diff options
174 files changed, 2916 insertions, 1402 deletions
diff --git a/CMakeLists.txt b/CMakeLists.txt index 812997a29..8835ee928 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -98,9 +98,11 @@ else() endif() option(EIGEN_BUILD_BTL "Build benchmark suite" OFF) -if(NOT WIN32) + +# Disable pkgconfig only for native Windows builds +if(NOT WIN32 OR NOT CMAKE_HOST_SYSTEM_NAME MATCHES Windows) option(EIGEN_BUILD_PKGCONFIG "Build pkg-config .pc file for Eigen" ON) -endif(NOT WIN32) +endif() set(CMAKE_INCLUDE_CURRENT_DIR ON) @@ -403,7 +405,7 @@ if(EIGEN_BUILD_PKGCONFIG) install(FILES ${CMAKE_CURRENT_BINARY_DIR}/eigen3.pc DESTINATION ${PKGCONFIG_INSTALL_DIR} ) -endif(EIGEN_BUILD_PKGCONFIG) +endif() add_subdirectory(Eigen) diff --git a/Eigen/CMakeLists.txt b/Eigen/CMakeLists.txt index a92dd6f6c..9eb502b79 100644 --- a/Eigen/CMakeLists.txt +++ b/Eigen/CMakeLists.txt @@ -16,4 +16,4 @@ install(FILES DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen COMPONENT Devel ) -add_subdirectory(src) +install(DIRECTORY src DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen COMPONENT Devel FILES_MATCHING PATTERN "*.h") diff --git a/Eigen/Core b/Eigen/Core index 3d2152acf..d89eee824 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -362,6 +362,7 @@ using std::ptrdiff_t; #include "src/Core/NumTraits.h" #include "src/Core/MathFunctions.h" #include "src/Core/GenericPacketMath.h" +#include "src/Core/MathFunctionsImpl.h" #if defined EIGEN_VECTORIZE_AVX // Use AVX for floats and doubles, SSE for integers diff --git a/Eigen/src/CMakeLists.txt b/Eigen/src/CMakeLists.txt deleted file mode 100644 index c326f374d..000000000 --- a/Eigen/src/CMakeLists.txt +++ /dev/null @@ -1,7 +0,0 @@ -file(GLOB Eigen_src_subdirectories "*") -escape_string_as_regex(ESCAPED_CMAKE_CURRENT_SOURCE_DIR "${CMAKE_CURRENT_SOURCE_DIR}") -foreach(f ${Eigen_src_subdirectories}) - if(NOT f MATCHES "\\.txt" AND NOT f MATCHES "${ESCAPED_CMAKE_CURRENT_SOURCE_DIR}/[.].+" ) - add_subdirectory(${f}) - endif() -endforeach() diff --git a/Eigen/src/Cholesky/CMakeLists.txt b/Eigen/src/Cholesky/CMakeLists.txt deleted file mode 100644 index d01488b41..000000000 --- a/Eigen/src/Cholesky/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_Cholesky_SRCS "*.h") - -INSTALL(FILES - ${Eigen_Cholesky_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Cholesky COMPONENT Devel - ) diff --git a/Eigen/src/Cholesky/LDLT.h b/Eigen/src/Cholesky/LDLT.h index 69e176110..fcee7b2e3 100644 --- a/Eigen/src/Cholesky/LDLT.h +++ b/Eigen/src/Cholesky/LDLT.h @@ -253,7 +253,7 @@ template<typename _MatrixType, int _UpLo> class LDLT ComputationInfo info() const { eigen_assert(m_isInitialized && "LDLT is not initialized."); - return Success; + return m_info; } #ifndef EIGEN_PARSED_BY_DOXYGEN @@ -281,6 +281,7 @@ template<typename _MatrixType, int _UpLo> class LDLT TmpMatrixType m_temporary; internal::SignMatrix m_sign; bool m_isInitialized; + ComputationInfo m_info; }; namespace internal { @@ -298,6 +299,8 @@ template<> struct ldlt_inplace<Lower> typedef typename TranspositionType::StorageIndex IndexType; eigen_assert(mat.rows()==mat.cols()); const Index size = mat.rows(); + bool found_zero_pivot = false; + bool ret = true; if (size <= 1) { @@ -356,9 +359,27 @@ template<> struct ldlt_inplace<Lower> // we should only make sure that we do not introduce INF or NaN values. // Remark that LAPACK also uses 0 as the cutoff value. RealScalar realAkk = numext::real(mat.coeffRef(k,k)); - if((rs>0) && (abs(realAkk) > RealScalar(0))) + bool pivot_is_valid = (abs(realAkk) > RealScalar(0)); + + if(k==0 && !pivot_is_valid) + { + // The entire diagonal is zero, there is nothing more to do + // except filling the transpositions, and checking whether the matrix is zero. + sign = ZeroSign; + for(Index j = 0; j<size; ++j) + { + transpositions.coeffRef(j) = IndexType(j); + ret = ret && (mat.col(j).tail(size-j-1).array()==Scalar(0)).all(); + } + return ret; + } + + if((rs>0) && pivot_is_valid) A21 /= realAkk; + if(found_zero_pivot && pivot_is_valid) ret = false; // factorization failed + else if(!pivot_is_valid) found_zero_pivot = true; + if (sign == PositiveSemiDef) { if (realAkk < static_cast<RealScalar>(0)) sign = Indefinite; } else if (sign == NegativeSemiDef) { @@ -369,7 +390,7 @@ template<> struct ldlt_inplace<Lower> } } - return true; + return ret; } // Reference for the algorithm: Davis and Hager, "Multiple Rank @@ -493,7 +514,7 @@ LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::compute(const EigenBase<InputTyp m_temporary.resize(size); m_sign = internal::ZeroSign; - internal::ldlt_inplace<UpLo>::unblocked(m_matrix, m_transpositions, m_temporary, m_sign); + m_info = internal::ldlt_inplace<UpLo>::unblocked(m_matrix, m_transpositions, m_temporary, m_sign) ? Success : NumericalIssue; m_isInitialized = true; return *this; @@ -621,7 +642,6 @@ MatrixType LDLT<MatrixType,_UpLo>::reconstructedMatrix() const return res; } -#ifndef __CUDACC__ /** \cholesky_module * \returns the Cholesky decomposition with full pivoting without square root of \c *this * \sa MatrixBase::ldlt() @@ -643,7 +663,6 @@ MatrixBase<Derived>::ldlt() const { return LDLT<PlainObject>(derived()); } -#endif // __CUDACC__ } // end namespace Eigen diff --git a/Eigen/src/Cholesky/LLT.h b/Eigen/src/Cholesky/LLT.h index bd966656d..ddf4875ab 100644 --- a/Eigen/src/Cholesky/LLT.h +++ b/Eigen/src/Cholesky/LLT.h @@ -507,7 +507,6 @@ MatrixType LLT<MatrixType,_UpLo>::reconstructedMatrix() const return matrixL() * matrixL().adjoint().toDenseMatrix(); } -#ifndef __CUDACC__ /** \cholesky_module * \returns the LLT decomposition of \c *this * \sa SelfAdjointView::llt() @@ -529,7 +528,6 @@ SelfAdjointView<MatrixType, UpLo>::llt() const { return LLT<PlainObject,UpLo>(m_matrix); } -#endif // __CUDACC__ } // end namespace Eigen diff --git a/Eigen/src/CholmodSupport/CMakeLists.txt b/Eigen/src/CholmodSupport/CMakeLists.txt deleted file mode 100644 index 814dfa613..000000000 --- a/Eigen/src/CholmodSupport/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_CholmodSupport_SRCS "*.h") - -INSTALL(FILES - ${Eigen_CholmodSupport_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/CholmodSupport COMPONENT Devel - ) diff --git a/Eigen/src/Core/Array.h b/Eigen/src/Core/Array.h index 7c2e0de16..0d34269fd 100644 --- a/Eigen/src/Core/Array.h +++ b/Eigen/src/Core/Array.h @@ -37,7 +37,7 @@ struct traits<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> > : tra * storage layout. * * This class can be extended with the help of the plugin mechanism described on the page - * \ref TopicCustomizingEigen by defining the preprocessor symbol \c EIGEN_ARRAY_PLUGIN. + * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_ARRAY_PLUGIN. * * \sa \blank \ref TutorialArrayClass, \ref TopicClassHierarchy */ diff --git a/Eigen/src/Core/ArrayBase.h b/Eigen/src/Core/ArrayBase.h index 62851a0c2..3a66f0e40 100644 --- a/Eigen/src/Core/ArrayBase.h +++ b/Eigen/src/Core/ArrayBase.h @@ -32,7 +32,7 @@ template<typename ExpressionType> class MatrixWrapper; * \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 TopicCustomizingEigen by defining the preprocessor symbol \c EIGEN_ARRAYBASE_PLUGIN. + * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_ARRAYBASE_PLUGIN. * * \sa class MatrixBase, \ref TopicClassHierarchy */ diff --git a/Eigen/src/Core/CMakeLists.txt b/Eigen/src/Core/CMakeLists.txt deleted file mode 100644 index 38c3afde9..000000000 --- a/Eigen/src/Core/CMakeLists.txt +++ /dev/null @@ -1,11 +0,0 @@ -FILE(GLOB Eigen_Core_SRCS "*.h") - -INSTALL(FILES - ${Eigen_Core_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core COMPONENT Devel - ) - -ADD_SUBDIRECTORY(products) -ADD_SUBDIRECTORY(util) -ADD_SUBDIRECTORY(arch) -ADD_SUBDIRECTORY(functors) diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index 7ba92963c..7a5540593 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -337,6 +337,120 @@ protected: // 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> > @@ -356,41 +470,44 @@ struct evaluator<CwiseNullaryOp<NullaryOp,PlainObjectType> > }; EIGEN_DEVICE_FUNC explicit evaluator(const XprType& n) - : m_functor(n.functor()) + : 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(Index row, Index col) const + CoeffReturnType coeff(IndexType row, IndexType col) const { - return m_functor(row, col); + return m_wrapper(m_functor, row, col); } + template <typename IndexType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE - CoeffReturnType coeff(Index index) const + CoeffReturnType coeff(IndexType index) const { - return m_functor(index); + return m_wrapper(m_functor,index); } - template<int LoadMode, typename PacketType> + template<int LoadMode, typename PacketType, typename IndexType> EIGEN_STRONG_INLINE - PacketType packet(Index row, Index col) const + PacketType packet(IndexType row, IndexType col) const { - return m_functor.template packetOp<Index,PacketType>(row, col); + return m_wrapper.template packetOp<PacketType>(m_functor, row, col); } - template<int LoadMode, typename PacketType> + template<int LoadMode, typename PacketType, typename IndexType> EIGEN_STRONG_INLINE - PacketType packet(Index index) const + PacketType packet(IndexType index) const { - return m_functor.template packetOp<Index,PacketType>(index); + return m_wrapper.template packetOp<PacketType>(m_functor, index); } protected: const NullaryOp m_functor; + const internal::nullary_wrapper<CoeffReturnType,NullaryOp> m_wrapper; }; // -------------------- CwiseUnaryOp -------------------- diff --git a/Eigen/src/Core/CwiseNullaryOp.h b/Eigen/src/Core/CwiseNullaryOp.h index 3c6508cd0..e3f20894d 100644 --- a/Eigen/src/Core/CwiseNullaryOp.h +++ b/Eigen/src/Core/CwiseNullaryOp.h @@ -20,7 +20,8 @@ struct traits<CwiseNullaryOp<NullaryOp, PlainObjectType> > : traits<PlainObjectT Flags = traits<PlainObjectType>::Flags & RowMajorBit }; }; -} + +} // namespace internal /** \class CwiseNullaryOp * \ingroup Core_Module @@ -37,7 +38,23 @@ struct traits<CwiseNullaryOp<NullaryOp, PlainObjectType> > : traits<PlainObjectT * However, if you want to write a function returning such an expression, you * will need to use this class. * - * \sa class CwiseUnaryOp, class CwiseBinaryOp, DenseBase::NullaryExpr() + * 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 @@ -62,30 +79,6 @@ class CwiseNullaryOp : public internal::dense_xpr_base< CwiseNullaryOp<NullaryOp EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index cols() const { return m_cols.value(); } - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE const Scalar coeff(Index rowId, Index colId) const - { - return m_functor(rowId, colId); - } - - template<int LoadMode> - EIGEN_STRONG_INLINE PacketScalar packet(Index rowId, Index colId) const - { - return m_functor.packetOp(rowId, colId); - } - - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE const Scalar coeff(Index index) const - { - return m_functor(index); - } - - template<int LoadMode> - EIGEN_STRONG_INLINE PacketScalar packet(Index index) const - { - return m_functor.packetOp(index); - } - /** \returns the functor representing the nullary operation */ EIGEN_DEVICE_FUNC const NullaryOp& functor() const { return m_functor; } diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h index a60e5cb00..0ede9b041 100644 --- a/Eigen/src/Core/DenseBase.h +++ b/Eigen/src/Core/DenseBase.h @@ -34,7 +34,7 @@ static inline void check_DenseIndex_is_signed() { * \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 TopicCustomizingEigen by defining the preprocessor symbol \c EIGEN_DENSEBASE_PLUGIN. + * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_DENSEBASE_PLUGIN. * * \sa \blank \ref TopicClassHierarchy */ diff --git a/Eigen/src/Core/MapBase.h b/Eigen/src/Core/MapBase.h index c351c6b92..020f939ad 100644 --- a/Eigen/src/Core/MapBase.h +++ b/Eigen/src/Core/MapBase.h @@ -26,7 +26,7 @@ namespace Eigen { * 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 TopicCustomizingEigen customizing Eigen \endlink for details. + * 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 diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h index 9934e5601..bf3044b96 100644 --- a/Eigen/src/Core/MathFunctions.h +++ b/Eigen/src/Core/MathFunctions.h @@ -459,30 +459,33 @@ struct arg_retval /**************************************************************************** * Implementation of log1p * ****************************************************************************/ -template<typename Scalar, bool isComplex = NumTraits<Scalar>::IsComplex > -struct log1p_impl -{ - static EIGEN_DEVICE_FUNC inline Scalar run(const Scalar& x) - { + +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)) ); } -}; +} -#if EIGEN_HAS_CXX11_MATH && !defined(__CUDACC__) template<typename Scalar> -struct log1p_impl<Scalar, false> { +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); } }; -#endif + template<typename Scalar> struct log1p_retval @@ -615,16 +618,18 @@ struct random_default_impl<Scalar, false, true> 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 = (range+RAND_MAX-1)/(range+1); - std::size_t multiplier = (range+RAND_MAX-1)/std::size_t(RAND_MAX); - + 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 ); + offset = (std::size_t(std::rand()) * multiplier) / divisor; } while (offset > range); - return Scalar(ScalarX(x) + offset); } @@ -785,6 +790,8 @@ template<typename T> EIGEN_DEVICE_FUNC bool isfinite_impl(const std::complex<T>& 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 /**************************************************************************** @@ -921,6 +928,14 @@ 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) @@ -1031,6 +1046,16 @@ 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(real(x), imag(x)); +} + +template <> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +double abs(const std::complex<double>& x) { + return ::hypot(real(x), imag(x)); +} #endif template<typename T> @@ -1176,6 +1201,11 @@ T tanh(const T &x) { 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); } diff --git a/Eigen/src/Core/MathFunctionsImpl.h b/Eigen/src/Core/MathFunctionsImpl.h new file mode 100644 index 000000000..0c77ee003 --- /dev/null +++ b/Eigen/src/Core/MathFunctionsImpl.h @@ -0,0 +1,74 @@ +// 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); + 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/Eigen/src/Core/Matrix.h b/Eigen/src/Core/Matrix.h index 502b7935a..90c336d8c 100644 --- a/Eigen/src/Core/Matrix.h +++ b/Eigen/src/Core/Matrix.h @@ -106,7 +106,7 @@ public: * \endcode * * This class can be extended with the help of the plugin mechanism described on the page - * \ref TopicCustomizingEigen by defining the preprocessor symbol \c EIGEN_MATRIX_PLUGIN. + * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_MATRIX_PLUGIN. * * <i><b>Some notes:</b></i> * diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h index d9d2426ad..334a4d71e 100644 --- a/Eigen/src/Core/MatrixBase.h +++ b/Eigen/src/Core/MatrixBase.h @@ -41,7 +41,7 @@ namespace Eigen { * \endcode * * This class can be extended with the help of the plugin mechanism described on the page - * \ref TopicCustomizingEigen by defining the preprocessor symbol \c EIGEN_MATRIXBASE_PLUGIN. + * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_MATRIXBASE_PLUGIN. * * \sa \blank \ref TopicClassHierarchy */ diff --git a/Eigen/src/Core/NumTraits.h b/Eigen/src/Core/NumTraits.h index 42cffbd3b..dd61195bc 100644 --- a/Eigen/src/Core/NumTraits.h +++ b/Eigen/src/Core/NumTraits.h @@ -97,23 +97,6 @@ template<typename T> struct GenericNumTraits MulCost = 1 }; - // Division is messy but important, because it is expensive and throughput - // varies significantly. The following numbers are based on min division - // throughput on Haswell. - template<bool Vectorized> - struct Div { - enum { -#ifdef EIGEN_VECTORIZE_AVX - AVX = true, -#else - AVX = false, -#endif - Cost = IsInteger ? (sizeof(T) == 8 ? (IsSigned ? 24 : 21) : (IsSigned ? 8 : 9)): - Vectorized ? (sizeof(T) == 8 ? (AVX ? 16 : 8) : (AVX ? 14 : 7)) : 8 - }; - }; - - typedef T Real; typedef typename internal::conditional< IsInteger, @@ -255,6 +238,9 @@ private: 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/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h index 64f5eb052..55b4ac057 100644 --- a/Eigen/src/Core/PlainObjectBase.h +++ b/Eigen/src/Core/PlainObjectBase.h @@ -63,7 +63,7 @@ template<typename MatrixTypeA, typename MatrixTypeB, bool SwapPointers> struct m * \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 TopicCustomizingEigen by defining the preprocessor symbol \c EIGEN_PLAINOBJECTBASE_PLUGIN. + * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_PLAINOBJECTBASE_PLUGIN. * * \sa \ref TopicClassHierarchy */ diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index 955668bef..b8f92a3dc 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -194,7 +194,6 @@ struct Assignment<DstXprType, CwiseBinaryOp<internal::scalar_product_op<ScalarBi //---------------------------------------- // 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 -// TODO enable it for "Dense ?= xpr - Product<>" as well. 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, @@ -203,10 +202,9 @@ struct evaluator_assume_aliasing<CwiseBinaryOp<internal::scalar_sum_op<typename }; template<typename DstXprType, typename OtherXpr, typename ProductType, typename Func1, typename Func2> -struct assignment_from_xpr_plus_product +struct assignment_from_xpr_op_product { - typedef CwiseBinaryOp<internal::scalar_sum_op<typename OtherXpr::Scalar,typename ProductType::Scalar>, const OtherXpr, const ProductType> SrcXprType; - template<typename InitialFunc> + template<typename SrcXprType, typename InitialFunc> static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const InitialFunc& /*func*/) { @@ -215,21 +213,21 @@ struct assignment_from_xpr_plus_product } }; -template< typename DstXprType, typename OtherXpr, typename Lhs, typename Rhs, typename DstScalar, typename SrcScalar, typename OtherScalar,typename ProdScalar> -struct Assignment<DstXprType, CwiseBinaryOp<internal::scalar_sum_op<OtherScalar,ProdScalar>, const OtherXpr, - const Product<Lhs,Rhs,DefaultProduct> >, internal::assign_op<DstScalar,SrcScalar>, Dense2Dense> - : assignment_from_xpr_plus_product<DstXprType, OtherXpr, Product<Lhs,Rhs,DefaultProduct>, internal::assign_op<DstScalar,OtherScalar>, internal::add_assign_op<DstScalar,ProdScalar> > -{}; -template< typename DstXprType, typename OtherXpr, typename Lhs, typename Rhs, typename DstScalar, typename SrcScalar, typename OtherScalar,typename ProdScalar> -struct Assignment<DstXprType, CwiseBinaryOp<internal::scalar_sum_op<OtherScalar,ProdScalar>, const OtherXpr, - const Product<Lhs,Rhs,DefaultProduct> >, internal::add_assign_op<DstScalar,SrcScalar>, Dense2Dense> - : assignment_from_xpr_plus_product<DstXprType, OtherXpr, Product<Lhs,Rhs,DefaultProduct>, internal::add_assign_op<DstScalar,OtherScalar>, internal::add_assign_op<DstScalar,ProdScalar> > -{}; -template< typename DstXprType, typename OtherXpr, typename Lhs, typename Rhs, typename DstScalar, typename SrcScalar, typename OtherScalar,typename ProdScalar> -struct Assignment<DstXprType, CwiseBinaryOp<internal::scalar_sum_op<OtherScalar,ProdScalar>, const OtherXpr, - const Product<Lhs,Rhs,DefaultProduct> >, internal::sub_assign_op<DstScalar,SrcScalar>, Dense2Dense> - : assignment_from_xpr_plus_product<DstXprType, OtherXpr, Product<Lhs,Rhs,DefaultProduct>, internal::sub_assign_op<DstScalar,OtherScalar>, internal::sub_assign_op<DstScalar,ProdScalar> > -{}; +#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> @@ -532,8 +530,8 @@ struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape, */ EIGEN_DEVICE_FUNC const CoeffReturnType coeff(Index index) const { - const Index row = RowsAtCompileTime == 1 ? 0 : index; - const Index col = RowsAtCompileTime == 1 ? index : 0; + 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(); } @@ -551,8 +549,8 @@ struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape, template<int LoadMode, typename PacketType> const PacketType packet(Index index) const { - const Index row = RowsAtCompileTime == 1 ? 0 : index; - const Index col = RowsAtCompileTime == 1 ? index : 0; + const Index row = (RowsAtCompileTime == 1 || MaxRowsAtCompileTime==1) ? 0 : index; + const Index col = (RowsAtCompileTime == 1 || MaxRowsAtCompileTime==1) ? index : 0; return packet<LoadMode,PacketType>(row,col); } diff --git a/Eigen/src/Core/Random.h b/Eigen/src/Core/Random.h index 02038e9e3..6faf789c7 100644 --- a/Eigen/src/Core/Random.h +++ b/Eigen/src/Core/Random.h @@ -16,8 +16,7 @@ namespace internal { template<typename Scalar> struct scalar_random_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_random_op) - template<typename Index> - inline const Scalar operator() (Index, Index = 0) const { return random<Scalar>(); } + inline const Scalar operator() () const { return random<Scalar>(); } }; template<typename Scalar> diff --git a/Eigen/src/Core/Ref.h b/Eigen/src/Core/Ref.h index 17065fdd5..bdf24f52a 100644 --- a/Eigen/src/Core/Ref.h +++ b/Eigen/src/Core/Ref.h @@ -35,7 +35,13 @@ struct traits<Ref<_PlainObjectType, _Options, _StrideType> > || (int(StrideType::InnerStrideAtCompileTime)==0 && int(Derived::InnerStrideAtCompileTime)==1), OuterStrideMatch = Derived::IsVectorAtCompileTime || int(StrideType::OuterStrideAtCompileTime)==int(Dynamic) || int(StrideType::OuterStrideAtCompileTime)==int(Derived::OuterStrideAtCompileTime), - AlignmentMatch = (int(traits<PlainObjectType>::Alignment)==int(Unaligned)) || (int(evaluator<Derived>::Alignment) >= int(Alignment)), // FIXME the first condition is not very clear, it should be replaced by the required alignment + // 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 }; diff --git a/Eigen/src/Core/arch/AVX/CMakeLists.txt b/Eigen/src/Core/arch/AVX/CMakeLists.txt deleted file mode 100644 index bdb71ab99..000000000 --- a/Eigen/src/Core/arch/AVX/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_Core_arch_AVX_SRCS "*.h") - -INSTALL(FILES - ${Eigen_Core_arch_AVX_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/arch/AVX COMPONENT Devel -) diff --git a/Eigen/src/Core/arch/AVX/MathFunctions.h b/Eigen/src/Core/arch/AVX/MathFunctions.h index 98d8e029f..d21ec39dd 100644 --- a/Eigen/src/Core/arch/AVX/MathFunctions.h +++ b/Eigen/src/Core/arch/AVX/MathFunctions.h @@ -266,52 +266,10 @@ pexp<Packet8f>(const Packet8f& _x) { } // Hyperbolic Tangent function. -// 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 -// fl(tanh(x)) = +/-1. template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f -ptanh<Packet8f>(const Packet8f& _x) { - // Clamp the inputs to the range [-9, 9] since anything outside - // this range is +/-1.0f in single-precision. - _EIGEN_DECLARE_CONST_Packet8f(plus_9, 9.0f); - _EIGEN_DECLARE_CONST_Packet8f(minus_9, -9.0f); - const Packet8f x = pmax(p8f_minus_9, pmin(p8f_plus_9, _x)); - - // The monomial coefficients of the numerator polynomial (odd). - _EIGEN_DECLARE_CONST_Packet8f(alpha_1, 4.89352455891786e-03f); - _EIGEN_DECLARE_CONST_Packet8f(alpha_3, 6.37261928875436e-04f); - _EIGEN_DECLARE_CONST_Packet8f(alpha_5, 1.48572235717979e-05f); - _EIGEN_DECLARE_CONST_Packet8f(alpha_7, 5.12229709037114e-08f); - _EIGEN_DECLARE_CONST_Packet8f(alpha_9, -8.60467152213735e-11f); - _EIGEN_DECLARE_CONST_Packet8f(alpha_11, 2.00018790482477e-13f); - _EIGEN_DECLARE_CONST_Packet8f(alpha_13, -2.76076847742355e-16f); - - // The monomial coefficients of the denominator polynomial (even). - _EIGEN_DECLARE_CONST_Packet8f(beta_0, 4.89352518554385e-03f); - _EIGEN_DECLARE_CONST_Packet8f(beta_2, 2.26843463243900e-03f); - _EIGEN_DECLARE_CONST_Packet8f(beta_4, 1.18534705686654e-04f); - _EIGEN_DECLARE_CONST_Packet8f(beta_6, 1.19825839466702e-06f); - - // Since the polynomials are odd/even, we need x^2. - const Packet8f x2 = pmul(x, x); - - // Evaluate the numerator polynomial p. - Packet8f p = pmadd(x2, p8f_alpha_13, p8f_alpha_11); - p = pmadd(x2, p, p8f_alpha_9); - p = pmadd(x2, p, p8f_alpha_7); - p = pmadd(x2, p, p8f_alpha_5); - p = pmadd(x2, p, p8f_alpha_3); - p = pmadd(x2, p, p8f_alpha_1); - p = pmul(x, p); - - // Evaluate the denominator polynomial p. - Packet8f q = pmadd(x2, p8f_beta_6, p8f_beta_4); - q = pmadd(x2, q, p8f_beta_2); - q = pmadd(x2, q, p8f_beta_0); - - // Divide the numerator by the denominator. - return pdiv(p, q); +ptanh<Packet8f>(const Packet8f& x) { + return internal::generic_fast_tanh_float(x); } template <> diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index 4fec14f44..dae0ca5d0 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -94,6 +94,9 @@ template<> struct packet_traits<double> : default_packet_traits }; }; +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 @@ -153,7 +156,7 @@ template<> EIGEN_STRONG_INLINE Packet8i pdiv<Packet8i>(const Packet8i& /*a*/, co #ifdef __FMA__ template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f& b, const Packet8f& c) { -#if EIGEN_COMP_GNUC || EIGEN_COMP_CLANG +#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 @@ -166,7 +169,7 @@ template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f& #endif } template<> EIGEN_STRONG_INLINE Packet4d pmadd(const Packet4d& a, const Packet4d& b, const Packet4d& c) { -#if EIGEN_COMP_GNUC || EIGEN_COMP_CLANG +#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)); diff --git a/Eigen/src/Core/arch/AltiVec/CMakeLists.txt b/Eigen/src/Core/arch/AltiVec/CMakeLists.txt deleted file mode 100644 index 9f8d2e9c4..000000000 --- a/Eigen/src/Core/arch/AltiVec/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_Core_arch_AltiVec_SRCS "*.h") - -INSTALL(FILES - ${Eigen_Core_arch_AltiVec_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/arch/AltiVec COMPONENT Devel -) diff --git a/Eigen/src/Core/arch/CMakeLists.txt b/Eigen/src/Core/arch/CMakeLists.txt deleted file mode 100644 index 42b0b486e..000000000 --- a/Eigen/src/Core/arch/CMakeLists.txt +++ /dev/null @@ -1,9 +0,0 @@ -ADD_SUBDIRECTORY(AltiVec) -ADD_SUBDIRECTORY(AVX) -ADD_SUBDIRECTORY(CUDA) -ADD_SUBDIRECTORY(Default) -ADD_SUBDIRECTORY(NEON) -ADD_SUBDIRECTORY(SSE) - - - diff --git a/Eigen/src/Core/arch/CUDA/CMakeLists.txt b/Eigen/src/Core/arch/CUDA/CMakeLists.txt deleted file mode 100644 index 7ba28da7c..000000000 --- a/Eigen/src/Core/arch/CUDA/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_Core_arch_CUDA_SRCS "*.h") - -INSTALL(FILES - ${Eigen_Core_arch_CUDA_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/arch/CUDA COMPONENT Devel -) diff --git a/Eigen/src/Core/arch/CUDA/Half.h b/Eigen/src/Core/arch/CUDA/Half.h index 4d91420d0..52892db38 100644 --- a/Eigen/src/Core/arch/CUDA/Half.h +++ b/Eigen/src/Core/arch/CUDA/Half.h @@ -389,10 +389,14 @@ 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(::log1pf(float(a))); + return half(numext::log1p(float(a))); } EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log10(const half& a) { return half(::log10f(float(a))); @@ -503,7 +507,11 @@ 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))); diff --git a/Eigen/src/Core/arch/Default/CMakeLists.txt b/Eigen/src/Core/arch/Default/CMakeLists.txt deleted file mode 100644 index 339c091d1..000000000 --- a/Eigen/src/Core/arch/Default/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_Core_arch_Default_SRCS "*.h") - -INSTALL(FILES - ${Eigen_Core_arch_Default_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/arch/Default COMPONENT Devel -) diff --git a/Eigen/src/Core/arch/NEON/CMakeLists.txt b/Eigen/src/Core/arch/NEON/CMakeLists.txt deleted file mode 100644 index fd4d4af50..000000000 --- a/Eigen/src/Core/arch/NEON/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_Core_arch_NEON_SRCS "*.h") - -INSTALL(FILES - ${Eigen_Core_arch_NEON_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/arch/NEON COMPONENT Devel -) diff --git a/Eigen/src/Core/arch/SSE/CMakeLists.txt b/Eigen/src/Core/arch/SSE/CMakeLists.txt deleted file mode 100644 index 46ea7cc62..000000000 --- a/Eigen/src/Core/arch/SSE/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_Core_arch_SSE_SRCS "*.h") - -INSTALL(FILES - ${Eigen_Core_arch_SSE_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/arch/SSE COMPONENT Devel -) diff --git a/Eigen/src/Core/arch/SSE/MathFunctions.h b/Eigen/src/Core/arch/SSE/MathFunctions.h index 28f103eeb..ac2fd8103 100644 --- a/Eigen/src/Core/arch/SSE/MathFunctions.h +++ b/Eigen/src/Core/arch/SSE/MathFunctions.h @@ -517,52 +517,10 @@ Packet2d prsqrt<Packet2d>(const Packet2d& x) { } // Hyperbolic Tangent function. -// 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 -// fl(tanh(x)) = +/-1. template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f -ptanh<Packet4f>(const Packet4f& _x) { - // Clamp the inputs to the range [-9, 9] since anything outside - // this range is +/-1.0f in single-precision. - _EIGEN_DECLARE_CONST_Packet4f(plus_9, 9.0f); - _EIGEN_DECLARE_CONST_Packet4f(minus_9, -9.0f); - const Packet4f x = pmax(p4f_minus_9, pmin(p4f_plus_9, _x)); - - // The monomial coefficients of the numerator polynomial (odd). - _EIGEN_DECLARE_CONST_Packet4f(alpha_1, 4.89352455891786e-03f); - _EIGEN_DECLARE_CONST_Packet4f(alpha_3, 6.37261928875436e-04f); - _EIGEN_DECLARE_CONST_Packet4f(alpha_5, 1.48572235717979e-05f); - _EIGEN_DECLARE_CONST_Packet4f(alpha_7, 5.12229709037114e-08f); - _EIGEN_DECLARE_CONST_Packet4f(alpha_9, -8.60467152213735e-11f); - _EIGEN_DECLARE_CONST_Packet4f(alpha_11, 2.00018790482477e-13f); - _EIGEN_DECLARE_CONST_Packet4f(alpha_13, -2.76076847742355e-16f); - - // The monomial coefficients of the denominator polynomial (even). - _EIGEN_DECLARE_CONST_Packet4f(beta_0, 4.89352518554385e-03f); - _EIGEN_DECLARE_CONST_Packet4f(beta_2, 2.26843463243900e-03f); - _EIGEN_DECLARE_CONST_Packet4f(beta_4, 1.18534705686654e-04f); - _EIGEN_DECLARE_CONST_Packet4f(beta_6, 1.19825839466702e-06f); - - // Since the polynomials are odd/even, we need x^2. - const Packet4f x2 = pmul(x, x); - - // Evaluate the numerator polynomial p. - Packet4f p = pmadd(x2, p4f_alpha_13, p4f_alpha_11); - p = pmadd(x2, p, p4f_alpha_9); - p = pmadd(x2, p, p4f_alpha_7); - p = pmadd(x2, p, p4f_alpha_5); - p = pmadd(x2, p, p4f_alpha_3); - p = pmadd(x2, p, p4f_alpha_1); - p = pmul(x, p); - - // Evaluate the denominator polynomial p. - Packet4f q = pmadd(x2, p4f_beta_6, p4f_beta_4); - q = pmadd(x2, q, p4f_beta_2); - q = pmadd(x2, q, p4f_beta_0); - - // Divide the numerator by the denominator. - return pdiv(p, q); +ptanh<Packet4f>(const Packet4f& x) { + return internal::generic_fast_tanh_float(x); } } // end namespace internal diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index 70839d68d..baad692e3 100755 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -162,6 +162,11 @@ template<> struct unpacket_traits<Packet4f> { typedef float type; enum {size=4, 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 @@ -813,6 +818,16 @@ template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, cons #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 diff --git a/Eigen/src/Core/arch/ZVector/CMakeLists.txt b/Eigen/src/Core/arch/ZVector/CMakeLists.txt deleted file mode 100644 index 5eb0957eb..000000000 --- a/Eigen/src/Core/arch/ZVector/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_Core_arch_ZVector_SRCS "*.h") - -INSTALL(FILES - ${Eigen_Core_arch_ZVector_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/arch/ZVector COMPONENT Devel -) diff --git a/Eigen/src/Core/functors/BinaryFunctors.h b/Eigen/src/Core/functors/BinaryFunctors.h index dc3690444..d82ffed02 100644 --- a/Eigen/src/Core/functors/BinaryFunctors.h +++ b/Eigen/src/Core/functors/BinaryFunctors.h @@ -287,7 +287,7 @@ struct functor_traits<scalar_hypot_op<Scalar,Scalar> > { { Cost = 3 * NumTraits<Scalar>::AddCost + 2 * NumTraits<Scalar>::MulCost + - 2 * NumTraits<Scalar>::template Div<false>::Cost, + 2 * scalar_div_cost<Scalar,false>::value, PacketAccess = false }; }; @@ -375,7 +375,7 @@ 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 = NumTraits<result_type>::template Div<PacketAccess>::Cost + Cost = scalar_div_cost<result_type,PacketAccess>::value }; }; diff --git a/Eigen/src/Core/functors/CMakeLists.txt b/Eigen/src/Core/functors/CMakeLists.txt deleted file mode 100644 index f4b99a9c3..000000000 --- a/Eigen/src/Core/functors/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_Core_Functor_SRCS "*.h") - -INSTALL(FILES - ${Eigen_Core_Functor_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/functors COMPONENT Devel - ) diff --git a/Eigen/src/Core/functors/NullaryFunctors.h b/Eigen/src/Core/functors/NullaryFunctors.h index eaa582f23..a2154d3b5 100644 --- a/Eigen/src/Core/functors/NullaryFunctors.h +++ b/Eigen/src/Core/functors/NullaryFunctors.h @@ -18,10 +18,9 @@ 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) { } - template<typename Index> - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index, Index = 0) const { return m_other; } - template<typename Index, typename PacketType> - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const PacketType packetOp(Index, Index = 0) const { return internal::pset1<PacketType>(m_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> @@ -31,8 +30,8 @@ struct functor_traits<scalar_constant_op<Scalar> > template<typename Scalar> struct scalar_identity_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_identity_op) - template<typename Index> - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index row, Index col) const { return row==col ? Scalar(1) : Scalar(0); } + 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> > @@ -56,15 +55,15 @@ struct linspaced_op_impl<Scalar,Packet,/*RandomAccess*/false,/*IsInteger*/false> m_packetStep(pset1<Packet>(unpacket_traits<Packet>::size*m_step)), m_base(padd(pset1<Packet>(low), pmul(pset1<Packet>(m_step),plset<Packet>(-unpacket_traits<Packet>::size)))) {} - template<typename Index> - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index i) const + template<typename IndexType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (IndexType i) const { m_base = padd(m_base, pset1<Packet>(m_step)); return m_low+Scalar(i)*m_step; } - template<typename Index> - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(Index) const { return m_base = padd(m_base,m_packetStep); } + template<typename IndexType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(IndexType) const { return m_base = padd(m_base,m_packetStep); } const Scalar m_low; const Scalar m_step; @@ -82,11 +81,11 @@ struct linspaced_op_impl<Scalar,Packet,/*RandomAccess*/true,/*IsInteger*/false> m_low(low), m_step(num_steps==1 ? Scalar() : (high-low)/Scalar(num_steps-1)), m_lowPacket(pset1<Packet>(m_low)), m_stepPacket(pset1<Packet>(m_step)), m_interPacket(plset<Packet>(0)) {} - template<typename Index> - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index i) const { return m_low+i*m_step; } + template<typename IndexType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (IndexType i) const { return m_low+i*m_step; } - template<typename Index> - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(Index i) const + template<typename IndexType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(IndexType i) const { return internal::padd(m_lowPacket, pmul(m_stepPacket, padd(pset1<Packet>(Scalar(i)),m_interPacket))); } const Scalar m_low; @@ -103,15 +102,15 @@ struct linspaced_op_impl<Scalar,Packet,/*RandomAccess*/true,/*IsInteger*/true> m_low(low), m_length(high-low), m_divisor(convert_index<Scalar>(num_steps==1?1:num_steps-1)), m_interPacket(plset<Packet>(0)) {} - template<typename Index> + template<typename IndexType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE - const Scalar operator() (Index i) const { + const Scalar operator() (IndexType i) const { return m_low + (m_length*Scalar(i))/m_divisor; } - template<typename Index> + template<typename IndexType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE - const Packet packetOp(Index i) const { + const Packet packetOp(IndexType i) const { return internal::padd(pset1<Packet>(m_low), pdiv(pmul(pset1<Packet>(m_length), padd(pset1<Packet>(Scalar(i)),m_interPacket)), pset1<Packet>(m_divisor))); } @@ -143,29 +142,11 @@ template <typename Scalar, typename PacketType, bool RandomAccess> struct linspa : impl((num_steps==1 ? high : low),high,num_steps) {} - template<typename Index> - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index i) const { return impl(i); } + template<typename IndexType> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (IndexType i) const { return impl(i); } - // We need this function when assigning e.g. a RowVectorXd to a MatrixXd since - // there row==0 and col is used for the actual iteration. - template<typename Index> - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index row, Index col) const - { - eigen_assert(col==0 || row==0); - return impl(col + row); - } - - template<typename Index, typename Packet> - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(Index i) const { return impl.packetOp(i); } - - // We need this function when assigning e.g. a RowVectorXd to a MatrixXd since - // there row==0 and col is used for the actual iteration. - template<typename Index, typename Packet> - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(Index row, Index col) const - { - eigen_assert(col==0 || row==0); - return impl.packetOp(col + row); - } + 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, the different // implementations (random vs. sequential access) as well as the @@ -175,11 +156,11 @@ template <typename Scalar, typename PacketType, bool RandomAccess> struct linspa const linspaced_op_impl<Scalar,PacketType,(NumTraits<Scalar>::IsInteger?true:RandomAccess),NumTraits<Scalar>::IsInteger> impl; }; -// all functors allow linear access, except scalar_identity_op. So we fix here a quick meta -// to indicate whether a functor allows linear access, just always answering 'yes' except for -// scalar_identity_op. -template<typename Functor> struct functor_has_linear_access { enum { ret = 1 }; }; -template<typename Scalar> struct functor_has_linear_access<scalar_identity_op<Scalar> > { enum { ret = 0 }; }; +// 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 struture. +template<typename Functor> struct functor_has_linear_access { enum { ret = !has_binary_operator<Functor>::value }; }; } // end namespace internal diff --git a/Eigen/src/Core/functors/UnaryFunctors.h b/Eigen/src/Core/functors/UnaryFunctors.h index e2f3d869f..2009f8e57 100644 --- a/Eigen/src/Core/functors/UnaryFunctors.h +++ b/Eigen/src/Core/functors/UnaryFunctors.h @@ -1,7 +1,7 @@ // 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) 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 @@ -248,7 +248,7 @@ struct functor_traits<scalar_exp_op<Scalar> > { // double: 7 pmadd, 5 pmul, 3 padd/psub, 1 div, 13 other : (14 * NumTraits<Scalar>::AddCost + 6 * NumTraits<Scalar>::MulCost + - NumTraits<Scalar>::template Div<packet_traits<Scalar>::HasDiv>::Cost)) + scalar_div_cost<Scalar,packet_traits<Scalar>::HasDiv>::value)) #else Cost = (sizeof(Scalar) == 4 @@ -257,7 +257,7 @@ struct functor_traits<scalar_exp_op<Scalar> > { // double: 7 pmadd, 5 pmul, 3 padd/psub, 1 div, 13 other : (23 * NumTraits<Scalar>::AddCost + 12 * NumTraits<Scalar>::MulCost + - NumTraits<Scalar>::template Div<packet_traits<Scalar>::HasDiv>::Cost)) + scalar_div_cost<Scalar,packet_traits<Scalar>::HasDiv>::value)) #endif }; }; @@ -498,138 +498,32 @@ struct functor_traits<scalar_atan_op<Scalar> > 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 { - /** \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 fl(tanh(x)) = +/-1. */ - - // Clamp the inputs to the range [-9, 9] since anything outside - // this range is +/-1.0f in single-precision. - const Scalar plus_9 = static_cast<Scalar>(9.0); - const Scalar minus_9 = static_cast<Scalar>(-9.0); - const Scalar x = numext::maxi(minus_9, numext::mini(plus_9, a)); - // Scalarhe monomial coefficients of the numerator polynomial (odd). - const Scalar alpha_1 = static_cast<Scalar>(4.89352455891786e-03); - const Scalar alpha_3 = static_cast<Scalar>(6.37261928875436e-04); - const Scalar alpha_5 = static_cast<Scalar>(1.48572235717979e-05); - const Scalar alpha_7 = static_cast<Scalar>(5.12229709037114e-08); - const Scalar alpha_9 = static_cast<Scalar>(-8.60467152213735e-11); - const Scalar alpha_11 = static_cast<Scalar>(2.00018790482477e-13); - const Scalar alpha_13 = static_cast<Scalar>(-2.76076847742355e-16); - // Scalarhe monomial coefficients of the denominator polynomial (even). - const Scalar beta_0 = static_cast<Scalar>(4.89352518554385e-03); - const Scalar beta_2 = static_cast<Scalar>(2.26843463243900e-03); - const Scalar beta_4 = static_cast<Scalar>(1.18534705686654e-04); - const Scalar beta_6 = static_cast<Scalar>(1.19825839466702e-06); - // Since the polynomials are odd/even, we need x^2. - const Scalar x2 = x * x; - // Evaluate the numerator polynomial p. - Scalar p = x2 * alpha_13 + alpha_11; - p = x2 * p + alpha_9; - p = x2 * p + alpha_7; - p = x2 * p + alpha_5; - p = x2 * p + alpha_3; - p = x2 * p + alpha_1; - p = x * p; - // Evaluate the denominator polynomial p. - Scalar q = x2 * beta_6 + beta_4; - q = x2 * q + beta_2; - q = x2 * q + beta_0; - // Divide the numerator by the denominator. - return p / q; - } + 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 { - /** \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 - fl(tanh(x)) = +/-1. */ - - // Clamp the inputs to the range [-9, 9] since anything outside - // this range is +/-1.0f in single-precision. - const Packet plus_9 = pset1<Packet>(9.0); - const Packet minus_9 = pset1<Packet>(-9.0); - const Packet x = pmax(minus_9, pmin(plus_9, _x)); - - // The monomial coefficients of the numerator polynomial (odd). - const Packet alpha_1 = pset1<Packet>(4.89352455891786e-03); - const Packet alpha_3 = pset1<Packet>(6.37261928875436e-04); - const Packet alpha_5 = pset1<Packet>(1.48572235717979e-05); - const Packet alpha_7 = pset1<Packet>(5.12229709037114e-08); - const Packet alpha_9 = pset1<Packet>(-8.60467152213735e-11); - const Packet alpha_11 = pset1<Packet>(2.00018790482477e-13); - const Packet alpha_13 = pset1<Packet>(-2.76076847742355e-16); - - // The monomial coefficients of the denominator polynomial (even). - const Packet beta_0 = pset1<Packet>(4.89352518554385e-03); - const Packet beta_2 = pset1<Packet>(2.26843463243900e-03); - const Packet beta_4 = pset1<Packet>(1.18534705686654e-04); - const Packet beta_6 = pset1<Packet>(1.19825839466702e-06); - - // Since the polynomials are odd/even, we need x^2. - const Packet x2 = pmul(x, x); - - // Evaluate the numerator polynomial p. - Packet 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. - Packet 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); - } -}; -template <> -struct scalar_tanh_op<std::complex<double> > { - EIGEN_DEVICE_FUNC inline const std::complex<double> operator()( - const std::complex<double>& a) const { - return numext::tanh(a); - } -}; -template <> -struct scalar_tanh_op<std::complex<float> > { - EIGEN_DEVICE_FUNC inline const std::complex<float> operator()( - const std::complex<float>& a) const { - return numext::tanh(a); - } + 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 = (PacketAccess && (!is_same<Scalar, std::complex<float> >::value) && - (!is_same<Scalar, std::complex<double> >::value) + 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 + - NumTraits<Scalar>::template Div< - packet_traits<Scalar>::HasDiv>::Cost) + scalar_div_cost<Scalar,packet_traits<Scalar>::HasDiv>::value) #else ? (11 * NumTraits<Scalar>::AddCost + 11 * NumTraits<Scalar>::MulCost + - NumTraits<Scalar>::template Div< - packet_traits<Scalar>::HasDiv>::Cost) + 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 * NumTraits<Scalar>::template Div< - packet_traits<Scalar>::HasDiv>::Cost + + 2 * scalar_div_cost<Scalar,packet_traits<Scalar>::HasDiv>::value + functor_traits<scalar_exp_op<Scalar> >::Cost)) }; }; diff --git a/Eigen/src/Core/products/CMakeLists.txt b/Eigen/src/Core/products/CMakeLists.txt deleted file mode 100644 index 21fc94ae3..000000000 --- a/Eigen/src/Core/products/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_Core_Product_SRCS "*.h") - -INSTALL(FILES - ${Eigen_Core_Product_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/products COMPONENT Devel - ) diff --git a/Eigen/src/Core/products/GeneralMatrixVector.h b/Eigen/src/Core/products/GeneralMatrixVector.h index 4a5cf3fb6..3c1a7fc40 100644 --- a/Eigen/src/Core/products/GeneralMatrixVector.h +++ b/Eigen/src/Core/products/GeneralMatrixVector.h @@ -183,8 +183,8 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,LhsMapper,C alignmentPattern = AllAligned; } - const Index offset1 = (FirstAligned && alignmentStep==1?3:1); - const Index offset3 = (FirstAligned && alignmentStep==1?1:3); + 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) @@ -457,8 +457,8 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,LhsMapper,R alignmentPattern = AllAligned; } - const Index offset1 = (FirstAligned && alignmentStep==1?3:1); - const Index offset3 = (FirstAligned && alignmentStep==1?1:3); + 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) diff --git a/Eigen/src/Core/util/BlasUtil.h b/Eigen/src/Core/util/BlasUtil.h index 8b3b44a58..6e6ee119b 100755 --- a/Eigen/src/Core/util/BlasUtil.h +++ b/Eigen/src/Core/util/BlasUtil.h @@ -44,16 +44,29 @@ template<bool Conjugate> struct conj_if; template<> struct conj_if<true> { template<typename T> - inline T operator()(const T& x) { return numext::conj(x); } + inline T operator()(const T& x) const { return numext::conj(x); } template<typename T> - inline T pconj(const T& x) { return internal::pconj(x); } + 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) { return x; } + inline const T& operator()(const T& x) const { return x; } template<typename T> - inline const T& pconj(const T& x) { return x; } + 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> @@ -315,6 +328,11 @@ struct blas_traits<CwiseBinaryOp<scalar_product_op<Scalar>, NestedXpr, const Cwi 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> diff --git a/Eigen/src/Core/util/CMakeLists.txt b/Eigen/src/Core/util/CMakeLists.txt deleted file mode 100644 index a1e2e521f..000000000 --- a/Eigen/src/Core/util/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_Core_util_SRCS "*.h") - -INSTALL(FILES - ${Eigen_Core_util_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/util COMPONENT Devel - ) diff --git a/Eigen/src/Core/util/DisableStupidWarnings.h b/Eigen/src/Core/util/DisableStupidWarnings.h index dd44c7cbc..b13e5da25 100755 --- a/Eigen/src/Core/util/DisableStupidWarnings.h +++ b/Eigen/src/Core/util/DisableStupidWarnings.h @@ -56,7 +56,11 @@ #pragma diag_suppress code_is_unreachable // Disable the "dynamic initialization in unreachable code" message #pragma diag_suppress initialization_not_reachable - // Disable the "calling a __host__ function from a __host__ __device__ function is not allowed" messages (yes, there are 4 of them) + // 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 diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index a7c4ec9a6..a9db2f4c7 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -28,9 +28,9 @@ #define EIGEN_COMP_GNUC 0 #endif -/// \internal EIGEN_COMP_CLANG set to 1 if the compiler is clang (alias for __clang__) +/// \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 1 + #define EIGEN_COMP_CLANG (__clang_major__*100+__clang_minor__) #else #define EIGEN_COMP_CLANG 0 #endif diff --git a/Eigen/src/Core/util/Meta.h b/Eigen/src/Core/util/Meta.h index 4b35761e0..d4460bb77 100755 --- a/Eigen/src/Core/util/Meta.h +++ b/Eigen/src/Core/util/Meta.h @@ -22,6 +22,16 @@ 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 @@ -358,17 +368,46 @@ struct result_of<Func(ArgType0,ArgType1,ArgType2)> { }; #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 { - typedef char yes[1]; - typedef char no[2]; + 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_ref(); + +template <typename T, typename IndexType=Index> +struct has_nullary_operator +{ + template <typename C> static meta_yes testFunctor(C const *,typename enable_if<(sizeof(return_ref<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_ref<C>().operator()(IndexType(0)))>0)>::type * = 0); + static meta_no testFunctor(...); + + enum { value = sizeof(testFunctor(static_cast<T*>(0))) == sizeof(meta_yes) }; +}; - template <typename C> static yes& testFunctor(C const *, typename C::ReturnType const * = 0); - static no& testFunctor(...); +template <typename T, typename IndexType=Index> +struct has_binary_operator +{ + template <typename C> static meta_yes testFunctor(C const *,typename enable_if<(sizeof(return_ref<C>().operator()(IndexType(0),IndexType(0)))>0)>::type * = 0); + static meta_no testFunctor(...); - static const bool value = sizeof(testFunctor(static_cast<T*>(0))) == sizeof(yes); + 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. diff --git a/Eigen/src/Core/util/ReenableStupidWarnings.h b/Eigen/src/Core/util/ReenableStupidWarnings.h index 5d1bbeef6..86b60f52f 100644 --- a/Eigen/src/Core/util/ReenableStupidWarnings.h +++ b/Eigen/src/Core/util/ReenableStupidWarnings.h @@ -15,7 +15,7 @@ #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 triggeredby nvcc. +// otherwise they'll be triggered by nvcc. // #pragma diag_default code_is_unreachable // #pragma diag_default initialization_not_reachable // #pragma diag_default 2651 diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h index a98ba6e86..fa60008ef 100644 --- a/Eigen/src/Core/util/XprHelper.h +++ b/Eigen/src/Core/util/XprHelper.h @@ -24,16 +24,6 @@ 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 { template<typename IndexDest, typename IndexSrc> @@ -674,6 +664,20 @@ bool is_same_dense(const T1 &, const T2 &, typename enable_if<!(has_direct_acces 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<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) { @@ -717,7 +721,7 @@ std::string demangle_flags(int f) * 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: + * You can let %Eigen knows that by defining: \code template<typename BinaryOp> struct ScalarBinaryOpTraits<U1,U2,BinaryOp> { typedef U3 ReturnType; }; @@ -735,6 +739,14 @@ std::string demangle_flags(int f) 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> > @@ -751,6 +763,17 @@ 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> @@ -772,18 +795,6 @@ struct ScalarBinaryOpTraits<void,void,BinaryOp> typedef void ReturnType; }; -template<typename T, typename BinaryOp> -struct ScalarBinaryOpTraits<T,std::complex<T>,BinaryOp> -{ - typedef std::complex<T> ReturnType; -}; - -template<typename T, typename BinaryOp> -struct ScalarBinaryOpTraits<std::complex<T>, T,BinaryOp> -{ - typedef std::complex<T> 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 diff --git a/Eigen/src/Eigenvalues/CMakeLists.txt b/Eigen/src/Eigenvalues/CMakeLists.txt deleted file mode 100644 index 193e02685..000000000 --- a/Eigen/src/Eigenvalues/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_EIGENVALUES_SRCS "*.h") - -INSTALL(FILES - ${Eigen_EIGENVALUES_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Eigenvalues COMPONENT Devel - ) diff --git a/Eigen/src/Eigenvalues/GeneralizedEigenSolver.h b/Eigen/src/Eigenvalues/GeneralizedEigenSolver.h index 6eeeb174e..36a91dffc 100644 --- a/Eigen/src/Eigenvalues/GeneralizedEigenSolver.h +++ b/Eigen/src/Eigenvalues/GeneralizedEigenSolver.h @@ -1,8 +1,9 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2012 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2012-2016 Gael Guennebaud <gael.guennebaud@inria.fr> // Copyright (C) 2010,2012 Jitse Niesen <jitse@maths.leeds.ac.uk> +// Copyright (C) 2016 Tobias Wood <tobias@spinicist.org.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 @@ -89,7 +90,7 @@ template<typename _MatrixType> class GeneralizedEigenSolver */ typedef Matrix<Scalar, ColsAtCompileTime, 1, Options & ~RowMajor, MaxColsAtCompileTime, 1> VectorType; - /** \brief Type for vector of complex scalar values eigenvalues as returned by betas(). + /** \brief Type for vector of complex scalar values eigenvalues as returned by alphas(). * * This is a column vector with entries of type #ComplexScalar. * The length of the vector is the size of #MatrixType. @@ -114,7 +115,14 @@ template<typename _MatrixType> class GeneralizedEigenSolver * * \sa compute() for an example. */ - GeneralizedEigenSolver() : m_eivec(), m_alphas(), m_betas(), m_isInitialized(false), m_realQZ(), m_matS(), m_tmp() {} + GeneralizedEigenSolver() + : m_eivec(), + m_alphas(), + m_betas(), + m_valuesOkay(false), + m_vectorsOkay(false), + m_realQZ() + {} /** \brief Default constructor with memory preallocation * @@ -126,10 +134,9 @@ template<typename _MatrixType> class GeneralizedEigenSolver : m_eivec(size, size), m_alphas(size), m_betas(size), - m_isInitialized(false), - m_eigenvectorsOk(false), + m_valuesOkay(false), + m_vectorsOkay(false), m_realQZ(size), - m_matS(size, size), m_tmp(size) {} @@ -149,10 +156,9 @@ template<typename _MatrixType> class GeneralizedEigenSolver : m_eivec(A.rows(), A.cols()), m_alphas(A.cols()), m_betas(A.cols()), - m_isInitialized(false), - m_eigenvectorsOk(false), + m_valuesOkay(false), + m_vectorsOkay(false), m_realQZ(A.cols()), - m_matS(A.rows(), A.cols()), m_tmp(A.cols()) { compute(A, B, computeEigenvectors); @@ -160,22 +166,20 @@ template<typename _MatrixType> class GeneralizedEigenSolver /* \brief Returns the computed generalized eigenvectors. * - * \returns %Matrix whose columns are the (possibly complex) eigenvectors. + * \returns %Matrix whose columns are the (possibly complex) right eigenvectors. + * i.e. the eigenvectors that solve (A - l*B)x = 0. The ordering matches the eigenvalues. * * \pre Either the constructor * GeneralizedEigenSolver(const MatrixType&,const MatrixType&, bool) or the member function * compute(const MatrixType&, const MatrixType& bool) has been called before, and * \p computeEigenvectors was set to true (the default). * - * Column \f$ k \f$ of the returned matrix is an eigenvector corresponding - * to eigenvalue number \f$ k \f$ as returned by eigenvalues(). The - * eigenvectors are normalized to have (Euclidean) norm equal to one. The - * matrix returned by this function is the matrix \f$ V \f$ in the - * generalized eigendecomposition \f$ A = B V D V^{-1} \f$, if it exists. - * * \sa eigenvalues() */ -// EigenvectorsType eigenvectors() const; + EigenvectorsType eigenvectors() const { + eigen_assert(m_vectorsOkay && "Eigenvectors for GeneralizedEigenSolver were not calculated."); + return m_eivec; + } /** \brief Returns an expression of the computed generalized eigenvalues. * @@ -197,7 +201,7 @@ template<typename _MatrixType> class GeneralizedEigenSolver */ EigenvalueType eigenvalues() const { - eigen_assert(m_isInitialized && "GeneralizedEigenSolver is not initialized."); + eigen_assert(m_valuesOkay && "GeneralizedEigenSolver is not initialized."); return EigenvalueType(m_alphas,m_betas); } @@ -208,7 +212,7 @@ template<typename _MatrixType> class GeneralizedEigenSolver * \sa betas(), eigenvalues() */ ComplexVectorType alphas() const { - eigen_assert(m_isInitialized && "GeneralizedEigenSolver is not initialized."); + eigen_assert(m_valuesOkay && "GeneralizedEigenSolver is not initialized."); return m_alphas; } @@ -219,7 +223,7 @@ template<typename _MatrixType> class GeneralizedEigenSolver * \sa alphas(), eigenvalues() */ VectorType betas() const { - eigen_assert(m_isInitialized && "GeneralizedEigenSolver is not initialized."); + eigen_assert(m_valuesOkay && "GeneralizedEigenSolver is not initialized."); return m_betas; } @@ -250,7 +254,7 @@ template<typename _MatrixType> class GeneralizedEigenSolver ComputationInfo info() const { - eigen_assert(m_isInitialized && "EigenSolver is not initialized."); + eigen_assert(m_valuesOkay && "EigenSolver is not initialized."); return m_realQZ.info(); } @@ -270,29 +274,14 @@ template<typename _MatrixType> class GeneralizedEigenSolver EIGEN_STATIC_ASSERT(!NumTraits<Scalar>::IsComplex, NUMERIC_TYPE_MUST_BE_REAL); } - MatrixType m_eivec; + EigenvectorsType m_eivec; ComplexVectorType m_alphas; VectorType m_betas; - bool m_isInitialized; - bool m_eigenvectorsOk; + bool m_valuesOkay, m_vectorsOkay; RealQZ<MatrixType> m_realQZ; - MatrixType m_matS; - - typedef Matrix<Scalar, ColsAtCompileTime, 1, Options & ~RowMajor, MaxColsAtCompileTime, 1> ColumnVectorType; - ColumnVectorType m_tmp; + ComplexVectorType m_tmp; }; -//template<typename MatrixType> -//typename GeneralizedEigenSolver<MatrixType>::EigenvectorsType GeneralizedEigenSolver<MatrixType>::eigenvectors() const -//{ -// eigen_assert(m_isInitialized && "EigenSolver is not initialized."); -// eigen_assert(m_eigenvectorsOk && "The eigenvectors have not been computed together with the eigenvalues."); -// Index n = m_eivec.cols(); -// EigenvectorsType matV(n,n); -// // TODO -// return matV; -//} - template<typename MatrixType> GeneralizedEigenSolver<MatrixType>& GeneralizedEigenSolver<MatrixType>::compute(const MatrixType& A, const MatrixType& B, bool computeEigenvectors) @@ -302,28 +291,70 @@ GeneralizedEigenSolver<MatrixType>::compute(const MatrixType& A, const MatrixTyp using std::sqrt; using std::abs; eigen_assert(A.cols() == A.rows() && B.cols() == A.rows() && B.cols() == B.rows()); - + Index size = A.cols(); + m_valuesOkay = false; + m_vectorsOkay = false; // Reduce to generalized real Schur form: // A = Q S Z and B = Q T Z m_realQZ.compute(A, B, computeEigenvectors); - if (m_realQZ.info() == Success) { - m_matS = m_realQZ.matrixS(); - const MatrixType &matT = m_realQZ.matrixT(); + // Resize storage + m_alphas.resize(size); + m_betas.resize(size); if (computeEigenvectors) - m_eivec = m_realQZ.matrixZ().transpose(); - - // Compute eigenvalues from matS - m_alphas.resize(A.cols()); - m_betas.resize(A.cols()); + { + m_eivec.resize(size,size); + m_tmp.resize(size); + } + + // Aliases: + Map<VectorType> v(reinterpret_cast<Scalar*>(m_tmp.data()), size); + ComplexVectorType &cv = m_tmp; + const MatrixType &mZ = m_realQZ.matrixZ(); + const MatrixType &mS = m_realQZ.matrixS(); + const MatrixType &mT = m_realQZ.matrixT(); + Index i = 0; - while (i < A.cols()) + while (i < size) { - if (i == A.cols() - 1 || m_matS.coeff(i+1, i) == Scalar(0)) + if (i == size - 1 || mS.coeff(i+1, i) == Scalar(0)) { - m_alphas.coeffRef(i) = m_matS.coeff(i, i); - m_betas.coeffRef(i) = matT.coeff(i,i); + // Real eigenvalue + m_alphas.coeffRef(i) = mS.diagonal().coeff(i); + m_betas.coeffRef(i) = mT.diagonal().coeff(i); + if (computeEigenvectors) + { + v.setConstant(Scalar(0.0)); + v.coeffRef(i) = Scalar(1.0); + // For singular eigenvalues do nothing more + if(abs(m_betas.coeffRef(i)) >= (std::numeric_limits<RealScalar>::min)()) + { + // Non-singular eigenvalue + const Scalar alpha = real(m_alphas.coeffRef(i)); + const Scalar beta = m_betas.coeffRef(i); + for (Index j = i-1; j >= 0; j--) + { + const Index st = j+1; + const Index sz = i-j; + if (j > 0 && mS.coeff(j, j-1) != Scalar(0)) + { + // 2x2 block + Matrix<Scalar, 2, 1> rhs = (alpha*mT.template block<2,Dynamic>(j-1,st,2,sz) - beta*mS.template block<2,Dynamic>(j-1,st,2,sz)) .lazyProduct( v.segment(st,sz) ); + Matrix<Scalar, 2, 2> lhs = beta * mS.template block<2,2>(j-1,j-1) - alpha * mT.template block<2,2>(j-1,j-1); + v.template segment<2>(j-1) = lhs.partialPivLu().solve(rhs); + j--; + } + else + { + v.coeffRef(j) = -v.segment(st,sz).transpose().cwiseProduct(beta*mS.block(j,st,1,sz) - alpha*mT.block(j,st,1,sz)).sum() / (beta*mS.coeffRef(j,j) - alpha*mT.coeffRef(j,j)); + } + } + } + m_eivec.col(i).real().noalias() = mZ.transpose() * v; + m_eivec.col(i).real().normalize(); + m_eivec.col(i).imag().setConstant(0); + } ++i; } else @@ -333,27 +364,53 @@ GeneralizedEigenSolver<MatrixType>::compute(const MatrixType& A, const MatrixTyp // T = [a 0] // [0 b] - RealScalar a = matT.diagonal().coeff(i), - b = matT.diagonal().coeff(i+1); + RealScalar a = mT.diagonal().coeff(i), + b = mT.diagonal().coeff(i+1); + const RealScalar beta = m_betas.coeffRef(i) = m_betas.coeffRef(i+1) = a*b; + // ^^ NOTE: using diagonal()(i) instead of coeff(i,i) workarounds a MSVC bug. - Matrix<RealScalar,2,2> S2 = m_matS.template block<2,2>(i,i) * Matrix<Scalar,2,1>(b,a).asDiagonal(); + Matrix<RealScalar,2,2> S2 = mS.template block<2,2>(i,i) * Matrix<Scalar,2,1>(b,a).asDiagonal(); Scalar p = Scalar(0.5) * (S2.coeff(0,0) - S2.coeff(1,1)); Scalar z = sqrt(abs(p * p + S2.coeff(1,0) * S2.coeff(0,1))); - m_alphas.coeffRef(i) = ComplexScalar(S2.coeff(1,1) + p, z); - m_alphas.coeffRef(i+1) = ComplexScalar(S2.coeff(1,1) + p, -z); - - m_betas.coeffRef(i) = - m_betas.coeffRef(i+1) = a*b; - + const ComplexScalar alpha = ComplexScalar(S2.coeff(1,1) + p, (beta > 0) ? z : -z); + m_alphas.coeffRef(i) = conj(alpha); + m_alphas.coeffRef(i+1) = alpha; + + if (computeEigenvectors) { + // Compute eigenvector in position (i+1) and then position (i) is just the conjugate + cv.setZero(); + cv.coeffRef(i+1) = Scalar(1.0); + // here, the "static_cast" workaound expression template issues. + cv.coeffRef(i) = -(static_cast<Scalar>(beta*mS.coeffRef(i,i+1)) - alpha*mT.coeffRef(i,i+1)) + / (static_cast<Scalar>(beta*mS.coeffRef(i,i)) - alpha*mT.coeffRef(i,i)); + for (Index j = i-1; j >= 0; j--) + { + const Index st = j+1; + const Index sz = i+1-j; + if (j > 0 && mS.coeff(j, j-1) != Scalar(0)) + { + // 2x2 block + Matrix<ComplexScalar, 2, 1> rhs = (alpha*mT.template block<2,Dynamic>(j-1,st,2,sz) - beta*mS.template block<2,Dynamic>(j-1,st,2,sz)) .lazyProduct( cv.segment(st,sz) ); + Matrix<ComplexScalar, 2, 2> lhs = beta * mS.template block<2,2>(j-1,j-1) - alpha * mT.template block<2,2>(j-1,j-1); + cv.template segment<2>(j-1) = lhs.partialPivLu().solve(rhs); + j--; + } else { + cv.coeffRef(j) = cv.segment(st,sz).transpose().cwiseProduct(beta*mS.block(j,st,1,sz) - alpha*mT.block(j,st,1,sz)).sum() + / (alpha*mT.coeffRef(j,j) - static_cast<Scalar>(beta*mS.coeffRef(j,j))); + } + } + m_eivec.col(i+1).noalias() = (mZ.transpose() * cv); + m_eivec.col(i+1).normalize(); + m_eivec.col(i) = m_eivec.col(i+1).conjugate(); + } i += 2; } } - } - - m_isInitialized = true; - m_eigenvectorsOk = false;//computeEigenvectors; + m_valuesOkay = true; + m_vectorsOkay = computeEigenvectors; + } return *this; } diff --git a/Eigen/src/Eigenvalues/RealSchur.h b/Eigen/src/Eigenvalues/RealSchur.h index 9d9063004..d6a339f07 100644 --- a/Eigen/src/Eigenvalues/RealSchur.h +++ b/Eigen/src/Eigenvalues/RealSchur.h @@ -236,7 +236,7 @@ template<typename _MatrixType> class RealSchur typedef Matrix<Scalar,3,1> Vector3s; Scalar computeNormOfT(); - Index findSmallSubdiagEntry(Index iu, const Scalar& maxDiagEntry); + Index findSmallSubdiagEntry(Index iu); void splitOffTwoRows(Index iu, bool computeU, const Scalar& exshift); void computeShift(Index iu, Index iter, Scalar& exshift, Vector3s& shiftInfo); void initFrancisQRStep(Index il, Index iu, const Vector3s& shiftInfo, Index& im, Vector3s& firstHouseholderVector); @@ -293,18 +293,14 @@ RealSchur<MatrixType>& RealSchur<MatrixType>::computeFromHessenberg(const HessMa if(norm!=0) { - Scalar maxDiagEntry = m_matT.cwiseAbs().diagonal().maxCoeff(); - while (iu >= 0) { - Index il = findSmallSubdiagEntry(iu,maxDiagEntry); + Index il = findSmallSubdiagEntry(iu); // Check for convergence if (il == iu) // One root found { m_matT.coeffRef(iu,iu) = m_matT.coeff(iu,iu) + exshift; - // keep track of the largest diagonal coefficient - maxDiagEntry = numext::maxi<Scalar>(maxDiagEntry,abs(m_matT.coeffRef(iu,iu))); if (iu > 0) m_matT.coeffRef(iu, iu-1) = Scalar(0); iu--; @@ -313,8 +309,6 @@ RealSchur<MatrixType>& RealSchur<MatrixType>::computeFromHessenberg(const HessMa else if (il == iu-1) // Two roots found { splitOffTwoRows(iu, computeU, exshift); - // keep track of the largest diagonal coefficient - maxDiagEntry = numext::maxi<Scalar>(maxDiagEntry,numext::maxi(abs(m_matT.coeff(iu,iu)), abs(m_matT.coeff(iu-1,iu-1)))); iu -= 2; iter = 0; } @@ -329,8 +323,6 @@ RealSchur<MatrixType>& RealSchur<MatrixType>::computeFromHessenberg(const HessMa Index im; initFrancisQRStep(il, iu, shiftInfo, im, firstHouseholderVector); performFrancisQRStep(il, im, iu, computeU, firstHouseholderVector, workspace); - // keep track of the largest diagonal coefficient - maxDiagEntry = numext::maxi(maxDiagEntry,m_matT.cwiseAbs().diagonal().segment(im,iu-im).maxCoeff()); } } } @@ -360,13 +352,14 @@ inline typename MatrixType::Scalar RealSchur<MatrixType>::computeNormOfT() /** \internal Look for single small sub-diagonal element and returns its index */ template<typename MatrixType> -inline Index RealSchur<MatrixType>::findSmallSubdiagEntry(Index iu, const Scalar& maxDiagEntry) +inline Index RealSchur<MatrixType>::findSmallSubdiagEntry(Index iu) { using std::abs; Index res = iu; while (res > 0) { - if (abs(m_matT.coeff(res,res-1)) <= NumTraits<Scalar>::epsilon() * maxDiagEntry) + Scalar s = abs(m_matT.coeff(res-1,res-1)) + abs(m_matT.coeff(res,res)); + if (abs(m_matT.coeff(res,res-1)) <= NumTraits<Scalar>::epsilon() * s) break; res--; } diff --git a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h index 8258fadbf..a9f56c4f5 100644 --- a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h +++ b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h @@ -501,7 +501,7 @@ ComputationInfo computeFromTridiagonal_impl(DiagType& diag, SubDiagType& subdiag subdiag[i] = 0; // find the largest unreduced block - while (end>0 && subdiag[end-1]==0) + while (end>0 && subdiag[end-1]==RealScalar(0)) { end--; } @@ -569,8 +569,8 @@ template<typename SolverType> struct direct_selfadjoint_eigenvalues<SolverType,3 EIGEN_USING_STD_MATH(atan2) EIGEN_USING_STD_MATH(cos) EIGEN_USING_STD_MATH(sin) - const Scalar s_inv3 = Scalar(1.0)/Scalar(3.0); - const Scalar s_sqrt3 = sqrt(Scalar(3.0)); + const Scalar s_inv3 = Scalar(1)/Scalar(3); + const Scalar s_sqrt3 = sqrt(Scalar(3)); // The characteristic equation is x^3 - c2*x^2 + c1*x - c0 = 0. The // eigenvalues are the roots to this equation, all guaranteed to be @@ -815,14 +815,14 @@ static void tridiagonal_qr_step(RealScalar* diag, RealScalar* subdiag, Index sta // RealScalar mu = diag[end] - e2 / (td + (td>0 ? 1 : -1) * sqrt(td*td + e2)); // This explain the following, somewhat more complicated, version: RealScalar mu = diag[end]; - if(td==0) + if(td==RealScalar(0)) mu -= abs(e); else { RealScalar e2 = numext::abs2(subdiag[end-1]); RealScalar h = numext::hypot(td,e); - if(e2==0) mu -= (e / (td + (td>0 ? 1 : -1))) * (e / h); - else mu -= e2 / (td + (td>0 ? h : -h)); + if(e2==RealScalar(0)) mu -= (e / (td + (td>RealScalar(0) ? RealScalar(1) : RealScalar(-1)))) * (e / h); + else mu -= e2 / (td + (td>RealScalar(0) ? h : -h)); } RealScalar x = diag[start] - mu; diff --git a/Eigen/src/Geometry/CMakeLists.txt b/Eigen/src/Geometry/CMakeLists.txt deleted file mode 100644 index f8f728b84..000000000 --- a/Eigen/src/Geometry/CMakeLists.txt +++ /dev/null @@ -1,8 +0,0 @@ -FILE(GLOB Eigen_Geometry_SRCS "*.h") - -INSTALL(FILES - ${Eigen_Geometry_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Geometry COMPONENT Devel - ) - -ADD_SUBDIRECTORY(arch) diff --git a/Eigen/src/Geometry/Transform.h b/Eigen/src/Geometry/Transform.h index 073f4dcd1..db5fd07c3 100644 --- a/Eigen/src/Geometry/Transform.h +++ b/Eigen/src/Geometry/Transform.h @@ -193,7 +193,7 @@ template<int Mode> struct transform_make_affine; * preprocessor token EIGEN_QT_SUPPORT is defined. * * This class can be extended with the help of the plugin mechanism described on the page - * \ref TopicCustomizingEigen by defining the preprocessor symbol \c EIGEN_TRANSFORM_PLUGIN. + * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_TRANSFORM_PLUGIN. * * \sa class Matrix, class Quaternion */ @@ -1073,7 +1073,7 @@ void Transform<Scalar,Dim,Mode,Options>::computeRotationScaling(RotationMatrixTy } } -/** decomposes the linear part of the transformation as a product rotation x scaling, the scaling being +/** decomposes the linear part of the transformation as a product scaling x rotation, the scaling being * not necessarily positive. * * If either pointer is zero, the corresponding computation is skipped. diff --git a/Eigen/src/Geometry/arch/CMakeLists.txt b/Eigen/src/Geometry/arch/CMakeLists.txt deleted file mode 100644 index 1267a79c7..000000000 --- a/Eigen/src/Geometry/arch/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_Geometry_arch_SRCS "*.h") - -INSTALL(FILES - ${Eigen_Geometry_arch_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Geometry/arch COMPONENT Devel - ) diff --git a/Eigen/src/Householder/CMakeLists.txt b/Eigen/src/Householder/CMakeLists.txt deleted file mode 100644 index ce4937db0..000000000 --- a/Eigen/src/Householder/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_Householder_SRCS "*.h") - -INSTALL(FILES - ${Eigen_Householder_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Householder COMPONENT Devel - ) diff --git a/Eigen/src/IterativeLinearSolvers/CMakeLists.txt b/Eigen/src/IterativeLinearSolvers/CMakeLists.txt deleted file mode 100644 index 59ccc0072..000000000 --- a/Eigen/src/IterativeLinearSolvers/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_IterativeLinearSolvers_SRCS "*.h") - -INSTALL(FILES - ${Eigen_IterativeLinearSolvers_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/IterativeLinearSolvers COMPONENT Devel - ) diff --git a/Eigen/src/Jacobi/CMakeLists.txt b/Eigen/src/Jacobi/CMakeLists.txt deleted file mode 100644 index 490dac626..000000000 --- a/Eigen/src/Jacobi/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_Jacobi_SRCS "*.h") - -INSTALL(FILES - ${Eigen_Jacobi_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Jacobi COMPONENT Devel - ) diff --git a/Eigen/src/LU/CMakeLists.txt b/Eigen/src/LU/CMakeLists.txt deleted file mode 100644 index e0d8d78c1..000000000 --- a/Eigen/src/LU/CMakeLists.txt +++ /dev/null @@ -1,8 +0,0 @@ -FILE(GLOB Eigen_LU_SRCS "*.h") - -INSTALL(FILES - ${Eigen_LU_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/LU COMPONENT Devel - ) - -ADD_SUBDIRECTORY(arch) diff --git a/Eigen/src/LU/FullPivLU.h b/Eigen/src/LU/FullPivLU.h index 1632d3ac3..2b30fc146 100644 --- a/Eigen/src/LU/FullPivLU.h +++ b/Eigen/src/LU/FullPivLU.h @@ -436,7 +436,7 @@ template<typename _MatrixType> class FullPivLU Index m_nonzero_pivots; RealScalar m_l1_norm; RealScalar m_maxpivot, m_prescribedThreshold; - char m_det_pq; + signed char m_det_pq; bool m_isInitialized, m_usePrescribedThreshold; }; @@ -879,14 +879,12 @@ struct Assignment<DstXprType, Inverse<FullPivLU<MatrixType> >, internal::assign_ * * \sa class FullPivLU */ -#ifndef __CUDACC__ template<typename Derived> inline const FullPivLU<typename MatrixBase<Derived>::PlainObject> MatrixBase<Derived>::fullPivLu() const { return FullPivLU<PlainObject>(eval()); } -#endif } // end namespace Eigen diff --git a/Eigen/src/LU/PartialPivLU.h b/Eigen/src/LU/PartialPivLU.h index 87ac6a281..d43961887 100644 --- a/Eigen/src/LU/PartialPivLU.h +++ b/Eigen/src/LU/PartialPivLU.h @@ -284,7 +284,7 @@ template<typename _MatrixType> class PartialPivLU PermutationType m_p; TranspositionType m_rowsTranspositions; RealScalar m_l1_norm; - char m_det_p; + signed char m_det_p; bool m_isInitialized; }; @@ -584,14 +584,12 @@ struct Assignment<DstXprType, Inverse<PartialPivLU<MatrixType> >, internal::assi * * \sa class PartialPivLU */ -#ifndef __CUDACC__ template<typename Derived> inline const PartialPivLU<typename MatrixBase<Derived>::PlainObject> MatrixBase<Derived>::partialPivLu() const { return PartialPivLU<PlainObject>(eval()); } -#endif /** \lu_module * @@ -601,14 +599,12 @@ MatrixBase<Derived>::partialPivLu() const * * \sa class PartialPivLU */ -#ifndef __CUDACC__ template<typename Derived> inline const PartialPivLU<typename MatrixBase<Derived>::PlainObject> MatrixBase<Derived>::lu() const { return PartialPivLU<PlainObject>(eval()); } -#endif } // end namespace Eigen diff --git a/Eigen/src/LU/arch/CMakeLists.txt b/Eigen/src/LU/arch/CMakeLists.txt deleted file mode 100644 index f6b7ed9ec..000000000 --- a/Eigen/src/LU/arch/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_LU_arch_SRCS "*.h") - -INSTALL(FILES - ${Eigen_LU_arch_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/LU/arch COMPONENT Devel - ) diff --git a/Eigen/src/LU/arch/Inverse_SSE.h b/Eigen/src/LU/arch/Inverse_SSE.h index e1470c664..ebb64a62b 100644 --- a/Eigen/src/LU/arch/Inverse_SSE.h +++ b/Eigen/src/LU/arch/Inverse_SSE.h @@ -153,10 +153,12 @@ struct compute_inverse_size4<Architecture::SSE, float, MatrixType, ResultType> iC = _mm_mul_ps(rd,iC); iD = _mm_mul_ps(rd,iD); - result.template writePacket<ResultAlignment>( 0, _mm_shuffle_ps(iA,iB,0x77)); - result.template writePacket<ResultAlignment>( 4, _mm_shuffle_ps(iA,iB,0x22)); - result.template writePacket<ResultAlignment>( 8, _mm_shuffle_ps(iC,iD,0x77)); - result.template writePacket<ResultAlignment>(12, _mm_shuffle_ps(iC,iD,0x22)); + Index res_stride = result.outerStride(); + float* res = result.data(); + pstoret<float, Packet4f, ResultAlignment>(res+0, _mm_shuffle_ps(iA,iB,0x77)); + pstoret<float, Packet4f, ResultAlignment>(res+res_stride, _mm_shuffle_ps(iA,iB,0x22)); + pstoret<float, Packet4f, ResultAlignment>(res+2*res_stride, _mm_shuffle_ps(iC,iD,0x77)); + pstoret<float, Packet4f, ResultAlignment>(res+3*res_stride, _mm_shuffle_ps(iC,iD,0x22)); } }; @@ -316,14 +318,16 @@ struct compute_inverse_size4<Architecture::SSE, double, MatrixType, ResultType> iC1 = _mm_sub_pd(_mm_mul_pd(B1, dC), iC1); iC2 = _mm_sub_pd(_mm_mul_pd(B2, dC), iC2); - result.template writePacket<ResultAlignment>( 0, _mm_mul_pd(_mm_shuffle_pd(iA2, iA1, 3), d1)); // iA# / det - result.template writePacket<ResultAlignment>( 4, _mm_mul_pd(_mm_shuffle_pd(iA2, iA1, 0), d2)); - result.template writePacket<ResultAlignment>( 2, _mm_mul_pd(_mm_shuffle_pd(iB2, iB1, 3), d1)); // iB# / det - result.template writePacket<ResultAlignment>( 6, _mm_mul_pd(_mm_shuffle_pd(iB2, iB1, 0), d2)); - result.template writePacket<ResultAlignment>( 8, _mm_mul_pd(_mm_shuffle_pd(iC2, iC1, 3), d1)); // iC# / det - result.template writePacket<ResultAlignment>(12, _mm_mul_pd(_mm_shuffle_pd(iC2, iC1, 0), d2)); - result.template writePacket<ResultAlignment>(10, _mm_mul_pd(_mm_shuffle_pd(iD2, iD1, 3), d1)); // iD# / det - result.template writePacket<ResultAlignment>(14, _mm_mul_pd(_mm_shuffle_pd(iD2, iD1, 0), d2)); + Index res_stride = result.outerStride(); + double* res = result.data(); + pstoret<double, Packet2d, ResultAlignment>(res+0, _mm_mul_pd(_mm_shuffle_pd(iA2, iA1, 3), d1)); + pstoret<double, Packet2d, ResultAlignment>(res+res_stride, _mm_mul_pd(_mm_shuffle_pd(iA2, iA1, 0), d2)); + pstoret<double, Packet2d, ResultAlignment>(res+2, _mm_mul_pd(_mm_shuffle_pd(iB2, iB1, 3), d1)); + pstoret<double, Packet2d, ResultAlignment>(res+res_stride+2, _mm_mul_pd(_mm_shuffle_pd(iB2, iB1, 0), d2)); + pstoret<double, Packet2d, ResultAlignment>(res+2*res_stride, _mm_mul_pd(_mm_shuffle_pd(iC2, iC1, 3), d1)); + pstoret<double, Packet2d, ResultAlignment>(res+3*res_stride, _mm_mul_pd(_mm_shuffle_pd(iC2, iC1, 0), d2)); + pstoret<double, Packet2d, ResultAlignment>(res+2*res_stride+2,_mm_mul_pd(_mm_shuffle_pd(iD2, iD1, 3), d1)); + pstoret<double, Packet2d, ResultAlignment>(res+3*res_stride+2,_mm_mul_pd(_mm_shuffle_pd(iD2, iD1, 0), d2)); } }; diff --git a/Eigen/src/MetisSupport/CMakeLists.txt b/Eigen/src/MetisSupport/CMakeLists.txt deleted file mode 100644 index 2bad31416..000000000 --- a/Eigen/src/MetisSupport/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_MetisSupport_SRCS "*.h") - -INSTALL(FILES - ${Eigen_MetisSupport_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/MetisSupport COMPONENT Devel - ) diff --git a/Eigen/src/OrderingMethods/CMakeLists.txt b/Eigen/src/OrderingMethods/CMakeLists.txt deleted file mode 100644 index 9f4bb2758..000000000 --- a/Eigen/src/OrderingMethods/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_OrderingMethods_SRCS "*.h") - -INSTALL(FILES - ${Eigen_OrderingMethods_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/OrderingMethods COMPONENT Devel - ) diff --git a/Eigen/src/PaStiXSupport/CMakeLists.txt b/Eigen/src/PaStiXSupport/CMakeLists.txt deleted file mode 100644 index 28c657e9b..000000000 --- a/Eigen/src/PaStiXSupport/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_PastixSupport_SRCS "*.h") - -INSTALL(FILES - ${Eigen_PastixSupport_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/PaStiXSupport COMPONENT Devel - ) diff --git a/Eigen/src/PardisoSupport/CMakeLists.txt b/Eigen/src/PardisoSupport/CMakeLists.txt deleted file mode 100644 index a097ab401..000000000 --- a/Eigen/src/PardisoSupport/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_PardisoSupport_SRCS "*.h") - -INSTALL(FILES - ${Eigen_PardisoSupport_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/PardisoSupport COMPONENT Devel - ) diff --git a/Eigen/src/QR/CMakeLists.txt b/Eigen/src/QR/CMakeLists.txt deleted file mode 100644 index 96f43d7f5..000000000 --- a/Eigen/src/QR/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_QR_SRCS "*.h") - -INSTALL(FILES - ${Eigen_QR_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/QR COMPONENT Devel - ) diff --git a/Eigen/src/QR/ColPivHouseholderQR.h b/Eigen/src/QR/ColPivHouseholderQR.h index ee6989897..9650781d6 100644 --- a/Eigen/src/QR/ColPivHouseholderQR.h +++ b/Eigen/src/QR/ColPivHouseholderQR.h @@ -163,9 +163,6 @@ template<typename _MatrixType> class ColPivHouseholderQR * * \returns a solution. * - * \note The case where b is a matrix is not yet implemented. Also, this - * code is space inefficient. - * * \note_about_checking_solutions * * \note_about_arbitrary_choice_of_solution @@ -640,7 +637,6 @@ typename ColPivHouseholderQR<MatrixType>::HouseholderSequenceType ColPivHousehol return HouseholderSequenceType(m_qr, m_hCoeffs.conjugate()); } -#ifndef __CUDACC__ /** \return the column-pivoting Householder QR decomposition of \c *this. * * \sa class ColPivHouseholderQR @@ -651,7 +647,6 @@ MatrixBase<Derived>::colPivHouseholderQr() const { return ColPivHouseholderQR<PlainObject>(eval()); } -#endif // __CUDACC__ } // end namespace Eigen diff --git a/Eigen/src/QR/CompleteOrthogonalDecomposition.h b/Eigen/src/QR/CompleteOrthogonalDecomposition.h index f299d3c00..41e4ecdfd 100644 --- a/Eigen/src/QR/CompleteOrthogonalDecomposition.h +++ b/Eigen/src/QR/CompleteOrthogonalDecomposition.h @@ -547,7 +547,6 @@ CompleteOrthogonalDecomposition<MatrixType>::householderQ() const { return m_cpqr.householderQ(); } -#ifndef __CUDACC__ /** \return the complete orthogonal decomposition of \c *this. * * \sa class CompleteOrthogonalDecomposition @@ -557,7 +556,6 @@ const CompleteOrthogonalDecomposition<typename MatrixBase<Derived>::PlainObject> MatrixBase<Derived>::completeOrthogonalDecomposition() const { return CompleteOrthogonalDecomposition<PlainObject>(eval()); } -#endif // __CUDACC__ } // end namespace Eigen diff --git a/Eigen/src/QR/FullPivHouseholderQR.h b/Eigen/src/QR/FullPivHouseholderQR.h index 4c85a4693..e0e15100d 100644 --- a/Eigen/src/QR/FullPivHouseholderQR.h +++ b/Eigen/src/QR/FullPivHouseholderQR.h @@ -164,9 +164,6 @@ template<typename _MatrixType> class FullPivHouseholderQR * \returns the exact or least-square solution if the rank is greater or equal to the number of columns of A, * and an arbitrary solution otherwise. * - * \note The case where b is a matrix is not yet implemented. Also, this - * code is space inefficient. - * * \note_about_checking_solutions * * \note_about_arbitrary_choice_of_solution @@ -663,7 +660,6 @@ inline typename FullPivHouseholderQR<MatrixType>::MatrixQReturnType FullPivHouse return MatrixQReturnType(m_qr, m_hCoeffs, m_rows_transpositions); } -#ifndef __CUDACC__ /** \return the full-pivoting Householder QR decomposition of \c *this. * * \sa class FullPivHouseholderQR @@ -674,7 +670,6 @@ MatrixBase<Derived>::fullPivHouseholderQr() const { return FullPivHouseholderQR<PlainObject>(eval()); } -#endif // __CUDACC__ } // end namespace Eigen diff --git a/Eigen/src/QR/HouseholderQR.h b/Eigen/src/QR/HouseholderQR.h index 2e64fa7fe..3513d995c 100644 --- a/Eigen/src/QR/HouseholderQR.h +++ b/Eigen/src/QR/HouseholderQR.h @@ -128,9 +128,6 @@ template<typename _MatrixType> class HouseholderQR * * \returns a solution. * - * \note The case where b is a matrix is not yet implemented. Also, this - * code is space inefficient. - * * \note_about_checking_solutions * * \note_about_arbitrary_choice_of_solution @@ -396,7 +393,6 @@ void HouseholderQR<MatrixType>::computeInPlace() m_isInitialized = true; } -#ifndef __CUDACC__ /** \return the Householder QR decomposition of \c *this. * * \sa class HouseholderQR @@ -407,7 +403,6 @@ MatrixBase<Derived>::householderQr() const { return HouseholderQR<PlainObject>(eval()); } -#endif // __CUDACC__ } // end namespace Eigen diff --git a/Eigen/src/SPQRSupport/CMakeLists.txt b/Eigen/src/SPQRSupport/CMakeLists.txt deleted file mode 100644 index 4968beaf2..000000000 --- a/Eigen/src/SPQRSupport/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_SPQRSupport_SRCS "*.h") - -INSTALL(FILES - ${Eigen_SPQRSupport_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/SPQRSupport/ COMPONENT Devel - ) diff --git a/Eigen/src/SVD/CMakeLists.txt b/Eigen/src/SVD/CMakeLists.txt deleted file mode 100644 index 55efc44b1..000000000 --- a/Eigen/src/SVD/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_SVD_SRCS "*.h") - -INSTALL(FILES - ${Eigen_SVD_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/SVD COMPONENT Devel - ) diff --git a/Eigen/src/SVD/JacobiSVD.h b/Eigen/src/SVD/JacobiSVD.h index 605c1a2a6..78dfd1d59 100644 --- a/Eigen/src/SVD/JacobiSVD.h +++ b/Eigen/src/SVD/JacobiSVD.h @@ -783,7 +783,6 @@ JacobiSVD<MatrixType, QRPreconditioner>::compute(const MatrixType& matrix, unsig return *this; } -#ifndef __CUDACC__ /** \svd_module * * \return the singular value decomposition of \c *this computed by two-sided @@ -797,7 +796,6 @@ MatrixBase<Derived>::jacobiSvd(unsigned int computationOptions) const { return JacobiSVD<PlainObject>(*this, computationOptions); } -#endif // __CUDACC__ } // end namespace Eigen diff --git a/Eigen/src/SparseCholesky/CMakeLists.txt b/Eigen/src/SparseCholesky/CMakeLists.txt deleted file mode 100644 index 375a59d7a..000000000 --- a/Eigen/src/SparseCholesky/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_SparseCholesky_SRCS "*.h") - -INSTALL(FILES - ${Eigen_SparseCholesky_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/SparseCholesky COMPONENT Devel - ) diff --git a/Eigen/src/SparseCore/CMakeLists.txt b/Eigen/src/SparseCore/CMakeLists.txt deleted file mode 100644 index d860452a6..000000000 --- a/Eigen/src/SparseCore/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_SparseCore_SRCS "*.h") - -INSTALL(FILES - ${Eigen_SparseCore_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/SparseCore COMPONENT Devel - ) diff --git a/Eigen/src/SparseCore/SparseCompressedBase.h b/Eigen/src/SparseCore/SparseCompressedBase.h index 15854a73b..55ad91f46 100644 --- a/Eigen/src/SparseCore/SparseCompressedBase.h +++ b/Eigen/src/SparseCore/SparseCompressedBase.h @@ -106,6 +106,25 @@ class SparseCompressedBase /** \returns whether \c *this is in compressed form. */ inline bool isCompressed() const { return innerNonZeroPtr()==0; } + /** \returns a read-only view of the stored coefficients as a 1D array expression. + * + * \warning this method is for \b compressed \b storage \b only, and it will trigger an assertion otherwise. + * + * \sa valuePtr(), isCompressed() */ + const Map<const Array<Scalar,Dynamic,1> > coeffs() const { eigen_assert(isCompressed()); return Array<Scalar,Dynamic,1>::Map(valuePtr(),nonZeros()); } + + /** \returns a read-write view of the stored coefficients as a 1D array expression + * + * \warning this method is for \b compressed \b storage \b only, and it will trigger an assertion otherwise. + * + * Here is an example: + * \include SparseMatrix_coeffs.cpp + * and the output is: + * \include SparseMatrix_coeffs.out + * + * \sa valuePtr(), isCompressed() */ + Map<Array<Scalar,Dynamic,1> > coeffs() { eigen_assert(isCompressed()); return Array<Scalar,Dynamic,1>::Map(valuePtr(),nonZeros()); } + protected: /** Default constructor. Do nothing. */ SparseCompressedBase() {} diff --git a/Eigen/src/SparseCore/SparseMatrix.h b/Eigen/src/SparseCore/SparseMatrix.h index 531fea399..64ca5fc44 100644 --- a/Eigen/src/SparseCore/SparseMatrix.h +++ b/Eigen/src/SparseCore/SparseMatrix.h @@ -35,7 +35,7 @@ namespace Eigen { * \tparam _Index the type of the indices. It has to be a \b signed type (e.g., short, int, std::ptrdiff_t). Default is \c int. * * This class can be extended with the help of the plugin mechanism described on the page - * \ref TopicCustomizingEigen by defining the preprocessor symbol \c EIGEN_SPARSEMATRIX_PLUGIN. + * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_SPARSEMATRIX_PLUGIN. */ namespace internal { diff --git a/Eigen/src/SparseCore/SparseMatrixBase.h b/Eigen/src/SparseCore/SparseMatrixBase.h index 45f64e7f2..96b1b0504 100644 --- a/Eigen/src/SparseCore/SparseMatrixBase.h +++ b/Eigen/src/SparseCore/SparseMatrixBase.h @@ -21,7 +21,7 @@ namespace Eigen { * \tparam Derived is the derived type, e.g. a sparse matrix type, or an expression, etc. * * This class can be extended with the help of the plugin mechanism described on the page - * \ref TopicCustomizingEigen by defining the preprocessor symbol \c EIGEN_SPARSEMATRIXBASE_PLUGIN. + * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_SPARSEMATRIXBASE_PLUGIN. */ template<typename Derived> class SparseMatrixBase : public EigenBase<Derived> diff --git a/Eigen/src/SparseCore/SparseSelfAdjointView.h b/Eigen/src/SparseCore/SparseSelfAdjointView.h index a48520c0c..d31d9babf 100644 --- a/Eigen/src/SparseCore/SparseSelfAdjointView.h +++ b/Eigen/src/SparseCore/SparseSelfAdjointView.h @@ -250,11 +250,11 @@ template<int Mode, typename SparseLhsType, typename DenseRhsType, typename Dense inline void sparse_selfadjoint_time_dense_product(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const AlphaType& alpha) { EIGEN_ONLY_USED_FOR_DEBUG(alpha); - // TODO use alpha - eigen_assert(alpha==AlphaType(1) && "alpha != 1 is not implemented yet, sorry"); - typedef evaluator<SparseLhsType> LhsEval; - typedef typename evaluator<SparseLhsType>::InnerIterator LhsIterator; + typedef typename internal::nested_eval<SparseLhsType,DenseRhsType::MaxColsAtCompileTime>::type SparseLhsTypeNested; + typedef typename internal::remove_all<SparseLhsTypeNested>::type SparseLhsTypeNestedCleaned; + typedef evaluator<SparseLhsTypeNestedCleaned> LhsEval; + typedef typename LhsEval::InnerIterator LhsIterator; typedef typename SparseLhsType::Scalar LhsScalar; enum { @@ -266,39 +266,53 @@ inline void sparse_selfadjoint_time_dense_product(const SparseLhsType& lhs, cons ProcessSecondHalf = !ProcessFirstHalf }; - LhsEval lhsEval(lhs); - - for (Index j=0; j<lhs.outerSize(); ++j) + SparseLhsTypeNested lhs_nested(lhs); + LhsEval lhsEval(lhs_nested); + + // work on one column at once + for (Index k=0; k<rhs.cols(); ++k) { - LhsIterator i(lhsEval,j); - if (ProcessSecondHalf) + for (Index j=0; j<lhs.outerSize(); ++j) { - while (i && i.index()<j) ++i; - if(i && i.index()==j) + LhsIterator i(lhsEval,j); + // handle diagonal coeff + if (ProcessSecondHalf) { - res.row(j) += i.value() * rhs.row(j); - ++i; + while (i && i.index()<j) ++i; + if(i && i.index()==j) + { + res(j,k) += alpha * i.value() * rhs(j,k); + ++i; + } } + + // premultiplied rhs for scatters + typename ScalarBinaryOpTraits<AlphaType, typename DenseRhsType::Scalar>::ReturnType rhs_j(alpha*rhs(j,k)); + // accumulator for partial scalar product + typename DenseResType::Scalar res_j(0); + for(; (ProcessFirstHalf ? i && i.index() < j : i) ; ++i) + { + LhsScalar lhs_ij = i.value(); + if(!LhsIsRowMajor) lhs_ij = numext::conj(lhs_ij); + res_j += lhs_ij * rhs(i.index(),k); + res(i.index(),k) += numext::conj(lhs_ij) * rhs_j; + } + res(j,k) += alpha * res_j; + + // handle diagonal coeff + if (ProcessFirstHalf && i && (i.index()==j)) + res(j,k) += alpha * i.value() * rhs(j,k); } - for(; (ProcessFirstHalf ? i && i.index() < j : i) ; ++i) - { - Index a = LhsIsRowMajor ? j : i.index(); - Index b = LhsIsRowMajor ? i.index() : j; - LhsScalar v = i.value(); - res.row(a) += (v) * rhs.row(b); - res.row(b) += numext::conj(v) * rhs.row(a); - } - if (ProcessFirstHalf && i && (i.index()==j)) - res.row(j) += i.value() * rhs.row(j); } } template<typename LhsView, typename Rhs, int ProductType> struct generic_product_impl<LhsView, Rhs, SparseSelfAdjointShape, DenseShape, ProductType> +: generic_product_impl_base<LhsView, Rhs, generic_product_impl<LhsView, Rhs, SparseSelfAdjointShape, DenseShape, ProductType> > { template<typename Dest> - static void evalTo(Dest& dst, const LhsView& lhsView, const Rhs& rhs) + static void scaleAndAddTo(Dest& dst, const LhsView& lhsView, const Rhs& rhs, const typename Dest::Scalar& alpha) { typedef typename LhsView::_MatrixTypeNested Lhs; typedef typename nested_eval<Lhs,Dynamic>::type LhsNested; @@ -306,16 +320,16 @@ struct generic_product_impl<LhsView, Rhs, SparseSelfAdjointShape, DenseShape, Pr LhsNested lhsNested(lhsView.matrix()); RhsNested rhsNested(rhs); - dst.setZero(); - internal::sparse_selfadjoint_time_dense_product<LhsView::Mode>(lhsNested, rhsNested, dst, typename Dest::Scalar(1)); + internal::sparse_selfadjoint_time_dense_product<LhsView::Mode>(lhsNested, rhsNested, dst, alpha); } }; template<typename Lhs, typename RhsView, int ProductType> struct generic_product_impl<Lhs, RhsView, DenseShape, SparseSelfAdjointShape, ProductType> +: generic_product_impl_base<Lhs, RhsView, generic_product_impl<Lhs, RhsView, DenseShape, SparseSelfAdjointShape, ProductType> > { template<typename Dest> - static void evalTo(Dest& dst, const Lhs& lhs, const RhsView& rhsView) + static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const RhsView& rhsView, const typename Dest::Scalar& alpha) { typedef typename RhsView::_MatrixTypeNested Rhs; typedef typename nested_eval<Lhs,Dynamic>::type LhsNested; @@ -323,10 +337,9 @@ struct generic_product_impl<Lhs, RhsView, DenseShape, SparseSelfAdjointShape, Pr LhsNested lhsNested(lhs); RhsNested rhsNested(rhsView.matrix()); - dst.setZero(); - // transpoe everything + // transpose everything Transpose<Dest> dstT(dst); - internal::sparse_selfadjoint_time_dense_product<RhsView::Mode>(rhsNested.transpose(), lhsNested.transpose(), dstT, typename Dest::Scalar(1)); + internal::sparse_selfadjoint_time_dense_product<RhsView::Mode>(rhsNested.transpose(), lhsNested.transpose(), dstT, alpha); } }; diff --git a/Eigen/src/SparseCore/SparseVector.h b/Eigen/src/SparseCore/SparseVector.h index 167a9886c..00ee6ec89 100644 --- a/Eigen/src/SparseCore/SparseVector.h +++ b/Eigen/src/SparseCore/SparseVector.h @@ -22,7 +22,7 @@ namespace Eigen { * See http://www.netlib.org/linalg/html_templates/node91.html for details on the storage scheme. * * This class can be extended with the help of the plugin mechanism described on the page - * \ref TopicCustomizingEigen by defining the preprocessor symbol \c EIGEN_SPARSEVECTOR_PLUGIN. + * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_SPARSEVECTOR_PLUGIN. */ namespace internal { diff --git a/Eigen/src/SparseLU/CMakeLists.txt b/Eigen/src/SparseLU/CMakeLists.txt deleted file mode 100644 index 69729ee89..000000000 --- a/Eigen/src/SparseLU/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_SparseLU_SRCS "*.h") - -INSTALL(FILES - ${Eigen_SparseLU_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/SparseLU COMPONENT Devel - ) diff --git a/Eigen/src/SparseQR/CMakeLists.txt b/Eigen/src/SparseQR/CMakeLists.txt deleted file mode 100644 index f9ddf2bdb..000000000 --- a/Eigen/src/SparseQR/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_SparseQR_SRCS "*.h") - -INSTALL(FILES - ${Eigen_SparseQR_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/SparseQR/ COMPONENT Devel - ) diff --git a/Eigen/src/StlSupport/CMakeLists.txt b/Eigen/src/StlSupport/CMakeLists.txt deleted file mode 100644 index 0f094f637..000000000 --- a/Eigen/src/StlSupport/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_StlSupport_SRCS "*.h") - -INSTALL(FILES - ${Eigen_StlSupport_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/StlSupport COMPONENT Devel - ) diff --git a/Eigen/src/SuperLUSupport/CMakeLists.txt b/Eigen/src/SuperLUSupport/CMakeLists.txt deleted file mode 100644 index b28ebe583..000000000 --- a/Eigen/src/SuperLUSupport/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_SuperLUSupport_SRCS "*.h") - -INSTALL(FILES - ${Eigen_SuperLUSupport_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/SuperLUSupport COMPONENT Devel - ) diff --git a/Eigen/src/UmfPackSupport/CMakeLists.txt b/Eigen/src/UmfPackSupport/CMakeLists.txt deleted file mode 100644 index a57de0020..000000000 --- a/Eigen/src/UmfPackSupport/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_UmfPackSupport_SRCS "*.h") - -INSTALL(FILES - ${Eigen_UmfPackSupport_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/UmfPackSupport COMPONENT Devel - ) diff --git a/Eigen/src/misc/CMakeLists.txt b/Eigen/src/misc/CMakeLists.txt deleted file mode 100644 index a58ffb745..000000000 --- a/Eigen/src/misc/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_misc_SRCS "*.h") - -INSTALL(FILES - ${Eigen_misc_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/misc COMPONENT Devel - ) diff --git a/Eigen/src/plugins/CMakeLists.txt b/Eigen/src/plugins/CMakeLists.txt deleted file mode 100644 index 1a1d3ffbd..000000000 --- a/Eigen/src/plugins/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_plugins_SRCS "*.h") - -INSTALL(FILES - ${Eigen_plugins_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/plugins COMPONENT Devel - ) diff --git a/cmake/FindEigen3.cmake b/cmake/FindEigen3.cmake index cea1afeab..9e9697860 100644 --- a/cmake/FindEigen3.cmake +++ b/cmake/FindEigen3.cmake @@ -66,16 +66,23 @@ if (EIGEN3_INCLUDE_DIR) set(EIGEN3_FOUND ${EIGEN3_VERSION_OK}) else (EIGEN3_INCLUDE_DIR) - - find_path(EIGEN3_INCLUDE_DIR NAMES signature_of_eigen3_matrix_library - HINTS - ENV EIGEN3_ROOT - ENV EIGEN3_ROOT_DIR - PATHS - ${CMAKE_INSTALL_PREFIX}/include - ${KDE4_INCLUDE_DIR} - PATH_SUFFIXES eigen3 eigen - ) + + # search first if an Eigen3Config.cmake is available in the system, + # if successful this would set EIGEN3_INCLUDE_DIR and the rest of + # the script will work as usual + find_package(Eigen3 ${Eigen3_FIND_VERSION} NO_MODULE QUIET) + + if(NOT EIGEN3_INCLUDE_DIR) + find_path(EIGEN3_INCLUDE_DIR NAMES signature_of_eigen3_matrix_library + HINTS + ENV EIGEN3_ROOT + ENV EIGEN3_ROOT_DIR + PATHS + ${CMAKE_INSTALL_PREFIX}/include + ${KDE4_INCLUDE_DIR} + PATH_SUFFIXES eigen3 eigen + ) + endif(NOT EIGEN3_INCLUDE_DIR) if(EIGEN3_INCLUDE_DIR) _eigen3_check_version() diff --git a/doc/CustomizingEigen.dox b/doc/CustomizingEigen.dox deleted file mode 100644 index 1b15c69a4..000000000 --- a/doc/CustomizingEigen.dox +++ /dev/null @@ -1,220 +0,0 @@ -namespace Eigen { - -/** \page TopicCustomizingEigen Customizing/Extending Eigen - -Eigen can be extended in several ways, for instance, by defining global methods, \ref ExtendingMatrixBase "by adding custom methods to MatrixBase", adding support to \ref CustomScalarType "custom types" etc. - -\eigenAutoToc - -\section ExtendingMatrixBase Extending MatrixBase (and other classes) - -In this section we will see how to add custom methods to MatrixBase. Since all expressions and matrix types inherit MatrixBase, adding a method to MatrixBase make it immediately available to all expressions ! A typical use case is, for instance, to make Eigen compatible with another API. - -You certainly know that in C++ it is not possible to add methods to an existing class. So how that's possible ? Here the trick is to include in the declaration of MatrixBase a file defined by the preprocessor token \c EIGEN_MATRIXBASE_PLUGIN: -\code -class MatrixBase { - // ... - #ifdef EIGEN_MATRIXBASE_PLUGIN - #include EIGEN_MATRIXBASE_PLUGIN - #endif -}; -\endcode -Therefore to extend MatrixBase with your own methods you just have to create a file with your method declaration and define EIGEN_MATRIXBASE_PLUGIN before you include any Eigen's header file. - -You can extend many of the other classes used in Eigen by defining similarly named preprocessor symbols. For instance, define \c EIGEN_ARRAYBASE_PLUGIN if you want to extend the ArrayBase class. A full list of classes that can be extended in this way and the corresponding preprocessor symbols can be found on our page \ref TopicPreprocessorDirectives. - -Here is an example of an extension file for adding methods to MatrixBase: \n -\b MatrixBaseAddons.h -\code -inline Scalar at(uint i, uint j) const { return this->operator()(i,j); } -inline Scalar& at(uint i, uint j) { return this->operator()(i,j); } -inline Scalar at(uint i) const { return this->operator[](i); } -inline Scalar& at(uint i) { return this->operator[](i); } - -inline RealScalar squaredLength() const { return squaredNorm(); } -inline RealScalar length() const { return norm(); } -inline RealScalar invLength(void) const { return fast_inv_sqrt(squaredNorm()); } - -template<typename OtherDerived> -inline Scalar squaredDistanceTo(const MatrixBase<OtherDerived>& other) const -{ return (derived() - other.derived()).squaredNorm(); } - -template<typename OtherDerived> -inline RealScalar distanceTo(const MatrixBase<OtherDerived>& other) const -{ return internal::sqrt(derived().squaredDistanceTo(other)); } - -inline void scaleTo(RealScalar l) { RealScalar vl = norm(); if (vl>1e-9) derived() *= (l/vl); } - -inline Transpose<Derived> transposed() {return this->transpose();} -inline const Transpose<Derived> transposed() const {return this->transpose();} - -inline uint minComponentId(void) const { int i; this->minCoeff(&i); return i; } -inline uint maxComponentId(void) const { int i; this->maxCoeff(&i); return i; } - -template<typename OtherDerived> -void makeFloor(const MatrixBase<OtherDerived>& other) { derived() = derived().cwiseMin(other.derived()); } -template<typename OtherDerived> -void makeCeil(const MatrixBase<OtherDerived>& other) { derived() = derived().cwiseMax(other.derived()); } - -const CwiseBinaryOp<internal::scalar_sum_op<Scalar>, const Derived, const ConstantReturnType> -operator+(const Scalar& scalar) const -{ return CwiseBinaryOp<internal::scalar_sum_op<Scalar>, const Derived, const ConstantReturnType>(derived(), Constant(rows(),cols(),scalar)); } - -friend const CwiseBinaryOp<internal::scalar_sum_op<Scalar>, const ConstantReturnType, Derived> -operator+(const Scalar& scalar, const MatrixBase<Derived>& mat) -{ return CwiseBinaryOp<internal::scalar_sum_op<Scalar>, const ConstantReturnType, Derived>(Constant(rows(),cols(),scalar), mat.derived()); } -\endcode - -Then one can the following declaration in the config.h or whatever prerequisites header file of his project: -\code -#define EIGEN_MATRIXBASE_PLUGIN "MatrixBaseAddons.h" -\endcode - -\section InheritingFromMatrix Inheriting from Matrix - -Before inheriting from Matrix, be really, I mean REALLY, sure that using -EIGEN_MATRIX_PLUGIN is not what you really want (see previous section). -If you just need to add few members to Matrix, this is the way to go. - -An example of when you actually need to inherit Matrix, is when you -have several layers of heritage such as -MyVerySpecificVector1, MyVerySpecificVector2 -> MyVector1 -> Matrix and -MyVerySpecificVector3, MyVerySpecificVector4 -> MyVector2 -> Matrix. - -In order for your object to work within the %Eigen framework, you need to -define a few members in your inherited class. - -Here is a minimalistic example: - -\include CustomizingEigen_Inheritance.cpp - -Output: \verbinclude CustomizingEigen_Inheritance.out - -This is the kind of error you can get if you don't provide those methods -\verbatim -error: no match for ‘operator=’ in ‘v = Eigen::operator*( -const Eigen::MatrixBase<Eigen::Matrix<double, -0x000000001, 1, 0, -0x000000001, 1> >::Scalar&, -const Eigen::MatrixBase<Eigen::Matrix<double, -0x000000001, 1> >::StorageBaseType&) -(((const Eigen::MatrixBase<Eigen::Matrix<double, -0x000000001, 1> >::StorageBaseType&) -((const Eigen::MatrixBase<Eigen::Matrix<double, -0x000000001, 1> >::StorageBaseType*)(& v))))’ -\endverbatim - -\anchor user_defined_scalars \section CustomScalarType Using custom scalar types - -By default, Eigen currently supports standard floating-point types (\c float, \c double, \c std::complex<float>, \c std::complex<double>, \c long \c double), as well as all native integer types (e.g., \c int, \c unsigned \c int, \c short, etc.), and \c bool. -On x86-64 systems, \c long \c double permits to locally enforces the use of x87 registers with extended accuracy (in comparison to SSE). - -In order to add support for a custom type \c T you need: --# make sure the common operator (+,-,*,/,etc.) are supported by the type \c T --# add a specialization of struct Eigen::NumTraits<T> (see \ref NumTraits) --# define the math functions that makes sense for your type. This includes standard ones like sqrt, pow, sin, tan, conj, real, imag, etc, as well as abs2 which is Eigen specific. - (see the file Eigen/src/Core/MathFunctions.h) - -The math function should be defined in the same namespace than \c T, or in the \c std namespace though that second approach is not recommended. - -Here is a concrete example adding support for the Adolc's \c adouble type. <a href="https://projects.coin-or.org/ADOL-C">Adolc</a> is an automatic differentiation library. The type \c adouble is basically a real value tracking the values of any number of partial derivatives. - -\code -#ifndef ADOLCSUPPORT_H -#define ADOLCSUPPORT_H - -#define ADOLC_TAPELESS -#include <adolc/adouble.h> -#include <Eigen/Core> - -namespace Eigen { - -template<> struct NumTraits<adtl::adouble> - : NumTraits<double> // permits to get the epsilon, dummy_precision, lowest, highest functions -{ - typedef adtl::adouble Real; - typedef adtl::adouble NonInteger; - typedef adtl::adouble Nested; - - enum { - IsComplex = 0, - IsInteger = 0, - IsSigned = 1, - RequireInitialization = 1, - ReadCost = 1, - AddCost = 3, - MulCost = 3 - }; -}; - -} - -namespace adtl { - -inline const adouble& conj(const adouble& x) { return x; } -inline const adouble& real(const adouble& x) { return x; } -inline adouble imag(const adouble&) { return 0.; } -inline adouble abs(const adouble& x) { return fabs(x); } -inline adouble abs2(const adouble& x) { return x*x; } - -} - -#endif // ADOLCSUPPORT_H -\endcode - -This other example adds support for the \c mpq_class type from <a href="https://gmplib.org/">GMP</a>. It shows in particular how to change the way Eigen picks the best pivot during LU factorization. It selects the coefficient with the highest score, where the score is by default the absolute value of a number, but we can define a different score, for instance to prefer pivots with a more compact representation (this is an example, not a recommendation). Note that the scores should always be non-negative and only zero is allowed to have a score of zero. Also, this can interact badly with thresholds for inexact scalar types. - -\code -#include <gmpxx.h> -#include <Eigen/Core> -#include <boost/operators.hpp> - -namespace Eigen { - template<> struct NumTraits<mpq_class> : GenericNumTraits<mpq_class> - { - typedef mpq_class Real; - typedef mpq_class NonInteger; - typedef mpq_class Nested; - - static inline Real epsilon() { return 0; } - static inline Real dummy_precision() { return 0; } - static inline Real digits10() { return 0; } - - enum { - IsInteger = 0, - IsSigned = 1, - IsComplex = 0, - RequireInitialization = 1, - ReadCost = 6, - AddCost = 150, - MulCost = 100 - }; - }; - - namespace internal { - - template<> struct scalar_score_coeff_op<mpq_class> { - struct result_type : boost::totally_ordered1<result_type> { - std::size_t len; - result_type(int i = 0) : len(i) {} // Eigen uses Score(0) and Score() - result_type(mpq_class const& q) : - len(mpz_size(q.get_num_mpz_t())+ - mpz_size(q.get_den_mpz_t())-1) {} - friend bool operator<(result_type x, result_type y) { - // 0 is the worst possible pivot - if (x.len == 0) return y.len > 0; - if (y.len == 0) return false; - // Prefer a pivot with a small representation - return x.len > y.len; - } - friend bool operator==(result_type x, result_type y) { - // Only used to test if the score is 0 - return x.len == y.len; - } - }; - result_type operator()(mpq_class const& x) const { return x; } - }; - } -} -\endcode - -\sa \ref TopicPreprocessorDirectives - -*/ - -} diff --git a/doc/CustomizingEigen_CustomScalar.dox b/doc/CustomizingEigen_CustomScalar.dox new file mode 100644 index 000000000..1ee78cbe5 --- /dev/null +++ b/doc/CustomizingEigen_CustomScalar.dox @@ -0,0 +1,120 @@ +namespace Eigen { + +/** \page TopicCustomizing_CustomScalar Using custom scalar types +\anchor user_defined_scalars + +By default, Eigen currently supports standard floating-point types (\c float, \c double, \c std::complex<float>, \c std::complex<double>, \c long \c double), as well as all native integer types (e.g., \c int, \c unsigned \c int, \c short, etc.), and \c bool. +On x86-64 systems, \c long \c double permits to locally enforces the use of x87 registers with extended accuracy (in comparison to SSE). + +In order to add support for a custom type \c T you need: +-# make sure the common operator (+,-,*,/,etc.) are supported by the type \c T +-# add a specialization of struct Eigen::NumTraits<T> (see \ref NumTraits) +-# define the math functions that makes sense for your type. This includes standard ones like sqrt, pow, sin, tan, conj, real, imag, etc, as well as abs2 which is Eigen specific. + (see the file Eigen/src/Core/MathFunctions.h) + +The math function should be defined in the same namespace than \c T, or in the \c std namespace though that second approach is not recommended. + +Here is a concrete example adding support for the Adolc's \c adouble type. <a href="https://projects.coin-or.org/ADOL-C">Adolc</a> is an automatic differentiation library. The type \c adouble is basically a real value tracking the values of any number of partial derivatives. + +\code +#ifndef ADOLCSUPPORT_H +#define ADOLCSUPPORT_H + +#define ADOLC_TAPELESS +#include <adolc/adouble.h> +#include <Eigen/Core> + +namespace Eigen { + +template<> struct NumTraits<adtl::adouble> + : NumTraits<double> // permits to get the epsilon, dummy_precision, lowest, highest functions +{ + typedef adtl::adouble Real; + typedef adtl::adouble NonInteger; + typedef adtl::adouble Nested; + + enum { + IsComplex = 0, + IsInteger = 0, + IsSigned = 1, + RequireInitialization = 1, + ReadCost = 1, + AddCost = 3, + MulCost = 3 + }; +}; + +} + +namespace adtl { + +inline const adouble& conj(const adouble& x) { return x; } +inline const adouble& real(const adouble& x) { return x; } +inline adouble imag(const adouble&) { return 0.; } +inline adouble abs(const adouble& x) { return fabs(x); } +inline adouble abs2(const adouble& x) { return x*x; } + +} + +#endif // ADOLCSUPPORT_H +\endcode + +This other example adds support for the \c mpq_class type from <a href="https://gmplib.org/">GMP</a>. It shows in particular how to change the way Eigen picks the best pivot during LU factorization. It selects the coefficient with the highest score, where the score is by default the absolute value of a number, but we can define a different score, for instance to prefer pivots with a more compact representation (this is an example, not a recommendation). Note that the scores should always be non-negative and only zero is allowed to have a score of zero. Also, this can interact badly with thresholds for inexact scalar types. + +\code +#include <gmpxx.h> +#include <Eigen/Core> +#include <boost/operators.hpp> + +namespace Eigen { + template<> struct NumTraits<mpq_class> : GenericNumTraits<mpq_class> + { + typedef mpq_class Real; + typedef mpq_class NonInteger; + typedef mpq_class Nested; + + static inline Real epsilon() { return 0; } + static inline Real dummy_precision() { return 0; } + static inline Real digits10() { return 0; } + + enum { + IsInteger = 0, + IsSigned = 1, + IsComplex = 0, + RequireInitialization = 1, + ReadCost = 6, + AddCost = 150, + MulCost = 100 + }; + }; + + namespace internal { + + template<> struct scalar_score_coeff_op<mpq_class> { + struct result_type : boost::totally_ordered1<result_type> { + std::size_t len; + result_type(int i = 0) : len(i) {} // Eigen uses Score(0) and Score() + result_type(mpq_class const& q) : + len(mpz_size(q.get_num_mpz_t())+ + mpz_size(q.get_den_mpz_t())-1) {} + friend bool operator<(result_type x, result_type y) { + // 0 is the worst possible pivot + if (x.len == 0) return y.len > 0; + if (y.len == 0) return false; + // Prefer a pivot with a small representation + return x.len > y.len; + } + friend bool operator==(result_type x, result_type y) { + // Only used to test if the score is 0 + return x.len == y.len; + } + }; + result_type operator()(mpq_class const& x) const { return x; } + }; + } +} +\endcode + +*/ + +} diff --git a/doc/CustomizingEigen_InheritingMatrix.dox b/doc/CustomizingEigen_InheritingMatrix.dox new file mode 100644 index 000000000..b21e55433 --- /dev/null +++ b/doc/CustomizingEigen_InheritingMatrix.dox @@ -0,0 +1,34 @@ +namespace Eigen { + +/** \page TopicCustomizing_InheritingMatrix Inheriting from Matrix + +Before inheriting from Matrix, be really, I mean REALLY, sure that using +EIGEN_MATRIX_PLUGIN is not what you really want (see previous section). +If you just need to add few members to Matrix, this is the way to go. + +An example of when you actually need to inherit Matrix, is when you +have several layers of heritage such as +MyVerySpecificVector1, MyVerySpecificVector2 -> MyVector1 -> Matrix and +MyVerySpecificVector3, MyVerySpecificVector4 -> MyVector2 -> Matrix. + +In order for your object to work within the %Eigen framework, you need to +define a few members in your inherited class. + +Here is a minimalistic example: + +\include CustomizingEigen_Inheritance.cpp + +Output: \verbinclude CustomizingEigen_Inheritance.out + +This is the kind of error you can get if you don't provide those methods +\verbatim +error: no match for ‘operator=’ in ‘v = Eigen::operator*( +const Eigen::MatrixBase<Eigen::Matrix<double, -0x000000001, 1, 0, -0x000000001, 1> >::Scalar&, +const Eigen::MatrixBase<Eigen::Matrix<double, -0x000000001, 1> >::StorageBaseType&) +(((const Eigen::MatrixBase<Eigen::Matrix<double, -0x000000001, 1> >::StorageBaseType&) +((const Eigen::MatrixBase<Eigen::Matrix<double, -0x000000001, 1> >::StorageBaseType*)(& v))))’ +\endverbatim + +*/ + +} diff --git a/doc/CustomizingEigen_NullaryExpr.dox b/doc/CustomizingEigen_NullaryExpr.dox new file mode 100644 index 000000000..d70f81065 --- /dev/null +++ b/doc/CustomizingEigen_NullaryExpr.dox @@ -0,0 +1,59 @@ +namespace Eigen { + +/** \page TopicCustomizing_NullaryExpr Matrix manipulation via nullary-expressions + + +The main purpose of the class CwiseNullaryOp is to define \em procedural matrices such as constant or random matrices as returned by the Ones(), Zero(), Constant(), Identity() and Random() methods. +Nevertheless, with some imagination it is possible to accomplish very sophisticated matrix manipulation with minimal efforts such that \ref TopicNewExpressionType "implementing new expression" is rarely needed. + +\section NullaryExpr_Circulant Example 1: circulant matrix + +To explore these possibilities let us start with the \em circulant example of the \ref TopicNewExpressionType "implementing new expression" topic. +Let us recall that a circulant matrix is a matrix where each column is the same as the +column to the left, except that it is cyclically shifted downwards. +For example, here is a 4-by-4 circulant matrix: +\f[ \begin{bmatrix} + 1 & 8 & 4 & 2 \\ + 2 & 1 & 8 & 4 \\ + 4 & 2 & 1 & 8 \\ + 8 & 4 & 2 & 1 +\end{bmatrix} \f] +A circulant matrix is uniquely determined by its first column. We wish +to write a function \c makeCirculant which, given the first column, +returns an expression representing the circulant matrix. + +For this exercise, the return type of \c makeCirculant will be a CwiseNullaryOp that we need to instantiate with: +1 - a proper \c circulant_functor storing the input vector and implementing the adequate coefficient accessor \c operator(i,j) +2 - a template instantiation of class Matrix conveying compile-time information such as the scalar type, sizes, and preferred storage layout. + +Calling \c ArgType the type of the input vector, we can construct the equivalent squared Matrix type as follows: + +\snippet make_circulant2.cpp square + +This little helper structure will help us to implement our \c makeCirculant function as follows: + +\snippet make_circulant2.cpp makeCirculant + +As usual, our function takes as argument a \c MatrixBase (see this \ref TopicFunctionTakingEigenTypes "page" for more details). +Then, the CwiseNullaryOp object is constructed through the DenseBase::NullaryExpr static method with the adequate runtime sizes. + +Then, we need to implement our \c circulant_functor, which is a straightforward exercise: + +\snippet make_circulant2.cpp circulant_func + +We are now all set to try our new feature: + +\snippet make_circulant2.cpp main + + +If all the fragments are combined, the following output is produced, +showing that the program works as expected: + +\include make_circulant2.out + +This implementation of \c makeCirculant is much simpler than \ref TopicNewExpressionType "defining a new expression" from scratch. + +*/ + +} + diff --git a/doc/CustomizingEigen_Plugins.dox b/doc/CustomizingEigen_Plugins.dox new file mode 100644 index 000000000..d88f2409b --- /dev/null +++ b/doc/CustomizingEigen_Plugins.dox @@ -0,0 +1,69 @@ +namespace Eigen { + +/** \page TopicCustomizing_Plugins Extending MatrixBase (and other classes) + +In this section we will see how to add custom methods to MatrixBase. Since all expressions and matrix types inherit MatrixBase, adding a method to MatrixBase make it immediately available to all expressions ! A typical use case is, for instance, to make Eigen compatible with another API. + +You certainly know that in C++ it is not possible to add methods to an existing class. So how that's possible ? Here the trick is to include in the declaration of MatrixBase a file defined by the preprocessor token \c EIGEN_MATRIXBASE_PLUGIN: +\code +class MatrixBase { + // ... + #ifdef EIGEN_MATRIXBASE_PLUGIN + #include EIGEN_MATRIXBASE_PLUGIN + #endif +}; +\endcode +Therefore to extend MatrixBase with your own methods you just have to create a file with your method declaration and define EIGEN_MATRIXBASE_PLUGIN before you include any Eigen's header file. + +You can extend many of the other classes used in Eigen by defining similarly named preprocessor symbols. For instance, define \c EIGEN_ARRAYBASE_PLUGIN if you want to extend the ArrayBase class. A full list of classes that can be extended in this way and the corresponding preprocessor symbols can be found on our page \ref TopicPreprocessorDirectives. + +Here is an example of an extension file for adding methods to MatrixBase: \n +\b MatrixBaseAddons.h +\code +inline Scalar at(uint i, uint j) const { return this->operator()(i,j); } +inline Scalar& at(uint i, uint j) { return this->operator()(i,j); } +inline Scalar at(uint i) const { return this->operator[](i); } +inline Scalar& at(uint i) { return this->operator[](i); } + +inline RealScalar squaredLength() const { return squaredNorm(); } +inline RealScalar length() const { return norm(); } +inline RealScalar invLength(void) const { return fast_inv_sqrt(squaredNorm()); } + +template<typename OtherDerived> +inline Scalar squaredDistanceTo(const MatrixBase<OtherDerived>& other) const +{ return (derived() - other.derived()).squaredNorm(); } + +template<typename OtherDerived> +inline RealScalar distanceTo(const MatrixBase<OtherDerived>& other) const +{ return internal::sqrt(derived().squaredDistanceTo(other)); } + +inline void scaleTo(RealScalar l) { RealScalar vl = norm(); if (vl>1e-9) derived() *= (l/vl); } + +inline Transpose<Derived> transposed() {return this->transpose();} +inline const Transpose<Derived> transposed() const {return this->transpose();} + +inline uint minComponentId(void) const { int i; this->minCoeff(&i); return i; } +inline uint maxComponentId(void) const { int i; this->maxCoeff(&i); return i; } + +template<typename OtherDerived> +void makeFloor(const MatrixBase<OtherDerived>& other) { derived() = derived().cwiseMin(other.derived()); } +template<typename OtherDerived> +void makeCeil(const MatrixBase<OtherDerived>& other) { derived() = derived().cwiseMax(other.derived()); } + +const CwiseBinaryOp<internal::scalar_sum_op<Scalar>, const Derived, const ConstantReturnType> +operator+(const Scalar& scalar) const +{ return CwiseBinaryOp<internal::scalar_sum_op<Scalar>, const Derived, const ConstantReturnType>(derived(), Constant(rows(),cols(),scalar)); } + +friend const CwiseBinaryOp<internal::scalar_sum_op<Scalar>, const ConstantReturnType, Derived> +operator+(const Scalar& scalar, const MatrixBase<Derived>& mat) +{ return CwiseBinaryOp<internal::scalar_sum_op<Scalar>, const ConstantReturnType, Derived>(Constant(rows(),cols(),scalar), mat.derived()); } +\endcode + +Then one can the following declaration in the config.h or whatever prerequisites header file of his project: +\code +#define EIGEN_MATRIXBASE_PLUGIN "MatrixBaseAddons.h" +\endcode + +*/ + +} diff --git a/doc/Doxyfile.in b/doc/Doxyfile.in index 058c88b97..6f8d6bc01 100644 --- a/doc/Doxyfile.in +++ b/doc/Doxyfile.in @@ -1609,7 +1609,10 @@ EXPAND_AS_DEFINED = EIGEN_MAKE_TYPEDEFS \ EIGEN_MATHFUNC_IMPL \ _EIGEN_GENERIC_PUBLIC_INTERFACE \ EIGEN_ARRAY_DECLARE_GLOBAL_UNARY \ - EIGEN_EMPTY + EIGEN_EMPTY \ + EIGEN_EULER_ANGLES_TYPEDEFS \ + EIGEN_EULER_ANGLES_SINGLE_TYPEDEF \ + EIGEN_EULER_SYSTEM_TYPEDEF # If the SKIP_FUNCTION_MACROS tag is set to YES (the default) then # doxygen's preprocessor will remove all references to function-like macros diff --git a/doc/Manual.dox b/doc/Manual.dox index db73e1199..a08609ad7 100644 --- a/doc/Manual.dox +++ b/doc/Manual.dox @@ -3,22 +3,31 @@ namespace Eigen { +/** \page UserManual_CustomizingEigen Extending/Customizing Eigen + %Eigen can be extended in several ways, for instance, by defining global methods, by inserting custom methods within main %Eigen's classes through the \ref TopicCustomizing_Plugins "plugin" mechanism, by adding support to \ref TopicCustomizing_CustomScalar "custom scalar types" etc. See below for the respective sub-topics. + - \subpage TopicCustomizing_Plugins + - \subpage TopicCustomizing_InheritingMatrix + - \subpage TopicCustomizing_CustomScalar + - \subpage TopicCustomizing_NullaryExpr + - \subpage TopicNewExpressionType + \sa \ref TopicPreprocessorDirectives +*/ + + /** \page UserManual_Generalities General topics - \subpage Eigen2ToEigen3 - \subpage TopicFunctionTakingEigenTypes - \subpage TopicPreprocessorDirectives - \subpage TopicAssertions - - \subpage TopicCustomizingEigen - \subpage TopicMultiThreading - \subpage TopicUsingBlasLapack - \subpage TopicUsingIntelMKL - \subpage TopicCUDA - \subpage TopicPitfalls - \subpage TopicTemplateKeyword - - \subpage TopicNewExpressionType - \subpage UserManual_UnderstandingEigen */ - + /** \page UserManual_UnderstandingEigen Understanding Eigen - \subpage TopicInsideEigenExample - \subpage TopicClassHierarchy diff --git a/doc/NewExpressionType.dox b/doc/NewExpressionType.dox index ad8b7f86b..c2f243312 100644 --- a/doc/NewExpressionType.dox +++ b/doc/NewExpressionType.dox @@ -2,6 +2,12 @@ namespace Eigen { /** \page TopicNewExpressionType Adding a new expression type +<!--<span style="font-size:130%; color:red; font-weight: 900;"></span>--> +\warning +Disclaimer: this page is tailored to very advanced users who are not afraid of dealing with some %Eigen's internal aspects. +In most cases, a custom expression can be avoided by either using custom \ref MatrixBase::unaryExpr "unary" or \ref MatrixBase::binaryExpr "binary" functors, +while extremely complex matrix manipulations can be achieved by a nullary functors as described in the \ref TopicCustomizing_NullaryExpr "previous page". + This page describes with the help of an example how to implement a new light-weight expression type in %Eigen. This consists of three parts: the expression type itself, a traits class containing compile-time @@ -130,7 +136,7 @@ function can be called. If all the fragments are combined, the following output is produced, showing that the program works as expected: -\verbinclude make_circulant.out +\include make_circulant.out */ } diff --git a/doc/Overview.dox b/doc/Overview.dox index 9ab96233a..dbb49bd21 100644 --- a/doc/Overview.dox +++ b/doc/Overview.dox @@ -17,7 +17,9 @@ You're a MatLab user? There is also a <a href="AsciiQuickReference.txt">short AS The \b main \b documentation is organized into \em chapters covering different domains of features. They are themselves composed of \em user \em manual pages describing the different features in a comprehensive way, and \em reference pages that gives you access to the API documentation through the related Eigen's \em modules and \em classes. -Under the \subpage UserManual_Generalities section, you will find documentation on more general topics such as preprocessor directives, controlling assertions, multi-threading, MKL support, some Eigen's internal insights, and much more... +Under the \subpage UserManual_CustomizingEigen section, you will find discussions and examples on extending %Eigen's features and supporting custom scalar types. + +Under the \subpage UserManual_Generalities section, you will find documentation on more general topics such as preprocessor directives, controlling assertions, multi-threading, MKL support, some Eigen's internal insights, and much more... Finally, do not miss the search engine, useful to quickly get to the documentation of a given class or function. diff --git a/doc/snippets/SparseMatrix_coeffs.cpp b/doc/snippets/SparseMatrix_coeffs.cpp new file mode 100644 index 000000000..f71a69b07 --- /dev/null +++ b/doc/snippets/SparseMatrix_coeffs.cpp @@ -0,0 +1,9 @@ +SparseMatrix<double> A(3,3); +A.insert(1,2) = 0; +A.insert(0,1) = 1; +A.insert(2,0) = 2; +A.makeCompressed(); +cout << "The matrix A is:" << endl << MatrixXd(A) << endl; +cout << "it has " << A.nonZeros() << " stored non zero coefficients that are: " << A.coeffs().transpose() << endl; +A.coeffs() += 10; +cout << "After adding 10 to every stored non zero coefficient, the matrix A is:" << endl << MatrixXd(A) << endl; diff --git a/doc/special_examples/random_cpp11.cpp b/doc/special_examples/random_cpp11.cpp index adc3c110c..33744c051 100644 --- a/doc/special_examples/random_cpp11.cpp +++ b/doc/special_examples/random_cpp11.cpp @@ -7,7 +7,7 @@ using namespace Eigen; int main() { std::default_random_engine generator; std::poisson_distribution<int> distribution(4.1); - auto poisson = [&] (Eigen::Index) {return distribution(generator);}; + auto poisson = [&] () {return distribution(generator);}; RowVectorXi v = RowVectorXi::NullaryExpr(10, poisson ); std::cout << v << "\n"; diff --git a/test/adjoint.cpp b/test/adjoint.cpp index 9c895e0ac..bdea51c10 100644 --- a/test/adjoint.cpp +++ b/test/adjoint.cpp @@ -169,7 +169,7 @@ void test_adjoint() // test a large static matrix only once CALL_SUBTEST_7( adjoint(Matrix<float, 100, 100>()) ); -#ifdef EIGEN_TEST_PART_4 +#ifdef EIGEN_TEST_PART_13 { MatrixXcf a(10,10), b(10,10); VERIFY_RAISES_ASSERT(a = a.transpose()); @@ -187,6 +187,13 @@ void test_adjoint() a.transpose() = a.adjoint(); a.transpose() += a.adjoint(); a.transpose() += a.adjoint() + b; + + // regression tests for check_for_aliasing + MatrixXd c(10,10); + c = 1.0 * MatrixXd::Ones(10,10) + c; + c = MatrixXd::Ones(10,10) * 1.0 + c; + c = c + MatrixXd::Ones(10,10) .cwiseProduct( MatrixXd::Zero(10,10) ); + c = MatrixXd::Ones(10,10) * MatrixXd::Zero(10,10); } #endif } diff --git a/test/cholesky.cpp b/test/cholesky.cpp index 12efd2d60..9a1f3792c 100644 --- a/test/cholesky.cpp +++ b/test/cholesky.cpp @@ -154,6 +154,7 @@ template<typename MatrixType> void cholesky(const MatrixType& m) SquareMatrixType symmLo = symm.template triangularView<Lower>(); LDLT<SquareMatrixType,Lower> ldltlo(symmLo); + VERIFY(ldltlo.info()==Success); VERIFY_IS_APPROX(symm, ldltlo.reconstructedMatrix()); vecX = ldltlo.solve(vecB); VERIFY_IS_APPROX(symm * vecX, vecB); @@ -170,6 +171,7 @@ template<typename MatrixType> void cholesky(const MatrixType& m) LDLT<SquareMatrixType,Upper> ldltup(symmUp); + VERIFY(ldltup.info()==Success); VERIFY_IS_APPROX(symm, ldltup.reconstructedMatrix()); vecX = ldltup.solve(vecB); VERIFY_IS_APPROX(symm * vecX, vecB); @@ -331,6 +333,7 @@ template<typename MatrixType> void cholesky_cplx(const MatrixType& m) RealMatrixType symmLo = symm.template triangularView<Lower>(); LDLT<RealMatrixType,Lower> ldltlo(symmLo); + VERIFY(ldltlo.info()==Success); VERIFY_IS_APPROX(symm, ldltlo.reconstructedMatrix()); vecX = ldltlo.solve(vecB); VERIFY_IS_APPROX(symm * vecX, vecB); @@ -367,35 +370,88 @@ template<typename MatrixType> void cholesky_definiteness(const MatrixType& m) { mat << 1, 0, 0, -1; ldlt.compute(mat); + VERIFY(ldlt.info()==Success); VERIFY(!ldlt.isNegative()); VERIFY(!ldlt.isPositive()); } { mat << 1, 2, 2, 1; ldlt.compute(mat); + VERIFY(ldlt.info()==Success); VERIFY(!ldlt.isNegative()); VERIFY(!ldlt.isPositive()); } { mat << 0, 0, 0, 0; ldlt.compute(mat); + VERIFY(ldlt.info()==Success); VERIFY(ldlt.isNegative()); VERIFY(ldlt.isPositive()); } { mat << 0, 0, 0, 1; ldlt.compute(mat); + VERIFY(ldlt.info()==Success); VERIFY(!ldlt.isNegative()); VERIFY(ldlt.isPositive()); } { mat << -1, 0, 0, 0; ldlt.compute(mat); + VERIFY(ldlt.info()==Success); VERIFY(ldlt.isNegative()); VERIFY(!ldlt.isPositive()); } } +template<typename> +void cholesky_faillure_cases() +{ + MatrixXd mat; + LDLT<MatrixXd> ldlt; + + { + mat.resize(2,2); + mat << 0, 1, 1, 0; + ldlt.compute(mat); + VERIFY_IS_NOT_APPROX(mat,ldlt.reconstructedMatrix()); + VERIFY(ldlt.info()==NumericalIssue); + } + { + mat.resize(3,3); + mat << -1, -3, 3, + -3, -8.9999999999999999999, 1, + 3, 1, 0; + ldlt.compute(mat); + VERIFY(ldlt.info()==NumericalIssue); + VERIFY_IS_NOT_APPROX(mat,ldlt.reconstructedMatrix()); + } + { + mat.resize(3,3); + mat << 1, 2, 3, + 2, 4, 1, + 3, 1, 0; + ldlt.compute(mat); + VERIFY(ldlt.info()==NumericalIssue); + VERIFY_IS_NOT_APPROX(mat,ldlt.reconstructedMatrix()); + } + + { + mat.resize(8,8); + mat << 0.1, 0, -0.1, 0, 0, 0, 1, 0, + 0, 4.24667, 0, 2.00333, 0, 0, 0, 0, + -0.1, 0, 0.2, 0, -0.1, 0, 0, 0, + 0, 2.00333, 0, 8.49333, 0, 2.00333, 0, 0, + 0, 0, -0.1, 0, 0.1, 0, 0, 1, + 0, 0, 0, 2.00333, 0, 4.24667, 0, 0, + 1, 0, 0, 0, 0, 0, 0, 0, + 0, 0, 0, 0, 1, 0, 0, 0; + ldlt.compute(mat); + VERIFY(ldlt.info()==NumericalIssue); + VERIFY_IS_NOT_APPROX(mat,ldlt.reconstructedMatrix()); + } +} + template<typename MatrixType> void cholesky_verify_assert() { MatrixType tmp; @@ -445,5 +501,7 @@ void test_cholesky() CALL_SUBTEST_9( LLT<MatrixXf>(10) ); CALL_SUBTEST_9( LDLT<MatrixXf>(10) ); + CALL_SUBTEST_2( cholesky_faillure_cases<void>() ); + TEST_SET_BUT_UNUSED_VARIABLE(nb_temporaries) } diff --git a/test/cuda_basic.cu b/test/cuda_basic.cu index b36ed888d..cb2e4167a 100644 --- a/test/cuda_basic.cu +++ b/test/cuda_basic.cu @@ -1,4 +1,11 @@ - +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2015-2016 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/. // workaround issue between gcc >= 4.7 and cuda 5.5 #if (defined __GNUC__) && (__GNUC__>4 || __GNUC_MINOR__>=7) @@ -12,10 +19,15 @@ #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int #include <math_constants.h> +#include <cuda.h> +#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500 +#include <cuda_fp16.h> +#endif #include "main.h" #include "cuda_common.h" -#include <Eigen/Eigenvalues> +// Check that dense modules can be properly parsed by nvcc +#include <Eigen/Dense> // struct Foo{ // EIGEN_DEVICE_FUNC diff --git a/test/eigensolver_generalized_real.cpp b/test/eigensolver_generalized_real.cpp index da14482de..9c0838ba4 100644 --- a/test/eigensolver_generalized_real.cpp +++ b/test/eigensolver_generalized_real.cpp @@ -7,6 +7,7 @@ // 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/. +#define EIGEN_RUNTIME_NO_MALLOC #include "main.h" #include <limits> #include <Eigen/Eigenvalues> @@ -42,18 +43,31 @@ template<typename MatrixType> void generalized_eigensolver_real(const MatrixType VectorType realEigenvalues = eig.eigenvalues().real(); std::sort(realEigenvalues.data(), realEigenvalues.data()+realEigenvalues.size()); VERIFY_IS_APPROX(realEigenvalues, symmEig.eigenvalues()); + + // check eigenvectors + typename GeneralizedEigenSolver<MatrixType>::EigenvectorsType D = eig.eigenvalues().asDiagonal(); + typename GeneralizedEigenSolver<MatrixType>::EigenvectorsType V = eig.eigenvectors(); + VERIFY_IS_APPROX(spdA*V, spdB*V*D); } // non symmetric case: { - GeneralizedEigenSolver<MatrixType> eig(a,b); + GeneralizedEigenSolver<MatrixType> eig(rows); + // TODO enable full-prealocation of required memory, this probably requires an in-place mode for HessenbergDecomposition + //Eigen::internal::set_is_malloc_allowed(false); + eig.compute(a,b); + //Eigen::internal::set_is_malloc_allowed(true); for(Index k=0; k<cols; ++k) { Matrix<ComplexScalar,Dynamic,Dynamic> tmp = (eig.betas()(k)*a).template cast<ComplexScalar>() - eig.alphas()(k)*b; - if(tmp.norm()>(std::numeric_limits<Scalar>::min)()) + if(tmp.size()>1 && tmp.norm()>(std::numeric_limits<Scalar>::min)()) tmp /= tmp.norm(); VERIFY_IS_MUCH_SMALLER_THAN( std::abs(tmp.determinant()), Scalar(1) ); } + // check eigenvectors + typename GeneralizedEigenSolver<MatrixType>::EigenvectorsType D = eig.eigenvalues().asDiagonal(); + typename GeneralizedEigenSolver<MatrixType>::EigenvectorsType V = eig.eigenvectors(); + VERIFY_IS_APPROX(a*V, b*V*D); } // regression test for bug 1098 diff --git a/test/integer_types.cpp b/test/integer_types.cpp index 950f8e9be..a21f73a81 100644 --- a/test/integer_types.cpp +++ b/test/integer_types.cpp @@ -158,4 +158,12 @@ void test_integer_types() CALL_SUBTEST_8( integer_type_tests(Matrix<unsigned long long, Dynamic, 5>(1, 5)) ); } +#ifdef EIGEN_TEST_PART_9 + VERIFY_IS_EQUAL(internal::scalar_div_cost<int>::value, 8); + VERIFY_IS_EQUAL(internal::scalar_div_cost<unsigned int>::value, 8); + if(sizeof(long)>sizeof(int)) { + VERIFY(internal::scalar_div_cost<long>::value > internal::scalar_div_cost<int>::value); + VERIFY(internal::scalar_div_cost<unsigned long>::value > internal::scalar_div_cost<int>::value); + } +#endif } diff --git a/test/nullary.cpp b/test/nullary.cpp index cb87695ee..9063c6de8 100644 --- a/test/nullary.cpp +++ b/test/nullary.cpp @@ -104,13 +104,29 @@ void testVectorType(const VectorType& base) template<typename MatrixType> void testMatrixType(const MatrixType& m) { + using std::abs; const Index rows = m.rows(); const Index cols = m.cols(); + typedef typename MatrixType::Scalar Scalar; + typedef typename MatrixType::RealScalar RealScalar; + + Scalar s1; + do { + s1 = internal::random<Scalar>(); + } while(abs(s1)<RealScalar(1e-5) && (!NumTraits<Scalar>::IsInteger)); MatrixType A; A.setIdentity(rows, cols); VERIFY(equalsIdentity(A)); VERIFY(equalsIdentity(MatrixType::Identity(rows, cols))); + + + A = MatrixType::Constant(rows,cols,s1); + Index i = internal::random<Index>(0,rows-1); + Index j = internal::random<Index>(0,cols-1); + VERIFY_IS_APPROX( MatrixType::Constant(rows,cols,s1)(i,j), s1 ); + VERIFY_IS_APPROX( MatrixType::Constant(rows,cols,s1).coeff(i,j), s1 ); + VERIFY_IS_APPROX( A(i,j), s1 ); } void test_nullary() @@ -137,4 +153,47 @@ void test_nullary() // Assignment of a RowVectorXd to a MatrixXd (regression test for bug #79). VERIFY( (MatrixXd(RowVectorXd::LinSpaced(3, 0, 1)) - RowVector3d(0, 0.5, 1)).norm() < std::numeric_limits<double>::epsilon() ); #endif + +#ifdef EIGEN_TEST_PART_10 + // check some internal logic + VERIFY(( internal::has_nullary_operator<internal::scalar_constant_op<double> >::value )); + VERIFY(( !internal::has_unary_operator<internal::scalar_constant_op<double> >::value )); + VERIFY(( !internal::has_binary_operator<internal::scalar_constant_op<double> >::value )); + VERIFY(( internal::functor_has_linear_access<internal::scalar_constant_op<double> >::ret )); + + VERIFY(( !internal::has_nullary_operator<internal::scalar_identity_op<double> >::value )); + VERIFY(( !internal::has_unary_operator<internal::scalar_identity_op<double> >::value )); + VERIFY(( internal::has_binary_operator<internal::scalar_identity_op<double> >::value )); + VERIFY(( !internal::functor_has_linear_access<internal::scalar_identity_op<double> >::ret )); + + VERIFY(( !internal::has_nullary_operator<internal::linspaced_op<float,float,false> >::value )); + VERIFY(( internal::has_unary_operator<internal::linspaced_op<float,float,false> >::value )); + VERIFY(( !internal::has_binary_operator<internal::linspaced_op<float,float,false> >::value )); + VERIFY(( internal::functor_has_linear_access<internal::linspaced_op<float,float,false> >::ret )); + + // Regression unit test for a weird MSVC bug. + // Search "nullary_wrapper_workaround_msvc" in CoreEvaluators.h for the details. + // See also traits<Ref>::match. + { + MatrixXf A = MatrixXf::Random(3,3); + Ref<const MatrixXf> R = 2.0*A; + VERIFY_IS_APPROX(R, A+A); + + Ref<const MatrixXf> R1 = MatrixXf::Random(3,3)+A; + + VectorXi V = VectorXi::Random(3); + Ref<const VectorXi> R2 = VectorXi::LinSpaced(3,1,3)+V; + VERIFY_IS_APPROX(R2, V+Vector3i(1,2,3)); + + VERIFY(( internal::has_nullary_operator<internal::scalar_constant_op<float> >::value )); + VERIFY(( !internal::has_unary_operator<internal::scalar_constant_op<float> >::value )); + VERIFY(( !internal::has_binary_operator<internal::scalar_constant_op<float> >::value )); + VERIFY(( internal::functor_has_linear_access<internal::scalar_constant_op<float> >::ret )); + + VERIFY(( !internal::has_nullary_operator<internal::linspaced_op<int,int,false> >::value )); + VERIFY(( internal::has_unary_operator<internal::linspaced_op<int,int,false> >::value )); + VERIFY(( !internal::has_binary_operator<internal::linspaced_op<int,int,false> >::value )); + VERIFY(( internal::functor_has_linear_access<internal::linspaced_op<int,int,false> >::ret )); + } +#endif } diff --git a/test/packetmath.cpp b/test/packetmath.cpp index 77cbf3e5b..77514d8a0 100644 --- a/test/packetmath.cpp +++ b/test/packetmath.cpp @@ -402,8 +402,8 @@ template<typename Scalar> void packetmath_real() data1[internal::random<int>(0, PacketSize)] = 0; CHECK_CWISE1_IF(PacketTraits::HasSqrt, std::sqrt, internal::psqrt); CHECK_CWISE1_IF(PacketTraits::HasLog, std::log, internal::plog); - CHECK_CWISE1_IF(PacketTraits::HasLog1p, std::log1p, internal::plog1p); #if EIGEN_HAS_C99_MATH && (__cplusplus > 199711L) + CHECK_CWISE1_IF(PacketTraits::HasLog1p, std::log1p, internal::plog1p); CHECK_CWISE1_IF(internal::packet_traits<Scalar>::HasLGamma, std::lgamma, internal::plgamma); CHECK_CWISE1_IF(internal::packet_traits<Scalar>::HasErf, std::erf, internal::perf); CHECK_CWISE1_IF(internal::packet_traits<Scalar>::HasErfc, std::erfc, internal::perfc); diff --git a/test/prec_inverse_4x4.cpp b/test/prec_inverse_4x4.cpp index c4ef2d4bd..eb6ad18c9 100644 --- a/test/prec_inverse_4x4.cpp +++ b/test/prec_inverse_4x4.cpp @@ -53,14 +53,29 @@ template<typename MatrixType> void inverse_general_4x4(int repeat) // FIXME that 1.25 used to be 1.2 until we tested gcc 4.1 on 30 June 2010 and got 1.21. VERIFY(error_avg < (NumTraits<Scalar>::IsComplex ? 8.0 : 1.25)); VERIFY(error_max < (NumTraits<Scalar>::IsComplex ? 64.0 : 20.0)); + + { + int s = 5;//internal::random<int>(4,10); + int i = 0;//internal::random<int>(0,s-4); + int j = 0;//internal::random<int>(0,s-4); + Matrix<Scalar,5,5> mat(s,s); + mat.setRandom(); + MatrixType submat = mat.template block<4,4>(i,j); + MatrixType mat_inv = mat.template block<4,4>(i,j).inverse(); + VERIFY_IS_APPROX(mat_inv, submat.inverse()); + mat.template block<4,4>(i,j) = submat.inverse(); + VERIFY_IS_APPROX(mat_inv, (mat.template block<4,4>(i,j))); + } } void test_prec_inverse_4x4() { CALL_SUBTEST_1((inverse_permutation_4x4<Matrix4f>())); CALL_SUBTEST_1(( inverse_general_4x4<Matrix4f>(200000 * g_repeat) )); + CALL_SUBTEST_1(( inverse_general_4x4<Matrix<float,4,4,RowMajor> >(200000 * g_repeat) )); CALL_SUBTEST_2((inverse_permutation_4x4<Matrix<double,4,4,RowMajor> >())); + CALL_SUBTEST_2(( inverse_general_4x4<Matrix<double,4,4,ColMajor> >(200000 * g_repeat) )); CALL_SUBTEST_2(( inverse_general_4x4<Matrix<double,4,4,RowMajor> >(200000 * g_repeat) )); CALL_SUBTEST_3((inverse_permutation_4x4<Matrix4cf>())); diff --git a/test/product.h b/test/product.h index 27976a4ae..3b6511270 100644 --- a/test/product.h +++ b/test/product.h @@ -119,6 +119,14 @@ template<typename MatrixType> void product(const MatrixType& m) res.noalias() -= square + m1 * m2.transpose(); VERIFY_IS_APPROX(res, square + m1 * m2.transpose()); + // test d ?= a-b*c rules + res.noalias() = square - m1 * m2.transpose(); + VERIFY_IS_APPROX(res, square - m1 * m2.transpose()); + res.noalias() += square - m1 * m2.transpose(); + VERIFY_IS_APPROX(res, 2*(square - m1 * m2.transpose())); + res.noalias() -= square - m1 * m2.transpose(); + VERIFY_IS_APPROX(res, square - m1 * m2.transpose()); + tm1 = m1; VERIFY_IS_APPROX(tm1.transpose() * v1, m1.transpose() * v1); @@ -160,6 +168,29 @@ template<typename MatrixType> void product(const MatrixType& m) VERIFY_IS_APPROX(res2.block(0,0,1,cols).noalias() = m1.block(0,0,1,cols) * square2, (ref2.row(0) = m1.row(0) * square2)); } + // vector.block() (see bug 1283) + { + RowVectorType w1(rows); + VERIFY_IS_APPROX(square * v1.block(0,0,rows,1), square * v1); + VERIFY_IS_APPROX(w1.noalias() = square * v1.block(0,0,rows,1), square * v1); + VERIFY_IS_APPROX(w1.block(0,0,rows,1).noalias() = square * v1.block(0,0,rows,1), square * v1); + + Matrix<Scalar,1,MatrixType::ColsAtCompileTime> w2(cols); + VERIFY_IS_APPROX(vc2.block(0,0,cols,1).transpose() * square2, vc2.transpose() * square2); + VERIFY_IS_APPROX(w2.noalias() = vc2.block(0,0,cols,1).transpose() * square2, vc2.transpose() * square2); + VERIFY_IS_APPROX(w2.block(0,0,1,cols).noalias() = vc2.block(0,0,cols,1).transpose() * square2, vc2.transpose() * square2); + + vc2 = square2.block(0,0,1,cols).transpose(); + VERIFY_IS_APPROX(square2.block(0,0,1,cols) * square2, vc2.transpose() * square2); + VERIFY_IS_APPROX(w2.noalias() = square2.block(0,0,1,cols) * square2, vc2.transpose() * square2); + VERIFY_IS_APPROX(w2.block(0,0,1,cols).noalias() = square2.block(0,0,1,cols) * square2, vc2.transpose() * square2); + + vc2 = square2.block(0,0,cols,1); + VERIFY_IS_APPROX(square2.block(0,0,cols,1).transpose() * square2, vc2.transpose() * square2); + VERIFY_IS_APPROX(w2.noalias() = square2.block(0,0,cols,1).transpose() * square2, vc2.transpose() * square2); + VERIFY_IS_APPROX(w2.block(0,0,1,cols).noalias() = square2.block(0,0,cols,1).transpose() * square2, vc2.transpose() * square2); + } + // inner product { Scalar x = square2.row(c) * square2.col(c2); @@ -196,4 +227,5 @@ template<typename MatrixType> void product(const MatrixType& m) VERIFY_IS_APPROX(square * (s1*(square*square)), s1 * square * square * square); VERIFY_IS_APPROX(square * (square*square).conjugate(), square * square.conjugate() * square.conjugate()); } + } diff --git a/test/product_notemporary.cpp b/test/product_notemporary.cpp index 5a3f3a01a..2bb19a681 100644 --- a/test/product_notemporary.cpp +++ b/test/product_notemporary.cpp @@ -56,6 +56,9 @@ template<typename MatrixType> void product_notemporary(const MatrixType& m) VERIFY_EVALUATION_COUNT( m3.noalias() = m3 + m1 * m2.transpose(), 0); VERIFY_EVALUATION_COUNT( m3.noalias() += m3 + m1 * m2.transpose(), 0); VERIFY_EVALUATION_COUNT( m3.noalias() -= m3 + m1 * m2.transpose(), 0); + VERIFY_EVALUATION_COUNT( m3.noalias() = m3 - m1 * m2.transpose(), 0); + VERIFY_EVALUATION_COUNT( m3.noalias() += m3 - m1 * m2.transpose(), 0); + VERIFY_EVALUATION_COUNT( m3.noalias() -= m3 - m1 * m2.transpose(), 0); VERIFY_EVALUATION_COUNT( m3.noalias() = s1 * m1 * s2 * m2.adjoint(), 0); VERIFY_EVALUATION_COUNT( m3.noalias() = s1 * m1 * s2 * (m1*s3+m2*s2).adjoint(), 1); diff --git a/test/rand.cpp b/test/rand.cpp index eeec34191..51cf01773 100644 --- a/test/rand.cpp +++ b/test/rand.cpp @@ -9,6 +9,8 @@ #include "main.h" +typedef long long int64; + template<typename Scalar> Scalar check_in_range(Scalar x, Scalar y) { Scalar r = internal::random<Scalar>(x,y); @@ -35,31 +37,49 @@ template<typename Scalar> void check_all_in_range(Scalar x, Scalar y) VERIFY( (mask>0).all() ); } +template<typename Scalar> void check_histogram(Scalar x, Scalar y, int bins) +{ + Array<int,1,Dynamic> hist(bins); + hist.fill(0); + int f = 100000; + int n = bins*f; + int64 range = int64(y)-int64(x); + int divisor = int((range+1)/bins); + assert(((range+1)%bins)==0); + for(int k=0; k<n; ++k) + { + Scalar r = check_in_range(x,y); + hist( int((int64(r)-int64(x))/divisor) )++; + } + VERIFY( (((hist.cast<double>()/double(f))-1.0).abs()<0.02).all() ); +} + void test_rand() { long long_ref = NumTraits<long>::highest()/10; signed char char_offset = (std::min)(g_repeat,64); signed char short_offset = (std::min)(g_repeat,16000); - - for(int i = 0; i < g_repeat*10; i++) { + + for(int i = 0; i < g_repeat*10000; i++) { CALL_SUBTEST(check_in_range<float>(10,11)); CALL_SUBTEST(check_in_range<float>(1.24234523,1.24234523)); CALL_SUBTEST(check_in_range<float>(-1,1)); CALL_SUBTEST(check_in_range<float>(-1432.2352,-1432.2352)); - + CALL_SUBTEST(check_in_range<double>(10,11)); CALL_SUBTEST(check_in_range<double>(1.24234523,1.24234523)); CALL_SUBTEST(check_in_range<double>(-1,1)); CALL_SUBTEST(check_in_range<double>(-1432.2352,-1432.2352)); - + CALL_SUBTEST(check_in_range<int>(0,-1)); CALL_SUBTEST(check_in_range<short>(0,-1)); CALL_SUBTEST(check_in_range<long>(0,-1)); CALL_SUBTEST(check_in_range<int>(-673456,673456)); + CALL_SUBTEST(check_in_range<int>(-RAND_MAX+10,RAND_MAX-10)); CALL_SUBTEST(check_in_range<short>(-24345,24345)); CALL_SUBTEST(check_in_range<long>(-long_ref,long_ref)); } - + CALL_SUBTEST(check_all_in_range<signed char>(11,11)); CALL_SUBTEST(check_all_in_range<signed char>(11,11+char_offset)); CALL_SUBTEST(check_all_in_range<signed char>(-5,5)); @@ -67,25 +87,32 @@ void test_rand() CALL_SUBTEST(check_all_in_range<signed char>(-126,-126+char_offset)); CALL_SUBTEST(check_all_in_range<signed char>(126-char_offset,126)); CALL_SUBTEST(check_all_in_range<signed char>(-126,126)); - + CALL_SUBTEST(check_all_in_range<short>(11,11)); CALL_SUBTEST(check_all_in_range<short>(11,11+short_offset)); CALL_SUBTEST(check_all_in_range<short>(-5,5)); CALL_SUBTEST(check_all_in_range<short>(-11-short_offset,-11)); CALL_SUBTEST(check_all_in_range<short>(-24345,-24345+short_offset)); CALL_SUBTEST(check_all_in_range<short>(24345,24345+short_offset)); - + CALL_SUBTEST(check_all_in_range<int>(11,11)); CALL_SUBTEST(check_all_in_range<int>(11,11+g_repeat)); CALL_SUBTEST(check_all_in_range<int>(-5,5)); CALL_SUBTEST(check_all_in_range<int>(-11-g_repeat,-11)); CALL_SUBTEST(check_all_in_range<int>(-673456,-673456+g_repeat)); CALL_SUBTEST(check_all_in_range<int>(673456,673456+g_repeat)); - + CALL_SUBTEST(check_all_in_range<long>(11,11)); CALL_SUBTEST(check_all_in_range<long>(11,11+g_repeat)); CALL_SUBTEST(check_all_in_range<long>(-5,5)); CALL_SUBTEST(check_all_in_range<long>(-11-g_repeat,-11)); CALL_SUBTEST(check_all_in_range<long>(-long_ref,-long_ref+g_repeat)); CALL_SUBTEST(check_all_in_range<long>( long_ref, long_ref+g_repeat)); + + CALL_SUBTEST(check_histogram<int>(-5,5,11)); + int bins = 100; + CALL_SUBTEST(check_histogram<int>(-3333,-3333+bins*(3333/bins)-1,bins)); + bins = 1000; + CALL_SUBTEST(check_histogram<int>(-RAND_MAX+10,-RAND_MAX+10+bins*(RAND_MAX/bins)-1,bins)); + CALL_SUBTEST(check_histogram<int>(-RAND_MAX+10,-int64(RAND_MAX)+10+bins*(2*int64(RAND_MAX)/bins)-1,bins)); } diff --git a/test/real_qz.cpp b/test/real_qz.cpp index 45ae8d763..99ac31235 100644 --- a/test/real_qz.cpp +++ b/test/real_qz.cpp @@ -7,6 +7,7 @@ // 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/. +#define EIGEN_RUNTIME_NO_MALLOC #include "main.h" #include <limits> #include <Eigen/Eigenvalues> @@ -41,7 +42,11 @@ template<typename MatrixType> void real_qz(const MatrixType& m) break; } - RealQZ<MatrixType> qz(A,B); + RealQZ<MatrixType> qz(dim); + // TODO enable full-prealocation of required memory, this probably requires an in-place mode for HessenbergDecomposition + //Eigen::internal::set_is_malloc_allowed(false); + qz.compute(A,B); + //Eigen::internal::set_is_malloc_allowed(true); VERIFY_IS_EQUAL(qz.info(), Success); // check for zeros diff --git a/test/sparse_basic.cpp b/test/sparse_basic.cpp index 10309b3a9..7b5f3eb38 100644 --- a/test/sparse_basic.cpp +++ b/test/sparse_basic.cpp @@ -207,6 +207,16 @@ template<typename SparseMatrixType> void sparse_basic(const SparseMatrixType& re VERIFY_IS_APPROX((m1 = m1.transpose()), (refM1 = refM1.transpose().eval())); VERIFY_IS_APPROX((m1 = -m1.transpose()), (refM1 = -refM1.transpose().eval())); VERIFY_IS_APPROX((m1 += -m1), (refM1 += -refM1)); + + if(m1.isCompressed()) + { + VERIFY_IS_APPROX(m1.coeffs().sum(), m1.sum()); + m1.coeffs() += s1; + for(Index j = 0; j<m1.outerSize(); ++j) + for(typename SparseMatrixType::InnerIterator it(m1,j); it; ++it) + refM1(it.row(), it.col()) += s1; + VERIFY_IS_APPROX(m1, refM1); + } } // test transpose diff --git a/test/sparse_product.cpp b/test/sparse_product.cpp index c518a6e55..c7c93373d 100644 --- a/test/sparse_product.cpp +++ b/test/sparse_product.cpp @@ -292,6 +292,10 @@ template<typename SparseMatrixType> void sparse_product() VERIFY_IS_APPROX(x=mUp.template selfadjointView<Upper>()*b, refX=refS*b); VERIFY_IS_APPROX(x=mLo.template selfadjointView<Lower>()*b, refX=refS*b); VERIFY_IS_APPROX(x=mS.template selfadjointView<Upper|Lower>()*b, refX=refS*b); + + VERIFY_IS_APPROX(x.noalias()+=mUp.template selfadjointView<Upper>()*b, refX+=refS*b); + VERIFY_IS_APPROX(x.noalias()-=mLo.template selfadjointView<Lower>()*b, refX-=refS*b); + VERIFY_IS_APPROX(x.noalias()+=mS.template selfadjointView<Upper|Lower>()*b, refX+=refS*b); // sparse selfadjointView with sparse matrices SparseMatrixType mSres(rows,rows); diff --git a/unsupported/Eigen/CMakeLists.txt b/unsupported/Eigen/CMakeLists.txt index 7478b6b0d..631a06014 100644 --- a/unsupported/Eigen/CMakeLists.txt +++ b/unsupported/Eigen/CMakeLists.txt @@ -4,6 +4,7 @@ set(Eigen_HEADERS ArpackSupport AutoDiff BVH + EulerAngles FFT IterativeSolvers KroneckerProduct @@ -26,5 +27,6 @@ install(FILES DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen COMPONENT Devel ) -add_subdirectory(src) -add_subdirectory(CXX11)
\ No newline at end of file +install(DIRECTORY src DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen COMPONENT Devel FILES_MATCHING PATTERN "*.h") + +add_subdirectory(CXX11) diff --git a/unsupported/Eigen/CXX11/CMakeLists.txt b/unsupported/Eigen/CXX11/CMakeLists.txt index a40bc4715..385ed240c 100644 --- a/unsupported/Eigen/CXX11/CMakeLists.txt +++ b/unsupported/Eigen/CXX11/CMakeLists.txt @@ -5,4 +5,4 @@ install(FILES DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/CXX11 COMPONENT Devel ) -add_subdirectory(src) +install(DIRECTORY src DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/CXX11 COMPONENT Devel FILES_MATCHING PATTERN "*.h") diff --git a/unsupported/Eigen/CXX11/src/CMakeLists.txt b/unsupported/Eigen/CXX11/src/CMakeLists.txt deleted file mode 100644 index 1734262bb..000000000 --- a/unsupported/Eigen/CXX11/src/CMakeLists.txt +++ /dev/null @@ -1,4 +0,0 @@ -add_subdirectory(util) -add_subdirectory(ThreadPool) -add_subdirectory(Tensor) -add_subdirectory(TensorSymmetry) diff --git a/unsupported/Eigen/CXX11/src/Tensor/CMakeLists.txt b/unsupported/Eigen/CXX11/src/Tensor/CMakeLists.txt deleted file mode 100644 index 6d4b3ea0d..000000000 --- a/unsupported/Eigen/CXX11/src/Tensor/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_CXX11_Tensor_SRCS "*.h") - -INSTALL(FILES - ${Eigen_CXX11_Tensor_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/CXX11/src/Tensor COMPONENT Devel - ) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h index e3880d2e0..3c8710255 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h @@ -192,6 +192,12 @@ class TensorBase<Derived, ReadOnlyAccessors> } EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_log1p_op<Scalar>, const Derived> + log1p() const { + return unaryExpr(internal::scalar_log1p_op<Scalar>()); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_abs_op<Scalar>, const Derived> abs() const { return unaryExpr(internal::scalar_abs_op<Scalar>()); diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h index 56d9c2025..20b29e5fd 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h @@ -25,8 +25,9 @@ template<typename Dimensions, typename LhsXprType, typename RhsXprType> struct traits<TensorContractionOp<Dimensions, LhsXprType, RhsXprType> > { // Type promotion to handle the case where the types of the lhs and the rhs are different. - typedef typename internal::promote_storage_type<typename LhsXprType::Scalar, - typename RhsXprType::Scalar>::ret Scalar; + typedef typename gebp_traits<typename remove_const<typename LhsXprType::Scalar>::type, + typename remove_const<typename RhsXprType::Scalar>::type>::ResScalar Scalar; + typedef typename promote_storage_type<typename traits<LhsXprType>::StorageKind, typename traits<RhsXprType>::StorageKind>::ret StorageKind; typedef typename promote_index_type<typename traits<LhsXprType>::Index, @@ -75,8 +76,8 @@ class TensorContractionOp : public TensorBase<TensorContractionOp<Indices, LhsXp { public: typedef typename Eigen::internal::traits<TensorContractionOp>::Scalar Scalar; - typedef typename internal::promote_storage_type<typename LhsXprType::CoeffReturnType, - typename RhsXprType::CoeffReturnType>::ret CoeffReturnType; + typedef typename internal::gebp_traits<typename LhsXprType::CoeffReturnType, + typename RhsXprType::CoeffReturnType>::ResScalar CoeffReturnType; typedef typename Eigen::internal::nested<TensorContractionOp>::type Nested; typedef typename Eigen::internal::traits<TensorContractionOp>::StorageKind StorageKind; typedef typename Eigen::internal::traits<TensorContractionOp>::Index Index; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorCostModel.h b/unsupported/Eigen/CXX11/src/Tensor/TensorCostModel.h index a76c8ca35..d66e45d50 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorCostModel.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorCostModel.h @@ -91,21 +91,21 @@ class TensorOpCost { } // TODO(rmlarsen): Define min in terms of total cost, not elementwise. - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost& cwiseMin( - const TensorOpCost& rhs) { - bytes_loaded_ = numext::mini(bytes_loaded_, rhs.bytes_loaded()); - bytes_stored_ = numext::mini(bytes_stored_, rhs.bytes_stored()); - compute_cycles_ = numext::mini(compute_cycles_, rhs.compute_cycles()); - return *this; + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost cwiseMin( + const TensorOpCost& rhs) const { + double bytes_loaded = numext::mini(bytes_loaded_, rhs.bytes_loaded()); + double bytes_stored = numext::mini(bytes_stored_, rhs.bytes_stored()); + double compute_cycles = numext::mini(compute_cycles_, rhs.compute_cycles()); + return TensorOpCost(bytes_loaded, bytes_stored, compute_cycles); } // TODO(rmlarsen): Define max in terms of total cost, not elementwise. - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost& cwiseMax( - const TensorOpCost& rhs) { - bytes_loaded_ = numext::maxi(bytes_loaded_, rhs.bytes_loaded()); - bytes_stored_ = numext::maxi(bytes_stored_, rhs.bytes_stored()); - compute_cycles_ = numext::maxi(compute_cycles_, rhs.compute_cycles()); - return *this; + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost cwiseMax( + const TensorOpCost& rhs) const { + double bytes_loaded = numext::maxi(bytes_loaded_, rhs.bytes_loaded()); + double bytes_stored = numext::maxi(bytes_stored_, rhs.bytes_stored()); + double compute_cycles = numext::maxi(compute_cycles_, rhs.compute_cycles()); + return TensorOpCost(bytes_loaded, bytes_stored, compute_cycles); } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost& operator+=( diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h index b2b4bcf62..834ce07df 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h @@ -239,7 +239,7 @@ struct TensorEvaluator<const TensorCwiseNullaryOp<NullaryOp, ArgType>, Device> EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const Device& device) - : m_functor(op.functor()), m_argImpl(op.nestedExpression(), device) + : m_functor(op.functor()), m_argImpl(op.nestedExpression(), device), m_wrapper() { } typedef typename XprType::Index Index; @@ -256,13 +256,13 @@ struct TensorEvaluator<const TensorCwiseNullaryOp<NullaryOp, ArgType>, Device> EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { - return m_functor(index); + return m_wrapper(m_functor, index); } template<int LoadMode> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const { - return m_functor.template packetOp<Index, PacketReturnType>(index); + return m_wrapper.template packetOp<PacketReturnType, Index>(m_functor, index); } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost @@ -282,6 +282,7 @@ struct TensorEvaluator<const TensorCwiseNullaryOp<NullaryOp, ArgType>, Device> private: const NullaryOp m_functor; TensorEvaluator<ArgType, Device> m_argImpl; + const internal::nullary_wrapper<CoeffReturnType,NullaryOp> m_wrapper; }; @@ -612,7 +613,7 @@ struct TensorEvaluator<const TensorSelectOp<IfArgType, ThenArgType, ElseArgType> .cwiseMax(m_elseImpl.costPerCoeff(vectorized)); } - EIGEN_DEVICE_FUNC CoeffReturnType* data() const { return NULL; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType* data() const { return NULL; } /// required by sycl in order to extract the accessor const TensorEvaluator<IfArgType, Device> & cond_impl() const { return m_condImpl; } /// required by sycl in order to extract the accessor diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h index a8e48fced..fc75dbb5c 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h @@ -25,7 +25,7 @@ struct scalar_mod_op { }; template <typename Scalar> struct functor_traits<scalar_mod_op<Scalar> > -{ enum { Cost = NumTraits<Scalar>::template Div<false>::Cost, PacketAccess = false }; }; +{ enum { Cost = scalar_div_cost<Scalar,false>::value, PacketAccess = false }; }; /** \internal @@ -38,7 +38,7 @@ struct scalar_mod2_op { }; template <typename Scalar> struct functor_traits<scalar_mod2_op<Scalar> > -{ enum { Cost = NumTraits<Scalar>::template Div<false>::Cost, PacketAccess = false }; }; +{ enum { Cost = scalar_div_cost<Scalar,false>::value, PacketAccess = false }; }; template <typename Scalar> struct scalar_fmod_op { @@ -188,6 +188,32 @@ struct reducer_traits<MeanReducer<T>, Device> { }; +template <typename T, bool IsMax = true, bool IsInteger = true> +struct MinMaxBottomValue { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE static T bottom_value() { + return Eigen::NumTraits<T>::lowest(); + } +}; +template <typename T> +struct MinMaxBottomValue<T, true, false> { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE static T bottom_value() { + return -Eigen::NumTraits<T>::infinity(); + } +}; +template <typename T> +struct MinMaxBottomValue<T, false, true> { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE static T bottom_value() { + return Eigen::NumTraits<T>::highest(); + } +}; +template <typename T> +struct MinMaxBottomValue<T, false, false> { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE static T bottom_value() { + return Eigen::NumTraits<T>::infinity(); + } +}; + + template <typename T> struct MaxReducer { static const bool PacketAccess = packet_traits<T>::HasMax; @@ -200,9 +226,8 @@ template <typename T> struct MaxReducer EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reducePacket(const Packet& p, Packet* accum) const { (*accum) = pmax<Packet>(*accum, p); } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T initialize() const { - return Eigen::NumTraits<T>::lowest(); + return MinMaxBottomValue<T, true, Eigen::NumTraits<T>::IsInteger>::bottom_value(); } template <typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet initializePacket() const { @@ -242,9 +267,8 @@ template <typename T> struct MinReducer EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reducePacket(const Packet& p, Packet* accum) const { (*accum) = pmin<Packet>(*accum, p); } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T initialize() const { - return Eigen::NumTraits<T>::highest(); + return MinMaxBottomValue<T, false, Eigen::NumTraits<T>::IsInteger>::bottom_value(); } template <typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet initializePacket() const { @@ -454,12 +478,11 @@ template <typename T> class UniformRandomGenerator { m_deterministic = other.m_deterministic; } - template<typename Index> - T operator()(Index) const { + T operator()() const { return random<T>(); } - template<typename Index, typename PacketType> - PacketType packetOp(Index) const { + template<typename PacketType> + PacketType packetOp() const { const int packetSize = internal::unpacket_traits<PacketType>::size; EIGEN_ALIGN_MAX T values[packetSize]; for (int i = 0; i < packetSize; ++i) { @@ -484,23 +507,22 @@ template <> class UniformRandomGenerator<float> { } UniformRandomGenerator(const UniformRandomGenerator<float>& other) { m_generator = new std::mt19937(); - m_generator->seed(other(0) * UINT_MAX); + m_generator->seed(other() * UINT_MAX); m_deterministic = other.m_deterministic; } ~UniformRandomGenerator() { delete m_generator; } - template<typename Index> - float operator()(Index) const { + float operator()() const { return m_distribution(*m_generator); } - template<typename Index, typename PacketType> - PacketType packetOp(Index i) const { + template<typename PacketType> + PacketType packetOp() const { const int packetSize = internal::unpacket_traits<PacketType>::size; EIGEN_ALIGN_MAX float values[packetSize]; for (int k = 0; k < packetSize; ++k) { - values[k] = this->operator()(i); + values[k] = this->operator()(); } return internal::pload<PacketType>(values); } @@ -525,23 +547,22 @@ template <> class UniformRandomGenerator<double> { } UniformRandomGenerator(const UniformRandomGenerator<double>& other) { m_generator = new std::mt19937(); - m_generator->seed(other(0) * UINT_MAX); + m_generator->seed(other() * UINT_MAX); m_deterministic = other.m_deterministic; } ~UniformRandomGenerator() { delete m_generator; } - template<typename Index> - double operator()(Index) const { + double operator()() const { return m_distribution(*m_generator); } - template<typename Index, typename PacketType> - PacketType packetOp(Index i) const { + template<typename PacketType> + PacketType packetOp() const { const int packetSize = internal::unpacket_traits<PacketType>::size; EIGEN_ALIGN_MAX double values[packetSize]; for (int k = 0; k < packetSize; ++k) { - values[k] = this->operator()(i); + values[k] = this->operator()(); } return internal::pload<PacketType>(values); } @@ -578,12 +599,11 @@ template <> class UniformRandomGenerator<float> { curand_init(seed, tid, 0, &m_state); } - template<typename Index> - __device__ float operator()(Index) const { + __device__ float operator()() const { return curand_uniform(&m_state); } - template<typename Index, typename PacketType> - __device__ float4 packetOp(Index) const { + template<typename PacketType> + __device__ float4 packetOp() const { EIGEN_STATIC_ASSERT((is_same<PacketType, float4>::value), YOU_MADE_A_PROGRAMMING_MISTAKE); return curand_uniform4(&m_state); } @@ -608,12 +628,11 @@ template <> class UniformRandomGenerator<double> { const int seed = m_deterministic ? 0 : get_random_seed(); curand_init(seed, tid, 0, &m_state); } - template<typename Index> - __device__ double operator()(Index) const { + __device__ double operator()() const { return curand_uniform_double(&m_state); } - template<typename Index, typename PacketType> - __device__ double2 packetOp(Index) const { + template<typename PacketType> + __device__ double2 packetOp() const { EIGEN_STATIC_ASSERT((is_same<PacketType, double2>::value), YOU_MADE_A_PROGRAMMING_MISTAKE); return curand_uniform2_double(&m_state); } @@ -638,8 +657,7 @@ template <> class UniformRandomGenerator<std::complex<float> > { const int seed = m_deterministic ? 0 : get_random_seed(); curand_init(seed, tid, 0, &m_state); } - template<typename Index> - __device__ std::complex<float> operator()(Index) const { + __device__ std::complex<float> operator()() const { float4 vals = curand_uniform4(&m_state); return std::complex<float>(vals.x, vals.y); } @@ -664,8 +682,7 @@ template <> class UniformRandomGenerator<std::complex<double> > { const int seed = m_deterministic ? 0 : get_random_seed(); curand_init(seed, tid, 0, &m_state); } - template<typename Index> - __device__ std::complex<double> operator()(Index) const { + __device__ std::complex<double> operator()() const { double2 vals = curand_uniform2_double(&m_state); return std::complex<double>(vals.x, vals.y); } @@ -701,17 +718,16 @@ template <typename T> class NormalRandomGenerator { } NormalRandomGenerator(const NormalRandomGenerator& other) : m_deterministic(other.m_deterministic), m_distribution(other.m_distribution), m_generator(new std::mt19937()) { - m_generator->seed(other(0) * UINT_MAX); + m_generator->seed(other() * UINT_MAX); } ~NormalRandomGenerator() { delete m_generator; } - template<typename Index> - T operator()(Index) const { + T operator()() const { return m_distribution(*m_generator); } - template<typename Index, typename PacketType> - PacketType packetOp(Index) const { + template<typename PacketType> + PacketType packetOp() const { const int packetSize = internal::unpacket_traits<PacketType>::size; EIGEN_ALIGN_MAX T values[packetSize]; for (int i = 0; i < packetSize; ++i) { @@ -749,12 +765,11 @@ template <> class NormalRandomGenerator<float> { const int seed = m_deterministic ? 0 : get_random_seed(); curand_init(seed, tid, 0, &m_state); } - template<typename Index> - __device__ float operator()(Index) const { + __device__ float operator()() const { return curand_normal(&m_state); } - template<typename Index, typename PacketType> - __device__ float4 packetOp(Index) const { + template<typename PacketType> + __device__ float4 packetOp() const { EIGEN_STATIC_ASSERT((is_same<PacketType, float4>::value), YOU_MADE_A_PROGRAMMING_MISTAKE); return curand_normal4(&m_state); } @@ -779,12 +794,11 @@ template <> class NormalRandomGenerator<double> { const int seed = m_deterministic ? 0 : get_random_seed(); curand_init(seed, tid, 0, &m_state); } - template<typename Index> - __device__ double operator()(Index) const { + __device__ double operator()() const { return curand_normal_double(&m_state); } - template<typename Index, typename PacketType> - __device__ double2 packetOp(Index) const { + template<typename PacketType> + __device__ double2 packetOp() const { EIGEN_STATIC_ASSERT((is_same<PacketType, double2>::value), YOU_MADE_A_PROGRAMMING_MISTAKE); return curand_normal2_double(&m_state); } @@ -809,8 +823,7 @@ template <> class NormalRandomGenerator<std::complex<float> > { const int seed = m_deterministic ? 0 : get_random_seed(); curand_init(seed, tid, 0, &m_state); } - template<typename Index> - __device__ std::complex<float> operator()(Index) const { + __device__ std::complex<float> operator()() const { float4 vals = curand_normal4(&m_state); return std::complex<float>(vals.x, vals.y); } @@ -835,8 +848,7 @@ template <> class NormalRandomGenerator<std::complex<double> > { const int seed = m_deterministic ? 0 : get_random_seed(); curand_init(seed, tid, 0, &m_state); } - template<typename Index> - __device__ std::complex<double> operator()(Index) const { + __device__ std::complex<double> operator()() const { double2 vals = curand_normal2_double(&m_state); return std::complex<double>(vals.x, vals.y); } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h index 9df697e4c..a87777b22 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h @@ -505,9 +505,14 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device> (reducing_inner_dims || ReducingInnerMostDims)) { const Index num_values_to_reduce = internal::array_prod(m_reducedDims); const Index num_coeffs_to_preserve = internal::array_prod(m_dimensions); - if (!data && num_coeffs_to_preserve < 1024 && num_values_to_reduce > num_coeffs_to_preserve && num_values_to_reduce > 128) { - data = static_cast<CoeffReturnType*>(m_device.allocate(sizeof(CoeffReturnType) * num_coeffs_to_preserve)); - m_result = data; + if (!data) { + if (num_coeffs_to_preserve < 1024 && num_values_to_reduce > num_coeffs_to_preserve && num_values_to_reduce > 128) { + data = static_cast<CoeffReturnType*>(m_device.allocate(sizeof(CoeffReturnType) * num_coeffs_to_preserve)); + m_result = data; + } + else { + return true; + } } Op reducer(m_reducer); if (internal::InnerReducer<Self, Op, Device>::run(*this, reducer, m_device, data, num_values_to_reduce, num_coeffs_to_preserve)) { @@ -533,9 +538,14 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device> preserving_inner_dims) { const Index num_values_to_reduce = internal::array_prod(m_reducedDims); const Index num_coeffs_to_preserve = internal::array_prod(m_dimensions); - if (!data && num_coeffs_to_preserve < 1024 && num_values_to_reduce > num_coeffs_to_preserve && num_values_to_reduce > 32) { - data = static_cast<CoeffReturnType*>(m_device.allocate(sizeof(CoeffReturnType) * num_coeffs_to_preserve)); - m_result = data; + if (!data) { + if (num_coeffs_to_preserve < 1024 && num_values_to_reduce > num_coeffs_to_preserve && num_values_to_reduce > 32) { + data = static_cast<CoeffReturnType*>(m_device.allocate(sizeof(CoeffReturnType) * num_coeffs_to_preserve)); + m_result = data; + } + else { + return true; + } } Op reducer(m_reducer); if (internal::OuterReducer<Self, Op, Device>::run(*this, reducer, m_device, data, num_values_to_reduce, num_coeffs_to_preserve)) { @@ -556,6 +566,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device> m_impl.cleanup(); if (m_result) { m_device.deallocate(m_result); + m_result = NULL; } } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h index 5e512490c..65638b6a8 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h @@ -67,11 +67,21 @@ __device__ EIGEN_ALWAYS_INLINE void atomicReduce(T* output, T accum, R& reducer) #endif } +// We extend atomicExch to support extra data types +template <typename Type> +__device__ inline Type atomicExchCustom(Type* address, Type val) { + return atomicExch(address, val); +} + +template <> +__device__ inline double atomicExchCustom(double* address, double val) { + unsigned long long int* address_as_ull = reinterpret_cast<unsigned long long int*>(address); + return __longlong_as_double(atomicExch(address_as_ull, __double_as_longlong(val))); +} #ifdef EIGEN_HAS_CUDA_FP16 template <template <typename T> class R> __device__ inline void atomicReduce(half2* output, half2 accum, R<half>& reducer) { -#if __CUDA_ARCH__ >= 300 unsigned int oldval = *reinterpret_cast<unsigned int*>(output); unsigned int newval = oldval; reducer.reducePacket(accum, reinterpret_cast<half2*>(&newval)); @@ -87,9 +97,6 @@ __device__ inline void atomicReduce(half2* output, half2 accum, R<half>& reducer return; } } -#else - assert(0 && "Shouldn't be called on unsupported device"); -#endif } #endif @@ -130,7 +137,7 @@ __global__ void FullReductionKernel(Reducer reducer, const Self input, Index num unsigned int block = atomicCAS(semaphore, 0u, 1u); if (block == 0) { // We're the first block to run, initialize the output value - atomicExch(output, reducer.initialize()); + atomicExchCustom(output, reducer.initialize()); __threadfence(); atomicExch(semaphore, 2u); } @@ -263,17 +270,22 @@ __global__ void ReductionCleanupKernelHalfFloat(Op& reducer, half* output, half2 #endif - -template <typename Self, typename Op, typename OutputType, bool PacketAccess> +template <typename Self, typename Op, typename OutputType, bool PacketAccess, typename Enabled = void> struct FullReductionLauncher { static void run(const Self&, Op&, const GpuDevice&, OutputType*, typename Self::Index) { - assert(false && "Should only be called on floats and half floats"); + assert(false && "Should only be called on doubles, floats and half floats"); } }; -template <typename Self, typename Op, bool PacketAccess> -struct FullReductionLauncher<Self, Op, float, PacketAccess> { - static void run(const Self& self, Op& reducer, const GpuDevice& device, float* output, typename Self::Index num_coeffs) { +// Specialization for float and double +template <typename Self, typename Op, typename OutputType, bool PacketAccess> +struct FullReductionLauncher< + Self, Op, OutputType, PacketAccess, + typename internal::enable_if< + internal::is_same<float, OutputType>::value || + internal::is_same<double, OutputType>::value, + void>::type> { + static void run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output, typename Self::Index num_coeffs) { typedef typename Self::Index Index; typedef typename Self::CoeffReturnType Scalar; const int block_size = 256; @@ -330,20 +342,22 @@ struct FullReductionLauncher<Self, Op, Eigen::half, true> { template <typename Self, typename Op, bool Vectorizable> struct FullReducer<Self, Op, GpuDevice, Vectorizable> { // Unfortunately nvidia doesn't support well exotic types such as complex, - // so reduce the scope of the optimized version of the code to the simple case - // of floats and half floats. + // so reduce the scope of the optimized version of the code to the simple cases + // of doubles, floats and half floats #ifdef EIGEN_HAS_CUDA_FP16 static const bool HasOptimizedImplementation = !Op::IsStateful && (internal::is_same<typename Self::CoeffReturnType, float>::value || + internal::is_same<typename Self::CoeffReturnType, double>::value || (internal::is_same<typename Self::CoeffReturnType, Eigen::half>::value && reducer_traits<Op, GpuDevice>::PacketAccess)); #else static const bool HasOptimizedImplementation = !Op::IsStateful && - internal::is_same<typename Self::CoeffReturnType, float>::value; + (internal::is_same<typename Self::CoeffReturnType, float>::value || + internal::is_same<typename Self::CoeffReturnType, double>::value); #endif template <typename OutputType> static void run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output) { - assert(HasOptimizedImplementation && "Should only be called on floats or half floats"); + assert(HasOptimizedImplementation && "Should only be called on doubles, floats or half floats"); const Index num_coeffs = array_prod(self.m_impl.dimensions()); // Don't crash when we're called with an input tensor of size 0. if (num_coeffs == 0) { @@ -360,6 +374,7 @@ template <int NumPerThread, typename Self, __global__ void InnerReductionKernel(Reducer reducer, const Self input, Index num_coeffs_to_reduce, Index num_preserved_coeffs, typename Self::CoeffReturnType* output) { #if __CUDA_ARCH__ >= 300 + typedef typename Self::CoeffReturnType Type; eigen_assert(blockDim.y == 1); eigen_assert(blockDim.z == 1); eigen_assert(gridDim.y == 1); @@ -389,13 +404,13 @@ __global__ void InnerReductionKernel(Reducer reducer, const Self input, Index nu const Index col_block = i % input_col_blocks; const Index col_begin = col_block * blockDim.x * NumPerThread + threadIdx.x; - float reduced_val = reducer.initialize(); + Type reduced_val = reducer.initialize(); for (Index j = 0; j < NumPerThread; j += unroll_times) { const Index last_col = col_begin + blockDim.x * (j + unroll_times - 1); if (last_col >= num_coeffs_to_reduce) { for (Index col = col_begin + blockDim.x * j; col < num_coeffs_to_reduce; col += blockDim.x) { - const float val = input.m_impl.coeff(row * num_coeffs_to_reduce + col); + const Type val = input.m_impl.coeff(row * num_coeffs_to_reduce + col); reducer.reduce(val, &reduced_val); } break; @@ -521,17 +536,23 @@ __global__ void InnerReductionKernelHalfFloat(Reducer reducer, const Self input, #endif -template <typename Self, typename Op, typename OutputType, bool PacketAccess> +template <typename Self, typename Op, typename OutputType, bool PacketAccess, typename Enabled = void> struct InnerReductionLauncher { static EIGEN_DEVICE_FUNC bool run(const Self&, Op&, const GpuDevice&, OutputType*, typename Self::Index, typename Self::Index) { - assert(false && "Should only be called to reduce floats and half floats on a gpu device"); + assert(false && "Should only be called to reduce doubles, floats and half floats on a gpu device"); return true; } }; -template <typename Self, typename Op, bool PacketAccess> -struct InnerReductionLauncher<Self, Op, float, PacketAccess> { - static bool run(const Self& self, Op& reducer, const GpuDevice& device, float* output, typename Self::Index num_coeffs_to_reduce, typename Self::Index num_preserved_vals) { +// Specialization for float and double +template <typename Self, typename Op, typename OutputType, bool PacketAccess> +struct InnerReductionLauncher< + Self, Op, OutputType, PacketAccess, + typename internal::enable_if< + internal::is_same<float, OutputType>::value || + internal::is_same<double, OutputType>::value, + void>::type> { + static bool run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output, typename Self::Index num_coeffs_to_reduce, typename Self::Index num_preserved_vals) { typedef typename Self::Index Index; const Index num_coeffs = num_coeffs_to_reduce * num_preserved_vals; @@ -549,7 +570,7 @@ struct InnerReductionLauncher<Self, Op, float, PacketAccess> { const int max_blocks = device.getNumCudaMultiProcessors() * device.maxCudaThreadsPerMultiProcessor() / 1024; const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks); - LAUNCH_CUDA_KERNEL((ReductionInitKernel<float, Index>), + LAUNCH_CUDA_KERNEL((ReductionInitKernel<OutputType, Index>), num_blocks, 1024, 0, device, reducer.initialize(), num_preserved_vals, output); } @@ -616,15 +637,17 @@ struct InnerReducer<Self, Op, GpuDevice> { #ifdef EIGEN_HAS_CUDA_FP16 static const bool HasOptimizedImplementation = !Op::IsStateful && (internal::is_same<typename Self::CoeffReturnType, float>::value || + internal::is_same<typename Self::CoeffReturnType, double>::value || (internal::is_same<typename Self::CoeffReturnType, Eigen::half>::value && reducer_traits<Op, GpuDevice>::PacketAccess)); #else static const bool HasOptimizedImplementation = !Op::IsStateful && - internal::is_same<typename Self::CoeffReturnType, float>::value; + (internal::is_same<typename Self::CoeffReturnType, float>::value || + internal::is_same<typename Self::CoeffReturnType, double>::value); #endif template <typename OutputType> static bool run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output, typename Self::Index num_coeffs_to_reduce, typename Self::Index num_preserved_vals) { - assert(HasOptimizedImplementation && "Should only be called on floats or half floats"); + assert(HasOptimizedImplementation && "Should only be called on doubles, floats or half floats"); const Index num_coeffs = array_prod(self.m_impl.dimensions()); // Don't crash when we're called with an input tensor of size 0. if (num_coeffs == 0) { @@ -675,11 +698,11 @@ struct OuterReducer<Self, Op, GpuDevice> { // so reduce the scope of the optimized version of the code to the simple case // of floats. static const bool HasOptimizedImplementation = !Op::IsStateful && - internal::is_same<typename Self::CoeffReturnType, float>::value; - + (internal::is_same<typename Self::CoeffReturnType, float>::value || + internal::is_same<typename Self::CoeffReturnType, double>::value); template <typename Device, typename OutputType> static EIGEN_DEVICE_FUNC bool run(const Self&, Op&, const Device&, OutputType*, typename Self::Index, typename Self::Index) { - assert(false && "Should only be called to reduce floats on a gpu device"); + assert(false && "Should only be called to reduce doubles or floats on a gpu device"); return true; } diff --git a/unsupported/Eigen/CXX11/src/TensorSymmetry/CMakeLists.txt b/unsupported/Eigen/CXX11/src/TensorSymmetry/CMakeLists.txt deleted file mode 100644 index 6e871a8da..000000000 --- a/unsupported/Eigen/CXX11/src/TensorSymmetry/CMakeLists.txt +++ /dev/null @@ -1,8 +0,0 @@ -FILE(GLOB Eigen_CXX11_TensorSymmetry_SRCS "*.h") - -INSTALL(FILES - ${Eigen_CXX11_TensorSymmetry_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/CXX11/src/TensorSymmetry COMPONENT Devel - ) - -add_subdirectory(util) diff --git a/unsupported/Eigen/CXX11/src/TensorSymmetry/util/CMakeLists.txt b/unsupported/Eigen/CXX11/src/TensorSymmetry/util/CMakeLists.txt deleted file mode 100644 index dc9fc78ec..000000000 --- a/unsupported/Eigen/CXX11/src/TensorSymmetry/util/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_CXX11_TensorSymmetry_util_SRCS "*.h") - -INSTALL(FILES - ${Eigen_CXX11_TensorSymmetry_util_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/CXX11/src/TensorSymmetry/util COMPONENT Devel - ) diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/CMakeLists.txt b/unsupported/Eigen/CXX11/src/ThreadPool/CMakeLists.txt deleted file mode 100644 index 88fef50c6..000000000 --- a/unsupported/Eigen/CXX11/src/ThreadPool/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_CXX11_ThreadPool_SRCS "*.h") - -INSTALL(FILES - ${Eigen_CXX11_ThreadPool_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/CXX11/src/ThreadPool COMPONENT Devel - ) diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h b/unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h index 12b80d6c4..71d55552d 100644 --- a/unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h +++ b/unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h @@ -50,7 +50,7 @@ class EventCount { public: class Waiter; - EventCount(std::vector<Waiter>& waiters) : waiters_(waiters) { + EventCount(MaxSizeVector<Waiter>& waiters) : waiters_(waiters) { eigen_assert(waiters.size() < (1 << kWaiterBits) - 1); // Initialize epoch to something close to overflow to test overflow. state_ = kStackMask | (kEpochMask - kEpochInc * waiters.size() * 2); @@ -199,7 +199,7 @@ class EventCount { static const uint64_t kEpochMask = ((1ull << kEpochBits) - 1) << kEpochShift; static const uint64_t kEpochInc = 1ull << kEpochShift; std::atomic<uint64_t> state_; - std::vector<Waiter>& waiters_; + MaxSizeVector<Waiter>& waiters_; void Park(Waiter* w) { std::unique_lock<std::mutex> lock(w->mu); diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h index 33ae45131..354bce52a 100644 --- a/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h +++ b/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h @@ -29,6 +29,8 @@ class NonBlockingThreadPoolTempl : public Eigen::ThreadPoolInterface { spinning_(0), done_(false), ec_(waiters_) { + waiters_.resize(num_threads); + // Calculate coprimes of num_threads. // Coprimes are used for a random walk over all threads in Steal // and NonEmptyQueueIndex. Iteration is based on the fact that if we take @@ -123,7 +125,7 @@ class NonBlockingThreadPoolTempl : public Eigen::ThreadPoolInterface { MaxSizeVector<Thread*> threads_; MaxSizeVector<Queue*> queues_; MaxSizeVector<unsigned> coprimes_; - std::vector<EventCount::Waiter> waiters_; + MaxSizeVector<EventCount::Waiter> waiters_; std::atomic<unsigned> blocked_; std::atomic<bool> spinning_; std::atomic<bool> done_; diff --git a/unsupported/Eigen/CXX11/src/util/CMakeLists.txt b/unsupported/Eigen/CXX11/src/util/CMakeLists.txt deleted file mode 100644 index 7eab492d6..000000000 --- a/unsupported/Eigen/CXX11/src/util/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_CXX11_util_SRCS "*.h") - -INSTALL(FILES - ${Eigen_CXX11_util_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/CXX11/src/util COMPONENT Devel - ) diff --git a/unsupported/Eigen/CXX11/src/util/MaxSizeVector.h b/unsupported/Eigen/CXX11/src/util/MaxSizeVector.h index 961456f10..4bc3dd1ba 100644 --- a/unsupported/Eigen/CXX11/src/util/MaxSizeVector.h +++ b/unsupported/Eigen/CXX11/src/util/MaxSizeVector.h @@ -55,6 +55,17 @@ class MaxSizeVector { internal::aligned_free(data_); } + void resize(size_t n) { + eigen_assert(n <= reserve_); + for (size_t i = size_; i < n; ++i) { + new (&data_[i]) T; + } + for (size_t i = n; i < size_; ++i) { + data_[i].~T(); + } + size_ = n; + } + // Append new elements (up to reserved size). EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void push_back(const T& t) { diff --git a/unsupported/Eigen/EulerAngles b/unsupported/Eigen/EulerAngles new file mode 100644 index 000000000..521fa3f76 --- /dev/null +++ b/unsupported/Eigen/EulerAngles @@ -0,0 +1,43 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2015 Tal Hadad <tal_hd@hotmail.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_EULERANGLES_MODULE_H +#define EIGEN_EULERANGLES_MODULE_H + + +#include "Eigen/Core" +#include "Eigen/Geometry" + +#include "Eigen/src/Core/util/DisableStupidWarnings.h" + +namespace Eigen { + +/** + * \defgroup EulerAngles_Module EulerAngles module + * \brief This module provides generic euler angles rotation. + * + * Euler angles are a way to represent 3D rotation. + * + * In order to use this module in your code, include this header: + * \code + * #include <unsupported/Eigen/EulerAngles> + * \endcode + * + * See \ref EulerAngles for more information. + * + */ + +} + +#include "src/EulerAngles/EulerSystem.h" +#include "src/EulerAngles/EulerAngles.h" + +#include "Eigen/src/Core/util/ReenableStupidWarnings.h" + +#endif // EIGEN_EULERANGLES_MODULE_H diff --git a/unsupported/Eigen/KroneckerProduct b/unsupported/Eigen/KroneckerProduct index c932c06a6..5f5afb8cf 100644 --- a/unsupported/Eigen/KroneckerProduct +++ b/unsupported/Eigen/KroneckerProduct @@ -13,6 +13,8 @@ #include "../../Eigen/src/Core/util/DisableStupidWarnings.h" +#include "../../Eigen/src/SparseCore/SparseUtil.h" + namespace Eigen { /** diff --git a/unsupported/Eigen/src/AutoDiff/CMakeLists.txt b/unsupported/Eigen/src/AutoDiff/CMakeLists.txt deleted file mode 100644 index ad91fd9c4..000000000 --- a/unsupported/Eigen/src/AutoDiff/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_AutoDiff_SRCS "*.h") - -INSTALL(FILES - ${Eigen_AutoDiff_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/src/AutoDiff COMPONENT Devel - ) diff --git a/unsupported/Eigen/src/BVH/CMakeLists.txt b/unsupported/Eigen/src/BVH/CMakeLists.txt deleted file mode 100644 index b377d865c..000000000 --- a/unsupported/Eigen/src/BVH/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_BVH_SRCS "*.h") - -INSTALL(FILES - ${Eigen_BVH_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/src/BVH COMPONENT Devel - ) diff --git a/unsupported/Eigen/src/CMakeLists.txt b/unsupported/Eigen/src/CMakeLists.txt deleted file mode 100644 index f42946793..000000000 --- a/unsupported/Eigen/src/CMakeLists.txt +++ /dev/null @@ -1,16 +0,0 @@ -ADD_SUBDIRECTORY(AutoDiff) -ADD_SUBDIRECTORY(BVH) -ADD_SUBDIRECTORY(Eigenvalues) -ADD_SUBDIRECTORY(FFT) -ADD_SUBDIRECTORY(IterativeSolvers) -ADD_SUBDIRECTORY(LevenbergMarquardt) -ADD_SUBDIRECTORY(MatrixFunctions) -ADD_SUBDIRECTORY(MoreVectorization) -ADD_SUBDIRECTORY(NonLinearOptimization) -ADD_SUBDIRECTORY(NumericalDiff) -ADD_SUBDIRECTORY(Polynomials) -ADD_SUBDIRECTORY(Skyline) -ADD_SUBDIRECTORY(SparseExtra) -ADD_SUBDIRECTORY(SpecialFunctions) -ADD_SUBDIRECTORY(KroneckerProduct) -ADD_SUBDIRECTORY(Splines) diff --git a/unsupported/Eigen/src/Eigenvalues/ArpackSelfAdjointEigenSolver.h b/unsupported/Eigen/src/Eigenvalues/ArpackSelfAdjointEigenSolver.h index 3b6a69aff..866a8a460 100644 --- a/unsupported/Eigen/src/Eigenvalues/ArpackSelfAdjointEigenSolver.h +++ b/unsupported/Eigen/src/Eigenvalues/ArpackSelfAdjointEigenSolver.h @@ -628,15 +628,15 @@ ArpackGeneralizedSelfAdjointEigenSolver<MatrixType, MatrixSolver, BisSPD>& m_info = Success; } - delete select; + delete[] select; } - delete v; - delete iparam; - delete ipntr; - delete workd; - delete workl; - delete resid; + delete[] v; + delete[] iparam; + delete[] ipntr; + delete[] workd; + delete[] workl; + delete[] resid; m_isInitialized = true; diff --git a/unsupported/Eigen/src/Eigenvalues/CMakeLists.txt b/unsupported/Eigen/src/Eigenvalues/CMakeLists.txt deleted file mode 100644 index 1d4387c82..000000000 --- a/unsupported/Eigen/src/Eigenvalues/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_Eigenvalues_SRCS "*.h") - -INSTALL(FILES - ${Eigen_Eigenvalues_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/src/Eigenvalues COMPONENT Devel - ) diff --git a/unsupported/Eigen/src/EulerAngles/CMakeLists.txt b/unsupported/Eigen/src/EulerAngles/CMakeLists.txt new file mode 100644 index 000000000..40af550e8 --- /dev/null +++ b/unsupported/Eigen/src/EulerAngles/CMakeLists.txt @@ -0,0 +1,6 @@ +FILE(GLOB Eigen_EulerAngles_SRCS "*.h") + +INSTALL(FILES + ${Eigen_EulerAngles_SRCS} + DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/src/EulerAngles COMPONENT Devel + ) diff --git a/unsupported/Eigen/src/EulerAngles/EulerAngles.h b/unsupported/Eigen/src/EulerAngles/EulerAngles.h new file mode 100644 index 000000000..13a0da1ab --- /dev/null +++ b/unsupported/Eigen/src/EulerAngles/EulerAngles.h @@ -0,0 +1,386 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2015 Tal Hadad <tal_hd@hotmail.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_EULERANGLESCLASS_H// TODO: Fix previous "EIGEN_EULERANGLES_H" definition? +#define EIGEN_EULERANGLESCLASS_H + +namespace Eigen +{ + /*template<typename Other, + int OtherRows=Other::RowsAtCompileTime, + int OtherCols=Other::ColsAtCompileTime> + struct ei_eulerangles_assign_impl;*/ + + /** \class EulerAngles + * + * \ingroup EulerAngles_Module + * + * \brief Represents a rotation in a 3 dimensional space as three Euler angles. + * + * Euler rotation is a set of three rotation of three angles over three fixed axes, defined by the EulerSystem given as a template parameter. + * + * Here is how intrinsic Euler angles works: + * - first, rotate the axes system over the alpha axis in angle alpha + * - then, rotate the axes system over the beta axis(which was rotated in the first stage) in angle beta + * - then, rotate the axes system over the gamma axis(which was rotated in the two stages above) in angle gamma + * + * \note This class support only intrinsic Euler angles for simplicity, + * see EulerSystem how to easily overcome this for extrinsic systems. + * + * ### Rotation representation and conversions ### + * + * It has been proved(see Wikipedia link below) that every rotation can be represented + * by Euler angles, but there is no singular representation (e.g. unlike rotation matrices). + * Therefore, you can convert from Eigen rotation and to them + * (including rotation matrices, which is not called "rotations" by Eigen design). + * + * Euler angles usually used for: + * - convenient human representation of rotation, especially in interactive GUI. + * - gimbal systems and robotics + * - efficient encoding(i.e. 3 floats only) of rotation for network protocols. + * + * However, Euler angles are slow comparing to quaternion or matrices, + * because their unnatural math definition, although it's simple for human. + * To overcome this, this class provide easy movement from the math friendly representation + * to the human friendly representation, and vise-versa. + * + * All the user need to do is a safe simple C++ type conversion, + * and this class take care for the math. + * Additionally, some axes related computation is done in compile time. + * + * #### Euler angles ranges in conversions #### + * + * When converting some rotation to Euler angles, there are some ways you can guarantee + * the Euler angles ranges. + * + * #### implicit ranges #### + * When using implicit ranges, all angles are guarantee to be in the range [-PI, +PI], + * unless you convert from some other Euler angles. + * In this case, the range is __undefined__ (might be even less than -PI or greater than +2*PI). + * \sa EulerAngles(const MatrixBase<Derived>&) + * \sa EulerAngles(const RotationBase<Derived, 3>&) + * + * #### explicit ranges #### + * When using explicit ranges, all angles are guarantee to be in the range you choose. + * In the range Boolean parameter, you're been ask whether you prefer the positive range or not: + * - _true_ - force the range between [0, +2*PI] + * - _false_ - force the range between [-PI, +PI] + * + * ##### compile time ranges ##### + * This is when you have compile time ranges and you prefer to + * use template parameter. (e.g. for performance) + * \sa FromRotation() + * + * ##### run-time time ranges ##### + * Run-time ranges are also supported. + * \sa EulerAngles(const MatrixBase<Derived>&, bool, bool, bool) + * \sa EulerAngles(const RotationBase<Derived, 3>&, bool, bool, bool) + * + * ### Convenient user typedefs ### + * + * Convenient typedefs for EulerAngles exist for float and double scalar, + * in a form of EulerAngles{A}{B}{C}{scalar}, + * e.g. \ref EulerAnglesXYZd, \ref EulerAnglesZYZf. + * + * Only for positive axes{+x,+y,+z} Euler systems are have convenient typedef. + * If you need negative axes{-x,-y,-z}, it is recommended to create you own typedef with + * a word that represent what you need. + * + * ### Example ### + * + * \include EulerAngles.cpp + * Output: \verbinclude EulerAngles.out + * + * ### Additional reading ### + * + * If you're want to get more idea about how Euler system work in Eigen see EulerSystem. + * + * More information about Euler angles: https://en.wikipedia.org/wiki/Euler_angles + * + * \tparam _Scalar the scalar type, i.e., the type of the angles. + * + * \tparam _System the EulerSystem to use, which represents the axes of rotation. + */ + template <typename _Scalar, class _System> + class EulerAngles : public RotationBase<EulerAngles<_Scalar, _System>, 3> + { + public: + /** the scalar type of the angles */ + typedef _Scalar Scalar; + + /** the EulerSystem to use, which represents the axes of rotation. */ + typedef _System System; + + typedef Matrix<Scalar,3,3> Matrix3; /*!< the equivalent rotation matrix type */ + typedef Matrix<Scalar,3,1> Vector3; /*!< the equivalent 3 dimension vector type */ + typedef Quaternion<Scalar> QuaternionType; /*!< the equivalent quaternion type */ + typedef AngleAxis<Scalar> AngleAxisType; /*!< the equivalent angle-axis type */ + + /** \returns the axis vector of the first (alpha) rotation */ + static Vector3 AlphaAxisVector() { + const Vector3& u = Vector3::Unit(System::AlphaAxisAbs - 1); + return System::IsAlphaOpposite ? -u : u; + } + + /** \returns the axis vector of the second (beta) rotation */ + static Vector3 BetaAxisVector() { + const Vector3& u = Vector3::Unit(System::BetaAxisAbs - 1); + return System::IsBetaOpposite ? -u : u; + } + + /** \returns the axis vector of the third (gamma) rotation */ + static Vector3 GammaAxisVector() { + const Vector3& u = Vector3::Unit(System::GammaAxisAbs - 1); + return System::IsGammaOpposite ? -u : u; + } + + private: + Vector3 m_angles; + + public: + /** Default constructor without initialization. */ + EulerAngles() {} + /** Constructs and initialize Euler angles(\p alpha, \p beta, \p gamma). */ + EulerAngles(const Scalar& alpha, const Scalar& beta, const Scalar& gamma) : + m_angles(alpha, beta, gamma) {} + + /** Constructs and initialize Euler angles from a 3x3 rotation matrix \p m. + * + * \note All angles will be in the range [-PI, PI]. + */ + template<typename Derived> + EulerAngles(const MatrixBase<Derived>& m) { *this = m; } + + /** Constructs and initialize Euler angles from a 3x3 rotation matrix \p m, + * with options to choose for each angle the requested range. + * + * If positive range is true, then the specified angle will be in the range [0, +2*PI]. + * Otherwise, the specified angle will be in the range [-PI, +PI]. + * + * \param m The 3x3 rotation matrix to convert + * \param positiveRangeAlpha If true, alpha will be in [0, 2*PI]. Otherwise, in [-PI, +PI]. + * \param positiveRangeBeta If true, beta will be in [0, 2*PI]. Otherwise, in [-PI, +PI]. + * \param positiveRangeGamma If true, gamma will be in [0, 2*PI]. Otherwise, in [-PI, +PI]. + */ + template<typename Derived> + EulerAngles( + const MatrixBase<Derived>& m, + bool positiveRangeAlpha, + bool positiveRangeBeta, + bool positiveRangeGamma) { + + System::CalcEulerAngles(*this, m, positiveRangeAlpha, positiveRangeBeta, positiveRangeGamma); + } + + /** Constructs and initialize Euler angles from a rotation \p rot. + * + * \note All angles will be in the range [-PI, PI], unless \p rot is an EulerAngles. + * If rot is an EulerAngles, expected EulerAngles range is __undefined__. + * (Use other functions here for enforcing range if this effect is desired) + */ + template<typename Derived> + EulerAngles(const RotationBase<Derived, 3>& rot) { *this = rot; } + + /** Constructs and initialize Euler angles from a rotation \p rot, + * with options to choose for each angle the requested range. + * + * If positive range is true, then the specified angle will be in the range [0, +2*PI]. + * Otherwise, the specified angle will be in the range [-PI, +PI]. + * + * \param rot The 3x3 rotation matrix to convert + * \param positiveRangeAlpha If true, alpha will be in [0, 2*PI]. Otherwise, in [-PI, +PI]. + * \param positiveRangeBeta If true, beta will be in [0, 2*PI]. Otherwise, in [-PI, +PI]. + * \param positiveRangeGamma If true, gamma will be in [0, 2*PI]. Otherwise, in [-PI, +PI]. + */ + template<typename Derived> + EulerAngles( + const RotationBase<Derived, 3>& rot, + bool positiveRangeAlpha, + bool positiveRangeBeta, + bool positiveRangeGamma) { + + System::CalcEulerAngles(*this, rot.toRotationMatrix(), positiveRangeAlpha, positiveRangeBeta, positiveRangeGamma); + } + + /** \returns The angle values stored in a vector (alpha, beta, gamma). */ + const Vector3& angles() const { return m_angles; } + /** \returns A read-write reference to the angle values stored in a vector (alpha, beta, gamma). */ + Vector3& angles() { return m_angles; } + + /** \returns The value of the first angle. */ + Scalar alpha() const { return m_angles[0]; } + /** \returns A read-write reference to the angle of the first angle. */ + Scalar& alpha() { return m_angles[0]; } + + /** \returns The value of the second angle. */ + Scalar beta() const { return m_angles[1]; } + /** \returns A read-write reference to the angle of the second angle. */ + Scalar& beta() { return m_angles[1]; } + + /** \returns The value of the third angle. */ + Scalar gamma() const { return m_angles[2]; } + /** \returns A read-write reference to the angle of the third angle. */ + Scalar& gamma() { return m_angles[2]; } + + /** \returns The Euler angles rotation inverse (which is as same as the negative), + * (-alpha, -beta, -gamma). + */ + EulerAngles inverse() const + { + EulerAngles res; + res.m_angles = -m_angles; + return res; + } + + /** \returns The Euler angles rotation negative (which is as same as the inverse), + * (-alpha, -beta, -gamma). + */ + EulerAngles operator -() const + { + return inverse(); + } + + /** Constructs and initialize Euler angles from a 3x3 rotation matrix \p m, + * with options to choose for each angle the requested range (__only in compile time__). + * + * If positive range is true, then the specified angle will be in the range [0, +2*PI]. + * Otherwise, the specified angle will be in the range [-PI, +PI]. + * + * \param m The 3x3 rotation matrix to convert + * \tparam positiveRangeAlpha If true, alpha will be in [0, 2*PI]. Otherwise, in [-PI, +PI]. + * \tparam positiveRangeBeta If true, beta will be in [0, 2*PI]. Otherwise, in [-PI, +PI]. + * \tparam positiveRangeGamma If true, gamma will be in [0, 2*PI]. Otherwise, in [-PI, +PI]. + */ + template< + bool PositiveRangeAlpha, + bool PositiveRangeBeta, + bool PositiveRangeGamma, + typename Derived> + static EulerAngles FromRotation(const MatrixBase<Derived>& m) + { + EIGEN_STATIC_ASSERT_MATRIX_SPECIFIC_SIZE(Derived, 3, 3) + + EulerAngles e; + System::template CalcEulerAngles< + PositiveRangeAlpha, PositiveRangeBeta, PositiveRangeGamma, _Scalar>(e, m); + return e; + } + + /** Constructs and initialize Euler angles from a rotation \p rot, + * with options to choose for each angle the requested range (__only in compile time__). + * + * If positive range is true, then the specified angle will be in the range [0, +2*PI]. + * Otherwise, the specified angle will be in the range [-PI, +PI]. + * + * \param rot The 3x3 rotation matrix to convert + * \tparam positiveRangeAlpha If true, alpha will be in [0, 2*PI]. Otherwise, in [-PI, +PI]. + * \tparam positiveRangeBeta If true, beta will be in [0, 2*PI]. Otherwise, in [-PI, +PI]. + * \tparam positiveRangeGamma If true, gamma will be in [0, 2*PI]. Otherwise, in [-PI, +PI]. + */ + template< + bool PositiveRangeAlpha, + bool PositiveRangeBeta, + bool PositiveRangeGamma, + typename Derived> + static EulerAngles FromRotation(const RotationBase<Derived, 3>& rot) + { + return FromRotation<PositiveRangeAlpha, PositiveRangeBeta, PositiveRangeGamma>(rot.toRotationMatrix()); + } + + /*EulerAngles& fromQuaternion(const QuaternionType& q) + { + // TODO: Implement it in a faster way for quaternions + // According to http://www.euclideanspace.com/maths/geometry/rotations/conversions/quaternionToEuler/ + // we can compute only the needed matrix cells and then convert to euler angles. (see ZYX example below) + // Currently we compute all matrix cells from quaternion. + + // Special case only for ZYX + //Scalar y2 = q.y() * q.y(); + //m_angles[0] = std::atan2(2*(q.w()*q.z() + q.x()*q.y()), (1 - 2*(y2 + q.z()*q.z()))); + //m_angles[1] = std::asin( 2*(q.w()*q.y() - q.z()*q.x())); + //m_angles[2] = std::atan2(2*(q.w()*q.x() + q.y()*q.z()), (1 - 2*(q.x()*q.x() + y2))); + }*/ + + /** Set \c *this from a rotation matrix(i.e. pure orthogonal matrix with determinant of +1). */ + template<typename Derived> + EulerAngles& operator=(const MatrixBase<Derived>& m) { + EIGEN_STATIC_ASSERT_MATRIX_SPECIFIC_SIZE(Derived, 3, 3) + + System::CalcEulerAngles(*this, m); + return *this; + } + + // TODO: Assign and construct from another EulerAngles (with different system) + + /** Set \c *this from a rotation. */ + template<typename Derived> + EulerAngles& operator=(const RotationBase<Derived, 3>& rot) { + System::CalcEulerAngles(*this, rot.toRotationMatrix()); + return *this; + } + + // TODO: Support isApprox function + + /** \returns an equivalent 3x3 rotation matrix. */ + Matrix3 toRotationMatrix() const + { + return static_cast<QuaternionType>(*this).toRotationMatrix(); + } + + /** Convert the Euler angles to quaternion. */ + operator QuaternionType() const + { + return + AngleAxisType(alpha(), AlphaAxisVector()) * + AngleAxisType(beta(), BetaAxisVector()) * + AngleAxisType(gamma(), GammaAxisVector()); + } + + friend std::ostream& operator<<(std::ostream& s, const EulerAngles<Scalar, System>& eulerAngles) + { + s << eulerAngles.angles().transpose(); + return s; + } + }; + +#define EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(AXES, SCALAR_TYPE, SCALAR_POSTFIX) \ + /** \ingroup EulerAngles_Module */ \ + typedef EulerAngles<SCALAR_TYPE, EulerSystem##AXES> EulerAngles##AXES##SCALAR_POSTFIX; + +#define EIGEN_EULER_ANGLES_TYPEDEFS(SCALAR_TYPE, SCALAR_POSTFIX) \ + EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(XYZ, SCALAR_TYPE, SCALAR_POSTFIX) \ + EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(XYX, SCALAR_TYPE, SCALAR_POSTFIX) \ + EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(XZY, SCALAR_TYPE, SCALAR_POSTFIX) \ + EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(XZX, SCALAR_TYPE, SCALAR_POSTFIX) \ + \ + EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(YZX, SCALAR_TYPE, SCALAR_POSTFIX) \ + EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(YZY, SCALAR_TYPE, SCALAR_POSTFIX) \ + EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(YXZ, SCALAR_TYPE, SCALAR_POSTFIX) \ + EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(YXY, SCALAR_TYPE, SCALAR_POSTFIX) \ + \ + EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(ZXY, SCALAR_TYPE, SCALAR_POSTFIX) \ + EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(ZXZ, SCALAR_TYPE, SCALAR_POSTFIX) \ + EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(ZYX, SCALAR_TYPE, SCALAR_POSTFIX) \ + EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(ZYZ, SCALAR_TYPE, SCALAR_POSTFIX) + +EIGEN_EULER_ANGLES_TYPEDEFS(float, f) +EIGEN_EULER_ANGLES_TYPEDEFS(double, d) + + namespace internal + { + template<typename _Scalar, class _System> + struct traits<EulerAngles<_Scalar, _System> > + { + typedef _Scalar Scalar; + }; + } + +} + +#endif // EIGEN_EULERANGLESCLASS_H diff --git a/unsupported/Eigen/src/EulerAngles/EulerSystem.h b/unsupported/Eigen/src/EulerAngles/EulerSystem.h new file mode 100644 index 000000000..82243e643 --- /dev/null +++ b/unsupported/Eigen/src/EulerAngles/EulerSystem.h @@ -0,0 +1,316 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2015 Tal Hadad <tal_hd@hotmail.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_EULERSYSTEM_H +#define EIGEN_EULERSYSTEM_H + +namespace Eigen +{ + // Forward declerations + template <typename _Scalar, class _System> + class EulerAngles; + + namespace internal + { + // TODO: Check if already exists on the rest API + template <int Num, bool IsPositive = (Num > 0)> + struct Abs + { + enum { value = Num }; + }; + + template <int Num> + struct Abs<Num, false> + { + enum { value = -Num }; + }; + + template <int Axis> + struct IsValidAxis + { + enum { value = Axis != 0 && Abs<Axis>::value <= 3 }; + }; + } + + #define EIGEN_EULER_ANGLES_CLASS_STATIC_ASSERT(COND,MSG) typedef char static_assertion_##MSG[(COND)?1:-1] + + /** \brief Representation of a fixed signed rotation axis for EulerSystem. + * + * \ingroup EulerAngles_Module + * + * Values here represent: + * - The axis of the rotation: X, Y or Z. + * - The sign (i.e. direction of the rotation along the axis): positive(+) or negative(-) + * + * Therefore, this could express all the axes {+X,+Y,+Z,-X,-Y,-Z} + * + * For positive axis, use +EULER_{axis}, and for negative axis use -EULER_{axis}. + */ + enum EulerAxis + { + EULER_X = 1, /*!< the X axis */ + EULER_Y = 2, /*!< the Y axis */ + EULER_Z = 3 /*!< the Z axis */ + }; + + /** \class EulerSystem + * + * \ingroup EulerAngles_Module + * + * \brief Represents a fixed Euler rotation system. + * + * This meta-class goal is to represent the Euler system in compilation time, for EulerAngles. + * + * You can use this class to get two things: + * - Build an Euler system, and then pass it as a template parameter to EulerAngles. + * - Query some compile time data about an Euler system. (e.g. Whether it's tait bryan) + * + * Euler rotation is a set of three rotation on fixed axes. (see \ref EulerAngles) + * This meta-class store constantly those signed axes. (see \ref EulerAxis) + * + * ### Types of Euler systems ### + * + * All and only valid 3 dimension Euler rotation over standard + * signed axes{+X,+Y,+Z,-X,-Y,-Z} are supported: + * - all axes X, Y, Z in each valid order (see below what order is valid) + * - rotation over the axis is supported both over the positive and negative directions. + * - both tait bryan and proper/classic Euler angles (i.e. the opposite). + * + * Since EulerSystem support both positive and negative directions, + * you may call this rotation distinction in other names: + * - _right handed_ or _left handed_ + * - _counterclockwise_ or _clockwise_ + * + * Notice all axed combination are valid, and would trigger a static assertion. + * Same unsigned axes can't be neighbors, e.g. {X,X,Y} is invalid. + * This yield two and only two classes: + * - _tait bryan_ - all unsigned axes are distinct, e.g. {X,Y,Z} + * - _proper/classic Euler angles_ - The first and the third unsigned axes is equal, + * and the second is different, e.g. {X,Y,X} + * + * ### Intrinsic vs extrinsic Euler systems ### + * + * Only intrinsic Euler systems are supported for simplicity. + * If you want to use extrinsic Euler systems, + * just use the equal intrinsic opposite order for axes and angles. + * I.e axes (A,B,C) becomes (C,B,A), and angles (a,b,c) becomes (c,b,a). + * + * ### Convenient user typedefs ### + * + * Convenient typedefs for EulerSystem exist (only for positive axes Euler systems), + * in a form of EulerSystem{A}{B}{C}, e.g. \ref EulerSystemXYZ. + * + * ### Additional reading ### + * + * More information about Euler angles: https://en.wikipedia.org/wiki/Euler_angles + * + * \tparam _AlphaAxis the first fixed EulerAxis + * + * \tparam _AlphaAxis the second fixed EulerAxis + * + * \tparam _AlphaAxis the third fixed EulerAxis + */ + template <int _AlphaAxis, int _BetaAxis, int _GammaAxis> + class EulerSystem + { + public: + // It's defined this way and not as enum, because I think + // that enum is not guerantee to support negative numbers + + /** The first rotation axis */ + static const int AlphaAxis = _AlphaAxis; + + /** The second rotation axis */ + static const int BetaAxis = _BetaAxis; + + /** The third rotation axis */ + static const int GammaAxis = _GammaAxis; + + enum + { + AlphaAxisAbs = internal::Abs<AlphaAxis>::value, /*!< the first rotation axis unsigned */ + BetaAxisAbs = internal::Abs<BetaAxis>::value, /*!< the second rotation axis unsigned */ + GammaAxisAbs = internal::Abs<GammaAxis>::value, /*!< the third rotation axis unsigned */ + + IsAlphaOpposite = (AlphaAxis < 0) ? 1 : 0, /*!< weather alpha axis is negative */ + IsBetaOpposite = (BetaAxis < 0) ? 1 : 0, /*!< weather beta axis is negative */ + IsGammaOpposite = (GammaAxis < 0) ? 1 : 0, /*!< weather gamma axis is negative */ + + IsOdd = ((AlphaAxisAbs)%3 == (BetaAxisAbs - 1)%3) ? 0 : 1, /*!< weather the Euler system is odd */ + IsEven = IsOdd ? 0 : 1, /*!< weather the Euler system is even */ + + IsTaitBryan = ((unsigned)AlphaAxisAbs != (unsigned)GammaAxisAbs) ? 1 : 0 /*!< weather the Euler system is tait bryan */ + }; + + private: + + EIGEN_EULER_ANGLES_CLASS_STATIC_ASSERT(internal::IsValidAxis<AlphaAxis>::value, + ALPHA_AXIS_IS_INVALID); + + EIGEN_EULER_ANGLES_CLASS_STATIC_ASSERT(internal::IsValidAxis<BetaAxis>::value, + BETA_AXIS_IS_INVALID); + + EIGEN_EULER_ANGLES_CLASS_STATIC_ASSERT(internal::IsValidAxis<GammaAxis>::value, + GAMMA_AXIS_IS_INVALID); + + EIGEN_EULER_ANGLES_CLASS_STATIC_ASSERT((unsigned)AlphaAxisAbs != (unsigned)BetaAxisAbs, + ALPHA_AXIS_CANT_BE_EQUAL_TO_BETA_AXIS); + + EIGEN_EULER_ANGLES_CLASS_STATIC_ASSERT((unsigned)BetaAxisAbs != (unsigned)GammaAxisAbs, + BETA_AXIS_CANT_BE_EQUAL_TO_GAMMA_AXIS); + + enum + { + // I, J, K are the pivot indexes permutation for the rotation matrix, that match this Euler system. + // They are used in this class converters. + // They are always different from each other, and their possible values are: 0, 1, or 2. + I = AlphaAxisAbs - 1, + J = (AlphaAxisAbs - 1 + 1 + IsOdd)%3, + K = (AlphaAxisAbs - 1 + 2 - IsOdd)%3 + }; + + // TODO: Get @mat parameter in form that avoids double evaluation. + template <typename Derived> + static void CalcEulerAngles_imp(Matrix<typename MatrixBase<Derived>::Scalar, 3, 1>& res, const MatrixBase<Derived>& mat, internal::true_type /*isTaitBryan*/) + { + using std::atan2; + using std::sin; + using std::cos; + + typedef typename Derived::Scalar Scalar; + typedef Matrix<Scalar,2,1> Vector2; + + res[0] = atan2(mat(J,K), mat(K,K)); + Scalar c2 = Vector2(mat(I,I), mat(I,J)).norm(); + if((IsOdd && res[0]<Scalar(0)) || ((!IsOdd) && res[0]>Scalar(0))) { + res[0] = (res[0] > Scalar(0)) ? res[0] - Scalar(EIGEN_PI) : res[0] + Scalar(EIGEN_PI); + res[1] = atan2(-mat(I,K), -c2); + } + else + res[1] = atan2(-mat(I,K), c2); + Scalar s1 = sin(res[0]); + Scalar c1 = cos(res[0]); + res[2] = atan2(s1*mat(K,I)-c1*mat(J,I), c1*mat(J,J) - s1 * mat(K,J)); + } + + template <typename Derived> + static void CalcEulerAngles_imp(Matrix<typename MatrixBase<Derived>::Scalar,3,1>& res, const MatrixBase<Derived>& mat, internal::false_type /*isTaitBryan*/) + { + using std::atan2; + using std::sin; + using std::cos; + + typedef typename Derived::Scalar Scalar; + typedef Matrix<Scalar,2,1> Vector2; + + res[0] = atan2(mat(J,I), mat(K,I)); + if((IsOdd && res[0]<Scalar(0)) || ((!IsOdd) && res[0]>Scalar(0))) + { + res[0] = (res[0] > Scalar(0)) ? res[0] - Scalar(EIGEN_PI) : res[0] + Scalar(EIGEN_PI); + Scalar s2 = Vector2(mat(J,I), mat(K,I)).norm(); + res[1] = -atan2(s2, mat(I,I)); + } + else + { + Scalar s2 = Vector2(mat(J,I), mat(K,I)).norm(); + res[1] = atan2(s2, mat(I,I)); + } + + // With a=(0,1,0), we have i=0; j=1; k=2, and after computing the first two angles, + // we can compute their respective rotation, and apply its inverse to M. Since the result must + // be a rotation around x, we have: + // + // c2 s1.s2 c1.s2 1 0 0 + // 0 c1 -s1 * M = 0 c3 s3 + // -s2 s1.c2 c1.c2 0 -s3 c3 + // + // Thus: m11.c1 - m21.s1 = c3 & m12.c1 - m22.s1 = s3 + + Scalar s1 = sin(res[0]); + Scalar c1 = cos(res[0]); + res[2] = atan2(c1*mat(J,K)-s1*mat(K,K), c1*mat(J,J) - s1 * mat(K,J)); + } + + template<typename Scalar> + static void CalcEulerAngles( + EulerAngles<Scalar, EulerSystem>& res, + const typename EulerAngles<Scalar, EulerSystem>::Matrix3& mat) + { + CalcEulerAngles(res, mat, false, false, false); + } + + template< + bool PositiveRangeAlpha, + bool PositiveRangeBeta, + bool PositiveRangeGamma, + typename Scalar> + static void CalcEulerAngles( + EulerAngles<Scalar, EulerSystem>& res, + const typename EulerAngles<Scalar, EulerSystem>::Matrix3& mat) + { + CalcEulerAngles(res, mat, PositiveRangeAlpha, PositiveRangeBeta, PositiveRangeGamma); + } + + template<typename Scalar> + static void CalcEulerAngles( + EulerAngles<Scalar, EulerSystem>& res, + const typename EulerAngles<Scalar, EulerSystem>::Matrix3& mat, + bool PositiveRangeAlpha, + bool PositiveRangeBeta, + bool PositiveRangeGamma) + { + CalcEulerAngles_imp( + res.angles(), mat, + typename internal::conditional<IsTaitBryan, internal::true_type, internal::false_type>::type()); + + if (IsAlphaOpposite == IsOdd) + res.alpha() = -res.alpha(); + + if (IsBetaOpposite == IsOdd) + res.beta() = -res.beta(); + + if (IsGammaOpposite == IsOdd) + res.gamma() = -res.gamma(); + + // Saturate results to the requested range + if (PositiveRangeAlpha && (res.alpha() < 0)) + res.alpha() += Scalar(2 * EIGEN_PI); + + if (PositiveRangeBeta && (res.beta() < 0)) + res.beta() += Scalar(2 * EIGEN_PI); + + if (PositiveRangeGamma && (res.gamma() < 0)) + res.gamma() += Scalar(2 * EIGEN_PI); + } + + template <typename _Scalar, class _System> + friend class Eigen::EulerAngles; + }; + +#define EIGEN_EULER_SYSTEM_TYPEDEF(A, B, C) \ + /** \ingroup EulerAngles_Module */ \ + typedef EulerSystem<EULER_##A, EULER_##B, EULER_##C> EulerSystem##A##B##C; + + EIGEN_EULER_SYSTEM_TYPEDEF(X,Y,Z) + EIGEN_EULER_SYSTEM_TYPEDEF(X,Y,X) + EIGEN_EULER_SYSTEM_TYPEDEF(X,Z,Y) + EIGEN_EULER_SYSTEM_TYPEDEF(X,Z,X) + + EIGEN_EULER_SYSTEM_TYPEDEF(Y,Z,X) + EIGEN_EULER_SYSTEM_TYPEDEF(Y,Z,Y) + EIGEN_EULER_SYSTEM_TYPEDEF(Y,X,Z) + EIGEN_EULER_SYSTEM_TYPEDEF(Y,X,Y) + + EIGEN_EULER_SYSTEM_TYPEDEF(Z,X,Y) + EIGEN_EULER_SYSTEM_TYPEDEF(Z,X,Z) + EIGEN_EULER_SYSTEM_TYPEDEF(Z,Y,X) + EIGEN_EULER_SYSTEM_TYPEDEF(Z,Y,Z) +} + +#endif // EIGEN_EULERSYSTEM_H diff --git a/unsupported/Eigen/src/FFT/CMakeLists.txt b/unsupported/Eigen/src/FFT/CMakeLists.txt deleted file mode 100644 index edcffcb18..000000000 --- a/unsupported/Eigen/src/FFT/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_FFT_SRCS "*.h") - -INSTALL(FILES - ${Eigen_FFT_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/src/FFT COMPONENT Devel - ) diff --git a/unsupported/Eigen/src/IterativeSolvers/CMakeLists.txt b/unsupported/Eigen/src/IterativeSolvers/CMakeLists.txt deleted file mode 100644 index 7986afc5e..000000000 --- a/unsupported/Eigen/src/IterativeSolvers/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_IterativeSolvers_SRCS "*.h") - -INSTALL(FILES - ${Eigen_IterativeSolvers_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/src/IterativeSolvers COMPONENT Devel - ) diff --git a/unsupported/Eigen/src/KroneckerProduct/CMakeLists.txt b/unsupported/Eigen/src/KroneckerProduct/CMakeLists.txt deleted file mode 100644 index 4daefebee..000000000 --- a/unsupported/Eigen/src/KroneckerProduct/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_KroneckerProduct_SRCS "*.h") - -INSTALL(FILES - ${Eigen_KroneckerProduct_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/src/KroneckerProduct COMPONENT Devel - ) diff --git a/unsupported/Eigen/src/LevenbergMarquardt/CMakeLists.txt b/unsupported/Eigen/src/LevenbergMarquardt/CMakeLists.txt deleted file mode 100644 index d9690854d..000000000 --- a/unsupported/Eigen/src/LevenbergMarquardt/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_LevenbergMarquardt_SRCS "*.h") - -INSTALL(FILES - ${Eigen_LevenbergMarquardt_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/src/LevenbergMarquardt COMPONENT Devel - ) diff --git a/unsupported/Eigen/src/MatrixFunctions/CMakeLists.txt b/unsupported/Eigen/src/MatrixFunctions/CMakeLists.txt deleted file mode 100644 index cdde64d2c..000000000 --- a/unsupported/Eigen/src/MatrixFunctions/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_MatrixFunctions_SRCS "*.h") - -INSTALL(FILES - ${Eigen_MatrixFunctions_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/src/MatrixFunctions COMPONENT Devel - ) diff --git a/unsupported/Eigen/src/MoreVectorization/CMakeLists.txt b/unsupported/Eigen/src/MoreVectorization/CMakeLists.txt deleted file mode 100644 index 1b887cc8e..000000000 --- a/unsupported/Eigen/src/MoreVectorization/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_MoreVectorization_SRCS "*.h") - -INSTALL(FILES - ${Eigen_MoreVectorization_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/src/MoreVectorization COMPONENT Devel - ) diff --git a/unsupported/Eigen/src/NonLinearOptimization/CMakeLists.txt b/unsupported/Eigen/src/NonLinearOptimization/CMakeLists.txt deleted file mode 100644 index 9322ddadf..000000000 --- a/unsupported/Eigen/src/NonLinearOptimization/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_NonLinearOptimization_SRCS "*.h") - -INSTALL(FILES - ${Eigen_NonLinearOptimization_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/src/NonLinearOptimization COMPONENT Devel - ) diff --git a/unsupported/Eigen/src/NumericalDiff/CMakeLists.txt b/unsupported/Eigen/src/NumericalDiff/CMakeLists.txt deleted file mode 100644 index 1199aca2f..000000000 --- a/unsupported/Eigen/src/NumericalDiff/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_NumericalDiff_SRCS "*.h") - -INSTALL(FILES - ${Eigen_NumericalDiff_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/src/NumericalDiff COMPONENT Devel - ) diff --git a/unsupported/Eigen/src/Polynomials/CMakeLists.txt b/unsupported/Eigen/src/Polynomials/CMakeLists.txt deleted file mode 100644 index 51f13f3cb..000000000 --- a/unsupported/Eigen/src/Polynomials/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_Polynomials_SRCS "*.h") - -INSTALL(FILES - ${Eigen_Polynomials_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/src/Polynomials COMPONENT Devel - ) diff --git a/unsupported/Eigen/src/Skyline/CMakeLists.txt b/unsupported/Eigen/src/Skyline/CMakeLists.txt deleted file mode 100644 index 3bf1b0dd4..000000000 --- a/unsupported/Eigen/src/Skyline/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_Skyline_SRCS "*.h") - -INSTALL(FILES - ${Eigen_Skyline_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/src/Skyline COMPONENT Devel - ) diff --git a/unsupported/Eigen/src/SparseExtra/CMakeLists.txt b/unsupported/Eigen/src/SparseExtra/CMakeLists.txt deleted file mode 100644 index 7ea32ca5e..000000000 --- a/unsupported/Eigen/src/SparseExtra/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_SparseExtra_SRCS "*.h") - -INSTALL(FILES - ${Eigen_SparseExtra_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/src/SparseExtra COMPONENT Devel - ) diff --git a/unsupported/Eigen/src/SpecialFunctions/CMakeLists.txt b/unsupported/Eigen/src/SpecialFunctions/CMakeLists.txt deleted file mode 100644 index 25df9439d..000000000 --- a/unsupported/Eigen/src/SpecialFunctions/CMakeLists.txt +++ /dev/null @@ -1,11 +0,0 @@ -FILE(GLOB Eigen_SpecialFunctions_SRCS "*.h") -INSTALL(FILES - ${Eigen_SpecialFunctions_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/src/SpecialFunctions COMPONENT Devel - ) - -FILE(GLOB Eigen_SpecialFunctions_arch_CUDA_SRCS "arch/CUDA/*.h") -INSTALL(FILES - ${Eigen_SpecialFunctions_arch_CUDA_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/src/SpecialFunctions/arch/CUDA COMPONENT Devel - )
\ No newline at end of file diff --git a/unsupported/Eigen/src/Splines/CMakeLists.txt b/unsupported/Eigen/src/Splines/CMakeLists.txt deleted file mode 100644 index 55c6271e9..000000000 --- a/unsupported/Eigen/src/Splines/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_Splines_SRCS "*.h") - -INSTALL(FILES - ${Eigen_Splines_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/src/Splines COMPONENT Devel - ) diff --git a/unsupported/doc/examples/EulerAngles.cpp b/unsupported/doc/examples/EulerAngles.cpp new file mode 100644 index 000000000..1ef6aee18 --- /dev/null +++ b/unsupported/doc/examples/EulerAngles.cpp @@ -0,0 +1,46 @@ +#include <unsupported/Eigen/EulerAngles> +#include <iostream> + +using namespace Eigen; + +int main() +{ + // A common Euler system by many armies around the world, + // where the first one is the azimuth(the angle from the north - + // the same angle that is show in compass) + // and the second one is elevation(the angle from the horizon) + // and the third one is roll(the angle between the horizontal body + // direction and the plane ground surface) + // Keep remembering we're using radian angles here! + typedef EulerSystem<-EULER_Z, EULER_Y, EULER_X> MyArmySystem; + typedef EulerAngles<double, MyArmySystem> MyArmyAngles; + + MyArmyAngles vehicleAngles( + 3.14/*PI*/ / 2, /* heading to east, notice that this angle is counter-clockwise */ + -0.3, /* going down from a mountain */ + 0.1); /* slightly rolled to the right */ + + // Some Euler angles representation that our plane use. + EulerAnglesZYZd planeAngles(0.78474, 0.5271, -0.513794); + + MyArmyAngles planeAnglesInMyArmyAngles = MyArmyAngles::FromRotation<true, false, false>(planeAngles); + + std::cout << "vehicle angles(MyArmy): " << vehicleAngles << std::endl; + std::cout << "plane angles(ZYZ): " << planeAngles << std::endl; + std::cout << "plane angles(MyArmy): " << planeAnglesInMyArmyAngles << std::endl; + + // Now lets rotate the plane a little bit + std::cout << "==========================================================\n"; + std::cout << "rotating plane now!\n"; + std::cout << "==========================================================\n"; + + Quaterniond planeRotated = AngleAxisd(-0.342, Vector3d::UnitY()) * planeAngles; + + planeAngles = planeRotated; + planeAnglesInMyArmyAngles = MyArmyAngles::FromRotation<true, false, false>(planeRotated); + + std::cout << "new plane angles(ZYZ): " << planeAngles << std::endl; + std::cout << "new plane angles(MyArmy): " << planeAnglesInMyArmyAngles << std::endl; + + return 0; +} diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index de9b5243a..0d7ed1db2 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -59,6 +59,8 @@ ei_add_test(alignedvector3) ei_add_test(FFT) +ei_add_test(EulerAngles) + find_package(MPFR 2.3.0) find_package(GMP) if(MPFR_FOUND AND EIGEN_COMPILER_SUPPORT_CXX11) @@ -230,20 +232,25 @@ if(CUDA_FOUND AND EIGEN_TEST_CUDA) cuda_include_directories("${CMAKE_CURRENT_BINARY_DIR}" "${CUDA_TOOLKIT_ROOT_DIR}/include") set(EIGEN_ADD_TEST_FILENAME_EXTENSION "cu") - ei_add_test(cxx11_tensor_device) - ei_add_test(cxx11_tensor_cuda) - ei_add_test(cxx11_tensor_contract_cuda) + ei_add_test(cxx11_tensor_complex_cuda) ei_add_test(cxx11_tensor_reduction_cuda) ei_add_test(cxx11_tensor_argmax_cuda) ei_add_test(cxx11_tensor_cast_float16_cuda) ei_add_test(cxx11_tensor_scan_cuda) + # Contractions require arch 3.0 or higher + if (${EIGEN_CUDA_COMPUTE_ARCH} GREATER 29) + ei_add_test(cxx11_tensor_device) + ei_add_test(cxx11_tensor_cuda) + ei_add_test(cxx11_tensor_contract_cuda) + ei_add_test(cxx11_tensor_of_float16_cuda) + endif() + # The random number generation code requires arch 3.5 or greater. if (${EIGEN_CUDA_COMPUTE_ARCH} GREATER 34) ei_add_test(cxx11_tensor_random_cuda) endif() - ei_add_test(cxx11_tensor_of_float16_cuda) unset(EIGEN_ADD_TEST_FILENAME_EXTENSION) endif() diff --git a/unsupported/test/EulerAngles.cpp b/unsupported/test/EulerAngles.cpp new file mode 100644 index 000000000..a8cb52864 --- /dev/null +++ b/unsupported/test/EulerAngles.cpp @@ -0,0 +1,208 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2015 Tal Hadad <tal_hd@hotmail.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/. + +#include "main.h" + +#include <unsupported/Eigen/EulerAngles> + +using namespace Eigen; + +template<typename EulerSystem, typename Scalar> +void verify_euler_ranged(const Matrix<Scalar,3,1>& ea, + bool positiveRangeAlpha, bool positiveRangeBeta, bool positiveRangeGamma) +{ + typedef EulerAngles<Scalar, EulerSystem> EulerAnglesType; + typedef Matrix<Scalar,3,3> Matrix3; + typedef Matrix<Scalar,3,1> Vector3; + typedef Quaternion<Scalar> QuaternionType; + typedef AngleAxis<Scalar> AngleAxisType; + using std::abs; + + Scalar alphaRangeStart, alphaRangeEnd; + Scalar betaRangeStart, betaRangeEnd; + Scalar gammaRangeStart, gammaRangeEnd; + + if (positiveRangeAlpha) + { + alphaRangeStart = Scalar(0); + alphaRangeEnd = Scalar(2 * EIGEN_PI); + } + else + { + alphaRangeStart = -Scalar(EIGEN_PI); + alphaRangeEnd = Scalar(EIGEN_PI); + } + + if (positiveRangeBeta) + { + betaRangeStart = Scalar(0); + betaRangeEnd = Scalar(2 * EIGEN_PI); + } + else + { + betaRangeStart = -Scalar(EIGEN_PI); + betaRangeEnd = Scalar(EIGEN_PI); + } + + if (positiveRangeGamma) + { + gammaRangeStart = Scalar(0); + gammaRangeEnd = Scalar(2 * EIGEN_PI); + } + else + { + gammaRangeStart = -Scalar(EIGEN_PI); + gammaRangeEnd = Scalar(EIGEN_PI); + } + + const int i = EulerSystem::AlphaAxisAbs - 1; + const int j = EulerSystem::BetaAxisAbs - 1; + const int k = EulerSystem::GammaAxisAbs - 1; + + const int iFactor = EulerSystem::IsAlphaOpposite ? -1 : 1; + const int jFactor = EulerSystem::IsBetaOpposite ? -1 : 1; + const int kFactor = EulerSystem::IsGammaOpposite ? -1 : 1; + + const Vector3 I = EulerAnglesType::AlphaAxisVector(); + const Vector3 J = EulerAnglesType::BetaAxisVector(); + const Vector3 K = EulerAnglesType::GammaAxisVector(); + + EulerAnglesType e(ea[0], ea[1], ea[2]); + + Matrix3 m(e); + Vector3 eabis = EulerAnglesType(m, positiveRangeAlpha, positiveRangeBeta, positiveRangeGamma).angles(); + + // Check that eabis in range + VERIFY(alphaRangeStart <= eabis[0] && eabis[0] <= alphaRangeEnd); + VERIFY(betaRangeStart <= eabis[1] && eabis[1] <= betaRangeEnd); + VERIFY(gammaRangeStart <= eabis[2] && eabis[2] <= gammaRangeEnd); + + Vector3 eabis2 = m.eulerAngles(i, j, k); + + // Invert the relevant axes + eabis2[0] *= iFactor; + eabis2[1] *= jFactor; + eabis2[2] *= kFactor; + + // Saturate the angles to the correct range + if (positiveRangeAlpha && (eabis2[0] < 0)) + eabis2[0] += Scalar(2 * EIGEN_PI); + if (positiveRangeBeta && (eabis2[1] < 0)) + eabis2[1] += Scalar(2 * EIGEN_PI); + if (positiveRangeGamma && (eabis2[2] < 0)) + eabis2[2] += Scalar(2 * EIGEN_PI); + + VERIFY_IS_APPROX(eabis, eabis2);// Verify that our estimation is the same as m.eulerAngles() is + + Matrix3 mbis(AngleAxisType(eabis[0], I) * AngleAxisType(eabis[1], J) * AngleAxisType(eabis[2], K)); + VERIFY_IS_APPROX(m, mbis); + + // Tests that are only relevant for no possitive range + if (!(positiveRangeAlpha || positiveRangeBeta || positiveRangeGamma)) + { + /* If I==K, and ea[1]==0, then there no unique solution. */ + /* The remark apply in the case where I!=K, and |ea[1]| is close to pi/2. */ + if( (i!=k || ea[1]!=0) && (i==k || !internal::isApprox(abs(ea[1]),Scalar(EIGEN_PI/2),test_precision<Scalar>())) ) + VERIFY((ea-eabis).norm() <= test_precision<Scalar>()); + + // approx_or_less_than does not work for 0 + VERIFY(0 < eabis[0] || test_isMuchSmallerThan(eabis[0], Scalar(1))); + } + + // Quaternions + QuaternionType q(e); + eabis = EulerAnglesType(q, positiveRangeAlpha, positiveRangeBeta, positiveRangeGamma).angles(); + VERIFY_IS_APPROX(eabis, eabis2);// Verify that the euler angles are still the same +} + +template<typename EulerSystem, typename Scalar> +void verify_euler(const Matrix<Scalar,3,1>& ea) +{ + verify_euler_ranged<EulerSystem>(ea, false, false, false); + verify_euler_ranged<EulerSystem>(ea, false, false, true); + verify_euler_ranged<EulerSystem>(ea, false, true, false); + verify_euler_ranged<EulerSystem>(ea, false, true, true); + verify_euler_ranged<EulerSystem>(ea, true, false, false); + verify_euler_ranged<EulerSystem>(ea, true, false, true); + verify_euler_ranged<EulerSystem>(ea, true, true, false); + verify_euler_ranged<EulerSystem>(ea, true, true, true); +} + +template<typename Scalar> void check_all_var(const Matrix<Scalar,3,1>& ea) +{ + verify_euler<EulerSystemXYZ>(ea); + verify_euler<EulerSystemXYX>(ea); + verify_euler<EulerSystemXZY>(ea); + verify_euler<EulerSystemXZX>(ea); + + verify_euler<EulerSystemYZX>(ea); + verify_euler<EulerSystemYZY>(ea); + verify_euler<EulerSystemYXZ>(ea); + verify_euler<EulerSystemYXY>(ea); + + verify_euler<EulerSystemZXY>(ea); + verify_euler<EulerSystemZXZ>(ea); + verify_euler<EulerSystemZYX>(ea); + verify_euler<EulerSystemZYZ>(ea); +} + +template<typename Scalar> void eulerangles() +{ + typedef Matrix<Scalar,3,3> Matrix3; + typedef Matrix<Scalar,3,1> Vector3; + typedef Array<Scalar,3,1> Array3; + typedef Quaternion<Scalar> Quaternionx; + typedef AngleAxis<Scalar> AngleAxisType; + + Scalar a = internal::random<Scalar>(-Scalar(EIGEN_PI), Scalar(EIGEN_PI)); + Quaternionx q1; + q1 = AngleAxisType(a, Vector3::Random().normalized()); + Matrix3 m; + m = q1; + + Vector3 ea = m.eulerAngles(0,1,2); + check_all_var(ea); + ea = m.eulerAngles(0,1,0); + check_all_var(ea); + + // Check with purely random Quaternion: + q1.coeffs() = Quaternionx::Coefficients::Random().normalized(); + m = q1; + ea = m.eulerAngles(0,1,2); + check_all_var(ea); + ea = m.eulerAngles(0,1,0); + check_all_var(ea); + + // Check with random angles in range [0:pi]x[-pi:pi]x[-pi:pi]. + ea = (Array3::Random() + Array3(1,0,0))*Scalar(EIGEN_PI)*Array3(0.5,1,1); + check_all_var(ea); + + ea[2] = ea[0] = internal::random<Scalar>(0,Scalar(EIGEN_PI)); + check_all_var(ea); + + ea[0] = ea[1] = internal::random<Scalar>(0,Scalar(EIGEN_PI)); + check_all_var(ea); + + ea[1] = 0; + check_all_var(ea); + + ea.head(2).setZero(); + check_all_var(ea); + + ea.setZero(); + check_all_var(ea); +} + +void test_EulerAngles() +{ + for(int i = 0; i < g_repeat; i++) { + CALL_SUBTEST_1( eulerangles<float>() ); + CALL_SUBTEST_2( eulerangles<double>() ); + } +} diff --git a/unsupported/test/cxx11_eventcount.cpp b/unsupported/test/cxx11_eventcount.cpp index f16cc6f07..3b598bf42 100644 --- a/unsupported/test/cxx11_eventcount.cpp +++ b/unsupported/test/cxx11_eventcount.cpp @@ -25,7 +25,8 @@ int rand_reentrant(unsigned int* s) { static void test_basic_eventcount() { - std::vector<EventCount::Waiter> waiters(1); + MaxSizeVector<EventCount::Waiter> waiters(1); + waiters.resize(1); EventCount ec(waiters); EventCount::Waiter& w = waiters[0]; ec.Notify(false); @@ -81,7 +82,8 @@ static void test_stress_eventcount() static const int kEvents = 1 << 16; static const int kQueues = 10; - std::vector<EventCount::Waiter> waiters(kThreads); + MaxSizeVector<EventCount::Waiter> waiters(kThreads); + waiters.resize(kThreads); EventCount ec(waiters); TestQueue queues[kQueues]; diff --git a/unsupported/test/cxx11_tensor_argmax_cuda.cu b/unsupported/test/cxx11_tensor_argmax_cuda.cu index 41ccbe974..6fe8982f2 100644 --- a/unsupported/test/cxx11_tensor_argmax_cuda.cu +++ b/unsupported/test/cxx11_tensor_argmax_cuda.cu @@ -12,6 +12,9 @@ #define EIGEN_TEST_FUNC cxx11_tensor_cuda #define EIGEN_USE_GPU +#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500 +#include <cuda_fp16.h> +#endif #include "main.h" #include <unsupported/Eigen/CXX11/Tensor> diff --git a/unsupported/test/cxx11_tensor_cast_float16_cuda.cu b/unsupported/test/cxx11_tensor_cast_float16_cuda.cu index f22b99de8..88c233994 100644 --- a/unsupported/test/cxx11_tensor_cast_float16_cuda.cu +++ b/unsupported/test/cxx11_tensor_cast_float16_cuda.cu @@ -13,7 +13,9 @@ #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int #define EIGEN_USE_GPU - +#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500 +#include <cuda_fp16.h> +#endif #include "main.h" #include <unsupported/Eigen/CXX11/Tensor> diff --git a/unsupported/test/cxx11_tensor_complex_cuda.cu b/unsupported/test/cxx11_tensor_complex_cuda.cu new file mode 100644 index 000000000..74befe670 --- /dev/null +++ b/unsupported/test/cxx11_tensor_complex_cuda.cu @@ -0,0 +1,78 @@ +// 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/. + +#define EIGEN_TEST_NO_LONGDOUBLE +#define EIGEN_TEST_FUNC cxx11_tensor_complex +#define EIGEN_USE_GPU + +#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500 +#include <cuda_fp16.h> +#endif +#include "main.h" +#include <unsupported/Eigen/CXX11/Tensor> + +using Eigen::Tensor; + +void test_cuda_nullary() { + Tensor<std::complex<float>, 1, 0, int> in1(2); + Tensor<std::complex<float>, 1, 0, int> in2(2); + in1.setRandom(); + in2.setRandom(); + + std::size_t float_bytes = in1.size() * sizeof(float); + std::size_t complex_bytes = in1.size() * sizeof(std::complex<float>); + + std::complex<float>* d_in1; + std::complex<float>* d_in2; + float* d_out2; + cudaMalloc((void**)(&d_in1), complex_bytes); + cudaMalloc((void**)(&d_in2), complex_bytes); + cudaMalloc((void**)(&d_out2), float_bytes); + cudaMemcpy(d_in1, in1.data(), complex_bytes, cudaMemcpyHostToDevice); + cudaMemcpy(d_in2, in2.data(), complex_bytes, cudaMemcpyHostToDevice); + + Eigen::CudaStreamDevice stream; + Eigen::GpuDevice gpu_device(&stream); + + Eigen::TensorMap<Eigen::Tensor<std::complex<float>, 1, 0, int>, Eigen::Aligned> gpu_in1( + d_in1, 2); + Eigen::TensorMap<Eigen::Tensor<std::complex<float>, 1, 0, int>, Eigen::Aligned> gpu_in2( + d_in2, 2); + Eigen::TensorMap<Eigen::Tensor<float, 1, 0, int>, Eigen::Aligned> gpu_out2( + d_out2, 2); + + gpu_in1.device(gpu_device) = gpu_in1.constant(std::complex<float>(3.14f, 2.7f)); + gpu_out2.device(gpu_device) = gpu_in2.abs(); + + Tensor<std::complex<float>, 1, 0, int> new1(2); + Tensor<float, 1, 0, int> new2(2); + + assert(cudaMemcpyAsync(new1.data(), d_in1, complex_bytes, cudaMemcpyDeviceToHost, + gpu_device.stream()) == cudaSuccess); + assert(cudaMemcpyAsync(new2.data(), d_out2, float_bytes, cudaMemcpyDeviceToHost, + gpu_device.stream()) == cudaSuccess); + + assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess); + + for (int i = 0; i < 2; ++i) { + VERIFY_IS_APPROX(new1(i), std::complex<float>(3.14f, 2.7f)); + VERIFY_IS_APPROX(new2(i), std::abs(in2(i))); + } + + cudaFree(d_in1); + cudaFree(d_in2); + cudaFree(d_out2); +} + + + +void test_cxx11_tensor_complex() +{ + CALL_SUBTEST(test_cuda_nullary()); +} diff --git a/unsupported/test/cxx11_tensor_contract_cuda.cu b/unsupported/test/cxx11_tensor_contract_cuda.cu index 98ac180ef..767e9c678 100644 --- a/unsupported/test/cxx11_tensor_contract_cuda.cu +++ b/unsupported/test/cxx11_tensor_contract_cuda.cu @@ -14,7 +14,9 @@ #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int #define EIGEN_USE_GPU - +#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500 +#include <cuda_fp16.h> +#endif #include "main.h" #include <unsupported/Eigen/CXX11/Tensor> diff --git a/unsupported/test/cxx11_tensor_contraction.cpp b/unsupported/test/cxx11_tensor_contraction.cpp index 73623b2ed..ace97057f 100644 --- a/unsupported/test/cxx11_tensor_contraction.cpp +++ b/unsupported/test/cxx11_tensor_contraction.cpp @@ -489,6 +489,27 @@ static void test_tensor_product() } +template<int DataLayout> +static void test_const_inputs() +{ + Tensor<float, 2, DataLayout> in1(2, 3); + Tensor<float, 2, DataLayout> in2(3, 2); + in1.setRandom(); + in2.setRandom(); + + TensorMap<Tensor<const float, 2, DataLayout> > mat1(in1.data(), 2, 3); + TensorMap<Tensor<const float, 2, DataLayout> > mat2(in2.data(), 3, 2); + Tensor<float, 2, DataLayout> mat3(2,2); + + Eigen::array<DimPair, 1> dims = {{DimPair(1, 0)}}; + mat3 = mat1.contract(mat2, dims); + + VERIFY_IS_APPROX(mat3(0,0), mat1(0,0)*mat2(0,0) + mat1(0,1)*mat2(1,0) + mat1(0,2)*mat2(2,0)); + VERIFY_IS_APPROX(mat3(0,1), mat1(0,0)*mat2(0,1) + mat1(0,1)*mat2(1,1) + mat1(0,2)*mat2(2,1)); + VERIFY_IS_APPROX(mat3(1,0), mat1(1,0)*mat2(0,0) + mat1(1,1)*mat2(1,0) + mat1(1,2)*mat2(2,0)); + VERIFY_IS_APPROX(mat3(1,1), mat1(1,0)*mat2(0,1) + mat1(1,1)*mat2(1,1) + mat1(1,2)*mat2(2,1)); +} + void test_cxx11_tensor_contraction() { CALL_SUBTEST(test_evals<ColMajor>()); @@ -519,4 +540,6 @@ void test_cxx11_tensor_contraction() CALL_SUBTEST(test_small_blocking_factors<RowMajor>()); CALL_SUBTEST(test_tensor_product<ColMajor>()); CALL_SUBTEST(test_tensor_product<RowMajor>()); + CALL_SUBTEST(test_const_inputs<ColMajor>()); + CALL_SUBTEST(test_const_inputs<RowMajor>()); } diff --git a/unsupported/test/cxx11_tensor_cuda.cu b/unsupported/test/cxx11_tensor_cuda.cu index 284b46803..bf216587a 100644 --- a/unsupported/test/cxx11_tensor_cuda.cu +++ b/unsupported/test/cxx11_tensor_cuda.cu @@ -10,19 +10,65 @@ #define EIGEN_TEST_NO_LONGDOUBLE #define EIGEN_TEST_NO_COMPLEX #define EIGEN_TEST_FUNC cxx11_tensor_cuda -#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int #define EIGEN_USE_GPU - +#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500 +#include <cuda_fp16.h> +#endif #include "main.h" #include <unsupported/Eigen/CXX11/Tensor> using Eigen::Tensor; +void test_cuda_nullary() { + Tensor<float, 1, 0, int> in1(2); + Tensor<float, 1, 0, int> in2(2); + in1.setRandom(); + in2.setRandom(); + + std::size_t tensor_bytes = in1.size() * sizeof(float); + + float* d_in1; + float* d_in2; + cudaMalloc((void**)(&d_in1), tensor_bytes); + cudaMalloc((void**)(&d_in2), tensor_bytes); + cudaMemcpy(d_in1, in1.data(), tensor_bytes, cudaMemcpyHostToDevice); + cudaMemcpy(d_in2, in2.data(), tensor_bytes, cudaMemcpyHostToDevice); + + Eigen::CudaStreamDevice stream; + Eigen::GpuDevice gpu_device(&stream); + + Eigen::TensorMap<Eigen::Tensor<float, 1, 0, int>, Eigen::Aligned> gpu_in1( + d_in1, 2); + Eigen::TensorMap<Eigen::Tensor<float, 1, 0, int>, Eigen::Aligned> gpu_in2( + d_in2, 2); + + gpu_in1.device(gpu_device) = gpu_in1.constant(3.14f); + gpu_in2.device(gpu_device) = gpu_in2.random(); + + Tensor<float, 1, 0, int> new1(2); + Tensor<float, 1, 0, int> new2(2); + + assert(cudaMemcpyAsync(new1.data(), d_in1, tensor_bytes, cudaMemcpyDeviceToHost, + gpu_device.stream()) == cudaSuccess); + assert(cudaMemcpyAsync(new2.data(), d_in2, tensor_bytes, cudaMemcpyDeviceToHost, + gpu_device.stream()) == cudaSuccess); + + assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess); + + for (int i = 0; i < 2; ++i) { + VERIFY_IS_APPROX(new1(i), 3.14f); + VERIFY_IS_NOT_EQUAL(new2(i), in2(i)); + } + + cudaFree(d_in1); + cudaFree(d_in2); +} + void test_cuda_elementwise_small() { - Tensor<float, 1> in1(Eigen::array<int, 1>(2)); - Tensor<float, 1> in2(Eigen::array<int, 1>(2)); - Tensor<float, 1> out(Eigen::array<int, 1>(2)); + Tensor<float, 1> in1(Eigen::array<Eigen::DenseIndex, 1>(2)); + Tensor<float, 1> in2(Eigen::array<Eigen::DenseIndex, 1>(2)); + Tensor<float, 1> out(Eigen::array<Eigen::DenseIndex, 1>(2)); in1.setRandom(); in2.setRandom(); @@ -44,11 +90,11 @@ void test_cuda_elementwise_small() { Eigen::GpuDevice gpu_device(&stream); Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_in1( - d_in1, Eigen::array<int, 1>(2)); + d_in1, Eigen::array<Eigen::DenseIndex, 1>(2)); Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_in2( - d_in2, Eigen::array<int, 1>(2)); + d_in2, Eigen::array<Eigen::DenseIndex, 1>(2)); Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_out( - d_out, Eigen::array<int, 1>(2)); + d_out, Eigen::array<Eigen::DenseIndex, 1>(2)); gpu_out.device(gpu_device) = gpu_in1 + gpu_in2; @@ -58,8 +104,8 @@ void test_cuda_elementwise_small() { for (int i = 0; i < 2; ++i) { VERIFY_IS_APPROX( - out(Eigen::array<int, 1>(i)), - in1(Eigen::array<int, 1>(i)) + in2(Eigen::array<int, 1>(i))); + out(Eigen::array<Eigen::DenseIndex, 1>(i)), + in1(Eigen::array<Eigen::DenseIndex, 1>(i)) + in2(Eigen::array<Eigen::DenseIndex, 1>(i))); } cudaFree(d_in1); @@ -69,10 +115,10 @@ void test_cuda_elementwise_small() { void test_cuda_elementwise() { - Tensor<float, 3> in1(Eigen::array<int, 3>(72,53,97)); - Tensor<float, 3> in2(Eigen::array<int, 3>(72,53,97)); - Tensor<float, 3> in3(Eigen::array<int, 3>(72,53,97)); - Tensor<float, 3> out(Eigen::array<int, 3>(72,53,97)); + Tensor<float, 3> in1(Eigen::array<Eigen::DenseIndex, 3>(72,53,97)); + Tensor<float, 3> in2(Eigen::array<Eigen::DenseIndex, 3>(72,53,97)); + Tensor<float, 3> in3(Eigen::array<Eigen::DenseIndex, 3>(72,53,97)); + Tensor<float, 3> out(Eigen::array<Eigen::DenseIndex, 3>(72,53,97)); in1.setRandom(); in2.setRandom(); in3.setRandom(); @@ -98,10 +144,10 @@ void test_cuda_elementwise() Eigen::CudaStreamDevice stream; Eigen::GpuDevice gpu_device(&stream); - Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_in1(d_in1, Eigen::array<int, 3>(72,53,97)); - Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_in2(d_in2, Eigen::array<int, 3>(72,53,97)); - Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_in3(d_in3, Eigen::array<int, 3>(72,53,97)); - Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_out(d_out, Eigen::array<int, 3>(72,53,97)); + Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_in1(d_in1, Eigen::array<Eigen::DenseIndex, 3>(72,53,97)); + Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_in2(d_in2, Eigen::array<Eigen::DenseIndex, 3>(72,53,97)); + Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_in3(d_in3, Eigen::array<Eigen::DenseIndex, 3>(72,53,97)); + Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_out(d_out, Eigen::array<Eigen::DenseIndex, 3>(72,53,97)); gpu_out.device(gpu_device) = gpu_in1 + gpu_in2 * gpu_in3; @@ -111,7 +157,7 @@ void test_cuda_elementwise() for (int i = 0; i < 72; ++i) { for (int j = 0; j < 53; ++j) { for (int k = 0; k < 97; ++k) { - VERIFY_IS_APPROX(out(Eigen::array<int, 3>(i,j,k)), in1(Eigen::array<int, 3>(i,j,k)) + in2(Eigen::array<int, 3>(i,j,k)) * in3(Eigen::array<int, 3>(i,j,k))); + VERIFY_IS_APPROX(out(Eigen::array<Eigen::DenseIndex, 3>(i,j,k)), in1(Eigen::array<Eigen::DenseIndex, 3>(i,j,k)) + in2(Eigen::array<Eigen::DenseIndex, 3>(i,j,k)) * in3(Eigen::array<Eigen::DenseIndex, 3>(i,j,k))); } } } @@ -181,7 +227,7 @@ void test_cuda_reduction() Eigen::TensorMap<Eigen::Tensor<float, 4> > gpu_in1(d_in1, 72,53,97,113); Eigen::TensorMap<Eigen::Tensor<float, 2> > gpu_out(d_out, 72,97); - array<int, 2> reduction_axis; + array<Eigen::DenseIndex, 2> reduction_axis; reduction_axis[0] = 1; reduction_axis[1] = 3; @@ -214,8 +260,8 @@ void test_cuda_contraction() // more than 30 * 1024, which is the number of threads in blocks on // a 15 SM GK110 GPU Tensor<float, 4, DataLayout> t_left(6, 50, 3, 31); - Tensor<float, 5, DataLayout> t_right(Eigen::array<int, 5>(3, 31, 7, 20, 1)); - Tensor<float, 5, DataLayout> t_result(Eigen::array<int, 5>(6, 50, 7, 20, 1)); + Tensor<float, 5, DataLayout> t_right(Eigen::array<Eigen::DenseIndex, 5>(3, 31, 7, 20, 1)); + Tensor<float, 5, DataLayout> t_result(Eigen::array<Eigen::DenseIndex, 5>(6, 50, 7, 20, 1)); t_left.setRandom(); t_right.setRandom(); @@ -299,7 +345,7 @@ void test_cuda_convolution_1d() Eigen::TensorMap<Eigen::Tensor<float, 1, DataLayout> > gpu_kernel(d_kernel, 4); Eigen::TensorMap<Eigen::Tensor<float, 4, DataLayout> > gpu_out(d_out, 74,34,11,137); - Eigen::array<int, 1> dims(1); + Eigen::array<Eigen::DenseIndex, 1> dims(1); gpu_out.device(gpu_device) = gpu_input.convolve(gpu_kernel, dims); assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess); @@ -352,7 +398,7 @@ void test_cuda_convolution_inner_dim_col_major_1d() Eigen::TensorMap<Eigen::Tensor<float, 1, ColMajor> > gpu_kernel(d_kernel,4); Eigen::TensorMap<Eigen::Tensor<float, 4, ColMajor> > gpu_out(d_out,71,9,11,7); - Eigen::array<int, 1> dims(0); + Eigen::array<Eigen::DenseIndex, 1> dims(0); gpu_out.device(gpu_device) = gpu_input.convolve(gpu_kernel, dims); assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess); @@ -405,7 +451,7 @@ void test_cuda_convolution_inner_dim_row_major_1d() Eigen::TensorMap<Eigen::Tensor<float, 1, RowMajor> > gpu_kernel(d_kernel, 4); Eigen::TensorMap<Eigen::Tensor<float, 4, RowMajor> > gpu_out(d_out, 7,9,11,71); - Eigen::array<int, 1> dims(3); + Eigen::array<Eigen::DenseIndex, 1> dims(3); gpu_out.device(gpu_device) = gpu_input.convolve(gpu_kernel, dims); assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess); @@ -459,7 +505,7 @@ void test_cuda_convolution_2d() Eigen::TensorMap<Eigen::Tensor<float, 2, DataLayout> > gpu_kernel(d_kernel,3,4); Eigen::TensorMap<Eigen::Tensor<float, 4, DataLayout> > gpu_out(d_out,74,35,8,137); - Eigen::array<int, 2> dims(1,2); + Eigen::array<Eigen::DenseIndex, 2> dims(1,2); gpu_out.device(gpu_device) = gpu_input.convolve(gpu_kernel, dims); assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess); @@ -496,9 +542,9 @@ void test_cuda_convolution_2d() template<int DataLayout> void test_cuda_convolution_3d() { - Tensor<float, 5, DataLayout> input(Eigen::array<int, 5>(74,37,11,137,17)); + Tensor<float, 5, DataLayout> input(Eigen::array<Eigen::DenseIndex, 5>(74,37,11,137,17)); Tensor<float, 3, DataLayout> kernel(3,4,2); - Tensor<float, 5, DataLayout> out(Eigen::array<int, 5>(74,35,8,136,17)); + Tensor<float, 5, DataLayout> out(Eigen::array<Eigen::DenseIndex, 5>(74,35,8,136,17)); input = input.constant(10.0f) + input.random(); kernel = kernel.constant(7.0f) + kernel.random(); @@ -523,7 +569,7 @@ void test_cuda_convolution_3d() Eigen::TensorMap<Eigen::Tensor<float, 3, DataLayout> > gpu_kernel(d_kernel,3,4,2); Eigen::TensorMap<Eigen::Tensor<float, 5, DataLayout> > gpu_out(d_out,74,35,8,136,17); - Eigen::array<int, 3> dims(1,2,3); + Eigen::array<Eigen::DenseIndex, 3> dims(1,2,3); gpu_out.device(gpu_device) = gpu_input.convolve(gpu_kernel, dims); assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess); @@ -1168,6 +1214,7 @@ void test_cuda_betainc() void test_cxx11_tensor_cuda() { + CALL_SUBTEST_1(test_cuda_nullary()); CALL_SUBTEST_1(test_cuda_elementwise_small()); CALL_SUBTEST_1(test_cuda_elementwise()); CALL_SUBTEST_1(test_cuda_props()); diff --git a/unsupported/test/cxx11_tensor_device.cu b/unsupported/test/cxx11_tensor_device.cu index b6ca54d93..fde20ddf2 100644 --- a/unsupported/test/cxx11_tensor_device.cu +++ b/unsupported/test/cxx11_tensor_device.cu @@ -13,7 +13,9 @@ #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int #define EIGEN_USE_GPU - +#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500 +#include <cuda_fp16.h> +#endif #include "main.h" #include <unsupported/Eigen/CXX11/Tensor> diff --git a/unsupported/test/cxx11_tensor_of_float16_cuda.cu b/unsupported/test/cxx11_tensor_of_float16_cuda.cu index 2f55f9361..cbf401c86 100644 --- a/unsupported/test/cxx11_tensor_of_float16_cuda.cu +++ b/unsupported/test/cxx11_tensor_of_float16_cuda.cu @@ -13,7 +13,9 @@ #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int #define EIGEN_USE_GPU - +#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500 +#include <cuda_fp16.h> +#endif #include "main.h" #include <unsupported/Eigen/CXX11/Tensor> @@ -181,30 +183,39 @@ void test_cuda_trancendental() { float* d_float1 = (float*)gpu_device.allocate(num_elem * sizeof(float)); float* d_float2 = (float*)gpu_device.allocate(num_elem * sizeof(float)); + float* d_float3 = (float*)gpu_device.allocate(num_elem * sizeof(float)); Eigen::half* d_res1_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half)); Eigen::half* d_res1_float = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half)); Eigen::half* d_res2_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half)); Eigen::half* d_res2_float = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half)); - - Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float1( - d_float1, num_elem); - Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float2( - d_float2, num_elem); - Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res1_half( - d_res1_half, num_elem); - Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res1_float( - d_res1_float, num_elem); - Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res2_half( - d_res2_half, num_elem); - Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res2_float( - d_res2_float, num_elem); + Eigen::half* d_res3_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half)); + Eigen::half* d_res3_float = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half)); + + Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float1(d_float1, num_elem); + Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float2(d_float2, num_elem); + Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float3(d_float3, num_elem); + Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res1_half(d_res1_half, num_elem); + Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res1_float(d_res1_float, num_elem); + Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res2_half(d_res2_half, num_elem); + Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res2_float(d_res2_float, num_elem); + Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res3_half(d_res3_half, num_elem); + Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res3_float(d_res3_float, num_elem); gpu_float1.device(gpu_device) = gpu_float1.random() - gpu_float1.constant(0.5f); gpu_float2.device(gpu_device) = gpu_float2.random() + gpu_float1.constant(0.5f); + gpu_float3.device(gpu_device) = gpu_float3.random(); gpu_res1_float.device(gpu_device) = gpu_float1.exp().cast<Eigen::half>(); gpu_res2_float.device(gpu_device) = gpu_float2.log().cast<Eigen::half>(); - gpu_res1_half.device(gpu_device) = gpu_float1.cast<Eigen::half>().exp(); - gpu_res2_half.device(gpu_device) = gpu_float2.cast<Eigen::half>().log(); + gpu_res3_float.device(gpu_device) = gpu_float3.log1p().cast<Eigen::half>(); + + gpu_res1_half.device(gpu_device) = gpu_float1.cast<Eigen::half>(); + gpu_res1_half.device(gpu_device) = gpu_res1_half.exp(); + + gpu_res2_half.device(gpu_device) = gpu_float2.cast<Eigen::half>(); + gpu_res2_half.device(gpu_device) = gpu_res2_half.log(); + + gpu_res3_half.device(gpu_device) = gpu_float3.cast<Eigen::half>(); + gpu_res3_half.device(gpu_device) = gpu_res3_half.log1p(); Tensor<float, 1> input1(num_elem); Tensor<Eigen::half, 1> half_prec1(num_elem); @@ -212,12 +223,18 @@ void test_cuda_trancendental() { Tensor<float, 1> input2(num_elem); Tensor<Eigen::half, 1> half_prec2(num_elem); Tensor<Eigen::half, 1> full_prec2(num_elem); + Tensor<float, 1> input3(num_elem); + Tensor<Eigen::half, 1> half_prec3(num_elem); + Tensor<Eigen::half, 1> full_prec3(num_elem); gpu_device.memcpyDeviceToHost(input1.data(), d_float1, num_elem*sizeof(float)); gpu_device.memcpyDeviceToHost(input2.data(), d_float2, num_elem*sizeof(float)); + gpu_device.memcpyDeviceToHost(input3.data(), d_float3, num_elem*sizeof(float)); gpu_device.memcpyDeviceToHost(half_prec1.data(), d_res1_half, num_elem*sizeof(Eigen::half)); gpu_device.memcpyDeviceToHost(full_prec1.data(), d_res1_float, num_elem*sizeof(Eigen::half)); gpu_device.memcpyDeviceToHost(half_prec2.data(), d_res2_half, num_elem*sizeof(Eigen::half)); gpu_device.memcpyDeviceToHost(full_prec2.data(), d_res2_float, num_elem*sizeof(Eigen::half)); + gpu_device.memcpyDeviceToHost(half_prec3.data(), d_res3_half, num_elem*sizeof(Eigen::half)); + gpu_device.memcpyDeviceToHost(full_prec3.data(), d_res3_float, num_elem*sizeof(Eigen::half)); gpu_device.synchronize(); for (int i = 0; i < num_elem; ++i) { @@ -231,12 +248,19 @@ void test_cuda_trancendental() { else VERIFY_IS_APPROX(full_prec2(i), half_prec2(i)); } + for (int i = 0; i < num_elem; ++i) { + std::cout << "Checking elemwise plog1 " << i << " input = " << input3(i) << " full = " << full_prec3(i) << " half = " << half_prec3(i) << std::endl; + VERIFY_IS_APPROX(full_prec3(i), half_prec3(i)); + } gpu_device.deallocate(d_float1); gpu_device.deallocate(d_float2); + gpu_device.deallocate(d_float3); gpu_device.deallocate(d_res1_half); gpu_device.deallocate(d_res1_float); gpu_device.deallocate(d_res2_half); gpu_device.deallocate(d_res2_float); + gpu_device.deallocate(d_res3_float); + gpu_device.deallocate(d_res3_half); } template<typename> diff --git a/unsupported/test/cxx11_tensor_random_cuda.cu b/unsupported/test/cxx11_tensor_random_cuda.cu index fa1a46732..b3be199e1 100644 --- a/unsupported/test/cxx11_tensor_random_cuda.cu +++ b/unsupported/test/cxx11_tensor_random_cuda.cu @@ -13,6 +13,9 @@ #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int #define EIGEN_USE_GPU +#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500 +#include <cuda_fp16.h> +#endif #include "main.h" #include <Eigen/CXX11/Tensor> diff --git a/unsupported/test/cxx11_tensor_reduction_cuda.cu b/unsupported/test/cxx11_tensor_reduction_cuda.cu index cad0c08e0..6858b43a7 100644 --- a/unsupported/test/cxx11_tensor_reduction_cuda.cu +++ b/unsupported/test/cxx11_tensor_reduction_cuda.cu @@ -12,11 +12,14 @@ #define EIGEN_TEST_FUNC cxx11_tensor_reduction_cuda #define EIGEN_USE_GPU +#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500 +#include <cuda_fp16.h> +#endif #include "main.h" #include <unsupported/Eigen/CXX11/Tensor> -template<int DataLayout> +template<typename Type, int DataLayout> static void test_full_reductions() { Eigen::CudaStreamDevice stream; @@ -25,24 +28,24 @@ static void test_full_reductions() { const int num_rows = internal::random<int>(1024, 5*1024); const int num_cols = internal::random<int>(1024, 5*1024); - Tensor<float, 2, DataLayout> in(num_rows, num_cols); + Tensor<Type, 2, DataLayout> in(num_rows, num_cols); in.setRandom(); - Tensor<float, 0, DataLayout> full_redux; + Tensor<Type, 0, DataLayout> full_redux; full_redux = in.sum(); - std::size_t in_bytes = in.size() * sizeof(float); - std::size_t out_bytes = full_redux.size() * sizeof(float); - float* gpu_in_ptr = static_cast<float*>(gpu_device.allocate(in_bytes)); - float* gpu_out_ptr = static_cast<float*>(gpu_device.allocate(out_bytes)); + std::size_t in_bytes = in.size() * sizeof(Type); + std::size_t out_bytes = full_redux.size() * sizeof(Type); + Type* gpu_in_ptr = static_cast<Type*>(gpu_device.allocate(in_bytes)); + Type* gpu_out_ptr = static_cast<Type*>(gpu_device.allocate(out_bytes)); gpu_device.memcpyHostToDevice(gpu_in_ptr, in.data(), in_bytes); - TensorMap<Tensor<float, 2, DataLayout> > in_gpu(gpu_in_ptr, num_rows, num_cols); - TensorMap<Tensor<float, 0, DataLayout> > out_gpu(gpu_out_ptr); + TensorMap<Tensor<Type, 2, DataLayout> > in_gpu(gpu_in_ptr, num_rows, num_cols); + TensorMap<Tensor<Type, 0, DataLayout> > out_gpu(gpu_out_ptr); out_gpu.device(gpu_device) = in_gpu.sum(); - Tensor<float, 0, DataLayout> full_redux_gpu; + Tensor<Type, 0, DataLayout> full_redux_gpu; gpu_device.memcpyDeviceToHost(full_redux_gpu.data(), gpu_out_ptr, out_bytes); gpu_device.synchronize(); @@ -53,7 +56,102 @@ static void test_full_reductions() { gpu_device.deallocate(gpu_out_ptr); } +template<typename Type, int DataLayout> +static void test_first_dim_reductions() { + int dim_x = 33; + int dim_y = 1; + int dim_z = 128; + + Tensor<Type, 3, DataLayout> in(dim_x, dim_y, dim_z); + in.setRandom(); + + Eigen::array<int, 1> red_axis; + red_axis[0] = 0; + Tensor<Type, 2, DataLayout> redux = in.sum(red_axis); + + // Create device + Eigen::CudaStreamDevice stream; + Eigen::GpuDevice dev(&stream); + + // Create data(T) + Type* in_data = (Type*)dev.allocate(dim_x*dim_y*dim_z*sizeof(Type)); + Type* out_data = (Type*)dev.allocate(dim_z*dim_y*sizeof(Type)); + Eigen::TensorMap<Eigen::Tensor<Type, 3, DataLayout> > gpu_in(in_data, dim_x, dim_y, dim_z); + Eigen::TensorMap<Eigen::Tensor<Type, 2, DataLayout> > gpu_out(out_data, dim_y, dim_z); + + // Perform operation + dev.memcpyHostToDevice(in_data, in.data(), in.size()*sizeof(Type)); + gpu_out.device(dev) = gpu_in.sum(red_axis); + gpu_out.device(dev) += gpu_in.sum(red_axis); + Tensor<Type, 2, DataLayout> redux_gpu(dim_y, dim_z); + dev.memcpyDeviceToHost(redux_gpu.data(), out_data, gpu_out.size()*sizeof(Type)); + dev.synchronize(); + + // Check that the CPU and GPU reductions return the same result. + for (int i = 0; i < gpu_out.size(); ++i) { + VERIFY_IS_APPROX(2*redux(i), redux_gpu(i)); + } + + dev.deallocate(in_data); + dev.deallocate(out_data); +} + +template<typename Type, int DataLayout> +static void test_last_dim_reductions() { + int dim_x = 128; + int dim_y = 1; + int dim_z = 33; + + Tensor<Type, 3, DataLayout> in(dim_x, dim_y, dim_z); + in.setRandom(); + + Eigen::array<int, 1> red_axis; + red_axis[0] = 2; + Tensor<Type, 2, DataLayout> redux = in.sum(red_axis); + + // Create device + Eigen::CudaStreamDevice stream; + Eigen::GpuDevice dev(&stream); + + // Create data + Type* in_data = (Type*)dev.allocate(dim_x*dim_y*dim_z*sizeof(Type)); + Type* out_data = (Type*)dev.allocate(dim_x*dim_y*sizeof(Type)); + Eigen::TensorMap<Eigen::Tensor<Type, 3, DataLayout> > gpu_in(in_data, dim_x, dim_y, dim_z); + Eigen::TensorMap<Eigen::Tensor<Type, 2, DataLayout> > gpu_out(out_data, dim_x, dim_y); + + // Perform operation + dev.memcpyHostToDevice(in_data, in.data(), in.size()*sizeof(Type)); + gpu_out.device(dev) = gpu_in.sum(red_axis); + gpu_out.device(dev) += gpu_in.sum(red_axis); + Tensor<Type, 2, DataLayout> redux_gpu(dim_x, dim_y); + dev.memcpyDeviceToHost(redux_gpu.data(), out_data, gpu_out.size()*sizeof(Type)); + dev.synchronize(); + + // Check that the CPU and GPU reductions return the same result. + for (int i = 0; i < gpu_out.size(); ++i) { + VERIFY_IS_APPROX(2*redux(i), redux_gpu(i)); + } + + dev.deallocate(in_data); + dev.deallocate(out_data); +} + + void test_cxx11_tensor_reduction_cuda() { - CALL_SUBTEST_1(test_full_reductions<ColMajor>()); - CALL_SUBTEST_2(test_full_reductions<RowMajor>()); + CALL_SUBTEST_1((test_full_reductions<float, ColMajor>())); + CALL_SUBTEST_1((test_full_reductions<double, ColMajor>())); + CALL_SUBTEST_2((test_full_reductions<float, RowMajor>())); + CALL_SUBTEST_2((test_full_reductions<double, RowMajor>())); + + CALL_SUBTEST_3((test_first_dim_reductions<float, ColMajor>())); + CALL_SUBTEST_3((test_first_dim_reductions<double, ColMajor>())); + CALL_SUBTEST_4((test_first_dim_reductions<float, RowMajor>())); +// Outer reductions of doubles aren't supported just yet. +// CALL_SUBTEST_4((test_first_dim_reductions<double, RowMajor>())) + + CALL_SUBTEST_5((test_last_dim_reductions<float, ColMajor>())); +// Outer reductions of doubles aren't supported just yet. +// CALL_SUBTEST_5((test_last_dim_reductions<double, ColMajor>())); + CALL_SUBTEST_6((test_last_dim_reductions<float, RowMajor>())); + CALL_SUBTEST_6((test_last_dim_reductions<double, RowMajor>())); } diff --git a/unsupported/test/cxx11_tensor_scan_cuda.cu b/unsupported/test/cxx11_tensor_scan_cuda.cu index 35e19e51c..761d11fd1 100644 --- a/unsupported/test/cxx11_tensor_scan_cuda.cu +++ b/unsupported/test/cxx11_tensor_scan_cuda.cu @@ -13,7 +13,9 @@ #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int #define EIGEN_USE_GPU - +#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500 +#include <cuda_fp16.h> +#endif #include "main.h" #include <unsupported/Eigen/CXX11/Tensor> diff --git a/unsupported/test/kronecker_product.cpp b/unsupported/test/kronecker_product.cpp index 02411a262..e770049e5 100644 --- a/unsupported/test/kronecker_product.cpp +++ b/unsupported/test/kronecker_product.cpp @@ -9,12 +9,12 @@ // 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/. +#ifdef EIGEN_TEST_PART_1 #include "sparse.h" #include <Eigen/SparseExtra> #include <Eigen/KroneckerProduct> - template<typename MatrixType> void check_dimension(const MatrixType& ab, const int rows, const int cols) { @@ -230,3 +230,23 @@ void test_kronecker_product() VERIFY_IS_APPROX(MatrixXf(sC2),dC); } } + +#endif + +#ifdef EIGEN_TEST_PART_2 + +// simply check that for a dense kronecker product, sparse module is not needed + +#include "main.h" +#include <Eigen/KroneckerProduct> + +void test_kronecker_product() +{ + MatrixXd a(2,2), b(3,3), c; + a.setRandom(); + b.setRandom(); + c = kroneckerProduct(a,b); + VERIFY_IS_APPROX(c.block(3,3,3,3), a(1,1)*b); +} + +#endif |