Eigen  3.2.92
Memory.h
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2008-2015 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 // Copyright (C) 2013 Pavel Holoborodko <pavel@holoborodko.com>
10 //
11 // This Source Code Form is subject to the terms of the Mozilla
12 // Public License v. 2.0. If a copy of the MPL was not distributed
13 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
14 
15 
16 /*****************************************************************************
17 *** Platform checks for aligned malloc functions ***
18 *****************************************************************************/
19 
20 #ifndef EIGEN_MEMORY_H
21 #define EIGEN_MEMORY_H
22 
23 #ifndef EIGEN_MALLOC_ALREADY_ALIGNED
24 
25 // Try to determine automatically if malloc is already aligned.
26 
27 // On 64-bit systems, glibc's malloc returns 16-byte-aligned pointers, see:
28 // http://www.gnu.org/s/libc/manual/html_node/Aligned-Memory-Blocks.html
29 // This is true at least since glibc 2.8.
30 // This leaves the question how to detect 64-bit. According to this document,
31 // http://gcc.fyxm.net/summit/2003/Porting%20to%2064%20bit.pdf
32 // page 114, "[The] LP64 model [...] is used by all 64-bit UNIX ports" so it's indeed
33 // quite safe, at least within the context of glibc, to equate 64-bit with LP64.
34 #if defined(__GLIBC__) && ((__GLIBC__>=2 && __GLIBC_MINOR__ >= 8) || __GLIBC__>2) \
35  && defined(__LP64__) && ! defined( __SANITIZE_ADDRESS__ ) && (EIGEN_DEFAULT_ALIGN_BYTES == 16)
36  #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 1
37 #else
38  #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 0
39 #endif
40 
41 // FreeBSD 6 seems to have 16-byte aligned malloc
42 // See http://svn.freebsd.org/viewvc/base/stable/6/lib/libc/stdlib/malloc.c?view=markup
43 // FreeBSD 7 seems to have 16-byte aligned malloc except on ARM and MIPS architectures
44 // See http://svn.freebsd.org/viewvc/base/stable/7/lib/libc/stdlib/malloc.c?view=markup
45 #if defined(__FreeBSD__) && !(EIGEN_ARCH_ARM || EIGEN_ARCH_MIPS) && (EIGEN_DEFAULT_ALIGN_BYTES == 16)
46  #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 1
47 #else
48  #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 0
49 #endif
50 
51 #if (EIGEN_OS_MAC && (EIGEN_DEFAULT_ALIGN_BYTES == 16)) \
52  || (EIGEN_OS_WIN64 && (EIGEN_DEFAULT_ALIGN_BYTES == 16)) \
53  || EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED \
54  || EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED
55  #define EIGEN_MALLOC_ALREADY_ALIGNED 1
56 #else
57  #define EIGEN_MALLOC_ALREADY_ALIGNED 0
58 #endif
59 
60 #endif
61 
62 #ifndef EIGEN_HAS_POSIX_MEMALIGN
63  // See bug 554 (http://eigen.tuxfamily.org/bz/show_bug.cgi?id=554)
64  // It seems to be unsafe to check _POSIX_ADVISORY_INFO without including unistd.h first.
65  // Currently, let's include it only on unix systems:
66  #if EIGEN_OS_UNIX && !(EIGEN_OS_SUN || EIGEN_OS_SOLARIS)
67  #include <unistd.h>
68  #if (EIGEN_OS_QNX || (defined _GNU_SOURCE) || EIGEN_COMP_PGI || ((defined _XOPEN_SOURCE) && (_XOPEN_SOURCE >= 600))) && (defined _POSIX_ADVISORY_INFO) && (_POSIX_ADVISORY_INFO > 0)
69  #define EIGEN_HAS_POSIX_MEMALIGN 1
70  #endif
71  #endif
72 
73  #ifndef EIGEN_HAS_POSIX_MEMALIGN
74  #define EIGEN_HAS_POSIX_MEMALIGN 0
75  #endif
76 #endif
77 
78 #if defined EIGEN_VECTORIZE_SSE || defined EIGEN_VECTORIZE_AVX
79  #define EIGEN_HAS_MM_MALLOC 1
80 #else
81  #define EIGEN_HAS_MM_MALLOC 0
82 #endif
83 
84 namespace Eigen {
85 
86 namespace internal {
87 
88 EIGEN_DEVICE_FUNC
89 inline void throw_std_bad_alloc()
90 {
91  #ifdef EIGEN_EXCEPTIONS
92  throw std::bad_alloc();
93  #else
94  std::size_t huge = static_cast<std::size_t>(-1);
95  new int[huge];
96  #endif
97 }
98 
99 /*****************************************************************************
100 *** Implementation of handmade aligned functions ***
101 *****************************************************************************/
102 
103 /* ----- Hand made implementations of aligned malloc/free and realloc ----- */
104 
108 inline void* handmade_aligned_malloc(std::size_t size)
109 {
110  void *original = std::malloc(size+EIGEN_DEFAULT_ALIGN_BYTES);
111  if (original == 0) return 0;
112  void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1))) + EIGEN_DEFAULT_ALIGN_BYTES);
113  *(reinterpret_cast<void**>(aligned) - 1) = original;
114  return aligned;
115 }
116 
118 inline void handmade_aligned_free(void *ptr)
119 {
120  if (ptr) std::free(*(reinterpret_cast<void**>(ptr) - 1));
121 }
122 
128 inline void* handmade_aligned_realloc(void* ptr, std::size_t size, std::size_t = 0)
129 {
130  if (ptr == 0) return handmade_aligned_malloc(size);
131  void *original = *(reinterpret_cast<void**>(ptr) - 1);
132  std::ptrdiff_t previous_offset = static_cast<char *>(ptr)-static_cast<char *>(original);
133  original = std::realloc(original,size+EIGEN_DEFAULT_ALIGN_BYTES);
134  if (original == 0) return 0;
135  void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1))) + EIGEN_DEFAULT_ALIGN_BYTES);
136  void *previous_aligned = static_cast<char *>(original)+previous_offset;
137  if(aligned!=previous_aligned)
138  std::memmove(aligned, previous_aligned, size);
139 
140  *(reinterpret_cast<void**>(aligned) - 1) = original;
141  return aligned;
142 }
143 
144 /*****************************************************************************
145 *** Implementation of generic aligned realloc (when no realloc can be used)***
146 *****************************************************************************/
147 
148 EIGEN_DEVICE_FUNC void* aligned_malloc(std::size_t size);
149 EIGEN_DEVICE_FUNC void aligned_free(void *ptr);
150 
156 inline void* generic_aligned_realloc(void* ptr, size_t size, size_t old_size)
157 {
158  if (ptr==0)
159  return aligned_malloc(size);
160 
161  if (size==0)
162  {
163  aligned_free(ptr);
164  return 0;
165  }
166 
167  void* newptr = aligned_malloc(size);
168  if (newptr == 0)
169  {
170  #ifdef EIGEN_HAS_ERRNO
171  errno = ENOMEM; // according to the standard
172  #endif
173  return 0;
174  }
175 
176  if (ptr != 0)
177  {
178  std::memcpy(newptr, ptr, (std::min)(size,old_size));
179  aligned_free(ptr);
180  }
181 
182  return newptr;
183 }
184 
185 /*****************************************************************************
186 *** Implementation of portable aligned versions of malloc/free/realloc ***
187 *****************************************************************************/
188 
189 #ifdef EIGEN_NO_MALLOC
190 EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed()
191 {
192  eigen_assert(false && "heap allocation is forbidden (EIGEN_NO_MALLOC is defined)");
193 }
194 #elif defined EIGEN_RUNTIME_NO_MALLOC
195 EIGEN_DEVICE_FUNC inline bool is_malloc_allowed_impl(bool update, bool new_value = false)
196 {
197  static bool value = true;
198  if (update == 1)
199  value = new_value;
200  return value;
201 }
202 EIGEN_DEVICE_FUNC inline bool is_malloc_allowed() { return is_malloc_allowed_impl(false); }
203 EIGEN_DEVICE_FUNC inline bool set_is_malloc_allowed(bool new_value) { return is_malloc_allowed_impl(true, new_value); }
204 EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed()
205 {
206  eigen_assert(is_malloc_allowed() && "heap allocation is forbidden (EIGEN_RUNTIME_NO_MALLOC is defined and g_is_malloc_allowed is false)");
207 }
208 #else
209 EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed()
210 {}
211 #endif
212 
216 EIGEN_DEVICE_FUNC inline void* aligned_malloc(size_t size)
217 {
218  check_that_malloc_is_allowed();
219 
220  void *result;
221  #if EIGEN_DEFAULT_ALIGN_BYTES==0
222  result = std::malloc(size);
223  #elif EIGEN_MALLOC_ALREADY_ALIGNED
224  result = std::malloc(size);
225  #elif EIGEN_HAS_POSIX_MEMALIGN
226  if(posix_memalign(&result, EIGEN_DEFAULT_ALIGN_BYTES, size)) result = 0;
227  #elif EIGEN_HAS_MM_MALLOC
228  result = _mm_malloc(size, EIGEN_DEFAULT_ALIGN_BYTES);
229  #elif EIGEN_OS_WIN_STRICT
230  result = _aligned_malloc(size, EIGEN_DEFAULT_ALIGN_BYTES);
231  #else
232  result = handmade_aligned_malloc(size);
233  #endif
234 
235  if(!result && size)
236  throw_std_bad_alloc();
237 
238  return result;
239 }
240 
242 EIGEN_DEVICE_FUNC inline void aligned_free(void *ptr)
243 {
244  #if EIGEN_DEFAULT_ALIGN_BYTES==0
245  std::free(ptr);
246  #elif EIGEN_MALLOC_ALREADY_ALIGNED
247  std::free(ptr);
248  #elif EIGEN_HAS_POSIX_MEMALIGN
249  std::free(ptr);
250  #elif EIGEN_HAS_MM_MALLOC
251  _mm_free(ptr);
252  #elif EIGEN_OS_WIN_STRICT
253  _aligned_free(ptr);
254  #else
255  handmade_aligned_free(ptr);
256  #endif
257 }
258 
264 inline void* aligned_realloc(void *ptr, size_t new_size, size_t old_size)
265 {
266  EIGEN_UNUSED_VARIABLE(old_size);
267 
268  void *result;
269 #if EIGEN_DEFAULT_ALIGN_BYTES==0
270  result = std::realloc(ptr,new_size);
271 #elif EIGEN_MALLOC_ALREADY_ALIGNED
272  result = std::realloc(ptr,new_size);
273 #elif EIGEN_HAS_POSIX_MEMALIGN
274  result = generic_aligned_realloc(ptr,new_size,old_size);
275 #elif EIGEN_HAS_MM_MALLOC
276  // The defined(_mm_free) is just here to verify that this MSVC version
277  // implements _mm_malloc/_mm_free based on the corresponding _aligned_
278  // functions. This may not always be the case and we just try to be safe.
279  #if EIGEN_OS_WIN_STRICT && defined(_mm_free)
280  result = _aligned_realloc(ptr,new_size,EIGEN_DEFAULT_ALIGN_BYTES);
281  #else
282  result = generic_aligned_realloc(ptr,new_size,old_size);
283  #endif
284 #elif EIGEN_OS_WIN_STRICT
285  result = _aligned_realloc(ptr,new_size,EIGEN_DEFAULT_ALIGN_BYTES);
286 #else
287  result = handmade_aligned_realloc(ptr,new_size,old_size);
288 #endif
289 
290  if (!result && new_size)
291  throw_std_bad_alloc();
292 
293  return result;
294 }
295 
296 /*****************************************************************************
297 *** Implementation of conditionally aligned functions ***
298 *****************************************************************************/
299 
303 template<bool Align> EIGEN_DEVICE_FUNC inline void* conditional_aligned_malloc(size_t size)
304 {
305  return aligned_malloc(size);
306 }
307 
308 template<> EIGEN_DEVICE_FUNC inline void* conditional_aligned_malloc<false>(size_t size)
309 {
310  check_that_malloc_is_allowed();
311 
312  void *result = std::malloc(size);
313  if(!result && size)
314  throw_std_bad_alloc();
315  return result;
316 }
317 
319 template<bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_free(void *ptr)
320 {
321  aligned_free(ptr);
322 }
323 
324 template<> EIGEN_DEVICE_FUNC inline void conditional_aligned_free<false>(void *ptr)
325 {
326  std::free(ptr);
327 }
328 
329 template<bool Align> inline void* conditional_aligned_realloc(void* ptr, size_t new_size, size_t old_size)
330 {
331  return aligned_realloc(ptr, new_size, old_size);
332 }
333 
334 template<> inline void* conditional_aligned_realloc<false>(void* ptr, size_t new_size, size_t)
335 {
336  return std::realloc(ptr, new_size);
337 }
338 
339 /*****************************************************************************
340 *** Construction/destruction of array elements ***
341 *****************************************************************************/
342 
346 template<typename T> EIGEN_DEVICE_FUNC inline void destruct_elements_of_array(T *ptr, size_t size)
347 {
348  // always destruct an array starting from the end.
349  if(ptr)
350  while(size) ptr[--size].~T();
351 }
352 
356 template<typename T> EIGEN_DEVICE_FUNC inline T* construct_elements_of_array(T *ptr, size_t size)
357 {
358  size_t i;
359  EIGEN_TRY
360  {
361  for (i = 0; i < size; ++i) ::new (ptr + i) T;
362  return ptr;
363  }
364  EIGEN_CATCH(...)
365  {
366  destruct_elements_of_array(ptr, i);
367  EIGEN_THROW;
368  }
369 }
370 
371 /*****************************************************************************
372 *** Implementation of aligned new/delete-like functions ***
373 *****************************************************************************/
374 
375 template<typename T>
376 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void check_size_for_overflow(size_t size)
377 {
378  if(size > size_t(-1) / sizeof(T))
379  throw_std_bad_alloc();
380 }
381 
386 template<typename T> EIGEN_DEVICE_FUNC inline T* aligned_new(size_t size)
387 {
388  check_size_for_overflow<T>(size);
389  T *result = reinterpret_cast<T*>(aligned_malloc(sizeof(T)*size));
390  EIGEN_TRY
391  {
392  return construct_elements_of_array(result, size);
393  }
394  EIGEN_CATCH(...)
395  {
396  aligned_free(result);
397  EIGEN_THROW;
398  }
399 }
400 
401 template<typename T, bool Align> EIGEN_DEVICE_FUNC inline T* conditional_aligned_new(size_t size)
402 {
403  check_size_for_overflow<T>(size);
404  T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
405  EIGEN_TRY
406  {
407  return construct_elements_of_array(result, size);
408  }
409  EIGEN_CATCH(...)
410  {
411  conditional_aligned_free<Align>(result);
412  EIGEN_THROW;
413  }
414 }
415 
419 template<typename T> EIGEN_DEVICE_FUNC inline void aligned_delete(T *ptr, size_t size)
420 {
421  destruct_elements_of_array<T>(ptr, size);
422  aligned_free(ptr);
423 }
424 
428 template<typename T, bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_delete(T *ptr, size_t size)
429 {
430  destruct_elements_of_array<T>(ptr, size);
431  conditional_aligned_free<Align>(ptr);
432 }
433 
434 template<typename T, bool Align> EIGEN_DEVICE_FUNC inline T* conditional_aligned_realloc_new(T* pts, size_t new_size, size_t old_size)
435 {
436  check_size_for_overflow<T>(new_size);
437  check_size_for_overflow<T>(old_size);
438  if(new_size < old_size)
439  destruct_elements_of_array(pts+new_size, old_size-new_size);
440  T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
441  if(new_size > old_size)
442  {
443  EIGEN_TRY
444  {
445  construct_elements_of_array(result+old_size, new_size-old_size);
446  }
447  EIGEN_CATCH(...)
448  {
449  conditional_aligned_free<Align>(result);
450  EIGEN_THROW;
451  }
452  }
453  return result;
454 }
455 
456 
457 template<typename T, bool Align> EIGEN_DEVICE_FUNC inline T* conditional_aligned_new_auto(size_t size)
458 {
459  if(size==0)
460  return 0; // short-cut. Also fixes Bug 884
461  check_size_for_overflow<T>(size);
462  T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
463  if(NumTraits<T>::RequireInitialization)
464  {
465  EIGEN_TRY
466  {
467  construct_elements_of_array(result, size);
468  }
469  EIGEN_CATCH(...)
470  {
471  conditional_aligned_free<Align>(result);
472  EIGEN_THROW;
473  }
474  }
475  return result;
476 }
477 
478 template<typename T, bool Align> inline T* conditional_aligned_realloc_new_auto(T* pts, size_t new_size, size_t old_size)
479 {
480  check_size_for_overflow<T>(new_size);
481  check_size_for_overflow<T>(old_size);
482  if(NumTraits<T>::RequireInitialization && (new_size < old_size))
483  destruct_elements_of_array(pts+new_size, old_size-new_size);
484  T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
485  if(NumTraits<T>::RequireInitialization && (new_size > old_size))
486  {
487  EIGEN_TRY
488  {
489  construct_elements_of_array(result+old_size, new_size-old_size);
490  }
491  EIGEN_CATCH(...)
492  {
493  conditional_aligned_free<Align>(result);
494  EIGEN_THROW;
495  }
496  }
497  return result;
498 }
499 
500 template<typename T, bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_delete_auto(T *ptr, size_t size)
501 {
502  if(NumTraits<T>::RequireInitialization)
503  destruct_elements_of_array<T>(ptr, size);
504  conditional_aligned_free<Align>(ptr);
505 }
506 
507 /****************************************************************************/
508 
526 template<int Alignment, typename Scalar, typename Index>
527 inline Index first_aligned(const Scalar* array, Index size)
528 {
529  static const Index ScalarSize = sizeof(Scalar);
530  static const Index AlignmentSize = Alignment / ScalarSize;
531  static const Index AlignmentMask = AlignmentSize-1;
532 
533  if(AlignmentSize<=1)
534  {
535  // Either the requested alignment if smaller than a scalar, or it exactly match a 1 scalar
536  // so that all elements of the array have the same alignment.
537  return 0;
538  }
539  else if( (std::size_t(array) & (sizeof(Scalar)-1)) || (Alignment%ScalarSize)!=0)
540  {
541  // The array is not aligned to the size of a single scalar, or the requested alignment is not a multiple of the scalar size.
542  // Consequently, no element of the array is well aligned.
543  return size;
544  }
545  else
546  {
547  return std::min<Index>( (AlignmentSize - (Index((std::size_t(array)/sizeof(Scalar))) & AlignmentMask)) & AlignmentMask, size);
548  }
549 }
550 
553 template<typename Scalar, typename Index>
554 inline Index first_default_aligned(const Scalar* array, Index size)
555 {
556  typedef typename packet_traits<Scalar>::type DefaultPacketType;
557  return first_aligned<unpacket_traits<DefaultPacketType>::alignment>(array, size);
558 }
559 
562 template<typename Index>
563 inline Index first_multiple(Index size, Index base)
564 {
565  return ((size+base-1)/base)*base;
566 }
567 
568 // std::copy is much slower than memcpy, so let's introduce a smart_copy which
569 // use memcpy on trivial types, i.e., on types that does not require an initialization ctor.
570 template<typename T, bool UseMemcpy> struct smart_copy_helper;
571 
572 template<typename T> EIGEN_DEVICE_FUNC void smart_copy(const T* start, const T* end, T* target)
573 {
574  smart_copy_helper<T,!NumTraits<T>::RequireInitialization>::run(start, end, target);
575 }
576 
577 template<typename T> struct smart_copy_helper<T,true> {
578  EIGEN_DEVICE_FUNC static inline void run(const T* start, const T* end, T* target)
579  { memcpy(target, start, std::ptrdiff_t(end)-std::ptrdiff_t(start)); }
580 };
581 
582 template<typename T> struct smart_copy_helper<T,false> {
583  EIGEN_DEVICE_FUNC static inline void run(const T* start, const T* end, T* target)
584  { std::copy(start, end, target); }
585 };
586 
587 // intelligent memmove. falls back to std::memmove for POD types, uses std::copy otherwise.
588 template<typename T, bool UseMemmove> struct smart_memmove_helper;
589 
590 template<typename T> void smart_memmove(const T* start, const T* end, T* target)
591 {
592  smart_memmove_helper<T,!NumTraits<T>::RequireInitialization>::run(start, end, target);
593 }
594 
595 template<typename T> struct smart_memmove_helper<T,true> {
596  static inline void run(const T* start, const T* end, T* target)
597  { std::memmove(target, start, std::ptrdiff_t(end)-std::ptrdiff_t(start)); }
598 };
599 
600 template<typename T> struct smart_memmove_helper<T,false> {
601  static inline void run(const T* start, const T* end, T* target)
602  {
603  if (uintptr_t(target) < uintptr_t(start))
604  {
605  std::copy(start, end, target);
606  }
607  else
608  {
609  std::ptrdiff_t count = (std::ptrdiff_t(end)-std::ptrdiff_t(start)) / sizeof(T);
610  std::copy_backward(start, end, target + count);
611  }
612  }
613 };
614 
615 
616 /*****************************************************************************
617 *** Implementation of runtime stack allocation (falling back to malloc) ***
618 *****************************************************************************/
619 
620 // you can overwrite Eigen's default behavior regarding alloca by defining EIGEN_ALLOCA
621 // to the appropriate stack allocation function
622 #ifndef EIGEN_ALLOCA
623  #if EIGEN_OS_LINUX || EIGEN_OS_MAC || (defined alloca)
624  #define EIGEN_ALLOCA alloca
625  #elif EIGEN_COMP_MSVC
626  #define EIGEN_ALLOCA _alloca
627  #endif
628 #endif
629 
630 // This helper class construct the allocated memory, and takes care of destructing and freeing the handled data
631 // at destruction time. In practice this helper class is mainly useful to avoid memory leak in case of exceptions.
632 template<typename T> class aligned_stack_memory_handler : noncopyable
633 {
634  public:
635  /* Creates a stack_memory_handler responsible for the buffer \a ptr of size \a size.
636  * Note that \a ptr can be 0 regardless of the other parameters.
637  * This constructor takes care of constructing/initializing the elements of the buffer if required by the scalar type T (see NumTraits<T>::RequireInitialization).
638  * In this case, the buffer elements will also be destructed when this handler will be destructed.
639  * Finally, if \a dealloc is true, then the pointer \a ptr is freed.
640  **/
641  aligned_stack_memory_handler(T* ptr, size_t size, bool dealloc)
642  : m_ptr(ptr), m_size(size), m_deallocate(dealloc)
643  {
644  if(NumTraits<T>::RequireInitialization && m_ptr)
645  Eigen::internal::construct_elements_of_array(m_ptr, size);
646  }
647  ~aligned_stack_memory_handler()
648  {
649  if(NumTraits<T>::RequireInitialization && m_ptr)
650  Eigen::internal::destruct_elements_of_array<T>(m_ptr, m_size);
651  if(m_deallocate)
652  Eigen::internal::aligned_free(m_ptr);
653  }
654  protected:
655  T* m_ptr;
656  size_t m_size;
657  bool m_deallocate;
658 };
659 
660 template<typename T> class scoped_array : noncopyable
661 {
662  T* m_ptr;
663 public:
664  explicit scoped_array(std::ptrdiff_t size)
665  {
666  m_ptr = new T[size];
667  }
668  ~scoped_array()
669  {
670  delete[] m_ptr;
671  }
672  T& operator[](std::ptrdiff_t i) { return m_ptr[i]; }
673  const T& operator[](std::ptrdiff_t i) const { return m_ptr[i]; }
674  T* &ptr() { return m_ptr; }
675  const T* ptr() const { return m_ptr; }
676  operator const T*() const { return m_ptr; }
677 };
678 
679 template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b)
680 {
681  std::swap(a.ptr(),b.ptr());
682 }
683 
684 } // end namespace internal
685 
701 #ifdef EIGEN_ALLOCA
702 
703  #if EIGEN_DEFAULT_ALIGN_BYTES>0
704  // We always manually re-align the result of EIGEN_ALLOCA.
705  // If alloca is already aligned, the compiler should be smart enough to optimize away the re-alignment.
706  #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast<void*>((reinterpret_cast<std::size_t>(EIGEN_ALLOCA(SIZE+EIGEN_DEFAULT_ALIGN_BYTES-1)) + EIGEN_DEFAULT_ALIGN_BYTES-1) & ~(std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1)))
707  #else
708  #define EIGEN_ALIGNED_ALLOCA(SIZE) EIGEN_ALLOCA(SIZE)
709  #endif
710 
711  #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
712  Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \
713  TYPE* NAME = (BUFFER)!=0 ? (BUFFER) \
714  : reinterpret_cast<TYPE*>( \
715  (sizeof(TYPE)*SIZE<=EIGEN_STACK_ALLOCATION_LIMIT) ? EIGEN_ALIGNED_ALLOCA(sizeof(TYPE)*SIZE) \
716  : Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE) ); \
717  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)
718 
719 #else
720 
721  #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
722  Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \
723  TYPE* NAME = (BUFFER)!=0 ? BUFFER : reinterpret_cast<TYPE*>(Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE)); \
724  Eigen::internal::aligned_stack_memory_handler<TYPE> EIGEN_CAT(NAME,_stack_memory_destructor)((BUFFER)==0 ? NAME : 0,SIZE,true)
725 
726 #endif
727 
728 
729 /*****************************************************************************
730 *** Implementation of EIGEN_MAKE_ALIGNED_OPERATOR_NEW [_IF] ***
731 *****************************************************************************/
732 
733 #if EIGEN_MAX_ALIGN_BYTES!=0
734  #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
735  void* operator new(size_t size, const std::nothrow_t&) EIGEN_NO_THROW { \
736  EIGEN_TRY { return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); } \
737  EIGEN_CATCH (...) { return 0; } \
738  }
739  #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign) \
740  void *operator new(size_t size) { \
741  return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
742  } \
743  void *operator new[](size_t size) { \
744  return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
745  } \
746  void operator delete(void * ptr) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
747  void operator delete[](void * ptr) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
748  void operator delete(void * ptr, std::size_t /* sz */) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
749  void operator delete[](void * ptr, std::size_t /* sz */) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
750  /* in-place new and delete. since (at least afaik) there is no actual */ \
751  /* memory allocated we can safely let the default implementation handle */ \
752  /* this particular case. */ \
753  static void *operator new(size_t size, void *ptr) { return ::operator new(size,ptr); } \
754  static void *operator new[](size_t size, void* ptr) { return ::operator new[](size,ptr); } \
755  void operator delete(void * memory, void *ptr) EIGEN_NO_THROW { return ::operator delete(memory,ptr); } \
756  void operator delete[](void * memory, void *ptr) EIGEN_NO_THROW { return ::operator delete[](memory,ptr); } \
757  /* nothrow-new (returns zero instead of std::bad_alloc) */ \
758  EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
759  void operator delete(void *ptr, const std::nothrow_t&) EIGEN_NO_THROW { \
760  Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); \
761  } \
762  typedef void eigen_aligned_operator_new_marker_type;
763 #else
764  #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)
765 #endif
766 
767 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(true)
768 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size) \
769  EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool(((Size)!=Eigen::Dynamic) && ((sizeof(Scalar)*(Size))%EIGEN_MAX_ALIGN_BYTES==0)))
770 
771 /****************************************************************************/
772 
789 template<class T>
790 class aligned_allocator : public std::allocator<T>
791 {
792 public:
793  typedef size_t size_type;
794  typedef std::ptrdiff_t difference_type;
795  typedef T* pointer;
796  typedef const T* const_pointer;
797  typedef T& reference;
798  typedef const T& const_reference;
799  typedef T value_type;
800 
801  template<class U>
802  struct rebind
803  {
804  typedef aligned_allocator<U> other;
805  };
806 
807  aligned_allocator() : std::allocator<T>() {}
808 
809  aligned_allocator(const aligned_allocator& other) : std::allocator<T>(other) {}
810 
811  template<class U>
812  aligned_allocator(const aligned_allocator<U>& other) : std::allocator<T>(other) {}
813 
814  ~aligned_allocator() {}
815 
816  pointer allocate(size_type num, const void* /*hint*/ = 0)
817  {
818  internal::check_size_for_overflow<T>(num);
819  return static_cast<pointer>( internal::aligned_malloc(num * sizeof(T)) );
820  }
821 
822  void deallocate(pointer p, size_type /*num*/)
823  {
824  internal::aligned_free(p);
825  }
826 };
827 
828 //---------- Cache sizes ----------
829 
830 #if !defined(EIGEN_NO_CPUID)
831 # if EIGEN_COMP_GNUC && EIGEN_ARCH_i386_OR_x86_64
832 # if defined(__PIC__) && EIGEN_ARCH_i386
833  // Case for x86 with PIC
834 # define EIGEN_CPUID(abcd,func,id) \
835  __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));
836 # elif defined(__PIC__) && EIGEN_ARCH_x86_64
837  // 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.
838  // However, we cannot detect which code model is used, and the xchg overhead is negligible anyway.
839 # define EIGEN_CPUID(abcd,func,id) \
840  __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));
841 # else
842  // Case for x86_64 or x86 w/o PIC
843 # define EIGEN_CPUID(abcd,func,id) \
844  __asm__ __volatile__ ("cpuid": "=a" (abcd[0]), "=b" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "0" (func), "2" (id) );
845 # endif
846 # elif EIGEN_COMP_MSVC
847 # if (EIGEN_COMP_MSVC > 1500) && EIGEN_ARCH_i386_OR_x86_64
848 # define EIGEN_CPUID(abcd,func,id) __cpuidex((int*)abcd,func,id)
849 # endif
850 # endif
851 #endif
852 
853 namespace internal {
854 
855 #ifdef EIGEN_CPUID
856 
857 inline bool cpuid_is_vendor(int abcd[4], const int vendor[3])
858 {
859  return abcd[1]==vendor[0] && abcd[3]==vendor[1] && abcd[2]==vendor[2];
860 }
861 
862 inline void queryCacheSizes_intel_direct(int& l1, int& l2, int& l3)
863 {
864  int abcd[4];
865  l1 = l2 = l3 = 0;
866  int cache_id = 0;
867  int cache_type = 0;
868  do {
869  abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
870  EIGEN_CPUID(abcd,0x4,cache_id);
871  cache_type = (abcd[0] & 0x0F) >> 0;
872  if(cache_type==1||cache_type==3) // data or unified cache
873  {
874  int cache_level = (abcd[0] & 0xE0) >> 5; // A[7:5]
875  int ways = (abcd[1] & 0xFFC00000) >> 22; // B[31:22]
876  int partitions = (abcd[1] & 0x003FF000) >> 12; // B[21:12]
877  int line_size = (abcd[1] & 0x00000FFF) >> 0; // B[11:0]
878  int sets = (abcd[2]); // C[31:0]
879 
880  int cache_size = (ways+1) * (partitions+1) * (line_size+1) * (sets+1);
881 
882  switch(cache_level)
883  {
884  case 1: l1 = cache_size; break;
885  case 2: l2 = cache_size; break;
886  case 3: l3 = cache_size; break;
887  default: break;
888  }
889  }
890  cache_id++;
891  } while(cache_type>0 && cache_id<16);
892 }
893 
894 inline void queryCacheSizes_intel_codes(int& l1, int& l2, int& l3)
895 {
896  int abcd[4];
897  abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
898  l1 = l2 = l3 = 0;
899  EIGEN_CPUID(abcd,0x00000002,0);
900  unsigned char * bytes = reinterpret_cast<unsigned char *>(abcd)+2;
901  bool check_for_p2_core2 = false;
902  for(int i=0; i<14; ++i)
903  {
904  switch(bytes[i])
905  {
906  case 0x0A: l1 = 8; break; // 0Ah data L1 cache, 8 KB, 2 ways, 32 byte lines
907  case 0x0C: l1 = 16; break; // 0Ch data L1 cache, 16 KB, 4 ways, 32 byte lines
908  case 0x0E: l1 = 24; break; // 0Eh data L1 cache, 24 KB, 6 ways, 64 byte lines
909  case 0x10: l1 = 16; break; // 10h data L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
910  case 0x15: l1 = 16; break; // 15h code L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
911  case 0x2C: l1 = 32; break; // 2Ch data L1 cache, 32 KB, 8 ways, 64 byte lines
912  case 0x30: l1 = 32; break; // 30h code L1 cache, 32 KB, 8 ways, 64 byte lines
913  case 0x60: l1 = 16; break; // 60h data L1 cache, 16 KB, 8 ways, 64 byte lines, sectored
914  case 0x66: l1 = 8; break; // 66h data L1 cache, 8 KB, 4 ways, 64 byte lines, sectored
915  case 0x67: l1 = 16; break; // 67h data L1 cache, 16 KB, 4 ways, 64 byte lines, sectored
916  case 0x68: l1 = 32; break; // 68h data L1 cache, 32 KB, 4 ways, 64 byte lines, sectored
917  case 0x1A: l2 = 96; break; // code and data L2 cache, 96 KB, 6 ways, 64 byte lines (IA-64)
918  case 0x22: l3 = 512; break; // code and data L3 cache, 512 KB, 4 ways (!), 64 byte lines, dual-sectored
919  case 0x23: l3 = 1024; break; // code and data L3 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
920  case 0x25: l3 = 2048; break; // code and data L3 cache, 2048 KB, 8 ways, 64 byte lines, dual-sectored
921  case 0x29: l3 = 4096; break; // code and data L3 cache, 4096 KB, 8 ways, 64 byte lines, dual-sectored
922  case 0x39: l2 = 128; break; // code and data L2 cache, 128 KB, 4 ways, 64 byte lines, sectored
923  case 0x3A: l2 = 192; break; // code and data L2 cache, 192 KB, 6 ways, 64 byte lines, sectored
924  case 0x3B: l2 = 128; break; // code and data L2 cache, 128 KB, 2 ways, 64 byte lines, sectored
925  case 0x3C: l2 = 256; break; // code and data L2 cache, 256 KB, 4 ways, 64 byte lines, sectored
926  case 0x3D: l2 = 384; break; // code and data L2 cache, 384 KB, 6 ways, 64 byte lines, sectored
927  case 0x3E: l2 = 512; break; // code and data L2 cache, 512 KB, 4 ways, 64 byte lines, sectored
928  case 0x40: l2 = 0; break; // no integrated L2 cache (P6 core) or L3 cache (P4 core)
929  case 0x41: l2 = 128; break; // code and data L2 cache, 128 KB, 4 ways, 32 byte lines
930  case 0x42: l2 = 256; break; // code and data L2 cache, 256 KB, 4 ways, 32 byte lines
931  case 0x43: l2 = 512; break; // code and data L2 cache, 512 KB, 4 ways, 32 byte lines
932  case 0x44: l2 = 1024; break; // code and data L2 cache, 1024 KB, 4 ways, 32 byte lines
933  case 0x45: l2 = 2048; break; // code and data L2 cache, 2048 KB, 4 ways, 32 byte lines
934  case 0x46: l3 = 4096; break; // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines
935  case 0x47: l3 = 8192; break; // code and data L3 cache, 8192 KB, 8 ways, 64 byte lines
936  case 0x48: l2 = 3072; break; // code and data L2 cache, 3072 KB, 12 ways, 64 byte lines
937  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
938  case 0x4A: l3 = 6144; break; // code and data L3 cache, 6144 KB, 12 ways, 64 byte lines
939  case 0x4B: l3 = 8192; break; // code and data L3 cache, 8192 KB, 16 ways, 64 byte lines
940  case 0x4C: l3 = 12288; break; // code and data L3 cache, 12288 KB, 12 ways, 64 byte lines
941  case 0x4D: l3 = 16384; break; // code and data L3 cache, 16384 KB, 16 ways, 64 byte lines
942  case 0x4E: l2 = 6144; break; // code and data L2 cache, 6144 KB, 24 ways, 64 byte lines
943  case 0x78: l2 = 1024; break; // code and data L2 cache, 1024 KB, 4 ways, 64 byte lines
944  case 0x79: l2 = 128; break; // code and data L2 cache, 128 KB, 8 ways, 64 byte lines, dual-sectored
945  case 0x7A: l2 = 256; break; // code and data L2 cache, 256 KB, 8 ways, 64 byte lines, dual-sectored
946  case 0x7B: l2 = 512; break; // code and data L2 cache, 512 KB, 8 ways, 64 byte lines, dual-sectored
947  case 0x7C: l2 = 1024; break; // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
948  case 0x7D: l2 = 2048; break; // code and data L2 cache, 2048 KB, 8 ways, 64 byte lines
949  case 0x7E: l2 = 256; break; // code and data L2 cache, 256 KB, 8 ways, 128 byte lines, sect. (IA-64)
950  case 0x7F: l2 = 512; break; // code and data L2 cache, 512 KB, 2 ways, 64 byte lines
951  case 0x80: l2 = 512; break; // code and data L2 cache, 512 KB, 8 ways, 64 byte lines
952  case 0x81: l2 = 128; break; // code and data L2 cache, 128 KB, 8 ways, 32 byte lines
953  case 0x82: l2 = 256; break; // code and data L2 cache, 256 KB, 8 ways, 32 byte lines
954  case 0x83: l2 = 512; break; // code and data L2 cache, 512 KB, 8 ways, 32 byte lines
955  case 0x84: l2 = 1024; break; // code and data L2 cache, 1024 KB, 8 ways, 32 byte lines
956  case 0x85: l2 = 2048; break; // code and data L2 cache, 2048 KB, 8 ways, 32 byte lines
957  case 0x86: l2 = 512; break; // code and data L2 cache, 512 KB, 4 ways, 64 byte lines
958  case 0x87: l2 = 1024; break; // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines
959  case 0x88: l3 = 2048; break; // code and data L3 cache, 2048 KB, 4 ways, 64 byte lines (IA-64)
960  case 0x89: l3 = 4096; break; // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines (IA-64)
961  case 0x8A: l3 = 8192; break; // code and data L3 cache, 8192 KB, 4 ways, 64 byte lines (IA-64)
962  case 0x8D: l3 = 3072; break; // code and data L3 cache, 3072 KB, 12 ways, 128 byte lines (IA-64)
963 
964  default: break;
965  }
966  }
967  if(check_for_p2_core2 && l2 == l3)
968  l3 = 0;
969  l1 *= 1024;
970  l2 *= 1024;
971  l3 *= 1024;
972 }
973 
974 inline void queryCacheSizes_intel(int& l1, int& l2, int& l3, int max_std_funcs)
975 {
976  if(max_std_funcs>=4)
977  queryCacheSizes_intel_direct(l1,l2,l3);
978  else
979  queryCacheSizes_intel_codes(l1,l2,l3);
980 }
981 
982 inline void queryCacheSizes_amd(int& l1, int& l2, int& l3)
983 {
984  int abcd[4];
985  abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
986  EIGEN_CPUID(abcd,0x80000005,0);
987  l1 = (abcd[2] >> 24) * 1024; // C[31:24] = L1 size in KB
988  abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
989  EIGEN_CPUID(abcd,0x80000006,0);
990  l2 = (abcd[2] >> 16) * 1024; // C[31;16] = l2 cache size in KB
991  l3 = ((abcd[3] & 0xFFFC000) >> 18) * 512 * 1024; // D[31;18] = l3 cache size in 512KB
992 }
993 #endif
994 
997 inline void queryCacheSizes(int& l1, int& l2, int& l3)
998 {
999  #ifdef EIGEN_CPUID
1000  int abcd[4];
1001  const int GenuineIntel[] = {0x756e6547, 0x49656e69, 0x6c65746e};
1002  const int AuthenticAMD[] = {0x68747541, 0x69746e65, 0x444d4163};
1003  const int AMDisbetter_[] = {0x69444d41, 0x74656273, 0x21726574}; // "AMDisbetter!"
1004 
1005  // identify the CPU vendor
1006  EIGEN_CPUID(abcd,0x0,0);
1007  int max_std_funcs = abcd[1];
1008  if(cpuid_is_vendor(abcd,GenuineIntel))
1009  queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
1010  else if(cpuid_is_vendor(abcd,AuthenticAMD) || cpuid_is_vendor(abcd,AMDisbetter_))
1011  queryCacheSizes_amd(l1,l2,l3);
1012  else
1013  // by default let's use Intel's API
1014  queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
1015 
1016  // here is the list of other vendors:
1017 // ||cpuid_is_vendor(abcd,"VIA VIA VIA ")
1018 // ||cpuid_is_vendor(abcd,"CyrixInstead")
1019 // ||cpuid_is_vendor(abcd,"CentaurHauls")
1020 // ||cpuid_is_vendor(abcd,"GenuineTMx86")
1021 // ||cpuid_is_vendor(abcd,"TransmetaCPU")
1022 // ||cpuid_is_vendor(abcd,"RiseRiseRise")
1023 // ||cpuid_is_vendor(abcd,"Geode by NSC")
1024 // ||cpuid_is_vendor(abcd,"SiS SiS SiS ")
1025 // ||cpuid_is_vendor(abcd,"UMC UMC UMC ")
1026 // ||cpuid_is_vendor(abcd,"NexGenDriven")
1027  #else
1028  l1 = l2 = l3 = -1;
1029  #endif
1030 }
1031 
1034 inline int queryL1CacheSize()
1035 {
1036  int l1(-1), l2, l3;
1037  queryCacheSizes(l1,l2,l3);
1038  return l1;
1039 }
1040 
1043 inline int queryTopLevelCacheSize()
1044 {
1045  int l1, l2(-1), l3(-1);
1046  queryCacheSizes(l1,l2,l3);
1047  return (std::max)(l2,l3);
1048 }
1049 
1050 } // end namespace internal
1051 
1052 } // end namespace Eigen
1053 
1054 #endif // EIGEN_MEMORY_H
Definition: LDLT.h:16
STL compatible allocator to use with with 16 byte aligned types.
Definition: Memory.h:790
Definition: Eigen_Colamd.h:54