs& b) { solver.template _check_solve_assertion<Transpose_>(b); } +}; + +template<typename Derived> +struct solve_assertion<Transpose<Derived> > +{ + typedef Transpose<Derived> type; + + template<bool Transpose_, typename Rhs> + static void run(const type& transpose, const Rhs& b) + { + internal::solve_assertion<typename internal::remove_all<Derived>::type>::template run<true>(transpose.nestedExpression(), b); + } +}; +template<typename Scalar, typename Derived> +struct solve_assertion<CwiseUnaryOp<Eigen::internal::scalar_conjugate_op<Scalar>, const Transpose<Derived> > > +{ + typedef CwiseUnaryOp<Eigen::internal::scalar_conjugate_op<Scalar>, const Transpose<Derived> > type; + template<bool Transpose_, typename Rhs> + static void run(const type& adjoint, const Rhs& b) + { + internal::solve_assertion<typename internal::remove_all<Transpose<Derived> >::type>::template run<true>(adjoint.nestedExpression(), b); + } +}; } // end namespace internal /** \class SolverBase @@ -35,7 +62,7 @@ namespace internal { * * \warning Currently, any other usage of transpose() and adjoint() are not supported and will produce compilation errors. * - * \sa class PartialPivLU, class FullPivLU + * \sa class PartialPivLU, class FullPivLU, class HouseholderQR, class ColPivHouseholderQR, class FullPivHouseholderQR, class CompleteOrthogonalDecomposition, class LLT, class LDLT, class SVDBase */ template<typename Derived> class SolverBase : public EigenBase<Derived> @@ -46,6 +73,9 @@ class SolverBase : public EigenBase<Derived> typedef typename internal::traits<Derived>::Scalar Scalar; typedef Scalar CoeffReturnType; + template<typename Derived_> + friend struct internal::solve_assertion; + enum { RowsAtCompileTime = internal::traits<Derived>::RowsAtCompileTime, ColsAtCompileTime = internal::traits<Derived>::ColsAtCompileTime, @@ -75,7 +105,7 @@ class SolverBase : public EigenBase<Derived> inline const Solve<Derived, Rhs> solve(const MatrixBase<Rhs>& b) const { - eigen_assert(derived().rows()==b.rows() && "solve(): invalid number of rows of the right hand side matrix b"); + internal::solve_assertion<typename internal::remove_all<Derived>::type>::template run<false>(derived(), b); return Solve<Derived, Rhs>(derived(), b.derived()); } @@ -113,6 +143,13 @@ class SolverBase : public EigenBase<Derived> } protected: + + template<bool Transpose_, typename Rhs> + void _check_solve_assertion(const Rhs& b) const { + EIGEN_ONLY_USED_FOR_DEBUG(b); + eigen_assert(derived().m_isInitialized && "Solver is not initialized."); + eigen_assert((Transpose_?derived().cols():derived().rows())==b.rows() && "SolverBase::solve(): invalid number of rows of the right hand side matrix b"); + } }; namespace internal { diff --git a/Eigen/src/Core/Swap.h b/Eigen/src/Core/Swap.h index d70200918..180a4e5ad 100644 --- a/Eigen/src/Core/Swap.h +++ b/Eigen/src/Core/Swap.h @@ -30,12 +30,13 @@ public: typedef typename Base::DstXprType DstXprType; typedef swap_assign_op<Scalar> Functor; - EIGEN_DEVICE_FUNC generic_dense_assignment_kernel(DstEvaluatorTypeT &dst, const SrcEvaluatorTypeT &src, const Functor &func, DstXprType& dstExpr) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + generic_dense_assignment_kernel(DstEvaluatorTypeT &dst, const SrcEvaluatorTypeT &src, const Functor &func, DstXprType& dstExpr) : Base(dst, src, func, dstExpr) {} template<int StoreMode, int LoadMode, typename PacketType> - void assignPacket(Index row, Index col) + EIGEN_STRONG_INLINE void assignPacket(Index row, Index col) { PacketType tmp = m_src.template packet<LoadMode,PacketType>(row,col); const_cast<SrcEvaluatorTypeT&>(m_src).template writePacket<LoadMode>(row,col, m_dst.template packet<StoreMode,PacketType>(row,col)); @@ -43,7 +44,7 @@ public: } template<int StoreMode, int LoadMode, typename PacketType> - void assignPacket(Index index) + EIGEN_STRONG_INLINE void assignPacket(Index index) { PacketType tmp = m_src.template packet<LoadMode,PacketType>(index); const_cast<SrcEvaluatorTypeT&>(m_src).template writePacket<LoadMode>(index, m_dst.template packet<StoreMode,PacketType>(index)); @@ -52,7 +53,7 @@ public: // TODO find a simple way not to have to copy/paste this function from generic_dense_assignment_kernel, by simple I mean no CRTP (Gael) template<int StoreMode, int LoadMode, typename PacketType> - void assignPacketByOuterInner(Index outer, Index inner) + EIGEN_STRONG_INLINE void assignPacketByOuterInner(Index outer, Index inner) { Index row = Base::rowIndexByOuterInner(outer, inner); Index col = Base::colIndexByOuterInner(outer, inner); diff --git a/Eigen/src/Core/Transpose.h b/Eigen/src/Core/Transpose.h index d7c204579..c513f7f7c 100644 --- a/Eigen/src/Core/Transpose.h +++ b/Eigen/src/Core/Transpose.h @@ -61,25 +61,27 @@ template<typename MatrixType> class Transpose typedef typename internal::remove_all<MatrixType>::type NestedExpression; EIGEN_DEVICE_FUNC - explicit inline Transpose(MatrixType& matrix) : m_matrix(matrix) {} + explicit EIGEN_STRONG_INLINE Transpose(MatrixType& matrix) : m_matrix(matrix) {} EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Transpose) - EIGEN_DEVICE_FUNC inline Index rows() const { return m_matrix.cols(); } - EIGEN_DEVICE_FUNC inline Index cols() const { return m_matrix.rows(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Index rows() const { return m_matrix.cols(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Index cols() const { return m_matrix.rows(); } /** \returns the nested expression */ - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename internal::remove_all<MatrixTypeNested>::type& nestedExpression() const { return m_matrix; } /** \returns the nested expression */ - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename internal::remove_reference<MatrixTypeNested>::type& nestedExpression() { return m_matrix; } /** \internal */ - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void resize(Index nrows, Index ncols) { m_matrix.resize(ncols,nrows); } @@ -123,8 +125,10 @@ template<typename MatrixType> class TransposeImpl<MatrixType,Dense> EIGEN_DENSE_PUBLIC_INTERFACE(Transpose<MatrixType>) EIGEN_INHERIT_ASSIGNMENT_OPERATORS(TransposeImpl) - EIGEN_DEVICE_FUNC inline Index innerStride() const { return derived().nestedExpression().innerStride(); } - EIGEN_DEVICE_FUNC inline Index outerStride() const { return derived().nestedExpression().outerStride(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Index innerStride() const { return derived().nestedExpression().innerStride(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Index outerStride() const { return derived().nestedExpression().outerStride(); } typedef typename internal::conditional< internal::is_lvalue<MatrixType>::value, @@ -132,18 +136,20 @@ template<typename MatrixType> class TransposeImpl<MatrixType,Dense> const Scalar >::type ScalarWithConstIfNotLvalue; - EIGEN_DEVICE_FUNC inline ScalarWithConstIfNotLvalue* data() { return derived().nestedExpression().data(); } - EIGEN_DEVICE_FUNC inline const Scalar* data() const { return derived().nestedExpression().data(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + ScalarWithConstIfNotLvalue* data() { return derived().nestedExpression().data(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const Scalar* data() const { return derived().nestedExpression().data(); } // FIXME: shall we keep the const version of coeffRef? - EIGEN_DEVICE_FUNC - inline const Scalar& coeffRef(Index rowId, Index colId) const + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const Scalar& coeffRef(Index rowId, Index colId) const { return derived().nestedExpression().coeffRef(colId, rowId); } - EIGEN_DEVICE_FUNC - inline const Scalar& coeffRef(Index index) const + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const Scalar& coeffRef(Index index) const { return derived().nestedExpression().coeffRef(index); } @@ -169,7 +175,8 @@ template<typename MatrixType> class TransposeImpl<MatrixType,Dense> * * \sa transposeInPlace(), adjoint() */ template<typename Derived> -EIGEN_DEVICE_FUNC inline Transpose<Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +Transpose<Derived> DenseBase<Derived>::transpose() { return TransposeReturnType(derived()); @@ -181,7 +188,8 @@ DenseBase<Derived>::transpose() * * \sa transposeInPlace(), adjoint() */ template<typename Derived> -EIGEN_DEVICE_FUNC inline typename DenseBase<Derived>::ConstTransposeReturnType +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +typename DenseBase<Derived>::ConstTransposeReturnType DenseBase<Derived>::transpose() const { return ConstTransposeReturnType(derived()); @@ -392,7 +400,8 @@ struct checkTransposeAliasing_impl<Derived, OtherDerived, false> template<typename Dst, typename Src> void check_for_aliasing(const Dst &dst, const Src &src) { - internal::checkTransposeAliasing_impl<Dst, Src>::run(dst, src); + if((!Dst::IsVectorAtCompileTime) && dst.rows()>1 && dst.cols()>1) + internal::checkTransposeAliasing_impl<Dst, Src>::run(dst, src); } } // end namespace internal diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index 521de6160..cf3532f06 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -198,6 +198,7 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularView typedef typename internal::traits<TriangularView>::MatrixTypeNestedNonRef MatrixTypeNestedNonRef; typedef typename internal::remove_all<typename MatrixType::ConjugateReturnType>::type MatrixConjugateReturnType; + typedef TriangularView<typename internal::add_const<MatrixType>::type, _Mode> ConstTriangularView; public: @@ -243,6 +244,18 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularView inline const ConjugateReturnType conjugate() const { return ConjugateReturnType(m_matrix.conjugate()); } + /** \returns an expression of the complex conjugate of \c *this if Cond==true, + * returns \c *this otherwise. + */ + template<bool Cond> + EIGEN_DEVICE_FUNC + inline typename internal::conditional<Cond,ConjugateReturnType,ConstTriangularView>::type + conjugateIf() const + { + typedef typename internal::conditional<Cond,ConjugateReturnType,ConstTriangularView>::type ReturnType; + return ReturnType(m_matrix.template conjugateIf<Cond>()); + } + typedef TriangularView<const typename MatrixType::AdjointReturnType,TransposeMode> AdjointReturnType; /** \sa MatrixBase::adjoint() const */ EIGEN_DEVICE_FUNC diff --git a/Eigen/src/Core/VectorBlock.h b/Eigen/src/Core/VectorBlock.h index 0ede5d58e..71c5b95ee 100644 --- a/Eigen/src/Core/VectorBlock.h +++ b/Eigen/src/Core/VectorBlock.h @@ -71,8 +71,8 @@ template<typename VectorType, int Size> class VectorBlock /** Dynamic-size constructor */ - EIGEN_DEVICE_FUNC - inline VectorBlock(VectorType& vector, Index start, Index size) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + VectorBlock(VectorType& vector, Index start, Index size) : Base(vector, IsColVector ? start : 0, IsColVector ? 0 : start, IsColVector ? size : 1, IsColVector ? 1 : size) @@ -82,8 +82,8 @@ template<typename VectorType, int Size> class VectorBlock /** Fixed-size constructor */ - EIGEN_DEVICE_FUNC - inline VectorBlock(VectorType& vector, Index start) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + VectorBlock(VectorType& vector, Index start) : Base(vector, IsColVector ? start : 0, IsColVector ? 0 : start) { EIGEN_STATIC_ASSERT_VECTOR_ONLY(VectorBlock); diff --git a/Eigen/src/Core/VectorwiseOp.h b/Eigen/src/Core/VectorwiseOp.h index a88b6e736..db0b9f8c4 100644 --- a/Eigen/src/Core/VectorwiseOp.h +++ b/Eigen/src/Core/VectorwiseOp.h @@ -173,6 +173,14 @@ struct member_redux { * Example: \include MatrixBase_colwise_iterator_cxx11.cpp * Output: \verbinclude MatrixBase_colwise_iterator_cxx11.out * + * For a partial reduction on an empty input, some rules apply. + * For the sake of clarity, let's consider a vertical reduction: + * - If the number of columns is zero, then a 1x0 row-major vector expression is returned. + * - Otherwise, if the number of rows is zero, then + * - a row vector of zeros is returned for sum-like reductions (sum, squaredNorm, norm, etc.) + * - a row vector of ones is returned for a product reduction (e.g., <code>MatrixXd(n,0).colwise().prod()</code>) + * - an assert is triggered for all other reductions (minCoeff,maxCoeff,redux(bin_op)) + * * \sa DenseBase::colwise(), DenseBase::rowwise(), class PartialReduxExpr */ template<typename ExpressionType, int Direction> class VectorwiseOp @@ -294,13 +302,19 @@ template<typename ExpressionType, int Direction> class VectorwiseOp * The template parameter \a BinaryOp is the type of the functor * of the custom redux operator. Note that func must be an associative operator. * + * \warning the size along the reduction direction must be strictly positive, + * otherwise an assertion is triggered. + * * \sa class VectorwiseOp, DenseBase::colwise(), DenseBase::rowwise() */ template<typename BinaryOp> EIGEN_DEVICE_FUNC const typename ReduxReturnType<BinaryOp>::Type redux(const BinaryOp& func = BinaryOp()) const - { return typename ReduxReturnType<BinaryOp>::Type(_expression(), internal::member_redux<BinaryOp,Scalar>(func)); } + { + eigen_assert(redux_length()>0 && "you are using an empty matrix"); + return typename ReduxReturnType<BinaryOp>::Type(_expression(), internal::member_redux<BinaryOp,Scalar>(func)); + } typedef typename ReturnType<internal::member_minCoeff>::Type MinCoeffReturnType; typedef typename ReturnType<internal::member_maxCoeff>::Type MaxCoeffReturnType; @@ -325,6 +339,9 @@ template<typename ExpressionType, int Direction> class VectorwiseOp /** \returns a row (or column) vector expression of the smallest coefficient * of each column (or row) of the referenced expression. * + * \warning the size along the reduction direction must be strictly positive, + * otherwise an assertion is triggered. + * * \warning the result is undefined if \c *this contains NaN. * * Example: \include PartialRedux_minCoeff.cpp @@ -333,11 +350,17 @@ template<typename ExpressionType, int Direction> class VectorwiseOp * \sa DenseBase::minCoeff() */ EIGEN_DEVICE_FUNC const MinCoeffReturnType minCoeff() const - { return MinCoeffReturnType(_expression()); } + { + eigen_assert(redux_length()>0 && "you are using an empty matrix"); + return MinCoeffReturnType(_expression()); + } /** \returns a row (or column) vector expression of the largest coefficient * of each column (or row) of the referenced expression. * + * \warning the size along the reduction direction must be strictly positive, + * otherwise an assertion is triggered. + * * \warning the result is undefined if \c *this contains NaN. * * Example: \include PartialRedux_maxCoeff.cpp @@ -346,7 +369,10 @@ template<typename ExpressionType, int Direction> class VectorwiseOp * \sa DenseBase::maxCoeff() */ EIGEN_DEVICE_FUNC const MaxCoeffReturnType maxCoeff() const - { return MaxCoeffReturnType(_expression()); } + { + eigen_assert(redux_length()>0 && "you are using an empty matrix"); + return MaxCoeffReturnType(_expression()); + } /** \returns a row (or column) vector expression of the squared norm * of each column (or row) of the referenced expression. @@ -531,7 +557,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived) EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived) //eigen_assert((m_matrix.isNull()) == (other.isNull())); FIXME - return const_cast<ExpressionType&>(m_matrix = extendedTo(other.derived())); + return m_matrix = extendedTo(other.derived()); } /** Adds the vector \a other to each subvector of \c *this */ @@ -541,7 +567,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp { EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived) EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived) - return const_cast<ExpressionType&>(m_matrix += extendedTo(other.derived())); + return m_matrix += extendedTo(other.derived()); } /** Substracts the vector \a other to each subvector of \c *this */ @@ -551,7 +577,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp { EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived) EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived) - return const_cast<ExpressionType&>(m_matrix -= extendedTo(other.derived())); + return m_matrix -= extendedTo(other.derived()); } /** Multiples each subvector of \c *this by the vector \a other */ @@ -563,7 +589,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp EIGEN_STATIC_ASSERT_ARRAYXPR(ExpressionType) EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived) m_matrix *= extendedTo(other.derived()); - return const_cast<ExpressionType&>(m_matrix); + return m_matrix; } /** Divides each subvector of \c *this by the vector \a other */ @@ -575,7 +601,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp EIGEN_STATIC_ASSERT_ARRAYXPR(ExpressionType) EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived) m_matrix /= extendedTo(other.derived()); - return const_cast<ExpressionType&>(m_matrix); + return m_matrix; } /** Returns the expression of the sum of the vector \a other to each subvector of \c *this */ @@ -690,6 +716,10 @@ template<typename ExpressionType, int Direction> class VectorwiseOp const HNormalizedReturnType hnormalized() const; protected: + Index redux_length() const + { + return Direction==Vertical ? m_matrix.rows() : m_matrix.cols(); + } ExpressionTypeNested m_matrix; }; diff --git a/Eigen/src/Core/Visitor.h b/Eigen/src/Core/Visitor.h index 54c1883d9..f67d83bd1 100644 --- a/Eigen/src/Core/Visitor.h +++ b/Eigen/src/Core/Visitor.h @@ -40,6 +40,14 @@ struct visitor_impl<Visitor, Derived, 1> } }; +// This specialization enables visitors on empty matrices at compile-time +template<typename Visitor, typename Derived> +struct visitor_impl<Visitor, Derived, 0> { + EIGEN_DEVICE_FUNC + static inline void run(const Derived &/*mat*/, Visitor& /*visitor*/) + {} +}; + template<typename Visitor, typename Derived> struct visitor_impl<Visitor, Derived, Dynamic> { @@ -98,6 +106,8 @@ protected: * * \note compared to one or two \em for \em loops, visitors offer automatic * unrolling for small fixed size matrix. + * + * \note if the matrix is empty, then the visitor is left unchanged. * * \sa minCoeff(Index*,Index*), maxCoeff(Index*,Index*), DenseBase::redux() */ @@ -106,6 +116,9 @@ template<typename Visitor> EIGEN_DEVICE_FUNC void DenseBase<Derived>::visit(Visitor& visitor) const { + if(size()==0) + return; + typedef typename internal::visitor_evaluator<Derived> ThisEvaluator; ThisEvaluator thisEval(derived()); @@ -124,6 +137,8 @@ namespace internal { template <typename Derived> struct coeff_visitor { + // default initialization to avoid countless invalid maybe-uninitialized warnings by gcc + coeff_visitor() : row(-1), col(-1), res(0) {} typedef typename Derived::Scalar Scalar; Index row, col; Scalar res; @@ -196,6 +211,9 @@ struct functor_traits<max_coeff_visitor<Scalar> > { /** \fn DenseBase<Derived>::minCoeff(IndexType* rowId, IndexType* colId) const * \returns the minimum of all coefficients of *this and puts in *row and *col its location. + * + * \warning the matrix must be not empty, otherwise an assertion is triggered. + * * \warning the result is undefined if \c *this contains NaN. * * \sa DenseBase::minCoeff(Index*), DenseBase::maxCoeff(Index*,Index*), DenseBase::visit(), DenseBase::minCoeff() @@ -206,6 +224,8 @@ EIGEN_DEVICE_FUNC typename internal::traits<Derived>::Scalar DenseBase<Derived>::minCoeff(IndexType* rowId, IndexType* colId) const { + eigen_assert(this->rows()>0 && this->cols()>0 && "you are using an empty matrix"); + internal::min_coeff_visitor<Derived> minVisitor; this->visit(minVisitor); *rowId = minVisitor.row; @@ -214,6 +234,9 @@ DenseBase<Derived>::minCoeff(IndexType* rowId, IndexType* colId) const } /** \returns the minimum of all coefficients of *this and puts in *index its location. + * + * \warning the matrix must be not empty, otherwise an assertion is triggered. + * * \warning the result is undefined if \c *this contains NaN. * * \sa DenseBase::minCoeff(IndexType*,IndexType*), DenseBase::maxCoeff(IndexType*,IndexType*), DenseBase::visit(), DenseBase::minCoeff() @@ -224,6 +247,8 @@ EIGEN_DEVICE_FUNC typename internal::traits<Derived>::Scalar DenseBase<Derived>::minCoeff(IndexType* index) const { + eigen_assert(this->rows()>0 && this->cols()>0 && "you are using an empty matrix"); + EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) internal::min_coeff_visitor<Derived> minVisitor; this->visit(minVisitor); @@ -233,6 +258,9 @@ DenseBase<Derived>::minCoeff(IndexType* index) const /** \fn DenseBase<Derived>::maxCoeff(IndexType* rowId, IndexType* colId) const * \returns the maximum of all coefficients of *this and puts in *row and *col its location. + * + * \warning the matrix must be not empty, otherwise an assertion is triggered. + * * \warning the result is undefined if \c *this contains NaN. * * \sa DenseBase::minCoeff(IndexType*,IndexType*), DenseBase::visit(), DenseBase::maxCoeff() @@ -243,6 +271,8 @@ EIGEN_DEVICE_FUNC typename internal::traits<Derived>::Scalar DenseBase<Derived>::maxCoeff(IndexType* rowPtr, IndexType* colPtr) const { + eigen_assert(this->rows()>0 && this->cols()>0 && "you are using an empty matrix"); + internal::max_coeff_visitor<Derived> maxVisitor; this->visit(maxVisitor); *rowPtr = maxVisitor.row; @@ -251,6 +281,9 @@ DenseBase<Derived>::maxCoeff(IndexType* rowPtr, IndexType* colPtr) const } /** \returns the maximum of all coefficients of *this and puts in *index its location. + * + * \warning the matrix must be not empty, otherwise an assertion is triggered. + * * \warning the result is undefined if \c *this contains NaN. * * \sa DenseBase::maxCoeff(IndexType*,IndexType*), DenseBase::minCoeff(IndexType*,IndexType*), DenseBase::visitor(), DenseBase::maxCoeff() @@ -261,6 +294,8 @@ EIGEN_DEVICE_FUNC typename internal::traits<Derived>::Scalar DenseBase<Derived>::maxCoeff(IndexType* index) const { + eigen_assert(this->rows()>0 && this->cols()>0 && "you are using an empty matrix"); + EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) internal::max_coeff_visitor<Derived> maxVisitor; this->visit(maxVisitor); diff --git a/Eigen/src/Core/arch/AVX/Complex.h b/Eigen/src/Core/arch/AVX/Complex.h index 7fa61969d..5b8ff59bd 100644 --- a/Eigen/src/Core/arch/AVX/Complex.h +++ b/Eigen/src/Core/arch/AVX/Complex.h @@ -22,6 +22,7 @@ struct Packet4cf __m256 v; }; +#ifndef EIGEN_VECTORIZE_AVX512 template<> struct packet_traits<std::complex<float> > : default_packet_traits { typedef Packet4cf type; @@ -44,8 +45,9 @@ template<> struct packet_traits<std::complex<float> > : default_packet_traits HasSetLinear = 0 }; }; +#endif -template<> struct unpacket_traits<Packet4cf> { typedef std::complex<float> type; enum {size=4, alignment=Aligned32}; typedef Packet2cf half; }; +template<> struct unpacket_traits<Packet4cf> { typedef std::complex<float> type; enum {size=4, alignment=Aligned32, vectorizable=true}; typedef Packet2cf half; }; template<> EIGEN_STRONG_INLINE Packet4cf padd<Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_add_ps(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet4cf psub<Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_sub_ps(a.v,b.v)); } @@ -67,10 +69,18 @@ template<> EIGEN_STRONG_INLINE Packet4cf pmul<Packet4cf>(const Packet4cf& a, con return Packet4cf(result); } +template <> +EIGEN_STRONG_INLINE Packet4cf pcmp_eq(const Packet4cf& a, const Packet4cf& b) { + __m256 eq = _mm256_cmp_ps(a.v, b.v, _CMP_EQ_OQ); + return Packet4cf(_mm256_and_ps(eq, _mm256_permute_ps(eq, 0xb1))); +} + +template<> EIGEN_STRONG_INLINE Packet4cf ptrue<Packet4cf>(const Packet4cf& a) { return Packet4cf(ptrue(Packet8f(a.v))); } +template<> EIGEN_STRONG_INLINE Packet4cf pnot<Packet4cf>(const Packet4cf& a) { return Packet4cf(pnot(Packet8f(a.v))); } template<> EIGEN_STRONG_INLINE Packet4cf pand <Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_and_ps(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet4cf por <Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_or_ps(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet4cf pxor <Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_xor_ps(a.v,b.v)); } -template<> EIGEN_STRONG_INLINE Packet4cf pandnot<Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_andnot_ps(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet4cf pandnot<Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_andnot_ps(b.v,a.v)); } template<> EIGEN_STRONG_INLINE Packet4cf pload <Packet4cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet4cf(pload<Packet8f>(&numext::real_ref(*from))); } template<> EIGEN_STRONG_INLINE Packet4cf ploadu<Packet4cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet4cf(ploadu<Packet8f>(&numext::real_ref(*from))); } @@ -228,6 +238,7 @@ struct Packet2cd __m256d v; }; +#ifndef EIGEN_VECTORIZE_AVX512 template<> struct packet_traits<std::complex<double> > : default_packet_traits { typedef Packet2cd type; @@ -250,8 +261,9 @@ template<> struct packet_traits<std::complex<double> > : default_packet_traits HasSetLinear = 0 }; }; +#endif -template<> struct unpacket_traits<Packet2cd> { typedef std::complex<double> type; enum {size=2, alignment=Aligned32}; typedef Packet1cd half; }; +template<> struct unpacket_traits<Packet2cd> { typedef std::complex<double> type; enum {size=2, alignment=Aligned32, vectorizable=true}; typedef Packet1cd half; }; template<> EIGEN_STRONG_INLINE Packet2cd padd<Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_add_pd(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet2cd psub<Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_sub_pd(a.v,b.v)); } @@ -272,10 +284,18 @@ template<> EIGEN_STRONG_INLINE Packet2cd pmul<Packet2cd>(const Packet2cd& a, con return Packet2cd(_mm256_addsub_pd(even, odd)); } +template <> +EIGEN_STRONG_INLINE Packet2cd pcmp_eq(const Packet2cd& a, const Packet2cd& b) { + __m256d eq = _mm256_cmp_pd(a.v, b.v, _CMP_EQ_OQ); + return Packet2cd(pand(eq, _mm256_permute_pd(eq, 0x5))); +} + +template<> EIGEN_STRONG_INLINE Packet2cd ptrue<Packet2cd>(const Packet2cd& a) { return Packet2cd(ptrue(Packet4d(a.v))); } +template<> EIGEN_STRONG_INLINE Packet2cd pnot<Packet2cd>(const Packet2cd& a) { return Packet2cd(pnot(Packet4d(a.v))); } template<> EIGEN_STRONG_INLINE Packet2cd pand <Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_and_pd(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet2cd por <Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_or_pd(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet2cd pxor <Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_xor_pd(a.v,b.v)); } -template<> EIGEN_STRONG_INLINE Packet2cd pandnot<Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_andnot_pd(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cd pandnot<Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_andnot_pd(b.v,a.v)); } template<> EIGEN_STRONG_INLINE Packet2cd pload <Packet2cd>(const std::complex<double>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cd(pload<Packet4d>((const double*)from)); } diff --git a/Eigen/src/Core/arch/AVX/MathFunctions.h b/Eigen/src/Core/arch/AVX/MathFunctions.h index 6af67ce2d..9f375ed98 100644 --- a/Eigen/src/Core/arch/AVX/MathFunctions.h +++ b/Eigen/src/Core/arch/AVX/MathFunctions.h @@ -10,7 +10,7 @@ #ifndef EIGEN_MATH_FUNCTIONS_AVX_H #define EIGEN_MATH_FUNCTIONS_AVX_H -/* The sin, cos, exp, and log functions of this file are loosely derived from +/* The sin and cos functions of this file are loosely derived from * Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/ */ @@ -18,187 +18,22 @@ namespace Eigen { namespace internal { -inline Packet8i pshiftleft(Packet8i v, int n) -{ -#ifdef EIGEN_VECTORIZE_AVX2 - return _mm256_slli_epi32(v, n); -#else - __m128i lo = _mm_slli_epi32(_mm256_extractf128_si256(v, 0), n); - __m128i hi = _mm_slli_epi32(_mm256_extractf128_si256(v, 1), n); - return _mm256_insertf128_si256(_mm256_castsi128_si256(lo), (hi), 1); -#endif -} - -inline Packet8f pshiftright(Packet8f v, int n) -{ -#ifdef EIGEN_VECTORIZE_AVX2 - return _mm256_cvtepi32_ps(_mm256_srli_epi32(_mm256_castps_si256(v), n)); -#else - __m128i lo = _mm_srli_epi32(_mm256_extractf128_si256(_mm256_castps_si256(v), 0), n); - __m128i hi = _mm_srli_epi32(_mm256_extractf128_si256(_mm256_castps_si256(v), 1), n); - return _mm256_cvtepi32_ps(_mm256_insertf128_si256(_mm256_castsi128_si256(lo), (hi), 1)); -#endif -} - -// Sine function -// Computes sin(x) by wrapping x to the interval [-Pi/4,3*Pi/4] and -// evaluating interpolants in [-Pi/4,Pi/4] or [Pi/4,3*Pi/4]. The interpolants -// are (anti-)symmetric and thus have only odd/even coefficients template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f psin<Packet8f>(const Packet8f& _x) { - Packet8f x = _x; - - // Some useful values. - _EIGEN_DECLARE_CONST_Packet8i(one, 1); - _EIGEN_DECLARE_CONST_Packet8f(one, 1.0f); - _EIGEN_DECLARE_CONST_Packet8f(two, 2.0f); - _EIGEN_DECLARE_CONST_Packet8f(one_over_four, 0.25f); - _EIGEN_DECLARE_CONST_Packet8f(one_over_pi, 3.183098861837907e-01f); - _EIGEN_DECLARE_CONST_Packet8f(neg_pi_first, -3.140625000000000e+00f); - _EIGEN_DECLARE_CONST_Packet8f(neg_pi_second, -9.670257568359375e-04f); - _EIGEN_DECLARE_CONST_Packet8f(neg_pi_third, -6.278329571784980e-07f); - _EIGEN_DECLARE_CONST_Packet8f(four_over_pi, 1.273239544735163e+00f); - - // Map x from [-Pi/4,3*Pi/4] to z in [-1,3] and subtract the shifted period. - Packet8f z = pmul(x, p8f_one_over_pi); - Packet8f shift = _mm256_floor_ps(padd(z, p8f_one_over_four)); - x = pmadd(shift, p8f_neg_pi_first, x); - x = pmadd(shift, p8f_neg_pi_second, x); - x = pmadd(shift, p8f_neg_pi_third, x); - z = pmul(x, p8f_four_over_pi); - - // Make a mask for the entries that need flipping, i.e. wherever the shift - // is odd. - Packet8i shift_ints = _mm256_cvtps_epi32(shift); - Packet8i shift_isodd = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(shift_ints), _mm256_castsi256_ps(p8i_one))); - Packet8i sign_flip_mask = pshiftleft(shift_isodd, 31); - - // Create a mask for which interpolant to use, i.e. if z > 1, then the mask - // is set to ones for that entry. - Packet8f ival_mask = _mm256_cmp_ps(z, p8f_one, _CMP_GT_OQ); - - // Evaluate the polynomial for the interval [1,3] in z. - _EIGEN_DECLARE_CONST_Packet8f(coeff_right_0, 9.999999724233232e-01f); - _EIGEN_DECLARE_CONST_Packet8f(coeff_right_2, -3.084242535619928e-01f); - _EIGEN_DECLARE_CONST_Packet8f(coeff_right_4, 1.584991525700324e-02f); - _EIGEN_DECLARE_CONST_Packet8f(coeff_right_6, -3.188805084631342e-04f); - Packet8f z_minus_two = psub(z, p8f_two); - Packet8f z_minus_two2 = pmul(z_minus_two, z_minus_two); - Packet8f right = pmadd(p8f_coeff_right_6, z_minus_two2, p8f_coeff_right_4); - right = pmadd(right, z_minus_two2, p8f_coeff_right_2); - right = pmadd(right, z_minus_two2, p8f_coeff_right_0); - - // Evaluate the polynomial for the interval [-1,1] in z. - _EIGEN_DECLARE_CONST_Packet8f(coeff_left_1, 7.853981525427295e-01f); - _EIGEN_DECLARE_CONST_Packet8f(coeff_left_3, -8.074536727092352e-02f); - _EIGEN_DECLARE_CONST_Packet8f(coeff_left_5, 2.489871967827018e-03f); - _EIGEN_DECLARE_CONST_Packet8f(coeff_left_7, -3.587725841214251e-05f); - Packet8f z2 = pmul(z, z); - Packet8f left = pmadd(p8f_coeff_left_7, z2, p8f_coeff_left_5); - left = pmadd(left, z2, p8f_coeff_left_3); - left = pmadd(left, z2, p8f_coeff_left_1); - left = pmul(left, z); - - // Assemble the results, i.e. select the left and right polynomials. - left = _mm256_andnot_ps(ival_mask, left); - right = _mm256_and_ps(ival_mask, right); - Packet8f res = _mm256_or_ps(left, right); + return psin_float(_x); +} - // Flip the sign on the odd intervals and return the result. - res = _mm256_xor_ps(res, _mm256_castsi256_ps(sign_flip_mask)); - return res; +template <> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f +pcos<Packet8f>(const Packet8f& _x) { + return pcos_float(_x); } -// Natural logarithm -// Computes log(x) as log(2^e * m) = C*e + log(m), where the constant C =log(2) -// and m is in the range [sqrt(1/2),sqrt(2)). In this range, the logarithm can -// be easily approximated by a polynomial centered on m=1 for stability. -// TODO(gonnet): Further reduce the interval allowing for lower-degree -// polynomial interpolants -> ... -> profit! template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f plog<Packet8f>(const Packet8f& _x) { - Packet8f x = _x; - _EIGEN_DECLARE_CONST_Packet8f(1, 1.0f); - _EIGEN_DECLARE_CONST_Packet8f(half, 0.5f); - _EIGEN_DECLARE_CONST_Packet8f(126f, 126.0f); - - _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(inv_mant_mask, ~0x7f800000); - - // The smallest non denormalized float number. - _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(min_norm_pos, 0x00800000); - _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(minus_inf, 0xff800000); - - // Polynomial coefficients. - _EIGEN_DECLARE_CONST_Packet8f(cephes_SQRTHF, 0.707106781186547524f); - _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p0, 7.0376836292E-2f); - _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p1, -1.1514610310E-1f); - _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p2, 1.1676998740E-1f); - _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p3, -1.2420140846E-1f); - _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p4, +1.4249322787E-1f); - _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p5, -1.6668057665E-1f); - _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p6, +2.0000714765E-1f); - _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p7, -2.4999993993E-1f); - _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p8, +3.3333331174E-1f); - _EIGEN_DECLARE_CONST_Packet8f(cephes_log_q1, -2.12194440e-4f); - _EIGEN_DECLARE_CONST_Packet8f(cephes_log_q2, 0.693359375f); - - Packet8f invalid_mask = _mm256_cmp_ps(x, _mm256_setzero_ps(), _CMP_NGE_UQ); // not greater equal is true if x is NaN - Packet8f iszero_mask = _mm256_cmp_ps(x, _mm256_setzero_ps(), _CMP_EQ_OQ); - - // Truncate input values to the minimum positive normal. - x = pmax(x, p8f_min_norm_pos); - - Packet8f emm0 = pshiftright(x,23); - Packet8f e = _mm256_sub_ps(emm0, p8f_126f); - - // Set the exponents to -1, i.e. x are in the range [0.5,1). - x = _mm256_and_ps(x, p8f_inv_mant_mask); - x = _mm256_or_ps(x, p8f_half); - - // part2: Shift the inputs from the range [0.5,1) to [sqrt(1/2),sqrt(2)) - // and shift by -1. The values are then centered around 0, which improves - // the stability of the polynomial evaluation. - // if( x < SQRTHF ) { - // e -= 1; - // x = x + x - 1.0; - // } else { x = x - 1.0; } - Packet8f mask = _mm256_cmp_ps(x, p8f_cephes_SQRTHF, _CMP_LT_OQ); - Packet8f tmp = _mm256_and_ps(x, mask); - x = psub(x, p8f_1); - e = psub(e, _mm256_and_ps(p8f_1, mask)); - x = padd(x, tmp); - - Packet8f x2 = pmul(x, x); - Packet8f x3 = pmul(x2, x); - - // Evaluate the polynomial approximant of degree 8 in three parts, probably - // to improve instruction-level parallelism. - Packet8f y, y1, y2; - y = pmadd(p8f_cephes_log_p0, x, p8f_cephes_log_p1); - y1 = pmadd(p8f_cephes_log_p3, x, p8f_cephes_log_p4); - y2 = pmadd(p8f_cephes_log_p6, x, p8f_cephes_log_p7); - y = pmadd(y, x, p8f_cephes_log_p2); - y1 = pmadd(y1, x, p8f_cephes_log_p5); - y2 = pmadd(y2, x, p8f_cephes_log_p8); - y = pmadd(y, x3, y1); - y = pmadd(y, x3, y2); - y = pmul(y, x3); - - // Add the logarithm of the exponent back to the result of the interpolation. - y1 = pmul(e, p8f_cephes_log_q1); - tmp = pmul(x2, p8f_half); - y = padd(y, y1); - x = psub(x, tmp); - y2 = pmul(e, p8f_cephes_log_q2); - x = padd(x, y); - x = padd(x, y2); - - // Filter out invalid inputs, i.e. negative arg will be NAN, 0 will be -INF. - return _mm256_or_ps( - _mm256_andnot_ps(iszero_mask, _mm256_or_ps(x, invalid_mask)), - _mm256_and_ps(iszero_mask, p8f_minus_inf)); + return plog_float(_x); } // Exponential function. Works by writing "x = m*log(2) + r" where @@ -207,62 +42,7 @@ plog<Packet8f>(const Packet8f& _x) { template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f pexp<Packet8f>(const Packet8f& _x) { - _EIGEN_DECLARE_CONST_Packet8f(1, 1.0f); - _EIGEN_DECLARE_CONST_Packet8f(half, 0.5f); - _EIGEN_DECLARE_CONST_Packet8f(127, 127.0f); - - _EIGEN_DECLARE_CONST_Packet8f(exp_hi, 88.3762626647950f); - _EIGEN_DECLARE_CONST_Packet8f(exp_lo, -88.3762626647949f); - - _EIGEN_DECLARE_CONST_Packet8f(cephes_LOG2EF, 1.44269504088896341f); - - _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p0, 1.9875691500E-4f); - _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p1, 1.3981999507E-3f); - _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p2, 8.3334519073E-3f); - _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p3, 4.1665795894E-2f); - _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p4, 1.6666665459E-1f); - _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p5, 5.0000001201E-1f); - - // Clamp x. - Packet8f x = pmax(pmin(_x, p8f_exp_hi), p8f_exp_lo); - - // Express exp(x) as exp(m*ln(2) + r), start by extracting - // m = floor(x/ln(2) + 0.5). - Packet8f m = _mm256_floor_ps(pmadd(x, p8f_cephes_LOG2EF, p8f_half)); - -// Get r = x - m*ln(2). If no FMA instructions are available, m*ln(2) is -// subtracted out in two parts, m*C1+m*C2 = m*ln(2), to avoid accumulating -// truncation errors. Note that we don't use the "pmadd" function here to -// ensure that a precision-preserving FMA instruction is used. -#ifdef EIGEN_VECTORIZE_FMA - _EIGEN_DECLARE_CONST_Packet8f(nln2, -0.6931471805599453f); - Packet8f r = _mm256_fmadd_ps(m, p8f_nln2, x); -#else - _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_C1, 0.693359375f); - _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_C2, -2.12194440e-4f); - Packet8f r = psub(x, pmul(m, p8f_cephes_exp_C1)); - r = psub(r, pmul(m, p8f_cephes_exp_C2)); -#endif - - Packet8f r2 = pmul(r, r); - - // TODO(gonnet): Split into odd/even polynomials and try to exploit - // instruction-level parallelism. - Packet8f y = p8f_cephes_exp_p0; - y = pmadd(y, r, p8f_cephes_exp_p1); - y = pmadd(y, r, p8f_cephes_exp_p2); - y = pmadd(y, r, p8f_cephes_exp_p3); - y = pmadd(y, r, p8f_cephes_exp_p4); - y = pmadd(y, r, p8f_cephes_exp_p5); - y = pmadd(y, r2, r); - y = padd(y, p8f_1); - - // Build emm0 = 2^m. - Packet8i emm0 = _mm256_cvttps_epi32(padd(m, p8f_127)); - emm0 = pshiftleft(emm0, 23); - - // Return 2^m * exp(r). - return pmax(pmul(y, _mm256_castsi256_ps(emm0)), _x); + return pexp_float(_x); } // Hyperbolic Tangent function. @@ -274,82 +54,8 @@ ptanh<Packet8f>(const Packet8f& x) { template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4d -pexp<Packet4d>(const Packet4d& _x) { - Packet4d x = _x; - - _EIGEN_DECLARE_CONST_Packet4d(1, 1.0); - _EIGEN_DECLARE_CONST_Packet4d(2, 2.0); - _EIGEN_DECLARE_CONST_Packet4d(half, 0.5); - - _EIGEN_DECLARE_CONST_Packet4d(exp_hi, 709.437); - _EIGEN_DECLARE_CONST_Packet4d(exp_lo, -709.436139303); - - _EIGEN_DECLARE_CONST_Packet4d(cephes_LOG2EF, 1.4426950408889634073599); - - _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_p0, 1.26177193074810590878e-4); - _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_p1, 3.02994407707441961300e-2); - _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_p2, 9.99999999999999999910e-1); - - _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_q0, 3.00198505138664455042e-6); - _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_q1, 2.52448340349684104192e-3); - _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_q2, 2.27265548208155028766e-1); - _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_q3, 2.00000000000000000009e0); - - _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_C1, 0.693145751953125); - _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_C2, 1.42860682030941723212e-6); - _EIGEN_DECLARE_CONST_Packet4i(1023, 1023); - - Packet4d tmp, fx; - - // clamp x - x = pmax(pmin(x, p4d_exp_hi), p4d_exp_lo); - // Express exp(x) as exp(g + n*log(2)). - fx = pmadd(p4d_cephes_LOG2EF, x, p4d_half); - - // Get the integer modulus of log(2), i.e. the "n" described above. - fx = _mm256_floor_pd(fx); - - // Get the remainder modulo log(2), i.e. the "g" described above. Subtract - // n*log(2) out in two steps, i.e. n*C1 + n*C2, C1+C2=log2 to get the last - // digits right. - tmp = pmul(fx, p4d_cephes_exp_C1); - Packet4d z = pmul(fx, p4d_cephes_exp_C2); - x = psub(x, tmp); - x = psub(x, z); - - Packet4d x2 = pmul(x, x); - - // Evaluate the numerator polynomial of the rational interpolant. - Packet4d px = p4d_cephes_exp_p0; - px = pmadd(px, x2, p4d_cephes_exp_p1); - px = pmadd(px, x2, p4d_cephes_exp_p2); - px = pmul(px, x); - - // Evaluate the denominator polynomial of the rational interpolant. - Packet4d qx = p4d_cephes_exp_q0; - qx = pmadd(qx, x2, p4d_cephes_exp_q1); - qx = pmadd(qx, x2, p4d_cephes_exp_q2); - qx = pmadd(qx, x2, p4d_cephes_exp_q3); - - // I don't really get this bit, copied from the SSE2 routines, so... - // TODO(gonnet): Figure out what is going on here, perhaps find a better - // rational interpolant? - x = _mm256_div_pd(px, psub(qx, px)); - x = pmadd(p4d_2, x, p4d_1); - - // Build e=2^n by constructing the exponents in a 128-bit vector and - // shifting them to where they belong in double-precision values. - __m128i emm0 = _mm256_cvtpd_epi32(fx); - emm0 = _mm_add_epi32(emm0, p4i_1023); - emm0 = _mm_shuffle_epi32(emm0, _MM_SHUFFLE(3, 1, 2, 0)); - __m128i lo = _mm_slli_epi64(emm0, 52); - __m128i hi = _mm_slli_epi64(_mm_srli_epi64(emm0, 32), 52); - __m256i e = _mm256_insertf128_si256(_mm256_setzero_si256(), lo, 0); - e = _mm256_insertf128_si256(e, hi, 1); - - // Construct the result 2^n * exp(g) = e * x. The max is used to catch - // non-finite values in the input. - return pmax(pmul(x, _mm256_castsi256_pd(e)), _x); +pexp<Packet4d>(const Packet4d& x) { + return pexp_double(x); } // Functions for sqrt. diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index 774e64981..f88e36024 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -18,11 +18,11 @@ namespace internal { #define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8 #endif -#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS -#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*)) +#if !defined(EIGEN_VECTORIZE_AVX512) && !defined(EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS) +#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 16 #endif -#ifdef __FMA__ +#ifdef EIGEN_VECTORIZE_FMA #ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD #define EIGEN_HAS_SINGLE_INSTRUCTION_MADD #endif @@ -63,7 +63,7 @@ template<> struct packet_traits<float> : default_packet_traits HasDiv = 1, HasSin = EIGEN_FAST_MATH, - HasCos = 0, + HasCos = EIGEN_FAST_MATH, HasLog = 1, HasExp = 1, HasSqrt = 1, @@ -113,14 +113,29 @@ template<> struct packet_traits<int> : default_packet_traits }; */ -template<> struct unpacket_traits<Packet8f> { typedef float type; typedef Packet4f half; enum {size=8, alignment=Aligned32}; }; -template<> struct unpacket_traits<Packet4d> { typedef double type; typedef Packet2d half; enum {size=4, alignment=Aligned32}; }; -template<> struct unpacket_traits<Packet8i> { typedef int type; typedef Packet4i half; enum {size=8, alignment=Aligned32}; }; +template<> struct unpacket_traits<Packet8f> { + typedef float type; + typedef Packet4f half; + typedef Packet8i integer_packet; + enum {size=8, alignment=Aligned32, vectorizable=true}; +}; +template<> struct unpacket_traits<Packet4d> { + typedef double type; + typedef Packet2d half; + enum {size=4, alignment=Aligned32, vectorizable=true}; +}; +template<> struct unpacket_traits<Packet8i> { typedef int type; typedef Packet4i half; enum {size=8, alignment=Aligned32, vectorizable=false}; }; template<> EIGEN_STRONG_INLINE Packet8f pset1<Packet8f>(const float& from) { return _mm256_set1_ps(from); } template<> EIGEN_STRONG_INLINE Packet4d pset1<Packet4d>(const double& from) { return _mm256_set1_pd(from); } template<> EIGEN_STRONG_INLINE Packet8i pset1<Packet8i>(const int& from) { return _mm256_set1_epi32(from); } +template<> EIGEN_STRONG_INLINE Packet8f pset1frombits<Packet8f>(unsigned int from) { return _mm256_castsi256_ps(pset1<Packet8i>(from)); } + +template<> EIGEN_STRONG_INLINE Packet8f pzero(const Packet8f& /*a*/) { return _mm256_setzero_ps(); } +template<> EIGEN_STRONG_INLINE Packet4d pzero(const Packet4d& /*a*/) { return _mm256_setzero_pd(); } +template<> EIGEN_STRONG_INLINE Packet8i pzero(const Packet8i& /*a*/) { return _mm256_setzero_si256(); } + template<> EIGEN_STRONG_INLINE Packet8f pload1<Packet8f>(const float* from) { return _mm256_broadcast_ss(from); } template<> EIGEN_STRONG_INLINE Packet4d pload1<Packet4d>(const double* from) { return _mm256_broadcast_sd(from); } @@ -129,6 +144,15 @@ template<> EIGEN_STRONG_INLINE Packet4d plset<Packet4d>(const double& a) { retur template<> EIGEN_STRONG_INLINE Packet8f padd<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_add_ps(a,b); } template<> EIGEN_STRONG_INLINE Packet4d padd<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_add_pd(a,b); } +template<> EIGEN_STRONG_INLINE Packet8i padd<Packet8i>(const Packet8i& a, const Packet8i& b) { +#ifdef EIGEN_VECTORIZE_AVX2 + return _mm256_add_epi32(a,b); +#else + __m128i lo = _mm_add_epi32(_mm256_extractf128_si256(a, 0), _mm256_extractf128_si256(b, 0)); + __m128i hi = _mm_add_epi32(_mm256_extractf128_si256(a, 1), _mm256_extractf128_si256(b, 1)); + return _mm256_insertf128_si256(_mm256_castsi128_si256(lo), (hi), 1); +#endif +} template<> EIGEN_STRONG_INLINE Packet8f psub<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_sub_ps(a,b); } template<> EIGEN_STRONG_INLINE Packet4d psub<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_sub_pd(a,b); } @@ -157,13 +181,14 @@ template<> EIGEN_STRONG_INLINE Packet8i pdiv<Packet8i>(const Packet8i& /*a*/, co return pset1<Packet8i>(0); } -#ifdef __FMA__ +#ifdef EIGEN_VECTORIZE_FMA template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f& b, const Packet8f& c) { -#if ( EIGEN_COMP_GNUC_STRICT || (EIGEN_COMP_CLANG && (EIGEN_COMP_CLANG<308)) ) - // clang stupidly generates a vfmadd213ps instruction plus some vmovaps on registers, - // and gcc stupidly generates a vfmadd132ps instruction, - // so let's enforce it to generate a vfmadd231ps instruction since the most common use case is to accumulate - // the result of the product. +#if ( (EIGEN_COMP_GNUC_STRICT && EIGEN_COMP_GNUC<80) || (EIGEN_COMP_CLANG) ) + // Clang stupidly generates a vfmadd213ps instruction plus some vmovaps on registers, + // and even register spilling with clang>=6.0 (bug 1637). + // Gcc stupidly generates a vfmadd132ps instruction. + // So let's enforce it to generate a vfmadd231ps instruction since the most common use + // case is to accumulate the result of the product. Packet8f res = c; __asm__("vfmadd231ps %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b)); return res; @@ -172,7 +197,7 @@ template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f& #endif } template<> EIGEN_STRONG_INLINE Packet4d pmadd(const Packet4d& a, const Packet4d& b, const Packet4d& c) { -#if ( EIGEN_COMP_GNUC_STRICT || (EIGEN_COMP_CLANG && (EIGEN_COMP_CLANG<308)) ) +#if ( (EIGEN_COMP_GNUC_STRICT && EIGEN_COMP_GNUC<80) || (EIGEN_COMP_CLANG) ) // see above Packet4d res = c; __asm__("vfmadd231pd %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b)); @@ -184,21 +209,69 @@ template<> EIGEN_STRONG_INLINE Packet4d pmadd(const Packet4d& a, const Packet4d& #endif template<> EIGEN_STRONG_INLINE Packet8f pmin<Packet8f>(const Packet8f& a, const Packet8f& b) { +#if EIGEN_COMP_GNUC && EIGEN_COMP_GNUC < 63 + // There appears to be a bug in GCC, by which the optimizer may flip + // the argument order in calls to _mm_min_ps/_mm_max_ps, so we have to + // resort to inline ASM here. This is supposed to be fixed in gcc6.3, + // see also: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72867 + Packet8f res; + asm("vminps %[a], %[b], %[res]" : [res] "=x" (res) : [a] "x" (a), [b] "x" (b)); + return res; +#else // Arguments are swapped to match NaN propagation behavior of std::min. return _mm256_min_ps(b,a); +#endif } template<> EIGEN_STRONG_INLINE Packet4d pmin<Packet4d>(const Packet4d& a, const Packet4d& b) { +#if EIGEN_COMP_GNUC && EIGEN_COMP_GNUC < 63 + // See pmin above + Packet4d res; + asm("vminpd %[a], %[b], %[res]" : [res] "=x" (res) : [a] "x" (a), [b] "x" (b)); + return res; +#else // Arguments are swapped to match NaN propagation behavior of std::min. return _mm256_min_pd(b,a); +#endif } template<> EIGEN_STRONG_INLINE Packet8f pmax<Packet8f>(const Packet8f& a, const Packet8f& b) { +#if EIGEN_COMP_GNUC && EIGEN_COMP_GNUC < 63 + // See pmin above + Packet8f res; + asm("vmaxps %[a], %[b], %[res]" : [res] "=x" (res) : [a] "x" (a), [b] "x" (b)); + return res; +#else // Arguments are swapped to match NaN propagation behavior of std::max. return _mm256_max_ps(b,a); +#endif } template<> EIGEN_STRONG_INLINE Packet4d pmax<Packet4d>(const Packet4d& a, const Packet4d& b) { +#if EIGEN_COMP_GNUC && EIGEN_COMP_GNUC < 63 + // See pmin above + Packet4d res; + asm("vmaxpd %[a], %[b], %[res]" : [res] "=x" (res) : [a] "x" (a), [b] "x" (b)); + return res; +#else // Arguments are swapped to match NaN propagation behavior of std::max. return _mm256_max_pd(b,a); +#endif } + +template<> EIGEN_STRONG_INLINE Packet8f pcmp_le(const Packet8f& a, const Packet8f& b) { return _mm256_cmp_ps(a,b,_CMP_LE_OQ); } +template<> EIGEN_STRONG_INLINE Packet8f pcmp_lt(const Packet8f& a, const Packet8f& b) { return _mm256_cmp_ps(a,b,_CMP_LT_OQ); } +template<> EIGEN_STRONG_INLINE Packet8f pcmp_eq(const Packet8f& a, const Packet8f& b) { return _mm256_cmp_ps(a,b,_CMP_EQ_OQ); } +template<> EIGEN_STRONG_INLINE Packet4d pcmp_eq(const Packet4d& a, const Packet4d& b) { return _mm256_cmp_pd(a,b,_CMP_EQ_OQ); } +template<> EIGEN_STRONG_INLINE Packet8f pcmp_lt_or_nan(const Packet8f& a, const Packet8f& b) { return _mm256_cmp_ps(a, b, _CMP_NGE_UQ); } + +template<> EIGEN_STRONG_INLINE Packet8i pcmp_eq(const Packet8i& a, const Packet8i& b) { +#ifdef EIGEN_VECTORIZE_AVX2 + return _mm256_cmpeq_epi32(a,b); +#else + __m128i lo = _mm_cmpeq_epi32(_mm256_extractf128_si256(a, 0), _mm256_extractf128_si256(b, 0)); + __m128i hi = _mm_cmpeq_epi32(_mm256_extractf128_si256(a, 1), _mm256_extractf128_si256(b, 1)); + return _mm256_insertf128_si256(_mm256_castsi128_si256(lo), (hi), 1); +#endif +} + template<> EIGEN_STRONG_INLINE Packet8f pround<Packet8f>(const Packet8f& a) { return _mm256_round_ps(a, _MM_FROUND_CUR_DIRECTION); } template<> EIGEN_STRONG_INLINE Packet4d pround<Packet4d>(const Packet4d& a) { return _mm256_round_pd(a, _MM_FROUND_CUR_DIRECTION); } @@ -208,17 +281,101 @@ template<> EIGEN_STRONG_INLINE Packet4d pceil<Packet4d>(const Packet4d& a) { ret template<> EIGEN_STRONG_INLINE Packet8f pfloor<Packet8f>(const Packet8f& a) { return _mm256_floor_ps(a); } template<> EIGEN_STRONG_INLINE Packet4d pfloor<Packet4d>(const Packet4d& a) { return _mm256_floor_pd(a); } + +template<> EIGEN_STRONG_INLINE Packet8i ptrue<Packet8i>(const Packet8i& a) { +#ifdef EIGEN_VECTORIZE_AVX2 + // vpcmpeqd has lower latency than the more general vcmpps + return _mm256_cmpeq_epi32(a,a); +#else + const __m256 b = _mm256_castsi256_ps(a); + return _mm256_castps_si256(_mm256_cmp_ps(b,b,_CMP_TRUE_UQ)); +#endif +} + +template<> EIGEN_STRONG_INLINE Packet8f ptrue<Packet8f>(const Packet8f& a) { +#ifdef EIGEN_VECTORIZE_AVX2 + // vpcmpeqd has lower latency than the more general vcmpps + const __m256i b = _mm256_castps_si256(a); + return _mm256_castsi256_ps(_mm256_cmpeq_epi32(b,b)); +#else + return _mm256_cmp_ps(a,a,_CMP_TRUE_UQ); +#endif +} + +template<> EIGEN_STRONG_INLINE Packet4d ptrue<Packet4d>(const Packet4d& a) { +#ifdef EIGEN_VECTORIZE_AVX2 + // vpcmpeqq has lower latency than the more general vcmppd + const __m256i b = _mm256_castpd_si256(a); + return _mm256_castsi256_pd(_mm256_cmpeq_epi64(b,b)); +#else + return _mm256_cmp_pd(a,a,_CMP_TRUE_UQ); +#endif +} + template<> EIGEN_STRONG_INLINE Packet8f pand<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_and_ps(a,b); } template<> EIGEN_STRONG_INLINE Packet4d pand<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_and_pd(a,b); } +template<> EIGEN_STRONG_INLINE Packet8i pand<Packet8i>(const Packet8i& a, const Packet8i& b) { +#ifdef EIGEN_VECTORIZE_AVX2 + return _mm256_and_si256(a,b); +#else + return _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(a),_mm256_castsi256_ps(b))); +#endif +} template<> EIGEN_STRONG_INLINE Packet8f por<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_or_ps(a,b); } template<> EIGEN_STRONG_INLINE Packet4d por<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_or_pd(a,b); } +template<> EIGEN_STRONG_INLINE Packet8i por<Packet8i>(const Packet8i& a, const Packet8i& b) { +#ifdef EIGEN_VECTORIZE_AVX2 + return _mm256_or_si256(a,b); +#else + return _mm256_castps_si256(_mm256_or_ps(_mm256_castsi256_ps(a),_mm256_castsi256_ps(b))); +#endif +} template<> EIGEN_STRONG_INLINE Packet8f pxor<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_xor_ps(a,b); } template<> EIGEN_STRONG_INLINE Packet4d pxor<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_xor_pd(a,b); } +template<> EIGEN_STRONG_INLINE Packet8i pxor<Packet8i>(const Packet8i& a, const Packet8i& b) { +#ifdef EIGEN_VECTORIZE_AVX2 + return _mm256_xor_si256(a,b); +#else + return _mm256_castps_si256(_mm256_xor_ps(_mm256_castsi256_ps(a),_mm256_castsi256_ps(b))); +#endif +} + +template<> EIGEN_STRONG_INLINE Packet8f pandnot<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_andnot_ps(b,a); } +template<> EIGEN_STRONG_INLINE Packet4d pandnot<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_andnot_pd(b,a); } +template<> EIGEN_STRONG_INLINE Packet8i pandnot<Packet8i>(const Packet8i& a, const Packet8i& b) { +#ifdef EIGEN_VECTORIZE_AVX2 + return _mm256_andnot_si256(b,a); +#else + return _mm256_castps_si256(_mm256_andnot_ps(_mm256_castsi256_ps(b),_mm256_castsi256_ps(a))); +#endif +} -template<> EIGEN_STRONG_INLINE Packet8f pandnot<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_andnot_ps(a,b); } -template<> EIGEN_STRONG_INLINE Packet4d pandnot<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_andnot_pd(a,b); } +template<> EIGEN_STRONG_INLINE Packet8f pselect<Packet8f>(const Packet8f& mask, const Packet8f& a, const Packet8f& b) +{ return _mm256_blendv_ps(b,a,mask); } +template<> EIGEN_STRONG_INLINE Packet4d pselect<Packet4d>(const Packet4d& mask, const Packet4d& a, const Packet4d& b) +{ return _mm256_blendv_pd(b,a,mask); } + +template<int N> EIGEN_STRONG_INLINE Packet8i pshiftright(Packet8i a) { +#ifdef EIGEN_VECTORIZE_AVX2 + return _mm256_srli_epi32(a, N); +#else + __m128i lo = _mm_srli_epi32(_mm256_extractf128_si256(a, 0), N); + __m128i hi = _mm_srli_epi32(_mm256_extractf128_si256(a, 1), N); + return _mm256_insertf128_si256(_mm256_castsi128_si256(lo), (hi), 1); +#endif +} + +template<int N> EIGEN_STRONG_INLINE Packet8i pshiftleft(Packet8i a) { +#ifdef EIGEN_VECTORIZE_AVX2 + return _mm256_slli_epi32(a, N); +#else + __m128i lo = _mm_slli_epi32(_mm256_extractf128_si256(a, 0), N); + __m128i hi = _mm_slli_epi32(_mm256_extractf128_si256(a, 1), N); + return _mm256_insertf128_si256(_mm256_castsi128_si256(lo), (hi), 1); +#endif +} template<> EIGEN_STRONG_INLINE Packet8f pload<Packet8f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm256_load_ps(from); } template<> EIGEN_STRONG_INLINE Packet4d pload<Packet4d>(const double* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm256_load_pd(from); } @@ -363,6 +520,28 @@ template<> EIGEN_STRONG_INLINE Packet4d pabs(const Packet4d& a) return _mm256_and_pd(a,mask); } +template<> EIGEN_STRONG_INLINE Packet8f pfrexp<Packet8f>(const Packet8f& a, Packet8f& exponent) { + return pfrexp_float(a,exponent); +} + +template<> EIGEN_STRONG_INLINE Packet8f pldexp<Packet8f>(const Packet8f& a, const Packet8f& exponent) { + return pldexp_float(a,exponent); +} + +template<> EIGEN_STRONG_INLINE Packet4d pldexp<Packet4d>(const Packet4d& a, const Packet4d& exponent) { + // Build e=2^n by constructing the exponents in a 128-bit vector and + // shifting them to where they belong in double-precision values. + Packet4i cst_1023 = pset1<Packet4i>(1023); + __m128i emm0 = _mm256_cvtpd_epi32(exponent); + emm0 = _mm_add_epi32(emm0, cst_1023); + emm0 = _mm_shuffle_epi32(emm0, _MM_SHUFFLE(3, 1, 2, 0)); + __m128i lo = _mm_slli_epi64(emm0, 52); + __m128i hi = _mm_slli_epi64(_mm_srli_epi64(emm0, 32), 52); + __m256i e = _mm256_insertf128_si256(_mm256_setzero_si256(), lo, 0); + e = _mm256_insertf128_si256(e, hi, 1); + return pmul(a,_mm256_castsi256_pd(e)); +} + // preduxp should be ok // FIXME: why is this ok? why isn't the simply implementation working as expected? template<> EIGEN_STRONG_INLINE Packet8f preduxp<Packet8f>(const Packet8f* vecs) @@ -459,6 +638,16 @@ template<> EIGEN_STRONG_INLINE double predux_max<Packet4d>(const Packet4d& a) return pfirst(_mm256_max_pd(tmp, _mm256_shuffle_pd(tmp, tmp, 1))); } +// not needed yet +// template<> EIGEN_STRONG_INLINE bool predux_all(const Packet8f& x) +// { +// return _mm256_movemask_ps(x)==0xFF; +// } + +template<> EIGEN_STRONG_INLINE bool predux_any(const Packet8f& x) +{ + return _mm256_movemask_ps(x)!=0; +} template<int Offset> struct palign_impl<Offset,Packet8f> diff --git a/Eigen/src/Core/arch/AVX/TypeCasting.h b/Eigen/src/Core/arch/AVX/TypeCasting.h index 83bfdc604..7d2e1e67f 100644 --- a/Eigen/src/Core/arch/AVX/TypeCasting.h +++ b/Eigen/src/Core/arch/AVX/TypeCasting.h @@ -37,13 +37,21 @@ struct type_casting_traits<int, float> { template<> EIGEN_STRONG_INLINE Packet8i pcast<Packet8f, Packet8i>(const Packet8f& a) { - return _mm256_cvtps_epi32(a); + return _mm256_cvttps_epi32(a); } template<> EIGEN_STRONG_INLINE Packet8f pcast<Packet8i, Packet8f>(const Packet8i& a) { return _mm256_cvtepi32_ps(a); } +template<> EIGEN_STRONG_INLINE Packet8i preinterpret<Packet8i,Packet8f>(const Packet8f& a) { + return _mm256_castps_si256(a); +} + +template<> EIGEN_STRONG_INLINE Packet8f preinterpret<Packet8f,Packet8i>(const Packet8i& a) { + return _mm256_castsi256_ps(a); +} + } // end namespace internal } // end namespace Eigen diff --git a/Eigen/src/Core/arch/AVX512/Complex.h b/Eigen/src/Core/arch/AVX512/Complex.h new file mode 100644 index 000000000..9a89dd01f --- /dev/null +++ b/Eigen/src/Core/arch/AVX512/Complex.h @@ -0,0 +1,488 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2018 Gael Guennebaud <gael.guennebaud@inria.fr> +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_COMPLEX_AVX512_H +#define EIGEN_COMPLEX_AVX512_H + +namespace Eigen { + +namespace internal { + +//---------- float ---------- +struct Packet8cf +{ + EIGEN_STRONG_INLINE Packet8cf() {} + EIGEN_STRONG_INLINE explicit Packet8cf(const __m512& a) : v(a) {} + __m512 v; +}; + +template<> struct packet_traits<std::complex<float> > : default_packet_traits +{ + typedef Packet8cf type; + typedef Packet4cf half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size = 8, + HasHalfPacket = 1, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasNegate = 1, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasSetLinear = 0, + HasReduxp = 0 + }; +}; + +template<> struct unpacket_traits<Packet8cf> { + typedef std::complex<float> type; + enum { + size = 8, + alignment=unpacket_traits<Packet16f>::alignment, + vectorizable=true + }; + typedef Packet4cf half; +}; + +template<> EIGEN_STRONG_INLINE Packet8cf ptrue<Packet8cf>(const Packet8cf& a) { return Packet8cf(ptrue(Packet16f(a.v))); } +template<> EIGEN_STRONG_INLINE Packet8cf pnot<Packet8cf>(const Packet8cf& a) { return Packet8cf(pnot(Packet16f(a.v))); } +template<> EIGEN_STRONG_INLINE Packet8cf padd<Packet8cf>(const Packet8cf& a, const Packet8cf& b) { return Packet8cf(_mm512_add_ps(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet8cf psub<Packet8cf>(const Packet8cf& a, const Packet8cf& b) { return Packet8cf(_mm512_sub_ps(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet8cf pnegate(const Packet8cf& a) +{ + return Packet8cf(pnegate(a.v)); +} +template<> EIGEN_STRONG_INLINE Packet8cf pconj(const Packet8cf& a) +{ + const __m512 mask = _mm512_castsi512_ps(_mm512_setr_epi32( + 0x00000000,0x80000000,0x00000000,0x80000000,0x00000000,0x80000000,0x00000000,0x80000000, + 0x00000000,0x80000000,0x00000000,0x80000000,0x00000000,0x80000000,0x00000000,0x80000000)); + return Packet8cf(pxor(a.v,mask)); +} + +template<> EIGEN_STRONG_INLINE Packet8cf pmul<Packet8cf>(const Packet8cf& a, const Packet8cf& b) +{ + __m512 tmp2 = _mm512_mul_ps(_mm512_movehdup_ps(a.v), _mm512_permute_ps(b.v, _MM_SHUFFLE(2,3,0,1))); + return Packet8cf(_mm512_fmaddsub_ps(_mm512_moveldup_ps(a.v), b.v, tmp2)); +} + +template<> EIGEN_STRONG_INLINE Packet8cf pand <Packet8cf>(const Packet8cf& a, const Packet8cf& b) { return Packet8cf(pand(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet8cf por <Packet8cf>(const Packet8cf& a, const Packet8cf& b) { return Packet8cf(por(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet8cf pxor <Packet8cf>(const Packet8cf& a, const Packet8cf& b) { return Packet8cf(pxor(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet8cf pandnot<Packet8cf>(const Packet8cf& a, const Packet8cf& b) { return Packet8cf(pandnot(a.v,b.v)); } + +template <> +EIGEN_STRONG_INLINE Packet8cf pcmp_eq(const Packet8cf& a, const Packet8cf& b) { + __m512 eq = pcmp_eq<Packet16f>(a.v, b.v); + return Packet8cf(pand(eq, _mm512_permute_ps(eq, 0xB1))); +} + +template<> EIGEN_STRONG_INLINE Packet8cf pload <Packet8cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet8cf(pload<Packet16f>(&numext::real_ref(*from))); } +template<> EIGEN_STRONG_INLINE Packet8cf ploadu<Packet8cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet8cf(ploadu<Packet16f>(&numext::real_ref(*from))); } + + +template<> EIGEN_STRONG_INLINE Packet8cf pset1<Packet8cf>(const std::complex<float>& from) +{ + return Packet8cf(_mm512_castpd_ps(pload1<Packet8d>((const double*)(const void*)&from))); +} + +template<> EIGEN_STRONG_INLINE Packet8cf ploaddup<Packet8cf>(const std::complex<float>* from) +{ + return Packet8cf( _mm512_castpd_ps( ploaddup<Packet8d>((const double*)(const void*)from )) ); +} +template<> EIGEN_STRONG_INLINE Packet8cf ploadquad<Packet8cf>(const std::complex<float>* from) +{ + return Packet8cf( _mm512_castpd_ps( ploadquad<Packet8d>((const double*)(const void*)from )) ); +} + +template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float>* to, const Packet8cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore(&numext::real_ref(*to), from.v); } +template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float>* to, const Packet8cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&numext::real_ref(*to), from.v); } + +template<> EIGEN_DEVICE_FUNC inline Packet8cf pgather<std::complex<float>, Packet8cf>(const std::complex<float>* from, Index stride) +{ + return Packet8cf(_mm512_castpd_ps(pgather<double,Packet8d>((const double*)(const void*)from, stride))); +} + +template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet8cf>(std::complex<float>* to, const Packet8cf& from, Index stride) +{ + pscatter((double*)(void*)to, _mm512_castps_pd(from.v), stride); +} + +template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet8cf>(const Packet8cf& a) +{ + return pfirst(Packet2cf(_mm512_castps512_ps128(a.v))); +} + +template<> EIGEN_STRONG_INLINE Packet8cf preverse(const Packet8cf& a) { + return Packet8cf(_mm512_castsi512_ps( + _mm512_permutexvar_epi64( _mm512_set_epi32(0, 0, 0, 1, 0, 2, 0, 3, 0, 4, 0, 5, 0, 6, 0, 7), + _mm512_castps_si512(a.v)))); +} + +template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet8cf>(const Packet8cf& a) +{ + return predux(padd(Packet4cf(extract256<0>(a.v)), + Packet4cf(extract256<1>(a.v)))); +} + +template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet8cf>(const Packet8cf& a) +{ + return predux_mul(pmul(Packet4cf(extract256<0>(a.v)), + Packet4cf(extract256<1>(a.v)))); +} + +template <> +EIGEN_STRONG_INLINE Packet4cf predux_half_dowto4<Packet8cf>(const Packet8cf& a) { + __m256 lane0 = extract256<0>(a.v); + __m256 lane1 = extract256<1>(a.v); + __m256 res = _mm256_add_ps(lane0, lane1); + return Packet4cf(res); +} + +template<int Offset> +struct palign_impl<Offset,Packet8cf> +{ + static EIGEN_STRONG_INLINE void run(Packet8cf& first, const Packet8cf& second) + { + if (Offset==0) return; + palign_impl<Offset*2,Packet16f>::run(first.v, second.v); + } +}; + +template<> struct conj_helper<Packet8cf, Packet8cf, false,true> +{ + EIGEN_STRONG_INLINE Packet8cf pmadd(const Packet8cf& x, const Packet8cf& y, const Packet8cf& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet8cf pmul(const Packet8cf& a, const Packet8cf& b) const + { + return internal::pmul(a, pconj(b)); + } +}; + +template<> struct conj_helper<Packet8cf, Packet8cf, true,false> +{ + EIGEN_STRONG_INLINE Packet8cf pmadd(const Packet8cf& x, const Packet8cf& y, const Packet8cf& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet8cf pmul(const Packet8cf& a, const Packet8cf& b) const + { + return internal::pmul(pconj(a), b); + } +}; + +template<> struct conj_helper<Packet8cf, Packet8cf, true,true> +{ + EIGEN_STRONG_INLINE Packet8cf pmadd(const Packet8cf& x, const Packet8cf& y, const Packet8cf& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet8cf pmul(const Packet8cf& a, const Packet8cf& b) const + { + return pconj(internal::pmul(a, b)); + } +}; + +EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet8cf,Packet16f) + +template<> EIGEN_STRONG_INLINE Packet8cf pdiv<Packet8cf>(const Packet8cf& a, const Packet8cf& b) +{ + Packet8cf num = pmul(a, pconj(b)); + __m512 tmp = _mm512_mul_ps(b.v, b.v); + __m512 tmp2 = _mm512_shuffle_ps(tmp,tmp,0xB1); + __m512 denom = _mm512_add_ps(tmp, tmp2); + return Packet8cf(_mm512_div_ps(num.v, denom)); +} + +template<> EIGEN_STRONG_INLINE Packet8cf pcplxflip<Packet8cf>(const Packet8cf& x) +{ + return Packet8cf(_mm512_shuffle_ps(x.v, x.v, _MM_SHUFFLE(2, 3, 0 ,1))); +} + +//---------- double ---------- +struct Packet4cd +{ + EIGEN_STRONG_INLINE Packet4cd() {} + EIGEN_STRONG_INLINE explicit Packet4cd(const __m512d& a) : v(a) {} + __m512d v; +}; + +template<> struct packet_traits<std::complex<double> > : default_packet_traits +{ + typedef Packet4cd type; + typedef Packet2cd half; + enum { + Vectorizable = 1, + AlignedOnScalar = 0, + size = 4, + HasHalfPacket = 1, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasNegate = 1, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasSetLinear = 0, + HasReduxp = 0 + }; +}; + +template<> struct unpacket_traits<Packet4cd> { + typedef std::complex<double> type; + enum { + size = 4, + alignment = unpacket_traits<Packet8d>::alignment, + vectorizable=true + }; + typedef Packet2cd half; +}; + +template<> EIGEN_STRONG_INLINE Packet4cd padd<Packet4cd>(const Packet4cd& a, const Packet4cd& b) { return Packet4cd(_mm512_add_pd(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet4cd psub<Packet4cd>(const Packet4cd& a, const Packet4cd& b) { return Packet4cd(_mm512_sub_pd(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet4cd pnegate(const Packet4cd& a) { return Packet4cd(pnegate(a.v)); } +template<> EIGEN_STRONG_INLINE Packet4cd pconj(const Packet4cd& a) +{ + const __m512d mask = _mm512_castsi512_pd( + _mm512_set_epi32(0x80000000,0x0,0x0,0x0,0x80000000,0x0,0x0,0x0, + 0x80000000,0x0,0x0,0x0,0x80000000,0x0,0x0,0x0)); + return Packet4cd(pxor(a.v,mask)); +} + +template<> EIGEN_STRONG_INLINE Packet4cd pmul<Packet4cd>(const Packet4cd& a, const Packet4cd& b) +{ + __m512d tmp1 = _mm512_shuffle_pd(a.v,a.v,0x0); + __m512d tmp2 = _mm512_shuffle_pd(a.v,a.v,0xFF); + __m512d tmp3 = _mm512_shuffle_pd(b.v,b.v,0x55); + __m512d odd = _mm512_mul_pd(tmp2, tmp3); + return Packet4cd(_mm512_fmaddsub_pd(tmp1, b.v, odd)); +} + +template<> EIGEN_STRONG_INLINE Packet4cd ptrue<Packet4cd>(const Packet4cd& a) { return Packet4cd(ptrue(Packet8d(a.v))); } +template<> EIGEN_STRONG_INLINE Packet4cd pnot<Packet4cd>(const Packet4cd& a) { return Packet4cd(pnot(Packet8d(a.v))); } +template<> EIGEN_STRONG_INLINE Packet4cd pand <Packet4cd>(const Packet4cd& a, const Packet4cd& b) { return Packet4cd(pand(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet4cd por <Packet4cd>(const Packet4cd& a, const Packet4cd& b) { return Packet4cd(por(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet4cd pxor <Packet4cd>(const Packet4cd& a, const Packet4cd& b) { return Packet4cd(pxor(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet4cd pandnot<Packet4cd>(const Packet4cd& a, const Packet4cd& b) { return Packet4cd(pandnot(a.v,b.v)); } + +template <> +EIGEN_STRONG_INLINE Packet4cd pcmp_eq(const Packet4cd& a, const Packet4cd& b) { + __m512d eq = pcmp_eq<Packet8d>(a.v, b.v); + return Packet4cd(pand(eq, _mm512_permute_pd(eq, 0x55))); +} + +template<> EIGEN_STRONG_INLINE Packet4cd pload <Packet4cd>(const std::complex<double>* from) +{ EIGEN_DEBUG_ALIGNED_LOAD return Packet4cd(pload<Packet8d>((const double*)from)); } +template<> EIGEN_STRONG_INLINE Packet4cd ploadu<Packet4cd>(const std::complex<double>* from) +{ EIGEN_DEBUG_UNALIGNED_LOAD return Packet4cd(ploadu<Packet8d>((const double*)from)); } + +template<> EIGEN_STRONG_INLINE Packet4cd pset1<Packet4cd>(const std::complex<double>& from) +{ + #ifdef EIGEN_VECTORIZE_AVX512DQ + return Packet4cd(_mm512_broadcast_f64x2(pset1<Packet1cd>(from).v)); + #else + return Packet4cd(_mm512_castps_pd(_mm512_broadcast_f32x4( _mm_castpd_ps(pset1<Packet1cd>(from).v)))); + #endif +} + +template<> EIGEN_STRONG_INLINE Packet4cd ploaddup<Packet4cd>(const std::complex<double>* from) { + return Packet4cd(_mm512_insertf64x4( + _mm512_castpd256_pd512(ploaddup<Packet2cd>(from).v), ploaddup<Packet2cd>(from+1).v, 1)); +} + +template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> * to, const Packet4cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); } +template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> * to, const Packet4cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); } + +template<> EIGEN_DEVICE_FUNC inline Packet4cd pgather<std::complex<double>, Packet4cd>(const std::complex<double>* from, Index stride) +{ + return Packet4cd(_mm512_insertf64x4(_mm512_castpd256_pd512( + _mm256_insertf128_pd(_mm256_castpd128_pd256(ploadu<Packet1cd>(from+0*stride).v), ploadu<Packet1cd>(from+1*stride).v,1)), + _mm256_insertf128_pd(_mm256_castpd128_pd256(ploadu<Packet1cd>(from+2*stride).v), ploadu<Packet1cd>(from+3*stride).v,1), 1)); +} + +template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet4cd>(std::complex<double>* to, const Packet4cd& from, Index stride) +{ + __m512i fromi = _mm512_castpd_si512(from.v); + double* tod = (double*)(void*)to; + _mm_storeu_pd(tod+0*stride, _mm_castsi128_pd(_mm512_extracti32x4_epi32(fromi,0)) ); + _mm_storeu_pd(tod+2*stride, _mm_castsi128_pd(_mm512_extracti32x4_epi32(fromi,1)) ); + _mm_storeu_pd(tod+4*stride, _mm_castsi128_pd(_mm512_extracti32x4_epi32(fromi,2)) ); + _mm_storeu_pd(tod+6*stride, _mm_castsi128_pd(_mm512_extracti32x4_epi32(fromi,3)) ); +} + +template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet4cd>(const Packet4cd& a) +{ + __m128d low = extract128<0>(a.v); + EIGEN_ALIGN16 double res[2]; + _mm_store_pd(res, low); + return std::complex<double>(res[0],res[1]); +} + +template<> EIGEN_STRONG_INLINE Packet4cd preverse(const Packet4cd& a) { + return Packet4cd(_mm512_shuffle_f64x2(a.v, a.v, EIGEN_SSE_SHUFFLE_MASK(3,2,1,0))); +} + +template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet4cd>(const Packet4cd& a) +{ + return predux(padd(Packet2cd(_mm512_extractf64x4_pd(a.v,0)), + Packet2cd(_mm512_extractf64x4_pd(a.v,1)))); +} + +template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet4cd>(const Packet4cd& a) +{ + return predux_mul(pmul(Packet2cd(_mm512_extractf64x4_pd(a.v,0)), + Packet2cd(_mm512_extractf64x4_pd(a.v,1)))); +} + +template<int Offset> +struct palign_impl<Offset,Packet4cd> +{ + static EIGEN_STRONG_INLINE void run(Packet4cd& first, const Packet4cd& second) + { + if (Offset==0) return; + palign_impl<Offset*2,Packet8d>::run(first.v, second.v); + } +}; + +template<> struct conj_helper<Packet4cd, Packet4cd, false,true> +{ + EIGEN_STRONG_INLINE Packet4cd pmadd(const Packet4cd& x, const Packet4cd& y, const Packet4cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet4cd pmul(const Packet4cd& a, const Packet4cd& b) const + { + return internal::pmul(a, pconj(b)); + } +}; + +template<> struct conj_helper<Packet4cd, Packet4cd, true,false> +{ + EIGEN_STRONG_INLINE Packet4cd pmadd(const Packet4cd& x, const Packet4cd& y, const Packet4cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet4cd pmul(const Packet4cd& a, const Packet4cd& b) const + { + return internal::pmul(pconj(a), b); + } +}; + +template<> struct conj_helper<Packet4cd, Packet4cd, true,true> +{ + EIGEN_STRONG_INLINE Packet4cd pmadd(const Packet4cd& x, const Packet4cd& y, const Packet4cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet4cd pmul(const Packet4cd& a, const Packet4cd& b) const + { + return pconj(internal::pmul(a, b)); + } +}; + +EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet4cd,Packet8d) + +template<> EIGEN_STRONG_INLINE Packet4cd pdiv<Packet4cd>(const Packet4cd& a, const Packet4cd& b) +{ + Packet4cd num = pmul(a, pconj(b)); + __m512d tmp = _mm512_mul_pd(b.v, b.v); + __m512d denom = padd(_mm512_permute_pd(tmp,0x55), tmp); + return Packet4cd(_mm512_div_pd(num.v, denom)); +} + +template<> EIGEN_STRONG_INLINE Packet4cd pcplxflip<Packet4cd>(const Packet4cd& x) +{ + return Packet4cd(_mm512_permute_pd(x.v,0x55)); +} + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet8cf,4>& kernel) { + PacketBlock<Packet8d,4> pb; + + pb.packet[0] = _mm512_castps_pd(kernel.packet[0].v); + pb.packet[1] = _mm512_castps_pd(kernel.packet[1].v); + pb.packet[2] = _mm512_castps_pd(kernel.packet[2].v); + pb.packet[3] = _mm512_castps_pd(kernel.packet[3].v); + ptranspose(pb); + kernel.packet[0].v = _mm512_castpd_ps(pb.packet[0]); + kernel.packet[1].v = _mm512_castpd_ps(pb.packet[1]); + kernel.packet[2].v = _mm512_castpd_ps(pb.packet[2]); + kernel.packet[3].v = _mm512_castpd_ps(pb.packet[3]); +} + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet8cf,8>& kernel) { + PacketBlock<Packet8d,8> pb; + + pb.packet[0] = _mm512_castps_pd(kernel.packet[0].v); + pb.packet[1] = _mm512_castps_pd(kernel.packet[1].v); + pb.packet[2] = _mm512_castps_pd(kernel.packet[2].v); + pb.packet[3] = _mm512_castps_pd(kernel.packet[3].v); + pb.packet[4] = _mm512_castps_pd(kernel.packet[4].v); + pb.packet[5] = _mm512_castps_pd(kernel.packet[5].v); + pb.packet[6] = _mm512_castps_pd(kernel.packet[6].v); + pb.packet[7] = _mm512_castps_pd(kernel.packet[7].v); + ptranspose(pb); + kernel.packet[0].v = _mm512_castpd_ps(pb.packet[0]); + kernel.packet[1].v = _mm512_castpd_ps(pb.packet[1]); + kernel.packet[2].v = _mm512_castpd_ps(pb.packet[2]); + kernel.packet[3].v = _mm512_castpd_ps(pb.packet[3]); + kernel.packet[4].v = _mm512_castpd_ps(pb.packet[4]); + kernel.packet[5].v = _mm512_castpd_ps(pb.packet[5]); + kernel.packet[6].v = _mm512_castpd_ps(pb.packet[6]); + kernel.packet[7].v = _mm512_castpd_ps(pb.packet[7]); +} + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet4cd,4>& kernel) { + __m512d T0 = _mm512_shuffle_f64x2(kernel.packet[0].v, kernel.packet[1].v, EIGEN_SSE_SHUFFLE_MASK(0,1,0,1)); // [a0 a1 b0 b1] + __m512d T1 = _mm512_shuffle_f64x2(kernel.packet[0].v, kernel.packet[1].v, EIGEN_SSE_SHUFFLE_MASK(2,3,2,3)); // [a2 a3 b2 b3] + __m512d T2 = _mm512_shuffle_f64x2(kernel.packet[2].v, kernel.packet[3].v, EIGEN_SSE_SHUFFLE_MASK(0,1,0,1)); // [c0 c1 d0 d1] + __m512d T3 = _mm512_shuffle_f64x2(kernel.packet[2].v, kernel.packet[3].v, EIGEN_SSE_SHUFFLE_MASK(2,3,2,3)); // [c2 c3 d2 d3] + + kernel.packet[3] = Packet4cd(_mm512_shuffle_f64x2(T1, T3, EIGEN_SSE_SHUFFLE_MASK(1,3,1,3))); // [a3 b3 c3 d3] + kernel.packet[2] = Packet4cd(_mm512_shuffle_f64x2(T1, T3, EIGEN_SSE_SHUFFLE_MASK(0,2,0,2))); // [a2 b2 c2 d2] + kernel.packet[1] = Packet4cd(_mm512_shuffle_f64x2(T0, T2, EIGEN_SSE_SHUFFLE_MASK(1,3,1,3))); // [a1 b1 c1 d1] + kernel.packet[0] = Packet4cd(_mm512_shuffle_f64x2(T0, T2, EIGEN_SSE_SHUFFLE_MASK(0,2,0,2))); // [a0 b0 c0 d0] +} + +template<> EIGEN_STRONG_INLINE Packet8cf pinsertfirst(const Packet8cf& a, std::complex<float> b) +{ + Packet2cf tmp = Packet2cf(_mm512_extractf32x4_ps(a.v,0)); + tmp = pinsertfirst(tmp, b); + return Packet8cf( _mm512_insertf32x4(a.v, tmp.v, 0) ); +} + +template<> EIGEN_STRONG_INLINE Packet4cd pinsertfirst(const Packet4cd& a, std::complex<double> b) +{ + return Packet4cd(_mm512_castsi512_pd( _mm512_inserti32x4(_mm512_castpd_si512(a.v), _mm_castpd_si128(pset1<Packet1cd>(b).v), 0) )); +} + +template<> EIGEN_STRONG_INLINE Packet8cf pinsertlast(const Packet8cf& a, std::complex<float> b) +{ + Packet2cf tmp = Packet2cf(_mm512_extractf32x4_ps(a.v,3) ); + tmp = pinsertlast(tmp, b); + return Packet8cf( _mm512_insertf32x4(a.v, tmp.v, 3) ); +} + +template<> EIGEN_STRONG_INLINE Packet4cd pinsertlast(const Packet4cd& a, std::complex<double> b) +{ + return Packet4cd(_mm512_castsi512_pd( _mm512_inserti32x4(_mm512_castpd_si512(a.v), _mm_castpd_si128(pset1<Packet1cd>(b).v), 3) )); +} + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_COMPLEX_AVX512_H diff --git a/Eigen/src/Core/arch/AVX512/MathFunctions.h b/Eigen/src/Core/arch/AVX512/MathFunctions.h index 93c5ec43f..c2158c538 100644 --- a/Eigen/src/Core/arch/AVX512/MathFunctions.h +++ b/Eigen/src/Core/arch/AVX512/MathFunctions.h @@ -47,6 +47,7 @@ plog<Packet16f>(const Packet16f& _x) { // The smallest non denormalized float number. _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(min_norm_pos, 0x00800000); _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(minus_inf, 0xff800000); + _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(pos_inf, 0x7f800000); _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(nan, 0x7fc00000); // Polynomial coefficients. @@ -116,10 +117,16 @@ plog<Packet16f>(const Packet16f& _x) { x = padd(x, y); x = padd(x, y2); - // Filter out invalid inputs, i.e. negative arg will be NAN, 0 will be -INF. + __mmask16 pos_inf_mask = _mm512_cmp_ps_mask(_x,p16f_pos_inf,_CMP_EQ_OQ); + // Filter out invalid inputs, i.e.: + // - negative arg will be NAN, + // - 0 will be -INF. + // - +INF will be +INF return _mm512_mask_blend_ps(iszero_mask, - _mm512_mask_blend_ps(invalid_mask, x, p16f_nan), - p16f_minus_inf); + _mm512_mask_blend_ps(invalid_mask, + _mm512_mask_blend_ps(pos_inf_mask,x,p16f_pos_inf), + p16f_nan), + p16f_minus_inf); } #endif @@ -373,6 +380,19 @@ EIGEN_STRONG_INLINE Packet16f prsqrt<Packet16f>(const Packet16f& x) { #endif #endif + +template <> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet16f +psin<Packet16f>(const Packet16f& _x) { + return psin_float(_x); +} + +template <> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet16f +pcos<Packet16f>(const Packet16f& _x) { + return pcos_float(_x); +} + } // end namespace internal } // end namespace Eigen diff --git a/Eigen/src/Core/arch/AVX512/PacketMath.h b/Eigen/src/Core/arch/AVX512/PacketMath.h index 86cefba92..60b723b08 100644 --- a/Eigen/src/Core/arch/AVX512/PacketMath.h +++ b/Eigen/src/Core/arch/AVX512/PacketMath.h @@ -19,10 +19,10 @@ namespace internal { #endif #ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS -#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*)) +#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 32 #endif -#ifdef __FMA__ +#ifdef EIGEN_VECTORIZE_FMA #ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD #define EIGEN_HAS_SINGLE_INSTRUCTION_MADD #endif @@ -55,7 +55,9 @@ template<> struct packet_traits<float> : default_packet_traits size = 16, HasHalfPacket = 1, HasBlend = 0, -#if EIGEN_GNUC_AT_LEAST(5, 3) || EIGEN_COMP_CLANG + HasSin = EIGEN_FAST_MATH, + HasCos = EIGEN_FAST_MATH, +#if EIGEN_GNUC_AT_LEAST(5, 3) || (!EIGEN_COMP_GNUC_STRICT) #ifdef EIGEN_VECTORIZE_AVX512DQ HasLog = 1, #endif @@ -75,7 +77,7 @@ template<> struct packet_traits<double> : default_packet_traits AlignedOnScalar = 1, size = 8, HasHalfPacket = 1, -#if EIGEN_GNUC_AT_LEAST(5, 3) +#if EIGEN_GNUC_AT_LEAST(5, 3) || (!EIGEN_COMP_GNUC_STRICT) HasSqrt = EIGEN_FAST_MATH, HasRsqrt = EIGEN_FAST_MATH, #endif @@ -99,19 +101,20 @@ template <> struct unpacket_traits<Packet16f> { typedef float type; typedef Packet8f half; - enum { size = 16, alignment=Aligned64 }; + typedef Packet16i integer_packet; + enum { size = 16, alignment=Aligned64, vectorizable=true }; }; template <> struct unpacket_traits<Packet8d> { typedef double type; typedef Packet4d half; - enum { size = 8, alignment=Aligned64 }; + enum { size = 8, alignment=Aligned64, vectorizable=true }; }; template <> struct unpacket_traits<Packet16i> { typedef int type; typedef Packet8i half; - enum { size = 16, alignment=Aligned64 }; + enum { size = 16, alignment=Aligned64, vectorizable=false }; }; template <> @@ -128,12 +131,17 @@ EIGEN_STRONG_INLINE Packet16i pset1<Packet16i>(const int& from) { } template <> +EIGEN_STRONG_INLINE Packet16f pset1frombits<Packet16f>(unsigned int from) { + return _mm512_castsi512_ps(_mm512_set1_epi32(from)); +} + +template <> EIGEN_STRONG_INLINE Packet16f pload1<Packet16f>(const float* from) { return _mm512_broadcastss_ps(_mm_load_ps1(from)); } template <> EIGEN_STRONG_INLINE Packet8d pload1<Packet8d>(const double* from) { - return _mm512_broadcastsd_pd(_mm_load_pd1(from)); + return _mm512_set1_pd(*from); } template <> @@ -159,6 +167,11 @@ EIGEN_STRONG_INLINE Packet8d padd<Packet8d>(const Packet8d& a, const Packet8d& b) { return _mm512_add_pd(a, b); } +template <> +EIGEN_STRONG_INLINE Packet16i padd<Packet16i>(const Packet16i& a, + const Packet16i& b) { + return _mm512_add_epi32(a, b); +} template <> EIGEN_STRONG_INLINE Packet16f psub<Packet16f>(const Packet16f& a, @@ -170,6 +183,11 @@ EIGEN_STRONG_INLINE Packet8d psub<Packet8d>(const Packet8d& a, const Packet8d& b) { return _mm512_sub_pd(a, b); } +template <> +EIGEN_STRONG_INLINE Packet16i psub<Packet16i>(const Packet16i& a, + const Packet16i& b) { + return _mm512_sub_epi32(a, b); +} template <> EIGEN_STRONG_INLINE Packet16f pnegate(const Packet16f& a) { @@ -203,6 +221,11 @@ EIGEN_STRONG_INLINE Packet8d pmul<Packet8d>(const Packet8d& a, const Packet8d& b) { return _mm512_mul_pd(a, b); } +template <> +EIGEN_STRONG_INLINE Packet16i pmul<Packet16i>(const Packet16i& a, + const Packet16i& b) { + return _mm512_mul_epi32(a, b); +} template <> EIGEN_STRONG_INLINE Packet16f pdiv<Packet16f>(const Packet16f& a, @@ -215,7 +238,7 @@ EIGEN_STRONG_INLINE Packet8d pdiv<Packet8d>(const Packet8d& a, return _mm512_div_pd(a, b); } -#ifdef __FMA__ +#ifdef EIGEN_VECTORIZE_FMA template <> EIGEN_STRONG_INLINE Packet16f pmadd(const Packet16f& a, const Packet16f& b, const Packet16f& c) { @@ -254,30 +277,92 @@ EIGEN_STRONG_INLINE Packet8d pmax<Packet8d>(const Packet8d& a, return _mm512_max_pd(b, a); } -template <> -EIGEN_STRONG_INLINE Packet16f pand<Packet16f>(const Packet16f& a, - const Packet16f& b) { #ifdef EIGEN_VECTORIZE_AVX512DQ - return _mm512_and_ps(a, b); +template<int I_> EIGEN_STRONG_INLINE Packet8f extract256(Packet16f x) { return _mm512_extractf32x8_ps(x,I_); } +template<int I_> EIGEN_STRONG_INLINE Packet2d extract128(Packet8d x) { return _mm512_extractf64x2_pd(x,I_); } +EIGEN_STRONG_INLINE Packet16f cat256(Packet8f a, Packet8f b) { return _mm512_insertf32x8(_mm512_castps256_ps512(a),b,1); } #else - Packet16f res = _mm512_undefined_ps(); - Packet4f lane0_a = _mm512_extractf32x4_ps(a, 0); - Packet4f lane0_b = _mm512_extractf32x4_ps(b, 0); - res = _mm512_insertf32x4(res, _mm_and_ps(lane0_a, lane0_b), 0); +// AVX512F does not define _mm512_extractf32x8_ps to extract _m256 from _m512 +template<int I_> EIGEN_STRONG_INLINE Packet8f extract256(Packet16f x) { + return _mm256_castsi256_ps(_mm512_extracti64x4_epi64( _mm512_castps_si512(x),I_)); +} - Packet4f lane1_a = _mm512_extractf32x4_ps(a, 1); - Packet4f lane1_b = _mm512_extractf32x4_ps(b, 1); - res = _mm512_insertf32x4(res, _mm_and_ps(lane1_a, lane1_b), 1); +// AVX512F does not define _mm512_extractf64x2_pd to extract _m128 from _m512 +template<int I_> EIGEN_STRONG_INLINE Packet2d extract128(Packet8d x) { + return _mm_castsi128_pd(_mm512_extracti32x4_epi32( _mm512_castpd_si512(x),I_)); +} - Packet4f lane2_a = _mm512_extractf32x4_ps(a, 2); - Packet4f lane2_b = _mm512_extractf32x4_ps(b, 2); - res = _mm512_insertf32x4(res, _mm_and_ps(lane2_a, lane2_b), 2); +EIGEN_STRONG_INLINE Packet16f cat256(Packet8f a, Packet8f b) { + return _mm512_castsi512_ps(_mm512_inserti64x4(_mm512_castsi256_si512(_mm256_castps_si256(a)), + _mm256_castps_si256(b),1)); +} +#endif - Packet4f lane3_a = _mm512_extractf32x4_ps(a, 3); - Packet4f lane3_b = _mm512_extractf32x4_ps(b, 3); - res = _mm512_insertf32x4(res, _mm_and_ps(lane3_a, lane3_b), 3); +template<> EIGEN_STRONG_INLINE Packet16f pcmp_le(const Packet16f& a, const Packet16f& b) { + __mmask16 mask = _mm512_cmp_ps_mask(a, b, _CMP_LE_OQ); + return _mm512_castsi512_ps( + _mm512_mask_set1_epi32(_mm512_set1_epi32(0), mask, 0xffffffffu)); +} - return res; +template<> EIGEN_STRONG_INLINE Packet16f pcmp_lt(const Packet16f& a, const Packet16f& b) { + __mmask16 mask = _mm512_cmp_ps_mask(a, b, _CMP_LT_OQ); + return _mm512_castsi512_ps( + _mm512_mask_set1_epi32(_mm512_set1_epi32(0), mask, 0xffffffffu)); +} + +template<> EIGEN_STRONG_INLINE Packet16f pcmp_lt_or_nan(const Packet16f& a, const Packet16f& b) { + __mmask16 mask = _mm512_cmp_ps_mask(a, b, _CMP_NGT_UQ); + return _mm512_castsi512_ps( + _mm512_mask_set1_epi32(_mm512_set1_epi32(0), mask, 0xffffffffu)); +} + +template<> EIGEN_STRONG_INLINE Packet16i pcmp_eq(const Packet16i& a, const Packet16i& b) { + __mmask16 mask = _mm512_cmp_epi32_mask(a, b, _CMP_EQ_OQ); + return _mm512_mask_set1_epi32(_mm512_set1_epi32(0), mask, 0xffffffffu); +} + +template <> +EIGEN_STRONG_INLINE Packet16f pcmp_eq(const Packet16f& a, const Packet16f& b) { + __mmask16 mask = _mm512_cmp_ps_mask(a, b, _CMP_EQ_OQ); + return _mm512_castsi512_ps( + _mm512_mask_set1_epi32(_mm512_set1_epi32(0), mask, 0xffffffffu)); +} + +template <> +EIGEN_STRONG_INLINE Packet8d pcmp_eq(const Packet8d& a, const Packet8d& b) { + __mmask8 mask = _mm512_cmp_pd_mask(a, b, _CMP_EQ_OQ); + return _mm512_castsi512_pd( + _mm512_mask_set1_epi64(_mm512_set1_epi64(0), mask, 0xffffffffffffffffu)); +} + +template <> +EIGEN_STRONG_INLINE Packet16i ptrue<Packet16i>(const Packet16i& /*a*/) { + return _mm512_set1_epi32(0xffffffffu); +} + +template <> +EIGEN_STRONG_INLINE Packet16f ptrue<Packet16f>(const Packet16f& a) { + return _mm512_castsi512_ps(ptrue<Packet16i>(_mm512_castps_si512(a))); +} + +template <> +EIGEN_STRONG_INLINE Packet8d ptrue<Packet8d>(const Packet8d& a) { + return _mm512_castsi512_pd(ptrue<Packet16i>(_mm512_castpd_si512(a))); +} + +template <> +EIGEN_STRONG_INLINE Packet16i pand<Packet16i>(const Packet16i& a, + const Packet16i& b) { + return _mm512_and_si512(a,b); +} + +template <> +EIGEN_STRONG_INLINE Packet16f pand<Packet16f>(const Packet16f& a, + const Packet16f& b) { +#ifdef EIGEN_VECTORIZE_AVX512DQ + return _mm512_and_ps(a, b); +#else + return _mm512_castsi512_ps(pand(_mm512_castps_si512(a),_mm512_castps_si512(b))); #endif } template <> @@ -298,30 +383,18 @@ EIGEN_STRONG_INLINE Packet8d pand<Packet8d>(const Packet8d& a, return res; #endif } + +template <> +EIGEN_STRONG_INLINE Packet16i por<Packet16i>(const Packet16i& a, const Packet16i& b) { + return _mm512_or_si512(a, b); +} + template <> -EIGEN_STRONG_INLINE Packet16f por<Packet16f>(const Packet16f& a, - const Packet16f& b) { +EIGEN_STRONG_INLINE Packet16f por<Packet16f>(const Packet16f& a, const Packet16f& b) { #ifdef EIGEN_VECTORIZE_AVX512DQ return _mm512_or_ps(a, b); #else - Packet16f res = _mm512_undefined_ps(); - Packet4f lane0_a = _mm512_extractf32x4_ps(a, 0); - Packet4f lane0_b = _mm512_extractf32x4_ps(b, 0); - res = _mm512_insertf32x4(res, _mm_or_ps(lane0_a, lane0_b), 0); - - Packet4f lane1_a = _mm512_extractf32x4_ps(a, 1); - Packet4f lane1_b = _mm512_extractf32x4_ps(b, 1); - res = _mm512_insertf32x4(res, _mm_or_ps(lane1_a, lane1_b), 1); - - Packet4f lane2_a = _mm512_extractf32x4_ps(a, 2); - Packet4f lane2_b = _mm512_extractf32x4_ps(b, 2); - res = _mm512_insertf32x4(res, _mm_or_ps(lane2_a, lane2_b), 2); - - Packet4f lane3_a = _mm512_extractf32x4_ps(a, 3); - Packet4f lane3_b = _mm512_extractf32x4_ps(b, 3); - res = _mm512_insertf32x4(res, _mm_or_ps(lane3_a, lane3_b), 3); - - return res; + return _mm512_castsi512_ps(por(_mm512_castps_si512(a),_mm512_castps_si512(b))); #endif } @@ -331,109 +404,59 @@ EIGEN_STRONG_INLINE Packet8d por<Packet8d>(const Packet8d& a, #ifdef EIGEN_VECTORIZE_AVX512DQ return _mm512_or_pd(a, b); #else - Packet8d res = _mm512_undefined_pd(); - Packet4d lane0_a = _mm512_extractf64x4_pd(a, 0); - Packet4d lane0_b = _mm512_extractf64x4_pd(b, 0); - res = _mm512_insertf64x4(res, _mm256_or_pd(lane0_a, lane0_b), 0); - - Packet4d lane1_a = _mm512_extractf64x4_pd(a, 1); - Packet4d lane1_b = _mm512_extractf64x4_pd(b, 1); - res = _mm512_insertf64x4(res, _mm256_or_pd(lane1_a, lane1_b), 1); - - return res; + return _mm512_castsi512_pd(por(_mm512_castpd_si512(a),_mm512_castpd_si512(b))); #endif } template <> -EIGEN_STRONG_INLINE Packet16f pxor<Packet16f>(const Packet16f& a, - const Packet16f& b) { +EIGEN_STRONG_INLINE Packet16i pxor<Packet16i>(const Packet16i& a, const Packet16i& b) { + return _mm512_xor_si512(a, b); +} + +template <> +EIGEN_STRONG_INLINE Packet16f pxor<Packet16f>(const Packet16f& a, const Packet16f& b) { #ifdef EIGEN_VECTORIZE_AVX512DQ return _mm512_xor_ps(a, b); #else - Packet16f res = _mm512_undefined_ps(); - Packet4f lane0_a = _mm512_extractf32x4_ps(a, 0); - Packet4f lane0_b = _mm512_extractf32x4_ps(b, 0); - res = _mm512_insertf32x4(res, _mm_xor_ps(lane0_a, lane0_b), 0); - - Packet4f lane1_a = _mm512_extractf32x4_ps(a, 1); - Packet4f lane1_b = _mm512_extractf32x4_ps(b, 1); - res = _mm512_insertf32x4(res, _mm_xor_ps(lane1_a, lane1_b), 1); - - Packet4f lane2_a = _mm512_extractf32x4_ps(a, 2); - Packet4f lane2_b = _mm512_extractf32x4_ps(b, 2); - res = _mm512_insertf32x4(res, _mm_xor_ps(lane2_a, lane2_b), 2); - - Packet4f lane3_a = _mm512_extractf32x4_ps(a, 3); - Packet4f lane3_b = _mm512_extractf32x4_ps(b, 3); - res = _mm512_insertf32x4(res, _mm_xor_ps(lane3_a, lane3_b), 3); - - return res; + return _mm512_castsi512_ps(pxor(_mm512_castps_si512(a),_mm512_castps_si512(b))); #endif } + template <> -EIGEN_STRONG_INLINE Packet8d pxor<Packet8d>(const Packet8d& a, - const Packet8d& b) { +EIGEN_STRONG_INLINE Packet8d pxor<Packet8d>(const Packet8d& a, const Packet8d& b) { #ifdef EIGEN_VECTORIZE_AVX512DQ return _mm512_xor_pd(a, b); #else - Packet8d res = _mm512_undefined_pd(); - Packet4d lane0_a = _mm512_extractf64x4_pd(a, 0); - Packet4d lane0_b = _mm512_extractf64x4_pd(b, 0); - res = _mm512_insertf64x4(res, _mm256_xor_pd(lane0_a, lane0_b), 0); - - Packet4d lane1_a = _mm512_extractf64x4_pd(a, 1); - Packet4d lane1_b = _mm512_extractf64x4_pd(b, 1); - res = _mm512_insertf64x4(res, _mm256_xor_pd(lane1_a, lane1_b), 1); - - return res; + return _mm512_castsi512_pd(pxor(_mm512_castpd_si512(a),_mm512_castpd_si512(b))); #endif } template <> -EIGEN_STRONG_INLINE Packet16f pandnot<Packet16f>(const Packet16f& a, - const Packet16f& b) { +EIGEN_STRONG_INLINE Packet16i pandnot<Packet16i>(const Packet16i& a, const Packet16i& b) { + return _mm512_andnot_si512(b, a); +} + +template <> +EIGEN_STRONG_INLINE Packet16f pandnot<Packet16f>(const Packet16f& a, const Packet16f& b) { #ifdef EIGEN_VECTORIZE_AVX512DQ - return _mm512_andnot_ps(a, b); + return _mm512_andnot_ps(b, a); #else - Packet16f res = _mm512_undefined_ps(); - Packet4f lane0_a = _mm512_extractf32x4_ps(a, 0); - Packet4f lane0_b = _mm512_extractf32x4_ps(b, 0); - res = _mm512_insertf32x4(res, _mm_andnot_ps(lane0_a, lane0_b), 0); - - Packet4f lane1_a = _mm512_extractf32x4_ps(a, 1); - Packet4f lane1_b = _mm512_extractf32x4_ps(b, 1); - res = _mm512_insertf32x4(res, _mm_andnot_ps(lane1_a, lane1_b), 1); - - Packet4f lane2_a = _mm512_extractf32x4_ps(a, 2); - Packet4f lane2_b = _mm512_extractf32x4_ps(b, 2); - res = _mm512_insertf32x4(res, _mm_andnot_ps(lane2_a, lane2_b), 2); - - Packet4f lane3_a = _mm512_extractf32x4_ps(a, 3); - Packet4f lane3_b = _mm512_extractf32x4_ps(b, 3); - res = _mm512_insertf32x4(res, _mm_andnot_ps(lane3_a, lane3_b), 3); - - return res; + return _mm512_castsi512_ps(pandnot(_mm512_castps_si512(a),_mm512_castps_si512(b))); #endif } template <> -EIGEN_STRONG_INLINE Packet8d pandnot<Packet8d>(const Packet8d& a, - const Packet8d& b) { +EIGEN_STRONG_INLINE Packet8d pandnot<Packet8d>(const Packet8d& a,const Packet8d& b) { #ifdef EIGEN_VECTORIZE_AVX512DQ - return _mm512_andnot_pd(a, b); + return _mm512_andnot_pd(b, a); #else - Packet8d res = _mm512_undefined_pd(); - Packet4d lane0_a = _mm512_extractf64x4_pd(a, 0); - Packet4d lane0_b = _mm512_extractf64x4_pd(b, 0); - res = _mm512_insertf64x4(res, _mm256_andnot_pd(lane0_a, lane0_b), 0); - - Packet4d lane1_a = _mm512_extractf64x4_pd(a, 1); - Packet4d lane1_b = _mm512_extractf64x4_pd(b, 1); - res = _mm512_insertf64x4(res, _mm256_andnot_pd(lane1_a, lane1_b), 1); - - return res; + return _mm512_castsi512_pd(pandnot(_mm512_castpd_si512(a),_mm512_castpd_si512(b))); #endif } +template<int N> EIGEN_STRONG_INLINE Packet16i pshiftleft(Packet16i a) { + return _mm512_slli_epi32(a, N); +} + template <> EIGEN_STRONG_INLINE Packet16f pload<Packet16f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm512_load_ps(from); @@ -475,6 +498,7 @@ EIGEN_STRONG_INLINE Packet16f ploaddup<Packet16f>(const float* from) { } #ifdef EIGEN_VECTORIZE_AVX512DQ +// FIXME: this does not look optimal, better load a Packet4d and shuffle... // Loads 4 doubles from memory a returns the packet {a0, a0 a1, a1, a2, a2, a3, // a3} template <> @@ -502,21 +526,17 @@ EIGEN_STRONG_INLINE Packet8d ploaddup<Packet8d>(const double* from) { // {a0, a0 a0, a0, a1, a1, a1, a1, a2, a2, a2, a2, a3, a3, a3, a3} template <> EIGEN_STRONG_INLINE Packet16f ploadquad<Packet16f>(const float* from) { - Packet16f tmp = _mm512_undefined_ps(); - tmp = _mm512_insertf32x4(tmp, _mm_load_ps1(from), 0); - tmp = _mm512_insertf32x4(tmp, _mm_load_ps1(from + 1), 1); - tmp = _mm512_insertf32x4(tmp, _mm_load_ps1(from + 2), 2); - tmp = _mm512_insertf32x4(tmp, _mm_load_ps1(from + 3), 3); - return tmp; + Packet16f tmp = _mm512_castps128_ps512(ploadu<Packet4f>(from)); + const Packet16i scatter_mask = _mm512_set_epi32(3,3,3,3, 2,2,2,2, 1,1,1,1, 0,0,0,0); + return _mm512_permutexvar_ps(scatter_mask, tmp); } + // Loads 2 doubles from memory a returns the packet // {a0, a0 a0, a0, a1, a1, a1, a1} template <> EIGEN_STRONG_INLINE Packet8d ploadquad<Packet8d>(const double* from) { - __m128d tmp0 = _mm_load_pd1(from); - __m256d lane0 = _mm256_broadcastsd_pd(tmp0); - __m128d tmp1 = _mm_load_pd1(from + 1); - __m256d lane1 = _mm256_broadcastsd_pd(tmp1); + __m256d lane0 = _mm256_set1_pd(*from); + __m256d lane1 = _mm256_set1_pd(*(from+1)); __m512d tmp = _mm512_undefined_pd(); tmp = _mm512_insertf64x4(tmp, lane0, 0); return _mm512_insertf64x4(tmp, lane1, 1); @@ -981,6 +1001,13 @@ EIGEN_STRONG_INLINE double predux_max<Packet8d>(const Packet8d& a) { return pfirst(_mm256_max_pd(res, _mm256_shuffle_pd(res, res, 1))); } +template<> EIGEN_STRONG_INLINE bool predux_any(const Packet16f& x) +{ + Packet16i xi = _mm512_castps_si512(x); + __mmask16 tmp = _mm512_test_epi32_mask(xi,xi); + return !_mm512_kortestz(tmp,tmp); +} + template <int Offset> struct palign_impl<Offset, Packet16f> { static EIGEN_STRONG_INLINE void run(Packet16f& first, @@ -1322,6 +1349,22 @@ template<> EIGEN_STRONG_INLINE Packet8d pinsertlast(const Packet8d& a, double b) return _mm512_mask_broadcastsd_pd(a, (1<<7), _mm_load_sd(&b)); } +template<> EIGEN_STRONG_INLINE Packet16i pcast<Packet16f, Packet16i>(const Packet16f& a) { + return _mm512_cvttps_epi32(a); +} + +template<> EIGEN_STRONG_INLINE Packet16f pcast<Packet16i, Packet16f>(const Packet16i& a) { + return _mm512_cvtepi32_ps(a); +} + +template<> EIGEN_STRONG_INLINE Packet16i preinterpret<Packet16i,Packet16f>(const Packet16f& a) { + return _mm512_castps_si512(a); +} + +template<> EIGEN_STRONG_INLINE Packet16f preinterpret<Packet16f,Packet16i>(const Packet16i& a) { + return _mm512_castsi512_ps(a); +} + } // end namespace internal } // end namespace Eigen diff --git a/Eigen/src/Core/arch/AltiVec/Complex.h b/Eigen/src/Core/arch/AltiVec/Complex.h index 3e665730c..440d058d8 100644 --- a/Eigen/src/Core/arch/AltiVec/Complex.h +++ b/Eigen/src/Core/arch/AltiVec/Complex.h @@ -60,7 +60,7 @@ template<> struct packet_traits<std::complex<float> > : default_packet_traits }; }; -template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16}; typedef Packet2cf half; }; +template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16, vectorizable=true}; typedef Packet2cf half; }; template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>& from) { @@ -82,14 +82,14 @@ template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<f template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from, Index stride) { - std::complex<float> EIGEN_ALIGN16 af[2]; + EIGEN_ALIGN16 std::complex<float> af[2]; af[0] = from[0*stride]; af[1] = from[1*stride]; return pload<Packet2cf>(af); } template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, Index stride) { - std::complex<float> EIGEN_ALIGN16 af[2]; + EIGEN_ALIGN16 std::complex<float> af[2]; pstore<std::complex<float> >((std::complex<float> *) af, from); to[0*stride] = af[0]; to[1*stride] = af[1]; @@ -128,7 +128,7 @@ template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::co template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a) { - std::complex<float> EIGEN_ALIGN16 res[2]; + EIGEN_ALIGN16 std::complex<float> res[2]; pstore((float *)&res, a.v); return res[0]; @@ -286,7 +286,7 @@ template<> struct packet_traits<std::complex<double> > : default_packet_traits }; }; -template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16}; typedef Packet1cd half; }; +template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16, vectorizable=true}; typedef Packet1cd half; }; template<> EIGEN_STRONG_INLINE Packet1cd pload <Packet1cd>(const std::complex<double>* from) { return Packet1cd(pload<Packet2d>((const double*)from)); } template<> EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from) { return Packet1cd(ploadu<Packet2d>((const double*)from)); } @@ -298,14 +298,14 @@ template<> EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<dou template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather<std::complex<double>, Packet1cd>(const std::complex<double>* from, Index stride) { - std::complex<double> EIGEN_ALIGN16 af[2]; + EIGEN_ALIGN16 std::complex<double> af[2]; af[0] = from[0*stride]; af[1] = from[1*stride]; return pload<Packet1cd>(af); } template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet1cd>(std::complex<double>* to, const Packet1cd& from, Index stride) { - std::complex<double> EIGEN_ALIGN16 af[2]; + EIGEN_ALIGN16 std::complex<double> af[2]; pstore<std::complex<double> >(af, from); to[0*stride] = af[0]; to[1*stride] = af[1]; @@ -345,7 +345,7 @@ template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::c template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Packet1cd& a) { - std::complex<double> EIGEN_ALIGN16 res[2]; + EIGEN_ALIGN16 std::complex<double> res[2]; pstore<std::complex<double> >(res, a); return res[0]; diff --git a/Eigen/src/Core/arch/AltiVec/MathFunctions.h b/Eigen/src/Core/arch/AltiVec/MathFunctions.h index c5e4bede7..81097e668 100644 --- a/Eigen/src/Core/arch/AltiVec/MathFunctions.h +++ b/Eigen/src/Core/arch/AltiVec/MathFunctions.h @@ -9,191 +9,37 @@ // 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/. -/* The sin, cos, exp, and log functions of this file come from - * Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/ - */ - #ifndef EIGEN_MATH_FUNCTIONS_ALTIVEC_H #define EIGEN_MATH_FUNCTIONS_ALTIVEC_H +#include "../Default/GenericPacketMathFunctions.h" + namespace Eigen { namespace internal { -static _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f); -static _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f); -static _EIGEN_DECLARE_CONST_Packet4i(0x7f, 0x7f); -static _EIGEN_DECLARE_CONST_Packet4i(23, 23); - -static _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(inv_mant_mask, ~0x7f800000); - -/* the smallest non denormalized float number */ -static _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(min_norm_pos, 0x00800000); -static _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(minus_inf, 0xff800000); // -1.f/0.f -static _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(minus_nan, 0xffffffff); - -/* natural logarithm computed for 4 simultaneous float - return NaN for x <= 0 -*/ -static _EIGEN_DECLARE_CONST_Packet4f(cephes_SQRTHF, 0.707106781186547524f); -static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p0, 7.0376836292E-2f); -static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p1, - 1.1514610310E-1f); -static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p2, 1.1676998740E-1f); -static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p3, - 1.2420140846E-1f); -static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p4, + 1.4249322787E-1f); -static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p5, - 1.6668057665E-1f); -static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p6, + 2.0000714765E-1f); -static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p7, - 2.4999993993E-1f); -static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p8, + 3.3333331174E-1f); -static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q1, -2.12194440e-4f); -static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q2, 0.693359375f); - -static _EIGEN_DECLARE_CONST_Packet4f(exp_hi, 88.3762626647950f); -static _EIGEN_DECLARE_CONST_Packet4f(exp_lo, -88.3762626647949f); - -static _EIGEN_DECLARE_CONST_Packet4f(cephes_LOG2EF, 1.44269504088896341f); -static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C1, 0.693359375f); -static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C2, -2.12194440e-4f); - -static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p0, 1.9875691500E-4f); -static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p1, 1.3981999507E-3f); -static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p2, 8.3334519073E-3f); -static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p3, 4.1665795894E-2f); -static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459E-1f); -static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201E-1f); - -#ifdef __VSX__ -static _EIGEN_DECLARE_CONST_Packet2d(1 , 1.0); -static _EIGEN_DECLARE_CONST_Packet2d(2 , 2.0); -static _EIGEN_DECLARE_CONST_Packet2d(half, 0.5); - -static _EIGEN_DECLARE_CONST_Packet2d(exp_hi, 709.437); -static _EIGEN_DECLARE_CONST_Packet2d(exp_lo, -709.436139303); - -static _EIGEN_DECLARE_CONST_Packet2d(cephes_LOG2EF, 1.4426950408889634073599); - -static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p0, 1.26177193074810590878e-4); -static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p1, 3.02994407707441961300e-2); -static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p2, 9.99999999999999999910e-1); - -static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q0, 3.00198505138664455042e-6); -static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q1, 2.52448340349684104192e-3); -static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q2, 2.27265548208155028766e-1); -static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q3, 2.00000000000000000009e0); - -static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C1, 0.693145751953125); -static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6); - -#ifdef __POWER8_VECTOR__ -static Packet2l p2l_1023 = { 1023, 1023 }; -static Packet2ul p2ul_52 = { 52, 52 }; -#endif - -#endif - template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f plog<Packet4f>(const Packet4f& _x) { - Packet4f x = _x; - - Packet4i emm0; - - /* isvalid_mask is 0 if x < 0 or x is NaN. */ - Packet4ui isvalid_mask = reinterpret_cast<Packet4ui>(vec_cmpge(x, p4f_ZERO)); - Packet4ui iszero_mask = reinterpret_cast<Packet4ui>(vec_cmpeq(x, p4f_ZERO)); - - x = pmax(x, p4f_min_norm_pos); /* cut off denormalized stuff */ - emm0 = vec_sr(reinterpret_cast<Packet4i>(x), - reinterpret_cast<Packet4ui>(p4i_23)); - - /* keep only the fractional part */ - x = pand(x, p4f_inv_mant_mask); - x = por(x, p4f_half); - - emm0 = psub(emm0, p4i_0x7f); - Packet4f e = padd(vec_ctf(emm0, 0), p4f_1); - - /* part2: - if( x < SQRTHF ) { - e -= 1; - x = x + x - 1.0; - } else { x = x - 1.0; } - */ - Packet4f mask = reinterpret_cast<Packet4f>(vec_cmplt(x, p4f_cephes_SQRTHF)); - Packet4f tmp = pand(x, mask); - x = psub(x, p4f_1); - e = psub(e, pand(p4f_1, mask)); - x = padd(x, tmp); - - Packet4f x2 = pmul(x,x); - Packet4f x3 = pmul(x2,x); - - Packet4f y, y1, y2; - y = pmadd(p4f_cephes_log_p0, x, p4f_cephes_log_p1); - y1 = pmadd(p4f_cephes_log_p3, x, p4f_cephes_log_p4); - y2 = pmadd(p4f_cephes_log_p6, x, p4f_cephes_log_p7); - y = pmadd(y , x, p4f_cephes_log_p2); - y1 = pmadd(y1, x, p4f_cephes_log_p5); - y2 = pmadd(y2, x, p4f_cephes_log_p8); - y = pmadd(y, x3, y1); - y = pmadd(y, x3, y2); - y = pmul(y, x3); - - y1 = pmul(e, p4f_cephes_log_q1); - tmp = pmul(x2, p4f_half); - y = padd(y, y1); - x = psub(x, tmp); - y2 = pmul(e, p4f_cephes_log_q2); - x = padd(x, y); - x = padd(x, y2); - // negative arg will be NAN, 0 will be -INF - x = vec_sel(x, p4f_minus_inf, iszero_mask); - x = vec_sel(p4f_minus_nan, x, isvalid_mask); - return x; + return plog_float(_x); } template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f pexp<Packet4f>(const Packet4f& _x) { - Packet4f x = _x; - - Packet4f tmp, fx; - Packet4i emm0; - - // clamp x - x = pmax(pmin(x, p4f_exp_hi), p4f_exp_lo); - - // express exp(x) as exp(g + n*log(2)) - fx = pmadd(x, p4f_cephes_LOG2EF, p4f_half); - - fx = pfloor(fx); - - tmp = pmul(fx, p4f_cephes_exp_C1); - Packet4f z = pmul(fx, p4f_cephes_exp_C2); - x = psub(x, tmp); - x = psub(x, z); - - z = pmul(x,x); - - Packet4f y = p4f_cephes_exp_p0; - y = pmadd(y, x, p4f_cephes_exp_p1); - y = pmadd(y, x, p4f_cephes_exp_p2); - y = pmadd(y, x, p4f_cephes_exp_p3); - y = pmadd(y, x, p4f_cephes_exp_p4); - y = pmadd(y, x, p4f_cephes_exp_p5); - y = pmadd(y, z, x); - y = padd(y, p4f_1); + return pexp_float(_x); +} - // build 2^n - emm0 = vec_cts(fx, 0); - emm0 = vec_add(emm0, p4i_0x7f); - emm0 = vec_sl(emm0, reinterpret_cast<Packet4ui>(p4i_23)); +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet4f psin<Packet4f>(const Packet4f& _x) +{ + return psin_float(_x); +} - // Altivec's max & min operators just drop silent NaNs. Check NaNs in - // inputs and return them unmodified. - Packet4ui isnumber_mask = reinterpret_cast<Packet4ui>(vec_cmpeq(_x, _x)); - return vec_sel(_x, pmax(pmul(y, reinterpret_cast<Packet4f>(emm0)), _x), - isnumber_mask); +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet4f pcos<Packet4f>(const Packet4f& _x) +{ + return pcos_float(_x); } #ifndef EIGEN_COMP_CLANG @@ -225,93 +71,10 @@ Packet2d psqrt<Packet2d>(const Packet2d& x) return vec_sqrt(x); } -// VSX support varies between different compilers and even different -// versions of the same compiler. For gcc version >= 4.9.3, we can use -// vec_cts to efficiently convert Packet2d to Packet2l. Otherwise, use -// a slow version that works with older compilers. -// Update: apparently vec_cts/vec_ctf intrinsics for 64-bit doubles -// are buggy, https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70963 -static inline Packet2l ConvertToPacket2l(const Packet2d& x) { -#if EIGEN_GNUC_AT_LEAST(5, 4) || \ - (EIGEN_GNUC_AT(6, 1) && __GNUC_PATCHLEVEL__ >= 1) - return vec_cts(x, 0); // TODO: check clang version. -#else - double tmp[2]; - memcpy(tmp, &x, sizeof(tmp)); - Packet2l l = { static_cast<long long>(tmp[0]), - static_cast<long long>(tmp[1]) }; - return l; -#endif -} - template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet2d pexp<Packet2d>(const Packet2d& _x) { - Packet2d x = _x; - - Packet2d tmp, fx; - Packet2l emm0; - - // clamp x - x = pmax(pmin(x, p2d_exp_hi), p2d_exp_lo); - - /* express exp(x) as exp(g + n*log(2)) */ - fx = pmadd(x, p2d_cephes_LOG2EF, p2d_half); - - fx = pfloor(fx); - - tmp = pmul(fx, p2d_cephes_exp_C1); - Packet2d z = pmul(fx, p2d_cephes_exp_C2); - x = psub(x, tmp); - x = psub(x, z); - - Packet2d x2 = pmul(x,x); - - Packet2d px = p2d_cephes_exp_p0; - px = pmadd(px, x2, p2d_cephes_exp_p1); - px = pmadd(px, x2, p2d_cephes_exp_p2); - px = pmul (px, x); - - Packet2d qx = p2d_cephes_exp_q0; - qx = pmadd(qx, x2, p2d_cephes_exp_q1); - qx = pmadd(qx, x2, p2d_cephes_exp_q2); - qx = pmadd(qx, x2, p2d_cephes_exp_q3); - - x = pdiv(px,psub(qx,px)); - x = pmadd(p2d_2,x,p2d_1); - - // build 2^n - emm0 = ConvertToPacket2l(fx); - -#ifdef __POWER8_VECTOR__ - emm0 = vec_add(emm0, p2l_1023); - emm0 = vec_sl(emm0, p2ul_52); -#else - // Code is a bit complex for POWER7. There is actually a - // vec_xxsldi intrinsic but it is not supported by some gcc versions. - // So we shift (52-32) bits and do a word swap with zeros. - _EIGEN_DECLARE_CONST_Packet4i(1023, 1023); - _EIGEN_DECLARE_CONST_Packet4i(20, 20); // 52 - 32 - - Packet4i emm04i = reinterpret_cast<Packet4i>(emm0); - emm04i = vec_add(emm04i, p4i_1023); - emm04i = vec_sl(emm04i, reinterpret_cast<Packet4ui>(p4i_20)); - static const Packet16uc perm = { - 0x14, 0x15, 0x16, 0x17, 0x00, 0x01, 0x02, 0x03, - 0x1c, 0x1d, 0x1e, 0x1f, 0x08, 0x09, 0x0a, 0x0b }; -#ifdef _BIG_ENDIAN - emm0 = reinterpret_cast<Packet2l>(vec_perm(p4i_ZERO, emm04i, perm)); -#else - emm0 = reinterpret_cast<Packet2l>(vec_perm(emm04i, p4i_ZERO, perm)); -#endif - -#endif - - // Altivec's max & min operators just drop silent NaNs. Check NaNs in - // inputs and return them unmodified. - Packet2ul isnumber_mask = reinterpret_cast<Packet2ul>(vec_cmpeq(_x, _x)); - return vec_sel(_x, pmax(pmul(x, reinterpret_cast<Packet2d>(emm0)), _x), - isnumber_mask); + return pexp_double(_x); } #endif diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h index 7f4e90f75..9535724eb 100755 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -146,9 +146,9 @@ template<> struct packet_traits<float> : default_packet_traits HasMin = 1, HasMax = 1, HasAbs = 1, - HasSin = 0, - HasCos = 0, - HasLog = 0, + HasSin = EIGEN_FAST_MATH, + HasCos = EIGEN_FAST_MATH, + HasLog = 1, HasExp = 1, #ifdef __VSX__ HasSqrt = 1, @@ -187,8 +187,19 @@ template<> struct packet_traits<int> : default_packet_traits }; -template<> struct unpacket_traits<Packet4f> { typedef float type; enum {size=4, alignment=Aligned16}; typedef Packet4f half; }; -template<> struct unpacket_traits<Packet4i> { typedef int type; enum {size=4, alignment=Aligned16}; typedef Packet4i half; }; +template<> struct unpacket_traits<Packet4f> +{ + typedef float type; + typedef Packet4f half; + typedef Packet4i integer_packet; + enum {size=4, alignment=Aligned16, vectorizable=true}; +}; +template<> struct unpacket_traits<Packet4i> +{ + typedef int type; + typedef Packet4i half; + enum {size=4, alignment=Aligned16, vectorizable=false}; +}; inline std::ostream & operator <<(std::ostream & s, const Packet16uc & v) { @@ -285,6 +296,11 @@ template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) { Packet4i v = {from, from, from, from}; return v; } + +template<> EIGEN_STRONG_INLINE Packet4f pset1frombits<Packet4f>(unsigned int from) { + return reinterpret_cast<Packet4f>(pset1<Packet4i>(from)); +} + template<> EIGEN_STRONG_INLINE void pbroadcast4<Packet4f>(const float *a, Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3) @@ -308,7 +324,7 @@ pbroadcast4<Packet4i>(const int *a, template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride) { - float EIGEN_ALIGN16 af[4]; + EIGEN_ALIGN16 float af[4]; af[0] = from[0*stride]; af[1] = from[1*stride]; af[2] = from[2*stride]; @@ -317,7 +333,7 @@ template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const floa } template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* from, Index stride) { - int EIGEN_ALIGN16 ai[4]; + EIGEN_ALIGN16 int ai[4]; ai[0] = from[0*stride]; ai[1] = from[1*stride]; ai[2] = from[2*stride]; @@ -326,7 +342,7 @@ template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* f } template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride) { - float EIGEN_ALIGN16 af[4]; + EIGEN_ALIGN16 float af[4]; pstore<float>(af, from); to[0*stride] = af[0]; to[1*stride] = af[1]; @@ -335,7 +351,7 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, co } template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const Packet4i& from, Index stride) { - int EIGEN_ALIGN16 ai[4]; + EIGEN_ALIGN16 int ai[4]; pstore<int>((int *)ai, from); to[0*stride] = ai[0]; to[1*stride] = ai[1]; @@ -414,6 +430,15 @@ template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const } template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_max(a, b); } +template<> EIGEN_STRONG_INLINE Packet4f pcmp_le(const Packet4f& a, const Packet4f& b) { return reinterpret_cast<Packet4f>(vec_cmple(a,b)); } +template<> EIGEN_STRONG_INLINE Packet4f pcmp_lt(const Packet4f& a, const Packet4f& b) { return reinterpret_cast<Packet4f>(vec_cmplt(a,b)); } +template<> EIGEN_STRONG_INLINE Packet4f pcmp_eq(const Packet4f& a, const Packet4f& b) { return reinterpret_cast<Packet4f>(vec_cmpeq(a,b)); } +template<> EIGEN_STRONG_INLINE Packet4f pcmp_lt_or_nan(const Packet4f& a, const Packet4f& b) { + Packet4f c = reinterpret_cast<Packet4f>(vec_cmpge(a,b)); + return vec_nor(c,c); +} +template<> EIGEN_STRONG_INLINE Packet4i pcmp_eq(const Packet4i& a, const Packet4i& b) { return reinterpret_cast<Packet4i>(vec_cmpeq(a,b)); } + template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_and(a, b); } template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, b); } @@ -426,6 +451,10 @@ template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_and(a, vec_nor(b, b)); } template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, vec_nor(b, b)); } +template<> EIGEN_STRONG_INLINE Packet4f pselect(const Packet4f& mask, const Packet4f& a, const Packet4f& b) { + return vec_sel(b, a, mask); +} + template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a) { return vec_round(a); } template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const Packet4f& a) { return vec_ceil(a); } template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) { return vec_floor(a); } @@ -536,8 +565,8 @@ template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { EIGEN_PPC_PREFETCH(addr); } template<> EIGEN_STRONG_INLINE void prefetch<int>(const int* addr) { EIGEN_PPC_PREFETCH(addr); } -template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { float EIGEN_ALIGN16 x; vec_ste(a, 0, &x); return x; } -template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { int EIGEN_ALIGN16 x; vec_ste(a, 0, &x); return x; } +template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { EIGEN_ALIGN16 float x; vec_ste(a, 0, &x); return x; } +template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { EIGEN_ALIGN16 int x; vec_ste(a, 0, &x); return x; } template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) { @@ -550,6 +579,19 @@ template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a) { return vec_abs(a); } template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) { return vec_abs(a); } +template<int N> EIGEN_STRONG_INLINE Packet4i pshiftright(Packet4i a) +{ return vec_sr(a,reinterpret_cast<Packet4ui>(pset1<Packet4i>(N))); } +template<int N> EIGEN_STRONG_INLINE Packet4i pshiftleft(Packet4i a) +{ return vec_sl(a,reinterpret_cast<Packet4ui>(pset1<Packet4i>(N))); } + +template<> EIGEN_STRONG_INLINE Packet4f pfrexp<Packet4f>(const Packet4f& a, Packet4f& exponent) { + return pfrexp_float(a,exponent); +} + +template<> EIGEN_STRONG_INLINE Packet4f pldexp<Packet4f>(const Packet4f& a, const Packet4f& exponent) { + return pldexp_float(a,exponent); +} + template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a) { Packet4f b, sum; @@ -678,6 +720,11 @@ template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a) return pfirst(res); } +template<> EIGEN_STRONG_INLINE bool predux_any(const Packet4f& x) +{ + return vec_any_ne(x, pzero(x)); +} + template<int Offset> struct palign_impl<Offset,Packet4f> { @@ -771,6 +818,43 @@ template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, cons } +template <> +struct type_casting_traits<float, int> { + enum { + VectorizedCast = 1, + SrcCoeffRatio = 1, + TgtCoeffRatio = 1 + }; +}; + +template <> +struct type_casting_traits<int, float> { + enum { + VectorizedCast = 1, + SrcCoeffRatio = 1, + TgtCoeffRatio = 1 + }; +}; + + +template<> EIGEN_STRONG_INLINE Packet4i pcast<Packet4f, Packet4i>(const Packet4f& a) { + return vec_cts(a,0); +} + +template<> EIGEN_STRONG_INLINE Packet4f pcast<Packet4i, Packet4f>(const Packet4i& a) { + return vec_ctf(a,0); +} + +template<> EIGEN_STRONG_INLINE Packet4i preinterpret<Packet4i,Packet4f>(const Packet4f& a) { + return reinterpret_cast<Packet4i>(a); +} + +template<> EIGEN_STRONG_INLINE Packet4f preinterpret<Packet4f,Packet4i>(const Packet4i& a) { + return reinterpret_cast<Packet4f>(a); +} + + + //---------- double ---------- #ifdef __VSX__ typedef __vector double Packet2d; @@ -837,7 +921,7 @@ template<> struct packet_traits<double> : default_packet_traits }; }; -template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16}; typedef Packet2d half; }; +template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16, vectorizable=true}; typedef Packet2d half; }; inline std::ostream & operator <<(std::ostream & s, const Packet2l & v) { @@ -901,14 +985,14 @@ pbroadcast4<Packet2d>(const double *a, template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride) { - double EIGEN_ALIGN16 af[2]; + EIGEN_ALIGN16 double af[2]; af[0] = from[0*stride]; af[1] = from[1*stride]; return pload<Packet2d>(af); } template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride) { - double EIGEN_ALIGN16 af[2]; + EIGEN_ALIGN16 double af[2]; pstore<double>(af, from); to[0*stride] = af[0]; to[1*stride] = af[1]; @@ -980,7 +1064,7 @@ template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d& template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { EIGEN_PPC_PREFETCH(addr); } -template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { double EIGEN_ALIGN16 x[2]; pstore<double>(x, a); return x[0]; } +template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { EIGEN_ALIGN16 double x[2]; pstore<double>(x, a); return x[0]; } template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a) { @@ -988,6 +1072,59 @@ template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a) } template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a) { return vec_abs(a); } +// VSX support varies between different compilers and even different +// versions of the same compiler. For gcc version >= 4.9.3, we can use +// vec_cts to efficiently convert Packet2d to Packet2l. Otherwise, use +// a slow version that works with older compilers. +// Update: apparently vec_cts/vec_ctf intrinsics for 64-bit doubles +// are buggy, https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70963 +static inline Packet2l ConvertToPacket2l(const Packet2d& x) { +#if EIGEN_GNUC_AT_LEAST(5, 4) || \ + (EIGEN_GNUC_AT(6, 1) && __GNUC_PATCHLEVEL__ >= 1) + return vec_cts(x, 0); // TODO: check clang version. +#else + double tmp[2]; + memcpy(tmp, &x, sizeof(tmp)); + Packet2l l = { static_cast<long long>(tmp[0]), + static_cast<long long>(tmp[1]) }; + return l; +#endif +} + +template<> EIGEN_STRONG_INLINE Packet2d pldexp<Packet2d>(const Packet2d& a, const Packet2d& exponent) { + + // build 2^n + Packet2l emm0 = ConvertToPacket2l(exponent); + +#ifdef __POWER8_VECTOR__ + const Packet2l p2l_1023 = { 1023, 1023 }; + const Packet2ul p2ul_52 = { 52, 52 }; + emm0 = vec_add(emm0, p2l_1023); + emm0 = vec_sl(emm0, p2ul_52); +#else + // Code is a bit complex for POWER7. There is actually a + // vec_xxsldi intrinsic but it is not supported by some gcc versions. + // So we shift (52-32) bits and do a word swap with zeros. + const Packet4i p4i_1023 = pset1<Packet4i>(1023); + const Packet4i p4i_20 = pset1<Packet4i>(20); // 52 - 32 + + Packet4i emm04i = reinterpret_cast<Packet4i>(emm0); + emm04i = vec_add(emm04i, p4i_1023); + emm04i = vec_sl(emm04i, reinterpret_cast<Packet4ui>(p4i_20)); + static const Packet16uc perm = { + 0x14, 0x15, 0x16, 0x17, 0x00, 0x01, 0x02, 0x03, + 0x1c, 0x1d, 0x1e, 0x1f, 0x08, 0x09, 0x0a, 0x0b }; +#ifdef _BIG_ENDIAN + emm0 = reinterpret_cast<Packet2l>(vec_perm(p4i_ZERO, emm04i, perm)); +#else + emm0 = reinterpret_cast<Packet2l>(vec_perm(emm04i, p4i_ZERO, perm)); +#endif + +#endif + + return pmul(a, reinterpret_cast<Packet2d>(emm0)); +} + template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) { Packet2d b, sum; diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h new file mode 100644 index 000000000..452b4c806 --- /dev/null +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -0,0 +1,471 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2007 Julien Pommier +// Copyright (C) 2014 Pedro Gonnet (pedro.gonnet@gmail.com) +// Copyright (C) 2009-2019 Gael Guennebaud <gael.guennebaud@inria.fr> +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +/* The exp and log functions of this file initially come from + * Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/ + */ + +namespace Eigen { +namespace internal { + +template<typename Packet> EIGEN_STRONG_INLINE Packet +pfrexp_float(const Packet& a, Packet& exponent) { + typedef typename unpacket_traits<Packet>::integer_packet PacketI; + const Packet cst_126f = pset1<Packet>(126.0f); + const Packet cst_half = pset1<Packet>(0.5f); + const Packet cst_inv_mant_mask = pset1frombits<Packet>(~0x7f800000u); + exponent = psub(pcast<PacketI,Packet>(pshiftright<23>(preinterpret<PacketI>(a))), cst_126f); + return por(pand(a, cst_inv_mant_mask), cst_half); +} + +template<typename Packet> EIGEN_STRONG_INLINE Packet +pldexp_float(Packet a, Packet exponent) +{ + typedef typename unpacket_traits<Packet>::integer_packet PacketI; + const Packet cst_127 = pset1<Packet>(127.f); + // return a * 2^exponent + PacketI ei = pcast<Packet,PacketI>(padd(exponent, cst_127)); + return pmul(a, preinterpret<Packet>(pshiftleft<23>(ei))); +} + +// Natural logarithm +// Computes log(x) as log(2^e * m) = C*e + log(m), where the constant C =log(2) +// and m is in the range [sqrt(1/2),sqrt(2)). In this range, the logarithm can +// be easily approximated by a polynomial centered on m=1 for stability. +// TODO(gonnet): Further reduce the interval allowing for lower-degree +// polynomial interpolants -> ... -> profit! +template <typename Packet> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +EIGEN_UNUSED +Packet plog_float(const Packet _x) +{ + Packet x = _x; + + const Packet cst_1 = pset1<Packet>(1.0f); + const Packet cst_half = pset1<Packet>(0.5f); + // The smallest non denormalized float number. + const Packet cst_min_norm_pos = pset1frombits<Packet>( 0x00800000u); + const Packet cst_minus_inf = pset1frombits<Packet>( 0xff800000u); + const Packet cst_pos_inf = pset1frombits<Packet>( 0x7f800000u); + + // Polynomial coefficients. + const Packet cst_cephes_SQRTHF = pset1<Packet>(0.707106781186547524f); + const Packet cst_cephes_log_p0 = pset1<Packet>(7.0376836292E-2f); + const Packet cst_cephes_log_p1 = pset1<Packet>(-1.1514610310E-1f); + const Packet cst_cephes_log_p2 = pset1<Packet>(1.1676998740E-1f); + const Packet cst_cephes_log_p3 = pset1<Packet>(-1.2420140846E-1f); + const Packet cst_cephes_log_p4 = pset1<Packet>(+1.4249322787E-1f); + const Packet cst_cephes_log_p5 = pset1<Packet>(-1.6668057665E-1f); + const Packet cst_cephes_log_p6 = pset1<Packet>(+2.0000714765E-1f); + const Packet cst_cephes_log_p7 = pset1<Packet>(-2.4999993993E-1f); + const Packet cst_cephes_log_p8 = pset1<Packet>(+3.3333331174E-1f); + const Packet cst_cephes_log_q1 = pset1<Packet>(-2.12194440e-4f); + const Packet cst_cephes_log_q2 = pset1<Packet>(0.693359375f); + + // Truncate input values to the minimum positive normal. + x = pmax(x, cst_min_norm_pos); + + Packet e; + // extract significant in the range [0.5,1) and exponent + x = pfrexp(x,e); + + // part2: Shift the inputs from the range [0.5,1) to [sqrt(1/2),sqrt(2)) + // and shift by -1. The values are then centered around 0, which improves + // the stability of the polynomial evaluation. + // if( x < SQRTHF ) { + // e -= 1; + // x = x + x - 1.0; + // } else { x = x - 1.0; } + Packet mask = pcmp_lt(x, cst_cephes_SQRTHF); + Packet tmp = pand(x, mask); + x = psub(x, cst_1); + e = psub(e, pand(cst_1, mask)); + x = padd(x, tmp); + + Packet x2 = pmul(x, x); + Packet x3 = pmul(x2, x); + + // Evaluate the polynomial approximant of degree 8 in three parts, probably + // to improve instruction-level parallelism. + Packet y, y1, y2; + y = pmadd(cst_cephes_log_p0, x, cst_cephes_log_p1); + y1 = pmadd(cst_cephes_log_p3, x, cst_cephes_log_p4); + y2 = pmadd(cst_cephes_log_p6, x, cst_cephes_log_p7); + y = pmadd(y, x, cst_cephes_log_p2); + y1 = pmadd(y1, x, cst_cephes_log_p5); + y2 = pmadd(y2, x, cst_cephes_log_p8); + y = pmadd(y, x3, y1); + y = pmadd(y, x3, y2); + y = pmul(y, x3); + + // Add the logarithm of the exponent back to the result of the interpolation. + y1 = pmul(e, cst_cephes_log_q1); + tmp = pmul(x2, cst_half); + y = padd(y, y1); + x = psub(x, tmp); + y2 = pmul(e, cst_cephes_log_q2); + x = padd(x, y); + x = padd(x, y2); + + Packet invalid_mask = pcmp_lt_or_nan(_x, pzero(_x)); + Packet iszero_mask = pcmp_eq(_x,pzero(_x)); + Packet pos_inf_mask = pcmp_eq(_x,cst_pos_inf); + // Filter out invalid inputs, i.e.: + // - negative arg will be NAN + // - 0 will be -INF + // - +INF will be +INF + return pselect(iszero_mask, cst_minus_inf, + por(pselect(pos_inf_mask,cst_pos_inf,x), invalid_mask)); +} + +// Exponential function. Works by writing "x = m*log(2) + r" where +// "m = floor(x/log(2)+1/2)" and "r" is the remainder. The result is then +// "exp(x) = 2^m*exp(r)" where exp(r) is in the range [-1,1). +template <typename Packet> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +EIGEN_UNUSED +Packet pexp_float(const Packet _x) +{ + const Packet cst_1 = pset1<Packet>(1.0f); + const Packet cst_half = pset1<Packet>(0.5f); + const Packet cst_exp_hi = pset1<Packet>( 88.3762626647950f); + const Packet cst_exp_lo = pset1<Packet>(-88.3762626647949f); + + const Packet cst_cephes_LOG2EF = pset1<Packet>(1.44269504088896341f); + const Packet cst_cephes_exp_p0 = pset1<Packet>(1.9875691500E-4f); + const Packet cst_cephes_exp_p1 = pset1<Packet>(1.3981999507E-3f); + const Packet cst_cephes_exp_p2 = pset1<Packet>(8.3334519073E-3f); + const Packet cst_cephes_exp_p3 = pset1<Packet>(4.1665795894E-2f); + const Packet cst_cephes_exp_p4 = pset1<Packet>(1.6666665459E-1f); + const Packet cst_cephes_exp_p5 = pset1<Packet>(5.0000001201E-1f); + + // Clamp x. + Packet x = pmax(pmin(_x, cst_exp_hi), cst_exp_lo); + + // Express exp(x) as exp(m*ln(2) + r), start by extracting + // m = floor(x/ln(2) + 0.5). + Packet m = pfloor(pmadd(x, cst_cephes_LOG2EF, cst_half)); + + // Get r = x - m*ln(2). If no FMA instructions are available, m*ln(2) is + // subtracted out in two parts, m*C1+m*C2 = m*ln(2), to avoid accumulating + // truncation errors. + Packet r; +#ifdef EIGEN_HAS_SINGLE_INSTRUCTION_MADD + const Packet cst_nln2 = pset1<Packet>(-0.6931471805599453f); + r = pmadd(m, cst_nln2, x); +#else + const Packet cst_cephes_exp_C1 = pset1<Packet>(0.693359375f); + const Packet cst_cephes_exp_C2 = pset1<Packet>(-2.12194440e-4f); + r = psub(x, pmul(m, cst_cephes_exp_C1)); + r = psub(r, pmul(m, cst_cephes_exp_C2)); +#endif + + Packet r2 = pmul(r, r); + + // TODO(gonnet): Split into odd/even polynomials and try to exploit + // instruction-level parallelism. + Packet y = cst_cephes_exp_p0; + y = pmadd(y, r, cst_cephes_exp_p1); + y = pmadd(y, r, cst_cephes_exp_p2); + y = pmadd(y, r, cst_cephes_exp_p3); + y = pmadd(y, r, cst_cephes_exp_p4); + y = pmadd(y, r, cst_cephes_exp_p5); + y = pmadd(y, r2, r); + y = padd(y, cst_1); + + // Return 2^m * exp(r). + return pmax(pldexp(y,m), _x); +} + +template <typename Packet> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +EIGEN_UNUSED +Packet pexp_double(const Packet _x) +{ + Packet x = _x; + + const Packet cst_1 = pset1<Packet>(1.0); + const Packet cst_2 = pset1<Packet>(2.0); + const Packet cst_half = pset1<Packet>(0.5); + + const Packet cst_exp_hi = pset1<Packet>(709.437); + const Packet cst_exp_lo = pset1<Packet>(-709.436139303); + + const Packet cst_cephes_LOG2EF = pset1<Packet>(1.4426950408889634073599); + const Packet cst_cephes_exp_p0 = pset1<Packet>(1.26177193074810590878e-4); + const Packet cst_cephes_exp_p1 = pset1<Packet>(3.02994407707441961300e-2); + const Packet cst_cephes_exp_p2 = pset1<Packet>(9.99999999999999999910e-1); + const Packet cst_cephes_exp_q0 = pset1<Packet>(3.00198505138664455042e-6); + const Packet cst_cephes_exp_q1 = pset1<Packet>(2.52448340349684104192e-3); + const Packet cst_cephes_exp_q2 = pset1<Packet>(2.27265548208155028766e-1); + const Packet cst_cephes_exp_q3 = pset1<Packet>(2.00000000000000000009e0); + const Packet cst_cephes_exp_C1 = pset1<Packet>(0.693145751953125); + const Packet cst_cephes_exp_C2 = pset1<Packet>(1.42860682030941723212e-6); + + Packet tmp, fx; + + // clamp x + x = pmax(pmin(x, cst_exp_hi), cst_exp_lo); + // Express exp(x) as exp(g + n*log(2)). + fx = pmadd(cst_cephes_LOG2EF, x, cst_half); + + // Get the integer modulus of log(2), i.e. the "n" described above. + fx = pfloor(fx); + + // Get the remainder modulo log(2), i.e. the "g" described above. Subtract + // n*log(2) out in two steps, i.e. n*C1 + n*C2, C1+C2=log2 to get the last + // digits right. + tmp = pmul(fx, cst_cephes_exp_C1); + Packet z = pmul(fx, cst_cephes_exp_C2); + x = psub(x, tmp); + x = psub(x, z); + + Packet x2 = pmul(x, x); + + // Evaluate the numerator polynomial of the rational interpolant. + Packet px = cst_cephes_exp_p0; + px = pmadd(px, x2, cst_cephes_exp_p1); + px = pmadd(px, x2, cst_cephes_exp_p2); + px = pmul(px, x); + + // Evaluate the denominator polynomial of the rational interpolant. + Packet qx = cst_cephes_exp_q0; + qx = pmadd(qx, x2, cst_cephes_exp_q1); + qx = pmadd(qx, x2, cst_cephes_exp_q2); + qx = pmadd(qx, x2, cst_cephes_exp_q3); + + // I don't really get this bit, copied from the SSE2 routines, so... + // TODO(gonnet): Figure out what is going on here, perhaps find a better + // rational interpolant? + x = pdiv(px, psub(qx, px)); + x = pmadd(cst_2, x, cst_1); + + // Construct the result 2^n * exp(g) = e * x. The max is used to catch + // non-finite values in the input. + return pmax(pldexp(x,fx), _x); +} + +// The following code is inspired by the following stack-overflow answer: +// https://stackoverflow.com/questions/30463616/payne-hanek-algorithm-implementation-in-c/30465751#30465751 +// It has been largely optimized: +// - By-pass calls to frexp. +// - Aligned loads of required 96 bits of 2/pi. This is accomplished by +// (1) balancing the mantissa and exponent to the required bits of 2/pi are +// aligned on 8-bits, and (2) replicating the storage of the bits of 2/pi. +// - Avoid a branch in rounding and extraction of the remaining fractional part. +// Overall, I measured a speed up higher than x2 on x86-64. +inline float trig_reduce_huge (float xf, int *quadrant) +{ + using Eigen::numext::int32_t; + using Eigen::numext::uint32_t; + using Eigen::numext::int64_t; + using Eigen::numext::uint64_t; + + const double pio2_62 = 3.4061215800865545e-19; // pi/2 * 2^-62 + const uint64_t zero_dot_five = uint64_t(1) << 61; // 0.5 in 2.62-bit fixed-point foramt + + // 192 bits of 2/pi for Payne-Hanek reduction + // Bits are introduced by packet of 8 to enable aligned reads. + static const uint32_t two_over_pi [] = + { + 0x00000028, 0x000028be, 0x0028be60, 0x28be60db, + 0xbe60db93, 0x60db9391, 0xdb939105, 0x9391054a, + 0x91054a7f, 0x054a7f09, 0x4a7f09d5, 0x7f09d5f4, + 0x09d5f47d, 0xd5f47d4d, 0xf47d4d37, 0x7d4d3770, + 0x4d377036, 0x377036d8, 0x7036d8a5, 0x36d8a566, + 0xd8a5664f, 0xa5664f10, 0x664f10e4, 0x4f10e410, + 0x10e41000, 0xe4100000 + }; + + uint32_t xi = numext::as_uint(xf); + // Below, -118 = -126 + 8. + // -126 is to get the exponent, + // +8 is to enable alignment of 2/pi's bits on 8 bits. + // This is possible because the fractional part of x as only 24 meaningful bits. + uint32_t e = (xi >> 23) - 118; + // Extract the mantissa and shift it to align it wrt the exponent + xi = ((xi & 0x007fffffu)| 0x00800000u) << (e & 0x7); + + uint32_t i = e >> 3; + uint32_t twoopi_1 = two_over_pi[i-1]; + uint32_t twoopi_2 = two_over_pi[i+3]; + uint32_t twoopi_3 = two_over_pi[i+7]; + + // Compute x * 2/pi in 2.62-bit fixed-point format. + uint64_t p; + p = uint64_t(xi) * twoopi_3; + p = uint64_t(xi) * twoopi_2 + (p >> 32); + p = (uint64_t(xi * twoopi_1) << 32) + p; + + // Round to nearest: add 0.5 and extract integral part. + uint64_t q = (p + zero_dot_five) >> 62; + *quadrant = int(q); + // Now it remains to compute "r = x - q*pi/2" with high accuracy, + // since we have p=x/(pi/2) with high accuracy, we can more efficiently compute r as: + // r = (p-q)*pi/2, + // where the product can be be carried out with sufficient accuracy using double precision. + p -= q<<62; + return float(double(int64_t(p)) * pio2_62); +} + +template<bool ComputeSine,typename Packet> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +EIGEN_UNUSED +#if EIGEN_GNUC_AT_LEAST(4,4) +__attribute__((optimize("-fno-unsafe-math-optimizations"))) +#endif +Packet psincos_float(const Packet& _x) +{ +// Workaround -ffast-math aggressive optimizations +// See bug 1674 +#if EIGEN_COMP_CLANG && defined(EIGEN_VECTORIZE_SSE) +#define EIGEN_SINCOS_DONT_OPT(X) __asm__ ("" : "+x" (X)); +#else +#define EIGEN_SINCOS_DONT_OPT(X) +#endif + + typedef typename unpacket_traits<Packet>::integer_packet PacketI; + + const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + const PacketI csti_1 = pset1<PacketI>(1); + const Packet cst_sign_mask = pset1frombits<Packet>(0x80000000u); + + Packet x = pabs(_x); + + // Scale x by 2/Pi to find x's octant. + Packet y = pmul(x, cst_2oPI); + + // Rounding trick: + Packet y_round = padd(y, cst_rounding_magic); + EIGEN_SINCOS_DONT_OPT(y_round) + PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + y = psub(y_round, cst_rounding_magic); // nearest integer to x*4/pi + + // Reduce x by y octants to get: -Pi/4 <= x <= +Pi/4 + // using "Extended precision modular arithmetic" + #if defined(EIGEN_HAS_SINGLE_INSTRUCTION_MADD) + // This version requires true FMA for high accuracy + // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + const float huge_th = ComputeSine ? 117435.992f : 71476.0625f; + x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + #else + // Without true FMA, the previous set of coefficients maintain 1ULP accuracy + // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + // and 2 ULP up to: + const float huge_th = ComputeSine ? 25966.f : 18838.f; + x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + EIGEN_SINCOS_DONT_OPT(x) + x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + EIGEN_SINCOS_DONT_OPT(x) + x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + // For the record, the following set of coefficients maintain 2ULP up + // to a slightly larger range: + // const float huge_th = ComputeSine ? 51981.f : 39086.125f; + // but it slightly fails to maintain 1ULP for two values of sin below pi. + // x = pmadd(y, pset1<Packet>(-3.140625/2.), x); + // x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + // x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + // x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + // For the record, with only 3 iterations it is possible to maintain + // 1 ULP up to 3PI (maybe more) and 2ULP up to 255. + // The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + #endif + + if(predux_any(pcmp_le(pset1<Packet>(huge_th),pabs(_x)))) + { + const int PacketSize = unpacket_traits<Packet>::size; + EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; + EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; + EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) int y_int2[PacketSize]; + pstoreu(vals, pabs(_x)); + pstoreu(x_cpy, x); + pstoreu(y_int2, y_int); + for(int k=0; k<PacketSize;++k) + { + float val = vals[k]; + if(val>=huge_th && (numext::isfinite)(val)) + x_cpy[k] = trig_reduce_huge(val,&y_int2[k]); + } + x = ploadu<Packet>(x_cpy); + y_int = ploadu<PacketI>(y_int2); + } + + // Compute the sign to apply to the polynomial. + // sin: sign = second_bit(y_int) xor signbit(_x) + // cos: sign = second_bit(y_int+1) + Packet sign_bit = ComputeSine ? pxor(_x, preinterpret<Packet>(pshiftleft<30>(y_int))) + : preinterpret<Packet>(pshiftleft<30>(padd(y_int,csti_1))); + sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + + // Get the polynomial selection mask from the second bit of y_int + // We'll calculate both (sin and cos) polynomials and then select from the two. + Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + Packet x2 = pmul(x,x); + + // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f )); + y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f )); + y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); + y1 = pmadd(y1, x2, pset1<Packet>(1.f)); + + // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + // octave/matlab code to compute those coefficients: + // x = (0:0.0001:pi/4)'; + // A = [x.^3 x.^5 x.^7]; + // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + // + Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + y2 = pmadd(y2, x2, pset1<Packet>( 0.0083326873655616851693794799871284340042620897293090820312500000f)); + y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + y2 = pmul(y2, x2); + y2 = pmadd(y2, x, x); + + // Select the correct result from the two polynomials. + y = ComputeSine ? pselect(poly_mask,y2,y1) + : pselect(poly_mask,y1,y2); + + // Update the sign and filter huge inputs + return pxor(y, sign_bit); + +#undef EIGEN_SINCOS_DONT_OPT +} + +template<typename Packet> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +EIGEN_UNUSED +Packet psin_float(const Packet& x) +{ + return psincos_float<true>(x); +} + +template<typename Packet> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +EIGEN_UNUSED +Packet pcos_float(const Packet& x) +{ + return psincos_float<false>(x); +} + +} // end namespace internal +} // end namespace Eigen diff --git a/Eigen/src/Core/arch/Default/Settings.h b/Eigen/src/Core/arch/Default/Settings.h index 097373c84..a5c3ada4c 100644 --- a/Eigen/src/Core/arch/Default/Settings.h +++ b/Eigen/src/Core/arch/Default/Settings.h @@ -21,7 +21,7 @@ * it does not correspond to the number of iterations or the number of instructions */ #ifndef EIGEN_UNROLLING_LIMIT -#define EIGEN_UNROLLING_LIMIT 100 +#define EIGEN_UNROLLING_LIMIT 110 #endif /** Defines the threshold between a "small" and a "large" matrix. diff --git a/Eigen/src/Core/arch/GPU/PacketMath.h b/Eigen/src/Core/arch/GPU/PacketMath.h index ddf37b9c1..cd4615a45 100644 --- a/Eigen/src/Core/arch/GPU/PacketMath.h +++ b/Eigen/src/Core/arch/GPU/PacketMath.h @@ -53,6 +53,7 @@ template<> struct packet_traits<float> : default_packet_traits HasBetaInc = 1, HasBlend = 0, + HasFloor = 1, }; }; @@ -86,12 +87,13 @@ template<> struct packet_traits<double> : default_packet_traits HasBetaInc = 1, HasBlend = 0, + HasFloor = 1, }; }; -template<> struct unpacket_traits<float4> { typedef float type; enum {size=4, alignment=Aligned16}; typedef float4 half; }; -template<> struct unpacket_traits<double2> { typedef double type; enum {size=2, alignment=Aligned16}; typedef double2 half; }; +template<> struct unpacket_traits<float4> { typedef float type; enum {size=4, alignment=Aligned16, vectorizable=true}; typedef float4 half; }; +template<> struct unpacket_traits<double2> { typedef double type; enum {size=2, alignment=Aligned16, vectorizable=true}; typedef double2 half; }; template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pset1<float4>(const float& from) { return make_float4(from, from, from, from); @@ -100,6 +102,117 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pset1<double2>(const do return make_double2(from, from); } +namespace { + +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float bitwise_and(const float& a, + const float& b) { + return __int_as_float(__float_as_int(a) & __float_as_int(b)); +} +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double bitwise_and(const double& a, + const double& b) { + return __longlong_as_double(__double_as_longlong(a) & + __double_as_longlong(b)); +} + +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float bitwise_or(const float& a, + const float& b) { + return __int_as_float(__float_as_int(a) | __float_as_int(b)); +} +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double bitwise_or(const double& a, + const double& b) { + return __longlong_as_double(__double_as_longlong(a) | + __double_as_longlong(b)); +} + +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float bitwise_xor(const float& a, + const float& b) { + return __int_as_float(__float_as_int(a) ^ __float_as_int(b)); +} +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double bitwise_xor(const double& a, + const double& b) { + return __longlong_as_double(__double_as_longlong(a) ^ + __double_as_longlong(b)); +} + +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float bitwise_andnot(const float& a, + const float& b) { + return __int_as_float(__float_as_int(a) & ~__float_as_int(b)); +} +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double bitwise_andnot(const double& a, + const double& b) { + return __longlong_as_double(__double_as_longlong(a) & + ~__double_as_longlong(b)); +} +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float eq_mask(const float& a, + const float& b) { + return __int_as_float(a == b ? 0xffffffffu : 0u); +} +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double eq_mask(const double& a, + const double& b) { + return __longlong_as_double(a == b ? 0xffffffffffffffffull : 0ull); +} + +} // namespace + +template <> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pand<float4>(const float4& a, + const float4& b) { + return make_float4(bitwise_and(a.x, b.x), bitwise_and(a.y, b.y), + bitwise_and(a.z, b.z), bitwise_and(a.w, b.w)); +} +template <> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pand<double2>(const double2& a, + const double2& b) { + return make_double2(bitwise_and(a.x, b.x), bitwise_and(a.y, b.y)); +} + +template <> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 por<float4>(const float4& a, + const float4& b) { + return make_float4(bitwise_or(a.x, b.x), bitwise_or(a.y, b.y), + bitwise_or(a.z, b.z), bitwise_or(a.w, b.w)); +} +template <> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 por<double2>(const double2& a, + const double2& b) { + return make_double2(bitwise_or(a.x, b.x), bitwise_or(a.y, b.y)); +} + +template <> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pxor<float4>(const float4& a, + const float4& b) { + return make_float4(bitwise_xor(a.x, b.x), bitwise_xor(a.y, b.y), + bitwise_xor(a.z, b.z), bitwise_xor(a.w, b.w)); +} +template <> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pxor<double2>(const double2& a, + const double2& b) { + return make_double2(bitwise_xor(a.x, b.x), bitwise_xor(a.y, b.y)); +} + +template <> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pandnot<float4>(const float4& a, + const float4& b) { + return make_float4(bitwise_andnot(a.x, b.x), bitwise_andnot(a.y, b.y), + bitwise_andnot(a.z, b.z), bitwise_andnot(a.w, b.w)); +} +template <> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 +pandnot<double2>(const double2& a, const double2& b) { + return make_double2(bitwise_andnot(a.x, b.x), bitwise_andnot(a.y, b.y)); +} + +template <> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pcmp_eq<float4>(const float4& a, + const float4& b) { + return make_float4(eq_mask(a.x, b.x), eq_mask(a.y, b.y), eq_mask(a.z, b.z), + eq_mask(a.w, b.w)); +} +template <> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 +pcmp_eq<double2>(const double2& a, const double2& b) { + return make_double2(eq_mask(a.x, b.x), eq_mask(a.y, b.y)); +} template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 plset<float4>(const float& a) { return make_float4(a, a+1, a+2, a+3); @@ -297,6 +410,13 @@ template<> EIGEN_DEVICE_FUNC inline double2 pabs<double2>(const double2& a) { return make_double2(fabs(a.x), fabs(a.y)); } +template<> EIGEN_DEVICE_FUNC inline float4 pfloor<float4>(const float4& a) { + return make_float4(floorf(a.x), floorf(a.y), floorf(a.z), floorf(a.w)); +} +template<> EIGEN_DEVICE_FUNC inline double2 pfloor<double2>(const double2& a) { + return make_double2(floor(a.x), floor(a.y)); +} + EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<float4,4>& kernel) { float tmp = kernel.packet[0].y; diff --git a/Eigen/src/Core/arch/GPU/PacketMathHalf.h b/Eigen/src/Core/arch/GPU/PacketMathHalf.h index 8787adcde..869fa7ec6 100644 --- a/Eigen/src/Core/arch/GPU/PacketMathHalf.h +++ b/Eigen/src/Core/arch/GPU/PacketMathHalf.h @@ -30,6 +30,7 @@ template<> struct packet_traits<Eigen::half> : default_packet_traits size=2, HasHalfPacket = 0, HasAdd = 1, + HasSub = 1, HasMul = 1, HasDiv = 1, HasSqrt = 1, @@ -41,7 +42,7 @@ template<> struct packet_traits<Eigen::half> : default_packet_traits }; }; -template<> struct unpacket_traits<half2> { typedef Eigen::half type; enum {size=2, alignment=Aligned16}; typedef half2 half; }; +template<> struct unpacket_traits<half2> { typedef Eigen::half type; enum {size=2, alignment=Aligned16, vectorizable=true}; typedef half2 half; }; template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pset1<half2>(const Eigen::half& from) { @@ -137,12 +138,22 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half pfirst<half2>(const } template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pabs<half2>(const half2& a) { - half2 result; - unsigned temp = *(reinterpret_cast<const unsigned*>(&(a))); - *(reinterpret_cast<unsigned*>(&(result))) = temp & 0x7FFF7FFF; - return result; + half a1 = __low2half(a); + half a2 = __high2half(a); + half result1 = half_impl::raw_uint16_to_half(a1.x & 0x7FFF); + half result2 = half_impl::raw_uint16_to_half(a2.x & 0x7FFF); + return __halves2half2(result1, result2); } +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 ptrue<half2>(const half2& a) { + half true_half = half_impl::raw_uint16_to_half(0xffffu); + return pset1<half2>(true_half); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pzero<half2>(const half2& a) { + half false_half = half_impl::raw_uint16_to_half(0x0000u); + return pset1<half2>(false_half); +} EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void ptranspose(PacketBlock<half2,2>& kernel) { @@ -171,6 +182,68 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 plset<half2>(const Eigen: #endif } +template <> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pcmp_eq<half2>(const half2& a, + const half2& b) { + half true_half = half_impl::raw_uint16_to_half(0xffffu); + half false_half = half_impl::raw_uint16_to_half(0x0000u); + half a1 = __low2half(a); + half a2 = __high2half(a); + half b1 = __low2half(b); + half b2 = __high2half(b); + half eq1 = __half2float(a1) == __half2float(b1) ? true_half : false_half; + half eq2 = __half2float(a2) == __half2float(b2) ? true_half : false_half; + return __halves2half2(eq1, eq2); +} + +template <> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pand<half2>(const half2& a, + const half2& b) { + half a1 = __low2half(a); + half a2 = __high2half(a); + half b1 = __low2half(b); + half b2 = __high2half(b); + half result1 = half_impl::raw_uint16_to_half(a1.x & b1.x); + half result2 = half_impl::raw_uint16_to_half(a2.x & b2.x); + return __halves2half2(result1, result2); +} + +template <> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 por<half2>(const half2& a, + const half2& b) { + half a1 = __low2half(a); + half a2 = __high2half(a); + half b1 = __low2half(b); + half b2 = __high2half(b); + half result1 = half_impl::raw_uint16_to_half(a1.x | b1.x); + half result2 = half_impl::raw_uint16_to_half(a2.x | b2.x); + return __halves2half2(result1, result2); +} + +template <> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pxor<half2>(const half2& a, + const half2& b) { + half a1 = __low2half(a); + half a2 = __high2half(a); + half b1 = __low2half(b); + half b2 = __high2half(b); + half result1 = half_impl::raw_uint16_to_half(a1.x ^ b1.x); + half result2 = half_impl::raw_uint16_to_half(a2.x ^ b2.x); + return __halves2half2(result1, result2); +} + +template <> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pandnot<half2>(const half2& a, + const half2& b) { + half a1 = __low2half(a); + half a2 = __high2half(a); + half b1 = __low2half(b); + half b2 = __high2half(b); + half result1 = half_impl::raw_uint16_to_half(a1.x & ~b1.x); + half result2 = half_impl::raw_uint16_to_half(a2.x & ~b2.x); + return __halves2half2(result1, result2); +} + template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 padd<half2>(const half2& a, const half2& b) { #if defined(EIGEN_HIP_DEVICE_COMPILE) @@ -500,6 +573,7 @@ struct packet_traits<half> : default_packet_traits { HasAdd = 1, HasSub = 1, HasMul = 1, + HasDiv = 1, HasNegate = 1, HasAbs = 0, HasAbs2 = 0, @@ -507,7 +581,6 @@ struct packet_traits<half> : default_packet_traits { HasMax = 0, HasConj = 0, HasSetLinear = 0, - HasDiv = 0, HasSqrt = 0, HasRsqrt = 0, HasExp = 0, @@ -517,7 +590,7 @@ struct packet_traits<half> : default_packet_traits { }; -template<> struct unpacket_traits<Packet16h> { typedef Eigen::half type; enum {size=16, alignment=Aligned32}; typedef Packet16h half; }; +template<> struct unpacket_traits<Packet16h> { typedef Eigen::half type; enum {size=16, alignment=Aligned32, vectorizable=true}; typedef Packet16h half; }; template<> EIGEN_STRONG_INLINE Packet16h pset1<Packet16h>(const Eigen::half& from) { Packet16h result; @@ -640,6 +713,36 @@ EIGEN_STRONG_INLINE Packet16h float2half(const Packet16f& a) { #endif } +template<> EIGEN_STRONG_INLINE Packet16h pnot(const Packet16h& a) { + Packet16h r; r.x = _mm256_xor_si256(a.x, pcmp_eq(a.x, a.x)); return r; +} + +template<> EIGEN_STRONG_INLINE Packet16h ptrue(const Packet16h& a) { + Packet16h r; r.x = Packet8i(ptrue(a.x)); return r; +} + +template<> EIGEN_STRONG_INLINE Packet16h por(const Packet16h& a,const Packet16h& b) { + // in some cases Packet8i is a wrapper around __m256i, so we need to + // cast to Packet8i to call the correct overload. + Packet16h r; r.x = por(Packet8i(a.x),Packet8i(b.x)); return r; +} +template<> EIGEN_STRONG_INLINE Packet16h pxor(const Packet16h& a,const Packet16h& b) { + Packet16h r; r.x = pxor(Packet8i(a.x),Packet8i(b.x)); return r; +} +template<> EIGEN_STRONG_INLINE Packet16h pand(const Packet16h& a,const Packet16h& b) { + Packet16h r; r.x = pand(Packet8i(a.x),Packet8i(b.x)); return r; +} +template<> EIGEN_STRONG_INLINE Packet16h pandnot(const Packet16h& a,const Packet16h& b) { + Packet16h r; r.x = pandnot(Packet8i(a.x),Packet8i(b.x)); return r; +} + +template<> EIGEN_STRONG_INLINE Packet16h pcmp_eq(const Packet16h& a,const Packet16h& b) { + Packet16f af = half2float(a); + Packet16f bf = half2float(b); + Packet16f rf = pcmp_eq(af, bf); + return float2half(rf); +} + template<> EIGEN_STRONG_INLINE Packet16h pnegate(const Packet16h& a) { // FIXME we could do that with bit manipulation Packet16f af = half2float(a); @@ -668,6 +771,13 @@ template<> EIGEN_STRONG_INLINE Packet16h pmul<Packet16h>(const Packet16h& a, con return float2half(rf); } +template<> EIGEN_STRONG_INLINE Packet16h pdiv<Packet16h>(const Packet16h& a, const Packet16h& b) { + Packet16f af = half2float(a); + Packet16f bf = half2float(b); + Packet16f rf = pdiv(af, bf); + return float2half(rf); +} + template<> EIGEN_STRONG_INLINE half predux<Packet16h>(const Packet16h& from) { Packet16f from_float = half2float(from); return half(predux(from_float)); @@ -952,6 +1062,7 @@ struct packet_traits<Eigen::half> : default_packet_traits { HasAdd = 1, HasSub = 1, HasMul = 1, + HasDiv = 1, HasNegate = 1, HasAbs = 0, HasAbs2 = 0, @@ -959,7 +1070,6 @@ struct packet_traits<Eigen::half> : default_packet_traits { HasMax = 0, HasConj = 0, HasSetLinear = 0, - HasDiv = 0, HasSqrt = 0, HasRsqrt = 0, HasExp = 0, @@ -969,7 +1079,7 @@ struct packet_traits<Eigen::half> : default_packet_traits { }; -template<> struct unpacket_traits<Packet8h> { typedef Eigen::half type; enum {size=8, alignment=Aligned16}; typedef Packet8h half; }; +template<> struct unpacket_traits<Packet8h> { typedef Eigen::half type; enum {size=8, alignment=Aligned16, vectorizable=true}; typedef Packet8h half; }; template<> EIGEN_STRONG_INLINE Packet8h pset1<Packet8h>(const Eigen::half& from) { Packet8h result; @@ -1063,6 +1173,32 @@ EIGEN_STRONG_INLINE Packet8h float2half(const Packet8f& a) { #endif } +template<> EIGEN_STRONG_INLINE Packet8h ptrue(const Packet8h& a) { + Packet8h r; r.x = _mm_cmpeq_epi32(a.x, a.x); return r; +} + +template<> EIGEN_STRONG_INLINE Packet8h por(const Packet8h& a,const Packet8h& b) { + // in some cases Packet4i is a wrapper around __m128i, so we either need to + // cast to Packet4i to directly call the intrinsics as below: + Packet8h r; r.x = _mm_or_si128(a.x,b.x); return r; +} +template<> EIGEN_STRONG_INLINE Packet8h pxor(const Packet8h& a,const Packet8h& b) { + Packet8h r; r.x = _mm_xor_si128(a.x,b.x); return r; +} +template<> EIGEN_STRONG_INLINE Packet8h pand(const Packet8h& a,const Packet8h& b) { + Packet8h r; r.x = _mm_and_si128(a.x,b.x); return r; +} +template<> EIGEN_STRONG_INLINE Packet8h pandnot(const Packet8h& a,const Packet8h& b) { + Packet8h r; r.x = _mm_andnot_si128(b.x,a.x); return r; +} + +template<> EIGEN_STRONG_INLINE Packet8h pcmp_eq(const Packet8h& a,const Packet8h& b) { + Packet8f af = half2float(a); + Packet8f bf = half2float(b); + Packet8f rf = pcmp_eq(af, bf); + return float2half(rf); +} + template<> EIGEN_STRONG_INLINE Packet8h pconj(const Packet8h& a) { return a; } template<> EIGEN_STRONG_INLINE Packet8h pnegate(const Packet8h& a) { @@ -1093,6 +1229,13 @@ template<> EIGEN_STRONG_INLINE Packet8h pmul<Packet8h>(const Packet8h& a, const return float2half(rf); } +template<> EIGEN_STRONG_INLINE Packet8h pdiv<Packet8h>(const Packet8h& a, const Packet8h& b) { + Packet8f af = half2float(a); + Packet8f bf = half2float(b); + Packet8f rf = pdiv(af, bf); + return float2half(rf); +} + template<> EIGEN_STRONG_INLINE Packet8h pgather<Eigen::half, Packet8h>(const Eigen::half* from, Index stride) { Packet8h result; @@ -1279,9 +1422,10 @@ struct packet_traits<Eigen::half> : default_packet_traits { AlignedOnScalar = 1, size = 4, HasHalfPacket = 0, - HasAdd = 0, - HasSub = 0, - HasMul = 0, + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, HasNegate = 0, HasAbs = 0, HasAbs2 = 0, @@ -1289,7 +1433,6 @@ struct packet_traits<Eigen::half> : default_packet_traits { HasMax = 0, HasConj = 0, HasSetLinear = 0, - HasDiv = 0, HasSqrt = 0, HasRsqrt = 0, HasExp = 0, @@ -1299,7 +1442,7 @@ struct packet_traits<Eigen::half> : default_packet_traits { }; -template<> struct unpacket_traits<Packet4h> { typedef Eigen::half type; enum {size=4, alignment=Aligned16}; typedef Packet4h half; }; +template<> struct unpacket_traits<Packet4h> { typedef Eigen::half type; enum {size=4, alignment=Aligned16, vectorizable=true}; typedef Packet4h half; }; template<> EIGEN_STRONG_INLINE Packet4h pset1<Packet4h>(const Eigen::half& from) { Packet4h result; @@ -1336,6 +1479,29 @@ template<> EIGEN_STRONG_INLINE Packet4h padd<Packet4h>(const Packet4h& a, const return result; } +template<> EIGEN_STRONG_INLINE Packet4h psub<Packet4h>(const Packet4h& a, const Packet4h& b) { + __int64_t a64 = _mm_cvtm64_si64(a.x); + __int64_t b64 = _mm_cvtm64_si64(b.x); + + Eigen::half h[4]; + + Eigen::half ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64)); + Eigen::half hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64)); + h[0] = ha - hb; + ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 16)); + hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 16)); + h[1] = ha - hb; + ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 32)); + hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 32)); + h[2] = ha - hb; + ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 48)); + hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 48)); + h[3] = ha - hb; + Packet4h result; + result.x = _mm_set_pi16(h[3].x, h[2].x, h[1].x, h[0].x); + return result; +} + template<> EIGEN_STRONG_INLINE Packet4h pmul<Packet4h>(const Packet4h& a, const Packet4h& b) { __int64_t a64 = _mm_cvtm64_si64(a.x); __int64_t b64 = _mm_cvtm64_si64(b.x); @@ -1359,6 +1525,29 @@ template<> EIGEN_STRONG_INLINE Packet4h pmul<Packet4h>(const Packet4h& a, const return result; } +template<> EIGEN_STRONG_INLINE Packet4h pdiv<Packet4h>(const Packet4h& a, const Packet4h& b) { + __int64_t a64 = _mm_cvtm64_si64(a.x); + __int64_t b64 = _mm_cvtm64_si64(b.x); + + Eigen::half h[4]; + + Eigen::half ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64)); + Eigen::half hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64)); + h[0] = ha / hb; + ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 16)); + hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 16)); + h[1] = ha / hb; + ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 32)); + hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 32)); + h[2] = ha / hb; + ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 48)); + hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 48)); + h[3] = ha / hb; + Packet4h result; + result.x = _mm_set_pi16(h[3].x, h[2].x, h[1].x, h[0].x); + return result; +} + template<> EIGEN_STRONG_INLINE Packet4h pload<Packet4h>(const Eigen::half* from) { Packet4h result; result.x = _mm_cvtsi64_m64(*reinterpret_cast<const __int64_t*>(from)); diff --git a/Eigen/src/Core/arch/MSA/Complex.h b/Eigen/src/Core/arch/MSA/Complex.h index 9a45cf51e..fa64d3564 100644 --- a/Eigen/src/Core/arch/MSA/Complex.h +++ b/Eigen/src/Core/arch/MSA/Complex.h @@ -127,7 +127,7 @@ struct packet_traits<std::complex<float> > : default_packet_traits { template <> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; - enum { size = 2, alignment = Aligned16 }; + enum { size = 2, alignment = Aligned16, vectorizable=true }; typedef Packet2cf half; }; @@ -500,7 +500,7 @@ struct packet_traits<std::complex<double> > : default_packet_traits { template <> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; - enum { size = 1, alignment = Aligned16 }; + enum { size = 1, alignment = Aligned16, vectorizable=true }; typedef Packet1cd half; }; diff --git a/Eigen/src/Core/arch/MSA/MathFunctions.h b/Eigen/src/Core/arch/MSA/MathFunctions.h index 98e23e36f..f5181b90e 100644 --- a/Eigen/src/Core/arch/MSA/MathFunctions.h +++ b/Eigen/src/Core/arch/MSA/MathFunctions.h @@ -261,7 +261,7 @@ Packet4f psincos_inner_msa_float(const Packet4f& _x) { // x's from odd-numbered octants will translate to octant -1: [-Pi/4, 0]. // Adjustment for odd-numbered octants: octant = (octant + 1) & (~1). Packet4i y_int1 = __builtin_msa_addvi_w(y_int, 1); - Packet4i y_int2 = (Packet4i)__builtin_msa_bclri_w((Packet4ui)y_int1, 0); + Packet4i y_int2 = (Packet4i)__builtin_msa_bclri_w((Packet4ui)y_int1, 0); // bclri = bit-clear y = __builtin_msa_ffint_s_w(y_int2); // Compute the sign to apply to the polynomial. @@ -305,7 +305,7 @@ Packet4f psincos_inner_msa_float(const Packet4f& _x) { // Update the sign. sign_mask = pxor(sign_mask, (Packet4i)y); - y = (Packet4f)__builtin_msa_binsli_w((v4u32)y, (v4u32)sign_mask, 0); + y = (Packet4f)__builtin_msa_binsli_w((v4u32)y, (v4u32)sign_mask, 0); // binsli = bit-insert-left return y; } diff --git a/Eigen/src/Core/arch/MSA/PacketMath.h b/Eigen/src/Core/arch/MSA/PacketMath.h index 094c874ee..a97156a84 100644 --- a/Eigen/src/Core/arch/MSA/PacketMath.h +++ b/Eigen/src/Core/arch/MSA/PacketMath.h @@ -117,14 +117,14 @@ struct packet_traits<int32_t> : default_packet_traits { template <> struct unpacket_traits<Packet4f> { typedef float type; - enum { size = 4, alignment = Aligned16 }; + enum { size = 4, alignment = Aligned16, vectorizable=true }; typedef Packet4f half; }; template <> struct unpacket_traits<Packet4i> { typedef int32_t type; - enum { size = 4, alignment = Aligned16 }; + enum { size = 4, alignment = Aligned16, vectorizable=true }; typedef Packet4i half; }; @@ -925,7 +925,7 @@ struct packet_traits<double> : default_packet_traits { template <> struct unpacket_traits<Packet2d> { typedef double type; - enum { size = 2, alignment = Aligned16 }; + enum { size = 2, alignment = Aligned16, vectorizable=true }; typedef Packet2d half; }; diff --git a/Eigen/src/Core/arch/NEON/Complex.h b/Eigen/src/Core/arch/NEON/Complex.h index 306a309be..f6c5c211c 100644 --- a/Eigen/src/Core/arch/NEON/Complex.h +++ b/Eigen/src/Core/arch/NEON/Complex.h @@ -62,7 +62,7 @@ template<> struct packet_traits<std::complex<float> > : default_packet_traits }; }; -template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16}; typedef Packet2cf half; }; +template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16, vectorizable=true}; typedef Packet2cf half; }; template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>& from) { @@ -101,6 +101,18 @@ template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, con return Packet2cf(vaddq_f32(v1, v2)); } +template<> EIGEN_STRONG_INLINE Packet2cf pcmp_eq(const Packet2cf& a, const Packet2cf& b) +{ + // Compare real and imaginary parts of a and b to get the mask vector: + // [re(a[0])==re(b[0]), im(a[0])==im(b[0]), re(a[1])==re(b[1]), im(a[1])==im(b[1])] + Packet4f eq = pcmp_eq<Packet4f>(a.v, b.v); + // Swap real/imag elements in the mask in to get: + // [im(a[0])==im(b[0]), re(a[0])==re(b[0]), im(a[1])==im(b[1]), re(a[1])==re(b[1])] + Packet4f eq_swapped = vrev64q_f32(eq); + // Return re(a)==re(b) && im(a)==im(b) by computing bitwise AND of eq and eq_swapped + return Packet2cf(pand<Packet4f>(eq, eq_swapped)); +} + template<> EIGEN_STRONG_INLINE Packet2cf pand <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v)))); @@ -146,7 +158,7 @@ template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::co template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a) { - std::complex<float> EIGEN_ALIGN16 x[2]; + EIGEN_ALIGN16 std::complex<float> x[2]; vst1q_f32((float *)x, a.v); return x[0]; } @@ -328,7 +340,7 @@ template<> struct packet_traits<std::complex<double> > : default_packet_traits }; }; -template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16}; typedef Packet1cd half; }; +template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16, vectorizable=true}; typedef Packet1cd half; }; template<> EIGEN_STRONG_INLINE Packet1cd pload<Packet1cd>(const std::complex<double>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(pload<Packet2d>((const double*)from)); } template<> EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu<Packet2d>((const double*)from)); } @@ -361,6 +373,18 @@ template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, con return Packet1cd(vaddq_f64(v1, v2)); } +template<> EIGEN_STRONG_INLINE Packet1cd pcmp_eq(const Packet1cd& a, const Packet1cd& b) +{ + // Compare real and imaginary parts of a and b to get the mask vector: + // [re(a)==re(b), im(a)==im(b)] + Packet2d eq = pcmp_eq<Packet2d>(a.v, b.v); + // Swap real/imag elements in the mask in to get: + // [im(a)==im(b), re(a)==re(b)] + Packet2d eq_swapped = vreinterpretq_f64_u32(vrev64q_u32(vreinterpretq_u32_f64(eq))); + // Return re(a)==re(b) & im(a)==im(b) by computing bitwise AND of eq and eq_swapped + return Packet1cd(pand<Packet2d>(eq, eq_swapped)); +} + template<> EIGEN_STRONG_INLINE Packet1cd pand <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vreinterpretq_f64_u64(vandq_u64(vreinterpretq_u64_f64(a.v),vreinterpretq_u64_f64(b.v)))); @@ -401,7 +425,7 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet1c template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Packet1cd& a) { - std::complex<double> EIGEN_ALIGN16 res; + EIGEN_ALIGN16 std::complex<double> res; pstore<std::complex<double> >(&res, a); return res; diff --git a/Eigen/src/Core/arch/NEON/MathFunctions.h b/Eigen/src/Core/arch/NEON/MathFunctions.h index c48c61023..2e7d0e944 100644 --- a/Eigen/src/Core/arch/NEON/MathFunctions.h +++ b/Eigen/src/Core/arch/NEON/MathFunctions.h @@ -5,175 +5,37 @@ // 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/. -/* The sin, cos, exp, and log functions of this file come from - * Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/ - */ - #ifndef EIGEN_MATH_FUNCTIONS_NEON_H #define EIGEN_MATH_FUNCTIONS_NEON_H +#include "../Default/GenericPacketMathFunctions.h" + namespace Eigen { namespace internal { template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED -Packet4f pexp<Packet4f>(const Packet4f& _x) +Packet4f pexp<Packet4f>(const Packet4f& x) { - Packet4f x = _x; - Packet4f tmp, fx; - - _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f); - _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f); - _EIGEN_DECLARE_CONST_Packet4i(0x7f, 0x7f); - _EIGEN_DECLARE_CONST_Packet4f(exp_hi, 88.3762626647950f); - _EIGEN_DECLARE_CONST_Packet4f(exp_lo, -88.3762626647949f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_LOG2EF, 1.44269504088896341f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C1, 0.693359375f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C2, -2.12194440e-4f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p0, 1.9875691500E-4f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p1, 1.3981999507E-3f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p2, 8.3334519073E-3f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p3, 4.1665795894E-2f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459E-1f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201E-1f); - - x = vminq_f32(x, p4f_exp_hi); - x = vmaxq_f32(x, p4f_exp_lo); - - /* express exp(x) as exp(g + n*log(2)) */ - fx = vmlaq_f32(p4f_half, x, p4f_cephes_LOG2EF); - - /* perform a floorf */ - tmp = vcvtq_f32_s32(vcvtq_s32_f32(fx)); - - /* if greater, substract 1 */ - Packet4ui mask = vcgtq_f32(tmp, fx); - mask = vandq_u32(mask, vreinterpretq_u32_f32(p4f_1)); - - fx = vsubq_f32(tmp, vreinterpretq_f32_u32(mask)); - - tmp = vmulq_f32(fx, p4f_cephes_exp_C1); - Packet4f z = vmulq_f32(fx, p4f_cephes_exp_C2); - x = vsubq_f32(x, tmp); - x = vsubq_f32(x, z); - - Packet4f y = vmulq_f32(p4f_cephes_exp_p0, x); - z = vmulq_f32(x, x); - y = vaddq_f32(y, p4f_cephes_exp_p1); - y = vmulq_f32(y, x); - y = vaddq_f32(y, p4f_cephes_exp_p2); - y = vmulq_f32(y, x); - y = vaddq_f32(y, p4f_cephes_exp_p3); - y = vmulq_f32(y, x); - y = vaddq_f32(y, p4f_cephes_exp_p4); - y = vmulq_f32(y, x); - y = vaddq_f32(y, p4f_cephes_exp_p5); - - y = vmulq_f32(y, z); - y = vaddq_f32(y, x); - y = vaddq_f32(y, p4f_1); - - /* build 2^n */ - int32x4_t mm; - mm = vcvtq_s32_f32(fx); - mm = vaddq_s32(mm, p4i_0x7f); - mm = vshlq_n_s32(mm, 23); - Packet4f pow2n = vreinterpretq_f32_s32(mm); - - y = vmulq_f32(y, pow2n); - return y; + return pexp_float(x); } template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED -Packet4f plog<Packet4f>(const Packet4f& _x) +Packet4f plog<Packet4f>(const Packet4f& x) { - Packet4f x = _x; - _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f); - _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f); - _EIGEN_DECLARE_CONST_Packet4i(0x7f, 0x7f); - - _EIGEN_DECLARE_CONST_Packet4i(inv_mant_mask, ~0x7f800000); - - /* natural logarithm computed for 4 simultaneous float - return NaN for x <= 0 - */ - _EIGEN_DECLARE_CONST_Packet4f(cephes_SQRTHF, 0.707106781186547524f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p0, 7.0376836292E-2f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p1, - 1.1514610310E-1f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p2, 1.1676998740E-1f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p3, - 1.2420140846E-1f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p4, + 1.4249322787E-1f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p5, - 1.6668057665E-1f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p6, + 2.0000714765E-1f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p7, - 2.4999993993E-1f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p8, + 3.3333331174E-1f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q1, -2.12194440e-4f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q2, 0.693359375f); - - x = vmaxq_f32(x, vdupq_n_f32(0)); /* force flush to zero on denormal values */ - Packet4ui invalid_mask = vcleq_f32(x, vdupq_n_f32(0)); - - Packet4i ux = vreinterpretq_s32_f32(x); - - Packet4i emm0 = vshrq_n_s32(ux, 23); - - /* keep only the fractional part */ - ux = vandq_s32(ux, p4i_inv_mant_mask); - ux = vorrq_s32(ux, vreinterpretq_s32_f32(p4f_half)); - x = vreinterpretq_f32_s32(ux); - - emm0 = vsubq_s32(emm0, p4i_0x7f); - Packet4f e = vcvtq_f32_s32(emm0); - - e = vaddq_f32(e, p4f_1); - - /* part2: - if( x < SQRTHF ) { - e -= 1; - x = x + x - 1.0; - } else { x = x - 1.0; } - */ - Packet4ui mask = vcltq_f32(x, p4f_cephes_SQRTHF); - Packet4f tmp = vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(x), mask)); - x = vsubq_f32(x, p4f_1); - e = vsubq_f32(e, vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(p4f_1), mask))); - x = vaddq_f32(x, tmp); - - Packet4f z = vmulq_f32(x,x); - - Packet4f y = p4f_cephes_log_p0; - y = vmulq_f32(y, x); - y = vaddq_f32(y, p4f_cephes_log_p1); - y = vmulq_f32(y, x); - y = vaddq_f32(y, p4f_cephes_log_p2); - y = vmulq_f32(y, x); - y = vaddq_f32(y, p4f_cephes_log_p3); - y = vmulq_f32(y, x); - y = vaddq_f32(y, p4f_cephes_log_p4); - y = vmulq_f32(y, x); - y = vaddq_f32(y, p4f_cephes_log_p5); - y = vmulq_f32(y, x); - y = vaddq_f32(y, p4f_cephes_log_p6); - y = vmulq_f32(y, x); - y = vaddq_f32(y, p4f_cephes_log_p7); - y = vmulq_f32(y, x); - y = vaddq_f32(y, p4f_cephes_log_p8); - y = vmulq_f32(y, x); - - y = vmulq_f32(y, z); - - tmp = vmulq_f32(e, p4f_cephes_log_q1); - y = vaddq_f32(y, tmp); - + return plog_float(x); +} - tmp = vmulq_f32(z, p4f_half); - y = vsubq_f32(y, tmp); +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet4f psin<Packet4f>(const Packet4f& x) +{ + return psin_float(x); +} - tmp = vmulq_f32(e, p4f_cephes_log_q2); - x = vaddq_f32(x, y); - x = vaddq_f32(x, tmp); - x = vreinterpretq_f32_u32(vorrq_u32(vreinterpretq_u32_f32(x), invalid_mask)); // negative arg will be NAN - return x; +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet4f pcos<Packet4f>(const Packet4f& x) +{ + return pcos_float(x); } } // end namespace internal diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index 010739380..e8b351849 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -108,10 +108,11 @@ template<> struct packet_traits<float> : default_packet_traits size = 4, HasHalfPacket=0, // Packet2f intrinsics not implemented yet - HasDiv = 1, + HasDiv = 1, + HasFloor = 1, // FIXME check the Has* - HasSin = 0, - HasCos = 0, + HasSin = EIGEN_FAST_MATH, + HasCos = EIGEN_FAST_MATH, HasLog = 1, HasExp = 1, HasSqrt = 0 @@ -139,12 +140,25 @@ EIGEN_STRONG_INLINE void vst1q_f32(float* to, float32x4_t from) { ::vst1q EIGEN_STRONG_INLINE void vst1_f32 (float* to, float32x2_t from) { ::vst1_f32 ((float32_t*)to,from); } #endif -template<> struct unpacket_traits<Packet4f> { typedef float type; enum {size=4, alignment=Aligned16}; typedef Packet4f half; }; -template<> struct unpacket_traits<Packet4i> { typedef int32_t type; enum {size=4, alignment=Aligned16}; typedef Packet4i half; }; +template<> struct unpacket_traits<Packet4f> +{ + typedef float type; + typedef Packet4f half; + typedef Packet4i integer_packet; + enum {size=4, alignment=Aligned16, vectorizable=true}; +}; +template<> struct unpacket_traits<Packet4i> +{ + typedef int32_t type; + typedef Packet4i half; + enum {size=4, alignment=Aligned16, vectorizable=true}; +}; template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) { return vdupq_n_f32(from); } template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int32_t& from) { return vdupq_n_s32(from); } +template<> EIGEN_STRONG_INLINE Packet4f pset1frombits<Packet4f>(unsigned int from) { return vreinterpretq_f32_u32(vdupq_n_u32(from)); } + template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a) { const float f[] = {0, 1, 2, 3}; @@ -249,6 +263,25 @@ template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) { return vmaxq_f32(a,b); } template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vmaxq_s32(a,b); } +template<> EIGEN_STRONG_INLINE Packet4f pcmp_le(const Packet4f& a, const Packet4f& b) { return vreinterpretq_f32_u32(vcleq_f32(a,b)); } +template<> EIGEN_STRONG_INLINE Packet4f pcmp_lt(const Packet4f& a, const Packet4f& b) { return vreinterpretq_f32_u32(vcltq_f32(a,b)); } +template<> EIGEN_STRONG_INLINE Packet4f pcmp_eq(const Packet4f& a, const Packet4f& b) { return vreinterpretq_f32_u32(vceqq_f32(a,b)); } +template<> EIGEN_STRONG_INLINE Packet4f pcmp_lt_or_nan(const Packet4f& a, const Packet4f& b) { return vreinterpretq_f32_u32(vmvnq_u32(vcgeq_f32(a,b))); } + +template<> EIGEN_STRONG_INLINE Packet4i pcmp_eq(const Packet4i& a, const Packet4i& b) { return vreinterpretq_s32_u32(vceqq_s32(a,b)); } + +template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) +{ + const Packet4f cst_1 = pset1<Packet4f>(1.0f); + /* perform a floorf */ + Packet4f tmp = vcvtq_f32_s32(vcvtq_s32_f32(a)); + + /* if greater, substract 1 */ + Packet4ui mask = vcgtq_f32(tmp, a); + mask = vandq_u32(mask, vreinterpretq_u32_f32(cst_1)); + return vsubq_f32(tmp, vreinterpretq_f32_u32(mask)); +} + // Logical Operations are not supported for float, so we have to reinterpret casts using NEON intrinsics template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { @@ -274,6 +307,9 @@ template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, con } template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return vbicq_s32(a,b); } +template<int N> EIGEN_STRONG_INLINE Packet4i pshiftright(Packet4i a) { return vshrq_n_s32(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet4i pshiftleft(Packet4i a) { return vshlq_n_s32(a,N); } + template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return vld1q_f32(from); } template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int32_t* from) { EIGEN_DEBUG_ALIGNED_LOAD return vld1q_s32(from); } @@ -339,8 +375,8 @@ template<> EIGEN_STRONG_INLINE void prefetch<float> (const float* addr) { EI template<> EIGEN_STRONG_INLINE void prefetch<int32_t>(const int32_t* addr) { EIGEN_ARM_PREFETCH(addr); } // FIXME only store the 2 first elements ? -template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { float EIGEN_ALIGN16 x[4]; vst1q_f32(x, a); return x[0]; } -template<> EIGEN_STRONG_INLINE int32_t pfirst<Packet4i>(const Packet4i& a) { int32_t EIGEN_ALIGN16 x[4]; vst1q_s32(x, a); return x[0]; } +template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { EIGEN_ALIGN16 float x[4]; vst1q_f32(x, a); return x[0]; } +template<> EIGEN_STRONG_INLINE int32_t pfirst<Packet4i>(const Packet4i& a) { EIGEN_ALIGN16 int32_t x[4]; vst1q_s32(x, a); return x[0]; } template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) { float32x2_t a_lo, a_hi; @@ -364,6 +400,14 @@ template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) { template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a) { return vabsq_f32(a); } template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) { return vabsq_s32(a); } +template<> EIGEN_STRONG_INLINE Packet4f pfrexp<Packet4f>(const Packet4f& a, Packet4f& exponent) { + return pfrexp_float(a,exponent); +} + +template<> EIGEN_STRONG_INLINE Packet4f pldexp<Packet4f>(const Packet4f& a, const Packet4f& exponent) { + return pldexp_float(a,exponent); +} + template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a) { float32x2_t a_lo, a_hi, sum; @@ -507,6 +551,13 @@ template<> EIGEN_STRONG_INLINE int32_t predux_max<Packet4i>(const Packet4i& a) return vget_lane_s32(max, 0); } +template<> EIGEN_STRONG_INLINE bool predux_any(const Packet4f& x) +{ + uint32x2_t tmp = vorr_u32(vget_low_u32( vreinterpretq_u32_f32(x)), + vget_high_u32(vreinterpretq_u32_f32(x))); + return vget_lane_u32(vpmax_u32(tmp,tmp),0); +} + // this PALIGN_NEON business is to work around a bug in LLVM Clang 3.0 causing incorrect compilation errors, // see bug 347 and this LLVM bug: http://llvm.org/bugs/show_bug.cgi?id=11074 #define PALIGN_NEON(Offset,Type,Command) \ @@ -606,7 +657,7 @@ template<> struct packet_traits<double> : default_packet_traits }; }; -template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16}; typedef Packet2d half; }; +template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16, vectorizable=true}; typedef Packet2d half; }; template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { return vdupq_n_f64(from); } @@ -660,6 +711,8 @@ template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, con return vreinterpretq_f64_u64(vbicq_u64(vreinterpretq_u64_f64(a),vreinterpretq_u64_f64(b))); } +template<> EIGEN_STRONG_INLINE Packet2d pcmp_eq(const Packet2d& a, const Packet2d& b) { return vreinterpretq_f64_u64(vceqq_f64(a,b)); } + template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from) { EIGEN_DEBUG_ALIGNED_LOAD return vld1q_f64(from); } template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from) { EIGEN_DEBUG_UNALIGNED_LOAD return vld1q_f64(from); } diff --git a/Eigen/src/Core/arch/NEON/TypeCasting.h b/Eigen/src/Core/arch/NEON/TypeCasting.h index 95d1fd0e4..20dbe1332 100644 --- a/Eigen/src/Core/arch/NEON/TypeCasting.h +++ b/Eigen/src/Core/arch/NEON/TypeCasting.h @@ -41,6 +41,14 @@ template<> EIGEN_STRONG_INLINE Packet4f pcast<Packet4i, Packet4f>(const Packet4i return vcvtq_f32_s32(a); } +template<> EIGEN_STRONG_INLINE Packet4i preinterpret<Packet4i,Packet4f>(const Packet4f& a) { + return vreinterpretq_s32_f32(a); +} + +template<> EIGEN_STRONG_INLINE Packet4f preinterpret<Packet4f,Packet4i>(const Packet4i& a) { + return vreinterpretq_f32_s32(a); +} + } // end namespace internal } // end namespace Eigen diff --git a/Eigen/src/Core/arch/SSE/Complex.h b/Eigen/src/Core/arch/SSE/Complex.h index d075043ce..f39988eac 100644 --- a/Eigen/src/Core/arch/SSE/Complex.h +++ b/Eigen/src/Core/arch/SSE/Complex.h @@ -50,7 +50,7 @@ template<> struct packet_traits<std::complex<float> > : default_packet_traits }; #endif -template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16}; typedef Packet2cf half; }; +template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16, vectorizable=true}; typedef Packet2cf half; }; template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_add_ps(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_sub_ps(a.v,b.v)); } @@ -82,10 +82,13 @@ template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, con #endif } +template<> EIGEN_STRONG_INLINE Packet2cf ptrue <Packet2cf>(const Packet2cf& a) { return Packet2cf(ptrue(Packet4f(a.v))); } +template<> EIGEN_STRONG_INLINE Packet2cf pnot <Packet2cf>(const Packet2cf& a) { return Packet2cf(pnot(Packet4f(a.v))); } + template<> EIGEN_STRONG_INLINE Packet2cf pand <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_and_ps(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet2cf por <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_or_ps(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet2cf pxor <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_xor_ps(a.v,b.v)); } -template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_andnot_ps(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_andnot_ps(b.v,a.v)); } template<> EIGEN_STRONG_INLINE Packet2cf pload <Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>(&numext::real_ref(*from))); } template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>(&numext::real_ref(*from))); } @@ -280,7 +283,7 @@ template<> struct packet_traits<std::complex<double> > : default_packet_traits }; #endif -template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16}; typedef Packet1cd half; }; +template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16, vectorizable=true}; typedef Packet1cd half; }; template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_add_pd(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_sub_pd(a.v,b.v)); } @@ -305,10 +308,12 @@ template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, con #endif } +template<> EIGEN_STRONG_INLINE Packet1cd ptrue <Packet1cd>(const Packet1cd& a) { return Packet1cd(ptrue(Packet2d(a.v))); } +template<> EIGEN_STRONG_INLINE Packet1cd pnot <Packet1cd>(const Packet1cd& a) { return Packet1cd(pnot(Packet2d(a.v))); } template<> EIGEN_STRONG_INLINE Packet1cd pand <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_and_pd(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet1cd por <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_or_pd(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet1cd pxor <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_xor_pd(a.v,b.v)); } -template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_andnot_pd(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_andnot_pd(b.v,a.v)); } // FIXME force unaligned load, this is a temporary fix template<> EIGEN_STRONG_INLINE Packet1cd pload <Packet1cd>(const std::complex<double>* from) @@ -439,6 +444,18 @@ ptranspose(PacketBlock<Packet2cf,2>& kernel) { kernel.packet[1].v = tmp; } +template<> EIGEN_STRONG_INLINE Packet2cf pcmp_eq(const Packet2cf& a, const Packet2cf& b) +{ + __m128 eq = _mm_cmpeq_ps(a.v, b.v); + return Packet2cf(pand<Packet4f>(eq, vec4f_swizzle1(eq, 1, 0, 3, 2))); +} + +template<> EIGEN_STRONG_INLINE Packet1cd pcmp_eq(const Packet1cd& a, const Packet1cd& b) +{ + __m128d eq = _mm_cmpeq_pd(a.v, b.v); + return Packet1cd(pand<Packet2d>(eq, vec2d_swizzle1(eq, 1, 0))); +} + template<> EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, const Packet2cf& thenPacket, const Packet2cf& elsePacket) { __m128d result = pblend<Packet2d>(ifPacket, _mm_castps_pd(thenPacket.v), _mm_castps_pd(elsePacket.v)); return Packet2cf(_mm_castpd_ps(result)); diff --git a/Eigen/src/Core/arch/SSE/MathFunctions.h b/Eigen/src/Core/arch/SSE/MathFunctions.h index 4af2c6cae..0d491ab88 100644 --- a/Eigen/src/Core/arch/SSE/MathFunctions.h +++ b/Eigen/src/Core/arch/SSE/MathFunctions.h @@ -8,13 +8,15 @@ // 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/. -/* The sin, cos, exp, and log functions of this file come from +/* The sin and cos and functions of this file come from * Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/ */ #ifndef EIGEN_MATH_FUNCTIONS_SSE_H #define EIGEN_MATH_FUNCTIONS_SSE_H +#include "../Default/GenericPacketMathFunctions.h" + namespace Eigen { namespace internal { @@ -22,424 +24,31 @@ namespace internal { template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f plog<Packet4f>(const Packet4f& _x) { - Packet4f x = _x; - _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f); - _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f); - _EIGEN_DECLARE_CONST_Packet4i(0x7f, 0x7f); - - _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(inv_mant_mask, ~0x7f800000); - - /* the smallest non denormalized float number */ - _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(min_norm_pos, 0x00800000); - _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(minus_inf, 0xff800000);//-1.f/0.f); - - /* natural logarithm computed for 4 simultaneous float - return NaN for x <= 0 - */ - _EIGEN_DECLARE_CONST_Packet4f(cephes_SQRTHF, 0.707106781186547524f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p0, 7.0376836292E-2f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p1, - 1.1514610310E-1f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p2, 1.1676998740E-1f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p3, - 1.2420140846E-1f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p4, + 1.4249322787E-1f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p5, - 1.6668057665E-1f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p6, + 2.0000714765E-1f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p7, - 2.4999993993E-1f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p8, + 3.3333331174E-1f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q1, -2.12194440e-4f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q2, 0.693359375f); - - - Packet4i emm0; - - Packet4f invalid_mask = _mm_cmpnge_ps(x, _mm_setzero_ps()); // not greater equal is true if x is NaN - Packet4f iszero_mask = _mm_cmpeq_ps(x, _mm_setzero_ps()); - - x = pmax(x, p4f_min_norm_pos); /* cut off denormalized stuff */ - emm0 = _mm_srli_epi32(_mm_castps_si128(x), 23); - - /* keep only the fractional part */ - x = _mm_and_ps(x, p4f_inv_mant_mask); - x = _mm_or_ps(x, p4f_half); - - emm0 = _mm_sub_epi32(emm0, p4i_0x7f); - Packet4f e = padd(Packet4f(_mm_cvtepi32_ps(emm0)), p4f_1); - - /* part2: - if( x < SQRTHF ) { - e -= 1; - x = x + x - 1.0; - } else { x = x - 1.0; } - */ - Packet4f mask = _mm_cmplt_ps(x, p4f_cephes_SQRTHF); - Packet4f tmp = pand(x, mask); - x = psub(x, p4f_1); - e = psub(e, pand(p4f_1, mask)); - x = padd(x, tmp); - - Packet4f x2 = pmul(x,x); - Packet4f x3 = pmul(x2,x); - - Packet4f y, y1, y2; - y = pmadd(p4f_cephes_log_p0, x, p4f_cephes_log_p1); - y1 = pmadd(p4f_cephes_log_p3, x, p4f_cephes_log_p4); - y2 = pmadd(p4f_cephes_log_p6, x, p4f_cephes_log_p7); - y = pmadd(y , x, p4f_cephes_log_p2); - y1 = pmadd(y1, x, p4f_cephes_log_p5); - y2 = pmadd(y2, x, p4f_cephes_log_p8); - y = pmadd(y, x3, y1); - y = pmadd(y, x3, y2); - y = pmul(y, x3); - - y1 = pmul(e, p4f_cephes_log_q1); - tmp = pmul(x2, p4f_half); - y = padd(y, y1); - x = psub(x, tmp); - y2 = pmul(e, p4f_cephes_log_q2); - x = padd(x, y); - x = padd(x, y2); - // negative arg will be NAN, 0 will be -INF - return _mm_or_ps(_mm_andnot_ps(iszero_mask, _mm_or_ps(x, invalid_mask)), - _mm_and_ps(iszero_mask, p4f_minus_inf)); + return plog_float(_x); } template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f pexp<Packet4f>(const Packet4f& _x) { - Packet4f x = _x; - _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f); - _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f); - _EIGEN_DECLARE_CONST_Packet4i(0x7f, 0x7f); - - - _EIGEN_DECLARE_CONST_Packet4f(exp_hi, 88.3762626647950f); - _EIGEN_DECLARE_CONST_Packet4f(exp_lo, -88.3762626647949f); - - _EIGEN_DECLARE_CONST_Packet4f(cephes_LOG2EF, 1.44269504088896341f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C1, 0.693359375f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C2, -2.12194440e-4f); - - _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p0, 1.9875691500E-4f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p1, 1.3981999507E-3f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p2, 8.3334519073E-3f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p3, 4.1665795894E-2f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459E-1f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201E-1f); - - Packet4f tmp, fx; - Packet4i emm0; - - // clamp x - x = pmax(pmin(x, p4f_exp_hi), p4f_exp_lo); - - /* express exp(x) as exp(g + n*log(2)) */ - fx = pmadd(x, p4f_cephes_LOG2EF, p4f_half); - -#ifdef EIGEN_VECTORIZE_SSE4_1 - fx = _mm_floor_ps(fx); -#else - emm0 = _mm_cvttps_epi32(fx); - tmp = _mm_cvtepi32_ps(emm0); - /* if greater, substract 1 */ - Packet4f mask = _mm_cmpgt_ps(tmp, fx); - mask = _mm_and_ps(mask, p4f_1); - fx = psub(tmp, mask); -#endif - - tmp = pmul(fx, p4f_cephes_exp_C1); - Packet4f z = pmul(fx, p4f_cephes_exp_C2); - x = psub(x, tmp); - x = psub(x, z); - - z = pmul(x,x); - - Packet4f y = p4f_cephes_exp_p0; - y = pmadd(y, x, p4f_cephes_exp_p1); - y = pmadd(y, x, p4f_cephes_exp_p2); - y = pmadd(y, x, p4f_cephes_exp_p3); - y = pmadd(y, x, p4f_cephes_exp_p4); - y = pmadd(y, x, p4f_cephes_exp_p5); - y = pmadd(y, z, x); - y = padd(y, p4f_1); - - // build 2^n - emm0 = _mm_cvttps_epi32(fx); - emm0 = _mm_add_epi32(emm0, p4i_0x7f); - emm0 = _mm_slli_epi32(emm0, 23); - return pmax(pmul(y, Packet4f(_mm_castsi128_ps(emm0))), _x); + return pexp_float(_x); } + template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED -Packet2d pexp<Packet2d>(const Packet2d& _x) +Packet2d pexp<Packet2d>(const Packet2d& x) { - Packet2d x = _x; - - _EIGEN_DECLARE_CONST_Packet2d(1 , 1.0); - _EIGEN_DECLARE_CONST_Packet2d(2 , 2.0); - _EIGEN_DECLARE_CONST_Packet2d(half, 0.5); - - _EIGEN_DECLARE_CONST_Packet2d(exp_hi, 709.437); - _EIGEN_DECLARE_CONST_Packet2d(exp_lo, -709.436139303); - - _EIGEN_DECLARE_CONST_Packet2d(cephes_LOG2EF, 1.4426950408889634073599); - - _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p0, 1.26177193074810590878e-4); - _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p1, 3.02994407707441961300e-2); - _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p2, 9.99999999999999999910e-1); - - _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q0, 3.00198505138664455042e-6); - _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q1, 2.52448340349684104192e-3); - _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q2, 2.27265548208155028766e-1); - _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q3, 2.00000000000000000009e0); - - _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C1, 0.693145751953125); - _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6); - static const __m128i p4i_1023_0 = _mm_setr_epi32(1023, 1023, 0, 0); - - Packet2d tmp, fx; - Packet4i emm0; - - // clamp x - x = pmax(pmin(x, p2d_exp_hi), p2d_exp_lo); - /* express exp(x) as exp(g + n*log(2)) */ - fx = pmadd(p2d_cephes_LOG2EF, x, p2d_half); - -#ifdef EIGEN_VECTORIZE_SSE4_1 - fx = _mm_floor_pd(fx); -#else - emm0 = _mm_cvttpd_epi32(fx); - tmp = _mm_cvtepi32_pd(emm0); - /* if greater, substract 1 */ - Packet2d mask = _mm_cmpgt_pd(tmp, fx); - mask = _mm_and_pd(mask, p2d_1); - fx = psub(tmp, mask); -#endif - - tmp = pmul(fx, p2d_cephes_exp_C1); - Packet2d z = pmul(fx, p2d_cephes_exp_C2); - x = psub(x, tmp); - x = psub(x, z); - - Packet2d x2 = pmul(x,x); - - Packet2d px = p2d_cephes_exp_p0; - px = pmadd(px, x2, p2d_cephes_exp_p1); - px = pmadd(px, x2, p2d_cephes_exp_p2); - px = pmul (px, x); - - Packet2d qx = p2d_cephes_exp_q0; - qx = pmadd(qx, x2, p2d_cephes_exp_q1); - qx = pmadd(qx, x2, p2d_cephes_exp_q2); - qx = pmadd(qx, x2, p2d_cephes_exp_q3); - - x = pdiv(px,psub(qx,px)); - x = pmadd(p2d_2,x,p2d_1); - - // build 2^n - emm0 = _mm_cvttpd_epi32(fx); - emm0 = _mm_add_epi32(emm0, p4i_1023_0); - emm0 = _mm_slli_epi32(emm0, 20); - emm0 = _mm_shuffle_epi32(emm0, _MM_SHUFFLE(1,2,0,3)); - return pmax(pmul(x, Packet2d(_mm_castsi128_pd(emm0))), _x); + return pexp_double(x); } -/* evaluation of 4 sines at once, using SSE2 intrinsics. - - The code is the exact rewriting of the cephes sinf function. - Precision is excellent as long as x < 8192 (I did not bother to - take into account the special handling they have for greater values - -- it does not return garbage for arguments over 8192, though, but - the extra precision is missing). - - Note that it is such that sinf((float)M_PI) = 8.74e-8, which is the - surprising but correct result. -*/ - template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f psin<Packet4f>(const Packet4f& _x) { - Packet4f x = _x; - _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f); - _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f); - - _EIGEN_DECLARE_CONST_Packet4i(1, 1); - _EIGEN_DECLARE_CONST_Packet4i(not1, ~1); - _EIGEN_DECLARE_CONST_Packet4i(2, 2); - _EIGEN_DECLARE_CONST_Packet4i(4, 4); - - _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(sign_mask, 0x80000000); - - _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP1,-0.78515625f); - _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP2, -2.4187564849853515625e-4f); - _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP3, -3.77489497744594108e-8f); - _EIGEN_DECLARE_CONST_Packet4f(sincof_p0, -1.9515295891E-4f); - _EIGEN_DECLARE_CONST_Packet4f(sincof_p1, 8.3321608736E-3f); - _EIGEN_DECLARE_CONST_Packet4f(sincof_p2, -1.6666654611E-1f); - _EIGEN_DECLARE_CONST_Packet4f(coscof_p0, 2.443315711809948E-005f); - _EIGEN_DECLARE_CONST_Packet4f(coscof_p1, -1.388731625493765E-003f); - _EIGEN_DECLARE_CONST_Packet4f(coscof_p2, 4.166664568298827E-002f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_FOPI, 1.27323954473516f); // 4 / M_PI - - Packet4f xmm1, xmm2, xmm3, sign_bit, y; - - Packet4i emm0, emm2; - sign_bit = x; - /* take the absolute value */ - x = pabs(x); - - /* take the modulo */ - - /* extract the sign bit (upper one) */ - sign_bit = _mm_and_ps(sign_bit, p4f_sign_mask); - - /* scale by 4/Pi */ - y = pmul(x, p4f_cephes_FOPI); - - /* store the integer part of y in mm0 */ - emm2 = _mm_cvttps_epi32(y); - /* j=(j+1) & (~1) (see the cephes sources) */ - emm2 = _mm_add_epi32(emm2, p4i_1); - emm2 = _mm_and_si128(emm2, p4i_not1); - y = _mm_cvtepi32_ps(emm2); - /* get the swap sign flag */ - emm0 = _mm_and_si128(emm2, p4i_4); - emm0 = _mm_slli_epi32(emm0, 29); - /* get the polynom selection mask - there is one polynom for 0 <= x <= Pi/4 - and another one for Pi/4<x<=Pi/2 - - Both branches will be computed. - */ - emm2 = _mm_and_si128(emm2, p4i_2); - emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128()); - - Packet4f swap_sign_bit = _mm_castsi128_ps(emm0); - Packet4f poly_mask = _mm_castsi128_ps(emm2); - sign_bit = _mm_xor_ps(sign_bit, swap_sign_bit); - - /* The magic pass: "Extended precision modular arithmetic" - x = ((x - y * DP1) - y * DP2) - y * DP3; */ - xmm1 = pmul(y, p4f_minus_cephes_DP1); - xmm2 = pmul(y, p4f_minus_cephes_DP2); - xmm3 = pmul(y, p4f_minus_cephes_DP3); - x = padd(x, xmm1); - x = padd(x, xmm2); - x = padd(x, xmm3); - - /* Evaluate the first polynom (0 <= x <= Pi/4) */ - y = p4f_coscof_p0; - Packet4f z = _mm_mul_ps(x,x); - - y = pmadd(y, z, p4f_coscof_p1); - y = pmadd(y, z, p4f_coscof_p2); - y = pmul(y, z); - y = pmul(y, z); - Packet4f tmp = pmul(z, p4f_half); - y = psub(y, tmp); - y = padd(y, p4f_1); - - /* Evaluate the second polynom (Pi/4 <= x <= 0) */ - - Packet4f y2 = p4f_sincof_p0; - y2 = pmadd(y2, z, p4f_sincof_p1); - y2 = pmadd(y2, z, p4f_sincof_p2); - y2 = pmul(y2, z); - y2 = pmul(y2, x); - y2 = padd(y2, x); - - /* select the correct result from the two polynoms */ - y2 = _mm_and_ps(poly_mask, y2); - y = _mm_andnot_ps(poly_mask, y); - y = _mm_or_ps(y,y2); - /* update the sign */ - return _mm_xor_ps(y, sign_bit); + return psin_float(_x); } -/* almost the same as psin */ template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f pcos<Packet4f>(const Packet4f& _x) { - Packet4f x = _x; - _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f); - _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f); - - _EIGEN_DECLARE_CONST_Packet4i(1, 1); - _EIGEN_DECLARE_CONST_Packet4i(not1, ~1); - _EIGEN_DECLARE_CONST_Packet4i(2, 2); - _EIGEN_DECLARE_CONST_Packet4i(4, 4); - - _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP1,-0.78515625f); - _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP2, -2.4187564849853515625e-4f); - _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP3, -3.77489497744594108e-8f); - _EIGEN_DECLARE_CONST_Packet4f(sincof_p0, -1.9515295891E-4f); - _EIGEN_DECLARE_CONST_Packet4f(sincof_p1, 8.3321608736E-3f); - _EIGEN_DECLARE_CONST_Packet4f(sincof_p2, -1.6666654611E-1f); - _EIGEN_DECLARE_CONST_Packet4f(coscof_p0, 2.443315711809948E-005f); - _EIGEN_DECLARE_CONST_Packet4f(coscof_p1, -1.388731625493765E-003f); - _EIGEN_DECLARE_CONST_Packet4f(coscof_p2, 4.166664568298827E-002f); - _EIGEN_DECLARE_CONST_Packet4f(cephes_FOPI, 1.27323954473516f); // 4 / M_PI - - Packet4f xmm1, xmm2, xmm3, y; - Packet4i emm0, emm2; - - x = pabs(x); - - /* scale by 4/Pi */ - y = pmul(x, p4f_cephes_FOPI); - - /* get the integer part of y */ - emm2 = _mm_cvttps_epi32(y); - /* j=(j+1) & (~1) (see the cephes sources) */ - emm2 = _mm_add_epi32(emm2, p4i_1); - emm2 = _mm_and_si128(emm2, p4i_not1); - y = _mm_cvtepi32_ps(emm2); - - emm2 = _mm_sub_epi32(emm2, p4i_2); - - /* get the swap sign flag */ - emm0 = _mm_andnot_si128(emm2, p4i_4); - emm0 = _mm_slli_epi32(emm0, 29); - /* get the polynom selection mask */ - emm2 = _mm_and_si128(emm2, p4i_2); - emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128()); - - Packet4f sign_bit = _mm_castsi128_ps(emm0); - Packet4f poly_mask = _mm_castsi128_ps(emm2); - - /* The magic pass: "Extended precision modular arithmetic" - x = ((x - y * DP1) - y * DP2) - y * DP3; */ - xmm1 = pmul(y, p4f_minus_cephes_DP1); - xmm2 = pmul(y, p4f_minus_cephes_DP2); - xmm3 = pmul(y, p4f_minus_cephes_DP3); - x = padd(x, xmm1); - x = padd(x, xmm2); - x = padd(x, xmm3); - - /* Evaluate the first polynom (0 <= x <= Pi/4) */ - y = p4f_coscof_p0; - Packet4f z = pmul(x,x); - - y = pmadd(y,z,p4f_coscof_p1); - y = pmadd(y,z,p4f_coscof_p2); - y = pmul(y, z); - y = pmul(y, z); - Packet4f tmp = _mm_mul_ps(z, p4f_half); - y = psub(y, tmp); - y = padd(y, p4f_1); - - /* Evaluate the second polynom (Pi/4 <= x <= 0) */ - Packet4f y2 = p4f_sincof_p0; - y2 = pmadd(y2, z, p4f_sincof_p1); - y2 = pmadd(y2, z, p4f_sincof_p2); - y2 = pmul(y2, z); - y2 = pmadd(y2, x, x); - - /* select the correct result from the two polynoms */ - y2 = _mm_and_ps(poly_mask, y2); - y = _mm_andnot_ps(poly_mask, y); - y = _mm_or_ps(y,y2); - - /* update the sign */ - return _mm_xor_ps(y, sign_bit); + return pcos_float(_x); } #if EIGEN_FAST_MATH @@ -482,11 +91,11 @@ Packet2d psqrt<Packet2d>(const Packet2d& x) { return _mm_sqrt_pd(x); } template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f prsqrt<Packet4f>(const Packet4f& _x) { - _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(inf, 0x7f800000); - _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(nan, 0x7fc00000); + _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(inf, 0x7f800000u); + _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(nan, 0x7fc00000u); _EIGEN_DECLARE_CONST_Packet4f(one_point_five, 1.5f); _EIGEN_DECLARE_CONST_Packet4f(minus_half, -0.5f); - _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(flt_min, 0x00800000); + _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(flt_min, 0x00800000u); Packet4f neg_half = pmul(_x, p4f_minus_half); diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index 99d55d5e9..9c3750af0 100755 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -18,11 +18,13 @@ namespace internal { #define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8 #endif -#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS +#if !defined(EIGEN_VECTORIZE_AVX) && !defined(EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS) +// 32 bits => 8 registers +// 64 bits => 16 registers #define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*)) #endif -#ifdef __FMA__ +#ifdef EIGEN_VECTORIZE_FMA #ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD #define EIGEN_HAS_SINGLE_INSTRUCTION_MADD 1 #endif @@ -61,20 +63,22 @@ template<> struct is_arithmetic<__m128> { enum { value = true }; }; template<> struct is_arithmetic<__m128i> { enum { value = true }; }; template<> struct is_arithmetic<__m128d> { enum { value = true }; }; +#define EIGEN_SSE_SHUFFLE_MASK(p,q,r,s) ((s)<<6|(r)<<4|(q)<<2|(p)) + #define vec4f_swizzle1(v,p,q,r,s) \ - (_mm_castsi128_ps(_mm_shuffle_epi32( _mm_castps_si128(v), ((s)<<6|(r)<<4|(q)<<2|(p))))) + (_mm_castsi128_ps(_mm_shuffle_epi32( _mm_castps_si128(v), EIGEN_SSE_SHUFFLE_MASK(p,q,r,s)))) #define vec4i_swizzle1(v,p,q,r,s) \ - (_mm_shuffle_epi32( v, ((s)<<6|(r)<<4|(q)<<2|(p)))) + (_mm_shuffle_epi32( v, EIGEN_SSE_SHUFFLE_MASK(p,q,r,s))) #define vec2d_swizzle1(v,p,q) \ - (_mm_castsi128_pd(_mm_shuffle_epi32( _mm_castpd_si128(v), ((q*2+1)<<6|(q*2)<<4|(p*2+1)<<2|(p*2))))) + (_mm_castsi128_pd(_mm_shuffle_epi32( _mm_castpd_si128(v), EIGEN_SSE_SHUFFLE_MASK(2*p,2*p+1,2*q,2*q+1)))) #define vec4f_swizzle2(a,b,p,q,r,s) \ - (_mm_shuffle_ps( (a), (b), ((s)<<6|(r)<<4|(q)<<2|(p)))) + (_mm_shuffle_ps( (a), (b), EIGEN_SSE_SHUFFLE_MASK(p,q,r,s))) #define vec4i_swizzle2(a,b,p,q,r,s) \ - (_mm_castps_si128( (_mm_shuffle_ps( _mm_castsi128_ps(a), _mm_castsi128_ps(b), ((s)<<6|(r)<<4|(q)<<2|(p)))))) + (_mm_castps_si128( (_mm_shuffle_ps( _mm_castsi128_ps(a), _mm_castsi128_ps(b), EIGEN_SSE_SHUFFLE_MASK(p,q,r,s))))) #define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \ const Packet4f p4f_##NAME = pset1<Packet4f>(X) @@ -83,7 +87,7 @@ template<> struct is_arithmetic<__m128d> { enum { value = true }; }; const Packet2d p2d_##NAME = pset1<Packet2d>(X) #define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \ - const Packet4f p4f_##NAME = _mm_castsi128_ps(pset1<Packet4i>(X)) + const Packet4f p4f_##NAME = pset1frombits<Packet4f>(X) #define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \ const Packet4i p4i_##NAME = pset1<Packet4i>(X) @@ -110,12 +114,12 @@ template<> struct packet_traits<float> : default_packet_traits HasSqrt = 1, HasRsqrt = 1, HasTanh = EIGEN_FAST_MATH, - HasBlend = 1 + HasBlend = 1, + HasFloor = 1 #ifdef EIGEN_VECTORIZE_SSE4_1 , HasRound = 1, - HasFloor = 1, HasCeil = 1 #endif }; @@ -158,9 +162,22 @@ template<> struct packet_traits<int> : default_packet_traits }; }; -template<> struct unpacket_traits<Packet4f> { typedef float type; enum {size=4, alignment=Aligned16}; typedef Packet4f half; }; -template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16}; typedef Packet2d half; }; -template<> struct unpacket_traits<Packet4i> { typedef int type; enum {size=4, alignment=Aligned16}; typedef Packet4i half; }; +template<> struct unpacket_traits<Packet4f> { + typedef float type; + typedef Packet4f half; + typedef Packet4i integer_packet; + enum {size=4, alignment=Aligned16, vectorizable=true}; +}; +template<> struct unpacket_traits<Packet2d> { + typedef double type; + typedef Packet2d half; + enum {size=2, alignment=Aligned16, vectorizable=true}; +}; +template<> struct unpacket_traits<Packet4i> { + typedef int type; + typedef Packet4i half; + enum {size=4, alignment=Aligned16, vectorizable=false}; +}; #ifndef EIGEN_VECTORIZE_AVX template<> struct scalar_div_cost<float,true> { enum { value = 7 }; }; @@ -180,6 +197,12 @@ template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { re template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) { return _mm_set1_epi32(from); } #endif +template<> EIGEN_STRONG_INLINE Packet4f pset1frombits<Packet4f>(unsigned int from) { return _mm_castsi128_ps(pset1<Packet4i>(from)); } + +template<> EIGEN_STRONG_INLINE Packet4f pzero(const Packet4f& /*a*/) { return _mm_setzero_ps(); } +template<> EIGEN_STRONG_INLINE Packet2d pzero(const Packet2d& /*a*/) { return _mm_setzero_pd(); } +template<> EIGEN_STRONG_INLINE Packet4i pzero(const Packet4i& /*a*/) { return _mm_setzero_si128(); } + // GCC generates a shufps instruction for _mm_set1_ps/_mm_load1_ps instead of the more efficient pshufd instruction. // However, using inrinsics for pset1 makes gcc to generate crappy code in some cases (see bug 203) // Using inline assembly is also not an option because then gcc fails to reorder properly the instructions. @@ -245,19 +268,24 @@ template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const // for some weird raisons, it has to be overloaded for packet of integers template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd(pmul(a,b), c); } -#ifdef __FMA__ +#ifdef EIGEN_VECTORIZE_FMA template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return _mm_fmadd_ps(a,b,c); } template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return _mm_fmadd_pd(a,b,c); } #endif template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { -#if EIGEN_COMP_GNUC +#if EIGEN_COMP_GNUC && EIGEN_COMP_GNUC < 63 // There appears to be a bug in GCC, by which the optimizer may // flip the argument order in calls to _mm_min_ps, so we have to // resort to inline ASM here. This is supposed to be fixed in gcc6.3, // see also: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72867 + #ifdef EIGEN_VECTORIZE_AVX + Packet4f res; + asm("vminps %[a], %[b], %[res]" : [res] "=x" (res) : [a] "x" (a), [b] "x" (b)); + #else Packet4f res = b; asm("minps %[a], %[res]" : [res] "+x" (res) : [a] "x" (a)); + #endif return res; #else // Arguments are reversed to match NaN propagation behavior of std::min. @@ -265,13 +293,18 @@ template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const #endif } template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { -#if EIGEN_COMP_GNUC +#if EIGEN_COMP_GNUC && EIGEN_COMP_GNUC < 63 // There appears to be a bug in GCC, by which the optimizer may // flip the argument order in calls to _mm_min_pd, so we have to // resort to inline ASM here. This is supposed to be fixed in gcc6.3, // see also: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72867 + #ifdef EIGEN_VECTORIZE_AVX + Packet2d res; + asm("vminpd %[a], %[b], %[res]" : [res] "=x" (res) : [a] "x" (a), [b] "x" (b)); + #else Packet2d res = b; asm("minpd %[a], %[res]" : [res] "+x" (res) : [a] "x" (a)); + #endif return res; #else // Arguments are reversed to match NaN propagation behavior of std::min. @@ -290,13 +323,18 @@ template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const } template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) { -#if EIGEN_COMP_GNUC +#if EIGEN_COMP_GNUC && EIGEN_COMP_GNUC < 63 // There appears to be a bug in GCC, by which the optimizer may // flip the argument order in calls to _mm_max_ps, so we have to // resort to inline ASM here. This is supposed to be fixed in gcc6.3, // see also: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72867 + #ifdef EIGEN_VECTORIZE_AVX + Packet4f res; + asm("vmaxps %[a], %[b], %[res]" : [res] "=x" (res) : [a] "x" (a), [b] "x" (b)); + #else Packet4f res = b; asm("maxps %[a], %[res]" : [res] "+x" (res) : [a] "x" (a)); + #endif return res; #else // Arguments are reversed to match NaN propagation behavior of std::max. @@ -304,13 +342,18 @@ template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const #endif } template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { -#if EIGEN_COMP_GNUC +#if EIGEN_COMP_GNUC && EIGEN_COMP_GNUC < 63 // There appears to be a bug in GCC, by which the optimizer may // flip the argument order in calls to _mm_max_pd, so we have to // resort to inline ASM here. This is supposed to be fixed in gcc6.3, // see also: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72867 + #ifdef EIGEN_VECTORIZE_AVX + Packet2d res; + asm("vmaxpd %[a], %[b], %[res]" : [res] "=x" (res) : [a] "x" (a), [b] "x" (b)); + #else Packet2d res = b; asm("maxpd %[a], %[res]" : [res] "+x" (res) : [a] "x" (a)); + #endif return res; #else // Arguments are reversed to match NaN propagation behavior of std::max. @@ -328,16 +371,24 @@ template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const #endif } -#ifdef EIGEN_VECTORIZE_SSE4_1 -template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a) { return _mm_round_ps(a, 0); } -template<> EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a) { return _mm_round_pd(a, 0); } - -template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const Packet4f& a) { return _mm_ceil_ps(a); } -template<> EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const Packet2d& a) { return _mm_ceil_pd(a); } - -template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) { return _mm_floor_ps(a); } -template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) { return _mm_floor_pd(a); } -#endif +template<> EIGEN_STRONG_INLINE Packet4f pcmp_le(const Packet4f& a, const Packet4f& b) { return _mm_cmple_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet4f pcmp_lt(const Packet4f& a, const Packet4f& b) { return _mm_cmplt_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet4f pcmp_eq(const Packet4f& a, const Packet4f& b) { return _mm_cmpeq_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i pcmp_eq(const Packet4i& a, const Packet4i& b) { return _mm_cmpeq_epi32(a,b); } +template<> EIGEN_STRONG_INLINE Packet2d pcmp_eq(const Packet2d& a, const Packet2d& b) { return _mm_cmpeq_pd(a,b); } +template<> EIGEN_STRONG_INLINE Packet4f pcmp_lt_or_nan(const Packet4f& a, const Packet4f& b) { return _mm_cmpnge_ps(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4i ptrue<Packet4i>(const Packet4i& a) { return _mm_cmpeq_epi32(a, a); } +template<> EIGEN_STRONG_INLINE Packet4f +ptrue<Packet4f>(const Packet4f& a) { + Packet4i b = _mm_castps_si128(a); + return _mm_castsi128_ps(_mm_cmpeq_epi32(b, b)); +} +template<> EIGEN_STRONG_INLINE Packet2d +ptrue<Packet2d>(const Packet2d& a) { + Packet4i b = _mm_castpd_si128(a); + return _mm_castsi128_pd(_mm_cmpeq_epi32(b, b)); +} template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_and_ps(a,b); } template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_and_pd(a,b); } @@ -351,9 +402,47 @@ template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_xor_pd(a,b); } template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_xor_si128(a,b); } -template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_andnot_ps(a,b); } -template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_andnot_pd(a,b); } -template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_andnot_si128(a,b); } +template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_andnot_ps(b,a); } +template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_andnot_pd(b,a); } +template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_andnot_si128(b,a); } + +template<int N> EIGEN_STRONG_INLINE Packet4i pshiftright(Packet4i a) { return _mm_srli_epi32(a,N); } +template<int N> EIGEN_STRONG_INLINE Packet4i pshiftleft(Packet4i a) { return _mm_slli_epi32(a,N); } + +#ifdef EIGEN_VECTORIZE_SSE4_1 +template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a) { return _mm_round_ps(a, 0); } +template<> EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a) { return _mm_round_pd(a, 0); } + +template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const Packet4f& a) { return _mm_ceil_ps(a); } +template<> EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const Packet2d& a) { return _mm_ceil_pd(a); } + +template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) { return _mm_floor_ps(a); } +template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) { return _mm_floor_pd(a); } +#else +template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) +{ + const Packet4f cst_1 = pset1<Packet4f>(1.0f); + Packet4i emm0 = _mm_cvttps_epi32(a); + Packet4f tmp = _mm_cvtepi32_ps(emm0); + /* if greater, substract 1 */ + Packet4f mask = _mm_cmpgt_ps(tmp, a); + mask = pand(mask, cst_1); + return psub(tmp, mask); +} + +// WARNING: this pfloor implementation makes sense for small inputs only, +// It is currently only used by pexp and not exposed through HasFloor. +template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) +{ + const Packet2d cst_1 = pset1<Packet2d>(1.0); + Packet4i emm0 = _mm_cvttpd_epi32(a); + Packet2d tmp = _mm_cvtepi32_pd(emm0); + /* if greater, substract 1 */ + Packet2d mask = _mm_cmpgt_pd(tmp, a); + mask = pand(mask, cst_1); + return psub(tmp, mask); +} +#endif template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_ps(from); } template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_pd(from); } @@ -517,6 +606,23 @@ template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) #endif } +template<> EIGEN_STRONG_INLINE Packet4f pfrexp<Packet4f>(const Packet4f& a, Packet4f& exponent) { + return pfrexp_float(a,exponent); +} + +template<> EIGEN_STRONG_INLINE Packet4f pldexp<Packet4f>(const Packet4f& a, const Packet4f& exponent) { + return pldexp_float(a,exponent); +} + +template<> EIGEN_STRONG_INLINE Packet2d pldexp<Packet2d>(const Packet2d& a, const Packet2d& exponent) { + const Packet4i cst_1023_0 = _mm_setr_epi32(1023, 1023, 0, 0); + Packet4i emm0 = _mm_cvttpd_epi32(exponent); + emm0 = padd(emm0, cst_1023_0); + emm0 = _mm_slli_epi32(emm0, 20); + emm0 = _mm_shuffle_epi32(emm0, _MM_SHUFFLE(1,2,0,3)); + return pmul(a, Packet2d(_mm_castsi128_pd(emm0))); +} + // with AVX, the default implementations based on pload1 are faster #ifndef __AVX__ template<> EIGEN_STRONG_INLINE void @@ -718,6 +824,17 @@ template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a) #endif // EIGEN_VECTORIZE_SSE4_1 } +// not needed yet +// template<> EIGEN_STRONG_INLINE bool predux_all(const Packet4f& x) +// { +// return _mm_movemask_ps(x) == 0xF; +// } + +template<> EIGEN_STRONG_INLINE bool predux_any(const Packet4f& x) +{ + return _mm_movemask_ps(x) != 0x0; +} + #if EIGEN_COMP_GNUC // template <> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) // { @@ -921,7 +1038,7 @@ template<> EIGEN_STRONG_INLINE Packet2d pinsertlast(const Packet2d& a, double b) } // Scalar path for pmadd with FMA to ensure consistency with vectorized path. -#ifdef __FMA__ +#ifdef EIGEN_VECTORIZE_FMA template<> EIGEN_STRONG_INLINE float pmadd(const float& a, const float& b, const float& c) { return ::fmaf(a,b,c); } diff --git a/Eigen/src/Core/arch/SSE/TypeCasting.h b/Eigen/src/Core/arch/SSE/TypeCasting.h index c6ca8c716..f607366f0 100644 --- a/Eigen/src/Core/arch/SSE/TypeCasting.h +++ b/Eigen/src/Core/arch/SSE/TypeCasting.h @@ -69,6 +69,13 @@ template<> EIGEN_STRONG_INLINE Packet2d pcast<Packet4f, Packet2d>(const Packet4f return _mm_cvtps_pd(a); } +template<> EIGEN_STRONG_INLINE Packet4i preinterpret<Packet4i,Packet4f>(const Packet4f& a) { + return _mm_castps_si128(a); +} + +template<> EIGEN_STRONG_INLINE Packet4f preinterpret<Packet4f,Packet4i>(const Packet4i& a) { + return _mm_castsi128_ps(a); +} } // end namespace internal diff --git a/Eigen/src/Core/arch/SYCL/InteropHeaders.h b/Eigen/src/Core/arch/SYCL/InteropHeaders.h index c1da40d14..294cb101a 100644 --- a/Eigen/src/Core/arch/SYCL/InteropHeaders.h +++ b/Eigen/src/Core/arch/SYCL/InteropHeaders.h @@ -88,7 +88,7 @@ SYCL_ARITHMETIC(cl::sycl::cl_double2) #define SYCL_UNPACKET_TRAITS(packet_type, unpacket_type, lengths)\ template<> struct unpacket_traits<packet_type> {\ typedef unpacket_type type;\ - enum {size=lengths, alignment=Aligned16};\ + enum {size=lengths, alignment=Aligned16, vectorizable=true};\ typedef packet_type half;\ }; SYCL_UNPACKET_TRAITS(cl::sycl::cl_float4, float, 4) diff --git a/Eigen/src/Core/arch/ZVector/Complex.h b/Eigen/src/Core/arch/ZVector/Complex.h index 95aba428f..167c3ee4c 100644 --- a/Eigen/src/Core/arch/ZVector/Complex.h +++ b/Eigen/src/Core/arch/ZVector/Complex.h @@ -91,8 +91,8 @@ template<> struct packet_traits<std::complex<double> > : default_packet_traits }; }; -template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16}; typedef Packet2cf half; }; -template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16}; typedef Packet1cd half; }; +template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16, vectorizable=true}; typedef Packet2cf half; }; +template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16, vectorizable=true}; typedef Packet1cd half; }; /* Forward declaration */ EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2cf,2>& kernel); diff --git a/Eigen/src/Core/arch/ZVector/PacketMath.h b/Eigen/src/Core/arch/ZVector/PacketMath.h index 0b37f4992..c8e90f1a8 100755 --- a/Eigen/src/Core/arch/ZVector/PacketMath.h +++ b/Eigen/src/Core/arch/ZVector/PacketMath.h @@ -239,9 +239,9 @@ template<> struct packet_traits<double> : default_packet_traits }; }; -template<> struct unpacket_traits<Packet4i> { typedef int type; enum {size=4, alignment=Aligned16}; typedef Packet4i half; }; -template<> struct unpacket_traits<Packet4f> { typedef float type; enum {size=4, alignment=Aligned16}; typedef Packet4f half; }; -template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16}; typedef Packet2d half; }; +template<> struct unpacket_traits<Packet4i> { typedef int type; enum {size=4, alignment=Aligned16, vectorizable=true}; typedef Packet4i half; }; +template<> struct unpacket_traits<Packet4f> { typedef float type; enum {size=4, alignment=Aligned16, vectorizable=true}; typedef Packet4f half; }; +template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16, vectorizable=true}; typedef Packet2d half; }; /* Forward declaration */ EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4f,4>& kernel); diff --git a/Eigen/src/Core/functors/AssignmentFunctors.h b/Eigen/src/Core/functors/AssignmentFunctors.h index 9765cc763..bf64ef4ed 100644 --- a/Eigen/src/Core/functors/AssignmentFunctors.h +++ b/Eigen/src/Core/functors/AssignmentFunctors.h @@ -157,7 +157,16 @@ template<typename Scalar> struct functor_traits<swap_assign_op<Scalar> > { enum { Cost = 3 * NumTraits<Scalar>::ReadCost, - PacketAccess = packet_traits<Scalar>::Vectorizable + PacketAccess = + #if defined(EIGEN_VECTORIZE_AVX) && EIGEN_COMP_CLANG && (EIGEN_COMP_CLANG<800 || defined(__apple_build_version__)) + // This is a partial workaround for a bug in clang generating bad code + // when mixing 256/512 bits loads and 128 bits moves. + // See http://eigen.tuxfamily.org/bz/show_bug.cgi?id=1684 + // https://bugs.llvm.org/show_bug.cgi?id=40815 + 0 + #else + packet_traits<Scalar>::Vectorizable + #endif }; }; diff --git a/Eigen/src/Core/functors/NullaryFunctors.h b/Eigen/src/Core/functors/NullaryFunctors.h index b03be0269..16b645f91 100644 --- a/Eigen/src/Core/functors/NullaryFunctors.h +++ b/Eigen/src/Core/functors/NullaryFunctors.h @@ -37,26 +37,27 @@ template<typename Scalar> struct functor_traits<scalar_identity_op<Scalar> > { enum { Cost = NumTraits<Scalar>::AddCost, PacketAccess = false, IsRepeatable = true }; }; -template <typename Scalar, typename Packet, bool IsInteger> struct linspaced_op_impl; +template <typename Scalar, bool IsInteger> struct linspaced_op_impl; -template <typename Scalar, typename Packet> -struct linspaced_op_impl<Scalar,Packet,/*IsInteger*/false> +template <typename Scalar> +struct linspaced_op_impl<Scalar,/*IsInteger*/false> { + typedef typename NumTraits<Scalar>::Real RealScalar; + 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_low(low), m_high(high), m_size1(num_steps==1 ? 1 : num_steps-1), m_step(num_steps==1 ? Scalar() : (high-low)/RealScalar(num_steps-1)), m_flip(numext::abs(high)<numext::abs(low)) {} template<typename IndexType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (IndexType i) const { - typedef typename NumTraits<Scalar>::Real RealScalar; if(m_flip) return (i==0)? m_low : (m_high - RealScalar(m_size1-i)*m_step); else return (i==m_size1)? m_high : (m_low + RealScalar(i)*m_step); } - template<typename IndexType> + template<typename Packet, typename IndexType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(IndexType i) const { // Principle: @@ -86,8 +87,8 @@ struct linspaced_op_impl<Scalar,Packet,/*IsInteger*/false> const bool m_flip; }; -template <typename Scalar, typename Packet> -struct linspaced_op_impl<Scalar,Packet,/*IsInteger*/true> +template <typename Scalar> +struct linspaced_op_impl<Scalar,/*IsInteger*/true> { linspaced_op_impl(const Scalar& low, const Scalar& high, Index num_steps) : m_low(low), @@ -115,8 +116,8 @@ struct linspaced_op_impl<Scalar,Packet,/*IsInteger*/true> // Forward declaration (we default to random access which does not really give // us a speed gain when using packet access but it allows to use the functor in // nested expressions). -template <typename Scalar, typename PacketType> struct linspaced_op; -template <typename Scalar, typename PacketType> struct functor_traits< linspaced_op<Scalar,PacketType> > +template <typename Scalar> struct linspaced_op; +template <typename Scalar> struct functor_traits< linspaced_op<Scalar> > { enum { @@ -126,7 +127,7 @@ template <typename Scalar, typename PacketType> struct functor_traits< linspaced IsRepeatable = true }; }; -template <typename Scalar, typename PacketType> struct linspaced_op +template <typename Scalar> struct linspaced_op { linspaced_op(const Scalar& low, const Scalar& high, Index num_steps) : impl((num_steps==1 ? high : low),high,num_steps) @@ -136,11 +137,11 @@ template <typename Scalar, typename PacketType> struct linspaced_op EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (IndexType i) const { return impl(i); } template<typename Packet,typename IndexType> - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(IndexType i) const { return impl.packetOp(i); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(IndexType i) const { return impl.template packetOp<Packet>(i); } // This proxy object handles the actual required temporaries and the different // implementations (integer vs. floating point). - const linspaced_op_impl<Scalar,PacketType,NumTraits<Scalar>::IsInteger> impl; + const linspaced_op_impl<Scalar,NumTraits<Scalar>::IsInteger> impl; }; // Linear access is automatically determined from the operator() prototypes available for the given functor. @@ -166,12 +167,12 @@ struct has_unary_operator<scalar_identity_op<Scalar>,IndexType> { enum { value = template<typename Scalar,typename IndexType> struct has_binary_operator<scalar_identity_op<Scalar>,IndexType> { enum { value = 1}; }; -template<typename Scalar, typename PacketType,typename IndexType> -struct has_nullary_operator<linspaced_op<Scalar,PacketType>,IndexType> { enum { value = 0}; }; -template<typename Scalar, typename PacketType,typename IndexType> -struct has_unary_operator<linspaced_op<Scalar,PacketType>,IndexType> { enum { value = 1}; }; -template<typename Scalar, typename PacketType,typename IndexType> -struct has_binary_operator<linspaced_op<Scalar,PacketType>,IndexType> { enum { value = 0}; }; +template<typename Scalar,typename IndexType> +struct has_nullary_operator<linspaced_op<Scalar>,IndexType> { enum { value = 0}; }; +template<typename Scalar,typename IndexType> +struct has_unary_operator<linspaced_op<Scalar>,IndexType> { enum { value = 1}; }; +template<typename Scalar,typename IndexType> +struct has_binary_operator<linspaced_op<Scalar>,IndexType> { enum { value = 0}; }; template<typename Scalar,typename IndexType> struct has_nullary_operator<scalar_random_op<Scalar>,IndexType> { enum { value = 1}; }; diff --git a/Eigen/src/Core/functors/UnaryFunctors.h b/Eigen/src/Core/functors/UnaryFunctors.h index 0c2d2cfca..03f167ac9 100644 --- a/Eigen/src/Core/functors/UnaryFunctors.h +++ b/Eigen/src/Core/functors/UnaryFunctors.h @@ -117,7 +117,15 @@ template<typename Scalar> struct functor_traits<scalar_conjugate_op<Scalar> > { enum { - Cost = NumTraits<Scalar>::IsComplex ? NumTraits<Scalar>::AddCost : 0, + Cost = 0, + // Yes the cost is zero even for complexes because in most cases for which + // the cost is used, conjugation turns to be a no-op. Some examples: + // cost(a*conj(b)) == cost(a*b) + // cost(a+conj(b)) == cost(a+b) + // <etc. + // If we don't set it to zero, then: + // A.conjugate().lazyProduct(B.conjugate()) + // will bake its operands. We definitely don't want that! PacketAccess = packet_traits<Scalar>::HasConj }; }; @@ -548,6 +556,23 @@ struct functor_traits<scalar_tanh_op<Scalar> > { }; }; +#if EIGEN_HAS_CXX11_MATH +/** \internal + * \brief Template functor to compute the atanh of a scalar + * \sa class CwiseUnaryOp, ArrayBase::atanh() + */ +template <typename Scalar> +struct scalar_atanh_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_atanh_op) + EIGEN_DEVICE_FUNC inline const Scalar operator()(const Scalar& a) const { return numext::atanh(a); } +}; + +template <typename Scalar> +struct functor_traits<scalar_atanh_op<Scalar> > { + enum { Cost = 5 * NumTraits<Scalar>::MulCost, PacketAccess = false }; +}; +#endif + /** \internal * \brief Template functor to compute the sinh of a scalar * \sa class CwiseUnaryOp, ArrayBase::sinh() @@ -567,6 +592,23 @@ struct functor_traits<scalar_sinh_op<Scalar> > }; }; +#if EIGEN_HAS_CXX11_MATH +/** \internal + * \brief Template functor to compute the asinh of a scalar + * \sa class CwiseUnaryOp, ArrayBase::asinh() + */ +template <typename Scalar> +struct scalar_asinh_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_asinh_op) + EIGEN_DEVICE_FUNC inline const Scalar operator()(const Scalar& a) const { return numext::asinh(a); } +}; + +template <typename Scalar> +struct functor_traits<scalar_asinh_op<Scalar> > { + enum { Cost = 5 * NumTraits<Scalar>::MulCost, PacketAccess = false }; +}; +#endif + /** \internal * \brief Template functor to compute the cosh of a scalar * \sa class CwiseUnaryOp, ArrayBase::cosh() @@ -586,6 +628,23 @@ struct functor_traits<scalar_cosh_op<Scalar> > }; }; +#if EIGEN_HAS_CXX11_MATH +/** \internal + * \brief Template functor to compute the acosh of a scalar + * \sa class CwiseUnaryOp, ArrayBase::acosh() + */ +template <typename Scalar> +struct scalar_acosh_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_acosh_op) + EIGEN_DEVICE_FUNC inline const Scalar operator()(const Scalar& a) const { return numext::acosh(a); } +}; + +template <typename Scalar> +struct functor_traits<scalar_acosh_op<Scalar> > { + enum { Cost = 5 * NumTraits<Scalar>::MulCost, PacketAccess = false }; +}; +#endif + /** \internal * \brief Template functor to compute the inverse of a scalar * \sa class CwiseUnaryOp, Cwise::inverse() @@ -598,9 +657,13 @@ struct scalar_inverse_op { EIGEN_DEVICE_FUNC inline const Packet packetOp(const Packet& a) const { return internal::pdiv(pset1<Packet>(Scalar(1)),a); } }; -template<typename Scalar> -struct functor_traits<scalar_inverse_op<Scalar> > -{ enum { Cost = NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasDiv }; }; +template <typename Scalar> +struct functor_traits<scalar_inverse_op<Scalar> > { + enum { + PacketAccess = packet_traits<Scalar>::HasDiv, + Cost = scalar_div_cost<Scalar, PacketAccess>::value + }; +}; /** \internal * \brief Template functor to compute the square of a scalar @@ -864,8 +927,9 @@ template <> struct scalar_logistic_op<float> { EIGEN_EMPTY_STRUCT_CTOR(scalar_logistic_op) EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float operator()(const float& x) const { - const float one = 1.0f; - return one / (one + numext::exp(-x)); + if (x < -18.0f) return 0.0f; + else if (x > 18.0f) return 1.0f; + else return 1.0f / (1.0f + numext::exp(-x)); } template <typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index e7cab4720..fdd0ec0e9 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -15,7 +15,13 @@ namespace Eigen { namespace internal { -template<typename _LhsScalar, typename _RhsScalar, bool _ConjLhs=false, bool _ConjRhs=false> +enum PacketSizeType { + PacketFull = 0, + PacketHalf, + PacketQuarter +}; + +template<typename _LhsScalar, typename _RhsScalar, bool _ConjLhs=false, bool _ConjRhs=false, int Arch=Architecture::Target, int _PacketSize=PacketFull> class gebp_traits; @@ -101,6 +107,16 @@ void evaluateProductBlockingSizesHeuristic(Index& k, Index& m, Index& n, Index n // at the register level. This small horizontal panel has to stay within L1 cache. std::ptrdiff_t l1, l2, l3; manage_caching_sizes(GetAction, &l1, &l2, &l3); + #ifdef EIGEN_VECTORIZE_AVX512 + // We need to find a rationale for that, but without this adjustment, + // performance with AVX512 is pretty bad, like -20% slower. + // One reason is that with increasing packet-size, the blocking size k + // has to become pretty small if we want that 1 lhs panel fit within L1. + // For instance, with the 3pX4 kernel and double, the size of the lhs+rhs panels are: + // k*(3*64 + 4*8) Bytes, with l1=32kBytes, and k%8=0, we have k=144. + // This is quite small for a good reuse of the accumulation registers. + l1 *= 4; + #endif if (num_threads > 1) { typedef typename Traits::ResScalar ResScalar; @@ -337,6 +353,61 @@ inline void computeProductBlockingSizes(Index& k, Index& m, Index& n, Index num_ // #define CJMADD(CJ,A,B,C,T) T = B; T = CJ.pmul(A,T); C = padd(C,T); #endif +template <typename RhsPacket, typename RhsPacketx4, int registers_taken> +struct RhsPanelHelper { + private: + static const int remaining_registers = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS - registers_taken; + public: + typedef typename conditional<remaining_registers>=4, RhsPacketx4, RhsPacket>::type type; +}; + +template <typename Packet> +struct QuadPacket +{ + Packet B_0, B1, B2, B3; + const Packet& get(const FixedInt<0>&) const { return B_0; } + const Packet& get(const FixedInt<1>&) const { return B1; } + const Packet& get(const FixedInt<2>&) const { return B2; } + const Packet& get(const FixedInt<3>&) const { return B3; } +}; + +template <int N, typename T1, typename T2, typename T3> +struct packet_conditional { typedef T3 type; }; + +template <typename T1, typename T2, typename T3> +struct packet_conditional<PacketFull, T1, T2, T3> { typedef T1 type; }; + +template <typename T1, typename T2, typename T3> +struct packet_conditional<PacketHalf, T1, T2, T3> { typedef T2 type; }; + +#define PACKET_DECL_COND_PREFIX(prefix, name, packet_size) \ + typedef typename packet_conditional<packet_size, \ + typename packet_traits<name ## Scalar>::type, \ + typename packet_traits<name ## Scalar>::half, \ + typename unpacket_traits<typename packet_traits<name ## Scalar>::half>::half>::type \ + prefix ## name ## Packet + +#define PACKET_DECL_COND(name, packet_size) \ + typedef typename packet_conditional<packet_size, \ + typename packet_traits<name ## Scalar>::type, \ + typename packet_traits<name ## Scalar>::half, \ + typename unpacket_traits<typename packet_traits<name ## Scalar>::half>::half>::type \ + name ## Packet + +#define PACKET_DECL_COND_SCALAR_PREFIX(prefix, packet_size) \ + typedef typename packet_conditional<packet_size, \ + typename packet_traits<Scalar>::type, \ + typename packet_traits<Scalar>::half, \ + typename unpacket_traits<typename packet_traits<Scalar>::half>::half>::type \ + prefix ## ScalarPacket + +#define PACKET_DECL_COND_SCALAR(packet_size) \ + typedef typename packet_conditional<packet_size, \ + typename packet_traits<Scalar>::type, \ + typename packet_traits<Scalar>::half, \ + typename unpacket_traits<typename packet_traits<Scalar>::half>::half>::type \ + ScalarPacket + /* Vectorization logic * real*real: unpack rhs to constant packets, ... * @@ -347,7 +418,7 @@ inline void computeProductBlockingSizes(Index& k, Index& m, Index& n, Index num_ * cplx*real : unpack rhs to constant packets, ... * real*cplx : load lhs as (a0,a0,a1,a1), and mul as usual */ -template<typename _LhsScalar, typename _RhsScalar, bool _ConjLhs, bool _ConjRhs> +template<typename _LhsScalar, typename _RhsScalar, bool _ConjLhs, bool _ConjRhs, int Arch, int _PacketSize> class gebp_traits { public: @@ -355,13 +426,17 @@ public: typedef _RhsScalar RhsScalar; typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar; + PACKET_DECL_COND_PREFIX(_, Lhs, _PacketSize); + PACKET_DECL_COND_PREFIX(_, Rhs, _PacketSize); + PACKET_DECL_COND_PREFIX(_, Res, _PacketSize); + enum { ConjLhs = _ConjLhs, ConjRhs = _ConjRhs, - Vectorizable = packet_traits<LhsScalar>::Vectorizable && packet_traits<RhsScalar>::Vectorizable, - LhsPacketSize = Vectorizable ? packet_traits<LhsScalar>::size : 1, - RhsPacketSize = Vectorizable ? packet_traits<RhsScalar>::size : 1, - ResPacketSize = Vectorizable ? packet_traits<ResScalar>::size : 1, + Vectorizable = unpacket_traits<_LhsPacket>::vectorizable && unpacket_traits<_RhsPacket>::vectorizable, + LhsPacketSize = Vectorizable ? unpacket_traits<_LhsPacket>::size : 1, + RhsPacketSize = Vectorizable ? unpacket_traits<_RhsPacket>::size : 1, + ResPacketSize = Vectorizable ? unpacket_traits<_ResPacket>::size : 1, NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS, @@ -370,10 +445,12 @@ public: // register block size along the M direction (currently, this one cannot be modified) default_mr = (EIGEN_PLAIN_ENUM_MIN(16,NumberOfRegisters)/2/nr)*LhsPacketSize, -#if defined(EIGEN_HAS_SINGLE_INSTRUCTION_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) && !defined(EIGEN_VECTORIZE_VSX) - // we assume 16 registers +#if defined(EIGEN_HAS_SINGLE_INSTRUCTION_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) && !defined(EIGEN_VECTORIZE_VSX) \ + && ((!EIGEN_COMP_MSVC) || (EIGEN_COMP_MSVC>=1914)) + // we assume 16 registers or more // See bug 992, if the scalar type is not vectorizable but that EIGEN_HAS_SINGLE_INSTRUCTION_MADD is defined, // then using 3*LhsPacketSize triggers non-implemented paths in syrk. + // Bug 1515: MSVC prior to v19.14 yields to register spilling. mr = Vectorizable ? 3*LhsPacketSize : default_mr, #else mr = default_mr, @@ -383,38 +460,41 @@ public: RhsProgress = 1 }; - typedef typename packet_traits<LhsScalar>::type _LhsPacket; - typedef typename packet_traits<RhsScalar>::type _RhsPacket; - typedef typename packet_traits<ResScalar>::type _ResPacket; typedef typename conditional<Vectorizable,_LhsPacket,LhsScalar>::type LhsPacket; typedef typename conditional<Vectorizable,_RhsPacket,RhsScalar>::type RhsPacket; typedef typename conditional<Vectorizable,_ResPacket,ResScalar>::type ResPacket; typedef LhsPacket LhsPacket4Packing; + typedef QuadPacket<RhsPacket> RhsPacketx4; typedef ResPacket AccPacket; EIGEN_STRONG_INLINE void initAcc(AccPacket& p) { p = pset1<ResPacket>(ResScalar(0)); } - - EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3) - { - pbroadcast4(b, b0, b1, b2, b3); - } - -// EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1) -// { -// pbroadcast2(b, b0, b1); -// } - + template<typename RhsPacketType> EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacketType& dest) const { dest = pset1<RhsPacketType>(*b); } - + + EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacketx4& dest) const + { + pbroadcast4(b, dest.B_0, dest.B1, dest.B2, dest.B3); + } + + template<typename RhsPacketType> + EIGEN_STRONG_INLINE void updateRhs(const RhsScalar* b, RhsPacketType& dest) const + { + loadRhs(b, dest); + } + + EIGEN_STRONG_INLINE void updateRhs(const RhsScalar*, RhsPacketx4&) const + { + } + EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, RhsPacket& dest) const { dest = ploadquad<RhsPacket>(b); @@ -432,8 +512,8 @@ public: dest = ploadu<LhsPacketType>(a); } - template<typename LhsPacketType, typename RhsPacketType, typename AccPacketType> - EIGEN_STRONG_INLINE void madd(const LhsPacketType& a, const RhsPacketType& b, AccPacketType& c, AccPacketType& tmp) const + template<typename LhsPacketType, typename RhsPacketType, typename AccPacketType, typename LaneIdType> + EIGEN_STRONG_INLINE void madd(const LhsPacketType& a, const RhsPacketType& b, AccPacketType& c, RhsPacketType& tmp, const LaneIdType&) const { conj_helper<LhsPacketType,RhsPacketType,ConjLhs,ConjRhs> cj; // It would be a lot cleaner to call pmadd all the time. Unfortunately if we @@ -448,6 +528,12 @@ public: #endif } + template<typename LhsPacketType, typename AccPacketType, typename LaneIdType> + EIGEN_STRONG_INLINE void madd(const LhsPacketType& a, const RhsPacketx4& b, AccPacketType& c, RhsPacket& tmp, const LaneIdType& lane) const + { + madd(a, b.get(lane), c, tmp, lane); + } + EIGEN_STRONG_INLINE void acc(const AccPacket& c, const ResPacket& alpha, ResPacket& r) const { r = pmadd(c,alpha,r); @@ -461,21 +547,25 @@ public: }; -template<typename RealScalar, bool _ConjLhs> -class gebp_traits<std::complex<RealScalar>, RealScalar, _ConjLhs, false> +template<typename RealScalar, bool _ConjLhs, int Arch, int _PacketSize> +class gebp_traits<std::complex<RealScalar>, RealScalar, _ConjLhs, false, Arch, _PacketSize> { public: typedef std::complex<RealScalar> LhsScalar; typedef RealScalar RhsScalar; typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar; + PACKET_DECL_COND_PREFIX(_, Lhs, _PacketSize); + PACKET_DECL_COND_PREFIX(_, Rhs, _PacketSize); + PACKET_DECL_COND_PREFIX(_, Res, _PacketSize); + enum { ConjLhs = _ConjLhs, ConjRhs = false, - Vectorizable = packet_traits<LhsScalar>::Vectorizable && packet_traits<RhsScalar>::Vectorizable, - LhsPacketSize = Vectorizable ? packet_traits<LhsScalar>::size : 1, - RhsPacketSize = Vectorizable ? packet_traits<RhsScalar>::size : 1, - ResPacketSize = Vectorizable ? packet_traits<ResScalar>::size : 1, + Vectorizable = unpacket_traits<_LhsPacket>::vectorizable && unpacket_traits<_RhsPacket>::vectorizable, + LhsPacketSize = Vectorizable ? unpacket_traits<_LhsPacket>::size : 1, + RhsPacketSize = Vectorizable ? unpacket_traits<_RhsPacket>::size : 1, + ResPacketSize = Vectorizable ? unpacket_traits<_ResPacket>::size : 1, NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS, nr = 4, @@ -490,15 +580,13 @@ public: RhsProgress = 1 }; - typedef typename packet_traits<LhsScalar>::type _LhsPacket; - typedef typename packet_traits<RhsScalar>::type _RhsPacket; - typedef typename packet_traits<ResScalar>::type _ResPacket; - typedef typename conditional<Vectorizable,_LhsPacket,LhsScalar>::type LhsPacket; typedef typename conditional<Vectorizable,_RhsPacket,RhsScalar>::type RhsPacket; typedef typename conditional<Vectorizable,_ResPacket,ResScalar>::type ResPacket; typedef LhsPacket LhsPacket4Packing; + typedef QuadPacket<RhsPacket> RhsPacketx4; + typedef ResPacket AccPacket; EIGEN_STRONG_INLINE void initAcc(AccPacket& p) @@ -506,42 +594,64 @@ public: p = pset1<ResPacket>(ResScalar(0)); } - EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacket& dest) const + template<typename RhsPacketType> + EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacketType& dest) const { - dest = pset1<RhsPacket>(*b); + dest = pset1<RhsPacketType>(*b); + } + + EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacketx4& dest) const + { + pbroadcast4(b, dest.B_0, dest.B1, dest.B2, dest.B3); + } + + template<typename RhsPacketType> + EIGEN_STRONG_INLINE void updateRhs(const RhsScalar* b, RhsPacketType& dest) const + { + loadRhs(b, dest); } + + EIGEN_STRONG_INLINE void updateRhs(const RhsScalar*, RhsPacketx4&) const + {} EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, RhsPacket& dest) const { - dest = pset1<RhsPacket>(*b); + loadRhsQuad_impl(b,dest, typename conditional<RhsPacketSize==16,true_type,false_type>::type()); } - EIGEN_STRONG_INLINE void loadLhs(const LhsScalar* a, LhsPacket& dest) const + EIGEN_STRONG_INLINE void loadRhsQuad_impl(const RhsScalar* b, RhsPacket& dest, const true_type&) const { - dest = pload<LhsPacket>(a); + // FIXME we can do better! + // what we want here is a ploadheight + RhsScalar tmp[4] = {b[0],b[0],b[1],b[1]}; + dest = ploadquad<RhsPacket>(tmp); } - EIGEN_STRONG_INLINE void loadLhsUnaligned(const LhsScalar* a, LhsPacket& dest) const + EIGEN_STRONG_INLINE void loadRhsQuad_impl(const RhsScalar* b, RhsPacket& dest, const false_type&) const { - dest = ploadu<LhsPacket>(a); + eigen_internal_assert(RhsPacketSize<=8); + dest = pset1<RhsPacket>(*b); } - EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3) + EIGEN_STRONG_INLINE void loadLhs(const LhsScalar* a, LhsPacket& dest) const { - pbroadcast4(b, b0, b1, b2, b3); + dest = pload<LhsPacket>(a); } - -// EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1) -// { -// pbroadcast2(b, b0, b1); -// } - EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp) const + template<typename LhsPacketType> + EIGEN_STRONG_INLINE void loadLhsUnaligned(const LhsScalar* a, LhsPacketType& dest) const + { + dest = ploadu<LhsPacketType>(a); + } + + template <typename LhsPacketType, typename RhsPacketType, typename AccPacketType, typename LaneIdType> + EIGEN_STRONG_INLINE void madd(const LhsPacketType& a, const RhsPacketType& b, AccPacketType& c, RhsPacketType& tmp, const LaneIdType&) const { madd_impl(a, b, c, tmp, typename conditional<Vectorizable,true_type,false_type>::type()); } - EIGEN_STRONG_INLINE void madd_impl(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp, const true_type&) const + template <typename LhsPacketType, typename RhsPacketType, typename AccPacketType> + EIGEN_STRONG_INLINE void madd_impl(const LhsPacketType& a, const RhsPacketType& b, AccPacketType& c, RhsPacketType& tmp, const true_type&) const { #ifdef EIGEN_HAS_SINGLE_INSTRUCTION_MADD EIGEN_UNUSED_VARIABLE(tmp); @@ -556,13 +666,20 @@ public: c += a * b; } - EIGEN_STRONG_INLINE void acc(const AccPacket& c, const ResPacket& alpha, ResPacket& r) const + template<typename LhsPacketType, typename AccPacketType, typename LaneIdType> + EIGEN_STRONG_INLINE void madd(const LhsPacketType& a, const RhsPacketx4& b, AccPacketType& c, RhsPacket& tmp, const LaneIdType& lane) const + { + madd(a, b.get(lane), c, tmp, lane); + } + + template <typename ResPacketType, typename AccPacketType> + EIGEN_STRONG_INLINE void acc(const AccPacketType& c, const ResPacketType& alpha, ResPacketType& r) const { + conj_helper<ResPacketType,ResPacketType,ConjLhs,false> cj; r = cj.pmadd(c,alpha,r); } protected: - conj_helper<ResPacket,ResPacket,ConjLhs,false> cj; }; template<typename Packet> @@ -581,13 +698,57 @@ DoublePacket<Packet> padd(const DoublePacket<Packet> &a, const DoublePacket<Pack return res; } +// note that for DoublePacket<RealPacket> the "4" in "downto4" +// corresponds to the number of complexes, so it means "8" +// it terms of real coefficients. + template<typename Packet> -const DoublePacket<Packet>& predux_half_dowto4(const DoublePacket<Packet> &a) +const DoublePacket<Packet>& +predux_half_dowto4(const DoublePacket<Packet> &a, + typename enable_if<unpacket_traits<Packet>::size<=8>::type* = 0) { return a; } -template<typename Packet> struct unpacket_traits<DoublePacket<Packet> > { typedef DoublePacket<Packet> half; }; +template<typename Packet> +DoublePacket<typename unpacket_traits<Packet>::half> +predux_half_dowto4(const DoublePacket<Packet> &a, + typename enable_if<unpacket_traits<Packet>::size==16>::type* = 0) +{ + // yes, that's pretty hackish :( + DoublePacket<typename unpacket_traits<Packet>::half> res; + typedef std::complex<typename unpacket_traits<Packet>::type> Cplx; + typedef typename packet_traits<Cplx>::type CplxPacket; + res.first = predux_half_dowto4(CplxPacket(a.first)).v; + res.second = predux_half_dowto4(CplxPacket(a.second)).v; + return res; +} + +// same here, "quad" actually means "8" in terms of real coefficients +template<typename Scalar, typename RealPacket> +void loadQuadToDoublePacket(const Scalar* b, DoublePacket<RealPacket>& dest, + typename enable_if<unpacket_traits<RealPacket>::size<=8>::type* = 0) +{ + dest.first = pset1<RealPacket>(real(*b)); + dest.second = pset1<RealPacket>(imag(*b)); +} + +template<typename Scalar, typename RealPacket> +void loadQuadToDoublePacket(const Scalar* b, DoublePacket<RealPacket>& dest, + typename enable_if<unpacket_traits<RealPacket>::size==16>::type* = 0) +{ + // yes, that's pretty hackish too :( + typedef typename NumTraits<Scalar>::Real RealScalar; + RealScalar r[4] = {real(b[0]), real(b[0]), real(b[1]), real(b[1])}; + RealScalar i[4] = {imag(b[0]), imag(b[0]), imag(b[1]), imag(b[1])}; + dest.first = ploadquad<RealPacket>(r); + dest.second = ploadquad<RealPacket>(i); +} + + +template<typename Packet> struct unpacket_traits<DoublePacket<Packet> > { + typedef DoublePacket<typename unpacket_traits<Packet>::half> half; +}; // template<typename Packet> // DoublePacket<Packet> pmadd(const DoublePacket<Packet> &a, const DoublePacket<Packet> &b) // { @@ -597,8 +758,8 @@ template<typename Packet> struct unpacket_traits<DoublePacket<Packet> > { typede // return res; // } -template<typename RealScalar, bool _ConjLhs, bool _ConjRhs> -class gebp_traits<std::complex<RealScalar>, std::complex<RealScalar>, _ConjLhs, _ConjRhs > +template<typename RealScalar, bool _ConjLhs, bool _ConjRhs, int Arch, int _PacketSize> +class gebp_traits<std::complex<RealScalar>, std::complex<RealScalar>, _ConjLhs, _ConjRhs, Arch, _PacketSize > { public: typedef std::complex<RealScalar> Scalar; @@ -606,15 +767,21 @@ public: typedef std::complex<RealScalar> RhsScalar; typedef std::complex<RealScalar> ResScalar; + PACKET_DECL_COND_PREFIX(_, Lhs, _PacketSize); + PACKET_DECL_COND_PREFIX(_, Rhs, _PacketSize); + PACKET_DECL_COND_PREFIX(_, Res, _PacketSize); + PACKET_DECL_COND(Real, _PacketSize); + PACKET_DECL_COND_SCALAR(_PacketSize); + enum { ConjLhs = _ConjLhs, ConjRhs = _ConjRhs, - Vectorizable = packet_traits<RealScalar>::Vectorizable - && packet_traits<Scalar>::Vectorizable, - RealPacketSize = Vectorizable ? packet_traits<RealScalar>::size : 1, - ResPacketSize = Vectorizable ? packet_traits<ResScalar>::size : 1, - LhsPacketSize = Vectorizable ? packet_traits<LhsScalar>::size : 1, - RhsPacketSize = Vectorizable ? packet_traits<RhsScalar>::size : 1, + Vectorizable = unpacket_traits<RealPacket>::vectorizable + && unpacket_traits<ScalarPacket>::vectorizable, + ResPacketSize = Vectorizable ? unpacket_traits<_ResPacket>::size : 1, + LhsPacketSize = Vectorizable ? unpacket_traits<_LhsPacket>::size : 1, + RhsPacketSize = Vectorizable ? unpacket_traits<RhsScalar>::size : 1, + RealPacketSize = Vectorizable ? unpacket_traits<RealPacket>::size : 1, // FIXME: should depend on NumberOfRegisters nr = 4, @@ -624,15 +791,16 @@ public: RhsProgress = 1 }; - typedef typename packet_traits<RealScalar>::type RealPacket; - typedef typename packet_traits<Scalar>::type ScalarPacket; - typedef DoublePacket<RealPacket> DoublePacketType; + typedef DoublePacket<RealPacket> DoublePacketType; typedef typename conditional<Vectorizable,ScalarPacket,Scalar>::type LhsPacket4Packing; typedef typename conditional<Vectorizable,RealPacket, Scalar>::type LhsPacket; typedef typename conditional<Vectorizable,DoublePacketType,Scalar>::type RhsPacket; typedef typename conditional<Vectorizable,ScalarPacket,Scalar>::type ResPacket; typedef typename conditional<Vectorizable,DoublePacketType,Scalar>::type AccPacket; + + // this actualy holds 8 packets! + typedef QuadPacket<RhsPacket> RhsPacketx4; EIGEN_STRONG_INLINE void initAcc(Scalar& p) { p = Scalar(0); } @@ -643,51 +811,49 @@ public: } // Scalar path - EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, ResPacket& dest) const + EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, ScalarPacket& dest) const { - dest = pset1<ResPacket>(*b); + dest = pset1<ScalarPacket>(*b); } // Vectorized path - EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, DoublePacketType& dest) const + template<typename RealPacketType> + EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, DoublePacket<RealPacketType>& dest) const { - dest.first = pset1<RealPacket>(real(*b)); - dest.second = pset1<RealPacket>(imag(*b)); + dest.first = pset1<RealPacketType>(real(*b)); + dest.second = pset1<RealPacketType>(imag(*b)); } - - EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, ResPacket& dest) const + + EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacketx4& dest) const { - loadRhs(b,dest); + loadRhs(b, dest.B_0); + loadRhs(b + 1, dest.B1); + loadRhs(b + 2, dest.B2); + loadRhs(b + 3, dest.B3); } - EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, DoublePacketType& dest) const + + // Scalar path + EIGEN_STRONG_INLINE void updateRhs(const RhsScalar* b, ScalarPacket& dest) const { - eigen_internal_assert(unpacket_traits<ScalarPacket>::size<=4); - loadRhs(b,dest); + loadRhs(b, dest); } - - EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3) + + // Vectorized path + template<typename RealPacketType> + EIGEN_STRONG_INLINE void updateRhs(const RhsScalar* b, DoublePacket<RealPacketType>& dest) const { - // FIXME not sure that's the best way to implement it! - loadRhs(b+0, b0); - loadRhs(b+1, b1); - loadRhs(b+2, b2); - loadRhs(b+3, b3); + loadRhs(b, dest); } + + EIGEN_STRONG_INLINE void updateRhs(const RhsScalar*, RhsPacketx4&) const {} - // Vectorized path - EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, DoublePacketType& b0, DoublePacketType& b1) + EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, ResPacket& dest) const { - // FIXME not sure that's the best way to implement it! - loadRhs(b+0, b0); - loadRhs(b+1, b1); + loadRhs(b,dest); } - - // Scalar path - EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsScalar& b0, RhsScalar& b1) + EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, DoublePacketType& dest) const { - // FIXME not sure that's the best way to implement it! - loadRhs(b+0, b0); - loadRhs(b+1, b1); + loadQuadToDoublePacket(b,dest); } // nothing special here @@ -696,47 +862,59 @@ public: dest = pload<LhsPacket>((const typename unpacket_traits<LhsPacket>::type*)(a)); } - EIGEN_STRONG_INLINE void loadLhsUnaligned(const LhsScalar* a, LhsPacket& dest) const + template<typename LhsPacketType> + EIGEN_STRONG_INLINE void loadLhsUnaligned(const LhsScalar* a, LhsPacketType& dest) const { - dest = ploadu<LhsPacket>((const typename unpacket_traits<LhsPacket>::type*)(a)); + dest = ploadu<LhsPacketType>((const typename unpacket_traits<LhsPacketType>::type*)(a)); } - EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, DoublePacketType& c, RhsPacket& /*tmp*/) const + template<typename LhsPacketType, typename RhsPacketType, typename ResPacketType, typename TmpType, typename LaneIdType> + EIGEN_STRONG_INLINE + typename enable_if<!is_same<RhsPacketType,RhsPacketx4>::value>::type + madd(const LhsPacketType& a, const RhsPacketType& b, DoublePacket<ResPacketType>& c, TmpType& /*tmp*/, const LaneIdType&) const { c.first = padd(pmul(a,b.first), c.first); c.second = padd(pmul(a,b.second),c.second); } - EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, ResPacket& c, RhsPacket& /*tmp*/) const + template<typename LaneIdType> + EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, ResPacket& c, RhsPacket& /*tmp*/, const LaneIdType&) const { c = cj.pmadd(a,b,c); } + + template<typename LhsPacketType, typename AccPacketType, typename LaneIdType> + EIGEN_STRONG_INLINE void madd(const LhsPacketType& a, const RhsPacketx4& b, AccPacketType& c, RhsPacket& tmp, const LaneIdType& lane) const + { + madd(a, b.get(lane), c, tmp, lane); + } EIGEN_STRONG_INLINE void acc(const Scalar& c, const Scalar& alpha, Scalar& r) const { r += alpha * c; } - EIGEN_STRONG_INLINE void acc(const DoublePacketType& c, const ResPacket& alpha, ResPacket& r) const + template<typename RealPacketType, typename ResPacketType> + EIGEN_STRONG_INLINE void acc(const DoublePacket<RealPacketType>& c, const ResPacketType& alpha, ResPacketType& r) const { // assemble c - ResPacket tmp; + ResPacketType tmp; if((!ConjLhs)&&(!ConjRhs)) { - tmp = pcplxflip(pconj(ResPacket(c.second))); - tmp = padd(ResPacket(c.first),tmp); + tmp = pcplxflip(pconj(ResPacketType(c.second))); + tmp = padd(ResPacketType(c.first),tmp); } else if((!ConjLhs)&&(ConjRhs)) { - tmp = pconj(pcplxflip(ResPacket(c.second))); - tmp = padd(ResPacket(c.first),tmp); + tmp = pconj(pcplxflip(ResPacketType(c.second))); + tmp = padd(ResPacketType(c.first),tmp); } else if((ConjLhs)&&(!ConjRhs)) { - tmp = pcplxflip(ResPacket(c.second)); - tmp = padd(pconj(ResPacket(c.first)),tmp); + tmp = pcplxflip(ResPacketType(c.second)); + tmp = padd(pconj(ResPacketType(c.first)),tmp); } else if((ConjLhs)&&(ConjRhs)) { - tmp = pcplxflip(ResPacket(c.second)); - tmp = psub(pconj(ResPacket(c.first)),tmp); + tmp = pcplxflip(ResPacketType(c.second)); + tmp = psub(pconj(ResPacketType(c.first)),tmp); } r = pmadd(tmp,alpha,r); @@ -746,8 +924,8 @@ protected: conj_helper<LhsScalar,RhsScalar,ConjLhs,ConjRhs> cj; }; -template<typename RealScalar, bool _ConjRhs> -class gebp_traits<RealScalar, std::complex<RealScalar>, false, _ConjRhs > +template<typename RealScalar, bool _ConjRhs, int Arch, int _PacketSize> +class gebp_traits<RealScalar, std::complex<RealScalar>, false, _ConjRhs, Arch, _PacketSize > { public: typedef std::complex<RealScalar> Scalar; @@ -755,14 +933,25 @@ public: typedef Scalar RhsScalar; typedef Scalar ResScalar; + PACKET_DECL_COND_PREFIX(_, Lhs, _PacketSize); + PACKET_DECL_COND_PREFIX(_, Rhs, _PacketSize); + PACKET_DECL_COND_PREFIX(_, Res, _PacketSize); + PACKET_DECL_COND_PREFIX(_, Real, _PacketSize); + PACKET_DECL_COND_SCALAR_PREFIX(_, _PacketSize); + +#undef PACKET_DECL_COND_SCALAR_PREFIX +#undef PACKET_DECL_COND_PREFIX +#undef PACKET_DECL_COND_SCALAR +#undef PACKET_DECL_COND + enum { ConjLhs = false, ConjRhs = _ConjRhs, - Vectorizable = packet_traits<RealScalar>::Vectorizable - && packet_traits<Scalar>::Vectorizable, - LhsPacketSize = Vectorizable ? packet_traits<LhsScalar>::size : 1, - RhsPacketSize = Vectorizable ? packet_traits<RhsScalar>::size : 1, - ResPacketSize = Vectorizable ? packet_traits<ResScalar>::size : 1, + Vectorizable = unpacket_traits<_RealPacket>::vectorizable + && unpacket_traits<_ScalarPacket>::vectorizable, + LhsPacketSize = Vectorizable ? unpacket_traits<_LhsPacket>::size : 1, + RhsPacketSize = Vectorizable ? unpacket_traits<_RhsPacket>::size : 1, + ResPacketSize = Vectorizable ? unpacket_traits<_ResPacket>::size : 1, NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS, // FIXME: should depend on NumberOfRegisters @@ -773,15 +962,11 @@ public: RhsProgress = 1 }; - typedef typename packet_traits<LhsScalar>::type _LhsPacket; - typedef typename packet_traits<RhsScalar>::type _RhsPacket; - typedef typename packet_traits<ResScalar>::type _ResPacket; - typedef typename conditional<Vectorizable,_LhsPacket,LhsScalar>::type LhsPacket; typedef typename conditional<Vectorizable,_RhsPacket,RhsScalar>::type RhsPacket; typedef typename conditional<Vectorizable,_ResPacket,ResScalar>::type ResPacket; typedef LhsPacket LhsPacket4Packing; - + typedef QuadPacket<RhsPacket> RhsPacketx4; typedef ResPacket AccPacket; EIGEN_STRONG_INLINE void initAcc(AccPacket& p) @@ -789,22 +974,25 @@ public: p = pset1<ResPacket>(ResScalar(0)); } - EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacket& dest) const + template<typename RhsPacketType> + EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacketType& dest) const { - dest = pset1<RhsPacket>(*b); + dest = pset1<RhsPacketType>(*b); } - - void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3) + + EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacketx4& dest) const { - pbroadcast4(b, b0, b1, b2, b3); + pbroadcast4(b, dest.B_0, dest.B1, dest.B2, dest.B3); } - -// EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1) -// { -// // FIXME not sure that's the best way to implement it! -// b0 = pload1<RhsPacket>(b+0); -// b1 = pload1<RhsPacket>(b+1); -// } + + template<typename RhsPacketType> + EIGEN_STRONG_INLINE void updateRhs(const RhsScalar* b, RhsPacketType& dest) const + { + loadRhs(b, dest); + } + + EIGEN_STRONG_INLINE void updateRhs(const RhsScalar*, RhsPacketx4&) const + {} EIGEN_STRONG_INLINE void loadLhs(const LhsScalar* a, LhsPacket& dest) const { @@ -813,21 +1001,23 @@ public: EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, RhsPacket& dest) const { - eigen_internal_assert(unpacket_traits<RhsPacket>::size<=4); - loadRhs(b,dest); + dest = ploadquad<RhsPacket>(b); } - EIGEN_STRONG_INLINE void loadLhsUnaligned(const LhsScalar* a, LhsPacket& dest) const + template<typename LhsPacketType> + EIGEN_STRONG_INLINE void loadLhsUnaligned(const LhsScalar* a, LhsPacketType& dest) const { - dest = ploaddup<LhsPacket>(a); + dest = ploaddup<LhsPacketType>(a); } - EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp) const + template <typename LhsPacketType, typename RhsPacketType, typename AccPacketType, typename LaneIdType> + EIGEN_STRONG_INLINE void madd(const LhsPacketType& a, const RhsPacketType& b, AccPacketType& c, RhsPacketType& tmp, const LaneIdType&) const { madd_impl(a, b, c, tmp, typename conditional<Vectorizable,true_type,false_type>::type()); } - EIGEN_STRONG_INLINE void madd_impl(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp, const true_type&) const + template <typename LhsPacketType, typename RhsPacketType, typename AccPacketType> + EIGEN_STRONG_INLINE void madd_impl(const LhsPacketType& a, const RhsPacketType& b, AccPacketType& c, RhsPacketType& tmp, const true_type&) const { #ifdef EIGEN_HAS_SINGLE_INSTRUCTION_MADD EIGEN_UNUSED_VARIABLE(tmp); @@ -843,16 +1033,166 @@ public: c += a * b; } - EIGEN_STRONG_INLINE void acc(const AccPacket& c, const ResPacket& alpha, ResPacket& r) const + template<typename LhsPacketType, typename AccPacketType, typename LaneIdType> + EIGEN_STRONG_INLINE void madd(const LhsPacketType& a, const RhsPacketx4& b, AccPacketType& c, RhsPacket& tmp, const LaneIdType& lane) const + { + madd(a, b.get(lane), c, tmp, lane); + } + + template <typename ResPacketType, typename AccPacketType> + EIGEN_STRONG_INLINE void acc(const AccPacketType& c, const ResPacketType& alpha, ResPacketType& r) const { + conj_helper<ResPacketType,ResPacketType,false,ConjRhs> cj; r = cj.pmadd(alpha,c,r); } protected: - conj_helper<ResPacket,ResPacket,false,ConjRhs> cj; + +}; + + +#if EIGEN_ARCH_ARM64 && defined EIGEN_VECTORIZE_NEON + +template<> +struct gebp_traits <float, float, false, false,Architecture::NEON,PacketFull> + : gebp_traits<float,float,false,false,Architecture::Generic,PacketFull> +{ + typedef float RhsPacket; + + typedef float32x4_t RhsPacketx4; + + EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacket& dest) const + { + dest = *b; + } + + EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacketx4& dest) const + { + dest = vld1q_f32(b); + } + + EIGEN_STRONG_INLINE void updateRhs(const RhsScalar* b, RhsPacket& dest) const + { + dest = *b; + } + + EIGEN_STRONG_INLINE void updateRhs(const RhsScalar* b, RhsPacketx4& dest) const + {} + + EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, RhsPacket& dest) const + { + loadRhs(b,dest); + } + + EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& /*tmp*/, const FixedInt<0>&) const + { + c = vfmaq_n_f32(c, a, b); + } + + // NOTE: Template parameter inference failed when compiled with Android NDK: + // "candidate template ignored: could not match 'FixedInt<N>' against 'Eigen::internal::FixedInt<0>". + + EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacketx4& b, AccPacket& c, RhsPacket& /*tmp*/, const FixedInt<0>&) const + { madd_helper<0>(a, b, c); } + EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacketx4& b, AccPacket& c, RhsPacket& /*tmp*/, const FixedInt<1>&) const + { madd_helper<1>(a, b, c); } + EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacketx4& b, AccPacket& c, RhsPacket& /*tmp*/, const FixedInt<2>&) const + { madd_helper<2>(a, b, c); } + EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacketx4& b, AccPacket& c, RhsPacket& /*tmp*/, const FixedInt<3>&) const + { madd_helper<3>(a, b, c); } + + private: + template<int LaneID> + EIGEN_STRONG_INLINE void madd_helper(const LhsPacket& a, const RhsPacketx4& b, AccPacket& c) const + { + #if EIGEN_COMP_GNUC_STRICT && !(EIGEN_GNUC_AT_LEAST(9,0)) + // workaround gcc issue https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89101 + // vfmaq_laneq_f32 is implemented through a costly dup + if(LaneID==0) asm("fmla %0.4s, %1.4s, %2.s[0]\n" : "+w" (c) : "w" (a), "w" (b) : ); + else if(LaneID==1) asm("fmla %0.4s, %1.4s, %2.s[1]\n" : "+w" (c) : "w" (a), "w" (b) : ); + else if(LaneID==2) asm("fmla %0.4s, %1.4s, %2.s[2]\n" : "+w" (c) : "w" (a), "w" (b) : ); + else if(LaneID==3) asm("fmla %0.4s, %1.4s, %2.s[3]\n" : "+w" (c) : "w" (a), "w" (b) : ); + #else + c = vfmaq_laneq_f32(c, a, b, LaneID); + #endif + } +}; + + +template<> +struct gebp_traits <double, double, false, false,Architecture::NEON> + : gebp_traits<double,double,false,false,Architecture::Generic> +{ + typedef double RhsPacket; + + struct RhsPacketx4 { + float64x2_t B_0, B_1; + }; + + EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacket& dest) const + { + dest = *b; + } + + EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacketx4& dest) const + { + dest.B_0 = vld1q_f64(b); + dest.B_1 = vld1q_f64(b+2); + } + + EIGEN_STRONG_INLINE void updateRhs(const RhsScalar* b, RhsPacket& dest) const + { + loadRhs(b,dest); + } + + EIGEN_STRONG_INLINE void updateRhs(const RhsScalar* b, RhsPacketx4& dest) const + {} + + EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, RhsPacket& dest) const + { + loadRhs(b,dest); + } + + EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& /*tmp*/, const FixedInt<0>&) const + { + c = vfmaq_n_f64(c, a, b); + } + + // NOTE: Template parameter inference failed when compiled with Android NDK: + // "candidate template ignored: could not match 'FixedInt<N>' against 'Eigen::internal::FixedInt<0>". + + EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacketx4& b, AccPacket& c, RhsPacket& /*tmp*/, const FixedInt<0>&) const + { madd_helper<0>(a, b, c); } + EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacketx4& b, AccPacket& c, RhsPacket& /*tmp*/, const FixedInt<1>&) const + { madd_helper<1>(a, b, c); } + EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacketx4& b, AccPacket& c, RhsPacket& /*tmp*/, const FixedInt<2>&) const + { madd_helper<2>(a, b, c); } + EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacketx4& b, AccPacket& c, RhsPacket& /*tmp*/, const FixedInt<3>&) const + { madd_helper<3>(a, b, c); } + + private: + template <int LaneID> + EIGEN_STRONG_INLINE void madd_helper(const LhsPacket& a, const RhsPacketx4& b, AccPacket& c) const + { + #if EIGEN_COMP_GNUC_STRICT && !(EIGEN_GNUC_AT_LEAST(9,0)) + // workaround gcc issue https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89101 + // vfmaq_laneq_f64 is implemented through a costly dup + if(LaneID==0) asm("fmla %0.2d, %1.2d, %2.d[0]\n" : "+w" (c) : "w" (a), "w" (b.B_0) : ); + else if(LaneID==1) asm("fmla %0.2d, %1.2d, %2.d[1]\n" : "+w" (c) : "w" (a), "w" (b.B_0) : ); + else if(LaneID==2) asm("fmla %0.2d, %1.2d, %2.d[0]\n" : "+w" (c) : "w" (a), "w" (b.B_1) : ); + else if(LaneID==3) asm("fmla %0.2d, %1.2d, %2.d[1]\n" : "+w" (c) : "w" (a), "w" (b.B_1) : ); + #else + if(LaneID==0) c = vfmaq_laneq_f64(c, a, b.B_0, 0); + else if(LaneID==1) c = vfmaq_laneq_f64(c, a, b.B_0, 1); + else if(LaneID==2) c = vfmaq_laneq_f64(c, a, b.B_1, 0); + else if(LaneID==3) c = vfmaq_laneq_f64(c, a, b.B_1, 1); + #endif + } }; -/* optimized GEneral packed Block * packed Panel product kernel +#endif + +/* optimized General packed Block * packed Panel product kernel * * Mixing type logic: C += A * B * | A | B | comments @@ -862,26 +1202,47 @@ protected: template<typename LhsScalar, typename RhsScalar, typename Index, typename DataMapper, int mr, int nr, bool ConjugateLhs, bool ConjugateRhs> struct gebp_kernel { - typedef gebp_traits<LhsScalar,RhsScalar,ConjugateLhs,ConjugateRhs> Traits; + typedef gebp_traits<LhsScalar,RhsScalar,ConjugateLhs,ConjugateRhs,Architecture::Target> Traits; + typedef gebp_traits<LhsScalar,RhsScalar,ConjugateLhs,ConjugateRhs,Architecture::Target,PacketHalf> HalfTraits; + typedef gebp_traits<LhsScalar,RhsScalar,ConjugateLhs,ConjugateRhs,Architecture::Target,PacketQuarter> QuarterTraits; + typedef typename Traits::ResScalar ResScalar; typedef typename Traits::LhsPacket LhsPacket; typedef typename Traits::RhsPacket RhsPacket; typedef typename Traits::ResPacket ResPacket; typedef typename Traits::AccPacket AccPacket; + typedef typename Traits::RhsPacketx4 RhsPacketx4; + + typedef typename RhsPanelHelper<RhsPacket, RhsPacketx4, 15>::type RhsPanel15; + + typedef gebp_traits<RhsScalar,LhsScalar,ConjugateRhs,ConjugateLhs,Architecture::Target> SwappedTraits; - typedef gebp_traits<RhsScalar,LhsScalar,ConjugateRhs,ConjugateLhs> SwappedTraits; typedef typename SwappedTraits::ResScalar SResScalar; typedef typename SwappedTraits::LhsPacket SLhsPacket; typedef typename SwappedTraits::RhsPacket SRhsPacket; typedef typename SwappedTraits::ResPacket SResPacket; typedef typename SwappedTraits::AccPacket SAccPacket; + typedef typename HalfTraits::LhsPacket LhsPacketHalf; + typedef typename HalfTraits::RhsPacket RhsPacketHalf; + typedef typename HalfTraits::ResPacket ResPacketHalf; + typedef typename HalfTraits::AccPacket AccPacketHalf; + + typedef typename QuarterTraits::LhsPacket LhsPacketQuarter; + typedef typename QuarterTraits::RhsPacket RhsPacketQuarter; + typedef typename QuarterTraits::ResPacket ResPacketQuarter; + typedef typename QuarterTraits::AccPacket AccPacketQuarter; + typedef typename DataMapper::LinearMapper LinearMapper; enum { Vectorizable = Traits::Vectorizable, LhsProgress = Traits::LhsProgress, + LhsProgressHalf = HalfTraits::LhsProgress, + LhsProgressQuarter = QuarterTraits::LhsProgress, RhsProgress = Traits::RhsProgress, + RhsProgressHalf = HalfTraits::RhsProgress, + RhsProgressQuarter = QuarterTraits::RhsProgress, ResPacketSize = Traits::ResPacketSize }; @@ -892,11 +1253,11 @@ struct gebp_kernel }; template<typename LhsScalar, typename RhsScalar, typename Index, typename DataMapper, int mr, int nr, bool ConjugateLhs, bool ConjugateRhs, - int SwappedLhsProgress = gebp_traits<RhsScalar,LhsScalar,ConjugateRhs,ConjugateLhs>::LhsProgress> +int SwappedLhsProgress = gebp_traits<RhsScalar,LhsScalar,ConjugateRhs,ConjugateLhs,Architecture::Target>::LhsProgress> struct last_row_process_16_packets { - typedef gebp_traits<LhsScalar,RhsScalar,ConjugateLhs,ConjugateRhs> Traits; - typedef gebp_traits<RhsScalar,LhsScalar,ConjugateRhs,ConjugateLhs> SwappedTraits; + typedef gebp_traits<LhsScalar,RhsScalar,ConjugateLhs,ConjugateRhs,Architecture::Target> Traits; + typedef gebp_traits<RhsScalar,LhsScalar,ConjugateRhs,ConjugateLhs,Architecture::Target> SwappedTraits; typedef typename Traits::ResScalar ResScalar; typedef typename SwappedTraits::LhsPacket SLhsPacket; @@ -924,8 +1285,8 @@ struct last_row_process_16_packets template<typename LhsScalar, typename RhsScalar, typename Index, typename DataMapper, int mr, int nr, bool ConjugateLhs, bool ConjugateRhs> struct last_row_process_16_packets<LhsScalar, RhsScalar, Index, DataMapper, mr, nr, ConjugateLhs, ConjugateRhs, 16> { - typedef gebp_traits<LhsScalar,RhsScalar,ConjugateLhs,ConjugateRhs> Traits; - typedef gebp_traits<RhsScalar,LhsScalar,ConjugateRhs,ConjugateLhs> SwappedTraits; + typedef gebp_traits<LhsScalar,RhsScalar,ConjugateLhs,ConjugateRhs,Architecture::Target> Traits; + typedef gebp_traits<RhsScalar,LhsScalar,ConjugateRhs,ConjugateLhs,Architecture::Target> SwappedTraits; typedef typename Traits::ResScalar ResScalar; typedef typename SwappedTraits::LhsPacket SLhsPacket; @@ -957,7 +1318,7 @@ struct last_row_process_16_packets<LhsScalar, RhsScalar, Index, DataMapper, mr, SRhsPacketQuarter b0; straits.loadLhsUnaligned(blB, a0); straits.loadRhs(blA, b0); - straits.madd(a0,b0,c0,b0); + straits.madd(a0,b0,c0,b0, fix<0>); blB += SwappedTraits::LhsProgress/4; blA += 1; } @@ -971,6 +1332,219 @@ struct last_row_process_16_packets<LhsScalar, RhsScalar, Index, DataMapper, mr, } }; +template<int nr, Index LhsProgress, Index RhsProgress, typename LhsScalar, typename RhsScalar, typename ResScalar, typename AccPacket, typename LhsPacket, typename RhsPacket, typename ResPacket, typename GEBPTraits, typename LinearMapper, typename DataMapper> +struct lhs_process_one_packet +{ + typedef typename GEBPTraits::RhsPacketx4 RhsPacketx4; + + EIGEN_STRONG_INLINE void peeled_kc_onestep(Index K, const LhsScalar* blA, const RhsScalar* blB, GEBPTraits traits, LhsPacket *A0, RhsPacketx4 *rhs_panel, RhsPacket *T0, AccPacket *C0, AccPacket *C1, AccPacket *C2, AccPacket *C3) + { + EIGEN_ASM_COMMENT("begin step of gebp micro kernel 1X4"); + EIGEN_ASM_COMMENT("Note: these asm comments work around bug 935!"); + traits.loadLhs(&blA[(0+1*K)*LhsProgress], *A0); + traits.loadRhs(&blB[(0+4*K)*RhsProgress], *rhs_panel); + traits.madd(*A0, *rhs_panel, *C0, *T0, fix<0>); + traits.madd(*A0, *rhs_panel, *C1, *T0, fix<1>); + traits.madd(*A0, *rhs_panel, *C2, *T0, fix<2>); + traits.madd(*A0, *rhs_panel, *C3, *T0, fix<3>); + #if EIGEN_GNUC_AT_LEAST(6,0) && defined(EIGEN_VECTORIZE_SSE) + __asm__ ("" : "+x,m" (*A0)); + #endif + EIGEN_ASM_COMMENT("end step of gebp micro kernel 1X4"); + } + + EIGEN_STRONG_INLINE void operator()( + const DataMapper& res, const LhsScalar* blockA, const RhsScalar* blockB, ResScalar alpha, + Index peelStart, Index peelEnd, Index strideA, Index strideB, Index offsetA, Index offsetB, + int prefetch_res_offset, Index peeled_kc, Index pk, Index cols, Index depth, Index packet_cols4) + { + GEBPTraits traits; + + // loops on each largest micro horizontal panel of lhs + // (LhsProgress x depth) + for(Index i=peelStart; i<peelEnd; i+=LhsProgress) + { + // loops on each largest micro vertical panel of rhs (depth * nr) + for(Index j2=0; j2<packet_cols4; j2+=nr) + { + // We select a LhsProgress x nr micro block of res + // which is entirely stored into 1 x nr registers. + + const LhsScalar* blA = &blockA[i*strideA+offsetA*(LhsProgress)]; + prefetch(&blA[0]); + + // gets res block as register + AccPacket C0, C1, C2, C3; + traits.initAcc(C0); + traits.initAcc(C1); + traits.initAcc(C2); + traits.initAcc(C3); + // To improve instruction pipelining, let's double the accumulation registers: + // even k will accumulate in C*, while odd k will accumulate in D*. + // This trick is crutial to get good performance with FMA, otherwise it is + // actually faster to perform separated MUL+ADD because of a naturally + // better instruction-level parallelism. + AccPacket D0, D1, D2, D3; + traits.initAcc(D0); + traits.initAcc(D1); + traits.initAcc(D2); + traits.initAcc(D3); + + LinearMapper r0 = res.getLinearMapper(i, j2 + 0); + LinearMapper r1 = res.getLinearMapper(i, j2 + 1); + LinearMapper r2 = res.getLinearMapper(i, j2 + 2); + LinearMapper r3 = res.getLinearMapper(i, j2 + 3); + + r0.prefetch(prefetch_res_offset); + r1.prefetch(prefetch_res_offset); + r2.prefetch(prefetch_res_offset); + r3.prefetch(prefetch_res_offset); + + // performs "inner" products + const RhsScalar* blB = &blockB[j2*strideB+offsetB*nr]; + prefetch(&blB[0]); + LhsPacket A0, A1; + + for(Index k=0; k<peeled_kc; k+=pk) + { + EIGEN_ASM_COMMENT("begin gebp micro kernel 1/half/quarterX4"); + RhsPacketx4 rhs_panel; + RhsPacket T0; + + internal::prefetch(blB+(48+0)); + peeled_kc_onestep(0, blA, blB, traits, &A0, &rhs_panel, &T0, &C0, &C1, &C2, &C3); + peeled_kc_onestep(1, blA, blB, traits, &A1, &rhs_panel, &T0, &D0, &D1, &D2, &D3); + peeled_kc_onestep(2, blA, blB, traits, &A0, &rhs_panel, &T0, &C0, &C1, &C2, &C3); + peeled_kc_onestep(3, blA, blB, traits, &A1, &rhs_panel, &T0, &D0, &D1, &D2, &D3); + internal::prefetch(blB+(48+16)); + peeled_kc_onestep(4, blA, blB, traits, &A0, &rhs_panel, &T0, &C0, &C1, &C2, &C3); + peeled_kc_onestep(5, blA, blB, traits, &A1, &rhs_panel, &T0, &D0, &D1, &D2, &D3); + peeled_kc_onestep(6, blA, blB, traits, &A0, &rhs_panel, &T0, &C0, &C1, &C2, &C3); + peeled_kc_onestep(7, blA, blB, traits, &A1, &rhs_panel, &T0, &D0, &D1, &D2, &D3); + + blB += pk*4*RhsProgress; + blA += pk*LhsProgress; + + EIGEN_ASM_COMMENT("end gebp micro kernel 1/half/quarterX4"); + } + C0 = padd(C0,D0); + C1 = padd(C1,D1); + C2 = padd(C2,D2); + C3 = padd(C3,D3); + + // process remaining peeled loop + for(Index k=peeled_kc; k<depth; k++) + { + RhsPacketx4 rhs_panel; + RhsPacket T0; + peeled_kc_onestep(0, blA, blB, traits, &A0, &rhs_panel, &T0, &C0, &C1, &C2, &C3); + blB += 4*RhsProgress; + blA += LhsProgress; + } + + ResPacket R0, R1; + ResPacket alphav = pset1<ResPacket>(alpha); + + R0 = r0.template loadPacket<ResPacket>(0); + R1 = r1.template loadPacket<ResPacket>(0); + traits.acc(C0, alphav, R0); + traits.acc(C1, alphav, R1); + r0.storePacket(0, R0); + r1.storePacket(0, R1); + + R0 = r2.template loadPacket<ResPacket>(0); + R1 = r3.template loadPacket<ResPacket>(0); + traits.acc(C2, alphav, R0); + traits.acc(C3, alphav, R1); + r2.storePacket(0, R0); + r3.storePacket(0, R1); + } + + // Deal with remaining columns of the rhs + for(Index j2=packet_cols4; j2<cols; j2++) + { + // One column at a time + const LhsScalar* blA = &blockA[i*strideA+offsetA*(LhsProgress)]; + prefetch(&blA[0]); + + // gets res block as register + AccPacket C0; + traits.initAcc(C0); + + LinearMapper r0 = res.getLinearMapper(i, j2); + + // performs "inner" products + const RhsScalar* blB = &blockB[j2*strideB+offsetB]; + LhsPacket A0; + + for(Index k= 0; k<peeled_kc; k+=pk) + { + EIGEN_ASM_COMMENT("begin gebp micro kernel 1/half/quarterX1"); + RhsPacket B_0; + +#define EIGEN_GEBGP_ONESTEP(K) \ + do { \ + EIGEN_ASM_COMMENT("begin step of gebp micro kernel 1/half/quarterX1"); \ + EIGEN_ASM_COMMENT("Note: these asm comments work around bug 935!"); \ + /* FIXME: why unaligned???? */ \ + traits.loadLhsUnaligned(&blA[(0+1*K)*LhsProgress], A0); \ + traits.loadRhs(&blB[(0+K)*RhsProgress], B_0); \ + traits.madd(A0, B_0, C0, B_0, fix<0>); \ + EIGEN_ASM_COMMENT("end step of gebp micro kernel 1/half/quarterX1"); \ + } while(false); + + EIGEN_GEBGP_ONESTEP(0); + EIGEN_GEBGP_ONESTEP(1); + EIGEN_GEBGP_ONESTEP(2); + EIGEN_GEBGP_ONESTEP(3); + EIGEN_GEBGP_ONESTEP(4); + EIGEN_GEBGP_ONESTEP(5); + EIGEN_GEBGP_ONESTEP(6); + EIGEN_GEBGP_ONESTEP(7); + + blB += pk*RhsProgress; + blA += pk*LhsProgress; + + EIGEN_ASM_COMMENT("end gebp micro kernel 1/half/quarterX1"); + } + + // process remaining peeled loop + for(Index k=peeled_kc; k<depth; k++) + { + RhsPacket B_0; + EIGEN_GEBGP_ONESTEP(0); + blB += RhsProgress; + blA += LhsProgress; + } +#undef EIGEN_GEBGP_ONESTEP + ResPacket R0; + ResPacket alphav = pset1<ResPacket>(alpha); + R0 = r0.template loadPacket<ResPacket>(0); + traits.acc(C0, alphav, R0); + r0.storePacket(0, R0); + } + } + } +}; + +template<int nr, Index LhsProgress, Index RhsProgress, typename LhsScalar, typename RhsScalar, typename ResScalar, typename AccPacket, typename LhsPacket, typename RhsPacket, typename ResPacket, typename GEBPTraits, typename LinearMapper, typename DataMapper> +struct lhs_process_fraction_of_packet : lhs_process_one_packet<nr, LhsProgress, RhsProgress, LhsScalar, RhsScalar, ResScalar, AccPacket, LhsPacket, RhsPacket, ResPacket, GEBPTraits, LinearMapper, DataMapper> +{ + +EIGEN_STRONG_INLINE void peeled_kc_onestep(Index K, const LhsScalar* blA, const RhsScalar* blB, GEBPTraits traits, LhsPacket *A0, RhsPacket *B_0, RhsPacket *B1, RhsPacket *B2, RhsPacket *B3, AccPacket *C0, AccPacket *C1, AccPacket *C2, AccPacket *C3) + { + EIGEN_ASM_COMMENT("begin step of gebp micro kernel 1X4"); + EIGEN_ASM_COMMENT("Note: these asm comments work around bug 935!"); + traits.loadLhsUnaligned(&blA[(0+1*K)*(LhsProgress)], *A0); + traits.broadcastRhs(&blB[(0+4*K)*RhsProgress], *B_0, *B1, *B2, *B3); + traits.madd(*A0, *B_0, *C0, *B_0); + traits.madd(*A0, *B1, *C1, *B1); + traits.madd(*A0, *B2, *C2, *B2); + traits.madd(*A0, *B3, *C3, *B3); + EIGEN_ASM_COMMENT("end step of gebp micro kernel 1X4"); + } +}; + template<typename LhsScalar, typename RhsScalar, typename Index, typename DataMapper, int mr, int nr, bool ConjugateLhs, bool ConjugateRhs> EIGEN_DONT_INLINE void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,ConjugateRhs> @@ -987,10 +1561,12 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga Index packet_cols4 = nr>=4 ? (cols/4) * 4 : 0; const Index peeled_mc3 = mr>=3*Traits::LhsProgress ? (rows/(3*LhsProgress))*(3*LhsProgress) : 0; const Index peeled_mc2 = mr>=2*Traits::LhsProgress ? peeled_mc3+((rows-peeled_mc3)/(2*LhsProgress))*(2*LhsProgress) : 0; - const Index peeled_mc1 = mr>=1*Traits::LhsProgress ? (rows/(1*LhsProgress))*(1*LhsProgress) : 0; + const Index peeled_mc1 = mr>=1*Traits::LhsProgress ? peeled_mc2+((rows-peeled_mc2)/(1*LhsProgress))*(1*LhsProgress) : 0; + const Index peeled_mc_half = mr>=LhsProgressHalf ? peeled_mc1+((rows-peeled_mc1)/(LhsProgressHalf))*(LhsProgressHalf) : 0; + const Index peeled_mc_quarter = mr>=LhsProgressQuarter ? peeled_mc_half+((rows-peeled_mc_half)/(LhsProgressQuarter))*(LhsProgressQuarter) : 0; enum { pk = 8 }; // NOTE Such a large peeling factor is important for large matrices (~ +5% when >1000 on Haswell) const Index peeled_kc = depth & ~(pk-1); - const Index prefetch_res_offset = 32/sizeof(ResScalar); + const int prefetch_res_offset = 32/sizeof(ResScalar); // const Index depth2 = depth & ~1; //---------- Process 3 * LhsProgress rows at once ---------- @@ -1048,36 +1624,48 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga for(Index k=0; k<peeled_kc; k+=pk) { EIGEN_ASM_COMMENT("begin gebp micro kernel 3pX4"); - RhsPacket B_0, T0; + // 15 registers are taken (12 for acc, 2 for lhs). + RhsPanel15 rhs_panel; + RhsPacket T0; LhsPacket A2; - -#define EIGEN_GEBP_ONESTEP(K) \ - do { \ - EIGEN_ASM_COMMENT("begin step of gebp micro kernel 3pX4"); \ + #if EIGEN_COMP_GNUC_STRICT && EIGEN_ARCH_ARM64 && defined(EIGEN_VECTORIZE_NEON) && !(EIGEN_GNUC_AT_LEAST(9,0)) + // see http://eigen.tuxfamily.org/bz/show_bug.cgi?id=1633 + // without this workaround A0, A1, and A2 are loaded in the same register, + // which is not good for pipelining + #define EIGEN_GEBP_3PX4_REGISTER_ALLOC_WORKAROUND __asm__ ("" : "+w,m" (A0), "+w,m" (A1), "+w,m" (A2)); + #else + #define EIGEN_GEBP_3PX4_REGISTER_ALLOC_WORKAROUND + #endif +#define EIGEN_GEBP_ONESTEP(K) \ + do { \ + EIGEN_ASM_COMMENT("begin step of gebp micro kernel 3pX4"); \ EIGEN_ASM_COMMENT("Note: these asm comments work around bug 935!"); \ - internal::prefetch(blA+(3*K+16)*LhsProgress); \ - if (EIGEN_ARCH_ARM || EIGEN_ARCH_MIPS) { internal::prefetch(blB+(4*K+16)*RhsProgress); } /* Bug 953 */ \ - traits.loadLhs(&blA[(0+3*K)*LhsProgress], A0); \ - traits.loadLhs(&blA[(1+3*K)*LhsProgress], A1); \ - traits.loadLhs(&blA[(2+3*K)*LhsProgress], A2); \ - traits.loadRhs(blB + (0+4*K)*Traits::RhsProgress, B_0); \ - traits.madd(A0, B_0, C0, T0); \ - traits.madd(A1, B_0, C4, T0); \ - traits.madd(A2, B_0, C8, B_0); \ - traits.loadRhs(blB + (1+4*K)*Traits::RhsProgress, B_0); \ - traits.madd(A0, B_0, C1, T0); \ - traits.madd(A1, B_0, C5, T0); \ - traits.madd(A2, B_0, C9, B_0); \ - traits.loadRhs(blB + (2+4*K)*Traits::RhsProgress, B_0); \ - traits.madd(A0, B_0, C2, T0); \ - traits.madd(A1, B_0, C6, T0); \ - traits.madd(A2, B_0, C10, B_0); \ - traits.loadRhs(blB + (3+4*K)*Traits::RhsProgress, B_0); \ - traits.madd(A0, B_0, C3 , T0); \ - traits.madd(A1, B_0, C7, T0); \ - traits.madd(A2, B_0, C11, B_0); \ - EIGEN_ASM_COMMENT("end step of gebp micro kernel 3pX4"); \ - } while(false) + internal::prefetch(blA + (3 * K + 16) * LhsProgress); \ + if (EIGEN_ARCH_ARM || EIGEN_ARCH_MIPS) { \ + internal::prefetch(blB + (4 * K + 16) * RhsProgress); \ + } /* Bug 953 */ \ + traits.loadLhs(&blA[(0 + 3 * K) * LhsProgress], A0); \ + traits.loadLhs(&blA[(1 + 3 * K) * LhsProgress], A1); \ + traits.loadLhs(&blA[(2 + 3 * K) * LhsProgress], A2); \ + EIGEN_GEBP_3PX4_REGISTER_ALLOC_WORKAROUND \ + traits.loadRhs(blB + (0+4*K) * Traits::RhsProgress, rhs_panel); \ + traits.madd(A0, rhs_panel, C0, T0, fix<0>); \ + traits.madd(A1, rhs_panel, C4, T0, fix<0>); \ + traits.madd(A2, rhs_panel, C8, T0, fix<0>); \ + traits.updateRhs(blB + (1+4*K) * Traits::RhsProgress, rhs_panel); \ + traits.madd(A0, rhs_panel, C1, T0, fix<1>); \ + traits.madd(A1, rhs_panel, C5, T0, fix<1>); \ + traits.madd(A2, rhs_panel, C9, T0, fix<1>); \ + traits.updateRhs(blB + (2+4*K) * Traits::RhsProgress, rhs_panel); \ + traits.madd(A0, rhs_panel, C2, T0, fix<2>); \ + traits.madd(A1, rhs_panel, C6, T0, fix<2>); \ + traits.madd(A2, rhs_panel, C10, T0, fix<2>); \ + traits.updateRhs(blB + (3+4*K) * Traits::RhsProgress, rhs_panel); \ + traits.madd(A0, rhs_panel, C3, T0, fix<3>); \ + traits.madd(A1, rhs_panel, C7, T0, fix<3>); \ + traits.madd(A2, rhs_panel, C11, T0, fix<3>); \ + EIGEN_ASM_COMMENT("end step of gebp micro kernel 3pX4"); \ + } while (false) internal::prefetch(blB); EIGEN_GEBP_ONESTEP(0); @@ -1097,7 +1685,8 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga // process remaining peeled loop for(Index k=peeled_kc; k<depth; k++) { - RhsPacket B_0, T0; + RhsPanel15 rhs_panel; + RhsPacket T0; LhsPacket A2; EIGEN_GEBP_ONESTEP(0); blB += 4*RhsProgress; @@ -1177,20 +1766,20 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga { EIGEN_ASM_COMMENT("begin gebp micro kernel 3pX1"); RhsPacket B_0; -#define EIGEN_GEBGP_ONESTEP(K) \ - do { \ - EIGEN_ASM_COMMENT("begin step of gebp micro kernel 3pX1"); \ +#define EIGEN_GEBGP_ONESTEP(K) \ + do { \ + EIGEN_ASM_COMMENT("begin step of gebp micro kernel 3pX1"); \ EIGEN_ASM_COMMENT("Note: these asm comments work around bug 935!"); \ - traits.loadLhs(&blA[(0+3*K)*LhsProgress], A0); \ - traits.loadLhs(&blA[(1+3*K)*LhsProgress], A1); \ - traits.loadLhs(&blA[(2+3*K)*LhsProgress], A2); \ - traits.loadRhs(&blB[(0+K)*RhsProgress], B_0); \ - traits.madd(A0, B_0, C0, B_0); \ - traits.madd(A1, B_0, C4, B_0); \ - traits.madd(A2, B_0, C8, B_0); \ - EIGEN_ASM_COMMENT("end step of gebp micro kernel 3pX1"); \ - } while(false) - + traits.loadLhs(&blA[(0 + 3 * K) * LhsProgress], A0); \ + traits.loadLhs(&blA[(1 + 3 * K) * LhsProgress], A1); \ + traits.loadLhs(&blA[(2 + 3 * K) * LhsProgress], A2); \ + traits.loadRhs(&blB[(0 + K) * RhsProgress], B_0); \ + traits.madd(A0, B_0, C0, B_0, fix<0>); \ + traits.madd(A1, B_0, C4, B_0, fix<0>); \ + traits.madd(A2, B_0, C8, B_0, fix<0>); \ + EIGEN_ASM_COMMENT("end step of gebp micro kernel 3pX1"); \ + } while (false) + EIGEN_GEBGP_ONESTEP(0); EIGEN_GEBGP_ONESTEP(1); EIGEN_GEBGP_ONESTEP(2); @@ -1279,26 +1868,34 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga for(Index k=0; k<peeled_kc; k+=pk) { EIGEN_ASM_COMMENT("begin gebp micro kernel 2pX4"); - RhsPacket B_0, B1, B2, B3, T0; + RhsPacketx4 rhs_panel; + RhsPacket T0; + + // NOTE: the begin/end asm comments below work around bug 935! + // but they are not enough for gcc>=6 without FMA (bug 1637) + #if EIGEN_GNUC_AT_LEAST(6,0) && defined(EIGEN_VECTORIZE_SSE) + #define EIGEN_GEBP_2PX4_SPILLING_WORKAROUND __asm__ ("" : [a0] "+x,m" (A0),[a1] "+x,m" (A1)); + #else + #define EIGEN_GEBP_2PX4_SPILLING_WORKAROUND + #endif +#define EIGEN_GEBGP_ONESTEP(K) \ + do { \ + EIGEN_ASM_COMMENT("begin step of gebp micro kernel 2pX4"); \ + traits.loadLhs(&blA[(0 + 2 * K) * LhsProgress], A0); \ + traits.loadLhs(&blA[(1 + 2 * K) * LhsProgress], A1); \ + traits.loadRhs(&blB[(0 + 4 * K) * RhsProgress], rhs_panel); \ + traits.madd(A0, rhs_panel, C0, T0, fix<0>); \ + traits.madd(A1, rhs_panel, C4, T0, fix<0>); \ + traits.madd(A0, rhs_panel, C1, T0, fix<1>); \ + traits.madd(A1, rhs_panel, C5, T0, fix<1>); \ + traits.madd(A0, rhs_panel, C2, T0, fix<2>); \ + traits.madd(A1, rhs_panel, C6, T0, fix<2>); \ + traits.madd(A0, rhs_panel, C3, T0, fix<3>); \ + traits.madd(A1, rhs_panel, C7, T0, fix<3>); \ + EIGEN_GEBP_2PX4_SPILLING_WORKAROUND \ + EIGEN_ASM_COMMENT("end step of gebp micro kernel 2pX4"); \ + } while (false) - #define EIGEN_GEBGP_ONESTEP(K) \ - do { \ - EIGEN_ASM_COMMENT("begin step of gebp micro kernel 2pX4"); \ - EIGEN_ASM_COMMENT("Note: these asm comments work around bug 935!"); \ - traits.loadLhs(&blA[(0+2*K)*LhsProgress], A0); \ - traits.loadLhs(&blA[(1+2*K)*LhsProgress], A1); \ - traits.broadcastRhs(&blB[(0+4*K)*RhsProgress], B_0, B1, B2, B3); \ - traits.madd(A0, B_0, C0, T0); \ - traits.madd(A1, B_0, C4, B_0); \ - traits.madd(A0, B1, C1, T0); \ - traits.madd(A1, B1, C5, B1); \ - traits.madd(A0, B2, C2, T0); \ - traits.madd(A1, B2, C6, B2); \ - traits.madd(A0, B3, C3, T0); \ - traits.madd(A1, B3, C7, B3); \ - EIGEN_ASM_COMMENT("end step of gebp micro kernel 2pX4"); \ - } while(false) - internal::prefetch(blB+(48+0)); EIGEN_GEBGP_ONESTEP(0); EIGEN_GEBGP_ONESTEP(1); @@ -1318,7 +1915,8 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga // process remaining peeled loop for(Index k=peeled_kc; k<depth; k++) { - RhsPacket B_0, B1, B2, B3, T0; + RhsPacketx4 rhs_panel; + RhsPacket T0; EIGEN_GEBGP_ONESTEP(0); blB += 4*RhsProgress; blA += 2*Traits::LhsProgress; @@ -1389,8 +1987,8 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga traits.loadLhs(&blA[(0+2*K)*LhsProgress], A0); \ traits.loadLhs(&blA[(1+2*K)*LhsProgress], A1); \ traits.loadRhs(&blB[(0+K)*RhsProgress], B_0); \ - traits.madd(A0, B_0, C0, B1); \ - traits.madd(A1, B_0, C4, B_0); \ + traits.madd(A0, B_0, C0, B1, fix<0>); \ + traits.madd(A1, B_0, C4, B_0, fix<0>); \ EIGEN_ASM_COMMENT("end step of gebp micro kernel 2pX1"); \ } while(false) @@ -1434,174 +2032,29 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga //---------- Process 1 * LhsProgress rows at once ---------- if(mr>=1*Traits::LhsProgress) { - // loops on each largest micro horizontal panel of lhs (1*LhsProgress x depth) - for(Index i=peeled_mc2; i<peeled_mc1; i+=1*LhsProgress) - { - // loops on each largest micro vertical panel of rhs (depth * nr) - for(Index j2=0; j2<packet_cols4; j2+=nr) - { - // We select a 1*Traits::LhsProgress x nr micro block of res which is entirely - // stored into 1 x nr registers. - - const LhsScalar* blA = &blockA[i*strideA+offsetA*(1*Traits::LhsProgress)]; - prefetch(&blA[0]); - - // gets res block as register - AccPacket C0, C1, C2, C3; - traits.initAcc(C0); - traits.initAcc(C1); - traits.initAcc(C2); - traits.initAcc(C3); - - LinearMapper r0 = res.getLinearMapper(i, j2 + 0); - LinearMapper r1 = res.getLinearMapper(i, j2 + 1); - LinearMapper r2 = res.getLinearMapper(i, j2 + 2); - LinearMapper r3 = res.getLinearMapper(i, j2 + 3); - - r0.prefetch(prefetch_res_offset); - r1.prefetch(prefetch_res_offset); - r2.prefetch(prefetch_res_offset); - r3.prefetch(prefetch_res_offset); - - // performs "inner" products - const RhsScalar* blB = &blockB[j2*strideB+offsetB*nr]; - prefetch(&blB[0]); - LhsPacket A0; - - for(Index k=0; k<peeled_kc; k+=pk) - { - EIGEN_ASM_COMMENT("begin gebp micro kernel 1pX4"); - RhsPacket B_0, B1, B2, B3; - -#define EIGEN_GEBGP_ONESTEP(K) \ - do { \ - EIGEN_ASM_COMMENT("begin step of gebp micro kernel 1pX4"); \ - EIGEN_ASM_COMMENT("Note: these asm comments work around bug 935!"); \ - traits.loadLhs(&blA[(0+1*K)*LhsProgress], A0); \ - traits.broadcastRhs(&blB[(0+4*K)*RhsProgress], B_0, B1, B2, B3); \ - traits.madd(A0, B_0, C0, B_0); \ - traits.madd(A0, B1, C1, B1); \ - traits.madd(A0, B2, C2, B2); \ - traits.madd(A0, B3, C3, B3); \ - EIGEN_ASM_COMMENT("end step of gebp micro kernel 1pX4"); \ - } while(false) - - internal::prefetch(blB+(48+0)); - EIGEN_GEBGP_ONESTEP(0); - EIGEN_GEBGP_ONESTEP(1); - EIGEN_GEBGP_ONESTEP(2); - EIGEN_GEBGP_ONESTEP(3); - internal::prefetch(blB+(48+16)); - EIGEN_GEBGP_ONESTEP(4); - EIGEN_GEBGP_ONESTEP(5); - EIGEN_GEBGP_ONESTEP(6); - EIGEN_GEBGP_ONESTEP(7); - - blB += pk*4*RhsProgress; - blA += pk*1*LhsProgress; - - EIGEN_ASM_COMMENT("end gebp micro kernel 1pX4"); - } - // process remaining peeled loop - for(Index k=peeled_kc; k<depth; k++) - { - RhsPacket B_0, B1, B2, B3; - EIGEN_GEBGP_ONESTEP(0); - blB += 4*RhsProgress; - blA += 1*LhsProgress; - } -#undef EIGEN_GEBGP_ONESTEP - - ResPacket R0, R1; - ResPacket alphav = pset1<ResPacket>(alpha); - - R0 = r0.template loadPacket<ResPacket>(0 * Traits::ResPacketSize); - R1 = r1.template loadPacket<ResPacket>(0 * Traits::ResPacketSize); - traits.acc(C0, alphav, R0); - traits.acc(C1, alphav, R1); - r0.storePacket(0 * Traits::ResPacketSize, R0); - r1.storePacket(0 * Traits::ResPacketSize, R1); - - R0 = r2.template loadPacket<ResPacket>(0 * Traits::ResPacketSize); - R1 = r3.template loadPacket<ResPacket>(0 * Traits::ResPacketSize); - traits.acc(C2, alphav, R0); - traits.acc(C3, alphav, R1); - r2.storePacket(0 * Traits::ResPacketSize, R0); - r3.storePacket(0 * Traits::ResPacketSize, R1); - } - - // Deal with remaining columns of the rhs - for(Index j2=packet_cols4; j2<cols; j2++) - { - // One column at a time - const LhsScalar* blA = &blockA[i*strideA+offsetA*(1*Traits::LhsProgress)]; - prefetch(&blA[0]); - - // gets res block as register - AccPacket C0; - traits.initAcc(C0); - - LinearMapper r0 = res.getLinearMapper(i, j2); - - // performs "inner" products - const RhsScalar* blB = &blockB[j2*strideB+offsetB]; - LhsPacket A0; - - for(Index k=0; k<peeled_kc; k+=pk) - { - EIGEN_ASM_COMMENT("begin gebp micro kernel 1pX1"); - RhsPacket B_0; - -#define EIGEN_GEBGP_ONESTEP(K) \ - do { \ - EIGEN_ASM_COMMENT("begin step of gebp micro kernel 1pX1"); \ - EIGEN_ASM_COMMENT("Note: these asm comments work around bug 935!"); \ - traits.loadLhs(&blA[(0+1*K)*LhsProgress], A0); \ - traits.loadRhs(&blB[(0+K)*RhsProgress], B_0); \ - traits.madd(A0, B_0, C0, B_0); \ - EIGEN_ASM_COMMENT("end step of gebp micro kernel 1pX1"); \ - } while(false); - - EIGEN_GEBGP_ONESTEP(0); - EIGEN_GEBGP_ONESTEP(1); - EIGEN_GEBGP_ONESTEP(2); - EIGEN_GEBGP_ONESTEP(3); - EIGEN_GEBGP_ONESTEP(4); - EIGEN_GEBGP_ONESTEP(5); - EIGEN_GEBGP_ONESTEP(6); - EIGEN_GEBGP_ONESTEP(7); - - blB += pk*RhsProgress; - blA += pk*1*Traits::LhsProgress; - - EIGEN_ASM_COMMENT("end gebp micro kernel 1pX1"); - } - - // process remaining peeled loop - for(Index k=peeled_kc; k<depth; k++) - { - RhsPacket B_0; - EIGEN_GEBGP_ONESTEP(0); - blB += RhsProgress; - blA += 1*Traits::LhsProgress; - } -#undef EIGEN_GEBGP_ONESTEP - ResPacket R0; - ResPacket alphav = pset1<ResPacket>(alpha); - R0 = r0.template loadPacket<ResPacket>(0 * Traits::ResPacketSize); - traits.acc(C0, alphav, R0); - r0.storePacket(0 * Traits::ResPacketSize, R0); - } - } + lhs_process_one_packet<nr, LhsProgress, RhsProgress, LhsScalar, RhsScalar, ResScalar, AccPacket, LhsPacket, RhsPacket, ResPacket, Traits, LinearMapper, DataMapper> p; + p(res, blockA, blockB, alpha, peeled_mc2, peeled_mc1, strideA, strideB, offsetA, offsetB, prefetch_res_offset, peeled_kc, pk, cols, depth, packet_cols4); + } + //---------- Process LhsProgressHalf rows at once ---------- + if((LhsProgressHalf < LhsProgress) && mr>=LhsProgressHalf) + { + lhs_process_fraction_of_packet<nr, LhsProgressHalf, RhsProgressHalf, LhsScalar, RhsScalar, ResScalar, AccPacketHalf, LhsPacketHalf, RhsPacketHalf, ResPacketHalf, HalfTraits, LinearMapper, DataMapper> p; + p(res, blockA, blockB, alpha, peeled_mc1, peeled_mc_half, strideA, strideB, offsetA, offsetB, prefetch_res_offset, peeled_kc, pk, cols, depth, packet_cols4); + } + //---------- Process LhsProgressQuarter rows at once ---------- + if((LhsProgressQuarter < LhsProgressHalf) && mr>=LhsProgressQuarter) + { + lhs_process_fraction_of_packet<nr, LhsProgressQuarter, RhsProgressQuarter, LhsScalar, RhsScalar, ResScalar, AccPacketQuarter, LhsPacketQuarter, RhsPacketQuarter, ResPacketQuarter, QuarterTraits, LinearMapper, DataMapper> p; + p(res, blockA, blockB, alpha, peeled_mc_half, peeled_mc_quarter, strideA, strideB, offsetA, offsetB, prefetch_res_offset, peeled_kc, pk, cols, depth, packet_cols4); } //---------- Process remaining rows, 1 at once ---------- - if(peeled_mc1<rows) + if(peeled_mc_quarter<rows) { // loop on each panel of the rhs for(Index j2=0; j2<packet_cols4; j2+=nr) { // loop on each row of the lhs (1*LhsProgress x depth) - for(Index i=peeled_mc1; i<rows; i+=1) + for(Index i=peeled_mc_quarter; i<rows; i+=1) { const LhsScalar* blA = &blockA[i*strideA+offsetA]; prefetch(&blA[0]); @@ -1614,7 +2067,7 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga const int SResPacketQuarterSize = unpacket_traits<typename unpacket_traits<typename unpacket_traits<SResPacket>::half>::half>::size; if ((SwappedTraits::LhsProgress % 4) == 0 && (SwappedTraits::LhsProgress<=16) && - (SwappedTraits::LhsProgress!=8 || SResPacketHalfSize==nr) && + (SwappedTraits::LhsProgress!=8 || SResPacketHalfSize==nr) && (SwappedTraits::LhsProgress!=16 || SResPacketQuarterSize==nr)) { SAccPacket C0, C1, C2, C3; @@ -1638,15 +2091,15 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga straits.loadRhsQuad(blA+0*spk, B_0); straits.loadRhsQuad(blA+1*spk, B_1); - straits.madd(A0,B_0,C0,B_0); - straits.madd(A1,B_1,C1,B_1); + straits.madd(A0,B_0,C0,B_0, fix<0>); + straits.madd(A1,B_1,C1,B_1, fix<0>); straits.loadLhsUnaligned(blB+2*SwappedTraits::LhsProgress, A0); straits.loadLhsUnaligned(blB+3*SwappedTraits::LhsProgress, A1); straits.loadRhsQuad(blA+2*spk, B_0); straits.loadRhsQuad(blA+3*spk, B_1); - straits.madd(A0,B_0,C2,B_0); - straits.madd(A1,B_1,C3,B_1); + straits.madd(A0,B_0,C2,B_0, fix<0>); + straits.madd(A1,B_1,C3,B_1, fix<0>); blB += 4*SwappedTraits::LhsProgress; blA += 4*spk; @@ -1659,7 +2112,7 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga straits.loadLhsUnaligned(blB, A0); straits.loadRhsQuad(blA, B_0); - straits.madd(A0,B_0,C0,B_0); + straits.madd(A0,B_0,C0,B_0, fix<0>); blB += SwappedTraits::LhsProgress; blA += spk; @@ -1669,7 +2122,7 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga // Special case where we have to first reduce the accumulation register C0 typedef typename conditional<SwappedTraits::LhsProgress>=8,typename unpacket_traits<SResPacket>::half,SResPacket>::type SResPacketHalf; typedef typename conditional<SwappedTraits::LhsProgress>=8,typename unpacket_traits<SLhsPacket>::half,SLhsPacket>::type SLhsPacketHalf; - typedef typename conditional<SwappedTraits::LhsProgress>=8,typename unpacket_traits<SLhsPacket>::half,SRhsPacket>::type SRhsPacketHalf; + typedef typename conditional<SwappedTraits::LhsProgress>=8,typename unpacket_traits<SRhsPacket>::half,SRhsPacket>::type SRhsPacketHalf; typedef typename conditional<SwappedTraits::LhsProgress>=8,typename unpacket_traits<SAccPacket>::half,SAccPacket>::type SAccPacketHalf; SResPacketHalf R = res.template gatherPacket<SResPacketHalf>(i, j2); @@ -1683,7 +2136,7 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga straits.loadLhsUnaligned(blB, a0); straits.loadRhs(blA, b0); SAccPacketHalf c0 = predux_half_dowto4(C0); - straits.madd(a0,b0,c0,b0); + straits.madd(a0,b0,c0,b0, fix<0>); straits.acc(c0, alphav, R); } else @@ -1699,7 +2152,7 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga // template form, so that LhsProgress < 16 paths don't // fail to compile last_row_process_16_packets<LhsScalar, RhsScalar, Index, DataMapper, mr, nr, ConjugateLhs, ConjugateRhs> p; - p(res, straits, blA, blB, depth, endk, i, j2,alpha, C0); + p(res, straits, blA, blB, depth, endk, i, j2,alpha, C0); } else { @@ -1744,7 +2197,7 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga for(Index j2=packet_cols4; j2<cols; j2++) { // loop on each row of the lhs (1*LhsProgress x depth) - for(Index i=peeled_mc1; i<rows; i+=1) + for(Index i=peeled_mc_quarter; i<rows; i+=1) { const LhsScalar* blA = &blockA[i*strideA+offsetA]; prefetch(&blA[0]); @@ -1791,7 +2244,13 @@ template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pa EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, Packet, ColMajor, Conjugate, PanelMode> ::operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride, Index offset) { - enum { PacketSize = unpacket_traits<Packet>::size }; + typedef typename unpacket_traits<Packet>::half HalfPacket; + typedef typename unpacket_traits<typename unpacket_traits<Packet>::half>::half QuarterPacket; + enum { PacketSize = unpacket_traits<Packet>::size, + HalfPacketSize = unpacket_traits<HalfPacket>::size, + QuarterPacketSize = unpacket_traits<QuarterPacket>::size, + HasHalf = (int)HalfPacketSize < (int)PacketSize, + HasQuarter = (int)QuarterPacketSize < (int)HalfPacketSize}; EIGEN_ASM_COMMENT("EIGEN PRODUCT PACK LHS"); EIGEN_UNUSED_VARIABLE(stride); @@ -1803,9 +2262,12 @@ EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, Pa const Index peeled_mc3 = Pack1>=3*PacketSize ? (rows/(3*PacketSize))*(3*PacketSize) : 0; const Index peeled_mc2 = Pack1>=2*PacketSize ? peeled_mc3+((rows-peeled_mc3)/(2*PacketSize))*(2*PacketSize) : 0; - const Index peeled_mc1 = Pack1>=1*PacketSize ? (rows/(1*PacketSize))*(1*PacketSize) : 0; - const Index peeled_mc0 = Pack2>=1*PacketSize ? peeled_mc1 - : Pack2>1 ? (rows/Pack2)*Pack2 : 0; + const Index peeled_mc1 = Pack1>=1*PacketSize ? peeled_mc2+((rows-peeled_mc2)/(1*PacketSize))*(1*PacketSize) : 0; + const Index peeled_mc_half = Pack1>=HalfPacketSize ? peeled_mc1+((rows-peeled_mc1)/(HalfPacketSize))*(HalfPacketSize) : 0; + const Index peeled_mc_quarter = Pack1>=QuarterPacketSize ? (rows/(QuarterPacketSize))*(QuarterPacketSize) : 0; + const Index last_lhs_progress = rows > peeled_mc_quarter ? (rows - peeled_mc_quarter) & ~1 : 0; + const Index peeled_mc0 = Pack2>=PacketSize ? peeled_mc_quarter + : Pack2>1 && last_lhs_progress ? (rows/last_lhs_progress)*last_lhs_progress : 0; Index i=0; @@ -1864,20 +2326,60 @@ EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, Pa if(PanelMode) count += (1*PacketSize) * (stride-offset-depth); } } - // Pack scalars + // Pack half packets + if(HasHalf && Pack1>=HalfPacketSize) + { + for(; i<peeled_mc_half; i+=HalfPacketSize) + { + if(PanelMode) count += (HalfPacketSize) * offset; + + for(Index k=0; k<depth; k++) + { + HalfPacket A; + A = lhs.template loadPacket<HalfPacket>(i+0*(HalfPacketSize), k); + pstoreu(blockA+count, cj.pconj(A)); + count+=HalfPacketSize; + } + if(PanelMode) count += (HalfPacketSize) * (stride-offset-depth); + } + } + // Pack quarter packets + if(HasQuarter && Pack1>=QuarterPacketSize) + { + for(; i<peeled_mc_quarter; i+=QuarterPacketSize) + { + if(PanelMode) count += (QuarterPacketSize) * offset; + + for(Index k=0; k<depth; k++) + { + QuarterPacket A; + A = lhs.template loadPacket<QuarterPacket>(i+0*(QuarterPacketSize), k); + pstoreu(blockA+count, cj.pconj(A)); + count+=QuarterPacketSize; + } + if(PanelMode) count += (QuarterPacketSize) * (stride-offset-depth); + } + } + // Pack2 may be *smaller* than PacketSize—that happens for + // products like real * complex, where we have to go half the + // progress on the lhs in order to duplicate those operands to + // address both real & imaginary parts on the rhs. This portion will + // pack those half ones until they match the number expected on the + // last peeling loop at this point (for the rhs). if(Pack2<PacketSize && Pack2>1) { - for(; i<peeled_mc0; i+=Pack2) + for(; i<peeled_mc0; i+=last_lhs_progress) { - if(PanelMode) count += Pack2 * offset; + if(PanelMode) count += last_lhs_progress * offset; for(Index k=0; k<depth; k++) - for(Index w=0; w<Pack2; w++) + for(Index w=0; w<last_lhs_progress; w++) blockA[count++] = cj(lhs(i+w, k)); - if(PanelMode) count += Pack2 * (stride-offset-depth); + if(PanelMode) count += last_lhs_progress * (stride-offset-depth); } } + // Pack scalars for(; i<rows; i++) { if(PanelMode) count += offset; @@ -1898,7 +2400,13 @@ template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pa EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, Packet, RowMajor, Conjugate, PanelMode> ::operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride, Index offset) { - enum { PacketSize = unpacket_traits<Packet>::size }; + typedef typename unpacket_traits<Packet>::half HalfPacket; + typedef typename unpacket_traits<typename unpacket_traits<Packet>::half>::half QuarterPacket; + enum { PacketSize = unpacket_traits<Packet>::size, + HalfPacketSize = unpacket_traits<HalfPacket>::size, + QuarterPacketSize = unpacket_traits<QuarterPacket>::size, + HasHalf = (int)HalfPacketSize < (int)PacketSize, + HasQuarter = (int)QuarterPacketSize < (int)HalfPacketSize}; EIGEN_ASM_COMMENT("EIGEN PRODUCT PACK LHS"); EIGEN_UNUSED_VARIABLE(stride); @@ -1906,37 +2414,51 @@ EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, Pa eigen_assert(((!PanelMode) && stride==0 && offset==0) || (PanelMode && stride>=depth && offset<=stride)); conj_if<NumTraits<Scalar>::IsComplex && Conjugate> cj; Index count = 0; + bool gone_half = false, gone_quarter = false, gone_last = false; -// const Index peeled_mc3 = Pack1>=3*PacketSize ? (rows/(3*PacketSize))*(3*PacketSize) : 0; -// const Index peeled_mc2 = Pack1>=2*PacketSize ? peeled_mc3+((rows-peeled_mc3)/(2*PacketSize))*(2*PacketSize) : 0; -// const Index peeled_mc1 = Pack1>=1*PacketSize ? (rows/(1*PacketSize))*(1*PacketSize) : 0; - - int pack = Pack1; Index i = 0; + int pack = Pack1; + int psize = PacketSize; while(pack>0) { Index remaining_rows = rows-i; - Index peeled_mc = i+(remaining_rows/pack)*pack; + Index peeled_mc = gone_last ? Pack2>1 ? (rows/pack)*pack : 0 : i+(remaining_rows/pack)*pack; + Index starting_pos = i; for(; i<peeled_mc; i+=pack) { if(PanelMode) count += pack * offset; - const Index peeled_k = (depth/PacketSize)*PacketSize; Index k=0; - if(pack>=PacketSize) + if(pack>=psize && psize >= QuarterPacketSize) { - for(; k<peeled_k; k+=PacketSize) + const Index peeled_k = (depth/psize)*psize; + for(; k<peeled_k; k+=psize) { - for (Index m = 0; m < pack; m += PacketSize) + for (Index m = 0; m < pack; m += psize) { - PacketBlock<Packet> kernel; - for (int p = 0; p < PacketSize; ++p) kernel.packet[p] = lhs.template loadPacket<Packet>(i+p+m, k); - ptranspose(kernel); - for (int p = 0; p < PacketSize; ++p) pstore(blockA+count+m+(pack)*p, cj.pconj(kernel.packet[p])); + if (psize == PacketSize) { + PacketBlock<Packet> kernel; + for (int p = 0; p < psize; ++p) kernel.packet[p] = lhs.template loadPacket<Packet>(i+p+m, k); + ptranspose(kernel); + for (int p = 0; p < psize; ++p) pstore(blockA+count+m+(pack)*p, cj.pconj(kernel.packet[p])); + } else if (HasHalf && psize == HalfPacketSize) { + gone_half = true; + PacketBlock<HalfPacket> kernel_half; + for (int p = 0; p < psize; ++p) kernel_half.packet[p] = lhs.template loadPacket<HalfPacket>(i+p+m, k); + ptranspose(kernel_half); + for (int p = 0; p < psize; ++p) pstore(blockA+count+m+(pack)*p, cj.pconj(kernel_half.packet[p])); + } else if (HasQuarter && psize == QuarterPacketSize) { + gone_quarter = true; + PacketBlock<QuarterPacket> kernel_quarter; + for (int p = 0; p < psize; ++p) kernel_quarter.packet[p] = lhs.template loadPacket<QuarterPacket>(i+p+m, k); + ptranspose(kernel_quarter); + for (int p = 0; p < psize; ++p) pstore(blockA+count+m+(pack)*p, cj.pconj(kernel_quarter.packet[p])); + } } - count += PacketSize*pack; + count += psize*pack; } } + for(; k<depth; k++) { Index w=0; @@ -1959,9 +2481,28 @@ EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, Pa if(PanelMode) count += pack * (stride-offset-depth); } - pack -= PacketSize; - if(pack<Pack2 && (pack+PacketSize)!=Pack2) - pack = Pack2; + pack -= psize; + Index left = rows - i; + if (pack <= 0) { + if (!gone_last && + (starting_pos == i || left >= psize/2 || left >= psize/4) && + ((psize/2 == HalfPacketSize && HasHalf && !gone_half) || + (psize/2 == QuarterPacketSize && HasQuarter && !gone_quarter))) { + psize /= 2; + pack = psize; + continue; + } + // Pack2 may be *smaller* than PacketSize—that happens for + // products like real * complex, where we have to go half the + // progress on the lhs in order to duplicate those operands to + // address both real & imaginary parts on the rhs. This portion will + // pack those half ones until they match the number expected on the + // last peeling loop at this point (for the rhs). + if (Pack2 < PacketSize && !gone_last) { + gone_last = true; + psize = pack = left & ~1; + } + } } for(; i<rows; i++) diff --git a/Eigen/src/Core/products/GeneralMatrixMatrix.h b/Eigen/src/Core/products/GeneralMatrixMatrix.h index f49abcad5..90c9c4647 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrix.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrix.h @@ -469,6 +469,20 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct> if(a_lhs.cols()==0 || a_lhs.rows()==0 || a_rhs.cols()==0) return; + // Fallback to GEMV if either the lhs or rhs is a runtime vector + if (dst.cols() == 1) + { + typename Dest::ColXpr dst_vec(dst.col(0)); + return internal::generic_product_impl<Lhs,typename Rhs::ConstColXpr,DenseShape,DenseShape,GemvProduct> + ::scaleAndAddTo(dst_vec, a_lhs, a_rhs.col(0), alpha); + } + else if (dst.rows() == 1) + { + typename Dest::RowXpr dst_vec(dst.row(0)); + return internal::generic_product_impl<typename Lhs::ConstRowXpr,Rhs,DenseShape,DenseShape,GemvProduct> + ::scaleAndAddTo(dst_vec, a_lhs.row(0), a_rhs, alpha); + } + typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(a_lhs); typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(a_rhs); diff --git a/Eigen/src/Core/products/Parallelizer.h b/Eigen/src/Core/products/Parallelizer.h index 92e9b0d9f..e01e798f1 100644 --- a/Eigen/src/Core/products/Parallelizer.h +++ b/Eigen/src/Core/products/Parallelizer.h @@ -21,7 +21,8 @@ namespace internal { /** \internal */ inline void manage_multi_threading(Action action, int* v) { - static EIGEN_UNUSED int m_maxThreads = -1; + static int m_maxThreads = -1; + EIGEN_UNUSED_VARIABLE(m_maxThreads); if(action==SetAction) { diff --git a/Eigen/src/Core/products/SelfadjointMatrixMatrix.h b/Eigen/src/Core/products/SelfadjointMatrixMatrix.h index c84c71609..673073601 100644 --- a/Eigen/src/Core/products/SelfadjointMatrixMatrix.h +++ b/Eigen/src/Core/products/SelfadjointMatrixMatrix.h @@ -45,14 +45,23 @@ struct symm_pack_lhs } void operator()(Scalar* blockA, const Scalar* _lhs, Index lhsStride, Index cols, Index rows) { - enum { PacketSize = packet_traits<Scalar>::size }; + typedef typename unpacket_traits<typename packet_traits<Scalar>::type>::half HalfPacket; + typedef typename unpacket_traits<typename unpacket_traits<typename packet_traits<Scalar>::type>::half>::half QuarterPacket; + enum { PacketSize = packet_traits<Scalar>::size, + HalfPacketSize = unpacket_traits<HalfPacket>::size, + QuarterPacketSize = unpacket_traits<QuarterPacket>::size, + HasHalf = (int)HalfPacketSize < (int)PacketSize, + HasQuarter = (int)QuarterPacketSize < (int)HalfPacketSize}; + const_blas_data_mapper<Scalar,Index,StorageOrder> lhs(_lhs,lhsStride); Index count = 0; //Index peeled_mc3 = (rows/Pack1)*Pack1; const Index peeled_mc3 = Pack1>=3*PacketSize ? (rows/(3*PacketSize))*(3*PacketSize) : 0; const Index peeled_mc2 = Pack1>=2*PacketSize ? peeled_mc3+((rows-peeled_mc3)/(2*PacketSize))*(2*PacketSize) : 0; - const Index peeled_mc1 = Pack1>=1*PacketSize ? (rows/(1*PacketSize))*(1*PacketSize) : 0; + const Index peeled_mc1 = Pack1>=1*PacketSize ? peeled_mc2+((rows-peeled_mc2)/(1*PacketSize))*(1*PacketSize) : 0; + const Index peeled_mc_half = Pack1>=HalfPacketSize ? peeled_mc1+((rows-peeled_mc1)/(HalfPacketSize))*(HalfPacketSize) : 0; + const Index peeled_mc_quarter = Pack1>=QuarterPacketSize ? peeled_mc_half+((rows-peeled_mc_half)/(QuarterPacketSize))*(QuarterPacketSize) : 0; if(Pack1>=3*PacketSize) for(Index i=0; i<peeled_mc3; i+=3*PacketSize) @@ -66,8 +75,16 @@ struct symm_pack_lhs for(Index i=peeled_mc2; i<peeled_mc1; i+=1*PacketSize) pack<1*PacketSize>(blockA, lhs, cols, i, count); + if(HasHalf && Pack1>=HalfPacketSize) + for(Index i=peeled_mc1; i<peeled_mc_half; i+=HalfPacketSize) + pack<HalfPacketSize>(blockA, lhs, cols, i, count); + + if(HasQuarter && Pack1>=QuarterPacketSize) + for(Index i=peeled_mc_half; i<peeled_mc_quarter; i+=QuarterPacketSize) + pack<QuarterPacketSize>(blockA, lhs, cols, i, count); + // do the same with mr==1 - for(Index i=peeled_mc1; i<rows; i++) + for(Index i=peeled_mc_quarter; i<rows; i++) { for(Index k=0; k<i; k++) blockA[count++] = lhs(i, k); // normal diff --git a/Eigen/src/Core/util/BlasUtil.h b/Eigen/src/Core/util/BlasUtil.h index a32630ed7..e6689c656 100755 --- a/Eigen/src/Core/util/BlasUtil.h +++ b/Eigen/src/Core/util/BlasUtil.h @@ -274,7 +274,8 @@ template<typename XprType> struct blas_traits HasUsableDirectAccess = ( (int(XprType::Flags)&DirectAccessBit) && ( bool(XprType::IsVectorAtCompileTime) || int(inner_stride_at_compile_time<XprType>::ret) == 1) - ) ? 1 : 0 + ) ? 1 : 0, + HasScalarFactor = false }; typedef typename conditional<bool(HasUsableDirectAccess), ExtractType, @@ -306,6 +307,9 @@ template<typename Scalar, typename NestedXpr, typename Plain> struct blas_traits<CwiseBinaryOp<scalar_product_op<Scalar>, const CwiseNullaryOp<scalar_constant_op<Scalar>,Plain>, NestedXpr> > : blas_traits<NestedXpr> { + enum { + HasScalarFactor = true + }; typedef blas_traits<NestedXpr> Base; typedef CwiseBinaryOp<scalar_product_op<Scalar>, const CwiseNullaryOp<scalar_constant_op<Scalar>,Plain>, NestedXpr> XprType; typedef typename Base::ExtractType ExtractType; @@ -317,6 +321,9 @@ template<typename Scalar, typename NestedXpr, typename Plain> struct blas_traits<CwiseBinaryOp<scalar_product_op<Scalar>, NestedXpr, const CwiseNullaryOp<scalar_constant_op<Scalar>,Plain> > > : blas_traits<NestedXpr> { + enum { + HasScalarFactor = true + }; typedef blas_traits<NestedXpr> Base; typedef CwiseBinaryOp<scalar_product_op<Scalar>, NestedXpr, const CwiseNullaryOp<scalar_constant_op<Scalar>,Plain> > XprType; typedef typename Base::ExtractType ExtractType; @@ -335,6 +342,9 @@ template<typename Scalar, typename NestedXpr> struct blas_traits<CwiseUnaryOp<scalar_opposite_op<Scalar>, NestedXpr> > : blas_traits<NestedXpr> { + enum { + HasScalarFactor = true + }; typedef blas_traits<NestedXpr> Base; typedef CwiseUnaryOp<scalar_opposite_op<Scalar>, NestedXpr> XprType; typedef typename Base::ExtractType ExtractType; diff --git a/Eigen/src/Core/util/ConfigureVectorization.h b/Eigen/src/Core/util/ConfigureVectorization.h index 263604597..b00d8b038 100644 --- a/Eigen/src/Core/util/ConfigureVectorization.h +++ b/Eigen/src/Core/util/ConfigureVectorization.h @@ -10,13 +10,6 @@ #ifndef EIGEN_CONFIGURE_VECTORIZATION_H #define EIGEN_CONFIGURE_VECTORIZATION_H -// FIXME: not sure why this is needed, perhaps it is not needed anymore. -#ifdef __NVCC__ - #ifndef EIGEN_DONT_VECTORIZE - #define EIGEN_DONT_VECTORIZE - #endif -#endif - //------------------------------------------------------------------------------------------ // Static and dynamic alignment control // @@ -36,10 +29,15 @@ * * If we made alignment depend on whether or not EIGEN_VECTORIZE is defined, it would be impossible to link * vectorized and non-vectorized code. + * + * FIXME: this code can be cleaned up once we switch to proper C++11 only. */ #if (defined EIGEN_CUDACC) #define EIGEN_ALIGN_TO_BOUNDARY(n) __align__(n) #define EIGEN_ALIGNOF(x) __alignof(x) +#elif EIGEN_HAS_ALIGNAS + #define EIGEN_ALIGN_TO_BOUNDARY(n) alignas(n) + #define EIGEN_ALIGNOF(x) alignof(x) #elif EIGEN_COMP_GNUC || EIGEN_COMP_PGI || EIGEN_COMP_IBM || EIGEN_COMP_ARM #define EIGEN_ALIGN_TO_BOUNDARY(n) __attribute__((aligned(n))) #define EIGEN_ALIGNOF(x) __alignof(x) @@ -51,12 +49,18 @@ #define EIGEN_ALIGN_TO_BOUNDARY(n) __attribute__((aligned(n))) #define EIGEN_ALIGNOF(x) __alignof(x) #else - #error Please tell me what is the equivalent of __attribute__((aligned(n))) and __alignof(x) for your compiler + #error Please tell me what is the equivalent of alignas(n) and alignof(x) for your compiler #endif // If the user explicitly disable vectorization, then we also disable alignment #if defined(EIGEN_DONT_VECTORIZE) - #define EIGEN_IDEAL_MAX_ALIGN_BYTES 0 + #if defined(EIGEN_GPUCC) + // GPU code is always vectorized and requires memory alignment for + // statically allocated buffers. + #define EIGEN_IDEAL_MAX_ALIGN_BYTES 16 + #else + #define EIGEN_IDEAL_MAX_ALIGN_BYTES 0 + #endif #elif defined(__AVX512F__) // 64 bytes static alignment is preferred only if really required #define EIGEN_IDEAL_MAX_ALIGN_BYTES 64 @@ -183,8 +187,6 @@ //---------------------------------------------------------------------- - - // if alignment is disabled, then disable vectorization. Note: EIGEN_MAX_ALIGN_BYTES is the proper check, it takes into // account both the user's will (EIGEN_MAX_ALIGN_BYTES,EIGEN_DONT_ALIGN) and our own platform checks #if EIGEN_MAX_ALIGN_BYTES==0 @@ -211,7 +213,7 @@ #endif -#ifndef EIGEN_DONT_VECTORIZE +#if !(defined(EIGEN_DONT_VECTORIZE) || defined(EIGEN_GPUCC)) #if defined (EIGEN_SSE2_ON_NON_MSVC_BUT_NOT_OLD_GCC) || defined(EIGEN_SSE2_ON_MSVC_2008_OR_LATER) @@ -253,10 +255,19 @@ #define EIGEN_VECTORIZE_SSE4_1 #define EIGEN_VECTORIZE_SSE4_2 #endif - #ifdef __FMA__ + #if defined(__FMA__) || (EIGEN_COMP_MSVC && defined(__AVX2__)) + // MSVC does not expose a switch dedicated for FMA + // For MSVC, AVX2 => FMA #define EIGEN_VECTORIZE_FMA #endif #if defined(__AVX512F__) + #ifndef EIGEN_VECTORIZE_FMA + #if EIGEN_COMP_GNUC + #error Please add -mfma to your compiler flags: compiling with -mavx512f alone without SSE/AVX FMA is not supported (bug 1638). + #else + #error Please enable FMA in your compiler flags (e.g. -mfma): compiling with AVX512 alone without SSE/AVX FMA is not supported (bug 1638). + #endif + #endif #define EIGEN_VECTORIZE_AVX512 #define EIGEN_VECTORIZE_AVX2 #define EIGEN_VECTORIZE_AVX @@ -375,7 +386,7 @@ #endif #if defined(EIGEN_HAS_CUDA_FP16) - #include <host_defines.h> + #include <cuda_runtime_api.h> #include <cuda_fp16.h> #endif diff --git a/Eigen/src/Core/util/ForwardDeclarations.h b/Eigen/src/Core/util/ForwardDeclarations.h index 3ab3a5f50..050d15e96 100644 --- a/Eigen/src/Core/util/ForwardDeclarations.h +++ b/Eigen/src/Core/util/ForwardDeclarations.h @@ -47,11 +47,7 @@ template<typename T> struct NumTraits; template<typename Derived> struct EigenBase; template<typename Derived> class DenseBase; template<typename Derived> class PlainObjectBase; - - -template<typename Derived, - int Level = internal::accessors_level<Derived>::value > -class DenseCoeffsBase; +template<typename Derived, int Level> class DenseCoeffsBase; template<typename _Scalar, int _Rows, int _Cols, int _Options = AutoAlign | @@ -260,6 +256,7 @@ template<typename MatrixType> class HouseholderQR; template<typename MatrixType> class ColPivHouseholderQR; template<typename MatrixType> class FullPivHouseholderQR; template<typename MatrixType> class CompleteOrthogonalDecomposition; +template<typename MatrixType> class SVDBase; template<typename MatrixType, int QRPreconditioner = ColPivHouseholderQRPreconditioner> class JacobiSVD; template<typename MatrixType> class BDCSVD; template<typename MatrixType, int UpLo = Lower> class LLT; diff --git a/Eigen/src/Core/util/IndexedViewHelper.h b/Eigen/src/Core/util/IndexedViewHelper.h index 40e16fdb4..1cda85060 100644 --- a/Eigen/src/Core/util/IndexedViewHelper.h +++ b/Eigen/src/Core/util/IndexedViewHelper.h @@ -23,7 +23,7 @@ struct symbolic_last_tag {}; * Can be used as a parameter to Eigen::seq and Eigen::seqN functions to symbolically reference the last element/row/columns * of the underlying vector or matrix once passed to DenseBase::operator()(const RowIndices&, const ColIndices&). * - * This symbolic placeholder support standard arithmetic operation. + * This symbolic placeholder supports standard arithmetic operations. * * A typical usage example would be: * \code @@ -44,7 +44,7 @@ static const symbolic::SymbolExpr<internal::symbolic_last_tag> last; // PLEASE u * reference the last+1 element/row/columns of the underlying vector or matrix once * passed to DenseBase::operator()(const RowIndices&, const ColIndices&). * - * This symbolic placeholder support standard arithmetic operation. + * This symbolic placeholder supports standard arithmetic operations. * It is essentially an alias to last+fix<1>. * * \sa last diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index a7c6f50c3..ce3633388 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -129,16 +129,21 @@ #define EIGEN_COMP_MSVC_STRICT 0 #endif -/// \internal EIGEN_COMP_IBM set to 1 if the compiler is IBM XL C++ -#if defined(__IBMCPP__) || defined(__xlc__) - #define EIGEN_COMP_IBM 1 +/// \internal EIGEN_COMP_IBM set to xlc version if the compiler is IBM XL C++ +// XLC version +// 3.1 0x0301 +// 4.5 0x0405 +// 5.0 0x0500 +// 12.1 0x0C01 +#if defined(__IBMCPP__) || defined(__xlc__) || defined(__ibmxl__) + #define EIGEN_COMP_IBM __xlC__ #else #define EIGEN_COMP_IBM 0 #endif -/// \internal EIGEN_COMP_PGI set to 1 if the compiler is Portland Group Compiler +/// \internal EIGEN_COMP_PGI set to PGI version if the compiler is Portland Group Compiler #if defined(__PGI) - #define EIGEN_COMP_PGI 1 + #define EIGEN_COMP_PGI (__PGIC__*100+__PGIC_MINOR__) #else #define EIGEN_COMP_PGI 0 #endif @@ -347,9 +352,17 @@ #define EIGEN_OS_WIN_STRICT 0 #endif -/// \internal EIGEN_OS_SUN set to 1 if the OS is SUN +/// \internal EIGEN_OS_SUN set to __SUNPRO_C if the OS is SUN +// compiler solaris __SUNPRO_C +// version studio +// 5.7 10 0x570 +// 5.8 11 0x580 +// 5.9 12 0x590 +// 5.10 12.1 0x5100 +// 5.11 12.2 0x5110 +// 5.12 12.3 0x5120 #if (defined(sun) || defined(__sun)) && !(defined(__SVR4) || defined(__svr4__)) - #define EIGEN_OS_SUN 1 + #define EIGEN_OS_SUN __SUNPRO_C #else #define EIGEN_OS_SUN 0 #endif @@ -495,13 +508,33 @@ #define EIGEN_HAS_STATIC_ARRAY_TEMPLATE 0 #endif -#if EIGEN_MAX_CPP_VER>=11 && (defined(__cplusplus) && (__cplusplus >= 201103L) || EIGEN_COMP_MSVC >= 1900) + +// The macro EIGEN_COMP_CXXVER defines the c++ verson expected by the compiler. +// For instance, if compiling with gcc and -std=c++17, then EIGEN_COMP_CXXVER +// is defined to 17. +#if (defined(__cplusplus) && (__cplusplus > 201402L) || EIGEN_COMP_MSVC >= 1914) +#define EIGEN_COMP_CXXVER 17 +#elif (defined(__cplusplus) && (__cplusplus > 201103L) || EIGEN_COMP_MSVC >= 1910) +#define EIGEN_COMP_CXXVER 14 +#elif (defined(__cplusplus) && (__cplusplus >= 201103L) || EIGEN_COMP_MSVC >= 1900) +#define EIGEN_COMP_CXXVER 11 +#else +#define EIGEN_COMP_CXXVER 03 +#endif + + +// The macros EIGEN_HAS_CXX?? defines a rough estimate of available c++ features +// but in practice we should not rely on them but rather on the availabilty of +// individual features as defined later. +// This is why there is no EIGEN_HAS_CXX17. +// FIXME: get rid of EIGEN_HAS_CXX14 and maybe even EIGEN_HAS_CXX11. +#if EIGEN_MAX_CPP_VER>=11 && EIGEN_COMP_CXXVER>=11 #define EIGEN_HAS_CXX11 1 #else #define EIGEN_HAS_CXX11 0 #endif -#if EIGEN_MAX_CPP_VER>=14 && (defined(__cplusplus) && (__cplusplus > 201103L) || EIGEN_COMP_MSVC >= 1910) +#if EIGEN_MAX_CPP_VER>=14 && EIGEN_COMP_CXXVER>=14 #define EIGEN_HAS_CXX14 1 #else #define EIGEN_HAS_CXX14 0 @@ -546,6 +579,22 @@ #endif #endif +#ifndef EIGEN_HAS_ALIGNAS +#if EIGEN_MAX_CPP_VER>=11 && EIGEN_HAS_CXX11 && \ + ( __has_feature(cxx_alignas) \ + || EIGEN_HAS_CXX14 \ + || (EIGEN_COMP_MSVC >= 1800) \ + || (EIGEN_GNUC_AT_LEAST(4,8)) \ + || (EIGEN_COMP_CLANG>=305) \ + || (EIGEN_COMP_ICC>=1500) \ + || (EIGEN_COMP_PGI>=1500) \ + || (EIGEN_COMP_SUNCC>=0x5130)) +#define EIGEN_HAS_ALIGNAS 1 +#else +#define EIGEN_HAS_ALIGNAS 0 +#endif +#endif + // Does the compiler support type_traits? // - full support of type traits was added only to GCC 5.1.0. // - 20150626 corresponds to the last release of 4.x libstdc++ @@ -649,6 +698,23 @@ #endif #endif +// NOTE: the required Apple's clang version is very conservative +// and it could be that XCode 9 works just fine. +// NOTE: the MSVC version is based on https://en.cppreference.com/w/cpp/compiler_support +// and not tested. +#ifndef EIGEN_HAS_CXX17_OVERALIGN +#if EIGEN_MAX_CPP_VER>=17 && EIGEN_COMP_CXXVER>=17 && ( \ + (EIGEN_COMP_MSVC >= 1912) \ + || (EIGEN_GNUC_AT_LEAST(7,0)) \ + || ((!defined(__apple_build_version__)) && (EIGEN_COMP_CLANG>=500)) \ + || (( defined(__apple_build_version__)) && (__apple_build_version__>=10000000)) \ + ) +#define EIGEN_HAS_CXX17_OVERALIGN 1 +#else +#define EIGEN_HAS_CXX17_OVERALIGN 0 +#endif +#endif + #if defined(EIGEN_CUDACC) && EIGEN_HAS_CONSTEXPR // While available already with c++11, this is useful mostly starting with c++14 and relaxed constexpr rules #if defined(__NVCC__) @@ -742,10 +808,6 @@ // All functions callable from CUDA/HIP code must be qualified with __device__ #ifdef EIGEN_GPUCC - #ifndef EIGEN_DONT_VECTORIZE - #define EIGEN_DONT_VECTORIZE - #endif - #define EIGEN_DEVICE_FUNC __host__ __device__ #else #define EIGEN_DEVICE_FUNC @@ -841,7 +903,7 @@ // Suppresses 'unused variable' warnings. namespace Eigen { namespace internal { - template<typename T> EIGEN_DEVICE_FUNC void ignore_unused_variable(const T&) {} + template<typename T> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void ignore_unused_variable(const T&) {} } } #define EIGEN_UNUSED_VARIABLE(var) Eigen::internal::ignore_unused_variable(var); @@ -1035,7 +1097,7 @@ namespace Eigen { #endif #define EIGEN_MAKE_SCALAR_BINARY_OP_ONTHERIGHT(METHOD,OPNAME) \ - template <typename T> EIGEN_DEVICE_FUNC inline \ + template <typename T> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE \ EIGEN_MSVC10_WORKAROUND_BINARYOP_RETURN_TYPE(const EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(Derived,typename internal::promote_scalar_arg<Scalar EIGEN_COMMA T EIGEN_COMMA EIGEN_SCALAR_BINARY_SUPPORTED(OPNAME,Scalar,T)>::type,OPNAME))\ (METHOD)(const T& scalar) const { \ typedef typename internal::promote_scalar_arg<Scalar,T,EIGEN_SCALAR_BINARY_SUPPORTED(OPNAME,Scalar,T)>::type PromotedT; \ @@ -1044,7 +1106,7 @@ namespace Eigen { } #define EIGEN_MAKE_SCALAR_BINARY_OP_ONTHELEFT(METHOD,OPNAME) \ - template <typename T> EIGEN_DEVICE_FUNC inline friend \ + template <typename T> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE friend \ EIGEN_MSVC10_WORKAROUND_BINARYOP_RETURN_TYPE(const EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(typename internal::promote_scalar_arg<Scalar EIGEN_COMMA T EIGEN_COMMA EIGEN_SCALAR_BINARY_SUPPORTED(OPNAME,T,Scalar)>::type,Derived,OPNAME)) \ (METHOD)(const T& scalar, const StorageBaseType& matrix) { \ typedef typename internal::promote_scalar_arg<Scalar,T,EIGEN_SCALAR_BINARY_SUPPORTED(OPNAME,T,Scalar)>::type PromotedT; \ @@ -1093,9 +1155,9 @@ namespace Eigen { # define EIGEN_NOEXCEPT # define EIGEN_NOEXCEPT_IF(x) # define EIGEN_NO_THROW throw() -# if EIGEN_COMP_MSVC +# if EIGEN_COMP_MSVC || EIGEN_COMP_CXXVER>=17 // MSVC does not support exception specifications (warning C4290), - // and they are deprecated in c++11 anyway. + // and they are deprecated in c++11 anyway. This is even an error in c++17. # define EIGEN_EXCEPTION_SPEC(X) throw() # else # define EIGEN_EXCEPTION_SPEC(X) throw(X) diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index 87b538658..1b12544d2 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -360,7 +360,7 @@ template<typename T, bool Align> EIGEN_DEVICE_FUNC inline T* conditional_aligned template<typename T> EIGEN_DEVICE_FUNC inline void aligned_delete(T *ptr, std::size_t size) { destruct_elements_of_array<T>(ptr, size); - aligned_free(ptr); + Eigen::internal::aligned_free(ptr); } /** \internal Deletes objects constructed with conditional_aligned_new @@ -768,6 +768,17 @@ template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b) *** Implementation of EIGEN_MAKE_ALIGNED_OPERATOR_NEW [_IF] *** *****************************************************************************/ +#if EIGEN_HAS_CXX17_OVERALIGN + +// C++17 -> no need to bother about alignment anymore :) + +#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) +#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign) +#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW +#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size) + +#else + #if EIGEN_MAX_ALIGN_BYTES!=0 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \ void* operator new(std::size_t size, const std::nothrow_t&) EIGEN_NO_THROW { \ @@ -810,6 +821,8 @@ template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b) ((EIGEN_MAX_ALIGN_BYTES>=32) && ((sizeof(Scalar)*(Size))%(EIGEN_MAX_ALIGN_BYTES/2)==0)) || \ ((EIGEN_MAX_ALIGN_BYTES>=64) && ((sizeof(Scalar)*(Size))%(EIGEN_MAX_ALIGN_BYTES/4)==0)) ))) +#endif + /****************************************************************************/ /** \class aligned_allocator diff --git a/Eigen/src/Core/util/Meta.h b/Eigen/src/Core/util/Meta.h index 1415b3fc1..8fcb18a94 100755 --- a/Eigen/src/Core/util/Meta.h +++ b/Eigen/src/Core/util/Meta.h @@ -636,8 +636,41 @@ template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool not_equal_strict(const double& x,const double& y) { return std::not_equal_to<double>()(x,y); } #endif +/** \internal extract the bits of the float \a x */ +inline unsigned int as_uint(float x) +{ + unsigned int ret; + std::memcpy(&ret, &x, sizeof(float)); + return ret; +} + } // end namespace numext } // end namespace Eigen +// Define portable (u)int{32,64} types +#if EIGEN_HAS_CXX11 +#include <cstdint> +namespace Eigen { +namespace numext { +typedef std::uint32_t uint32_t; +typedef std::int32_t int32_t; +typedef std::uint64_t uint64_t; +typedef std::int64_t int64_t; +} +} +#else +// Without c++11, all compilers able to compile Eigen also +// provides the C99 stdint.h header file. +#include <stdint.h> +namespace Eigen { +namespace numext { +typedef ::uint32_t uint32_t; +typedef ::int32_t int32_t; +typedef ::uint64_t uint64_t; +typedef ::int64_t int64_t; +} +} +#endif + #endif // EIGEN_META_H diff --git a/Eigen/src/Core/util/StaticAssert.h b/Eigen/src/Core/util/StaticAssert.h index b2f95153e..67714e444 100644 --- a/Eigen/src/Core/util/StaticAssert.h +++ b/Eigen/src/Core/util/StaticAssert.h @@ -104,7 +104,8 @@ STORAGE_INDEX_MUST_MATCH=1, CHOLMOD_SUPPORTS_DOUBLE_PRECISION_ONLY=1, SELFADJOINTVIEW_ACCEPTS_UPPER_AND_LOWER_MODE_ONLY=1, - INVALID_TEMPLATE_PARAMETER=1 + INVALID_TEMPLATE_PARAMETER=1, + GPU_TENSOR_CONTRACTION_DOES_NOT_SUPPORT_OUTPUT_KERNELS=1 }; }; diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h index 836ff4711..91c2e42e4 100644 --- a/Eigen/src/Core/util/XprHelper.h +++ b/Eigen/src/Core/util/XprHelper.h @@ -184,7 +184,8 @@ template<typename T> struct unpacket_traits enum { size = 1, - alignment = 1 + alignment = 1, + vectorizable = false }; }; |