Memory.h

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 // This Source Code Form is subject to the terms of the Mozilla
00011 // Public License v. 2.0. If a copy of the MPL was not distributed
00012 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
00013 
00014 
00015 /*****************************************************************************
00016 *** Platform checks for aligned malloc functions                           ***
00017 *****************************************************************************/
00018 
00019 #ifndef EIGEN_MEMORY_H
00020 #define EIGEN_MEMORY_H
00021 
00022 #ifndef EIGEN_MALLOC_ALREADY_ALIGNED
00023 
00024 // Try to determine automatically if malloc is already aligned.
00025 
00026 // On 64-bit systems, glibc's malloc returns 16-byte-aligned pointers, see:
00027 //   http://www.gnu.org/s/libc/manual/html_node/Aligned-Memory-Blocks.html
00028 // This is true at least since glibc 2.8.
00029 // This leaves the question how to detect 64-bit. According to this document,
00030 //   http://gcc.fyxm.net/summit/2003/Porting%20to%2064%20bit.pdf
00031 // page 114, "[The] LP64 model [...] is used by all 64-bit UNIX ports" so it's indeed
00032 // quite safe, at least within the context of glibc, to equate 64-bit with LP64.
00033 #if defined(__GLIBC__) && ((__GLIBC__>=2 && __GLIBC_MINOR__ >= 8) || __GLIBC__>2) \
00034  && defined(__LP64__) && ! defined( __SANITIZE_ADDRESS__ )
00035   #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 1
00036 #else
00037   #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 0
00038 #endif
00039 
00040 // FreeBSD 6 seems to have 16-byte aligned malloc
00041 //   See http://svn.freebsd.org/viewvc/base/stable/6/lib/libc/stdlib/malloc.c?view=markup
00042 // FreeBSD 7 seems to have 16-byte aligned malloc except on ARM and MIPS architectures
00043 //   See http://svn.freebsd.org/viewvc/base/stable/7/lib/libc/stdlib/malloc.c?view=markup
00044 #if defined(__FreeBSD__) && !defined(__arm__) && !defined(__mips__)
00045   #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 1
00046 #else
00047   #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 0
00048 #endif
00049 
00050 #if defined(__APPLE__) \
00051  || defined(_WIN64) \
00052  || EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED \
00053  || EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED
00054   #define EIGEN_MALLOC_ALREADY_ALIGNED 1
00055 #else
00056   #define EIGEN_MALLOC_ALREADY_ALIGNED 0
00057 #endif
00058 
00059 #endif
00060 
00061 // See bug 554 (http://eigen.tuxfamily.org/bz/show_bug.cgi?id=554)
00062 // It seems to be unsafe to check _POSIX_ADVISORY_INFO without including unistd.h first.
00063 // Currently, let's include it only on unix systems:
00064 #if defined(__unix__) || defined(__unix)
00065   #include <unistd.h>
00066   #if ((defined __QNXNTO__) || (defined _GNU_SOURCE) || (defined __PGI) || ((defined _XOPEN_SOURCE) && (_XOPEN_SOURCE >= 600))) && (defined _POSIX_ADVISORY_INFO) && (_POSIX_ADVISORY_INFO > 0)
00067     #define EIGEN_HAS_POSIX_MEMALIGN 1
00068   #endif
00069 #endif
00070 
00071 #ifndef EIGEN_HAS_POSIX_MEMALIGN
00072   #define EIGEN_HAS_POSIX_MEMALIGN 0
00073 #endif
00074 
00075 #ifdef EIGEN_VECTORIZE_SSE
00076   #define EIGEN_HAS_MM_MALLOC 1
00077 #else
00078   #define EIGEN_HAS_MM_MALLOC 0
00079 #endif
00080 
00081 namespace Eigen {
00082 
00083 namespace internal {
00084 
00085 inline void throw_std_bad_alloc()
00086 {
00087   #ifdef EIGEN_EXCEPTIONS
00088     throw std::bad_alloc();
00089   #else
00090     std::size_t huge = -1;
00091     new int[huge];
00092   #endif
00093 }
00094 
00095 /*****************************************************************************
00096 *** Implementation of handmade aligned functions                           ***
00097 *****************************************************************************/
00098 
00099 /* ----- Hand made implementations of aligned malloc/free and realloc ----- */
00100 
00101 /** \internal Like malloc, but the returned pointer is guaranteed to be 16-byte aligned.
00102   * Fast, but wastes 16 additional bytes of memory. Does not throw any exception.
00103   */
00104 inline void* handmade_aligned_malloc(std::size_t size)
00105 {
00106   void *original = std::malloc(size+16);
00107   if (original == 0) return 0;
00108   void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(15))) + 16);
00109   *(reinterpret_cast<void**>(aligned) - 1) = original;
00110   return aligned;
00111 }
00112 
00113 /** \internal Frees memory allocated with handmade_aligned_malloc */
00114 inline void handmade_aligned_free(void *ptr)
00115 {
00116   if (ptr) std::free(*(reinterpret_cast<void**>(ptr) - 1));
00117 }
00118 
00119 /** \internal
00120   * \brief Reallocates aligned memory.
00121   * Since we know that our handmade version is based on std::realloc
00122   * we can use std::realloc to implement efficient reallocation.
00123   */
00124 inline void* handmade_aligned_realloc(void* ptr, std::size_t size, std::size_t = 0)
00125 {
00126   if (ptr == 0) return handmade_aligned_malloc(size);
00127   void *original = *(reinterpret_cast<void**>(ptr) - 1);
00128   std::ptrdiff_t previous_offset = static_cast<char *>(ptr)-static_cast<char *>(original);
00129   original = std::realloc(original,size+16);
00130   if (original == 0) return 0;
00131   void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(15))) + 16);
00132   void *previous_aligned = static_cast<char *>(original)+previous_offset;
00133   if(aligned!=previous_aligned)
00134     std::memmove(aligned, previous_aligned, size);
00135   
00136   *(reinterpret_cast<void**>(aligned) - 1) = original;
00137   return aligned;
00138 }
00139 
00140 /*****************************************************************************
00141 *** Implementation of generic aligned realloc (when no realloc can be used)***
00142 *****************************************************************************/
00143 
00144 void* aligned_malloc(std::size_t size);
00145 void  aligned_free(void *ptr);
00146 
00147 /** \internal
00148   * \brief Reallocates aligned memory.
00149   * Allows reallocation with aligned ptr types. This implementation will
00150   * always create a new memory chunk and copy the old data.
00151   */
00152 inline void* generic_aligned_realloc(void* ptr, size_t size, size_t old_size)
00153 {
00154   if (ptr==0)
00155     return aligned_malloc(size);
00156 
00157   if (size==0)
00158   {
00159     aligned_free(ptr);
00160     return 0;
00161   }
00162 
00163   void* newptr = aligned_malloc(size);
00164   if (newptr == 0)
00165   {
00166     #ifdef EIGEN_HAS_ERRNO
00167     errno = ENOMEM; // according to the standard
00168     #endif
00169     return 0;
00170   }
00171 
00172   if (ptr != 0)
00173   {
00174     std::memcpy(newptr, ptr, (std::min)(size,old_size));
00175     aligned_free(ptr);
00176   }
00177 
00178   return newptr;
00179 }
00180 
00181 /*****************************************************************************
00182 *** Implementation of portable aligned versions of malloc/free/realloc     ***
00183 *****************************************************************************/
00184 
00185 #ifdef EIGEN_NO_MALLOC
00186 inline void check_that_malloc_is_allowed()
00187 {
00188   eigen_assert(false && "heap allocation is forbidden (EIGEN_NO_MALLOC is defined)");
00189 }
00190 #elif defined EIGEN_RUNTIME_NO_MALLOC
00191 inline bool is_malloc_allowed_impl(bool update, bool new_value = false)
00192 {
00193   static bool value = true;
00194   if (update == 1)
00195     value = new_value;
00196   return value;
00197 }
00198 inline bool is_malloc_allowed() { return is_malloc_allowed_impl(false); }
00199 inline bool set_is_malloc_allowed(bool new_value) { return is_malloc_allowed_impl(true, new_value); }
00200 inline void check_that_malloc_is_allowed()
00201 {
00202   eigen_assert(is_malloc_allowed() && "heap allocation is forbidden (EIGEN_RUNTIME_NO_MALLOC is defined and g_is_malloc_allowed is false)");
00203 }
00204 #else 
00205 inline void check_that_malloc_is_allowed()
00206 {}
00207 #endif
00208 
00209 /** \internal Allocates \a size bytes. The returned pointer is guaranteed to have 16 bytes alignment.
00210   * On allocation error, the returned pointer is null, and std::bad_alloc is thrown.
00211   */
00212 inline void* aligned_malloc(size_t size)
00213 {
00214   check_that_malloc_is_allowed();
00215 
00216   void *result;
00217   #if !EIGEN_ALIGN
00218     result = std::malloc(size);
00219   #elif EIGEN_MALLOC_ALREADY_ALIGNED
00220     result = std::malloc(size);
00221   #elif EIGEN_HAS_POSIX_MEMALIGN
00222     if(posix_memalign(&result, 16, size)) result = 0;
00223   #elif EIGEN_HAS_MM_MALLOC
00224     result = _mm_malloc(size, 16);
00225   #elif defined(_MSC_VER) && (!defined(_WIN32_WCE))
00226     result = _aligned_malloc(size, 16);
00227   #else
00228     result = handmade_aligned_malloc(size);
00229   #endif
00230 
00231   if(!result && size)
00232     throw_std_bad_alloc();
00233 
00234   return result;
00235 }
00236 
00237 /** \internal Frees memory allocated with aligned_malloc. */
00238 inline void aligned_free(void *ptr)
00239 {
00240   #if !EIGEN_ALIGN
00241     std::free(ptr);
00242   #elif EIGEN_MALLOC_ALREADY_ALIGNED
00243     std::free(ptr);
00244   #elif EIGEN_HAS_POSIX_MEMALIGN
00245     std::free(ptr);
00246   #elif EIGEN_HAS_MM_MALLOC
00247     _mm_free(ptr);
00248   #elif defined(_MSC_VER) && (!defined(_WIN32_WCE))
00249     _aligned_free(ptr);
00250   #else
00251     handmade_aligned_free(ptr);
00252   #endif
00253 }
00254 
00255 /**
00256 * \internal
00257 * \brief Reallocates an aligned block of memory.
00258 * \throws std::bad_alloc on allocation failure
00259 **/
00260 inline void* aligned_realloc(void *ptr, size_t new_size, size_t old_size)
00261 {
00262   EIGEN_UNUSED_VARIABLE(old_size);
00263 
00264   void *result;
00265 #if !EIGEN_ALIGN
00266   result = std::realloc(ptr,new_size);
00267 #elif EIGEN_MALLOC_ALREADY_ALIGNED
00268   result = std::realloc(ptr,new_size);
00269 #elif EIGEN_HAS_POSIX_MEMALIGN
00270   result = generic_aligned_realloc(ptr,new_size,old_size);
00271 #elif EIGEN_HAS_MM_MALLOC
00272   // The defined(_mm_free) is just here to verify that this MSVC version
00273   // implements _mm_malloc/_mm_free based on the corresponding _aligned_
00274   // functions. This may not always be the case and we just try to be safe.
00275   #if defined(_MSC_VER) && (!defined(_WIN32_WCE)) && defined(_mm_free)
00276     result = _aligned_realloc(ptr,new_size,16);
00277   #else
00278     result = generic_aligned_realloc(ptr,new_size,old_size);
00279   #endif
00280 #elif defined(_MSC_VER) && (!defined(_WIN32_WCE))
00281   result = _aligned_realloc(ptr,new_size,16);
00282 #else
00283   result = handmade_aligned_realloc(ptr,new_size,old_size);
00284 #endif
00285 
00286   if (!result && new_size)
00287     throw_std_bad_alloc();
00288 
00289   return result;
00290 }
00291 
00292 /*****************************************************************************
00293 *** Implementation of conditionally aligned functions                      ***
00294 *****************************************************************************/
00295 
00296 /** \internal Allocates \a size bytes. If Align is true, then the returned ptr is 16-byte-aligned.
00297   * On allocation error, the returned pointer is null, and a std::bad_alloc is thrown.
00298   */
00299 template<bool Align> inline void* conditional_aligned_malloc(size_t size)
00300 {
00301   return aligned_malloc(size);
00302 }
00303 
00304 template<> inline void* conditional_aligned_malloc<false>(size_t size)
00305 {
00306   check_that_malloc_is_allowed();
00307 
00308   void *result = std::malloc(size);
00309   if(!result && size)
00310     throw_std_bad_alloc();
00311   return result;
00312 }
00313 
00314 /** \internal Frees memory allocated with conditional_aligned_malloc */
00315 template<bool Align> inline void conditional_aligned_free(void *ptr)
00316 {
00317   aligned_free(ptr);
00318 }
00319 
00320 template<> inline void conditional_aligned_free<false>(void *ptr)
00321 {
00322   std::free(ptr);
00323 }
00324 
00325 template<bool Align> inline void* conditional_aligned_realloc(void* ptr, size_t new_size, size_t old_size)
00326 {
00327   return aligned_realloc(ptr, new_size, old_size);
00328 }
00329 
00330 template<> inline void* conditional_aligned_realloc<false>(void* ptr, size_t new_size, size_t)
00331 {
00332   return std::realloc(ptr, new_size);
00333 }
00334 
00335 /*****************************************************************************
00336 *** Construction/destruction of array elements                             ***
00337 *****************************************************************************/
00338 
00339 /** \internal Constructs the elements of an array.
00340   * The \a size parameter tells on how many objects to call the constructor of T.
00341   */
00342 template<typename T> inline T* construct_elements_of_array(T *ptr, size_t size)
00343 {
00344   for (size_t i=0; i < size; ++i) ::new (ptr + i) T;
00345   return ptr;
00346 }
00347 
00348 /** \internal Destructs the elements of an array.
00349   * The \a size parameters tells on how many objects to call the destructor of T.
00350   */
00351 template<typename T> inline void destruct_elements_of_array(T *ptr, size_t size)
00352 {
00353   // always destruct an array starting from the end.
00354   if(ptr)
00355     while(size) ptr[--size].~T();
00356 }
00357 
00358 /*****************************************************************************
00359 *** Implementation of aligned new/delete-like functions                    ***
00360 *****************************************************************************/
00361 
00362 template<typename T>
00363 EIGEN_ALWAYS_INLINE void check_size_for_overflow(size_t size)
00364 {
00365   if(size > size_t(-1) / sizeof(T))
00366     throw_std_bad_alloc();
00367 }
00368 
00369 /** \internal Allocates \a size objects of type T. The returned pointer is guaranteed to have 16 bytes alignment.
00370   * On allocation error, the returned pointer is undefined, but a std::bad_alloc is thrown.
00371   * The default constructor of T is called.
00372   */
00373 template<typename T> inline T* aligned_new(size_t size)
00374 {
00375   check_size_for_overflow<T>(size);
00376   T *result = reinterpret_cast<T*>(aligned_malloc(sizeof(T)*size));
00377   return construct_elements_of_array(result, size);
00378 }
00379 
00380 template<typename T, bool Align> inline T* conditional_aligned_new(size_t size)
00381 {
00382   check_size_for_overflow<T>(size);
00383   T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
00384   return construct_elements_of_array(result, size);
00385 }
00386 
00387 /** \internal Deletes objects constructed with aligned_new
00388   * The \a size parameters tells on how many objects to call the destructor of T.
00389   */
00390 template<typename T> inline void aligned_delete(T *ptr, size_t size)
00391 {
00392   destruct_elements_of_array<T>(ptr, size);
00393   aligned_free(ptr);
00394 }
00395 
00396 /** \internal Deletes objects constructed with conditional_aligned_new
00397   * The \a size parameters tells on how many objects to call the destructor of T.
00398   */
00399 template<typename T, bool Align> inline void conditional_aligned_delete(T *ptr, size_t size)
00400 {
00401   destruct_elements_of_array<T>(ptr, size);
00402   conditional_aligned_free<Align>(ptr);
00403 }
00404 
00405 template<typename T, bool Align> inline T* conditional_aligned_realloc_new(T* pts, size_t new_size, size_t old_size)
00406 {
00407   check_size_for_overflow<T>(new_size);
00408   check_size_for_overflow<T>(old_size);
00409   if(new_size < old_size)
00410     destruct_elements_of_array(pts+new_size, old_size-new_size);
00411   T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
00412   if(new_size > old_size)
00413     construct_elements_of_array(result+old_size, new_size-old_size);
00414   return result;
00415 }
00416 
00417 
00418 template<typename T, bool Align> inline T* conditional_aligned_new_auto(size_t size)
00419 {
00420   if(size==0)
00421     return 0; // short-cut. Also fixes Bug 884
00422   check_size_for_overflow<T>(size);
00423   T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
00424   if(NumTraits<T>::RequireInitialization)
00425     construct_elements_of_array(result, size);
00426   return result;
00427 }
00428 
00429 template<typename T, bool Align> inline T* conditional_aligned_realloc_new_auto(T* pts, size_t new_size, size_t old_size)
00430 {
00431   check_size_for_overflow<T>(new_size);
00432   check_size_for_overflow<T>(old_size);
00433   if(NumTraits<T>::RequireInitialization && (new_size < old_size))
00434     destruct_elements_of_array(pts+new_size, old_size-new_size);
00435   T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
00436   if(NumTraits<T>::RequireInitialization && (new_size > old_size))
00437     construct_elements_of_array(result+old_size, new_size-old_size);
00438   return result;
00439 }
00440 
00441 template<typename T, bool Align> inline void conditional_aligned_delete_auto(T *ptr, size_t size)
00442 {
00443   if(NumTraits<T>::RequireInitialization)
00444     destruct_elements_of_array<T>(ptr, size);
00445   conditional_aligned_free<Align>(ptr);
00446 }
00447 
00448 /****************************************************************************/
00449 
00450 /** \internal Returns the index of the first element of the array that is well aligned for vectorization.
00451   *
00452   * \param array the address of the start of the array
00453   * \param size the size of the array
00454   *
00455   * \note If no element of the array is well aligned, the size of the array is returned. Typically,
00456   * for example with SSE, "well aligned" means 16-byte-aligned. If vectorization is disabled or if the
00457   * packet size for the given scalar type is 1, then everything is considered well-aligned.
00458   *
00459   * \note If the scalar type is vectorizable, we rely on the following assumptions: sizeof(Scalar) is a
00460   * power of 2, the packet size in bytes is also a power of 2, and is a multiple of sizeof(Scalar). On the
00461   * other hand, we do not assume that the array address is a multiple of sizeof(Scalar), as that fails for
00462   * example with Scalar=double on certain 32-bit platforms, see bug #79.
00463   *
00464   * There is also the variant first_aligned(const MatrixBase&) defined in DenseCoeffsBase.h.
00465   */
00466 template<typename Scalar, typename Index>
00467 static inline Index first_aligned(const Scalar* array, Index size)
00468 {
00469   static const Index PacketSize = packet_traits<Scalar>::size;
00470   static const Index PacketAlignedMask = PacketSize-1;
00471 
00472   if(PacketSize==1)
00473   {
00474     // Either there is no vectorization, or a packet consists of exactly 1 scalar so that all elements
00475     // of the array have the same alignment.
00476     return 0;
00477   }
00478   else if(size_t(array) & (sizeof(Scalar)-1))
00479   {
00480     // There is vectorization for this scalar type, but the array is not aligned to the size of a single scalar.
00481     // Consequently, no element of the array is well aligned.
00482     return size;
00483   }
00484   else
00485   {
00486     return std::min<Index>( (PacketSize - (Index((size_t(array)/sizeof(Scalar))) & PacketAlignedMask))
00487                            & PacketAlignedMask, size);
00488   }
00489 }
00490 
00491 /** \internal Returns the smallest integer multiple of \a base and greater or equal to \a size
00492   */ 
00493 template<typename Index> 
00494 inline static Index first_multiple(Index size, Index base)
00495 {
00496   return ((size+base-1)/base)*base;
00497 }
00498 
00499 // std::copy is much slower than memcpy, so let's introduce a smart_copy which
00500 // use memcpy on trivial types, i.e., on types that does not require an initialization ctor.
00501 template<typename T, bool UseMemcpy> struct smart_copy_helper;
00502 
00503 template<typename T> void smart_copy(const T* start, const T* end, T* target)
00504 {
00505   smart_copy_helper<T,!NumTraits<T>::RequireInitialization>::run(start, end, target);
00506 }
00507 
00508 template<typename T> struct smart_copy_helper<T,true> {
00509   static inline void run(const T* start, const T* end, T* target)
00510   { memcpy(target, start, std::ptrdiff_t(end)-std::ptrdiff_t(start)); }
00511 };
00512 
00513 template<typename T> struct smart_copy_helper<T,false> {
00514   static inline void run(const T* start, const T* end, T* target)
00515   { std::copy(start, end, target); }
00516 };
00517 
00518 
00519 /*****************************************************************************
00520 *** Implementation of runtime stack allocation (falling back to malloc)    ***
00521 *****************************************************************************/
00522 
00523 // you can overwrite Eigen's default behavior regarding alloca by defining EIGEN_ALLOCA
00524 // to the appropriate stack allocation function
00525 #ifndef EIGEN_ALLOCA
00526   #if (defined __linux__) || (defined __APPLE__) || (defined alloca)
00527     #define EIGEN_ALLOCA alloca
00528   #elif defined(_MSC_VER)
00529     #define EIGEN_ALLOCA _alloca
00530   #endif
00531 #endif
00532 
00533 // This helper class construct the allocated memory, and takes care of destructing and freeing the handled data
00534 // at destruction time. In practice this helper class is mainly useful to avoid memory leak in case of exceptions.
00535 template<typename T> class aligned_stack_memory_handler
00536 {
00537   public:
00538     /* Creates a stack_memory_handler responsible for the buffer \a ptr of size \a size.
00539      * Note that \a ptr can be 0 regardless of the other parameters.
00540      * This constructor takes care of constructing/initializing the elements of the buffer if required by the scalar type T (see NumTraits<T>::RequireInitialization).
00541      * In this case, the buffer elements will also be destructed when this handler will be destructed.
00542      * Finally, if \a dealloc is true, then the pointer \a ptr is freed.
00543      **/
00544     aligned_stack_memory_handler(T* ptr, size_t size, bool dealloc)
00545       : m_ptr(ptr), m_size(size), m_deallocate(dealloc)
00546     {
00547       if(NumTraits<T>::RequireInitialization && m_ptr)
00548         Eigen::internal::construct_elements_of_array(m_ptr, size);
00549     }
00550     ~aligned_stack_memory_handler()
00551     {
00552       if(NumTraits<T>::RequireInitialization && m_ptr)
00553         Eigen::internal::destruct_elements_of_array<T>(m_ptr, m_size);
00554       if(m_deallocate)
00555         Eigen::internal::aligned_free(m_ptr);
00556     }
00557   protected:
00558     T* m_ptr;
00559     size_t m_size;
00560     bool m_deallocate;
00561 };
00562 
00563 } // end namespace internal
00564 
00565 /** \internal
00566   * Declares, allocates and construct an aligned buffer named NAME of SIZE elements of type TYPE on the stack
00567   * if SIZE is smaller than EIGEN_STACK_ALLOCATION_LIMIT, and if stack allocation is supported by the platform
00568   * (currently, this is Linux and Visual Studio only). Otherwise the memory is allocated on the heap.
00569   * The allocated buffer is automatically deleted when exiting the scope of this declaration.
00570   * If BUFFER is non null, then the declared variable is simply an alias for BUFFER, and no allocation/deletion occurs.
00571   * Here is an example:
00572   * \code
00573   * {
00574   *   ei_declare_aligned_stack_constructed_variable(float,data,size,0);
00575   *   // use data[0] to data[size-1]
00576   * }
00577   * \endcode
00578   * The underlying stack allocation function can controlled with the EIGEN_ALLOCA preprocessor token.
00579   */
00580 #ifdef EIGEN_ALLOCA
00581 
00582   #if defined(__arm__) || defined(_WIN32)
00583     #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast<void*>((reinterpret_cast<size_t>(EIGEN_ALLOCA(SIZE+16)) & ~(size_t(15))) + 16)
00584   #else
00585     #define EIGEN_ALIGNED_ALLOCA EIGEN_ALLOCA
00586   #endif
00587 
00588   #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
00589     Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \
00590     TYPE* NAME = (BUFFER)!=0 ? (BUFFER) \
00591                : reinterpret_cast<TYPE*>( \
00592                       (sizeof(TYPE)*SIZE<=EIGEN_STACK_ALLOCATION_LIMIT) ? EIGEN_ALIGNED_ALLOCA(sizeof(TYPE)*SIZE) \
00593                     : Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE) );  \
00594     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)
00595 
00596 #else
00597 
00598   #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
00599     Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \
00600     TYPE* NAME = (BUFFER)!=0 ? BUFFER : reinterpret_cast<TYPE*>(Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE));    \
00601     Eigen::internal::aligned_stack_memory_handler<TYPE> EIGEN_CAT(NAME,_stack_memory_destructor)((BUFFER)==0 ? NAME : 0,SIZE,true)
00602     
00603 #endif
00604 
00605 
00606 /*****************************************************************************
00607 *** Implementation of EIGEN_MAKE_ALIGNED_OPERATOR_NEW [_IF]                ***
00608 *****************************************************************************/
00609 
00610 #if EIGEN_ALIGN
00611   #ifdef EIGEN_EXCEPTIONS
00612     #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
00613       void* operator new(size_t size, const std::nothrow_t&) throw() { \
00614         try { return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); } \
00615         catch (...) { return 0; } \
00616       }
00617   #else
00618     #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
00619       void* operator new(size_t size, const std::nothrow_t&) throw() { \
00620         return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
00621       }
00622   #endif
00623 
00624   #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign) \
00625       void *operator new(size_t size) { \
00626         return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
00627       } \
00628       void *operator new[](size_t size) { \
00629         return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
00630       } \
00631       void operator delete(void * ptr) throw() { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
00632       void operator delete[](void * ptr) throw() { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
00633       void operator delete(void * ptr, std::size_t /* sz */) throw() { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
00634       void operator delete[](void * ptr, std::size_t /* sz */) throw() { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
00635       /* in-place new and delete. since (at least afaik) there is no actual   */ \
00636       /* memory allocated we can safely let the default implementation handle */ \
00637       /* this particular case. */ \
00638       static void *operator new(size_t size, void *ptr) { return ::operator new(size,ptr); } \
00639       static void *operator new[](size_t size, void* ptr) { return ::operator new[](size,ptr); } \
00640       void operator delete(void * memory, void *ptr) throw() { return ::operator delete(memory,ptr); } \
00641       void operator delete[](void * memory, void *ptr) throw() { return ::operator delete[](memory,ptr); } \
00642       /* nothrow-new (returns zero instead of std::bad_alloc) */ \
00643       EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
00644       void operator delete(void *ptr, const std::nothrow_t&) throw() { \
00645         Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); \
00646       } \
00647       typedef void eigen_aligned_operator_new_marker_type;
00648 #else
00649   #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)
00650 #endif
00651 
00652 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(true)
00653 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size) \
00654   EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool(((Size)!=Eigen::Dynamic) && ((sizeof(Scalar)*(Size))%16==0)))
00655 
00656 /****************************************************************************/
00657 
00658 /** \class aligned_allocator
00659 * \ingroup Core_Module
00660 *
00661 * \brief STL compatible allocator to use with with 16 byte aligned types
00662 *
00663 * Example:
00664 * \code
00665 * // Matrix4f requires 16 bytes alignment:
00666 * std::map< int, Matrix4f, std::less<int>, 
00667 *           aligned_allocator<std::pair<const int, Matrix4f> > > my_map_mat4;
00668 * // Vector3f does not require 16 bytes alignment, no need to use Eigen's allocator:
00669 * std::map< int, Vector3f > my_map_vec3;
00670 * \endcode
00671 *
00672 * \sa \ref TopicStlContainers.
00673 */
00674 template<class T>
00675 class aligned_allocator
00676 {
00677 public:
00678     typedef size_t    size_type;
00679     typedef std::ptrdiff_t difference_type;
00680     typedef T*        pointer;
00681     typedef const T*  const_pointer;
00682     typedef T&        reference;
00683     typedef const T&  const_reference;
00684     typedef T         value_type;
00685 
00686     template<class U>
00687     struct rebind
00688     {
00689         typedef aligned_allocator<U> other;
00690     };
00691 
00692     pointer address( reference value ) const
00693     {
00694         return &value;
00695     }
00696 
00697     const_pointer address( const_reference value ) const
00698     {
00699         return &value;
00700     }
00701 
00702     aligned_allocator()
00703     {
00704     }
00705 
00706     aligned_allocator( const aligned_allocator& )
00707     {
00708     }
00709 
00710     template<class U>
00711     aligned_allocator( const aligned_allocator<U>& )
00712     {
00713     }
00714 
00715     ~aligned_allocator()
00716     {
00717     }
00718 
00719     size_type max_size() const
00720     {
00721         return (std::numeric_limits<size_type>::max)();
00722     }
00723 
00724     pointer allocate( size_type num, const void* hint = 0 )
00725     {
00726         EIGEN_UNUSED_VARIABLE(hint);
00727         internal::check_size_for_overflow<T>(num);
00728         return static_cast<pointer>( internal::aligned_malloc( num * sizeof(T) ) );
00729     }
00730 
00731     void construct( pointer p, const T& value )
00732     {
00733         ::new( p ) T( value );
00734     }
00735 
00736     void destroy( pointer p )
00737     {
00738         p->~T();
00739     }
00740 
00741     void deallocate( pointer p, size_type /*num*/ )
00742     {
00743         internal::aligned_free( p );
00744     }
00745 
00746     bool operator!=(const aligned_allocator<T>& ) const
00747     { return false; }
00748 
00749     bool operator==(const aligned_allocator<T>& ) const
00750     { return true; }
00751 };
00752 
00753 //---------- Cache sizes ----------
00754 
00755 #if !defined(EIGEN_NO_CPUID)
00756 #  if defined(__GNUC__) && ( defined(__i386__) || defined(__x86_64__) )
00757 #    if defined(__PIC__) && defined(__i386__)
00758        // Case for x86 with PIC
00759 #      define EIGEN_CPUID(abcd,func,id) \
00760          __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));
00761 #    elif defined(__PIC__) && defined(__x86_64__)
00762        // 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.
00763        // However, we cannot detect which code model is used, and the xchg overhead is negligible anyway.
00764 #      define EIGEN_CPUID(abcd,func,id) \
00765         __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));
00766 #    else
00767        // Case for x86_64 or x86 w/o PIC
00768 #      define EIGEN_CPUID(abcd,func,id) \
00769          __asm__ __volatile__ ("cpuid": "=a" (abcd[0]), "=b" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "0" (func), "2" (id) );
00770 #    endif
00771 #  elif defined(_MSC_VER)
00772 #    if (_MSC_VER > 1500) && ( defined(_M_IX86) || defined(_M_X64) )
00773 #      define EIGEN_CPUID(abcd,func,id) __cpuidex((int*)abcd,func,id)
00774 #    endif
00775 #  endif
00776 #endif
00777 
00778 namespace internal {
00779 
00780 #ifdef EIGEN_CPUID
00781 
00782 inline bool cpuid_is_vendor(int abcd[4], const int vendor[3])
00783 {
00784   return abcd[1]==vendor[0] && abcd[3]==vendor[1] && abcd[2]==vendor[2];
00785 }
00786 
00787 inline void queryCacheSizes_intel_direct(int& l1, int& l2, int& l3)
00788 {
00789   int abcd[4];
00790   l1 = l2 = l3 = 0;
00791   int cache_id = 0;
00792   int cache_type = 0;
00793   do {
00794     abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
00795     EIGEN_CPUID(abcd,0x4,cache_id);
00796     cache_type  = (abcd[0] & 0x0F) >> 0;
00797     if(cache_type==1||cache_type==3) // data or unified cache
00798     {
00799       int cache_level = (abcd[0] & 0xE0) >> 5;  // A[7:5]
00800       int ways        = (abcd[1] & 0xFFC00000) >> 22; // B[31:22]
00801       int partitions  = (abcd[1] & 0x003FF000) >> 12; // B[21:12]
00802       int line_size   = (abcd[1] & 0x00000FFF) >>  0; // B[11:0]
00803       int sets        = (abcd[2]);                    // C[31:0]
00804 
00805       int cache_size = (ways+1) * (partitions+1) * (line_size+1) * (sets+1);
00806 
00807       switch(cache_level)
00808       {
00809         case 1: l1 = cache_size; break;
00810         case 2: l2 = cache_size; break;
00811         case 3: l3 = cache_size; break;
00812         default: break;
00813       }
00814     }
00815     cache_id++;
00816   } while(cache_type>0 && cache_id<16);
00817 }
00818 
00819 inline void queryCacheSizes_intel_codes(int& l1, int& l2, int& l3)
00820 {
00821   int abcd[4];
00822   abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
00823   l1 = l2 = l3 = 0;
00824   EIGEN_CPUID(abcd,0x00000002,0);
00825   unsigned char * bytes = reinterpret_cast<unsigned char *>(abcd)+2;
00826   bool check_for_p2_core2 = false;
00827   for(int i=0; i<14; ++i)
00828   {
00829     switch(bytes[i])
00830     {
00831       case 0x0A: l1 = 8; break;   // 0Ah   data L1 cache, 8 KB, 2 ways, 32 byte lines
00832       case 0x0C: l1 = 16; break;  // 0Ch   data L1 cache, 16 KB, 4 ways, 32 byte lines
00833       case 0x0E: l1 = 24; break;  // 0Eh   data L1 cache, 24 KB, 6 ways, 64 byte lines
00834       case 0x10: l1 = 16; break;  // 10h   data L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
00835       case 0x15: l1 = 16; break;  // 15h   code L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
00836       case 0x2C: l1 = 32; break;  // 2Ch   data L1 cache, 32 KB, 8 ways, 64 byte lines
00837       case 0x30: l1 = 32; break;  // 30h   code L1 cache, 32 KB, 8 ways, 64 byte lines
00838       case 0x60: l1 = 16; break;  // 60h   data L1 cache, 16 KB, 8 ways, 64 byte lines, sectored
00839       case 0x66: l1 = 8; break;   // 66h   data L1 cache, 8 KB, 4 ways, 64 byte lines, sectored
00840       case 0x67: l1 = 16; break;  // 67h   data L1 cache, 16 KB, 4 ways, 64 byte lines, sectored
00841       case 0x68: l1 = 32; break;  // 68h   data L1 cache, 32 KB, 4 ways, 64 byte lines, sectored
00842       case 0x1A: l2 = 96; break;   // code and data L2 cache, 96 KB, 6 ways, 64 byte lines (IA-64)
00843       case 0x22: l3 = 512; break;   // code and data L3 cache, 512 KB, 4 ways (!), 64 byte lines, dual-sectored
00844       case 0x23: l3 = 1024; break;   // code and data L3 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
00845       case 0x25: l3 = 2048; break;   // code and data L3 cache, 2048 KB, 8 ways, 64 byte lines, dual-sectored
00846       case 0x29: l3 = 4096; break;   // code and data L3 cache, 4096 KB, 8 ways, 64 byte lines, dual-sectored
00847       case 0x39: l2 = 128; break;   // code and data L2 cache, 128 KB, 4 ways, 64 byte lines, sectored
00848       case 0x3A: l2 = 192; break;   // code and data L2 cache, 192 KB, 6 ways, 64 byte lines, sectored
00849       case 0x3B: l2 = 128; break;   // code and data L2 cache, 128 KB, 2 ways, 64 byte lines, sectored
00850       case 0x3C: l2 = 256; break;   // code and data L2 cache, 256 KB, 4 ways, 64 byte lines, sectored
00851       case 0x3D: l2 = 384; break;   // code and data L2 cache, 384 KB, 6 ways, 64 byte lines, sectored
00852       case 0x3E: l2 = 512; break;   // code and data L2 cache, 512 KB, 4 ways, 64 byte lines, sectored
00853       case 0x40: l2 = 0; break;   // no integrated L2 cache (P6 core) or L3 cache (P4 core)
00854       case 0x41: l2 = 128; break;   // code and data L2 cache, 128 KB, 4 ways, 32 byte lines
00855       case 0x42: l2 = 256; break;   // code and data L2 cache, 256 KB, 4 ways, 32 byte lines
00856       case 0x43: l2 = 512; break;   // code and data L2 cache, 512 KB, 4 ways, 32 byte lines
00857       case 0x44: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 4 ways, 32 byte lines
00858       case 0x45: l2 = 2048; break;   // code and data L2 cache, 2048 KB, 4 ways, 32 byte lines
00859       case 0x46: l3 = 4096; break;   // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines
00860       case 0x47: l3 = 8192; break;   // code and data L3 cache, 8192 KB, 8 ways, 64 byte lines
00861       case 0x48: l2 = 3072; break;   // code and data L2 cache, 3072 KB, 12 ways, 64 byte lines
00862       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
00863       case 0x4A: l3 = 6144; break;   // code and data L3 cache, 6144 KB, 12 ways, 64 byte lines
00864       case 0x4B: l3 = 8192; break;   // code and data L3 cache, 8192 KB, 16 ways, 64 byte lines
00865       case 0x4C: l3 = 12288; break;   // code and data L3 cache, 12288 KB, 12 ways, 64 byte lines
00866       case 0x4D: l3 = 16384; break;   // code and data L3 cache, 16384 KB, 16 ways, 64 byte lines
00867       case 0x4E: l2 = 6144; break;   // code and data L2 cache, 6144 KB, 24 ways, 64 byte lines
00868       case 0x78: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 4 ways, 64 byte lines
00869       case 0x79: l2 = 128; break;   // code and data L2 cache, 128 KB, 8 ways, 64 byte lines, dual-sectored
00870       case 0x7A: l2 = 256; break;   // code and data L2 cache, 256 KB, 8 ways, 64 byte lines, dual-sectored
00871       case 0x7B: l2 = 512; break;   // code and data L2 cache, 512 KB, 8 ways, 64 byte lines, dual-sectored
00872       case 0x7C: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
00873       case 0x7D: l2 = 2048; break;   // code and data L2 cache, 2048 KB, 8 ways, 64 byte lines
00874       case 0x7E: l2 = 256; break;   // code and data L2 cache, 256 KB, 8 ways, 128 byte lines, sect. (IA-64)
00875       case 0x7F: l2 = 512; break;   // code and data L2 cache, 512 KB, 2 ways, 64 byte lines
00876       case 0x80: l2 = 512; break;   // code and data L2 cache, 512 KB, 8 ways, 64 byte lines
00877       case 0x81: l2 = 128; break;   // code and data L2 cache, 128 KB, 8 ways, 32 byte lines
00878       case 0x82: l2 = 256; break;   // code and data L2 cache, 256 KB, 8 ways, 32 byte lines
00879       case 0x83: l2 = 512; break;   // code and data L2 cache, 512 KB, 8 ways, 32 byte lines
00880       case 0x84: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 8 ways, 32 byte lines
00881       case 0x85: l2 = 2048; break;   // code and data L2 cache, 2048 KB, 8 ways, 32 byte lines
00882       case 0x86: l2 = 512; break;   // code and data L2 cache, 512 KB, 4 ways, 64 byte lines
00883       case 0x87: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines
00884       case 0x88: l3 = 2048; break;   // code and data L3 cache, 2048 KB, 4 ways, 64 byte lines (IA-64)
00885       case 0x89: l3 = 4096; break;   // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines (IA-64)
00886       case 0x8A: l3 = 8192; break;   // code and data L3 cache, 8192 KB, 4 ways, 64 byte lines (IA-64)
00887       case 0x8D: l3 = 3072; break;   // code and data L3 cache, 3072 KB, 12 ways, 128 byte lines (IA-64)
00888 
00889       default: break;
00890     }
00891   }
00892   if(check_for_p2_core2 && l2 == l3)
00893     l3 = 0;
00894   l1 *= 1024;
00895   l2 *= 1024;
00896   l3 *= 1024;
00897 }
00898 
00899 inline void queryCacheSizes_intel(int& l1, int& l2, int& l3, int max_std_funcs)
00900 {
00901   if(max_std_funcs>=4)
00902     queryCacheSizes_intel_direct(l1,l2,l3);
00903   else
00904     queryCacheSizes_intel_codes(l1,l2,l3);
00905 }
00906 
00907 inline void queryCacheSizes_amd(int& l1, int& l2, int& l3)
00908 {
00909   int abcd[4];
00910   abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
00911   EIGEN_CPUID(abcd,0x80000005,0);
00912   l1 = (abcd[2] >> 24) * 1024; // C[31:24] = L1 size in KB
00913   abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
00914   EIGEN_CPUID(abcd,0x80000006,0);
00915   l2 = (abcd[2] >> 16) * 1024; // C[31;16] = l2 cache size in KB
00916   l3 = ((abcd[3] & 0xFFFC000) >> 18) * 512 * 1024; // D[31;18] = l3 cache size in 512KB
00917 }
00918 #endif
00919 
00920 /** \internal
00921  * Queries and returns the cache sizes in Bytes of the L1, L2, and L3 data caches respectively */
00922 inline void queryCacheSizes(int& l1, int& l2, int& l3)
00923 {
00924   #ifdef EIGEN_CPUID
00925   int abcd[4];
00926   const int GenuineIntel[] = {0x756e6547, 0x49656e69, 0x6c65746e};
00927   const int AuthenticAMD[] = {0x68747541, 0x69746e65, 0x444d4163};
00928   const int AMDisbetter_[] = {0x69444d41, 0x74656273, 0x21726574}; // "AMDisbetter!"
00929 
00930   // identify the CPU vendor
00931   EIGEN_CPUID(abcd,0x0,0);
00932   int max_std_funcs = abcd[1];
00933   if(cpuid_is_vendor(abcd,GenuineIntel))
00934     queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
00935   else if(cpuid_is_vendor(abcd,AuthenticAMD) || cpuid_is_vendor(abcd,AMDisbetter_))
00936     queryCacheSizes_amd(l1,l2,l3);
00937   else
00938     // by default let's use Intel's API
00939     queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
00940 
00941   // here is the list of other vendors:
00942 //   ||cpuid_is_vendor(abcd,"VIA VIA VIA ")
00943 //   ||cpuid_is_vendor(abcd,"CyrixInstead")
00944 //   ||cpuid_is_vendor(abcd,"CentaurHauls")
00945 //   ||cpuid_is_vendor(abcd,"GenuineTMx86")
00946 //   ||cpuid_is_vendor(abcd,"TransmetaCPU")
00947 //   ||cpuid_is_vendor(abcd,"RiseRiseRise")
00948 //   ||cpuid_is_vendor(abcd,"Geode by NSC")
00949 //   ||cpuid_is_vendor(abcd,"SiS SiS SiS ")
00950 //   ||cpuid_is_vendor(abcd,"UMC UMC UMC ")
00951 //   ||cpuid_is_vendor(abcd,"NexGenDriven")
00952   #else
00953   l1 = l2 = l3 = -1;
00954   #endif
00955 }
00956 
00957 /** \internal
00958  * \returns the size in Bytes of the L1 data cache */
00959 inline int queryL1CacheSize()
00960 {
00961   int l1(-1), l2, l3;
00962   queryCacheSizes(l1,l2,l3);
00963   return l1;
00964 }
00965 
00966 /** \internal
00967  * \returns the size in Bytes of the L2 or L3 cache if this later is present */
00968 inline int queryTopLevelCacheSize()
00969 {
00970   int l1, l2(-1), l3(-1);
00971   queryCacheSizes(l1,l2,l3);
00972   return (std::max)(l2,l3);
00973 }
00974 
00975 } // end namespace internal
00976 
00977 } // end namespace Eigen
00978 
00979 #endif // EIGEN_MEMORY_H