Memory.h
Go to the documentation of this file.
00001 // This file is part of Eigen, a lightweight C++ template library
00002 // for linear algebra.
00003 //
00004 // Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
00005 // Copyright (C) 2008-2009 Benoit Jacob <jacob.benoit.1@gmail.com>
00006 // Copyright (C) 2009 Kenneth Riddile <kfriddile@yahoo.com>
00007 // Copyright (C) 2010 Hauke Heibel <hauke.heibel@gmail.com>
00008 // Copyright (C) 2010 Thomas Capricelli <orzel@freehackers.org>
00009 //
00010 // Eigen is free software; you can redistribute it and/or
00011 // modify it under the terms of the GNU Lesser General Public
00012 // License as published by the Free Software Foundation; either
00013 // version 3 of the License, or (at your option) any later version.
00014 //
00015 // Alternatively, you can redistribute it and/or
00016 // modify it under the terms of the GNU General Public License as
00017 // published by the Free Software Foundation; either version 2 of
00018 // the License, or (at your option) any later version.
00019 //
00020 // Eigen is distributed in the hope that it will be useful, but WITHOUT ANY
00021 // WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
00022 // FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License or the
00023 // GNU General Public License for more details.
00024 //
00025 // You should have received a copy of the GNU Lesser General Public
00026 // License and a copy of the GNU General Public License along with
00027 // Eigen. If not, see <http://www.gnu.org/licenses/>.
00028 
00029 
00030 /*****************************************************************************
00031 *** Platform checks for aligned malloc functions                           ***
00032 *****************************************************************************/
00033 
00034 #ifndef EIGEN_MEMORY_H
00035 #define EIGEN_MEMORY_H
00036 
00037 // On 64-bit systems, glibc's malloc returns 16-byte-aligned pointers, see:
00038 //   http://www.gnu.org/s/libc/manual/html_node/Aligned-Memory-Blocks.html
00039 // This is true at least since glibc 2.8.
00040 // This leaves the question how to detect 64-bit. According to this document,
00041 //   http://gcc.fyxm.net/summit/2003/Porting%20to%2064%20bit.pdf
00042 // page 114, "[The] LP64 model [...] is used by all 64-bit UNIX ports" so it's indeed
00043 // quite safe, at least within the context of glibc, to equate 64-bit with LP64.
00044 #if defined(__GLIBC__) && ((__GLIBC__>=2 && __GLIBC_MINOR__ >= 8) || __GLIBC__>2) \
00045  && defined(__LP64__)
00046   #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 1
00047 #else
00048   #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 0
00049 #endif
00050 
00051 // FreeBSD 6 seems to have 16-byte aligned malloc
00052 //   See http://svn.freebsd.org/viewvc/base/stable/6/lib/libc/stdlib/malloc.c?view=markup
00053 // FreeBSD 7 seems to have 16-byte aligned malloc except on ARM and MIPS architectures
00054 //   See http://svn.freebsd.org/viewvc/base/stable/7/lib/libc/stdlib/malloc.c?view=markup
00055 #if defined(__FreeBSD__) && !defined(__arm__) && !defined(__mips__)
00056   #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 1
00057 #else
00058   #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 0
00059 #endif
00060 
00061 #if defined(__APPLE__) \
00062  || defined(_WIN64) \
00063  || EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED \
00064  || EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED
00065   #define EIGEN_MALLOC_ALREADY_ALIGNED 1
00066 #else
00067   #define EIGEN_MALLOC_ALREADY_ALIGNED 0
00068 #endif
00069 
00070 #if ((defined __QNXNTO__) || (defined _GNU_SOURCE) || ((defined _XOPEN_SOURCE) && (_XOPEN_SOURCE >= 600))) \
00071  && (defined _POSIX_ADVISORY_INFO) && (_POSIX_ADVISORY_INFO > 0)
00072   #define EIGEN_HAS_POSIX_MEMALIGN 1
00073 #else
00074   #define EIGEN_HAS_POSIX_MEMALIGN 0
00075 #endif
00076 
00077 #ifdef EIGEN_VECTORIZE_SSE
00078   #define EIGEN_HAS_MM_MALLOC 1
00079 #else
00080   #define EIGEN_HAS_MM_MALLOC 0
00081 #endif
00082 
00083 namespace internal {
00084 
00085 /*****************************************************************************
00086 *** Implementation of handmade aligned functions                           ***
00087 *****************************************************************************/
00088 
00089 /* ----- Hand made implementations of aligned malloc/free and realloc ----- */
00090 
00094 inline void* handmade_aligned_malloc(size_t size)
00095 {
00096   void *original = std::malloc(size+16);
00097   if (original == 0) return 0;
00098   void *aligned = reinterpret_cast<void*>((reinterpret_cast<size_t>(original) & ~(size_t(15))) + 16);
00099   *(reinterpret_cast<void**>(aligned) - 1) = original;
00100   return aligned;
00101 }
00102 
00104 inline void handmade_aligned_free(void *ptr)
00105 {
00106   if (ptr) std::free(*(reinterpret_cast<void**>(ptr) - 1));
00107 }
00108 
00114 inline void* handmade_aligned_realloc(void* ptr, size_t size, size_t = 0)
00115 {
00116   if (ptr == 0) return handmade_aligned_malloc(size);
00117   void *original = *(reinterpret_cast<void**>(ptr) - 1);
00118   original = std::realloc(original,size+16);
00119   if (original == 0) return 0;
00120   void *aligned = reinterpret_cast<void*>((reinterpret_cast<size_t>(original) & ~(size_t(15))) + 16);
00121   *(reinterpret_cast<void**>(aligned) - 1) = original;
00122   return aligned;
00123 }
00124 
00125 /*****************************************************************************
00126 *** Implementation of generic aligned realloc (when no realloc can be used)***
00127 *****************************************************************************/
00128 
00129 void* aligned_malloc(size_t size);
00130 void  aligned_free(void *ptr);
00131 
00137 inline void* generic_aligned_realloc(void* ptr, size_t size, size_t old_size)
00138 {
00139   if (ptr==0)
00140     return aligned_malloc(size);
00141 
00142   if (size==0)
00143   {
00144     aligned_free(ptr);
00145     return 0;
00146   }
00147 
00148   void* newptr = aligned_malloc(size);
00149   if (newptr == 0)
00150   {
00151     #ifdef EIGEN_HAS_ERRNO
00152     errno = ENOMEM; // according to the standard
00153     #endif
00154     return 0;
00155   }
00156 
00157   if (ptr != 0)
00158   {
00159     std::memcpy(newptr, ptr, (std::min)(size,old_size));
00160     aligned_free(ptr);
00161   }
00162 
00163   return newptr;
00164 }
00165 
00166 /*****************************************************************************
00167 *** Implementation of portable aligned versions of malloc/free/realloc     ***
00168 *****************************************************************************/
00169 
00170 #ifdef EIGEN_NO_MALLOC
00171 inline void check_that_malloc_is_allowed()
00172 {
00173   eigen_assert(false && "heap allocation is forbidden (EIGEN_NO_MALLOC is defined)");
00174 }
00175 #elif defined EIGEN_RUNTIME_NO_MALLOC
00176 inline bool is_malloc_allowed_impl(bool update, bool new_value = false)
00177 {
00178   static bool value = true;
00179   if (update == 1)
00180     value = new_value;
00181   return value;
00182 }
00183 inline bool is_malloc_allowed() { return is_malloc_allowed_impl(false); }
00184 inline bool set_is_malloc_allowed(bool new_value) { return is_malloc_allowed_impl(true, new_value); }
00185 inline void check_that_malloc_is_allowed()
00186 {
00187   eigen_assert(is_malloc_allowed() && "heap allocation is forbidden (EIGEN_RUNTIME_NO_MALLOC is defined and g_is_malloc_allowed is false)");
00188 }
00189 #else 
00190 inline void check_that_malloc_is_allowed()
00191 {}
00192 #endif
00193 
00197 inline void* aligned_malloc(size_t size)
00198 {
00199   check_that_malloc_is_allowed();
00200 
00201   void *result;
00202   #if !EIGEN_ALIGN
00203     result = std::malloc(size);
00204   #elif EIGEN_MALLOC_ALREADY_ALIGNED
00205     result = std::malloc(size);
00206   #elif EIGEN_HAS_POSIX_MEMALIGN
00207     if(posix_memalign(&result, 16, size)) result = 0;
00208   #elif EIGEN_HAS_MM_MALLOC
00209     result = _mm_malloc(size, 16);
00210   #elif (defined _MSC_VER)
00211     result = _aligned_malloc(size, 16);
00212   #else
00213     result = handmade_aligned_malloc(size);
00214   #endif
00215 
00216   #ifdef EIGEN_EXCEPTIONS
00217     if(result == 0)
00218       throw std::bad_alloc();
00219   #endif
00220   return result;
00221 }
00222 
00224 inline void aligned_free(void *ptr)
00225 {
00226   #if !EIGEN_ALIGN
00227     std::free(ptr);
00228   #elif EIGEN_MALLOC_ALREADY_ALIGNED
00229     std::free(ptr);
00230   #elif EIGEN_HAS_POSIX_MEMALIGN
00231     std::free(ptr);
00232   #elif EIGEN_HAS_MM_MALLOC
00233     _mm_free(ptr);
00234   #elif defined(_MSC_VER)
00235     _aligned_free(ptr);
00236   #else
00237     handmade_aligned_free(ptr);
00238   #endif
00239 }
00240 
00246 inline void* aligned_realloc(void *ptr, size_t new_size, size_t old_size)
00247 {
00248   EIGEN_UNUSED_VARIABLE(old_size);
00249 
00250   void *result;
00251 #if !EIGEN_ALIGN
00252   result = std::realloc(ptr,new_size);
00253 #elif EIGEN_MALLOC_ALREADY_ALIGNED
00254   result = std::realloc(ptr,new_size);
00255 #elif EIGEN_HAS_POSIX_MEMALIGN
00256   result = generic_aligned_realloc(ptr,new_size,old_size);
00257 #elif EIGEN_HAS_MM_MALLOC
00258   // The defined(_mm_free) is just here to verify that this MSVC version
00259   // implements _mm_malloc/_mm_free based on the corresponding _aligned_
00260   // functions. This may not always be the case and we just try to be safe.
00261   #if defined(_MSC_VER) && defined(_mm_free)
00262     result = _aligned_realloc(ptr,new_size,16);
00263   #else
00264     result = generic_aligned_realloc(ptr,new_size,old_size);
00265   #endif
00266 #elif defined(_MSC_VER)
00267   result = _aligned_realloc(ptr,new_size,16);
00268 #else
00269   result = handmade_aligned_realloc(ptr,new_size,old_size);
00270 #endif
00271 
00272 #ifdef EIGEN_EXCEPTIONS
00273   if (result==0 && new_size!=0)
00274     throw std::bad_alloc();
00275 #endif
00276   return result;
00277 }
00278 
00279 /*****************************************************************************
00280 *** Implementation of conditionally aligned functions                      ***
00281 *****************************************************************************/
00282 
00286 template<bool Align> inline void* conditional_aligned_malloc(size_t size)
00287 {
00288   return aligned_malloc(size);
00289 }
00290 
00291 template<> inline void* conditional_aligned_malloc<false>(size_t size)
00292 {
00293   check_that_malloc_is_allowed();
00294 
00295   void *result = std::malloc(size);
00296   #ifdef EIGEN_EXCEPTIONS
00297     if(!result) throw std::bad_alloc();
00298   #endif
00299   return result;
00300 }
00301 
00303 template<bool Align> inline void conditional_aligned_free(void *ptr)
00304 {
00305   aligned_free(ptr);
00306 }
00307 
00308 template<> inline void conditional_aligned_free<false>(void *ptr)
00309 {
00310   std::free(ptr);
00311 }
00312 
00313 template<bool Align> inline void* conditional_aligned_realloc(void* ptr, size_t new_size, size_t old_size)
00314 {
00315   return aligned_realloc(ptr, new_size, old_size);
00316 }
00317 
00318 template<> inline void* conditional_aligned_realloc<false>(void* ptr, size_t new_size, size_t)
00319 {
00320   return std::realloc(ptr, new_size);
00321 }
00322 
00323 /*****************************************************************************
00324 *** Construction/destruction of array elements                             ***
00325 *****************************************************************************/
00326 
00330 template<typename T> inline T* construct_elements_of_array(T *ptr, size_t size)
00331 {
00332   for (size_t i=0; i < size; ++i) ::new (ptr + i) T;
00333   return ptr;
00334 }
00335 
00339 template<typename T> inline void destruct_elements_of_array(T *ptr, size_t size)
00340 {
00341   // always destruct an array starting from the end.
00342   if(ptr)
00343     while(size) ptr[--size].~T();
00344 }
00345 
00346 /*****************************************************************************
00347 *** Implementation of aligned new/delete-like functions                    ***
00348 *****************************************************************************/
00349 
00354 template<typename T> inline T* aligned_new(size_t size)
00355 {
00356   T *result = reinterpret_cast<T*>(aligned_malloc(sizeof(T)*size));
00357   return construct_elements_of_array(result, size);
00358 }
00359 
00360 template<typename T, bool Align> inline T* conditional_aligned_new(size_t size)
00361 {
00362   T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
00363   return construct_elements_of_array(result, size);
00364 }
00365 
00369 template<typename T> inline void aligned_delete(T *ptr, size_t size)
00370 {
00371   destruct_elements_of_array<T>(ptr, size);
00372   aligned_free(ptr);
00373 }
00374 
00378 template<typename T, bool Align> inline void conditional_aligned_delete(T *ptr, size_t size)
00379 {
00380   destruct_elements_of_array<T>(ptr, size);
00381   conditional_aligned_free<Align>(ptr);
00382 }
00383 
00384 template<typename T, bool Align> inline T* conditional_aligned_realloc_new(T* pts, size_t new_size, size_t old_size)
00385 {
00386   if(new_size < old_size)
00387     destruct_elements_of_array(pts+new_size, old_size-new_size);
00388   T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
00389   if(new_size > old_size)
00390     construct_elements_of_array(result+old_size, new_size-old_size);
00391   return result;
00392 }
00393 
00394 
00395 template<typename T, bool Align> inline T* conditional_aligned_new_auto(size_t size)
00396 {
00397   T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
00398   if(NumTraits<T>::RequireInitialization)
00399     construct_elements_of_array(result, size);
00400   return result;
00401 }
00402 
00403 template<typename T, bool Align> inline T* conditional_aligned_realloc_new_auto(T* pts, size_t new_size, size_t old_size)
00404 {
00405   if(NumTraits<T>::RequireInitialization && (new_size < old_size))
00406     destruct_elements_of_array(pts+new_size, old_size-new_size);
00407   T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
00408   if(NumTraits<T>::RequireInitialization && (new_size > old_size))
00409     construct_elements_of_array(result+old_size, new_size-old_size);
00410   return result;
00411 }
00412 
00413 template<typename T, bool Align> inline void conditional_aligned_delete_auto(T *ptr, size_t size)
00414 {
00415   if(NumTraits<T>::RequireInitialization)
00416     destruct_elements_of_array<T>(ptr, size);
00417   conditional_aligned_free<Align>(ptr);
00418 }
00419 
00420 /****************************************************************************/
00421 
00438 template<typename Scalar, typename Index>
00439 inline static Index first_aligned(const Scalar* array, Index size)
00440 {
00441   typedef typename packet_traits<Scalar>::type Packet;
00442   enum { PacketSize = packet_traits<Scalar>::size,
00443          PacketAlignedMask = PacketSize-1
00444   };
00445 
00446   if(PacketSize==1)
00447   {
00448     // Either there is no vectorization, or a packet consists of exactly 1 scalar so that all elements
00449     // of the array have the same alignment.
00450     return 0;
00451   }
00452   else if(size_t(array) & (sizeof(Scalar)-1))
00453   {
00454     // There is vectorization for this scalar type, but the array is not aligned to the size of a single scalar.
00455     // Consequently, no element of the array is well aligned.
00456     return size;
00457   }
00458   else
00459   {
00460     return std::min<Index>( (PacketSize - (Index((size_t(array)/sizeof(Scalar))) & PacketAlignedMask))
00461                            & PacketAlignedMask, size);
00462   }
00463 }
00464 
00465 } // end namespace internal
00466 
00467 /*****************************************************************************
00468 *** Implementation of runtime stack allocation (falling back to malloc)    ***
00469 *****************************************************************************/
00470 
00471 // you can overwrite Eigen's default behavior regarding alloca by defining EIGEN_ALLOCA
00472 // to the appropriate stack allocation function
00473 #ifndef EIGEN_ALLOCA
00474   #if (defined __linux__)
00475     #define EIGEN_ALLOCA alloca
00476   #elif defined(_MSC_VER)
00477     #define EIGEN_ALLOCA _alloca
00478   #endif
00479 #endif
00480 
00481 namespace internal {
00482 
00483 // This helper class construct the allocated memory, and takes care of destructing and freeing the handled data
00484 // at destruction time. In practice this helper class is mainly useful to avoid memory leak in case of exceptions.
00485 template<typename T> class aligned_stack_memory_handler
00486 {
00487   public:
00488     /* Creates a stack_memory_handler responsible for the buffer \a ptr of size \a size.
00489      * Note that \a ptr can be 0 regardless of the other parameters.
00490      * This constructor takes care of constructing/initializing the elements of the buffer if required by the scalar type T (see NumTraits<T>::RequireInitialization).
00491      * In this case, the buffer elements will also be destructed when this handler will be destructed.
00492      * Finally, if \a dealloc is true, then the pointer \a ptr is freed.
00493      **/
00494     aligned_stack_memory_handler(T* ptr, size_t size, bool dealloc)
00495       : m_ptr(ptr), m_size(size), m_deallocate(dealloc)
00496     {
00497       if(NumTraits<T>::RequireInitialization && m_ptr)
00498         Eigen::internal::construct_elements_of_array(m_ptr, size);
00499     }
00500     ~aligned_stack_memory_handler()
00501     {
00502       if(NumTraits<T>::RequireInitialization && m_ptr)
00503         Eigen::internal::destruct_elements_of_array<T>(m_ptr, m_size);
00504       if(m_deallocate)
00505         Eigen::internal::aligned_free(m_ptr);
00506     }
00507   protected:
00508     T* m_ptr;
00509     size_t m_size;
00510     bool m_deallocate;
00511 };
00512 
00513 }
00514 
00530 #ifdef EIGEN_ALLOCA
00531 
00532   #ifdef __arm__
00533     #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast<void*>((reinterpret_cast<size_t>(EIGEN_ALLOCA(SIZE+16)) & ~(size_t(15))) + 16)
00534   #else
00535     #define EIGEN_ALIGNED_ALLOCA EIGEN_ALLOCA
00536   #endif
00537 
00538   #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
00539     TYPE* NAME = (BUFFER)!=0 ? (BUFFER) \
00540                : reinterpret_cast<TYPE*>( \
00541                       (sizeof(TYPE)*SIZE<=EIGEN_STACK_ALLOCATION_LIMIT) ? EIGEN_ALIGNED_ALLOCA(sizeof(TYPE)*SIZE) \
00542                     : Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE) );  \
00543     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)
00544 
00545 #else
00546 
00547   #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
00548     TYPE* NAME = (BUFFER)!=0 ? BUFFER : reinterpret_cast<TYPE*>(Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE));    \
00549     Eigen::internal::aligned_stack_memory_handler<TYPE> EIGEN_CAT(NAME,_stack_memory_destructor)((BUFFER)==0 ? NAME : 0,SIZE,true)
00550     
00551 #endif
00552 
00553 
00554 /*****************************************************************************
00555 *** Implementation of EIGEN_MAKE_ALIGNED_OPERATOR_NEW [_IF]                ***
00556 *****************************************************************************/
00557 
00558 #if EIGEN_ALIGN
00559   #ifdef EIGEN_EXCEPTIONS
00560     #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
00561       void* operator new(size_t size, const std::nothrow_t&) throw() { \
00562         try { return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); } \
00563         catch (...) { return 0; } \
00564         return 0; \
00565       }
00566   #else
00567     #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
00568       void* operator new(size_t size, const std::nothrow_t&) throw() { \
00569         return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
00570       }
00571   #endif
00572 
00573   #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign) \
00574       void *operator new(size_t size) { \
00575         return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
00576       } \
00577       void *operator new[](size_t size) { \
00578         return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
00579       } \
00580       void operator delete(void * ptr) throw() { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
00581       void operator delete[](void * ptr) throw() { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
00582       /* in-place new and delete. since (at least afaik) there is no actual   */ \
00583       /* memory allocated we can safely let the default implementation handle */ \
00584       /* this particular case. */ \
00585       static void *operator new(size_t size, void *ptr) { return ::operator new(size,ptr); } \
00586       void operator delete(void * memory, void *ptr) throw() { return ::operator delete(memory,ptr); } \
00587       /* nothrow-new (returns zero instead of std::bad_alloc) */ \
00588       EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
00589       void operator delete(void *ptr, const std::nothrow_t&) throw() { \
00590         Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); \
00591       } \
00592       typedef void eigen_aligned_operator_new_marker_type;
00593 #else
00594   #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)
00595 #endif
00596 
00597 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(true)
00598 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size) \
00599   EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(((Size)!=Eigen::Dynamic) && ((sizeof(Scalar)*(Size))%16==0))
00600 
00601 /****************************************************************************/
00602 
00619 template<class T>
00620 class aligned_allocator
00621 {
00622 public:
00623     typedef size_t    size_type;
00624     typedef std::ptrdiff_t difference_type;
00625     typedef T*        pointer;
00626     typedef const T*  const_pointer;
00627     typedef T&        reference;
00628     typedef const T&  const_reference;
00629     typedef T         value_type;
00630 
00631     template<class U>
00632     struct rebind
00633     {
00634         typedef aligned_allocator<U> other;
00635     };
00636 
00637     pointer address( reference value ) const
00638     {
00639         return &value;
00640     }
00641 
00642     const_pointer address( const_reference value ) const
00643     {
00644         return &value;
00645     }
00646 
00647     aligned_allocator() throw()
00648     {
00649     }
00650 
00651     aligned_allocator( const aligned_allocator& ) throw()
00652     {
00653     }
00654 
00655     template<class U>
00656     aligned_allocator( const aligned_allocator<U>& ) throw()
00657     {
00658     }
00659 
00660     ~aligned_allocator() throw()
00661     {
00662     }
00663 
00664     size_type max_size() const throw()
00665     {
00666         return (std::numeric_limits<size_type>::max)();
00667     }
00668 
00669     pointer allocate( size_type num, const void* hint = 0 )
00670     {
00671         EIGEN_UNUSED_VARIABLE(hint);
00672         return static_cast<pointer>( internal::aligned_malloc( num * sizeof(T) ) );
00673     }
00674 
00675     void construct( pointer p, const T& value )
00676     {
00677         ::new( p ) T( value );
00678     }
00679 
00680     void destroy( pointer p )
00681     {
00682         p->~T();
00683     }
00684 
00685     void deallocate( pointer p, size_type /*num*/ )
00686     {
00687         internal::aligned_free( p );
00688     }
00689 
00690     bool operator!=(const aligned_allocator<T>& ) const
00691     { return false; }
00692 
00693     bool operator==(const aligned_allocator<T>& ) const
00694     { return true; }
00695 };
00696 
00697 //---------- Cache sizes ----------
00698 
00699 #if defined(__GNUC__) && ( defined(__i386__) || defined(__x86_64__) )
00700 #  if defined(__PIC__) && defined(__i386__)
00701      // Case for x86 with PIC
00702 #    define EIGEN_CPUID(abcd,func,id) \
00703        __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));
00704 #  else
00705      // Case for x86_64 or x86 w/o PIC
00706 #    define EIGEN_CPUID(abcd,func,id) \
00707        __asm__ __volatile__ ("cpuid": "=a" (abcd[0]), "=b" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "a" (func), "c" (id) );
00708 #  endif
00709 #elif defined(_MSC_VER)
00710 #  if (_MSC_VER > 1500)
00711 #    define EIGEN_CPUID(abcd,func,id) __cpuidex((int*)abcd,func,id)
00712 #  endif
00713 #endif
00714 
00715 namespace internal {
00716 
00717 #ifdef EIGEN_CPUID
00718 
00719 inline bool cpuid_is_vendor(int abcd[4], const char* vendor)
00720 {
00721   return abcd[1]==((int*)(vendor))[0] && abcd[3]==((int*)(vendor))[1] && abcd[2]==((int*)(vendor))[2];
00722 }
00723 
00724 inline void queryCacheSizes_intel_direct(int& l1, int& l2, int& l3)
00725 {
00726   int abcd[4];
00727   l1 = l2 = l3 = 0;
00728   int cache_id = 0;
00729   int cache_type = 0;
00730   do {
00731     abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
00732     EIGEN_CPUID(abcd,0x4,cache_id);
00733     cache_type  = (abcd[0] & 0x0F) >> 0;
00734     if(cache_type==1||cache_type==3) // data or unified cache
00735     {
00736       int cache_level = (abcd[0] & 0xE0) >> 5;  // A[7:5]
00737       int ways        = (abcd[1] & 0xFFC00000) >> 22; // B[31:22]
00738       int partitions  = (abcd[1] & 0x003FF000) >> 12; // B[21:12]
00739       int line_size   = (abcd[1] & 0x00000FFF) >>  0; // B[11:0]
00740       int sets        = (abcd[2]);                    // C[31:0]
00741 
00742       int cache_size = (ways+1) * (partitions+1) * (line_size+1) * (sets+1);
00743 
00744       switch(cache_level)
00745       {
00746         case 1: l1 = cache_size; break;
00747         case 2: l2 = cache_size; break;
00748         case 3: l3 = cache_size; break;
00749         default: break;
00750       }
00751     }
00752     cache_id++;
00753   } while(cache_type>0 && cache_id<16);
00754 }
00755 
00756 inline void queryCacheSizes_intel_codes(int& l1, int& l2, int& l3)
00757 {
00758   int abcd[4];
00759   abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
00760   l1 = l2 = l3 = 0;
00761   EIGEN_CPUID(abcd,0x00000002,0);
00762   unsigned char * bytes = reinterpret_cast<unsigned char *>(abcd)+2;
00763   bool check_for_p2_core2 = false;
00764   for(int i=0; i<14; ++i)
00765   {
00766     switch(bytes[i])
00767     {
00768       case 0x0A: l1 = 8; break;   // 0Ah   data L1 cache, 8 KB, 2 ways, 32 byte lines
00769       case 0x0C: l1 = 16; break;  // 0Ch   data L1 cache, 16 KB, 4 ways, 32 byte lines
00770       case 0x0E: l1 = 24; break;  // 0Eh   data L1 cache, 24 KB, 6 ways, 64 byte lines
00771       case 0x10: l1 = 16; break;  // 10h   data L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
00772       case 0x15: l1 = 16; break;  // 15h   code L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
00773       case 0x2C: l1 = 32; break;  // 2Ch   data L1 cache, 32 KB, 8 ways, 64 byte lines
00774       case 0x30: l1 = 32; break;  // 30h   code L1 cache, 32 KB, 8 ways, 64 byte lines
00775       case 0x60: l1 = 16; break;  // 60h   data L1 cache, 16 KB, 8 ways, 64 byte lines, sectored
00776       case 0x66: l1 = 8; break;   // 66h   data L1 cache, 8 KB, 4 ways, 64 byte lines, sectored
00777       case 0x67: l1 = 16; break;  // 67h   data L1 cache, 16 KB, 4 ways, 64 byte lines, sectored
00778       case 0x68: l1 = 32; break;  // 68h   data L1 cache, 32 KB, 4 ways, 64 byte lines, sectored
00779       case 0x1A: l2 = 96; break;   // code and data L2 cache, 96 KB, 6 ways, 64 byte lines (IA-64)
00780       case 0x22: l3 = 512; break;   // code and data L3 cache, 512 KB, 4 ways (!), 64 byte lines, dual-sectored
00781       case 0x23: l3 = 1024; break;   // code and data L3 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
00782       case 0x25: l3 = 2048; break;   // code and data L3 cache, 2048 KB, 8 ways, 64 byte lines, dual-sectored
00783       case 0x29: l3 = 4096; break;   // code and data L3 cache, 4096 KB, 8 ways, 64 byte lines, dual-sectored
00784       case 0x39: l2 = 128; break;   // code and data L2 cache, 128 KB, 4 ways, 64 byte lines, sectored
00785       case 0x3A: l2 = 192; break;   // code and data L2 cache, 192 KB, 6 ways, 64 byte lines, sectored
00786       case 0x3B: l2 = 128; break;   // code and data L2 cache, 128 KB, 2 ways, 64 byte lines, sectored
00787       case 0x3C: l2 = 256; break;   // code and data L2 cache, 256 KB, 4 ways, 64 byte lines, sectored
00788       case 0x3D: l2 = 384; break;   // code and data L2 cache, 384 KB, 6 ways, 64 byte lines, sectored
00789       case 0x3E: l2 = 512; break;   // code and data L2 cache, 512 KB, 4 ways, 64 byte lines, sectored
00790       case 0x40: l2 = 0; break;   // no integrated L2 cache (P6 core) or L3 cache (P4 core)
00791       case 0x41: l2 = 128; break;   // code and data L2 cache, 128 KB, 4 ways, 32 byte lines
00792       case 0x42: l2 = 256; break;   // code and data L2 cache, 256 KB, 4 ways, 32 byte lines
00793       case 0x43: l2 = 512; break;   // code and data L2 cache, 512 KB, 4 ways, 32 byte lines
00794       case 0x44: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 4 ways, 32 byte lines
00795       case 0x45: l2 = 2048; break;   // code and data L2 cache, 2048 KB, 4 ways, 32 byte lines
00796       case 0x46: l3 = 4096; break;   // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines
00797       case 0x47: l3 = 8192; break;   // code and data L3 cache, 8192 KB, 8 ways, 64 byte lines
00798       case 0x48: l2 = 3072; break;   // code and data L2 cache, 3072 KB, 12 ways, 64 byte lines
00799       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
00800       case 0x4A: l3 = 6144; break;   // code and data L3 cache, 6144 KB, 12 ways, 64 byte lines
00801       case 0x4B: l3 = 8192; break;   // code and data L3 cache, 8192 KB, 16 ways, 64 byte lines
00802       case 0x4C: l3 = 12288; break;   // code and data L3 cache, 12288 KB, 12 ways, 64 byte lines
00803       case 0x4D: l3 = 16384; break;   // code and data L3 cache, 16384 KB, 16 ways, 64 byte lines
00804       case 0x4E: l2 = 6144; break;   // code and data L2 cache, 6144 KB, 24 ways, 64 byte lines
00805       case 0x78: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 4 ways, 64 byte lines
00806       case 0x79: l2 = 128; break;   // code and data L2 cache, 128 KB, 8 ways, 64 byte lines, dual-sectored
00807       case 0x7A: l2 = 256; break;   // code and data L2 cache, 256 KB, 8 ways, 64 byte lines, dual-sectored
00808       case 0x7B: l2 = 512; break;   // code and data L2 cache, 512 KB, 8 ways, 64 byte lines, dual-sectored
00809       case 0x7C: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
00810       case 0x7D: l2 = 2048; break;   // code and data L2 cache, 2048 KB, 8 ways, 64 byte lines
00811       case 0x7E: l2 = 256; break;   // code and data L2 cache, 256 KB, 8 ways, 128 byte lines, sect. (IA-64)
00812       case 0x7F: l2 = 512; break;   // code and data L2 cache, 512 KB, 2 ways, 64 byte lines
00813       case 0x80: l2 = 512; break;   // code and data L2 cache, 512 KB, 8 ways, 64 byte lines
00814       case 0x81: l2 = 128; break;   // code and data L2 cache, 128 KB, 8 ways, 32 byte lines
00815       case 0x82: l2 = 256; break;   // code and data L2 cache, 256 KB, 8 ways, 32 byte lines
00816       case 0x83: l2 = 512; break;   // code and data L2 cache, 512 KB, 8 ways, 32 byte lines
00817       case 0x84: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 8 ways, 32 byte lines
00818       case 0x85: l2 = 2048; break;   // code and data L2 cache, 2048 KB, 8 ways, 32 byte lines
00819       case 0x86: l2 = 512; break;   // code and data L2 cache, 512 KB, 4 ways, 64 byte lines
00820       case 0x87: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines
00821       case 0x88: l3 = 2048; break;   // code and data L3 cache, 2048 KB, 4 ways, 64 byte lines (IA-64)
00822       case 0x89: l3 = 4096; break;   // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines (IA-64)
00823       case 0x8A: l3 = 8192; break;   // code and data L3 cache, 8192 KB, 4 ways, 64 byte lines (IA-64)
00824       case 0x8D: l3 = 3072; break;   // code and data L3 cache, 3072 KB, 12 ways, 128 byte lines (IA-64)
00825 
00826       default: break;
00827     }
00828   }
00829   if(check_for_p2_core2 && l2 == l3)
00830     l3 = 0;
00831   l1 *= 1024;
00832   l2 *= 1024;
00833   l3 *= 1024;
00834 }
00835 
00836 inline void queryCacheSizes_intel(int& l1, int& l2, int& l3, int max_std_funcs)
00837 {
00838   if(max_std_funcs>=4)
00839     queryCacheSizes_intel_direct(l1,l2,l3);
00840   else
00841     queryCacheSizes_intel_codes(l1,l2,l3);
00842 }
00843 
00844 inline void queryCacheSizes_amd(int& l1, int& l2, int& l3)
00845 {
00846   int abcd[4];
00847   abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
00848   EIGEN_CPUID(abcd,0x80000005,0);
00849   l1 = (abcd[2] >> 24) * 1024; // C[31:24] = L1 size in KB
00850   abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
00851   EIGEN_CPUID(abcd,0x80000006,0);
00852   l2 = (abcd[2] >> 16) * 1024; // C[31;16] = l2 cache size in KB
00853   l3 = ((abcd[3] & 0xFFFC000) >> 18) * 512 * 1024; // D[31;18] = l3 cache size in 512KB
00854 }
00855 #endif
00856 
00859 inline void queryCacheSizes(int& l1, int& l2, int& l3)
00860 {
00861   #ifdef EIGEN_CPUID
00862   int abcd[4];
00863 
00864   // identify the CPU vendor
00865   EIGEN_CPUID(abcd,0x0,0);
00866   int max_std_funcs = abcd[1];
00867   if(cpuid_is_vendor(abcd,"GenuineIntel"))
00868     queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
00869   else if(cpuid_is_vendor(abcd,"AuthenticAMD") || cpuid_is_vendor(abcd,"AMDisbetter!"))
00870     queryCacheSizes_amd(l1,l2,l3);
00871   else
00872     // by default let's use Intel's API
00873     queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
00874 
00875   // here is the list of other vendors:
00876 //   ||cpuid_is_vendor(abcd,"VIA VIA VIA ")
00877 //   ||cpuid_is_vendor(abcd,"CyrixInstead")
00878 //   ||cpuid_is_vendor(abcd,"CentaurHauls")
00879 //   ||cpuid_is_vendor(abcd,"GenuineTMx86")
00880 //   ||cpuid_is_vendor(abcd,"TransmetaCPU")
00881 //   ||cpuid_is_vendor(abcd,"RiseRiseRise")
00882 //   ||cpuid_is_vendor(abcd,"Geode by NSC")
00883 //   ||cpuid_is_vendor(abcd,"SiS SiS SiS ")
00884 //   ||cpuid_is_vendor(abcd,"UMC UMC UMC ")
00885 //   ||cpuid_is_vendor(abcd,"NexGenDriven")
00886   #else
00887   l1 = l2 = l3 = -1;
00888   #endif
00889 }
00890 
00893 inline int queryL1CacheSize()
00894 {
00895   int l1(-1), l2, l3;
00896   queryCacheSizes(l1,l2,l3);
00897   return l1;
00898 }
00899 
00902 inline int queryTopLevelCacheSize()
00903 {
00904   int l1, l2(-1), l3(-1);
00905   queryCacheSizes(l1,l2,l3);
00906   return (std::max)(l2,l3);
00907 }
00908 
00909 } // end namespace internal
00910 
00911 #endif // EIGEN_MEMORY_H


libicr
Author(s): Robert Krug
autogenerated on Mon Jan 6 2014 11:33:07