Eigen  3.4.90 (git rev a4098ac676528a83cfb73d4d26ce1b42ec05f47c)
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#include "../InternalHeaderCheck.h"
63
64namespace Eigen {
65
66namespace internal {
67
68EIGEN_DEVICE_FUNC
69inline void throw_std_bad_alloc()
70{
71 #ifdef EIGEN_EXCEPTIONS
72 throw std::bad_alloc();
73 #else
74 std::size_t huge = static_cast<std::size_t>(-1);
75 #if defined(EIGEN_HIPCC)
76 //
77 // calls to "::operator new" are to be treated as opaque function calls (i.e no inlining),
78 // and as a consequence the code in the #else block triggers the hipcc warning :
79 // "no overloaded function has restriction specifiers that are compatible with the ambient context"
80 //
81 // "throw_std_bad_alloc" has the EIGEN_DEVICE_FUNC attribute, so it seems that hipcc expects
82 // the same on "operator new"
83 // Reverting code back to the old version in this #if block for the hipcc compiler
84 //
85 new int[huge];
86 #else
87 void* unused = ::operator new(huge);
88 EIGEN_UNUSED_VARIABLE(unused);
89 #endif
90 #endif
91}
92
93/*****************************************************************************
94*** Implementation of handmade aligned functions ***
95*****************************************************************************/
96
97/* ----- Hand made implementations of aligned malloc/free and realloc ----- */
98
102EIGEN_DEVICE_FUNC inline void* handmade_aligned_malloc(std::size_t size, std::size_t alignment = EIGEN_DEFAULT_ALIGN_BYTES)
103{
104 eigen_assert(alignment >= sizeof(void*) && (alignment & (alignment-1)) == 0 && "Alignment must be at least sizeof(void*) and a power of 2");
105
106 EIGEN_USING_STD(malloc)
107 void *original = malloc(size+alignment);
108
109 if (original == 0) return 0;
110 void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(alignment-1))) + alignment);
111 *(reinterpret_cast<void**>(aligned) - 1) = original;
112 return aligned;
113}
114
116EIGEN_DEVICE_FUNC inline void handmade_aligned_free(void *ptr)
117{
118 if (ptr) {
119 EIGEN_USING_STD(free)
120 free(*(reinterpret_cast<void**>(ptr) - 1));
121 }
122}
123
129inline void* handmade_aligned_realloc(void* ptr, std::size_t size, std::size_t = 0)
130{
131 if (ptr == 0) return handmade_aligned_malloc(size);
132 void *original = *(reinterpret_cast<void**>(ptr) - 1);
133 std::ptrdiff_t previous_offset = static_cast<char *>(ptr)-static_cast<char *>(original);
134 original = std::realloc(original,size+EIGEN_DEFAULT_ALIGN_BYTES);
135 if (original == 0) return 0;
136 void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1))) + EIGEN_DEFAULT_ALIGN_BYTES);
137 void *previous_aligned = static_cast<char *>(original)+previous_offset;
138 if(aligned!=previous_aligned)
139 std::memmove(aligned, previous_aligned, size);
140
141 *(reinterpret_cast<void**>(aligned) - 1) = original;
142 return aligned;
143}
144
145/*****************************************************************************
146*** Implementation of portable aligned versions of malloc/free/realloc ***
147*****************************************************************************/
148
149#ifdef EIGEN_NO_MALLOC
150EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed()
151{
152 eigen_assert(false && "heap allocation is forbidden (EIGEN_NO_MALLOC is defined)");
153}
154#elif defined EIGEN_RUNTIME_NO_MALLOC
155EIGEN_DEVICE_FUNC inline bool is_malloc_allowed_impl(bool update, bool new_value = false)
156{
157 static bool value = true;
158 if (update == 1)
159 value = new_value;
160 return value;
161}
162EIGEN_DEVICE_FUNC inline bool is_malloc_allowed() { return is_malloc_allowed_impl(false); }
163EIGEN_DEVICE_FUNC inline bool set_is_malloc_allowed(bool new_value) { return is_malloc_allowed_impl(true, new_value); }
164EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed()
165{
166 eigen_assert(is_malloc_allowed() && "heap allocation is forbidden (EIGEN_RUNTIME_NO_MALLOC is defined and g_is_malloc_allowed is false)");
167}
168#else
169EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed()
170{}
171#endif
172
176EIGEN_DEVICE_FUNC inline void* aligned_malloc(std::size_t size)
177{
178 check_that_malloc_is_allowed();
179
180 void *result;
181 #if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED
182
183 EIGEN_USING_STD(malloc)
184 result = malloc(size);
185
186 #if EIGEN_DEFAULT_ALIGN_BYTES==16
187 eigen_assert((size<16 || (std::size_t(result)%16)==0) && "System's malloc returned an unaligned pointer. Compile with EIGEN_MALLOC_ALREADY_ALIGNED=0 to fallback to handmade aligned memory allocator.");
188 #endif
189 #else
190 result = handmade_aligned_malloc(size);
191 #endif
192
193 if(!result && size)
194 throw_std_bad_alloc();
195
196 return result;
197}
198
200EIGEN_DEVICE_FUNC inline void aligned_free(void *ptr)
201{
202 #if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED
203
204 EIGEN_USING_STD(free)
205 free(ptr);
206
207 #else
208 handmade_aligned_free(ptr);
209 #endif
210}
211
217inline void* aligned_realloc(void *ptr, std::size_t new_size, std::size_t old_size)
218{
219 EIGEN_UNUSED_VARIABLE(old_size)
220
221 void *result;
222#if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED
223 result = std::realloc(ptr,new_size);
224#else
225 result = handmade_aligned_realloc(ptr,new_size,old_size);
226#endif
227
228 if (!result && new_size)
229 throw_std_bad_alloc();
230
231 return result;
232}
233
234/*****************************************************************************
235*** Implementation of conditionally aligned functions ***
236*****************************************************************************/
237
241template<bool Align> EIGEN_DEVICE_FUNC inline void* conditional_aligned_malloc(std::size_t size)
242{
243 return aligned_malloc(size);
244}
245
246template<> EIGEN_DEVICE_FUNC inline void* conditional_aligned_malloc<false>(std::size_t size)
247{
248 check_that_malloc_is_allowed();
249
250 EIGEN_USING_STD(malloc)
251 void *result = malloc(size);
252
253 if(!result && size)
254 throw_std_bad_alloc();
255 return result;
256}
257
259template<bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_free(void *ptr)
260{
261 aligned_free(ptr);
262}
263
264template<> EIGEN_DEVICE_FUNC inline void conditional_aligned_free<false>(void *ptr)
265{
266 EIGEN_USING_STD(free)
267 free(ptr);
268}
269
270template<bool Align> inline void* conditional_aligned_realloc(void* ptr, std::size_t new_size, std::size_t old_size)
271{
272 return aligned_realloc(ptr, new_size, old_size);
273}
274
275template<> inline void* conditional_aligned_realloc<false>(void* ptr, std::size_t new_size, std::size_t)
276{
277 return std::realloc(ptr, new_size);
278}
279
280/*****************************************************************************
281*** Construction/destruction of array elements ***
282*****************************************************************************/
283
287template<typename T> EIGEN_DEVICE_FUNC inline void destruct_elements_of_array(T *ptr, std::size_t size)
288{
289 // always destruct an array starting from the end.
290 if(ptr)
291 while(size) ptr[--size].~T();
292}
293
297template<typename T> EIGEN_DEVICE_FUNC inline T* construct_elements_of_array(T *ptr, std::size_t size)
298{
299 std::size_t i;
300 EIGEN_TRY
301 {
302 for (i = 0; i < size; ++i) ::new (ptr + i) T;
303 return ptr;
304 }
305 EIGEN_CATCH(...)
306 {
307 destruct_elements_of_array(ptr, i);
308 EIGEN_THROW;
309 }
310 return NULL;
311}
312
313/*****************************************************************************
314*** Implementation of aligned new/delete-like functions ***
315*****************************************************************************/
316
317template<typename T>
318EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void check_size_for_overflow(std::size_t size)
319{
320 if(size > std::size_t(-1) / sizeof(T))
321 throw_std_bad_alloc();
322}
323
328template<typename T> EIGEN_DEVICE_FUNC inline T* aligned_new(std::size_t size)
329{
330 check_size_for_overflow<T>(size);
331 T *result = reinterpret_cast<T*>(aligned_malloc(sizeof(T)*size));
332 EIGEN_TRY
333 {
334 return construct_elements_of_array(result, size);
335 }
336 EIGEN_CATCH(...)
337 {
338 aligned_free(result);
339 EIGEN_THROW;
340 }
341 return result;
342}
343
344template<typename T, bool Align> EIGEN_DEVICE_FUNC inline T* conditional_aligned_new(std::size_t size)
345{
346 check_size_for_overflow<T>(size);
347 T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
348 EIGEN_TRY
349 {
350 return construct_elements_of_array(result, size);
351 }
352 EIGEN_CATCH(...)
353 {
354 conditional_aligned_free<Align>(result);
355 EIGEN_THROW;
356 }
357 return result;
358}
359
363template<typename T> EIGEN_DEVICE_FUNC inline void aligned_delete(T *ptr, std::size_t size)
364{
365 destruct_elements_of_array<T>(ptr, size);
366 Eigen::internal::aligned_free(ptr);
367}
368
372template<typename T, bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_delete(T *ptr, std::size_t size)
373{
374 destruct_elements_of_array<T>(ptr, size);
375 conditional_aligned_free<Align>(ptr);
376}
377
378template<typename T, bool Align> EIGEN_DEVICE_FUNC inline T* conditional_aligned_realloc_new(T* pts, std::size_t new_size, std::size_t old_size)
379{
380 check_size_for_overflow<T>(new_size);
381 check_size_for_overflow<T>(old_size);
382 if(new_size < old_size)
383 destruct_elements_of_array(pts+new_size, old_size-new_size);
384 T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
385 if(new_size > old_size)
386 {
387 EIGEN_TRY
388 {
389 construct_elements_of_array(result+old_size, new_size-old_size);
390 }
391 EIGEN_CATCH(...)
392 {
393 conditional_aligned_free<Align>(result);
394 EIGEN_THROW;
395 }
396 }
397 return result;
398}
399
400
401template<typename T, bool Align> EIGEN_DEVICE_FUNC inline T* conditional_aligned_new_auto(std::size_t size)
402{
403 if(size==0)
404 return 0; // short-cut. Also fixes Bug 884
405 check_size_for_overflow<T>(size);
406 T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
407 if(NumTraits<T>::RequireInitialization)
408 {
409 EIGEN_TRY
410 {
411 construct_elements_of_array(result, size);
412 }
413 EIGEN_CATCH(...)
414 {
415 conditional_aligned_free<Align>(result);
416 EIGEN_THROW;
417 }
418 }
419 return result;
420}
421
422template<typename T, bool Align> inline T* conditional_aligned_realloc_new_auto(T* pts, std::size_t new_size, std::size_t old_size)
423{
424 check_size_for_overflow<T>(new_size);
425 check_size_for_overflow<T>(old_size);
426 if(NumTraits<T>::RequireInitialization && (new_size < old_size))
427 destruct_elements_of_array(pts+new_size, old_size-new_size);
428 T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
429 if(NumTraits<T>::RequireInitialization && (new_size > old_size))
430 {
431 EIGEN_TRY
432 {
433 construct_elements_of_array(result+old_size, new_size-old_size);
434 }
435 EIGEN_CATCH(...)
436 {
437 conditional_aligned_free<Align>(result);
438 EIGEN_THROW;
439 }
440 }
441 return result;
442}
443
444template<typename T, bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_delete_auto(T *ptr, std::size_t size)
445{
446 if(NumTraits<T>::RequireInitialization)
447 destruct_elements_of_array<T>(ptr, size);
448 conditional_aligned_free<Align>(ptr);
449}
450
451/****************************************************************************/
452
470template<int Alignment, typename Scalar, typename Index>
471EIGEN_DEVICE_FUNC inline Index first_aligned(const Scalar* array, Index size)
472{
473 const Index ScalarSize = sizeof(Scalar);
474 const Index AlignmentSize = Alignment / ScalarSize;
475 const Index AlignmentMask = AlignmentSize-1;
476
477 if(AlignmentSize<=1)
478 {
479 // Either the requested alignment if smaller than a scalar, or it exactly match a 1 scalar
480 // so that all elements of the array have the same alignment.
481 return 0;
482 }
483 else if( (UIntPtr(array) & (sizeof(Scalar)-1)) || (Alignment%ScalarSize)!=0)
484 {
485 // The array is not aligned to the size of a single scalar, or the requested alignment is not a multiple of the scalar size.
486 // Consequently, no element of the array is well aligned.
487 return size;
488 }
489 else
490 {
491 Index first = (AlignmentSize - (Index((UIntPtr(array)/sizeof(Scalar))) & AlignmentMask)) & AlignmentMask;
492 return (first < size) ? first : size;
493 }
494}
495
498template<typename Scalar, typename Index>
499EIGEN_DEVICE_FUNC inline Index first_default_aligned(const Scalar* array, Index size)
500{
501 typedef typename packet_traits<Scalar>::type DefaultPacketType;
502 return first_aligned<unpacket_traits<DefaultPacketType>::alignment>(array, size);
503}
504
507template<typename Index>
508inline Index first_multiple(Index size, Index base)
509{
510 return ((size+base-1)/base)*base;
511}
512
513// std::copy is much slower than memcpy, so let's introduce a smart_copy which
514// use memcpy on trivial types, i.e., on types that does not require an initialization ctor.
515template<typename T, bool UseMemcpy> struct smart_copy_helper;
516
517template<typename T> EIGEN_DEVICE_FUNC void smart_copy(const T* start, const T* end, T* target)
518{
519 smart_copy_helper<T,!NumTraits<T>::RequireInitialization>::run(start, end, target);
520}
521
522template<typename T> struct smart_copy_helper<T,true> {
523 EIGEN_DEVICE_FUNC static inline void run(const T* start, const T* end, T* target)
524 {
525 IntPtr size = IntPtr(end)-IntPtr(start);
526 if(size==0) return;
527 eigen_internal_assert(start!=0 && end!=0 && target!=0);
528 EIGEN_USING_STD(memcpy)
529 memcpy(target, start, size);
530 }
531};
532
533template<typename T> struct smart_copy_helper<T,false> {
534 EIGEN_DEVICE_FUNC static inline void run(const T* start, const T* end, T* target)
535 { std::copy(start, end, target); }
536};
537
538// intelligent memmove. falls back to std::memmove for POD types, uses std::copy otherwise.
539template<typename T, bool UseMemmove> struct smart_memmove_helper;
540
541template<typename T> void smart_memmove(const T* start, const T* end, T* target)
542{
543 smart_memmove_helper<T,!NumTraits<T>::RequireInitialization>::run(start, end, target);
544}
545
546template<typename T> struct smart_memmove_helper<T,true> {
547 static inline void run(const T* start, const T* end, T* target)
548 {
549 IntPtr size = IntPtr(end)-IntPtr(start);
550 if(size==0) return;
551 eigen_internal_assert(start!=0 && end!=0 && target!=0);
552 std::memmove(target, start, size);
553 }
554};
555
556template<typename T> struct smart_memmove_helper<T,false> {
557 static inline void run(const T* start, const T* end, T* target)
558 {
559 if (UIntPtr(target) < UIntPtr(start))
560 {
561 std::copy(start, end, target);
562 }
563 else
564 {
565 std::ptrdiff_t count = (std::ptrdiff_t(end)-std::ptrdiff_t(start)) / sizeof(T);
566 std::copy_backward(start, end, target + count);
567 }
568 }
569};
570
571template<typename T> EIGEN_DEVICE_FUNC T* smart_move(T* start, T* end, T* target)
572{
573 return std::move(start, end, target);
574}
575
576/*****************************************************************************
577*** Implementation of runtime stack allocation (falling back to malloc) ***
578*****************************************************************************/
579
580// you can overwrite Eigen's default behavior regarding alloca by defining EIGEN_ALLOCA
581// to the appropriate stack allocation function
582#if ! defined EIGEN_ALLOCA && ! defined EIGEN_GPU_COMPILE_PHASE
583 #if EIGEN_OS_LINUX || EIGEN_OS_MAC || (defined alloca)
584 #define EIGEN_ALLOCA alloca
585 #elif EIGEN_COMP_MSVC
586 #define EIGEN_ALLOCA _alloca
587 #endif
588#endif
589
590// With clang -Oz -mthumb, alloca changes the stack pointer in a way that is
591// not allowed in Thumb2. -DEIGEN_STACK_ALLOCATION_LIMIT=0 doesn't work because
592// the compiler still emits bad code because stack allocation checks use "<=".
593// TODO: Eliminate after https://bugs.llvm.org/show_bug.cgi?id=23772
594// is fixed.
595#if defined(__clang__) && defined(__thumb__)
596 #undef EIGEN_ALLOCA
597#endif
598
599// This helper class construct the allocated memory, and takes care of destructing and freeing the handled data
600// at destruction time. In practice this helper class is mainly useful to avoid memory leak in case of exceptions.
601template<typename T> class aligned_stack_memory_handler : noncopyable
602{
603 public:
604 /* Creates a stack_memory_handler responsible for the buffer \a ptr of size \a size.
605 * Note that \a ptr can be 0 regardless of the other parameters.
606 * This constructor takes care of constructing/initializing the elements of the buffer if required by the scalar type T (see NumTraits<T>::RequireInitialization).
607 * In this case, the buffer elements will also be destructed when this handler will be destructed.
608 * Finally, if \a dealloc is true, then the pointer \a ptr is freed.
609 **/
610 EIGEN_DEVICE_FUNC
611 aligned_stack_memory_handler(T* ptr, std::size_t size, bool dealloc)
612 : m_ptr(ptr), m_size(size), m_deallocate(dealloc)
613 {
614 if(NumTraits<T>::RequireInitialization && m_ptr)
615 Eigen::internal::construct_elements_of_array(m_ptr, size);
616 }
617 EIGEN_DEVICE_FUNC
618 ~aligned_stack_memory_handler()
619 {
620 if(NumTraits<T>::RequireInitialization && m_ptr)
621 Eigen::internal::destruct_elements_of_array<T>(m_ptr, m_size);
622 if(m_deallocate)
623 Eigen::internal::aligned_free(m_ptr);
624 }
625 protected:
626 T* m_ptr;
627 std::size_t m_size;
628 bool m_deallocate;
629};
630
631#ifdef EIGEN_ALLOCA
632
633template<typename Xpr, int NbEvaluations,
634 bool MapExternalBuffer = nested_eval<Xpr,NbEvaluations>::Evaluate && Xpr::MaxSizeAtCompileTime==Dynamic
635 >
636struct local_nested_eval_wrapper
637{
638 static const bool NeedExternalBuffer = false;
639 typedef typename Xpr::Scalar Scalar;
640 typedef typename nested_eval<Xpr,NbEvaluations>::type ObjectType;
641 ObjectType object;
642
643 EIGEN_DEVICE_FUNC
644 local_nested_eval_wrapper(const Xpr& xpr, Scalar* ptr) : object(xpr)
645 {
646 EIGEN_UNUSED_VARIABLE(ptr);
647 eigen_internal_assert(ptr==0);
648 }
649};
650
651template<typename Xpr, int NbEvaluations>
652struct local_nested_eval_wrapper<Xpr,NbEvaluations,true>
653{
654 static const bool NeedExternalBuffer = true;
655 typedef typename Xpr::Scalar Scalar;
656 typedef typename plain_object_eval<Xpr>::type PlainObject;
657 typedef Map<PlainObject,EIGEN_DEFAULT_ALIGN_BYTES> ObjectType;
658 ObjectType object;
659
660 EIGEN_DEVICE_FUNC
661 local_nested_eval_wrapper(const Xpr& xpr, Scalar* ptr)
662 : object(ptr==0 ? reinterpret_cast<Scalar*>(Eigen::internal::aligned_malloc(sizeof(Scalar)*xpr.size())) : ptr, xpr.rows(), xpr.cols()),
663 m_deallocate(ptr==0)
664 {
665 if(NumTraits<Scalar>::RequireInitialization && object.data())
666 Eigen::internal::construct_elements_of_array(object.data(), object.size());
667 object = xpr;
668 }
669
670 EIGEN_DEVICE_FUNC
671 ~local_nested_eval_wrapper()
672 {
673 if(NumTraits<Scalar>::RequireInitialization && object.data())
674 Eigen::internal::destruct_elements_of_array(object.data(), object.size());
675 if(m_deallocate)
676 Eigen::internal::aligned_free(object.data());
677 }
678
679private:
680 bool m_deallocate;
681};
682
683#endif // EIGEN_ALLOCA
684
685template<typename T> class scoped_array : noncopyable
686{
687 T* m_ptr;
688public:
689 explicit scoped_array(std::ptrdiff_t size)
690 {
691 m_ptr = new T[size];
692 }
693 ~scoped_array()
694 {
695 delete[] m_ptr;
696 }
697 T& operator[](std::ptrdiff_t i) { return m_ptr[i]; }
698 const T& operator[](std::ptrdiff_t i) const { return m_ptr[i]; }
699 T* &ptr() { return m_ptr; }
700 const T* ptr() const { return m_ptr; }
701 operator const T*() const { return m_ptr; }
702};
703
704template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b)
705{
706 std::swap(a.ptr(),b.ptr());
707}
708
709} // end namespace internal
710
736#ifdef EIGEN_ALLOCA
737
738 #if EIGEN_DEFAULT_ALIGN_BYTES>0
739 // We always manually re-align the result of EIGEN_ALLOCA.
740 // If alloca is already aligned, the compiler should be smart enough to optimize away the re-alignment.
741 #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast<void*>((internal::UIntPtr(EIGEN_ALLOCA(SIZE+EIGEN_DEFAULT_ALIGN_BYTES-1)) + EIGEN_DEFAULT_ALIGN_BYTES-1) & ~(std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1)))
742 #else
743 #define EIGEN_ALIGNED_ALLOCA(SIZE) EIGEN_ALLOCA(SIZE)
744 #endif
745
746 #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
747 Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \
748 TYPE* NAME = (BUFFER)!=0 ? (BUFFER) \
749 : reinterpret_cast<TYPE*>( \
750 (sizeof(TYPE)*SIZE<=EIGEN_STACK_ALLOCATION_LIMIT) ? EIGEN_ALIGNED_ALLOCA(sizeof(TYPE)*SIZE) \
751 : Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE) ); \
752 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)
753
754
755 #define ei_declare_local_nested_eval(XPR_T,XPR,N,NAME) \
756 Eigen::internal::local_nested_eval_wrapper<XPR_T,N> EIGEN_CAT(NAME,_wrapper)(XPR, reinterpret_cast<typename XPR_T::Scalar*>( \
757 ( (Eigen::internal::local_nested_eval_wrapper<XPR_T,N>::NeedExternalBuffer) && ((sizeof(typename XPR_T::Scalar)*XPR.size())<=EIGEN_STACK_ALLOCATION_LIMIT) ) \
758 ? EIGEN_ALIGNED_ALLOCA( sizeof(typename XPR_T::Scalar)*XPR.size() ) : 0 ) ) ; \
759 typename Eigen::internal::local_nested_eval_wrapper<XPR_T,N>::ObjectType NAME(EIGEN_CAT(NAME,_wrapper).object)
760
761#else
762
763 #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
764 Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \
765 TYPE* NAME = (BUFFER)!=0 ? BUFFER : reinterpret_cast<TYPE*>(Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE)); \
766 Eigen::internal::aligned_stack_memory_handler<TYPE> EIGEN_CAT(NAME,_stack_memory_destructor)((BUFFER)==0 ? NAME : 0,SIZE,true)
767
768
769#define ei_declare_local_nested_eval(XPR_T,XPR,N,NAME) typename Eigen::internal::nested_eval<XPR_T,N>::type NAME(XPR)
770
771#endif
772
773
774/*****************************************************************************
775*** Implementation of EIGEN_MAKE_ALIGNED_OPERATOR_NEW [_IF] ***
776*****************************************************************************/
777
778#if EIGEN_HAS_CXX17_OVERALIGN
779
780// C++17 -> no need to bother about alignment anymore :)
781
782#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign)
783#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)
784#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW
785#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size)
786
787#else
788
789// HIP does not support new/delete on device.
790#if EIGEN_MAX_ALIGN_BYTES!=0 && !defined(EIGEN_HIP_DEVICE_COMPILE)
791 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
792 EIGEN_DEVICE_FUNC \
793 void* operator new(std::size_t size, const std::nothrow_t&) EIGEN_NO_THROW { \
794 EIGEN_TRY { return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); } \
795 EIGEN_CATCH (...) { return 0; } \
796 }
797 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign) \
798 EIGEN_DEVICE_FUNC \
799 void *operator new(std::size_t size) { \
800 return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
801 } \
802 EIGEN_DEVICE_FUNC \
803 void *operator new[](std::size_t size) { \
804 return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
805 } \
806 EIGEN_DEVICE_FUNC \
807 void operator delete(void * ptr) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
808 EIGEN_DEVICE_FUNC \
809 void operator delete[](void * ptr) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
810 EIGEN_DEVICE_FUNC \
811 void operator delete(void * ptr, std::size_t /* sz */) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
812 EIGEN_DEVICE_FUNC \
813 void operator delete[](void * ptr, std::size_t /* sz */) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
814 /* in-place new and delete. since (at least afaik) there is no actual */ \
815 /* memory allocated we can safely let the default implementation handle */ \
816 /* this particular case. */ \
817 EIGEN_DEVICE_FUNC \
818 static void *operator new(std::size_t size, void *ptr) { return ::operator new(size,ptr); } \
819 EIGEN_DEVICE_FUNC \
820 static void *operator new[](std::size_t size, void* ptr) { return ::operator new[](size,ptr); } \
821 EIGEN_DEVICE_FUNC \
822 void operator delete(void * memory, void *ptr) EIGEN_NO_THROW { return ::operator delete(memory,ptr); } \
823 EIGEN_DEVICE_FUNC \
824 void operator delete[](void * memory, void *ptr) EIGEN_NO_THROW { return ::operator delete[](memory,ptr); } \
825 /* nothrow-new (returns zero instead of std::bad_alloc) */ \
826 EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
827 EIGEN_DEVICE_FUNC \
828 void operator delete(void *ptr, const std::nothrow_t&) EIGEN_NO_THROW { \
829 Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); \
830 } \
831 typedef void eigen_aligned_operator_new_marker_type;
832#else
833 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)
834#endif
835
836#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(true)
837#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size) \
838 EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool( \
839 ((Size)!=Eigen::Dynamic) && \
840 (((EIGEN_MAX_ALIGN_BYTES>=16) && ((sizeof(Scalar)*(Size))%(EIGEN_MAX_ALIGN_BYTES )==0)) || \
841 ((EIGEN_MAX_ALIGN_BYTES>=32) && ((sizeof(Scalar)*(Size))%(EIGEN_MAX_ALIGN_BYTES/2)==0)) || \
842 ((EIGEN_MAX_ALIGN_BYTES>=64) && ((sizeof(Scalar)*(Size))%(EIGEN_MAX_ALIGN_BYTES/4)==0)) )))
843
844#endif
845
846/****************************************************************************/
847
872template<class T>
873class aligned_allocator : public std::allocator<T>
874{
875public:
876 typedef std::size_t size_type;
877 typedef std::ptrdiff_t difference_type;
878 typedef T* pointer;
879 typedef const T* const_pointer;
880 typedef T& reference;
881 typedef const T& const_reference;
882 typedef T value_type;
883
884 template<class U>
885 struct rebind
886 {
887 typedef aligned_allocator<U> other;
888 };
889
890 aligned_allocator() : std::allocator<T>() {}
891
892 aligned_allocator(const aligned_allocator& other) : std::allocator<T>(other) {}
893
894 template<class U>
895 aligned_allocator(const aligned_allocator<U>& other) : std::allocator<T>(other) {}
896
898
899 #if EIGEN_COMP_GNUC_STRICT && EIGEN_GNUC_AT_LEAST(7,0)
900 // In gcc std::allocator::max_size() is bugged making gcc triggers a warning:
901 // eigen/Eigen/src/Core/util/Memory.h:189:12: warning: argument 1 value '18446744073709551612' exceeds maximum object size 9223372036854775807
902 // See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87544
903 size_type max_size() const {
904 return (std::numeric_limits<std::ptrdiff_t>::max)()/sizeof(T);
905 }
906 #endif
907
908 pointer allocate(size_type num, const void* /*hint*/ = 0)
909 {
910 internal::check_size_for_overflow<T>(num);
911 return static_cast<pointer>( internal::aligned_malloc(num * sizeof(T)) );
912 }
913
914 void deallocate(pointer p, size_type /*num*/)
915 {
916 internal::aligned_free(p);
917 }
918};
919
920//---------- Cache sizes ----------
921
922#if !defined(EIGEN_NO_CPUID)
923# if EIGEN_COMP_GNUC && EIGEN_ARCH_i386_OR_x86_64
924# if defined(__PIC__) && EIGEN_ARCH_i386
925 // Case for x86 with PIC
926# define EIGEN_CPUID(abcd,func,id) \
927 __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));
928# elif defined(__PIC__) && EIGEN_ARCH_x86_64
929 // 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.
930 // However, we cannot detect which code model is used, and the xchg overhead is negligible anyway.
931# define EIGEN_CPUID(abcd,func,id) \
932 __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));
933# else
934 // Case for x86_64 or x86 w/o PIC
935# define EIGEN_CPUID(abcd,func,id) \
936 __asm__ __volatile__ ("cpuid": "=a" (abcd[0]), "=b" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "0" (func), "2" (id) );
937# endif
938# elif EIGEN_COMP_MSVC
939# if EIGEN_ARCH_i386_OR_x86_64
940# define EIGEN_CPUID(abcd,func,id) __cpuidex((int*)abcd,func,id)
941# endif
942# endif
943#endif
944
945namespace internal {
946
947#ifdef EIGEN_CPUID
948
949inline bool cpuid_is_vendor(int abcd[4], const int vendor[3])
950{
951 return abcd[1]==vendor[0] && abcd[3]==vendor[1] && abcd[2]==vendor[2];
952}
953
954inline void queryCacheSizes_intel_direct(int& l1, int& l2, int& l3)
955{
956 int abcd[4];
957 l1 = l2 = l3 = 0;
958 int cache_id = 0;
959 int cache_type = 0;
960 do {
961 abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
962 EIGEN_CPUID(abcd,0x4,cache_id);
963 cache_type = (abcd[0] & 0x0F) >> 0;
964 if(cache_type==1||cache_type==3) // data or unified cache
965 {
966 int cache_level = (abcd[0] & 0xE0) >> 5; // A[7:5]
967 int ways = (abcd[1] & 0xFFC00000) >> 22; // B[31:22]
968 int partitions = (abcd[1] & 0x003FF000) >> 12; // B[21:12]
969 int line_size = (abcd[1] & 0x00000FFF) >> 0; // B[11:0]
970 int sets = (abcd[2]); // C[31:0]
971
972 int cache_size = (ways+1) * (partitions+1) * (line_size+1) * (sets+1);
973
974 switch(cache_level)
975 {
976 case 1: l1 = cache_size; break;
977 case 2: l2 = cache_size; break;
978 case 3: l3 = cache_size; break;
979 default: break;
980 }
981 }
982 cache_id++;
983 } while(cache_type>0 && cache_id<16);
984}
985
986inline void queryCacheSizes_intel_codes(int& l1, int& l2, int& l3)
987{
988 int abcd[4];
989 abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
990 l1 = l2 = l3 = 0;
991 EIGEN_CPUID(abcd,0x00000002,0);
992 unsigned char * bytes = reinterpret_cast<unsigned char *>(abcd)+2;
993 bool check_for_p2_core2 = false;
994 for(int i=0; i<14; ++i)
995 {
996 switch(bytes[i])
997 {
998 case 0x0A: l1 = 8; break; // 0Ah data L1 cache, 8 KB, 2 ways, 32 byte lines
999 case 0x0C: l1 = 16; break; // 0Ch data L1 cache, 16 KB, 4 ways, 32 byte lines
1000 case 0x0E: l1 = 24; break; // 0Eh data L1 cache, 24 KB, 6 ways, 64 byte lines
1001 case 0x10: l1 = 16; break; // 10h data L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
1002 case 0x15: l1 = 16; break; // 15h code L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
1003 case 0x2C: l1 = 32; break; // 2Ch data L1 cache, 32 KB, 8 ways, 64 byte lines
1004 case 0x30: l1 = 32; break; // 30h code L1 cache, 32 KB, 8 ways, 64 byte lines
1005 case 0x60: l1 = 16; break; // 60h data L1 cache, 16 KB, 8 ways, 64 byte lines, sectored
1006 case 0x66: l1 = 8; break; // 66h data L1 cache, 8 KB, 4 ways, 64 byte lines, sectored
1007 case 0x67: l1 = 16; break; // 67h data L1 cache, 16 KB, 4 ways, 64 byte lines, sectored
1008 case 0x68: l1 = 32; break; // 68h data L1 cache, 32 KB, 4 ways, 64 byte lines, sectored
1009 case 0x1A: l2 = 96; break; // code and data L2 cache, 96 KB, 6 ways, 64 byte lines (IA-64)
1010 case 0x22: l3 = 512; break; // code and data L3 cache, 512 KB, 4 ways (!), 64 byte lines, dual-sectored
1011 case 0x23: l3 = 1024; break; // code and data L3 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
1012 case 0x25: l3 = 2048; break; // code and data L3 cache, 2048 KB, 8 ways, 64 byte lines, dual-sectored
1013 case 0x29: l3 = 4096; break; // code and data L3 cache, 4096 KB, 8 ways, 64 byte lines, dual-sectored
1014 case 0x39: l2 = 128; break; // code and data L2 cache, 128 KB, 4 ways, 64 byte lines, sectored
1015 case 0x3A: l2 = 192; break; // code and data L2 cache, 192 KB, 6 ways, 64 byte lines, sectored
1016 case 0x3B: l2 = 128; break; // code and data L2 cache, 128 KB, 2 ways, 64 byte lines, sectored
1017 case 0x3C: l2 = 256; break; // code and data L2 cache, 256 KB, 4 ways, 64 byte lines, sectored
1018 case 0x3D: l2 = 384; break; // code and data L2 cache, 384 KB, 6 ways, 64 byte lines, sectored
1019 case 0x3E: l2 = 512; break; // code and data L2 cache, 512 KB, 4 ways, 64 byte lines, sectored
1020 case 0x40: l2 = 0; break; // no integrated L2 cache (P6 core) or L3 cache (P4 core)
1021 case 0x41: l2 = 128; break; // code and data L2 cache, 128 KB, 4 ways, 32 byte lines
1022 case 0x42: l2 = 256; break; // code and data L2 cache, 256 KB, 4 ways, 32 byte lines
1023 case 0x43: l2 = 512; break; // code and data L2 cache, 512 KB, 4 ways, 32 byte lines
1024 case 0x44: l2 = 1024; break; // code and data L2 cache, 1024 KB, 4 ways, 32 byte lines
1025 case 0x45: l2 = 2048; break; // code and data L2 cache, 2048 KB, 4 ways, 32 byte lines
1026 case 0x46: l3 = 4096; break; // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines
1027 case 0x47: l3 = 8192; break; // code and data L3 cache, 8192 KB, 8 ways, 64 byte lines
1028 case 0x48: l2 = 3072; break; // code and data L2 cache, 3072 KB, 12 ways, 64 byte lines
1029 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
1030 case 0x4A: l3 = 6144; break; // code and data L3 cache, 6144 KB, 12 ways, 64 byte lines
1031 case 0x4B: l3 = 8192; break; // code and data L3 cache, 8192 KB, 16 ways, 64 byte lines
1032 case 0x4C: l3 = 12288; break; // code and data L3 cache, 12288 KB, 12 ways, 64 byte lines
1033 case 0x4D: l3 = 16384; break; // code and data L3 cache, 16384 KB, 16 ways, 64 byte lines
1034 case 0x4E: l2 = 6144; break; // code and data L2 cache, 6144 KB, 24 ways, 64 byte lines
1035 case 0x78: l2 = 1024; break; // code and data L2 cache, 1024 KB, 4 ways, 64 byte lines
1036 case 0x79: l2 = 128; break; // code and data L2 cache, 128 KB, 8 ways, 64 byte lines, dual-sectored
1037 case 0x7A: l2 = 256; break; // code and data L2 cache, 256 KB, 8 ways, 64 byte lines, dual-sectored
1038 case 0x7B: l2 = 512; break; // code and data L2 cache, 512 KB, 8 ways, 64 byte lines, dual-sectored
1039 case 0x7C: l2 = 1024; break; // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
1040 case 0x7D: l2 = 2048; break; // code and data L2 cache, 2048 KB, 8 ways, 64 byte lines
1041 case 0x7E: l2 = 256; break; // code and data L2 cache, 256 KB, 8 ways, 128 byte lines, sect. (IA-64)
1042 case 0x7F: l2 = 512; break; // code and data L2 cache, 512 KB, 2 ways, 64 byte lines
1043 case 0x80: l2 = 512; break; // code and data L2 cache, 512 KB, 8 ways, 64 byte lines
1044 case 0x81: l2 = 128; break; // code and data L2 cache, 128 KB, 8 ways, 32 byte lines
1045 case 0x82: l2 = 256; break; // code and data L2 cache, 256 KB, 8 ways, 32 byte lines
1046 case 0x83: l2 = 512; break; // code and data L2 cache, 512 KB, 8 ways, 32 byte lines
1047 case 0x84: l2 = 1024; break; // code and data L2 cache, 1024 KB, 8 ways, 32 byte lines
1048 case 0x85: l2 = 2048; break; // code and data L2 cache, 2048 KB, 8 ways, 32 byte lines
1049 case 0x86: l2 = 512; break; // code and data L2 cache, 512 KB, 4 ways, 64 byte lines
1050 case 0x87: l2 = 1024; break; // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines
1051 case 0x88: l3 = 2048; break; // code and data L3 cache, 2048 KB, 4 ways, 64 byte lines (IA-64)
1052 case 0x89: l3 = 4096; break; // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines (IA-64)
1053 case 0x8A: l3 = 8192; break; // code and data L3 cache, 8192 KB, 4 ways, 64 byte lines (IA-64)
1054 case 0x8D: l3 = 3072; break; // code and data L3 cache, 3072 KB, 12 ways, 128 byte lines (IA-64)
1055
1056 default: break;
1057 }
1058 }
1059 if(check_for_p2_core2 && l2 == l3)
1060 l3 = 0;
1061 l1 *= 1024;
1062 l2 *= 1024;
1063 l3 *= 1024;
1064}
1065
1066inline void queryCacheSizes_intel(int& l1, int& l2, int& l3, int max_std_funcs)
1067{
1068 if(max_std_funcs>=4)
1069 queryCacheSizes_intel_direct(l1,l2,l3);
1070 else if(max_std_funcs>=2)
1071 queryCacheSizes_intel_codes(l1,l2,l3);
1072 else
1073 l1 = l2 = l3 = 0;
1074}
1075
1076inline void queryCacheSizes_amd(int& l1, int& l2, int& l3)
1077{
1078 int abcd[4];
1079 abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
1080
1081 // First query the max supported function.
1082 EIGEN_CPUID(abcd,0x80000000,0);
1083 if(static_cast<numext::uint32_t>(abcd[0]) >= static_cast<numext::uint32_t>(0x80000006))
1084 {
1085 EIGEN_CPUID(abcd,0x80000005,0);
1086 l1 = (abcd[2] >> 24) * 1024; // C[31:24] = L1 size in KB
1087 abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
1088 EIGEN_CPUID(abcd,0x80000006,0);
1089 l2 = (abcd[2] >> 16) * 1024; // C[31;16] = l2 cache size in KB
1090 l3 = ((abcd[3] & 0xFFFC000) >> 18) * 512 * 1024; // D[31;18] = l3 cache size in 512KB
1091 }
1092 else
1093 {
1094 l1 = l2 = l3 = 0;
1095 }
1096}
1097#endif
1098
1101inline void queryCacheSizes(int& l1, int& l2, int& l3)
1102{
1103 #ifdef EIGEN_CPUID
1104 int abcd[4];
1105 const int GenuineIntel[] = {0x756e6547, 0x49656e69, 0x6c65746e};
1106 const int AuthenticAMD[] = {0x68747541, 0x69746e65, 0x444d4163};
1107 const int AMDisbetter_[] = {0x69444d41, 0x74656273, 0x21726574}; // "AMDisbetter!"
1108
1109 // identify the CPU vendor
1110 EIGEN_CPUID(abcd,0x0,0);
1111 int max_std_funcs = abcd[0];
1112 if(cpuid_is_vendor(abcd,GenuineIntel))
1113 queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
1114 else if(cpuid_is_vendor(abcd,AuthenticAMD) || cpuid_is_vendor(abcd,AMDisbetter_))
1115 queryCacheSizes_amd(l1,l2,l3);
1116 else
1117 // by default let's use Intel's API
1118 queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
1119
1120 // here is the list of other vendors:
1121// ||cpuid_is_vendor(abcd,"VIA VIA VIA ")
1122// ||cpuid_is_vendor(abcd,"CyrixInstead")
1123// ||cpuid_is_vendor(abcd,"CentaurHauls")
1124// ||cpuid_is_vendor(abcd,"GenuineTMx86")
1125// ||cpuid_is_vendor(abcd,"TransmetaCPU")
1126// ||cpuid_is_vendor(abcd,"RiseRiseRise")
1127// ||cpuid_is_vendor(abcd,"Geode by NSC")
1128// ||cpuid_is_vendor(abcd,"SiS SiS SiS ")
1129// ||cpuid_is_vendor(abcd,"UMC UMC UMC ")
1130// ||cpuid_is_vendor(abcd,"NexGenDriven")
1131 #else
1132 l1 = l2 = l3 = -1;
1133 #endif
1134}
1135
1138inline int queryL1CacheSize()
1139{
1140 int l1(-1), l2, l3;
1141 queryCacheSizes(l1,l2,l3);
1142 return l1;
1143}
1144
1147inline int queryTopLevelCacheSize()
1148{
1149 int l1, l2(-1), l3(-1);
1150 queryCacheSizes(l1,l2,l3);
1151 return (std::max)(l2,l3);
1152}
1153
1154} // end namespace internal
1155
1156} // end namespace Eigen
1157
1158#endif // EIGEN_MEMORY_H
STL compatible allocator to use with types requiring a non-standard alignment.
Definition: Memory.h:874
static const lastp1_t end
Definition: IndexedViewHelper.h:183
Namespace containing all symbols from the Eigen library.
Definition: B01_Experimental.dox:1
EIGEN_DEFAULT_DENSE_INDEX_TYPE Index
The Index type as used for the API.
Definition: Meta.h:59
const int Dynamic
Definition: Constants.h:24