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@inria.fr>
5 // Copyright (C) 2008-2009 Benoit Jacob <jacob.benoit.1@gmail.com>
6 // Copyright (C) 2009 Kenneth Riddile <kfriddile@yahoo.com>
7 // Copyright (C) 2010 Hauke Heibel <hauke.heibel@gmail.com>
8 // Copyright (C) 2010 Thomas Capricelli <orzel@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 #ifndef EIGEN_MALLOC_ALREADY_ALIGNED
23
24 // Try to determine automatically if malloc is already aligned.
25
26 // On 64-bit systems, glibc's malloc returns 16-byte-aligned pointers, see:
27 // http://www.gnu.org/s/libc/manual/html_node/Aligned-Memory-Blocks.html
28 // This is true at least since glibc 2.8.
29 // This leaves the question how to detect 64-bit. According to this document,
30 // http://gcc.fyxm.net/summit/2003/Porting%20to%2064%20bit.pdf
31 // page 114, "[The] LP64 model [...] is used by all 64-bit UNIX ports" so it's indeed
32 // quite safe, at least within the context of glibc, to equate 64-bit with LP64.
33 #if defined(__GLIBC__) && ((__GLIBC__>=2 && __GLIBC_MINOR__ >= 8) || __GLIBC__>2) \
34 && defined(__LP64__) && ! defined( __SANITIZE_ADDRESS__ )
35 #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 1
36 #else
37 #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 0
38 #endif
39
40 // FreeBSD 6 seems to have 16-byte aligned malloc
41 // See http://svn.freebsd.org/viewvc/base/stable/6/lib/libc/stdlib/malloc.c?view=markup
42 // FreeBSD 7 seems to have 16-byte aligned malloc except on ARM and MIPS architectures
43 // See http://svn.freebsd.org/viewvc/base/stable/7/lib/libc/stdlib/malloc.c?view=markup
44 #if defined(__FreeBSD__) && !defined(__arm__) && !defined(__mips__)
45 #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 1
46 #else
47 #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 0
48 #endif
49
50 #if defined(__APPLE__) \
51 || defined(_WIN64) \
52 || EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED \
53 || EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED
54 #define EIGEN_MALLOC_ALREADY_ALIGNED 1
55 #else
56 #define EIGEN_MALLOC_ALREADY_ALIGNED 0
57 #endif
58
59 #endif
60
61 // See bug 554 (http://eigen.tuxfamily.org/bz/show_bug.cgi?id=554)
62 // It seems to be unsafe to check _POSIX_ADVISORY_INFO without including unistd.h first.
63 // Currently, let's include it only on unix systems:
64 #if defined(__unix__) || defined(__unix)
65 #include <unistd.h>
66 #if ((defined __QNXNTO__) || (defined _GNU_SOURCE) || (defined __PGI) || ((defined _XOPEN_SOURCE) && (_XOPEN_SOURCE >= 600))) && (defined _POSIX_ADVISORY_INFO) && (_POSIX_ADVISORY_INFO > 0)
67 #define EIGEN_HAS_POSIX_MEMALIGN 1
68 #endif
69 #endif
70
71 #ifndef EIGEN_HAS_POSIX_MEMALIGN
72 #define EIGEN_HAS_POSIX_MEMALIGN 0
73 #endif
74
75 #if defined(EIGEN_VECTORIZE_SSE) && !defined(EIGEN_ANDROID_POSIX_MEMALIGN_WR)
76 #define EIGEN_HAS_MM_MALLOC 1
77 #else
78 #define EIGEN_HAS_MM_MALLOC 0
79 #endif
80
81 namespace Eigen {
82
83 namespace internal {
84
throw_std_bad_alloc()85 inline void throw_std_bad_alloc()
86 {
87 #ifdef EIGEN_EXCEPTIONS
88 throw std::bad_alloc();
89 #else
90 std::size_t huge = -1;
91 new int[huge];
92 #endif
93 }
94
95 /*****************************************************************************
96 *** Implementation of handmade aligned functions ***
97 *****************************************************************************/
98
99 /* ----- Hand made implementations of aligned malloc/free and realloc ----- */
100
101 /** \internal Like malloc, but the returned pointer is guaranteed to be 16-byte aligned.
102 * Fast, but wastes 16 additional bytes of memory. Does not throw any exception.
103 */
handmade_aligned_malloc(std::size_t size)104 inline void* handmade_aligned_malloc(std::size_t size)
105 {
106 void *original = std::malloc(size+16);
107 if (original == 0) return 0;
108 void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(15))) + 16);
109 *(reinterpret_cast<void**>(aligned) - 1) = original;
110 return aligned;
111 }
112
113 /** \internal Frees memory allocated with handmade_aligned_malloc */
handmade_aligned_free(void * ptr)114 inline void handmade_aligned_free(void *ptr)
115 {
116 if (ptr) std::free(*(reinterpret_cast<void**>(ptr) - 1));
117 }
118
119 /** \internal
120 * \brief Reallocates aligned memory.
121 * Since we know that our handmade version is based on std::realloc
122 * we can use std::realloc to implement efficient reallocation.
123 */
124 inline void* handmade_aligned_realloc(void* ptr, std::size_t size, std::size_t = 0)
125 {
126 if (ptr == 0) return handmade_aligned_malloc(size);
127 void *original = *(reinterpret_cast<void**>(ptr) - 1);
128 std::ptrdiff_t previous_offset = static_cast<char *>(ptr)-static_cast<char *>(original);
129 original = std::realloc(original,size+16);
130 if (original == 0) return 0;
131 void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(15))) + 16);
132 void *previous_aligned = static_cast<char *>(original)+previous_offset;
133 if(aligned!=previous_aligned)
134 std::memmove(aligned, previous_aligned, size);
135
136 *(reinterpret_cast<void**>(aligned) - 1) = original;
137 return aligned;
138 }
139
140 /*****************************************************************************
141 *** Implementation of generic aligned realloc (when no realloc can be used)***
142 *****************************************************************************/
143
144 void* aligned_malloc(std::size_t size);
145 void aligned_free(void *ptr);
146
147 /** \internal
148 * \brief Reallocates aligned memory.
149 * Allows reallocation with aligned ptr types. This implementation will
150 * always create a new memory chunk and copy the old data.
151 */
generic_aligned_realloc(void * ptr,size_t size,size_t old_size)152 inline void* generic_aligned_realloc(void* ptr, size_t size, size_t old_size)
153 {
154 if (ptr==0)
155 return aligned_malloc(size);
156
157 if (size==0)
158 {
159 aligned_free(ptr);
160 return 0;
161 }
162
163 void* newptr = aligned_malloc(size);
164 if (newptr == 0)
165 {
166 #ifdef EIGEN_HAS_ERRNO
167 errno = ENOMEM; // according to the standard
168 #endif
169 return 0;
170 }
171
172 if (ptr != 0)
173 {
174 std::memcpy(newptr, ptr, (std::min)(size,old_size));
175 aligned_free(ptr);
176 }
177
178 return newptr;
179 }
180
181 /*****************************************************************************
182 *** Implementation of portable aligned versions of malloc/free/realloc ***
183 *****************************************************************************/
184
185 #ifdef EIGEN_NO_MALLOC
check_that_malloc_is_allowed()186 inline void check_that_malloc_is_allowed()
187 {
188 eigen_assert(false && "heap allocation is forbidden (EIGEN_NO_MALLOC is defined)");
189 }
190 #elif defined EIGEN_RUNTIME_NO_MALLOC
191 inline bool is_malloc_allowed_impl(bool update, bool new_value = false)
192 {
193 static bool value = true;
194 if (update == 1)
195 value = new_value;
196 return value;
197 }
is_malloc_allowed()198 inline bool is_malloc_allowed() { return is_malloc_allowed_impl(false); }
set_is_malloc_allowed(bool new_value)199 inline bool set_is_malloc_allowed(bool new_value) { return is_malloc_allowed_impl(true, new_value); }
check_that_malloc_is_allowed()200 inline void check_that_malloc_is_allowed()
201 {
202 eigen_assert(is_malloc_allowed() && "heap allocation is forbidden (EIGEN_RUNTIME_NO_MALLOC is defined and g_is_malloc_allowed is false)");
203 }
204 #else
check_that_malloc_is_allowed()205 inline void check_that_malloc_is_allowed()
206 {}
207 #endif
208
209 /** \internal Allocates \a size bytes. The returned pointer is guaranteed to have 16 bytes alignment.
210 * On allocation error, the returned pointer is null, and std::bad_alloc is thrown.
211 */
aligned_malloc(size_t size)212 inline void* aligned_malloc(size_t size)
213 {
214 check_that_malloc_is_allowed();
215
216 void *result;
217 #if !EIGEN_ALIGN
218 result = std::malloc(size);
219 #elif EIGEN_MALLOC_ALREADY_ALIGNED
220 result = std::malloc(size);
221 #elif EIGEN_HAS_POSIX_MEMALIGN
222 if(posix_memalign(&result, 16, size)) result = 0;
223 #elif EIGEN_HAS_MM_MALLOC
224 result = _mm_malloc(size, 16);
225 #elif defined(_MSC_VER) && (!defined(_WIN32_WCE))
226 result = _aligned_malloc(size, 16);
227 #else
228 result = handmade_aligned_malloc(size);
229 #endif
230
231 if(!result && size)
232 throw_std_bad_alloc();
233
234 return result;
235 }
236
237 /** \internal Frees memory allocated with aligned_malloc. */
aligned_free(void * ptr)238 inline void aligned_free(void *ptr)
239 {
240 #if !EIGEN_ALIGN
241 std::free(ptr);
242 #elif EIGEN_MALLOC_ALREADY_ALIGNED
243 std::free(ptr);
244 #elif EIGEN_HAS_POSIX_MEMALIGN
245 std::free(ptr);
246 #elif EIGEN_HAS_MM_MALLOC
247 _mm_free(ptr);
248 #elif defined(_MSC_VER) && (!defined(_WIN32_WCE))
249 _aligned_free(ptr);
250 #else
251 handmade_aligned_free(ptr);
252 #endif
253 }
254
255 /**
256 * \internal
257 * \brief Reallocates an aligned block of memory.
258 * \throws std::bad_alloc on allocation failure
259 **/
aligned_realloc(void * ptr,size_t new_size,size_t old_size)260 inline void* aligned_realloc(void *ptr, size_t new_size, size_t old_size)
261 {
262 EIGEN_UNUSED_VARIABLE(old_size);
263
264 void *result;
265 #if !EIGEN_ALIGN
266 result = std::realloc(ptr,new_size);
267 #elif EIGEN_MALLOC_ALREADY_ALIGNED
268 result = std::realloc(ptr,new_size);
269 #elif EIGEN_HAS_POSIX_MEMALIGN
270 result = generic_aligned_realloc(ptr,new_size,old_size);
271 #elif EIGEN_HAS_MM_MALLOC
272 // The defined(_mm_free) is just here to verify that this MSVC version
273 // implements _mm_malloc/_mm_free based on the corresponding _aligned_
274 // functions. This may not always be the case and we just try to be safe.
275 #if defined(_MSC_VER) && (!defined(_WIN32_WCE)) && defined(_mm_free)
276 result = _aligned_realloc(ptr,new_size,16);
277 #else
278 result = generic_aligned_realloc(ptr,new_size,old_size);
279 #endif
280 #elif defined(_MSC_VER) && (!defined(_WIN32_WCE))
281 result = _aligned_realloc(ptr,new_size,16);
282 #else
283 result = handmade_aligned_realloc(ptr,new_size,old_size);
284 #endif
285
286 if (!result && new_size)
287 throw_std_bad_alloc();
288
289 return result;
290 }
291
292 /*****************************************************************************
293 *** Implementation of conditionally aligned functions ***
294 *****************************************************************************/
295
296 /** \internal Allocates \a size bytes. If Align is true, then the returned ptr is 16-byte-aligned.
297 * On allocation error, the returned pointer is null, and a std::bad_alloc is thrown.
298 */
conditional_aligned_malloc(size_t size)299 template<bool Align> inline void* conditional_aligned_malloc(size_t size)
300 {
301 return aligned_malloc(size);
302 }
303
304 template<> inline void* conditional_aligned_malloc<false>(size_t size)
305 {
306 check_that_malloc_is_allowed();
307
308 void *result = std::malloc(size);
309 if(!result && size)
310 throw_std_bad_alloc();
311 return result;
312 }
313
314 /** \internal Frees memory allocated with conditional_aligned_malloc */
conditional_aligned_free(void * ptr)315 template<bool Align> inline void conditional_aligned_free(void *ptr)
316 {
317 aligned_free(ptr);
318 }
319
320 template<> inline void conditional_aligned_free<false>(void *ptr)
321 {
322 std::free(ptr);
323 }
324
conditional_aligned_realloc(void * ptr,size_t new_size,size_t old_size)325 template<bool Align> inline void* conditional_aligned_realloc(void* ptr, size_t new_size, size_t old_size)
326 {
327 return aligned_realloc(ptr, new_size, old_size);
328 }
329
330 template<> inline void* conditional_aligned_realloc<false>(void* ptr, size_t new_size, size_t)
331 {
332 return std::realloc(ptr, new_size);
333 }
334
335 /*****************************************************************************
336 *** Construction/destruction of array elements ***
337 *****************************************************************************/
338
339 /** \internal Constructs the elements of an array.
340 * The \a size parameter tells on how many objects to call the constructor of T.
341 */
construct_elements_of_array(T * ptr,size_t size)342 template<typename T> inline T* construct_elements_of_array(T *ptr, size_t size)
343 {
344 for (size_t i=0; i < size; ++i) ::new (ptr + i) T;
345 return ptr;
346 }
347
348 /** \internal Destructs the elements of an array.
349 * The \a size parameters tells on how many objects to call the destructor of T.
350 */
destruct_elements_of_array(T * ptr,size_t size)351 template<typename T> inline void destruct_elements_of_array(T *ptr, size_t size)
352 {
353 // always destruct an array starting from the end.
354 if(ptr)
355 while(size) ptr[--size].~T();
356 }
357
358 /*****************************************************************************
359 *** Implementation of aligned new/delete-like functions ***
360 *****************************************************************************/
361
362 template<typename T>
check_size_for_overflow(size_t size)363 EIGEN_ALWAYS_INLINE void check_size_for_overflow(size_t size)
364 {
365 if(size > size_t(-1) / sizeof(T))
366 throw_std_bad_alloc();
367 }
368
369 /** \internal Allocates \a size objects of type T. The returned pointer is guaranteed to have 16 bytes alignment.
370 * On allocation error, the returned pointer is undefined, but a std::bad_alloc is thrown.
371 * The default constructor of T is called.
372 */
aligned_new(size_t size)373 template<typename T> inline T* aligned_new(size_t size)
374 {
375 check_size_for_overflow<T>(size);
376 T *result = reinterpret_cast<T*>(aligned_malloc(sizeof(T)*size));
377 return construct_elements_of_array(result, size);
378 }
379
conditional_aligned_new(size_t size)380 template<typename T, bool Align> inline T* conditional_aligned_new(size_t size)
381 {
382 check_size_for_overflow<T>(size);
383 T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
384 return construct_elements_of_array(result, size);
385 }
386
387 /** \internal Deletes objects constructed with aligned_new
388 * The \a size parameters tells on how many objects to call the destructor of T.
389 */
aligned_delete(T * ptr,size_t size)390 template<typename T> inline void aligned_delete(T *ptr, size_t size)
391 {
392 destruct_elements_of_array<T>(ptr, size);
393 aligned_free(ptr);
394 }
395
396 /** \internal Deletes objects constructed with conditional_aligned_new
397 * The \a size parameters tells on how many objects to call the destructor of T.
398 */
conditional_aligned_delete(T * ptr,size_t size)399 template<typename T, bool Align> inline void conditional_aligned_delete(T *ptr, size_t size)
400 {
401 destruct_elements_of_array<T>(ptr, size);
402 conditional_aligned_free<Align>(ptr);
403 }
404
conditional_aligned_realloc_new(T * pts,size_t new_size,size_t old_size)405 template<typename T, bool Align> inline T* conditional_aligned_realloc_new(T* pts, size_t new_size, size_t old_size)
406 {
407 check_size_for_overflow<T>(new_size);
408 check_size_for_overflow<T>(old_size);
409 if(new_size < old_size)
410 destruct_elements_of_array(pts+new_size, old_size-new_size);
411 T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
412 if(new_size > old_size)
413 construct_elements_of_array(result+old_size, new_size-old_size);
414 return result;
415 }
416
417
conditional_aligned_new_auto(size_t size)418 template<typename T, bool Align> inline T* conditional_aligned_new_auto(size_t size)
419 {
420 if(size==0)
421 return 0; // short-cut. Also fixes Bug 884
422 check_size_for_overflow<T>(size);
423 T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
424 if(NumTraits<T>::RequireInitialization)
425 construct_elements_of_array(result, size);
426 return result;
427 }
428
conditional_aligned_realloc_new_auto(T * pts,size_t new_size,size_t old_size)429 template<typename T, bool Align> inline T* conditional_aligned_realloc_new_auto(T* pts, size_t new_size, size_t old_size)
430 {
431 check_size_for_overflow<T>(new_size);
432 check_size_for_overflow<T>(old_size);
433 if(NumTraits<T>::RequireInitialization && (new_size < old_size))
434 destruct_elements_of_array(pts+new_size, old_size-new_size);
435 T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
436 if(NumTraits<T>::RequireInitialization && (new_size > old_size))
437 construct_elements_of_array(result+old_size, new_size-old_size);
438 return result;
439 }
440
conditional_aligned_delete_auto(T * ptr,size_t size)441 template<typename T, bool Align> inline void conditional_aligned_delete_auto(T *ptr, size_t size)
442 {
443 if(NumTraits<T>::RequireInitialization)
444 destruct_elements_of_array<T>(ptr, size);
445 conditional_aligned_free<Align>(ptr);
446 }
447
448 /****************************************************************************/
449
450 /** \internal Returns the index of the first element of the array that is well aligned for vectorization.
451 *
452 * \param array the address of the start of the array
453 * \param size the size of the array
454 *
455 * \note If no element of the array is well aligned, the size of the array is returned. Typically,
456 * for example with SSE, "well aligned" means 16-byte-aligned. If vectorization is disabled or if the
457 * packet size for the given scalar type is 1, then everything is considered well-aligned.
458 *
459 * \note If the scalar type is vectorizable, we rely on the following assumptions: sizeof(Scalar) is a
460 * power of 2, the packet size in bytes is also a power of 2, and is a multiple of sizeof(Scalar). On the
461 * other hand, we do not assume that the array address is a multiple of sizeof(Scalar), as that fails for
462 * example with Scalar=double on certain 32-bit platforms, see bug #79.
463 *
464 * There is also the variant first_aligned(const MatrixBase&) defined in DenseCoeffsBase.h.
465 */
466 template<typename Scalar, typename Index>
first_aligned(const Scalar * array,Index size)467 static inline Index first_aligned(const Scalar* array, Index size)
468 {
469 static const Index PacketSize = packet_traits<Scalar>::size;
470 static const Index PacketAlignedMask = PacketSize-1;
471
472 if(PacketSize==1)
473 {
474 // Either there is no vectorization, or a packet consists of exactly 1 scalar so that all elements
475 // of the array have the same alignment.
476 return 0;
477 }
478 else if(size_t(array) & (sizeof(Scalar)-1))
479 {
480 // There is vectorization for this scalar type, but the array is not aligned to the size of a single scalar.
481 // Consequently, no element of the array is well aligned.
482 return size;
483 }
484 else
485 {
486 return std::min<Index>( (PacketSize - (Index((size_t(array)/sizeof(Scalar))) & PacketAlignedMask))
487 & PacketAlignedMask, size);
488 }
489 }
490
491 /** \internal Returns the smallest integer multiple of \a base and greater or equal to \a size
492 */
493 template<typename Index>
first_multiple(Index size,Index base)494 inline static Index first_multiple(Index size, Index base)
495 {
496 return ((size+base-1)/base)*base;
497 }
498
499 // std::copy is much slower than memcpy, so let's introduce a smart_copy which
500 // use memcpy on trivial types, i.e., on types that does not require an initialization ctor.
501 template<typename T, bool UseMemcpy> struct smart_copy_helper;
502
smart_copy(const T * start,const T * end,T * target)503 template<typename T> void smart_copy(const T* start, const T* end, T* target)
504 {
505 smart_copy_helper<T,!NumTraits<T>::RequireInitialization>::run(start, end, target);
506 }
507
508 template<typename T> struct smart_copy_helper<T,true> {
509 static inline void run(const T* start, const T* end, T* target)
510 { memcpy(target, start, std::ptrdiff_t(end)-std::ptrdiff_t(start)); }
511 };
512
513 template<typename T> struct smart_copy_helper<T,false> {
514 static inline void run(const T* start, const T* end, T* target)
515 { std::copy(start, end, target); }
516 };
517
518
519 /*****************************************************************************
520 *** Implementation of runtime stack allocation (falling back to malloc) ***
521 *****************************************************************************/
522
523 // you can overwrite Eigen's default behavior regarding alloca by defining EIGEN_ALLOCA
524 // to the appropriate stack allocation function
525 #ifndef EIGEN_ALLOCA
526 #if (defined __linux__) || (defined __APPLE__) || (defined alloca)
527 #define EIGEN_ALLOCA alloca
528 #elif defined(_MSC_VER)
529 #define EIGEN_ALLOCA _alloca
530 #endif
531 #endif
532
533 // This helper class construct the allocated memory, and takes care of destructing and freeing the handled data
534 // at destruction time. In practice this helper class is mainly useful to avoid memory leak in case of exceptions.
535 template<typename T> class aligned_stack_memory_handler
536 {
537 public:
538 /* Creates a stack_memory_handler responsible for the buffer \a ptr of size \a size.
539 * Note that \a ptr can be 0 regardless of the other parameters.
540 * This constructor takes care of constructing/initializing the elements of the buffer if required by the scalar type T (see NumTraits<T>::RequireInitialization).
541 * In this case, the buffer elements will also be destructed when this handler will be destructed.
542 * Finally, if \a dealloc is true, then the pointer \a ptr is freed.
543 **/
544 aligned_stack_memory_handler(T* ptr, size_t size, bool dealloc)
545 : m_ptr(ptr), m_size(size), m_deallocate(dealloc)
546 {
547 if(NumTraits<T>::RequireInitialization && m_ptr)
548 Eigen::internal::construct_elements_of_array(m_ptr, size);
549 }
550 ~aligned_stack_memory_handler()
551 {
552 if(NumTraits<T>::RequireInitialization && m_ptr)
553 Eigen::internal::destruct_elements_of_array<T>(m_ptr, m_size);
554 if(m_deallocate)
555 Eigen::internal::aligned_free(m_ptr);
556 }
557 protected:
558 T* m_ptr;
559 size_t m_size;
560 bool m_deallocate;
561 };
562
563 } // end namespace internal
564
565 /** \internal
566 * Declares, allocates and construct an aligned buffer named NAME of SIZE elements of type TYPE on the stack
567 * if SIZE is smaller than EIGEN_STACK_ALLOCATION_LIMIT, and if stack allocation is supported by the platform
568 * (currently, this is Linux and Visual Studio only). Otherwise the memory is allocated on the heap.
569 * The allocated buffer is automatically deleted when exiting the scope of this declaration.
570 * If BUFFER is non null, then the declared variable is simply an alias for BUFFER, and no allocation/deletion occurs.
571 * Here is an example:
572 * \code
573 * {
574 * ei_declare_aligned_stack_constructed_variable(float,data,size,0);
575 * // use data[0] to data[size-1]
576 * }
577 * \endcode
578 * The underlying stack allocation function can controlled with the EIGEN_ALLOCA preprocessor token.
579 */
580 #ifdef EIGEN_ALLOCA
581
582 #if defined(__arm__) || defined(_WIN32)
583 #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast<void*>((reinterpret_cast<size_t>(EIGEN_ALLOCA(SIZE+16)) & ~(size_t(15))) + 16)
584 #else
585 #define EIGEN_ALIGNED_ALLOCA EIGEN_ALLOCA
586 #endif
587
588 #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
589 Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \
590 TYPE* NAME = (BUFFER)!=0 ? (BUFFER) \
591 : reinterpret_cast<TYPE*>( \
592 (sizeof(TYPE)*SIZE<=EIGEN_STACK_ALLOCATION_LIMIT) ? EIGEN_ALIGNED_ALLOCA(sizeof(TYPE)*SIZE) \
593 : Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE) ); \
594 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)
595
596 #else
597
598 #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
599 Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \
600 TYPE* NAME = (BUFFER)!=0 ? BUFFER : reinterpret_cast<TYPE*>(Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE)); \
601 Eigen::internal::aligned_stack_memory_handler<TYPE> EIGEN_CAT(NAME,_stack_memory_destructor)((BUFFER)==0 ? NAME : 0,SIZE,true)
602
603 #endif
604
605
606 /*****************************************************************************
607 *** Implementation of EIGEN_MAKE_ALIGNED_OPERATOR_NEW [_IF] ***
608 *****************************************************************************/
609
610 #if EIGEN_ALIGN
611 #ifdef EIGEN_EXCEPTIONS
612 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
613 void* operator new(size_t size, const std::nothrow_t&) throw() { \
614 try { return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); } \
615 catch (...) { return 0; } \
616 }
617 #else
618 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
619 void* operator new(size_t size, const std::nothrow_t&) throw() { \
620 return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
621 }
622 #endif
623
624 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign) \
625 void *operator new(size_t size) { \
626 return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
627 } \
628 void *operator new[](size_t size) { \
629 return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
630 } \
631 void operator delete(void * ptr) throw() { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
632 void operator delete[](void * ptr) throw() { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
633 /* in-place new and delete. since (at least afaik) there is no actual */ \
634 /* memory allocated we can safely let the default implementation handle */ \
635 /* this particular case. */ \
636 static void *operator new(size_t size, void *ptr) { return ::operator new(size,ptr); } \
637 static void *operator new[](size_t size, void* ptr) { return ::operator new[](size,ptr); } \
638 void operator delete(void * memory, void *ptr) throw() { return ::operator delete(memory,ptr); } \
639 void operator delete[](void * memory, void *ptr) throw() { return ::operator delete[](memory,ptr); } \
640 /* nothrow-new (returns zero instead of std::bad_alloc) */ \
641 EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
642 void operator delete(void *ptr, const std::nothrow_t&) throw() { \
643 Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); \
644 } \
645 typedef void eigen_aligned_operator_new_marker_type;
646 #else
647 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)
648 #endif
649
650 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(true)
651 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size) \
652 EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool(((Size)!=Eigen::Dynamic) && ((sizeof(Scalar)*(Size))%16==0)))
653
654 /****************************************************************************/
655
656 /** \class aligned_allocator
657 * \ingroup Core_Module
658 *
659 * \brief STL compatible allocator to use with with 16 byte aligned types
660 *
661 * Example:
662 * \code
663 * // Matrix4f requires 16 bytes alignment:
664 * std::map< int, Matrix4f, std::less<int>,
665 * aligned_allocator<std::pair<const int, Matrix4f> > > my_map_mat4;
666 * // Vector3f does not require 16 bytes alignment, no need to use Eigen's allocator:
667 * std::map< int, Vector3f > my_map_vec3;
668 * \endcode
669 *
670 * \sa \ref TopicStlContainers.
671 */
672 template<class T>
673 class aligned_allocator
674 {
675 public:
676 typedef size_t size_type;
677 typedef std::ptrdiff_t difference_type;
678 typedef T* pointer;
679 typedef const T* const_pointer;
680 typedef T& reference;
681 typedef const T& const_reference;
682 typedef T value_type;
683
684 template<class U>
685 struct rebind
686 {
687 typedef aligned_allocator<U> other;
688 };
689
690 pointer address( reference value ) const
691 {
692 return &value;
693 }
694
695 const_pointer address( const_reference value ) const
696 {
697 return &value;
698 }
699
700 aligned_allocator()
701 {
702 }
703
704 aligned_allocator( const aligned_allocator& )
705 {
706 }
707
708 template<class U>
709 aligned_allocator( const aligned_allocator<U>& )
710 {
711 }
712
713 ~aligned_allocator()
714 {
715 }
716
717 size_type max_size() const
718 {
719 return (std::numeric_limits<size_type>::max)();
720 }
721
722 pointer allocate( size_type num, const void* hint = 0 )
723 {
724 EIGEN_UNUSED_VARIABLE(hint);
725 internal::check_size_for_overflow<T>(num);
726 return static_cast<pointer>( internal::aligned_malloc( num * sizeof(T) ) );
727 }
728
729 void construct( pointer p, const T& value )
730 {
731 ::new( p ) T( value );
732 }
733
734 void destroy( pointer p )
735 {
736 p->~T();
737 }
738
739 void deallocate( pointer p, size_type /*num*/ )
740 {
741 internal::aligned_free( p );
742 }
743
744 bool operator!=(const aligned_allocator<T>& ) const
745 { return false; }
746
747 bool operator==(const aligned_allocator<T>& ) const
748 { return true; }
749 };
750
751 //---------- Cache sizes ----------
752
753 #if !defined(EIGEN_NO_CPUID)
754 # if defined(__GNUC__) && ( defined(__i386__) || defined(__x86_64__) )
755 # if defined(__PIC__) && defined(__i386__)
756 // Case for x86 with PIC
757 # define EIGEN_CPUID(abcd,func,id) \
758 __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));
759 # elif defined(__PIC__) && defined(__x86_64__)
760 // 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.
761 // However, we cannot detect which code model is used, and the xchg overhead is negligible anyway.
762 # define EIGEN_CPUID(abcd,func,id) \
763 __asm__ __volatile__ ("xchg{q}\t{%%}rbx, %q1; cpuid; xchg{q}\t{%%}rbx, %q1": "=a" (abcd[0]), "=&r" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "0" (func), "2" (id));
764 # else
765 // Case for x86_64 or x86 w/o PIC
766 # define EIGEN_CPUID(abcd,func,id) \
767 __asm__ __volatile__ ("cpuid": "=a" (abcd[0]), "=b" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "0" (func), "2" (id) );
768 # endif
769 # elif defined(_MSC_VER)
770 # if (_MSC_VER > 1500) && ( defined(_M_IX86) || defined(_M_X64) )
771 # define EIGEN_CPUID(abcd,func,id) __cpuidex((int*)abcd,func,id)
772 # endif
773 # endif
774 #endif
775
776 namespace internal {
777
778 #ifdef EIGEN_CPUID
779
780 inline bool cpuid_is_vendor(int abcd[4], const int vendor[3])
781 {
782 return abcd[1]==vendor[0] && abcd[3]==vendor[1] && abcd[2]==vendor[2];
783 }
784
785 inline void queryCacheSizes_intel_direct(int& l1, int& l2, int& l3)
786 {
787 int abcd[4];
788 l1 = l2 = l3 = 0;
789 int cache_id = 0;
790 int cache_type = 0;
791 do {
792 abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
793 EIGEN_CPUID(abcd,0x4,cache_id);
794 cache_type = (abcd[0] & 0x0F) >> 0;
795 if(cache_type==1||cache_type==3) // data or unified cache
796 {
797 int cache_level = (abcd[0] & 0xE0) >> 5; // A[7:5]
798 int ways = (abcd[1] & 0xFFC00000) >> 22; // B[31:22]
799 int partitions = (abcd[1] & 0x003FF000) >> 12; // B[21:12]
800 int line_size = (abcd[1] & 0x00000FFF) >> 0; // B[11:0]
801 int sets = (abcd[2]); // C[31:0]
802
803 int cache_size = (ways+1) * (partitions+1) * (line_size+1) * (sets+1);
804
805 switch(cache_level)
806 {
807 case 1: l1 = cache_size; break;
808 case 2: l2 = cache_size; break;
809 case 3: l3 = cache_size; break;
810 default: break;
811 }
812 }
813 cache_id++;
814 } while(cache_type>0 && cache_id<16);
815 }
816
817 inline void queryCacheSizes_intel_codes(int& l1, int& l2, int& l3)
818 {
819 int abcd[4];
820 abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
821 l1 = l2 = l3 = 0;
822 EIGEN_CPUID(abcd,0x00000002,0);
823 unsigned char * bytes = reinterpret_cast<unsigned char *>(abcd)+2;
824 bool check_for_p2_core2 = false;
825 for(int i=0; i<14; ++i)
826 {
827 switch(bytes[i])
828 {
829 case 0x0A: l1 = 8; break; // 0Ah data L1 cache, 8 KB, 2 ways, 32 byte lines
830 case 0x0C: l1 = 16; break; // 0Ch data L1 cache, 16 KB, 4 ways, 32 byte lines
831 case 0x0E: l1 = 24; break; // 0Eh data L1 cache, 24 KB, 6 ways, 64 byte lines
832 case 0x10: l1 = 16; break; // 10h data L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
833 case 0x15: l1 = 16; break; // 15h code L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
834 case 0x2C: l1 = 32; break; // 2Ch data L1 cache, 32 KB, 8 ways, 64 byte lines
835 case 0x30: l1 = 32; break; // 30h code L1 cache, 32 KB, 8 ways, 64 byte lines
836 case 0x60: l1 = 16; break; // 60h data L1 cache, 16 KB, 8 ways, 64 byte lines, sectored
837 case 0x66: l1 = 8; break; // 66h data L1 cache, 8 KB, 4 ways, 64 byte lines, sectored
838 case 0x67: l1 = 16; break; // 67h data L1 cache, 16 KB, 4 ways, 64 byte lines, sectored
839 case 0x68: l1 = 32; break; // 68h data L1 cache, 32 KB, 4 ways, 64 byte lines, sectored
840 case 0x1A: l2 = 96; break; // code and data L2 cache, 96 KB, 6 ways, 64 byte lines (IA-64)
841 case 0x22: l3 = 512; break; // code and data L3 cache, 512 KB, 4 ways (!), 64 byte lines, dual-sectored
842 case 0x23: l3 = 1024; break; // code and data L3 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
843 case 0x25: l3 = 2048; break; // code and data L3 cache, 2048 KB, 8 ways, 64 byte lines, dual-sectored
844 case 0x29: l3 = 4096; break; // code and data L3 cache, 4096 KB, 8 ways, 64 byte lines, dual-sectored
845 case 0x39: l2 = 128; break; // code and data L2 cache, 128 KB, 4 ways, 64 byte lines, sectored
846 case 0x3A: l2 = 192; break; // code and data L2 cache, 192 KB, 6 ways, 64 byte lines, sectored
847 case 0x3B: l2 = 128; break; // code and data L2 cache, 128 KB, 2 ways, 64 byte lines, sectored
848 case 0x3C: l2 = 256; break; // code and data L2 cache, 256 KB, 4 ways, 64 byte lines, sectored
849 case 0x3D: l2 = 384; break; // code and data L2 cache, 384 KB, 6 ways, 64 byte lines, sectored
850 case 0x3E: l2 = 512; break; // code and data L2 cache, 512 KB, 4 ways, 64 byte lines, sectored
851 case 0x40: l2 = 0; break; // no integrated L2 cache (P6 core) or L3 cache (P4 core)
852 case 0x41: l2 = 128; break; // code and data L2 cache, 128 KB, 4 ways, 32 byte lines
853 case 0x42: l2 = 256; break; // code and data L2 cache, 256 KB, 4 ways, 32 byte lines
854 case 0x43: l2 = 512; break; // code and data L2 cache, 512 KB, 4 ways, 32 byte lines
855 case 0x44: l2 = 1024; break; // code and data L2 cache, 1024 KB, 4 ways, 32 byte lines
856 case 0x45: l2 = 2048; break; // code and data L2 cache, 2048 KB, 4 ways, 32 byte lines
857 case 0x46: l3 = 4096; break; // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines
858 case 0x47: l3 = 8192; break; // code and data L3 cache, 8192 KB, 8 ways, 64 byte lines
859 case 0x48: l2 = 3072; break; // code and data L2 cache, 3072 KB, 12 ways, 64 byte lines
860 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
861 case 0x4A: l3 = 6144; break; // code and data L3 cache, 6144 KB, 12 ways, 64 byte lines
862 case 0x4B: l3 = 8192; break; // code and data L3 cache, 8192 KB, 16 ways, 64 byte lines
863 case 0x4C: l3 = 12288; break; // code and data L3 cache, 12288 KB, 12 ways, 64 byte lines
864 case 0x4D: l3 = 16384; break; // code and data L3 cache, 16384 KB, 16 ways, 64 byte lines
865 case 0x4E: l2 = 6144; break; // code and data L2 cache, 6144 KB, 24 ways, 64 byte lines
866 case 0x78: l2 = 1024; break; // code and data L2 cache, 1024 KB, 4 ways, 64 byte lines
867 case 0x79: l2 = 128; break; // code and data L2 cache, 128 KB, 8 ways, 64 byte lines, dual-sectored
868 case 0x7A: l2 = 256; break; // code and data L2 cache, 256 KB, 8 ways, 64 byte lines, dual-sectored
869 case 0x7B: l2 = 512; break; // code and data L2 cache, 512 KB, 8 ways, 64 byte lines, dual-sectored
870 case 0x7C: l2 = 1024; break; // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
871 case 0x7D: l2 = 2048; break; // code and data L2 cache, 2048 KB, 8 ways, 64 byte lines
872 case 0x7E: l2 = 256; break; // code and data L2 cache, 256 KB, 8 ways, 128 byte lines, sect. (IA-64)
873 case 0x7F: l2 = 512; break; // code and data L2 cache, 512 KB, 2 ways, 64 byte lines
874 case 0x80: l2 = 512; break; // code and data L2 cache, 512 KB, 8 ways, 64 byte lines
875 case 0x81: l2 = 128; break; // code and data L2 cache, 128 KB, 8 ways, 32 byte lines
876 case 0x82: l2 = 256; break; // code and data L2 cache, 256 KB, 8 ways, 32 byte lines
877 case 0x83: l2 = 512; break; // code and data L2 cache, 512 KB, 8 ways, 32 byte lines
878 case 0x84: l2 = 1024; break; // code and data L2 cache, 1024 KB, 8 ways, 32 byte lines
879 case 0x85: l2 = 2048; break; // code and data L2 cache, 2048 KB, 8 ways, 32 byte lines
880 case 0x86: l2 = 512; break; // code and data L2 cache, 512 KB, 4 ways, 64 byte lines
881 case 0x87: l2 = 1024; break; // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines
882 case 0x88: l3 = 2048; break; // code and data L3 cache, 2048 KB, 4 ways, 64 byte lines (IA-64)
883 case 0x89: l3 = 4096; break; // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines (IA-64)
884 case 0x8A: l3 = 8192; break; // code and data L3 cache, 8192 KB, 4 ways, 64 byte lines (IA-64)
885 case 0x8D: l3 = 3072; break; // code and data L3 cache, 3072 KB, 12 ways, 128 byte lines (IA-64)
886
887 default: break;
888 }
889 }
890 if(check_for_p2_core2 && l2 == l3)
891 l3 = 0;
892 l1 *= 1024;
893 l2 *= 1024;
894 l3 *= 1024;
895 }
896
897 inline void queryCacheSizes_intel(int& l1, int& l2, int& l3, int max_std_funcs)
898 {
899 if(max_std_funcs>=4)
900 queryCacheSizes_intel_direct(l1,l2,l3);
901 else
902 queryCacheSizes_intel_codes(l1,l2,l3);
903 }
904
905 inline void queryCacheSizes_amd(int& l1, int& l2, int& l3)
906 {
907 int abcd[4];
908 abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
909 EIGEN_CPUID(abcd,0x80000005,0);
910 l1 = (abcd[2] >> 24) * 1024; // C[31:24] = L1 size in KB
911 abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
912 EIGEN_CPUID(abcd,0x80000006,0);
913 l2 = (abcd[2] >> 16) * 1024; // C[31;16] = l2 cache size in KB
914 l3 = ((abcd[3] & 0xFFFC000) >> 18) * 512 * 1024; // D[31;18] = l3 cache size in 512KB
915 }
916 #endif
917
918 /** \internal
919 * Queries and returns the cache sizes in Bytes of the L1, L2, and L3 data caches respectively */
920 inline void queryCacheSizes(int& l1, int& l2, int& l3)
921 {
922 #ifdef EIGEN_CPUID
923 int abcd[4];
924 const int GenuineIntel[] = {0x756e6547, 0x49656e69, 0x6c65746e};
925 const int AuthenticAMD[] = {0x68747541, 0x69746e65, 0x444d4163};
926 const int AMDisbetter_[] = {0x69444d41, 0x74656273, 0x21726574}; // "AMDisbetter!"
927
928 // identify the CPU vendor
929 EIGEN_CPUID(abcd,0x0,0);
930 int max_std_funcs = abcd[1];
931 if(cpuid_is_vendor(abcd,GenuineIntel))
932 queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
933 else if(cpuid_is_vendor(abcd,AuthenticAMD) || cpuid_is_vendor(abcd,AMDisbetter_))
934 queryCacheSizes_amd(l1,l2,l3);
935 else
936 // by default let's use Intel's API
937 queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
938
939 // here is the list of other vendors:
940 // ||cpuid_is_vendor(abcd,"VIA VIA VIA ")
941 // ||cpuid_is_vendor(abcd,"CyrixInstead")
942 // ||cpuid_is_vendor(abcd,"CentaurHauls")
943 // ||cpuid_is_vendor(abcd,"GenuineTMx86")
944 // ||cpuid_is_vendor(abcd,"TransmetaCPU")
945 // ||cpuid_is_vendor(abcd,"RiseRiseRise")
946 // ||cpuid_is_vendor(abcd,"Geode by NSC")
947 // ||cpuid_is_vendor(abcd,"SiS SiS SiS ")
948 // ||cpuid_is_vendor(abcd,"UMC UMC UMC ")
949 // ||cpuid_is_vendor(abcd,"NexGenDriven")
950 #else
951 l1 = l2 = l3 = -1;
952 #endif
953 }
954
955 /** \internal
956 * \returns the size in Bytes of the L1 data cache */
957 inline int queryL1CacheSize()
958 {
959 int l1(-1), l2, l3;
960 queryCacheSizes(l1,l2,l3);
961 return l1;
962 }
963
964 /** \internal
965 * \returns the size in Bytes of the L2 or L3 cache if this later is present */
966 inline int queryTopLevelCacheSize()
967 {
968 int l1, l2(-1), l3(-1);
969 queryCacheSizes(l1,l2,l3);
970 return (std::max)(l2,l3);
971 }
972
973 } // end namespace internal
974
975 } // end namespace Eigen
976
977 #endif // EIGEN_MEMORY_H
978