1
// This file is part of Eigen, a lightweight C++ template library
4
// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
5
// Copyright (C) 2008-2009 Benoit Jacob <jacob.benoit.1@gmail.com>
6
// Copyright (C) 2009 Kenneth Riddile <kfriddile@yahoo.com>
7
// Copyright (C) 2010 Hauke Heibel <hauke.heibel@gmail.com>
8
// Copyright (C) 2010 Thomas Capricelli <orzel@freehackers.org>
10
// Eigen is free software; you can redistribute it and/or
11
// modify it under the terms of the GNU Lesser General Public
12
// License as published by the Free Software Foundation; either
13
// version 3 of the License, or (at your option) any later version.
15
// Alternatively, you can redistribute it and/or
16
// modify it under the terms of the GNU General Public License as
17
// published by the Free Software Foundation; either version 2 of
18
// the License, or (at your option) any later version.
20
// Eigen is distributed in the hope that it will be useful, but WITHOUT ANY
21
// WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
22
// FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License or the
23
// GNU General Public License for more details.
25
// You should have received a copy of the GNU Lesser General Public
26
// License and a copy of the GNU General Public License along with
27
// Eigen. If not, see <http://www.gnu.org/licenses/>.
30
/*****************************************************************************
31
*** Platform checks for aligned malloc functions ***
32
*****************************************************************************/
34
#ifndef EIGEN_MEMORY_H
35
#define EIGEN_MEMORY_H
37
// On 64-bit systems, glibc's malloc returns 16-byte-aligned pointers, see:
38
// http://www.gnu.org/s/libc/manual/html_node/Aligned-Memory-Blocks.html
39
// This is true at least since glibc 2.8.
40
// This leaves the question how to detect 64-bit. According to this document,
41
// http://gcc.fyxm.net/summit/2003/Porting%20to%2064%20bit.pdf
42
// page 114, "[The] LP64 model [...] is used by all 64-bit UNIX ports" so it's indeed
43
// quite safe, at least within the context of glibc, to equate 64-bit with LP64.
44
#if defined(__GLIBC__) && ((__GLIBC__>=2 && __GLIBC_MINOR__ >= 8) || __GLIBC__>2) \
46
#define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 1
48
#define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 0
51
// FreeBSD 6 seems to have 16-byte aligned malloc
52
// See http://svn.freebsd.org/viewvc/base/stable/6/lib/libc/stdlib/malloc.c?view=markup
53
// FreeBSD 7 seems to have 16-byte aligned malloc except on ARM and MIPS architectures
54
// See http://svn.freebsd.org/viewvc/base/stable/7/lib/libc/stdlib/malloc.c?view=markup
55
#if defined(__FreeBSD__) && !defined(__arm__) && !defined(__mips__)
56
#define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 1
58
#define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 0
61
#if defined(__APPLE__) \
63
|| EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED \
64
|| EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED
65
#define EIGEN_MALLOC_ALREADY_ALIGNED 1
67
#define EIGEN_MALLOC_ALREADY_ALIGNED 0
70
#if ((defined __QNXNTO__) || (defined _GNU_SOURCE) || ((defined _XOPEN_SOURCE) && (_XOPEN_SOURCE >= 600))) \
71
&& (defined _POSIX_ADVISORY_INFO) && (_POSIX_ADVISORY_INFO > 0)
72
#define EIGEN_HAS_POSIX_MEMALIGN 1
74
#define EIGEN_HAS_POSIX_MEMALIGN 0
77
#ifdef EIGEN_VECTORIZE_SSE
78
#define EIGEN_HAS_MM_MALLOC 1
80
#define EIGEN_HAS_MM_MALLOC 0
85
/*****************************************************************************
86
*** Implementation of handmade aligned functions ***
87
*****************************************************************************/
89
/* ----- Hand made implementations of aligned malloc/free and realloc ----- */
91
/** \internal Like malloc, but the returned pointer is guaranteed to be 16-byte aligned.
92
* Fast, but wastes 16 additional bytes of memory. Does not throw any exception.
94
inline void* handmade_aligned_malloc(size_t size)
96
void *original = std::malloc(size+16);
97
if (original == 0) return 0;
98
void *aligned = reinterpret_cast<void*>((reinterpret_cast<size_t>(original) & ~(size_t(15))) + 16);
99
*(reinterpret_cast<void**>(aligned) - 1) = original;
103
/** \internal Frees memory allocated with handmade_aligned_malloc */
104
inline void handmade_aligned_free(void *ptr)
106
if (ptr) std::free(*(reinterpret_cast<void**>(ptr) - 1));
110
* \brief Reallocates aligned memory.
111
* Since we know that our handmade version is based on std::realloc
112
* we can use std::realloc to implement efficient reallocation.
114
inline void* handmade_aligned_realloc(void* ptr, size_t size, size_t = 0)
116
if (ptr == 0) return handmade_aligned_malloc(size);
117
void *original = *(reinterpret_cast<void**>(ptr) - 1);
118
original = std::realloc(original,size+16);
119
if (original == 0) return 0;
120
void *aligned = reinterpret_cast<void*>((reinterpret_cast<size_t>(original) & ~(size_t(15))) + 16);
121
*(reinterpret_cast<void**>(aligned) - 1) = original;
125
/*****************************************************************************
126
*** Implementation of generic aligned realloc (when no realloc can be used)***
127
*****************************************************************************/
129
void* aligned_malloc(size_t size);
130
void aligned_free(void *ptr);
133
* \brief Reallocates aligned memory.
134
* Allows reallocation with aligned ptr types. This implementation will
135
* always create a new memory chunk and copy the old data.
137
inline void* generic_aligned_realloc(void* ptr, size_t size, size_t old_size)
140
return aligned_malloc(size);
148
void* newptr = aligned_malloc(size);
151
#ifdef EIGEN_HAS_ERRNO
152
errno = ENOMEM; // according to the standard
159
std::memcpy(newptr, ptr, (std::min)(size,old_size));
166
/*****************************************************************************
167
*** Implementation of portable aligned versions of malloc/free/realloc ***
168
*****************************************************************************/
170
#ifdef EIGEN_NO_MALLOC
171
inline void check_that_malloc_is_allowed()
173
eigen_assert(false && "heap allocation is forbidden (EIGEN_NO_MALLOC is defined)");
175
#elif defined EIGEN_RUNTIME_NO_MALLOC
176
inline bool is_malloc_allowed_impl(bool update, bool new_value = false)
178
static bool value = true;
183
inline bool is_malloc_allowed() { return is_malloc_allowed_impl(false); }
184
inline bool set_is_malloc_allowed(bool new_value) { return is_malloc_allowed_impl(true, new_value); }
185
inline void check_that_malloc_is_allowed()
187
eigen_assert(is_malloc_allowed() && "heap allocation is forbidden (EIGEN_RUNTIME_NO_MALLOC is defined and g_is_malloc_allowed is false)");
190
inline void check_that_malloc_is_allowed()
194
/** \internal Allocates \a size bytes. The returned pointer is guaranteed to have 16 bytes alignment.
195
* On allocation error, the returned pointer is null, and if exceptions are enabled then a std::bad_alloc is thrown.
197
inline void* aligned_malloc(size_t size)
199
check_that_malloc_is_allowed();
203
result = std::malloc(size);
204
#elif EIGEN_MALLOC_ALREADY_ALIGNED
205
result = std::malloc(size);
206
#elif EIGEN_HAS_POSIX_MEMALIGN
207
if(posix_memalign(&result, 16, size)) result = 0;
208
#elif EIGEN_HAS_MM_MALLOC
209
result = _mm_malloc(size, 16);
210
#elif (defined _MSC_VER)
211
result = _aligned_malloc(size, 16);
213
result = handmade_aligned_malloc(size);
216
#ifdef EIGEN_EXCEPTIONS
218
throw std::bad_alloc();
223
/** \internal Frees memory allocated with aligned_malloc. */
224
inline void aligned_free(void *ptr)
228
#elif EIGEN_MALLOC_ALREADY_ALIGNED
230
#elif EIGEN_HAS_POSIX_MEMALIGN
232
#elif EIGEN_HAS_MM_MALLOC
234
#elif defined(_MSC_VER)
237
handmade_aligned_free(ptr);
243
* \brief Reallocates an aligned block of memory.
244
* \throws std::bad_alloc if EIGEN_EXCEPTIONS are defined.
246
inline void* aligned_realloc(void *ptr, size_t new_size, size_t old_size)
248
EIGEN_UNUSED_VARIABLE(old_size);
252
result = std::realloc(ptr,new_size);
253
#elif EIGEN_MALLOC_ALREADY_ALIGNED
254
result = std::realloc(ptr,new_size);
255
#elif EIGEN_HAS_POSIX_MEMALIGN
256
result = generic_aligned_realloc(ptr,new_size,old_size);
257
#elif EIGEN_HAS_MM_MALLOC
258
// The defined(_mm_free) is just here to verify that this MSVC version
259
// implements _mm_malloc/_mm_free based on the corresponding _aligned_
260
// functions. This may not always be the case and we just try to be safe.
261
#if defined(_MSC_VER) && defined(_mm_free)
262
result = _aligned_realloc(ptr,new_size,16);
264
result = generic_aligned_realloc(ptr,new_size,old_size);
266
#elif defined(_MSC_VER)
267
result = _aligned_realloc(ptr,new_size,16);
269
result = handmade_aligned_realloc(ptr,new_size,old_size);
272
#ifdef EIGEN_EXCEPTIONS
273
if (result==0 && new_size!=0)
274
throw std::bad_alloc();
279
/*****************************************************************************
280
*** Implementation of conditionally aligned functions ***
281
*****************************************************************************/
283
/** \internal Allocates \a size bytes. If Align is true, then the returned ptr is 16-byte-aligned.
284
* On allocation error, the returned pointer is null, and if exceptions are enabled then a std::bad_alloc is thrown.
286
template<bool Align> inline void* conditional_aligned_malloc(size_t size)
288
return aligned_malloc(size);
291
template<> inline void* conditional_aligned_malloc<false>(size_t size)
293
check_that_malloc_is_allowed();
295
void *result = std::malloc(size);
296
#ifdef EIGEN_EXCEPTIONS
297
if(!result) throw std::bad_alloc();
302
/** \internal Frees memory allocated with conditional_aligned_malloc */
303
template<bool Align> inline void conditional_aligned_free(void *ptr)
308
template<> inline void conditional_aligned_free<false>(void *ptr)
313
template<bool Align> inline void* conditional_aligned_realloc(void* ptr, size_t new_size, size_t old_size)
315
return aligned_realloc(ptr, new_size, old_size);
318
template<> inline void* conditional_aligned_realloc<false>(void* ptr, size_t new_size, size_t)
320
return std::realloc(ptr, new_size);
323
/*****************************************************************************
324
*** Construction/destruction of array elements ***
325
*****************************************************************************/
327
/** \internal Constructs the elements of an array.
328
* The \a size parameter tells on how many objects to call the constructor of T.
330
template<typename T> inline T* construct_elements_of_array(T *ptr, size_t size)
332
for (size_t i=0; i < size; ++i) ::new (ptr + i) T;
336
/** \internal Destructs the elements of an array.
337
* The \a size parameters tells on how many objects to call the destructor of T.
339
template<typename T> inline void destruct_elements_of_array(T *ptr, size_t size)
341
// always destruct an array starting from the end.
343
while(size) ptr[--size].~T();
346
/*****************************************************************************
347
*** Implementation of aligned new/delete-like functions ***
348
*****************************************************************************/
350
/** \internal Allocates \a size objects of type T. The returned pointer is guaranteed to have 16 bytes alignment.
351
* On allocation error, the returned pointer is undefined, but if exceptions are enabled then a std::bad_alloc is thrown.
352
* The default constructor of T is called.
354
template<typename T> inline T* aligned_new(size_t size)
356
T *result = reinterpret_cast<T*>(aligned_malloc(sizeof(T)*size));
357
return construct_elements_of_array(result, size);
360
template<typename T, bool Align> inline T* conditional_aligned_new(size_t size)
362
T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
363
return construct_elements_of_array(result, size);
366
/** \internal Deletes objects constructed with aligned_new
367
* The \a size parameters tells on how many objects to call the destructor of T.
369
template<typename T> inline void aligned_delete(T *ptr, size_t size)
371
destruct_elements_of_array<T>(ptr, size);
375
/** \internal Deletes objects constructed with conditional_aligned_new
376
* The \a size parameters tells on how many objects to call the destructor of T.
378
template<typename T, bool Align> inline void conditional_aligned_delete(T *ptr, size_t size)
380
destruct_elements_of_array<T>(ptr, size);
381
conditional_aligned_free<Align>(ptr);
384
template<typename T, bool Align> inline T* conditional_aligned_realloc_new(T* pts, size_t new_size, size_t old_size)
386
if(new_size < old_size)
387
destruct_elements_of_array(pts+new_size, old_size-new_size);
388
T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
389
if(new_size > old_size)
390
construct_elements_of_array(result+old_size, new_size-old_size);
395
template<typename T, bool Align> inline T* conditional_aligned_new_auto(size_t size)
397
T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
398
if(NumTraits<T>::RequireInitialization)
399
construct_elements_of_array(result, size);
403
template<typename T, bool Align> inline T* conditional_aligned_realloc_new_auto(T* pts, size_t new_size, size_t old_size)
405
if(NumTraits<T>::RequireInitialization && (new_size < old_size))
406
destruct_elements_of_array(pts+new_size, old_size-new_size);
407
T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
408
if(NumTraits<T>::RequireInitialization && (new_size > old_size))
409
construct_elements_of_array(result+old_size, new_size-old_size);
413
template<typename T, bool Align> inline void conditional_aligned_delete_auto(T *ptr, size_t size)
415
if(NumTraits<T>::RequireInitialization)
416
destruct_elements_of_array<T>(ptr, size);
417
conditional_aligned_free<Align>(ptr);
420
/****************************************************************************/
422
/** \internal Returns the index of the first element of the array that is well aligned for vectorization.
424
* \param array the address of the start of the array
425
* \param size the size of the array
427
* \note If no element of the array is well aligned, the size of the array is returned. Typically,
428
* for example with SSE, "well aligned" means 16-byte-aligned. If vectorization is disabled or if the
429
* packet size for the given scalar type is 1, then everything is considered well-aligned.
431
* \note If the scalar type is vectorizable, we rely on the following assumptions: sizeof(Scalar) is a
432
* power of 2, the packet size in bytes is also a power of 2, and is a multiple of sizeof(Scalar). On the
433
* other hand, we do not assume that the array address is a multiple of sizeof(Scalar), as that fails for
434
* example with Scalar=double on certain 32-bit platforms, see bug #79.
436
* There is also the variant first_aligned(const MatrixBase&) defined in DenseCoeffsBase.h.
438
template<typename Scalar, typename Index>
439
inline static Index first_aligned(const Scalar* array, Index size)
441
typedef typename packet_traits<Scalar>::type Packet;
442
enum { PacketSize = packet_traits<Scalar>::size,
443
PacketAlignedMask = PacketSize-1
448
// Either there is no vectorization, or a packet consists of exactly 1 scalar so that all elements
449
// of the array have the same alignment.
452
else if(size_t(array) & (sizeof(Scalar)-1))
454
// There is vectorization for this scalar type, but the array is not aligned to the size of a single scalar.
455
// Consequently, no element of the array is well aligned.
460
return std::min<Index>( (PacketSize - (Index((size_t(array)/sizeof(Scalar))) & PacketAlignedMask))
461
& PacketAlignedMask, size);
465
} // end namespace internal
467
/*****************************************************************************
468
*** Implementation of runtime stack allocation (falling back to malloc) ***
469
*****************************************************************************/
471
// you can overwrite Eigen's default behavior regarding alloca by defining EIGEN_ALLOCA
472
// to the appropriate stack allocation function
474
#if (defined __linux__)
475
#define EIGEN_ALLOCA alloca
476
#elif defined(_MSC_VER)
477
#define EIGEN_ALLOCA _alloca
483
// This helper class construct the allocated memory, and takes care of destructing and freeing the handled data
484
// at destruction time. In practice this helper class is mainly useful to avoid memory leak in case of exceptions.
485
template<typename T> class aligned_stack_memory_handler
488
/* Creates a stack_memory_handler responsible for the buffer \a ptr of size \a size.
489
* Note that \a ptr can be 0 regardless of the other parameters.
490
* This constructor takes care of constructing/initializing the elements of the buffer if required by the scalar type T (see NumTraits<T>::RequireInitialization).
491
* In this case, the buffer elements will also be destructed when this handler will be destructed.
492
* Finally, if \a dealloc is true, then the pointer \a ptr is freed.
494
aligned_stack_memory_handler(T* ptr, size_t size, bool dealloc)
495
: m_ptr(ptr), m_size(size), m_deallocate(dealloc)
497
if(NumTraits<T>::RequireInitialization && m_ptr)
498
Eigen::internal::construct_elements_of_array(m_ptr, size);
500
~aligned_stack_memory_handler()
502
if(NumTraits<T>::RequireInitialization && m_ptr)
503
Eigen::internal::destruct_elements_of_array<T>(m_ptr, m_size);
505
Eigen::internal::aligned_free(m_ptr);
516
* Declares, allocates and construct an aligned buffer named NAME of SIZE elements of type TYPE on the stack
517
* if SIZE is smaller than EIGEN_STACK_ALLOCATION_LIMIT, and if stack allocation is supported by the platform
518
* (currently, this is Linux and Visual Studio only). Otherwise the memory is allocated on the heap.
519
* The allocated buffer is automatically deleted when exiting the scope of this declaration.
520
* If BUFFER is non nul, then the declared variable is simply an alias for BUFFER, and no allocation/deletion occurs.
521
* Here is an example:
524
* ei_declare_aligned_stack_constructed_variable(float,data,size,0);
525
* // use data[0] to data[size-1]
528
* The underlying stack allocation function can controlled with the EIGEN_ALLOCA preprocessor token.
533
#define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast<void*>((reinterpret_cast<size_t>(EIGEN_ALLOCA(SIZE+16)) & ~(size_t(15))) + 16)
535
#define EIGEN_ALIGNED_ALLOCA EIGEN_ALLOCA
538
#define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
539
TYPE* NAME = (BUFFER)!=0 ? (BUFFER) \
540
: reinterpret_cast<TYPE*>( \
541
(sizeof(TYPE)*SIZE<=EIGEN_STACK_ALLOCATION_LIMIT) ? EIGEN_ALIGNED_ALLOCA(sizeof(TYPE)*SIZE) \
542
: Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE) ); \
543
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)
547
#define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
548
TYPE* NAME = (BUFFER)!=0 ? BUFFER : reinterpret_cast<TYPE*>(Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE)); \
549
Eigen::internal::aligned_stack_memory_handler<TYPE> EIGEN_CAT(NAME,_stack_memory_destructor)((BUFFER)==0 ? NAME : 0,SIZE,true)
554
/*****************************************************************************
555
*** Implementation of EIGEN_MAKE_ALIGNED_OPERATOR_NEW [_IF] ***
556
*****************************************************************************/
559
#ifdef EIGEN_EXCEPTIONS
560
#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
561
void* operator new(size_t size, const std::nothrow_t&) throw() { \
562
try { return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); } \
563
catch (...) { return 0; } \
567
#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
568
void* operator new(size_t size, const std::nothrow_t&) throw() { \
569
return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
573
#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign) \
574
void *operator new(size_t size) { \
575
return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
577
void *operator new[](size_t size) { \
578
return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
580
void operator delete(void * ptr) throw() { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
581
void operator delete[](void * ptr) throw() { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
582
/* in-place new and delete. since (at least afaik) there is no actual */ \
583
/* memory allocated we can safely let the default implementation handle */ \
584
/* this particular case. */ \
585
static void *operator new(size_t size, void *ptr) { return ::operator new(size,ptr); } \
586
void operator delete(void * memory, void *ptr) throw() { return ::operator delete(memory,ptr); } \
587
/* nothrow-new (returns zero instead of std::bad_alloc) */ \
588
EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
589
void operator delete(void *ptr, const std::nothrow_t&) throw() { \
590
Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); \
592
typedef void eigen_aligned_operator_new_marker_type;
594
#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)
597
#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(true)
598
#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size) \
599
EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(((Size)!=Eigen::Dynamic) && ((sizeof(Scalar)*(Size))%16==0))
601
/****************************************************************************/
603
/** \class aligned_allocator
604
* \ingroup Core_Module
606
* \brief STL compatible allocator to use with with 16 byte aligned types
610
* // Matrix4f requires 16 bytes alignment:
611
* std::map< int, Matrix4f, std::less<int>,
612
* aligned_allocator<std::pair<const int, Matrix4f> > > my_map_mat4;
613
* // Vector3f does not require 16 bytes alignment, no need to use Eigen's allocator:
614
* std::map< int, Vector3f > my_map_vec3;
617
* \sa \ref TopicStlContainers.
620
class aligned_allocator
623
typedef size_t size_type;
624
typedef std::ptrdiff_t difference_type;
626
typedef const T* const_pointer;
627
typedef T& reference;
628
typedef const T& const_reference;
629
typedef T value_type;
634
typedef aligned_allocator<U> other;
637
pointer address( reference value ) const
642
const_pointer address( const_reference value ) const
647
aligned_allocator() throw()
651
aligned_allocator( const aligned_allocator& ) throw()
656
aligned_allocator( const aligned_allocator<U>& ) throw()
660
~aligned_allocator() throw()
664
size_type max_size() const throw()
666
return (std::numeric_limits<size_type>::max)();
669
pointer allocate( size_type num, const void* hint = 0 )
671
EIGEN_UNUSED_VARIABLE(hint);
672
return static_cast<pointer>( internal::aligned_malloc( num * sizeof(T) ) );
675
void construct( pointer p, const T& value )
677
::new( p ) T( value );
680
void destroy( pointer p )
685
void deallocate( pointer p, size_type /*num*/ )
687
internal::aligned_free( p );
690
bool operator!=(const aligned_allocator<T>& ) const
693
bool operator==(const aligned_allocator<T>& ) const
697
//---------- Cache sizes ----------
699
#if defined(__GNUC__) && ( defined(__i386__) || defined(__x86_64__) )
700
# if defined(__PIC__) && defined(__i386__)
701
// Case for x86 with PIC
702
# define EIGEN_CPUID(abcd,func,id) \
703
__asm__ __volatile__ ("xchgl %%ebx, %%esi;cpuid; xchgl %%ebx,%%esi": "=a" (abcd[0]), "=S" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "a" (func), "c" (id));
705
// Case for x86_64 or x86 w/o PIC
706
# define EIGEN_CPUID(abcd,func,id) \
707
__asm__ __volatile__ ("cpuid": "=a" (abcd[0]), "=b" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "a" (func), "c" (id) );
709
#elif defined(_MSC_VER)
710
# if (_MSC_VER > 1500)
711
# define EIGEN_CPUID(abcd,func,id) __cpuidex((int*)abcd,func,id)
719
inline bool cpuid_is_vendor(int abcd[4], const char* vendor)
721
return abcd[1]==((int*)(vendor))[0] && abcd[3]==((int*)(vendor))[1] && abcd[2]==((int*)(vendor))[2];
724
inline void queryCacheSizes_intel_direct(int& l1, int& l2, int& l3)
731
abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
732
EIGEN_CPUID(abcd,0x4,cache_id);
733
cache_type = (abcd[0] & 0x0F) >> 0;
734
if(cache_type==1||cache_type==3) // data or unified cache
736
int cache_level = (abcd[0] & 0xE0) >> 5; // A[7:5]
737
int ways = (abcd[1] & 0xFFC00000) >> 22; // B[31:22]
738
int partitions = (abcd[1] & 0x003FF000) >> 12; // B[21:12]
739
int line_size = (abcd[1] & 0x00000FFF) >> 0; // B[11:0]
740
int sets = (abcd[2]); // C[31:0]
742
int cache_size = (ways+1) * (partitions+1) * (line_size+1) * (sets+1);
746
case 1: l1 = cache_size; break;
747
case 2: l2 = cache_size; break;
748
case 3: l3 = cache_size; break;
753
} while(cache_type>0 && cache_id<16);
756
inline void queryCacheSizes_intel_codes(int& l1, int& l2, int& l3)
759
abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
761
EIGEN_CPUID(abcd,0x00000002,0);
762
unsigned char * bytes = reinterpret_cast<unsigned char *>(abcd)+2;
763
bool check_for_p2_core2 = false;
764
for(int i=0; i<14; ++i)
768
case 0x0A: l1 = 8; break; // 0Ah data L1 cache, 8 KB, 2 ways, 32 byte lines
769
case 0x0C: l1 = 16; break; // 0Ch data L1 cache, 16 KB, 4 ways, 32 byte lines
770
case 0x0E: l1 = 24; break; // 0Eh data L1 cache, 24 KB, 6 ways, 64 byte lines
771
case 0x10: l1 = 16; break; // 10h data L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
772
case 0x15: l1 = 16; break; // 15h code L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
773
case 0x2C: l1 = 32; break; // 2Ch data L1 cache, 32 KB, 8 ways, 64 byte lines
774
case 0x30: l1 = 32; break; // 30h code L1 cache, 32 KB, 8 ways, 64 byte lines
775
case 0x60: l1 = 16; break; // 60h data L1 cache, 16 KB, 8 ways, 64 byte lines, sectored
776
case 0x66: l1 = 8; break; // 66h data L1 cache, 8 KB, 4 ways, 64 byte lines, sectored
777
case 0x67: l1 = 16; break; // 67h data L1 cache, 16 KB, 4 ways, 64 byte lines, sectored
778
case 0x68: l1 = 32; break; // 68h data L1 cache, 32 KB, 4 ways, 64 byte lines, sectored
779
case 0x1A: l2 = 96; break; // code and data L2 cache, 96 KB, 6 ways, 64 byte lines (IA-64)
780
case 0x22: l3 = 512; break; // code and data L3 cache, 512 KB, 4 ways (!), 64 byte lines, dual-sectored
781
case 0x23: l3 = 1024; break; // code and data L3 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
782
case 0x25: l3 = 2048; break; // code and data L3 cache, 2048 KB, 8 ways, 64 byte lines, dual-sectored
783
case 0x29: l3 = 4096; break; // code and data L3 cache, 4096 KB, 8 ways, 64 byte lines, dual-sectored
784
case 0x39: l2 = 128; break; // code and data L2 cache, 128 KB, 4 ways, 64 byte lines, sectored
785
case 0x3A: l2 = 192; break; // code and data L2 cache, 192 KB, 6 ways, 64 byte lines, sectored
786
case 0x3B: l2 = 128; break; // code and data L2 cache, 128 KB, 2 ways, 64 byte lines, sectored
787
case 0x3C: l2 = 256; break; // code and data L2 cache, 256 KB, 4 ways, 64 byte lines, sectored
788
case 0x3D: l2 = 384; break; // code and data L2 cache, 384 KB, 6 ways, 64 byte lines, sectored
789
case 0x3E: l2 = 512; break; // code and data L2 cache, 512 KB, 4 ways, 64 byte lines, sectored
790
case 0x40: l2 = 0; break; // no integrated L2 cache (P6 core) or L3 cache (P4 core)
791
case 0x41: l2 = 128; break; // code and data L2 cache, 128 KB, 4 ways, 32 byte lines
792
case 0x42: l2 = 256; break; // code and data L2 cache, 256 KB, 4 ways, 32 byte lines
793
case 0x43: l2 = 512; break; // code and data L2 cache, 512 KB, 4 ways, 32 byte lines
794
case 0x44: l2 = 1024; break; // code and data L2 cache, 1024 KB, 4 ways, 32 byte lines
795
case 0x45: l2 = 2048; break; // code and data L2 cache, 2048 KB, 4 ways, 32 byte lines
796
case 0x46: l3 = 4096; break; // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines
797
case 0x47: l3 = 8192; break; // code and data L3 cache, 8192 KB, 8 ways, 64 byte lines
798
case 0x48: l2 = 3072; break; // code and data L2 cache, 3072 KB, 12 ways, 64 byte lines
799
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
800
case 0x4A: l3 = 6144; break; // code and data L3 cache, 6144 KB, 12 ways, 64 byte lines
801
case 0x4B: l3 = 8192; break; // code and data L3 cache, 8192 KB, 16 ways, 64 byte lines
802
case 0x4C: l3 = 12288; break; // code and data L3 cache, 12288 KB, 12 ways, 64 byte lines
803
case 0x4D: l3 = 16384; break; // code and data L3 cache, 16384 KB, 16 ways, 64 byte lines
804
case 0x4E: l2 = 6144; break; // code and data L2 cache, 6144 KB, 24 ways, 64 byte lines
805
case 0x78: l2 = 1024; break; // code and data L2 cache, 1024 KB, 4 ways, 64 byte lines
806
case 0x79: l2 = 128; break; // code and data L2 cache, 128 KB, 8 ways, 64 byte lines, dual-sectored
807
case 0x7A: l2 = 256; break; // code and data L2 cache, 256 KB, 8 ways, 64 byte lines, dual-sectored
808
case 0x7B: l2 = 512; break; // code and data L2 cache, 512 KB, 8 ways, 64 byte lines, dual-sectored
809
case 0x7C: l2 = 1024; break; // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
810
case 0x7D: l2 = 2048; break; // code and data L2 cache, 2048 KB, 8 ways, 64 byte lines
811
case 0x7E: l2 = 256; break; // code and data L2 cache, 256 KB, 8 ways, 128 byte lines, sect. (IA-64)
812
case 0x7F: l2 = 512; break; // code and data L2 cache, 512 KB, 2 ways, 64 byte lines
813
case 0x80: l2 = 512; break; // code and data L2 cache, 512 KB, 8 ways, 64 byte lines
814
case 0x81: l2 = 128; break; // code and data L2 cache, 128 KB, 8 ways, 32 byte lines
815
case 0x82: l2 = 256; break; // code and data L2 cache, 256 KB, 8 ways, 32 byte lines
816
case 0x83: l2 = 512; break; // code and data L2 cache, 512 KB, 8 ways, 32 byte lines
817
case 0x84: l2 = 1024; break; // code and data L2 cache, 1024 KB, 8 ways, 32 byte lines
818
case 0x85: l2 = 2048; break; // code and data L2 cache, 2048 KB, 8 ways, 32 byte lines
819
case 0x86: l2 = 512; break; // code and data L2 cache, 512 KB, 4 ways, 64 byte lines
820
case 0x87: l2 = 1024; break; // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines
821
case 0x88: l3 = 2048; break; // code and data L3 cache, 2048 KB, 4 ways, 64 byte lines (IA-64)
822
case 0x89: l3 = 4096; break; // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines (IA-64)
823
case 0x8A: l3 = 8192; break; // code and data L3 cache, 8192 KB, 4 ways, 64 byte lines (IA-64)
824
case 0x8D: l3 = 3072; break; // code and data L3 cache, 3072 KB, 12 ways, 128 byte lines (IA-64)
829
if(check_for_p2_core2 && l2 == l3)
836
inline void queryCacheSizes_intel(int& l1, int& l2, int& l3, int max_std_funcs)
839
queryCacheSizes_intel_direct(l1,l2,l3);
841
queryCacheSizes_intel_codes(l1,l2,l3);
844
inline void queryCacheSizes_amd(int& l1, int& l2, int& l3)
847
abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
848
EIGEN_CPUID(abcd,0x80000005,0);
849
l1 = (abcd[2] >> 24) * 1024; // C[31:24] = L1 size in KB
850
abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
851
EIGEN_CPUID(abcd,0x80000006,0);
852
l2 = (abcd[2] >> 16) * 1024; // C[31;16] = l2 cache size in KB
853
l3 = ((abcd[3] & 0xFFFC000) >> 18) * 512 * 1024; // D[31;18] = l3 cache size in 512KB
858
* Queries and returns the cache sizes in Bytes of the L1, L2, and L3 data caches respectively */
859
inline void queryCacheSizes(int& l1, int& l2, int& l3)
864
// identify the CPU vendor
865
EIGEN_CPUID(abcd,0x0,0);
866
int max_std_funcs = abcd[1];
867
if(cpuid_is_vendor(abcd,"GenuineIntel"))
868
queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
869
else if(cpuid_is_vendor(abcd,"AuthenticAMD") || cpuid_is_vendor(abcd,"AMDisbetter!"))
870
queryCacheSizes_amd(l1,l2,l3);
872
// by default let's use Intel's API
873
queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
875
// here is the list of other vendors:
876
// ||cpuid_is_vendor(abcd,"VIA VIA VIA ")
877
// ||cpuid_is_vendor(abcd,"CyrixInstead")
878
// ||cpuid_is_vendor(abcd,"CentaurHauls")
879
// ||cpuid_is_vendor(abcd,"GenuineTMx86")
880
// ||cpuid_is_vendor(abcd,"TransmetaCPU")
881
// ||cpuid_is_vendor(abcd,"RiseRiseRise")
882
// ||cpuid_is_vendor(abcd,"Geode by NSC")
883
// ||cpuid_is_vendor(abcd,"SiS SiS SiS ")
884
// ||cpuid_is_vendor(abcd,"UMC UMC UMC ")
885
// ||cpuid_is_vendor(abcd,"NexGenDriven")
892
* \returns the size in Bytes of the L1 data cache */
893
inline int queryL1CacheSize()
896
queryCacheSizes(l1,l2,l3);
901
* \returns the size in Bytes of the L2 or L3 cache if this later is present */
902
inline int queryTopLevelCacheSize()
904
int l1, l2(-1), l3(-1);
905
queryCacheSizes(l1,l2,l3);
906
return (std::max)(l2,l3);
909
} // end namespace internal
911
#endif // EIGEN_MEMORY_H