Eigen  5.0.0-dev
Loading...
Searching...
No Matches
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*** 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) && defined(__LP64__) && \
34 !defined(__SANITIZE_ADDRESS__) && (EIGEN_DEFAULT_ALIGN_BYTES == 16)
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__) && !(EIGEN_ARCH_ARM || EIGEN_ARCH_MIPS) && (EIGEN_DEFAULT_ALIGN_BYTES == 16)
45#define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 1
46#else
47#define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 0
48#endif
49
50#if (EIGEN_OS_MAC && (EIGEN_DEFAULT_ALIGN_BYTES == 16)) || (EIGEN_OS_WIN64 && (EIGEN_DEFAULT_ALIGN_BYTES == 16)) || \
51 EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED || EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED
52#define EIGEN_MALLOC_ALREADY_ALIGNED 1
53#else
54#define EIGEN_MALLOC_ALREADY_ALIGNED 0
55#endif
56
57#endif
58
59#ifndef EIGEN_MALLOC_CHECK_THREAD_LOCAL
60
61// Check whether we can use the thread_local keyword to allow or disallow
62// allocating memory with per-thread granularity, by means of the
63// set_is_malloc_allowed() function.
64#ifndef EIGEN_AVOID_THREAD_LOCAL
65
66#if ((EIGEN_COMP_GNUC) || __has_feature(cxx_thread_local) || EIGEN_COMP_MSVC >= 1900) && \
67 !defined(EIGEN_GPU_COMPILE_PHASE)
68#define EIGEN_MALLOC_CHECK_THREAD_LOCAL thread_local
69#else
70#define EIGEN_MALLOC_CHECK_THREAD_LOCAL
71#endif
72
73#else // EIGEN_AVOID_THREAD_LOCAL
74#define EIGEN_MALLOC_CHECK_THREAD_LOCAL
75#endif // EIGEN_AVOID_THREAD_LOCAL
76
77#endif
78
79// IWYU pragma: private
80#include "../InternalHeaderCheck.h"
81
82namespace Eigen {
83
84namespace internal {
85
86/*****************************************************************************
87*** Implementation of portable aligned versions of malloc/free/realloc ***
88*****************************************************************************/
89
90#ifdef EIGEN_NO_MALLOC
91EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed() {
92 eigen_assert(false && "heap allocation is forbidden (EIGEN_NO_MALLOC is defined)");
93}
94#elif defined EIGEN_RUNTIME_NO_MALLOC
95EIGEN_DEVICE_FUNC inline bool is_malloc_allowed_impl(bool update, bool new_value = false) {
96 EIGEN_MALLOC_CHECK_THREAD_LOCAL static bool value = true;
97 if (update == 1) value = new_value;
98 return value;
99}
100EIGEN_DEVICE_FUNC inline bool is_malloc_allowed() { return is_malloc_allowed_impl(false); }
101EIGEN_DEVICE_FUNC inline bool set_is_malloc_allowed(bool new_value) { return is_malloc_allowed_impl(true, new_value); }
102EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed() {
103 eigen_assert(is_malloc_allowed() &&
104 "heap allocation is forbidden (EIGEN_RUNTIME_NO_MALLOC is defined and g_is_malloc_allowed is false)");
105}
106#else
107EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed() {}
108#endif
109
110EIGEN_DEVICE_FUNC inline void throw_std_bad_alloc() {
111#ifdef EIGEN_EXCEPTIONS
112 throw std::bad_alloc();
113#else
114 std::size_t huge = static_cast<std::size_t>(-1);
115#if defined(EIGEN_HIPCC)
116 //
117 // calls to "::operator new" are to be treated as opaque function calls (i.e no inlining),
118 // and as a consequence the code in the #else block triggers the hipcc warning :
119 // "no overloaded function has restriction specifiers that are compatible with the ambient context"
120 //
121 // "throw_std_bad_alloc" has the EIGEN_DEVICE_FUNC attribute, so it seems that hipcc expects
122 // the same on "operator new"
123 // Reverting code back to the old version in this #if block for the hipcc compiler
124 //
125 new int[huge];
126#else
127 void* unused = ::operator new(huge);
128 EIGEN_UNUSED_VARIABLE(unused);
129#endif
130#endif
131}
132
133/*****************************************************************************
134*** Implementation of handmade aligned functions ***
135*****************************************************************************/
136
137/* ----- Hand made implementations of aligned malloc/free and realloc ----- */
138
142EIGEN_DEVICE_FUNC inline void* handmade_aligned_malloc(std::size_t size,
143 std::size_t alignment = EIGEN_DEFAULT_ALIGN_BYTES) {
144 eigen_assert(alignment >= sizeof(void*) && alignment <= 256 && (alignment & (alignment - 1)) == 0 &&
145 "Alignment must be at least sizeof(void*), less than or equal to 256, and a power of 2");
146
147 check_that_malloc_is_allowed();
148 EIGEN_USING_STD(malloc)
149 void* original = malloc(size + alignment);
150 if (original == nullptr) return nullptr;
151 std::size_t offset = alignment - (reinterpret_cast<std::size_t>(original) & (alignment - 1));
152 void* aligned = static_cast<void*>(static_cast<uint8_t*>(original) + offset);
153 // Store offset - 1, since it is guaranteed to be at least 1.
154 *(static_cast<uint8_t*>(aligned) - 1) = static_cast<uint8_t>(offset - 1);
155 return aligned;
156}
157
159EIGEN_DEVICE_FUNC inline void handmade_aligned_free(void* ptr) {
160 if (ptr != nullptr) {
161 std::size_t offset = static_cast<std::size_t>(*(static_cast<uint8_t*>(ptr) - 1)) + 1;
162 void* original = static_cast<void*>(static_cast<uint8_t*>(ptr) - offset);
163
164 check_that_malloc_is_allowed();
165 EIGEN_USING_STD(free)
166 free(original);
167 }
168}
169
175EIGEN_DEVICE_FUNC inline void* handmade_aligned_realloc(void* ptr, std::size_t new_size, std::size_t old_size,
176 std::size_t alignment = EIGEN_DEFAULT_ALIGN_BYTES) {
177 if (ptr == nullptr) return handmade_aligned_malloc(new_size, alignment);
178 std::size_t old_offset = static_cast<std::size_t>(*(static_cast<uint8_t*>(ptr) - 1)) + 1;
179 void* old_original = static_cast<uint8_t*>(ptr) - old_offset;
180
181 check_that_malloc_is_allowed();
182 EIGEN_USING_STD(realloc)
183 void* original = realloc(old_original, new_size + alignment);
184 if (original == nullptr) return nullptr;
185 if (original == old_original) return ptr;
186 std::size_t offset = alignment - (reinterpret_cast<std::size_t>(original) & (alignment - 1));
187 void* aligned = static_cast<void*>(static_cast<uint8_t*>(original) + offset);
188 if (offset != old_offset) {
189 const void* src = static_cast<const void*>(static_cast<uint8_t*>(original) + old_offset);
190 std::size_t count = (std::min)(new_size, old_size);
191 std::memmove(aligned, src, count);
192 }
193 // Store offset - 1, since it is guaranteed to be at least 1.
194 *(static_cast<uint8_t*>(aligned) - 1) = static_cast<uint8_t>(offset - 1);
195 return aligned;
196}
197
201EIGEN_DEVICE_FUNC inline void* aligned_malloc(std::size_t size) {
202 if (size == 0) return nullptr;
203
204 void* result;
205#if (EIGEN_DEFAULT_ALIGN_BYTES == 0) || EIGEN_MALLOC_ALREADY_ALIGNED
206
207 check_that_malloc_is_allowed();
208 EIGEN_USING_STD(malloc)
209 result = malloc(size);
210
211#if EIGEN_DEFAULT_ALIGN_BYTES == 16
212 eigen_assert((size < 16 || (std::size_t(result) % 16) == 0) &&
213 "System's malloc returned an unaligned pointer. Compile with EIGEN_MALLOC_ALREADY_ALIGNED=0 to fallback "
214 "to handmade aligned memory allocator.");
215#endif
216#else
217 result = handmade_aligned_malloc(size);
218#endif
219
220 if (!result && size) throw_std_bad_alloc();
221
222 return result;
223}
224
226EIGEN_DEVICE_FUNC inline void aligned_free(void* ptr) {
227#if (EIGEN_DEFAULT_ALIGN_BYTES == 0) || EIGEN_MALLOC_ALREADY_ALIGNED
228
229 if (ptr != nullptr) {
230 check_that_malloc_is_allowed();
231 EIGEN_USING_STD(free)
232 free(ptr);
233 }
234
235#else
236 handmade_aligned_free(ptr);
237#endif
238}
239
245EIGEN_DEVICE_FUNC inline void* aligned_realloc(void* ptr, std::size_t new_size, std::size_t old_size) {
246 if (ptr == nullptr) return aligned_malloc(new_size);
247 if (old_size == new_size) return ptr;
248 if (new_size == 0) {
249 aligned_free(ptr);
250 return nullptr;
251 }
252
253 void* result;
254#if (EIGEN_DEFAULT_ALIGN_BYTES == 0) || EIGEN_MALLOC_ALREADY_ALIGNED
255 EIGEN_UNUSED_VARIABLE(old_size)
256
257 check_that_malloc_is_allowed();
258 EIGEN_USING_STD(realloc)
259 result = realloc(ptr, new_size);
260#else
261 result = handmade_aligned_realloc(ptr, new_size, old_size);
262#endif
263
264 if (!result && new_size) throw_std_bad_alloc();
265
266 return result;
267}
268
269/*****************************************************************************
270*** Implementation of conditionally aligned functions ***
271*****************************************************************************/
272
276template <bool Align>
277EIGEN_DEVICE_FUNC inline void* conditional_aligned_malloc(std::size_t size) {
278 return aligned_malloc(size);
279}
280
281template <>
282EIGEN_DEVICE_FUNC inline void* conditional_aligned_malloc<false>(std::size_t size) {
283 if (size == 0) return nullptr;
284
285 check_that_malloc_is_allowed();
286 EIGEN_USING_STD(malloc)
287 void* result = malloc(size);
288
289 if (!result && size) throw_std_bad_alloc();
290 return result;
291}
292
294template <bool Align>
295EIGEN_DEVICE_FUNC inline void conditional_aligned_free(void* ptr) {
296 aligned_free(ptr);
297}
298
299template <>
300EIGEN_DEVICE_FUNC inline void conditional_aligned_free<false>(void* ptr) {
301 if (ptr != nullptr) {
302 check_that_malloc_is_allowed();
303 EIGEN_USING_STD(free)
304 free(ptr);
305 }
306}
307
308template <bool Align>
309EIGEN_DEVICE_FUNC inline void* conditional_aligned_realloc(void* ptr, std::size_t new_size, std::size_t old_size) {
310 return aligned_realloc(ptr, new_size, old_size);
311}
312
313template <>
314EIGEN_DEVICE_FUNC inline void* conditional_aligned_realloc<false>(void* ptr, std::size_t new_size,
315 std::size_t old_size) {
316 if (ptr == nullptr) return conditional_aligned_malloc<false>(new_size);
317 if (old_size == new_size) return ptr;
318 if (new_size == 0) {
319 conditional_aligned_free<false>(ptr);
320 return nullptr;
321 }
322
323 check_that_malloc_is_allowed();
324 EIGEN_USING_STD(realloc)
325 return realloc(ptr, new_size);
326}
327
328/*****************************************************************************
329*** Construction/destruction of array elements ***
330*****************************************************************************/
331
335template <typename T>
336EIGEN_DEVICE_FUNC inline void destruct_elements_of_array(T* ptr, std::size_t size) {
337 // always destruct an array starting from the end.
338 if (ptr)
339 while (size) ptr[--size].~T();
340}
341
345template <typename T>
346EIGEN_DEVICE_FUNC inline T* default_construct_elements_of_array(T* ptr, std::size_t size) {
347 std::size_t i = 0;
348 EIGEN_TRY {
349 for (i = 0; i < size; ++i) ::new (ptr + i) T;
350 }
351 EIGEN_CATCH(...) {
352 destruct_elements_of_array(ptr, i);
353 EIGEN_THROW;
354 }
355 return ptr;
356}
357
361template <typename T>
362EIGEN_DEVICE_FUNC inline T* copy_construct_elements_of_array(T* ptr, const T* src, std::size_t size) {
363 std::size_t i = 0;
364 EIGEN_TRY {
365 for (i = 0; i < size; ++i) ::new (ptr + i) T(*(src + i));
366 }
367 EIGEN_CATCH(...) {
368 destruct_elements_of_array(ptr, i);
369 EIGEN_THROW;
370 }
371 return ptr;
372}
373
377template <typename T>
378EIGEN_DEVICE_FUNC inline T* move_construct_elements_of_array(T* ptr, T* src, std::size_t size) {
379 std::size_t i = 0;
380 EIGEN_TRY {
381 for (i = 0; i < size; ++i) ::new (ptr + i) T(std::move(*(src + i)));
382 }
383 EIGEN_CATCH(...) {
384 destruct_elements_of_array(ptr, i);
385 EIGEN_THROW;
386 }
387 return ptr;
388}
389
390/*****************************************************************************
391*** Implementation of aligned new/delete-like functions ***
392*****************************************************************************/
393
394template <typename T>
395EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void check_size_for_overflow(std::size_t size) {
396 constexpr std::size_t max_elements = (std::numeric_limits<std::ptrdiff_t>::max)() / sizeof(T);
397 if (size > max_elements) throw_std_bad_alloc();
398}
399
404template <typename T>
405EIGEN_DEVICE_FUNC inline T* aligned_new(std::size_t size) {
406 check_size_for_overflow<T>(size);
407 T* result = static_cast<T*>(aligned_malloc(sizeof(T) * size));
408 EIGEN_TRY { return default_construct_elements_of_array(result, size); }
409 EIGEN_CATCH(...) {
410 aligned_free(result);
411 EIGEN_THROW;
412 }
413 return result;
414}
415
416template <typename T, bool Align>
417EIGEN_DEVICE_FUNC inline T* conditional_aligned_new(std::size_t size) {
418 check_size_for_overflow<T>(size);
419 T* result = static_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T) * size));
420 EIGEN_TRY { return default_construct_elements_of_array(result, size); }
421 EIGEN_CATCH(...) {
422 conditional_aligned_free<Align>(result);
423 EIGEN_THROW;
424 }
425 return result;
426}
427
431template <typename T>
432EIGEN_DEVICE_FUNC inline void aligned_delete(T* ptr, std::size_t size) {
433 destruct_elements_of_array<T>(ptr, size);
434 aligned_free(ptr);
435}
436
440template <typename T, bool Align>
441EIGEN_DEVICE_FUNC inline void conditional_aligned_delete(T* ptr, std::size_t size) {
442 destruct_elements_of_array<T>(ptr, size);
443 conditional_aligned_free<Align>(ptr);
444}
445
446template <typename T, bool Align>
447EIGEN_DEVICE_FUNC inline T* conditional_aligned_realloc_new(T* pts, std::size_t new_size, std::size_t old_size) {
448 check_size_for_overflow<T>(new_size);
449 check_size_for_overflow<T>(old_size);
450
451 // If elements need to be explicitly initialized, we cannot simply realloc
452 // (or memcpy) the memory block - each element needs to be reconstructed.
453 // Otherwise, objects that contain internal pointers like mpfr or
454 // AnnoyingScalar can be pointing to the wrong thing.
455 T* result = static_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T) * new_size));
456 EIGEN_TRY {
457 // Move-construct initial elements.
458 std::size_t copy_size = (std::min)(old_size, new_size);
459 move_construct_elements_of_array(result, pts, copy_size);
460
461 // Default-construct remaining elements.
462 if (new_size > old_size) {
463 default_construct_elements_of_array(result + copy_size, new_size - old_size);
464 }
465
466 // Delete old elements.
467 conditional_aligned_delete<T, Align>(pts, old_size);
468 }
469 EIGEN_CATCH(...) {
470 conditional_aligned_free<Align>(result);
471 EIGEN_THROW;
472 }
473
474 return result;
475}
476
477template <typename T, bool Align>
478EIGEN_DEVICE_FUNC inline T* conditional_aligned_new_auto(std::size_t size) {
479 if (size == 0) return nullptr; // short-cut. Also fixes Bug 884
480 check_size_for_overflow<T>(size);
481 T* result = static_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T) * size));
482 if (NumTraits<T>::RequireInitialization) {
483 EIGEN_TRY { default_construct_elements_of_array(result, size); }
484 EIGEN_CATCH(...) {
485 conditional_aligned_free<Align>(result);
486 EIGEN_THROW;
487 }
488 }
489 return result;
490}
491
492template <typename T, bool Align>
493EIGEN_DEVICE_FUNC inline T* conditional_aligned_realloc_new_auto(T* pts, std::size_t new_size, std::size_t old_size) {
494 if (NumTraits<T>::RequireInitialization) {
495 return conditional_aligned_realloc_new<T, Align>(pts, new_size, old_size);
496 }
497
498 check_size_for_overflow<T>(new_size);
499 check_size_for_overflow<T>(old_size);
500 return static_cast<T*>(
501 conditional_aligned_realloc<Align>(static_cast<void*>(pts), sizeof(T) * new_size, sizeof(T) * old_size));
502}
503
504template <typename T, bool Align>
505EIGEN_DEVICE_FUNC inline void conditional_aligned_delete_auto(T* ptr, std::size_t size) {
506 if (NumTraits<T>::RequireInitialization) destruct_elements_of_array<T>(ptr, size);
507 conditional_aligned_free<Align>(ptr);
508}
509
510/****************************************************************************/
511
530template <int Alignment, typename Scalar, typename Index>
531EIGEN_DEVICE_FUNC inline Index first_aligned(const Scalar* array, Index size) {
532 const Index ScalarSize = sizeof(Scalar);
533 const Index AlignmentSize = Alignment / ScalarSize;
534 const Index AlignmentMask = AlignmentSize - 1;
535
536 if (AlignmentSize <= 1) {
537 // Either the requested alignment if smaller than a scalar, or it exactly match a 1 scalar
538 // so that all elements of the array have the same alignment.
539 return 0;
540 } else if ((std::uintptr_t(array) & (sizeof(Scalar) - 1)) || (Alignment % ScalarSize) != 0) {
541 // The array is not aligned to the size of a single scalar, or the requested alignment is not a multiple of the
542 // scalar size. Consequently, no element of the array is well aligned.
543 return size;
544 } else {
545 Index first = (AlignmentSize - (Index((std::uintptr_t(array) / sizeof(Scalar))) & AlignmentMask)) & AlignmentMask;
546 return (first < size) ? first : size;
547 }
548}
549
552template <typename Scalar, typename Index>
553EIGEN_DEVICE_FUNC inline Index first_default_aligned(const Scalar* array, Index size) {
554 typedef typename packet_traits<Scalar>::type DefaultPacketType;
555 return first_aligned<unpacket_traits<DefaultPacketType>::alignment>(array, size);
556}
557
560template <typename Index>
561inline Index first_multiple(Index size, Index base) {
562 return ((size + base - 1) / base) * base;
563}
564
565// std::copy is much slower than memcpy, so let's introduce a smart_copy which
566// use memcpy on trivial types, i.e., on types that does not require an initialization ctor.
567template <typename T, bool UseMemcpy>
568struct smart_copy_helper;
569
570template <typename T>
571EIGEN_DEVICE_FUNC void smart_copy(const T* start, const T* end, T* target) {
572 smart_copy_helper<T, !NumTraits<T>::RequireInitialization>::run(start, end, target);
573}
574
575template <typename T>
576struct smart_copy_helper<T, true> {
577 EIGEN_DEVICE_FUNC static inline void run(const T* start, const T* end, T* target) {
578 std::intptr_t size = std::intptr_t(end) - std::intptr_t(start);
579 if (size == 0) return;
580 eigen_internal_assert(start != 0 && end != 0 && target != 0);
581 EIGEN_USING_STD(memcpy)
582 memcpy(target, start, size);
583 }
584};
585
586template <typename T>
587struct smart_copy_helper<T, false> {
588 EIGEN_DEVICE_FUNC static inline void run(const T* start, const T* end, T* target) { std::copy(start, end, target); }
589};
590
591// intelligent memmove. falls back to std::memmove for POD types, uses std::copy otherwise.
592template <typename T, bool UseMemmove>
593struct smart_memmove_helper;
594
595template <typename T>
596void smart_memmove(const T* start, const T* end, T* target) {
597 smart_memmove_helper<T, !NumTraits<T>::RequireInitialization>::run(start, end, target);
598}
599
600template <typename T>
601struct smart_memmove_helper<T, true> {
602 static inline void run(const T* start, const T* end, T* target) {
603 std::intptr_t size = std::intptr_t(end) - std::intptr_t(start);
604 if (size == 0) return;
605 eigen_internal_assert(start != 0 && end != 0 && target != 0);
606 std::memmove(target, start, size);
607 }
608};
609
610template <typename T>
611struct smart_memmove_helper<T, false> {
612 static inline void run(const T* start, const T* end, T* target) {
613 if (std::uintptr_t(target) < std::uintptr_t(start)) {
614 std::copy(start, end, target);
615 } else {
616 std::ptrdiff_t count = (std::ptrdiff_t(end) - std::ptrdiff_t(start)) / sizeof(T);
617 std::copy_backward(start, end, target + count);
618 }
619 }
620};
621
622template <typename T>
623EIGEN_DEVICE_FUNC T* smart_move(T* start, T* end, T* target) {
624 return std::move(start, end, target);
625}
626
627/*****************************************************************************
628*** Implementation of runtime stack allocation (falling back to malloc) ***
629*****************************************************************************/
630
631// you can overwrite Eigen's default behavior regarding alloca by defining EIGEN_ALLOCA
632// to the appropriate stack allocation function
633#if !defined EIGEN_ALLOCA && !defined EIGEN_GPU_COMPILE_PHASE
634#if EIGEN_OS_LINUX || EIGEN_OS_MAC || (defined alloca)
635#define EIGEN_ALLOCA alloca
636#elif EIGEN_COMP_MSVC
637#define EIGEN_ALLOCA _alloca
638#endif
639#endif
640
641// With clang -Oz -mthumb, alloca changes the stack pointer in a way that is
642// not allowed in Thumb2. -DEIGEN_STACK_ALLOCATION_LIMIT=0 doesn't work because
643// the compiler still emits bad code because stack allocation checks use "<=".
644// TODO: Eliminate after https://bugs.llvm.org/show_bug.cgi?id=23772
645// is fixed.
646#if defined(__clang__) && defined(__thumb__)
647#undef EIGEN_ALLOCA
648#endif
649
650// This helper class construct the allocated memory, and takes care of destructing and freeing the handled data
651// at destruction time. In practice this helper class is mainly useful to avoid memory leak in case of exceptions.
652template <typename T>
653class aligned_stack_memory_handler : noncopyable {
654 public:
655 /* Creates a stack_memory_handler responsible for the buffer \a ptr of size \a size.
656 * Note that \a ptr can be 0 regardless of the other parameters.
657 * This constructor takes care of constructing/initializing the elements of the buffer if required by the scalar type
658 *T (see NumTraits<T>::RequireInitialization). In this case, the buffer elements will also be destructed when this
659 *handler will be destructed. Finally, if \a dealloc is true, then the pointer \a ptr is freed.
660 **/
661 EIGEN_DEVICE_FUNC aligned_stack_memory_handler(T* ptr, std::size_t size, bool dealloc)
662 : m_ptr(ptr), m_size(size), m_deallocate(dealloc) {
663 if (NumTraits<T>::RequireInitialization && m_ptr) Eigen::internal::default_construct_elements_of_array(m_ptr, size);
664 }
665 EIGEN_DEVICE_FUNC ~aligned_stack_memory_handler() {
666 if (NumTraits<T>::RequireInitialization && m_ptr) Eigen::internal::destruct_elements_of_array<T>(m_ptr, m_size);
667 if (m_deallocate) Eigen::internal::aligned_free(m_ptr);
668 }
669
670 protected:
671 T* m_ptr;
672 std::size_t m_size;
673 bool m_deallocate;
674};
675
676#ifdef EIGEN_ALLOCA
677
678template <typename Xpr, int NbEvaluations,
679 bool MapExternalBuffer = nested_eval<Xpr, NbEvaluations>::Evaluate && Xpr::MaxSizeAtCompileTime == Dynamic>
680struct local_nested_eval_wrapper {
681 static constexpr bool NeedExternalBuffer = false;
682 typedef typename Xpr::Scalar Scalar;
683 typedef typename nested_eval<Xpr, NbEvaluations>::type ObjectType;
684 ObjectType object;
685
686 EIGEN_DEVICE_FUNC local_nested_eval_wrapper(const Xpr& xpr, Scalar* ptr) : object(xpr) {
687 EIGEN_UNUSED_VARIABLE(ptr);
688 eigen_internal_assert(ptr == 0);
689 }
690};
691
692template <typename Xpr, int NbEvaluations>
693struct local_nested_eval_wrapper<Xpr, NbEvaluations, true> {
694 static constexpr bool NeedExternalBuffer = true;
695 typedef typename Xpr::Scalar Scalar;
696 typedef typename plain_object_eval<Xpr>::type PlainObject;
697 typedef Map<PlainObject, EIGEN_DEFAULT_ALIGN_BYTES> ObjectType;
698 ObjectType object;
699
700 EIGEN_DEVICE_FUNC local_nested_eval_wrapper(const Xpr& xpr, Scalar* ptr)
701 : object(ptr == 0 ? reinterpret_cast<Scalar*>(Eigen::internal::aligned_malloc(sizeof(Scalar) * xpr.size())) : ptr,
702 xpr.rows(), xpr.cols()),
703 m_deallocate(ptr == 0) {
704 if (NumTraits<Scalar>::RequireInitialization && object.data())
705 Eigen::internal::default_construct_elements_of_array(object.data(), object.size());
706 object = xpr;
707 }
708
709 EIGEN_DEVICE_FUNC ~local_nested_eval_wrapper() {
710 if (NumTraits<Scalar>::RequireInitialization && object.data())
711 Eigen::internal::destruct_elements_of_array(object.data(), object.size());
712 if (m_deallocate) Eigen::internal::aligned_free(object.data());
713 }
714
715 private:
716 bool m_deallocate;
717};
718
719#endif // EIGEN_ALLOCA
720
721template <typename T>
722class scoped_array : noncopyable {
723 T* m_ptr;
724
725 public:
726 explicit scoped_array(std::ptrdiff_t size) { m_ptr = new T[size]; }
727 ~scoped_array() { delete[] m_ptr; }
728 T& operator[](std::ptrdiff_t i) { return m_ptr[i]; }
729 const T& operator[](std::ptrdiff_t i) const { return m_ptr[i]; }
730 T*& ptr() { return m_ptr; }
731 const T* ptr() const { return m_ptr; }
732 operator const T*() const { return m_ptr; }
733};
734
735template <typename T>
736void swap(scoped_array<T>& a, scoped_array<T>& b) {
737 std::swap(a.ptr(), b.ptr());
738}
739
740} // end namespace internal
741
765#if defined(EIGEN_ALLOCA) && !defined(EIGEN_NO_ALLOCA)
766
767#if EIGEN_DEFAULT_ALIGN_BYTES > 0
768// We always manually re-align the result of EIGEN_ALLOCA.
769// If alloca is already aligned, the compiler should be smart enough to optimize away the re-alignment.
770
771#if ((EIGEN_COMP_GNUC || EIGEN_COMP_CLANG) && !EIGEN_COMP_NVHPC)
772#define EIGEN_ALIGNED_ALLOCA(SIZE) __builtin_alloca_with_align(SIZE, CHAR_BIT* EIGEN_DEFAULT_ALIGN_BYTES)
773#else
774EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void* eigen_aligned_alloca_helper(void* ptr) {
775 constexpr std::uintptr_t mask = EIGEN_DEFAULT_ALIGN_BYTES - 1;
776 std::uintptr_t ptr_int = std::uintptr_t(ptr);
777 std::uintptr_t aligned_ptr_int = (ptr_int + mask) & ~mask;
778 std::uintptr_t offset = aligned_ptr_int - ptr_int;
779 return static_cast<void*>(static_cast<uint8_t*>(ptr) + offset);
780}
781#define EIGEN_ALIGNED_ALLOCA(SIZE) eigen_aligned_alloca_helper(EIGEN_ALLOCA(SIZE + EIGEN_DEFAULT_ALIGN_BYTES - 1))
782#endif
783
784#else
785#define EIGEN_ALIGNED_ALLOCA(SIZE) EIGEN_ALLOCA(SIZE)
786#endif
787
788#define ei_declare_aligned_stack_constructed_variable(TYPE, NAME, SIZE, BUFFER) \
789 Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \
790 TYPE* NAME = (BUFFER) != 0 ? (BUFFER) \
791 : reinterpret_cast<TYPE*>((sizeof(TYPE) * (SIZE) <= EIGEN_STACK_ALLOCATION_LIMIT) \
792 ? EIGEN_ALIGNED_ALLOCA(sizeof(TYPE) * (SIZE)) \
793 : Eigen::internal::aligned_malloc(sizeof(TYPE) * (SIZE))); \
794 Eigen::internal::aligned_stack_memory_handler<TYPE> EIGEN_CAT(NAME, _stack_memory_destructor)( \
795 (BUFFER) == 0 ? NAME : 0, SIZE, sizeof(TYPE) * (SIZE) > EIGEN_STACK_ALLOCATION_LIMIT)
796
797#define ei_declare_local_nested_eval(XPR_T, XPR, N, NAME) \
798 Eigen::internal::local_nested_eval_wrapper<XPR_T, N> EIGEN_CAT(NAME, _wrapper)( \
799 XPR, reinterpret_cast<typename XPR_T::Scalar*>( \
800 ((Eigen::internal::local_nested_eval_wrapper<XPR_T, N>::NeedExternalBuffer) && \
801 ((sizeof(typename XPR_T::Scalar) * XPR.size()) <= EIGEN_STACK_ALLOCATION_LIMIT)) \
802 ? EIGEN_ALIGNED_ALLOCA(sizeof(typename XPR_T::Scalar) * XPR.size()) \
803 : 0)); \
804 typename Eigen::internal::local_nested_eval_wrapper<XPR_T, N>::ObjectType NAME(EIGEN_CAT(NAME, _wrapper).object)
805
806#else
807
808#define ei_declare_aligned_stack_constructed_variable(TYPE, NAME, SIZE, BUFFER) \
809 Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \
810 TYPE* NAME = \
811 (BUFFER) != 0 ? BUFFER : reinterpret_cast<TYPE*>(Eigen::internal::aligned_malloc(sizeof(TYPE) * (SIZE))); \
812 Eigen::internal::aligned_stack_memory_handler<TYPE> EIGEN_CAT(NAME, _stack_memory_destructor)( \
813 (BUFFER) == 0 ? NAME : 0, SIZE, true)
814
815#define ei_declare_local_nested_eval(XPR_T, XPR, N, NAME) \
816 typename Eigen::internal::nested_eval<XPR_T, N>::type NAME(XPR)
817
818#endif
819
820/*****************************************************************************
821*** Implementation of EIGEN_MAKE_ALIGNED_OPERATOR_NEW [_IF] ***
822*****************************************************************************/
823
824#if EIGEN_HAS_CXX17_OVERALIGN
825
826// C++17 -> no need to bother about alignment anymore :)
827
828#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign)
829#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)
830#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW
831#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar, Size)
832
833#else
834
835// HIP does not support new/delete on device.
836#if EIGEN_MAX_ALIGN_BYTES != 0 && !defined(EIGEN_HIP_DEVICE_COMPILE)
837#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
838 EIGEN_DEVICE_FUNC void* operator new(std::size_t size, const std::nothrow_t&) noexcept { \
839 EIGEN_TRY { return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); } \
840 EIGEN_CATCH(...) { return 0; } \
841 }
842#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign) \
843 EIGEN_DEVICE_FUNC void* operator new(std::size_t size) { \
844 return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
845 } \
846 EIGEN_DEVICE_FUNC void* operator new[](std::size_t size) { \
847 return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
848 } \
849 EIGEN_DEVICE_FUNC void operator delete(void* ptr) noexcept { \
850 Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); \
851 } \
852 EIGEN_DEVICE_FUNC void operator delete[](void* ptr) noexcept { \
853 Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); \
854 } \
855 EIGEN_DEVICE_FUNC void operator delete(void* ptr, std::size_t /* sz */) noexcept { \
856 Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); \
857 } \
858 EIGEN_DEVICE_FUNC void operator delete[](void* ptr, std::size_t /* sz */) noexcept { \
859 Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); \
860 } \
861 /* in-place new and delete. since (at least afaik) there is no actual */ \
862 /* memory allocated we can safely let the default implementation handle */ \
863 /* this particular case. */ \
864 EIGEN_DEVICE_FUNC static void* operator new(std::size_t size, void* ptr) { return ::operator new(size, ptr); } \
865 EIGEN_DEVICE_FUNC static void* operator new[](std::size_t size, void* ptr) { return ::operator new[](size, ptr); } \
866 EIGEN_DEVICE_FUNC void operator delete(void* memory, void* ptr) noexcept { return ::operator delete(memory, ptr); } \
867 EIGEN_DEVICE_FUNC void operator delete[](void* memory, void* ptr) noexcept { \
868 return ::operator delete[](memory, ptr); \
869 } \
870 /* nothrow-new (returns zero instead of std::bad_alloc) */ \
871 EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
872 EIGEN_DEVICE_FUNC void operator delete(void* ptr, const std::nothrow_t&) noexcept { \
873 Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); \
874 } \
875 typedef void eigen_aligned_operator_new_marker_type;
876#else
877#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)
878#endif
879
880#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(true)
881#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar, Size) \
882 EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF( \
883 bool(((Size) != Eigen::Dynamic) && \
884 (((EIGEN_MAX_ALIGN_BYTES >= 16) && ((sizeof(Scalar) * (Size)) % (EIGEN_MAX_ALIGN_BYTES) == 0)) || \
885 ((EIGEN_MAX_ALIGN_BYTES >= 32) && ((sizeof(Scalar) * (Size)) % (EIGEN_MAX_ALIGN_BYTES / 2) == 0)) || \
886 ((EIGEN_MAX_ALIGN_BYTES >= 64) && ((sizeof(Scalar) * (Size)) % (EIGEN_MAX_ALIGN_BYTES / 4) == 0)))))
887
888#endif
889
890/****************************************************************************/
891
916template <class T>
917class aligned_allocator {
918 public:
919 typedef std::size_t size_type;
920 typedef std::ptrdiff_t difference_type;
921 typedef T* pointer;
922 typedef const T* const_pointer;
923 typedef T& reference;
924 typedef const T& const_reference;
925 typedef T value_type;
926
927 template <class U>
928 struct rebind {
929 typedef aligned_allocator<U> other;
930 };
931
932 aligned_allocator() = default;
933
934 aligned_allocator(const aligned_allocator&) = default;
935
936 template <class U>
937 aligned_allocator(const aligned_allocator<U>&) {}
938
939 template <class U>
940 constexpr bool operator==(const aligned_allocator<U>&) const noexcept {
941 return true;
942 }
943 template <class U>
944 constexpr bool operator!=(const aligned_allocator<U>&) const noexcept {
945 return false;
946 }
947
948#if EIGEN_COMP_GNUC_STRICT && EIGEN_GNUC_STRICT_AT_LEAST(7, 0, 0)
949 // In gcc std::allocator::max_size() is bugged making gcc triggers a warning:
950 // eigen/Eigen/src/Core/util/Memory.h:189:12: warning: argument 1 value '18446744073709551612' exceeds maximum object
951 // size 9223372036854775807 See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87544
952 size_type max_size() const { return (std::numeric_limits<std::ptrdiff_t>::max)() / sizeof(T); }
953#endif
954
955 pointer allocate(size_type num, const void* /*hint*/ = 0) {
956 internal::check_size_for_overflow<T>(num);
957 return static_cast<pointer>(internal::aligned_malloc(num * sizeof(T)));
958 }
959
960 void deallocate(pointer p, size_type /*num*/) { internal::aligned_free(p); }
961};
962
963//---------- Cache sizes ----------
964
965#if !defined(EIGEN_NO_CPUID)
966#if EIGEN_COMP_GNUC && EIGEN_ARCH_i386_OR_x86_64
967#if defined(__PIC__) && EIGEN_ARCH_i386
968// Case for x86 with PIC
969#define EIGEN_CPUID(abcd, func, id) \
970 __asm__ __volatile__("xchgl %%ebx, %k1;cpuid; xchgl %%ebx,%k1" \
971 : "=a"(abcd[0]), "=&r"(abcd[1]), "=c"(abcd[2]), "=d"(abcd[3]) \
972 : "a"(func), "c"(id));
973#elif defined(__PIC__) && EIGEN_ARCH_x86_64
974// Case for x64 with PIC. In theory this is only a problem with recent gcc and with medium or large code model, not with
975// the default small code model. However, we cannot detect which code model is used, and the xchg overhead is negligible
976// anyway.
977#define EIGEN_CPUID(abcd, func, id) \
978 __asm__ __volatile__("xchg{q}\t{%%}rbx, %q1; cpuid; xchg{q}\t{%%}rbx, %q1" \
979 : "=a"(abcd[0]), "=&r"(abcd[1]), "=c"(abcd[2]), "=d"(abcd[3]) \
980 : "0"(func), "2"(id));
981#else
982// Case for x86_64 or x86 w/o PIC
983#define EIGEN_CPUID(abcd, func, id) \
984 __asm__ __volatile__("cpuid" : "=a"(abcd[0]), "=b"(abcd[1]), "=c"(abcd[2]), "=d"(abcd[3]) : "0"(func), "2"(id));
985#endif
986#elif EIGEN_COMP_MSVC
987#if EIGEN_ARCH_i386_OR_x86_64
988#define EIGEN_CPUID(abcd, func, id) __cpuidex((int*)abcd, func, id)
989#endif
990#endif
991#endif
992
993namespace internal {
994
995#ifdef EIGEN_CPUID
996
997inline bool cpuid_is_vendor(int abcd[4], const int vendor[3]) {
998 return abcd[1] == vendor[0] && abcd[3] == vendor[1] && abcd[2] == vendor[2];
999}
1000
1001inline void queryCacheSizes_intel_direct(int& l1, int& l2, int& l3) {
1002 int abcd[4];
1003 l1 = l2 = l3 = 0;
1004 int cache_id = 0;
1005 int cache_type = 0;
1006 do {
1007 abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
1008 EIGEN_CPUID(abcd, 0x4, cache_id);
1009 cache_type = (abcd[0] & 0x0F) >> 0;
1010 if (cache_type == 1 || cache_type == 3) // data or unified cache
1011 {
1012 int cache_level = (abcd[0] & 0xE0) >> 5; // A[7:5]
1013 int ways = (abcd[1] & 0xFFC00000) >> 22; // B[31:22]
1014 int partitions = (abcd[1] & 0x003FF000) >> 12; // B[21:12]
1015 int line_size = (abcd[1] & 0x00000FFF) >> 0; // B[11:0]
1016 int sets = (abcd[2]); // C[31:0]
1017
1018 int cache_size = (ways + 1) * (partitions + 1) * (line_size + 1) * (sets + 1);
1019
1020 switch (cache_level) {
1021 case 1:
1022 l1 = cache_size;
1023 break;
1024 case 2:
1025 l2 = cache_size;
1026 break;
1027 case 3:
1028 l3 = cache_size;
1029 break;
1030 default:
1031 break;
1032 }
1033 }
1034 cache_id++;
1035 } while (cache_type > 0 && cache_id < 16);
1036}
1037
1038inline void queryCacheSizes_intel_codes(int& l1, int& l2, int& l3) {
1039 int abcd[4];
1040 abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
1041 l1 = l2 = l3 = 0;
1042 EIGEN_CPUID(abcd, 0x00000002, 0);
1043 unsigned char* bytes = reinterpret_cast<unsigned char*>(abcd) + 2;
1044 bool check_for_p2_core2 = false;
1045 for (int i = 0; i < 14; ++i) {
1046 switch (bytes[i]) {
1047 case 0x0A:
1048 l1 = 8;
1049 break; // 0Ah data L1 cache, 8 KB, 2 ways, 32 byte lines
1050 case 0x0C:
1051 l1 = 16;
1052 break; // 0Ch data L1 cache, 16 KB, 4 ways, 32 byte lines
1053 case 0x0E:
1054 l1 = 24;
1055 break; // 0Eh data L1 cache, 24 KB, 6 ways, 64 byte lines
1056 case 0x10:
1057 l1 = 16;
1058 break; // 10h data L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
1059 case 0x15:
1060 l1 = 16;
1061 break; // 15h code L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
1062 case 0x2C:
1063 l1 = 32;
1064 break; // 2Ch data L1 cache, 32 KB, 8 ways, 64 byte lines
1065 case 0x30:
1066 l1 = 32;
1067 break; // 30h code L1 cache, 32 KB, 8 ways, 64 byte lines
1068 case 0x60:
1069 l1 = 16;
1070 break; // 60h data L1 cache, 16 KB, 8 ways, 64 byte lines, sectored
1071 case 0x66:
1072 l1 = 8;
1073 break; // 66h data L1 cache, 8 KB, 4 ways, 64 byte lines, sectored
1074 case 0x67:
1075 l1 = 16;
1076 break; // 67h data L1 cache, 16 KB, 4 ways, 64 byte lines, sectored
1077 case 0x68:
1078 l1 = 32;
1079 break; // 68h data L1 cache, 32 KB, 4 ways, 64 byte lines, sectored
1080 case 0x1A:
1081 l2 = 96;
1082 break; // code and data L2 cache, 96 KB, 6 ways, 64 byte lines (IA-64)
1083 case 0x22:
1084 l3 = 512;
1085 break; // code and data L3 cache, 512 KB, 4 ways (!), 64 byte lines, dual-sectored
1086 case 0x23:
1087 l3 = 1024;
1088 break; // code and data L3 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
1089 case 0x25:
1090 l3 = 2048;
1091 break; // code and data L3 cache, 2048 KB, 8 ways, 64 byte lines, dual-sectored
1092 case 0x29:
1093 l3 = 4096;
1094 break; // code and data L3 cache, 4096 KB, 8 ways, 64 byte lines, dual-sectored
1095 case 0x39:
1096 l2 = 128;
1097 break; // code and data L2 cache, 128 KB, 4 ways, 64 byte lines, sectored
1098 case 0x3A:
1099 l2 = 192;
1100 break; // code and data L2 cache, 192 KB, 6 ways, 64 byte lines, sectored
1101 case 0x3B:
1102 l2 = 128;
1103 break; // code and data L2 cache, 128 KB, 2 ways, 64 byte lines, sectored
1104 case 0x3C:
1105 l2 = 256;
1106 break; // code and data L2 cache, 256 KB, 4 ways, 64 byte lines, sectored
1107 case 0x3D:
1108 l2 = 384;
1109 break; // code and data L2 cache, 384 KB, 6 ways, 64 byte lines, sectored
1110 case 0x3E:
1111 l2 = 512;
1112 break; // code and data L2 cache, 512 KB, 4 ways, 64 byte lines, sectored
1113 case 0x40:
1114 l2 = 0;
1115 break; // no integrated L2 cache (P6 core) or L3 cache (P4 core)
1116 case 0x41:
1117 l2 = 128;
1118 break; // code and data L2 cache, 128 KB, 4 ways, 32 byte lines
1119 case 0x42:
1120 l2 = 256;
1121 break; // code and data L2 cache, 256 KB, 4 ways, 32 byte lines
1122 case 0x43:
1123 l2 = 512;
1124 break; // code and data L2 cache, 512 KB, 4 ways, 32 byte lines
1125 case 0x44:
1126 l2 = 1024;
1127 break; // code and data L2 cache, 1024 KB, 4 ways, 32 byte lines
1128 case 0x45:
1129 l2 = 2048;
1130 break; // code and data L2 cache, 2048 KB, 4 ways, 32 byte lines
1131 case 0x46:
1132 l3 = 4096;
1133 break; // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines
1134 case 0x47:
1135 l3 = 8192;
1136 break; // code and data L3 cache, 8192 KB, 8 ways, 64 byte lines
1137 case 0x48:
1138 l2 = 3072;
1139 break; // code and data L2 cache, 3072 KB, 12 ways, 64 byte lines
1140 case 0x49:
1141 if (l2 != 0)
1142 l3 = 4096;
1143 else {
1144 check_for_p2_core2 = true;
1145 l3 = l2 = 4096;
1146 }
1147 break; // code and data L3 cache, 4096 KB, 16 ways, 64 byte lines (P4) or L2 for core2
1148 case 0x4A:
1149 l3 = 6144;
1150 break; // code and data L3 cache, 6144 KB, 12 ways, 64 byte lines
1151 case 0x4B:
1152 l3 = 8192;
1153 break; // code and data L3 cache, 8192 KB, 16 ways, 64 byte lines
1154 case 0x4C:
1155 l3 = 12288;
1156 break; // code and data L3 cache, 12288 KB, 12 ways, 64 byte lines
1157 case 0x4D:
1158 l3 = 16384;
1159 break; // code and data L3 cache, 16384 KB, 16 ways, 64 byte lines
1160 case 0x4E:
1161 l2 = 6144;
1162 break; // code and data L2 cache, 6144 KB, 24 ways, 64 byte lines
1163 case 0x78:
1164 l2 = 1024;
1165 break; // code and data L2 cache, 1024 KB, 4 ways, 64 byte lines
1166 case 0x79:
1167 l2 = 128;
1168 break; // code and data L2 cache, 128 KB, 8 ways, 64 byte lines, dual-sectored
1169 case 0x7A:
1170 l2 = 256;
1171 break; // code and data L2 cache, 256 KB, 8 ways, 64 byte lines, dual-sectored
1172 case 0x7B:
1173 l2 = 512;
1174 break; // code and data L2 cache, 512 KB, 8 ways, 64 byte lines, dual-sectored
1175 case 0x7C:
1176 l2 = 1024;
1177 break; // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
1178 case 0x7D:
1179 l2 = 2048;
1180 break; // code and data L2 cache, 2048 KB, 8 ways, 64 byte lines
1181 case 0x7E:
1182 l2 = 256;
1183 break; // code and data L2 cache, 256 KB, 8 ways, 128 byte lines, sect. (IA-64)
1184 case 0x7F:
1185 l2 = 512;
1186 break; // code and data L2 cache, 512 KB, 2 ways, 64 byte lines
1187 case 0x80:
1188 l2 = 512;
1189 break; // code and data L2 cache, 512 KB, 8 ways, 64 byte lines
1190 case 0x81:
1191 l2 = 128;
1192 break; // code and data L2 cache, 128 KB, 8 ways, 32 byte lines
1193 case 0x82:
1194 l2 = 256;
1195 break; // code and data L2 cache, 256 KB, 8 ways, 32 byte lines
1196 case 0x83:
1197 l2 = 512;
1198 break; // code and data L2 cache, 512 KB, 8 ways, 32 byte lines
1199 case 0x84:
1200 l2 = 1024;
1201 break; // code and data L2 cache, 1024 KB, 8 ways, 32 byte lines
1202 case 0x85:
1203 l2 = 2048;
1204 break; // code and data L2 cache, 2048 KB, 8 ways, 32 byte lines
1205 case 0x86:
1206 l2 = 512;
1207 break; // code and data L2 cache, 512 KB, 4 ways, 64 byte lines
1208 case 0x87:
1209 l2 = 1024;
1210 break; // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines
1211 case 0x88:
1212 l3 = 2048;
1213 break; // code and data L3 cache, 2048 KB, 4 ways, 64 byte lines (IA-64)
1214 case 0x89:
1215 l3 = 4096;
1216 break; // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines (IA-64)
1217 case 0x8A:
1218 l3 = 8192;
1219 break; // code and data L3 cache, 8192 KB, 4 ways, 64 byte lines (IA-64)
1220 case 0x8D:
1221 l3 = 3072;
1222 break; // code and data L3 cache, 3072 KB, 12 ways, 128 byte lines (IA-64)
1223
1224 default:
1225 break;
1226 }
1227 }
1228 if (check_for_p2_core2 && l2 == l3) l3 = 0;
1229 l1 *= 1024;
1230 l2 *= 1024;
1231 l3 *= 1024;
1232}
1233
1234inline void queryCacheSizes_intel(int& l1, int& l2, int& l3, int max_std_funcs) {
1235 if (max_std_funcs >= 4)
1236 queryCacheSizes_intel_direct(l1, l2, l3);
1237 else if (max_std_funcs >= 2)
1238 queryCacheSizes_intel_codes(l1, l2, l3);
1239 else
1240 l1 = l2 = l3 = 0;
1241}
1242
1243inline void queryCacheSizes_amd(int& l1, int& l2, int& l3) {
1244 int abcd[4];
1245 abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
1246
1247 // First query the max supported function.
1248 EIGEN_CPUID(abcd, 0x80000000, 0);
1249 if (static_cast<numext::uint32_t>(abcd[0]) >= static_cast<numext::uint32_t>(0x80000006)) {
1250 EIGEN_CPUID(abcd, 0x80000005, 0);
1251 l1 = (abcd[2] >> 24) * 1024; // C[31:24] = L1 size in KB
1252 abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
1253 EIGEN_CPUID(abcd, 0x80000006, 0);
1254 l2 = (abcd[2] >> 16) * 1024; // C[31;16] = l2 cache size in KB
1255 l3 = ((abcd[3] & 0xFFFC000) >> 18) * 512 * 1024; // D[31;18] = l3 cache size in 512KB
1256 } else {
1257 l1 = l2 = l3 = 0;
1258 }
1259}
1260#endif
1261
1264inline void queryCacheSizes(int& l1, int& l2, int& l3) {
1265#ifdef EIGEN_CPUID
1266 int abcd[4];
1267 const int GenuineIntel[] = {0x756e6547, 0x49656e69, 0x6c65746e};
1268 const int AuthenticAMD[] = {0x68747541, 0x69746e65, 0x444d4163};
1269 const int AMDisbetter_[] = {0x69444d41, 0x74656273, 0x21726574}; // "AMDisbetter!"
1270
1271 // identify the CPU vendor
1272 EIGEN_CPUID(abcd, 0x0, 0);
1273 int max_std_funcs = abcd[0];
1274 if (cpuid_is_vendor(abcd, GenuineIntel))
1275 queryCacheSizes_intel(l1, l2, l3, max_std_funcs);
1276 else if (cpuid_is_vendor(abcd, AuthenticAMD) || cpuid_is_vendor(abcd, AMDisbetter_))
1277 queryCacheSizes_amd(l1, l2, l3);
1278 else
1279 // by default let's use Intel's API
1280 queryCacheSizes_intel(l1, l2, l3, max_std_funcs);
1281
1282 // here is the list of other vendors:
1283 // ||cpuid_is_vendor(abcd,"VIA VIA VIA ")
1284 // ||cpuid_is_vendor(abcd,"CyrixInstead")
1285 // ||cpuid_is_vendor(abcd,"CentaurHauls")
1286 // ||cpuid_is_vendor(abcd,"GenuineTMx86")
1287 // ||cpuid_is_vendor(abcd,"TransmetaCPU")
1288 // ||cpuid_is_vendor(abcd,"RiseRiseRise")
1289 // ||cpuid_is_vendor(abcd,"Geode by NSC")
1290 // ||cpuid_is_vendor(abcd,"SiS SiS SiS ")
1291 // ||cpuid_is_vendor(abcd,"UMC UMC UMC ")
1292 // ||cpuid_is_vendor(abcd,"NexGenDriven")
1293#else
1294 l1 = l2 = l3 = -1;
1295#endif
1296}
1297
1300inline int queryL1CacheSize() {
1301 int l1(-1), l2, l3;
1302 queryCacheSizes(l1, l2, l3);
1303 return l1;
1304}
1305
1308inline int queryTopLevelCacheSize() {
1309 int l1, l2(-1), l3(-1);
1310 queryCacheSizes(l1, l2, l3);
1311 return (std::max)(l2, l3);
1312}
1313
1317
1318#if EIGEN_COMP_CXXVER >= 20 && defined(__cpp_lib_constexpr_dynamic_alloc) && \
1319 __cpp_lib_constexpr_dynamic_alloc >= 201907L
1320using std::construct_at;
1321#else
1322template <class T, class... Args>
1323EIGEN_DEVICE_FUNC T* construct_at(T* p, Args&&... args) {
1324 return ::new (const_cast<void*>(static_cast<const volatile void*>(p))) T(std::forward<Args>(args)...);
1325}
1326#endif
1327
1333#if EIGEN_COMP_CXXVER >= 17
1334using std::destroy_at;
1335#else
1336template <class T>
1337EIGEN_DEVICE_FUNC void destroy_at(T* p) {
1338 p->~T();
1339}
1340#endif
1341
1345#ifndef EIGEN_ASSUME_ALIGNED
1346#if defined(__cpp_lib_assume_aligned) && (__cpp_lib_assume_aligned >= 201811L)
1347#define EIGEN_ASSUME_ALIGNED(PTR, ALIGN_BYTES) \
1348 { PTR = std::assume_aligned<8 * (ALIGN_BYTES)>(PTR); }
1349#elif EIGEN_HAS_BUILTIN(__builtin_assume_aligned)
1350#define EIGEN_ASSUME_ALIGNED(PTR, ALIGN_BYTES) \
1351 { PTR = static_cast<decltype(PTR)>(__builtin_assume_aligned(PTR, (ALIGN_BYTES))); }
1352#else
1353#define EIGEN_ASSUME_ALIGNED(PTR, ALIGN_BYTES) /* do nothing */
1354#endif
1355#endif
1356
1357} // end namespace internal
1358
1359} // end namespace Eigen
1360
1361#endif // EIGEN_MEMORY_H
Eigen
Namespace containing all symbols from the Eigen library.
Definition B01_Experimental.dox:1
Eigen::Index
EIGEN_DEFAULT_DENSE_INDEX_TYPE Index
The Index type as used for the API.
Definition Meta.h:82
Eigen::Dynamic
const int Dynamic
Definition Constants.h:25