1 // This file is part of Eigen, a lightweight C++ template library 2 // for linear algebra. 3 // 4 // Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud (at) inria.fr> 5 // Copyright (C) 2008-2009 Benoit Jacob <jacob.benoit.1 (at) gmail.com> 6 // Copyright (C) 2009 Kenneth Riddile <kfriddile (at) yahoo.com> 7 // Copyright (C) 2010 Hauke Heibel <hauke.heibel (at) gmail.com> 8 // Copyright (C) 2010 Thomas Capricelli <orzel (at) freehackers.org> 9 // 10 // This Source Code Form is subject to the terms of the Mozilla 11 // Public License v. 2.0. If a copy of the MPL was not distributed 12 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. 13 14 15 /***************************************************************************** 16 *** Platform checks for aligned malloc functions *** 17 *****************************************************************************/ 18 19 #ifndef EIGEN_MEMORY_H 20 #define EIGEN_MEMORY_H 21 22 // On 64-bit systems, glibc's malloc returns 16-byte-aligned pointers, see: 23 // http://www.gnu.org/s/libc/manual/html_node/Aligned-Memory-Blocks.html 24 // This is true at least since glibc 2.8. 25 // This leaves the question how to detect 64-bit. According to this document, 26 // http://gcc.fyxm.net/summit/2003/Porting%20to%2064%20bit.pdf 27 // page 114, "[The] LP64 model [...] is used by all 64-bit UNIX ports" so it's indeed 28 // quite safe, at least within the context of glibc, to equate 64-bit with LP64. 29 #if defined(__GLIBC__) && ((__GLIBC__>=2 && __GLIBC_MINOR__ >= 8) || __GLIBC__>2) \ 30 && defined(__LP64__) 31 #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 1 32 #else 33 #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 0 34 #endif 35 36 // FreeBSD 6 seems to have 16-byte aligned malloc 37 // See http://svn.freebsd.org/viewvc/base/stable/6/lib/libc/stdlib/malloc.c?view=markup 38 // FreeBSD 7 seems to have 16-byte aligned malloc except on ARM and MIPS architectures 39 // See http://svn.freebsd.org/viewvc/base/stable/7/lib/libc/stdlib/malloc.c?view=markup 40 #if defined(__FreeBSD__) && !defined(__arm__) && !defined(__mips__) 41 #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 1 42 #else 43 #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 0 44 #endif 45 46 #if defined(__APPLE__) \ 47 || defined(_WIN64) \ 48 || EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED \ 49 || EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 50 #define EIGEN_MALLOC_ALREADY_ALIGNED 1 51 #else 52 #define EIGEN_MALLOC_ALREADY_ALIGNED 0 53 #endif 54 55 #if ((defined __QNXNTO__) || (defined _GNU_SOURCE) || ((defined _XOPEN_SOURCE) && (_XOPEN_SOURCE >= 600))) \ 56 && (defined _POSIX_ADVISORY_INFO) && (_POSIX_ADVISORY_INFO > 0) 57 #define EIGEN_HAS_POSIX_MEMALIGN 1 58 #else 59 #define EIGEN_HAS_POSIX_MEMALIGN 0 60 #endif 61 62 #ifdef EIGEN_VECTORIZE_SSE 63 #define EIGEN_HAS_MM_MALLOC 1 64 #else 65 #define EIGEN_HAS_MM_MALLOC 0 66 #endif 67 68 namespace Eigen { 69 70 namespace internal { 71 72 inline void throw_std_bad_alloc() 73 { 74 #ifdef EIGEN_EXCEPTIONS 75 throw std::bad_alloc(); 76 #else 77 std::size_t huge = -1; 78 new int[huge]; 79 #endif 80 } 81 82 /***************************************************************************** 83 *** Implementation of handmade aligned functions *** 84 *****************************************************************************/ 85 86 /* ----- Hand made implementations of aligned malloc/free and realloc ----- */ 87 88 /** \internal Like malloc, but the returned pointer is guaranteed to be 16-byte aligned. 89 * Fast, but wastes 16 additional bytes of memory. Does not throw any exception. 90 */ 91 inline void* handmade_aligned_malloc(size_t size) 92 { 93 void *original = std::malloc(size+16); 94 if (original == 0) return 0; 95 void *aligned = reinterpret_cast<void*>((reinterpret_cast<size_t>(original) & ~(size_t(15))) + 16); 96 *(reinterpret_cast<void**>(aligned) - 1) = original; 97 return aligned; 98 } 99 100 /** \internal Frees memory allocated with handmade_aligned_malloc */ 101 inline void handmade_aligned_free(void *ptr) 102 { 103 if (ptr) std::free(*(reinterpret_cast<void**>(ptr) - 1)); 104 } 105 106 /** \internal 107 * \brief Reallocates aligned memory. 108 * Since we know that our handmade version is based on std::realloc 109 * we can use std::realloc to implement efficient reallocation. 110 */ 111 inline void* handmade_aligned_realloc(void* ptr, size_t size, size_t = 0) 112 { 113 if (ptr == 0) return handmade_aligned_malloc(size); 114 void *original = *(reinterpret_cast<void**>(ptr) - 1); 115 original = std::realloc(original,size+16); 116 if (original == 0) return 0; 117 void *aligned = reinterpret_cast<void*>((reinterpret_cast<size_t>(original) & ~(size_t(15))) + 16); 118 *(reinterpret_cast<void**>(aligned) - 1) = original; 119 return aligned; 120 } 121 122 /***************************************************************************** 123 *** Implementation of generic aligned realloc (when no realloc can be used)*** 124 *****************************************************************************/ 125 126 void* aligned_malloc(size_t size); 127 void aligned_free(void *ptr); 128 129 /** \internal 130 * \brief Reallocates aligned memory. 131 * Allows reallocation with aligned ptr types. This implementation will 132 * always create a new memory chunk and copy the old data. 133 */ 134 inline void* generic_aligned_realloc(void* ptr, size_t size, size_t old_size) 135 { 136 if (ptr==0) 137 return aligned_malloc(size); 138 139 if (size==0) 140 { 141 aligned_free(ptr); 142 return 0; 143 } 144 145 void* newptr = aligned_malloc(size); 146 if (newptr == 0) 147 { 148 #ifdef EIGEN_HAS_ERRNO 149 errno = ENOMEM; // according to the standard 150 #endif 151 return 0; 152 } 153 154 if (ptr != 0) 155 { 156 std::memcpy(newptr, ptr, (std::min)(size,old_size)); 157 aligned_free(ptr); 158 } 159 160 return newptr; 161 } 162 163 /***************************************************************************** 164 *** Implementation of portable aligned versions of malloc/free/realloc *** 165 *****************************************************************************/ 166 167 #ifdef EIGEN_NO_MALLOC 168 inline void check_that_malloc_is_allowed() 169 { 170 eigen_assert(false && "heap allocation is forbidden (EIGEN_NO_MALLOC is defined)"); 171 } 172 #elif defined EIGEN_RUNTIME_NO_MALLOC 173 inline bool is_malloc_allowed_impl(bool update, bool new_value = false) 174 { 175 static bool value = true; 176 if (update == 1) 177 value = new_value; 178 return value; 179 } 180 inline bool is_malloc_allowed() { return is_malloc_allowed_impl(false); } 181 inline bool set_is_malloc_allowed(bool new_value) { return is_malloc_allowed_impl(true, new_value); } 182 inline void check_that_malloc_is_allowed() 183 { 184 eigen_assert(is_malloc_allowed() && "heap allocation is forbidden (EIGEN_RUNTIME_NO_MALLOC is defined and g_is_malloc_allowed is false)"); 185 } 186 #else 187 inline void check_that_malloc_is_allowed() 188 {} 189 #endif 190 191 /** \internal Allocates \a size bytes. The returned pointer is guaranteed to have 16 bytes alignment. 192 * On allocation error, the returned pointer is null, and std::bad_alloc is thrown. 193 */ 194 inline void* aligned_malloc(size_t size) 195 { 196 check_that_malloc_is_allowed(); 197 198 void *result; 199 #if !EIGEN_ALIGN 200 result = std::malloc(size); 201 #elif EIGEN_MALLOC_ALREADY_ALIGNED 202 result = std::malloc(size); 203 #elif EIGEN_HAS_POSIX_MEMALIGN 204 if(posix_memalign(&result, 16, size)) result = 0; 205 #elif EIGEN_HAS_MM_MALLOC 206 result = _mm_malloc(size, 16); 207 #elif (defined _MSC_VER) 208 result = _aligned_malloc(size, 16); 209 #else 210 result = handmade_aligned_malloc(size); 211 #endif 212 213 if(!result && size) 214 throw_std_bad_alloc(); 215 216 return result; 217 } 218 219 /** \internal Frees memory allocated with aligned_malloc. */ 220 inline void aligned_free(void *ptr) 221 { 222 #if !EIGEN_ALIGN 223 std::free(ptr); 224 #elif EIGEN_MALLOC_ALREADY_ALIGNED 225 std::free(ptr); 226 #elif EIGEN_HAS_POSIX_MEMALIGN 227 std::free(ptr); 228 #elif EIGEN_HAS_MM_MALLOC 229 _mm_free(ptr); 230 #elif defined(_MSC_VER) 231 _aligned_free(ptr); 232 #else 233 handmade_aligned_free(ptr); 234 #endif 235 } 236 237 /** 238 * \internal 239 * \brief Reallocates an aligned block of memory. 240 * \throws std::bad_alloc on allocation failure 241 **/ 242 inline void* aligned_realloc(void *ptr, size_t new_size, size_t old_size) 243 { 244 EIGEN_UNUSED_VARIABLE(old_size); 245 246 void *result; 247 #if !EIGEN_ALIGN 248 result = std::realloc(ptr,new_size); 249 #elif EIGEN_MALLOC_ALREADY_ALIGNED 250 result = std::realloc(ptr,new_size); 251 #elif EIGEN_HAS_POSIX_MEMALIGN 252 result = generic_aligned_realloc(ptr,new_size,old_size); 253 #elif EIGEN_HAS_MM_MALLOC 254 // The defined(_mm_free) is just here to verify that this MSVC version 255 // implements _mm_malloc/_mm_free based on the corresponding _aligned_ 256 // functions. This may not always be the case and we just try to be safe. 257 #if defined(_MSC_VER) && defined(_mm_free) 258 result = _aligned_realloc(ptr,new_size,16); 259 #else 260 result = generic_aligned_realloc(ptr,new_size,old_size); 261 #endif 262 #elif defined(_MSC_VER) 263 result = _aligned_realloc(ptr,new_size,16); 264 #else 265 result = handmade_aligned_realloc(ptr,new_size,old_size); 266 #endif 267 268 if (!result && new_size) 269 throw_std_bad_alloc(); 270 271 return result; 272 } 273 274 /***************************************************************************** 275 *** Implementation of conditionally aligned functions *** 276 *****************************************************************************/ 277 278 /** \internal Allocates \a size bytes. If Align is true, then the returned ptr is 16-byte-aligned. 279 * On allocation error, the returned pointer is null, and a std::bad_alloc is thrown. 280 */ 281 template<bool Align> inline void* conditional_aligned_malloc(size_t size) 282 { 283 return aligned_malloc(size); 284 } 285 286 template<> inline void* conditional_aligned_malloc<false>(size_t size) 287 { 288 check_that_malloc_is_allowed(); 289 290 void *result = std::malloc(size); 291 if(!result && size) 292 throw_std_bad_alloc(); 293 return result; 294 } 295 296 /** \internal Frees memory allocated with conditional_aligned_malloc */ 297 template<bool Align> inline void conditional_aligned_free(void *ptr) 298 { 299 aligned_free(ptr); 300 } 301 302 template<> inline void conditional_aligned_free<false>(void *ptr) 303 { 304 std::free(ptr); 305 } 306 307 template<bool Align> inline void* conditional_aligned_realloc(void* ptr, size_t new_size, size_t old_size) 308 { 309 return aligned_realloc(ptr, new_size, old_size); 310 } 311 312 template<> inline void* conditional_aligned_realloc<false>(void* ptr, size_t new_size, size_t) 313 { 314 return std::realloc(ptr, new_size); 315 } 316 317 /***************************************************************************** 318 *** Construction/destruction of array elements *** 319 *****************************************************************************/ 320 321 /** \internal Constructs the elements of an array. 322 * The \a size parameter tells on how many objects to call the constructor of T. 323 */ 324 template<typename T> inline T* construct_elements_of_array(T *ptr, size_t size) 325 { 326 for (size_t i=0; i < size; ++i) ::new (ptr + i) T; 327 return ptr; 328 } 329 330 /** \internal Destructs the elements of an array. 331 * The \a size parameters tells on how many objects to call the destructor of T. 332 */ 333 template<typename T> inline void destruct_elements_of_array(T *ptr, size_t size) 334 { 335 // always destruct an array starting from the end. 336 if(ptr) 337 while(size) ptr[--size].~T(); 338 } 339 340 /***************************************************************************** 341 *** Implementation of aligned new/delete-like functions *** 342 *****************************************************************************/ 343 344 template<typename T> 345 EIGEN_ALWAYS_INLINE void check_size_for_overflow(size_t size) 346 { 347 if(size > size_t(-1) / sizeof(T)) 348 throw_std_bad_alloc(); 349 } 350 351 /** \internal Allocates \a size objects of type T. The returned pointer is guaranteed to have 16 bytes alignment. 352 * On allocation error, the returned pointer is undefined, but a std::bad_alloc is thrown. 353 * The default constructor of T is called. 354 */ 355 template<typename T> inline T* aligned_new(size_t size) 356 { 357 check_size_for_overflow<T>(size); 358 T *result = reinterpret_cast<T*>(aligned_malloc(sizeof(T)*size)); 359 return construct_elements_of_array(result, size); 360 } 361 362 template<typename T, bool Align> inline T* conditional_aligned_new(size_t size) 363 { 364 check_size_for_overflow<T>(size); 365 T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size)); 366 return construct_elements_of_array(result, size); 367 } 368 369 /** \internal Deletes objects constructed with aligned_new 370 * The \a size parameters tells on how many objects to call the destructor of T. 371 */ 372 template<typename T> inline void aligned_delete(T *ptr, size_t size) 373 { 374 destruct_elements_of_array<T>(ptr, size); 375 aligned_free(ptr); 376 } 377 378 /** \internal Deletes objects constructed with conditional_aligned_new 379 * The \a size parameters tells on how many objects to call the destructor of T. 380 */ 381 template<typename T, bool Align> inline void conditional_aligned_delete(T *ptr, size_t size) 382 { 383 destruct_elements_of_array<T>(ptr, size); 384 conditional_aligned_free<Align>(ptr); 385 } 386 387 template<typename T, bool Align> inline T* conditional_aligned_realloc_new(T* pts, size_t new_size, size_t old_size) 388 { 389 check_size_for_overflow<T>(new_size); 390 check_size_for_overflow<T>(old_size); 391 if(new_size < old_size) 392 destruct_elements_of_array(pts+new_size, old_size-new_size); 393 T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size)); 394 if(new_size > old_size) 395 construct_elements_of_array(result+old_size, new_size-old_size); 396 return result; 397 } 398 399 400 template<typename T, bool Align> inline T* conditional_aligned_new_auto(size_t size) 401 { 402 check_size_for_overflow<T>(size); 403 T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size)); 404 if(NumTraits<T>::RequireInitialization) 405 construct_elements_of_array(result, size); 406 return result; 407 } 408 409 template<typename T, bool Align> inline T* conditional_aligned_realloc_new_auto(T* pts, size_t new_size, size_t old_size) 410 { 411 check_size_for_overflow<T>(new_size); 412 check_size_for_overflow<T>(old_size); 413 if(NumTraits<T>::RequireInitialization && (new_size < old_size)) 414 destruct_elements_of_array(pts+new_size, old_size-new_size); 415 T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size)); 416 if(NumTraits<T>::RequireInitialization && (new_size > old_size)) 417 construct_elements_of_array(result+old_size, new_size-old_size); 418 return result; 419 } 420 421 template<typename T, bool Align> inline void conditional_aligned_delete_auto(T *ptr, size_t size) 422 { 423 if(NumTraits<T>::RequireInitialization) 424 destruct_elements_of_array<T>(ptr, size); 425 conditional_aligned_free<Align>(ptr); 426 } 427 428 /****************************************************************************/ 429 430 /** \internal Returns the index of the first element of the array that is well aligned for vectorization. 431 * 432 * \param array the address of the start of the array 433 * \param size the size of the array 434 * 435 * \note If no element of the array is well aligned, the size of the array is returned. Typically, 436 * for example with SSE, "well aligned" means 16-byte-aligned. If vectorization is disabled or if the 437 * packet size for the given scalar type is 1, then everything is considered well-aligned. 438 * 439 * \note If the scalar type is vectorizable, we rely on the following assumptions: sizeof(Scalar) is a 440 * power of 2, the packet size in bytes is also a power of 2, and is a multiple of sizeof(Scalar). On the 441 * other hand, we do not assume that the array address is a multiple of sizeof(Scalar), as that fails for 442 * example with Scalar=double on certain 32-bit platforms, see bug #79. 443 * 444 * There is also the variant first_aligned(const MatrixBase&) defined in DenseCoeffsBase.h. 445 */ 446 template<typename Scalar, typename Index> 447 static inline Index first_aligned(const Scalar* array, Index size) 448 { 449 typedef typename packet_traits<Scalar>::type Packet; 450 enum { PacketSize = packet_traits<Scalar>::size, 451 PacketAlignedMask = PacketSize-1 452 }; 453 454 if(PacketSize==1) 455 { 456 // Either there is no vectorization, or a packet consists of exactly 1 scalar so that all elements 457 // of the array have the same alignment. 458 return 0; 459 } 460 else if(size_t(array) & (sizeof(Scalar)-1)) 461 { 462 // There is vectorization for this scalar type, but the array is not aligned to the size of a single scalar. 463 // Consequently, no element of the array is well aligned. 464 return size; 465 } 466 else 467 { 468 return std::min<Index>( (PacketSize - (Index((size_t(array)/sizeof(Scalar))) & PacketAlignedMask)) 469 & PacketAlignedMask, size); 470 } 471 } 472 473 474 // std::copy is much slower than memcpy, so let's introduce a smart_copy which 475 // use memcpy on trivial types, i.e., on types that does not require an initialization ctor. 476 template<typename T, bool UseMemcpy> struct smart_copy_helper; 477 478 template<typename T> void smart_copy(const T* start, const T* end, T* target) 479 { 480 smart_copy_helper<T,!NumTraits<T>::RequireInitialization>::run(start, end, target); 481 } 482 483 template<typename T> struct smart_copy_helper<T,true> { 484 static inline void run(const T* start, const T* end, T* target) 485 { memcpy(target, start, std::ptrdiff_t(end)-std::ptrdiff_t(start)); } 486 }; 487 488 template<typename T> struct smart_copy_helper<T,false> { 489 static inline void run(const T* start, const T* end, T* target) 490 { std::copy(start, end, target); } 491 }; 492 493 494 /***************************************************************************** 495 *** Implementation of runtime stack allocation (falling back to malloc) *** 496 *****************************************************************************/ 497 498 // you can overwrite Eigen's default behavior regarding alloca by defining EIGEN_ALLOCA 499 // to the appropriate stack allocation function 500 #ifndef EIGEN_ALLOCA 501 #if (defined __linux__) 502 #define EIGEN_ALLOCA alloca 503 #elif defined(_MSC_VER) 504 #define EIGEN_ALLOCA _alloca 505 #endif 506 #endif 507 508 // This helper class construct the allocated memory, and takes care of destructing and freeing the handled data 509 // at destruction time. In practice this helper class is mainly useful to avoid memory leak in case of exceptions. 510 template<typename T> class aligned_stack_memory_handler 511 { 512 public: 513 /* Creates a stack_memory_handler responsible for the buffer \a ptr of size \a size. 514 * Note that \a ptr can be 0 regardless of the other parameters. 515 * This constructor takes care of constructing/initializing the elements of the buffer if required by the scalar type T (see NumTraits<T>::RequireInitialization). 516 * In this case, the buffer elements will also be destructed when this handler will be destructed. 517 * Finally, if \a dealloc is true, then the pointer \a ptr is freed. 518 **/ 519 aligned_stack_memory_handler(T* ptr, size_t size, bool dealloc) 520 : m_ptr(ptr), m_size(size), m_deallocate(dealloc) 521 { 522 if(NumTraits<T>::RequireInitialization && m_ptr) 523 Eigen::internal::construct_elements_of_array(m_ptr, size); 524 } 525 ~aligned_stack_memory_handler() 526 { 527 if(NumTraits<T>::RequireInitialization && m_ptr) 528 Eigen::internal::destruct_elements_of_array<T>(m_ptr, m_size); 529 if(m_deallocate) 530 Eigen::internal::aligned_free(m_ptr); 531 } 532 protected: 533 T* m_ptr; 534 size_t m_size; 535 bool m_deallocate; 536 }; 537 538 } // end namespace internal 539 540 /** \internal 541 * Declares, allocates and construct an aligned buffer named NAME of SIZE elements of type TYPE on the stack 542 * if SIZE is smaller than EIGEN_STACK_ALLOCATION_LIMIT, and if stack allocation is supported by the platform 543 * (currently, this is Linux and Visual Studio only). Otherwise the memory is allocated on the heap. 544 * The allocated buffer is automatically deleted when exiting the scope of this declaration. 545 * If BUFFER is non null, then the declared variable is simply an alias for BUFFER, and no allocation/deletion occurs. 546 * Here is an example: 547 * \code 548 * { 549 * ei_declare_aligned_stack_constructed_variable(float,data,size,0); 550 * // use data[0] to data[size-1] 551 * } 552 * \endcode 553 * The underlying stack allocation function can controlled with the EIGEN_ALLOCA preprocessor token. 554 */ 555 #ifdef EIGEN_ALLOCA 556 557 #ifdef __arm__ 558 #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast<void*>((reinterpret_cast<size_t>(EIGEN_ALLOCA(SIZE+16)) & ~(size_t(15))) + 16) 559 #else 560 #define EIGEN_ALIGNED_ALLOCA EIGEN_ALLOCA 561 #endif 562 563 #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \ 564 Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \ 565 TYPE* NAME = (BUFFER)!=0 ? (BUFFER) \ 566 : reinterpret_cast<TYPE*>( \ 567 (sizeof(TYPE)*SIZE<=EIGEN_STACK_ALLOCATION_LIMIT) ? EIGEN_ALIGNED_ALLOCA(sizeof(TYPE)*SIZE) \ 568 : Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE) ); \ 569 Eigen::internal::aligned_stack_memory_handler<TYPE> EIGEN_CAT(NAME,_stack_memory_destructor)((BUFFER)==0 ? NAME : 0,SIZE,sizeof(TYPE)*SIZE>EIGEN_STACK_ALLOCATION_LIMIT) 570 571 #else 572 573 #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \ 574 Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \ 575 TYPE* NAME = (BUFFER)!=0 ? BUFFER : reinterpret_cast<TYPE*>(Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE)); \ 576 Eigen::internal::aligned_stack_memory_handler<TYPE> EIGEN_CAT(NAME,_stack_memory_destructor)((BUFFER)==0 ? NAME : 0,SIZE,true) 577 578 #endif 579 580 581 /***************************************************************************** 582 *** Implementation of EIGEN_MAKE_ALIGNED_OPERATOR_NEW [_IF] *** 583 *****************************************************************************/ 584 585 #if EIGEN_ALIGN 586 #ifdef EIGEN_EXCEPTIONS 587 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \ 588 void* operator new(size_t size, const std::nothrow_t&) throw() { \ 589 try { return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); } \ 590 catch (...) { return 0; } \ 591 return 0; \ 592 } 593 #else 594 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \ 595 void* operator new(size_t size, const std::nothrow_t&) throw() { \ 596 return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \ 597 } 598 #endif 599 600 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign) \ 601 void *operator new(size_t size) { \ 602 return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \ 603 } \ 604 void *operator new[](size_t size) { \ 605 return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \ 606 } \ 607 void operator delete(void * ptr) throw() { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \ 608 void operator delete[](void * ptr) throw() { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \ 609 /* in-place new and delete. since (at least afaik) there is no actual */ \ 610 /* memory allocated we can safely let the default implementation handle */ \ 611 /* this particular case. */ \ 612 static void *operator new(size_t size, void *ptr) { return ::operator new(size,ptr); } \ 613 void operator delete(void * memory, void *ptr) throw() { return ::operator delete(memory,ptr); } \ 614 /* nothrow-new (returns zero instead of std::bad_alloc) */ \ 615 EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \ 616 void operator delete(void *ptr, const std::nothrow_t&) throw() { \ 617 Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); \ 618 } \ 619 typedef void eigen_aligned_operator_new_marker_type; 620 #else 621 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign) 622 #endif 623 624 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(true) 625 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size) \ 626 EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool(((Size)!=Eigen::Dynamic) && ((sizeof(Scalar)*(Size))%16==0))) 627 628 /****************************************************************************/ 629 630 /** \class aligned_allocator 631 * \ingroup Core_Module 632 * 633 * \brief STL compatible allocator to use with with 16 byte aligned types 634 * 635 * Example: 636 * \code 637 * // Matrix4f requires 16 bytes alignment: 638 * std::map< int, Matrix4f, std::less<int>, 639 * aligned_allocator<std::pair<const int, Matrix4f> > > my_map_mat4; 640 * // Vector3f does not require 16 bytes alignment, no need to use Eigen's allocator: 641 * std::map< int, Vector3f > my_map_vec3; 642 * \endcode 643 * 644 * \sa \ref TopicStlContainers. 645 */ 646 template<class T> 647 class aligned_allocator 648 { 649 public: 650 typedef size_t size_type; 651 typedef std::ptrdiff_t difference_type; 652 typedef T* pointer; 653 typedef const T* const_pointer; 654 typedef T& reference; 655 typedef const T& const_reference; 656 typedef T value_type; 657 658 template<class U> 659 struct rebind 660 { 661 typedef aligned_allocator<U> other; 662 }; 663 664 pointer address( reference value ) const 665 { 666 return &value; 667 } 668 669 const_pointer address( const_reference value ) const 670 { 671 return &value; 672 } 673 674 aligned_allocator() 675 { 676 } 677 678 aligned_allocator( const aligned_allocator& ) 679 { 680 } 681 682 template<class U> 683 aligned_allocator( const aligned_allocator<U>& ) 684 { 685 } 686 687 ~aligned_allocator() 688 { 689 } 690 691 size_type max_size() const 692 { 693 return (std::numeric_limits<size_type>::max)(); 694 } 695 696 pointer allocate( size_type num, const void* hint = 0 ) 697 { 698 EIGEN_UNUSED_VARIABLE(hint); 699 internal::check_size_for_overflow<T>(num); 700 return static_cast<pointer>( internal::aligned_malloc( num * sizeof(T) ) ); 701 } 702 703 void construct( pointer p, const T& value ) 704 { 705 ::new( p ) T( value ); 706 } 707 708 // Support for c++11 709 #if (__cplusplus >= 201103L) 710 template<typename... Args> 711 void construct(pointer p, Args&&... args) 712 { 713 ::new(p) T(std::forward<Args>(args)...); 714 } 715 #endif 716 717 void destroy( pointer p ) 718 { 719 p->~T(); 720 } 721 722 void deallocate( pointer p, size_type /*num*/ ) 723 { 724 internal::aligned_free( p ); 725 } 726 727 bool operator!=(const aligned_allocator<T>& ) const 728 { return false; } 729 730 bool operator==(const aligned_allocator<T>& ) const 731 { return true; } 732 }; 733 734 //---------- Cache sizes ---------- 735 736 #if !defined(EIGEN_NO_CPUID) 737 # if defined(__GNUC__) && ( defined(__i386__) || defined(__x86_64__) ) 738 # if defined(__PIC__) && defined(__i386__) 739 // Case for x86 with PIC 740 # define EIGEN_CPUID(abcd,func,id) \ 741 __asm__ __volatile__ ("xchgl %%ebx, %%esi;cpuid; xchgl %%ebx,%%esi": "=a" (abcd[0]), "=S" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "a" (func), "c" (id)); 742 # else 743 // Case for x86_64 or x86 w/o PIC 744 # define EIGEN_CPUID(abcd,func,id) \ 745 __asm__ __volatile__ ("cpuid": "=a" (abcd[0]), "=b" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "a" (func), "c" (id) ); 746 # endif 747 # elif defined(_MSC_VER) 748 # if (_MSC_VER > 1500) 749 # define EIGEN_CPUID(abcd,func,id) __cpuidex((int*)abcd,func,id) 750 # endif 751 # endif 752 #endif 753 754 namespace internal { 755 756 #ifdef EIGEN_CPUID 757 758 inline bool cpuid_is_vendor(int abcd[4], const char* vendor) 759 { 760 return abcd[1]==(reinterpret_cast<const int*>(vendor))[0] && abcd[3]==(reinterpret_cast<const int*>(vendor))[1] && abcd[2]==(reinterpret_cast<const int*>(vendor))[2]; 761 } 762 763 inline void queryCacheSizes_intel_direct(int& l1, int& l2, int& l3) 764 { 765 int abcd[4]; 766 l1 = l2 = l3 = 0; 767 int cache_id = 0; 768 int cache_type = 0; 769 do { 770 abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0; 771 EIGEN_CPUID(abcd,0x4,cache_id); 772 cache_type = (abcd[0] & 0x0F) >> 0; 773 if(cache_type==1||cache_type==3) // data or unified cache 774 { 775 int cache_level = (abcd[0] & 0xE0) >> 5; // A[7:5] 776 int ways = (abcd[1] & 0xFFC00000) >> 22; // B[31:22] 777 int partitions = (abcd[1] & 0x003FF000) >> 12; // B[21:12] 778 int line_size = (abcd[1] & 0x00000FFF) >> 0; // B[11:0] 779 int sets = (abcd[2]); // C[31:0] 780 781 int cache_size = (ways+1) * (partitions+1) * (line_size+1) * (sets+1); 782 783 switch(cache_level) 784 { 785 case 1: l1 = cache_size; break; 786 case 2: l2 = cache_size; break; 787 case 3: l3 = cache_size; break; 788 default: break; 789 } 790 } 791 cache_id++; 792 } while(cache_type>0 && cache_id<16); 793 } 794 795 inline void queryCacheSizes_intel_codes(int& l1, int& l2, int& l3) 796 { 797 int abcd[4]; 798 abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0; 799 l1 = l2 = l3 = 0; 800 EIGEN_CPUID(abcd,0x00000002,0); 801 unsigned char * bytes = reinterpret_cast<unsigned char *>(abcd)+2; 802 bool check_for_p2_core2 = false; 803 for(int i=0; i<14; ++i) 804 { 805 switch(bytes[i]) 806 { 807 case 0x0A: l1 = 8; break; // 0Ah data L1 cache, 8 KB, 2 ways, 32 byte lines 808 case 0x0C: l1 = 16; break; // 0Ch data L1 cache, 16 KB, 4 ways, 32 byte lines 809 case 0x0E: l1 = 24; break; // 0Eh data L1 cache, 24 KB, 6 ways, 64 byte lines 810 case 0x10: l1 = 16; break; // 10h data L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64) 811 case 0x15: l1 = 16; break; // 15h code L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64) 812 case 0x2C: l1 = 32; break; // 2Ch data L1 cache, 32 KB, 8 ways, 64 byte lines 813 case 0x30: l1 = 32; break; // 30h code L1 cache, 32 KB, 8 ways, 64 byte lines 814 case 0x60: l1 = 16; break; // 60h data L1 cache, 16 KB, 8 ways, 64 byte lines, sectored 815 case 0x66: l1 = 8; break; // 66h data L1 cache, 8 KB, 4 ways, 64 byte lines, sectored 816 case 0x67: l1 = 16; break; // 67h data L1 cache, 16 KB, 4 ways, 64 byte lines, sectored 817 case 0x68: l1 = 32; break; // 68h data L1 cache, 32 KB, 4 ways, 64 byte lines, sectored 818 case 0x1A: l2 = 96; break; // code and data L2 cache, 96 KB, 6 ways, 64 byte lines (IA-64) 819 case 0x22: l3 = 512; break; // code and data L3 cache, 512 KB, 4 ways (!), 64 byte lines, dual-sectored 820 case 0x23: l3 = 1024; break; // code and data L3 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored 821 case 0x25: l3 = 2048; break; // code and data L3 cache, 2048 KB, 8 ways, 64 byte lines, dual-sectored 822 case 0x29: l3 = 4096; break; // code and data L3 cache, 4096 KB, 8 ways, 64 byte lines, dual-sectored 823 case 0x39: l2 = 128; break; // code and data L2 cache, 128 KB, 4 ways, 64 byte lines, sectored 824 case 0x3A: l2 = 192; break; // code and data L2 cache, 192 KB, 6 ways, 64 byte lines, sectored 825 case 0x3B: l2 = 128; break; // code and data L2 cache, 128 KB, 2 ways, 64 byte lines, sectored 826 case 0x3C: l2 = 256; break; // code and data L2 cache, 256 KB, 4 ways, 64 byte lines, sectored 827 case 0x3D: l2 = 384; break; // code and data L2 cache, 384 KB, 6 ways, 64 byte lines, sectored 828 case 0x3E: l2 = 512; break; // code and data L2 cache, 512 KB, 4 ways, 64 byte lines, sectored 829 case 0x40: l2 = 0; break; // no integrated L2 cache (P6 core) or L3 cache (P4 core) 830 case 0x41: l2 = 128; break; // code and data L2 cache, 128 KB, 4 ways, 32 byte lines 831 case 0x42: l2 = 256; break; // code and data L2 cache, 256 KB, 4 ways, 32 byte lines 832 case 0x43: l2 = 512; break; // code and data L2 cache, 512 KB, 4 ways, 32 byte lines 833 case 0x44: l2 = 1024; break; // code and data L2 cache, 1024 KB, 4 ways, 32 byte lines 834 case 0x45: l2 = 2048; break; // code and data L2 cache, 2048 KB, 4 ways, 32 byte lines 835 case 0x46: l3 = 4096; break; // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines 836 case 0x47: l3 = 8192; break; // code and data L3 cache, 8192 KB, 8 ways, 64 byte lines 837 case 0x48: l2 = 3072; break; // code and data L2 cache, 3072 KB, 12 ways, 64 byte lines 838 case 0x49: if(l2!=0) l3 = 4096; else {check_for_p2_core2=true; l3 = l2 = 4096;} break;// code and data L3 cache, 4096 KB, 16 ways, 64 byte lines (P4) or L2 for core2 839 case 0x4A: l3 = 6144; break; // code and data L3 cache, 6144 KB, 12 ways, 64 byte lines 840 case 0x4B: l3 = 8192; break; // code and data L3 cache, 8192 KB, 16 ways, 64 byte lines 841 case 0x4C: l3 = 12288; break; // code and data L3 cache, 12288 KB, 12 ways, 64 byte lines 842 case 0x4D: l3 = 16384; break; // code and data L3 cache, 16384 KB, 16 ways, 64 byte lines 843 case 0x4E: l2 = 6144; break; // code and data L2 cache, 6144 KB, 24 ways, 64 byte lines 844 case 0x78: l2 = 1024; break; // code and data L2 cache, 1024 KB, 4 ways, 64 byte lines 845 case 0x79: l2 = 128; break; // code and data L2 cache, 128 KB, 8 ways, 64 byte lines, dual-sectored 846 case 0x7A: l2 = 256; break; // code and data L2 cache, 256 KB, 8 ways, 64 byte lines, dual-sectored 847 case 0x7B: l2 = 512; break; // code and data L2 cache, 512 KB, 8 ways, 64 byte lines, dual-sectored 848 case 0x7C: l2 = 1024; break; // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored 849 case 0x7D: l2 = 2048; break; // code and data L2 cache, 2048 KB, 8 ways, 64 byte lines 850 case 0x7E: l2 = 256; break; // code and data L2 cache, 256 KB, 8 ways, 128 byte lines, sect. (IA-64) 851 case 0x7F: l2 = 512; break; // code and data L2 cache, 512 KB, 2 ways, 64 byte lines 852 case 0x80: l2 = 512; break; // code and data L2 cache, 512 KB, 8 ways, 64 byte lines 853 case 0x81: l2 = 128; break; // code and data L2 cache, 128 KB, 8 ways, 32 byte lines 854 case 0x82: l2 = 256; break; // code and data L2 cache, 256 KB, 8 ways, 32 byte lines 855 case 0x83: l2 = 512; break; // code and data L2 cache, 512 KB, 8 ways, 32 byte lines 856 case 0x84: l2 = 1024; break; // code and data L2 cache, 1024 KB, 8 ways, 32 byte lines 857 case 0x85: l2 = 2048; break; // code and data L2 cache, 2048 KB, 8 ways, 32 byte lines 858 case 0x86: l2 = 512; break; // code and data L2 cache, 512 KB, 4 ways, 64 byte lines 859 case 0x87: l2 = 1024; break; // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines 860 case 0x88: l3 = 2048; break; // code and data L3 cache, 2048 KB, 4 ways, 64 byte lines (IA-64) 861 case 0x89: l3 = 4096; break; // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines (IA-64) 862 case 0x8A: l3 = 8192; break; // code and data L3 cache, 8192 KB, 4 ways, 64 byte lines (IA-64) 863 case 0x8D: l3 = 3072; break; // code and data L3 cache, 3072 KB, 12 ways, 128 byte lines (IA-64) 864 865 default: break; 866 } 867 } 868 if(check_for_p2_core2 && l2 == l3) 869 l3 = 0; 870 l1 *= 1024; 871 l2 *= 1024; 872 l3 *= 1024; 873 } 874 875 inline void queryCacheSizes_intel(int& l1, int& l2, int& l3, int max_std_funcs) 876 { 877 if(max_std_funcs>=4) 878 queryCacheSizes_intel_direct(l1,l2,l3); 879 else 880 queryCacheSizes_intel_codes(l1,l2,l3); 881 } 882 883 inline void queryCacheSizes_amd(int& l1, int& l2, int& l3) 884 { 885 int abcd[4]; 886 abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0; 887 EIGEN_CPUID(abcd,0x80000005,0); 888 l1 = (abcd[2] >> 24) * 1024; // C[31:24] = L1 size in KB 889 abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0; 890 EIGEN_CPUID(abcd,0x80000006,0); 891 l2 = (abcd[2] >> 16) * 1024; // C[31;16] = l2 cache size in KB 892 l3 = ((abcd[3] & 0xFFFC000) >> 18) * 512 * 1024; // D[31;18] = l3 cache size in 512KB 893 } 894 #endif 895 896 /** \internal 897 * Queries and returns the cache sizes in Bytes of the L1, L2, and L3 data caches respectively */ 898 inline void queryCacheSizes(int& l1, int& l2, int& l3) 899 { 900 #ifdef EIGEN_CPUID 901 int abcd[4]; 902 903 // identify the CPU vendor 904 EIGEN_CPUID(abcd,0x0,0); 905 int max_std_funcs = abcd[1]; 906 if(cpuid_is_vendor(abcd,"GenuineIntel")) 907 queryCacheSizes_intel(l1,l2,l3,max_std_funcs); 908 else if(cpuid_is_vendor(abcd,"AuthenticAMD") || cpuid_is_vendor(abcd,"AMDisbetter!")) 909 queryCacheSizes_amd(l1,l2,l3); 910 else 911 // by default let's use Intel's API 912 queryCacheSizes_intel(l1,l2,l3,max_std_funcs); 913 914 // here is the list of other vendors: 915 // ||cpuid_is_vendor(abcd,"VIA VIA VIA ") 916 // ||cpuid_is_vendor(abcd,"CyrixInstead") 917 // ||cpuid_is_vendor(abcd,"CentaurHauls") 918 // ||cpuid_is_vendor(abcd,"GenuineTMx86") 919 // ||cpuid_is_vendor(abcd,"TransmetaCPU") 920 // ||cpuid_is_vendor(abcd,"RiseRiseRise") 921 // ||cpuid_is_vendor(abcd,"Geode by NSC") 922 // ||cpuid_is_vendor(abcd,"SiS SiS SiS ") 923 // ||cpuid_is_vendor(abcd,"UMC UMC UMC ") 924 // ||cpuid_is_vendor(abcd,"NexGenDriven") 925 #else 926 l1 = l2 = l3 = -1; 927 #endif 928 } 929 930 /** \internal 931 * \returns the size in Bytes of the L1 data cache */ 932 inline int queryL1CacheSize() 933 { 934 int l1(-1), l2, l3; 935 queryCacheSizes(l1,l2,l3); 936 return l1; 937 } 938 939 /** \internal 940 * \returns the size in Bytes of the L2 or L3 cache if this later is present */ 941 inline int queryTopLevelCacheSize() 942 { 943 int l1, l2(-1), l3(-1); 944 queryCacheSizes(l1,l2,l3); 945 return (std::max)(l2,l3); 946 } 947 948 } // end namespace internal 949 950 } // end namespace Eigen 951 952 #endif // EIGEN_MEMORY_H 953
