diff options
Diffstat (limited to 'extern/Eigen3/Eigen/src/Core/util/Memory.h')
-rw-r--r-- | extern/Eigen3/Eigen/src/Core/util/Memory.h | 580 |
1 files changed, 298 insertions, 282 deletions
diff --git a/extern/Eigen3/Eigen/src/Core/util/Memory.h b/extern/Eigen3/Eigen/src/Core/util/Memory.h index b9af5cf8c7b..291383c581a 100644 --- a/extern/Eigen3/Eigen/src/Core/util/Memory.h +++ b/extern/Eigen3/Eigen/src/Core/util/Memory.h @@ -1,11 +1,12 @@ // 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) 2008-2015 Gael Guennebaud <gael.guennebaud@inria.fr> // Copyright (C) 2008-2009 Benoit Jacob <jacob.benoit.1@gmail.com> // Copyright (C) 2009 Kenneth Riddile <kfriddile@yahoo.com> // Copyright (C) 2010 Hauke Heibel <hauke.heibel@gmail.com> // Copyright (C) 2010 Thomas Capricelli <orzel@freehackers.org> +// Copyright (C) 2013 Pavel Holoborodko <pavel@holoborodko.com> // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -31,7 +32,7 @@ // page 114, "[The] LP64 model [...] is used by all 64-bit UNIX ports" so it's indeed // quite safe, at least within the context of glibc, to equate 64-bit with LP64. #if defined(__GLIBC__) && ((__GLIBC__>=2 && __GLIBC_MINOR__ >= 8) || __GLIBC__>2) \ - && defined(__LP64__) && ! defined( __SANITIZE_ADDRESS__ ) + && defined(__LP64__) && ! defined( __SANITIZE_ADDRESS__ ) && (EIGEN_DEFAULT_ALIGN_BYTES == 16) #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 1 #else #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 0 @@ -41,15 +42,15 @@ // See http://svn.freebsd.org/viewvc/base/stable/6/lib/libc/stdlib/malloc.c?view=markup // FreeBSD 7 seems to have 16-byte aligned malloc except on ARM and MIPS architectures // See http://svn.freebsd.org/viewvc/base/stable/7/lib/libc/stdlib/malloc.c?view=markup -#if defined(__FreeBSD__) && !defined(__arm__) && !defined(__mips__) +#if defined(__FreeBSD__) && !(EIGEN_ARCH_ARM || EIGEN_ARCH_MIPS) && (EIGEN_DEFAULT_ALIGN_BYTES == 16) #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 1 #else #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 0 #endif -#if defined(__APPLE__) \ - || defined(_WIN64) \ - || EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED \ +#if (EIGEN_OS_MAC && (EIGEN_DEFAULT_ALIGN_BYTES == 16)) \ + || (EIGEN_OS_WIN64 && (EIGEN_DEFAULT_ALIGN_BYTES == 16)) \ + || EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED \ || EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED #define EIGEN_MALLOC_ALREADY_ALIGNED 1 #else @@ -58,37 +59,18 @@ #endif -// See bug 554 (http://eigen.tuxfamily.org/bz/show_bug.cgi?id=554) -// It seems to be unsafe to check _POSIX_ADVISORY_INFO without including unistd.h first. -// Currently, let's include it only on unix systems: -#if defined(__unix__) || defined(__unix) - #include <unistd.h> - #if ((defined __QNXNTO__) || (defined _GNU_SOURCE) || (defined __PGI) || ((defined _XOPEN_SOURCE) && (_XOPEN_SOURCE >= 600))) && (defined _POSIX_ADVISORY_INFO) && (_POSIX_ADVISORY_INFO > 0) - #define EIGEN_HAS_POSIX_MEMALIGN 1 - #endif -#endif - -#ifndef EIGEN_HAS_POSIX_MEMALIGN - #define EIGEN_HAS_POSIX_MEMALIGN 0 -#endif - -#ifdef EIGEN_VECTORIZE_SSE - #define EIGEN_HAS_MM_MALLOC 1 -#else - #define EIGEN_HAS_MM_MALLOC 0 -#endif - namespace Eigen { namespace internal { +EIGEN_DEVICE_FUNC inline void throw_std_bad_alloc() { #ifdef EIGEN_EXCEPTIONS throw std::bad_alloc(); #else - std::size_t huge = -1; - new int[huge]; + std::size_t huge = static_cast<std::size_t>(-1); + ::operator new(huge); #endif } @@ -103,9 +85,9 @@ inline void throw_std_bad_alloc() */ inline void* handmade_aligned_malloc(std::size_t size) { - void *original = std::malloc(size+16); + void *original = std::malloc(size+EIGEN_DEFAULT_ALIGN_BYTES); if (original == 0) return 0; - void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(15))) + 16); + void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1))) + EIGEN_DEFAULT_ALIGN_BYTES); *(reinterpret_cast<void**>(aligned) - 1) = original; return aligned; } @@ -118,7 +100,7 @@ inline void handmade_aligned_free(void *ptr) /** \internal * \brief Reallocates aligned memory. - * Since we know that our handmade version is based on std::realloc + * Since we know that our handmade version is based on std::malloc * we can use std::realloc to implement efficient reallocation. */ inline void* handmade_aligned_realloc(void* ptr, std::size_t size, std::size_t = 0) @@ -126,9 +108,9 @@ inline void* handmade_aligned_realloc(void* ptr, std::size_t size, std::size_t = if (ptr == 0) return handmade_aligned_malloc(size); void *original = *(reinterpret_cast<void**>(ptr) - 1); std::ptrdiff_t previous_offset = static_cast<char *>(ptr)-static_cast<char *>(original); - original = std::realloc(original,size+16); + original = std::realloc(original,size+EIGEN_DEFAULT_ALIGN_BYTES); if (original == 0) return 0; - void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(15))) + 16); + void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1))) + EIGEN_DEFAULT_ALIGN_BYTES); void *previous_aligned = static_cast<char *>(original)+previous_offset; if(aligned!=previous_aligned) std::memmove(aligned, previous_aligned, size); @@ -138,92 +120,46 @@ inline void* handmade_aligned_realloc(void* ptr, std::size_t size, std::size_t = } /***************************************************************************** -*** Implementation of generic aligned realloc (when no realloc can be used)*** -*****************************************************************************/ - -void* aligned_malloc(std::size_t size); -void aligned_free(void *ptr); - -/** \internal - * \brief Reallocates aligned memory. - * Allows reallocation with aligned ptr types. This implementation will - * always create a new memory chunk and copy the old data. - */ -inline void* generic_aligned_realloc(void* ptr, size_t size, size_t old_size) -{ - if (ptr==0) - return aligned_malloc(size); - - if (size==0) - { - aligned_free(ptr); - return 0; - } - - void* newptr = aligned_malloc(size); - if (newptr == 0) - { - #ifdef EIGEN_HAS_ERRNO - errno = ENOMEM; // according to the standard - #endif - return 0; - } - - if (ptr != 0) - { - std::memcpy(newptr, ptr, (std::min)(size,old_size)); - aligned_free(ptr); - } - - return newptr; -} - -/***************************************************************************** *** Implementation of portable aligned versions of malloc/free/realloc *** *****************************************************************************/ #ifdef EIGEN_NO_MALLOC -inline void check_that_malloc_is_allowed() +EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed() { eigen_assert(false && "heap allocation is forbidden (EIGEN_NO_MALLOC is defined)"); } #elif defined EIGEN_RUNTIME_NO_MALLOC -inline bool is_malloc_allowed_impl(bool update, bool new_value = false) +EIGEN_DEVICE_FUNC inline bool is_malloc_allowed_impl(bool update, bool new_value = false) { static bool value = true; if (update == 1) value = new_value; return value; } -inline bool is_malloc_allowed() { return is_malloc_allowed_impl(false); } -inline bool set_is_malloc_allowed(bool new_value) { return is_malloc_allowed_impl(true, new_value); } -inline void check_that_malloc_is_allowed() +EIGEN_DEVICE_FUNC inline bool is_malloc_allowed() { return is_malloc_allowed_impl(false); } +EIGEN_DEVICE_FUNC inline bool set_is_malloc_allowed(bool new_value) { return is_malloc_allowed_impl(true, new_value); } +EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed() { eigen_assert(is_malloc_allowed() && "heap allocation is forbidden (EIGEN_RUNTIME_NO_MALLOC is defined and g_is_malloc_allowed is false)"); } #else -inline void check_that_malloc_is_allowed() +EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed() {} #endif -/** \internal Allocates \a size bytes. The returned pointer is guaranteed to have 16 bytes alignment. +/** \internal Allocates \a size bytes. The returned pointer is guaranteed to have 16 or 32 bytes alignment depending on the requirements. * On allocation error, the returned pointer is null, and std::bad_alloc is thrown. */ -inline void* aligned_malloc(size_t size) +EIGEN_DEVICE_FUNC inline void* aligned_malloc(std::size_t size) { check_that_malloc_is_allowed(); void *result; - #if !EIGEN_ALIGN - result = std::malloc(size); - #elif EIGEN_MALLOC_ALREADY_ALIGNED + #if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED result = std::malloc(size); - #elif EIGEN_HAS_POSIX_MEMALIGN - if(posix_memalign(&result, 16, size)) result = 0; - #elif EIGEN_HAS_MM_MALLOC - result = _mm_malloc(size, 16); - #elif defined(_MSC_VER) && (!defined(_WIN32_WCE)) - result = _aligned_malloc(size, 16); + #if EIGEN_DEFAULT_ALIGN_BYTES==16 + eigen_assert((size<16 || (std::size_t(result)%16)==0) && "System's malloc returned an unaligned pointer. Compile with EIGEN_MALLOC_ALREADY_ALIGNED=0 to fallback to handmade alignd memory allocator."); + #endif #else result = handmade_aligned_malloc(size); #endif @@ -235,50 +171,27 @@ inline void* aligned_malloc(size_t size) } /** \internal Frees memory allocated with aligned_malloc. */ -inline void aligned_free(void *ptr) +EIGEN_DEVICE_FUNC inline void aligned_free(void *ptr) { - #if !EIGEN_ALIGN - std::free(ptr); - #elif EIGEN_MALLOC_ALREADY_ALIGNED - std::free(ptr); - #elif EIGEN_HAS_POSIX_MEMALIGN + #if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED std::free(ptr); - #elif EIGEN_HAS_MM_MALLOC - _mm_free(ptr); - #elif defined(_MSC_VER) && (!defined(_WIN32_WCE)) - _aligned_free(ptr); #else handmade_aligned_free(ptr); #endif } /** -* \internal -* \brief Reallocates an aligned block of memory. -* \throws std::bad_alloc on allocation failure -**/ -inline void* aligned_realloc(void *ptr, size_t new_size, size_t old_size) + * \internal + * \brief Reallocates an aligned block of memory. + * \throws std::bad_alloc on allocation failure + */ +inline void* aligned_realloc(void *ptr, std::size_t new_size, std::size_t old_size) { EIGEN_UNUSED_VARIABLE(old_size); void *result; -#if !EIGEN_ALIGN - result = std::realloc(ptr,new_size); -#elif EIGEN_MALLOC_ALREADY_ALIGNED +#if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED result = std::realloc(ptr,new_size); -#elif EIGEN_HAS_POSIX_MEMALIGN - result = generic_aligned_realloc(ptr,new_size,old_size); -#elif EIGEN_HAS_MM_MALLOC - // The defined(_mm_free) is just here to verify that this MSVC version - // implements _mm_malloc/_mm_free based on the corresponding _aligned_ - // functions. This may not always be the case and we just try to be safe. - #if defined(_MSC_VER) && (!defined(_WIN32_WCE)) && defined(_mm_free) - result = _aligned_realloc(ptr,new_size,16); - #else - result = generic_aligned_realloc(ptr,new_size,old_size); - #endif -#elif defined(_MSC_VER) && (!defined(_WIN32_WCE)) - result = _aligned_realloc(ptr,new_size,16); #else result = handmade_aligned_realloc(ptr,new_size,old_size); #endif @@ -296,12 +209,12 @@ inline void* aligned_realloc(void *ptr, size_t new_size, size_t old_size) /** \internal Allocates \a size bytes. If Align is true, then the returned ptr is 16-byte-aligned. * On allocation error, the returned pointer is null, and a std::bad_alloc is thrown. */ -template<bool Align> inline void* conditional_aligned_malloc(size_t size) +template<bool Align> EIGEN_DEVICE_FUNC inline void* conditional_aligned_malloc(std::size_t size) { return aligned_malloc(size); } -template<> inline void* conditional_aligned_malloc<false>(size_t size) +template<> EIGEN_DEVICE_FUNC inline void* conditional_aligned_malloc<false>(std::size_t size) { check_that_malloc_is_allowed(); @@ -312,22 +225,22 @@ template<> inline void* conditional_aligned_malloc<false>(size_t size) } /** \internal Frees memory allocated with conditional_aligned_malloc */ -template<bool Align> inline void conditional_aligned_free(void *ptr) +template<bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_free(void *ptr) { aligned_free(ptr); } -template<> inline void conditional_aligned_free<false>(void *ptr) +template<> EIGEN_DEVICE_FUNC inline void conditional_aligned_free<false>(void *ptr) { std::free(ptr); } -template<bool Align> inline void* conditional_aligned_realloc(void* ptr, size_t new_size, size_t old_size) +template<bool Align> inline void* conditional_aligned_realloc(void* ptr, std::size_t new_size, std::size_t old_size) { return aligned_realloc(ptr, new_size, old_size); } -template<> inline void* conditional_aligned_realloc<false>(void* ptr, size_t new_size, size_t) +template<> inline void* conditional_aligned_realloc<false>(void* ptr, std::size_t new_size, std::size_t) { return std::realloc(ptr, new_size); } @@ -336,33 +249,43 @@ template<> inline void* conditional_aligned_realloc<false>(void* ptr, size_t new *** Construction/destruction of array elements *** *****************************************************************************/ -/** \internal Constructs the elements of an array. - * The \a size parameter tells on how many objects to call the constructor of T. - */ -template<typename T> inline T* construct_elements_of_array(T *ptr, size_t size) -{ - for (size_t i=0; i < size; ++i) ::new (ptr + i) T; - return ptr; -} - /** \internal Destructs the elements of an array. * The \a size parameters tells on how many objects to call the destructor of T. */ -template<typename T> inline void destruct_elements_of_array(T *ptr, size_t size) +template<typename T> EIGEN_DEVICE_FUNC inline void destruct_elements_of_array(T *ptr, std::size_t size) { // always destruct an array starting from the end. if(ptr) while(size) ptr[--size].~T(); } +/** \internal Constructs the elements of an array. + * The \a size parameter tells on how many objects to call the constructor of T. + */ +template<typename T> EIGEN_DEVICE_FUNC inline T* construct_elements_of_array(T *ptr, std::size_t size) +{ + std::size_t i; + EIGEN_TRY + { + for (i = 0; i < size; ++i) ::new (ptr + i) T; + return ptr; + } + EIGEN_CATCH(...) + { + destruct_elements_of_array(ptr, i); + EIGEN_THROW; + } + return NULL; +} + /***************************************************************************** *** Implementation of aligned new/delete-like functions *** *****************************************************************************/ template<typename T> -EIGEN_ALWAYS_INLINE void check_size_for_overflow(size_t size) +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void check_size_for_overflow(std::size_t size) { - if(size > size_t(-1) / sizeof(T)) + if(size > std::size_t(-1) / sizeof(T)) throw_std_bad_alloc(); } @@ -370,24 +293,42 @@ EIGEN_ALWAYS_INLINE void check_size_for_overflow(size_t size) * On allocation error, the returned pointer is undefined, but a std::bad_alloc is thrown. * The default constructor of T is called. */ -template<typename T> inline T* aligned_new(size_t size) +template<typename T> EIGEN_DEVICE_FUNC inline T* aligned_new(std::size_t size) { check_size_for_overflow<T>(size); T *result = reinterpret_cast<T*>(aligned_malloc(sizeof(T)*size)); - return construct_elements_of_array(result, size); + EIGEN_TRY + { + return construct_elements_of_array(result, size); + } + EIGEN_CATCH(...) + { + aligned_free(result); + EIGEN_THROW; + } + return result; } -template<typename T, bool Align> inline T* conditional_aligned_new(size_t size) +template<typename T, bool Align> EIGEN_DEVICE_FUNC inline T* conditional_aligned_new(std::size_t size) { check_size_for_overflow<T>(size); T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size)); - return construct_elements_of_array(result, size); + EIGEN_TRY + { + return construct_elements_of_array(result, size); + } + EIGEN_CATCH(...) + { + conditional_aligned_free<Align>(result); + EIGEN_THROW; + } + return result; } /** \internal Deletes objects constructed with aligned_new * The \a size parameters tells on how many objects to call the destructor of T. */ -template<typename T> inline void aligned_delete(T *ptr, size_t size) +template<typename T> EIGEN_DEVICE_FUNC inline void aligned_delete(T *ptr, std::size_t size) { destruct_elements_of_array<T>(ptr, size); aligned_free(ptr); @@ -396,13 +337,13 @@ template<typename T> inline void aligned_delete(T *ptr, size_t size) /** \internal Deletes objects constructed with conditional_aligned_new * The \a size parameters tells on how many objects to call the destructor of T. */ -template<typename T, bool Align> inline void conditional_aligned_delete(T *ptr, size_t size) +template<typename T, bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_delete(T *ptr, std::size_t size) { destruct_elements_of_array<T>(ptr, size); conditional_aligned_free<Align>(ptr); } -template<typename T, bool Align> inline T* conditional_aligned_realloc_new(T* pts, size_t new_size, size_t old_size) +template<typename T, bool Align> EIGEN_DEVICE_FUNC inline T* conditional_aligned_realloc_new(T* pts, std::size_t new_size, std::size_t old_size) { check_size_for_overflow<T>(new_size); check_size_for_overflow<T>(old_size); @@ -410,23 +351,43 @@ template<typename T, bool Align> inline T* conditional_aligned_realloc_new(T* pt destruct_elements_of_array(pts+new_size, old_size-new_size); T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size)); if(new_size > old_size) - construct_elements_of_array(result+old_size, new_size-old_size); + { + EIGEN_TRY + { + construct_elements_of_array(result+old_size, new_size-old_size); + } + EIGEN_CATCH(...) + { + conditional_aligned_free<Align>(result); + EIGEN_THROW; + } + } return result; } -template<typename T, bool Align> inline T* conditional_aligned_new_auto(size_t size) +template<typename T, bool Align> EIGEN_DEVICE_FUNC inline T* conditional_aligned_new_auto(std::size_t size) { if(size==0) return 0; // short-cut. Also fixes Bug 884 check_size_for_overflow<T>(size); T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size)); if(NumTraits<T>::RequireInitialization) - construct_elements_of_array(result, size); + { + EIGEN_TRY + { + construct_elements_of_array(result, size); + } + EIGEN_CATCH(...) + { + conditional_aligned_free<Align>(result); + EIGEN_THROW; + } + } return result; } -template<typename T, bool Align> inline T* conditional_aligned_realloc_new_auto(T* pts, size_t new_size, size_t old_size) +template<typename T, bool Align> inline T* conditional_aligned_realloc_new_auto(T* pts, std::size_t new_size, std::size_t old_size) { check_size_for_overflow<T>(new_size); check_size_for_overflow<T>(old_size); @@ -434,11 +395,21 @@ template<typename T, bool Align> inline T* conditional_aligned_realloc_new_auto( destruct_elements_of_array(pts+new_size, old_size-new_size); T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size)); if(NumTraits<T>::RequireInitialization && (new_size > old_size)) - construct_elements_of_array(result+old_size, new_size-old_size); + { + EIGEN_TRY + { + construct_elements_of_array(result+old_size, new_size-old_size); + } + EIGEN_CATCH(...) + { + conditional_aligned_free<Align>(result); + EIGEN_THROW; + } + } return result; } -template<typename T, bool Align> inline void conditional_aligned_delete_auto(T *ptr, size_t size) +template<typename T, bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_delete_auto(T *ptr, std::size_t size) { if(NumTraits<T>::RequireInitialization) destruct_elements_of_array<T>(ptr, size); @@ -447,51 +418,62 @@ template<typename T, bool Align> inline void conditional_aligned_delete_auto(T * /****************************************************************************/ -/** \internal Returns the index of the first element of the array that is well aligned for vectorization. +/** \internal Returns the index of the first element of the array that is well aligned with respect to the requested \a Alignment. * + * \tparam Alignment requested alignment in Bytes. * \param array the address of the start of the array * \param size the size of the array * - * \note If no element of the array is well aligned, the size of the array is returned. Typically, - * for example with SSE, "well aligned" means 16-byte-aligned. If vectorization is disabled or if the + * \note If no element of the array is well aligned or the requested alignment is not a multiple of a scalar, + * the size of the array is returned. For example with SSE, the requested alignment is typically 16-bytes. If * packet size for the given scalar type is 1, then everything is considered well-aligned. * - * \note If the scalar type is vectorizable, we rely on the following assumptions: sizeof(Scalar) is a - * power of 2, the packet size in bytes is also a power of 2, and is a multiple of sizeof(Scalar). On the - * other hand, we do not assume that the array address is a multiple of sizeof(Scalar), as that fails for + * \note Otherwise, if the Alignment is larger that the scalar size, we rely on the assumptions that sizeof(Scalar) is a + * power of 2. On the other hand, we do not assume that the array address is a multiple of sizeof(Scalar), as that fails for * example with Scalar=double on certain 32-bit platforms, see bug #79. * * There is also the variant first_aligned(const MatrixBase&) defined in DenseCoeffsBase.h. + * \sa first_default_aligned() */ -template<typename Scalar, typename Index> -static inline Index first_aligned(const Scalar* array, Index size) +template<int Alignment, typename Scalar, typename Index> +EIGEN_DEVICE_FUNC inline Index first_aligned(const Scalar* array, Index size) { - static const Index PacketSize = packet_traits<Scalar>::size; - static const Index PacketAlignedMask = PacketSize-1; + const Index ScalarSize = sizeof(Scalar); + const Index AlignmentSize = Alignment / ScalarSize; + const Index AlignmentMask = AlignmentSize-1; - if(PacketSize==1) + if(AlignmentSize<=1) { - // Either there is no vectorization, or a packet consists of exactly 1 scalar so that all elements - // of the array have the same alignment. + // Either the requested alignment if smaller than a scalar, or it exactly match a 1 scalar + // so that all elements of the array have the same alignment. return 0; } - else if(size_t(array) & (sizeof(Scalar)-1)) + else if( (UIntPtr(array) & (sizeof(Scalar)-1)) || (Alignment%ScalarSize)!=0) { - // There is vectorization for this scalar type, but the array is not aligned to the size of a single scalar. + // The array is not aligned to the size of a single scalar, or the requested alignment is not a multiple of the scalar size. // Consequently, no element of the array is well aligned. return size; } else { - return std::min<Index>( (PacketSize - (Index((size_t(array)/sizeof(Scalar))) & PacketAlignedMask)) - & PacketAlignedMask, size); + Index first = (AlignmentSize - (Index((UIntPtr(array)/sizeof(Scalar))) & AlignmentMask)) & AlignmentMask; + return (first < size) ? first : size; } } +/** \internal Returns the index of the first element of the array that is well aligned with respect the largest packet requirement. + * \sa first_aligned(Scalar*,Index) and first_default_aligned(DenseBase<Derived>) */ +template<typename Scalar, typename Index> +EIGEN_DEVICE_FUNC inline Index first_default_aligned(const Scalar* array, Index size) +{ + typedef typename packet_traits<Scalar>::type DefaultPacketType; + return first_aligned<unpacket_traits<DefaultPacketType>::alignment>(array, size); +} + /** \internal Returns the smallest integer multiple of \a base and greater or equal to \a size */ template<typename Index> -inline static Index first_multiple(Index size, Index base) +inline Index first_multiple(Index size, Index base) { return ((size+base-1)/base)*base; } @@ -500,21 +482,59 @@ inline static Index first_multiple(Index size, Index base) // use memcpy on trivial types, i.e., on types that does not require an initialization ctor. template<typename T, bool UseMemcpy> struct smart_copy_helper; -template<typename T> void smart_copy(const T* start, const T* end, T* target) +template<typename T> EIGEN_DEVICE_FUNC void smart_copy(const T* start, const T* end, T* target) { smart_copy_helper<T,!NumTraits<T>::RequireInitialization>::run(start, end, target); } template<typename T> struct smart_copy_helper<T,true> { - static inline void run(const T* start, const T* end, T* target) - { memcpy(target, start, std::ptrdiff_t(end)-std::ptrdiff_t(start)); } + EIGEN_DEVICE_FUNC static inline void run(const T* start, const T* end, T* target) + { + IntPtr size = IntPtr(end)-IntPtr(start); + if(size==0) return; + eigen_internal_assert(start!=0 && end!=0 && target!=0); + std::memcpy(target, start, size); + } }; template<typename T> struct smart_copy_helper<T,false> { - static inline void run(const T* start, const T* end, T* target) + EIGEN_DEVICE_FUNC static inline void run(const T* start, const T* end, T* target) { std::copy(start, end, target); } }; +// intelligent memmove. falls back to std::memmove for POD types, uses std::copy otherwise. +template<typename T, bool UseMemmove> struct smart_memmove_helper; + +template<typename T> void smart_memmove(const T* start, const T* end, T* target) +{ + smart_memmove_helper<T,!NumTraits<T>::RequireInitialization>::run(start, end, target); +} + +template<typename T> struct smart_memmove_helper<T,true> { + static inline void run(const T* start, const T* end, T* target) + { + IntPtr size = IntPtr(end)-IntPtr(start); + if(size==0) return; + eigen_internal_assert(start!=0 && end!=0 && target!=0); + std::memmove(target, start, size); + } +}; + +template<typename T> struct smart_memmove_helper<T,false> { + static inline void run(const T* start, const T* end, T* target) + { + if (UIntPtr(target) < UIntPtr(start)) + { + std::copy(start, end, target); + } + else + { + std::ptrdiff_t count = (std::ptrdiff_t(end)-std::ptrdiff_t(start)) / sizeof(T); + std::copy_backward(start, end, target + count); + } + } +}; + /***************************************************************************** *** Implementation of runtime stack allocation (falling back to malloc) *** @@ -523,16 +543,16 @@ template<typename T> struct smart_copy_helper<T,false> { // you can overwrite Eigen's default behavior regarding alloca by defining EIGEN_ALLOCA // to the appropriate stack allocation function #ifndef EIGEN_ALLOCA - #if (defined __linux__) || (defined __APPLE__) || (defined alloca) + #if EIGEN_OS_LINUX || EIGEN_OS_MAC || (defined alloca) #define EIGEN_ALLOCA alloca - #elif defined(_MSC_VER) + #elif EIGEN_COMP_MSVC #define EIGEN_ALLOCA _alloca #endif #endif // This helper class construct the allocated memory, and takes care of destructing and freeing the handled data // at destruction time. In practice this helper class is mainly useful to avoid memory leak in case of exceptions. -template<typename T> class aligned_stack_memory_handler +template<typename T> class aligned_stack_memory_handler : noncopyable { public: /* Creates a stack_memory_handler responsible for the buffer \a ptr of size \a size. @@ -541,7 +561,7 @@ template<typename T> class aligned_stack_memory_handler * In this case, the buffer elements will also be destructed when this handler will be destructed. * Finally, if \a dealloc is true, then the pointer \a ptr is freed. **/ - aligned_stack_memory_handler(T* ptr, size_t size, bool dealloc) + aligned_stack_memory_handler(T* ptr, std::size_t size, bool dealloc) : m_ptr(ptr), m_size(size), m_deallocate(dealloc) { if(NumTraits<T>::RequireInitialization && m_ptr) @@ -556,10 +576,34 @@ template<typename T> class aligned_stack_memory_handler } protected: T* m_ptr; - size_t m_size; + std::size_t m_size; bool m_deallocate; }; +template<typename T> class scoped_array : noncopyable +{ + T* m_ptr; +public: + explicit scoped_array(std::ptrdiff_t size) + { + m_ptr = new T[size]; + } + ~scoped_array() + { + delete[] m_ptr; + } + T& operator[](std::ptrdiff_t i) { return m_ptr[i]; } + const T& operator[](std::ptrdiff_t i) const { return m_ptr[i]; } + T* &ptr() { return m_ptr; } + const T* ptr() const { return m_ptr; } + operator const T*() const { return m_ptr; } +}; + +template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b) +{ + std::swap(a.ptr(),b.ptr()); +} + } // end namespace internal /** \internal @@ -578,11 +622,13 @@ template<typename T> class aligned_stack_memory_handler * The underlying stack allocation function can controlled with the EIGEN_ALLOCA preprocessor token. */ #ifdef EIGEN_ALLOCA - - #if defined(__arm__) || defined(_WIN32) - #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast<void*>((reinterpret_cast<size_t>(EIGEN_ALLOCA(SIZE+16)) & ~(size_t(15))) + 16) + + #if EIGEN_DEFAULT_ALIGN_BYTES>0 + // We always manually re-align the result of EIGEN_ALLOCA. + // If alloca is already aligned, the compiler should be smart enough to optimize away the re-alignment. + #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast<void*>((internal::UIntPtr(EIGEN_ALLOCA(SIZE+EIGEN_DEFAULT_ALIGN_BYTES-1)) + EIGEN_DEFAULT_ALIGN_BYTES-1) & ~(std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1))) #else - #define EIGEN_ALIGNED_ALLOCA EIGEN_ALLOCA + #define EIGEN_ALIGNED_ALLOCA(SIZE) EIGEN_ALLOCA(SIZE) #endif #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \ @@ -607,39 +653,33 @@ template<typename T> class aligned_stack_memory_handler *** Implementation of EIGEN_MAKE_ALIGNED_OPERATOR_NEW [_IF] *** *****************************************************************************/ -#if EIGEN_ALIGN - #ifdef EIGEN_EXCEPTIONS - #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \ - void* operator new(size_t size, const std::nothrow_t&) throw() { \ - try { return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); } \ - catch (...) { return 0; } \ +#if EIGEN_MAX_ALIGN_BYTES!=0 + #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \ + void* operator new(std::size_t size, const std::nothrow_t&) EIGEN_NO_THROW { \ + EIGEN_TRY { return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); } \ + EIGEN_CATCH (...) { return 0; } \ } - #else - #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \ - void* operator new(size_t size, const std::nothrow_t&) throw() { \ - return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \ - } - #endif - #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign) \ - void *operator new(size_t size) { \ + void *operator new(std::size_t size) { \ return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \ } \ - void *operator new[](size_t size) { \ + void *operator new[](std::size_t size) { \ return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \ } \ - void operator delete(void * ptr) throw() { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \ - void operator delete[](void * ptr) throw() { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \ + void operator delete(void * ptr) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \ + void operator delete[](void * ptr) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \ + void operator delete(void * ptr, std::size_t /* sz */) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \ + void operator delete[](void * ptr, std::size_t /* sz */) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \ /* in-place new and delete. since (at least afaik) there is no actual */ \ /* memory allocated we can safely let the default implementation handle */ \ /* this particular case. */ \ - static void *operator new(size_t size, void *ptr) { return ::operator new(size,ptr); } \ - static void *operator new[](size_t size, void* ptr) { return ::operator new[](size,ptr); } \ - void operator delete(void * memory, void *ptr) throw() { return ::operator delete(memory,ptr); } \ - void operator delete[](void * memory, void *ptr) throw() { return ::operator delete[](memory,ptr); } \ + static void *operator new(std::size_t size, void *ptr) { return ::operator new(size,ptr); } \ + static void *operator new[](std::size_t size, void* ptr) { return ::operator new[](size,ptr); } \ + void operator delete(void * memory, void *ptr) EIGEN_NO_THROW { return ::operator delete(memory,ptr); } \ + void operator delete[](void * memory, void *ptr) EIGEN_NO_THROW { return ::operator delete[](memory,ptr); } \ /* nothrow-new (returns zero instead of std::bad_alloc) */ \ EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \ - void operator delete(void *ptr, const std::nothrow_t&) throw() { \ + void operator delete(void *ptr, const std::nothrow_t&) EIGEN_NO_THROW { \ Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); \ } \ typedef void eigen_aligned_operator_new_marker_type; @@ -649,14 +689,22 @@ template<typename T> class aligned_stack_memory_handler #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(true) #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size) \ - EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool(((Size)!=Eigen::Dynamic) && ((sizeof(Scalar)*(Size))%16==0))) + EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool(((Size)!=Eigen::Dynamic) && ((sizeof(Scalar)*(Size))%EIGEN_MAX_ALIGN_BYTES==0))) /****************************************************************************/ /** \class aligned_allocator * \ingroup Core_Module * -* \brief STL compatible allocator to use with with 16 byte aligned types +* \brief STL compatible allocator to use with types requiring a non standrad alignment. +* +* The memory is aligned as for dynamically aligned matrix/array types such as MatrixXd. +* By default, it will thus provide at least 16 bytes alignment and more in following cases: +* - 32 bytes alignment if AVX is enabled. +* - 64 bytes alignment if AVX512 is enabled. +* +* This can be controled using the \c EIGEN_MAX_ALIGN_BYTES macro as documented +* \link TopicPreprocessorDirectivesPerformance there \endlink. * * Example: * \code @@ -667,96 +715,64 @@ template<typename T> class aligned_stack_memory_handler * std::map< int, Vector3f > my_map_vec3; * \endcode * -* \sa \ref TopicStlContainers. +* \sa \blank \ref TopicStlContainers. */ template<class T> -class aligned_allocator +class aligned_allocator : public std::allocator<T> { public: - typedef size_t size_type; - typedef std::ptrdiff_t difference_type; - typedef T* pointer; - typedef const T* const_pointer; - typedef T& reference; - typedef const T& const_reference; - typedef T value_type; - - template<class U> - struct rebind - { - typedef aligned_allocator<U> other; - }; - - pointer address( reference value ) const - { - return &value; - } - - const_pointer address( const_reference value ) const - { - return &value; - } - - aligned_allocator() - { - } - - aligned_allocator( const aligned_allocator& ) - { - } - - template<class U> - aligned_allocator( const aligned_allocator<U>& ) - { - } - - ~aligned_allocator() - { - } - - size_type max_size() const - { - return (std::numeric_limits<size_type>::max)(); - } + typedef std::size_t size_type; + typedef std::ptrdiff_t difference_type; + typedef T* pointer; + typedef const T* const_pointer; + typedef T& reference; + typedef const T& const_reference; + typedef T value_type; + + template<class U> + struct rebind + { + typedef aligned_allocator<U> other; + }; - pointer allocate( size_type num, const void* hint = 0 ) - { - EIGEN_UNUSED_VARIABLE(hint); - internal::check_size_for_overflow<T>(num); - return static_cast<pointer>( internal::aligned_malloc( num * sizeof(T) ) ); - } + aligned_allocator() : std::allocator<T>() {} - void construct( pointer p, const T& value ) - { - ::new( p ) T( value ); - } + aligned_allocator(const aligned_allocator& other) : std::allocator<T>(other) {} - void destroy( pointer p ) - { - p->~T(); - } + template<class U> + aligned_allocator(const aligned_allocator<U>& other) : std::allocator<T>(other) {} - void deallocate( pointer p, size_type /*num*/ ) - { - internal::aligned_free( p ); - } + ~aligned_allocator() {} - bool operator!=(const aligned_allocator<T>& ) const - { return false; } + pointer allocate(size_type num, const void* /*hint*/ = 0) + { + internal::check_size_for_overflow<T>(num); + size_type size = num * sizeof(T); +#if EIGEN_COMP_GNUC_STRICT && EIGEN_GNUC_AT_LEAST(7,0) + // workaround gcc bug https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87544 + // It triggered eigen/Eigen/src/Core/util/Memory.h:189:12: warning: argument 1 value '18446744073709551612' exceeds maximum object size 9223372036854775807 + if(size>=std::size_t((std::numeric_limits<std::ptrdiff_t>::max)())) + return 0; + else +#endif + return static_cast<pointer>( internal::aligned_malloc(size) ); + } - bool operator==(const aligned_allocator<T>& ) const - { return true; } + void deallocate(pointer p, size_type /*num*/) + { + internal::aligned_free(p); + } }; //---------- Cache sizes ---------- #if !defined(EIGEN_NO_CPUID) -# if defined(__GNUC__) && ( defined(__i386__) || defined(__x86_64__) ) -# if defined(__PIC__) && defined(__i386__) +# if EIGEN_COMP_GNUC && EIGEN_ARCH_i386_OR_x86_64 +# if defined(__PIC__) && EIGEN_ARCH_i386 // Case for x86 with PIC # define EIGEN_CPUID(abcd,func,id) \ __asm__ __volatile__ ("xchgl %%ebx, %k1;cpuid; xchgl %%ebx,%k1": "=a" (abcd[0]), "=&r" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "a" (func), "c" (id)); -# elif defined(__PIC__) && defined(__x86_64__) +# elif defined(__PIC__) && EIGEN_ARCH_x86_64 // Case for x64 with PIC. In theory this is only a problem with recent gcc and with medium or large code model, not with the default small code model. // However, we cannot detect which code model is used, and the xchg overhead is negligible anyway. # define EIGEN_CPUID(abcd,func,id) \ @@ -766,8 +782,8 @@ public: # define EIGEN_CPUID(abcd,func,id) \ __asm__ __volatile__ ("cpuid": "=a" (abcd[0]), "=b" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "0" (func), "2" (id) ); # endif -# elif defined(_MSC_VER) -# if (_MSC_VER > 1500) && ( defined(_M_IX86) || defined(_M_X64) ) +# elif EIGEN_COMP_MSVC +# if (EIGEN_COMP_MSVC > 1500) && EIGEN_ARCH_i386_OR_x86_64 # define EIGEN_CPUID(abcd,func,id) __cpuidex((int*)abcd,func,id) # endif # endif |