diff options
Diffstat (limited to 'runtimes/nn/depend/external/eigen/Eigen/src/Core')
132 files changed, 50322 insertions, 0 deletions
diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/Array.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Array.h new file mode 100644 index 000000000..e10020d4f --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Array.h @@ -0,0 +1,331 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2009 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_ARRAY_H +#define EIGEN_ARRAY_H + +namespace Eigen { + +namespace internal { +template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols> +struct traits<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> > : traits<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> > +{ + typedef ArrayXpr XprKind; + typedef ArrayBase<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> > XprBase; +}; +} + +/** \class Array + * \ingroup Core_Module + * + * \brief General-purpose arrays with easy API for coefficient-wise operations + * + * The %Array class is very similar to the Matrix class. It provides + * general-purpose one- and two-dimensional arrays. The difference between the + * %Array and the %Matrix class is primarily in the API: the API for the + * %Array class provides easy access to coefficient-wise operations, while the + * API for the %Matrix class provides easy access to linear-algebra + * operations. + * + * See documentation of class Matrix for detailed information on the template parameters + * storage layout. + * + * This class can be extended with the help of the plugin mechanism described on the page + * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_ARRAY_PLUGIN. + * + * \sa \blank \ref TutorialArrayClass, \ref TopicClassHierarchy + */ +template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols> +class Array + : public PlainObjectBase<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> > +{ + public: + + typedef PlainObjectBase<Array> Base; + EIGEN_DENSE_PUBLIC_INTERFACE(Array) + + enum { Options = _Options }; + typedef typename Base::PlainObject PlainObject; + + protected: + template <typename Derived, typename OtherDerived, bool IsVector> + friend struct internal::conservative_resize_like_impl; + + using Base::m_storage; + + public: + + using Base::base; + using Base::coeff; + using Base::coeffRef; + + /** + * The usage of + * using Base::operator=; + * fails on MSVC. Since the code below is working with GCC and MSVC, we skipped + * the usage of 'using'. This should be done only for operator=. + */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Array& operator=(const EigenBase<OtherDerived> &other) + { + return Base::operator=(other); + } + + /** Set all the entries to \a value. + * \sa DenseBase::setConstant(), DenseBase::fill() + */ + /* This overload is needed because the usage of + * using Base::operator=; + * fails on MSVC. Since the code below is working with GCC and MSVC, we skipped + * the usage of 'using'. This should be done only for operator=. + */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Array& operator=(const Scalar &value) + { + Base::setConstant(value); + return *this; + } + + /** Copies the value of the expression \a other into \c *this with automatic resizing. + * + * *this might be resized to match the dimensions of \a other. If *this was a null matrix (not already initialized), + * it will be initialized. + * + * Note that copying a row-vector into a vector (and conversely) is allowed. + * The resizing, if any, is then done in the appropriate way so that row-vectors + * remain row-vectors and vectors remain vectors. + */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Array& operator=(const DenseBase<OtherDerived>& other) + { + return Base::_set(other); + } + + /** This is a special case of the templated operator=. Its purpose is to + * prevent a default operator= from hiding the templated operator=. + */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Array& operator=(const Array& other) + { + return Base::_set(other); + } + + /** Default constructor. + * + * For fixed-size matrices, does nothing. + * + * For dynamic-size matrices, creates an empty matrix of size 0. Does not allocate any array. Such a matrix + * is called a null matrix. This constructor is the unique way to create null matrices: resizing + * a matrix to 0 is not supported. + * + * \sa resize(Index,Index) + */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Array() : Base() + { + Base::_check_template_params(); + EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED + } + +#ifndef EIGEN_PARSED_BY_DOXYGEN + // FIXME is it still needed ?? + /** \internal */ + EIGEN_DEVICE_FUNC + Array(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 + } +#endif + +#if EIGEN_HAS_RVALUE_REFERENCES + EIGEN_DEVICE_FUNC + Array(Array&& 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 + Array& operator=(Array&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_assignable<Scalar>::value) + { + other.swap(*this); + return *this; + } +#endif + + #ifndef EIGEN_PARSED_BY_DOXYGEN + template<typename T> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE explicit Array(const T& x) + { + Base::_check_template_params(); + Base::template _init1<T>(x); + } + + template<typename T0, typename T1> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Array(const T0& val0, const T1& val1) + { + Base::_check_template_params(); + this->template _init2<T0,T1>(val0, val1); + } + #else + /** \brief Constructs a fixed-sized array initialized with coefficients starting at \a data */ + EIGEN_DEVICE_FUNC explicit Array(const Scalar *data); + /** Constructs a vector or row-vector with given dimension. \only_for_vectors + * + * Note that this is only useful for dynamic-size vectors. For fixed-size vectors, + * it is redundant to pass the dimension here, so it makes more sense to use the default + * constructor Array() instead. + */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE explicit Array(Index dim); + /** constructs an initialized 1x1 Array with the given coefficient */ + Array(const Scalar& value); + /** constructs an uninitialized array with \a rows rows and \a cols columns. + * + * This is useful for dynamic-size arrays. For fixed-size arrays, + * 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 */ + Array(const Scalar& val0, const Scalar& val1); + #endif + + /** constructs an initialized 3D vector with given coefficients */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Array(const Scalar& val0, const Scalar& val1, const Scalar& val2) + { + Base::_check_template_params(); + EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(Array, 3) + m_storage.data()[0] = val0; + m_storage.data()[1] = val1; + m_storage.data()[2] = val2; + } + /** constructs an initialized 4D vector with given coefficients */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Array(const Scalar& val0, const Scalar& val1, const Scalar& val2, const Scalar& val3) + { + Base::_check_template_params(); + EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(Array, 4) + m_storage.data()[0] = val0; + m_storage.data()[1] = val1; + m_storage.data()[2] = val2; + m_storage.data()[3] = val3; + } + + /** Copy constructor */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Array(const Array& other) + : Base(other) + { } + + private: + struct PrivateType {}; + public: + + /** \sa MatrixBase::operator=(const EigenBase<OtherDerived>&) */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Array(const EigenBase<OtherDerived> &other, + typename internal::enable_if<internal::is_convertible<typename OtherDerived::Scalar,Scalar>::value, + PrivateType>::type = PrivateType()) + : Base(other.derived()) + { } + + EIGEN_DEVICE_FUNC inline Index innerStride() const { return 1; } + EIGEN_DEVICE_FUNC inline Index outerStride() const { return this->innerSize(); } + + #ifdef EIGEN_ARRAY_PLUGIN + #include EIGEN_ARRAY_PLUGIN + #endif + + private: + + template<typename MatrixType, typename OtherDerived, bool SwapPointers> + friend struct internal::matrix_swap_impl; +}; + +/** \defgroup arraytypedefs Global array typedefs + * \ingroup Core_Module + * + * Eigen defines several typedef shortcuts for most common 1D and 2D array types. + * + * The general patterns are the following: + * + * \c ArrayRowsColsType where \c Rows and \c Cols can be \c 2,\c 3,\c 4 for fixed size square matrices or \c X for dynamic size, + * and where \c Type can be \c i for integer, \c f for float, \c d for double, \c cf for complex float, \c cd + * for complex double. + * + * For example, \c Array33d is a fixed-size 3x3 array type of doubles, and \c ArrayXXf is a dynamic-size matrix of floats. + * + * There are also \c ArraySizeType which are self-explanatory. For example, \c Array4cf is + * a fixed-size 1D array of 4 complex floats. + * + * \sa class Array + */ + +#define EIGEN_MAKE_ARRAY_TYPEDEFS(Type, TypeSuffix, Size, SizeSuffix) \ +/** \ingroup arraytypedefs */ \ +typedef Array<Type, Size, Size> Array##SizeSuffix##SizeSuffix##TypeSuffix; \ +/** \ingroup arraytypedefs */ \ +typedef Array<Type, Size, 1> Array##SizeSuffix##TypeSuffix; + +#define EIGEN_MAKE_ARRAY_FIXED_TYPEDEFS(Type, TypeSuffix, Size) \ +/** \ingroup arraytypedefs */ \ +typedef Array<Type, Size, Dynamic> Array##Size##X##TypeSuffix; \ +/** \ingroup arraytypedefs */ \ +typedef Array<Type, Dynamic, Size> Array##X##Size##TypeSuffix; + +#define EIGEN_MAKE_ARRAY_TYPEDEFS_ALL_SIZES(Type, TypeSuffix) \ +EIGEN_MAKE_ARRAY_TYPEDEFS(Type, TypeSuffix, 2, 2) \ +EIGEN_MAKE_ARRAY_TYPEDEFS(Type, TypeSuffix, 3, 3) \ +EIGEN_MAKE_ARRAY_TYPEDEFS(Type, TypeSuffix, 4, 4) \ +EIGEN_MAKE_ARRAY_TYPEDEFS(Type, TypeSuffix, Dynamic, X) \ +EIGEN_MAKE_ARRAY_FIXED_TYPEDEFS(Type, TypeSuffix, 2) \ +EIGEN_MAKE_ARRAY_FIXED_TYPEDEFS(Type, TypeSuffix, 3) \ +EIGEN_MAKE_ARRAY_FIXED_TYPEDEFS(Type, TypeSuffix, 4) + +EIGEN_MAKE_ARRAY_TYPEDEFS_ALL_SIZES(int, i) +EIGEN_MAKE_ARRAY_TYPEDEFS_ALL_SIZES(float, f) +EIGEN_MAKE_ARRAY_TYPEDEFS_ALL_SIZES(double, d) +EIGEN_MAKE_ARRAY_TYPEDEFS_ALL_SIZES(std::complex<float>, cf) +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_TYPEDEFS_LARGE + +#define EIGEN_USING_ARRAY_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, SizeSuffix) \ +using Eigen::Matrix##SizeSuffix##TypeSuffix; \ +using Eigen::Vector##SizeSuffix##TypeSuffix; \ +using Eigen::RowVector##SizeSuffix##TypeSuffix; + +#define EIGEN_USING_ARRAY_TYPEDEFS_FOR_TYPE(TypeSuffix) \ +EIGEN_USING_ARRAY_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, 2) \ +EIGEN_USING_ARRAY_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, 3) \ +EIGEN_USING_ARRAY_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, 4) \ +EIGEN_USING_ARRAY_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, X) \ + +#define EIGEN_USING_ARRAY_TYPEDEFS \ +EIGEN_USING_ARRAY_TYPEDEFS_FOR_TYPE(i) \ +EIGEN_USING_ARRAY_TYPEDEFS_FOR_TYPE(f) \ +EIGEN_USING_ARRAY_TYPEDEFS_FOR_TYPE(d) \ +EIGEN_USING_ARRAY_TYPEDEFS_FOR_TYPE(cf) \ +EIGEN_USING_ARRAY_TYPEDEFS_FOR_TYPE(cd) + +} // end namespace Eigen + +#endif // EIGEN_ARRAY_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/ArrayBase.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/ArrayBase.h new file mode 100644 index 000000000..3dbc7084c --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/ArrayBase.h @@ -0,0 +1,226 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2009 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_ARRAYBASE_H +#define EIGEN_ARRAYBASE_H + +namespace Eigen { + +template<typename ExpressionType> class MatrixWrapper; + +/** \class ArrayBase + * \ingroup Core_Module + * + * \brief Base class for all 1D and 2D array, and related expressions + * + * An array is similar to a dense vector or matrix. While matrices are mathematical + * objects with well defined linear algebra operators, an array is just a collection + * of scalar values arranged in a one or two dimensionnal fashion. As the main consequence, + * all operations applied to an array are performed coefficient wise. Furthermore, + * arrays support scalar math functions of the c++ standard library (e.g., std::sin(x)), and convenient + * constructors allowing to easily write generic code working for both scalar values + * and arrays. + * + * This class is the base that is inherited by all array expression types. + * + * \tparam Derived is the derived type, e.g., an array or an expression type. + * + * This class can be extended with the help of the plugin mechanism described on the page + * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_ARRAYBASE_PLUGIN. + * + * \sa class MatrixBase, \ref TopicClassHierarchy + */ +template<typename Derived> class ArrayBase + : public DenseBase<Derived> +{ + public: +#ifndef EIGEN_PARSED_BY_DOXYGEN + /** The base class for a given storage type. */ + typedef ArrayBase StorageBaseType; + + typedef ArrayBase Eigen_BaseClassForSpecializationOfGlobalMathFuncImpl; + + typedef typename internal::traits<Derived>::StorageKind StorageKind; + typedef typename internal::traits<Derived>::Scalar Scalar; + typedef typename internal::packet_traits<Scalar>::type PacketScalar; + typedef typename NumTraits<Scalar>::Real RealScalar; + + typedef DenseBase<Derived> Base; + using Base::RowsAtCompileTime; + using Base::ColsAtCompileTime; + using Base::SizeAtCompileTime; + using Base::MaxRowsAtCompileTime; + using Base::MaxColsAtCompileTime; + using Base::MaxSizeAtCompileTime; + using Base::IsVectorAtCompileTime; + using Base::Flags; + + using Base::derived; + using Base::const_cast_derived; + using Base::rows; + using Base::cols; + using Base::size; + using Base::coeff; + using Base::coeffRef; + using Base::lazyAssign; + using Base::operator=; + using Base::operator+=; + using Base::operator-=; + using Base::operator*=; + using Base::operator/=; + + typedef typename Base::CoeffReturnType CoeffReturnType; + +#endif // not EIGEN_PARSED_BY_DOXYGEN + +#ifndef EIGEN_PARSED_BY_DOXYGEN + typedef typename Base::PlainObject PlainObject; + + /** \internal Represents a matrix with all coefficients equal to one another*/ + typedef CwiseNullaryOp<internal::scalar_constant_op<Scalar>,PlainObject> ConstantReturnType; +#endif // not EIGEN_PARSED_BY_DOXYGEN + +#define EIGEN_CURRENT_STORAGE_BASE_CLASS Eigen::ArrayBase +#define EIGEN_DOC_UNARY_ADDONS(X,Y) +# include "../plugins/CommonCwiseUnaryOps.h" +# include "../plugins/MatrixCwiseUnaryOps.h" +# include "../plugins/ArrayCwiseUnaryOps.h" +# include "../plugins/CommonCwiseBinaryOps.h" +# include "../plugins/MatrixCwiseBinaryOps.h" +# include "../plugins/ArrayCwiseBinaryOps.h" +# ifdef EIGEN_ARRAYBASE_PLUGIN +# include EIGEN_ARRAYBASE_PLUGIN +# endif +#undef EIGEN_CURRENT_STORAGE_BASE_CLASS +#undef EIGEN_DOC_UNARY_ADDONS + + /** Special case of the template operator=, in order to prevent the compiler + * from generating a default operator= (issue hit with g++ 4.1) + */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Derived& operator=(const ArrayBase& other) + { + internal::call_assignment(derived(), other.derived()); + return derived(); + } + + /** Set all the entries to \a value. + * \sa DenseBase::setConstant(), DenseBase::fill() */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Derived& operator=(const Scalar &value) + { Base::setConstant(value); return derived(); } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Derived& operator+=(const Scalar& scalar); + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Derived& operator-=(const Scalar& scalar); + + template<typename OtherDerived> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Derived& operator+=(const ArrayBase<OtherDerived>& other); + template<typename OtherDerived> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Derived& operator-=(const ArrayBase<OtherDerived>& other); + + template<typename OtherDerived> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Derived& operator*=(const ArrayBase<OtherDerived>& other); + + template<typename OtherDerived> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Derived& operator/=(const ArrayBase<OtherDerived>& other); + + public: + EIGEN_DEVICE_FUNC + ArrayBase<Derived>& array() { return *this; } + EIGEN_DEVICE_FUNC + const ArrayBase<Derived>& array() const { return *this; } + + /** \returns an \link Eigen::MatrixBase Matrix \endlink expression of this array + * \sa MatrixBase::array() */ + EIGEN_DEVICE_FUNC + MatrixWrapper<Derived> matrix() { return MatrixWrapper<Derived>(derived()); } + EIGEN_DEVICE_FUNC + const MatrixWrapper<const Derived> matrix() const { return MatrixWrapper<const Derived>(derived()); } + +// template<typename Dest> +// inline void evalTo(Dest& dst) const { dst = matrix(); } + + protected: + EIGEN_DEVICE_FUNC + ArrayBase() : Base() {} + + private: + explicit ArrayBase(Index); + ArrayBase(Index,Index); + template<typename OtherDerived> explicit ArrayBase(const ArrayBase<OtherDerived>&); + protected: + // mixing arrays and matrices is not legal + template<typename OtherDerived> Derived& operator+=(const MatrixBase<OtherDerived>& ) + {EIGEN_STATIC_ASSERT(std::ptrdiff_t(sizeof(typename OtherDerived::Scalar))==-1,YOU_CANNOT_MIX_ARRAYS_AND_MATRICES); return *this;} + // mixing arrays and matrices is not legal + template<typename OtherDerived> Derived& operator-=(const MatrixBase<OtherDerived>& ) + {EIGEN_STATIC_ASSERT(std::ptrdiff_t(sizeof(typename OtherDerived::Scalar))==-1,YOU_CANNOT_MIX_ARRAYS_AND_MATRICES); return *this;} +}; + +/** replaces \c *this by \c *this - \a other. + * + * \returns a reference to \c *this + */ +template<typename Derived> +template<typename OtherDerived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived & +ArrayBase<Derived>::operator-=(const ArrayBase<OtherDerived> &other) +{ + call_assignment(derived(), other.derived(), internal::sub_assign_op<Scalar,typename OtherDerived::Scalar>()); + return derived(); +} + +/** replaces \c *this by \c *this + \a other. + * + * \returns a reference to \c *this + */ +template<typename Derived> +template<typename OtherDerived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived & +ArrayBase<Derived>::operator+=(const ArrayBase<OtherDerived>& other) +{ + call_assignment(derived(), other.derived(), internal::add_assign_op<Scalar,typename OtherDerived::Scalar>()); + return derived(); +} + +/** replaces \c *this by \c *this * \a other coefficient wise. + * + * \returns a reference to \c *this + */ +template<typename Derived> +template<typename OtherDerived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived & +ArrayBase<Derived>::operator*=(const ArrayBase<OtherDerived>& other) +{ + call_assignment(derived(), other.derived(), internal::mul_assign_op<Scalar,typename OtherDerived::Scalar>()); + return derived(); +} + +/** replaces \c *this by \c *this / \a other coefficient wise. + * + * \returns a reference to \c *this + */ +template<typename Derived> +template<typename OtherDerived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived & +ArrayBase<Derived>::operator/=(const ArrayBase<OtherDerived>& other) +{ + call_assignment(derived(), other.derived(), internal::div_assign_op<Scalar,typename OtherDerived::Scalar>()); + return derived(); +} + +} // end namespace Eigen + +#endif // EIGEN_ARRAYBASE_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/ArrayWrapper.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/ArrayWrapper.h new file mode 100644 index 000000000..688aadd62 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/ArrayWrapper.h @@ -0,0 +1,209 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2009-2010 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_ARRAYWRAPPER_H +#define EIGEN_ARRAYWRAPPER_H + +namespace Eigen { + +/** \class ArrayWrapper + * \ingroup Core_Module + * + * \brief Expression of a mathematical vector or matrix as an array object + * + * This class is the return type of MatrixBase::array(), and most of the time + * this is the only way it is use. + * + * \sa MatrixBase::array(), class MatrixWrapper + */ + +namespace internal { +template<typename ExpressionType> +struct traits<ArrayWrapper<ExpressionType> > + : public traits<typename remove_all<typename ExpressionType::Nested>::type > +{ + typedef ArrayXpr XprKind; + // Let's remove NestByRefBit + enum { + Flags0 = traits<typename remove_all<typename ExpressionType::Nested>::type >::Flags, + LvalueBitFlag = is_lvalue<ExpressionType>::value ? LvalueBit : 0, + Flags = (Flags0 & ~(NestByRefBit | LvalueBit)) | LvalueBitFlag + }; +}; +} + +template<typename ExpressionType> +class ArrayWrapper : public ArrayBase<ArrayWrapper<ExpressionType> > +{ + public: + typedef ArrayBase<ArrayWrapper> Base; + EIGEN_DENSE_PUBLIC_INTERFACE(ArrayWrapper) + EIGEN_INHERIT_ASSIGNMENT_OPERATORS(ArrayWrapper) + typedef typename internal::remove_all<ExpressionType>::type NestedExpression; + + typedef typename internal::conditional< + internal::is_lvalue<ExpressionType>::value, + Scalar, + const Scalar + >::type ScalarWithConstIfNotLvalue; + + typedef typename internal::ref_selector<ExpressionType>::non_const_type NestedExpressionType; + + using Base::coeffRef; + + EIGEN_DEVICE_FUNC + explicit EIGEN_STRONG_INLINE ArrayWrapper(ExpressionType& matrix) : m_expression(matrix) {} + + 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 ScalarWithConstIfNotLvalue* data() { return m_expression.data(); } + EIGEN_DEVICE_FUNC + inline const Scalar* data() const { return m_expression.data(); } + + EIGEN_DEVICE_FUNC + inline const Scalar& coeffRef(Index rowId, Index colId) const + { + return m_expression.coeffRef(rowId, colId); + } + + EIGEN_DEVICE_FUNC + inline const Scalar& coeffRef(Index index) const + { + return m_expression.coeffRef(index); + } + + template<typename Dest> + EIGEN_DEVICE_FUNC + inline void evalTo(Dest& dst) const { dst = m_expression; } + + const typename internal::remove_all<NestedExpressionType>::type& + EIGEN_DEVICE_FUNC + nestedExpression() const + { + return m_expression; + } + + /** Forwards the resizing request to the nested expression + * \sa DenseBase::resize(Index) */ + EIGEN_DEVICE_FUNC + void resize(Index newSize) { m_expression.resize(newSize); } + /** Forwards the resizing request to the nested expression + * \sa DenseBase::resize(Index,Index)*/ + EIGEN_DEVICE_FUNC + void resize(Index rows, Index cols) { m_expression.resize(rows,cols); } + + protected: + NestedExpressionType m_expression; +}; + +/** \class MatrixWrapper + * \ingroup Core_Module + * + * \brief Expression of an array as a mathematical vector or matrix + * + * This class is the return type of ArrayBase::matrix(), and most of the time + * this is the only way it is use. + * + * \sa MatrixBase::matrix(), class ArrayWrapper + */ + +namespace internal { +template<typename ExpressionType> +struct traits<MatrixWrapper<ExpressionType> > + : public traits<typename remove_all<typename ExpressionType::Nested>::type > +{ + typedef MatrixXpr XprKind; + // Let's remove NestByRefBit + enum { + Flags0 = traits<typename remove_all<typename ExpressionType::Nested>::type >::Flags, + LvalueBitFlag = is_lvalue<ExpressionType>::value ? LvalueBit : 0, + Flags = (Flags0 & ~(NestByRefBit | LvalueBit)) | LvalueBitFlag + }; +}; +} + +template<typename ExpressionType> +class MatrixWrapper : public MatrixBase<MatrixWrapper<ExpressionType> > +{ + public: + typedef MatrixBase<MatrixWrapper<ExpressionType> > Base; + EIGEN_DENSE_PUBLIC_INTERFACE(MatrixWrapper) + EIGEN_INHERIT_ASSIGNMENT_OPERATORS(MatrixWrapper) + typedef typename internal::remove_all<ExpressionType>::type NestedExpression; + + typedef typename internal::conditional< + internal::is_lvalue<ExpressionType>::value, + Scalar, + const Scalar + >::type ScalarWithConstIfNotLvalue; + + typedef typename internal::ref_selector<ExpressionType>::non_const_type NestedExpressionType; + + using Base::coeffRef; + + EIGEN_DEVICE_FUNC + explicit inline MatrixWrapper(ExpressionType& matrix) : m_expression(matrix) {} + + 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 ScalarWithConstIfNotLvalue* data() { return m_expression.data(); } + EIGEN_DEVICE_FUNC + inline const Scalar* data() const { return m_expression.data(); } + + EIGEN_DEVICE_FUNC + inline const Scalar& coeffRef(Index rowId, Index colId) const + { + return m_expression.derived().coeffRef(rowId, colId); + } + + EIGEN_DEVICE_FUNC + inline const Scalar& coeffRef(Index index) const + { + return m_expression.coeffRef(index); + } + + EIGEN_DEVICE_FUNC + const typename internal::remove_all<NestedExpressionType>::type& + nestedExpression() const + { + return m_expression; + } + + /** Forwards the resizing request to the nested expression + * \sa DenseBase::resize(Index) */ + EIGEN_DEVICE_FUNC + void resize(Index newSize) { m_expression.resize(newSize); } + /** Forwards the resizing request to the nested expression + * \sa DenseBase::resize(Index,Index)*/ + EIGEN_DEVICE_FUNC + void resize(Index rows, Index cols) { m_expression.resize(rows,cols); } + + protected: + NestedExpressionType m_expression; +}; + +} // end namespace Eigen + +#endif // EIGEN_ARRAYWRAPPER_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/Assign.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Assign.h new file mode 100644 index 000000000..53806ba33 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Assign.h @@ -0,0 +1,90 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2007 Michael Olbrich <michael.olbrich@gmx.net> +// Copyright (C) 2006-2010 Benoit Jacob <jacob.benoit.1@gmail.com> +// Copyright (C) 2008 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_ASSIGN_H +#define EIGEN_ASSIGN_H + +namespace Eigen { + +template<typename Derived> +template<typename OtherDerived> +EIGEN_STRONG_INLINE Derived& DenseBase<Derived> + ::lazyAssign(const DenseBase<OtherDerived>& other) +{ + enum{ + SameType = internal::is_same<typename Derived::Scalar,typename OtherDerived::Scalar>::value + }; + + EIGEN_STATIC_ASSERT_LVALUE(Derived) + EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Derived,OtherDerived) + EIGEN_STATIC_ASSERT(SameType,YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) + + eigen_assert(rows() == other.rows() && cols() == other.cols()); + internal::call_assignment_no_alias(derived(),other.derived()); + + return derived(); +} + +template<typename Derived> +template<typename OtherDerived> +EIGEN_DEVICE_FUNC +EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator=(const DenseBase<OtherDerived>& other) +{ + internal::call_assignment(derived(), other.derived()); + return derived(); +} + +template<typename Derived> +EIGEN_DEVICE_FUNC +EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator=(const DenseBase& other) +{ + internal::call_assignment(derived(), other.derived()); + return derived(); +} + +template<typename Derived> +EIGEN_DEVICE_FUNC +EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const MatrixBase& other) +{ + internal::call_assignment(derived(), other.derived()); + return derived(); +} + +template<typename Derived> +template <typename OtherDerived> +EIGEN_DEVICE_FUNC +EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const DenseBase<OtherDerived>& other) +{ + internal::call_assignment(derived(), other.derived()); + return derived(); +} + +template<typename Derived> +template <typename OtherDerived> +EIGEN_DEVICE_FUNC +EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const EigenBase<OtherDerived>& other) +{ + internal::call_assignment(derived(), other.derived()); + return derived(); +} + +template<typename Derived> +template<typename OtherDerived> +EIGEN_DEVICE_FUNC +EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const ReturnByValue<OtherDerived>& other) +{ + other.derived().evalTo(derived()); + return derived(); +} + +} // end namespace Eigen + +#endif // EIGEN_ASSIGN_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/AssignEvaluator.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/AssignEvaluator.h new file mode 100644 index 000000000..b0ec7b7ca --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/AssignEvaluator.h @@ -0,0 +1,935 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2011 Benoit Jacob <jacob.benoit.1@gmail.com> +// Copyright (C) 2011-2014 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2011-2012 Jitse Niesen <jitse@maths.leeds.ac.uk> +// +// 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_ASSIGN_EVALUATOR_H +#define EIGEN_ASSIGN_EVALUATOR_H + +namespace Eigen { + +// This implementation is based on Assign.h + +namespace internal { + +/*************************************************************************** +* Part 1 : the logic deciding a strategy for traversal and unrolling * +***************************************************************************/ + +// copy_using_evaluator_traits is based on assign_traits + +template <typename DstEvaluator, typename SrcEvaluator, typename AssignFunc> +struct copy_using_evaluator_traits +{ + typedef typename DstEvaluator::XprType Dst; + typedef typename Dst::Scalar DstScalar; + + enum { + DstFlags = DstEvaluator::Flags, + SrcFlags = SrcEvaluator::Flags + }; + +public: + enum { + DstAlignment = DstEvaluator::Alignment, + SrcAlignment = SrcEvaluator::Alignment, + DstHasDirectAccess = DstFlags & DirectAccessBit, + JointAlignment = EIGEN_PLAIN_ENUM_MIN(DstAlignment,SrcAlignment) + }; + +private: + enum { + InnerSize = int(Dst::IsVectorAtCompileTime) ? int(Dst::SizeAtCompileTime) + : int(DstFlags)&RowMajorBit ? int(Dst::ColsAtCompileTime) + : int(Dst::RowsAtCompileTime), + InnerMaxSize = int(Dst::IsVectorAtCompileTime) ? int(Dst::MaxSizeAtCompileTime) + : int(DstFlags)&RowMajorBit ? int(Dst::MaxColsAtCompileTime) + : int(Dst::MaxRowsAtCompileTime), + OuterStride = int(outer_stride_at_compile_time<Dst>::ret), + MaxSizeAtCompileTime = Dst::SizeAtCompileTime + }; + + // TODO distinguish between linear traversal and inner-traversals + typedef typename find_best_packet<DstScalar,Dst::SizeAtCompileTime>::type LinearPacketType; + typedef typename find_best_packet<DstScalar,InnerSize>::type InnerPacketType; + + enum { + LinearPacketSize = unpacket_traits<LinearPacketType>::size, + InnerPacketSize = unpacket_traits<InnerPacketType>::size + }; + +public: + enum { + LinearRequiredAlignment = unpacket_traits<LinearPacketType>::alignment, + InnerRequiredAlignment = unpacket_traits<InnerPacketType>::alignment + }; + +private: + enum { + DstIsRowMajor = DstFlags&RowMajorBit, + SrcIsRowMajor = SrcFlags&RowMajorBit, + StorageOrdersAgree = (int(DstIsRowMajor) == int(SrcIsRowMajor)), + MightVectorize = bool(StorageOrdersAgree) + && (int(DstFlags) & int(SrcFlags) & ActualPacketAccessBit) + && bool(functor_traits<AssignFunc>::PacketAccess), + MayInnerVectorize = MightVectorize + && int(InnerSize)!=Dynamic && int(InnerSize)%int(InnerPacketSize)==0 + && int(OuterStride)!=Dynamic && int(OuterStride)%int(InnerPacketSize)==0 + && (EIGEN_UNALIGNED_VECTORIZE || int(JointAlignment)>=int(InnerRequiredAlignment)), + MayLinearize = bool(StorageOrdersAgree) && (int(DstFlags) & int(SrcFlags) & LinearAccessBit), + MayLinearVectorize = bool(MightVectorize) && MayLinearize && DstHasDirectAccess + && (EIGEN_UNALIGNED_VECTORIZE || (int(DstAlignment)>=int(LinearRequiredAlignment)) || MaxSizeAtCompileTime == Dynamic), + /* If the destination isn't aligned, we have to do runtime checks and we don't unroll, + so it's only good for large enough sizes. */ + MaySliceVectorize = bool(MightVectorize) && bool(DstHasDirectAccess) + && (int(InnerMaxSize)==Dynamic || int(InnerMaxSize)>=(EIGEN_UNALIGNED_VECTORIZE?InnerPacketSize:(3*InnerPacketSize))) + /* slice vectorization can be slow, so we only want it if the slices are big, which is + indicated by InnerMaxSize rather than InnerSize, think of the case of a dynamic block + in a fixed-size matrix + However, with EIGEN_UNALIGNED_VECTORIZE and unrolling, slice vectorization is still worth it */ + }; + +public: + enum { + Traversal = int(MayLinearVectorize) && (LinearPacketSize>InnerPacketSize) ? int(LinearVectorizedTraversal) + : int(MayInnerVectorize) ? int(InnerVectorizedTraversal) + : int(MayLinearVectorize) ? int(LinearVectorizedTraversal) + : int(MaySliceVectorize) ? int(SliceVectorizedTraversal) + : int(MayLinearize) ? int(LinearTraversal) + : int(DefaultTraversal), + Vectorized = int(Traversal) == InnerVectorizedTraversal + || int(Traversal) == LinearVectorizedTraversal + || int(Traversal) == SliceVectorizedTraversal + }; + + typedef typename conditional<int(Traversal)==LinearVectorizedTraversal, LinearPacketType, InnerPacketType>::type PacketType; + +private: + enum { + ActualPacketSize = int(Traversal)==LinearVectorizedTraversal ? LinearPacketSize + : Vectorized ? InnerPacketSize + : 1, + UnrollingLimit = EIGEN_UNROLLING_LIMIT * ActualPacketSize, + MayUnrollCompletely = int(Dst::SizeAtCompileTime) != Dynamic + && int(Dst::SizeAtCompileTime) * (int(DstEvaluator::CoeffReadCost)+int(SrcEvaluator::CoeffReadCost)) <= int(UnrollingLimit), + MayUnrollInner = int(InnerSize) != Dynamic + && int(InnerSize) * (int(DstEvaluator::CoeffReadCost)+int(SrcEvaluator::CoeffReadCost)) <= int(UnrollingLimit) + }; + +public: + enum { + Unrolling = (int(Traversal) == int(InnerVectorizedTraversal) || int(Traversal) == int(DefaultTraversal)) + ? ( + int(MayUnrollCompletely) ? int(CompleteUnrolling) + : int(MayUnrollInner) ? int(InnerUnrolling) + : int(NoUnrolling) + ) + : int(Traversal) == int(LinearVectorizedTraversal) + ? ( bool(MayUnrollCompletely) && ( EIGEN_UNALIGNED_VECTORIZE || (int(DstAlignment)>=int(LinearRequiredAlignment))) + ? int(CompleteUnrolling) + : int(NoUnrolling) ) + : int(Traversal) == int(LinearTraversal) + ? ( bool(MayUnrollCompletely) ? int(CompleteUnrolling) + : int(NoUnrolling) ) +#if EIGEN_UNALIGNED_VECTORIZE + : int(Traversal) == int(SliceVectorizedTraversal) + ? ( bool(MayUnrollInner) ? int(InnerUnrolling) + : int(NoUnrolling) ) +#endif + : int(NoUnrolling) + }; + +#ifdef EIGEN_DEBUG_ASSIGN + static void debug() + { + std::cerr << "DstXpr: " << typeid(typename DstEvaluator::XprType).name() << std::endl; + std::cerr << "SrcXpr: " << typeid(typename SrcEvaluator::XprType).name() << std::endl; + std::cerr.setf(std::ios::hex, std::ios::basefield); + std::cerr << "DstFlags" << " = " << DstFlags << " (" << demangle_flags(DstFlags) << " )" << std::endl; + std::cerr << "SrcFlags" << " = " << SrcFlags << " (" << demangle_flags(SrcFlags) << " )" << std::endl; + std::cerr.unsetf(std::ios::hex); + EIGEN_DEBUG_VAR(DstAlignment) + EIGEN_DEBUG_VAR(SrcAlignment) + EIGEN_DEBUG_VAR(LinearRequiredAlignment) + EIGEN_DEBUG_VAR(InnerRequiredAlignment) + EIGEN_DEBUG_VAR(JointAlignment) + EIGEN_DEBUG_VAR(InnerSize) + EIGEN_DEBUG_VAR(InnerMaxSize) + EIGEN_DEBUG_VAR(LinearPacketSize) + EIGEN_DEBUG_VAR(InnerPacketSize) + EIGEN_DEBUG_VAR(ActualPacketSize) + EIGEN_DEBUG_VAR(StorageOrdersAgree) + EIGEN_DEBUG_VAR(MightVectorize) + EIGEN_DEBUG_VAR(MayLinearize) + EIGEN_DEBUG_VAR(MayInnerVectorize) + EIGEN_DEBUG_VAR(MayLinearVectorize) + EIGEN_DEBUG_VAR(MaySliceVectorize) + std::cerr << "Traversal" << " = " << Traversal << " (" << demangle_traversal(Traversal) << ")" << std::endl; + EIGEN_DEBUG_VAR(SrcEvaluator::CoeffReadCost) + EIGEN_DEBUG_VAR(UnrollingLimit) + EIGEN_DEBUG_VAR(MayUnrollCompletely) + EIGEN_DEBUG_VAR(MayUnrollInner) + std::cerr << "Unrolling" << " = " << Unrolling << " (" << demangle_unrolling(Unrolling) << ")" << std::endl; + std::cerr << std::endl; + } +#endif +}; + +/*************************************************************************** +* Part 2 : meta-unrollers +***************************************************************************/ + +/************************ +*** Default traversal *** +************************/ + +template<typename Kernel, int Index, int Stop> +struct copy_using_evaluator_DefaultTraversal_CompleteUnrolling +{ + // FIXME: this is not very clean, perhaps this information should be provided by the kernel? + typedef typename Kernel::DstEvaluatorType DstEvaluatorType; + typedef typename DstEvaluatorType::XprType DstXprType; + + enum { + outer = Index / DstXprType::InnerSizeAtCompileTime, + inner = Index % DstXprType::InnerSizeAtCompileTime + }; + + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel) + { + kernel.assignCoeffByOuterInner(outer, inner); + copy_using_evaluator_DefaultTraversal_CompleteUnrolling<Kernel, Index+1, Stop>::run(kernel); + } +}; + +template<typename Kernel, int Stop> +struct copy_using_evaluator_DefaultTraversal_CompleteUnrolling<Kernel, Stop, Stop> +{ + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&) { } +}; + +template<typename Kernel, int Index_, int Stop> +struct copy_using_evaluator_DefaultTraversal_InnerUnrolling +{ + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel, Index outer) + { + kernel.assignCoeffByOuterInner(outer, Index_); + copy_using_evaluator_DefaultTraversal_InnerUnrolling<Kernel, Index_+1, Stop>::run(kernel, outer); + } +}; + +template<typename Kernel, int Stop> +struct copy_using_evaluator_DefaultTraversal_InnerUnrolling<Kernel, Stop, Stop> +{ + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&, Index) { } +}; + +/*********************** +*** Linear traversal *** +***********************/ + +template<typename Kernel, int Index, int Stop> +struct copy_using_evaluator_LinearTraversal_CompleteUnrolling +{ + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel& kernel) + { + kernel.assignCoeff(Index); + copy_using_evaluator_LinearTraversal_CompleteUnrolling<Kernel, Index+1, Stop>::run(kernel); + } +}; + +template<typename Kernel, int Stop> +struct copy_using_evaluator_LinearTraversal_CompleteUnrolling<Kernel, Stop, Stop> +{ + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&) { } +}; + +/************************** +*** Inner vectorization *** +**************************/ + +template<typename Kernel, int Index, int Stop> +struct copy_using_evaluator_innervec_CompleteUnrolling +{ + // FIXME: this is not very clean, perhaps this information should be provided by the kernel? + typedef typename Kernel::DstEvaluatorType DstEvaluatorType; + typedef typename DstEvaluatorType::XprType DstXprType; + typedef typename Kernel::PacketType PacketType; + + enum { + outer = Index / DstXprType::InnerSizeAtCompileTime, + inner = Index % DstXprType::InnerSizeAtCompileTime, + SrcAlignment = Kernel::AssignmentTraits::SrcAlignment, + DstAlignment = Kernel::AssignmentTraits::DstAlignment + }; + + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel) + { + kernel.template assignPacketByOuterInner<DstAlignment, SrcAlignment, PacketType>(outer, inner); + enum { NextIndex = Index + unpacket_traits<PacketType>::size }; + copy_using_evaluator_innervec_CompleteUnrolling<Kernel, NextIndex, Stop>::run(kernel); + } +}; + +template<typename Kernel, int Stop> +struct copy_using_evaluator_innervec_CompleteUnrolling<Kernel, Stop, Stop> +{ + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&) { } +}; + +template<typename Kernel, int Index_, int Stop, int SrcAlignment, int DstAlignment> +struct copy_using_evaluator_innervec_InnerUnrolling +{ + typedef typename Kernel::PacketType PacketType; + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel, Index outer) + { + kernel.template assignPacketByOuterInner<DstAlignment, SrcAlignment, PacketType>(outer, Index_); + enum { NextIndex = Index_ + unpacket_traits<PacketType>::size }; + copy_using_evaluator_innervec_InnerUnrolling<Kernel, NextIndex, Stop, SrcAlignment, DstAlignment>::run(kernel, outer); + } +}; + +template<typename Kernel, int Stop, int SrcAlignment, int DstAlignment> +struct copy_using_evaluator_innervec_InnerUnrolling<Kernel, Stop, Stop, SrcAlignment, DstAlignment> +{ + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &, Index) { } +}; + +/*************************************************************************** +* Part 3 : implementation of all cases +***************************************************************************/ + +// dense_assignment_loop is based on assign_impl + +template<typename Kernel, + int Traversal = Kernel::AssignmentTraits::Traversal, + int Unrolling = Kernel::AssignmentTraits::Unrolling> +struct dense_assignment_loop; + +/************************ +*** Default traversal *** +************************/ + +template<typename Kernel> +struct dense_assignment_loop<Kernel, DefaultTraversal, NoUnrolling> +{ + EIGEN_DEVICE_FUNC static void EIGEN_STRONG_INLINE run(Kernel &kernel) + { + for(Index outer = 0; outer < kernel.outerSize(); ++outer) { + for(Index inner = 0; inner < kernel.innerSize(); ++inner) { + kernel.assignCoeffByOuterInner(outer, inner); + } + } + } +}; + +template<typename Kernel> +struct dense_assignment_loop<Kernel, DefaultTraversal, CompleteUnrolling> +{ + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel) + { + typedef typename Kernel::DstEvaluatorType::XprType DstXprType; + copy_using_evaluator_DefaultTraversal_CompleteUnrolling<Kernel, 0, DstXprType::SizeAtCompileTime>::run(kernel); + } +}; + +template<typename Kernel> +struct dense_assignment_loop<Kernel, DefaultTraversal, InnerUnrolling> +{ + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel) + { + typedef typename Kernel::DstEvaluatorType::XprType DstXprType; + + const Index outerSize = kernel.outerSize(); + for(Index outer = 0; outer < outerSize; ++outer) + copy_using_evaluator_DefaultTraversal_InnerUnrolling<Kernel, 0, DstXprType::InnerSizeAtCompileTime>::run(kernel, outer); + } +}; + +/*************************** +*** Linear vectorization *** +***************************/ + + +// The goal of unaligned_dense_assignment_loop is simply to factorize the handling +// of the non vectorizable beginning and ending parts + +template <bool IsAligned = false> +struct unaligned_dense_assignment_loop +{ + // if IsAligned = true, then do nothing + template <typename Kernel> + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&, Index, Index) {} +}; + +template <> +struct unaligned_dense_assignment_loop<false> +{ + // MSVC must not inline this functions. If it does, it fails to optimize the + // packet access path. + // FIXME check which version exhibits this issue +#if EIGEN_COMP_MSVC + template <typename Kernel> + static EIGEN_DONT_INLINE void run(Kernel &kernel, + Index start, + Index end) +#else + template <typename Kernel> + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel, + Index start, + Index end) +#endif + { + for (Index index = start; index < end; ++index) + kernel.assignCoeff(index); + } +}; + +template<typename Kernel> +struct dense_assignment_loop<Kernel, LinearVectorizedTraversal, NoUnrolling> +{ + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel) + { + const Index size = kernel.size(); + typedef typename Kernel::Scalar Scalar; + typedef typename Kernel::PacketType PacketType; + enum { + requestedAlignment = Kernel::AssignmentTraits::LinearRequiredAlignment, + packetSize = unpacket_traits<PacketType>::size, + dstIsAligned = int(Kernel::AssignmentTraits::DstAlignment)>=int(requestedAlignment), + dstAlignment = packet_traits<Scalar>::AlignedOnScalar ? int(requestedAlignment) + : int(Kernel::AssignmentTraits::DstAlignment), + srcAlignment = Kernel::AssignmentTraits::JointAlignment + }; + const Index alignedStart = dstIsAligned ? 0 : internal::first_aligned<requestedAlignment>(kernel.dstDataPtr(), size); + const Index alignedEnd = alignedStart + ((size-alignedStart)/packetSize)*packetSize; + + unaligned_dense_assignment_loop<dstIsAligned!=0>::run(kernel, 0, alignedStart); + + for(Index index = alignedStart; index < alignedEnd; index += packetSize) + kernel.template assignPacket<dstAlignment, srcAlignment, PacketType>(index); + + unaligned_dense_assignment_loop<>::run(kernel, alignedEnd, size); + } +}; + +template<typename Kernel> +struct dense_assignment_loop<Kernel, LinearVectorizedTraversal, CompleteUnrolling> +{ + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel) + { + typedef typename Kernel::DstEvaluatorType::XprType DstXprType; + typedef typename Kernel::PacketType PacketType; + + enum { size = DstXprType::SizeAtCompileTime, + packetSize =unpacket_traits<PacketType>::size, + alignedSize = (size/packetSize)*packetSize }; + + copy_using_evaluator_innervec_CompleteUnrolling<Kernel, 0, alignedSize>::run(kernel); + copy_using_evaluator_DefaultTraversal_CompleteUnrolling<Kernel, alignedSize, size>::run(kernel); + } +}; + +/************************** +*** Inner vectorization *** +**************************/ + +template<typename Kernel> +struct dense_assignment_loop<Kernel, InnerVectorizedTraversal, NoUnrolling> +{ + typedef typename Kernel::PacketType PacketType; + enum { + SrcAlignment = Kernel::AssignmentTraits::SrcAlignment, + DstAlignment = Kernel::AssignmentTraits::DstAlignment + }; + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel) + { + const Index innerSize = kernel.innerSize(); + const Index outerSize = kernel.outerSize(); + const Index packetSize = unpacket_traits<PacketType>::size; + for(Index outer = 0; outer < outerSize; ++outer) + for(Index inner = 0; inner < innerSize; inner+=packetSize) + kernel.template assignPacketByOuterInner<DstAlignment, SrcAlignment, PacketType>(outer, inner); + } +}; + +template<typename Kernel> +struct dense_assignment_loop<Kernel, InnerVectorizedTraversal, CompleteUnrolling> +{ + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel) + { + typedef typename Kernel::DstEvaluatorType::XprType DstXprType; + copy_using_evaluator_innervec_CompleteUnrolling<Kernel, 0, DstXprType::SizeAtCompileTime>::run(kernel); + } +}; + +template<typename Kernel> +struct dense_assignment_loop<Kernel, InnerVectorizedTraversal, InnerUnrolling> +{ + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel) + { + typedef typename Kernel::DstEvaluatorType::XprType DstXprType; + typedef typename Kernel::AssignmentTraits Traits; + const Index outerSize = kernel.outerSize(); + for(Index outer = 0; outer < outerSize; ++outer) + copy_using_evaluator_innervec_InnerUnrolling<Kernel, 0, DstXprType::InnerSizeAtCompileTime, + Traits::SrcAlignment, Traits::DstAlignment>::run(kernel, outer); + } +}; + +/*********************** +*** Linear traversal *** +***********************/ + +template<typename Kernel> +struct dense_assignment_loop<Kernel, LinearTraversal, NoUnrolling> +{ + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel) + { + const Index size = kernel.size(); + for(Index i = 0; i < size; ++i) + kernel.assignCoeff(i); + } +}; + +template<typename Kernel> +struct dense_assignment_loop<Kernel, LinearTraversal, CompleteUnrolling> +{ + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel) + { + typedef typename Kernel::DstEvaluatorType::XprType DstXprType; + copy_using_evaluator_LinearTraversal_CompleteUnrolling<Kernel, 0, DstXprType::SizeAtCompileTime>::run(kernel); + } +}; + +/************************** +*** Slice vectorization *** +***************************/ + +template<typename Kernel> +struct dense_assignment_loop<Kernel, SliceVectorizedTraversal, NoUnrolling> +{ + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel) + { + typedef typename Kernel::Scalar Scalar; + typedef typename Kernel::PacketType PacketType; + enum { + packetSize = unpacket_traits<PacketType>::size, + requestedAlignment = int(Kernel::AssignmentTraits::InnerRequiredAlignment), + alignable = packet_traits<Scalar>::AlignedOnScalar || int(Kernel::AssignmentTraits::DstAlignment)>=sizeof(Scalar), + dstIsAligned = int(Kernel::AssignmentTraits::DstAlignment)>=int(requestedAlignment), + dstAlignment = alignable ? int(requestedAlignment) + : int(Kernel::AssignmentTraits::DstAlignment) + }; + const Scalar *dst_ptr = kernel.dstDataPtr(); + if((!bool(dstIsAligned)) && (UIntPtr(dst_ptr) % sizeof(Scalar))>0) + { + // the pointer is not aligend-on scalar, so alignment is not possible + return dense_assignment_loop<Kernel,DefaultTraversal,NoUnrolling>::run(kernel); + } + const Index packetAlignedMask = packetSize - 1; + const Index innerSize = kernel.innerSize(); + const Index outerSize = kernel.outerSize(); + const Index alignedStep = alignable ? (packetSize - kernel.outerStride() % packetSize) & packetAlignedMask : 0; + Index alignedStart = ((!alignable) || bool(dstIsAligned)) ? 0 : internal::first_aligned<requestedAlignment>(dst_ptr, innerSize); + + for(Index outer = 0; outer < outerSize; ++outer) + { + const Index alignedEnd = alignedStart + ((innerSize-alignedStart) & ~packetAlignedMask); + // do the non-vectorizable part of the assignment + for(Index inner = 0; inner<alignedStart ; ++inner) + kernel.assignCoeffByOuterInner(outer, inner); + + // do the vectorizable part of the assignment + for(Index inner = alignedStart; inner<alignedEnd; inner+=packetSize) + kernel.template assignPacketByOuterInner<dstAlignment, Unaligned, PacketType>(outer, inner); + + // do the non-vectorizable part of the assignment + for(Index inner = alignedEnd; inner<innerSize ; ++inner) + kernel.assignCoeffByOuterInner(outer, inner); + + alignedStart = numext::mini((alignedStart+alignedStep)%packetSize, innerSize); + } + } +}; + +#if EIGEN_UNALIGNED_VECTORIZE +template<typename Kernel> +struct dense_assignment_loop<Kernel, SliceVectorizedTraversal, InnerUnrolling> +{ + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel) + { + typedef typename Kernel::DstEvaluatorType::XprType DstXprType; + typedef typename Kernel::PacketType PacketType; + + enum { size = DstXprType::InnerSizeAtCompileTime, + packetSize =unpacket_traits<PacketType>::size, + vectorizableSize = (size/packetSize)*packetSize }; + + for(Index outer = 0; outer < kernel.outerSize(); ++outer) + { + copy_using_evaluator_innervec_InnerUnrolling<Kernel, 0, vectorizableSize, 0, 0>::run(kernel, outer); + copy_using_evaluator_DefaultTraversal_InnerUnrolling<Kernel, vectorizableSize, size>::run(kernel, outer); + } + } +}; +#endif + + +/*************************************************************************** +* Part 4 : Generic dense assignment kernel +***************************************************************************/ + +// This class generalize the assignment of a coefficient (or packet) from one dense evaluator +// to another dense writable evaluator. +// It is parametrized by the two evaluators, and the actual assignment functor. +// This abstraction level permits to keep the evaluation loops as simple and as generic as possible. +// One can customize the assignment using this generic dense_assignment_kernel with different +// functors, or by completely overloading it, by-passing a functor. +template<typename DstEvaluatorTypeT, typename SrcEvaluatorTypeT, typename Functor, int Version = Specialized> +class generic_dense_assignment_kernel +{ +protected: + typedef typename DstEvaluatorTypeT::XprType DstXprType; + typedef typename SrcEvaluatorTypeT::XprType SrcXprType; +public: + + typedef DstEvaluatorTypeT DstEvaluatorType; + typedef SrcEvaluatorTypeT SrcEvaluatorType; + typedef typename DstEvaluatorType::Scalar Scalar; + typedef copy_using_evaluator_traits<DstEvaluatorTypeT, SrcEvaluatorTypeT, Functor> AssignmentTraits; + typedef typename AssignmentTraits::PacketType PacketType; + + + EIGEN_DEVICE_FUNC 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 + AssignmentTraits::debug(); + #endif + } + + EIGEN_DEVICE_FUNC Index size() const { return m_dstExpr.size(); } + EIGEN_DEVICE_FUNC Index innerSize() const { return m_dstExpr.innerSize(); } + EIGEN_DEVICE_FUNC Index outerSize() const { return m_dstExpr.outerSize(); } + EIGEN_DEVICE_FUNC Index rows() const { return m_dstExpr.rows(); } + EIGEN_DEVICE_FUNC Index cols() const { return m_dstExpr.cols(); } + EIGEN_DEVICE_FUNC Index outerStride() const { return m_dstExpr.outerStride(); } + + EIGEN_DEVICE_FUNC DstEvaluatorType& dstEvaluator() { return m_dst; } + EIGEN_DEVICE_FUNC const SrcEvaluatorType& srcEvaluator() const { return m_src; } + + /// Assign src(row,col) to dst(row,col) through the assignment functor. + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(Index row, Index col) + { + m_functor.assignCoeff(m_dst.coeffRef(row,col), m_src.coeff(row,col)); + } + + /// \sa assignCoeff(Index,Index) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(Index index) + { + m_functor.assignCoeff(m_dst.coeffRef(index), m_src.coeff(index)); + } + + /// \sa assignCoeff(Index,Index) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeffByOuterInner(Index outer, Index inner) + { + Index row = rowIndexByOuterInner(outer, inner); + Index col = colIndexByOuterInner(outer, inner); + assignCoeff(row, col); + } + + + template<int StoreMode, int LoadMode, typename PacketType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignPacket(Index row, Index col) + { + m_functor.template assignPacket<StoreMode>(&m_dst.coeffRef(row,col), m_src.template packet<LoadMode,PacketType>(row,col)); + } + + template<int StoreMode, int LoadMode, typename PacketType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignPacket(Index index) + { + m_functor.template assignPacket<StoreMode>(&m_dst.coeffRef(index), m_src.template packet<LoadMode,PacketType>(index)); + } + + template<int StoreMode, int LoadMode, typename PacketType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignPacketByOuterInner(Index outer, Index inner) + { + Index row = rowIndexByOuterInner(outer, inner); + Index col = colIndexByOuterInner(outer, inner); + assignPacket<StoreMode,LoadMode,PacketType>(row, col); + } + + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Index rowIndexByOuterInner(Index outer, Index inner) + { + typedef typename DstEvaluatorType::ExpressionTraits Traits; + return int(Traits::RowsAtCompileTime) == 1 ? 0 + : int(Traits::ColsAtCompileTime) == 1 ? inner + : int(DstEvaluatorType::Flags)&RowMajorBit ? outer + : inner; + } + + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Index colIndexByOuterInner(Index outer, Index inner) + { + typedef typename DstEvaluatorType::ExpressionTraits Traits; + return int(Traits::ColsAtCompileTime) == 1 ? 0 + : int(Traits::RowsAtCompileTime) == 1 ? inner + : int(DstEvaluatorType::Flags)&RowMajorBit ? inner + : outer; + } + + EIGEN_DEVICE_FUNC const Scalar* dstDataPtr() const + { + return m_dstExpr.data(); + } + +protected: + DstEvaluatorType& m_dst; + const SrcEvaluatorType& m_src; + const Functor &m_functor; + // TODO find a way to avoid the needs of the original expression + DstXprType& m_dstExpr; +}; + +/*************************************************************************** +* Part 5 : Entry point for dense rectangular assignment +***************************************************************************/ + +template<typename DstXprType,typename SrcXprType, typename Functor> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +void resize_if_allowed(DstXprType &dst, const SrcXprType& src, const Functor &/*func*/) +{ + EIGEN_ONLY_USED_FOR_DEBUG(dst); + EIGEN_ONLY_USED_FOR_DEBUG(src); + eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); +} + +template<typename DstXprType,typename SrcXprType, typename T1, typename T2> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +void resize_if_allowed(DstXprType &dst, const SrcXprType& src, const internal::assign_op<T1,T2> &/*func*/) +{ + Index dstRows = src.rows(); + Index dstCols = src.cols(); + if(((dst.rows()!=dstRows) || (dst.cols()!=dstCols))) + dst.resize(dstRows, dstCols); + eigen_assert(dst.rows() == dstRows && dst.cols() == dstCols); +} + +template<typename DstXprType, typename SrcXprType, typename Functor> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void call_dense_assignment_loop(DstXprType& dst, const SrcXprType& src, const Functor &func) +{ + typedef evaluator<DstXprType> DstEvaluatorType; + typedef evaluator<SrcXprType> SrcEvaluatorType; + + SrcEvaluatorType srcEvaluator(src); + + // NOTE To properly handle A = (A*A.transpose())/s with A rectangular, + // we need to resize the destination after the source evaluator has been created. + resize_if_allowed(dst, src, func); + + DstEvaluatorType dstEvaluator(dst); + + typedef generic_dense_assignment_kernel<DstEvaluatorType,SrcEvaluatorType,Functor> Kernel; + Kernel kernel(dstEvaluator, srcEvaluator, func, dst.const_cast_derived()); + + dense_assignment_loop<Kernel>::run(kernel); +} + +template<typename DstXprType, typename SrcXprType> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void call_dense_assignment_loop(DstXprType& dst, const SrcXprType& src) +{ + call_dense_assignment_loop(dst, src, internal::assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar>()); +} + +/*************************************************************************** +* Part 6 : Generic assignment +***************************************************************************/ + +// Based on the respective shapes of the destination and source, +// the class AssignmentKind determine the kind of assignment mechanism. +// AssignmentKind must define a Kind typedef. +template<typename DstShape, typename SrcShape> struct AssignmentKind; + +// Assignement kind defined in this file: +struct Dense2Dense {}; +struct EigenBase2EigenBase {}; + +template<typename,typename> struct AssignmentKind { typedef EigenBase2EigenBase Kind; }; +template<> struct AssignmentKind<DenseShape,DenseShape> { typedef Dense2Dense Kind; }; + +// This is the main assignment class +template< typename DstXprType, typename SrcXprType, typename Functor, + typename Kind = typename AssignmentKind< typename evaluator_traits<DstXprType>::Shape , typename evaluator_traits<SrcXprType>::Shape >::Kind, + typename EnableIf = void> +struct Assignment; + + +// The only purpose of this call_assignment() function is to deal with noalias() / "assume-aliasing" and automatic transposition. +// Indeed, I (Gael) think that this concept of "assume-aliasing" was a mistake, and it makes thing quite complicated. +// So this intermediate function removes everything related to "assume-aliasing" such that Assignment +// does not has to bother about these annoying details. + +template<typename Dst, typename Src> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +void call_assignment(Dst& dst, const Src& src) +{ + call_assignment(dst, src, internal::assign_op<typename Dst::Scalar,typename Src::Scalar>()); +} +template<typename Dst, typename Src> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +void call_assignment(const Dst& dst, const Src& src) +{ + call_assignment(dst, src, internal::assign_op<typename Dst::Scalar,typename Src::Scalar>()); +} + +// Deal with "assume-aliasing" +template<typename Dst, typename Src, typename Func> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable_if< evaluator_assume_aliasing<Src>::value, void*>::type = 0) +{ + typename plain_matrix_type<Src>::type tmp(src); + call_assignment_no_alias(dst, tmp, func); +} + +template<typename Dst, typename Src, typename Func> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable_if<!evaluator_assume_aliasing<Src>::value, void*>::type = 0) +{ + call_assignment_no_alias(dst, src, func); +} + +// by-pass "assume-aliasing" +// When there is no aliasing, we require that 'dst' has been properly resized +template<typename Dst, template <typename> class StorageBase, typename Src, typename Func> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +void call_assignment(NoAlias<Dst,StorageBase>& dst, const Src& src, const Func& func) +{ + call_assignment_no_alias(dst.expression(), src, func); +} + + +template<typename Dst, typename Src, typename Func> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +void call_assignment_no_alias(Dst& dst, const Src& src, const Func& func) +{ + enum { + NeedToTranspose = ( (int(Dst::RowsAtCompileTime) == 1 && int(Src::ColsAtCompileTime) == 1) + || (int(Dst::ColsAtCompileTime) == 1 && int(Src::RowsAtCompileTime) == 1) + ) && int(Dst::SizeAtCompileTime) != 1 + }; + + typedef typename internal::conditional<NeedToTranspose, Transpose<Dst>, Dst>::type ActualDstTypeCleaned; + typedef typename internal::conditional<NeedToTranspose, Transpose<Dst>, Dst&>::type ActualDstType; + ActualDstType actualDst(dst); + + // TODO check whether this is the right place to perform these checks: + EIGEN_STATIC_ASSERT_LVALUE(Dst) + EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(ActualDstTypeCleaned,Src) + EIGEN_CHECK_BINARY_COMPATIBILIY(Func,typename ActualDstTypeCleaned::Scalar,typename Src::Scalar); + + Assignment<ActualDstTypeCleaned,Src,Func>::run(actualDst, src, func); +} +template<typename Dst, typename Src> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +void call_assignment_no_alias(Dst& dst, const Src& src) +{ + call_assignment_no_alias(dst, src, internal::assign_op<typename Dst::Scalar,typename Src::Scalar>()); +} + +template<typename Dst, typename Src, typename Func> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +void call_assignment_no_alias_no_transpose(Dst& dst, const Src& src, const Func& func) +{ + // TODO check whether this is the right place to perform these checks: + EIGEN_STATIC_ASSERT_LVALUE(Dst) + EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Dst,Src) + EIGEN_CHECK_BINARY_COMPATIBILIY(Func,typename Dst::Scalar,typename Src::Scalar); + + Assignment<Dst,Src,Func>::run(dst, src, func); +} +template<typename Dst, typename Src> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +void call_assignment_no_alias_no_transpose(Dst& dst, const Src& src) +{ + call_assignment_no_alias_no_transpose(dst, src, internal::assign_op<typename Dst::Scalar,typename Src::Scalar>()); +} + +// forward declaration +template<typename Dst, typename Src> void check_for_aliasing(const Dst &dst, const Src &src); + +// Generic Dense to Dense assignment +// Note that the last template argument "Weak" is needed to make it possible to perform +// both partial specialization+SFINAE without ambiguous specialization +template< typename DstXprType, typename SrcXprType, typename Functor, typename Weak> +struct Assignment<DstXprType, SrcXprType, Functor, Dense2Dense, Weak> +{ + EIGEN_DEVICE_FUNC + static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const Functor &func) + { +#ifndef EIGEN_NO_DEBUG + internal::check_for_aliasing(dst, src); +#endif + + call_dense_assignment_loop(dst, src, func); + } +}; + +// Generic assignment through evalTo. +// TODO: not sure we have to keep that one, but it helps porting current code to new evaluator mechanism. +// Note that the last template argument "Weak" is needed to make it possible to perform +// both partial specialization+SFINAE without ambiguous specialization +template< typename DstXprType, typename SrcXprType, typename Functor, typename Weak> +struct Assignment<DstXprType, SrcXprType, Functor, EigenBase2EigenBase, Weak> +{ + EIGEN_DEVICE_FUNC + static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &/*func*/) + { + Index dstRows = src.rows(); + Index dstCols = src.cols(); + if((dst.rows()!=dstRows) || (dst.cols()!=dstCols)) + dst.resize(dstRows, dstCols); + + eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); + src.evalTo(dst); + } + + // NOTE The following two functions are templated to avoid their instanciation if not needed + // This is needed because some expressions supports evalTo only and/or have 'void' as scalar type. + template<typename SrcScalarType> + EIGEN_DEVICE_FUNC + static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op<typename DstXprType::Scalar,SrcScalarType> &/*func*/) + { + Index dstRows = src.rows(); + Index dstCols = src.cols(); + if((dst.rows()!=dstRows) || (dst.cols()!=dstCols)) + dst.resize(dstRows, dstCols); + + eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); + src.addTo(dst); + } + + template<typename SrcScalarType> + EIGEN_DEVICE_FUNC + static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op<typename DstXprType::Scalar,SrcScalarType> &/*func*/) + { + Index dstRows = src.rows(); + Index dstCols = src.cols(); + if((dst.rows()!=dstRows) || (dst.cols()!=dstCols)) + dst.resize(dstRows, dstCols); + + eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); + src.subTo(dst); + } +}; + +} // namespace internal + +} // end namespace Eigen + +#endif // EIGEN_ASSIGN_EVALUATOR_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/Assign_MKL.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Assign_MKL.h new file mode 100755 index 000000000..6c2ab9264 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Assign_MKL.h @@ -0,0 +1,176 @@ +/* + Copyright (c) 2011, Intel Corporation. All rights reserved. + Copyright (C) 2015 Gael Guennebaud <gael.guennebaud@inria.fr> + + Redistribution and use in source and binary forms, with or without modification, + are permitted provided that the following conditions are met: + + * Redistributions of source code must retain the above copyright notice, this + list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above copyright notice, + this list of conditions and the following disclaimer in the documentation + and/or other materials provided with the distribution. + * Neither the name of Intel Corporation nor the names of its contributors may + be used to endorse or promote products derived from this software without + specific prior written permission. + + THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND + ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED + WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE + DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR + ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES + (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; + LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON + ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS + SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + + ******************************************************************************** + * Content : Eigen bindings to Intel(R) MKL + * MKL VML support for coefficient-wise unary Eigen expressions like a=b.sin() + ******************************************************************************** +*/ + +#ifndef EIGEN_ASSIGN_VML_H +#define EIGEN_ASSIGN_VML_H + +namespace Eigen { + +namespace internal { + +template<typename Dst, typename Src> +class vml_assign_traits +{ + private: + enum { + DstHasDirectAccess = Dst::Flags & DirectAccessBit, + SrcHasDirectAccess = Src::Flags & DirectAccessBit, + StorageOrdersAgree = (int(Dst::IsRowMajor) == int(Src::IsRowMajor)), + InnerSize = int(Dst::IsVectorAtCompileTime) ? int(Dst::SizeAtCompileTime) + : int(Dst::Flags)&RowMajorBit ? int(Dst::ColsAtCompileTime) + : int(Dst::RowsAtCompileTime), + InnerMaxSize = int(Dst::IsVectorAtCompileTime) ? int(Dst::MaxSizeAtCompileTime) + : int(Dst::Flags)&RowMajorBit ? int(Dst::MaxColsAtCompileTime) + : int(Dst::MaxRowsAtCompileTime), + MaxSizeAtCompileTime = Dst::SizeAtCompileTime, + + MightEnableVml = StorageOrdersAgree && DstHasDirectAccess && SrcHasDirectAccess && Src::InnerStrideAtCompileTime==1 && Dst::InnerStrideAtCompileTime==1, + MightLinearize = MightEnableVml && (int(Dst::Flags) & int(Src::Flags) & LinearAccessBit), + VmlSize = MightLinearize ? MaxSizeAtCompileTime : InnerMaxSize, + LargeEnough = VmlSize==Dynamic || VmlSize>=EIGEN_MKL_VML_THRESHOLD + }; + public: + enum { + EnableVml = MightEnableVml && LargeEnough, + Traversal = MightLinearize ? LinearTraversal : DefaultTraversal + }; +}; + +#define EIGEN_PP_EXPAND(ARG) ARG +#if !defined (EIGEN_FAST_MATH) || (EIGEN_FAST_MATH != 1) +#define EIGEN_VMLMODE_EXPAND_LA , VML_HA +#else +#define EIGEN_VMLMODE_EXPAND_LA , VML_LA +#endif + +#define EIGEN_VMLMODE_EXPAND__ + +#define EIGEN_VMLMODE_PREFIX_LA vm +#define EIGEN_VMLMODE_PREFIX__ v +#define EIGEN_VMLMODE_PREFIX(VMLMODE) EIGEN_CAT(EIGEN_VMLMODE_PREFIX_,VMLMODE) + +#define EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE, VMLMODE) \ + template< typename DstXprType, typename SrcXprNested> \ + struct Assignment<DstXprType, CwiseUnaryOp<scalar_##EIGENOP##_op<EIGENTYPE>, SrcXprNested>, assign_op<EIGENTYPE,EIGENTYPE>, \ + Dense2Dense, typename enable_if<vml_assign_traits<DstXprType,SrcXprNested>::EnableVml>::type> { \ + typedef CwiseUnaryOp<scalar_##EIGENOP##_op<EIGENTYPE>, SrcXprNested> SrcXprType; \ + static void run(DstXprType &dst, const SrcXprType &src, const assign_op<EIGENTYPE,EIGENTYPE> &/*func*/) { \ + 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) ); \ + } else { \ + const Index outerSize = dst.outerSize(); \ + for(Index outer = 0; outer < outerSize; ++outer) { \ + const EIGENTYPE *src_ptr = src.IsRowMajor ? &(src.nestedExpression().coeffRef(outer,0)) : \ + &(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)); \ + } \ + } \ + } \ + }; \ + + +#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(EIGENOP, VMLOP, VMLMODE) \ + EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, EIGEN_CAT(EIGEN_VMLMODE_PREFIX(VMLMODE),s##VMLOP), float, float, VMLMODE) \ + EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, EIGEN_CAT(EIGEN_VMLMODE_PREFIX(VMLMODE),d##VMLOP), double, double, VMLMODE) + +#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_CPLX(EIGENOP, VMLOP, VMLMODE) \ + EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, EIGEN_CAT(EIGEN_VMLMODE_PREFIX(VMLMODE),c##VMLOP), scomplex, MKL_Complex8, VMLMODE) \ + EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, EIGEN_CAT(EIGEN_VMLMODE_PREFIX(VMLMODE),z##VMLOP), dcomplex, MKL_Complex16, VMLMODE) + +#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS(EIGENOP, VMLOP, VMLMODE) \ + EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(EIGENOP, VMLOP, VMLMODE) \ + EIGEN_MKL_VML_DECLARE_UNARY_CALLS_CPLX(EIGENOP, VMLOP, VMLMODE) + + +EIGEN_MKL_VML_DECLARE_UNARY_CALLS(sin, Sin, LA) +EIGEN_MKL_VML_DECLARE_UNARY_CALLS(asin, Asin, LA) +EIGEN_MKL_VML_DECLARE_UNARY_CALLS(sinh, Sinh, LA) +EIGEN_MKL_VML_DECLARE_UNARY_CALLS(cos, Cos, LA) +EIGEN_MKL_VML_DECLARE_UNARY_CALLS(acos, Acos, LA) +EIGEN_MKL_VML_DECLARE_UNARY_CALLS(cosh, Cosh, LA) +EIGEN_MKL_VML_DECLARE_UNARY_CALLS(tan, Tan, LA) +EIGEN_MKL_VML_DECLARE_UNARY_CALLS(atan, Atan, LA) +EIGEN_MKL_VML_DECLARE_UNARY_CALLS(tanh, Tanh, LA) +// EIGEN_MKL_VML_DECLARE_UNARY_CALLS(abs, Abs, _) +EIGEN_MKL_VML_DECLARE_UNARY_CALLS(exp, Exp, LA) +EIGEN_MKL_VML_DECLARE_UNARY_CALLS(log, Ln, LA) +EIGEN_MKL_VML_DECLARE_UNARY_CALLS(log10, Log10, LA) +EIGEN_MKL_VML_DECLARE_UNARY_CALLS(sqrt, Sqrt, _) + +EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(square, Sqr, _) +EIGEN_MKL_VML_DECLARE_UNARY_CALLS_CPLX(arg, Arg, _) +EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(round, Round, _) +EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(floor, Floor, _) +EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(ceil, Ceil, _) + +#define EIGEN_MKL_VML_DECLARE_POW_CALL(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE, VMLMODE) \ + template< typename DstXprType, typename SrcXprNested, typename Plain> \ + struct Assignment<DstXprType, CwiseBinaryOp<scalar_##EIGENOP##_op<EIGENTYPE,EIGENTYPE>, SrcXprNested, \ + const CwiseNullaryOp<internal::scalar_constant_op<EIGENTYPE>,Plain> >, assign_op<EIGENTYPE,EIGENTYPE>, \ + Dense2Dense, typename enable_if<vml_assign_traits<DstXprType,SrcXprNested>::EnableVml>::type> { \ + typedef CwiseBinaryOp<scalar_##EIGENOP##_op<EIGENTYPE,EIGENTYPE>, SrcXprNested, \ + const CwiseNullaryOp<internal::scalar_constant_op<EIGENTYPE>,Plain> > SrcXprType; \ + static void run(DstXprType &dst, const SrcXprType &src, const assign_op<EIGENTYPE,EIGENTYPE> &/*func*/) { \ + eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); \ + VMLTYPE exponent = reinterpret_cast<const VMLTYPE&>(src.rhs().functor().m_other); \ + 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) ); \ + } else { \ + const Index outerSize = dst.outerSize(); \ + for(Index outer = 0; outer < outerSize; ++outer) { \ + const EIGENTYPE *src_ptr = src.IsRowMajor ? &(src.lhs().coeffRef(outer,0)) : \ + &(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)); \ + } \ + } \ + } \ + }; + +EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmsPowx, float, float, LA) +EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmdPowx, double, double, LA) +EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmcPowx, scomplex, MKL_Complex8, LA) +EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmzPowx, dcomplex, MKL_Complex16, LA) + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_ASSIGN_VML_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/BandMatrix.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/BandMatrix.h new file mode 100644 index 000000000..4978c9140 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/BandMatrix.h @@ -0,0 +1,353 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2009 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_BANDMATRIX_H +#define EIGEN_BANDMATRIX_H + +namespace Eigen { + +namespace internal { + +template<typename Derived> +class BandMatrixBase : public EigenBase<Derived> +{ + public: + + enum { + Flags = internal::traits<Derived>::Flags, + CoeffReadCost = internal::traits<Derived>::CoeffReadCost, + RowsAtCompileTime = internal::traits<Derived>::RowsAtCompileTime, + ColsAtCompileTime = internal::traits<Derived>::ColsAtCompileTime, + MaxRowsAtCompileTime = internal::traits<Derived>::MaxRowsAtCompileTime, + MaxColsAtCompileTime = internal::traits<Derived>::MaxColsAtCompileTime, + Supers = internal::traits<Derived>::Supers, + Subs = internal::traits<Derived>::Subs, + Options = internal::traits<Derived>::Options + }; + typedef typename internal::traits<Derived>::Scalar Scalar; + typedef Matrix<Scalar,RowsAtCompileTime,ColsAtCompileTime> DenseMatrixType; + typedef typename DenseMatrixType::StorageIndex StorageIndex; + typedef typename internal::traits<Derived>::CoefficientsType CoefficientsType; + typedef EigenBase<Derived> Base; + + protected: + enum { + DataRowsAtCompileTime = ((Supers!=Dynamic) && (Subs!=Dynamic)) + ? 1 + Supers + Subs + : Dynamic, + SizeAtCompileTime = EIGEN_SIZE_MIN_PREFER_DYNAMIC(RowsAtCompileTime,ColsAtCompileTime) + }; + + public: + + using Base::derived; + using Base::rows; + using Base::cols; + + /** \returns the number of super diagonals */ + inline Index supers() const { return derived().supers(); } + + /** \returns the number of sub diagonals */ + inline Index subs() const { return derived().subs(); } + + /** \returns an expression of the underlying coefficient matrix */ + inline const CoefficientsType& coeffs() const { return derived().coeffs(); } + + /** \returns an expression of the underlying coefficient matrix */ + inline CoefficientsType& coeffs() { return derived().coeffs(); } + + /** \returns a vector expression of the \a i -th column, + * only the meaningful part is returned. + * \warning the internal storage must be column major. */ + inline Block<CoefficientsType,Dynamic,1> col(Index i) + { + EIGEN_STATIC_ASSERT((Options&RowMajor)==0,THIS_METHOD_IS_ONLY_FOR_COLUMN_MAJOR_MATRICES); + Index start = 0; + Index len = coeffs().rows(); + if (i<=supers()) + { + start = supers()-i; + len = (std::min)(rows(),std::max<Index>(0,coeffs().rows() - (supers()-i))); + } + else if (i>=rows()-subs()) + len = std::max<Index>(0,coeffs().rows() - (i + 1 - rows() + subs())); + return Block<CoefficientsType,Dynamic,1>(coeffs(), start, i, len, 1); + } + + /** \returns a vector expression of the main diagonal */ + inline Block<CoefficientsType,1,SizeAtCompileTime> diagonal() + { return Block<CoefficientsType,1,SizeAtCompileTime>(coeffs(),supers(),0,1,(std::min)(rows(),cols())); } + + /** \returns a vector expression of the main diagonal (const version) */ + inline const Block<const CoefficientsType,1,SizeAtCompileTime> diagonal() const + { return Block<const CoefficientsType,1,SizeAtCompileTime>(coeffs(),supers(),0,1,(std::min)(rows(),cols())); } + + template<int Index> struct DiagonalIntReturnType { + enum { + ReturnOpposite = (Options&SelfAdjoint) && (((Index)>0 && Supers==0) || ((Index)<0 && Subs==0)), + Conjugate = ReturnOpposite && NumTraits<Scalar>::IsComplex, + ActualIndex = ReturnOpposite ? -Index : Index, + DiagonalSize = (RowsAtCompileTime==Dynamic || ColsAtCompileTime==Dynamic) + ? Dynamic + : (ActualIndex<0 + ? EIGEN_SIZE_MIN_PREFER_DYNAMIC(ColsAtCompileTime, RowsAtCompileTime + ActualIndex) + : EIGEN_SIZE_MIN_PREFER_DYNAMIC(RowsAtCompileTime, ColsAtCompileTime - ActualIndex)) + }; + typedef Block<CoefficientsType,1, DiagonalSize> BuildType; + typedef typename internal::conditional<Conjugate, + CwiseUnaryOp<internal::scalar_conjugate_op<Scalar>,BuildType >, + BuildType>::type Type; + }; + + /** \returns a vector expression of the \a N -th sub or super diagonal */ + template<int N> inline typename DiagonalIntReturnType<N>::Type diagonal() + { + return typename DiagonalIntReturnType<N>::BuildType(coeffs(), supers()-N, (std::max)(0,N), 1, diagonalLength(N)); + } + + /** \returns a vector expression of the \a N -th sub or super diagonal */ + template<int N> inline const typename DiagonalIntReturnType<N>::Type diagonal() const + { + return typename DiagonalIntReturnType<N>::BuildType(coeffs(), supers()-N, (std::max)(0,N), 1, diagonalLength(N)); + } + + /** \returns a vector expression of the \a i -th sub or super diagonal */ + inline Block<CoefficientsType,1,Dynamic> diagonal(Index i) + { + eigen_assert((i<0 && -i<=subs()) || (i>=0 && i<=supers())); + return Block<CoefficientsType,1,Dynamic>(coeffs(), supers()-i, std::max<Index>(0,i), 1, diagonalLength(i)); + } + + /** \returns a vector expression of the \a i -th sub or super diagonal */ + inline const Block<const CoefficientsType,1,Dynamic> diagonal(Index i) const + { + eigen_assert((i<0 && -i<=subs()) || (i>=0 && i<=supers())); + return Block<const CoefficientsType,1,Dynamic>(coeffs(), supers()-i, std::max<Index>(0,i), 1, diagonalLength(i)); + } + + template<typename Dest> inline void evalTo(Dest& dst) const + { + dst.resize(rows(),cols()); + dst.setZero(); + dst.diagonal() = diagonal(); + for (Index i=1; i<=supers();++i) + dst.diagonal(i) = diagonal(i); + for (Index i=1; i<=subs();++i) + dst.diagonal(-i) = diagonal(-i); + } + + DenseMatrixType toDenseMatrix() const + { + DenseMatrixType res(rows(),cols()); + evalTo(res); + return res; + } + + protected: + + inline Index diagonalLength(Index i) const + { return i<0 ? (std::min)(cols(),rows()+i) : (std::min)(rows(),cols()-i); } +}; + +/** + * \class BandMatrix + * \ingroup Core_Module + * + * \brief Represents a rectangular matrix with a banded storage + * + * \tparam _Scalar Numeric type, i.e. float, double, int + * \tparam _Rows Number of rows, or \b Dynamic + * \tparam _Cols Number of columns, or \b Dynamic + * \tparam _Supers Number of super diagonal + * \tparam _Subs Number of sub diagonal + * \tparam _Options A combination of either \b #RowMajor or \b #ColMajor, and of \b #SelfAdjoint + * The former controls \ref TopicStorageOrders "storage order", and defaults to + * column-major. The latter controls whether the matrix represents a selfadjoint + * matrix in which case either Supers of Subs have to be null. + * + * \sa class TridiagonalMatrix + */ + +template<typename _Scalar, int _Rows, int _Cols, int _Supers, int _Subs, int _Options> +struct traits<BandMatrix<_Scalar,_Rows,_Cols,_Supers,_Subs,_Options> > +{ + typedef _Scalar Scalar; + typedef Dense StorageKind; + typedef Eigen::Index StorageIndex; + enum { + CoeffReadCost = NumTraits<Scalar>::ReadCost, + RowsAtCompileTime = _Rows, + ColsAtCompileTime = _Cols, + MaxRowsAtCompileTime = _Rows, + MaxColsAtCompileTime = _Cols, + Flags = LvalueBit, + Supers = _Supers, + Subs = _Subs, + Options = _Options, + DataRowsAtCompileTime = ((Supers!=Dynamic) && (Subs!=Dynamic)) ? 1 + Supers + Subs : Dynamic + }; + typedef Matrix<Scalar,DataRowsAtCompileTime,ColsAtCompileTime,Options&RowMajor?RowMajor:ColMajor> CoefficientsType; +}; + +template<typename _Scalar, int Rows, int Cols, int Supers, int Subs, int Options> +class BandMatrix : public BandMatrixBase<BandMatrix<_Scalar,Rows,Cols,Supers,Subs,Options> > +{ + public: + + typedef typename internal::traits<BandMatrix>::Scalar Scalar; + typedef typename internal::traits<BandMatrix>::StorageIndex StorageIndex; + typedef typename internal::traits<BandMatrix>::CoefficientsType CoefficientsType; + + explicit inline BandMatrix(Index rows=Rows, Index cols=Cols, Index supers=Supers, Index subs=Subs) + : m_coeffs(1+supers+subs,cols), + m_rows(rows), m_supers(supers), m_subs(subs) + { + } + + /** \returns the number of columns */ + inline Index rows() const { return m_rows.value(); } + + /** \returns the number of rows */ + inline Index cols() const { return m_coeffs.cols(); } + + /** \returns the number of super diagonals */ + inline Index supers() const { return m_supers.value(); } + + /** \returns the number of sub diagonals */ + inline Index subs() const { return m_subs.value(); } + + inline const CoefficientsType& coeffs() const { return m_coeffs; } + inline CoefficientsType& coeffs() { return m_coeffs; } + + protected: + + CoefficientsType m_coeffs; + internal::variable_if_dynamic<Index, Rows> m_rows; + internal::variable_if_dynamic<Index, Supers> m_supers; + internal::variable_if_dynamic<Index, Subs> m_subs; +}; + +template<typename _CoefficientsType,int _Rows, int _Cols, int _Supers, int _Subs,int _Options> +class BandMatrixWrapper; + +template<typename _CoefficientsType,int _Rows, int _Cols, int _Supers, int _Subs,int _Options> +struct traits<BandMatrixWrapper<_CoefficientsType,_Rows,_Cols,_Supers,_Subs,_Options> > +{ + typedef typename _CoefficientsType::Scalar Scalar; + typedef typename _CoefficientsType::StorageKind StorageKind; + typedef typename _CoefficientsType::StorageIndex StorageIndex; + enum { + CoeffReadCost = internal::traits<_CoefficientsType>::CoeffReadCost, + RowsAtCompileTime = _Rows, + ColsAtCompileTime = _Cols, + MaxRowsAtCompileTime = _Rows, + MaxColsAtCompileTime = _Cols, + Flags = LvalueBit, + Supers = _Supers, + Subs = _Subs, + Options = _Options, + DataRowsAtCompileTime = ((Supers!=Dynamic) && (Subs!=Dynamic)) ? 1 + Supers + Subs : Dynamic + }; + typedef _CoefficientsType CoefficientsType; +}; + +template<typename _CoefficientsType,int _Rows, int _Cols, int _Supers, int _Subs,int _Options> +class BandMatrixWrapper : public BandMatrixBase<BandMatrixWrapper<_CoefficientsType,_Rows,_Cols,_Supers,_Subs,_Options> > +{ + public: + + typedef typename internal::traits<BandMatrixWrapper>::Scalar Scalar; + typedef typename internal::traits<BandMatrixWrapper>::CoefficientsType CoefficientsType; + typedef typename internal::traits<BandMatrixWrapper>::StorageIndex StorageIndex; + + explicit inline BandMatrixWrapper(const CoefficientsType& coeffs, Index rows=_Rows, Index cols=_Cols, Index supers=_Supers, Index subs=_Subs) + : m_coeffs(coeffs), + m_rows(rows), m_supers(supers), m_subs(subs) + { + EIGEN_UNUSED_VARIABLE(cols); + //internal::assert(coeffs.cols()==cols() && (supers()+subs()+1)==coeffs.rows()); + } + + /** \returns the number of columns */ + inline Index rows() const { return m_rows.value(); } + + /** \returns the number of rows */ + inline Index cols() const { return m_coeffs.cols(); } + + /** \returns the number of super diagonals */ + inline Index supers() const { return m_supers.value(); } + + /** \returns the number of sub diagonals */ + inline Index subs() const { return m_subs.value(); } + + inline const CoefficientsType& coeffs() const { return m_coeffs; } + + protected: + + const CoefficientsType& m_coeffs; + internal::variable_if_dynamic<Index, _Rows> m_rows; + internal::variable_if_dynamic<Index, _Supers> m_supers; + internal::variable_if_dynamic<Index, _Subs> m_subs; +}; + +/** + * \class TridiagonalMatrix + * \ingroup Core_Module + * + * \brief Represents a tridiagonal matrix with a compact banded storage + * + * \tparam Scalar Numeric type, i.e. float, double, int + * \tparam Size Number of rows and cols, or \b Dynamic + * \tparam Options Can be 0 or \b SelfAdjoint + * + * \sa class BandMatrix + */ +template<typename Scalar, int Size, int Options> +class TridiagonalMatrix : public BandMatrix<Scalar,Size,Size,Options&SelfAdjoint?0:1,1,Options|RowMajor> +{ + typedef BandMatrix<Scalar,Size,Size,Options&SelfAdjoint?0:1,1,Options|RowMajor> Base; + typedef typename Base::StorageIndex StorageIndex; + public: + explicit TridiagonalMatrix(Index size = Size) : Base(size,size,Options&SelfAdjoint?0:1,1) {} + + inline typename Base::template DiagonalIntReturnType<1>::Type super() + { return Base::template diagonal<1>(); } + inline const typename Base::template DiagonalIntReturnType<1>::Type super() const + { return Base::template diagonal<1>(); } + inline typename Base::template DiagonalIntReturnType<-1>::Type sub() + { return Base::template diagonal<-1>(); } + inline const typename Base::template DiagonalIntReturnType<-1>::Type sub() const + { return Base::template diagonal<-1>(); } + protected: +}; + + +struct BandShape {}; + +template<typename _Scalar, int _Rows, int _Cols, int _Supers, int _Subs, int _Options> +struct evaluator_traits<BandMatrix<_Scalar,_Rows,_Cols,_Supers,_Subs,_Options> > + : public evaluator_traits_base<BandMatrix<_Scalar,_Rows,_Cols,_Supers,_Subs,_Options> > +{ + typedef BandShape Shape; +}; + +template<typename _CoefficientsType,int _Rows, int _Cols, int _Supers, int _Subs,int _Options> +struct evaluator_traits<BandMatrixWrapper<_CoefficientsType,_Rows,_Cols,_Supers,_Subs,_Options> > + : public evaluator_traits_base<BandMatrixWrapper<_CoefficientsType,_Rows,_Cols,_Supers,_Subs,_Options> > +{ + typedef BandShape Shape; +}; + +template<> struct AssignmentKind<DenseShape,BandShape> { typedef EigenBase2EigenBase Kind; }; + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_BANDMATRIX_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/Block.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Block.h new file mode 100644 index 000000000..11de45c2e --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Block.h @@ -0,0 +1,452 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2006-2010 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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_BLOCK_H +#define EIGEN_BLOCK_H + +namespace Eigen { + +namespace internal { +template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel> +struct traits<Block<XprType, BlockRows, BlockCols, InnerPanel> > : traits<XprType> +{ + typedef typename traits<XprType>::Scalar Scalar; + typedef typename traits<XprType>::StorageKind StorageKind; + typedef typename traits<XprType>::XprKind XprKind; + typedef typename ref_selector<XprType>::type XprTypeNested; + typedef typename remove_reference<XprTypeNested>::type _XprTypeNested; + enum{ + MatrixRows = traits<XprType>::RowsAtCompileTime, + MatrixCols = traits<XprType>::ColsAtCompileTime, + RowsAtCompileTime = MatrixRows == 0 ? 0 : BlockRows, + ColsAtCompileTime = MatrixCols == 0 ? 0 : BlockCols, + MaxRowsAtCompileTime = BlockRows==0 ? 0 + : RowsAtCompileTime != Dynamic ? int(RowsAtCompileTime) + : int(traits<XprType>::MaxRowsAtCompileTime), + MaxColsAtCompileTime = BlockCols==0 ? 0 + : ColsAtCompileTime != Dynamic ? int(ColsAtCompileTime) + : int(traits<XprType>::MaxColsAtCompileTime), + + XprTypeIsRowMajor = (int(traits<XprType>::Flags)&RowMajorBit) != 0, + IsRowMajor = (MaxRowsAtCompileTime==1&&MaxColsAtCompileTime!=1) ? 1 + : (MaxColsAtCompileTime==1&&MaxRowsAtCompileTime!=1) ? 0 + : XprTypeIsRowMajor, + HasSameStorageOrderAsXprType = (IsRowMajor == XprTypeIsRowMajor), + InnerSize = IsRowMajor ? int(ColsAtCompileTime) : int(RowsAtCompileTime), + InnerStrideAtCompileTime = HasSameStorageOrderAsXprType + ? int(inner_stride_at_compile_time<XprType>::ret) + : int(outer_stride_at_compile_time<XprType>::ret), + OuterStrideAtCompileTime = HasSameStorageOrderAsXprType + ? int(outer_stride_at_compile_time<XprType>::ret) + : int(inner_stride_at_compile_time<XprType>::ret), + + // FIXME, this traits is rather specialized for dense object and it needs to be cleaned further + FlagsLvalueBit = is_lvalue<XprType>::value ? LvalueBit : 0, + FlagsRowMajorBit = IsRowMajor ? RowMajorBit : 0, + Flags = (traits<XprType>::Flags & (DirectAccessBit | (InnerPanel?CompressedAccessBit:0))) | FlagsLvalueBit | FlagsRowMajorBit, + // FIXME DirectAccessBit should not be handled by expressions + // + // Alignment is needed by MapBase's assertions + // We can sefely set it to false here. Internal alignment errors will be detected by an eigen_internal_assert in the respective evaluator + Alignment = 0 + }; +}; + +template<typename XprType, int BlockRows=Dynamic, int BlockCols=Dynamic, bool InnerPanel = false, + bool HasDirectAccess = internal::has_direct_access<XprType>::ret> class BlockImpl_dense; + +} // end namespace internal + +template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel, typename StorageKind> class BlockImpl; + +/** \class Block + * \ingroup Core_Module + * + * \brief Expression of a fixed-size or dynamic-size block + * + * \tparam XprType the type of the expression in which we are taking a block + * \tparam BlockRows the number of rows of the block we are taking at compile time (optional) + * \tparam BlockCols the number of columns of the block we are taking at compile time (optional) + * \tparam InnerPanel is true, if the block maps to a set of rows of a row major matrix or + * to set of columns of a column major matrix (optional). The parameter allows to determine + * at compile time whether aligned access is possible on the block expression. + * + * This class represents an expression of either a fixed-size or dynamic-size block. It is the return + * type of DenseBase::block(Index,Index,Index,Index) and DenseBase::block<int,int>(Index,Index) and + * most of the time this is the only way it is used. + * + * However, if you want to directly maniputate block expressions, + * for instance if you want to write a function returning such an expression, you + * will need to use this class. + * + * Here is an example illustrating the dynamic case: + * \include class_Block.cpp + * Output: \verbinclude class_Block.out + * + * \note Even though this expression has dynamic size, in the case where \a XprType + * has fixed size, this expression inherits a fixed maximal size which means that evaluating + * it does not cause a dynamic memory allocation. + * + * Here is an example illustrating the fixed-size case: + * \include class_FixedBlock.cpp + * Output: \verbinclude class_FixedBlock.out + * + * \sa DenseBase::block(Index,Index,Index,Index), DenseBase::block(Index,Index), class VectorBlock + */ +template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel> class Block + : public BlockImpl<XprType, BlockRows, BlockCols, InnerPanel, typename internal::traits<XprType>::StorageKind> +{ + typedef BlockImpl<XprType, BlockRows, BlockCols, InnerPanel, typename internal::traits<XprType>::StorageKind> Impl; + public: + //typedef typename Impl::Base Base; + typedef Impl Base; + EIGEN_GENERIC_PUBLIC_INTERFACE(Block) + EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Block) + + typedef typename internal::remove_all<XprType>::type NestedExpression; + + /** Column or Row constructor + */ + EIGEN_DEVICE_FUNC + inline Block(XprType& xpr, Index i) : Impl(xpr,i) + { + eigen_assert( (i>=0) && ( + ((BlockRows==1) && (BlockCols==XprType::ColsAtCompileTime) && i<xpr.rows()) + ||((BlockRows==XprType::RowsAtCompileTime) && (BlockCols==1) && i<xpr.cols()))); + } + + /** Fixed-size constructor + */ + EIGEN_DEVICE_FUNC + 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) + eigen_assert(startRow >= 0 && BlockRows >= 0 && startRow + BlockRows <= xpr.rows() + && startCol >= 0 && BlockCols >= 0 && startCol + BlockCols <= xpr.cols()); + } + + /** Dynamic-size constructor + */ + EIGEN_DEVICE_FUNC + inline Block(XprType& xpr, + Index startRow, Index startCol, + Index blockRows, Index blockCols) + : Impl(xpr, startRow, startCol, blockRows, blockCols) + { + eigen_assert((RowsAtCompileTime==Dynamic || RowsAtCompileTime==blockRows) + && (ColsAtCompileTime==Dynamic || ColsAtCompileTime==blockCols)); + eigen_assert(startRow >= 0 && blockRows >= 0 && startRow <= xpr.rows() - blockRows + && startCol >= 0 && blockCols >= 0 && startCol <= xpr.cols() - blockCols); + } +}; + +// The generic default implementation for dense block simplu forward to the internal::BlockImpl_dense +// that must be specialized for direct and non-direct access... +template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel> +class BlockImpl<XprType, BlockRows, BlockCols, InnerPanel, Dense> + : public internal::BlockImpl_dense<XprType, BlockRows, BlockCols, InnerPanel> +{ + typedef internal::BlockImpl_dense<XprType, BlockRows, BlockCols, InnerPanel> Impl; + typedef typename XprType::StorageIndex StorageIndex; + 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 + inline BlockImpl(XprType& xpr, Index startRow, Index startCol, Index blockRows, Index blockCols) + : Impl(xpr, startRow, startCol, blockRows, blockCols) {} +}; + +namespace internal { + +/** \internal Internal implementation of dense Blocks in the general case. */ +template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel, bool HasDirectAccess> class BlockImpl_dense + : public internal::dense_xpr_base<Block<XprType, BlockRows, BlockCols, InnerPanel> >::type +{ + typedef Block<XprType, BlockRows, BlockCols, InnerPanel> BlockType; + typedef typename internal::ref_selector<XprType>::non_const_type XprTypeNested; + public: + + typedef typename internal::dense_xpr_base<BlockType>::type Base; + EIGEN_DENSE_PUBLIC_INTERFACE(BlockType) + EIGEN_INHERIT_ASSIGNMENT_OPERATORS(BlockImpl_dense) + + // class InnerIterator; // FIXME apparently never used + + /** Column or Row constructor + */ + EIGEN_DEVICE_FUNC + inline BlockImpl_dense(XprType& xpr, Index i) + : m_xpr(xpr), + // It is a row if and only if BlockRows==1 and BlockCols==XprType::ColsAtCompileTime, + // and it is a column if and only if BlockRows==XprType::RowsAtCompileTime and BlockCols==1, + // all other cases are invalid. + // The case a 1x1 matrix seems ambiguous, but the result is the same anyway. + m_startRow( (BlockRows==1) && (BlockCols==XprType::ColsAtCompileTime) ? i : 0), + m_startCol( (BlockRows==XprType::RowsAtCompileTime) && (BlockCols==1) ? i : 0), + m_blockRows(BlockRows==1 ? 1 : xpr.rows()), + m_blockCols(BlockCols==1 ? 1 : xpr.cols()) + {} + + /** Fixed-size constructor + */ + EIGEN_DEVICE_FUNC + inline BlockImpl_dense(XprType& xpr, Index startRow, Index startCol) + : m_xpr(xpr), m_startRow(startRow), m_startCol(startCol), + m_blockRows(BlockRows), m_blockCols(BlockCols) + {} + + /** Dynamic-size constructor + */ + EIGEN_DEVICE_FUNC + inline BlockImpl_dense(XprType& xpr, + Index startRow, Index startCol, + Index blockRows, Index blockCols) + : m_xpr(xpr), m_startRow(startRow), m_startCol(startCol), + m_blockRows(blockRows), m_blockCols(blockCols) + {} + + EIGEN_DEVICE_FUNC inline Index rows() const { return m_blockRows.value(); } + EIGEN_DEVICE_FUNC inline Index cols() const { return m_blockCols.value(); } + + EIGEN_DEVICE_FUNC + inline Scalar& coeffRef(Index rowId, Index colId) + { + EIGEN_STATIC_ASSERT_LVALUE(XprType) + return m_xpr.coeffRef(rowId + m_startRow.value(), colId + m_startCol.value()); + } + + EIGEN_DEVICE_FUNC + inline const Scalar& coeffRef(Index rowId, Index colId) const + { + return m_xpr.derived().coeffRef(rowId + m_startRow.value(), colId + m_startCol.value()); + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const CoeffReturnType coeff(Index rowId, Index colId) const + { + return m_xpr.coeff(rowId + m_startRow.value(), colId + m_startCol.value()); + } + + EIGEN_DEVICE_FUNC + inline Scalar& coeffRef(Index index) + { + EIGEN_STATIC_ASSERT_LVALUE(XprType) + return m_xpr.coeffRef(m_startRow.value() + (RowsAtCompileTime == 1 ? 0 : index), + m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0)); + } + + EIGEN_DEVICE_FUNC + inline const Scalar& coeffRef(Index index) const + { + return m_xpr.coeffRef(m_startRow.value() + (RowsAtCompileTime == 1 ? 0 : index), + m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0)); + } + + EIGEN_DEVICE_FUNC + inline const CoeffReturnType coeff(Index index) const + { + return m_xpr.coeff(m_startRow.value() + (RowsAtCompileTime == 1 ? 0 : index), + m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0)); + } + + template<int LoadMode> + inline PacketScalar packet(Index rowId, Index colId) const + { + return m_xpr.template packet<Unaligned>(rowId + m_startRow.value(), colId + m_startCol.value()); + } + + template<int LoadMode> + inline void writePacket(Index rowId, Index colId, const PacketScalar& val) + { + m_xpr.template writePacket<Unaligned>(rowId + m_startRow.value(), colId + m_startCol.value(), val); + } + + template<int LoadMode> + inline PacketScalar packet(Index index) const + { + return m_xpr.template packet<Unaligned> + (m_startRow.value() + (RowsAtCompileTime == 1 ? 0 : index), + m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0)); + } + + template<int LoadMode> + inline void writePacket(Index index, const PacketScalar& val) + { + m_xpr.template writePacket<Unaligned> + (m_startRow.value() + (RowsAtCompileTime == 1 ? 0 : index), + m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0), val); + } + + #ifdef EIGEN_PARSED_BY_DOXYGEN + /** \sa MapBase::data() */ + EIGEN_DEVICE_FUNC inline const Scalar* data() const; + EIGEN_DEVICE_FUNC inline Index innerStride() const; + EIGEN_DEVICE_FUNC inline Index outerStride() const; + #endif + + EIGEN_DEVICE_FUNC + const typename internal::remove_all<XprTypeNested>::type& nestedExpression() const + { + return m_xpr; + } + + EIGEN_DEVICE_FUNC + XprType& nestedExpression() { return m_xpr; } + + EIGEN_DEVICE_FUNC + StorageIndex startRow() const + { + return m_startRow.value(); + } + + EIGEN_DEVICE_FUNC + StorageIndex startCol() const + { + return m_startCol.value(); + } + + protected: + + XprTypeNested m_xpr; + const internal::variable_if_dynamic<StorageIndex, (XprType::RowsAtCompileTime == 1 && BlockRows==1) ? 0 : Dynamic> m_startRow; + const internal::variable_if_dynamic<StorageIndex, (XprType::ColsAtCompileTime == 1 && BlockCols==1) ? 0 : Dynamic> m_startCol; + const internal::variable_if_dynamic<StorageIndex, RowsAtCompileTime> m_blockRows; + const internal::variable_if_dynamic<StorageIndex, ColsAtCompileTime> m_blockCols; +}; + +/** \internal Internal implementation of dense Blocks in the direct access case.*/ +template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel> +class BlockImpl_dense<XprType,BlockRows,BlockCols, InnerPanel,true> + : public MapBase<Block<XprType, BlockRows, BlockCols, InnerPanel> > +{ + typedef Block<XprType, BlockRows, BlockCols, InnerPanel> BlockType; + typedef typename internal::ref_selector<XprType>::non_const_type XprTypeNested; + enum { + XprTypeIsRowMajor = (int(traits<XprType>::Flags)&RowMajorBit) != 0 + }; + public: + + typedef MapBase<BlockType> Base; + EIGEN_DENSE_PUBLIC_INTERFACE(BlockType) + EIGEN_INHERIT_ASSIGNMENT_OPERATORS(BlockImpl_dense) + + /** Column or Row constructor + */ + EIGEN_DEVICE_FUNC + 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(), + BlockCols==1 ? 1 : xpr.cols()), + m_xpr(xpr), + m_startRow( (BlockRows==1) && (BlockCols==XprType::ColsAtCompileTime) ? i : 0), + m_startCol( (BlockRows==XprType::RowsAtCompileTime) && (BlockCols==1) ? i : 0) + { + init(); + } + + /** Fixed-size constructor + */ + EIGEN_DEVICE_FUNC + 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) + { + init(); + } + + /** Dynamic-size constructor + */ + EIGEN_DEVICE_FUNC + 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), + m_xpr(xpr), m_startRow(startRow), m_startCol(startCol) + { + init(); + } + + EIGEN_DEVICE_FUNC + const typename internal::remove_all<XprTypeNested>::type& nestedExpression() const + { + return m_xpr; + } + + EIGEN_DEVICE_FUNC + XprType& nestedExpression() { return m_xpr; } + + /** \sa MapBase::innerStride() */ + EIGEN_DEVICE_FUNC + inline Index innerStride() const + { + return internal::traits<BlockType>::HasSameStorageOrderAsXprType + ? m_xpr.innerStride() + : m_xpr.outerStride(); + } + + /** \sa MapBase::outerStride() */ + EIGEN_DEVICE_FUNC + inline Index outerStride() const + { + return m_outerStride; + } + + EIGEN_DEVICE_FUNC + StorageIndex startRow() const + { + return m_startRow.value(); + } + + EIGEN_DEVICE_FUNC + StorageIndex startCol() const + { + return m_startCol.value(); + } + + #ifndef __SUNPRO_CC + // FIXME sunstudio is not friendly with the above friend... + // META-FIXME there is no 'friend' keyword around here. Is this obsolete? + protected: + #endif + + #ifndef EIGEN_PARSED_BY_DOXYGEN + /** \internal used by allowAligned() */ + EIGEN_DEVICE_FUNC + inline BlockImpl_dense(XprType& xpr, const Scalar* data, Index blockRows, Index blockCols) + : Base(data, blockRows, blockCols), m_xpr(xpr) + { + init(); + } + #endif + + protected: + EIGEN_DEVICE_FUNC + void init() + { + m_outerStride = internal::traits<BlockType>::HasSameStorageOrderAsXprType + ? m_xpr.outerStride() + : m_xpr.innerStride(); + } + + XprTypeNested m_xpr; + const internal::variable_if_dynamic<StorageIndex, (XprType::RowsAtCompileTime == 1 && BlockRows==1) ? 0 : Dynamic> m_startRow; + const internal::variable_if_dynamic<StorageIndex, (XprType::ColsAtCompileTime == 1 && BlockCols==1) ? 0 : Dynamic> m_startCol; + Index m_outerStride; +}; + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_BLOCK_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/BooleanRedux.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/BooleanRedux.h new file mode 100644 index 000000000..8409d8749 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/BooleanRedux.h @@ -0,0 +1,164 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008 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_ALLANDANY_H +#define EIGEN_ALLANDANY_H + +namespace Eigen { + +namespace internal { + +template<typename Derived, int UnrollCount> +struct all_unroller +{ + typedef typename Derived::ExpressionTraits Traits; + enum { + col = (UnrollCount-1) / Traits::RowsAtCompileTime, + row = (UnrollCount-1) % Traits::RowsAtCompileTime + }; + + static inline bool run(const Derived &mat) + { + return all_unroller<Derived, UnrollCount-1>::run(mat) && mat.coeff(row, col); + } +}; + +template<typename Derived> +struct all_unroller<Derived, 0> +{ + static inline bool run(const Derived &/*mat*/) { return true; } +}; + +template<typename Derived> +struct all_unroller<Derived, Dynamic> +{ + static inline bool run(const Derived &) { return false; } +}; + +template<typename Derived, int UnrollCount> +struct any_unroller +{ + typedef typename Derived::ExpressionTraits Traits; + enum { + col = (UnrollCount-1) / Traits::RowsAtCompileTime, + row = (UnrollCount-1) % Traits::RowsAtCompileTime + }; + + static inline bool run(const Derived &mat) + { + return any_unroller<Derived, UnrollCount-1>::run(mat) || mat.coeff(row, col); + } +}; + +template<typename Derived> +struct any_unroller<Derived, 0> +{ + static inline bool run(const Derived & /*mat*/) { return false; } +}; + +template<typename Derived> +struct any_unroller<Derived, Dynamic> +{ + static inline bool run(const Derived &) { return false; } +}; + +} // end namespace internal + +/** \returns true if all coefficients are true + * + * Example: \include MatrixBase_all.cpp + * Output: \verbinclude MatrixBase_all.out + * + * \sa any(), Cwise::operator<() + */ +template<typename Derived> +inline bool DenseBase<Derived>::all() const +{ + typedef internal::evaluator<Derived> Evaluator; + enum { + unroll = SizeAtCompileTime != Dynamic + && SizeAtCompileTime * (Evaluator::CoeffReadCost + NumTraits<Scalar>::AddCost) <= EIGEN_UNROLLING_LIMIT + }; + Evaluator evaluator(derived()); + if(unroll) + return internal::all_unroller<Evaluator, unroll ? int(SizeAtCompileTime) : Dynamic>::run(evaluator); + else + { + for(Index j = 0; j < cols(); ++j) + for(Index i = 0; i < rows(); ++i) + if (!evaluator.coeff(i, j)) return false; + return true; + } +} + +/** \returns true if at least one coefficient is true + * + * \sa all() + */ +template<typename Derived> +inline bool DenseBase<Derived>::any() const +{ + typedef internal::evaluator<Derived> Evaluator; + enum { + unroll = SizeAtCompileTime != Dynamic + && SizeAtCompileTime * (Evaluator::CoeffReadCost + NumTraits<Scalar>::AddCost) <= EIGEN_UNROLLING_LIMIT + }; + Evaluator evaluator(derived()); + if(unroll) + return internal::any_unroller<Evaluator, unroll ? int(SizeAtCompileTime) : Dynamic>::run(evaluator); + else + { + for(Index j = 0; j < cols(); ++j) + for(Index i = 0; i < rows(); ++i) + if (evaluator.coeff(i, j)) return true; + return false; + } +} + +/** \returns the number of coefficients which evaluate to true + * + * \sa all(), any() + */ +template<typename Derived> +inline Eigen::Index DenseBase<Derived>::count() const +{ + return derived().template cast<bool>().template cast<Index>().sum(); +} + +/** \returns true is \c *this contains at least one Not A Number (NaN). + * + * \sa allFinite() + */ +template<typename Derived> +inline bool DenseBase<Derived>::hasNaN() const +{ +#if EIGEN_COMP_MSVC || (defined __FAST_MATH__) + return derived().array().isNaN().any(); +#else + return !((derived().array()==derived().array()).all()); +#endif +} + +/** \returns true if \c *this contains only finite numbers, i.e., no NaN and no +/-INF values. + * + * \sa hasNaN() + */ +template<typename Derived> +inline bool DenseBase<Derived>::allFinite() const +{ +#if EIGEN_COMP_MSVC || (defined __FAST_MATH__) + return derived().array().isFinite().all(); +#else + return !((derived()-derived()).hasNaN()); +#endif +} + +} // end namespace Eigen + +#endif // EIGEN_ALLANDANY_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/CommaInitializer.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/CommaInitializer.h new file mode 100644 index 000000000..d218e9814 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/CommaInitializer.h @@ -0,0 +1,160 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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_COMMAINITIALIZER_H +#define EIGEN_COMMAINITIALIZER_H + +namespace Eigen { + +/** \class CommaInitializer + * \ingroup Core_Module + * + * \brief Helper class used by the comma initializer operator + * + * This class is internally used to implement the comma initializer feature. It is + * the return type of MatrixBase::operator<<, and most of the time this is the only + * way it is used. + * + * \sa \blank \ref MatrixBaseCommaInitRef "MatrixBase::operator<<", CommaInitializer::finished() + */ +template<typename XprType> +struct CommaInitializer +{ + typedef typename XprType::Scalar Scalar; + + EIGEN_DEVICE_FUNC + inline CommaInitializer(XprType& xpr, const Scalar& s) + : m_xpr(xpr), m_row(0), m_col(1), m_currentBlockRows(1) + { + m_xpr.coeffRef(0,0) = s; + } + + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + inline CommaInitializer(XprType& xpr, const DenseBase<OtherDerived>& other) + : m_xpr(xpr), m_row(0), m_col(other.cols()), m_currentBlockRows(other.rows()) + { + m_xpr.block(0, 0, other.rows(), other.cols()) = other; + } + + /* Copy/Move constructor which transfers ownership. This is crucial in + * absence of return value optimization to avoid assertions during destruction. */ + // FIXME in C++11 mode this could be replaced by a proper RValue constructor + EIGEN_DEVICE_FUNC + inline CommaInitializer(const CommaInitializer& o) + : m_xpr(o.m_xpr), m_row(o.m_row), m_col(o.m_col), m_currentBlockRows(o.m_currentBlockRows) { + // Mark original object as finished. In absence of R-value references we need to const_cast: + const_cast<CommaInitializer&>(o).m_row = m_xpr.rows(); + const_cast<CommaInitializer&>(o).m_col = m_xpr.cols(); + const_cast<CommaInitializer&>(o).m_currentBlockRows = 0; + } + + /* inserts a scalar value in the target matrix */ + EIGEN_DEVICE_FUNC + CommaInitializer& operator,(const Scalar& s) + { + if (m_col==m_xpr.cols()) + { + m_row+=m_currentBlockRows; + m_col = 0; + m_currentBlockRows = 1; + eigen_assert(m_row<m_xpr.rows() + && "Too many rows passed to comma initializer (operator<<)"); + } + eigen_assert(m_col<m_xpr.cols() + && "Too many coefficients passed to comma initializer (operator<<)"); + eigen_assert(m_currentBlockRows==1); + m_xpr.coeffRef(m_row, m_col++) = s; + return *this; + } + + /* inserts a matrix expression in the target matrix */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + CommaInitializer& operator,(const DenseBase<OtherDerived>& other) + { + if (m_col==m_xpr.cols() && (other.cols()!=0 || other.rows()!=m_currentBlockRows)) + { + m_row+=m_currentBlockRows; + m_col = 0; + m_currentBlockRows = other.rows(); + eigen_assert(m_row+m_currentBlockRows<=m_xpr.rows() + && "Too many rows passed to comma initializer (operator<<)"); + } + eigen_assert((m_col + other.cols() <= m_xpr.cols()) + && "Too many coefficients passed to comma initializer (operator<<)"); + eigen_assert(m_currentBlockRows==other.rows()); + m_xpr.template block<OtherDerived::RowsAtCompileTime, OtherDerived::ColsAtCompileTime> + (m_row, m_col, other.rows(), other.cols()) = other; + m_col += other.cols(); + return *this; + } + + EIGEN_DEVICE_FUNC + inline ~CommaInitializer() +#if defined VERIFY_RAISES_ASSERT && (!defined EIGEN_NO_ASSERTION_CHECKING) && defined EIGEN_EXCEPTIONS + EIGEN_EXCEPTION_SPEC(Eigen::eigen_assert_exception) +#endif + { + finished(); + } + + /** \returns the built matrix once all its coefficients have been set. + * Calling finished is 100% optional. Its purpose is to write expressions + * like this: + * \code + * quaternion.fromRotationMatrix((Matrix3f() << axis0, axis1, axis2).finished()); + * \endcode + */ + EIGEN_DEVICE_FUNC + inline XprType& finished() { + eigen_assert(((m_row+m_currentBlockRows) == m_xpr.rows() || m_xpr.cols() == 0) + && m_col == m_xpr.cols() + && "Too few coefficients passed to comma initializer (operator<<)"); + return m_xpr; + } + + XprType& m_xpr; // target expression + Index m_row; // current row id + Index m_col; // current col id + Index m_currentBlockRows; // current block height +}; + +/** \anchor MatrixBaseCommaInitRef + * Convenient operator to set the coefficients of a matrix. + * + * The coefficients must be provided in a row major order and exactly match + * the size of the matrix. Otherwise an assertion is raised. + * + * Example: \include MatrixBase_set.cpp + * Output: \verbinclude MatrixBase_set.out + * + * \note According the c++ standard, the argument expressions of this comma initializer are evaluated in arbitrary order. + * + * \sa CommaInitializer::finished(), class CommaInitializer + */ +template<typename Derived> +inline CommaInitializer<Derived> DenseBase<Derived>::operator<< (const Scalar& s) +{ + return CommaInitializer<Derived>(*static_cast<Derived*>(this), s); +} + +/** \sa operator<<(const Scalar&) */ +template<typename Derived> +template<typename OtherDerived> +inline CommaInitializer<Derived> +DenseBase<Derived>::operator<<(const DenseBase<OtherDerived>& other) +{ + return CommaInitializer<Derived>(*static_cast<Derived *>(this), other); +} + +} // end namespace Eigen + +#endif // EIGEN_COMMAINITIALIZER_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/ConditionEstimator.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/ConditionEstimator.h new file mode 100644 index 000000000..aa7efdc76 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/ConditionEstimator.h @@ -0,0 +1,175 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2016 Rasmus Munk Larsen (rmlarsen@google.com) +// +// 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_CONDITIONESTIMATOR_H +#define EIGEN_CONDITIONESTIMATOR_H + +namespace Eigen { + +namespace internal { + +template <typename Vector, typename RealVector, bool IsComplex> +struct rcond_compute_sign { + static inline Vector run(const Vector& v) { + const RealVector v_abs = v.cwiseAbs(); + return (v_abs.array() == static_cast<typename Vector::RealScalar>(0)) + .select(Vector::Ones(v.size()), v.cwiseQuotient(v_abs)); + } +}; + +// Partial specialization to avoid elementwise division for real vectors. +template <typename Vector> +struct rcond_compute_sign<Vector, Vector, false> { + static inline Vector run(const Vector& v) { + return (v.array() < static_cast<typename Vector::RealScalar>(0)) + .select(-Vector::Ones(v.size()), Vector::Ones(v.size())); + } +}; + +/** + * \returns an estimate of ||inv(matrix)||_1 given a decomposition of + * \a matrix that implements .solve() and .adjoint().solve() methods. + * + * This function implements Algorithms 4.1 and 5.1 from + * http://www.maths.manchester.ac.uk/~higham/narep/narep135.pdf + * which also forms the basis for the condition number estimators in + * LAPACK. Since at most 10 calls to the solve method of dec are + * performed, the total cost is O(dims^2), as opposed to O(dims^3) + * needed to compute the inverse matrix explicitly. + * + * The most common usage is in estimating the condition number + * ||matrix||_1 * ||inv(matrix)||_1. The first term ||matrix||_1 can be + * computed directly in O(n^2) operations. + * + * Supports the following decompositions: FullPivLU, PartialPivLU, LDLT, and + * LLT. + * + * \sa FullPivLU, PartialPivLU, LDLT, LLT. + */ +template <typename Decomposition> +typename Decomposition::RealScalar rcond_invmatrix_L1_norm_estimate(const Decomposition& dec) +{ + typedef typename Decomposition::MatrixType MatrixType; + typedef typename Decomposition::Scalar Scalar; + typedef typename Decomposition::RealScalar RealScalar; + typedef typename internal::plain_col_type<MatrixType>::type Vector; + typedef typename internal::plain_col_type<MatrixType, RealScalar>::type RealVector; + const bool is_complex = (NumTraits<Scalar>::IsComplex != 0); + + eigen_assert(dec.rows() == dec.cols()); + const Index n = dec.rows(); + if (n == 0) + return 0; + + // Disable Index to float conversion warning +#ifdef __INTEL_COMPILER + #pragma warning push + #pragma warning ( disable : 2259 ) +#endif + Vector v = dec.solve(Vector::Ones(n) / Scalar(n)); +#ifdef __INTEL_COMPILER + #pragma warning pop +#endif + + // lower_bound is a lower bound on + // ||inv(matrix)||_1 = sup_v ||inv(matrix) v||_1 / ||v||_1 + // and is the objective maximized by the ("super-") gradient ascent + // algorithm below. + RealScalar lower_bound = v.template lpNorm<1>(); + if (n == 1) + return lower_bound; + + // Gradient ascent algorithm follows: We know that the optimum is achieved at + // one of the simplices v = e_i, so in each iteration we follow a + // super-gradient to move towards the optimal one. + RealScalar old_lower_bound = lower_bound; + Vector sign_vector(n); + Vector old_sign_vector; + Index v_max_abs_index = -1; + Index old_v_max_abs_index = v_max_abs_index; + for (int k = 0; k < 4; ++k) + { + sign_vector = internal::rcond_compute_sign<Vector, RealVector, is_complex>::run(v); + if (k > 0 && !is_complex && sign_vector == old_sign_vector) { + // Break if the solution stagnated. + break; + } + // v_max_abs_index = argmax |real( inv(matrix)^T * sign_vector )| + v = dec.adjoint().solve(sign_vector); + v.real().cwiseAbs().maxCoeff(&v_max_abs_index); + if (v_max_abs_index == old_v_max_abs_index) { + // Break if the solution stagnated. + break; + } + // Move to the new simplex e_j, where j = v_max_abs_index. + v = dec.solve(Vector::Unit(n, v_max_abs_index)); // v = inv(matrix) * e_j. + lower_bound = v.template lpNorm<1>(); + if (lower_bound <= old_lower_bound) { + // Break if the gradient step did not increase the lower_bound. + break; + } + if (!is_complex) { + old_sign_vector = sign_vector; + } + old_v_max_abs_index = v_max_abs_index; + old_lower_bound = lower_bound; + } + // The following calculates an independent estimate of ||matrix||_1 by + // multiplying matrix by a vector with entries of slowly increasing + // magnitude and alternating sign: + // v_i = (-1)^{i} (1 + (i / (dim-1))), i = 0,...,dim-1. + // This improvement to Hager's algorithm above is due to Higham. It was + // added to make the algorithm more robust in certain corner cases where + // large elements in the matrix might otherwise escape detection due to + // exact cancellation (especially when op and op_adjoint correspond to a + // sequence of backsubstitutions and permutations), which could cause + // Hager's algorithm to vastly underestimate ||matrix||_1. + Scalar alternating_sign(RealScalar(1)); + for (Index i = 0; i < n; ++i) { + // The static_cast is needed when Scalar is a complex and RealScalar implements expression templates + v[i] = alternating_sign * static_cast<RealScalar>(RealScalar(1) + (RealScalar(i) / (RealScalar(n - 1)))); + alternating_sign = -alternating_sign; + } + v = dec.solve(v); + const RealScalar alternate_lower_bound = (2 * v.template lpNorm<1>()) / (3 * RealScalar(n)); + return numext::maxi(lower_bound, alternate_lower_bound); +} + +/** \brief Reciprocal condition number estimator. + * + * Computing a decomposition of a dense matrix takes O(n^3) operations, while + * this method estimates the condition number quickly and reliably in O(n^2) + * operations. + * + * \returns an estimate of the reciprocal condition number + * (1 / (||matrix||_1 * ||inv(matrix)||_1)) of matrix, given ||matrix||_1 and + * its decomposition. Supports the following decompositions: FullPivLU, + * PartialPivLU, LDLT, and LLT. + * + * \sa FullPivLU, PartialPivLU, LDLT, LLT. + */ +template <typename Decomposition> +typename Decomposition::RealScalar +rcond_estimate_helper(typename Decomposition::RealScalar matrix_norm, const Decomposition& dec) +{ + typedef typename Decomposition::RealScalar RealScalar; + eigen_assert(dec.rows() == dec.cols()); + if (dec.rows() == 0) return RealScalar(1); + 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); + return (inverse_matrix_norm == RealScalar(0) ? RealScalar(0) + : (RealScalar(1) / inverse_matrix_norm) / matrix_norm); +} + +} // namespace internal + +} // namespace Eigen + +#endif diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/CoreEvaluators.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/CoreEvaluators.h new file mode 100644 index 000000000..f7c1effca --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/CoreEvaluators.h @@ -0,0 +1,1671 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2011 Benoit Jacob <jacob.benoit.1@gmail.com> +// Copyright (C) 2011-2014 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2011-2012 Jitse Niesen <jitse@maths.leeds.ac.uk> +// +// 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_COREEVALUATORS_H +#define EIGEN_COREEVALUATORS_H + +namespace Eigen { + +namespace internal { + +// This class returns the evaluator kind from the expression storage kind. +// Default assumes index based accessors +template<typename StorageKind> +struct storage_kind_to_evaluator_kind { + typedef IndexBased Kind; +}; + +// This class returns the evaluator shape from the expression storage kind. +// It can be Dense, Sparse, Triangular, Diagonal, SelfAdjoint, Band, etc. +template<typename StorageKind> struct storage_kind_to_shape; + +template<> struct storage_kind_to_shape<Dense> { typedef DenseShape Shape; }; +template<> struct storage_kind_to_shape<SolverStorage> { typedef SolverShape Shape; }; +template<> struct storage_kind_to_shape<PermutationStorage> { typedef PermutationShape Shape; }; +template<> struct storage_kind_to_shape<TranspositionsStorage> { typedef TranspositionsShape Shape; }; + +// Evaluators have to be specialized with respect to various criteria such as: +// - storage/structure/shape +// - scalar type +// - etc. +// Therefore, we need specialization of evaluator providing additional template arguments for each kind of evaluators. +// We currently distinguish the following kind of evaluators: +// - unary_evaluator for expressions taking only one arguments (CwiseUnaryOp, CwiseUnaryView, Transpose, MatrixWrapper, ArrayWrapper, Reverse, Replicate) +// - binary_evaluator for expression taking two arguments (CwiseBinaryOp) +// - ternary_evaluator for expression taking three arguments (CwiseTernaryOp) +// - product_evaluator for linear algebra products (Product); special case of binary_evaluator because it requires additional tags for dispatching. +// - mapbase_evaluator for Map, Block, Ref +// - block_evaluator for Block (special dispatching to a mapbase_evaluator or unary_evaluator) + +template< typename T, + typename Arg1Kind = typename evaluator_traits<typename T::Arg1>::Kind, + typename Arg2Kind = typename evaluator_traits<typename T::Arg2>::Kind, + typename Arg3Kind = typename evaluator_traits<typename T::Arg3>::Kind, + typename Arg1Scalar = typename traits<typename T::Arg1>::Scalar, + typename Arg2Scalar = typename traits<typename T::Arg2>::Scalar, + typename Arg3Scalar = typename traits<typename T::Arg3>::Scalar> struct ternary_evaluator; + +template< typename T, + typename LhsKind = typename evaluator_traits<typename T::Lhs>::Kind, + typename RhsKind = typename evaluator_traits<typename T::Rhs>::Kind, + typename LhsScalar = typename traits<typename T::Lhs>::Scalar, + typename RhsScalar = typename traits<typename T::Rhs>::Scalar> struct binary_evaluator; + +template< typename T, + typename Kind = typename evaluator_traits<typename T::NestedExpression>::Kind, + typename Scalar = typename T::Scalar> struct unary_evaluator; + +// evaluator_traits<T> contains traits for evaluator<T> + +template<typename T> +struct evaluator_traits_base +{ + // by default, get evaluator kind and shape from storage + typedef typename storage_kind_to_evaluator_kind<typename traits<T>::StorageKind>::Kind Kind; + typedef typename storage_kind_to_shape<typename traits<T>::StorageKind>::Shape Shape; +}; + +// Default evaluator traits +template<typename T> +struct evaluator_traits : public evaluator_traits_base<T> +{ +}; + +template<typename T, typename Shape = typename evaluator_traits<T>::Shape > +struct evaluator_assume_aliasing { + static const bool value = false; +}; + +// By default, we assume a unary expression: +template<typename T> +struct evaluator : public unary_evaluator<T> +{ + typedef unary_evaluator<T> Base; + EIGEN_DEVICE_FUNC explicit evaluator(const T& xpr) : Base(xpr) {} +}; + + +// TODO: Think about const-correctness +template<typename T> +struct evaluator<const T> + : evaluator<T> +{ + EIGEN_DEVICE_FUNC + explicit evaluator(const T& xpr) : evaluator<T>(xpr) {} +}; + +// ---------- base class for all evaluators ---------- + +template<typename ExpressionType> +struct evaluator_base : public noncopyable +{ + // TODO that's not very nice to have to propagate all these traits. They are currently only needed to handle outer,inner indices. + typedef traits<ExpressionType> ExpressionTraits; + + enum { + Alignment = 0 + }; +}; + +// -------------------- Matrix and Array -------------------- +// +// evaluator<PlainObjectBase> is a common base class for the +// Matrix and Array evaluators. +// Here we directly specialize evaluator. This is not really a unary expression, and it is, by definition, dense, +// so no need for more sophisticated dispatching. + +template<typename Derived> +struct evaluator<PlainObjectBase<Derived> > + : evaluator_base<Derived> +{ + typedef PlainObjectBase<Derived> PlainObjectType; + typedef typename PlainObjectType::Scalar Scalar; + typedef typename PlainObjectType::CoeffReturnType CoeffReturnType; + + enum { + IsRowMajor = PlainObjectType::IsRowMajor, + IsVectorAtCompileTime = PlainObjectType::IsVectorAtCompileTime, + RowsAtCompileTime = PlainObjectType::RowsAtCompileTime, + ColsAtCompileTime = PlainObjectType::ColsAtCompileTime, + + CoeffReadCost = NumTraits<Scalar>::ReadCost, + Flags = traits<Derived>::EvaluatorFlags, + Alignment = traits<Derived>::Alignment + }; + + EIGEN_DEVICE_FUNC evaluator() + : m_data(0), + m_outerStride(IsVectorAtCompileTime ? 0 + : int(IsRowMajor) ? ColsAtCompileTime + : RowsAtCompileTime) + { + EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost); + } + + EIGEN_DEVICE_FUNC explicit evaluator(const PlainObjectType& m) + : m_data(m.data()), m_outerStride(IsVectorAtCompileTime ? 0 : m.outerStride()) + { + EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index row, Index col) const + { + if (IsRowMajor) + return m_data[row * m_outerStride.value() + col]; + else + return m_data[row + col * m_outerStride.value()]; + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index index) const + { + return m_data[index]; + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Scalar& coeffRef(Index row, Index col) + { + if (IsRowMajor) + return const_cast<Scalar*>(m_data)[row * m_outerStride.value() + col]; + else + return const_cast<Scalar*>(m_data)[row + col * m_outerStride.value()]; + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Scalar& coeffRef(Index index) + { + return const_cast<Scalar*>(m_data)[index]; + } + + template<int LoadMode, typename PacketType> + EIGEN_STRONG_INLINE + PacketType packet(Index row, Index col) const + { + if (IsRowMajor) + return ploadt<PacketType, LoadMode>(m_data + row * m_outerStride.value() + col); + else + return ploadt<PacketType, LoadMode>(m_data + row + col * m_outerStride.value()); + } + + template<int LoadMode, typename PacketType> + EIGEN_STRONG_INLINE + PacketType packet(Index index) const + { + return ploadt<PacketType, LoadMode>(m_data + index); + } + + template<int StoreMode,typename PacketType> + EIGEN_STRONG_INLINE + void writePacket(Index row, Index col, const PacketType& x) + { + if (IsRowMajor) + return pstoret<Scalar, PacketType, StoreMode> + (const_cast<Scalar*>(m_data) + row * m_outerStride.value() + col, x); + else + return pstoret<Scalar, PacketType, StoreMode> + (const_cast<Scalar*>(m_data) + row + col * m_outerStride.value(), x); + } + + template<int StoreMode, typename PacketType> + EIGEN_STRONG_INLINE + void writePacket(Index index, const PacketType& x) + { + return pstoret<Scalar, PacketType, StoreMode>(const_cast<Scalar*>(m_data) + index, x); + } + +protected: + const Scalar *m_data; + + // We do not need to know the outer stride for vectors + variable_if_dynamic<Index, IsVectorAtCompileTime ? 0 + : int(IsRowMajor) ? ColsAtCompileTime + : RowsAtCompileTime> m_outerStride; +}; + +template<typename Scalar, int Rows, int Cols, int Options, int MaxRows, int MaxCols> +struct evaluator<Matrix<Scalar, Rows, Cols, Options, MaxRows, MaxCols> > + : evaluator<PlainObjectBase<Matrix<Scalar, Rows, Cols, Options, MaxRows, MaxCols> > > +{ + typedef Matrix<Scalar, Rows, Cols, Options, MaxRows, MaxCols> XprType; + + EIGEN_DEVICE_FUNC evaluator() {} + + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& m) + : evaluator<PlainObjectBase<XprType> >(m) + { } +}; + +template<typename Scalar, int Rows, int Cols, int Options, int MaxRows, int MaxCols> +struct evaluator<Array<Scalar, Rows, Cols, Options, MaxRows, MaxCols> > + : evaluator<PlainObjectBase<Array<Scalar, Rows, Cols, Options, MaxRows, MaxCols> > > +{ + typedef Array<Scalar, Rows, Cols, Options, MaxRows, MaxCols> XprType; + + EIGEN_DEVICE_FUNC evaluator() {} + + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& m) + : evaluator<PlainObjectBase<XprType> >(m) + { } +}; + +// -------------------- Transpose -------------------- + +template<typename ArgType> +struct unary_evaluator<Transpose<ArgType>, IndexBased> + : evaluator_base<Transpose<ArgType> > +{ + typedef Transpose<ArgType> XprType; + + enum { + CoeffReadCost = evaluator<ArgType>::CoeffReadCost, + Flags = evaluator<ArgType>::Flags ^ RowMajorBit, + Alignment = evaluator<ArgType>::Alignment + }; + + EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& t) : m_argImpl(t.nestedExpression()) {} + + typedef typename XprType::Scalar Scalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index row, Index col) const + { + return m_argImpl.coeff(col, row); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index index) const + { + return m_argImpl.coeff(index); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Scalar& coeffRef(Index row, Index col) + { + return m_argImpl.coeffRef(col, row); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + typename XprType::Scalar& coeffRef(Index index) + { + return m_argImpl.coeffRef(index); + } + + template<int LoadMode, typename PacketType> + EIGEN_STRONG_INLINE + PacketType packet(Index row, Index col) const + { + return m_argImpl.template packet<LoadMode,PacketType>(col, row); + } + + template<int LoadMode, typename PacketType> + EIGEN_STRONG_INLINE + PacketType packet(Index index) const + { + return m_argImpl.template packet<LoadMode,PacketType>(index); + } + + template<int StoreMode, typename PacketType> + EIGEN_STRONG_INLINE + void writePacket(Index row, Index col, const PacketType& x) + { + m_argImpl.template writePacket<StoreMode,PacketType>(col, row, x); + } + + template<int StoreMode, typename PacketType> + EIGEN_STRONG_INLINE + void writePacket(Index index, const PacketType& x) + { + m_argImpl.template writePacket<StoreMode,PacketType>(index, x); + } + +protected: + evaluator<ArgType> m_argImpl; +}; + +// -------------------- CwiseNullaryOp -------------------- +// Like Matrix and Array, this is not really a unary expression, so we directly specialize evaluator. +// Likewise, there is not need to more sophisticated dispatching here. + +template<typename Scalar,typename NullaryOp, + bool has_nullary = has_nullary_operator<NullaryOp>::value, + bool has_unary = has_unary_operator<NullaryOp>::value, + bool has_binary = has_binary_operator<NullaryOp>::value> +struct nullary_wrapper +{ + template <typename IndexType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i, IndexType j) const { return op(i,j); } + template <typename IndexType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i) const { return op(i); } + + template <typename T, typename IndexType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i, IndexType j) const { return op.template packetOp<T>(i,j); } + template <typename T, typename IndexType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i) const { return op.template packetOp<T>(i); } +}; + +template<typename Scalar,typename NullaryOp> +struct nullary_wrapper<Scalar,NullaryOp,true,false,false> +{ + template <typename IndexType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType=0, IndexType=0) const { return op(); } + template <typename T, typename IndexType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType=0, IndexType=0) const { return op.template packetOp<T>(); } +}; + +template<typename Scalar,typename NullaryOp> +struct nullary_wrapper<Scalar,NullaryOp,false,false,true> +{ + template <typename IndexType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i, IndexType j=0) const { return op(i,j); } + template <typename T, typename IndexType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i, IndexType j=0) const { return op.template packetOp<T>(i,j); } +}; + +// We need the following specialization for vector-only functors assigned to a runtime vector, +// for instance, using linspace and assigning a RowVectorXd to a MatrixXd or even a row of a MatrixXd. +// In this case, i==0 and j is used for the actual iteration. +template<typename Scalar,typename NullaryOp> +struct nullary_wrapper<Scalar,NullaryOp,false,true,false> +{ + template <typename IndexType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i, IndexType j) const { + eigen_assert(i==0 || j==0); + return op(i+j); + } + template <typename T, typename IndexType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i, IndexType j) const { + eigen_assert(i==0 || j==0); + return op.template packetOp<T>(i+j); + } + + template <typename IndexType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i) const { return op(i); } + template <typename T, typename IndexType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i) const { return op.template packetOp<T>(i); } +}; + +template<typename Scalar,typename NullaryOp> +struct nullary_wrapper<Scalar,NullaryOp,false,false,false> {}; + +#if 0 && EIGEN_COMP_MSVC>0 +// Disable this ugly workaround. This is now handled in traits<Ref>::match, +// but this piece of code might still become handly if some other weird compilation +// erros pop up again. + +// MSVC exhibits a weird compilation error when +// compiling: +// Eigen::MatrixXf A = MatrixXf::Random(3,3); +// Ref<const MatrixXf> R = 2.f*A; +// and that has_*ary_operator<scalar_constant_op<float>> have not been instantiated yet. +// The "problem" is that evaluator<2.f*A> is instantiated by traits<Ref>::match<2.f*A> +// and at that time has_*ary_operator<T> returns true regardless of T. +// Then nullary_wrapper is badly instantiated as nullary_wrapper<.,.,true,true,true>. +// The trick is thus to defer the proper instantiation of nullary_wrapper when coeff(), +// and packet() are really instantiated as implemented below: + +// This is a simple wrapper around Index to enforce the re-instantiation of +// has_*ary_operator when needed. +template<typename T> struct nullary_wrapper_workaround_msvc { + nullary_wrapper_workaround_msvc(const T&); + operator T()const; +}; + +template<typename Scalar,typename NullaryOp> +struct nullary_wrapper<Scalar,NullaryOp,true,true,true> +{ + template <typename IndexType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i, IndexType j) const { + return nullary_wrapper<Scalar,NullaryOp, + has_nullary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value, + has_unary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value, + has_binary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value>().operator()(op,i,j); + } + template <typename IndexType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i) const { + return nullary_wrapper<Scalar,NullaryOp, + has_nullary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value, + has_unary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value, + has_binary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value>().operator()(op,i); + } + + template <typename T, typename IndexType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i, IndexType j) const { + return nullary_wrapper<Scalar,NullaryOp, + has_nullary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value, + has_unary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value, + has_binary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value>().template packetOp<T>(op,i,j); + } + template <typename T, typename IndexType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i) const { + return nullary_wrapper<Scalar,NullaryOp, + has_nullary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value, + has_unary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value, + has_binary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value>().template packetOp<T>(op,i); + } +}; +#endif // MSVC workaround + +template<typename NullaryOp, typename PlainObjectType> +struct evaluator<CwiseNullaryOp<NullaryOp,PlainObjectType> > + : evaluator_base<CwiseNullaryOp<NullaryOp,PlainObjectType> > +{ + typedef CwiseNullaryOp<NullaryOp,PlainObjectType> XprType; + typedef typename internal::remove_all<PlainObjectType>::type PlainObjectTypeCleaned; + + enum { + CoeffReadCost = internal::functor_traits<NullaryOp>::Cost, + + Flags = (evaluator<PlainObjectTypeCleaned>::Flags + & ( HereditaryBits + | (functor_has_linear_access<NullaryOp>::ret ? LinearAccessBit : 0) + | (functor_traits<NullaryOp>::PacketAccess ? PacketAccessBit : 0))) + | (functor_traits<NullaryOp>::IsRepeatable ? 0 : EvalBeforeNestingBit), + Alignment = AlignedMax + }; + + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& n) + : m_functor(n.functor()), m_wrapper() + { + EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost); + } + + typedef typename XprType::CoeffReturnType CoeffReturnType; + + template <typename IndexType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(IndexType row, IndexType col) const + { + return m_wrapper(m_functor, row, col); + } + + template <typename IndexType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(IndexType index) const + { + return m_wrapper(m_functor,index); + } + + template<int LoadMode, typename PacketType, typename IndexType> + EIGEN_STRONG_INLINE + PacketType packet(IndexType row, IndexType col) const + { + return m_wrapper.template packetOp<PacketType>(m_functor, row, col); + } + + template<int LoadMode, typename PacketType, typename IndexType> + EIGEN_STRONG_INLINE + PacketType packet(IndexType index) const + { + return m_wrapper.template packetOp<PacketType>(m_functor, index); + } + +protected: + const NullaryOp m_functor; + const internal::nullary_wrapper<CoeffReturnType,NullaryOp> m_wrapper; +}; + +// -------------------- CwiseUnaryOp -------------------- + +template<typename UnaryOp, typename ArgType> +struct unary_evaluator<CwiseUnaryOp<UnaryOp, ArgType>, IndexBased > + : evaluator_base<CwiseUnaryOp<UnaryOp, ArgType> > +{ + typedef CwiseUnaryOp<UnaryOp, ArgType> XprType; + + enum { + CoeffReadCost = evaluator<ArgType>::CoeffReadCost + functor_traits<UnaryOp>::Cost, + + Flags = evaluator<ArgType>::Flags + & (HereditaryBits | LinearAccessBit | (functor_traits<UnaryOp>::PacketAccess ? PacketAccessBit : 0)), + Alignment = evaluator<ArgType>::Alignment + }; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit unary_evaluator(const XprType& op) + : m_functor(op.functor()), + m_argImpl(op.nestedExpression()) + { + EIGEN_INTERNAL_CHECK_COST_VALUE(functor_traits<UnaryOp>::Cost); + EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost); + } + + typedef typename XprType::CoeffReturnType CoeffReturnType; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index row, Index col) const + { + return m_functor(m_argImpl.coeff(row, col)); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index index) const + { + return m_functor(m_argImpl.coeff(index)); + } + + template<int LoadMode, typename PacketType> + EIGEN_STRONG_INLINE + PacketType packet(Index row, Index col) const + { + return m_functor.packetOp(m_argImpl.template packet<LoadMode, PacketType>(row, col)); + } + + template<int LoadMode, typename PacketType> + EIGEN_STRONG_INLINE + PacketType packet(Index index) const + { + return m_functor.packetOp(m_argImpl.template packet<LoadMode, PacketType>(index)); + } + +protected: + const UnaryOp m_functor; + evaluator<ArgType> m_argImpl; +}; + +// -------------------- CwiseTernaryOp -------------------- + +// this is a ternary expression +template<typename TernaryOp, typename Arg1, typename Arg2, typename Arg3> +struct evaluator<CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> > + : public ternary_evaluator<CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> > +{ + typedef CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> XprType; + typedef ternary_evaluator<CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> > Base; + + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr) : Base(xpr) {} +}; + +template<typename TernaryOp, typename Arg1, typename Arg2, typename Arg3> +struct ternary_evaluator<CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3>, IndexBased, IndexBased> + : evaluator_base<CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> > +{ + typedef CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> XprType; + + enum { + CoeffReadCost = evaluator<Arg1>::CoeffReadCost + evaluator<Arg2>::CoeffReadCost + evaluator<Arg3>::CoeffReadCost + functor_traits<TernaryOp>::Cost, + + Arg1Flags = evaluator<Arg1>::Flags, + Arg2Flags = evaluator<Arg2>::Flags, + Arg3Flags = evaluator<Arg3>::Flags, + SameType = is_same<typename Arg1::Scalar,typename Arg2::Scalar>::value && is_same<typename Arg1::Scalar,typename Arg3::Scalar>::value, + StorageOrdersAgree = (int(Arg1Flags)&RowMajorBit)==(int(Arg2Flags)&RowMajorBit) && (int(Arg1Flags)&RowMajorBit)==(int(Arg3Flags)&RowMajorBit), + Flags0 = (int(Arg1Flags) | int(Arg2Flags) | int(Arg3Flags)) & ( + HereditaryBits + | (int(Arg1Flags) & int(Arg2Flags) & int(Arg3Flags) & + ( (StorageOrdersAgree ? LinearAccessBit : 0) + | (functor_traits<TernaryOp>::PacketAccess && StorageOrdersAgree && SameType ? PacketAccessBit : 0) + ) + ) + ), + Flags = (Flags0 & ~RowMajorBit) | (Arg1Flags & RowMajorBit), + Alignment = EIGEN_PLAIN_ENUM_MIN( + EIGEN_PLAIN_ENUM_MIN(evaluator<Arg1>::Alignment, evaluator<Arg2>::Alignment), + evaluator<Arg3>::Alignment) + }; + + EIGEN_DEVICE_FUNC explicit ternary_evaluator(const XprType& xpr) + : m_functor(xpr.functor()), + m_arg1Impl(xpr.arg1()), + m_arg2Impl(xpr.arg2()), + m_arg3Impl(xpr.arg3()) + { + EIGEN_INTERNAL_CHECK_COST_VALUE(functor_traits<TernaryOp>::Cost); + EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost); + } + + typedef typename XprType::CoeffReturnType CoeffReturnType; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index row, Index col) const + { + return m_functor(m_arg1Impl.coeff(row, col), m_arg2Impl.coeff(row, col), m_arg3Impl.coeff(row, col)); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index index) const + { + return m_functor(m_arg1Impl.coeff(index), m_arg2Impl.coeff(index), m_arg3Impl.coeff(index)); + } + + template<int LoadMode, typename PacketType> + EIGEN_STRONG_INLINE + PacketType packet(Index row, Index col) const + { + return m_functor.packetOp(m_arg1Impl.template packet<LoadMode,PacketType>(row, col), + m_arg2Impl.template packet<LoadMode,PacketType>(row, col), + m_arg3Impl.template packet<LoadMode,PacketType>(row, col)); + } + + template<int LoadMode, typename PacketType> + EIGEN_STRONG_INLINE + PacketType packet(Index index) const + { + return m_functor.packetOp(m_arg1Impl.template packet<LoadMode,PacketType>(index), + m_arg2Impl.template packet<LoadMode,PacketType>(index), + m_arg3Impl.template packet<LoadMode,PacketType>(index)); + } + +protected: + const TernaryOp m_functor; + evaluator<Arg1> m_arg1Impl; + evaluator<Arg2> m_arg2Impl; + evaluator<Arg3> m_arg3Impl; +}; + +// -------------------- CwiseBinaryOp -------------------- + +// this is a binary expression +template<typename BinaryOp, typename Lhs, typename Rhs> +struct evaluator<CwiseBinaryOp<BinaryOp, Lhs, Rhs> > + : public binary_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) {} +}; + +template<typename BinaryOp, typename Lhs, typename Rhs> +struct binary_evaluator<CwiseBinaryOp<BinaryOp, Lhs, Rhs>, IndexBased, IndexBased> + : evaluator_base<CwiseBinaryOp<BinaryOp, Lhs, Rhs> > +{ + typedef CwiseBinaryOp<BinaryOp, Lhs, Rhs> XprType; + + enum { + CoeffReadCost = evaluator<Lhs>::CoeffReadCost + evaluator<Rhs>::CoeffReadCost + functor_traits<BinaryOp>::Cost, + + LhsFlags = evaluator<Lhs>::Flags, + RhsFlags = evaluator<Rhs>::Flags, + SameType = is_same<typename Lhs::Scalar,typename Rhs::Scalar>::value, + StorageOrdersAgree = (int(LhsFlags)&RowMajorBit)==(int(RhsFlags)&RowMajorBit), + Flags0 = (int(LhsFlags) | int(RhsFlags)) & ( + HereditaryBits + | (int(LhsFlags) & int(RhsFlags) & + ( (StorageOrdersAgree ? LinearAccessBit : 0) + | (functor_traits<BinaryOp>::PacketAccess && StorageOrdersAgree && SameType ? PacketAccessBit : 0) + ) + ) + ), + Flags = (Flags0 & ~RowMajorBit) | (LhsFlags & RowMajorBit), + Alignment = EIGEN_PLAIN_ENUM_MIN(evaluator<Lhs>::Alignment,evaluator<Rhs>::Alignment) + }; + + EIGEN_DEVICE_FUNC explicit binary_evaluator(const XprType& xpr) + : m_functor(xpr.functor()), + m_lhsImpl(xpr.lhs()), + m_rhsImpl(xpr.rhs()) + { + EIGEN_INTERNAL_CHECK_COST_VALUE(functor_traits<BinaryOp>::Cost); + EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost); + } + + typedef typename XprType::CoeffReturnType CoeffReturnType; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index row, Index col) const + { + return m_functor(m_lhsImpl.coeff(row, col), m_rhsImpl.coeff(row, col)); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index index) const + { + return m_functor(m_lhsImpl.coeff(index), m_rhsImpl.coeff(index)); + } + + template<int LoadMode, typename PacketType> + EIGEN_STRONG_INLINE + PacketType packet(Index row, Index col) const + { + return m_functor.packetOp(m_lhsImpl.template packet<LoadMode,PacketType>(row, col), + m_rhsImpl.template packet<LoadMode,PacketType>(row, col)); + } + + template<int LoadMode, typename PacketType> + EIGEN_STRONG_INLINE + PacketType packet(Index index) const + { + return m_functor.packetOp(m_lhsImpl.template packet<LoadMode,PacketType>(index), + m_rhsImpl.template packet<LoadMode,PacketType>(index)); + } + +protected: + const BinaryOp m_functor; + evaluator<Lhs> m_lhsImpl; + evaluator<Rhs> m_rhsImpl; +}; + +// -------------------- CwiseUnaryView -------------------- + +template<typename UnaryOp, typename ArgType> +struct unary_evaluator<CwiseUnaryView<UnaryOp, ArgType>, IndexBased> + : evaluator_base<CwiseUnaryView<UnaryOp, ArgType> > +{ + typedef CwiseUnaryView<UnaryOp, ArgType> XprType; + + enum { + CoeffReadCost = evaluator<ArgType>::CoeffReadCost + functor_traits<UnaryOp>::Cost, + + Flags = (evaluator<ArgType>::Flags & (HereditaryBits | LinearAccessBit | DirectAccessBit)), + + Alignment = 0 // FIXME it is not very clear why alignment is necessarily lost... + }; + + EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& op) + : m_unaryOp(op.functor()), + m_argImpl(op.nestedExpression()) + { + EIGEN_INTERNAL_CHECK_COST_VALUE(functor_traits<UnaryOp>::Cost); + EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost); + } + + typedef typename XprType::Scalar Scalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index row, Index col) const + { + return m_unaryOp(m_argImpl.coeff(row, col)); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index index) const + { + return m_unaryOp(m_argImpl.coeff(index)); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Scalar& coeffRef(Index row, Index col) + { + return m_unaryOp(m_argImpl.coeffRef(row, col)); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Scalar& coeffRef(Index index) + { + return m_unaryOp(m_argImpl.coeffRef(index)); + } + +protected: + const UnaryOp m_unaryOp; + evaluator<ArgType> m_argImpl; +}; + +// -------------------- Map -------------------- + +// FIXME perhaps the PlainObjectType could be provided by Derived::PlainObject ? +// but that might complicate template specialization +template<typename Derived, typename PlainObjectType> +struct mapbase_evaluator; + +template<typename Derived, typename PlainObjectType> +struct mapbase_evaluator : evaluator_base<Derived> +{ + typedef Derived XprType; + typedef typename XprType::PointerType PointerType; + typedef typename XprType::Scalar Scalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + + enum { + IsRowMajor = XprType::RowsAtCompileTime, + ColsAtCompileTime = XprType::ColsAtCompileTime, + CoeffReadCost = NumTraits<Scalar>::ReadCost + }; + + EIGEN_DEVICE_FUNC explicit mapbase_evaluator(const XprType& map) + : m_data(const_cast<PointerType>(map.data())), + m_innerStride(map.innerStride()), + m_outerStride(map.outerStride()) + { + EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(evaluator<Derived>::Flags&PacketAccessBit, internal::inner_stride_at_compile_time<Derived>::ret==1), + PACKET_ACCESS_REQUIRES_TO_HAVE_INNER_STRIDE_FIXED_TO_1); + EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index row, Index col) const + { + return m_data[col * colStride() + row * rowStride()]; + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index index) const + { + return m_data[index * m_innerStride.value()]; + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Scalar& coeffRef(Index row, Index col) + { + return m_data[col * colStride() + row * rowStride()]; + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Scalar& coeffRef(Index index) + { + return m_data[index * m_innerStride.value()]; + } + + template<int LoadMode, typename PacketType> + EIGEN_STRONG_INLINE + PacketType packet(Index row, Index col) const + { + PointerType ptr = m_data + row * rowStride() + col * colStride(); + return internal::ploadt<PacketType, LoadMode>(ptr); + } + + template<int LoadMode, typename PacketType> + EIGEN_STRONG_INLINE + PacketType packet(Index index) const + { + return internal::ploadt<PacketType, LoadMode>(m_data + index * m_innerStride.value()); + } + + template<int StoreMode, typename PacketType> + EIGEN_STRONG_INLINE + void writePacket(Index row, Index col, const PacketType& x) + { + PointerType ptr = m_data + row * rowStride() + col * colStride(); + return internal::pstoret<Scalar, PacketType, StoreMode>(ptr, x); + } + + template<int StoreMode, typename PacketType> + EIGEN_STRONG_INLINE + void writePacket(Index index, const PacketType& x) + { + 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(); } + + PointerType m_data; + const internal::variable_if_dynamic<Index, XprType::InnerStrideAtCompileTime> m_innerStride; + const internal::variable_if_dynamic<Index, XprType::OuterStrideAtCompileTime> m_outerStride; +}; + +template<typename PlainObjectType, int MapOptions, typename StrideType> +struct evaluator<Map<PlainObjectType, MapOptions, StrideType> > + : public mapbase_evaluator<Map<PlainObjectType, MapOptions, StrideType>, PlainObjectType> +{ + typedef Map<PlainObjectType, MapOptions, StrideType> XprType; + typedef typename XprType::Scalar Scalar; + // TODO: should check for smaller packet types once we can handle multi-sized packet types + typedef typename packet_traits<Scalar>::type PacketScalar; + + enum { + InnerStrideAtCompileTime = StrideType::InnerStrideAtCompileTime == 0 + ? int(PlainObjectType::InnerStrideAtCompileTime) + : int(StrideType::InnerStrideAtCompileTime), + OuterStrideAtCompileTime = StrideType::OuterStrideAtCompileTime == 0 + ? int(PlainObjectType::OuterStrideAtCompileTime) + : int(StrideType::OuterStrideAtCompileTime), + HasNoInnerStride = InnerStrideAtCompileTime == 1, + HasNoOuterStride = StrideType::OuterStrideAtCompileTime == 0, + HasNoStride = HasNoInnerStride && HasNoOuterStride, + IsDynamicSize = PlainObjectType::SizeAtCompileTime==Dynamic, + + PacketAccessMask = bool(HasNoInnerStride) ? ~int(0) : ~int(PacketAccessBit), + LinearAccessMask = bool(HasNoStride) || bool(PlainObjectType::IsVectorAtCompileTime) ? ~int(0) : ~int(LinearAccessBit), + Flags = int( evaluator<PlainObjectType>::Flags) & (LinearAccessMask&PacketAccessMask), + + Alignment = int(MapOptions)&int(AlignedMask) + }; + + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& map) + : mapbase_evaluator<XprType, PlainObjectType>(map) + { } +}; + +// -------------------- Ref -------------------- + +template<typename PlainObjectType, int RefOptions, typename StrideType> +struct evaluator<Ref<PlainObjectType, RefOptions, StrideType> > + : public mapbase_evaluator<Ref<PlainObjectType, RefOptions, StrideType>, PlainObjectType> +{ + typedef Ref<PlainObjectType, RefOptions, StrideType> XprType; + + enum { + Flags = evaluator<Map<PlainObjectType, RefOptions, StrideType> >::Flags, + Alignment = evaluator<Map<PlainObjectType, RefOptions, StrideType> >::Alignment + }; + + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& ref) + : mapbase_evaluator<XprType, PlainObjectType>(ref) + { } +}; + +// -------------------- Block -------------------- + +template<typename ArgType, int BlockRows, int BlockCols, bool InnerPanel, + bool HasDirectAccess = internal::has_direct_access<ArgType>::ret> struct block_evaluator; + +template<typename ArgType, int BlockRows, int BlockCols, bool InnerPanel> +struct evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel> > + : block_evaluator<ArgType, BlockRows, BlockCols, InnerPanel> +{ + typedef Block<ArgType, BlockRows, BlockCols, InnerPanel> XprType; + typedef typename XprType::Scalar Scalar; + // TODO: should check for smaller packet types once we can handle multi-sized packet types + typedef typename packet_traits<Scalar>::type PacketScalar; + + enum { + CoeffReadCost = evaluator<ArgType>::CoeffReadCost, + + RowsAtCompileTime = traits<XprType>::RowsAtCompileTime, + ColsAtCompileTime = traits<XprType>::ColsAtCompileTime, + MaxRowsAtCompileTime = traits<XprType>::MaxRowsAtCompileTime, + MaxColsAtCompileTime = traits<XprType>::MaxColsAtCompileTime, + + ArgTypeIsRowMajor = (int(evaluator<ArgType>::Flags)&RowMajorBit) != 0, + IsRowMajor = (MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1) ? 1 + : (MaxColsAtCompileTime==1 && MaxRowsAtCompileTime!=1) ? 0 + : ArgTypeIsRowMajor, + HasSameStorageOrderAsArgType = (IsRowMajor == ArgTypeIsRowMajor), + InnerSize = IsRowMajor ? int(ColsAtCompileTime) : int(RowsAtCompileTime), + InnerStrideAtCompileTime = HasSameStorageOrderAsArgType + ? int(inner_stride_at_compile_time<ArgType>::ret) + : int(outer_stride_at_compile_time<ArgType>::ret), + OuterStrideAtCompileTime = HasSameStorageOrderAsArgType + ? int(outer_stride_at_compile_time<ArgType>::ret) + : int(inner_stride_at_compile_time<ArgType>::ret), + MaskPacketAccessBit = (InnerStrideAtCompileTime == 1) ? PacketAccessBit : 0, + + FlagsLinearAccessBit = (RowsAtCompileTime == 1 || ColsAtCompileTime == 1 || (InnerPanel && (evaluator<ArgType>::Flags&LinearAccessBit))) ? LinearAccessBit : 0, + FlagsRowMajorBit = XprType::Flags&RowMajorBit, + Flags0 = evaluator<ArgType>::Flags & ( (HereditaryBits & ~RowMajorBit) | + DirectAccessBit | + MaskPacketAccessBit), + Flags = Flags0 | FlagsLinearAccessBit | FlagsRowMajorBit, + + PacketAlignment = unpacket_traits<PacketScalar>::alignment, + Alignment0 = (InnerPanel && (OuterStrideAtCompileTime!=Dynamic) && (((OuterStrideAtCompileTime * int(sizeof(Scalar))) % int(PacketAlignment)) == 0)) ? int(PacketAlignment) : 0, + 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_INTERNAL_CHECK_COST_VALUE(CoeffReadCost); + } +}; + +// no direct-access => dispatch to a unary evaluator +template<typename ArgType, int BlockRows, int BlockCols, bool InnerPanel> +struct block_evaluator<ArgType, BlockRows, BlockCols, InnerPanel, /*HasDirectAccess*/ false> + : unary_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel> > +{ + typedef Block<ArgType, BlockRows, BlockCols, InnerPanel> XprType; + + EIGEN_DEVICE_FUNC explicit block_evaluator(const XprType& block) + : unary_evaluator<XprType>(block) + {} +}; + +template<typename ArgType, int BlockRows, int BlockCols, bool InnerPanel> +struct unary_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel>, IndexBased> + : evaluator_base<Block<ArgType, BlockRows, BlockCols, InnerPanel> > +{ + typedef Block<ArgType, BlockRows, BlockCols, InnerPanel> XprType; + + EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& block) + : m_argImpl(block.nestedExpression()), + m_startRow(block.startRow()), + m_startCol(block.startCol()) + { } + + typedef typename XprType::Scalar Scalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + + enum { + RowsAtCompileTime = XprType::RowsAtCompileTime + }; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index row, Index col) const + { + return m_argImpl.coeff(m_startRow.value() + row, m_startCol.value() + col); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index index) const + { + return coeff(RowsAtCompileTime == 1 ? 0 : index, RowsAtCompileTime == 1 ? index : 0); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Scalar& coeffRef(Index row, Index col) + { + return m_argImpl.coeffRef(m_startRow.value() + row, m_startCol.value() + col); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Scalar& coeffRef(Index index) + { + return coeffRef(RowsAtCompileTime == 1 ? 0 : index, RowsAtCompileTime == 1 ? index : 0); + } + + template<int LoadMode, typename PacketType> + EIGEN_STRONG_INLINE + PacketType packet(Index row, Index col) const + { + return m_argImpl.template packet<LoadMode,PacketType>(m_startRow.value() + row, m_startCol.value() + col); + } + + template<int LoadMode, typename PacketType> + EIGEN_STRONG_INLINE + PacketType packet(Index index) const + { + return packet<LoadMode,PacketType>(RowsAtCompileTime == 1 ? 0 : index, + RowsAtCompileTime == 1 ? index : 0); + } + + template<int StoreMode, typename PacketType> + EIGEN_STRONG_INLINE + void writePacket(Index row, Index col, const PacketType& x) + { + return m_argImpl.template writePacket<StoreMode,PacketType>(m_startRow.value() + row, m_startCol.value() + col, x); + } + + template<int StoreMode, typename PacketType> + EIGEN_STRONG_INLINE + void writePacket(Index index, const PacketType& x) + { + return writePacket<StoreMode,PacketType>(RowsAtCompileTime == 1 ? 0 : index, + RowsAtCompileTime == 1 ? index : 0, + x); + } + +protected: + evaluator<ArgType> m_argImpl; + const variable_if_dynamic<Index, (ArgType::RowsAtCompileTime == 1 && BlockRows==1) ? 0 : Dynamic> m_startRow; + const variable_if_dynamic<Index, (ArgType::ColsAtCompileTime == 1 && BlockCols==1) ? 0 : Dynamic> m_startCol; +}; + +// TODO: This evaluator does not actually use the child evaluator; +// all action is via the data() as returned by the Block expression. + +template<typename ArgType, int BlockRows, int BlockCols, bool InnerPanel> +struct block_evaluator<ArgType, BlockRows, BlockCols, InnerPanel, /* HasDirectAccess */ true> + : mapbase_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel>, + typename Block<ArgType, BlockRows, BlockCols, InnerPanel>::PlainObject> +{ + typedef Block<ArgType, BlockRows, BlockCols, InnerPanel> XprType; + typedef typename XprType::Scalar Scalar; + + EIGEN_DEVICE_FUNC 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 + eigen_assert(((internal::UIntPtr(block.data()) % EIGEN_PLAIN_ENUM_MAX(1,evaluator<XprType>::Alignment)) == 0) && "data is not aligned"); + } +}; + + +// -------------------- Select -------------------- +// NOTE shall we introduce a ternary_evaluator? + +// TODO enable vectorization for Select +template<typename ConditionMatrixType, typename ThenMatrixType, typename ElseMatrixType> +struct evaluator<Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> > + : evaluator_base<Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> > +{ + typedef Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> XprType; + enum { + CoeffReadCost = evaluator<ConditionMatrixType>::CoeffReadCost + + EIGEN_PLAIN_ENUM_MAX(evaluator<ThenMatrixType>::CoeffReadCost, + evaluator<ElseMatrixType>::CoeffReadCost), + + Flags = (unsigned int)evaluator<ThenMatrixType>::Flags & evaluator<ElseMatrixType>::Flags & HereditaryBits, + + Alignment = EIGEN_PLAIN_ENUM_MIN(evaluator<ThenMatrixType>::Alignment, evaluator<ElseMatrixType>::Alignment) + }; + + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& select) + : m_conditionImpl(select.conditionMatrix()), + m_thenImpl(select.thenMatrix()), + m_elseImpl(select.elseMatrix()) + { + EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost); + } + + typedef typename XprType::CoeffReturnType CoeffReturnType; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index row, Index col) const + { + if (m_conditionImpl.coeff(row, col)) + return m_thenImpl.coeff(row, col); + else + return m_elseImpl.coeff(row, col); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index index) const + { + if (m_conditionImpl.coeff(index)) + return m_thenImpl.coeff(index); + else + return m_elseImpl.coeff(index); + } + +protected: + evaluator<ConditionMatrixType> m_conditionImpl; + evaluator<ThenMatrixType> m_thenImpl; + evaluator<ElseMatrixType> m_elseImpl; +}; + + +// -------------------- Replicate -------------------- + +template<typename ArgType, int RowFactor, int ColFactor> +struct unary_evaluator<Replicate<ArgType, RowFactor, ColFactor> > + : evaluator_base<Replicate<ArgType, RowFactor, ColFactor> > +{ + typedef Replicate<ArgType, RowFactor, ColFactor> XprType; + typedef typename XprType::CoeffReturnType CoeffReturnType; + enum { + Factor = (RowFactor==Dynamic || ColFactor==Dynamic) ? Dynamic : RowFactor*ColFactor + }; + typedef typename internal::nested_eval<ArgType,Factor>::type ArgTypeNested; + typedef typename internal::remove_all<ArgTypeNested>::type ArgTypeNestedCleaned; + + enum { + CoeffReadCost = evaluator<ArgTypeNestedCleaned>::CoeffReadCost, + LinearAccessMask = XprType::IsVectorAtCompileTime ? LinearAccessBit : 0, + Flags = (evaluator<ArgTypeNestedCleaned>::Flags & (HereditaryBits|LinearAccessMask) & ~RowMajorBit) | (traits<XprType>::Flags & RowMajorBit), + + Alignment = evaluator<ArgTypeNestedCleaned>::Alignment + }; + + EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& replicate) + : m_arg(replicate.nestedExpression()), + m_argImpl(m_arg), + m_rows(replicate.nestedExpression().rows()), + m_cols(replicate.nestedExpression().cols()) + {} + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index row, Index col) const + { + // try to avoid using modulo; this is a pure optimization strategy + const Index actual_row = internal::traits<XprType>::RowsAtCompileTime==1 ? 0 + : RowFactor==1 ? row + : row % m_rows.value(); + const Index actual_col = internal::traits<XprType>::ColsAtCompileTime==1 ? 0 + : ColFactor==1 ? col + : col % m_cols.value(); + + return m_argImpl.coeff(actual_row, actual_col); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index index) const + { + // try to avoid using modulo; this is a pure optimization strategy + const Index actual_index = internal::traits<XprType>::RowsAtCompileTime==1 + ? (ColFactor==1 ? index : index%m_cols.value()) + : (RowFactor==1 ? index : index%m_rows.value()); + + return m_argImpl.coeff(actual_index); + } + + template<int LoadMode, typename PacketType> + EIGEN_STRONG_INLINE + PacketType packet(Index row, Index col) const + { + const Index actual_row = internal::traits<XprType>::RowsAtCompileTime==1 ? 0 + : RowFactor==1 ? row + : row % m_rows.value(); + const Index actual_col = internal::traits<XprType>::ColsAtCompileTime==1 ? 0 + : ColFactor==1 ? col + : col % m_cols.value(); + + return m_argImpl.template packet<LoadMode,PacketType>(actual_row, actual_col); + } + + template<int LoadMode, typename PacketType> + EIGEN_STRONG_INLINE + PacketType packet(Index index) const + { + const Index actual_index = internal::traits<XprType>::RowsAtCompileTime==1 + ? (ColFactor==1 ? index : index%m_cols.value()) + : (RowFactor==1 ? index : index%m_rows.value()); + + return m_argImpl.template packet<LoadMode,PacketType>(actual_index); + } + +protected: + const ArgTypeNested m_arg; + evaluator<ArgTypeNestedCleaned> m_argImpl; + const variable_if_dynamic<Index, ArgType::RowsAtCompileTime> m_rows; + const variable_if_dynamic<Index, ArgType::ColsAtCompileTime> m_cols; +}; + + +// -------------------- PartialReduxExpr -------------------- + +template< typename ArgType, typename MemberOp, int Direction> +struct evaluator<PartialReduxExpr<ArgType, MemberOp, Direction> > + : evaluator_base<PartialReduxExpr<ArgType, MemberOp, Direction> > +{ + typedef PartialReduxExpr<ArgType, MemberOp, Direction> XprType; + typedef typename internal::nested_eval<ArgType,1>::type ArgTypeNested; + typedef typename internal::remove_all<ArgTypeNested>::type ArgTypeNestedCleaned; + typedef typename ArgType::Scalar InputScalar; + typedef typename XprType::Scalar Scalar; + enum { + TraversalSize = Direction==int(Vertical) ? int(ArgType::RowsAtCompileTime) : int(ArgType::ColsAtCompileTime) + }; + typedef typename MemberOp::template Cost<InputScalar,int(TraversalSize)> CostOpType; + enum { + CoeffReadCost = TraversalSize==Dynamic ? HugeCost + : TraversalSize * evaluator<ArgType>::CoeffReadCost + int(CostOpType::value), + + Flags = (traits<XprType>::Flags&RowMajorBit) | (evaluator<ArgType>::Flags&(HereditaryBits&(~RowMajorBit))) | LinearAccessBit, + + Alignment = 0 // FIXME this will need to be improved once PartialReduxExpr is vectorized + }; + + EIGEN_DEVICE_FUNC explicit evaluator(const XprType xpr) + : m_arg(xpr.nestedExpression()), m_functor(xpr.functor()) + { + EIGEN_INTERNAL_CHECK_COST_VALUE(TraversalSize==Dynamic ? HugeCost : int(CostOpType::value)); + EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost); + } + + typedef typename XprType::CoeffReturnType CoeffReturnType; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const Scalar coeff(Index i, Index j) const + { + if (Direction==Vertical) + return m_functor(m_arg.col(j)); + else + return m_functor(m_arg.row(i)); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const Scalar coeff(Index index) const + { + if (Direction==Vertical) + return m_functor(m_arg.col(index)); + else + return m_functor(m_arg.row(index)); + } + +protected: + typename internal::add_const_on_value_type<ArgTypeNested>::type m_arg; + const MemberOp m_functor; +}; + + +// -------------------- MatrixWrapper and ArrayWrapper -------------------- +// +// evaluator_wrapper_base<T> is a common base class for the +// MatrixWrapper and ArrayWrapper evaluators. + +template<typename XprType> +struct evaluator_wrapper_base + : evaluator_base<XprType> +{ + typedef typename remove_all<typename XprType::NestedExpressionType>::type ArgType; + enum { + CoeffReadCost = evaluator<ArgType>::CoeffReadCost, + Flags = evaluator<ArgType>::Flags, + Alignment = evaluator<ArgType>::Alignment + }; + + EIGEN_DEVICE_FUNC explicit evaluator_wrapper_base(const ArgType& arg) : m_argImpl(arg) {} + + typedef typename ArgType::Scalar Scalar; + typedef typename ArgType::CoeffReturnType CoeffReturnType; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index row, Index col) const + { + return m_argImpl.coeff(row, col); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index index) const + { + return m_argImpl.coeff(index); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Scalar& coeffRef(Index row, Index col) + { + return m_argImpl.coeffRef(row, col); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Scalar& coeffRef(Index index) + { + return m_argImpl.coeffRef(index); + } + + template<int LoadMode, typename PacketType> + EIGEN_STRONG_INLINE + PacketType packet(Index row, Index col) const + { + return m_argImpl.template packet<LoadMode,PacketType>(row, col); + } + + template<int LoadMode, typename PacketType> + EIGEN_STRONG_INLINE + PacketType packet(Index index) const + { + return m_argImpl.template packet<LoadMode,PacketType>(index); + } + + template<int StoreMode, typename PacketType> + EIGEN_STRONG_INLINE + void writePacket(Index row, Index col, const PacketType& x) + { + m_argImpl.template writePacket<StoreMode>(row, col, x); + } + + template<int StoreMode, typename PacketType> + EIGEN_STRONG_INLINE + void writePacket(Index index, const PacketType& x) + { + m_argImpl.template writePacket<StoreMode>(index, x); + } + +protected: + evaluator<ArgType> m_argImpl; +}; + +template<typename TArgType> +struct unary_evaluator<MatrixWrapper<TArgType> > + : evaluator_wrapper_base<MatrixWrapper<TArgType> > +{ + typedef MatrixWrapper<TArgType> XprType; + + EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& wrapper) + : evaluator_wrapper_base<MatrixWrapper<TArgType> >(wrapper.nestedExpression()) + { } +}; + +template<typename TArgType> +struct unary_evaluator<ArrayWrapper<TArgType> > + : evaluator_wrapper_base<ArrayWrapper<TArgType> > +{ + typedef ArrayWrapper<TArgType> XprType; + + EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& wrapper) + : evaluator_wrapper_base<ArrayWrapper<TArgType> >(wrapper.nestedExpression()) + { } +}; + + +// -------------------- Reverse -------------------- + +// defined in Reverse.h: +template<typename PacketType, bool ReversePacket> struct reverse_packet_cond; + +template<typename ArgType, int Direction> +struct unary_evaluator<Reverse<ArgType, Direction> > + : evaluator_base<Reverse<ArgType, Direction> > +{ + typedef Reverse<ArgType, Direction> XprType; + typedef typename XprType::Scalar Scalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + + enum { + IsRowMajor = XprType::IsRowMajor, + IsColMajor = !IsRowMajor, + ReverseRow = (Direction == Vertical) || (Direction == BothDirections), + ReverseCol = (Direction == Horizontal) || (Direction == BothDirections), + ReversePacket = (Direction == BothDirections) + || ((Direction == Vertical) && IsColMajor) + || ((Direction == Horizontal) && IsRowMajor), + + CoeffReadCost = evaluator<ArgType>::CoeffReadCost, + + // let's enable LinearAccess only with vectorization because of the product overhead + // FIXME enable DirectAccess with negative strides? + Flags0 = evaluator<ArgType>::Flags, + LinearAccess = ( (Direction==BothDirections) && (int(Flags0)&PacketAccessBit) ) + || ((ReverseRow && XprType::ColsAtCompileTime==1) || (ReverseCol && XprType::RowsAtCompileTime==1)) + ? LinearAccessBit : 0, + + Flags = int(Flags0) & (HereditaryBits | PacketAccessBit | LinearAccess), + + Alignment = 0 // FIXME in some rare cases, Alignment could be preserved, like a Vector4f. + }; + + EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& reverse) + : m_argImpl(reverse.nestedExpression()), + m_rows(ReverseRow ? reverse.nestedExpression().rows() : 1), + m_cols(ReverseCol ? reverse.nestedExpression().cols() : 1) + { } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index row, Index col) const + { + return m_argImpl.coeff(ReverseRow ? m_rows.value() - row - 1 : row, + ReverseCol ? m_cols.value() - col - 1 : col); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index index) const + { + return m_argImpl.coeff(m_rows.value() * m_cols.value() - index - 1); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Scalar& coeffRef(Index row, Index col) + { + return m_argImpl.coeffRef(ReverseRow ? m_rows.value() - row - 1 : row, + ReverseCol ? m_cols.value() - col - 1 : col); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Scalar& coeffRef(Index index) + { + return m_argImpl.coeffRef(m_rows.value() * m_cols.value() - index - 1); + } + + template<int LoadMode, typename PacketType> + EIGEN_STRONG_INLINE + PacketType packet(Index row, Index col) const + { + enum { + PacketSize = unpacket_traits<PacketType>::size, + OffsetRow = ReverseRow && IsColMajor ? PacketSize : 1, + OffsetCol = ReverseCol && IsRowMajor ? PacketSize : 1 + }; + typedef internal::reverse_packet_cond<PacketType,ReversePacket> reverse_packet; + return reverse_packet::run(m_argImpl.template packet<LoadMode,PacketType>( + ReverseRow ? m_rows.value() - row - OffsetRow : row, + ReverseCol ? m_cols.value() - col - OffsetCol : col)); + } + + template<int LoadMode, typename PacketType> + EIGEN_STRONG_INLINE + PacketType packet(Index index) const + { + enum { PacketSize = unpacket_traits<PacketType>::size }; + return preverse(m_argImpl.template packet<LoadMode,PacketType>(m_rows.value() * m_cols.value() - index - PacketSize)); + } + + template<int LoadMode, typename PacketType> + EIGEN_STRONG_INLINE + void writePacket(Index row, Index col, const PacketType& x) + { + // FIXME we could factorize some code with packet(i,j) + enum { + PacketSize = unpacket_traits<PacketType>::size, + OffsetRow = ReverseRow && IsColMajor ? PacketSize : 1, + OffsetCol = ReverseCol && IsRowMajor ? PacketSize : 1 + }; + typedef internal::reverse_packet_cond<PacketType,ReversePacket> reverse_packet; + m_argImpl.template writePacket<LoadMode>( + ReverseRow ? m_rows.value() - row - OffsetRow : row, + ReverseCol ? m_cols.value() - col - OffsetCol : col, + reverse_packet::run(x)); + } + + template<int LoadMode, typename PacketType> + EIGEN_STRONG_INLINE + void writePacket(Index index, const PacketType& x) + { + enum { PacketSize = unpacket_traits<PacketType>::size }; + m_argImpl.template writePacket<LoadMode> + (m_rows.value() * m_cols.value() - index - PacketSize, preverse(x)); + } + +protected: + evaluator<ArgType> m_argImpl; + + // If we do not reverse rows, then we do not need to know the number of rows; same for columns + // Nonetheless, in this case it is important to set to 1 such that the coeff(index) method works fine for vectors. + const variable_if_dynamic<Index, ReverseRow ? ArgType::RowsAtCompileTime : 1> m_rows; + const variable_if_dynamic<Index, ReverseCol ? ArgType::ColsAtCompileTime : 1> m_cols; +}; + + +// -------------------- Diagonal -------------------- + +template<typename ArgType, int DiagIndex> +struct evaluator<Diagonal<ArgType, DiagIndex> > + : evaluator_base<Diagonal<ArgType, DiagIndex> > +{ + typedef Diagonal<ArgType, DiagIndex> XprType; + + enum { + CoeffReadCost = evaluator<ArgType>::CoeffReadCost, + + Flags = (unsigned int)(evaluator<ArgType>::Flags & (HereditaryBits | DirectAccessBit) & ~RowMajorBit) | LinearAccessBit, + + Alignment = 0 + }; + + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& diagonal) + : m_argImpl(diagonal.nestedExpression()), + m_index(diagonal.index()) + { } + + typedef typename XprType::Scalar Scalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index row, Index) const + { + return m_argImpl.coeff(row + rowOffset(), row + colOffset()); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + CoeffReturnType coeff(Index index) const + { + return m_argImpl.coeff(index + rowOffset(), index + colOffset()); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Scalar& coeffRef(Index row, Index) + { + return m_argImpl.coeffRef(row + rowOffset(), row + colOffset()); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Scalar& coeffRef(Index index) + { + return m_argImpl.coeffRef(index + rowOffset(), index + colOffset()); + } + +protected: + evaluator<ArgType> m_argImpl; + const internal::variable_if_dynamicindex<Index, XprType::DiagIndex> m_index; + +private: + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index rowOffset() const { return m_index.value() > 0 ? 0 : -m_index.value(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index colOffset() const { return m_index.value() > 0 ? m_index.value() : 0; } +}; + + +//---------------------------------------------------------------------- +// deprecated code +//---------------------------------------------------------------------- + +// -------------------- EvalToTemp -------------------- + +// expression class for evaluating nested expression to a temporary + +template<typename ArgType> class EvalToTemp; + +template<typename ArgType> +struct traits<EvalToTemp<ArgType> > + : public traits<ArgType> +{ }; + +template<typename ArgType> +class EvalToTemp + : public dense_xpr_base<EvalToTemp<ArgType> >::type +{ + public: + + typedef typename dense_xpr_base<EvalToTemp>::type Base; + EIGEN_GENERIC_PUBLIC_INTERFACE(EvalToTemp) + + explicit EvalToTemp(const ArgType& arg) + : m_arg(arg) + { } + + const ArgType& arg() const + { + return m_arg; + } + + Index rows() const + { + return m_arg.rows(); + } + + Index cols() const + { + return m_arg.cols(); + } + + private: + const ArgType& m_arg; +}; + +template<typename ArgType> +struct evaluator<EvalToTemp<ArgType> > + : public evaluator<typename ArgType::PlainObject> +{ + typedef EvalToTemp<ArgType> XprType; + typedef typename ArgType::PlainObject PlainObject; + typedef evaluator<PlainObject> Base; + + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr) + : m_result(xpr.arg()) + { + ::new (static_cast<Base*>(this)) Base(m_result); + } + + // This constructor is used when nesting an EvalTo evaluator in another evaluator + EIGEN_DEVICE_FUNC evaluator(const ArgType& arg) + : m_result(arg) + { + ::new (static_cast<Base*>(this)) Base(m_result); + } + +protected: + PlainObject m_result; +}; + +} // namespace internal + +} // end namespace Eigen + +#endif // EIGEN_COREEVALUATORS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/CoreIterators.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/CoreIterators.h new file mode 100644 index 000000000..4eb42b93a --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/CoreIterators.h @@ -0,0 +1,127 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2014 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_COREITERATORS_H +#define EIGEN_COREITERATORS_H + +namespace Eigen { + +/* This file contains the respective InnerIterator definition of the expressions defined in Eigen/Core + */ + +namespace internal { + +template<typename XprType, typename EvaluatorKind> +class inner_iterator_selector; + +} + +/** \class InnerIterator + * \brief An InnerIterator allows to loop over the element of any matrix expression. + * + * \warning To be used with care because an evaluator is constructed every time an InnerIterator iterator is constructed. + * + * TODO: add a usage example + */ +template<typename XprType> +class InnerIterator +{ +protected: + typedef internal::inner_iterator_selector<XprType, typename internal::evaluator_traits<XprType>::Kind> IteratorType; + typedef internal::evaluator<XprType> EvaluatorType; + typedef typename internal::traits<XprType>::Scalar Scalar; +public: + /** Construct an iterator over the \a outerId -th row or column of \a xpr */ + InnerIterator(const XprType &xpr, const Index &outerId) + : m_eval(xpr), m_iter(m_eval, outerId, xpr.innerSize()) + {} + + /// \returns the value of the current coefficient. + EIGEN_STRONG_INLINE Scalar value() const { return m_iter.value(); } + /** Increment the iterator \c *this to the next non-zero coefficient. + * Explicit zeros are not skipped over. To skip explicit zeros, see class SparseView + */ + EIGEN_STRONG_INLINE InnerIterator& operator++() { m_iter.operator++(); return *this; } + /// \returns the column or row index of the current coefficient. + EIGEN_STRONG_INLINE Index index() const { return m_iter.index(); } + /// \returns the row index of the current coefficient. + EIGEN_STRONG_INLINE Index row() const { return m_iter.row(); } + /// \returns the column index of the current coefficient. + EIGEN_STRONG_INLINE Index col() const { return m_iter.col(); } + /// \returns \c true if the iterator \c *this still references a valid coefficient. + EIGEN_STRONG_INLINE operator bool() const { return m_iter; } + +protected: + EvaluatorType m_eval; + IteratorType m_iter; +private: + // If you get here, then you're not using the right InnerIterator type, e.g.: + // SparseMatrix<double,RowMajor> A; + // SparseMatrix<double>::InnerIterator it(A,0); + template<typename T> InnerIterator(const EigenBase<T>&,Index outer); +}; + +namespace internal { + +// Generic inner iterator implementation for dense objects +template<typename XprType> +class inner_iterator_selector<XprType, IndexBased> +{ +protected: + typedef evaluator<XprType> EvaluatorType; + typedef typename traits<XprType>::Scalar Scalar; + enum { IsRowMajor = (XprType::Flags&RowMajorBit)==RowMajorBit }; + +public: + EIGEN_STRONG_INLINE inner_iterator_selector(const EvaluatorType &eval, const Index &outerId, const Index &innerSize) + : m_eval(eval), m_inner(0), m_outer(outerId), m_end(innerSize) + {} + + EIGEN_STRONG_INLINE Scalar value() const + { + return (IsRowMajor) ? m_eval.coeff(m_outer, m_inner) + : m_eval.coeff(m_inner, m_outer); + } + + EIGEN_STRONG_INLINE inner_iterator_selector& operator++() { m_inner++; return *this; } + + EIGEN_STRONG_INLINE Index index() const { return m_inner; } + inline Index row() const { return IsRowMajor ? m_outer : index(); } + inline Index col() const { return IsRowMajor ? index() : m_outer; } + + EIGEN_STRONG_INLINE operator bool() const { return m_inner < m_end && m_inner>=0; } + +protected: + const EvaluatorType& m_eval; + Index m_inner; + const Index m_outer; + const Index m_end; +}; + +// For iterator-based evaluator, inner-iterator is already implemented as +// evaluator<>::InnerIterator +template<typename XprType> +class inner_iterator_selector<XprType, IteratorBased> + : public evaluator<XprType>::InnerIterator +{ +protected: + typedef typename evaluator<XprType>::InnerIterator Base; + typedef evaluator<XprType> EvaluatorType; + +public: + EIGEN_STRONG_INLINE inner_iterator_selector(const EvaluatorType &eval, const Index &outerId, const Index &/*innerSize*/) + : Base(eval, outerId) + {} +}; + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_COREITERATORS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/CwiseBinaryOp.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/CwiseBinaryOp.h new file mode 100644 index 000000000..a36765e39 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/CwiseBinaryOp.h @@ -0,0 +1,184 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2014 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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_CWISE_BINARY_OP_H +#define EIGEN_CWISE_BINARY_OP_H + +namespace Eigen { + +namespace internal { +template<typename BinaryOp, typename Lhs, typename Rhs> +struct traits<CwiseBinaryOp<BinaryOp, Lhs, Rhs> > +{ + // we must not inherit from traits<Lhs> since it has + // the potential to cause problems with MSVC + typedef typename remove_all<Lhs>::type Ancestor; + typedef typename traits<Ancestor>::XprKind XprKind; + enum { + RowsAtCompileTime = traits<Ancestor>::RowsAtCompileTime, + ColsAtCompileTime = traits<Ancestor>::ColsAtCompileTime, + MaxRowsAtCompileTime = traits<Ancestor>::MaxRowsAtCompileTime, + MaxColsAtCompileTime = traits<Ancestor>::MaxColsAtCompileTime + }; + + // even though we require Lhs and Rhs to have the same scalar type (see CwiseBinaryOp constructor), + // we still want to handle the case when the result type is different. + typedef typename result_of< + BinaryOp( + const typename Lhs::Scalar&, + const typename Rhs::Scalar& + ) + >::type Scalar; + typedef typename cwise_promote_storage_type<typename traits<Lhs>::StorageKind, + typename traits<Rhs>::StorageKind, + BinaryOp>::ret StorageKind; + typedef typename promote_index_type<typename traits<Lhs>::StorageIndex, + typename traits<Rhs>::StorageIndex>::type StorageIndex; + typedef typename Lhs::Nested LhsNested; + typedef typename Rhs::Nested RhsNested; + typedef typename remove_reference<LhsNested>::type _LhsNested; + typedef typename remove_reference<RhsNested>::type _RhsNested; + enum { + Flags = cwise_promote_storage_order<typename traits<Lhs>::StorageKind,typename traits<Rhs>::StorageKind,_LhsNested::Flags & RowMajorBit,_RhsNested::Flags & RowMajorBit>::value + }; +}; +} // end namespace internal + +template<typename BinaryOp, typename Lhs, typename Rhs, typename StorageKind> +class CwiseBinaryOpImpl; + +/** \class CwiseBinaryOp + * \ingroup Core_Module + * + * \brief Generic expression where a coefficient-wise binary operator is applied to two expressions + * + * \tparam BinaryOp template functor implementing the operator + * \tparam LhsType the type of the left-hand side + * \tparam RhsType the type of the right-hand side + * + * This class represents an expression where a coefficient-wise binary operator is applied to two expressions. + * It is the return type of binary operators, by which we mean only those binary operators where + * both the left-hand side and the right-hand side are Eigen expressions. + * For example, the return type of matrix1+matrix2 is a CwiseBinaryOp. + * + * Most of the time, this is the only way that it is used, so you typically don't have to name + * CwiseBinaryOp types explicitly. + * + * \sa MatrixBase::binaryExpr(const MatrixBase<OtherDerived> &,const CustomBinaryOp &) const, class CwiseUnaryOp, class CwiseNullaryOp + */ +template<typename BinaryOp, typename LhsType, typename RhsType> +class CwiseBinaryOp : + public CwiseBinaryOpImpl< + BinaryOp, LhsType, RhsType, + typename internal::cwise_promote_storage_type<typename internal::traits<LhsType>::StorageKind, + typename internal::traits<RhsType>::StorageKind, + BinaryOp>::ret>, + internal::no_assignment_operator +{ + public: + + typedef typename internal::remove_all<BinaryOp>::type Functor; + typedef typename internal::remove_all<LhsType>::type Lhs; + typedef typename internal::remove_all<RhsType>::type Rhs; + + typedef typename CwiseBinaryOpImpl< + BinaryOp, LhsType, RhsType, + typename internal::cwise_promote_storage_type<typename internal::traits<LhsType>::StorageKind, + typename internal::traits<Rhs>::StorageKind, + BinaryOp>::ret>::Base Base; + EIGEN_GENERIC_PUBLIC_INTERFACE(CwiseBinaryOp) + + typedef typename internal::ref_selector<LhsType>::type LhsNested; + typedef typename internal::ref_selector<RhsType>::type RhsNested; + 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()) + : m_lhs(aLhs), m_rhs(aRhs), m_functor(func) + { + EIGEN_CHECK_BINARY_COMPATIBILIY(BinaryOp,typename Lhs::Scalar,typename Rhs::Scalar); + // require the sizes to match + EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Lhs, Rhs) + eigen_assert(aLhs.rows() == aRhs.rows() && aLhs.cols() == aRhs.cols()); + } + + 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 { + // 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(); + else + return m_lhs.cols(); + } + + /** \returns the left hand side nested expression */ + EIGEN_DEVICE_FUNC + const _LhsNested& lhs() const { return m_lhs; } + /** \returns the right hand side nested expression */ + EIGEN_DEVICE_FUNC + const _RhsNested& rhs() const { return m_rhs; } + /** \returns the functor representing the binary operation */ + EIGEN_DEVICE_FUNC + const BinaryOp& functor() const { return m_functor; } + + protected: + LhsNested m_lhs; + RhsNested m_rhs; + const BinaryOp m_functor; +}; + +// Generic API dispatcher +template<typename BinaryOp, typename Lhs, typename Rhs, typename StorageKind> +class CwiseBinaryOpImpl + : public internal::generic_xpr_base<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >::type +{ +public: + typedef typename internal::generic_xpr_base<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >::type Base; +}; + +/** replaces \c *this by \c *this - \a other. + * + * \returns a reference to \c *this + */ +template<typename Derived> +template<typename OtherDerived> +EIGEN_STRONG_INLINE Derived & +MatrixBase<Derived>::operator-=(const MatrixBase<OtherDerived> &other) +{ + call_assignment(derived(), other.derived(), internal::sub_assign_op<Scalar,typename OtherDerived::Scalar>()); + return derived(); +} + +/** replaces \c *this by \c *this + \a other. + * + * \returns a reference to \c *this + */ +template<typename Derived> +template<typename OtherDerived> +EIGEN_STRONG_INLINE Derived & +MatrixBase<Derived>::operator+=(const MatrixBase<OtherDerived>& other) +{ + call_assignment(derived(), other.derived(), internal::add_assign_op<Scalar,typename OtherDerived::Scalar>()); + return derived(); +} + +} // end namespace Eigen + +#endif // EIGEN_CWISE_BINARY_OP_H + diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/CwiseNullaryOp.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/CwiseNullaryOp.h new file mode 100644 index 000000000..ddd607e38 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/CwiseNullaryOp.h @@ -0,0 +1,866 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2010 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_CWISE_NULLARY_OP_H +#define EIGEN_CWISE_NULLARY_OP_H + +namespace Eigen { + +namespace internal { +template<typename NullaryOp, typename PlainObjectType> +struct traits<CwiseNullaryOp<NullaryOp, PlainObjectType> > : traits<PlainObjectType> +{ + enum { + Flags = traits<PlainObjectType>::Flags & RowMajorBit + }; +}; + +} // namespace internal + +/** \class CwiseNullaryOp + * \ingroup Core_Module + * + * \brief Generic expression of a matrix where all coefficients are defined by a functor + * + * \tparam NullaryOp template functor implementing the operator + * \tparam PlainObjectType the underlying plain matrix/array type + * + * This class represents an expression of a generic nullary operator. + * It is the return type of the Ones(), Zero(), Constant(), Identity() and Random() methods, + * and most of the time this is the only way it is used. + * + * However, if you want to write a function returning such an expression, you + * will need to use this class. + * + * The functor NullaryOp must expose one of the following method: + <table class="manual"> + <tr ><td>\c operator()() </td><td>if the procedural generation does not depend on the coefficient entries (e.g., random numbers)</td></tr> + <tr class="alt"><td>\c operator()(Index i)</td><td>if the procedural generation makes sense for vectors only and that it depends on the coefficient index \c i (e.g., linspace) </td></tr> + <tr ><td>\c operator()(Index i,Index j)</td><td>if the procedural generation depends on the matrix coordinates \c i, \c j (e.g., to generate a checkerboard with 0 and 1)</td></tr> + </table> + * It is also possible to expose the last two operators if the generation makes sense for matrices but can be optimized for vectors. + * + * See DenseBase::NullaryExpr(Index,const CustomNullaryOp&) for an example binding + * C++11 random number generators. + * + * A nullary expression can also be used to implement custom sophisticated matrix manipulations + * that cannot be covered by the existing set of natively supported matrix manipulations. + * See this \ref TopicCustomizing_NullaryExpr "page" for some examples and additional explanations + * on the behavior of CwiseNullaryOp. + * + * \sa class CwiseUnaryOp, class CwiseBinaryOp, DenseBase::NullaryExpr + */ +template<typename NullaryOp, typename PlainObjectType> +class CwiseNullaryOp : public internal::dense_xpr_base< CwiseNullaryOp<NullaryOp, PlainObjectType> >::type, internal::no_assignment_operator +{ + public: + + typedef typename internal::dense_xpr_base<CwiseNullaryOp>::type Base; + EIGEN_DENSE_PUBLIC_INTERFACE(CwiseNullaryOp) + + EIGEN_DEVICE_FUNC + CwiseNullaryOp(Index rows, Index cols, const NullaryOp& func = NullaryOp()) + : m_rows(rows), m_cols(cols), m_functor(func) + { + eigen_assert(rows >= 0 + && (RowsAtCompileTime == Dynamic || RowsAtCompileTime == rows) + && cols >= 0 + && (ColsAtCompileTime == Dynamic || ColsAtCompileTime == cols)); + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Index rows() const { return m_rows.value(); } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Index cols() const { return m_cols.value(); } + + /** \returns the functor representing the nullary operation */ + EIGEN_DEVICE_FUNC + const NullaryOp& functor() const { return m_functor; } + + protected: + const internal::variable_if_dynamic<Index, RowsAtCompileTime> m_rows; + const internal::variable_if_dynamic<Index, ColsAtCompileTime> m_cols; + const NullaryOp m_functor; +}; + + +/** \returns an expression of a matrix defined by a custom functor \a func + * + * The parameters \a rows and \a cols are the number of rows and of columns of + * the returned matrix. Must be compatible with this MatrixBase type. + * + * This variant is meant to be used for dynamic-size matrix types. For fixed-size types, + * it is redundant to pass \a rows and \a cols as arguments, so Zero() should be used + * instead. + * + * The template parameter \a CustomNullaryOp is the type of the functor. + * + * \sa class CwiseNullaryOp + */ +template<typename Derived> +template<typename CustomNullaryOp> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, typename DenseBase<Derived>::PlainObject> +DenseBase<Derived>::NullaryExpr(Index rows, Index cols, const CustomNullaryOp& func) +{ + return CwiseNullaryOp<CustomNullaryOp, PlainObject>(rows, cols, func); +} + +/** \returns an expression of a matrix defined by a custom functor \a func + * + * The parameter \a size is the size of the returned vector. + * Must be compatible with this MatrixBase type. + * + * \only_for_vectors + * + * This variant is meant to be used for dynamic-size vector types. For fixed-size types, + * it is redundant to pass \a size as argument, so Zero() should be used + * instead. + * + * The template parameter \a CustomNullaryOp is the type of the functor. + * + * Here is an example with C++11 random generators: \include random_cpp11.cpp + * Output: \verbinclude random_cpp11.out + * + * \sa class CwiseNullaryOp + */ +template<typename Derived> +template<typename CustomNullaryOp> +EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, typename DenseBase<Derived>::PlainObject> +DenseBase<Derived>::NullaryExpr(Index size, const CustomNullaryOp& func) +{ + EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) + if(RowsAtCompileTime == 1) return CwiseNullaryOp<CustomNullaryOp, PlainObject>(1, size, func); + else return CwiseNullaryOp<CustomNullaryOp, PlainObject>(size, 1, func); +} + +/** \returns an expression of a matrix defined by a custom functor \a func + * + * This variant is only for fixed-size DenseBase types. For dynamic-size types, you + * need to use the variants taking size arguments. + * + * The template parameter \a CustomNullaryOp is the type of the functor. + * + * \sa class CwiseNullaryOp + */ +template<typename Derived> +template<typename CustomNullaryOp> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, typename DenseBase<Derived>::PlainObject> +DenseBase<Derived>::NullaryExpr(const CustomNullaryOp& func) +{ + return CwiseNullaryOp<CustomNullaryOp, PlainObject>(RowsAtCompileTime, ColsAtCompileTime, func); +} + +/** \returns an expression of a constant matrix of value \a value + * + * The parameters \a rows and \a cols are the number of rows and of columns of + * the returned matrix. Must be compatible with this DenseBase type. + * + * This variant is meant to be used for dynamic-size matrix types. For fixed-size types, + * it is redundant to pass \a rows and \a cols as arguments, so Zero() should be used + * instead. + * + * The template parameter \a CustomNullaryOp is the type of the functor. + * + * \sa class CwiseNullaryOp + */ +template<typename Derived> +EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType +DenseBase<Derived>::Constant(Index rows, Index cols, const Scalar& value) +{ + return DenseBase<Derived>::NullaryExpr(rows, cols, internal::scalar_constant_op<Scalar>(value)); +} + +/** \returns an expression of a constant matrix of value \a value + * + * The parameter \a size is the size of the returned vector. + * Must be compatible with this DenseBase type. + * + * \only_for_vectors + * + * This variant is meant to be used for dynamic-size vector types. For fixed-size types, + * it is redundant to pass \a size as argument, so Zero() should be used + * instead. + * + * The template parameter \a CustomNullaryOp is the type of the functor. + * + * \sa class CwiseNullaryOp + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType +DenseBase<Derived>::Constant(Index size, const Scalar& value) +{ + return DenseBase<Derived>::NullaryExpr(size, internal::scalar_constant_op<Scalar>(value)); +} + +/** \returns an expression of a constant matrix of value \a value + * + * This variant is only for fixed-size DenseBase types. For dynamic-size types, you + * need to use the variants taking size arguments. + * + * The template parameter \a CustomNullaryOp is the type of the functor. + * + * \sa class CwiseNullaryOp + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType +DenseBase<Derived>::Constant(const Scalar& value) +{ + EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived) + return DenseBase<Derived>::NullaryExpr(RowsAtCompileTime, ColsAtCompileTime, internal::scalar_constant_op<Scalar>(value)); +} + +/** \deprecated because of accuracy loss. In Eigen 3.3, it is an alias for LinSpaced(Index,const Scalar&,const Scalar&) + * + * \sa LinSpaced(Index,Scalar,Scalar), setLinSpaced(Index,const Scalar&,const Scalar&) + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedReturnType +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)); +} + +/** \deprecated because of accuracy loss. In Eigen 3.3, it is an alias for LinSpaced(const Scalar&,const Scalar&) + * + * \sa LinSpaced(Scalar,Scalar) + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedReturnType +DenseBase<Derived>::LinSpaced(Sequential_t, 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)); +} + +/** + * \brief Sets a linearly spaced vector. + * + * The function generates 'size' equally spaced values in the closed interval [low,high]. + * When size is set to 1, a vector of length 1 containing 'high' is returned. + * + * \only_for_vectors + * + * Example: \include DenseBase_LinSpaced.cpp + * Output: \verbinclude DenseBase_LinSpaced.out + * + * For integer scalar types, an even spacing is possible if and only if the length of the range, + * i.e., \c high-low is a scalar multiple of \c size-1, or if \c size is a scalar multiple of the + * number of values \c high-low+1 (meaning each value can be repeated the same number of time). + * If one of these two considions is not satisfied, then \c high is lowered to the largest value + * satisfying one of this constraint. + * Here are some examples: + * + * Example: \include DenseBase_LinSpacedInt.cpp + * Output: \verbinclude DenseBase_LinSpacedInt.out + * + * \sa setLinSpaced(Index,const Scalar&,const Scalar&), CwiseNullaryOp + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedReturnType +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)); +} + +/** + * \copydoc DenseBase::LinSpaced(Index, const Scalar&, const Scalar&) + * Special version for fixed size types which does not require the size parameter. + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedReturnType +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)); +} + +/** \returns true if all coefficients in this matrix are approximately equal to \a val, to within precision \a prec */ +template<typename Derived> +EIGEN_DEVICE_FUNC bool DenseBase<Derived>::isApproxToConstant +(const Scalar& val, const RealScalar& prec) const +{ + typename internal::nested_eval<Derived,1>::type self(derived()); + for(Index j = 0; j < cols(); ++j) + for(Index i = 0; i < rows(); ++i) + if(!internal::isApprox(self.coeff(i, j), val, prec)) + return false; + return true; +} + +/** This is just an alias for isApproxToConstant(). + * + * \returns true if all coefficients in this matrix are approximately equal to \a value, to within precision \a prec */ +template<typename Derived> +EIGEN_DEVICE_FUNC bool DenseBase<Derived>::isConstant +(const Scalar& val, const RealScalar& prec) const +{ + return isApproxToConstant(val, prec); +} + +/** Alias for setConstant(): sets all coefficients in this expression to \a val. + * + * \sa setConstant(), Constant(), class CwiseNullaryOp + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void DenseBase<Derived>::fill(const Scalar& val) +{ + setConstant(val); +} + +/** Sets all coefficients in this expression to value \a val. + * + * \sa fill(), setConstant(Index,const Scalar&), setConstant(Index,Index,const Scalar&), setZero(), setOnes(), Constant(), class CwiseNullaryOp, setZero(), setOnes() + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setConstant(const Scalar& val) +{ + return derived() = Constant(rows(), cols(), val); +} + +/** Resizes to the given \a size, and sets all coefficients in this expression to the given value \a val. + * + * \only_for_vectors + * + * Example: \include Matrix_setConstant_int.cpp + * Output: \verbinclude Matrix_setConstant_int.out + * + * \sa MatrixBase::setConstant(const Scalar&), setConstant(Index,Index,const Scalar&), class CwiseNullaryOp, MatrixBase::Constant(const Scalar&) + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& +PlainObjectBase<Derived>::setConstant(Index size, const Scalar& val) +{ + resize(size); + return setConstant(val); +} + +/** Resizes to the given size, and sets all coefficients in this expression to the given value \a val. + * + * \param rows the new number of rows + * \param cols the new number of columns + * \param val the value to which all coefficients are set + * + * Example: \include Matrix_setConstant_int_int.cpp + * Output: \verbinclude Matrix_setConstant_int_int.out + * + * \sa MatrixBase::setConstant(const Scalar&), setConstant(Index,const Scalar&), class CwiseNullaryOp, MatrixBase::Constant(const Scalar&) + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& +PlainObjectBase<Derived>::setConstant(Index rows, Index cols, const Scalar& val) +{ + resize(rows, cols); + return setConstant(val); +} + +/** + * \brief Sets a linearly spaced vector. + * + * The function generates 'size' equally spaced values in the closed interval [low,high]. + * When size is set to 1, a vector of length 1 containing 'high' is returned. + * + * \only_for_vectors + * + * Example: \include DenseBase_setLinSpaced.cpp + * Output: \verbinclude DenseBase_setLinSpaced.out + * + * For integer scalar types, do not miss the explanations on the definition + * of \link LinSpaced(Index,const Scalar&,const Scalar&) even spacing \endlink. + * + * \sa LinSpaced(Index,const Scalar&,const Scalar&), CwiseNullaryOp + */ +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)); +} + +/** + * \brief Sets a linearly spaced vector. + * + * The function fills \c *this with equally spaced values in the closed interval [low,high]. + * When size is set to 1, a vector of length 1 containing 'high' is returned. + * + * \only_for_vectors + * + * For integer scalar types, do not miss the explanations on the definition + * of \link LinSpaced(Index,const Scalar&,const Scalar&) even spacing \endlink. + * + * \sa LinSpaced(Index,const Scalar&,const Scalar&), setLinSpaced(Index, const Scalar&, const Scalar&), CwiseNullaryOp + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setLinSpaced(const Scalar& low, const Scalar& high) +{ + EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) + return setLinSpaced(size(), low, high); +} + +// zero: + +/** \returns an expression of a zero matrix. + * + * The parameters \a rows and \a cols are the number of rows and of columns of + * the returned matrix. Must be compatible with this MatrixBase type. + * + * This variant is meant to be used for dynamic-size matrix types. For fixed-size types, + * it is redundant to pass \a rows and \a cols as arguments, so Zero() should be used + * instead. + * + * Example: \include MatrixBase_zero_int_int.cpp + * Output: \verbinclude MatrixBase_zero_int_int.out + * + * \sa Zero(), Zero(Index) + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType +DenseBase<Derived>::Zero(Index rows, Index cols) +{ + return Constant(rows, cols, Scalar(0)); +} + +/** \returns an expression of a zero vector. + * + * The parameter \a size is the size of the returned vector. + * Must be compatible with this MatrixBase type. + * + * \only_for_vectors + * + * This variant is meant to be used for dynamic-size vector types. For fixed-size types, + * it is redundant to pass \a size as argument, so Zero() should be used + * instead. + * + * Example: \include MatrixBase_zero_int.cpp + * Output: \verbinclude MatrixBase_zero_int.out + * + * \sa Zero(), Zero(Index,Index) + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType +DenseBase<Derived>::Zero(Index size) +{ + return Constant(size, Scalar(0)); +} + +/** \returns an expression of a fixed-size zero matrix or vector. + * + * This variant is only for fixed-size MatrixBase types. For dynamic-size types, you + * need to use the variants taking size arguments. + * + * Example: \include MatrixBase_zero.cpp + * Output: \verbinclude MatrixBase_zero.out + * + * \sa Zero(Index), Zero(Index,Index) + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType +DenseBase<Derived>::Zero() +{ + return Constant(Scalar(0)); +} + +/** \returns true if *this is approximately equal to the zero matrix, + * within the precision given by \a prec. + * + * Example: \include MatrixBase_isZero.cpp + * Output: \verbinclude MatrixBase_isZero.out + * + * \sa class CwiseNullaryOp, Zero() + */ +template<typename Derived> +EIGEN_DEVICE_FUNC bool DenseBase<Derived>::isZero(const RealScalar& prec) const +{ + typename internal::nested_eval<Derived,1>::type self(derived()); + for(Index j = 0; j < cols(); ++j) + for(Index i = 0; i < rows(); ++i) + if(!internal::isMuchSmallerThan(self.coeff(i, j), static_cast<Scalar>(1), prec)) + return false; + return true; +} + +/** Sets all coefficients in this expression to zero. + * + * Example: \include MatrixBase_setZero.cpp + * Output: \verbinclude MatrixBase_setZero.out + * + * \sa class CwiseNullaryOp, Zero() + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setZero() +{ + return setConstant(Scalar(0)); +} + +/** Resizes to the given \a size, and sets all coefficients in this expression to zero. + * + * \only_for_vectors + * + * Example: \include Matrix_setZero_int.cpp + * Output: \verbinclude Matrix_setZero_int.out + * + * \sa DenseBase::setZero(), setZero(Index,Index), class CwiseNullaryOp, DenseBase::Zero() + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& +PlainObjectBase<Derived>::setZero(Index newSize) +{ + resize(newSize); + return setConstant(Scalar(0)); +} + +/** Resizes to the given size, and sets all coefficients in this expression to zero. + * + * \param rows the new number of rows + * \param cols the new number of columns + * + * Example: \include Matrix_setZero_int_int.cpp + * Output: \verbinclude Matrix_setZero_int_int.out + * + * \sa DenseBase::setZero(), setZero(Index), class CwiseNullaryOp, DenseBase::Zero() + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& +PlainObjectBase<Derived>::setZero(Index rows, Index cols) +{ + resize(rows, cols); + return setConstant(Scalar(0)); +} + +// ones: + +/** \returns an expression of a matrix where all coefficients equal one. + * + * The parameters \a rows and \a cols are the number of rows and of columns of + * the returned matrix. Must be compatible with this MatrixBase type. + * + * This variant is meant to be used for dynamic-size matrix types. For fixed-size types, + * it is redundant to pass \a rows and \a cols as arguments, so Ones() should be used + * instead. + * + * Example: \include MatrixBase_ones_int_int.cpp + * Output: \verbinclude MatrixBase_ones_int_int.out + * + * \sa Ones(), Ones(Index), isOnes(), class Ones + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType +DenseBase<Derived>::Ones(Index rows, Index cols) +{ + return Constant(rows, cols, Scalar(1)); +} + +/** \returns an expression of a vector where all coefficients equal one. + * + * The parameter \a newSize is the size of the returned vector. + * Must be compatible with this MatrixBase type. + * + * \only_for_vectors + * + * This variant is meant to be used for dynamic-size vector types. For fixed-size types, + * it is redundant to pass \a size as argument, so Ones() should be used + * instead. + * + * Example: \include MatrixBase_ones_int.cpp + * Output: \verbinclude MatrixBase_ones_int.out + * + * \sa Ones(), Ones(Index,Index), isOnes(), class Ones + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType +DenseBase<Derived>::Ones(Index newSize) +{ + return Constant(newSize, Scalar(1)); +} + +/** \returns an expression of a fixed-size matrix or vector where all coefficients equal one. + * + * This variant is only for fixed-size MatrixBase types. For dynamic-size types, you + * need to use the variants taking size arguments. + * + * Example: \include MatrixBase_ones.cpp + * Output: \verbinclude MatrixBase_ones.out + * + * \sa Ones(Index), Ones(Index,Index), isOnes(), class Ones + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType +DenseBase<Derived>::Ones() +{ + return Constant(Scalar(1)); +} + +/** \returns true if *this is approximately equal to the matrix where all coefficients + * are equal to 1, within the precision given by \a prec. + * + * Example: \include MatrixBase_isOnes.cpp + * Output: \verbinclude MatrixBase_isOnes.out + * + * \sa class CwiseNullaryOp, Ones() + */ +template<typename Derived> +EIGEN_DEVICE_FUNC bool DenseBase<Derived>::isOnes +(const RealScalar& prec) const +{ + return isApproxToConstant(Scalar(1), prec); +} + +/** Sets all coefficients in this expression to one. + * + * Example: \include MatrixBase_setOnes.cpp + * Output: \verbinclude MatrixBase_setOnes.out + * + * \sa class CwiseNullaryOp, Ones() + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setOnes() +{ + return setConstant(Scalar(1)); +} + +/** Resizes to the given \a newSize, and sets all coefficients in this expression to one. + * + * \only_for_vectors + * + * Example: \include Matrix_setOnes_int.cpp + * Output: \verbinclude Matrix_setOnes_int.out + * + * \sa MatrixBase::setOnes(), setOnes(Index,Index), class CwiseNullaryOp, MatrixBase::Ones() + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& +PlainObjectBase<Derived>::setOnes(Index newSize) +{ + resize(newSize); + return setConstant(Scalar(1)); +} + +/** Resizes to the given size, and sets all coefficients in this expression to one. + * + * \param rows the new number of rows + * \param cols the new number of columns + * + * Example: \include Matrix_setOnes_int_int.cpp + * Output: \verbinclude Matrix_setOnes_int_int.out + * + * \sa MatrixBase::setOnes(), setOnes(Index), class CwiseNullaryOp, MatrixBase::Ones() + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& +PlainObjectBase<Derived>::setOnes(Index rows, Index cols) +{ + resize(rows, cols); + return setConstant(Scalar(1)); +} + +// Identity: + +/** \returns an expression of the identity matrix (not necessarily square). + * + * The parameters \a rows and \a cols are the number of rows and of columns of + * the returned matrix. Must be compatible with this MatrixBase type. + * + * This variant is meant to be used for dynamic-size matrix types. For fixed-size types, + * it is redundant to pass \a rows and \a cols as arguments, so Identity() should be used + * instead. + * + * Example: \include MatrixBase_identity_int_int.cpp + * Output: \verbinclude MatrixBase_identity_int_int.out + * + * \sa Identity(), setIdentity(), isIdentity() + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::IdentityReturnType +MatrixBase<Derived>::Identity(Index rows, Index cols) +{ + return DenseBase<Derived>::NullaryExpr(rows, cols, internal::scalar_identity_op<Scalar>()); +} + +/** \returns an expression of the identity matrix (not necessarily square). + * + * This variant is only for fixed-size MatrixBase types. For dynamic-size types, you + * need to use the variant taking size arguments. + * + * Example: \include MatrixBase_identity.cpp + * Output: \verbinclude MatrixBase_identity.out + * + * \sa Identity(Index,Index), setIdentity(), isIdentity() + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::IdentityReturnType +MatrixBase<Derived>::Identity() +{ + EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived) + return MatrixBase<Derived>::NullaryExpr(RowsAtCompileTime, ColsAtCompileTime, internal::scalar_identity_op<Scalar>()); +} + +/** \returns true if *this is approximately equal to the identity matrix + * (not necessarily square), + * within the precision given by \a prec. + * + * Example: \include MatrixBase_isIdentity.cpp + * Output: \verbinclude MatrixBase_isIdentity.out + * + * \sa class CwiseNullaryOp, Identity(), Identity(Index,Index), setIdentity() + */ +template<typename Derived> +bool MatrixBase<Derived>::isIdentity +(const RealScalar& prec) const +{ + typename internal::nested_eval<Derived,1>::type self(derived()); + for(Index j = 0; j < cols(); ++j) + { + for(Index i = 0; i < rows(); ++i) + { + if(i == j) + { + if(!internal::isApprox(self.coeff(i, j), static_cast<Scalar>(1), prec)) + return false; + } + else + { + if(!internal::isMuchSmallerThan(self.coeff(i, j), static_cast<RealScalar>(1), prec)) + return false; + } + } + } + return true; +} + +namespace internal { + +template<typename Derived, bool Big = (Derived::SizeAtCompileTime>=16)> +struct setIdentity_impl +{ + EIGEN_DEVICE_FUNC + static EIGEN_STRONG_INLINE Derived& run(Derived& m) + { + return m = Derived::Identity(m.rows(), m.cols()); + } +}; + +template<typename Derived> +struct setIdentity_impl<Derived, true> +{ + EIGEN_DEVICE_FUNC + static EIGEN_STRONG_INLINE Derived& run(Derived& m) + { + m.setZero(); + const Index size = numext::mini(m.rows(), m.cols()); + for(Index i = 0; i < size; ++i) m.coeffRef(i,i) = typename Derived::Scalar(1); + return m; + } +}; + +} // end namespace internal + +/** Writes the identity expression (not necessarily square) into *this. + * + * Example: \include MatrixBase_setIdentity.cpp + * Output: \verbinclude MatrixBase_setIdentity.out + * + * \sa class CwiseNullaryOp, Identity(), Identity(Index,Index), isIdentity() + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setIdentity() +{ + return internal::setIdentity_impl<Derived>::run(derived()); +} + +/** \brief Resizes to the given size, and writes the identity expression (not necessarily square) into *this. + * + * \param rows the new number of rows + * \param cols the new number of columns + * + * Example: \include Matrix_setIdentity_int_int.cpp + * Output: \verbinclude Matrix_setIdentity_int_int.out + * + * \sa MatrixBase::setIdentity(), class CwiseNullaryOp, MatrixBase::Identity() + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setIdentity(Index rows, Index cols) +{ + derived().resize(rows, cols); + return setIdentity(); +} + +/** \returns an expression of the i-th unit (basis) vector. + * + * \only_for_vectors + * + * \sa MatrixBase::Unit(Index), MatrixBase::UnitX(), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW() + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::Unit(Index newSize, Index i) +{ + EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) + return BasisReturnType(SquareMatrixType::Identity(newSize,newSize), i); +} + +/** \returns an expression of the i-th unit (basis) vector. + * + * \only_for_vectors + * + * This variant is for fixed-size vector only. + * + * \sa MatrixBase::Unit(Index,Index), MatrixBase::UnitX(), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW() + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::Unit(Index i) +{ + EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) + return BasisReturnType(SquareMatrixType::Identity(),i); +} + +/** \returns an expression of the X axis unit vector (1{,0}^*) + * + * \only_for_vectors + * + * \sa MatrixBase::Unit(Index,Index), MatrixBase::Unit(Index), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW() + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::UnitX() +{ return Derived::Unit(0); } + +/** \returns an expression of the Y axis unit vector (0,1{,0}^*) + * + * \only_for_vectors + * + * \sa MatrixBase::Unit(Index,Index), MatrixBase::Unit(Index), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW() + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::UnitY() +{ return Derived::Unit(1); } + +/** \returns an expression of the Z axis unit vector (0,0,1{,0}^*) + * + * \only_for_vectors + * + * \sa MatrixBase::Unit(Index,Index), MatrixBase::Unit(Index), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW() + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::UnitZ() +{ return Derived::Unit(2); } + +/** \returns an expression of the W axis unit vector (0,0,0,1) + * + * \only_for_vectors + * + * \sa MatrixBase::Unit(Index,Index), MatrixBase::Unit(Index), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW() + */ +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::UnitW() +{ return Derived::Unit(3); } + +} // end namespace Eigen + +#endif // EIGEN_CWISE_NULLARY_OP_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/CwiseTernaryOp.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/CwiseTernaryOp.h new file mode 100644 index 000000000..9f3576fec --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/CwiseTernaryOp.h @@ -0,0 +1,197 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2014 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com> +// Copyright (C) 2016 Eugene Brevdo <ebrevdo@gmail.com> +// +// 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_CWISE_TERNARY_OP_H +#define EIGEN_CWISE_TERNARY_OP_H + +namespace Eigen { + +namespace internal { +template <typename TernaryOp, typename Arg1, typename Arg2, typename Arg3> +struct traits<CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> > { + // we must not inherit from traits<Arg1> since it has + // the potential to cause problems with MSVC + typedef typename remove_all<Arg1>::type Ancestor; + typedef typename traits<Ancestor>::XprKind XprKind; + enum { + RowsAtCompileTime = traits<Ancestor>::RowsAtCompileTime, + ColsAtCompileTime = traits<Ancestor>::ColsAtCompileTime, + MaxRowsAtCompileTime = traits<Ancestor>::MaxRowsAtCompileTime, + MaxColsAtCompileTime = traits<Ancestor>::MaxColsAtCompileTime + }; + + // even though we require Arg1, Arg2, and Arg3 to have the same scalar type + // (see CwiseTernaryOp constructor), + // we still want to handle the case when the result type is different. + typedef typename result_of<TernaryOp( + const typename Arg1::Scalar&, const typename Arg2::Scalar&, + const typename Arg3::Scalar&)>::type Scalar; + + typedef typename internal::traits<Arg1>::StorageKind StorageKind; + typedef typename internal::traits<Arg1>::StorageIndex StorageIndex; + + typedef typename Arg1::Nested Arg1Nested; + typedef typename Arg2::Nested Arg2Nested; + typedef typename Arg3::Nested Arg3Nested; + typedef typename remove_reference<Arg1Nested>::type _Arg1Nested; + typedef typename remove_reference<Arg2Nested>::type _Arg2Nested; + typedef typename remove_reference<Arg3Nested>::type _Arg3Nested; + enum { Flags = _Arg1Nested::Flags & RowMajorBit }; +}; +} // end namespace internal + +template <typename TernaryOp, typename Arg1, typename Arg2, typename Arg3, + typename StorageKind> +class CwiseTernaryOpImpl; + +/** \class CwiseTernaryOp + * \ingroup Core_Module + * + * \brief Generic expression where a coefficient-wise ternary operator is + * applied to two expressions + * + * \tparam TernaryOp template functor implementing the operator + * \tparam Arg1Type the type of the first argument + * \tparam Arg2Type the type of the second argument + * \tparam Arg3Type the type of the third argument + * + * This class represents an expression where a coefficient-wise ternary + * operator is applied to three expressions. + * It is the return type of ternary operators, by which we mean only those + * ternary operators where + * all three arguments are Eigen expressions. + * For example, the return type of betainc(matrix1, matrix2, matrix3) is a + * CwiseTernaryOp. + * + * Most of the time, this is the only way that it is used, so you typically + * don't have to name + * CwiseTernaryOp types explicitly. + * + * \sa MatrixBase::ternaryExpr(const MatrixBase<Argument2> &, const + * MatrixBase<Argument3> &, const CustomTernaryOp &) const, class CwiseBinaryOp, + * class CwiseUnaryOp, class CwiseNullaryOp + */ +template <typename TernaryOp, typename Arg1Type, typename Arg2Type, + typename Arg3Type> +class CwiseTernaryOp : public CwiseTernaryOpImpl< + TernaryOp, Arg1Type, Arg2Type, Arg3Type, + typename internal::traits<Arg1Type>::StorageKind>, + internal::no_assignment_operator +{ + public: + typedef typename internal::remove_all<Arg1Type>::type Arg1; + typedef typename internal::remove_all<Arg2Type>::type Arg2; + typedef typename internal::remove_all<Arg3Type>::type Arg3; + + typedef typename CwiseTernaryOpImpl< + TernaryOp, Arg1Type, Arg2Type, Arg3Type, + typename internal::traits<Arg1Type>::StorageKind>::Base Base; + EIGEN_GENERIC_PUBLIC_INTERFACE(CwiseTernaryOp) + + typedef typename internal::ref_selector<Arg1Type>::type Arg1Nested; + typedef typename internal::ref_selector<Arg2Type>::type Arg2Nested; + typedef typename internal::ref_selector<Arg3Type>::type Arg3Nested; + typedef typename internal::remove_reference<Arg1Nested>::type _Arg1Nested; + typedef typename internal::remove_reference<Arg2Nested>::type _Arg2Nested; + typedef typename internal::remove_reference<Arg3Nested>::type _Arg3Nested; + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE CwiseTernaryOp(const Arg1& a1, const Arg2& a2, + const Arg3& a3, + const TernaryOp& func = TernaryOp()) + : m_arg1(a1), m_arg2(a2), m_arg3(a3), m_functor(func) { + // require the sizes to match + EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Arg1, Arg2) + EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Arg1, Arg3) + + // The index types should match + EIGEN_STATIC_ASSERT((internal::is_same< + typename internal::traits<Arg1Type>::StorageKind, + typename internal::traits<Arg2Type>::StorageKind>::value), + STORAGE_KIND_MUST_MATCH) + EIGEN_STATIC_ASSERT((internal::is_same< + typename internal::traits<Arg1Type>::StorageKind, + typename internal::traits<Arg3Type>::StorageKind>::value), + STORAGE_KIND_MUST_MATCH) + + eigen_assert(a1.rows() == a2.rows() && a1.cols() == a2.cols() && + a1.rows() == a3.rows() && a1.cols() == a3.cols()); + } + + 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<Arg1Nested>::type>:: + RowsAtCompileTime == Dynamic && + internal::traits<typename internal::remove_all<Arg2Nested>::type>:: + RowsAtCompileTime == Dynamic) + return m_arg3.rows(); + else if (internal::traits<typename internal::remove_all<Arg1Nested>::type>:: + RowsAtCompileTime == Dynamic && + internal::traits<typename internal::remove_all<Arg3Nested>::type>:: + RowsAtCompileTime == Dynamic) + return m_arg2.rows(); + else + return m_arg1.rows(); + } + 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<Arg1Nested>::type>:: + ColsAtCompileTime == Dynamic && + internal::traits<typename internal::remove_all<Arg2Nested>::type>:: + ColsAtCompileTime == Dynamic) + return m_arg3.cols(); + else if (internal::traits<typename internal::remove_all<Arg1Nested>::type>:: + ColsAtCompileTime == Dynamic && + internal::traits<typename internal::remove_all<Arg3Nested>::type>:: + ColsAtCompileTime == Dynamic) + return m_arg2.cols(); + else + return m_arg1.cols(); + } + + /** \returns the first argument nested expression */ + EIGEN_DEVICE_FUNC + const _Arg1Nested& arg1() const { return m_arg1; } + /** \returns the first argument nested expression */ + EIGEN_DEVICE_FUNC + const _Arg2Nested& arg2() const { return m_arg2; } + /** \returns the third argument nested expression */ + EIGEN_DEVICE_FUNC + const _Arg3Nested& arg3() const { return m_arg3; } + /** \returns the functor representing the ternary operation */ + EIGEN_DEVICE_FUNC + const TernaryOp& functor() const { return m_functor; } + + protected: + Arg1Nested m_arg1; + Arg2Nested m_arg2; + Arg3Nested m_arg3; + const TernaryOp m_functor; +}; + +// Generic API dispatcher +template <typename TernaryOp, typename Arg1, typename Arg2, typename Arg3, + typename StorageKind> +class CwiseTernaryOpImpl + : public internal::generic_xpr_base< + CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> >::type { + public: + typedef typename internal::generic_xpr_base< + CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> >::type Base; +}; + +} // end namespace Eigen + +#endif // EIGEN_CWISE_TERNARY_OP_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/CwiseUnaryOp.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/CwiseUnaryOp.h new file mode 100644 index 000000000..1d2dd19f2 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/CwiseUnaryOp.h @@ -0,0 +1,103 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2014 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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_CWISE_UNARY_OP_H +#define EIGEN_CWISE_UNARY_OP_H + +namespace Eigen { + +namespace internal { +template<typename UnaryOp, typename XprType> +struct traits<CwiseUnaryOp<UnaryOp, XprType> > + : traits<XprType> +{ + typedef typename result_of< + UnaryOp(const typename XprType::Scalar&) + >::type Scalar; + typedef typename XprType::Nested XprTypeNested; + typedef typename remove_reference<XprTypeNested>::type _XprTypeNested; + enum { + Flags = _XprTypeNested::Flags & RowMajorBit + }; +}; +} + +template<typename UnaryOp, typename XprType, typename StorageKind> +class CwiseUnaryOpImpl; + +/** \class CwiseUnaryOp + * \ingroup Core_Module + * + * \brief Generic expression where a coefficient-wise unary operator is applied to an expression + * + * \tparam UnaryOp template functor implementing the operator + * \tparam XprType the type of the expression to which we are applying the unary operator + * + * This class represents an expression where a unary operator is applied to an expression. + * It is the return type of all operations taking exactly 1 input expression, regardless of the + * presence of other inputs such as scalars. For example, the operator* in the expression 3*matrix + * is considered unary, because only the right-hand side is an expression, and its + * return type is a specialization of CwiseUnaryOp. + * + * Most of the time, this is the only way that it is used, so you typically don't have to name + * CwiseUnaryOp types explicitly. + * + * \sa MatrixBase::unaryExpr(const CustomUnaryOp &) const, class CwiseBinaryOp, class CwiseNullaryOp + */ +template<typename UnaryOp, typename XprType> +class CwiseUnaryOp : public CwiseUnaryOpImpl<UnaryOp, XprType, typename internal::traits<XprType>::StorageKind>, internal::no_assignment_operator +{ + public: + + typedef typename CwiseUnaryOpImpl<UnaryOp, XprType,typename internal::traits<XprType>::StorageKind>::Base Base; + EIGEN_GENERIC_PUBLIC_INTERFACE(CwiseUnaryOp) + typedef typename internal::ref_selector<XprType>::type XprTypeNested; + typedef typename internal::remove_all<XprType>::type NestedExpression; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit CwiseUnaryOp(const XprType& xpr, const UnaryOp& func = UnaryOp()) + : m_xpr(xpr), m_functor(func) {} + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Index rows() const { return m_xpr.rows(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Index cols() const { return m_xpr.cols(); } + + /** \returns the functor representing the unary operation */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const UnaryOp& functor() const { return m_functor; } + + /** \returns the nested expression */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const typename internal::remove_all<XprTypeNested>::type& + nestedExpression() const { return m_xpr; } + + /** \returns the nested expression */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + typename internal::remove_all<XprTypeNested>::type& + nestedExpression() { return m_xpr; } + + protected: + XprTypeNested m_xpr; + const UnaryOp m_functor; +}; + +// Generic API dispatcher +template<typename UnaryOp, typename XprType, typename StorageKind> +class CwiseUnaryOpImpl + : public internal::generic_xpr_base<CwiseUnaryOp<UnaryOp, XprType> >::type +{ +public: + typedef typename internal::generic_xpr_base<CwiseUnaryOp<UnaryOp, XprType> >::type Base; +}; + +} // end namespace Eigen + +#endif // EIGEN_CWISE_UNARY_OP_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/CwiseUnaryView.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/CwiseUnaryView.h new file mode 100644 index 000000000..271033056 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/CwiseUnaryView.h @@ -0,0 +1,128 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2009-2010 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_CWISE_UNARY_VIEW_H +#define EIGEN_CWISE_UNARY_VIEW_H + +namespace Eigen { + +namespace internal { +template<typename ViewOp, typename MatrixType> +struct traits<CwiseUnaryView<ViewOp, MatrixType> > + : traits<MatrixType> +{ + typedef typename result_of< + ViewOp(const typename traits<MatrixType>::Scalar&) + >::type Scalar; + typedef typename MatrixType::Nested MatrixTypeNested; + typedef typename remove_all<MatrixTypeNested>::type _MatrixTypeNested; + enum { + FlagsLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0, + Flags = traits<_MatrixTypeNested>::Flags & (RowMajorBit | FlagsLvalueBit | DirectAccessBit), // FIXME DirectAccessBit should not be handled by expressions + MatrixTypeInnerStride = inner_stride_at_compile_time<MatrixType>::ret, + // need to cast the sizeof's from size_t to int explicitly, otherwise: + // "error: no integral type can represent all of the enumerator values + InnerStrideAtCompileTime = MatrixTypeInnerStride == Dynamic + ? int(Dynamic) + : int(MatrixTypeInnerStride) * int(sizeof(typename traits<MatrixType>::Scalar) / sizeof(Scalar)), + OuterStrideAtCompileTime = outer_stride_at_compile_time<MatrixType>::ret == Dynamic + ? int(Dynamic) + : outer_stride_at_compile_time<MatrixType>::ret * int(sizeof(typename traits<MatrixType>::Scalar) / sizeof(Scalar)) + }; +}; +} + +template<typename ViewOp, typename MatrixType, typename StorageKind> +class CwiseUnaryViewImpl; + +/** \class CwiseUnaryView + * \ingroup Core_Module + * + * \brief Generic lvalue expression of a coefficient-wise unary operator of a matrix or a vector + * + * \tparam ViewOp template functor implementing the view + * \tparam MatrixType the type of the matrix we are applying the unary operator + * + * This class represents a lvalue expression of a generic unary view operator of a matrix or a vector. + * It is the return type of real() and imag(), and most of the time this is the only way it is used. + * + * \sa MatrixBase::unaryViewExpr(const CustomUnaryOp &) const, class CwiseUnaryOp + */ +template<typename ViewOp, typename MatrixType> +class CwiseUnaryView : public CwiseUnaryViewImpl<ViewOp, MatrixType, typename internal::traits<MatrixType>::StorageKind> +{ + public: + + typedef typename CwiseUnaryViewImpl<ViewOp, MatrixType,typename internal::traits<MatrixType>::StorageKind>::Base Base; + EIGEN_GENERIC_PUBLIC_INTERFACE(CwiseUnaryView) + typedef typename internal::ref_selector<MatrixType>::non_const_type MatrixTypeNested; + typedef typename internal::remove_all<MatrixType>::type NestedExpression; + + explicit inline CwiseUnaryView(MatrixType& mat, const ViewOp& func = ViewOp()) + : m_matrix(mat), m_functor(func) {} + + EIGEN_INHERIT_ASSIGNMENT_OPERATORS(CwiseUnaryView) + + EIGEN_STRONG_INLINE Index rows() const { return m_matrix.rows(); } + EIGEN_STRONG_INLINE Index cols() const { return m_matrix.cols(); } + + /** \returns the functor representing unary operation */ + const ViewOp& functor() const { return m_functor; } + + /** \returns the nested expression */ + const typename internal::remove_all<MatrixTypeNested>::type& + nestedExpression() const { return m_matrix; } + + /** \returns the nested expression */ + typename internal::remove_reference<MatrixTypeNested>::type& + nestedExpression() { return m_matrix.const_cast_derived(); } + + protected: + MatrixTypeNested m_matrix; + ViewOp m_functor; +}; + +// Generic API dispatcher +template<typename ViewOp, typename XprType, typename StorageKind> +class CwiseUnaryViewImpl + : public internal::generic_xpr_base<CwiseUnaryView<ViewOp, XprType> >::type +{ +public: + typedef typename internal::generic_xpr_base<CwiseUnaryView<ViewOp, XprType> >::type Base; +}; + +template<typename ViewOp, typename MatrixType> +class CwiseUnaryViewImpl<ViewOp,MatrixType,Dense> + : public internal::dense_xpr_base< CwiseUnaryView<ViewOp, MatrixType> >::type +{ + public: + + typedef CwiseUnaryView<ViewOp, MatrixType> Derived; + typedef typename internal::dense_xpr_base< CwiseUnaryView<ViewOp, MatrixType> >::type Base; + + EIGEN_DENSE_PUBLIC_INTERFACE(Derived) + EIGEN_INHERIT_ASSIGNMENT_OPERATORS(CwiseUnaryViewImpl) + + EIGEN_DEVICE_FUNC inline Scalar* data() { return &(this->coeffRef(0)); } + EIGEN_DEVICE_FUNC inline const Scalar* data() const { return &(this->coeff(0)); } + + EIGEN_DEVICE_FUNC inline Index innerStride() const + { + return derived().nestedExpression().innerStride() * sizeof(typename internal::traits<MatrixType>::Scalar) / sizeof(Scalar); + } + + EIGEN_DEVICE_FUNC inline Index outerStride() const + { + return derived().nestedExpression().outerStride() * sizeof(typename internal::traits<MatrixType>::Scalar) / sizeof(Scalar); + } +}; + +} // end namespace Eigen + +#endif // EIGEN_CWISE_UNARY_VIEW_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/DenseBase.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/DenseBase.h new file mode 100644 index 000000000..90066ae73 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/DenseBase.h @@ -0,0 +1,611 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2007-2010 Benoit Jacob <jacob.benoit.1@gmail.com> +// Copyright (C) 2008-2010 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_DENSEBASE_H +#define EIGEN_DENSEBASE_H + +namespace Eigen { + +namespace internal { + +// The index type defined by EIGEN_DEFAULT_DENSE_INDEX_TYPE must be a signed type. +// This dummy function simply aims at checking that at compile time. +static inline void check_DenseIndex_is_signed() { + EIGEN_STATIC_ASSERT(NumTraits<DenseIndex>::IsSigned,THE_INDEX_TYPE_MUST_BE_A_SIGNED_TYPE); +} + +} // end namespace internal + +/** \class DenseBase + * \ingroup Core_Module + * + * \brief Base class for all dense matrices, vectors, and arrays + * + * This class is the base that is inherited by all dense objects (matrix, vector, arrays, + * and related expression types). The common Eigen API for dense objects is contained in this class. + * + * \tparam Derived is the derived type, e.g., a matrix type or an expression. + * + * This class can be extended with the help of the plugin mechanism described on the page + * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_DENSEBASE_PLUGIN. + * + * \sa \blank \ref TopicClassHierarchy + */ +template<typename Derived> class DenseBase +#ifndef EIGEN_PARSED_BY_DOXYGEN + : public DenseCoeffsBase<Derived> +#else + : public DenseCoeffsBase<Derived,DirectWriteAccessors> +#endif // not EIGEN_PARSED_BY_DOXYGEN +{ + public: + + /** Inner iterator type to iterate over the coefficients of a row or column. + * \sa class InnerIterator + */ + typedef Eigen::InnerIterator<Derived> InnerIterator; + + typedef typename internal::traits<Derived>::StorageKind StorageKind; + + /** + * \brief The type used to store indices + * \details This typedef is relevant for types that store multiple indices such as + * PermutationMatrix or Transpositions, otherwise it defaults to Eigen::Index + * \sa \blank \ref TopicPreprocessorDirectives, Eigen::Index, SparseMatrixBase. + */ + typedef typename internal::traits<Derived>::StorageIndex StorageIndex; + + /** The numeric type of the expression' coefficients, e.g. float, double, int or std::complex<float>, etc. */ + typedef typename internal::traits<Derived>::Scalar Scalar; + + /** The numeric type of the expression' coefficients, e.g. float, double, int or std::complex<float>, etc. + * + * It is an alias for the Scalar type */ + typedef Scalar value_type; + + typedef typename NumTraits<Scalar>::Real RealScalar; + typedef DenseCoeffsBase<Derived> Base; + + using Base::derived; + using Base::const_cast_derived; + using Base::rows; + using Base::cols; + using Base::size; + using Base::rowIndexByOuterInner; + using Base::colIndexByOuterInner; + using Base::coeff; + using Base::coeffByOuterInner; + using Base::operator(); + using Base::operator[]; + using Base::x; + using Base::y; + using Base::z; + using Base::w; + using Base::stride; + using Base::innerStride; + using Base::outerStride; + using Base::rowStride; + using Base::colStride; + typedef typename Base::CoeffReturnType CoeffReturnType; + + enum { + + RowsAtCompileTime = internal::traits<Derived>::RowsAtCompileTime, + /**< The number of rows at compile-time. This is just a copy of the value provided + * by the \a Derived type. If a value is not known at compile-time, + * it is set to the \a Dynamic constant. + * \sa MatrixBase::rows(), MatrixBase::cols(), ColsAtCompileTime, SizeAtCompileTime */ + + ColsAtCompileTime = internal::traits<Derived>::ColsAtCompileTime, + /**< The number of columns at compile-time. This is just a copy of the value provided + * by the \a Derived type. If a value is not known at compile-time, + * it is set to the \a Dynamic constant. + * \sa MatrixBase::rows(), MatrixBase::cols(), RowsAtCompileTime, SizeAtCompileTime */ + + + SizeAtCompileTime = (internal::size_at_compile_time<internal::traits<Derived>::RowsAtCompileTime, + internal::traits<Derived>::ColsAtCompileTime>::ret), + /**< This is equal to the number of coefficients, i.e. the number of + * rows times the number of columns, or to \a Dynamic if this is not + * known at compile-time. \sa RowsAtCompileTime, ColsAtCompileTime */ + + MaxRowsAtCompileTime = internal::traits<Derived>::MaxRowsAtCompileTime, + /**< This value is equal to the maximum possible number of rows that this expression + * might have. If this expression might have an arbitrarily high number of rows, + * this value is set to \a Dynamic. + * + * This value is useful to know when evaluating an expression, in order to determine + * whether it is possible to avoid doing a dynamic memory allocation. + * + * \sa RowsAtCompileTime, MaxColsAtCompileTime, MaxSizeAtCompileTime + */ + + MaxColsAtCompileTime = internal::traits<Derived>::MaxColsAtCompileTime, + /**< This value is equal to the maximum possible number of columns that this expression + * might have. If this expression might have an arbitrarily high number of columns, + * this value is set to \a Dynamic. + * + * This value is useful to know when evaluating an expression, in order to determine + * whether it is possible to avoid doing a dynamic memory allocation. + * + * \sa ColsAtCompileTime, MaxRowsAtCompileTime, MaxSizeAtCompileTime + */ + + MaxSizeAtCompileTime = (internal::size_at_compile_time<internal::traits<Derived>::MaxRowsAtCompileTime, + internal::traits<Derived>::MaxColsAtCompileTime>::ret), + /**< This value is equal to the maximum possible number of coefficients that this expression + * might have. If this expression might have an arbitrarily high number of coefficients, + * this value is set to \a Dynamic. + * + * This value is useful to know when evaluating an expression, in order to determine + * whether it is possible to avoid doing a dynamic memory allocation. + * + * \sa SizeAtCompileTime, MaxRowsAtCompileTime, MaxColsAtCompileTime + */ + + IsVectorAtCompileTime = internal::traits<Derived>::MaxRowsAtCompileTime == 1 + || internal::traits<Derived>::MaxColsAtCompileTime == 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 + * a row-vector (if there is only one row). */ + + Flags = internal::traits<Derived>::Flags, + /**< This stores expression \ref flags flags which may or may not be inherited by new expressions + * constructed from this one. See the \ref flags "list of flags". + */ + + IsRowMajor = int(Flags) & RowMajorBit, /**< True if this expression has row-major storage order. */ + + InnerSizeAtCompileTime = int(IsVectorAtCompileTime) ? int(SizeAtCompileTime) + : int(IsRowMajor) ? int(ColsAtCompileTime) : int(RowsAtCompileTime), + + InnerStrideAtCompileTime = internal::inner_stride_at_compile_time<Derived>::ret, + OuterStrideAtCompileTime = internal::outer_stride_at_compile_time<Derived>::ret + }; + + typedef typename internal::find_best_packet<Scalar,SizeAtCompileTime>::type PacketScalar; + + enum { IsPlainObjectBase = 0 }; + + /** The plain matrix type corresponding to this expression. + * \sa PlainObject */ + typedef Matrix<typename internal::traits<Derived>::Scalar, + internal::traits<Derived>::RowsAtCompileTime, + internal::traits<Derived>::ColsAtCompileTime, + AutoAlign | (internal::traits<Derived>::Flags&RowMajorBit ? RowMajor : ColMajor), + internal::traits<Derived>::MaxRowsAtCompileTime, + internal::traits<Derived>::MaxColsAtCompileTime + > PlainMatrix; + + /** The plain array type corresponding to this expression. + * \sa PlainObject */ + typedef Array<typename internal::traits<Derived>::Scalar, + internal::traits<Derived>::RowsAtCompileTime, + internal::traits<Derived>::ColsAtCompileTime, + AutoAlign | (internal::traits<Derived>::Flags&RowMajorBit ? RowMajor : ColMajor), + internal::traits<Derived>::MaxRowsAtCompileTime, + internal::traits<Derived>::MaxColsAtCompileTime + > PlainArray; + + /** \brief The plain matrix or array type corresponding to this expression. + * + * This is not necessarily exactly the return type of eval(). In the case of plain matrices, + * the return type of eval() is a const reference to a matrix, not a matrix! It is however guaranteed + * that the return type of eval() is either PlainObject or const PlainObject&. + */ + typedef typename internal::conditional<internal::is_same<typename internal::traits<Derived>::XprKind,MatrixXpr >::value, + PlainMatrix, PlainArray>::type PlainObject; + + /** \returns the number of nonzero coefficients which is in practice the number + * of stored coefficients. */ + EIGEN_DEVICE_FUNC + inline Index nonZeros() const { return size(); } + + /** \returns the outer size. + * + * \note For a vector, this returns just 1. For a matrix (non-vector), this is the major dimension + * with respect to the \ref TopicStorageOrders "storage order", i.e., the number of columns for a + * column-major matrix, and the number of rows for a row-major matrix. */ + EIGEN_DEVICE_FUNC + Index outerSize() const + { + return IsVectorAtCompileTime ? 1 + : int(IsRowMajor) ? this->rows() : this->cols(); + } + + /** \returns the inner size. + * + * \note For a vector, this is just the size. For a matrix (non-vector), this is the minor dimension + * with respect to the \ref TopicStorageOrders "storage order", i.e., the number of rows for a + * column-major matrix, and the number of columns for a row-major matrix. */ + EIGEN_DEVICE_FUNC + Index innerSize() const + { + return IsVectorAtCompileTime ? this->size() + : int(IsRowMajor) ? this->cols() : this->rows(); + } + + /** Only plain matrices/arrays, not expressions, may be resized; therefore the only useful resize methods are + * Matrix::resize() and Array::resize(). The present method only asserts that the new size equals the old size, and does + * nothing else. + */ + EIGEN_DEVICE_FUNC + void resize(Index newSize) + { + EIGEN_ONLY_USED_FOR_DEBUG(newSize); + eigen_assert(newSize == this->size() + && "DenseBase::resize() does not actually allow to resize."); + } + /** Only plain matrices/arrays, not expressions, may be resized; therefore the only useful resize methods are + * Matrix::resize() and Array::resize(). The present method only asserts that the new size equals the old size, and does + * nothing else. + */ + EIGEN_DEVICE_FUNC + void resize(Index rows, Index cols) + { + EIGEN_ONLY_USED_FOR_DEBUG(rows); + EIGEN_ONLY_USED_FOR_DEBUG(cols); + eigen_assert(rows == this->rows() && cols == this->cols() + && "DenseBase::resize() does not actually allow to resize."); + } + +#ifndef EIGEN_PARSED_BY_DOXYGEN + /** \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; + /** \internal Represents a vector with linearly spaced coefficients that allows random access. */ + typedef CwiseNullaryOp<internal::linspaced_op<Scalar,PacketScalar>,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; + +#endif // not EIGEN_PARSED_BY_DOXYGEN + + /** Copies \a other into *this. \returns a reference to *this. */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Derived& operator=(const DenseBase<OtherDerived>& other); + + /** Special case of the template operator=, in order to prevent the compiler + * from generating a default operator= (issue hit with g++ 4.1) + */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Derived& operator=(const DenseBase& other); + + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + Derived& operator=(const EigenBase<OtherDerived> &other); + + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + Derived& operator+=(const EigenBase<OtherDerived> &other); + + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + Derived& operator-=(const EigenBase<OtherDerived> &other); + + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + Derived& operator=(const ReturnByValue<OtherDerived>& func); + + /** \internal + * Copies \a other into *this without evaluating other. \returns a reference to *this. + * \deprecated */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + Derived& lazyAssign(const DenseBase<OtherDerived>& other); + + EIGEN_DEVICE_FUNC + CommaInitializer<Derived> operator<< (const Scalar& s); + + /** \deprecated it now returns \c *this */ + template<unsigned int Added,unsigned int Removed> + EIGEN_DEPRECATED + const Derived& flagged() const + { return derived(); } + + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + CommaInitializer<Derived> operator<< (const DenseBase<OtherDerived>& other); + + typedef Transpose<Derived> TransposeReturnType; + EIGEN_DEVICE_FUNC + TransposeReturnType transpose(); + typedef typename internal::add_const<Transpose<const Derived> >::type ConstTransposeReturnType; + EIGEN_DEVICE_FUNC + ConstTransposeReturnType transpose() const; + EIGEN_DEVICE_FUNC + void transposeInPlace(); + + EIGEN_DEVICE_FUNC static const ConstantReturnType + Constant(Index rows, Index cols, const Scalar& value); + EIGEN_DEVICE_FUNC static const ConstantReturnType + Constant(Index size, const Scalar& value); + EIGEN_DEVICE_FUNC static const ConstantReturnType + Constant(const Scalar& value); + + EIGEN_DEVICE_FUNC static const SequentialLinSpacedReturnType + LinSpaced(Sequential_t, Index size, const Scalar& low, const Scalar& high); + EIGEN_DEVICE_FUNC static const RandomAccessLinSpacedReturnType + LinSpaced(Index size, const Scalar& low, const Scalar& high); + EIGEN_DEVICE_FUNC static const SequentialLinSpacedReturnType + LinSpaced(Sequential_t, const Scalar& low, const Scalar& high); + EIGEN_DEVICE_FUNC static const RandomAccessLinSpacedReturnType + LinSpaced(const Scalar& low, const Scalar& high); + + template<typename CustomNullaryOp> EIGEN_DEVICE_FUNC + static const CwiseNullaryOp<CustomNullaryOp, PlainObject> + NullaryExpr(Index rows, Index cols, const CustomNullaryOp& func); + template<typename CustomNullaryOp> EIGEN_DEVICE_FUNC + static const CwiseNullaryOp<CustomNullaryOp, PlainObject> + NullaryExpr(Index size, const CustomNullaryOp& func); + template<typename CustomNullaryOp> EIGEN_DEVICE_FUNC + static const CwiseNullaryOp<CustomNullaryOp, PlainObject> + NullaryExpr(const CustomNullaryOp& func); + + EIGEN_DEVICE_FUNC static const ConstantReturnType Zero(Index rows, Index cols); + EIGEN_DEVICE_FUNC static const ConstantReturnType Zero(Index size); + EIGEN_DEVICE_FUNC static const ConstantReturnType Zero(); + EIGEN_DEVICE_FUNC static const ConstantReturnType Ones(Index rows, Index cols); + EIGEN_DEVICE_FUNC static const ConstantReturnType Ones(Index size); + EIGEN_DEVICE_FUNC static const ConstantReturnType Ones(); + + EIGEN_DEVICE_FUNC void fill(const Scalar& value); + EIGEN_DEVICE_FUNC Derived& setConstant(const Scalar& value); + EIGEN_DEVICE_FUNC Derived& setLinSpaced(Index size, const Scalar& low, const Scalar& high); + EIGEN_DEVICE_FUNC Derived& setLinSpaced(const Scalar& low, const Scalar& high); + EIGEN_DEVICE_FUNC Derived& setZero(); + EIGEN_DEVICE_FUNC Derived& setOnes(); + EIGEN_DEVICE_FUNC Derived& setRandom(); + + template<typename OtherDerived> EIGEN_DEVICE_FUNC + bool isApprox(const DenseBase<OtherDerived>& other, + const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const; + EIGEN_DEVICE_FUNC + bool isMuchSmallerThan(const RealScalar& other, + const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const; + template<typename OtherDerived> EIGEN_DEVICE_FUNC + bool isMuchSmallerThan(const DenseBase<OtherDerived>& other, + const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const; + + EIGEN_DEVICE_FUNC bool isApproxToConstant(const Scalar& value, const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const; + EIGEN_DEVICE_FUNC bool isConstant(const Scalar& value, const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const; + EIGEN_DEVICE_FUNC bool isZero(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const; + EIGEN_DEVICE_FUNC bool isOnes(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const; + + inline bool hasNaN() const; + inline bool allFinite() const; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Derived& operator*=(const Scalar& other); + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Derived& operator/=(const Scalar& other); + + typedef typename internal::add_const_on_value_type<typename internal::eval<Derived>::type>::type EvalReturnType; + /** \returns the matrix or vector obtained by evaluating this expression. + * + * Notice that in the case of a plain matrix or vector (not an expression) this function just returns + * a const reference, in order to avoid a useless copy. + * + * \warning Be carefull with eval() and the auto C++ keyword, as detailed in this \link TopicPitfalls_auto_keyword page \endlink. + */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE EvalReturnType eval() const + { + // Even though MSVC does not honor strong inlining when the return type + // is a dynamic matrix, we desperately need strong inlining for fixed + // size types on MSVC. + return typename internal::eval<Derived>::type(derived()); + } + + /** swaps *this with the expression \a other. + * + */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + void swap(const DenseBase<OtherDerived>& other) + { + EIGEN_STATIC_ASSERT(!OtherDerived::IsPlainObjectBase,THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY); + eigen_assert(rows()==other.rows() && cols()==other.cols()); + call_assignment(derived(), other.const_cast_derived(), internal::swap_assign_op<Scalar>()); + } + + /** swaps *this with the matrix or array \a other. + * + */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + void swap(PlainObjectBase<OtherDerived>& other) + { + eigen_assert(rows()==other.rows() && cols()==other.cols()); + call_assignment(derived(), other.derived(), internal::swap_assign_op<Scalar>()); + } + + EIGEN_DEVICE_FUNC inline const NestByValue<Derived> nestByValue() const; + EIGEN_DEVICE_FUNC inline const ForceAlignedAccess<Derived> forceAlignedAccess() const; + EIGEN_DEVICE_FUNC inline ForceAlignedAccess<Derived> forceAlignedAccess(); + template<bool Enable> EIGEN_DEVICE_FUNC + inline const typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type forceAlignedAccessIf() const; + template<bool Enable> EIGEN_DEVICE_FUNC + inline typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type forceAlignedAccessIf(); + + EIGEN_DEVICE_FUNC Scalar sum() const; + EIGEN_DEVICE_FUNC Scalar mean() const; + EIGEN_DEVICE_FUNC Scalar trace() const; + + EIGEN_DEVICE_FUNC Scalar prod() const; + + EIGEN_DEVICE_FUNC typename internal::traits<Derived>::Scalar minCoeff() const; + EIGEN_DEVICE_FUNC typename internal::traits<Derived>::Scalar maxCoeff() const; + + template<typename IndexType> EIGEN_DEVICE_FUNC + typename internal::traits<Derived>::Scalar minCoeff(IndexType* row, IndexType* col) const; + template<typename IndexType> EIGEN_DEVICE_FUNC + typename internal::traits<Derived>::Scalar maxCoeff(IndexType* row, IndexType* col) const; + template<typename IndexType> EIGEN_DEVICE_FUNC + typename internal::traits<Derived>::Scalar minCoeff(IndexType* index) const; + template<typename IndexType> EIGEN_DEVICE_FUNC + typename internal::traits<Derived>::Scalar maxCoeff(IndexType* index) const; + + template<typename BinaryOp> + EIGEN_DEVICE_FUNC + Scalar redux(const BinaryOp& func) const; + + template<typename Visitor> + EIGEN_DEVICE_FUNC + void visit(Visitor& func) const; + + /** \returns a WithFormat proxy object allowing to print a matrix the with given + * format \a fmt. + * + * See class IOFormat for some examples. + * + * \sa class IOFormat, class WithFormat + */ + inline const WithFormat<Derived> format(const IOFormat& fmt) const + { + return WithFormat<Derived>(derived(), fmt); + } + + /** \returns the unique coefficient of a 1x1 expression */ + EIGEN_DEVICE_FUNC + CoeffReturnType value() const + { + EIGEN_STATIC_ASSERT_SIZE_1x1(Derived) + eigen_assert(this->rows() == 1 && this->cols() == 1); + return derived().coeff(0,0); + } + + EIGEN_DEVICE_FUNC bool all() const; + EIGEN_DEVICE_FUNC bool any() const; + EIGEN_DEVICE_FUNC Index count() const; + + typedef VectorwiseOp<Derived, Horizontal> RowwiseReturnType; + typedef const VectorwiseOp<const Derived, Horizontal> ConstRowwiseReturnType; + typedef VectorwiseOp<Derived, Vertical> ColwiseReturnType; + typedef const VectorwiseOp<const Derived, Vertical> ConstColwiseReturnType; + + /** \returns a VectorwiseOp wrapper of *this providing additional partial reduction operations + * + * Example: \include MatrixBase_rowwise.cpp + * Output: \verbinclude MatrixBase_rowwise.out + * + * \sa colwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting + */ + //Code moved here due to a CUDA compiler bug + EIGEN_DEVICE_FUNC inline ConstRowwiseReturnType rowwise() const { + return ConstRowwiseReturnType(derived()); + } + EIGEN_DEVICE_FUNC RowwiseReturnType rowwise(); + + /** \returns a VectorwiseOp wrapper of *this providing additional partial reduction operations + * + * Example: \include MatrixBase_colwise.cpp + * Output: \verbinclude MatrixBase_colwise.out + * + * \sa rowwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting + */ + EIGEN_DEVICE_FUNC inline ConstColwiseReturnType colwise() const { + return ConstColwiseReturnType(derived()); + } + EIGEN_DEVICE_FUNC ColwiseReturnType colwise(); + + typedef CwiseNullaryOp<internal::scalar_random_op<Scalar>,PlainObject> RandomReturnType; + static const RandomReturnType Random(Index rows, Index cols); + static const RandomReturnType Random(Index size); + static const RandomReturnType Random(); + + template<typename ThenDerived,typename ElseDerived> + const Select<Derived,ThenDerived,ElseDerived> + select(const DenseBase<ThenDerived>& thenMatrix, + const DenseBase<ElseDerived>& elseMatrix) const; + + template<typename ThenDerived> + inline const Select<Derived,ThenDerived, typename ThenDerived::ConstantReturnType> + select(const DenseBase<ThenDerived>& thenMatrix, const typename ThenDerived::Scalar& elseScalar) const; + + template<typename ElseDerived> + inline const Select<Derived, typename ElseDerived::ConstantReturnType, ElseDerived > + select(const typename ElseDerived::Scalar& thenScalar, const DenseBase<ElseDerived>& elseMatrix) const; + + template<int p> RealScalar lpNorm() const; + + template<int RowFactor, int ColFactor> + EIGEN_DEVICE_FUNC + const Replicate<Derived,RowFactor,ColFactor> replicate() const; + /** + * \return an expression of the replication of \c *this + * + * Example: \include MatrixBase_replicate_int_int.cpp + * Output: \verbinclude MatrixBase_replicate_int_int.out + * + * \sa VectorwiseOp::replicate(), DenseBase::replicate<int,int>(), class Replicate + */ + //Code moved here due to a CUDA compiler bug + EIGEN_DEVICE_FUNC + const Replicate<Derived, Dynamic, Dynamic> replicate(Index rowFactor, Index colFactor) const + { + return Replicate<Derived, Dynamic, Dynamic>(derived(), rowFactor, colFactor); + } + + typedef Reverse<Derived, BothDirections> ReverseReturnType; + typedef const Reverse<const Derived, BothDirections> ConstReverseReturnType; + EIGEN_DEVICE_FUNC ReverseReturnType reverse(); + /** This is the const version of reverse(). */ + //Code moved here due to a CUDA compiler bug + EIGEN_DEVICE_FUNC ConstReverseReturnType reverse() const + { + return ConstReverseReturnType(derived()); + } + EIGEN_DEVICE_FUNC void reverseInPlace(); + +#define EIGEN_CURRENT_STORAGE_BASE_CLASS Eigen::DenseBase +#define EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL +#define EIGEN_DOC_BLOCK_ADDONS_INNER_PANEL_IF(COND) +# include "../plugins/BlockMethods.h" +# ifdef EIGEN_DENSEBASE_PLUGIN +# include EIGEN_DENSEBASE_PLUGIN +# endif +#undef EIGEN_CURRENT_STORAGE_BASE_CLASS +#undef EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL +#undef EIGEN_DOC_BLOCK_ADDONS_INNER_PANEL_IF + + // disable the use of evalTo for dense objects with a nice compilation error + template<typename Dest> + EIGEN_DEVICE_FUNC + inline void evalTo(Dest& ) const + { + EIGEN_STATIC_ASSERT((internal::is_same<Dest,void>::value),THE_EVAL_EVALTO_FUNCTION_SHOULD_NEVER_BE_CALLED_FOR_DENSE_OBJECTS); + } + + protected: + /** Default constructor. Do nothing. */ + EIGEN_DEVICE_FUNC DenseBase() + { + /* Just checks for self-consistency of the flags. + * Only do it when debugging Eigen, as this borders on paranoiac and could slow compilation down + */ +#ifdef EIGEN_INTERNAL_DEBUGGING + EIGEN_STATIC_ASSERT((EIGEN_IMPLIES(MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1, int(IsRowMajor)) + && EIGEN_IMPLIES(MaxColsAtCompileTime==1 && MaxRowsAtCompileTime!=1, int(!IsRowMajor))), + INVALID_STORAGE_ORDER_FOR_THIS_VECTOR_EXPRESSION) +#endif + } + + private: + EIGEN_DEVICE_FUNC explicit DenseBase(int); + EIGEN_DEVICE_FUNC DenseBase(int,int); + template<typename OtherDerived> EIGEN_DEVICE_FUNC explicit DenseBase(const DenseBase<OtherDerived>&); +}; + +} // end namespace Eigen + +#endif // EIGEN_DENSEBASE_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/DenseCoeffsBase.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/DenseCoeffsBase.h new file mode 100644 index 000000000..c4af48ab6 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/DenseCoeffsBase.h @@ -0,0 +1,681 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2006-2010 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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_DENSECOEFFSBASE_H +#define EIGEN_DENSECOEFFSBASE_H + +namespace Eigen { + +namespace internal { +template<typename T> struct add_const_on_value_type_if_arithmetic +{ + typedef typename conditional<is_arithmetic<T>::value, T, typename add_const_on_value_type<T>::type>::type type; +}; +} + +/** \brief Base class providing read-only coefficient access to matrices and arrays. + * \ingroup Core_Module + * \tparam Derived Type of the derived class + * \tparam #ReadOnlyAccessors Constant indicating read-only access + * + * This class defines the \c operator() \c const function and friends, which can be used to read specific + * entries of a matrix or array. + * + * \sa DenseCoeffsBase<Derived, WriteAccessors>, DenseCoeffsBase<Derived, DirectAccessors>, + * \ref TopicClassHierarchy + */ +template<typename Derived> +class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived> +{ + public: + typedef typename internal::traits<Derived>::StorageKind StorageKind; + typedef typename internal::traits<Derived>::Scalar Scalar; + typedef typename internal::packet_traits<Scalar>::type PacketScalar; + + // Explanation for this CoeffReturnType typedef. + // - This is the return type of the coeff() method. + // - The LvalueBit means exactly that we can offer a coeffRef() method, which means exactly that we can get references + // to coeffs, which means exactly that we can have coeff() return a const reference (as opposed to returning a value). + // - The is_artihmetic check is required since "const int", "const double", etc. will cause warnings on some systems + // while the declaration of "const T", where T is a non arithmetic type does not. Always returning "const Scalar&" is + // not possible, since the underlying expressions might not offer a valid address the reference could be referring to. + typedef typename internal::conditional<bool(internal::traits<Derived>::Flags&LvalueBit), + const Scalar&, + typename internal::conditional<internal::is_arithmetic<Scalar>::value, Scalar, const Scalar>::type + >::type CoeffReturnType; + + typedef typename internal::add_const_on_value_type_if_arithmetic< + typename internal::packet_traits<Scalar>::type + >::type PacketReturnType; + + typedef EigenBase<Derived> Base; + using Base::rows; + using Base::cols; + using Base::size; + using Base::derived; + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Index rowIndexByOuterInner(Index outer, Index inner) const + { + return int(Derived::RowsAtCompileTime) == 1 ? 0 + : int(Derived::ColsAtCompileTime) == 1 ? inner + : int(Derived::Flags)&RowMajorBit ? outer + : inner; + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Index colIndexByOuterInner(Index outer, Index inner) const + { + return int(Derived::ColsAtCompileTime) == 1 ? 0 + : int(Derived::RowsAtCompileTime) == 1 ? inner + : int(Derived::Flags)&RowMajorBit ? inner + : outer; + } + + /** Short version: don't use this function, use + * \link operator()(Index,Index) const \endlink instead. + * + * Long version: this function is similar to + * \link operator()(Index,Index) const \endlink, but without the assertion. + * Use this for limiting the performance cost of debugging code when doing + * repeated coefficient access. Only use this when it is guaranteed that the + * parameters \a row and \a col are in range. + * + * If EIGEN_INTERNAL_DEBUGGING is defined, an assertion will be made, making this + * function equivalent to \link operator()(Index,Index) const \endlink. + * + * \sa operator()(Index,Index) const, coeffRef(Index,Index), coeff(Index) const + */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE CoeffReturnType coeff(Index row, Index col) const + { + eigen_internal_assert(row >= 0 && row < rows() + && col >= 0 && col < cols()); + return internal::evaluator<Derived>(derived()).coeff(row,col); + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE CoeffReturnType coeffByOuterInner(Index outer, Index inner) const + { + return coeff(rowIndexByOuterInner(outer, inner), + colIndexByOuterInner(outer, inner)); + } + + /** \returns the coefficient at given the given row and column. + * + * \sa operator()(Index,Index), operator[](Index) + */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE CoeffReturnType operator()(Index row, Index col) const + { + eigen_assert(row >= 0 && row < rows() + && col >= 0 && col < cols()); + return coeff(row, col); + } + + /** Short version: don't use this function, use + * \link operator[](Index) const \endlink instead. + * + * Long version: this function is similar to + * \link operator[](Index) const \endlink, but without the assertion. + * Use this for limiting the performance cost of debugging code when doing + * repeated coefficient access. Only use this when it is guaranteed that the + * parameter \a index is in range. + * + * If EIGEN_INTERNAL_DEBUGGING is defined, an assertion will be made, making this + * function equivalent to \link operator[](Index) const \endlink. + * + * \sa operator[](Index) const, coeffRef(Index), coeff(Index,Index) const + */ + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE CoeffReturnType + coeff(Index index) const + { + EIGEN_STATIC_ASSERT(internal::evaluator<Derived>::Flags & LinearAccessBit, + THIS_COEFFICIENT_ACCESSOR_TAKING_ONE_ACCESS_IS_ONLY_FOR_EXPRESSIONS_ALLOWING_LINEAR_ACCESS) + eigen_internal_assert(index >= 0 && index < size()); + return internal::evaluator<Derived>(derived()).coeff(index); + } + + + /** \returns the coefficient at given index. + * + * This method is allowed only for vector expressions, and for matrix expressions having the LinearAccessBit. + * + * \sa operator[](Index), operator()(Index,Index) const, x() const, y() const, + * z() const, w() const + */ + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE CoeffReturnType + operator[](Index index) const + { + EIGEN_STATIC_ASSERT(Derived::IsVectorAtCompileTime, + THE_BRACKET_OPERATOR_IS_ONLY_FOR_VECTORS__USE_THE_PARENTHESIS_OPERATOR_INSTEAD) + eigen_assert(index >= 0 && index < size()); + return coeff(index); + } + + /** \returns the coefficient at given index. + * + * This is synonymous to operator[](Index) const. + * + * This method is allowed only for vector expressions, and for matrix expressions having the LinearAccessBit. + * + * \sa operator[](Index), operator()(Index,Index) const, x() const, y() const, + * z() const, w() const + */ + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE CoeffReturnType + operator()(Index index) const + { + eigen_assert(index >= 0 && index < size()); + return coeff(index); + } + + /** equivalent to operator[](0). */ + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE CoeffReturnType + x() const { return (*this)[0]; } + + /** equivalent to operator[](1). */ + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE CoeffReturnType + y() const + { + EIGEN_STATIC_ASSERT(Derived::SizeAtCompileTime==-1 || Derived::SizeAtCompileTime>=2, OUT_OF_RANGE_ACCESS); + return (*this)[1]; + } + + /** equivalent to operator[](2). */ + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE CoeffReturnType + z() const + { + EIGEN_STATIC_ASSERT(Derived::SizeAtCompileTime==-1 || Derived::SizeAtCompileTime>=3, OUT_OF_RANGE_ACCESS); + return (*this)[2]; + } + + /** equivalent to operator[](3). */ + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE CoeffReturnType + w() const + { + EIGEN_STATIC_ASSERT(Derived::SizeAtCompileTime==-1 || Derived::SizeAtCompileTime>=4, OUT_OF_RANGE_ACCESS); + return (*this)[3]; + } + + /** \internal + * \returns the packet of coefficients starting at the given row and column. It is your responsibility + * to ensure that a packet really starts there. This method is only available on expressions having the + * PacketAccessBit. + * + * The \a LoadMode parameter may have the value \a #Aligned or \a #Unaligned. Its effect is to select + * the appropriate vectorization instruction. Aligned access is faster, but is only possible for packets + * starting at an address which is a multiple of the packet size. + */ + + template<int LoadMode> + EIGEN_STRONG_INLINE PacketReturnType packet(Index row, Index col) const + { + typedef typename internal::packet_traits<Scalar>::type DefaultPacketType; + eigen_internal_assert(row >= 0 && row < rows() && col >= 0 && col < cols()); + return internal::evaluator<Derived>(derived()).template packet<LoadMode,DefaultPacketType>(row,col); + } + + + /** \internal */ + template<int LoadMode> + EIGEN_STRONG_INLINE PacketReturnType packetByOuterInner(Index outer, Index inner) const + { + return packet<LoadMode>(rowIndexByOuterInner(outer, inner), + colIndexByOuterInner(outer, inner)); + } + + /** \internal + * \returns the packet of coefficients starting at the given index. It is your responsibility + * to ensure that a packet really starts there. This method is only available on expressions having the + * PacketAccessBit and the LinearAccessBit. + * + * The \a LoadMode parameter may have the value \a #Aligned or \a #Unaligned. Its effect is to select + * the appropriate vectorization instruction. Aligned access is faster, but is only possible for packets + * starting at an address which is a multiple of the packet size. + */ + + template<int LoadMode> + EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const + { + EIGEN_STATIC_ASSERT(internal::evaluator<Derived>::Flags & LinearAccessBit, + THIS_COEFFICIENT_ACCESSOR_TAKING_ONE_ACCESS_IS_ONLY_FOR_EXPRESSIONS_ALLOWING_LINEAR_ACCESS) + typedef typename internal::packet_traits<Scalar>::type DefaultPacketType; + eigen_internal_assert(index >= 0 && index < size()); + return internal::evaluator<Derived>(derived()).template packet<LoadMode,DefaultPacketType>(index); + } + + protected: + // explanation: DenseBase is doing "using ..." on the methods from DenseCoeffsBase. + // But some methods are only available in the DirectAccess case. + // So we add dummy methods here with these names, so that "using... " doesn't fail. + // It's not private so that the child class DenseBase can access them, and it's not public + // either since it's an implementation detail, so has to be protected. + void coeffRef(); + void coeffRefByOuterInner(); + void writePacket(); + void writePacketByOuterInner(); + void copyCoeff(); + void copyCoeffByOuterInner(); + void copyPacket(); + void copyPacketByOuterInner(); + void stride(); + void innerStride(); + void outerStride(); + void rowStride(); + void colStride(); +}; + +/** \brief Base class providing read/write coefficient access to matrices and arrays. + * \ingroup Core_Module + * \tparam Derived Type of the derived class + * \tparam #WriteAccessors Constant indicating read/write access + * + * This class defines the non-const \c operator() function and friends, which can be used to write specific + * entries of a matrix or array. This class inherits DenseCoeffsBase<Derived, ReadOnlyAccessors> which + * defines the const variant for reading specific entries. + * + * \sa DenseCoeffsBase<Derived, DirectAccessors>, \ref TopicClassHierarchy + */ +template<typename Derived> +class DenseCoeffsBase<Derived, WriteAccessors> : public DenseCoeffsBase<Derived, ReadOnlyAccessors> +{ + public: + + typedef DenseCoeffsBase<Derived, ReadOnlyAccessors> Base; + + typedef typename internal::traits<Derived>::StorageKind StorageKind; + typedef typename internal::traits<Derived>::Scalar Scalar; + typedef typename internal::packet_traits<Scalar>::type PacketScalar; + typedef typename NumTraits<Scalar>::Real RealScalar; + + using Base::coeff; + using Base::rows; + using Base::cols; + using Base::size; + using Base::derived; + using Base::rowIndexByOuterInner; + using Base::colIndexByOuterInner; + using Base::operator[]; + using Base::operator(); + using Base::x; + using Base::y; + using Base::z; + using Base::w; + + /** Short version: don't use this function, use + * \link operator()(Index,Index) \endlink instead. + * + * Long version: this function is similar to + * \link operator()(Index,Index) \endlink, but without the assertion. + * Use this for limiting the performance cost of debugging code when doing + * repeated coefficient access. Only use this when it is guaranteed that the + * parameters \a row and \a col are in range. + * + * If EIGEN_INTERNAL_DEBUGGING is defined, an assertion will be made, making this + * function equivalent to \link operator()(Index,Index) \endlink. + * + * \sa operator()(Index,Index), coeff(Index, Index) const, coeffRef(Index) + */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar& coeffRef(Index row, Index col) + { + eigen_internal_assert(row >= 0 && row < rows() + && col >= 0 && col < cols()); + return internal::evaluator<Derived>(derived()).coeffRef(row,col); + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar& + coeffRefByOuterInner(Index outer, Index inner) + { + return coeffRef(rowIndexByOuterInner(outer, inner), + colIndexByOuterInner(outer, inner)); + } + + /** \returns a reference to the coefficient at given the given row and column. + * + * \sa operator[](Index) + */ + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar& + operator()(Index row, Index col) + { + eigen_assert(row >= 0 && row < rows() + && col >= 0 && col < cols()); + return coeffRef(row, col); + } + + + /** Short version: don't use this function, use + * \link operator[](Index) \endlink instead. + * + * Long version: this function is similar to + * \link operator[](Index) \endlink, but without the assertion. + * Use this for limiting the performance cost of debugging code when doing + * repeated coefficient access. Only use this when it is guaranteed that the + * parameters \a row and \a col are in range. + * + * If EIGEN_INTERNAL_DEBUGGING is defined, an assertion will be made, making this + * function equivalent to \link operator[](Index) \endlink. + * + * \sa operator[](Index), coeff(Index) const, coeffRef(Index,Index) + */ + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar& + coeffRef(Index index) + { + EIGEN_STATIC_ASSERT(internal::evaluator<Derived>::Flags & LinearAccessBit, + THIS_COEFFICIENT_ACCESSOR_TAKING_ONE_ACCESS_IS_ONLY_FOR_EXPRESSIONS_ALLOWING_LINEAR_ACCESS) + eigen_internal_assert(index >= 0 && index < size()); + return internal::evaluator<Derived>(derived()).coeffRef(index); + } + + /** \returns a reference to the coefficient at given index. + * + * This method is allowed only for vector expressions, and for matrix expressions having the LinearAccessBit. + * + * \sa operator[](Index) const, operator()(Index,Index), x(), y(), z(), w() + */ + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar& + operator[](Index index) + { + EIGEN_STATIC_ASSERT(Derived::IsVectorAtCompileTime, + THE_BRACKET_OPERATOR_IS_ONLY_FOR_VECTORS__USE_THE_PARENTHESIS_OPERATOR_INSTEAD) + eigen_assert(index >= 0 && index < size()); + return coeffRef(index); + } + + /** \returns a reference to the coefficient at given index. + * + * This is synonymous to operator[](Index). + * + * This method is allowed only for vector expressions, and for matrix expressions having the LinearAccessBit. + * + * \sa operator[](Index) const, operator()(Index,Index), x(), y(), z(), w() + */ + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar& + operator()(Index index) + { + eigen_assert(index >= 0 && index < size()); + return coeffRef(index); + } + + /** equivalent to operator[](0). */ + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar& + x() { return (*this)[0]; } + + /** equivalent to operator[](1). */ + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar& + y() + { + EIGEN_STATIC_ASSERT(Derived::SizeAtCompileTime==-1 || Derived::SizeAtCompileTime>=2, OUT_OF_RANGE_ACCESS); + return (*this)[1]; + } + + /** equivalent to operator[](2). */ + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar& + z() + { + EIGEN_STATIC_ASSERT(Derived::SizeAtCompileTime==-1 || Derived::SizeAtCompileTime>=3, OUT_OF_RANGE_ACCESS); + return (*this)[2]; + } + + /** equivalent to operator[](3). */ + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar& + w() + { + EIGEN_STATIC_ASSERT(Derived::SizeAtCompileTime==-1 || Derived::SizeAtCompileTime>=4, OUT_OF_RANGE_ACCESS); + return (*this)[3]; + } +}; + +/** \brief Base class providing direct read-only coefficient access to matrices and arrays. + * \ingroup Core_Module + * \tparam Derived Type of the derived class + * \tparam #DirectAccessors Constant indicating direct access + * + * This class defines functions to work with strides which can be used to access entries directly. This class + * inherits DenseCoeffsBase<Derived, ReadOnlyAccessors> which defines functions to access entries read-only using + * \c operator() . + * + * \sa \blank \ref TopicClassHierarchy + */ +template<typename Derived> +class DenseCoeffsBase<Derived, DirectAccessors> : public DenseCoeffsBase<Derived, ReadOnlyAccessors> +{ + public: + + typedef DenseCoeffsBase<Derived, ReadOnlyAccessors> Base; + typedef typename internal::traits<Derived>::Scalar Scalar; + typedef typename NumTraits<Scalar>::Real RealScalar; + + using Base::rows; + using Base::cols; + using Base::size; + using Base::derived; + + /** \returns the pointer increment between two consecutive elements within a slice in the inner direction. + * + * \sa outerStride(), rowStride(), colStride() + */ + EIGEN_DEVICE_FUNC + inline Index innerStride() const + { + return derived().innerStride(); + } + + /** \returns the pointer increment between two consecutive inner slices (for example, between two consecutive columns + * in a column-major matrix). + * + * \sa innerStride(), rowStride(), colStride() + */ + EIGEN_DEVICE_FUNC + inline Index outerStride() const + { + return derived().outerStride(); + } + + // FIXME shall we remove it ? + inline Index stride() const + { + return Derived::IsVectorAtCompileTime ? innerStride() : outerStride(); + } + + /** \returns the pointer increment between two consecutive rows. + * + * \sa innerStride(), outerStride(), colStride() + */ + EIGEN_DEVICE_FUNC + inline Index rowStride() const + { + return Derived::IsRowMajor ? outerStride() : innerStride(); + } + + /** \returns the pointer increment between two consecutive columns. + * + * \sa innerStride(), outerStride(), rowStride() + */ + EIGEN_DEVICE_FUNC + inline Index colStride() const + { + return Derived::IsRowMajor ? innerStride() : outerStride(); + } +}; + +/** \brief Base class providing direct read/write coefficient access to matrices and arrays. + * \ingroup Core_Module + * \tparam Derived Type of the derived class + * \tparam #DirectWriteAccessors Constant indicating direct access + * + * This class defines functions to work with strides which can be used to access entries directly. This class + * inherits DenseCoeffsBase<Derived, WriteAccessors> which defines functions to access entries read/write using + * \c operator(). + * + * \sa \blank \ref TopicClassHierarchy + */ +template<typename Derived> +class DenseCoeffsBase<Derived, DirectWriteAccessors> + : public DenseCoeffsBase<Derived, WriteAccessors> +{ + public: + + typedef DenseCoeffsBase<Derived, WriteAccessors> Base; + typedef typename internal::traits<Derived>::Scalar Scalar; + typedef typename NumTraits<Scalar>::Real RealScalar; + + using Base::rows; + using Base::cols; + using Base::size; + using Base::derived; + + /** \returns the pointer increment between two consecutive elements within a slice in the inner direction. + * + * \sa outerStride(), rowStride(), colStride() + */ + EIGEN_DEVICE_FUNC + inline Index innerStride() const + { + return derived().innerStride(); + } + + /** \returns the pointer increment between two consecutive inner slices (for example, between two consecutive columns + * in a column-major matrix). + * + * \sa innerStride(), rowStride(), colStride() + */ + EIGEN_DEVICE_FUNC + inline Index outerStride() const + { + return derived().outerStride(); + } + + // FIXME shall we remove it ? + inline Index stride() const + { + return Derived::IsVectorAtCompileTime ? innerStride() : outerStride(); + } + + /** \returns the pointer increment between two consecutive rows. + * + * \sa innerStride(), outerStride(), colStride() + */ + EIGEN_DEVICE_FUNC + inline Index rowStride() const + { + return Derived::IsRowMajor ? outerStride() : innerStride(); + } + + /** \returns the pointer increment between two consecutive columns. + * + * \sa innerStride(), outerStride(), rowStride() + */ + EIGEN_DEVICE_FUNC + inline Index colStride() const + { + return Derived::IsRowMajor ? innerStride() : outerStride(); + } +}; + +namespace internal { + +template<int Alignment, typename Derived, bool JustReturnZero> +struct first_aligned_impl +{ + static inline Index run(const Derived&) + { return 0; } +}; + +template<int Alignment, typename Derived> +struct first_aligned_impl<Alignment, Derived, false> +{ + static inline Index run(const Derived& m) + { + return internal::first_aligned<Alignment>(m.data(), m.size()); + } +}; + +/** \internal \returns the index of the first element of the array stored by \a m that is properly aligned with respect to \a Alignment for vectorization. + * + * \tparam Alignment requested alignment in Bytes. + * + * There is also the variant first_aligned(const Scalar*, Integer) defined in Memory.h. See it for more + * documentation. + */ +template<int Alignment, typename Derived> +static inline Index first_aligned(const DenseBase<Derived>& m) +{ + enum { ReturnZero = (int(evaluator<Derived>::Alignment) >= Alignment) || !(Derived::Flags & DirectAccessBit) }; + return first_aligned_impl<Alignment, Derived, ReturnZero>::run(m.derived()); +} + +template<typename Derived> +static inline Index first_default_aligned(const DenseBase<Derived>& m) +{ + typedef typename Derived::Scalar Scalar; + typedef typename packet_traits<Scalar>::type DefaultPacketType; + return internal::first_aligned<int(unpacket_traits<DefaultPacketType>::alignment),Derived>(m); +} + +template<typename Derived, bool HasDirectAccess = has_direct_access<Derived>::ret> +struct inner_stride_at_compile_time +{ + enum { ret = traits<Derived>::InnerStrideAtCompileTime }; +}; + +template<typename Derived> +struct inner_stride_at_compile_time<Derived, false> +{ + enum { ret = 0 }; +}; + +template<typename Derived, bool HasDirectAccess = has_direct_access<Derived>::ret> +struct outer_stride_at_compile_time +{ + enum { ret = traits<Derived>::OuterStrideAtCompileTime }; +}; + +template<typename Derived> +struct outer_stride_at_compile_time<Derived, false> +{ + enum { ret = 0 }; +}; + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_DENSECOEFFSBASE_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/DenseStorage.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/DenseStorage.h new file mode 100644 index 000000000..7958feeb9 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/DenseStorage.h @@ -0,0 +1,570 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2006-2009 Benoit Jacob <jacob.benoit.1@gmail.com> +// Copyright (C) 2010-2013 Hauke Heibel <hauke.heibel@gmail.com> +// +// 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_MATRIXSTORAGE_H +#define EIGEN_MATRIXSTORAGE_H + +#ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN + #define EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(X) X; EIGEN_DENSE_STORAGE_CTOR_PLUGIN; +#else + #define EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(X) +#endif + +namespace Eigen { + +namespace internal { + +struct constructor_without_unaligned_array_assert {}; + +template<typename T, int Size> +EIGEN_DEVICE_FUNC +void check_static_allocation_size() +{ + // if EIGEN_STACK_ALLOCATION_LIMIT is defined to 0, then no limit + #if EIGEN_STACK_ALLOCATION_LIMIT + EIGEN_STATIC_ASSERT(Size * sizeof(T) <= EIGEN_STACK_ALLOCATION_LIMIT, OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG); + #endif +} + +/** \internal + * Static array. If the MatrixOrArrayOptions require auto-alignment, the array will be automatically aligned: + * to 16 bytes boundary if the total size is a multiple of 16 bytes. + */ +template <typename T, int Size, int MatrixOrArrayOptions, + int Alignment = (MatrixOrArrayOptions&DontAlign) ? 0 + : compute_default_alignment<T,Size>::value > +struct plain_array +{ + T array[Size]; + + EIGEN_DEVICE_FUNC + plain_array() + { + check_static_allocation_size<T,Size>(); + } + + EIGEN_DEVICE_FUNC + plain_array(constructor_without_unaligned_array_assert) + { + check_static_allocation_size<T,Size>(); + } +}; + +#if defined(EIGEN_DISABLE_UNALIGNED_ARRAY_ASSERT) + #define EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(sizemask) +#elif EIGEN_GNUC_AT_LEAST(4,7) + // GCC 4.7 is too aggressive in its optimizations and remove the alignement test based on the fact the array is declared to be aligned. + // See this bug report: http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53900 + // Hiding the origin of the array pointer behind a function argument seems to do the trick even if the function is inlined: + template<typename PtrType> + EIGEN_ALWAYS_INLINE PtrType eigen_unaligned_array_assert_workaround_gcc47(PtrType array) { return array; } + #define EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(sizemask) \ + eigen_assert((internal::UIntPtr(eigen_unaligned_array_assert_workaround_gcc47(array)) & (sizemask)) == 0 \ + && "this assertion is explained here: " \ + "http://eigen.tuxfamily.org/dox-devel/group__TopicUnalignedArrayAssert.html" \ + " **** READ THIS WEB PAGE !!! ****"); +#else + #define EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(sizemask) \ + eigen_assert((internal::UIntPtr(array) & (sizemask)) == 0 \ + && "this assertion is explained here: " \ + "http://eigen.tuxfamily.org/dox-devel/group__TopicUnalignedArrayAssert.html" \ + " **** READ THIS WEB PAGE !!! ****"); +#endif + +template <typename T, int Size, int MatrixOrArrayOptions> +struct plain_array<T, Size, MatrixOrArrayOptions, 8> +{ + EIGEN_ALIGN_TO_BOUNDARY(8) T array[Size]; + + EIGEN_DEVICE_FUNC + plain_array() + { + EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(7); + check_static_allocation_size<T,Size>(); + } + + EIGEN_DEVICE_FUNC + plain_array(constructor_without_unaligned_array_assert) + { + check_static_allocation_size<T,Size>(); + } +}; + +template <typename T, int Size, int MatrixOrArrayOptions> +struct plain_array<T, Size, MatrixOrArrayOptions, 16> +{ + EIGEN_ALIGN_TO_BOUNDARY(16) T array[Size]; + + EIGEN_DEVICE_FUNC + plain_array() + { + EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(15); + check_static_allocation_size<T,Size>(); + } + + EIGEN_DEVICE_FUNC + plain_array(constructor_without_unaligned_array_assert) + { + check_static_allocation_size<T,Size>(); + } +}; + +template <typename T, int Size, int MatrixOrArrayOptions> +struct plain_array<T, Size, MatrixOrArrayOptions, 32> +{ + EIGEN_ALIGN_TO_BOUNDARY(32) T array[Size]; + + EIGEN_DEVICE_FUNC + plain_array() + { + EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(31); + check_static_allocation_size<T,Size>(); + } + + EIGEN_DEVICE_FUNC + plain_array(constructor_without_unaligned_array_assert) + { + check_static_allocation_size<T,Size>(); + } +}; + +template <typename T, int Size, int MatrixOrArrayOptions> +struct plain_array<T, Size, MatrixOrArrayOptions, 64> +{ + EIGEN_ALIGN_TO_BOUNDARY(64) T array[Size]; + + EIGEN_DEVICE_FUNC + plain_array() + { + EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(63); + check_static_allocation_size<T,Size>(); + } + + EIGEN_DEVICE_FUNC + plain_array(constructor_without_unaligned_array_assert) + { + check_static_allocation_size<T,Size>(); + } +}; + +template <typename T, int MatrixOrArrayOptions, int Alignment> +struct plain_array<T, 0, MatrixOrArrayOptions, Alignment> +{ + T array[1]; + EIGEN_DEVICE_FUNC plain_array() {} + EIGEN_DEVICE_FUNC plain_array(constructor_without_unaligned_array_assert) {} +}; + +} // end namespace internal + +/** \internal + * + * \class DenseStorage + * \ingroup Core_Module + * + * \brief Stores the data of a matrix + * + * This class stores the data of fixed-size, dynamic-size or mixed matrices + * in a way as compact as possible. + * + * \sa Matrix + */ +template<typename T, int Size, int _Rows, int _Cols, int _Options> class DenseStorage; + +// purely fixed-size matrix +template<typename T, int Size, int _Rows, int _Cols, int _Options> class DenseStorage +{ + internal::plain_array<T,Size,_Options> m_data; + public: + EIGEN_DEVICE_FUNC DenseStorage() { + EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(Index size = Size) + } + EIGEN_DEVICE_FUNC + explicit DenseStorage(internal::constructor_without_unaligned_array_assert) + : m_data(internal::constructor_without_unaligned_array_assert()) {} + EIGEN_DEVICE_FUNC + DenseStorage(const DenseStorage& other) : m_data(other.m_data) { + EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(Index size = Size) + } + EIGEN_DEVICE_FUNC + DenseStorage& operator=(const DenseStorage& other) + { + if (this != &other) m_data = other.m_data; + return *this; + } + EIGEN_DEVICE_FUNC DenseStorage(Index size, Index rows, Index cols) { + EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({}) + eigen_internal_assert(size==rows*cols && rows==_Rows && cols==_Cols); + EIGEN_UNUSED_VARIABLE(size); + EIGEN_UNUSED_VARIABLE(rows); + EIGEN_UNUSED_VARIABLE(cols); + } + EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); } + EIGEN_DEVICE_FUNC static Index rows(void) {return _Rows;} + EIGEN_DEVICE_FUNC static Index cols(void) {return _Cols;} + EIGEN_DEVICE_FUNC void conservativeResize(Index,Index,Index) {} + EIGEN_DEVICE_FUNC void resize(Index,Index,Index) {} + EIGEN_DEVICE_FUNC const T *data() const { return m_data.array; } + EIGEN_DEVICE_FUNC T *data() { return m_data.array; } +}; + +// null matrix +template<typename T, int _Rows, int _Cols, int _Options> class DenseStorage<T, 0, _Rows, _Cols, _Options> +{ + public: + EIGEN_DEVICE_FUNC DenseStorage() {} + EIGEN_DEVICE_FUNC explicit DenseStorage(internal::constructor_without_unaligned_array_assert) {} + EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage&) {} + EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage&) { return *this; } + EIGEN_DEVICE_FUNC DenseStorage(Index,Index,Index) {} + EIGEN_DEVICE_FUNC void swap(DenseStorage& ) {} + EIGEN_DEVICE_FUNC static Index rows(void) {return _Rows;} + EIGEN_DEVICE_FUNC static Index cols(void) {return _Cols;} + EIGEN_DEVICE_FUNC void conservativeResize(Index,Index,Index) {} + EIGEN_DEVICE_FUNC void resize(Index,Index,Index) {} + EIGEN_DEVICE_FUNC const T *data() const { return 0; } + EIGEN_DEVICE_FUNC T *data() { return 0; } +}; + +// more specializations for null matrices; these are necessary to resolve ambiguities +template<typename T, int _Options> class DenseStorage<T, 0, Dynamic, Dynamic, _Options> +: public DenseStorage<T, 0, 0, 0, _Options> { }; + +template<typename T, int _Rows, int _Options> class DenseStorage<T, 0, _Rows, Dynamic, _Options> +: public DenseStorage<T, 0, 0, 0, _Options> { }; + +template<typename T, int _Cols, int _Options> class DenseStorage<T, 0, Dynamic, _Cols, _Options> +: public DenseStorage<T, 0, 0, 0, _Options> { }; + +// dynamic-size matrix with fixed-size storage +template<typename T, int Size, int _Options> class DenseStorage<T, Size, Dynamic, Dynamic, _Options> +{ + internal::plain_array<T,Size,_Options> m_data; + Index m_rows; + Index m_cols; + public: + EIGEN_DEVICE_FUNC DenseStorage() : m_rows(0), m_cols(0) {} + EIGEN_DEVICE_FUNC explicit DenseStorage(internal::constructor_without_unaligned_array_assert) + : m_data(internal::constructor_without_unaligned_array_assert()), m_rows(0), m_cols(0) {} + EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_rows(other.m_rows), m_cols(other.m_cols) {} + EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other) + { + if (this != &other) + { + m_data = other.m_data; + m_rows = other.m_rows; + m_cols = other.m_cols; + } + return *this; + } + EIGEN_DEVICE_FUNC DenseStorage(Index, Index rows, Index cols) : m_rows(rows), m_cols(cols) {} + EIGEN_DEVICE_FUNC void swap(DenseStorage& other) + { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); std::swap(m_cols,other.m_cols); } + EIGEN_DEVICE_FUNC Index rows() const {return m_rows;} + EIGEN_DEVICE_FUNC Index cols() const {return m_cols;} + EIGEN_DEVICE_FUNC void conservativeResize(Index, Index rows, Index cols) { m_rows = rows; m_cols = cols; } + EIGEN_DEVICE_FUNC void resize(Index, Index rows, Index cols) { m_rows = rows; m_cols = cols; } + EIGEN_DEVICE_FUNC const T *data() const { return m_data.array; } + EIGEN_DEVICE_FUNC T *data() { return m_data.array; } +}; + +// dynamic-size matrix with fixed-size storage and fixed width +template<typename T, int Size, int _Cols, int _Options> class DenseStorage<T, Size, Dynamic, _Cols, _Options> +{ + internal::plain_array<T,Size,_Options> m_data; + Index m_rows; + public: + EIGEN_DEVICE_FUNC DenseStorage() : m_rows(0) {} + EIGEN_DEVICE_FUNC explicit DenseStorage(internal::constructor_without_unaligned_array_assert) + : m_data(internal::constructor_without_unaligned_array_assert()), m_rows(0) {} + EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_rows(other.m_rows) {} + EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other) + { + if (this != &other) + { + m_data = other.m_data; + m_rows = other.m_rows; + } + return *this; + } + EIGEN_DEVICE_FUNC DenseStorage(Index, Index rows, Index) : m_rows(rows) {} + EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); } + EIGEN_DEVICE_FUNC Index rows(void) const {return m_rows;} + EIGEN_DEVICE_FUNC Index cols(void) const {return _Cols;} + EIGEN_DEVICE_FUNC void conservativeResize(Index, Index rows, Index) { m_rows = rows; } + EIGEN_DEVICE_FUNC void resize(Index, Index rows, Index) { m_rows = rows; } + EIGEN_DEVICE_FUNC const T *data() const { return m_data.array; } + EIGEN_DEVICE_FUNC T *data() { return m_data.array; } +}; + +// dynamic-size matrix with fixed-size storage and fixed height +template<typename T, int Size, int _Rows, int _Options> class DenseStorage<T, Size, _Rows, Dynamic, _Options> +{ + internal::plain_array<T,Size,_Options> m_data; + Index m_cols; + public: + EIGEN_DEVICE_FUNC DenseStorage() : m_cols(0) {} + EIGEN_DEVICE_FUNC explicit DenseStorage(internal::constructor_without_unaligned_array_assert) + : m_data(internal::constructor_without_unaligned_array_assert()), m_cols(0) {} + EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_cols(other.m_cols) {} + EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other) + { + if (this != &other) + { + m_data = other.m_data; + m_cols = other.m_cols; + } + return *this; + } + EIGEN_DEVICE_FUNC DenseStorage(Index, Index, Index cols) : m_cols(cols) {} + EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); } + EIGEN_DEVICE_FUNC Index rows(void) const {return _Rows;} + EIGEN_DEVICE_FUNC Index cols(void) const {return m_cols;} + void conservativeResize(Index, Index, Index cols) { m_cols = cols; } + void resize(Index, Index, Index cols) { m_cols = cols; } + EIGEN_DEVICE_FUNC const T *data() const { return m_data.array; } + EIGEN_DEVICE_FUNC T *data() { return m_data.array; } +}; + +// purely dynamic matrix. +template<typename T, int _Options> class DenseStorage<T, Dynamic, Dynamic, Dynamic, _Options> +{ + T *m_data; + Index m_rows; + Index m_cols; + public: + EIGEN_DEVICE_FUNC DenseStorage() : m_data(0), m_rows(0), m_cols(0) {} + EIGEN_DEVICE_FUNC explicit DenseStorage(internal::constructor_without_unaligned_array_assert) + : m_data(0), m_rows(0), m_cols(0) {} + EIGEN_DEVICE_FUNC DenseStorage(Index size, Index rows, Index cols) + : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_rows(rows), m_cols(cols) + { + EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({}) + eigen_internal_assert(size==rows*cols && rows>=0 && cols >=0); + } + EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other) + : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(other.m_rows*other.m_cols)) + , m_rows(other.m_rows) + , m_cols(other.m_cols) + { + EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(Index size = m_rows*m_cols) + internal::smart_copy(other.m_data, other.m_data+other.m_rows*other.m_cols, m_data); + } + EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other) + { + if (this != &other) + { + DenseStorage tmp(other); + this->swap(tmp); + } + return *this; + } +#if EIGEN_HAS_RVALUE_REFERENCES + EIGEN_DEVICE_FUNC + DenseStorage(DenseStorage&& other) EIGEN_NOEXCEPT + : m_data(std::move(other.m_data)) + , m_rows(std::move(other.m_rows)) + , m_cols(std::move(other.m_cols)) + { + other.m_data = nullptr; + other.m_rows = 0; + other.m_cols = 0; + } + EIGEN_DEVICE_FUNC + DenseStorage& operator=(DenseStorage&& other) EIGEN_NOEXCEPT + { + using std::swap; + swap(m_data, other.m_data); + swap(m_rows, other.m_rows); + swap(m_cols, other.m_cols); + return *this; + } +#endif + EIGEN_DEVICE_FUNC ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, m_rows*m_cols); } + EIGEN_DEVICE_FUNC void swap(DenseStorage& other) + { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); std::swap(m_cols,other.m_cols); } + EIGEN_DEVICE_FUNC Index rows(void) const {return m_rows;} + EIGEN_DEVICE_FUNC Index cols(void) const {return m_cols;} + void conservativeResize(Index size, Index rows, Index cols) + { + m_data = internal::conditional_aligned_realloc_new_auto<T,(_Options&DontAlign)==0>(m_data, size, m_rows*m_cols); + m_rows = rows; + m_cols = cols; + } + EIGEN_DEVICE_FUNC void resize(Index size, Index rows, Index cols) + { + if(size != m_rows*m_cols) + { + internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, m_rows*m_cols); + if (size) + m_data = internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size); + else + m_data = 0; + EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({}) + } + m_rows = rows; + m_cols = cols; + } + EIGEN_DEVICE_FUNC const T *data() const { return m_data; } + EIGEN_DEVICE_FUNC T *data() { return m_data; } +}; + +// matrix with dynamic width and fixed height (so that matrix has dynamic size). +template<typename T, int _Rows, int _Options> class DenseStorage<T, Dynamic, _Rows, Dynamic, _Options> +{ + T *m_data; + Index m_cols; + public: + EIGEN_DEVICE_FUNC DenseStorage() : m_data(0), m_cols(0) {} + explicit DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_cols(0) {} + EIGEN_DEVICE_FUNC DenseStorage(Index size, Index rows, Index cols) : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_cols(cols) + { + EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({}) + eigen_internal_assert(size==rows*cols && rows==_Rows && cols >=0); + EIGEN_UNUSED_VARIABLE(rows); + } + EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other) + : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(_Rows*other.m_cols)) + , m_cols(other.m_cols) + { + EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(Index size = m_cols*_Rows) + internal::smart_copy(other.m_data, other.m_data+_Rows*m_cols, m_data); + } + EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other) + { + if (this != &other) + { + DenseStorage tmp(other); + this->swap(tmp); + } + return *this; + } +#if EIGEN_HAS_RVALUE_REFERENCES + EIGEN_DEVICE_FUNC + DenseStorage(DenseStorage&& other) EIGEN_NOEXCEPT + : m_data(std::move(other.m_data)) + , m_cols(std::move(other.m_cols)) + { + other.m_data = nullptr; + other.m_cols = 0; + } + EIGEN_DEVICE_FUNC + DenseStorage& operator=(DenseStorage&& other) EIGEN_NOEXCEPT + { + using std::swap; + swap(m_data, other.m_data); + swap(m_cols, other.m_cols); + return *this; + } +#endif + EIGEN_DEVICE_FUNC ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Rows*m_cols); } + EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); } + EIGEN_DEVICE_FUNC static Index rows(void) {return _Rows;} + EIGEN_DEVICE_FUNC Index cols(void) const {return m_cols;} + EIGEN_DEVICE_FUNC void conservativeResize(Index size, Index, Index cols) + { + m_data = internal::conditional_aligned_realloc_new_auto<T,(_Options&DontAlign)==0>(m_data, size, _Rows*m_cols); + m_cols = cols; + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void resize(Index size, Index, Index cols) + { + if(size != _Rows*m_cols) + { + internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Rows*m_cols); + if (size) + m_data = internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size); + else + m_data = 0; + EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({}) + } + m_cols = cols; + } + EIGEN_DEVICE_FUNC const T *data() const { return m_data; } + EIGEN_DEVICE_FUNC T *data() { return m_data; } +}; + +// matrix with dynamic height and fixed width (so that matrix has dynamic size). +template<typename T, int _Cols, int _Options> class DenseStorage<T, Dynamic, Dynamic, _Cols, _Options> +{ + T *m_data; + Index m_rows; + public: + EIGEN_DEVICE_FUNC DenseStorage() : m_data(0), m_rows(0) {} + explicit DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_rows(0) {} + EIGEN_DEVICE_FUNC DenseStorage(Index size, Index rows, Index cols) : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_rows(rows) + { + EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({}) + eigen_internal_assert(size==rows*cols && rows>=0 && cols == _Cols); + EIGEN_UNUSED_VARIABLE(cols); + } + EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other) + : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(other.m_rows*_Cols)) + , m_rows(other.m_rows) + { + EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(Index size = m_rows*_Cols) + internal::smart_copy(other.m_data, other.m_data+other.m_rows*_Cols, m_data); + } + EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other) + { + if (this != &other) + { + DenseStorage tmp(other); + this->swap(tmp); + } + return *this; + } +#if EIGEN_HAS_RVALUE_REFERENCES + EIGEN_DEVICE_FUNC + DenseStorage(DenseStorage&& other) EIGEN_NOEXCEPT + : m_data(std::move(other.m_data)) + , m_rows(std::move(other.m_rows)) + { + other.m_data = nullptr; + other.m_rows = 0; + } + EIGEN_DEVICE_FUNC + DenseStorage& operator=(DenseStorage&& other) EIGEN_NOEXCEPT + { + using std::swap; + swap(m_data, other.m_data); + swap(m_rows, other.m_rows); + return *this; + } +#endif + EIGEN_DEVICE_FUNC ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Cols*m_rows); } + EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); } + EIGEN_DEVICE_FUNC Index rows(void) const {return m_rows;} + EIGEN_DEVICE_FUNC static Index cols(void) {return _Cols;} + void conservativeResize(Index size, Index rows, Index) + { + m_data = internal::conditional_aligned_realloc_new_auto<T,(_Options&DontAlign)==0>(m_data, size, m_rows*_Cols); + m_rows = rows; + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void resize(Index size, Index rows, Index) + { + if(size != m_rows*_Cols) + { + internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Cols*m_rows); + if (size) + m_data = internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size); + else + m_data = 0; + EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({}) + } + m_rows = rows; + } + EIGEN_DEVICE_FUNC const T *data() const { return m_data; } + EIGEN_DEVICE_FUNC T *data() { return m_data; } +}; + +} // end namespace Eigen + +#endif // EIGEN_MATRIX_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/Diagonal.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Diagonal.h new file mode 100644 index 000000000..49e711257 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Diagonal.h @@ -0,0 +1,257 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2007-2009 Benoit Jacob <jacob.benoit.1@gmail.com> +// Copyright (C) 2009-2010 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_DIAGONAL_H +#define EIGEN_DIAGONAL_H + +namespace Eigen { + +/** \class Diagonal + * \ingroup Core_Module + * + * \brief Expression of a diagonal/subdiagonal/superdiagonal in a matrix + * + * \param MatrixType the type of the object in which we are taking a sub/main/super diagonal + * \param DiagIndex the index of the sub/super diagonal. The default is 0 and it means the main diagonal. + * A positive value means a superdiagonal, a negative value means a subdiagonal. + * You can also use DynamicIndex so the index can be set at runtime. + * + * The matrix is not required to be square. + * + * This class represents an expression of the main diagonal, or any sub/super diagonal + * of a square matrix. It is the return type of MatrixBase::diagonal() and MatrixBase::diagonal(Index) and most of the + * time this is the only way it is used. + * + * \sa MatrixBase::diagonal(), MatrixBase::diagonal(Index) + */ + +namespace internal { +template<typename MatrixType, int DiagIndex> +struct traits<Diagonal<MatrixType,DiagIndex> > + : traits<MatrixType> +{ + typedef typename ref_selector<MatrixType>::type MatrixTypeNested; + typedef typename remove_reference<MatrixTypeNested>::type _MatrixTypeNested; + typedef typename MatrixType::StorageKind StorageKind; + enum { + RowsAtCompileTime = (int(DiagIndex) == DynamicIndex || int(MatrixType::SizeAtCompileTime) == Dynamic) ? Dynamic + : (EIGEN_PLAIN_ENUM_MIN(MatrixType::RowsAtCompileTime - EIGEN_PLAIN_ENUM_MAX(-DiagIndex, 0), + MatrixType::ColsAtCompileTime - EIGEN_PLAIN_ENUM_MAX( DiagIndex, 0))), + ColsAtCompileTime = 1, + MaxRowsAtCompileTime = int(MatrixType::MaxSizeAtCompileTime) == Dynamic ? Dynamic + : DiagIndex == DynamicIndex ? EIGEN_SIZE_MIN_PREFER_FIXED(MatrixType::MaxRowsAtCompileTime, + MatrixType::MaxColsAtCompileTime) + : (EIGEN_PLAIN_ENUM_MIN(MatrixType::MaxRowsAtCompileTime - EIGEN_PLAIN_ENUM_MAX(-DiagIndex, 0), + MatrixType::MaxColsAtCompileTime - EIGEN_PLAIN_ENUM_MAX( DiagIndex, 0))), + MaxColsAtCompileTime = 1, + MaskLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0, + Flags = (unsigned int)_MatrixTypeNested::Flags & (RowMajorBit | MaskLvalueBit | DirectAccessBit) & ~RowMajorBit, // FIXME DirectAccessBit should not be handled by expressions + MatrixTypeOuterStride = outer_stride_at_compile_time<MatrixType>::ret, + InnerStrideAtCompileTime = MatrixTypeOuterStride == Dynamic ? Dynamic : MatrixTypeOuterStride+1, + OuterStrideAtCompileTime = 0 + }; +}; +} + +template<typename MatrixType, int _DiagIndex> class Diagonal + : public internal::dense_xpr_base< Diagonal<MatrixType,_DiagIndex> >::type +{ + public: + + enum { DiagIndex = _DiagIndex }; + typedef typename internal::dense_xpr_base<Diagonal>::type Base; + EIGEN_DENSE_PUBLIC_INTERFACE(Diagonal) + + EIGEN_DEVICE_FUNC + explicit inline Diagonal(MatrixType& matrix, Index a_index = DiagIndex) : m_matrix(matrix), m_index(a_index) {} + + EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Diagonal) + + EIGEN_DEVICE_FUNC + inline Index rows() const + { + return m_index.value()<0 ? numext::mini<Index>(m_matrix.cols(),m_matrix.rows()+m_index.value()) + : numext::mini<Index>(m_matrix.rows(),m_matrix.cols()-m_index.value()); + } + + EIGEN_DEVICE_FUNC + inline Index cols() const { return 1; } + + EIGEN_DEVICE_FUNC + inline Index innerStride() const + { + return m_matrix.outerStride() + 1; + } + + EIGEN_DEVICE_FUNC + inline Index outerStride() const + { + return 0; + } + + typedef typename internal::conditional< + internal::is_lvalue<MatrixType>::value, + Scalar, + const Scalar + >::type ScalarWithConstIfNotLvalue; + + EIGEN_DEVICE_FUNC + inline ScalarWithConstIfNotLvalue* data() { return &(m_matrix.coeffRef(rowOffset(), colOffset())); } + EIGEN_DEVICE_FUNC + inline const Scalar* data() const { return &(m_matrix.coeffRef(rowOffset(), colOffset())); } + + EIGEN_DEVICE_FUNC + inline Scalar& coeffRef(Index row, Index) + { + EIGEN_STATIC_ASSERT_LVALUE(MatrixType) + return m_matrix.coeffRef(row+rowOffset(), row+colOffset()); + } + + EIGEN_DEVICE_FUNC + inline const Scalar& coeffRef(Index row, Index) const + { + return m_matrix.coeffRef(row+rowOffset(), row+colOffset()); + } + + EIGEN_DEVICE_FUNC + inline CoeffReturnType coeff(Index row, Index) const + { + return m_matrix.coeff(row+rowOffset(), row+colOffset()); + } + + EIGEN_DEVICE_FUNC + inline Scalar& coeffRef(Index idx) + { + EIGEN_STATIC_ASSERT_LVALUE(MatrixType) + return m_matrix.coeffRef(idx+rowOffset(), idx+colOffset()); + } + + EIGEN_DEVICE_FUNC + inline const Scalar& coeffRef(Index idx) const + { + return m_matrix.coeffRef(idx+rowOffset(), idx+colOffset()); + } + + EIGEN_DEVICE_FUNC + inline CoeffReturnType coeff(Index idx) const + { + return m_matrix.coeff(idx+rowOffset(), idx+colOffset()); + } + + EIGEN_DEVICE_FUNC + inline const typename internal::remove_all<typename MatrixType::Nested>::type& + nestedExpression() const + { + return m_matrix; + } + + EIGEN_DEVICE_FUNC + inline Index index() const + { + return m_index.value(); + } + + protected: + typename internal::ref_selector<MatrixType>::non_const_type m_matrix; + const internal::variable_if_dynamicindex<Index, DiagIndex> m_index; + + private: + // some compilers may fail to optimize std::max etc in case of compile-time constants... + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Index absDiagIndex() const { return m_index.value()>0 ? m_index.value() : -m_index.value(); } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Index rowOffset() const { return m_index.value()>0 ? 0 : -m_index.value(); } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Index colOffset() const { return m_index.value()>0 ? m_index.value() : 0; } + // trigger a compile-time error if someone try to call packet + template<int LoadMode> typename MatrixType::PacketReturnType packet(Index) const; + template<int LoadMode> typename MatrixType::PacketReturnType packet(Index,Index) const; +}; + +/** \returns an expression of the main diagonal of the matrix \c *this + * + * \c *this is not required to be square. + * + * Example: \include MatrixBase_diagonal.cpp + * Output: \verbinclude MatrixBase_diagonal.out + * + * \sa class Diagonal */ +template<typename Derived> +inline typename MatrixBase<Derived>::DiagonalReturnType +MatrixBase<Derived>::diagonal() +{ + return DiagonalReturnType(derived()); +} + +/** This is the const version of diagonal(). */ +template<typename Derived> +inline typename MatrixBase<Derived>::ConstDiagonalReturnType +MatrixBase<Derived>::diagonal() const +{ + return ConstDiagonalReturnType(derived()); +} + +/** \returns an expression of the \a DiagIndex-th sub or super diagonal of the matrix \c *this + * + * \c *this is not required to be square. + * + * The template parameter \a DiagIndex represent a super diagonal if \a DiagIndex > 0 + * and a sub diagonal otherwise. \a DiagIndex == 0 is equivalent to the main diagonal. + * + * Example: \include MatrixBase_diagonal_int.cpp + * Output: \verbinclude MatrixBase_diagonal_int.out + * + * \sa MatrixBase::diagonal(), class Diagonal */ +template<typename Derived> +inline typename MatrixBase<Derived>::DiagonalDynamicIndexReturnType +MatrixBase<Derived>::diagonal(Index index) +{ + return DiagonalDynamicIndexReturnType(derived(), index); +} + +/** This is the const version of diagonal(Index). */ +template<typename Derived> +inline typename MatrixBase<Derived>::ConstDiagonalDynamicIndexReturnType +MatrixBase<Derived>::diagonal(Index index) const +{ + return ConstDiagonalDynamicIndexReturnType(derived(), index); +} + +/** \returns an expression of the \a DiagIndex-th sub or super diagonal of the matrix \c *this + * + * \c *this is not required to be square. + * + * The template parameter \a DiagIndex represent a super diagonal if \a DiagIndex > 0 + * and a sub diagonal otherwise. \a DiagIndex == 0 is equivalent to the main diagonal. + * + * Example: \include MatrixBase_diagonal_template_int.cpp + * Output: \verbinclude MatrixBase_diagonal_template_int.out + * + * \sa MatrixBase::diagonal(), class Diagonal */ +template<typename Derived> +template<int Index_> +inline typename MatrixBase<Derived>::template DiagonalIndexReturnType<Index_>::Type +MatrixBase<Derived>::diagonal() +{ + return typename DiagonalIndexReturnType<Index_>::Type(derived()); +} + +/** This is the const version of diagonal<int>(). */ +template<typename Derived> +template<int Index_> +inline typename MatrixBase<Derived>::template ConstDiagonalIndexReturnType<Index_>::Type +MatrixBase<Derived>::diagonal() const +{ + return typename ConstDiagonalIndexReturnType<Index_>::Type(derived()); +} + +} // end namespace Eigen + +#endif // EIGEN_DIAGONAL_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/DiagonalMatrix.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/DiagonalMatrix.h new file mode 100644 index 000000000..ecfdce8ef --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/DiagonalMatrix.h @@ -0,0 +1,343 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2007-2009 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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_DIAGONALMATRIX_H +#define EIGEN_DIAGONALMATRIX_H + +namespace Eigen { + +#ifndef EIGEN_PARSED_BY_DOXYGEN +template<typename Derived> +class DiagonalBase : public EigenBase<Derived> +{ + public: + typedef typename internal::traits<Derived>::DiagonalVectorType DiagonalVectorType; + typedef typename DiagonalVectorType::Scalar Scalar; + typedef typename DiagonalVectorType::RealScalar RealScalar; + typedef typename internal::traits<Derived>::StorageKind StorageKind; + typedef typename internal::traits<Derived>::StorageIndex StorageIndex; + + enum { + RowsAtCompileTime = DiagonalVectorType::SizeAtCompileTime, + ColsAtCompileTime = DiagonalVectorType::SizeAtCompileTime, + MaxRowsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime, + MaxColsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime, + IsVectorAtCompileTime = 0, + Flags = NoPreferredStorageOrderBit + }; + + typedef Matrix<Scalar, RowsAtCompileTime, ColsAtCompileTime, 0, MaxRowsAtCompileTime, MaxColsAtCompileTime> DenseMatrixType; + typedef DenseMatrixType DenseType; + typedef DiagonalMatrix<Scalar,DiagonalVectorType::SizeAtCompileTime,DiagonalVectorType::MaxSizeAtCompileTime> PlainObject; + + EIGEN_DEVICE_FUNC + inline const Derived& derived() const { return *static_cast<const Derived*>(this); } + EIGEN_DEVICE_FUNC + inline Derived& derived() { return *static_cast<Derived*>(this); } + + EIGEN_DEVICE_FUNC + DenseMatrixType toDenseMatrix() const { return derived(); } + + EIGEN_DEVICE_FUNC + inline const DiagonalVectorType& diagonal() const { return derived().diagonal(); } + EIGEN_DEVICE_FUNC + inline DiagonalVectorType& diagonal() { return derived().diagonal(); } + + EIGEN_DEVICE_FUNC + inline Index rows() const { return diagonal().size(); } + EIGEN_DEVICE_FUNC + inline Index cols() const { return diagonal().size(); } + + template<typename MatrixDerived> + EIGEN_DEVICE_FUNC + const Product<Derived,MatrixDerived,LazyProduct> + operator*(const MatrixBase<MatrixDerived> &matrix) const + { + return Product<Derived, MatrixDerived, LazyProduct>(derived(),matrix.derived()); + } + + typedef DiagonalWrapper<const CwiseUnaryOp<internal::scalar_inverse_op<Scalar>, const DiagonalVectorType> > InverseReturnType; + EIGEN_DEVICE_FUNC + inline const InverseReturnType + inverse() const + { + return InverseReturnType(diagonal().cwiseInverse()); + } + + EIGEN_DEVICE_FUNC + inline const DiagonalWrapper<const EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(DiagonalVectorType,Scalar,product) > + operator*(const Scalar& scalar) const + { + return DiagonalWrapper<const EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(DiagonalVectorType,Scalar,product) >(diagonal() * scalar); + } + EIGEN_DEVICE_FUNC + friend inline const DiagonalWrapper<const EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(Scalar,DiagonalVectorType,product) > + operator*(const Scalar& scalar, const DiagonalBase& other) + { + return DiagonalWrapper<const EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(Scalar,DiagonalVectorType,product) >(scalar * other.diagonal()); + } +}; + +#endif + +/** \class DiagonalMatrix + * \ingroup Core_Module + * + * \brief Represents a diagonal matrix with its storage + * + * \param _Scalar the type of coefficients + * \param SizeAtCompileTime the dimension of the matrix, or Dynamic + * \param MaxSizeAtCompileTime the dimension of the matrix, or Dynamic. This parameter is optional and defaults + * to SizeAtCompileTime. Most of the time, you do not need to specify it. + * + * \sa class DiagonalWrapper + */ + +namespace internal { +template<typename _Scalar, int SizeAtCompileTime, int MaxSizeAtCompileTime> +struct traits<DiagonalMatrix<_Scalar,SizeAtCompileTime,MaxSizeAtCompileTime> > + : traits<Matrix<_Scalar,SizeAtCompileTime,SizeAtCompileTime,0,MaxSizeAtCompileTime,MaxSizeAtCompileTime> > +{ + typedef Matrix<_Scalar,SizeAtCompileTime,1,0,MaxSizeAtCompileTime,1> DiagonalVectorType; + typedef DiagonalShape StorageKind; + enum { + Flags = LvalueBit | NoPreferredStorageOrderBit + }; +}; +} +template<typename _Scalar, int SizeAtCompileTime, int MaxSizeAtCompileTime> +class DiagonalMatrix + : public DiagonalBase<DiagonalMatrix<_Scalar,SizeAtCompileTime,MaxSizeAtCompileTime> > +{ + public: + #ifndef EIGEN_PARSED_BY_DOXYGEN + typedef typename internal::traits<DiagonalMatrix>::DiagonalVectorType DiagonalVectorType; + typedef const DiagonalMatrix& Nested; + typedef _Scalar Scalar; + typedef typename internal::traits<DiagonalMatrix>::StorageKind StorageKind; + typedef typename internal::traits<DiagonalMatrix>::StorageIndex StorageIndex; + #endif + + protected: + + DiagonalVectorType m_diagonal; + + public: + + /** const version of diagonal(). */ + EIGEN_DEVICE_FUNC + inline const DiagonalVectorType& diagonal() const { return m_diagonal; } + /** \returns a reference to the stored vector of diagonal coefficients. */ + EIGEN_DEVICE_FUNC + inline DiagonalVectorType& diagonal() { return m_diagonal; } + + /** Default constructor without initialization */ + EIGEN_DEVICE_FUNC + inline DiagonalMatrix() {} + + /** Constructs a diagonal matrix with given dimension */ + EIGEN_DEVICE_FUNC + explicit inline DiagonalMatrix(Index dim) : m_diagonal(dim) {} + + /** 2D constructor. */ + EIGEN_DEVICE_FUNC + inline DiagonalMatrix(const Scalar& x, const Scalar& y) : m_diagonal(x,y) {} + + /** 3D constructor. */ + EIGEN_DEVICE_FUNC + inline DiagonalMatrix(const Scalar& x, const Scalar& y, const Scalar& z) : m_diagonal(x,y,z) {} + + /** Copy constructor. */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + inline DiagonalMatrix(const DiagonalBase<OtherDerived>& other) : m_diagonal(other.diagonal()) {} + + #ifndef EIGEN_PARSED_BY_DOXYGEN + /** copy constructor. prevent a default copy constructor from hiding the other templated constructor */ + inline DiagonalMatrix(const DiagonalMatrix& other) : m_diagonal(other.diagonal()) {} + #endif + + /** generic constructor from expression of the diagonal coefficients */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + explicit inline DiagonalMatrix(const MatrixBase<OtherDerived>& other) : m_diagonal(other) + {} + + /** Copy operator. */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + DiagonalMatrix& operator=(const DiagonalBase<OtherDerived>& other) + { + m_diagonal = other.diagonal(); + return *this; + } + + #ifndef EIGEN_PARSED_BY_DOXYGEN + /** This is a special case of the templated operator=. Its purpose is to + * prevent a default operator= from hiding the templated operator=. + */ + EIGEN_DEVICE_FUNC + DiagonalMatrix& operator=(const DiagonalMatrix& other) + { + m_diagonal = other.diagonal(); + return *this; + } + #endif + + /** Resizes to given size. */ + EIGEN_DEVICE_FUNC + inline void resize(Index size) { m_diagonal.resize(size); } + /** Sets all coefficients to zero. */ + EIGEN_DEVICE_FUNC + inline void setZero() { m_diagonal.setZero(); } + /** Resizes and sets all coefficients to zero. */ + EIGEN_DEVICE_FUNC + inline void setZero(Index size) { m_diagonal.setZero(size); } + /** Sets this matrix to be the identity matrix of the current size. */ + EIGEN_DEVICE_FUNC + inline void setIdentity() { m_diagonal.setOnes(); } + /** Sets this matrix to be the identity matrix of the given size. */ + EIGEN_DEVICE_FUNC + inline void setIdentity(Index size) { m_diagonal.setOnes(size); } +}; + +/** \class DiagonalWrapper + * \ingroup Core_Module + * + * \brief Expression of a diagonal matrix + * + * \param _DiagonalVectorType the type of the vector of diagonal coefficients + * + * This class is an expression of a diagonal matrix, but not storing its own vector of diagonal coefficients, + * instead wrapping an existing vector expression. It is the return type of MatrixBase::asDiagonal() + * and most of the time this is the only way that it is used. + * + * \sa class DiagonalMatrix, class DiagonalBase, MatrixBase::asDiagonal() + */ + +namespace internal { +template<typename _DiagonalVectorType> +struct traits<DiagonalWrapper<_DiagonalVectorType> > +{ + typedef _DiagonalVectorType DiagonalVectorType; + typedef typename DiagonalVectorType::Scalar Scalar; + typedef typename DiagonalVectorType::StorageIndex StorageIndex; + typedef DiagonalShape StorageKind; + typedef typename traits<DiagonalVectorType>::XprKind XprKind; + enum { + RowsAtCompileTime = DiagonalVectorType::SizeAtCompileTime, + ColsAtCompileTime = DiagonalVectorType::SizeAtCompileTime, + MaxRowsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime, + MaxColsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime, + Flags = (traits<DiagonalVectorType>::Flags & LvalueBit) | NoPreferredStorageOrderBit + }; +}; +} + +template<typename _DiagonalVectorType> +class DiagonalWrapper + : public DiagonalBase<DiagonalWrapper<_DiagonalVectorType> >, internal::no_assignment_operator +{ + public: + #ifndef EIGEN_PARSED_BY_DOXYGEN + typedef _DiagonalVectorType DiagonalVectorType; + typedef DiagonalWrapper Nested; + #endif + + /** Constructor from expression of diagonal coefficients to wrap. */ + EIGEN_DEVICE_FUNC + explicit inline DiagonalWrapper(DiagonalVectorType& a_diagonal) : m_diagonal(a_diagonal) {} + + /** \returns a const reference to the wrapped expression of diagonal coefficients. */ + EIGEN_DEVICE_FUNC + const DiagonalVectorType& diagonal() const { return m_diagonal; } + + protected: + typename DiagonalVectorType::Nested m_diagonal; +}; + +/** \returns a pseudo-expression of a diagonal matrix with *this as vector of diagonal coefficients + * + * \only_for_vectors + * + * Example: \include MatrixBase_asDiagonal.cpp + * Output: \verbinclude MatrixBase_asDiagonal.out + * + * \sa class DiagonalWrapper, class DiagonalMatrix, diagonal(), isDiagonal() + **/ +template<typename Derived> +inline const DiagonalWrapper<const Derived> +MatrixBase<Derived>::asDiagonal() const +{ + return DiagonalWrapper<const Derived>(derived()); +} + +/** \returns true if *this is approximately equal to a diagonal matrix, + * within the precision given by \a prec. + * + * Example: \include MatrixBase_isDiagonal.cpp + * Output: \verbinclude MatrixBase_isDiagonal.out + * + * \sa asDiagonal() + */ +template<typename Derived> +bool MatrixBase<Derived>::isDiagonal(const RealScalar& prec) const +{ + if(cols() != rows()) return false; + RealScalar maxAbsOnDiagonal = static_cast<RealScalar>(-1); + for(Index j = 0; j < cols(); ++j) + { + RealScalar absOnDiagonal = numext::abs(coeff(j,j)); + if(absOnDiagonal > maxAbsOnDiagonal) maxAbsOnDiagonal = absOnDiagonal; + } + for(Index j = 0; j < cols(); ++j) + for(Index i = 0; i < j; ++i) + { + if(!internal::isMuchSmallerThan(coeff(i, j), maxAbsOnDiagonal, prec)) return false; + if(!internal::isMuchSmallerThan(coeff(j, i), maxAbsOnDiagonal, prec)) return false; + } + return true; +} + +namespace internal { + +template<> struct storage_kind_to_shape<DiagonalShape> { typedef DiagonalShape Shape; }; + +struct Diagonal2Dense {}; + +template<> struct AssignmentKind<DenseShape,DiagonalShape> { typedef Diagonal2Dense Kind; }; + +// Diagonal matrix to Dense assignment +template< typename DstXprType, typename SrcXprType, typename Functor> +struct Assignment<DstXprType, SrcXprType, Functor, Diagonal2Dense> +{ + static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &/*func*/) + { + Index dstRows = src.rows(); + Index dstCols = src.cols(); + if((dst.rows()!=dstRows) || (dst.cols()!=dstCols)) + dst.resize(dstRows, dstCols); + + dst.setZero(); + dst.diagonal() = src.diagonal(); + } + + static void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &/*func*/) + { dst.diagonal() += src.diagonal(); } + + static void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &/*func*/) + { dst.diagonal() -= src.diagonal(); } +}; + +} // namespace internal + +} // end namespace Eigen + +#endif // EIGEN_DIAGONALMATRIX_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/DiagonalProduct.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/DiagonalProduct.h new file mode 100644 index 000000000..d372b938f --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/DiagonalProduct.h @@ -0,0 +1,28 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2007-2009 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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_DIAGONALPRODUCT_H +#define EIGEN_DIAGONALPRODUCT_H + +namespace Eigen { + +/** \returns the diagonal matrix product of \c *this by the diagonal matrix \a diagonal. + */ +template<typename Derived> +template<typename DiagonalDerived> +inline const Product<Derived, DiagonalDerived, LazyProduct> +MatrixBase<Derived>::operator*(const DiagonalBase<DiagonalDerived> &a_diagonal) const +{ + return Product<Derived, DiagonalDerived, LazyProduct>(derived(),a_diagonal.derived()); +} + +} // end namespace Eigen + +#endif // EIGEN_DIAGONALPRODUCT_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/Dot.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Dot.h new file mode 100644 index 000000000..06ef18b8b --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Dot.h @@ -0,0 +1,315 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2006-2008, 2010 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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_DOT_H +#define EIGEN_DOT_H + +namespace Eigen { + +namespace internal { + +// helper function for dot(). The problem is that if we put that in the body of dot(), then upon calling dot +// with mismatched types, the compiler emits errors about failing to instantiate cwiseProduct BEFORE +// looking at the static assertions. Thus this is a trick to get better compile errors. +template<typename T, typename U, +// the NeedToTranspose condition here is taken straight from Assign.h + bool NeedToTranspose = T::IsVectorAtCompileTime + && U::IsVectorAtCompileTime + && ((int(T::RowsAtCompileTime) == 1 && int(U::ColsAtCompileTime) == 1) + | // FIXME | instead of || to please GCC 4.4.0 stupid warning "suggest parentheses around &&". + // revert to || as soon as not needed anymore. + (int(T::ColsAtCompileTime) == 1 && int(U::RowsAtCompileTime) == 1)) +> +struct dot_nocheck +{ + typedef scalar_conj_product_op<typename traits<T>::Scalar,typename traits<U>::Scalar> conj_prod; + typedef typename conj_prod::result_type ResScalar; + EIGEN_DEVICE_FUNC + static inline ResScalar run(const MatrixBase<T>& a, const MatrixBase<U>& b) + { + return a.template binaryExpr<conj_prod>(b).sum(); + } +}; + +template<typename T, typename U> +struct dot_nocheck<T, U, true> +{ + typedef scalar_conj_product_op<typename traits<T>::Scalar,typename traits<U>::Scalar> conj_prod; + typedef typename conj_prod::result_type ResScalar; + EIGEN_DEVICE_FUNC + static inline ResScalar run(const MatrixBase<T>& a, const MatrixBase<U>& b) + { + return a.transpose().template binaryExpr<conj_prod>(b).sum(); + } +}; + +} // end namespace internal + +/** \fn MatrixBase::dot + * \returns the dot product of *this with other. + * + * \only_for_vectors + * + * \note If the scalar type is complex numbers, then this function returns the hermitian + * (sesquilinear) dot product, conjugate-linear in the first variable and linear in the + * second variable. + * + * \sa squaredNorm(), norm() + */ +template<typename Derived> +template<typename OtherDerived> +EIGEN_DEVICE_FUNC +typename ScalarBinaryOpTraits<typename internal::traits<Derived>::Scalar,typename internal::traits<OtherDerived>::Scalar>::ReturnType +MatrixBase<Derived>::dot(const MatrixBase<OtherDerived>& other) const +{ + EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) + EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived) + EIGEN_STATIC_ASSERT_SAME_VECTOR_SIZE(Derived,OtherDerived) +#if !(defined(EIGEN_NO_STATIC_ASSERT) && defined(EIGEN_NO_DEBUG)) + typedef internal::scalar_conj_product_op<Scalar,typename OtherDerived::Scalar> func; + EIGEN_CHECK_BINARY_COMPATIBILIY(func,Scalar,typename OtherDerived::Scalar); +#endif + + eigen_assert(size() == other.size()); + + return internal::dot_nocheck<Derived,OtherDerived>::run(*this, other); +} + +//---------- implementation of L2 norm and related functions ---------- + +/** \returns, for vectors, the squared \em l2 norm of \c *this, and for matrices the Frobenius norm. + * In both cases, it consists in the sum of the square of all the matrix entries. + * For vectors, this is also equals to the dot product of \c *this with itself. + * + * \sa dot(), norm(), lpNorm() + */ +template<typename Derived> +EIGEN_STRONG_INLINE typename NumTraits<typename internal::traits<Derived>::Scalar>::Real MatrixBase<Derived>::squaredNorm() const +{ + return numext::real((*this).cwiseAbs2().sum()); +} + +/** \returns, for vectors, the \em l2 norm of \c *this, and for matrices the Frobenius norm. + * In both cases, it consists in the square root of the sum of the square of all the matrix entries. + * For vectors, this is also equals to the square root of the dot product of \c *this with itself. + * + * \sa lpNorm(), dot(), squaredNorm() + */ +template<typename Derived> +inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real MatrixBase<Derived>::norm() const +{ + return numext::sqrt(squaredNorm()); +} + +/** \returns an expression of the quotient of \c *this by its own norm. + * + * \warning If the input vector is too small (i.e., this->norm()==0), + * then this function returns a copy of the input. + * + * \only_for_vectors + * + * \sa norm(), normalize() + */ +template<typename Derived> +inline const typename MatrixBase<Derived>::PlainObject +MatrixBase<Derived>::normalized() const +{ + typedef typename internal::nested_eval<Derived,2>::type _Nested; + _Nested n(derived()); + RealScalar z = n.squaredNorm(); + // NOTE: after extensive benchmarking, this conditional does not impact performance, at least on recent x86 CPU + if(z>RealScalar(0)) + return n / numext::sqrt(z); + else + return n; +} + +/** Normalizes the vector, i.e. divides it by its own norm. + * + * \only_for_vectors + * + * \warning If the input vector is too small (i.e., this->norm()==0), then \c *this is left unchanged. + * + * \sa norm(), normalized() + */ +template<typename Derived> +inline void MatrixBase<Derived>::normalize() +{ + RealScalar z = squaredNorm(); + // NOTE: after extensive benchmarking, this conditional does not impact performance, at least on recent x86 CPU + if(z>RealScalar(0)) + derived() /= numext::sqrt(z); +} + +/** \returns an expression of the quotient of \c *this by its own norm while avoiding underflow and overflow. + * + * \only_for_vectors + * + * This method is analogue to the normalized() method, but it reduces the risk of + * underflow and overflow when computing the norm. + * + * \warning If the input vector is too small (i.e., this->norm()==0), + * then this function returns a copy of the input. + * + * \sa stableNorm(), stableNormalize(), normalized() + */ +template<typename Derived> +inline const typename MatrixBase<Derived>::PlainObject +MatrixBase<Derived>::stableNormalized() const +{ + typedef typename internal::nested_eval<Derived,3>::type _Nested; + _Nested n(derived()); + RealScalar w = n.cwiseAbs().maxCoeff(); + RealScalar z = (n/w).squaredNorm(); + if(z>RealScalar(0)) + return n / (numext::sqrt(z)*w); + else + return n; +} + +/** Normalizes the vector while avoid underflow and overflow + * + * \only_for_vectors + * + * This method is analogue to the normalize() method, but it reduces the risk of + * underflow and overflow when computing the norm. + * + * \warning If the input vector is too small (i.e., this->norm()==0), then \c *this is left unchanged. + * + * \sa stableNorm(), stableNormalized(), normalize() + */ +template<typename Derived> +inline void MatrixBase<Derived>::stableNormalize() +{ + RealScalar w = cwiseAbs().maxCoeff(); + RealScalar z = (derived()/w).squaredNorm(); + if(z>RealScalar(0)) + derived() /= numext::sqrt(z)*w; +} + +//---------- implementation of other norms ---------- + +namespace internal { + +template<typename Derived, int p> +struct lpNorm_selector +{ + typedef typename NumTraits<typename traits<Derived>::Scalar>::Real RealScalar; + EIGEN_DEVICE_FUNC + static inline RealScalar run(const MatrixBase<Derived>& m) + { + EIGEN_USING_STD_MATH(pow) + return pow(m.cwiseAbs().array().pow(p).sum(), RealScalar(1)/p); + } +}; + +template<typename Derived> +struct lpNorm_selector<Derived, 1> +{ + EIGEN_DEVICE_FUNC + static inline typename NumTraits<typename traits<Derived>::Scalar>::Real run(const MatrixBase<Derived>& m) + { + return m.cwiseAbs().sum(); + } +}; + +template<typename Derived> +struct lpNorm_selector<Derived, 2> +{ + EIGEN_DEVICE_FUNC + static inline typename NumTraits<typename traits<Derived>::Scalar>::Real run(const MatrixBase<Derived>& m) + { + return m.norm(); + } +}; + +template<typename Derived> +struct lpNorm_selector<Derived, Infinity> +{ + typedef typename NumTraits<typename traits<Derived>::Scalar>::Real RealScalar; + EIGEN_DEVICE_FUNC + static inline RealScalar run(const MatrixBase<Derived>& m) + { + if(Derived::SizeAtCompileTime==0 || (Derived::SizeAtCompileTime==Dynamic && m.size()==0)) + return RealScalar(0); + return m.cwiseAbs().maxCoeff(); + } +}; + +} // end namespace internal + +/** \returns the \b coefficient-wise \f$ \ell^p \f$ norm of \c *this, that is, returns the p-th root of the sum of the p-th powers of the absolute values + * of the coefficients of \c *this. If \a p is the special value \a Eigen::Infinity, this function returns the \f$ \ell^\infty \f$ + * norm, that is the maximum of the absolute values of the coefficients of \c *this. + * + * In all cases, if \c *this is empty, then the value 0 is returned. + * + * \note For matrices, this function does not compute the <a href="https://en.wikipedia.org/wiki/Operator_norm">operator-norm</a>. That is, if \c *this is a matrix, then its coefficients are interpreted as a 1D vector. Nonetheless, you can easily compute the 1-norm and \f$\infty\f$-norm matrix operator norms using \link TutorialReductionsVisitorsBroadcastingReductionsNorm partial reductions \endlink. + * + * \sa norm() + */ +template<typename Derived> +template<int p> +#ifndef EIGEN_PARSED_BY_DOXYGEN +inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real +#else +MatrixBase<Derived>::RealScalar +#endif +MatrixBase<Derived>::lpNorm() const +{ + return internal::lpNorm_selector<Derived, p>::run(*this); +} + +//---------- implementation of isOrthogonal / isUnitary ---------- + +/** \returns true if *this is approximately orthogonal to \a other, + * within the precision given by \a prec. + * + * Example: \include MatrixBase_isOrthogonal.cpp + * Output: \verbinclude MatrixBase_isOrthogonal.out + */ +template<typename Derived> +template<typename OtherDerived> +bool MatrixBase<Derived>::isOrthogonal +(const MatrixBase<OtherDerived>& other, const RealScalar& prec) const +{ + typename internal::nested_eval<Derived,2>::type nested(derived()); + typename internal::nested_eval<OtherDerived,2>::type otherNested(other.derived()); + return numext::abs2(nested.dot(otherNested)) <= prec * prec * nested.squaredNorm() * otherNested.squaredNorm(); +} + +/** \returns true if *this is approximately an unitary matrix, + * within the precision given by \a prec. In the case where the \a Scalar + * type is real numbers, a unitary matrix is an orthogonal matrix, whence the name. + * + * \note This can be used to check whether a family of vectors forms an orthonormal basis. + * Indeed, \c m.isUnitary() returns true if and only if the columns (equivalently, the rows) of m form an + * orthonormal basis. + * + * Example: \include MatrixBase_isUnitary.cpp + * Output: \verbinclude MatrixBase_isUnitary.out + */ +template<typename Derived> +bool MatrixBase<Derived>::isUnitary(const RealScalar& prec) const +{ + typename internal::nested_eval<Derived,1>::type self(derived()); + for(Index i = 0; i < cols(); ++i) + { + if(!internal::isApprox(self.col(i).squaredNorm(), static_cast<RealScalar>(1), prec)) + return false; + for(Index j = 0; j < i; ++j) + if(!internal::isMuchSmallerThan(self.col(i).dot(self.col(j)), static_cast<Scalar>(1), prec)) + return false; + } + return true; +} + +} // end namespace Eigen + +#endif // EIGEN_DOT_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/EigenBase.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/EigenBase.h new file mode 100644 index 000000000..b195506a9 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/EigenBase.h @@ -0,0 +1,159 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2009 Benoit Jacob <jacob.benoit.1@gmail.com> +// Copyright (C) 2009 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_EIGENBASE_H +#define EIGEN_EIGENBASE_H + +namespace Eigen { + +/** \class EigenBase + * \ingroup Core_Module + * + * Common base class for all classes T such that MatrixBase has an operator=(T) and a constructor MatrixBase(T). + * + * In other words, an EigenBase object is an object that can be copied into a MatrixBase. + * + * Besides MatrixBase-derived classes, this also includes special matrix classes such as diagonal matrices, etc. + * + * Notice that this class is trivial, it is only used to disambiguate overloaded functions. + * + * \sa \blank \ref TopicClassHierarchy + */ +template<typename Derived> struct EigenBase +{ +// typedef typename internal::plain_matrix_type<Derived>::type PlainObject; + + /** \brief The interface type of indices + * \details To change this, \c \#define the preprocessor symbol \c EIGEN_DEFAULT_DENSE_INDEX_TYPE. + * \deprecated Since Eigen 3.3, its usage is deprecated. Use Eigen::Index instead. + * \sa StorageIndex, \ref TopicPreprocessorDirectives. + */ + typedef Eigen::Index Index; + + // FIXME is it needed? + typedef typename internal::traits<Derived>::StorageKind StorageKind; + + /** \returns a reference to the derived object */ + EIGEN_DEVICE_FUNC + Derived& derived() { return *static_cast<Derived*>(this); } + /** \returns a const reference to the derived object */ + EIGEN_DEVICE_FUNC + const Derived& derived() const { return *static_cast<const Derived*>(this); } + + EIGEN_DEVICE_FUNC + inline Derived& const_cast_derived() const + { return *static_cast<Derived*>(const_cast<EigenBase*>(this)); } + EIGEN_DEVICE_FUNC + inline const Derived& const_derived() const + { return *static_cast<const Derived*>(this); } + + /** \returns the number of rows. \sa cols(), RowsAtCompileTime */ + EIGEN_DEVICE_FUNC + inline Index rows() const { return derived().rows(); } + /** \returns the number of columns. \sa rows(), ColsAtCompileTime*/ + EIGEN_DEVICE_FUNC + inline Index cols() const { return derived().cols(); } + /** \returns the number of coefficients, which is rows()*cols(). + * \sa rows(), cols(), SizeAtCompileTime. */ + EIGEN_DEVICE_FUNC + inline Index size() const { return rows() * cols(); } + + /** \internal Don't use it, but do the equivalent: \code dst = *this; \endcode */ + template<typename Dest> + EIGEN_DEVICE_FUNC + inline void evalTo(Dest& dst) const + { derived().evalTo(dst); } + + /** \internal Don't use it, but do the equivalent: \code dst += *this; \endcode */ + template<typename Dest> + EIGEN_DEVICE_FUNC + inline void addTo(Dest& dst) const + { + // This is the default implementation, + // derived class can reimplement it in a more optimized way. + typename Dest::PlainObject res(rows(),cols()); + evalTo(res); + dst += res; + } + + /** \internal Don't use it, but do the equivalent: \code dst -= *this; \endcode */ + template<typename Dest> + EIGEN_DEVICE_FUNC + inline void subTo(Dest& dst) const + { + // This is the default implementation, + // derived class can reimplement it in a more optimized way. + typename Dest::PlainObject res(rows(),cols()); + evalTo(res); + dst -= res; + } + + /** \internal Don't use it, but do the equivalent: \code dst.applyOnTheRight(*this); \endcode */ + template<typename Dest> + EIGEN_DEVICE_FUNC inline void applyThisOnTheRight(Dest& dst) const + { + // This is the default implementation, + // derived class can reimplement it in a more optimized way. + dst = dst * this->derived(); + } + + /** \internal Don't use it, but do the equivalent: \code dst.applyOnTheLeft(*this); \endcode */ + template<typename Dest> + EIGEN_DEVICE_FUNC inline void applyThisOnTheLeft(Dest& dst) const + { + // This is the default implementation, + // derived class can reimplement it in a more optimized way. + dst = this->derived() * dst; + } + +}; + +/*************************************************************************** +* Implementation of matrix base methods +***************************************************************************/ + +/** \brief Copies the generic expression \a other into *this. + * + * \details The expression must provide a (templated) evalTo(Derived& dst) const + * function which does the actual job. In practice, this allows any user to write + * its own special matrix without having to modify MatrixBase + * + * \returns a reference to *this. + */ +template<typename Derived> +template<typename OtherDerived> +EIGEN_DEVICE_FUNC +Derived& DenseBase<Derived>::operator=(const EigenBase<OtherDerived> &other) +{ + call_assignment(derived(), other.derived()); + return derived(); +} + +template<typename Derived> +template<typename OtherDerived> +EIGEN_DEVICE_FUNC +Derived& DenseBase<Derived>::operator+=(const EigenBase<OtherDerived> &other) +{ + call_assignment(derived(), other.derived(), internal::add_assign_op<Scalar,typename OtherDerived::Scalar>()); + return derived(); +} + +template<typename Derived> +template<typename OtherDerived> +EIGEN_DEVICE_FUNC +Derived& DenseBase<Derived>::operator-=(const EigenBase<OtherDerived> &other) +{ + call_assignment(derived(), other.derived(), internal::sub_assign_op<Scalar,typename OtherDerived::Scalar>()); + return derived(); +} + +} // end namespace Eigen + +#endif // EIGEN_EIGENBASE_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/ForceAlignedAccess.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/ForceAlignedAccess.h new file mode 100644 index 000000000..7b08b45e6 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/ForceAlignedAccess.h @@ -0,0 +1,146 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2009-2010 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_FORCEALIGNEDACCESS_H +#define EIGEN_FORCEALIGNEDACCESS_H + +namespace Eigen { + +/** \class ForceAlignedAccess + * \ingroup Core_Module + * + * \brief Enforce aligned packet loads and stores regardless of what is requested + * + * \param ExpressionType the type of the object of which we are forcing aligned packet access + * + * This class is the return type of MatrixBase::forceAlignedAccess() + * and most of the time this is the only way it is used. + * + * \sa MatrixBase::forceAlignedAccess() + */ + +namespace internal { +template<typename ExpressionType> +struct traits<ForceAlignedAccess<ExpressionType> > : public traits<ExpressionType> +{}; +} + +template<typename ExpressionType> class ForceAlignedAccess + : public internal::dense_xpr_base< ForceAlignedAccess<ExpressionType> >::type +{ + public: + + typedef typename internal::dense_xpr_base<ForceAlignedAccess>::type Base; + EIGEN_DENSE_PUBLIC_INTERFACE(ForceAlignedAccess) + + EIGEN_DEVICE_FUNC explicit inline ForceAlignedAccess(const ExpressionType& matrix) : m_expression(matrix) {} + + 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> + inline const PacketScalar packet(Index row, Index col) const + { + return m_expression.template packet<Aligned>(row, col); + } + + template<int LoadMode> + inline void writePacket(Index row, Index col, const PacketScalar& x) + { + m_expression.const_cast_derived().template writePacket<Aligned>(row, col, x); + } + + template<int LoadMode> + inline const PacketScalar packet(Index index) const + { + return m_expression.template packet<Aligned>(index); + } + + template<int LoadMode> + inline void writePacket(Index index, const PacketScalar& x) + { + m_expression.const_cast_derived().template writePacket<Aligned>(index, x); + } + + EIGEN_DEVICE_FUNC operator const ExpressionType&() const { return m_expression; } + + protected: + const ExpressionType& m_expression; + + private: + ForceAlignedAccess& operator=(const ForceAlignedAccess&); +}; + +/** \returns an expression of *this with forced aligned access + * \sa forceAlignedAccessIf(),class ForceAlignedAccess + */ +template<typename Derived> +inline const ForceAlignedAccess<Derived> +MatrixBase<Derived>::forceAlignedAccess() const +{ + return ForceAlignedAccess<Derived>(derived()); +} + +/** \returns an expression of *this with forced aligned access + * \sa forceAlignedAccessIf(), class ForceAlignedAccess + */ +template<typename Derived> +inline ForceAlignedAccess<Derived> +MatrixBase<Derived>::forceAlignedAccess() +{ + return ForceAlignedAccess<Derived>(derived()); +} + +/** \returns an expression of *this with forced aligned access if \a Enable is true. + * \sa forceAlignedAccess(), class ForceAlignedAccess + */ +template<typename Derived> +template<bool Enable> +inline typename internal::add_const_on_value_type<typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type>::type +MatrixBase<Derived>::forceAlignedAccessIf() const +{ + return derived(); // FIXME This should not work but apparently is never used +} + +/** \returns an expression of *this with forced aligned access if \a Enable is true. + * \sa forceAlignedAccess(), class ForceAlignedAccess + */ +template<typename Derived> +template<bool Enable> +inline typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type +MatrixBase<Derived>::forceAlignedAccessIf() +{ + return derived(); // FIXME This should not work but apparently is never used +} + +} // end namespace Eigen + +#endif // EIGEN_FORCEALIGNEDACCESS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/Fuzzy.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Fuzzy.h new file mode 100644 index 000000000..3e403a09d --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Fuzzy.h @@ -0,0 +1,155 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com> +// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr> +// +// 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_FUZZY_H +#define EIGEN_FUZZY_H + +namespace Eigen { + +namespace internal +{ + +template<typename Derived, typename OtherDerived, bool is_integer = NumTraits<typename Derived::Scalar>::IsInteger> +struct isApprox_selector +{ + EIGEN_DEVICE_FUNC + static bool run(const Derived& x, const OtherDerived& y, const typename Derived::RealScalar& prec) + { + typename internal::nested_eval<Derived,2>::type nested(x); + typename internal::nested_eval<OtherDerived,2>::type otherNested(y); + return (nested - otherNested).cwiseAbs2().sum() <= prec * prec * numext::mini(nested.cwiseAbs2().sum(), otherNested.cwiseAbs2().sum()); + } +}; + +template<typename Derived, typename OtherDerived> +struct isApprox_selector<Derived, OtherDerived, true> +{ + EIGEN_DEVICE_FUNC + static bool run(const Derived& x, const OtherDerived& y, const typename Derived::RealScalar&) + { + return x.matrix() == y.matrix(); + } +}; + +template<typename Derived, typename OtherDerived, bool is_integer = NumTraits<typename Derived::Scalar>::IsInteger> +struct isMuchSmallerThan_object_selector +{ + EIGEN_DEVICE_FUNC + static bool run(const Derived& x, const OtherDerived& y, const typename Derived::RealScalar& prec) + { + return x.cwiseAbs2().sum() <= numext::abs2(prec) * y.cwiseAbs2().sum(); + } +}; + +template<typename Derived, typename OtherDerived> +struct isMuchSmallerThan_object_selector<Derived, OtherDerived, true> +{ + EIGEN_DEVICE_FUNC + static bool run(const Derived& x, const OtherDerived&, const typename Derived::RealScalar&) + { + return x.matrix() == Derived::Zero(x.rows(), x.cols()).matrix(); + } +}; + +template<typename Derived, bool is_integer = NumTraits<typename Derived::Scalar>::IsInteger> +struct isMuchSmallerThan_scalar_selector +{ + EIGEN_DEVICE_FUNC + static bool run(const Derived& x, const typename Derived::RealScalar& y, const typename Derived::RealScalar& prec) + { + return x.cwiseAbs2().sum() <= numext::abs2(prec * y); + } +}; + +template<typename Derived> +struct isMuchSmallerThan_scalar_selector<Derived, true> +{ + EIGEN_DEVICE_FUNC + static bool run(const Derived& x, const typename Derived::RealScalar&, const typename Derived::RealScalar&) + { + return x.matrix() == Derived::Zero(x.rows(), x.cols()).matrix(); + } +}; + +} // end namespace internal + + +/** \returns \c true if \c *this is approximately equal to \a other, within the precision + * determined by \a prec. + * + * \note The fuzzy compares are done multiplicatively. Two vectors \f$ v \f$ and \f$ w \f$ + * are considered to be approximately equal within precision \f$ p \f$ if + * \f[ \Vert v - w \Vert \leqslant p\,\min(\Vert v\Vert, \Vert w\Vert). \f] + * For matrices, the comparison is done using the Hilbert-Schmidt norm (aka Frobenius norm + * L2 norm). + * + * \note Because of the multiplicativeness of this comparison, one can't use this function + * to check whether \c *this is approximately equal to the zero matrix or vector. + * Indeed, \c isApprox(zero) returns false unless \c *this itself is exactly the zero matrix + * or vector. If you want to test whether \c *this is zero, use internal::isMuchSmallerThan(const + * RealScalar&, RealScalar) instead. + * + * \sa internal::isMuchSmallerThan(const RealScalar&, RealScalar) const + */ +template<typename Derived> +template<typename OtherDerived> +bool DenseBase<Derived>::isApprox( + const DenseBase<OtherDerived>& other, + const RealScalar& prec +) const +{ + return internal::isApprox_selector<Derived, OtherDerived>::run(derived(), other.derived(), prec); +} + +/** \returns \c true if the norm of \c *this is much smaller than \a other, + * within the precision determined by \a prec. + * + * \note The fuzzy compares are done multiplicatively. A vector \f$ v \f$ is + * considered to be much smaller than \f$ x \f$ within precision \f$ p \f$ if + * \f[ \Vert v \Vert \leqslant p\,\vert x\vert. \f] + * + * For matrices, the comparison is done using the Hilbert-Schmidt norm. For this reason, + * the value of the reference scalar \a other should come from the Hilbert-Schmidt norm + * of a reference matrix of same dimensions. + * + * \sa isApprox(), isMuchSmallerThan(const DenseBase<OtherDerived>&, RealScalar) const + */ +template<typename Derived> +bool DenseBase<Derived>::isMuchSmallerThan( + const typename NumTraits<Scalar>::Real& other, + const RealScalar& prec +) const +{ + return internal::isMuchSmallerThan_scalar_selector<Derived>::run(derived(), other, prec); +} + +/** \returns \c true if the norm of \c *this is much smaller than the norm of \a other, + * within the precision determined by \a prec. + * + * \note The fuzzy compares are done multiplicatively. A vector \f$ v \f$ is + * considered to be much smaller than a vector \f$ w \f$ within precision \f$ p \f$ if + * \f[ \Vert v \Vert \leqslant p\,\Vert w\Vert. \f] + * For matrices, the comparison is done using the Hilbert-Schmidt norm. + * + * \sa isApprox(), isMuchSmallerThan(const RealScalar&, RealScalar) const + */ +template<typename Derived> +template<typename OtherDerived> +bool DenseBase<Derived>::isMuchSmallerThan( + const DenseBase<OtherDerived>& other, + const RealScalar& prec +) const +{ + return internal::isMuchSmallerThan_object_selector<Derived, OtherDerived>::run(derived(), other.derived(), prec); +} + +} // end namespace Eigen + +#endif // EIGEN_FUZZY_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/GeneralProduct.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/GeneralProduct.h new file mode 100644 index 000000000..0f16cd8e3 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/GeneralProduct.h @@ -0,0 +1,454 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com> +// Copyright (C) 2008-2011 Gael Guennebaud <gael.guennebaud@inria.fr> +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_GENERAL_PRODUCT_H +#define EIGEN_GENERAL_PRODUCT_H + +namespace Eigen { + +enum { + Large = 2, + Small = 3 +}; + +namespace internal { + +template<int Rows, int Cols, int Depth> struct product_type_selector; + +template<int Size, int MaxSize> struct product_size_category +{ + enum { is_large = MaxSize == Dynamic || + Size >= EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD || + (Size==Dynamic && MaxSize>=EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD), + value = is_large ? Large + : Size == 1 ? 1 + : Small + }; +}; + +template<typename Lhs, typename Rhs> struct product_type +{ + typedef typename remove_all<Lhs>::type _Lhs; + typedef typename remove_all<Rhs>::type _Rhs; + enum { + MaxRows = traits<_Lhs>::MaxRowsAtCompileTime, + Rows = traits<_Lhs>::RowsAtCompileTime, + MaxCols = traits<_Rhs>::MaxColsAtCompileTime, + Cols = traits<_Rhs>::ColsAtCompileTime, + MaxDepth = EIGEN_SIZE_MIN_PREFER_FIXED(traits<_Lhs>::MaxColsAtCompileTime, + traits<_Rhs>::MaxRowsAtCompileTime), + Depth = EIGEN_SIZE_MIN_PREFER_FIXED(traits<_Lhs>::ColsAtCompileTime, + traits<_Rhs>::RowsAtCompileTime) + }; + + // the splitting into different lines of code here, introducing the _select enums and the typedef below, + // is to work around an internal compiler error with gcc 4.1 and 4.2. +private: + enum { + rows_select = product_size_category<Rows,MaxRows>::value, + cols_select = product_size_category<Cols,MaxCols>::value, + depth_select = product_size_category<Depth,MaxDepth>::value + }; + typedef product_type_selector<rows_select, cols_select, depth_select> selector; + +public: + enum { + value = selector::ret, + ret = selector::ret + }; +#ifdef EIGEN_DEBUG_PRODUCT + static void debug() + { + EIGEN_DEBUG_VAR(Rows); + EIGEN_DEBUG_VAR(Cols); + EIGEN_DEBUG_VAR(Depth); + EIGEN_DEBUG_VAR(rows_select); + EIGEN_DEBUG_VAR(cols_select); + EIGEN_DEBUG_VAR(depth_select); + EIGEN_DEBUG_VAR(value); + } +#endif +}; + +/* The following allows to select the kind of product at compile time + * based on the three dimensions of the product. + * This is a compile time mapping from {1,Small,Large}^3 -> {product types} */ +// FIXME I'm not sure the current mapping is the ideal one. +template<int M, int N> struct product_type_selector<M,N,1> { enum { ret = OuterProduct }; }; +template<int M> struct product_type_selector<M, 1, 1> { enum { ret = LazyCoeffBasedProductMode }; }; +template<int N> struct product_type_selector<1, N, 1> { enum { ret = LazyCoeffBasedProductMode }; }; +template<int Depth> struct product_type_selector<1, 1, Depth> { enum { ret = InnerProduct }; }; +template<> struct product_type_selector<1, 1, 1> { enum { ret = InnerProduct }; }; +template<> struct product_type_selector<Small,1, Small> { enum { ret = CoeffBasedProductMode }; }; +template<> struct product_type_selector<1, Small,Small> { enum { ret = CoeffBasedProductMode }; }; +template<> struct product_type_selector<Small,Small,Small> { enum { ret = CoeffBasedProductMode }; }; +template<> struct product_type_selector<Small, Small, 1> { enum { ret = LazyCoeffBasedProductMode }; }; +template<> struct product_type_selector<Small, Large, 1> { enum { ret = LazyCoeffBasedProductMode }; }; +template<> struct product_type_selector<Large, Small, 1> { enum { ret = LazyCoeffBasedProductMode }; }; +template<> struct product_type_selector<1, Large,Small> { enum { ret = CoeffBasedProductMode }; }; +template<> struct product_type_selector<1, Large,Large> { enum { ret = GemvProduct }; }; +template<> struct product_type_selector<1, Small,Large> { enum { ret = CoeffBasedProductMode }; }; +template<> struct product_type_selector<Large,1, Small> { enum { ret = CoeffBasedProductMode }; }; +template<> struct product_type_selector<Large,1, Large> { enum { ret = GemvProduct }; }; +template<> struct product_type_selector<Small,1, Large> { enum { ret = CoeffBasedProductMode }; }; +template<> struct product_type_selector<Small,Small,Large> { enum { ret = GemmProduct }; }; +template<> struct product_type_selector<Large,Small,Large> { enum { ret = GemmProduct }; }; +template<> struct product_type_selector<Small,Large,Large> { enum { ret = GemmProduct }; }; +template<> struct product_type_selector<Large,Large,Large> { enum { ret = GemmProduct }; }; +template<> struct product_type_selector<Large,Small,Small> { enum { ret = CoeffBasedProductMode }; }; +template<> struct product_type_selector<Small,Large,Small> { enum { ret = CoeffBasedProductMode }; }; +template<> struct product_type_selector<Large,Large,Small> { enum { ret = GemmProduct }; }; + +} // end namespace internal + +/*********************************************************************** +* Implementation of Inner Vector Vector Product +***********************************************************************/ + +// FIXME : maybe the "inner product" could return a Scalar +// instead of a 1x1 matrix ?? +// Pro: more natural for the user +// Cons: this could be a problem if in a meta unrolled algorithm a matrix-matrix +// product ends up to a row-vector times col-vector product... To tackle this use +// case, we could have a specialization for Block<MatrixType,1,1> with: operator=(Scalar x); + +/*********************************************************************** +* Implementation of Outer Vector Vector Product +***********************************************************************/ + +/*********************************************************************** +* Implementation of General Matrix Vector Product +***********************************************************************/ + +/* According to the shape/flags of the matrix we have to distinghish 3 different cases: + * 1 - the matrix is col-major, BLAS compatible and M is large => call fast BLAS-like colmajor routine + * 2 - the matrix is row-major, BLAS compatible and N is large => call fast BLAS-like rowmajor routine + * 3 - all other cases are handled using a simple loop along the outer-storage direction. + * Therefore we need a lower level meta selector. + * Furthermore, if the matrix is the rhs, then the product has to be transposed. + */ +namespace internal { + +template<int Side, int StorageOrder, bool BlasCompatible> +struct gemv_dense_selector; + +} // end namespace internal + +namespace internal { + +template<typename Scalar,int Size,int MaxSize,bool Cond> struct gemv_static_vector_if; + +template<typename Scalar,int Size,int MaxSize> +struct gemv_static_vector_if<Scalar,Size,MaxSize,false> +{ + EIGEN_STRONG_INLINE Scalar* data() { eigen_internal_assert(false && "should never be called"); return 0; } +}; + +template<typename Scalar,int Size> +struct gemv_static_vector_if<Scalar,Size,Dynamic,true> +{ + EIGEN_STRONG_INLINE Scalar* data() { return 0; } +}; + +template<typename Scalar,int Size,int MaxSize> +struct gemv_static_vector_if<Scalar,Size,MaxSize,true> +{ + enum { + ForceAlignment = internal::packet_traits<Scalar>::Vectorizable, + PacketSize = internal::packet_traits<Scalar>::size + }; + #if EIGEN_MAX_STATIC_ALIGN_BYTES!=0 + internal::plain_array<Scalar,EIGEN_SIZE_MIN_PREFER_FIXED(Size,MaxSize),0,EIGEN_PLAIN_ENUM_MIN(AlignedMax,PacketSize)> m_data; + EIGEN_STRONG_INLINE Scalar* data() { return m_data.array; } + #else + // Some architectures cannot align on the stack, + // => let's manually enforce alignment by allocating more data and return the address of the first aligned element. + internal::plain_array<Scalar,EIGEN_SIZE_MIN_PREFER_FIXED(Size,MaxSize)+(ForceAlignment?EIGEN_MAX_ALIGN_BYTES:0),0> m_data; + EIGEN_STRONG_INLINE Scalar* data() { + return ForceAlignment + ? reinterpret_cast<Scalar*>((internal::UIntPtr(m_data.array) & ~(std::size_t(EIGEN_MAX_ALIGN_BYTES-1))) + EIGEN_MAX_ALIGN_BYTES) + : m_data.array; + } + #endif +}; + +// The vector is on the left => transposition +template<int StorageOrder, bool BlasCompatible> +struct gemv_dense_selector<OnTheLeft,StorageOrder,BlasCompatible> +{ + template<typename Lhs, typename Rhs, typename Dest> + static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha) + { + Transpose<Dest> destT(dest); + enum { OtherStorageOrder = StorageOrder == RowMajor ? ColMajor : RowMajor }; + gemv_dense_selector<OnTheRight,OtherStorageOrder,BlasCompatible> + ::run(rhs.transpose(), lhs.transpose(), destT, alpha); + } +}; + +template<> struct gemv_dense_selector<OnTheRight,ColMajor,true> +{ + template<typename Lhs, typename Rhs, typename Dest> + static inline void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha) + { + typedef typename Lhs::Scalar LhsScalar; + typedef typename Rhs::Scalar RhsScalar; + typedef typename Dest::Scalar ResScalar; + typedef typename Dest::RealScalar RealScalar; + + typedef internal::blas_traits<Lhs> LhsBlasTraits; + typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType; + typedef internal::blas_traits<Rhs> RhsBlasTraits; + typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType; + + typedef Map<Matrix<ResScalar,Dynamic,1>, EIGEN_PLAIN_ENUM_MIN(AlignedMax,internal::packet_traits<ResScalar>::size)> MappedDest; + + ActualLhsType actualLhs = LhsBlasTraits::extract(lhs); + ActualRhsType actualRhs = RhsBlasTraits::extract(rhs); + + ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(lhs) + * RhsBlasTraits::extractScalarFactor(rhs); + + // make sure Dest is a compile-time vector type (bug 1166) + typedef typename conditional<Dest::IsVectorAtCompileTime, Dest, typename Dest::ColXpr>::type ActualDest; + + enum { + // FIXME find a way to allow an inner stride on the result if packet_traits<Scalar>::size==1 + // on, the other hand it is good for the cache to pack the vector anyways... + EvalToDestAtCompileTime = (ActualDest::InnerStrideAtCompileTime==1), + ComplexByReal = (NumTraits<LhsScalar>::IsComplex) && (!NumTraits<RhsScalar>::IsComplex), + MightCannotUseDest = (!EvalToDestAtCompileTime) || ComplexByReal + }; + + typedef const_blas_data_mapper<LhsScalar,Index,ColMajor> LhsMapper; + typedef const_blas_data_mapper<RhsScalar,Index,RowMajor> RhsMapper; + RhsScalar compatibleAlpha = get_factor<ResScalar,RhsScalar>::run(actualAlpha); + + if(!MightCannotUseDest) + { + // shortcut if we are sure to be able to use dest directly, + // this ease the compiler to generate cleaner and more optimzized code for most common cases + general_matrix_vector_product + <Index,LhsScalar,LhsMapper,ColMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsMapper,RhsBlasTraits::NeedToConjugate>::run( + actualLhs.rows(), actualLhs.cols(), + LhsMapper(actualLhs.data(), actualLhs.outerStride()), + RhsMapper(actualRhs.data(), actualRhs.innerStride()), + dest.data(), 1, + compatibleAlpha); + } + else + { + gemv_static_vector_if<ResScalar,ActualDest::SizeAtCompileTime,ActualDest::MaxSizeAtCompileTime,MightCannotUseDest> static_dest; + + const bool alphaIsCompatible = (!ComplexByReal) || (numext::imag(actualAlpha)==RealScalar(0)); + const bool evalToDest = EvalToDestAtCompileTime && alphaIsCompatible; + + ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(), + evalToDest ? dest.data() : static_dest.data()); + + if(!evalToDest) + { + #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN + Index size = dest.size(); + EIGEN_DENSE_STORAGE_CTOR_PLUGIN + #endif + if(!alphaIsCompatible) + { + MappedDest(actualDestPtr, dest.size()).setZero(); + compatibleAlpha = RhsScalar(1); + } + else + MappedDest(actualDestPtr, dest.size()) = dest; + } + + general_matrix_vector_product + <Index,LhsScalar,LhsMapper,ColMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsMapper,RhsBlasTraits::NeedToConjugate>::run( + actualLhs.rows(), actualLhs.cols(), + LhsMapper(actualLhs.data(), actualLhs.outerStride()), + RhsMapper(actualRhs.data(), actualRhs.innerStride()), + actualDestPtr, 1, + compatibleAlpha); + + if (!evalToDest) + { + if(!alphaIsCompatible) + dest.matrix() += actualAlpha * MappedDest(actualDestPtr, dest.size()); + else + dest = MappedDest(actualDestPtr, dest.size()); + } + } + } +}; + +template<> struct gemv_dense_selector<OnTheRight,RowMajor,true> +{ + template<typename Lhs, typename Rhs, typename Dest> + static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha) + { + typedef typename Lhs::Scalar LhsScalar; + typedef typename Rhs::Scalar RhsScalar; + typedef typename Dest::Scalar ResScalar; + + typedef internal::blas_traits<Lhs> LhsBlasTraits; + typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType; + typedef internal::blas_traits<Rhs> RhsBlasTraits; + typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType; + typedef typename internal::remove_all<ActualRhsType>::type ActualRhsTypeCleaned; + + typename add_const<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(lhs); + typename add_const<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(rhs); + + ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(lhs) + * RhsBlasTraits::extractScalarFactor(rhs); + + enum { + // FIXME find a way to allow an inner stride on the result if packet_traits<Scalar>::size==1 + // on, the other hand it is good for the cache to pack the vector anyways... + DirectlyUseRhs = ActualRhsTypeCleaned::InnerStrideAtCompileTime==1 + }; + + gemv_static_vector_if<RhsScalar,ActualRhsTypeCleaned::SizeAtCompileTime,ActualRhsTypeCleaned::MaxSizeAtCompileTime,!DirectlyUseRhs> static_rhs; + + ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhsPtr,actualRhs.size(), + DirectlyUseRhs ? const_cast<RhsScalar*>(actualRhs.data()) : static_rhs.data()); + + if(!DirectlyUseRhs) + { + #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN + Index size = actualRhs.size(); + EIGEN_DENSE_STORAGE_CTOR_PLUGIN + #endif + Map<typename ActualRhsTypeCleaned::PlainObject>(actualRhsPtr, actualRhs.size()) = actualRhs; + } + + typedef const_blas_data_mapper<LhsScalar,Index,RowMajor> LhsMapper; + typedef const_blas_data_mapper<RhsScalar,Index,ColMajor> RhsMapper; + general_matrix_vector_product + <Index,LhsScalar,LhsMapper,RowMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsMapper,RhsBlasTraits::NeedToConjugate>::run( + actualLhs.rows(), actualLhs.cols(), + LhsMapper(actualLhs.data(), actualLhs.outerStride()), + RhsMapper(actualRhsPtr, 1), + dest.data(), dest.col(0).innerStride(), //NOTE if dest is not a vector at compile-time, then dest.innerStride() might be wrong. (bug 1166) + actualAlpha); + } +}; + +template<> struct gemv_dense_selector<OnTheRight,ColMajor,false> +{ + template<typename Lhs, typename Rhs, typename Dest> + static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha) + { + EIGEN_STATIC_ASSERT((!nested_eval<Lhs,1>::Evaluate),EIGEN_INTERNAL_COMPILATION_ERROR_OR_YOU_MADE_A_PROGRAMMING_MISTAKE); + // TODO if rhs is large enough it might be beneficial to make sure that dest is sequentially stored in memory, otherwise use a temp + typename nested_eval<Rhs,1>::type actual_rhs(rhs); + const Index size = rhs.rows(); + for(Index k=0; k<size; ++k) + dest += (alpha*actual_rhs.coeff(k)) * lhs.col(k); + } +}; + +template<> struct gemv_dense_selector<OnTheRight,RowMajor,false> +{ + template<typename Lhs, typename Rhs, typename Dest> + static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha) + { + EIGEN_STATIC_ASSERT((!nested_eval<Lhs,1>::Evaluate),EIGEN_INTERNAL_COMPILATION_ERROR_OR_YOU_MADE_A_PROGRAMMING_MISTAKE); + typename nested_eval<Rhs,Lhs::RowsAtCompileTime>::type actual_rhs(rhs); + const Index rows = dest.rows(); + for(Index i=0; i<rows; ++i) + dest.coeffRef(i) += alpha * (lhs.row(i).cwiseProduct(actual_rhs.transpose())).sum(); + } +}; + +} // end namespace internal + +/*************************************************************************** +* Implementation of matrix base methods +***************************************************************************/ + +/** \returns the matrix product of \c *this and \a other. + * + * \note If instead of the matrix product you want the coefficient-wise product, see Cwise::operator*(). + * + * \sa lazyProduct(), operator*=(const MatrixBase&), Cwise::operator*() + */ +#ifndef __CUDACC__ + +template<typename Derived> +template<typename OtherDerived> +inline const Product<Derived, OtherDerived> +MatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const +{ + // A note regarding the function declaration: In MSVC, this function will sometimes + // not be inlined since DenseStorage is an unwindable object for dynamic + // matrices and product types are holding a member to store the result. + // Thus it does not help tagging this function with EIGEN_STRONG_INLINE. + enum { + ProductIsValid = Derived::ColsAtCompileTime==Dynamic + || OtherDerived::RowsAtCompileTime==Dynamic + || int(Derived::ColsAtCompileTime)==int(OtherDerived::RowsAtCompileTime), + AreVectors = Derived::IsVectorAtCompileTime && OtherDerived::IsVectorAtCompileTime, + SameSizes = EIGEN_PREDICATE_SAME_MATRIX_SIZE(Derived,OtherDerived) + }; + // note to the lost user: + // * for a dot product use: v1.dot(v2) + // * for a coeff-wise product use: v1.cwiseProduct(v2) + EIGEN_STATIC_ASSERT(ProductIsValid || !(AreVectors && SameSizes), + INVALID_VECTOR_VECTOR_PRODUCT__IF_YOU_WANTED_A_DOT_OR_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTIONS) + EIGEN_STATIC_ASSERT(ProductIsValid || !(SameSizes && !AreVectors), + INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION) + EIGEN_STATIC_ASSERT(ProductIsValid || SameSizes, INVALID_MATRIX_PRODUCT) +#ifdef EIGEN_DEBUG_PRODUCT + internal::product_type<Derived,OtherDerived>::debug(); +#endif + + return Product<Derived, OtherDerived>(derived(), other.derived()); +} + +#endif // __CUDACC__ + +/** \returns an expression of the matrix product of \c *this and \a other without implicit evaluation. + * + * The returned product will behave like any other expressions: the coefficients of the product will be + * computed once at a time as requested. This might be useful in some extremely rare cases when only + * a small and no coherent fraction of the result's coefficients have to be computed. + * + * \warning This version of the matrix product can be much much slower. So use it only if you know + * what you are doing and that you measured a true speed improvement. + * + * \sa operator*(const MatrixBase&) + */ +template<typename Derived> +template<typename OtherDerived> +const Product<Derived,OtherDerived,LazyProduct> +MatrixBase<Derived>::lazyProduct(const MatrixBase<OtherDerived> &other) const +{ + enum { + ProductIsValid = Derived::ColsAtCompileTime==Dynamic + || OtherDerived::RowsAtCompileTime==Dynamic + || int(Derived::ColsAtCompileTime)==int(OtherDerived::RowsAtCompileTime), + AreVectors = Derived::IsVectorAtCompileTime && OtherDerived::IsVectorAtCompileTime, + SameSizes = EIGEN_PREDICATE_SAME_MATRIX_SIZE(Derived,OtherDerived) + }; + // note to the lost user: + // * for a dot product use: v1.dot(v2) + // * for a coeff-wise product use: v1.cwiseProduct(v2) + EIGEN_STATIC_ASSERT(ProductIsValid || !(AreVectors && SameSizes), + INVALID_VECTOR_VECTOR_PRODUCT__IF_YOU_WANTED_A_DOT_OR_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTIONS) + EIGEN_STATIC_ASSERT(ProductIsValid || !(SameSizes && !AreVectors), + INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION) + EIGEN_STATIC_ASSERT(ProductIsValid || SameSizes, INVALID_MATRIX_PRODUCT) + + return Product<Derived,OtherDerived,LazyProduct>(derived(), other.derived()); +} + +} // end namespace Eigen + +#endif // EIGEN_PRODUCT_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/GenericPacketMath.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/GenericPacketMath.h new file mode 100644 index 000000000..029f8ac36 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/GenericPacketMath.h @@ -0,0 +1,593 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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_GENERIC_PACKET_MATH_H +#define EIGEN_GENERIC_PACKET_MATH_H + +namespace Eigen { + +namespace internal { + +/** \internal + * \file GenericPacketMath.h + * + * Default implementation for types not supported by the vectorization. + * In practice these functions are provided to make easier the writing + * of generic vectorized code. + */ + +#ifndef EIGEN_DEBUG_ALIGNED_LOAD +#define EIGEN_DEBUG_ALIGNED_LOAD +#endif + +#ifndef EIGEN_DEBUG_UNALIGNED_LOAD +#define EIGEN_DEBUG_UNALIGNED_LOAD +#endif + +#ifndef EIGEN_DEBUG_ALIGNED_STORE +#define EIGEN_DEBUG_ALIGNED_STORE +#endif + +#ifndef EIGEN_DEBUG_UNALIGNED_STORE +#define EIGEN_DEBUG_UNALIGNED_STORE +#endif + +struct default_packet_traits +{ + enum { + HasHalfPacket = 0, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasNegate = 1, + HasAbs = 1, + HasArg = 0, + HasAbs2 = 1, + HasMin = 1, + HasMax = 1, + HasConj = 1, + HasSetLinear = 1, + HasBlend = 0, + + HasDiv = 0, + HasSqrt = 0, + HasRsqrt = 0, + HasExp = 0, + HasLog = 0, + HasLog1p = 0, + HasLog10 = 0, + HasPow = 0, + + HasSin = 0, + HasCos = 0, + HasTan = 0, + HasASin = 0, + HasACos = 0, + HasATan = 0, + HasSinh = 0, + HasCosh = 0, + HasTanh = 0, + HasLGamma = 0, + HasDiGamma = 0, + HasZeta = 0, + HasPolygamma = 0, + HasErf = 0, + HasErfc = 0, + HasIGamma = 0, + HasIGammac = 0, + HasBetaInc = 0, + + HasRound = 0, + HasFloor = 0, + HasCeil = 0, + + HasSign = 0 + }; +}; + +template<typename T> struct packet_traits : default_packet_traits +{ + typedef T type; + typedef T half; + enum { + Vectorizable = 0, + size = 1, + AlignedOnScalar = 0, + HasHalfPacket = 0 + }; + enum { + HasAdd = 0, + HasSub = 0, + HasMul = 0, + HasNegate = 0, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasConj = 0, + HasSetLinear = 0 + }; +}; + +template<typename T> struct packet_traits<const T> : packet_traits<T> { }; + +template <typename Src, typename Tgt> struct type_casting_traits { + enum { + VectorizedCast = 0, + SrcCoeffRatio = 1, + TgtCoeffRatio = 1 + }; +}; + + +/** \internal \returns static_cast<TgtType>(a) (coeff-wise) */ +template <typename SrcPacket, typename TgtPacket> +EIGEN_DEVICE_FUNC inline TgtPacket +pcast(const SrcPacket& a) { + return static_cast<TgtPacket>(a); +} +template <typename SrcPacket, typename TgtPacket> +EIGEN_DEVICE_FUNC inline TgtPacket +pcast(const SrcPacket& a, const SrcPacket& /*b*/) { + return static_cast<TgtPacket>(a); +} + +template <typename SrcPacket, typename TgtPacket> +EIGEN_DEVICE_FUNC inline TgtPacket +pcast(const SrcPacket& a, const SrcPacket& /*b*/, const SrcPacket& /*c*/, const SrcPacket& /*d*/) { + return static_cast<TgtPacket>(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; } + +/** \internal \returns a - b (coeff-wise) */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +psub(const Packet& a, + const Packet& b) { return a-b; } + +/** \internal \returns -a (coeff-wise) */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +pnegate(const Packet& a) { return -a; } + +/** \internal \returns conj(a) (coeff-wise) */ + +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +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; } + +/** \internal \returns a / b (coeff-wise) */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +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); } + +/** \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); } + +/** \internal \returns the absolute value of \a a */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +pabs(const Packet& a) { using std::abs; return abs(a); } + +/** \internal \returns the phase angle of \a a */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +parg(const Packet& a) { using numext::arg; return arg(a); } + +/** \internal \returns the bitwise and of \a a and \a b */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +pand(const Packet& a, const Packet& b) { return a & b; } + +/** \internal \returns the bitwise or of \a a and \a b */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +por(const Packet& a, const Packet& b) { return a | b; } + +/** \internal \returns the bitwise xor of \a a and \a b */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +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); } + +/** \internal \returns a packet version of \a *from, from must be 16 bytes aligned */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +pload(const typename unpacket_traits<Packet>::type* from) { return *from; } + +/** \internal \returns a packet version of \a *from, (un-aligned load) */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +ploadu(const typename unpacket_traits<Packet>::type* from) { return *from; } + +/** \internal \returns a packet with constant coefficients \a a, e.g.: (a,a,a,a) */ +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 \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); } + +/** \internal \returns a packet with elements of \a *from duplicated. + * For instance, for a packet of 8 elements, 4 scalars will be read from \a *from and + * duplicated to form: {from[0],from[0],from[1],from[1],from[2],from[2],from[3],from[3]} + * Currently, this function is only used for scalar * complex products. + */ +template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet +ploaddup(const typename unpacket_traits<Packet>::type* from) { return *from; } + +/** \internal \returns a packet with elements of \a *from quadrupled. + * For instance, for a packet of 8 elements, 2 scalars will be read from \a *from and + * replicated to form: {from[0],from[0],from[0],from[0],from[1],from[1],from[1],from[1]} + * Currently, this function is only used in matrix products. + * For packet-size smaller or equal to 4, this function is equivalent to pload1 + */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +ploadquad(const typename unpacket_traits<Packet>::type* from) +{ return pload1<Packet>(from); } + +/** \internal equivalent to + * \code + * a0 = pload1(a+0); + * a1 = pload1(a+1); + * a2 = pload1(a+2); + * a3 = pload1(a+3); + * \endcode + * \sa pset1, pload1, ploaddup, pbroadcast2 + */ +template<typename Packet> EIGEN_DEVICE_FUNC +inline void pbroadcast4(const typename unpacket_traits<Packet>::type *a, + Packet& a0, Packet& a1, Packet& a2, Packet& a3) +{ + a0 = pload1<Packet>(a+0); + a1 = pload1<Packet>(a+1); + a2 = pload1<Packet>(a+2); + a3 = pload1<Packet>(a+3); +} + +/** \internal equivalent to + * \code + * a0 = pload1(a+0); + * a1 = pload1(a+1); + * \endcode + * \sa pset1, pload1, ploaddup, pbroadcast4 + */ +template<typename Packet> EIGEN_DEVICE_FUNC +inline void pbroadcast2(const typename unpacket_traits<Packet>::type *a, + Packet& a0, Packet& a1) +{ + a0 = pload1<Packet>(a+0); + a1 = pload1<Packet>(a+1); +} + +/** \internal \brief Returns a packet with coefficients (a,a+1,...,a+packet_size-1). */ +template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet +plset(const typename unpacket_traits<Packet>::type& a) { return a; } + +/** \internal copy the packet \a from to \a *to, \a to must be 16 bytes aligned */ +template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline void pstore(Scalar* to, const Packet& from) +{ (*to) = from; } + +/** \internal copy the packet \a from to \a *to, (un-aligned store) */ +template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline void pstoreu(Scalar* to, const Packet& from) +{ (*to) = from; } + + template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline Packet pgather(const Scalar* from, Index /*stride*/) + { return ploadu<Packet>(from); } + + template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline void pscatter(Scalar* to, const Packet& from, Index /*stride*/) + { pstore(to, from); } + +/** \internal tries to do cache prefetching of \a addr */ +template<typename Scalar> EIGEN_DEVICE_FUNC inline void prefetch(const Scalar* addr) +{ +#ifdef __CUDA_ARCH__ +#if defined(__LP64__) + // 64-bit pointer operand constraint for inlined asm + asm(" prefetch.L1 [ %1 ];" : "=l"(addr) : "l"(addr)); +#else + // 32-bit pointer operand constraint for inlined asm + asm(" prefetch.L1 [ %1 ];" : "=r"(addr) : "r"(addr)); +#endif +#elif (!EIGEN_COMP_MSVC) && (EIGEN_COMP_GNUC || EIGEN_COMP_CLANG || EIGEN_COMP_ICC) + __builtin_prefetch(addr); +#endif +} + +/** \internal \returns the first element of a packet */ +template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type pfirst(const Packet& a) +{ return a; } + +/** \internal \returns a packet where the element i contains the sum of the packet of \a vec[i] */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +preduxp(const Packet* vecs) { return vecs[0]; } + +/** \internal \returns the sum of the elements of \a a*/ +template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type predux(const Packet& a) +{ return a; } + +/** \internal \returns the sum of the elements of \a a by block of 4 elements. + * For a packet {a0, a1, a2, a3, a4, a5, a6, a7}, it returns a half packet {a0+a4, a1+a5, a2+a6, a3+a7} + * For packet-size smaller or equal to 4, this boils down to a noop. + */ +template<typename Packet> EIGEN_DEVICE_FUNC inline +typename conditional<(unpacket_traits<Packet>::size%8)==0,typename unpacket_traits<Packet>::half,Packet>::type +predux_downto4(const Packet& a) +{ return 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*/ +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*/ +template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type predux_max(const Packet& a) +{ return a; } + +/** \internal \returns the reversed elements of \a a*/ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet preverse(const Packet& a) +{ return a; } + +/** \internal \returns \a a with real and imaginary part flipped (for complex type only) */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet pcplxflip(const Packet& a) +{ + // FIXME: uncomment the following in case we drop the internal imag and real functions. +// using std::imag; +// using std::real; + return Packet(imag(a),real(a)); +} + +/************************** +* Special math functions +***************************/ + +/** \internal \returns the sine of \a a (coeff-wise) */ +template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +Packet psin(const Packet& a) { using std::sin; return sin(a); } + +/** \internal \returns the cosine of \a a (coeff-wise) */ +template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +Packet pcos(const Packet& a) { using std::cos; return cos(a); } + +/** \internal \returns the tan of \a a (coeff-wise) */ +template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +Packet ptan(const Packet& a) { using std::tan; return tan(a); } + +/** \internal \returns the arc sine of \a a (coeff-wise) */ +template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +Packet pasin(const Packet& a) { using std::asin; return asin(a); } + +/** \internal \returns the arc cosine of \a a (coeff-wise) */ +template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +Packet pacos(const Packet& a) { using std::acos; return acos(a); } + +/** \internal \returns the arc tangent of \a a (coeff-wise) */ +template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +Packet patan(const Packet& a) { using std::atan; return atan(a); } + +/** \internal \returns the hyperbolic sine of \a a (coeff-wise) */ +template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +Packet psinh(const Packet& a) { using std::sinh; return sinh(a); } + +/** \internal \returns the hyperbolic cosine of \a a (coeff-wise) */ +template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +Packet pcosh(const Packet& a) { using std::cosh; return cosh(a); } + +/** \internal \returns the hyperbolic tan of \a a (coeff-wise) */ +template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +Packet ptanh(const Packet& a) { using std::tanh; return tanh(a); } + +/** \internal \returns the exp of \a a (coeff-wise) */ +template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +Packet pexp(const Packet& a) { using std::exp; return exp(a); } + +/** \internal \returns the log of \a a (coeff-wise) */ +template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +Packet plog(const Packet& a) { using std::log; return log(a); } + +/** \internal \returns the log1p of \a a (coeff-wise) */ +template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +Packet plog1p(const Packet& a) { return numext::log1p(a); } + +/** \internal \returns the log10 of \a a (coeff-wise) */ +template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +Packet plog10(const Packet& a) { using std::log10; return log10(a); } + +/** \internal \returns the square-root of \a a (coeff-wise) */ +template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +Packet psqrt(const Packet& a) { using std::sqrt; return sqrt(a); } + +/** \internal \returns the reciprocal square-root of \a a (coeff-wise) */ +template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +Packet prsqrt(const Packet& a) { + return pdiv(pset1<Packet>(1), psqrt(a)); +} + +/** \internal \returns the rounded value of \a a (coeff-wise) */ +template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +Packet pround(const Packet& a) { using numext::round; return round(a); } + +/** \internal \returns the floor of \a a (coeff-wise) */ +template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +Packet pfloor(const Packet& a) { using numext::floor; return floor(a); } + +/** \internal \returns the ceil of \a a (coeff-wise) */ +template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +Packet pceil(const Packet& a) { using numext::ceil; return ceil(a); } + +/*************************************************************************** +* The following functions might not have to be overwritten for vectorized types +***************************************************************************/ + +/** \internal copy a packet with constant coeficient \a a (e.g., [a,a,a,a]) to \a *to. \a to must be 16 bytes aligned */ +// NOTE: this function must really be templated on the packet type (think about different packet types for the same scalar type) +template<typename Packet> +inline void pstore1(typename unpacket_traits<Packet>::type* to, const typename unpacket_traits<Packet>::type& a) +{ + pstore(to, pset1<Packet>(a)); +} + +/** \internal \returns a * b + c (coeff-wise) */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +pmadd(const Packet& a, + const Packet& b, + const Packet& c) +{ return padd(pmul(a, b),c); } + +/** \internal \returns a packet version of \a *from. + * The pointer \a from must be aligned on a \a Alignment bytes boundary. */ +template<typename Packet, int Alignment> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet ploadt(const typename unpacket_traits<Packet>::type* from) +{ + if(Alignment >= unpacket_traits<Packet>::alignment) + return pload<Packet>(from); + else + return ploadu<Packet>(from); +} + +/** \internal copy the packet \a from to \a *to. + * The pointer \a from must be aligned on a \a Alignment bytes boundary. */ +template<typename Scalar, typename Packet, int Alignment> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void pstoret(Scalar* to, const Packet& from) +{ + if(Alignment >= unpacket_traits<Packet>::alignment) + pstore(to, from); + else + pstoreu(to, from); +} + +/** \internal \returns a packet version of \a *from. + * Unlike ploadt, ploadt_ro takes advantage of the read-only memory path on the + * hardware if available to speedup the loading of data that won't be modified + * by the current computation. + */ +template<typename Packet, int LoadMode> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet ploadt_ro(const typename unpacket_traits<Packet>::type* from) +{ + return ploadt<Packet, LoadMode>(from); +} + +/** \internal default implementation of palign() allowing partial specialization */ +template<int Offset,typename PacketType> +struct palign_impl +{ + // by default data are aligned, so there is nothing to be done :) + static inline void run(PacketType&, const PacketType&) {} +}; + +/** \internal update \a first using the concatenation of the packet_size minus \a Offset last elements + * of \a first and \a Offset first elements of \a second. + * + * This function is currently only used to optimize matrix-vector products on unligned matrices. + * It takes 2 packets that represent a contiguous memory array, and returns a packet starting + * at the position \a Offset. For instance, for packets of 4 elements, we have: + * Input: + * - first = {f0,f1,f2,f3} + * - second = {s0,s1,s2,s3} + * Output: + * - if Offset==0 then {f0,f1,f2,f3} + * - if Offset==1 then {f1,f2,f3,s0} + * - if Offset==2 then {f2,f3,s0,s1} + * - if Offset==3 then {f3,s0,s1,s3} + */ +template<int Offset,typename PacketType> +inline void palign(PacketType& first, const PacketType& second) +{ + palign_impl<Offset,PacketType>::run(first,second); +} + +/*************************************************************************** +* Fast complex products (GCC generates a function call which is very slow) +***************************************************************************/ + +// Eigen+CUDA does not support complexes. +#ifndef __CUDACC__ + +template<> inline std::complex<float> pmul(const std::complex<float>& a, const std::complex<float>& b) +{ return std::complex<float>(real(a)*real(b) - imag(a)*imag(b), imag(a)*real(b) + real(a)*imag(b)); } + +template<> inline std::complex<double> pmul(const std::complex<double>& a, const std::complex<double>& b) +{ return std::complex<double>(real(a)*real(b) - imag(a)*imag(b), imag(a)*real(b) + real(a)*imag(b)); } + +#endif + + +/*************************************************************************** + * PacketBlock, that is a collection of N packets where the number of words + * in the packet is a multiple of N. +***************************************************************************/ +template <typename Packet,int N=unpacket_traits<Packet>::size> struct PacketBlock { + Packet packet[N]; +}; + +template<typename Packet> EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet,1>& /*kernel*/) { + // Nothing to do in the scalar case, i.e. a 1x1 matrix. +} + +/*************************************************************************** + * Selector, i.e. vector of N boolean values used to select (i.e. blend) + * words from 2 packets. +***************************************************************************/ +template <size_t N> struct Selector { + bool select[N]; +}; + +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +pblend(const Selector<unpacket_traits<Packet>::size>& ifPacket, const Packet& thenPacket, const Packet& elsePacket) { + return ifPacket.select[0] ? thenPacket : elsePacket; +} + +/** \internal \returns \a a with the first coefficient replaced by the scalar b */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +pinsertfirst(const Packet& a, typename unpacket_traits<Packet>::type b) +{ + // Default implementation based on pblend. + // It must be specialized for higher performance. + Selector<unpacket_traits<Packet>::size> mask; + mask.select[0] = true; + // This for loop should be optimized away by the compiler. + for(Index i=1; i<unpacket_traits<Packet>::size; ++i) + mask.select[i] = false; + return pblend(mask, pset1<Packet>(b), a); +} + +/** \internal \returns \a a with the last coefficient replaced by the scalar b */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +pinsertlast(const Packet& a, typename unpacket_traits<Packet>::type b) +{ + // Default implementation based on pblend. + // It must be specialized for higher performance. + Selector<unpacket_traits<Packet>::size> mask; + // This for loop should be optimized away by the compiler. + for(Index i=0; i<unpacket_traits<Packet>::size-1; ++i) + mask.select[i] = false; + mask.select[unpacket_traits<Packet>::size-1] = true; + return pblend(mask, pset1<Packet>(b), a); +} + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_GENERIC_PACKET_MATH_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/GlobalFunctions.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/GlobalFunctions.h new file mode 100644 index 000000000..769dc255c --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/GlobalFunctions.h @@ -0,0 +1,187 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2010-2016 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2010 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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_GLOBAL_FUNCTIONS_H +#define EIGEN_GLOBAL_FUNCTIONS_H + +#ifdef EIGEN_PARSED_BY_DOXYGEN + +#define EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(NAME,FUNCTOR,DOC_OP,DOC_DETAILS) \ + /** \returns an expression of the coefficient-wise DOC_OP of \a x + + DOC_DETAILS + + \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_##NAME">Math functions</a>, class CwiseUnaryOp + */ \ + template<typename Derived> \ + inline const Eigen::CwiseUnaryOp<Eigen::internal::FUNCTOR<typename Derived::Scalar>, const Derived> \ + NAME(const Eigen::ArrayBase<Derived>& x); + +#else + +#define EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(NAME,FUNCTOR,DOC_OP,DOC_DETAILS) \ + template<typename Derived> \ + inline const Eigen::CwiseUnaryOp<Eigen::internal::FUNCTOR<typename Derived::Scalar>, const Derived> \ + (NAME)(const Eigen::ArrayBase<Derived>& x) { \ + return Eigen::CwiseUnaryOp<Eigen::internal::FUNCTOR<typename Derived::Scalar>, const Derived>(x.derived()); \ + } + +#endif // EIGEN_PARSED_BY_DOXYGEN + +#define EIGEN_ARRAY_DECLARE_GLOBAL_EIGEN_UNARY(NAME,FUNCTOR) \ + \ + template<typename Derived> \ + struct NAME##_retval<ArrayBase<Derived> > \ + { \ + typedef const Eigen::CwiseUnaryOp<Eigen::internal::FUNCTOR<typename Derived::Scalar>, const Derived> type; \ + }; \ + template<typename Derived> \ + struct NAME##_impl<ArrayBase<Derived> > \ + { \ + static inline typename NAME##_retval<ArrayBase<Derived> >::type run(const Eigen::ArrayBase<Derived>& x) \ + { \ + return typename NAME##_retval<ArrayBase<Derived> >::type(x.derived()); \ + } \ + }; + +namespace Eigen +{ + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(real,scalar_real_op,real part,\sa ArrayBase::real) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(imag,scalar_imag_op,imaginary part,\sa ArrayBase::imag) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(conj,scalar_conjugate_op,complex conjugate,\sa ArrayBase::conjugate) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(inverse,scalar_inverse_op,inverse,\sa ArrayBase::inverse) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sin,scalar_sin_op,sine,\sa ArrayBase::sin) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(cos,scalar_cos_op,cosine,\sa ArrayBase::cos) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(tan,scalar_tan_op,tangent,\sa ArrayBase::tan) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(atan,scalar_atan_op,arc-tangent,\sa ArrayBase::atan) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(asin,scalar_asin_op,arc-sine,\sa ArrayBase::asin) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(acos,scalar_acos_op,arc-consine,\sa ArrayBase::acos) + 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) + 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) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(erf,scalar_erf_op,error function,\sa ArrayBase::erf) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(erfc,scalar_erfc_op,complement error function,\sa ArrayBase::erfc) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(exp,scalar_exp_op,exponential,\sa ArrayBase::exp) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(log,scalar_log_op,natural logarithm,\sa Eigen::log10 DOXCOMMA ArrayBase::log) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(log1p,scalar_log1p_op,natural logarithm of 1 plus the value,\sa ArrayBase::log1p) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(log10,scalar_log10_op,base 10 logarithm,\sa Eigen::log DOXCOMMA ArrayBase::log) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(abs,scalar_abs_op,absolute value,\sa ArrayBase::abs DOXCOMMA MatrixBase::cwiseAbs) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(abs2,scalar_abs2_op,squared absolute value,\sa ArrayBase::abs2 DOXCOMMA MatrixBase::cwiseAbs2) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(arg,scalar_arg_op,complex argument,\sa ArrayBase::arg) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sqrt,scalar_sqrt_op,square root,\sa ArrayBase::sqrt DOXCOMMA MatrixBase::cwiseSqrt) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(rsqrt,scalar_rsqrt_op,reciprocal square root,\sa ArrayBase::rsqrt) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(square,scalar_square_op,square (power 2),\sa Eigen::abs2 DOXCOMMA Eigen::pow DOXCOMMA ArrayBase::square) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(cube,scalar_cube_op,cube (power 3),\sa Eigen::pow DOXCOMMA ArrayBase::cube) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(round,scalar_round_op,nearest integer,\sa Eigen::floor DOXCOMMA Eigen::ceil DOXCOMMA ArrayBase::round) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(floor,scalar_floor_op,nearest integer not greater than the giben value,\sa Eigen::ceil DOXCOMMA ArrayBase::floor) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(ceil,scalar_ceil_op,nearest integer not less than the giben value,\sa Eigen::floor DOXCOMMA ArrayBase::ceil) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(isnan,scalar_isnan_op,not-a-number test,\sa Eigen::isinf DOXCOMMA Eigen::isfinite DOXCOMMA ArrayBase::isnan) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(isinf,scalar_isinf_op,infinite value test,\sa Eigen::isnan DOXCOMMA Eigen::isfinite DOXCOMMA ArrayBase::isinf) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(isfinite,scalar_isfinite_op,finite value test,\sa Eigen::isinf DOXCOMMA Eigen::isnan DOXCOMMA ArrayBase::isfinite) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sign,scalar_sign_op,sign (or 0),\sa ArrayBase::sign) + + /** \returns an expression of the coefficient-wise power of \a x to the given constant \a exponent. + * + * \tparam ScalarExponent is the scalar type of \a exponent. It must be compatible with the scalar type of the given expression (\c Derived::Scalar). + * + * \sa ArrayBase::pow() + * + * \relates ArrayBase + */ +#ifdef EIGEN_PARSED_BY_DOXYGEN + template<typename Derived,typename ScalarExponent> + inline const CwiseBinaryOp<internal::scalar_pow_op<Derived::Scalar,ScalarExponent>,Derived,Constant<ScalarExponent> > + pow(const Eigen::ArrayBase<Derived>& x, const ScalarExponent& exponent); +#else + template<typename Derived,typename ScalarExponent> + inline typename internal::enable_if< !(internal::is_same<typename Derived::Scalar,ScalarExponent>::value) && EIGEN_SCALAR_BINARY_SUPPORTED(pow,typename Derived::Scalar,ScalarExponent), + const EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(Derived,ScalarExponent,pow) >::type + pow(const Eigen::ArrayBase<Derived>& x, const ScalarExponent& exponent) { + return x.derived().pow(exponent); + } + + template<typename Derived> + inline const EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(Derived,typename Derived::Scalar,pow) + pow(const Eigen::ArrayBase<Derived>& x, const typename Derived::Scalar& exponent) { + return x.derived().pow(exponent); + } +#endif + + /** \returns an expression of the coefficient-wise power of \a x to the given array of \a exponents. + * + * This function computes the coefficient-wise power. + * + * Example: \include Cwise_array_power_array.cpp + * Output: \verbinclude Cwise_array_power_array.out + * + * \sa ArrayBase::pow() + * + * \relates ArrayBase + */ + template<typename Derived,typename ExponentDerived> + inline const Eigen::CwiseBinaryOp<Eigen::internal::scalar_pow_op<typename Derived::Scalar, typename ExponentDerived::Scalar>, const Derived, const ExponentDerived> + pow(const Eigen::ArrayBase<Derived>& x, const Eigen::ArrayBase<ExponentDerived>& exponents) + { + return Eigen::CwiseBinaryOp<Eigen::internal::scalar_pow_op<typename Derived::Scalar, typename ExponentDerived::Scalar>, const Derived, const ExponentDerived>( + x.derived(), + exponents.derived() + ); + } + + /** \returns an expression of the coefficient-wise power of the scalar \a x to the given array of \a exponents. + * + * This function computes the coefficient-wise power between a scalar and an array of exponents. + * + * \tparam Scalar is the scalar type of \a x. It must be compatible with the scalar type of the given array expression (\c Derived::Scalar). + * + * Example: \include Cwise_scalar_power_array.cpp + * Output: \verbinclude Cwise_scalar_power_array.out + * + * \sa ArrayBase::pow() + * + * \relates ArrayBase + */ +#ifdef EIGEN_PARSED_BY_DOXYGEN + template<typename Scalar,typename Derived> + inline const CwiseBinaryOp<internal::scalar_pow_op<Scalar,Derived::Scalar>,Constant<Scalar>,Derived> + pow(const Scalar& x,const Eigen::ArrayBase<Derived>& x); +#else + template<typename Scalar, typename Derived> + inline typename internal::enable_if< !(internal::is_same<typename Derived::Scalar,Scalar>::value) && EIGEN_SCALAR_BINARY_SUPPORTED(pow,Scalar,typename Derived::Scalar), + const EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(Scalar,Derived,pow) >::type + pow(const Scalar& x, const Eigen::ArrayBase<Derived>& exponents) + { + return EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(Scalar,Derived,pow)( + typename internal::plain_constant_type<Derived,Scalar>::type(exponents.rows(), exponents.cols(), x), exponents.derived() ); + } + + template<typename Derived> + inline const EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(typename Derived::Scalar,Derived,pow) + pow(const typename Derived::Scalar& x, const Eigen::ArrayBase<Derived>& exponents) + { + return EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(typename Derived::Scalar,Derived,pow)( + typename internal::plain_constant_type<Derived,typename Derived::Scalar>::type(exponents.rows(), exponents.cols(), x), exponents.derived() ); + } +#endif + + + namespace internal + { + EIGEN_ARRAY_DECLARE_GLOBAL_EIGEN_UNARY(real,scalar_real_op) + EIGEN_ARRAY_DECLARE_GLOBAL_EIGEN_UNARY(imag,scalar_imag_op) + EIGEN_ARRAY_DECLARE_GLOBAL_EIGEN_UNARY(abs2,scalar_abs2_op) + } +} + +// TODO: cleanly disable those functions that are not supported on Array (numext::real_ref, internal::random, internal::isApprox...) + +#endif // EIGEN_GLOBAL_FUNCTIONS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/IO.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/IO.h new file mode 100644 index 000000000..da7fd6cce --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/IO.h @@ -0,0 +1,225 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com> +// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr> +// +// 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_IO_H +#define EIGEN_IO_H + +namespace Eigen { + +enum { DontAlignCols = 1 }; +enum { StreamPrecision = -1, + FullPrecision = -2 }; + +namespace internal { +template<typename Derived> +std::ostream & print_matrix(std::ostream & s, const Derived& _m, const IOFormat& fmt); +} + +/** \class IOFormat + * \ingroup Core_Module + * + * \brief Stores a set of parameters controlling the way matrices are printed + * + * List of available parameters: + * - \b precision number of digits for floating point values, or one of the special constants \c StreamPrecision and \c FullPrecision. + * The default is the special value \c StreamPrecision which means to use the + * stream's own precision setting, as set for instance using \c cout.precision(3). The other special value + * \c FullPrecision means that the number of digits will be computed to match the full precision of each floating-point + * type. + * - \b flags an OR-ed combination of flags, the default value is 0, the only currently available flag is \c DontAlignCols which + * allows to disable the alignment of columns, resulting in faster code. + * - \b coeffSeparator string printed between two coefficients of the same row + * - \b rowSeparator string printed between two rows + * - \b rowPrefix string printed at the beginning of each row + * - \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 + * + * Example: \include IOFormat.cpp + * Output: \verbinclude IOFormat.out + * + * \sa DenseBase::format(), class WithFormat + */ +struct IOFormat +{ + /** Default constructor, see class IOFormat for the meaning of the parameters */ + 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="") + : matPrefix(_matPrefix), matSuffix(_matSuffix), rowPrefix(_rowPrefix), rowSuffix(_rowSuffix), rowSeparator(_rowSeparator), + rowSpacer(""), coeffSeparator(_coeffSeparator), 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 + if((flags & DontAlignCols)) + return; + int i = int(matSuffix.length())-1; + while (i>=0 && matSuffix[i]!='\n') + { + rowSpacer += ' '; + i--; + } + } + std::string matPrefix, matSuffix; + std::string rowPrefix, rowSuffix, rowSeparator, rowSpacer; + std::string coeffSeparator; + int precision; + int flags; +}; + +/** \class WithFormat + * \ingroup Core_Module + * + * \brief Pseudo expression providing matrix output with given format + * + * \tparam ExpressionType the type of the object on which IO stream operations are performed + * + * This class represents an expression with stream operators controlled by a given IOFormat. + * It is the return type of DenseBase::format() + * and most of the time this is the only way it is used. + * + * See class IOFormat for some examples. + * + * \sa DenseBase::format(), class IOFormat + */ +template<typename ExpressionType> +class WithFormat +{ + public: + + WithFormat(const ExpressionType& matrix, const IOFormat& format) + : m_matrix(matrix), m_format(format) + {} + + friend std::ostream & operator << (std::ostream & s, const WithFormat& wf) + { + return internal::print_matrix(s, wf.m_matrix.eval(), wf.m_format); + } + + protected: + typename ExpressionType::Nested m_matrix; + IOFormat m_format; +}; + +namespace internal { + +// NOTE: This helper is kept for backward compatibility with previous code specializing +// this internal::significant_decimals_impl structure. In the future we should directly +// call digits10() which has been introduced in July 2016 in 3.3. +template<typename Scalar> +struct significant_decimals_impl +{ + static inline int run() + { + return NumTraits<Scalar>::digits10(); + } +}; + +/** \internal + * print the matrix \a _m to the output stream \a s using the output format \a fmt */ +template<typename Derived> +std::ostream & print_matrix(std::ostream & s, const Derived& _m, const IOFormat& fmt) +{ + if(_m.size() == 0) + { + s << fmt.matPrefix << fmt.matSuffix; + return s; + } + + typename Derived::Nested m = _m; + typedef typename Derived::Scalar Scalar; + + Index width = 0; + + std::streamsize explicit_precision; + if(fmt.precision == StreamPrecision) + { + explicit_precision = 0; + } + else if(fmt.precision == FullPrecision) + { + if (NumTraits<Scalar>::IsInteger) + { + explicit_precision = 0; + } + else + { + explicit_precision = significant_decimals_impl<Scalar>::run(); + } + } + else + { + explicit_precision = fmt.precision; + } + + std::streamsize old_precision = 0; + if(explicit_precision) old_precision = s.precision(explicit_precision); + + bool align_cols = !(fmt.flags & DontAlignCols); + if(align_cols) + { + // compute the largest width + for(Index j = 0; j < m.cols(); ++j) + for(Index i = 0; i < m.rows(); ++i) + { + std::stringstream sstr; + sstr.copyfmt(s); + sstr << m.coeff(i,j); + width = std::max<Index>(width, Index(sstr.str().length())); + } + } + s << fmt.matPrefix; + for(Index i = 0; i < m.rows(); ++i) + { + if (i) + s << fmt.rowSpacer; + s << fmt.rowPrefix; + if(width) s.width(width); + s << m.coeff(i, 0); + for(Index j = 1; j < m.cols(); ++j) + { + s << fmt.coeffSeparator; + if (width) s.width(width); + s << m.coeff(i, j); + } + s << fmt.rowSuffix; + if( i < m.rows() - 1) + s << fmt.rowSeparator; + } + s << fmt.matSuffix; + if(explicit_precision) s.precision(old_precision); + return s; +} + +} // end namespace internal + +/** \relates DenseBase + * + * Outputs the matrix, to the given stream. + * + * If you wish to print the matrix with a format different than the default, use DenseBase::format(). + * + * It is also possible to change the default format by defining EIGEN_DEFAULT_IO_FORMAT before including Eigen headers. + * If not defined, this will automatically be defined to Eigen::IOFormat(), that is the Eigen::IOFormat with default parameters. + * + * \sa DenseBase::format() + */ +template<typename Derived> +std::ostream & operator << +(std::ostream & s, + const DenseBase<Derived> & m) +{ + return internal::print_matrix(s, m.eval(), EIGEN_DEFAULT_IO_FORMAT); +} + +} // end namespace Eigen + +#endif // EIGEN_IO_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/Inverse.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Inverse.h new file mode 100644 index 000000000..b76f0439d --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Inverse.h @@ -0,0 +1,118 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 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_INVERSE_H +#define EIGEN_INVERSE_H + +namespace Eigen { + +template<typename XprType,typename StorageKind> class InverseImpl; + +namespace internal { + +template<typename XprType> +struct traits<Inverse<XprType> > + : traits<typename XprType::PlainObject> +{ + typedef typename XprType::PlainObject PlainObject; + typedef traits<PlainObject> BaseTraits; + enum { + Flags = BaseTraits::Flags & RowMajorBit + }; +}; + +} // end namespace internal + +/** \class Inverse + * + * \brief Expression of the inverse of another expression + * + * \tparam XprType the type of the expression we are taking the inverse + * + * This class represents an abstract expression of A.inverse() + * and most of the time this is the only way it is used. + * + */ +template<typename XprType> +class Inverse : public InverseImpl<XprType,typename internal::traits<XprType>::StorageKind> +{ +public: + typedef typename XprType::StorageIndex StorageIndex; + typedef typename XprType::PlainObject PlainObject; + typedef typename XprType::Scalar Scalar; + typedef typename internal::ref_selector<XprType>::type XprTypeNested; + typedef typename internal::remove_all<XprTypeNested>::type XprTypeNestedCleaned; + typedef typename internal::ref_selector<Inverse>::type Nested; + typedef typename internal::remove_all<XprType>::type NestedExpression; + + explicit EIGEN_DEVICE_FUNC Inverse(const XprType &xpr) + : m_xpr(xpr) + {} + + EIGEN_DEVICE_FUNC Index rows() const { return m_xpr.rows(); } + EIGEN_DEVICE_FUNC Index cols() const { return m_xpr.cols(); } + + EIGEN_DEVICE_FUNC const XprTypeNestedCleaned& nestedExpression() const { return m_xpr; } + +protected: + XprTypeNested m_xpr; +}; + +// Generic API dispatcher +template<typename XprType, typename StorageKind> +class InverseImpl + : public internal::generic_xpr_base<Inverse<XprType> >::type +{ +public: + typedef typename internal::generic_xpr_base<Inverse<XprType> >::type Base; + typedef typename XprType::Scalar Scalar; +private: + + Scalar coeff(Index row, Index col) const; + Scalar coeff(Index i) const; +}; + +namespace internal { + +/** \internal + * \brief Default evaluator for Inverse expression. + * + * This default evaluator for Inverse expression simply evaluate the inverse into a temporary + * by a call to internal::call_assignment_no_alias. + * Therefore, inverse implementers only have to specialize Assignment<Dst,Inverse<...>, ...> for + * there own nested expression. + * + * \sa class Inverse + */ +template<typename ArgType> +struct unary_evaluator<Inverse<ArgType> > + : public evaluator<typename Inverse<ArgType>::PlainObject> +{ + typedef Inverse<ArgType> InverseType; + typedef typename InverseType::PlainObject PlainObject; + typedef evaluator<PlainObject> Base; + + enum { Flags = Base::Flags | EvalBeforeNestingBit }; + + unary_evaluator(const InverseType& inv_xpr) + : m_result(inv_xpr.rows(), inv_xpr.cols()) + { + ::new (static_cast<Base*>(this)) Base(m_result); + internal::call_assignment_no_alias(m_result, inv_xpr); + } + +protected: + PlainObject m_result; +}; + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_INVERSE_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/Map.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Map.h new file mode 100644 index 000000000..06d196702 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Map.h @@ -0,0 +1,164 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2007-2010 Benoit Jacob <jacob.benoit.1@gmail.com> +// Copyright (C) 2008 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_MAP_H +#define EIGEN_MAP_H + +namespace Eigen { + +namespace internal { +template<typename PlainObjectType, int MapOptions, typename StrideType> +struct traits<Map<PlainObjectType, MapOptions, StrideType> > + : public traits<PlainObjectType> +{ + typedef traits<PlainObjectType> TraitsBase; + enum { + InnerStrideAtCompileTime = StrideType::InnerStrideAtCompileTime == 0 + ? int(PlainObjectType::InnerStrideAtCompileTime) + : int(StrideType::InnerStrideAtCompileTime), + OuterStrideAtCompileTime = StrideType::OuterStrideAtCompileTime == 0 + ? int(PlainObjectType::OuterStrideAtCompileTime) + : int(StrideType::OuterStrideAtCompileTime), + Alignment = int(MapOptions)&int(AlignedMask), + Flags0 = TraitsBase::Flags & (~NestByRefBit), + Flags = is_lvalue<PlainObjectType>::value ? int(Flags0) : (int(Flags0) & ~LvalueBit) + }; +private: + enum { Options }; // Expressions don't have Options +}; +} + +/** \class Map + * \ingroup Core_Module + * + * \brief A matrix or vector expression mapping an existing array of data. + * + * \tparam PlainObjectType the equivalent matrix type of the mapped data + * \tparam MapOptions specifies the pointer alignment in bytes. It can be: \c #Aligned128, , \c #Aligned64, \c #Aligned32, \c #Aligned16, \c #Aligned8 or \c #Unaligned. + * The default is \c #Unaligned. + * \tparam StrideType optionally specifies strides. By default, Map assumes the memory layout + * of an ordinary, contiguous array. This can be overridden by specifying strides. + * The type passed here must be a specialization of the Stride template, see examples below. + * + * This class represents a matrix or vector expression mapping an existing array of data. + * It can be used to let Eigen interface without any overhead with non-Eigen data structures, + * such as plain C arrays or structures from other libraries. By default, it assumes that the + * data is laid out contiguously in memory. You can however override this by explicitly specifying + * inner and outer strides. + * + * Here's an example of simply mapping a contiguous array as a \ref TopicStorageOrders "column-major" matrix: + * \include Map_simple.cpp + * Output: \verbinclude Map_simple.out + * + * If you need to map non-contiguous arrays, you can do so by specifying strides: + * + * Here's an example of mapping an array as a vector, specifying an inner stride, that is, the pointer + * increment between two consecutive coefficients. Here, we're specifying the inner stride as a compile-time + * fixed value. + * \include Map_inner_stride.cpp + * Output: \verbinclude Map_inner_stride.out + * + * Here's an example of mapping an array while specifying an outer stride. Here, since we're mapping + * as a column-major matrix, 'outer stride' means the pointer increment between two consecutive columns. + * Here, we're specifying the outer stride as a runtime parameter. Note that here \c OuterStride<> is + * a short version of \c OuterStride<Dynamic> because the default template parameter of OuterStride + * is \c Dynamic + * \include Map_outer_stride.cpp + * Output: \verbinclude Map_outer_stride.out + * + * For more details and for an example of specifying both an inner and an outer stride, see class Stride. + * + * \b Tip: to change the array of data mapped by a Map object, you can use the C++ + * placement new syntax: + * + * Example: \include Map_placement_new.cpp + * Output: \verbinclude Map_placement_new.out + * + * This class is the return type of PlainObjectBase::Map() but can also be used directly. + * + * \sa PlainObjectBase::Map(), \ref TopicStorageOrders + */ +template<typename PlainObjectType, int MapOptions, typename StrideType> class Map + : public MapBase<Map<PlainObjectType, MapOptions, StrideType> > +{ + public: + + typedef MapBase<Map> Base; + EIGEN_DENSE_PUBLIC_INTERFACE(Map) + + typedef typename Base::PointerType PointerType; + typedef PointerType PointerArgType; + EIGEN_DEVICE_FUNC + inline PointerType cast_to_pointer_type(PointerArgType ptr) { return ptr; } + + EIGEN_DEVICE_FUNC + inline Index innerStride() const + { + return StrideType::InnerStrideAtCompileTime != 0 ? m_stride.inner() : 1; + } + + EIGEN_DEVICE_FUNC + inline Index outerStride() const + { + return StrideType::OuterStrideAtCompileTime != 0 ? m_stride.outer() + : IsVectorAtCompileTime ? this->size() + : int(Flags)&RowMajorBit ? this->cols() + : this->rows(); + } + + /** Constructor in the fixed-size case. + * + * \param dataPtr pointer to the array to map + * \param stride optional Stride object, passing the strides. + */ + EIGEN_DEVICE_FUNC + explicit inline Map(PointerArgType dataPtr, const StrideType& stride = StrideType()) + : Base(cast_to_pointer_type(dataPtr)), m_stride(stride) + { + PlainObjectType::Base::_check_template_params(); + } + + /** Constructor in the dynamic-size vector case. + * + * \param dataPtr pointer to the array to map + * \param size the size of the vector expression + * \param stride optional Stride object, passing the strides. + */ + EIGEN_DEVICE_FUNC + inline Map(PointerArgType dataPtr, Index size, const StrideType& stride = StrideType()) + : Base(cast_to_pointer_type(dataPtr), size), m_stride(stride) + { + PlainObjectType::Base::_check_template_params(); + } + + /** Constructor in the dynamic-size matrix case. + * + * \param dataPtr pointer to the array to map + * \param rows the number of rows of the matrix expression + * \param cols the number of columns of the matrix expression + * \param stride optional Stride object, passing the strides. + */ + EIGEN_DEVICE_FUNC + inline Map(PointerArgType dataPtr, Index rows, Index cols, const StrideType& stride = StrideType()) + : Base(cast_to_pointer_type(dataPtr), rows, cols), m_stride(stride) + { + PlainObjectType::Base::_check_template_params(); + } + + EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Map) + + protected: + StrideType m_stride; +}; + + +} // end namespace Eigen + +#endif // EIGEN_MAP_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/MapBase.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/MapBase.h new file mode 100644 index 000000000..020f939ad --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/MapBase.h @@ -0,0 +1,299 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2007-2010 Benoit Jacob <jacob.benoit.1@gmail.com> +// Copyright (C) 2008 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_MAPBASE_H +#define EIGEN_MAPBASE_H + +#define EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived) \ + EIGEN_STATIC_ASSERT((int(internal::evaluator<Derived>::Flags) & LinearAccessBit) || Derived::IsVectorAtCompileTime, \ + YOU_ARE_TRYING_TO_USE_AN_INDEX_BASED_ACCESSOR_ON_AN_EXPRESSION_THAT_DOES_NOT_SUPPORT_THAT) + +namespace Eigen { + +/** \ingroup Core_Module + * + * \brief Base class for dense Map and Block expression with direct access + * + * This base class provides the const low-level accessors (e.g. coeff, coeffRef) of dense + * Map and Block objects with direct access. + * Typical users do not have to directly deal with this class. + * + * This class can be extended by through the macro plugin \c EIGEN_MAPBASE_PLUGIN. + * See \link TopicCustomizing_Plugins customizing Eigen \endlink for details. + * + * The \c Derived class has to provide the following two methods describing the memory layout: + * \code Index innerStride() const; \endcode + * \code Index outerStride() const; \endcode + * + * \sa class Map, class Block + */ +template<typename Derived> class MapBase<Derived, ReadOnlyAccessors> + : public internal::dense_xpr_base<Derived>::type +{ + public: + + typedef typename internal::dense_xpr_base<Derived>::type Base; + enum { + RowsAtCompileTime = internal::traits<Derived>::RowsAtCompileTime, + ColsAtCompileTime = internal::traits<Derived>::ColsAtCompileTime, + SizeAtCompileTime = Base::SizeAtCompileTime + }; + + typedef typename internal::traits<Derived>::StorageKind StorageKind; + typedef typename internal::traits<Derived>::Scalar Scalar; + typedef typename internal::packet_traits<Scalar>::type PacketScalar; + typedef typename NumTraits<Scalar>::Real RealScalar; + typedef typename internal::conditional< + bool(internal::is_lvalue<Derived>::value), + Scalar *, + const Scalar *>::type + PointerType; + + using Base::derived; +// using Base::RowsAtCompileTime; +// using Base::ColsAtCompileTime; +// using Base::SizeAtCompileTime; + using Base::MaxRowsAtCompileTime; + using Base::MaxColsAtCompileTime; + using Base::MaxSizeAtCompileTime; + using Base::IsVectorAtCompileTime; + using Base::Flags; + using Base::IsRowMajor; + + using Base::rows; + using Base::cols; + using Base::size; + using Base::coeff; + using Base::coeffRef; + using Base::lazyAssign; + using Base::eval; + + using Base::innerStride; + using Base::outerStride; + using Base::rowStride; + using Base::colStride; + + // bug 217 - compile error on ICC 11.1 + using Base::operator=; + + typedef typename Base::CoeffReturnType CoeffReturnType; + + /** \copydoc DenseBase::rows() */ + EIGEN_DEVICE_FUNC inline Index rows() const { return m_rows.value(); } + /** \copydoc DenseBase::cols() */ + EIGEN_DEVICE_FUNC inline Index cols() const { return m_cols.value(); } + + /** Returns a pointer to the first coefficient of the matrix or vector. + * + * \note When addressing this data, make sure to honor the strides returned by innerStride() and outerStride(). + * + * \sa innerStride(), outerStride() + */ + EIGEN_DEVICE_FUNC inline const Scalar* data() const { return m_data; } + + /** \copydoc PlainObjectBase::coeff(Index,Index) const */ + EIGEN_DEVICE_FUNC + inline const Scalar& coeff(Index rowId, Index colId) const + { + return m_data[colId * colStride() + rowId * rowStride()]; + } + + /** \copydoc PlainObjectBase::coeff(Index) const */ + EIGEN_DEVICE_FUNC + inline const Scalar& coeff(Index index) const + { + EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived) + return m_data[index * innerStride()]; + } + + /** \copydoc PlainObjectBase::coeffRef(Index,Index) const */ + EIGEN_DEVICE_FUNC + inline const Scalar& coeffRef(Index rowId, Index colId) const + { + return this->m_data[colId * colStride() + rowId * rowStride()]; + } + + /** \copydoc PlainObjectBase::coeffRef(Index) const */ + EIGEN_DEVICE_FUNC + inline const Scalar& coeffRef(Index index) const + { + EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived) + return this->m_data[index * innerStride()]; + } + + /** \internal */ + template<int LoadMode> + inline PacketScalar packet(Index rowId, Index colId) const + { + return internal::ploadt<PacketScalar, LoadMode> + (m_data + (colId * colStride() + rowId * rowStride())); + } + + /** \internal */ + template<int LoadMode> + inline PacketScalar packet(Index index) const + { + EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived) + return internal::ploadt<PacketScalar, LoadMode>(m_data + index * innerStride()); + } + + /** \internal Constructor for fixed size matrices or vectors */ + EIGEN_DEVICE_FUNC + explicit inline MapBase(PointerType dataPtr) : m_data(dataPtr), m_rows(RowsAtCompileTime), m_cols(ColsAtCompileTime) + { + EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived) + checkSanity<Derived>(); + } + + /** \internal Constructor for dynamically sized vectors */ + EIGEN_DEVICE_FUNC + inline MapBase(PointerType dataPtr, Index vecSize) + : m_data(dataPtr), + m_rows(RowsAtCompileTime == Dynamic ? vecSize : Index(RowsAtCompileTime)), + m_cols(ColsAtCompileTime == Dynamic ? vecSize : Index(ColsAtCompileTime)) + { + EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) + eigen_assert(vecSize >= 0); + eigen_assert(dataPtr == 0 || SizeAtCompileTime == Dynamic || SizeAtCompileTime == vecSize); + checkSanity<Derived>(); + } + + /** \internal Constructor for dynamically sized matrices */ + EIGEN_DEVICE_FUNC + inline MapBase(PointerType dataPtr, Index rows, Index cols) + : m_data(dataPtr), m_rows(rows), m_cols(cols) + { + eigen_assert( (dataPtr == 0) + || ( rows >= 0 && (RowsAtCompileTime == Dynamic || RowsAtCompileTime == rows) + && cols >= 0 && (ColsAtCompileTime == Dynamic || ColsAtCompileTime == cols))); + checkSanity<Derived>(); + } + + #ifdef EIGEN_MAPBASE_PLUGIN + #include EIGEN_MAPBASE_PLUGIN + #endif + + protected: + + template<typename T> + EIGEN_DEVICE_FUNC + void checkSanity(typename internal::enable_if<(internal::traits<T>::Alignment>0),void*>::type = 0) const + { +#if EIGEN_MAX_ALIGN_BYTES>0 + eigen_assert(( ((internal::UIntPtr(m_data) % internal::traits<Derived>::Alignment) == 0) + || (cols() * rows() * innerStride() * sizeof(Scalar)) < internal::traits<Derived>::Alignment ) && "data is not aligned"); +#endif + } + + template<typename T> + EIGEN_DEVICE_FUNC + void checkSanity(typename internal::enable_if<internal::traits<T>::Alignment==0,void*>::type = 0) const + {} + + PointerType m_data; + const internal::variable_if_dynamic<Index, RowsAtCompileTime> m_rows; + const internal::variable_if_dynamic<Index, ColsAtCompileTime> m_cols; +}; + +/** \ingroup Core_Module + * + * \brief Base class for non-const dense Map and Block expression with direct access + * + * This base class provides the non-const low-level accessors (e.g. coeff and coeffRef) of + * dense Map and Block objects with direct access. + * It inherits MapBase<Derived, ReadOnlyAccessors> which defines the const variant for reading specific entries. + * + * \sa class Map, class Block + */ +template<typename Derived> class MapBase<Derived, WriteAccessors> + : public MapBase<Derived, ReadOnlyAccessors> +{ + typedef MapBase<Derived, ReadOnlyAccessors> ReadOnlyMapBase; + public: + + typedef MapBase<Derived, ReadOnlyAccessors> Base; + + typedef typename Base::Scalar Scalar; + typedef typename Base::PacketScalar PacketScalar; + typedef typename Base::StorageIndex StorageIndex; + typedef typename Base::PointerType PointerType; + + using Base::derived; + using Base::rows; + using Base::cols; + using Base::size; + using Base::coeff; + using Base::coeffRef; + + using Base::innerStride; + using Base::outerStride; + using Base::rowStride; + using Base::colStride; + + typedef typename internal::conditional< + internal::is_lvalue<Derived>::value, + Scalar, + const Scalar + >::type ScalarWithConstIfNotLvalue; + + EIGEN_DEVICE_FUNC + inline const Scalar* data() const { return this->m_data; } + EIGEN_DEVICE_FUNC + inline ScalarWithConstIfNotLvalue* data() { return this->m_data; } // no const-cast here so non-const-correct code will give a compile error + + EIGEN_DEVICE_FUNC + inline ScalarWithConstIfNotLvalue& coeffRef(Index row, Index col) + { + return this->m_data[col * colStride() + row * rowStride()]; + } + + EIGEN_DEVICE_FUNC + inline ScalarWithConstIfNotLvalue& coeffRef(Index index) + { + EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived) + return this->m_data[index * innerStride()]; + } + + template<int StoreMode> + inline void writePacket(Index row, Index col, const PacketScalar& val) + { + internal::pstoret<Scalar, PacketScalar, StoreMode> + (this->m_data + (col * colStride() + row * rowStride()), val); + } + + template<int StoreMode> + inline void writePacket(Index index, const PacketScalar& val) + { + EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived) + internal::pstoret<Scalar, PacketScalar, StoreMode> + (this->m_data + index * innerStride(), val); + } + + EIGEN_DEVICE_FUNC explicit inline MapBase(PointerType dataPtr) : Base(dataPtr) {} + EIGEN_DEVICE_FUNC inline MapBase(PointerType dataPtr, Index vecSize) : Base(dataPtr, vecSize) {} + EIGEN_DEVICE_FUNC inline MapBase(PointerType dataPtr, Index rows, Index cols) : Base(dataPtr, rows, cols) {} + + EIGEN_DEVICE_FUNC + Derived& operator=(const MapBase& other) + { + ReadOnlyMapBase::Base::operator=(other); + return derived(); + } + + // In theory we could simply refer to Base:Base::operator=, but MSVC does not like Base::Base, + // see bugs 821 and 920. + using ReadOnlyMapBase::Base::operator=; +}; + +#undef EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS + +} // end namespace Eigen + +#endif // EIGEN_MAPBASE_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/MathFunctions.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/MathFunctions.h new file mode 100644 index 000000000..a648aa0fa --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/MathFunctions.h @@ -0,0 +1,1431 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2006-2010 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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_MATHFUNCTIONS_H +#define EIGEN_MATHFUNCTIONS_H + +// source: http://www.geom.uiuc.edu/~huberty/math5337/groupe/digits.html +// TODO this should better be moved to NumTraits +#define EIGEN_PI 3.141592653589793238462643383279502884197169399375105820974944592307816406L + + +namespace Eigen { + +// On WINCE, std::abs is defined for int only, so let's defined our own overloads: +// This issue has been confirmed with MSVC 2008 only, but the issue might exist for more recent versions too. +#if EIGEN_OS_WINCE && EIGEN_COMP_MSVC && EIGEN_COMP_MSVC<=1500 +long abs(long x) { return (labs(x)); } +double abs(double x) { return (fabs(x)); } +float abs(float x) { return (fabsf(x)); } +long double abs(long double x) { return (fabsl(x)); } +#endif + +namespace internal { + +/** \internal \class global_math_functions_filtering_base + * + * What it does: + * Defines a typedef 'type' as follows: + * - if type T has a member typedef Eigen_BaseClassForSpecializationOfGlobalMathFuncImpl, then + * global_math_functions_filtering_base<T>::type is a typedef for it. + * - otherwise, global_math_functions_filtering_base<T>::type is a typedef for T. + * + * How it's used: + * To allow to defined the global math functions (like sin...) in certain cases, like the Array expressions. + * When you do sin(array1+array2), the object array1+array2 has a complicated expression type, all what you want to know + * is that it inherits ArrayBase. So we implement a partial specialization of sin_impl for ArrayBase<Derived>. + * So we must make sure to use sin_impl<ArrayBase<Derived> > and not sin_impl<Derived>, otherwise our partial specialization + * won't be used. How does sin know that? That's exactly what global_math_functions_filtering_base tells it. + * + * How it's implemented: + * SFINAE in the style of enable_if. Highly susceptible of breaking compilers. With GCC, it sure does work, but if you replace + * the typename dummy by an integer template parameter, it doesn't work anymore! + */ + +template<typename T, typename dummy = void> +struct global_math_functions_filtering_base +{ + typedef T type; +}; + +template<typename T> struct always_void { typedef void type; }; + +template<typename T> +struct global_math_functions_filtering_base + <T, + typename always_void<typename T::Eigen_BaseClassForSpecializationOfGlobalMathFuncImpl>::type + > +{ + typedef typename T::Eigen_BaseClassForSpecializationOfGlobalMathFuncImpl type; +}; + +#define EIGEN_MATHFUNC_IMPL(func, scalar) Eigen::internal::func##_impl<typename Eigen::internal::global_math_functions_filtering_base<scalar>::type> +#define EIGEN_MATHFUNC_RETVAL(func, scalar) typename Eigen::internal::func##_retval<typename Eigen::internal::global_math_functions_filtering_base<scalar>::type>::type + +/**************************************************************************** +* Implementation of real * +****************************************************************************/ + +template<typename Scalar, bool IsComplex = NumTraits<Scalar>::IsComplex> +struct real_default_impl +{ + typedef typename NumTraits<Scalar>::Real RealScalar; + EIGEN_DEVICE_FUNC + static inline RealScalar run(const Scalar& x) + { + return x; + } +}; + +template<typename Scalar> +struct real_default_impl<Scalar,true> +{ + typedef typename NumTraits<Scalar>::Real RealScalar; + EIGEN_DEVICE_FUNC + static inline RealScalar run(const Scalar& x) + { + using std::real; + return real(x); + } +}; + +template<typename Scalar> struct real_impl : real_default_impl<Scalar> {}; + +#ifdef __CUDA_ARCH__ +template<typename T> +struct real_impl<std::complex<T> > +{ + typedef T RealScalar; + EIGEN_DEVICE_FUNC + static inline T run(const std::complex<T>& x) + { + return x.real(); + } +}; +#endif + +template<typename Scalar> +struct real_retval +{ + typedef typename NumTraits<Scalar>::Real type; +}; + +/**************************************************************************** +* Implementation of imag * +****************************************************************************/ + +template<typename Scalar, bool IsComplex = NumTraits<Scalar>::IsComplex> +struct imag_default_impl +{ + typedef typename NumTraits<Scalar>::Real RealScalar; + EIGEN_DEVICE_FUNC + static inline RealScalar run(const Scalar&) + { + return RealScalar(0); + } +}; + +template<typename Scalar> +struct imag_default_impl<Scalar,true> +{ + typedef typename NumTraits<Scalar>::Real RealScalar; + EIGEN_DEVICE_FUNC + static inline RealScalar run(const Scalar& x) + { + using std::imag; + return imag(x); + } +}; + +template<typename Scalar> struct imag_impl : imag_default_impl<Scalar> {}; + +#ifdef __CUDA_ARCH__ +template<typename T> +struct imag_impl<std::complex<T> > +{ + typedef T RealScalar; + EIGEN_DEVICE_FUNC + static inline T run(const std::complex<T>& x) + { + return x.imag(); + } +}; +#endif + +template<typename Scalar> +struct imag_retval +{ + typedef typename NumTraits<Scalar>::Real type; +}; + +/**************************************************************************** +* Implementation of real_ref * +****************************************************************************/ + +template<typename Scalar> +struct real_ref_impl +{ + typedef typename NumTraits<Scalar>::Real RealScalar; + EIGEN_DEVICE_FUNC + static inline RealScalar& run(Scalar& x) + { + return reinterpret_cast<RealScalar*>(&x)[0]; + } + EIGEN_DEVICE_FUNC + static inline const RealScalar& run(const Scalar& x) + { + return reinterpret_cast<const RealScalar*>(&x)[0]; + } +}; + +template<typename Scalar> +struct real_ref_retval +{ + typedef typename NumTraits<Scalar>::Real & type; +}; + +/**************************************************************************** +* Implementation of imag_ref * +****************************************************************************/ + +template<typename Scalar, bool IsComplex> +struct imag_ref_default_impl +{ + typedef typename NumTraits<Scalar>::Real RealScalar; + EIGEN_DEVICE_FUNC + static inline RealScalar& run(Scalar& x) + { + return reinterpret_cast<RealScalar*>(&x)[1]; + } + EIGEN_DEVICE_FUNC + static inline const RealScalar& run(const Scalar& x) + { + return reinterpret_cast<RealScalar*>(&x)[1]; + } +}; + +template<typename Scalar> +struct imag_ref_default_impl<Scalar, false> +{ + EIGEN_DEVICE_FUNC + static inline Scalar run(Scalar&) + { + return Scalar(0); + } + EIGEN_DEVICE_FUNC + static inline const Scalar run(const Scalar&) + { + return Scalar(0); + } +}; + +template<typename Scalar> +struct imag_ref_impl : imag_ref_default_impl<Scalar, NumTraits<Scalar>::IsComplex> {}; + +template<typename Scalar> +struct imag_ref_retval +{ + typedef typename NumTraits<Scalar>::Real & type; +}; + +/**************************************************************************** +* Implementation of conj * +****************************************************************************/ + +template<typename Scalar, bool IsComplex = NumTraits<Scalar>::IsComplex> +struct conj_impl +{ + EIGEN_DEVICE_FUNC + static inline Scalar run(const Scalar& x) + { + return x; + } +}; + +template<typename Scalar> +struct conj_impl<Scalar,true> +{ + EIGEN_DEVICE_FUNC + static inline Scalar run(const Scalar& x) + { + using std::conj; + return conj(x); + } +}; + +template<typename Scalar> +struct conj_retval +{ + typedef Scalar type; +}; + +/**************************************************************************** +* Implementation of abs2 * +****************************************************************************/ + +template<typename Scalar,bool IsComplex> +struct abs2_impl_default +{ + typedef typename NumTraits<Scalar>::Real RealScalar; + EIGEN_DEVICE_FUNC + static inline RealScalar run(const Scalar& x) + { + return x*x; + } +}; + +template<typename Scalar> +struct abs2_impl_default<Scalar, true> // IsComplex +{ + typedef typename NumTraits<Scalar>::Real RealScalar; + EIGEN_DEVICE_FUNC + static inline RealScalar run(const Scalar& x) + { + return real(x)*real(x) + imag(x)*imag(x); + } +}; + +template<typename Scalar> +struct abs2_impl +{ + typedef typename NumTraits<Scalar>::Real RealScalar; + EIGEN_DEVICE_FUNC + static inline RealScalar run(const Scalar& x) + { + return abs2_impl_default<Scalar,NumTraits<Scalar>::IsComplex>::run(x); + } +}; + +template<typename Scalar> +struct abs2_retval +{ + typedef typename NumTraits<Scalar>::Real type; +}; + +/**************************************************************************** +* Implementation of norm1 * +****************************************************************************/ + +template<typename Scalar, bool IsComplex> +struct norm1_default_impl +{ + typedef typename NumTraits<Scalar>::Real RealScalar; + EIGEN_DEVICE_FUNC + static inline RealScalar run(const Scalar& x) + { + EIGEN_USING_STD_MATH(abs); + return abs(real(x)) + abs(imag(x)); + } +}; + +template<typename Scalar> +struct norm1_default_impl<Scalar, false> +{ + EIGEN_DEVICE_FUNC + static inline Scalar run(const Scalar& x) + { + EIGEN_USING_STD_MATH(abs); + return abs(x); + } +}; + +template<typename Scalar> +struct norm1_impl : norm1_default_impl<Scalar, NumTraits<Scalar>::IsComplex> {}; + +template<typename Scalar> +struct norm1_retval +{ + typedef typename NumTraits<Scalar>::Real type; +}; + +/**************************************************************************** +* Implementation of hypot * +****************************************************************************/ + +template<typename Scalar> +struct hypot_impl +{ + typedef typename NumTraits<Scalar>::Real RealScalar; + static inline RealScalar run(const Scalar& x, const Scalar& y) + { + EIGEN_USING_STD_MATH(abs); + EIGEN_USING_STD_MATH(sqrt); + RealScalar _x = abs(x); + RealScalar _y = abs(y); + Scalar p, qp; + if(_x>_y) + { + p = _x; + qp = _y / p; + } + else + { + p = _y; + qp = _x / p; + } + if(p==RealScalar(0)) return RealScalar(0); + return p * sqrt(RealScalar(1) + qp*qp); + } +}; + +template<typename Scalar> +struct hypot_retval +{ + typedef typename NumTraits<Scalar>::Real type; +}; + +/**************************************************************************** +* Implementation of cast * +****************************************************************************/ + +template<typename OldType, typename NewType> +struct cast_impl +{ + EIGEN_DEVICE_FUNC + static inline NewType run(const OldType& x) + { + return static_cast<NewType>(x); + } +}; + +// here, for once, we're plainly returning NewType: we don't want cast to do weird things. + +template<typename OldType, typename NewType> +EIGEN_DEVICE_FUNC +inline NewType cast(const OldType& x) +{ + return cast_impl<OldType, NewType>::run(x); +} + +/**************************************************************************** +* Implementation of round * +****************************************************************************/ + +#if EIGEN_HAS_CXX11_MATH + template<typename Scalar> + struct round_impl { + static inline Scalar run(const Scalar& x) + { + EIGEN_STATIC_ASSERT((!NumTraits<Scalar>::IsComplex), NUMERIC_TYPE_MUST_BE_REAL) + using std::round; + return round(x); + } + }; +#else + template<typename Scalar> + struct round_impl + { + static inline Scalar run(const Scalar& x) + { + EIGEN_STATIC_ASSERT((!NumTraits<Scalar>::IsComplex), NUMERIC_TYPE_MUST_BE_REAL) + EIGEN_USING_STD_MATH(floor); + EIGEN_USING_STD_MATH(ceil); + return (x > Scalar(0)) ? floor(x + Scalar(0.5)) : ceil(x - Scalar(0.5)); + } + }; +#endif + +template<typename Scalar> +struct round_retval +{ + typedef Scalar type; +}; + +/**************************************************************************** +* Implementation of arg * +****************************************************************************/ + +#if EIGEN_HAS_CXX11_MATH + template<typename Scalar> + struct arg_impl { + static inline Scalar run(const Scalar& x) + { + EIGEN_USING_STD_MATH(arg); + return arg(x); + } + }; +#else + template<typename Scalar, bool IsComplex = NumTraits<Scalar>::IsComplex> + struct arg_default_impl + { + typedef typename NumTraits<Scalar>::Real RealScalar; + EIGEN_DEVICE_FUNC + static inline RealScalar run(const Scalar& x) + { + return (x < Scalar(0)) ? Scalar(EIGEN_PI) : Scalar(0); } + }; + + template<typename Scalar> + struct arg_default_impl<Scalar,true> + { + typedef typename NumTraits<Scalar>::Real RealScalar; + EIGEN_DEVICE_FUNC + static inline RealScalar run(const Scalar& x) + { + EIGEN_USING_STD_MATH(arg); + return arg(x); + } + }; + + template<typename Scalar> struct arg_impl : arg_default_impl<Scalar> {}; +#endif + +template<typename Scalar> +struct arg_retval +{ + typedef typename NumTraits<Scalar>::Real type; +}; + +/**************************************************************************** +* Implementation of log1p * +****************************************************************************/ + +namespace std_fallback { + // fallback log1p implementation in case there is no log1p(Scalar) function in namespace of Scalar, + // or that there is no suitable std::log1p function available + template<typename Scalar> + EIGEN_DEVICE_FUNC inline Scalar log1p(const Scalar& x) { + EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar) + typedef typename NumTraits<Scalar>::Real RealScalar; + EIGEN_USING_STD_MATH(log); + Scalar x1p = RealScalar(1) + x; + return ( x1p == Scalar(1) ) ? x : x * ( log(x1p) / (x1p - RealScalar(1)) ); + } +} + +template<typename Scalar> +struct log1p_impl { + static inline Scalar run(const Scalar& x) + { + EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar) + #if EIGEN_HAS_CXX11_MATH + using std::log1p; + #endif + using std_fallback::log1p; + return log1p(x); + } +}; + + +template<typename Scalar> +struct log1p_retval +{ + typedef Scalar type; +}; + +/**************************************************************************** +* Implementation of pow * +****************************************************************************/ + +template<typename ScalarX,typename ScalarY, bool IsInteger = NumTraits<ScalarX>::IsInteger&&NumTraits<ScalarY>::IsInteger> +struct pow_impl +{ + //typedef Scalar retval; + typedef typename ScalarBinaryOpTraits<ScalarX,ScalarY,internal::scalar_pow_op<ScalarX,ScalarY> >::ReturnType result_type; + static EIGEN_DEVICE_FUNC inline result_type run(const ScalarX& x, const ScalarY& y) + { + EIGEN_USING_STD_MATH(pow); + return pow(x, y); + } +}; + +template<typename ScalarX,typename ScalarY> +struct pow_impl<ScalarX,ScalarY, true> +{ + typedef ScalarX result_type; + static EIGEN_DEVICE_FUNC inline ScalarX run(ScalarX x, ScalarY y) + { + ScalarX res(1); + eigen_assert(!NumTraits<ScalarY>::IsSigned || y >= 0); + if(y & 1) res *= x; + y >>= 1; + while(y) + { + x *= x; + if(y&1) res *= x; + y >>= 1; + } + return res; + } +}; + +/**************************************************************************** +* Implementation of random * +****************************************************************************/ + +template<typename Scalar, + bool IsComplex, + bool IsInteger> +struct random_default_impl {}; + +template<typename Scalar> +struct random_impl : random_default_impl<Scalar, NumTraits<Scalar>::IsComplex, NumTraits<Scalar>::IsInteger> {}; + +template<typename Scalar> +struct random_retval +{ + typedef Scalar type; +}; + +template<typename Scalar> inline EIGEN_MATHFUNC_RETVAL(random, Scalar) random(const Scalar& x, const Scalar& y); +template<typename Scalar> inline EIGEN_MATHFUNC_RETVAL(random, Scalar) random(); + +template<typename Scalar> +struct random_default_impl<Scalar, false, false> +{ + static inline Scalar run(const Scalar& x, const Scalar& y) + { + return x + (y-x) * Scalar(std::rand()) / Scalar(RAND_MAX); + } + static inline Scalar run() + { + return run(Scalar(NumTraits<Scalar>::IsSigned ? -1 : 0), Scalar(1)); + } +}; + +enum { + meta_floor_log2_terminate, + meta_floor_log2_move_up, + meta_floor_log2_move_down, + meta_floor_log2_bogus +}; + +template<unsigned int n, int lower, int upper> struct meta_floor_log2_selector +{ + enum { middle = (lower + upper) / 2, + value = (upper <= lower + 1) ? int(meta_floor_log2_terminate) + : (n < (1 << middle)) ? int(meta_floor_log2_move_down) + : (n==0) ? int(meta_floor_log2_bogus) + : int(meta_floor_log2_move_up) + }; +}; + +template<unsigned int n, + int lower = 0, + int upper = sizeof(unsigned int) * CHAR_BIT - 1, + int selector = meta_floor_log2_selector<n, lower, upper>::value> +struct meta_floor_log2 {}; + +template<unsigned int n, int lower, int upper> +struct meta_floor_log2<n, lower, upper, meta_floor_log2_move_down> +{ + enum { value = meta_floor_log2<n, lower, meta_floor_log2_selector<n, lower, upper>::middle>::value }; +}; + +template<unsigned int n, int lower, int upper> +struct meta_floor_log2<n, lower, upper, meta_floor_log2_move_up> +{ + enum { value = meta_floor_log2<n, meta_floor_log2_selector<n, lower, upper>::middle, upper>::value }; +}; + +template<unsigned int n, int lower, int upper> +struct meta_floor_log2<n, lower, upper, meta_floor_log2_terminate> +{ + enum { value = (n >= ((unsigned int)(1) << (lower+1))) ? lower+1 : lower }; +}; + +template<unsigned int n, int lower, int upper> +struct meta_floor_log2<n, lower, upper, meta_floor_log2_bogus> +{ + // no value, error at compile time +}; + +template<typename Scalar> +struct random_default_impl<Scalar, false, true> +{ + static inline Scalar run(const Scalar& x, const Scalar& y) + { + typedef typename conditional<NumTraits<Scalar>::IsSigned,std::ptrdiff_t,std::size_t>::type ScalarX; + if(y<x) + return x; + // the following difference might overflow on a 32 bits system, + // but since y>=x the result converted to an unsigned long is still correct. + std::size_t range = ScalarX(y)-ScalarX(x); + std::size_t offset = 0; + // rejection sampling + std::size_t divisor = 1; + std::size_t multiplier = 1; + if(range<RAND_MAX) divisor = (std::size_t(RAND_MAX)+1)/(range+1); + else multiplier = 1 + range/(std::size_t(RAND_MAX)+1); + do { + offset = (std::size_t(std::rand()) * multiplier) / divisor; + } while (offset > range); + return Scalar(ScalarX(x) + offset); + } + + static inline Scalar run() + { +#ifdef EIGEN_MAKING_DOCS + return run(Scalar(NumTraits<Scalar>::IsSigned ? -10 : 0), Scalar(10)); +#else + enum { rand_bits = meta_floor_log2<(unsigned int)(RAND_MAX)+1>::value, + scalar_bits = sizeof(Scalar) * CHAR_BIT, + shift = EIGEN_PLAIN_ENUM_MAX(0, int(rand_bits) - int(scalar_bits)), + offset = NumTraits<Scalar>::IsSigned ? (1 << (EIGEN_PLAIN_ENUM_MIN(rand_bits,scalar_bits)-1)) : 0 + }; + return Scalar((std::rand() >> shift) - offset); +#endif + } +}; + +template<typename Scalar> +struct random_default_impl<Scalar, true, false> +{ + static inline Scalar run(const Scalar& x, const Scalar& y) + { + return Scalar(random(real(x), real(y)), + random(imag(x), imag(y))); + } + static inline Scalar run() + { + typedef typename NumTraits<Scalar>::Real RealScalar; + return Scalar(random<RealScalar>(), random<RealScalar>()); + } +}; + +template<typename Scalar> +inline EIGEN_MATHFUNC_RETVAL(random, Scalar) random(const Scalar& x, const Scalar& y) +{ + return EIGEN_MATHFUNC_IMPL(random, Scalar)::run(x, y); +} + +template<typename Scalar> +inline EIGEN_MATHFUNC_RETVAL(random, Scalar) random() +{ + return EIGEN_MATHFUNC_IMPL(random, Scalar)::run(); +} + +// Implementatin of is* functions + +// std::is* do not work with fast-math and gcc, std::is* are available on MSVC 2013 and newer, as well as in clang. +#if (EIGEN_HAS_CXX11_MATH && !(EIGEN_COMP_GNUC_STRICT && __FINITE_MATH_ONLY__)) || (EIGEN_COMP_MSVC>=1800) || (EIGEN_COMP_CLANG) +#define EIGEN_USE_STD_FPCLASSIFY 1 +#else +#define EIGEN_USE_STD_FPCLASSIFY 0 +#endif + +template<typename T> +EIGEN_DEVICE_FUNC +typename internal::enable_if<internal::is_integral<T>::value,bool>::type +isnan_impl(const T&) { return false; } + +template<typename T> +EIGEN_DEVICE_FUNC +typename internal::enable_if<internal::is_integral<T>::value,bool>::type +isinf_impl(const T&) { return false; } + +template<typename T> +EIGEN_DEVICE_FUNC +typename internal::enable_if<internal::is_integral<T>::value,bool>::type +isfinite_impl(const T&) { return true; } + +template<typename T> +EIGEN_DEVICE_FUNC +typename internal::enable_if<(!internal::is_integral<T>::value)&&(!NumTraits<T>::IsComplex),bool>::type +isfinite_impl(const T& x) +{ + #ifdef __CUDA_ARCH__ + return (::isfinite)(x); + #elif EIGEN_USE_STD_FPCLASSIFY + using std::isfinite; + return isfinite EIGEN_NOT_A_MACRO (x); + #else + return x<=NumTraits<T>::highest() && x>=NumTraits<T>::lowest(); + #endif +} + +template<typename T> +EIGEN_DEVICE_FUNC +typename internal::enable_if<(!internal::is_integral<T>::value)&&(!NumTraits<T>::IsComplex),bool>::type +isinf_impl(const T& x) +{ + #ifdef __CUDA_ARCH__ + return (::isinf)(x); + #elif EIGEN_USE_STD_FPCLASSIFY + using std::isinf; + return isinf EIGEN_NOT_A_MACRO (x); + #else + return x>NumTraits<T>::highest() || x<NumTraits<T>::lowest(); + #endif +} + +template<typename T> +EIGEN_DEVICE_FUNC +typename internal::enable_if<(!internal::is_integral<T>::value)&&(!NumTraits<T>::IsComplex),bool>::type +isnan_impl(const T& x) +{ + #ifdef __CUDA_ARCH__ + return (::isnan)(x); + #elif EIGEN_USE_STD_FPCLASSIFY + using std::isnan; + return isnan EIGEN_NOT_A_MACRO (x); + #else + return x != x; + #endif +} + +#if (!EIGEN_USE_STD_FPCLASSIFY) + +#if EIGEN_COMP_MSVC + +template<typename T> EIGEN_DEVICE_FUNC bool isinf_msvc_helper(T x) +{ + return _fpclass(x)==_FPCLASS_NINF || _fpclass(x)==_FPCLASS_PINF; +} + +//MSVC defines a _isnan builtin function, but for double only +EIGEN_DEVICE_FUNC inline bool isnan_impl(const long double& x) { return _isnan(x)!=0; } +EIGEN_DEVICE_FUNC inline bool isnan_impl(const double& x) { return _isnan(x)!=0; } +EIGEN_DEVICE_FUNC inline bool isnan_impl(const float& x) { return _isnan(x)!=0; } + +EIGEN_DEVICE_FUNC inline bool isinf_impl(const long double& x) { return isinf_msvc_helper(x); } +EIGEN_DEVICE_FUNC inline bool isinf_impl(const double& x) { return isinf_msvc_helper(x); } +EIGEN_DEVICE_FUNC inline bool isinf_impl(const float& x) { return isinf_msvc_helper(x); } + +#elif (defined __FINITE_MATH_ONLY__ && __FINITE_MATH_ONLY__ && EIGEN_COMP_GNUC) + +#if EIGEN_GNUC_AT_LEAST(5,0) + #define EIGEN_TMP_NOOPT_ATTRIB EIGEN_DEVICE_FUNC inline __attribute__((optimize("no-finite-math-only"))) +#else + // NOTE the inline qualifier and noinline attribute are both needed: the former is to avoid linking issue (duplicate symbol), + // while the second prevent too aggressive optimizations in fast-math mode: + #define EIGEN_TMP_NOOPT_ATTRIB EIGEN_DEVICE_FUNC inline __attribute__((noinline,optimize("no-finite-math-only"))) +#endif + +template<> EIGEN_TMP_NOOPT_ATTRIB bool isnan_impl(const long double& x) { return __builtin_isnan(x); } +template<> EIGEN_TMP_NOOPT_ATTRIB bool isnan_impl(const double& x) { return __builtin_isnan(x); } +template<> EIGEN_TMP_NOOPT_ATTRIB bool isnan_impl(const float& x) { return __builtin_isnan(x); } +template<> EIGEN_TMP_NOOPT_ATTRIB bool isinf_impl(const double& x) { return __builtin_isinf(x); } +template<> EIGEN_TMP_NOOPT_ATTRIB bool isinf_impl(const float& x) { return __builtin_isinf(x); } +template<> EIGEN_TMP_NOOPT_ATTRIB bool isinf_impl(const long double& x) { return __builtin_isinf(x); } + +#undef EIGEN_TMP_NOOPT_ATTRIB + +#endif + +#endif + +// The following overload are defined at the end of this file +template<typename T> EIGEN_DEVICE_FUNC bool isfinite_impl(const std::complex<T>& x); +template<typename T> EIGEN_DEVICE_FUNC bool isnan_impl(const std::complex<T>& x); +template<typename T> EIGEN_DEVICE_FUNC bool isinf_impl(const std::complex<T>& x); + +template<typename T> T generic_fast_tanh_float(const T& a_x); + +} // end namespace internal + +/**************************************************************************** +* Generic math functions * +****************************************************************************/ + +namespace numext { + +#ifndef __CUDA_ARCH__ +template<typename T> +EIGEN_DEVICE_FUNC +EIGEN_ALWAYS_INLINE T mini(const T& x, const T& y) +{ + EIGEN_USING_STD_MATH(min); + return min EIGEN_NOT_A_MACRO (x,y); +} + +template<typename T> +EIGEN_DEVICE_FUNC +EIGEN_ALWAYS_INLINE T maxi(const T& x, const T& y) +{ + EIGEN_USING_STD_MATH(max); + return max EIGEN_NOT_A_MACRO (x,y); +} +#else +template<typename T> +EIGEN_DEVICE_FUNC +EIGEN_ALWAYS_INLINE T mini(const T& x, const T& y) +{ + return y < x ? y : x; +} +template<> +EIGEN_DEVICE_FUNC +EIGEN_ALWAYS_INLINE float mini(const float& x, const float& y) +{ + return fminf(x, y); +} +template<typename T> +EIGEN_DEVICE_FUNC +EIGEN_ALWAYS_INLINE T maxi(const T& x, const T& y) +{ + return x < y ? y : x; +} +template<> +EIGEN_DEVICE_FUNC +EIGEN_ALWAYS_INLINE float maxi(const float& x, const float& y) +{ + return fmaxf(x, y); +} +#endif + + +template<typename Scalar> +EIGEN_DEVICE_FUNC +inline EIGEN_MATHFUNC_RETVAL(real, Scalar) real(const Scalar& x) +{ + return EIGEN_MATHFUNC_IMPL(real, Scalar)::run(x); +} + +template<typename Scalar> +EIGEN_DEVICE_FUNC +inline typename internal::add_const_on_value_type< EIGEN_MATHFUNC_RETVAL(real_ref, Scalar) >::type real_ref(const Scalar& x) +{ + return internal::real_ref_impl<Scalar>::run(x); +} + +template<typename Scalar> +EIGEN_DEVICE_FUNC +inline EIGEN_MATHFUNC_RETVAL(real_ref, Scalar) real_ref(Scalar& x) +{ + return EIGEN_MATHFUNC_IMPL(real_ref, Scalar)::run(x); +} + +template<typename Scalar> +EIGEN_DEVICE_FUNC +inline EIGEN_MATHFUNC_RETVAL(imag, Scalar) imag(const Scalar& x) +{ + return EIGEN_MATHFUNC_IMPL(imag, Scalar)::run(x); +} + +template<typename Scalar> +EIGEN_DEVICE_FUNC +inline EIGEN_MATHFUNC_RETVAL(arg, Scalar) arg(const Scalar& x) +{ + return EIGEN_MATHFUNC_IMPL(arg, Scalar)::run(x); +} + +template<typename Scalar> +EIGEN_DEVICE_FUNC +inline typename internal::add_const_on_value_type< EIGEN_MATHFUNC_RETVAL(imag_ref, Scalar) >::type imag_ref(const Scalar& x) +{ + return internal::imag_ref_impl<Scalar>::run(x); +} + +template<typename Scalar> +EIGEN_DEVICE_FUNC +inline EIGEN_MATHFUNC_RETVAL(imag_ref, Scalar) imag_ref(Scalar& x) +{ + return EIGEN_MATHFUNC_IMPL(imag_ref, Scalar)::run(x); +} + +template<typename Scalar> +EIGEN_DEVICE_FUNC +inline EIGEN_MATHFUNC_RETVAL(conj, Scalar) conj(const Scalar& x) +{ + return EIGEN_MATHFUNC_IMPL(conj, Scalar)::run(x); +} + +template<typename Scalar> +EIGEN_DEVICE_FUNC +inline EIGEN_MATHFUNC_RETVAL(abs2, Scalar) abs2(const Scalar& x) +{ + return EIGEN_MATHFUNC_IMPL(abs2, Scalar)::run(x); +} + +template<typename Scalar> +EIGEN_DEVICE_FUNC +inline EIGEN_MATHFUNC_RETVAL(norm1, Scalar) norm1(const Scalar& x) +{ + return EIGEN_MATHFUNC_IMPL(norm1, Scalar)::run(x); +} + +template<typename Scalar> +EIGEN_DEVICE_FUNC +inline EIGEN_MATHFUNC_RETVAL(hypot, Scalar) hypot(const Scalar& x, const Scalar& y) +{ + return EIGEN_MATHFUNC_IMPL(hypot, Scalar)::run(x, y); +} + +template<typename Scalar> +EIGEN_DEVICE_FUNC +inline EIGEN_MATHFUNC_RETVAL(log1p, Scalar) log1p(const Scalar& x) +{ + return EIGEN_MATHFUNC_IMPL(log1p, Scalar)::run(x); +} + +#ifdef __CUDACC__ +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +float log1p(const float &x) { return ::log1pf(x); } + +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +double log1p(const double &x) { return ::log1p(x); } +#endif + +template<typename ScalarX,typename ScalarY> +EIGEN_DEVICE_FUNC +inline typename internal::pow_impl<ScalarX,ScalarY>::result_type pow(const ScalarX& x, const ScalarY& y) +{ + return internal::pow_impl<ScalarX,ScalarY>::run(x, y); +} + +template<typename T> EIGEN_DEVICE_FUNC bool (isnan) (const T &x) { return internal::isnan_impl(x); } +template<typename T> EIGEN_DEVICE_FUNC bool (isinf) (const T &x) { return internal::isinf_impl(x); } +template<typename T> EIGEN_DEVICE_FUNC bool (isfinite)(const T &x) { return internal::isfinite_impl(x); } + +template<typename Scalar> +EIGEN_DEVICE_FUNC +inline EIGEN_MATHFUNC_RETVAL(round, Scalar) round(const Scalar& x) +{ + return EIGEN_MATHFUNC_IMPL(round, Scalar)::run(x); +} + +template<typename T> +EIGEN_DEVICE_FUNC +T (floor)(const T& x) +{ + EIGEN_USING_STD_MATH(floor); + return floor(x); +} + +#ifdef __CUDACC__ +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +float floor(const float &x) { return ::floorf(x); } + +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +double floor(const double &x) { return ::floor(x); } +#endif + +template<typename T> +EIGEN_DEVICE_FUNC +T (ceil)(const T& x) +{ + EIGEN_USING_STD_MATH(ceil); + return ceil(x); +} + +#ifdef __CUDACC__ +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +float ceil(const float &x) { return ::ceilf(x); } + +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +double ceil(const double &x) { return ::ceil(x); } +#endif + + +/** Log base 2 for 32 bits positive integers. + * Conveniently returns 0 for x==0. */ +inline int log2(int x) +{ + eigen_assert(x>=0); + unsigned int v(x); + static const int table[32] = { 0, 9, 1, 10, 13, 21, 2, 29, 11, 14, 16, 18, 22, 25, 3, 30, 8, 12, 20, 28, 15, 17, 24, 7, 19, 27, 23, 6, 26, 5, 4, 31 }; + v |= v >> 1; + v |= v >> 2; + v |= v >> 4; + v |= v >> 8; + v |= v >> 16; + return table[(v * 0x07C4ACDDU) >> 27]; +} + +/** \returns the square root of \a x. + * + * It is essentially equivalent to \code using std::sqrt; return sqrt(x); \endcode, + * but slightly faster for float/double and some compilers (e.g., gcc), thanks to + * specializations when SSE is enabled. + * + * It's usage is justified in performance critical functions, like norm/normalize. + */ +template<typename T> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +T sqrt(const T &x) +{ + EIGEN_USING_STD_MATH(sqrt); + return sqrt(x); +} + +template<typename T> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +T log(const T &x) { + EIGEN_USING_STD_MATH(log); + return log(x); +} + +#ifdef __CUDACC__ +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +float log(const float &x) { return ::logf(x); } + +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +double log(const double &x) { return ::log(x); } +#endif + +template<typename T> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +typename internal::enable_if<NumTraits<T>::IsSigned || NumTraits<T>::IsComplex,typename NumTraits<T>::Real>::type +abs(const T &x) { + EIGEN_USING_STD_MATH(abs); + return abs(x); +} + +template<typename T> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +typename internal::enable_if<!(NumTraits<T>::IsSigned || NumTraits<T>::IsComplex),typename NumTraits<T>::Real>::type +abs(const T &x) { + return x; +} + +#if defined(__SYCL_DEVICE_ONLY__) +EIGEN_ALWAYS_INLINE float abs(float x) { return cl::sycl::fabs(x); } +EIGEN_ALWAYS_INLINE double abs(double x) { return cl::sycl::fabs(x); } +#endif // defined(__SYCL_DEVICE_ONLY__) + +#ifdef __CUDACC__ +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +float abs(const float &x) { return ::fabsf(x); } + +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +double abs(const double &x) { return ::fabs(x); } + +template <> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +float abs(const std::complex<float>& x) { + return ::hypotf(x.real(), x.imag()); +} + +template <> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +double abs(const std::complex<double>& x) { + return ::hypot(x.real(), x.imag()); +} +#endif + +template<typename T> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +T exp(const T &x) { + EIGEN_USING_STD_MATH(exp); + return exp(x); +} + +#ifdef __CUDACC__ +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +float exp(const float &x) { return ::expf(x); } + +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +double exp(const double &x) { return ::exp(x); } +#endif + +template<typename T> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +T cos(const T &x) { + EIGEN_USING_STD_MATH(cos); + return cos(x); +} + +#ifdef __CUDACC__ +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +float cos(const float &x) { return ::cosf(x); } + +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +double cos(const double &x) { return ::cos(x); } +#endif + +template<typename T> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +T sin(const T &x) { + EIGEN_USING_STD_MATH(sin); + return sin(x); +} + +#ifdef __CUDACC__ +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +float sin(const float &x) { return ::sinf(x); } + +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +double sin(const double &x) { return ::sin(x); } +#endif + +template<typename T> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +T tan(const T &x) { + EIGEN_USING_STD_MATH(tan); + return tan(x); +} + +#ifdef __CUDACC__ +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +float tan(const float &x) { return ::tanf(x); } + +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +double tan(const double &x) { return ::tan(x); } +#endif + +template<typename T> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +T acos(const T &x) { + EIGEN_USING_STD_MATH(acos); + return acos(x); +} + +#ifdef __CUDACC__ +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +float acos(const float &x) { return ::acosf(x); } + +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +double acos(const double &x) { return ::acos(x); } +#endif + +template<typename T> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +T asin(const T &x) { + EIGEN_USING_STD_MATH(asin); + return asin(x); +} + +#ifdef __CUDACC__ +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +float asin(const float &x) { return ::asinf(x); } + +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +double asin(const double &x) { return ::asin(x); } +#endif + +template<typename T> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +T atan(const T &x) { + EIGEN_USING_STD_MATH(atan); + return atan(x); +} + +#ifdef __CUDACC__ +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +float atan(const float &x) { return ::atanf(x); } + +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +double atan(const double &x) { return ::atan(x); } +#endif + + +template<typename T> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +T cosh(const T &x) { + EIGEN_USING_STD_MATH(cosh); + return cosh(x); +} + +#ifdef __CUDACC__ +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +float cosh(const float &x) { return ::coshf(x); } + +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +double cosh(const double &x) { return ::cosh(x); } +#endif + +template<typename T> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +T sinh(const T &x) { + EIGEN_USING_STD_MATH(sinh); + return sinh(x); +} + +#ifdef __CUDACC__ +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +float sinh(const float &x) { return ::sinhf(x); } + +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +double sinh(const double &x) { return ::sinh(x); } +#endif + +template<typename T> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +T tanh(const T &x) { + EIGEN_USING_STD_MATH(tanh); + return tanh(x); +} + +#if (!defined(__CUDACC__)) && EIGEN_FAST_MATH +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +float tanh(float x) { return internal::generic_fast_tanh_float(x); } +#endif + +#ifdef __CUDACC__ +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +float tanh(const float &x) { return ::tanhf(x); } + +template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +double tanh(const double &x) { return ::tanh(x); } +#endif + +template <typename T> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +T fmod(const T& a, const T& b) { + EIGEN_USING_STD_MATH(fmod); + return fmod(a, b); +} + +#ifdef __CUDACC__ +template <> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +float fmod(const float& a, const float& b) { + return ::fmodf(a, b); +} + +template <> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +double fmod(const double& a, const double& b) { + return ::fmod(a, b); +} +#endif + +} // end namespace numext + +namespace internal { + +template<typename T> +EIGEN_DEVICE_FUNC bool isfinite_impl(const std::complex<T>& x) +{ + return (numext::isfinite)(numext::real(x)) && (numext::isfinite)(numext::imag(x)); +} + +template<typename T> +EIGEN_DEVICE_FUNC bool isnan_impl(const std::complex<T>& x) +{ + return (numext::isnan)(numext::real(x)) || (numext::isnan)(numext::imag(x)); +} + +template<typename T> +EIGEN_DEVICE_FUNC bool isinf_impl(const std::complex<T>& x) +{ + return ((numext::isinf)(numext::real(x)) || (numext::isinf)(numext::imag(x))) && (!(numext::isnan)(x)); +} + +/**************************************************************************** +* Implementation of fuzzy comparisons * +****************************************************************************/ + +template<typename Scalar, + bool IsComplex, + bool IsInteger> +struct scalar_fuzzy_default_impl {}; + +template<typename Scalar> +struct scalar_fuzzy_default_impl<Scalar, false, false> +{ + typedef typename NumTraits<Scalar>::Real RealScalar; + template<typename OtherScalar> EIGEN_DEVICE_FUNC + static inline bool isMuchSmallerThan(const Scalar& x, const OtherScalar& y, const RealScalar& prec) + { + return numext::abs(x) <= numext::abs(y) * prec; + } + EIGEN_DEVICE_FUNC + static inline bool isApprox(const Scalar& x, const Scalar& y, const RealScalar& prec) + { + return numext::abs(x - y) <= numext::mini(numext::abs(x), numext::abs(y)) * prec; + } + EIGEN_DEVICE_FUNC + static inline bool isApproxOrLessThan(const Scalar& x, const Scalar& y, const RealScalar& prec) + { + return x <= y || isApprox(x, y, prec); + } +}; + +template<typename Scalar> +struct scalar_fuzzy_default_impl<Scalar, false, true> +{ + typedef typename NumTraits<Scalar>::Real RealScalar; + template<typename OtherScalar> EIGEN_DEVICE_FUNC + static inline bool isMuchSmallerThan(const Scalar& x, const Scalar&, const RealScalar&) + { + return x == Scalar(0); + } + EIGEN_DEVICE_FUNC + static inline bool isApprox(const Scalar& x, const Scalar& y, const RealScalar&) + { + return x == y; + } + EIGEN_DEVICE_FUNC + static inline bool isApproxOrLessThan(const Scalar& x, const Scalar& y, const RealScalar&) + { + return x <= y; + } +}; + +template<typename Scalar> +struct scalar_fuzzy_default_impl<Scalar, true, false> +{ + typedef typename NumTraits<Scalar>::Real RealScalar; + template<typename OtherScalar> EIGEN_DEVICE_FUNC + static inline bool isMuchSmallerThan(const Scalar& x, const OtherScalar& y, const RealScalar& prec) + { + return numext::abs2(x) <= numext::abs2(y) * prec * prec; + } + EIGEN_DEVICE_FUNC + static inline bool isApprox(const Scalar& x, const Scalar& y, const RealScalar& prec) + { + return numext::abs2(x - y) <= numext::mini(numext::abs2(x), numext::abs2(y)) * prec * prec; + } +}; + +template<typename Scalar> +struct scalar_fuzzy_impl : scalar_fuzzy_default_impl<Scalar, NumTraits<Scalar>::IsComplex, NumTraits<Scalar>::IsInteger> {}; + +template<typename Scalar, typename OtherScalar> EIGEN_DEVICE_FUNC +inline bool isMuchSmallerThan(const Scalar& x, const OtherScalar& y, + const typename NumTraits<Scalar>::Real &precision = NumTraits<Scalar>::dummy_precision()) +{ + return scalar_fuzzy_impl<Scalar>::template isMuchSmallerThan<OtherScalar>(x, y, precision); +} + +template<typename Scalar> EIGEN_DEVICE_FUNC +inline bool isApprox(const Scalar& x, const Scalar& y, + const typename NumTraits<Scalar>::Real &precision = NumTraits<Scalar>::dummy_precision()) +{ + return scalar_fuzzy_impl<Scalar>::isApprox(x, y, precision); +} + +template<typename Scalar> EIGEN_DEVICE_FUNC +inline bool isApproxOrLessThan(const Scalar& x, const Scalar& y, + const typename NumTraits<Scalar>::Real &precision = NumTraits<Scalar>::dummy_precision()) +{ + return scalar_fuzzy_impl<Scalar>::isApproxOrLessThan(x, y, precision); +} + +/****************************************** +*** The special case of the bool type *** +******************************************/ + +template<> struct random_impl<bool> +{ + static inline bool run() + { + return random<int>(0,1)==0 ? false : true; + } +}; + +template<> struct scalar_fuzzy_impl<bool> +{ + typedef bool RealScalar; + + template<typename OtherScalar> EIGEN_DEVICE_FUNC + static inline bool isMuchSmallerThan(const bool& x, const bool&, const bool&) + { + return !x; + } + + EIGEN_DEVICE_FUNC + static inline bool isApprox(bool x, bool y, bool) + { + return x == y; + } + + EIGEN_DEVICE_FUNC + static inline bool isApproxOrLessThan(const bool& x, const bool& y, const bool&) + { + return (!x) || y; + } + +}; + + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_MATHFUNCTIONS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/MathFunctionsImpl.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/MathFunctionsImpl.h new file mode 100644 index 000000000..3c9ef22fa --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/MathFunctionsImpl.h @@ -0,0 +1,78 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Pedro Gonnet (pedro.gonnet@gmail.com) +// Copyright (C) 2016 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_MATHFUNCTIONSIMPL_H +#define EIGEN_MATHFUNCTIONSIMPL_H + +namespace Eigen { + +namespace internal { + +/** \internal \returns the hyperbolic tan of \a a (coeff-wise) + Doesn't do anything fancy, just a 13/6-degree rational interpolant which + is accurate up to a couple of ulp in the range [-9, 9], outside of which + the tanh(x) = +/-1. + + This implementation works on both scalars and packets. +*/ +template<typename T> +T generic_fast_tanh_float(const T& a_x) +{ + // Clamp the inputs to the range [-9, 9] since anything outside + // this range is +/-1.0f in single-precision. + const T plus_9 = pset1<T>(9.f); + const T minus_9 = pset1<T>(-9.f); + // NOTE GCC prior to 6.3 might improperly optimize this max/min + // step such that if a_x is nan, x will be either 9 or -9, + // and tanh will return 1 or -1 instead of nan. + // This is supposed to be fixed in gcc6.3, + // see: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72867 + const T x = pmax(minus_9,pmin(plus_9,a_x)); + // The monomial coefficients of the numerator polynomial (odd). + const T alpha_1 = pset1<T>(4.89352455891786e-03f); + const T alpha_3 = pset1<T>(6.37261928875436e-04f); + const T alpha_5 = pset1<T>(1.48572235717979e-05f); + const T alpha_7 = pset1<T>(5.12229709037114e-08f); + const T alpha_9 = pset1<T>(-8.60467152213735e-11f); + const T alpha_11 = pset1<T>(2.00018790482477e-13f); + const T alpha_13 = pset1<T>(-2.76076847742355e-16f); + + // The monomial coefficients of the denominator polynomial (even). + const T beta_0 = pset1<T>(4.89352518554385e-03f); + const T beta_2 = pset1<T>(2.26843463243900e-03f); + const T beta_4 = pset1<T>(1.18534705686654e-04f); + const T beta_6 = pset1<T>(1.19825839466702e-06f); + + // Since the polynomials are odd/even, we need x^2. + const T x2 = pmul(x, x); + + // Evaluate the numerator polynomial p. + T p = pmadd(x2, alpha_13, alpha_11); + p = pmadd(x2, p, alpha_9); + p = pmadd(x2, p, alpha_7); + p = pmadd(x2, p, alpha_5); + p = pmadd(x2, p, alpha_3); + p = pmadd(x2, p, alpha_1); + p = pmul(x, p); + + // Evaluate the denominator polynomial p. + T q = pmadd(x2, beta_6, beta_4); + q = pmadd(x2, q, beta_2); + q = pmadd(x2, q, beta_0); + + // Divide the numerator by the denominator. + return pdiv(p, q); +} + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_MATHFUNCTIONSIMPL_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/Matrix.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Matrix.h new file mode 100644 index 000000000..90c336d8c --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Matrix.h @@ -0,0 +1,461 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2006-2010 Benoit Jacob <jacob.benoit.1@gmail.com> +// Copyright (C) 2008-2009 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_MATRIX_H +#define EIGEN_MATRIX_H + +namespace Eigen { + +namespace internal { +template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols> +struct traits<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> > +{ +private: + enum { size = internal::size_at_compile_time<_Rows,_Cols>::ret }; + typedef typename find_best_packet<_Scalar,size>::type PacketScalar; + enum { + row_major_bit = _Options&RowMajor ? RowMajorBit : 0, + is_dynamic_size_storage = _MaxRows==Dynamic || _MaxCols==Dynamic, + max_size = is_dynamic_size_storage ? Dynamic : _MaxRows*_MaxCols, + default_alignment = compute_default_alignment<_Scalar,max_size>::value, + actual_alignment = ((_Options&DontAlign)==0) ? default_alignment : 0, + required_alignment = unpacket_traits<PacketScalar>::alignment, + packet_access_bit = (packet_traits<_Scalar>::Vectorizable && (EIGEN_UNALIGNED_VECTORIZE || (actual_alignment>=required_alignment))) ? PacketAccessBit : 0 + }; + +public: + typedef _Scalar Scalar; + typedef Dense StorageKind; + typedef Eigen::Index StorageIndex; + typedef MatrixXpr XprKind; + enum { + RowsAtCompileTime = _Rows, + ColsAtCompileTime = _Cols, + MaxRowsAtCompileTime = _MaxRows, + MaxColsAtCompileTime = _MaxCols, + Flags = compute_matrix_flags<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>::ret, + Options = _Options, + InnerStrideAtCompileTime = 1, + OuterStrideAtCompileTime = (Options&RowMajor) ? ColsAtCompileTime : RowsAtCompileTime, + + // FIXME, the following flag in only used to define NeedsToAlign in PlainObjectBase + EvaluatorFlags = LinearAccessBit | DirectAccessBit | packet_access_bit | row_major_bit, + Alignment = actual_alignment + }; +}; +} + +/** \class Matrix + * \ingroup Core_Module + * + * \brief The matrix class, also used for vectors and row-vectors + * + * The %Matrix class is the work-horse for all \em dense (\ref dense "note") matrices and vectors within Eigen. + * Vectors are matrices with one column, and row-vectors are matrices with one row. + * + * The %Matrix class encompasses \em both fixed-size and dynamic-size objects (\ref fixedsize "note"). + * + * The first three template parameters are required: + * \tparam _Scalar Numeric type, e.g. float, double, int or std::complex<float>. + * User defined scalar types are supported as well (see \ref user_defined_scalars "here"). + * \tparam _Rows Number of rows, or \b Dynamic + * \tparam _Cols Number of columns, or \b Dynamic + * + * The remaining template parameters are optional -- in most cases you don't have to worry about them. + * \tparam _Options A combination of either \b #RowMajor or \b #ColMajor, and of either + * \b #AutoAlign or \b #DontAlign. + * The former controls \ref TopicStorageOrders "storage order", and defaults to column-major. The latter controls alignment, which is required + * for vectorization. It defaults to aligning matrices except for fixed sizes that aren't a multiple of the packet size. + * \tparam _MaxRows Maximum number of rows. Defaults to \a _Rows (\ref maxrows "note"). + * \tparam _MaxCols Maximum number of columns. Defaults to \a _Cols (\ref maxrows "note"). + * + * Eigen provides a number of typedefs covering the usual cases. Here are some examples: + * + * \li \c Matrix2d is a 2x2 square matrix of doubles (\c Matrix<double, 2, 2>) + * \li \c Vector4f is a vector of 4 floats (\c Matrix<float, 4, 1>) + * \li \c RowVector3i is a row-vector of 3 ints (\c Matrix<int, 1, 3>) + * + * \li \c MatrixXf is a dynamic-size matrix of floats (\c Matrix<float, Dynamic, Dynamic>) + * \li \c VectorXf is a dynamic-size vector of floats (\c Matrix<float, Dynamic, 1>) + * + * \li \c Matrix2Xf is a partially fixed-size (dynamic-size) matrix of floats (\c Matrix<float, 2, Dynamic>) + * \li \c MatrixX3d is a partially dynamic-size (fixed-size) matrix of double (\c Matrix<double, Dynamic, 3>) + * + * See \link matrixtypedefs this page \endlink for a complete list of predefined \em %Matrix and \em Vector typedefs. + * + * You can access elements of vectors and matrices using normal subscripting: + * + * \code + * Eigen::VectorXd v(10); + * v[0] = 0.1; + * v[1] = 0.2; + * v(0) = 0.3; + * v(1) = 0.4; + * + * Eigen::MatrixXi m(10, 10); + * m(0, 1) = 1; + * m(0, 2) = 2; + * m(0, 3) = 3; + * \endcode + * + * This class can be extended with the help of the plugin mechanism described on the page + * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_MATRIX_PLUGIN. + * + * <i><b>Some notes:</b></i> + * + * <dl> + * <dt><b>\anchor dense Dense versus sparse:</b></dt> + * <dd>This %Matrix class handles dense, not sparse matrices and vectors. For sparse matrices and vectors, see the Sparse module. + * + * Dense matrices and vectors are plain usual arrays of coefficients. All the coefficients are stored, in an ordinary contiguous array. + * This is unlike Sparse matrices and vectors where the coefficients are stored as a list of nonzero coefficients.</dd> + * + * <dt><b>\anchor fixedsize Fixed-size versus dynamic-size:</b></dt> + * <dd>Fixed-size means that the numbers of rows and columns are known are compile-time. In this case, Eigen allocates the array + * of coefficients as a fixed-size array, as a class member. This makes sense for very small matrices, typically up to 4x4, sometimes up + * to 16x16. Larger matrices should be declared as dynamic-size even if one happens to know their size at compile-time. + * + * Dynamic-size means that the numbers of rows or columns are not necessarily known at compile-time. In this case they are runtime + * variables, and the array of coefficients is allocated dynamically on the heap. + * + * Note that \em dense matrices, be they Fixed-size or Dynamic-size, <em>do not</em> expand dynamically in the sense of a std::map. + * If you want this behavior, see the Sparse module.</dd> + * + * <dt><b>\anchor maxrows _MaxRows and _MaxCols:</b></dt> + * <dd>In most cases, one just leaves these parameters to the default values. + * These parameters mean the maximum size of rows and columns that the matrix may have. They are useful in cases + * when the exact numbers of rows and columns are not known are compile-time, but it is known at compile-time that they cannot + * exceed a certain value. This happens when taking dynamic-size blocks inside fixed-size matrices: in this case _MaxRows and _MaxCols + * are the dimensions of the original matrix, while _Rows and _Cols are Dynamic.</dd> + * </dl> + * + * <i><b>ABI and storage layout</b></i> + * + * The table below summarizes the ABI of some possible Matrix instances which is fixed thorough the lifetime of Eigen 3. + * <table class="manual"> + * <tr><th>Matrix type</th><th>Equivalent C structure</th></tr> + * <tr><td>\code Matrix<T,Dynamic,Dynamic> \endcode</td><td>\code + * struct { + * T *data; // with (size_t(data)%EIGEN_MAX_ALIGN_BYTES)==0 + * Eigen::Index rows, cols; + * }; + * \endcode</td></tr> + * <tr class="alt"><td>\code + * Matrix<T,Dynamic,1> + * Matrix<T,1,Dynamic> \endcode</td><td>\code + * struct { + * T *data; // with (size_t(data)%EIGEN_MAX_ALIGN_BYTES)==0 + * Eigen::Index size; + * }; + * \endcode</td></tr> + * <tr><td>\code Matrix<T,Rows,Cols> \endcode</td><td>\code + * struct { + * T data[Rows*Cols]; // with (size_t(data)%A(Rows*Cols*sizeof(T)))==0 + * }; + * \endcode</td></tr> + * <tr class="alt"><td>\code Matrix<T,Dynamic,Dynamic,0,MaxRows,MaxCols> \endcode</td><td>\code + * struct { + * T data[MaxRows*MaxCols]; // with (size_t(data)%A(MaxRows*MaxCols*sizeof(T)))==0 + * Eigen::Index rows, cols; + * }; + * \endcode</td></tr> + * </table> + * Note that in this table Rows, Cols, MaxRows and MaxCols are all positive integers. A(S) is defined to the largest possible power-of-two + * smaller to EIGEN_MAX_STATIC_ALIGN_BYTES. + * + * \see MatrixBase for the majority of the API methods for matrices, \ref TopicClassHierarchy, + * \ref TopicStorageOrders + */ + +template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols> +class Matrix + : public PlainObjectBase<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> > +{ + public: + + /** \brief Base class typedef. + * \sa PlainObjectBase + */ + typedef PlainObjectBase<Matrix> Base; + + enum { Options = _Options }; + + EIGEN_DENSE_PUBLIC_INTERFACE(Matrix) + + typedef typename Base::PlainObject PlainObject; + + using Base::base; + using Base::coeffRef; + + /** + * \brief Assigns matrices to each other. + * + * \note This is a special case of the templated operator=. Its purpose is + * to prevent a default operator= from hiding the templated operator=. + * + * \callgraph + */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Matrix& operator=(const Matrix& other) + { + return Base::_set(other); + } + + /** \internal + * \brief Copies the value of the expression \a other into \c *this with automatic resizing. + * + * *this might be resized to match the dimensions of \a other. If *this was a null matrix (not already initialized), + * it will be initialized. + * + * Note that copying a row-vector into a vector (and conversely) is allowed. + * The resizing, if any, is then done in the appropriate way so that row-vectors + * remain row-vectors and vectors remain vectors. + */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Matrix& operator=(const DenseBase<OtherDerived>& other) + { + return Base::_set(other); + } + + /* Here, doxygen failed to copy the brief information when using \copydoc */ + + /** + * \brief Copies the generic expression \a other into *this. + * \copydetails DenseBase::operator=(const EigenBase<OtherDerived> &other) + */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Matrix& operator=(const EigenBase<OtherDerived> &other) + { + return Base::operator=(other); + } + + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Matrix& operator=(const ReturnByValue<OtherDerived>& func) + { + return Base::operator=(func); + } + + /** \brief Default constructor. + * + * For fixed-size matrices, does nothing. + * + * For dynamic-size matrices, creates an empty matrix of size 0. Does not allocate any array. Such a matrix + * is called a null matrix. This constructor is the unique way to create null matrices: resizing + * a matrix to 0 is not supported. + * + * \sa resize(Index,Index) + */ + 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 + 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 + 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 + Matrix& operator=(Matrix&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_assignable<Scalar>::value) + { + other.swap(*this); + return *this; + } +#endif + + #ifndef EIGEN_PARSED_BY_DOXYGEN + + // This constructor is for both 1x1 matrices and dynamic vectors + template<typename T> + 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) + { + Base::_check_template_params(); + Base::template _init2<T0,T1>(x, y); + } + #else + /** \brief Constructs a fixed-sized matrix initialized with coefficients starting at \a data */ + EIGEN_DEVICE_FUNC + explicit Matrix(const Scalar *data); + + /** \brief Constructs a vector or row-vector with given dimension. \only_for_vectors + * + * This is useful for dynamic-size vectors. For fixed-size vectors, + * it is redundant to pass these parameters, so one should use the default constructor + * Matrix() instead. + * + * \warning This constructor is disabled for fixed-size \c 1x1 matrices. For instance, + * calling Matrix<double,1,1>(1) will call the initialization constructor: Matrix(const Scalar&). + * For fixed-size \c 1x1 matrices it is therefore recommended to use the default + * constructor Matrix() instead, especially when using one of the non standard + * \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 */ + Matrix(const Scalar& x); + /** \brief Constructs an uninitialized matrix with \a rows rows and \a cols columns. + * + * This is useful for dynamic-size matrices. For fixed-size matrices, + * it is redundant to pass these parameters, so one should use the default constructor + * Matrix() instead. + * + * \warning This constructor is disabled for fixed-size \c 1x2 and \c 2x1 vectors. For instance, + * calling Matrix2f(2,1) will call the initialization constructor: Matrix(const Scalar& x, const Scalar& y). + * For fixed-size \c 1x2 or \c 2x1 vectors it is therefore recommended to use the default + * constructor Matrix() instead, especially when using one of the non standard + * \c EIGEN_INITIALIZE_MATRICES_BY_{ZERO,\c NAN} macros (see \ref TopicPreprocessorDirectives). + */ + EIGEN_DEVICE_FUNC + Matrix(Index rows, Index cols); + + /** \brief Constructs an initialized 2D vector with given coefficients */ + Matrix(const Scalar& x, const Scalar& y); + #endif + + /** \brief Constructs an initialized 3D vector with given coefficients */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Matrix(const Scalar& x, const Scalar& y, const Scalar& z) + { + Base::_check_template_params(); + EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(Matrix, 3) + m_storage.data()[0] = x; + m_storage.data()[1] = y; + m_storage.data()[2] = z; + } + /** \brief Constructs an initialized 4D vector with given coefficients */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Matrix(const Scalar& x, const Scalar& y, const Scalar& z, const Scalar& w) + { + Base::_check_template_params(); + EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(Matrix, 4) + m_storage.data()[0] = x; + m_storage.data()[1] = y; + m_storage.data()[2] = z; + m_storage.data()[3] = w; + } + + + /** \brief Copy constructor */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Matrix(const Matrix& other) : Base(other) + { } + + /** \brief Copy constructor for generic expressions. + * \sa MatrixBase::operator=(const EigenBase<OtherDerived>&) + */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Matrix(const EigenBase<OtherDerived> &other) + : Base(other.derived()) + { } + + EIGEN_DEVICE_FUNC inline Index innerStride() const { return 1; } + EIGEN_DEVICE_FUNC inline Index outerStride() const { return this->innerSize(); } + + /////////// Geometry module /////////// + + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + explicit Matrix(const RotationBase<OtherDerived,ColsAtCompileTime>& r); + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + Matrix& operator=(const RotationBase<OtherDerived,ColsAtCompileTime>& r); + + // allow to extend Matrix outside Eigen + #ifdef EIGEN_MATRIX_PLUGIN + #include EIGEN_MATRIX_PLUGIN + #endif + + protected: + template <typename Derived, typename OtherDerived, bool IsVector> + friend struct internal::conservative_resize_like_impl; + + using Base::m_storage; +}; + +/** \defgroup matrixtypedefs Global matrix typedefs + * + * \ingroup Core_Module + * + * Eigen defines several typedef shortcuts for most common matrix and vector types. + * + * The general patterns are the following: + * + * \c MatrixSizeType where \c Size can be \c 2,\c 3,\c 4 for fixed size square matrices or \c X for dynamic size, + * and where \c Type can be \c i for integer, \c f for float, \c d for double, \c cf for complex float, \c cd + * for complex double. + * + * For example, \c Matrix3d is a fixed-size 3x3 matrix type of doubles, and \c MatrixXf is a dynamic-size matrix of floats. + * + * 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. + * + * \sa class Matrix + */ + +#define EIGEN_MAKE_TYPEDEFS(Type, TypeSuffix, Size, SizeSuffix) \ +/** \ingroup matrixtypedefs */ \ +typedef Matrix<Type, Size, Size> Matrix##SizeSuffix##TypeSuffix; \ +/** \ingroup matrixtypedefs */ \ +typedef Matrix<Type, Size, 1> Vector##SizeSuffix##TypeSuffix; \ +/** \ingroup matrixtypedefs */ \ +typedef Matrix<Type, 1, Size> RowVector##SizeSuffix##TypeSuffix; + +#define EIGEN_MAKE_FIXED_TYPEDEFS(Type, TypeSuffix, Size) \ +/** \ingroup matrixtypedefs */ \ +typedef Matrix<Type, Size, Dynamic> Matrix##Size##X##TypeSuffix; \ +/** \ingroup matrixtypedefs */ \ +typedef Matrix<Type, Dynamic, Size> Matrix##X##Size##TypeSuffix; + +#define EIGEN_MAKE_TYPEDEFS_ALL_SIZES(Type, TypeSuffix) \ +EIGEN_MAKE_TYPEDEFS(Type, TypeSuffix, 2, 2) \ +EIGEN_MAKE_TYPEDEFS(Type, TypeSuffix, 3, 3) \ +EIGEN_MAKE_TYPEDEFS(Type, TypeSuffix, 4, 4) \ +EIGEN_MAKE_TYPEDEFS(Type, TypeSuffix, Dynamic, X) \ +EIGEN_MAKE_FIXED_TYPEDEFS(Type, TypeSuffix, 2) \ +EIGEN_MAKE_FIXED_TYPEDEFS(Type, TypeSuffix, 3) \ +EIGEN_MAKE_FIXED_TYPEDEFS(Type, TypeSuffix, 4) + +EIGEN_MAKE_TYPEDEFS_ALL_SIZES(int, i) +EIGEN_MAKE_TYPEDEFS_ALL_SIZES(float, f) +EIGEN_MAKE_TYPEDEFS_ALL_SIZES(double, d) +EIGEN_MAKE_TYPEDEFS_ALL_SIZES(std::complex<float>, cf) +EIGEN_MAKE_TYPEDEFS_ALL_SIZES(std::complex<double>, cd) + +#undef EIGEN_MAKE_TYPEDEFS_ALL_SIZES +#undef EIGEN_MAKE_TYPEDEFS +#undef EIGEN_MAKE_FIXED_TYPEDEFS + +} // end namespace Eigen + +#endif // EIGEN_MATRIX_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/MatrixBase.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/MatrixBase.h new file mode 100644 index 000000000..ce412180a --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/MatrixBase.h @@ -0,0 +1,530 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2006-2009 Benoit Jacob <jacob.benoit.1@gmail.com> +// Copyright (C) 2008 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_MATRIXBASE_H +#define EIGEN_MATRIXBASE_H + +namespace Eigen { + +/** \class MatrixBase + * \ingroup Core_Module + * + * \brief Base class for all dense matrices, vectors, and expressions + * + * This class is the base that is inherited by all matrix, vector, and related expression + * types. Most of the Eigen API is contained in this class, and its base classes. Other important + * classes for the Eigen API are Matrix, and VectorwiseOp. + * + * Note that some methods are defined in other modules such as the \ref LU_Module LU module + * for all functions related to matrix inversions. + * + * \tparam Derived is the derived type, e.g. a matrix type, or an expression, etc. + * + * When writing a function taking Eigen objects as argument, if you want your function + * to take as argument any matrix, vector, or expression, just let it take a + * MatrixBase argument. As an example, here is a function printFirstRow which, given + * a matrix, vector, or expression \a x, prints the first row of \a x. + * + * \code + template<typename Derived> + void printFirstRow(const Eigen::MatrixBase<Derived>& x) + { + cout << x.row(0) << endl; + } + * \endcode + * + * This class can be extended with the help of the plugin mechanism described on the page + * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_MATRIXBASE_PLUGIN. + * + * \sa \blank \ref TopicClassHierarchy + */ +template<typename Derived> class MatrixBase + : public DenseBase<Derived> +{ + public: +#ifndef EIGEN_PARSED_BY_DOXYGEN + typedef MatrixBase StorageBaseType; + typedef typename internal::traits<Derived>::StorageKind StorageKind; + typedef typename internal::traits<Derived>::StorageIndex StorageIndex; + typedef typename internal::traits<Derived>::Scalar Scalar; + typedef typename internal::packet_traits<Scalar>::type PacketScalar; + typedef typename NumTraits<Scalar>::Real RealScalar; + + typedef DenseBase<Derived> Base; + using Base::RowsAtCompileTime; + using Base::ColsAtCompileTime; + using Base::SizeAtCompileTime; + using Base::MaxRowsAtCompileTime; + using Base::MaxColsAtCompileTime; + using Base::MaxSizeAtCompileTime; + using Base::IsVectorAtCompileTime; + using Base::Flags; + + using Base::derived; + using Base::const_cast_derived; + using Base::rows; + using Base::cols; + using Base::size; + using Base::coeff; + using Base::coeffRef; + using Base::lazyAssign; + using Base::eval; + using Base::operator+=; + using Base::operator-=; + using Base::operator*=; + using Base::operator/=; + + typedef typename Base::CoeffReturnType CoeffReturnType; + typedef typename Base::ConstTransposeReturnType ConstTransposeReturnType; + typedef typename Base::RowXpr RowXpr; + typedef typename Base::ColXpr ColXpr; +#endif // not EIGEN_PARSED_BY_DOXYGEN + + + +#ifndef EIGEN_PARSED_BY_DOXYGEN + /** type of the equivalent square matrix */ + typedef Matrix<Scalar,EIGEN_SIZE_MAX(RowsAtCompileTime,ColsAtCompileTime), + EIGEN_SIZE_MAX(RowsAtCompileTime,ColsAtCompileTime)> SquareMatrixType; +#endif // not EIGEN_PARSED_BY_DOXYGEN + + /** \returns the size of the main diagonal, which is min(rows(),cols()). + * \sa rows(), cols(), SizeAtCompileTime. */ + EIGEN_DEVICE_FUNC + inline Index diagonalSize() const { return (numext::mini)(rows(),cols()); } + + typedef typename Base::PlainObject PlainObject; + +#ifndef EIGEN_PARSED_BY_DOXYGEN + /** \internal Represents a matrix with all coefficients equal to one another*/ + typedef CwiseNullaryOp<internal::scalar_constant_op<Scalar>,PlainObject> ConstantReturnType; + /** \internal the return type of MatrixBase::adjoint() */ + typedef typename internal::conditional<NumTraits<Scalar>::IsComplex, + CwiseUnaryOp<internal::scalar_conjugate_op<Scalar>, ConstTransposeReturnType>, + ConstTransposeReturnType + >::type AdjointReturnType; + /** \internal Return type of eigenvalues() */ + typedef Matrix<std::complex<RealScalar>, internal::traits<Derived>::ColsAtCompileTime, 1, ColMajor> EigenvaluesReturnType; + /** \internal the return type of identity */ + typedef CwiseNullaryOp<internal::scalar_identity_op<Scalar>,PlainObject> IdentityReturnType; + /** \internal the return type of unit vectors */ + typedef Block<const CwiseNullaryOp<internal::scalar_identity_op<Scalar>, SquareMatrixType>, + internal::traits<Derived>::RowsAtCompileTime, + internal::traits<Derived>::ColsAtCompileTime> BasisReturnType; +#endif // not EIGEN_PARSED_BY_DOXYGEN + +#define EIGEN_CURRENT_STORAGE_BASE_CLASS Eigen::MatrixBase +#define EIGEN_DOC_UNARY_ADDONS(X,Y) +# include "../plugins/CommonCwiseUnaryOps.h" +# include "../plugins/CommonCwiseBinaryOps.h" +# include "../plugins/MatrixCwiseUnaryOps.h" +# include "../plugins/MatrixCwiseBinaryOps.h" +# ifdef EIGEN_MATRIXBASE_PLUGIN +# include EIGEN_MATRIXBASE_PLUGIN +# endif +#undef EIGEN_CURRENT_STORAGE_BASE_CLASS +#undef EIGEN_DOC_UNARY_ADDONS + + /** Special case of the template operator=, in order to prevent the compiler + * from generating a default operator= (issue hit with g++ 4.1) + */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Derived& operator=(const MatrixBase& other); + + // We cannot inherit here via Base::operator= since it is causing + // trouble with MSVC. + + template <typename OtherDerived> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Derived& operator=(const DenseBase<OtherDerived>& other); + + template <typename OtherDerived> + EIGEN_DEVICE_FUNC + Derived& operator=(const EigenBase<OtherDerived>& other); + + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + Derived& operator=(const ReturnByValue<OtherDerived>& other); + + template<typename OtherDerived> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Derived& operator+=(const MatrixBase<OtherDerived>& other); + template<typename OtherDerived> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Derived& operator-=(const MatrixBase<OtherDerived>& other); + +#ifdef __CUDACC__ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + const Product<Derived,OtherDerived,LazyProduct> + operator*(const MatrixBase<OtherDerived> &other) const + { return this->lazyProduct(other); } +#else + + template<typename OtherDerived> + const Product<Derived,OtherDerived> + operator*(const MatrixBase<OtherDerived> &other) const; + +#endif + + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + const Product<Derived,OtherDerived,LazyProduct> + lazyProduct(const MatrixBase<OtherDerived> &other) const; + + template<typename OtherDerived> + Derived& operator*=(const EigenBase<OtherDerived>& other); + + template<typename OtherDerived> + void applyOnTheLeft(const EigenBase<OtherDerived>& other); + + template<typename OtherDerived> + void applyOnTheRight(const EigenBase<OtherDerived>& other); + + template<typename DiagonalDerived> + EIGEN_DEVICE_FUNC + const Product<Derived, DiagonalDerived, LazyProduct> + operator*(const DiagonalBase<DiagonalDerived> &diagonal) const; + + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + typename ScalarBinaryOpTraits<typename internal::traits<Derived>::Scalar,typename internal::traits<OtherDerived>::Scalar>::ReturnType + dot(const MatrixBase<OtherDerived>& other) const; + + EIGEN_DEVICE_FUNC RealScalar squaredNorm() const; + EIGEN_DEVICE_FUNC RealScalar norm() const; + RealScalar stableNorm() const; + RealScalar blueNorm() const; + RealScalar hypotNorm() const; + EIGEN_DEVICE_FUNC const PlainObject normalized() const; + EIGEN_DEVICE_FUNC const PlainObject stableNormalized() const; + EIGEN_DEVICE_FUNC void normalize(); + EIGEN_DEVICE_FUNC void stableNormalize(); + + EIGEN_DEVICE_FUNC const AdjointReturnType adjoint() const; + EIGEN_DEVICE_FUNC void adjointInPlace(); + + typedef Diagonal<Derived> DiagonalReturnType; + EIGEN_DEVICE_FUNC + DiagonalReturnType diagonal(); + + typedef typename internal::add_const<Diagonal<const Derived> >::type ConstDiagonalReturnType; + EIGEN_DEVICE_FUNC + ConstDiagonalReturnType diagonal() const; + + template<int Index> struct DiagonalIndexReturnType { typedef Diagonal<Derived,Index> Type; }; + template<int Index> struct ConstDiagonalIndexReturnType { typedef const Diagonal<const Derived,Index> Type; }; + + template<int Index> + EIGEN_DEVICE_FUNC + typename DiagonalIndexReturnType<Index>::Type diagonal(); + + template<int Index> + EIGEN_DEVICE_FUNC + typename ConstDiagonalIndexReturnType<Index>::Type diagonal() const; + + typedef Diagonal<Derived,DynamicIndex> DiagonalDynamicIndexReturnType; + typedef typename internal::add_const<Diagonal<const Derived,DynamicIndex> >::type ConstDiagonalDynamicIndexReturnType; + + EIGEN_DEVICE_FUNC + DiagonalDynamicIndexReturnType diagonal(Index index); + EIGEN_DEVICE_FUNC + ConstDiagonalDynamicIndexReturnType diagonal(Index index) const; + + template<unsigned int Mode> struct TriangularViewReturnType { typedef TriangularView<Derived, Mode> Type; }; + template<unsigned int Mode> struct ConstTriangularViewReturnType { typedef const TriangularView<const Derived, Mode> Type; }; + + template<unsigned int Mode> + EIGEN_DEVICE_FUNC + typename TriangularViewReturnType<Mode>::Type triangularView(); + template<unsigned int Mode> + EIGEN_DEVICE_FUNC + typename ConstTriangularViewReturnType<Mode>::Type triangularView() const; + + template<unsigned int UpLo> struct SelfAdjointViewReturnType { typedef SelfAdjointView<Derived, UpLo> Type; }; + template<unsigned int UpLo> struct ConstSelfAdjointViewReturnType { typedef const SelfAdjointView<const Derived, UpLo> Type; }; + + template<unsigned int UpLo> + EIGEN_DEVICE_FUNC + typename SelfAdjointViewReturnType<UpLo>::Type selfadjointView(); + template<unsigned int UpLo> + EIGEN_DEVICE_FUNC + typename ConstSelfAdjointViewReturnType<UpLo>::Type selfadjointView() const; + + const SparseView<Derived> sparseView(const Scalar& m_reference = Scalar(0), + const typename NumTraits<Scalar>::Real& m_epsilon = NumTraits<Scalar>::dummy_precision()) const; + EIGEN_DEVICE_FUNC static const IdentityReturnType Identity(); + EIGEN_DEVICE_FUNC static const IdentityReturnType Identity(Index rows, Index cols); + EIGEN_DEVICE_FUNC static const BasisReturnType Unit(Index size, Index i); + EIGEN_DEVICE_FUNC static const BasisReturnType Unit(Index i); + EIGEN_DEVICE_FUNC static const BasisReturnType UnitX(); + EIGEN_DEVICE_FUNC static const BasisReturnType UnitY(); + EIGEN_DEVICE_FUNC static const BasisReturnType UnitZ(); + EIGEN_DEVICE_FUNC static const BasisReturnType UnitW(); + + EIGEN_DEVICE_FUNC + const DiagonalWrapper<const Derived> asDiagonal() const; + const PermutationWrapper<const Derived> asPermutation() const; + + EIGEN_DEVICE_FUNC + Derived& setIdentity(); + EIGEN_DEVICE_FUNC + Derived& setIdentity(Index rows, Index cols); + + bool isIdentity(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const; + bool isDiagonal(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const; + + bool isUpperTriangular(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const; + bool isLowerTriangular(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const; + + template<typename OtherDerived> + bool isOrthogonal(const MatrixBase<OtherDerived>& other, + const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const; + bool isUnitary(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const; + + /** \returns true if each coefficients of \c *this and \a other are all exactly equal. + * \warning When using floating point scalar values you probably should rather use a + * fuzzy comparison such as isApprox() + * \sa isApprox(), operator!= */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC inline bool operator==(const MatrixBase<OtherDerived>& other) const + { return cwiseEqual(other).all(); } + + /** \returns true if at least one pair of coefficients of \c *this and \a other are not exactly equal to each other. + * \warning When using floating point scalar values you probably should rather use a + * fuzzy comparison such as isApprox() + * \sa isApprox(), operator== */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC inline bool operator!=(const MatrixBase<OtherDerived>& other) const + { return cwiseNotEqual(other).any(); } + + NoAlias<Derived,Eigen::MatrixBase > noalias(); + + // TODO forceAlignedAccess is temporarily disabled + // Need to find a nicer workaround. + inline const Derived& forceAlignedAccess() const { return derived(); } + inline Derived& forceAlignedAccess() { return derived(); } + template<bool Enable> inline const Derived& forceAlignedAccessIf() const { return derived(); } + template<bool Enable> inline Derived& forceAlignedAccessIf() { return derived(); } + + EIGEN_DEVICE_FUNC Scalar trace() const; + + template<int p> EIGEN_DEVICE_FUNC RealScalar lpNorm() const; + + EIGEN_DEVICE_FUNC MatrixBase<Derived>& matrix() { return *this; } + EIGEN_DEVICE_FUNC const MatrixBase<Derived>& matrix() const { return *this; } + + /** \returns an \link Eigen::ArrayBase Array \endlink expression of this matrix + * \sa ArrayBase::matrix() */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ArrayWrapper<Derived> array() { return ArrayWrapper<Derived>(derived()); } + /** \returns a const \link Eigen::ArrayBase Array \endlink expression of this matrix + * \sa ArrayBase::matrix() */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const ArrayWrapper<const Derived> array() const { return ArrayWrapper<const Derived>(derived()); } + +/////////// LU module /////////// + + inline const FullPivLU<PlainObject> fullPivLu() const; + inline const PartialPivLU<PlainObject> partialPivLu() const; + + inline const PartialPivLU<PlainObject> lu() const; + + inline const Inverse<Derived> inverse() const; + + template<typename ResultType> + inline void computeInverseAndDetWithCheck( + ResultType& inverse, + typename ResultType::Scalar& determinant, + bool& invertible, + const RealScalar& absDeterminantThreshold = NumTraits<Scalar>::dummy_precision() + ) const; + template<typename ResultType> + inline void computeInverseWithCheck( + ResultType& inverse, + bool& invertible, + const RealScalar& absDeterminantThreshold = NumTraits<Scalar>::dummy_precision() + ) const; + Scalar determinant() const; + +/////////// Cholesky module /////////// + + inline const LLT<PlainObject> llt() const; + inline const LDLT<PlainObject> ldlt() const; + +/////////// QR module /////////// + + inline const HouseholderQR<PlainObject> householderQr() const; + inline const ColPivHouseholderQR<PlainObject> colPivHouseholderQr() const; + inline const FullPivHouseholderQR<PlainObject> fullPivHouseholderQr() const; + inline const CompleteOrthogonalDecomposition<PlainObject> completeOrthogonalDecomposition() const; + +/////////// Eigenvalues module /////////// + + inline EigenvaluesReturnType eigenvalues() const; + inline RealScalar operatorNorm() const; + +/////////// SVD module /////////// + + inline JacobiSVD<PlainObject> jacobiSvd(unsigned int computationOptions = 0) const; + inline BDCSVD<PlainObject> bdcSvd(unsigned int computationOptions = 0) const; + +/////////// Geometry module /////////// + + #ifndef EIGEN_PARSED_BY_DOXYGEN + /// \internal helper struct to form the return type of the cross product + template<typename OtherDerived> struct cross_product_return_type { + typedef typename ScalarBinaryOpTraits<typename internal::traits<Derived>::Scalar,typename internal::traits<OtherDerived>::Scalar>::ReturnType Scalar; + typedef Matrix<Scalar,MatrixBase::RowsAtCompileTime,MatrixBase::ColsAtCompileTime> type; + }; + #endif // EIGEN_PARSED_BY_DOXYGEN + template<typename OtherDerived> + EIGEN_DEVICE_FUNC +#ifndef EIGEN_PARSED_BY_DOXYGEN + inline typename cross_product_return_type<OtherDerived>::type +#else + inline PlainObject +#endif + cross(const MatrixBase<OtherDerived>& other) const; + + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + inline PlainObject cross3(const MatrixBase<OtherDerived>& other) const; + + EIGEN_DEVICE_FUNC + inline PlainObject unitOrthogonal(void) const; + + EIGEN_DEVICE_FUNC + inline Matrix<Scalar,3,1> eulerAngles(Index a0, Index a1, Index a2) const; + + // put this as separate enum value to work around possible GCC 4.3 bug (?) + enum { HomogeneousReturnTypeDirection = ColsAtCompileTime==1&&RowsAtCompileTime==1 ? ((internal::traits<Derived>::Flags&RowMajorBit)==RowMajorBit ? Horizontal : Vertical) + : ColsAtCompileTime==1 ? Vertical : Horizontal }; + typedef Homogeneous<Derived, HomogeneousReturnTypeDirection> HomogeneousReturnType; + EIGEN_DEVICE_FUNC + inline HomogeneousReturnType homogeneous() const; + + enum { + SizeMinusOne = SizeAtCompileTime==Dynamic ? Dynamic : SizeAtCompileTime-1 + }; + typedef Block<const Derived, + internal::traits<Derived>::ColsAtCompileTime==1 ? SizeMinusOne : 1, + internal::traits<Derived>::ColsAtCompileTime==1 ? 1 : SizeMinusOne> ConstStartMinusOne; + typedef EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(ConstStartMinusOne,Scalar,quotient) HNormalizedReturnType; + EIGEN_DEVICE_FUNC + inline const HNormalizedReturnType hnormalized() const; + +////////// Householder module /////////// + + void makeHouseholderInPlace(Scalar& tau, RealScalar& beta); + template<typename EssentialPart> + void makeHouseholder(EssentialPart& essential, + Scalar& tau, RealScalar& beta) const; + template<typename EssentialPart> + void applyHouseholderOnTheLeft(const EssentialPart& essential, + const Scalar& tau, + Scalar* workspace); + template<typename EssentialPart> + void applyHouseholderOnTheRight(const EssentialPart& essential, + const Scalar& tau, + Scalar* workspace); + +///////// Jacobi module ///////// + + template<typename OtherScalar> + void applyOnTheLeft(Index p, Index q, const JacobiRotation<OtherScalar>& j); + template<typename OtherScalar> + void applyOnTheRight(Index p, Index q, const JacobiRotation<OtherScalar>& j); + +///////// SparseCore module ///////// + + template<typename OtherDerived> + EIGEN_STRONG_INLINE const typename SparseMatrixBase<OtherDerived>::template CwiseProductDenseReturnType<Derived>::Type + cwiseProduct(const SparseMatrixBase<OtherDerived> &other) const + { + return other.cwiseProduct(derived()); + } + +///////// MatrixFunctions module ///////// + + typedef typename internal::stem_function<Scalar>::type StemFunction; + const MatrixExponentialReturnValue<Derived> exp() const; + const MatrixFunctionReturnValue<Derived> matrixFunction(StemFunction f) const; + const MatrixFunctionReturnValue<Derived> cosh() const; + const MatrixFunctionReturnValue<Derived> sinh() const; + const MatrixFunctionReturnValue<Derived> cos() const; + const MatrixFunctionReturnValue<Derived> sin() const; + const MatrixSquareRootReturnValue<Derived> sqrt() const; + const MatrixLogarithmReturnValue<Derived> log() const; + const MatrixPowerReturnValue<Derived> pow(const RealScalar& p) const; + const MatrixComplexPowerReturnValue<Derived> pow(const std::complex<RealScalar>& p) const; + + protected: + EIGEN_DEVICE_FUNC MatrixBase() : Base() {} + + private: + EIGEN_DEVICE_FUNC explicit MatrixBase(int); + EIGEN_DEVICE_FUNC MatrixBase(int,int); + template<typename OtherDerived> EIGEN_DEVICE_FUNC explicit MatrixBase(const MatrixBase<OtherDerived>&); + protected: + // mixing arrays and matrices is not legal + template<typename OtherDerived> Derived& operator+=(const ArrayBase<OtherDerived>& ) + {EIGEN_STATIC_ASSERT(std::ptrdiff_t(sizeof(typename OtherDerived::Scalar))==-1,YOU_CANNOT_MIX_ARRAYS_AND_MATRICES); return *this;} + // mixing arrays and matrices is not legal + template<typename OtherDerived> Derived& operator-=(const ArrayBase<OtherDerived>& ) + {EIGEN_STATIC_ASSERT(std::ptrdiff_t(sizeof(typename OtherDerived::Scalar))==-1,YOU_CANNOT_MIX_ARRAYS_AND_MATRICES); return *this;} +}; + + +/*************************************************************************** +* Implementation of matrix base methods +***************************************************************************/ + +/** replaces \c *this by \c *this * \a other. + * + * \returns a reference to \c *this + * + * Example: \include MatrixBase_applyOnTheRight.cpp + * Output: \verbinclude MatrixBase_applyOnTheRight.out + */ +template<typename Derived> +template<typename OtherDerived> +inline Derived& +MatrixBase<Derived>::operator*=(const EigenBase<OtherDerived> &other) +{ + other.derived().applyThisOnTheRight(derived()); + return derived(); +} + +/** replaces \c *this by \c *this * \a other. It is equivalent to MatrixBase::operator*=(). + * + * Example: \include MatrixBase_applyOnTheRight.cpp + * Output: \verbinclude MatrixBase_applyOnTheRight.out + */ +template<typename Derived> +template<typename OtherDerived> +inline void MatrixBase<Derived>::applyOnTheRight(const EigenBase<OtherDerived> &other) +{ + other.derived().applyThisOnTheRight(derived()); +} + +/** replaces \c *this by \a other * \c *this. + * + * Example: \include MatrixBase_applyOnTheLeft.cpp + * Output: \verbinclude MatrixBase_applyOnTheLeft.out + */ +template<typename Derived> +template<typename OtherDerived> +inline void MatrixBase<Derived>::applyOnTheLeft(const EigenBase<OtherDerived> &other) +{ + other.derived().applyThisOnTheLeft(derived()); +} + +} // end namespace Eigen + +#endif // EIGEN_MATRIXBASE_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/NestByValue.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/NestByValue.h new file mode 100644 index 000000000..13adf070e --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/NestByValue.h @@ -0,0 +1,110 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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_NESTBYVALUE_H +#define EIGEN_NESTBYVALUE_H + +namespace Eigen { + +namespace internal { +template<typename ExpressionType> +struct traits<NestByValue<ExpressionType> > : public traits<ExpressionType> +{}; +} + +/** \class NestByValue + * \ingroup Core_Module + * + * \brief Expression which must be nested by value + * + * \tparam ExpressionType the type of the object of which we are requiring nesting-by-value + * + * This class is the return type of MatrixBase::nestByValue() + * and most of the time this is the only way it is used. + * + * \sa MatrixBase::nestByValue() + */ +template<typename ExpressionType> class NestByValue + : public internal::dense_xpr_base< NestByValue<ExpressionType> >::type +{ + public: + + typedef typename internal::dense_xpr_base<NestByValue>::type Base; + EIGEN_DENSE_PUBLIC_INTERFACE(NestByValue) + + EIGEN_DEVICE_FUNC explicit inline NestByValue(const ExpressionType& matrix) : m_expression(matrix) {} + + 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> + inline const PacketScalar packet(Index row, Index col) const + { + return m_expression.template packet<LoadMode>(row, col); + } + + template<int LoadMode> + inline void writePacket(Index row, Index col, const PacketScalar& x) + { + m_expression.const_cast_derived().template writePacket<LoadMode>(row, col, x); + } + + template<int LoadMode> + inline const PacketScalar packet(Index index) const + { + return m_expression.template packet<LoadMode>(index); + } + + template<int LoadMode> + 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; } + + protected: + const ExpressionType m_expression; +}; + +/** \returns an expression of the temporary version of *this. + */ +template<typename Derived> +inline const NestByValue<Derived> +DenseBase<Derived>::nestByValue() const +{ + return NestByValue<Derived>(derived()); +} + +} // end namespace Eigen + +#endif // EIGEN_NESTBYVALUE_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/NoAlias.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/NoAlias.h new file mode 100644 index 000000000..33908010b --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/NoAlias.h @@ -0,0 +1,108 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2009 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_NOALIAS_H +#define EIGEN_NOALIAS_H + +namespace Eigen { + +/** \class NoAlias + * \ingroup Core_Module + * + * \brief Pseudo expression providing an operator = assuming no aliasing + * + * \tparam ExpressionType the type of the object on which to do the lazy assignment + * + * This class represents an expression with special assignment operators + * assuming no aliasing between the target expression and the source expression. + * More precisely it alloas to bypass the EvalBeforeAssignBit flag of the source expression. + * It is the return type of MatrixBase::noalias() + * and most of the time this is the only way it is used. + * + * \sa MatrixBase::noalias() + */ +template<typename ExpressionType, template <typename> class StorageBase> +class NoAlias +{ + public: + typedef typename ExpressionType::Scalar Scalar; + + explicit NoAlias(ExpressionType& expression) : m_expression(expression) {} + + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE ExpressionType& operator=(const StorageBase<OtherDerived>& other) + { + call_assignment_no_alias(m_expression, other.derived(), internal::assign_op<Scalar,typename OtherDerived::Scalar>()); + return m_expression; + } + + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE ExpressionType& operator+=(const StorageBase<OtherDerived>& other) + { + call_assignment_no_alias(m_expression, other.derived(), internal::add_assign_op<Scalar,typename OtherDerived::Scalar>()); + return m_expression; + } + + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE ExpressionType& operator-=(const StorageBase<OtherDerived>& other) + { + call_assignment_no_alias(m_expression, other.derived(), internal::sub_assign_op<Scalar,typename OtherDerived::Scalar>()); + return m_expression; + } + + EIGEN_DEVICE_FUNC + ExpressionType& expression() const + { + return m_expression; + } + + protected: + ExpressionType& m_expression; +}; + +/** \returns a pseudo expression of \c *this with an operator= assuming + * no aliasing between \c *this and the source expression. + * + * More precisely, noalias() allows to bypass the EvalBeforeAssignBit flag. + * Currently, even though several expressions may alias, only product + * expressions have this flag. Therefore, noalias() is only usefull when + * the source expression contains a matrix product. + * + * Here are some examples where noalias is usefull: + * \code + * D.noalias() = A * B; + * D.noalias() += A.transpose() * B; + * D.noalias() -= 2 * A * B.adjoint(); + * \endcode + * + * On the other hand the following example will lead to a \b wrong result: + * \code + * A.noalias() = A * B; + * \endcode + * because the result matrix A is also an operand of the matrix product. Therefore, + * there is no alternative than evaluating A * B in a temporary, that is the default + * behavior when you write: + * \code + * A = A * B; + * \endcode + * + * \sa class NoAlias + */ +template<typename Derived> +NoAlias<Derived,MatrixBase> MatrixBase<Derived>::noalias() +{ + return NoAlias<Derived, Eigen::MatrixBase >(derived()); +} + +} // end namespace Eigen + +#endif // EIGEN_NOALIAS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/NumTraits.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/NumTraits.h new file mode 100644 index 000000000..daf489878 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/NumTraits.h @@ -0,0 +1,248 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2006-2010 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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_NUMTRAITS_H +#define EIGEN_NUMTRAITS_H + +namespace Eigen { + +namespace internal { + +// default implementation of digits10(), based on numeric_limits if specialized, +// 0 for integer types, and log10(epsilon()) otherwise. +template< typename T, + bool use_numeric_limits = std::numeric_limits<T>::is_specialized, + bool is_integer = NumTraits<T>::IsInteger> +struct default_digits10_impl +{ + static int run() { return std::numeric_limits<T>::digits10; } +}; + +template<typename T> +struct default_digits10_impl<T,false,false> // Floating point +{ + static int run() { + using std::log10; + using std::ceil; + typedef typename NumTraits<T>::Real Real; + return int(ceil(-log10(NumTraits<Real>::epsilon()))); + } +}; + +template<typename T> +struct default_digits10_impl<T,false,true> // Integer +{ + static int run() { return 0; } +}; + +} // end namespace internal + +/** \class NumTraits + * \ingroup Core_Module + * + * \brief Holds information about the various numeric (i.e. scalar) types allowed by Eigen. + * + * \tparam T the numeric type at hand + * + * This class stores enums, typedefs and static methods giving information about a numeric type. + * + * The provided data consists of: + * \li A typedef \c Real, giving the "real part" type of \a T. If \a T is already real, + * then \c Real is just a typedef to \a T. If \a T is \c std::complex<U> then \c Real + * is a typedef to \a U. + * \li A typedef \c NonInteger, giving the type that should be used for operations producing non-integral values, + * such as quotients, square roots, etc. If \a T is a floating-point type, then this typedef just gives + * \a T again. Note however that many Eigen functions such as internal::sqrt simply refuse to + * take integers. Outside of a few cases, Eigen doesn't do automatic type promotion. Thus, this typedef is + * only intended as a helper for code that needs to explicitly promote types. + * \li A typedef \c Literal giving the type to use for numeric literals such as "2" or "0.5". For instance, for \c std::complex<U>, Literal is defined as \c U. + * Of course, this type must be fully compatible with \a T. In doubt, just use \a T here. + * \li A typedef \a Nested giving the type to use to nest a value inside of the expression tree. If you don't know what + * this means, just use \a T here. + * \li An enum value \a IsComplex. It is equal to 1 if \a T is a \c std::complex + * type, and to 0 otherwise. + * \li An enum value \a IsInteger. It is equal to \c 1 if \a T is an integer type such as \c int, + * and to \c 0 otherwise. + * \li Enum values ReadCost, AddCost and MulCost representing a rough estimate of the number of CPU cycles needed + * to by move / add / mul instructions respectively, assuming the data is already stored in CPU registers. + * Stay vague here. No need to do architecture-specific stuff. + * \li An enum value \a IsSigned. It is equal to \c 1 if \a T is a signed type and to 0 if \a T is unsigned. + * \li An enum value \a RequireInitialization. It is equal to \c 1 if the constructor of the numeric type \a T must + * be called, and to 0 if it is safe not to call it. Default is 0 if \a T is an arithmetic type, and 1 otherwise. + * \li An epsilon() function which, unlike <a href="http://en.cppreference.com/w/cpp/types/numeric_limits/epsilon">std::numeric_limits::epsilon()</a>, + * it returns a \a Real instead of a \a T. + * \li A dummy_precision() function returning a weak epsilon value. It is mainly used as a default + * value by the fuzzy comparison operators. + * \li highest() and lowest() functions returning the highest and lowest possible values respectively. + * \li digits10() function returning the number of decimal digits that can be represented without change. This is + * the analogue of <a href="http://en.cppreference.com/w/cpp/types/numeric_limits/digits10">std::numeric_limits<T>::digits10</a> + * which is used as the default implementation if specialized. + */ + +template<typename T> struct GenericNumTraits +{ + enum { + IsInteger = std::numeric_limits<T>::is_integer, + IsSigned = std::numeric_limits<T>::is_signed, + IsComplex = 0, + RequireInitialization = internal::is_arithmetic<T>::value ? 0 : 1, + ReadCost = 1, + AddCost = 1, + MulCost = 1 + }; + + typedef T Real; + typedef typename internal::conditional< + IsInteger, + typename internal::conditional<sizeof(T)<=2, float, double>::type, + T + >::type NonInteger; + typedef T Nested; + typedef T Literal; + + EIGEN_DEVICE_FUNC + static inline Real epsilon() + { + return numext::numeric_limits<T>::epsilon(); + } + + EIGEN_DEVICE_FUNC + static inline int digits10() + { + return internal::default_digits10_impl<T>::run(); + } + + EIGEN_DEVICE_FUNC + static inline Real dummy_precision() + { + // make sure to override this for floating-point types + return Real(0); + } + + + EIGEN_DEVICE_FUNC + static inline T highest() { + return (numext::numeric_limits<T>::max)(); + } + + EIGEN_DEVICE_FUNC + static inline T lowest() { + return IsInteger ? (numext::numeric_limits<T>::min)() : (-(numext::numeric_limits<T>::max)()); + } + + EIGEN_DEVICE_FUNC + static inline T infinity() { + return numext::numeric_limits<T>::infinity(); + } + + EIGEN_DEVICE_FUNC + static inline T quiet_NaN() { + return numext::numeric_limits<T>::quiet_NaN(); + } +}; + +template<typename T> struct NumTraits : GenericNumTraits<T> +{}; + +template<> struct NumTraits<float> + : GenericNumTraits<float> +{ + EIGEN_DEVICE_FUNC + static inline float dummy_precision() { return 1e-5f; } +}; + +template<> struct NumTraits<double> : GenericNumTraits<double> +{ + EIGEN_DEVICE_FUNC + static inline double dummy_precision() { return 1e-12; } +}; + +template<> struct NumTraits<long double> + : GenericNumTraits<long double> +{ + static inline long double dummy_precision() { return 1e-15l; } +}; + +template<typename _Real> struct NumTraits<std::complex<_Real> > + : GenericNumTraits<std::complex<_Real> > +{ + typedef _Real Real; + typedef typename NumTraits<_Real>::Literal Literal; + enum { + IsComplex = 1, + RequireInitialization = NumTraits<_Real>::RequireInitialization, + ReadCost = 2 * NumTraits<_Real>::ReadCost, + AddCost = 2 * NumTraits<Real>::AddCost, + MulCost = 4 * NumTraits<Real>::MulCost + 2 * NumTraits<Real>::AddCost + }; + + EIGEN_DEVICE_FUNC + static inline Real epsilon() { return NumTraits<Real>::epsilon(); } + EIGEN_DEVICE_FUNC + static inline Real dummy_precision() { return NumTraits<Real>::dummy_precision(); } + EIGEN_DEVICE_FUNC + static inline int digits10() { return NumTraits<Real>::digits10(); } +}; + +template<typename Scalar, int Rows, int Cols, int Options, int MaxRows, int MaxCols> +struct NumTraits<Array<Scalar, Rows, Cols, Options, MaxRows, MaxCols> > +{ + typedef Array<Scalar, Rows, Cols, Options, MaxRows, MaxCols> ArrayType; + typedef typename NumTraits<Scalar>::Real RealScalar; + typedef Array<RealScalar, Rows, Cols, Options, MaxRows, MaxCols> Real; + typedef typename NumTraits<Scalar>::NonInteger NonIntegerScalar; + typedef Array<NonIntegerScalar, Rows, Cols, Options, MaxRows, MaxCols> NonInteger; + typedef ArrayType & Nested; + typedef typename NumTraits<Scalar>::Literal Literal; + + enum { + IsComplex = NumTraits<Scalar>::IsComplex, + IsInteger = NumTraits<Scalar>::IsInteger, + IsSigned = NumTraits<Scalar>::IsSigned, + RequireInitialization = 1, + ReadCost = ArrayType::SizeAtCompileTime==Dynamic ? HugeCost : ArrayType::SizeAtCompileTime * NumTraits<Scalar>::ReadCost, + AddCost = ArrayType::SizeAtCompileTime==Dynamic ? HugeCost : ArrayType::SizeAtCompileTime * NumTraits<Scalar>::AddCost, + MulCost = ArrayType::SizeAtCompileTime==Dynamic ? HugeCost : ArrayType::SizeAtCompileTime * NumTraits<Scalar>::MulCost + }; + + EIGEN_DEVICE_FUNC + static inline RealScalar epsilon() { return NumTraits<RealScalar>::epsilon(); } + EIGEN_DEVICE_FUNC + static inline RealScalar dummy_precision() { return NumTraits<RealScalar>::dummy_precision(); } + + static inline int digits10() { return NumTraits<Scalar>::digits10(); } +}; + +template<> struct NumTraits<std::string> + : GenericNumTraits<std::string> +{ + enum { + RequireInitialization = 1, + ReadCost = HugeCost, + AddCost = HugeCost, + MulCost = HugeCost + }; + + static inline int digits10() { return 0; } + +private: + static inline std::string epsilon(); + static inline std::string dummy_precision(); + static inline std::string lowest(); + static inline std::string highest(); + static inline std::string infinity(); + static inline std::string quiet_NaN(); +}; + +// Empty specialization for void to allow template specialization based on NumTraits<T>::Real with T==void and SFINAE. +template<> struct NumTraits<void> {}; + +} // end namespace Eigen + +#endif // EIGEN_NUMTRAITS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/PermutationMatrix.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/PermutationMatrix.h new file mode 100644 index 000000000..b1fb455b9 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/PermutationMatrix.h @@ -0,0 +1,633 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2009 Benoit Jacob <jacob.benoit.1@gmail.com> +// Copyright (C) 2009-2015 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_PERMUTATIONMATRIX_H +#define EIGEN_PERMUTATIONMATRIX_H + +namespace Eigen { + +namespace internal { + +enum PermPermProduct_t {PermPermProduct}; + +} // end namespace internal + +/** \class PermutationBase + * \ingroup Core_Module + * + * \brief Base class for permutations + * + * \tparam Derived the derived class + * + * This class is the base class for all expressions representing a permutation matrix, + * internally stored as a vector of integers. + * The convention followed here is that if \f$ \sigma \f$ is a permutation, the corresponding permutation matrix + * \f$ P_\sigma \f$ is such that if \f$ (e_1,\ldots,e_p) \f$ is the canonical basis, we have: + * \f[ P_\sigma(e_i) = e_{\sigma(i)}. \f] + * This convention ensures that for any two permutations \f$ \sigma, \tau \f$, we have: + * \f[ P_{\sigma\circ\tau} = P_\sigma P_\tau. \f] + * + * Permutation matrices are square and invertible. + * + * Notice that in addition to the member functions and operators listed here, there also are non-member + * operator* to multiply any kind of permutation object with any kind of matrix expression (MatrixBase) + * on either side. + * + * \sa class PermutationMatrix, class PermutationWrapper + */ +template<typename Derived> +class PermutationBase : public EigenBase<Derived> +{ + typedef internal::traits<Derived> Traits; + typedef EigenBase<Derived> Base; + public: + + #ifndef EIGEN_PARSED_BY_DOXYGEN + typedef typename Traits::IndicesType IndicesType; + enum { + Flags = Traits::Flags, + RowsAtCompileTime = Traits::RowsAtCompileTime, + ColsAtCompileTime = Traits::ColsAtCompileTime, + MaxRowsAtCompileTime = Traits::MaxRowsAtCompileTime, + MaxColsAtCompileTime = Traits::MaxColsAtCompileTime + }; + typedef typename Traits::StorageIndex StorageIndex; + typedef Matrix<StorageIndex,RowsAtCompileTime,ColsAtCompileTime,0,MaxRowsAtCompileTime,MaxColsAtCompileTime> + DenseMatrixType; + typedef PermutationMatrix<IndicesType::SizeAtCompileTime,IndicesType::MaxSizeAtCompileTime,StorageIndex> + PlainPermutationType; + typedef PlainPermutationType PlainObject; + using Base::derived; + typedef Inverse<Derived> InverseReturnType; + typedef void Scalar; + #endif + + /** Copies the other permutation into *this */ + template<typename OtherDerived> + Derived& operator=(const PermutationBase<OtherDerived>& other) + { + indices() = other.indices(); + return derived(); + } + + /** Assignment from the Transpositions \a tr */ + template<typename OtherDerived> + Derived& operator=(const TranspositionsBase<OtherDerived>& tr) + { + setIdentity(tr.size()); + for(Index k=size()-1; k>=0; --k) + applyTranspositionOnTheRight(k,tr.coeff(k)); + return derived(); + } + + #ifndef EIGEN_PARSED_BY_DOXYGEN + /** This is a special case of the templated operator=. Its purpose is to + * prevent a default operator= from hiding the templated operator=. + */ + Derived& operator=(const PermutationBase& other) + { + indices() = other.indices(); + return derived(); + } + #endif + + /** \returns the number of rows */ + inline Index rows() const { return Index(indices().size()); } + + /** \returns the number of columns */ + inline Index cols() const { return Index(indices().size()); } + + /** \returns the size of a side of the respective square matrix, i.e., the number of indices */ + inline Index size() const { return Index(indices().size()); } + + #ifndef EIGEN_PARSED_BY_DOXYGEN + template<typename DenseDerived> + void evalTo(MatrixBase<DenseDerived>& other) const + { + other.setZero(); + for (Index i=0; i<rows(); ++i) + other.coeffRef(indices().coeff(i),i) = typename DenseDerived::Scalar(1); + } + #endif + + /** \returns a Matrix object initialized from this permutation matrix. Notice that it + * is inefficient to return this Matrix object by value. For efficiency, favor using + * the Matrix constructor taking EigenBase objects. + */ + DenseMatrixType toDenseMatrix() const + { + return derived(); + } + + /** const version of indices(). */ + const IndicesType& indices() const { return derived().indices(); } + /** \returns a reference to the stored array representing the permutation. */ + IndicesType& indices() { return derived().indices(); } + + /** Resizes to given size. + */ + inline void resize(Index newSize) + { + indices().resize(newSize); + } + + /** Sets *this to be the identity permutation matrix */ + void setIdentity() + { + StorageIndex n = StorageIndex(size()); + for(StorageIndex i = 0; i < n; ++i) + indices().coeffRef(i) = i; + } + + /** Sets *this to be the identity permutation matrix of given size. + */ + void setIdentity(Index newSize) + { + resize(newSize); + setIdentity(); + } + + /** Multiplies *this by the transposition \f$(ij)\f$ on the left. + * + * \returns a reference to *this. + * + * \warning This is much slower than applyTranspositionOnTheRight(Index,Index): + * this has linear complexity and requires a lot of branching. + * + * \sa applyTranspositionOnTheRight(Index,Index) + */ + Derived& applyTranspositionOnTheLeft(Index i, Index j) + { + eigen_assert(i>=0 && j>=0 && i<size() && j<size()); + for(Index k = 0; k < size(); ++k) + { + if(indices().coeff(k) == i) indices().coeffRef(k) = StorageIndex(j); + else if(indices().coeff(k) == j) indices().coeffRef(k) = StorageIndex(i); + } + return derived(); + } + + /** Multiplies *this by the transposition \f$(ij)\f$ on the right. + * + * \returns a reference to *this. + * + * This is a fast operation, it only consists in swapping two indices. + * + * \sa applyTranspositionOnTheLeft(Index,Index) + */ + Derived& applyTranspositionOnTheRight(Index i, Index j) + { + eigen_assert(i>=0 && j>=0 && i<size() && j<size()); + std::swap(indices().coeffRef(i), indices().coeffRef(j)); + return derived(); + } + + /** \returns the inverse permutation matrix. + * + * \note \blank \note_try_to_help_rvo + */ + inline InverseReturnType inverse() const + { return InverseReturnType(derived()); } + /** \returns the tranpose permutation matrix. + * + * \note \blank \note_try_to_help_rvo + */ + inline InverseReturnType transpose() const + { return InverseReturnType(derived()); } + + /**** multiplication helpers to hopefully get RVO ****/ + + +#ifndef EIGEN_PARSED_BY_DOXYGEN + protected: + template<typename OtherDerived> + void assignTranspose(const PermutationBase<OtherDerived>& other) + { + for (Index i=0; i<rows();++i) indices().coeffRef(other.indices().coeff(i)) = i; + } + template<typename Lhs,typename Rhs> + void assignProduct(const Lhs& lhs, const Rhs& rhs) + { + eigen_assert(lhs.cols() == rhs.rows()); + for (Index i=0; i<rows();++i) indices().coeffRef(i) = lhs.indices().coeff(rhs.indices().coeff(i)); + } +#endif + + public: + + /** \returns the product permutation matrix. + * + * \note \blank \note_try_to_help_rvo + */ + template<typename Other> + inline PlainPermutationType operator*(const PermutationBase<Other>& other) const + { return PlainPermutationType(internal::PermPermProduct, derived(), other.derived()); } + + /** \returns the product of a permutation with another inverse permutation. + * + * \note \blank \note_try_to_help_rvo + */ + template<typename Other> + inline PlainPermutationType operator*(const InverseImpl<Other,PermutationStorage>& other) const + { return PlainPermutationType(internal::PermPermProduct, *this, other.eval()); } + + /** \returns the product of an inverse permutation with another permutation. + * + * \note \blank \note_try_to_help_rvo + */ + template<typename Other> friend + inline PlainPermutationType operator*(const InverseImpl<Other, PermutationStorage>& other, const PermutationBase& perm) + { return PlainPermutationType(internal::PermPermProduct, other.eval(), perm); } + + /** \returns the determinant of the permutation matrix, which is either 1 or -1 depending on the parity of the permutation. + * + * This function is O(\c n) procedure allocating a buffer of \c n booleans. + */ + Index determinant() const + { + Index res = 1; + Index n = size(); + Matrix<bool,RowsAtCompileTime,1,0,MaxRowsAtCompileTime> mask(n); + mask.fill(false); + Index r = 0; + while(r < n) + { + // search for the next seed + while(r<n && mask[r]) r++; + if(r>=n) + break; + // we got one, let's follow it until we are back to the seed + Index k0 = r++; + mask.coeffRef(k0) = true; + for(Index k=indices().coeff(k0); k!=k0; k=indices().coeff(k)) + { + mask.coeffRef(k) = true; + res = -res; + } + } + return res; + } + + protected: + +}; + +namespace internal { +template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex> +struct traits<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndex> > + : traits<Matrix<_StorageIndex,SizeAtCompileTime,SizeAtCompileTime,0,MaxSizeAtCompileTime,MaxSizeAtCompileTime> > +{ + typedef PermutationStorage StorageKind; + typedef Matrix<_StorageIndex, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1> IndicesType; + typedef _StorageIndex StorageIndex; + typedef void Scalar; +}; +} + +/** \class PermutationMatrix + * \ingroup Core_Module + * + * \brief Permutation matrix + * + * \tparam SizeAtCompileTime the number of rows/cols, or Dynamic + * \tparam MaxSizeAtCompileTime the maximum number of rows/cols, or Dynamic. This optional parameter defaults to SizeAtCompileTime. Most of the time, you should not have to specify it. + * \tparam _StorageIndex the integer type of the indices + * + * This class represents a permutation matrix, internally stored as a vector of integers. + * + * \sa class PermutationBase, class PermutationWrapper, class DiagonalMatrix + */ +template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex> +class PermutationMatrix : public PermutationBase<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndex> > +{ + typedef PermutationBase<PermutationMatrix> Base; + typedef internal::traits<PermutationMatrix> Traits; + public: + + typedef const PermutationMatrix& Nested; + + #ifndef EIGEN_PARSED_BY_DOXYGEN + typedef typename Traits::IndicesType IndicesType; + typedef typename Traits::StorageIndex StorageIndex; + #endif + + inline PermutationMatrix() + {} + + /** Constructs an uninitialized permutation matrix of given size. + */ + explicit inline PermutationMatrix(Index size) : m_indices(size) + { + eigen_internal_assert(size <= NumTraits<StorageIndex>::highest()); + } + + /** Copy constructor. */ + template<typename OtherDerived> + inline PermutationMatrix(const PermutationBase<OtherDerived>& other) + : m_indices(other.indices()) {} + + #ifndef EIGEN_PARSED_BY_DOXYGEN + /** Standard copy constructor. Defined only to prevent a default copy constructor + * from hiding the other templated constructor */ + inline PermutationMatrix(const PermutationMatrix& other) : m_indices(other.indices()) {} + #endif + + /** Generic constructor from expression of the indices. The indices + * array has the meaning that the permutations sends each integer i to indices[i]. + * + * \warning It is your responsibility to check that the indices array that you passes actually + * describes a permutation, i.e., each value between 0 and n-1 occurs exactly once, where n is the + * array's size. + */ + template<typename Other> + explicit inline PermutationMatrix(const MatrixBase<Other>& indices) : m_indices(indices) + {} + + /** Convert the Transpositions \a tr to a permutation matrix */ + template<typename Other> + explicit PermutationMatrix(const TranspositionsBase<Other>& tr) + : m_indices(tr.size()) + { + *this = tr; + } + + /** Copies the other permutation into *this */ + template<typename Other> + PermutationMatrix& operator=(const PermutationBase<Other>& other) + { + m_indices = other.indices(); + return *this; + } + + /** Assignment from the Transpositions \a tr */ + template<typename Other> + PermutationMatrix& operator=(const TranspositionsBase<Other>& tr) + { + return Base::operator=(tr.derived()); + } + + #ifndef EIGEN_PARSED_BY_DOXYGEN + /** This is a special case of the templated operator=. Its purpose is to + * prevent a default operator= from hiding the templated operator=. + */ + PermutationMatrix& operator=(const PermutationMatrix& other) + { + m_indices = other.m_indices; + return *this; + } + #endif + + /** const version of indices(). */ + const IndicesType& indices() const { return m_indices; } + /** \returns a reference to the stored array representing the permutation. */ + IndicesType& indices() { return m_indices; } + + + /**** multiplication helpers to hopefully get RVO ****/ + +#ifndef EIGEN_PARSED_BY_DOXYGEN + template<typename Other> + PermutationMatrix(const InverseImpl<Other,PermutationStorage>& other) + : m_indices(other.derived().nestedExpression().size()) + { + eigen_internal_assert(m_indices.size() <= NumTraits<StorageIndex>::highest()); + StorageIndex end = StorageIndex(m_indices.size()); + for (StorageIndex i=0; i<end;++i) + m_indices.coeffRef(other.derived().nestedExpression().indices().coeff(i)) = i; + } + template<typename Lhs,typename Rhs> + PermutationMatrix(internal::PermPermProduct_t, const Lhs& lhs, const Rhs& rhs) + : m_indices(lhs.indices().size()) + { + Base::assignProduct(lhs,rhs); + } +#endif + + protected: + + IndicesType m_indices; +}; + + +namespace internal { +template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex, int _PacketAccess> +struct traits<Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndex>,_PacketAccess> > + : traits<Matrix<_StorageIndex,SizeAtCompileTime,SizeAtCompileTime,0,MaxSizeAtCompileTime,MaxSizeAtCompileTime> > +{ + typedef PermutationStorage StorageKind; + typedef Map<const Matrix<_StorageIndex, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1>, _PacketAccess> IndicesType; + typedef _StorageIndex StorageIndex; + typedef void Scalar; +}; +} + +template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex, int _PacketAccess> +class Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndex>,_PacketAccess> + : public PermutationBase<Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndex>,_PacketAccess> > +{ + typedef PermutationBase<Map> Base; + typedef internal::traits<Map> Traits; + public: + + #ifndef EIGEN_PARSED_BY_DOXYGEN + typedef typename Traits::IndicesType IndicesType; + typedef typename IndicesType::Scalar StorageIndex; + #endif + + inline Map(const StorageIndex* indicesPtr) + : m_indices(indicesPtr) + {} + + inline Map(const StorageIndex* indicesPtr, Index size) + : m_indices(indicesPtr,size) + {} + + /** Copies the other permutation into *this */ + template<typename Other> + Map& operator=(const PermutationBase<Other>& other) + { return Base::operator=(other.derived()); } + + /** Assignment from the Transpositions \a tr */ + template<typename Other> + Map& operator=(const TranspositionsBase<Other>& tr) + { return Base::operator=(tr.derived()); } + + #ifndef EIGEN_PARSED_BY_DOXYGEN + /** This is a special case of the templated operator=. Its purpose is to + * prevent a default operator= from hiding the templated operator=. + */ + Map& operator=(const Map& other) + { + m_indices = other.m_indices; + return *this; + } + #endif + + /** const version of indices(). */ + const IndicesType& indices() const { return m_indices; } + /** \returns a reference to the stored array representing the permutation. */ + IndicesType& indices() { return m_indices; } + + protected: + + IndicesType m_indices; +}; + +template<typename _IndicesType> class TranspositionsWrapper; +namespace internal { +template<typename _IndicesType> +struct traits<PermutationWrapper<_IndicesType> > +{ + typedef PermutationStorage StorageKind; + typedef void Scalar; + typedef typename _IndicesType::Scalar StorageIndex; + typedef _IndicesType IndicesType; + enum { + RowsAtCompileTime = _IndicesType::SizeAtCompileTime, + ColsAtCompileTime = _IndicesType::SizeAtCompileTime, + MaxRowsAtCompileTime = IndicesType::MaxSizeAtCompileTime, + MaxColsAtCompileTime = IndicesType::MaxSizeAtCompileTime, + Flags = 0 + }; +}; +} + +/** \class PermutationWrapper + * \ingroup Core_Module + * + * \brief Class to view a vector of integers as a permutation matrix + * + * \tparam _IndicesType the type of the vector of integer (can be any compatible expression) + * + * This class allows to view any vector expression of integers as a permutation matrix. + * + * \sa class PermutationBase, class PermutationMatrix + */ +template<typename _IndicesType> +class PermutationWrapper : public PermutationBase<PermutationWrapper<_IndicesType> > +{ + typedef PermutationBase<PermutationWrapper> Base; + typedef internal::traits<PermutationWrapper> Traits; + public: + + #ifndef EIGEN_PARSED_BY_DOXYGEN + typedef typename Traits::IndicesType IndicesType; + #endif + + inline PermutationWrapper(const IndicesType& indices) + : m_indices(indices) + {} + + /** const version of indices(). */ + const typename internal::remove_all<typename IndicesType::Nested>::type& + indices() const { return m_indices; } + + protected: + + typename IndicesType::Nested m_indices; +}; + + +/** \returns the matrix with the permutation applied to the columns. + */ +template<typename MatrixDerived, typename PermutationDerived> +EIGEN_DEVICE_FUNC +const Product<MatrixDerived, PermutationDerived, AliasFreeProduct> +operator*(const MatrixBase<MatrixDerived> &matrix, + const PermutationBase<PermutationDerived>& permutation) +{ + return Product<MatrixDerived, PermutationDerived, AliasFreeProduct> + (matrix.derived(), permutation.derived()); +} + +/** \returns the matrix with the permutation applied to the rows. + */ +template<typename PermutationDerived, typename MatrixDerived> +EIGEN_DEVICE_FUNC +const Product<PermutationDerived, MatrixDerived, AliasFreeProduct> +operator*(const PermutationBase<PermutationDerived> &permutation, + const MatrixBase<MatrixDerived>& matrix) +{ + return Product<PermutationDerived, MatrixDerived, AliasFreeProduct> + (permutation.derived(), matrix.derived()); +} + + +template<typename PermutationType> +class InverseImpl<PermutationType, PermutationStorage> + : public EigenBase<Inverse<PermutationType> > +{ + typedef typename PermutationType::PlainPermutationType PlainPermutationType; + typedef internal::traits<PermutationType> PermTraits; + protected: + InverseImpl() {} + public: + typedef Inverse<PermutationType> InverseType; + using EigenBase<Inverse<PermutationType> >::derived; + + #ifndef EIGEN_PARSED_BY_DOXYGEN + typedef typename PermutationType::DenseMatrixType DenseMatrixType; + enum { + RowsAtCompileTime = PermTraits::RowsAtCompileTime, + ColsAtCompileTime = PermTraits::ColsAtCompileTime, + MaxRowsAtCompileTime = PermTraits::MaxRowsAtCompileTime, + MaxColsAtCompileTime = PermTraits::MaxColsAtCompileTime + }; + #endif + + #ifndef EIGEN_PARSED_BY_DOXYGEN + template<typename DenseDerived> + void evalTo(MatrixBase<DenseDerived>& other) const + { + other.setZero(); + for (Index i=0; i<derived().rows();++i) + other.coeffRef(i, derived().nestedExpression().indices().coeff(i)) = typename DenseDerived::Scalar(1); + } + #endif + + /** \return the equivalent permutation matrix */ + PlainPermutationType eval() const { return derived(); } + + DenseMatrixType toDenseMatrix() const { return derived(); } + + /** \returns the matrix with the inverse permutation applied to the columns. + */ + template<typename OtherDerived> friend + const Product<OtherDerived, InverseType, AliasFreeProduct> + operator*(const MatrixBase<OtherDerived>& matrix, const InverseType& trPerm) + { + return Product<OtherDerived, InverseType, AliasFreeProduct>(matrix.derived(), trPerm.derived()); + } + + /** \returns the matrix with the inverse permutation applied to the rows. + */ + template<typename OtherDerived> + const Product<InverseType, OtherDerived, AliasFreeProduct> + operator*(const MatrixBase<OtherDerived>& matrix) const + { + return Product<InverseType, OtherDerived, AliasFreeProduct>(derived(), matrix.derived()); + } +}; + +template<typename Derived> +const PermutationWrapper<const Derived> MatrixBase<Derived>::asPermutation() const +{ + return derived(); +} + +namespace internal { + +template<> struct AssignmentKind<DenseShape,PermutationShape> { typedef EigenBase2EigenBase Kind; }; + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_PERMUTATIONMATRIX_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/PlainObjectBase.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/PlainObjectBase.h new file mode 100644 index 000000000..77f4f6066 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/PlainObjectBase.h @@ -0,0 +1,1031 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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_DENSESTORAGEBASE_H +#define EIGEN_DENSESTORAGEBASE_H + +#if defined(EIGEN_INITIALIZE_MATRICES_BY_ZERO) +# define EIGEN_INITIALIZE_COEFFS +# define EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED for(int i=0;i<base().size();++i) coeffRef(i)=Scalar(0); +#elif defined(EIGEN_INITIALIZE_MATRICES_BY_NAN) +# define EIGEN_INITIALIZE_COEFFS +# define EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED for(int i=0;i<base().size();++i) coeffRef(i)=std::numeric_limits<Scalar>::quiet_NaN(); +#else +# undef EIGEN_INITIALIZE_COEFFS +# define EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED +#endif + +namespace Eigen { + +namespace internal { + +template<int MaxSizeAtCompileTime> struct check_rows_cols_for_overflow { + template<typename Index> + EIGEN_DEVICE_FUNC + static EIGEN_ALWAYS_INLINE void run(Index, Index) + { + } +}; + +template<> struct check_rows_cols_for_overflow<Dynamic> { + template<typename Index> + EIGEN_DEVICE_FUNC + static EIGEN_ALWAYS_INLINE void run(Index rows, Index cols) + { + // http://hg.mozilla.org/mozilla-central/file/6c8a909977d3/xpcom/ds/CheckedInt.h#l242 + // we assume Index is signed + Index max_index = (std::size_t(1) << (8 * sizeof(Index) - 1)) - 1; // assume Index is signed + bool error = (rows == 0 || cols == 0) ? false + : (rows > max_index / cols); + if (error) + throw_std_bad_alloc(); + } +}; + +template <typename Derived, + typename OtherDerived = Derived, + bool IsVector = bool(Derived::IsVectorAtCompileTime) && bool(OtherDerived::IsVectorAtCompileTime)> +struct conservative_resize_like_impl; + +template<typename MatrixTypeA, typename MatrixTypeB, bool SwapPointers> struct matrix_swap_impl; + +} // end namespace internal + +#ifdef EIGEN_PARSED_BY_DOXYGEN +namespace doxygen { + +// This is a workaround to doxygen not being able to understand the inheritance logic +// when it is hidden by the dense_xpr_base helper struct. +// Moreover, doxygen fails to include members that are not documented in the declaration body of +// MatrixBase if we inherits MatrixBase<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >, +// this is why we simply inherits MatrixBase, though this does not make sense. + +/** This class is just a workaround for Doxygen and it does not not actually exist. */ +template<typename Derived> struct dense_xpr_base_dispatcher; +/** This class is just a workaround for Doxygen and it does not not actually exist. */ +template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols> +struct dense_xpr_base_dispatcher<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> > + : public MatrixBase {}; +/** This class is just a workaround for Doxygen and it does not not actually exist. */ +template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols> +struct dense_xpr_base_dispatcher<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> > + : public ArrayBase {}; + +} // namespace doxygen + +/** \class PlainObjectBase + * \ingroup Core_Module + * \brief %Dense storage base class for matrices and arrays. + * + * This class can be extended with the help of the plugin mechanism described on the page + * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_PLAINOBJECTBASE_PLUGIN. + * + * \tparam Derived is the derived type, e.g., a Matrix or Array + * + * \sa \ref TopicClassHierarchy + */ +template<typename Derived> +class PlainObjectBase : public doxygen::dense_xpr_base_dispatcher<Derived> +#else +template<typename Derived> +class PlainObjectBase : public internal::dense_xpr_base<Derived>::type +#endif +{ + public: + enum { Options = internal::traits<Derived>::Options }; + typedef typename internal::dense_xpr_base<Derived>::type Base; + + typedef typename internal::traits<Derived>::StorageKind StorageKind; + typedef typename internal::traits<Derived>::Scalar Scalar; + + typedef typename internal::packet_traits<Scalar>::type PacketScalar; + typedef typename NumTraits<Scalar>::Real RealScalar; + typedef Derived DenseType; + + using Base::RowsAtCompileTime; + using Base::ColsAtCompileTime; + using Base::SizeAtCompileTime; + using Base::MaxRowsAtCompileTime; + using Base::MaxColsAtCompileTime; + using Base::MaxSizeAtCompileTime; + using Base::IsVectorAtCompileTime; + using Base::Flags; + + template<typename PlainObjectType, int MapOptions, typename StrideType> friend class Eigen::Map; + friend class Eigen::Map<Derived, Unaligned>; + typedef Eigen::Map<Derived, Unaligned> MapType; + friend class Eigen::Map<const Derived, Unaligned>; + typedef const Eigen::Map<const Derived, Unaligned> ConstMapType; +#if EIGEN_MAX_ALIGN_BYTES>0 + // for EIGEN_MAX_ALIGN_BYTES==0, AlignedMax==Unaligned, and many compilers generate warnings for friend-ing a class twice. + friend class Eigen::Map<Derived, AlignedMax>; + friend class Eigen::Map<const Derived, AlignedMax>; +#endif + typedef Eigen::Map<Derived, AlignedMax> AlignedMapType; + typedef const Eigen::Map<const Derived, AlignedMax> ConstAlignedMapType; + template<typename StrideType> struct StridedMapType { typedef Eigen::Map<Derived, Unaligned, StrideType> type; }; + template<typename StrideType> struct StridedConstMapType { typedef Eigen::Map<const Derived, Unaligned, StrideType> type; }; + template<typename StrideType> struct StridedAlignedMapType { typedef Eigen::Map<Derived, AlignedMax, StrideType> type; }; + template<typename StrideType> struct StridedConstAlignedMapType { typedef Eigen::Map<const Derived, AlignedMax, StrideType> type; }; + + protected: + DenseStorage<Scalar, Base::MaxSizeAtCompileTime, Base::RowsAtCompileTime, Base::ColsAtCompileTime, Options> m_storage; + + public: + enum { NeedsToAlign = (SizeAtCompileTime != Dynamic) && (internal::traits<Derived>::Alignment>0) }; + EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign) + + EIGEN_DEVICE_FUNC + Base& base() { return *static_cast<Base*>(this); } + EIGEN_DEVICE_FUNC + const Base& base() const { return *static_cast<const Base*>(this); } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Index rows() const { return m_storage.rows(); } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Index cols() const { return m_storage.cols(); } + + /** This is an overloaded version of DenseCoeffsBase<Derived,ReadOnlyAccessors>::coeff(Index,Index) const + * provided to by-pass the creation of an evaluator of the expression, thus saving compilation efforts. + * + * See DenseCoeffsBase<Derived,ReadOnlyAccessors>::coeff(Index) const for details. */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar& coeff(Index rowId, Index colId) const + { + if(Flags & RowMajorBit) + return m_storage.data()[colId + rowId * m_storage.cols()]; + else // column-major + return m_storage.data()[rowId + colId * m_storage.rows()]; + } + + /** This is an overloaded version of DenseCoeffsBase<Derived,ReadOnlyAccessors>::coeff(Index) const + * provided to by-pass the creation of an evaluator of the expression, thus saving compilation efforts. + * + * See DenseCoeffsBase<Derived,ReadOnlyAccessors>::coeff(Index) const for details. */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar& coeff(Index index) const + { + return m_storage.data()[index]; + } + + /** This is an overloaded version of DenseCoeffsBase<Derived,WriteAccessors>::coeffRef(Index,Index) const + * provided to by-pass the creation of an evaluator of the expression, thus saving compilation efforts. + * + * See DenseCoeffsBase<Derived,WriteAccessors>::coeffRef(Index,Index) const for details. */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar& coeffRef(Index rowId, Index colId) + { + if(Flags & RowMajorBit) + return m_storage.data()[colId + rowId * m_storage.cols()]; + else // column-major + return m_storage.data()[rowId + colId * m_storage.rows()]; + } + + /** This is an overloaded version of DenseCoeffsBase<Derived,WriteAccessors>::coeffRef(Index) const + * provided to by-pass the creation of an evaluator of the expression, thus saving compilation efforts. + * + * See DenseCoeffsBase<Derived,WriteAccessors>::coeffRef(Index) const for details. */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar& coeffRef(Index index) + { + return m_storage.data()[index]; + } + + /** This is the const version of coeffRef(Index,Index) which is thus synonym of coeff(Index,Index). + * It is provided for convenience. */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar& coeffRef(Index rowId, Index colId) const + { + if(Flags & RowMajorBit) + return m_storage.data()[colId + rowId * m_storage.cols()]; + else // column-major + return m_storage.data()[rowId + colId * m_storage.rows()]; + } + + /** This is the const version of coeffRef(Index) which is thus synonym of coeff(Index). + * It is provided for convenience. */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar& coeffRef(Index index) const + { + return m_storage.data()[index]; + } + + /** \internal */ + template<int LoadMode> + EIGEN_STRONG_INLINE PacketScalar packet(Index rowId, Index colId) const + { + return internal::ploadt<PacketScalar, LoadMode> + (m_storage.data() + (Flags & RowMajorBit + ? colId + rowId * m_storage.cols() + : rowId + colId * m_storage.rows())); + } + + /** \internal */ + template<int LoadMode> + EIGEN_STRONG_INLINE PacketScalar packet(Index index) const + { + return internal::ploadt<PacketScalar, LoadMode>(m_storage.data() + index); + } + + /** \internal */ + template<int StoreMode> + EIGEN_STRONG_INLINE void writePacket(Index rowId, Index colId, const PacketScalar& val) + { + internal::pstoret<Scalar, PacketScalar, StoreMode> + (m_storage.data() + (Flags & RowMajorBit + ? colId + rowId * m_storage.cols() + : rowId + colId * m_storage.rows()), val); + } + + /** \internal */ + template<int StoreMode> + EIGEN_STRONG_INLINE void writePacket(Index index, const PacketScalar& val) + { + internal::pstoret<Scalar, PacketScalar, StoreMode>(m_storage.data() + index, val); + } + + /** \returns a const pointer to the data array of this matrix */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar *data() const + { return m_storage.data(); } + + /** \returns a pointer to the data array of this matrix */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar *data() + { return m_storage.data(); } + + /** Resizes \c *this to a \a rows x \a cols matrix. + * + * This method is intended for dynamic-size matrices, although it is legal to call it on any + * matrix as long as fixed dimensions are left unchanged. If you only want to change the number + * of rows and/or of columns, you can use resize(NoChange_t, Index), resize(Index, NoChange_t). + * + * If the current number of coefficients of \c *this exactly matches the + * product \a rows * \a cols, then no memory allocation is performed and + * the current values are left unchanged. In all other cases, including + * shrinking, the data is reallocated and all previous values are lost. + * + * Example: \include Matrix_resize_int_int.cpp + * Output: \verbinclude Matrix_resize_int_int.out + * + * \sa resize(Index) for vectors, resize(NoChange_t, Index), resize(Index, NoChange_t) + */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void resize(Index rows, Index cols) + { + eigen_assert( EIGEN_IMPLIES(RowsAtCompileTime!=Dynamic,rows==RowsAtCompileTime) + && EIGEN_IMPLIES(ColsAtCompileTime!=Dynamic,cols==ColsAtCompileTime) + && EIGEN_IMPLIES(RowsAtCompileTime==Dynamic && MaxRowsAtCompileTime!=Dynamic,rows<=MaxRowsAtCompileTime) + && EIGEN_IMPLIES(ColsAtCompileTime==Dynamic && MaxColsAtCompileTime!=Dynamic,cols<=MaxColsAtCompileTime) + && rows>=0 && cols>=0 && "Invalid sizes when resizing a matrix or array."); + internal::check_rows_cols_for_overflow<MaxSizeAtCompileTime>::run(rows, cols); + #ifdef EIGEN_INITIALIZE_COEFFS + Index size = rows*cols; + bool size_changed = size != this->size(); + m_storage.resize(size, rows, cols); + if(size_changed) EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED + #else + m_storage.resize(rows*cols, rows, cols); + #endif + } + + /** Resizes \c *this to a vector of length \a size + * + * \only_for_vectors. This method does not work for + * partially dynamic matrices when the static dimension is anything other + * than 1. For example it will not work with Matrix<double, 2, Dynamic>. + * + * Example: \include Matrix_resize_int.cpp + * Output: \verbinclude Matrix_resize_int.out + * + * \sa resize(Index,Index), resize(NoChange_t, Index), resize(Index, NoChange_t) + */ + EIGEN_DEVICE_FUNC + inline void resize(Index size) + { + EIGEN_STATIC_ASSERT_VECTOR_ONLY(PlainObjectBase) + eigen_assert(((SizeAtCompileTime == Dynamic && (MaxSizeAtCompileTime==Dynamic || size<=MaxSizeAtCompileTime)) || SizeAtCompileTime == size) && size>=0); + #ifdef EIGEN_INITIALIZE_COEFFS + bool size_changed = size != this->size(); + #endif + if(RowsAtCompileTime == 1) + m_storage.resize(size, 1, size); + else + m_storage.resize(size, size, 1); + #ifdef EIGEN_INITIALIZE_COEFFS + if(size_changed) EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED + #endif + } + + /** Resizes the matrix, changing only the number of columns. For the parameter of type NoChange_t, just pass the special value \c NoChange + * as in the example below. + * + * Example: \include Matrix_resize_NoChange_int.cpp + * Output: \verbinclude Matrix_resize_NoChange_int.out + * + * \sa resize(Index,Index) + */ + EIGEN_DEVICE_FUNC + inline void resize(NoChange_t, Index cols) + { + resize(rows(), cols); + } + + /** Resizes the matrix, changing only the number of rows. For the parameter of type NoChange_t, just pass the special value \c NoChange + * as in the example below. + * + * Example: \include Matrix_resize_int_NoChange.cpp + * Output: \verbinclude Matrix_resize_int_NoChange.out + * + * \sa resize(Index,Index) + */ + EIGEN_DEVICE_FUNC + inline void resize(Index rows, NoChange_t) + { + resize(rows, cols()); + } + + /** Resizes \c *this to have the same dimensions as \a other. + * Takes care of doing all the checking that's needed. + * + * Note that copying a row-vector into a vector (and conversely) is allowed. + * The resizing, if any, is then done in the appropriate way so that row-vectors + * remain row-vectors and vectors remain vectors. + */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void resizeLike(const EigenBase<OtherDerived>& _other) + { + const OtherDerived& other = _other.derived(); + internal::check_rows_cols_for_overflow<MaxSizeAtCompileTime>::run(other.rows(), other.cols()); + const Index othersize = other.rows()*other.cols(); + if(RowsAtCompileTime == 1) + { + eigen_assert(other.rows() == 1 || other.cols() == 1); + resize(1, othersize); + } + else if(ColsAtCompileTime == 1) + { + eigen_assert(other.rows() == 1 || other.cols() == 1); + resize(othersize, 1); + } + else resize(other.rows(), other.cols()); + } + + /** Resizes the matrix to \a rows x \a cols while leaving old values untouched. + * + * The method is intended for matrices of dynamic size. If you only want to change the number + * of rows and/or of columns, you can use conservativeResize(NoChange_t, Index) or + * conservativeResize(Index, NoChange_t). + * + * Matrices are resized relative to the top-left element. In case values need to be + * appended to the matrix they will be uninitialized. + */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void conservativeResize(Index rows, Index cols) + { + internal::conservative_resize_like_impl<Derived>::run(*this, rows, cols); + } + + /** Resizes the matrix to \a rows x \a cols while leaving old values untouched. + * + * As opposed to conservativeResize(Index rows, Index cols), this version leaves + * the number of columns unchanged. + * + * In case the matrix is growing, new rows will be uninitialized. + */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void conservativeResize(Index rows, NoChange_t) + { + // Note: see the comment in conservativeResize(Index,Index) + conservativeResize(rows, cols()); + } + + /** Resizes the matrix to \a rows x \a cols while leaving old values untouched. + * + * As opposed to conservativeResize(Index rows, Index cols), this version leaves + * the number of rows unchanged. + * + * In case the matrix is growing, new columns will be uninitialized. + */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void conservativeResize(NoChange_t, Index cols) + { + // Note: see the comment in conservativeResize(Index,Index) + conservativeResize(rows(), cols); + } + + /** Resizes the vector to \a size while retaining old values. + * + * \only_for_vectors. This method does not work for + * partially dynamic matrices when the static dimension is anything other + * than 1. For example it will not work with Matrix<double, 2, Dynamic>. + * + * When values are appended, they will be uninitialized. + */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void conservativeResize(Index size) + { + internal::conservative_resize_like_impl<Derived>::run(*this, size); + } + + /** Resizes the matrix to \a rows x \a cols of \c other, while leaving old values untouched. + * + * The method is intended for matrices of dynamic size. If you only want to change the number + * of rows and/or of columns, you can use conservativeResize(NoChange_t, Index) or + * conservativeResize(Index, NoChange_t). + * + * Matrices are resized relative to the top-left element. In case values need to be + * appended to the matrix they will copied from \c other. + */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void conservativeResizeLike(const DenseBase<OtherDerived>& other) + { + internal::conservative_resize_like_impl<Derived,OtherDerived>::run(*this, other); + } + + /** This is a special case of the templated operator=. Its purpose is to + * prevent a default operator= from hiding the templated operator=. + */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Derived& operator=(const PlainObjectBase& other) + { + return _set(other); + } + + /** \sa MatrixBase::lazyAssign() */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Derived& lazyAssign(const DenseBase<OtherDerived>& other) + { + _resize_to_match(other); + return Base::lazyAssign(other.derived()); + } + + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Derived& operator=(const ReturnByValue<OtherDerived>& func) + { + resize(func.rows(), func.cols()); + return Base::operator=(func); + } + + // Prevent user from trying to instantiate PlainObjectBase objects + // by making all its constructor protected. See bug 1074. + protected: + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE PlainObjectBase() : m_storage() + { +// _check_template_params(); +// EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED + } + +#ifndef EIGEN_PARSED_BY_DOXYGEN + // FIXME is it still needed ? + /** \internal */ + EIGEN_DEVICE_FUNC + explicit PlainObjectBase(internal::constructor_without_unaligned_array_assert) + : m_storage(internal::constructor_without_unaligned_array_assert()) + { +// _check_template_params(); EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED + } +#endif + +#if EIGEN_HAS_RVALUE_REFERENCES + EIGEN_DEVICE_FUNC + PlainObjectBase(PlainObjectBase&& other) EIGEN_NOEXCEPT + : m_storage( std::move(other.m_storage) ) + { + } + + EIGEN_DEVICE_FUNC + PlainObjectBase& operator=(PlainObjectBase&& other) EIGEN_NOEXCEPT + { + using std::swap; + swap(m_storage, other.m_storage); + return *this; + } +#endif + + /** Copy constructor */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE PlainObjectBase(const PlainObjectBase& other) + : Base(), m_storage(other.m_storage) { } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE PlainObjectBase(Index size, Index rows, Index cols) + : m_storage(size, rows, cols) + { +// _check_template_params(); +// EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED + } + + /** \sa PlainObjectBase::operator=(const EigenBase<OtherDerived>&) */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE PlainObjectBase(const DenseBase<OtherDerived> &other) + : m_storage() + { + _check_template_params(); + resizeLike(other); + _set_noalias(other); + } + + /** \sa PlainObjectBase::operator=(const EigenBase<OtherDerived>&) */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE PlainObjectBase(const EigenBase<OtherDerived> &other) + : m_storage() + { + _check_template_params(); + resizeLike(other); + *this = other.derived(); + } + /** \brief Copy constructor with in-place evaluation */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE PlainObjectBase(const ReturnByValue<OtherDerived>& other) + { + _check_template_params(); + // FIXME this does not automatically transpose vectors if necessary + resize(other.rows(), other.cols()); + other.evalTo(this->derived()); + } + + public: + + /** \brief Copies the generic expression \a other into *this. + * \copydetails DenseBase::operator=(const EigenBase<OtherDerived> &other) + */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Derived& operator=(const EigenBase<OtherDerived> &other) + { + _resize_to_match(other); + Base::operator=(other.derived()); + return this->derived(); + } + + /** \name Map + * These are convenience functions returning Map objects. The Map() static functions return unaligned Map objects, + * while the AlignedMap() functions return aligned Map objects and thus should be called only with 16-byte-aligned + * \a data pointers. + * + * \see class Map + */ + //@{ + static inline ConstMapType Map(const Scalar* data) + { return ConstMapType(data); } + static inline MapType Map(Scalar* data) + { return MapType(data); } + static inline ConstMapType Map(const Scalar* data, Index size) + { return ConstMapType(data, size); } + static inline MapType Map(Scalar* data, Index size) + { return MapType(data, size); } + static inline ConstMapType Map(const Scalar* data, Index rows, Index cols) + { return ConstMapType(data, rows, cols); } + static inline MapType Map(Scalar* data, Index rows, Index cols) + { return MapType(data, rows, cols); } + + static inline ConstAlignedMapType MapAligned(const Scalar* data) + { return ConstAlignedMapType(data); } + static inline AlignedMapType MapAligned(Scalar* data) + { return AlignedMapType(data); } + static inline ConstAlignedMapType MapAligned(const Scalar* data, Index size) + { return ConstAlignedMapType(data, size); } + static inline AlignedMapType MapAligned(Scalar* data, Index size) + { return AlignedMapType(data, size); } + static inline ConstAlignedMapType MapAligned(const Scalar* data, Index rows, Index cols) + { return ConstAlignedMapType(data, rows, cols); } + static inline AlignedMapType MapAligned(Scalar* data, Index rows, Index cols) + { return AlignedMapType(data, rows, cols); } + + template<int Outer, int Inner> + static inline typename StridedConstMapType<Stride<Outer, Inner> >::type Map(const Scalar* data, const Stride<Outer, Inner>& stride) + { return typename StridedConstMapType<Stride<Outer, Inner> >::type(data, stride); } + template<int Outer, int Inner> + static inline typename StridedMapType<Stride<Outer, Inner> >::type Map(Scalar* data, const Stride<Outer, Inner>& stride) + { return typename StridedMapType<Stride<Outer, Inner> >::type(data, stride); } + template<int Outer, int Inner> + static inline typename StridedConstMapType<Stride<Outer, Inner> >::type Map(const Scalar* data, Index size, const Stride<Outer, Inner>& stride) + { return typename StridedConstMapType<Stride<Outer, Inner> >::type(data, size, stride); } + template<int Outer, int Inner> + static inline typename StridedMapType<Stride<Outer, Inner> >::type Map(Scalar* data, Index size, const Stride<Outer, Inner>& stride) + { return typename StridedMapType<Stride<Outer, Inner> >::type(data, size, stride); } + template<int Outer, int Inner> + static inline typename StridedConstMapType<Stride<Outer, Inner> >::type Map(const Scalar* data, Index rows, Index cols, const Stride<Outer, Inner>& stride) + { return typename StridedConstMapType<Stride<Outer, Inner> >::type(data, rows, cols, stride); } + template<int Outer, int Inner> + static inline typename StridedMapType<Stride<Outer, Inner> >::type Map(Scalar* data, Index rows, Index cols, const Stride<Outer, Inner>& stride) + { return typename StridedMapType<Stride<Outer, Inner> >::type(data, rows, cols, stride); } + + template<int Outer, int Inner> + static inline typename StridedConstAlignedMapType<Stride<Outer, Inner> >::type MapAligned(const Scalar* data, const Stride<Outer, Inner>& stride) + { return typename StridedConstAlignedMapType<Stride<Outer, Inner> >::type(data, stride); } + template<int Outer, int Inner> + static inline typename StridedAlignedMapType<Stride<Outer, Inner> >::type MapAligned(Scalar* data, const Stride<Outer, Inner>& stride) + { return typename StridedAlignedMapType<Stride<Outer, Inner> >::type(data, stride); } + template<int Outer, int Inner> + static inline typename StridedConstAlignedMapType<Stride<Outer, Inner> >::type MapAligned(const Scalar* data, Index size, const Stride<Outer, Inner>& stride) + { return typename StridedConstAlignedMapType<Stride<Outer, Inner> >::type(data, size, stride); } + template<int Outer, int Inner> + static inline typename StridedAlignedMapType<Stride<Outer, Inner> >::type MapAligned(Scalar* data, Index size, const Stride<Outer, Inner>& stride) + { return typename StridedAlignedMapType<Stride<Outer, Inner> >::type(data, size, stride); } + template<int Outer, int Inner> + static inline typename StridedConstAlignedMapType<Stride<Outer, Inner> >::type MapAligned(const Scalar* data, Index rows, Index cols, const Stride<Outer, Inner>& stride) + { return typename StridedConstAlignedMapType<Stride<Outer, Inner> >::type(data, rows, cols, stride); } + template<int Outer, int Inner> + static inline typename StridedAlignedMapType<Stride<Outer, Inner> >::type MapAligned(Scalar* data, Index rows, Index cols, const Stride<Outer, Inner>& stride) + { return typename StridedAlignedMapType<Stride<Outer, Inner> >::type(data, rows, cols, stride); } + //@} + + using Base::setConstant; + EIGEN_DEVICE_FUNC Derived& setConstant(Index size, const Scalar& val); + EIGEN_DEVICE_FUNC Derived& setConstant(Index rows, Index cols, const Scalar& val); + + using Base::setZero; + EIGEN_DEVICE_FUNC Derived& setZero(Index size); + EIGEN_DEVICE_FUNC Derived& setZero(Index rows, Index cols); + + using Base::setOnes; + EIGEN_DEVICE_FUNC Derived& setOnes(Index size); + EIGEN_DEVICE_FUNC Derived& setOnes(Index rows, Index cols); + + using Base::setRandom; + Derived& setRandom(Index size); + Derived& setRandom(Index rows, Index cols); + + #ifdef EIGEN_PLAINOBJECTBASE_PLUGIN + #include EIGEN_PLAINOBJECTBASE_PLUGIN + #endif + + protected: + /** \internal Resizes *this in preparation for assigning \a other to it. + * Takes care of doing all the checking that's needed. + * + * Note that copying a row-vector into a vector (and conversely) is allowed. + * The resizing, if any, is then done in the appropriate way so that row-vectors + * remain row-vectors and vectors remain vectors. + */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _resize_to_match(const EigenBase<OtherDerived>& other) + { + #ifdef EIGEN_NO_AUTOMATIC_RESIZING + eigen_assert((this->size()==0 || (IsVectorAtCompileTime ? (this->size() == other.size()) + : (rows() == other.rows() && cols() == other.cols()))) + && "Size mismatch. Automatic resizing is disabled because EIGEN_NO_AUTOMATIC_RESIZING is defined"); + EIGEN_ONLY_USED_FOR_DEBUG(other); + #else + resizeLike(other); + #endif + } + + /** + * \brief Copies the value of the expression \a other into \c *this with automatic resizing. + * + * *this might be resized to match the dimensions of \a other. If *this was a null matrix (not already initialized), + * it will be initialized. + * + * Note that copying a row-vector into a vector (and conversely) is allowed. + * The resizing, if any, is then done in the appropriate way so that row-vectors + * remain row-vectors and vectors remain vectors. + * + * \sa operator=(const MatrixBase<OtherDerived>&), _set_noalias() + * + * \internal + */ + // aliasing is dealt once in internall::call_assignment + // so at this stage we have to assume aliasing... and resising has to be done later. + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Derived& _set(const DenseBase<OtherDerived>& other) + { + internal::call_assignment(this->derived(), other.derived()); + return this->derived(); + } + + /** \internal Like _set() but additionally makes the assumption that no aliasing effect can happen (which + * is the case when creating a new matrix) so one can enforce lazy evaluation. + * + * \sa operator=(const MatrixBase<OtherDerived>&), _set() + */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Derived& _set_noalias(const DenseBase<OtherDerived>& other) + { + // I don't think we need this resize call since the lazyAssign will anyways resize + // and lazyAssign will be called by the assign selector. + //_resize_to_match(other); + // the 'false' below means to enforce lazy evaluation. We don't use lazyAssign() because + // it wouldn't allow to copy a row-vector into a column-vector. + internal::call_assignment_no_alias(this->derived(), other.derived(), internal::assign_op<Scalar,typename OtherDerived::Scalar>()); + return this->derived(); + } + + template<typename T0, typename T1> + 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), + FLOATING_POINT_ARGUMENT_PASSED__INTEGER_WAS_EXPECTED) + resize(rows,cols); + } + + template<typename T0, typename T1> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _init2(const T0& val0, const T1& val1, typename internal::enable_if<Base::SizeAtCompileTime==2,T0>::type* = 0) + { + EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, 2) + m_storage.data()[0] = Scalar(val0); + m_storage.data()[1] = Scalar(val1); + } + + template<typename T0, typename T1> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _init2(const Index& val0, const Index& val1, + typename internal::enable_if< (!internal::is_same<Index,Scalar>::value) + && (internal::is_same<T0,Index>::value) + && (internal::is_same<T1,Index>::value) + && Base::SizeAtCompileTime==2,T1>::type* = 0) + { + EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, 2) + m_storage.data()[0] = Scalar(val0); + m_storage.data()[1] = Scalar(val1); + } + + // The argument is convertible to the Index type and we either have a non 1x1 Matrix, or a dynamic-sized Array, + // then the argument is meant to be the size of the object. + template<typename T> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _init1(Index size, typename internal::enable_if< (Base::SizeAtCompileTime!=1 || !internal::is_convertible<T, Scalar>::value) + && ((!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, + FLOATING_POINT_ARGUMENT_PASSED__INTEGER_WAS_EXPECTED) + resize(size); + } + + // We have a 1x1 matrix/array => the argument is interpreted as the value of the unique coefficient (case where scalar type can be implicitely converted) + template<typename T> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _init1(const Scalar& val0, typename internal::enable_if<Base::SizeAtCompileTime==1 && internal::is_convertible<T, Scalar>::value,T>::type* = 0) + { + EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, 1) + m_storage.data()[0] = val0; + } + + // We have a 1x1 matrix/array => the argument is interpreted as the value of the unique coefficient (case where scalar type match the index type) + template<typename T> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _init1(const Index& val0, + typename internal::enable_if< (!internal::is_same<Index,Scalar>::value) + && (internal::is_same<Index,T>::value) + && Base::SizeAtCompileTime==1 + && internal::is_convertible<T, Scalar>::value,T*>::type* = 0) + { + EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, 1) + m_storage.data()[0] = Scalar(val0); + } + + // Initialize a fixed size matrix from a pointer to raw data + template<typename T> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _init1(const Scalar* data){ + this->_set_noalias(ConstMapType(data)); + } + + // Initialize an arbitrary matrix from a dense expression + template<typename T, typename OtherDerived> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _init1(const DenseBase<OtherDerived>& other){ + this->_set_noalias(other); + } + + // Initialize an arbitrary matrix from an object convertible to the Derived type. + template<typename T> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _init1(const Derived& other){ + this->_set_noalias(other); + } + + // Initialize an arbitrary matrix from a generic Eigen expression + template<typename T, typename OtherDerived> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _init1(const EigenBase<OtherDerived>& other){ + this->derived() = other; + } + + template<typename T, typename OtherDerived> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _init1(const ReturnByValue<OtherDerived>& other) + { + resize(other.rows(), other.cols()); + other.evalTo(this->derived()); + } + + template<typename T, typename OtherDerived, int ColsAtCompileTime> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _init1(const RotationBase<OtherDerived,ColsAtCompileTime>& r) + { + this->derived() = r; + } + + // For fixed-size Array<Scalar,...> + template<typename T> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _init1(const Scalar& val0, + typename internal::enable_if< Base::SizeAtCompileTime!=Dynamic + && Base::SizeAtCompileTime!=1 + && internal::is_convertible<T, Scalar>::value + && internal::is_same<typename internal::traits<Derived>::XprKind,ArrayXpr>::value,T>::type* = 0) + { + Base::setConstant(val0); + } + + // For fixed-size Array<Index,...> + template<typename T> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _init1(const Index& val0, + typename internal::enable_if< (!internal::is_same<Index,Scalar>::value) + && (internal::is_same<Index,T>::value) + && Base::SizeAtCompileTime!=Dynamic + && Base::SizeAtCompileTime!=1 + && internal::is_convertible<T, Scalar>::value + && internal::is_same<typename internal::traits<Derived>::XprKind,ArrayXpr>::value,T*>::type* = 0) + { + Base::setConstant(val0); + } + + template<typename MatrixTypeA, typename MatrixTypeB, bool SwapPointers> + friend struct internal::matrix_swap_impl; + + public: + +#ifndef EIGEN_PARSED_BY_DOXYGEN + /** \internal + * \brief Override DenseBase::swap() since for dynamic-sized matrices + * of same type it is enough to swap the data pointers. + */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + void swap(DenseBase<OtherDerived> & other) + { + enum { SwapPointers = internal::is_same<Derived, OtherDerived>::value && Base::SizeAtCompileTime==Dynamic }; + internal::matrix_swap_impl<Derived, OtherDerived, bool(SwapPointers)>::run(this->derived(), other.derived()); + } + + /** \internal + * \brief const version forwarded to DenseBase::swap + */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + void swap(DenseBase<OtherDerived> const & other) + { Base::swap(other.derived()); } + + EIGEN_DEVICE_FUNC + static EIGEN_STRONG_INLINE void _check_template_params() + { + EIGEN_STATIC_ASSERT((EIGEN_IMPLIES(MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1, (Options&RowMajor)==RowMajor) + && EIGEN_IMPLIES(MaxColsAtCompileTime==1 && MaxRowsAtCompileTime!=1, (Options&RowMajor)==0) + && ((RowsAtCompileTime == Dynamic) || (RowsAtCompileTime >= 0)) + && ((ColsAtCompileTime == Dynamic) || (ColsAtCompileTime >= 0)) + && ((MaxRowsAtCompileTime == Dynamic) || (MaxRowsAtCompileTime >= 0)) + && ((MaxColsAtCompileTime == Dynamic) || (MaxColsAtCompileTime >= 0)) + && (MaxRowsAtCompileTime == RowsAtCompileTime || RowsAtCompileTime==Dynamic) + && (MaxColsAtCompileTime == ColsAtCompileTime || ColsAtCompileTime==Dynamic) + && (Options & (DontAlign|RowMajor)) == Options), + INVALID_MATRIX_TEMPLATE_PARAMETERS) + } + + enum { IsPlainObjectBase = 1 }; +#endif +}; + +namespace internal { + +template <typename Derived, typename OtherDerived, bool IsVector> +struct conservative_resize_like_impl +{ + static void run(DenseBase<Derived>& _this, Index rows, Index cols) + { + if (_this.rows() == rows && _this.cols() == cols) return; + EIGEN_STATIC_ASSERT_DYNAMIC_SIZE(Derived) + + if ( ( Derived::IsRowMajor && _this.cols() == cols) || // row-major and we change only the number of rows + (!Derived::IsRowMajor && _this.rows() == rows) ) // column-major and we change only the number of columns + { + internal::check_rows_cols_for_overflow<Derived::MaxSizeAtCompileTime>::run(rows, cols); + _this.derived().m_storage.conservativeResize(rows*cols,rows,cols); + } + else + { + // The storage order does not allow us to use reallocation. + typename Derived::PlainObject tmp(rows,cols); + const Index common_rows = numext::mini(rows, _this.rows()); + const Index common_cols = numext::mini(cols, _this.cols()); + tmp.block(0,0,common_rows,common_cols) = _this.block(0,0,common_rows,common_cols); + _this.derived().swap(tmp); + } + } + + static void run(DenseBase<Derived>& _this, const DenseBase<OtherDerived>& other) + { + if (_this.rows() == other.rows() && _this.cols() == other.cols()) return; + + // Note: Here is space for improvement. Basically, for conservativeResize(Index,Index), + // neither RowsAtCompileTime or ColsAtCompileTime must be Dynamic. If only one of the + // dimensions is dynamic, one could use either conservativeResize(Index rows, NoChange_t) or + // conservativeResize(NoChange_t, Index cols). For these methods new static asserts like + // EIGEN_STATIC_ASSERT_DYNAMIC_ROWS and EIGEN_STATIC_ASSERT_DYNAMIC_COLS would be good. + EIGEN_STATIC_ASSERT_DYNAMIC_SIZE(Derived) + EIGEN_STATIC_ASSERT_DYNAMIC_SIZE(OtherDerived) + + if ( ( Derived::IsRowMajor && _this.cols() == other.cols()) || // row-major and we change only the number of rows + (!Derived::IsRowMajor && _this.rows() == other.rows()) ) // column-major and we change only the number of columns + { + const Index new_rows = other.rows() - _this.rows(); + const Index new_cols = other.cols() - _this.cols(); + _this.derived().m_storage.conservativeResize(other.size(),other.rows(),other.cols()); + if (new_rows>0) + _this.bottomRightCorner(new_rows, other.cols()) = other.bottomRows(new_rows); + else if (new_cols>0) + _this.bottomRightCorner(other.rows(), new_cols) = other.rightCols(new_cols); + } + else + { + // The storage order does not allow us to use reallocation. + typename Derived::PlainObject tmp(other); + const Index common_rows = numext::mini(tmp.rows(), _this.rows()); + const Index common_cols = numext::mini(tmp.cols(), _this.cols()); + tmp.block(0,0,common_rows,common_cols) = _this.block(0,0,common_rows,common_cols); + _this.derived().swap(tmp); + } + } +}; + +// Here, the specialization for vectors inherits from the general matrix case +// to allow calling .conservativeResize(rows,cols) on vectors. +template <typename Derived, typename OtherDerived> +struct conservative_resize_like_impl<Derived,OtherDerived,true> + : conservative_resize_like_impl<Derived,OtherDerived,false> +{ + using conservative_resize_like_impl<Derived,OtherDerived,false>::run; + + static void run(DenseBase<Derived>& _this, Index size) + { + const Index new_rows = Derived::RowsAtCompileTime==1 ? 1 : size; + const Index new_cols = Derived::RowsAtCompileTime==1 ? size : 1; + _this.derived().m_storage.conservativeResize(size,new_rows,new_cols); + } + + static void run(DenseBase<Derived>& _this, const DenseBase<OtherDerived>& other) + { + if (_this.rows() == other.rows() && _this.cols() == other.cols()) return; + + const Index num_new_elements = other.size() - _this.size(); + + const Index new_rows = Derived::RowsAtCompileTime==1 ? 1 : other.rows(); + const Index new_cols = Derived::RowsAtCompileTime==1 ? other.cols() : 1; + _this.derived().m_storage.conservativeResize(other.size(),new_rows,new_cols); + + if (num_new_elements > 0) + _this.tail(num_new_elements) = other.tail(num_new_elements); + } +}; + +template<typename MatrixTypeA, typename MatrixTypeB, bool SwapPointers> +struct matrix_swap_impl +{ + EIGEN_DEVICE_FUNC + static inline void run(MatrixTypeA& a, MatrixTypeB& b) + { + a.base().swap(b); + } +}; + +template<typename MatrixTypeA, typename MatrixTypeB> +struct matrix_swap_impl<MatrixTypeA, MatrixTypeB, true> +{ + EIGEN_DEVICE_FUNC + static inline void run(MatrixTypeA& a, MatrixTypeB& b) + { + static_cast<typename MatrixTypeA::Base&>(a).m_storage.swap(static_cast<typename MatrixTypeB::Base&>(b).m_storage); + } +}; + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_DENSESTORAGEBASE_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/Product.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Product.h new file mode 100644 index 000000000..ae0c94b38 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Product.h @@ -0,0 +1,186 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2011 Gael Guennebaud <gael.guennebaud@inria.fr> +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_PRODUCT_H +#define EIGEN_PRODUCT_H + +namespace Eigen { + +template<typename Lhs, typename Rhs, int Option, typename StorageKind> class ProductImpl; + +namespace internal { + +template<typename Lhs, typename Rhs, int Option> +struct traits<Product<Lhs, Rhs, Option> > +{ + typedef typename remove_all<Lhs>::type LhsCleaned; + typedef typename remove_all<Rhs>::type RhsCleaned; + typedef traits<LhsCleaned> LhsTraits; + typedef traits<RhsCleaned> RhsTraits; + + typedef MatrixXpr XprKind; + + typedef typename ScalarBinaryOpTraits<typename traits<LhsCleaned>::Scalar, typename traits<RhsCleaned>::Scalar>::ReturnType Scalar; + typedef typename product_promote_storage_type<typename LhsTraits::StorageKind, + typename RhsTraits::StorageKind, + internal::product_type<Lhs,Rhs>::ret>::ret StorageKind; + typedef typename promote_index_type<typename LhsTraits::StorageIndex, + typename RhsTraits::StorageIndex>::type StorageIndex; + + enum { + RowsAtCompileTime = LhsTraits::RowsAtCompileTime, + ColsAtCompileTime = RhsTraits::ColsAtCompileTime, + MaxRowsAtCompileTime = LhsTraits::MaxRowsAtCompileTime, + MaxColsAtCompileTime = RhsTraits::MaxColsAtCompileTime, + + // FIXME: only needed by GeneralMatrixMatrixTriangular + InnerSize = EIGEN_SIZE_MIN_PREFER_FIXED(LhsTraits::ColsAtCompileTime, RhsTraits::RowsAtCompileTime), + + // The storage order is somewhat arbitrary here. The correct one will be determined through the evaluator. + Flags = (MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1) ? RowMajorBit + : (MaxColsAtCompileTime==1 && MaxRowsAtCompileTime!=1) ? 0 + : ( ((LhsTraits::Flags&NoPreferredStorageOrderBit) && (RhsTraits::Flags&RowMajorBit)) + || ((RhsTraits::Flags&NoPreferredStorageOrderBit) && (LhsTraits::Flags&RowMajorBit)) ) ? RowMajorBit + : NoPreferredStorageOrderBit + }; +}; + +} // end namespace internal + +/** \class Product + * \ingroup Core_Module + * + * \brief Expression of the product of two arbitrary matrices or vectors + * + * \tparam _Lhs the type of the left-hand side expression + * \tparam _Rhs the type of the right-hand side expression + * + * This class represents an expression of the product of two arbitrary matrices. + * + * The other template parameters are: + * \tparam Option can be DefaultProduct, AliasFreeProduct, or LazyProduct + * + */ +template<typename _Lhs, typename _Rhs, int Option> +class Product : public ProductImpl<_Lhs,_Rhs,Option, + typename internal::product_promote_storage_type<typename internal::traits<_Lhs>::StorageKind, + typename internal::traits<_Rhs>::StorageKind, + internal::product_type<_Lhs,_Rhs>::ret>::ret> +{ + public: + + typedef _Lhs Lhs; + typedef _Rhs Rhs; + + typedef typename ProductImpl< + Lhs, Rhs, Option, + typename internal::product_promote_storage_type<typename internal::traits<Lhs>::StorageKind, + typename internal::traits<Rhs>::StorageKind, + internal::product_type<Lhs,Rhs>::ret>::ret>::Base Base; + EIGEN_GENERIC_PUBLIC_INTERFACE(Product) + + typedef typename internal::ref_selector<Lhs>::type LhsNested; + typedef typename internal::ref_selector<Rhs>::type RhsNested; + 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_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 inline Index rows() const { return m_lhs.rows(); } + EIGEN_DEVICE_FUNC 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; } + + protected: + + LhsNested m_lhs; + RhsNested m_rhs; +}; + +namespace internal { + +template<typename Lhs, typename Rhs, int Option, int ProductTag = internal::product_type<Lhs,Rhs>::ret> +class dense_product_base + : public internal::dense_xpr_base<Product<Lhs,Rhs,Option> >::type +{}; + +/** Convertion to scalar for inner-products */ +template<typename Lhs, typename Rhs, int Option> +class dense_product_base<Lhs, Rhs, Option, InnerProduct> + : public internal::dense_xpr_base<Product<Lhs,Rhs,Option> >::type +{ + typedef Product<Lhs,Rhs,Option> ProductXpr; + typedef typename internal::dense_xpr_base<ProductXpr>::type Base; +public: + using Base::derived; + typedef typename Base::Scalar Scalar; + + operator const Scalar() const + { + return internal::evaluator<ProductXpr>(derived()).coeff(0,0); + } +}; + +} // namespace internal + +// Generic API dispatcher +template<typename Lhs, typename Rhs, int Option, typename StorageKind> +class ProductImpl : public internal::generic_xpr_base<Product<Lhs,Rhs,Option>, MatrixXpr, StorageKind>::type +{ + public: + typedef typename internal::generic_xpr_base<Product<Lhs,Rhs,Option>, MatrixXpr, StorageKind>::type Base; +}; + +template<typename Lhs, typename Rhs, int Option> +class ProductImpl<Lhs,Rhs,Option,Dense> + : public internal::dense_product_base<Lhs,Rhs,Option> +{ + typedef Product<Lhs, Rhs, Option> Derived; + + public: + + typedef typename internal::dense_product_base<Lhs, Rhs, Option> Base; + EIGEN_DENSE_PUBLIC_INTERFACE(Derived) + protected: + enum { + IsOneByOne = (RowsAtCompileTime == 1 || RowsAtCompileTime == Dynamic) && + (ColsAtCompileTime == 1 || ColsAtCompileTime == Dynamic), + EnableCoeff = IsOneByOne || Option==LazyProduct + }; + + public: + + EIGEN_DEVICE_FUNC Scalar coeff(Index row, Index col) const + { + EIGEN_STATIC_ASSERT(EnableCoeff, THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS); + eigen_assert( (Option==LazyProduct) || (this->rows() == 1 && this->cols() == 1) ); + + return internal::evaluator<Derived>(derived()).coeff(row,col); + } + + EIGEN_DEVICE_FUNC Scalar coeff(Index i) const + { + EIGEN_STATIC_ASSERT(EnableCoeff, THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS); + eigen_assert( (Option==LazyProduct) || (this->rows() == 1 && this->cols() == 1) ); + + return internal::evaluator<Derived>(derived()).coeff(i); + } + + +}; + +} // end namespace Eigen + +#endif // EIGEN_PRODUCT_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/ProductEvaluators.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/ProductEvaluators.h new file mode 100644 index 000000000..c42725dbd --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/ProductEvaluators.h @@ -0,0 +1,1105 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com> +// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2011 Jitse Niesen <jitse@maths.leeds.ac.uk> +// +// 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_PRODUCTEVALUATORS_H +#define EIGEN_PRODUCTEVALUATORS_H + +namespace Eigen { + +namespace internal { + +/** \internal + * Evaluator of a product expression. + * Since products require special treatments to handle all possible cases, + * we simply deffer the evaluation logic to a product_evaluator class + * which offers more partial specialization possibilities. + * + * \sa class product_evaluator + */ +template<typename Lhs, typename Rhs, int Options> +struct evaluator<Product<Lhs, Rhs, Options> > + : public product_evaluator<Product<Lhs, Rhs, Options> > +{ + typedef Product<Lhs, Rhs, Options> XprType; + typedef product_evaluator<XprType> Base; + + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr) : Base(xpr) {} +}; + +// Catch "scalar * ( A * B )" and transform it to "(A*scalar) * B" +// TODO we should apply that rule only if that's really helpful +template<typename Lhs, typename Rhs, typename Scalar1, typename Scalar2, typename Plain1> +struct evaluator_assume_aliasing<CwiseBinaryOp<internal::scalar_product_op<Scalar1,Scalar2>, + const CwiseNullaryOp<internal::scalar_constant_op<Scalar1>, Plain1>, + const Product<Lhs, Rhs, DefaultProduct> > > +{ + static const bool value = true; +}; +template<typename Lhs, typename Rhs, typename Scalar1, typename Scalar2, typename Plain1> +struct evaluator<CwiseBinaryOp<internal::scalar_product_op<Scalar1,Scalar2>, + const CwiseNullaryOp<internal::scalar_constant_op<Scalar1>, Plain1>, + const Product<Lhs, Rhs, DefaultProduct> > > + : public evaluator<Product<EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(Scalar1,Lhs,product), Rhs, DefaultProduct> > +{ + typedef CwiseBinaryOp<internal::scalar_product_op<Scalar1,Scalar2>, + const CwiseNullaryOp<internal::scalar_constant_op<Scalar1>, Plain1>, + const Product<Lhs, Rhs, DefaultProduct> > XprType; + typedef evaluator<Product<EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(Scalar1,Lhs,product), Rhs, DefaultProduct> > Base; + + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr) + : Base(xpr.lhs().functor().m_other * xpr.rhs().lhs() * xpr.rhs().rhs()) + {} +}; + + +template<typename Lhs, typename Rhs, int DiagIndex> +struct evaluator<Diagonal<const Product<Lhs, Rhs, DefaultProduct>, DiagIndex> > + : public evaluator<Diagonal<const Product<Lhs, Rhs, LazyProduct>, DiagIndex> > +{ + typedef Diagonal<const Product<Lhs, Rhs, DefaultProduct>, DiagIndex> XprType; + typedef evaluator<Diagonal<const Product<Lhs, Rhs, LazyProduct>, DiagIndex> > Base; + + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr) + : Base(Diagonal<const Product<Lhs, Rhs, LazyProduct>, DiagIndex>( + Product<Lhs, Rhs, LazyProduct>(xpr.nestedExpression().lhs(), xpr.nestedExpression().rhs()), + xpr.index() )) + {} +}; + + +// Helper class to perform a matrix product with the destination at hand. +// Depending on the sizes of the factors, there are different evaluation strategies +// as controlled by internal::product_type. +template< typename Lhs, typename Rhs, + typename LhsShape = typename evaluator_traits<Lhs>::Shape, + typename RhsShape = typename evaluator_traits<Rhs>::Shape, + int ProductType = internal::product_type<Lhs,Rhs>::value> +struct generic_product_impl; + +template<typename Lhs, typename Rhs> +struct evaluator_assume_aliasing<Product<Lhs, Rhs, DefaultProduct> > { + static const bool value = true; +}; + +// This is the default evaluator implementation for products: +// It creates a temporary and call generic_product_impl +template<typename Lhs, typename Rhs, int Options, int ProductTag, typename LhsShape, typename RhsShape> +struct product_evaluator<Product<Lhs, Rhs, Options>, ProductTag, LhsShape, RhsShape> + : public evaluator<typename Product<Lhs, Rhs, Options>::PlainObject> +{ + typedef Product<Lhs, Rhs, Options> XprType; + typedef typename XprType::PlainObject PlainObject; + typedef evaluator<PlainObject> Base; + enum { + Flags = Base::Flags | EvalBeforeNestingBit + }; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit product_evaluator(const XprType& xpr) + : m_result(xpr.rows(), xpr.cols()) + { + ::new (static_cast<Base*>(this)) Base(m_result); + +// FIXME shall we handle nested_eval here?, +// if so, then we must take care at removing the call to nested_eval in the specializations (e.g., in permutation_matrix_product, transposition_matrix_product, etc.) +// typedef typename internal::nested_eval<Lhs,Rhs::ColsAtCompileTime>::type LhsNested; +// typedef typename internal::nested_eval<Rhs,Lhs::RowsAtCompileTime>::type RhsNested; +// typedef typename internal::remove_all<LhsNested>::type LhsNestedCleaned; +// typedef typename internal::remove_all<RhsNested>::type RhsNestedCleaned; +// +// const LhsNested lhs(xpr.lhs()); +// const RhsNested rhs(xpr.rhs()); +// +// generic_product_impl<LhsNestedCleaned, RhsNestedCleaned>::evalTo(m_result, lhs, rhs); + + generic_product_impl<Lhs, Rhs, LhsShape, RhsShape, ProductTag>::evalTo(m_result, xpr.lhs(), xpr.rhs()); + } + +protected: + PlainObject m_result; +}; + +// The following three shortcuts are enabled only if the scalar types match excatly. +// TODO: we could enable them for different scalar types when the product is not vectorized. + +// Dense = Product +template< typename DstXprType, typename Lhs, typename Rhs, int Options, typename Scalar> +struct Assignment<DstXprType, Product<Lhs,Rhs,Options>, internal::assign_op<Scalar,Scalar>, Dense2Dense, + typename enable_if<(Options==DefaultProduct || Options==AliasFreeProduct)>::type> +{ + typedef Product<Lhs,Rhs,Options> SrcXprType; + static EIGEN_STRONG_INLINE + void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,Scalar> &) + { + Index dstRows = src.rows(); + Index dstCols = src.cols(); + if((dst.rows()!=dstRows) || (dst.cols()!=dstCols)) + dst.resize(dstRows, dstCols); + // FIXME shall we handle nested_eval here? + generic_product_impl<Lhs, Rhs>::evalTo(dst, src.lhs(), src.rhs()); + } +}; + +// Dense += Product +template< typename DstXprType, typename Lhs, typename Rhs, int Options, typename Scalar> +struct Assignment<DstXprType, Product<Lhs,Rhs,Options>, internal::add_assign_op<Scalar,Scalar>, Dense2Dense, + typename enable_if<(Options==DefaultProduct || Options==AliasFreeProduct)>::type> +{ + typedef Product<Lhs,Rhs,Options> SrcXprType; + static EIGEN_STRONG_INLINE + void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op<Scalar,Scalar> &) + { + eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); + // FIXME shall we handle nested_eval here? + generic_product_impl<Lhs, Rhs>::addTo(dst, src.lhs(), src.rhs()); + } +}; + +// Dense -= Product +template< typename DstXprType, typename Lhs, typename Rhs, int Options, typename Scalar> +struct Assignment<DstXprType, Product<Lhs,Rhs,Options>, internal::sub_assign_op<Scalar,Scalar>, Dense2Dense, + typename enable_if<(Options==DefaultProduct || Options==AliasFreeProduct)>::type> +{ + typedef Product<Lhs,Rhs,Options> SrcXprType; + static EIGEN_STRONG_INLINE + void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op<Scalar,Scalar> &) + { + eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); + // FIXME shall we handle nested_eval here? + generic_product_impl<Lhs, Rhs>::subTo(dst, src.lhs(), src.rhs()); + } +}; + + +// Dense ?= scalar * Product +// TODO we should apply that rule if that's really helpful +// for instance, this is not good for inner products +template< typename DstXprType, typename Lhs, typename Rhs, typename AssignFunc, typename Scalar, typename ScalarBis, typename Plain> +struct Assignment<DstXprType, CwiseBinaryOp<internal::scalar_product_op<ScalarBis,Scalar>, const CwiseNullaryOp<internal::scalar_constant_op<ScalarBis>,Plain>, + const Product<Lhs,Rhs,DefaultProduct> >, AssignFunc, Dense2Dense> +{ + typedef CwiseBinaryOp<internal::scalar_product_op<ScalarBis,Scalar>, + const CwiseNullaryOp<internal::scalar_constant_op<ScalarBis>,Plain>, + const Product<Lhs,Rhs,DefaultProduct> > SrcXprType; + static EIGEN_STRONG_INLINE + void run(DstXprType &dst, const SrcXprType &src, const AssignFunc& func) + { + call_assignment_no_alias(dst, (src.lhs().functor().m_other * src.rhs().lhs())*src.rhs().rhs(), func); + } +}; + +//---------------------------------------- +// Catch "Dense ?= xpr + Product<>" expression to save one temporary +// FIXME we could probably enable these rules for any product, i.e., not only Dense and DefaultProduct + +template<typename OtherXpr, typename Lhs, typename Rhs> +struct evaluator_assume_aliasing<CwiseBinaryOp<internal::scalar_sum_op<typename OtherXpr::Scalar,typename Product<Lhs,Rhs,DefaultProduct>::Scalar>, const OtherXpr, + const Product<Lhs,Rhs,DefaultProduct> >, DenseShape > { + static const bool value = true; +}; + +template<typename OtherXpr, typename Lhs, typename Rhs> +struct evaluator_assume_aliasing<CwiseBinaryOp<internal::scalar_difference_op<typename OtherXpr::Scalar,typename Product<Lhs,Rhs,DefaultProduct>::Scalar>, const OtherXpr, + const Product<Lhs,Rhs,DefaultProduct> >, DenseShape > { + static const bool value = true; +}; + +template<typename DstXprType, typename OtherXpr, typename ProductType, typename Func1, typename Func2> +struct assignment_from_xpr_op_product +{ + template<typename SrcXprType, typename InitialFunc> + static EIGEN_STRONG_INLINE + void run(DstXprType &dst, const SrcXprType &src, const InitialFunc& /*func*/) + { + call_assignment_no_alias(dst, src.lhs(), Func1()); + call_assignment_no_alias(dst, src.rhs(), Func2()); + } +}; + +#define EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(ASSIGN_OP,BINOP,ASSIGN_OP2) \ + template< typename DstXprType, typename OtherXpr, typename Lhs, typename Rhs, typename DstScalar, typename SrcScalar, typename OtherScalar,typename ProdScalar> \ + struct Assignment<DstXprType, CwiseBinaryOp<internal::BINOP<OtherScalar,ProdScalar>, const OtherXpr, \ + const Product<Lhs,Rhs,DefaultProduct> >, internal::ASSIGN_OP<DstScalar,SrcScalar>, Dense2Dense> \ + : assignment_from_xpr_op_product<DstXprType, OtherXpr, Product<Lhs,Rhs,DefaultProduct>, internal::ASSIGN_OP<DstScalar,OtherScalar>, internal::ASSIGN_OP2<DstScalar,ProdScalar> > \ + {} + +EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(assign_op, scalar_sum_op,add_assign_op); +EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(add_assign_op,scalar_sum_op,add_assign_op); +EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(sub_assign_op,scalar_sum_op,sub_assign_op); + +EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(assign_op, scalar_difference_op,sub_assign_op); +EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(add_assign_op,scalar_difference_op,sub_assign_op); +EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(sub_assign_op,scalar_difference_op,add_assign_op); + +//---------------------------------------- + +template<typename Lhs, typename Rhs> +struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,InnerProduct> +{ + template<typename Dst> + static inline void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { + dst.coeffRef(0,0) = (lhs.transpose().cwiseProduct(rhs)).sum(); + } + + template<typename Dst> + static inline void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { + dst.coeffRef(0,0) += (lhs.transpose().cwiseProduct(rhs)).sum(); + } + + template<typename Dst> + static void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { dst.coeffRef(0,0) -= (lhs.transpose().cwiseProduct(rhs)).sum(); } +}; + + +/*********************************************************************** +* Implementation of outer dense * dense vector product +***********************************************************************/ + +// Column major result +template<typename Dst, typename Lhs, typename Rhs, typename Func> +void outer_product_selector_run(Dst& dst, const Lhs &lhs, const Rhs &rhs, const Func& func, const false_type&) +{ + evaluator<Rhs> rhsEval(rhs); + typename nested_eval<Lhs,Rhs::SizeAtCompileTime>::type actual_lhs(lhs); + // FIXME if cols is large enough, then it might be useful to make sure that lhs is sequentially stored + // FIXME not very good if rhs is real and lhs complex while alpha is real too + const Index cols = dst.cols(); + for (Index j=0; j<cols; ++j) + func(dst.col(j), rhsEval.coeff(Index(0),j) * actual_lhs); +} + +// Row major result +template<typename Dst, typename Lhs, typename Rhs, typename Func> +void outer_product_selector_run(Dst& dst, const Lhs &lhs, const Rhs &rhs, const Func& func, const true_type&) +{ + evaluator<Lhs> lhsEval(lhs); + typename nested_eval<Rhs,Lhs::SizeAtCompileTime>::type actual_rhs(rhs); + // FIXME if rows is large enough, then it might be useful to make sure that rhs is sequentially stored + // FIXME not very good if lhs is real and rhs complex while alpha is real too + const Index rows = dst.rows(); + for (Index i=0; i<rows; ++i) + func(dst.row(i), lhsEval.coeff(i,Index(0)) * actual_rhs); +} + +template<typename Lhs, typename Rhs> +struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,OuterProduct> +{ + template<typename T> struct is_row_major : internal::conditional<(int(T::Flags)&RowMajorBit), internal::true_type, internal::false_type>::type {}; + typedef typename Product<Lhs,Rhs>::Scalar Scalar; + + // TODO it would be nice to be able to exploit our *_assign_op functors for that purpose + struct set { template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived() = src; } }; + struct add { template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived() += src; } }; + struct sub { template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived() -= src; } }; + struct adds { + Scalar m_scale; + explicit adds(const Scalar& s) : m_scale(s) {} + template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const { + dst.const_cast_derived() += m_scale * src; + } + }; + + template<typename Dst> + static inline void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { + internal::outer_product_selector_run(dst, lhs, rhs, set(), is_row_major<Dst>()); + } + + template<typename Dst> + static inline void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { + internal::outer_product_selector_run(dst, lhs, rhs, add(), is_row_major<Dst>()); + } + + template<typename Dst> + static inline void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { + internal::outer_product_selector_run(dst, lhs, rhs, sub(), is_row_major<Dst>()); + } + + template<typename Dst> + static inline void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) + { + internal::outer_product_selector_run(dst, lhs, rhs, adds(alpha), is_row_major<Dst>()); + } + +}; + + +// This base class provides default implementations for evalTo, addTo, subTo, in terms of scaleAndAddTo +template<typename Lhs, typename Rhs, typename Derived> +struct generic_product_impl_base +{ + typedef typename Product<Lhs,Rhs>::Scalar Scalar; + + template<typename Dst> + static EIGEN_STRONG_INLINE void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { dst.setZero(); scaleAndAddTo(dst, lhs, rhs, Scalar(1)); } + + template<typename Dst> + static EIGEN_STRONG_INLINE void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { scaleAndAddTo(dst,lhs, rhs, Scalar(1)); } + + template<typename Dst> + static EIGEN_STRONG_INLINE void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { scaleAndAddTo(dst, lhs, rhs, Scalar(-1)); } + + template<typename Dst> + static EIGEN_STRONG_INLINE void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) + { Derived::scaleAndAddTo(dst,lhs,rhs,alpha); } + +}; + +template<typename Lhs, typename Rhs> +struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemvProduct> + : generic_product_impl_base<Lhs,Rhs,generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemvProduct> > +{ + typedef typename nested_eval<Lhs,1>::type LhsNested; + typedef typename nested_eval<Rhs,1>::type RhsNested; + typedef typename Product<Lhs,Rhs>::Scalar Scalar; + enum { Side = Lhs::IsVectorAtCompileTime ? OnTheLeft : OnTheRight }; + typedef typename internal::remove_all<typename internal::conditional<int(Side)==OnTheRight,LhsNested,RhsNested>::type>::type MatrixType; + + template<typename Dest> + static EIGEN_STRONG_INLINE void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) + { + LhsNested actual_lhs(lhs); + RhsNested actual_rhs(rhs); + internal::gemv_dense_selector<Side, + (int(MatrixType::Flags)&RowMajorBit) ? RowMajor : ColMajor, + bool(internal::blas_traits<MatrixType>::HasUsableDirectAccess) + >::run(actual_lhs, actual_rhs, dst, alpha); + } +}; + +template<typename Lhs, typename Rhs> +struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,CoeffBasedProductMode> +{ + typedef typename Product<Lhs,Rhs>::Scalar Scalar; + + template<typename Dst> + static EIGEN_STRONG_INLINE void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { + // Same as: dst.noalias() = lhs.lazyProduct(rhs); + // but easier on the compiler side + call_assignment_no_alias(dst, lhs.lazyProduct(rhs), internal::assign_op<typename Dst::Scalar,Scalar>()); + } + + template<typename Dst> + static EIGEN_STRONG_INLINE void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { + // dst.noalias() += lhs.lazyProduct(rhs); + call_assignment_no_alias(dst, lhs.lazyProduct(rhs), internal::add_assign_op<typename Dst::Scalar,Scalar>()); + } + + template<typename Dst> + static EIGEN_STRONG_INLINE void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { + // dst.noalias() -= lhs.lazyProduct(rhs); + call_assignment_no_alias(dst, lhs.lazyProduct(rhs), internal::sub_assign_op<typename Dst::Scalar,Scalar>()); + } + +// template<typename Dst> +// static inline void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) +// { dst.noalias() += alpha * lhs.lazyProduct(rhs); } +}; + +// This specialization enforces the use of a coefficient-based evaluation strategy +template<typename Lhs, typename Rhs> +struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,LazyCoeffBasedProductMode> + : generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,CoeffBasedProductMode> {}; + +// Case 2: Evaluate coeff by coeff +// +// This is mostly taken from CoeffBasedProduct.h +// The main difference is that we add an extra argument to the etor_product_*_impl::run() function +// for the inner dimension of the product, because evaluator object do not know their size. + +template<int Traversal, int UnrollingIndex, typename Lhs, typename Rhs, typename RetScalar> +struct etor_product_coeff_impl; + +template<int StorageOrder, int UnrollingIndex, typename Lhs, typename Rhs, typename Packet, int LoadMode> +struct etor_product_packet_impl; + +template<typename Lhs, typename Rhs, int ProductTag> +struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape, DenseShape> + : evaluator_base<Product<Lhs, Rhs, LazyProduct> > +{ + typedef Product<Lhs, Rhs, LazyProduct> XprType; + typedef typename XprType::Scalar Scalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + explicit product_evaluator(const XprType& xpr) + : m_lhs(xpr.lhs()), + m_rhs(xpr.rhs()), + m_lhsImpl(m_lhs), // FIXME the creation of the evaluator objects should result in a no-op, but check that! + m_rhsImpl(m_rhs), // Moreover, they are only useful for the packet path, so we could completely disable them when not needed, + // or perhaps declare them on the fly on the packet method... We have experiment to check what's best. + m_innerDim(xpr.lhs().cols()) + { + EIGEN_INTERNAL_CHECK_COST_VALUE(NumTraits<Scalar>::MulCost); + EIGEN_INTERNAL_CHECK_COST_VALUE(NumTraits<Scalar>::AddCost); + EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost); +#if 0 + std::cerr << "LhsOuterStrideBytes= " << LhsOuterStrideBytes << "\n"; + std::cerr << "RhsOuterStrideBytes= " << RhsOuterStrideBytes << "\n"; + std::cerr << "LhsAlignment= " << LhsAlignment << "\n"; + std::cerr << "RhsAlignment= " << RhsAlignment << "\n"; + std::cerr << "CanVectorizeLhs= " << CanVectorizeLhs << "\n"; + std::cerr << "CanVectorizeRhs= " << CanVectorizeRhs << "\n"; + std::cerr << "CanVectorizeInner= " << CanVectorizeInner << "\n"; + std::cerr << "EvalToRowMajor= " << EvalToRowMajor << "\n"; + std::cerr << "Alignment= " << Alignment << "\n"; + std::cerr << "Flags= " << Flags << "\n"; +#endif + } + + // Everything below here is taken from CoeffBasedProduct.h + + typedef typename internal::nested_eval<Lhs,Rhs::ColsAtCompileTime>::type LhsNested; + typedef typename internal::nested_eval<Rhs,Lhs::RowsAtCompileTime>::type RhsNested; + + typedef typename internal::remove_all<LhsNested>::type LhsNestedCleaned; + typedef typename internal::remove_all<RhsNested>::type RhsNestedCleaned; + + typedef evaluator<LhsNestedCleaned> LhsEtorType; + typedef evaluator<RhsNestedCleaned> RhsEtorType; + + enum { + RowsAtCompileTime = LhsNestedCleaned::RowsAtCompileTime, + ColsAtCompileTime = RhsNestedCleaned::ColsAtCompileTime, + InnerSize = EIGEN_SIZE_MIN_PREFER_FIXED(LhsNestedCleaned::ColsAtCompileTime, RhsNestedCleaned::RowsAtCompileTime), + MaxRowsAtCompileTime = LhsNestedCleaned::MaxRowsAtCompileTime, + MaxColsAtCompileTime = RhsNestedCleaned::MaxColsAtCompileTime + }; + + typedef typename find_best_packet<Scalar,RowsAtCompileTime>::type LhsVecPacketType; + typedef typename find_best_packet<Scalar,ColsAtCompileTime>::type RhsVecPacketType; + + enum { + + LhsCoeffReadCost = LhsEtorType::CoeffReadCost, + RhsCoeffReadCost = RhsEtorType::CoeffReadCost, + CoeffReadCost = InnerSize==0 ? NumTraits<Scalar>::ReadCost + : InnerSize == Dynamic ? HugeCost + : InnerSize * (NumTraits<Scalar>::MulCost + LhsCoeffReadCost + RhsCoeffReadCost) + + (InnerSize - 1) * NumTraits<Scalar>::AddCost, + + Unroll = CoeffReadCost <= EIGEN_UNROLLING_LIMIT, + + LhsFlags = LhsEtorType::Flags, + RhsFlags = RhsEtorType::Flags, + + LhsRowMajor = LhsFlags & RowMajorBit, + RhsRowMajor = RhsFlags & RowMajorBit, + + LhsVecPacketSize = unpacket_traits<LhsVecPacketType>::size, + RhsVecPacketSize = unpacket_traits<RhsVecPacketType>::size, + + // Here, we don't care about alignment larger than the usable packet size. + LhsAlignment = EIGEN_PLAIN_ENUM_MIN(LhsEtorType::Alignment,LhsVecPacketSize*int(sizeof(typename LhsNestedCleaned::Scalar))), + RhsAlignment = EIGEN_PLAIN_ENUM_MIN(RhsEtorType::Alignment,RhsVecPacketSize*int(sizeof(typename RhsNestedCleaned::Scalar))), + + SameType = is_same<typename LhsNestedCleaned::Scalar,typename RhsNestedCleaned::Scalar>::value, + + CanVectorizeRhs = bool(RhsRowMajor) && (RhsFlags & PacketAccessBit) && (ColsAtCompileTime!=1), + CanVectorizeLhs = (!LhsRowMajor) && (LhsFlags & PacketAccessBit) && (RowsAtCompileTime!=1), + + EvalToRowMajor = (MaxRowsAtCompileTime==1&&MaxColsAtCompileTime!=1) ? 1 + : (MaxColsAtCompileTime==1&&MaxRowsAtCompileTime!=1) ? 0 + : (bool(RhsRowMajor) && !CanVectorizeLhs), + + Flags = ((unsigned int)(LhsFlags | RhsFlags) & HereditaryBits & ~RowMajorBit) + | (EvalToRowMajor ? RowMajorBit : 0) + // TODO enable vectorization for mixed types + | (SameType && (CanVectorizeLhs || CanVectorizeRhs) ? PacketAccessBit : 0) + | (XprType::IsVectorAtCompileTime ? LinearAccessBit : 0), + + LhsOuterStrideBytes = int(LhsNestedCleaned::OuterStrideAtCompileTime) * int(sizeof(typename LhsNestedCleaned::Scalar)), + RhsOuterStrideBytes = int(RhsNestedCleaned::OuterStrideAtCompileTime) * int(sizeof(typename RhsNestedCleaned::Scalar)), + + Alignment = bool(CanVectorizeLhs) ? (LhsOuterStrideBytes<=0 || (int(LhsOuterStrideBytes) % EIGEN_PLAIN_ENUM_MAX(1,LhsAlignment))!=0 ? 0 : LhsAlignment) + : bool(CanVectorizeRhs) ? (RhsOuterStrideBytes<=0 || (int(RhsOuterStrideBytes) % EIGEN_PLAIN_ENUM_MAX(1,RhsAlignment))!=0 ? 0 : RhsAlignment) + : 0, + + /* CanVectorizeInner deserves special explanation. It does not affect the product flags. It is not used outside + * of Product. If the Product itself is not a packet-access expression, there is still a chance that the inner + * loop of the product might be vectorized. This is the meaning of CanVectorizeInner. Since it doesn't affect + * the Flags, it is safe to make this value depend on ActualPacketAccessBit, that doesn't affect the ABI. + */ + CanVectorizeInner = SameType + && LhsRowMajor + && (!RhsRowMajor) + && (LhsFlags & RhsFlags & ActualPacketAccessBit) + && (InnerSize % packet_traits<Scalar>::size == 0) + }; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CoeffReturnType coeff(Index row, Index col) const + { + return (m_lhs.row(row).transpose().cwiseProduct( m_rhs.col(col) )).sum(); + } + + /* Allow index-based non-packet access. It is impossible though to allow index-based packed access, + * 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 + { + const Index row = (RowsAtCompileTime == 1 || MaxRowsAtCompileTime==1) ? 0 : index; + const Index col = (RowsAtCompileTime == 1 || MaxRowsAtCompileTime==1) ? index : 0; + return (m_lhs.row(row).transpose().cwiseProduct( m_rhs.col(col) )).sum(); + } + + template<int LoadMode, typename PacketType> + const PacketType packet(Index row, Index col) const + { + PacketType res; + typedef etor_product_packet_impl<bool(int(Flags)&RowMajorBit) ? RowMajor : ColMajor, + Unroll ? int(InnerSize) : Dynamic, + LhsEtorType, RhsEtorType, PacketType, LoadMode> PacketImpl; + PacketImpl::run(row, col, m_lhsImpl, m_rhsImpl, m_innerDim, res); + return res; + } + + template<int LoadMode, typename PacketType> + const PacketType packet(Index index) const + { + const Index row = (RowsAtCompileTime == 1 || MaxRowsAtCompileTime==1) ? 0 : index; + const Index col = (RowsAtCompileTime == 1 || MaxRowsAtCompileTime==1) ? index : 0; + return packet<LoadMode,PacketType>(row,col); + } + +protected: + typename internal::add_const_on_value_type<LhsNested>::type m_lhs; + typename internal::add_const_on_value_type<RhsNested>::type m_rhs; + + LhsEtorType m_lhsImpl; + RhsEtorType m_rhsImpl; + + // TODO: Get rid of m_innerDim if known at compile time + Index m_innerDim; +}; + +template<typename Lhs, typename Rhs> +struct product_evaluator<Product<Lhs, Rhs, DefaultProduct>, LazyCoeffBasedProductMode, DenseShape, DenseShape> + : product_evaluator<Product<Lhs, Rhs, LazyProduct>, CoeffBasedProductMode, DenseShape, DenseShape> +{ + typedef Product<Lhs, Rhs, DefaultProduct> XprType; + typedef Product<Lhs, Rhs, LazyProduct> BaseProduct; + typedef product_evaluator<BaseProduct, CoeffBasedProductMode, DenseShape, DenseShape> Base; + enum { + Flags = Base::Flags | EvalBeforeNestingBit + }; + EIGEN_DEVICE_FUNC explicit product_evaluator(const XprType& xpr) + : Base(BaseProduct(xpr.lhs(),xpr.rhs())) + {} +}; + +/**************************************** +*** Coeff based product, Packet path *** +****************************************/ + +template<int UnrollingIndex, typename Lhs, typename Rhs, typename Packet, int LoadMode> +struct etor_product_packet_impl<RowMajor, UnrollingIndex, Lhs, Rhs, Packet, LoadMode> +{ + static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, Packet &res) + { + etor_product_packet_impl<RowMajor, UnrollingIndex-1, Lhs, Rhs, Packet, LoadMode>::run(row, col, lhs, rhs, innerDim, res); + res = pmadd(pset1<Packet>(lhs.coeff(row, Index(UnrollingIndex-1))), rhs.template packet<LoadMode,Packet>(Index(UnrollingIndex-1), col), res); + } +}; + +template<int UnrollingIndex, typename Lhs, typename Rhs, typename Packet, int LoadMode> +struct etor_product_packet_impl<ColMajor, UnrollingIndex, Lhs, Rhs, Packet, LoadMode> +{ + static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, Packet &res) + { + etor_product_packet_impl<ColMajor, UnrollingIndex-1, Lhs, Rhs, Packet, LoadMode>::run(row, col, lhs, rhs, innerDim, res); + res = pmadd(lhs.template packet<LoadMode,Packet>(row, Index(UnrollingIndex-1)), pset1<Packet>(rhs.coeff(Index(UnrollingIndex-1), col)), res); + } +}; + +template<typename Lhs, typename Rhs, typename Packet, int LoadMode> +struct etor_product_packet_impl<RowMajor, 1, Lhs, Rhs, Packet, LoadMode> +{ + static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index /*innerDim*/, Packet &res) + { + res = pmul(pset1<Packet>(lhs.coeff(row, Index(0))),rhs.template packet<LoadMode,Packet>(Index(0), col)); + } +}; + +template<typename Lhs, typename Rhs, typename Packet, int LoadMode> +struct etor_product_packet_impl<ColMajor, 1, Lhs, Rhs, Packet, LoadMode> +{ + static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index /*innerDim*/, Packet &res) + { + res = pmul(lhs.template packet<LoadMode,Packet>(row, Index(0)), pset1<Packet>(rhs.coeff(Index(0), col))); + } +}; + +template<typename Lhs, typename Rhs, typename Packet, int LoadMode> +struct etor_product_packet_impl<RowMajor, 0, Lhs, Rhs, Packet, LoadMode> +{ + static EIGEN_STRONG_INLINE void run(Index /*row*/, Index /*col*/, const Lhs& /*lhs*/, const Rhs& /*rhs*/, Index /*innerDim*/, Packet &res) + { + res = pset1<Packet>(typename unpacket_traits<Packet>::type(0)); + } +}; + +template<typename Lhs, typename Rhs, typename Packet, int LoadMode> +struct etor_product_packet_impl<ColMajor, 0, Lhs, Rhs, Packet, LoadMode> +{ + static EIGEN_STRONG_INLINE void run(Index /*row*/, Index /*col*/, const Lhs& /*lhs*/, const Rhs& /*rhs*/, Index /*innerDim*/, Packet &res) + { + res = pset1<Packet>(typename unpacket_traits<Packet>::type(0)); + } +}; + +template<typename Lhs, typename Rhs, typename Packet, int LoadMode> +struct etor_product_packet_impl<RowMajor, Dynamic, Lhs, Rhs, Packet, LoadMode> +{ + static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, Packet& res) + { + res = pset1<Packet>(typename unpacket_traits<Packet>::type(0)); + for(Index i = 0; i < innerDim; ++i) + res = pmadd(pset1<Packet>(lhs.coeff(row, i)), rhs.template packet<LoadMode,Packet>(i, col), res); + } +}; + +template<typename Lhs, typename Rhs, typename Packet, int LoadMode> +struct etor_product_packet_impl<ColMajor, Dynamic, Lhs, Rhs, Packet, LoadMode> +{ + static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, Packet& res) + { + res = pset1<Packet>(typename unpacket_traits<Packet>::type(0)); + for(Index i = 0; i < innerDim; ++i) + res = pmadd(lhs.template packet<LoadMode,Packet>(row, i), pset1<Packet>(rhs.coeff(i, col)), res); + } +}; + + +/*************************************************************************** +* Triangular products +***************************************************************************/ +template<int Mode, bool LhsIsTriangular, + typename Lhs, bool LhsIsVector, + typename Rhs, bool RhsIsVector> +struct triangular_product_impl; + +template<typename Lhs, typename Rhs, int ProductTag> +struct generic_product_impl<Lhs,Rhs,TriangularShape,DenseShape,ProductTag> + : generic_product_impl_base<Lhs,Rhs,generic_product_impl<Lhs,Rhs,TriangularShape,DenseShape,ProductTag> > +{ + typedef typename Product<Lhs,Rhs>::Scalar Scalar; + + template<typename Dest> + static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) + { + triangular_product_impl<Lhs::Mode,true,typename Lhs::MatrixType,false,Rhs, Rhs::ColsAtCompileTime==1> + ::run(dst, lhs.nestedExpression(), rhs, alpha); + } +}; + +template<typename Lhs, typename Rhs, int ProductTag> +struct generic_product_impl<Lhs,Rhs,DenseShape,TriangularShape,ProductTag> +: generic_product_impl_base<Lhs,Rhs,generic_product_impl<Lhs,Rhs,DenseShape,TriangularShape,ProductTag> > +{ + typedef typename Product<Lhs,Rhs>::Scalar Scalar; + + template<typename Dest> + static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) + { + triangular_product_impl<Rhs::Mode,false,Lhs,Lhs::RowsAtCompileTime==1, typename Rhs::MatrixType, false>::run(dst, lhs, rhs.nestedExpression(), alpha); + } +}; + + +/*************************************************************************** +* SelfAdjoint products +***************************************************************************/ +template <typename Lhs, int LhsMode, bool LhsIsVector, + typename Rhs, int RhsMode, bool RhsIsVector> +struct selfadjoint_product_impl; + +template<typename Lhs, typename Rhs, int ProductTag> +struct generic_product_impl<Lhs,Rhs,SelfAdjointShape,DenseShape,ProductTag> + : generic_product_impl_base<Lhs,Rhs,generic_product_impl<Lhs,Rhs,SelfAdjointShape,DenseShape,ProductTag> > +{ + typedef typename Product<Lhs,Rhs>::Scalar Scalar; + + template<typename Dest> + static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) + { + selfadjoint_product_impl<typename Lhs::MatrixType,Lhs::Mode,false,Rhs,0,Rhs::IsVectorAtCompileTime>::run(dst, lhs.nestedExpression(), rhs, alpha); + } +}; + +template<typename Lhs, typename Rhs, int ProductTag> +struct generic_product_impl<Lhs,Rhs,DenseShape,SelfAdjointShape,ProductTag> +: generic_product_impl_base<Lhs,Rhs,generic_product_impl<Lhs,Rhs,DenseShape,SelfAdjointShape,ProductTag> > +{ + typedef typename Product<Lhs,Rhs>::Scalar Scalar; + + template<typename Dest> + static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) + { + selfadjoint_product_impl<Lhs,0,Lhs::IsVectorAtCompileTime,typename Rhs::MatrixType,Rhs::Mode,false>::run(dst, lhs, rhs.nestedExpression(), alpha); + } +}; + + +/*************************************************************************** +* Diagonal products +***************************************************************************/ + +template<typename MatrixType, typename DiagonalType, typename Derived, int ProductOrder> +struct diagonal_product_evaluator_base + : evaluator_base<Derived> +{ + typedef typename ScalarBinaryOpTraits<typename MatrixType::Scalar, typename DiagonalType::Scalar>::ReturnType Scalar; +public: + enum { + CoeffReadCost = NumTraits<Scalar>::MulCost + evaluator<MatrixType>::CoeffReadCost + evaluator<DiagonalType>::CoeffReadCost, + + MatrixFlags = evaluator<MatrixType>::Flags, + DiagFlags = evaluator<DiagonalType>::Flags, + _StorageOrder = MatrixFlags & RowMajorBit ? RowMajor : ColMajor, + _ScalarAccessOnDiag = !((int(_StorageOrder) == ColMajor && int(ProductOrder) == OnTheLeft) + ||(int(_StorageOrder) == RowMajor && int(ProductOrder) == OnTheRight)), + _SameTypes = is_same<typename MatrixType::Scalar, typename DiagonalType::Scalar>::value, + // FIXME currently we need same types, but in the future the next rule should be the one + //_Vectorizable = bool(int(MatrixFlags)&PacketAccessBit) && ((!_PacketOnDiag) || (_SameTypes && bool(int(DiagFlags)&PacketAccessBit))), + _Vectorizable = bool(int(MatrixFlags)&PacketAccessBit) && _SameTypes && (_ScalarAccessOnDiag || (bool(int(DiagFlags)&PacketAccessBit))), + _LinearAccessMask = (MatrixType::RowsAtCompileTime==1 || MatrixType::ColsAtCompileTime==1) ? LinearAccessBit : 0, + Flags = ((HereditaryBits|_LinearAccessMask) & (unsigned int)(MatrixFlags)) | (_Vectorizable ? PacketAccessBit : 0), + Alignment = evaluator<MatrixType>::Alignment + }; + + diagonal_product_evaluator_base(const MatrixType &mat, const DiagonalType &diag) + : m_diagImpl(diag), m_matImpl(mat) + { + EIGEN_INTERNAL_CHECK_COST_VALUE(NumTraits<Scalar>::MulCost); + EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index idx) const + { + return m_diagImpl.coeff(idx) * m_matImpl.coeff(idx); + } + +protected: + template<int LoadMode,typename PacketType> + EIGEN_STRONG_INLINE PacketType packet_impl(Index row, Index col, Index id, internal::true_type) const + { + return internal::pmul(m_matImpl.template packet<LoadMode,PacketType>(row, col), + internal::pset1<PacketType>(m_diagImpl.coeff(id))); + } + + template<int LoadMode,typename PacketType> + EIGEN_STRONG_INLINE PacketType packet_impl(Index row, Index col, Index id, internal::false_type) const + { + enum { + InnerSize = (MatrixType::Flags & RowMajorBit) ? MatrixType::ColsAtCompileTime : MatrixType::RowsAtCompileTime, + DiagonalPacketLoadMode = EIGEN_PLAIN_ENUM_MIN(LoadMode,((InnerSize%16) == 0) ? int(Aligned16) : int(evaluator<DiagonalType>::Alignment)) // FIXME hardcoded 16!! + }; + return internal::pmul(m_matImpl.template packet<LoadMode,PacketType>(row, col), + m_diagImpl.template packet<DiagonalPacketLoadMode,PacketType>(id)); + } + + evaluator<DiagonalType> m_diagImpl; + evaluator<MatrixType> m_matImpl; +}; + +// diagonal * dense +template<typename Lhs, typename Rhs, int ProductKind, int ProductTag> +struct product_evaluator<Product<Lhs, Rhs, ProductKind>, ProductTag, DiagonalShape, DenseShape> + : diagonal_product_evaluator_base<Rhs, typename Lhs::DiagonalVectorType, Product<Lhs, Rhs, LazyProduct>, OnTheLeft> +{ + typedef diagonal_product_evaluator_base<Rhs, typename Lhs::DiagonalVectorType, Product<Lhs, Rhs, LazyProduct>, OnTheLeft> Base; + using Base::m_diagImpl; + using Base::m_matImpl; + using Base::coeff; + typedef typename Base::Scalar Scalar; + + typedef Product<Lhs, Rhs, ProductKind> XprType; + typedef typename XprType::PlainObject PlainObject; + + enum { + StorageOrder = int(Rhs::Flags) & RowMajorBit ? RowMajor : ColMajor + }; + + EIGEN_DEVICE_FUNC explicit product_evaluator(const XprType& xpr) + : Base(xpr.rhs(), xpr.lhs().diagonal()) + { + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index row, Index col) const + { + return m_diagImpl.coeff(row) * m_matImpl.coeff(row, col); + } + +#ifndef __CUDACC__ + template<int LoadMode,typename PacketType> + EIGEN_STRONG_INLINE PacketType packet(Index row, Index col) const + { + // FIXME: NVCC used to complain about the template keyword, but we have to check whether this is still the case. + // See also similar calls below. + return this->template packet_impl<LoadMode,PacketType>(row,col, row, + typename internal::conditional<int(StorageOrder)==RowMajor, internal::true_type, internal::false_type>::type()); + } + + template<int LoadMode,typename PacketType> + EIGEN_STRONG_INLINE PacketType packet(Index idx) const + { + return packet<LoadMode,PacketType>(int(StorageOrder)==ColMajor?idx:0,int(StorageOrder)==ColMajor?0:idx); + } +#endif +}; + +// dense * diagonal +template<typename Lhs, typename Rhs, int ProductKind, int ProductTag> +struct product_evaluator<Product<Lhs, Rhs, ProductKind>, ProductTag, DenseShape, DiagonalShape> + : diagonal_product_evaluator_base<Lhs, typename Rhs::DiagonalVectorType, Product<Lhs, Rhs, LazyProduct>, OnTheRight> +{ + typedef diagonal_product_evaluator_base<Lhs, typename Rhs::DiagonalVectorType, Product<Lhs, Rhs, LazyProduct>, OnTheRight> Base; + using Base::m_diagImpl; + using Base::m_matImpl; + using Base::coeff; + typedef typename Base::Scalar Scalar; + + typedef Product<Lhs, Rhs, ProductKind> XprType; + typedef typename XprType::PlainObject PlainObject; + + enum { StorageOrder = int(Lhs::Flags) & RowMajorBit ? RowMajor : ColMajor }; + + EIGEN_DEVICE_FUNC explicit product_evaluator(const XprType& xpr) + : Base(xpr.lhs(), xpr.rhs().diagonal()) + { + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index row, Index col) const + { + return m_matImpl.coeff(row, col) * m_diagImpl.coeff(col); + } + +#ifndef __CUDACC__ + template<int LoadMode,typename PacketType> + EIGEN_STRONG_INLINE PacketType packet(Index row, Index col) const + { + return this->template packet_impl<LoadMode,PacketType>(row,col, col, + typename internal::conditional<int(StorageOrder)==ColMajor, internal::true_type, internal::false_type>::type()); + } + + template<int LoadMode,typename PacketType> + EIGEN_STRONG_INLINE PacketType packet(Index idx) const + { + return packet<LoadMode,PacketType>(int(StorageOrder)==ColMajor?idx:0,int(StorageOrder)==ColMajor?0:idx); + } +#endif +}; + +/*************************************************************************** +* Products with permutation matrices +***************************************************************************/ + +/** \internal + * \class permutation_matrix_product + * Internal helper class implementing the product between a permutation matrix and a matrix. + * This class is specialized for DenseShape below and for SparseShape in SparseCore/SparsePermutation.h + */ +template<typename ExpressionType, int Side, bool Transposed, typename ExpressionShape> +struct permutation_matrix_product; + +template<typename ExpressionType, int Side, bool Transposed> +struct permutation_matrix_product<ExpressionType, Side, Transposed, DenseShape> +{ + typedef typename nested_eval<ExpressionType, 1>::type MatrixType; + typedef typename remove_all<MatrixType>::type MatrixTypeCleaned; + + template<typename Dest, typename PermutationType> + static inline void run(Dest& dst, const PermutationType& perm, const ExpressionType& xpr) + { + MatrixType mat(xpr); + const Index n = Side==OnTheLeft ? mat.rows() : mat.cols(); + // FIXME we need an is_same for expression that is not sensitive to constness. For instance + // is_same_xpr<Block<const Matrix>, Block<Matrix> >::value should be true. + //if(is_same<MatrixTypeCleaned,Dest>::value && extract_data(dst) == extract_data(mat)) + if(is_same_dense(dst, mat)) + { + // apply the permutation inplace + Matrix<bool,PermutationType::RowsAtCompileTime,1,0,PermutationType::MaxRowsAtCompileTime> mask(perm.size()); + mask.fill(false); + Index r = 0; + while(r < perm.size()) + { + // search for the next seed + while(r<perm.size() && mask[r]) r++; + if(r>=perm.size()) + break; + // we got one, let's follow it until we are back to the seed + Index k0 = r++; + Index kPrev = k0; + mask.coeffRef(k0) = true; + for(Index k=perm.indices().coeff(k0); k!=k0; k=perm.indices().coeff(k)) + { + Block<Dest, Side==OnTheLeft ? 1 : Dest::RowsAtCompileTime, Side==OnTheRight ? 1 : Dest::ColsAtCompileTime>(dst, k) + .swap(Block<Dest, Side==OnTheLeft ? 1 : Dest::RowsAtCompileTime, Side==OnTheRight ? 1 : Dest::ColsAtCompileTime> + (dst,((Side==OnTheLeft) ^ Transposed) ? k0 : kPrev)); + + mask.coeffRef(k) = true; + kPrev = k; + } + } + } + else + { + for(Index i = 0; i < n; ++i) + { + Block<Dest, Side==OnTheLeft ? 1 : Dest::RowsAtCompileTime, Side==OnTheRight ? 1 : Dest::ColsAtCompileTime> + (dst, ((Side==OnTheLeft) ^ Transposed) ? perm.indices().coeff(i) : i) + + = + + Block<const MatrixTypeCleaned,Side==OnTheLeft ? 1 : MatrixTypeCleaned::RowsAtCompileTime,Side==OnTheRight ? 1 : MatrixTypeCleaned::ColsAtCompileTime> + (mat, ((Side==OnTheRight) ^ Transposed) ? perm.indices().coeff(i) : i); + } + } + } +}; + +template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape> +struct generic_product_impl<Lhs, Rhs, PermutationShape, MatrixShape, ProductTag> +{ + template<typename Dest> + static void evalTo(Dest& dst, const Lhs& lhs, const Rhs& rhs) + { + permutation_matrix_product<Rhs, OnTheLeft, false, MatrixShape>::run(dst, lhs, rhs); + } +}; + +template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape> +struct generic_product_impl<Lhs, Rhs, MatrixShape, PermutationShape, ProductTag> +{ + template<typename Dest> + static void evalTo(Dest& dst, const Lhs& lhs, const Rhs& rhs) + { + permutation_matrix_product<Lhs, OnTheRight, false, MatrixShape>::run(dst, rhs, lhs); + } +}; + +template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape> +struct generic_product_impl<Inverse<Lhs>, Rhs, PermutationShape, MatrixShape, ProductTag> +{ + template<typename Dest> + static void evalTo(Dest& dst, const Inverse<Lhs>& lhs, const Rhs& rhs) + { + permutation_matrix_product<Rhs, OnTheLeft, true, MatrixShape>::run(dst, lhs.nestedExpression(), rhs); + } +}; + +template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape> +struct generic_product_impl<Lhs, Inverse<Rhs>, MatrixShape, PermutationShape, ProductTag> +{ + template<typename Dest> + static void evalTo(Dest& dst, const Lhs& lhs, const Inverse<Rhs>& rhs) + { + permutation_matrix_product<Lhs, OnTheRight, true, MatrixShape>::run(dst, rhs.nestedExpression(), lhs); + } +}; + + +/*************************************************************************** +* Products with transpositions matrices +***************************************************************************/ + +// FIXME could we unify Transpositions and Permutation into a single "shape"?? + +/** \internal + * \class transposition_matrix_product + * Internal helper class implementing the product between a permutation matrix and a matrix. + */ +template<typename ExpressionType, int Side, bool Transposed, typename ExpressionShape> +struct transposition_matrix_product +{ + typedef typename nested_eval<ExpressionType, 1>::type MatrixType; + typedef typename remove_all<MatrixType>::type MatrixTypeCleaned; + + template<typename Dest, typename TranspositionType> + static inline void run(Dest& dst, const TranspositionType& tr, const ExpressionType& xpr) + { + MatrixType mat(xpr); + typedef typename TranspositionType::StorageIndex StorageIndex; + const Index size = tr.size(); + StorageIndex j = 0; + + if(!is_same_dense(dst,mat)) + dst = mat; + + for(Index k=(Transposed?size-1:0) ; Transposed?k>=0:k<size ; Transposed?--k:++k) + if(Index(j=tr.coeff(k))!=k) + { + if(Side==OnTheLeft) dst.row(k).swap(dst.row(j)); + else if(Side==OnTheRight) dst.col(k).swap(dst.col(j)); + } + } +}; + +template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape> +struct generic_product_impl<Lhs, Rhs, TranspositionsShape, MatrixShape, ProductTag> +{ + template<typename Dest> + static void evalTo(Dest& dst, const Lhs& lhs, const Rhs& rhs) + { + transposition_matrix_product<Rhs, OnTheLeft, false, MatrixShape>::run(dst, lhs, rhs); + } +}; + +template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape> +struct generic_product_impl<Lhs, Rhs, MatrixShape, TranspositionsShape, ProductTag> +{ + template<typename Dest> + static void evalTo(Dest& dst, const Lhs& lhs, const Rhs& rhs) + { + transposition_matrix_product<Lhs, OnTheRight, false, MatrixShape>::run(dst, rhs, lhs); + } +}; + + +template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape> +struct generic_product_impl<Transpose<Lhs>, Rhs, TranspositionsShape, MatrixShape, ProductTag> +{ + template<typename Dest> + static void evalTo(Dest& dst, const Transpose<Lhs>& lhs, const Rhs& rhs) + { + transposition_matrix_product<Rhs, OnTheLeft, true, MatrixShape>::run(dst, lhs.nestedExpression(), rhs); + } +}; + +template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape> +struct generic_product_impl<Lhs, Transpose<Rhs>, MatrixShape, TranspositionsShape, ProductTag> +{ + template<typename Dest> + static void evalTo(Dest& dst, const Lhs& lhs, const Transpose<Rhs>& rhs) + { + transposition_matrix_product<Lhs, OnTheRight, true, MatrixShape>::run(dst, rhs.nestedExpression(), lhs); + } +}; + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_PRODUCT_EVALUATORS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/Random.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Random.h new file mode 100644 index 000000000..6faf789c7 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Random.h @@ -0,0 +1,182 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008 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_RANDOM_H +#define EIGEN_RANDOM_H + +namespace Eigen { + +namespace internal { + +template<typename Scalar> struct scalar_random_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_random_op) + inline const Scalar operator() () const { return random<Scalar>(); } +}; + +template<typename Scalar> +struct functor_traits<scalar_random_op<Scalar> > +{ enum { Cost = 5 * NumTraits<Scalar>::MulCost, PacketAccess = false, IsRepeatable = false }; }; + +} // end namespace internal + +/** \returns a random matrix expression + * + * Numbers are uniformly spread through their whole definition range for integer types, + * and in the [-1:1] range for floating point scalar types. + * + * The parameters \a rows and \a cols are the number of rows and of columns of + * the returned matrix. Must be compatible with this MatrixBase type. + * + * \not_reentrant + * + * This variant is meant to be used for dynamic-size matrix types. For fixed-size types, + * it is redundant to pass \a rows and \a cols as arguments, so Random() should be used + * instead. + * + * + * Example: \include MatrixBase_random_int_int.cpp + * Output: \verbinclude MatrixBase_random_int_int.out + * + * This expression has the "evaluate before nesting" flag so that it will be evaluated into + * a temporary matrix whenever it is nested in a larger expression. This prevents unexpected + * behavior with expressions involving random matrices. + * + * See DenseBase::NullaryExpr(Index, const CustomNullaryOp&) for an example using C++11 random generators. + * + * \sa DenseBase::setRandom(), DenseBase::Random(Index), DenseBase::Random() + */ +template<typename Derived> +inline const typename DenseBase<Derived>::RandomReturnType +DenseBase<Derived>::Random(Index rows, Index cols) +{ + return NullaryExpr(rows, cols, internal::scalar_random_op<Scalar>()); +} + +/** \returns a random vector expression + * + * Numbers are uniformly spread through their whole definition range for integer types, + * and in the [-1:1] range for floating point scalar types. + * + * The parameter \a size is the size of the returned vector. + * Must be compatible with this MatrixBase type. + * + * \only_for_vectors + * \not_reentrant + * + * This variant is meant to be used for dynamic-size vector types. For fixed-size types, + * it is redundant to pass \a size as argument, so Random() should be used + * instead. + * + * Example: \include MatrixBase_random_int.cpp + * Output: \verbinclude MatrixBase_random_int.out + * + * This expression has the "evaluate before nesting" flag so that it will be evaluated into + * a temporary vector whenever it is nested in a larger expression. This prevents unexpected + * behavior with expressions involving random matrices. + * + * \sa DenseBase::setRandom(), DenseBase::Random(Index,Index), DenseBase::Random() + */ +template<typename Derived> +inline const typename DenseBase<Derived>::RandomReturnType +DenseBase<Derived>::Random(Index size) +{ + return NullaryExpr(size, internal::scalar_random_op<Scalar>()); +} + +/** \returns a fixed-size random matrix or vector expression + * + * Numbers are uniformly spread through their whole definition range for integer types, + * and in the [-1:1] range for floating point scalar types. + * + * This variant is only for fixed-size MatrixBase types. For dynamic-size types, you + * need to use the variants taking size arguments. + * + * Example: \include MatrixBase_random.cpp + * Output: \verbinclude MatrixBase_random.out + * + * This expression has the "evaluate before nesting" flag so that it will be evaluated into + * a temporary matrix whenever it is nested in a larger expression. This prevents unexpected + * behavior with expressions involving random matrices. + * + * \not_reentrant + * + * \sa DenseBase::setRandom(), DenseBase::Random(Index,Index), DenseBase::Random(Index) + */ +template<typename Derived> +inline const typename DenseBase<Derived>::RandomReturnType +DenseBase<Derived>::Random() +{ + return NullaryExpr(RowsAtCompileTime, ColsAtCompileTime, internal::scalar_random_op<Scalar>()); +} + +/** Sets all coefficients in this expression to random values. + * + * Numbers are uniformly spread through their whole definition range for integer types, + * and in the [-1:1] range for floating point scalar types. + * + * \not_reentrant + * + * Example: \include MatrixBase_setRandom.cpp + * Output: \verbinclude MatrixBase_setRandom.out + * + * \sa class CwiseNullaryOp, setRandom(Index), setRandom(Index,Index) + */ +template<typename Derived> +inline Derived& DenseBase<Derived>::setRandom() +{ + return *this = Random(rows(), cols()); +} + +/** Resizes to the given \a newSize, and sets all coefficients in this expression to random values. + * + * Numbers are uniformly spread through their whole definition range for integer types, + * and in the [-1:1] range for floating point scalar types. + * + * \only_for_vectors + * \not_reentrant + * + * Example: \include Matrix_setRandom_int.cpp + * Output: \verbinclude Matrix_setRandom_int.out + * + * \sa DenseBase::setRandom(), setRandom(Index,Index), class CwiseNullaryOp, DenseBase::Random() + */ +template<typename Derived> +EIGEN_STRONG_INLINE Derived& +PlainObjectBase<Derived>::setRandom(Index newSize) +{ + resize(newSize); + return setRandom(); +} + +/** Resizes to the given size, and sets all coefficients in this expression to random values. + * + * Numbers are uniformly spread through their whole definition range for integer types, + * and in the [-1:1] range for floating point scalar types. + * + * \not_reentrant + * + * \param rows the new number of rows + * \param cols the new number of columns + * + * Example: \include Matrix_setRandom_int_int.cpp + * Output: \verbinclude Matrix_setRandom_int_int.out + * + * \sa DenseBase::setRandom(), setRandom(Index), class CwiseNullaryOp, DenseBase::Random() + */ +template<typename Derived> +EIGEN_STRONG_INLINE Derived& +PlainObjectBase<Derived>::setRandom(Index rows, Index cols) +{ + resize(rows, cols); + return setRandom(); +} + +} // end namespace Eigen + +#endif // EIGEN_RANDOM_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/Redux.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Redux.h new file mode 100644 index 000000000..b6e8f8887 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Redux.h @@ -0,0 +1,505 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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_REDUX_H +#define EIGEN_REDUX_H + +namespace Eigen { + +namespace internal { + +// TODO +// * implement other kind of vectorization +// * factorize code + +/*************************************************************************** +* Part 1 : the logic deciding a strategy for vectorization and unrolling +***************************************************************************/ + +template<typename Func, typename Derived> +struct redux_traits +{ +public: + typedef typename find_best_packet<typename Derived::Scalar,Derived::SizeAtCompileTime>::type PacketType; + enum { + PacketSize = unpacket_traits<PacketType>::size, + InnerMaxSize = int(Derived::IsRowMajor) + ? Derived::MaxColsAtCompileTime + : Derived::MaxRowsAtCompileTime + }; + + enum { + MightVectorize = (int(Derived::Flags)&ActualPacketAccessBit) + && (functor_traits<Func>::PacketAccess), + MayLinearVectorize = bool(MightVectorize) && (int(Derived::Flags)&LinearAccessBit), + MaySliceVectorize = bool(MightVectorize) && int(InnerMaxSize)>=3*PacketSize + }; + +public: + enum { + Traversal = int(MayLinearVectorize) ? int(LinearVectorizedTraversal) + : int(MaySliceVectorize) ? int(SliceVectorizedTraversal) + : int(DefaultTraversal) + }; + +public: + enum { + Cost = Derived::SizeAtCompileTime == Dynamic ? HugeCost + : Derived::SizeAtCompileTime * Derived::CoeffReadCost + (Derived::SizeAtCompileTime-1) * functor_traits<Func>::Cost, + UnrollingLimit = EIGEN_UNROLLING_LIMIT * (int(Traversal) == int(DefaultTraversal) ? 1 : int(PacketSize)) + }; + +public: + enum { + Unrolling = Cost <= UnrollingLimit ? CompleteUnrolling : NoUnrolling + }; + +#ifdef EIGEN_DEBUG_ASSIGN + static void debug() + { + std::cerr << "Xpr: " << typeid(typename Derived::XprType).name() << std::endl; + std::cerr.setf(std::ios::hex, std::ios::basefield); + EIGEN_DEBUG_VAR(Derived::Flags) + std::cerr.unsetf(std::ios::hex); + EIGEN_DEBUG_VAR(InnerMaxSize) + EIGEN_DEBUG_VAR(PacketSize) + EIGEN_DEBUG_VAR(MightVectorize) + EIGEN_DEBUG_VAR(MayLinearVectorize) + EIGEN_DEBUG_VAR(MaySliceVectorize) + EIGEN_DEBUG_VAR(Traversal) + EIGEN_DEBUG_VAR(UnrollingLimit) + EIGEN_DEBUG_VAR(Unrolling) + std::cerr << std::endl; + } +#endif +}; + +/*************************************************************************** +* Part 2 : unrollers +***************************************************************************/ + +/*** no vectorization ***/ + +template<typename Func, typename Derived, int Start, int Length> +struct redux_novec_unroller +{ + enum { + HalfLength = Length/2 + }; + + typedef typename Derived::Scalar Scalar; + + EIGEN_DEVICE_FUNC + static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func) + { + return func(redux_novec_unroller<Func, Derived, Start, HalfLength>::run(mat,func), + redux_novec_unroller<Func, Derived, Start+HalfLength, Length-HalfLength>::run(mat,func)); + } +}; + +template<typename Func, typename Derived, int Start> +struct redux_novec_unroller<Func, Derived, Start, 1> +{ + enum { + outer = Start / Derived::InnerSizeAtCompileTime, + inner = Start % Derived::InnerSizeAtCompileTime + }; + + typedef typename Derived::Scalar Scalar; + + EIGEN_DEVICE_FUNC + static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func&) + { + return mat.coeffByOuterInner(outer, inner); + } +}; + +// This is actually dead code and will never be called. It is required +// to prevent false warnings regarding failed inlining though +// for 0 length run() will never be called at all. +template<typename Func, typename Derived, int Start> +struct redux_novec_unroller<Func, Derived, Start, 0> +{ + typedef typename Derived::Scalar Scalar; + EIGEN_DEVICE_FUNC + static EIGEN_STRONG_INLINE Scalar run(const Derived&, const Func&) { return Scalar(); } +}; + +/*** vectorization ***/ + +template<typename Func, typename Derived, int Start, int Length> +struct redux_vec_unroller +{ + enum { + PacketSize = redux_traits<Func, Derived>::PacketSize, + HalfLength = Length/2 + }; + + typedef typename Derived::Scalar Scalar; + typedef typename redux_traits<Func, Derived>::PacketType PacketScalar; + + static EIGEN_STRONG_INLINE PacketScalar run(const Derived &mat, const Func& func) + { + return func.packetOp( + redux_vec_unroller<Func, Derived, Start, HalfLength>::run(mat,func), + redux_vec_unroller<Func, Derived, Start+HalfLength, Length-HalfLength>::run(mat,func) ); + } +}; + +template<typename Func, typename Derived, int Start> +struct redux_vec_unroller<Func, Derived, Start, 1> +{ + enum { + index = Start * redux_traits<Func, Derived>::PacketSize, + outer = index / int(Derived::InnerSizeAtCompileTime), + inner = index % int(Derived::InnerSizeAtCompileTime), + alignment = Derived::Alignment + }; + + typedef typename Derived::Scalar Scalar; + typedef typename redux_traits<Func, Derived>::PacketType PacketScalar; + + static EIGEN_STRONG_INLINE PacketScalar run(const Derived &mat, const Func&) + { + return mat.template packetByOuterInner<alignment,PacketScalar>(outer, inner); + } +}; + +/*************************************************************************** +* Part 3 : implementation of all cases +***************************************************************************/ + +template<typename Func, typename Derived, + int Traversal = redux_traits<Func, Derived>::Traversal, + int Unrolling = redux_traits<Func, Derived>::Unrolling +> +struct redux_impl; + +template<typename Func, typename Derived> +struct redux_impl<Func, Derived, DefaultTraversal, NoUnrolling> +{ + typedef typename Derived::Scalar Scalar; + EIGEN_DEVICE_FUNC + static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func) + { + eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix"); + Scalar res; + res = mat.coeffByOuterInner(0, 0); + for(Index i = 1; i < mat.innerSize(); ++i) + res = func(res, mat.coeffByOuterInner(0, i)); + for(Index i = 1; i < mat.outerSize(); ++i) + for(Index j = 0; j < mat.innerSize(); ++j) + res = func(res, mat.coeffByOuterInner(i, j)); + return res; + } +}; + +template<typename Func, typename Derived> +struct redux_impl<Func,Derived, DefaultTraversal, CompleteUnrolling> + : public redux_novec_unroller<Func,Derived, 0, Derived::SizeAtCompileTime> +{}; + +template<typename Func, typename Derived> +struct redux_impl<Func, Derived, LinearVectorizedTraversal, NoUnrolling> +{ + typedef typename Derived::Scalar Scalar; + typedef typename redux_traits<Func, Derived>::PacketType PacketScalar; + + static Scalar run(const Derived &mat, const Func& func) + { + const Index size = mat.size(); + + const Index packetSize = redux_traits<Func, Derived>::PacketSize; + const int packetAlignment = unpacket_traits<PacketScalar>::alignment; + enum { + alignment0 = (bool(Derived::Flags & DirectAccessBit) && bool(packet_traits<Scalar>::AlignedOnScalar)) ? int(packetAlignment) : int(Unaligned), + alignment = EIGEN_PLAIN_ENUM_MAX(alignment0, Derived::Alignment) + }; + const Index alignedStart = internal::first_default_aligned(mat.nestedExpression()); + const Index alignedSize2 = ((size-alignedStart)/(2*packetSize))*(2*packetSize); + const Index alignedSize = ((size-alignedStart)/(packetSize))*(packetSize); + const Index alignedEnd2 = alignedStart + alignedSize2; + const Index alignedEnd = alignedStart + alignedSize; + Scalar res; + if(alignedSize) + { + PacketScalar packet_res0 = mat.template packet<alignment,PacketScalar>(alignedStart); + if(alignedSize>packetSize) // we have at least two packets to partly unroll the loop + { + PacketScalar packet_res1 = mat.template packet<alignment,PacketScalar>(alignedStart+packetSize); + for(Index index = alignedStart + 2*packetSize; index < alignedEnd2; index += 2*packetSize) + { + packet_res0 = func.packetOp(packet_res0, mat.template packet<alignment,PacketScalar>(index)); + packet_res1 = func.packetOp(packet_res1, mat.template packet<alignment,PacketScalar>(index+packetSize)); + } + + packet_res0 = func.packetOp(packet_res0,packet_res1); + if(alignedEnd>alignedEnd2) + packet_res0 = func.packetOp(packet_res0, mat.template packet<alignment,PacketScalar>(alignedEnd2)); + } + res = func.predux(packet_res0); + + for(Index index = 0; index < alignedStart; ++index) + res = func(res,mat.coeff(index)); + + for(Index index = alignedEnd; index < size; ++index) + res = func(res,mat.coeff(index)); + } + else // too small to vectorize anything. + // since this is dynamic-size hence inefficient anyway for such small sizes, don't try to optimize. + { + res = mat.coeff(0); + for(Index index = 1; index < size; ++index) + res = func(res,mat.coeff(index)); + } + + return res; + } +}; + +// NOTE: for SliceVectorizedTraversal we simply bypass unrolling +template<typename Func, typename Derived, int Unrolling> +struct redux_impl<Func, Derived, SliceVectorizedTraversal, Unrolling> +{ + typedef typename Derived::Scalar Scalar; + typedef typename redux_traits<Func, Derived>::PacketType PacketType; + + EIGEN_DEVICE_FUNC static Scalar run(const Derived &mat, const Func& func) + { + eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix"); + const Index innerSize = mat.innerSize(); + const Index outerSize = mat.outerSize(); + enum { + packetSize = redux_traits<Func, Derived>::PacketSize + }; + const Index packetedInnerSize = ((innerSize)/packetSize)*packetSize; + Scalar res; + if(packetedInnerSize) + { + PacketType packet_res = mat.template packet<Unaligned,PacketType>(0,0); + for(Index j=0; j<outerSize; ++j) + for(Index i=(j==0?packetSize:0); i<packetedInnerSize; i+=Index(packetSize)) + packet_res = func.packetOp(packet_res, mat.template packetByOuterInner<Unaligned,PacketType>(j,i)); + + res = func.predux(packet_res); + for(Index j=0; j<outerSize; ++j) + for(Index i=packetedInnerSize; i<innerSize; ++i) + res = func(res, mat.coeffByOuterInner(j,i)); + } + else // too small to vectorize anything. + // since this is dynamic-size hence inefficient anyway for such small sizes, don't try to optimize. + { + res = redux_impl<Func, Derived, DefaultTraversal, NoUnrolling>::run(mat, func); + } + + return res; + } +}; + +template<typename Func, typename Derived> +struct redux_impl<Func, Derived, LinearVectorizedTraversal, CompleteUnrolling> +{ + typedef typename Derived::Scalar Scalar; + + typedef typename redux_traits<Func, Derived>::PacketType PacketScalar; + enum { + PacketSize = redux_traits<Func, Derived>::PacketSize, + Size = Derived::SizeAtCompileTime, + VectorizedSize = (Size / PacketSize) * PacketSize + }; + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func) + { + eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix"); + if (VectorizedSize > 0) { + Scalar res = func.predux(redux_vec_unroller<Func, Derived, 0, Size / PacketSize>::run(mat,func)); + if (VectorizedSize != Size) + res = func(res,redux_novec_unroller<Func, Derived, VectorizedSize, Size-VectorizedSize>::run(mat,func)); + return res; + } + else { + return redux_novec_unroller<Func, Derived, 0, Size>::run(mat,func); + } + } +}; + +// evaluator adaptor +template<typename _XprType> +class redux_evaluator +{ +public: + typedef _XprType XprType; + EIGEN_DEVICE_FUNC explicit redux_evaluator(const XprType &xpr) : m_evaluator(xpr), m_xpr(xpr) {} + + typedef typename XprType::Scalar Scalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketScalar PacketScalar; + typedef typename XprType::PacketReturnType PacketReturnType; + + enum { + MaxRowsAtCompileTime = XprType::MaxRowsAtCompileTime, + MaxColsAtCompileTime = XprType::MaxColsAtCompileTime, + // TODO we should not remove DirectAccessBit and rather find an elegant way to query the alignment offset at runtime from the evaluator + Flags = evaluator<XprType>::Flags & ~DirectAccessBit, + IsRowMajor = XprType::IsRowMajor, + SizeAtCompileTime = XprType::SizeAtCompileTime, + InnerSizeAtCompileTime = XprType::InnerSizeAtCompileTime, + CoeffReadCost = evaluator<XprType>::CoeffReadCost, + Alignment = evaluator<XprType>::Alignment + }; + + EIGEN_DEVICE_FUNC Index rows() const { return m_xpr.rows(); } + EIGEN_DEVICE_FUNC Index cols() const { return m_xpr.cols(); } + EIGEN_DEVICE_FUNC Index size() const { return m_xpr.size(); } + EIGEN_DEVICE_FUNC Index innerSize() const { return m_xpr.innerSize(); } + EIGEN_DEVICE_FUNC Index outerSize() const { return m_xpr.outerSize(); } + + EIGEN_DEVICE_FUNC + CoeffReturnType coeff(Index row, Index col) const + { return m_evaluator.coeff(row, col); } + + EIGEN_DEVICE_FUNC + CoeffReturnType coeff(Index index) const + { return m_evaluator.coeff(index); } + + template<int LoadMode, typename PacketType> + PacketType packet(Index row, Index col) const + { return m_evaluator.template packet<LoadMode,PacketType>(row, col); } + + template<int LoadMode, typename PacketType> + PacketType packet(Index index) const + { return m_evaluator.template packet<LoadMode,PacketType>(index); } + + EIGEN_DEVICE_FUNC + CoeffReturnType coeffByOuterInner(Index outer, Index inner) const + { return m_evaluator.coeff(IsRowMajor ? outer : inner, IsRowMajor ? inner : outer); } + + template<int LoadMode, typename PacketType> + PacketType packetByOuterInner(Index outer, Index inner) const + { return m_evaluator.template packet<LoadMode,PacketType>(IsRowMajor ? outer : inner, IsRowMajor ? inner : outer); } + + const XprType & nestedExpression() const { return m_xpr; } + +protected: + internal::evaluator<XprType> m_evaluator; + const XprType &m_xpr; +}; + +} // end namespace internal + +/*************************************************************************** +* Part 4 : public API +***************************************************************************/ + + +/** \returns the result of a full redux operation on the whole matrix or vector using \a func + * + * 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. + * + * \sa DenseBase::sum(), DenseBase::minCoeff(), DenseBase::maxCoeff(), MatrixBase::colwise(), MatrixBase::rowwise() + */ +template<typename Derived> +template<typename Func> +typename internal::traits<Derived>::Scalar +DenseBase<Derived>::redux(const Func& func) const +{ + eigen_assert(this->rows()>0 && this->cols()>0 && "you are using an empty matrix"); + + typedef typename internal::redux_evaluator<Derived> ThisEvaluator; + ThisEvaluator thisEval(derived()); + + return internal::redux_impl<Func, ThisEvaluator>::run(thisEval, func); +} + +/** \returns the minimum of all coefficients of \c *this. + * \warning the result is undefined if \c *this contains NaN. + */ +template<typename Derived> +EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar +DenseBase<Derived>::minCoeff() const +{ + return derived().redux(Eigen::internal::scalar_min_op<Scalar,Scalar>()); +} + +/** \returns the maximum of all coefficients of \c *this. + * \warning the result is undefined if \c *this contains NaN. + */ +template<typename Derived> +EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar +DenseBase<Derived>::maxCoeff() const +{ + return derived().redux(Eigen::internal::scalar_max_op<Scalar,Scalar>()); +} + +/** \returns the sum of all coefficients of \c *this + * + * If \c *this is empty, then the value 0 is returned. + * + * \sa trace(), prod(), mean() + */ +template<typename Derived> +EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar +DenseBase<Derived>::sum() const +{ + if(SizeAtCompileTime==0 || (SizeAtCompileTime==Dynamic && size()==0)) + return Scalar(0); + return derived().redux(Eigen::internal::scalar_sum_op<Scalar,Scalar>()); +} + +/** \returns the mean of all coefficients of *this +* +* \sa trace(), prod(), sum() +*/ +template<typename Derived> +EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar +DenseBase<Derived>::mean() const +{ +#ifdef __INTEL_COMPILER + #pragma warning push + #pragma warning ( disable : 2259 ) +#endif + return Scalar(derived().redux(Eigen::internal::scalar_sum_op<Scalar,Scalar>())) / Scalar(this->size()); +#ifdef __INTEL_COMPILER + #pragma warning pop +#endif +} + +/** \returns the product of all coefficients of *this + * + * Example: \include MatrixBase_prod.cpp + * Output: \verbinclude MatrixBase_prod.out + * + * \sa sum(), mean(), trace() + */ +template<typename Derived> +EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar +DenseBase<Derived>::prod() const +{ + if(SizeAtCompileTime==0 || (SizeAtCompileTime==Dynamic && size()==0)) + return Scalar(1); + return derived().redux(Eigen::internal::scalar_product_op<Scalar>()); +} + +/** \returns the trace of \c *this, i.e. the sum of the coefficients on the main diagonal. + * + * \c *this can be any matrix, not necessarily square. + * + * \sa diagonal(), sum() + */ +template<typename Derived> +EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar +MatrixBase<Derived>::trace() const +{ + return derived().diagonal().sum(); +} + +} // end namespace Eigen + +#endif // EIGEN_REDUX_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/Ref.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Ref.h new file mode 100644 index 000000000..bdf24f52a --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Ref.h @@ -0,0 +1,281 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2012 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_REF_H +#define EIGEN_REF_H + +namespace Eigen { + +namespace internal { + +template<typename _PlainObjectType, int _Options, typename _StrideType> +struct traits<Ref<_PlainObjectType, _Options, _StrideType> > + : public traits<Map<_PlainObjectType, _Options, _StrideType> > +{ + typedef _PlainObjectType PlainObjectType; + typedef _StrideType StrideType; + enum { + Options = _Options, + Flags = traits<Map<_PlainObjectType, _Options, _StrideType> >::Flags | NestByRefBit, + Alignment = traits<Map<_PlainObjectType, _Options, _StrideType> >::Alignment + }; + + template<typename Derived> struct match { + enum { + HasDirectAccess = internal::has_direct_access<Derived>::ret, + StorageOrderMatch = PlainObjectType::IsVectorAtCompileTime || Derived::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 + || 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 + // of has_*ary_operator in evaluator<CwiseNullaryOp>. + // This line is surprisingly very sensitive. For instance, simply adding parenthesis + // as "DerivedAlignment = (int(evaluator<Derived>::Alignment))," will make MSVC fail... + DerivedAlignment = int(evaluator<Derived>::Alignment), + AlignmentMatch = (int(traits<PlainObjectType>::Alignment)==int(Unaligned)) || (DerivedAlignment >= int(Alignment)), // FIXME the first condition is not very clear, it should be replaced by the required alignment + ScalarTypeMatch = internal::is_same<typename PlainObjectType::Scalar, typename Derived::Scalar>::value, + MatchAtCompileTime = HasDirectAccess && StorageOrderMatch && InnerStrideMatch && OuterStrideMatch && AlignmentMatch && ScalarTypeMatch + }; + typedef typename internal::conditional<MatchAtCompileTime,internal::true_type,internal::false_type>::type type; + }; + +}; + +template<typename Derived> +struct traits<RefBase<Derived> > : public traits<Derived> {}; + +} + +template<typename Derived> class RefBase + : public MapBase<Derived> +{ + typedef typename internal::traits<Derived>::PlainObjectType PlainObjectType; + typedef typename internal::traits<Derived>::StrideType StrideType; + +public: + + typedef MapBase<Derived> Base; + EIGEN_DENSE_PUBLIC_INTERFACE(RefBase) + + EIGEN_DEVICE_FUNC inline Index innerStride() const + { + return StrideType::InnerStrideAtCompileTime != 0 ? m_stride.inner() : 1; + } + + EIGEN_DEVICE_FUNC inline Index outerStride() const + { + return StrideType::OuterStrideAtCompileTime != 0 ? m_stride.outer() + : IsVectorAtCompileTime ? this->size() + : int(Flags)&RowMajorBit ? this->cols() + : this->rows(); + } + + EIGEN_DEVICE_FUNC RefBase() + : Base(0,RowsAtCompileTime==Dynamic?0:RowsAtCompileTime,ColsAtCompileTime==Dynamic?0:ColsAtCompileTime), + // Stride<> does not allow default ctor for Dynamic strides, so let' initialize it with dummy values: + m_stride(StrideType::OuterStrideAtCompileTime==Dynamic?0:StrideType::OuterStrideAtCompileTime, + StrideType::InnerStrideAtCompileTime==Dynamic?0:StrideType::InnerStrideAtCompileTime) + {} + + EIGEN_INHERIT_ASSIGNMENT_OPERATORS(RefBase) + +protected: + + typedef Stride<StrideType::OuterStrideAtCompileTime,StrideType::InnerStrideAtCompileTime> StrideBase; + + template<typename Expression> + EIGEN_DEVICE_FUNC void construct(Expression& expr) + { + if(PlainObjectType::RowsAtCompileTime==1) + { + eigen_assert(expr.rows()==1 || expr.cols()==1); + ::new (static_cast<Base*>(this)) Base(expr.data(), 1, expr.size()); + } + else if(PlainObjectType::ColsAtCompileTime==1) + { + eigen_assert(expr.rows()==1 || expr.cols()==1); + ::new (static_cast<Base*>(this)) Base(expr.data(), expr.size(), 1); + } + else + ::new (static_cast<Base*>(this)) Base(expr.data(), expr.rows(), expr.cols()); + + if(Expression::IsVectorAtCompileTime && (!PlainObjectType::IsVectorAtCompileTime) && ((Expression::Flags&RowMajorBit)!=(PlainObjectType::Flags&RowMajorBit))) + ::new (&m_stride) StrideBase(expr.innerStride(), StrideType::InnerStrideAtCompileTime==0?0:1); + else + ::new (&m_stride) StrideBase(StrideType::OuterStrideAtCompileTime==0?0:expr.outerStride(), + StrideType::InnerStrideAtCompileTime==0?0:expr.innerStride()); + } + + StrideBase m_stride; +}; + +/** \class Ref + * \ingroup Core_Module + * + * \brief A matrix or vector expression mapping an existing expression + * + * \tparam PlainObjectType the equivalent matrix type of the mapped data + * \tparam Options specifies the pointer alignment in bytes. It can be: \c #Aligned128, , \c #Aligned64, \c #Aligned32, \c #Aligned16, \c #Aligned8 or \c #Unaligned. + * The default is \c #Unaligned. + * \tparam StrideType optionally specifies strides. By default, Ref implies a contiguous storage along the inner dimension (inner stride==1), + * but accepts a variable outer stride (leading dimension). + * This can be overridden by specifying strides. + * The type passed here must be a specialization of the Stride template, see examples below. + * + * This class provides a way to write non-template functions taking Eigen objects as parameters while limiting the number of copies. + * A Ref<> object can represent either a const expression or a l-value: + * \code + * // in-out argument: + * void foo1(Ref<VectorXf> x); + * + * // read-only const argument: + * void foo2(const Ref<const VectorXf>& x); + * \endcode + * + * In the in-out case, the input argument must satisfy the constraints of the actual Ref<> type, otherwise a compilation issue will be triggered. + * By default, a Ref<VectorXf> can reference any dense vector expression of float having a contiguous memory layout. + * Likewise, a Ref<MatrixXf> can reference any column-major dense matrix expression of float whose column's elements are contiguously stored with + * the possibility to have a constant space in-between each column, i.e. the inner stride must be equal to 1, but the outer stride (or leading dimension) + * can be greater than the number of rows. + * + * In the const case, if the input expression does not match the above requirement, then it is evaluated into a temporary before being passed to the function. + * Here are some examples: + * \code + * MatrixXf A; + * VectorXf a; + * foo1(a.head()); // OK + * foo1(A.col()); // OK + * foo1(A.row()); // Compilation error because here innerstride!=1 + * foo2(A.row()); // Compilation error because A.row() is a 1xN object while foo2 is expecting a Nx1 object + * foo2(A.row().transpose()); // The row is copied into a contiguous temporary + * foo2(2*a); // The expression is evaluated into a temporary + * foo2(A.col().segment(2,4)); // No temporary + * \endcode + * + * The range of inputs that can be referenced without temporary can be enlarged using the last two template parameters. + * Here is an example accepting an innerstride!=1: + * \code + * // in-out argument: + * void foo3(Ref<VectorXf,0,InnerStride<> > x); + * foo3(A.row()); // OK + * \endcode + * The downside here is that the function foo3 might be significantly slower than foo1 because it won't be able to exploit vectorization, and will involve more + * expensive address computations even if the input is contiguously stored in memory. To overcome this issue, one might propose to overload internally calling a + * template function, e.g.: + * \code + * // in the .h: + * void foo(const Ref<MatrixXf>& A); + * void foo(const Ref<MatrixXf,0,Stride<> >& A); + * + * // in the .cpp: + * template<typename TypeOfA> void foo_impl(const TypeOfA& A) { + * ... // crazy code goes here + * } + * void foo(const Ref<MatrixXf>& A) { foo_impl(A); } + * void foo(const Ref<MatrixXf,0,Stride<> >& A) { foo_impl(A); } + * \endcode + * + * + * \sa PlainObjectBase::Map(), \ref TopicStorageOrders + */ +template<typename PlainObjectType, int Options, typename StrideType> class Ref + : public RefBase<Ref<PlainObjectType, Options, StrideType> > +{ + private: + typedef internal::traits<Ref> Traits; + template<typename Derived> + EIGEN_DEVICE_FUNC inline Ref(const PlainObjectBase<Derived>& expr, + typename internal::enable_if<bool(Traits::template match<Derived>::MatchAtCompileTime),Derived>::type* = 0); + public: + + typedef RefBase<Ref> Base; + EIGEN_DENSE_PUBLIC_INTERFACE(Ref) + + + #ifndef EIGEN_PARSED_BY_DOXYGEN + template<typename Derived> + EIGEN_DEVICE_FUNC inline Ref(PlainObjectBase<Derived>& expr, + typename internal::enable_if<bool(Traits::template match<Derived>::MatchAtCompileTime),Derived>::type* = 0) + { + EIGEN_STATIC_ASSERT(bool(Traits::template match<Derived>::MatchAtCompileTime), STORAGE_LAYOUT_DOES_NOT_MATCH); + Base::construct(expr.derived()); + } + template<typename Derived> + EIGEN_DEVICE_FUNC inline Ref(const DenseBase<Derived>& expr, + typename internal::enable_if<bool(Traits::template match<Derived>::MatchAtCompileTime),Derived>::type* = 0) + #else + /** Implicit constructor from any dense expression */ + template<typename Derived> + inline Ref(DenseBase<Derived>& expr) + #endif + { + EIGEN_STATIC_ASSERT(bool(internal::is_lvalue<Derived>::value), THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY); + EIGEN_STATIC_ASSERT(bool(Traits::template match<Derived>::MatchAtCompileTime), STORAGE_LAYOUT_DOES_NOT_MATCH); + EIGEN_STATIC_ASSERT(!Derived::IsPlainObjectBase,THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY); + Base::construct(expr.const_cast_derived()); + } + + EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Ref) + +}; + +// this is the const ref version +template<typename TPlainObjectType, int Options, typename StrideType> class Ref<const TPlainObjectType, Options, StrideType> + : public RefBase<Ref<const TPlainObjectType, Options, StrideType> > +{ + typedef internal::traits<Ref> Traits; + public: + + typedef RefBase<Ref> Base; + EIGEN_DENSE_PUBLIC_INTERFACE(Ref) + + template<typename Derived> + EIGEN_DEVICE_FUNC inline Ref(const DenseBase<Derived>& expr, + typename internal::enable_if<bool(Traits::template match<Derived>::ScalarTypeMatch),Derived>::type* = 0) + { +// std::cout << match_helper<Derived>::HasDirectAccess << "," << match_helper<Derived>::OuterStrideMatch << "," << match_helper<Derived>::InnerStrideMatch << "\n"; +// std::cout << int(StrideType::OuterStrideAtCompileTime) << " - " << int(Derived::OuterStrideAtCompileTime) << "\n"; +// std::cout << int(StrideType::InnerStrideAtCompileTime) << " - " << int(Derived::InnerStrideAtCompileTime) << "\n"; + construct(expr.derived(), typename Traits::template match<Derived>::type()); + } + + EIGEN_DEVICE_FUNC inline Ref(const Ref& other) : Base(other) { + // copy constructor shall not copy the m_object, to avoid unnecessary malloc and copy + } + + template<typename OtherRef> + EIGEN_DEVICE_FUNC inline Ref(const RefBase<OtherRef>& other) { + construct(other.derived(), typename Traits::template match<OtherRef>::type()); + } + + protected: + + template<typename Expression> + EIGEN_DEVICE_FUNC void construct(const Expression& expr,internal::true_type) + { + Base::construct(expr); + } + + template<typename Expression> + EIGEN_DEVICE_FUNC void construct(const Expression& expr, internal::false_type) + { + internal::call_assignment_no_alias(m_object,expr,internal::assign_op<Scalar,Scalar>()); + Base::construct(m_object); + } + + protected: + TPlainObjectType m_object; +}; + +} // end namespace Eigen + +#endif // EIGEN_REF_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/Replicate.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Replicate.h new file mode 100644 index 000000000..9960ef884 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Replicate.h @@ -0,0 +1,142 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2009-2010 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_REPLICATE_H +#define EIGEN_REPLICATE_H + +namespace Eigen { + +namespace internal { +template<typename MatrixType,int RowFactor,int ColFactor> +struct traits<Replicate<MatrixType,RowFactor,ColFactor> > + : traits<MatrixType> +{ + typedef typename MatrixType::Scalar Scalar; + typedef typename traits<MatrixType>::StorageKind StorageKind; + typedef typename traits<MatrixType>::XprKind XprKind; + typedef typename ref_selector<MatrixType>::type MatrixTypeNested; + typedef typename remove_reference<MatrixTypeNested>::type _MatrixTypeNested; + enum { + RowsAtCompileTime = RowFactor==Dynamic || int(MatrixType::RowsAtCompileTime)==Dynamic + ? Dynamic + : RowFactor * MatrixType::RowsAtCompileTime, + ColsAtCompileTime = ColFactor==Dynamic || int(MatrixType::ColsAtCompileTime)==Dynamic + ? Dynamic + : ColFactor * MatrixType::ColsAtCompileTime, + //FIXME we don't propagate the max sizes !!! + MaxRowsAtCompileTime = RowsAtCompileTime, + MaxColsAtCompileTime = ColsAtCompileTime, + IsRowMajor = MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1 ? 1 + : MaxColsAtCompileTime==1 && MaxRowsAtCompileTime!=1 ? 0 + : (MatrixType::Flags & RowMajorBit) ? 1 : 0, + + // FIXME enable DirectAccess with negative strides? + Flags = IsRowMajor ? RowMajorBit : 0 + }; +}; +} + +/** + * \class Replicate + * \ingroup Core_Module + * + * \brief Expression of the multiple replication of a matrix or vector + * + * \tparam MatrixType the type of the object we are replicating + * \tparam RowFactor number of repetitions at compile time along the vertical direction, can be Dynamic. + * \tparam ColFactor number of repetitions at compile time along the horizontal direction, can be Dynamic. + * + * This class represents an expression of the multiple replication of a matrix or vector. + * It is the return type of DenseBase::replicate() and most of the time + * this is the only way it is used. + * + * \sa DenseBase::replicate() + */ +template<typename MatrixType,int RowFactor,int ColFactor> class Replicate + : public internal::dense_xpr_base< Replicate<MatrixType,RowFactor,ColFactor> >::type +{ + typedef typename internal::traits<Replicate>::MatrixTypeNested MatrixTypeNested; + typedef typename internal::traits<Replicate>::_MatrixTypeNested _MatrixTypeNested; + public: + + typedef typename internal::dense_xpr_base<Replicate>::type Base; + EIGEN_DENSE_PUBLIC_INTERFACE(Replicate) + typedef typename internal::remove_all<MatrixType>::type NestedExpression; + + template<typename OriginalMatrixType> + EIGEN_DEVICE_FUNC + inline explicit Replicate(const OriginalMatrixType& matrix) + : m_matrix(matrix), m_rowFactor(RowFactor), m_colFactor(ColFactor) + { + EIGEN_STATIC_ASSERT((internal::is_same<typename internal::remove_const<MatrixType>::type,OriginalMatrixType>::value), + THE_MATRIX_OR_EXPRESSION_THAT_YOU_PASSED_DOES_NOT_HAVE_THE_EXPECTED_TYPE) + eigen_assert(RowFactor!=Dynamic && ColFactor!=Dynamic); + } + + template<typename OriginalMatrixType> + EIGEN_DEVICE_FUNC + inline Replicate(const OriginalMatrixType& matrix, Index rowFactor, Index colFactor) + : m_matrix(matrix), m_rowFactor(rowFactor), m_colFactor(colFactor) + { + EIGEN_STATIC_ASSERT((internal::is_same<typename internal::remove_const<MatrixType>::type,OriginalMatrixType>::value), + THE_MATRIX_OR_EXPRESSION_THAT_YOU_PASSED_DOES_NOT_HAVE_THE_EXPECTED_TYPE) + } + + EIGEN_DEVICE_FUNC + inline Index rows() const { return m_matrix.rows() * m_rowFactor.value(); } + EIGEN_DEVICE_FUNC + inline Index cols() const { return m_matrix.cols() * m_colFactor.value(); } + + EIGEN_DEVICE_FUNC + const _MatrixTypeNested& nestedExpression() const + { + return m_matrix; + } + + protected: + MatrixTypeNested m_matrix; + const internal::variable_if_dynamic<Index, RowFactor> m_rowFactor; + const internal::variable_if_dynamic<Index, ColFactor> m_colFactor; +}; + +/** + * \return an expression of the replication of \c *this + * + * Example: \include MatrixBase_replicate.cpp + * Output: \verbinclude MatrixBase_replicate.out + * + * \sa VectorwiseOp::replicate(), DenseBase::replicate(Index,Index), class Replicate + */ +template<typename Derived> +template<int RowFactor, int ColFactor> +const Replicate<Derived,RowFactor,ColFactor> +DenseBase<Derived>::replicate() const +{ + return Replicate<Derived,RowFactor,ColFactor>(derived()); +} + +/** + * \return an expression of the replication of each column (or row) of \c *this + * + * Example: \include DirectionWise_replicate_int.cpp + * Output: \verbinclude DirectionWise_replicate_int.out + * + * \sa VectorwiseOp::replicate(), DenseBase::replicate(), class Replicate + */ +template<typename ExpressionType, int Direction> +const typename VectorwiseOp<ExpressionType,Direction>::ReplicateReturnType +VectorwiseOp<ExpressionType,Direction>::replicate(Index factor) const +{ + return typename VectorwiseOp<ExpressionType,Direction>::ReplicateReturnType + (_expression(),Direction==Vertical?factor:1,Direction==Horizontal?factor:1); +} + +} // end namespace Eigen + +#endif // EIGEN_REPLICATE_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/ReturnByValue.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/ReturnByValue.h new file mode 100644 index 000000000..c44b7673b --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/ReturnByValue.h @@ -0,0 +1,117 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2009-2010 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2009-2010 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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_RETURNBYVALUE_H +#define EIGEN_RETURNBYVALUE_H + +namespace Eigen { + +namespace internal { + +template<typename Derived> +struct traits<ReturnByValue<Derived> > + : public traits<typename traits<Derived>::ReturnType> +{ + enum { + // We're disabling the DirectAccess because e.g. the constructor of + // the Block-with-DirectAccess expression requires to have a coeffRef method. + // Also, we don't want to have to implement the stride stuff. + Flags = (traits<typename traits<Derived>::ReturnType>::Flags + | EvalBeforeNestingBit) & ~DirectAccessBit + }; +}; + +/* The ReturnByValue object doesn't even have a coeff() method. + * So the only way that nesting it in an expression can work, is by evaluating it into a plain matrix. + * So internal::nested always gives the plain return matrix type. + * + * FIXME: I don't understand why we need this specialization: isn't this taken care of by the EvalBeforeNestingBit ?? + * Answer: EvalBeforeNestingBit should be deprecated since we have the evaluators + */ +template<typename Derived,int n,typename PlainObject> +struct nested_eval<ReturnByValue<Derived>, n, PlainObject> +{ + typedef typename traits<Derived>::ReturnType type; +}; + +} // end namespace internal + +/** \class ReturnByValue + * \ingroup Core_Module + * + */ +template<typename Derived> class ReturnByValue + : public internal::dense_xpr_base< ReturnByValue<Derived> >::type, internal::no_assignment_operator +{ + public: + typedef typename internal::traits<Derived>::ReturnType ReturnType; + + typedef typename internal::dense_xpr_base<ReturnByValue>::type Base; + EIGEN_DENSE_PUBLIC_INTERFACE(ReturnByValue) + + template<typename Dest> + EIGEN_DEVICE_FUNC + inline void evalTo(Dest& dst) const + { static_cast<const Derived*>(this)->evalTo(dst); } + EIGEN_DEVICE_FUNC inline Index rows() const { return static_cast<const Derived*>(this)->rows(); } + EIGEN_DEVICE_FUNC inline Index cols() const { return static_cast<const Derived*>(this)->cols(); } + +#ifndef EIGEN_PARSED_BY_DOXYGEN +#define Unusable YOU_ARE_TRYING_TO_ACCESS_A_SINGLE_COEFFICIENT_IN_A_SPECIAL_EXPRESSION_WHERE_THAT_IS_NOT_ALLOWED_BECAUSE_THAT_WOULD_BE_INEFFICIENT + class Unusable{ + Unusable(const Unusable&) {} + Unusable& operator=(const Unusable&) {return *this;} + }; + const Unusable& coeff(Index) const { return *reinterpret_cast<const Unusable*>(this); } + const Unusable& coeff(Index,Index) const { return *reinterpret_cast<const Unusable*>(this); } + Unusable& coeffRef(Index) { return *reinterpret_cast<Unusable*>(this); } + Unusable& coeffRef(Index,Index) { return *reinterpret_cast<Unusable*>(this); } +#undef Unusable +#endif +}; + +template<typename Derived> +template<typename OtherDerived> +Derived& DenseBase<Derived>::operator=(const ReturnByValue<OtherDerived>& other) +{ + other.evalTo(derived()); + return derived(); +} + +namespace internal { + +// Expression is evaluated in a temporary; default implementation of Assignment is bypassed so that +// when a ReturnByValue expression is assigned, the evaluator is not constructed. +// TODO: Finalize port to new regime; ReturnByValue should not exist in the expression world + +template<typename Derived> +struct evaluator<ReturnByValue<Derived> > + : public evaluator<typename internal::traits<Derived>::ReturnType> +{ + typedef ReturnByValue<Derived> XprType; + typedef typename internal::traits<Derived>::ReturnType PlainObject; + typedef evaluator<PlainObject> Base; + + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr) + : m_result(xpr.rows(), xpr.cols()) + { + ::new (static_cast<Base*>(this)) Base(m_result); + xpr.evalTo(m_result); + } + +protected: + PlainObject m_result; +}; + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_RETURNBYVALUE_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/Reverse.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Reverse.h new file mode 100644 index 000000000..0640cda2a --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Reverse.h @@ -0,0 +1,211 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com> +// Copyright (C) 2009 Ricard Marxer <email@ricardmarxer.com> +// Copyright (C) 2009-2010 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_REVERSE_H +#define EIGEN_REVERSE_H + +namespace Eigen { + +namespace internal { + +template<typename MatrixType, int Direction> +struct traits<Reverse<MatrixType, Direction> > + : traits<MatrixType> +{ + typedef typename MatrixType::Scalar Scalar; + typedef typename traits<MatrixType>::StorageKind StorageKind; + typedef typename traits<MatrixType>::XprKind XprKind; + typedef typename ref_selector<MatrixType>::type MatrixTypeNested; + typedef typename remove_reference<MatrixTypeNested>::type _MatrixTypeNested; + enum { + RowsAtCompileTime = MatrixType::RowsAtCompileTime, + ColsAtCompileTime = MatrixType::ColsAtCompileTime, + MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime, + MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime, + Flags = _MatrixTypeNested::Flags & (RowMajorBit | LvalueBit) + }; +}; + +template<typename PacketType, bool ReversePacket> struct reverse_packet_cond +{ + static inline PacketType run(const PacketType& x) { return preverse(x); } +}; + +template<typename PacketType> struct reverse_packet_cond<PacketType,false> +{ + static inline PacketType run(const PacketType& x) { return x; } +}; + +} // end namespace internal + +/** \class Reverse + * \ingroup Core_Module + * + * \brief Expression of the reverse of a vector or matrix + * + * \tparam MatrixType the type of the object of which we are taking the reverse + * \tparam Direction defines the direction of the reverse operation, can be Vertical, Horizontal, or BothDirections + * + * This class represents an expression of the reverse of a vector. + * It is the return type of MatrixBase::reverse() and VectorwiseOp::reverse() + * and most of the time this is the only way it is used. + * + * \sa MatrixBase::reverse(), VectorwiseOp::reverse() + */ +template<typename MatrixType, int Direction> class Reverse + : public internal::dense_xpr_base< Reverse<MatrixType, Direction> >::type +{ + public: + + typedef typename internal::dense_xpr_base<Reverse>::type Base; + EIGEN_DENSE_PUBLIC_INTERFACE(Reverse) + typedef typename internal::remove_all<MatrixType>::type NestedExpression; + using Base::IsRowMajor; + + protected: + enum { + PacketSize = internal::packet_traits<Scalar>::size, + IsColMajor = !IsRowMajor, + ReverseRow = (Direction == Vertical) || (Direction == BothDirections), + ReverseCol = (Direction == Horizontal) || (Direction == BothDirections), + OffsetRow = ReverseRow && IsColMajor ? PacketSize : 1, + OffsetCol = ReverseCol && IsRowMajor ? PacketSize : 1, + ReversePacket = (Direction == BothDirections) + || ((Direction == Vertical) && IsColMajor) + || ((Direction == Horizontal) && IsRowMajor) + }; + typedef internal::reverse_packet_cond<PacketScalar,ReversePacket> reverse_packet; + public: + + EIGEN_DEVICE_FUNC explicit inline Reverse(const MatrixType& matrix) : m_matrix(matrix) { } + + EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Reverse) + + EIGEN_DEVICE_FUNC inline Index rows() const { return m_matrix.rows(); } + EIGEN_DEVICE_FUNC inline Index cols() const { return m_matrix.cols(); } + + EIGEN_DEVICE_FUNC inline Index innerStride() const + { + return -m_matrix.innerStride(); + } + + EIGEN_DEVICE_FUNC const typename internal::remove_all<typename MatrixType::Nested>::type& + nestedExpression() const + { + return m_matrix; + } + + protected: + typename MatrixType::Nested m_matrix; +}; + +/** \returns an expression of the reverse of *this. + * + * Example: \include MatrixBase_reverse.cpp + * Output: \verbinclude MatrixBase_reverse.out + * + */ +template<typename Derived> +inline typename DenseBase<Derived>::ReverseReturnType +DenseBase<Derived>::reverse() +{ + return ReverseReturnType(derived()); +} + + +//reverse const overload moved DenseBase.h due to a CUDA compiler bug + +/** This is the "in place" version of reverse: it reverses \c *this. + * + * In most cases it is probably better to simply use the reversed expression + * of a matrix. However, when reversing the matrix data itself is really needed, + * then this "in-place" version is probably the right choice because it provides + * the following additional benefits: + * - less error prone: doing the same operation with .reverse() requires special care: + * \code m = m.reverse().eval(); \endcode + * - this API enables reverse operations without the need for a temporary + * - it allows future optimizations (cache friendliness, etc.) + * + * \sa VectorwiseOp::reverseInPlace(), reverse() */ +template<typename Derived> +inline void DenseBase<Derived>::reverseInPlace() +{ + if(cols()>rows()) + { + Index half = cols()/2; + leftCols(half).swap(rightCols(half).reverse()); + if((cols()%2)==1) + { + Index half2 = rows()/2; + col(half).head(half2).swap(col(half).tail(half2).reverse()); + } + } + else + { + Index half = rows()/2; + topRows(half).swap(bottomRows(half).reverse()); + if((rows()%2)==1) + { + Index half2 = cols()/2; + row(half).head(half2).swap(row(half).tail(half2).reverse()); + } + } +} + +namespace internal { + +template<int Direction> +struct vectorwise_reverse_inplace_impl; + +template<> +struct vectorwise_reverse_inplace_impl<Vertical> +{ + template<typename ExpressionType> + static void run(ExpressionType &xpr) + { + Index half = xpr.rows()/2; + xpr.topRows(half).swap(xpr.bottomRows(half).colwise().reverse()); + } +}; + +template<> +struct vectorwise_reverse_inplace_impl<Horizontal> +{ + template<typename ExpressionType> + static void run(ExpressionType &xpr) + { + Index half = xpr.cols()/2; + xpr.leftCols(half).swap(xpr.rightCols(half).rowwise().reverse()); + } +}; + +} // end namespace internal + +/** This is the "in place" version of VectorwiseOp::reverse: it reverses each column or row of \c *this. + * + * In most cases it is probably better to simply use the reversed expression + * of a matrix. However, when reversing the matrix data itself is really needed, + * then this "in-place" version is probably the right choice because it provides + * the following additional benefits: + * - less error prone: doing the same operation with .reverse() requires special care: + * \code m = m.reverse().eval(); \endcode + * - this API enables reverse operations without the need for a temporary + * + * \sa DenseBase::reverseInPlace(), reverse() */ +template<typename ExpressionType, int Direction> +void VectorwiseOp<ExpressionType,Direction>::reverseInPlace() +{ + internal::vectorwise_reverse_inplace_impl<Direction>::run(_expression().const_cast_derived()); +} + +} // end namespace Eigen + +#endif // EIGEN_REVERSE_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/Select.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Select.h new file mode 100644 index 000000000..79eec1b5b --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Select.h @@ -0,0 +1,162 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2010 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_SELECT_H +#define EIGEN_SELECT_H + +namespace Eigen { + +/** \class Select + * \ingroup Core_Module + * + * \brief Expression of a coefficient wise version of the C++ ternary operator ?: + * + * \param ConditionMatrixType the type of the \em condition expression which must be a boolean matrix + * \param ThenMatrixType the type of the \em then expression + * \param ElseMatrixType the type of the \em else expression + * + * This class represents an expression of a coefficient wise version of the C++ ternary operator ?:. + * It is the return type of DenseBase::select() and most of the time this is the only way it is used. + * + * \sa DenseBase::select(const DenseBase<ThenDerived>&, const DenseBase<ElseDerived>&) const + */ + +namespace internal { +template<typename ConditionMatrixType, typename ThenMatrixType, typename ElseMatrixType> +struct traits<Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> > + : traits<ThenMatrixType> +{ + typedef typename traits<ThenMatrixType>::Scalar Scalar; + typedef Dense StorageKind; + typedef typename traits<ThenMatrixType>::XprKind XprKind; + typedef typename ConditionMatrixType::Nested ConditionMatrixNested; + typedef typename ThenMatrixType::Nested ThenMatrixNested; + typedef typename ElseMatrixType::Nested ElseMatrixNested; + enum { + RowsAtCompileTime = ConditionMatrixType::RowsAtCompileTime, + ColsAtCompileTime = ConditionMatrixType::ColsAtCompileTime, + MaxRowsAtCompileTime = ConditionMatrixType::MaxRowsAtCompileTime, + MaxColsAtCompileTime = ConditionMatrixType::MaxColsAtCompileTime, + Flags = (unsigned int)ThenMatrixType::Flags & ElseMatrixType::Flags & RowMajorBit + }; +}; +} + +template<typename ConditionMatrixType, typename ThenMatrixType, typename ElseMatrixType> +class Select : public internal::dense_xpr_base< Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> >::type, + internal::no_assignment_operator +{ + public: + + typedef typename internal::dense_xpr_base<Select>::type Base; + EIGEN_DENSE_PUBLIC_INTERFACE(Select) + + inline EIGEN_DEVICE_FUNC + Select(const ConditionMatrixType& a_conditionMatrix, + const ThenMatrixType& a_thenMatrix, + const ElseMatrixType& a_elseMatrix) + : m_condition(a_conditionMatrix), m_then(a_thenMatrix), m_else(a_elseMatrix) + { + eigen_assert(m_condition.rows() == m_then.rows() && m_condition.rows() == m_else.rows()); + eigen_assert(m_condition.cols() == m_then.cols() && m_condition.cols() == m_else.cols()); + } + + inline EIGEN_DEVICE_FUNC Index rows() const { return m_condition.rows(); } + inline EIGEN_DEVICE_FUNC Index cols() const { return m_condition.cols(); } + + inline EIGEN_DEVICE_FUNC + const Scalar coeff(Index i, Index j) const + { + if (m_condition.coeff(i,j)) + return m_then.coeff(i,j); + else + return m_else.coeff(i,j); + } + + inline EIGEN_DEVICE_FUNC + const Scalar coeff(Index i) const + { + if (m_condition.coeff(i)) + return m_then.coeff(i); + else + return m_else.coeff(i); + } + + inline EIGEN_DEVICE_FUNC const ConditionMatrixType& conditionMatrix() const + { + return m_condition; + } + + inline EIGEN_DEVICE_FUNC const ThenMatrixType& thenMatrix() const + { + return m_then; + } + + inline EIGEN_DEVICE_FUNC const ElseMatrixType& elseMatrix() const + { + return m_else; + } + + protected: + typename ConditionMatrixType::Nested m_condition; + typename ThenMatrixType::Nested m_then; + typename ElseMatrixType::Nested m_else; +}; + + +/** \returns a matrix where each coefficient (i,j) is equal to \a thenMatrix(i,j) + * if \c *this(i,j), and \a elseMatrix(i,j) otherwise. + * + * Example: \include MatrixBase_select.cpp + * Output: \verbinclude MatrixBase_select.out + * + * \sa class Select + */ +template<typename Derived> +template<typename ThenDerived,typename ElseDerived> +inline const Select<Derived,ThenDerived,ElseDerived> +DenseBase<Derived>::select(const DenseBase<ThenDerived>& thenMatrix, + const DenseBase<ElseDerived>& elseMatrix) const +{ + return Select<Derived,ThenDerived,ElseDerived>(derived(), thenMatrix.derived(), elseMatrix.derived()); +} + +/** Version of DenseBase::select(const DenseBase&, const DenseBase&) with + * the \em else expression being a scalar value. + * + * \sa DenseBase::select(const DenseBase<ThenDerived>&, const DenseBase<ElseDerived>&) const, class Select + */ +template<typename Derived> +template<typename ThenDerived> +inline const Select<Derived,ThenDerived, typename ThenDerived::ConstantReturnType> +DenseBase<Derived>::select(const DenseBase<ThenDerived>& thenMatrix, + const typename ThenDerived::Scalar& elseScalar) const +{ + return Select<Derived,ThenDerived,typename ThenDerived::ConstantReturnType>( + derived(), thenMatrix.derived(), ThenDerived::Constant(rows(),cols(),elseScalar)); +} + +/** Version of DenseBase::select(const DenseBase&, const DenseBase&) with + * the \em then expression being a scalar value. + * + * \sa DenseBase::select(const DenseBase<ThenDerived>&, const DenseBase<ElseDerived>&) const, class Select + */ +template<typename Derived> +template<typename ElseDerived> +inline const Select<Derived, typename ElseDerived::ConstantReturnType, ElseDerived > +DenseBase<Derived>::select(const typename ElseDerived::Scalar& thenScalar, + const DenseBase<ElseDerived>& elseMatrix) const +{ + return Select<Derived,typename ElseDerived::ConstantReturnType,ElseDerived>( + derived(), ElseDerived::Constant(rows(),cols(),thenScalar), elseMatrix.derived()); +} + +} // end namespace Eigen + +#endif // EIGEN_SELECT_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/SelfAdjointView.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/SelfAdjointView.h new file mode 100644 index 000000000..504c98f0e --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/SelfAdjointView.h @@ -0,0 +1,350 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2009 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_SELFADJOINTMATRIX_H +#define EIGEN_SELFADJOINTMATRIX_H + +namespace Eigen { + +/** \class SelfAdjointView + * \ingroup Core_Module + * + * + * \brief Expression of a selfadjoint matrix from a triangular part of a dense matrix + * + * \param MatrixType the type of the dense matrix storing the coefficients + * \param TriangularPart can be either \c #Lower or \c #Upper + * + * This class is an expression of a sefladjoint matrix from a triangular part of a matrix + * with given dense storage of the coefficients. It is the return type of MatrixBase::selfadjointView() + * and most of the time this is the only way that it is used. + * + * \sa class TriangularBase, MatrixBase::selfadjointView() + */ + +namespace internal { +template<typename MatrixType, unsigned int UpLo> +struct traits<SelfAdjointView<MatrixType, UpLo> > : traits<MatrixType> +{ + typedef typename ref_selector<MatrixType>::non_const_type MatrixTypeNested; + typedef typename remove_all<MatrixTypeNested>::type MatrixTypeNestedCleaned; + typedef MatrixType ExpressionType; + typedef typename MatrixType::PlainObject FullMatrixType; + enum { + Mode = UpLo | SelfAdjoint, + FlagsLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0, + Flags = MatrixTypeNestedCleaned::Flags & (HereditaryBits|FlagsLvalueBit) + & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit)) // FIXME these flags should be preserved + }; +}; +} + + +template<typename _MatrixType, unsigned int UpLo> class SelfAdjointView + : public TriangularBase<SelfAdjointView<_MatrixType, UpLo> > +{ + public: + + typedef _MatrixType MatrixType; + typedef TriangularBase<SelfAdjointView> Base; + typedef typename internal::traits<SelfAdjointView>::MatrixTypeNested MatrixTypeNested; + typedef typename internal::traits<SelfAdjointView>::MatrixTypeNestedCleaned MatrixTypeNestedCleaned; + typedef MatrixTypeNestedCleaned NestedExpression; + + /** \brief The type of coefficients in this matrix */ + typedef typename internal::traits<SelfAdjointView>::Scalar Scalar; + typedef typename MatrixType::StorageIndex StorageIndex; + typedef typename internal::remove_all<typename MatrixType::ConjugateReturnType>::type MatrixConjugateReturnType; + + enum { + Mode = internal::traits<SelfAdjointView>::Mode, + Flags = internal::traits<SelfAdjointView>::Flags, + TransposeMode = ((Mode & Upper) ? Lower : 0) | ((Mode & Lower) ? Upper : 0) + }; + typedef typename MatrixType::PlainObject PlainObject; + + EIGEN_DEVICE_FUNC + explicit inline SelfAdjointView(MatrixType& matrix) : m_matrix(matrix) + {} + + EIGEN_DEVICE_FUNC + inline Index rows() const { return m_matrix.rows(); } + EIGEN_DEVICE_FUNC + inline Index cols() const { return m_matrix.cols(); } + EIGEN_DEVICE_FUNC + inline Index outerStride() const { return m_matrix.outerStride(); } + EIGEN_DEVICE_FUNC + inline Index innerStride() const { return m_matrix.innerStride(); } + + /** \sa MatrixBase::coeff() + * \warning the coordinates must fit into the referenced triangular part + */ + EIGEN_DEVICE_FUNC + inline Scalar coeff(Index row, Index col) const + { + Base::check_coordinates_internal(row, col); + return m_matrix.coeff(row, col); + } + + /** \sa MatrixBase::coeffRef() + * \warning the coordinates must fit into the referenced triangular part + */ + EIGEN_DEVICE_FUNC + inline Scalar& coeffRef(Index row, Index col) + { + EIGEN_STATIC_ASSERT_LVALUE(SelfAdjointView); + Base::check_coordinates_internal(row, col); + return m_matrix.coeffRef(row, col); + } + + /** \internal */ + EIGEN_DEVICE_FUNC + const MatrixTypeNestedCleaned& _expression() const { return m_matrix; } + + EIGEN_DEVICE_FUNC + const MatrixTypeNestedCleaned& nestedExpression() const { return m_matrix; } + EIGEN_DEVICE_FUNC + MatrixTypeNestedCleaned& nestedExpression() { return m_matrix; } + + /** Efficient triangular matrix times vector/matrix product */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + const Product<SelfAdjointView,OtherDerived> + operator*(const MatrixBase<OtherDerived>& rhs) const + { + return Product<SelfAdjointView,OtherDerived>(*this, rhs.derived()); + } + + /** Efficient vector/matrix times triangular matrix product */ + template<typename OtherDerived> friend + EIGEN_DEVICE_FUNC + const Product<OtherDerived,SelfAdjointView> + operator*(const MatrixBase<OtherDerived>& lhs, const SelfAdjointView& rhs) + { + return Product<OtherDerived,SelfAdjointView>(lhs.derived(),rhs); + } + + friend EIGEN_DEVICE_FUNC + const SelfAdjointView<const EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(Scalar,MatrixType,product),UpLo> + operator*(const Scalar& s, const SelfAdjointView& mat) + { + return (s*mat.nestedExpression()).template selfadjointView<UpLo>(); + } + + /** Perform a symmetric rank 2 update of the selfadjoint matrix \c *this: + * \f$ this = this + \alpha u v^* + conj(\alpha) v u^* \f$ + * \returns a reference to \c *this + * + * The vectors \a u and \c v \b must be column vectors, however they can be + * a adjoint expression without any overhead. Only the meaningful triangular + * part of the matrix is updated, the rest is left unchanged. + * + * \sa rankUpdate(const MatrixBase<DerivedU>&, Scalar) + */ + template<typename DerivedU, typename DerivedV> + EIGEN_DEVICE_FUNC + SelfAdjointView& rankUpdate(const MatrixBase<DerivedU>& u, const MatrixBase<DerivedV>& v, const Scalar& alpha = Scalar(1)); + + /** Perform a symmetric rank K update of the selfadjoint matrix \c *this: + * \f$ this = this + \alpha ( u u^* ) \f$ where \a u is a vector or matrix. + * + * \returns a reference to \c *this + * + * Note that to perform \f$ this = this + \alpha ( u^* u ) \f$ you can simply + * call this function with u.adjoint(). + * + * \sa rankUpdate(const MatrixBase<DerivedU>&, const MatrixBase<DerivedV>&, Scalar) + */ + template<typename DerivedU> + EIGEN_DEVICE_FUNC + SelfAdjointView& rankUpdate(const MatrixBase<DerivedU>& u, const Scalar& alpha = Scalar(1)); + + /** \returns an expression of a triangular view extracted from the current selfadjoint view of a given triangular part + * + * The parameter \a TriMode can have the following values: \c #Upper, \c #StrictlyUpper, \c #UnitUpper, + * \c #Lower, \c #StrictlyLower, \c #UnitLower. + * + * If \c TriMode references the same triangular part than \c *this, then this method simply return a \c TriangularView of the nested expression, + * otherwise, the nested expression is first transposed, thus returning a \c TriangularView<Transpose<MatrixType>> object. + * + * \sa MatrixBase::triangularView(), class TriangularView + */ + template<unsigned int TriMode> + EIGEN_DEVICE_FUNC + typename internal::conditional<(TriMode&(Upper|Lower))==(UpLo&(Upper|Lower)), + TriangularView<MatrixType,TriMode>, + TriangularView<typename MatrixType::AdjointReturnType,TriMode> >::type + triangularView() const + { + typename internal::conditional<(TriMode&(Upper|Lower))==(UpLo&(Upper|Lower)), MatrixType&, typename MatrixType::ConstTransposeReturnType>::type tmp1(m_matrix); + typename internal::conditional<(TriMode&(Upper|Lower))==(UpLo&(Upper|Lower)), MatrixType&, typename MatrixType::AdjointReturnType>::type tmp2(tmp1); + return typename internal::conditional<(TriMode&(Upper|Lower))==(UpLo&(Upper|Lower)), + TriangularView<MatrixType,TriMode>, + TriangularView<typename MatrixType::AdjointReturnType,TriMode> >::type(tmp2); + } + + typedef SelfAdjointView<const MatrixConjugateReturnType,Mode> ConjugateReturnType; + /** \sa MatrixBase::conjugate() const */ + EIGEN_DEVICE_FUNC + inline const ConjugateReturnType conjugate() const + { return ConjugateReturnType(m_matrix.conjugate()); } + + typedef SelfAdjointView<const typename MatrixType::AdjointReturnType,TransposeMode> AdjointReturnType; + /** \sa MatrixBase::adjoint() const */ + EIGEN_DEVICE_FUNC + inline const AdjointReturnType adjoint() const + { return AdjointReturnType(m_matrix.adjoint()); } + + typedef SelfAdjointView<typename MatrixType::TransposeReturnType,TransposeMode> TransposeReturnType; + /** \sa MatrixBase::transpose() */ + EIGEN_DEVICE_FUNC + inline TransposeReturnType transpose() + { + EIGEN_STATIC_ASSERT_LVALUE(MatrixType) + typename MatrixType::TransposeReturnType tmp(m_matrix); + return TransposeReturnType(tmp); + } + + typedef SelfAdjointView<const typename MatrixType::ConstTransposeReturnType,TransposeMode> ConstTransposeReturnType; + /** \sa MatrixBase::transpose() const */ + EIGEN_DEVICE_FUNC + inline const ConstTransposeReturnType transpose() const + { + return ConstTransposeReturnType(m_matrix.transpose()); + } + + /** \returns a const expression of the main diagonal of the matrix \c *this + * + * This method simply returns the diagonal of the nested expression, thus by-passing the SelfAdjointView decorator. + * + * \sa MatrixBase::diagonal(), class Diagonal */ + EIGEN_DEVICE_FUNC + typename MatrixType::ConstDiagonalReturnType diagonal() const + { + return typename MatrixType::ConstDiagonalReturnType(m_matrix); + } + +/////////// Cholesky module /////////// + + const LLT<PlainObject, UpLo> llt() const; + const LDLT<PlainObject, UpLo> ldlt() const; + +/////////// Eigenvalue module /////////// + + /** Real part of #Scalar */ + typedef typename NumTraits<Scalar>::Real RealScalar; + /** Return type of eigenvalues() */ + typedef Matrix<RealScalar, internal::traits<MatrixType>::ColsAtCompileTime, 1> EigenvaluesReturnType; + + EIGEN_DEVICE_FUNC + EigenvaluesReturnType eigenvalues() const; + EIGEN_DEVICE_FUNC + RealScalar operatorNorm() const; + + protected: + MatrixTypeNested m_matrix; +}; + + +// template<typename OtherDerived, typename MatrixType, unsigned int UpLo> +// internal::selfadjoint_matrix_product_returntype<OtherDerived,SelfAdjointView<MatrixType,UpLo> > +// operator*(const MatrixBase<OtherDerived>& lhs, const SelfAdjointView<MatrixType,UpLo>& rhs) +// { +// return internal::matrix_selfadjoint_product_returntype<OtherDerived,SelfAdjointView<MatrixType,UpLo> >(lhs.derived(),rhs); +// } + +// selfadjoint to dense matrix + +namespace internal { + +// TODO currently a selfadjoint expression has the form SelfAdjointView<.,.> +// in the future selfadjoint-ness should be defined by the expression traits +// such that Transpose<SelfAdjointView<.,.> > is valid. (currently TriangularBase::transpose() is overloaded to make it work) +template<typename MatrixType, unsigned int Mode> +struct evaluator_traits<SelfAdjointView<MatrixType,Mode> > +{ + typedef typename storage_kind_to_evaluator_kind<typename MatrixType::StorageKind>::Kind Kind; + typedef SelfAdjointShape Shape; +}; + +template<int UpLo, int SetOpposite, typename DstEvaluatorTypeT, typename SrcEvaluatorTypeT, typename Functor, int Version> +class triangular_dense_assignment_kernel<UpLo,SelfAdjoint,SetOpposite,DstEvaluatorTypeT,SrcEvaluatorTypeT,Functor,Version> + : public generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, Functor, Version> +{ +protected: + typedef generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, Functor, Version> Base; + typedef typename Base::DstXprType DstXprType; + typedef typename Base::SrcXprType SrcXprType; + using Base::m_dst; + using Base::m_src; + using Base::m_functor; +public: + + typedef typename Base::DstEvaluatorType DstEvaluatorType; + typedef typename Base::SrcEvaluatorType SrcEvaluatorType; + typedef typename Base::Scalar Scalar; + typedef typename Base::AssignmentTraits AssignmentTraits; + + + EIGEN_DEVICE_FUNC triangular_dense_assignment_kernel(DstEvaluatorType &dst, const SrcEvaluatorType &src, const Functor &func, DstXprType& dstExpr) + : Base(dst, src, func, dstExpr) + {} + + EIGEN_DEVICE_FUNC void assignCoeff(Index row, Index col) + { + eigen_internal_assert(row!=col); + Scalar tmp = m_src.coeff(row,col); + m_functor.assignCoeff(m_dst.coeffRef(row,col), tmp); + m_functor.assignCoeff(m_dst.coeffRef(col,row), numext::conj(tmp)); + } + + EIGEN_DEVICE_FUNC void assignDiagonalCoeff(Index id) + { + Base::assignCoeff(id,id); + } + + EIGEN_DEVICE_FUNC void assignOppositeCoeff(Index, Index) + { eigen_internal_assert(false && "should never be called"); } +}; + +} // end namespace internal + +/*************************************************************************** +* Implementation of MatrixBase methods +***************************************************************************/ + +/** This is the const version of MatrixBase::selfadjointView() */ +template<typename Derived> +template<unsigned int UpLo> +typename MatrixBase<Derived>::template ConstSelfAdjointViewReturnType<UpLo>::Type +MatrixBase<Derived>::selfadjointView() const +{ + return typename ConstSelfAdjointViewReturnType<UpLo>::Type(derived()); +} + +/** \returns an expression of a symmetric/self-adjoint view extracted from the upper or lower triangular part of the current matrix + * + * The parameter \a UpLo can be either \c #Upper or \c #Lower + * + * Example: \include MatrixBase_selfadjointView.cpp + * Output: \verbinclude MatrixBase_selfadjointView.out + * + * \sa class SelfAdjointView + */ +template<typename Derived> +template<unsigned int UpLo> +typename MatrixBase<Derived>::template SelfAdjointViewReturnType<UpLo>::Type +MatrixBase<Derived>::selfadjointView() +{ + return typename SelfAdjointViewReturnType<UpLo>::Type(derived()); +} + +} // end namespace Eigen + +#endif // EIGEN_SELFADJOINTMATRIX_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/SelfCwiseBinaryOp.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/SelfCwiseBinaryOp.h new file mode 100644 index 000000000..50099df82 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/SelfCwiseBinaryOp.h @@ -0,0 +1,51 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2009-2010 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_SELFCWISEBINARYOP_H +#define EIGEN_SELFCWISEBINARYOP_H + +namespace Eigen { + +// TODO generalize the scalar type of 'other' + +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator*=(const Scalar& other) +{ + typedef typename Derived::PlainObject PlainObject; + internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::mul_assign_op<Scalar,Scalar>()); + return derived(); +} + +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& ArrayBase<Derived>::operator+=(const Scalar& other) +{ + typedef typename Derived::PlainObject PlainObject; + internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::add_assign_op<Scalar,Scalar>()); + return derived(); +} + +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& ArrayBase<Derived>::operator-=(const Scalar& other) +{ + typedef typename Derived::PlainObject PlainObject; + internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::sub_assign_op<Scalar,Scalar>()); + return derived(); +} + +template<typename Derived> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator/=(const Scalar& other) +{ + typedef typename Derived::PlainObject PlainObject; + internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::div_assign_op<Scalar,Scalar>()); + return derived(); +} + +} // end namespace Eigen + +#endif // EIGEN_SELFCWISEBINARYOP_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/Solve.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Solve.h new file mode 100644 index 000000000..a8daea511 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Solve.h @@ -0,0 +1,188 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 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_SOLVE_H +#define EIGEN_SOLVE_H + +namespace Eigen { + +template<typename Decomposition, typename RhsType, typename StorageKind> class SolveImpl; + +/** \class Solve + * \ingroup Core_Module + * + * \brief Pseudo expression representing a solving operation + * + * \tparam Decomposition the type of the matrix or decomposion object + * \tparam Rhstype the type of the right-hand side + * + * This class represents an expression of A.solve(B) + * and most of the time this is the only way it is used. + * + */ +namespace internal { + +// this solve_traits class permits to determine the evaluation type with respect to storage kind (Dense vs Sparse) +template<typename Decomposition, typename RhsType,typename StorageKind> struct solve_traits; + +template<typename Decomposition, typename RhsType> +struct solve_traits<Decomposition,RhsType,Dense> +{ + typedef typename make_proper_matrix_type<typename RhsType::Scalar, + Decomposition::ColsAtCompileTime, + RhsType::ColsAtCompileTime, + RhsType::PlainObject::Options, + Decomposition::MaxColsAtCompileTime, + RhsType::MaxColsAtCompileTime>::type PlainObject; +}; + +template<typename Decomposition, typename RhsType> +struct traits<Solve<Decomposition, RhsType> > + : traits<typename solve_traits<Decomposition,RhsType,typename internal::traits<RhsType>::StorageKind>::PlainObject> +{ + typedef typename solve_traits<Decomposition,RhsType,typename internal::traits<RhsType>::StorageKind>::PlainObject PlainObject; + typedef typename promote_index_type<typename Decomposition::StorageIndex, typename RhsType::StorageIndex>::type StorageIndex; + typedef traits<PlainObject> BaseTraits; + enum { + Flags = BaseTraits::Flags & RowMajorBit, + CoeffReadCost = HugeCost + }; +}; + +} + + +template<typename Decomposition, typename RhsType> +class Solve : public SolveImpl<Decomposition,RhsType,typename internal::traits<RhsType>::StorageKind> +{ +public: + typedef typename internal::traits<Solve>::PlainObject PlainObject; + typedef typename internal::traits<Solve>::StorageIndex StorageIndex; + + Solve(const Decomposition &dec, const RhsType &rhs) + : m_dec(dec), m_rhs(rhs) + {} + + EIGEN_DEVICE_FUNC Index rows() const { return m_dec.cols(); } + EIGEN_DEVICE_FUNC Index cols() const { return m_rhs.cols(); } + + EIGEN_DEVICE_FUNC const Decomposition& dec() const { return m_dec; } + EIGEN_DEVICE_FUNC const RhsType& rhs() const { return m_rhs; } + +protected: + const Decomposition &m_dec; + const RhsType &m_rhs; +}; + + +// Specialization of the Solve expression for dense results +template<typename Decomposition, typename RhsType> +class SolveImpl<Decomposition,RhsType,Dense> + : public MatrixBase<Solve<Decomposition,RhsType> > +{ + typedef Solve<Decomposition,RhsType> Derived; + +public: + + typedef MatrixBase<Solve<Decomposition,RhsType> > Base; + EIGEN_DENSE_PUBLIC_INTERFACE(Derived) + +private: + + Scalar coeff(Index row, Index col) const; + Scalar coeff(Index i) const; +}; + +// Generic API dispatcher +template<typename Decomposition, typename RhsType, typename StorageKind> +class SolveImpl : public internal::generic_xpr_base<Solve<Decomposition,RhsType>, MatrixXpr, StorageKind>::type +{ + public: + typedef typename internal::generic_xpr_base<Solve<Decomposition,RhsType>, MatrixXpr, StorageKind>::type Base; +}; + +namespace internal { + +// Evaluator of Solve -> eval into a temporary +template<typename Decomposition, typename RhsType> +struct evaluator<Solve<Decomposition,RhsType> > + : public evaluator<typename Solve<Decomposition,RhsType>::PlainObject> +{ + typedef Solve<Decomposition,RhsType> SolveType; + typedef typename SolveType::PlainObject PlainObject; + typedef evaluator<PlainObject> Base; + + enum { Flags = Base::Flags | EvalBeforeNestingBit }; + + EIGEN_DEVICE_FUNC explicit evaluator(const SolveType& solve) + : m_result(solve.rows(), solve.cols()) + { + ::new (static_cast<Base*>(this)) Base(m_result); + solve.dec()._solve_impl(solve.rhs(), m_result); + } + +protected: + PlainObject m_result; +}; + +// Specialization for "dst = dec.solve(rhs)" +// NOTE we need to specialize it for Dense2Dense to avoid ambiguous specialization error and a Sparse2Sparse specialization must exist somewhere +template<typename DstXprType, typename DecType, typename RhsType, typename Scalar> +struct Assignment<DstXprType, Solve<DecType,RhsType>, internal::assign_op<Scalar,Scalar>, Dense2Dense> +{ + typedef Solve<DecType,RhsType> SrcXprType; + static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,Scalar> &) + { + Index dstRows = src.rows(); + Index dstCols = src.cols(); + if((dst.rows()!=dstRows) || (dst.cols()!=dstCols)) + dst.resize(dstRows, dstCols); + + src.dec()._solve_impl(src.rhs(), dst); + } +}; + +// Specialization for "dst = dec.transpose().solve(rhs)" +template<typename DstXprType, typename DecType, typename RhsType, typename Scalar> +struct Assignment<DstXprType, Solve<Transpose<const DecType>,RhsType>, internal::assign_op<Scalar,Scalar>, Dense2Dense> +{ + typedef Solve<Transpose<const DecType>,RhsType> SrcXprType; + static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,Scalar> &) + { + Index dstRows = src.rows(); + Index dstCols = src.cols(); + if((dst.rows()!=dstRows) || (dst.cols()!=dstCols)) + dst.resize(dstRows, dstCols); + + src.dec().nestedExpression().template _solve_impl_transposed<false>(src.rhs(), dst); + } +}; + +// Specialization for "dst = dec.adjoint().solve(rhs)" +template<typename DstXprType, typename DecType, typename RhsType, typename Scalar> +struct Assignment<DstXprType, Solve<CwiseUnaryOp<internal::scalar_conjugate_op<typename DecType::Scalar>, const Transpose<const DecType> >,RhsType>, + internal::assign_op<Scalar,Scalar>, Dense2Dense> +{ + typedef Solve<CwiseUnaryOp<internal::scalar_conjugate_op<typename DecType::Scalar>, const Transpose<const DecType> >,RhsType> SrcXprType; + static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,Scalar> &) + { + Index dstRows = src.rows(); + Index dstCols = src.cols(); + if((dst.rows()!=dstRows) || (dst.cols()!=dstCols)) + dst.resize(dstRows, dstCols); + + src.dec().nestedExpression().nestedExpression().template _solve_impl_transposed<true>(src.rhs(), dst); + } +}; + +} // end namepsace internal + +} // end namespace Eigen + +#endif // EIGEN_SOLVE_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/SolveTriangular.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/SolveTriangular.h new file mode 100644 index 000000000..049890b25 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/SolveTriangular.h @@ -0,0 +1,232 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2009 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_SOLVETRIANGULAR_H +#define EIGEN_SOLVETRIANGULAR_H + +namespace Eigen { + +namespace internal { + +// Forward declarations: +// The following two routines are implemented in the products/TriangularSolver*.h files +template<typename LhsScalar, typename RhsScalar, typename Index, int Side, int Mode, bool Conjugate, int StorageOrder> +struct triangular_solve_vector; + +template <typename Scalar, typename Index, int Side, int Mode, bool Conjugate, int TriStorageOrder, int OtherStorageOrder> +struct triangular_solve_matrix; + +// small helper struct extracting some traits on the underlying solver operation +template<typename Lhs, typename Rhs, int Side> +class trsolve_traits +{ + private: + enum { + RhsIsVectorAtCompileTime = (Side==OnTheLeft ? Rhs::ColsAtCompileTime : Rhs::RowsAtCompileTime)==1 + }; + public: + enum { + Unrolling = (RhsIsVectorAtCompileTime && Rhs::SizeAtCompileTime != Dynamic && Rhs::SizeAtCompileTime <= 8) + ? CompleteUnrolling : NoUnrolling, + RhsVectors = RhsIsVectorAtCompileTime ? 1 : Dynamic + }; +}; + +template<typename Lhs, typename Rhs, + int Side, // can be OnTheLeft/OnTheRight + int Mode, // can be Upper/Lower | UnitDiag + int Unrolling = trsolve_traits<Lhs,Rhs,Side>::Unrolling, + int RhsVectors = trsolve_traits<Lhs,Rhs,Side>::RhsVectors + > +struct triangular_solver_selector; + +template<typename Lhs, typename Rhs, int Side, int Mode> +struct triangular_solver_selector<Lhs,Rhs,Side,Mode,NoUnrolling,1> +{ + typedef typename Lhs::Scalar LhsScalar; + typedef typename Rhs::Scalar RhsScalar; + typedef blas_traits<Lhs> LhsProductTraits; + typedef typename LhsProductTraits::ExtractType ActualLhsType; + typedef Map<Matrix<RhsScalar,Dynamic,1>, Aligned> MappedRhs; + static void run(const Lhs& lhs, Rhs& rhs) + { + ActualLhsType actualLhs = LhsProductTraits::extract(lhs); + + // FIXME find a way to allow an inner stride if packet_traits<Scalar>::size==1 + + bool useRhsDirectly = Rhs::InnerStrideAtCompileTime==1 || rhs.innerStride()==1; + + ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhs,rhs.size(), + (useRhsDirectly ? rhs.data() : 0)); + + if(!useRhsDirectly) + MappedRhs(actualRhs,rhs.size()) = rhs; + + triangular_solve_vector<LhsScalar, RhsScalar, Index, Side, Mode, LhsProductTraits::NeedToConjugate, + (int(Lhs::Flags) & RowMajorBit) ? RowMajor : ColMajor> + ::run(actualLhs.cols(), actualLhs.data(), actualLhs.outerStride(), actualRhs); + + if(!useRhsDirectly) + rhs = MappedRhs(actualRhs, rhs.size()); + } +}; + +// the rhs is a matrix +template<typename Lhs, typename Rhs, int Side, int Mode> +struct triangular_solver_selector<Lhs,Rhs,Side,Mode,NoUnrolling,Dynamic> +{ + typedef typename Rhs::Scalar Scalar; + typedef blas_traits<Lhs> LhsProductTraits; + typedef typename LhsProductTraits::DirectLinearAccessType ActualLhsType; + + static void run(const Lhs& lhs, Rhs& rhs) + { + typename internal::add_const_on_value_type<ActualLhsType>::type actualLhs = LhsProductTraits::extract(lhs); + + const Index size = lhs.rows(); + const Index othersize = Side==OnTheLeft? rhs.cols() : rhs.rows(); + + typedef internal::gemm_blocking_space<(Rhs::Flags&RowMajorBit) ? RowMajor : ColMajor,Scalar,Scalar, + Rhs::MaxRowsAtCompileTime, Rhs::MaxColsAtCompileTime, Lhs::MaxRowsAtCompileTime,4> BlockingType; + + BlockingType blocking(rhs.rows(), rhs.cols(), size, 1, false); + + triangular_solve_matrix<Scalar,Index,Side,Mode,LhsProductTraits::NeedToConjugate,(int(Lhs::Flags) & RowMajorBit) ? RowMajor : ColMajor, + (Rhs::Flags&RowMajorBit) ? RowMajor : ColMajor> + ::run(size, othersize, &actualLhs.coeffRef(0,0), actualLhs.outerStride(), &rhs.coeffRef(0,0), rhs.outerStride(), blocking); + } +}; + +/*************************************************************************** +* meta-unrolling implementation +***************************************************************************/ + +template<typename Lhs, typename Rhs, int Mode, int LoopIndex, int Size, + bool Stop = LoopIndex==Size> +struct triangular_solver_unroller; + +template<typename Lhs, typename Rhs, int Mode, int LoopIndex, int Size> +struct triangular_solver_unroller<Lhs,Rhs,Mode,LoopIndex,Size,false> { + enum { + IsLower = ((Mode&Lower)==Lower), + DiagIndex = IsLower ? LoopIndex : Size - LoopIndex - 1, + StartIndex = IsLower ? 0 : DiagIndex+1 + }; + static void run(const Lhs& lhs, Rhs& rhs) + { + if (LoopIndex>0) + rhs.coeffRef(DiagIndex) -= lhs.row(DiagIndex).template segment<LoopIndex>(StartIndex).transpose() + .cwiseProduct(rhs.template segment<LoopIndex>(StartIndex)).sum(); + + if(!(Mode & UnitDiag)) + rhs.coeffRef(DiagIndex) /= lhs.coeff(DiagIndex,DiagIndex); + + triangular_solver_unroller<Lhs,Rhs,Mode,LoopIndex+1,Size>::run(lhs,rhs); + } +}; + +template<typename Lhs, typename Rhs, int Mode, int LoopIndex, int Size> +struct triangular_solver_unroller<Lhs,Rhs,Mode,LoopIndex,Size,true> { + static void run(const Lhs&, Rhs&) {} +}; + +template<typename Lhs, typename Rhs, int Mode> +struct triangular_solver_selector<Lhs,Rhs,OnTheLeft,Mode,CompleteUnrolling,1> { + static void run(const Lhs& lhs, Rhs& rhs) + { triangular_solver_unroller<Lhs,Rhs,Mode,0,Rhs::SizeAtCompileTime>::run(lhs,rhs); } +}; + +template<typename Lhs, typename Rhs, int Mode> +struct triangular_solver_selector<Lhs,Rhs,OnTheRight,Mode,CompleteUnrolling,1> { + static void run(const Lhs& lhs, Rhs& rhs) + { + Transpose<const Lhs> trLhs(lhs); + Transpose<Rhs> trRhs(rhs); + + triangular_solver_unroller<Transpose<const Lhs>,Transpose<Rhs>, + ((Mode&Upper)==Upper ? Lower : Upper) | (Mode&UnitDiag), + 0,Rhs::SizeAtCompileTime>::run(trLhs,trRhs); + } +}; + +} // end namespace internal + +/*************************************************************************** +* TriangularView methods +***************************************************************************/ + +#ifndef EIGEN_PARSED_BY_DOXYGEN +template<typename MatrixType, unsigned int Mode> +template<int Side, typename OtherDerived> +void TriangularViewImpl<MatrixType,Mode,Dense>::solveInPlace(const MatrixBase<OtherDerived>& _other) const +{ + OtherDerived& other = _other.const_cast_derived(); + eigen_assert( derived().cols() == derived().rows() && ((Side==OnTheLeft && derived().cols() == other.rows()) || (Side==OnTheRight && derived().cols() == other.cols())) ); + eigen_assert((!(Mode & ZeroDiag)) && bool(Mode & (Upper|Lower))); + + enum { copy = (internal::traits<OtherDerived>::Flags & RowMajorBit) && OtherDerived::IsVectorAtCompileTime && OtherDerived::SizeAtCompileTime!=1}; + typedef typename internal::conditional<copy, + typename internal::plain_matrix_type_column_major<OtherDerived>::type, OtherDerived&>::type OtherCopy; + OtherCopy otherCopy(other); + + internal::triangular_solver_selector<MatrixType, typename internal::remove_reference<OtherCopy>::type, + Side, Mode>::run(derived().nestedExpression(), otherCopy); + + if (copy) + other = otherCopy; +} + +template<typename Derived, unsigned int Mode> +template<int Side, typename Other> +const internal::triangular_solve_retval<Side,TriangularView<Derived,Mode>,Other> +TriangularViewImpl<Derived,Mode,Dense>::solve(const MatrixBase<Other>& other) const +{ + return internal::triangular_solve_retval<Side,TriangularViewType,Other>(derived(), other.derived()); +} +#endif + +namespace internal { + + +template<int Side, typename TriangularType, typename Rhs> +struct traits<triangular_solve_retval<Side, TriangularType, Rhs> > +{ + typedef typename internal::plain_matrix_type_column_major<Rhs>::type ReturnType; +}; + +template<int Side, typename TriangularType, typename Rhs> struct triangular_solve_retval + : public ReturnByValue<triangular_solve_retval<Side, TriangularType, Rhs> > +{ + typedef typename remove_all<typename Rhs::Nested>::type RhsNestedCleaned; + typedef ReturnByValue<triangular_solve_retval> Base; + + triangular_solve_retval(const TriangularType& tri, const Rhs& rhs) + : m_triangularMatrix(tri), m_rhs(rhs) + {} + + inline Index rows() const { return m_rhs.rows(); } + inline Index cols() const { return m_rhs.cols(); } + + template<typename Dest> inline void evalTo(Dest& dst) const + { + if(!is_same_dense(dst,m_rhs)) + dst = m_rhs; + m_triangularMatrix.template solveInPlace<Side>(dst); + } + + protected: + const TriangularType& m_triangularMatrix; + typename Rhs::Nested m_rhs; +}; + +} // namespace internal + +} // end namespace Eigen + +#endif // EIGEN_SOLVETRIANGULAR_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/SolverBase.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/SolverBase.h new file mode 100644 index 000000000..8a4adc229 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/SolverBase.h @@ -0,0 +1,130 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2015 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_SOLVERBASE_H +#define EIGEN_SOLVERBASE_H + +namespace Eigen { + +namespace internal { + + + +} // end namespace internal + +/** \class SolverBase + * \brief A base class for matrix decomposition and solvers + * + * \tparam Derived the actual type of the decomposition/solver. + * + * Any matrix decomposition inheriting this base class provide the following API: + * + * \code + * MatrixType A, b, x; + * DecompositionType dec(A); + * x = dec.solve(b); // solve A * x = b + * x = dec.transpose().solve(b); // solve A^T * x = b + * x = dec.adjoint().solve(b); // solve A' * x = b + * \endcode + * + * \warning Currently, any other usage of transpose() and adjoint() are not supported and will produce compilation errors. + * + * \sa class PartialPivLU, class FullPivLU + */ +template<typename Derived> +class SolverBase : public EigenBase<Derived> +{ + public: + + typedef EigenBase<Derived> Base; + typedef typename internal::traits<Derived>::Scalar Scalar; + typedef Scalar CoeffReturnType; + + enum { + RowsAtCompileTime = internal::traits<Derived>::RowsAtCompileTime, + ColsAtCompileTime = internal::traits<Derived>::ColsAtCompileTime, + SizeAtCompileTime = (internal::size_at_compile_time<internal::traits<Derived>::RowsAtCompileTime, + internal::traits<Derived>::ColsAtCompileTime>::ret), + MaxRowsAtCompileTime = internal::traits<Derived>::MaxRowsAtCompileTime, + MaxColsAtCompileTime = internal::traits<Derived>::MaxColsAtCompileTime, + MaxSizeAtCompileTime = (internal::size_at_compile_time<internal::traits<Derived>::MaxRowsAtCompileTime, + internal::traits<Derived>::MaxColsAtCompileTime>::ret), + IsVectorAtCompileTime = internal::traits<Derived>::MaxRowsAtCompileTime == 1 + || internal::traits<Derived>::MaxColsAtCompileTime == 1 + }; + + /** Default constructor */ + SolverBase() + {} + + ~SolverBase() + {} + + using Base::derived; + + /** \returns an expression of the solution x of \f$ A x = b \f$ using the current decomposition of A. + */ + template<typename Rhs> + 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"); + return Solve<Derived, Rhs>(derived(), b.derived()); + } + + /** \internal the return type of transpose() */ + typedef typename internal::add_const<Transpose<const Derived> >::type ConstTransposeReturnType; + /** \returns an expression of the transposed of the factored matrix. + * + * A typical usage is to solve for the transposed problem A^T x = b: + * \code x = dec.transpose().solve(b); \endcode + * + * \sa adjoint(), solve() + */ + inline ConstTransposeReturnType transpose() const + { + return ConstTransposeReturnType(derived()); + } + + /** \internal the return type of adjoint() */ + typedef typename internal::conditional<NumTraits<Scalar>::IsComplex, + CwiseUnaryOp<internal::scalar_conjugate_op<Scalar>, ConstTransposeReturnType>, + ConstTransposeReturnType + >::type AdjointReturnType; + /** \returns an expression of the adjoint of the factored matrix + * + * A typical usage is to solve for the adjoint problem A' x = b: + * \code x = dec.adjoint().solve(b); \endcode + * + * For real scalar types, this function is equivalent to transpose(). + * + * \sa transpose(), solve() + */ + inline AdjointReturnType adjoint() const + { + return AdjointReturnType(derived().transpose()); + } + + protected: +}; + +namespace internal { + +template<typename Derived> +struct generic_xpr_base<Derived, MatrixXpr, SolverStorage> +{ + typedef SolverBase<Derived> type; + +}; + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_SOLVERBASE_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/StableNorm.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/StableNorm.h new file mode 100644 index 000000000..be04ed44d --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/StableNorm.h @@ -0,0 +1,221 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2009 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_STABLENORM_H +#define EIGEN_STABLENORM_H + +namespace Eigen { + +namespace internal { + +template<typename ExpressionType, typename Scalar> +inline void stable_norm_kernel(const ExpressionType& bl, Scalar& ssq, Scalar& scale, Scalar& invScale) +{ + Scalar maxCoeff = bl.cwiseAbs().maxCoeff(); + + if(maxCoeff>scale) + { + ssq = ssq * numext::abs2(scale/maxCoeff); + Scalar tmp = Scalar(1)/maxCoeff; + if(tmp > NumTraits<Scalar>::highest()) + { + invScale = NumTraits<Scalar>::highest(); + scale = Scalar(1)/invScale; + } + else if(maxCoeff>NumTraits<Scalar>::highest()) // we got a INF + { + invScale = Scalar(1); + scale = maxCoeff; + } + else + { + scale = maxCoeff; + invScale = tmp; + } + } + else if(maxCoeff!=maxCoeff) // we got a NaN + { + scale = maxCoeff; + } + + // TODO if the maxCoeff is much much smaller than the current scale, + // then we can neglect this sub vector + if(scale>Scalar(0)) // if scale==0, then bl is 0 + ssq += (bl*invScale).squaredNorm(); +} + +template<typename Derived> +inline typename NumTraits<typename traits<Derived>::Scalar>::Real +blueNorm_impl(const EigenBase<Derived>& _vec) +{ + typedef typename Derived::RealScalar RealScalar; + using std::pow; + using std::sqrt; + using std::abs; + const Derived& vec(_vec.derived()); + static bool initialized = false; + static RealScalar b1, b2, s1m, s2m, rbig, relerr; + if(!initialized) + { + int ibeta, it, iemin, iemax, iexp; + RealScalar eps; + // This program calculates the machine-dependent constants + // bl, b2, slm, s2m, relerr overfl + // from the "basic" machine-dependent numbers + // nbig, ibeta, it, iemin, iemax, rbig. + // The following define the basic machine-dependent constants. + // For portability, the PORT subprograms "ilmaeh" and "rlmach" + // are used. For any specific computer, each of the assignment + // statements can be replaced + ibeta = std::numeric_limits<RealScalar>::radix; // base for floating-point numbers + it = std::numeric_limits<RealScalar>::digits; // number of base-beta digits in mantissa + iemin = std::numeric_limits<RealScalar>::min_exponent; // minimum exponent + iemax = std::numeric_limits<RealScalar>::max_exponent; // maximum exponent + rbig = (std::numeric_limits<RealScalar>::max)(); // largest floating-point number + + iexp = -((1-iemin)/2); + b1 = RealScalar(pow(RealScalar(ibeta),RealScalar(iexp))); // lower boundary of midrange + iexp = (iemax + 1 - it)/2; + b2 = RealScalar(pow(RealScalar(ibeta),RealScalar(iexp))); // upper boundary of midrange + + iexp = (2-iemin)/2; + s1m = RealScalar(pow(RealScalar(ibeta),RealScalar(iexp))); // scaling factor for lower range + iexp = - ((iemax+it)/2); + s2m = RealScalar(pow(RealScalar(ibeta),RealScalar(iexp))); // scaling factor for upper range + + eps = RealScalar(pow(double(ibeta), 1-it)); + relerr = sqrt(eps); // tolerance for neglecting asml + initialized = true; + } + Index n = vec.size(); + RealScalar ab2 = b2 / RealScalar(n); + RealScalar asml = RealScalar(0); + RealScalar amed = RealScalar(0); + RealScalar abig = RealScalar(0); + for(typename Derived::InnerIterator it(vec, 0); it; ++it) + { + RealScalar ax = abs(it.value()); + if(ax > ab2) abig += numext::abs2(ax*s2m); + else if(ax < b1) asml += numext::abs2(ax*s1m); + else amed += numext::abs2(ax); + } + if(amed!=amed) + return amed; // we got a NaN + if(abig > RealScalar(0)) + { + abig = sqrt(abig); + if(abig > rbig) // overflow, or *this contains INF values + return abig; // return INF + if(amed > RealScalar(0)) + { + abig = abig/s2m; + amed = sqrt(amed); + } + else + return abig/s2m; + } + else if(asml > RealScalar(0)) + { + if (amed > RealScalar(0)) + { + abig = sqrt(amed); + amed = sqrt(asml) / s1m; + } + else + return sqrt(asml)/s1m; + } + else + return sqrt(amed); + asml = numext::mini(abig, amed); + abig = numext::maxi(abig, amed); + if(asml <= abig*relerr) + return abig; + else + return abig * sqrt(RealScalar(1) + numext::abs2(asml/abig)); +} + +} // end namespace internal + +/** \returns the \em l2 norm of \c *this avoiding underflow and overflow. + * This version use a blockwise two passes algorithm: + * 1 - find the absolute largest coefficient \c s + * 2 - compute \f$ s \Vert \frac{*this}{s} \Vert \f$ in a standard way + * + * For architecture/scalar types supporting vectorization, this version + * is faster than blueNorm(). Otherwise the blueNorm() is much faster. + * + * \sa norm(), blueNorm(), hypotNorm() + */ +template<typename Derived> +inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real +MatrixBase<Derived>::stableNorm() const +{ + using std::sqrt; + using std::abs; + const Index blockSize = 4096; + RealScalar scale(0); + RealScalar invScale(1); + RealScalar ssq(0); // sum of square + + typedef typename internal::nested_eval<Derived,2>::type DerivedCopy; + typedef typename internal::remove_all<DerivedCopy>::type DerivedCopyClean; + DerivedCopy copy(derived()); + + enum { + CanAlign = ( (int(DerivedCopyClean::Flags)&DirectAccessBit) + || (int(internal::evaluator<DerivedCopyClean>::Alignment)>0) // FIXME Alignment)>0 might not be enough + ) && (blockSize*sizeof(Scalar)*2<EIGEN_STACK_ALLOCATION_LIMIT) + && (EIGEN_MAX_STATIC_ALIGN_BYTES>0) // if we cannot allocate on the stack, then let's not bother about this optimization + }; + typedef typename internal::conditional<CanAlign, Ref<const Matrix<Scalar,Dynamic,1,0,blockSize,1>, internal::evaluator<DerivedCopyClean>::Alignment>, + typename DerivedCopyClean::ConstSegmentReturnType>::type SegmentWrapper; + Index n = size(); + + if(n==1) + return abs(this->coeff(0)); + + Index bi = internal::first_default_aligned(copy); + if (bi>0) + internal::stable_norm_kernel(copy.head(bi), ssq, scale, invScale); + for (; bi<n; bi+=blockSize) + internal::stable_norm_kernel(SegmentWrapper(copy.segment(bi,numext::mini(blockSize, n - bi))), ssq, scale, invScale); + return scale * sqrt(ssq); +} + +/** \returns the \em l2 norm of \c *this using the Blue's algorithm. + * A Portable Fortran Program to Find the Euclidean Norm of a Vector, + * ACM TOMS, Vol 4, Issue 1, 1978. + * + * For architecture/scalar types without vectorization, this version + * is much faster than stableNorm(). Otherwise the stableNorm() is faster. + * + * \sa norm(), stableNorm(), hypotNorm() + */ +template<typename Derived> +inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real +MatrixBase<Derived>::blueNorm() const +{ + return internal::blueNorm_impl(*this); +} + +/** \returns the \em l2 norm of \c *this avoiding undeflow and overflow. + * This version use a concatenation of hypot() calls, and it is very slow. + * + * \sa norm(), stableNorm() + */ +template<typename Derived> +inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real +MatrixBase<Derived>::hypotNorm() const +{ + return this->cwiseAbs().redux(internal::scalar_hypot_op<RealScalar>()); +} + +} // end namespace Eigen + +#endif // EIGEN_STABLENORM_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/Stride.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Stride.h new file mode 100644 index 000000000..513742f34 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Stride.h @@ -0,0 +1,111 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2010 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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_STRIDE_H +#define EIGEN_STRIDE_H + +namespace Eigen { + +/** \class Stride + * \ingroup Core_Module + * + * \brief Holds strides information for Map + * + * This class holds the strides information for mapping arrays with strides with class Map. + * + * It holds two values: the inner stride and the outer stride. + * + * The inner stride is the pointer increment between two consecutive entries within a given row of a + * row-major matrix or within a given column of a column-major matrix. + * + * The outer stride is the pointer increment between two consecutive rows of a row-major matrix or + * between two consecutive columns of a column-major matrix. + * + * These two values can be passed either at compile-time as template parameters, or at runtime as + * arguments to the constructor. + * + * Indeed, this class takes two template parameters: + * \tparam _OuterStrideAtCompileTime the outer stride, or Dynamic if you want to specify it at runtime. + * \tparam _InnerStrideAtCompileTime the inner stride, or Dynamic if you want to specify it at runtime. + * + * Here is an example: + * \include Map_general_stride.cpp + * Output: \verbinclude Map_general_stride.out + * + * \sa class InnerStride, class OuterStride, \ref TopicStorageOrders + */ +template<int _OuterStrideAtCompileTime, int _InnerStrideAtCompileTime> +class Stride +{ + public: + typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3 + enum { + InnerStrideAtCompileTime = _InnerStrideAtCompileTime, + OuterStrideAtCompileTime = _OuterStrideAtCompileTime + }; + + /** Default constructor, for use when strides are fixed at compile time */ + EIGEN_DEVICE_FUNC + Stride() + : m_outer(OuterStrideAtCompileTime), m_inner(InnerStrideAtCompileTime) + { + eigen_assert(InnerStrideAtCompileTime != Dynamic && OuterStrideAtCompileTime != Dynamic); + } + + /** Constructor allowing to pass the strides at runtime */ + EIGEN_DEVICE_FUNC + Stride(Index outerStride, Index innerStride) + : m_outer(outerStride), m_inner(innerStride) + { + eigen_assert(innerStride>=0 && outerStride>=0); + } + + /** Copy constructor */ + EIGEN_DEVICE_FUNC + Stride(const Stride& other) + : m_outer(other.outer()), m_inner(other.inner()) + {} + + /** \returns the outer stride */ + EIGEN_DEVICE_FUNC + inline Index outer() const { return m_outer.value(); } + /** \returns the inner stride */ + EIGEN_DEVICE_FUNC + inline Index inner() const { return m_inner.value(); } + + protected: + internal::variable_if_dynamic<Index, OuterStrideAtCompileTime> m_outer; + internal::variable_if_dynamic<Index, InnerStrideAtCompileTime> m_inner; +}; + +/** \brief Convenience specialization of Stride to specify only an inner stride + * See class Map for some examples */ +template<int Value> +class InnerStride : public Stride<0, Value> +{ + typedef Stride<0, Value> Base; + public: + EIGEN_DEVICE_FUNC InnerStride() : Base() {} + EIGEN_DEVICE_FUNC InnerStride(Index v) : Base(0, v) {} // FIXME making this explicit could break valid code +}; + +/** \brief Convenience specialization of Stride to specify only an outer stride + * See class Map for some examples */ +template<int Value> +class OuterStride : public Stride<Value, 0> +{ + typedef Stride<Value, 0> Base; + public: + EIGEN_DEVICE_FUNC OuterStride() : Base() {} + EIGEN_DEVICE_FUNC OuterStride(Index v) : Base(v,0) {} // FIXME making this explicit could break valid code +}; + +} // end namespace Eigen + +#endif // EIGEN_STRIDE_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/Swap.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Swap.h new file mode 100644 index 000000000..d70200918 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Swap.h @@ -0,0 +1,67 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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_SWAP_H +#define EIGEN_SWAP_H + +namespace Eigen { + +namespace internal { + +// Overload default assignPacket behavior for swapping them +template<typename DstEvaluatorTypeT, typename SrcEvaluatorTypeT> +class generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, swap_assign_op<typename DstEvaluatorTypeT::Scalar>, Specialized> + : public generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, swap_assign_op<typename DstEvaluatorTypeT::Scalar>, BuiltIn> +{ +protected: + typedef generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, swap_assign_op<typename DstEvaluatorTypeT::Scalar>, BuiltIn> Base; + using Base::m_dst; + using Base::m_src; + using Base::m_functor; + +public: + typedef typename Base::Scalar Scalar; + 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) + : Base(dst, src, func, dstExpr) + {} + + template<int StoreMode, int LoadMode, typename PacketType> + 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)); + m_dst.template writePacket<StoreMode>(row,col,tmp); + } + + template<int StoreMode, int LoadMode, typename PacketType> + 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)); + m_dst.template writePacket<StoreMode>(index,tmp); + } + + // 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) + { + Index row = Base::rowIndexByOuterInner(outer, inner); + Index col = Base::colIndexByOuterInner(outer, inner); + assignPacket<StoreMode,LoadMode,PacketType>(row, col); + } +}; + +} // namespace internal + +} // end namespace Eigen + +#endif // EIGEN_SWAP_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/Transpose.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Transpose.h new file mode 100644 index 000000000..79b767bcc --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Transpose.h @@ -0,0 +1,403 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com> +// Copyright (C) 2009-2014 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_TRANSPOSE_H +#define EIGEN_TRANSPOSE_H + +namespace Eigen { + +namespace internal { +template<typename MatrixType> +struct traits<Transpose<MatrixType> > : public traits<MatrixType> +{ + typedef typename ref_selector<MatrixType>::type MatrixTypeNested; + typedef typename remove_reference<MatrixTypeNested>::type MatrixTypeNestedPlain; + enum { + RowsAtCompileTime = MatrixType::ColsAtCompileTime, + ColsAtCompileTime = MatrixType::RowsAtCompileTime, + MaxRowsAtCompileTime = MatrixType::MaxColsAtCompileTime, + MaxColsAtCompileTime = MatrixType::MaxRowsAtCompileTime, + FlagsLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0, + Flags0 = traits<MatrixTypeNestedPlain>::Flags & ~(LvalueBit | NestByRefBit), + Flags1 = Flags0 | FlagsLvalueBit, + Flags = Flags1 ^ RowMajorBit, + InnerStrideAtCompileTime = inner_stride_at_compile_time<MatrixType>::ret, + OuterStrideAtCompileTime = outer_stride_at_compile_time<MatrixType>::ret + }; +}; +} + +template<typename MatrixType, typename StorageKind> class TransposeImpl; + +/** \class Transpose + * \ingroup Core_Module + * + * \brief Expression of the transpose of a matrix + * + * \tparam MatrixType the type of the object of which we are taking the transpose + * + * This class represents an expression of the transpose of a matrix. + * It is the return type of MatrixBase::transpose() and MatrixBase::adjoint() + * and most of the time this is the only way it is used. + * + * \sa MatrixBase::transpose(), MatrixBase::adjoint() + */ +template<typename MatrixType> class Transpose + : public TransposeImpl<MatrixType,typename internal::traits<MatrixType>::StorageKind> +{ + public: + + typedef typename internal::ref_selector<MatrixType>::non_const_type MatrixTypeNested; + + typedef typename TransposeImpl<MatrixType,typename internal::traits<MatrixType>::StorageKind>::Base Base; + EIGEN_GENERIC_PUBLIC_INTERFACE(Transpose) + typedef typename internal::remove_all<MatrixType>::type NestedExpression; + + EIGEN_DEVICE_FUNC + explicit 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(); } + + /** \returns the nested expression */ + EIGEN_DEVICE_FUNC + const typename internal::remove_all<MatrixTypeNested>::type& + nestedExpression() const { return m_matrix; } + + /** \returns the nested expression */ + EIGEN_DEVICE_FUNC + typename internal::remove_reference<MatrixTypeNested>::type& + nestedExpression() { return m_matrix; } + + /** \internal */ + void resize(Index nrows, Index ncols) { + m_matrix.resize(ncols,nrows); + } + + protected: + typename internal::ref_selector<MatrixType>::non_const_type m_matrix; +}; + +namespace internal { + +template<typename MatrixType, bool HasDirectAccess = has_direct_access<MatrixType>::ret> +struct TransposeImpl_base +{ + typedef typename dense_xpr_base<Transpose<MatrixType> >::type type; +}; + +template<typename MatrixType> +struct TransposeImpl_base<MatrixType, false> +{ + typedef typename dense_xpr_base<Transpose<MatrixType> >::type type; +}; + +} // end namespace internal + +// Generic API dispatcher +template<typename XprType, typename StorageKind> +class TransposeImpl + : public internal::generic_xpr_base<Transpose<XprType> >::type +{ +public: + typedef typename internal::generic_xpr_base<Transpose<XprType> >::type Base; +}; + +template<typename MatrixType> class TransposeImpl<MatrixType,Dense> + : public internal::TransposeImpl_base<MatrixType>::type +{ + public: + + typedef typename internal::TransposeImpl_base<MatrixType>::type Base; + using Base::coeffRef; + 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(); } + + typedef typename internal::conditional< + internal::is_lvalue<MatrixType>::value, + Scalar, + 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(); } + + // FIXME: shall we keep the const version of coeffRef? + EIGEN_DEVICE_FUNC + 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 + { + return derived().nestedExpression().coeffRef(index); + } +}; + +/** \returns an expression of the transpose of *this. + * + * Example: \include MatrixBase_transpose.cpp + * Output: \verbinclude MatrixBase_transpose.out + * + * \warning If you want to replace a matrix by its own transpose, do \b NOT do this: + * \code + * m = m.transpose(); // bug!!! caused by aliasing effect + * \endcode + * Instead, use the transposeInPlace() method: + * \code + * m.transposeInPlace(); + * \endcode + * which gives Eigen good opportunities for optimization, or alternatively you can also do: + * \code + * m = m.transpose().eval(); + * \endcode + * + * \sa transposeInPlace(), adjoint() */ +template<typename Derived> +inline Transpose<Derived> +DenseBase<Derived>::transpose() +{ + return TransposeReturnType(derived()); +} + +/** This is the const version of transpose(). + * + * Make sure you read the warning for transpose() ! + * + * \sa transposeInPlace(), adjoint() */ +template<typename Derived> +inline typename DenseBase<Derived>::ConstTransposeReturnType +DenseBase<Derived>::transpose() const +{ + return ConstTransposeReturnType(derived()); +} + +/** \returns an expression of the adjoint (i.e. conjugate transpose) of *this. + * + * Example: \include MatrixBase_adjoint.cpp + * Output: \verbinclude MatrixBase_adjoint.out + * + * \warning If you want to replace a matrix by its own adjoint, do \b NOT do this: + * \code + * m = m.adjoint(); // bug!!! caused by aliasing effect + * \endcode + * Instead, use the adjointInPlace() method: + * \code + * m.adjointInPlace(); + * \endcode + * which gives Eigen good opportunities for optimization, or alternatively you can also do: + * \code + * m = m.adjoint().eval(); + * \endcode + * + * \sa adjointInPlace(), transpose(), conjugate(), class Transpose, class internal::scalar_conjugate_op */ +template<typename Derived> +inline const typename MatrixBase<Derived>::AdjointReturnType +MatrixBase<Derived>::adjoint() const +{ + return AdjointReturnType(this->transpose()); +} + +/*************************************************************************** +* "in place" transpose implementation +***************************************************************************/ + +namespace internal { + +template<typename MatrixType, + bool IsSquare = (MatrixType::RowsAtCompileTime == MatrixType::ColsAtCompileTime) && MatrixType::RowsAtCompileTime!=Dynamic, + bool MatchPacketSize = + (int(MatrixType::RowsAtCompileTime) == int(internal::packet_traits<typename MatrixType::Scalar>::size)) + && (internal::evaluator<MatrixType>::Flags&PacketAccessBit) > +struct inplace_transpose_selector; + +template<typename MatrixType> +struct inplace_transpose_selector<MatrixType,true,false> { // square matrix + static void run(MatrixType& m) { + m.matrix().template triangularView<StrictlyUpper>().swap(m.matrix().transpose()); + } +}; + +// TODO: vectorized path is currently limited to LargestPacketSize x LargestPacketSize cases only. +template<typename MatrixType> +struct inplace_transpose_selector<MatrixType,true,true> { // PacketSize x PacketSize + static void run(MatrixType& m) { + typedef typename MatrixType::Scalar Scalar; + typedef typename internal::packet_traits<typename MatrixType::Scalar>::type Packet; + const Index PacketSize = internal::packet_traits<Scalar>::size; + const Index Alignment = internal::evaluator<MatrixType>::Alignment; + PacketBlock<Packet> A; + for (Index i=0; i<PacketSize; ++i) + A.packet[i] = m.template packetByOuterInner<Alignment>(i,0); + internal::ptranspose(A); + for (Index i=0; i<PacketSize; ++i) + m.template writePacket<Alignment>(m.rowIndexByOuterInner(i,0), m.colIndexByOuterInner(i,0), A.packet[i]); + } +}; + +template<typename MatrixType,bool MatchPacketSize> +struct inplace_transpose_selector<MatrixType,false,MatchPacketSize> { // non square matrix + static void run(MatrixType& m) { + if (m.rows()==m.cols()) + m.matrix().template triangularView<StrictlyUpper>().swap(m.matrix().transpose()); + else + m = m.transpose().eval(); + } +}; + +} // end namespace internal + +/** This is the "in place" version of transpose(): it replaces \c *this by its own transpose. + * Thus, doing + * \code + * m.transposeInPlace(); + * \endcode + * has the same effect on m as doing + * \code + * m = m.transpose().eval(); + * \endcode + * and is faster and also safer because in the latter line of code, forgetting the eval() results + * in a bug caused by \ref TopicAliasing "aliasing". + * + * Notice however that this method is only useful if you want to replace a matrix by its own transpose. + * If you just need the transpose of a matrix, use transpose(). + * + * \note if the matrix is not square, then \c *this must be a resizable matrix. + * This excludes (non-square) fixed-size matrices, block-expressions and maps. + * + * \sa transpose(), adjoint(), adjointInPlace() */ +template<typename Derived> +inline void DenseBase<Derived>::transposeInPlace() +{ + eigen_assert((rows() == cols() || (RowsAtCompileTime == Dynamic && ColsAtCompileTime == Dynamic)) + && "transposeInPlace() called on a non-square non-resizable matrix"); + internal::inplace_transpose_selector<Derived>::run(derived()); +} + +/*************************************************************************** +* "in place" adjoint implementation +***************************************************************************/ + +/** This is the "in place" version of adjoint(): it replaces \c *this by its own transpose. + * Thus, doing + * \code + * m.adjointInPlace(); + * \endcode + * has the same effect on m as doing + * \code + * m = m.adjoint().eval(); + * \endcode + * and is faster and also safer because in the latter line of code, forgetting the eval() results + * in a bug caused by aliasing. + * + * Notice however that this method is only useful if you want to replace a matrix by its own adjoint. + * If you just need the adjoint of a matrix, use adjoint(). + * + * \note if the matrix is not square, then \c *this must be a resizable matrix. + * This excludes (non-square) fixed-size matrices, block-expressions and maps. + * + * \sa transpose(), adjoint(), transposeInPlace() */ +template<typename Derived> +inline void MatrixBase<Derived>::adjointInPlace() +{ + derived() = adjoint().eval(); +} + +#ifndef EIGEN_NO_DEBUG + +// The following is to detect aliasing problems in most common cases. + +namespace internal { + +template<bool DestIsTransposed, typename OtherDerived> +struct check_transpose_aliasing_compile_time_selector +{ + enum { ret = bool(blas_traits<OtherDerived>::IsTransposed) != DestIsTransposed }; +}; + +template<bool DestIsTransposed, typename BinOp, typename DerivedA, typename DerivedB> +struct check_transpose_aliasing_compile_time_selector<DestIsTransposed,CwiseBinaryOp<BinOp,DerivedA,DerivedB> > +{ + enum { ret = bool(blas_traits<DerivedA>::IsTransposed) != DestIsTransposed + || bool(blas_traits<DerivedB>::IsTransposed) != DestIsTransposed + }; +}; + +template<typename Scalar, bool DestIsTransposed, typename OtherDerived> +struct check_transpose_aliasing_run_time_selector +{ + static bool run(const Scalar* dest, const OtherDerived& src) + { + return (bool(blas_traits<OtherDerived>::IsTransposed) != DestIsTransposed) && (dest!=0 && dest==(const Scalar*)extract_data(src)); + } +}; + +template<typename Scalar, bool DestIsTransposed, typename BinOp, typename DerivedA, typename DerivedB> +struct check_transpose_aliasing_run_time_selector<Scalar,DestIsTransposed,CwiseBinaryOp<BinOp,DerivedA,DerivedB> > +{ + static bool run(const Scalar* dest, const CwiseBinaryOp<BinOp,DerivedA,DerivedB>& src) + { + return ((blas_traits<DerivedA>::IsTransposed != DestIsTransposed) && (dest!=0 && dest==(const Scalar*)extract_data(src.lhs()))) + || ((blas_traits<DerivedB>::IsTransposed != DestIsTransposed) && (dest!=0 && dest==(const Scalar*)extract_data(src.rhs()))); + } +}; + +// the following selector, checkTransposeAliasing_impl, based on MightHaveTransposeAliasing, +// is because when the condition controlling the assert is known at compile time, ICC emits a warning. +// This is actually a good warning: in expressions that don't have any transposing, the condition is +// known at compile time to be false, and using that, we can avoid generating the code of the assert again +// and again for all these expressions that don't need it. + +template<typename Derived, typename OtherDerived, + bool MightHaveTransposeAliasing + = check_transpose_aliasing_compile_time_selector + <blas_traits<Derived>::IsTransposed,OtherDerived>::ret + > +struct checkTransposeAliasing_impl +{ + static void run(const Derived& dst, const OtherDerived& other) + { + eigen_assert((!check_transpose_aliasing_run_time_selector + <typename Derived::Scalar,blas_traits<Derived>::IsTransposed,OtherDerived> + ::run(extract_data(dst), other)) + && "aliasing detected during transposition, use transposeInPlace() " + "or evaluate the rhs into a temporary using .eval()"); + + } +}; + +template<typename Derived, typename OtherDerived> +struct checkTransposeAliasing_impl<Derived, OtherDerived, false> +{ + static void run(const Derived&, const OtherDerived&) + { + } +}; + +template<typename Dst, typename Src> +void check_for_aliasing(const Dst &dst, const Src &src) +{ + internal::checkTransposeAliasing_impl<Dst, Src>::run(dst, src); +} + +} // end namespace internal + +#endif // EIGEN_NO_DEBUG + +} // end namespace Eigen + +#endif // EIGEN_TRANSPOSE_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/Transpositions.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Transpositions.h new file mode 100644 index 000000000..19c17bb4a --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Transpositions.h @@ -0,0 +1,407 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2010-2011 Gael Guennebaud <gael.guennebaud@inria.fr> +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_TRANSPOSITIONS_H +#define EIGEN_TRANSPOSITIONS_H + +namespace Eigen { + +template<typename Derived> +class TranspositionsBase +{ + typedef internal::traits<Derived> Traits; + + public: + + typedef typename Traits::IndicesType IndicesType; + typedef typename IndicesType::Scalar StorageIndex; + typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3 + + Derived& derived() { return *static_cast<Derived*>(this); } + const Derived& derived() const { return *static_cast<const Derived*>(this); } + + /** Copies the \a other transpositions into \c *this */ + template<typename OtherDerived> + Derived& operator=(const TranspositionsBase<OtherDerived>& other) + { + indices() = other.indices(); + return derived(); + } + + #ifndef EIGEN_PARSED_BY_DOXYGEN + /** This is a special case of the templated operator=. Its purpose is to + * prevent a default operator= from hiding the templated operator=. + */ + Derived& operator=(const TranspositionsBase& other) + { + indices() = other.indices(); + return derived(); + } + #endif + + /** \returns the number of transpositions */ + Index size() const { return indices().size(); } + /** \returns the number of rows of the equivalent permutation matrix */ + Index rows() const { return indices().size(); } + /** \returns the number of columns of the equivalent permutation matrix */ + Index cols() const { return indices().size(); } + + /** Direct access to the underlying index vector */ + inline const StorageIndex& coeff(Index i) const { return indices().coeff(i); } + /** Direct access to the underlying index vector */ + inline StorageIndex& coeffRef(Index i) { return indices().coeffRef(i); } + /** Direct access to the underlying index vector */ + inline const StorageIndex& operator()(Index i) const { return indices()(i); } + /** Direct access to the underlying index vector */ + inline StorageIndex& operator()(Index i) { return indices()(i); } + /** Direct access to the underlying index vector */ + inline const StorageIndex& operator[](Index i) const { return indices()(i); } + /** Direct access to the underlying index vector */ + inline StorageIndex& operator[](Index i) { return indices()(i); } + + /** const version of indices(). */ + const IndicesType& indices() const { return derived().indices(); } + /** \returns a reference to the stored array representing the transpositions. */ + IndicesType& indices() { return derived().indices(); } + + /** Resizes to given size. */ + inline void resize(Index newSize) + { + indices().resize(newSize); + } + + /** Sets \c *this to represents an identity transformation */ + void setIdentity() + { + for(StorageIndex i = 0; i < indices().size(); ++i) + coeffRef(i) = i; + } + + // FIXME: do we want such methods ? + // might be usefull when the target matrix expression is complex, e.g.: + // object.matrix().block(..,..,..,..) = trans * object.matrix().block(..,..,..,..); + /* + template<typename MatrixType> + void applyForwardToRows(MatrixType& mat) const + { + for(Index k=0 ; k<size() ; ++k) + if(m_indices(k)!=k) + mat.row(k).swap(mat.row(m_indices(k))); + } + + template<typename MatrixType> + void applyBackwardToRows(MatrixType& mat) const + { + for(Index k=size()-1 ; k>=0 ; --k) + if(m_indices(k)!=k) + mat.row(k).swap(mat.row(m_indices(k))); + } + */ + + /** \returns the inverse transformation */ + inline Transpose<TranspositionsBase> inverse() const + { return Transpose<TranspositionsBase>(derived()); } + + /** \returns the tranpose transformation */ + inline Transpose<TranspositionsBase> transpose() const + { return Transpose<TranspositionsBase>(derived()); } + + protected: +}; + +namespace internal { +template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex> +struct traits<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex> > + : traits<PermutationMatrix<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex> > +{ + typedef Matrix<_StorageIndex, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1> IndicesType; + typedef TranspositionsStorage StorageKind; +}; +} + +/** \class Transpositions + * \ingroup Core_Module + * + * \brief Represents a sequence of transpositions (row/column interchange) + * + * \tparam SizeAtCompileTime the number of transpositions, or Dynamic + * \tparam MaxSizeAtCompileTime the maximum number of transpositions, or Dynamic. This optional parameter defaults to SizeAtCompileTime. Most of the time, you should not have to specify it. + * + * This class represents a permutation transformation as a sequence of \em n transpositions + * \f$[T_{n-1} \ldots T_{i} \ldots T_{0}]\f$. It is internally stored as a vector of integers \c indices. + * Each transposition \f$ T_{i} \f$ applied on the left of a matrix (\f$ T_{i} M\f$) interchanges + * the rows \c i and \c indices[i] of the matrix \c M. + * A transposition applied on the right (e.g., \f$ M T_{i}\f$) yields a column interchange. + * + * Compared to the class PermutationMatrix, such a sequence of transpositions is what is + * computed during a decomposition with pivoting, and it is faster when applying the permutation in-place. + * + * To apply a sequence of transpositions to a matrix, simply use the operator * as in the following example: + * \code + * Transpositions tr; + * MatrixXf mat; + * mat = tr * mat; + * \endcode + * In this example, we detect that the matrix appears on both side, and so the transpositions + * are applied in-place without any temporary or extra copy. + * + * \sa class PermutationMatrix + */ + +template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex> +class Transpositions : public TranspositionsBase<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex> > +{ + typedef internal::traits<Transpositions> Traits; + public: + + typedef TranspositionsBase<Transpositions> Base; + typedef typename Traits::IndicesType IndicesType; + typedef typename IndicesType::Scalar StorageIndex; + + inline Transpositions() {} + + /** Copy constructor. */ + template<typename OtherDerived> + inline Transpositions(const TranspositionsBase<OtherDerived>& other) + : m_indices(other.indices()) {} + + #ifndef EIGEN_PARSED_BY_DOXYGEN + /** Standard copy constructor. Defined only to prevent a default copy constructor + * from hiding the other templated constructor */ + inline Transpositions(const Transpositions& other) : m_indices(other.indices()) {} + #endif + + /** Generic constructor from expression of the transposition indices. */ + template<typename Other> + explicit inline Transpositions(const MatrixBase<Other>& indices) : m_indices(indices) + {} + + /** Copies the \a other transpositions into \c *this */ + template<typename OtherDerived> + Transpositions& operator=(const TranspositionsBase<OtherDerived>& other) + { + return Base::operator=(other); + } + + #ifndef EIGEN_PARSED_BY_DOXYGEN + /** This is a special case of the templated operator=. Its purpose is to + * prevent a default operator= from hiding the templated operator=. + */ + Transpositions& operator=(const Transpositions& other) + { + m_indices = other.m_indices; + return *this; + } + #endif + + /** Constructs an uninitialized permutation matrix of given size. + */ + inline Transpositions(Index size) : m_indices(size) + {} + + /** const version of indices(). */ + const IndicesType& indices() const { return m_indices; } + /** \returns a reference to the stored array representing the transpositions. */ + IndicesType& indices() { return m_indices; } + + protected: + + IndicesType m_indices; +}; + + +namespace internal { +template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex, int _PacketAccess> +struct traits<Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex>,_PacketAccess> > + : traits<PermutationMatrix<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex> > +{ + typedef Map<const Matrix<_StorageIndex,SizeAtCompileTime,1,0,MaxSizeAtCompileTime,1>, _PacketAccess> IndicesType; + typedef _StorageIndex StorageIndex; + typedef TranspositionsStorage StorageKind; +}; +} + +template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex, int PacketAccess> +class Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex>,PacketAccess> + : public TranspositionsBase<Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex>,PacketAccess> > +{ + typedef internal::traits<Map> Traits; + public: + + typedef TranspositionsBase<Map> Base; + typedef typename Traits::IndicesType IndicesType; + typedef typename IndicesType::Scalar StorageIndex; + + explicit inline Map(const StorageIndex* indicesPtr) + : m_indices(indicesPtr) + {} + + inline Map(const StorageIndex* indicesPtr, Index size) + : m_indices(indicesPtr,size) + {} + + /** Copies the \a other transpositions into \c *this */ + template<typename OtherDerived> + Map& operator=(const TranspositionsBase<OtherDerived>& other) + { + return Base::operator=(other); + } + + #ifndef EIGEN_PARSED_BY_DOXYGEN + /** This is a special case of the templated operator=. Its purpose is to + * prevent a default operator= from hiding the templated operator=. + */ + Map& operator=(const Map& other) + { + m_indices = other.m_indices; + return *this; + } + #endif + + /** const version of indices(). */ + const IndicesType& indices() const { return m_indices; } + + /** \returns a reference to the stored array representing the transpositions. */ + IndicesType& indices() { return m_indices; } + + protected: + + IndicesType m_indices; +}; + +namespace internal { +template<typename _IndicesType> +struct traits<TranspositionsWrapper<_IndicesType> > + : traits<PermutationWrapper<_IndicesType> > +{ + typedef TranspositionsStorage StorageKind; +}; +} + +template<typename _IndicesType> +class TranspositionsWrapper + : public TranspositionsBase<TranspositionsWrapper<_IndicesType> > +{ + typedef internal::traits<TranspositionsWrapper> Traits; + public: + + typedef TranspositionsBase<TranspositionsWrapper> Base; + typedef typename Traits::IndicesType IndicesType; + typedef typename IndicesType::Scalar StorageIndex; + + explicit inline TranspositionsWrapper(IndicesType& indices) + : m_indices(indices) + {} + + /** Copies the \a other transpositions into \c *this */ + template<typename OtherDerived> + TranspositionsWrapper& operator=(const TranspositionsBase<OtherDerived>& other) + { + return Base::operator=(other); + } + + #ifndef EIGEN_PARSED_BY_DOXYGEN + /** This is a special case of the templated operator=. Its purpose is to + * prevent a default operator= from hiding the templated operator=. + */ + TranspositionsWrapper& operator=(const TranspositionsWrapper& other) + { + m_indices = other.m_indices; + return *this; + } + #endif + + /** const version of indices(). */ + const IndicesType& indices() const { return m_indices; } + + /** \returns a reference to the stored array representing the transpositions. */ + IndicesType& indices() { return m_indices; } + + protected: + + typename IndicesType::Nested m_indices; +}; + + + +/** \returns the \a matrix with the \a transpositions applied to the columns. + */ +template<typename MatrixDerived, typename TranspositionsDerived> +EIGEN_DEVICE_FUNC +const Product<MatrixDerived, TranspositionsDerived, AliasFreeProduct> +operator*(const MatrixBase<MatrixDerived> &matrix, + const TranspositionsBase<TranspositionsDerived>& transpositions) +{ + return Product<MatrixDerived, TranspositionsDerived, AliasFreeProduct> + (matrix.derived(), transpositions.derived()); +} + +/** \returns the \a matrix with the \a transpositions applied to the rows. + */ +template<typename TranspositionsDerived, typename MatrixDerived> +EIGEN_DEVICE_FUNC +const Product<TranspositionsDerived, MatrixDerived, AliasFreeProduct> +operator*(const TranspositionsBase<TranspositionsDerived> &transpositions, + const MatrixBase<MatrixDerived>& matrix) +{ + return Product<TranspositionsDerived, MatrixDerived, AliasFreeProduct> + (transpositions.derived(), matrix.derived()); +} + +// Template partial specialization for transposed/inverse transpositions + +namespace internal { + +template<typename Derived> +struct traits<Transpose<TranspositionsBase<Derived> > > + : traits<Derived> +{}; + +} // end namespace internal + +template<typename TranspositionsDerived> +class Transpose<TranspositionsBase<TranspositionsDerived> > +{ + typedef TranspositionsDerived TranspositionType; + typedef typename TranspositionType::IndicesType IndicesType; + public: + + explicit Transpose(const TranspositionType& t) : m_transpositions(t) {} + + Index size() const { return m_transpositions.size(); } + Index rows() const { return m_transpositions.size(); } + Index cols() const { return m_transpositions.size(); } + + /** \returns the \a matrix with the inverse transpositions applied to the columns. + */ + template<typename OtherDerived> friend + const Product<OtherDerived, Transpose, AliasFreeProduct> + operator*(const MatrixBase<OtherDerived>& matrix, const Transpose& trt) + { + return Product<OtherDerived, Transpose, AliasFreeProduct>(matrix.derived(), trt.derived()); + } + + /** \returns the \a matrix with the inverse transpositions applied to the rows. + */ + template<typename OtherDerived> + const Product<Transpose, OtherDerived, AliasFreeProduct> + operator*(const MatrixBase<OtherDerived>& matrix) const + { + return Product<Transpose, OtherDerived, AliasFreeProduct>(*this, matrix.derived()); + } + + const TranspositionType& nestedExpression() const { return m_transpositions; } + + protected: + const TranspositionType& m_transpositions; +}; + +} // end namespace Eigen + +#endif // EIGEN_TRANSPOSITIONS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/TriangularMatrix.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/TriangularMatrix.h new file mode 100644 index 000000000..667ef09dc --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/TriangularMatrix.h @@ -0,0 +1,983 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008 Benoit Jacob <jacob.benoit.1@gmail.com> +// Copyright (C) 2008-2009 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_TRIANGULARMATRIX_H +#define EIGEN_TRIANGULARMATRIX_H + +namespace Eigen { + +namespace internal { + +template<int Side, typename TriangularType, typename Rhs> struct triangular_solve_retval; + +} + +/** \class TriangularBase + * \ingroup Core_Module + * + * \brief Base class for triangular part in a matrix + */ +template<typename Derived> class TriangularBase : public EigenBase<Derived> +{ + public: + + enum { + Mode = internal::traits<Derived>::Mode, + RowsAtCompileTime = internal::traits<Derived>::RowsAtCompileTime, + ColsAtCompileTime = internal::traits<Derived>::ColsAtCompileTime, + MaxRowsAtCompileTime = internal::traits<Derived>::MaxRowsAtCompileTime, + MaxColsAtCompileTime = internal::traits<Derived>::MaxColsAtCompileTime, + + SizeAtCompileTime = (internal::size_at_compile_time<internal::traits<Derived>::RowsAtCompileTime, + internal::traits<Derived>::ColsAtCompileTime>::ret), + /**< This is equal to the number of coefficients, i.e. the number of + * rows times the number of columns, or to \a Dynamic if this is not + * known at compile-time. \sa RowsAtCompileTime, ColsAtCompileTime */ + + MaxSizeAtCompileTime = (internal::size_at_compile_time<internal::traits<Derived>::MaxRowsAtCompileTime, + internal::traits<Derived>::MaxColsAtCompileTime>::ret) + + }; + typedef typename internal::traits<Derived>::Scalar Scalar; + typedef typename internal::traits<Derived>::StorageKind StorageKind; + typedef typename internal::traits<Derived>::StorageIndex StorageIndex; + typedef typename internal::traits<Derived>::FullMatrixType DenseMatrixType; + typedef DenseMatrixType DenseType; + typedef Derived const& Nested; + + EIGEN_DEVICE_FUNC + inline TriangularBase() { eigen_assert(!((Mode&UnitDiag) && (Mode&ZeroDiag))); } + + EIGEN_DEVICE_FUNC + inline Index rows() const { return derived().rows(); } + EIGEN_DEVICE_FUNC + inline Index cols() const { return derived().cols(); } + EIGEN_DEVICE_FUNC + inline Index outerStride() const { return derived().outerStride(); } + EIGEN_DEVICE_FUNC + inline Index innerStride() const { return derived().innerStride(); } + + // dummy resize function + void resize(Index rows, Index cols) + { + EIGEN_UNUSED_VARIABLE(rows); + EIGEN_UNUSED_VARIABLE(cols); + eigen_assert(rows==this->rows() && cols==this->cols()); + } + + EIGEN_DEVICE_FUNC + inline Scalar coeff(Index row, Index col) const { return derived().coeff(row,col); } + EIGEN_DEVICE_FUNC + inline Scalar& coeffRef(Index row, Index col) { return derived().coeffRef(row,col); } + + /** \see MatrixBase::copyCoeff(row,col) + */ + template<typename Other> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void copyCoeff(Index row, Index col, Other& other) + { + derived().coeffRef(row, col) = other.coeff(row, col); + } + + EIGEN_DEVICE_FUNC + inline Scalar operator()(Index row, Index col) const + { + check_coordinates(row, col); + return coeff(row,col); + } + EIGEN_DEVICE_FUNC + inline Scalar& operator()(Index row, Index col) + { + check_coordinates(row, col); + return coeffRef(row,col); + } + + #ifndef EIGEN_PARSED_BY_DOXYGEN + EIGEN_DEVICE_FUNC + inline const Derived& derived() const { return *static_cast<const Derived*>(this); } + EIGEN_DEVICE_FUNC + inline Derived& derived() { return *static_cast<Derived*>(this); } + #endif // not EIGEN_PARSED_BY_DOXYGEN + + template<typename DenseDerived> + EIGEN_DEVICE_FUNC + void evalTo(MatrixBase<DenseDerived> &other) const; + template<typename DenseDerived> + EIGEN_DEVICE_FUNC + void evalToLazy(MatrixBase<DenseDerived> &other) const; + + EIGEN_DEVICE_FUNC + DenseMatrixType toDenseMatrix() const + { + DenseMatrixType res(rows(), cols()); + evalToLazy(res); + return res; + } + + protected: + + void check_coordinates(Index row, Index col) const + { + EIGEN_ONLY_USED_FOR_DEBUG(row); + EIGEN_ONLY_USED_FOR_DEBUG(col); + eigen_assert(col>=0 && col<cols() && row>=0 && row<rows()); + const int mode = int(Mode) & ~SelfAdjoint; + EIGEN_ONLY_USED_FOR_DEBUG(mode); + eigen_assert((mode==Upper && col>=row) + || (mode==Lower && col<=row) + || ((mode==StrictlyUpper || mode==UnitUpper) && col>row) + || ((mode==StrictlyLower || mode==UnitLower) && col<row)); + } + + #ifdef EIGEN_INTERNAL_DEBUGGING + void check_coordinates_internal(Index row, Index col) const + { + check_coordinates(row, col); + } + #else + void check_coordinates_internal(Index , Index ) const {} + #endif + +}; + +/** \class TriangularView + * \ingroup Core_Module + * + * \brief Expression of a triangular part in a matrix + * + * \param MatrixType the type of the object in which we are taking the triangular part + * \param Mode the kind of triangular matrix expression to construct. Can be #Upper, + * #Lower, #UnitUpper, #UnitLower, #StrictlyUpper, or #StrictlyLower. + * This is in fact a bit field; it must have either #Upper or #Lower, + * and additionally it may have #UnitDiag or #ZeroDiag or neither. + * + * This class represents a triangular part of a matrix, not necessarily square. Strictly speaking, for rectangular + * matrices one should speak of "trapezoid" parts. This class is the return type + * of MatrixBase::triangularView() and SparseMatrixBase::triangularView(), and most of the time this is the only way it is used. + * + * \sa MatrixBase::triangularView() + */ +namespace internal { +template<typename MatrixType, unsigned int _Mode> +struct traits<TriangularView<MatrixType, _Mode> > : traits<MatrixType> +{ + typedef typename ref_selector<MatrixType>::non_const_type MatrixTypeNested; + typedef typename remove_reference<MatrixTypeNested>::type MatrixTypeNestedNonRef; + typedef typename remove_all<MatrixTypeNested>::type MatrixTypeNestedCleaned; + typedef typename MatrixType::PlainObject FullMatrixType; + typedef MatrixType ExpressionType; + enum { + Mode = _Mode, + FlagsLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0, + Flags = (MatrixTypeNestedCleaned::Flags & (HereditaryBits | FlagsLvalueBit) & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit))) + }; +}; +} + +template<typename _MatrixType, unsigned int _Mode, typename StorageKind> class TriangularViewImpl; + +template<typename _MatrixType, unsigned int _Mode> class TriangularView + : public TriangularViewImpl<_MatrixType, _Mode, typename internal::traits<_MatrixType>::StorageKind > +{ + public: + + typedef TriangularViewImpl<_MatrixType, _Mode, typename internal::traits<_MatrixType>::StorageKind > Base; + typedef typename internal::traits<TriangularView>::Scalar Scalar; + typedef _MatrixType MatrixType; + + protected: + typedef typename internal::traits<TriangularView>::MatrixTypeNested MatrixTypeNested; + typedef typename internal::traits<TriangularView>::MatrixTypeNestedNonRef MatrixTypeNestedNonRef; + + typedef typename internal::remove_all<typename MatrixType::ConjugateReturnType>::type MatrixConjugateReturnType; + + public: + + typedef typename internal::traits<TriangularView>::StorageKind StorageKind; + typedef typename internal::traits<TriangularView>::MatrixTypeNestedCleaned NestedExpression; + + enum { + Mode = _Mode, + Flags = internal::traits<TriangularView>::Flags, + TransposeMode = (Mode & Upper ? Lower : 0) + | (Mode & Lower ? Upper : 0) + | (Mode & (UnitDiag)) + | (Mode & (ZeroDiag)), + IsVectorAtCompileTime = false + }; + + EIGEN_DEVICE_FUNC + explicit inline TriangularView(MatrixType& matrix) : m_matrix(matrix) + {} + + using Base::operator=; + TriangularView& operator=(const TriangularView &other) + { return Base::operator=(other); } + + /** \copydoc EigenBase::rows() */ + EIGEN_DEVICE_FUNC + inline Index rows() const { return m_matrix.rows(); } + /** \copydoc EigenBase::cols() */ + EIGEN_DEVICE_FUNC + inline Index cols() const { return m_matrix.cols(); } + + /** \returns a const reference to the nested expression */ + EIGEN_DEVICE_FUNC + const NestedExpression& nestedExpression() const { return m_matrix; } + + /** \returns a reference to the nested expression */ + EIGEN_DEVICE_FUNC + NestedExpression& nestedExpression() { return m_matrix; } + + typedef TriangularView<const MatrixConjugateReturnType,Mode> ConjugateReturnType; + /** \sa MatrixBase::conjugate() const */ + EIGEN_DEVICE_FUNC + inline const ConjugateReturnType conjugate() const + { return ConjugateReturnType(m_matrix.conjugate()); } + + typedef TriangularView<const typename MatrixType::AdjointReturnType,TransposeMode> AdjointReturnType; + /** \sa MatrixBase::adjoint() const */ + EIGEN_DEVICE_FUNC + inline const AdjointReturnType adjoint() const + { return AdjointReturnType(m_matrix.adjoint()); } + + typedef TriangularView<typename MatrixType::TransposeReturnType,TransposeMode> TransposeReturnType; + /** \sa MatrixBase::transpose() */ + EIGEN_DEVICE_FUNC + inline TransposeReturnType transpose() + { + EIGEN_STATIC_ASSERT_LVALUE(MatrixType) + typename MatrixType::TransposeReturnType tmp(m_matrix); + return TransposeReturnType(tmp); + } + + typedef TriangularView<const typename MatrixType::ConstTransposeReturnType,TransposeMode> ConstTransposeReturnType; + /** \sa MatrixBase::transpose() const */ + EIGEN_DEVICE_FUNC + inline const ConstTransposeReturnType transpose() const + { + return ConstTransposeReturnType(m_matrix.transpose()); + } + + template<typename Other> + EIGEN_DEVICE_FUNC + inline const Solve<TriangularView, Other> + solve(const MatrixBase<Other>& other) const + { return Solve<TriangularView, Other>(*this, other.derived()); } + + // workaround MSVC ICE + #if EIGEN_COMP_MSVC + template<int Side, typename Other> + EIGEN_DEVICE_FUNC + inline const internal::triangular_solve_retval<Side,TriangularView, Other> + solve(const MatrixBase<Other>& other) const + { return Base::template solve<Side>(other); } + #else + using Base::solve; + #endif + + /** \returns a selfadjoint view of the referenced triangular part which must be either \c #Upper or \c #Lower. + * + * This is a shortcut for \code this->nestedExpression().selfadjointView<(*this)::Mode>() \endcode + * \sa MatrixBase::selfadjointView() */ + EIGEN_DEVICE_FUNC + SelfAdjointView<MatrixTypeNestedNonRef,Mode> selfadjointView() + { + EIGEN_STATIC_ASSERT((Mode&(UnitDiag|ZeroDiag))==0,PROGRAMMING_ERROR); + return SelfAdjointView<MatrixTypeNestedNonRef,Mode>(m_matrix); + } + + /** This is the const version of selfadjointView() */ + EIGEN_DEVICE_FUNC + const SelfAdjointView<MatrixTypeNestedNonRef,Mode> selfadjointView() const + { + EIGEN_STATIC_ASSERT((Mode&(UnitDiag|ZeroDiag))==0,PROGRAMMING_ERROR); + return SelfAdjointView<MatrixTypeNestedNonRef,Mode>(m_matrix); + } + + + /** \returns the determinant of the triangular matrix + * \sa MatrixBase::determinant() */ + EIGEN_DEVICE_FUNC + Scalar determinant() const + { + if (Mode & UnitDiag) + return 1; + else if (Mode & ZeroDiag) + return 0; + else + return m_matrix.diagonal().prod(); + } + + protected: + + MatrixTypeNested m_matrix; +}; + +/** \ingroup Core_Module + * + * \brief Base class for a triangular part in a \b dense matrix + * + * This class is an abstract base class of class TriangularView, and objects of type TriangularViewImpl cannot be instantiated. + * It extends class TriangularView with additional methods which available for dense expressions only. + * + * \sa class TriangularView, MatrixBase::triangularView() + */ +template<typename _MatrixType, unsigned int _Mode> class TriangularViewImpl<_MatrixType,_Mode,Dense> + : public TriangularBase<TriangularView<_MatrixType, _Mode> > +{ + public: + + typedef TriangularView<_MatrixType, _Mode> TriangularViewType; + typedef TriangularBase<TriangularViewType> Base; + typedef typename internal::traits<TriangularViewType>::Scalar Scalar; + + typedef _MatrixType MatrixType; + typedef typename MatrixType::PlainObject DenseMatrixType; + typedef DenseMatrixType PlainObject; + + public: + using Base::evalToLazy; + using Base::derived; + + typedef typename internal::traits<TriangularViewType>::StorageKind StorageKind; + + enum { + Mode = _Mode, + Flags = internal::traits<TriangularViewType>::Flags + }; + + /** \returns the outer-stride of the underlying dense matrix + * \sa DenseCoeffsBase::outerStride() */ + EIGEN_DEVICE_FUNC + inline Index outerStride() const { return derived().nestedExpression().outerStride(); } + /** \returns the inner-stride of the underlying dense matrix + * \sa DenseCoeffsBase::innerStride() */ + EIGEN_DEVICE_FUNC + inline Index innerStride() const { return derived().nestedExpression().innerStride(); } + + /** \sa MatrixBase::operator+=() */ + template<typename Other> + EIGEN_DEVICE_FUNC + TriangularViewType& operator+=(const DenseBase<Other>& other) { + internal::call_assignment_no_alias(derived(), other.derived(), internal::add_assign_op<Scalar,typename Other::Scalar>()); + return derived(); + } + /** \sa MatrixBase::operator-=() */ + template<typename Other> + EIGEN_DEVICE_FUNC + TriangularViewType& operator-=(const DenseBase<Other>& other) { + internal::call_assignment_no_alias(derived(), other.derived(), internal::sub_assign_op<Scalar,typename Other::Scalar>()); + return derived(); + } + + /** \sa MatrixBase::operator*=() */ + EIGEN_DEVICE_FUNC + TriangularViewType& operator*=(const typename internal::traits<MatrixType>::Scalar& other) { return *this = derived().nestedExpression() * other; } + /** \sa DenseBase::operator/=() */ + EIGEN_DEVICE_FUNC + TriangularViewType& operator/=(const typename internal::traits<MatrixType>::Scalar& other) { return *this = derived().nestedExpression() / other; } + + /** \sa MatrixBase::fill() */ + EIGEN_DEVICE_FUNC + void fill(const Scalar& value) { setConstant(value); } + /** \sa MatrixBase::setConstant() */ + EIGEN_DEVICE_FUNC + TriangularViewType& setConstant(const Scalar& value) + { return *this = MatrixType::Constant(derived().rows(), derived().cols(), value); } + /** \sa MatrixBase::setZero() */ + EIGEN_DEVICE_FUNC + TriangularViewType& setZero() { return setConstant(Scalar(0)); } + /** \sa MatrixBase::setOnes() */ + EIGEN_DEVICE_FUNC + TriangularViewType& setOnes() { return setConstant(Scalar(1)); } + + /** \sa MatrixBase::coeff() + * \warning the coordinates must fit into the referenced triangular part + */ + EIGEN_DEVICE_FUNC + inline Scalar coeff(Index row, Index col) const + { + Base::check_coordinates_internal(row, col); + return derived().nestedExpression().coeff(row, col); + } + + /** \sa MatrixBase::coeffRef() + * \warning the coordinates must fit into the referenced triangular part + */ + EIGEN_DEVICE_FUNC + inline Scalar& coeffRef(Index row, Index col) + { + EIGEN_STATIC_ASSERT_LVALUE(TriangularViewType); + Base::check_coordinates_internal(row, col); + return derived().nestedExpression().coeffRef(row, col); + } + + /** Assigns a triangular matrix to a triangular part of a dense matrix */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + TriangularViewType& operator=(const TriangularBase<OtherDerived>& other); + + /** Shortcut for\code *this = other.other.triangularView<(*this)::Mode>() \endcode */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + TriangularViewType& operator=(const MatrixBase<OtherDerived>& other); + +#ifndef EIGEN_PARSED_BY_DOXYGEN + EIGEN_DEVICE_FUNC + TriangularViewType& operator=(const TriangularViewImpl& other) + { return *this = other.derived().nestedExpression(); } + + /** \deprecated */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + void lazyAssign(const TriangularBase<OtherDerived>& other); + + /** \deprecated */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + void lazyAssign(const MatrixBase<OtherDerived>& other); +#endif + + /** Efficient triangular matrix times vector/matrix product */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + const Product<TriangularViewType,OtherDerived> + operator*(const MatrixBase<OtherDerived>& rhs) const + { + return Product<TriangularViewType,OtherDerived>(derived(), rhs.derived()); + } + + /** Efficient vector/matrix times triangular matrix product */ + template<typename OtherDerived> friend + EIGEN_DEVICE_FUNC + const Product<OtherDerived,TriangularViewType> + operator*(const MatrixBase<OtherDerived>& lhs, const TriangularViewImpl& rhs) + { + return Product<OtherDerived,TriangularViewType>(lhs.derived(),rhs.derived()); + } + + /** \returns the product of the inverse of \c *this with \a other, \a *this being triangular. + * + * This function computes the inverse-matrix matrix product inverse(\c *this) * \a other if + * \a Side==OnTheLeft (the default), or the right-inverse-multiply \a other * inverse(\c *this) if + * \a Side==OnTheRight. + * + * Note that the template parameter \c Side can be ommitted, in which case \c Side==OnTheLeft + * + * The matrix \c *this must be triangular and invertible (i.e., all the coefficients of the + * diagonal must be non zero). It works as a forward (resp. backward) substitution if \c *this + * is an upper (resp. lower) triangular matrix. + * + * Example: \include Triangular_solve.cpp + * Output: \verbinclude Triangular_solve.out + * + * This function returns an expression of the inverse-multiply and can works in-place if it is assigned + * to the same matrix or vector \a other. + * + * For users coming from BLAS, this function (and more specifically solveInPlace()) offer + * all the operations supported by the \c *TRSV and \c *TRSM BLAS routines. + * + * \sa TriangularView::solveInPlace() + */ + template<int Side, typename Other> + EIGEN_DEVICE_FUNC + inline const internal::triangular_solve_retval<Side,TriangularViewType, Other> + solve(const MatrixBase<Other>& other) const; + + /** "in-place" version of TriangularView::solve() where the result is written in \a other + * + * \warning The parameter is only marked 'const' to make the C++ compiler accept a temporary expression here. + * This function will const_cast it, so constness isn't honored here. + * + * Note that the template parameter \c Side can be ommitted, in which case \c Side==OnTheLeft + * + * See TriangularView:solve() for the details. + */ + template<int Side, typename OtherDerived> + EIGEN_DEVICE_FUNC + void solveInPlace(const MatrixBase<OtherDerived>& other) const; + + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + void solveInPlace(const MatrixBase<OtherDerived>& other) const + { return solveInPlace<OnTheLeft>(other); } + + /** Swaps the coefficients of the common triangular parts of two matrices */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC +#ifdef EIGEN_PARSED_BY_DOXYGEN + void swap(TriangularBase<OtherDerived> &other) +#else + void swap(TriangularBase<OtherDerived> const & other) +#endif + { + EIGEN_STATIC_ASSERT_LVALUE(OtherDerived); + call_assignment(derived(), other.const_cast_derived(), internal::swap_assign_op<Scalar>()); + } + + /** \deprecated + * Shortcut for \code (*this).swap(other.triangularView<(*this)::Mode>()) \endcode */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + void swap(MatrixBase<OtherDerived> const & other) + { + EIGEN_STATIC_ASSERT_LVALUE(OtherDerived); + call_assignment(derived(), other.const_cast_derived(), internal::swap_assign_op<Scalar>()); + } + + template<typename RhsType, typename DstType> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _solve_impl(const RhsType &rhs, DstType &dst) const { + if(!internal::is_same_dense(dst,rhs)) + dst = rhs; + this->solveInPlace(dst); + } + + template<typename ProductType> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE TriangularViewType& _assignProduct(const ProductType& prod, const Scalar& alpha, bool beta); +}; + +/*************************************************************************** +* Implementation of triangular evaluation/assignment +***************************************************************************/ + +#ifndef EIGEN_PARSED_BY_DOXYGEN +// FIXME should we keep that possibility +template<typename MatrixType, unsigned int Mode> +template<typename OtherDerived> +inline TriangularView<MatrixType, Mode>& +TriangularViewImpl<MatrixType, Mode, Dense>::operator=(const MatrixBase<OtherDerived>& other) +{ + internal::call_assignment_no_alias(derived(), other.derived(), internal::assign_op<Scalar,typename OtherDerived::Scalar>()); + return derived(); +} + +// FIXME should we keep that possibility +template<typename MatrixType, unsigned int Mode> +template<typename OtherDerived> +void TriangularViewImpl<MatrixType, Mode, Dense>::lazyAssign(const MatrixBase<OtherDerived>& other) +{ + internal::call_assignment_no_alias(derived(), other.template triangularView<Mode>()); +} + + + +template<typename MatrixType, unsigned int Mode> +template<typename OtherDerived> +inline TriangularView<MatrixType, Mode>& +TriangularViewImpl<MatrixType, Mode, Dense>::operator=(const TriangularBase<OtherDerived>& other) +{ + eigen_assert(Mode == int(OtherDerived::Mode)); + internal::call_assignment(derived(), other.derived()); + return derived(); +} + +template<typename MatrixType, unsigned int Mode> +template<typename OtherDerived> +void TriangularViewImpl<MatrixType, Mode, Dense>::lazyAssign(const TriangularBase<OtherDerived>& other) +{ + eigen_assert(Mode == int(OtherDerived::Mode)); + internal::call_assignment_no_alias(derived(), other.derived()); +} +#endif + +/*************************************************************************** +* Implementation of TriangularBase methods +***************************************************************************/ + +/** Assigns a triangular or selfadjoint matrix to a dense matrix. + * If the matrix is triangular, the opposite part is set to zero. */ +template<typename Derived> +template<typename DenseDerived> +void TriangularBase<Derived>::evalTo(MatrixBase<DenseDerived> &other) const +{ + evalToLazy(other.derived()); +} + +/*************************************************************************** +* Implementation of TriangularView methods +***************************************************************************/ + +/*************************************************************************** +* Implementation of MatrixBase methods +***************************************************************************/ + +/** + * \returns an expression of a triangular view extracted from the current matrix + * + * The parameter \a Mode can have the following values: \c #Upper, \c #StrictlyUpper, \c #UnitUpper, + * \c #Lower, \c #StrictlyLower, \c #UnitLower. + * + * Example: \include MatrixBase_triangularView.cpp + * Output: \verbinclude MatrixBase_triangularView.out + * + * \sa class TriangularView + */ +template<typename Derived> +template<unsigned int Mode> +typename MatrixBase<Derived>::template TriangularViewReturnType<Mode>::Type +MatrixBase<Derived>::triangularView() +{ + return typename TriangularViewReturnType<Mode>::Type(derived()); +} + +/** This is the const version of MatrixBase::triangularView() */ +template<typename Derived> +template<unsigned int Mode> +typename MatrixBase<Derived>::template ConstTriangularViewReturnType<Mode>::Type +MatrixBase<Derived>::triangularView() const +{ + return typename ConstTriangularViewReturnType<Mode>::Type(derived()); +} + +/** \returns true if *this is approximately equal to an upper triangular matrix, + * within the precision given by \a prec. + * + * \sa isLowerTriangular() + */ +template<typename Derived> +bool MatrixBase<Derived>::isUpperTriangular(const RealScalar& prec) const +{ + RealScalar maxAbsOnUpperPart = static_cast<RealScalar>(-1); + for(Index j = 0; j < cols(); ++j) + { + Index maxi = numext::mini(j, rows()-1); + for(Index i = 0; i <= maxi; ++i) + { + RealScalar absValue = numext::abs(coeff(i,j)); + if(absValue > maxAbsOnUpperPart) maxAbsOnUpperPart = absValue; + } + } + RealScalar threshold = maxAbsOnUpperPart * prec; + for(Index j = 0; j < cols(); ++j) + for(Index i = j+1; i < rows(); ++i) + if(numext::abs(coeff(i, j)) > threshold) return false; + return true; +} + +/** \returns true if *this is approximately equal to a lower triangular matrix, + * within the precision given by \a prec. + * + * \sa isUpperTriangular() + */ +template<typename Derived> +bool MatrixBase<Derived>::isLowerTriangular(const RealScalar& prec) const +{ + RealScalar maxAbsOnLowerPart = static_cast<RealScalar>(-1); + for(Index j = 0; j < cols(); ++j) + for(Index i = j; i < rows(); ++i) + { + RealScalar absValue = numext::abs(coeff(i,j)); + if(absValue > maxAbsOnLowerPart) maxAbsOnLowerPart = absValue; + } + RealScalar threshold = maxAbsOnLowerPart * prec; + for(Index j = 1; j < cols(); ++j) + { + Index maxi = numext::mini(j, rows()-1); + for(Index i = 0; i < maxi; ++i) + if(numext::abs(coeff(i, j)) > threshold) return false; + } + return true; +} + + +/*************************************************************************** +**************************************************************************** +* Evaluators and Assignment of triangular expressions +*************************************************************************** +***************************************************************************/ + +namespace internal { + + +// TODO currently a triangular expression has the form TriangularView<.,.> +// in the future triangular-ness should be defined by the expression traits +// such that Transpose<TriangularView<.,.> > is valid. (currently TriangularBase::transpose() is overloaded to make it work) +template<typename MatrixType, unsigned int Mode> +struct evaluator_traits<TriangularView<MatrixType,Mode> > +{ + typedef typename storage_kind_to_evaluator_kind<typename MatrixType::StorageKind>::Kind Kind; + typedef typename glue_shapes<typename evaluator_traits<MatrixType>::Shape, TriangularShape>::type Shape; +}; + +template<typename MatrixType, unsigned int Mode> +struct unary_evaluator<TriangularView<MatrixType,Mode>, IndexBased> + : evaluator<typename internal::remove_all<MatrixType>::type> +{ + typedef TriangularView<MatrixType,Mode> XprType; + typedef evaluator<typename internal::remove_all<MatrixType>::type> Base; + unary_evaluator(const XprType &xpr) : Base(xpr.nestedExpression()) {} +}; + +// Additional assignment kinds: +struct Triangular2Triangular {}; +struct Triangular2Dense {}; +struct Dense2Triangular {}; + + +template<typename Kernel, unsigned int Mode, int UnrollCount, bool ClearOpposite> struct triangular_assignment_loop; + + +/** \internal Specialization of the dense assignment kernel for triangular matrices. + * The main difference is that the triangular, diagonal, and opposite parts are processed through three different functions. + * \tparam UpLo must be either Lower or Upper + * \tparam Mode must be either 0, UnitDiag, ZeroDiag, or SelfAdjoint + */ +template<int UpLo, int Mode, int SetOpposite, typename DstEvaluatorTypeT, typename SrcEvaluatorTypeT, typename Functor, int Version = Specialized> +class triangular_dense_assignment_kernel : public generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, Functor, Version> +{ +protected: + typedef generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, Functor, Version> Base; + typedef typename Base::DstXprType DstXprType; + typedef typename Base::SrcXprType SrcXprType; + using Base::m_dst; + using Base::m_src; + using Base::m_functor; +public: + + typedef typename Base::DstEvaluatorType DstEvaluatorType; + typedef typename Base::SrcEvaluatorType SrcEvaluatorType; + typedef typename Base::Scalar Scalar; + typedef typename Base::AssignmentTraits AssignmentTraits; + + + EIGEN_DEVICE_FUNC triangular_dense_assignment_kernel(DstEvaluatorType &dst, const SrcEvaluatorType &src, const Functor &func, DstXprType& dstExpr) + : Base(dst, src, func, dstExpr) + {} + +#ifdef EIGEN_INTERNAL_DEBUGGING + EIGEN_DEVICE_FUNC void assignCoeff(Index row, Index col) + { + eigen_internal_assert(row!=col); + Base::assignCoeff(row,col); + } +#else + using Base::assignCoeff; +#endif + + EIGEN_DEVICE_FUNC void assignDiagonalCoeff(Index id) + { + if(Mode==UnitDiag && SetOpposite) m_functor.assignCoeff(m_dst.coeffRef(id,id), Scalar(1)); + else if(Mode==ZeroDiag && SetOpposite) m_functor.assignCoeff(m_dst.coeffRef(id,id), Scalar(0)); + else if(Mode==0) Base::assignCoeff(id,id); + } + + EIGEN_DEVICE_FUNC void assignOppositeCoeff(Index row, Index col) + { + eigen_internal_assert(row!=col); + if(SetOpposite) + m_functor.assignCoeff(m_dst.coeffRef(row,col), Scalar(0)); + } +}; + +template<int Mode, bool SetOpposite, typename DstXprType, typename SrcXprType, typename Functor> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +void call_triangular_assignment_loop(DstXprType& dst, const SrcXprType& src, const Functor &func) +{ + typedef evaluator<DstXprType> DstEvaluatorType; + typedef evaluator<SrcXprType> SrcEvaluatorType; + + SrcEvaluatorType srcEvaluator(src); + + Index dstRows = src.rows(); + Index dstCols = src.cols(); + if((dst.rows()!=dstRows) || (dst.cols()!=dstCols)) + dst.resize(dstRows, dstCols); + DstEvaluatorType dstEvaluator(dst); + + typedef triangular_dense_assignment_kernel< Mode&(Lower|Upper),Mode&(UnitDiag|ZeroDiag|SelfAdjoint),SetOpposite, + DstEvaluatorType,SrcEvaluatorType,Functor> Kernel; + Kernel kernel(dstEvaluator, srcEvaluator, func, dst.const_cast_derived()); + + enum { + unroll = DstXprType::SizeAtCompileTime != Dynamic + && SrcEvaluatorType::CoeffReadCost < HugeCost + && DstXprType::SizeAtCompileTime * (DstEvaluatorType::CoeffReadCost+SrcEvaluatorType::CoeffReadCost) / 2 <= EIGEN_UNROLLING_LIMIT + }; + + triangular_assignment_loop<Kernel, Mode, unroll ? int(DstXprType::SizeAtCompileTime) : Dynamic, SetOpposite>::run(kernel); +} + +template<int Mode, bool SetOpposite, typename DstXprType, typename SrcXprType> +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +void call_triangular_assignment_loop(DstXprType& dst, const SrcXprType& src) +{ + call_triangular_assignment_loop<Mode,SetOpposite>(dst, src, internal::assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar>()); +} + +template<> struct AssignmentKind<TriangularShape,TriangularShape> { typedef Triangular2Triangular Kind; }; +template<> struct AssignmentKind<DenseShape,TriangularShape> { typedef Triangular2Dense Kind; }; +template<> struct AssignmentKind<TriangularShape,DenseShape> { typedef Dense2Triangular Kind; }; + + +template< typename DstXprType, typename SrcXprType, typename Functor> +struct Assignment<DstXprType, SrcXprType, Functor, Triangular2Triangular> +{ + EIGEN_DEVICE_FUNC static void run(DstXprType &dst, const SrcXprType &src, const Functor &func) + { + eigen_assert(int(DstXprType::Mode) == int(SrcXprType::Mode)); + + call_triangular_assignment_loop<DstXprType::Mode, false>(dst, src, func); + } +}; + +template< typename DstXprType, typename SrcXprType, typename Functor> +struct Assignment<DstXprType, SrcXprType, Functor, Triangular2Dense> +{ + EIGEN_DEVICE_FUNC static void run(DstXprType &dst, const SrcXprType &src, const Functor &func) + { + call_triangular_assignment_loop<SrcXprType::Mode, (SrcXprType::Mode&SelfAdjoint)==0>(dst, src, func); + } +}; + +template< typename DstXprType, typename SrcXprType, typename Functor> +struct Assignment<DstXprType, SrcXprType, Functor, Dense2Triangular> +{ + EIGEN_DEVICE_FUNC static void run(DstXprType &dst, const SrcXprType &src, const Functor &func) + { + call_triangular_assignment_loop<DstXprType::Mode, false>(dst, src, func); + } +}; + + +template<typename Kernel, unsigned int Mode, int UnrollCount, bool SetOpposite> +struct triangular_assignment_loop +{ + // FIXME: this is not very clean, perhaps this information should be provided by the kernel? + typedef typename Kernel::DstEvaluatorType DstEvaluatorType; + typedef typename DstEvaluatorType::XprType DstXprType; + + enum { + col = (UnrollCount-1) / DstXprType::RowsAtCompileTime, + row = (UnrollCount-1) % DstXprType::RowsAtCompileTime + }; + + typedef typename Kernel::Scalar Scalar; + + EIGEN_DEVICE_FUNC + static inline void run(Kernel &kernel) + { + triangular_assignment_loop<Kernel, Mode, UnrollCount-1, SetOpposite>::run(kernel); + + if(row==col) + kernel.assignDiagonalCoeff(row); + else if( ((Mode&Lower) && row>col) || ((Mode&Upper) && row<col) ) + kernel.assignCoeff(row,col); + else if(SetOpposite) + kernel.assignOppositeCoeff(row,col); + } +}; + +// prevent buggy user code from causing an infinite recursion +template<typename Kernel, unsigned int Mode, bool SetOpposite> +struct triangular_assignment_loop<Kernel, Mode, 0, SetOpposite> +{ + EIGEN_DEVICE_FUNC + static inline void run(Kernel &) {} +}; + + + +// TODO: experiment with a recursive assignment procedure splitting the current +// triangular part into one rectangular and two triangular parts. + + +template<typename Kernel, unsigned int Mode, bool SetOpposite> +struct triangular_assignment_loop<Kernel, Mode, Dynamic, SetOpposite> +{ + typedef typename Kernel::Scalar Scalar; + EIGEN_DEVICE_FUNC + static inline void run(Kernel &kernel) + { + for(Index j = 0; j < kernel.cols(); ++j) + { + Index maxi = numext::mini(j, kernel.rows()); + Index i = 0; + if (((Mode&Lower) && SetOpposite) || (Mode&Upper)) + { + for(; i < maxi; ++i) + if(Mode&Upper) kernel.assignCoeff(i, j); + else kernel.assignOppositeCoeff(i, j); + } + else + i = maxi; + + if(i<kernel.rows()) // then i==j + kernel.assignDiagonalCoeff(i++); + + if (((Mode&Upper) && SetOpposite) || (Mode&Lower)) + { + for(; i < kernel.rows(); ++i) + if(Mode&Lower) kernel.assignCoeff(i, j); + else kernel.assignOppositeCoeff(i, j); + } + } + } +}; + +} // end namespace internal + +/** Assigns a triangular or selfadjoint matrix to a dense matrix. + * If the matrix is triangular, the opposite part is set to zero. */ +template<typename Derived> +template<typename DenseDerived> +void TriangularBase<Derived>::evalToLazy(MatrixBase<DenseDerived> &other) const +{ + other.derived().resize(this->rows(), this->cols()); + internal::call_triangular_assignment_loop<Derived::Mode,(Derived::Mode&SelfAdjoint)==0 /* SetOpposite */>(other.derived(), derived().nestedExpression()); +} + +namespace internal { + +// Triangular = Product +template< typename DstXprType, typename Lhs, typename Rhs, typename Scalar> +struct Assignment<DstXprType, Product<Lhs,Rhs,DefaultProduct>, internal::assign_op<Scalar,typename Product<Lhs,Rhs,DefaultProduct>::Scalar>, Dense2Triangular> +{ + typedef Product<Lhs,Rhs,DefaultProduct> SrcXprType; + static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,typename SrcXprType::Scalar> &) + { + Index dstRows = src.rows(); + Index dstCols = src.cols(); + if((dst.rows()!=dstRows) || (dst.cols()!=dstCols)) + dst.resize(dstRows, dstCols); + + dst._assignProduct(src, 1, 0); + } +}; + +// Triangular += Product +template< typename DstXprType, typename Lhs, typename Rhs, typename Scalar> +struct Assignment<DstXprType, Product<Lhs,Rhs,DefaultProduct>, internal::add_assign_op<Scalar,typename Product<Lhs,Rhs,DefaultProduct>::Scalar>, Dense2Triangular> +{ + typedef Product<Lhs,Rhs,DefaultProduct> SrcXprType; + static void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op<Scalar,typename SrcXprType::Scalar> &) + { + dst._assignProduct(src, 1, 1); + } +}; + +// Triangular -= Product +template< typename DstXprType, typename Lhs, typename Rhs, typename Scalar> +struct Assignment<DstXprType, Product<Lhs,Rhs,DefaultProduct>, internal::sub_assign_op<Scalar,typename Product<Lhs,Rhs,DefaultProduct>::Scalar>, Dense2Triangular> +{ + typedef Product<Lhs,Rhs,DefaultProduct> SrcXprType; + static void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op<Scalar,typename SrcXprType::Scalar> &) + { + dst._assignProduct(src, -1, 1); + } +}; + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_TRIANGULARMATRIX_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/VectorBlock.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/VectorBlock.h new file mode 100644 index 000000000..d72fbf7e9 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/VectorBlock.h @@ -0,0 +1,96 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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_VECTORBLOCK_H +#define EIGEN_VECTORBLOCK_H + +namespace Eigen { + +namespace internal { +template<typename VectorType, int Size> +struct traits<VectorBlock<VectorType, Size> > + : public traits<Block<VectorType, + traits<VectorType>::Flags & RowMajorBit ? 1 : Size, + traits<VectorType>::Flags & RowMajorBit ? Size : 1> > +{ +}; +} + +/** \class VectorBlock + * \ingroup Core_Module + * + * \brief Expression of a fixed-size or dynamic-size sub-vector + * + * \tparam VectorType the type of the object in which we are taking a sub-vector + * \tparam Size size of the sub-vector we are taking at compile time (optional) + * + * This class represents an expression of either a fixed-size or dynamic-size sub-vector. + * It is the return type of DenseBase::segment(Index,Index) and DenseBase::segment<int>(Index) and + * most of the time this is the only way it is used. + * + * However, if you want to directly maniputate sub-vector expressions, + * for instance if you want to write a function returning such an expression, you + * will need to use this class. + * + * Here is an example illustrating the dynamic case: + * \include class_VectorBlock.cpp + * Output: \verbinclude class_VectorBlock.out + * + * \note Even though this expression has dynamic size, in the case where \a VectorType + * has fixed size, this expression inherits a fixed maximal size which means that evaluating + * it does not cause a dynamic memory allocation. + * + * Here is an example illustrating the fixed-size case: + * \include class_FixedVectorBlock.cpp + * Output: \verbinclude class_FixedVectorBlock.out + * + * \sa class Block, DenseBase::segment(Index,Index,Index,Index), DenseBase::segment(Index,Index) + */ +template<typename VectorType, int Size> class VectorBlock + : public Block<VectorType, + internal::traits<VectorType>::Flags & RowMajorBit ? 1 : Size, + internal::traits<VectorType>::Flags & RowMajorBit ? Size : 1> +{ + typedef Block<VectorType, + internal::traits<VectorType>::Flags & RowMajorBit ? 1 : Size, + internal::traits<VectorType>::Flags & RowMajorBit ? Size : 1> Base; + enum { + IsColVector = !(internal::traits<VectorType>::Flags & RowMajorBit) + }; + public: + EIGEN_DENSE_PUBLIC_INTERFACE(VectorBlock) + + using Base::operator=; + + /** Dynamic-size constructor + */ + EIGEN_DEVICE_FUNC + inline VectorBlock(VectorType& vector, Index start, Index size) + : Base(vector, + IsColVector ? start : 0, IsColVector ? 0 : start, + IsColVector ? size : 1, IsColVector ? 1 : size) + { + EIGEN_STATIC_ASSERT_VECTOR_ONLY(VectorBlock); + } + + /** Fixed-size constructor + */ + EIGEN_DEVICE_FUNC + inline VectorBlock(VectorType& vector, Index start) + : Base(vector, IsColVector ? start : 0, IsColVector ? 0 : start) + { + EIGEN_STATIC_ASSERT_VECTOR_ONLY(VectorBlock); + } +}; + + +} // end namespace Eigen + +#endif // EIGEN_VECTORBLOCK_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/VectorwiseOp.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/VectorwiseOp.h new file mode 100644 index 000000000..4fe267e9f --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/VectorwiseOp.h @@ -0,0 +1,695 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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_PARTIAL_REDUX_H +#define EIGEN_PARTIAL_REDUX_H + +namespace Eigen { + +/** \class PartialReduxExpr + * \ingroup Core_Module + * + * \brief Generic expression of a partially reduxed matrix + * + * \tparam MatrixType the type of the matrix we are applying the redux operation + * \tparam MemberOp type of the member functor + * \tparam Direction indicates the direction of the redux (#Vertical or #Horizontal) + * + * This class represents an expression of a partial redux operator of a matrix. + * It is the return type of some VectorwiseOp functions, + * and most of the time this is the only way it is used. + * + * \sa class VectorwiseOp + */ + +template< typename MatrixType, typename MemberOp, int Direction> +class PartialReduxExpr; + +namespace internal { +template<typename MatrixType, typename MemberOp, int Direction> +struct traits<PartialReduxExpr<MatrixType, MemberOp, Direction> > + : traits<MatrixType> +{ + typedef typename MemberOp::result_type Scalar; + typedef typename traits<MatrixType>::StorageKind StorageKind; + typedef typename traits<MatrixType>::XprKind XprKind; + typedef typename MatrixType::Scalar InputScalar; + enum { + RowsAtCompileTime = Direction==Vertical ? 1 : MatrixType::RowsAtCompileTime, + ColsAtCompileTime = Direction==Horizontal ? 1 : MatrixType::ColsAtCompileTime, + MaxRowsAtCompileTime = Direction==Vertical ? 1 : MatrixType::MaxRowsAtCompileTime, + MaxColsAtCompileTime = Direction==Horizontal ? 1 : MatrixType::MaxColsAtCompileTime, + Flags = RowsAtCompileTime == 1 ? RowMajorBit : 0, + TraversalSize = Direction==Vertical ? MatrixType::RowsAtCompileTime : MatrixType::ColsAtCompileTime + }; +}; +} + +template< typename MatrixType, typename MemberOp, int Direction> +class PartialReduxExpr : public internal::dense_xpr_base< PartialReduxExpr<MatrixType, MemberOp, Direction> >::type, + internal::no_assignment_operator +{ + public: + + typedef typename internal::dense_xpr_base<PartialReduxExpr>::type Base; + EIGEN_DENSE_PUBLIC_INTERFACE(PartialReduxExpr) + + EIGEN_DEVICE_FUNC + explicit PartialReduxExpr(const MatrixType& mat, const MemberOp& func = MemberOp()) + : m_matrix(mat), m_functor(func) {} + + EIGEN_DEVICE_FUNC + Index rows() const { return (Direction==Vertical ? 1 : m_matrix.rows()); } + EIGEN_DEVICE_FUNC + Index cols() const { return (Direction==Horizontal ? 1 : m_matrix.cols()); } + + EIGEN_DEVICE_FUNC + typename MatrixType::Nested nestedExpression() const { return m_matrix; } + + EIGEN_DEVICE_FUNC + const MemberOp& functor() const { return m_functor; } + + protected: + typename MatrixType::Nested m_matrix; + const MemberOp m_functor; +}; + +#define EIGEN_MEMBER_FUNCTOR(MEMBER,COST) \ + template <typename ResultType> \ + struct member_##MEMBER { \ + EIGEN_EMPTY_STRUCT_CTOR(member_##MEMBER) \ + typedef ResultType result_type; \ + template<typename Scalar, int Size> struct Cost \ + { enum { value = COST }; }; \ + template<typename XprType> \ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE \ + ResultType operator()(const XprType& mat) const \ + { return mat.MEMBER(); } \ + } + +namespace internal { + +EIGEN_MEMBER_FUNCTOR(squaredNorm, Size * NumTraits<Scalar>::MulCost + (Size-1)*NumTraits<Scalar>::AddCost); +EIGEN_MEMBER_FUNCTOR(norm, (Size+5) * NumTraits<Scalar>::MulCost + (Size-1)*NumTraits<Scalar>::AddCost); +EIGEN_MEMBER_FUNCTOR(stableNorm, (Size+5) * NumTraits<Scalar>::MulCost + (Size-1)*NumTraits<Scalar>::AddCost); +EIGEN_MEMBER_FUNCTOR(blueNorm, (Size+5) * NumTraits<Scalar>::MulCost + (Size-1)*NumTraits<Scalar>::AddCost); +EIGEN_MEMBER_FUNCTOR(hypotNorm, (Size-1) * functor_traits<scalar_hypot_op<Scalar> >::Cost ); +EIGEN_MEMBER_FUNCTOR(sum, (Size-1)*NumTraits<Scalar>::AddCost); +EIGEN_MEMBER_FUNCTOR(mean, (Size-1)*NumTraits<Scalar>::AddCost + NumTraits<Scalar>::MulCost); +EIGEN_MEMBER_FUNCTOR(minCoeff, (Size-1)*NumTraits<Scalar>::AddCost); +EIGEN_MEMBER_FUNCTOR(maxCoeff, (Size-1)*NumTraits<Scalar>::AddCost); +EIGEN_MEMBER_FUNCTOR(all, (Size-1)*NumTraits<Scalar>::AddCost); +EIGEN_MEMBER_FUNCTOR(any, (Size-1)*NumTraits<Scalar>::AddCost); +EIGEN_MEMBER_FUNCTOR(count, (Size-1)*NumTraits<Scalar>::AddCost); +EIGEN_MEMBER_FUNCTOR(prod, (Size-1)*NumTraits<Scalar>::MulCost); + +template <int p, typename ResultType> +struct member_lpnorm { + typedef ResultType result_type; + template<typename Scalar, int Size> struct Cost + { enum { value = (Size+5) * NumTraits<Scalar>::MulCost + (Size-1)*NumTraits<Scalar>::AddCost }; }; + EIGEN_DEVICE_FUNC member_lpnorm() {} + template<typename XprType> + EIGEN_DEVICE_FUNC inline ResultType operator()(const XprType& mat) const + { return mat.template lpNorm<p>(); } +}; + +template <typename BinaryOp, typename Scalar> +struct member_redux { + typedef typename result_of< + BinaryOp(const Scalar&,const Scalar&) + >::type result_type; + template<typename _Scalar, int Size> struct Cost + { enum { value = (Size-1) * functor_traits<BinaryOp>::Cost }; }; + EIGEN_DEVICE_FUNC explicit member_redux(const BinaryOp func) : m_functor(func) {} + template<typename Derived> + EIGEN_DEVICE_FUNC inline result_type operator()(const DenseBase<Derived>& mat) const + { return mat.redux(m_functor); } + const BinaryOp m_functor; +}; +} + +/** \class VectorwiseOp + * \ingroup Core_Module + * + * \brief Pseudo expression providing partial reduction operations + * + * \tparam ExpressionType the type of the object on which to do partial reductions + * \tparam Direction indicates the direction of the redux (#Vertical or #Horizontal) + * + * This class represents a pseudo expression with partial reduction features. + * It is the return type of DenseBase::colwise() and DenseBase::rowwise() + * and most of the time this is the only way it is used. + * + * Example: \include MatrixBase_colwise.cpp + * Output: \verbinclude MatrixBase_colwise.out + * + * \sa DenseBase::colwise(), DenseBase::rowwise(), class PartialReduxExpr + */ +template<typename ExpressionType, int Direction> class VectorwiseOp +{ + public: + + typedef typename ExpressionType::Scalar Scalar; + typedef typename ExpressionType::RealScalar RealScalar; + typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3 + typedef typename internal::ref_selector<ExpressionType>::non_const_type ExpressionTypeNested; + typedef typename internal::remove_all<ExpressionTypeNested>::type ExpressionTypeNestedCleaned; + + template<template<typename _Scalar> class Functor, + typename Scalar_=Scalar> struct ReturnType + { + typedef PartialReduxExpr<ExpressionType, + Functor<Scalar_>, + Direction + > Type; + }; + + template<typename BinaryOp> struct ReduxReturnType + { + typedef PartialReduxExpr<ExpressionType, + internal::member_redux<BinaryOp,Scalar>, + Direction + > Type; + }; + + enum { + isVertical = (Direction==Vertical) ? 1 : 0, + isHorizontal = (Direction==Horizontal) ? 1 : 0 + }; + + protected: + + typedef typename internal::conditional<isVertical, + typename ExpressionType::ColXpr, + typename ExpressionType::RowXpr>::type SubVector; + /** \internal + * \returns the i-th subvector according to the \c Direction */ + EIGEN_DEVICE_FUNC + SubVector subVector(Index i) + { + return SubVector(m_matrix.derived(),i); + } + + /** \internal + * \returns the number of subvectors in the direction \c Direction */ + EIGEN_DEVICE_FUNC + Index subVectors() const + { return isVertical?m_matrix.cols():m_matrix.rows(); } + + template<typename OtherDerived> struct ExtendedType { + typedef Replicate<OtherDerived, + isVertical ? 1 : ExpressionType::RowsAtCompileTime, + isHorizontal ? 1 : ExpressionType::ColsAtCompileTime> Type; + }; + + /** \internal + * Replicates a vector to match the size of \c *this */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + typename ExtendedType<OtherDerived>::Type + extendedTo(const DenseBase<OtherDerived>& other) const + { + EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(isVertical, OtherDerived::MaxColsAtCompileTime==1), + YOU_PASSED_A_ROW_VECTOR_BUT_A_COLUMN_VECTOR_WAS_EXPECTED) + EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(isHorizontal, OtherDerived::MaxRowsAtCompileTime==1), + YOU_PASSED_A_COLUMN_VECTOR_BUT_A_ROW_VECTOR_WAS_EXPECTED) + return typename ExtendedType<OtherDerived>::Type + (other.derived(), + isVertical ? 1 : m_matrix.rows(), + isHorizontal ? 1 : m_matrix.cols()); + } + + template<typename OtherDerived> struct OppositeExtendedType { + typedef Replicate<OtherDerived, + isHorizontal ? 1 : ExpressionType::RowsAtCompileTime, + isVertical ? 1 : ExpressionType::ColsAtCompileTime> Type; + }; + + /** \internal + * Replicates a vector in the opposite direction to match the size of \c *this */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + typename OppositeExtendedType<OtherDerived>::Type + extendedToOpposite(const DenseBase<OtherDerived>& other) const + { + EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(isHorizontal, OtherDerived::MaxColsAtCompileTime==1), + YOU_PASSED_A_ROW_VECTOR_BUT_A_COLUMN_VECTOR_WAS_EXPECTED) + EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(isVertical, OtherDerived::MaxRowsAtCompileTime==1), + YOU_PASSED_A_COLUMN_VECTOR_BUT_A_ROW_VECTOR_WAS_EXPECTED) + return typename OppositeExtendedType<OtherDerived>::Type + (other.derived(), + isHorizontal ? 1 : m_matrix.rows(), + isVertical ? 1 : m_matrix.cols()); + } + + public: + EIGEN_DEVICE_FUNC + explicit inline VectorwiseOp(ExpressionType& matrix) : m_matrix(matrix) {} + + /** \internal */ + EIGEN_DEVICE_FUNC + inline const ExpressionType& _expression() const { return m_matrix; } + + /** \returns a row or column vector expression of \c *this reduxed by \a func + * + * The template parameter \a BinaryOp is the type of the functor + * of the custom redux operator. Note that func must be an associative operator. + * + * \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)); } + + typedef typename ReturnType<internal::member_minCoeff>::Type MinCoeffReturnType; + typedef typename ReturnType<internal::member_maxCoeff>::Type MaxCoeffReturnType; + typedef typename ReturnType<internal::member_squaredNorm,RealScalar>::Type SquaredNormReturnType; + typedef typename ReturnType<internal::member_norm,RealScalar>::Type NormReturnType; + typedef typename ReturnType<internal::member_blueNorm,RealScalar>::Type BlueNormReturnType; + typedef typename ReturnType<internal::member_stableNorm,RealScalar>::Type StableNormReturnType; + typedef typename ReturnType<internal::member_hypotNorm,RealScalar>::Type HypotNormReturnType; + typedef typename ReturnType<internal::member_sum>::Type SumReturnType; + typedef typename ReturnType<internal::member_mean>::Type MeanReturnType; + typedef typename ReturnType<internal::member_all>::Type AllReturnType; + typedef typename ReturnType<internal::member_any>::Type AnyReturnType; + typedef PartialReduxExpr<ExpressionType, internal::member_count<Index>, Direction> CountReturnType; + typedef typename ReturnType<internal::member_prod>::Type ProdReturnType; + typedef Reverse<const ExpressionType, Direction> ConstReverseReturnType; + typedef Reverse<ExpressionType, Direction> ReverseReturnType; + + template<int p> struct LpNormReturnType { + typedef PartialReduxExpr<ExpressionType, internal::member_lpnorm<p,RealScalar>,Direction> Type; + }; + + /** \returns a row (or column) vector expression of the smallest coefficient + * of each column (or row) of the referenced expression. + * + * \warning the result is undefined if \c *this contains NaN. + * + * Example: \include PartialRedux_minCoeff.cpp + * Output: \verbinclude PartialRedux_minCoeff.out + * + * \sa DenseBase::minCoeff() */ + EIGEN_DEVICE_FUNC + const MinCoeffReturnType minCoeff() const + { 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 result is undefined if \c *this contains NaN. + * + * Example: \include PartialRedux_maxCoeff.cpp + * Output: \verbinclude PartialRedux_maxCoeff.out + * + * \sa DenseBase::maxCoeff() */ + EIGEN_DEVICE_FUNC + const MaxCoeffReturnType maxCoeff() const + { return MaxCoeffReturnType(_expression()); } + + /** \returns a row (or column) vector expression of the squared norm + * of each column (or row) of the referenced expression. + * This is a vector with real entries, even if the original matrix has complex entries. + * + * Example: \include PartialRedux_squaredNorm.cpp + * Output: \verbinclude PartialRedux_squaredNorm.out + * + * \sa DenseBase::squaredNorm() */ + EIGEN_DEVICE_FUNC + const SquaredNormReturnType squaredNorm() const + { return SquaredNormReturnType(_expression()); } + + /** \returns a row (or column) vector expression of the norm + * of each column (or row) of the referenced expression. + * This is a vector with real entries, even if the original matrix has complex entries. + * + * Example: \include PartialRedux_norm.cpp + * Output: \verbinclude PartialRedux_norm.out + * + * \sa DenseBase::norm() */ + EIGEN_DEVICE_FUNC + const NormReturnType norm() const + { return NormReturnType(_expression()); } + + /** \returns a row (or column) vector expression of the norm + * of each column (or row) of the referenced expression. + * This is a vector with real entries, even if the original matrix has complex entries. + * + * Example: \include PartialRedux_norm.cpp + * Output: \verbinclude PartialRedux_norm.out + * + * \sa DenseBase::norm() */ + template<int p> + EIGEN_DEVICE_FUNC + const typename LpNormReturnType<p>::Type lpNorm() const + { return typename LpNormReturnType<p>::Type(_expression()); } + + + /** \returns a row (or column) vector expression of the norm + * of each column (or row) of the referenced expression, using + * Blue's algorithm. + * This is a vector with real entries, even if the original matrix has complex entries. + * + * \sa DenseBase::blueNorm() */ + EIGEN_DEVICE_FUNC + const BlueNormReturnType blueNorm() const + { return BlueNormReturnType(_expression()); } + + + /** \returns a row (or column) vector expression of the norm + * of each column (or row) of the referenced expression, avoiding + * underflow and overflow. + * This is a vector with real entries, even if the original matrix has complex entries. + * + * \sa DenseBase::stableNorm() */ + EIGEN_DEVICE_FUNC + const StableNormReturnType stableNorm() const + { return StableNormReturnType(_expression()); } + + + /** \returns a row (or column) vector expression of the norm + * of each column (or row) of the referenced expression, avoiding + * underflow and overflow using a concatenation of hypot() calls. + * This is a vector with real entries, even if the original matrix has complex entries. + * + * \sa DenseBase::hypotNorm() */ + EIGEN_DEVICE_FUNC + const HypotNormReturnType hypotNorm() const + { return HypotNormReturnType(_expression()); } + + /** \returns a row (or column) vector expression of the sum + * of each column (or row) of the referenced expression. + * + * Example: \include PartialRedux_sum.cpp + * Output: \verbinclude PartialRedux_sum.out + * + * \sa DenseBase::sum() */ + EIGEN_DEVICE_FUNC + const SumReturnType sum() const + { return SumReturnType(_expression()); } + + /** \returns a row (or column) vector expression of the mean + * of each column (or row) of the referenced expression. + * + * \sa DenseBase::mean() */ + EIGEN_DEVICE_FUNC + const MeanReturnType mean() const + { return MeanReturnType(_expression()); } + + /** \returns a row (or column) vector expression representing + * whether \b all coefficients of each respective column (or row) are \c true. + * This expression can be assigned to a vector with entries of type \c bool. + * + * \sa DenseBase::all() */ + EIGEN_DEVICE_FUNC + const AllReturnType all() const + { return AllReturnType(_expression()); } + + /** \returns a row (or column) vector expression representing + * whether \b at \b least one coefficient of each respective column (or row) is \c true. + * This expression can be assigned to a vector with entries of type \c bool. + * + * \sa DenseBase::any() */ + EIGEN_DEVICE_FUNC + const AnyReturnType any() const + { return AnyReturnType(_expression()); } + + /** \returns a row (or column) vector expression representing + * the number of \c true coefficients of each respective column (or row). + * This expression can be assigned to a vector whose entries have the same type as is used to + * index entries of the original matrix; for dense matrices, this is \c std::ptrdiff_t . + * + * Example: \include PartialRedux_count.cpp + * Output: \verbinclude PartialRedux_count.out + * + * \sa DenseBase::count() */ + EIGEN_DEVICE_FUNC + const CountReturnType count() const + { return CountReturnType(_expression()); } + + /** \returns a row (or column) vector expression of the product + * of each column (or row) of the referenced expression. + * + * Example: \include PartialRedux_prod.cpp + * Output: \verbinclude PartialRedux_prod.out + * + * \sa DenseBase::prod() */ + EIGEN_DEVICE_FUNC + const ProdReturnType prod() const + { return ProdReturnType(_expression()); } + + + /** \returns a matrix expression + * where each column (or row) are reversed. + * + * Example: \include Vectorwise_reverse.cpp + * Output: \verbinclude Vectorwise_reverse.out + * + * \sa DenseBase::reverse() */ + EIGEN_DEVICE_FUNC + const ConstReverseReturnType reverse() const + { return ConstReverseReturnType( _expression() ); } + + /** \returns a writable matrix expression + * where each column (or row) are reversed. + * + * \sa reverse() const */ + EIGEN_DEVICE_FUNC + ReverseReturnType reverse() + { return ReverseReturnType( _expression() ); } + + typedef Replicate<ExpressionType,(isVertical?Dynamic:1),(isHorizontal?Dynamic:1)> ReplicateReturnType; + EIGEN_DEVICE_FUNC + const ReplicateReturnType replicate(Index factor) const; + + /** + * \return an expression of the replication of each column (or row) of \c *this + * + * Example: \include DirectionWise_replicate.cpp + * Output: \verbinclude DirectionWise_replicate.out + * + * \sa VectorwiseOp::replicate(Index), DenseBase::replicate(), class Replicate + */ + // NOTE implemented here because of sunstudio's compilation errors + // isVertical*Factor+isHorizontal instead of (isVertical?Factor:1) to handle CUDA bug with ternary operator + template<int Factor> const Replicate<ExpressionType,isVertical*Factor+isHorizontal,isHorizontal*Factor+isVertical> + EIGEN_DEVICE_FUNC + replicate(Index factor = Factor) const + { + return Replicate<ExpressionType,(isVertical?Factor:1),(isHorizontal?Factor:1)> + (_expression(),isVertical?factor:1,isHorizontal?factor:1); + } + +/////////// Artithmetic operators /////////// + + /** Copies the vector \a other to each subvector of \c *this */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + ExpressionType& operator=(const DenseBase<OtherDerived>& other) + { + 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())); + } + + /** Adds the vector \a other to each subvector of \c *this */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + ExpressionType& operator+=(const DenseBase<OtherDerived>& other) + { + EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived) + EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived) + return const_cast<ExpressionType&>(m_matrix += extendedTo(other.derived())); + } + + /** Substracts the vector \a other to each subvector of \c *this */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + ExpressionType& operator-=(const DenseBase<OtherDerived>& other) + { + EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived) + EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived) + return const_cast<ExpressionType&>(m_matrix -= extendedTo(other.derived())); + } + + /** Multiples each subvector of \c *this by the vector \a other */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + ExpressionType& operator*=(const DenseBase<OtherDerived>& other) + { + EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived) + EIGEN_STATIC_ASSERT_ARRAYXPR(ExpressionType) + EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived) + m_matrix *= extendedTo(other.derived()); + return const_cast<ExpressionType&>(m_matrix); + } + + /** Divides each subvector of \c *this by the vector \a other */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + ExpressionType& operator/=(const DenseBase<OtherDerived>& other) + { + EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived) + EIGEN_STATIC_ASSERT_ARRAYXPR(ExpressionType) + EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived) + m_matrix /= extendedTo(other.derived()); + return const_cast<ExpressionType&>(m_matrix); + } + + /** Returns the expression of the sum of the vector \a other to each subvector of \c *this */ + template<typename OtherDerived> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC + CwiseBinaryOp<internal::scalar_sum_op<Scalar,typename OtherDerived::Scalar>, + const ExpressionTypeNestedCleaned, + const typename ExtendedType<OtherDerived>::Type> + operator+(const DenseBase<OtherDerived>& other) const + { + EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived) + EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived) + return m_matrix + extendedTo(other.derived()); + } + + /** Returns the expression of the difference between each subvector of \c *this and the vector \a other */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + CwiseBinaryOp<internal::scalar_difference_op<Scalar,typename OtherDerived::Scalar>, + const ExpressionTypeNestedCleaned, + const typename ExtendedType<OtherDerived>::Type> + operator-(const DenseBase<OtherDerived>& other) const + { + EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived) + EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived) + return m_matrix - extendedTo(other.derived()); + } + + /** Returns the expression where each subvector is the product of the vector \a other + * by the corresponding subvector of \c *this */ + template<typename OtherDerived> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC + CwiseBinaryOp<internal::scalar_product_op<Scalar>, + const ExpressionTypeNestedCleaned, + const typename ExtendedType<OtherDerived>::Type> + EIGEN_DEVICE_FUNC + operator*(const DenseBase<OtherDerived>& other) const + { + EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived) + EIGEN_STATIC_ASSERT_ARRAYXPR(ExpressionType) + EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived) + return m_matrix * extendedTo(other.derived()); + } + + /** Returns the expression where each subvector is the quotient of the corresponding + * subvector of \c *this by the vector \a other */ + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + CwiseBinaryOp<internal::scalar_quotient_op<Scalar>, + const ExpressionTypeNestedCleaned, + const typename ExtendedType<OtherDerived>::Type> + operator/(const DenseBase<OtherDerived>& other) const + { + EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived) + EIGEN_STATIC_ASSERT_ARRAYXPR(ExpressionType) + EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived) + return m_matrix / extendedTo(other.derived()); + } + + /** \returns an expression where each column (or row) of the referenced matrix are normalized. + * The referenced matrix is \b not modified. + * \sa MatrixBase::normalized(), normalize() + */ + EIGEN_DEVICE_FUNC + CwiseBinaryOp<internal::scalar_quotient_op<Scalar>, + const ExpressionTypeNestedCleaned, + const typename OppositeExtendedType<typename ReturnType<internal::member_norm,RealScalar>::Type>::Type> + normalized() const { return m_matrix.cwiseQuotient(extendedToOpposite(this->norm())); } + + + /** Normalize in-place each row or columns of the referenced matrix. + * \sa MatrixBase::normalize(), normalized() + */ + EIGEN_DEVICE_FUNC void normalize() { + m_matrix = this->normalized(); + } + + EIGEN_DEVICE_FUNC inline void reverseInPlace(); + +/////////// Geometry module /////////// + + typedef Homogeneous<ExpressionType,Direction> HomogeneousReturnType; + EIGEN_DEVICE_FUNC + HomogeneousReturnType homogeneous() const; + + typedef typename ExpressionType::PlainObject CrossReturnType; + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + const CrossReturnType cross(const MatrixBase<OtherDerived>& other) const; + + enum { + HNormalized_Size = Direction==Vertical ? internal::traits<ExpressionType>::RowsAtCompileTime + : internal::traits<ExpressionType>::ColsAtCompileTime, + HNormalized_SizeMinusOne = HNormalized_Size==Dynamic ? Dynamic : HNormalized_Size-1 + }; + typedef Block<const ExpressionType, + Direction==Vertical ? int(HNormalized_SizeMinusOne) + : int(internal::traits<ExpressionType>::RowsAtCompileTime), + Direction==Horizontal ? int(HNormalized_SizeMinusOne) + : int(internal::traits<ExpressionType>::ColsAtCompileTime)> + HNormalized_Block; + typedef Block<const ExpressionType, + Direction==Vertical ? 1 : int(internal::traits<ExpressionType>::RowsAtCompileTime), + Direction==Horizontal ? 1 : int(internal::traits<ExpressionType>::ColsAtCompileTime)> + HNormalized_Factors; + typedef CwiseBinaryOp<internal::scalar_quotient_op<typename internal::traits<ExpressionType>::Scalar>, + const HNormalized_Block, + const Replicate<HNormalized_Factors, + Direction==Vertical ? HNormalized_SizeMinusOne : 1, + Direction==Horizontal ? HNormalized_SizeMinusOne : 1> > + HNormalizedReturnType; + + EIGEN_DEVICE_FUNC + const HNormalizedReturnType hnormalized() const; + + protected: + ExpressionTypeNested m_matrix; +}; + +//const colwise moved to DenseBase.h due to CUDA compiler bug + + +/** \returns a writable VectorwiseOp wrapper of *this providing additional partial reduction operations + * + * \sa rowwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting + */ +template<typename Derived> +inline typename DenseBase<Derived>::ColwiseReturnType +DenseBase<Derived>::colwise() +{ + return ColwiseReturnType(derived()); +} + +//const rowwise moved to DenseBase.h due to CUDA compiler bug + + +/** \returns a writable VectorwiseOp wrapper of *this providing additional partial reduction operations + * + * \sa colwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting + */ +template<typename Derived> +inline typename DenseBase<Derived>::RowwiseReturnType +DenseBase<Derived>::rowwise() +{ + return RowwiseReturnType(derived()); +} + +} // end namespace Eigen + +#endif // EIGEN_PARTIAL_REDUX_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/Visitor.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Visitor.h new file mode 100644 index 000000000..54c1883d9 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/Visitor.h @@ -0,0 +1,273 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008 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_VISITOR_H +#define EIGEN_VISITOR_H + +namespace Eigen { + +namespace internal { + +template<typename Visitor, typename Derived, int UnrollCount> +struct visitor_impl +{ + enum { + col = (UnrollCount-1) / Derived::RowsAtCompileTime, + row = (UnrollCount-1) % Derived::RowsAtCompileTime + }; + + EIGEN_DEVICE_FUNC + static inline void run(const Derived &mat, Visitor& visitor) + { + visitor_impl<Visitor, Derived, UnrollCount-1>::run(mat, visitor); + visitor(mat.coeff(row, col), row, col); + } +}; + +template<typename Visitor, typename Derived> +struct visitor_impl<Visitor, Derived, 1> +{ + EIGEN_DEVICE_FUNC + static inline void run(const Derived &mat, Visitor& visitor) + { + return visitor.init(mat.coeff(0, 0), 0, 0); + } +}; + +template<typename Visitor, typename Derived> +struct visitor_impl<Visitor, Derived, Dynamic> +{ + EIGEN_DEVICE_FUNC + static inline void run(const Derived& mat, Visitor& visitor) + { + visitor.init(mat.coeff(0,0), 0, 0); + for(Index i = 1; i < mat.rows(); ++i) + visitor(mat.coeff(i, 0), i, 0); + for(Index j = 1; j < mat.cols(); ++j) + for(Index i = 0; i < mat.rows(); ++i) + visitor(mat.coeff(i, j), i, j); + } +}; + +// evaluator adaptor +template<typename XprType> +class visitor_evaluator +{ +public: + EIGEN_DEVICE_FUNC + explicit visitor_evaluator(const XprType &xpr) : m_evaluator(xpr), m_xpr(xpr) {} + + typedef typename XprType::Scalar Scalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + + enum { + RowsAtCompileTime = XprType::RowsAtCompileTime, + CoeffReadCost = internal::evaluator<XprType>::CoeffReadCost + }; + + EIGEN_DEVICE_FUNC Index rows() const { return m_xpr.rows(); } + EIGEN_DEVICE_FUNC Index cols() const { return m_xpr.cols(); } + EIGEN_DEVICE_FUNC Index size() const { return m_xpr.size(); } + + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index col) const + { return m_evaluator.coeff(row, col); } + +protected: + internal::evaluator<XprType> m_evaluator; + const XprType &m_xpr; +}; +} // end namespace internal + +/** Applies the visitor \a visitor to the whole coefficients of the matrix or vector. + * + * The template parameter \a Visitor is the type of the visitor and provides the following interface: + * \code + * struct MyVisitor { + * // called for the first coefficient + * void init(const Scalar& value, Index i, Index j); + * // called for all other coefficients + * void operator() (const Scalar& value, Index i, Index j); + * }; + * \endcode + * + * \note compared to one or two \em for \em loops, visitors offer automatic + * unrolling for small fixed size matrix. + * + * \sa minCoeff(Index*,Index*), maxCoeff(Index*,Index*), DenseBase::redux() + */ +template<typename Derived> +template<typename Visitor> +EIGEN_DEVICE_FUNC +void DenseBase<Derived>::visit(Visitor& visitor) const +{ + typedef typename internal::visitor_evaluator<Derived> ThisEvaluator; + ThisEvaluator thisEval(derived()); + + enum { + unroll = SizeAtCompileTime != Dynamic + && SizeAtCompileTime * ThisEvaluator::CoeffReadCost + (SizeAtCompileTime-1) * internal::functor_traits<Visitor>::Cost <= EIGEN_UNROLLING_LIMIT + }; + return internal::visitor_impl<Visitor, ThisEvaluator, unroll ? int(SizeAtCompileTime) : Dynamic>::run(thisEval, visitor); +} + +namespace internal { + +/** \internal + * \brief Base class to implement min and max visitors + */ +template <typename Derived> +struct coeff_visitor +{ + typedef typename Derived::Scalar Scalar; + Index row, col; + Scalar res; + EIGEN_DEVICE_FUNC + inline void init(const Scalar& value, Index i, Index j) + { + res = value; + row = i; + col = j; + } +}; + +/** \internal + * \brief Visitor computing the min coefficient with its value and coordinates + * + * \sa DenseBase::minCoeff(Index*, Index*) + */ +template <typename Derived> +struct min_coeff_visitor : coeff_visitor<Derived> +{ + typedef typename Derived::Scalar Scalar; + EIGEN_DEVICE_FUNC + void operator() (const Scalar& value, Index i, Index j) + { + if(value < this->res) + { + this->res = value; + this->row = i; + this->col = j; + } + } +}; + +template<typename Scalar> +struct functor_traits<min_coeff_visitor<Scalar> > { + enum { + Cost = NumTraits<Scalar>::AddCost + }; +}; + +/** \internal + * \brief Visitor computing the max coefficient with its value and coordinates + * + * \sa DenseBase::maxCoeff(Index*, Index*) + */ +template <typename Derived> +struct max_coeff_visitor : coeff_visitor<Derived> +{ + typedef typename Derived::Scalar Scalar; + EIGEN_DEVICE_FUNC + void operator() (const Scalar& value, Index i, Index j) + { + if(value > this->res) + { + this->res = value; + this->row = i; + this->col = j; + } + } +}; + +template<typename Scalar> +struct functor_traits<max_coeff_visitor<Scalar> > { + enum { + Cost = NumTraits<Scalar>::AddCost + }; +}; + +} // end namespace internal + +/** \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 result is undefined if \c *this contains NaN. + * + * \sa DenseBase::minCoeff(Index*), DenseBase::maxCoeff(Index*,Index*), DenseBase::visit(), DenseBase::minCoeff() + */ +template<typename Derived> +template<typename IndexType> +EIGEN_DEVICE_FUNC +typename internal::traits<Derived>::Scalar +DenseBase<Derived>::minCoeff(IndexType* rowId, IndexType* colId) const +{ + internal::min_coeff_visitor<Derived> minVisitor; + this->visit(minVisitor); + *rowId = minVisitor.row; + if (colId) *colId = minVisitor.col; + return minVisitor.res; +} + +/** \returns the minimum of all coefficients of *this and puts in *index its location. + * \warning the result is undefined if \c *this contains NaN. + * + * \sa DenseBase::minCoeff(IndexType*,IndexType*), DenseBase::maxCoeff(IndexType*,IndexType*), DenseBase::visit(), DenseBase::minCoeff() + */ +template<typename Derived> +template<typename IndexType> +EIGEN_DEVICE_FUNC +typename internal::traits<Derived>::Scalar +DenseBase<Derived>::minCoeff(IndexType* index) const +{ + EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) + internal::min_coeff_visitor<Derived> minVisitor; + this->visit(minVisitor); + *index = IndexType((RowsAtCompileTime==1) ? minVisitor.col : minVisitor.row); + return minVisitor.res; +} + +/** \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 result is undefined if \c *this contains NaN. + * + * \sa DenseBase::minCoeff(IndexType*,IndexType*), DenseBase::visit(), DenseBase::maxCoeff() + */ +template<typename Derived> +template<typename IndexType> +EIGEN_DEVICE_FUNC +typename internal::traits<Derived>::Scalar +DenseBase<Derived>::maxCoeff(IndexType* rowPtr, IndexType* colPtr) const +{ + internal::max_coeff_visitor<Derived> maxVisitor; + this->visit(maxVisitor); + *rowPtr = maxVisitor.row; + if (colPtr) *colPtr = maxVisitor.col; + return maxVisitor.res; +} + +/** \returns the maximum of all coefficients of *this and puts in *index its location. + * \warning the result is undefined if \c *this contains NaN. + * + * \sa DenseBase::maxCoeff(IndexType*,IndexType*), DenseBase::minCoeff(IndexType*,IndexType*), DenseBase::visitor(), DenseBase::maxCoeff() + */ +template<typename Derived> +template<typename IndexType> +EIGEN_DEVICE_FUNC +typename internal::traits<Derived>::Scalar +DenseBase<Derived>::maxCoeff(IndexType* index) const +{ + EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) + internal::max_coeff_visitor<Derived> maxVisitor; + this->visit(maxVisitor); + *index = (RowsAtCompileTime==1) ? maxVisitor.col : maxVisitor.row; + return maxVisitor.res; +} + +} // end namespace Eigen + +#endif // EIGEN_VISITOR_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AVX/Complex.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AVX/Complex.h new file mode 100644 index 000000000..99439c8aa --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AVX/Complex.h @@ -0,0 +1,483 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner (benoit.steiner.goog@gmail.com) +// +// 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_AVX_H +#define EIGEN_COMPLEX_AVX_H + +namespace Eigen { + +namespace internal { + +//---------- float ---------- +struct Packet4cf +{ + EIGEN_STRONG_INLINE Packet4cf() {} + EIGEN_STRONG_INLINE explicit Packet4cf(const __m256& a) : v(a) {} + __m256 v; +}; + +template<> struct packet_traits<std::complex<float> > : default_packet_traits +{ + typedef Packet4cf type; + typedef Packet2cf half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size = 4, + HasHalfPacket = 1, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasNegate = 1, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasSetLinear = 0 + }; +}; + +template<> struct unpacket_traits<Packet4cf> { typedef std::complex<float> type; enum {size=4, alignment=Aligned32}; 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)); } +template<> EIGEN_STRONG_INLINE Packet4cf pnegate(const Packet4cf& a) +{ + return Packet4cf(pnegate(a.v)); +} +template<> EIGEN_STRONG_INLINE Packet4cf pconj(const Packet4cf& a) +{ + const __m256 mask = _mm256_castsi256_ps(_mm256_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000,0x00000000,0x80000000,0x00000000,0x80000000)); + return Packet4cf(_mm256_xor_ps(a.v,mask)); +} + +template<> EIGEN_STRONG_INLINE Packet4cf pmul<Packet4cf>(const Packet4cf& a, const Packet4cf& b) +{ + __m256 tmp1 = _mm256_mul_ps(_mm256_moveldup_ps(a.v), b.v); + __m256 tmp2 = _mm256_mul_ps(_mm256_movehdup_ps(a.v), _mm256_permute_ps(b.v, _MM_SHUFFLE(2,3,0,1))); + __m256 result = _mm256_addsub_ps(tmp1, tmp2); + return Packet4cf(result); +} + +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 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))); } + + +template<> EIGEN_STRONG_INLINE Packet4cf pset1<Packet4cf>(const std::complex<float>& from) +{ + return Packet4cf(_mm256_castpd_ps(_mm256_broadcast_sd((const double*)(const void*)&from))); +} + +template<> EIGEN_STRONG_INLINE Packet4cf ploaddup<Packet4cf>(const std::complex<float>* from) +{ + // FIXME The following might be optimized using _mm256_movedup_pd + Packet2cf a = ploaddup<Packet2cf>(from); + Packet2cf b = ploaddup<Packet2cf>(from+1); + return Packet4cf(_mm256_insertf128_ps(_mm256_castps128_ps256(a.v), b.v, 1)); +} + +template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float>* to, const Packet4cf& 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 Packet4cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&numext::real_ref(*to), from.v); } + +template<> EIGEN_DEVICE_FUNC inline Packet4cf pgather<std::complex<float>, Packet4cf>(const std::complex<float>* from, Index stride) +{ + return Packet4cf(_mm256_set_ps(std::imag(from[3*stride]), std::real(from[3*stride]), + std::imag(from[2*stride]), std::real(from[2*stride]), + std::imag(from[1*stride]), std::real(from[1*stride]), + std::imag(from[0*stride]), std::real(from[0*stride]))); +} + +template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet4cf>(std::complex<float>* to, const Packet4cf& from, Index stride) +{ + __m128 low = _mm256_extractf128_ps(from.v, 0); + to[stride*0] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(low, low, 0)), + _mm_cvtss_f32(_mm_shuffle_ps(low, low, 1))); + to[stride*1] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(low, low, 2)), + _mm_cvtss_f32(_mm_shuffle_ps(low, low, 3))); + + __m128 high = _mm256_extractf128_ps(from.v, 1); + to[stride*2] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(high, high, 0)), + _mm_cvtss_f32(_mm_shuffle_ps(high, high, 1))); + to[stride*3] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(high, high, 2)), + _mm_cvtss_f32(_mm_shuffle_ps(high, high, 3))); + +} + +template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet4cf>(const Packet4cf& a) +{ + return pfirst(Packet2cf(_mm256_castps256_ps128(a.v))); +} + +template<> EIGEN_STRONG_INLINE Packet4cf preverse(const Packet4cf& a) { + __m128 low = _mm256_extractf128_ps(a.v, 0); + __m128 high = _mm256_extractf128_ps(a.v, 1); + __m128d lowd = _mm_castps_pd(low); + __m128d highd = _mm_castps_pd(high); + low = _mm_castpd_ps(_mm_shuffle_pd(lowd,lowd,0x1)); + high = _mm_castpd_ps(_mm_shuffle_pd(highd,highd,0x1)); + __m256 result = _mm256_setzero_ps(); + result = _mm256_insertf128_ps(result, low, 1); + result = _mm256_insertf128_ps(result, high, 0); + return Packet4cf(result); +} + +template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet4cf>(const Packet4cf& a) +{ + return predux(padd(Packet2cf(_mm256_extractf128_ps(a.v,0)), + Packet2cf(_mm256_extractf128_ps(a.v,1)))); +} + +template<> EIGEN_STRONG_INLINE Packet4cf preduxp<Packet4cf>(const Packet4cf* vecs) +{ + Packet8f t0 = _mm256_shuffle_ps(vecs[0].v, vecs[0].v, _MM_SHUFFLE(3, 1, 2 ,0)); + Packet8f t1 = _mm256_shuffle_ps(vecs[1].v, vecs[1].v, _MM_SHUFFLE(3, 1, 2 ,0)); + t0 = _mm256_hadd_ps(t0,t1); + Packet8f t2 = _mm256_shuffle_ps(vecs[2].v, vecs[2].v, _MM_SHUFFLE(3, 1, 2 ,0)); + Packet8f t3 = _mm256_shuffle_ps(vecs[3].v, vecs[3].v, _MM_SHUFFLE(3, 1, 2 ,0)); + t2 = _mm256_hadd_ps(t2,t3); + + t1 = _mm256_permute2f128_ps(t0,t2, 0 + (2<<4)); + t3 = _mm256_permute2f128_ps(t0,t2, 1 + (3<<4)); + + return Packet4cf(_mm256_add_ps(t1,t3)); +} + +template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet4cf>(const Packet4cf& a) +{ + return predux_mul(pmul(Packet2cf(_mm256_extractf128_ps(a.v, 0)), + Packet2cf(_mm256_extractf128_ps(a.v, 1)))); +} + +template<int Offset> +struct palign_impl<Offset,Packet4cf> +{ + static EIGEN_STRONG_INLINE void run(Packet4cf& first, const Packet4cf& second) + { + if (Offset==0) return; + palign_impl<Offset*2,Packet8f>::run(first.v, second.v); + } +}; + +template<> struct conj_helper<Packet4cf, Packet4cf, false,true> +{ + EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet4cf& x, const Packet4cf& y, const Packet4cf& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet4cf pmul(const Packet4cf& a, const Packet4cf& b) const + { + return internal::pmul(a, pconj(b)); + } +}; + +template<> struct conj_helper<Packet4cf, Packet4cf, true,false> +{ + EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet4cf& x, const Packet4cf& y, const Packet4cf& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet4cf pmul(const Packet4cf& a, const Packet4cf& b) const + { + return internal::pmul(pconj(a), b); + } +}; + +template<> struct conj_helper<Packet4cf, Packet4cf, true,true> +{ + EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet4cf& x, const Packet4cf& y, const Packet4cf& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet4cf pmul(const Packet4cf& a, const Packet4cf& b) const + { + return pconj(internal::pmul(a, b)); + } +}; + +template<> struct conj_helper<Packet8f, Packet4cf, false,false> +{ + EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet8f& x, const Packet4cf& y, const Packet4cf& c) const + { return padd(c, pmul(x,y)); } + + EIGEN_STRONG_INLINE Packet4cf pmul(const Packet8f& x, const Packet4cf& y) const + { return Packet4cf(Eigen::internal::pmul(x, y.v)); } +}; + +template<> struct conj_helper<Packet4cf, Packet8f, false,false> +{ + EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet4cf& x, const Packet8f& y, const Packet4cf& c) const + { return padd(c, pmul(x,y)); } + + EIGEN_STRONG_INLINE Packet4cf pmul(const Packet4cf& x, const Packet8f& y) const + { return Packet4cf(Eigen::internal::pmul(x.v, y)); } +}; + +template<> EIGEN_STRONG_INLINE Packet4cf pdiv<Packet4cf>(const Packet4cf& a, const Packet4cf& b) +{ + Packet4cf num = pmul(a, pconj(b)); + __m256 tmp = _mm256_mul_ps(b.v, b.v); + __m256 tmp2 = _mm256_shuffle_ps(tmp,tmp,0xB1); + __m256 denom = _mm256_add_ps(tmp, tmp2); + return Packet4cf(_mm256_div_ps(num.v, denom)); +} + +template<> EIGEN_STRONG_INLINE Packet4cf pcplxflip<Packet4cf>(const Packet4cf& x) +{ + return Packet4cf(_mm256_shuffle_ps(x.v, x.v, _MM_SHUFFLE(2, 3, 0 ,1))); +} + +//---------- double ---------- +struct Packet2cd +{ + EIGEN_STRONG_INLINE Packet2cd() {} + EIGEN_STRONG_INLINE explicit Packet2cd(const __m256d& a) : v(a) {} + __m256d v; +}; + +template<> struct packet_traits<std::complex<double> > : default_packet_traits +{ + typedef Packet2cd type; + typedef Packet1cd half; + enum { + Vectorizable = 1, + AlignedOnScalar = 0, + size = 2, + HasHalfPacket = 1, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasNegate = 1, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasSetLinear = 0 + }; +}; + +template<> struct unpacket_traits<Packet2cd> { typedef std::complex<double> type; enum {size=2, alignment=Aligned32}; 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)); } +template<> EIGEN_STRONG_INLINE Packet2cd pnegate(const Packet2cd& a) { return Packet2cd(pnegate(a.v)); } +template<> EIGEN_STRONG_INLINE Packet2cd pconj(const Packet2cd& a) +{ + const __m256d mask = _mm256_castsi256_pd(_mm256_set_epi32(0x80000000,0x0,0x0,0x0,0x80000000,0x0,0x0,0x0)); + return Packet2cd(_mm256_xor_pd(a.v,mask)); +} + +template<> EIGEN_STRONG_INLINE Packet2cd pmul<Packet2cd>(const Packet2cd& a, const Packet2cd& b) +{ + __m256d tmp1 = _mm256_shuffle_pd(a.v,a.v,0x0); + __m256d even = _mm256_mul_pd(tmp1, b.v); + __m256d tmp2 = _mm256_shuffle_pd(a.v,a.v,0xF); + __m256d tmp3 = _mm256_shuffle_pd(b.v,b.v,0x5); + __m256d odd = _mm256_mul_pd(tmp2, tmp3); + return Packet2cd(_mm256_addsub_pd(even, odd)); +} + +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 pload <Packet2cd>(const std::complex<double>* from) +{ EIGEN_DEBUG_ALIGNED_LOAD return Packet2cd(pload<Packet4d>((const double*)from)); } +template<> EIGEN_STRONG_INLINE Packet2cd ploadu<Packet2cd>(const std::complex<double>* from) +{ EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cd(ploadu<Packet4d>((const double*)from)); } + +template<> EIGEN_STRONG_INLINE Packet2cd pset1<Packet2cd>(const std::complex<double>& from) +{ + // in case casting to a __m128d* is really not safe, then we can still fallback to this version: (much slower though) +// return Packet2cd(_mm256_loadu2_m128d((const double*)&from,(const double*)&from)); + return Packet2cd(_mm256_broadcast_pd((const __m128d*)(const void*)&from)); +} + +template<> EIGEN_STRONG_INLINE Packet2cd ploaddup<Packet2cd>(const std::complex<double>* from) { return pset1<Packet2cd>(*from); } + +template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> * to, const Packet2cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); } +template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> * to, const Packet2cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); } + +template<> EIGEN_DEVICE_FUNC inline Packet2cd pgather<std::complex<double>, Packet2cd>(const std::complex<double>* from, Index stride) +{ + return Packet2cd(_mm256_set_pd(std::imag(from[1*stride]), std::real(from[1*stride]), + std::imag(from[0*stride]), std::real(from[0*stride]))); +} + +template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet2cd>(std::complex<double>* to, const Packet2cd& from, Index stride) +{ + __m128d low = _mm256_extractf128_pd(from.v, 0); + to[stride*0] = std::complex<double>(_mm_cvtsd_f64(low), _mm_cvtsd_f64(_mm_shuffle_pd(low, low, 1))); + __m128d high = _mm256_extractf128_pd(from.v, 1); + to[stride*1] = std::complex<double>(_mm_cvtsd_f64(high), _mm_cvtsd_f64(_mm_shuffle_pd(high, high, 1))); +} + +template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet2cd>(const Packet2cd& a) +{ + __m128d low = _mm256_extractf128_pd(a.v, 0); + EIGEN_ALIGN16 double res[2]; + _mm_store_pd(res, low); + return std::complex<double>(res[0],res[1]); +} + +template<> EIGEN_STRONG_INLINE Packet2cd preverse(const Packet2cd& a) { + __m256d result = _mm256_permute2f128_pd(a.v, a.v, 1); + return Packet2cd(result); +} + +template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet2cd>(const Packet2cd& a) +{ + return predux(padd(Packet1cd(_mm256_extractf128_pd(a.v,0)), + Packet1cd(_mm256_extractf128_pd(a.v,1)))); +} + +template<> EIGEN_STRONG_INLINE Packet2cd preduxp<Packet2cd>(const Packet2cd* vecs) +{ + Packet4d t0 = _mm256_permute2f128_pd(vecs[0].v,vecs[1].v, 0 + (2<<4)); + Packet4d t1 = _mm256_permute2f128_pd(vecs[0].v,vecs[1].v, 1 + (3<<4)); + + return Packet2cd(_mm256_add_pd(t0,t1)); +} + +template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet2cd>(const Packet2cd& a) +{ + return predux(pmul(Packet1cd(_mm256_extractf128_pd(a.v,0)), + Packet1cd(_mm256_extractf128_pd(a.v,1)))); +} + +template<int Offset> +struct palign_impl<Offset,Packet2cd> +{ + static EIGEN_STRONG_INLINE void run(Packet2cd& first, const Packet2cd& second) + { + if (Offset==0) return; + palign_impl<Offset*2,Packet4d>::run(first.v, second.v); + } +}; + +template<> struct conj_helper<Packet2cd, Packet2cd, false,true> +{ + EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet2cd& x, const Packet2cd& y, const Packet2cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet2cd pmul(const Packet2cd& a, const Packet2cd& b) const + { + return internal::pmul(a, pconj(b)); + } +}; + +template<> struct conj_helper<Packet2cd, Packet2cd, true,false> +{ + EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet2cd& x, const Packet2cd& y, const Packet2cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet2cd pmul(const Packet2cd& a, const Packet2cd& b) const + { + return internal::pmul(pconj(a), b); + } +}; + +template<> struct conj_helper<Packet2cd, Packet2cd, true,true> +{ + EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet2cd& x, const Packet2cd& y, const Packet2cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet2cd pmul(const Packet2cd& a, const Packet2cd& b) const + { + return pconj(internal::pmul(a, b)); + } +}; + +template<> struct conj_helper<Packet4d, Packet2cd, false,false> +{ + EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet4d& x, const Packet2cd& y, const Packet2cd& c) const + { return padd(c, pmul(x,y)); } + + EIGEN_STRONG_INLINE Packet2cd pmul(const Packet4d& x, const Packet2cd& y) const + { return Packet2cd(Eigen::internal::pmul(x, y.v)); } +}; + +template<> struct conj_helper<Packet2cd, Packet4d, false,false> +{ + EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet2cd& x, const Packet4d& y, const Packet2cd& c) const + { return padd(c, pmul(x,y)); } + + EIGEN_STRONG_INLINE Packet2cd pmul(const Packet2cd& x, const Packet4d& y) const + { return Packet2cd(Eigen::internal::pmul(x.v, y)); } +}; + +template<> EIGEN_STRONG_INLINE Packet2cd pdiv<Packet2cd>(const Packet2cd& a, const Packet2cd& b) +{ + Packet2cd num = pmul(a, pconj(b)); + __m256d tmp = _mm256_mul_pd(b.v, b.v); + __m256d denom = _mm256_hadd_pd(tmp, tmp); + return Packet2cd(_mm256_div_pd(num.v, denom)); +} + +template<> EIGEN_STRONG_INLINE Packet2cd pcplxflip<Packet2cd>(const Packet2cd& x) +{ + return Packet2cd(_mm256_shuffle_pd(x.v, x.v, 0x5)); +} + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet4cf,4>& kernel) { + __m256d P0 = _mm256_castps_pd(kernel.packet[0].v); + __m256d P1 = _mm256_castps_pd(kernel.packet[1].v); + __m256d P2 = _mm256_castps_pd(kernel.packet[2].v); + __m256d P3 = _mm256_castps_pd(kernel.packet[3].v); + + __m256d T0 = _mm256_shuffle_pd(P0, P1, 15); + __m256d T1 = _mm256_shuffle_pd(P0, P1, 0); + __m256d T2 = _mm256_shuffle_pd(P2, P3, 15); + __m256d T3 = _mm256_shuffle_pd(P2, P3, 0); + + kernel.packet[1].v = _mm256_castpd_ps(_mm256_permute2f128_pd(T0, T2, 32)); + kernel.packet[3].v = _mm256_castpd_ps(_mm256_permute2f128_pd(T0, T2, 49)); + kernel.packet[0].v = _mm256_castpd_ps(_mm256_permute2f128_pd(T1, T3, 32)); + kernel.packet[2].v = _mm256_castpd_ps(_mm256_permute2f128_pd(T1, T3, 49)); +} + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet2cd,2>& kernel) { + __m256d tmp = _mm256_permute2f128_pd(kernel.packet[0].v, kernel.packet[1].v, 0+(2<<4)); + kernel.packet[1].v = _mm256_permute2f128_pd(kernel.packet[0].v, kernel.packet[1].v, 1+(3<<4)); + kernel.packet[0].v = tmp; +} + +template<> EIGEN_STRONG_INLINE Packet4cf pinsertfirst(const Packet4cf& a, std::complex<float> b) +{ + return Packet4cf(_mm256_blend_ps(a.v,pset1<Packet4cf>(b).v,1|2)); +} + +template<> EIGEN_STRONG_INLINE Packet2cd pinsertfirst(const Packet2cd& a, std::complex<double> b) +{ + return Packet2cd(_mm256_blend_pd(a.v,pset1<Packet2cd>(b).v,1|2)); +} + +template<> EIGEN_STRONG_INLINE Packet4cf pinsertlast(const Packet4cf& a, std::complex<float> b) +{ + return Packet4cf(_mm256_blend_ps(a.v,pset1<Packet4cf>(b).v,(1<<7)|(1<<6))); +} + +template<> EIGEN_STRONG_INLINE Packet2cd pinsertlast(const Packet2cd& a, std::complex<double> b) +{ + return Packet2cd(_mm256_blend_pd(a.v,pset1<Packet2cd>(b).v,(1<<3)|(1<<2))); +} + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_COMPLEX_AVX_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AVX/MathFunctions.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AVX/MathFunctions.h new file mode 100644 index 000000000..6af67ce2d --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AVX/MathFunctions.h @@ -0,0 +1,439 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Pedro Gonnet (pedro.gonnet@gmail.com) +// +// 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_MATH_FUNCTIONS_AVX_H +#define EIGEN_MATH_FUNCTIONS_AVX_H + +/* The sin, cos, exp, and log functions of this file are loosely derived from + * Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/ + */ + +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); + + // Flip the sign on the odd intervals and return the result. + res = _mm256_xor_ps(res, _mm256_castsi256_ps(sign_flip_mask)); + return res; +} + +// 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)); +} + +// 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 <> +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); +} + +// Hyperbolic Tangent function. +template <> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f +ptanh<Packet8f>(const Packet8f& x) { + return internal::generic_fast_tanh_float(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); +} + +// Functions for sqrt. +// The EIGEN_FAST_MATH version uses the _mm_rsqrt_ps approximation and one step +// of Newton's method, at a cost of 1-2 bits of precision as opposed to the +// exact solution. It does not handle +inf, or denormalized numbers correctly. +// The main advantage of this approach is not just speed, but also the fact that +// it can be inlined and pipelined with other computations, further reducing its +// effective latency. This is similar to Quake3's fast inverse square root. +// For detail see here: http://www.beyond3d.com/content/articles/8/ +#if EIGEN_FAST_MATH +template <> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f +psqrt<Packet8f>(const Packet8f& _x) { + Packet8f half = pmul(_x, pset1<Packet8f>(.5f)); + Packet8f denormal_mask = _mm256_and_ps( + _mm256_cmp_ps(_x, pset1<Packet8f>((std::numeric_limits<float>::min)()), + _CMP_LT_OQ), + _mm256_cmp_ps(_x, _mm256_setzero_ps(), _CMP_GE_OQ)); + + // Compute approximate reciprocal sqrt. + Packet8f x = _mm256_rsqrt_ps(_x); + // Do a single step of Newton's iteration. + x = pmul(x, psub(pset1<Packet8f>(1.5f), pmul(half, pmul(x,x)))); + // Flush results for denormals to zero. + return _mm256_andnot_ps(denormal_mask, pmul(_x,x)); +} +#else +template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet8f psqrt<Packet8f>(const Packet8f& x) { + return _mm256_sqrt_ps(x); +} +#endif +template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet4d psqrt<Packet4d>(const Packet4d& x) { + return _mm256_sqrt_pd(x); +} +#if EIGEN_FAST_MATH + +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet8f prsqrt<Packet8f>(const Packet8f& _x) { + _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(inf, 0x7f800000); + _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(nan, 0x7fc00000); + _EIGEN_DECLARE_CONST_Packet8f(one_point_five, 1.5f); + _EIGEN_DECLARE_CONST_Packet8f(minus_half, -0.5f); + _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(flt_min, 0x00800000); + + Packet8f neg_half = pmul(_x, p8f_minus_half); + + // select only the inverse sqrt of positive normal inputs (denormals are + // flushed to zero and cause infs as well). + Packet8f le_zero_mask = _mm256_cmp_ps(_x, p8f_flt_min, _CMP_LT_OQ); + Packet8f x = _mm256_andnot_ps(le_zero_mask, _mm256_rsqrt_ps(_x)); + + // Fill in NaNs and Infs for the negative/zero entries. + Packet8f neg_mask = _mm256_cmp_ps(_x, _mm256_setzero_ps(), _CMP_LT_OQ); + Packet8f zero_mask = _mm256_andnot_ps(neg_mask, le_zero_mask); + Packet8f infs_and_nans = _mm256_or_ps(_mm256_and_ps(neg_mask, p8f_nan), + _mm256_and_ps(zero_mask, p8f_inf)); + + // Do a single step of Newton's iteration. + x = pmul(x, pmadd(neg_half, pmul(x, x), p8f_one_point_five)); + + // Insert NaNs and Infs in all the right places. + return _mm256_or_ps(x, infs_and_nans); +} + +#else +template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet8f prsqrt<Packet8f>(const Packet8f& x) { + _EIGEN_DECLARE_CONST_Packet8f(one, 1.0f); + return _mm256_div_ps(p8f_one, _mm256_sqrt_ps(x)); +} +#endif + +template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet4d prsqrt<Packet4d>(const Packet4d& x) { + _EIGEN_DECLARE_CONST_Packet4d(one, 1.0); + return _mm256_div_pd(p4d_one, _mm256_sqrt_pd(x)); +} + + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_MATH_FUNCTIONS_AVX_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AVX/PacketMath.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AVX/PacketMath.h new file mode 100644 index 000000000..195d40fb4 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AVX/PacketMath.h @@ -0,0 +1,633 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner (benoit.steiner.goog@gmail.com) +// +// 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_PACKET_MATH_AVX_H +#define EIGEN_PACKET_MATH_AVX_H + +namespace Eigen { + +namespace internal { + +#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD +#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8 +#endif + +#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS +#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*)) +#endif + +#ifdef __FMA__ +#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD +#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD +#endif +#endif + +typedef __m256 Packet8f; +typedef __m256i Packet8i; +typedef __m256d Packet4d; + +template<> struct is_arithmetic<__m256> { enum { value = true }; }; +template<> struct is_arithmetic<__m256i> { enum { value = true }; }; +template<> struct is_arithmetic<__m256d> { enum { value = true }; }; + +#define _EIGEN_DECLARE_CONST_Packet8f(NAME,X) \ + const Packet8f p8f_##NAME = pset1<Packet8f>(X) + +#define _EIGEN_DECLARE_CONST_Packet4d(NAME,X) \ + const Packet4d p4d_##NAME = pset1<Packet4d>(X) + +#define _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(NAME,X) \ + const Packet8f p8f_##NAME = _mm256_castsi256_ps(pset1<Packet8i>(X)) + +#define _EIGEN_DECLARE_CONST_Packet8i(NAME,X) \ + const Packet8i p8i_##NAME = pset1<Packet8i>(X) + +// Use the packet_traits defined in AVX512/PacketMath.h instead if we're going +// to leverage AVX512 instructions. +#ifndef EIGEN_VECTORIZE_AVX512 +template<> struct packet_traits<float> : default_packet_traits +{ + typedef Packet8f type; + typedef Packet4f half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=8, + HasHalfPacket = 1, + + HasDiv = 1, + HasSin = EIGEN_FAST_MATH, + HasCos = 0, + HasLog = 1, + HasExp = 1, + HasSqrt = 1, + HasRsqrt = 1, + HasTanh = EIGEN_FAST_MATH, + HasBlend = 1, + HasRound = 1, + HasFloor = 1, + HasCeil = 1 + }; +}; +template<> struct packet_traits<double> : default_packet_traits +{ + typedef Packet4d type; + typedef Packet2d half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=4, + HasHalfPacket = 1, + + HasDiv = 1, + HasExp = 1, + HasSqrt = 1, + HasRsqrt = 1, + HasBlend = 1, + HasRound = 1, + HasFloor = 1, + HasCeil = 1 + }; +}; +#endif + +template<> struct scalar_div_cost<float,true> { enum { value = 14 }; }; +template<> struct scalar_div_cost<double,true> { enum { value = 16 }; }; + +/* Proper support for integers is only provided by AVX2. In the meantime, we'll + use SSE instructions and packets to deal with integers. +template<> struct packet_traits<int> : default_packet_traits +{ + typedef Packet8i type; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=8 + }; +}; +*/ + +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<> 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 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); } + +template<> EIGEN_STRONG_INLINE Packet8f plset<Packet8f>(const float& a) { return _mm256_add_ps(_mm256_set1_ps(a), _mm256_set_ps(7.0,6.0,5.0,4.0,3.0,2.0,1.0,0.0)); } +template<> EIGEN_STRONG_INLINE Packet4d plset<Packet4d>(const double& a) { return _mm256_add_pd(_mm256_set1_pd(a), _mm256_set_pd(3.0,2.0,1.0,0.0)); } + +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 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); } + +template<> EIGEN_STRONG_INLINE Packet8f pnegate(const Packet8f& a) +{ + return _mm256_sub_ps(_mm256_set1_ps(0.0),a); +} +template<> EIGEN_STRONG_INLINE Packet4d pnegate(const Packet4d& a) +{ + return _mm256_sub_pd(_mm256_set1_pd(0.0),a); +} + +template<> EIGEN_STRONG_INLINE Packet8f pconj(const Packet8f& a) { return a; } +template<> EIGEN_STRONG_INLINE Packet4d pconj(const Packet4d& a) { return a; } +template<> EIGEN_STRONG_INLINE Packet8i pconj(const Packet8i& a) { return a; } + +template<> EIGEN_STRONG_INLINE Packet8f pmul<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_mul_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet4d pmul<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_mul_pd(a,b); } + + +template<> EIGEN_STRONG_INLINE Packet8f pdiv<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_div_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet4d pdiv<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_div_pd(a,b); } +template<> EIGEN_STRONG_INLINE Packet8i pdiv<Packet8i>(const Packet8i& /*a*/, const Packet8i& /*b*/) +{ eigen_assert(false && "packet integer division are not supported by AVX"); + return pset1<Packet8i>(0); +} + +#ifdef __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. + Packet8f res = c; + __asm__("vfmadd231ps %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b)); + return res; +#else + return _mm256_fmadd_ps(a,b,c); +#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)) ) + // see above + Packet4d res = c; + __asm__("vfmadd231pd %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b)); + return res; +#else + return _mm256_fmadd_pd(a,b,c); +#endif +} +#endif + +template<> EIGEN_STRONG_INLINE Packet8f pmin<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_min_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet4d pmin<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_min_pd(a,b); } + +template<> EIGEN_STRONG_INLINE Packet8f pmax<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_max_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet4d pmax<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_max_pd(a,b); } + +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); } + +template<> EIGEN_STRONG_INLINE Packet8f pceil<Packet8f>(const Packet8f& a) { return _mm256_ceil_ps(a); } +template<> EIGEN_STRONG_INLINE Packet4d pceil<Packet4d>(const Packet4d& a) { return _mm256_ceil_pd(a); } + +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 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 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 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 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 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); } +template<> EIGEN_STRONG_INLINE Packet8i pload<Packet8i>(const int* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm256_load_si256(reinterpret_cast<const __m256i*>(from)); } + +template<> EIGEN_STRONG_INLINE Packet8f ploadu<Packet8f>(const float* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm256_loadu_ps(from); } +template<> EIGEN_STRONG_INLINE Packet4d ploadu<Packet4d>(const double* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm256_loadu_pd(from); } +template<> EIGEN_STRONG_INLINE Packet8i ploadu<Packet8i>(const int* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm256_loadu_si256(reinterpret_cast<const __m256i*>(from)); } + +// Loads 4 floats from memory a returns the packet {a0, a0 a1, a1, a2, a2, a3, a3} +template<> EIGEN_STRONG_INLINE Packet8f ploaddup<Packet8f>(const float* from) +{ + // TODO try to find a way to avoid the need of a temporary register +// Packet8f tmp = _mm256_castps128_ps256(_mm_loadu_ps(from)); +// tmp = _mm256_insertf128_ps(tmp, _mm_movehl_ps(_mm256_castps256_ps128(tmp),_mm256_castps256_ps128(tmp)), 1); +// return _mm256_unpacklo_ps(tmp,tmp); + + // _mm256_insertf128_ps is very slow on Haswell, thus: + Packet8f tmp = _mm256_broadcast_ps((const __m128*)(const void*)from); + // mimic an "inplace" permutation of the lower 128bits using a blend + tmp = _mm256_blend_ps(tmp,_mm256_castps128_ps256(_mm_permute_ps( _mm256_castps256_ps128(tmp), _MM_SHUFFLE(1,0,1,0))), 15); + // then we can perform a consistent permutation on the global register to get everything in shape: + return _mm256_permute_ps(tmp, _MM_SHUFFLE(3,3,2,2)); +} +// Loads 2 doubles from memory a returns the packet {a0, a0 a1, a1} +template<> EIGEN_STRONG_INLINE Packet4d ploaddup<Packet4d>(const double* from) +{ + Packet4d tmp = _mm256_broadcast_pd((const __m128d*)(const void*)from); + return _mm256_permute_pd(tmp, 3<<2); +} + +// Loads 2 floats from memory a returns the packet {a0, a0 a0, a0, a1, a1, a1, a1} +template<> EIGEN_STRONG_INLINE Packet8f ploadquad<Packet8f>(const float* from) +{ + Packet8f tmp = _mm256_castps128_ps256(_mm_broadcast_ss(from)); + return _mm256_insertf128_ps(tmp, _mm_broadcast_ss(from+1), 1); +} + +template<> EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet8f& from) { EIGEN_DEBUG_ALIGNED_STORE _mm256_store_ps(to, from); } +template<> EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet4d& from) { EIGEN_DEBUG_ALIGNED_STORE _mm256_store_pd(to, from); } +template<> EIGEN_STRONG_INLINE void pstore<int>(int* to, const Packet8i& from) { EIGEN_DEBUG_ALIGNED_STORE _mm256_storeu_si256(reinterpret_cast<__m256i*>(to), from); } + +template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet8f& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_ps(to, from); } +template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet4d& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_pd(to, from); } +template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet8i& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_si256(reinterpret_cast<__m256i*>(to), from); } + +// NOTE: leverage _mm256_i32gather_ps and _mm256_i32gather_pd if AVX2 instructions are available +// NOTE: for the record the following seems to be slower: return _mm256_i32gather_ps(from, _mm256_set1_epi32(stride), 4); +template<> EIGEN_DEVICE_FUNC inline Packet8f pgather<float, Packet8f>(const float* from, Index stride) +{ + return _mm256_set_ps(from[7*stride], from[6*stride], from[5*stride], from[4*stride], + from[3*stride], from[2*stride], from[1*stride], from[0*stride]); +} +template<> EIGEN_DEVICE_FUNC inline Packet4d pgather<double, Packet4d>(const double* from, Index stride) +{ + return _mm256_set_pd(from[3*stride], from[2*stride], from[1*stride], from[0*stride]); +} + +template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet8f>(float* to, const Packet8f& from, Index stride) +{ + __m128 low = _mm256_extractf128_ps(from, 0); + to[stride*0] = _mm_cvtss_f32(low); + to[stride*1] = _mm_cvtss_f32(_mm_shuffle_ps(low, low, 1)); + to[stride*2] = _mm_cvtss_f32(_mm_shuffle_ps(low, low, 2)); + to[stride*3] = _mm_cvtss_f32(_mm_shuffle_ps(low, low, 3)); + + __m128 high = _mm256_extractf128_ps(from, 1); + to[stride*4] = _mm_cvtss_f32(high); + to[stride*5] = _mm_cvtss_f32(_mm_shuffle_ps(high, high, 1)); + to[stride*6] = _mm_cvtss_f32(_mm_shuffle_ps(high, high, 2)); + to[stride*7] = _mm_cvtss_f32(_mm_shuffle_ps(high, high, 3)); +} +template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet4d>(double* to, const Packet4d& from, Index stride) +{ + __m128d low = _mm256_extractf128_pd(from, 0); + to[stride*0] = _mm_cvtsd_f64(low); + to[stride*1] = _mm_cvtsd_f64(_mm_shuffle_pd(low, low, 1)); + __m128d high = _mm256_extractf128_pd(from, 1); + to[stride*2] = _mm_cvtsd_f64(high); + to[stride*3] = _mm_cvtsd_f64(_mm_shuffle_pd(high, high, 1)); +} + +template<> EIGEN_STRONG_INLINE void pstore1<Packet8f>(float* to, const float& a) +{ + Packet8f pa = pset1<Packet8f>(a); + pstore(to, pa); +} +template<> EIGEN_STRONG_INLINE void pstore1<Packet4d>(double* to, const double& a) +{ + Packet4d pa = pset1<Packet4d>(a); + pstore(to, pa); +} +template<> EIGEN_STRONG_INLINE void pstore1<Packet8i>(int* to, const int& a) +{ + Packet8i pa = pset1<Packet8i>(a); + pstore(to, pa); +} + +#ifndef EIGEN_VECTORIZE_AVX512 +template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } +template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } +template<> EIGEN_STRONG_INLINE void prefetch<int>(const int* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } +#endif + +template<> EIGEN_STRONG_INLINE float pfirst<Packet8f>(const Packet8f& a) { + return _mm_cvtss_f32(_mm256_castps256_ps128(a)); +} +template<> EIGEN_STRONG_INLINE double pfirst<Packet4d>(const Packet4d& a) { + return _mm_cvtsd_f64(_mm256_castpd256_pd128(a)); +} +template<> EIGEN_STRONG_INLINE int pfirst<Packet8i>(const Packet8i& a) { + return _mm_cvtsi128_si32(_mm256_castsi256_si128(a)); +} + + +template<> EIGEN_STRONG_INLINE Packet8f preverse(const Packet8f& a) +{ + __m256 tmp = _mm256_shuffle_ps(a,a,0x1b); + return _mm256_permute2f128_ps(tmp, tmp, 1); +} +template<> EIGEN_STRONG_INLINE Packet4d preverse(const Packet4d& a) +{ + __m256d tmp = _mm256_shuffle_pd(a,a,5); + return _mm256_permute2f128_pd(tmp, tmp, 1); + + __m256d swap_halves = _mm256_permute2f128_pd(a,a,1); + return _mm256_permute_pd(swap_halves,5); +} + +// pabs should be ok +template<> EIGEN_STRONG_INLINE Packet8f pabs(const Packet8f& a) +{ + const Packet8f mask = _mm256_castsi256_ps(_mm256_setr_epi32(0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF)); + return _mm256_and_ps(a,mask); +} +template<> EIGEN_STRONG_INLINE Packet4d pabs(const Packet4d& a) +{ + const Packet4d mask = _mm256_castsi256_pd(_mm256_setr_epi32(0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF)); + return _mm256_and_pd(a,mask); +} + +// 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) +{ + __m256 hsum1 = _mm256_hadd_ps(vecs[0], vecs[1]); + __m256 hsum2 = _mm256_hadd_ps(vecs[2], vecs[3]); + __m256 hsum3 = _mm256_hadd_ps(vecs[4], vecs[5]); + __m256 hsum4 = _mm256_hadd_ps(vecs[6], vecs[7]); + + __m256 hsum5 = _mm256_hadd_ps(hsum1, hsum1); + __m256 hsum6 = _mm256_hadd_ps(hsum2, hsum2); + __m256 hsum7 = _mm256_hadd_ps(hsum3, hsum3); + __m256 hsum8 = _mm256_hadd_ps(hsum4, hsum4); + + __m256 perm1 = _mm256_permute2f128_ps(hsum5, hsum5, 0x23); + __m256 perm2 = _mm256_permute2f128_ps(hsum6, hsum6, 0x23); + __m256 perm3 = _mm256_permute2f128_ps(hsum7, hsum7, 0x23); + __m256 perm4 = _mm256_permute2f128_ps(hsum8, hsum8, 0x23); + + __m256 sum1 = _mm256_add_ps(perm1, hsum5); + __m256 sum2 = _mm256_add_ps(perm2, hsum6); + __m256 sum3 = _mm256_add_ps(perm3, hsum7); + __m256 sum4 = _mm256_add_ps(perm4, hsum8); + + __m256 blend1 = _mm256_blend_ps(sum1, sum2, 0xcc); + __m256 blend2 = _mm256_blend_ps(sum3, sum4, 0xcc); + + __m256 final = _mm256_blend_ps(blend1, blend2, 0xf0); + return final; +} +template<> EIGEN_STRONG_INLINE Packet4d preduxp<Packet4d>(const Packet4d* vecs) +{ + Packet4d tmp0, tmp1; + + tmp0 = _mm256_hadd_pd(vecs[0], vecs[1]); + tmp0 = _mm256_add_pd(tmp0, _mm256_permute2f128_pd(tmp0, tmp0, 1)); + + tmp1 = _mm256_hadd_pd(vecs[2], vecs[3]); + tmp1 = _mm256_add_pd(tmp1, _mm256_permute2f128_pd(tmp1, tmp1, 1)); + + return _mm256_blend_pd(tmp0, tmp1, 0xC); +} + +template<> EIGEN_STRONG_INLINE float predux<Packet8f>(const Packet8f& a) +{ + return predux(Packet4f(_mm_add_ps(_mm256_castps256_ps128(a),_mm256_extractf128_ps(a,1)))); +} +template<> EIGEN_STRONG_INLINE double predux<Packet4d>(const Packet4d& a) +{ + return predux(Packet2d(_mm_add_pd(_mm256_castpd256_pd128(a),_mm256_extractf128_pd(a,1)))); +} + +template<> EIGEN_STRONG_INLINE Packet4f predux_downto4<Packet8f>(const Packet8f& a) +{ + return _mm_add_ps(_mm256_castps256_ps128(a),_mm256_extractf128_ps(a,1)); +} + +template<> EIGEN_STRONG_INLINE float predux_mul<Packet8f>(const Packet8f& a) +{ + Packet8f tmp; + tmp = _mm256_mul_ps(a, _mm256_permute2f128_ps(a,a,1)); + tmp = _mm256_mul_ps(tmp, _mm256_shuffle_ps(tmp,tmp,_MM_SHUFFLE(1,0,3,2))); + return pfirst(_mm256_mul_ps(tmp, _mm256_shuffle_ps(tmp,tmp,1))); +} +template<> EIGEN_STRONG_INLINE double predux_mul<Packet4d>(const Packet4d& a) +{ + Packet4d tmp; + tmp = _mm256_mul_pd(a, _mm256_permute2f128_pd(a,a,1)); + return pfirst(_mm256_mul_pd(tmp, _mm256_shuffle_pd(tmp,tmp,1))); +} + +template<> EIGEN_STRONG_INLINE float predux_min<Packet8f>(const Packet8f& a) +{ + Packet8f tmp = _mm256_min_ps(a, _mm256_permute2f128_ps(a,a,1)); + tmp = _mm256_min_ps(tmp, _mm256_shuffle_ps(tmp,tmp,_MM_SHUFFLE(1,0,3,2))); + return pfirst(_mm256_min_ps(tmp, _mm256_shuffle_ps(tmp,tmp,1))); +} +template<> EIGEN_STRONG_INLINE double predux_min<Packet4d>(const Packet4d& a) +{ + Packet4d tmp = _mm256_min_pd(a, _mm256_permute2f128_pd(a,a,1)); + return pfirst(_mm256_min_pd(tmp, _mm256_shuffle_pd(tmp, tmp, 1))); +} + +template<> EIGEN_STRONG_INLINE float predux_max<Packet8f>(const Packet8f& a) +{ + Packet8f tmp = _mm256_max_ps(a, _mm256_permute2f128_ps(a,a,1)); + tmp = _mm256_max_ps(tmp, _mm256_shuffle_ps(tmp,tmp,_MM_SHUFFLE(1,0,3,2))); + return pfirst(_mm256_max_ps(tmp, _mm256_shuffle_ps(tmp,tmp,1))); +} + +template<> EIGEN_STRONG_INLINE double predux_max<Packet4d>(const Packet4d& a) +{ + Packet4d tmp = _mm256_max_pd(a, _mm256_permute2f128_pd(a,a,1)); + return pfirst(_mm256_max_pd(tmp, _mm256_shuffle_pd(tmp, tmp, 1))); +} + + +template<int Offset> +struct palign_impl<Offset,Packet8f> +{ + static EIGEN_STRONG_INLINE void run(Packet8f& first, const Packet8f& second) + { + if (Offset==1) + { + first = _mm256_blend_ps(first, second, 1); + Packet8f tmp1 = _mm256_permute_ps (first, _MM_SHUFFLE(0,3,2,1)); + Packet8f tmp2 = _mm256_permute2f128_ps (tmp1, tmp1, 1); + first = _mm256_blend_ps(tmp1, tmp2, 0x88); + } + else if (Offset==2) + { + first = _mm256_blend_ps(first, second, 3); + Packet8f tmp1 = _mm256_permute_ps (first, _MM_SHUFFLE(1,0,3,2)); + Packet8f tmp2 = _mm256_permute2f128_ps (tmp1, tmp1, 1); + first = _mm256_blend_ps(tmp1, tmp2, 0xcc); + } + else if (Offset==3) + { + first = _mm256_blend_ps(first, second, 7); + Packet8f tmp1 = _mm256_permute_ps (first, _MM_SHUFFLE(2,1,0,3)); + Packet8f tmp2 = _mm256_permute2f128_ps (tmp1, tmp1, 1); + first = _mm256_blend_ps(tmp1, tmp2, 0xee); + } + else if (Offset==4) + { + first = _mm256_blend_ps(first, second, 15); + Packet8f tmp1 = _mm256_permute_ps (first, _MM_SHUFFLE(3,2,1,0)); + Packet8f tmp2 = _mm256_permute2f128_ps (tmp1, tmp1, 1); + first = _mm256_permute_ps(tmp2, _MM_SHUFFLE(3,2,1,0)); + } + else if (Offset==5) + { + first = _mm256_blend_ps(first, second, 31); + first = _mm256_permute2f128_ps(first, first, 1); + Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(0,3,2,1)); + first = _mm256_permute2f128_ps(tmp, tmp, 1); + first = _mm256_blend_ps(tmp, first, 0x88); + } + else if (Offset==6) + { + first = _mm256_blend_ps(first, second, 63); + first = _mm256_permute2f128_ps(first, first, 1); + Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(1,0,3,2)); + first = _mm256_permute2f128_ps(tmp, tmp, 1); + first = _mm256_blend_ps(tmp, first, 0xcc); + } + else if (Offset==7) + { + first = _mm256_blend_ps(first, second, 127); + first = _mm256_permute2f128_ps(first, first, 1); + Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(2,1,0,3)); + first = _mm256_permute2f128_ps(tmp, tmp, 1); + first = _mm256_blend_ps(tmp, first, 0xee); + } + } +}; + +template<int Offset> +struct palign_impl<Offset,Packet4d> +{ + static EIGEN_STRONG_INLINE void run(Packet4d& first, const Packet4d& second) + { + if (Offset==1) + { + first = _mm256_blend_pd(first, second, 1); + __m256d tmp = _mm256_permute_pd(first, 5); + first = _mm256_permute2f128_pd(tmp, tmp, 1); + first = _mm256_blend_pd(tmp, first, 0xA); + } + else if (Offset==2) + { + first = _mm256_blend_pd(first, second, 3); + first = _mm256_permute2f128_pd(first, first, 1); + } + else if (Offset==3) + { + first = _mm256_blend_pd(first, second, 7); + __m256d tmp = _mm256_permute_pd(first, 5); + first = _mm256_permute2f128_pd(tmp, tmp, 1); + first = _mm256_blend_pd(tmp, first, 5); + } + } +}; + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet8f,8>& kernel) { + __m256 T0 = _mm256_unpacklo_ps(kernel.packet[0], kernel.packet[1]); + __m256 T1 = _mm256_unpackhi_ps(kernel.packet[0], kernel.packet[1]); + __m256 T2 = _mm256_unpacklo_ps(kernel.packet[2], kernel.packet[3]); + __m256 T3 = _mm256_unpackhi_ps(kernel.packet[2], kernel.packet[3]); + __m256 T4 = _mm256_unpacklo_ps(kernel.packet[4], kernel.packet[5]); + __m256 T5 = _mm256_unpackhi_ps(kernel.packet[4], kernel.packet[5]); + __m256 T6 = _mm256_unpacklo_ps(kernel.packet[6], kernel.packet[7]); + __m256 T7 = _mm256_unpackhi_ps(kernel.packet[6], kernel.packet[7]); + __m256 S0 = _mm256_shuffle_ps(T0,T2,_MM_SHUFFLE(1,0,1,0)); + __m256 S1 = _mm256_shuffle_ps(T0,T2,_MM_SHUFFLE(3,2,3,2)); + __m256 S2 = _mm256_shuffle_ps(T1,T3,_MM_SHUFFLE(1,0,1,0)); + __m256 S3 = _mm256_shuffle_ps(T1,T3,_MM_SHUFFLE(3,2,3,2)); + __m256 S4 = _mm256_shuffle_ps(T4,T6,_MM_SHUFFLE(1,0,1,0)); + __m256 S5 = _mm256_shuffle_ps(T4,T6,_MM_SHUFFLE(3,2,3,2)); + __m256 S6 = _mm256_shuffle_ps(T5,T7,_MM_SHUFFLE(1,0,1,0)); + __m256 S7 = _mm256_shuffle_ps(T5,T7,_MM_SHUFFLE(3,2,3,2)); + kernel.packet[0] = _mm256_permute2f128_ps(S0, S4, 0x20); + kernel.packet[1] = _mm256_permute2f128_ps(S1, S5, 0x20); + kernel.packet[2] = _mm256_permute2f128_ps(S2, S6, 0x20); + kernel.packet[3] = _mm256_permute2f128_ps(S3, S7, 0x20); + kernel.packet[4] = _mm256_permute2f128_ps(S0, S4, 0x31); + kernel.packet[5] = _mm256_permute2f128_ps(S1, S5, 0x31); + kernel.packet[6] = _mm256_permute2f128_ps(S2, S6, 0x31); + kernel.packet[7] = _mm256_permute2f128_ps(S3, S7, 0x31); +} + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet8f,4>& kernel) { + __m256 T0 = _mm256_unpacklo_ps(kernel.packet[0], kernel.packet[1]); + __m256 T1 = _mm256_unpackhi_ps(kernel.packet[0], kernel.packet[1]); + __m256 T2 = _mm256_unpacklo_ps(kernel.packet[2], kernel.packet[3]); + __m256 T3 = _mm256_unpackhi_ps(kernel.packet[2], kernel.packet[3]); + + __m256 S0 = _mm256_shuffle_ps(T0,T2,_MM_SHUFFLE(1,0,1,0)); + __m256 S1 = _mm256_shuffle_ps(T0,T2,_MM_SHUFFLE(3,2,3,2)); + __m256 S2 = _mm256_shuffle_ps(T1,T3,_MM_SHUFFLE(1,0,1,0)); + __m256 S3 = _mm256_shuffle_ps(T1,T3,_MM_SHUFFLE(3,2,3,2)); + + kernel.packet[0] = _mm256_permute2f128_ps(S0, S1, 0x20); + kernel.packet[1] = _mm256_permute2f128_ps(S2, S3, 0x20); + kernel.packet[2] = _mm256_permute2f128_ps(S0, S1, 0x31); + kernel.packet[3] = _mm256_permute2f128_ps(S2, S3, 0x31); +} + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet4d,4>& kernel) { + __m256d T0 = _mm256_shuffle_pd(kernel.packet[0], kernel.packet[1], 15); + __m256d T1 = _mm256_shuffle_pd(kernel.packet[0], kernel.packet[1], 0); + __m256d T2 = _mm256_shuffle_pd(kernel.packet[2], kernel.packet[3], 15); + __m256d T3 = _mm256_shuffle_pd(kernel.packet[2], kernel.packet[3], 0); + + kernel.packet[1] = _mm256_permute2f128_pd(T0, T2, 32); + kernel.packet[3] = _mm256_permute2f128_pd(T0, T2, 49); + kernel.packet[0] = _mm256_permute2f128_pd(T1, T3, 32); + kernel.packet[2] = _mm256_permute2f128_pd(T1, T3, 49); +} + +template<> EIGEN_STRONG_INLINE Packet8f pblend(const Selector<8>& ifPacket, const Packet8f& thenPacket, const Packet8f& elsePacket) { + const __m256 zero = _mm256_setzero_ps(); + const __m256 select = _mm256_set_ps(ifPacket.select[7], ifPacket.select[6], ifPacket.select[5], ifPacket.select[4], ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]); + __m256 false_mask = _mm256_cmp_ps(select, zero, _CMP_EQ_UQ); + return _mm256_blendv_ps(thenPacket, elsePacket, false_mask); +} +template<> EIGEN_STRONG_INLINE Packet4d pblend(const Selector<4>& ifPacket, const Packet4d& thenPacket, const Packet4d& elsePacket) { + const __m256d zero = _mm256_setzero_pd(); + const __m256d select = _mm256_set_pd(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]); + __m256d false_mask = _mm256_cmp_pd(select, zero, _CMP_EQ_UQ); + return _mm256_blendv_pd(thenPacket, elsePacket, false_mask); +} + +template<> EIGEN_STRONG_INLINE Packet8f pinsertfirst(const Packet8f& a, float b) +{ + return _mm256_blend_ps(a,pset1<Packet8f>(b),1); +} + +template<> EIGEN_STRONG_INLINE Packet4d pinsertfirst(const Packet4d& a, double b) +{ + return _mm256_blend_pd(a,pset1<Packet4d>(b),1); +} + +template<> EIGEN_STRONG_INLINE Packet8f pinsertlast(const Packet8f& a, float b) +{ + return _mm256_blend_ps(a,pset1<Packet8f>(b),(1<<7)); +} + +template<> EIGEN_STRONG_INLINE Packet4d pinsertlast(const Packet4d& a, double b) +{ + return _mm256_blend_pd(a,pset1<Packet4d>(b),(1<<3)); +} + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_PACKET_MATH_AVX_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AVX/TypeCasting.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AVX/TypeCasting.h new file mode 100644 index 000000000..83bfdc604 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AVX/TypeCasting.h @@ -0,0 +1,51 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2015 Benoit Steiner <benoit.steiner.goog@gmail.com> +// +// 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_TYPE_CASTING_AVX_H +#define EIGEN_TYPE_CASTING_AVX_H + +namespace Eigen { + +namespace internal { + +// For now we use SSE to handle integers, so we can't use AVX instructions to cast +// from int to float +template <> +struct type_casting_traits<float, int> { + enum { + VectorizedCast = 0, + SrcCoeffRatio = 1, + TgtCoeffRatio = 1 + }; +}; + +template <> +struct type_casting_traits<int, float> { + enum { + VectorizedCast = 0, + SrcCoeffRatio = 1, + TgtCoeffRatio = 1 + }; +}; + + + +template<> EIGEN_STRONG_INLINE Packet8i pcast<Packet8f, Packet8i>(const Packet8f& a) { + return _mm256_cvtps_epi32(a); +} + +template<> EIGEN_STRONG_INLINE Packet8f pcast<Packet8i, Packet8f>(const Packet8i& a) { + return _mm256_cvtepi32_ps(a); +} + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_TYPE_CASTING_AVX_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AVX512/MathFunctions.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AVX512/MathFunctions.h new file mode 100644 index 000000000..399be0ee4 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AVX512/MathFunctions.h @@ -0,0 +1,396 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2016 Pedro Gonnet (pedro.gonnet@gmail.com) +// +// 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 THIRD_PARTY_EIGEN3_EIGEN_SRC_CORE_ARCH_AVX512_MATHFUNCTIONS_H_ +#define THIRD_PARTY_EIGEN3_EIGEN_SRC_CORE_ARCH_AVX512_MATHFUNCTIONS_H_ + +namespace Eigen { + +namespace internal { + +// Disable the code for older versions of gcc that don't support many of the required avx512 instrinsics. +#if EIGEN_GNUC_AT_LEAST(5, 3) + +#define _EIGEN_DECLARE_CONST_Packet16f(NAME, X) \ + const Packet16f p16f_##NAME = pset1<Packet16f>(X) + +#define _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(NAME, X) \ + const Packet16f p16f_##NAME = (__m512)pset1<Packet16i>(X) + +#define _EIGEN_DECLARE_CONST_Packet8d(NAME, X) \ + const Packet8d p8d_##NAME = pset1<Packet8d>(X) + +#define _EIGEN_DECLARE_CONST_Packet8d_FROM_INT64(NAME, X) \ + const Packet8d p8d_##NAME = _mm512_castsi512_pd(_mm512_set1_epi64(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. +#if defined(EIGEN_VECTORIZE_AVX512DQ) +template <> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet16f +plog<Packet16f>(const Packet16f& _x) { + Packet16f x = _x; + _EIGEN_DECLARE_CONST_Packet16f(1, 1.0f); + _EIGEN_DECLARE_CONST_Packet16f(half, 0.5f); + _EIGEN_DECLARE_CONST_Packet16f(126f, 126.0f); + + _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(inv_mant_mask, ~0x7f800000); + + // 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(nan, 0x7fc00000); + + // Polynomial coefficients. + _EIGEN_DECLARE_CONST_Packet16f(cephes_SQRTHF, 0.707106781186547524f); + _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p0, 7.0376836292E-2f); + _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p1, -1.1514610310E-1f); + _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p2, 1.1676998740E-1f); + _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p3, -1.2420140846E-1f); + _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p4, +1.4249322787E-1f); + _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p5, -1.6668057665E-1f); + _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p6, +2.0000714765E-1f); + _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p7, -2.4999993993E-1f); + _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p8, +3.3333331174E-1f); + _EIGEN_DECLARE_CONST_Packet16f(cephes_log_q1, -2.12194440e-4f); + _EIGEN_DECLARE_CONST_Packet16f(cephes_log_q2, 0.693359375f); + + // invalid_mask is set to true when x is NaN + __mmask16 invalid_mask = + _mm512_cmp_ps_mask(x, _mm512_setzero_ps(), _CMP_NGE_UQ); + __mmask16 iszero_mask = + _mm512_cmp_ps_mask(x, _mm512_setzero_ps(), _CMP_EQ_UQ); + + // Truncate input values to the minimum positive normal. + x = pmax(x, p16f_min_norm_pos); + + // Extract the shifted exponents. + Packet16f emm0 = _mm512_cvtepi32_ps(_mm512_srli_epi32((__m512i)x, 23)); + Packet16f e = _mm512_sub_ps(emm0, p16f_126f); + + // Set the exponents to -1, i.e. x are in the range [0.5,1). + x = _mm512_and_ps(x, p16f_inv_mant_mask); + x = _mm512_or_ps(x, p16f_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; } + __mmask16 mask = _mm512_cmp_ps_mask(x, p16f_cephes_SQRTHF, _CMP_LT_OQ); + Packet16f tmp = _mm512_mask_blend_ps(mask, x, _mm512_setzero_ps()); + x = psub(x, p16f_1); + e = psub(e, _mm512_mask_blend_ps(mask, p16f_1, _mm512_setzero_ps())); + x = padd(x, tmp); + + Packet16f x2 = pmul(x, x); + Packet16f x3 = pmul(x2, x); + + // Evaluate the polynomial approximant of degree 8 in three parts, probably + // to improve instruction-level parallelism. + Packet16f y, y1, y2; + y = pmadd(p16f_cephes_log_p0, x, p16f_cephes_log_p1); + y1 = pmadd(p16f_cephes_log_p3, x, p16f_cephes_log_p4); + y2 = pmadd(p16f_cephes_log_p6, x, p16f_cephes_log_p7); + y = pmadd(y, x, p16f_cephes_log_p2); + y1 = pmadd(y1, x, p16f_cephes_log_p5); + y2 = pmadd(y2, x, p16f_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, p16f_cephes_log_q1); + tmp = pmul(x2, p16f_half); + y = padd(y, y1); + x = psub(x, tmp); + y2 = pmul(e, p16f_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 _mm512_mask_blend_ps(iszero_mask, p16f_minus_inf, + _mm512_mask_blend_ps(invalid_mask, p16f_nan, x)); +} +#endif + +// 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 <> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet16f +pexp<Packet16f>(const Packet16f& _x) { + _EIGEN_DECLARE_CONST_Packet16f(1, 1.0f); + _EIGEN_DECLARE_CONST_Packet16f(half, 0.5f); + _EIGEN_DECLARE_CONST_Packet16f(127, 127.0f); + + _EIGEN_DECLARE_CONST_Packet16f(exp_hi, 88.3762626647950f); + _EIGEN_DECLARE_CONST_Packet16f(exp_lo, -88.3762626647949f); + + _EIGEN_DECLARE_CONST_Packet16f(cephes_LOG2EF, 1.44269504088896341f); + + _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p0, 1.9875691500E-4f); + _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p1, 1.3981999507E-3f); + _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p2, 8.3334519073E-3f); + _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p3, 4.1665795894E-2f); + _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p4, 1.6666665459E-1f); + _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p5, 5.0000001201E-1f); + + // Clamp x. + Packet16f x = pmax(pmin(_x, p16f_exp_hi), p16f_exp_lo); + + // Express exp(x) as exp(m*ln(2) + r), start by extracting + // m = floor(x/ln(2) + 0.5). + Packet16f m = _mm512_floor_ps(pmadd(x, p16f_cephes_LOG2EF, p16f_half)); + + // Get r = x - m*ln(2). Note that we can do this without losing more than one + // ulp precision due to the FMA instruction. + _EIGEN_DECLARE_CONST_Packet16f(nln2, -0.6931471805599453f); + Packet16f r = _mm512_fmadd_ps(m, p16f_nln2, x); + Packet16f r2 = pmul(r, r); + + // TODO(gonnet): Split into odd/even polynomials and try to exploit + // instruction-level parallelism. + Packet16f y = p16f_cephes_exp_p0; + y = pmadd(y, r, p16f_cephes_exp_p1); + y = pmadd(y, r, p16f_cephes_exp_p2); + y = pmadd(y, r, p16f_cephes_exp_p3); + y = pmadd(y, r, p16f_cephes_exp_p4); + y = pmadd(y, r, p16f_cephes_exp_p5); + y = pmadd(y, r2, r); + y = padd(y, p16f_1); + + // Build emm0 = 2^m. + Packet16i emm0 = _mm512_cvttps_epi32(padd(m, p16f_127)); + emm0 = _mm512_slli_epi32(emm0, 23); + + // Return 2^m * exp(r). + return pmax(pmul(y, _mm512_castsi512_ps(emm0)), _x); +} + +/*template <> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8d +pexp<Packet8d>(const Packet8d& _x) { + Packet8d x = _x; + + _EIGEN_DECLARE_CONST_Packet8d(1, 1.0); + _EIGEN_DECLARE_CONST_Packet8d(2, 2.0); + + _EIGEN_DECLARE_CONST_Packet8d(exp_hi, 709.437); + _EIGEN_DECLARE_CONST_Packet8d(exp_lo, -709.436139303); + + _EIGEN_DECLARE_CONST_Packet8d(cephes_LOG2EF, 1.4426950408889634073599); + + _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_p0, 1.26177193074810590878e-4); + _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_p1, 3.02994407707441961300e-2); + _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_p2, 9.99999999999999999910e-1); + + _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_q0, 3.00198505138664455042e-6); + _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_q1, 2.52448340349684104192e-3); + _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_q2, 2.27265548208155028766e-1); + _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_q3, 2.00000000000000000009e0); + + _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_C1, 0.693145751953125); + _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_C2, 1.42860682030941723212e-6); + + // clamp x + x = pmax(pmin(x, p8d_exp_hi), p8d_exp_lo); + + // Express exp(x) as exp(g + n*log(2)). + const Packet8d n = + _mm512_mul_round_pd(p8d_cephes_LOG2EF, x, _MM_FROUND_TO_NEAREST_INT); + + // 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. + const Packet8d nC1 = pmul(n, p8d_cephes_exp_C1); + const Packet8d nC2 = pmul(n, p8d_cephes_exp_C2); + x = psub(x, nC1); + x = psub(x, nC2); + + const Packet8d x2 = pmul(x, x); + + // Evaluate the numerator polynomial of the rational interpolant. + Packet8d px = p8d_cephes_exp_p0; + px = pmadd(px, x2, p8d_cephes_exp_p1); + px = pmadd(px, x2, p8d_cephes_exp_p2); + px = pmul(px, x); + + // Evaluate the denominator polynomial of the rational interpolant. + Packet8d qx = p8d_cephes_exp_q0; + qx = pmadd(qx, x2, p8d_cephes_exp_q1); + qx = pmadd(qx, x2, p8d_cephes_exp_q2); + qx = pmadd(qx, x2, p8d_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 = _mm512_div_pd(px, psub(qx, px)); + x = pmadd(p8d_2, x, p8d_1); + + // Build e=2^n. + const Packet8d e = _mm512_castsi512_pd(_mm512_slli_epi64( + _mm512_add_epi64(_mm512_cvtpd_epi64(n), _mm512_set1_epi64(1023)), 52)); + + // 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, e), _x); + }*/ + +// Functions for sqrt. +// The EIGEN_FAST_MATH version uses the _mm_rsqrt_ps approximation and one step +// of Newton's method, at a cost of 1-2 bits of precision as opposed to the +// exact solution. The main advantage of this approach is not just speed, but +// also the fact that it can be inlined and pipelined with other computations, +// further reducing its effective latency. +#if EIGEN_FAST_MATH +template <> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet16f +psqrt<Packet16f>(const Packet16f& _x) { + _EIGEN_DECLARE_CONST_Packet16f(one_point_five, 1.5f); + _EIGEN_DECLARE_CONST_Packet16f(minus_half, -0.5f); + _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(flt_min, 0x00800000); + + Packet16f neg_half = pmul(_x, p16f_minus_half); + + // select only the inverse sqrt of positive normal inputs (denormals are + // flushed to zero and cause infs as well). + __mmask16 non_zero_mask = _mm512_cmp_ps_mask(_x, p16f_flt_min, _CMP_GE_OQ); + Packet16f x = _mm512_mask_blend_ps(non_zero_mask, _mm512_rsqrt14_ps(_x), + _mm512_setzero_ps()); + + // Do a single step of Newton's iteration. + x = pmul(x, pmadd(neg_half, pmul(x, x), p16f_one_point_five)); + + // Multiply the original _x by it's reciprocal square root to extract the + // square root. + return pmul(_x, x); +} + +template <> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8d +psqrt<Packet8d>(const Packet8d& _x) { + _EIGEN_DECLARE_CONST_Packet8d(one_point_five, 1.5); + _EIGEN_DECLARE_CONST_Packet8d(minus_half, -0.5); + _EIGEN_DECLARE_CONST_Packet8d_FROM_INT64(dbl_min, 0x0010000000000000LL); + + Packet8d neg_half = pmul(_x, p8d_minus_half); + + // select only the inverse sqrt of positive normal inputs (denormals are + // flushed to zero and cause infs as well). + __mmask8 non_zero_mask = _mm512_cmp_pd_mask(_x, p8d_dbl_min, _CMP_GE_OQ); + Packet8d x = _mm512_mask_blend_pd(non_zero_mask, _mm512_rsqrt14_pd(_x), + _mm512_setzero_pd()); + + // Do a first step of Newton's iteration. + x = pmul(x, pmadd(neg_half, pmul(x, x), p8d_one_point_five)); + + // Do a second step of Newton's iteration. + x = pmul(x, pmadd(neg_half, pmul(x, x), p8d_one_point_five)); + + // Multiply the original _x by it's reciprocal square root to extract the + // square root. + return pmul(_x, x); +} +#else +template <> +EIGEN_STRONG_INLINE Packet16f psqrt<Packet16f>(const Packet16f& x) { + return _mm512_sqrt_ps(x); +} +template <> +EIGEN_STRONG_INLINE Packet8d psqrt<Packet8d>(const Packet8d& x) { + return _mm512_sqrt_pd(x); +} +#endif + +// Functions for rsqrt. +// Almost identical to the sqrt routine, just leave out the last multiplication +// and fill in NaN/Inf where needed. Note that this function only exists as an +// iterative version for doubles since there is no instruction for diretly +// computing the reciprocal square root in AVX-512. +#ifdef EIGEN_FAST_MATH +template <> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet16f +prsqrt<Packet16f>(const Packet16f& _x) { + _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(inf, 0x7f800000); + _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(nan, 0x7fc00000); + _EIGEN_DECLARE_CONST_Packet16f(one_point_five, 1.5f); + _EIGEN_DECLARE_CONST_Packet16f(minus_half, -0.5f); + _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(flt_min, 0x00800000); + + Packet16f neg_half = pmul(_x, p16f_minus_half); + + // select only the inverse sqrt of positive normal inputs (denormals are + // flushed to zero and cause infs as well). + __mmask16 le_zero_mask = _mm512_cmp_ps_mask(_x, p16f_flt_min, _CMP_LT_OQ); + Packet16f x = _mm512_mask_blend_ps(le_zero_mask, _mm512_setzero_ps(), + _mm512_rsqrt14_ps(_x)); + + // Fill in NaNs and Infs for the negative/zero entries. + __mmask16 neg_mask = _mm512_cmp_ps_mask(_x, _mm512_setzero_ps(), _CMP_LT_OQ); + Packet16f infs_and_nans = _mm512_mask_blend_ps( + neg_mask, p16f_nan, + _mm512_mask_blend_ps(le_zero_mask, p16f_inf, _mm512_setzero_ps())); + + // Do a single step of Newton's iteration. + x = pmul(x, pmadd(neg_half, pmul(x, x), p16f_one_point_five)); + + // Insert NaNs and Infs in all the right places. + return _mm512_mask_blend_ps(le_zero_mask, infs_and_nans, x); +} + +template <> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8d +prsqrt<Packet8d>(const Packet8d& _x) { + _EIGEN_DECLARE_CONST_Packet8d_FROM_INT64(inf, 0x7ff0000000000000LL); + _EIGEN_DECLARE_CONST_Packet8d_FROM_INT64(nan, 0x7ff1000000000000LL); + _EIGEN_DECLARE_CONST_Packet8d(one_point_five, 1.5); + _EIGEN_DECLARE_CONST_Packet8d(minus_half, -0.5); + _EIGEN_DECLARE_CONST_Packet8d_FROM_INT64(dbl_min, 0x0010000000000000LL); + + Packet8d neg_half = pmul(_x, p8d_minus_half); + + // select only the inverse sqrt of positive normal inputs (denormals are + // flushed to zero and cause infs as well). + __mmask8 le_zero_mask = _mm512_cmp_pd_mask(_x, p8d_dbl_min, _CMP_LT_OQ); + Packet8d x = _mm512_mask_blend_pd(le_zero_mask, _mm512_setzero_pd(), + _mm512_rsqrt14_pd(_x)); + + // Fill in NaNs and Infs for the negative/zero entries. + __mmask8 neg_mask = _mm512_cmp_pd_mask(_x, _mm512_setzero_pd(), _CMP_LT_OQ); + Packet8d infs_and_nans = _mm512_mask_blend_pd( + neg_mask, p8d_nan, + _mm512_mask_blend_pd(le_zero_mask, p8d_inf, _mm512_setzero_pd())); + + // Do a first step of Newton's iteration. + x = pmul(x, pmadd(neg_half, pmul(x, x), p8d_one_point_five)); + + // Do a second step of Newton's iteration. + x = pmul(x, pmadd(neg_half, pmul(x, x), p8d_one_point_five)); + + // Insert NaNs and Infs in all the right places. + return _mm512_mask_blend_pd(le_zero_mask, infs_and_nans, x); +} +#else +template <> +EIGEN_STRONG_INLINE Packet16f prsqrt<Packet16f>(const Packet16f& x) { + return _mm512_rsqrt28_ps(x); +} +#endif +#endif + +} // end namespace internal + +} // end namespace Eigen + +#endif // THIRD_PARTY_EIGEN3_EIGEN_SRC_CORE_ARCH_AVX512_MATHFUNCTIONS_H_ diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AVX512/PacketMath.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AVX512/PacketMath.h new file mode 100644 index 000000000..f6500a16e --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AVX512/PacketMath.h @@ -0,0 +1,1316 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2016 Benoit Steiner (benoit.steiner.goog@gmail.com) +// +// 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_PACKET_MATH_AVX512_H +#define EIGEN_PACKET_MATH_AVX512_H + +namespace Eigen { + +namespace internal { + +#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD +#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8 +#endif + +#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS +#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*)) +#endif + +#ifdef __FMA__ +#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD +#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD +#endif +#endif + +typedef __m512 Packet16f; +typedef __m512i Packet16i; +typedef __m512d Packet8d; + +template <> +struct is_arithmetic<__m512> { + enum { value = true }; +}; +template <> +struct is_arithmetic<__m512i> { + enum { value = true }; +}; +template <> +struct is_arithmetic<__m512d> { + enum { value = true }; +}; + +template<> struct packet_traits<float> : default_packet_traits +{ + typedef Packet16f type; + typedef Packet8f half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size = 16, + HasHalfPacket = 1, +#if EIGEN_GNUC_AT_LEAST(5, 3) +#ifdef EIGEN_VECTORIZE_AVX512DQ + HasLog = 1, +#endif + HasExp = 1, + HasSqrt = 1, + HasRsqrt = 1, +#endif + HasDiv = 1 + }; + }; +template<> struct packet_traits<double> : default_packet_traits +{ + typedef Packet8d type; + typedef Packet4d half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size = 8, + HasHalfPacket = 1, +#if EIGEN_GNUC_AT_LEAST(5, 3) + HasSqrt = 1, + HasRsqrt = EIGEN_FAST_MATH, +#endif + HasDiv = 1 + }; +}; + +/* TODO Implement AVX512 for integers +template<> struct packet_traits<int> : default_packet_traits +{ + typedef Packet16i type; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=8 + }; +}; +*/ + +template <> +struct unpacket_traits<Packet16f> { + typedef float type; + typedef Packet8f half; + enum { size = 16, alignment=Aligned64 }; +}; +template <> +struct unpacket_traits<Packet8d> { + typedef double type; + typedef Packet4d half; + enum { size = 8, alignment=Aligned64 }; +}; +template <> +struct unpacket_traits<Packet16i> { + typedef int type; + typedef Packet8i half; + enum { size = 16, alignment=Aligned64 }; +}; + +template <> +EIGEN_STRONG_INLINE Packet16f pset1<Packet16f>(const float& from) { + return _mm512_set1_ps(from); +} +template <> +EIGEN_STRONG_INLINE Packet8d pset1<Packet8d>(const double& from) { + return _mm512_set1_pd(from); +} +template <> +EIGEN_STRONG_INLINE Packet16i pset1<Packet16i>(const int& from) { + return _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)); +} + +template <> +EIGEN_STRONG_INLINE Packet16f plset<Packet16f>(const float& a) { + return _mm512_add_ps( + _mm512_set1_ps(a), + _mm512_set_ps(15.0f, 14.0f, 13.0f, 12.0f, 11.0f, 10.0f, 9.0f, 8.0f, 7.0f, 6.0f, 5.0f, + 4.0f, 3.0f, 2.0f, 1.0f, 0.0f)); +} +template <> +EIGEN_STRONG_INLINE Packet8d plset<Packet8d>(const double& a) { + return _mm512_add_pd(_mm512_set1_pd(a), + _mm512_set_pd(7.0, 6.0, 5.0, 4.0, 3.0, 2.0, 1.0, 0.0)); +} + +template <> +EIGEN_STRONG_INLINE Packet16f padd<Packet16f>(const Packet16f& a, + const Packet16f& b) { + return _mm512_add_ps(a, b); +} +template <> +EIGEN_STRONG_INLINE Packet8d padd<Packet8d>(const Packet8d& a, + const Packet8d& b) { + return _mm512_add_pd(a, b); +} + +template <> +EIGEN_STRONG_INLINE Packet16f psub<Packet16f>(const Packet16f& a, + const Packet16f& b) { + return _mm512_sub_ps(a, b); +} +template <> +EIGEN_STRONG_INLINE Packet8d psub<Packet8d>(const Packet8d& a, + const Packet8d& b) { + return _mm512_sub_pd(a, b); +} + +template <> +EIGEN_STRONG_INLINE Packet16f pnegate(const Packet16f& a) { + return _mm512_sub_ps(_mm512_set1_ps(0.0), a); +} +template <> +EIGEN_STRONG_INLINE Packet8d pnegate(const Packet8d& a) { + return _mm512_sub_pd(_mm512_set1_pd(0.0), a); +} + +template <> +EIGEN_STRONG_INLINE Packet16f pconj(const Packet16f& a) { + return a; +} +template <> +EIGEN_STRONG_INLINE Packet8d pconj(const Packet8d& a) { + return a; +} +template <> +EIGEN_STRONG_INLINE Packet16i pconj(const Packet16i& a) { + return a; +} + +template <> +EIGEN_STRONG_INLINE Packet16f pmul<Packet16f>(const Packet16f& a, + const Packet16f& b) { + return _mm512_mul_ps(a, b); +} +template <> +EIGEN_STRONG_INLINE Packet8d pmul<Packet8d>(const Packet8d& a, + const Packet8d& b) { + return _mm512_mul_pd(a, b); +} + +template <> +EIGEN_STRONG_INLINE Packet16f pdiv<Packet16f>(const Packet16f& a, + const Packet16f& b) { + return _mm512_div_ps(a, b); +} +template <> +EIGEN_STRONG_INLINE Packet8d pdiv<Packet8d>(const Packet8d& a, + const Packet8d& b) { + return _mm512_div_pd(a, b); +} + +#ifdef __FMA__ +template <> +EIGEN_STRONG_INLINE Packet16f pmadd(const Packet16f& a, const Packet16f& b, + const Packet16f& c) { + return _mm512_fmadd_ps(a, b, c); +} +template <> +EIGEN_STRONG_INLINE Packet8d pmadd(const Packet8d& a, const Packet8d& b, + const Packet8d& c) { + return _mm512_fmadd_pd(a, b, c); +} +#endif + +template <> +EIGEN_STRONG_INLINE Packet16f pmin<Packet16f>(const Packet16f& a, + const Packet16f& b) { + return _mm512_min_ps(a, b); +} +template <> +EIGEN_STRONG_INLINE Packet8d pmin<Packet8d>(const Packet8d& a, + const Packet8d& b) { + return _mm512_min_pd(a, b); +} + +template <> +EIGEN_STRONG_INLINE Packet16f pmax<Packet16f>(const Packet16f& a, + const Packet16f& b) { + return _mm512_max_ps(a, b); +} +template <> +EIGEN_STRONG_INLINE Packet8d pmax<Packet8d>(const Packet8d& a, + const Packet8d& b) { + return _mm512_max_pd(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 + 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); + + 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); + + 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); + + 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); + + return res; +#endif +} +template <> +EIGEN_STRONG_INLINE Packet8d pand<Packet8d>(const Packet8d& a, + const Packet8d& b) { +#ifdef EIGEN_VECTORIZE_AVX512DQ + return _mm512_and_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_and_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_and_pd(lane1_a, lane1_b), 1); + + return res; +#endif +} +template <> +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; +#endif +} + +template <> +EIGEN_STRONG_INLINE Packet8d por<Packet8d>(const Packet8d& a, + const Packet8d& b) { +#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; +#endif +} + +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; +#endif +} +template <> +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; +#endif +} + +template <> +EIGEN_STRONG_INLINE Packet16f pandnot<Packet16f>(const Packet16f& a, + const Packet16f& b) { +#ifdef EIGEN_VECTORIZE_AVX512DQ + return _mm512_andnot_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_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; +#endif +} +template <> +EIGEN_STRONG_INLINE Packet8d pandnot<Packet8d>(const Packet8d& a, + const Packet8d& b) { +#ifdef EIGEN_VECTORIZE_AVX512DQ + return _mm512_andnot_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_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; +#endif +} + +template <> +EIGEN_STRONG_INLINE Packet16f pload<Packet16f>(const float* from) { + EIGEN_DEBUG_ALIGNED_LOAD return _mm512_load_ps(from); +} +template <> +EIGEN_STRONG_INLINE Packet8d pload<Packet8d>(const double* from) { + EIGEN_DEBUG_ALIGNED_LOAD return _mm512_load_pd(from); +} +template <> +EIGEN_STRONG_INLINE Packet16i pload<Packet16i>(const int* from) { + EIGEN_DEBUG_ALIGNED_LOAD return _mm512_load_si512( + reinterpret_cast<const __m512i*>(from)); +} + +template <> +EIGEN_STRONG_INLINE Packet16f ploadu<Packet16f>(const float* from) { + EIGEN_DEBUG_UNALIGNED_LOAD return _mm512_loadu_ps(from); +} +template <> +EIGEN_STRONG_INLINE Packet8d ploadu<Packet8d>(const double* from) { + EIGEN_DEBUG_UNALIGNED_LOAD return _mm512_loadu_pd(from); +} +template <> +EIGEN_STRONG_INLINE Packet16i ploadu<Packet16i>(const int* from) { + EIGEN_DEBUG_UNALIGNED_LOAD return _mm512_loadu_si512( + reinterpret_cast<const __m512i*>(from)); +} + +// Loads 8 floats from memory a returns the packet +// {a0, a0 a1, a1, a2, a2, a3, a3, a4, a4, a5, a5, a6, a6, a7, a7} +template <> +EIGEN_STRONG_INLINE Packet16f ploaddup<Packet16f>(const float* from) { + Packet8f lane0 = _mm256_broadcast_ps((const __m128*)(const void*)from); + // mimic an "inplace" permutation of the lower 128bits using a blend + lane0 = _mm256_blend_ps( + lane0, _mm256_castps128_ps256(_mm_permute_ps( + _mm256_castps256_ps128(lane0), _MM_SHUFFLE(1, 0, 1, 0))), + 15); + // then we can perform a consistent permutation on the global register to get + // everything in shape: + lane0 = _mm256_permute_ps(lane0, _MM_SHUFFLE(3, 3, 2, 2)); + + Packet8f lane1 = _mm256_broadcast_ps((const __m128*)(const void*)(from + 4)); + // mimic an "inplace" permutation of the lower 128bits using a blend + lane1 = _mm256_blend_ps( + lane1, _mm256_castps128_ps256(_mm_permute_ps( + _mm256_castps256_ps128(lane1), _MM_SHUFFLE(1, 0, 1, 0))), + 15); + // then we can perform a consistent permutation on the global register to get + // everything in shape: + lane1 = _mm256_permute_ps(lane1, _MM_SHUFFLE(3, 3, 2, 2)); + +#ifdef EIGEN_VECTORIZE_AVX512DQ + Packet16f res = _mm512_undefined_ps(); + return _mm512_insertf32x8(res, lane0, 0); + return _mm512_insertf32x8(res, lane1, 1); + return res; +#else + Packet16f res = _mm512_undefined_ps(); + res = _mm512_insertf32x4(res, _mm256_extractf128_ps(lane0, 0), 0); + res = _mm512_insertf32x4(res, _mm256_extractf128_ps(lane0, 1), 1); + res = _mm512_insertf32x4(res, _mm256_extractf128_ps(lane1, 0), 2); + res = _mm512_insertf32x4(res, _mm256_extractf128_ps(lane1, 1), 3); + return res; +#endif +} +// Loads 4 doubles from memory a returns the packet {a0, a0 a1, a1, a2, a2, a3, +// a3} +template <> +EIGEN_STRONG_INLINE Packet8d ploaddup<Packet8d>(const double* from) { + Packet4d lane0 = _mm256_broadcast_pd((const __m128d*)(const void*)from); + lane0 = _mm256_permute_pd(lane0, 3 << 2); + + Packet4d lane1 = _mm256_broadcast_pd((const __m128d*)(const void*)(from + 2)); + lane1 = _mm256_permute_pd(lane1, 3 << 2); + + Packet8d res = _mm512_undefined_pd(); + res = _mm512_insertf64x4(res, lane0, 0); + return _mm512_insertf64x4(res, lane1, 1); +} + +// Loads 4 floats from memory a returns the packet +// {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; +} +// 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) { + Packet8d tmp = _mm512_undefined_pd(); + Packet2d tmp0 = _mm_load_pd1(from); + Packet2d tmp1 = _mm_load_pd1(from + 1); + Packet4d lane0 = _mm256_broadcastsd_pd(tmp0); + Packet4d lane1 = _mm256_broadcastsd_pd(tmp1); + tmp = _mm512_insertf64x4(tmp, lane0, 0); + return _mm512_insertf64x4(tmp, lane1, 1); +} + +template <> +EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet16f& from) { + EIGEN_DEBUG_ALIGNED_STORE _mm512_store_ps(to, from); +} +template <> +EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet8d& from) { + EIGEN_DEBUG_ALIGNED_STORE _mm512_store_pd(to, from); +} +template <> +EIGEN_STRONG_INLINE void pstore<int>(int* to, const Packet16i& from) { + EIGEN_DEBUG_ALIGNED_STORE _mm512_storeu_si512(reinterpret_cast<__m512i*>(to), + from); +} + +template <> +EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet16f& from) { + EIGEN_DEBUG_UNALIGNED_STORE _mm512_storeu_ps(to, from); +} +template <> +EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet8d& from) { + EIGEN_DEBUG_UNALIGNED_STORE _mm512_storeu_pd(to, from); +} +template <> +EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet16i& from) { + EIGEN_DEBUG_UNALIGNED_STORE _mm512_storeu_si512( + reinterpret_cast<__m512i*>(to), from); +} + +template <> +EIGEN_DEVICE_FUNC inline Packet16f pgather<float, Packet16f>(const float* from, + Index stride) { + Packet16i stride_vector = _mm512_set1_epi32(stride); + Packet16i stride_multiplier = + _mm512_set_epi32(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0); + Packet16i indices = _mm512_mullo_epi32(stride_vector, stride_multiplier); + + return _mm512_i32gather_ps(indices, from, 4); +} +template <> +EIGEN_DEVICE_FUNC inline Packet8d pgather<double, Packet8d>(const double* from, + Index stride) { + Packet8i stride_vector = _mm256_set1_epi32(stride); + Packet8i stride_multiplier = _mm256_set_epi32(7, 6, 5, 4, 3, 2, 1, 0); + Packet8i indices = _mm256_mullo_epi32(stride_vector, stride_multiplier); + + return _mm512_i32gather_pd(indices, from, 8); +} + +template <> +EIGEN_DEVICE_FUNC inline void pscatter<float, Packet16f>(float* to, + const Packet16f& from, + Index stride) { + Packet16i stride_vector = _mm512_set1_epi32(stride); + Packet16i stride_multiplier = + _mm512_set_epi32(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0); + Packet16i indices = _mm512_mullo_epi32(stride_vector, stride_multiplier); + _mm512_i32scatter_ps(to, indices, from, 4); +} +template <> +EIGEN_DEVICE_FUNC inline void pscatter<double, Packet8d>(double* to, + const Packet8d& from, + Index stride) { + Packet8i stride_vector = _mm256_set1_epi32(stride); + Packet8i stride_multiplier = _mm256_set_epi32(7, 6, 5, 4, 3, 2, 1, 0); + Packet8i indices = _mm256_mullo_epi32(stride_vector, stride_multiplier); + _mm512_i32scatter_pd(to, indices, from, 8); +} + +template <> +EIGEN_STRONG_INLINE void pstore1<Packet16f>(float* to, const float& a) { + Packet16f pa = pset1<Packet16f>(a); + pstore(to, pa); +} +template <> +EIGEN_STRONG_INLINE void pstore1<Packet8d>(double* to, const double& a) { + Packet8d pa = pset1<Packet8d>(a); + pstore(to, pa); +} +template <> +EIGEN_STRONG_INLINE void pstore1<Packet16i>(int* to, const int& a) { + Packet16i pa = pset1<Packet16i>(a); + pstore(to, pa); +} + +template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } +template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } +template<> EIGEN_STRONG_INLINE void prefetch<int>(const int* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } + +template <> +EIGEN_STRONG_INLINE float pfirst<Packet16f>(const Packet16f& a) { + return _mm_cvtss_f32(_mm512_extractf32x4_ps(a, 0)); +} +template <> +EIGEN_STRONG_INLINE double pfirst<Packet8d>(const Packet8d& a) { + return _mm_cvtsd_f64(_mm256_extractf128_pd(_mm512_extractf64x4_pd(a, 0), 0)); +} +template <> +EIGEN_STRONG_INLINE int pfirst<Packet16i>(const Packet16i& a) { + return _mm_extract_epi32(_mm512_extracti32x4_epi32(a, 0), 0); +} + +template<> EIGEN_STRONG_INLINE Packet16f preverse(const Packet16f& a) +{ + return _mm512_permutexvar_ps(_mm512_set_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15), a); +} + +template<> EIGEN_STRONG_INLINE Packet8d preverse(const Packet8d& a) +{ + return _mm512_permutexvar_pd(_mm512_set_epi32(0, 0, 0, 1, 0, 2, 0, 3, 0, 4, 0, 5, 0, 6, 0, 7), a); +} + +template<> EIGEN_STRONG_INLINE Packet16f pabs(const Packet16f& a) +{ + // _mm512_abs_ps intrinsic not found, so hack around it + return (__m512)_mm512_and_si512((__m512i)a, _mm512_set1_epi32(0x7fffffff)); +} +template <> +EIGEN_STRONG_INLINE Packet8d pabs(const Packet8d& a) { + // _mm512_abs_ps intrinsic not found, so hack around it + return (__m512d)_mm512_and_si512((__m512i)a, + _mm512_set1_epi64(0x7fffffffffffffff)); +} + +#ifdef EIGEN_VECTORIZE_AVX512DQ +// AVX512F does not define _mm512_extractf32x8_ps to extract _m256 from _m512 +#define EIGEN_EXTRACT_8f_FROM_16f(INPUT, OUTPUT) \ + __m256 OUTPUT##_0 = _mm512_extractf32x8_ps(INPUT, 0) __m256 OUTPUT##_1 = \ + _mm512_extractf32x8_ps(INPUT, 1) +#else +#define EIGEN_EXTRACT_8f_FROM_16f(INPUT, OUTPUT) \ + __m256 OUTPUT##_0 = _mm256_insertf128_ps( \ + _mm256_castps128_ps256(_mm512_extractf32x4_ps(INPUT, 0)), \ + _mm512_extractf32x4_ps(INPUT, 1), 1); \ + __m256 OUTPUT##_1 = _mm256_insertf128_ps( \ + _mm256_castps128_ps256(_mm512_extractf32x4_ps(INPUT, 2)), \ + _mm512_extractf32x4_ps(INPUT, 3), 1); +#endif + +#ifdef EIGEN_VECTORIZE_AVX512DQ +#define EIGEN_INSERT_8f_INTO_16f(OUTPUT, INPUTA, INPUTB) \ + OUTPUT = _mm512_insertf32x8(OUTPUT, INPUTA, 0); \ + OUTPUT = _mm512_insertf32x8(OUTPUT, INPUTB, 1); +#else +#define EIGEN_INSERT_8f_INTO_16f(OUTPUT, INPUTA, INPUTB) \ + OUTPUT = _mm512_insertf32x4(OUTPUT, _mm256_extractf128_ps(INPUTA, 0), 0); \ + OUTPUT = _mm512_insertf32x4(OUTPUT, _mm256_extractf128_ps(INPUTA, 1), 1); \ + OUTPUT = _mm512_insertf32x4(OUTPUT, _mm256_extractf128_ps(INPUTB, 0), 2); \ + OUTPUT = _mm512_insertf32x4(OUTPUT, _mm256_extractf128_ps(INPUTB, 1), 3); +#endif +template<> EIGEN_STRONG_INLINE Packet16f preduxp<Packet16f>(const Packet16f* +vecs) +{ + EIGEN_EXTRACT_8f_FROM_16f(vecs[0], vecs0); + EIGEN_EXTRACT_8f_FROM_16f(vecs[1], vecs1); + EIGEN_EXTRACT_8f_FROM_16f(vecs[2], vecs2); + EIGEN_EXTRACT_8f_FROM_16f(vecs[3], vecs3); + EIGEN_EXTRACT_8f_FROM_16f(vecs[4], vecs4); + EIGEN_EXTRACT_8f_FROM_16f(vecs[5], vecs5); + EIGEN_EXTRACT_8f_FROM_16f(vecs[6], vecs6); + EIGEN_EXTRACT_8f_FROM_16f(vecs[7], vecs7); + EIGEN_EXTRACT_8f_FROM_16f(vecs[8], vecs8); + EIGEN_EXTRACT_8f_FROM_16f(vecs[9], vecs9); + EIGEN_EXTRACT_8f_FROM_16f(vecs[10], vecs10); + EIGEN_EXTRACT_8f_FROM_16f(vecs[11], vecs11); + EIGEN_EXTRACT_8f_FROM_16f(vecs[12], vecs12); + EIGEN_EXTRACT_8f_FROM_16f(vecs[13], vecs13); + EIGEN_EXTRACT_8f_FROM_16f(vecs[14], vecs14); + EIGEN_EXTRACT_8f_FROM_16f(vecs[15], vecs15); + + __m256 hsum1 = _mm256_hadd_ps(vecs0_0, vecs1_0); + __m256 hsum2 = _mm256_hadd_ps(vecs2_0, vecs3_0); + __m256 hsum3 = _mm256_hadd_ps(vecs4_0, vecs5_0); + __m256 hsum4 = _mm256_hadd_ps(vecs6_0, vecs7_0); + + __m256 hsum5 = _mm256_hadd_ps(hsum1, hsum1); + __m256 hsum6 = _mm256_hadd_ps(hsum2, hsum2); + __m256 hsum7 = _mm256_hadd_ps(hsum3, hsum3); + __m256 hsum8 = _mm256_hadd_ps(hsum4, hsum4); + + __m256 perm1 = _mm256_permute2f128_ps(hsum5, hsum5, 0x23); + __m256 perm2 = _mm256_permute2f128_ps(hsum6, hsum6, 0x23); + __m256 perm3 = _mm256_permute2f128_ps(hsum7, hsum7, 0x23); + __m256 perm4 = _mm256_permute2f128_ps(hsum8, hsum8, 0x23); + + __m256 sum1 = _mm256_add_ps(perm1, hsum5); + __m256 sum2 = _mm256_add_ps(perm2, hsum6); + __m256 sum3 = _mm256_add_ps(perm3, hsum7); + __m256 sum4 = _mm256_add_ps(perm4, hsum8); + + __m256 blend1 = _mm256_blend_ps(sum1, sum2, 0xcc); + __m256 blend2 = _mm256_blend_ps(sum3, sum4, 0xcc); + + __m256 final = _mm256_blend_ps(blend1, blend2, 0xf0); + + hsum1 = _mm256_hadd_ps(vecs0_1, vecs1_1); + hsum2 = _mm256_hadd_ps(vecs2_1, vecs3_1); + hsum3 = _mm256_hadd_ps(vecs4_1, vecs5_1); + hsum4 = _mm256_hadd_ps(vecs6_1, vecs7_1); + + hsum5 = _mm256_hadd_ps(hsum1, hsum1); + hsum6 = _mm256_hadd_ps(hsum2, hsum2); + hsum7 = _mm256_hadd_ps(hsum3, hsum3); + hsum8 = _mm256_hadd_ps(hsum4, hsum4); + + perm1 = _mm256_permute2f128_ps(hsum5, hsum5, 0x23); + perm2 = _mm256_permute2f128_ps(hsum6, hsum6, 0x23); + perm3 = _mm256_permute2f128_ps(hsum7, hsum7, 0x23); + perm4 = _mm256_permute2f128_ps(hsum8, hsum8, 0x23); + + sum1 = _mm256_add_ps(perm1, hsum5); + sum2 = _mm256_add_ps(perm2, hsum6); + sum3 = _mm256_add_ps(perm3, hsum7); + sum4 = _mm256_add_ps(perm4, hsum8); + + blend1 = _mm256_blend_ps(sum1, sum2, 0xcc); + blend2 = _mm256_blend_ps(sum3, sum4, 0xcc); + + final = padd(final, _mm256_blend_ps(blend1, blend2, 0xf0)); + + hsum1 = _mm256_hadd_ps(vecs8_0, vecs9_0); + hsum2 = _mm256_hadd_ps(vecs10_0, vecs11_0); + hsum3 = _mm256_hadd_ps(vecs12_0, vecs13_0); + hsum4 = _mm256_hadd_ps(vecs14_0, vecs15_0); + + hsum5 = _mm256_hadd_ps(hsum1, hsum1); + hsum6 = _mm256_hadd_ps(hsum2, hsum2); + hsum7 = _mm256_hadd_ps(hsum3, hsum3); + hsum8 = _mm256_hadd_ps(hsum4, hsum4); + + perm1 = _mm256_permute2f128_ps(hsum5, hsum5, 0x23); + perm2 = _mm256_permute2f128_ps(hsum6, hsum6, 0x23); + perm3 = _mm256_permute2f128_ps(hsum7, hsum7, 0x23); + perm4 = _mm256_permute2f128_ps(hsum8, hsum8, 0x23); + + sum1 = _mm256_add_ps(perm1, hsum5); + sum2 = _mm256_add_ps(perm2, hsum6); + sum3 = _mm256_add_ps(perm3, hsum7); + sum4 = _mm256_add_ps(perm4, hsum8); + + blend1 = _mm256_blend_ps(sum1, sum2, 0xcc); + blend2 = _mm256_blend_ps(sum3, sum4, 0xcc); + + __m256 final_1 = _mm256_blend_ps(blend1, blend2, 0xf0); + + hsum1 = _mm256_hadd_ps(vecs8_1, vecs9_1); + hsum2 = _mm256_hadd_ps(vecs10_1, vecs11_1); + hsum3 = _mm256_hadd_ps(vecs12_1, vecs13_1); + hsum4 = _mm256_hadd_ps(vecs14_1, vecs15_1); + + hsum5 = _mm256_hadd_ps(hsum1, hsum1); + hsum6 = _mm256_hadd_ps(hsum2, hsum2); + hsum7 = _mm256_hadd_ps(hsum3, hsum3); + hsum8 = _mm256_hadd_ps(hsum4, hsum4); + + perm1 = _mm256_permute2f128_ps(hsum5, hsum5, 0x23); + perm2 = _mm256_permute2f128_ps(hsum6, hsum6, 0x23); + perm3 = _mm256_permute2f128_ps(hsum7, hsum7, 0x23); + perm4 = _mm256_permute2f128_ps(hsum8, hsum8, 0x23); + + sum1 = _mm256_add_ps(perm1, hsum5); + sum2 = _mm256_add_ps(perm2, hsum6); + sum3 = _mm256_add_ps(perm3, hsum7); + sum4 = _mm256_add_ps(perm4, hsum8); + + blend1 = _mm256_blend_ps(sum1, sum2, 0xcc); + blend2 = _mm256_blend_ps(sum3, sum4, 0xcc); + + final_1 = padd(final_1, _mm256_blend_ps(blend1, blend2, 0xf0)); + + __m512 final_output; + + EIGEN_INSERT_8f_INTO_16f(final_output, final, final_1); + return final_output; +} + +template<> EIGEN_STRONG_INLINE Packet8d preduxp<Packet8d>(const Packet8d* vecs) +{ + Packet4d vecs0_0 = _mm512_extractf64x4_pd(vecs[0], 0); + Packet4d vecs0_1 = _mm512_extractf64x4_pd(vecs[0], 1); + + Packet4d vecs1_0 = _mm512_extractf64x4_pd(vecs[1], 0); + Packet4d vecs1_1 = _mm512_extractf64x4_pd(vecs[1], 1); + + Packet4d vecs2_0 = _mm512_extractf64x4_pd(vecs[2], 0); + Packet4d vecs2_1 = _mm512_extractf64x4_pd(vecs[2], 1); + + Packet4d vecs3_0 = _mm512_extractf64x4_pd(vecs[3], 0); + Packet4d vecs3_1 = _mm512_extractf64x4_pd(vecs[3], 1); + + Packet4d vecs4_0 = _mm512_extractf64x4_pd(vecs[4], 0); + Packet4d vecs4_1 = _mm512_extractf64x4_pd(vecs[4], 1); + + Packet4d vecs5_0 = _mm512_extractf64x4_pd(vecs[5], 0); + Packet4d vecs5_1 = _mm512_extractf64x4_pd(vecs[5], 1); + + Packet4d vecs6_0 = _mm512_extractf64x4_pd(vecs[6], 0); + Packet4d vecs6_1 = _mm512_extractf64x4_pd(vecs[6], 1); + + Packet4d vecs7_0 = _mm512_extractf64x4_pd(vecs[7], 0); + Packet4d vecs7_1 = _mm512_extractf64x4_pd(vecs[7], 1); + + Packet4d tmp0, tmp1; + + tmp0 = _mm256_hadd_pd(vecs0_0, vecs1_0); + tmp0 = _mm256_add_pd(tmp0, _mm256_permute2f128_pd(tmp0, tmp0, 1)); + + tmp1 = _mm256_hadd_pd(vecs2_0, vecs3_0); + tmp1 = _mm256_add_pd(tmp1, _mm256_permute2f128_pd(tmp1, tmp1, 1)); + + __m256d final_0 = _mm256_blend_pd(tmp0, tmp1, 0xC); + + tmp0 = _mm256_hadd_pd(vecs0_1, vecs1_1); + tmp0 = _mm256_add_pd(tmp0, _mm256_permute2f128_pd(tmp0, tmp0, 1)); + + tmp1 = _mm256_hadd_pd(vecs2_1, vecs3_1); + tmp1 = _mm256_add_pd(tmp1, _mm256_permute2f128_pd(tmp1, tmp1, 1)); + + final_0 = padd(final_0, _mm256_blend_pd(tmp0, tmp1, 0xC)); + + tmp0 = _mm256_hadd_pd(vecs4_0, vecs5_0); + tmp0 = _mm256_add_pd(tmp0, _mm256_permute2f128_pd(tmp0, tmp0, 1)); + + tmp1 = _mm256_hadd_pd(vecs6_0, vecs7_0); + tmp1 = _mm256_add_pd(tmp1, _mm256_permute2f128_pd(tmp1, tmp1, 1)); + + __m256d final_1 = _mm256_blend_pd(tmp0, tmp1, 0xC); + + tmp0 = _mm256_hadd_pd(vecs4_1, vecs5_1); + tmp0 = _mm256_add_pd(tmp0, _mm256_permute2f128_pd(tmp0, tmp0, 1)); + + tmp1 = _mm256_hadd_pd(vecs6_1, vecs7_1); + tmp1 = _mm256_add_pd(tmp1, _mm256_permute2f128_pd(tmp1, tmp1, 1)); + + final_1 = padd(final_1, _mm256_blend_pd(tmp0, tmp1, 0xC)); + + __m512d final_output = _mm512_insertf64x4(final_output, final_0, 0); + + return _mm512_insertf64x4(final_output, final_1, 1); +} + +template <> +EIGEN_STRONG_INLINE float predux<Packet16f>(const Packet16f& a) { + //#ifdef EIGEN_VECTORIZE_AVX512DQ +#if 0 + Packet8f lane0 = _mm512_extractf32x8_ps(a, 0); + Packet8f lane1 = _mm512_extractf32x8_ps(a, 1); + Packet8f sum = padd(lane0, lane1); + Packet8f tmp0 = _mm256_hadd_ps(sum, _mm256_permute2f128_ps(a, a, 1)); + tmp0 = _mm256_hadd_ps(tmp0, tmp0); + return pfirst(_mm256_hadd_ps(tmp0, tmp0)); +#else + Packet4f lane0 = _mm512_extractf32x4_ps(a, 0); + Packet4f lane1 = _mm512_extractf32x4_ps(a, 1); + Packet4f lane2 = _mm512_extractf32x4_ps(a, 2); + Packet4f lane3 = _mm512_extractf32x4_ps(a, 3); + Packet4f sum = padd(padd(lane0, lane1), padd(lane2, lane3)); + sum = _mm_hadd_ps(sum, sum); + sum = _mm_hadd_ps(sum, _mm_permute_ps(sum, 1)); + return pfirst(sum); +#endif +} +template <> +EIGEN_STRONG_INLINE double predux<Packet8d>(const Packet8d& a) { + Packet4d lane0 = _mm512_extractf64x4_pd(a, 0); + Packet4d lane1 = _mm512_extractf64x4_pd(a, 1); + Packet4d sum = padd(lane0, lane1); + Packet4d tmp0 = _mm256_hadd_pd(sum, _mm256_permute2f128_pd(sum, sum, 1)); + return pfirst(_mm256_hadd_pd(tmp0, tmp0)); +} + +template <> +EIGEN_STRONG_INLINE Packet8f predux_downto4<Packet16f>(const Packet16f& a) { +#ifdef EIGEN_VECTORIZE_AVX512DQ + Packet8f lane0 = _mm512_extractf32x8_ps(a, 0); + Packet8f lane1 = _mm512_extractf32x8_ps(a, 1); + return padd(lane0, lane1); +#else + Packet4f lane0 = _mm512_extractf32x4_ps(a, 0); + Packet4f lane1 = _mm512_extractf32x4_ps(a, 1); + Packet4f lane2 = _mm512_extractf32x4_ps(a, 2); + Packet4f lane3 = _mm512_extractf32x4_ps(a, 3); + Packet4f sum0 = padd(lane0, lane2); + Packet4f sum1 = padd(lane1, lane3); + return _mm256_insertf128_ps(_mm256_castps128_ps256(sum0), sum1, 1); +#endif +} +template <> +EIGEN_STRONG_INLINE Packet4d predux_downto4<Packet8d>(const Packet8d& a) { + Packet4d lane0 = _mm512_extractf64x4_pd(a, 0); + Packet4d lane1 = _mm512_extractf64x4_pd(a, 1); + Packet4d res = padd(lane0, lane1); + return res; +} + +template <> +EIGEN_STRONG_INLINE float predux_mul<Packet16f>(const Packet16f& a) { +//#ifdef EIGEN_VECTORIZE_AVX512DQ +#if 0 + Packet8f lane0 = _mm512_extractf32x8_ps(a, 0); + Packet8f lane1 = _mm512_extractf32x8_ps(a, 1); + Packet8f res = pmul(lane0, lane1); + res = pmul(res, _mm256_permute2f128_ps(res, res, 1)); + res = pmul(res, _mm_permute_ps(res, _MM_SHUFFLE(0, 0, 3, 2))); + return pfirst(pmul(res, _mm_permute_ps(res, _MM_SHUFFLE(0, 0, 0, 1)))); +#else + Packet4f lane0 = _mm512_extractf32x4_ps(a, 0); + Packet4f lane1 = _mm512_extractf32x4_ps(a, 1); + Packet4f lane2 = _mm512_extractf32x4_ps(a, 2); + Packet4f lane3 = _mm512_extractf32x4_ps(a, 3); + Packet4f res = pmul(pmul(lane0, lane1), pmul(lane2, lane3)); + res = pmul(res, _mm_permute_ps(res, _MM_SHUFFLE(0, 0, 3, 2))); + return pfirst(pmul(res, _mm_permute_ps(res, _MM_SHUFFLE(0, 0, 0, 1)))); +#endif +} +template <> +EIGEN_STRONG_INLINE double predux_mul<Packet8d>(const Packet8d& a) { + Packet4d lane0 = _mm512_extractf64x4_pd(a, 0); + Packet4d lane1 = _mm512_extractf64x4_pd(a, 1); + Packet4d res = pmul(lane0, lane1); + res = pmul(res, _mm256_permute2f128_pd(res, res, 1)); + return pfirst(pmul(res, _mm256_shuffle_pd(res, res, 1))); +} + +template <> +EIGEN_STRONG_INLINE float predux_min<Packet16f>(const Packet16f& a) { + Packet4f lane0 = _mm512_extractf32x4_ps(a, 0); + Packet4f lane1 = _mm512_extractf32x4_ps(a, 1); + Packet4f lane2 = _mm512_extractf32x4_ps(a, 2); + Packet4f lane3 = _mm512_extractf32x4_ps(a, 3); + Packet4f res = _mm_min_ps(_mm_min_ps(lane0, lane1), _mm_min_ps(lane2, lane3)); + res = _mm_min_ps(res, _mm_permute_ps(res, _MM_SHUFFLE(0, 0, 3, 2))); + return pfirst(_mm_min_ps(res, _mm_permute_ps(res, _MM_SHUFFLE(0, 0, 0, 1)))); +} +template <> +EIGEN_STRONG_INLINE double predux_min<Packet8d>(const Packet8d& a) { + Packet4d lane0 = _mm512_extractf64x4_pd(a, 0); + Packet4d lane1 = _mm512_extractf64x4_pd(a, 1); + Packet4d res = _mm256_min_pd(lane0, lane1); + res = _mm256_min_pd(res, _mm256_permute2f128_pd(res, res, 1)); + return pfirst(_mm256_min_pd(res, _mm256_shuffle_pd(res, res, 1))); +} + +template <> +EIGEN_STRONG_INLINE float predux_max<Packet16f>(const Packet16f& a) { + Packet4f lane0 = _mm512_extractf32x4_ps(a, 0); + Packet4f lane1 = _mm512_extractf32x4_ps(a, 1); + Packet4f lane2 = _mm512_extractf32x4_ps(a, 2); + Packet4f lane3 = _mm512_extractf32x4_ps(a, 3); + Packet4f res = _mm_max_ps(_mm_max_ps(lane0, lane1), _mm_max_ps(lane2, lane3)); + res = _mm_max_ps(res, _mm_permute_ps(res, _MM_SHUFFLE(0, 0, 3, 2))); + return pfirst(_mm_max_ps(res, _mm_permute_ps(res, _MM_SHUFFLE(0, 0, 0, 1)))); +} +template <> +EIGEN_STRONG_INLINE double predux_max<Packet8d>(const Packet8d& a) { + Packet4d lane0 = _mm512_extractf64x4_pd(a, 0); + Packet4d lane1 = _mm512_extractf64x4_pd(a, 1); + Packet4d res = _mm256_max_pd(lane0, lane1); + res = _mm256_max_pd(res, _mm256_permute2f128_pd(res, res, 1)); + return pfirst(_mm256_max_pd(res, _mm256_shuffle_pd(res, res, 1))); +} + +template <int Offset> +struct palign_impl<Offset, Packet16f> { + static EIGEN_STRONG_INLINE void run(Packet16f& first, + const Packet16f& second) { + if (Offset != 0) { + __m512i first_idx = _mm512_set_epi32( + Offset + 15, Offset + 14, Offset + 13, Offset + 12, Offset + 11, + Offset + 10, Offset + 9, Offset + 8, Offset + 7, Offset + 6, + Offset + 5, Offset + 4, Offset + 3, Offset + 2, Offset + 1, Offset); + + __m512i second_idx = + _mm512_set_epi32(Offset - 1, Offset - 2, Offset - 3, Offset - 4, + Offset - 5, Offset - 6, Offset - 7, Offset - 8, + Offset - 9, Offset - 10, Offset - 11, Offset - 12, + Offset - 13, Offset - 14, Offset - 15, Offset - 16); + + unsigned short mask = 0xFFFF; + mask <<= (16 - Offset); + + first = _mm512_permutexvar_ps(first_idx, first); + Packet16f tmp = _mm512_permutexvar_ps(second_idx, second); + first = _mm512_mask_blend_ps(mask, first, tmp); + } + } +}; +template <int Offset> +struct palign_impl<Offset, Packet8d> { + static EIGEN_STRONG_INLINE void run(Packet8d& first, const Packet8d& second) { + if (Offset != 0) { + __m512i first_idx = _mm512_set_epi32( + 0, Offset + 7, 0, Offset + 6, 0, Offset + 5, 0, Offset + 4, 0, + Offset + 3, 0, Offset + 2, 0, Offset + 1, 0, Offset); + + __m512i second_idx = _mm512_set_epi32( + 0, Offset - 1, 0, Offset - 2, 0, Offset - 3, 0, Offset - 4, 0, + Offset - 5, 0, Offset - 6, 0, Offset - 7, 0, Offset - 8); + + unsigned char mask = 0xFF; + mask <<= (8 - Offset); + + first = _mm512_permutexvar_pd(first_idx, first); + Packet8d tmp = _mm512_permutexvar_pd(second_idx, second); + first = _mm512_mask_blend_pd(mask, first, tmp); + } + } +}; + + +#define PACK_OUTPUT(OUTPUT, INPUT, INDEX, STRIDE) \ + EIGEN_INSERT_8f_INTO_16f(OUTPUT[INDEX], INPUT[INDEX], INPUT[INDEX + STRIDE]); + +EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet16f, 16>& kernel) { + __m512 T0 = _mm512_unpacklo_ps(kernel.packet[0], kernel.packet[1]); + __m512 T1 = _mm512_unpackhi_ps(kernel.packet[0], kernel.packet[1]); + __m512 T2 = _mm512_unpacklo_ps(kernel.packet[2], kernel.packet[3]); + __m512 T3 = _mm512_unpackhi_ps(kernel.packet[2], kernel.packet[3]); + __m512 T4 = _mm512_unpacklo_ps(kernel.packet[4], kernel.packet[5]); + __m512 T5 = _mm512_unpackhi_ps(kernel.packet[4], kernel.packet[5]); + __m512 T6 = _mm512_unpacklo_ps(kernel.packet[6], kernel.packet[7]); + __m512 T7 = _mm512_unpackhi_ps(kernel.packet[6], kernel.packet[7]); + __m512 T8 = _mm512_unpacklo_ps(kernel.packet[8], kernel.packet[9]); + __m512 T9 = _mm512_unpackhi_ps(kernel.packet[8], kernel.packet[9]); + __m512 T10 = _mm512_unpacklo_ps(kernel.packet[10], kernel.packet[11]); + __m512 T11 = _mm512_unpackhi_ps(kernel.packet[10], kernel.packet[11]); + __m512 T12 = _mm512_unpacklo_ps(kernel.packet[12], kernel.packet[13]); + __m512 T13 = _mm512_unpackhi_ps(kernel.packet[12], kernel.packet[13]); + __m512 T14 = _mm512_unpacklo_ps(kernel.packet[14], kernel.packet[15]); + __m512 T15 = _mm512_unpackhi_ps(kernel.packet[14], kernel.packet[15]); + __m512 S0 = _mm512_shuffle_ps(T0, T2, _MM_SHUFFLE(1, 0, 1, 0)); + __m512 S1 = _mm512_shuffle_ps(T0, T2, _MM_SHUFFLE(3, 2, 3, 2)); + __m512 S2 = _mm512_shuffle_ps(T1, T3, _MM_SHUFFLE(1, 0, 1, 0)); + __m512 S3 = _mm512_shuffle_ps(T1, T3, _MM_SHUFFLE(3, 2, 3, 2)); + __m512 S4 = _mm512_shuffle_ps(T4, T6, _MM_SHUFFLE(1, 0, 1, 0)); + __m512 S5 = _mm512_shuffle_ps(T4, T6, _MM_SHUFFLE(3, 2, 3, 2)); + __m512 S6 = _mm512_shuffle_ps(T5, T7, _MM_SHUFFLE(1, 0, 1, 0)); + __m512 S7 = _mm512_shuffle_ps(T5, T7, _MM_SHUFFLE(3, 2, 3, 2)); + __m512 S8 = _mm512_shuffle_ps(T8, T10, _MM_SHUFFLE(1, 0, 1, 0)); + __m512 S9 = _mm512_shuffle_ps(T8, T10, _MM_SHUFFLE(3, 2, 3, 2)); + __m512 S10 = _mm512_shuffle_ps(T9, T11, _MM_SHUFFLE(1, 0, 1, 0)); + __m512 S11 = _mm512_shuffle_ps(T9, T11, _MM_SHUFFLE(3, 2, 3, 2)); + __m512 S12 = _mm512_shuffle_ps(T12, T14, _MM_SHUFFLE(1, 0, 1, 0)); + __m512 S13 = _mm512_shuffle_ps(T12, T14, _MM_SHUFFLE(3, 2, 3, 2)); + __m512 S14 = _mm512_shuffle_ps(T13, T15, _MM_SHUFFLE(1, 0, 1, 0)); + __m512 S15 = _mm512_shuffle_ps(T13, T15, _MM_SHUFFLE(3, 2, 3, 2)); + + EIGEN_EXTRACT_8f_FROM_16f(S0, S0); + EIGEN_EXTRACT_8f_FROM_16f(S1, S1); + EIGEN_EXTRACT_8f_FROM_16f(S2, S2); + EIGEN_EXTRACT_8f_FROM_16f(S3, S3); + EIGEN_EXTRACT_8f_FROM_16f(S4, S4); + EIGEN_EXTRACT_8f_FROM_16f(S5, S5); + EIGEN_EXTRACT_8f_FROM_16f(S6, S6); + EIGEN_EXTRACT_8f_FROM_16f(S7, S7); + EIGEN_EXTRACT_8f_FROM_16f(S8, S8); + EIGEN_EXTRACT_8f_FROM_16f(S9, S9); + EIGEN_EXTRACT_8f_FROM_16f(S10, S10); + EIGEN_EXTRACT_8f_FROM_16f(S11, S11); + EIGEN_EXTRACT_8f_FROM_16f(S12, S12); + EIGEN_EXTRACT_8f_FROM_16f(S13, S13); + EIGEN_EXTRACT_8f_FROM_16f(S14, S14); + EIGEN_EXTRACT_8f_FROM_16f(S15, S15); + + PacketBlock<Packet8f, 32> tmp; + + tmp.packet[0] = _mm256_permute2f128_ps(S0_0, S4_0, 0x20); + tmp.packet[1] = _mm256_permute2f128_ps(S1_0, S5_0, 0x20); + tmp.packet[2] = _mm256_permute2f128_ps(S2_0, S6_0, 0x20); + tmp.packet[3] = _mm256_permute2f128_ps(S3_0, S7_0, 0x20); + tmp.packet[4] = _mm256_permute2f128_ps(S0_0, S4_0, 0x31); + tmp.packet[5] = _mm256_permute2f128_ps(S1_0, S5_0, 0x31); + tmp.packet[6] = _mm256_permute2f128_ps(S2_0, S6_0, 0x31); + tmp.packet[7] = _mm256_permute2f128_ps(S3_0, S7_0, 0x31); + + tmp.packet[8] = _mm256_permute2f128_ps(S0_1, S4_1, 0x20); + tmp.packet[9] = _mm256_permute2f128_ps(S1_1, S5_1, 0x20); + tmp.packet[10] = _mm256_permute2f128_ps(S2_1, S6_1, 0x20); + tmp.packet[11] = _mm256_permute2f128_ps(S3_1, S7_1, 0x20); + tmp.packet[12] = _mm256_permute2f128_ps(S0_1, S4_1, 0x31); + tmp.packet[13] = _mm256_permute2f128_ps(S1_1, S5_1, 0x31); + tmp.packet[14] = _mm256_permute2f128_ps(S2_1, S6_1, 0x31); + tmp.packet[15] = _mm256_permute2f128_ps(S3_1, S7_1, 0x31); + + // Second set of _m256 outputs + tmp.packet[16] = _mm256_permute2f128_ps(S8_0, S12_0, 0x20); + tmp.packet[17] = _mm256_permute2f128_ps(S9_0, S13_0, 0x20); + tmp.packet[18] = _mm256_permute2f128_ps(S10_0, S14_0, 0x20); + tmp.packet[19] = _mm256_permute2f128_ps(S11_0, S15_0, 0x20); + tmp.packet[20] = _mm256_permute2f128_ps(S8_0, S12_0, 0x31); + tmp.packet[21] = _mm256_permute2f128_ps(S9_0, S13_0, 0x31); + tmp.packet[22] = _mm256_permute2f128_ps(S10_0, S14_0, 0x31); + tmp.packet[23] = _mm256_permute2f128_ps(S11_0, S15_0, 0x31); + + tmp.packet[24] = _mm256_permute2f128_ps(S8_1, S12_1, 0x20); + tmp.packet[25] = _mm256_permute2f128_ps(S9_1, S13_1, 0x20); + tmp.packet[26] = _mm256_permute2f128_ps(S10_1, S14_1, 0x20); + tmp.packet[27] = _mm256_permute2f128_ps(S11_1, S15_1, 0x20); + tmp.packet[28] = _mm256_permute2f128_ps(S8_1, S12_1, 0x31); + tmp.packet[29] = _mm256_permute2f128_ps(S9_1, S13_1, 0x31); + tmp.packet[30] = _mm256_permute2f128_ps(S10_1, S14_1, 0x31); + tmp.packet[31] = _mm256_permute2f128_ps(S11_1, S15_1, 0x31); + + // Pack them into the output + PACK_OUTPUT(kernel.packet, tmp.packet, 0, 16); + PACK_OUTPUT(kernel.packet, tmp.packet, 1, 16); + PACK_OUTPUT(kernel.packet, tmp.packet, 2, 16); + PACK_OUTPUT(kernel.packet, tmp.packet, 3, 16); + + PACK_OUTPUT(kernel.packet, tmp.packet, 4, 16); + PACK_OUTPUT(kernel.packet, tmp.packet, 5, 16); + PACK_OUTPUT(kernel.packet, tmp.packet, 6, 16); + PACK_OUTPUT(kernel.packet, tmp.packet, 7, 16); + + PACK_OUTPUT(kernel.packet, tmp.packet, 8, 16); + PACK_OUTPUT(kernel.packet, tmp.packet, 9, 16); + PACK_OUTPUT(kernel.packet, tmp.packet, 10, 16); + PACK_OUTPUT(kernel.packet, tmp.packet, 11, 16); + + PACK_OUTPUT(kernel.packet, tmp.packet, 12, 16); + PACK_OUTPUT(kernel.packet, tmp.packet, 13, 16); + PACK_OUTPUT(kernel.packet, tmp.packet, 14, 16); + PACK_OUTPUT(kernel.packet, tmp.packet, 15, 16); +} +#define PACK_OUTPUT_2(OUTPUT, INPUT, INDEX, STRIDE) \ + EIGEN_INSERT_8f_INTO_16f(OUTPUT[INDEX], INPUT[2 * INDEX], \ + INPUT[2 * INDEX + STRIDE]); + +EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet16f, 4>& kernel) { + __m512 T0 = _mm512_unpacklo_ps(kernel.packet[0], kernel.packet[1]); + __m512 T1 = _mm512_unpackhi_ps(kernel.packet[0], kernel.packet[1]); + __m512 T2 = _mm512_unpacklo_ps(kernel.packet[2], kernel.packet[3]); + __m512 T3 = _mm512_unpackhi_ps(kernel.packet[2], kernel.packet[3]); + + __m512 S0 = _mm512_shuffle_ps(T0, T2, _MM_SHUFFLE(1, 0, 1, 0)); + __m512 S1 = _mm512_shuffle_ps(T0, T2, _MM_SHUFFLE(3, 2, 3, 2)); + __m512 S2 = _mm512_shuffle_ps(T1, T3, _MM_SHUFFLE(1, 0, 1, 0)); + __m512 S3 = _mm512_shuffle_ps(T1, T3, _MM_SHUFFLE(3, 2, 3, 2)); + + EIGEN_EXTRACT_8f_FROM_16f(S0, S0); + EIGEN_EXTRACT_8f_FROM_16f(S1, S1); + EIGEN_EXTRACT_8f_FROM_16f(S2, S2); + EIGEN_EXTRACT_8f_FROM_16f(S3, S3); + + PacketBlock<Packet8f, 8> tmp; + + tmp.packet[0] = _mm256_permute2f128_ps(S0_0, S1_0, 0x20); + tmp.packet[1] = _mm256_permute2f128_ps(S2_0, S3_0, 0x20); + tmp.packet[2] = _mm256_permute2f128_ps(S0_0, S1_0, 0x31); + tmp.packet[3] = _mm256_permute2f128_ps(S2_0, S3_0, 0x31); + + tmp.packet[4] = _mm256_permute2f128_ps(S0_1, S1_1, 0x20); + tmp.packet[5] = _mm256_permute2f128_ps(S2_1, S3_1, 0x20); + tmp.packet[6] = _mm256_permute2f128_ps(S0_1, S1_1, 0x31); + tmp.packet[7] = _mm256_permute2f128_ps(S2_1, S3_1, 0x31); + + PACK_OUTPUT_2(kernel.packet, tmp.packet, 0, 1); + PACK_OUTPUT_2(kernel.packet, tmp.packet, 1, 1); + PACK_OUTPUT_2(kernel.packet, tmp.packet, 2, 1); + PACK_OUTPUT_2(kernel.packet, tmp.packet, 3, 1); +} + +#define PACK_OUTPUT_SQ_D(OUTPUT, INPUT, INDEX, STRIDE) \ + OUTPUT[INDEX] = _mm512_insertf64x4(OUTPUT[INDEX], INPUT[INDEX], 0); \ + OUTPUT[INDEX] = _mm512_insertf64x4(OUTPUT[INDEX], INPUT[INDEX + STRIDE], 1); + +#define PACK_OUTPUT_D(OUTPUT, INPUT, INDEX, STRIDE) \ + OUTPUT[INDEX] = _mm512_insertf64x4(OUTPUT[INDEX], INPUT[(2 * INDEX)], 0); \ + OUTPUT[INDEX] = \ + _mm512_insertf64x4(OUTPUT[INDEX], INPUT[(2 * INDEX) + STRIDE], 1); + +EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8d, 4>& kernel) { + __m512d T0 = _mm512_shuffle_pd(kernel.packet[0], kernel.packet[1], 0); + __m512d T1 = _mm512_shuffle_pd(kernel.packet[0], kernel.packet[1], 0xff); + __m512d T2 = _mm512_shuffle_pd(kernel.packet[2], kernel.packet[3], 0); + __m512d T3 = _mm512_shuffle_pd(kernel.packet[2], kernel.packet[3], 0xff); + + PacketBlock<Packet4d, 8> tmp; + + tmp.packet[0] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T0, 0), + _mm512_extractf64x4_pd(T2, 0), 0x20); + tmp.packet[1] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T1, 0), + _mm512_extractf64x4_pd(T3, 0), 0x20); + tmp.packet[2] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T0, 0), + _mm512_extractf64x4_pd(T2, 0), 0x31); + tmp.packet[3] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T1, 0), + _mm512_extractf64x4_pd(T3, 0), 0x31); + + tmp.packet[4] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T0, 1), + _mm512_extractf64x4_pd(T2, 1), 0x20); + tmp.packet[5] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T1, 1), + _mm512_extractf64x4_pd(T3, 1), 0x20); + tmp.packet[6] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T0, 1), + _mm512_extractf64x4_pd(T2, 1), 0x31); + tmp.packet[7] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T1, 1), + _mm512_extractf64x4_pd(T3, 1), 0x31); + + PACK_OUTPUT_D(kernel.packet, tmp.packet, 0, 1); + PACK_OUTPUT_D(kernel.packet, tmp.packet, 1, 1); + PACK_OUTPUT_D(kernel.packet, tmp.packet, 2, 1); + PACK_OUTPUT_D(kernel.packet, tmp.packet, 3, 1); +} + +EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8d, 8>& kernel) { + __m512d T0 = _mm512_unpacklo_pd(kernel.packet[0], kernel.packet[1]); + __m512d T1 = _mm512_unpackhi_pd(kernel.packet[0], kernel.packet[1]); + __m512d T2 = _mm512_unpacklo_pd(kernel.packet[2], kernel.packet[3]); + __m512d T3 = _mm512_unpackhi_pd(kernel.packet[2], kernel.packet[3]); + __m512d T4 = _mm512_unpacklo_pd(kernel.packet[4], kernel.packet[5]); + __m512d T5 = _mm512_unpackhi_pd(kernel.packet[4], kernel.packet[5]); + __m512d T6 = _mm512_unpacklo_pd(kernel.packet[6], kernel.packet[7]); + __m512d T7 = _mm512_unpackhi_pd(kernel.packet[6], kernel.packet[7]); + + PacketBlock<Packet4d, 16> tmp; + + tmp.packet[0] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T0, 0), + _mm512_extractf64x4_pd(T2, 0), 0x20); + tmp.packet[1] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T1, 0), + _mm512_extractf64x4_pd(T3, 0), 0x20); + tmp.packet[2] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T0, 0), + _mm512_extractf64x4_pd(T2, 0), 0x31); + tmp.packet[3] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T1, 0), + _mm512_extractf64x4_pd(T3, 0), 0x31); + + tmp.packet[4] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T0, 1), + _mm512_extractf64x4_pd(T2, 1), 0x20); + tmp.packet[5] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T1, 1), + _mm512_extractf64x4_pd(T3, 1), 0x20); + tmp.packet[6] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T0, 1), + _mm512_extractf64x4_pd(T2, 1), 0x31); + tmp.packet[7] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T1, 1), + _mm512_extractf64x4_pd(T3, 1), 0x31); + + tmp.packet[8] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T4, 0), + _mm512_extractf64x4_pd(T6, 0), 0x20); + tmp.packet[9] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T5, 0), + _mm512_extractf64x4_pd(T7, 0), 0x20); + tmp.packet[10] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T4, 0), + _mm512_extractf64x4_pd(T6, 0), 0x31); + tmp.packet[11] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T5, 0), + _mm512_extractf64x4_pd(T7, 0), 0x31); + + tmp.packet[12] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T4, 1), + _mm512_extractf64x4_pd(T6, 1), 0x20); + tmp.packet[13] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T5, 1), + _mm512_extractf64x4_pd(T7, 1), 0x20); + tmp.packet[14] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T4, 1), + _mm512_extractf64x4_pd(T6, 1), 0x31); + tmp.packet[15] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T5, 1), + _mm512_extractf64x4_pd(T7, 1), 0x31); + + PACK_OUTPUT_SQ_D(kernel.packet, tmp.packet, 0, 8); + PACK_OUTPUT_SQ_D(kernel.packet, tmp.packet, 1, 8); + PACK_OUTPUT_SQ_D(kernel.packet, tmp.packet, 2, 8); + PACK_OUTPUT_SQ_D(kernel.packet, tmp.packet, 3, 8); + + PACK_OUTPUT_SQ_D(kernel.packet, tmp.packet, 4, 8); + PACK_OUTPUT_SQ_D(kernel.packet, tmp.packet, 5, 8); + PACK_OUTPUT_SQ_D(kernel.packet, tmp.packet, 6, 8); + PACK_OUTPUT_SQ_D(kernel.packet, tmp.packet, 7, 8); +} +template <> +EIGEN_STRONG_INLINE Packet16f pblend(const Selector<16>& /*ifPacket*/, + const Packet16f& /*thenPacket*/, + const Packet16f& /*elsePacket*/) { + assert(false && "To be implemented"); + return Packet16f(); +} +template <> +EIGEN_STRONG_INLINE Packet8d pblend(const Selector<8>& /*ifPacket*/, + const Packet8d& /*thenPacket*/, + const Packet8d& /*elsePacket*/) { + assert(false && "To be implemented"); + return Packet8d(); +} + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_PACKET_MATH_AVX512_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AltiVec/Complex.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AltiVec/Complex.h new file mode 100644 index 000000000..67db2f8ee --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AltiVec/Complex.h @@ -0,0 +1,461 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2010-2016 Konstantinos Margaritis <markos@freevec.org> +// +// 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_COMPLEX32_ALTIVEC_H +#define EIGEN_COMPLEX32_ALTIVEC_H + +namespace Eigen { + +namespace internal { + +static Packet4ui p4ui_CONJ_XOR = vec_mergeh((Packet4ui)p4i_ZERO, (Packet4ui)p4f_MZERO);//{ 0x00000000, 0x80000000, 0x00000000, 0x80000000 }; +#ifdef __VSX__ +#if defined(_BIG_ENDIAN) +static Packet2ul p2ul_CONJ_XOR1 = (Packet2ul) vec_sld((Packet4ui) p2d_MZERO, (Packet4ui) p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 }; +static Packet2ul p2ul_CONJ_XOR2 = (Packet2ul) vec_sld((Packet4ui) p2l_ZERO, (Packet4ui) p2d_MZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 }; +#else +static Packet2ul p2ul_CONJ_XOR1 = (Packet2ul) vec_sld((Packet4ui) p2l_ZERO, (Packet4ui) p2d_MZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 }; +static Packet2ul p2ul_CONJ_XOR2 = (Packet2ul) vec_sld((Packet4ui) p2d_MZERO, (Packet4ui) p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 }; +#endif +#endif + +//---------- float ---------- +struct Packet2cf +{ + EIGEN_STRONG_INLINE explicit Packet2cf() : v(p4f_ZERO) {} + EIGEN_STRONG_INLINE explicit Packet2cf(const Packet4f& a) : v(a) {} + Packet4f v; +}; + +template<> struct packet_traits<std::complex<float> > : default_packet_traits +{ + typedef Packet2cf type; + typedef Packet2cf half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size = 2, + HasHalfPacket = 0, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasNegate = 1, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, +#ifdef __VSX__ + HasBlend = 1, +#endif + HasSetLinear = 0 + }; +}; + +template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16}; typedef Packet2cf half; }; + +template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>& from) +{ + Packet2cf res; + if((std::ptrdiff_t(&from) % 16) == 0) + res.v = pload<Packet4f>((const float *)&from); + else + res.v = ploadu<Packet4f>((const float *)&from); + res.v = vec_perm(res.v, res.v, p16uc_PSET64_HI); + return res; +} + +template<> EIGEN_STRONG_INLINE Packet2cf pload<Packet2cf>(const std::complex<float>* from) { return Packet2cf(pload<Packet4f>((const float *) from)); } +template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from) { return Packet2cf(ploadu<Packet4f>((const float*) from)); } +template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from) { return pset1<Packet2cf>(*from); } + +template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { pstore((float*)to, from.v); } +template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { pstoreu((float*)to, from.v); } + +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]; + 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]; + pstore<std::complex<float> >((std::complex<float> *) af, from); + to[0*stride] = af[0]; + to[1*stride] = af[1]; +} + +template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(a.v + b.v); } +template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(a.v - b.v); } +template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) { return Packet2cf(pnegate(a.v)); } +template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a) { return Packet2cf(pxor<Packet4f>(a.v, reinterpret_cast<Packet4f>(p4ui_CONJ_XOR))); } + +template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b) +{ + Packet4f v1, v2; + + // Permute and multiply the real parts of a and b + v1 = vec_perm(a.v, a.v, p16uc_PSET32_WODD); + // Get the imaginary parts of a + v2 = vec_perm(a.v, a.v, p16uc_PSET32_WEVEN); + // multiply a_re * b + v1 = vec_madd(v1, b.v, p4f_ZERO); + // multiply a_im * b and get the conjugate result + v2 = vec_madd(v2, b.v, p4f_ZERO); + v2 = reinterpret_cast<Packet4f>(pxor(v2, reinterpret_cast<Packet4f>(p4ui_CONJ_XOR))); + // permute back to a proper order + v2 = vec_perm(v2, v2, p16uc_COMPLEX32_REV); + + return Packet2cf(padd<Packet4f>(v1, v2)); +} + +template<> EIGEN_STRONG_INLINE Packet2cf pand <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pand<Packet4f>(a.v, b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cf por <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(por<Packet4f>(a.v, b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cf pxor <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pxor<Packet4f>(a.v, b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pandnot<Packet4f>(a.v, b.v)); } + +template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> * addr) { EIGEN_PPC_PREFETCH(addr); } + +template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a) +{ + std::complex<float> EIGEN_ALIGN16 res[2]; + pstore((float *)&res, a.v); + + return res[0]; +} + +template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) +{ + Packet4f rev_a; + rev_a = vec_perm(a.v, a.v, p16uc_COMPLEX32_REV2); + return Packet2cf(rev_a); +} + +template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a) +{ + Packet4f b; + b = vec_sld(a.v, a.v, 8); + b = padd<Packet4f>(a.v, b); + return pfirst<Packet2cf>(Packet2cf(b)); +} + +template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs) +{ + Packet4f b1, b2; +#ifdef _BIG_ENDIAN + b1 = vec_sld(vecs[0].v, vecs[1].v, 8); + b2 = vec_sld(vecs[1].v, vecs[0].v, 8); +#else + b1 = vec_sld(vecs[1].v, vecs[0].v, 8); + b2 = vec_sld(vecs[0].v, vecs[1].v, 8); +#endif + b2 = vec_sld(b2, b2, 8); + b2 = padd<Packet4f>(b1, b2); + + return Packet2cf(b2); +} + +template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a) +{ + Packet4f b; + Packet2cf prod; + b = vec_sld(a.v, a.v, 8); + prod = pmul<Packet2cf>(a, Packet2cf(b)); + + return pfirst<Packet2cf>(prod); +} + +template<int Offset> +struct palign_impl<Offset,Packet2cf> +{ + static EIGEN_STRONG_INLINE void run(Packet2cf& first, const Packet2cf& second) + { + if (Offset==1) + { +#ifdef _BIG_ENDIAN + first.v = vec_sld(first.v, second.v, 8); +#else + first.v = vec_sld(second.v, first.v, 8); +#endif + } + } +}; + +template<> struct conj_helper<Packet2cf, Packet2cf, false,true> +{ + EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const + { + return internal::pmul(a, pconj(b)); + } +}; + +template<> struct conj_helper<Packet2cf, Packet2cf, true,false> +{ + EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const + { + return internal::pmul(pconj(a), b); + } +}; + +template<> struct conj_helper<Packet2cf, Packet2cf, true,true> +{ + EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const + { + return pconj(internal::pmul(a, b)); + } +}; + +template<> struct conj_helper<Packet4f, Packet2cf, false,false> +{ + EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet4f& x, const Packet2cf& y, const Packet2cf& c) const + { return padd(c, pmul(x,y)); } + + EIGEN_STRONG_INLINE Packet2cf pmul(const Packet4f& x, const Packet2cf& y) const + { return Packet2cf(internal::pmul<Packet4f>(x, y.v)); } +}; + +template<> struct conj_helper<Packet2cf, Packet4f, false,false> +{ + EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet4f& y, const Packet2cf& c) const + { return padd(c, pmul(x,y)); } + + EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& x, const Packet4f& y) const + { return Packet2cf(internal::pmul<Packet4f>(x.v, y)); } +}; + +template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b) +{ + // TODO optimize it for AltiVec + Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a, b); + Packet4f s = pmul<Packet4f>(b.v, b.v); + return Packet2cf(pdiv(res.v, padd<Packet4f>(s, vec_perm(s, s, p16uc_COMPLEX32_REV)))); +} + +template<> EIGEN_STRONG_INLINE Packet2cf pcplxflip<Packet2cf>(const Packet2cf& x) +{ + return Packet2cf(vec_perm(x.v, x.v, p16uc_COMPLEX32_REV)); +} + +EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2cf,2>& kernel) +{ + Packet4f tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_HI); + kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO); + kernel.packet[0].v = tmp; +} + +#ifdef __VSX__ +template<> EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, const Packet2cf& thenPacket, const Packet2cf& elsePacket) { + Packet2cf result; + result.v = reinterpret_cast<Packet4f>(pblend<Packet2d>(ifPacket, reinterpret_cast<Packet2d>(thenPacket.v), reinterpret_cast<Packet2d>(elsePacket.v))); + return result; +} +#endif + +//---------- double ---------- +#ifdef __VSX__ +struct Packet1cd +{ + EIGEN_STRONG_INLINE Packet1cd() {} + EIGEN_STRONG_INLINE explicit Packet1cd(const Packet2d& a) : v(a) {} + Packet2d v; +}; + +template<> struct packet_traits<std::complex<double> > : default_packet_traits +{ + typedef Packet1cd type; + typedef Packet1cd half; + enum { + Vectorizable = 1, + AlignedOnScalar = 0, + size = 1, + HasHalfPacket = 0, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasNegate = 1, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasSetLinear = 0 + }; +}; + +template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16}; 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)); } +template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> * to, const Packet1cd& from) { pstore((double*)to, from.v); } +template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> * to, const Packet1cd& from) { pstoreu((double*)to, from.v); } + +template<> EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<double>& from) +{ /* here we really have to use unaligned loads :( */ return ploadu<Packet1cd>(&from); } + +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]; + 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]; + pstore<std::complex<double> >(af, from); + to[0*stride] = af[0]; + to[1*stride] = af[1]; +} + +template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(a.v + b.v); } +template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(a.v - b.v); } +template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(Packet2d(a.v))); } +template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) { return Packet1cd(pxor(a.v, reinterpret_cast<Packet2d>(p2ul_CONJ_XOR2))); } + +template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b) +{ + Packet2d a_re, a_im, v1, v2; + + // Permute and multiply the real parts of a and b + a_re = vec_perm(a.v, a.v, p16uc_PSET64_HI); + // Get the imaginary parts of a + a_im = vec_perm(a.v, a.v, p16uc_PSET64_LO); + // multiply a_re * b + v1 = vec_madd(a_re, b.v, p2d_ZERO); + // multiply a_im * b and get the conjugate result + v2 = vec_madd(a_im, b.v, p2d_ZERO); + v2 = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(v2), reinterpret_cast<Packet4ui>(v2), 8)); + v2 = pxor(v2, reinterpret_cast<Packet2d>(p2ul_CONJ_XOR1)); + + return Packet1cd(padd<Packet2d>(v1, v2)); +} + +template<> EIGEN_STRONG_INLINE Packet1cd pand <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(pand(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd por <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(por(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd pxor <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(pxor(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(pandnot(a.v, b.v)); } + +template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>* from) { return pset1<Packet1cd>(*from); } + +template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> * addr) { EIGEN_PPC_PREFETCH(addr); } + +template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Packet1cd& a) +{ + std::complex<double> EIGEN_ALIGN16 res[2]; + pstore<std::complex<double> >(res, a); + + return res[0]; +} + +template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; } + +template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a) { return pfirst(a); } +template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs) { return vecs[0]; } + +template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a) { return pfirst(a); } + +template<int Offset> +struct palign_impl<Offset,Packet1cd> +{ + static EIGEN_STRONG_INLINE void run(Packet1cd& /*first*/, const Packet1cd& /*second*/) + { + // FIXME is it sure we never have to align a Packet1cd? + // Even though a std::complex<double> has 16 bytes, it is not necessarily aligned on a 16 bytes boundary... + } +}; + +template<> struct conj_helper<Packet1cd, Packet1cd, false,true> +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const + { + return internal::pmul(a, pconj(b)); + } +}; + +template<> struct conj_helper<Packet1cd, Packet1cd, true,false> +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const + { + return internal::pmul(pconj(a), b); + } +}; + +template<> struct conj_helper<Packet1cd, Packet1cd, true,true> +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const + { + return pconj(internal::pmul(a, b)); + } +}; +template<> struct conj_helper<Packet2d, Packet1cd, false,false> +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet2d& x, const Packet1cd& y, const Packet1cd& c) const + { return padd(c, pmul(x,y)); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet2d& x, const Packet1cd& y) const + { return Packet1cd(internal::pmul<Packet2d>(x, y.v)); } +}; + +template<> struct conj_helper<Packet1cd, Packet2d, false,false> +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet2d& y, const Packet1cd& c) const + { return padd(c, pmul(x,y)); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& x, const Packet2d& y) const + { return Packet1cd(internal::pmul<Packet2d>(x.v, y)); } +}; + +template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b) +{ + // TODO optimize it for AltiVec + Packet1cd res = conj_helper<Packet1cd,Packet1cd,false,true>().pmul(a,b); + Packet2d s = pmul<Packet2d>(b.v, b.v); + return Packet1cd(pdiv(res.v, padd<Packet2d>(s, vec_perm(s, s, p16uc_REVERSE64)))); +} + +EIGEN_STRONG_INLINE Packet1cd pcplxflip/*<Packet1cd>*/(const Packet1cd& x) +{ + return Packet1cd(preverse(Packet2d(x.v))); +} + +EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet1cd,2>& kernel) +{ + Packet2d tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_HI); + kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO); + kernel.packet[0].v = tmp; +} +#endif // __VSX__ +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_COMPLEX32_ALTIVEC_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AltiVec/MathFunctions.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AltiVec/MathFunctions.h new file mode 100644 index 000000000..c5e4bede7 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AltiVec/MathFunctions.h @@ -0,0 +1,322 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2007 Julien Pommier +// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2016 Konstantinos Margaritis <markos@freevec.org> +// +// 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 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 + +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; +} + +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); + + // build 2^n + emm0 = vec_cts(fx, 0); + emm0 = vec_add(emm0, p4i_0x7f); + emm0 = vec_sl(emm0, reinterpret_cast<Packet4ui>(p4i_23)); + + // 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); +} + +#ifndef EIGEN_COMP_CLANG +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet4f prsqrt<Packet4f>(const Packet4f& x) +{ + return vec_rsqrt(x); +} +#endif + +#ifdef __VSX__ +#ifndef EIGEN_COMP_CLANG +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet2d prsqrt<Packet2d>(const Packet2d& x) +{ + return vec_rsqrt(x); +} +#endif + +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet4f psqrt<Packet4f>(const Packet4f& x) +{ + return vec_sqrt(x); +} + +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +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); +} +#endif + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_MATH_FUNCTIONS_ALTIVEC_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AltiVec/PacketMath.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AltiVec/PacketMath.h new file mode 100755 index 000000000..b3f1ea199 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -0,0 +1,1033 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2016 Konstantinos Margaritis <markos@freevec.org> +// +// 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_PACKET_MATH_ALTIVEC_H +#define EIGEN_PACKET_MATH_ALTIVEC_H + +namespace Eigen { + +namespace internal { + +#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD +#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 4 +#endif + +#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD +#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD +#endif + +#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD +#define EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD +#endif + +// NOTE Altivec has 32 registers, but Eigen only accepts a value of 8 or 16 +#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS +#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 32 +#endif + +typedef __vector float Packet4f; +typedef __vector int Packet4i; +typedef __vector unsigned int Packet4ui; +typedef __vector __bool int Packet4bi; +typedef __vector short int Packet8i; +typedef __vector unsigned char Packet16uc; + +// We don't want to write the same code all the time, but we need to reuse the constants +// and it doesn't really work to declare them global, so we define macros instead + +#define _EIGEN_DECLARE_CONST_FAST_Packet4f(NAME,X) \ + Packet4f p4f_##NAME = reinterpret_cast<Packet4f>(vec_splat_s32(X)) + +#define _EIGEN_DECLARE_CONST_FAST_Packet4i(NAME,X) \ + Packet4i p4i_##NAME = vec_splat_s32(X) + +#define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \ + Packet4f p4f_##NAME = pset1<Packet4f>(X) + +#define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \ + Packet4i p4i_##NAME = pset1<Packet4i>(X) + +#define _EIGEN_DECLARE_CONST_Packet2d(NAME,X) \ + Packet2d p2d_##NAME = pset1<Packet2d>(X) + +#define _EIGEN_DECLARE_CONST_Packet2l(NAME,X) \ + Packet2l p2l_##NAME = pset1<Packet2l>(X) + +#define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \ + const Packet4f p4f_##NAME = reinterpret_cast<Packet4f>(pset1<Packet4i>(X)) + +#define DST_CHAN 1 +#define DST_CTRL(size, count, stride) (((size) << 24) | ((count) << 16) | (stride)) + + +// These constants are endian-agnostic +static _EIGEN_DECLARE_CONST_FAST_Packet4f(ZERO, 0); //{ 0.0, 0.0, 0.0, 0.0} +static _EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0); //{ 0, 0, 0, 0,} +static _EIGEN_DECLARE_CONST_FAST_Packet4i(ONE,1); //{ 1, 1, 1, 1} +static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS16,-16); //{ -16, -16, -16, -16} +static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS1,-1); //{ -1, -1, -1, -1} +static Packet4f p4f_MZERO = (Packet4f) vec_sl((Packet4ui)p4i_MINUS1, (Packet4ui)p4i_MINUS1); //{ 0x80000000, 0x80000000, 0x80000000, 0x80000000} +#ifndef __VSX__ +static Packet4f p4f_ONE = vec_ctf(p4i_ONE, 0); //{ 1.0, 1.0, 1.0, 1.0} +#endif + +static Packet4f p4f_COUNTDOWN = { 0.0, 1.0, 2.0, 3.0 }; +static Packet4i p4i_COUNTDOWN = { 0, 1, 2, 3 }; + +static Packet16uc p16uc_REVERSE32 = { 12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3 }; +static Packet16uc p16uc_DUPLICATE32_HI = { 0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7 }; + +// Mask alignment +#ifdef __PPC64__ +#define _EIGEN_MASK_ALIGNMENT 0xfffffffffffffff0 +#else +#define _EIGEN_MASK_ALIGNMENT 0xfffffff0 +#endif + +#define _EIGEN_ALIGNED_PTR(x) ((std::ptrdiff_t)(x) & _EIGEN_MASK_ALIGNMENT) + +// Handle endianness properly while loading constants +// Define global static constants: +#ifdef _BIG_ENDIAN +static Packet16uc p16uc_FORWARD = vec_lvsl(0, (float*)0); +#ifdef __VSX__ +static Packet16uc p16uc_REVERSE64 = { 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 }; +#endif +static Packet16uc p16uc_PSET32_WODD = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 }; +static Packet16uc p16uc_PSET32_WEVEN = vec_sld(p16uc_DUPLICATE32_HI, (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 }; +static Packet16uc p16uc_HALF64_0_16 = vec_sld((Packet16uc)p4i_ZERO, vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 3), 8); //{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16}; +#else +static Packet16uc p16uc_FORWARD = p16uc_REVERSE32; +static Packet16uc p16uc_REVERSE64 = { 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 }; +static Packet16uc p16uc_PSET32_WODD = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 1), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 }; +static Packet16uc p16uc_PSET32_WEVEN = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 }; +static Packet16uc p16uc_HALF64_0_16 = vec_sld(vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 0), (Packet16uc)p4i_ZERO, 8); //{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16}; +#endif // _BIG_ENDIAN + +static Packet16uc p16uc_PSET64_HI = (Packet16uc) vec_mergeh((Packet4ui)p16uc_PSET32_WODD, (Packet4ui)p16uc_PSET32_WEVEN); //{ 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 }; +static Packet16uc p16uc_PSET64_LO = (Packet16uc) vec_mergel((Packet4ui)p16uc_PSET32_WODD, (Packet4ui)p16uc_PSET32_WEVEN); //{ 8,9,10,11, 12,13,14,15, 8,9,10,11, 12,13,14,15 }; +static Packet16uc p16uc_TRANSPOSE64_HI = p16uc_PSET64_HI + p16uc_HALF64_0_16; //{ 0,1,2,3, 4,5,6,7, 16,17,18,19, 20,21,22,23}; +static Packet16uc p16uc_TRANSPOSE64_LO = p16uc_PSET64_LO + p16uc_HALF64_0_16; //{ 8,9,10,11, 12,13,14,15, 24,25,26,27, 28,29,30,31}; + +static Packet16uc p16uc_COMPLEX32_REV = vec_sld(p16uc_REVERSE32, p16uc_REVERSE32, 8); //{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 }; + +#ifdef _BIG_ENDIAN +static Packet16uc p16uc_COMPLEX32_REV2 = vec_sld(p16uc_FORWARD, p16uc_FORWARD, 8); //{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 }; +#else +static Packet16uc p16uc_COMPLEX32_REV2 = vec_sld(p16uc_PSET64_HI, p16uc_PSET64_LO, 8); //{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 }; +#endif // _BIG_ENDIAN + +#if EIGEN_HAS_BUILTIN(__builtin_prefetch) || EIGEN_COMP_GNUC + #define EIGEN_PPC_PREFETCH(ADDR) __builtin_prefetch(ADDR); +#else + #define EIGEN_PPC_PREFETCH(ADDR) asm( " dcbt [%[addr]]\n" :: [addr] "r" (ADDR) : "cc" ); +#endif + +template<> struct packet_traits<float> : default_packet_traits +{ + typedef Packet4f type; + typedef Packet4f half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=4, + HasHalfPacket = 1, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasMin = 1, + HasMax = 1, + HasAbs = 1, + HasSin = 0, + HasCos = 0, + HasLog = 0, + HasExp = 1, +#ifdef __VSX__ + HasSqrt = 1, +#if !EIGEN_COMP_CLANG + HasRsqrt = 1, +#else + HasRsqrt = 0, +#endif +#else + HasSqrt = 0, + HasRsqrt = 0, +#endif + HasRound = 1, + HasFloor = 1, + HasCeil = 1, + HasNegate = 1, + HasBlend = 1 + }; +}; +template<> struct packet_traits<int> : default_packet_traits +{ + typedef Packet4i type; + typedef Packet4i half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size = 4, + HasHalfPacket = 0, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 0, + HasBlend = 1 + }; +}; + + +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; }; + +inline std::ostream & operator <<(std::ostream & s, const Packet16uc & v) +{ + union { + Packet16uc v; + unsigned char n[16]; + } vt; + vt.v = v; + for (int i=0; i< 16; i++) + s << (int)vt.n[i] << ", "; + return s; +} + +inline std::ostream & operator <<(std::ostream & s, const Packet4f & v) +{ + union { + Packet4f v; + float n[4]; + } vt; + vt.v = v; + s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3]; + return s; +} + +inline std::ostream & operator <<(std::ostream & s, const Packet4i & v) +{ + union { + Packet4i v; + int n[4]; + } vt; + vt.v = v; + s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3]; + return s; +} + +inline std::ostream & operator <<(std::ostream & s, const Packet4ui & v) +{ + union { + Packet4ui v; + unsigned int n[4]; + } vt; + vt.v = v; + s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3]; + return s; +} + +// Need to define them first or we get specialization after instantiation errors +template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) +{ + EIGEN_DEBUG_ALIGNED_LOAD +#ifdef __VSX__ + return vec_vsx_ld(0, from); +#else + return vec_ld(0, from); +#endif +} + +template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int* from) +{ + EIGEN_DEBUG_ALIGNED_LOAD +#ifdef __VSX__ + return vec_vsx_ld(0, from); +#else + return vec_ld(0, from); +#endif +} + +template<> EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet4f& from) +{ + EIGEN_DEBUG_ALIGNED_STORE +#ifdef __VSX__ + vec_vsx_st(from, 0, to); +#else + vec_st(from, 0, to); +#endif +} + +template<> EIGEN_STRONG_INLINE void pstore<int>(int* to, const Packet4i& from) +{ + EIGEN_DEBUG_ALIGNED_STORE +#ifdef __VSX__ + vec_vsx_st(from, 0, to); +#else + vec_st(from, 0, to); +#endif +} + +template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) { + Packet4f v = {from, from, from, from}; + return v; +} + +template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) { + Packet4i v = {from, from, from, from}; + return v; +} +template<> EIGEN_STRONG_INLINE void +pbroadcast4<Packet4f>(const float *a, + Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3) +{ + a3 = pload<Packet4f>(a); + a0 = vec_splat(a3, 0); + a1 = vec_splat(a3, 1); + a2 = vec_splat(a3, 2); + a3 = vec_splat(a3, 3); +} +template<> EIGEN_STRONG_INLINE void +pbroadcast4<Packet4i>(const int *a, + Packet4i& a0, Packet4i& a1, Packet4i& a2, Packet4i& a3) +{ + a3 = pload<Packet4i>(a); + a0 = vec_splat(a3, 0); + a1 = vec_splat(a3, 1); + a2 = vec_splat(a3, 2); + a3 = vec_splat(a3, 3); +} + +template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride) +{ + float EIGEN_ALIGN16 af[4]; + af[0] = from[0*stride]; + af[1] = from[1*stride]; + af[2] = from[2*stride]; + af[3] = from[3*stride]; + return pload<Packet4f>(af); +} +template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* from, Index stride) +{ + int EIGEN_ALIGN16 ai[4]; + ai[0] = from[0*stride]; + ai[1] = from[1*stride]; + ai[2] = from[2*stride]; + ai[3] = from[3*stride]; + return pload<Packet4i>(ai); +} +template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride) +{ + float EIGEN_ALIGN16 af[4]; + pstore<float>(af, from); + to[0*stride] = af[0]; + to[1*stride] = af[1]; + to[2*stride] = af[2]; + to[3*stride] = af[3]; +} +template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const Packet4i& from, Index stride) +{ + int EIGEN_ALIGN16 ai[4]; + pstore<int>((int *)ai, from); + to[0*stride] = ai[0]; + to[1*stride] = ai[1]; + to[2*stride] = ai[2]; + to[3*stride] = ai[3]; +} + +template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a) { return pset1<Packet4f>(a) + p4f_COUNTDOWN; } +template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int& a) { return pset1<Packet4i>(a) + p4i_COUNTDOWN; } + +template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return a + b; } +template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return a + b; } + +template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { return a - b; } +template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return a - b; } + +template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a) { return p4f_ZERO - a; } +template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return p4i_ZERO - a; } + +template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; } +template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; } + +template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_madd(a,b, p4f_MZERO); } +template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) { return a * b; } + +template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) +{ +#ifndef __VSX__ // VSX actually provides a div instruction + Packet4f t, y_0, y_1; + + // Altivec does not offer a divide instruction, we have to do a reciprocal approximation + y_0 = vec_re(b); + + // Do one Newton-Raphson iteration to get the needed accuracy + t = vec_nmsub(y_0, b, p4f_ONE); + y_1 = vec_madd(y_0, t, y_0); + + return vec_madd(a, y_1, p4f_MZERO); +#else + return vec_div(a, b); +#endif +} + +template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, const Packet4i& /*b*/) +{ eigen_assert(false && "packet integer division are not supported by AltiVec"); + return pset1<Packet4i>(0); +} + +// for some weird raisons, it has to be overloaded for packet of integers +template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vec_madd(a,b,c); } +template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return a*b + c; } + +template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_min(a, b); } +template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_min(a, b); } + +template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_max(a, b); } +template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_max(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); } + +template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_or(a, b); } +template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_or(a, b); } + +template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_xor(a, b); } +template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_xor(a, b); } + +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 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); } + +#ifdef _BIG_ENDIAN +template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) +{ + EIGEN_DEBUG_ALIGNED_LOAD + Packet16uc MSQ, LSQ; + Packet16uc mask; + MSQ = vec_ld(0, (unsigned char *)from); // most significant quadword + LSQ = vec_ld(15, (unsigned char *)from); // least significant quadword + mask = vec_lvsl(0, from); // create the permute mask + return (Packet4f) vec_perm(MSQ, LSQ, mask); // align the data + +} +template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from) +{ + EIGEN_DEBUG_ALIGNED_LOAD + // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html + Packet16uc MSQ, LSQ; + Packet16uc mask; + MSQ = vec_ld(0, (unsigned char *)from); // most significant quadword + LSQ = vec_ld(15, (unsigned char *)from); // least significant quadword + mask = vec_lvsl(0, from); // create the permute mask + return (Packet4i) vec_perm(MSQ, LSQ, mask); // align the data +} +#else +// We also need ot redefine little endian loading of Packet4i/Packet4f using VSX +template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from) +{ + EIGEN_DEBUG_UNALIGNED_LOAD + return (Packet4i) vec_vsx_ld((long)from & 15, (const int*) _EIGEN_ALIGNED_PTR(from)); +} +template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) +{ + EIGEN_DEBUG_UNALIGNED_LOAD + return (Packet4f) vec_vsx_ld((long)from & 15, (const float*) _EIGEN_ALIGNED_PTR(from)); +} +#endif + +template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from) +{ + Packet4f p; + if((std::ptrdiff_t(from) % 16) == 0) p = pload<Packet4f>(from); + else p = ploadu<Packet4f>(from); + return vec_perm(p, p, p16uc_DUPLICATE32_HI); +} +template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int* from) +{ + Packet4i p; + if((std::ptrdiff_t(from) % 16) == 0) p = pload<Packet4i>(from); + else p = ploadu<Packet4i>(from); + return vec_perm(p, p, p16uc_DUPLICATE32_HI); +} + +#ifdef _BIG_ENDIAN +template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from) +{ + EIGEN_DEBUG_UNALIGNED_STORE + // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html + // Warning: not thread safe! + Packet16uc MSQ, LSQ, edges; + Packet16uc edgeAlign, align; + + MSQ = vec_ld(0, (unsigned char *)to); // most significant quadword + LSQ = vec_ld(15, (unsigned char *)to); // least significant quadword + edgeAlign = vec_lvsl(0, to); // permute map to extract edges + edges=vec_perm(LSQ,MSQ,edgeAlign); // extract the edges + align = vec_lvsr( 0, to ); // permute map to misalign data + MSQ = vec_perm(edges,(Packet16uc)from,align); // misalign the data (MSQ) + LSQ = vec_perm((Packet16uc)from,edges,align); // misalign the data (LSQ) + vec_st( LSQ, 15, (unsigned char *)to ); // Store the LSQ part first + vec_st( MSQ, 0, (unsigned char *)to ); // Store the MSQ part +} +template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet4i& from) +{ + EIGEN_DEBUG_UNALIGNED_STORE + // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html + // Warning: not thread safe! + Packet16uc MSQ, LSQ, edges; + Packet16uc edgeAlign, align; + + MSQ = vec_ld(0, (unsigned char *)to); // most significant quadword + LSQ = vec_ld(15, (unsigned char *)to); // least significant quadword + edgeAlign = vec_lvsl(0, to); // permute map to extract edges + edges=vec_perm(LSQ, MSQ, edgeAlign); // extract the edges + align = vec_lvsr( 0, to ); // permute map to misalign data + MSQ = vec_perm(edges, (Packet16uc) from, align); // misalign the data (MSQ) + LSQ = vec_perm((Packet16uc) from, edges, align); // misalign the data (LSQ) + vec_st( LSQ, 15, (unsigned char *)to ); // Store the LSQ part first + vec_st( MSQ, 0, (unsigned char *)to ); // Store the MSQ part +} +#else +// We also need ot redefine little endian loading of Packet4i/Packet4f using VSX +template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet4i& from) +{ + EIGEN_DEBUG_ALIGNED_STORE + vec_vsx_st(from, (long)to & 15, (int*) _EIGEN_ALIGNED_PTR(to)); +} +template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from) +{ + EIGEN_DEBUG_ALIGNED_STORE + vec_vsx_st(from, (long)to & 15, (float*) _EIGEN_ALIGNED_PTR(to)); +} +#endif + +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 Packet4f preverse(const Packet4f& a) +{ + return reinterpret_cast<Packet4f>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE32)); +} +template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) +{ + return reinterpret_cast<Packet4i>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE32)); } + +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<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a) +{ + Packet4f b, sum; + b = vec_sld(a, a, 8); + sum = a + b; + b = vec_sld(sum, sum, 4); + sum += b; + return pfirst(sum); +} + +template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs) +{ + Packet4f v[4], sum[4]; + + // It's easier and faster to transpose then add as columns + // Check: http://www.freevec.org/function/matrix_4x4_transpose_floats for explanation + // Do the transpose, first set of moves + v[0] = vec_mergeh(vecs[0], vecs[2]); + v[1] = vec_mergel(vecs[0], vecs[2]); + v[2] = vec_mergeh(vecs[1], vecs[3]); + v[3] = vec_mergel(vecs[1], vecs[3]); + // Get the resulting vectors + sum[0] = vec_mergeh(v[0], v[2]); + sum[1] = vec_mergel(v[0], v[2]); + sum[2] = vec_mergeh(v[1], v[3]); + sum[3] = vec_mergel(v[1], v[3]); + + // Now do the summation: + // Lines 0+1 + sum[0] = sum[0] + sum[1]; + // Lines 2+3 + sum[1] = sum[2] + sum[3]; + // Add the results + sum[0] = sum[0] + sum[1]; + + return sum[0]; +} + +template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a) +{ + Packet4i sum; + sum = vec_sums(a, p4i_ZERO); +#ifdef _BIG_ENDIAN + sum = vec_sld(sum, p4i_ZERO, 12); +#else + sum = vec_sld(p4i_ZERO, sum, 4); +#endif + return pfirst(sum); +} + +template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs) +{ + Packet4i v[4], sum[4]; + + // It's easier and faster to transpose then add as columns + // Check: http://www.freevec.org/function/matrix_4x4_transpose_floats for explanation + // Do the transpose, first set of moves + v[0] = vec_mergeh(vecs[0], vecs[2]); + v[1] = vec_mergel(vecs[0], vecs[2]); + v[2] = vec_mergeh(vecs[1], vecs[3]); + v[3] = vec_mergel(vecs[1], vecs[3]); + // Get the resulting vectors + sum[0] = vec_mergeh(v[0], v[2]); + sum[1] = vec_mergel(v[0], v[2]); + sum[2] = vec_mergeh(v[1], v[3]); + sum[3] = vec_mergel(v[1], v[3]); + + // Now do the summation: + // Lines 0+1 + sum[0] = sum[0] + sum[1]; + // Lines 2+3 + sum[1] = sum[2] + sum[3]; + // Add the results + sum[0] = sum[0] + sum[1]; + + return sum[0]; +} + +// Other reduction functions: +// mul +template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a) +{ + Packet4f prod; + prod = pmul(a, vec_sld(a, a, 8)); + return pfirst(pmul(prod, vec_sld(prod, prod, 4))); +} + +template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a) +{ + EIGEN_ALIGN16 int aux[4]; + pstore(aux, a); + return aux[0] * aux[1] * aux[2] * aux[3]; +} + +// min +template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a) +{ + Packet4f b, res; + b = vec_min(a, vec_sld(a, a, 8)); + res = vec_min(b, vec_sld(b, b, 4)); + return pfirst(res); +} + +template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a) +{ + Packet4i b, res; + b = vec_min(a, vec_sld(a, a, 8)); + res = vec_min(b, vec_sld(b, b, 4)); + return pfirst(res); +} + +// max +template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a) +{ + Packet4f b, res; + b = vec_max(a, vec_sld(a, a, 8)); + res = vec_max(b, vec_sld(b, b, 4)); + return pfirst(res); +} + +template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a) +{ + Packet4i b, res; + b = vec_max(a, vec_sld(a, a, 8)); + res = vec_max(b, vec_sld(b, b, 4)); + return pfirst(res); +} + +template<int Offset> +struct palign_impl<Offset,Packet4f> +{ + static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second) + { +#ifdef _BIG_ENDIAN + switch (Offset % 4) { + case 1: + first = vec_sld(first, second, 4); break; + case 2: + first = vec_sld(first, second, 8); break; + case 3: + first = vec_sld(first, second, 12); break; + } +#else + switch (Offset % 4) { + case 1: + first = vec_sld(second, first, 12); break; + case 2: + first = vec_sld(second, first, 8); break; + case 3: + first = vec_sld(second, first, 4); break; + } +#endif + } +}; + +template<int Offset> +struct palign_impl<Offset,Packet4i> +{ + static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second) + { +#ifdef _BIG_ENDIAN + switch (Offset % 4) { + case 1: + first = vec_sld(first, second, 4); break; + case 2: + first = vec_sld(first, second, 8); break; + case 3: + first = vec_sld(first, second, 12); break; + } +#else + switch (Offset % 4) { + case 1: + first = vec_sld(second, first, 12); break; + case 2: + first = vec_sld(second, first, 8); break; + case 3: + first = vec_sld(second, first, 4); break; + } +#endif + } +}; + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet4f,4>& kernel) { + Packet4f t0, t1, t2, t3; + t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]); + t1 = vec_mergel(kernel.packet[0], kernel.packet[2]); + t2 = vec_mergeh(kernel.packet[1], kernel.packet[3]); + t3 = vec_mergel(kernel.packet[1], kernel.packet[3]); + kernel.packet[0] = vec_mergeh(t0, t2); + kernel.packet[1] = vec_mergel(t0, t2); + kernel.packet[2] = vec_mergeh(t1, t3); + kernel.packet[3] = vec_mergel(t1, t3); +} + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet4i,4>& kernel) { + Packet4i t0, t1, t2, t3; + t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]); + t1 = vec_mergel(kernel.packet[0], kernel.packet[2]); + t2 = vec_mergeh(kernel.packet[1], kernel.packet[3]); + t3 = vec_mergel(kernel.packet[1], kernel.packet[3]); + kernel.packet[0] = vec_mergeh(t0, t2); + kernel.packet[1] = vec_mergel(t0, t2); + kernel.packet[2] = vec_mergeh(t1, t3); + kernel.packet[3] = vec_mergel(t1, t3); +} + +template<> EIGEN_STRONG_INLINE Packet4i pblend(const Selector<4>& ifPacket, const Packet4i& thenPacket, const Packet4i& elsePacket) { + Packet4ui select = { ifPacket.select[0], ifPacket.select[1], ifPacket.select[2], ifPacket.select[3] }; + Packet4ui mask = reinterpret_cast<Packet4ui>(vec_cmpeq(reinterpret_cast<Packet4ui>(select), reinterpret_cast<Packet4ui>(p4i_ONE))); + return vec_sel(elsePacket, thenPacket, mask); +} + +template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket, const Packet4f& elsePacket) { + Packet4ui select = { ifPacket.select[0], ifPacket.select[1], ifPacket.select[2], ifPacket.select[3] }; + Packet4ui mask = reinterpret_cast<Packet4ui>(vec_cmpeq(reinterpret_cast<Packet4ui>(select), reinterpret_cast<Packet4ui>(p4i_ONE))); + return vec_sel(elsePacket, thenPacket, mask); +} + + +//---------- double ---------- +#ifdef __VSX__ +typedef __vector double Packet2d; +typedef __vector unsigned long long Packet2ul; +typedef __vector long long Packet2l; +#if EIGEN_COMP_CLANG +typedef Packet2ul Packet2bl; +#else +typedef __vector __bool long Packet2bl; +#endif + +static Packet2l p2l_ONE = { 1, 1 }; +static Packet2l p2l_ZERO = reinterpret_cast<Packet2l>(p4i_ZERO); +static Packet2d p2d_ONE = { 1.0, 1.0 }; +static Packet2d p2d_ZERO = reinterpret_cast<Packet2d>(p4f_ZERO); +static Packet2d p2d_MZERO = { -0.0, -0.0 }; + +#ifdef _BIG_ENDIAN +static Packet2d p2d_COUNTDOWN = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(p2d_ZERO), reinterpret_cast<Packet4f>(p2d_ONE), 8)); +#else +static Packet2d p2d_COUNTDOWN = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(p2d_ONE), reinterpret_cast<Packet4f>(p2d_ZERO), 8)); +#endif + +template<int index> Packet2d vec_splat_dbl(Packet2d& a); + +template<> EIGEN_STRONG_INLINE Packet2d vec_splat_dbl<0>(Packet2d& a) +{ + return reinterpret_cast<Packet2d>(vec_perm(a, a, p16uc_PSET64_HI)); +} + +template<> EIGEN_STRONG_INLINE Packet2d vec_splat_dbl<1>(Packet2d& a) +{ + return reinterpret_cast<Packet2d>(vec_perm(a, a, p16uc_PSET64_LO)); +} + +template<> struct packet_traits<double> : default_packet_traits +{ + typedef Packet2d type; + typedef Packet2d half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=2, + HasHalfPacket = 1, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasMin = 1, + HasMax = 1, + HasAbs = 1, + HasSin = 0, + HasCos = 0, + HasLog = 0, + HasExp = 1, + HasSqrt = 1, + HasRsqrt = 1, + HasRound = 1, + HasFloor = 1, + HasCeil = 1, + HasNegate = 1, + HasBlend = 1 + }; +}; + +template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16}; typedef Packet2d half; }; + +inline std::ostream & operator <<(std::ostream & s, const Packet2l & v) +{ + union { + Packet2l v; + int64_t n[2]; + } vt; + vt.v = v; + s << vt.n[0] << ", " << vt.n[1]; + return s; +} + +inline std::ostream & operator <<(std::ostream & s, const Packet2d & v) +{ + union { + Packet2d v; + double n[2]; + } vt; + vt.v = v; + s << vt.n[0] << ", " << vt.n[1]; + return s; +} + +// Need to define them first or we get specialization after instantiation errors +template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from) +{ + EIGEN_DEBUG_ALIGNED_LOAD +#ifdef __VSX__ + return vec_vsx_ld(0, from); +#else + return vec_ld(0, from); +#endif +} + +template<> EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet2d& from) +{ + EIGEN_DEBUG_ALIGNED_STORE +#ifdef __VSX__ + vec_vsx_st(from, 0, to); +#else + vec_st(from, 0, to); +#endif +} + +template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { + Packet2d v = {from, from}; + return v; +} + +template<> EIGEN_STRONG_INLINE void +pbroadcast4<Packet2d>(const double *a, + Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3) +{ + a1 = pload<Packet2d>(a); + a0 = vec_splat_dbl<0>(a1); + a1 = vec_splat_dbl<1>(a1); + a3 = pload<Packet2d>(a+2); + a2 = vec_splat_dbl<0>(a3); + a3 = vec_splat_dbl<1>(a3); +} + +template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride) +{ + double EIGEN_ALIGN16 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]; + pstore<double>(af, from); + to[0*stride] = af[0]; + to[1*stride] = af[1]; +} + +template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a) { return pset1<Packet2d>(a) + p2d_COUNTDOWN; } + +template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return a + b; } + +template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return a - b; } + +template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) { return p2d_ZERO - a; } + +template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; } + +template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_madd(a,b,p2d_MZERO); } +template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_div(a,b); } + +// for some weird raisons, it has to be overloaded for packet of integers +template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vec_madd(a, b, c); } + +template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_min(a, b); } + +template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_max(a, b); } + +template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, b); } + +template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_or(a, b); } + +template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_xor(a, b); } + +template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, vec_nor(b, b)); } + +template<> EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a) { return vec_round(a); } +template<> EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const Packet2d& a) { return vec_ceil(a); } +template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) { return vec_floor(a); } + +template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from) +{ + EIGEN_DEBUG_ALIGNED_LOAD + return (Packet2d) vec_vsx_ld((long)from & 15, (const double*) _EIGEN_ALIGNED_PTR(from)); +} + +template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double* from) +{ + Packet2d p; + if((std::ptrdiff_t(from) % 16) == 0) p = pload<Packet2d>(from); + else p = ploadu<Packet2d>(from); + return vec_splat_dbl<0>(p); +} + +template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d& from) +{ + EIGEN_DEBUG_ALIGNED_STORE + vec_vsx_st((Packet4f)from, (long)to & 15, (float*) _EIGEN_ALIGNED_PTR(to)); +} + +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 Packet2d preverse(const Packet2d& a) +{ + return reinterpret_cast<Packet2d>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE64)); +} +template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a) { return vec_abs(a); } + +template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) +{ + Packet2d b, sum; + b = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(a), reinterpret_cast<Packet4f>(a), 8)); + sum = a + b; + return pfirst<Packet2d>(sum); +} + +template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs) +{ + Packet2d v[2], sum; + v[0] = vecs[0] + reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(vecs[0]), reinterpret_cast<Packet4f>(vecs[0]), 8)); + v[1] = vecs[1] + reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(vecs[1]), reinterpret_cast<Packet4f>(vecs[1]), 8)); + +#ifdef _BIG_ENDIAN + sum = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(v[0]), reinterpret_cast<Packet4f>(v[1]), 8)); +#else + sum = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(v[1]), reinterpret_cast<Packet4f>(v[0]), 8)); +#endif + + return sum; +} +// Other reduction functions: +// mul +template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a) +{ + return pfirst(pmul(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(a), reinterpret_cast<Packet4ui>(a), 8)))); +} + +// min +template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a) +{ + return pfirst(pmin(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(a), reinterpret_cast<Packet4ui>(a), 8)))); +} + +// max +template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a) +{ + return pfirst(pmax(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(a), reinterpret_cast<Packet4ui>(a), 8)))); +} + +template<int Offset> +struct palign_impl<Offset,Packet2d> +{ + static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second) + { + if (Offset == 1) +#ifdef _BIG_ENDIAN + first = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(first), reinterpret_cast<Packet4ui>(second), 8)); +#else + first = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(second), reinterpret_cast<Packet4ui>(first), 8)); +#endif + } +}; + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet2d,2>& kernel) { + Packet2d t0, t1; + t0 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_HI); + t1 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_LO); + kernel.packet[0] = t0; + kernel.packet[1] = t1; +} + +template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket, const Packet2d& elsePacket) { + Packet2l select = { ifPacket.select[0], ifPacket.select[1] }; + Packet2bl mask = vec_cmpeq(reinterpret_cast<Packet2d>(select), reinterpret_cast<Packet2d>(p2l_ONE)); + return vec_sel(elsePacket, thenPacket, mask); +} +#endif // __VSX__ +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_PACKET_MATH_ALTIVEC_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/Complex.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/Complex.h new file mode 100644 index 000000000..9c2536509 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/Complex.h @@ -0,0 +1,103 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com> +// +// 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_CUDA_H +#define EIGEN_COMPLEX_CUDA_H + +// clang-format off + +namespace Eigen { + +namespace internal { + +#if defined(__CUDACC__) && defined(EIGEN_USE_GPU) + +// Many std::complex methods such as operator+, operator-, operator* and +// operator/ are not constexpr. Due to this, clang does not treat them as device +// functions and thus Eigen functors making use of these operators fail to +// compile. Here, we manually specialize these functors for complex types when +// building for CUDA to avoid non-constexpr methods. + +// Sum +template<typename T> struct scalar_sum_op<const std::complex<T>, const std::complex<T> > : binary_op_base<const std::complex<T>, const std::complex<T> > { + typedef typename std::complex<T> result_type; + + EIGEN_EMPTY_STRUCT_CTOR(scalar_sum_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::complex<T> operator() (const std::complex<T>& a, const std::complex<T>& b) const { + return std::complex<T>(numext::real(a) + numext::real(b), + numext::imag(a) + numext::imag(b)); + } +}; + +template<typename T> struct scalar_sum_op<std::complex<T>, std::complex<T> > : scalar_sum_op<const std::complex<T>, const std::complex<T> > {}; + + +// Difference +template<typename T> struct scalar_difference_op<const std::complex<T>, const std::complex<T> > : binary_op_base<const std::complex<T>, const std::complex<T> > { + typedef typename std::complex<T> result_type; + + EIGEN_EMPTY_STRUCT_CTOR(scalar_difference_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::complex<T> operator() (const std::complex<T>& a, const std::complex<T>& b) const { + return std::complex<T>(numext::real(a) - numext::real(b), + numext::imag(a) - numext::imag(b)); + } +}; + +template<typename T> struct scalar_difference_op<std::complex<T>, std::complex<T> > : scalar_difference_op<const std::complex<T>, const std::complex<T> > {}; + + +// Product +template<typename T> struct scalar_product_op<const std::complex<T>, const std::complex<T> > : binary_op_base<const std::complex<T>, const std::complex<T> > { + enum { + Vectorizable = packet_traits<std::complex<T>>::HasMul + }; + typedef typename std::complex<T> result_type; + + EIGEN_EMPTY_STRUCT_CTOR(scalar_product_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::complex<T> operator() (const std::complex<T>& a, const std::complex<T>& b) const { + const T a_real = numext::real(a); + const T a_imag = numext::imag(a); + const T b_real = numext::real(b); + const T b_imag = numext::imag(b); + return std::complex<T>(a_real * b_real - a_imag * b_imag, + a_real * b_imag + a_imag * b_real); + } +}; + +template<typename T> struct scalar_product_op<std::complex<T>, std::complex<T> > : scalar_product_op<const std::complex<T>, const std::complex<T> > {}; + + +// Quotient +template<typename T> struct scalar_quotient_op<const std::complex<T>, const std::complex<T> > : binary_op_base<const std::complex<T>, const std::complex<T> > { + enum { + Vectorizable = packet_traits<std::complex<T>>::HasDiv + }; + typedef typename std::complex<T> result_type; + + EIGEN_EMPTY_STRUCT_CTOR(scalar_quotient_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::complex<T> operator() (const std::complex<T>& a, const std::complex<T>& b) const { + const T a_real = numext::real(a); + const T a_imag = numext::imag(a); + const T b_real = numext::real(b); + const T b_imag = numext::imag(b); + const T norm = T(1) / (b_real * b_real + b_imag * b_imag); + return std::complex<T>((a_real * b_real + a_imag * b_imag) * norm, + (a_imag * b_real - a_real * b_imag) * norm); + } +}; + +template<typename T> struct scalar_quotient_op<std::complex<T>, std::complex<T> > : scalar_quotient_op<const std::complex<T>, const std::complex<T> > {}; + +#endif + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_COMPLEX_CUDA_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/Half.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/Half.h new file mode 100644 index 000000000..294c517ea --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/Half.h @@ -0,0 +1,635 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// 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 conversion routines are Copyright (c) Fabian Giesen, 2016. +// The original license follows: +// +// Copyright (c) Fabian Giesen, 2016 +// All rights reserved. +// Redistribution and use in source and binary forms, with or without +// modification, are permitted. +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +// HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + + +// Standard 16-bit float type, mostly useful for GPUs. Defines a new +// type Eigen::half (inheriting from CUDA's __half struct) with +// operator overloads such that it behaves basically as an arithmetic +// type. It will be quite slow on CPUs (so it is recommended to stay +// in fp32 for CPUs, except for simple parameter conversions, I/O +// to disk and the likes), but fast on GPUs. + + +#ifndef EIGEN_HALF_CUDA_H +#define EIGEN_HALF_CUDA_H + +#if __cplusplus > 199711L +#define EIGEN_EXPLICIT_CAST(tgt_type) explicit operator tgt_type() +#else +#define EIGEN_EXPLICIT_CAST(tgt_type) operator tgt_type() +#endif + + +namespace Eigen { + +struct half; + +namespace half_impl { + +#if !defined(EIGEN_HAS_CUDA_FP16) + +// Make our own __half definition that is similar to CUDA's. +struct __half { + EIGEN_DEVICE_FUNC __half() {} + explicit EIGEN_DEVICE_FUNC __half(unsigned short raw) : x(raw) {} + unsigned short x; +}; + +#endif + +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half raw_uint16_to_half(unsigned short x); +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half float_to_half_rtne(float ff); +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half h); + +struct half_base : public __half { + EIGEN_DEVICE_FUNC half_base() {} + EIGEN_DEVICE_FUNC half_base(const half_base& h) : __half(h) {} + EIGEN_DEVICE_FUNC half_base(const __half& h) : __half(h) {} +}; + +} // namespace half_impl + +// Class definition. +struct half : public half_impl::half_base { + #if !defined(EIGEN_HAS_CUDA_FP16) + typedef half_impl::__half __half; + #endif + + EIGEN_DEVICE_FUNC half() {} + + EIGEN_DEVICE_FUNC half(const __half& h) : half_impl::half_base(h) {} + EIGEN_DEVICE_FUNC half(const half& h) : half_impl::half_base(h) {} + + explicit EIGEN_DEVICE_FUNC half(bool b) + : half_impl::half_base(half_impl::raw_uint16_to_half(b ? 0x3c00 : 0)) {} + template<class T> + explicit EIGEN_DEVICE_FUNC half(const T& val) + : half_impl::half_base(half_impl::float_to_half_rtne(static_cast<float>(val))) {} + explicit EIGEN_DEVICE_FUNC half(float f) + : half_impl::half_base(half_impl::float_to_half_rtne(f)) {} + + EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(bool) const { + // +0.0 and -0.0 become false, everything else becomes true. + return (x & 0x7fff) != 0; + } + EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(signed char) const { + return static_cast<signed char>(half_impl::half_to_float(*this)); + } + EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned char) const { + return static_cast<unsigned char>(half_impl::half_to_float(*this)); + } + EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(short) const { + return static_cast<short>(half_impl::half_to_float(*this)); + } + EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned short) const { + return static_cast<unsigned short>(half_impl::half_to_float(*this)); + } + EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(int) const { + return static_cast<int>(half_impl::half_to_float(*this)); + } + EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned int) const { + return static_cast<unsigned int>(half_impl::half_to_float(*this)); + } + EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(long) const { + return static_cast<long>(half_impl::half_to_float(*this)); + } + EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned long) const { + return static_cast<unsigned long>(half_impl::half_to_float(*this)); + } + EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(long long) const { + return static_cast<long long>(half_impl::half_to_float(*this)); + } + EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned long long) const { + return static_cast<unsigned long long>(half_to_float(*this)); + } + EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(float) const { + return half_impl::half_to_float(*this); + } + EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(double) const { + return static_cast<double>(half_impl::half_to_float(*this)); + } + + EIGEN_DEVICE_FUNC half& operator=(const half& other) { + x = other.x; + return *this; + } +}; + +namespace half_impl { + +#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530 + +// Intrinsics for native fp16 support. Note that on current hardware, +// these are no faster than fp32 arithmetic (you need to use the half2 +// versions to get the ALU speed increased), but you do save the +// conversion steps back and forth. + +__device__ half operator + (const half& a, const half& b) { + return __hadd(a, b); +} +__device__ half operator * (const half& a, const half& b) { + return __hmul(a, b); +} +__device__ half operator - (const half& a, const half& b) { + return __hsub(a, b); +} +__device__ half operator / (const half& a, const half& b) { + float num = __half2float(a); + float denom = __half2float(b); + return __float2half(num / denom); +} +__device__ half operator - (const half& a) { + return __hneg(a); +} +__device__ half& operator += (half& a, const half& b) { + a = a + b; + return a; +} +__device__ half& operator *= (half& a, const half& b) { + a = a * b; + return a; +} +__device__ half& operator -= (half& a, const half& b) { + a = a - b; + return a; +} +__device__ half& operator /= (half& a, const half& b) { + a = a / b; + return a; +} +__device__ bool operator == (const half& a, const half& b) { + return __heq(a, b); +} +__device__ bool operator != (const half& a, const half& b) { + return __hne(a, b); +} +__device__ bool operator < (const half& a, const half& b) { + return __hlt(a, b); +} +__device__ bool operator <= (const half& a, const half& b) { + return __hle(a, b); +} +__device__ bool operator > (const half& a, const half& b) { + return __hgt(a, b); +} +__device__ bool operator >= (const half& a, const half& b) { + return __hge(a, b); +} + +#else // Emulate support for half floats + +// Definitions for CPUs and older CUDA, mostly working through conversion +// to/from fp32. + +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator + (const half& a, const half& b) { + return half(float(a) + float(b)); +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator * (const half& a, const half& b) { + return half(float(a) * float(b)); +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator - (const half& a, const half& b) { + return half(float(a) - float(b)); +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator / (const half& a, const half& b) { + return half(float(a) / float(b)); +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator - (const half& a) { + half result; + result.x = a.x ^ 0x8000; + return result; +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator += (half& a, const half& b) { + a = half(float(a) + float(b)); + return a; +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator *= (half& a, const half& b) { + a = half(float(a) * float(b)); + return a; +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator -= (half& a, const half& b) { + a = half(float(a) - float(b)); + return a; +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator /= (half& a, const half& b) { + a = half(float(a) / float(b)); + return a; +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator == (const half& a, const half& b) { + return float(a) == float(b); +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator != (const half& a, const half& b) { + return float(a) != float(b); +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator < (const half& a, const half& b) { + return float(a) < float(b); +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator <= (const half& a, const half& b) { + return float(a) <= float(b); +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator > (const half& a, const half& b) { + return float(a) > float(b); +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator >= (const half& a, const half& b) { + return float(a) >= float(b); +} + +#endif // Emulate support for half floats + +// Division by an index. Do it in full float precision to avoid accuracy +// issues in converting the denominator to half. +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator / (const half& a, Index b) { + return half(static_cast<float>(a) / static_cast<float>(b)); +} + +// Conversion routines, including fallbacks for the host or older CUDA. +// Note that newer Intel CPUs (Haswell or newer) have vectorized versions of +// these in hardware. If we need more performance on older/other CPUs, they are +// also possible to vectorize directly. + +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half raw_uint16_to_half(unsigned short x) { + __half h; + h.x = x; + return h; +} + +union FP32 { + unsigned int u; + float f; +}; + +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half float_to_half_rtne(float ff) { +#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300 + return __float2half(ff); + +#elif defined(EIGEN_HAS_FP16_C) + __half h; + h.x = _cvtss_sh(ff, 0); + return h; + +#else + FP32 f; f.f = ff; + + const FP32 f32infty = { 255 << 23 }; + const FP32 f16max = { (127 + 16) << 23 }; + const FP32 denorm_magic = { ((127 - 15) + (23 - 10) + 1) << 23 }; + unsigned int sign_mask = 0x80000000u; + __half o; + o.x = static_cast<unsigned short>(0x0u); + + unsigned int sign = f.u & sign_mask; + f.u ^= sign; + + // NOTE all the integer compares in this function can be safely + // compiled into signed compares since all operands are below + // 0x80000000. Important if you want fast straight SSE2 code + // (since there's no unsigned PCMPGTD). + + if (f.u >= f16max.u) { // result is Inf or NaN (all exponent bits set) + o.x = (f.u > f32infty.u) ? 0x7e00 : 0x7c00; // NaN->qNaN and Inf->Inf + } else { // (De)normalized number or zero + if (f.u < (113 << 23)) { // resulting FP16 is subnormal or zero + // use a magic value to align our 10 mantissa bits at the bottom of + // the float. as long as FP addition is round-to-nearest-even this + // just works. + f.f += denorm_magic.f; + + // and one integer subtract of the bias later, we have our final float! + o.x = static_cast<unsigned short>(f.u - denorm_magic.u); + } else { + unsigned int mant_odd = (f.u >> 13) & 1; // resulting mantissa is odd + + // update exponent, rounding bias part 1 + f.u += ((unsigned int)(15 - 127) << 23) + 0xfff; + // rounding bias part 2 + f.u += mant_odd; + // take the bits! + o.x = static_cast<unsigned short>(f.u >> 13); + } + } + + o.x |= static_cast<unsigned short>(sign >> 16); + return o; +#endif +} + +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half h) { +#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300 + return __half2float(h); + +#elif defined(EIGEN_HAS_FP16_C) + return _cvtsh_ss(h.x); + +#else + const FP32 magic = { 113 << 23 }; + const unsigned int shifted_exp = 0x7c00 << 13; // exponent mask after shift + FP32 o; + + o.u = (h.x & 0x7fff) << 13; // exponent/mantissa bits + unsigned int exp = shifted_exp & o.u; // just the exponent + o.u += (127 - 15) << 23; // exponent adjust + + // handle exponent special cases + if (exp == shifted_exp) { // Inf/NaN? + o.u += (128 - 16) << 23; // extra exp adjust + } else if (exp == 0) { // Zero/Denormal? + o.u += 1 << 23; // extra exp adjust + o.f -= magic.f; // renormalize + } + + o.u |= (h.x & 0x8000) << 16; // sign bit + return o.f; +#endif +} + +// --- standard functions --- + +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isinf)(const half& a) { + return (a.x & 0x7fff) == 0x7c00; +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isnan)(const half& a) { +#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530 + return __hisnan(a); +#else + return (a.x & 0x7fff) > 0x7c00; +#endif +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isfinite)(const half& a) { + return !(isinf EIGEN_NOT_A_MACRO (a)) && !(isnan EIGEN_NOT_A_MACRO (a)); +} + +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half abs(const half& a) { + half result; + result.x = a.x & 0x7FFF; + return result; +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half exp(const half& a) { + return half(::expf(float(a))); +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log(const half& a) { +#if defined(EIGEN_HAS_CUDA_FP16) && defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530 + return Eigen::half(::hlog(a)); +#else + return half(::logf(float(a))); +#endif +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log1p(const half& a) { + return half(numext::log1p(float(a))); +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log10(const half& a) { + return half(::log10f(float(a))); +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half sqrt(const half& a) { + return half(::sqrtf(float(a))); +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half pow(const half& a, const half& b) { + return half(::powf(float(a), float(b))); +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half sin(const half& a) { + return half(::sinf(float(a))); +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half cos(const half& a) { + return half(::cosf(float(a))); +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half tan(const half& a) { + return half(::tanf(float(a))); +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half tanh(const half& a) { + return half(::tanhf(float(a))); +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half floor(const half& a) { + return half(::floorf(float(a))); +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half ceil(const half& a) { + return half(::ceilf(float(a))); +} + +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half (min)(const half& a, const half& b) { +#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530 + return __hlt(b, a) ? b : a; +#else + const float f1 = static_cast<float>(a); + const float f2 = static_cast<float>(b); + return f2 < f1 ? b : a; +#endif +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half (max)(const half& a, const half& b) { +#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530 + return __hlt(a, b) ? b : a; +#else + const float f1 = static_cast<float>(a); + const float f2 = static_cast<float>(b); + return f1 < f2 ? b : a; +#endif +} + +EIGEN_ALWAYS_INLINE std::ostream& operator << (std::ostream& os, const half& v) { + os << static_cast<float>(v); + return os; +} + +} // end namespace half_impl + +// import Eigen::half_impl::half into Eigen namespace +// using half_impl::half; + +namespace internal { + +template<> +struct random_default_impl<half, false, false> +{ + static inline half run(const half& x, const half& y) + { + return x + (y-x) * half(float(std::rand()) / float(RAND_MAX)); + } + static inline half run() + { + return run(half(-1.f), half(1.f)); + } +}; + +template<> struct is_arithmetic<half> { enum { value = true }; }; + +} // end namespace internal + +} // end namespace Eigen + +namespace std { +template<> +struct numeric_limits<Eigen::half> { + static const bool is_specialized = true; + static const bool is_signed = true; + static const bool is_integer = false; + static const bool is_exact = false; + static const bool has_infinity = true; + static const bool has_quiet_NaN = true; + static const bool has_signaling_NaN = true; + static const float_denorm_style has_denorm = denorm_present; + static const bool has_denorm_loss = false; + static const std::float_round_style round_style = std::round_to_nearest; + static const bool is_iec559 = false; + static const bool is_bounded = false; + static const bool is_modulo = false; + static const int digits = 11; + static const int digits10 = 2; + //static const int max_digits10 = ; + static const int radix = 2; + static const int min_exponent = -13; + static const int min_exponent10 = -4; + static const int max_exponent = 16; + static const int max_exponent10 = 4; + static const bool traps = true; + static const bool tinyness_before = false; + + static Eigen::half (min)() { return Eigen::half_impl::raw_uint16_to_half(0x400); } + static Eigen::half lowest() { return Eigen::half_impl::raw_uint16_to_half(0xfbff); } + static Eigen::half (max)() { return Eigen::half_impl::raw_uint16_to_half(0x7bff); } + static Eigen::half epsilon() { return Eigen::half_impl::raw_uint16_to_half(0x0800); } + static Eigen::half round_error() { return Eigen::half(0.5); } + static Eigen::half infinity() { return Eigen::half_impl::raw_uint16_to_half(0x7c00); } + static Eigen::half quiet_NaN() { return Eigen::half_impl::raw_uint16_to_half(0x7e00); } + static Eigen::half signaling_NaN() { return Eigen::half_impl::raw_uint16_to_half(0x7e00); } + static Eigen::half denorm_min() { return Eigen::half_impl::raw_uint16_to_half(0x1); } +}; +} + +namespace Eigen { + +template<> struct NumTraits<Eigen::half> + : GenericNumTraits<Eigen::half> +{ + enum { + IsSigned = true, + IsInteger = false, + IsComplex = false, + RequireInitialization = false + }; + + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half epsilon() { + return half_impl::raw_uint16_to_half(0x0800); + } + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half dummy_precision() { return Eigen::half(1e-2f); } + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half highest() { + return half_impl::raw_uint16_to_half(0x7bff); + } + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half lowest() { + return half_impl::raw_uint16_to_half(0xfbff); + } + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half infinity() { + return half_impl::raw_uint16_to_half(0x7c00); + } + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half quiet_NaN() { + return half_impl::raw_uint16_to_half(0x7c01); + } +}; + +} // end namespace Eigen + +// C-like standard mathematical functions and trancendentals. +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half fabsh(const Eigen::half& a) { + Eigen::half result; + result.x = a.x & 0x7FFF; + return result; +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half exph(const Eigen::half& a) { + return Eigen::half(::expf(float(a))); +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half logh(const Eigen::half& a) { +#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530 + return Eigen::half(::hlog(a)); +#else + return Eigen::half(::logf(float(a))); +#endif +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half sqrth(const Eigen::half& a) { + return Eigen::half(::sqrtf(float(a))); +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half powh(const Eigen::half& a, const Eigen::half& b) { + return Eigen::half(::powf(float(a), float(b))); +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half floorh(const Eigen::half& a) { + return Eigen::half(::floorf(float(a))); +} +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half ceilh(const Eigen::half& a) { + return Eigen::half(::ceilf(float(a))); +} + +namespace std { + +#if __cplusplus > 199711L +template <> +struct hash<Eigen::half> { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::size_t operator()(const Eigen::half& a) const { + return static_cast<std::size_t>(a.x); + } +}; +#endif + +} // end namespace std + + +// Add the missing shfl_xor intrinsic +#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300 +__device__ EIGEN_STRONG_INLINE Eigen::half __shfl_xor(Eigen::half var, int laneMask, int width=warpSize) { + return static_cast<Eigen::half>(__shfl_xor(static_cast<float>(var), laneMask, width)); +} +#endif + +// ldg() has an overload for __half, but we also need one for Eigen::half. +#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350 +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half __ldg(const Eigen::half* ptr) { + return Eigen::half_impl::raw_uint16_to_half( + __ldg(reinterpret_cast<const unsigned short*>(ptr))); +} +#endif + + +#if defined(__CUDA_ARCH__) +namespace Eigen { +namespace numext { + +template<> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +bool (isnan)(const Eigen::half& h) { + return (half_impl::isnan)(h); +} + +template<> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +bool (isinf)(const Eigen::half& h) { + return (half_impl::isinf)(h); +} + +template<> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +bool (isfinite)(const Eigen::half& h) { + return (half_impl::isfinite)(h); +} + +} // namespace Eigen +} // namespace numext +#endif + +#endif // EIGEN_HALF_CUDA_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/MathFunctions.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/MathFunctions.h new file mode 100644 index 000000000..0348b41db --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/MathFunctions.h @@ -0,0 +1,91 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com> +// +// 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_MATH_FUNCTIONS_CUDA_H +#define EIGEN_MATH_FUNCTIONS_CUDA_H + +namespace Eigen { + +namespace internal { + +// Make sure this is only available when targeting a GPU: we don't want to +// introduce conflicts between these packet_traits definitions and the ones +// we'll use on the host side (SSE, AVX, ...) +#if defined(__CUDACC__) && defined(EIGEN_USE_GPU) +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +float4 plog<float4>(const float4& a) +{ + return make_float4(logf(a.x), logf(a.y), logf(a.z), logf(a.w)); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +double2 plog<double2>(const double2& a) +{ + using ::log; + return make_double2(log(a.x), log(a.y)); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +float4 plog1p<float4>(const float4& a) +{ + return make_float4(log1pf(a.x), log1pf(a.y), log1pf(a.z), log1pf(a.w)); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +double2 plog1p<double2>(const double2& a) +{ + return make_double2(log1p(a.x), log1p(a.y)); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +float4 pexp<float4>(const float4& a) +{ + return make_float4(expf(a.x), expf(a.y), expf(a.z), expf(a.w)); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +double2 pexp<double2>(const double2& a) +{ + using ::exp; + return make_double2(exp(a.x), exp(a.y)); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +float4 psqrt<float4>(const float4& a) +{ + return make_float4(sqrtf(a.x), sqrtf(a.y), sqrtf(a.z), sqrtf(a.w)); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +double2 psqrt<double2>(const double2& a) +{ + using ::sqrt; + return make_double2(sqrt(a.x), sqrt(a.y)); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +float4 prsqrt<float4>(const float4& a) +{ + return make_float4(rsqrtf(a.x), rsqrtf(a.y), rsqrtf(a.z), rsqrtf(a.w)); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +double2 prsqrt<double2>(const double2& a) +{ + return make_double2(rsqrt(a.x), rsqrt(a.y)); +} + + +#endif + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_MATH_FUNCTIONS_CUDA_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/PacketMath.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/PacketMath.h new file mode 100644 index 000000000..4dda63188 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/PacketMath.h @@ -0,0 +1,333 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com> +// +// 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_PACKET_MATH_CUDA_H +#define EIGEN_PACKET_MATH_CUDA_H + +namespace Eigen { + +namespace internal { + +// Make sure this is only available when targeting a GPU: we don't want to +// introduce conflicts between these packet_traits definitions and the ones +// we'll use on the host side (SSE, AVX, ...) +#if defined(__CUDACC__) && defined(EIGEN_USE_GPU) +template<> struct is_arithmetic<float4> { enum { value = true }; }; +template<> struct is_arithmetic<double2> { enum { value = true }; }; + +template<> struct packet_traits<float> : default_packet_traits +{ + typedef float4 type; + typedef float4 half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=4, + HasHalfPacket = 0, + + HasDiv = 1, + HasSin = 0, + HasCos = 0, + HasLog = 1, + HasExp = 1, + HasSqrt = 1, + HasRsqrt = 1, + HasLGamma = 1, + HasDiGamma = 1, + HasZeta = 1, + HasPolygamma = 1, + HasErf = 1, + HasErfc = 1, + HasIGamma = 1, + HasIGammac = 1, + HasBetaInc = 1, + + HasBlend = 0, + }; +}; + +template<> struct packet_traits<double> : default_packet_traits +{ + typedef double2 type; + typedef double2 half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=2, + HasHalfPacket = 0, + + HasDiv = 1, + HasLog = 1, + HasExp = 1, + HasSqrt = 1, + HasRsqrt = 1, + HasLGamma = 1, + HasDiGamma = 1, + HasZeta = 1, + HasPolygamma = 1, + HasErf = 1, + HasErfc = 1, + HasIGamma = 1, + HasIGammac = 1, + HasBetaInc = 1, + + HasBlend = 0, + }; +}; + + +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<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pset1<float4>(const float& from) { + return make_float4(from, from, from, from); +} +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pset1<double2>(const double& from) { + return make_double2(from, from); +} + + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 plset<float4>(const float& a) { + return make_float4(a, a+1, a+2, a+3); +} +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 plset<double2>(const double& a) { + return make_double2(a, a+1); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 padd<float4>(const float4& a, const float4& b) { + return make_float4(a.x+b.x, a.y+b.y, a.z+b.z, a.w+b.w); +} +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 padd<double2>(const double2& a, const double2& b) { + return make_double2(a.x+b.x, a.y+b.y); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 psub<float4>(const float4& a, const float4& b) { + return make_float4(a.x-b.x, a.y-b.y, a.z-b.z, a.w-b.w); +} +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 psub<double2>(const double2& a, const double2& b) { + return make_double2(a.x-b.x, a.y-b.y); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pnegate(const float4& a) { + return make_float4(-a.x, -a.y, -a.z, -a.w); +} +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pnegate(const double2& a) { + return make_double2(-a.x, -a.y); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pconj(const float4& a) { return a; } +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pconj(const double2& a) { return a; } + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pmul<float4>(const float4& a, const float4& b) { + return make_float4(a.x*b.x, a.y*b.y, a.z*b.z, a.w*b.w); +} +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pmul<double2>(const double2& a, const double2& b) { + return make_double2(a.x*b.x, a.y*b.y); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pdiv<float4>(const float4& a, const float4& b) { + return make_float4(a.x/b.x, a.y/b.y, a.z/b.z, a.w/b.w); +} +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pdiv<double2>(const double2& a, const double2& b) { + return make_double2(a.x/b.x, a.y/b.y); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pmin<float4>(const float4& a, const float4& b) { + return make_float4(fminf(a.x, b.x), fminf(a.y, b.y), fminf(a.z, b.z), fminf(a.w, b.w)); +} +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pmin<double2>(const double2& a, const double2& b) { + return make_double2(fmin(a.x, b.x), fmin(a.y, b.y)); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pmax<float4>(const float4& a, const float4& b) { + return make_float4(fmaxf(a.x, b.x), fmaxf(a.y, b.y), fmaxf(a.z, b.z), fmaxf(a.w, b.w)); +} +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pmax<double2>(const double2& a, const double2& b) { + return make_double2(fmax(a.x, b.x), fmax(a.y, b.y)); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pload<float4>(const float* from) { + return *reinterpret_cast<const float4*>(from); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pload<double2>(const double* from) { + return *reinterpret_cast<const double2*>(from); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 ploadu<float4>(const float* from) { + return make_float4(from[0], from[1], from[2], from[3]); +} +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 ploadu<double2>(const double* from) { + return make_double2(from[0], from[1]); +} + +template<> EIGEN_STRONG_INLINE float4 ploaddup<float4>(const float* from) { + return make_float4(from[0], from[0], from[1], from[1]); +} +template<> EIGEN_STRONG_INLINE double2 ploaddup<double2>(const double* from) { + return make_double2(from[0], from[0]); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstore<float>(float* to, const float4& from) { + *reinterpret_cast<float4*>(to) = from; +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstore<double>(double* to, const double2& from) { + *reinterpret_cast<double2*>(to) = from; +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const float4& from) { + to[0] = from.x; + to[1] = from.y; + to[2] = from.z; + to[3] = from.w; +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const double2& from) { + to[0] = from.x; + to[1] = from.y; +} + +template<> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE float4 ploadt_ro<float4, Aligned>(const float* from) { +#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350 + return __ldg((const float4*)from); +#else + return make_float4(from[0], from[1], from[2], from[3]); +#endif +} +template<> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE double2 ploadt_ro<double2, Aligned>(const double* from) { +#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350 + return __ldg((const double2*)from); +#else + return make_double2(from[0], from[1]); +#endif +} + +template<> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE float4 ploadt_ro<float4, Unaligned>(const float* from) { +#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350 + return make_float4(__ldg(from+0), __ldg(from+1), __ldg(from+2), __ldg(from+3)); +#else + return make_float4(from[0], from[1], from[2], from[3]); +#endif +} +template<> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE double2 ploadt_ro<double2, Unaligned>(const double* from) { +#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350 + return make_double2(__ldg(from+0), __ldg(from+1)); +#else + return make_double2(from[0], from[1]); +#endif +} + +template<> EIGEN_DEVICE_FUNC inline float4 pgather<float, float4>(const float* from, Index stride) { + return make_float4(from[0*stride], from[1*stride], from[2*stride], from[3*stride]); +} + +template<> EIGEN_DEVICE_FUNC inline double2 pgather<double, double2>(const double* from, Index stride) { + return make_double2(from[0*stride], from[1*stride]); +} + +template<> EIGEN_DEVICE_FUNC inline void pscatter<float, float4>(float* to, const float4& from, Index stride) { + to[stride*0] = from.x; + to[stride*1] = from.y; + to[stride*2] = from.z; + to[stride*3] = from.w; +} +template<> EIGEN_DEVICE_FUNC inline void pscatter<double, double2>(double* to, const double2& from, Index stride) { + to[stride*0] = from.x; + to[stride*1] = from.y; +} + +template<> EIGEN_DEVICE_FUNC inline float pfirst<float4>(const float4& a) { + return a.x; +} +template<> EIGEN_DEVICE_FUNC inline double pfirst<double2>(const double2& a) { + return a.x; +} + +template<> EIGEN_DEVICE_FUNC inline float predux<float4>(const float4& a) { + return a.x + a.y + a.z + a.w; +} +template<> EIGEN_DEVICE_FUNC inline double predux<double2>(const double2& a) { + return a.x + a.y; +} + +template<> EIGEN_DEVICE_FUNC inline float predux_max<float4>(const float4& a) { + return fmaxf(fmaxf(a.x, a.y), fmaxf(a.z, a.w)); +} +template<> EIGEN_DEVICE_FUNC inline double predux_max<double2>(const double2& a) { + return fmax(a.x, a.y); +} + +template<> EIGEN_DEVICE_FUNC inline float predux_min<float4>(const float4& a) { + return fminf(fminf(a.x, a.y), fminf(a.z, a.w)); +} +template<> EIGEN_DEVICE_FUNC inline double predux_min<double2>(const double2& a) { + return fmin(a.x, a.y); +} + +template<> EIGEN_DEVICE_FUNC inline float predux_mul<float4>(const float4& a) { + return a.x * a.y * a.z * a.w; +} +template<> EIGEN_DEVICE_FUNC inline double predux_mul<double2>(const double2& a) { + return a.x * a.y; +} + +template<> EIGEN_DEVICE_FUNC inline float4 pabs<float4>(const float4& a) { + return make_float4(fabsf(a.x), fabsf(a.y), fabsf(a.z), fabsf(a.w)); +} +template<> EIGEN_DEVICE_FUNC inline double2 pabs<double2>(const double2& a) { + return make_double2(fabs(a.x), fabs(a.y)); +} + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<float4,4>& kernel) { + float tmp = kernel.packet[0].y; + kernel.packet[0].y = kernel.packet[1].x; + kernel.packet[1].x = tmp; + + tmp = kernel.packet[0].z; + kernel.packet[0].z = kernel.packet[2].x; + kernel.packet[2].x = tmp; + + tmp = kernel.packet[0].w; + kernel.packet[0].w = kernel.packet[3].x; + kernel.packet[3].x = tmp; + + tmp = kernel.packet[1].z; + kernel.packet[1].z = kernel.packet[2].y; + kernel.packet[2].y = tmp; + + tmp = kernel.packet[1].w; + kernel.packet[1].w = kernel.packet[3].y; + kernel.packet[3].y = tmp; + + tmp = kernel.packet[2].w; + kernel.packet[2].w = kernel.packet[3].z; + kernel.packet[3].z = tmp; +} + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<double2,2>& kernel) { + double tmp = kernel.packet[0].y; + kernel.packet[0].y = kernel.packet[1].x; + kernel.packet[1].x = tmp; +} + +#endif + +} // end namespace internal + +} // end namespace Eigen + + +#endif // EIGEN_PACKET_MATH_CUDA_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/PacketMathHalf.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/PacketMathHalf.h new file mode 100644 index 000000000..ae54225f8 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/PacketMathHalf.h @@ -0,0 +1,1123 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com> +// +// 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_PACKET_MATH_HALF_CUDA_H +#define EIGEN_PACKET_MATH_HALF_CUDA_H + + +namespace Eigen { +namespace internal { + +// Most of the following operations require arch >= 3.0 +#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDACC__) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300 + +template<> struct is_arithmetic<half2> { enum { value = true }; }; + +template<> struct packet_traits<Eigen::half> : default_packet_traits +{ + typedef half2 type; + typedef half2 half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=2, + HasHalfPacket = 0, + HasAdd = 1, + HasMul = 1, + HasDiv = 1, + HasSqrt = 1, + HasRsqrt = 1, + HasExp = 1, + HasLog = 1, + HasLog1p = 1 + }; +}; + +template<> struct unpacket_traits<half2> { typedef Eigen::half type; enum {size=2, alignment=Aligned16}; typedef half2 half; }; + +template<> __device__ EIGEN_STRONG_INLINE half2 pset1<half2>(const Eigen::half& from) { + return __half2half2(from); +} + +template<> __device__ EIGEN_STRONG_INLINE half2 pload<half2>(const Eigen::half* from) { + return *reinterpret_cast<const half2*>(from); +} + +template<> __device__ EIGEN_STRONG_INLINE half2 ploadu<half2>(const Eigen::half* from) { + return __halves2half2(from[0], from[1]); +} + +template<> EIGEN_STRONG_INLINE half2 ploaddup<half2>(const Eigen::half* from) { + return __halves2half2(from[0], from[0]); +} + +template<> __device__ EIGEN_STRONG_INLINE void pstore<Eigen::half>(Eigen::half* to, const half2& from) { + *reinterpret_cast<half2*>(to) = from; +} + +template<> __device__ EIGEN_STRONG_INLINE void pstoreu<Eigen::half>(Eigen::half* to, const half2& from) { + to[0] = __low2half(from); + to[1] = __high2half(from); +} + +template<> + __device__ EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Aligned>(const Eigen::half* from) { +#if __CUDA_ARCH__ >= 350 + return __ldg((const half2*)from); +#else + return __halves2half2(*(from+0), *(from+1)); +#endif +} + +template<> +__device__ EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Unaligned>(const Eigen::half* from) { +#if __CUDA_ARCH__ >= 350 + return __halves2half2(__ldg(from+0), __ldg(from+1)); +#else + return __halves2half2(*(from+0), *(from+1)); +#endif +} + +template<> __device__ EIGEN_STRONG_INLINE half2 pgather<Eigen::half, half2>(const Eigen::half* from, Index stride) { + return __halves2half2(from[0*stride], from[1*stride]); +} + +template<> __device__ EIGEN_STRONG_INLINE void pscatter<Eigen::half, half2>(Eigen::half* to, const half2& from, Index stride) { + to[stride*0] = __low2half(from); + to[stride*1] = __high2half(from); +} + +template<> __device__ EIGEN_STRONG_INLINE Eigen::half pfirst<half2>(const half2& a) { + return __low2half(a); +} + +template<> __device__ EIGEN_STRONG_INLINE half2 pabs<half2>(const half2& a) { + half2 result; + result.x = a.x & 0x7FFF7FFF; + return result; +} + + +__device__ EIGEN_STRONG_INLINE void +ptranspose(PacketBlock<half2,2>& kernel) { + __half a1 = __low2half(kernel.packet[0]); + __half a2 = __high2half(kernel.packet[0]); + __half b1 = __low2half(kernel.packet[1]); + __half b2 = __high2half(kernel.packet[1]); + kernel.packet[0] = __halves2half2(a1, b1); + kernel.packet[1] = __halves2half2(a2, b2); +} + +template<> __device__ EIGEN_STRONG_INLINE half2 plset<half2>(const Eigen::half& a) { +#if __CUDA_ARCH__ >= 530 + return __halves2half2(a, __hadd(a, __float2half(1.0f))); +#else + float f = __half2float(a) + 1.0f; + return __halves2half2(a, __float2half(f)); +#endif +} + +template<> __device__ EIGEN_STRONG_INLINE half2 padd<half2>(const half2& a, const half2& b) { +#if __CUDA_ARCH__ >= 530 + return __hadd2(a, b); +#else + float a1 = __low2float(a); + float a2 = __high2float(a); + float b1 = __low2float(b); + float b2 = __high2float(b); + float r1 = a1 + b1; + float r2 = a2 + b2; + return __floats2half2_rn(r1, r2); +#endif +} + +template<> __device__ EIGEN_STRONG_INLINE half2 psub<half2>(const half2& a, const half2& b) { +#if __CUDA_ARCH__ >= 530 + return __hsub2(a, b); +#else + float a1 = __low2float(a); + float a2 = __high2float(a); + float b1 = __low2float(b); + float b2 = __high2float(b); + float r1 = a1 - b1; + float r2 = a2 - b2; + return __floats2half2_rn(r1, r2); +#endif +} + +template<> __device__ EIGEN_STRONG_INLINE half2 pnegate(const half2& a) { +#if __CUDA_ARCH__ >= 530 + return __hneg2(a); +#else + float a1 = __low2float(a); + float a2 = __high2float(a); + return __floats2half2_rn(-a1, -a2); +#endif +} + +template<> __device__ EIGEN_STRONG_INLINE half2 pconj(const half2& a) { return a; } + +template<> __device__ EIGEN_STRONG_INLINE half2 pmul<half2>(const half2& a, const half2& b) { +#if __CUDA_ARCH__ >= 530 + return __hmul2(a, b); +#else + float a1 = __low2float(a); + float a2 = __high2float(a); + float b1 = __low2float(b); + float b2 = __high2float(b); + float r1 = a1 * b1; + float r2 = a2 * b2; + return __floats2half2_rn(r1, r2); +#endif +} + +template<> __device__ EIGEN_STRONG_INLINE half2 pmadd<half2>(const half2& a, const half2& b, const half2& c) { +#if __CUDA_ARCH__ >= 530 + return __hfma2(a, b, c); +#else + float a1 = __low2float(a); + float a2 = __high2float(a); + float b1 = __low2float(b); + float b2 = __high2float(b); + float c1 = __low2float(c); + float c2 = __high2float(c); + float r1 = a1 * b1 + c1; + float r2 = a2 * b2 + c2; + return __floats2half2_rn(r1, r2); +#endif +} + +template<> __device__ EIGEN_STRONG_INLINE half2 pdiv<half2>(const half2& a, const half2& b) { + float a1 = __low2float(a); + float a2 = __high2float(a); + float b1 = __low2float(b); + float b2 = __high2float(b); + float r1 = a1 / b1; + float r2 = a2 / b2; + return __floats2half2_rn(r1, r2); +} + +template<> __device__ EIGEN_STRONG_INLINE half2 pmin<half2>(const half2& a, const half2& b) { + float a1 = __low2float(a); + float a2 = __high2float(a); + float b1 = __low2float(b); + float b2 = __high2float(b); + __half r1 = a1 < b1 ? __low2half(a) : __low2half(b); + __half r2 = a2 < b2 ? __high2half(a) : __high2half(b); + return __halves2half2(r1, r2); +} + +template<> __device__ EIGEN_STRONG_INLINE half2 pmax<half2>(const half2& a, const half2& b) { + float a1 = __low2float(a); + float a2 = __high2float(a); + float b1 = __low2float(b); + float b2 = __high2float(b); + __half r1 = a1 > b1 ? __low2half(a) : __low2half(b); + __half r2 = a2 > b2 ? __high2half(a) : __high2half(b); + return __halves2half2(r1, r2); +} + +template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux<half2>(const half2& a) { +#if __CUDA_ARCH__ >= 530 + return __hadd(__low2half(a), __high2half(a)); +#else + float a1 = __low2float(a); + float a2 = __high2float(a); + return Eigen::half(half_impl::raw_uint16_to_half(__float2half_rn(a1 + a2))); +#endif +} + +template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_max<half2>(const half2& a) { +#if __CUDA_ARCH__ >= 530 + __half first = __low2half(a); + __half second = __high2half(a); + return __hgt(first, second) ? first : second; +#else + float a1 = __low2float(a); + float a2 = __high2float(a); + return a1 > a2 ? __low2half(a) : __high2half(a); +#endif +} + +template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_min<half2>(const half2& a) { +#if __CUDA_ARCH__ >= 530 + __half first = __low2half(a); + __half second = __high2half(a); + return __hlt(first, second) ? first : second; +#else + float a1 = __low2float(a); + float a2 = __high2float(a); + return a1 < a2 ? __low2half(a) : __high2half(a); +#endif +} + +template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_mul<half2>(const half2& a) { +#if __CUDA_ARCH__ >= 530 + return __hmul(__low2half(a), __high2half(a)); +#else + float a1 = __low2float(a); + float a2 = __high2float(a); + return Eigen::half(half_impl::raw_uint16_to_half(__float2half_rn(a1 * a2))); +#endif +} + +template<> __device__ EIGEN_STRONG_INLINE half2 plog1p<half2>(const half2& a) { + float a1 = __low2float(a); + float a2 = __high2float(a); + float r1 = log1pf(a1); + float r2 = log1pf(a2); + return __floats2half2_rn(r1, r2); +} + +#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 530 + +template<> __device__ EIGEN_STRONG_INLINE +half2 plog<half2>(const half2& a) { + return h2log(a); +} + +template<> __device__ EIGEN_STRONG_INLINE +half2 pexp<half2>(const half2& a) { + return h2exp(a); +} + +template<> __device__ EIGEN_STRONG_INLINE +half2 psqrt<half2>(const half2& a) { + return h2sqrt(a); +} + +template<> __device__ EIGEN_STRONG_INLINE +half2 prsqrt<half2>(const half2& a) { + return h2rsqrt(a); +} + +#else + +template<> __device__ EIGEN_STRONG_INLINE half2 plog<half2>(const half2& a) { + float a1 = __low2float(a); + float a2 = __high2float(a); + float r1 = logf(a1); + float r2 = logf(a2); + return __floats2half2_rn(r1, r2); +} + +template<> __device__ EIGEN_STRONG_INLINE half2 pexp<half2>(const half2& a) { + float a1 = __low2float(a); + float a2 = __high2float(a); + float r1 = expf(a1); + float r2 = expf(a2); + return __floats2half2_rn(r1, r2); +} + +template<> __device__ EIGEN_STRONG_INLINE half2 psqrt<half2>(const half2& a) { + float a1 = __low2float(a); + float a2 = __high2float(a); + float r1 = sqrtf(a1); + float r2 = sqrtf(a2); + return __floats2half2_rn(r1, r2); +} + +template<> __device__ EIGEN_STRONG_INLINE half2 prsqrt<half2>(const half2& a) { + float a1 = __low2float(a); + float a2 = __high2float(a); + float r1 = rsqrtf(a1); + float r2 = rsqrtf(a2); + return __floats2half2_rn(r1, r2); +} + +#endif + +#elif defined EIGEN_VECTORIZE_AVX512 + +typedef struct { + __m256i x; +} Packet16h; + + +template<> struct is_arithmetic<Packet16h> { enum { value = true }; }; + +template <> +struct packet_traits<half> : default_packet_traits { + typedef Packet16h type; + // There is no half-size packet for Packet16h. + typedef Packet16h half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size = 16, + HasHalfPacket = 0, + HasAdd = 0, + HasSub = 0, + HasMul = 0, + HasNegate = 0, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasConj = 0, + HasSetLinear = 0, + HasDiv = 0, + HasSqrt = 0, + HasRsqrt = 0, + HasExp = 0, + HasLog = 0, + HasBlend = 0 + }; +}; + + +template<> struct unpacket_traits<Packet16h> { typedef Eigen::half type; enum {size=16, alignment=Aligned32}; typedef Packet16h half; }; + +template<> EIGEN_STRONG_INLINE Packet16h pset1<Packet16h>(const Eigen::half& from) { + Packet16h result; + result.x = _mm256_set1_epi16(from.x); + return result; +} + +template<> EIGEN_STRONG_INLINE Eigen::half pfirst<Packet16h>(const Packet16h& from) { + return half_impl::raw_uint16_to_half(static_cast<unsigned short>(_mm256_extract_epi16(from.x, 0))); +} + +template<> EIGEN_STRONG_INLINE Packet16h pload<Packet16h>(const Eigen::half* from) { + Packet16h result; + result.x = _mm256_load_si256(reinterpret_cast<const __m256i*>(from)); + return result; +} + +template<> EIGEN_STRONG_INLINE Packet16h ploadu<Packet16h>(const Eigen::half* from) { + Packet16h result; + result.x = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(from)); + return result; +} + +template<> EIGEN_STRONG_INLINE void pstore<half>(Eigen::half* to, const Packet16h& from) { + _mm256_store_si256((__m256i*)to, from.x); +} + +template<> EIGEN_STRONG_INLINE void pstoreu<half>(Eigen::half* to, const Packet16h& from) { + _mm256_storeu_si256((__m256i*)to, from.x); +} + +template<> EIGEN_STRONG_INLINE Packet16h +ploadquad(const Eigen::half* from) { + Packet16h result; + unsigned short a = from[0].x; + unsigned short b = from[1].x; + unsigned short c = from[2].x; + unsigned short d = from[3].x; + result.x = _mm256_set_epi16(d, d, d, d, c, c, c, c, b, b, b, b, a, a, a, a); + return result; +} + +EIGEN_STRONG_INLINE Packet16f half2float(const Packet16h& a) { +#ifdef EIGEN_HAS_FP16_C + return _mm512_cvtph_ps(a.x); +#else + EIGEN_ALIGN64 half aux[16]; + pstore(aux, a); + float f0(aux[0]); + float f1(aux[1]); + float f2(aux[2]); + float f3(aux[3]); + float f4(aux[4]); + float f5(aux[5]); + float f6(aux[6]); + float f7(aux[7]); + float f8(aux[8]); + float f9(aux[9]); + float fa(aux[10]); + float fb(aux[11]); + float fc(aux[12]); + float fd(aux[13]); + float fe(aux[14]); + float ff(aux[15]); + + return _mm512_set_ps( + ff, fe, fd, fc, fb, fa, f9, f8, f7, f6, f5, f4, f3, f2, f1, f0); +#endif +} + +EIGEN_STRONG_INLINE Packet16h float2half(const Packet16f& a) { +#ifdef EIGEN_HAS_FP16_C + Packet16h result; + result.x = _mm512_cvtps_ph(a, _MM_FROUND_TO_NEAREST_INT|_MM_FROUND_NO_EXC); + return result; +#else + EIGEN_ALIGN64 float aux[16]; + pstore(aux, a); + half h0(aux[0]); + half h1(aux[1]); + half h2(aux[2]); + half h3(aux[3]); + half h4(aux[4]); + half h5(aux[5]); + half h6(aux[6]); + half h7(aux[7]); + half h8(aux[8]); + half h9(aux[9]); + half ha(aux[10]); + half hb(aux[11]); + half hc(aux[12]); + half hd(aux[13]); + half he(aux[14]); + half hf(aux[15]); + + Packet16h result; + result.x = _mm256_set_epi16( + hf.x, he.x, hd.x, hc.x, hb.x, ha.x, h9.x, h8.x, + h7.x, h6.x, h5.x, h4.x, h3.x, h2.x, h1.x, h0.x); + return result; +#endif +} + +template<> EIGEN_STRONG_INLINE Packet16h padd<Packet16h>(const Packet16h& a, const Packet16h& b) { + Packet16f af = half2float(a); + Packet16f bf = half2float(b); + Packet16f rf = padd(af, bf); + return float2half(rf); +} + +template<> EIGEN_STRONG_INLINE Packet16h pmul<Packet16h>(const Packet16h& a, const Packet16h& b) { + Packet16f af = half2float(a); + Packet16f bf = half2float(b); + Packet16f rf = pmul(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)); +} + +template<> EIGEN_STRONG_INLINE Packet16h pgather<Eigen::half, Packet16h>(const Eigen::half* from, Index stride) +{ + Packet16h result; + result.x = _mm256_set_epi16( + from[15*stride].x, from[14*stride].x, from[13*stride].x, from[12*stride].x, + from[11*stride].x, from[10*stride].x, from[9*stride].x, from[8*stride].x, + from[7*stride].x, from[6*stride].x, from[5*stride].x, from[4*stride].x, + from[3*stride].x, from[2*stride].x, from[1*stride].x, from[0*stride].x); + return result; +} + +template<> EIGEN_STRONG_INLINE void pscatter<half, Packet16h>(half* to, const Packet16h& from, Index stride) +{ + EIGEN_ALIGN64 half aux[16]; + pstore(aux, from); + to[stride*0].x = aux[0].x; + to[stride*1].x = aux[1].x; + to[stride*2].x = aux[2].x; + to[stride*3].x = aux[3].x; + to[stride*4].x = aux[4].x; + to[stride*5].x = aux[5].x; + to[stride*6].x = aux[6].x; + to[stride*7].x = aux[7].x; + to[stride*8].x = aux[8].x; + to[stride*9].x = aux[9].x; + to[stride*10].x = aux[10].x; + to[stride*11].x = aux[11].x; + to[stride*12].x = aux[12].x; + to[stride*13].x = aux[13].x; + to[stride*14].x = aux[14].x; + to[stride*15].x = aux[15].x; +} + +EIGEN_STRONG_INLINE void +ptranspose(PacketBlock<Packet16h,16>& kernel) { + __m256i a = kernel.packet[0].x; + __m256i b = kernel.packet[1].x; + __m256i c = kernel.packet[2].x; + __m256i d = kernel.packet[3].x; + __m256i e = kernel.packet[4].x; + __m256i f = kernel.packet[5].x; + __m256i g = kernel.packet[6].x; + __m256i h = kernel.packet[7].x; + __m256i i = kernel.packet[8].x; + __m256i j = kernel.packet[9].x; + __m256i k = kernel.packet[10].x; + __m256i l = kernel.packet[11].x; + __m256i m = kernel.packet[12].x; + __m256i n = kernel.packet[13].x; + __m256i o = kernel.packet[14].x; + __m256i p = kernel.packet[15].x; + + __m256i ab_07 = _mm256_unpacklo_epi16(a, b); + __m256i cd_07 = _mm256_unpacklo_epi16(c, d); + __m256i ef_07 = _mm256_unpacklo_epi16(e, f); + __m256i gh_07 = _mm256_unpacklo_epi16(g, h); + __m256i ij_07 = _mm256_unpacklo_epi16(i, j); + __m256i kl_07 = _mm256_unpacklo_epi16(k, l); + __m256i mn_07 = _mm256_unpacklo_epi16(m, n); + __m256i op_07 = _mm256_unpacklo_epi16(o, p); + + __m256i ab_8f = _mm256_unpackhi_epi16(a, b); + __m256i cd_8f = _mm256_unpackhi_epi16(c, d); + __m256i ef_8f = _mm256_unpackhi_epi16(e, f); + __m256i gh_8f = _mm256_unpackhi_epi16(g, h); + __m256i ij_8f = _mm256_unpackhi_epi16(i, j); + __m256i kl_8f = _mm256_unpackhi_epi16(k, l); + __m256i mn_8f = _mm256_unpackhi_epi16(m, n); + __m256i op_8f = _mm256_unpackhi_epi16(o, p); + + __m256i abcd_03 = _mm256_unpacklo_epi32(ab_07, cd_07); + __m256i abcd_47 = _mm256_unpackhi_epi32(ab_07, cd_07); + __m256i efgh_03 = _mm256_unpacklo_epi32(ef_07, gh_07); + __m256i efgh_47 = _mm256_unpackhi_epi32(ef_07, gh_07); + __m256i ijkl_03 = _mm256_unpacklo_epi32(ij_07, kl_07); + __m256i ijkl_47 = _mm256_unpackhi_epi32(ij_07, kl_07); + __m256i mnop_03 = _mm256_unpacklo_epi32(mn_07, op_07); + __m256i mnop_47 = _mm256_unpackhi_epi32(mn_07, op_07); + + __m256i abcd_8b = _mm256_unpacklo_epi32(ab_8f, cd_8f); + __m256i abcd_cf = _mm256_unpackhi_epi32(ab_8f, cd_8f); + __m256i efgh_8b = _mm256_unpacklo_epi32(ef_8f, gh_8f); + __m256i efgh_cf = _mm256_unpackhi_epi32(ef_8f, gh_8f); + __m256i ijkl_8b = _mm256_unpacklo_epi32(ij_8f, kl_8f); + __m256i ijkl_cf = _mm256_unpackhi_epi32(ij_8f, kl_8f); + __m256i mnop_8b = _mm256_unpacklo_epi32(mn_8f, op_8f); + __m256i mnop_cf = _mm256_unpackhi_epi32(mn_8f, op_8f); + + __m256i abcdefgh_01 = _mm256_unpacklo_epi64(abcd_03, efgh_03); + __m256i abcdefgh_23 = _mm256_unpackhi_epi64(abcd_03, efgh_03); + __m256i ijklmnop_01 = _mm256_unpacklo_epi64(ijkl_03, mnop_03); + __m256i ijklmnop_23 = _mm256_unpackhi_epi64(ijkl_03, mnop_03); + __m256i abcdefgh_45 = _mm256_unpacklo_epi64(abcd_47, efgh_47); + __m256i abcdefgh_67 = _mm256_unpackhi_epi64(abcd_47, efgh_47); + __m256i ijklmnop_45 = _mm256_unpacklo_epi64(ijkl_47, mnop_47); + __m256i ijklmnop_67 = _mm256_unpackhi_epi64(ijkl_47, mnop_47); + __m256i abcdefgh_89 = _mm256_unpacklo_epi64(abcd_8b, efgh_8b); + __m256i abcdefgh_ab = _mm256_unpackhi_epi64(abcd_8b, efgh_8b); + __m256i ijklmnop_89 = _mm256_unpacklo_epi64(ijkl_8b, mnop_8b); + __m256i ijklmnop_ab = _mm256_unpackhi_epi64(ijkl_8b, mnop_8b); + __m256i abcdefgh_cd = _mm256_unpacklo_epi64(abcd_cf, efgh_cf); + __m256i abcdefgh_ef = _mm256_unpackhi_epi64(abcd_cf, efgh_cf); + __m256i ijklmnop_cd = _mm256_unpacklo_epi64(ijkl_cf, mnop_cf); + __m256i ijklmnop_ef = _mm256_unpackhi_epi64(ijkl_cf, mnop_cf); + + // NOTE: no unpacklo/hi instr in this case, so using permute instr. + __m256i a_p_0 = _mm256_permute2x128_si256(abcdefgh_01, ijklmnop_01, 0x20); + __m256i a_p_1 = _mm256_permute2x128_si256(abcdefgh_01, ijklmnop_01, 0x31); + __m256i a_p_2 = _mm256_permute2x128_si256(abcdefgh_23, ijklmnop_23, 0x20); + __m256i a_p_3 = _mm256_permute2x128_si256(abcdefgh_23, ijklmnop_23, 0x31); + __m256i a_p_4 = _mm256_permute2x128_si256(abcdefgh_45, ijklmnop_45, 0x20); + __m256i a_p_5 = _mm256_permute2x128_si256(abcdefgh_45, ijklmnop_45, 0x31); + __m256i a_p_6 = _mm256_permute2x128_si256(abcdefgh_67, ijklmnop_67, 0x20); + __m256i a_p_7 = _mm256_permute2x128_si256(abcdefgh_67, ijklmnop_67, 0x31); + __m256i a_p_8 = _mm256_permute2x128_si256(abcdefgh_89, ijklmnop_89, 0x20); + __m256i a_p_9 = _mm256_permute2x128_si256(abcdefgh_89, ijklmnop_89, 0x31); + __m256i a_p_a = _mm256_permute2x128_si256(abcdefgh_ab, ijklmnop_ab, 0x20); + __m256i a_p_b = _mm256_permute2x128_si256(abcdefgh_ab, ijklmnop_ab, 0x31); + __m256i a_p_c = _mm256_permute2x128_si256(abcdefgh_cd, ijklmnop_cd, 0x20); + __m256i a_p_d = _mm256_permute2x128_si256(abcdefgh_cd, ijklmnop_cd, 0x31); + __m256i a_p_e = _mm256_permute2x128_si256(abcdefgh_ef, ijklmnop_ef, 0x20); + __m256i a_p_f = _mm256_permute2x128_si256(abcdefgh_ef, ijklmnop_ef, 0x31); + + kernel.packet[0].x = a_p_0; + kernel.packet[1].x = a_p_1; + kernel.packet[2].x = a_p_2; + kernel.packet[3].x = a_p_3; + kernel.packet[4].x = a_p_4; + kernel.packet[5].x = a_p_5; + kernel.packet[6].x = a_p_6; + kernel.packet[7].x = a_p_7; + kernel.packet[8].x = a_p_8; + kernel.packet[9].x = a_p_9; + kernel.packet[10].x = a_p_a; + kernel.packet[11].x = a_p_b; + kernel.packet[12].x = a_p_c; + kernel.packet[13].x = a_p_d; + kernel.packet[14].x = a_p_e; + kernel.packet[15].x = a_p_f; +} + +EIGEN_STRONG_INLINE void +ptranspose(PacketBlock<Packet16h,8>& kernel) { + EIGEN_ALIGN64 half in[8][16]; + pstore<half>(in[0], kernel.packet[0]); + pstore<half>(in[1], kernel.packet[1]); + pstore<half>(in[2], kernel.packet[2]); + pstore<half>(in[3], kernel.packet[3]); + pstore<half>(in[4], kernel.packet[4]); + pstore<half>(in[5], kernel.packet[5]); + pstore<half>(in[6], kernel.packet[6]); + pstore<half>(in[7], kernel.packet[7]); + + EIGEN_ALIGN64 half out[8][16]; + + for (int i = 0; i < 8; ++i) { + for (int j = 0; j < 8; ++j) { + out[i][j] = in[j][2*i]; + } + for (int j = 0; j < 8; ++j) { + out[i][j+8] = in[j][2*i+1]; + } + } + + kernel.packet[0] = pload<Packet16h>(out[0]); + kernel.packet[1] = pload<Packet16h>(out[1]); + kernel.packet[2] = pload<Packet16h>(out[2]); + kernel.packet[3] = pload<Packet16h>(out[3]); + kernel.packet[4] = pload<Packet16h>(out[4]); + kernel.packet[5] = pload<Packet16h>(out[5]); + kernel.packet[6] = pload<Packet16h>(out[6]); + kernel.packet[7] = pload<Packet16h>(out[7]); +} + +EIGEN_STRONG_INLINE void +ptranspose(PacketBlock<Packet16h,4>& kernel) { + EIGEN_ALIGN64 half in[4][16]; + pstore<half>(in[0], kernel.packet[0]); + pstore<half>(in[1], kernel.packet[1]); + pstore<half>(in[2], kernel.packet[2]); + pstore<half>(in[3], kernel.packet[3]); + + EIGEN_ALIGN64 half out[4][16]; + + for (int i = 0; i < 4; ++i) { + for (int j = 0; j < 4; ++j) { + out[i][j] = in[j][4*i]; + } + for (int j = 0; j < 4; ++j) { + out[i][j+4] = in[j][4*i+1]; + } + for (int j = 0; j < 4; ++j) { + out[i][j+8] = in[j][4*i+2]; + } + for (int j = 0; j < 4; ++j) { + out[i][j+12] = in[j][4*i+3]; + } + } + + kernel.packet[0] = pload<Packet16h>(out[0]); + kernel.packet[1] = pload<Packet16h>(out[1]); + kernel.packet[2] = pload<Packet16h>(out[2]); + kernel.packet[3] = pload<Packet16h>(out[3]); +} + + +#elif defined EIGEN_VECTORIZE_AVX + +typedef struct { + __m128i x; +} Packet8h; + + +template<> struct is_arithmetic<Packet8h> { enum { value = true }; }; + +template <> +struct packet_traits<Eigen::half> : default_packet_traits { + typedef Packet8h type; + // There is no half-size packet for Packet8h. + typedef Packet8h half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size = 8, + HasHalfPacket = 0, + HasAdd = 0, + HasSub = 0, + HasMul = 0, + HasNegate = 0, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasConj = 0, + HasSetLinear = 0, + HasDiv = 0, + HasSqrt = 0, + HasRsqrt = 0, + HasExp = 0, + HasLog = 0, + HasBlend = 0 + }; +}; + + +template<> struct unpacket_traits<Packet8h> { typedef Eigen::half type; enum {size=8, alignment=Aligned16}; typedef Packet8h half; }; + +template<> EIGEN_STRONG_INLINE Packet8h pset1<Packet8h>(const Eigen::half& from) { + Packet8h result; + result.x = _mm_set1_epi16(from.x); + return result; +} + +template<> EIGEN_STRONG_INLINE Eigen::half pfirst<Packet8h>(const Packet8h& from) { + return half_impl::raw_uint16_to_half(static_cast<unsigned short>(_mm_extract_epi16(from.x, 0))); +} + +template<> EIGEN_STRONG_INLINE Packet8h pload<Packet8h>(const Eigen::half* from) { + Packet8h result; + result.x = _mm_load_si128(reinterpret_cast<const __m128i*>(from)); + return result; +} + +template<> EIGEN_STRONG_INLINE Packet8h ploadu<Packet8h>(const Eigen::half* from) { + Packet8h result; + result.x = _mm_loadu_si128(reinterpret_cast<const __m128i*>(from)); + return result; +} + +template<> EIGEN_STRONG_INLINE void pstore<Eigen::half>(Eigen::half* to, const Packet8h& from) { + _mm_store_si128(reinterpret_cast<__m128i*>(to), from.x); +} + +template<> EIGEN_STRONG_INLINE void pstoreu<Eigen::half>(Eigen::half* to, const Packet8h& from) { + _mm_storeu_si128(reinterpret_cast<__m128i*>(to), from.x); +} + +template<> EIGEN_STRONG_INLINE Packet8h +ploadquad<Packet8h>(const Eigen::half* from) { + Packet8h result; + unsigned short a = from[0].x; + unsigned short b = from[1].x; + result.x = _mm_set_epi16(b, b, b, b, a, a, a, a); + return result; +} + +EIGEN_STRONG_INLINE Packet8f half2float(const Packet8h& a) { +#ifdef EIGEN_HAS_FP16_C + return _mm256_cvtph_ps(a.x); +#else + EIGEN_ALIGN32 Eigen::half aux[8]; + pstore(aux, a); + float f0(aux[0]); + float f1(aux[1]); + float f2(aux[2]); + float f3(aux[3]); + float f4(aux[4]); + float f5(aux[5]); + float f6(aux[6]); + float f7(aux[7]); + + return _mm256_set_ps(f7, f6, f5, f4, f3, f2, f1, f0); +#endif +} + +EIGEN_STRONG_INLINE Packet8h float2half(const Packet8f& a) { +#ifdef EIGEN_HAS_FP16_C + Packet8h result; + result.x = _mm256_cvtps_ph(a, _MM_FROUND_TO_NEAREST_INT|_MM_FROUND_NO_EXC); + return result; +#else + EIGEN_ALIGN32 float aux[8]; + pstore(aux, a); + Eigen::half h0(aux[0]); + Eigen::half h1(aux[1]); + Eigen::half h2(aux[2]); + Eigen::half h3(aux[3]); + Eigen::half h4(aux[4]); + Eigen::half h5(aux[5]); + Eigen::half h6(aux[6]); + Eigen::half h7(aux[7]); + + Packet8h result; + result.x = _mm_set_epi16(h7.x, h6.x, h5.x, h4.x, h3.x, h2.x, h1.x, h0.x); + return result; +#endif +} + +template<> EIGEN_STRONG_INLINE Packet8h pconj(const Packet8h& a) { return a; } + +template<> EIGEN_STRONG_INLINE Packet8h padd<Packet8h>(const Packet8h& a, const Packet8h& b) { + Packet8f af = half2float(a); + Packet8f bf = half2float(b); + Packet8f rf = padd(af, bf); + return float2half(rf); +} + +template<> EIGEN_STRONG_INLINE Packet8h pmul<Packet8h>(const Packet8h& a, const Packet8h& b) { + Packet8f af = half2float(a); + Packet8f bf = half2float(b); + Packet8f rf = pmul(af, bf); + return float2half(rf); +} + +template<> EIGEN_STRONG_INLINE Packet8h pgather<Eigen::half, Packet8h>(const Eigen::half* from, Index stride) +{ + Packet8h result; + result.x = _mm_set_epi16(from[7*stride].x, from[6*stride].x, from[5*stride].x, from[4*stride].x, from[3*stride].x, from[2*stride].x, from[1*stride].x, from[0*stride].x); + return result; +} + +template<> EIGEN_STRONG_INLINE void pscatter<Eigen::half, Packet8h>(Eigen::half* to, const Packet8h& from, Index stride) +{ + EIGEN_ALIGN32 Eigen::half aux[8]; + pstore(aux, from); + to[stride*0].x = aux[0].x; + to[stride*1].x = aux[1].x; + to[stride*2].x = aux[2].x; + to[stride*3].x = aux[3].x; + to[stride*4].x = aux[4].x; + to[stride*5].x = aux[5].x; + to[stride*6].x = aux[6].x; + to[stride*7].x = aux[7].x; +} + +template<> EIGEN_STRONG_INLINE Eigen::half predux<Packet8h>(const Packet8h& a) { + Packet8f af = half2float(a); + float reduced = predux<Packet8f>(af); + return Eigen::half(reduced); +} + +template<> EIGEN_STRONG_INLINE Eigen::half predux_max<Packet8h>(const Packet8h& a) { + Packet8f af = half2float(a); + float reduced = predux_max<Packet8f>(af); + return Eigen::half(reduced); +} + +template<> EIGEN_STRONG_INLINE Eigen::half predux_min<Packet8h>(const Packet8h& a) { + Packet8f af = half2float(a); + float reduced = predux_min<Packet8f>(af); + return Eigen::half(reduced); +} + +template<> EIGEN_STRONG_INLINE Eigen::half predux_mul<Packet8h>(const Packet8h& a) { + Packet8f af = half2float(a); + float reduced = predux_mul<Packet8f>(af); + return Eigen::half(reduced); +} + +EIGEN_STRONG_INLINE void +ptranspose(PacketBlock<Packet8h,8>& kernel) { + __m128i a = kernel.packet[0].x; + __m128i b = kernel.packet[1].x; + __m128i c = kernel.packet[2].x; + __m128i d = kernel.packet[3].x; + __m128i e = kernel.packet[4].x; + __m128i f = kernel.packet[5].x; + __m128i g = kernel.packet[6].x; + __m128i h = kernel.packet[7].x; + + __m128i a03b03 = _mm_unpacklo_epi16(a, b); + __m128i c03d03 = _mm_unpacklo_epi16(c, d); + __m128i e03f03 = _mm_unpacklo_epi16(e, f); + __m128i g03h03 = _mm_unpacklo_epi16(g, h); + __m128i a47b47 = _mm_unpackhi_epi16(a, b); + __m128i c47d47 = _mm_unpackhi_epi16(c, d); + __m128i e47f47 = _mm_unpackhi_epi16(e, f); + __m128i g47h47 = _mm_unpackhi_epi16(g, h); + + __m128i a01b01c01d01 = _mm_unpacklo_epi32(a03b03, c03d03); + __m128i a23b23c23d23 = _mm_unpackhi_epi32(a03b03, c03d03); + __m128i e01f01g01h01 = _mm_unpacklo_epi32(e03f03, g03h03); + __m128i e23f23g23h23 = _mm_unpackhi_epi32(e03f03, g03h03); + __m128i a45b45c45d45 = _mm_unpacklo_epi32(a47b47, c47d47); + __m128i a67b67c67d67 = _mm_unpackhi_epi32(a47b47, c47d47); + __m128i e45f45g45h45 = _mm_unpacklo_epi32(e47f47, g47h47); + __m128i e67f67g67h67 = _mm_unpackhi_epi32(e47f47, g47h47); + + __m128i a0b0c0d0e0f0g0h0 = _mm_unpacklo_epi64(a01b01c01d01, e01f01g01h01); + __m128i a1b1c1d1e1f1g1h1 = _mm_unpackhi_epi64(a01b01c01d01, e01f01g01h01); + __m128i a2b2c2d2e2f2g2h2 = _mm_unpacklo_epi64(a23b23c23d23, e23f23g23h23); + __m128i a3b3c3d3e3f3g3h3 = _mm_unpackhi_epi64(a23b23c23d23, e23f23g23h23); + __m128i a4b4c4d4e4f4g4h4 = _mm_unpacklo_epi64(a45b45c45d45, e45f45g45h45); + __m128i a5b5c5d5e5f5g5h5 = _mm_unpackhi_epi64(a45b45c45d45, e45f45g45h45); + __m128i a6b6c6d6e6f6g6h6 = _mm_unpacklo_epi64(a67b67c67d67, e67f67g67h67); + __m128i a7b7c7d7e7f7g7h7 = _mm_unpackhi_epi64(a67b67c67d67, e67f67g67h67); + + kernel.packet[0].x = a0b0c0d0e0f0g0h0; + kernel.packet[1].x = a1b1c1d1e1f1g1h1; + kernel.packet[2].x = a2b2c2d2e2f2g2h2; + kernel.packet[3].x = a3b3c3d3e3f3g3h3; + kernel.packet[4].x = a4b4c4d4e4f4g4h4; + kernel.packet[5].x = a5b5c5d5e5f5g5h5; + kernel.packet[6].x = a6b6c6d6e6f6g6h6; + kernel.packet[7].x = a7b7c7d7e7f7g7h7; +} + +EIGEN_STRONG_INLINE void +ptranspose(PacketBlock<Packet8h,4>& kernel) { + EIGEN_ALIGN32 Eigen::half in[4][8]; + pstore<Eigen::half>(in[0], kernel.packet[0]); + pstore<Eigen::half>(in[1], kernel.packet[1]); + pstore<Eigen::half>(in[2], kernel.packet[2]); + pstore<Eigen::half>(in[3], kernel.packet[3]); + + EIGEN_ALIGN32 Eigen::half out[4][8]; + + for (int i = 0; i < 4; ++i) { + for (int j = 0; j < 4; ++j) { + out[i][j] = in[j][2*i]; + } + for (int j = 0; j < 4; ++j) { + out[i][j+4] = in[j][2*i+1]; + } + } + + kernel.packet[0] = pload<Packet8h>(out[0]); + kernel.packet[1] = pload<Packet8h>(out[1]); + kernel.packet[2] = pload<Packet8h>(out[2]); + kernel.packet[3] = pload<Packet8h>(out[3]); +} + + +// Disable the following code since it's broken on too many platforms / compilers. +//#elif defined(EIGEN_VECTORIZE_SSE) && (!EIGEN_ARCH_x86_64) && (!EIGEN_COMP_MSVC) +#elif 0 + +typedef struct { + __m64 x; +} Packet4h; + + +template<> struct is_arithmetic<Packet4h> { enum { value = true }; }; + +template <> +struct packet_traits<Eigen::half> : default_packet_traits { + typedef Packet4h type; + // There is no half-size packet for Packet4h. + typedef Packet4h half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size = 4, + HasHalfPacket = 0, + HasAdd = 0, + HasSub = 0, + HasMul = 0, + HasNegate = 0, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasConj = 0, + HasSetLinear = 0, + HasDiv = 0, + HasSqrt = 0, + HasRsqrt = 0, + HasExp = 0, + HasLog = 0, + HasBlend = 0 + }; +}; + + +template<> struct unpacket_traits<Packet4h> { typedef Eigen::half type; enum {size=4, alignment=Aligned16}; typedef Packet4h half; }; + +template<> EIGEN_STRONG_INLINE Packet4h pset1<Packet4h>(const Eigen::half& from) { + Packet4h result; + result.x = _mm_set1_pi16(from.x); + return result; +} + +template<> EIGEN_STRONG_INLINE Eigen::half pfirst<Packet4h>(const Packet4h& from) { + return half_impl::raw_uint16_to_half(static_cast<unsigned short>(_mm_cvtsi64_si32(from.x))); +} + +template<> EIGEN_STRONG_INLINE Packet4h pconj(const Packet4h& a) { return a; } + +template<> EIGEN_STRONG_INLINE Packet4h padd<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); + + 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)); + return result; +} + +template<> EIGEN_STRONG_INLINE Packet4h ploadu<Packet4h>(const Eigen::half* from) { + Packet4h result; + result.x = _mm_cvtsi64_m64(*reinterpret_cast<const __int64_t*>(from)); + return result; +} + +template<> EIGEN_STRONG_INLINE void pstore<Eigen::half>(Eigen::half* to, const Packet4h& from) { + __int64_t r = _mm_cvtm64_si64(from.x); + *(reinterpret_cast<__int64_t*>(to)) = r; +} + +template<> EIGEN_STRONG_INLINE void pstoreu<Eigen::half>(Eigen::half* to, const Packet4h& from) { + __int64_t r = _mm_cvtm64_si64(from.x); + *(reinterpret_cast<__int64_t*>(to)) = r; +} + +template<> EIGEN_STRONG_INLINE Packet4h +ploadquad<Packet4h>(const Eigen::half* from) { + return pset1<Packet4h>(*from); +} + +template<> EIGEN_STRONG_INLINE Packet4h pgather<Eigen::half, Packet4h>(const Eigen::half* from, Index stride) +{ + Packet4h result; + result.x = _mm_set_pi16(from[3*stride].x, from[2*stride].x, from[1*stride].x, from[0*stride].x); + return result; +} + +template<> EIGEN_STRONG_INLINE void pscatter<Eigen::half, Packet4h>(Eigen::half* to, const Packet4h& from, Index stride) +{ + __int64_t a = _mm_cvtm64_si64(from.x); + to[stride*0].x = static_cast<unsigned short>(a); + to[stride*1].x = static_cast<unsigned short>(a >> 16); + to[stride*2].x = static_cast<unsigned short>(a >> 32); + to[stride*3].x = static_cast<unsigned short>(a >> 48); +} + +EIGEN_STRONG_INLINE void +ptranspose(PacketBlock<Packet4h,4>& kernel) { + __m64 T0 = _mm_unpacklo_pi16(kernel.packet[0].x, kernel.packet[1].x); + __m64 T1 = _mm_unpacklo_pi16(kernel.packet[2].x, kernel.packet[3].x); + __m64 T2 = _mm_unpackhi_pi16(kernel.packet[0].x, kernel.packet[1].x); + __m64 T3 = _mm_unpackhi_pi16(kernel.packet[2].x, kernel.packet[3].x); + + kernel.packet[0].x = _mm_unpacklo_pi32(T0, T1); + kernel.packet[1].x = _mm_unpackhi_pi32(T0, T1); + kernel.packet[2].x = _mm_unpacklo_pi32(T2, T3); + kernel.packet[3].x = _mm_unpackhi_pi32(T2, T3); +} + +#endif + +} +} + +#endif // EIGEN_PACKET_MATH_HALF_CUDA_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/TypeCasting.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/TypeCasting.h new file mode 100644 index 000000000..aa5fbce8e --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/TypeCasting.h @@ -0,0 +1,212 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com> +// +// 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_TYPE_CASTING_CUDA_H +#define EIGEN_TYPE_CASTING_CUDA_H + +namespace Eigen { + +namespace internal { + +template<> +struct scalar_cast_op<float, Eigen::half> { + EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op) + typedef Eigen::half result_type; + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half operator() (const float& a) const { + #if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300 + return __float2half(a); + #else + return Eigen::half(a); + #endif + } +}; + +template<> +struct functor_traits<scalar_cast_op<float, Eigen::half> > +{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; }; + + +template<> +struct scalar_cast_op<int, Eigen::half> { + EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op) + typedef Eigen::half result_type; + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half operator() (const int& a) const { + #if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300 + return __float2half(static_cast<float>(a)); + #else + return Eigen::half(static_cast<float>(a)); + #endif + } +}; + +template<> +struct functor_traits<scalar_cast_op<int, Eigen::half> > +{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; }; + + +template<> +struct scalar_cast_op<Eigen::half, float> { + EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op) + typedef float result_type; + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float operator() (const Eigen::half& a) const { + #if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300 + return __half2float(a); + #else + return static_cast<float>(a); + #endif + } +}; + +template<> +struct functor_traits<scalar_cast_op<Eigen::half, float> > +{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; }; + + + +#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300 + +template <> +struct type_casting_traits<Eigen::half, float> { + enum { + VectorizedCast = 1, + SrcCoeffRatio = 2, + TgtCoeffRatio = 1 + }; +}; + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pcast<half2, float4>(const half2& a, const half2& b) { + float2 r1 = __half22float2(a); + float2 r2 = __half22float2(b); + return make_float4(r1.x, r1.y, r2.x, r2.y); +} + +template <> +struct type_casting_traits<float, Eigen::half> { + enum { + VectorizedCast = 1, + SrcCoeffRatio = 1, + TgtCoeffRatio = 2 + }; +}; + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pcast<float4, half2>(const float4& a) { + // Simply discard the second half of the input + return __floats2half2_rn(a.x, a.y); +} + +#elif defined EIGEN_VECTORIZE_AVX512 +template <> +struct type_casting_traits<half, float> { + enum { + VectorizedCast = 1, + SrcCoeffRatio = 1, + TgtCoeffRatio = 1 + }; +}; + +template<> EIGEN_STRONG_INLINE Packet16f pcast<Packet16h, Packet16f>(const Packet16h& a) { + return half2float(a); +} + +template <> +struct type_casting_traits<float, half> { + enum { + VectorizedCast = 1, + SrcCoeffRatio = 1, + TgtCoeffRatio = 1 + }; +}; + +template<> EIGEN_STRONG_INLINE Packet16h pcast<Packet16f, Packet16h>(const Packet16f& a) { + return float2half(a); +} + +#elif defined EIGEN_VECTORIZE_AVX + +template <> +struct type_casting_traits<Eigen::half, float> { + enum { + VectorizedCast = 1, + SrcCoeffRatio = 1, + TgtCoeffRatio = 1 + }; +}; + +template<> EIGEN_STRONG_INLINE Packet8f pcast<Packet8h, Packet8f>(const Packet8h& a) { + return half2float(a); +} + +template <> +struct type_casting_traits<float, Eigen::half> { + enum { + VectorizedCast = 1, + SrcCoeffRatio = 1, + TgtCoeffRatio = 1 + }; +}; + +template<> EIGEN_STRONG_INLINE Packet8h pcast<Packet8f, Packet8h>(const Packet8f& a) { + return float2half(a); +} + +// Disable the following code since it's broken on too many platforms / compilers. +//#elif defined(EIGEN_VECTORIZE_SSE) && (!EIGEN_ARCH_x86_64) && (!EIGEN_COMP_MSVC) +#elif 0 + +template <> +struct type_casting_traits<Eigen::half, float> { + enum { + VectorizedCast = 1, + SrcCoeffRatio = 1, + TgtCoeffRatio = 1 + }; +}; + +template<> EIGEN_STRONG_INLINE Packet4f pcast<Packet4h, Packet4f>(const Packet4h& a) { + __int64_t a64 = _mm_cvtm64_si64(a.x); + Eigen::half h = raw_uint16_to_half(static_cast<unsigned short>(a64)); + float f1 = static_cast<float>(h); + h = raw_uint16_to_half(static_cast<unsigned short>(a64 >> 16)); + float f2 = static_cast<float>(h); + h = raw_uint16_to_half(static_cast<unsigned short>(a64 >> 32)); + float f3 = static_cast<float>(h); + h = raw_uint16_to_half(static_cast<unsigned short>(a64 >> 48)); + float f4 = static_cast<float>(h); + return _mm_set_ps(f4, f3, f2, f1); +} + +template <> +struct type_casting_traits<float, Eigen::half> { + enum { + VectorizedCast = 1, + SrcCoeffRatio = 1, + TgtCoeffRatio = 1 + }; +}; + +template<> EIGEN_STRONG_INLINE Packet4h pcast<Packet4f, Packet4h>(const Packet4f& a) { + EIGEN_ALIGN16 float aux[4]; + pstore(aux, a); + Eigen::half h0(aux[0]); + Eigen::half h1(aux[1]); + Eigen::half h2(aux[2]); + Eigen::half h3(aux[3]); + + Packet4h result; + result.x = _mm_set_pi16(h3.x, h2.x, h1.x, h0.x); + return result; +} + +#endif + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_TYPE_CASTING_CUDA_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/Default/Settings.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/Default/Settings.h new file mode 100644 index 000000000..097373c84 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/Default/Settings.h @@ -0,0 +1,49 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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/. + + +/* All the parameters defined in this file can be specialized in the + * architecture specific files, and/or by the user. + * More to come... */ + +#ifndef EIGEN_DEFAULT_SETTINGS_H +#define EIGEN_DEFAULT_SETTINGS_H + +/** Defines the maximal loop size to enable meta unrolling of loops. + * Note that the value here is expressed in Eigen's own notion of "number of FLOPS", + * it does not correspond to the number of iterations or the number of instructions + */ +#ifndef EIGEN_UNROLLING_LIMIT +#define EIGEN_UNROLLING_LIMIT 100 +#endif + +/** Defines the threshold between a "small" and a "large" matrix. + * This threshold is mainly used to select the proper product implementation. + */ +#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD +#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8 +#endif + +/** Defines the maximal width of the blocks used in the triangular product and solver + * for vectors (level 2 blas xTRMV and xTRSV). The default is 8. + */ +#ifndef EIGEN_TUNE_TRIANGULAR_PANEL_WIDTH +#define EIGEN_TUNE_TRIANGULAR_PANEL_WIDTH 8 +#endif + + +/** Defines the default number of registers available for that architecture. + * Currently it must be 8 or 16. Other values will fail. + */ +#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS +#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 8 +#endif + +#endif // EIGEN_DEFAULT_SETTINGS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/NEON/Complex.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/NEON/Complex.h new file mode 100644 index 000000000..57e9b431f --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/NEON/Complex.h @@ -0,0 +1,486 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2010 Konstantinos Margaritis <markos@freevec.org> +// +// 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_NEON_H +#define EIGEN_COMPLEX_NEON_H + +namespace Eigen { + +namespace internal { + +inline uint32x4_t p4ui_CONJ_XOR() { +// See bug 1325, clang fails to call vld1q_u64. +#if EIGEN_COMP_CLANG + uint32x4_t ret = { 0x00000000, 0x80000000, 0x00000000, 0x80000000 }; + return ret; +#else + static const uint32_t conj_XOR_DATA[] = { 0x00000000, 0x80000000, 0x00000000, 0x80000000 }; + return vld1q_u32( conj_XOR_DATA ); +#endif +} + +inline uint32x2_t p2ui_CONJ_XOR() { + static const uint32_t conj_XOR_DATA[] = { 0x00000000, 0x80000000 }; + return vld1_u32( conj_XOR_DATA ); +} + +//---------- float ---------- +struct Packet2cf +{ + EIGEN_STRONG_INLINE Packet2cf() {} + EIGEN_STRONG_INLINE explicit Packet2cf(const Packet4f& a) : v(a) {} + Packet4f v; +}; + +template<> struct packet_traits<std::complex<float> > : default_packet_traits +{ + typedef Packet2cf type; + typedef Packet2cf half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size = 2, + HasHalfPacket = 0, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasNegate = 1, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasSetLinear = 0 + }; +}; + +template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16}; typedef Packet2cf half; }; + +template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>& from) +{ + float32x2_t r64; + r64 = vld1_f32((float *)&from); + + return Packet2cf(vcombine_f32(r64, r64)); +} + +template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(padd<Packet4f>(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(psub<Packet4f>(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) { return Packet2cf(pnegate<Packet4f>(a.v)); } +template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a) +{ + Packet4ui b = vreinterpretq_u32_f32(a.v); + return Packet2cf(vreinterpretq_f32_u32(veorq_u32(b, p4ui_CONJ_XOR()))); +} + +template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b) +{ + Packet4f v1, v2; + + // Get the real values of a | a1_re | a1_re | a2_re | a2_re | + v1 = vcombine_f32(vdup_lane_f32(vget_low_f32(a.v), 0), vdup_lane_f32(vget_high_f32(a.v), 0)); + // Get the imag values of a | a1_im | a1_im | a2_im | a2_im | + v2 = vcombine_f32(vdup_lane_f32(vget_low_f32(a.v), 1), vdup_lane_f32(vget_high_f32(a.v), 1)); + // Multiply the real a with b + v1 = vmulq_f32(v1, b.v); + // Multiply the imag a with b + v2 = vmulq_f32(v2, b.v); + // Conjugate v2 + v2 = vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(v2), p4ui_CONJ_XOR())); + // Swap real/imag elements in v2. + v2 = vrev64q_f32(v2); + // Add and return the result + return Packet2cf(vaddq_f32(v1, v2)); +} + +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)))); +} +template<> EIGEN_STRONG_INLINE Packet2cf por <Packet2cf>(const Packet2cf& a, const Packet2cf& b) +{ + return Packet2cf(vreinterpretq_f32_u32(vorrq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v)))); +} +template<> EIGEN_STRONG_INLINE Packet2cf pxor <Packet2cf>(const Packet2cf& a, const Packet2cf& b) +{ + return Packet2cf(vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v)))); +} +template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) +{ + return Packet2cf(vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v)))); +} + +template<> EIGEN_STRONG_INLINE Packet2cf pload<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>((const float*)from)); } +template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>((const float*)from)); } + +template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from) { return pset1<Packet2cf>(*from); } + +template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((float*)to, from.v); } +template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((float*)to, from.v); } + +template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from, Index stride) +{ + Packet4f res = pset1<Packet4f>(0.f); + res = vsetq_lane_f32(std::real(from[0*stride]), res, 0); + res = vsetq_lane_f32(std::imag(from[0*stride]), res, 1); + res = vsetq_lane_f32(std::real(from[1*stride]), res, 2); + res = vsetq_lane_f32(std::imag(from[1*stride]), res, 3); + return Packet2cf(res); +} + +template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, Index stride) +{ + to[stride*0] = std::complex<float>(vgetq_lane_f32(from.v, 0), vgetq_lane_f32(from.v, 1)); + to[stride*1] = std::complex<float>(vgetq_lane_f32(from.v, 2), vgetq_lane_f32(from.v, 3)); +} + +template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> * addr) { EIGEN_ARM_PREFETCH((float *)addr); } + +template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a) +{ + std::complex<float> EIGEN_ALIGN16 x[2]; + vst1q_f32((float *)x, a.v); + return x[0]; +} + +template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) +{ + float32x2_t a_lo, a_hi; + Packet4f a_r128; + + a_lo = vget_low_f32(a.v); + a_hi = vget_high_f32(a.v); + a_r128 = vcombine_f32(a_hi, a_lo); + + return Packet2cf(a_r128); +} + +template<> EIGEN_STRONG_INLINE Packet2cf pcplxflip<Packet2cf>(const Packet2cf& a) +{ + return Packet2cf(vrev64q_f32(a.v)); +} + +template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a) +{ + float32x2_t a1, a2; + std::complex<float> s; + + a1 = vget_low_f32(a.v); + a2 = vget_high_f32(a.v); + a2 = vadd_f32(a1, a2); + vst1_f32((float *)&s, a2); + + return s; +} + +template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs) +{ + Packet4f sum1, sum2, sum; + + // Add the first two 64-bit float32x2_t of vecs[0] + sum1 = vcombine_f32(vget_low_f32(vecs[0].v), vget_low_f32(vecs[1].v)); + sum2 = vcombine_f32(vget_high_f32(vecs[0].v), vget_high_f32(vecs[1].v)); + sum = vaddq_f32(sum1, sum2); + + return Packet2cf(sum); +} + +template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a) +{ + float32x2_t a1, a2, v1, v2, prod; + std::complex<float> s; + + a1 = vget_low_f32(a.v); + a2 = vget_high_f32(a.v); + // Get the real values of a | a1_re | a1_re | a2_re | a2_re | + v1 = vdup_lane_f32(a1, 0); + // Get the real values of a | a1_im | a1_im | a2_im | a2_im | + v2 = vdup_lane_f32(a1, 1); + // Multiply the real a with b + v1 = vmul_f32(v1, a2); + // Multiply the imag a with b + v2 = vmul_f32(v2, a2); + // Conjugate v2 + v2 = vreinterpret_f32_u32(veor_u32(vreinterpret_u32_f32(v2), p2ui_CONJ_XOR())); + // Swap real/imag elements in v2. + v2 = vrev64_f32(v2); + // Add v1, v2 + prod = vadd_f32(v1, v2); + + vst1_f32((float *)&s, prod); + + return s; +} + +template<int Offset> +struct palign_impl<Offset,Packet2cf> +{ + EIGEN_STRONG_INLINE static void run(Packet2cf& first, const Packet2cf& second) + { + if (Offset==1) + { + first.v = vextq_f32(first.v, second.v, 2); + } + } +}; + +template<> struct conj_helper<Packet2cf, Packet2cf, false,true> +{ + EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const + { + return internal::pmul(a, pconj(b)); + } +}; + +template<> struct conj_helper<Packet2cf, Packet2cf, true,false> +{ + EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const + { + return internal::pmul(pconj(a), b); + } +}; + +template<> struct conj_helper<Packet2cf, Packet2cf, true,true> +{ + EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const + { + return pconj(internal::pmul(a, b)); + } +}; + +template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b) +{ + // TODO optimize it for NEON + Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a,b); + Packet4f s, rev_s; + + // this computes the norm + s = vmulq_f32(b.v, b.v); + rev_s = vrev64q_f32(s); + + return Packet2cf(pdiv(res.v, vaddq_f32(s,rev_s))); +} + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet2cf,2>& kernel) { + Packet4f tmp = vcombine_f32(vget_high_f32(kernel.packet[0].v), vget_high_f32(kernel.packet[1].v)); + kernel.packet[0].v = vcombine_f32(vget_low_f32(kernel.packet[0].v), vget_low_f32(kernel.packet[1].v)); + kernel.packet[1].v = tmp; +} + +//---------- double ---------- +#if EIGEN_ARCH_ARM64 && !EIGEN_APPLE_DOUBLE_NEON_BUG + +// See bug 1325, clang fails to call vld1q_u64. +#if EIGEN_COMP_CLANG + static uint64x2_t p2ul_CONJ_XOR = {0x0, 0x8000000000000000}; +#else + const uint64_t p2ul_conj_XOR_DATA[] = { 0x0, 0x8000000000000000 }; + static uint64x2_t p2ul_CONJ_XOR = vld1q_u64( p2ul_conj_XOR_DATA ); +#endif + +struct Packet1cd +{ + EIGEN_STRONG_INLINE Packet1cd() {} + EIGEN_STRONG_INLINE explicit Packet1cd(const Packet2d& a) : v(a) {} + Packet2d v; +}; + +template<> struct packet_traits<std::complex<double> > : default_packet_traits +{ + typedef Packet1cd type; + typedef Packet1cd half; + enum { + Vectorizable = 1, + AlignedOnScalar = 0, + size = 1, + HasHalfPacket = 0, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasNegate = 1, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasSetLinear = 0 + }; +}; + +template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16}; 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)); } + +template<> EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<double>& from) +{ /* here we really have to use unaligned loads :( */ return ploadu<Packet1cd>(&from); } + +template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(padd<Packet2d>(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(psub<Packet2d>(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate<Packet2d>(a.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) { return Packet1cd(vreinterpretq_f64_u64(veorq_u64(vreinterpretq_u64_f64(a.v), p2ul_CONJ_XOR))); } + +template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b) +{ + Packet2d v1, v2; + + // Get the real values of a + v1 = vdupq_lane_f64(vget_low_f64(a.v), 0); + // Get the imag values of a + v2 = vdupq_lane_f64(vget_high_f64(a.v), 0); + // Multiply the real a with b + v1 = vmulq_f64(v1, b.v); + // Multiply the imag a with b + v2 = vmulq_f64(v2, b.v); + // Conjugate v2 + v2 = vreinterpretq_f64_u64(veorq_u64(vreinterpretq_u64_f64(v2), p2ul_CONJ_XOR)); + // Swap real/imag elements in v2. + v2 = preverse<Packet2d>(v2); + // Add and return the result + return Packet1cd(vaddq_f64(v1, v2)); +} + +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)))); +} +template<> EIGEN_STRONG_INLINE Packet1cd por <Packet1cd>(const Packet1cd& a, const Packet1cd& b) +{ + return Packet1cd(vreinterpretq_f64_u64(vorrq_u64(vreinterpretq_u64_f64(a.v),vreinterpretq_u64_f64(b.v)))); +} +template<> EIGEN_STRONG_INLINE Packet1cd pxor <Packet1cd>(const Packet1cd& a, const Packet1cd& b) +{ + return Packet1cd(vreinterpretq_f64_u64(veorq_u64(vreinterpretq_u64_f64(a.v),vreinterpretq_u64_f64(b.v)))); +} +template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) +{ + return Packet1cd(vreinterpretq_f64_u64(vbicq_u64(vreinterpretq_u64_f64(a.v),vreinterpretq_u64_f64(b.v)))); +} + +template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>* from) { return pset1<Packet1cd>(*from); } + +template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> * to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); } +template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> * to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); } + +template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> * addr) { EIGEN_ARM_PREFETCH((double *)addr); } + +template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather<std::complex<double>, Packet1cd>(const std::complex<double>* from, Index stride) +{ + Packet2d res = pset1<Packet2d>(0.0); + res = vsetq_lane_f64(std::real(from[0*stride]), res, 0); + res = vsetq_lane_f64(std::imag(from[0*stride]), res, 1); + return Packet1cd(res); +} + +template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet1cd>(std::complex<double>* to, const Packet1cd& from, Index stride) +{ + to[stride*0] = std::complex<double>(vgetq_lane_f64(from.v, 0), vgetq_lane_f64(from.v, 1)); +} + + +template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Packet1cd& a) +{ + std::complex<double> EIGEN_ALIGN16 res; + pstore<std::complex<double> >(&res, a); + + return res; +} + +template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; } + +template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a) { return pfirst(a); } + +template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs) { return vecs[0]; } + +template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a) { return pfirst(a); } + +template<int Offset> +struct palign_impl<Offset,Packet1cd> +{ + static EIGEN_STRONG_INLINE void run(Packet1cd& /*first*/, const Packet1cd& /*second*/) + { + // FIXME is it sure we never have to align a Packet1cd? + // Even though a std::complex<double> has 16 bytes, it is not necessarily aligned on a 16 bytes boundary... + } +}; + +template<> struct conj_helper<Packet1cd, Packet1cd, false,true> +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const + { + return internal::pmul(a, pconj(b)); + } +}; + +template<> struct conj_helper<Packet1cd, Packet1cd, true,false> +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const + { + return internal::pmul(pconj(a), b); + } +}; + +template<> struct conj_helper<Packet1cd, Packet1cd, true,true> +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const + { + return pconj(internal::pmul(a, b)); + } +}; + +template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b) +{ + // TODO optimize it for NEON + Packet1cd res = conj_helper<Packet1cd,Packet1cd,false,true>().pmul(a,b); + Packet2d s = pmul<Packet2d>(b.v, b.v); + Packet2d rev_s = preverse<Packet2d>(s); + + return Packet1cd(pdiv(res.v, padd<Packet2d>(s,rev_s))); +} + +EIGEN_STRONG_INLINE Packet1cd pcplxflip/*<Packet1cd>*/(const Packet1cd& x) +{ + return Packet1cd(preverse(Packet2d(x.v))); +} + +EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet1cd,2>& kernel) +{ + Packet2d tmp = vcombine_f64(vget_high_f64(kernel.packet[0].v), vget_high_f64(kernel.packet[1].v)); + kernel.packet[0].v = vcombine_f64(vget_low_f64(kernel.packet[0].v), vget_low_f64(kernel.packet[1].v)); + kernel.packet[1].v = tmp; +} +#endif // EIGEN_ARCH_ARM64 + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_COMPLEX_NEON_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/NEON/MathFunctions.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/NEON/MathFunctions.h new file mode 100644 index 000000000..6bb05bb92 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/NEON/MathFunctions.h @@ -0,0 +1,91 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// 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 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 + +namespace Eigen { + +namespace internal { + +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +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; +} + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_MATH_FUNCTIONS_NEON_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/NEON/PacketMath.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/NEON/PacketMath.h new file mode 100644 index 000000000..836fbc0dd --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/NEON/PacketMath.h @@ -0,0 +1,729 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2010 Konstantinos Margaritis <markos@freevec.org> +// Heavily based on Gael's SSE version. +// +// 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_PACKET_MATH_NEON_H +#define EIGEN_PACKET_MATH_NEON_H + +namespace Eigen { + +namespace internal { + +#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD +#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8 +#endif + +#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD +#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD +#endif + +#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD +#define EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD +#endif + +#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS +#if EIGEN_ARCH_ARM64 +#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 32 +#else +#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 16 +#endif +#endif + +typedef float32x2_t Packet2f; +typedef float32x4_t Packet4f; +typedef int32x4_t Packet4i; +typedef int32x2_t Packet2i; +typedef uint32x4_t Packet4ui; + +#define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \ + const Packet4f p4f_##NAME = pset1<Packet4f>(X) + +#define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \ + const Packet4f p4f_##NAME = vreinterpretq_f32_u32(pset1<int32_t>(X)) + +#define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \ + const Packet4i p4i_##NAME = pset1<Packet4i>(X) + +#if EIGEN_ARCH_ARM64 + // __builtin_prefetch tends to do nothing on ARM64 compilers because the + // prefetch instructions there are too detailed for __builtin_prefetch to map + // meaningfully to them. + #define EIGEN_ARM_PREFETCH(ADDR) __asm__ __volatile__("prfm pldl1keep, [%[addr]]\n" ::[addr] "r"(ADDR) : ); +#elif EIGEN_HAS_BUILTIN(__builtin_prefetch) || EIGEN_COMP_GNUC + #define EIGEN_ARM_PREFETCH(ADDR) __builtin_prefetch(ADDR); +#elif defined __pld + #define EIGEN_ARM_PREFETCH(ADDR) __pld(ADDR) +#elif EIGEN_ARCH_ARM32 + #define EIGEN_ARM_PREFETCH(ADDR) __asm__ __volatile__ ("pld [%[addr]]\n" :: [addr] "r" (ADDR) : ); +#else + // by default no explicit prefetching + #define EIGEN_ARM_PREFETCH(ADDR) +#endif + +template<> struct packet_traits<float> : default_packet_traits +{ + typedef Packet4f type; + typedef Packet4f half; // Packet2f intrinsics not implemented yet + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size = 4, + HasHalfPacket=0, // Packet2f intrinsics not implemented yet + + HasDiv = 1, + // FIXME check the Has* + HasSin = 0, + HasCos = 0, + HasLog = 0, + HasExp = 1, + HasSqrt = 0 + }; +}; +template<> struct packet_traits<int32_t> : default_packet_traits +{ + typedef Packet4i type; + typedef Packet4i half; // Packet2i intrinsics not implemented yet + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=4, + HasHalfPacket=0 // Packet2i intrinsics not implemented yet + // FIXME check the Has* + }; +}; + +#if EIGEN_GNUC_AT_MOST(4,4) && !EIGEN_COMP_LLVM +// workaround gcc 4.2, 4.3 and 4.4 compilatin issue +EIGEN_STRONG_INLINE float32x4_t vld1q_f32(const float* x) { return ::vld1q_f32((const float32_t*)x); } +EIGEN_STRONG_INLINE float32x2_t vld1_f32 (const float* x) { return ::vld1_f32 ((const float32_t*)x); } +EIGEN_STRONG_INLINE float32x2_t vld1_dup_f32 (const float* x) { return ::vld1_dup_f32 ((const float32_t*)x); } +EIGEN_STRONG_INLINE void vst1q_f32(float* to, float32x4_t from) { ::vst1q_f32((float32_t*)to,from); } +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<> 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 plset<Packet4f>(const float& a) +{ + const float f[] = {0, 1, 2, 3}; + Packet4f countdown = vld1q_f32(f); + return vaddq_f32(pset1<Packet4f>(a), countdown); +} +template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int32_t& a) +{ + const int32_t i[] = {0, 1, 2, 3}; + Packet4i countdown = vld1q_s32(i); + return vaddq_s32(pset1<Packet4i>(a), countdown); +} + +template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return vaddq_f32(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return vaddq_s32(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { return vsubq_f32(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return vsubq_s32(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a) { return vnegq_f32(a); } +template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return vnegq_s32(a); } + +template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; } +template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; } + +template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return vmulq_f32(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) { return vmulq_s32(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) +{ +#if EIGEN_ARCH_ARM64 + return vdivq_f32(a,b); +#else + Packet4f inv, restep, div; + + // NEON does not offer a divide instruction, we have to do a reciprocal approximation + // However NEON in contrast to other SIMD engines (AltiVec/SSE), offers + // a reciprocal estimate AND a reciprocal step -which saves a few instructions + // vrecpeq_f32() returns an estimate to 1/b, which we will finetune with + // Newton-Raphson and vrecpsq_f32() + inv = vrecpeq_f32(b); + + // This returns a differential, by which we will have to multiply inv to get a better + // approximation of 1/b. + restep = vrecpsq_f32(b, inv); + inv = vmulq_f32(restep, inv); + + // Finally, multiply a by 1/b and get the wanted result of the division. + div = vmulq_f32(a, inv); + + return div; +#endif +} + +template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, const Packet4i& /*b*/) +{ eigen_assert(false && "packet integer division are not supported by NEON"); + return pset1<Packet4i>(0); +} + +// Clang/ARM wrongly advertises __ARM_FEATURE_FMA even when it's not available, +// then implements a slow software scalar fallback calling fmaf()! +// Filed LLVM bug: +// https://llvm.org/bugs/show_bug.cgi?id=27216 +#if (defined __ARM_FEATURE_FMA) && !(EIGEN_COMP_CLANG && EIGEN_ARCH_ARM) +// See bug 936. +// FMA is available on VFPv4 i.e. when compiling with -mfpu=neon-vfpv4. +// FMA is a true fused multiply-add i.e. only 1 rounding at the end, no intermediate rounding. +// MLA is not fused i.e. does 2 roundings. +// In addition to giving better accuracy, FMA also gives better performance here on a Krait (Nexus 4): +// MLA: 10 GFlop/s ; FMA: 12 GFlops/s. +template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vfmaq_f32(c,a,b); } +#else +template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { +#if EIGEN_COMP_CLANG && EIGEN_ARCH_ARM + // Clang/ARM will replace VMLA by VMUL+VADD at least for some values of -mcpu, + // at least -mcpu=cortex-a8 and -mcpu=cortex-a7. Since the former is the default on + // -march=armv7-a, that is a very common case. + // See e.g. this thread: + // http://lists.llvm.org/pipermail/llvm-dev/2013-December/068806.html + // Filed LLVM bug: + // https://llvm.org/bugs/show_bug.cgi?id=27219 + Packet4f r = c; + asm volatile( + "vmla.f32 %q[r], %q[a], %q[b]" + : [r] "+w" (r) + : [a] "w" (a), + [b] "w" (b) + : ); + return r; +#else + return vmlaq_f32(c,a,b); +#endif +} +#endif + +// No FMA instruction for int, so use MLA unconditionally. +template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return vmlaq_s32(c,a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { return vminq_f32(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) { return vminq_s32(a,b); } + +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); } + +// 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) +{ + return vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(a),vreinterpretq_u32_f32(b))); +} +template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return vandq_s32(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b) +{ + return vreinterpretq_f32_u32(vorrq_u32(vreinterpretq_u32_f32(a),vreinterpretq_u32_f32(b))); +} +template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return vorrq_s32(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b) +{ + return vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(a),vreinterpretq_u32_f32(b))); +} +template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return veorq_s32(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) +{ + return vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(a),vreinterpretq_u32_f32(b))); +} +template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return vbicq_s32(a,b); } + +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); } + +template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) { EIGEN_DEBUG_UNALIGNED_LOAD return vld1q_f32(from); } +template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int32_t* from) { EIGEN_DEBUG_UNALIGNED_LOAD return vld1q_s32(from); } + +template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from) +{ + float32x2_t lo, hi; + lo = vld1_dup_f32(from); + hi = vld1_dup_f32(from+1); + return vcombine_f32(lo, hi); +} +template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int32_t* from) +{ + int32x2_t lo, hi; + lo = vld1_dup_s32(from); + hi = vld1_dup_s32(from+1); + return vcombine_s32(lo, hi); +} + +template<> EIGEN_STRONG_INLINE void pstore<float> (float* to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_f32(to, from); } +template<> EIGEN_STRONG_INLINE void pstore<int32_t>(int32_t* to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_s32(to, from); } + +template<> EIGEN_STRONG_INLINE void pstoreu<float> (float* to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_f32(to, from); } +template<> EIGEN_STRONG_INLINE void pstoreu<int32_t>(int32_t* to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_s32(to, from); } + +template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride) +{ + Packet4f res = pset1<Packet4f>(0.f); + res = vsetq_lane_f32(from[0*stride], res, 0); + res = vsetq_lane_f32(from[1*stride], res, 1); + res = vsetq_lane_f32(from[2*stride], res, 2); + res = vsetq_lane_f32(from[3*stride], res, 3); + return res; +} +template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int32_t, Packet4i>(const int32_t* from, Index stride) +{ + Packet4i res = pset1<Packet4i>(0); + res = vsetq_lane_s32(from[0*stride], res, 0); + res = vsetq_lane_s32(from[1*stride], res, 1); + res = vsetq_lane_s32(from[2*stride], res, 2); + res = vsetq_lane_s32(from[3*stride], res, 3); + return res; +} + +template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride) +{ + to[stride*0] = vgetq_lane_f32(from, 0); + to[stride*1] = vgetq_lane_f32(from, 1); + to[stride*2] = vgetq_lane_f32(from, 2); + to[stride*3] = vgetq_lane_f32(from, 3); +} +template<> EIGEN_DEVICE_FUNC inline void pscatter<int32_t, Packet4i>(int32_t* to, const Packet4i& from, Index stride) +{ + to[stride*0] = vgetq_lane_s32(from, 0); + to[stride*1] = vgetq_lane_s32(from, 1); + to[stride*2] = vgetq_lane_s32(from, 2); + to[stride*3] = vgetq_lane_s32(from, 3); +} + +template<> EIGEN_STRONG_INLINE void prefetch<float> (const float* addr) { EIGEN_ARM_PREFETCH(addr); } +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 Packet4f preverse(const Packet4f& a) { + float32x2_t a_lo, a_hi; + Packet4f a_r64; + + a_r64 = vrev64q_f32(a); + a_lo = vget_low_f32(a_r64); + a_hi = vget_high_f32(a_r64); + return vcombine_f32(a_hi, a_lo); +} +template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) { + int32x2_t a_lo, a_hi; + Packet4i a_r64; + + a_r64 = vrev64q_s32(a); + a_lo = vget_low_s32(a_r64); + a_hi = vget_high_s32(a_r64); + return vcombine_s32(a_hi, a_lo); +} + +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 float predux<Packet4f>(const Packet4f& a) +{ + float32x2_t a_lo, a_hi, sum; + + a_lo = vget_low_f32(a); + a_hi = vget_high_f32(a); + sum = vpadd_f32(a_lo, a_hi); + sum = vpadd_f32(sum, sum); + return vget_lane_f32(sum, 0); +} + +template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs) +{ + float32x4x2_t vtrn1, vtrn2, res1, res2; + Packet4f sum1, sum2, sum; + + // NEON zip performs interleaving of the supplied vectors. + // We perform two interleaves in a row to acquire the transposed vector + vtrn1 = vzipq_f32(vecs[0], vecs[2]); + vtrn2 = vzipq_f32(vecs[1], vecs[3]); + res1 = vzipq_f32(vtrn1.val[0], vtrn2.val[0]); + res2 = vzipq_f32(vtrn1.val[1], vtrn2.val[1]); + + // Do the addition of the resulting vectors + sum1 = vaddq_f32(res1.val[0], res1.val[1]); + sum2 = vaddq_f32(res2.val[0], res2.val[1]); + sum = vaddq_f32(sum1, sum2); + + return sum; +} + +template<> EIGEN_STRONG_INLINE int32_t predux<Packet4i>(const Packet4i& a) +{ + int32x2_t a_lo, a_hi, sum; + + a_lo = vget_low_s32(a); + a_hi = vget_high_s32(a); + sum = vpadd_s32(a_lo, a_hi); + sum = vpadd_s32(sum, sum); + return vget_lane_s32(sum, 0); +} + +template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs) +{ + int32x4x2_t vtrn1, vtrn2, res1, res2; + Packet4i sum1, sum2, sum; + + // NEON zip performs interleaving of the supplied vectors. + // We perform two interleaves in a row to acquire the transposed vector + vtrn1 = vzipq_s32(vecs[0], vecs[2]); + vtrn2 = vzipq_s32(vecs[1], vecs[3]); + res1 = vzipq_s32(vtrn1.val[0], vtrn2.val[0]); + res2 = vzipq_s32(vtrn1.val[1], vtrn2.val[1]); + + // Do the addition of the resulting vectors + sum1 = vaddq_s32(res1.val[0], res1.val[1]); + sum2 = vaddq_s32(res2.val[0], res2.val[1]); + sum = vaddq_s32(sum1, sum2); + + return sum; +} + +// Other reduction functions: +// mul +template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a) +{ + float32x2_t a_lo, a_hi, prod; + + // Get a_lo = |a1|a2| and a_hi = |a3|a4| + a_lo = vget_low_f32(a); + a_hi = vget_high_f32(a); + // Get the product of a_lo * a_hi -> |a1*a3|a2*a4| + prod = vmul_f32(a_lo, a_hi); + // Multiply prod with its swapped value |a2*a4|a1*a3| + prod = vmul_f32(prod, vrev64_f32(prod)); + + return vget_lane_f32(prod, 0); +} +template<> EIGEN_STRONG_INLINE int32_t predux_mul<Packet4i>(const Packet4i& a) +{ + int32x2_t a_lo, a_hi, prod; + + // Get a_lo = |a1|a2| and a_hi = |a3|a4| + a_lo = vget_low_s32(a); + a_hi = vget_high_s32(a); + // Get the product of a_lo * a_hi -> |a1*a3|a2*a4| + prod = vmul_s32(a_lo, a_hi); + // Multiply prod with its swapped value |a2*a4|a1*a3| + prod = vmul_s32(prod, vrev64_s32(prod)); + + return vget_lane_s32(prod, 0); +} + +// min +template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a) +{ + float32x2_t a_lo, a_hi, min; + + a_lo = vget_low_f32(a); + a_hi = vget_high_f32(a); + min = vpmin_f32(a_lo, a_hi); + min = vpmin_f32(min, min); + + return vget_lane_f32(min, 0); +} + +template<> EIGEN_STRONG_INLINE int32_t predux_min<Packet4i>(const Packet4i& a) +{ + int32x2_t a_lo, a_hi, min; + + a_lo = vget_low_s32(a); + a_hi = vget_high_s32(a); + min = vpmin_s32(a_lo, a_hi); + min = vpmin_s32(min, min); + + return vget_lane_s32(min, 0); +} + +// max +template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a) +{ + float32x2_t a_lo, a_hi, max; + + a_lo = vget_low_f32(a); + a_hi = vget_high_f32(a); + max = vpmax_f32(a_lo, a_hi); + max = vpmax_f32(max, max); + + return vget_lane_f32(max, 0); +} + +template<> EIGEN_STRONG_INLINE int32_t predux_max<Packet4i>(const Packet4i& a) +{ + int32x2_t a_lo, a_hi, max; + + a_lo = vget_low_s32(a); + a_hi = vget_high_s32(a); + max = vpmax_s32(a_lo, a_hi); + max = vpmax_s32(max, max); + + return vget_lane_s32(max, 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) \ +template<>\ +struct palign_impl<Offset,Type>\ +{\ + EIGEN_STRONG_INLINE static void run(Type& first, const Type& second)\ + {\ + if (Offset!=0)\ + first = Command(first, second, Offset);\ + }\ +};\ + +PALIGN_NEON(0,Packet4f,vextq_f32) +PALIGN_NEON(1,Packet4f,vextq_f32) +PALIGN_NEON(2,Packet4f,vextq_f32) +PALIGN_NEON(3,Packet4f,vextq_f32) +PALIGN_NEON(0,Packet4i,vextq_s32) +PALIGN_NEON(1,Packet4i,vextq_s32) +PALIGN_NEON(2,Packet4i,vextq_s32) +PALIGN_NEON(3,Packet4i,vextq_s32) + +#undef PALIGN_NEON + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet4f,4>& kernel) { + float32x4x2_t tmp1 = vzipq_f32(kernel.packet[0], kernel.packet[1]); + float32x4x2_t tmp2 = vzipq_f32(kernel.packet[2], kernel.packet[3]); + + kernel.packet[0] = vcombine_f32(vget_low_f32(tmp1.val[0]), vget_low_f32(tmp2.val[0])); + kernel.packet[1] = vcombine_f32(vget_high_f32(tmp1.val[0]), vget_high_f32(tmp2.val[0])); + kernel.packet[2] = vcombine_f32(vget_low_f32(tmp1.val[1]), vget_low_f32(tmp2.val[1])); + kernel.packet[3] = vcombine_f32(vget_high_f32(tmp1.val[1]), vget_high_f32(tmp2.val[1])); +} + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet4i,4>& kernel) { + int32x4x2_t tmp1 = vzipq_s32(kernel.packet[0], kernel.packet[1]); + int32x4x2_t tmp2 = vzipq_s32(kernel.packet[2], kernel.packet[3]); + kernel.packet[0] = vcombine_s32(vget_low_s32(tmp1.val[0]), vget_low_s32(tmp2.val[0])); + kernel.packet[1] = vcombine_s32(vget_high_s32(tmp1.val[0]), vget_high_s32(tmp2.val[0])); + kernel.packet[2] = vcombine_s32(vget_low_s32(tmp1.val[1]), vget_low_s32(tmp2.val[1])); + kernel.packet[3] = vcombine_s32(vget_high_s32(tmp1.val[1]), vget_high_s32(tmp2.val[1])); +} + +//---------- double ---------- + +// Clang 3.5 in the iOS toolchain has an ICE triggered by NEON intrisics for double. +// Confirmed at least with __apple_build_version__ = 6000054. +#ifdef __apple_build_version__ +// Let's hope that by the time __apple_build_version__ hits the 601* range, the bug will be fixed. +// https://gist.github.com/yamaya/2924292 suggests that the 3 first digits are only updated with +// major toolchain updates. +#define EIGEN_APPLE_DOUBLE_NEON_BUG (__apple_build_version__ < 6010000) +#else +#define EIGEN_APPLE_DOUBLE_NEON_BUG 0 +#endif + +#if EIGEN_ARCH_ARM64 && !EIGEN_APPLE_DOUBLE_NEON_BUG + +// Bug 907: workaround missing declarations of the following two functions in the ADK +// Defining these functions as templates ensures that if these intrinsics are +// already defined in arm_neon.h, then our workaround doesn't cause a conflict +// and has lower priority in overload resolution. +template <typename T> +uint64x2_t vreinterpretq_u64_f64(T a) +{ + return (uint64x2_t) a; +} + +template <typename T> +float64x2_t vreinterpretq_f64_u64(T a) +{ + return (float64x2_t) a; +} + +typedef float64x2_t Packet2d; +typedef float64x1_t Packet1d; + +template<> struct packet_traits<double> : default_packet_traits +{ + typedef Packet2d type; + typedef Packet2d half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size = 2, + HasHalfPacket=0, + + HasDiv = 1, + // FIXME check the Has* + HasSin = 0, + HasCos = 0, + HasLog = 0, + HasExp = 0, + HasSqrt = 0 + }; +}; + +template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16}; typedef Packet2d half; }; + +template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { return vdupq_n_f64(from); } + +template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a) +{ + const double countdown_raw[] = {0.0,1.0}; + const Packet2d countdown = vld1q_f64(countdown_raw); + return vaddq_f64(pset1<Packet2d>(a), countdown); +} +template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return vaddq_f64(a,b); } + +template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return vsubq_f64(a,b); } + +template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) { return vnegq_f64(a); } + +template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; } + +template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return vmulq_f64(a,b); } + +template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return vdivq_f64(a,b); } + +#ifdef __ARM_FEATURE_FMA +// See bug 936. See above comment about FMA for float. +template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vfmaq_f64(c,a,b); } +#else +template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vmlaq_f64(c,a,b); } +#endif + +template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return vminq_f64(a,b); } + +template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return vmaxq_f64(a,b); } + +// Logical Operations are not supported for float, so we have to reinterpret casts using NEON intrinsics +template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) +{ + return vreinterpretq_f64_u64(vandq_u64(vreinterpretq_u64_f64(a),vreinterpretq_u64_f64(b))); +} + +template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) +{ + return vreinterpretq_f64_u64(vorrq_u64(vreinterpretq_u64_f64(a),vreinterpretq_u64_f64(b))); +} + +template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) +{ + return vreinterpretq_f64_u64(veorq_u64(vreinterpretq_u64_f64(a),vreinterpretq_u64_f64(b))); +} + +template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) +{ + return vreinterpretq_f64_u64(vbicq_u64(vreinterpretq_u64_f64(a),vreinterpretq_u64_f64(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); } + +template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double* from) +{ + return vld1q_dup_f64(from); +} +template<> EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet2d& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_f64(to, from); } + +template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_f64(to, from); } + +template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride) +{ + Packet2d res = pset1<Packet2d>(0.0); + res = vsetq_lane_f64(from[0*stride], res, 0); + res = vsetq_lane_f64(from[1*stride], res, 1); + return res; +} +template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride) +{ + to[stride*0] = vgetq_lane_f64(from, 0); + to[stride*1] = vgetq_lane_f64(from, 1); +} +template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { EIGEN_ARM_PREFETCH(addr); } + +// FIXME only store the 2 first elements ? +template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { return vgetq_lane_f64(a, 0); } + +template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a) { return vcombine_f64(vget_high_f64(a), vget_low_f64(a)); } + +template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a) { return vabsq_f64(a); } + +#if EIGEN_COMP_CLANG && defined(__apple_build_version__) +// workaround ICE, see bug 907 +template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) { return (vget_low_f64(a) + vget_high_f64(a))[0]; } +#else +template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) { return vget_lane_f64(vget_low_f64(a) + vget_high_f64(a), 0); } +#endif + +template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs) +{ + float64x2_t trn1, trn2; + + // NEON zip performs interleaving of the supplied vectors. + // We perform two interleaves in a row to acquire the transposed vector + trn1 = vzip1q_f64(vecs[0], vecs[1]); + trn2 = vzip2q_f64(vecs[0], vecs[1]); + + // Do the addition of the resulting vectors + return vaddq_f64(trn1, trn2); +} +// Other reduction functions: +// mul +#if EIGEN_COMP_CLANG && defined(__apple_build_version__) +template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a) { return (vget_low_f64(a) * vget_high_f64(a))[0]; } +#else +template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a) { return vget_lane_f64(vget_low_f64(a) * vget_high_f64(a), 0); } +#endif + +// min +template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a) { return vgetq_lane_f64(vpminq_f64(a, a), 0); } + +// max +template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a) { return vgetq_lane_f64(vpmaxq_f64(a, a), 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) \ +template<>\ +struct palign_impl<Offset,Type>\ +{\ + EIGEN_STRONG_INLINE static void run(Type& first, const Type& second)\ + {\ + if (Offset!=0)\ + first = Command(first, second, Offset);\ + }\ +};\ + +PALIGN_NEON(0,Packet2d,vextq_f64) +PALIGN_NEON(1,Packet2d,vextq_f64) +#undef PALIGN_NEON + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet2d,2>& kernel) { + float64x2_t trn1 = vzip1q_f64(kernel.packet[0], kernel.packet[1]); + float64x2_t trn2 = vzip2q_f64(kernel.packet[0], kernel.packet[1]); + + kernel.packet[0] = trn1; + kernel.packet[1] = trn2; +} +#endif // EIGEN_ARCH_ARM64 + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_PACKET_MATH_NEON_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/SSE/Complex.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/SSE/Complex.h new file mode 100644 index 000000000..5607fe0ab --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/SSE/Complex.h @@ -0,0 +1,503 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2010 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_SSE_H +#define EIGEN_COMPLEX_SSE_H + +namespace Eigen { + +namespace internal { + +//---------- float ---------- +struct Packet2cf +{ + EIGEN_STRONG_INLINE Packet2cf() {} + EIGEN_STRONG_INLINE explicit Packet2cf(const __m128& a) : v(a) {} + __m128 v; +}; + +// Use the packet_traits defined in AVX/PacketMath.h instead if we're going +// to leverage AVX instructions. +#ifndef EIGEN_VECTORIZE_AVX +template<> struct packet_traits<std::complex<float> > : default_packet_traits +{ + typedef Packet2cf type; + typedef Packet2cf half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size = 2, + HasHalfPacket = 0, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasNegate = 1, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasSetLinear = 0, + HasBlend = 1 + }; +}; +#endif + +template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16}; 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)); } +template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) +{ + const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000,0x80000000,0x80000000,0x80000000)); + return Packet2cf(_mm_xor_ps(a.v,mask)); +} +template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a) +{ + const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000)); + return Packet2cf(_mm_xor_ps(a.v,mask)); +} + +template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b) +{ + #ifdef EIGEN_VECTORIZE_SSE3 + return Packet2cf(_mm_addsub_ps(_mm_mul_ps(_mm_moveldup_ps(a.v), b.v), + _mm_mul_ps(_mm_movehdup_ps(a.v), + vec4f_swizzle1(b.v, 1, 0, 3, 2)))); +// return Packet2cf(_mm_addsub_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), +// _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3), +// vec4f_swizzle1(b.v, 1, 0, 3, 2)))); + #else + const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000,0x00000000,0x80000000,0x00000000)); + return Packet2cf(_mm_add_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), + _mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3), + vec4f_swizzle1(b.v, 1, 0, 3, 2)), mask))); + #endif +} + +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 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))); } + +template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>& from) +{ + Packet2cf res; +#if EIGEN_GNUC_AT_MOST(4,2) + // Workaround annoying "may be used uninitialized in this function" warning with gcc 4.2 + res.v = _mm_loadl_pi(_mm_set1_ps(0.0f), reinterpret_cast<const __m64*>(&from)); +#elif EIGEN_GNUC_AT_LEAST(4,6) + // Suppress annoying "may be used uninitialized in this function" warning with gcc >= 4.6 + #pragma GCC diagnostic push + #pragma GCC diagnostic ignored "-Wuninitialized" + res.v = _mm_loadl_pi(res.v, (const __m64*)&from); + #pragma GCC diagnostic pop +#else + res.v = _mm_loadl_pi(res.v, (const __m64*)&from); +#endif + return Packet2cf(_mm_movelh_ps(res.v,res.v)); +} + +template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from) { return pset1<Packet2cf>(*from); } + +template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore(&numext::real_ref(*to), Packet4f(from.v)); } +template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&numext::real_ref(*to), Packet4f(from.v)); } + + +template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from, Index stride) +{ + return Packet2cf(_mm_set_ps(std::imag(from[1*stride]), std::real(from[1*stride]), + std::imag(from[0*stride]), std::real(from[0*stride]))); +} + +template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, Index stride) +{ + to[stride*0] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 0)), + _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 1))); + to[stride*1] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 2)), + _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 3))); +} + +template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> * addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } + +template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a) +{ + #if EIGEN_GNUC_AT_MOST(4,3) + // Workaround gcc 4.2 ICE - this is not performance wise ideal, but who cares... + // This workaround also fix invalid code generation with gcc 4.3 + EIGEN_ALIGN16 std::complex<float> res[2]; + _mm_store_ps((float*)res, a.v); + return res[0]; + #else + std::complex<float> res; + _mm_storel_pi((__m64*)&res, a.v); + return res; + #endif +} + +template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) { return Packet2cf(_mm_castpd_ps(preverse(Packet2d(_mm_castps_pd(a.v))))); } + +template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a) +{ + return pfirst(Packet2cf(_mm_add_ps(a.v, _mm_movehl_ps(a.v,a.v)))); +} + +template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs) +{ + return Packet2cf(_mm_add_ps(_mm_movelh_ps(vecs[0].v,vecs[1].v), _mm_movehl_ps(vecs[1].v,vecs[0].v))); +} + +template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a) +{ + return pfirst(pmul(a, Packet2cf(_mm_movehl_ps(a.v,a.v)))); +} + +template<int Offset> +struct palign_impl<Offset,Packet2cf> +{ + static EIGEN_STRONG_INLINE void run(Packet2cf& first, const Packet2cf& second) + { + if (Offset==1) + { + first.v = _mm_movehl_ps(first.v, first.v); + first.v = _mm_movelh_ps(first.v, second.v); + } + } +}; + +template<> struct conj_helper<Packet2cf, Packet2cf, false,true> +{ + EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const + { + #ifdef EIGEN_VECTORIZE_SSE3 + return internal::pmul(a, pconj(b)); + #else + const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000)); + return Packet2cf(_mm_add_ps(_mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), mask), + _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3), + vec4f_swizzle1(b.v, 1, 0, 3, 2)))); + #endif + } +}; + +template<> struct conj_helper<Packet2cf, Packet2cf, true,false> +{ + EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const + { + #ifdef EIGEN_VECTORIZE_SSE3 + return internal::pmul(pconj(a), b); + #else + const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000)); + return Packet2cf(_mm_add_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), + _mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3), + vec4f_swizzle1(b.v, 1, 0, 3, 2)), mask))); + #endif + } +}; + +template<> struct conj_helper<Packet2cf, Packet2cf, true,true> +{ + EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const + { + #ifdef EIGEN_VECTORIZE_SSE3 + return pconj(internal::pmul(a, b)); + #else + const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000)); + return Packet2cf(_mm_sub_ps(_mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), mask), + _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3), + vec4f_swizzle1(b.v, 1, 0, 3, 2)))); + #endif + } +}; + +template<> struct conj_helper<Packet4f, Packet2cf, false,false> +{ + EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet4f& x, const Packet2cf& y, const Packet2cf& c) const + { return padd(c, pmul(x,y)); } + + EIGEN_STRONG_INLINE Packet2cf pmul(const Packet4f& x, const Packet2cf& y) const + { return Packet2cf(Eigen::internal::pmul<Packet4f>(x, y.v)); } +}; + +template<> struct conj_helper<Packet2cf, Packet4f, false,false> +{ + EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet4f& y, const Packet2cf& c) const + { return padd(c, pmul(x,y)); } + + EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& x, const Packet4f& y) const + { return Packet2cf(Eigen::internal::pmul<Packet4f>(x.v, y)); } +}; + +template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b) +{ + // TODO optimize it for SSE3 and 4 + Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a,b); + __m128 s = _mm_mul_ps(b.v,b.v); + return Packet2cf(_mm_div_ps(res.v,_mm_add_ps(s,_mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(s), 0xb1))))); +} + +EIGEN_STRONG_INLINE Packet2cf pcplxflip/* <Packet2cf> */(const Packet2cf& x) +{ + return Packet2cf(vec4f_swizzle1(x.v, 1, 0, 3, 2)); +} + + +//---------- double ---------- +struct Packet1cd +{ + EIGEN_STRONG_INLINE Packet1cd() {} + EIGEN_STRONG_INLINE explicit Packet1cd(const __m128d& a) : v(a) {} + __m128d v; +}; + +// Use the packet_traits defined in AVX/PacketMath.h instead if we're going +// to leverage AVX instructions. +#ifndef EIGEN_VECTORIZE_AVX +template<> struct packet_traits<std::complex<double> > : default_packet_traits +{ + typedef Packet1cd type; + typedef Packet1cd half; + enum { + Vectorizable = 1, + AlignedOnScalar = 0, + size = 1, + HasHalfPacket = 0, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasNegate = 1, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasSetLinear = 0 + }; +}; +#endif + +template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16}; 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)); } +template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(Packet2d(a.v))); } +template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) +{ + const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0)); + return Packet1cd(_mm_xor_pd(a.v,mask)); +} + +template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b) +{ + #ifdef EIGEN_VECTORIZE_SSE3 + return Packet1cd(_mm_addsub_pd(_mm_mul_pd(_mm_movedup_pd(a.v), b.v), + _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1), + vec2d_swizzle1(b.v, 1, 0)))); + #else + const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x0,0x0,0x80000000,0x0)); + return Packet1cd(_mm_add_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), + _mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 1, 1), + vec2d_swizzle1(b.v, 1, 0)), mask))); + #endif +} + +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)); } + +// FIXME force unaligned load, this is a temporary fix +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)); } +template<> EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<double>& from) +{ /* here we really have to use unaligned loads :( */ return ploadu<Packet1cd>(&from); } + +template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>* from) { return pset1<Packet1cd>(*from); } + +// FIXME force unaligned store, this is a temporary fix +template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> * to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, Packet2d(from.v)); } +template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> * to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, Packet2d(from.v)); } + +template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> * addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } + +template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Packet1cd& a) +{ + EIGEN_ALIGN16 double res[2]; + _mm_store_pd(res, a.v); + return std::complex<double>(res[0],res[1]); +} + +template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; } + +template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a) +{ + return pfirst(a); +} + +template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs) +{ + return vecs[0]; +} + +template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a) +{ + return pfirst(a); +} + +template<int Offset> +struct palign_impl<Offset,Packet1cd> +{ + static EIGEN_STRONG_INLINE void run(Packet1cd& /*first*/, const Packet1cd& /*second*/) + { + // FIXME is it sure we never have to align a Packet1cd? + // Even though a std::complex<double> has 16 bytes, it is not necessarily aligned on a 16 bytes boundary... + } +}; + +template<> struct conj_helper<Packet1cd, Packet1cd, false,true> +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const + { + #ifdef EIGEN_VECTORIZE_SSE3 + return internal::pmul(a, pconj(b)); + #else + const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0)); + return Packet1cd(_mm_add_pd(_mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), mask), + _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1), + vec2d_swizzle1(b.v, 1, 0)))); + #endif + } +}; + +template<> struct conj_helper<Packet1cd, Packet1cd, true,false> +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const + { + #ifdef EIGEN_VECTORIZE_SSE3 + return internal::pmul(pconj(a), b); + #else + const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0)); + return Packet1cd(_mm_add_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), + _mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 1, 1), + vec2d_swizzle1(b.v, 1, 0)), mask))); + #endif + } +}; + +template<> struct conj_helper<Packet1cd, Packet1cd, true,true> +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const + { + #ifdef EIGEN_VECTORIZE_SSE3 + return pconj(internal::pmul(a, b)); + #else + const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0)); + return Packet1cd(_mm_sub_pd(_mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), mask), + _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1), + vec2d_swizzle1(b.v, 1, 0)))); + #endif + } +}; + +template<> struct conj_helper<Packet2d, Packet1cd, false,false> +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet2d& x, const Packet1cd& y, const Packet1cd& c) const + { return padd(c, pmul(x,y)); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet2d& x, const Packet1cd& y) const + { return Packet1cd(Eigen::internal::pmul<Packet2d>(x, y.v)); } +}; + +template<> struct conj_helper<Packet1cd, Packet2d, false,false> +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet2d& y, const Packet1cd& c) const + { return padd(c, pmul(x,y)); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& x, const Packet2d& y) const + { return Packet1cd(Eigen::internal::pmul<Packet2d>(x.v, y)); } +}; + +template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b) +{ + // TODO optimize it for SSE3 and 4 + Packet1cd res = conj_helper<Packet1cd,Packet1cd,false,true>().pmul(a,b); + __m128d s = _mm_mul_pd(b.v,b.v); + return Packet1cd(_mm_div_pd(res.v, _mm_add_pd(s,_mm_shuffle_pd(s, s, 0x1)))); +} + +EIGEN_STRONG_INLINE Packet1cd pcplxflip/* <Packet1cd> */(const Packet1cd& x) +{ + return Packet1cd(preverse(Packet2d(x.v))); +} + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet2cf,2>& kernel) { + __m128d w1 = _mm_castps_pd(kernel.packet[0].v); + __m128d w2 = _mm_castps_pd(kernel.packet[1].v); + + __m128 tmp = _mm_castpd_ps(_mm_unpackhi_pd(w1, w2)); + kernel.packet[0].v = _mm_castpd_ps(_mm_unpacklo_pd(w1, w2)); + kernel.packet[1].v = tmp; +} + +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)); +} + +template<> EIGEN_STRONG_INLINE Packet2cf pinsertfirst(const Packet2cf& a, std::complex<float> b) +{ + return Packet2cf(_mm_loadl_pi(a.v, reinterpret_cast<const __m64*>(&b))); +} + +template<> EIGEN_STRONG_INLINE Packet1cd pinsertfirst(const Packet1cd&, std::complex<double> b) +{ + return pset1<Packet1cd>(b); +} + +template<> EIGEN_STRONG_INLINE Packet2cf pinsertlast(const Packet2cf& a, std::complex<float> b) +{ + return Packet2cf(_mm_loadh_pi(a.v, reinterpret_cast<const __m64*>(&b))); +} + +template<> EIGEN_STRONG_INLINE Packet1cd pinsertlast(const Packet1cd&, std::complex<double> b) +{ + return pset1<Packet1cd>(b); +} + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_COMPLEX_SSE_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/SSE/MathFunctions.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/SSE/MathFunctions.h new file mode 100644 index 000000000..7b5f948e1 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/SSE/MathFunctions.h @@ -0,0 +1,562 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2007 Julien Pommier +// Copyright (C) 2009 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 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_SSE_H +#define EIGEN_MATH_FUNCTIONS_SSE_H + +namespace Eigen { + +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)); +} + +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); +} +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +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); +} + +/* evaluation of 4 sines at onces, 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); +} + +/* 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); +} + +#if EIGEN_FAST_MATH + +// Functions for sqrt. +// The EIGEN_FAST_MATH version uses the _mm_rsqrt_ps approximation and one step +// of Newton's method, at a cost of 1-2 bits of precision as opposed to the +// exact solution. It does not handle +inf, or denormalized numbers correctly. +// The main advantage of this approach is not just speed, but also the fact that +// it can be inlined and pipelined with other computations, further reducing its +// effective latency. This is similar to Quake3's fast inverse square root. +// For detail see here: http://www.beyond3d.com/content/articles/8/ +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet4f psqrt<Packet4f>(const Packet4f& _x) +{ + Packet4f half = pmul(_x, pset1<Packet4f>(.5f)); + Packet4f denormal_mask = _mm_and_ps( + _mm_cmpge_ps(_x, _mm_setzero_ps()), + _mm_cmplt_ps(_x, pset1<Packet4f>((std::numeric_limits<float>::min)()))); + + // Compute approximate reciprocal sqrt. + Packet4f x = _mm_rsqrt_ps(_x); + // Do a single step of Newton's iteration. + x = pmul(x, psub(pset1<Packet4f>(1.5f), pmul(half, pmul(x,x)))); + // Flush results for denormals to zero. + return _mm_andnot_ps(denormal_mask, pmul(_x,x)); +} + +#else + +template<>EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet4f psqrt<Packet4f>(const Packet4f& x) { return _mm_sqrt_ps(x); } + +#endif + +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet2d psqrt<Packet2d>(const Packet2d& x) { return _mm_sqrt_pd(x); } + +#if EIGEN_FAST_MATH + +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(one_point_five, 1.5f); + _EIGEN_DECLARE_CONST_Packet4f(minus_half, -0.5f); + _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(flt_min, 0x00800000); + + Packet4f neg_half = pmul(_x, p4f_minus_half); + + // select only the inverse sqrt of positive normal inputs (denormals are + // flushed to zero and cause infs as well). + Packet4f le_zero_mask = _mm_cmple_ps(_x, p4f_flt_min); + Packet4f x = _mm_andnot_ps(le_zero_mask, _mm_rsqrt_ps(_x)); + + // Fill in NaNs and Infs for the negative/zero entries. + Packet4f neg_mask = _mm_cmplt_ps(_x, _mm_setzero_ps()); + Packet4f zero_mask = _mm_andnot_ps(neg_mask, le_zero_mask); + Packet4f infs_and_nans = _mm_or_ps(_mm_and_ps(neg_mask, p4f_nan), + _mm_and_ps(zero_mask, p4f_inf)); + + // Do a single step of Newton's iteration. + x = pmul(x, pmadd(neg_half, pmul(x, x), p4f_one_point_five)); + + // Insert NaNs and Infs in all the right places. + return _mm_or_ps(x, infs_and_nans); +} + +#else + +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet4f prsqrt<Packet4f>(const Packet4f& x) { + // Unfortunately we can't use the much faster mm_rqsrt_ps since it only provides an approximation. + return _mm_div_ps(pset1<Packet4f>(1.0f), _mm_sqrt_ps(x)); +} + +#endif + +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet2d prsqrt<Packet2d>(const Packet2d& x) { + // Unfortunately we can't use the much faster mm_rqsrt_pd since it only provides an approximation. + return _mm_div_pd(pset1<Packet2d>(1.0), _mm_sqrt_pd(x)); +} + +// Hyperbolic Tangent function. +template <> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f +ptanh<Packet4f>(const Packet4f& x) { + return internal::generic_fast_tanh_float(x); +} + +} // end namespace internal + +namespace numext { + +template<> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +float sqrt(const float &x) +{ + return internal::pfirst(internal::Packet4f(_mm_sqrt_ss(_mm_set_ss(x)))); +} + +template<> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +double sqrt(const double &x) +{ +#if EIGEN_COMP_GNUC_STRICT + // This works around a GCC bug generating poor code for _mm_sqrt_pd + // See https://bitbucket.org/eigen/eigen/commits/14f468dba4d350d7c19c9b93072e19f7b3df563b + return internal::pfirst(internal::Packet2d(__builtin_ia32_sqrtsd(_mm_set_sd(x)))); +#else + return internal::pfirst(internal::Packet2d(_mm_sqrt_pd(_mm_set_sd(x)))); +#endif +} + +} // end namespace numex + +} // end namespace Eigen + +#endif // EIGEN_MATH_FUNCTIONS_SSE_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/SSE/PacketMath.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/SSE/PacketMath.h new file mode 100755 index 000000000..3832de147 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/SSE/PacketMath.h @@ -0,0 +1,879 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2009 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_PACKET_MATH_SSE_H +#define EIGEN_PACKET_MATH_SSE_H + +namespace Eigen { + +namespace internal { + +#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD +#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8 +#endif + +#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS +#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*)) +#endif + +#ifdef __FMA__ +#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD +#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD 1 +#endif +#endif + +#if (defined EIGEN_VECTORIZE_AVX) && (EIGEN_COMP_GNUC_STRICT || EIGEN_COMP_MINGW) && (__GXX_ABI_VERSION < 1004) +// With GCC's default ABI version, a __m128 or __m256 are the same types and therefore we cannot +// have overloads for both types without linking error. +// One solution is to increase ABI version using -fabi-version=4 (or greater). +// Otherwise, we workaround this inconvenience by wrapping 128bit types into the following helper +// structure: +template<typename T> +struct eigen_packet_wrapper +{ + EIGEN_ALWAYS_INLINE operator T&() { return m_val; } + EIGEN_ALWAYS_INLINE operator const T&() const { return m_val; } + EIGEN_ALWAYS_INLINE eigen_packet_wrapper() {} + EIGEN_ALWAYS_INLINE eigen_packet_wrapper(const T &v) : m_val(v) {} + EIGEN_ALWAYS_INLINE eigen_packet_wrapper& operator=(const T &v) { + m_val = v; + return *this; + } + + T m_val; +}; +typedef eigen_packet_wrapper<__m128> Packet4f; +typedef eigen_packet_wrapper<__m128i> Packet4i; +typedef eigen_packet_wrapper<__m128d> Packet2d; +#else +typedef __m128 Packet4f; +typedef __m128i Packet4i; +typedef __m128d Packet2d; +#endif + +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 vec4f_swizzle1(v,p,q,r,s) \ + (_mm_castsi128_ps(_mm_shuffle_epi32( _mm_castps_si128(v), ((s)<<6|(r)<<4|(q)<<2|(p))))) + +#define vec4i_swizzle1(v,p,q,r,s) \ + (_mm_shuffle_epi32( v, ((s)<<6|(r)<<4|(q)<<2|(p)))) + +#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))))) + +#define vec4f_swizzle2(a,b,p,q,r,s) \ + (_mm_shuffle_ps( (a), (b), ((s)<<6|(r)<<4|(q)<<2|(p)))) + +#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)))))) + +#define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \ + const Packet4f p4f_##NAME = pset1<Packet4f>(X) + +#define _EIGEN_DECLARE_CONST_Packet2d(NAME,X) \ + 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)) + +#define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \ + const Packet4i p4i_##NAME = pset1<Packet4i>(X) + + +// Use the packet_traits defined in AVX/PacketMath.h instead if we're going +// to leverage AVX instructions. +#ifndef EIGEN_VECTORIZE_AVX +template<> struct packet_traits<float> : default_packet_traits +{ + typedef Packet4f type; + typedef Packet4f half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=4, + HasHalfPacket = 0, + + HasDiv = 1, + HasSin = EIGEN_FAST_MATH, + HasCos = EIGEN_FAST_MATH, + HasLog = 1, + HasExp = 1, + HasSqrt = 1, + HasRsqrt = 1, + HasTanh = EIGEN_FAST_MATH, + HasBlend = 1 + +#ifdef EIGEN_VECTORIZE_SSE4_1 + , + HasRound = 1, + HasFloor = 1, + HasCeil = 1 +#endif + }; +}; +template<> struct packet_traits<double> : default_packet_traits +{ + typedef Packet2d type; + typedef Packet2d half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=2, + HasHalfPacket = 0, + + HasDiv = 1, + HasExp = 1, + HasSqrt = 1, + HasRsqrt = 1, + HasBlend = 1 + +#ifdef EIGEN_VECTORIZE_SSE4_1 + , + HasRound = 1, + HasFloor = 1, + HasCeil = 1 +#endif + }; +}; +#endif +template<> struct packet_traits<int> : default_packet_traits +{ + typedef Packet4i type; + typedef Packet4i half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=4, + + HasBlend = 1 + }; +}; + +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; }; + +#ifndef EIGEN_VECTORIZE_AVX +template<> struct scalar_div_cost<float,true> { enum { value = 7 }; }; +template<> struct scalar_div_cost<double,true> { enum { value = 8 }; }; +#endif + +#if EIGEN_COMP_MSVC==1500 +// Workaround MSVC 9 internal compiler error. +// TODO: It has been detected with win64 builds (amd64), so let's check whether it also happens in 32bits+SSE mode +// TODO: let's check whether there does not exist a better fix, like adding a pset0() function. (it crashed on pset1(0)). +template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) { return _mm_set_ps(from,from,from,from); } +template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { return _mm_set_pd(from,from); } +template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) { return _mm_set_epi32(from,from,from,from); } +#else +template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) { return _mm_set_ps1(from); } +template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { return _mm_set1_pd(from); } +template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) { return _mm_set1_epi32(from); } +#endif + +// 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. +// Therefore, we introduced the pload1 functions to be used in product kernels for which bug 203 does not apply. +// Also note that with AVX, we want it to generate a vbroadcastss. +#if EIGEN_COMP_GNUC_STRICT && (!defined __AVX__) +template<> EIGEN_STRONG_INLINE Packet4f pload1<Packet4f>(const float *from) { + return vec4f_swizzle1(_mm_load_ss(from),0,0,0,0); +} +#endif + +template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a) { return _mm_add_ps(pset1<Packet4f>(a), _mm_set_ps(3,2,1,0)); } +template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a) { return _mm_add_pd(pset1<Packet2d>(a),_mm_set_pd(1,0)); } +template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int& a) { return _mm_add_epi32(pset1<Packet4i>(a),_mm_set_epi32(3,2,1,0)); } + +template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_add_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_add_pd(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_add_epi32(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_sub_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_sub_pd(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_sub_epi32(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a) +{ + const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000,0x80000000,0x80000000,0x80000000)); + return _mm_xor_ps(a,mask); +} +template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) +{ + const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0x0,0x80000000,0x0,0x80000000)); + return _mm_xor_pd(a,mask); +} +template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) +{ + return psub(Packet4i(_mm_setr_epi32(0,0,0,0)), a); +} + +template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; } +template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; } +template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; } + +template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_mul_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_mul_pd(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) +{ +#ifdef EIGEN_VECTORIZE_SSE4_1 + return _mm_mullo_epi32(a,b); +#else + // this version is slightly faster than 4 scalar products + return vec4i_swizzle1( + vec4i_swizzle2( + _mm_mul_epu32(a,b), + _mm_mul_epu32(vec4i_swizzle1(a,1,0,3,2), + vec4i_swizzle1(b,1,0,3,2)), + 0,2,0,2), + 0,2,1,3); +#endif +} + +template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_div_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_div_pd(a,b); } + +// 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__ +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) { return _mm_min_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_min_pd(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) +{ +#ifdef EIGEN_VECTORIZE_SSE4_1 + return _mm_min_epi32(a,b); +#else + // after some bench, this version *is* faster than a scalar implementation + Packet4i mask = _mm_cmplt_epi32(a,b); + return _mm_or_si128(_mm_and_si128(mask,a),_mm_andnot_si128(mask,b)); +#endif +} + +template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_max_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_max_pd(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) +{ +#ifdef EIGEN_VECTORIZE_SSE4_1 + return _mm_max_epi32(a,b); +#else + // after some bench, this version *is* faster than a scalar implementation + Packet4i mask = _mm_cmpgt_epi32(a,b); + return _mm_or_si128(_mm_and_si128(mask,a),_mm_andnot_si128(mask,b)); +#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 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); } +template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_and_si128(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_or_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_or_pd(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_or_si128(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_xor_ps(a,b); } +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 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); } +template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_si128(reinterpret_cast<const __m128i*>(from)); } + +#if EIGEN_COMP_MSVC + template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) { + EIGEN_DEBUG_UNALIGNED_LOAD + #if (EIGEN_COMP_MSVC==1600) + // NOTE Some version of MSVC10 generates bad code when using _mm_loadu_ps + // (i.e., it does not generate an unaligned load!! + __m128 res = _mm_loadl_pi(_mm_set1_ps(0.0f), (const __m64*)(from)); + res = _mm_loadh_pi(res, (const __m64*)(from+2)); + return res; + #else + return _mm_loadu_ps(from); + #endif + } +#else +// NOTE: with the code below, MSVC's compiler crashes! + +template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) +{ + EIGEN_DEBUG_UNALIGNED_LOAD + return _mm_loadu_ps(from); +} +#endif + +template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from) +{ + EIGEN_DEBUG_UNALIGNED_LOAD + return _mm_loadu_pd(from); +} +template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from) +{ + EIGEN_DEBUG_UNALIGNED_LOAD + return _mm_loadu_si128(reinterpret_cast<const __m128i*>(from)); +} + + +template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from) +{ + return vec4f_swizzle1(_mm_castpd_ps(_mm_load_sd(reinterpret_cast<const double*>(from))), 0, 0, 1, 1); +} +template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double* from) +{ return pset1<Packet2d>(from[0]); } +template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int* from) +{ + Packet4i tmp; + tmp = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(from)); + return vec4i_swizzle1(tmp, 0, 0, 1, 1); +} + +template<> EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_ps(to, from); } +template<> EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet2d& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_pd(to, from); } +template<> EIGEN_STRONG_INLINE void pstore<int>(int* to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_si128(reinterpret_cast<__m128i*>(to), from); } + +template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_pd(to, from); } +template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_ps(to, from); } +template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_si128(reinterpret_cast<__m128i*>(to), from); } + +template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride) +{ + return _mm_set_ps(from[3*stride], from[2*stride], from[1*stride], from[0*stride]); +} +template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride) +{ + return _mm_set_pd(from[1*stride], from[0*stride]); +} +template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* from, Index stride) +{ + return _mm_set_epi32(from[3*stride], from[2*stride], from[1*stride], from[0*stride]); + } + +template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride) +{ + to[stride*0] = _mm_cvtss_f32(from); + to[stride*1] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 1)); + to[stride*2] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 2)); + to[stride*3] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 3)); +} +template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride) +{ + to[stride*0] = _mm_cvtsd_f64(from); + to[stride*1] = _mm_cvtsd_f64(_mm_shuffle_pd(from, from, 1)); +} +template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const Packet4i& from, Index stride) +{ + to[stride*0] = _mm_cvtsi128_si32(from); + to[stride*1] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 1)); + to[stride*2] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 2)); + to[stride*3] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 3)); +} + +// some compilers might be tempted to perform multiple moves instead of using a vector path. +template<> EIGEN_STRONG_INLINE void pstore1<Packet4f>(float* to, const float& a) +{ + Packet4f pa = _mm_set_ss(a); + pstore(to, Packet4f(vec4f_swizzle1(pa,0,0,0,0))); +} +// some compilers might be tempted to perform multiple moves instead of using a vector path. +template<> EIGEN_STRONG_INLINE void pstore1<Packet2d>(double* to, const double& a) +{ + Packet2d pa = _mm_set_sd(a); + pstore(to, Packet2d(vec2d_swizzle1(pa,0,0))); +} + +#ifndef EIGEN_VECTORIZE_AVX +template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } +template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } +template<> EIGEN_STRONG_INLINE void prefetch<int>(const int* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } +#endif + +#if EIGEN_COMP_MSVC_STRICT && EIGEN_OS_WIN64 +// The temporary variable fixes an internal compilation error in vs <= 2008 and a wrong-result bug in vs 2010 +// Direct of the struct members fixed bug #62. +template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { return a.m128_f32[0]; } +template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { return a.m128d_f64[0]; } +template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { int x = _mm_cvtsi128_si32(a); return x; } +#elif EIGEN_COMP_MSVC_STRICT +// The temporary variable fixes an internal compilation error in vs <= 2008 and a wrong-result bug in vs 2010 +template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { float x = _mm_cvtss_f32(a); return x; } +template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { double x = _mm_cvtsd_f64(a); return x; } +template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { int x = _mm_cvtsi128_si32(a); return x; } +#else +template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { return _mm_cvtss_f32(a); } +template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { return _mm_cvtsd_f64(a); } +template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { return _mm_cvtsi128_si32(a); } +#endif + +template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) +{ return _mm_shuffle_ps(a,a,0x1B); } +template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a) +{ return _mm_shuffle_pd(a,a,0x1); } +template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) +{ return _mm_shuffle_epi32(a,0x1B); } + +template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a) +{ + const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF)); + return _mm_and_ps(a,mask); +} +template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a) +{ + const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF)); + return _mm_and_pd(a,mask); +} +template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) +{ + #ifdef EIGEN_VECTORIZE_SSSE3 + return _mm_abs_epi32(a); + #else + Packet4i aux = _mm_srai_epi32(a,31); + return _mm_sub_epi32(_mm_xor_si128(a,aux),aux); + #endif +} + +// with AVX, the default implementations based on pload1 are faster +#ifndef __AVX__ +template<> EIGEN_STRONG_INLINE void +pbroadcast4<Packet4f>(const float *a, + Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3) +{ + a3 = pload<Packet4f>(a); + a0 = vec4f_swizzle1(a3, 0,0,0,0); + a1 = vec4f_swizzle1(a3, 1,1,1,1); + a2 = vec4f_swizzle1(a3, 2,2,2,2); + a3 = vec4f_swizzle1(a3, 3,3,3,3); +} +template<> EIGEN_STRONG_INLINE void +pbroadcast4<Packet2d>(const double *a, + Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3) +{ +#ifdef EIGEN_VECTORIZE_SSE3 + a0 = _mm_loaddup_pd(a+0); + a1 = _mm_loaddup_pd(a+1); + a2 = _mm_loaddup_pd(a+2); + a3 = _mm_loaddup_pd(a+3); +#else + a1 = pload<Packet2d>(a); + a0 = vec2d_swizzle1(a1, 0,0); + a1 = vec2d_swizzle1(a1, 1,1); + a3 = pload<Packet2d>(a+2); + a2 = vec2d_swizzle1(a3, 0,0); + a3 = vec2d_swizzle1(a3, 1,1); +#endif +} +#endif + +EIGEN_STRONG_INLINE void punpackp(Packet4f* vecs) +{ + vecs[1] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0x55)); + vecs[2] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0xAA)); + vecs[3] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0xFF)); + vecs[0] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0x00)); +} + +#ifdef EIGEN_VECTORIZE_SSE3 +template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs) +{ + return _mm_hadd_ps(_mm_hadd_ps(vecs[0], vecs[1]),_mm_hadd_ps(vecs[2], vecs[3])); +} + +template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs) +{ + return _mm_hadd_pd(vecs[0], vecs[1]); +} + +#else +template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs) +{ + Packet4f tmp0, tmp1, tmp2; + tmp0 = _mm_unpacklo_ps(vecs[0], vecs[1]); + tmp1 = _mm_unpackhi_ps(vecs[0], vecs[1]); + tmp2 = _mm_unpackhi_ps(vecs[2], vecs[3]); + tmp0 = _mm_add_ps(tmp0, tmp1); + tmp1 = _mm_unpacklo_ps(vecs[2], vecs[3]); + tmp1 = _mm_add_ps(tmp1, tmp2); + tmp2 = _mm_movehl_ps(tmp1, tmp0); + tmp0 = _mm_movelh_ps(tmp0, tmp1); + return _mm_add_ps(tmp0, tmp2); +} + +template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs) +{ + return _mm_add_pd(_mm_unpacklo_pd(vecs[0], vecs[1]), _mm_unpackhi_pd(vecs[0], vecs[1])); +} +#endif // SSE3 + +template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a) +{ + // Disable SSE3 _mm_hadd_pd that is extremely slow on all existing Intel's architectures + // (from Nehalem to Haswell) +// #ifdef EIGEN_VECTORIZE_SSE3 +// Packet4f tmp = _mm_add_ps(a, vec4f_swizzle1(a,2,3,2,3)); +// return pfirst<Packet4f>(_mm_hadd_ps(tmp, tmp)); +// #else + Packet4f tmp = _mm_add_ps(a, _mm_movehl_ps(a,a)); + return pfirst<Packet4f>(_mm_add_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); +// #endif +} + +template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) +{ + // Disable SSE3 _mm_hadd_pd that is extremely slow on all existing Intel's architectures + // (from Nehalem to Haswell) +// #ifdef EIGEN_VECTORIZE_SSE3 +// return pfirst<Packet2d>(_mm_hadd_pd(a, a)); +// #else + return pfirst<Packet2d>(_mm_add_sd(a, _mm_unpackhi_pd(a,a))); +// #endif +} + +#ifdef EIGEN_VECTORIZE_SSSE3 +template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs) +{ + return _mm_hadd_epi32(_mm_hadd_epi32(vecs[0], vecs[1]),_mm_hadd_epi32(vecs[2], vecs[3])); +} +template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a) +{ + Packet4i tmp0 = _mm_hadd_epi32(a,a); + return pfirst<Packet4i>(_mm_hadd_epi32(tmp0,tmp0)); +} +#else +template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a) +{ + Packet4i tmp = _mm_add_epi32(a, _mm_unpackhi_epi64(a,a)); + return pfirst(tmp) + pfirst<Packet4i>(_mm_shuffle_epi32(tmp, 1)); +} + +template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs) +{ + Packet4i tmp0, tmp1, tmp2; + tmp0 = _mm_unpacklo_epi32(vecs[0], vecs[1]); + tmp1 = _mm_unpackhi_epi32(vecs[0], vecs[1]); + tmp2 = _mm_unpackhi_epi32(vecs[2], vecs[3]); + tmp0 = _mm_add_epi32(tmp0, tmp1); + tmp1 = _mm_unpacklo_epi32(vecs[2], vecs[3]); + tmp1 = _mm_add_epi32(tmp1, tmp2); + tmp2 = _mm_unpacklo_epi64(tmp0, tmp1); + tmp0 = _mm_unpackhi_epi64(tmp0, tmp1); + return _mm_add_epi32(tmp0, tmp2); +} +#endif +// Other reduction functions: + +// mul +template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a) +{ + Packet4f tmp = _mm_mul_ps(a, _mm_movehl_ps(a,a)); + return pfirst<Packet4f>(_mm_mul_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); +} +template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a) +{ + return pfirst<Packet2d>(_mm_mul_sd(a, _mm_unpackhi_pd(a,a))); +} +template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a) +{ + // after some experiments, it is seems this is the fastest way to implement it + // for GCC (eg., reusing pmul is very slow !) + // TODO try to call _mm_mul_epu32 directly + EIGEN_ALIGN16 int aux[4]; + pstore(aux, a); + return (aux[0] * aux[1]) * (aux[2] * aux[3]);; +} + +// min +template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a) +{ + Packet4f tmp = _mm_min_ps(a, _mm_movehl_ps(a,a)); + return pfirst<Packet4f>(_mm_min_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); +} +template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a) +{ + return pfirst<Packet2d>(_mm_min_sd(a, _mm_unpackhi_pd(a,a))); +} +template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a) +{ +#ifdef EIGEN_VECTORIZE_SSE4_1 + Packet4i tmp = _mm_min_epi32(a, _mm_shuffle_epi32(a, _MM_SHUFFLE(0,0,3,2))); + return pfirst<Packet4i>(_mm_min_epi32(tmp,_mm_shuffle_epi32(tmp, 1))); +#else + // after some experiments, it is seems this is the fastest way to implement it + // for GCC (eg., it does not like using std::min after the pstore !!) + EIGEN_ALIGN16 int aux[4]; + pstore(aux, a); + int aux0 = aux[0]<aux[1] ? aux[0] : aux[1]; + int aux2 = aux[2]<aux[3] ? aux[2] : aux[3]; + return aux0<aux2 ? aux0 : aux2; +#endif // EIGEN_VECTORIZE_SSE4_1 +} + +// max +template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a) +{ + Packet4f tmp = _mm_max_ps(a, _mm_movehl_ps(a,a)); + return pfirst<Packet4f>(_mm_max_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); +} +template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a) +{ + return pfirst<Packet2d>(_mm_max_sd(a, _mm_unpackhi_pd(a,a))); +} +template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a) +{ +#ifdef EIGEN_VECTORIZE_SSE4_1 + Packet4i tmp = _mm_max_epi32(a, _mm_shuffle_epi32(a, _MM_SHUFFLE(0,0,3,2))); + return pfirst<Packet4i>(_mm_max_epi32(tmp,_mm_shuffle_epi32(tmp, 1))); +#else + // after some experiments, it is seems this is the fastest way to implement it + // for GCC (eg., it does not like using std::min after the pstore !!) + EIGEN_ALIGN16 int aux[4]; + pstore(aux, a); + int aux0 = aux[0]>aux[1] ? aux[0] : aux[1]; + int aux2 = aux[2]>aux[3] ? aux[2] : aux[3]; + return aux0>aux2 ? aux0 : aux2; +#endif // EIGEN_VECTORIZE_SSE4_1 +} + +#if EIGEN_COMP_GNUC +// template <> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) +// { +// Packet4f res = b; +// asm("mulps %[a], %[b] \n\taddps %[c], %[b]" : [b] "+x" (res) : [a] "x" (a), [c] "x" (c)); +// return res; +// } +// EIGEN_STRONG_INLINE Packet4i _mm_alignr_epi8(const Packet4i& a, const Packet4i& b, const int i) +// { +// Packet4i res = a; +// asm("palignr %[i], %[a], %[b] " : [b] "+x" (res) : [a] "x" (a), [i] "i" (i)); +// return res; +// } +#endif + +#ifdef EIGEN_VECTORIZE_SSSE3 +// SSSE3 versions +template<int Offset> +struct palign_impl<Offset,Packet4f> +{ + static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second) + { + if (Offset!=0) + first = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(second), _mm_castps_si128(first), Offset*4)); + } +}; + +template<int Offset> +struct palign_impl<Offset,Packet4i> +{ + static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second) + { + if (Offset!=0) + first = _mm_alignr_epi8(second,first, Offset*4); + } +}; + +template<int Offset> +struct palign_impl<Offset,Packet2d> +{ + static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second) + { + if (Offset==1) + first = _mm_castsi128_pd(_mm_alignr_epi8(_mm_castpd_si128(second), _mm_castpd_si128(first), 8)); + } +}; +#else +// SSE2 versions +template<int Offset> +struct palign_impl<Offset,Packet4f> +{ + static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second) + { + if (Offset==1) + { + first = _mm_move_ss(first,second); + first = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(first),0x39)); + } + else if (Offset==2) + { + first = _mm_movehl_ps(first,first); + first = _mm_movelh_ps(first,second); + } + else if (Offset==3) + { + first = _mm_move_ss(first,second); + first = _mm_shuffle_ps(first,second,0x93); + } + } +}; + +template<int Offset> +struct palign_impl<Offset,Packet4i> +{ + static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second) + { + if (Offset==1) + { + first = _mm_castps_si128(_mm_move_ss(_mm_castsi128_ps(first),_mm_castsi128_ps(second))); + first = _mm_shuffle_epi32(first,0x39); + } + else if (Offset==2) + { + first = _mm_castps_si128(_mm_movehl_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(first))); + first = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(second))); + } + else if (Offset==3) + { + first = _mm_castps_si128(_mm_move_ss(_mm_castsi128_ps(first),_mm_castsi128_ps(second))); + first = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(second),0x93)); + } + } +}; + +template<int Offset> +struct palign_impl<Offset,Packet2d> +{ + static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second) + { + if (Offset==1) + { + first = _mm_castps_pd(_mm_movehl_ps(_mm_castpd_ps(first),_mm_castpd_ps(first))); + first = _mm_castps_pd(_mm_movelh_ps(_mm_castpd_ps(first),_mm_castpd_ps(second))); + } + } +}; +#endif + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet4f,4>& kernel) { + _MM_TRANSPOSE4_PS(kernel.packet[0], kernel.packet[1], kernel.packet[2], kernel.packet[3]); +} + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet2d,2>& kernel) { + __m128d tmp = _mm_unpackhi_pd(kernel.packet[0], kernel.packet[1]); + kernel.packet[0] = _mm_unpacklo_pd(kernel.packet[0], kernel.packet[1]); + kernel.packet[1] = tmp; +} + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet4i,4>& kernel) { + __m128i T0 = _mm_unpacklo_epi32(kernel.packet[0], kernel.packet[1]); + __m128i T1 = _mm_unpacklo_epi32(kernel.packet[2], kernel.packet[3]); + __m128i T2 = _mm_unpackhi_epi32(kernel.packet[0], kernel.packet[1]); + __m128i T3 = _mm_unpackhi_epi32(kernel.packet[2], kernel.packet[3]); + + kernel.packet[0] = _mm_unpacklo_epi64(T0, T1); + kernel.packet[1] = _mm_unpackhi_epi64(T0, T1); + kernel.packet[2] = _mm_unpacklo_epi64(T2, T3); + kernel.packet[3] = _mm_unpackhi_epi64(T2, T3); +} + +template<> EIGEN_STRONG_INLINE Packet4i pblend(const Selector<4>& ifPacket, const Packet4i& thenPacket, const Packet4i& elsePacket) { + const __m128i zero = _mm_setzero_si128(); + const __m128i select = _mm_set_epi32(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]); + __m128i false_mask = _mm_cmpeq_epi32(select, zero); +#ifdef EIGEN_VECTORIZE_SSE4_1 + return _mm_blendv_epi8(thenPacket, elsePacket, false_mask); +#else + return _mm_or_si128(_mm_andnot_si128(false_mask, thenPacket), _mm_and_si128(false_mask, elsePacket)); +#endif +} +template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket, const Packet4f& elsePacket) { + const __m128 zero = _mm_setzero_ps(); + const __m128 select = _mm_set_ps(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]); + __m128 false_mask = _mm_cmpeq_ps(select, zero); +#ifdef EIGEN_VECTORIZE_SSE4_1 + return _mm_blendv_ps(thenPacket, elsePacket, false_mask); +#else + return _mm_or_ps(_mm_andnot_ps(false_mask, thenPacket), _mm_and_ps(false_mask, elsePacket)); +#endif +} +template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket, const Packet2d& elsePacket) { + const __m128d zero = _mm_setzero_pd(); + const __m128d select = _mm_set_pd(ifPacket.select[1], ifPacket.select[0]); + __m128d false_mask = _mm_cmpeq_pd(select, zero); +#ifdef EIGEN_VECTORIZE_SSE4_1 + return _mm_blendv_pd(thenPacket, elsePacket, false_mask); +#else + return _mm_or_pd(_mm_andnot_pd(false_mask, thenPacket), _mm_and_pd(false_mask, elsePacket)); +#endif +} + +template<> EIGEN_STRONG_INLINE Packet4f pinsertfirst(const Packet4f& a, float b) +{ +#ifdef EIGEN_VECTORIZE_SSE4_1 + return _mm_blend_ps(a,pset1<Packet4f>(b),1); +#else + return _mm_move_ss(a, _mm_load_ss(&b)); +#endif +} + +template<> EIGEN_STRONG_INLINE Packet2d pinsertfirst(const Packet2d& a, double b) +{ +#ifdef EIGEN_VECTORIZE_SSE4_1 + return _mm_blend_pd(a,pset1<Packet2d>(b),1); +#else + return _mm_move_sd(a, _mm_load_sd(&b)); +#endif +} + +template<> EIGEN_STRONG_INLINE Packet4f pinsertlast(const Packet4f& a, float b) +{ +#ifdef EIGEN_VECTORIZE_SSE4_1 + return _mm_blend_ps(a,pset1<Packet4f>(b),(1<<3)); +#else + const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x0,0x0,0x0,0xFFFFFFFF)); + return _mm_or_ps(_mm_andnot_ps(mask, a), _mm_and_ps(mask, pset1<Packet4f>(b))); +#endif +} + +template<> EIGEN_STRONG_INLINE Packet2d pinsertlast(const Packet2d& a, double b) +{ +#ifdef EIGEN_VECTORIZE_SSE4_1 + return _mm_blend_pd(a,pset1<Packet2d>(b),(1<<1)); +#else + const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0x0,0x0,0xFFFFFFFF,0xFFFFFFFF)); + return _mm_or_pd(_mm_andnot_pd(mask, a), _mm_and_pd(mask, pset1<Packet2d>(b))); +#endif +} + +// Scalar path for pmadd with FMA to ensure consistency with vectorized path. +#ifdef __FMA__ +template<> EIGEN_STRONG_INLINE float pmadd(const float& a, const float& b, const float& c) { + return ::fmaf(a,b,c); +} +template<> EIGEN_STRONG_INLINE double pmadd(const double& a, const double& b, const double& c) { + return ::fma(a,b,c); +} +#endif + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_PACKET_MATH_SSE_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/SSE/TypeCasting.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/SSE/TypeCasting.h new file mode 100644 index 000000000..c84893230 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/SSE/TypeCasting.h @@ -0,0 +1,77 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2015 Benoit Steiner <benoit.steiner.goog@gmail.com> +// +// 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_TYPE_CASTING_SSE_H +#define EIGEN_TYPE_CASTING_SSE_H + +namespace Eigen { + +namespace internal { + +template <> +struct type_casting_traits<float, int> { + enum { + VectorizedCast = 1, + SrcCoeffRatio = 1, + TgtCoeffRatio = 1 + }; +}; + +template<> EIGEN_STRONG_INLINE Packet4i pcast<Packet4f, Packet4i>(const Packet4f& a) { + return _mm_cvttps_epi32(a); +} + + +template <> +struct type_casting_traits<int, float> { + enum { + VectorizedCast = 1, + SrcCoeffRatio = 1, + TgtCoeffRatio = 1 + }; +}; + +template<> EIGEN_STRONG_INLINE Packet4f pcast<Packet4i, Packet4f>(const Packet4i& a) { + return _mm_cvtepi32_ps(a); +} + + +template <> +struct type_casting_traits<double, float> { + enum { + VectorizedCast = 1, + SrcCoeffRatio = 2, + TgtCoeffRatio = 1 + }; +}; + +template<> EIGEN_STRONG_INLINE Packet4f pcast<Packet2d, Packet4f>(const Packet2d& a, const Packet2d& b) { + return _mm_shuffle_ps(_mm_cvtpd_ps(a), _mm_cvtpd_ps(b), (1 << 2) | (1 << 6)); +} + +template <> +struct type_casting_traits<float, double> { + enum { + VectorizedCast = 1, + SrcCoeffRatio = 1, + TgtCoeffRatio = 2 + }; +}; + +template<> EIGEN_STRONG_INLINE Packet2d pcast<Packet4f, Packet2d>(const Packet4f& a) { + // Simply discard the second half of the input + return _mm_cvtps_pd(a); +} + + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_TYPE_CASTING_SSE_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/ZVector/Complex.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/ZVector/Complex.h new file mode 100644 index 000000000..d39d2d105 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/ZVector/Complex.h @@ -0,0 +1,394 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2016 Konstantinos Margaritis <markos@freevec.org> +// +// 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_COMPLEX32_ALTIVEC_H +#define EIGEN_COMPLEX32_ALTIVEC_H + +namespace Eigen { + +namespace internal { + +static Packet2ul p2ul_CONJ_XOR1 = (Packet2ul) vec_sld((Packet4ui) p2d_ZERO_, (Packet4ui) p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 }; +static Packet2ul p2ul_CONJ_XOR2 = (Packet2ul) vec_sld((Packet4ui) p2l_ZERO, (Packet4ui) p2d_ZERO_, 8);//{ 0x8000000000000000, 0x0000000000000000 }; + +struct Packet1cd +{ + EIGEN_STRONG_INLINE Packet1cd() {} + EIGEN_STRONG_INLINE explicit Packet1cd(const Packet2d& a) : v(a) {} + Packet2d v; +}; + +struct Packet2cf +{ + EIGEN_STRONG_INLINE Packet2cf() {} + EIGEN_STRONG_INLINE explicit Packet2cf(const Packet4f& a) : v(a) {} + union { + Packet4f v; + Packet1cd cd[2]; + }; +}; + +template<> struct packet_traits<std::complex<float> > : default_packet_traits +{ + typedef Packet2cf type; + typedef Packet2cf half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size = 2, + HasHalfPacket = 0, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasNegate = 1, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasBlend = 1, + HasSetLinear = 0 + }; +}; + + +template<> struct packet_traits<std::complex<double> > : default_packet_traits +{ + typedef Packet1cd type; + typedef Packet1cd half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size = 1, + HasHalfPacket = 0, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasNegate = 1, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasSetLinear = 0 + }; +}; + +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; }; + +/* Forward declaration */ +EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2cf,2>& kernel); + +template<> EIGEN_STRONG_INLINE Packet2cf pload <Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>((const float*)from)); } +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 Packet2cf ploadu<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>((const float*)from)); } +template<> EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu<Packet2d>((const double*)from)); } +template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((float*)to, from.v); } +template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> * to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); } +template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((float*)to, from.v); } +template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> * to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); } + +template<> EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<double>& from) +{ /* here we really have to use unaligned loads :( */ return ploadu<Packet1cd>(&from); } + +template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>& from) +{ + Packet2cf res; + res.cd[0] = Packet1cd(vec_ld2f((const float *)&from)); + res.cd[1] = res.cd[0]; + return res; +} +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]; + af[0] = from[0*stride]; + af[1] = from[1*stride]; + return pload<Packet2cf>(af); +} +template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather<std::complex<double>, Packet1cd>(const std::complex<double>* from, Index stride EIGEN_UNUSED) +{ + return pload<Packet1cd>(from); +} +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]; + pstore<std::complex<float> >((std::complex<float> *) af, from); + to[0*stride] = af[0]; + to[1*stride] = af[1]; +} +template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet1cd>(std::complex<double>* to, const Packet1cd& from, Index stride EIGEN_UNUSED) +{ + pstore<std::complex<double> >(to, from); +} + +template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(padd<Packet4f>(a.v, b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(a.v + b.v); } +template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(psub<Packet4f>(a.v, b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(a.v - b.v); } +template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(Packet2d(a.v))); } +template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) { return Packet2cf(pnegate(Packet4f(a.v))); } +template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) { return Packet1cd((Packet2d)vec_xor((Packet2d)a.v, (Packet2d)p2ul_CONJ_XOR2)); } +template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a) +{ + Packet2cf res; + res.v.v4f[0] = pconj(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[0]))).v; + res.v.v4f[1] = pconj(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[1]))).v; + return res; +} + +template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b) +{ + Packet2d a_re, a_im, v1, v2; + + // Permute and multiply the real parts of a and b + a_re = vec_perm(a.v, a.v, p16uc_PSET64_HI); + // Get the imaginary parts of a + a_im = vec_perm(a.v, a.v, p16uc_PSET64_LO); + // multiply a_re * b + v1 = vec_madd(a_re, b.v, p2d_ZERO); + // multiply a_im * b and get the conjugate result + v2 = vec_madd(a_im, b.v, p2d_ZERO); + v2 = (Packet2d) vec_sld((Packet4ui)v2, (Packet4ui)v2, 8); + v2 = (Packet2d) vec_xor((Packet2d)v2, (Packet2d) p2ul_CONJ_XOR1); + + return Packet1cd(v1 + v2); +} +template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b) +{ + Packet2cf res; + res.v.v4f[0] = pmul(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[0])), Packet1cd(reinterpret_cast<Packet2d>(b.v.v4f[0]))).v; + res.v.v4f[1] = pmul(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[1])), Packet1cd(reinterpret_cast<Packet2d>(b.v.v4f[1]))).v; + return res; +} + +template<> EIGEN_STRONG_INLINE Packet1cd pand <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cf pand <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pand<Packet4f>(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd por <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_or(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cf por <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(por<Packet4f>(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd pxor <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_xor(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cf pxor <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pxor<Packet4f>(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v, vec_nor(b.v,b.v))); } +template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pandnot<Packet4f>(a.v,b.v)); } + +template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>* from) { return pset1<Packet1cd>(*from); } +template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from) { return pset1<Packet2cf>(*from); } + +template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> * addr) { EIGEN_ZVECTOR_PREFETCH(addr); } +template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> * addr) { EIGEN_ZVECTOR_PREFETCH(addr); } + +template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Packet1cd& a) +{ + std::complex<double> EIGEN_ALIGN16 res; + pstore<std::complex<double> >(&res, a); + + return res; +} +template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a) +{ + std::complex<float> EIGEN_ALIGN16 res[2]; + pstore<std::complex<float> >(res, a); + + return res[0]; +} + +template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; } +template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) +{ + Packet2cf res; + res.cd[0] = a.cd[1]; + res.cd[1] = a.cd[0]; + return res; +} + +template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a) +{ + return pfirst(a); +} +template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a) +{ + std::complex<float> res; + Packet1cd b = padd<Packet1cd>(a.cd[0], a.cd[1]); + vec_st2f(b.v, (float*)&res); + return res; +} + +template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs) +{ + return vecs[0]; +} +template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs) +{ + PacketBlock<Packet2cf,2> transpose; + transpose.packet[0] = vecs[0]; + transpose.packet[1] = vecs[1]; + ptranspose(transpose); + + return padd<Packet2cf>(transpose.packet[0], transpose.packet[1]); +} + +template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a) +{ + return pfirst(a); +} +template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a) +{ + std::complex<float> res; + Packet1cd b = pmul<Packet1cd>(a.cd[0], a.cd[1]); + vec_st2f(b.v, (float*)&res); + return res; +} + +template<int Offset> +struct palign_impl<Offset,Packet1cd> +{ + static EIGEN_STRONG_INLINE void run(Packet1cd& /*first*/, const Packet1cd& /*second*/) + { + // FIXME is it sure we never have to align a Packet1cd? + // Even though a std::complex<double> has 16 bytes, it is not necessarily aligned on a 16 bytes boundary... + } +}; + +template<int Offset> +struct palign_impl<Offset,Packet2cf> +{ + static EIGEN_STRONG_INLINE void run(Packet2cf& first, const Packet2cf& second) + { + if (Offset == 1) { + first.cd[0] = first.cd[1]; + first.cd[1] = second.cd[0]; + } + } +}; + +template<> struct conj_helper<Packet1cd, Packet1cd, false,true> +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const + { + return internal::pmul(a, pconj(b)); + } +}; + +template<> struct conj_helper<Packet1cd, Packet1cd, true,false> +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const + { + return internal::pmul(pconj(a), b); + } +}; + +template<> struct conj_helper<Packet1cd, Packet1cd, true,true> +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const + { + return pconj(internal::pmul(a, b)); + } +}; + +template<> struct conj_helper<Packet2cf, Packet2cf, false,true> +{ + EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const + { + return internal::pmul(a, pconj(b)); + } +}; + +template<> struct conj_helper<Packet2cf, Packet2cf, true,false> +{ + EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const + { + return internal::pmul(pconj(a), b); + } +}; + +template<> struct conj_helper<Packet2cf, Packet2cf, true,true> +{ + EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const + { + return pconj(internal::pmul(a, b)); + } +}; + +template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b) +{ + // TODO optimize it for AltiVec + Packet1cd res = conj_helper<Packet1cd,Packet1cd,false,true>().pmul(a,b); + Packet2d s = vec_madd(b.v, b.v, p2d_ZERO_); + return Packet1cd(pdiv(res.v, s + vec_perm(s, s, p16uc_REVERSE64))); +} + +template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b) +{ + // TODO optimize it for AltiVec + Packet2cf res; + res.cd[0] = pdiv<Packet1cd>(a.cd[0], b.cd[0]); + res.cd[1] = pdiv<Packet1cd>(a.cd[1], b.cd[1]); + return res; +} + +EIGEN_STRONG_INLINE Packet1cd pcplxflip/*<Packet1cd>*/(const Packet1cd& x) +{ + return Packet1cd(preverse(Packet2d(x.v))); +} + +EIGEN_STRONG_INLINE Packet2cf pcplxflip/*<Packet2cf>*/(const Packet2cf& x) +{ + Packet2cf res; + res.cd[0] = pcplxflip(x.cd[0]); + res.cd[1] = pcplxflip(x.cd[1]); + return res; +} + +EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet1cd,2>& kernel) +{ + Packet2d tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_HI); + kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO); + kernel.packet[0].v = tmp; +} + +EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2cf,2>& kernel) +{ + Packet1cd tmp = kernel.packet[0].cd[1]; + kernel.packet[0].cd[1] = kernel.packet[1].cd[0]; + kernel.packet[1].cd[0] = tmp; +} + +template<> EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, const Packet2cf& thenPacket, const Packet2cf& elsePacket) { + Packet2cf result; + const Selector<4> ifPacket4 = { ifPacket.select[0], ifPacket.select[0], ifPacket.select[1], ifPacket.select[1] }; + result.v = pblend<Packet4f>(ifPacket4, thenPacket.v, elsePacket.v); + return result; +} + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_COMPLEX32_ALTIVEC_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/ZVector/MathFunctions.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/ZVector/MathFunctions.h new file mode 100644 index 000000000..5c7aa7256 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/ZVector/MathFunctions.h @@ -0,0 +1,137 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2007 Julien Pommier +// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2016 Konstantinos Margaritis <markos@freevec.org> +// +// 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 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 + +namespace Eigen { + +namespace internal { + +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); + +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(p2d_cephes_LOG2EF, x, p2d_half); + + fx = vec_floor(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 = vec_ctsl(fx, 0); + + static const Packet2l p2l_1023 = { 1023, 1023 }; + static const Packet2ul p2ul_52 = { 52, 52 }; + + emm0 = emm0 + p2l_1023; + emm0 = emm0 << reinterpret_cast<Packet2l>(p2ul_52); + + // 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); +} + +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet4f pexp<Packet4f>(const Packet4f& x) +{ + Packet4f res; + res.v4f[0] = pexp<Packet2d>(x.v4f[0]); + res.v4f[1] = pexp<Packet2d>(x.v4f[1]); + return res; +} + +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet2d psqrt<Packet2d>(const Packet2d& x) +{ + return __builtin_s390_vfsqdb(x); +} + +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet4f psqrt<Packet4f>(const Packet4f& x) +{ + Packet4f res; + res.v4f[0] = psqrt<Packet2d>(x.v4f[0]); + res.v4f[1] = psqrt<Packet2d>(x.v4f[1]); + return res; +} + +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet2d prsqrt<Packet2d>(const Packet2d& x) { + // Unfortunately we can't use the much faster mm_rqsrt_pd since it only provides an approximation. + return pset1<Packet2d>(1.0) / psqrt<Packet2d>(x); +} + +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet4f prsqrt<Packet4f>(const Packet4f& x) { + Packet4f res; + res.v4f[0] = prsqrt<Packet2d>(x.v4f[0]); + res.v4f[1] = prsqrt<Packet2d>(x.v4f[1]); + return res; +} + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_MATH_FUNCTIONS_ALTIVEC_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/ZVector/PacketMath.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/ZVector/PacketMath.h new file mode 100755 index 000000000..57b01fc63 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/arch/ZVector/PacketMath.h @@ -0,0 +1,945 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2016 Konstantinos Margaritis <markos@freevec.org> +// +// 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_PACKET_MATH_ZVECTOR_H +#define EIGEN_PACKET_MATH_ZVECTOR_H + +#include <stdint.h> + +namespace Eigen { + +namespace internal { + +#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD +#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 4 +#endif + +#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD +#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD +#endif + +#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD +#define EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD +#endif + +#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS +#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 16 +#endif + +typedef __vector int Packet4i; +typedef __vector unsigned int Packet4ui; +typedef __vector __bool int Packet4bi; +typedef __vector short int Packet8i; +typedef __vector unsigned char Packet16uc; +typedef __vector double Packet2d; +typedef __vector unsigned long long Packet2ul; +typedef __vector long long Packet2l; + +typedef struct { + Packet2d v4f[2]; +} Packet4f; + +typedef union { + int32_t i[4]; + uint32_t ui[4]; + int64_t l[2]; + uint64_t ul[2]; + double d[2]; + Packet4i v4i; + Packet4ui v4ui; + Packet2l v2l; + Packet2ul v2ul; + Packet2d v2d; +} Packet; + +// We don't want to write the same code all the time, but we need to reuse the constants +// and it doesn't really work to declare them global, so we define macros instead + +#define _EIGEN_DECLARE_CONST_FAST_Packet4i(NAME,X) \ + Packet4i p4i_##NAME = reinterpret_cast<Packet4i>(vec_splat_s32(X)) + +#define _EIGEN_DECLARE_CONST_FAST_Packet2d(NAME,X) \ + Packet2d p2d_##NAME = reinterpret_cast<Packet2d>(vec_splat_s64(X)) + +#define _EIGEN_DECLARE_CONST_FAST_Packet2l(NAME,X) \ + Packet2l p2l_##NAME = reinterpret_cast<Packet2l>(vec_splat_s64(X)) + +#define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \ + Packet4i p4i_##NAME = pset1<Packet4i>(X) + +#define _EIGEN_DECLARE_CONST_Packet2d(NAME,X) \ + Packet2d p2d_##NAME = pset1<Packet2d>(X) + +#define _EIGEN_DECLARE_CONST_Packet2l(NAME,X) \ + Packet2l p2l_##NAME = pset1<Packet2l>(X) + +// These constants are endian-agnostic +//static _EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0); //{ 0, 0, 0, 0,} +static _EIGEN_DECLARE_CONST_FAST_Packet4i(ONE, 1); //{ 1, 1, 1, 1} + +static _EIGEN_DECLARE_CONST_FAST_Packet2d(ZERO, 0); +static _EIGEN_DECLARE_CONST_FAST_Packet2l(ZERO, 0); +static _EIGEN_DECLARE_CONST_FAST_Packet2l(ONE, 1); + +static Packet2d p2d_ONE = { 1.0, 1.0 }; +static Packet2d p2d_ZERO_ = { -0.0, -0.0 }; + +static Packet4i p4i_COUNTDOWN = { 0, 1, 2, 3 }; +static Packet4f p4f_COUNTDOWN = { 0.0, 1.0, 2.0, 3.0 }; +static Packet2d p2d_COUNTDOWN = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet16uc>(p2d_ZERO), reinterpret_cast<Packet16uc>(p2d_ONE), 8)); + +static Packet16uc p16uc_PSET64_HI = { 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 }; +static Packet16uc p16uc_DUPLICATE32_HI = { 0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7 }; + +// Mask alignment +#define _EIGEN_MASK_ALIGNMENT 0xfffffffffffffff0 + +#define _EIGEN_ALIGNED_PTR(x) ((std::ptrdiff_t)(x) & _EIGEN_MASK_ALIGNMENT) + +// Handle endianness properly while loading constants +// Define global static constants: + +static Packet16uc p16uc_FORWARD = { 0,1,2,3, 4,5,6,7, 8,9,10,11, 12,13,14,15 }; +static Packet16uc p16uc_REVERSE32 = { 12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3 }; +static Packet16uc p16uc_REVERSE64 = { 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 }; + +static Packet16uc p16uc_PSET32_WODD = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 }; +static Packet16uc p16uc_PSET32_WEVEN = vec_sld(p16uc_DUPLICATE32_HI, (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 }; +/*static Packet16uc p16uc_HALF64_0_16 = vec_sld((Packet16uc)p4i_ZERO, vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 3), 8); //{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16}; + +static Packet16uc p16uc_PSET64_HI = (Packet16uc) vec_mergeh((Packet4ui)p16uc_PSET32_WODD, (Packet4ui)p16uc_PSET32_WEVEN); //{ 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 };*/ +static Packet16uc p16uc_PSET64_LO = (Packet16uc) vec_mergel((Packet4ui)p16uc_PSET32_WODD, (Packet4ui)p16uc_PSET32_WEVEN); //{ 8,9,10,11, 12,13,14,15, 8,9,10,11, 12,13,14,15 }; +/*static Packet16uc p16uc_TRANSPOSE64_HI = vec_add(p16uc_PSET64_HI, p16uc_HALF64_0_16); //{ 0,1,2,3, 4,5,6,7, 16,17,18,19, 20,21,22,23}; +static Packet16uc p16uc_TRANSPOSE64_LO = vec_add(p16uc_PSET64_LO, p16uc_HALF64_0_16); //{ 8,9,10,11, 12,13,14,15, 24,25,26,27, 28,29,30,31};*/ +static Packet16uc p16uc_TRANSPOSE64_HI = { 0,1,2,3, 4,5,6,7, 16,17,18,19, 20,21,22,23}; +static Packet16uc p16uc_TRANSPOSE64_LO = { 8,9,10,11, 12,13,14,15, 24,25,26,27, 28,29,30,31}; + +//static Packet16uc p16uc_COMPLEX32_REV = vec_sld(p16uc_REVERSE32, p16uc_REVERSE32, 8); //{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 }; + +//static Packet16uc p16uc_COMPLEX32_REV2 = vec_sld(p16uc_FORWARD, p16uc_FORWARD, 8); //{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 }; + + +#if EIGEN_HAS_BUILTIN(__builtin_prefetch) || EIGEN_COMP_GNUC + #define EIGEN_ZVECTOR_PREFETCH(ADDR) __builtin_prefetch(ADDR); +#else + #define EIGEN_ZVECTOR_PREFETCH(ADDR) asm( " pfd [%[addr]]\n" :: [addr] "r" (ADDR) : "cc" ); +#endif + +template<> struct packet_traits<int> : default_packet_traits +{ + typedef Packet4i type; + typedef Packet4i half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size = 4, + HasHalfPacket = 0, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasBlend = 1 + }; +}; + +template<> struct packet_traits<float> : default_packet_traits +{ + typedef Packet4f type; + typedef Packet4f half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=4, + HasHalfPacket = 0, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasMin = 1, + HasMax = 1, + HasAbs = 1, + HasSin = 0, + HasCos = 0, + HasLog = 0, + HasExp = 1, + HasSqrt = 1, + HasRsqrt = 1, + HasRound = 1, + HasFloor = 1, + HasCeil = 1, + HasNegate = 1, + HasBlend = 1 + }; +}; + +template<> struct packet_traits<double> : default_packet_traits +{ + typedef Packet2d type; + typedef Packet2d half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=2, + HasHalfPacket = 1, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasMin = 1, + HasMax = 1, + HasAbs = 1, + HasSin = 0, + HasCos = 0, + HasLog = 0, + HasExp = 1, + HasSqrt = 1, + HasRsqrt = 1, + HasRound = 1, + HasFloor = 1, + HasCeil = 1, + HasNegate = 1, + HasBlend = 1 + }; +}; + +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; }; + +/* Forward declaration */ +EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4f,4>& kernel); + +inline std::ostream & operator <<(std::ostream & s, const Packet4i & v) +{ + Packet vt; + vt.v4i = v; + s << vt.i[0] << ", " << vt.i[1] << ", " << vt.i[2] << ", " << vt.i[3]; + return s; +} + +inline std::ostream & operator <<(std::ostream & s, const Packet4ui & v) +{ + Packet vt; + vt.v4ui = v; + s << vt.ui[0] << ", " << vt.ui[1] << ", " << vt.ui[2] << ", " << vt.ui[3]; + return s; +} + +inline std::ostream & operator <<(std::ostream & s, const Packet2l & v) +{ + Packet vt; + vt.v2l = v; + s << vt.l[0] << ", " << vt.l[1]; + return s; +} + +inline std::ostream & operator <<(std::ostream & s, const Packet2ul & v) +{ + Packet vt; + vt.v2ul = v; + s << vt.ul[0] << ", " << vt.ul[1] ; + return s; +} + +inline std::ostream & operator <<(std::ostream & s, const Packet2d & v) +{ + Packet vt; + vt.v2d = v; + s << vt.d[0] << ", " << vt.d[1]; + return s; +} + +/* Helper function to simulate a vec_splat_packet4f + */ +template<int element> EIGEN_STRONG_INLINE Packet4f vec_splat_packet4f(const Packet4f& from) +{ + Packet4f splat; + switch (element) { + case 0: + splat.v4f[0] = vec_splat(from.v4f[0], 0); + splat.v4f[1] = splat.v4f[0]; + break; + case 1: + splat.v4f[0] = vec_splat(from.v4f[0], 1); + splat.v4f[1] = splat.v4f[0]; + break; + case 2: + splat.v4f[0] = vec_splat(from.v4f[1], 0); + splat.v4f[1] = splat.v4f[0]; + break; + case 3: + splat.v4f[0] = vec_splat(from.v4f[1], 1); + splat.v4f[1] = splat.v4f[0]; + break; + } + return splat; +} + +template<int Offset> +struct palign_impl<Offset,Packet4i> +{ + static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second) + { + switch (Offset % 4) { + case 1: + first = vec_sld(first, second, 4); break; + case 2: + first = vec_sld(first, second, 8); break; + case 3: + first = vec_sld(first, second, 12); break; + } + } +}; + +/* This is a tricky one, we have to translate float alignment to vector elements of sizeof double + */ +template<int Offset> +struct palign_impl<Offset,Packet4f> +{ + static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second) + { + switch (Offset % 4) { + case 1: + first.v4f[0] = vec_sld(first.v4f[0], first.v4f[1], 8); + first.v4f[1] = vec_sld(first.v4f[1], second.v4f[0], 8); + break; + case 2: + first.v4f[0] = first.v4f[1]; + first.v4f[1] = second.v4f[0]; + break; + case 3: + first.v4f[0] = vec_sld(first.v4f[1], second.v4f[0], 8); + first.v4f[1] = vec_sld(second.v4f[0], second.v4f[1], 8); + break; + } + } +}; + + +template<int Offset> +struct palign_impl<Offset,Packet2d> +{ + static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second) + { + if (Offset == 1) + first = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(first), reinterpret_cast<Packet4i>(second), 8)); + } +}; + +template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int* from) +{ + // FIXME: No intrinsic yet + EIGEN_DEBUG_ALIGNED_LOAD + Packet *vfrom; + vfrom = (Packet *) from; + return vfrom->v4i; +} + +template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) +{ + // FIXME: No intrinsic yet + EIGEN_DEBUG_ALIGNED_LOAD + Packet4f vfrom; + vfrom.v4f[0] = vec_ld2f(&from[0]); + vfrom.v4f[1] = vec_ld2f(&from[2]); + return vfrom; +} + +template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from) +{ + // FIXME: No intrinsic yet + EIGEN_DEBUG_ALIGNED_LOAD + Packet *vfrom; + vfrom = (Packet *) from; + return vfrom->v2d; +} + +template<> EIGEN_STRONG_INLINE void pstore<int>(int* to, const Packet4i& from) +{ + // FIXME: No intrinsic yet + EIGEN_DEBUG_ALIGNED_STORE + Packet *vto; + vto = (Packet *) to; + vto->v4i = from; +} + +template<> EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet4f& from) +{ + // FIXME: No intrinsic yet + EIGEN_DEBUG_ALIGNED_STORE + vec_st2f(from.v4f[0], &to[0]); + vec_st2f(from.v4f[1], &to[2]); +} + + +template<> EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet2d& from) +{ + // FIXME: No intrinsic yet + EIGEN_DEBUG_ALIGNED_STORE + Packet *vto; + vto = (Packet *) to; + vto->v2d = from; +} + +template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) +{ + return vec_splats(from); +} +template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { + return vec_splats(from); +} +template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) +{ + Packet4f to; + to.v4f[0] = pset1<Packet2d>(static_cast<const double&>(from)); + to.v4f[1] = to.v4f[0]; + return to; +} + +template<> EIGEN_STRONG_INLINE void +pbroadcast4<Packet4i>(const int *a, + Packet4i& a0, Packet4i& a1, Packet4i& a2, Packet4i& a3) +{ + a3 = pload<Packet4i>(a); + a0 = vec_splat(a3, 0); + a1 = vec_splat(a3, 1); + a2 = vec_splat(a3, 2); + a3 = vec_splat(a3, 3); +} + +template<> EIGEN_STRONG_INLINE void +pbroadcast4<Packet4f>(const float *a, + Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3) +{ + a3 = pload<Packet4f>(a); + a0 = vec_splat_packet4f<0>(a3); + a1 = vec_splat_packet4f<1>(a3); + a2 = vec_splat_packet4f<2>(a3); + a3 = vec_splat_packet4f<3>(a3); +} + +template<> EIGEN_STRONG_INLINE void +pbroadcast4<Packet2d>(const double *a, + Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3) +{ + a1 = pload<Packet2d>(a); + a0 = vec_splat(a1, 0); + a1 = vec_splat(a1, 1); + a3 = pload<Packet2d>(a+2); + a2 = vec_splat(a3, 0); + a3 = vec_splat(a3, 1); +} + +template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* from, Index stride) +{ + int EIGEN_ALIGN16 ai[4]; + ai[0] = from[0*stride]; + ai[1] = from[1*stride]; + ai[2] = from[2*stride]; + ai[3] = from[3*stride]; + return pload<Packet4i>(ai); +} + +template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride) +{ + float EIGEN_ALIGN16 ai[4]; + ai[0] = from[0*stride]; + ai[1] = from[1*stride]; + ai[2] = from[2*stride]; + ai[3] = from[3*stride]; + return pload<Packet4f>(ai); +} + +template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride) +{ + double EIGEN_ALIGN16 af[2]; + af[0] = from[0*stride]; + af[1] = from[1*stride]; + return pload<Packet2d>(af); +} + +template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const Packet4i& from, Index stride) +{ + int EIGEN_ALIGN16 ai[4]; + pstore<int>((int *)ai, from); + to[0*stride] = ai[0]; + to[1*stride] = ai[1]; + to[2*stride] = ai[2]; + to[3*stride] = ai[3]; +} + +template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride) +{ + float EIGEN_ALIGN16 ai[4]; + pstore<float>((float *)ai, from); + to[0*stride] = ai[0]; + to[1*stride] = ai[1]; + to[2*stride] = ai[2]; + to[3*stride] = ai[3]; +} + +template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride) +{ + double EIGEN_ALIGN16 af[2]; + pstore<double>(af, from); + to[0*stride] = af[0]; + to[1*stride] = af[1]; +} + +template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a + b); } +template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) +{ + Packet4f c; + c.v4f[0] = a.v4f[0] + b.v4f[0]; + c.v4f[1] = a.v4f[1] + b.v4f[1]; + return c; +} +template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a + b); } + +template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a - b); } +template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) +{ + Packet4f c; + c.v4f[0] = a.v4f[0] - b.v4f[0]; + c.v4f[1] = a.v4f[1] - b.v4f[1]; + return c; +} +template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a - b); } + +template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a * b); } +template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) +{ + Packet4f c; + c.v4f[0] = a.v4f[0] * b.v4f[0]; + c.v4f[1] = a.v4f[1] * b.v4f[1]; + return c; +} +template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a * b); } + +template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a / b); } +template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) +{ + Packet4f c; + c.v4f[0] = a.v4f[0] / b.v4f[0]; + c.v4f[1] = a.v4f[1] / b.v4f[1]; + return c; +} +template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a / b); } + +template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return (-a); } +template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a) +{ + Packet4f c; + c.v4f[0] = -a.v4f[0]; + c.v4f[1] = -a.v4f[1]; + return c; +} +template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) { return (-a); } + +template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; } +template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; } +template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; } + +template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd<Packet4i>(pmul<Packet4i>(a, b), c); } +template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) +{ + Packet4f res; + res.v4f[0] = vec_madd(a.v4f[0], b.v4f[0], c.v4f[0]); + res.v4f[1] = vec_madd(a.v4f[1], b.v4f[1], c.v4f[1]); + return res; +} +template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vec_madd(a, b, c); } + +template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int& a) { return padd<Packet4i>(pset1<Packet4i>(a), p4i_COUNTDOWN); } +template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a) { return padd<Packet4f>(pset1<Packet4f>(a), p4f_COUNTDOWN); } +template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a) { return padd<Packet2d>(pset1<Packet2d>(a), p2d_COUNTDOWN); } + +template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_min(a, b); } +template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_min(a, b); } +template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) +{ + Packet4f res; + res.v4f[0] = pmin(a.v4f[0], b.v4f[0]); + res.v4f[1] = pmin(a.v4f[1], b.v4f[1]); + return res; +} + +template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_max(a, b); } +template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_max(a, b); } +template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) +{ + Packet4f res; + res.v4f[0] = pmax(a.v4f[0], b.v4f[0]); + res.v4f[1] = pmax(a.v4f[1], b.v4f[1]); + return res; +} + +template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, b); } +template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, b); } +template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) +{ + Packet4f res; + res.v4f[0] = pand(a.v4f[0], b.v4f[0]); + res.v4f[1] = pand(a.v4f[1], b.v4f[1]); + return res; +} + +template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_or(a, b); } +template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_or(a, b); } +template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b) +{ + Packet4f res; + res.v4f[0] = pand(a.v4f[0], b.v4f[0]); + res.v4f[1] = pand(a.v4f[1], b.v4f[1]); + return res; +} + +template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_xor(a, b); } +template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_xor(a, b); } +template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b) +{ + Packet4f res; + res.v4f[0] = pand(a.v4f[0], b.v4f[0]); + res.v4f[1] = pand(a.v4f[1], b.v4f[1]); + return res; +} + +template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return pand<Packet4i>(a, vec_nor(b, b)); } +template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, vec_nor(b, b)); } +template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) +{ + Packet4f res; + res.v4f[0] = pandnot(a.v4f[0], b.v4f[0]); + res.v4f[1] = pandnot(a.v4f[1], b.v4f[1]); + return res; +} + +template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a) +{ + Packet4f res; + res.v4f[0] = vec_round(a.v4f[0]); + res.v4f[1] = vec_round(a.v4f[1]); + return res; +} +template<> EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a) { return vec_round(a); } +template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const Packet4f& a) +{ + Packet4f res; + res.v4f[0] = vec_ceil(a.v4f[0]); + res.v4f[1] = vec_ceil(a.v4f[1]); + return res; +} +template<> EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const Packet2d& a) { return vec_ceil(a); } +template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) +{ + Packet4f res; + res.v4f[0] = vec_floor(a.v4f[0]); + res.v4f[1] = vec_floor(a.v4f[1]); + return res; +} +template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) { return vec_floor(a); } + +template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from) { return pload<Packet4i>(from); } +template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) { return pload<Packet4f>(from); } +template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from) { return pload<Packet2d>(from); } + + +template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int* from) +{ + Packet4i p = pload<Packet4i>(from); + return vec_perm(p, p, p16uc_DUPLICATE32_HI); +} + +template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from) +{ + Packet4f p = pload<Packet4f>(from); + p.v4f[1] = vec_splat(p.v4f[0], 1); + p.v4f[0] = vec_splat(p.v4f[0], 0); + return p; +} + +template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double* from) +{ + Packet2d p = pload<Packet2d>(from); + return vec_perm(p, p, p16uc_PSET64_HI); +} + +template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet4i& from) { pstore<int>(to, from); } +template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from) { pstore<float>(to, from); } +template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d& from) { pstore<double>(to, from); } + +template<> EIGEN_STRONG_INLINE void prefetch<int>(const int* addr) { EIGEN_ZVECTOR_PREFETCH(addr); } +template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { EIGEN_ZVECTOR_PREFETCH(addr); } +template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { EIGEN_ZVECTOR_PREFETCH(addr); } + +template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { int EIGEN_ALIGN16 x[4]; pstore(x, a); return x[0]; } +template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { float EIGEN_ALIGN16 x[2]; vec_st2f(a.v4f[0], &x[0]); return x[0]; } +template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { double EIGEN_ALIGN16 x[2]; pstore(x, a); return x[0]; } + +template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) +{ + return reinterpret_cast<Packet4i>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE32)); +} + +template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a) +{ + return reinterpret_cast<Packet2d>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE64)); +} + +template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) +{ + Packet4f rev; + rev.v4f[0] = preverse<Packet2d>(a.v4f[1]); + rev.v4f[1] = preverse<Packet2d>(a.v4f[0]); + return rev; +} + +template<> EIGEN_STRONG_INLINE Packet4i pabs<Packet4i>(const Packet4i& a) { return vec_abs(a); } +template<> EIGEN_STRONG_INLINE Packet2d pabs<Packet2d>(const Packet2d& a) { return vec_abs(a); } +template<> EIGEN_STRONG_INLINE Packet4f pabs<Packet4f>(const Packet4f& a) +{ + Packet4f res; + res.v4f[0] = pabs(a.v4f[0]); + res.v4f[1] = pabs(a.v4f[1]); + return res; +} + +template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a) +{ + Packet4i b, sum; + b = vec_sld(a, a, 8); + sum = padd<Packet4i>(a, b); + b = vec_sld(sum, sum, 4); + sum = padd<Packet4i>(sum, b); + return pfirst(sum); +} + +template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) +{ + Packet2d b, sum; + b = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8)); + sum = padd<Packet2d>(a, b); + return pfirst(sum); +} +template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a) +{ + Packet2d sum; + sum = padd<Packet2d>(a.v4f[0], a.v4f[1]); + double first = predux<Packet2d>(sum); + return static_cast<float>(first); +} + +template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs) +{ + Packet4i v[4], sum[4]; + + // It's easier and faster to transpose then add as columns + // Check: http://www.freevec.org/function/matrix_4x4_transpose_floats for explanation + // Do the transpose, first set of moves + v[0] = vec_mergeh(vecs[0], vecs[2]); + v[1] = vec_mergel(vecs[0], vecs[2]); + v[2] = vec_mergeh(vecs[1], vecs[3]); + v[3] = vec_mergel(vecs[1], vecs[3]); + // Get the resulting vectors + sum[0] = vec_mergeh(v[0], v[2]); + sum[1] = vec_mergel(v[0], v[2]); + sum[2] = vec_mergeh(v[1], v[3]); + sum[3] = vec_mergel(v[1], v[3]); + + // Now do the summation: + // Lines 0+1 + sum[0] = padd<Packet4i>(sum[0], sum[1]); + // Lines 2+3 + sum[1] = padd<Packet4i>(sum[2], sum[3]); + // Add the results + sum[0] = padd<Packet4i>(sum[0], sum[1]); + + return sum[0]; +} + +template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs) +{ + Packet2d v[2], sum; + v[0] = padd<Packet2d>(vecs[0], reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(vecs[0]), reinterpret_cast<Packet4ui>(vecs[0]), 8))); + v[1] = padd<Packet2d>(vecs[1], reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(vecs[1]), reinterpret_cast<Packet4ui>(vecs[1]), 8))); + + sum = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(v[0]), reinterpret_cast<Packet4ui>(v[1]), 8)); + + return sum; +} + +template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs) +{ + PacketBlock<Packet4f,4> transpose; + transpose.packet[0] = vecs[0]; + transpose.packet[1] = vecs[1]; + transpose.packet[2] = vecs[2]; + transpose.packet[3] = vecs[3]; + ptranspose(transpose); + + Packet4f sum = padd(transpose.packet[0], transpose.packet[1]); + sum = padd(sum, transpose.packet[2]); + sum = padd(sum, transpose.packet[3]); + return sum; +} + +// Other reduction functions: +// mul +template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a) +{ + EIGEN_ALIGN16 int aux[4]; + pstore(aux, a); + return aux[0] * aux[1] * aux[2] * aux[3]; +} + +template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a) +{ + return pfirst(pmul(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8)))); +} + +template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a) +{ + // Return predux_mul<Packet2d> of the subvectors product + return static_cast<float>(pfirst(predux_mul(pmul(a.v4f[0], a.v4f[1])))); +} + +// min +template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a) +{ + Packet4i b, res; + b = pmin<Packet4i>(a, vec_sld(a, a, 8)); + res = pmin<Packet4i>(b, vec_sld(b, b, 4)); + return pfirst(res); +} + +template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a) +{ + return pfirst(pmin<Packet2d>(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8)))); +} + +template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a) +{ + Packet2d b, res; + b = pmin<Packet2d>(a.v4f[0], a.v4f[1]); + res = pmin<Packet2d>(b, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(b), reinterpret_cast<Packet4i>(b), 8))); + return static_cast<float>(pfirst(res)); +} + +// max +template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a) +{ + Packet4i b, res; + b = pmax<Packet4i>(a, vec_sld(a, a, 8)); + res = pmax<Packet4i>(b, vec_sld(b, b, 4)); + return pfirst(res); +} + +// max +template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a) +{ + return pfirst(pmax<Packet2d>(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8)))); +} + +template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a) +{ + Packet2d b, res; + b = pmax<Packet2d>(a.v4f[0], a.v4f[1]); + res = pmax<Packet2d>(b, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(b), reinterpret_cast<Packet4i>(b), 8))); + return static_cast<float>(pfirst(res)); +} + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet4i,4>& kernel) { + Packet4i t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]); + Packet4i t1 = vec_mergel(kernel.packet[0], kernel.packet[2]); + Packet4i t2 = vec_mergeh(kernel.packet[1], kernel.packet[3]); + Packet4i t3 = vec_mergel(kernel.packet[1], kernel.packet[3]); + kernel.packet[0] = vec_mergeh(t0, t2); + kernel.packet[1] = vec_mergel(t0, t2); + kernel.packet[2] = vec_mergeh(t1, t3); + kernel.packet[3] = vec_mergel(t1, t3); +} + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet2d,2>& kernel) { + Packet2d t0 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_HI); + Packet2d t1 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_LO); + kernel.packet[0] = t0; + kernel.packet[1] = t1; +} + +/* Split the Packet4f PacketBlock into 4 Packet2d PacketBlocks and transpose each one + */ +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet4f,4>& kernel) { + PacketBlock<Packet2d,2> t0,t1,t2,t3; + // copy top-left 2x2 Packet2d block + t0.packet[0] = kernel.packet[0].v4f[0]; + t0.packet[1] = kernel.packet[1].v4f[0]; + + // copy top-right 2x2 Packet2d block + t1.packet[0] = kernel.packet[0].v4f[1]; + t1.packet[1] = kernel.packet[1].v4f[1]; + + // copy bottom-left 2x2 Packet2d block + t2.packet[0] = kernel.packet[2].v4f[0]; + t2.packet[1] = kernel.packet[3].v4f[0]; + + // copy bottom-right 2x2 Packet2d block + t3.packet[0] = kernel.packet[2].v4f[1]; + t3.packet[1] = kernel.packet[3].v4f[1]; + + // Transpose all 2x2 blocks + ptranspose(t0); + ptranspose(t1); + ptranspose(t2); + ptranspose(t3); + + // Copy back transposed blocks, but exchange t1 and t2 due to transposition + kernel.packet[0].v4f[0] = t0.packet[0]; + kernel.packet[0].v4f[1] = t2.packet[0]; + kernel.packet[1].v4f[0] = t0.packet[1]; + kernel.packet[1].v4f[1] = t2.packet[1]; + kernel.packet[2].v4f[0] = t1.packet[0]; + kernel.packet[2].v4f[1] = t3.packet[0]; + kernel.packet[3].v4f[0] = t1.packet[1]; + kernel.packet[3].v4f[1] = t3.packet[1]; +} + +template<> EIGEN_STRONG_INLINE Packet4i pblend(const Selector<4>& ifPacket, const Packet4i& thenPacket, const Packet4i& elsePacket) { + Packet4ui select = { ifPacket.select[0], ifPacket.select[1], ifPacket.select[2], ifPacket.select[3] }; + Packet4ui mask = vec_cmpeq(select, reinterpret_cast<Packet4ui>(p4i_ONE)); + return vec_sel(elsePacket, thenPacket, mask); +} + +template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket, const Packet4f& elsePacket) { + Packet2ul select_hi = { ifPacket.select[0], ifPacket.select[1] }; + Packet2ul select_lo = { ifPacket.select[2], ifPacket.select[3] }; + Packet2ul mask_hi = vec_cmpeq(select_hi, reinterpret_cast<Packet2ul>(p2l_ONE)); + Packet2ul mask_lo = vec_cmpeq(select_lo, reinterpret_cast<Packet2ul>(p2l_ONE)); + Packet4f result; + result.v4f[0] = vec_sel(elsePacket.v4f[0], thenPacket.v4f[0], mask_hi); + result.v4f[1] = vec_sel(elsePacket.v4f[1], thenPacket.v4f[1], mask_lo); + return result; +} + +template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket, const Packet2d& elsePacket) { + Packet2ul select = { ifPacket.select[0], ifPacket.select[1] }; + Packet2ul mask = vec_cmpeq(select, reinterpret_cast<Packet2ul>(p2l_ONE)); + return vec_sel(elsePacket, thenPacket, mask); +} + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_PACKET_MATH_ZVECTOR_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/functors/AssignmentFunctors.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/functors/AssignmentFunctors.h new file mode 100644 index 000000000..4153b877c --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/functors/AssignmentFunctors.h @@ -0,0 +1,168 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2010 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_ASSIGNMENT_FUNCTORS_H +#define EIGEN_ASSIGNMENT_FUNCTORS_H + +namespace Eigen { + +namespace internal { + +/** \internal + * \brief Template functor for scalar/packet assignment + * + */ +template<typename DstScalar,typename SrcScalar> struct assign_op { + + EIGEN_EMPTY_STRUCT_CTOR(assign_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(DstScalar& a, const SrcScalar& b) const { a = b; } + + template<int Alignment, typename Packet> + EIGEN_STRONG_INLINE void assignPacket(DstScalar* a, const Packet& b) const + { internal::pstoret<DstScalar,Packet,Alignment>(a,b); } +}; + +// Empty overload for void type (used by PermutationMatrix) +template<typename DstScalar> struct assign_op<DstScalar,void> {}; + +template<typename DstScalar,typename SrcScalar> +struct functor_traits<assign_op<DstScalar,SrcScalar> > { + enum { + Cost = NumTraits<DstScalar>::ReadCost, + PacketAccess = is_same<DstScalar,SrcScalar>::value && packet_traits<DstScalar>::Vectorizable && packet_traits<SrcScalar>::Vectorizable + }; +}; + +/** \internal + * \brief Template functor for scalar/packet assignment with addition + * + */ +template<typename DstScalar,typename SrcScalar> struct add_assign_op { + + EIGEN_EMPTY_STRUCT_CTOR(add_assign_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(DstScalar& a, const SrcScalar& b) const { a += b; } + + template<int Alignment, typename Packet> + EIGEN_STRONG_INLINE void assignPacket(DstScalar* a, const Packet& b) const + { internal::pstoret<DstScalar,Packet,Alignment>(a,internal::padd(internal::ploadt<Packet,Alignment>(a),b)); } +}; +template<typename DstScalar,typename SrcScalar> +struct functor_traits<add_assign_op<DstScalar,SrcScalar> > { + enum { + Cost = NumTraits<DstScalar>::ReadCost + NumTraits<DstScalar>::AddCost, + PacketAccess = is_same<DstScalar,SrcScalar>::value && packet_traits<DstScalar>::HasAdd + }; +}; + +/** \internal + * \brief Template functor for scalar/packet assignment with subtraction + * + */ +template<typename DstScalar,typename SrcScalar> struct sub_assign_op { + + EIGEN_EMPTY_STRUCT_CTOR(sub_assign_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(DstScalar& a, const SrcScalar& b) const { a -= b; } + + template<int Alignment, typename Packet> + EIGEN_STRONG_INLINE void assignPacket(DstScalar* a, const Packet& b) const + { internal::pstoret<DstScalar,Packet,Alignment>(a,internal::psub(internal::ploadt<Packet,Alignment>(a),b)); } +}; +template<typename DstScalar,typename SrcScalar> +struct functor_traits<sub_assign_op<DstScalar,SrcScalar> > { + enum { + Cost = NumTraits<DstScalar>::ReadCost + NumTraits<DstScalar>::AddCost, + PacketAccess = is_same<DstScalar,SrcScalar>::value && packet_traits<DstScalar>::HasSub + }; +}; + +/** \internal + * \brief Template functor for scalar/packet assignment with multiplication + * + */ +template<typename DstScalar, typename SrcScalar=DstScalar> +struct mul_assign_op { + + EIGEN_EMPTY_STRUCT_CTOR(mul_assign_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(DstScalar& a, const SrcScalar& b) const { a *= b; } + + template<int Alignment, typename Packet> + EIGEN_STRONG_INLINE void assignPacket(DstScalar* a, const Packet& b) const + { internal::pstoret<DstScalar,Packet,Alignment>(a,internal::pmul(internal::ploadt<Packet,Alignment>(a),b)); } +}; +template<typename DstScalar, typename SrcScalar> +struct functor_traits<mul_assign_op<DstScalar,SrcScalar> > { + enum { + Cost = NumTraits<DstScalar>::ReadCost + NumTraits<DstScalar>::MulCost, + PacketAccess = is_same<DstScalar,SrcScalar>::value && packet_traits<DstScalar>::HasMul + }; +}; + +/** \internal + * \brief Template functor for scalar/packet assignment with diviving + * + */ +template<typename DstScalar, typename SrcScalar=DstScalar> struct div_assign_op { + + EIGEN_EMPTY_STRUCT_CTOR(div_assign_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(DstScalar& a, const SrcScalar& b) const { a /= b; } + + template<int Alignment, typename Packet> + EIGEN_STRONG_INLINE void assignPacket(DstScalar* a, const Packet& b) const + { internal::pstoret<DstScalar,Packet,Alignment>(a,internal::pdiv(internal::ploadt<Packet,Alignment>(a),b)); } +}; +template<typename DstScalar, typename SrcScalar> +struct functor_traits<div_assign_op<DstScalar,SrcScalar> > { + enum { + Cost = NumTraits<DstScalar>::ReadCost + NumTraits<DstScalar>::MulCost, + PacketAccess = is_same<DstScalar,SrcScalar>::value && packet_traits<DstScalar>::HasDiv + }; +}; + +/** \internal + * \brief Template functor for scalar/packet assignment with swapping + * + * It works as follow. For a non-vectorized evaluation loop, we have: + * for(i) func(A.coeffRef(i), B.coeff(i)); + * where B is a SwapWrapper expression. The trick is to make SwapWrapper::coeff behaves like a non-const coeffRef. + * Actually, SwapWrapper might not even be needed since even if B is a plain expression, since it has to be writable + * B.coeff already returns a const reference to the underlying scalar value. + * + * The case of a vectorized loop is more tricky: + * for(i,j) func.assignPacket<A_Align>(&A.coeffRef(i,j), B.packet<B_Align>(i,j)); + * Here, B must be a SwapWrapper whose packet function actually returns a proxy object holding a Scalar*, + * the actual alignment and Packet type. + * + */ +template<typename Scalar> struct swap_assign_op { + + EIGEN_EMPTY_STRUCT_CTOR(swap_assign_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(Scalar& a, const Scalar& b) const + { +#ifdef __CUDACC__ + // FIXME is there some kind of cuda::swap? + Scalar t=b; const_cast<Scalar&>(b)=a; a=t; +#else + using std::swap; + swap(a,const_cast<Scalar&>(b)); +#endif + } +}; +template<typename Scalar> +struct functor_traits<swap_assign_op<Scalar> > { + enum { + Cost = 3 * NumTraits<Scalar>::ReadCost, + PacketAccess = packet_traits<Scalar>::Vectorizable + }; +}; + +} // namespace internal + +} // namespace Eigen + +#endif // EIGEN_ASSIGNMENT_FUNCTORS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/functors/BinaryFunctors.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/functors/BinaryFunctors.h new file mode 100644 index 000000000..96747bac7 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/functors/BinaryFunctors.h @@ -0,0 +1,482 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2010 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_BINARY_FUNCTORS_H +#define EIGEN_BINARY_FUNCTORS_H + +namespace Eigen { + +namespace internal { + +//---------- associative binary functors ---------- + +template<typename Arg1, typename Arg2> +struct binary_op_base +{ + typedef Arg1 first_argument_type; + typedef Arg2 second_argument_type; +}; + +/** \internal + * \brief Template functor to compute the sum of two scalars + * + * \sa class CwiseBinaryOp, MatrixBase::operator+, class VectorwiseOp, DenseBase::sum() + */ +template<typename LhsScalar,typename RhsScalar> +struct scalar_sum_op : binary_op_base<LhsScalar,RhsScalar> +{ + typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar,scalar_sum_op>::ReturnType result_type; +#ifndef EIGEN_SCALAR_BINARY_OP_PLUGIN + EIGEN_EMPTY_STRUCT_CTOR(scalar_sum_op) +#else + scalar_sum_op() { + EIGEN_SCALAR_BINARY_OP_PLUGIN + } +#endif + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a + b; } + template<typename Packet> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const + { return internal::padd(a,b); } + template<typename Packet> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type predux(const Packet& a) const + { return internal::predux(a); } +}; +template<typename LhsScalar,typename RhsScalar> +struct functor_traits<scalar_sum_op<LhsScalar,RhsScalar> > { + enum { + Cost = (NumTraits<LhsScalar>::AddCost+NumTraits<RhsScalar>::AddCost)/2, // rough estimate! + PacketAccess = is_same<LhsScalar,RhsScalar>::value && packet_traits<LhsScalar>::HasAdd && packet_traits<RhsScalar>::HasAdd + // TODO vectorize mixed sum + }; +}; + +/** \internal + * \brief Template specialization to deprecate the summation of boolean expressions. + * This is required to solve Bug 426. + * \sa DenseBase::count(), DenseBase::any(), ArrayBase::cast(), MatrixBase::cast() + */ +template<> struct scalar_sum_op<bool,bool> : scalar_sum_op<int,int> { + EIGEN_DEPRECATED + scalar_sum_op() {} +}; + + +/** \internal + * \brief Template functor to compute the product of two scalars + * + * \sa class CwiseBinaryOp, Cwise::operator*(), class VectorwiseOp, MatrixBase::redux() + */ +template<typename LhsScalar,typename RhsScalar> +struct scalar_product_op : binary_op_base<LhsScalar,RhsScalar> +{ + typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar,scalar_product_op>::ReturnType result_type; +#ifndef EIGEN_SCALAR_BINARY_OP_PLUGIN + EIGEN_EMPTY_STRUCT_CTOR(scalar_product_op) +#else + scalar_product_op() { + EIGEN_SCALAR_BINARY_OP_PLUGIN + } +#endif + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a * b; } + template<typename Packet> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const + { return internal::pmul(a,b); } + template<typename Packet> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type predux(const Packet& a) const + { return internal::predux_mul(a); } +}; +template<typename LhsScalar,typename RhsScalar> +struct functor_traits<scalar_product_op<LhsScalar,RhsScalar> > { + enum { + Cost = (NumTraits<LhsScalar>::MulCost + NumTraits<RhsScalar>::MulCost)/2, // rough estimate! + PacketAccess = is_same<LhsScalar,RhsScalar>::value && packet_traits<LhsScalar>::HasMul && packet_traits<RhsScalar>::HasMul + // TODO vectorize mixed product + }; +}; + +/** \internal + * \brief Template functor to compute the conjugate product of two scalars + * + * This is a short cut for conj(x) * y which is needed for optimization purpose; in Eigen2 support mode, this becomes x * conj(y) + */ +template<typename LhsScalar,typename RhsScalar> +struct scalar_conj_product_op : binary_op_base<LhsScalar,RhsScalar> +{ + + enum { + Conj = NumTraits<LhsScalar>::IsComplex + }; + + typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar,scalar_conj_product_op>::ReturnType result_type; + + EIGEN_EMPTY_STRUCT_CTOR(scalar_conj_product_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const + { return conj_helper<LhsScalar,RhsScalar,Conj,false>().pmul(a,b); } + + template<typename Packet> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const + { return conj_helper<Packet,Packet,Conj,false>().pmul(a,b); } +}; +template<typename LhsScalar,typename RhsScalar> +struct functor_traits<scalar_conj_product_op<LhsScalar,RhsScalar> > { + enum { + Cost = NumTraits<LhsScalar>::MulCost, + PacketAccess = internal::is_same<LhsScalar, RhsScalar>::value && packet_traits<LhsScalar>::HasMul + }; +}; + +/** \internal + * \brief Template functor to compute the min of two scalars + * + * \sa class CwiseBinaryOp, MatrixBase::cwiseMin, class VectorwiseOp, MatrixBase::minCoeff() + */ +template<typename LhsScalar,typename RhsScalar> +struct scalar_min_op : binary_op_base<LhsScalar,RhsScalar> +{ + typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar,scalar_min_op>::ReturnType result_type; + EIGEN_EMPTY_STRUCT_CTOR(scalar_min_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return numext::mini(a, b); } + template<typename Packet> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const + { return internal::pmin(a,b); } + template<typename Packet> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type predux(const Packet& a) const + { return internal::predux_min(a); } +}; +template<typename LhsScalar,typename RhsScalar> +struct functor_traits<scalar_min_op<LhsScalar,RhsScalar> > { + enum { + Cost = (NumTraits<LhsScalar>::AddCost+NumTraits<RhsScalar>::AddCost)/2, + PacketAccess = internal::is_same<LhsScalar, RhsScalar>::value && packet_traits<LhsScalar>::HasMin + }; +}; + +/** \internal + * \brief Template functor to compute the max of two scalars + * + * \sa class CwiseBinaryOp, MatrixBase::cwiseMax, class VectorwiseOp, MatrixBase::maxCoeff() + */ +template<typename LhsScalar,typename RhsScalar> +struct scalar_max_op : binary_op_base<LhsScalar,RhsScalar> +{ + typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar,scalar_max_op>::ReturnType result_type; + EIGEN_EMPTY_STRUCT_CTOR(scalar_max_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return numext::maxi(a, b); } + template<typename Packet> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const + { return internal::pmax(a,b); } + template<typename Packet> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type predux(const Packet& a) const + { return internal::predux_max(a); } +}; +template<typename LhsScalar,typename RhsScalar> +struct functor_traits<scalar_max_op<LhsScalar,RhsScalar> > { + enum { + Cost = (NumTraits<LhsScalar>::AddCost+NumTraits<RhsScalar>::AddCost)/2, + PacketAccess = internal::is_same<LhsScalar, RhsScalar>::value && packet_traits<LhsScalar>::HasMax + }; +}; + +/** \internal + * \brief Template functors for comparison of two scalars + * \todo Implement packet-comparisons + */ +template<typename LhsScalar, typename RhsScalar, ComparisonName cmp> struct scalar_cmp_op; + +template<typename LhsScalar, typename RhsScalar, ComparisonName cmp> +struct functor_traits<scalar_cmp_op<LhsScalar,RhsScalar, cmp> > { + enum { + Cost = (NumTraits<LhsScalar>::AddCost+NumTraits<RhsScalar>::AddCost)/2, + PacketAccess = false + }; +}; + +template<ComparisonName Cmp, typename LhsScalar, typename RhsScalar> +struct result_of<scalar_cmp_op<LhsScalar, RhsScalar, Cmp>(LhsScalar,RhsScalar)> { + typedef bool type; +}; + + +template<typename LhsScalar, typename RhsScalar> +struct scalar_cmp_op<LhsScalar,RhsScalar, cmp_EQ> : binary_op_base<LhsScalar,RhsScalar> +{ + typedef bool result_type; + EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator()(const LhsScalar& a, const RhsScalar& b) const {return a==b;} +}; +template<typename LhsScalar, typename RhsScalar> +struct scalar_cmp_op<LhsScalar,RhsScalar, cmp_LT> : binary_op_base<LhsScalar,RhsScalar> +{ + typedef bool result_type; + EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator()(const LhsScalar& a, const RhsScalar& b) const {return a<b;} +}; +template<typename LhsScalar, typename RhsScalar> +struct scalar_cmp_op<LhsScalar,RhsScalar, cmp_LE> : binary_op_base<LhsScalar,RhsScalar> +{ + typedef bool result_type; + EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator()(const LhsScalar& a, const RhsScalar& b) const {return a<=b;} +}; +template<typename LhsScalar, typename RhsScalar> +struct scalar_cmp_op<LhsScalar,RhsScalar, cmp_GT> : binary_op_base<LhsScalar,RhsScalar> +{ + typedef bool result_type; + EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator()(const LhsScalar& a, const RhsScalar& b) const {return a>b;} +}; +template<typename LhsScalar, typename RhsScalar> +struct scalar_cmp_op<LhsScalar,RhsScalar, cmp_GE> : binary_op_base<LhsScalar,RhsScalar> +{ + typedef bool result_type; + EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator()(const LhsScalar& a, const RhsScalar& b) const {return a>=b;} +}; +template<typename LhsScalar, typename RhsScalar> +struct scalar_cmp_op<LhsScalar,RhsScalar, cmp_UNORD> : binary_op_base<LhsScalar,RhsScalar> +{ + typedef bool result_type; + EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator()(const LhsScalar& a, const RhsScalar& b) const {return !(a<=b || b<=a);} +}; +template<typename LhsScalar, typename RhsScalar> +struct scalar_cmp_op<LhsScalar,RhsScalar, cmp_NEQ> : binary_op_base<LhsScalar,RhsScalar> +{ + typedef bool result_type; + EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator()(const LhsScalar& a, const RhsScalar& b) const {return a!=b;} +}; + + +/** \internal + * \brief Template functor to compute the hypot of two scalars + * + * \sa MatrixBase::stableNorm(), class Redux + */ +template<typename Scalar> +struct scalar_hypot_op<Scalar,Scalar> : binary_op_base<Scalar,Scalar> +{ + EIGEN_EMPTY_STRUCT_CTOR(scalar_hypot_op) +// typedef typename NumTraits<Scalar>::Real result_type; + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& _x, const Scalar& _y) const + { + EIGEN_USING_STD_MATH(sqrt) + Scalar p, qp; + if(_x>_y) + { + p = _x; + qp = _y / p; + } + else + { + p = _y; + qp = _x / p; + } + return p * sqrt(Scalar(1) + qp*qp); + } +}; +template<typename Scalar> +struct functor_traits<scalar_hypot_op<Scalar,Scalar> > { + enum + { + Cost = 3 * NumTraits<Scalar>::AddCost + + 2 * NumTraits<Scalar>::MulCost + + 2 * scalar_div_cost<Scalar,false>::value, + PacketAccess = false + }; +}; + +/** \internal + * \brief Template functor to compute the pow of two scalars + */ +template<typename Scalar, typename Exponent> +struct scalar_pow_op : binary_op_base<Scalar,Exponent> +{ + typedef typename ScalarBinaryOpTraits<Scalar,Exponent,scalar_pow_op>::ReturnType result_type; +#ifndef EIGEN_SCALAR_BINARY_OP_PLUGIN + EIGEN_EMPTY_STRUCT_CTOR(scalar_pow_op) +#else + scalar_pow_op() { + typedef Scalar LhsScalar; + typedef Exponent RhsScalar; + EIGEN_SCALAR_BINARY_OP_PLUGIN + } +#endif + EIGEN_DEVICE_FUNC + inline result_type operator() (const Scalar& a, const Exponent& b) const { return numext::pow(a, b); } +}; +template<typename Scalar, typename Exponent> +struct functor_traits<scalar_pow_op<Scalar,Exponent> > { + enum { Cost = 5 * NumTraits<Scalar>::MulCost, PacketAccess = false }; +}; + + + +//---------- non associative binary functors ---------- + +/** \internal + * \brief Template functor to compute the difference of two scalars + * + * \sa class CwiseBinaryOp, MatrixBase::operator- + */ +template<typename LhsScalar,typename RhsScalar> +struct scalar_difference_op : binary_op_base<LhsScalar,RhsScalar> +{ + typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar,scalar_difference_op>::ReturnType result_type; +#ifndef EIGEN_SCALAR_BINARY_OP_PLUGIN + EIGEN_EMPTY_STRUCT_CTOR(scalar_difference_op) +#else + scalar_difference_op() { + EIGEN_SCALAR_BINARY_OP_PLUGIN + } +#endif + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a - b; } + template<typename Packet> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const + { return internal::psub(a,b); } +}; +template<typename LhsScalar,typename RhsScalar> +struct functor_traits<scalar_difference_op<LhsScalar,RhsScalar> > { + enum { + Cost = (NumTraits<LhsScalar>::AddCost+NumTraits<RhsScalar>::AddCost)/2, + PacketAccess = is_same<LhsScalar,RhsScalar>::value && packet_traits<LhsScalar>::HasSub && packet_traits<RhsScalar>::HasSub + }; +}; + +/** \internal + * \brief Template functor to compute the quotient of two scalars + * + * \sa class CwiseBinaryOp, Cwise::operator/() + */ +template<typename LhsScalar,typename RhsScalar> +struct scalar_quotient_op : binary_op_base<LhsScalar,RhsScalar> +{ + typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar,scalar_quotient_op>::ReturnType result_type; +#ifndef EIGEN_SCALAR_BINARY_OP_PLUGIN + EIGEN_EMPTY_STRUCT_CTOR(scalar_quotient_op) +#else + scalar_quotient_op() { + EIGEN_SCALAR_BINARY_OP_PLUGIN + } +#endif + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a / b; } + template<typename Packet> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const + { return internal::pdiv(a,b); } +}; +template<typename LhsScalar,typename RhsScalar> +struct functor_traits<scalar_quotient_op<LhsScalar,RhsScalar> > { + typedef typename scalar_quotient_op<LhsScalar,RhsScalar>::result_type result_type; + enum { + PacketAccess = is_same<LhsScalar,RhsScalar>::value && packet_traits<LhsScalar>::HasDiv && packet_traits<RhsScalar>::HasDiv, + Cost = scalar_div_cost<result_type,PacketAccess>::value + }; +}; + + + +/** \internal + * \brief Template functor to compute the and of two booleans + * + * \sa class CwiseBinaryOp, ArrayBase::operator&& + */ +struct scalar_boolean_and_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_boolean_and_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator() (const bool& a, const bool& b) const { return a && b; } +}; +template<> struct functor_traits<scalar_boolean_and_op> { + enum { + Cost = NumTraits<bool>::AddCost, + PacketAccess = false + }; +}; + +/** \internal + * \brief Template functor to compute the or of two booleans + * + * \sa class CwiseBinaryOp, ArrayBase::operator|| + */ +struct scalar_boolean_or_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_boolean_or_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator() (const bool& a, const bool& b) const { return a || b; } +}; +template<> struct functor_traits<scalar_boolean_or_op> { + enum { + Cost = NumTraits<bool>::AddCost, + PacketAccess = false + }; +}; + +/** \internal + * \brief Template functor to compute the xor of two booleans + * + * \sa class CwiseBinaryOp, ArrayBase::operator^ + */ +struct scalar_boolean_xor_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_boolean_xor_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator() (const bool& a, const bool& b) const { return a ^ b; } +}; +template<> struct functor_traits<scalar_boolean_xor_op> { + enum { + Cost = NumTraits<bool>::AddCost, + PacketAccess = false + }; +}; + + + +//---------- binary functors bound to a constant, thus appearing as a unary functor ---------- + +// The following two classes permits to turn any binary functor into a unary one with one argument bound to a constant value. +// They are analogues to std::binder1st/binder2nd but with the following differences: +// - they are compatible with packetOp +// - they are portable across C++ versions (the std::binder* are deprecated in C++11) +template<typename BinaryOp> struct bind1st_op : BinaryOp { + + typedef typename BinaryOp::first_argument_type first_argument_type; + typedef typename BinaryOp::second_argument_type second_argument_type; + typedef typename BinaryOp::result_type result_type; + + bind1st_op(const first_argument_type &val) : m_value(val) {} + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const second_argument_type& b) const { return BinaryOp::operator()(m_value,b); } + + template<typename Packet> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& b) const + { return BinaryOp::packetOp(internal::pset1<Packet>(m_value), b); } + + first_argument_type m_value; +}; +template<typename BinaryOp> struct functor_traits<bind1st_op<BinaryOp> > : functor_traits<BinaryOp> {}; + + +template<typename BinaryOp> struct bind2nd_op : BinaryOp { + + typedef typename BinaryOp::first_argument_type first_argument_type; + typedef typename BinaryOp::second_argument_type second_argument_type; + typedef typename BinaryOp::result_type result_type; + + bind2nd_op(const second_argument_type &val) : m_value(val) {} + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const first_argument_type& a) const { return BinaryOp::operator()(a,m_value); } + + template<typename Packet> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const + { return BinaryOp::packetOp(a,internal::pset1<Packet>(m_value)); } + + second_argument_type m_value; +}; +template<typename BinaryOp> struct functor_traits<bind2nd_op<BinaryOp> > : functor_traits<BinaryOp> {}; + + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_BINARY_FUNCTORS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/functors/NullaryFunctors.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/functors/NullaryFunctors.h new file mode 100644 index 000000000..b03be0269 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/functors/NullaryFunctors.h @@ -0,0 +1,188 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2016 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_NULLARY_FUNCTORS_H +#define EIGEN_NULLARY_FUNCTORS_H + +namespace Eigen { + +namespace internal { + +template<typename Scalar> +struct scalar_constant_op { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_constant_op(const scalar_constant_op& other) : m_other(other.m_other) { } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_constant_op(const Scalar& other) : m_other(other) { } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() () const { return m_other; } + template<typename PacketType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const PacketType packetOp() const { return internal::pset1<PacketType>(m_other); } + const Scalar m_other; +}; +template<typename Scalar> +struct functor_traits<scalar_constant_op<Scalar> > +{ enum { Cost = 0 /* as the constant value should be loaded in register only once for the whole expression */, + PacketAccess = packet_traits<Scalar>::Vectorizable, IsRepeatable = true }; }; + +template<typename Scalar> struct scalar_identity_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_identity_op) + template<typename IndexType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (IndexType row, IndexType col) const { return row==col ? Scalar(1) : Scalar(0); } +}; +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, typename Packet> +struct linspaced_op_impl<Scalar,Packet,/*IsInteger*/false> +{ + 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_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> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(IndexType i) const + { + // Principle: + // [low, ..., low] + ( [step, ..., step] * ( [i, ..., i] + [0, ..., size] ) ) + if(m_flip) + { + Packet pi = plset<Packet>(Scalar(i-m_size1)); + Packet res = padd(pset1<Packet>(m_high), pmul(pset1<Packet>(m_step), pi)); + if(i==0) + res = pinsertfirst(res, m_low); + return res; + } + else + { + Packet pi = plset<Packet>(Scalar(i)); + Packet res = padd(pset1<Packet>(m_low), pmul(pset1<Packet>(m_step), pi)); + if(i==m_size1-unpacket_traits<Packet>::size+1) + res = pinsertlast(res, m_high); + return res; + } + } + + const Scalar m_low; + const Scalar m_high; + const Index m_size1; + const Scalar m_step; + const bool m_flip; +}; + +template <typename Scalar, typename Packet> +struct linspaced_op_impl<Scalar,Packet,/*IsInteger*/true> +{ + linspaced_op_impl(const Scalar& low, const Scalar& high, Index num_steps) : + m_low(low), + m_multiplier((high-low)/convert_index<Scalar>(num_steps<=1 ? 1 : num_steps-1)), + m_divisor(convert_index<Scalar>((high>=low?num_steps:-num_steps)+(high-low))/((numext::abs(high-low)+1)==0?1:(numext::abs(high-low)+1))), + m_use_divisor(num_steps>1 && (numext::abs(high-low)+1)<num_steps) + {} + + template<typename IndexType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const Scalar operator() (IndexType i) const + { + if(m_use_divisor) return m_low + convert_index<Scalar>(i)/m_divisor; + else return m_low + convert_index<Scalar>(i)*m_multiplier; + } + + const Scalar m_low; + const Scalar m_multiplier; + const Scalar m_divisor; + const bool m_use_divisor; +}; + +// ----- Linspace functor ---------------------------------------------------------------- + +// 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> > +{ + enum + { + Cost = 1, + PacketAccess = (!NumTraits<Scalar>::IsInteger) && packet_traits<Scalar>::HasSetLinear && packet_traits<Scalar>::HasBlend, + /*&& ((!NumTraits<Scalar>::IsInteger) || packet_traits<Scalar>::HasDiv),*/ // <- vectorization for integer is currently disabled + IsRepeatable = true + }; +}; +template <typename Scalar, typename PacketType> struct linspaced_op +{ + linspaced_op(const Scalar& low, const Scalar& high, Index num_steps) + : impl((num_steps==1 ? high : low),high,num_steps) + {} + + template<typename IndexType> + 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); } + + // 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; +}; + +// Linear access is automatically determined from the operator() prototypes available for the given functor. +// If it exposes an operator()(i,j), then we assume the i and j coefficients are required independently +// and linear access is not possible. In all other cases, linear access is enabled. +// Users should not have to deal with this structure. +template<typename Functor> struct functor_has_linear_access { enum { ret = !has_binary_operator<Functor>::value }; }; + +// For unreliable compilers, let's specialize the has_*ary_operator +// helpers so that at least built-in nullary functors work fine. +#if !( (EIGEN_COMP_MSVC>1600) || (EIGEN_GNUC_AT_LEAST(4,8)) || (EIGEN_COMP_ICC>=1600)) +template<typename Scalar,typename IndexType> +struct has_nullary_operator<scalar_constant_op<Scalar>,IndexType> { enum { value = 1}; }; +template<typename Scalar,typename IndexType> +struct has_unary_operator<scalar_constant_op<Scalar>,IndexType> { enum { value = 0}; }; +template<typename Scalar,typename IndexType> +struct has_binary_operator<scalar_constant_op<Scalar>,IndexType> { enum { value = 0}; }; + +template<typename Scalar,typename IndexType> +struct has_nullary_operator<scalar_identity_op<Scalar>,IndexType> { enum { value = 0}; }; +template<typename Scalar,typename IndexType> +struct has_unary_operator<scalar_identity_op<Scalar>,IndexType> { enum { value = 0}; }; +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<scalar_random_op<Scalar>,IndexType> { enum { value = 1}; }; +template<typename Scalar,typename IndexType> +struct has_unary_operator<scalar_random_op<Scalar>,IndexType> { enum { value = 0}; }; +template<typename Scalar,typename IndexType> +struct has_binary_operator<scalar_random_op<Scalar>,IndexType> { enum { value = 0}; }; +#endif + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_NULLARY_FUNCTORS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/functors/StlFunctors.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/functors/StlFunctors.h new file mode 100644 index 000000000..6df3fa501 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/functors/StlFunctors.h @@ -0,0 +1,132 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2010 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_STL_FUNCTORS_H +#define EIGEN_STL_FUNCTORS_H + +namespace Eigen { + +namespace internal { + +// default functor traits for STL functors: + +template<typename T> +struct functor_traits<std::multiplies<T> > +{ enum { Cost = NumTraits<T>::MulCost, PacketAccess = false }; }; + +template<typename T> +struct functor_traits<std::divides<T> > +{ enum { Cost = NumTraits<T>::MulCost, PacketAccess = false }; }; + +template<typename T> +struct functor_traits<std::plus<T> > +{ enum { Cost = NumTraits<T>::AddCost, PacketAccess = false }; }; + +template<typename T> +struct functor_traits<std::minus<T> > +{ enum { Cost = NumTraits<T>::AddCost, PacketAccess = false }; }; + +template<typename T> +struct functor_traits<std::negate<T> > +{ enum { Cost = NumTraits<T>::AddCost, PacketAccess = false }; }; + +template<typename T> +struct functor_traits<std::logical_or<T> > +{ enum { Cost = 1, PacketAccess = false }; }; + +template<typename T> +struct functor_traits<std::logical_and<T> > +{ enum { Cost = 1, PacketAccess = false }; }; + +template<typename T> +struct functor_traits<std::logical_not<T> > +{ enum { Cost = 1, PacketAccess = false }; }; + +template<typename T> +struct functor_traits<std::greater<T> > +{ enum { Cost = 1, PacketAccess = false }; }; + +template<typename T> +struct functor_traits<std::less<T> > +{ enum { Cost = 1, PacketAccess = false }; }; + +template<typename T> +struct functor_traits<std::greater_equal<T> > +{ enum { Cost = 1, PacketAccess = false }; }; + +template<typename T> +struct functor_traits<std::less_equal<T> > +{ enum { Cost = 1, PacketAccess = false }; }; + +template<typename T> +struct functor_traits<std::equal_to<T> > +{ enum { Cost = 1, PacketAccess = false }; }; + +template<typename T> +struct functor_traits<std::not_equal_to<T> > +{ enum { Cost = 1, PacketAccess = false }; }; + +#if (__cplusplus < 201103L) && (EIGEN_COMP_MSVC <= 1900) +// std::binder* are deprecated since c++11 and will be removed in c++17 +template<typename T> +struct functor_traits<std::binder2nd<T> > +{ enum { Cost = functor_traits<T>::Cost, PacketAccess = false }; }; + +template<typename T> +struct functor_traits<std::binder1st<T> > +{ enum { Cost = functor_traits<T>::Cost, PacketAccess = false }; }; +#endif + +template<typename T> +struct functor_traits<std::unary_negate<T> > +{ enum { Cost = 1 + functor_traits<T>::Cost, PacketAccess = false }; }; + +template<typename T> +struct functor_traits<std::binary_negate<T> > +{ enum { Cost = 1 + functor_traits<T>::Cost, PacketAccess = false }; }; + +#ifdef EIGEN_STDEXT_SUPPORT + +template<typename T0,typename T1> +struct functor_traits<std::project1st<T0,T1> > +{ enum { Cost = 0, PacketAccess = false }; }; + +template<typename T0,typename T1> +struct functor_traits<std::project2nd<T0,T1> > +{ enum { Cost = 0, PacketAccess = false }; }; + +template<typename T0,typename T1> +struct functor_traits<std::select2nd<std::pair<T0,T1> > > +{ enum { Cost = 0, PacketAccess = false }; }; + +template<typename T0,typename T1> +struct functor_traits<std::select1st<std::pair<T0,T1> > > +{ enum { Cost = 0, PacketAccess = false }; }; + +template<typename T0,typename T1> +struct functor_traits<std::unary_compose<T0,T1> > +{ enum { Cost = functor_traits<T0>::Cost + functor_traits<T1>::Cost, PacketAccess = false }; }; + +template<typename T0,typename T1,typename T2> +struct functor_traits<std::binary_compose<T0,T1,T2> > +{ enum { Cost = functor_traits<T0>::Cost + functor_traits<T1>::Cost + functor_traits<T2>::Cost, PacketAccess = false }; }; + +#endif // EIGEN_STDEXT_SUPPORT + +// allow to add new functors and specializations of functor_traits from outside Eigen. +// this macro is really needed because functor_traits must be specialized after it is declared but before it is used... +#ifdef EIGEN_FUNCTORS_PLUGIN +#include EIGEN_FUNCTORS_PLUGIN +#endif + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_STL_FUNCTORS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/functors/TernaryFunctors.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/functors/TernaryFunctors.h new file mode 100644 index 000000000..b254e96c6 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/functors/TernaryFunctors.h @@ -0,0 +1,25 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2016 Eugene Brevdo <ebrevdo@gmail.com> +// +// 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_TERNARY_FUNCTORS_H +#define EIGEN_TERNARY_FUNCTORS_H + +namespace Eigen { + +namespace internal { + +//---------- associative ternary functors ---------- + + + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_TERNARY_FUNCTORS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/functors/UnaryFunctors.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/functors/UnaryFunctors.h new file mode 100644 index 000000000..581a3c93a --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/functors/UnaryFunctors.h @@ -0,0 +1,823 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2016 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_UNARY_FUNCTORS_H +#define EIGEN_UNARY_FUNCTORS_H + +namespace Eigen { + +namespace internal { + +// Copied from unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h +// TODO: remove or update after upstream +/** \internal + * \brief Template functor to compute the sigmoid of a scalar + * \sa class CwiseUnaryOp, ArrayBase::sigmoid() + */ +template <typename T> +struct scalar_sigmoid_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_sigmoid_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T operator()(const T& x) const { + const T one = T(1); + return one / (one + numext::exp(-x)); + } + + template <typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Packet packetOp(const Packet& x) const { + const Packet one = pset1<Packet>(T(1)); + return pdiv(one, padd(one, pexp(pnegate(x)))); + } +}; + +template <typename T> +struct functor_traits<scalar_sigmoid_op<T> > { + enum { + Cost = NumTraits<T>::AddCost * 2 + NumTraits<T>::MulCost * 6, + PacketAccess = packet_traits<T>::HasAdd && packet_traits<T>::HasDiv && + packet_traits<T>::HasNegate && packet_traits<T>::HasExp + }; +}; + + +/** \internal + * \brief Template functor to compute the opposite of a scalar + * + * \sa class CwiseUnaryOp, MatrixBase::operator- + */ +template<typename Scalar> struct scalar_opposite_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_opposite_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { return -a; } + template<typename Packet> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const + { return internal::pnegate(a); } +}; +template<typename Scalar> +struct functor_traits<scalar_opposite_op<Scalar> > +{ enum { + Cost = NumTraits<Scalar>::AddCost, + PacketAccess = packet_traits<Scalar>::HasNegate }; +}; + +/** \internal + * \brief Template functor to compute the absolute value of a scalar + * + * \sa class CwiseUnaryOp, Cwise::abs + */ +template<typename Scalar> struct scalar_abs_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_abs_op) + typedef typename NumTraits<Scalar>::Real result_type; + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a) const { return numext::abs(a); } + template<typename Packet> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const + { return internal::pabs(a); } +}; +template<typename Scalar> +struct functor_traits<scalar_abs_op<Scalar> > +{ + enum { + Cost = NumTraits<Scalar>::AddCost, + PacketAccess = packet_traits<Scalar>::HasAbs + }; +}; + +/** \internal + * \brief Template functor to compute the score of a scalar, to chose a pivot + * + * \sa class CwiseUnaryOp + */ +template<typename Scalar> struct scalar_score_coeff_op : scalar_abs_op<Scalar> +{ + typedef void Score_is_abs; +}; +template<typename Scalar> +struct functor_traits<scalar_score_coeff_op<Scalar> > : functor_traits<scalar_abs_op<Scalar> > {}; + +/* Avoid recomputing abs when we know the score and they are the same. Not a true Eigen functor. */ +template<typename Scalar, typename=void> struct abs_knowing_score +{ + EIGEN_EMPTY_STRUCT_CTOR(abs_knowing_score) + typedef typename NumTraits<Scalar>::Real result_type; + template<typename Score> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a, const Score&) const { return numext::abs(a); } +}; +template<typename Scalar> struct abs_knowing_score<Scalar, typename scalar_score_coeff_op<Scalar>::Score_is_abs> +{ + EIGEN_EMPTY_STRUCT_CTOR(abs_knowing_score) + typedef typename NumTraits<Scalar>::Real result_type; + template<typename Scal> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const Scal&, const result_type& a) const { return a; } +}; + +/** \internal + * \brief Template functor to compute the squared absolute value of a scalar + * + * \sa class CwiseUnaryOp, Cwise::abs2 + */ +template<typename Scalar> struct scalar_abs2_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_abs2_op) + typedef typename NumTraits<Scalar>::Real result_type; + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a) const { return numext::abs2(a); } + template<typename Packet> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const + { return internal::pmul(a,a); } +}; +template<typename Scalar> +struct functor_traits<scalar_abs2_op<Scalar> > +{ enum { Cost = NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasAbs2 }; }; + +/** \internal + * \brief Template functor to compute the conjugate of a complex value + * + * \sa class CwiseUnaryOp, MatrixBase::conjugate() + */ +template<typename Scalar> struct scalar_conjugate_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_conjugate_op) + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { using numext::conj; return conj(a); } + template<typename Packet> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const { return internal::pconj(a); } +}; +template<typename Scalar> +struct functor_traits<scalar_conjugate_op<Scalar> > +{ + enum { + Cost = NumTraits<Scalar>::IsComplex ? NumTraits<Scalar>::AddCost : 0, + PacketAccess = packet_traits<Scalar>::HasConj + }; +}; + +/** \internal + * \brief Template functor to compute the phase angle of a complex + * + * \sa class CwiseUnaryOp, Cwise::arg + */ +template<typename Scalar> struct scalar_arg_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_arg_op) + typedef typename NumTraits<Scalar>::Real result_type; + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a) const { using numext::arg; return arg(a); } + template<typename Packet> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const + { return internal::parg(a); } +}; +template<typename Scalar> +struct functor_traits<scalar_arg_op<Scalar> > +{ + enum { + Cost = NumTraits<Scalar>::IsComplex ? 5 * NumTraits<Scalar>::MulCost : NumTraits<Scalar>::AddCost, + PacketAccess = packet_traits<Scalar>::HasArg + }; +}; +/** \internal + * \brief Template functor to cast a scalar to another type + * + * \sa class CwiseUnaryOp, MatrixBase::cast() + */ +template<typename Scalar, typename NewType> +struct scalar_cast_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op) + typedef NewType result_type; + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const NewType operator() (const Scalar& a) const { return cast<Scalar, NewType>(a); } +}; +template<typename Scalar, typename NewType> +struct functor_traits<scalar_cast_op<Scalar,NewType> > +{ enum { Cost = is_same<Scalar, NewType>::value ? 0 : NumTraits<NewType>::AddCost, PacketAccess = false }; }; + +/** \internal + * \brief Template functor to extract the real part of a complex + * + * \sa class CwiseUnaryOp, MatrixBase::real() + */ +template<typename Scalar> +struct scalar_real_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_real_op) + typedef typename NumTraits<Scalar>::Real result_type; + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const { return numext::real(a); } +}; +template<typename Scalar> +struct functor_traits<scalar_real_op<Scalar> > +{ enum { Cost = 0, PacketAccess = false }; }; + +/** \internal + * \brief Template functor to extract the imaginary part of a complex + * + * \sa class CwiseUnaryOp, MatrixBase::imag() + */ +template<typename Scalar> +struct scalar_imag_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_imag_op) + typedef typename NumTraits<Scalar>::Real result_type; + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const { return numext::imag(a); } +}; +template<typename Scalar> +struct functor_traits<scalar_imag_op<Scalar> > +{ enum { Cost = 0, PacketAccess = false }; }; + +/** \internal + * \brief Template functor to extract the real part of a complex as a reference + * + * \sa class CwiseUnaryOp, MatrixBase::real() + */ +template<typename Scalar> +struct scalar_real_ref_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_real_ref_op) + typedef typename NumTraits<Scalar>::Real result_type; + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE result_type& operator() (const Scalar& a) const { return numext::real_ref(*const_cast<Scalar*>(&a)); } +}; +template<typename Scalar> +struct functor_traits<scalar_real_ref_op<Scalar> > +{ enum { Cost = 0, PacketAccess = false }; }; + +/** \internal + * \brief Template functor to extract the imaginary part of a complex as a reference + * + * \sa class CwiseUnaryOp, MatrixBase::imag() + */ +template<typename Scalar> +struct scalar_imag_ref_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_imag_ref_op) + typedef typename NumTraits<Scalar>::Real result_type; + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE result_type& operator() (const Scalar& a) const { return numext::imag_ref(*const_cast<Scalar*>(&a)); } +}; +template<typename Scalar> +struct functor_traits<scalar_imag_ref_op<Scalar> > +{ enum { Cost = 0, PacketAccess = false }; }; + +/** \internal + * + * \brief Template functor to compute the exponential of a scalar + * + * \sa class CwiseUnaryOp, Cwise::exp() + */ +template<typename Scalar> struct scalar_exp_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_exp_op) + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::exp(a); } + template <typename Packet> + EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pexp(a); } +}; +template <typename Scalar> +struct functor_traits<scalar_exp_op<Scalar> > { + enum { + PacketAccess = packet_traits<Scalar>::HasExp, + // The following numbers are based on the AVX implementation. +#ifdef EIGEN_VECTORIZE_FMA + // Haswell can issue 2 add/mul/madd per cycle. + Cost = + (sizeof(Scalar) == 4 + // float: 8 pmadd, 4 pmul, 2 padd/psub, 6 other + ? (8 * NumTraits<Scalar>::AddCost + 6 * NumTraits<Scalar>::MulCost) + // double: 7 pmadd, 5 pmul, 3 padd/psub, 1 div, 13 other + : (14 * NumTraits<Scalar>::AddCost + + 6 * NumTraits<Scalar>::MulCost + + scalar_div_cost<Scalar,packet_traits<Scalar>::HasDiv>::value)) +#else + Cost = + (sizeof(Scalar) == 4 + // float: 7 pmadd, 6 pmul, 4 padd/psub, 10 other + ? (21 * NumTraits<Scalar>::AddCost + 13 * NumTraits<Scalar>::MulCost) + // double: 7 pmadd, 5 pmul, 3 padd/psub, 1 div, 13 other + : (23 * NumTraits<Scalar>::AddCost + + 12 * NumTraits<Scalar>::MulCost + + scalar_div_cost<Scalar,packet_traits<Scalar>::HasDiv>::value)) +#endif + }; +}; + +/** \internal + * + * \brief Template functor to compute the logarithm of a scalar + * + * \sa class CwiseUnaryOp, ArrayBase::log() + */ +template<typename Scalar> struct scalar_log_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_log_op) + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::log(a); } + template <typename Packet> + EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::plog(a); } +}; +template <typename Scalar> +struct functor_traits<scalar_log_op<Scalar> > { + enum { + PacketAccess = packet_traits<Scalar>::HasLog, + Cost = + (PacketAccess + // The following numbers are based on the AVX implementation. +#ifdef EIGEN_VECTORIZE_FMA + // 8 pmadd, 6 pmul, 8 padd/psub, 16 other, can issue 2 add/mul/madd per cycle. + ? (20 * NumTraits<Scalar>::AddCost + 7 * NumTraits<Scalar>::MulCost) +#else + // 8 pmadd, 6 pmul, 8 padd/psub, 20 other + ? (36 * NumTraits<Scalar>::AddCost + 14 * NumTraits<Scalar>::MulCost) +#endif + // Measured cost of std::log. + : sizeof(Scalar)==4 ? 40 : 85) + }; +}; + +/** \internal + * + * \brief Template functor to compute the logarithm of 1 plus a scalar value + * + * \sa class CwiseUnaryOp, ArrayBase::log1p() + */ +template<typename Scalar> struct scalar_log1p_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_log1p_op) + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::log1p(a); } + template <typename Packet> + EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::plog1p(a); } +}; +template <typename Scalar> +struct functor_traits<scalar_log1p_op<Scalar> > { + enum { + PacketAccess = packet_traits<Scalar>::HasLog1p, + Cost = functor_traits<scalar_log_op<Scalar> >::Cost // TODO measure cost of log1p + }; +}; + +/** \internal + * + * \brief Template functor to compute the base-10 logarithm of a scalar + * + * \sa class CwiseUnaryOp, Cwise::log10() + */ +template<typename Scalar> struct scalar_log10_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_log10_op) + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { EIGEN_USING_STD_MATH(log10) return log10(a); } + template <typename Packet> + EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::plog10(a); } +}; +template<typename Scalar> +struct functor_traits<scalar_log10_op<Scalar> > +{ enum { Cost = 5 * NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasLog10 }; }; + +/** \internal + * \brief Template functor to compute the square root of a scalar + * \sa class CwiseUnaryOp, Cwise::sqrt() + */ +template<typename Scalar> struct scalar_sqrt_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_sqrt_op) + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::sqrt(a); } + template <typename Packet> + EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::psqrt(a); } +}; +template <typename Scalar> +struct functor_traits<scalar_sqrt_op<Scalar> > { + enum { +#if EIGEN_FAST_MATH + // The following numbers are based on the AVX implementation. + Cost = (sizeof(Scalar) == 8 ? 28 + // 4 pmul, 1 pmadd, 3 other + : (3 * NumTraits<Scalar>::AddCost + + 5 * NumTraits<Scalar>::MulCost)), +#else + // The following numbers are based on min VSQRT throughput on Haswell. + Cost = (sizeof(Scalar) == 8 ? 28 : 14), +#endif + PacketAccess = packet_traits<Scalar>::HasSqrt + }; +}; + +/** \internal + * \brief Template functor to compute the reciprocal square root of a scalar + * \sa class CwiseUnaryOp, Cwise::rsqrt() + */ +template<typename Scalar> struct scalar_rsqrt_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_rsqrt_op) + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return Scalar(1)/numext::sqrt(a); } + template <typename Packet> + EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::prsqrt(a); } +}; + +template<typename Scalar> +struct functor_traits<scalar_rsqrt_op<Scalar> > +{ enum { + Cost = 5 * NumTraits<Scalar>::MulCost, + PacketAccess = packet_traits<Scalar>::HasRsqrt + }; +}; + +/** \internal + * \brief Template functor to compute the cosine of a scalar + * \sa class CwiseUnaryOp, ArrayBase::cos() + */ +template<typename Scalar> struct scalar_cos_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_cos_op) + EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return numext::cos(a); } + template <typename Packet> + EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pcos(a); } +}; +template<typename Scalar> +struct functor_traits<scalar_cos_op<Scalar> > +{ + enum { + Cost = 5 * NumTraits<Scalar>::MulCost, + PacketAccess = packet_traits<Scalar>::HasCos + }; +}; + +/** \internal + * \brief Template functor to compute the sine of a scalar + * \sa class CwiseUnaryOp, ArrayBase::sin() + */ +template<typename Scalar> struct scalar_sin_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_sin_op) + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::sin(a); } + template <typename Packet> + EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::psin(a); } +}; +template<typename Scalar> +struct functor_traits<scalar_sin_op<Scalar> > +{ + enum { + Cost = 5 * NumTraits<Scalar>::MulCost, + PacketAccess = packet_traits<Scalar>::HasSin + }; +}; + + +/** \internal + * \brief Template functor to compute the tan of a scalar + * \sa class CwiseUnaryOp, ArrayBase::tan() + */ +template<typename Scalar> struct scalar_tan_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_tan_op) + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::tan(a); } + template <typename Packet> + EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::ptan(a); } +}; +template<typename Scalar> +struct functor_traits<scalar_tan_op<Scalar> > +{ + enum { + Cost = 5 * NumTraits<Scalar>::MulCost, + PacketAccess = packet_traits<Scalar>::HasTan + }; +}; + +/** \internal + * \brief Template functor to compute the arc cosine of a scalar + * \sa class CwiseUnaryOp, ArrayBase::acos() + */ +template<typename Scalar> struct scalar_acos_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_acos_op) + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::acos(a); } + template <typename Packet> + EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pacos(a); } +}; +template<typename Scalar> +struct functor_traits<scalar_acos_op<Scalar> > +{ + enum { + Cost = 5 * NumTraits<Scalar>::MulCost, + PacketAccess = packet_traits<Scalar>::HasACos + }; +}; + +/** \internal + * \brief Template functor to compute the arc sine of a scalar + * \sa class CwiseUnaryOp, ArrayBase::asin() + */ +template<typename Scalar> struct scalar_asin_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_asin_op) + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::asin(a); } + template <typename Packet> + EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pasin(a); } +}; +template<typename Scalar> +struct functor_traits<scalar_asin_op<Scalar> > +{ + enum { + Cost = 5 * NumTraits<Scalar>::MulCost, + PacketAccess = packet_traits<Scalar>::HasASin + }; +}; + + +/** \internal + * \brief Template functor to compute the atan of a scalar + * \sa class CwiseUnaryOp, ArrayBase::atan() + */ +template<typename Scalar> struct scalar_atan_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_atan_op) + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::atan(a); } + template <typename Packet> + EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::patan(a); } +}; +template<typename Scalar> +struct functor_traits<scalar_atan_op<Scalar> > +{ + enum { + Cost = 5 * NumTraits<Scalar>::MulCost, + PacketAccess = packet_traits<Scalar>::HasATan + }; +}; + +/** \internal + * \brief Template functor to compute the tanh of a scalar + * \sa class CwiseUnaryOp, ArrayBase::tanh() + */ +template <typename Scalar> +struct scalar_tanh_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_tanh_op) + EIGEN_DEVICE_FUNC inline const Scalar operator()(const Scalar& a) const { return numext::tanh(a); } + template <typename Packet> + EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& x) const { return ptanh(x); } +}; + +template <typename Scalar> +struct functor_traits<scalar_tanh_op<Scalar> > { + enum { + PacketAccess = packet_traits<Scalar>::HasTanh, + Cost = ( (EIGEN_FAST_MATH && is_same<Scalar,float>::value) +// The following numbers are based on the AVX implementation, +#ifdef EIGEN_VECTORIZE_FMA + // Haswell can issue 2 add/mul/madd per cycle. + // 9 pmadd, 2 pmul, 1 div, 2 other + ? (2 * NumTraits<Scalar>::AddCost + + 6 * NumTraits<Scalar>::MulCost + + scalar_div_cost<Scalar,packet_traits<Scalar>::HasDiv>::value) +#else + ? (11 * NumTraits<Scalar>::AddCost + + 11 * NumTraits<Scalar>::MulCost + + scalar_div_cost<Scalar,packet_traits<Scalar>::HasDiv>::value) +#endif + // This number assumes a naive implementation of tanh + : (6 * NumTraits<Scalar>::AddCost + + 3 * NumTraits<Scalar>::MulCost + + 2 * scalar_div_cost<Scalar,packet_traits<Scalar>::HasDiv>::value + + functor_traits<scalar_exp_op<Scalar> >::Cost)) + }; +}; + +/** \internal + * \brief Template functor to compute the sinh of a scalar + * \sa class CwiseUnaryOp, ArrayBase::sinh() + */ +template<typename Scalar> struct scalar_sinh_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_sinh_op) + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::sinh(a); } + template <typename Packet> + EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::psinh(a); } +}; +template<typename Scalar> +struct functor_traits<scalar_sinh_op<Scalar> > +{ + enum { + Cost = 5 * NumTraits<Scalar>::MulCost, + PacketAccess = packet_traits<Scalar>::HasSinh + }; +}; + +/** \internal + * \brief Template functor to compute the cosh of a scalar + * \sa class CwiseUnaryOp, ArrayBase::cosh() + */ +template<typename Scalar> struct scalar_cosh_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_cosh_op) + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::cosh(a); } + template <typename Packet> + EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pcosh(a); } +}; +template<typename Scalar> +struct functor_traits<scalar_cosh_op<Scalar> > +{ + enum { + Cost = 5 * NumTraits<Scalar>::MulCost, + PacketAccess = packet_traits<Scalar>::HasCosh + }; +}; + +/** \internal + * \brief Template functor to compute the inverse of a scalar + * \sa class CwiseUnaryOp, Cwise::inverse() + */ +template<typename Scalar> +struct scalar_inverse_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_inverse_op) + EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return Scalar(1)/a; } + template<typename Packet> + 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 }; }; + +/** \internal + * \brief Template functor to compute the square of a scalar + * \sa class CwiseUnaryOp, Cwise::square() + */ +template<typename Scalar> +struct scalar_square_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_square_op) + EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return a*a; } + template<typename Packet> + EIGEN_DEVICE_FUNC inline const Packet packetOp(const Packet& a) const + { return internal::pmul(a,a); } +}; +template<typename Scalar> +struct functor_traits<scalar_square_op<Scalar> > +{ enum { Cost = NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasMul }; }; + +/** \internal + * \brief Template functor to compute the cube of a scalar + * \sa class CwiseUnaryOp, Cwise::cube() + */ +template<typename Scalar> +struct scalar_cube_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_cube_op) + EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return a*a*a; } + template<typename Packet> + EIGEN_DEVICE_FUNC inline const Packet packetOp(const Packet& a) const + { return internal::pmul(a,pmul(a,a)); } +}; +template<typename Scalar> +struct functor_traits<scalar_cube_op<Scalar> > +{ enum { Cost = 2*NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasMul }; }; + +/** \internal + * \brief Template functor to compute the rounded value of a scalar + * \sa class CwiseUnaryOp, ArrayBase::round() + */ +template<typename Scalar> struct scalar_round_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_round_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { return numext::round(a); } + template <typename Packet> + EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pround(a); } +}; +template<typename Scalar> +struct functor_traits<scalar_round_op<Scalar> > +{ + enum { + Cost = NumTraits<Scalar>::MulCost, + PacketAccess = packet_traits<Scalar>::HasRound + }; +}; + +/** \internal + * \brief Template functor to compute the floor of a scalar + * \sa class CwiseUnaryOp, ArrayBase::floor() + */ +template<typename Scalar> struct scalar_floor_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_floor_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { return numext::floor(a); } + template <typename Packet> + EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pfloor(a); } +}; +template<typename Scalar> +struct functor_traits<scalar_floor_op<Scalar> > +{ + enum { + Cost = NumTraits<Scalar>::MulCost, + PacketAccess = packet_traits<Scalar>::HasFloor + }; +}; + +/** \internal + * \brief Template functor to compute the ceil of a scalar + * \sa class CwiseUnaryOp, ArrayBase::ceil() + */ +template<typename Scalar> struct scalar_ceil_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_ceil_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { return numext::ceil(a); } + template <typename Packet> + EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pceil(a); } +}; +template<typename Scalar> +struct functor_traits<scalar_ceil_op<Scalar> > +{ + enum { + Cost = NumTraits<Scalar>::MulCost, + PacketAccess = packet_traits<Scalar>::HasCeil + }; +}; + +/** \internal + * \brief Template functor to compute whether a scalar is NaN + * \sa class CwiseUnaryOp, ArrayBase::isnan() + */ +template<typename Scalar> struct scalar_isnan_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_isnan_op) + typedef bool result_type; + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const { return (numext::isnan)(a); } +}; +template<typename Scalar> +struct functor_traits<scalar_isnan_op<Scalar> > +{ + enum { + Cost = NumTraits<Scalar>::MulCost, + PacketAccess = false + }; +}; + +/** \internal + * \brief Template functor to check whether a scalar is +/-inf + * \sa class CwiseUnaryOp, ArrayBase::isinf() + */ +template<typename Scalar> struct scalar_isinf_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_isinf_op) + typedef bool result_type; + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const { return (numext::isinf)(a); } +}; +template<typename Scalar> +struct functor_traits<scalar_isinf_op<Scalar> > +{ + enum { + Cost = NumTraits<Scalar>::MulCost, + PacketAccess = false + }; +}; + +/** \internal + * \brief Template functor to check whether a scalar has a finite value + * \sa class CwiseUnaryOp, ArrayBase::isfinite() + */ +template<typename Scalar> struct scalar_isfinite_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_isfinite_op) + typedef bool result_type; + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const { return (numext::isfinite)(a); } +}; +template<typename Scalar> +struct functor_traits<scalar_isfinite_op<Scalar> > +{ + enum { + Cost = NumTraits<Scalar>::MulCost, + PacketAccess = false + }; +}; + +/** \internal + * \brief Template functor to compute the logical not of a boolean + * + * \sa class CwiseUnaryOp, ArrayBase::operator! + */ +template<typename Scalar> struct scalar_boolean_not_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_boolean_not_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator() (const bool& a) const { return !a; } +}; +template<typename Scalar> +struct functor_traits<scalar_boolean_not_op<Scalar> > { + enum { + Cost = NumTraits<bool>::AddCost, + PacketAccess = false + }; +}; + +/** \internal + * \brief Template functor to compute the signum of a scalar + * \sa class CwiseUnaryOp, Cwise::sign() + */ +template<typename Scalar,bool iscpx=(NumTraits<Scalar>::IsComplex!=0) > struct scalar_sign_op; +template<typename Scalar> +struct scalar_sign_op<Scalar,false> { + EIGEN_EMPTY_STRUCT_CTOR(scalar_sign_op) + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const + { + return Scalar( (a>Scalar(0)) - (a<Scalar(0)) ); + } + //TODO + //template <typename Packet> + //EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::psign(a); } +}; +template<typename Scalar> +struct scalar_sign_op<Scalar,true> { + EIGEN_EMPTY_STRUCT_CTOR(scalar_sign_op) + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const + { + typedef typename NumTraits<Scalar>::Real real_type; + real_type aa = numext::abs(a); + if (aa==real_type(0)) + return Scalar(0); + aa = real_type(1)/aa; + return Scalar(real(a)*aa, imag(a)*aa ); + } + //TODO + //template <typename Packet> + //EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::psign(a); } +}; +template<typename Scalar> +struct functor_traits<scalar_sign_op<Scalar> > +{ enum { + Cost = + NumTraits<Scalar>::IsComplex + ? ( 8*NumTraits<Scalar>::MulCost ) // roughly + : ( 3*NumTraits<Scalar>::AddCost), + PacketAccess = packet_traits<Scalar>::HasSign + }; +}; + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_FUNCTORS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/GeneralBlockPanelKernel.h new file mode 100644 index 000000000..45230bce5 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -0,0 +1,2149 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2009 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_GENERAL_BLOCK_PANEL_H +#define EIGEN_GENERAL_BLOCK_PANEL_H + + +namespace Eigen { + +namespace internal { + +template<typename _LhsScalar, typename _RhsScalar, bool _ConjLhs=false, bool _ConjRhs=false> +class gebp_traits; + + +/** \internal \returns b if a<=0, and returns a otherwise. */ +inline std::ptrdiff_t manage_caching_sizes_helper(std::ptrdiff_t a, std::ptrdiff_t b) +{ + return a<=0 ? b : a; +} + +#if EIGEN_ARCH_i386_OR_x86_64 +const std::ptrdiff_t defaultL1CacheSize = 32*1024; +const std::ptrdiff_t defaultL2CacheSize = 256*1024; +const std::ptrdiff_t defaultL3CacheSize = 2*1024*1024; +#else +const std::ptrdiff_t defaultL1CacheSize = 16*1024; +const std::ptrdiff_t defaultL2CacheSize = 512*1024; +const std::ptrdiff_t defaultL3CacheSize = 512*1024; +#endif + +/** \internal */ +struct CacheSizes { + CacheSizes(): m_l1(-1),m_l2(-1),m_l3(-1) { + int l1CacheSize, l2CacheSize, l3CacheSize; + queryCacheSizes(l1CacheSize, l2CacheSize, l3CacheSize); + m_l1 = manage_caching_sizes_helper(l1CacheSize, defaultL1CacheSize); + m_l2 = manage_caching_sizes_helper(l2CacheSize, defaultL2CacheSize); + m_l3 = manage_caching_sizes_helper(l3CacheSize, defaultL3CacheSize); + } + + std::ptrdiff_t m_l1; + std::ptrdiff_t m_l2; + std::ptrdiff_t m_l3; +}; + + +/** \internal */ +inline void manage_caching_sizes(Action action, std::ptrdiff_t* l1, std::ptrdiff_t* l2, std::ptrdiff_t* l3) +{ + static CacheSizes m_cacheSizes; + + if(action==SetAction) + { + // set the cpu cache size and cache all block sizes from a global cache size in byte + eigen_internal_assert(l1!=0 && l2!=0); + m_cacheSizes.m_l1 = *l1; + m_cacheSizes.m_l2 = *l2; + m_cacheSizes.m_l3 = *l3; + } + else if(action==GetAction) + { + eigen_internal_assert(l1!=0 && l2!=0); + *l1 = m_cacheSizes.m_l1; + *l2 = m_cacheSizes.m_l2; + *l3 = m_cacheSizes.m_l3; + } + else + { + eigen_internal_assert(false); + } +} + +/* Helper for computeProductBlockingSizes. + * + * Given a m x k times k x n matrix product of scalar types \c LhsScalar and \c RhsScalar, + * this function computes the blocking size parameters along the respective dimensions + * for matrix products and related algorithms. The blocking sizes depends on various + * parameters: + * - the L1 and L2 cache sizes, + * - the register level blocking sizes defined by gebp_traits, + * - the number of scalars that fit into a packet (when vectorization is enabled). + * + * \sa setCpuCacheSizes */ + +template<typename LhsScalar, typename RhsScalar, int KcFactor, typename Index> +void evaluateProductBlockingSizesHeuristic(Index& k, Index& m, Index& n, Index num_threads = 1) +{ + typedef gebp_traits<LhsScalar,RhsScalar> Traits; + + // Explanations: + // Let's recall that the product algorithms form mc x kc vertical panels A' on the lhs and + // kc x nc blocks B' on the rhs. B' has to fit into L2/L3 cache. Moreover, A' is processed + // per mr x kc horizontal small panels where mr is the blocking size along the m dimension + // 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); + + if (num_threads > 1) { + typedef typename Traits::ResScalar ResScalar; + enum { + kdiv = KcFactor * (Traits::mr * sizeof(LhsScalar) + Traits::nr * sizeof(RhsScalar)), + ksub = Traits::mr * Traits::nr * sizeof(ResScalar), + kr = 8, + mr = Traits::mr, + nr = Traits::nr + }; + // Increasing k gives us more time to prefetch the content of the "C" + // registers. However once the latency is hidden there is no point in + // increasing the value of k, so we'll cap it at 320 (value determined + // experimentally). + const Index k_cache = (numext::mini<Index>)((l1-ksub)/kdiv, 320); + if (k_cache < k) { + k = k_cache - (k_cache % kr); + eigen_internal_assert(k > 0); + } + + const Index n_cache = (l2-l1) / (nr * sizeof(RhsScalar) * k); + const Index n_per_thread = numext::div_ceil(n, num_threads); + if (n_cache <= n_per_thread) { + // Don't exceed the capacity of the l2 cache. + eigen_internal_assert(n_cache >= static_cast<Index>(nr)); + n = n_cache - (n_cache % nr); + eigen_internal_assert(n > 0); + } else { + n = (numext::mini<Index>)(n, (n_per_thread + nr - 1) - ((n_per_thread + nr - 1) % nr)); + } + + if (l3 > l2) { + // l3 is shared between all cores, so we'll give each thread its own chunk of l3. + const Index m_cache = (l3-l2) / (sizeof(LhsScalar) * k * num_threads); + const Index m_per_thread = numext::div_ceil(m, num_threads); + if(m_cache < m_per_thread && m_cache >= static_cast<Index>(mr)) { + m = m_cache - (m_cache % mr); + eigen_internal_assert(m > 0); + } else { + m = (numext::mini<Index>)(m, (m_per_thread + mr - 1) - ((m_per_thread + mr - 1) % mr)); + } + } + } + else { + // In unit tests we do not want to use extra large matrices, + // so we reduce the cache size to check the blocking strategy is not flawed +#ifdef EIGEN_DEBUG_SMALL_PRODUCT_BLOCKS + l1 = 9*1024; + l2 = 32*1024; + l3 = 512*1024; +#endif + + // Early return for small problems because the computation below are time consuming for small problems. + // Perhaps it would make more sense to consider k*n*m?? + // Note that for very tiny problem, this function should be bypassed anyway + // because we use the coefficient-based implementation for them. + if((numext::maxi)(k,(numext::maxi)(m,n))<48) + return; + + typedef typename Traits::ResScalar ResScalar; + enum { + k_peeling = 8, + k_div = KcFactor * (Traits::mr * sizeof(LhsScalar) + Traits::nr * sizeof(RhsScalar)), + k_sub = Traits::mr * Traits::nr * sizeof(ResScalar) + }; + + // ---- 1st level of blocking on L1, yields kc ---- + + // Blocking on the third dimension (i.e., k) is chosen so that an horizontal panel + // of size mr x kc of the lhs plus a vertical panel of kc x nr of the rhs both fits within L1 cache. + // We also include a register-level block of the result (mx x nr). + // (In an ideal world only the lhs panel would stay in L1) + // Moreover, kc has to be a multiple of 8 to be compatible with loop peeling, leading to a maximum blocking size of: + const Index max_kc = numext::maxi<Index>(((l1-k_sub)/k_div) & (~(k_peeling-1)),1); + const Index old_k = k; + if(k>max_kc) + { + // We are really blocking on the third dimension: + // -> reduce blocking size to make sure the last block is as large as possible + // while keeping the same number of sweeps over the result. + k = (k%max_kc)==0 ? max_kc + : max_kc - k_peeling * ((max_kc-1-(k%max_kc))/(k_peeling*(k/max_kc+1))); + + eigen_internal_assert(((old_k/k) == (old_k/max_kc)) && "the number of sweeps has to remain the same"); + } + + // ---- 2nd level of blocking on max(L2,L3), yields nc ---- + + // TODO find a reliable way to get the actual amount of cache per core to use for 2nd level blocking, that is: + // actual_l2 = max(l2, l3/nb_core_sharing_l3) + // The number below is quite conservative: it is better to underestimate the cache size rather than overestimating it) + // For instance, it corresponds to 6MB of L3 shared among 4 cores. + #ifdef EIGEN_DEBUG_SMALL_PRODUCT_BLOCKS + const Index actual_l2 = l3; + #else + const Index actual_l2 = 1572864; // == 1.5 MB + #endif + + // Here, nc is chosen such that a block of kc x nc of the rhs fit within half of L2. + // The second half is implicitly reserved to access the result and lhs coefficients. + // When k<max_kc, then nc can arbitrarily growth. In practice, it seems to be fruitful + // to limit this growth: we bound nc to growth by a factor x1.5. + // However, if the entire lhs block fit within L1, then we are not going to block on the rows at all, + // and it becomes fruitful to keep the packed rhs blocks in L1 if there is enough remaining space. + Index max_nc; + const Index lhs_bytes = m * k * sizeof(LhsScalar); + const Index remaining_l1 = l1- k_sub - lhs_bytes; + if(remaining_l1 >= Index(Traits::nr*sizeof(RhsScalar))*k) + { + // L1 blocking + max_nc = remaining_l1 / (k*sizeof(RhsScalar)); + } + else + { + // L2 blocking + max_nc = (3*actual_l2)/(2*2*max_kc*sizeof(RhsScalar)); + } + // WARNING Below, we assume that Traits::nr is a power of two. + Index nc = numext::mini<Index>(actual_l2/(2*k*sizeof(RhsScalar)), max_nc) & (~(Traits::nr-1)); + if(n>nc) + { + // We are really blocking over the columns: + // -> reduce blocking size to make sure the last block is as large as possible + // while keeping the same number of sweeps over the packed lhs. + // Here we allow one more sweep if this gives us a perfect match, thus the commented "-1" + n = (n%nc)==0 ? nc + : (nc - Traits::nr * ((nc/*-1*/-(n%nc))/(Traits::nr*(n/nc+1)))); + } + else if(old_k==k) + { + // So far, no blocking at all, i.e., kc==k, and nc==n. + // In this case, let's perform a blocking over the rows such that the packed lhs data is kept in cache L1/L2 + // TODO: part of this blocking strategy is now implemented within the kernel itself, so the L1-based heuristic here should be obsolete. + Index problem_size = k*n*sizeof(LhsScalar); + Index actual_lm = actual_l2; + Index max_mc = m; + if(problem_size<=1024) + { + // problem is small enough to keep in L1 + // Let's choose m such that lhs's block fit in 1/3 of L1 + actual_lm = l1; + } + else if(l3!=0 && problem_size<=32768) + { + // we have both L2 and L3, and problem is small enough to be kept in L2 + // Let's choose m such that lhs's block fit in 1/3 of L2 + actual_lm = l2; + max_mc = (numext::mini<Index>)(576,max_mc); + } + Index mc = (numext::mini<Index>)(actual_lm/(3*k*sizeof(LhsScalar)), max_mc); + if (mc > Traits::mr) mc -= mc % Traits::mr; + else if (mc==0) return; + m = (m%mc)==0 ? mc + : (mc - Traits::mr * ((mc/*-1*/-(m%mc))/(Traits::mr*(m/mc+1)))); + } + } +} + +template <typename Index> +inline bool useSpecificBlockingSizes(Index& k, Index& m, Index& n) +{ +#ifdef EIGEN_TEST_SPECIFIC_BLOCKING_SIZES + if (EIGEN_TEST_SPECIFIC_BLOCKING_SIZES) { + k = numext::mini<Index>(k, EIGEN_TEST_SPECIFIC_BLOCKING_SIZE_K); + m = numext::mini<Index>(m, EIGEN_TEST_SPECIFIC_BLOCKING_SIZE_M); + n = numext::mini<Index>(n, EIGEN_TEST_SPECIFIC_BLOCKING_SIZE_N); + return true; + } +#else + EIGEN_UNUSED_VARIABLE(k) + EIGEN_UNUSED_VARIABLE(m) + EIGEN_UNUSED_VARIABLE(n) +#endif + return false; +} + +/** \brief Computes the blocking parameters for a m x k times k x n matrix product + * + * \param[in,out] k Input: the third dimension of the product. Output: the blocking size along the same dimension. + * \param[in,out] m Input: the number of rows of the left hand side. Output: the blocking size along the same dimension. + * \param[in,out] n Input: the number of columns of the right hand side. Output: the blocking size along the same dimension. + * + * Given a m x k times k x n matrix product of scalar types \c LhsScalar and \c RhsScalar, + * this function computes the blocking size parameters along the respective dimensions + * for matrix products and related algorithms. + * + * The blocking size parameters may be evaluated: + * - either by a heuristic based on cache sizes; + * - or using fixed prescribed values (for testing purposes). + * + * \sa setCpuCacheSizes */ + +template<typename LhsScalar, typename RhsScalar, int KcFactor, typename Index> +void computeProductBlockingSizes(Index& k, Index& m, Index& n, Index num_threads = 1) +{ + if (!useSpecificBlockingSizes(k, m, n)) { + evaluateProductBlockingSizesHeuristic<LhsScalar, RhsScalar, KcFactor, Index>(k, m, n, num_threads); + } +} + +template<typename LhsScalar, typename RhsScalar, typename Index> +inline void computeProductBlockingSizes(Index& k, Index& m, Index& n, Index num_threads = 1) +{ + computeProductBlockingSizes<LhsScalar,RhsScalar,1,Index>(k, m, n, num_threads); +} + +#ifdef EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD + #define CJMADD(CJ,A,B,C,T) C = CJ.pmadd(A,B,C); +#else + + // FIXME (a bit overkill maybe ?) + + template<typename CJ, typename A, typename B, typename C, typename T> struct gebp_madd_selector { + EIGEN_ALWAYS_INLINE static void run(const CJ& cj, A& a, B& b, C& c, T& /*t*/) + { + c = cj.pmadd(a,b,c); + } + }; + + template<typename CJ, typename T> struct gebp_madd_selector<CJ,T,T,T,T> { + EIGEN_ALWAYS_INLINE static void run(const CJ& cj, T& a, T& b, T& c, T& t) + { + t = b; t = cj.pmul(a,t); c = padd(c,t); + } + }; + + template<typename CJ, typename A, typename B, typename C, typename T> + EIGEN_STRONG_INLINE void gebp_madd(const CJ& cj, A& a, B& b, C& c, T& t) + { + gebp_madd_selector<CJ,A,B,C,T>::run(cj,a,b,c,t); + } + + #define CJMADD(CJ,A,B,C,T) gebp_madd(CJ,A,B,C,T); +// #define CJMADD(CJ,A,B,C,T) T = B; T = CJ.pmul(A,T); C = padd(C,T); +#endif + +/* Vectorization logic + * real*real: unpack rhs to constant packets, ... + * + * cd*cd : unpack rhs to (b_r,b_r), (b_i,b_i), mul to get (a_r b_r,a_i b_r) (a_r b_i,a_i b_i), + * storing each res packet into two packets (2x2), + * at the end combine them: swap the second and addsub them + * cf*cf : same but with 2x4 blocks + * 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> +class gebp_traits +{ +public: + typedef _LhsScalar LhsScalar; + typedef _RhsScalar RhsScalar; + typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar; + + 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, + + NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS, + + // register block size along the N direction must be 1 or 4 + nr = 4, + + // 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 + // 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. + mr = Vectorizable ? 3*LhsPacketSize : default_mr, +#else + mr = default_mr, +#endif + + LhsProgress = LhsPacketSize, + 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 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 loadRhsQuad(const RhsScalar* b, RhsPacket& dest) const + { + dest = ploadquad<RhsPacket>(b); + } + + template<typename LhsPacketType> + EIGEN_STRONG_INLINE void loadLhs(const LhsScalar* a, LhsPacketType& dest) const + { + dest = pload<LhsPacketType>(a); + } + + 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> + EIGEN_STRONG_INLINE void madd(const LhsPacketType& a, const RhsPacketType& b, AccPacketType& c, AccPacketType& tmp) const + { + conj_helper<LhsPacketType,RhsPacketType,ConjLhs,ConjRhs> cj; + // It would be a lot cleaner to call pmadd all the time. Unfortunately if we + // let gcc allocate the register in which to store the result of the pmul + // (in the case where there is no FMA) gcc fails to figure out how to avoid + // spilling register. +#ifdef EIGEN_HAS_SINGLE_INSTRUCTION_MADD + EIGEN_UNUSED_VARIABLE(tmp); + c = cj.pmadd(a,b,c); +#else + tmp = b; tmp = cj.pmul(a,tmp); c = padd(c,tmp); +#endif + } + + EIGEN_STRONG_INLINE void acc(const AccPacket& c, const ResPacket& alpha, ResPacket& r) const + { + r = pmadd(c,alpha,r); + } + + template<typename ResPacketHalf> + EIGEN_STRONG_INLINE void acc(const ResPacketHalf& c, const ResPacketHalf& alpha, ResPacketHalf& r) const + { + r = pmadd(c,alpha,r); + } + +}; + +template<typename RealScalar, bool _ConjLhs> +class gebp_traits<std::complex<RealScalar>, RealScalar, _ConjLhs, false> +{ +public: + typedef std::complex<RealScalar> LhsScalar; + typedef RealScalar RhsScalar; + typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar; + + 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, + + NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS, + nr = 4, +#if defined(EIGEN_HAS_SINGLE_INSTRUCTION_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) && !defined(EIGEN_VECTORIZE_VSX) + // we assume 16 registers + mr = 3*LhsPacketSize, +#else + mr = (EIGEN_PLAIN_ENUM_MIN(16,NumberOfRegisters)/2/nr)*LhsPacketSize, +#endif + + LhsProgress = LhsPacketSize, + 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 ResPacket AccPacket; + + EIGEN_STRONG_INLINE void initAcc(AccPacket& p) + { + p = pset1<ResPacket>(ResScalar(0)); + } + + EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacket& dest) const + { + dest = pset1<RhsPacket>(*b); + } + + EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, RhsPacket& dest) const + { + dest = pset1<RhsPacket>(*b); + } + + EIGEN_STRONG_INLINE void loadLhs(const LhsScalar* a, LhsPacket& dest) const + { + dest = pload<LhsPacket>(a); + } + + EIGEN_STRONG_INLINE void loadLhsUnaligned(const LhsScalar* a, LhsPacket& dest) const + { + dest = ploadu<LhsPacket>(a); + } + + 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); +// } + + EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp) 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 + { +#ifdef EIGEN_HAS_SINGLE_INSTRUCTION_MADD + EIGEN_UNUSED_VARIABLE(tmp); + c.v = pmadd(a.v,b,c.v); +#else + tmp = b; tmp = pmul(a.v,tmp); c.v = padd(c.v,tmp); +#endif + } + + EIGEN_STRONG_INLINE void madd_impl(const LhsScalar& a, const RhsScalar& b, ResScalar& c, RhsScalar& /*tmp*/, const false_type&) const + { + c += a * b; + } + + EIGEN_STRONG_INLINE void acc(const AccPacket& c, const ResPacket& alpha, ResPacket& r) const + { + r = cj.pmadd(c,alpha,r); + } + +protected: + conj_helper<ResPacket,ResPacket,ConjLhs,false> cj; +}; + +template<typename Packet> +struct DoublePacket +{ + Packet first; + Packet second; +}; + +template<typename Packet> +DoublePacket<Packet> padd(const DoublePacket<Packet> &a, const DoublePacket<Packet> &b) +{ + DoublePacket<Packet> res; + res.first = padd(a.first, b.first); + res.second = padd(a.second,b.second); + return res; +} + +template<typename Packet> +const DoublePacket<Packet>& predux_downto4(const DoublePacket<Packet> &a) +{ + return a; +} + +template<typename Packet> struct unpacket_traits<DoublePacket<Packet> > { typedef DoublePacket<Packet> half; }; +// template<typename Packet> +// DoublePacket<Packet> pmadd(const DoublePacket<Packet> &a, const DoublePacket<Packet> &b) +// { +// DoublePacket<Packet> res; +// res.first = padd(a.first, b.first); +// res.second = padd(a.second,b.second); +// return res; +// } + +template<typename RealScalar, bool _ConjLhs, bool _ConjRhs> +class gebp_traits<std::complex<RealScalar>, std::complex<RealScalar>, _ConjLhs, _ConjRhs > +{ +public: + typedef std::complex<RealScalar> Scalar; + typedef std::complex<RealScalar> LhsScalar; + typedef std::complex<RealScalar> RhsScalar; + typedef std::complex<RealScalar> ResScalar; + + 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, + + // FIXME: should depend on NumberOfRegisters + nr = 4, + mr = ResPacketSize, + + LhsProgress = ResPacketSize, + RhsProgress = 1 + }; + + typedef typename packet_traits<RealScalar>::type RealPacket; + typedef typename packet_traits<Scalar>::type ScalarPacket; + typedef DoublePacket<RealPacket> DoublePacketType; + + 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; + + EIGEN_STRONG_INLINE void initAcc(Scalar& p) { p = Scalar(0); } + + EIGEN_STRONG_INLINE void initAcc(DoublePacketType& p) + { + p.first = pset1<RealPacket>(RealScalar(0)); + p.second = pset1<RealPacket>(RealScalar(0)); + } + + // Scalar path + EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, ResPacket& dest) const + { + dest = pset1<ResPacket>(*b); + } + + // Vectorized path + EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, DoublePacketType& dest) const + { + dest.first = pset1<RealPacket>(real(*b)); + dest.second = pset1<RealPacket>(imag(*b)); + } + + EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, ResPacket& dest) const + { + loadRhs(b,dest); + } + EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, DoublePacketType& dest) const + { + eigen_internal_assert(unpacket_traits<ScalarPacket>::size<=4); + loadRhs(b,dest); + } + + EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3) + { + // 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); + } + + // Vectorized path + EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, DoublePacketType& b0, DoublePacketType& b1) + { + // FIXME not sure that's the best way to implement it! + loadRhs(b+0, b0); + loadRhs(b+1, b1); + } + + // Scalar path + EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsScalar& b0, RhsScalar& b1) + { + // FIXME not sure that's the best way to implement it! + loadRhs(b+0, b0); + loadRhs(b+1, b1); + } + + // nothing special here + EIGEN_STRONG_INLINE void loadLhs(const LhsScalar* a, LhsPacket& dest) const + { + dest = pload<LhsPacket>((const typename unpacket_traits<LhsPacket>::type*)(a)); + } + + EIGEN_STRONG_INLINE void loadLhsUnaligned(const LhsScalar* a, LhsPacket& dest) const + { + dest = ploadu<LhsPacket>((const typename unpacket_traits<LhsPacket>::type*)(a)); + } + + EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, DoublePacketType& c, RhsPacket& /*tmp*/) 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 + { + c = cj.pmadd(a,b,c); + } + + 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 + { + // assemble c + ResPacket tmp; + if((!ConjLhs)&&(!ConjRhs)) + { + tmp = pcplxflip(pconj(ResPacket(c.second))); + tmp = padd(ResPacket(c.first),tmp); + } + else if((!ConjLhs)&&(ConjRhs)) + { + tmp = pconj(pcplxflip(ResPacket(c.second))); + tmp = padd(ResPacket(c.first),tmp); + } + else if((ConjLhs)&&(!ConjRhs)) + { + tmp = pcplxflip(ResPacket(c.second)); + tmp = padd(pconj(ResPacket(c.first)),tmp); + } + else if((ConjLhs)&&(ConjRhs)) + { + tmp = pcplxflip(ResPacket(c.second)); + tmp = psub(pconj(ResPacket(c.first)),tmp); + } + + r = pmadd(tmp,alpha,r); + } + +protected: + conj_helper<LhsScalar,RhsScalar,ConjLhs,ConjRhs> cj; +}; + +template<typename RealScalar, bool _ConjRhs> +class gebp_traits<RealScalar, std::complex<RealScalar>, false, _ConjRhs > +{ +public: + typedef std::complex<RealScalar> Scalar; + typedef RealScalar LhsScalar; + typedef Scalar RhsScalar; + typedef Scalar ResScalar; + + 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, + + NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS, + // FIXME: should depend on NumberOfRegisters + nr = 4, + mr = (EIGEN_PLAIN_ENUM_MIN(16,NumberOfRegisters)/2/nr)*ResPacketSize, + + LhsProgress = ResPacketSize, + 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 ResPacket AccPacket; + + EIGEN_STRONG_INLINE void initAcc(AccPacket& p) + { + p = pset1<ResPacket>(ResScalar(0)); + } + + EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacket& dest) const + { + dest = pset1<RhsPacket>(*b); + } + + 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) +// { +// // FIXME not sure that's the best way to implement it! +// b0 = pload1<RhsPacket>(b+0); +// b1 = pload1<RhsPacket>(b+1); +// } + + EIGEN_STRONG_INLINE void loadLhs(const LhsScalar* a, LhsPacket& dest) const + { + dest = ploaddup<LhsPacket>(a); + } + + EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, RhsPacket& dest) const + { + eigen_internal_assert(unpacket_traits<RhsPacket>::size<=4); + loadRhs(b,dest); + } + + EIGEN_STRONG_INLINE void loadLhsUnaligned(const LhsScalar* a, LhsPacket& dest) const + { + dest = ploaddup<LhsPacket>(a); + } + + EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp) 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 + { +#ifdef EIGEN_HAS_SINGLE_INSTRUCTION_MADD + EIGEN_UNUSED_VARIABLE(tmp); + c.v = pmadd(a,b.v,c.v); +#else + tmp = b; tmp.v = pmul(a,tmp.v); c = padd(c,tmp); +#endif + + } + + EIGEN_STRONG_INLINE void madd_impl(const LhsScalar& a, const RhsScalar& b, ResScalar& c, RhsScalar& /*tmp*/, const false_type&) const + { + c += a * b; + } + + EIGEN_STRONG_INLINE void acc(const AccPacket& c, const ResPacket& alpha, ResPacket& r) const + { + r = cj.pmadd(alpha,c,r); + } + +protected: + conj_helper<ResPacket,ResPacket,false,ConjRhs> cj; +}; + +/* optimized GEneral packed Block * packed Panel product kernel + * + * Mixing type logic: C += A * B + * | A | B | comments + * |real |cplx | no vectorization yet, would require to pack A with duplication + * |cplx |real | easy vectorization + */ +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 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 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 DataMapper::LinearMapper LinearMapper; + + enum { + Vectorizable = Traits::Vectorizable, + LhsProgress = Traits::LhsProgress, + RhsProgress = Traits::RhsProgress, + ResPacketSize = Traits::ResPacketSize + }; + + EIGEN_DONT_INLINE + void operator()(const DataMapper& res, const LhsScalar* blockA, const RhsScalar* blockB, + Index rows, Index depth, Index cols, ResScalar alpha, + Index strideA=-1, Index strideB=-1, Index offsetA=0, Index offsetB=0); +}; + +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> + ::operator()(const DataMapper& res, const LhsScalar* blockA, const RhsScalar* blockB, + Index rows, Index depth, Index cols, ResScalar alpha, + Index strideA, Index strideB, Index offsetA, Index offsetB) + { + Traits traits; + SwappedTraits straits; + + if(strideA==-1) strideA = depth; + if(strideB==-1) strideB = depth; + conj_helper<LhsScalar,RhsScalar,ConjugateLhs,ConjugateRhs> cj; + 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; + 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 Index depth2 = depth & ~1; + + //---------- Process 3 * LhsProgress rows at once ---------- + // This corresponds to 3*LhsProgress x nr register blocks. + // Usually, make sense only with FMA + if(mr>=3*Traits::LhsProgress) + { + // Here, the general idea is to loop on each largest micro horizontal panel of the lhs (3*Traits::LhsProgress x depth) + // and on each largest micro vertical panel of the rhs (depth * nr). + // Blocking sizes, i.e., 'depth' has been computed so that the micro horizontal panel of the lhs fit in L1. + // However, if depth is too small, we can extend the number of rows of these horizontal panels. + // This actual number of rows is computed as follow: + const Index l1 = defaultL1CacheSize; // in Bytes, TODO, l1 should be passed to this function. + // The max(1, ...) here is needed because we may be using blocking params larger than what our known l1 cache size + // suggests we should be using: either because our known l1 cache size is inaccurate (e.g. on Android, we can only guess), + // or because we are testing specific blocking sizes. + const Index actual_panel_rows = (3*LhsProgress) * std::max<Index>(1,( (l1 - sizeof(ResScalar)*mr*nr - depth*nr*sizeof(RhsScalar)) / (depth * sizeof(LhsScalar) * 3*LhsProgress) )); + for(Index i1=0; i1<peeled_mc3; i1+=actual_panel_rows) + { + const Index actual_panel_end = (std::min)(i1+actual_panel_rows, peeled_mc3); + for(Index j2=0; j2<packet_cols4; j2+=nr) + { + for(Index i=i1; i<actual_panel_end; i+=3*LhsProgress) + { + + // We selected a 3*Traits::LhsProgress x nr micro block of res which is entirely + // stored into 3 x nr registers. + + const LhsScalar* blA = &blockA[i*strideA+offsetA*(3*LhsProgress)]; + prefetch(&blA[0]); + + // gets res block as register + AccPacket C0, C1, C2, C3, + C4, C5, C6, C7, + C8, C9, C10, C11; + traits.initAcc(C0); traits.initAcc(C1); traits.initAcc(C2); traits.initAcc(C3); + traits.initAcc(C4); traits.initAcc(C5); traits.initAcc(C6); traits.initAcc(C7); + traits.initAcc(C8); traits.initAcc(C9); traits.initAcc(C10); traits.initAcc(C11); + + 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(0); + r1.prefetch(0); + r2.prefetch(0); + r3.prefetch(0); + + // 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 3pX4"); + RhsPacket B_0, T0; + LhsPacket A2; + +#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) { 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(blB); + EIGEN_GEBP_ONESTEP(0); + EIGEN_GEBP_ONESTEP(1); + EIGEN_GEBP_ONESTEP(2); + EIGEN_GEBP_ONESTEP(3); + EIGEN_GEBP_ONESTEP(4); + EIGEN_GEBP_ONESTEP(5); + EIGEN_GEBP_ONESTEP(6); + EIGEN_GEBP_ONESTEP(7); + + blB += pk*4*RhsProgress; + blA += pk*3*Traits::LhsProgress; + + EIGEN_ASM_COMMENT("end gebp micro kernel 3pX4"); + } + // process remaining peeled loop + for(Index k=peeled_kc; k<depth; k++) + { + RhsPacket B_0, T0; + LhsPacket A2; + EIGEN_GEBP_ONESTEP(0); + blB += 4*RhsProgress; + blA += 3*Traits::LhsProgress; + } + +#undef EIGEN_GEBP_ONESTEP + + ResPacket R0, R1, R2; + ResPacket alphav = pset1<ResPacket>(alpha); + + R0 = r0.loadPacket(0 * Traits::ResPacketSize); + R1 = r0.loadPacket(1 * Traits::ResPacketSize); + R2 = r0.loadPacket(2 * Traits::ResPacketSize); + traits.acc(C0, alphav, R0); + traits.acc(C4, alphav, R1); + traits.acc(C8, alphav, R2); + r0.storePacket(0 * Traits::ResPacketSize, R0); + r0.storePacket(1 * Traits::ResPacketSize, R1); + r0.storePacket(2 * Traits::ResPacketSize, R2); + + R0 = r1.loadPacket(0 * Traits::ResPacketSize); + R1 = r1.loadPacket(1 * Traits::ResPacketSize); + R2 = r1.loadPacket(2 * Traits::ResPacketSize); + traits.acc(C1, alphav, R0); + traits.acc(C5, alphav, R1); + traits.acc(C9, alphav, R2); + r1.storePacket(0 * Traits::ResPacketSize, R0); + r1.storePacket(1 * Traits::ResPacketSize, R1); + r1.storePacket(2 * Traits::ResPacketSize, R2); + + R0 = r2.loadPacket(0 * Traits::ResPacketSize); + R1 = r2.loadPacket(1 * Traits::ResPacketSize); + R2 = r2.loadPacket(2 * Traits::ResPacketSize); + traits.acc(C2, alphav, R0); + traits.acc(C6, alphav, R1); + traits.acc(C10, alphav, R2); + r2.storePacket(0 * Traits::ResPacketSize, R0); + r2.storePacket(1 * Traits::ResPacketSize, R1); + r2.storePacket(2 * Traits::ResPacketSize, R2); + + R0 = r3.loadPacket(0 * Traits::ResPacketSize); + R1 = r3.loadPacket(1 * Traits::ResPacketSize); + R2 = r3.loadPacket(2 * Traits::ResPacketSize); + traits.acc(C3, alphav, R0); + traits.acc(C7, alphav, R1); + traits.acc(C11, alphav, R2); + r3.storePacket(0 * Traits::ResPacketSize, R0); + r3.storePacket(1 * Traits::ResPacketSize, R1); + r3.storePacket(2 * Traits::ResPacketSize, R2); + } + } + + // Deal with remaining columns of the rhs + for(Index j2=packet_cols4; j2<cols; j2++) + { + for(Index i=i1; i<actual_panel_end; i+=3*LhsProgress) + { + // One column at a time + const LhsScalar* blA = &blockA[i*strideA+offsetA*(3*Traits::LhsProgress)]; + prefetch(&blA[0]); + + // gets res block as register + AccPacket C0, C4, C8; + traits.initAcc(C0); + traits.initAcc(C4); + traits.initAcc(C8); + + LinearMapper r0 = res.getLinearMapper(i, j2); + r0.prefetch(0); + + // performs "inner" products + const RhsScalar* blB = &blockB[j2*strideB+offsetB]; + LhsPacket A0, A1, A2; + + for(Index k=0; k<peeled_kc; k+=pk) + { + 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"); \ + 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) + + 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*3*Traits::LhsProgress; + + EIGEN_ASM_COMMENT("end gebp micro kernel 3pX1"); + } + + // process remaining peeled loop + for(Index k=peeled_kc; k<depth; k++) + { + RhsPacket B_0; + EIGEN_GEBGP_ONESTEP(0); + blB += RhsProgress; + blA += 3*Traits::LhsProgress; + } +#undef EIGEN_GEBGP_ONESTEP + ResPacket R0, R1, R2; + ResPacket alphav = pset1<ResPacket>(alpha); + + R0 = r0.loadPacket(0 * Traits::ResPacketSize); + R1 = r0.loadPacket(1 * Traits::ResPacketSize); + R2 = r0.loadPacket(2 * Traits::ResPacketSize); + traits.acc(C0, alphav, R0); + traits.acc(C4, alphav, R1); + traits.acc(C8, alphav, R2); + r0.storePacket(0 * Traits::ResPacketSize, R0); + r0.storePacket(1 * Traits::ResPacketSize, R1); + r0.storePacket(2 * Traits::ResPacketSize, R2); + } + } + } + } + + //---------- Process 2 * LhsProgress rows at once ---------- + if(mr>=2*Traits::LhsProgress) + { + const Index l1 = defaultL1CacheSize; // in Bytes, TODO, l1 should be passed to this function. + // The max(1, ...) here is needed because we may be using blocking params larger than what our known l1 cache size + // suggests we should be using: either because our known l1 cache size is inaccurate (e.g. on Android, we can only guess), + // or because we are testing specific blocking sizes. + Index actual_panel_rows = (2*LhsProgress) * std::max<Index>(1,( (l1 - sizeof(ResScalar)*mr*nr - depth*nr*sizeof(RhsScalar)) / (depth * sizeof(LhsScalar) * 2*LhsProgress) )); + + for(Index i1=peeled_mc3; i1<peeled_mc2; i1+=actual_panel_rows) + { + Index actual_panel_end = (std::min)(i1+actual_panel_rows, peeled_mc2); + for(Index j2=0; j2<packet_cols4; j2+=nr) + { + for(Index i=i1; i<actual_panel_end; i+=2*LhsProgress) + { + + // We selected a 2*Traits::LhsProgress x nr micro block of res which is entirely + // stored into 2 x nr registers. + + const LhsScalar* blA = &blockA[i*strideA+offsetA*(2*Traits::LhsProgress)]; + prefetch(&blA[0]); + + // gets res block as register + AccPacket C0, C1, C2, C3, + C4, C5, C6, C7; + traits.initAcc(C0); traits.initAcc(C1); traits.initAcc(C2); traits.initAcc(C3); + traits.initAcc(C4); traits.initAcc(C5); traits.initAcc(C6); traits.initAcc(C7); + + 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 2pX4"); + RhsPacket B_0, B1, B2, B3, T0; + + #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); + 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*(2*Traits::LhsProgress); + + EIGEN_ASM_COMMENT("end gebp micro kernel 2pX4"); + } + // process remaining peeled loop + for(Index k=peeled_kc; k<depth; k++) + { + RhsPacket B_0, B1, B2, B3, T0; + EIGEN_GEBGP_ONESTEP(0); + blB += 4*RhsProgress; + blA += 2*Traits::LhsProgress; + } +#undef EIGEN_GEBGP_ONESTEP + + ResPacket R0, R1, R2, R3; + ResPacket alphav = pset1<ResPacket>(alpha); + + R0 = r0.loadPacket(0 * Traits::ResPacketSize); + R1 = r0.loadPacket(1 * Traits::ResPacketSize); + R2 = r1.loadPacket(0 * Traits::ResPacketSize); + R3 = r1.loadPacket(1 * Traits::ResPacketSize); + traits.acc(C0, alphav, R0); + traits.acc(C4, alphav, R1); + traits.acc(C1, alphav, R2); + traits.acc(C5, alphav, R3); + r0.storePacket(0 * Traits::ResPacketSize, R0); + r0.storePacket(1 * Traits::ResPacketSize, R1); + r1.storePacket(0 * Traits::ResPacketSize, R2); + r1.storePacket(1 * Traits::ResPacketSize, R3); + + R0 = r2.loadPacket(0 * Traits::ResPacketSize); + R1 = r2.loadPacket(1 * Traits::ResPacketSize); + R2 = r3.loadPacket(0 * Traits::ResPacketSize); + R3 = r3.loadPacket(1 * Traits::ResPacketSize); + traits.acc(C2, alphav, R0); + traits.acc(C6, alphav, R1); + traits.acc(C3, alphav, R2); + traits.acc(C7, alphav, R3); + r2.storePacket(0 * Traits::ResPacketSize, R0); + r2.storePacket(1 * Traits::ResPacketSize, R1); + r3.storePacket(0 * Traits::ResPacketSize, R2); + r3.storePacket(1 * Traits::ResPacketSize, R3); + } + } + + // Deal with remaining columns of the rhs + for(Index j2=packet_cols4; j2<cols; j2++) + { + for(Index i=i1; i<actual_panel_end; i+=2*LhsProgress) + { + // One column at a time + const LhsScalar* blA = &blockA[i*strideA+offsetA*(2*Traits::LhsProgress)]; + prefetch(&blA[0]); + + // gets res block as register + AccPacket C0, C4; + traits.initAcc(C0); + traits.initAcc(C4); + + LinearMapper r0 = res.getLinearMapper(i, j2); + r0.prefetch(prefetch_res_offset); + + // performs "inner" products + const RhsScalar* blB = &blockB[j2*strideB+offsetB]; + LhsPacket A0, A1; + + for(Index k=0; k<peeled_kc; k+=pk) + { + EIGEN_ASM_COMMENT("begin gebp micro kernel 2pX1"); + RhsPacket B_0, B1; + +#define EIGEN_GEBGP_ONESTEP(K) \ + do { \ + EIGEN_ASM_COMMENT("begin step of gebp micro kernel 2pX1"); \ + 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.loadRhs(&blB[(0+K)*RhsProgress], B_0); \ + traits.madd(A0, B_0, C0, B1); \ + traits.madd(A1, B_0, C4, B_0); \ + EIGEN_ASM_COMMENT("end step of gebp micro kernel 2pX1"); \ + } 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*2*Traits::LhsProgress; + + EIGEN_ASM_COMMENT("end gebp micro kernel 2pX1"); + } + + // process remaining peeled loop + for(Index k=peeled_kc; k<depth; k++) + { + RhsPacket B_0, B1; + EIGEN_GEBGP_ONESTEP(0); + blB += RhsProgress; + blA += 2*Traits::LhsProgress; + } +#undef EIGEN_GEBGP_ONESTEP + ResPacket R0, R1; + ResPacket alphav = pset1<ResPacket>(alpha); + + R0 = r0.loadPacket(0 * Traits::ResPacketSize); + R1 = r0.loadPacket(1 * Traits::ResPacketSize); + traits.acc(C0, alphav, R0); + traits.acc(C4, alphav, R1); + r0.storePacket(0 * Traits::ResPacketSize, R0); + r0.storePacket(1 * Traits::ResPacketSize, R1); + } + } + } + } + //---------- 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.loadPacket(0 * Traits::ResPacketSize); + R1 = r1.loadPacket(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.loadPacket(0 * Traits::ResPacketSize); + R1 = r3.loadPacket(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.loadPacket(0 * Traits::ResPacketSize); + traits.acc(C0, alphav, R0); + r0.storePacket(0 * Traits::ResPacketSize, R0); + } + } + } + //---------- Process remaining rows, 1 at once ---------- + if(peeled_mc1<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) + { + const LhsScalar* blA = &blockA[i*strideA+offsetA]; + prefetch(&blA[0]); + const RhsScalar* blB = &blockB[j2*strideB+offsetB*nr]; + + // The following piece of code wont work for 512 bit registers + // Moreover, if LhsProgress==8 it assumes that there is a half packet of the same size + // as nr (which is currently 4) for the return type. + typedef typename unpacket_traits<SResPacket>::half SResPacketHalf; + if ((SwappedTraits::LhsProgress % 4) == 0 && + (SwappedTraits::LhsProgress <= 8) && + (SwappedTraits::LhsProgress!=8 || unpacket_traits<SResPacketHalf>::size==nr)) + { + SAccPacket C0, C1, C2, C3; + straits.initAcc(C0); + straits.initAcc(C1); + straits.initAcc(C2); + straits.initAcc(C3); + + const Index spk = (std::max)(1,SwappedTraits::LhsProgress/4); + const Index endk = (depth/spk)*spk; + const Index endk4 = (depth/(spk*4))*(spk*4); + + Index k=0; + for(; k<endk4; k+=4*spk) + { + SLhsPacket A0,A1; + SRhsPacket B_0,B_1; + + straits.loadLhsUnaligned(blB+0*SwappedTraits::LhsProgress, A0); + straits.loadLhsUnaligned(blB+1*SwappedTraits::LhsProgress, A1); + + 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.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); + + blB += 4*SwappedTraits::LhsProgress; + blA += 4*spk; + } + C0 = padd(padd(C0,C1),padd(C2,C3)); + for(; k<endk; k+=spk) + { + SLhsPacket A0; + SRhsPacket B_0; + + straits.loadLhsUnaligned(blB, A0); + straits.loadRhsQuad(blA, B_0); + straits.madd(A0,B_0,C0,B_0); + + blB += SwappedTraits::LhsProgress; + blA += spk; + } + if(SwappedTraits::LhsProgress==8) + { + // 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<SAccPacket>::half,SAccPacket>::type SAccPacketHalf; + + SResPacketHalf R = res.template gatherPacket<SResPacketHalf>(i, j2); + SResPacketHalf alphav = pset1<SResPacketHalf>(alpha); + + if(depth-endk>0) + { + // We have to handle the last row of the rhs which corresponds to a half-packet + SLhsPacketHalf a0; + SRhsPacketHalf b0; + straits.loadLhsUnaligned(blB, a0); + straits.loadRhs(blA, b0); + SAccPacketHalf c0 = predux_downto4(C0); + straits.madd(a0,b0,c0,b0); + straits.acc(c0, alphav, R); + } + else + { + straits.acc(predux_downto4(C0), alphav, R); + } + res.scatterPacket(i, j2, R); + } + else + { + SResPacket R = res.template gatherPacket<SResPacket>(i, j2); + SResPacket alphav = pset1<SResPacket>(alpha); + straits.acc(C0, alphav, R); + res.scatterPacket(i, j2, R); + } + } + else // scalar path + { + // get a 1 x 4 res block as registers + ResScalar C0(0), C1(0), C2(0), C3(0); + + for(Index k=0; k<depth; k++) + { + LhsScalar A0; + RhsScalar B_0, B_1; + + A0 = blA[k]; + + B_0 = blB[0]; + B_1 = blB[1]; + CJMADD(cj,A0,B_0,C0, B_0); + CJMADD(cj,A0,B_1,C1, B_1); + + B_0 = blB[2]; + B_1 = blB[3]; + CJMADD(cj,A0,B_0,C2, B_0); + CJMADD(cj,A0,B_1,C3, B_1); + + blB += 4; + } + res(i, j2 + 0) += alpha * C0; + res(i, j2 + 1) += alpha * C1; + res(i, j2 + 2) += alpha * C2; + res(i, j2 + 3) += alpha * C3; + } + } + } + // remaining columns + 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) + { + const LhsScalar* blA = &blockA[i*strideA+offsetA]; + prefetch(&blA[0]); + // gets a 1 x 1 res block as registers + ResScalar C0(0); + const RhsScalar* blB = &blockB[j2*strideB+offsetB]; + for(Index k=0; k<depth; k++) + { + LhsScalar A0 = blA[k]; + RhsScalar B_0 = blB[k]; + CJMADD(cj, A0, B_0, C0, B_0); + } + res(i, j2) += alpha * C0; + } + } + } + } + + +#undef CJMADD + +// pack a block of the lhs +// The traversal is as follow (mr==4): +// 0 4 8 12 ... +// 1 5 9 13 ... +// 2 6 10 14 ... +// 3 7 11 15 ... +// +// 16 20 24 28 ... +// 17 21 25 29 ... +// 18 22 26 30 ... +// 19 23 27 31 ... +// +// 32 33 34 35 ... +// 36 36 38 39 ... +template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, bool Conjugate, bool PanelMode> +struct gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, ColMajor, Conjugate, PanelMode> +{ + typedef typename DataMapper::LinearMapper LinearMapper; + EIGEN_DONT_INLINE void operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride=0, Index offset=0); +}; + +template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, bool Conjugate, bool PanelMode> +EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, ColMajor, Conjugate, PanelMode> + ::operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride, Index offset) +{ + typedef typename packet_traits<Scalar>::type Packet; + enum { PacketSize = packet_traits<Scalar>::size }; + + EIGEN_ASM_COMMENT("EIGEN PRODUCT PACK LHS"); + EIGEN_UNUSED_VARIABLE(stride); + EIGEN_UNUSED_VARIABLE(offset); + eigen_assert(((!PanelMode) && stride==0 && offset==0) || (PanelMode && stride>=depth && offset<=stride)); + eigen_assert( ((Pack1%PacketSize)==0 && Pack1<=4*PacketSize) || (Pack1<=4) ); + conj_if<NumTraits<Scalar>::IsComplex && Conjugate> cj; + Index count = 0; + + 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; + + Index i=0; + + // Pack 3 packets + if(Pack1>=3*PacketSize) + { + for(; i<peeled_mc3; i+=3*PacketSize) + { + if(PanelMode) count += (3*PacketSize) * offset; + + for(Index k=0; k<depth; k++) + { + Packet A, B, C; + A = lhs.loadPacket(i+0*PacketSize, k); + B = lhs.loadPacket(i+1*PacketSize, k); + C = lhs.loadPacket(i+2*PacketSize, k); + pstore(blockA+count, cj.pconj(A)); count+=PacketSize; + pstore(blockA+count, cj.pconj(B)); count+=PacketSize; + pstore(blockA+count, cj.pconj(C)); count+=PacketSize; + } + if(PanelMode) count += (3*PacketSize) * (stride-offset-depth); + } + } + // Pack 2 packets + if(Pack1>=2*PacketSize) + { + for(; i<peeled_mc2; i+=2*PacketSize) + { + if(PanelMode) count += (2*PacketSize) * offset; + + for(Index k=0; k<depth; k++) + { + Packet A, B; + A = lhs.loadPacket(i+0*PacketSize, k); + B = lhs.loadPacket(i+1*PacketSize, k); + pstore(blockA+count, cj.pconj(A)); count+=PacketSize; + pstore(blockA+count, cj.pconj(B)); count+=PacketSize; + } + if(PanelMode) count += (2*PacketSize) * (stride-offset-depth); + } + } + // Pack 1 packets + if(Pack1>=1*PacketSize) + { + for(; i<peeled_mc1; i+=1*PacketSize) + { + if(PanelMode) count += (1*PacketSize) * offset; + + for(Index k=0; k<depth; k++) + { + Packet A; + A = lhs.loadPacket(i+0*PacketSize, k); + pstore(blockA+count, cj.pconj(A)); + count+=PacketSize; + } + if(PanelMode) count += (1*PacketSize) * (stride-offset-depth); + } + } + // Pack scalars + if(Pack2<PacketSize && Pack2>1) + { + for(; i<peeled_mc0; i+=Pack2) + { + if(PanelMode) count += Pack2 * offset; + + for(Index k=0; k<depth; k++) + for(Index w=0; w<Pack2; w++) + blockA[count++] = cj(lhs(i+w, k)); + + if(PanelMode) count += Pack2 * (stride-offset-depth); + } + } + for(; i<rows; i++) + { + if(PanelMode) count += offset; + for(Index k=0; k<depth; k++) + blockA[count++] = cj(lhs(i, k)); + if(PanelMode) count += (stride-offset-depth); + } +} + +template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, bool Conjugate, bool PanelMode> +struct gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, RowMajor, Conjugate, PanelMode> +{ + typedef typename DataMapper::LinearMapper LinearMapper; + EIGEN_DONT_INLINE void operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride=0, Index offset=0); +}; + +template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, bool Conjugate, bool PanelMode> +EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, RowMajor, Conjugate, PanelMode> + ::operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride, Index offset) +{ + typedef typename packet_traits<Scalar>::type Packet; + enum { PacketSize = packet_traits<Scalar>::size }; + + EIGEN_ASM_COMMENT("EIGEN PRODUCT PACK LHS"); + EIGEN_UNUSED_VARIABLE(stride); + EIGEN_UNUSED_VARIABLE(offset); + eigen_assert(((!PanelMode) && stride==0 && offset==0) || (PanelMode && stride>=depth && offset<=stride)); + conj_if<NumTraits<Scalar>::IsComplex && Conjugate> cj; + Index count = 0; + +// 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; + while(pack>0) + { + Index remaining_rows = rows-i; + Index peeled_mc = i+(remaining_rows/pack)*pack; + for(; i<peeled_mc; i+=pack) + { + if(PanelMode) count += pack * offset; + + const Index peeled_k = (depth/PacketSize)*PacketSize; + Index k=0; + if(pack>=PacketSize) + { + for(; k<peeled_k; k+=PacketSize) + { + for (Index m = 0; m < pack; m += PacketSize) + { + PacketBlock<Packet> kernel; + for (int p = 0; p < PacketSize; ++p) kernel.packet[p] = lhs.loadPacket(i+p+m, k); + ptranspose(kernel); + for (int p = 0; p < PacketSize; ++p) pstore(blockA+count+m+(pack)*p, cj.pconj(kernel.packet[p])); + } + count += PacketSize*pack; + } + } + for(; k<depth; k++) + { + Index w=0; + for(; w<pack-3; w+=4) + { + Scalar a(cj(lhs(i+w+0, k))), + b(cj(lhs(i+w+1, k))), + c(cj(lhs(i+w+2, k))), + d(cj(lhs(i+w+3, k))); + blockA[count++] = a; + blockA[count++] = b; + blockA[count++] = c; + blockA[count++] = d; + } + if(pack%4) + for(;w<pack;++w) + blockA[count++] = cj(lhs(i+w, k)); + } + + if(PanelMode) count += pack * (stride-offset-depth); + } + + pack -= PacketSize; + if(pack<Pack2 && (pack+PacketSize)!=Pack2) + pack = Pack2; + } + + for(; i<rows; i++) + { + if(PanelMode) count += offset; + for(Index k=0; k<depth; k++) + blockA[count++] = cj(lhs(i, k)); + if(PanelMode) count += (stride-offset-depth); + } +} + +// copy a complete panel of the rhs +// this version is optimized for column major matrices +// The traversal order is as follow: (nr==4): +// 0 1 2 3 12 13 14 15 24 27 +// 4 5 6 7 16 17 18 19 25 28 +// 8 9 10 11 20 21 22 23 26 29 +// . . . . . . . . . . +template<typename Scalar, typename Index, typename DataMapper, int nr, bool Conjugate, bool PanelMode> +struct gemm_pack_rhs<Scalar, Index, DataMapper, nr, ColMajor, Conjugate, PanelMode> +{ + typedef typename packet_traits<Scalar>::type Packet; + typedef typename DataMapper::LinearMapper LinearMapper; + enum { PacketSize = packet_traits<Scalar>::size }; + EIGEN_DONT_INLINE void operator()(Scalar* blockB, const DataMapper& rhs, Index depth, Index cols, Index stride=0, Index offset=0); +}; + +template<typename Scalar, typename Index, typename DataMapper, int nr, bool Conjugate, bool PanelMode> +EIGEN_DONT_INLINE void gemm_pack_rhs<Scalar, Index, DataMapper, nr, ColMajor, Conjugate, PanelMode> + ::operator()(Scalar* blockB, const DataMapper& rhs, Index depth, Index cols, Index stride, Index offset) +{ + EIGEN_ASM_COMMENT("EIGEN PRODUCT PACK RHS COLMAJOR"); + EIGEN_UNUSED_VARIABLE(stride); + EIGEN_UNUSED_VARIABLE(offset); + eigen_assert(((!PanelMode) && stride==0 && offset==0) || (PanelMode && stride>=depth && offset<=stride)); + conj_if<NumTraits<Scalar>::IsComplex && Conjugate> cj; + Index packet_cols8 = nr>=8 ? (cols/8) * 8 : 0; + Index packet_cols4 = nr>=4 ? (cols/4) * 4 : 0; + Index count = 0; + const Index peeled_k = (depth/PacketSize)*PacketSize; +// if(nr>=8) +// { +// for(Index j2=0; j2<packet_cols8; j2+=8) +// { +// // skip what we have before +// if(PanelMode) count += 8 * offset; +// const Scalar* b0 = &rhs[(j2+0)*rhsStride]; +// const Scalar* b1 = &rhs[(j2+1)*rhsStride]; +// const Scalar* b2 = &rhs[(j2+2)*rhsStride]; +// const Scalar* b3 = &rhs[(j2+3)*rhsStride]; +// const Scalar* b4 = &rhs[(j2+4)*rhsStride]; +// const Scalar* b5 = &rhs[(j2+5)*rhsStride]; +// const Scalar* b6 = &rhs[(j2+6)*rhsStride]; +// const Scalar* b7 = &rhs[(j2+7)*rhsStride]; +// Index k=0; +// if(PacketSize==8) // TODO enbale vectorized transposition for PacketSize==4 +// { +// for(; k<peeled_k; k+=PacketSize) { +// PacketBlock<Packet> kernel; +// for (int p = 0; p < PacketSize; ++p) { +// kernel.packet[p] = ploadu<Packet>(&rhs[(j2+p)*rhsStride+k]); +// } +// ptranspose(kernel); +// for (int p = 0; p < PacketSize; ++p) { +// pstoreu(blockB+count, cj.pconj(kernel.packet[p])); +// count+=PacketSize; +// } +// } +// } +// for(; k<depth; k++) +// { +// blockB[count+0] = cj(b0[k]); +// blockB[count+1] = cj(b1[k]); +// blockB[count+2] = cj(b2[k]); +// blockB[count+3] = cj(b3[k]); +// blockB[count+4] = cj(b4[k]); +// blockB[count+5] = cj(b5[k]); +// blockB[count+6] = cj(b6[k]); +// blockB[count+7] = cj(b7[k]); +// count += 8; +// } +// // skip what we have after +// if(PanelMode) count += 8 * (stride-offset-depth); +// } +// } + + if(nr>=4) + { + for(Index j2=packet_cols8; j2<packet_cols4; j2+=4) + { + // skip what we have before + if(PanelMode) count += 4 * offset; + const LinearMapper dm0 = rhs.getLinearMapper(0, j2 + 0); + const LinearMapper dm1 = rhs.getLinearMapper(0, j2 + 1); + const LinearMapper dm2 = rhs.getLinearMapper(0, j2 + 2); + const LinearMapper dm3 = rhs.getLinearMapper(0, j2 + 3); + + Index k=0; + if((PacketSize%4)==0) // TODO enable vectorized transposition for PacketSize==2 ?? + { + for(; k<peeled_k; k+=PacketSize) { + PacketBlock<Packet,(PacketSize%4)==0?4:PacketSize> kernel; + kernel.packet[0] = dm0.loadPacket(k); + kernel.packet[1%PacketSize] = dm1.loadPacket(k); + kernel.packet[2%PacketSize] = dm2.loadPacket(k); + kernel.packet[3%PacketSize] = dm3.loadPacket(k); + ptranspose(kernel); + pstoreu(blockB+count+0*PacketSize, cj.pconj(kernel.packet[0])); + pstoreu(blockB+count+1*PacketSize, cj.pconj(kernel.packet[1%PacketSize])); + pstoreu(blockB+count+2*PacketSize, cj.pconj(kernel.packet[2%PacketSize])); + pstoreu(blockB+count+3*PacketSize, cj.pconj(kernel.packet[3%PacketSize])); + count+=4*PacketSize; + } + } + for(; k<depth; k++) + { + blockB[count+0] = cj(dm0(k)); + blockB[count+1] = cj(dm1(k)); + blockB[count+2] = cj(dm2(k)); + blockB[count+3] = cj(dm3(k)); + count += 4; + } + // skip what we have after + if(PanelMode) count += 4 * (stride-offset-depth); + } + } + + // copy the remaining columns one at a time (nr==1) + for(Index j2=packet_cols4; j2<cols; ++j2) + { + if(PanelMode) count += offset; + const LinearMapper dm0 = rhs.getLinearMapper(0, j2); + for(Index k=0; k<depth; k++) + { + blockB[count] = cj(dm0(k)); + count += 1; + } + if(PanelMode) count += (stride-offset-depth); + } +} + +// this version is optimized for row major matrices +template<typename Scalar, typename Index, typename DataMapper, int nr, bool Conjugate, bool PanelMode> +struct gemm_pack_rhs<Scalar, Index, DataMapper, nr, RowMajor, Conjugate, PanelMode> +{ + typedef typename packet_traits<Scalar>::type Packet; + typedef typename DataMapper::LinearMapper LinearMapper; + enum { PacketSize = packet_traits<Scalar>::size }; + EIGEN_DONT_INLINE void operator()(Scalar* blockB, const DataMapper& rhs, Index depth, Index cols, Index stride=0, Index offset=0); +}; + +template<typename Scalar, typename Index, typename DataMapper, int nr, bool Conjugate, bool PanelMode> +EIGEN_DONT_INLINE void gemm_pack_rhs<Scalar, Index, DataMapper, nr, RowMajor, Conjugate, PanelMode> + ::operator()(Scalar* blockB, const DataMapper& rhs, Index depth, Index cols, Index stride, Index offset) +{ + EIGEN_ASM_COMMENT("EIGEN PRODUCT PACK RHS ROWMAJOR"); + EIGEN_UNUSED_VARIABLE(stride); + EIGEN_UNUSED_VARIABLE(offset); + eigen_assert(((!PanelMode) && stride==0 && offset==0) || (PanelMode && stride>=depth && offset<=stride)); + conj_if<NumTraits<Scalar>::IsComplex && Conjugate> cj; + Index packet_cols8 = nr>=8 ? (cols/8) * 8 : 0; + Index packet_cols4 = nr>=4 ? (cols/4) * 4 : 0; + Index count = 0; + +// if(nr>=8) +// { +// for(Index j2=0; j2<packet_cols8; j2+=8) +// { +// // skip what we have before +// if(PanelMode) count += 8 * offset; +// for(Index k=0; k<depth; k++) +// { +// if (PacketSize==8) { +// Packet A = ploadu<Packet>(&rhs[k*rhsStride + j2]); +// pstoreu(blockB+count, cj.pconj(A)); +// } else if (PacketSize==4) { +// Packet A = ploadu<Packet>(&rhs[k*rhsStride + j2]); +// Packet B = ploadu<Packet>(&rhs[k*rhsStride + j2 + PacketSize]); +// pstoreu(blockB+count, cj.pconj(A)); +// pstoreu(blockB+count+PacketSize, cj.pconj(B)); +// } else { +// const Scalar* b0 = &rhs[k*rhsStride + j2]; +// blockB[count+0] = cj(b0[0]); +// blockB[count+1] = cj(b0[1]); +// blockB[count+2] = cj(b0[2]); +// blockB[count+3] = cj(b0[3]); +// blockB[count+4] = cj(b0[4]); +// blockB[count+5] = cj(b0[5]); +// blockB[count+6] = cj(b0[6]); +// blockB[count+7] = cj(b0[7]); +// } +// count += 8; +// } +// // skip what we have after +// if(PanelMode) count += 8 * (stride-offset-depth); +// } +// } + if(nr>=4) + { + for(Index j2=packet_cols8; j2<packet_cols4; j2+=4) + { + // skip what we have before + if(PanelMode) count += 4 * offset; + for(Index k=0; k<depth; k++) + { + if (PacketSize==4) { + Packet A = rhs.loadPacket(k, j2); + pstoreu(blockB+count, cj.pconj(A)); + count += PacketSize; + } else { + const LinearMapper dm0 = rhs.getLinearMapper(k, j2); + blockB[count+0] = cj(dm0(0)); + blockB[count+1] = cj(dm0(1)); + blockB[count+2] = cj(dm0(2)); + blockB[count+3] = cj(dm0(3)); + count += 4; + } + } + // skip what we have after + if(PanelMode) count += 4 * (stride-offset-depth); + } + } + // copy the remaining columns one at a time (nr==1) + for(Index j2=packet_cols4; j2<cols; ++j2) + { + if(PanelMode) count += offset; + for(Index k=0; k<depth; k++) + { + blockB[count] = cj(rhs(k, j2)); + count += 1; + } + if(PanelMode) count += stride-offset-depth; + } +} + +} // end namespace internal + +/** \returns the currently set level 1 cpu cache size (in bytes) used to estimate the ideal blocking size parameters. + * \sa setCpuCacheSize */ +inline std::ptrdiff_t l1CacheSize() +{ + std::ptrdiff_t l1, l2, l3; + internal::manage_caching_sizes(GetAction, &l1, &l2, &l3); + return l1; +} + +/** \returns the currently set level 2 cpu cache size (in bytes) used to estimate the ideal blocking size parameters. + * \sa setCpuCacheSize */ +inline std::ptrdiff_t l2CacheSize() +{ + std::ptrdiff_t l1, l2, l3; + internal::manage_caching_sizes(GetAction, &l1, &l2, &l3); + return l2; +} + +/** \returns the currently set level 3 cpu cache size (in bytes) used to estimate the ideal blocking size paramete\ +rs. +* \sa setCpuCacheSize */ +inline std::ptrdiff_t l3CacheSize() +{ + std::ptrdiff_t l1, l2, l3; + internal::manage_caching_sizes(GetAction, &l1, &l2, &l3); + return l3; +} + +/** Set the cpu L1 and L2 cache sizes (in bytes). + * These values are use to adjust the size of the blocks + * for the algorithms working per blocks. + * + * \sa computeProductBlockingSizes */ +inline void setCpuCacheSizes(std::ptrdiff_t l1, std::ptrdiff_t l2, std::ptrdiff_t l3) +{ + internal::manage_caching_sizes(SetAction, &l1, &l2, &l3); +} + +} // end namespace Eigen + +#endif // EIGEN_GENERAL_BLOCK_PANEL_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixMatrix.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixMatrix.h new file mode 100644 index 000000000..6440e1d09 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixMatrix.h @@ -0,0 +1,492 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2009 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_GENERAL_MATRIX_MATRIX_H +#define EIGEN_GENERAL_MATRIX_MATRIX_H + +namespace Eigen { + +namespace internal { + +template<typename _LhsScalar, typename _RhsScalar> class level3_blocking; + +/* Specialization for a row-major destination matrix => simple transposition of the product */ +template< + typename Index, + typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs, + typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs> +struct general_matrix_matrix_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,RowMajor> +{ + typedef gebp_traits<RhsScalar,LhsScalar> Traits; + + typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar; + static EIGEN_STRONG_INLINE void run( + Index rows, Index cols, Index depth, + const LhsScalar* lhs, Index lhsStride, + const RhsScalar* rhs, Index rhsStride, + ResScalar* res, Index resStride, + ResScalar alpha, + level3_blocking<RhsScalar,LhsScalar>& blocking, + GemmParallelInfo<Index>* info = 0) + { + // transpose the product such that the result is column major + general_matrix_matrix_product<Index, + RhsScalar, RhsStorageOrder==RowMajor ? ColMajor : RowMajor, ConjugateRhs, + LhsScalar, LhsStorageOrder==RowMajor ? ColMajor : RowMajor, ConjugateLhs, + ColMajor> + ::run(cols,rows,depth,rhs,rhsStride,lhs,lhsStride,res,resStride,alpha,blocking,info); + } +}; + +/* Specialization for a col-major destination matrix + * => Blocking algorithm following Goto's paper */ +template< + typename Index, + typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs, + typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs> +struct general_matrix_matrix_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,ColMajor> +{ + +typedef gebp_traits<LhsScalar,RhsScalar> Traits; + +typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar; +static void run(Index rows, Index cols, Index depth, + const LhsScalar* _lhs, Index lhsStride, + const RhsScalar* _rhs, Index rhsStride, + ResScalar* _res, Index resStride, + ResScalar alpha, + level3_blocking<LhsScalar,RhsScalar>& blocking, + GemmParallelInfo<Index>* info = 0) +{ + typedef const_blas_data_mapper<LhsScalar, Index, LhsStorageOrder> LhsMapper; + typedef const_blas_data_mapper<RhsScalar, Index, RhsStorageOrder> RhsMapper; + typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper; + LhsMapper lhs(_lhs,lhsStride); + RhsMapper rhs(_rhs,rhsStride); + ResMapper res(_res, resStride); + + Index kc = blocking.kc(); // cache block size along the K direction + Index mc = (std::min)(rows,blocking.mc()); // cache block size along the M direction + Index nc = (std::min)(cols,blocking.nc()); // cache block size along the N direction + + gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs; + gemm_pack_rhs<RhsScalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs; + gebp_kernel<LhsScalar, RhsScalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp; + +#ifdef EIGEN_HAS_OPENMP + if(info) + { + // this is the parallel version! + int tid = omp_get_thread_num(); + int threads = omp_get_num_threads(); + + LhsScalar* blockA = blocking.blockA(); + eigen_internal_assert(blockA!=0); + + std::size_t sizeB = kc*nc; + ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, sizeB, 0); + + // For each horizontal panel of the rhs, and corresponding vertical panel of the lhs... + for(Index k=0; k<depth; k+=kc) + { + const Index actual_kc = (std::min)(k+kc,depth)-k; // => rows of B', and cols of the A' + + // In order to reduce the chance that a thread has to wait for the other, + // let's start by packing B'. + pack_rhs(blockB, rhs.getSubMapper(k,0), actual_kc, nc); + + // Pack A_k to A' in a parallel fashion: + // each thread packs the sub block A_k,i to A'_i where i is the thread id. + + // However, before copying to A'_i, we have to make sure that no other thread is still using it, + // i.e., we test that info[tid].users equals 0. + // Then, we set info[tid].users to the number of threads to mark that all other threads are going to use it. + while(info[tid].users!=0) {} + info[tid].users += threads; + + pack_lhs(blockA+info[tid].lhs_start*actual_kc, lhs.getSubMapper(info[tid].lhs_start,k), actual_kc, info[tid].lhs_length); + + // Notify the other threads that the part A'_i is ready to go. + info[tid].sync = k; + + // Computes C_i += A' * B' per A'_i + for(int shift=0; shift<threads; ++shift) + { + int i = (tid+shift)%threads; + + // At this point we have to make sure that A'_i has been updated by the thread i, + // we use testAndSetOrdered to mimic a volatile access. + // However, no need to wait for the B' part which has been updated by the current thread! + if (shift>0) { + while(info[i].sync!=k) { + } + } + + gebp(res.getSubMapper(info[i].lhs_start, 0), blockA+info[i].lhs_start*actual_kc, blockB, info[i].lhs_length, actual_kc, nc, alpha); + } + + // Then keep going as usual with the remaining B' + for(Index j=nc; j<cols; j+=nc) + { + const Index actual_nc = (std::min)(j+nc,cols)-j; + + // pack B_k,j to B' + pack_rhs(blockB, rhs.getSubMapper(k,j), actual_kc, actual_nc); + + // C_j += A' * B' + gebp(res.getSubMapper(0, j), blockA, blockB, rows, actual_kc, actual_nc, alpha); + } + + // Release all the sub blocks A'_i of A' for the current thread, + // i.e., we simply decrement the number of users by 1 + for(Index i=0; i<threads; ++i) + #pragma omp atomic + info[i].users -= 1; + } + } + else +#endif // EIGEN_HAS_OPENMP + { + EIGEN_UNUSED_VARIABLE(info); + + // this is the sequential version! + std::size_t sizeA = kc*mc; + std::size_t sizeB = kc*nc; + + ei_declare_aligned_stack_constructed_variable(LhsScalar, blockA, sizeA, blocking.blockA()); + ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, sizeB, blocking.blockB()); + + const bool pack_rhs_once = mc!=rows && kc==depth && nc==cols; + + // For each horizontal panel of the rhs, and corresponding panel of the lhs... + for(Index i2=0; i2<rows; i2+=mc) + { + const Index actual_mc = (std::min)(i2+mc,rows)-i2; + + for(Index k2=0; k2<depth; k2+=kc) + { + const Index actual_kc = (std::min)(k2+kc,depth)-k2; + + // OK, here we have selected one horizontal panel of rhs and one vertical panel of lhs. + // => Pack lhs's panel into a sequential chunk of memory (L2/L3 caching) + // Note that this panel will be read as many times as the number of blocks in the rhs's + // horizontal panel which is, in practice, a very low number. + pack_lhs(blockA, lhs.getSubMapper(i2,k2), actual_kc, actual_mc); + + // For each kc x nc block of the rhs's horizontal panel... + for(Index j2=0; j2<cols; j2+=nc) + { + const Index actual_nc = (std::min)(j2+nc,cols)-j2; + + // We pack the rhs's block into a sequential chunk of memory (L2 caching) + // Note that this block will be read a very high number of times, which is equal to the number of + // micro horizontal panel of the large rhs's panel (e.g., rows/12 times). + if((!pack_rhs_once) || i2==0) + pack_rhs(blockB, rhs.getSubMapper(k2,j2), actual_kc, actual_nc); + + // Everything is packed, we can now call the panel * block kernel: + gebp(res.getSubMapper(i2, j2), blockA, blockB, actual_mc, actual_kc, actual_nc, alpha); + } + } + } + } +} + +}; + +/********************************************************************************* +* Specialization of generic_product_impl for "large" GEMM, i.e., +* implementation of the high level wrapper to general_matrix_matrix_product +**********************************************************************************/ + +template<typename Scalar, typename Index, typename Gemm, typename Lhs, typename Rhs, typename Dest, typename BlockingType> +struct gemm_functor +{ + gemm_functor(const Lhs& lhs, const Rhs& rhs, Dest& dest, const Scalar& actualAlpha, BlockingType& blocking) + : m_lhs(lhs), m_rhs(rhs), m_dest(dest), m_actualAlpha(actualAlpha), m_blocking(blocking) + {} + + void initParallelSession(Index num_threads) const + { + m_blocking.initParallel(m_lhs.rows(), m_rhs.cols(), m_lhs.cols(), num_threads); + m_blocking.allocateA(); + } + + void operator() (Index row, Index rows, Index col=0, Index cols=-1, GemmParallelInfo<Index>* info=0) const + { + if(cols==-1) + cols = m_rhs.cols(); + + Gemm::run(rows, cols, m_lhs.cols(), + &m_lhs.coeffRef(row,0), m_lhs.outerStride(), + &m_rhs.coeffRef(0,col), m_rhs.outerStride(), + (Scalar*)&(m_dest.coeffRef(row,col)), m_dest.outerStride(), + m_actualAlpha, m_blocking, info); + } + + typedef typename Gemm::Traits Traits; + + protected: + const Lhs& m_lhs; + const Rhs& m_rhs; + Dest& m_dest; + Scalar m_actualAlpha; + BlockingType& m_blocking; +}; + +template<int StorageOrder, typename LhsScalar, typename RhsScalar, int MaxRows, int MaxCols, int MaxDepth, int KcFactor=1, +bool FiniteAtCompileTime = MaxRows!=Dynamic && MaxCols!=Dynamic && MaxDepth != Dynamic> class gemm_blocking_space; + +template<typename _LhsScalar, typename _RhsScalar> +class level3_blocking +{ + typedef _LhsScalar LhsScalar; + typedef _RhsScalar RhsScalar; + + protected: + LhsScalar* m_blockA; + RhsScalar* m_blockB; + + Index m_mc; + Index m_nc; + Index m_kc; + + public: + + level3_blocking() + : m_blockA(0), m_blockB(0), m_mc(0), m_nc(0), m_kc(0) + {} + + inline Index mc() const { return m_mc; } + inline Index nc() const { return m_nc; } + inline Index kc() const { return m_kc; } + + inline LhsScalar* blockA() { return m_blockA; } + inline RhsScalar* blockB() { return m_blockB; } +}; + +template<int StorageOrder, typename _LhsScalar, typename _RhsScalar, int MaxRows, int MaxCols, int MaxDepth, int KcFactor> +class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, MaxDepth, KcFactor, true /* == FiniteAtCompileTime */> + : public level3_blocking< + typename conditional<StorageOrder==RowMajor,_RhsScalar,_LhsScalar>::type, + typename conditional<StorageOrder==RowMajor,_LhsScalar,_RhsScalar>::type> +{ + enum { + Transpose = StorageOrder==RowMajor, + ActualRows = Transpose ? MaxCols : MaxRows, + ActualCols = Transpose ? MaxRows : MaxCols + }; + typedef typename conditional<Transpose,_RhsScalar,_LhsScalar>::type LhsScalar; + typedef typename conditional<Transpose,_LhsScalar,_RhsScalar>::type RhsScalar; + typedef gebp_traits<LhsScalar,RhsScalar> Traits; + enum { + SizeA = ActualRows * MaxDepth, + SizeB = ActualCols * MaxDepth + }; + +#if EIGEN_MAX_STATIC_ALIGN_BYTES >= EIGEN_DEFAULT_ALIGN_BYTES + EIGEN_ALIGN_MAX LhsScalar m_staticA[SizeA]; + EIGEN_ALIGN_MAX RhsScalar m_staticB[SizeB]; +#else + EIGEN_ALIGN_MAX char m_staticA[SizeA * sizeof(LhsScalar) + EIGEN_DEFAULT_ALIGN_BYTES-1]; + EIGEN_ALIGN_MAX char m_staticB[SizeB * sizeof(RhsScalar) + EIGEN_DEFAULT_ALIGN_BYTES-1]; +#endif + + public: + + gemm_blocking_space(Index /*rows*/, Index /*cols*/, Index /*depth*/, Index /*num_threads*/, bool /*full_rows = false*/) + { + this->m_mc = ActualRows; + this->m_nc = ActualCols; + this->m_kc = MaxDepth; +#if EIGEN_MAX_STATIC_ALIGN_BYTES >= EIGEN_DEFAULT_ALIGN_BYTES + this->m_blockA = m_staticA; + this->m_blockB = m_staticB; +#else + this->m_blockA = reinterpret_cast<LhsScalar*>((internal::UIntPtr(m_staticA) + (EIGEN_DEFAULT_ALIGN_BYTES-1)) & ~std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1)); + this->m_blockB = reinterpret_cast<RhsScalar*>((internal::UIntPtr(m_staticB) + (EIGEN_DEFAULT_ALIGN_BYTES-1)) & ~std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1)); +#endif + } + + void initParallel(Index, Index, Index, Index) + {} + + inline void allocateA() {} + inline void allocateB() {} + inline void allocateAll() {} +}; + +template<int StorageOrder, typename _LhsScalar, typename _RhsScalar, int MaxRows, int MaxCols, int MaxDepth, int KcFactor> +class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, MaxDepth, KcFactor, false> + : public level3_blocking< + typename conditional<StorageOrder==RowMajor,_RhsScalar,_LhsScalar>::type, + typename conditional<StorageOrder==RowMajor,_LhsScalar,_RhsScalar>::type> +{ + enum { + Transpose = StorageOrder==RowMajor + }; + typedef typename conditional<Transpose,_RhsScalar,_LhsScalar>::type LhsScalar; + typedef typename conditional<Transpose,_LhsScalar,_RhsScalar>::type RhsScalar; + typedef gebp_traits<LhsScalar,RhsScalar> Traits; + + Index m_sizeA; + Index m_sizeB; + + public: + + gemm_blocking_space(Index rows, Index cols, Index depth, Index num_threads, bool l3_blocking) + { + this->m_mc = Transpose ? cols : rows; + this->m_nc = Transpose ? rows : cols; + this->m_kc = depth; + + if(l3_blocking) + { + computeProductBlockingSizes<LhsScalar,RhsScalar,KcFactor>(this->m_kc, this->m_mc, this->m_nc, num_threads); + } + else // no l3 blocking + { + Index n = this->m_nc; + computeProductBlockingSizes<LhsScalar,RhsScalar,KcFactor>(this->m_kc, this->m_mc, n, num_threads); + } + + m_sizeA = this->m_mc * this->m_kc; + m_sizeB = this->m_kc * this->m_nc; + } + + void initParallel(Index rows, Index cols, Index depth, Index num_threads) + { + this->m_mc = Transpose ? cols : rows; + this->m_nc = Transpose ? rows : cols; + this->m_kc = depth; + + eigen_internal_assert(this->m_blockA==0 && this->m_blockB==0); + Index m = this->m_mc; + computeProductBlockingSizes<LhsScalar,RhsScalar,KcFactor>(this->m_kc, m, this->m_nc, num_threads); + m_sizeA = this->m_mc * this->m_kc; + m_sizeB = this->m_kc * this->m_nc; + } + + void allocateA() + { + if(this->m_blockA==0) + this->m_blockA = aligned_new<LhsScalar>(m_sizeA); + } + + void allocateB() + { + if(this->m_blockB==0) + this->m_blockB = aligned_new<RhsScalar>(m_sizeB); + } + + void allocateAll() + { + allocateA(); + allocateB(); + } + + ~gemm_blocking_space() + { + aligned_delete(this->m_blockA, m_sizeA); + aligned_delete(this->m_blockB, m_sizeB); + } +}; + +} // end namespace internal + +namespace internal { + +template<typename Lhs, typename Rhs> +struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct> + : generic_product_impl_base<Lhs,Rhs,generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct> > +{ + typedef typename Product<Lhs,Rhs>::Scalar Scalar; + typedef typename Lhs::Scalar LhsScalar; + typedef typename Rhs::Scalar RhsScalar; + + typedef internal::blas_traits<Lhs> LhsBlasTraits; + typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType; + typedef typename internal::remove_all<ActualLhsType>::type ActualLhsTypeCleaned; + + typedef internal::blas_traits<Rhs> RhsBlasTraits; + typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType; + typedef typename internal::remove_all<ActualRhsType>::type ActualRhsTypeCleaned; + + enum { + MaxDepthAtCompileTime = EIGEN_SIZE_MIN_PREFER_FIXED(Lhs::MaxColsAtCompileTime,Rhs::MaxRowsAtCompileTime) + }; + + typedef generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,CoeffBasedProductMode> lazyproduct; + + template<typename Dst> + static void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { + if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0) + lazyproduct::evalTo(dst, lhs, rhs); + else + { + dst.setZero(); + scaleAndAddTo(dst, lhs, rhs, Scalar(1)); + } + } + + template<typename Dst> + static void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { + if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0) + lazyproduct::addTo(dst, lhs, rhs); + else + scaleAndAddTo(dst,lhs, rhs, Scalar(1)); + } + + template<typename Dst> + static void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { + if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0) + lazyproduct::subTo(dst, lhs, rhs); + else + scaleAndAddTo(dst, lhs, rhs, Scalar(-1)); + } + + template<typename Dest> + static void scaleAndAddTo(Dest& dst, const Lhs& a_lhs, const Rhs& a_rhs, const Scalar& alpha) + { + eigen_assert(dst.rows()==a_lhs.rows() && dst.cols()==a_rhs.cols()); + if(a_lhs.cols()==0 || a_lhs.rows()==0 || a_rhs.cols()==0) + return; + + typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(a_lhs); + typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(a_rhs); + + Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(a_lhs) + * RhsBlasTraits::extractScalarFactor(a_rhs); + + typedef internal::gemm_blocking_space<(Dest::Flags&RowMajorBit) ? RowMajor : ColMajor,LhsScalar,RhsScalar, + Dest::MaxRowsAtCompileTime,Dest::MaxColsAtCompileTime,MaxDepthAtCompileTime> BlockingType; + + typedef internal::gemm_functor< + Scalar, Index, + internal::general_matrix_matrix_product< + Index, + LhsScalar, (ActualLhsTypeCleaned::Flags&RowMajorBit) ? RowMajor : ColMajor, bool(LhsBlasTraits::NeedToConjugate), + RhsScalar, (ActualRhsTypeCleaned::Flags&RowMajorBit) ? RowMajor : ColMajor, bool(RhsBlasTraits::NeedToConjugate), + (Dest::Flags&RowMajorBit) ? RowMajor : ColMajor>, + ActualLhsTypeCleaned, ActualRhsTypeCleaned, Dest, BlockingType> GemmFunctor; + + BlockingType blocking(dst.rows(), dst.cols(), lhs.cols(), 1, true); + internal::parallelize_gemm<(Dest::MaxRowsAtCompileTime>32 || Dest::MaxRowsAtCompileTime==Dynamic)> + (GemmFunctor(lhs, rhs, dst, actualAlpha, blocking), a_lhs.rows(), a_rhs.cols(), a_lhs.cols(), Dest::Flags&RowMajorBit); + } +}; + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_GENERAL_MATRIX_MATRIX_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h new file mode 100644 index 000000000..e844e37d1 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h @@ -0,0 +1,311 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2009-2010 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_GENERAL_MATRIX_MATRIX_TRIANGULAR_H +#define EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_H + +namespace Eigen { + +template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool ConjLhs, bool ConjRhs> +struct selfadjoint_rank1_update; + +namespace internal { + +/********************************************************************** +* This file implements a general A * B product while +* evaluating only one triangular part of the product. +* This is a more general version of self adjoint product (C += A A^T) +* as the level 3 SYRK Blas routine. +**********************************************************************/ + +// forward declarations (defined at the end of this file) +template<typename LhsScalar, typename RhsScalar, typename Index, int mr, int nr, bool ConjLhs, bool ConjRhs, int UpLo> +struct tribb_kernel; + +/* Optimized matrix-matrix product evaluating only one triangular half */ +template <typename Index, + typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs, + typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs, + int ResStorageOrder, int UpLo, int Version = Specialized> +struct general_matrix_matrix_triangular_product; + +// as usual if the result is row major => we transpose the product +template <typename Index, typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs, + typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs, int UpLo, int Version> +struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,RowMajor,UpLo,Version> +{ + typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar; + static EIGEN_STRONG_INLINE void run(Index size, Index depth,const LhsScalar* lhs, Index lhsStride, + const RhsScalar* rhs, Index rhsStride, ResScalar* res, Index resStride, + const ResScalar& alpha, level3_blocking<RhsScalar,LhsScalar>& blocking) + { + general_matrix_matrix_triangular_product<Index, + RhsScalar, RhsStorageOrder==RowMajor ? ColMajor : RowMajor, ConjugateRhs, + LhsScalar, LhsStorageOrder==RowMajor ? ColMajor : RowMajor, ConjugateLhs, + ColMajor, UpLo==Lower?Upper:Lower> + ::run(size,depth,rhs,rhsStride,lhs,lhsStride,res,resStride,alpha,blocking); + } +}; + +template <typename Index, typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs, + typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs, int UpLo, int Version> +struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,ColMajor,UpLo,Version> +{ + typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar; + static EIGEN_STRONG_INLINE void run(Index size, Index depth,const LhsScalar* _lhs, Index lhsStride, + const RhsScalar* _rhs, Index rhsStride, ResScalar* _res, Index resStride, + const ResScalar& alpha, level3_blocking<LhsScalar,RhsScalar>& blocking) + { + typedef gebp_traits<LhsScalar,RhsScalar> Traits; + + typedef const_blas_data_mapper<LhsScalar, Index, LhsStorageOrder> LhsMapper; + typedef const_blas_data_mapper<RhsScalar, Index, RhsStorageOrder> RhsMapper; + typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper; + LhsMapper lhs(_lhs,lhsStride); + RhsMapper rhs(_rhs,rhsStride); + ResMapper res(_res, resStride); + + Index kc = blocking.kc(); + Index mc = (std::min)(size,blocking.mc()); + + // !!! mc must be a multiple of nr: + if(mc > Traits::nr) + mc = (mc/Traits::nr)*Traits::nr; + + std::size_t sizeA = kc*mc; + std::size_t sizeB = kc*size; + + ei_declare_aligned_stack_constructed_variable(LhsScalar, blockA, sizeA, blocking.blockA()); + ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, sizeB, blocking.blockB()); + + gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs; + gemm_pack_rhs<RhsScalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs; + gebp_kernel<LhsScalar, RhsScalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp; + tribb_kernel<LhsScalar, RhsScalar, Index, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs, UpLo> sybb; + + for(Index k2=0; k2<depth; k2+=kc) + { + const Index actual_kc = (std::min)(k2+kc,depth)-k2; + + // note that the actual rhs is the transpose/adjoint of mat + pack_rhs(blockB, rhs.getSubMapper(k2,0), actual_kc, size); + + for(Index i2=0; i2<size; i2+=mc) + { + const Index actual_mc = (std::min)(i2+mc,size)-i2; + + pack_lhs(blockA, lhs.getSubMapper(i2, k2), actual_kc, actual_mc); + + // the selected actual_mc * size panel of res is split into three different part: + // 1 - before the diagonal => processed with gebp or skipped + // 2 - the actual_mc x actual_mc symmetric block => processed with a special kernel + // 3 - after the diagonal => processed with gebp or skipped + if (UpLo==Lower) + gebp(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc, + (std::min)(size,i2), alpha, -1, -1, 0, 0); + + + sybb(_res+resStride*i2 + i2, resStride, blockA, blockB + actual_kc*i2, actual_mc, actual_kc, alpha); + + if (UpLo==Upper) + { + Index j2 = i2+actual_mc; + gebp(res.getSubMapper(i2, j2), blockA, blockB+actual_kc*j2, actual_mc, + actual_kc, (std::max)(Index(0), size-j2), alpha, -1, -1, 0, 0); + } + } + } + } +}; + +// Optimized packed Block * packed Block product kernel evaluating only one given triangular part +// This kernel is built on top of the gebp kernel: +// - the current destination block is processed per panel of actual_mc x BlockSize +// where BlockSize is set to the minimal value allowing gebp to be as fast as possible +// - then, as usual, each panel is split into three parts along the diagonal, +// the sub blocks above and below the diagonal are processed as usual, +// while the triangular block overlapping the diagonal is evaluated into a +// small temporary buffer which is then accumulated into the result using a +// triangular traversal. +template<typename LhsScalar, typename RhsScalar, typename Index, int mr, int nr, bool ConjLhs, bool ConjRhs, int UpLo> +struct tribb_kernel +{ + typedef gebp_traits<LhsScalar,RhsScalar,ConjLhs,ConjRhs> Traits; + typedef typename Traits::ResScalar ResScalar; + + enum { + BlockSize = meta_least_common_multiple<EIGEN_PLAIN_ENUM_MAX(mr,nr),EIGEN_PLAIN_ENUM_MIN(mr,nr)>::ret + }; + void operator()(ResScalar* _res, Index resStride, const LhsScalar* blockA, const RhsScalar* blockB, Index size, Index depth, const ResScalar& alpha) + { + typedef blas_data_mapper<ResScalar, Index, ColMajor> ResMapper; + ResMapper res(_res, resStride); + gebp_kernel<LhsScalar, RhsScalar, Index, ResMapper, mr, nr, ConjLhs, ConjRhs> gebp_kernel; + + Matrix<ResScalar,BlockSize,BlockSize,ColMajor> buffer((internal::constructor_without_unaligned_array_assert())); + + // let's process the block per panel of actual_mc x BlockSize, + // again, each is split into three parts, etc. + for (Index j=0; j<size; j+=BlockSize) + { + Index actualBlockSize = std::min<Index>(BlockSize,size - j); + const RhsScalar* actual_b = blockB+j*depth; + + if(UpLo==Upper) + gebp_kernel(res.getSubMapper(0, j), blockA, actual_b, j, depth, actualBlockSize, alpha, + -1, -1, 0, 0); + + // selfadjoint micro block + { + Index i = j; + buffer.setZero(); + // 1 - apply the kernel on the temporary buffer + gebp_kernel(ResMapper(buffer.data(), BlockSize), blockA+depth*i, actual_b, actualBlockSize, depth, actualBlockSize, alpha, + -1, -1, 0, 0); + // 2 - triangular accumulation + for(Index j1=0; j1<actualBlockSize; ++j1) + { + ResScalar* r = &res(i, j + j1); + for(Index i1=UpLo==Lower ? j1 : 0; + UpLo==Lower ? i1<actualBlockSize : i1<=j1; ++i1) + r[i1] += buffer(i1,j1); + } + } + + if(UpLo==Lower) + { + Index i = j+actualBlockSize; + gebp_kernel(res.getSubMapper(i, j), blockA+depth*i, actual_b, size-i, + depth, actualBlockSize, alpha, -1, -1, 0, 0); + } + } + } +}; + +} // end namespace internal + +// high level API + +template<typename MatrixType, typename ProductType, int UpLo, bool IsOuterProduct> +struct general_product_to_triangular_selector; + + +template<typename MatrixType, typename ProductType, int UpLo> +struct general_product_to_triangular_selector<MatrixType,ProductType,UpLo,true> +{ + static void run(MatrixType& mat, const ProductType& prod, const typename MatrixType::Scalar& alpha, bool beta) + { + typedef typename MatrixType::Scalar Scalar; + + typedef typename internal::remove_all<typename ProductType::LhsNested>::type Lhs; + typedef internal::blas_traits<Lhs> LhsBlasTraits; + typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhs; + typedef typename internal::remove_all<ActualLhs>::type _ActualLhs; + typename internal::add_const_on_value_type<ActualLhs>::type actualLhs = LhsBlasTraits::extract(prod.lhs()); + + typedef typename internal::remove_all<typename ProductType::RhsNested>::type Rhs; + typedef internal::blas_traits<Rhs> RhsBlasTraits; + typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhs; + typedef typename internal::remove_all<ActualRhs>::type _ActualRhs; + typename internal::add_const_on_value_type<ActualRhs>::type actualRhs = RhsBlasTraits::extract(prod.rhs()); + + Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs().derived()) * RhsBlasTraits::extractScalarFactor(prod.rhs().derived()); + + if(!beta) + mat.template triangularView<UpLo>().setZero(); + + enum { + StorageOrder = (internal::traits<MatrixType>::Flags&RowMajorBit) ? RowMajor : ColMajor, + UseLhsDirectly = _ActualLhs::InnerStrideAtCompileTime==1, + UseRhsDirectly = _ActualRhs::InnerStrideAtCompileTime==1 + }; + + internal::gemv_static_vector_if<Scalar,Lhs::SizeAtCompileTime,Lhs::MaxSizeAtCompileTime,!UseLhsDirectly> static_lhs; + ei_declare_aligned_stack_constructed_variable(Scalar, actualLhsPtr, actualLhs.size(), + (UseLhsDirectly ? const_cast<Scalar*>(actualLhs.data()) : static_lhs.data())); + if(!UseLhsDirectly) Map<typename _ActualLhs::PlainObject>(actualLhsPtr, actualLhs.size()) = actualLhs; + + internal::gemv_static_vector_if<Scalar,Rhs::SizeAtCompileTime,Rhs::MaxSizeAtCompileTime,!UseRhsDirectly> static_rhs; + ei_declare_aligned_stack_constructed_variable(Scalar, actualRhsPtr, actualRhs.size(), + (UseRhsDirectly ? const_cast<Scalar*>(actualRhs.data()) : static_rhs.data())); + if(!UseRhsDirectly) Map<typename _ActualRhs::PlainObject>(actualRhsPtr, actualRhs.size()) = actualRhs; + + + selfadjoint_rank1_update<Scalar,Index,StorageOrder,UpLo, + LhsBlasTraits::NeedToConjugate && NumTraits<Scalar>::IsComplex, + RhsBlasTraits::NeedToConjugate && NumTraits<Scalar>::IsComplex> + ::run(actualLhs.size(), mat.data(), mat.outerStride(), actualLhsPtr, actualRhsPtr, actualAlpha); + } +}; + +template<typename MatrixType, typename ProductType, int UpLo> +struct general_product_to_triangular_selector<MatrixType,ProductType,UpLo,false> +{ + static void run(MatrixType& mat, const ProductType& prod, const typename MatrixType::Scalar& alpha, bool beta) + { + typedef typename internal::remove_all<typename ProductType::LhsNested>::type Lhs; + typedef internal::blas_traits<Lhs> LhsBlasTraits; + typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhs; + typedef typename internal::remove_all<ActualLhs>::type _ActualLhs; + typename internal::add_const_on_value_type<ActualLhs>::type actualLhs = LhsBlasTraits::extract(prod.lhs()); + + typedef typename internal::remove_all<typename ProductType::RhsNested>::type Rhs; + typedef internal::blas_traits<Rhs> RhsBlasTraits; + typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhs; + typedef typename internal::remove_all<ActualRhs>::type _ActualRhs; + typename internal::add_const_on_value_type<ActualRhs>::type actualRhs = RhsBlasTraits::extract(prod.rhs()); + + typename ProductType::Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs().derived()) * RhsBlasTraits::extractScalarFactor(prod.rhs().derived()); + + if(!beta) + mat.template triangularView<UpLo>().setZero(); + + enum { + IsRowMajor = (internal::traits<MatrixType>::Flags&RowMajorBit) ? 1 : 0, + LhsIsRowMajor = _ActualLhs::Flags&RowMajorBit ? 1 : 0, + RhsIsRowMajor = _ActualRhs::Flags&RowMajorBit ? 1 : 0, + SkipDiag = (UpLo&(UnitDiag|ZeroDiag))!=0 + }; + + Index size = mat.cols(); + if(SkipDiag) + size--; + Index depth = actualLhs.cols(); + + typedef internal::gemm_blocking_space<IsRowMajor ? RowMajor : ColMajor,typename Lhs::Scalar,typename Rhs::Scalar, + MatrixType::MaxColsAtCompileTime, MatrixType::MaxColsAtCompileTime, _ActualRhs::MaxColsAtCompileTime> BlockingType; + + BlockingType blocking(size, size, depth, 1, false); + + internal::general_matrix_matrix_triangular_product<Index, + typename Lhs::Scalar, LhsIsRowMajor ? RowMajor : ColMajor, LhsBlasTraits::NeedToConjugate, + typename Rhs::Scalar, RhsIsRowMajor ? RowMajor : ColMajor, RhsBlasTraits::NeedToConjugate, + IsRowMajor ? RowMajor : ColMajor, UpLo&(Lower|Upper)> + ::run(size, depth, + &actualLhs.coeffRef(SkipDiag&&(UpLo&Lower)==Lower ? 1 : 0,0), actualLhs.outerStride(), + &actualRhs.coeffRef(0,SkipDiag&&(UpLo&Upper)==Upper ? 1 : 0), actualRhs.outerStride(), + mat.data() + (SkipDiag ? (bool(IsRowMajor) != ((UpLo&Lower)==Lower) ? 1 : mat.outerStride() ) : 0), mat.outerStride(), actualAlpha, blocking); + } +}; + +template<typename MatrixType, unsigned int UpLo> +template<typename ProductType> +TriangularView<MatrixType,UpLo>& TriangularViewImpl<MatrixType,UpLo,Dense>::_assignProduct(const ProductType& prod, const Scalar& alpha, bool beta) +{ + EIGEN_STATIC_ASSERT((UpLo&UnitDiag)==0, WRITING_TO_TRIANGULAR_PART_WITH_UNIT_DIAGONAL_IS_NOT_SUPPORTED); + eigen_assert(derived().nestedExpression().rows() == prod.rows() && derived().cols() == prod.cols()); + + general_product_to_triangular_selector<MatrixType, ProductType, UpLo, internal::traits<ProductType>::InnerSize==1>::run(derived().nestedExpression().const_cast_derived(), prod, alpha, beta); + + return derived(); +} + +} // end namespace Eigen + +#endif // EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h new file mode 100644 index 000000000..41e18ff07 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h @@ -0,0 +1,141 @@ +/* + Copyright (c) 2011, Intel Corporation. All rights reserved. + + Redistribution and use in source and binary forms, with or without modification, + are permitted provided that the following conditions are met: + + * Redistributions of source code must retain the above copyright notice, this + list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above copyright notice, + this list of conditions and the following disclaimer in the documentation + and/or other materials provided with the distribution. + * Neither the name of Intel Corporation nor the names of its contributors may + be used to endorse or promote products derived from this software without + specific prior written permission. + + THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND + ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED + WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE + DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR + ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES + (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; + LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON + ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS + SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + + ******************************************************************************** + * Content : Eigen bindings to BLAS F77 + * Level 3 BLAS SYRK/HERK implementation. + ******************************************************************************** +*/ + +#ifndef EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_BLAS_H +#define EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_BLAS_H + +namespace Eigen { + +namespace internal { + +template <typename Index, typename Scalar, int AStorageOrder, bool ConjugateA, int ResStorageOrder, int UpLo> +struct general_matrix_matrix_rankupdate : + general_matrix_matrix_triangular_product< + Index,Scalar,AStorageOrder,ConjugateA,Scalar,AStorageOrder,ConjugateA,ResStorageOrder,UpLo,BuiltIn> {}; + + +// try to go to BLAS specialization +#define EIGEN_BLAS_RANKUPDATE_SPECIALIZE(Scalar) \ +template <typename Index, int LhsStorageOrder, bool ConjugateLhs, \ + int RhsStorageOrder, bool ConjugateRhs, int UpLo> \ +struct general_matrix_matrix_triangular_product<Index,Scalar,LhsStorageOrder,ConjugateLhs, \ + Scalar,RhsStorageOrder,ConjugateRhs,ColMajor,UpLo,Specialized> { \ + static EIGEN_STRONG_INLINE void run(Index size, Index depth,const Scalar* lhs, Index lhsStride, \ + const Scalar* rhs, Index rhsStride, Scalar* res, Index resStride, Scalar alpha, level3_blocking<Scalar, Scalar>& blocking) \ + { \ + if ( lhs==rhs && ((UpLo&(Lower|Upper)==UpLo)) ) { \ + general_matrix_matrix_rankupdate<Index,Scalar,LhsStorageOrder,ConjugateLhs,ColMajor,UpLo> \ + ::run(size,depth,lhs,lhsStride,rhs,rhsStride,res,resStride,alpha,blocking); \ + } else { \ + general_matrix_matrix_triangular_product<Index, \ + Scalar, LhsStorageOrder, ConjugateLhs, \ + Scalar, RhsStorageOrder, ConjugateRhs, \ + ColMajor, UpLo, BuiltIn> \ + ::run(size,depth,lhs,lhsStride,rhs,rhsStride,res,resStride,alpha,blocking); \ + } \ + } \ +}; + +EIGEN_BLAS_RANKUPDATE_SPECIALIZE(double) +EIGEN_BLAS_RANKUPDATE_SPECIALIZE(float) +// TODO handle complex cases +// EIGEN_BLAS_RANKUPDATE_SPECIALIZE(dcomplex) +// EIGEN_BLAS_RANKUPDATE_SPECIALIZE(scomplex) + +// SYRK for float/double +#define EIGEN_BLAS_RANKUPDATE_R(EIGTYPE, BLASTYPE, BLASFUNC) \ +template <typename Index, int AStorageOrder, bool ConjugateA, int UpLo> \ +struct general_matrix_matrix_rankupdate<Index,EIGTYPE,AStorageOrder,ConjugateA,ColMajor,UpLo> { \ + enum { \ + IsLower = (UpLo&Lower) == Lower, \ + LowUp = IsLower ? Lower : Upper, \ + conjA = ((AStorageOrder==ColMajor) && ConjugateA) ? 1 : 0 \ + }; \ + static EIGEN_STRONG_INLINE void run(Index size, Index depth,const EIGTYPE* lhs, Index lhsStride, \ + const EIGTYPE* /*rhs*/, Index /*rhsStride*/, EIGTYPE* res, Index resStride, EIGTYPE alpha, level3_blocking<EIGTYPE, EIGTYPE>& /*blocking*/) \ + { \ + /* typedef Matrix<EIGTYPE, Dynamic, Dynamic, RhsStorageOrder> MatrixRhs;*/ \ +\ + BlasIndex lda=convert_index<BlasIndex>(lhsStride), ldc=convert_index<BlasIndex>(resStride), n=convert_index<BlasIndex>(size), k=convert_index<BlasIndex>(depth); \ + char uplo=((IsLower) ? 'L' : 'U'), trans=((AStorageOrder==RowMajor) ? 'T':'N'); \ + EIGTYPE beta(1); \ + BLASFUNC(&uplo, &trans, &n, &k, &numext::real_ref(alpha), lhs, &lda, &numext::real_ref(beta), res, &ldc); \ + } \ +}; + +// HERK for complex data +#define EIGEN_BLAS_RANKUPDATE_C(EIGTYPE, BLASTYPE, RTYPE, BLASFUNC) \ +template <typename Index, int AStorageOrder, bool ConjugateA, int UpLo> \ +struct general_matrix_matrix_rankupdate<Index,EIGTYPE,AStorageOrder,ConjugateA,ColMajor,UpLo> { \ + enum { \ + IsLower = (UpLo&Lower) == Lower, \ + LowUp = IsLower ? Lower : Upper, \ + conjA = (((AStorageOrder==ColMajor) && ConjugateA) || ((AStorageOrder==RowMajor) && !ConjugateA)) ? 1 : 0 \ + }; \ + static EIGEN_STRONG_INLINE void run(Index size, Index depth,const EIGTYPE* lhs, Index lhsStride, \ + const EIGTYPE* /*rhs*/, Index /*rhsStride*/, EIGTYPE* res, Index resStride, EIGTYPE alpha, level3_blocking<EIGTYPE, EIGTYPE>& /*blocking*/) \ + { \ + typedef Matrix<EIGTYPE, Dynamic, Dynamic, AStorageOrder> MatrixType; \ +\ + BlasIndex lda=convert_index<BlasIndex>(lhsStride), ldc=convert_index<BlasIndex>(resStride), n=convert_index<BlasIndex>(size), k=convert_index<BlasIndex>(depth); \ + char uplo=((IsLower) ? 'L' : 'U'), trans=((AStorageOrder==RowMajor) ? 'C':'N'); \ + RTYPE alpha_, beta_; \ + const EIGTYPE* a_ptr; \ +\ + alpha_ = alpha.real(); \ + beta_ = 1.0; \ +/* Copy with conjugation in some cases*/ \ + MatrixType a; \ + if (conjA) { \ + Map<const MatrixType, 0, OuterStride<> > mapA(lhs,n,k,OuterStride<>(lhsStride)); \ + a = mapA.conjugate(); \ + lda = a.outerStride(); \ + a_ptr = a.data(); \ + } else a_ptr=lhs; \ + BLASFUNC(&uplo, &trans, &n, &k, &alpha_, (BLASTYPE*)a_ptr, &lda, &beta_, (BLASTYPE*)res, &ldc); \ + } \ +}; + + +EIGEN_BLAS_RANKUPDATE_R(double, double, dsyrk_) +EIGEN_BLAS_RANKUPDATE_R(float, float, ssyrk_) + +// TODO hanlde complex cases +// EIGEN_BLAS_RANKUPDATE_C(dcomplex, double, double, zherk_) +// EIGEN_BLAS_RANKUPDATE_C(scomplex, float, float, cherk_) + + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_BLAS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixMatrix_BLAS.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixMatrix_BLAS.h new file mode 100644 index 000000000..7a3bdbf20 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixMatrix_BLAS.h @@ -0,0 +1,115 @@ +/* + Copyright (c) 2011, Intel Corporation. All rights reserved. + + Redistribution and use in source and binary forms, with or without modification, + are permitted provided that the following conditions are met: + + * Redistributions of source code must retain the above copyright notice, this + list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above copyright notice, + this list of conditions and the following disclaimer in the documentation + and/or other materials provided with the distribution. + * Neither the name of Intel Corporation nor the names of its contributors may + be used to endorse or promote products derived from this software without + specific prior written permission. + + THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND + ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED + WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE + DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR + ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES + (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; + LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON + ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS + SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + + ******************************************************************************** + * Content : Eigen bindings to BLAS F77 + * General matrix-matrix product functionality based on ?GEMM. + ******************************************************************************** +*/ + +#ifndef EIGEN_GENERAL_MATRIX_MATRIX_BLAS_H +#define EIGEN_GENERAL_MATRIX_MATRIX_BLAS_H + +namespace Eigen { + +namespace internal { + +/********************************************************************** +* This file implements general matrix-matrix multiplication using BLAS +* gemm function via partial specialization of +* general_matrix_matrix_product::run(..) method for float, double, +* std::complex<float> and std::complex<double> types +**********************************************************************/ + +// gemm specialization + +#define GEMM_SPECIALIZATION(EIGTYPE, EIGPREFIX, BLASTYPE, BLASPREFIX) \ +template< \ + typename Index, \ + int LhsStorageOrder, bool ConjugateLhs, \ + int RhsStorageOrder, bool ConjugateRhs> \ +struct general_matrix_matrix_product<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,RhsStorageOrder,ConjugateRhs,ColMajor> \ +{ \ +typedef gebp_traits<EIGTYPE,EIGTYPE> Traits; \ +\ +static void run(Index rows, Index cols, Index depth, \ + const EIGTYPE* _lhs, Index lhsStride, \ + const EIGTYPE* _rhs, Index rhsStride, \ + EIGTYPE* res, Index resStride, \ + EIGTYPE alpha, \ + level3_blocking<EIGTYPE, EIGTYPE>& /*blocking*/, \ + GemmParallelInfo<Index>* /*info = 0*/) \ +{ \ + using std::conj; \ +\ + char transa, transb; \ + BlasIndex m, n, k, lda, ldb, ldc; \ + const EIGTYPE *a, *b; \ + EIGTYPE beta(1); \ + MatrixX##EIGPREFIX a_tmp, b_tmp; \ +\ +/* Set transpose options */ \ + transa = (LhsStorageOrder==RowMajor) ? ((ConjugateLhs) ? 'C' : 'T') : 'N'; \ + transb = (RhsStorageOrder==RowMajor) ? ((ConjugateRhs) ? 'C' : 'T') : 'N'; \ +\ +/* Set m, n, k */ \ + m = convert_index<BlasIndex>(rows); \ + n = convert_index<BlasIndex>(cols); \ + k = convert_index<BlasIndex>(depth); \ +\ +/* Set lda, ldb, ldc */ \ + lda = convert_index<BlasIndex>(lhsStride); \ + ldb = convert_index<BlasIndex>(rhsStride); \ + ldc = convert_index<BlasIndex>(resStride); \ +\ +/* Set a, b, c */ \ + if ((LhsStorageOrder==ColMajor) && (ConjugateLhs)) { \ + Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > lhs(_lhs,m,k,OuterStride<>(lhsStride)); \ + a_tmp = lhs.conjugate(); \ + a = a_tmp.data(); \ + lda = convert_index<BlasIndex>(a_tmp.outerStride()); \ + } else a = _lhs; \ +\ + if ((RhsStorageOrder==ColMajor) && (ConjugateRhs)) { \ + Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > rhs(_rhs,k,n,OuterStride<>(rhsStride)); \ + b_tmp = rhs.conjugate(); \ + b = b_tmp.data(); \ + ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \ + } else b = _rhs; \ +\ + BLASPREFIX##gemm_(&transa, &transb, &m, &n, &k, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, &numext::real_ref(beta), (BLASTYPE*)res, &ldc); \ +}}; + +GEMM_SPECIALIZATION(double, d, double, d) +GEMM_SPECIALIZATION(float, f, float, s) +GEMM_SPECIALIZATION(dcomplex, cd, double, z) +GEMM_SPECIALIZATION(scomplex, cf, float, c) + +} // end namespase internal + +} // end namespace Eigen + +#endif // EIGEN_GENERAL_MATRIX_MATRIX_BLAS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixVector.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixVector.h new file mode 100644 index 000000000..3c1a7fc40 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixVector.h @@ -0,0 +1,619 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2009 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_GENERAL_MATRIX_VECTOR_H +#define EIGEN_GENERAL_MATRIX_VECTOR_H + +namespace Eigen { + +namespace internal { + +/* Optimized col-major matrix * vector product: + * This algorithm processes 4 columns at onces that allows to both reduce + * the number of load/stores of the result by a factor 4 and to reduce + * the instruction dependency. Moreover, we know that all bands have the + * same alignment pattern. + * + * Mixing type logic: C += alpha * A * B + * | A | B |alpha| comments + * |real |cplx |cplx | no vectorization + * |real |cplx |real | alpha is converted to a cplx when calling the run function, no vectorization + * |cplx |real |cplx | invalid, the caller has to do tmp: = A * B; C += alpha*tmp + * |cplx |real |real | optimal case, vectorization possible via real-cplx mul + * + * Accesses to the matrix coefficients follow the following logic: + * + * - if all columns have the same alignment then + * - if the columns have the same alignment as the result vector, then easy! (-> AllAligned case) + * - otherwise perform unaligned loads only (-> NoneAligned case) + * - otherwise + * - if even columns have the same alignment then + * // odd columns are guaranteed to have the same alignment too + * - if even or odd columns have the same alignment as the result, then + * // for a register size of 2 scalars, this is guarantee to be the case (e.g., SSE with double) + * - perform half aligned and half unaligned loads (-> EvenAligned case) + * - otherwise perform unaligned loads only (-> NoneAligned case) + * - otherwise, if the register size is 4 scalars (e.g., SSE with float) then + * - one over 4 consecutive columns is guaranteed to be aligned with the result vector, + * perform simple aligned loads for this column and aligned loads plus re-alignment for the other. (-> FirstAligned case) + * // this re-alignment is done by the palign function implemented for SSE in Eigen/src/Core/arch/SSE/PacketMath.h + * - otherwise, + * // if we get here, this means the register size is greater than 4 (e.g., AVX with floats), + * // we currently fall back to the NoneAligned case + * + * The same reasoning apply for the transposed case. + * + * The last case (PacketSize>4) could probably be improved by generalizing the FirstAligned case, but since we do not support AVX yet... + * One might also wonder why in the EvenAligned case we perform unaligned loads instead of using the aligned-loads plus re-alignment + * strategy as in the FirstAligned case. The reason is that we observed that unaligned loads on a 8 byte boundary are not too slow + * compared to unaligned loads on a 4 byte boundary. + * + */ +template<typename Index, typename LhsScalar, typename LhsMapper, bool ConjugateLhs, typename RhsScalar, typename RhsMapper, bool ConjugateRhs, int Version> +struct general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,ConjugateLhs,RhsScalar,RhsMapper,ConjugateRhs,Version> +{ + typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar; + +enum { + Vectorizable = packet_traits<LhsScalar>::Vectorizable && packet_traits<RhsScalar>::Vectorizable + && int(packet_traits<LhsScalar>::size)==int(packet_traits<RhsScalar>::size), + LhsPacketSize = Vectorizable ? packet_traits<LhsScalar>::size : 1, + RhsPacketSize = Vectorizable ? packet_traits<RhsScalar>::size : 1, + ResPacketSize = Vectorizable ? packet_traits<ResScalar>::size : 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; + +EIGEN_DONT_INLINE static void run( + Index rows, Index cols, + const LhsMapper& lhs, + const RhsMapper& rhs, + ResScalar* res, Index resIncr, + RhsScalar alpha); +}; + +template<typename Index, typename LhsScalar, typename LhsMapper, bool ConjugateLhs, typename RhsScalar, typename RhsMapper, bool ConjugateRhs, int Version> +EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,ConjugateLhs,RhsScalar,RhsMapper,ConjugateRhs,Version>::run( + Index rows, Index cols, + const LhsMapper& lhs, + const RhsMapper& rhs, + ResScalar* res, Index resIncr, + RhsScalar alpha) +{ + EIGEN_UNUSED_VARIABLE(resIncr); + eigen_internal_assert(resIncr==1); + #ifdef _EIGEN_ACCUMULATE_PACKETS + #error _EIGEN_ACCUMULATE_PACKETS has already been defined + #endif + #define _EIGEN_ACCUMULATE_PACKETS(Alignment0,Alignment13,Alignment2) \ + pstore(&res[j], \ + padd(pload<ResPacket>(&res[j]), \ + padd( \ + padd(pcj.pmul(lhs0.template load<LhsPacket, Alignment0>(j), ptmp0), \ + pcj.pmul(lhs1.template load<LhsPacket, Alignment13>(j), ptmp1)), \ + padd(pcj.pmul(lhs2.template load<LhsPacket, Alignment2>(j), ptmp2), \ + pcj.pmul(lhs3.template load<LhsPacket, Alignment13>(j), ptmp3)) ))) + + typedef typename LhsMapper::VectorMapper LhsScalars; + + conj_helper<LhsScalar,RhsScalar,ConjugateLhs,ConjugateRhs> cj; + conj_helper<LhsPacket,RhsPacket,ConjugateLhs,ConjugateRhs> pcj; + if(ConjugateRhs) + alpha = numext::conj(alpha); + + enum { AllAligned = 0, EvenAligned, FirstAligned, NoneAligned }; + const Index columnsAtOnce = 4; + const Index peels = 2; + const Index LhsPacketAlignedMask = LhsPacketSize-1; + const Index ResPacketAlignedMask = ResPacketSize-1; +// const Index PeelAlignedMask = ResPacketSize*peels-1; + const Index size = rows; + + const Index lhsStride = lhs.stride(); + + // How many coeffs of the result do we have to skip to be aligned. + // Here we assume data are at least aligned on the base scalar type. + Index alignedStart = internal::first_default_aligned(res,size); + Index alignedSize = ResPacketSize>1 ? alignedStart + ((size-alignedStart) & ~ResPacketAlignedMask) : 0; + const Index peeledSize = alignedSize - RhsPacketSize*peels - RhsPacketSize + 1; + + const Index alignmentStep = LhsPacketSize>1 ? (LhsPacketSize - lhsStride % LhsPacketSize) & LhsPacketAlignedMask : 0; + Index alignmentPattern = alignmentStep==0 ? AllAligned + : alignmentStep==(LhsPacketSize/2) ? EvenAligned + : FirstAligned; + + // we cannot assume the first element is aligned because of sub-matrices + const Index lhsAlignmentOffset = lhs.firstAligned(size); + + // find how many columns do we have to skip to be aligned with the result (if possible) + Index skipColumns = 0; + // if the data cannot be aligned (TODO add some compile time tests when possible, e.g. for floats) + if( (lhsAlignmentOffset < 0) || (lhsAlignmentOffset == size) || (UIntPtr(res)%sizeof(ResScalar)) ) + { + alignedSize = 0; + alignedStart = 0; + alignmentPattern = NoneAligned; + } + else if(LhsPacketSize > 4) + { + // TODO: extend the code to support aligned loads whenever possible when LhsPacketSize > 4. + // Currently, it seems to be better to perform unaligned loads anyway + alignmentPattern = NoneAligned; + } + else if (LhsPacketSize>1) + { + // eigen_internal_assert(size_t(firstLhs+lhsAlignmentOffset)%sizeof(LhsPacket)==0 || size<LhsPacketSize); + + while (skipColumns<LhsPacketSize && + alignedStart != ((lhsAlignmentOffset + alignmentStep*skipColumns)%LhsPacketSize)) + ++skipColumns; + if (skipColumns==LhsPacketSize) + { + // nothing can be aligned, no need to skip any column + alignmentPattern = NoneAligned; + skipColumns = 0; + } + else + { + skipColumns = (std::min)(skipColumns,cols); + // note that the skiped columns are processed later. + } + + /* eigen_internal_assert( (alignmentPattern==NoneAligned) + || (skipColumns + columnsAtOnce >= cols) + || LhsPacketSize > size + || (size_t(firstLhs+alignedStart+lhsStride*skipColumns)%sizeof(LhsPacket))==0);*/ + } + else if(Vectorizable) + { + alignedStart = 0; + alignedSize = size; + alignmentPattern = AllAligned; + } + + const Index offset1 = (FirstAligned && alignmentStep==1)?3:1; + const Index offset3 = (FirstAligned && alignmentStep==1)?1:3; + + Index columnBound = ((cols-skipColumns)/columnsAtOnce)*columnsAtOnce + skipColumns; + for (Index i=skipColumns; i<columnBound; i+=columnsAtOnce) + { + RhsPacket ptmp0 = pset1<RhsPacket>(alpha*rhs(i, 0)), + ptmp1 = pset1<RhsPacket>(alpha*rhs(i+offset1, 0)), + ptmp2 = pset1<RhsPacket>(alpha*rhs(i+2, 0)), + ptmp3 = pset1<RhsPacket>(alpha*rhs(i+offset3, 0)); + + // this helps a lot generating better binary code + const LhsScalars lhs0 = lhs.getVectorMapper(0, i+0), lhs1 = lhs.getVectorMapper(0, i+offset1), + lhs2 = lhs.getVectorMapper(0, i+2), lhs3 = lhs.getVectorMapper(0, i+offset3); + + if (Vectorizable) + { + /* explicit vectorization */ + // process initial unaligned coeffs + for (Index j=0; j<alignedStart; ++j) + { + res[j] = cj.pmadd(lhs0(j), pfirst(ptmp0), res[j]); + res[j] = cj.pmadd(lhs1(j), pfirst(ptmp1), res[j]); + res[j] = cj.pmadd(lhs2(j), pfirst(ptmp2), res[j]); + res[j] = cj.pmadd(lhs3(j), pfirst(ptmp3), res[j]); + } + + if (alignedSize>alignedStart) + { + switch(alignmentPattern) + { + case AllAligned: + for (Index j = alignedStart; j<alignedSize; j+=ResPacketSize) + _EIGEN_ACCUMULATE_PACKETS(Aligned,Aligned,Aligned); + break; + case EvenAligned: + for (Index j = alignedStart; j<alignedSize; j+=ResPacketSize) + _EIGEN_ACCUMULATE_PACKETS(Aligned,Unaligned,Aligned); + break; + case FirstAligned: + { + Index j = alignedStart; + if(peels>1) + { + LhsPacket A00, A01, A02, A03, A10, A11, A12, A13; + ResPacket T0, T1; + + A01 = lhs1.template load<LhsPacket, Aligned>(alignedStart-1); + A02 = lhs2.template load<LhsPacket, Aligned>(alignedStart-2); + A03 = lhs3.template load<LhsPacket, Aligned>(alignedStart-3); + + for (; j<peeledSize; j+=peels*ResPacketSize) + { + A11 = lhs1.template load<LhsPacket, Aligned>(j-1+LhsPacketSize); palign<1>(A01,A11); + A12 = lhs2.template load<LhsPacket, Aligned>(j-2+LhsPacketSize); palign<2>(A02,A12); + A13 = lhs3.template load<LhsPacket, Aligned>(j-3+LhsPacketSize); palign<3>(A03,A13); + + A00 = lhs0.template load<LhsPacket, Aligned>(j); + A10 = lhs0.template load<LhsPacket, Aligned>(j+LhsPacketSize); + T0 = pcj.pmadd(A00, ptmp0, pload<ResPacket>(&res[j])); + T1 = pcj.pmadd(A10, ptmp0, pload<ResPacket>(&res[j+ResPacketSize])); + + T0 = pcj.pmadd(A01, ptmp1, T0); + A01 = lhs1.template load<LhsPacket, Aligned>(j-1+2*LhsPacketSize); palign<1>(A11,A01); + T0 = pcj.pmadd(A02, ptmp2, T0); + A02 = lhs2.template load<LhsPacket, Aligned>(j-2+2*LhsPacketSize); palign<2>(A12,A02); + T0 = pcj.pmadd(A03, ptmp3, T0); + pstore(&res[j],T0); + A03 = lhs3.template load<LhsPacket, Aligned>(j-3+2*LhsPacketSize); palign<3>(A13,A03); + T1 = pcj.pmadd(A11, ptmp1, T1); + T1 = pcj.pmadd(A12, ptmp2, T1); + T1 = pcj.pmadd(A13, ptmp3, T1); + pstore(&res[j+ResPacketSize],T1); + } + } + for (; j<alignedSize; j+=ResPacketSize) + _EIGEN_ACCUMULATE_PACKETS(Aligned,Unaligned,Unaligned); + break; + } + default: + for (Index j = alignedStart; j<alignedSize; j+=ResPacketSize) + _EIGEN_ACCUMULATE_PACKETS(Unaligned,Unaligned,Unaligned); + break; + } + } + } // end explicit vectorization + + /* process remaining coeffs (or all if there is no explicit vectorization) */ + for (Index j=alignedSize; j<size; ++j) + { + res[j] = cj.pmadd(lhs0(j), pfirst(ptmp0), res[j]); + res[j] = cj.pmadd(lhs1(j), pfirst(ptmp1), res[j]); + res[j] = cj.pmadd(lhs2(j), pfirst(ptmp2), res[j]); + res[j] = cj.pmadd(lhs3(j), pfirst(ptmp3), res[j]); + } + } + + // process remaining first and last columns (at most columnsAtOnce-1) + Index end = cols; + Index start = columnBound; + do + { + for (Index k=start; k<end; ++k) + { + RhsPacket ptmp0 = pset1<RhsPacket>(alpha*rhs(k, 0)); + const LhsScalars lhs0 = lhs.getVectorMapper(0, k); + + if (Vectorizable) + { + /* explicit vectorization */ + // process first unaligned result's coeffs + for (Index j=0; j<alignedStart; ++j) + res[j] += cj.pmul(lhs0(j), pfirst(ptmp0)); + // process aligned result's coeffs + if (lhs0.template aligned<LhsPacket>(alignedStart)) + for (Index i = alignedStart;i<alignedSize;i+=ResPacketSize) + pstore(&res[i], pcj.pmadd(lhs0.template load<LhsPacket, Aligned>(i), ptmp0, pload<ResPacket>(&res[i]))); + else + for (Index i = alignedStart;i<alignedSize;i+=ResPacketSize) + pstore(&res[i], pcj.pmadd(lhs0.template load<LhsPacket, Unaligned>(i), ptmp0, pload<ResPacket>(&res[i]))); + } + + // process remaining scalars (or all if no explicit vectorization) + for (Index i=alignedSize; i<size; ++i) + res[i] += cj.pmul(lhs0(i), pfirst(ptmp0)); + } + if (skipColumns) + { + start = 0; + end = skipColumns; + skipColumns = 0; + } + else + break; + } while(Vectorizable); + #undef _EIGEN_ACCUMULATE_PACKETS +} + +/* Optimized row-major matrix * vector product: + * This algorithm processes 4 rows at onces that allows to both reduce + * the number of load/stores of the result by a factor 4 and to reduce + * the instruction dependency. Moreover, we know that all bands have the + * same alignment pattern. + * + * Mixing type logic: + * - alpha is always a complex (or converted to a complex) + * - no vectorization + */ +template<typename Index, typename LhsScalar, typename LhsMapper, bool ConjugateLhs, typename RhsScalar, typename RhsMapper, bool ConjugateRhs, int Version> +struct general_matrix_vector_product<Index,LhsScalar,LhsMapper,RowMajor,ConjugateLhs,RhsScalar,RhsMapper,ConjugateRhs,Version> +{ +typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar; + +enum { + Vectorizable = packet_traits<LhsScalar>::Vectorizable && packet_traits<RhsScalar>::Vectorizable + && int(packet_traits<LhsScalar>::size)==int(packet_traits<RhsScalar>::size), + LhsPacketSize = Vectorizable ? packet_traits<LhsScalar>::size : 1, + RhsPacketSize = Vectorizable ? packet_traits<RhsScalar>::size : 1, + ResPacketSize = Vectorizable ? packet_traits<ResScalar>::size : 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; + +EIGEN_DONT_INLINE static void run( + Index rows, Index cols, + const LhsMapper& lhs, + const RhsMapper& rhs, + ResScalar* res, Index resIncr, + ResScalar alpha); +}; + +template<typename Index, typename LhsScalar, typename LhsMapper, bool ConjugateLhs, typename RhsScalar, typename RhsMapper, bool ConjugateRhs, int Version> +EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,LhsMapper,RowMajor,ConjugateLhs,RhsScalar,RhsMapper,ConjugateRhs,Version>::run( + Index rows, Index cols, + const LhsMapper& lhs, + const RhsMapper& rhs, + ResScalar* res, Index resIncr, + ResScalar alpha) +{ + eigen_internal_assert(rhs.stride()==1); + + #ifdef _EIGEN_ACCUMULATE_PACKETS + #error _EIGEN_ACCUMULATE_PACKETS has already been defined + #endif + + #define _EIGEN_ACCUMULATE_PACKETS(Alignment0,Alignment13,Alignment2) {\ + RhsPacket b = rhs.getVectorMapper(j, 0).template load<RhsPacket, Aligned>(0); \ + ptmp0 = pcj.pmadd(lhs0.template load<LhsPacket, Alignment0>(j), b, ptmp0); \ + ptmp1 = pcj.pmadd(lhs1.template load<LhsPacket, Alignment13>(j), b, ptmp1); \ + ptmp2 = pcj.pmadd(lhs2.template load<LhsPacket, Alignment2>(j), b, ptmp2); \ + ptmp3 = pcj.pmadd(lhs3.template load<LhsPacket, Alignment13>(j), b, ptmp3); } + + conj_helper<LhsScalar,RhsScalar,ConjugateLhs,ConjugateRhs> cj; + conj_helper<LhsPacket,RhsPacket,ConjugateLhs,ConjugateRhs> pcj; + + typedef typename LhsMapper::VectorMapper LhsScalars; + + enum { AllAligned=0, EvenAligned=1, FirstAligned=2, NoneAligned=3 }; + const Index rowsAtOnce = 4; + const Index peels = 2; + const Index RhsPacketAlignedMask = RhsPacketSize-1; + const Index LhsPacketAlignedMask = LhsPacketSize-1; + const Index depth = cols; + const Index lhsStride = lhs.stride(); + + // How many coeffs of the result do we have to skip to be aligned. + // Here we assume data are at least aligned on the base scalar type + // if that's not the case then vectorization is discarded, see below. + Index alignedStart = rhs.firstAligned(depth); + Index alignedSize = RhsPacketSize>1 ? alignedStart + ((depth-alignedStart) & ~RhsPacketAlignedMask) : 0; + const Index peeledSize = alignedSize - RhsPacketSize*peels - RhsPacketSize + 1; + + const Index alignmentStep = LhsPacketSize>1 ? (LhsPacketSize - lhsStride % LhsPacketSize) & LhsPacketAlignedMask : 0; + Index alignmentPattern = alignmentStep==0 ? AllAligned + : alignmentStep==(LhsPacketSize/2) ? EvenAligned + : FirstAligned; + + // we cannot assume the first element is aligned because of sub-matrices + const Index lhsAlignmentOffset = lhs.firstAligned(depth); + const Index rhsAlignmentOffset = rhs.firstAligned(rows); + + // find how many rows do we have to skip to be aligned with rhs (if possible) + Index skipRows = 0; + // if the data cannot be aligned (TODO add some compile time tests when possible, e.g. for floats) + if( (sizeof(LhsScalar)!=sizeof(RhsScalar)) || + (lhsAlignmentOffset < 0) || (lhsAlignmentOffset == depth) || + (rhsAlignmentOffset < 0) || (rhsAlignmentOffset == rows) ) + { + alignedSize = 0; + alignedStart = 0; + alignmentPattern = NoneAligned; + } + else if(LhsPacketSize > 4) + { + // TODO: extend the code to support aligned loads whenever possible when LhsPacketSize > 4. + alignmentPattern = NoneAligned; + } + else if (LhsPacketSize>1) + { + // eigen_internal_assert(size_t(firstLhs+lhsAlignmentOffset)%sizeof(LhsPacket)==0 || depth<LhsPacketSize); + + while (skipRows<LhsPacketSize && + alignedStart != ((lhsAlignmentOffset + alignmentStep*skipRows)%LhsPacketSize)) + ++skipRows; + if (skipRows==LhsPacketSize) + { + // nothing can be aligned, no need to skip any column + alignmentPattern = NoneAligned; + skipRows = 0; + } + else + { + skipRows = (std::min)(skipRows,Index(rows)); + // note that the skiped columns are processed later. + } + /* eigen_internal_assert( alignmentPattern==NoneAligned + || LhsPacketSize==1 + || (skipRows + rowsAtOnce >= rows) + || LhsPacketSize > depth + || (size_t(firstLhs+alignedStart+lhsStride*skipRows)%sizeof(LhsPacket))==0);*/ + } + else if(Vectorizable) + { + alignedStart = 0; + alignedSize = depth; + alignmentPattern = AllAligned; + } + + const Index offset1 = (FirstAligned && alignmentStep==1)?3:1; + const Index offset3 = (FirstAligned && alignmentStep==1)?1:3; + + Index rowBound = ((rows-skipRows)/rowsAtOnce)*rowsAtOnce + skipRows; + for (Index i=skipRows; i<rowBound; i+=rowsAtOnce) + { + // FIXME: what is the purpose of this EIGEN_ALIGN_DEFAULT ?? + EIGEN_ALIGN_MAX ResScalar tmp0 = ResScalar(0); + ResScalar tmp1 = ResScalar(0), tmp2 = ResScalar(0), tmp3 = ResScalar(0); + + // this helps the compiler generating good binary code + const LhsScalars lhs0 = lhs.getVectorMapper(i+0, 0), lhs1 = lhs.getVectorMapper(i+offset1, 0), + lhs2 = lhs.getVectorMapper(i+2, 0), lhs3 = lhs.getVectorMapper(i+offset3, 0); + + if (Vectorizable) + { + /* explicit vectorization */ + ResPacket ptmp0 = pset1<ResPacket>(ResScalar(0)), ptmp1 = pset1<ResPacket>(ResScalar(0)), + ptmp2 = pset1<ResPacket>(ResScalar(0)), ptmp3 = pset1<ResPacket>(ResScalar(0)); + + // process initial unaligned coeffs + // FIXME this loop get vectorized by the compiler ! + for (Index j=0; j<alignedStart; ++j) + { + RhsScalar b = rhs(j, 0); + tmp0 += cj.pmul(lhs0(j),b); tmp1 += cj.pmul(lhs1(j),b); + tmp2 += cj.pmul(lhs2(j),b); tmp3 += cj.pmul(lhs3(j),b); + } + + if (alignedSize>alignedStart) + { + switch(alignmentPattern) + { + case AllAligned: + for (Index j = alignedStart; j<alignedSize; j+=RhsPacketSize) + _EIGEN_ACCUMULATE_PACKETS(Aligned,Aligned,Aligned); + break; + case EvenAligned: + for (Index j = alignedStart; j<alignedSize; j+=RhsPacketSize) + _EIGEN_ACCUMULATE_PACKETS(Aligned,Unaligned,Aligned); + break; + case FirstAligned: + { + Index j = alignedStart; + if (peels>1) + { + /* Here we proccess 4 rows with with two peeled iterations to hide + * the overhead of unaligned loads. Moreover unaligned loads are handled + * using special shift/move operations between the two aligned packets + * overlaping the desired unaligned packet. This is *much* more efficient + * than basic unaligned loads. + */ + LhsPacket A01, A02, A03, A11, A12, A13; + A01 = lhs1.template load<LhsPacket, Aligned>(alignedStart-1); + A02 = lhs2.template load<LhsPacket, Aligned>(alignedStart-2); + A03 = lhs3.template load<LhsPacket, Aligned>(alignedStart-3); + + for (; j<peeledSize; j+=peels*RhsPacketSize) + { + RhsPacket b = rhs.getVectorMapper(j, 0).template load<RhsPacket, Aligned>(0); + A11 = lhs1.template load<LhsPacket, Aligned>(j-1+LhsPacketSize); palign<1>(A01,A11); + A12 = lhs2.template load<LhsPacket, Aligned>(j-2+LhsPacketSize); palign<2>(A02,A12); + A13 = lhs3.template load<LhsPacket, Aligned>(j-3+LhsPacketSize); palign<3>(A03,A13); + + ptmp0 = pcj.pmadd(lhs0.template load<LhsPacket, Aligned>(j), b, ptmp0); + ptmp1 = pcj.pmadd(A01, b, ptmp1); + A01 = lhs1.template load<LhsPacket, Aligned>(j-1+2*LhsPacketSize); palign<1>(A11,A01); + ptmp2 = pcj.pmadd(A02, b, ptmp2); + A02 = lhs2.template load<LhsPacket, Aligned>(j-2+2*LhsPacketSize); palign<2>(A12,A02); + ptmp3 = pcj.pmadd(A03, b, ptmp3); + A03 = lhs3.template load<LhsPacket, Aligned>(j-3+2*LhsPacketSize); palign<3>(A13,A03); + + b = rhs.getVectorMapper(j+RhsPacketSize, 0).template load<RhsPacket, Aligned>(0); + ptmp0 = pcj.pmadd(lhs0.template load<LhsPacket, Aligned>(j+LhsPacketSize), b, ptmp0); + ptmp1 = pcj.pmadd(A11, b, ptmp1); + ptmp2 = pcj.pmadd(A12, b, ptmp2); + ptmp3 = pcj.pmadd(A13, b, ptmp3); + } + } + for (; j<alignedSize; j+=RhsPacketSize) + _EIGEN_ACCUMULATE_PACKETS(Aligned,Unaligned,Unaligned); + break; + } + default: + for (Index j = alignedStart; j<alignedSize; j+=RhsPacketSize) + _EIGEN_ACCUMULATE_PACKETS(Unaligned,Unaligned,Unaligned); + break; + } + tmp0 += predux(ptmp0); + tmp1 += predux(ptmp1); + tmp2 += predux(ptmp2); + tmp3 += predux(ptmp3); + } + } // end explicit vectorization + + // process remaining coeffs (or all if no explicit vectorization) + // FIXME this loop get vectorized by the compiler ! + for (Index j=alignedSize; j<depth; ++j) + { + RhsScalar b = rhs(j, 0); + tmp0 += cj.pmul(lhs0(j),b); tmp1 += cj.pmul(lhs1(j),b); + tmp2 += cj.pmul(lhs2(j),b); tmp3 += cj.pmul(lhs3(j),b); + } + res[i*resIncr] += alpha*tmp0; + res[(i+offset1)*resIncr] += alpha*tmp1; + res[(i+2)*resIncr] += alpha*tmp2; + res[(i+offset3)*resIncr] += alpha*tmp3; + } + + // process remaining first and last rows (at most columnsAtOnce-1) + Index end = rows; + Index start = rowBound; + do + { + for (Index i=start; i<end; ++i) + { + EIGEN_ALIGN_MAX ResScalar tmp0 = ResScalar(0); + ResPacket ptmp0 = pset1<ResPacket>(tmp0); + const LhsScalars lhs0 = lhs.getVectorMapper(i, 0); + // process first unaligned result's coeffs + // FIXME this loop get vectorized by the compiler ! + for (Index j=0; j<alignedStart; ++j) + tmp0 += cj.pmul(lhs0(j), rhs(j, 0)); + + if (alignedSize>alignedStart) + { + // process aligned rhs coeffs + if (lhs0.template aligned<LhsPacket>(alignedStart)) + for (Index j = alignedStart;j<alignedSize;j+=RhsPacketSize) + ptmp0 = pcj.pmadd(lhs0.template load<LhsPacket, Aligned>(j), rhs.getVectorMapper(j, 0).template load<RhsPacket, Aligned>(0), ptmp0); + else + for (Index j = alignedStart;j<alignedSize;j+=RhsPacketSize) + ptmp0 = pcj.pmadd(lhs0.template load<LhsPacket, Unaligned>(j), rhs.getVectorMapper(j, 0).template load<RhsPacket, Aligned>(0), ptmp0); + tmp0 += predux(ptmp0); + } + + // process remaining scalars + // FIXME this loop get vectorized by the compiler ! + for (Index j=alignedSize; j<depth; ++j) + tmp0 += cj.pmul(lhs0(j), rhs(j, 0)); + res[i*resIncr] += alpha*tmp0; + } + if (skipRows) + { + start = 0; + end = skipRows; + skipRows = 0; + } + else + break; + } while(Vectorizable); + + #undef _EIGEN_ACCUMULATE_PACKETS +} + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_GENERAL_MATRIX_VECTOR_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixVector_BLAS.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixVector_BLAS.h new file mode 100644 index 000000000..e3a5d5892 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixVector_BLAS.h @@ -0,0 +1,129 @@ +/* + Copyright (c) 2011, Intel Corporation. All rights reserved. + + Redistribution and use in source and binary forms, with or without modification, + are permitted provided that the following conditions are met: + + * Redistributions of source code must retain the above copyright notice, this + list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above copyright notice, + this list of conditions and the following disclaimer in the documentation + and/or other materials provided with the distribution. + * Neither the name of Intel Corporation nor the names of its contributors may + be used to endorse or promote products derived from this software without + specific prior written permission. + + THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND + ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED + WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE + DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR + ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES + (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; + LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON + ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS + SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + + ******************************************************************************** + * Content : Eigen bindings to BLAS F77 + * General matrix-vector product functionality based on ?GEMV. + ******************************************************************************** +*/ + +#ifndef EIGEN_GENERAL_MATRIX_VECTOR_BLAS_H +#define EIGEN_GENERAL_MATRIX_VECTOR_BLAS_H + +namespace Eigen { + +namespace internal { + +/********************************************************************** +* This file implements general matrix-vector multiplication using BLAS +* gemv function via partial specialization of +* general_matrix_vector_product::run(..) method for float, double, +* std::complex<float> and std::complex<double> types +**********************************************************************/ + +// gemv specialization + +template<typename Index, typename LhsScalar, int StorageOrder, bool ConjugateLhs, typename RhsScalar, bool ConjugateRhs> +struct general_matrix_vector_product_gemv; + +#define EIGEN_BLAS_GEMV_SPECIALIZE(Scalar) \ +template<typename Index, bool ConjugateLhs, bool ConjugateRhs> \ +struct general_matrix_vector_product<Index,Scalar,const_blas_data_mapper<Scalar,Index,ColMajor>,ColMajor,ConjugateLhs,Scalar,const_blas_data_mapper<Scalar,Index,RowMajor>,ConjugateRhs,Specialized> { \ +static void run( \ + Index rows, Index cols, \ + const const_blas_data_mapper<Scalar,Index,ColMajor> &lhs, \ + const const_blas_data_mapper<Scalar,Index,RowMajor> &rhs, \ + Scalar* res, Index resIncr, Scalar alpha) \ +{ \ + if (ConjugateLhs) { \ + general_matrix_vector_product<Index,Scalar,const_blas_data_mapper<Scalar,Index,ColMajor>,ColMajor,ConjugateLhs,Scalar,const_blas_data_mapper<Scalar,Index,RowMajor>,ConjugateRhs,BuiltIn>::run( \ + rows, cols, lhs, rhs, res, resIncr, alpha); \ + } else { \ + general_matrix_vector_product_gemv<Index,Scalar,ColMajor,ConjugateLhs,Scalar,ConjugateRhs>::run( \ + rows, cols, lhs.data(), lhs.stride(), rhs.data(), rhs.stride(), res, resIncr, alpha); \ + } \ +} \ +}; \ +template<typename Index, bool ConjugateLhs, bool ConjugateRhs> \ +struct general_matrix_vector_product<Index,Scalar,const_blas_data_mapper<Scalar,Index,RowMajor>,RowMajor,ConjugateLhs,Scalar,const_blas_data_mapper<Scalar,Index,ColMajor>,ConjugateRhs,Specialized> { \ +static void run( \ + Index rows, Index cols, \ + const const_blas_data_mapper<Scalar,Index,RowMajor> &lhs, \ + const const_blas_data_mapper<Scalar,Index,ColMajor> &rhs, \ + Scalar* res, Index resIncr, Scalar alpha) \ +{ \ + general_matrix_vector_product_gemv<Index,Scalar,RowMajor,ConjugateLhs,Scalar,ConjugateRhs>::run( \ + rows, cols, lhs.data(), lhs.stride(), rhs.data(), rhs.stride(), res, resIncr, alpha); \ +} \ +}; \ + +EIGEN_BLAS_GEMV_SPECIALIZE(double) +EIGEN_BLAS_GEMV_SPECIALIZE(float) +EIGEN_BLAS_GEMV_SPECIALIZE(dcomplex) +EIGEN_BLAS_GEMV_SPECIALIZE(scomplex) + +#define EIGEN_BLAS_GEMV_SPECIALIZATION(EIGTYPE,BLASTYPE,BLASPREFIX) \ +template<typename Index, int LhsStorageOrder, bool ConjugateLhs, bool ConjugateRhs> \ +struct general_matrix_vector_product_gemv<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,ConjugateRhs> \ +{ \ +typedef Matrix<EIGTYPE,Dynamic,1,ColMajor> GEMVVector;\ +\ +static void run( \ + Index rows, Index cols, \ + const EIGTYPE* lhs, Index lhsStride, \ + const EIGTYPE* rhs, Index rhsIncr, \ + EIGTYPE* res, Index resIncr, EIGTYPE alpha) \ +{ \ + BlasIndex m=convert_index<BlasIndex>(rows), n=convert_index<BlasIndex>(cols), \ + lda=convert_index<BlasIndex>(lhsStride), incx=convert_index<BlasIndex>(rhsIncr), incy=convert_index<BlasIndex>(resIncr); \ + const EIGTYPE beta(1); \ + const EIGTYPE *x_ptr; \ + char trans=(LhsStorageOrder==ColMajor) ? 'N' : (ConjugateLhs) ? 'C' : 'T'; \ + if (LhsStorageOrder==RowMajor) { \ + m = convert_index<BlasIndex>(cols); \ + n = convert_index<BlasIndex>(rows); \ + }\ + GEMVVector x_tmp; \ + if (ConjugateRhs) { \ + Map<const GEMVVector, 0, InnerStride<> > map_x(rhs,cols,1,InnerStride<>(incx)); \ + x_tmp=map_x.conjugate(); \ + x_ptr=x_tmp.data(); \ + incx=1; \ + } else x_ptr=rhs; \ + BLASPREFIX##gemv_(&trans, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)lhs, &lda, (const BLASTYPE*)x_ptr, &incx, &numext::real_ref(beta), (BLASTYPE*)res, &incy); \ +}\ +}; + +EIGEN_BLAS_GEMV_SPECIALIZATION(double, double, d) +EIGEN_BLAS_GEMV_SPECIALIZATION(float, float, s) +EIGEN_BLAS_GEMV_SPECIALIZATION(dcomplex, double, z) +EIGEN_BLAS_GEMV_SPECIALIZATION(scomplex, float, c) + +} // end namespase internal + +} // end namespace Eigen + +#endif // EIGEN_GENERAL_MATRIX_VECTOR_BLAS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/Parallelizer.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/Parallelizer.h new file mode 100644 index 000000000..c2f084c82 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/Parallelizer.h @@ -0,0 +1,163 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2010 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_PARALLELIZER_H +#define EIGEN_PARALLELIZER_H + +namespace Eigen { + +namespace internal { + +/** \internal */ +inline void manage_multi_threading(Action action, int* v) +{ + static EIGEN_UNUSED int m_maxThreads = -1; + + if(action==SetAction) + { + eigen_internal_assert(v!=0); + m_maxThreads = *v; + } + else if(action==GetAction) + { + eigen_internal_assert(v!=0); + #ifdef EIGEN_HAS_OPENMP + if(m_maxThreads>0) + *v = m_maxThreads; + else + *v = omp_get_max_threads(); + #else + *v = 1; + #endif + } + else + { + eigen_internal_assert(false); + } +} + +} + +/** Must be call first when calling Eigen from multiple threads */ +inline void initParallel() +{ + int nbt; + internal::manage_multi_threading(GetAction, &nbt); + std::ptrdiff_t l1, l2, l3; + internal::manage_caching_sizes(GetAction, &l1, &l2, &l3); +} + +/** \returns the max number of threads reserved for Eigen + * \sa setNbThreads */ +inline int nbThreads() +{ + int ret; + internal::manage_multi_threading(GetAction, &ret); + return ret; +} + +/** Sets the max number of threads reserved for Eigen + * \sa nbThreads */ +inline void setNbThreads(int v) +{ + internal::manage_multi_threading(SetAction, &v); +} + +namespace internal { + +template<typename Index> struct GemmParallelInfo +{ + GemmParallelInfo() : sync(-1), users(0), lhs_start(0), lhs_length(0) {} + + Index volatile sync; + int volatile users; + + Index lhs_start; + Index lhs_length; +}; + +template<bool Condition, typename Functor, typename Index> +void parallelize_gemm(const Functor& func, Index rows, Index cols, Index depth, bool transpose) +{ + // TODO when EIGEN_USE_BLAS is defined, + // we should still enable OMP for other scalar types +#if !(defined (EIGEN_HAS_OPENMP)) || defined (EIGEN_USE_BLAS) + // FIXME the transpose variable is only needed to properly split + // the matrix product when multithreading is enabled. This is a temporary + // fix to support row-major destination matrices. This whole + // parallelizer mechanism has to be redisigned anyway. + EIGEN_UNUSED_VARIABLE(depth); + EIGEN_UNUSED_VARIABLE(transpose); + func(0,rows, 0,cols); +#else + + // Dynamically check whether we should enable or disable OpenMP. + // The conditions are: + // - the max number of threads we can create is greater than 1 + // - we are not already in a parallel code + // - the sizes are large enough + + // compute the maximal number of threads from the size of the product: + // This first heuristic takes into account that the product kernel is fully optimized when working with nr columns at once. + Index size = transpose ? rows : cols; + Index pb_max_threads = std::max<Index>(1,size / Functor::Traits::nr); + + // compute the maximal number of threads from the total amount of work: + double work = static_cast<double>(rows) * static_cast<double>(cols) * + static_cast<double>(depth); + double kMinTaskSize = 50000; // FIXME improve this heuristic. + pb_max_threads = std::max<Index>(1, std::min<Index>(pb_max_threads, work / kMinTaskSize)); + + // compute the number of threads we are going to use + Index threads = std::min<Index>(nbThreads(), pb_max_threads); + + // if multi-threading is explicitely disabled, not useful, or if we already are in a parallel session, + // then abort multi-threading + // FIXME omp_get_num_threads()>1 only works for openmp, what if the user does not use openmp? + if((!Condition) || (threads==1) || (omp_get_num_threads()>1)) + return func(0,rows, 0,cols); + + Eigen::initParallel(); + func.initParallelSession(threads); + + if(transpose) + std::swap(rows,cols); + + ei_declare_aligned_stack_constructed_variable(GemmParallelInfo<Index>,info,threads,0); + + #pragma omp parallel num_threads(threads) + { + Index i = omp_get_thread_num(); + // Note that the actual number of threads might be lower than the number of request ones. + Index actual_threads = omp_get_num_threads(); + + Index blockCols = (cols / actual_threads) & ~Index(0x3); + Index blockRows = (rows / actual_threads); + blockRows = (blockRows/Functor::Traits::mr)*Functor::Traits::mr; + + Index r0 = i*blockRows; + Index actualBlockRows = (i+1==actual_threads) ? rows-r0 : blockRows; + + Index c0 = i*blockCols; + Index actualBlockCols = (i+1==actual_threads) ? cols-c0 : blockCols; + + info[i].lhs_start = r0; + info[i].lhs_length = actualBlockRows; + + if(transpose) func(c0, actualBlockCols, 0, rows, info); + else func(0, rows, c0, actualBlockCols, info); + } +#endif +} + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_PARALLELIZER_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointMatrixMatrix.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointMatrixMatrix.h new file mode 100644 index 000000000..da6f82abc --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointMatrixMatrix.h @@ -0,0 +1,521 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2009 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_SELFADJOINT_MATRIX_MATRIX_H +#define EIGEN_SELFADJOINT_MATRIX_MATRIX_H + +namespace Eigen { + +namespace internal { + +// pack a selfadjoint block diagonal for use with the gebp_kernel +template<typename Scalar, typename Index, int Pack1, int Pack2_dummy, int StorageOrder> +struct symm_pack_lhs +{ + template<int BlockRows> inline + void pack(Scalar* blockA, const const_blas_data_mapper<Scalar,Index,StorageOrder>& lhs, Index cols, Index i, Index& count) + { + // normal copy + for(Index k=0; k<i; k++) + for(Index w=0; w<BlockRows; w++) + blockA[count++] = lhs(i+w,k); // normal + // symmetric copy + Index h = 0; + for(Index k=i; k<i+BlockRows; k++) + { + for(Index w=0; w<h; w++) + blockA[count++] = numext::conj(lhs(k, i+w)); // transposed + + blockA[count++] = numext::real(lhs(k,k)); // real (diagonal) + + for(Index w=h+1; w<BlockRows; w++) + blockA[count++] = lhs(i+w, k); // normal + ++h; + } + // transposed copy + for(Index k=i+BlockRows; k<cols; k++) + for(Index w=0; w<BlockRows; w++) + blockA[count++] = numext::conj(lhs(k, i+w)); // transposed + } + void operator()(Scalar* blockA, const Scalar* _lhs, Index lhsStride, Index cols, Index rows) + { + enum { PacketSize = packet_traits<Scalar>::size }; + 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; + + if(Pack1>=3*PacketSize) + for(Index i=0; i<peeled_mc3; i+=3*PacketSize) + pack<3*PacketSize>(blockA, lhs, cols, i, count); + + if(Pack1>=2*PacketSize) + for(Index i=peeled_mc3; i<peeled_mc2; i+=2*PacketSize) + pack<2*PacketSize>(blockA, lhs, cols, i, count); + + if(Pack1>=1*PacketSize) + for(Index i=peeled_mc2; i<peeled_mc1; i+=1*PacketSize) + pack<1*PacketSize>(blockA, lhs, cols, i, count); + + // do the same with mr==1 + for(Index i=peeled_mc1; i<rows; i++) + { + for(Index k=0; k<i; k++) + blockA[count++] = lhs(i, k); // normal + + blockA[count++] = numext::real(lhs(i, i)); // real (diagonal) + + for(Index k=i+1; k<cols; k++) + blockA[count++] = numext::conj(lhs(k, i)); // transposed + } + } +}; + +template<typename Scalar, typename Index, int nr, int StorageOrder> +struct symm_pack_rhs +{ + enum { PacketSize = packet_traits<Scalar>::size }; + void operator()(Scalar* blockB, const Scalar* _rhs, Index rhsStride, Index rows, Index cols, Index k2) + { + Index end_k = k2 + rows; + Index count = 0; + const_blas_data_mapper<Scalar,Index,StorageOrder> rhs(_rhs,rhsStride); + Index packet_cols8 = nr>=8 ? (cols/8) * 8 : 0; + Index packet_cols4 = nr>=4 ? (cols/4) * 4 : 0; + + // first part: normal case + for(Index j2=0; j2<k2; j2+=nr) + { + for(Index k=k2; k<end_k; k++) + { + blockB[count+0] = rhs(k,j2+0); + blockB[count+1] = rhs(k,j2+1); + if (nr>=4) + { + blockB[count+2] = rhs(k,j2+2); + blockB[count+3] = rhs(k,j2+3); + } + if (nr>=8) + { + blockB[count+4] = rhs(k,j2+4); + blockB[count+5] = rhs(k,j2+5); + blockB[count+6] = rhs(k,j2+6); + blockB[count+7] = rhs(k,j2+7); + } + count += nr; + } + } + + // second part: diagonal block + Index end8 = nr>=8 ? (std::min)(k2+rows,packet_cols8) : k2; + if(nr>=8) + { + for(Index j2=k2; j2<end8; j2+=8) + { + // again we can split vertically in three different parts (transpose, symmetric, normal) + // transpose + for(Index k=k2; k<j2; k++) + { + blockB[count+0] = numext::conj(rhs(j2+0,k)); + blockB[count+1] = numext::conj(rhs(j2+1,k)); + blockB[count+2] = numext::conj(rhs(j2+2,k)); + blockB[count+3] = numext::conj(rhs(j2+3,k)); + blockB[count+4] = numext::conj(rhs(j2+4,k)); + blockB[count+5] = numext::conj(rhs(j2+5,k)); + blockB[count+6] = numext::conj(rhs(j2+6,k)); + blockB[count+7] = numext::conj(rhs(j2+7,k)); + count += 8; + } + // symmetric + Index h = 0; + for(Index k=j2; k<j2+8; k++) + { + // normal + for (Index w=0 ; w<h; ++w) + blockB[count+w] = rhs(k,j2+w); + + blockB[count+h] = numext::real(rhs(k,k)); + + // transpose + for (Index w=h+1 ; w<8; ++w) + blockB[count+w] = numext::conj(rhs(j2+w,k)); + count += 8; + ++h; + } + // normal + for(Index k=j2+8; k<end_k; k++) + { + blockB[count+0] = rhs(k,j2+0); + blockB[count+1] = rhs(k,j2+1); + blockB[count+2] = rhs(k,j2+2); + blockB[count+3] = rhs(k,j2+3); + blockB[count+4] = rhs(k,j2+4); + blockB[count+5] = rhs(k,j2+5); + blockB[count+6] = rhs(k,j2+6); + blockB[count+7] = rhs(k,j2+7); + count += 8; + } + } + } + if(nr>=4) + { + for(Index j2=end8; j2<(std::min)(k2+rows,packet_cols4); j2+=4) + { + // again we can split vertically in three different parts (transpose, symmetric, normal) + // transpose + for(Index k=k2; k<j2; k++) + { + blockB[count+0] = numext::conj(rhs(j2+0,k)); + blockB[count+1] = numext::conj(rhs(j2+1,k)); + blockB[count+2] = numext::conj(rhs(j2+2,k)); + blockB[count+3] = numext::conj(rhs(j2+3,k)); + count += 4; + } + // symmetric + Index h = 0; + for(Index k=j2; k<j2+4; k++) + { + // normal + for (Index w=0 ; w<h; ++w) + blockB[count+w] = rhs(k,j2+w); + + blockB[count+h] = numext::real(rhs(k,k)); + + // transpose + for (Index w=h+1 ; w<4; ++w) + blockB[count+w] = numext::conj(rhs(j2+w,k)); + count += 4; + ++h; + } + // normal + for(Index k=j2+4; k<end_k; k++) + { + blockB[count+0] = rhs(k,j2+0); + blockB[count+1] = rhs(k,j2+1); + blockB[count+2] = rhs(k,j2+2); + blockB[count+3] = rhs(k,j2+3); + count += 4; + } + } + } + + // third part: transposed + if(nr>=8) + { + for(Index j2=k2+rows; j2<packet_cols8; j2+=8) + { + for(Index k=k2; k<end_k; k++) + { + blockB[count+0] = numext::conj(rhs(j2+0,k)); + blockB[count+1] = numext::conj(rhs(j2+1,k)); + blockB[count+2] = numext::conj(rhs(j2+2,k)); + blockB[count+3] = numext::conj(rhs(j2+3,k)); + blockB[count+4] = numext::conj(rhs(j2+4,k)); + blockB[count+5] = numext::conj(rhs(j2+5,k)); + blockB[count+6] = numext::conj(rhs(j2+6,k)); + blockB[count+7] = numext::conj(rhs(j2+7,k)); + count += 8; + } + } + } + if(nr>=4) + { + for(Index j2=(std::max)(packet_cols8,k2+rows); j2<packet_cols4; j2+=4) + { + for(Index k=k2; k<end_k; k++) + { + blockB[count+0] = numext::conj(rhs(j2+0,k)); + blockB[count+1] = numext::conj(rhs(j2+1,k)); + blockB[count+2] = numext::conj(rhs(j2+2,k)); + blockB[count+3] = numext::conj(rhs(j2+3,k)); + count += 4; + } + } + } + + // copy the remaining columns one at a time (=> the same with nr==1) + for(Index j2=packet_cols4; j2<cols; ++j2) + { + // transpose + Index half = (std::min)(end_k,j2); + for(Index k=k2; k<half; k++) + { + blockB[count] = numext::conj(rhs(j2,k)); + count += 1; + } + + if(half==j2 && half<k2+rows) + { + blockB[count] = numext::real(rhs(j2,j2)); + count += 1; + } + else + half--; + + // normal + for(Index k=half+1; k<k2+rows; k++) + { + blockB[count] = rhs(k,j2); + count += 1; + } + } + } +}; + +/* Optimized selfadjoint matrix * matrix (_SYMM) product built on top of + * the general matrix matrix product. + */ +template <typename Scalar, typename Index, + int LhsStorageOrder, bool LhsSelfAdjoint, bool ConjugateLhs, + int RhsStorageOrder, bool RhsSelfAdjoint, bool ConjugateRhs, + int ResStorageOrder> +struct product_selfadjoint_matrix; + +template <typename Scalar, typename Index, + int LhsStorageOrder, bool LhsSelfAdjoint, bool ConjugateLhs, + int RhsStorageOrder, bool RhsSelfAdjoint, bool ConjugateRhs> +struct product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,LhsSelfAdjoint,ConjugateLhs, RhsStorageOrder,RhsSelfAdjoint,ConjugateRhs,RowMajor> +{ + + static EIGEN_STRONG_INLINE void run( + Index rows, Index cols, + const Scalar* lhs, Index lhsStride, + const Scalar* rhs, Index rhsStride, + Scalar* res, Index resStride, + const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking) + { + product_selfadjoint_matrix<Scalar, Index, + EIGEN_LOGICAL_XOR(RhsSelfAdjoint,RhsStorageOrder==RowMajor) ? ColMajor : RowMajor, + RhsSelfAdjoint, NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(RhsSelfAdjoint,ConjugateRhs), + EIGEN_LOGICAL_XOR(LhsSelfAdjoint,LhsStorageOrder==RowMajor) ? ColMajor : RowMajor, + LhsSelfAdjoint, NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(LhsSelfAdjoint,ConjugateLhs), + ColMajor> + ::run(cols, rows, rhs, rhsStride, lhs, lhsStride, res, resStride, alpha, blocking); + } +}; + +template <typename Scalar, typename Index, + int LhsStorageOrder, bool ConjugateLhs, + int RhsStorageOrder, bool ConjugateRhs> +struct product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,true,ConjugateLhs, RhsStorageOrder,false,ConjugateRhs,ColMajor> +{ + + static EIGEN_DONT_INLINE void run( + Index rows, Index cols, + const Scalar* _lhs, Index lhsStride, + const Scalar* _rhs, Index rhsStride, + Scalar* res, Index resStride, + const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking); +}; + +template <typename Scalar, typename Index, + int LhsStorageOrder, bool ConjugateLhs, + int RhsStorageOrder, bool ConjugateRhs> +EIGEN_DONT_INLINE void product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,true,ConjugateLhs, RhsStorageOrder,false,ConjugateRhs,ColMajor>::run( + Index rows, Index cols, + const Scalar* _lhs, Index lhsStride, + const Scalar* _rhs, Index rhsStride, + Scalar* _res, Index resStride, + const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking) + { + Index size = rows; + + typedef gebp_traits<Scalar,Scalar> Traits; + + typedef const_blas_data_mapper<Scalar, Index, LhsStorageOrder> LhsMapper; + typedef const_blas_data_mapper<Scalar, Index, (LhsStorageOrder == RowMajor) ? ColMajor : RowMajor> LhsTransposeMapper; + typedef const_blas_data_mapper<Scalar, Index, RhsStorageOrder> RhsMapper; + typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper; + LhsMapper lhs(_lhs,lhsStride); + LhsTransposeMapper lhs_transpose(_lhs,lhsStride); + RhsMapper rhs(_rhs,rhsStride); + ResMapper res(_res, resStride); + + Index kc = blocking.kc(); // cache block size along the K direction + Index mc = (std::min)(rows,blocking.mc()); // cache block size along the M direction + // kc must be smaller than mc + kc = (std::min)(kc,mc); + std::size_t sizeA = kc*mc; + std::size_t sizeB = kc*cols; + ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA()); + ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB()); + + gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel; + symm_pack_lhs<Scalar, Index, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs; + gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder> pack_rhs; + gemm_pack_lhs<Scalar, Index, LhsTransposeMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder==RowMajor?ColMajor:RowMajor, true> pack_lhs_transposed; + + for(Index k2=0; k2<size; k2+=kc) + { + const Index actual_kc = (std::min)(k2+kc,size)-k2; + + // we have selected one row panel of rhs and one column panel of lhs + // pack rhs's panel into a sequential chunk of memory + // and expand each coeff to a constant packet for further reuse + pack_rhs(blockB, rhs.getSubMapper(k2,0), actual_kc, cols); + + // the select lhs's panel has to be split in three different parts: + // 1 - the transposed panel above the diagonal block => transposed packed copy + // 2 - the diagonal block => special packed copy + // 3 - the panel below the diagonal block => generic packed copy + for(Index i2=0; i2<k2; i2+=mc) + { + const Index actual_mc = (std::min)(i2+mc,k2)-i2; + // transposed packed copy + pack_lhs_transposed(blockA, lhs_transpose.getSubMapper(i2, k2), actual_kc, actual_mc); + + gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc, cols, alpha); + } + // the block diagonal + { + const Index actual_mc = (std::min)(k2+kc,size)-k2; + // symmetric packed copy + pack_lhs(blockA, &lhs(k2,k2), lhsStride, actual_kc, actual_mc); + + gebp_kernel(res.getSubMapper(k2, 0), blockA, blockB, actual_mc, actual_kc, cols, alpha); + } + + for(Index i2=k2+kc; i2<size; i2+=mc) + { + const Index actual_mc = (std::min)(i2+mc,size)-i2; + gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder,false>() + (blockA, lhs.getSubMapper(i2, k2), actual_kc, actual_mc); + + gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc, cols, alpha); + } + } + } + +// matrix * selfadjoint product +template <typename Scalar, typename Index, + int LhsStorageOrder, bool ConjugateLhs, + int RhsStorageOrder, bool ConjugateRhs> +struct product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,false,ConjugateLhs, RhsStorageOrder,true,ConjugateRhs,ColMajor> +{ + + static EIGEN_DONT_INLINE void run( + Index rows, Index cols, + const Scalar* _lhs, Index lhsStride, + const Scalar* _rhs, Index rhsStride, + Scalar* res, Index resStride, + const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking); +}; + +template <typename Scalar, typename Index, + int LhsStorageOrder, bool ConjugateLhs, + int RhsStorageOrder, bool ConjugateRhs> +EIGEN_DONT_INLINE void product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,false,ConjugateLhs, RhsStorageOrder,true,ConjugateRhs,ColMajor>::run( + Index rows, Index cols, + const Scalar* _lhs, Index lhsStride, + const Scalar* _rhs, Index rhsStride, + Scalar* _res, Index resStride, + const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking) + { + Index size = cols; + + typedef gebp_traits<Scalar,Scalar> Traits; + + typedef const_blas_data_mapper<Scalar, Index, LhsStorageOrder> LhsMapper; + typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper; + LhsMapper lhs(_lhs,lhsStride); + ResMapper res(_res,resStride); + + Index kc = blocking.kc(); // cache block size along the K direction + Index mc = (std::min)(rows,blocking.mc()); // cache block size along the M direction + std::size_t sizeA = kc*mc; + std::size_t sizeB = kc*cols; + ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA()); + ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB()); + + gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel; + gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs; + symm_pack_rhs<Scalar, Index, Traits::nr,RhsStorageOrder> pack_rhs; + + for(Index k2=0; k2<size; k2+=kc) + { + const Index actual_kc = (std::min)(k2+kc,size)-k2; + + pack_rhs(blockB, _rhs, rhsStride, actual_kc, cols, k2); + + // => GEPP + for(Index i2=0; i2<rows; i2+=mc) + { + const Index actual_mc = (std::min)(i2+mc,rows)-i2; + pack_lhs(blockA, lhs.getSubMapper(i2, k2), actual_kc, actual_mc); + + gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc, cols, alpha); + } + } + } + +} // end namespace internal + +/*************************************************************************** +* Wrapper to product_selfadjoint_matrix +***************************************************************************/ + +namespace internal { + +template<typename Lhs, int LhsMode, typename Rhs, int RhsMode> +struct selfadjoint_product_impl<Lhs,LhsMode,false,Rhs,RhsMode,false> +{ + typedef typename Product<Lhs,Rhs>::Scalar Scalar; + + typedef internal::blas_traits<Lhs> LhsBlasTraits; + typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType; + typedef internal::blas_traits<Rhs> RhsBlasTraits; + typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType; + + enum { + LhsIsUpper = (LhsMode&(Upper|Lower))==Upper, + LhsIsSelfAdjoint = (LhsMode&SelfAdjoint)==SelfAdjoint, + RhsIsUpper = (RhsMode&(Upper|Lower))==Upper, + RhsIsSelfAdjoint = (RhsMode&SelfAdjoint)==SelfAdjoint + }; + + template<typename Dest> + static void run(Dest &dst, const Lhs &a_lhs, const Rhs &a_rhs, const Scalar& alpha) + { + eigen_assert(dst.rows()==a_lhs.rows() && dst.cols()==a_rhs.cols()); + + typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(a_lhs); + typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(a_rhs); + + Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(a_lhs) + * RhsBlasTraits::extractScalarFactor(a_rhs); + + typedef internal::gemm_blocking_space<(Dest::Flags&RowMajorBit) ? RowMajor : ColMajor,Scalar,Scalar, + Lhs::MaxRowsAtCompileTime, Rhs::MaxColsAtCompileTime, Lhs::MaxColsAtCompileTime,1> BlockingType; + + BlockingType blocking(lhs.rows(), rhs.cols(), lhs.cols(), 1, false); + + internal::product_selfadjoint_matrix<Scalar, Index, + EIGEN_LOGICAL_XOR(LhsIsUpper,internal::traits<Lhs>::Flags &RowMajorBit) ? RowMajor : ColMajor, LhsIsSelfAdjoint, + NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(LhsIsUpper,bool(LhsBlasTraits::NeedToConjugate)), + EIGEN_LOGICAL_XOR(RhsIsUpper,internal::traits<Rhs>::Flags &RowMajorBit) ? RowMajor : ColMajor, RhsIsSelfAdjoint, + NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(RhsIsUpper,bool(RhsBlasTraits::NeedToConjugate)), + internal::traits<Dest>::Flags&RowMajorBit ? RowMajor : ColMajor> + ::run( + lhs.rows(), rhs.cols(), // sizes + &lhs.coeffRef(0,0), lhs.outerStride(), // lhs info + &rhs.coeffRef(0,0), rhs.outerStride(), // rhs info + &dst.coeffRef(0,0), dst.outerStride(), // result info + actualAlpha, blocking // alpha + ); + } +}; + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_SELFADJOINT_MATRIX_MATRIX_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointMatrixMatrix_BLAS.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointMatrixMatrix_BLAS.h new file mode 100644 index 000000000..a45238d69 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointMatrixMatrix_BLAS.h @@ -0,0 +1,275 @@ +/* + Copyright (c) 2011, Intel Corporation. All rights reserved. + + Redistribution and use in source and binary forms, with or without modification, + are permitted provided that the following conditions are met: + + * Redistributions of source code must retain the above copyright notice, this + list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above copyright notice, + this list of conditions and the following disclaimer in the documentation + and/or other materials provided with the distribution. + * Neither the name of Intel Corporation nor the names of its contributors may + be used to endorse or promote products derived from this software without + specific prior written permission. + + THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND + ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED + WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE + DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR + ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES + (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; + LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON + ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS + SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. +// + ******************************************************************************** + * Content : Eigen bindings to BLAS F77 + * Self adjoint matrix * matrix product functionality based on ?SYMM/?HEMM. + ******************************************************************************** +*/ + +#ifndef EIGEN_SELFADJOINT_MATRIX_MATRIX_BLAS_H +#define EIGEN_SELFADJOINT_MATRIX_MATRIX_BLAS_H + +namespace Eigen { + +namespace internal { + + +/* Optimized selfadjoint matrix * matrix (?SYMM/?HEMM) product */ + +#define EIGEN_BLAS_SYMM_L(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \ +template <typename Index, \ + int LhsStorageOrder, bool ConjugateLhs, \ + int RhsStorageOrder, bool ConjugateRhs> \ +struct product_selfadjoint_matrix<EIGTYPE,Index,LhsStorageOrder,true,ConjugateLhs,RhsStorageOrder,false,ConjugateRhs,ColMajor> \ +{\ +\ + static void run( \ + Index rows, Index cols, \ + const EIGTYPE* _lhs, Index lhsStride, \ + const EIGTYPE* _rhs, Index rhsStride, \ + EIGTYPE* res, Index resStride, \ + EIGTYPE alpha, level3_blocking<EIGTYPE, EIGTYPE>& /*blocking*/) \ + { \ + char side='L', uplo='L'; \ + BlasIndex m, n, lda, ldb, ldc; \ + const EIGTYPE *a, *b; \ + EIGTYPE beta(1); \ + MatrixX##EIGPREFIX b_tmp; \ +\ +/* Set transpose options */ \ +/* Set m, n, k */ \ + m = convert_index<BlasIndex>(rows); \ + n = convert_index<BlasIndex>(cols); \ +\ +/* Set lda, ldb, ldc */ \ + lda = convert_index<BlasIndex>(lhsStride); \ + ldb = convert_index<BlasIndex>(rhsStride); \ + ldc = convert_index<BlasIndex>(resStride); \ +\ +/* Set a, b, c */ \ + if (LhsStorageOrder==RowMajor) uplo='U'; \ + a = _lhs; \ +\ + if (RhsStorageOrder==RowMajor) { \ + Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > rhs(_rhs,n,m,OuterStride<>(rhsStride)); \ + b_tmp = rhs.adjoint(); \ + b = b_tmp.data(); \ + ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \ + } else b = _rhs; \ +\ + BLASPREFIX##symm_(&side, &uplo, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, &numext::real_ref(beta), (BLASTYPE*)res, &ldc); \ +\ + } \ +}; + + +#define EIGEN_BLAS_HEMM_L(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \ +template <typename Index, \ + int LhsStorageOrder, bool ConjugateLhs, \ + int RhsStorageOrder, bool ConjugateRhs> \ +struct product_selfadjoint_matrix<EIGTYPE,Index,LhsStorageOrder,true,ConjugateLhs,RhsStorageOrder,false,ConjugateRhs,ColMajor> \ +{\ + static void run( \ + Index rows, Index cols, \ + const EIGTYPE* _lhs, Index lhsStride, \ + const EIGTYPE* _rhs, Index rhsStride, \ + EIGTYPE* res, Index resStride, \ + EIGTYPE alpha, level3_blocking<EIGTYPE, EIGTYPE>& /*blocking*/) \ + { \ + char side='L', uplo='L'; \ + BlasIndex m, n, lda, ldb, ldc; \ + const EIGTYPE *a, *b; \ + EIGTYPE beta(1); \ + MatrixX##EIGPREFIX b_tmp; \ + Matrix<EIGTYPE, Dynamic, Dynamic, LhsStorageOrder> a_tmp; \ +\ +/* Set transpose options */ \ +/* Set m, n, k */ \ + m = convert_index<BlasIndex>(rows); \ + n = convert_index<BlasIndex>(cols); \ +\ +/* Set lda, ldb, ldc */ \ + lda = convert_index<BlasIndex>(lhsStride); \ + ldb = convert_index<BlasIndex>(rhsStride); \ + ldc = convert_index<BlasIndex>(resStride); \ +\ +/* Set a, b, c */ \ + if (((LhsStorageOrder==ColMajor) && ConjugateLhs) || ((LhsStorageOrder==RowMajor) && (!ConjugateLhs))) { \ + Map<const Matrix<EIGTYPE, Dynamic, Dynamic, LhsStorageOrder>, 0, OuterStride<> > lhs(_lhs,m,m,OuterStride<>(lhsStride)); \ + a_tmp = lhs.conjugate(); \ + a = a_tmp.data(); \ + lda = convert_index<BlasIndex>(a_tmp.outerStride()); \ + } else a = _lhs; \ + if (LhsStorageOrder==RowMajor) uplo='U'; \ +\ + if (RhsStorageOrder==ColMajor && (!ConjugateRhs)) { \ + b = _rhs; } \ + else { \ + if (RhsStorageOrder==ColMajor && ConjugateRhs) { \ + Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > rhs(_rhs,m,n,OuterStride<>(rhsStride)); \ + b_tmp = rhs.conjugate(); \ + } else \ + if (ConjugateRhs) { \ + Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > rhs(_rhs,n,m,OuterStride<>(rhsStride)); \ + b_tmp = rhs.adjoint(); \ + } else { \ + Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > rhs(_rhs,n,m,OuterStride<>(rhsStride)); \ + b_tmp = rhs.transpose(); \ + } \ + b = b_tmp.data(); \ + ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \ + } \ +\ + BLASPREFIX##hemm_(&side, &uplo, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, &numext::real_ref(beta), (BLASTYPE*)res, &ldc); \ +\ + } \ +}; + +EIGEN_BLAS_SYMM_L(double, double, d, d) +EIGEN_BLAS_SYMM_L(float, float, f, s) +EIGEN_BLAS_HEMM_L(dcomplex, double, cd, z) +EIGEN_BLAS_HEMM_L(scomplex, float, cf, c) + + +/* Optimized matrix * selfadjoint matrix (?SYMM/?HEMM) product */ + +#define EIGEN_BLAS_SYMM_R(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \ +template <typename Index, \ + int LhsStorageOrder, bool ConjugateLhs, \ + int RhsStorageOrder, bool ConjugateRhs> \ +struct product_selfadjoint_matrix<EIGTYPE,Index,LhsStorageOrder,false,ConjugateLhs,RhsStorageOrder,true,ConjugateRhs,ColMajor> \ +{\ +\ + static void run( \ + Index rows, Index cols, \ + const EIGTYPE* _lhs, Index lhsStride, \ + const EIGTYPE* _rhs, Index rhsStride, \ + EIGTYPE* res, Index resStride, \ + EIGTYPE alpha, level3_blocking<EIGTYPE, EIGTYPE>& /*blocking*/) \ + { \ + char side='R', uplo='L'; \ + BlasIndex m, n, lda, ldb, ldc; \ + const EIGTYPE *a, *b; \ + EIGTYPE beta(1); \ + MatrixX##EIGPREFIX b_tmp; \ +\ +/* Set m, n, k */ \ + m = convert_index<BlasIndex>(rows); \ + n = convert_index<BlasIndex>(cols); \ +\ +/* Set lda, ldb, ldc */ \ + lda = convert_index<BlasIndex>(rhsStride); \ + ldb = convert_index<BlasIndex>(lhsStride); \ + ldc = convert_index<BlasIndex>(resStride); \ +\ +/* Set a, b, c */ \ + if (RhsStorageOrder==RowMajor) uplo='U'; \ + a = _rhs; \ +\ + if (LhsStorageOrder==RowMajor) { \ + Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > lhs(_lhs,n,m,OuterStride<>(rhsStride)); \ + b_tmp = lhs.adjoint(); \ + b = b_tmp.data(); \ + ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \ + } else b = _lhs; \ +\ + BLASPREFIX##symm_(&side, &uplo, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, &numext::real_ref(beta), (BLASTYPE*)res, &ldc); \ +\ + } \ +}; + + +#define EIGEN_BLAS_HEMM_R(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \ +template <typename Index, \ + int LhsStorageOrder, bool ConjugateLhs, \ + int RhsStorageOrder, bool ConjugateRhs> \ +struct product_selfadjoint_matrix<EIGTYPE,Index,LhsStorageOrder,false,ConjugateLhs,RhsStorageOrder,true,ConjugateRhs,ColMajor> \ +{\ + static void run( \ + Index rows, Index cols, \ + const EIGTYPE* _lhs, Index lhsStride, \ + const EIGTYPE* _rhs, Index rhsStride, \ + EIGTYPE* res, Index resStride, \ + EIGTYPE alpha, level3_blocking<EIGTYPE, EIGTYPE>& /*blocking*/) \ + { \ + char side='R', uplo='L'; \ + BlasIndex m, n, lda, ldb, ldc; \ + const EIGTYPE *a, *b; \ + EIGTYPE beta(1); \ + MatrixX##EIGPREFIX b_tmp; \ + Matrix<EIGTYPE, Dynamic, Dynamic, RhsStorageOrder> a_tmp; \ +\ +/* Set m, n, k */ \ + m = convert_index<BlasIndex>(rows); \ + n = convert_index<BlasIndex>(cols); \ +\ +/* Set lda, ldb, ldc */ \ + lda = convert_index<BlasIndex>(rhsStride); \ + ldb = convert_index<BlasIndex>(lhsStride); \ + ldc = convert_index<BlasIndex>(resStride); \ +\ +/* Set a, b, c */ \ + if (((RhsStorageOrder==ColMajor) && ConjugateRhs) || ((RhsStorageOrder==RowMajor) && (!ConjugateRhs))) { \ + Map<const Matrix<EIGTYPE, Dynamic, Dynamic, RhsStorageOrder>, 0, OuterStride<> > rhs(_rhs,n,n,OuterStride<>(rhsStride)); \ + a_tmp = rhs.conjugate(); \ + a = a_tmp.data(); \ + lda = convert_index<BlasIndex>(a_tmp.outerStride()); \ + } else a = _rhs; \ + if (RhsStorageOrder==RowMajor) uplo='U'; \ +\ + if (LhsStorageOrder==ColMajor && (!ConjugateLhs)) { \ + b = _lhs; } \ + else { \ + if (LhsStorageOrder==ColMajor && ConjugateLhs) { \ + Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > lhs(_lhs,m,n,OuterStride<>(lhsStride)); \ + b_tmp = lhs.conjugate(); \ + } else \ + if (ConjugateLhs) { \ + Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > lhs(_lhs,n,m,OuterStride<>(lhsStride)); \ + b_tmp = lhs.adjoint(); \ + } else { \ + Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > lhs(_lhs,n,m,OuterStride<>(lhsStride)); \ + b_tmp = lhs.transpose(); \ + } \ + b = b_tmp.data(); \ + ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \ + } \ +\ + BLASPREFIX##hemm_(&side, &uplo, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, &numext::real_ref(beta), (BLASTYPE*)res, &ldc); \ + } \ +}; + +EIGEN_BLAS_SYMM_R(double, double, d, d) +EIGEN_BLAS_SYMM_R(float, float, f, s) +EIGEN_BLAS_HEMM_R(dcomplex, double, cd, z) +EIGEN_BLAS_HEMM_R(scomplex, float, cf, c) + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_SELFADJOINT_MATRIX_MATRIX_BLAS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointMatrixVector.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointMatrixVector.h new file mode 100644 index 000000000..3fd180e6c --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointMatrixVector.h @@ -0,0 +1,260 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2009 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_SELFADJOINT_MATRIX_VECTOR_H +#define EIGEN_SELFADJOINT_MATRIX_VECTOR_H + +namespace Eigen { + +namespace internal { + +/* Optimized selfadjoint matrix * vector product: + * This algorithm processes 2 columns at onces that allows to both reduce + * the number of load/stores of the result by a factor 2 and to reduce + * the instruction dependency. + */ + +template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs, int Version=Specialized> +struct selfadjoint_matrix_vector_product; + +template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs, int Version> +struct selfadjoint_matrix_vector_product + +{ +static EIGEN_DONT_INLINE void run( + Index size, + const Scalar* lhs, Index lhsStride, + const Scalar* rhs, + Scalar* res, + Scalar alpha); +}; + +template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs, int Version> +EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs,Version>::run( + Index size, + const Scalar* lhs, Index lhsStride, + const Scalar* rhs, + Scalar* res, + Scalar alpha) +{ + typedef typename packet_traits<Scalar>::type Packet; + typedef typename NumTraits<Scalar>::Real RealScalar; + const Index PacketSize = sizeof(Packet)/sizeof(Scalar); + + enum { + IsRowMajor = StorageOrder==RowMajor ? 1 : 0, + IsLower = UpLo == Lower ? 1 : 0, + FirstTriangular = IsRowMajor == IsLower + }; + + conj_helper<Scalar,Scalar,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs, IsRowMajor), ConjugateRhs> cj0; + conj_helper<Scalar,Scalar,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs, !IsRowMajor), ConjugateRhs> cj1; + conj_helper<RealScalar,Scalar,false, ConjugateRhs> cjd; + + conj_helper<Packet,Packet,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs, IsRowMajor), ConjugateRhs> pcj0; + conj_helper<Packet,Packet,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs, !IsRowMajor), ConjugateRhs> pcj1; + + Scalar cjAlpha = ConjugateRhs ? numext::conj(alpha) : alpha; + + + Index bound = (std::max)(Index(0),size-8) & 0xfffffffe; + if (FirstTriangular) + bound = size - bound; + + for (Index j=FirstTriangular ? bound : 0; + j<(FirstTriangular ? size : bound);j+=2) + { + const Scalar* EIGEN_RESTRICT A0 = lhs + j*lhsStride; + const Scalar* EIGEN_RESTRICT A1 = lhs + (j+1)*lhsStride; + + Scalar t0 = cjAlpha * rhs[j]; + Packet ptmp0 = pset1<Packet>(t0); + Scalar t1 = cjAlpha * rhs[j+1]; + Packet ptmp1 = pset1<Packet>(t1); + + Scalar t2(0); + Packet ptmp2 = pset1<Packet>(t2); + Scalar t3(0); + Packet ptmp3 = pset1<Packet>(t3); + + Index starti = FirstTriangular ? 0 : j+2; + Index endi = FirstTriangular ? j : size; + Index alignedStart = (starti) + internal::first_default_aligned(&res[starti], endi-starti); + Index alignedEnd = alignedStart + ((endi-alignedStart)/(PacketSize))*(PacketSize); + + res[j] += cjd.pmul(numext::real(A0[j]), t0); + res[j+1] += cjd.pmul(numext::real(A1[j+1]), t1); + if(FirstTriangular) + { + res[j] += cj0.pmul(A1[j], t1); + t3 += cj1.pmul(A1[j], rhs[j]); + } + else + { + res[j+1] += cj0.pmul(A0[j+1],t0); + t2 += cj1.pmul(A0[j+1], rhs[j+1]); + } + + for (Index i=starti; i<alignedStart; ++i) + { + res[i] += cj0.pmul(A0[i], t0) + cj0.pmul(A1[i],t1); + t2 += cj1.pmul(A0[i], rhs[i]); + t3 += cj1.pmul(A1[i], rhs[i]); + } + // Yes this an optimization for gcc 4.3 and 4.4 (=> huge speed up) + // gcc 4.2 does this optimization automatically. + const Scalar* EIGEN_RESTRICT a0It = A0 + alignedStart; + const Scalar* EIGEN_RESTRICT a1It = A1 + alignedStart; + const Scalar* EIGEN_RESTRICT rhsIt = rhs + alignedStart; + Scalar* EIGEN_RESTRICT resIt = res + alignedStart; + for (Index i=alignedStart; i<alignedEnd; i+=PacketSize) + { + Packet A0i = ploadu<Packet>(a0It); a0It += PacketSize; + Packet A1i = ploadu<Packet>(a1It); a1It += PacketSize; + Packet Bi = ploadu<Packet>(rhsIt); rhsIt += PacketSize; // FIXME should be aligned in most cases + Packet Xi = pload <Packet>(resIt); + + Xi = pcj0.pmadd(A0i,ptmp0, pcj0.pmadd(A1i,ptmp1,Xi)); + ptmp2 = pcj1.pmadd(A0i, Bi, ptmp2); + ptmp3 = pcj1.pmadd(A1i, Bi, ptmp3); + pstore(resIt,Xi); resIt += PacketSize; + } + for (Index i=alignedEnd; i<endi; i++) + { + res[i] += cj0.pmul(A0[i], t0) + cj0.pmul(A1[i],t1); + t2 += cj1.pmul(A0[i], rhs[i]); + t3 += cj1.pmul(A1[i], rhs[i]); + } + + res[j] += alpha * (t2 + predux(ptmp2)); + res[j+1] += alpha * (t3 + predux(ptmp3)); + } + for (Index j=FirstTriangular ? 0 : bound;j<(FirstTriangular ? bound : size);j++) + { + const Scalar* EIGEN_RESTRICT A0 = lhs + j*lhsStride; + + Scalar t1 = cjAlpha * rhs[j]; + Scalar t2(0); + res[j] += cjd.pmul(numext::real(A0[j]), t1); + for (Index i=FirstTriangular ? 0 : j+1; i<(FirstTriangular ? j : size); i++) + { + res[i] += cj0.pmul(A0[i], t1); + t2 += cj1.pmul(A0[i], rhs[i]); + } + res[j] += alpha * t2; + } +} + +} // end namespace internal + +/*************************************************************************** +* Wrapper to product_selfadjoint_vector +***************************************************************************/ + +namespace internal { + +template<typename Lhs, int LhsMode, typename Rhs> +struct selfadjoint_product_impl<Lhs,LhsMode,false,Rhs,0,true> +{ + typedef typename Product<Lhs,Rhs>::Scalar Scalar; + + typedef internal::blas_traits<Lhs> LhsBlasTraits; + typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType; + typedef typename internal::remove_all<ActualLhsType>::type ActualLhsTypeCleaned; + + typedef internal::blas_traits<Rhs> RhsBlasTraits; + typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType; + typedef typename internal::remove_all<ActualRhsType>::type ActualRhsTypeCleaned; + + enum { LhsUpLo = LhsMode&(Upper|Lower) }; + + template<typename Dest> + static void run(Dest& dest, const Lhs &a_lhs, const Rhs &a_rhs, const Scalar& alpha) + { + typedef typename Dest::Scalar ResScalar; + typedef typename Rhs::Scalar RhsScalar; + typedef Map<Matrix<ResScalar,Dynamic,1>, EIGEN_PLAIN_ENUM_MIN(AlignedMax,internal::packet_traits<ResScalar>::size)> MappedDest; + + eigen_assert(dest.rows()==a_lhs.rows() && dest.cols()==a_rhs.cols()); + + typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(a_lhs); + typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(a_rhs); + + Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(a_lhs) + * RhsBlasTraits::extractScalarFactor(a_rhs); + + enum { + EvalToDest = (Dest::InnerStrideAtCompileTime==1), + UseRhs = (ActualRhsTypeCleaned::InnerStrideAtCompileTime==1) + }; + + internal::gemv_static_vector_if<ResScalar,Dest::SizeAtCompileTime,Dest::MaxSizeAtCompileTime,!EvalToDest> static_dest; + internal::gemv_static_vector_if<RhsScalar,ActualRhsTypeCleaned::SizeAtCompileTime,ActualRhsTypeCleaned::MaxSizeAtCompileTime,!UseRhs> static_rhs; + + ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(), + EvalToDest ? dest.data() : static_dest.data()); + + ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhsPtr,rhs.size(), + UseRhs ? const_cast<RhsScalar*>(rhs.data()) : static_rhs.data()); + + if(!EvalToDest) + { + #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN + Index size = dest.size(); + EIGEN_DENSE_STORAGE_CTOR_PLUGIN + #endif + MappedDest(actualDestPtr, dest.size()) = dest; + } + + if(!UseRhs) + { + #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN + Index size = rhs.size(); + EIGEN_DENSE_STORAGE_CTOR_PLUGIN + #endif + Map<typename ActualRhsTypeCleaned::PlainObject>(actualRhsPtr, rhs.size()) = rhs; + } + + + internal::selfadjoint_matrix_vector_product<Scalar, Index, (internal::traits<ActualLhsTypeCleaned>::Flags&RowMajorBit) ? RowMajor : ColMajor, + int(LhsUpLo), bool(LhsBlasTraits::NeedToConjugate), bool(RhsBlasTraits::NeedToConjugate)>::run + ( + lhs.rows(), // size + &lhs.coeffRef(0,0), lhs.outerStride(), // lhs info + actualRhsPtr, // rhs info + actualDestPtr, // result info + actualAlpha // scale factor + ); + + if(!EvalToDest) + dest = MappedDest(actualDestPtr, dest.size()); + } +}; + +template<typename Lhs, typename Rhs, int RhsMode> +struct selfadjoint_product_impl<Lhs,0,true,Rhs,RhsMode,false> +{ + typedef typename Product<Lhs,Rhs>::Scalar Scalar; + enum { RhsUpLo = RhsMode&(Upper|Lower) }; + + template<typename Dest> + static void run(Dest& dest, const Lhs &a_lhs, const Rhs &a_rhs, const Scalar& alpha) + { + // let's simply transpose the product + Transpose<Dest> destT(dest); + selfadjoint_product_impl<Transpose<const Rhs>, int(RhsUpLo)==Upper ? Lower : Upper, false, + Transpose<const Lhs>, 0, true>::run(destT, a_rhs.transpose(), a_lhs.transpose(), alpha); + } +}; + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_SELFADJOINT_MATRIX_VECTOR_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointMatrixVector_BLAS.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointMatrixVector_BLAS.h new file mode 100644 index 000000000..38f23accf --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointMatrixVector_BLAS.h @@ -0,0 +1,111 @@ +/* + Copyright (c) 2011, Intel Corporation. All rights reserved. + + Redistribution and use in source and binary forms, with or without modification, + are permitted provided that the following conditions are met: + + * Redistributions of source code must retain the above copyright notice, this + list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above copyright notice, + this list of conditions and the following disclaimer in the documentation + and/or other materials provided with the distribution. + * Neither the name of Intel Corporation nor the names of its contributors may + be used to endorse or promote products derived from this software without + specific prior written permission. + + THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND + ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED + WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE + DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR + ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES + (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; + LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON + ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS + SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + + ******************************************************************************** + * Content : Eigen bindings to BLAS F77 + * Selfadjoint matrix-vector product functionality based on ?SYMV/HEMV. + ******************************************************************************** +*/ + +#ifndef EIGEN_SELFADJOINT_MATRIX_VECTOR_BLAS_H +#define EIGEN_SELFADJOINT_MATRIX_VECTOR_BLAS_H + +namespace Eigen { + +namespace internal { + +/********************************************************************** +* This file implements selfadjoint matrix-vector multiplication using BLAS +**********************************************************************/ + +// symv/hemv specialization + +template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs> +struct selfadjoint_matrix_vector_product_symv : + selfadjoint_matrix_vector_product<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs,BuiltIn> {}; + +#define EIGEN_BLAS_SYMV_SPECIALIZE(Scalar) \ +template<typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs> \ +struct selfadjoint_matrix_vector_product<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs,Specialized> { \ +static void run( \ + Index size, const Scalar* lhs, Index lhsStride, \ + const Scalar* _rhs, Scalar* res, Scalar alpha) { \ + enum {\ + IsColMajor = StorageOrder==ColMajor \ + }; \ + if (IsColMajor == ConjugateLhs) {\ + selfadjoint_matrix_vector_product<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs,BuiltIn>::run( \ + size, lhs, lhsStride, _rhs, res, alpha); \ + } else {\ + selfadjoint_matrix_vector_product_symv<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs>::run( \ + size, lhs, lhsStride, _rhs, res, alpha); \ + }\ + } \ +}; \ + +EIGEN_BLAS_SYMV_SPECIALIZE(double) +EIGEN_BLAS_SYMV_SPECIALIZE(float) +EIGEN_BLAS_SYMV_SPECIALIZE(dcomplex) +EIGEN_BLAS_SYMV_SPECIALIZE(scomplex) + +#define EIGEN_BLAS_SYMV_SPECIALIZATION(EIGTYPE,BLASTYPE,BLASFUNC) \ +template<typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs> \ +struct selfadjoint_matrix_vector_product_symv<EIGTYPE,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs> \ +{ \ +typedef Matrix<EIGTYPE,Dynamic,1,ColMajor> SYMVVector;\ +\ +static void run( \ +Index size, const EIGTYPE* lhs, Index lhsStride, \ +const EIGTYPE* _rhs, EIGTYPE* res, EIGTYPE alpha) \ +{ \ + enum {\ + IsRowMajor = StorageOrder==RowMajor ? 1 : 0, \ + IsLower = UpLo == Lower ? 1 : 0 \ + }; \ + BlasIndex n=convert_index<BlasIndex>(size), lda=convert_index<BlasIndex>(lhsStride), incx=1, incy=1; \ + EIGTYPE beta(1); \ + const EIGTYPE *x_ptr; \ + char uplo=(IsRowMajor) ? (IsLower ? 'U' : 'L') : (IsLower ? 'L' : 'U'); \ + SYMVVector x_tmp; \ + if (ConjugateRhs) { \ + Map<const SYMVVector, 0 > map_x(_rhs,size,1); \ + x_tmp=map_x.conjugate(); \ + x_ptr=x_tmp.data(); \ + } else x_ptr=_rhs; \ + BLASFUNC(&uplo, &n, &numext::real_ref(alpha), (const BLASTYPE*)lhs, &lda, (const BLASTYPE*)x_ptr, &incx, &numext::real_ref(beta), (BLASTYPE*)res, &incy); \ +}\ +}; + +EIGEN_BLAS_SYMV_SPECIALIZATION(double, double, dsymv_) +EIGEN_BLAS_SYMV_SPECIALIZATION(float, float, ssymv_) +EIGEN_BLAS_SYMV_SPECIALIZATION(dcomplex, double, zhemv_) +EIGEN_BLAS_SYMV_SPECIALIZATION(scomplex, float, chemv_) + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_SELFADJOINT_MATRIX_VECTOR_BLAS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointProduct.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointProduct.h new file mode 100644 index 000000000..f038d686f --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointProduct.h @@ -0,0 +1,133 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2009 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_SELFADJOINT_PRODUCT_H +#define EIGEN_SELFADJOINT_PRODUCT_H + +/********************************************************************** +* This file implements a self adjoint product: C += A A^T updating only +* half of the selfadjoint matrix C. +* It corresponds to the level 3 SYRK and level 2 SYR Blas routines. +**********************************************************************/ + +namespace Eigen { + + +template<typename Scalar, typename Index, int UpLo, bool ConjLhs, bool ConjRhs> +struct selfadjoint_rank1_update<Scalar,Index,ColMajor,UpLo,ConjLhs,ConjRhs> +{ + static void run(Index size, Scalar* mat, Index stride, const Scalar* vecX, const Scalar* vecY, const Scalar& alpha) + { + internal::conj_if<ConjRhs> cj; + typedef Map<const Matrix<Scalar,Dynamic,1> > OtherMap; + typedef typename internal::conditional<ConjLhs,typename OtherMap::ConjugateReturnType,const OtherMap&>::type ConjLhsType; + for (Index i=0; i<size; ++i) + { + Map<Matrix<Scalar,Dynamic,1> >(mat+stride*i+(UpLo==Lower ? i : 0), (UpLo==Lower ? size-i : (i+1))) + += (alpha * cj(vecY[i])) * ConjLhsType(OtherMap(vecX+(UpLo==Lower ? i : 0),UpLo==Lower ? size-i : (i+1))); + } + } +}; + +template<typename Scalar, typename Index, int UpLo, bool ConjLhs, bool ConjRhs> +struct selfadjoint_rank1_update<Scalar,Index,RowMajor,UpLo,ConjLhs,ConjRhs> +{ + static void run(Index size, Scalar* mat, Index stride, const Scalar* vecX, const Scalar* vecY, const Scalar& alpha) + { + selfadjoint_rank1_update<Scalar,Index,ColMajor,UpLo==Lower?Upper:Lower,ConjRhs,ConjLhs>::run(size,mat,stride,vecY,vecX,alpha); + } +}; + +template<typename MatrixType, typename OtherType, int UpLo, bool OtherIsVector = OtherType::IsVectorAtCompileTime> +struct selfadjoint_product_selector; + +template<typename MatrixType, typename OtherType, int UpLo> +struct selfadjoint_product_selector<MatrixType,OtherType,UpLo,true> +{ + static void run(MatrixType& mat, const OtherType& other, const typename MatrixType::Scalar& alpha) + { + typedef typename MatrixType::Scalar Scalar; + typedef internal::blas_traits<OtherType> OtherBlasTraits; + typedef typename OtherBlasTraits::DirectLinearAccessType ActualOtherType; + typedef typename internal::remove_all<ActualOtherType>::type _ActualOtherType; + typename internal::add_const_on_value_type<ActualOtherType>::type actualOther = OtherBlasTraits::extract(other.derived()); + + Scalar actualAlpha = alpha * OtherBlasTraits::extractScalarFactor(other.derived()); + + enum { + StorageOrder = (internal::traits<MatrixType>::Flags&RowMajorBit) ? RowMajor : ColMajor, + UseOtherDirectly = _ActualOtherType::InnerStrideAtCompileTime==1 + }; + internal::gemv_static_vector_if<Scalar,OtherType::SizeAtCompileTime,OtherType::MaxSizeAtCompileTime,!UseOtherDirectly> static_other; + + ei_declare_aligned_stack_constructed_variable(Scalar, actualOtherPtr, other.size(), + (UseOtherDirectly ? const_cast<Scalar*>(actualOther.data()) : static_other.data())); + + if(!UseOtherDirectly) + Map<typename _ActualOtherType::PlainObject>(actualOtherPtr, actualOther.size()) = actualOther; + + selfadjoint_rank1_update<Scalar,Index,StorageOrder,UpLo, + OtherBlasTraits::NeedToConjugate && NumTraits<Scalar>::IsComplex, + (!OtherBlasTraits::NeedToConjugate) && NumTraits<Scalar>::IsComplex> + ::run(other.size(), mat.data(), mat.outerStride(), actualOtherPtr, actualOtherPtr, actualAlpha); + } +}; + +template<typename MatrixType, typename OtherType, int UpLo> +struct selfadjoint_product_selector<MatrixType,OtherType,UpLo,false> +{ + static void run(MatrixType& mat, const OtherType& other, const typename MatrixType::Scalar& alpha) + { + typedef typename MatrixType::Scalar Scalar; + typedef internal::blas_traits<OtherType> OtherBlasTraits; + typedef typename OtherBlasTraits::DirectLinearAccessType ActualOtherType; + typedef typename internal::remove_all<ActualOtherType>::type _ActualOtherType; + typename internal::add_const_on_value_type<ActualOtherType>::type actualOther = OtherBlasTraits::extract(other.derived()); + + Scalar actualAlpha = alpha * OtherBlasTraits::extractScalarFactor(other.derived()); + + enum { + IsRowMajor = (internal::traits<MatrixType>::Flags&RowMajorBit) ? 1 : 0, + OtherIsRowMajor = _ActualOtherType::Flags&RowMajorBit ? 1 : 0 + }; + + Index size = mat.cols(); + Index depth = actualOther.cols(); + + typedef internal::gemm_blocking_space<IsRowMajor ? RowMajor : ColMajor,Scalar,Scalar, + MatrixType::MaxColsAtCompileTime, MatrixType::MaxColsAtCompileTime, _ActualOtherType::MaxColsAtCompileTime> BlockingType; + + BlockingType blocking(size, size, depth, 1, false); + + + internal::general_matrix_matrix_triangular_product<Index, + Scalar, OtherIsRowMajor ? RowMajor : ColMajor, OtherBlasTraits::NeedToConjugate && NumTraits<Scalar>::IsComplex, + Scalar, OtherIsRowMajor ? ColMajor : RowMajor, (!OtherBlasTraits::NeedToConjugate) && NumTraits<Scalar>::IsComplex, + IsRowMajor ? RowMajor : ColMajor, UpLo> + ::run(size, depth, + &actualOther.coeffRef(0,0), actualOther.outerStride(), &actualOther.coeffRef(0,0), actualOther.outerStride(), + mat.data(), mat.outerStride(), actualAlpha, blocking); + } +}; + +// high level API + +template<typename MatrixType, unsigned int UpLo> +template<typename DerivedU> +SelfAdjointView<MatrixType,UpLo>& SelfAdjointView<MatrixType,UpLo> +::rankUpdate(const MatrixBase<DerivedU>& u, const Scalar& alpha) +{ + selfadjoint_product_selector<MatrixType,DerivedU,UpLo>::run(_expression().const_cast_derived(), u.derived(), alpha); + + return *this; +} + +} // end namespace Eigen + +#endif // EIGEN_SELFADJOINT_PRODUCT_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointRank2Update.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointRank2Update.h new file mode 100644 index 000000000..2ae364111 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointRank2Update.h @@ -0,0 +1,93 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2009 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_SELFADJOINTRANK2UPTADE_H +#define EIGEN_SELFADJOINTRANK2UPTADE_H + +namespace Eigen { + +namespace internal { + +/* Optimized selfadjoint matrix += alpha * uv' + conj(alpha)*vu' + * It corresponds to the Level2 syr2 BLAS routine + */ + +template<typename Scalar, typename Index, typename UType, typename VType, int UpLo> +struct selfadjoint_rank2_update_selector; + +template<typename Scalar, typename Index, typename UType, typename VType> +struct selfadjoint_rank2_update_selector<Scalar,Index,UType,VType,Lower> +{ + static void run(Scalar* mat, Index stride, const UType& u, const VType& v, const Scalar& alpha) + { + const Index size = u.size(); + for (Index i=0; i<size; ++i) + { + Map<Matrix<Scalar,Dynamic,1> >(mat+stride*i+i, size-i) += + (numext::conj(alpha) * numext::conj(u.coeff(i))) * v.tail(size-i) + + (alpha * numext::conj(v.coeff(i))) * u.tail(size-i); + } + } +}; + +template<typename Scalar, typename Index, typename UType, typename VType> +struct selfadjoint_rank2_update_selector<Scalar,Index,UType,VType,Upper> +{ + static void run(Scalar* mat, Index stride, const UType& u, const VType& v, const Scalar& alpha) + { + const Index size = u.size(); + for (Index i=0; i<size; ++i) + Map<Matrix<Scalar,Dynamic,1> >(mat+stride*i, i+1) += + (numext::conj(alpha) * numext::conj(u.coeff(i))) * v.head(i+1) + + (alpha * numext::conj(v.coeff(i))) * u.head(i+1); + } +}; + +template<bool Cond, typename T> struct conj_expr_if + : conditional<!Cond, const T&, + CwiseUnaryOp<scalar_conjugate_op<typename traits<T>::Scalar>,T> > {}; + +} // end namespace internal + +template<typename MatrixType, unsigned int UpLo> +template<typename DerivedU, typename DerivedV> +SelfAdjointView<MatrixType,UpLo>& SelfAdjointView<MatrixType,UpLo> +::rankUpdate(const MatrixBase<DerivedU>& u, const MatrixBase<DerivedV>& v, const Scalar& alpha) +{ + typedef internal::blas_traits<DerivedU> UBlasTraits; + typedef typename UBlasTraits::DirectLinearAccessType ActualUType; + typedef typename internal::remove_all<ActualUType>::type _ActualUType; + typename internal::add_const_on_value_type<ActualUType>::type actualU = UBlasTraits::extract(u.derived()); + + typedef internal::blas_traits<DerivedV> VBlasTraits; + typedef typename VBlasTraits::DirectLinearAccessType ActualVType; + typedef typename internal::remove_all<ActualVType>::type _ActualVType; + typename internal::add_const_on_value_type<ActualVType>::type actualV = VBlasTraits::extract(v.derived()); + + // If MatrixType is row major, then we use the routine for lower triangular in the upper triangular case and + // vice versa, and take the complex conjugate of all coefficients and vector entries. + + enum { IsRowMajor = (internal::traits<MatrixType>::Flags&RowMajorBit) ? 1 : 0 }; + Scalar actualAlpha = alpha * UBlasTraits::extractScalarFactor(u.derived()) + * numext::conj(VBlasTraits::extractScalarFactor(v.derived())); + if (IsRowMajor) + actualAlpha = numext::conj(actualAlpha); + + typedef typename internal::remove_all<typename internal::conj_expr_if<IsRowMajor ^ UBlasTraits::NeedToConjugate,_ActualUType>::type>::type UType; + typedef typename internal::remove_all<typename internal::conj_expr_if<IsRowMajor ^ VBlasTraits::NeedToConjugate,_ActualVType>::type>::type VType; + internal::selfadjoint_rank2_update_selector<Scalar, Index, UType, VType, + (IsRowMajor ? int(UpLo==Upper ? Lower : Upper) : UpLo)> + ::run(_expression().const_cast_derived().data(),_expression().outerStride(),UType(actualU),VType(actualV),actualAlpha); + + return *this; +} + +} // end namespace Eigen + +#endif // EIGEN_SELFADJOINTRANK2UPTADE_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/TriangularMatrixMatrix.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/TriangularMatrixMatrix.h new file mode 100644 index 000000000..6ec5a8a0b --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/TriangularMatrixMatrix.h @@ -0,0 +1,441 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2009 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_TRIANGULAR_MATRIX_MATRIX_H +#define EIGEN_TRIANGULAR_MATRIX_MATRIX_H + +namespace Eigen { + +namespace internal { + +// template<typename Scalar, int mr, int StorageOrder, bool Conjugate, int Mode> +// struct gemm_pack_lhs_triangular +// { +// Matrix<Scalar,mr,mr, +// void operator()(Scalar* blockA, const EIGEN_RESTRICT Scalar* _lhs, int lhsStride, int depth, int rows) +// { +// conj_if<NumTraits<Scalar>::IsComplex && Conjugate> cj; +// const_blas_data_mapper<Scalar, StorageOrder> lhs(_lhs,lhsStride); +// int count = 0; +// const int peeled_mc = (rows/mr)*mr; +// for(int i=0; i<peeled_mc; i+=mr) +// { +// for(int k=0; k<depth; k++) +// for(int w=0; w<mr; w++) +// blockA[count++] = cj(lhs(i+w, k)); +// } +// for(int i=peeled_mc; i<rows; i++) +// { +// for(int k=0; k<depth; k++) +// blockA[count++] = cj(lhs(i, k)); +// } +// } +// }; + +/* Optimized triangular matrix * matrix (_TRMM++) product built on top of + * the general matrix matrix product. + */ +template <typename Scalar, typename Index, + int Mode, bool LhsIsTriangular, + int LhsStorageOrder, bool ConjugateLhs, + int RhsStorageOrder, bool ConjugateRhs, + int ResStorageOrder, int Version = Specialized> +struct product_triangular_matrix_matrix; + +template <typename Scalar, typename Index, + int Mode, bool LhsIsTriangular, + int LhsStorageOrder, bool ConjugateLhs, + int RhsStorageOrder, bool ConjugateRhs, int Version> +struct product_triangular_matrix_matrix<Scalar,Index,Mode,LhsIsTriangular, + LhsStorageOrder,ConjugateLhs, + RhsStorageOrder,ConjugateRhs,RowMajor,Version> +{ + static EIGEN_STRONG_INLINE void run( + Index rows, Index cols, Index depth, + const Scalar* lhs, Index lhsStride, + const Scalar* rhs, Index rhsStride, + Scalar* res, Index resStride, + const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking) + { + product_triangular_matrix_matrix<Scalar, Index, + (Mode&(UnitDiag|ZeroDiag)) | ((Mode&Upper) ? Lower : Upper), + (!LhsIsTriangular), + RhsStorageOrder==RowMajor ? ColMajor : RowMajor, + ConjugateRhs, + LhsStorageOrder==RowMajor ? ColMajor : RowMajor, + ConjugateLhs, + ColMajor> + ::run(cols, rows, depth, rhs, rhsStride, lhs, lhsStride, res, resStride, alpha, blocking); + } +}; + +// implements col-major += alpha * op(triangular) * op(general) +template <typename Scalar, typename Index, int Mode, + int LhsStorageOrder, bool ConjugateLhs, + int RhsStorageOrder, bool ConjugateRhs, int Version> +struct product_triangular_matrix_matrix<Scalar,Index,Mode,true, + LhsStorageOrder,ConjugateLhs, + RhsStorageOrder,ConjugateRhs,ColMajor,Version> +{ + + typedef gebp_traits<Scalar,Scalar> Traits; + enum { + SmallPanelWidth = 2 * EIGEN_PLAIN_ENUM_MAX(Traits::mr,Traits::nr), + IsLower = (Mode&Lower) == Lower, + SetDiag = (Mode&(ZeroDiag|UnitDiag)) ? 0 : 1 + }; + + static EIGEN_DONT_INLINE void run( + Index _rows, Index _cols, Index _depth, + const Scalar* _lhs, Index lhsStride, + const Scalar* _rhs, Index rhsStride, + Scalar* res, Index resStride, + const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking); +}; + +template <typename Scalar, typename Index, int Mode, + int LhsStorageOrder, bool ConjugateLhs, + int RhsStorageOrder, bool ConjugateRhs, int Version> +EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true, + LhsStorageOrder,ConjugateLhs, + RhsStorageOrder,ConjugateRhs,ColMajor,Version>::run( + Index _rows, Index _cols, Index _depth, + const Scalar* _lhs, Index lhsStride, + const Scalar* _rhs, Index rhsStride, + Scalar* _res, Index resStride, + const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking) + { + // strip zeros + Index diagSize = (std::min)(_rows,_depth); + Index rows = IsLower ? _rows : diagSize; + Index depth = IsLower ? diagSize : _depth; + Index cols = _cols; + + typedef const_blas_data_mapper<Scalar, Index, LhsStorageOrder> LhsMapper; + typedef const_blas_data_mapper<Scalar, Index, RhsStorageOrder> RhsMapper; + typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper; + LhsMapper lhs(_lhs,lhsStride); + RhsMapper rhs(_rhs,rhsStride); + ResMapper res(_res, resStride); + + Index kc = blocking.kc(); // cache block size along the K direction + Index mc = (std::min)(rows,blocking.mc()); // cache block size along the M direction + // The small panel size must not be larger than blocking size. + // Usually this should never be the case because SmallPanelWidth^2 is very small + // compared to L2 cache size, but let's be safe: + Index panelWidth = (std::min)(Index(SmallPanelWidth),(std::min)(kc,mc)); + + std::size_t sizeA = kc*mc; + std::size_t sizeB = kc*cols; + + ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA()); + ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB()); + + Matrix<Scalar,SmallPanelWidth,SmallPanelWidth,LhsStorageOrder> triangularBuffer((internal::constructor_without_unaligned_array_assert())); + triangularBuffer.setZero(); + if((Mode&ZeroDiag)==ZeroDiag) + triangularBuffer.diagonal().setZero(); + else + triangularBuffer.diagonal().setOnes(); + + gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel; + gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs; + gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder> pack_rhs; + + for(Index k2=IsLower ? depth : 0; + IsLower ? k2>0 : k2<depth; + IsLower ? k2-=kc : k2+=kc) + { + Index actual_kc = (std::min)(IsLower ? k2 : depth-k2, kc); + Index actual_k2 = IsLower ? k2-actual_kc : k2; + + // align blocks with the end of the triangular part for trapezoidal lhs + if((!IsLower)&&(k2<rows)&&(k2+actual_kc>rows)) + { + actual_kc = rows-k2; + k2 = k2+actual_kc-kc; + } + + pack_rhs(blockB, rhs.getSubMapper(actual_k2,0), actual_kc, cols); + + // the selected lhs's panel has to be split in three different parts: + // 1 - the part which is zero => skip it + // 2 - the diagonal block => special kernel + // 3 - the dense panel below (lower case) or above (upper case) the diagonal block => GEPP + + // the block diagonal, if any: + if(IsLower || actual_k2<rows) + { + // for each small vertical panels of lhs + for (Index k1=0; k1<actual_kc; k1+=panelWidth) + { + Index actualPanelWidth = std::min<Index>(actual_kc-k1, panelWidth); + Index lengthTarget = IsLower ? actual_kc-k1-actualPanelWidth : k1; + Index startBlock = actual_k2+k1; + Index blockBOffset = k1; + + // => GEBP with the micro triangular block + // The trick is to pack this micro block while filling the opposite triangular part with zeros. + // To this end we do an extra triangular copy to a small temporary buffer + for (Index k=0;k<actualPanelWidth;++k) + { + if (SetDiag) + triangularBuffer.coeffRef(k,k) = lhs(startBlock+k,startBlock+k); + for (Index i=IsLower ? k+1 : 0; IsLower ? i<actualPanelWidth : i<k; ++i) + triangularBuffer.coeffRef(i,k) = lhs(startBlock+i,startBlock+k); + } + pack_lhs(blockA, LhsMapper(triangularBuffer.data(), triangularBuffer.outerStride()), actualPanelWidth, actualPanelWidth); + + gebp_kernel(res.getSubMapper(startBlock, 0), blockA, blockB, + actualPanelWidth, actualPanelWidth, cols, alpha, + actualPanelWidth, actual_kc, 0, blockBOffset); + + // GEBP with remaining micro panel + if (lengthTarget>0) + { + Index startTarget = IsLower ? actual_k2+k1+actualPanelWidth : actual_k2; + + pack_lhs(blockA, lhs.getSubMapper(startTarget,startBlock), actualPanelWidth, lengthTarget); + + gebp_kernel(res.getSubMapper(startTarget, 0), blockA, blockB, + lengthTarget, actualPanelWidth, cols, alpha, + actualPanelWidth, actual_kc, 0, blockBOffset); + } + } + } + // the part below (lower case) or above (upper case) the diagonal => GEPP + { + Index start = IsLower ? k2 : 0; + Index end = IsLower ? rows : (std::min)(actual_k2,rows); + for(Index i2=start; i2<end; i2+=mc) + { + const Index actual_mc = (std::min)(i2+mc,end)-i2; + gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr,Traits::LhsProgress, LhsStorageOrder,false>() + (blockA, lhs.getSubMapper(i2, actual_k2), actual_kc, actual_mc); + + gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, + actual_kc, cols, alpha, -1, -1, 0, 0); + } + } + } + } + +// implements col-major += alpha * op(general) * op(triangular) +template <typename Scalar, typename Index, int Mode, + int LhsStorageOrder, bool ConjugateLhs, + int RhsStorageOrder, bool ConjugateRhs, int Version> +struct product_triangular_matrix_matrix<Scalar,Index,Mode,false, + LhsStorageOrder,ConjugateLhs, + RhsStorageOrder,ConjugateRhs,ColMajor,Version> +{ + typedef gebp_traits<Scalar,Scalar> Traits; + enum { + SmallPanelWidth = EIGEN_PLAIN_ENUM_MAX(Traits::mr,Traits::nr), + IsLower = (Mode&Lower) == Lower, + SetDiag = (Mode&(ZeroDiag|UnitDiag)) ? 0 : 1 + }; + + static EIGEN_DONT_INLINE void run( + Index _rows, Index _cols, Index _depth, + const Scalar* _lhs, Index lhsStride, + const Scalar* _rhs, Index rhsStride, + Scalar* res, Index resStride, + const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking); +}; + +template <typename Scalar, typename Index, int Mode, + int LhsStorageOrder, bool ConjugateLhs, + int RhsStorageOrder, bool ConjugateRhs, int Version> +EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false, + LhsStorageOrder,ConjugateLhs, + RhsStorageOrder,ConjugateRhs,ColMajor,Version>::run( + Index _rows, Index _cols, Index _depth, + const Scalar* _lhs, Index lhsStride, + const Scalar* _rhs, Index rhsStride, + Scalar* _res, Index resStride, + const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking) + { + const Index PacketBytes = packet_traits<Scalar>::size*sizeof(Scalar); + // strip zeros + Index diagSize = (std::min)(_cols,_depth); + Index rows = _rows; + Index depth = IsLower ? _depth : diagSize; + Index cols = IsLower ? diagSize : _cols; + + typedef const_blas_data_mapper<Scalar, Index, LhsStorageOrder> LhsMapper; + typedef const_blas_data_mapper<Scalar, Index, RhsStorageOrder> RhsMapper; + typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper; + LhsMapper lhs(_lhs,lhsStride); + RhsMapper rhs(_rhs,rhsStride); + ResMapper res(_res, resStride); + + Index kc = blocking.kc(); // cache block size along the K direction + Index mc = (std::min)(rows,blocking.mc()); // cache block size along the M direction + + std::size_t sizeA = kc*mc; + std::size_t sizeB = kc*cols+EIGEN_MAX_ALIGN_BYTES/sizeof(Scalar); + + ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA()); + ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB()); + + Matrix<Scalar,SmallPanelWidth,SmallPanelWidth,RhsStorageOrder> triangularBuffer((internal::constructor_without_unaligned_array_assert())); + triangularBuffer.setZero(); + if((Mode&ZeroDiag)==ZeroDiag) + triangularBuffer.diagonal().setZero(); + else + triangularBuffer.diagonal().setOnes(); + + gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel; + gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs; + gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder> pack_rhs; + gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder,false,true> pack_rhs_panel; + + for(Index k2=IsLower ? 0 : depth; + IsLower ? k2<depth : k2>0; + IsLower ? k2+=kc : k2-=kc) + { + Index actual_kc = (std::min)(IsLower ? depth-k2 : k2, kc); + Index actual_k2 = IsLower ? k2 : k2-actual_kc; + + // align blocks with the end of the triangular part for trapezoidal rhs + if(IsLower && (k2<cols) && (actual_k2+actual_kc>cols)) + { + actual_kc = cols-k2; + k2 = actual_k2 + actual_kc - kc; + } + + // remaining size + Index rs = IsLower ? (std::min)(cols,actual_k2) : cols - k2; + // size of the triangular part + Index ts = (IsLower && actual_k2>=cols) ? 0 : actual_kc; + + Scalar* geb = blockB+ts*ts; + geb = geb + internal::first_aligned<PacketBytes>(geb,PacketBytes/sizeof(Scalar)); + + pack_rhs(geb, rhs.getSubMapper(actual_k2,IsLower ? 0 : k2), actual_kc, rs); + + // pack the triangular part of the rhs padding the unrolled blocks with zeros + if(ts>0) + { + for (Index j2=0; j2<actual_kc; j2+=SmallPanelWidth) + { + Index actualPanelWidth = std::min<Index>(actual_kc-j2, SmallPanelWidth); + Index actual_j2 = actual_k2 + j2; + Index panelOffset = IsLower ? j2+actualPanelWidth : 0; + Index panelLength = IsLower ? actual_kc-j2-actualPanelWidth : j2; + // general part + pack_rhs_panel(blockB+j2*actual_kc, + rhs.getSubMapper(actual_k2+panelOffset, actual_j2), + panelLength, actualPanelWidth, + actual_kc, panelOffset); + + // append the triangular part via a temporary buffer + for (Index j=0;j<actualPanelWidth;++j) + { + if (SetDiag) + triangularBuffer.coeffRef(j,j) = rhs(actual_j2+j,actual_j2+j); + for (Index k=IsLower ? j+1 : 0; IsLower ? k<actualPanelWidth : k<j; ++k) + triangularBuffer.coeffRef(k,j) = rhs(actual_j2+k,actual_j2+j); + } + + pack_rhs_panel(blockB+j2*actual_kc, + RhsMapper(triangularBuffer.data(), triangularBuffer.outerStride()), + actualPanelWidth, actualPanelWidth, + actual_kc, j2); + } + } + + for (Index i2=0; i2<rows; i2+=mc) + { + const Index actual_mc = (std::min)(mc,rows-i2); + pack_lhs(blockA, lhs.getSubMapper(i2, actual_k2), actual_kc, actual_mc); + + // triangular kernel + if(ts>0) + { + for (Index j2=0; j2<actual_kc; j2+=SmallPanelWidth) + { + Index actualPanelWidth = std::min<Index>(actual_kc-j2, SmallPanelWidth); + Index panelLength = IsLower ? actual_kc-j2 : j2+actualPanelWidth; + Index blockOffset = IsLower ? j2 : 0; + + gebp_kernel(res.getSubMapper(i2, actual_k2 + j2), + blockA, blockB+j2*actual_kc, + actual_mc, panelLength, actualPanelWidth, + alpha, + actual_kc, actual_kc, // strides + blockOffset, blockOffset);// offsets + } + } + gebp_kernel(res.getSubMapper(i2, IsLower ? 0 : k2), + blockA, geb, actual_mc, actual_kc, rs, + alpha, + -1, -1, 0, 0); + } + } + } + +/*************************************************************************** +* Wrapper to product_triangular_matrix_matrix +***************************************************************************/ + +} // end namespace internal + +namespace internal { +template<int Mode, bool LhsIsTriangular, typename Lhs, typename Rhs> +struct triangular_product_impl<Mode,LhsIsTriangular,Lhs,false,Rhs,false> +{ + template<typename Dest> static void run(Dest& dst, const Lhs &a_lhs, const Rhs &a_rhs, const typename Dest::Scalar& alpha) + { + typedef typename Dest::Scalar Scalar; + + typedef internal::blas_traits<Lhs> LhsBlasTraits; + typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType; + typedef typename internal::remove_all<ActualLhsType>::type ActualLhsTypeCleaned; + typedef internal::blas_traits<Rhs> RhsBlasTraits; + typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType; + typedef typename internal::remove_all<ActualRhsType>::type ActualRhsTypeCleaned; + + typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(a_lhs); + typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(a_rhs); + + Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(a_lhs) + * RhsBlasTraits::extractScalarFactor(a_rhs); + + typedef internal::gemm_blocking_space<(Dest::Flags&RowMajorBit) ? RowMajor : ColMajor,Scalar,Scalar, + Lhs::MaxRowsAtCompileTime, Rhs::MaxColsAtCompileTime, Lhs::MaxColsAtCompileTime,4> BlockingType; + + enum { IsLower = (Mode&Lower) == Lower }; + Index stripedRows = ((!LhsIsTriangular) || (IsLower)) ? lhs.rows() : (std::min)(lhs.rows(),lhs.cols()); + Index stripedCols = ((LhsIsTriangular) || (!IsLower)) ? rhs.cols() : (std::min)(rhs.cols(),rhs.rows()); + Index stripedDepth = LhsIsTriangular ? ((!IsLower) ? lhs.cols() : (std::min)(lhs.cols(),lhs.rows())) + : ((IsLower) ? rhs.rows() : (std::min)(rhs.rows(),rhs.cols())); + + BlockingType blocking(stripedRows, stripedCols, stripedDepth, 1, false); + + internal::product_triangular_matrix_matrix<Scalar, Index, + Mode, LhsIsTriangular, + (internal::traits<ActualLhsTypeCleaned>::Flags&RowMajorBit) ? RowMajor : ColMajor, LhsBlasTraits::NeedToConjugate, + (internal::traits<ActualRhsTypeCleaned>::Flags&RowMajorBit) ? RowMajor : ColMajor, RhsBlasTraits::NeedToConjugate, + (internal::traits<Dest >::Flags&RowMajorBit) ? RowMajor : ColMajor> + ::run( + stripedRows, stripedCols, stripedDepth, // sizes + &lhs.coeffRef(0,0), lhs.outerStride(), // lhs info + &rhs.coeffRef(0,0), rhs.outerStride(), // rhs info + &dst.coeffRef(0,0), dst.outerStride(), // result info + actualAlpha, blocking + ); + } +}; + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_TRIANGULAR_MATRIX_MATRIX_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/TriangularMatrixMatrix_BLAS.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/TriangularMatrixMatrix_BLAS.h new file mode 100644 index 000000000..aecded6bb --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/TriangularMatrixMatrix_BLAS.h @@ -0,0 +1,302 @@ +/* + Copyright (c) 2011, Intel Corporation. All rights reserved. + + Redistribution and use in source and binary forms, with or without modification, + are permitted provided that the following conditions are met: + + * Redistributions of source code must retain the above copyright notice, this + list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above copyright notice, + this list of conditions and the following disclaimer in the documentation + and/or other materials provided with the distribution. + * Neither the name of Intel Corporation nor the names of its contributors may + be used to endorse or promote products derived from this software without + specific prior written permission. + + THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND + ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED + WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE + DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR + ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES + (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; + LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON + ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS + SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + + ******************************************************************************** + * Content : Eigen bindings to BLAS F77 + * Triangular matrix * matrix product functionality based on ?TRMM. + ******************************************************************************** +*/ + +#ifndef EIGEN_TRIANGULAR_MATRIX_MATRIX_BLAS_H +#define EIGEN_TRIANGULAR_MATRIX_MATRIX_BLAS_H + +namespace Eigen { + +namespace internal { + + +template <typename Scalar, typename Index, + int Mode, bool LhsIsTriangular, + int LhsStorageOrder, bool ConjugateLhs, + int RhsStorageOrder, bool ConjugateRhs, + int ResStorageOrder> +struct product_triangular_matrix_matrix_trmm : + product_triangular_matrix_matrix<Scalar,Index,Mode, + LhsIsTriangular,LhsStorageOrder,ConjugateLhs, + RhsStorageOrder, ConjugateRhs, ResStorageOrder, BuiltIn> {}; + + +// try to go to BLAS specialization +#define EIGEN_BLAS_TRMM_SPECIALIZE(Scalar, LhsIsTriangular) \ +template <typename Index, int Mode, \ + int LhsStorageOrder, bool ConjugateLhs, \ + int RhsStorageOrder, bool ConjugateRhs> \ +struct product_triangular_matrix_matrix<Scalar,Index, Mode, LhsIsTriangular, \ + LhsStorageOrder,ConjugateLhs, RhsStorageOrder,ConjugateRhs,ColMajor,Specialized> { \ + static inline void run(Index _rows, Index _cols, Index _depth, const Scalar* _lhs, Index lhsStride,\ + const Scalar* _rhs, Index rhsStride, Scalar* res, Index resStride, Scalar alpha, level3_blocking<Scalar,Scalar>& blocking) { \ + product_triangular_matrix_matrix_trmm<Scalar,Index,Mode, \ + LhsIsTriangular,LhsStorageOrder,ConjugateLhs, \ + RhsStorageOrder, ConjugateRhs, ColMajor>::run( \ + _rows, _cols, _depth, _lhs, lhsStride, _rhs, rhsStride, res, resStride, alpha, blocking); \ + } \ +}; + +EIGEN_BLAS_TRMM_SPECIALIZE(double, true) +EIGEN_BLAS_TRMM_SPECIALIZE(double, false) +EIGEN_BLAS_TRMM_SPECIALIZE(dcomplex, true) +EIGEN_BLAS_TRMM_SPECIALIZE(dcomplex, false) +EIGEN_BLAS_TRMM_SPECIALIZE(float, true) +EIGEN_BLAS_TRMM_SPECIALIZE(float, false) +EIGEN_BLAS_TRMM_SPECIALIZE(scomplex, true) +EIGEN_BLAS_TRMM_SPECIALIZE(scomplex, false) + +// implements col-major += alpha * op(triangular) * op(general) +#define EIGEN_BLAS_TRMM_L(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \ +template <typename Index, int Mode, \ + int LhsStorageOrder, bool ConjugateLhs, \ + int RhsStorageOrder, bool ConjugateRhs> \ +struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,true, \ + LhsStorageOrder,ConjugateLhs,RhsStorageOrder,ConjugateRhs,ColMajor> \ +{ \ + enum { \ + IsLower = (Mode&Lower) == Lower, \ + SetDiag = (Mode&(ZeroDiag|UnitDiag)) ? 0 : 1, \ + IsUnitDiag = (Mode&UnitDiag) ? 1 : 0, \ + IsZeroDiag = (Mode&ZeroDiag) ? 1 : 0, \ + LowUp = IsLower ? Lower : Upper, \ + conjA = ((LhsStorageOrder==ColMajor) && ConjugateLhs) ? 1 : 0 \ + }; \ +\ + static void run( \ + Index _rows, Index _cols, Index _depth, \ + const EIGTYPE* _lhs, Index lhsStride, \ + const EIGTYPE* _rhs, Index rhsStride, \ + EIGTYPE* res, Index resStride, \ + EIGTYPE alpha, level3_blocking<EIGTYPE,EIGTYPE>& blocking) \ + { \ + Index diagSize = (std::min)(_rows,_depth); \ + Index rows = IsLower ? _rows : diagSize; \ + Index depth = IsLower ? diagSize : _depth; \ + Index cols = _cols; \ +\ + typedef Matrix<EIGTYPE, Dynamic, Dynamic, LhsStorageOrder> MatrixLhs; \ + typedef Matrix<EIGTYPE, Dynamic, Dynamic, RhsStorageOrder> MatrixRhs; \ +\ +/* Non-square case - doesn't fit to BLAS ?TRMM. Fall to default triangular product or call BLAS ?GEMM*/ \ + if (rows != depth) { \ +\ + /* FIXME handle mkl_domain_get_max_threads */ \ + /*int nthr = mkl_domain_get_max_threads(EIGEN_BLAS_DOMAIN_BLAS);*/ int nthr = 1;\ +\ + if (((nthr==1) && (((std::max)(rows,depth)-diagSize)/(double)diagSize < 0.5))) { \ + /* Most likely no benefit to call TRMM or GEMM from BLAS */ \ + product_triangular_matrix_matrix<EIGTYPE,Index,Mode,true, \ + LhsStorageOrder,ConjugateLhs, RhsStorageOrder, ConjugateRhs, ColMajor, BuiltIn>::run( \ + _rows, _cols, _depth, _lhs, lhsStride, _rhs, rhsStride, res, resStride, alpha, blocking); \ + /*std::cout << "TRMM_L: A is not square! Go to Eigen TRMM implementation!\n";*/ \ + } else { \ + /* Make sense to call GEMM */ \ + Map<const MatrixLhs, 0, OuterStride<> > lhsMap(_lhs,rows,depth,OuterStride<>(lhsStride)); \ + MatrixLhs aa_tmp=lhsMap.template triangularView<Mode>(); \ + BlasIndex aStride = convert_index<BlasIndex>(aa_tmp.outerStride()); \ + gemm_blocking_space<ColMajor,EIGTYPE,EIGTYPE,Dynamic,Dynamic,Dynamic> gemm_blocking(_rows,_cols,_depth, 1, true); \ + general_matrix_matrix_product<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,RhsStorageOrder,ConjugateRhs,ColMajor>::run( \ + rows, cols, depth, aa_tmp.data(), aStride, _rhs, rhsStride, res, resStride, alpha, gemm_blocking, 0); \ +\ + /*std::cout << "TRMM_L: A is not square! Go to BLAS GEMM implementation! " << nthr<<" \n";*/ \ + } \ + return; \ + } \ + char side = 'L', transa, uplo, diag = 'N'; \ + EIGTYPE *b; \ + const EIGTYPE *a; \ + BlasIndex m, n, lda, ldb; \ +\ +/* Set m, n */ \ + m = convert_index<BlasIndex>(diagSize); \ + n = convert_index<BlasIndex>(cols); \ +\ +/* Set trans */ \ + transa = (LhsStorageOrder==RowMajor) ? ((ConjugateLhs) ? 'C' : 'T') : 'N'; \ +\ +/* Set b, ldb */ \ + Map<const MatrixRhs, 0, OuterStride<> > rhs(_rhs,depth,cols,OuterStride<>(rhsStride)); \ + MatrixX##EIGPREFIX b_tmp; \ +\ + if (ConjugateRhs) b_tmp = rhs.conjugate(); else b_tmp = rhs; \ + b = b_tmp.data(); \ + ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \ +\ +/* Set uplo */ \ + uplo = IsLower ? 'L' : 'U'; \ + if (LhsStorageOrder==RowMajor) uplo = (uplo == 'L') ? 'U' : 'L'; \ +/* Set a, lda */ \ + Map<const MatrixLhs, 0, OuterStride<> > lhs(_lhs,rows,depth,OuterStride<>(lhsStride)); \ + MatrixLhs a_tmp; \ +\ + if ((conjA!=0) || (SetDiag==0)) { \ + if (conjA) a_tmp = lhs.conjugate(); else a_tmp = lhs; \ + if (IsZeroDiag) \ + a_tmp.diagonal().setZero(); \ + else if (IsUnitDiag) \ + a_tmp.diagonal().setOnes();\ + a = a_tmp.data(); \ + lda = convert_index<BlasIndex>(a_tmp.outerStride()); \ + } else { \ + a = _lhs; \ + lda = convert_index<BlasIndex>(lhsStride); \ + } \ + /*std::cout << "TRMM_L: A is square! Go to BLAS TRMM implementation! \n";*/ \ +/* call ?trmm*/ \ + BLASPREFIX##trmm_(&side, &uplo, &transa, &diag, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (BLASTYPE*)b, &ldb); \ +\ +/* Add op(a_triangular)*b into res*/ \ + Map<MatrixX##EIGPREFIX, 0, OuterStride<> > res_tmp(res,rows,cols,OuterStride<>(resStride)); \ + res_tmp=res_tmp+b_tmp; \ + } \ +}; + +EIGEN_BLAS_TRMM_L(double, double, d, d) +EIGEN_BLAS_TRMM_L(dcomplex, double, cd, z) +EIGEN_BLAS_TRMM_L(float, float, f, s) +EIGEN_BLAS_TRMM_L(scomplex, float, cf, c) + +// implements col-major += alpha * op(general) * op(triangular) +#define EIGEN_BLAS_TRMM_R(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \ +template <typename Index, int Mode, \ + int LhsStorageOrder, bool ConjugateLhs, \ + int RhsStorageOrder, bool ConjugateRhs> \ +struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,false, \ + LhsStorageOrder,ConjugateLhs,RhsStorageOrder,ConjugateRhs,ColMajor> \ +{ \ + enum { \ + IsLower = (Mode&Lower) == Lower, \ + SetDiag = (Mode&(ZeroDiag|UnitDiag)) ? 0 : 1, \ + IsUnitDiag = (Mode&UnitDiag) ? 1 : 0, \ + IsZeroDiag = (Mode&ZeroDiag) ? 1 : 0, \ + LowUp = IsLower ? Lower : Upper, \ + conjA = ((RhsStorageOrder==ColMajor) && ConjugateRhs) ? 1 : 0 \ + }; \ +\ + static void run( \ + Index _rows, Index _cols, Index _depth, \ + const EIGTYPE* _lhs, Index lhsStride, \ + const EIGTYPE* _rhs, Index rhsStride, \ + EIGTYPE* res, Index resStride, \ + EIGTYPE alpha, level3_blocking<EIGTYPE,EIGTYPE>& blocking) \ + { \ + Index diagSize = (std::min)(_cols,_depth); \ + Index rows = _rows; \ + Index depth = IsLower ? _depth : diagSize; \ + Index cols = IsLower ? diagSize : _cols; \ +\ + typedef Matrix<EIGTYPE, Dynamic, Dynamic, LhsStorageOrder> MatrixLhs; \ + typedef Matrix<EIGTYPE, Dynamic, Dynamic, RhsStorageOrder> MatrixRhs; \ +\ +/* Non-square case - doesn't fit to BLAS ?TRMM. Fall to default triangular product or call BLAS ?GEMM*/ \ + if (cols != depth) { \ +\ + int nthr = 1 /*mkl_domain_get_max_threads(EIGEN_BLAS_DOMAIN_BLAS)*/; \ +\ + if ((nthr==1) && (((std::max)(cols,depth)-diagSize)/(double)diagSize < 0.5)) { \ + /* Most likely no benefit to call TRMM or GEMM from BLAS*/ \ + product_triangular_matrix_matrix<EIGTYPE,Index,Mode,false, \ + LhsStorageOrder,ConjugateLhs, RhsStorageOrder, ConjugateRhs, ColMajor, BuiltIn>::run( \ + _rows, _cols, _depth, _lhs, lhsStride, _rhs, rhsStride, res, resStride, alpha, blocking); \ + /*std::cout << "TRMM_R: A is not square! Go to Eigen TRMM implementation!\n";*/ \ + } else { \ + /* Make sense to call GEMM */ \ + Map<const MatrixRhs, 0, OuterStride<> > rhsMap(_rhs,depth,cols, OuterStride<>(rhsStride)); \ + MatrixRhs aa_tmp=rhsMap.template triangularView<Mode>(); \ + BlasIndex aStride = convert_index<BlasIndex>(aa_tmp.outerStride()); \ + gemm_blocking_space<ColMajor,EIGTYPE,EIGTYPE,Dynamic,Dynamic,Dynamic> gemm_blocking(_rows,_cols,_depth, 1, true); \ + general_matrix_matrix_product<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,RhsStorageOrder,ConjugateRhs,ColMajor>::run( \ + rows, cols, depth, _lhs, lhsStride, aa_tmp.data(), aStride, res, resStride, alpha, gemm_blocking, 0); \ +\ + /*std::cout << "TRMM_R: A is not square! Go to BLAS GEMM implementation! " << nthr<<" \n";*/ \ + } \ + return; \ + } \ + char side = 'R', transa, uplo, diag = 'N'; \ + EIGTYPE *b; \ + const EIGTYPE *a; \ + BlasIndex m, n, lda, ldb; \ +\ +/* Set m, n */ \ + m = convert_index<BlasIndex>(rows); \ + n = convert_index<BlasIndex>(diagSize); \ +\ +/* Set trans */ \ + transa = (RhsStorageOrder==RowMajor) ? ((ConjugateRhs) ? 'C' : 'T') : 'N'; \ +\ +/* Set b, ldb */ \ + Map<const MatrixLhs, 0, OuterStride<> > lhs(_lhs,rows,depth,OuterStride<>(lhsStride)); \ + MatrixX##EIGPREFIX b_tmp; \ +\ + if (ConjugateLhs) b_tmp = lhs.conjugate(); else b_tmp = lhs; \ + b = b_tmp.data(); \ + ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \ +\ +/* Set uplo */ \ + uplo = IsLower ? 'L' : 'U'; \ + if (RhsStorageOrder==RowMajor) uplo = (uplo == 'L') ? 'U' : 'L'; \ +/* Set a, lda */ \ + Map<const MatrixRhs, 0, OuterStride<> > rhs(_rhs,depth,cols, OuterStride<>(rhsStride)); \ + MatrixRhs a_tmp; \ +\ + if ((conjA!=0) || (SetDiag==0)) { \ + if (conjA) a_tmp = rhs.conjugate(); else a_tmp = rhs; \ + if (IsZeroDiag) \ + a_tmp.diagonal().setZero(); \ + else if (IsUnitDiag) \ + a_tmp.diagonal().setOnes();\ + a = a_tmp.data(); \ + lda = convert_index<BlasIndex>(a_tmp.outerStride()); \ + } else { \ + a = _rhs; \ + lda = convert_index<BlasIndex>(rhsStride); \ + } \ + /*std::cout << "TRMM_R: A is square! Go to BLAS TRMM implementation! \n";*/ \ +/* call ?trmm*/ \ + BLASPREFIX##trmm_(&side, &uplo, &transa, &diag, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (BLASTYPE*)b, &ldb); \ +\ +/* Add op(a_triangular)*b into res*/ \ + Map<MatrixX##EIGPREFIX, 0, OuterStride<> > res_tmp(res,rows,cols,OuterStride<>(resStride)); \ + res_tmp=res_tmp+b_tmp; \ + } \ +}; + +EIGEN_BLAS_TRMM_R(double, double, d, d) +EIGEN_BLAS_TRMM_R(dcomplex, double, cd, z) +EIGEN_BLAS_TRMM_R(float, float, f, s) +EIGEN_BLAS_TRMM_R(scomplex, float, cf, c) + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_TRIANGULAR_MATRIX_MATRIX_BLAS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/TriangularMatrixVector.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/TriangularMatrixVector.h new file mode 100644 index 000000000..4b292e74d --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/TriangularMatrixVector.h @@ -0,0 +1,336 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2009 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_TRIANGULARMATRIXVECTOR_H +#define EIGEN_TRIANGULARMATRIXVECTOR_H + +namespace Eigen { + +namespace internal { + +template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs, int StorageOrder, int Version=Specialized> +struct triangular_matrix_vector_product; + +template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs, int Version> +struct triangular_matrix_vector_product<Index,Mode,LhsScalar,ConjLhs,RhsScalar,ConjRhs,ColMajor,Version> +{ + typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar; + enum { + IsLower = ((Mode&Lower)==Lower), + HasUnitDiag = (Mode & UnitDiag)==UnitDiag, + HasZeroDiag = (Mode & ZeroDiag)==ZeroDiag + }; + static EIGEN_DONT_INLINE void run(Index _rows, Index _cols, const LhsScalar* _lhs, Index lhsStride, + const RhsScalar* _rhs, Index rhsIncr, ResScalar* _res, Index resIncr, const RhsScalar& alpha); +}; + +template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs, int Version> +EIGEN_DONT_INLINE void triangular_matrix_vector_product<Index,Mode,LhsScalar,ConjLhs,RhsScalar,ConjRhs,ColMajor,Version> + ::run(Index _rows, Index _cols, const LhsScalar* _lhs, Index lhsStride, + const RhsScalar* _rhs, Index rhsIncr, ResScalar* _res, Index resIncr, const RhsScalar& alpha) + { + static const Index PanelWidth = EIGEN_TUNE_TRIANGULAR_PANEL_WIDTH; + Index size = (std::min)(_rows,_cols); + Index rows = IsLower ? _rows : (std::min)(_rows,_cols); + Index cols = IsLower ? (std::min)(_rows,_cols) : _cols; + + typedef Map<const Matrix<LhsScalar,Dynamic,Dynamic,ColMajor>, 0, OuterStride<> > LhsMap; + const LhsMap lhs(_lhs,rows,cols,OuterStride<>(lhsStride)); + typename conj_expr_if<ConjLhs,LhsMap>::type cjLhs(lhs); + + typedef Map<const Matrix<RhsScalar,Dynamic,1>, 0, InnerStride<> > RhsMap; + const RhsMap rhs(_rhs,cols,InnerStride<>(rhsIncr)); + typename conj_expr_if<ConjRhs,RhsMap>::type cjRhs(rhs); + + typedef Map<Matrix<ResScalar,Dynamic,1> > ResMap; + ResMap res(_res,rows); + + typedef const_blas_data_mapper<LhsScalar,Index,ColMajor> LhsMapper; + typedef const_blas_data_mapper<RhsScalar,Index,RowMajor> RhsMapper; + + for (Index pi=0; pi<size; pi+=PanelWidth) + { + Index actualPanelWidth = (std::min)(PanelWidth, size-pi); + for (Index k=0; k<actualPanelWidth; ++k) + { + Index i = pi + k; + Index s = IsLower ? ((HasUnitDiag||HasZeroDiag) ? i+1 : i ) : pi; + Index r = IsLower ? actualPanelWidth-k : k+1; + if ((!(HasUnitDiag||HasZeroDiag)) || (--r)>0) + res.segment(s,r) += (alpha * cjRhs.coeff(i)) * cjLhs.col(i).segment(s,r); + if (HasUnitDiag) + res.coeffRef(i) += alpha * cjRhs.coeff(i); + } + Index r = IsLower ? rows - pi - actualPanelWidth : pi; + if (r>0) + { + Index s = IsLower ? pi+actualPanelWidth : 0; + general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,ConjLhs,RhsScalar,RhsMapper,ConjRhs,BuiltIn>::run( + r, actualPanelWidth, + LhsMapper(&lhs.coeffRef(s,pi), lhsStride), + RhsMapper(&rhs.coeffRef(pi), rhsIncr), + &res.coeffRef(s), resIncr, alpha); + } + } + if((!IsLower) && cols>size) + { + general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,ConjLhs,RhsScalar,RhsMapper,ConjRhs>::run( + rows, cols-size, + LhsMapper(&lhs.coeffRef(0,size), lhsStride), + RhsMapper(&rhs.coeffRef(size), rhsIncr), + _res, resIncr, alpha); + } + } + +template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs,int Version> +struct triangular_matrix_vector_product<Index,Mode,LhsScalar,ConjLhs,RhsScalar,ConjRhs,RowMajor,Version> +{ + typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar; + enum { + IsLower = ((Mode&Lower)==Lower), + HasUnitDiag = (Mode & UnitDiag)==UnitDiag, + HasZeroDiag = (Mode & ZeroDiag)==ZeroDiag + }; + static EIGEN_DONT_INLINE void run(Index _rows, Index _cols, const LhsScalar* _lhs, Index lhsStride, + const RhsScalar* _rhs, Index rhsIncr, ResScalar* _res, Index resIncr, const ResScalar& alpha); +}; + +template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs,int Version> +EIGEN_DONT_INLINE void triangular_matrix_vector_product<Index,Mode,LhsScalar,ConjLhs,RhsScalar,ConjRhs,RowMajor,Version> + ::run(Index _rows, Index _cols, const LhsScalar* _lhs, Index lhsStride, + const RhsScalar* _rhs, Index rhsIncr, ResScalar* _res, Index resIncr, const ResScalar& alpha) + { + static const Index PanelWidth = EIGEN_TUNE_TRIANGULAR_PANEL_WIDTH; + Index diagSize = (std::min)(_rows,_cols); + Index rows = IsLower ? _rows : diagSize; + Index cols = IsLower ? diagSize : _cols; + + typedef Map<const Matrix<LhsScalar,Dynamic,Dynamic,RowMajor>, 0, OuterStride<> > LhsMap; + const LhsMap lhs(_lhs,rows,cols,OuterStride<>(lhsStride)); + typename conj_expr_if<ConjLhs,LhsMap>::type cjLhs(lhs); + + typedef Map<const Matrix<RhsScalar,Dynamic,1> > RhsMap; + const RhsMap rhs(_rhs,cols); + typename conj_expr_if<ConjRhs,RhsMap>::type cjRhs(rhs); + + typedef Map<Matrix<ResScalar,Dynamic,1>, 0, InnerStride<> > ResMap; + ResMap res(_res,rows,InnerStride<>(resIncr)); + + typedef const_blas_data_mapper<LhsScalar,Index,RowMajor> LhsMapper; + typedef const_blas_data_mapper<RhsScalar,Index,RowMajor> RhsMapper; + + for (Index pi=0; pi<diagSize; pi+=PanelWidth) + { + Index actualPanelWidth = (std::min)(PanelWidth, diagSize-pi); + for (Index k=0; k<actualPanelWidth; ++k) + { + Index i = pi + k; + Index s = IsLower ? pi : ((HasUnitDiag||HasZeroDiag) ? i+1 : i); + Index r = IsLower ? k+1 : actualPanelWidth-k; + if ((!(HasUnitDiag||HasZeroDiag)) || (--r)>0) + res.coeffRef(i) += alpha * (cjLhs.row(i).segment(s,r).cwiseProduct(cjRhs.segment(s,r).transpose())).sum(); + if (HasUnitDiag) + res.coeffRef(i) += alpha * cjRhs.coeff(i); + } + Index r = IsLower ? pi : cols - pi - actualPanelWidth; + if (r>0) + { + Index s = IsLower ? 0 : pi + actualPanelWidth; + general_matrix_vector_product<Index,LhsScalar,LhsMapper,RowMajor,ConjLhs,RhsScalar,RhsMapper,ConjRhs,BuiltIn>::run( + actualPanelWidth, r, + LhsMapper(&lhs.coeffRef(pi,s), lhsStride), + RhsMapper(&rhs.coeffRef(s), rhsIncr), + &res.coeffRef(pi), resIncr, alpha); + } + } + if(IsLower && rows>diagSize) + { + general_matrix_vector_product<Index,LhsScalar,LhsMapper,RowMajor,ConjLhs,RhsScalar,RhsMapper,ConjRhs>::run( + rows-diagSize, cols, + LhsMapper(&lhs.coeffRef(diagSize,0), lhsStride), + RhsMapper(&rhs.coeffRef(0), rhsIncr), + &res.coeffRef(diagSize), resIncr, alpha); + } + } + +/*************************************************************************** +* Wrapper to product_triangular_vector +***************************************************************************/ + +template<int Mode,int StorageOrder> +struct trmv_selector; + +} // end namespace internal + +namespace internal { + +template<int Mode, typename Lhs, typename Rhs> +struct triangular_product_impl<Mode,true,Lhs,false,Rhs,true> +{ + template<typename Dest> static void run(Dest& dst, const Lhs &lhs, const Rhs &rhs, const typename Dest::Scalar& alpha) + { + eigen_assert(dst.rows()==lhs.rows() && dst.cols()==rhs.cols()); + + internal::trmv_selector<Mode,(int(internal::traits<Lhs>::Flags)&RowMajorBit) ? RowMajor : ColMajor>::run(lhs, rhs, dst, alpha); + } +}; + +template<int Mode, typename Lhs, typename Rhs> +struct triangular_product_impl<Mode,false,Lhs,true,Rhs,false> +{ + template<typename Dest> static void run(Dest& dst, const Lhs &lhs, const Rhs &rhs, const typename Dest::Scalar& alpha) + { + eigen_assert(dst.rows()==lhs.rows() && dst.cols()==rhs.cols()); + + Transpose<Dest> dstT(dst); + internal::trmv_selector<(Mode & (UnitDiag|ZeroDiag)) | ((Mode & Lower) ? Upper : Lower), + (int(internal::traits<Rhs>::Flags)&RowMajorBit) ? ColMajor : RowMajor> + ::run(rhs.transpose(),lhs.transpose(), dstT, alpha); + } +}; + +} // end namespace internal + +namespace internal { + +// TODO: find a way to factorize this piece of code with gemv_selector since the logic is exactly the same. + +template<int Mode> struct trmv_selector<Mode,ColMajor> +{ + template<typename Lhs, typename Rhs, typename Dest> + static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha) + { + typedef typename Lhs::Scalar LhsScalar; + typedef typename Rhs::Scalar RhsScalar; + typedef typename Dest::Scalar ResScalar; + typedef typename Dest::RealScalar RealScalar; + + typedef internal::blas_traits<Lhs> LhsBlasTraits; + typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType; + typedef internal::blas_traits<Rhs> RhsBlasTraits; + typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType; + + typedef Map<Matrix<ResScalar,Dynamic,1>, EIGEN_PLAIN_ENUM_MIN(AlignedMax,internal::packet_traits<ResScalar>::size)> MappedDest; + + typename internal::add_const_on_value_type<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(lhs); + typename internal::add_const_on_value_type<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(rhs); + + ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(lhs) + * RhsBlasTraits::extractScalarFactor(rhs); + + enum { + // FIXME find a way to allow an inner stride on the result if packet_traits<Scalar>::size==1 + // on, the other hand it is good for the cache to pack the vector anyways... + EvalToDestAtCompileTime = Dest::InnerStrideAtCompileTime==1, + ComplexByReal = (NumTraits<LhsScalar>::IsComplex) && (!NumTraits<RhsScalar>::IsComplex), + MightCannotUseDest = (Dest::InnerStrideAtCompileTime!=1) || ComplexByReal + }; + + gemv_static_vector_if<ResScalar,Dest::SizeAtCompileTime,Dest::MaxSizeAtCompileTime,MightCannotUseDest> static_dest; + + bool alphaIsCompatible = (!ComplexByReal) || (numext::imag(actualAlpha)==RealScalar(0)); + bool evalToDest = EvalToDestAtCompileTime && alphaIsCompatible; + + RhsScalar compatibleAlpha = get_factor<ResScalar,RhsScalar>::run(actualAlpha); + + ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(), + evalToDest ? dest.data() : static_dest.data()); + + if(!evalToDest) + { + #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN + Index size = dest.size(); + EIGEN_DENSE_STORAGE_CTOR_PLUGIN + #endif + if(!alphaIsCompatible) + { + MappedDest(actualDestPtr, dest.size()).setZero(); + compatibleAlpha = RhsScalar(1); + } + else + MappedDest(actualDestPtr, dest.size()) = dest; + } + + internal::triangular_matrix_vector_product + <Index,Mode, + LhsScalar, LhsBlasTraits::NeedToConjugate, + RhsScalar, RhsBlasTraits::NeedToConjugate, + ColMajor> + ::run(actualLhs.rows(),actualLhs.cols(), + actualLhs.data(),actualLhs.outerStride(), + actualRhs.data(),actualRhs.innerStride(), + actualDestPtr,1,compatibleAlpha); + + if (!evalToDest) + { + if(!alphaIsCompatible) + dest += actualAlpha * MappedDest(actualDestPtr, dest.size()); + else + dest = MappedDest(actualDestPtr, dest.size()); + } + } +}; + +template<int Mode> struct trmv_selector<Mode,RowMajor> +{ + template<typename Lhs, typename Rhs, typename Dest> + static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha) + { + typedef typename Lhs::Scalar LhsScalar; + typedef typename Rhs::Scalar RhsScalar; + typedef typename Dest::Scalar ResScalar; + + typedef internal::blas_traits<Lhs> LhsBlasTraits; + typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType; + typedef internal::blas_traits<Rhs> RhsBlasTraits; + typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType; + typedef typename internal::remove_all<ActualRhsType>::type ActualRhsTypeCleaned; + + typename add_const<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(lhs); + typename add_const<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(rhs); + + ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(lhs) + * RhsBlasTraits::extractScalarFactor(rhs); + + enum { + DirectlyUseRhs = ActualRhsTypeCleaned::InnerStrideAtCompileTime==1 + }; + + gemv_static_vector_if<RhsScalar,ActualRhsTypeCleaned::SizeAtCompileTime,ActualRhsTypeCleaned::MaxSizeAtCompileTime,!DirectlyUseRhs> static_rhs; + + ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhsPtr,actualRhs.size(), + DirectlyUseRhs ? const_cast<RhsScalar*>(actualRhs.data()) : static_rhs.data()); + + if(!DirectlyUseRhs) + { + #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN + Index size = actualRhs.size(); + EIGEN_DENSE_STORAGE_CTOR_PLUGIN + #endif + Map<typename ActualRhsTypeCleaned::PlainObject>(actualRhsPtr, actualRhs.size()) = actualRhs; + } + + internal::triangular_matrix_vector_product + <Index,Mode, + LhsScalar, LhsBlasTraits::NeedToConjugate, + RhsScalar, RhsBlasTraits::NeedToConjugate, + RowMajor> + ::run(actualLhs.rows(),actualLhs.cols(), + actualLhs.data(),actualLhs.outerStride(), + actualRhsPtr,1, + dest.data(),dest.innerStride(), + actualAlpha); + } +}; + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_TRIANGULARMATRIXVECTOR_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/TriangularMatrixVector_BLAS.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/TriangularMatrixVector_BLAS.h new file mode 100644 index 000000000..07bf26ce5 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/TriangularMatrixVector_BLAS.h @@ -0,0 +1,241 @@ +/* + Copyright (c) 2011, Intel Corporation. All rights reserved. + + Redistribution and use in source and binary forms, with or without modification, + are permitted provided that the following conditions are met: + + * Redistributions of source code must retain the above copyright notice, this + list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above copyright notice, + this list of conditions and the following disclaimer in the documentation + and/or other materials provided with the distribution. + * Neither the name of Intel Corporation nor the names of its contributors may + be used to endorse or promote products derived from this software without + specific prior written permission. + + THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND + ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED + WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE + DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR + ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES + (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; + LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON + ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS + SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + + ******************************************************************************** + * Content : Eigen bindings to BLAS F77 + * Triangular matrix-vector product functionality based on ?TRMV. + ******************************************************************************** +*/ + +#ifndef EIGEN_TRIANGULAR_MATRIX_VECTOR_BLAS_H +#define EIGEN_TRIANGULAR_MATRIX_VECTOR_BLAS_H + +namespace Eigen { + +namespace internal { + +/********************************************************************** +* This file implements triangular matrix-vector multiplication using BLAS +**********************************************************************/ + +// trmv/hemv specialization + +template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs, int StorageOrder> +struct triangular_matrix_vector_product_trmv : + triangular_matrix_vector_product<Index,Mode,LhsScalar,ConjLhs,RhsScalar,ConjRhs,StorageOrder,BuiltIn> {}; + +#define EIGEN_BLAS_TRMV_SPECIALIZE(Scalar) \ +template<typename Index, int Mode, bool ConjLhs, bool ConjRhs> \ +struct triangular_matrix_vector_product<Index,Mode,Scalar,ConjLhs,Scalar,ConjRhs,ColMajor,Specialized> { \ + static void run(Index _rows, Index _cols, const Scalar* _lhs, Index lhsStride, \ + const Scalar* _rhs, Index rhsIncr, Scalar* _res, Index resIncr, Scalar alpha) { \ + triangular_matrix_vector_product_trmv<Index,Mode,Scalar,ConjLhs,Scalar,ConjRhs,ColMajor>::run( \ + _rows, _cols, _lhs, lhsStride, _rhs, rhsIncr, _res, resIncr, alpha); \ + } \ +}; \ +template<typename Index, int Mode, bool ConjLhs, bool ConjRhs> \ +struct triangular_matrix_vector_product<Index,Mode,Scalar,ConjLhs,Scalar,ConjRhs,RowMajor,Specialized> { \ + static void run(Index _rows, Index _cols, const Scalar* _lhs, Index lhsStride, \ + const Scalar* _rhs, Index rhsIncr, Scalar* _res, Index resIncr, Scalar alpha) { \ + triangular_matrix_vector_product_trmv<Index,Mode,Scalar,ConjLhs,Scalar,ConjRhs,RowMajor>::run( \ + _rows, _cols, _lhs, lhsStride, _rhs, rhsIncr, _res, resIncr, alpha); \ + } \ +}; + +EIGEN_BLAS_TRMV_SPECIALIZE(double) +EIGEN_BLAS_TRMV_SPECIALIZE(float) +EIGEN_BLAS_TRMV_SPECIALIZE(dcomplex) +EIGEN_BLAS_TRMV_SPECIALIZE(scomplex) + +// implements col-major: res += alpha * op(triangular) * vector +#define EIGEN_BLAS_TRMV_CM(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \ +template<typename Index, int Mode, bool ConjLhs, bool ConjRhs> \ +struct triangular_matrix_vector_product_trmv<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,ConjRhs,ColMajor> { \ + enum { \ + IsLower = (Mode&Lower) == Lower, \ + SetDiag = (Mode&(ZeroDiag|UnitDiag)) ? 0 : 1, \ + IsUnitDiag = (Mode&UnitDiag) ? 1 : 0, \ + IsZeroDiag = (Mode&ZeroDiag) ? 1 : 0, \ + LowUp = IsLower ? Lower : Upper \ + }; \ + static void run(Index _rows, Index _cols, const EIGTYPE* _lhs, Index lhsStride, \ + const EIGTYPE* _rhs, Index rhsIncr, EIGTYPE* _res, Index resIncr, EIGTYPE alpha) \ + { \ + if (ConjLhs || IsZeroDiag) { \ + triangular_matrix_vector_product<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,ConjRhs,ColMajor,BuiltIn>::run( \ + _rows, _cols, _lhs, lhsStride, _rhs, rhsIncr, _res, resIncr, alpha); \ + return; \ + }\ + Index size = (std::min)(_rows,_cols); \ + Index rows = IsLower ? _rows : size; \ + Index cols = IsLower ? size : _cols; \ +\ + typedef VectorX##EIGPREFIX VectorRhs; \ + EIGTYPE *x, *y;\ +\ +/* Set x*/ \ + Map<const VectorRhs, 0, InnerStride<> > rhs(_rhs,cols,InnerStride<>(rhsIncr)); \ + VectorRhs x_tmp; \ + if (ConjRhs) x_tmp = rhs.conjugate(); else x_tmp = rhs; \ + x = x_tmp.data(); \ +\ +/* Square part handling */\ +\ + char trans, uplo, diag; \ + BlasIndex m, n, lda, incx, incy; \ + EIGTYPE const *a; \ + EIGTYPE beta(1); \ +\ +/* Set m, n */ \ + n = convert_index<BlasIndex>(size); \ + lda = convert_index<BlasIndex>(lhsStride); \ + incx = 1; \ + incy = convert_index<BlasIndex>(resIncr); \ +\ +/* Set uplo, trans and diag*/ \ + trans = 'N'; \ + uplo = IsLower ? 'L' : 'U'; \ + diag = IsUnitDiag ? 'U' : 'N'; \ +\ +/* call ?TRMV*/ \ + BLASPREFIX##trmv_(&uplo, &trans, &diag, &n, (const BLASTYPE*)_lhs, &lda, (BLASTYPE*)x, &incx); \ +\ +/* Add op(a_tr)rhs into res*/ \ + BLASPREFIX##axpy_(&n, &numext::real_ref(alpha),(const BLASTYPE*)x, &incx, (BLASTYPE*)_res, &incy); \ +/* Non-square case - doesn't fit to BLAS ?TRMV. Fall to default triangular product*/ \ + if (size<(std::max)(rows,cols)) { \ + if (ConjRhs) x_tmp = rhs.conjugate(); else x_tmp = rhs; \ + x = x_tmp.data(); \ + if (size<rows) { \ + y = _res + size*resIncr; \ + a = _lhs + size; \ + m = convert_index<BlasIndex>(rows-size); \ + n = convert_index<BlasIndex>(size); \ + } \ + else { \ + x += size; \ + y = _res; \ + a = _lhs + size*lda; \ + m = convert_index<BlasIndex>(size); \ + n = convert_index<BlasIndex>(cols-size); \ + } \ + BLASPREFIX##gemv_(&trans, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)x, &incx, &numext::real_ref(beta), (BLASTYPE*)y, &incy); \ + } \ + } \ +}; + +EIGEN_BLAS_TRMV_CM(double, double, d, d) +EIGEN_BLAS_TRMV_CM(dcomplex, double, cd, z) +EIGEN_BLAS_TRMV_CM(float, float, f, s) +EIGEN_BLAS_TRMV_CM(scomplex, float, cf, c) + +// implements row-major: res += alpha * op(triangular) * vector +#define EIGEN_BLAS_TRMV_RM(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \ +template<typename Index, int Mode, bool ConjLhs, bool ConjRhs> \ +struct triangular_matrix_vector_product_trmv<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,ConjRhs,RowMajor> { \ + enum { \ + IsLower = (Mode&Lower) == Lower, \ + SetDiag = (Mode&(ZeroDiag|UnitDiag)) ? 0 : 1, \ + IsUnitDiag = (Mode&UnitDiag) ? 1 : 0, \ + IsZeroDiag = (Mode&ZeroDiag) ? 1 : 0, \ + LowUp = IsLower ? Lower : Upper \ + }; \ + static void run(Index _rows, Index _cols, const EIGTYPE* _lhs, Index lhsStride, \ + const EIGTYPE* _rhs, Index rhsIncr, EIGTYPE* _res, Index resIncr, EIGTYPE alpha) \ + { \ + if (IsZeroDiag) { \ + triangular_matrix_vector_product<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,ConjRhs,RowMajor,BuiltIn>::run( \ + _rows, _cols, _lhs, lhsStride, _rhs, rhsIncr, _res, resIncr, alpha); \ + return; \ + }\ + Index size = (std::min)(_rows,_cols); \ + Index rows = IsLower ? _rows : size; \ + Index cols = IsLower ? size : _cols; \ +\ + typedef VectorX##EIGPREFIX VectorRhs; \ + EIGTYPE *x, *y;\ +\ +/* Set x*/ \ + Map<const VectorRhs, 0, InnerStride<> > rhs(_rhs,cols,InnerStride<>(rhsIncr)); \ + VectorRhs x_tmp; \ + if (ConjRhs) x_tmp = rhs.conjugate(); else x_tmp = rhs; \ + x = x_tmp.data(); \ +\ +/* Square part handling */\ +\ + char trans, uplo, diag; \ + BlasIndex m, n, lda, incx, incy; \ + EIGTYPE const *a; \ + EIGTYPE beta(1); \ +\ +/* Set m, n */ \ + n = convert_index<BlasIndex>(size); \ + lda = convert_index<BlasIndex>(lhsStride); \ + incx = 1; \ + incy = convert_index<BlasIndex>(resIncr); \ +\ +/* Set uplo, trans and diag*/ \ + trans = ConjLhs ? 'C' : 'T'; \ + uplo = IsLower ? 'U' : 'L'; \ + diag = IsUnitDiag ? 'U' : 'N'; \ +\ +/* call ?TRMV*/ \ + BLASPREFIX##trmv_(&uplo, &trans, &diag, &n, (const BLASTYPE*)_lhs, &lda, (BLASTYPE*)x, &incx); \ +\ +/* Add op(a_tr)rhs into res*/ \ + BLASPREFIX##axpy_(&n, &numext::real_ref(alpha),(const BLASTYPE*)x, &incx, (BLASTYPE*)_res, &incy); \ +/* Non-square case - doesn't fit to BLAS ?TRMV. Fall to default triangular product*/ \ + if (size<(std::max)(rows,cols)) { \ + if (ConjRhs) x_tmp = rhs.conjugate(); else x_tmp = rhs; \ + x = x_tmp.data(); \ + if (size<rows) { \ + y = _res + size*resIncr; \ + a = _lhs + size*lda; \ + m = convert_index<BlasIndex>(rows-size); \ + n = convert_index<BlasIndex>(size); \ + } \ + else { \ + x += size; \ + y = _res; \ + a = _lhs + size; \ + m = convert_index<BlasIndex>(size); \ + n = convert_index<BlasIndex>(cols-size); \ + } \ + BLASPREFIX##gemv_(&trans, &n, &m, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)x, &incx, &numext::real_ref(beta), (BLASTYPE*)y, &incy); \ + } \ + } \ +}; + +EIGEN_BLAS_TRMV_RM(double, double, d, d) +EIGEN_BLAS_TRMV_RM(dcomplex, double, cd, z) +EIGEN_BLAS_TRMV_RM(float, float, f, s) +EIGEN_BLAS_TRMV_RM(scomplex, float, cf, c) + +} // end namespase internal + +} // end namespace Eigen + +#endif // EIGEN_TRIANGULAR_MATRIX_VECTOR_BLAS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/TriangularSolverMatrix.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/TriangularSolverMatrix.h new file mode 100644 index 000000000..223c38b86 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/TriangularSolverMatrix.h @@ -0,0 +1,335 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2009 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_TRIANGULAR_SOLVER_MATRIX_H +#define EIGEN_TRIANGULAR_SOLVER_MATRIX_H + +namespace Eigen { + +namespace internal { + +// if the rhs is row major, let's transpose the product +template <typename Scalar, typename Index, int Side, int Mode, bool Conjugate, int TriStorageOrder> +struct triangular_solve_matrix<Scalar,Index,Side,Mode,Conjugate,TriStorageOrder,RowMajor> +{ + static void run( + Index size, Index cols, + const Scalar* tri, Index triStride, + Scalar* _other, Index otherStride, + level3_blocking<Scalar,Scalar>& blocking) + { + triangular_solve_matrix< + Scalar, Index, Side==OnTheLeft?OnTheRight:OnTheLeft, + (Mode&UnitDiag) | ((Mode&Upper) ? Lower : Upper), + NumTraits<Scalar>::IsComplex && Conjugate, + TriStorageOrder==RowMajor ? ColMajor : RowMajor, ColMajor> + ::run(size, cols, tri, triStride, _other, otherStride, blocking); + } +}; + +/* Optimized triangular solver with multiple right hand side and the triangular matrix on the left + */ +template <typename Scalar, typename Index, int Mode, bool Conjugate, int TriStorageOrder> +struct triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conjugate,TriStorageOrder,ColMajor> +{ + static EIGEN_DONT_INLINE void run( + Index size, Index otherSize, + const Scalar* _tri, Index triStride, + Scalar* _other, Index otherStride, + level3_blocking<Scalar,Scalar>& blocking); +}; +template <typename Scalar, typename Index, int Mode, bool Conjugate, int TriStorageOrder> +EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conjugate,TriStorageOrder,ColMajor>::run( + Index size, Index otherSize, + const Scalar* _tri, Index triStride, + Scalar* _other, Index otherStride, + level3_blocking<Scalar,Scalar>& blocking) + { + Index cols = otherSize; + + typedef const_blas_data_mapper<Scalar, Index, TriStorageOrder> TriMapper; + typedef blas_data_mapper<Scalar, Index, ColMajor> OtherMapper; + TriMapper tri(_tri, triStride); + OtherMapper other(_other, otherStride); + + typedef gebp_traits<Scalar,Scalar> Traits; + + enum { + SmallPanelWidth = EIGEN_PLAIN_ENUM_MAX(Traits::mr,Traits::nr), + IsLower = (Mode&Lower) == Lower + }; + + Index kc = blocking.kc(); // cache block size along the K direction + Index mc = (std::min)(size,blocking.mc()); // cache block size along the M direction + + std::size_t sizeA = kc*mc; + std::size_t sizeB = kc*cols; + + ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA()); + ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB()); + + conj_if<Conjugate> conj; + gebp_kernel<Scalar, Scalar, Index, OtherMapper, Traits::mr, Traits::nr, Conjugate, false> gebp_kernel; + gemm_pack_lhs<Scalar, Index, TriMapper, Traits::mr, Traits::LhsProgress, TriStorageOrder> pack_lhs; + gemm_pack_rhs<Scalar, Index, OtherMapper, Traits::nr, ColMajor, false, true> pack_rhs; + + // the goal here is to subdivise the Rhs panels such that we keep some cache + // coherence when accessing the rhs elements + std::ptrdiff_t l1, l2, l3; + manage_caching_sizes(GetAction, &l1, &l2, &l3); + Index subcols = cols>0 ? l2/(4 * sizeof(Scalar) * std::max<Index>(otherStride,size)) : 0; + subcols = std::max<Index>((subcols/Traits::nr)*Traits::nr, Traits::nr); + + for(Index k2=IsLower ? 0 : size; + IsLower ? k2<size : k2>0; + IsLower ? k2+=kc : k2-=kc) + { + const Index actual_kc = (std::min)(IsLower ? size-k2 : k2, kc); + + // We have selected and packed a big horizontal panel R1 of rhs. Let B be the packed copy of this panel, + // and R2 the remaining part of rhs. The corresponding vertical panel of lhs is split into + // A11 (the triangular part) and A21 the remaining rectangular part. + // Then the high level algorithm is: + // - B = R1 => general block copy (done during the next step) + // - R1 = A11^-1 B => tricky part + // - update B from the new R1 => actually this has to be performed continuously during the above step + // - R2 -= A21 * B => GEPP + + // The tricky part: compute R1 = A11^-1 B while updating B from R1 + // The idea is to split A11 into multiple small vertical panels. + // Each panel can be split into a small triangular part T1k which is processed without optimization, + // and the remaining small part T2k which is processed using gebp with appropriate block strides + for(Index j2=0; j2<cols; j2+=subcols) + { + Index actual_cols = (std::min)(cols-j2,subcols); + // for each small vertical panels [T1k^T, T2k^T]^T of lhs + for (Index k1=0; k1<actual_kc; k1+=SmallPanelWidth) + { + Index actualPanelWidth = std::min<Index>(actual_kc-k1, SmallPanelWidth); + // tr solve + for (Index k=0; k<actualPanelWidth; ++k) + { + // TODO write a small kernel handling this (can be shared with trsv) + Index i = IsLower ? k2+k1+k : k2-k1-k-1; + Index rs = actualPanelWidth - k - 1; // remaining size + Index s = TriStorageOrder==RowMajor ? (IsLower ? k2+k1 : i+1) + : IsLower ? i+1 : i-rs; + + Scalar a = (Mode & UnitDiag) ? Scalar(1) : Scalar(1)/conj(tri(i,i)); + for (Index j=j2; j<j2+actual_cols; ++j) + { + if (TriStorageOrder==RowMajor) + { + Scalar b(0); + const Scalar* l = &tri(i,s); + Scalar* r = &other(s,j); + for (Index i3=0; i3<k; ++i3) + b += conj(l[i3]) * r[i3]; + + other(i,j) = (other(i,j) - b)*a; + } + else + { + Scalar b = (other(i,j) *= a); + Scalar* r = &other(s,j); + const Scalar* l = &tri(s,i); + for (Index i3=0;i3<rs;++i3) + r[i3] -= b * conj(l[i3]); + } + } + } + + Index lengthTarget = actual_kc-k1-actualPanelWidth; + Index startBlock = IsLower ? k2+k1 : k2-k1-actualPanelWidth; + Index blockBOffset = IsLower ? k1 : lengthTarget; + + // update the respective rows of B from other + pack_rhs(blockB+actual_kc*j2, other.getSubMapper(startBlock,j2), actualPanelWidth, actual_cols, actual_kc, blockBOffset); + + // GEBP + if (lengthTarget>0) + { + Index startTarget = IsLower ? k2+k1+actualPanelWidth : k2-actual_kc; + + pack_lhs(blockA, tri.getSubMapper(startTarget,startBlock), actualPanelWidth, lengthTarget); + + gebp_kernel(other.getSubMapper(startTarget,j2), blockA, blockB+actual_kc*j2, lengthTarget, actualPanelWidth, actual_cols, Scalar(-1), + actualPanelWidth, actual_kc, 0, blockBOffset); + } + } + } + + // R2 -= A21 * B => GEPP + { + Index start = IsLower ? k2+kc : 0; + Index end = IsLower ? size : k2-kc; + for(Index i2=start; i2<end; i2+=mc) + { + const Index actual_mc = (std::min)(mc,end-i2); + if (actual_mc>0) + { + pack_lhs(blockA, tri.getSubMapper(i2, IsLower ? k2 : k2-kc), actual_kc, actual_mc); + + gebp_kernel(other.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc, cols, Scalar(-1), -1, -1, 0, 0); + } + } + } + } + } + +/* Optimized triangular solver with multiple left hand sides and the triangular matrix on the right + */ +template <typename Scalar, typename Index, int Mode, bool Conjugate, int TriStorageOrder> +struct triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conjugate,TriStorageOrder,ColMajor> +{ + static EIGEN_DONT_INLINE void run( + Index size, Index otherSize, + const Scalar* _tri, Index triStride, + Scalar* _other, Index otherStride, + level3_blocking<Scalar,Scalar>& blocking); +}; +template <typename Scalar, typename Index, int Mode, bool Conjugate, int TriStorageOrder> +EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conjugate,TriStorageOrder,ColMajor>::run( + Index size, Index otherSize, + const Scalar* _tri, Index triStride, + Scalar* _other, Index otherStride, + level3_blocking<Scalar,Scalar>& blocking) + { + Index rows = otherSize; + typedef typename NumTraits<Scalar>::Real RealScalar; + + typedef blas_data_mapper<Scalar, Index, ColMajor> LhsMapper; + typedef const_blas_data_mapper<Scalar, Index, TriStorageOrder> RhsMapper; + LhsMapper lhs(_other, otherStride); + RhsMapper rhs(_tri, triStride); + + typedef gebp_traits<Scalar,Scalar> Traits; + enum { + RhsStorageOrder = TriStorageOrder, + SmallPanelWidth = EIGEN_PLAIN_ENUM_MAX(Traits::mr,Traits::nr), + IsLower = (Mode&Lower) == Lower + }; + + Index kc = blocking.kc(); // cache block size along the K direction + Index mc = (std::min)(rows,blocking.mc()); // cache block size along the M direction + + std::size_t sizeA = kc*mc; + std::size_t sizeB = kc*size; + + ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA()); + ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB()); + + conj_if<Conjugate> conj; + gebp_kernel<Scalar, Scalar, Index, LhsMapper, Traits::mr, Traits::nr, false, Conjugate> gebp_kernel; + gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs; + gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr, RhsStorageOrder,false,true> pack_rhs_panel; + gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, ColMajor, false, true> pack_lhs_panel; + + for(Index k2=IsLower ? size : 0; + IsLower ? k2>0 : k2<size; + IsLower ? k2-=kc : k2+=kc) + { + const Index actual_kc = (std::min)(IsLower ? k2 : size-k2, kc); + Index actual_k2 = IsLower ? k2-actual_kc : k2 ; + + Index startPanel = IsLower ? 0 : k2+actual_kc; + Index rs = IsLower ? actual_k2 : size - actual_k2 - actual_kc; + Scalar* geb = blockB+actual_kc*actual_kc; + + if (rs>0) pack_rhs(geb, rhs.getSubMapper(actual_k2,startPanel), actual_kc, rs); + + // triangular packing (we only pack the panels off the diagonal, + // neglecting the blocks overlapping the diagonal + { + for (Index j2=0; j2<actual_kc; j2+=SmallPanelWidth) + { + Index actualPanelWidth = std::min<Index>(actual_kc-j2, SmallPanelWidth); + Index actual_j2 = actual_k2 + j2; + Index panelOffset = IsLower ? j2+actualPanelWidth : 0; + Index panelLength = IsLower ? actual_kc-j2-actualPanelWidth : j2; + + if (panelLength>0) + pack_rhs_panel(blockB+j2*actual_kc, + rhs.getSubMapper(actual_k2+panelOffset, actual_j2), + panelLength, actualPanelWidth, + actual_kc, panelOffset); + } + } + + for(Index i2=0; i2<rows; i2+=mc) + { + const Index actual_mc = (std::min)(mc,rows-i2); + + // triangular solver kernel + { + // for each small block of the diagonal (=> vertical panels of rhs) + for (Index j2 = IsLower + ? (actual_kc - ((actual_kc%SmallPanelWidth) ? Index(actual_kc%SmallPanelWidth) + : Index(SmallPanelWidth))) + : 0; + IsLower ? j2>=0 : j2<actual_kc; + IsLower ? j2-=SmallPanelWidth : j2+=SmallPanelWidth) + { + Index actualPanelWidth = std::min<Index>(actual_kc-j2, SmallPanelWidth); + Index absolute_j2 = actual_k2 + j2; + Index panelOffset = IsLower ? j2+actualPanelWidth : 0; + Index panelLength = IsLower ? actual_kc - j2 - actualPanelWidth : j2; + + // GEBP + if(panelLength>0) + { + gebp_kernel(lhs.getSubMapper(i2,absolute_j2), + blockA, blockB+j2*actual_kc, + actual_mc, panelLength, actualPanelWidth, + Scalar(-1), + actual_kc, actual_kc, // strides + panelOffset, panelOffset); // offsets + } + + // unblocked triangular solve + for (Index k=0; k<actualPanelWidth; ++k) + { + Index j = IsLower ? absolute_j2+actualPanelWidth-k-1 : absolute_j2+k; + + Scalar* r = &lhs(i2,j); + for (Index k3=0; k3<k; ++k3) + { + Scalar b = conj(rhs(IsLower ? j+1+k3 : absolute_j2+k3,j)); + Scalar* a = &lhs(i2,IsLower ? j+1+k3 : absolute_j2+k3); + for (Index i=0; i<actual_mc; ++i) + r[i] -= a[i] * b; + } + if((Mode & UnitDiag)==0) + { + Scalar inv_rjj = RealScalar(1)/conj(rhs(j,j)); + for (Index i=0; i<actual_mc; ++i) + r[i] *= inv_rjj; + } + } + + // pack the just computed part of lhs to A + pack_lhs_panel(blockA, LhsMapper(_other+absolute_j2*otherStride+i2, otherStride), + actualPanelWidth, actual_mc, + actual_kc, j2); + } + } + + if (rs>0) + gebp_kernel(lhs.getSubMapper(i2, startPanel), blockA, geb, + actual_mc, actual_kc, rs, Scalar(-1), + -1, -1, 0, 0); + } + } + } + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_TRIANGULAR_SOLVER_MATRIX_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/TriangularSolverMatrix_BLAS.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/TriangularSolverMatrix_BLAS.h new file mode 100644 index 000000000..88c0fb794 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/TriangularSolverMatrix_BLAS.h @@ -0,0 +1,151 @@ +/* + Copyright (c) 2011, Intel Corporation. All rights reserved. + + Redistribution and use in source and binary forms, with or without modification, + are permitted provided that the following conditions are met: + + * Redistributions of source code must retain the above copyright notice, this + list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above copyright notice, + this list of conditions and the following disclaimer in the documentation + and/or other materials provided with the distribution. + * Neither the name of Intel Corporation nor the names of its contributors may + be used to endorse or promote products derived from this software without + specific prior written permission. + + THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND + ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED + WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE + DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR + ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES + (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; + LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON + ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS + SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + + ******************************************************************************** + * Content : Eigen bindings to BLAS F77 + * Triangular matrix * matrix product functionality based on ?TRMM. + ******************************************************************************** +*/ + +#ifndef EIGEN_TRIANGULAR_SOLVER_MATRIX_BLAS_H +#define EIGEN_TRIANGULAR_SOLVER_MATRIX_BLAS_H + +namespace Eigen { + +namespace internal { + +// implements LeftSide op(triangular)^-1 * general +#define EIGEN_BLAS_TRSM_L(EIGTYPE, BLASTYPE, BLASPREFIX) \ +template <typename Index, int Mode, bool Conjugate, int TriStorageOrder> \ +struct triangular_solve_matrix<EIGTYPE,Index,OnTheLeft,Mode,Conjugate,TriStorageOrder,ColMajor> \ +{ \ + enum { \ + IsLower = (Mode&Lower) == Lower, \ + IsUnitDiag = (Mode&UnitDiag) ? 1 : 0, \ + IsZeroDiag = (Mode&ZeroDiag) ? 1 : 0, \ + conjA = ((TriStorageOrder==ColMajor) && Conjugate) ? 1 : 0 \ + }; \ + static void run( \ + Index size, Index otherSize, \ + const EIGTYPE* _tri, Index triStride, \ + EIGTYPE* _other, Index otherStride, level3_blocking<EIGTYPE,EIGTYPE>& /*blocking*/) \ + { \ + BlasIndex m = convert_index<BlasIndex>(size), n = convert_index<BlasIndex>(otherSize), lda, ldb; \ + char side = 'L', uplo, diag='N', transa; \ + /* Set alpha_ */ \ + EIGTYPE alpha(1); \ + ldb = convert_index<BlasIndex>(otherStride);\ +\ + const EIGTYPE *a; \ +/* Set trans */ \ + transa = (TriStorageOrder==RowMajor) ? ((Conjugate) ? 'C' : 'T') : 'N'; \ +/* Set uplo */ \ + uplo = IsLower ? 'L' : 'U'; \ + if (TriStorageOrder==RowMajor) uplo = (uplo == 'L') ? 'U' : 'L'; \ +/* Set a, lda */ \ + typedef Matrix<EIGTYPE, Dynamic, Dynamic, TriStorageOrder> MatrixTri; \ + Map<const MatrixTri, 0, OuterStride<> > tri(_tri,size,size,OuterStride<>(triStride)); \ + MatrixTri a_tmp; \ +\ + if (conjA) { \ + a_tmp = tri.conjugate(); \ + a = a_tmp.data(); \ + lda = convert_index<BlasIndex>(a_tmp.outerStride()); \ + } else { \ + a = _tri; \ + lda = convert_index<BlasIndex>(triStride); \ + } \ + if (IsUnitDiag) diag='U'; \ +/* call ?trsm*/ \ + BLASPREFIX##trsm_(&side, &uplo, &transa, &diag, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (BLASTYPE*)_other, &ldb); \ + } \ +}; + +EIGEN_BLAS_TRSM_L(double, double, d) +EIGEN_BLAS_TRSM_L(dcomplex, double, z) +EIGEN_BLAS_TRSM_L(float, float, s) +EIGEN_BLAS_TRSM_L(scomplex, float, c) + + +// implements RightSide general * op(triangular)^-1 +#define EIGEN_BLAS_TRSM_R(EIGTYPE, BLASTYPE, BLASPREFIX) \ +template <typename Index, int Mode, bool Conjugate, int TriStorageOrder> \ +struct triangular_solve_matrix<EIGTYPE,Index,OnTheRight,Mode,Conjugate,TriStorageOrder,ColMajor> \ +{ \ + enum { \ + IsLower = (Mode&Lower) == Lower, \ + IsUnitDiag = (Mode&UnitDiag) ? 1 : 0, \ + IsZeroDiag = (Mode&ZeroDiag) ? 1 : 0, \ + conjA = ((TriStorageOrder==ColMajor) && Conjugate) ? 1 : 0 \ + }; \ + static void run( \ + Index size, Index otherSize, \ + const EIGTYPE* _tri, Index triStride, \ + EIGTYPE* _other, Index otherStride, level3_blocking<EIGTYPE,EIGTYPE>& /*blocking*/) \ + { \ + BlasIndex m = convert_index<BlasIndex>(otherSize), n = convert_index<BlasIndex>(size), lda, ldb; \ + char side = 'R', uplo, diag='N', transa; \ + /* Set alpha_ */ \ + EIGTYPE alpha(1); \ + ldb = convert_index<BlasIndex>(otherStride);\ +\ + const EIGTYPE *a; \ +/* Set trans */ \ + transa = (TriStorageOrder==RowMajor) ? ((Conjugate) ? 'C' : 'T') : 'N'; \ +/* Set uplo */ \ + uplo = IsLower ? 'L' : 'U'; \ + if (TriStorageOrder==RowMajor) uplo = (uplo == 'L') ? 'U' : 'L'; \ +/* Set a, lda */ \ + typedef Matrix<EIGTYPE, Dynamic, Dynamic, TriStorageOrder> MatrixTri; \ + Map<const MatrixTri, 0, OuterStride<> > tri(_tri,size,size,OuterStride<>(triStride)); \ + MatrixTri a_tmp; \ +\ + if (conjA) { \ + a_tmp = tri.conjugate(); \ + a = a_tmp.data(); \ + lda = convert_index<BlasIndex>(a_tmp.outerStride()); \ + } else { \ + a = _tri; \ + lda = convert_index<BlasIndex>(triStride); \ + } \ + if (IsUnitDiag) diag='U'; \ +/* call ?trsm*/ \ + BLASPREFIX##trsm_(&side, &uplo, &transa, &diag, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (BLASTYPE*)_other, &ldb); \ + /*std::cout << "TRMS_L specialization!\n";*/ \ + } \ +}; + +EIGEN_BLAS_TRSM_R(double, double, d) +EIGEN_BLAS_TRSM_R(dcomplex, double, z) +EIGEN_BLAS_TRSM_R(float, float, s) +EIGEN_BLAS_TRSM_R(scomplex, float, c) + + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_TRIANGULAR_SOLVER_MATRIX_BLAS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/TriangularSolverVector.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/TriangularSolverVector.h new file mode 100644 index 000000000..b994759b2 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/products/TriangularSolverVector.h @@ -0,0 +1,145 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2010 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_TRIANGULAR_SOLVER_VECTOR_H +#define EIGEN_TRIANGULAR_SOLVER_VECTOR_H + +namespace Eigen { + +namespace internal { + +template<typename LhsScalar, typename RhsScalar, typename Index, int Mode, bool Conjugate, int StorageOrder> +struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheRight, Mode, Conjugate, StorageOrder> +{ + static void run(Index size, const LhsScalar* _lhs, Index lhsStride, RhsScalar* rhs) + { + triangular_solve_vector<LhsScalar,RhsScalar,Index,OnTheLeft, + ((Mode&Upper)==Upper ? Lower : Upper) | (Mode&UnitDiag), + Conjugate,StorageOrder==RowMajor?ColMajor:RowMajor + >::run(size, _lhs, lhsStride, rhs); + } +}; + +// forward and backward substitution, row-major, rhs is a vector +template<typename LhsScalar, typename RhsScalar, typename Index, int Mode, bool Conjugate> +struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheLeft, Mode, Conjugate, RowMajor> +{ + enum { + IsLower = ((Mode&Lower)==Lower) + }; + static void run(Index size, const LhsScalar* _lhs, Index lhsStride, RhsScalar* rhs) + { + typedef Map<const Matrix<LhsScalar,Dynamic,Dynamic,RowMajor>, 0, OuterStride<> > LhsMap; + const LhsMap lhs(_lhs,size,size,OuterStride<>(lhsStride)); + + typedef const_blas_data_mapper<LhsScalar,Index,RowMajor> LhsMapper; + typedef const_blas_data_mapper<RhsScalar,Index,ColMajor> RhsMapper; + + typename internal::conditional< + Conjugate, + const CwiseUnaryOp<typename internal::scalar_conjugate_op<LhsScalar>,LhsMap>, + const LhsMap&> + ::type cjLhs(lhs); + static const Index PanelWidth = EIGEN_TUNE_TRIANGULAR_PANEL_WIDTH; + for(Index pi=IsLower ? 0 : size; + IsLower ? pi<size : pi>0; + IsLower ? pi+=PanelWidth : pi-=PanelWidth) + { + Index actualPanelWidth = (std::min)(IsLower ? size - pi : pi, PanelWidth); + + Index r = IsLower ? pi : size - pi; // remaining size + if (r > 0) + { + // let's directly call the low level product function because: + // 1 - it is faster to compile + // 2 - it is slighlty faster at runtime + Index startRow = IsLower ? pi : pi-actualPanelWidth; + Index startCol = IsLower ? 0 : pi; + + general_matrix_vector_product<Index,LhsScalar,LhsMapper,RowMajor,Conjugate,RhsScalar,RhsMapper,false>::run( + actualPanelWidth, r, + LhsMapper(&lhs.coeffRef(startRow,startCol), lhsStride), + RhsMapper(rhs + startCol, 1), + rhs + startRow, 1, + RhsScalar(-1)); + } + + for(Index k=0; k<actualPanelWidth; ++k) + { + Index i = IsLower ? pi+k : pi-k-1; + Index s = IsLower ? pi : i+1; + if (k>0) + rhs[i] -= (cjLhs.row(i).segment(s,k).transpose().cwiseProduct(Map<const Matrix<RhsScalar,Dynamic,1> >(rhs+s,k))).sum(); + + if(!(Mode & UnitDiag)) + rhs[i] /= cjLhs(i,i); + } + } + } +}; + +// forward and backward substitution, column-major, rhs is a vector +template<typename LhsScalar, typename RhsScalar, typename Index, int Mode, bool Conjugate> +struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheLeft, Mode, Conjugate, ColMajor> +{ + enum { + IsLower = ((Mode&Lower)==Lower) + }; + static void run(Index size, const LhsScalar* _lhs, Index lhsStride, RhsScalar* rhs) + { + typedef Map<const Matrix<LhsScalar,Dynamic,Dynamic,ColMajor>, 0, OuterStride<> > LhsMap; + const LhsMap lhs(_lhs,size,size,OuterStride<>(lhsStride)); + typedef const_blas_data_mapper<LhsScalar,Index,ColMajor> LhsMapper; + typedef const_blas_data_mapper<RhsScalar,Index,ColMajor> RhsMapper; + typename internal::conditional<Conjugate, + const CwiseUnaryOp<typename internal::scalar_conjugate_op<LhsScalar>,LhsMap>, + const LhsMap& + >::type cjLhs(lhs); + static const Index PanelWidth = EIGEN_TUNE_TRIANGULAR_PANEL_WIDTH; + + for(Index pi=IsLower ? 0 : size; + IsLower ? pi<size : pi>0; + IsLower ? pi+=PanelWidth : pi-=PanelWidth) + { + Index actualPanelWidth = (std::min)(IsLower ? size - pi : pi, PanelWidth); + Index startBlock = IsLower ? pi : pi-actualPanelWidth; + Index endBlock = IsLower ? pi + actualPanelWidth : 0; + + for(Index k=0; k<actualPanelWidth; ++k) + { + Index i = IsLower ? pi+k : pi-k-1; + if(!(Mode & UnitDiag)) + rhs[i] /= cjLhs.coeff(i,i); + + Index r = actualPanelWidth - k - 1; // remaining size + Index s = IsLower ? i+1 : i-r; + if (r>0) + Map<Matrix<RhsScalar,Dynamic,1> >(rhs+s,r) -= rhs[i] * cjLhs.col(i).segment(s,r); + } + Index r = IsLower ? size - endBlock : startBlock; // remaining size + if (r > 0) + { + // let's directly call the low level product function because: + // 1 - it is faster to compile + // 2 - it is slighlty faster at runtime + general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,Conjugate,RhsScalar,RhsMapper,false>::run( + r, actualPanelWidth, + LhsMapper(&lhs.coeffRef(endBlock,startBlock), lhsStride), + RhsMapper(rhs+startBlock, 1), + rhs+endBlock, 1, RhsScalar(-1)); + } + } + } +}; + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_TRIANGULAR_SOLVER_VECTOR_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/BlasUtil.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/BlasUtil.h new file mode 100755 index 000000000..6e6ee119b --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/BlasUtil.h @@ -0,0 +1,398 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2009-2010 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_BLASUTIL_H +#define EIGEN_BLASUTIL_H + +// This file contains many lightweight helper classes used to +// implement and control fast level 2 and level 3 BLAS-like routines. + +namespace Eigen { + +namespace internal { + +// forward declarations +template<typename LhsScalar, typename RhsScalar, typename Index, typename DataMapper, int mr, int nr, bool ConjugateLhs=false, bool ConjugateRhs=false> +struct gebp_kernel; + +template<typename Scalar, typename Index, typename DataMapper, int nr, int StorageOrder, bool Conjugate = false, bool PanelMode=false> +struct gemm_pack_rhs; + +template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, int StorageOrder, bool Conjugate = false, bool PanelMode = false> +struct gemm_pack_lhs; + +template< + typename Index, + typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs, + typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs, + int ResStorageOrder> +struct general_matrix_matrix_product; + +template<typename Index, + typename LhsScalar, typename LhsMapper, int LhsStorageOrder, bool ConjugateLhs, + typename RhsScalar, typename RhsMapper, bool ConjugateRhs, int Version=Specialized> +struct general_matrix_vector_product; + + +template<bool Conjugate> struct conj_if; + +template<> struct conj_if<true> { + template<typename T> + inline T operator()(const T& x) const { return numext::conj(x); } + template<typename T> + inline T pconj(const T& x) const { return internal::pconj(x); } +}; + +template<> struct conj_if<false> { + template<typename T> + inline const T& operator()(const T& x) const { return x; } + template<typename T> + inline const T& pconj(const T& x) const { return x; } +}; + +// Generic implementation for custom complex types. +template<typename LhsScalar, typename RhsScalar, bool ConjLhs, bool ConjRhs> +struct conj_helper +{ + typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar>::ReturnType Scalar; + + EIGEN_STRONG_INLINE Scalar pmadd(const LhsScalar& x, const RhsScalar& y, const Scalar& c) const + { return padd(c, pmul(x,y)); } + + EIGEN_STRONG_INLINE Scalar pmul(const LhsScalar& x, const RhsScalar& y) const + { return conj_if<ConjLhs>()(x) * conj_if<ConjRhs>()(y); } +}; + +template<typename Scalar> struct conj_helper<Scalar,Scalar,false,false> +{ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar pmadd(const Scalar& x, const Scalar& y, const Scalar& c) const { return internal::pmadd(x,y,c); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar pmul(const Scalar& x, const Scalar& y) const { return internal::pmul(x,y); } +}; + +template<typename RealScalar> struct conj_helper<std::complex<RealScalar>, std::complex<RealScalar>, false,true> +{ + typedef std::complex<RealScalar> Scalar; + EIGEN_STRONG_INLINE Scalar pmadd(const Scalar& x, const Scalar& y, const Scalar& c) const + { return c + pmul(x,y); } + + EIGEN_STRONG_INLINE Scalar pmul(const Scalar& x, const Scalar& y) const + { return Scalar(numext::real(x)*numext::real(y) + numext::imag(x)*numext::imag(y), numext::imag(x)*numext::real(y) - numext::real(x)*numext::imag(y)); } +}; + +template<typename RealScalar> struct conj_helper<std::complex<RealScalar>, std::complex<RealScalar>, true,false> +{ + typedef std::complex<RealScalar> Scalar; + EIGEN_STRONG_INLINE Scalar pmadd(const Scalar& x, const Scalar& y, const Scalar& c) const + { return c + pmul(x,y); } + + EIGEN_STRONG_INLINE Scalar pmul(const Scalar& x, const Scalar& y) const + { return Scalar(numext::real(x)*numext::real(y) + numext::imag(x)*numext::imag(y), numext::real(x)*numext::imag(y) - numext::imag(x)*numext::real(y)); } +}; + +template<typename RealScalar> struct conj_helper<std::complex<RealScalar>, std::complex<RealScalar>, true,true> +{ + typedef std::complex<RealScalar> Scalar; + EIGEN_STRONG_INLINE Scalar pmadd(const Scalar& x, const Scalar& y, const Scalar& c) const + { return c + pmul(x,y); } + + EIGEN_STRONG_INLINE Scalar pmul(const Scalar& x, const Scalar& y) const + { return Scalar(numext::real(x)*numext::real(y) - numext::imag(x)*numext::imag(y), - numext::real(x)*numext::imag(y) - numext::imag(x)*numext::real(y)); } +}; + +template<typename RealScalar,bool Conj> struct conj_helper<std::complex<RealScalar>, RealScalar, Conj,false> +{ + typedef std::complex<RealScalar> Scalar; + EIGEN_STRONG_INLINE Scalar pmadd(const Scalar& x, const RealScalar& y, const Scalar& c) const + { return padd(c, pmul(x,y)); } + EIGEN_STRONG_INLINE Scalar pmul(const Scalar& x, const RealScalar& y) const + { return conj_if<Conj>()(x)*y; } +}; + +template<typename RealScalar,bool Conj> struct conj_helper<RealScalar, std::complex<RealScalar>, false,Conj> +{ + typedef std::complex<RealScalar> Scalar; + EIGEN_STRONG_INLINE Scalar pmadd(const RealScalar& x, const Scalar& y, const Scalar& c) const + { return padd(c, pmul(x,y)); } + EIGEN_STRONG_INLINE Scalar pmul(const RealScalar& x, const Scalar& y) const + { return x*conj_if<Conj>()(y); } +}; + +template<typename From,typename To> struct get_factor { + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE To run(const From& x) { return To(x); } +}; + +template<typename Scalar> struct get_factor<Scalar,typename NumTraits<Scalar>::Real> { + EIGEN_DEVICE_FUNC + static EIGEN_STRONG_INLINE typename NumTraits<Scalar>::Real run(const Scalar& x) { return numext::real(x); } +}; + + +template<typename Scalar, typename Index> +class BlasVectorMapper { + public: + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE BlasVectorMapper(Scalar *data) : m_data(data) {} + + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Scalar operator()(Index i) const { + return m_data[i]; + } + template <typename Packet, int AlignmentType> + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet load(Index i) const { + return ploadt<Packet, AlignmentType>(m_data + i); + } + + template <typename Packet> + EIGEN_DEVICE_FUNC bool aligned(Index i) const { + return (UIntPtr(m_data+i)%sizeof(Packet))==0; + } + + protected: + Scalar* m_data; +}; + +template<typename Scalar, typename Index, int AlignmentType> +class BlasLinearMapper { + public: + typedef typename packet_traits<Scalar>::type Packet; + typedef typename packet_traits<Scalar>::half HalfPacket; + + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE BlasLinearMapper(Scalar *data) : m_data(data) {} + + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void prefetch(int i) const { + internal::prefetch(&operator()(i)); + } + + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Scalar& operator()(Index i) const { + return m_data[i]; + } + + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet loadPacket(Index i) const { + return ploadt<Packet, AlignmentType>(m_data + i); + } + + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE HalfPacket loadHalfPacket(Index i) const { + return ploadt<HalfPacket, AlignmentType>(m_data + i); + } + + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void storePacket(Index i, const Packet &p) const { + pstoret<Scalar, Packet, AlignmentType>(m_data + i, p); + } + + protected: + Scalar *m_data; +}; + +// Lightweight helper class to access matrix coefficients. +template<typename Scalar, typename Index, int StorageOrder, int AlignmentType = Unaligned> +class blas_data_mapper { + public: + typedef typename packet_traits<Scalar>::type Packet; + typedef typename packet_traits<Scalar>::half HalfPacket; + + typedef BlasLinearMapper<Scalar, Index, AlignmentType> LinearMapper; + typedef BlasVectorMapper<Scalar, Index> VectorMapper; + + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE blas_data_mapper(Scalar* data, Index stride) : m_data(data), m_stride(stride) {} + + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE blas_data_mapper<Scalar, Index, StorageOrder, AlignmentType> + getSubMapper(Index i, Index j) const { + return blas_data_mapper<Scalar, Index, StorageOrder, AlignmentType>(&operator()(i, j), m_stride); + } + + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE LinearMapper getLinearMapper(Index i, Index j) const { + return LinearMapper(&operator()(i, j)); + } + + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE VectorMapper getVectorMapper(Index i, Index j) const { + return VectorMapper(&operator()(i, j)); + } + + + EIGEN_DEVICE_FUNC + EIGEN_ALWAYS_INLINE Scalar& operator()(Index i, Index j) const { + return m_data[StorageOrder==RowMajor ? j + i*m_stride : i + j*m_stride]; + } + + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet loadPacket(Index i, Index j) const { + return ploadt<Packet, AlignmentType>(&operator()(i, j)); + } + + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE HalfPacket loadHalfPacket(Index i, Index j) const { + return ploadt<HalfPacket, AlignmentType>(&operator()(i, j)); + } + + template<typename SubPacket> + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void scatterPacket(Index i, Index j, const SubPacket &p) const { + pscatter<Scalar, SubPacket>(&operator()(i, j), p, m_stride); + } + + template<typename SubPacket> + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE SubPacket gatherPacket(Index i, Index j) const { + return pgather<Scalar, SubPacket>(&operator()(i, j), m_stride); + } + + EIGEN_DEVICE_FUNC const Index stride() const { return m_stride; } + EIGEN_DEVICE_FUNC const Scalar* data() const { return m_data; } + + EIGEN_DEVICE_FUNC Index firstAligned(Index size) const { + if (UIntPtr(m_data)%sizeof(Scalar)) { + return -1; + } + return internal::first_default_aligned(m_data, size); + } + + protected: + Scalar* EIGEN_RESTRICT m_data; + const Index m_stride; +}; + +// lightweight helper class to access matrix coefficients (const version) +template<typename Scalar, typename Index, int StorageOrder> +class const_blas_data_mapper : public blas_data_mapper<const Scalar, Index, StorageOrder> { + public: + EIGEN_ALWAYS_INLINE const_blas_data_mapper(const Scalar *data, Index stride) : blas_data_mapper<const Scalar, Index, StorageOrder>(data, stride) {} + + EIGEN_ALWAYS_INLINE const_blas_data_mapper<Scalar, Index, StorageOrder> getSubMapper(Index i, Index j) const { + return const_blas_data_mapper<Scalar, Index, StorageOrder>(&(this->operator()(i, j)), this->m_stride); + } +}; + + +/* Helper class to analyze the factors of a Product expression. + * In particular it allows to pop out operator-, scalar multiples, + * and conjugate */ +template<typename XprType> struct blas_traits +{ + typedef typename traits<XprType>::Scalar Scalar; + typedef const XprType& ExtractType; + typedef XprType _ExtractType; + enum { + IsComplex = NumTraits<Scalar>::IsComplex, + IsTransposed = false, + NeedToConjugate = false, + HasUsableDirectAccess = ( (int(XprType::Flags)&DirectAccessBit) + && ( bool(XprType::IsVectorAtCompileTime) + || int(inner_stride_at_compile_time<XprType>::ret) == 1) + ) ? 1 : 0 + }; + typedef typename conditional<bool(HasUsableDirectAccess), + ExtractType, + typename _ExtractType::PlainObject + >::type DirectLinearAccessType; + static inline ExtractType extract(const XprType& x) { return x; } + static inline const Scalar extractScalarFactor(const XprType&) { return Scalar(1); } +}; + +// pop conjugate +template<typename Scalar, typename NestedXpr> +struct blas_traits<CwiseUnaryOp<scalar_conjugate_op<Scalar>, NestedXpr> > + : blas_traits<NestedXpr> +{ + typedef blas_traits<NestedXpr> Base; + typedef CwiseUnaryOp<scalar_conjugate_op<Scalar>, NestedXpr> XprType; + typedef typename Base::ExtractType ExtractType; + + enum { + IsComplex = NumTraits<Scalar>::IsComplex, + NeedToConjugate = Base::NeedToConjugate ? 0 : IsComplex + }; + static inline ExtractType extract(const XprType& x) { return Base::extract(x.nestedExpression()); } + static inline Scalar extractScalarFactor(const XprType& x) { return conj(Base::extractScalarFactor(x.nestedExpression())); } +}; + +// pop scalar multiple +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> +{ + 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; + static inline ExtractType extract(const XprType& x) { return Base::extract(x.rhs()); } + static inline Scalar extractScalarFactor(const XprType& x) + { return x.lhs().functor().m_other * Base::extractScalarFactor(x.rhs()); } +}; +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> +{ + 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; + static inline ExtractType extract(const XprType& x) { return Base::extract(x.lhs()); } + static inline Scalar extractScalarFactor(const XprType& x) + { return Base::extractScalarFactor(x.lhs()) * x.rhs().functor().m_other; } +}; +template<typename Scalar, typename Plain1, typename Plain2> +struct blas_traits<CwiseBinaryOp<scalar_product_op<Scalar>, const CwiseNullaryOp<scalar_constant_op<Scalar>,Plain1>, + const CwiseNullaryOp<scalar_constant_op<Scalar>,Plain2> > > + : blas_traits<CwiseNullaryOp<scalar_constant_op<Scalar>,Plain1> > +{}; + +// pop opposite +template<typename Scalar, typename NestedXpr> +struct blas_traits<CwiseUnaryOp<scalar_opposite_op<Scalar>, NestedXpr> > + : blas_traits<NestedXpr> +{ + typedef blas_traits<NestedXpr> Base; + typedef CwiseUnaryOp<scalar_opposite_op<Scalar>, NestedXpr> XprType; + typedef typename Base::ExtractType ExtractType; + static inline ExtractType extract(const XprType& x) { return Base::extract(x.nestedExpression()); } + static inline Scalar extractScalarFactor(const XprType& x) + { return - Base::extractScalarFactor(x.nestedExpression()); } +}; + +// pop/push transpose +template<typename NestedXpr> +struct blas_traits<Transpose<NestedXpr> > + : blas_traits<NestedXpr> +{ + typedef typename NestedXpr::Scalar Scalar; + typedef blas_traits<NestedXpr> Base; + typedef Transpose<NestedXpr> XprType; + typedef Transpose<const typename Base::_ExtractType> ExtractType; // const to get rid of a compile error; anyway blas traits are only used on the RHS + typedef Transpose<const typename Base::_ExtractType> _ExtractType; + typedef typename conditional<bool(Base::HasUsableDirectAccess), + ExtractType, + typename ExtractType::PlainObject + >::type DirectLinearAccessType; + enum { + IsTransposed = Base::IsTransposed ? 0 : 1 + }; + static inline ExtractType extract(const XprType& x) { return ExtractType(Base::extract(x.nestedExpression())); } + static inline Scalar extractScalarFactor(const XprType& x) { return Base::extractScalarFactor(x.nestedExpression()); } +}; + +template<typename T> +struct blas_traits<const T> + : blas_traits<T> +{}; + +template<typename T, bool HasUsableDirectAccess=blas_traits<T>::HasUsableDirectAccess> +struct extract_data_selector { + static const typename T::Scalar* run(const T& m) + { + return blas_traits<T>::extract(m).data(); + } +}; + +template<typename T> +struct extract_data_selector<T,false> { + static typename T::Scalar* run(const T&) { return 0; } +}; + +template<typename T> const typename T::Scalar* extract_data(const T& m) +{ + return extract_data_selector<T>::run(m); +} + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_BLASUTIL_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/Constants.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/Constants.h new file mode 100644 index 000000000..7587d6842 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/Constants.h @@ -0,0 +1,547 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2015 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2007-2009 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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_CONSTANTS_H +#define EIGEN_CONSTANTS_H + +namespace Eigen { + +/** This value means that a positive quantity (e.g., a size) is not known at compile-time, and that instead the value is + * stored in some runtime variable. + * + * Changing the value of Dynamic breaks the ABI, as Dynamic is often used as a template parameter for Matrix. + */ +const int Dynamic = -1; + +/** This value means that a signed quantity (e.g., a signed index) is not known at compile-time, and that instead its value + * has to be specified at runtime. + */ +const int DynamicIndex = 0xffffff; + +/** This value means +Infinity; it is currently used only as the p parameter to MatrixBase::lpNorm<int>(). + * The value Infinity there means the L-infinity norm. + */ +const int Infinity = -1; + +/** This value means that the cost to evaluate an expression coefficient is either very expensive or + * cannot be known at compile time. + * + * This value has to be positive to (1) simplify cost computation, and (2) allow to distinguish between a very expensive and very very expensive expressions. + * It thus must also be large enough to make sure unrolling won't happen and that sub expressions will be evaluated, but not too large to avoid overflow. + */ +const int HugeCost = 10000; + +/** \defgroup flags Flags + * \ingroup Core_Module + * + * These are the possible bits which can be OR'ed to constitute the flags of a matrix or + * expression. + * + * It is important to note that these flags are a purely compile-time notion. They are a compile-time property of + * an expression type, implemented as enum's. They are not stored in memory at runtime, and they do not incur any + * runtime overhead. + * + * \sa MatrixBase::Flags + */ + +/** \ingroup flags + * + * for a matrix, this means that the storage order is row-major. + * If this bit is not set, the storage order is column-major. + * For an expression, this determines the storage order of + * the matrix created by evaluation of that expression. + * \sa \blank \ref TopicStorageOrders */ +const unsigned int RowMajorBit = 0x1; + +/** \ingroup flags + * means the expression should be evaluated by the calling expression */ +const unsigned int EvalBeforeNestingBit = 0x2; + +/** \ingroup flags + * \deprecated + * means the expression should be evaluated before any assignment */ +EIGEN_DEPRECATED +const unsigned int EvalBeforeAssigningBit = 0x4; // FIXME deprecated + +/** \ingroup flags + * + * Short version: means the expression might be vectorized + * + * Long version: means that the coefficients can be handled by packets + * and start at a memory location whose alignment meets the requirements + * of the present CPU architecture for optimized packet access. In the fixed-size + * case, there is the additional condition that it be possible to access all the + * coefficients by packets (this implies the requirement that the size be a multiple of 16 bytes, + * and that any nontrivial strides don't break the alignment). In the dynamic-size case, + * there is no such condition on the total size and strides, so it might not be possible to access + * all coeffs by packets. + * + * \note This bit can be set regardless of whether vectorization is actually enabled. + * To check for actual vectorizability, see \a ActualPacketAccessBit. + */ +const unsigned int PacketAccessBit = 0x8; + +#ifdef EIGEN_VECTORIZE +/** \ingroup flags + * + * If vectorization is enabled (EIGEN_VECTORIZE is defined) this constant + * is set to the value \a PacketAccessBit. + * + * If vectorization is not enabled (EIGEN_VECTORIZE is not defined) this constant + * is set to the value 0. + */ +const unsigned int ActualPacketAccessBit = PacketAccessBit; +#else +const unsigned int ActualPacketAccessBit = 0x0; +#endif + +/** \ingroup flags + * + * Short version: means the expression can be seen as 1D vector. + * + * Long version: means that one can access the coefficients + * of this expression by coeff(int), and coeffRef(int) in the case of a lvalue expression. These + * index-based access methods are guaranteed + * to not have to do any runtime computation of a (row, col)-pair from the index, so that it + * is guaranteed that whenever it is available, index-based access is at least as fast as + * (row,col)-based access. Expressions for which that isn't possible don't have the LinearAccessBit. + * + * If both PacketAccessBit and LinearAccessBit are set, then the + * packets of this expression can be accessed by packet(int), and writePacket(int) in the case of a + * lvalue expression. + * + * Typically, all vector expressions have the LinearAccessBit, but there is one exception: + * Product expressions don't have it, because it would be troublesome for vectorization, even when the + * Product is a vector expression. Thus, vector Product expressions allow index-based coefficient access but + * not index-based packet access, so they don't have the LinearAccessBit. + */ +const unsigned int LinearAccessBit = 0x10; + +/** \ingroup flags + * + * Means the expression has a coeffRef() method, i.e. is writable as its individual coefficients are directly addressable. + * This rules out read-only expressions. + * + * Note that DirectAccessBit and LvalueBit are mutually orthogonal, as there are examples of expression having one but note + * the other: + * \li writable expressions that don't have a very simple memory layout as a strided array, have LvalueBit but not DirectAccessBit + * \li Map-to-const expressions, for example Map<const Matrix>, have DirectAccessBit but not LvalueBit + * + * Expressions having LvalueBit also have their coeff() method returning a const reference instead of returning a new value. + */ +const unsigned int LvalueBit = 0x20; + +/** \ingroup flags + * + * Means that the underlying array of coefficients can be directly accessed as a plain strided array. The memory layout + * of the array of coefficients must be exactly the natural one suggested by rows(), cols(), + * outerStride(), innerStride(), and the RowMajorBit. This rules out expressions such as Diagonal, whose coefficients, + * though referencable, do not have such a regular memory layout. + * + * See the comment on LvalueBit for an explanation of how LvalueBit and DirectAccessBit are mutually orthogonal. + */ +const unsigned int DirectAccessBit = 0x40; + +/** \deprecated \ingroup flags + * + * means the first coefficient packet is guaranteed to be aligned. + * An expression cannot has the AlignedBit without the PacketAccessBit flag. + * In other words, this means we are allow to perform an aligned packet access to the first element regardless + * of the expression kind: + * \code + * expression.packet<Aligned>(0); + * \endcode + */ +EIGEN_DEPRECATED const unsigned int AlignedBit = 0x80; + +const unsigned int NestByRefBit = 0x100; + +/** \ingroup flags + * + * for an expression, this means that the storage order + * can be either row-major or column-major. + * The precise choice will be decided at evaluation time or when + * combined with other expressions. + * \sa \blank \ref RowMajorBit, \ref TopicStorageOrders */ +const unsigned int NoPreferredStorageOrderBit = 0x200; + +/** \ingroup flags + * + * Means that the underlying coefficients can be accessed through pointers to the sparse (un)compressed storage format, + * that is, the expression provides: + * \code + inline const Scalar* valuePtr() const; + inline const Index* innerIndexPtr() const; + inline const Index* outerIndexPtr() const; + inline const Index* innerNonZeroPtr() const; + \endcode + */ +const unsigned int CompressedAccessBit = 0x400; + + +// list of flags that are inherited by default +const unsigned int HereditaryBits = RowMajorBit + | EvalBeforeNestingBit; + +/** \defgroup enums Enumerations + * \ingroup Core_Module + * + * Various enumerations used in %Eigen. Many of these are used as template parameters. + */ + +/** \ingroup enums + * Enum containing possible values for the \c Mode or \c UpLo parameter of + * MatrixBase::selfadjointView() and MatrixBase::triangularView(), and selfadjoint solvers. */ +enum UpLoType { + /** View matrix as a lower triangular matrix. */ + Lower=0x1, + /** View matrix as an upper triangular matrix. */ + Upper=0x2, + /** %Matrix has ones on the diagonal; to be used in combination with #Lower or #Upper. */ + UnitDiag=0x4, + /** %Matrix has zeros on the diagonal; to be used in combination with #Lower or #Upper. */ + ZeroDiag=0x8, + /** View matrix as a lower triangular matrix with ones on the diagonal. */ + UnitLower=UnitDiag|Lower, + /** View matrix as an upper triangular matrix with ones on the diagonal. */ + UnitUpper=UnitDiag|Upper, + /** View matrix as a lower triangular matrix with zeros on the diagonal. */ + StrictlyLower=ZeroDiag|Lower, + /** View matrix as an upper triangular matrix with zeros on the diagonal. */ + StrictlyUpper=ZeroDiag|Upper, + /** Used in BandMatrix and SelfAdjointView to indicate that the matrix is self-adjoint. */ + SelfAdjoint=0x10, + /** Used to support symmetric, non-selfadjoint, complex matrices. */ + Symmetric=0x20 +}; + +/** \ingroup enums + * Enum for indicating whether a buffer is aligned or not. */ +enum AlignmentType { + Unaligned=0, /**< Data pointer has no specific alignment. */ + Aligned8=8, /**< Data pointer is aligned on a 8 bytes boundary. */ + Aligned16=16, /**< Data pointer is aligned on a 16 bytes boundary. */ + Aligned32=32, /**< Data pointer is aligned on a 32 bytes boundary. */ + Aligned64=64, /**< Data pointer is aligned on a 64 bytes boundary. */ + Aligned128=128, /**< Data pointer is aligned on a 128 bytes boundary. */ + AlignedMask=255, + Aligned=16, /**< \deprecated Synonym for Aligned16. */ +#if EIGEN_MAX_ALIGN_BYTES==128 + AlignedMax = Aligned128 +#elif EIGEN_MAX_ALIGN_BYTES==64 + AlignedMax = Aligned64 +#elif EIGEN_MAX_ALIGN_BYTES==32 + AlignedMax = Aligned32 +#elif EIGEN_MAX_ALIGN_BYTES==16 + AlignedMax = Aligned16 +#elif EIGEN_MAX_ALIGN_BYTES==8 + AlignedMax = Aligned8 +#elif EIGEN_MAX_ALIGN_BYTES==0 + AlignedMax = Unaligned +#else +#error Invalid value for EIGEN_MAX_ALIGN_BYTES +#endif +}; + +/** \ingroup enums + * Enum used by DenseBase::corner() in Eigen2 compatibility mode. */ +// FIXME after the corner() API change, this was not needed anymore, except by AlignedBox +// TODO: find out what to do with that. Adapt the AlignedBox API ? +enum CornerType { TopLeft, TopRight, BottomLeft, BottomRight }; + +/** \ingroup enums + * Enum containing possible values for the \p Direction parameter of + * Reverse, PartialReduxExpr and VectorwiseOp. */ +enum DirectionType { + /** For Reverse, all columns are reversed; + * for PartialReduxExpr and VectorwiseOp, act on columns. */ + Vertical, + /** For Reverse, all rows are reversed; + * for PartialReduxExpr and VectorwiseOp, act on rows. */ + Horizontal, + /** For Reverse, both rows and columns are reversed; + * not used for PartialReduxExpr and VectorwiseOp. */ + BothDirections +}; + +/** \internal \ingroup enums + * Enum to specify how to traverse the entries of a matrix. */ +enum TraversalType { + /** \internal Default traversal, no vectorization, no index-based access */ + DefaultTraversal, + /** \internal No vectorization, use index-based access to have only one for loop instead of 2 nested loops */ + LinearTraversal, + /** \internal Equivalent to a slice vectorization for fixed-size matrices having good alignment + * and good size */ + InnerVectorizedTraversal, + /** \internal Vectorization path using a single loop plus scalar loops for the + * unaligned boundaries */ + LinearVectorizedTraversal, + /** \internal Generic vectorization path using one vectorized loop per row/column with some + * scalar loops to handle the unaligned boundaries */ + SliceVectorizedTraversal, + /** \internal Special case to properly handle incompatible scalar types or other defecting cases*/ + InvalidTraversal, + /** \internal Evaluate all entries at once */ + AllAtOnceTraversal +}; + +/** \internal \ingroup enums + * Enum to specify whether to unroll loops when traversing over the entries of a matrix. */ +enum UnrollingType { + /** \internal Do not unroll loops. */ + NoUnrolling, + /** \internal Unroll only the inner loop, but not the outer loop. */ + InnerUnrolling, + /** \internal Unroll both the inner and the outer loop. If there is only one loop, + * because linear traversal is used, then unroll that loop. */ + CompleteUnrolling +}; + +/** \internal \ingroup enums + * Enum to specify whether to use the default (built-in) implementation or the specialization. */ +enum SpecializedType { + Specialized, + BuiltIn +}; + +/** \ingroup enums + * Enum containing possible values for the \p _Options template parameter of + * Matrix, Array and BandMatrix. */ +enum StorageOptions { + /** Storage order is column major (see \ref TopicStorageOrders). */ + ColMajor = 0, + /** Storage order is row major (see \ref TopicStorageOrders). */ + RowMajor = 0x1, // it is only a coincidence that this is equal to RowMajorBit -- don't rely on that + /** Align the matrix itself if it is vectorizable fixed-size */ + AutoAlign = 0, + /** Don't require alignment for the matrix itself (the array of coefficients, if dynamically allocated, may still be requested to be aligned) */ // FIXME --- clarify the situation + DontAlign = 0x2 +}; + +/** \ingroup enums + * Enum for specifying whether to apply or solve on the left or right. */ +enum SideType { + /** Apply transformation on the left. */ + OnTheLeft = 1, + /** Apply transformation on the right. */ + OnTheRight = 2 +}; + +/* the following used to be written as: + * + * struct NoChange_t {}; + * namespace { + * EIGEN_UNUSED NoChange_t NoChange; + * } + * + * on the ground that it feels dangerous to disambiguate overloaded functions on enum/integer types. + * However, this leads to "variable declared but never referenced" warnings on Intel Composer XE, + * and we do not know how to get rid of them (bug 450). + */ + +enum NoChange_t { NoChange }; +enum Sequential_t { Sequential }; +enum Default_t { Default }; + +/** \internal \ingroup enums + * Used in AmbiVector. */ +enum AmbiVectorMode { + IsDense = 0, + IsSparse +}; + +/** \ingroup enums + * Used as template parameter in DenseCoeffBase and MapBase to indicate + * which accessors should be provided. */ +enum AccessorLevels { + /** Read-only access via a member function. */ + ReadOnlyAccessors, + /** Read/write access via member functions. */ + WriteAccessors, + /** Direct read-only access to the coefficients. */ + DirectAccessors, + /** Direct read/write access to the coefficients. */ + DirectWriteAccessors +}; + +/** \ingroup enums + * Enum with options to give to various decompositions. */ +enum DecompositionOptions { + /** \internal Not used (meant for LDLT?). */ + Pivoting = 0x01, + /** \internal Not used (meant for LDLT?). */ + NoPivoting = 0x02, + /** Used in JacobiSVD to indicate that the square matrix U is to be computed. */ + ComputeFullU = 0x04, + /** Used in JacobiSVD to indicate that the thin matrix U is to be computed. */ + ComputeThinU = 0x08, + /** Used in JacobiSVD to indicate that the square matrix V is to be computed. */ + ComputeFullV = 0x10, + /** Used in JacobiSVD to indicate that the thin matrix V is to be computed. */ + ComputeThinV = 0x20, + /** Used in SelfAdjointEigenSolver and GeneralizedSelfAdjointEigenSolver to specify + * that only the eigenvalues are to be computed and not the eigenvectors. */ + EigenvaluesOnly = 0x40, + /** Used in SelfAdjointEigenSolver and GeneralizedSelfAdjointEigenSolver to specify + * that both the eigenvalues and the eigenvectors are to be computed. */ + ComputeEigenvectors = 0x80, + /** \internal */ + EigVecMask = EigenvaluesOnly | ComputeEigenvectors, + /** Used in GeneralizedSelfAdjointEigenSolver to indicate that it should + * solve the generalized eigenproblem \f$ Ax = \lambda B x \f$. */ + Ax_lBx = 0x100, + /** Used in GeneralizedSelfAdjointEigenSolver to indicate that it should + * solve the generalized eigenproblem \f$ ABx = \lambda x \f$. */ + ABx_lx = 0x200, + /** Used in GeneralizedSelfAdjointEigenSolver to indicate that it should + * solve the generalized eigenproblem \f$ BAx = \lambda x \f$. */ + BAx_lx = 0x400, + /** \internal */ + GenEigMask = Ax_lBx | ABx_lx | BAx_lx +}; + +/** \ingroup enums + * Possible values for the \p QRPreconditioner template parameter of JacobiSVD. */ +enum QRPreconditioners { + /** Do not specify what is to be done if the SVD of a non-square matrix is asked for. */ + NoQRPreconditioner, + /** Use a QR decomposition without pivoting as the first step. */ + HouseholderQRPreconditioner, + /** Use a QR decomposition with column pivoting as the first step. */ + ColPivHouseholderQRPreconditioner, + /** Use a QR decomposition with full pivoting as the first step. */ + FullPivHouseholderQRPreconditioner +}; + +#ifdef Success +#error The preprocessor symbol 'Success' is defined, possibly by the X11 header file X.h +#endif + +/** \ingroup enums + * Enum for reporting the status of a computation. */ +enum ComputationInfo { + /** Computation was successful. */ + Success = 0, + /** The provided data did not satisfy the prerequisites. */ + NumericalIssue = 1, + /** Iterative procedure did not converge. */ + NoConvergence = 2, + /** The inputs are invalid, or the algorithm has been improperly called. + * When assertions are enabled, such errors trigger an assert. */ + InvalidInput = 3 +}; + +/** \ingroup enums + * Enum used to specify how a particular transformation is stored in a matrix. + * \sa Transform, Hyperplane::transform(). */ +enum TransformTraits { + /** Transformation is an isometry. */ + Isometry = 0x1, + /** Transformation is an affine transformation stored as a (Dim+1)^2 matrix whose last row is + * assumed to be [0 ... 0 1]. */ + Affine = 0x2, + /** Transformation is an affine transformation stored as a (Dim) x (Dim+1) matrix. */ + AffineCompact = 0x10 | Affine, + /** Transformation is a general projective transformation stored as a (Dim+1)^2 matrix. */ + Projective = 0x20 +}; + +/** \internal \ingroup enums + * Enum used to choose between implementation depending on the computer architecture. */ +namespace Architecture +{ + enum Type { + Generic = 0x0, + SSE = 0x1, + AltiVec = 0x2, + VSX = 0x3, + NEON = 0x4, +#if defined EIGEN_VECTORIZE_SSE + Target = SSE +#elif defined EIGEN_VECTORIZE_ALTIVEC + Target = AltiVec +#elif defined EIGEN_VECTORIZE_VSX + Target = VSX +#elif defined EIGEN_VECTORIZE_NEON + Target = NEON +#else + Target = Generic +#endif + }; +} + +/** \internal \ingroup enums + * Enum used as template parameter in Product and product evaluators. */ +enum ProductImplType +{ DefaultProduct=0, LazyProduct, AliasFreeProduct, CoeffBasedProductMode, LazyCoeffBasedProductMode, OuterProduct, InnerProduct, GemvProduct, GemmProduct }; + +/** \internal \ingroup enums + * Enum used in experimental parallel implementation. */ +enum Action {GetAction, SetAction}; + +/** The type used to identify a dense storage. */ +struct Dense {}; + +/** The type used to identify a general sparse storage. */ +struct Sparse {}; + +/** The type used to identify a general solver (factored) storage. */ +struct SolverStorage {}; + +/** The type used to identify a permutation storage. */ +struct PermutationStorage {}; + +/** The type used to identify a permutation storage. */ +struct TranspositionsStorage {}; + +/** The type used to identify a matrix expression */ +struct MatrixXpr {}; + +/** The type used to identify an array expression */ +struct ArrayXpr {}; + +// An evaluator must define its shape. By default, it can be one of the following: +struct DenseShape { static std::string debugName() { return "DenseShape"; } }; +struct SolverShape { static std::string debugName() { return "SolverShape"; } }; +struct HomogeneousShape { static std::string debugName() { return "HomogeneousShape"; } }; +struct DiagonalShape { static std::string debugName() { return "DiagonalShape"; } }; +struct BandShape { static std::string debugName() { return "BandShape"; } }; +struct TriangularShape { static std::string debugName() { return "TriangularShape"; } }; +struct SelfAdjointShape { static std::string debugName() { return "SelfAdjointShape"; } }; +struct PermutationShape { static std::string debugName() { return "PermutationShape"; } }; +struct TranspositionsShape { static std::string debugName() { return "TranspositionsShape"; } }; +struct SparseShape { static std::string debugName() { return "SparseShape"; } }; + +namespace internal { + + // random access iterators based on coeff*() accessors. +struct IndexBased {}; + +// evaluator based on iterators to access coefficients. +struct IteratorBased {}; + +/** \internal + * Constants for comparison functors + */ +enum ComparisonName { + cmp_EQ = 0, + cmp_LT = 1, + cmp_LE = 2, + cmp_UNORD = 3, + cmp_NEQ = 4, + cmp_GT = 5, + cmp_GE = 6 +}; +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_CONSTANTS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/DisableStupidWarnings.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/DisableStupidWarnings.h new file mode 100755 index 000000000..7559e129c --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/DisableStupidWarnings.h @@ -0,0 +1,75 @@ +#ifndef EIGEN_WARNINGS_DISABLED +#define EIGEN_WARNINGS_DISABLED + +#ifdef _MSC_VER + // 4100 - unreferenced formal parameter (occurred e.g. in aligned_allocator::destroy(pointer p)) + // 4101 - unreferenced local variable + // 4127 - conditional expression is constant + // 4181 - qualifier applied to reference type ignored + // 4211 - nonstandard extension used : redefined extern to static + // 4244 - 'argument' : conversion from 'type1' to 'type2', possible loss of data + // 4273 - QtAlignedMalloc, inconsistent DLL linkage + // 4324 - structure was padded due to declspec(align()) + // 4503 - decorated name length exceeded, name was truncated + // 4512 - assignment operator could not be generated + // 4522 - 'class' : multiple assignment operators specified + // 4700 - uninitialized local variable 'xyz' used + // 4714 - function marked as __forceinline not inlined + // 4717 - 'function' : recursive on all control paths, function will cause runtime stack overflow + // 4800 - 'type' : forcing value to bool 'true' or 'false' (performance warning) + #ifndef EIGEN_PERMANENTLY_DISABLE_STUPID_WARNINGS + #pragma warning( push ) + #endif + #pragma warning( disable : 4100 4101 4127 4181 4211 4244 4273 4324 4503 4512 4522 4700 4714 4717 4800) + +#elif defined __INTEL_COMPILER + // 2196 - routine is both "inline" and "noinline" ("noinline" assumed) + // ICC 12 generates this warning even without any inline keyword, when defining class methods 'inline' i.e. inside of class body + // typedef that may be a reference type. + // 279 - controlling expression is constant + // ICC 12 generates this warning on assert(constant_expression_depending_on_template_params) and frankly this is a legitimate use case. + // 1684 - conversion from pointer to same-sized integral type (potential portability problem) + // 2259 - non-pointer conversion from "Eigen::Index={ptrdiff_t={long}}" to "int" may lose significant bits + #ifndef EIGEN_PERMANENTLY_DISABLE_STUPID_WARNINGS + #pragma warning push + #endif + #pragma warning disable 2196 279 1684 2259 + +#elif defined __clang__ + // -Wconstant-logical-operand - warning: use of logical && with constant operand; switch to bitwise & or remove constant + // this is really a stupid warning as it warns on compile-time expressions involving enums + #ifndef EIGEN_PERMANENTLY_DISABLE_STUPID_WARNINGS + #pragma clang diagnostic push + #endif + #pragma clang diagnostic ignored "-Wconstant-logical-operand" + +#elif defined __GNUC__ && __GNUC__>=6 + + #ifndef EIGEN_PERMANENTLY_DISABLE_STUPID_WARNINGS + #pragma GCC diagnostic push + #endif + #pragma GCC diagnostic ignored "-Wignored-attributes" + +#endif + +#if defined __NVCC__ + // Disable the "statement is unreachable" message + #pragma diag_suppress code_is_unreachable + // Disable the "dynamic initialization in unreachable code" message + #pragma diag_suppress initialization_not_reachable + // Disable the "invalid error number" message that we get with older versions of nvcc + #pragma diag_suppress 1222 + // Disable the "calling a __host__ function from a __host__ __device__ function is not allowed" messages (yes, there are many of them and they seem to change with every version of the compiler) + #pragma diag_suppress 2527 + #pragma diag_suppress 2529 + #pragma diag_suppress 2651 + #pragma diag_suppress 2653 + #pragma diag_suppress 2668 + #pragma diag_suppress 2669 + #pragma diag_suppress 2670 + #pragma diag_suppress 2671 + #pragma diag_suppress 2735 + #pragma diag_suppress 2737 +#endif + +#endif // not EIGEN_WARNINGS_DISABLED diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/ForwardDeclarations.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/ForwardDeclarations.h new file mode 100644 index 000000000..ea107393a --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/ForwardDeclarations.h @@ -0,0 +1,302 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2007-2010 Benoit Jacob <jacob.benoit.1@gmail.com> +// Copyright (C) 2008-2009 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_FORWARDDECLARATIONS_H +#define EIGEN_FORWARDDECLARATIONS_H + +namespace Eigen { +namespace internal { + +template<typename T> struct traits; + +// here we say once and for all that traits<const T> == traits<T> +// When constness must affect traits, it has to be constness on template parameters on which T itself depends. +// For example, traits<Map<const T> > != traits<Map<T> >, but +// traits<const Map<T> > == traits<Map<T> > +template<typename T> struct traits<const T> : traits<T> {}; + +template<typename Derived> struct has_direct_access +{ + enum { ret = (traits<Derived>::Flags & DirectAccessBit) ? 1 : 0 }; +}; + +template<typename Derived> struct accessors_level +{ + enum { has_direct_access = (traits<Derived>::Flags & DirectAccessBit) ? 1 : 0, + has_write_access = (traits<Derived>::Flags & LvalueBit) ? 1 : 0, + value = has_direct_access ? (has_write_access ? DirectWriteAccessors : DirectAccessors) + : (has_write_access ? WriteAccessors : ReadOnlyAccessors) + }; +}; + +template<typename T> struct evaluator_traits; + +template< typename T> struct evaluator; + +} // end namespace internal + +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 _Scalar, int _Rows, int _Cols, + int _Options = AutoAlign | +#if EIGEN_GNUC_AT(3,4) + // workaround a bug in at least gcc 3.4.6 + // the innermost ?: ternary operator is misparsed. We write it slightly + // differently and this makes gcc 3.4.6 happy, but it's ugly. + // The error would only show up with EIGEN_DEFAULT_TO_ROW_MAJOR is defined + // (when EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION is RowMajor) + ( (_Rows==1 && _Cols!=1) ? Eigen::RowMajor + : !(_Cols==1 && _Rows!=1) ? EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION + : Eigen::ColMajor ), +#else + ( (_Rows==1 && _Cols!=1) ? Eigen::RowMajor + : (_Cols==1 && _Rows!=1) ? Eigen::ColMajor + : EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION ), +#endif + int _MaxRows = _Rows, + int _MaxCols = _Cols +> class Matrix; + +template<typename Derived> class MatrixBase; +template<typename Derived> class ArrayBase; + +template<typename ExpressionType, unsigned int Added, unsigned int Removed> class Flagged; +template<typename ExpressionType, template <typename> class StorageBase > class NoAlias; +template<typename ExpressionType> class NestByValue; +template<typename ExpressionType> class ForceAlignedAccess; +template<typename ExpressionType> class SwapWrapper; + +template<typename XprType, int BlockRows=Dynamic, int BlockCols=Dynamic, bool InnerPanel = false> class Block; + +template<typename MatrixType, int Size=Dynamic> class VectorBlock; +template<typename MatrixType> class Transpose; +template<typename MatrixType> class Conjugate; +template<typename NullaryOp, typename MatrixType> class CwiseNullaryOp; +template<typename UnaryOp, typename MatrixType> class CwiseUnaryOp; +template<typename ViewOp, typename MatrixType> class CwiseUnaryView; +template<typename BinaryOp, typename Lhs, typename Rhs> class CwiseBinaryOp; +template<typename TernaryOp, typename Arg1, typename Arg2, typename Arg3> class CwiseTernaryOp; +template<typename Decomposition, typename Rhstype> class Solve; +template<typename XprType> class Inverse; + +template<typename Lhs, typename Rhs, int Option = DefaultProduct> class Product; + +template<typename Derived> class DiagonalBase; +template<typename _DiagonalVectorType> class DiagonalWrapper; +template<typename _Scalar, int SizeAtCompileTime, int MaxSizeAtCompileTime=SizeAtCompileTime> class DiagonalMatrix; +template<typename MatrixType, typename DiagonalType, int ProductOrder> class DiagonalProduct; +template<typename MatrixType, int Index = 0> class Diagonal; +template<int SizeAtCompileTime, int MaxSizeAtCompileTime = SizeAtCompileTime, typename IndexType=int> class PermutationMatrix; +template<int SizeAtCompileTime, int MaxSizeAtCompileTime = SizeAtCompileTime, typename IndexType=int> class Transpositions; +template<typename Derived> class PermutationBase; +template<typename Derived> class TranspositionsBase; +template<typename _IndicesType> class PermutationWrapper; +template<typename _IndicesType> class TranspositionsWrapper; + +template<typename Derived, + int Level = internal::accessors_level<Derived>::has_write_access ? WriteAccessors : ReadOnlyAccessors +> class MapBase; +template<int InnerStrideAtCompileTime, int OuterStrideAtCompileTime> class Stride; +template<int Value = Dynamic> class InnerStride; +template<int Value = Dynamic> class OuterStride; +template<typename MatrixType, int MapOptions=Unaligned, typename StrideType = Stride<0,0> > class Map; +template<typename Derived> class RefBase; +template<typename PlainObjectType, int Options = 0, + typename StrideType = typename internal::conditional<PlainObjectType::IsVectorAtCompileTime,InnerStride<1>,OuterStride<> >::type > class Ref; + +template<typename Derived> class TriangularBase; +template<typename MatrixType, unsigned int Mode> class TriangularView; +template<typename MatrixType, unsigned int Mode> class SelfAdjointView; +template<typename MatrixType> class SparseView; +template<typename ExpressionType> class WithFormat; +template<typename MatrixType> struct CommaInitializer; +template<typename Derived> class ReturnByValue; +template<typename ExpressionType> class ArrayWrapper; +template<typename ExpressionType> class MatrixWrapper; +template<typename Derived> class SolverBase; +template<typename XprType> class InnerIterator; + +namespace internal { +template<typename DecompositionType> struct kernel_retval_base; +template<typename DecompositionType> struct kernel_retval; +template<typename DecompositionType> struct image_retval_base; +template<typename DecompositionType> struct image_retval; +} // end namespace internal + +namespace internal { +template<typename _Scalar, int Rows=Dynamic, int Cols=Dynamic, int Supers=Dynamic, int Subs=Dynamic, int Options=0> class BandMatrix; +} + +namespace internal { +template<typename Lhs, typename Rhs> struct product_type; + +template<bool> struct EnableIf; + +/** \internal + * \class product_evaluator + * Products need their own evaluator with more template arguments allowing for + * easier partial template specializations. + */ +template< typename T, + int ProductTag = internal::product_type<typename T::Lhs,typename T::Rhs>::ret, + typename LhsShape = typename evaluator_traits<typename T::Lhs>::Shape, + typename RhsShape = typename evaluator_traits<typename T::Rhs>::Shape, + typename LhsScalar = typename traits<typename T::Lhs>::Scalar, + typename RhsScalar = typename traits<typename T::Rhs>::Scalar + > struct product_evaluator; +} + +template<typename Lhs, typename Rhs, + int ProductType = internal::product_type<Lhs,Rhs>::value> +struct ProductReturnType; + +// this is a workaround for sun CC +template<typename Lhs, typename Rhs> struct LazyProductReturnType; + +namespace internal { + +// Provides scalar/packet-wise product and product with accumulation +// with optional conjugation of the arguments. +template<typename LhsScalar, typename RhsScalar, bool ConjLhs=false, bool ConjRhs=false> struct conj_helper; + +template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_sum_op; +template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_difference_op; +template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_conj_product_op; +template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_min_op; +template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_max_op; +template<typename Scalar> struct scalar_opposite_op; +template<typename Scalar> struct scalar_conjugate_op; +template<typename Scalar> struct scalar_real_op; +template<typename Scalar> struct scalar_imag_op; +template<typename Scalar> struct scalar_abs_op; +template<typename Scalar> struct scalar_abs2_op; +template<typename Scalar> struct scalar_sqrt_op; +template<typename Scalar> struct scalar_rsqrt_op; +template<typename Scalar> struct scalar_exp_op; +template<typename Scalar> struct scalar_log_op; +template<typename Scalar> struct scalar_cos_op; +template<typename Scalar> struct scalar_sin_op; +template<typename Scalar> struct scalar_acos_op; +template<typename Scalar> struct scalar_asin_op; +template<typename Scalar> struct scalar_tan_op; +template<typename Scalar> struct scalar_inverse_op; +template<typename Scalar> struct scalar_square_op; +template<typename Scalar> struct scalar_cube_op; +template<typename Scalar, typename NewType> struct scalar_cast_op; +template<typename Scalar> struct scalar_random_op; +template<typename Scalar> struct scalar_constant_op; +template<typename Scalar> struct scalar_identity_op; +template<typename Scalar,bool iscpx> struct scalar_sign_op; +template<typename Scalar,typename ScalarExponent> struct scalar_pow_op; +template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_hypot_op; +template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_product_op; +template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_quotient_op; + +// SpecialFunctions module +template<typename Scalar> struct scalar_lgamma_op; +template<typename Scalar> struct scalar_digamma_op; +template<typename Scalar> struct scalar_erf_op; +template<typename Scalar> struct scalar_erfc_op; +template<typename Scalar> struct scalar_igamma_op; +template<typename Scalar> struct scalar_igammac_op; +template<typename Scalar> struct scalar_zeta_op; +template<typename Scalar> struct scalar_betainc_op; + +} // end namespace internal + +struct IOFormat; + +// Array module +template<typename _Scalar, int _Rows, int _Cols, + int _Options = AutoAlign | +#if EIGEN_GNUC_AT(3,4) + // workaround a bug in at least gcc 3.4.6 + // the innermost ?: ternary operator is misparsed. We write it slightly + // differently and this makes gcc 3.4.6 happy, but it's ugly. + // The error would only show up with EIGEN_DEFAULT_TO_ROW_MAJOR is defined + // (when EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION is RowMajor) + ( (_Rows==1 && _Cols!=1) ? Eigen::RowMajor + : !(_Cols==1 && _Rows!=1) ? EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION + : Eigen::ColMajor ), +#else + ( (_Rows==1 && _Cols!=1) ? Eigen::RowMajor + : (_Cols==1 && _Rows!=1) ? Eigen::ColMajor + : EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION ), +#endif + int _MaxRows = _Rows, int _MaxCols = _Cols> class Array; +template<typename ConditionMatrixType, typename ThenMatrixType, typename ElseMatrixType> class Select; +template<typename MatrixType, typename BinaryOp, int Direction> class PartialReduxExpr; +template<typename ExpressionType, int Direction> class VectorwiseOp; +template<typename MatrixType,int RowFactor,int ColFactor> class Replicate; +template<typename MatrixType, int Direction = BothDirections> class Reverse; + +template<typename MatrixType> class FullPivLU; +template<typename MatrixType> class PartialPivLU; +namespace internal { +template<typename MatrixType> struct inverse_impl; +} +template<typename MatrixType> class HouseholderQR; +template<typename MatrixType> class ColPivHouseholderQR; +template<typename MatrixType> class FullPivHouseholderQR; +template<typename MatrixType> class CompleteOrthogonalDecomposition; +template<typename MatrixType, int QRPreconditioner = ColPivHouseholderQRPreconditioner> class JacobiSVD; +template<typename MatrixType> class BDCSVD; +template<typename MatrixType, int UpLo = Lower> class LLT; +template<typename MatrixType, int UpLo = Lower> class LDLT; +template<typename VectorsType, typename CoeffsType, int Side=OnTheLeft> class HouseholderSequence; +template<typename Scalar> class JacobiRotation; + +// Geometry module: +template<typename Derived, int _Dim> class RotationBase; +template<typename Lhs, typename Rhs> class Cross; +template<typename Derived> class QuaternionBase; +template<typename Scalar> class Rotation2D; +template<typename Scalar> class AngleAxis; +template<typename Scalar,int Dim> class Translation; +template<typename Scalar,int Dim> class AlignedBox; +template<typename Scalar, int Options = AutoAlign> class Quaternion; +template<typename Scalar,int Dim,int Mode,int _Options=AutoAlign> class Transform; +template <typename _Scalar, int _AmbientDim, int Options=AutoAlign> class ParametrizedLine; +template <typename _Scalar, int _AmbientDim, int Options=AutoAlign> class Hyperplane; +template<typename Scalar> class UniformScaling; +template<typename MatrixType,int Direction> class Homogeneous; + +// Sparse module: +template<typename Derived> class SparseMatrixBase; + +// MatrixFunctions module +template<typename Derived> struct MatrixExponentialReturnValue; +template<typename Derived> class MatrixFunctionReturnValue; +template<typename Derived> class MatrixSquareRootReturnValue; +template<typename Derived> class MatrixLogarithmReturnValue; +template<typename Derived> class MatrixPowerReturnValue; +template<typename Derived> class MatrixComplexPowerReturnValue; + +namespace internal { +template <typename Scalar> +struct stem_function +{ + typedef std::complex<typename NumTraits<Scalar>::Real> ComplexScalar; + typedef ComplexScalar type(ComplexScalar, int); +}; +} + +} // end namespace Eigen + +#endif // EIGEN_FORWARDDECLARATIONS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/MKL_support.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/MKL_support.h new file mode 100755 index 000000000..26b59669e --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/MKL_support.h @@ -0,0 +1,128 @@ +/* + Copyright (c) 2011, Intel Corporation. All rights reserved. + + Redistribution and use in source and binary forms, with or without modification, + are permitted provided that the following conditions are met: + + * Redistributions of source code must retain the above copyright notice, this + list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above copyright notice, + this list of conditions and the following disclaimer in the documentation + and/or other materials provided with the distribution. + * Neither the name of Intel Corporation nor the names of its contributors may + be used to endorse or promote products derived from this software without + specific prior written permission. + + THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND + ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED + WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE + DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR + ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES + (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; + LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON + ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS + SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + + ******************************************************************************** + * Content : Eigen bindings to Intel(R) MKL + * Include file with common MKL declarations + ******************************************************************************** +*/ + +#ifndef EIGEN_MKL_SUPPORT_H +#define EIGEN_MKL_SUPPORT_H + +#ifdef EIGEN_USE_MKL_ALL + #ifndef EIGEN_USE_BLAS + #define EIGEN_USE_BLAS + #endif + #ifndef EIGEN_USE_LAPACKE + #define EIGEN_USE_LAPACKE + #endif + #ifndef EIGEN_USE_MKL_VML + #define EIGEN_USE_MKL_VML + #endif +#endif + +#ifdef EIGEN_USE_LAPACKE_STRICT + #define EIGEN_USE_LAPACKE +#endif + +#if defined(EIGEN_USE_MKL_VML) + #define EIGEN_USE_MKL +#endif + +#if defined EIGEN_USE_MKL +# include <mkl.h> +/*Check IMKL version for compatibility: < 10.3 is not usable with Eigen*/ +# ifndef INTEL_MKL_VERSION +# undef EIGEN_USE_MKL /* INTEL_MKL_VERSION is not even defined on older versions */ +# elif INTEL_MKL_VERSION < 100305 /* the intel-mkl-103-release-notes say this was when the lapacke.h interface was added*/ +# undef EIGEN_USE_MKL +# endif +# ifndef EIGEN_USE_MKL + /*If the MKL version is too old, undef everything*/ +# undef EIGEN_USE_MKL_ALL +# undef EIGEN_USE_LAPACKE +# undef EIGEN_USE_MKL_VML +# undef EIGEN_USE_LAPACKE_STRICT +# undef EIGEN_USE_LAPACKE +# endif +#endif + +#if defined EIGEN_USE_MKL + +#define EIGEN_MKL_VML_THRESHOLD 128 + +/* MKL_DOMAIN_BLAS, etc are defined only in 10.3 update 7 */ +/* MKL_BLAS, etc are not defined in 11.2 */ +#ifdef MKL_DOMAIN_ALL +#define EIGEN_MKL_DOMAIN_ALL MKL_DOMAIN_ALL +#else +#define EIGEN_MKL_DOMAIN_ALL MKL_ALL +#endif + +#ifdef MKL_DOMAIN_BLAS +#define EIGEN_MKL_DOMAIN_BLAS MKL_DOMAIN_BLAS +#else +#define EIGEN_MKL_DOMAIN_BLAS MKL_BLAS +#endif + +#ifdef MKL_DOMAIN_FFT +#define EIGEN_MKL_DOMAIN_FFT MKL_DOMAIN_FFT +#else +#define EIGEN_MKL_DOMAIN_FFT MKL_FFT +#endif + +#ifdef MKL_DOMAIN_VML +#define EIGEN_MKL_DOMAIN_VML MKL_DOMAIN_VML +#else +#define EIGEN_MKL_DOMAIN_VML MKL_VML +#endif + +#ifdef MKL_DOMAIN_PARDISO +#define EIGEN_MKL_DOMAIN_PARDISO MKL_DOMAIN_PARDISO +#else +#define EIGEN_MKL_DOMAIN_PARDISO MKL_PARDISO +#endif +#endif + +namespace Eigen { + +typedef std::complex<double> dcomplex; +typedef std::complex<float> scomplex; + +#if defined(EIGEN_USE_MKL) +typedef MKL_INT BlasIndex; +#else +typedef int BlasIndex; +#endif + +} // end namespace Eigen + +#if defined(EIGEN_USE_BLAS) +#include "../../misc/blas.h" +#endif + +#endif // EIGEN_MKL_SUPPORT_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/Macros.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/Macros.h new file mode 100644 index 000000000..38d6ddb9a --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/Macros.h @@ -0,0 +1,992 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2015 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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_MACROS_H +#define EIGEN_MACROS_H + +#define EIGEN_WORLD_VERSION 3 +#define EIGEN_MAJOR_VERSION 3 +#define EIGEN_MINOR_VERSION 4 + +#define EIGEN_VERSION_AT_LEAST(x,y,z) (EIGEN_WORLD_VERSION>x || (EIGEN_WORLD_VERSION>=x && \ + (EIGEN_MAJOR_VERSION>y || (EIGEN_MAJOR_VERSION>=y && \ + EIGEN_MINOR_VERSION>=z)))) + +// Compiler identification, EIGEN_COMP_* + +/// \internal EIGEN_COMP_GNUC set to 1 for all compilers compatible with GCC +#ifdef __GNUC__ + #define EIGEN_COMP_GNUC 1 +#else + #define EIGEN_COMP_GNUC 0 +#endif + +/// \internal EIGEN_COMP_CLANG set to major+minor version (e.g., 307 for clang 3.7) if the compiler is clang +#if defined(__clang__) + #define EIGEN_COMP_CLANG (__clang_major__*100+__clang_minor__) +#else + #define EIGEN_COMP_CLANG 0 +#endif + + +/// \internal EIGEN_COMP_LLVM set to 1 if the compiler backend is llvm +#if defined(__llvm__) + #define EIGEN_COMP_LLVM 1 +#else + #define EIGEN_COMP_LLVM 0 +#endif + +/// \internal EIGEN_COMP_ICC set to __INTEL_COMPILER if the compiler is Intel compiler, 0 otherwise +#if defined(__INTEL_COMPILER) + #define EIGEN_COMP_ICC __INTEL_COMPILER +#else + #define EIGEN_COMP_ICC 0 +#endif + +/// \internal EIGEN_COMP_MINGW set to 1 if the compiler is mingw +#if defined(__MINGW32__) + #define EIGEN_COMP_MINGW 1 +#else + #define EIGEN_COMP_MINGW 0 +#endif + +/// \internal EIGEN_COMP_SUNCC set to 1 if the compiler is Solaris Studio +#if defined(__SUNPRO_CC) + #define EIGEN_COMP_SUNCC 1 +#else + #define EIGEN_COMP_SUNCC 0 +#endif + +/// \internal EIGEN_COMP_MSVC set to _MSC_VER if the compiler is Microsoft Visual C++, 0 otherwise. +#if defined(_MSC_VER) + #define EIGEN_COMP_MSVC _MSC_VER +#else + #define EIGEN_COMP_MSVC 0 +#endif + +// For the record, here is a table summarizing the possible values for EIGEN_COMP_MSVC: +// name ver MSC_VER +// 2008 9 1500 +// 2010 10 1600 +// 2012 11 1700 +// 2013 12 1800 +// 2015 14 1900 +// "15" 15 1900 + +/// \internal EIGEN_COMP_MSVC_STRICT set to 1 if the compiler is really Microsoft Visual C++ and not ,e.g., ICC or clang-cl +#if EIGEN_COMP_MSVC && !(EIGEN_COMP_ICC || EIGEN_COMP_LLVM || EIGEN_COMP_CLANG) + #define EIGEN_COMP_MSVC_STRICT _MSC_VER +#else + #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 +#else + #define EIGEN_COMP_IBM 0 +#endif + +/// \internal EIGEN_COMP_PGI set to 1 if the compiler is Portland Group Compiler +#if defined(__PGI) + #define EIGEN_COMP_PGI 1 +#else + #define EIGEN_COMP_PGI 0 +#endif + +/// \internal EIGEN_COMP_ARM set to 1 if the compiler is ARM Compiler +#if defined(__CC_ARM) || defined(__ARMCC_VERSION) + #define EIGEN_COMP_ARM 1 +#else + #define EIGEN_COMP_ARM 0 +#endif + +/// \internal EIGEN_COMP_ARM set to 1 if the compiler is ARM Compiler +#if defined(__EMSCRIPTEN__) + #define EIGEN_COMP_EMSCRIPTEN 1 +#else + #define EIGEN_COMP_EMSCRIPTEN 0 +#endif + + +/// \internal EIGEN_GNUC_STRICT set to 1 if the compiler is really GCC and not a compatible compiler (e.g., ICC, clang, mingw, etc.) +#if EIGEN_COMP_GNUC && !(EIGEN_COMP_CLANG || EIGEN_COMP_ICC || EIGEN_COMP_MINGW || EIGEN_COMP_PGI || EIGEN_COMP_IBM || EIGEN_COMP_ARM || EIGEN_COMP_EMSCRIPTEN) + #define EIGEN_COMP_GNUC_STRICT 1 +#else + #define EIGEN_COMP_GNUC_STRICT 0 +#endif + + +#if EIGEN_COMP_GNUC + #define EIGEN_GNUC_AT_LEAST(x,y) ((__GNUC__==x && __GNUC_MINOR__>=y) || __GNUC__>x) + #define EIGEN_GNUC_AT_MOST(x,y) ((__GNUC__==x && __GNUC_MINOR__<=y) || __GNUC__<x) + #define EIGEN_GNUC_AT(x,y) ( __GNUC__==x && __GNUC_MINOR__==y ) +#else + #define EIGEN_GNUC_AT_LEAST(x,y) 0 + #define EIGEN_GNUC_AT_MOST(x,y) 0 + #define EIGEN_GNUC_AT(x,y) 0 +#endif + +// FIXME: could probably be removed as we do not support gcc 3.x anymore +#if EIGEN_COMP_GNUC && (__GNUC__ <= 3) +#define EIGEN_GCC3_OR_OLDER 1 +#else +#define EIGEN_GCC3_OR_OLDER 0 +#endif + + +// Architecture identification, EIGEN_ARCH_* + +#if defined(__x86_64__) || defined(_M_X64) || defined(__amd64) + #define EIGEN_ARCH_x86_64 1 +#else + #define EIGEN_ARCH_x86_64 0 +#endif + +#if defined(__i386__) || defined(_M_IX86) || defined(_X86_) || defined(__i386) + #define EIGEN_ARCH_i386 1 +#else + #define EIGEN_ARCH_i386 0 +#endif + +#if EIGEN_ARCH_x86_64 || EIGEN_ARCH_i386 + #define EIGEN_ARCH_i386_OR_x86_64 1 +#else + #define EIGEN_ARCH_i386_OR_x86_64 0 +#endif + +/// \internal EIGEN_ARCH_ARM set to 1 if the architecture is ARM +#if defined(__arm__) + #define EIGEN_ARCH_ARM 1 +#else + #define EIGEN_ARCH_ARM 0 +#endif + +/// \internal EIGEN_ARCH_ARM64 set to 1 if the architecture is ARM64 +#if defined(__aarch64__) + #define EIGEN_ARCH_ARM64 1 +#else + #define EIGEN_ARCH_ARM64 0 +#endif + +#if EIGEN_ARCH_ARM || EIGEN_ARCH_ARM64 + #define EIGEN_ARCH_ARM_OR_ARM64 1 +#else + #define EIGEN_ARCH_ARM_OR_ARM64 0 +#endif + +/// \internal EIGEN_ARCH_MIPS set to 1 if the architecture is MIPS +#if defined(__mips__) || defined(__mips) + #define EIGEN_ARCH_MIPS 1 +#else + #define EIGEN_ARCH_MIPS 0 +#endif + +/// \internal EIGEN_ARCH_SPARC set to 1 if the architecture is SPARC +#if defined(__sparc__) || defined(__sparc) + #define EIGEN_ARCH_SPARC 1 +#else + #define EIGEN_ARCH_SPARC 0 +#endif + +/// \internal EIGEN_ARCH_IA64 set to 1 if the architecture is Intel Itanium +#if defined(__ia64__) + #define EIGEN_ARCH_IA64 1 +#else + #define EIGEN_ARCH_IA64 0 +#endif + +/// \internal EIGEN_ARCH_PPC set to 1 if the architecture is PowerPC +#if defined(__powerpc__) || defined(__ppc__) || defined(_M_PPC) + #define EIGEN_ARCH_PPC 1 +#else + #define EIGEN_ARCH_PPC 0 +#endif + + + +// Operating system identification, EIGEN_OS_* + +/// \internal EIGEN_OS_UNIX set to 1 if the OS is a unix variant +#if defined(__unix__) || defined(__unix) + #define EIGEN_OS_UNIX 1 +#else + #define EIGEN_OS_UNIX 0 +#endif + +/// \internal EIGEN_OS_LINUX set to 1 if the OS is based on Linux kernel +#if defined(__linux__) + #define EIGEN_OS_LINUX 1 +#else + #define EIGEN_OS_LINUX 0 +#endif + +/// \internal EIGEN_OS_ANDROID set to 1 if the OS is Android +// note: ANDROID is defined when using ndk_build, __ANDROID__ is defined when using a standalone toolchain. +#if defined(__ANDROID__) || defined(ANDROID) + #define EIGEN_OS_ANDROID 1 +#else + #define EIGEN_OS_ANDROID 0 +#endif + +/// \internal EIGEN_OS_GNULINUX set to 1 if the OS is GNU Linux and not Linux-based OS (e.g., not android) +#if defined(__gnu_linux__) && !(EIGEN_OS_ANDROID) + #define EIGEN_OS_GNULINUX 1 +#else + #define EIGEN_OS_GNULINUX 0 +#endif + +/// \internal EIGEN_OS_BSD set to 1 if the OS is a BSD variant +#if defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) || defined(__bsdi__) || defined(__DragonFly__) + #define EIGEN_OS_BSD 1 +#else + #define EIGEN_OS_BSD 0 +#endif + +/// \internal EIGEN_OS_MAC set to 1 if the OS is MacOS +#if defined(__APPLE__) + #define EIGEN_OS_MAC 1 +#else + #define EIGEN_OS_MAC 0 +#endif + +/// \internal EIGEN_OS_QNX set to 1 if the OS is QNX +#if defined(__QNX__) + #define EIGEN_OS_QNX 1 +#else + #define EIGEN_OS_QNX 0 +#endif + +/// \internal EIGEN_OS_WIN set to 1 if the OS is Windows based +#if defined(_WIN32) + #define EIGEN_OS_WIN 1 +#else + #define EIGEN_OS_WIN 0 +#endif + +/// \internal EIGEN_OS_WIN64 set to 1 if the OS is Windows 64bits +#if defined(_WIN64) + #define EIGEN_OS_WIN64 1 +#else + #define EIGEN_OS_WIN64 0 +#endif + +/// \internal EIGEN_OS_WINCE set to 1 if the OS is Windows CE +#if defined(_WIN32_WCE) + #define EIGEN_OS_WINCE 1 +#else + #define EIGEN_OS_WINCE 0 +#endif + +/// \internal EIGEN_OS_CYGWIN set to 1 if the OS is Windows/Cygwin +#if defined(__CYGWIN__) + #define EIGEN_OS_CYGWIN 1 +#else + #define EIGEN_OS_CYGWIN 0 +#endif + +/// \internal EIGEN_OS_WIN_STRICT set to 1 if the OS is really Windows and not some variants +#if EIGEN_OS_WIN && !( EIGEN_OS_WINCE || EIGEN_OS_CYGWIN ) + #define EIGEN_OS_WIN_STRICT 1 +#else + #define EIGEN_OS_WIN_STRICT 0 +#endif + +/// \internal EIGEN_OS_SUN set to 1 if the OS is SUN +#if (defined(sun) || defined(__sun)) && !(defined(__SVR4) || defined(__svr4__)) + #define EIGEN_OS_SUN 1 +#else + #define EIGEN_OS_SUN 0 +#endif + +/// \internal EIGEN_OS_SOLARIS set to 1 if the OS is Solaris +#if (defined(sun) || defined(__sun)) && (defined(__SVR4) || defined(__svr4__)) + #define EIGEN_OS_SOLARIS 1 +#else + #define EIGEN_OS_SOLARIS 0 +#endif + + + +#if EIGEN_GNUC_AT_MOST(4,3) && !EIGEN_COMP_CLANG + // see bug 89 + #define EIGEN_SAFE_TO_USE_STANDARD_ASSERT_MACRO 0 +#else + #define EIGEN_SAFE_TO_USE_STANDARD_ASSERT_MACRO 1 +#endif + +// This macro can be used to prevent from macro expansion, e.g.: +// std::max EIGEN_NOT_A_MACRO(a,b) +#define EIGEN_NOT_A_MACRO + +#ifdef EIGEN_DEFAULT_TO_ROW_MAJOR +#define EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION Eigen::RowMajor +#else +#define EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION Eigen::ColMajor +#endif + +#ifndef EIGEN_DEFAULT_DENSE_INDEX_TYPE +#define EIGEN_DEFAULT_DENSE_INDEX_TYPE std::ptrdiff_t +#endif + +// Cross compiler wrapper around LLVM's __has_builtin +#ifdef __has_builtin +# define EIGEN_HAS_BUILTIN(x) __has_builtin(x) +#else +# define EIGEN_HAS_BUILTIN(x) 0 +#endif + +// A Clang feature extension to determine compiler features. +// We use it to determine 'cxx_rvalue_references' +#ifndef __has_feature +# define __has_feature(x) 0 +#endif + +// Upperbound on the C++ version to use. +// Expected values are 03, 11, 14, 17, etc. +// By default, let's use an arbitrarily large C++ version. +#ifndef EIGEN_MAX_CPP_VER +#define EIGEN_MAX_CPP_VER 99 +#endif + +#if EIGEN_MAX_CPP_VER>=11 && (defined(__cplusplus) && (__cplusplus >= 201103L) || EIGEN_COMP_MSVC >= 1900) +#define EIGEN_HAS_CXX11 1 +#else +#define EIGEN_HAS_CXX11 0 +#endif + + +// Do we support r-value references? +#ifndef EIGEN_HAS_RVALUE_REFERENCES +#if EIGEN_MAX_CPP_VER>=11 && \ + (__has_feature(cxx_rvalue_references) || \ + (defined(__cplusplus) && __cplusplus >= 201103L) || \ + (EIGEN_COMP_MSVC >= 1600)) + #define EIGEN_HAS_RVALUE_REFERENCES 1 +#else + #define EIGEN_HAS_RVALUE_REFERENCES 0 +#endif +#endif + +// Does the compiler support C99? +#ifndef EIGEN_HAS_C99_MATH +#if EIGEN_MAX_CPP_VER>=11 && \ + ((defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901)) \ + || (defined(__GNUC__) && defined(_GLIBCXX_USE_C99)) \ + || (defined(_LIBCPP_VERSION) && !defined(_MSC_VER))) + #define EIGEN_HAS_C99_MATH 1 +#else + #define EIGEN_HAS_C99_MATH 0 +#endif +#endif + +// Does the compiler support result_of? +#ifndef EIGEN_HAS_STD_RESULT_OF +#if EIGEN_MAX_CPP_VER>=11 && ((__has_feature(cxx_lambdas) || (defined(__cplusplus) && __cplusplus >= 201103L))) +#define EIGEN_HAS_STD_RESULT_OF 1 +#else +#define EIGEN_HAS_STD_RESULT_OF 0 +#endif +#endif + +// Does the compiler support variadic templates? +#ifndef EIGEN_HAS_VARIADIC_TEMPLATES +#if EIGEN_MAX_CPP_VER>=11 && (__cplusplus > 199711L || EIGEN_COMP_MSVC >= 1900) \ + && ( !defined(__NVCC__) || !EIGEN_ARCH_ARM_OR_ARM64 || (defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000) ) + // ^^ Disable the use of variadic templates when compiling with versions of nvcc older than 8.0 on ARM devices: + // this prevents nvcc from crashing when compiling Eigen on Tegra X1 +#define EIGEN_HAS_VARIADIC_TEMPLATES 1 +#else +#define EIGEN_HAS_VARIADIC_TEMPLATES 0 +#endif +#endif + +// Does the compiler fully support const expressions? (as in c++14) +#ifndef EIGEN_HAS_CONSTEXPR + +#ifdef __CUDACC__ +// Const expressions are supported provided that c++11 is enabled and we're using either clang or nvcc 7.5 or above +#if EIGEN_MAX_CPP_VER>=14 && (__cplusplus > 199711L && defined(__CUDACC_VER__) && (EIGEN_COMP_CLANG || __CUDACC_VER__ >= 70500)) + #define EIGEN_HAS_CONSTEXPR 1 +#endif +#elif EIGEN_MAX_CPP_VER>=14 && (__has_feature(cxx_relaxed_constexpr) || (defined(__cplusplus) && __cplusplus >= 201402L) || \ + (EIGEN_GNUC_AT_LEAST(4,8) && (__cplusplus > 199711L))) +#define EIGEN_HAS_CONSTEXPR 1 +#endif + +#ifndef EIGEN_HAS_CONSTEXPR +#define EIGEN_HAS_CONSTEXPR 0 +#endif + +#endif + +// Does the compiler support C++11 math? +// Let's be conservative and enable the default C++11 implementation only if we are sure it exists +#ifndef EIGEN_HAS_CXX11_MATH + #if EIGEN_MAX_CPP_VER>=11 && ((__cplusplus > 201103L) || (__cplusplus >= 201103L) && (EIGEN_COMP_GNUC_STRICT || EIGEN_COMP_CLANG || EIGEN_COMP_MSVC || EIGEN_COMP_ICC) \ + && (EIGEN_ARCH_i386_OR_x86_64) && (EIGEN_OS_GNULINUX || EIGEN_OS_WIN_STRICT || EIGEN_OS_MAC)) + #define EIGEN_HAS_CXX11_MATH 1 + #else + #define EIGEN_HAS_CXX11_MATH 0 + #endif +#endif + +// Does the compiler support proper C++11 containers? +#ifndef EIGEN_HAS_CXX11_CONTAINERS + #if EIGEN_MAX_CPP_VER>=11 && \ + ((__cplusplus > 201103L) \ + || ((__cplusplus >= 201103L) && (EIGEN_COMP_GNUC_STRICT || EIGEN_COMP_CLANG || EIGEN_COMP_ICC>=1400)) \ + || EIGEN_COMP_MSVC >= 1900) + #define EIGEN_HAS_CXX11_CONTAINERS 1 + #else + #define EIGEN_HAS_CXX11_CONTAINERS 0 + #endif +#endif + +// Does the compiler support C++11 noexcept? +#ifndef EIGEN_HAS_CXX11_NOEXCEPT + #if EIGEN_MAX_CPP_VER>=11 && \ + (__has_feature(cxx_noexcept) \ + || (__cplusplus > 201103L) \ + || ((__cplusplus >= 201103L) && (EIGEN_COMP_GNUC_STRICT || EIGEN_COMP_CLANG || EIGEN_COMP_ICC>=1400)) \ + || EIGEN_COMP_MSVC >= 1900) + #define EIGEN_HAS_CXX11_NOEXCEPT 1 + #else + #define EIGEN_HAS_CXX11_NOEXCEPT 0 + #endif +#endif + +/** Allows to disable some optimizations which might affect the accuracy of the result. + * Such optimization are enabled by default, and set EIGEN_FAST_MATH to 0 to disable them. + * They currently include: + * - single precision ArrayBase::sin() and ArrayBase::cos() for SSE and AVX vectorization. + */ +#ifndef EIGEN_FAST_MATH +#define EIGEN_FAST_MATH 1 +#endif + +#define EIGEN_DEBUG_VAR(x) std::cerr << #x << " = " << x << std::endl; + +// concatenate two tokens +#define EIGEN_CAT2(a,b) a ## b +#define EIGEN_CAT(a,b) EIGEN_CAT2(a,b) + +#define EIGEN_COMMA , + +// convert a token to a string +#define EIGEN_MAKESTRING2(a) #a +#define EIGEN_MAKESTRING(a) EIGEN_MAKESTRING2(a) + +// EIGEN_STRONG_INLINE is a stronger version of the inline, using __forceinline on MSVC, +// but it still doesn't use GCC's always_inline. This is useful in (common) situations where MSVC needs forceinline +// but GCC is still doing fine with just inline. +#if EIGEN_COMP_MSVC || EIGEN_COMP_ICC +#define EIGEN_STRONG_INLINE __forceinline +#else +#define EIGEN_STRONG_INLINE inline +#endif + +// EIGEN_ALWAYS_INLINE is the stronget, it has the effect of making the function inline and adding every possible +// attribute to maximize inlining. This should only be used when really necessary: in particular, +// it uses __attribute__((always_inline)) on GCC, which most of the time is useless and can severely harm compile times. +// FIXME with the always_inline attribute, +// gcc 3.4.x and 4.1 reports the following compilation error: +// Eval.h:91: sorry, unimplemented: inlining failed in call to 'const Eigen::Eval<Derived> Eigen::MatrixBase<Scalar, Derived>::eval() const' +// : function body not available +// See also bug 1367 +#if EIGEN_GNUC_AT_LEAST(4,2) +#define EIGEN_ALWAYS_INLINE __attribute__((always_inline)) inline +#else +#define EIGEN_ALWAYS_INLINE EIGEN_STRONG_INLINE +#endif + +#if EIGEN_COMP_GNUC +#define EIGEN_DONT_INLINE __attribute__((noinline)) +#elif EIGEN_COMP_MSVC +#define EIGEN_DONT_INLINE __declspec(noinline) +#else +#define EIGEN_DONT_INLINE +#endif + +#if EIGEN_COMP_GNUC +#define EIGEN_PERMISSIVE_EXPR __extension__ +#else +#define EIGEN_PERMISSIVE_EXPR +#endif + +// this macro allows to get rid of linking errors about multiply defined functions. +// - static is not very good because it prevents definitions from different object files to be merged. +// So static causes the resulting linked executable to be bloated with multiple copies of the same function. +// - inline is not perfect either as it unwantedly hints the compiler toward inlining the function. +#define EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +#define EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS inline + +#ifdef NDEBUG +# ifndef EIGEN_NO_DEBUG +# define EIGEN_NO_DEBUG +# endif +#endif + +// eigen_plain_assert is where we implement the workaround for the assert() bug in GCC <= 4.3, see bug 89 +#ifdef EIGEN_NO_DEBUG + #define eigen_plain_assert(x) +#else + #if EIGEN_SAFE_TO_USE_STANDARD_ASSERT_MACRO + namespace Eigen { + namespace internal { + inline bool copy_bool(bool b) { return b; } + } + } + #define eigen_plain_assert(x) assert(x) + #else + // work around bug 89 + #include <cstdlib> // for abort + #include <iostream> // for std::cerr + + namespace Eigen { + namespace internal { + // trivial function copying a bool. Must be EIGEN_DONT_INLINE, so we implement it after including Eigen headers. + // see bug 89. + namespace { + EIGEN_DONT_INLINE bool copy_bool(bool b) { return b; } + } + inline void assert_fail(const char *condition, const char *function, const char *file, int line) + { + std::cerr << "assertion failed: " << condition << " in function " << function << " at " << file << ":" << line << std::endl; + abort(); + } + } + } + #define eigen_plain_assert(x) \ + do { \ + if(!Eigen::internal::copy_bool(x)) \ + Eigen::internal::assert_fail(EIGEN_MAKESTRING(x), __PRETTY_FUNCTION__, __FILE__, __LINE__); \ + } while(false) + #endif +#endif + +// eigen_assert can be overridden +#ifndef eigen_assert +#define eigen_assert(x) eigen_plain_assert(x) +#endif + +#ifdef EIGEN_INTERNAL_DEBUGGING +#define eigen_internal_assert(x) eigen_assert(x) +#else +#define eigen_internal_assert(x) +#endif + +#ifdef EIGEN_NO_DEBUG +#define EIGEN_ONLY_USED_FOR_DEBUG(x) EIGEN_UNUSED_VARIABLE(x) +#else +#define EIGEN_ONLY_USED_FOR_DEBUG(x) +#endif + +#ifndef EIGEN_NO_DEPRECATED_WARNING + #if EIGEN_COMP_GNUC + #define EIGEN_DEPRECATED __attribute__((deprecated)) + #elif EIGEN_COMP_MSVC + #define EIGEN_DEPRECATED __declspec(deprecated) + #else + #define EIGEN_DEPRECATED + #endif +#else + #define EIGEN_DEPRECATED +#endif + +#if EIGEN_COMP_GNUC +#define EIGEN_UNUSED __attribute__((unused)) +#else +#define EIGEN_UNUSED +#endif + +// Suppresses 'unused variable' warnings. +namespace Eigen { + namespace internal { + template<typename T> EIGEN_DEVICE_FUNC void ignore_unused_variable(const T&) {} + } +} +#define EIGEN_UNUSED_VARIABLE(var) Eigen::internal::ignore_unused_variable(var); + +#if !defined(EIGEN_ASM_COMMENT) + #if EIGEN_COMP_GNUC && (EIGEN_ARCH_i386_OR_x86_64 || EIGEN_ARCH_ARM_OR_ARM64) + #define EIGEN_ASM_COMMENT(X) __asm__("#" X) + #else + #define EIGEN_ASM_COMMENT(X) + #endif +#endif + + +//------------------------------------------------------------------------------------------ +// Static and dynamic alignment control +// +// The main purpose of this section is to define EIGEN_MAX_ALIGN_BYTES and EIGEN_MAX_STATIC_ALIGN_BYTES +// as the maximal boundary in bytes on which dynamically and statically allocated data may be alignment respectively. +// The values of EIGEN_MAX_ALIGN_BYTES and EIGEN_MAX_STATIC_ALIGN_BYTES can be specified by the user. If not, +// a default value is automatically computed based on architecture, compiler, and OS. +// +// This section also defines macros EIGEN_ALIGN_TO_BOUNDARY(N) and the shortcuts EIGEN_ALIGN{8,16,32,_MAX} +// to be used to declare statically aligned buffers. +//------------------------------------------------------------------------------------------ + + +/* EIGEN_ALIGN_TO_BOUNDARY(n) forces data to be n-byte aligned. This is used to satisfy SIMD requirements. + * However, we do that EVEN if vectorization (EIGEN_VECTORIZE) is disabled, + * so that vectorization doesn't affect binary compatibility. + * + * If we made alignment depend on whether or not EIGEN_VECTORIZE is defined, it would be impossible to link + * vectorized and non-vectorized code. + */ +#if (defined __CUDACC__) + #define EIGEN_ALIGN_TO_BOUNDARY(n) __align__(n) +#elif EIGEN_COMP_GNUC || EIGEN_COMP_PGI || EIGEN_COMP_IBM || EIGEN_COMP_ARM + #define EIGEN_ALIGN_TO_BOUNDARY(n) __attribute__((aligned(n))) +#elif EIGEN_COMP_MSVC + #define EIGEN_ALIGN_TO_BOUNDARY(n) __declspec(align(n)) +#elif EIGEN_COMP_SUNCC + // FIXME not sure about this one: + #define EIGEN_ALIGN_TO_BOUNDARY(n) __attribute__((aligned(n))) +#else + #error Please tell me what is the equivalent of __attribute__((aligned(n))) 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 +#elif defined(EIGEN_VECTORIZE_AVX512) + // 64 bytes static alignmeent is preferred only if really required + #define EIGEN_IDEAL_MAX_ALIGN_BYTES 64 +#elif defined(__AVX__) + // 32 bytes static alignmeent is preferred only if really required + #define EIGEN_IDEAL_MAX_ALIGN_BYTES 32 +#else + #define EIGEN_IDEAL_MAX_ALIGN_BYTES 16 +#endif + + +// EIGEN_MIN_ALIGN_BYTES defines the minimal value for which the notion of explicit alignment makes sense +#define EIGEN_MIN_ALIGN_BYTES 16 + +// Defined the boundary (in bytes) on which the data needs to be aligned. Note +// that unless EIGEN_ALIGN is defined and not equal to 0, the data may not be +// aligned at all regardless of the value of this #define. + +#if (defined(EIGEN_DONT_ALIGN_STATICALLY) || defined(EIGEN_DONT_ALIGN)) && defined(EIGEN_MAX_STATIC_ALIGN_BYTES) && EIGEN_MAX_STATIC_ALIGN_BYTES>0 +#error EIGEN_MAX_STATIC_ALIGN_BYTES and EIGEN_DONT_ALIGN[_STATICALLY] are both defined with EIGEN_MAX_STATIC_ALIGN_BYTES!=0. Use EIGEN_MAX_STATIC_ALIGN_BYTES=0 as a synonym of EIGEN_DONT_ALIGN_STATICALLY. +#endif + +// EIGEN_DONT_ALIGN_STATICALLY and EIGEN_DONT_ALIGN are deprectated +// They imply EIGEN_MAX_STATIC_ALIGN_BYTES=0 +#if defined(EIGEN_DONT_ALIGN_STATICALLY) || defined(EIGEN_DONT_ALIGN) + #ifdef EIGEN_MAX_STATIC_ALIGN_BYTES + #undef EIGEN_MAX_STATIC_ALIGN_BYTES + #endif + #define EIGEN_MAX_STATIC_ALIGN_BYTES 0 +#endif + +#ifndef EIGEN_MAX_STATIC_ALIGN_BYTES + + // Try to automatically guess what is the best default value for EIGEN_MAX_STATIC_ALIGN_BYTES + + // 16 byte alignment is only useful for vectorization. Since it affects the ABI, we need to enable + // 16 byte alignment on all platforms where vectorization might be enabled. In theory we could always + // enable alignment, but it can be a cause of problems on some platforms, so we just disable it in + // certain common platform (compiler+architecture combinations) to avoid these problems. + // Only static alignment is really problematic (relies on nonstandard compiler extensions), + // try to keep heap alignment even when we have to disable static alignment. + #if EIGEN_COMP_GNUC && !(EIGEN_ARCH_i386_OR_x86_64 || EIGEN_ARCH_ARM_OR_ARM64 || EIGEN_ARCH_PPC || EIGEN_ARCH_IA64) + #define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 1 + #elif EIGEN_ARCH_ARM_OR_ARM64 && EIGEN_COMP_GNUC_STRICT && EIGEN_GNUC_AT_MOST(4, 6) + // Old versions of GCC on ARM, at least 4.4, were once seen to have buggy static alignment support. + // Not sure which version fixed it, hopefully it doesn't affect 4.7, which is still somewhat in use. + // 4.8 and newer seem definitely unaffected. + #define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 1 + #else + #define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 0 + #endif + + // static alignment is completely disabled with GCC 3, Sun Studio, and QCC/QNX + #if !EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT \ + && !EIGEN_GCC3_OR_OLDER \ + && !EIGEN_COMP_SUNCC \ + && !EIGEN_OS_QNX + #define EIGEN_ARCH_WANTS_STACK_ALIGNMENT 1 + #else + #define EIGEN_ARCH_WANTS_STACK_ALIGNMENT 0 + #endif + + #if EIGEN_ARCH_WANTS_STACK_ALIGNMENT + #define EIGEN_MAX_STATIC_ALIGN_BYTES EIGEN_IDEAL_MAX_ALIGN_BYTES + #else + #define EIGEN_MAX_STATIC_ALIGN_BYTES 0 + #endif + +#endif + +// If EIGEN_MAX_ALIGN_BYTES is defined, then it is considered as an upper bound for EIGEN_MAX_ALIGN_BYTES +#if defined(EIGEN_MAX_ALIGN_BYTES) && EIGEN_MAX_ALIGN_BYTES<EIGEN_MAX_STATIC_ALIGN_BYTES +#undef EIGEN_MAX_STATIC_ALIGN_BYTES +#define EIGEN_MAX_STATIC_ALIGN_BYTES EIGEN_MAX_ALIGN_BYTES +#endif + +#if EIGEN_MAX_STATIC_ALIGN_BYTES==0 && !defined(EIGEN_DISABLE_UNALIGNED_ARRAY_ASSERT) + #define EIGEN_DISABLE_UNALIGNED_ARRAY_ASSERT +#endif + +// At this stage, EIGEN_MAX_STATIC_ALIGN_BYTES>0 is the true test whether we want to align arrays on the stack or not. +// It takes into account both the user choice to explicitly enable/disable alignment (by settting EIGEN_MAX_STATIC_ALIGN_BYTES) +// and the architecture config (EIGEN_ARCH_WANTS_STACK_ALIGNMENT). +// Henceforth, only EIGEN_MAX_STATIC_ALIGN_BYTES should be used. + + +// Shortcuts to EIGEN_ALIGN_TO_BOUNDARY +#define EIGEN_ALIGN8 EIGEN_ALIGN_TO_BOUNDARY(8) +#define EIGEN_ALIGN16 EIGEN_ALIGN_TO_BOUNDARY(16) +#define EIGEN_ALIGN32 EIGEN_ALIGN_TO_BOUNDARY(32) +#define EIGEN_ALIGN64 EIGEN_ALIGN_TO_BOUNDARY(64) +#if EIGEN_MAX_STATIC_ALIGN_BYTES>0 +#define EIGEN_ALIGN_MAX EIGEN_ALIGN_TO_BOUNDARY(EIGEN_MAX_STATIC_ALIGN_BYTES) +#else +#define EIGEN_ALIGN_MAX +#endif + + +// Dynamic alignment control + +#if defined(EIGEN_DONT_ALIGN) && defined(EIGEN_MAX_ALIGN_BYTES) && EIGEN_MAX_ALIGN_BYTES>0 +#error EIGEN_MAX_ALIGN_BYTES and EIGEN_DONT_ALIGN are both defined with EIGEN_MAX_ALIGN_BYTES!=0. Use EIGEN_MAX_ALIGN_BYTES=0 as a synonym of EIGEN_DONT_ALIGN. +#endif + +#ifdef EIGEN_DONT_ALIGN + #ifdef EIGEN_MAX_ALIGN_BYTES + #undef EIGEN_MAX_ALIGN_BYTES + #endif + #define EIGEN_MAX_ALIGN_BYTES 0 +#elif !defined(EIGEN_MAX_ALIGN_BYTES) + #define EIGEN_MAX_ALIGN_BYTES EIGEN_IDEAL_MAX_ALIGN_BYTES +#endif + +#if EIGEN_IDEAL_MAX_ALIGN_BYTES > EIGEN_MAX_ALIGN_BYTES +#define EIGEN_DEFAULT_ALIGN_BYTES EIGEN_IDEAL_MAX_ALIGN_BYTES +#else +#define EIGEN_DEFAULT_ALIGN_BYTES EIGEN_MAX_ALIGN_BYTES +#endif + + +#ifndef EIGEN_UNALIGNED_VECTORIZE +#define EIGEN_UNALIGNED_VECTORIZE 1 +#endif + +//---------------------------------------------------------------------- + + +#ifdef EIGEN_DONT_USE_RESTRICT_KEYWORD + #define EIGEN_RESTRICT +#endif +#ifndef EIGEN_RESTRICT + #define EIGEN_RESTRICT __restrict +#endif + +#ifndef EIGEN_STACK_ALLOCATION_LIMIT +// 131072 == 128 KB +#define EIGEN_STACK_ALLOCATION_LIMIT 131072 +#endif + +#ifndef EIGEN_DEFAULT_IO_FORMAT +#ifdef EIGEN_MAKING_DOCS +// format used in Eigen's documentation +// needed to define it here as escaping characters in CMake add_definition's argument seems very problematic. +#define EIGEN_DEFAULT_IO_FORMAT Eigen::IOFormat(3, 0, " ", "\n", "", "") +#else +#define EIGEN_DEFAULT_IO_FORMAT Eigen::IOFormat() +#endif +#endif + +// just an empty macro ! +#define EIGEN_EMPTY + +#if EIGEN_COMP_MSVC_STRICT && (EIGEN_COMP_MSVC < 1900 || defined(__CUDACC_VER__)) // for older MSVC versions, as well as 1900 && CUDA 8, using the base operator is sufficient (cf Bugs 1000, 1324) + #define EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived) \ + using Base::operator =; +#elif EIGEN_COMP_CLANG // workaround clang bug (see http://forum.kde.org/viewtopic.php?f=74&t=102653) + #define EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived) \ + using Base::operator =; \ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& operator=(const Derived& other) { Base::operator=(other); return *this; } \ + template <typename OtherDerived> \ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& operator=(const DenseBase<OtherDerived>& other) { Base::operator=(other.derived()); return *this; } +#else + #define EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived) \ + using Base::operator =; \ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& operator=(const Derived& other) \ + { \ + Base::operator=(other); \ + return *this; \ + } +#endif + + +/** \internal + * \brief Macro to manually inherit assignment operators. + * This is necessary, because the implicitly defined assignment operator gets deleted when a custom operator= is defined. + */ +#define EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Derived) EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived) + +/** +* Just a side note. Commenting within defines works only by documenting +* behind the object (via '!<'). Comments cannot be multi-line and thus +* we have these extra long lines. What is confusing doxygen over here is +* that we use '\' and basically have a bunch of typedefs with their +* documentation in a single line. +**/ + +#define EIGEN_GENERIC_PUBLIC_INTERFACE(Derived) \ + typedef typename Eigen::internal::traits<Derived>::Scalar Scalar; /*!< \brief Numeric type, e.g. float, double, int or std::complex<float>. */ \ + typedef typename Eigen::NumTraits<Scalar>::Real RealScalar; /*!< \brief The underlying numeric type for composed scalar types. \details In cases where Scalar is e.g. std::complex<T>, T were corresponding to RealScalar. */ \ + typedef typename Base::CoeffReturnType CoeffReturnType; /*!< \brief The return type for coefficient access. \details Depending on whether the object allows direct coefficient access (e.g. for a MatrixXd), this type is either 'const Scalar&' or simply 'Scalar' for objects that do not allow direct coefficient access. */ \ + typedef typename Eigen::internal::ref_selector<Derived>::type Nested; \ + typedef typename Eigen::internal::traits<Derived>::StorageKind StorageKind; \ + typedef typename Eigen::internal::traits<Derived>::StorageIndex StorageIndex; \ + enum { RowsAtCompileTime = Eigen::internal::traits<Derived>::RowsAtCompileTime, \ + ColsAtCompileTime = Eigen::internal::traits<Derived>::ColsAtCompileTime, \ + Flags = Eigen::internal::traits<Derived>::Flags, \ + SizeAtCompileTime = Base::SizeAtCompileTime, \ + MaxSizeAtCompileTime = Base::MaxSizeAtCompileTime, \ + IsVectorAtCompileTime = Base::IsVectorAtCompileTime }; \ + using Base::derived; \ + using Base::const_cast_derived; + + +// FIXME Maybe the EIGEN_DENSE_PUBLIC_INTERFACE could be removed as importing PacketScalar is rarely needed +#define EIGEN_DENSE_PUBLIC_INTERFACE(Derived) \ + EIGEN_GENERIC_PUBLIC_INTERFACE(Derived) \ + typedef typename Base::PacketScalar PacketScalar; + + +#define EIGEN_PLAIN_ENUM_MIN(a,b) (((int)a <= (int)b) ? (int)a : (int)b) +#define EIGEN_PLAIN_ENUM_MAX(a,b) (((int)a >= (int)b) ? (int)a : (int)b) + +// EIGEN_SIZE_MIN_PREFER_DYNAMIC gives the min between compile-time sizes. 0 has absolute priority, followed by 1, +// followed by Dynamic, followed by other finite values. The reason for giving Dynamic the priority over +// finite values is that min(3, Dynamic) should be Dynamic, since that could be anything between 0 and 3. +#define EIGEN_SIZE_MIN_PREFER_DYNAMIC(a,b) (((int)a == 0 || (int)b == 0) ? 0 \ + : ((int)a == 1 || (int)b == 1) ? 1 \ + : ((int)a == Dynamic || (int)b == Dynamic) ? Dynamic \ + : ((int)a <= (int)b) ? (int)a : (int)b) + +// EIGEN_SIZE_MIN_PREFER_FIXED is a variant of EIGEN_SIZE_MIN_PREFER_DYNAMIC comparing MaxSizes. The difference is that finite values +// now have priority over Dynamic, so that min(3, Dynamic) gives 3. Indeed, whatever the actual value is +// (between 0 and 3), it is not more than 3. +#define EIGEN_SIZE_MIN_PREFER_FIXED(a,b) (((int)a == 0 || (int)b == 0) ? 0 \ + : ((int)a == 1 || (int)b == 1) ? 1 \ + : ((int)a == Dynamic && (int)b == Dynamic) ? Dynamic \ + : ((int)a == Dynamic) ? (int)b \ + : ((int)b == Dynamic) ? (int)a \ + : ((int)a <= (int)b) ? (int)a : (int)b) + +// see EIGEN_SIZE_MIN_PREFER_DYNAMIC. No need for a separate variant for MaxSizes here. +#define EIGEN_SIZE_MAX(a,b) (((int)a == Dynamic || (int)b == Dynamic) ? Dynamic \ + : ((int)a >= (int)b) ? (int)a : (int)b) + +#define EIGEN_LOGICAL_XOR(a,b) (((a) || (b)) && !((a) && (b))) + +#define EIGEN_IMPLIES(a,b) (!(a) || (b)) + +// the expression type of a standard coefficient wise binary operation +#define EIGEN_CWISE_BINARY_RETURN_TYPE(LHS,RHS,OPNAME) \ + CwiseBinaryOp< \ + EIGEN_CAT(EIGEN_CAT(internal::scalar_,OPNAME),_op)< \ + typename internal::traits<LHS>::Scalar, \ + typename internal::traits<RHS>::Scalar \ + >, \ + const LHS, \ + const RHS \ + > + +#define EIGEN_MAKE_CWISE_BINARY_OP(METHOD,OPNAME) \ + template<typename OtherDerived> \ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const EIGEN_CWISE_BINARY_RETURN_TYPE(Derived,OtherDerived,OPNAME) \ + (METHOD)(const EIGEN_CURRENT_STORAGE_BASE_CLASS<OtherDerived> &other) const \ + { \ + return EIGEN_CWISE_BINARY_RETURN_TYPE(Derived,OtherDerived,OPNAME)(derived(), other.derived()); \ + } + +#define EIGEN_SCALAR_BINARY_SUPPORTED(OPNAME,TYPEA,TYPEB) \ + (Eigen::internal::has_ReturnType<Eigen::ScalarBinaryOpTraits<TYPEA,TYPEB,EIGEN_CAT(EIGEN_CAT(Eigen::internal::scalar_,OPNAME),_op)<TYPEA,TYPEB> > >::value) + +#define EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(EXPR,SCALAR,OPNAME) \ + CwiseBinaryOp<EIGEN_CAT(EIGEN_CAT(internal::scalar_,OPNAME),_op)<typename internal::traits<EXPR>::Scalar,SCALAR>, const EXPR, \ + const typename internal::plain_constant_type<EXPR,SCALAR>::type> + +#define EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(SCALAR,EXPR,OPNAME) \ + CwiseBinaryOp<EIGEN_CAT(EIGEN_CAT(internal::scalar_,OPNAME),_op)<SCALAR,typename internal::traits<EXPR>::Scalar>, \ + const typename internal::plain_constant_type<EXPR,SCALAR>::type, const EXPR> + +// Workaround for MSVC 2010 (see ML thread "patch with compile for for MSVC 2010") +#if EIGEN_COMP_MSVC_STRICT<=1600 +#define EIGEN_MSVC10_WORKAROUND_BINARYOP_RETURN_TYPE(X) typename internal::enable_if<true,X>::type +#else +#define EIGEN_MSVC10_WORKAROUND_BINARYOP_RETURN_TYPE(X) X +#endif + +#define EIGEN_MAKE_SCALAR_BINARY_OP_ONTHERIGHT(METHOD,OPNAME) \ + template <typename T> EIGEN_DEVICE_FUNC 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; \ + return EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(Derived,PromotedT,OPNAME)(derived(), \ + typename internal::plain_constant_type<Derived,PromotedT>::type(derived().rows(), derived().cols(), internal::scalar_constant_op<PromotedT>(scalar))); \ + } + +#define EIGEN_MAKE_SCALAR_BINARY_OP_ONTHELEFT(METHOD,OPNAME) \ + template <typename T> EIGEN_DEVICE_FUNC 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; \ + return EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(PromotedT,Derived,OPNAME)( \ + typename internal::plain_constant_type<Derived,PromotedT>::type(matrix.derived().rows(), matrix.derived().cols(), internal::scalar_constant_op<PromotedT>(scalar)), matrix.derived()); \ + } + +#define EIGEN_MAKE_SCALAR_BINARY_OP(METHOD,OPNAME) \ + EIGEN_MAKE_SCALAR_BINARY_OP_ONTHELEFT(METHOD,OPNAME) \ + EIGEN_MAKE_SCALAR_BINARY_OP_ONTHERIGHT(METHOD,OPNAME) + + +#ifdef EIGEN_EXCEPTIONS +# define EIGEN_THROW_X(X) throw X +# define EIGEN_THROW throw +# define EIGEN_TRY try +# define EIGEN_CATCH(X) catch (X) +#else +# ifdef __CUDA_ARCH__ +# define EIGEN_THROW_X(X) asm("trap;") +# define EIGEN_THROW asm("trap;") +# else +# define EIGEN_THROW_X(X) std::abort() +# define EIGEN_THROW std::abort() +# endif +# define EIGEN_TRY if (true) +# define EIGEN_CATCH(X) else +#endif + + +#if EIGEN_HAS_CXX11_NOEXCEPT +# define EIGEN_INCLUDE_TYPE_TRAITS +# define EIGEN_NOEXCEPT noexcept +# define EIGEN_NOEXCEPT_IF(x) noexcept(x) +# define EIGEN_NO_THROW noexcept(true) +# define EIGEN_EXCEPTION_SPEC(X) noexcept(false) +#else +# define EIGEN_NOEXCEPT +# define EIGEN_NOEXCEPT_IF(x) +# define EIGEN_NO_THROW throw() +# define EIGEN_EXCEPTION_SPEC(X) throw(X) +#endif + +#endif // EIGEN_MACROS_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/Memory.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/Memory.h new file mode 100644 index 000000000..c634d7ea0 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/Memory.h @@ -0,0 +1,977 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2015 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2008-2009 Benoit Jacob <jacob.benoit.1@gmail.com> +// Copyright (C) 2009 Kenneth Riddile <kfriddile@yahoo.com> +// Copyright (C) 2010 Hauke Heibel <hauke.heibel@gmail.com> +// Copyright (C) 2010 Thomas Capricelli <orzel@freehackers.org> +// Copyright (C) 2013 Pavel Holoborodko <pavel@holoborodko.com> +// +// 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/. + + +/***************************************************************************** +*** Platform checks for aligned malloc functions *** +*****************************************************************************/ + +#ifndef EIGEN_MEMORY_H +#define EIGEN_MEMORY_H + +#ifndef EIGEN_MALLOC_ALREADY_ALIGNED + +// Try to determine automatically if malloc is already aligned. + +// On 64-bit systems, glibc's malloc returns 16-byte-aligned pointers, see: +// http://www.gnu.org/s/libc/manual/html_node/Aligned-Memory-Blocks.html +// This is true at least since glibc 2.8. +// This leaves the question how to detect 64-bit. According to this document, +// http://gcc.fyxm.net/summit/2003/Porting%20to%2064%20bit.pdf +// page 114, "[The] LP64 model [...] is used by all 64-bit UNIX ports" so it's indeed +// quite safe, at least within the context of glibc, to equate 64-bit with LP64. +#if defined(__GLIBC__) && ((__GLIBC__>=2 && __GLIBC_MINOR__ >= 8) || __GLIBC__>2) \ + && defined(__LP64__) && ! defined( __SANITIZE_ADDRESS__ ) && (EIGEN_DEFAULT_ALIGN_BYTES == 16) + #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 1 +#else + #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 0 +#endif + +// FreeBSD 6 seems to have 16-byte aligned malloc +// See http://svn.freebsd.org/viewvc/base/stable/6/lib/libc/stdlib/malloc.c?view=markup +// FreeBSD 7 seems to have 16-byte aligned malloc except on ARM and MIPS architectures +// See http://svn.freebsd.org/viewvc/base/stable/7/lib/libc/stdlib/malloc.c?view=markup +#if defined(__FreeBSD__) && !(EIGEN_ARCH_ARM || EIGEN_ARCH_MIPS) && (EIGEN_DEFAULT_ALIGN_BYTES == 16) + #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 1 +#else + #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 0 +#endif + +#if (EIGEN_OS_MAC && (EIGEN_DEFAULT_ALIGN_BYTES == 16)) \ + || (EIGEN_OS_WIN64 && (EIGEN_DEFAULT_ALIGN_BYTES == 16)) \ + || EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED \ + || EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED + #define EIGEN_MALLOC_ALREADY_ALIGNED 1 +#else + #define EIGEN_MALLOC_ALREADY_ALIGNED 0 +#endif + +#endif + +namespace Eigen { + +namespace internal { + +EIGEN_DEVICE_FUNC +inline void throw_std_bad_alloc() +{ + #ifdef EIGEN_EXCEPTIONS + throw std::bad_alloc(); + #else + std::size_t huge = static_cast<std::size_t>(-1); + new int[huge]; + #endif +} + +/***************************************************************************** +*** Implementation of handmade aligned functions *** +*****************************************************************************/ + +/* ----- Hand made implementations of aligned malloc/free and realloc ----- */ + +/** \internal Like malloc, but the returned pointer is guaranteed to be 16-byte aligned. + * Fast, but wastes 16 additional bytes of memory. Does not throw any exception. + */ +inline void* handmade_aligned_malloc(std::size_t size) +{ + void *original = std::malloc(size+EIGEN_DEFAULT_ALIGN_BYTES); + if (original == 0) return 0; + void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1))) + EIGEN_DEFAULT_ALIGN_BYTES); + *(reinterpret_cast<void**>(aligned) - 1) = original; + return aligned; +} + +/** \internal Frees memory allocated with handmade_aligned_malloc */ +inline void handmade_aligned_free(void *ptr) +{ + if (ptr) std::free(*(reinterpret_cast<void**>(ptr) - 1)); +} + +/** \internal + * \brief Reallocates aligned memory. + * Since we know that our handmade version is based on std::malloc + * we can use std::realloc to implement efficient reallocation. + */ +inline void* handmade_aligned_realloc(void* ptr, std::size_t size, std::size_t = 0) +{ + if (ptr == 0) return handmade_aligned_malloc(size); + void *original = *(reinterpret_cast<void**>(ptr) - 1); + std::ptrdiff_t previous_offset = static_cast<char *>(ptr)-static_cast<char *>(original); + original = std::realloc(original,size+EIGEN_DEFAULT_ALIGN_BYTES); + if (original == 0) return 0; + void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1))) + EIGEN_DEFAULT_ALIGN_BYTES); + void *previous_aligned = static_cast<char *>(original)+previous_offset; + if(aligned!=previous_aligned) + std::memmove(aligned, previous_aligned, size); + + *(reinterpret_cast<void**>(aligned) - 1) = original; + return aligned; +} + +/***************************************************************************** +*** Implementation of portable aligned versions of malloc/free/realloc *** +*****************************************************************************/ + +#ifdef EIGEN_NO_MALLOC +EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed() +{ + eigen_assert(false && "heap allocation is forbidden (EIGEN_NO_MALLOC is defined)"); +} +#elif defined EIGEN_RUNTIME_NO_MALLOC +EIGEN_DEVICE_FUNC inline bool is_malloc_allowed_impl(bool update, bool new_value = false) +{ + static bool value = true; + if (update == 1) + value = new_value; + return value; +} +EIGEN_DEVICE_FUNC inline bool is_malloc_allowed() { return is_malloc_allowed_impl(false); } +EIGEN_DEVICE_FUNC inline bool set_is_malloc_allowed(bool new_value) { return is_malloc_allowed_impl(true, new_value); } +EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed() +{ + eigen_assert(is_malloc_allowed() && "heap allocation is forbidden (EIGEN_RUNTIME_NO_MALLOC is defined and g_is_malloc_allowed is false)"); +} +#else +EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed() +{} +#endif + +/** \internal Allocates \a size bytes. The returned pointer is guaranteed to have 16 or 32 bytes alignment depending on the requirements. + * On allocation error, the returned pointer is null, and std::bad_alloc is thrown. + */ +EIGEN_DEVICE_FUNC inline void* aligned_malloc(std::size_t size) +{ + check_that_malloc_is_allowed(); + + void *result; + #if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED + result = std::malloc(size); + #if EIGEN_DEFAULT_ALIGN_BYTES==16 + eigen_assert((size<16 || (std::size_t(result)%16)==0) && "System's malloc returned an unaligned pointer. Compile with EIGEN_MALLOC_ALREADY_ALIGNED=0 to fallback to handmade alignd memory allocator."); + #endif + #else + result = handmade_aligned_malloc(size); + #endif + + if(!result && size) + throw_std_bad_alloc(); + + return result; +} + +/** \internal Frees memory allocated with aligned_malloc. */ +EIGEN_DEVICE_FUNC inline void aligned_free(void *ptr) +{ + #if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED + std::free(ptr); + #else + handmade_aligned_free(ptr); + #endif +} + +/** + * \internal + * \brief Reallocates an aligned block of memory. + * \throws std::bad_alloc on allocation failure + */ +inline void* aligned_realloc(void *ptr, std::size_t new_size, std::size_t old_size) +{ + EIGEN_UNUSED_VARIABLE(old_size); + + void *result; +#if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED + result = std::realloc(ptr,new_size); +#else + result = handmade_aligned_realloc(ptr,new_size,old_size); +#endif + + if (!result && new_size) + throw_std_bad_alloc(); + + return result; +} + +/***************************************************************************** +*** Implementation of conditionally aligned functions *** +*****************************************************************************/ + +/** \internal Allocates \a size bytes. If Align is true, then the returned ptr is 16-byte-aligned. + * On allocation error, the returned pointer is null, and a std::bad_alloc is thrown. + */ +template<bool Align> EIGEN_DEVICE_FUNC inline void* conditional_aligned_malloc(std::size_t size) +{ + return aligned_malloc(size); +} + +template<> EIGEN_DEVICE_FUNC inline void* conditional_aligned_malloc<false>(std::size_t size) +{ + check_that_malloc_is_allowed(); + + void *result = std::malloc(size); + if(!result && size) + throw_std_bad_alloc(); + return result; +} + +/** \internal Frees memory allocated with conditional_aligned_malloc */ +template<bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_free(void *ptr) +{ + aligned_free(ptr); +} + +template<> EIGEN_DEVICE_FUNC inline void conditional_aligned_free<false>(void *ptr) +{ + std::free(ptr); +} + +template<bool Align> inline void* conditional_aligned_realloc(void* ptr, std::size_t new_size, std::size_t old_size) +{ + return aligned_realloc(ptr, new_size, old_size); +} + +template<> inline void* conditional_aligned_realloc<false>(void* ptr, std::size_t new_size, std::size_t) +{ + return std::realloc(ptr, new_size); +} + +/***************************************************************************** +*** Construction/destruction of array elements *** +*****************************************************************************/ + +/** \internal Destructs the elements of an array. + * The \a size parameters tells on how many objects to call the destructor of T. + */ +template<typename T> EIGEN_DEVICE_FUNC inline void destruct_elements_of_array(T *ptr, std::size_t size) +{ + // always destruct an array starting from the end. + if(ptr) + while(size) ptr[--size].~T(); +} + +/** \internal Constructs the elements of an array. + * The \a size parameter tells on how many objects to call the constructor of T. + */ +template<typename T> EIGEN_DEVICE_FUNC inline T* construct_elements_of_array(T *ptr, std::size_t size) +{ + std::size_t i; + EIGEN_TRY + { + for (i = 0; i < size; ++i) ::new (ptr + i) T; + return ptr; + } + EIGEN_CATCH(...) + { + destruct_elements_of_array(ptr, i); + EIGEN_THROW; + } + return NULL; +} + +/***************************************************************************** +*** Implementation of aligned new/delete-like functions *** +*****************************************************************************/ + +template<typename T> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void check_size_for_overflow(std::size_t size) +{ + if(size > std::size_t(-1) / sizeof(T)) + throw_std_bad_alloc(); +} + +/** \internal Allocates \a size objects of type T. The returned pointer is guaranteed to have 16 bytes alignment. + * On allocation error, the returned pointer is undefined, but a std::bad_alloc is thrown. + * The default constructor of T is called. + */ +template<typename T> EIGEN_DEVICE_FUNC inline T* aligned_new(std::size_t size) +{ + check_size_for_overflow<T>(size); + T *result = reinterpret_cast<T*>(aligned_malloc(sizeof(T)*size)); + EIGEN_TRY + { + return construct_elements_of_array(result, size); + } + EIGEN_CATCH(...) + { + aligned_free(result); + EIGEN_THROW; + } + return result; +} + +template<typename T, bool Align> EIGEN_DEVICE_FUNC inline T* conditional_aligned_new(std::size_t size) +{ + check_size_for_overflow<T>(size); + T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size)); + EIGEN_TRY + { + return construct_elements_of_array(result, size); + } + EIGEN_CATCH(...) + { + conditional_aligned_free<Align>(result); + EIGEN_THROW; + } + return result; +} + +/** \internal Deletes objects constructed with aligned_new + * The \a size parameters tells on how many objects to call the destructor of T. + */ +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); +} + +/** \internal Deletes objects constructed with conditional_aligned_new + * The \a size parameters tells on how many objects to call the destructor of T. + */ +template<typename T, bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_delete(T *ptr, std::size_t size) +{ + destruct_elements_of_array<T>(ptr, size); + conditional_aligned_free<Align>(ptr); +} + +template<typename T, bool Align> EIGEN_DEVICE_FUNC inline T* conditional_aligned_realloc_new(T* pts, std::size_t new_size, std::size_t old_size) +{ + check_size_for_overflow<T>(new_size); + check_size_for_overflow<T>(old_size); + if(new_size < old_size) + destruct_elements_of_array(pts+new_size, old_size-new_size); + T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size)); + if(new_size > old_size) + { + EIGEN_TRY + { + construct_elements_of_array(result+old_size, new_size-old_size); + } + EIGEN_CATCH(...) + { + conditional_aligned_free<Align>(result); + EIGEN_THROW; + } + } + return result; +} + + +template<typename T, bool Align> EIGEN_DEVICE_FUNC inline T* conditional_aligned_new_auto(std::size_t size) +{ + if(size==0) + return 0; // short-cut. Also fixes Bug 884 + check_size_for_overflow<T>(size); + T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size)); + if(NumTraits<T>::RequireInitialization) + { + EIGEN_TRY + { + construct_elements_of_array(result, size); + } + EIGEN_CATCH(...) + { + conditional_aligned_free<Align>(result); + EIGEN_THROW; + } + } + return result; +} + +template<typename T, bool Align> inline T* conditional_aligned_realloc_new_auto(T* pts, std::size_t new_size, std::size_t old_size) +{ + check_size_for_overflow<T>(new_size); + check_size_for_overflow<T>(old_size); + if(NumTraits<T>::RequireInitialization && (new_size < old_size)) + destruct_elements_of_array(pts+new_size, old_size-new_size); + T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size)); + if(NumTraits<T>::RequireInitialization && (new_size > old_size)) + { + EIGEN_TRY + { + construct_elements_of_array(result+old_size, new_size-old_size); + } + EIGEN_CATCH(...) + { + conditional_aligned_free<Align>(result); + EIGEN_THROW; + } + } + return result; +} + +template<typename T, bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_delete_auto(T *ptr, std::size_t size) +{ + if(NumTraits<T>::RequireInitialization) + destruct_elements_of_array<T>(ptr, size); + conditional_aligned_free<Align>(ptr); +} + +/****************************************************************************/ + +/** \internal Returns the index of the first element of the array that is well aligned with respect to the requested \a Alignment. + * + * \tparam Alignment requested alignment in Bytes. + * \param array the address of the start of the array + * \param size the size of the array + * + * \note If no element of the array is well aligned or the requested alignment is not a multiple of a scalar, + * the size of the array is returned. For example with SSE, the requested alignment is typically 16-bytes. If + * packet size for the given scalar type is 1, then everything is considered well-aligned. + * + * \note Otherwise, if the Alignment is larger that the scalar size, we rely on the assumptions that sizeof(Scalar) is a + * power of 2. On the other hand, we do not assume that the array address is a multiple of sizeof(Scalar), as that fails for + * example with Scalar=double on certain 32-bit platforms, see bug #79. + * + * There is also the variant first_aligned(const MatrixBase&) defined in DenseCoeffsBase.h. + * \sa first_default_aligned() + */ +template<int Alignment, typename Scalar, typename Index> +EIGEN_DEVICE_FUNC inline Index first_aligned(const Scalar* array, Index size) +{ + const Index ScalarSize = sizeof(Scalar); + const Index AlignmentSize = Alignment / ScalarSize; + const Index AlignmentMask = AlignmentSize-1; + + if(AlignmentSize<=1) + { + // Either the requested alignment if smaller than a scalar, or it exactly match a 1 scalar + // so that all elements of the array have the same alignment. + return 0; + } + else if( (UIntPtr(array) & (sizeof(Scalar)-1)) || (Alignment%ScalarSize)!=0) + { + // The array is not aligned to the size of a single scalar, or the requested alignment is not a multiple of the scalar size. + // Consequently, no element of the array is well aligned. + return size; + } + else + { + Index first = (AlignmentSize - (Index((UIntPtr(array)/sizeof(Scalar))) & AlignmentMask)) & AlignmentMask; + return (first < size) ? first : size; + } +} + +/** \internal Returns the index of the first element of the array that is well aligned with respect the largest packet requirement. + * \sa first_aligned(Scalar*,Index) and first_default_aligned(DenseBase<Derived>) */ +template<typename Scalar, typename Index> +EIGEN_DEVICE_FUNC inline Index first_default_aligned(const Scalar* array, Index size) +{ + typedef typename packet_traits<Scalar>::type DefaultPacketType; + return first_aligned<unpacket_traits<DefaultPacketType>::alignment>(array, size); +} + +/** \internal Returns the smallest integer multiple of \a base and greater or equal to \a size + */ +template<typename Index> +inline Index first_multiple(Index size, Index base) +{ + return ((size+base-1)/base)*base; +} + +// std::copy is much slower than memcpy, so let's introduce a smart_copy which +// use memcpy on trivial types, i.e., on types that does not require an initialization ctor. +template<typename T, bool UseMemcpy> struct smart_copy_helper; + +template<typename T> EIGEN_DEVICE_FUNC void smart_copy(const T* start, const T* end, T* target) +{ + smart_copy_helper<T,!NumTraits<T>::RequireInitialization>::run(start, end, target); +} + +template<typename T> struct smart_copy_helper<T,true> { + EIGEN_DEVICE_FUNC static inline void run(const T* start, const T* end, T* target) + { + IntPtr size = IntPtr(end)-IntPtr(start); + if(size==0) return; + eigen_internal_assert(start!=0 && end!=0 && target!=0); + memcpy(target, start, size); + } +}; + +template<typename T> struct smart_copy_helper<T,false> { + EIGEN_DEVICE_FUNC static inline void run(const T* start, const T* end, T* target) + { std::copy(start, end, target); } +}; + +// intelligent memmove. falls back to std::memmove for POD types, uses std::copy otherwise. +template<typename T, bool UseMemmove> struct smart_memmove_helper; + +template<typename T> void smart_memmove(const T* start, const T* end, T* target) +{ + smart_memmove_helper<T,!NumTraits<T>::RequireInitialization>::run(start, end, target); +} + +template<typename T> struct smart_memmove_helper<T,true> { + static inline void run(const T* start, const T* end, T* target) + { + IntPtr size = IntPtr(end)-IntPtr(start); + if(size==0) return; + eigen_internal_assert(start!=0 && end!=0 && target!=0); + std::memmove(target, start, size); + } +}; + +template<typename T> struct smart_memmove_helper<T,false> { + static inline void run(const T* start, const T* end, T* target) + { + if (UIntPtr(target) < UIntPtr(start)) + { + std::copy(start, end, target); + } + else + { + std::ptrdiff_t count = (std::ptrdiff_t(end)-std::ptrdiff_t(start)) / sizeof(T); + std::copy_backward(start, end, target + count); + } + } +}; + + +/***************************************************************************** +*** Implementation of runtime stack allocation (falling back to malloc) *** +*****************************************************************************/ + +// you can overwrite Eigen's default behavior regarding alloca by defining EIGEN_ALLOCA +// to the appropriate stack allocation function +#ifndef EIGEN_ALLOCA + #if EIGEN_OS_LINUX || EIGEN_OS_MAC || (defined alloca) + #define EIGEN_ALLOCA alloca + #elif EIGEN_COMP_MSVC + #define EIGEN_ALLOCA _alloca + #endif +#endif + +// This helper class construct the allocated memory, and takes care of destructing and freeing the handled data +// at destruction time. In practice this helper class is mainly useful to avoid memory leak in case of exceptions. +template<typename T> class aligned_stack_memory_handler : noncopyable +{ + public: + /* Creates a stack_memory_handler responsible for the buffer \a ptr of size \a size. + * Note that \a ptr can be 0 regardless of the other parameters. + * This constructor takes care of constructing/initializing the elements of the buffer if required by the scalar type T (see NumTraits<T>::RequireInitialization). + * In this case, the buffer elements will also be destructed when this handler will be destructed. + * Finally, if \a dealloc is true, then the pointer \a ptr is freed. + **/ + aligned_stack_memory_handler(T* ptr, std::size_t size, bool dealloc) + : m_ptr(ptr), m_size(size), m_deallocate(dealloc) + { + if(NumTraits<T>::RequireInitialization && m_ptr) + Eigen::internal::construct_elements_of_array(m_ptr, size); + } + ~aligned_stack_memory_handler() + { + if(NumTraits<T>::RequireInitialization && m_ptr) + Eigen::internal::destruct_elements_of_array<T>(m_ptr, m_size); + if(m_deallocate) + Eigen::internal::aligned_free(m_ptr); + } + protected: + T* m_ptr; + std::size_t m_size; + bool m_deallocate; +}; + +template<typename T> class scoped_array : noncopyable +{ + T* m_ptr; +public: + explicit scoped_array(std::ptrdiff_t size) + { + m_ptr = new T[size]; + } + ~scoped_array() + { + delete[] m_ptr; + } + T& operator[](std::ptrdiff_t i) { return m_ptr[i]; } + const T& operator[](std::ptrdiff_t i) const { return m_ptr[i]; } + T* &ptr() { return m_ptr; } + const T* ptr() const { return m_ptr; } + operator const T*() const { return m_ptr; } +}; + +template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b) +{ + std::swap(a.ptr(),b.ptr()); +} + +} // end namespace internal + +/** \internal + * Declares, allocates and construct an aligned buffer named NAME of SIZE elements of type TYPE on the stack + * if SIZE is smaller than EIGEN_STACK_ALLOCATION_LIMIT, and if stack allocation is supported by the platform + * (currently, this is Linux and Visual Studio only). Otherwise the memory is allocated on the heap. + * The allocated buffer is automatically deleted when exiting the scope of this declaration. + * If BUFFER is non null, then the declared variable is simply an alias for BUFFER, and no allocation/deletion occurs. + * Here is an example: + * \code + * { + * ei_declare_aligned_stack_constructed_variable(float,data,size,0); + * // use data[0] to data[size-1] + * } + * \endcode + * The underlying stack allocation function can controlled with the EIGEN_ALLOCA preprocessor token. + */ +#ifdef EIGEN_ALLOCA + + #if EIGEN_DEFAULT_ALIGN_BYTES>0 + // We always manually re-align the result of EIGEN_ALLOCA. + // If alloca is already aligned, the compiler should be smart enough to optimize away the re-alignment. + #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast<void*>((internal::UIntPtr(EIGEN_ALLOCA(SIZE+EIGEN_DEFAULT_ALIGN_BYTES-1)) + EIGEN_DEFAULT_ALIGN_BYTES-1) & ~(std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1))) + #else + #define EIGEN_ALIGNED_ALLOCA(SIZE) EIGEN_ALLOCA(SIZE) + #endif + + #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \ + Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \ + TYPE* NAME = (BUFFER)!=0 ? (BUFFER) \ + : reinterpret_cast<TYPE*>( \ + (sizeof(TYPE)*SIZE<=EIGEN_STACK_ALLOCATION_LIMIT) ? EIGEN_ALIGNED_ALLOCA(sizeof(TYPE)*SIZE) \ + : Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE) ); \ + Eigen::internal::aligned_stack_memory_handler<TYPE> EIGEN_CAT(NAME,_stack_memory_destructor)((BUFFER)==0 ? NAME : 0,SIZE,sizeof(TYPE)*SIZE>EIGEN_STACK_ALLOCATION_LIMIT) + +#else + + #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \ + Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \ + TYPE* NAME = (BUFFER)!=0 ? BUFFER : reinterpret_cast<TYPE*>(Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE)); \ + Eigen::internal::aligned_stack_memory_handler<TYPE> EIGEN_CAT(NAME,_stack_memory_destructor)((BUFFER)==0 ? NAME : 0,SIZE,true) + +#endif + + +/***************************************************************************** +*** Implementation of EIGEN_MAKE_ALIGNED_OPERATOR_NEW [_IF] *** +*****************************************************************************/ + +#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 { \ + EIGEN_TRY { return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); } \ + EIGEN_CATCH (...) { return 0; } \ + } + #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign) \ + void *operator new(std::size_t size) { \ + return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \ + } \ + void *operator new[](std::size_t size) { \ + return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \ + } \ + void operator delete(void * ptr) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \ + void operator delete[](void * ptr) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \ + void operator delete(void * ptr, std::size_t /* sz */) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \ + void operator delete[](void * ptr, std::size_t /* sz */) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \ + /* in-place new and delete. since (at least afaik) there is no actual */ \ + /* memory allocated we can safely let the default implementation handle */ \ + /* this particular case. */ \ + static void *operator new(std::size_t size, void *ptr) { return ::operator new(size,ptr); } \ + static void *operator new[](std::size_t size, void* ptr) { return ::operator new[](size,ptr); } \ + void operator delete(void * memory, void *ptr) EIGEN_NO_THROW { return ::operator delete(memory,ptr); } \ + void operator delete[](void * memory, void *ptr) EIGEN_NO_THROW { return ::operator delete[](memory,ptr); } \ + /* nothrow-new (returns zero instead of std::bad_alloc) */ \ + EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \ + void operator delete(void *ptr, const std::nothrow_t&) EIGEN_NO_THROW { \ + Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); \ + } \ + typedef void eigen_aligned_operator_new_marker_type; +#else + #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign) +#endif + +#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(true) +#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size) \ + EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool(((Size)!=Eigen::Dynamic) && ((sizeof(Scalar)*(Size))%EIGEN_MAX_ALIGN_BYTES==0))) + +/****************************************************************************/ + +/** \class aligned_allocator +* \ingroup Core_Module +* +* \brief STL compatible allocator to use with with 16 byte aligned types +* +* Example: +* \code +* // Matrix4f requires 16 bytes alignment: +* std::map< int, Matrix4f, std::less<int>, +* aligned_allocator<std::pair<const int, Matrix4f> > > my_map_mat4; +* // Vector3f does not require 16 bytes alignment, no need to use Eigen's allocator: +* std::map< int, Vector3f > my_map_vec3; +* \endcode +* +* \sa \blank \ref TopicStlContainers. +*/ +template<class T> +class aligned_allocator : public std::allocator<T> +{ +public: + typedef std::size_t size_type; + typedef std::ptrdiff_t difference_type; + typedef T* pointer; + typedef const T* const_pointer; + typedef T& reference; + typedef const T& const_reference; + typedef T value_type; + + template<class U> + struct rebind + { + typedef aligned_allocator<U> other; + }; + + aligned_allocator() : std::allocator<T>() {} + + aligned_allocator(const aligned_allocator& other) : std::allocator<T>(other) {} + + template<class U> + aligned_allocator(const aligned_allocator<U>& other) : std::allocator<T>(other) {} + + ~aligned_allocator() {} + + pointer allocate(size_type num, const void* /*hint*/ = 0) + { + internal::check_size_for_overflow<T>(num); + return static_cast<pointer>( internal::aligned_malloc(num * sizeof(T)) ); + } + + void deallocate(pointer p, size_type /*num*/) + { + internal::aligned_free(p); + } +}; + +//---------- Cache sizes ---------- + +#if !defined(EIGEN_NO_CPUID) +# if EIGEN_COMP_GNUC && EIGEN_ARCH_i386_OR_x86_64 +# if defined(__PIC__) && EIGEN_ARCH_i386 + // Case for x86 with PIC +# define EIGEN_CPUID(abcd,func,id) \ + __asm__ __volatile__ ("xchgl %%ebx, %k1;cpuid; xchgl %%ebx,%k1": "=a" (abcd[0]), "=&r" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "a" (func), "c" (id)); +# elif defined(__PIC__) && EIGEN_ARCH_x86_64 + // Case for x64 with PIC. In theory this is only a problem with recent gcc and with medium or large code model, not with the default small code model. + // However, we cannot detect which code model is used, and the xchg overhead is negligible anyway. +# define EIGEN_CPUID(abcd,func,id) \ + __asm__ __volatile__ ("xchg{q}\t{%%}rbx, %q1; cpuid; xchg{q}\t{%%}rbx, %q1": "=a" (abcd[0]), "=&r" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "0" (func), "2" (id)); +# else + // Case for x86_64 or x86 w/o PIC +# define EIGEN_CPUID(abcd,func,id) \ + __asm__ __volatile__ ("cpuid": "=a" (abcd[0]), "=b" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "0" (func), "2" (id) ); +# endif +# elif EIGEN_COMP_MSVC +# if (EIGEN_COMP_MSVC > 1500) && EIGEN_ARCH_i386_OR_x86_64 +# define EIGEN_CPUID(abcd,func,id) __cpuidex((int*)abcd,func,id) +# endif +# endif +#endif + +namespace internal { + +#ifdef EIGEN_CPUID + +inline bool cpuid_is_vendor(int abcd[4], const int vendor[3]) +{ + return abcd[1]==vendor[0] && abcd[3]==vendor[1] && abcd[2]==vendor[2]; +} + +inline void queryCacheSizes_intel_direct(int& l1, int& l2, int& l3) +{ + int abcd[4]; + l1 = l2 = l3 = 0; + int cache_id = 0; + int cache_type = 0; + do { + abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0; + EIGEN_CPUID(abcd,0x4,cache_id); + cache_type = (abcd[0] & 0x0F) >> 0; + if(cache_type==1||cache_type==3) // data or unified cache + { + int cache_level = (abcd[0] & 0xE0) >> 5; // A[7:5] + int ways = (abcd[1] & 0xFFC00000) >> 22; // B[31:22] + int partitions = (abcd[1] & 0x003FF000) >> 12; // B[21:12] + int line_size = (abcd[1] & 0x00000FFF) >> 0; // B[11:0] + int sets = (abcd[2]); // C[31:0] + + int cache_size = (ways+1) * (partitions+1) * (line_size+1) * (sets+1); + + switch(cache_level) + { + case 1: l1 = cache_size; break; + case 2: l2 = cache_size; break; + case 3: l3 = cache_size; break; + default: break; + } + } + cache_id++; + } while(cache_type>0 && cache_id<16); +} + +inline void queryCacheSizes_intel_codes(int& l1, int& l2, int& l3) +{ + int abcd[4]; + abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0; + l1 = l2 = l3 = 0; + EIGEN_CPUID(abcd,0x00000002,0); + unsigned char * bytes = reinterpret_cast<unsigned char *>(abcd)+2; + bool check_for_p2_core2 = false; + for(int i=0; i<14; ++i) + { + switch(bytes[i]) + { + case 0x0A: l1 = 8; break; // 0Ah data L1 cache, 8 KB, 2 ways, 32 byte lines + case 0x0C: l1 = 16; break; // 0Ch data L1 cache, 16 KB, 4 ways, 32 byte lines + case 0x0E: l1 = 24; break; // 0Eh data L1 cache, 24 KB, 6 ways, 64 byte lines + case 0x10: l1 = 16; break; // 10h data L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64) + case 0x15: l1 = 16; break; // 15h code L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64) + case 0x2C: l1 = 32; break; // 2Ch data L1 cache, 32 KB, 8 ways, 64 byte lines + case 0x30: l1 = 32; break; // 30h code L1 cache, 32 KB, 8 ways, 64 byte lines + case 0x60: l1 = 16; break; // 60h data L1 cache, 16 KB, 8 ways, 64 byte lines, sectored + case 0x66: l1 = 8; break; // 66h data L1 cache, 8 KB, 4 ways, 64 byte lines, sectored + case 0x67: l1 = 16; break; // 67h data L1 cache, 16 KB, 4 ways, 64 byte lines, sectored + case 0x68: l1 = 32; break; // 68h data L1 cache, 32 KB, 4 ways, 64 byte lines, sectored + case 0x1A: l2 = 96; break; // code and data L2 cache, 96 KB, 6 ways, 64 byte lines (IA-64) + case 0x22: l3 = 512; break; // code and data L3 cache, 512 KB, 4 ways (!), 64 byte lines, dual-sectored + case 0x23: l3 = 1024; break; // code and data L3 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored + case 0x25: l3 = 2048; break; // code and data L3 cache, 2048 KB, 8 ways, 64 byte lines, dual-sectored + case 0x29: l3 = 4096; break; // code and data L3 cache, 4096 KB, 8 ways, 64 byte lines, dual-sectored + case 0x39: l2 = 128; break; // code and data L2 cache, 128 KB, 4 ways, 64 byte lines, sectored + case 0x3A: l2 = 192; break; // code and data L2 cache, 192 KB, 6 ways, 64 byte lines, sectored + case 0x3B: l2 = 128; break; // code and data L2 cache, 128 KB, 2 ways, 64 byte lines, sectored + case 0x3C: l2 = 256; break; // code and data L2 cache, 256 KB, 4 ways, 64 byte lines, sectored + case 0x3D: l2 = 384; break; // code and data L2 cache, 384 KB, 6 ways, 64 byte lines, sectored + case 0x3E: l2 = 512; break; // code and data L2 cache, 512 KB, 4 ways, 64 byte lines, sectored + case 0x40: l2 = 0; break; // no integrated L2 cache (P6 core) or L3 cache (P4 core) + case 0x41: l2 = 128; break; // code and data L2 cache, 128 KB, 4 ways, 32 byte lines + case 0x42: l2 = 256; break; // code and data L2 cache, 256 KB, 4 ways, 32 byte lines + case 0x43: l2 = 512; break; // code and data L2 cache, 512 KB, 4 ways, 32 byte lines + case 0x44: l2 = 1024; break; // code and data L2 cache, 1024 KB, 4 ways, 32 byte lines + case 0x45: l2 = 2048; break; // code and data L2 cache, 2048 KB, 4 ways, 32 byte lines + case 0x46: l3 = 4096; break; // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines + case 0x47: l3 = 8192; break; // code and data L3 cache, 8192 KB, 8 ways, 64 byte lines + case 0x48: l2 = 3072; break; // code and data L2 cache, 3072 KB, 12 ways, 64 byte lines + case 0x49: if(l2!=0) l3 = 4096; else {check_for_p2_core2=true; l3 = l2 = 4096;} break;// code and data L3 cache, 4096 KB, 16 ways, 64 byte lines (P4) or L2 for core2 + case 0x4A: l3 = 6144; break; // code and data L3 cache, 6144 KB, 12 ways, 64 byte lines + case 0x4B: l3 = 8192; break; // code and data L3 cache, 8192 KB, 16 ways, 64 byte lines + case 0x4C: l3 = 12288; break; // code and data L3 cache, 12288 KB, 12 ways, 64 byte lines + case 0x4D: l3 = 16384; break; // code and data L3 cache, 16384 KB, 16 ways, 64 byte lines + case 0x4E: l2 = 6144; break; // code and data L2 cache, 6144 KB, 24 ways, 64 byte lines + case 0x78: l2 = 1024; break; // code and data L2 cache, 1024 KB, 4 ways, 64 byte lines + case 0x79: l2 = 128; break; // code and data L2 cache, 128 KB, 8 ways, 64 byte lines, dual-sectored + case 0x7A: l2 = 256; break; // code and data L2 cache, 256 KB, 8 ways, 64 byte lines, dual-sectored + case 0x7B: l2 = 512; break; // code and data L2 cache, 512 KB, 8 ways, 64 byte lines, dual-sectored + case 0x7C: l2 = 1024; break; // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored + case 0x7D: l2 = 2048; break; // code and data L2 cache, 2048 KB, 8 ways, 64 byte lines + case 0x7E: l2 = 256; break; // code and data L2 cache, 256 KB, 8 ways, 128 byte lines, sect. (IA-64) + case 0x7F: l2 = 512; break; // code and data L2 cache, 512 KB, 2 ways, 64 byte lines + case 0x80: l2 = 512; break; // code and data L2 cache, 512 KB, 8 ways, 64 byte lines + case 0x81: l2 = 128; break; // code and data L2 cache, 128 KB, 8 ways, 32 byte lines + case 0x82: l2 = 256; break; // code and data L2 cache, 256 KB, 8 ways, 32 byte lines + case 0x83: l2 = 512; break; // code and data L2 cache, 512 KB, 8 ways, 32 byte lines + case 0x84: l2 = 1024; break; // code and data L2 cache, 1024 KB, 8 ways, 32 byte lines + case 0x85: l2 = 2048; break; // code and data L2 cache, 2048 KB, 8 ways, 32 byte lines + case 0x86: l2 = 512; break; // code and data L2 cache, 512 KB, 4 ways, 64 byte lines + case 0x87: l2 = 1024; break; // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines + case 0x88: l3 = 2048; break; // code and data L3 cache, 2048 KB, 4 ways, 64 byte lines (IA-64) + case 0x89: l3 = 4096; break; // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines (IA-64) + case 0x8A: l3 = 8192; break; // code and data L3 cache, 8192 KB, 4 ways, 64 byte lines (IA-64) + case 0x8D: l3 = 3072; break; // code and data L3 cache, 3072 KB, 12 ways, 128 byte lines (IA-64) + + default: break; + } + } + if(check_for_p2_core2 && l2 == l3) + l3 = 0; + l1 *= 1024; + l2 *= 1024; + l3 *= 1024; +} + +inline void queryCacheSizes_intel(int& l1, int& l2, int& l3, int max_std_funcs) +{ + if(max_std_funcs>=4) + queryCacheSizes_intel_direct(l1,l2,l3); + else + queryCacheSizes_intel_codes(l1,l2,l3); +} + +inline void queryCacheSizes_amd(int& l1, int& l2, int& l3) +{ + int abcd[4]; + abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0; + EIGEN_CPUID(abcd,0x80000005,0); + l1 = (abcd[2] >> 24) * 1024; // C[31:24] = L1 size in KB + abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0; + EIGEN_CPUID(abcd,0x80000006,0); + l2 = (abcd[2] >> 16) * 1024; // C[31;16] = l2 cache size in KB + l3 = ((abcd[3] & 0xFFFC000) >> 18) * 512 * 1024; // D[31;18] = l3 cache size in 512KB +} +#endif + +/** \internal + * Queries and returns the cache sizes in Bytes of the L1, L2, and L3 data caches respectively */ +inline void queryCacheSizes(int& l1, int& l2, int& l3) +{ + #ifdef EIGEN_CPUID + int abcd[4]; + const int GenuineIntel[] = {0x756e6547, 0x49656e69, 0x6c65746e}; + const int AuthenticAMD[] = {0x68747541, 0x69746e65, 0x444d4163}; + const int AMDisbetter_[] = {0x69444d41, 0x74656273, 0x21726574}; // "AMDisbetter!" + + // identify the CPU vendor + EIGEN_CPUID(abcd,0x0,0); + int max_std_funcs = abcd[1]; + if(cpuid_is_vendor(abcd,GenuineIntel)) + queryCacheSizes_intel(l1,l2,l3,max_std_funcs); + else if(cpuid_is_vendor(abcd,AuthenticAMD) || cpuid_is_vendor(abcd,AMDisbetter_)) + queryCacheSizes_amd(l1,l2,l3); + else + // by default let's use Intel's API + queryCacheSizes_intel(l1,l2,l3,max_std_funcs); + + // here is the list of other vendors: +// ||cpuid_is_vendor(abcd,"VIA VIA VIA ") +// ||cpuid_is_vendor(abcd,"CyrixInstead") +// ||cpuid_is_vendor(abcd,"CentaurHauls") +// ||cpuid_is_vendor(abcd,"GenuineTMx86") +// ||cpuid_is_vendor(abcd,"TransmetaCPU") +// ||cpuid_is_vendor(abcd,"RiseRiseRise") +// ||cpuid_is_vendor(abcd,"Geode by NSC") +// ||cpuid_is_vendor(abcd,"SiS SiS SiS ") +// ||cpuid_is_vendor(abcd,"UMC UMC UMC ") +// ||cpuid_is_vendor(abcd,"NexGenDriven") + #else + l1 = l2 = l3 = -1; + #endif +} + +/** \internal + * \returns the size in Bytes of the L1 data cache */ +inline int queryL1CacheSize() +{ + int l1(-1), l2, l3; + queryCacheSizes(l1,l2,l3); + return l1; +} + +/** \internal + * \returns the size in Bytes of the L2 or L3 cache if this later is present */ +inline int queryTopLevelCacheSize() +{ + int l1, l2(-1), l3(-1); + queryCacheSizes(l1,l2,l3); + return (std::max)(l2,l3); +} + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_MEMORY_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/Meta.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/Meta.h new file mode 100755 index 000000000..7f6370755 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/Meta.h @@ -0,0 +1,492 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2015 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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_META_H +#define EIGEN_META_H + +#if defined(__CUDA_ARCH__) +#include <cfloat> +#include <math_constants.h> +#endif + +#if EIGEN_COMP_ICC>=1600 && __cplusplus >= 201103L +#include <cstdint> +#endif + +namespace Eigen { + +typedef EIGEN_DEFAULT_DENSE_INDEX_TYPE DenseIndex; + +/** + * \brief The Index type as used for the API. + * \details To change this, \c \#define the preprocessor symbol \c EIGEN_DEFAULT_DENSE_INDEX_TYPE. + * \sa \blank \ref TopicPreprocessorDirectives, StorageIndex. + */ + +typedef EIGEN_DEFAULT_DENSE_INDEX_TYPE Index; + +namespace internal { + +/** \internal + * \file Meta.h + * This file contains generic metaprogramming classes which are not specifically related to Eigen. + * \note In case you wonder, yes we're aware that Boost already provides all these features, + * we however don't want to add a dependency to Boost. + */ + +// Only recent versions of ICC complain about using ptrdiff_t to hold pointers, +// and older versions do not provide *intptr_t types. +#if EIGEN_COMP_ICC>=1600 && __cplusplus >= 201103L +typedef std::intptr_t IntPtr; +typedef std::uintptr_t UIntPtr; +#else +typedef std::ptrdiff_t IntPtr; +typedef std::size_t UIntPtr; +#endif + +struct true_type { enum { value = 1 }; }; +struct false_type { enum { value = 0 }; }; + +template<bool Condition, typename Then, typename Else> +struct conditional { typedef Then type; }; + +template<typename Then, typename Else> +struct conditional <false, Then, Else> { typedef Else type; }; + +template<typename T, typename U> struct is_same { enum { value = 0 }; }; +template<typename T> struct is_same<T,T> { enum { value = 1 }; }; + +template<typename T> struct remove_reference { typedef T type; }; +template<typename T> struct remove_reference<T&> { typedef T type; }; + +template<typename T> struct remove_pointer { typedef T type; }; +template<typename T> struct remove_pointer<T*> { typedef T type; }; +template<typename T> struct remove_pointer<T*const> { typedef T type; }; + +template <class T> struct remove_const { typedef T type; }; +template <class T> struct remove_const<const T> { typedef T type; }; +template <class T> struct remove_const<const T[]> { typedef T type[]; }; +template <class T, unsigned int Size> struct remove_const<const T[Size]> { typedef T type[Size]; }; + +template<typename T> struct remove_all { typedef T type; }; +template<typename T> struct remove_all<const T> { typedef typename remove_all<T>::type type; }; +template<typename T> struct remove_all<T const&> { typedef typename remove_all<T>::type type; }; +template<typename T> struct remove_all<T&> { typedef typename remove_all<T>::type type; }; +template<typename T> struct remove_all<T const*> { typedef typename remove_all<T>::type type; }; +template<typename T> struct remove_all<T*> { typedef typename remove_all<T>::type type; }; + +template<typename T> struct is_arithmetic { enum { value = false }; }; +template<> struct is_arithmetic<float> { enum { value = true }; }; +template<> struct is_arithmetic<double> { enum { value = true }; }; +template<> struct is_arithmetic<long double> { enum { value = true }; }; +template<> struct is_arithmetic<bool> { enum { value = true }; }; +template<> struct is_arithmetic<char> { enum { value = true }; }; +template<> struct is_arithmetic<signed char> { enum { value = true }; }; +template<> struct is_arithmetic<unsigned char> { enum { value = true }; }; +template<> struct is_arithmetic<signed short> { enum { value = true }; }; +template<> struct is_arithmetic<unsigned short>{ enum { value = true }; }; +template<> struct is_arithmetic<signed int> { enum { value = true }; }; +template<> struct is_arithmetic<unsigned int> { enum { value = true }; }; +template<> struct is_arithmetic<signed long> { enum { value = true }; }; +template<> struct is_arithmetic<unsigned long> { enum { value = true }; }; + +template<typename T> struct is_integral { enum { value = false }; }; +template<> struct is_integral<bool> { enum { value = true }; }; +template<> struct is_integral<char> { enum { value = true }; }; +template<> struct is_integral<signed char> { enum { value = true }; }; +template<> struct is_integral<unsigned char> { enum { value = true }; }; +template<> struct is_integral<signed short> { enum { value = true }; }; +template<> struct is_integral<unsigned short> { enum { value = true }; }; +template<> struct is_integral<signed int> { enum { value = true }; }; +template<> struct is_integral<unsigned int> { enum { value = true }; }; +template<> struct is_integral<signed long> { enum { value = true }; }; +template<> struct is_integral<unsigned long> { enum { value = true }; }; + +template <typename T> struct add_const { typedef const T type; }; +template <typename T> struct add_const<T&> { typedef T& type; }; + +template <typename T> struct is_const { enum { value = 0 }; }; +template <typename T> struct is_const<T const> { enum { value = 1 }; }; + +template<typename T> struct add_const_on_value_type { typedef const T type; }; +template<typename T> struct add_const_on_value_type<T&> { typedef T const& type; }; +template<typename T> struct add_const_on_value_type<T*> { typedef T const* type; }; +template<typename T> struct add_const_on_value_type<T* const> { typedef T const* const type; }; +template<typename T> struct add_const_on_value_type<T const* const> { typedef T const* const type; }; + + +template<typename From, typename To> +struct is_convertible_impl +{ +private: + struct any_conversion + { + template <typename T> any_conversion(const volatile T&); + template <typename T> any_conversion(T&); + }; + struct yes {int a[1];}; + struct no {int a[2];}; + + static yes test(const To&, int); + static no test(any_conversion, ...); + +public: + static From ms_from; +#ifdef __INTEL_COMPILER + #pragma warning push + #pragma warning ( disable : 2259 ) +#endif + enum { value = sizeof(test(ms_from, 0))==sizeof(yes) }; +#ifdef __INTEL_COMPILER + #pragma warning pop +#endif +}; + +template<typename From, typename To> +struct is_convertible +{ + enum { value = is_convertible_impl<typename remove_all<From>::type, + typename remove_all<To >::type>::value }; +}; + +/** \internal Allows to enable/disable an overload + * according to a compile time condition. + */ +template<bool Condition, typename T=void> struct enable_if; + +template<typename T> struct enable_if<true,T> +{ typedef T type; }; + +#if defined(__CUDA_ARCH__) +#if !defined(__FLT_EPSILON__) +#define __FLT_EPSILON__ FLT_EPSILON +#define __DBL_EPSILON__ DBL_EPSILON +#endif + +namespace device { + +template<typename T> struct numeric_limits +{ + EIGEN_DEVICE_FUNC + static T epsilon() { return 0; } + static T (max)() { assert(false && "Highest not supported for this type"); } + static T (min)() { assert(false && "Lowest not supported for this type"); } + static T infinity() { assert(false && "Infinity not supported for this type"); } + static T quiet_NaN() { assert(false && "quiet_NaN not supported for this type"); } +}; +template<> struct numeric_limits<float> +{ + EIGEN_DEVICE_FUNC + static float epsilon() { return __FLT_EPSILON__; } + EIGEN_DEVICE_FUNC + static float (max)() { return CUDART_MAX_NORMAL_F; } + EIGEN_DEVICE_FUNC + static float (min)() { return FLT_MIN; } + EIGEN_DEVICE_FUNC + static float infinity() { return CUDART_INF_F; } + EIGEN_DEVICE_FUNC + static float quiet_NaN() { return CUDART_NAN_F; } +}; +template<> struct numeric_limits<double> +{ + EIGEN_DEVICE_FUNC + static double epsilon() { return __DBL_EPSILON__; } + EIGEN_DEVICE_FUNC + static double (max)() { return DBL_MAX; } + EIGEN_DEVICE_FUNC + static double (min)() { return DBL_MIN; } + EIGEN_DEVICE_FUNC + static double infinity() { return CUDART_INF; } + EIGEN_DEVICE_FUNC + static double quiet_NaN() { return CUDART_NAN; } +}; +template<> struct numeric_limits<int> +{ + EIGEN_DEVICE_FUNC + static int epsilon() { return 0; } + EIGEN_DEVICE_FUNC + static int (max)() { return INT_MAX; } + EIGEN_DEVICE_FUNC + static int (min)() { return INT_MIN; } +}; +template<> struct numeric_limits<unsigned int> +{ + EIGEN_DEVICE_FUNC + static unsigned int epsilon() { return 0; } + EIGEN_DEVICE_FUNC + static unsigned int (max)() { return UINT_MAX; } + EIGEN_DEVICE_FUNC + static unsigned int (min)() { return 0; } +}; +template<> struct numeric_limits<long> +{ + EIGEN_DEVICE_FUNC + static long epsilon() { return 0; } + EIGEN_DEVICE_FUNC + static long (max)() { return LONG_MAX; } + EIGEN_DEVICE_FUNC + static long (min)() { return LONG_MIN; } +}; +template<> struct numeric_limits<unsigned long> +{ + EIGEN_DEVICE_FUNC + static unsigned long epsilon() { return 0; } + EIGEN_DEVICE_FUNC + static unsigned long (max)() { return ULONG_MAX; } + EIGEN_DEVICE_FUNC + static unsigned long (min)() { return 0; } +}; +template<> struct numeric_limits<long long> +{ + EIGEN_DEVICE_FUNC + static long long epsilon() { return 0; } + EIGEN_DEVICE_FUNC + static long long (max)() { return LLONG_MAX; } + EIGEN_DEVICE_FUNC + static long long (min)() { return LLONG_MIN; } +}; +template<> struct numeric_limits<unsigned long long> +{ + EIGEN_DEVICE_FUNC + static unsigned long long epsilon() { return 0; } + EIGEN_DEVICE_FUNC + static unsigned long long (max)() { return ULLONG_MAX; } + EIGEN_DEVICE_FUNC + static unsigned long long (min)() { return 0; } +}; + +} + +#endif + +/** \internal + * A base class do disable default copy ctor and copy assignement operator. + */ +class noncopyable +{ + EIGEN_DEVICE_FUNC noncopyable(const noncopyable&); + EIGEN_DEVICE_FUNC const noncopyable& operator=(const noncopyable&); +protected: + EIGEN_DEVICE_FUNC noncopyable() {} + EIGEN_DEVICE_FUNC ~noncopyable() {} +}; + +/** \internal + * Convenient struct to get the result type of a unary or binary functor. + * + * It supports both the current STL mechanism (using the result_type member) as well as + * upcoming next STL generation (using a templated result member). + * If none of these members is provided, then the type of the first argument is returned. FIXME, that behavior is a pretty bad hack. + */ +#if EIGEN_HAS_STD_RESULT_OF +template<typename T> struct result_of { + typedef typename std::result_of<T>::type type1; + typedef typename remove_all<type1>::type type; +}; +#else +template<typename T> struct result_of { }; + +struct has_none {int a[1];}; +struct has_std_result_type {int a[2];}; +struct has_tr1_result {int a[3];}; + +template<typename Func, typename ArgType, int SizeOf=sizeof(has_none)> +struct unary_result_of_select {typedef typename internal::remove_all<ArgType>::type type;}; + +template<typename Func, typename ArgType> +struct unary_result_of_select<Func, ArgType, sizeof(has_std_result_type)> {typedef typename Func::result_type type;}; + +template<typename Func, typename ArgType> +struct unary_result_of_select<Func, ArgType, sizeof(has_tr1_result)> {typedef typename Func::template result<Func(ArgType)>::type type;}; + +template<typename Func, typename ArgType> +struct result_of<Func(ArgType)> { + template<typename T> + static has_std_result_type testFunctor(T const *, typename T::result_type const * = 0); + template<typename T> + static has_tr1_result testFunctor(T const *, typename T::template result<T(ArgType)>::type const * = 0); + static has_none testFunctor(...); + + // note that the following indirection is needed for gcc-3.3 + enum {FunctorType = sizeof(testFunctor(static_cast<Func*>(0)))}; + typedef typename unary_result_of_select<Func, ArgType, FunctorType>::type type; +}; + +template<typename Func, typename ArgType0, typename ArgType1, int SizeOf=sizeof(has_none)> +struct binary_result_of_select {typedef typename internal::remove_all<ArgType0>::type type;}; + +template<typename Func, typename ArgType0, typename ArgType1> +struct binary_result_of_select<Func, ArgType0, ArgType1, sizeof(has_std_result_type)> +{typedef typename Func::result_type type;}; + +template<typename Func, typename ArgType0, typename ArgType1> +struct binary_result_of_select<Func, ArgType0, ArgType1, sizeof(has_tr1_result)> +{typedef typename Func::template result<Func(ArgType0,ArgType1)>::type type;}; + +template<typename Func, typename ArgType0, typename ArgType1> +struct result_of<Func(ArgType0,ArgType1)> { + template<typename T> + static has_std_result_type testFunctor(T const *, typename T::result_type const * = 0); + template<typename T> + static has_tr1_result testFunctor(T const *, typename T::template result<T(ArgType0,ArgType1)>::type const * = 0); + static has_none testFunctor(...); + + // note that the following indirection is needed for gcc-3.3 + enum {FunctorType = sizeof(testFunctor(static_cast<Func*>(0)))}; + typedef typename binary_result_of_select<Func, ArgType0, ArgType1, FunctorType>::type type; +}; + +template<typename Func, typename ArgType0, typename ArgType1, typename ArgType2, int SizeOf=sizeof(has_none)> +struct ternary_result_of_select {typedef typename internal::remove_all<ArgType0>::type type;}; + +template<typename Func, typename ArgType0, typename ArgType1, typename ArgType2> +struct ternary_result_of_select<Func, ArgType0, ArgType1, ArgType2, sizeof(has_std_result_type)> +{typedef typename Func::result_type type;}; + +template<typename Func, typename ArgType0, typename ArgType1, typename ArgType2> +struct ternary_result_of_select<Func, ArgType0, ArgType1, ArgType2, sizeof(has_tr1_result)> +{typedef typename Func::template result<Func(ArgType0,ArgType1,ArgType2)>::type type;}; + +template<typename Func, typename ArgType0, typename ArgType1, typename ArgType2> +struct result_of<Func(ArgType0,ArgType1,ArgType2)> { + template<typename T> + static has_std_result_type testFunctor(T const *, typename T::result_type const * = 0); + template<typename T> + static has_tr1_result testFunctor(T const *, typename T::template result<T(ArgType0,ArgType1,ArgType2)>::type const * = 0); + static has_none testFunctor(...); + + // note that the following indirection is needed for gcc-3.3 + enum {FunctorType = sizeof(testFunctor(static_cast<Func*>(0)))}; + typedef typename ternary_result_of_select<Func, ArgType0, ArgType1, ArgType2, FunctorType>::type type; +}; +#endif + +struct meta_yes { char a[1]; }; +struct meta_no { char a[2]; }; + +// Check whether T::ReturnType does exist +template <typename T> +struct has_ReturnType +{ + template <typename C> static meta_yes testFunctor(typename C::ReturnType const *); + template <typename C> static meta_no testFunctor(...); + + enum { value = sizeof(testFunctor<T>(0)) == sizeof(meta_yes) }; +}; + +template<typename T> const T* return_ptr(); + +template <typename T, typename IndexType=Index> +struct has_nullary_operator +{ + template <typename C> static meta_yes testFunctor(C const *,typename enable_if<(sizeof(return_ptr<C>()->operator()())>0)>::type * = 0); + static meta_no testFunctor(...); + + enum { value = sizeof(testFunctor(static_cast<T*>(0))) == sizeof(meta_yes) }; +}; + +template <typename T, typename IndexType=Index> +struct has_unary_operator +{ + template <typename C> static meta_yes testFunctor(C const *,typename enable_if<(sizeof(return_ptr<C>()->operator()(IndexType(0)))>0)>::type * = 0); + static meta_no testFunctor(...); + + enum { value = sizeof(testFunctor(static_cast<T*>(0))) == sizeof(meta_yes) }; +}; + +template <typename T, typename IndexType=Index> +struct has_binary_operator +{ + template <typename C> static meta_yes testFunctor(C const *,typename enable_if<(sizeof(return_ptr<C>()->operator()(IndexType(0),IndexType(0)))>0)>::type * = 0); + static meta_no testFunctor(...); + + enum { value = sizeof(testFunctor(static_cast<T*>(0))) == sizeof(meta_yes) }; +}; + +/** \internal In short, it computes int(sqrt(\a Y)) with \a Y an integer. + * Usage example: \code meta_sqrt<1023>::ret \endcode + */ +template<int Y, + int InfX = 0, + int SupX = ((Y==1) ? 1 : Y/2), + bool Done = ((SupX-InfX)<=1 ? true : ((SupX*SupX <= Y) && ((SupX+1)*(SupX+1) > Y))) > + // use ?: instead of || just to shut up a stupid gcc 4.3 warning +class meta_sqrt +{ + enum { + MidX = (InfX+SupX)/2, + TakeInf = MidX*MidX > Y ? 1 : 0, + NewInf = int(TakeInf) ? InfX : int(MidX), + NewSup = int(TakeInf) ? int(MidX) : SupX + }; + public: + enum { ret = meta_sqrt<Y,NewInf,NewSup>::ret }; +}; + +template<int Y, int InfX, int SupX> +class meta_sqrt<Y, InfX, SupX, true> { public: enum { ret = (SupX*SupX <= Y) ? SupX : InfX }; }; + + +/** \internal Computes the least common multiple of two positive integer A and B + * at compile-time. It implements a naive algorithm testing all multiples of A. + * It thus works better if A>=B. + */ +template<int A, int B, int K=1, bool Done = ((A*K)%B)==0> +struct meta_least_common_multiple +{ + enum { ret = meta_least_common_multiple<A,B,K+1>::ret }; +}; +template<int A, int B, int K> +struct meta_least_common_multiple<A,B,K,true> +{ + enum { ret = A*K }; +}; + +/** \internal determines whether the product of two numeric types is allowed and what the return type is */ +template<typename T, typename U> struct scalar_product_traits +{ + enum { Defined = 0 }; +}; + +// FIXME quick workaround around current limitation of result_of +// template<typename Scalar, typename ArgType0, typename ArgType1> +// struct result_of<scalar_product_op<Scalar>(ArgType0,ArgType1)> { +// typedef typename scalar_product_traits<typename remove_all<ArgType0>::type, typename remove_all<ArgType1>::type>::ReturnType type; +// }; + +} // end namespace internal + +namespace numext { + +#if defined(__CUDA_ARCH__) +template<typename T> EIGEN_DEVICE_FUNC void swap(T &a, T &b) { T tmp = b; b = a; a = tmp; } +#else +template<typename T> EIGEN_STRONG_INLINE void swap(T &a, T &b) { std::swap(a,b); } +#endif + +#if defined(__CUDA_ARCH__) +using internal::device::numeric_limits; +#else +using std::numeric_limits; +#endif + +// Integer division with rounding up. +// T is assumed to be an integer type with a>=0, and b>0 +template<typename T> +T div_ceil(const T &a, const T &b) +{ + return (a+b-1) / b; +} + +} // end namespace numext + +} // end namespace Eigen + +#endif // EIGEN_META_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/NonMPL2.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/NonMPL2.h new file mode 100644 index 000000000..1af67cf18 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/NonMPL2.h @@ -0,0 +1,3 @@ +#ifdef EIGEN_MPL2_ONLY +#error Including non-MPL2 code in EIGEN_MPL2_ONLY mode +#endif diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/ReenableStupidWarnings.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/ReenableStupidWarnings.h new file mode 100644 index 000000000..86b60f52f --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/ReenableStupidWarnings.h @@ -0,0 +1,27 @@ +#ifdef EIGEN_WARNINGS_DISABLED +#undef EIGEN_WARNINGS_DISABLED + +#ifndef EIGEN_PERMANENTLY_DISABLE_STUPID_WARNINGS + #ifdef _MSC_VER + #pragma warning( pop ) + #elif defined __INTEL_COMPILER + #pragma warning pop + #elif defined __clang__ + #pragma clang diagnostic pop + #elif defined __GNUC__ && __GNUC__>=6 + #pragma GCC diagnostic pop + #endif + + #if defined __NVCC__ +// Don't reenable the diagnostic messages, as it turns out these messages need +// to be disabled at the point of the template instantiation (i.e the user code) +// otherwise they'll be triggered by nvcc. +// #pragma diag_default code_is_unreachable +// #pragma diag_default initialization_not_reachable +// #pragma diag_default 2651 +// #pragma diag_default 2653 + #endif + +#endif + +#endif // EIGEN_WARNINGS_DISABLED diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/StaticAssert.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/StaticAssert.h new file mode 100644 index 000000000..983361a45 --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/StaticAssert.h @@ -0,0 +1,216 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2008 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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_STATIC_ASSERT_H +#define EIGEN_STATIC_ASSERT_H + +/* Some notes on Eigen's static assertion mechanism: + * + * - in EIGEN_STATIC_ASSERT(CONDITION,MSG) the parameter CONDITION must be a compile time boolean + * expression, and MSG an enum listed in struct internal::static_assertion<true> + * + * - define EIGEN_NO_STATIC_ASSERT to disable them (and save compilation time) + * in that case, the static assertion is converted to the following runtime assert: + * eigen_assert(CONDITION && "MSG") + * + * - currently EIGEN_STATIC_ASSERT can only be used in function scope + * + */ + +#ifndef EIGEN_NO_STATIC_ASSERT + + #if EIGEN_MAX_CPP_VER>=11 && (__has_feature(cxx_static_assert) || (defined(__cplusplus) && __cplusplus >= 201103L) || (EIGEN_COMP_MSVC >= 1600)) + + // if native static_assert is enabled, let's use it + #define EIGEN_STATIC_ASSERT(X,MSG) static_assert(X,#MSG); + + #else // not CXX0X + + namespace Eigen { + + namespace internal { + + template<bool condition> + struct static_assertion {}; + + template<> + struct static_assertion<true> + { + enum { + YOU_TRIED_CALLING_A_VECTOR_METHOD_ON_A_MATRIX, + YOU_MIXED_VECTORS_OF_DIFFERENT_SIZES, + YOU_MIXED_MATRICES_OF_DIFFERENT_SIZES, + THIS_METHOD_IS_ONLY_FOR_VECTORS_OF_A_SPECIFIC_SIZE, + THIS_METHOD_IS_ONLY_FOR_MATRICES_OF_A_SPECIFIC_SIZE, + THIS_METHOD_IS_ONLY_FOR_OBJECTS_OF_A_SPECIFIC_SIZE, + OUT_OF_RANGE_ACCESS, + YOU_MADE_A_PROGRAMMING_MISTAKE, + EIGEN_INTERNAL_ERROR_PLEASE_FILE_A_BUG_REPORT, + EIGEN_INTERNAL_COMPILATION_ERROR_OR_YOU_MADE_A_PROGRAMMING_MISTAKE, + YOU_CALLED_A_FIXED_SIZE_METHOD_ON_A_DYNAMIC_SIZE_MATRIX_OR_VECTOR, + YOU_CALLED_A_DYNAMIC_SIZE_METHOD_ON_A_FIXED_SIZE_MATRIX_OR_VECTOR, + UNALIGNED_LOAD_AND_STORE_OPERATIONS_UNIMPLEMENTED_ON_ALTIVEC, + THIS_FUNCTION_IS_NOT_FOR_INTEGER_NUMERIC_TYPES, + FLOATING_POINT_ARGUMENT_PASSED__INTEGER_WAS_EXPECTED, + NUMERIC_TYPE_MUST_BE_REAL, + COEFFICIENT_WRITE_ACCESS_TO_SELFADJOINT_NOT_SUPPORTED, + WRITING_TO_TRIANGULAR_PART_WITH_UNIT_DIAGONAL_IS_NOT_SUPPORTED, + THIS_METHOD_IS_ONLY_FOR_FIXED_SIZE, + INVALID_MATRIX_PRODUCT, + INVALID_VECTOR_VECTOR_PRODUCT__IF_YOU_WANTED_A_DOT_OR_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTIONS, + INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION, + YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY, + THIS_METHOD_IS_ONLY_FOR_COLUMN_MAJOR_MATRICES, + THIS_METHOD_IS_ONLY_FOR_ROW_MAJOR_MATRICES, + INVALID_MATRIX_TEMPLATE_PARAMETERS, + INVALID_MATRIXBASE_TEMPLATE_PARAMETERS, + BOTH_MATRICES_MUST_HAVE_THE_SAME_STORAGE_ORDER, + THIS_METHOD_IS_ONLY_FOR_DIAGONAL_MATRIX, + THE_MATRIX_OR_EXPRESSION_THAT_YOU_PASSED_DOES_NOT_HAVE_THE_EXPECTED_TYPE, + THIS_METHOD_IS_ONLY_FOR_EXPRESSIONS_WITH_DIRECT_MEMORY_ACCESS_SUCH_AS_MAP_OR_PLAIN_MATRICES, + YOU_ALREADY_SPECIFIED_THIS_STRIDE, + INVALID_STORAGE_ORDER_FOR_THIS_VECTOR_EXPRESSION, + THE_BRACKET_OPERATOR_IS_ONLY_FOR_VECTORS__USE_THE_PARENTHESIS_OPERATOR_INSTEAD, + PACKET_ACCESS_REQUIRES_TO_HAVE_INNER_STRIDE_FIXED_TO_1, + THIS_METHOD_IS_ONLY_FOR_SPECIFIC_TRANSFORMATIONS, + YOU_CANNOT_MIX_ARRAYS_AND_MATRICES, + YOU_PERFORMED_AN_INVALID_TRANSFORMATION_CONVERSION, + THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY, + YOU_ARE_TRYING_TO_USE_AN_INDEX_BASED_ACCESSOR_ON_AN_EXPRESSION_THAT_DOES_NOT_SUPPORT_THAT, + THIS_METHOD_IS_ONLY_FOR_1x1_EXPRESSIONS, + THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS, + THIS_METHOD_IS_ONLY_FOR_EXPRESSIONS_OF_BOOL, + THIS_METHOD_IS_ONLY_FOR_ARRAYS_NOT_MATRICES, + YOU_PASSED_A_ROW_VECTOR_BUT_A_COLUMN_VECTOR_WAS_EXPECTED, + YOU_PASSED_A_COLUMN_VECTOR_BUT_A_ROW_VECTOR_WAS_EXPECTED, + THE_INDEX_TYPE_MUST_BE_A_SIGNED_TYPE, + THE_STORAGE_ORDER_OF_BOTH_SIDES_MUST_MATCH, + OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG, + IMPLICIT_CONVERSION_TO_SCALAR_IS_FOR_INNER_PRODUCT_ONLY, + STORAGE_LAYOUT_DOES_NOT_MATCH, + EIGEN_INTERNAL_ERROR_PLEASE_FILE_A_BUG_REPORT__INVALID_COST_VALUE, + THIS_COEFFICIENT_ACCESSOR_TAKING_ONE_ACCESS_IS_ONLY_FOR_EXPRESSIONS_ALLOWING_LINEAR_ACCESS, + MATRIX_FREE_CONJUGATE_GRADIENT_IS_COMPATIBLE_WITH_UPPER_UNION_LOWER_MODE_ONLY, + THIS_TYPE_IS_NOT_SUPPORTED, + STORAGE_KIND_MUST_MATCH, + STORAGE_INDEX_MUST_MATCH, + CHOLMOD_SUPPORTS_DOUBLE_PRECISION_ONLY + }; + }; + + } // end namespace internal + + } // end namespace Eigen + + // Specialized implementation for MSVC to avoid "conditional + // expression is constant" warnings. This implementation doesn't + // appear to work under GCC, hence the multiple implementations. + #if EIGEN_COMP_MSVC + + #define EIGEN_STATIC_ASSERT(CONDITION,MSG) \ + {Eigen::internal::static_assertion<bool(CONDITION)>::MSG;} + + #else + // In some cases clang interprets bool(CONDITION) as function declaration + #define EIGEN_STATIC_ASSERT(CONDITION,MSG) \ + if (Eigen::internal::static_assertion<static_cast<bool>(CONDITION)>::MSG) {} + + #endif + + #endif // not CXX0X + +#else // EIGEN_NO_STATIC_ASSERT + + #define EIGEN_STATIC_ASSERT(CONDITION,MSG) eigen_assert((CONDITION) && #MSG); + +#endif // EIGEN_NO_STATIC_ASSERT + + +// static assertion failing if the type \a TYPE is not a vector type +#define EIGEN_STATIC_ASSERT_VECTOR_ONLY(TYPE) \ + EIGEN_STATIC_ASSERT(TYPE::IsVectorAtCompileTime, \ + YOU_TRIED_CALLING_A_VECTOR_METHOD_ON_A_MATRIX) + +// static assertion failing if the type \a TYPE is not fixed-size +#define EIGEN_STATIC_ASSERT_FIXED_SIZE(TYPE) \ + EIGEN_STATIC_ASSERT(TYPE::SizeAtCompileTime!=Eigen::Dynamic, \ + YOU_CALLED_A_FIXED_SIZE_METHOD_ON_A_DYNAMIC_SIZE_MATRIX_OR_VECTOR) + +// static assertion failing if the type \a TYPE is not dynamic-size +#define EIGEN_STATIC_ASSERT_DYNAMIC_SIZE(TYPE) \ + EIGEN_STATIC_ASSERT(TYPE::SizeAtCompileTime==Eigen::Dynamic, \ + YOU_CALLED_A_DYNAMIC_SIZE_METHOD_ON_A_FIXED_SIZE_MATRIX_OR_VECTOR) + +// static assertion failing if the type \a TYPE is not a vector type of the given size +#define EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(TYPE, SIZE) \ + EIGEN_STATIC_ASSERT(TYPE::IsVectorAtCompileTime && TYPE::SizeAtCompileTime==SIZE, \ + THIS_METHOD_IS_ONLY_FOR_VECTORS_OF_A_SPECIFIC_SIZE) + +// static assertion failing if the type \a TYPE is not a vector type of the given size +#define EIGEN_STATIC_ASSERT_MATRIX_SPECIFIC_SIZE(TYPE, ROWS, COLS) \ + EIGEN_STATIC_ASSERT(TYPE::RowsAtCompileTime==ROWS && TYPE::ColsAtCompileTime==COLS, \ + THIS_METHOD_IS_ONLY_FOR_MATRICES_OF_A_SPECIFIC_SIZE) + +// static assertion failing if the two vector expression types are not compatible (same fixed-size or dynamic size) +#define EIGEN_STATIC_ASSERT_SAME_VECTOR_SIZE(TYPE0,TYPE1) \ + EIGEN_STATIC_ASSERT( \ + (int(TYPE0::SizeAtCompileTime)==Eigen::Dynamic \ + || int(TYPE1::SizeAtCompileTime)==Eigen::Dynamic \ + || int(TYPE0::SizeAtCompileTime)==int(TYPE1::SizeAtCompileTime)),\ + YOU_MIXED_VECTORS_OF_DIFFERENT_SIZES) + +#define EIGEN_PREDICATE_SAME_MATRIX_SIZE(TYPE0,TYPE1) \ + ( \ + (int(Eigen::internal::size_of_xpr_at_compile_time<TYPE0>::ret)==0 && int(Eigen::internal::size_of_xpr_at_compile_time<TYPE1>::ret)==0) \ + || (\ + (int(TYPE0::RowsAtCompileTime)==Eigen::Dynamic \ + || int(TYPE1::RowsAtCompileTime)==Eigen::Dynamic \ + || int(TYPE0::RowsAtCompileTime)==int(TYPE1::RowsAtCompileTime)) \ + && (int(TYPE0::ColsAtCompileTime)==Eigen::Dynamic \ + || int(TYPE1::ColsAtCompileTime)==Eigen::Dynamic \ + || int(TYPE0::ColsAtCompileTime)==int(TYPE1::ColsAtCompileTime))\ + ) \ + ) + +#define EIGEN_STATIC_ASSERT_NON_INTEGER(TYPE) \ + EIGEN_STATIC_ASSERT(!NumTraits<TYPE>::IsInteger, THIS_FUNCTION_IS_NOT_FOR_INTEGER_NUMERIC_TYPES) + + +// static assertion failing if it is guaranteed at compile-time that the two matrix expression types have different sizes +#define EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(TYPE0,TYPE1) \ + EIGEN_STATIC_ASSERT( \ + EIGEN_PREDICATE_SAME_MATRIX_SIZE(TYPE0,TYPE1),\ + YOU_MIXED_MATRICES_OF_DIFFERENT_SIZES) + +#define EIGEN_STATIC_ASSERT_SIZE_1x1(TYPE) \ + EIGEN_STATIC_ASSERT((TYPE::RowsAtCompileTime == 1 || TYPE::RowsAtCompileTime == Dynamic) && \ + (TYPE::ColsAtCompileTime == 1 || TYPE::ColsAtCompileTime == Dynamic), \ + THIS_METHOD_IS_ONLY_FOR_1x1_EXPRESSIONS) + +#define EIGEN_STATIC_ASSERT_LVALUE(Derived) \ + EIGEN_STATIC_ASSERT(Eigen::internal::is_lvalue<Derived>::value, \ + THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY) + +#define EIGEN_STATIC_ASSERT_ARRAYXPR(Derived) \ + EIGEN_STATIC_ASSERT((Eigen::internal::is_same<typename Eigen::internal::traits<Derived>::XprKind, ArrayXpr>::value), \ + THIS_METHOD_IS_ONLY_FOR_ARRAYS_NOT_MATRICES) + +#define EIGEN_STATIC_ASSERT_SAME_XPR_KIND(Derived1, Derived2) \ + EIGEN_STATIC_ASSERT((Eigen::internal::is_same<typename Eigen::internal::traits<Derived1>::XprKind, \ + typename Eigen::internal::traits<Derived2>::XprKind \ + >::value), \ + YOU_CANNOT_MIX_ARRAYS_AND_MATRICES) + +// Check that a cost value is positive, and that is stay within a reasonable range +// TODO this check could be enabled for internal debugging only +#define EIGEN_INTERNAL_CHECK_COST_VALUE(C) \ + EIGEN_STATIC_ASSERT((C)>=0 && (C)<=HugeCost*HugeCost, EIGEN_INTERNAL_ERROR_PLEASE_FILE_A_BUG_REPORT__INVALID_COST_VALUE); + +#endif // EIGEN_STATIC_ASSERT_H diff --git a/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/XprHelper.h b/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/XprHelper.h new file mode 100644 index 000000000..ba5bd186d --- /dev/null +++ b/runtimes/nn/depend/external/eigen/Eigen/src/Core/util/XprHelper.h @@ -0,0 +1,821 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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_XPRHELPER_H +#define EIGEN_XPRHELPER_H + +// just a workaround because GCC seems to not really like empty structs +// FIXME: gcc 4.3 generates bad code when strict-aliasing is enabled +// so currently we simply disable this optimization for gcc 4.3 +#if EIGEN_COMP_GNUC && !EIGEN_GNUC_AT(4,3) + #define EIGEN_EMPTY_STRUCT_CTOR(X) \ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE X() {} \ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE X(const X& ) {} +#else + #define EIGEN_EMPTY_STRUCT_CTOR(X) +#endif + +namespace Eigen { + +namespace internal { + +template<typename IndexDest, typename IndexSrc> +EIGEN_DEVICE_FUNC +inline IndexDest convert_index(const IndexSrc& idx) { + // for sizeof(IndexDest)>=sizeof(IndexSrc) compilers should be able to optimize this away: + eigen_internal_assert(idx <= NumTraits<IndexDest>::highest() && "Index value to big for target type"); + return IndexDest(idx); +} + + +// promote_scalar_arg is an helper used in operation between an expression and a scalar, like: +// expression * scalar +// Its role is to determine how the type T of the scalar operand should be promoted given the scalar type ExprScalar of the given expression. +// The IsSupported template parameter must be provided by the caller as: internal::has_ReturnType<ScalarBinaryOpTraits<ExprScalar,T,op> >::value using the proper order for ExprScalar and T. +// Then the logic is as follows: +// - if the operation is natively supported as defined by IsSupported, then the scalar type is not promoted, and T is returned. +// - otherwise, NumTraits<ExprScalar>::Literal is returned if T is implicitly convertible to NumTraits<ExprScalar>::Literal AND that this does not imply a float to integer conversion. +// - otherwise, ExprScalar is returned if T is implicitly convertible to ExprScalar AND that this does not imply a float to integer conversion. +// - In all other cases, the promoted type is not defined, and the respective operation is thus invalid and not available (SFINAE). +template<typename ExprScalar,typename T, bool IsSupported> +struct promote_scalar_arg; + +template<typename S,typename T> +struct promote_scalar_arg<S,T,true> +{ + typedef T type; +}; + +// Recursively check safe conversion to PromotedType, and then ExprScalar if they are different. +template<typename ExprScalar,typename T,typename PromotedType, + bool ConvertibleToLiteral = internal::is_convertible<T,PromotedType>::value, + bool IsSafe = NumTraits<T>::IsInteger || !NumTraits<PromotedType>::IsInteger> +struct promote_scalar_arg_unsupported; + +// Start recursion with NumTraits<ExprScalar>::Literal +template<typename S,typename T> +struct promote_scalar_arg<S,T,false> : promote_scalar_arg_unsupported<S,T,typename NumTraits<S>::Literal> {}; + +// We found a match! +template<typename S,typename T, typename PromotedType> +struct promote_scalar_arg_unsupported<S,T,PromotedType,true,true> +{ + typedef PromotedType type; +}; + +// No match, but no real-to-integer issues, and ExprScalar and current PromotedType are different, +// so let's try to promote to ExprScalar +template<typename ExprScalar,typename T, typename PromotedType> +struct promote_scalar_arg_unsupported<ExprScalar,T,PromotedType,false,true> + : promote_scalar_arg_unsupported<ExprScalar,T,ExprScalar> +{}; + +// Unsafe real-to-integer, let's stop. +template<typename S,typename T, typename PromotedType, bool ConvertibleToLiteral> +struct promote_scalar_arg_unsupported<S,T,PromotedType,ConvertibleToLiteral,false> {}; + +// T is not even convertible to ExprScalar, let's stop. +template<typename S,typename T> +struct promote_scalar_arg_unsupported<S,T,S,false,true> {}; + +//classes inheriting no_assignment_operator don't generate a default operator=. +class no_assignment_operator +{ + private: + no_assignment_operator& operator=(const no_assignment_operator&); +}; + +/** \internal return the index type with the largest number of bits */ +template<typename I1, typename I2> +struct promote_index_type +{ + typedef typename conditional<(sizeof(I1)<sizeof(I2)), I2, I1>::type type; +}; + +/** \internal If the template parameter Value is Dynamic, this class is just a wrapper around a T variable that + * can be accessed using value() and setValue(). + * Otherwise, this class is an empty structure and value() just returns the template parameter Value. + */ +template<typename T, int Value> class variable_if_dynamic +{ + public: + EIGEN_EMPTY_STRUCT_CTOR(variable_if_dynamic) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit variable_if_dynamic(T v) { EIGEN_ONLY_USED_FOR_DEBUG(v); eigen_assert(v == T(Value)); } + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE T value() { return T(Value); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void setValue(T) {} +}; + +template<typename T> class variable_if_dynamic<T, Dynamic> +{ + T m_value; + EIGEN_DEVICE_FUNC variable_if_dynamic() { eigen_assert(false); } + public: + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit variable_if_dynamic(T value) : m_value(value) {} + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T value() const { return m_value; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void setValue(T value) { m_value = value; } +}; + +/** \internal like variable_if_dynamic but for DynamicIndex + */ +template<typename T, int Value> class variable_if_dynamicindex +{ + public: + EIGEN_EMPTY_STRUCT_CTOR(variable_if_dynamicindex) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit variable_if_dynamicindex(T v) { EIGEN_ONLY_USED_FOR_DEBUG(v); eigen_assert(v == T(Value)); } + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE T value() { return T(Value); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void setValue(T) {} +}; + +template<typename T> class variable_if_dynamicindex<T, DynamicIndex> +{ + T m_value; + EIGEN_DEVICE_FUNC variable_if_dynamicindex() { eigen_assert(false); } + public: + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit variable_if_dynamicindex(T value) : m_value(value) {} + EIGEN_DEVICE_FUNC T EIGEN_STRONG_INLINE value() const { return m_value; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void setValue(T value) { m_value = value; } +}; + +template<typename T> struct functor_traits +{ + enum + { + Cost = 10, + PacketAccess = false, + IsRepeatable = false + }; +}; + +template<typename T> struct packet_traits; + +template<typename T> struct unpacket_traits +{ + typedef T type; + typedef T half; + enum + { + size = 1, + alignment = 1 + }; +}; + +template<int Size, typename PacketType, + bool Stop = Size==Dynamic || (Size%unpacket_traits<PacketType>::size)==0 || is_same<PacketType,typename unpacket_traits<PacketType>::half>::value> +struct find_best_packet_helper; + +template< int Size, typename PacketType> +struct find_best_packet_helper<Size,PacketType,true> +{ + typedef PacketType type; +}; + +template<int Size, typename PacketType> +struct find_best_packet_helper<Size,PacketType,false> +{ + typedef typename find_best_packet_helper<Size,typename unpacket_traits<PacketType>::half>::type type; +}; + +template<typename T, int Size> +struct find_best_packet +{ + typedef typename find_best_packet_helper<Size,typename packet_traits<T>::type>::type type; +}; + +#if EIGEN_MAX_STATIC_ALIGN_BYTES>0 +template<int ArrayBytes, int AlignmentBytes, + bool Match = bool((ArrayBytes%AlignmentBytes)==0), + bool TryHalf = bool(EIGEN_MIN_ALIGN_BYTES<AlignmentBytes) > +struct compute_default_alignment_helper +{ + enum { value = 0 }; +}; + +template<int ArrayBytes, int AlignmentBytes, bool TryHalf> +struct compute_default_alignment_helper<ArrayBytes, AlignmentBytes, true, TryHalf> // Match +{ + enum { value = AlignmentBytes }; +}; + +template<int ArrayBytes, int AlignmentBytes> +struct compute_default_alignment_helper<ArrayBytes, AlignmentBytes, false, true> // Try-half +{ + // current packet too large, try with an half-packet + enum { value = compute_default_alignment_helper<ArrayBytes, AlignmentBytes/2>::value }; +}; +#else +// If static alignment is disabled, no need to bother. +// This also avoids a division by zero in "bool Match = bool((ArrayBytes%AlignmentBytes)==0)" +template<int ArrayBytes, int AlignmentBytes> +struct compute_default_alignment_helper +{ + enum { value = 0 }; +}; +#endif + +template<typename T, int Size> struct compute_default_alignment { + enum { value = compute_default_alignment_helper<Size*sizeof(T),EIGEN_MAX_STATIC_ALIGN_BYTES>::value }; +}; + +template<typename T> struct compute_default_alignment<T,Dynamic> { + enum { value = EIGEN_MAX_ALIGN_BYTES }; +}; + +template<typename _Scalar, int _Rows, int _Cols, + int _Options = AutoAlign | + ( (_Rows==1 && _Cols!=1) ? RowMajor + : (_Cols==1 && _Rows!=1) ? ColMajor + : EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION ), + int _MaxRows = _Rows, + int _MaxCols = _Cols +> class make_proper_matrix_type +{ + enum { + IsColVector = _Cols==1 && _Rows!=1, + IsRowVector = _Rows==1 && _Cols!=1, + Options = IsColVector ? (_Options | ColMajor) & ~RowMajor + : IsRowVector ? (_Options | RowMajor) & ~ColMajor + : _Options + }; + public: + typedef Matrix<_Scalar, _Rows, _Cols, Options, _MaxRows, _MaxCols> type; +}; + +template<typename Scalar, int Rows, int Cols, int Options, int MaxRows, int MaxCols> +class compute_matrix_flags +{ + enum { row_major_bit = Options&RowMajor ? RowMajorBit : 0 }; + public: + // FIXME currently we still have to handle DirectAccessBit at the expression level to handle DenseCoeffsBase<> + // and then propagate this information to the evaluator's flags. + // However, I (Gael) think that DirectAccessBit should only matter at the evaluation stage. + enum { ret = DirectAccessBit | LvalueBit | NestByRefBit | row_major_bit }; +}; + +template<int _Rows, int _Cols> struct size_at_compile_time +{ + enum { ret = (_Rows==Dynamic || _Cols==Dynamic) ? Dynamic : _Rows * _Cols }; +}; + +template<typename XprType> struct size_of_xpr_at_compile_time +{ + enum { ret = size_at_compile_time<traits<XprType>::RowsAtCompileTime,traits<XprType>::ColsAtCompileTime>::ret }; +}; + +/* plain_matrix_type : the difference from eval is that plain_matrix_type is always a plain matrix type, + * whereas eval is a const reference in the case of a matrix + */ + +template<typename T, typename StorageKind = typename traits<T>::StorageKind> struct plain_matrix_type; +template<typename T, typename BaseClassType, int Flags> struct plain_matrix_type_dense; +template<typename T> struct plain_matrix_type<T,Dense> +{ + typedef typename plain_matrix_type_dense<T,typename traits<T>::XprKind, traits<T>::Flags>::type type; +}; +template<typename T> struct plain_matrix_type<T,DiagonalShape> +{ + typedef typename T::PlainObject type; +}; + +template<typename T, int Flags> struct plain_matrix_type_dense<T,MatrixXpr,Flags> +{ + typedef Matrix<typename traits<T>::Scalar, + traits<T>::RowsAtCompileTime, + traits<T>::ColsAtCompileTime, + AutoAlign | (Flags&RowMajorBit ? RowMajor : ColMajor), + traits<T>::MaxRowsAtCompileTime, + traits<T>::MaxColsAtCompileTime + > type; +}; + +template<typename T, int Flags> struct plain_matrix_type_dense<T,ArrayXpr,Flags> +{ + typedef Array<typename traits<T>::Scalar, + traits<T>::RowsAtCompileTime, + traits<T>::ColsAtCompileTime, + AutoAlign | (Flags&RowMajorBit ? RowMajor : ColMajor), + traits<T>::MaxRowsAtCompileTime, + traits<T>::MaxColsAtCompileTime + > type; +}; + +/* eval : the return type of eval(). For matrices, this is just a const reference + * in order to avoid a useless copy + */ + +template<typename T, typename StorageKind = typename traits<T>::StorageKind> struct eval; + +template<typename T> struct eval<T,Dense> +{ + typedef typename plain_matrix_type<T>::type type; +// typedef typename T::PlainObject type; +// typedef T::Matrix<typename traits<T>::Scalar, +// traits<T>::RowsAtCompileTime, +// traits<T>::ColsAtCompileTime, +// AutoAlign | (traits<T>::Flags&RowMajorBit ? RowMajor : ColMajor), +// traits<T>::MaxRowsAtCompileTime, +// traits<T>::MaxColsAtCompileTime +// > type; +}; + +template<typename T> struct eval<T,DiagonalShape> +{ + typedef typename plain_matrix_type<T>::type type; +}; + +// for matrices, no need to evaluate, just use a const reference to avoid a useless copy +template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols> +struct eval<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>, Dense> +{ + typedef const Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>& type; +}; + +template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols> +struct eval<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>, Dense> +{ + typedef const Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>& type; +}; + + +/* similar to plain_matrix_type, but using the evaluator's Flags */ +template<typename T, typename StorageKind = typename traits<T>::StorageKind> struct plain_object_eval; + +template<typename T> +struct plain_object_eval<T,Dense> +{ + typedef typename plain_matrix_type_dense<T,typename traits<T>::XprKind, evaluator<T>::Flags>::type type; +}; + + +/* plain_matrix_type_column_major : same as plain_matrix_type but guaranteed to be column-major + */ +template<typename T> struct plain_matrix_type_column_major +{ + enum { Rows = traits<T>::RowsAtCompileTime, + Cols = traits<T>::ColsAtCompileTime, + MaxRows = traits<T>::MaxRowsAtCompileTime, + MaxCols = traits<T>::MaxColsAtCompileTime + }; + typedef Matrix<typename traits<T>::Scalar, + Rows, + Cols, + (MaxRows==1&&MaxCols!=1) ? RowMajor : ColMajor, + MaxRows, + MaxCols + > type; +}; + +/* plain_matrix_type_row_major : same as plain_matrix_type but guaranteed to be row-major + */ +template<typename T> struct plain_matrix_type_row_major +{ + enum { Rows = traits<T>::RowsAtCompileTime, + Cols = traits<T>::ColsAtCompileTime, + MaxRows = traits<T>::MaxRowsAtCompileTime, + MaxCols = traits<T>::MaxColsAtCompileTime + }; + typedef Matrix<typename traits<T>::Scalar, + Rows, + Cols, + (MaxCols==1&&MaxRows!=1) ? RowMajor : ColMajor, + MaxRows, + MaxCols + > type; +}; + +/** \internal The reference selector for template expressions. The idea is that we don't + * need to use references for expressions since they are light weight proxy + * objects which should generate no copying overhead. */ +template <typename T> +struct ref_selector +{ + typedef typename conditional< + bool(traits<T>::Flags & NestByRefBit), + T const&, + const T + >::type type; + + typedef typename conditional< + bool(traits<T>::Flags & NestByRefBit), + T &, + T + >::type non_const_type; +}; + +/** \internal Adds the const qualifier on the value-type of T2 if and only if T1 is a const type */ +template<typename T1, typename T2> +struct transfer_constness +{ + typedef typename conditional< + bool(internal::is_const<T1>::value), + typename internal::add_const_on_value_type<T2>::type, + T2 + >::type type; +}; + + +// However, we still need a mechanism to detect whether an expression which is evaluated multiple time +// has to be evaluated into a temporary. +// That's the purpose of this new nested_eval helper: +/** \internal Determines how a given expression should be nested when evaluated multiple times. + * For example, when you do a * (b+c), Eigen will determine how the expression b+c should be + * evaluated into the bigger product expression. The choice is between nesting the expression b+c as-is, or + * evaluating that expression b+c into a temporary variable d, and nest d so that the resulting expression is + * a*d. Evaluating can be beneficial for example if every coefficient access in the resulting expression causes + * many coefficient accesses in the nested expressions -- as is the case with matrix product for example. + * + * \tparam T the type of the expression being nested. + * \tparam n the number of coefficient accesses in the nested expression for each coefficient access in the bigger expression. + * \tparam PlainObject the type of the temporary if needed. + */ +template<typename T, int n, typename PlainObject = typename plain_object_eval<T>::type> struct nested_eval +{ + enum { + ScalarReadCost = NumTraits<typename traits<T>::Scalar>::ReadCost, + CoeffReadCost = evaluator<T>::CoeffReadCost, // NOTE What if an evaluator evaluate itself into a tempory? + // Then CoeffReadCost will be small (e.g., 1) but we still have to evaluate, especially if n>1. + // This situation is already taken care by the EvalBeforeNestingBit flag, which is turned ON + // for all evaluator creating a temporary. This flag is then propagated by the parent evaluators. + // Another solution could be to count the number of temps? + NAsInteger = n == Dynamic ? HugeCost : n, + CostEval = (NAsInteger+1) * ScalarReadCost + CoeffReadCost, + CostNoEval = NAsInteger * CoeffReadCost, + Evaluate = (int(evaluator<T>::Flags) & EvalBeforeNestingBit) || (int(CostEval) < int(CostNoEval)) + }; + + typedef typename conditional<Evaluate, PlainObject, typename ref_selector<T>::type>::type type; +}; + +template<typename T> +EIGEN_DEVICE_FUNC +inline T* const_cast_ptr(const T* ptr) +{ + return const_cast<T*>(ptr); +} + +template<typename Derived, typename XprKind = typename traits<Derived>::XprKind> +struct dense_xpr_base +{ + /* dense_xpr_base should only ever be used on dense expressions, thus falling either into the MatrixXpr or into the ArrayXpr cases */ +}; + +template<typename Derived> +struct dense_xpr_base<Derived, MatrixXpr> +{ + typedef MatrixBase<Derived> type; +}; + +template<typename Derived> +struct dense_xpr_base<Derived, ArrayXpr> +{ + typedef ArrayBase<Derived> type; +}; + +template<typename Derived, typename XprKind = typename traits<Derived>::XprKind, typename StorageKind = typename traits<Derived>::StorageKind> +struct generic_xpr_base; + +template<typename Derived, typename XprKind> +struct generic_xpr_base<Derived, XprKind, Dense> +{ + typedef typename dense_xpr_base<Derived,XprKind>::type type; +}; + +template<typename XprType, typename CastType> struct cast_return_type +{ + typedef typename XprType::Scalar CurrentScalarType; + typedef typename remove_all<CastType>::type _CastType; + typedef typename _CastType::Scalar NewScalarType; + typedef typename conditional<is_same<CurrentScalarType,NewScalarType>::value, + const XprType&,CastType>::type type; +}; + +template <typename A, typename B> struct promote_storage_type; + +template <typename A> struct promote_storage_type<A,A> +{ + typedef A ret; +}; +template <typename A> struct promote_storage_type<A, const A> +{ + typedef A ret; +}; +template <typename A> struct promote_storage_type<const A, A> +{ + typedef A ret; +}; + +/** \internal Specify the "storage kind" of applying a coefficient-wise + * binary operations between two expressions of kinds A and B respectively. + * The template parameter Functor permits to specialize the resulting storage kind wrt to + * the functor. + * The default rules are as follows: + * \code + * A op A -> A + * A op dense -> dense + * dense op B -> dense + * sparse op dense -> sparse + * dense op sparse -> sparse + * \endcode + */ +template <typename A, typename B, typename Functor> struct cwise_promote_storage_type; + +template <typename A, typename Functor> struct cwise_promote_storage_type<A,A,Functor> { typedef A ret; }; +template <typename Functor> struct cwise_promote_storage_type<Dense,Dense,Functor> { typedef Dense ret; }; +template <typename A, typename Functor> struct cwise_promote_storage_type<A,Dense,Functor> { typedef Dense ret; }; +template <typename B, typename Functor> struct cwise_promote_storage_type<Dense,B,Functor> { typedef Dense ret; }; +template <typename Functor> struct cwise_promote_storage_type<Sparse,Dense,Functor> { typedef Sparse ret; }; +template <typename Functor> struct cwise_promote_storage_type<Dense,Sparse,Functor> { typedef Sparse ret; }; + +template <typename LhsKind, typename RhsKind, int LhsOrder, int RhsOrder> struct cwise_promote_storage_order { + enum { value = LhsOrder }; +}; + +template <typename LhsKind, int LhsOrder, int RhsOrder> struct cwise_promote_storage_order<LhsKind,Sparse,LhsOrder,RhsOrder> { enum { value = RhsOrder }; }; +template <typename RhsKind, int LhsOrder, int RhsOrder> struct cwise_promote_storage_order<Sparse,RhsKind,LhsOrder,RhsOrder> { enum { value = LhsOrder }; }; +template <int Order> struct cwise_promote_storage_order<Sparse,Sparse,Order,Order> { enum { value = Order }; }; + + +/** \internal Specify the "storage kind" of multiplying an expression of kind A with kind B. + * The template parameter ProductTag permits to specialize the resulting storage kind wrt to + * some compile-time properties of the product: GemmProduct, GemvProduct, OuterProduct, InnerProduct. + * The default rules are as follows: + * \code + * K * K -> K + * dense * K -> dense + * K * dense -> dense + * diag * K -> K + * K * diag -> K + * Perm * K -> K + * K * Perm -> K + * \endcode + */ +template <typename A, typename B, int ProductTag> struct product_promote_storage_type; + +template <typename A, int ProductTag> struct product_promote_storage_type<A, A, ProductTag> { typedef A ret;}; +template <int ProductTag> struct product_promote_storage_type<Dense, Dense, ProductTag> { typedef Dense ret;}; +template <typename A, int ProductTag> struct product_promote_storage_type<A, Dense, ProductTag> { typedef Dense ret; }; +template <typename B, int ProductTag> struct product_promote_storage_type<Dense, B, ProductTag> { typedef Dense ret; }; + +template <typename A, int ProductTag> struct product_promote_storage_type<A, DiagonalShape, ProductTag> { typedef A ret; }; +template <typename B, int ProductTag> struct product_promote_storage_type<DiagonalShape, B, ProductTag> { typedef B ret; }; +template <int ProductTag> struct product_promote_storage_type<Dense, DiagonalShape, ProductTag> { typedef Dense ret; }; +template <int ProductTag> struct product_promote_storage_type<DiagonalShape, Dense, ProductTag> { typedef Dense ret; }; + +template <typename A, int ProductTag> struct product_promote_storage_type<A, PermutationStorage, ProductTag> { typedef A ret; }; +template <typename B, int ProductTag> struct product_promote_storage_type<PermutationStorage, B, ProductTag> { typedef B ret; }; +template <int ProductTag> struct product_promote_storage_type<Dense, PermutationStorage, ProductTag> { typedef Dense ret; }; +template <int ProductTag> struct product_promote_storage_type<PermutationStorage, Dense, ProductTag> { typedef Dense ret; }; + +/** \internal gives the plain matrix or array type to store a row/column/diagonal of a matrix type. + * \tparam Scalar optional parameter allowing to pass a different scalar type than the one of the MatrixType. + */ +template<typename ExpressionType, typename Scalar = typename ExpressionType::Scalar> +struct plain_row_type +{ + typedef Matrix<Scalar, 1, ExpressionType::ColsAtCompileTime, + ExpressionType::PlainObject::Options | RowMajor, 1, ExpressionType::MaxColsAtCompileTime> MatrixRowType; + typedef Array<Scalar, 1, ExpressionType::ColsAtCompileTime, + ExpressionType::PlainObject::Options | RowMajor, 1, ExpressionType::MaxColsAtCompileTime> ArrayRowType; + + typedef typename conditional< + is_same< typename traits<ExpressionType>::XprKind, MatrixXpr >::value, + MatrixRowType, + ArrayRowType + >::type type; +}; + +template<typename ExpressionType, typename Scalar = typename ExpressionType::Scalar> +struct plain_col_type +{ + typedef Matrix<Scalar, ExpressionType::RowsAtCompileTime, 1, + ExpressionType::PlainObject::Options & ~RowMajor, ExpressionType::MaxRowsAtCompileTime, 1> MatrixColType; + typedef Array<Scalar, ExpressionType::RowsAtCompileTime, 1, + ExpressionType::PlainObject::Options & ~RowMajor, ExpressionType::MaxRowsAtCompileTime, 1> ArrayColType; + + typedef typename conditional< + is_same< typename traits<ExpressionType>::XprKind, MatrixXpr >::value, + MatrixColType, + ArrayColType + >::type type; +}; + +template<typename ExpressionType, typename Scalar = typename ExpressionType::Scalar> +struct plain_diag_type +{ + enum { diag_size = EIGEN_SIZE_MIN_PREFER_DYNAMIC(ExpressionType::RowsAtCompileTime, ExpressionType::ColsAtCompileTime), + max_diag_size = EIGEN_SIZE_MIN_PREFER_FIXED(ExpressionType::MaxRowsAtCompileTime, ExpressionType::MaxColsAtCompileTime) + }; + typedef Matrix<Scalar, diag_size, 1, ExpressionType::PlainObject::Options & ~RowMajor, max_diag_size, 1> MatrixDiagType; + typedef Array<Scalar, diag_size, 1, ExpressionType::PlainObject::Options & ~RowMajor, max_diag_size, 1> ArrayDiagType; + + typedef typename conditional< + is_same< typename traits<ExpressionType>::XprKind, MatrixXpr >::value, + MatrixDiagType, + ArrayDiagType + >::type type; +}; + +template<typename Expr,typename Scalar = typename Expr::Scalar> +struct plain_constant_type +{ + enum { Options = (traits<Expr>::Flags&RowMajorBit)?RowMajor:0 }; + + typedef Array<Scalar, traits<Expr>::RowsAtCompileTime, traits<Expr>::ColsAtCompileTime, + Options, traits<Expr>::MaxRowsAtCompileTime,traits<Expr>::MaxColsAtCompileTime> array_type; + + typedef Matrix<Scalar, traits<Expr>::RowsAtCompileTime, traits<Expr>::ColsAtCompileTime, + Options, traits<Expr>::MaxRowsAtCompileTime,traits<Expr>::MaxColsAtCompileTime> matrix_type; + + typedef CwiseNullaryOp<scalar_constant_op<Scalar>, const typename conditional<is_same< typename traits<Expr>::XprKind, MatrixXpr >::value, matrix_type, array_type>::type > type; +}; + +template<typename ExpressionType> +struct is_lvalue +{ + enum { value = (!bool(is_const<ExpressionType>::value)) && + bool(traits<ExpressionType>::Flags & LvalueBit) }; +}; + +template<typename T> struct is_diagonal +{ enum { ret = false }; }; + +template<typename T> struct is_diagonal<DiagonalBase<T> > +{ enum { ret = true }; }; + +template<typename T> struct is_diagonal<DiagonalWrapper<T> > +{ enum { ret = true }; }; + +template<typename T, int S> struct is_diagonal<DiagonalMatrix<T,S> > +{ enum { ret = true }; }; + +template<typename S1, typename S2> struct glue_shapes; +template<> struct glue_shapes<DenseShape,TriangularShape> { typedef TriangularShape type; }; + +template<typename T1, typename T2> +bool is_same_dense(const T1 &mat1, const T2 &mat2, typename enable_if<has_direct_access<T1>::ret&&has_direct_access<T2>::ret, T1>::type * = 0) +{ + return (mat1.data()==mat2.data()) && (mat1.innerStride()==mat2.innerStride()) && (mat1.outerStride()==mat2.outerStride()); +} + +template<typename T1, typename T2> +bool is_same_dense(const T1 &, const T2 &, typename enable_if<!(has_direct_access<T1>::ret&&has_direct_access<T2>::ret), T1>::type * = 0) +{ + return false; +} + +// Internal helper defining the cost of a scalar division for the type T. +// The default heuristic can be specialized for each scalar type and architecture. +template<typename T,bool Vectorized=false,typename EnaleIf = void> +struct scalar_div_cost { + enum { value = 8*NumTraits<T>::MulCost }; +}; + +template<typename T,bool Vectorized> +struct scalar_div_cost<std::complex<T>, Vectorized> { + enum { value = 2*scalar_div_cost<T>::value + + 6*NumTraits<T>::MulCost + + 3*NumTraits<T>::AddCost + }; +}; + + +template<bool Vectorized> +struct scalar_div_cost<signed long,Vectorized,typename conditional<sizeof(long)==8,void,false_type>::type> { enum { value = 24 }; }; +template<bool Vectorized> +struct scalar_div_cost<unsigned long,Vectorized,typename conditional<sizeof(long)==8,void,false_type>::type> { enum { value = 21 }; }; + + +#ifdef EIGEN_DEBUG_ASSIGN +std::string demangle_traversal(int t) +{ + if(t==DefaultTraversal) return "DefaultTraversal"; + if(t==LinearTraversal) return "LinearTraversal"; + if(t==InnerVectorizedTraversal) return "InnerVectorizedTraversal"; + if(t==LinearVectorizedTraversal) return "LinearVectorizedTraversal"; + if(t==SliceVectorizedTraversal) return "SliceVectorizedTraversal"; + return "?"; +} +std::string demangle_unrolling(int t) +{ + if(t==NoUnrolling) return "NoUnrolling"; + if(t==InnerUnrolling) return "InnerUnrolling"; + if(t==CompleteUnrolling) return "CompleteUnrolling"; + return "?"; +} +std::string demangle_flags(int f) +{ + std::string res; + if(f&RowMajorBit) res += " | RowMajor"; + if(f&PacketAccessBit) res += " | Packet"; + if(f&LinearAccessBit) res += " | Linear"; + if(f&LvalueBit) res += " | Lvalue"; + if(f&DirectAccessBit) res += " | Direct"; + if(f&NestByRefBit) res += " | NestByRef"; + if(f&NoPreferredStorageOrderBit) res += " | NoPreferredStorageOrderBit"; + + return res; +} +#endif + +} // end namespace internal + + +/** \class ScalarBinaryOpTraits + * \ingroup Core_Module + * + * \brief Determines whether the given binary operation of two numeric types is allowed and what the scalar return type is. + * + * This class permits to control the scalar return type of any binary operation performed on two different scalar types through (partial) template specializations. + * + * For instance, let \c U1, \c U2 and \c U3 be three user defined scalar types for which most operations between instances of \c U1 and \c U2 returns an \c U3. + * You can let %Eigen knows that by defining: + \code + template<typename BinaryOp> + struct ScalarBinaryOpTraits<U1,U2,BinaryOp> { typedef U3 ReturnType; }; + template<typename BinaryOp> + struct ScalarBinaryOpTraits<U2,U1,BinaryOp> { typedef U3 ReturnType; }; + \endcode + * You can then explicitly disable some particular operations to get more explicit error messages: + \code + template<> + struct ScalarBinaryOpTraits<U1,U2,internal::scalar_max_op<U1,U2> > {}; + \endcode + * Or customize the return type for individual operation: + \code + template<> + struct ScalarBinaryOpTraits<U1,U2,internal::scalar_sum_op<U1,U2> > { typedef U1 ReturnType; }; + \endcode + * + * By default, the following generic combinations are supported: + <table class="manual"> + <tr><th>ScalarA</th><th>ScalarB</th><th>BinaryOp</th><th>ReturnType</th><th>Note</th></tr> + <tr ><td>\c T </td><td>\c T </td><td>\c * </td><td>\c T </td><td></td></tr> + <tr class="alt"><td>\c NumTraits<T>::Real </td><td>\c T </td><td>\c * </td><td>\c T </td><td>Only if \c NumTraits<T>::IsComplex </td></tr> + <tr ><td>\c T </td><td>\c NumTraits<T>::Real </td><td>\c * </td><td>\c T </td><td>Only if \c NumTraits<T>::IsComplex </td></tr> + </table> + * + * \sa CwiseBinaryOp + */ +template<typename ScalarA, typename ScalarB, typename BinaryOp=internal::scalar_product_op<ScalarA,ScalarB> > +struct ScalarBinaryOpTraits +#ifndef EIGEN_PARSED_BY_DOXYGEN + // for backward compatibility, use the hints given by the (deprecated) internal::scalar_product_traits class. + : internal::scalar_product_traits<ScalarA,ScalarB> +#endif // EIGEN_PARSED_BY_DOXYGEN +{}; + +template<typename T, typename BinaryOp> +struct ScalarBinaryOpTraits<T,T,BinaryOp> +{ + typedef T ReturnType; +}; + +template <typename T, typename BinaryOp> +struct ScalarBinaryOpTraits<T, typename NumTraits<typename internal::enable_if<NumTraits<T>::IsComplex,T>::type>::Real, BinaryOp> +{ + typedef T ReturnType; +}; +template <typename T, typename BinaryOp> +struct ScalarBinaryOpTraits<typename NumTraits<typename internal::enable_if<NumTraits<T>::IsComplex,T>::type>::Real, T, BinaryOp> +{ + typedef T ReturnType; +}; + +// For Matrix * Permutation +template<typename T, typename BinaryOp> +struct ScalarBinaryOpTraits<T,void,BinaryOp> +{ + typedef T ReturnType; +}; + +// For Permutation * Matrix +template<typename T, typename BinaryOp> +struct ScalarBinaryOpTraits<void,T,BinaryOp> +{ + typedef T ReturnType; +}; + +// for Permutation*Permutation +template<typename BinaryOp> +struct ScalarBinaryOpTraits<void,void,BinaryOp> +{ + typedef void ReturnType; +}; + +// We require Lhs and Rhs to have "compatible" scalar types. +// It is tempting to always allow mixing different types but remember that this is often impossible in the vectorized paths. +// So allowing mixing different types gives very unexpected errors when enabling vectorization, when the user tries to +// add together a float matrix and a double matrix. +#define EIGEN_CHECK_BINARY_COMPATIBILIY(BINOP,LHS,RHS) \ + EIGEN_STATIC_ASSERT((Eigen::internal::has_ReturnType<ScalarBinaryOpTraits<LHS, RHS,BINOP> >::value), \ + YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) + +} // end namespace Eigen + +#endif // EIGEN_XPRHELPER_H |