cart-elc

Source code for CART-ELC
git clone git://git.laack.co/cart-elc.git
Log | Files | Refs | README | LICENSE

Memory.h (46661B)


      1 // This file is part of Eigen, a lightweight C++ template library
      2 // for linear algebra.
      3 //
      4 // Copyright (C) 2008-2015 Gael Guennebaud <gael.guennebaud@inria.fr>
      5 // Copyright (C) 2008-2009 Benoit Jacob <jacob.benoit.1@gmail.com>
      6 // Copyright (C) 2009 Kenneth Riddile <kfriddile@yahoo.com>
      7 // Copyright (C) 2010 Hauke Heibel <hauke.heibel@gmail.com>
      8 // Copyright (C) 2010 Thomas Capricelli <orzel@freehackers.org>
      9 // Copyright (C) 2013 Pavel Holoborodko <pavel@holoborodko.com>
     10 //
     11 // This Source Code Form is subject to the terms of the Mozilla
     12 // Public License v. 2.0. If a copy of the MPL was not distributed
     13 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
     14 
     15 
     16 /*****************************************************************************
     17 *** Platform checks for aligned malloc functions                           ***
     18 *****************************************************************************/
     19 
     20 #ifndef EIGEN_MEMORY_H
     21 #define EIGEN_MEMORY_H
     22 
     23 #ifndef EIGEN_MALLOC_ALREADY_ALIGNED
     24 
     25 // Try to determine automatically if malloc is already aligned.
     26 
     27 // On 64-bit systems, glibc's malloc returns 16-byte-aligned pointers, see:
     28 //   http://www.gnu.org/s/libc/manual/html_node/Aligned-Memory-Blocks.html
     29 // This is true at least since glibc 2.8.
     30 // This leaves the question how to detect 64-bit. According to this document,
     31 //   http://gcc.fyxm.net/summit/2003/Porting%20to%2064%20bit.pdf
     32 // page 114, "[The] LP64 model [...] is used by all 64-bit UNIX ports" so it's indeed
     33 // quite safe, at least within the context of glibc, to equate 64-bit with LP64.
     34 #if defined(__GLIBC__) && ((__GLIBC__>=2 && __GLIBC_MINOR__ >= 8) || __GLIBC__>2) \
     35  && defined(__LP64__) && ! defined( __SANITIZE_ADDRESS__ ) && (EIGEN_DEFAULT_ALIGN_BYTES == 16)
     36   #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 1
     37 #else
     38   #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 0
     39 #endif
     40 
     41 // FreeBSD 6 seems to have 16-byte aligned malloc
     42 //   See http://svn.freebsd.org/viewvc/base/stable/6/lib/libc/stdlib/malloc.c?view=markup
     43 // FreeBSD 7 seems to have 16-byte aligned malloc except on ARM and MIPS architectures
     44 //   See http://svn.freebsd.org/viewvc/base/stable/7/lib/libc/stdlib/malloc.c?view=markup
     45 #if defined(__FreeBSD__) && !(EIGEN_ARCH_ARM || EIGEN_ARCH_MIPS) && (EIGEN_DEFAULT_ALIGN_BYTES == 16)
     46   #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 1
     47 #else
     48   #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 0
     49 #endif
     50 
     51 #if (EIGEN_OS_MAC && (EIGEN_DEFAULT_ALIGN_BYTES == 16))     \
     52  || (EIGEN_OS_WIN64 && (EIGEN_DEFAULT_ALIGN_BYTES == 16))   \
     53  || EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED              \
     54  || EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED
     55   #define EIGEN_MALLOC_ALREADY_ALIGNED 1
     56 #else
     57   #define EIGEN_MALLOC_ALREADY_ALIGNED 0
     58 #endif
     59 
     60 #endif
     61 
     62 namespace Eigen {
     63 
     64 namespace internal {
     65 
     66 EIGEN_DEVICE_FUNC
     67 inline void throw_std_bad_alloc()
     68 {
     69   #ifdef EIGEN_EXCEPTIONS
     70     throw std::bad_alloc();
     71   #else
     72     std::size_t huge = static_cast<std::size_t>(-1);
     73     #if defined(EIGEN_HIPCC)
     74     //
     75     // calls to "::operator new" are to be treated as opaque function calls (i.e no inlining),
     76     // and as a consequence the code in the #else block triggers the hipcc warning :
     77     // "no overloaded function has restriction specifiers that are compatible with the ambient context"
     78     //
     79     // "throw_std_bad_alloc" has the EIGEN_DEVICE_FUNC attribute, so it seems that hipcc expects
     80     // the same on "operator new"
     81     // Reverting code back to the old version in this #if block for the hipcc compiler
     82     //
     83     new int[huge];
     84     #else
     85     void* unused = ::operator new(huge);
     86     EIGEN_UNUSED_VARIABLE(unused);
     87     #endif
     88   #endif
     89 }
     90 
     91 /*****************************************************************************
     92 *** Implementation of handmade aligned functions                           ***
     93 *****************************************************************************/
     94 
     95 /* ----- Hand made implementations of aligned malloc/free and realloc ----- */
     96 
     97 /** \internal Like malloc, but the returned pointer is guaranteed to be 16-byte aligned.
     98   * Fast, but wastes 16 additional bytes of memory. Does not throw any exception.
     99   */
    100 EIGEN_DEVICE_FUNC inline void* handmade_aligned_malloc(std::size_t size, std::size_t alignment = EIGEN_DEFAULT_ALIGN_BYTES)
    101 {
    102   eigen_assert(alignment >= sizeof(void*) && (alignment & (alignment-1)) == 0 && "Alignment must be at least sizeof(void*) and a power of 2");
    103 
    104   EIGEN_USING_STD(malloc)
    105   void *original = malloc(size+alignment);
    106   
    107   if (original == 0) return 0;
    108   void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(alignment-1))) + alignment);
    109   *(reinterpret_cast<void**>(aligned) - 1) = original;
    110   return aligned;
    111 }
    112 
    113 /** \internal Frees memory allocated with handmade_aligned_malloc */
    114 EIGEN_DEVICE_FUNC inline void handmade_aligned_free(void *ptr)
    115 {
    116   if (ptr) {
    117     EIGEN_USING_STD(free)
    118     free(*(reinterpret_cast<void**>(ptr) - 1));
    119   }
    120 }
    121 
    122 /** \internal
    123   * \brief Reallocates aligned memory.
    124   * Since we know that our handmade version is based on std::malloc
    125   * we can use std::realloc to implement efficient reallocation.
    126   */
    127 inline void* handmade_aligned_realloc(void* ptr, std::size_t size, std::size_t = 0)
    128 {
    129   if (ptr == 0) return handmade_aligned_malloc(size);
    130   void *original = *(reinterpret_cast<void**>(ptr) - 1);
    131   std::ptrdiff_t previous_offset = static_cast<char *>(ptr)-static_cast<char *>(original);
    132   original = std::realloc(original,size+EIGEN_DEFAULT_ALIGN_BYTES);
    133   if (original == 0) return 0;
    134   void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1))) + EIGEN_DEFAULT_ALIGN_BYTES);
    135   void *previous_aligned = static_cast<char *>(original)+previous_offset;
    136   if(aligned!=previous_aligned)
    137     std::memmove(aligned, previous_aligned, size);
    138 
    139   *(reinterpret_cast<void**>(aligned) - 1) = original;
    140   return aligned;
    141 }
    142 
    143 /*****************************************************************************
    144 *** Implementation of portable aligned versions of malloc/free/realloc     ***
    145 *****************************************************************************/
    146 
    147 #ifdef EIGEN_NO_MALLOC
    148 EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed()
    149 {
    150   eigen_assert(false && "heap allocation is forbidden (EIGEN_NO_MALLOC is defined)");
    151 }
    152 #elif defined EIGEN_RUNTIME_NO_MALLOC
    153 EIGEN_DEVICE_FUNC inline bool is_malloc_allowed_impl(bool update, bool new_value = false)
    154 {
    155   static bool value = true;
    156   if (update == 1)
    157     value = new_value;
    158   return value;
    159 }
    160 EIGEN_DEVICE_FUNC inline bool is_malloc_allowed() { return is_malloc_allowed_impl(false); }
    161 EIGEN_DEVICE_FUNC inline bool set_is_malloc_allowed(bool new_value) { return is_malloc_allowed_impl(true, new_value); }
    162 EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed()
    163 {
    164   eigen_assert(is_malloc_allowed() && "heap allocation is forbidden (EIGEN_RUNTIME_NO_MALLOC is defined and g_is_malloc_allowed is false)");
    165 }
    166 #else
    167 EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed()
    168 {}
    169 #endif
    170 
    171 /** \internal Allocates \a size bytes. The returned pointer is guaranteed to have 16 or 32 bytes alignment depending on the requirements.
    172   * On allocation error, the returned pointer is null, and std::bad_alloc is thrown.
    173   */
    174 EIGEN_DEVICE_FUNC inline void* aligned_malloc(std::size_t size)
    175 {
    176   check_that_malloc_is_allowed();
    177 
    178   void *result;
    179   #if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED
    180 
    181     EIGEN_USING_STD(malloc)
    182     result = malloc(size);
    183 
    184     #if EIGEN_DEFAULT_ALIGN_BYTES==16
    185     eigen_assert((size<16 || (std::size_t(result)%16)==0) && "System's malloc returned an unaligned pointer. Compile with EIGEN_MALLOC_ALREADY_ALIGNED=0 to fallback to handmade aligned memory allocator.");
    186     #endif
    187   #else
    188     result = handmade_aligned_malloc(size);
    189   #endif
    190 
    191   if(!result && size)
    192     throw_std_bad_alloc();
    193 
    194   return result;
    195 }
    196 
    197 /** \internal Frees memory allocated with aligned_malloc. */
    198 EIGEN_DEVICE_FUNC inline void aligned_free(void *ptr)
    199 {
    200   #if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED
    201 
    202     EIGEN_USING_STD(free)
    203     free(ptr);
    204 
    205   #else
    206     handmade_aligned_free(ptr);
    207   #endif
    208 }
    209 
    210 /**
    211   * \internal
    212   * \brief Reallocates an aligned block of memory.
    213   * \throws std::bad_alloc on allocation failure
    214   */
    215 inline void* aligned_realloc(void *ptr, std::size_t new_size, std::size_t old_size)
    216 {
    217   EIGEN_UNUSED_VARIABLE(old_size)
    218 
    219   void *result;
    220 #if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED
    221   result = std::realloc(ptr,new_size);
    222 #else
    223   result = handmade_aligned_realloc(ptr,new_size,old_size);
    224 #endif
    225 
    226   if (!result && new_size)
    227     throw_std_bad_alloc();
    228 
    229   return result;
    230 }
    231 
    232 /*****************************************************************************
    233 *** Implementation of conditionally aligned functions                      ***
    234 *****************************************************************************/
    235 
    236 /** \internal Allocates \a size bytes. If Align is true, then the returned ptr is 16-byte-aligned.
    237   * On allocation error, the returned pointer is null, and a std::bad_alloc is thrown.
    238   */
    239 template<bool Align> EIGEN_DEVICE_FUNC inline void* conditional_aligned_malloc(std::size_t size)
    240 {
    241   return aligned_malloc(size);
    242 }
    243 
    244 template<> EIGEN_DEVICE_FUNC inline void* conditional_aligned_malloc<false>(std::size_t size)
    245 {
    246   check_that_malloc_is_allowed();
    247 
    248   EIGEN_USING_STD(malloc)
    249   void *result = malloc(size);
    250 
    251   if(!result && size)
    252     throw_std_bad_alloc();
    253   return result;
    254 }
    255 
    256 /** \internal Frees memory allocated with conditional_aligned_malloc */
    257 template<bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_free(void *ptr)
    258 {
    259   aligned_free(ptr);
    260 }
    261 
    262 template<> EIGEN_DEVICE_FUNC inline void conditional_aligned_free<false>(void *ptr)
    263 {
    264   EIGEN_USING_STD(free)
    265   free(ptr);
    266 }
    267 
    268 template<bool Align> inline void* conditional_aligned_realloc(void* ptr, std::size_t new_size, std::size_t old_size)
    269 {
    270   return aligned_realloc(ptr, new_size, old_size);
    271 }
    272 
    273 template<> inline void* conditional_aligned_realloc<false>(void* ptr, std::size_t new_size, std::size_t)
    274 {
    275   return std::realloc(ptr, new_size);
    276 }
    277 
    278 /*****************************************************************************
    279 *** Construction/destruction of array elements                             ***
    280 *****************************************************************************/
    281 
    282 /** \internal Destructs the elements of an array.
    283   * The \a size parameters tells on how many objects to call the destructor of T.
    284   */
    285 template<typename T> EIGEN_DEVICE_FUNC inline void destruct_elements_of_array(T *ptr, std::size_t size)
    286 {
    287   // always destruct an array starting from the end.
    288   if(ptr)
    289     while(size) ptr[--size].~T();
    290 }
    291 
    292 /** \internal Constructs the elements of an array.
    293   * The \a size parameter tells on how many objects to call the constructor of T.
    294   */
    295 template<typename T> EIGEN_DEVICE_FUNC inline T* construct_elements_of_array(T *ptr, std::size_t size)
    296 {
    297   std::size_t i;
    298   EIGEN_TRY
    299   {
    300       for (i = 0; i < size; ++i) ::new (ptr + i) T;
    301       return ptr;
    302   }
    303   EIGEN_CATCH(...)
    304   {
    305     destruct_elements_of_array(ptr, i);
    306     EIGEN_THROW;
    307   }
    308   return NULL;
    309 }
    310 
    311 /*****************************************************************************
    312 *** Implementation of aligned new/delete-like functions                    ***
    313 *****************************************************************************/
    314 
    315 template<typename T>
    316 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void check_size_for_overflow(std::size_t size)
    317 {
    318   if(size > std::size_t(-1) / sizeof(T))
    319     throw_std_bad_alloc();
    320 }
    321 
    322 /** \internal Allocates \a size objects of type T. The returned pointer is guaranteed to have 16 bytes alignment.
    323   * On allocation error, the returned pointer is undefined, but a std::bad_alloc is thrown.
    324   * The default constructor of T is called.
    325   */
    326 template<typename T> EIGEN_DEVICE_FUNC inline T* aligned_new(std::size_t size)
    327 {
    328   check_size_for_overflow<T>(size);
    329   T *result = reinterpret_cast<T*>(aligned_malloc(sizeof(T)*size));
    330   EIGEN_TRY
    331   {
    332     return construct_elements_of_array(result, size);
    333   }
    334   EIGEN_CATCH(...)
    335   {
    336     aligned_free(result);
    337     EIGEN_THROW;
    338   }
    339   return result;
    340 }
    341 
    342 template<typename T, bool Align> EIGEN_DEVICE_FUNC inline T* conditional_aligned_new(std::size_t size)
    343 {
    344   check_size_for_overflow<T>(size);
    345   T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
    346   EIGEN_TRY
    347   {
    348     return construct_elements_of_array(result, size);
    349   }
    350   EIGEN_CATCH(...)
    351   {
    352     conditional_aligned_free<Align>(result);
    353     EIGEN_THROW;
    354   }
    355   return result;
    356 }
    357 
    358 /** \internal Deletes objects constructed with aligned_new
    359   * The \a size parameters tells on how many objects to call the destructor of T.
    360   */
    361 template<typename T> EIGEN_DEVICE_FUNC inline void aligned_delete(T *ptr, std::size_t size)
    362 {
    363   destruct_elements_of_array<T>(ptr, size);
    364   Eigen::internal::aligned_free(ptr);
    365 }
    366 
    367 /** \internal Deletes objects constructed with conditional_aligned_new
    368   * The \a size parameters tells on how many objects to call the destructor of T.
    369   */
    370 template<typename T, bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_delete(T *ptr, std::size_t size)
    371 {
    372   destruct_elements_of_array<T>(ptr, size);
    373   conditional_aligned_free<Align>(ptr);
    374 }
    375 
    376 template<typename T, bool Align> EIGEN_DEVICE_FUNC inline T* conditional_aligned_realloc_new(T* pts, std::size_t new_size, std::size_t old_size)
    377 {
    378   check_size_for_overflow<T>(new_size);
    379   check_size_for_overflow<T>(old_size);
    380   if(new_size < old_size)
    381     destruct_elements_of_array(pts+new_size, old_size-new_size);
    382   T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
    383   if(new_size > old_size)
    384   {
    385     EIGEN_TRY
    386     {
    387       construct_elements_of_array(result+old_size, new_size-old_size);
    388     }
    389     EIGEN_CATCH(...)
    390     {
    391       conditional_aligned_free<Align>(result);
    392       EIGEN_THROW;
    393     }
    394   }
    395   return result;
    396 }
    397 
    398 
    399 template<typename T, bool Align> EIGEN_DEVICE_FUNC inline T* conditional_aligned_new_auto(std::size_t size)
    400 {
    401   if(size==0)
    402     return 0; // short-cut. Also fixes Bug 884
    403   check_size_for_overflow<T>(size);
    404   T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
    405   if(NumTraits<T>::RequireInitialization)
    406   {
    407     EIGEN_TRY
    408     {
    409       construct_elements_of_array(result, size);
    410     }
    411     EIGEN_CATCH(...)
    412     {
    413       conditional_aligned_free<Align>(result);
    414       EIGEN_THROW;
    415     }
    416   }
    417   return result;
    418 }
    419 
    420 template<typename T, bool Align> inline T* conditional_aligned_realloc_new_auto(T* pts, std::size_t new_size, std::size_t old_size)
    421 {
    422   check_size_for_overflow<T>(new_size);
    423   check_size_for_overflow<T>(old_size);
    424   if(NumTraits<T>::RequireInitialization && (new_size < old_size))
    425     destruct_elements_of_array(pts+new_size, old_size-new_size);
    426   T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
    427   if(NumTraits<T>::RequireInitialization && (new_size > old_size))
    428   {
    429     EIGEN_TRY
    430     {
    431       construct_elements_of_array(result+old_size, new_size-old_size);
    432     }
    433     EIGEN_CATCH(...)
    434     {
    435       conditional_aligned_free<Align>(result);
    436       EIGEN_THROW;
    437     }
    438   }
    439   return result;
    440 }
    441 
    442 template<typename T, bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_delete_auto(T *ptr, std::size_t size)
    443 {
    444   if(NumTraits<T>::RequireInitialization)
    445     destruct_elements_of_array<T>(ptr, size);
    446   conditional_aligned_free<Align>(ptr);
    447 }
    448 
    449 /****************************************************************************/
    450 
    451 /** \internal Returns the index of the first element of the array that is well aligned with respect to the requested \a Alignment.
    452   *
    453   * \tparam Alignment requested alignment in Bytes.
    454   * \param array the address of the start of the array
    455   * \param size the size of the array
    456   *
    457   * \note If no element of the array is well aligned or the requested alignment is not a multiple of a scalar,
    458   * the size of the array is returned. For example with SSE, the requested alignment is typically 16-bytes. If
    459   * packet size for the given scalar type is 1, then everything is considered well-aligned.
    460   *
    461   * \note Otherwise, if the Alignment is larger that the scalar size, we rely on the assumptions that sizeof(Scalar) is a
    462   * power of 2. On the other hand, we do not assume that the array address is a multiple of sizeof(Scalar), as that fails for
    463   * example with Scalar=double on certain 32-bit platforms, see bug #79.
    464   *
    465   * There is also the variant first_aligned(const MatrixBase&) defined in DenseCoeffsBase.h.
    466   * \sa first_default_aligned()
    467   */
    468 template<int Alignment, typename Scalar, typename Index>
    469 EIGEN_DEVICE_FUNC inline Index first_aligned(const Scalar* array, Index size)
    470 {
    471   const Index ScalarSize = sizeof(Scalar);
    472   const Index AlignmentSize = Alignment / ScalarSize;
    473   const Index AlignmentMask = AlignmentSize-1;
    474 
    475   if(AlignmentSize<=1)
    476   {
    477     // Either the requested alignment if smaller than a scalar, or it exactly match a 1 scalar
    478     // so that all elements of the array have the same alignment.
    479     return 0;
    480   }
    481   else if( (UIntPtr(array) & (sizeof(Scalar)-1)) || (Alignment%ScalarSize)!=0)
    482   {
    483     // The array is not aligned to the size of a single scalar, or the requested alignment is not a multiple of the scalar size.
    484     // Consequently, no element of the array is well aligned.
    485     return size;
    486   }
    487   else
    488   {
    489     Index first = (AlignmentSize - (Index((UIntPtr(array)/sizeof(Scalar))) & AlignmentMask)) & AlignmentMask;
    490     return (first < size) ? first : size;
    491   }
    492 }
    493 
    494 /** \internal Returns the index of the first element of the array that is well aligned with respect the largest packet requirement.
    495    * \sa first_aligned(Scalar*,Index) and first_default_aligned(DenseBase<Derived>) */
    496 template<typename Scalar, typename Index>
    497 EIGEN_DEVICE_FUNC inline Index first_default_aligned(const Scalar* array, Index size)
    498 {
    499   typedef typename packet_traits<Scalar>::type DefaultPacketType;
    500   return first_aligned<unpacket_traits<DefaultPacketType>::alignment>(array, size);
    501 }
    502 
    503 /** \internal Returns the smallest integer multiple of \a base and greater or equal to \a size
    504   */
    505 template<typename Index>
    506 inline Index first_multiple(Index size, Index base)
    507 {
    508   return ((size+base-1)/base)*base;
    509 }
    510 
    511 // std::copy is much slower than memcpy, so let's introduce a smart_copy which
    512 // use memcpy on trivial types, i.e., on types that does not require an initialization ctor.
    513 template<typename T, bool UseMemcpy> struct smart_copy_helper;
    514 
    515 template<typename T> EIGEN_DEVICE_FUNC void smart_copy(const T* start, const T* end, T* target)
    516 {
    517   smart_copy_helper<T,!NumTraits<T>::RequireInitialization>::run(start, end, target);
    518 }
    519 
    520 template<typename T> struct smart_copy_helper<T,true> {
    521   EIGEN_DEVICE_FUNC static inline void run(const T* start, const T* end, T* target)
    522   {
    523     IntPtr size = IntPtr(end)-IntPtr(start);
    524     if(size==0) return;
    525     eigen_internal_assert(start!=0 && end!=0 && target!=0);
    526     EIGEN_USING_STD(memcpy)
    527     memcpy(target, start, size);
    528   }
    529 };
    530 
    531 template<typename T> struct smart_copy_helper<T,false> {
    532   EIGEN_DEVICE_FUNC static inline void run(const T* start, const T* end, T* target)
    533   { std::copy(start, end, target); }
    534 };
    535 
    536 // intelligent memmove. falls back to std::memmove for POD types, uses std::copy otherwise.
    537 template<typename T, bool UseMemmove> struct smart_memmove_helper;
    538 
    539 template<typename T> void smart_memmove(const T* start, const T* end, T* target)
    540 {
    541   smart_memmove_helper<T,!NumTraits<T>::RequireInitialization>::run(start, end, target);
    542 }
    543 
    544 template<typename T> struct smart_memmove_helper<T,true> {
    545   static inline void run(const T* start, const T* end, T* target)
    546   {
    547     IntPtr size = IntPtr(end)-IntPtr(start);
    548     if(size==0) return;
    549     eigen_internal_assert(start!=0 && end!=0 && target!=0);
    550     std::memmove(target, start, size);
    551   }
    552 };
    553 
    554 template<typename T> struct smart_memmove_helper<T,false> {
    555   static inline void run(const T* start, const T* end, T* target)
    556   {
    557     if (UIntPtr(target) < UIntPtr(start))
    558     {
    559       std::copy(start, end, target);
    560     }
    561     else
    562     {
    563       std::ptrdiff_t count = (std::ptrdiff_t(end)-std::ptrdiff_t(start)) / sizeof(T);
    564       std::copy_backward(start, end, target + count);
    565     }
    566   }
    567 };
    568 
    569 #if EIGEN_HAS_RVALUE_REFERENCES
    570 template<typename T> EIGEN_DEVICE_FUNC T* smart_move(T* start, T* end, T* target)
    571 {
    572   return std::move(start, end, target);
    573 }
    574 #else
    575 template<typename T> EIGEN_DEVICE_FUNC T* smart_move(T* start, T* end, T* target)
    576 {
    577   return std::copy(start, end, target);
    578 }
    579 #endif
    580 
    581 /*****************************************************************************
    582 *** Implementation of runtime stack allocation (falling back to malloc)    ***
    583 *****************************************************************************/
    584 
    585 // you can overwrite Eigen's default behavior regarding alloca by defining EIGEN_ALLOCA
    586 // to the appropriate stack allocation function
    587 #if ! defined EIGEN_ALLOCA && ! defined EIGEN_GPU_COMPILE_PHASE
    588   #if EIGEN_OS_LINUX || EIGEN_OS_MAC || (defined alloca)
    589     #define EIGEN_ALLOCA alloca
    590   #elif EIGEN_COMP_MSVC
    591     #define EIGEN_ALLOCA _alloca
    592   #endif
    593 #endif
    594 
    595 // With clang -Oz -mthumb, alloca changes the stack pointer in a way that is
    596 // not allowed in Thumb2. -DEIGEN_STACK_ALLOCATION_LIMIT=0 doesn't work because
    597 // the compiler still emits bad code because stack allocation checks use "<=".
    598 // TODO: Eliminate after https://bugs.llvm.org/show_bug.cgi?id=23772
    599 // is fixed.
    600 #if defined(__clang__) && defined(__thumb__)
    601   #undef EIGEN_ALLOCA
    602 #endif
    603 
    604 // This helper class construct the allocated memory, and takes care of destructing and freeing the handled data
    605 // at destruction time. In practice this helper class is mainly useful to avoid memory leak in case of exceptions.
    606 template<typename T> class aligned_stack_memory_handler : noncopyable
    607 {
    608   public:
    609     /* Creates a stack_memory_handler responsible for the buffer \a ptr of size \a size.
    610      * Note that \a ptr can be 0 regardless of the other parameters.
    611      * This constructor takes care of constructing/initializing the elements of the buffer if required by the scalar type T (see NumTraits<T>::RequireInitialization).
    612      * In this case, the buffer elements will also be destructed when this handler will be destructed.
    613      * Finally, if \a dealloc is true, then the pointer \a ptr is freed.
    614      **/
    615     EIGEN_DEVICE_FUNC
    616     aligned_stack_memory_handler(T* ptr, std::size_t size, bool dealloc)
    617       : m_ptr(ptr), m_size(size), m_deallocate(dealloc)
    618     {
    619       if(NumTraits<T>::RequireInitialization && m_ptr)
    620         Eigen::internal::construct_elements_of_array(m_ptr, size);
    621     }
    622     EIGEN_DEVICE_FUNC
    623     ~aligned_stack_memory_handler()
    624     {
    625       if(NumTraits<T>::RequireInitialization && m_ptr)
    626         Eigen::internal::destruct_elements_of_array<T>(m_ptr, m_size);
    627       if(m_deallocate)
    628         Eigen::internal::aligned_free(m_ptr);
    629     }
    630   protected:
    631     T* m_ptr;
    632     std::size_t m_size;
    633     bool m_deallocate;
    634 };
    635 
    636 #ifdef EIGEN_ALLOCA
    637 
    638 template<typename Xpr, int NbEvaluations,
    639          bool MapExternalBuffer = nested_eval<Xpr,NbEvaluations>::Evaluate && Xpr::MaxSizeAtCompileTime==Dynamic
    640          >
    641 struct local_nested_eval_wrapper
    642 {
    643   static const bool NeedExternalBuffer = false;
    644   typedef typename Xpr::Scalar Scalar;
    645   typedef typename nested_eval<Xpr,NbEvaluations>::type ObjectType;
    646   ObjectType object;
    647 
    648   EIGEN_DEVICE_FUNC
    649   local_nested_eval_wrapper(const Xpr& xpr, Scalar* ptr) : object(xpr)
    650   {
    651     EIGEN_UNUSED_VARIABLE(ptr);
    652     eigen_internal_assert(ptr==0);
    653   }
    654 };
    655 
    656 template<typename Xpr, int NbEvaluations>
    657 struct local_nested_eval_wrapper<Xpr,NbEvaluations,true>
    658 {
    659   static const bool NeedExternalBuffer = true;
    660   typedef typename Xpr::Scalar Scalar;
    661   typedef typename plain_object_eval<Xpr>::type PlainObject;
    662   typedef Map<PlainObject,EIGEN_DEFAULT_ALIGN_BYTES> ObjectType;
    663   ObjectType object;
    664 
    665   EIGEN_DEVICE_FUNC
    666   local_nested_eval_wrapper(const Xpr& xpr, Scalar* ptr)
    667     : object(ptr==0 ? reinterpret_cast<Scalar*>(Eigen::internal::aligned_malloc(sizeof(Scalar)*xpr.size())) : ptr, xpr.rows(), xpr.cols()),
    668       m_deallocate(ptr==0)
    669   {
    670     if(NumTraits<Scalar>::RequireInitialization && object.data())
    671       Eigen::internal::construct_elements_of_array(object.data(), object.size());
    672     object = xpr;
    673   }
    674 
    675   EIGEN_DEVICE_FUNC
    676   ~local_nested_eval_wrapper()
    677   {
    678     if(NumTraits<Scalar>::RequireInitialization && object.data())
    679       Eigen::internal::destruct_elements_of_array(object.data(), object.size());
    680     if(m_deallocate)
    681       Eigen::internal::aligned_free(object.data());
    682   }
    683 
    684 private:
    685   bool m_deallocate;
    686 };
    687 
    688 #endif // EIGEN_ALLOCA
    689 
    690 template<typename T> class scoped_array : noncopyable
    691 {
    692   T* m_ptr;
    693 public:
    694   explicit scoped_array(std::ptrdiff_t size)
    695   {
    696     m_ptr = new T[size];
    697   }
    698   ~scoped_array()
    699   {
    700     delete[] m_ptr;
    701   }
    702   T& operator[](std::ptrdiff_t i) { return m_ptr[i]; }
    703   const T& operator[](std::ptrdiff_t i) const { return m_ptr[i]; }
    704   T* &ptr() { return m_ptr; }
    705   const T* ptr() const { return m_ptr; }
    706   operator const T*() const { return m_ptr; }
    707 };
    708 
    709 template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b)
    710 {
    711   std::swap(a.ptr(),b.ptr());
    712 }
    713 
    714 } // end namespace internal
    715 
    716 /** \internal
    717   *
    718   * The macro ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) declares, allocates,
    719   * and construct an aligned buffer named NAME of SIZE elements of type TYPE on the stack
    720   * if the size in bytes is smaller than EIGEN_STACK_ALLOCATION_LIMIT, and if stack allocation is supported by the platform
    721   * (currently, this is Linux, OSX and Visual Studio only). Otherwise the memory is allocated on the heap.
    722   * The allocated buffer is automatically deleted when exiting the scope of this declaration.
    723   * If BUFFER is non null, then the declared variable is simply an alias for BUFFER, and no allocation/deletion occurs.
    724   * Here is an example:
    725   * \code
    726   * {
    727   *   ei_declare_aligned_stack_constructed_variable(float,data,size,0);
    728   *   // use data[0] to data[size-1]
    729   * }
    730   * \endcode
    731   * The underlying stack allocation function can controlled with the EIGEN_ALLOCA preprocessor token.
    732   *
    733   * The macro ei_declare_local_nested_eval(XPR_T,XPR,N,NAME) is analogue to
    734   * \code
    735   *   typename internal::nested_eval<XPRT_T,N>::type NAME(XPR);
    736   * \endcode
    737   * with the advantage of using aligned stack allocation even if the maximal size of XPR at compile time is unknown.
    738   * This is accomplished through alloca if this later is supported and if the required number of bytes
    739   * is below EIGEN_STACK_ALLOCATION_LIMIT.
    740   */
    741 #ifdef EIGEN_ALLOCA
    742 
    743   #if EIGEN_DEFAULT_ALIGN_BYTES>0
    744     // We always manually re-align the result of EIGEN_ALLOCA.
    745     // If alloca is already aligned, the compiler should be smart enough to optimize away the re-alignment.
    746     #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast<void*>((internal::UIntPtr(EIGEN_ALLOCA(SIZE+EIGEN_DEFAULT_ALIGN_BYTES-1)) + EIGEN_DEFAULT_ALIGN_BYTES-1) & ~(std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1)))
    747   #else
    748     #define EIGEN_ALIGNED_ALLOCA(SIZE) EIGEN_ALLOCA(SIZE)
    749   #endif
    750 
    751   #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
    752     Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \
    753     TYPE* NAME = (BUFFER)!=0 ? (BUFFER) \
    754                : reinterpret_cast<TYPE*>( \
    755                       (sizeof(TYPE)*SIZE<=EIGEN_STACK_ALLOCATION_LIMIT) ? EIGEN_ALIGNED_ALLOCA(sizeof(TYPE)*SIZE) \
    756                     : Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE) );  \
    757     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)
    758 
    759 
    760   #define ei_declare_local_nested_eval(XPR_T,XPR,N,NAME) \
    761     Eigen::internal::local_nested_eval_wrapper<XPR_T,N> EIGEN_CAT(NAME,_wrapper)(XPR, reinterpret_cast<typename XPR_T::Scalar*>( \
    762       ( (Eigen::internal::local_nested_eval_wrapper<XPR_T,N>::NeedExternalBuffer) && ((sizeof(typename XPR_T::Scalar)*XPR.size())<=EIGEN_STACK_ALLOCATION_LIMIT) ) \
    763         ? EIGEN_ALIGNED_ALLOCA( sizeof(typename XPR_T::Scalar)*XPR.size() ) : 0 ) ) ; \
    764     typename Eigen::internal::local_nested_eval_wrapper<XPR_T,N>::ObjectType NAME(EIGEN_CAT(NAME,_wrapper).object)
    765 
    766 #else
    767 
    768   #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
    769     Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \
    770     TYPE* NAME = (BUFFER)!=0 ? BUFFER : reinterpret_cast<TYPE*>(Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE));    \
    771     Eigen::internal::aligned_stack_memory_handler<TYPE> EIGEN_CAT(NAME,_stack_memory_destructor)((BUFFER)==0 ? NAME : 0,SIZE,true)
    772 
    773 
    774 #define ei_declare_local_nested_eval(XPR_T,XPR,N,NAME) typename Eigen::internal::nested_eval<XPR_T,N>::type NAME(XPR)
    775 
    776 #endif
    777 
    778 
    779 /*****************************************************************************
    780 *** Implementation of EIGEN_MAKE_ALIGNED_OPERATOR_NEW [_IF]                ***
    781 *****************************************************************************/
    782 
    783 #if EIGEN_HAS_CXX17_OVERALIGN
    784 
    785 // C++17 -> no need to bother about alignment anymore :)
    786 
    787 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign)
    788 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)
    789 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW
    790 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size)
    791 
    792 #else
    793 
    794 // HIP does not support new/delete on device.
    795 #if EIGEN_MAX_ALIGN_BYTES!=0 && !defined(EIGEN_HIP_DEVICE_COMPILE)
    796   #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
    797       EIGEN_DEVICE_FUNC \
    798       void* operator new(std::size_t size, const std::nothrow_t&) EIGEN_NO_THROW { \
    799         EIGEN_TRY { return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); } \
    800         EIGEN_CATCH (...) { return 0; } \
    801       }
    802   #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign) \
    803       EIGEN_DEVICE_FUNC \
    804       void *operator new(std::size_t size) { \
    805         return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
    806       } \
    807       EIGEN_DEVICE_FUNC \
    808       void *operator new[](std::size_t size) { \
    809         return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
    810       } \
    811       EIGEN_DEVICE_FUNC \
    812       void operator delete(void * ptr) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
    813       EIGEN_DEVICE_FUNC \
    814       void operator delete[](void * ptr) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
    815       EIGEN_DEVICE_FUNC \
    816       void operator delete(void * ptr, std::size_t /* sz */) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
    817       EIGEN_DEVICE_FUNC \
    818       void operator delete[](void * ptr, std::size_t /* sz */) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
    819       /* in-place new and delete. since (at least afaik) there is no actual   */ \
    820       /* memory allocated we can safely let the default implementation handle */ \
    821       /* this particular case. */ \
    822       EIGEN_DEVICE_FUNC \
    823       static void *operator new(std::size_t size, void *ptr) { return ::operator new(size,ptr); } \
    824       EIGEN_DEVICE_FUNC \
    825       static void *operator new[](std::size_t size, void* ptr) { return ::operator new[](size,ptr); } \
    826       EIGEN_DEVICE_FUNC \
    827       void operator delete(void * memory, void *ptr) EIGEN_NO_THROW { return ::operator delete(memory,ptr); } \
    828       EIGEN_DEVICE_FUNC \
    829       void operator delete[](void * memory, void *ptr) EIGEN_NO_THROW { return ::operator delete[](memory,ptr); } \
    830       /* nothrow-new (returns zero instead of std::bad_alloc) */ \
    831       EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
    832       EIGEN_DEVICE_FUNC \
    833       void operator delete(void *ptr, const std::nothrow_t&) EIGEN_NO_THROW { \
    834         Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); \
    835       } \
    836       typedef void eigen_aligned_operator_new_marker_type;
    837 #else
    838   #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)
    839 #endif
    840 
    841 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(true)
    842 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size)                        \
    843   EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool(                                                             \
    844         ((Size)!=Eigen::Dynamic) &&                                                                    \
    845         (((EIGEN_MAX_ALIGN_BYTES>=16) && ((sizeof(Scalar)*(Size))%(EIGEN_MAX_ALIGN_BYTES  )==0)) ||    \
    846          ((EIGEN_MAX_ALIGN_BYTES>=32) && ((sizeof(Scalar)*(Size))%(EIGEN_MAX_ALIGN_BYTES/2)==0)) ||    \
    847          ((EIGEN_MAX_ALIGN_BYTES>=64) && ((sizeof(Scalar)*(Size))%(EIGEN_MAX_ALIGN_BYTES/4)==0))   )))
    848 
    849 #endif
    850 
    851 /****************************************************************************/
    852 
    853 /** \class aligned_allocator
    854 * \ingroup Core_Module
    855 *
    856 * \brief STL compatible allocator to use with types requiring a non standrad alignment.
    857 *
    858 * The memory is aligned as for dynamically aligned matrix/array types such as MatrixXd.
    859 * By default, it will thus provide at least 16 bytes alignment and more in following cases:
    860 *  - 32 bytes alignment if AVX is enabled.
    861 *  - 64 bytes alignment if AVX512 is enabled.
    862 *
    863 * This can be controlled using the \c EIGEN_MAX_ALIGN_BYTES macro as documented
    864 * \link TopicPreprocessorDirectivesPerformance there \endlink.
    865 *
    866 * Example:
    867 * \code
    868 * // Matrix4f requires 16 bytes alignment:
    869 * std::map< int, Matrix4f, std::less<int>,
    870 *           aligned_allocator<std::pair<const int, Matrix4f> > > my_map_mat4;
    871 * // Vector3f does not require 16 bytes alignment, no need to use Eigen's allocator:
    872 * std::map< int, Vector3f > my_map_vec3;
    873 * \endcode
    874 *
    875 * \sa \blank \ref TopicStlContainers.
    876 */
    877 template<class T>
    878 class aligned_allocator : public std::allocator<T>
    879 {
    880 public:
    881   typedef std::size_t     size_type;
    882   typedef std::ptrdiff_t  difference_type;
    883   typedef T*              pointer;
    884   typedef const T*        const_pointer;
    885   typedef T&              reference;
    886   typedef const T&        const_reference;
    887   typedef T               value_type;
    888 
    889   template<class U>
    890   struct rebind
    891   {
    892     typedef aligned_allocator<U> other;
    893   };
    894 
    895   aligned_allocator() : std::allocator<T>() {}
    896 
    897   aligned_allocator(const aligned_allocator& other) : std::allocator<T>(other) {}
    898 
    899   template<class U>
    900   aligned_allocator(const aligned_allocator<U>& other) : std::allocator<T>(other) {}
    901 
    902   ~aligned_allocator() {}
    903 
    904   #if EIGEN_COMP_GNUC_STRICT && EIGEN_GNUC_AT_LEAST(7,0)
    905   // In gcc std::allocator::max_size() is bugged making gcc triggers a warning:
    906   // eigen/Eigen/src/Core/util/Memory.h:189:12: warning: argument 1 value '18446744073709551612' exceeds maximum object size 9223372036854775807
    907   // See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87544
    908   size_type max_size() const {
    909     return (std::numeric_limits<std::ptrdiff_t>::max)()/sizeof(T);
    910   }
    911   #endif
    912 
    913   pointer allocate(size_type num, const void* /*hint*/ = 0)
    914   {
    915     internal::check_size_for_overflow<T>(num);
    916     return static_cast<pointer>( internal::aligned_malloc(num * sizeof(T)) );
    917   }
    918 
    919   void deallocate(pointer p, size_type /*num*/)
    920   {
    921     internal::aligned_free(p);
    922   }
    923 };
    924 
    925 //---------- Cache sizes ----------
    926 
    927 #if !defined(EIGEN_NO_CPUID)
    928 #  if EIGEN_COMP_GNUC && EIGEN_ARCH_i386_OR_x86_64
    929 #    if defined(__PIC__) && EIGEN_ARCH_i386
    930        // Case for x86 with PIC
    931 #      define EIGEN_CPUID(abcd,func,id) \
    932          __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));
    933 #    elif defined(__PIC__) && EIGEN_ARCH_x86_64
    934        // 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.
    935        // However, we cannot detect which code model is used, and the xchg overhead is negligible anyway.
    936 #      define EIGEN_CPUID(abcd,func,id) \
    937         __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));
    938 #    else
    939        // Case for x86_64 or x86 w/o PIC
    940 #      define EIGEN_CPUID(abcd,func,id) \
    941          __asm__ __volatile__ ("cpuid": "=a" (abcd[0]), "=b" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "0" (func), "2" (id) );
    942 #    endif
    943 #  elif EIGEN_COMP_MSVC
    944 #    if (EIGEN_COMP_MSVC > 1500) && EIGEN_ARCH_i386_OR_x86_64
    945 #      define EIGEN_CPUID(abcd,func,id) __cpuidex((int*)abcd,func,id)
    946 #    endif
    947 #  endif
    948 #endif
    949 
    950 namespace internal {
    951 
    952 #ifdef EIGEN_CPUID
    953 
    954 inline bool cpuid_is_vendor(int abcd[4], const int vendor[3])
    955 {
    956   return abcd[1]==vendor[0] && abcd[3]==vendor[1] && abcd[2]==vendor[2];
    957 }
    958 
    959 inline void queryCacheSizes_intel_direct(int& l1, int& l2, int& l3)
    960 {
    961   int abcd[4];
    962   l1 = l2 = l3 = 0;
    963   int cache_id = 0;
    964   int cache_type = 0;
    965   do {
    966     abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
    967     EIGEN_CPUID(abcd,0x4,cache_id);
    968     cache_type  = (abcd[0] & 0x0F) >> 0;
    969     if(cache_type==1||cache_type==3) // data or unified cache
    970     {
    971       int cache_level = (abcd[0] & 0xE0) >> 5;  // A[7:5]
    972       int ways        = (abcd[1] & 0xFFC00000) >> 22; // B[31:22]
    973       int partitions  = (abcd[1] & 0x003FF000) >> 12; // B[21:12]
    974       int line_size   = (abcd[1] & 0x00000FFF) >>  0; // B[11:0]
    975       int sets        = (abcd[2]);                    // C[31:0]
    976 
    977       int cache_size = (ways+1) * (partitions+1) * (line_size+1) * (sets+1);
    978 
    979       switch(cache_level)
    980       {
    981         case 1: l1 = cache_size; break;
    982         case 2: l2 = cache_size; break;
    983         case 3: l3 = cache_size; break;
    984         default: break;
    985       }
    986     }
    987     cache_id++;
    988   } while(cache_type>0 && cache_id<16);
    989 }
    990 
    991 inline void queryCacheSizes_intel_codes(int& l1, int& l2, int& l3)
    992 {
    993   int abcd[4];
    994   abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
    995   l1 = l2 = l3 = 0;
    996   EIGEN_CPUID(abcd,0x00000002,0);
    997   unsigned char * bytes = reinterpret_cast<unsigned char *>(abcd)+2;
    998   bool check_for_p2_core2 = false;
    999   for(int i=0; i<14; ++i)
   1000   {
   1001     switch(bytes[i])
   1002     {
   1003       case 0x0A: l1 = 8; break;   // 0Ah   data L1 cache, 8 KB, 2 ways, 32 byte lines
   1004       case 0x0C: l1 = 16; break;  // 0Ch   data L1 cache, 16 KB, 4 ways, 32 byte lines
   1005       case 0x0E: l1 = 24; break;  // 0Eh   data L1 cache, 24 KB, 6 ways, 64 byte lines
   1006       case 0x10: l1 = 16; break;  // 10h   data L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
   1007       case 0x15: l1 = 16; break;  // 15h   code L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
   1008       case 0x2C: l1 = 32; break;  // 2Ch   data L1 cache, 32 KB, 8 ways, 64 byte lines
   1009       case 0x30: l1 = 32; break;  // 30h   code L1 cache, 32 KB, 8 ways, 64 byte lines
   1010       case 0x60: l1 = 16; break;  // 60h   data L1 cache, 16 KB, 8 ways, 64 byte lines, sectored
   1011       case 0x66: l1 = 8; break;   // 66h   data L1 cache, 8 KB, 4 ways, 64 byte lines, sectored
   1012       case 0x67: l1 = 16; break;  // 67h   data L1 cache, 16 KB, 4 ways, 64 byte lines, sectored
   1013       case 0x68: l1 = 32; break;  // 68h   data L1 cache, 32 KB, 4 ways, 64 byte lines, sectored
   1014       case 0x1A: l2 = 96; break;   // code and data L2 cache, 96 KB, 6 ways, 64 byte lines (IA-64)
   1015       case 0x22: l3 = 512; break;   // code and data L3 cache, 512 KB, 4 ways (!), 64 byte lines, dual-sectored
   1016       case 0x23: l3 = 1024; break;   // code and data L3 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
   1017       case 0x25: l3 = 2048; break;   // code and data L3 cache, 2048 KB, 8 ways, 64 byte lines, dual-sectored
   1018       case 0x29: l3 = 4096; break;   // code and data L3 cache, 4096 KB, 8 ways, 64 byte lines, dual-sectored
   1019       case 0x39: l2 = 128; break;   // code and data L2 cache, 128 KB, 4 ways, 64 byte lines, sectored
   1020       case 0x3A: l2 = 192; break;   // code and data L2 cache, 192 KB, 6 ways, 64 byte lines, sectored
   1021       case 0x3B: l2 = 128; break;   // code and data L2 cache, 128 KB, 2 ways, 64 byte lines, sectored
   1022       case 0x3C: l2 = 256; break;   // code and data L2 cache, 256 KB, 4 ways, 64 byte lines, sectored
   1023       case 0x3D: l2 = 384; break;   // code and data L2 cache, 384 KB, 6 ways, 64 byte lines, sectored
   1024       case 0x3E: l2 = 512; break;   // code and data L2 cache, 512 KB, 4 ways, 64 byte lines, sectored
   1025       case 0x40: l2 = 0; break;   // no integrated L2 cache (P6 core) or L3 cache (P4 core)
   1026       case 0x41: l2 = 128; break;   // code and data L2 cache, 128 KB, 4 ways, 32 byte lines
   1027       case 0x42: l2 = 256; break;   // code and data L2 cache, 256 KB, 4 ways, 32 byte lines
   1028       case 0x43: l2 = 512; break;   // code and data L2 cache, 512 KB, 4 ways, 32 byte lines
   1029       case 0x44: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 4 ways, 32 byte lines
   1030       case 0x45: l2 = 2048; break;   // code and data L2 cache, 2048 KB, 4 ways, 32 byte lines
   1031       case 0x46: l3 = 4096; break;   // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines
   1032       case 0x47: l3 = 8192; break;   // code and data L3 cache, 8192 KB, 8 ways, 64 byte lines
   1033       case 0x48: l2 = 3072; break;   // code and data L2 cache, 3072 KB, 12 ways, 64 byte lines
   1034       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
   1035       case 0x4A: l3 = 6144; break;   // code and data L3 cache, 6144 KB, 12 ways, 64 byte lines
   1036       case 0x4B: l3 = 8192; break;   // code and data L3 cache, 8192 KB, 16 ways, 64 byte lines
   1037       case 0x4C: l3 = 12288; break;   // code and data L3 cache, 12288 KB, 12 ways, 64 byte lines
   1038       case 0x4D: l3 = 16384; break;   // code and data L3 cache, 16384 KB, 16 ways, 64 byte lines
   1039       case 0x4E: l2 = 6144; break;   // code and data L2 cache, 6144 KB, 24 ways, 64 byte lines
   1040       case 0x78: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 4 ways, 64 byte lines
   1041       case 0x79: l2 = 128; break;   // code and data L2 cache, 128 KB, 8 ways, 64 byte lines, dual-sectored
   1042       case 0x7A: l2 = 256; break;   // code and data L2 cache, 256 KB, 8 ways, 64 byte lines, dual-sectored
   1043       case 0x7B: l2 = 512; break;   // code and data L2 cache, 512 KB, 8 ways, 64 byte lines, dual-sectored
   1044       case 0x7C: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
   1045       case 0x7D: l2 = 2048; break;   // code and data L2 cache, 2048 KB, 8 ways, 64 byte lines
   1046       case 0x7E: l2 = 256; break;   // code and data L2 cache, 256 KB, 8 ways, 128 byte lines, sect. (IA-64)
   1047       case 0x7F: l2 = 512; break;   // code and data L2 cache, 512 KB, 2 ways, 64 byte lines
   1048       case 0x80: l2 = 512; break;   // code and data L2 cache, 512 KB, 8 ways, 64 byte lines
   1049       case 0x81: l2 = 128; break;   // code and data L2 cache, 128 KB, 8 ways, 32 byte lines
   1050       case 0x82: l2 = 256; break;   // code and data L2 cache, 256 KB, 8 ways, 32 byte lines
   1051       case 0x83: l2 = 512; break;   // code and data L2 cache, 512 KB, 8 ways, 32 byte lines
   1052       case 0x84: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 8 ways, 32 byte lines
   1053       case 0x85: l2 = 2048; break;   // code and data L2 cache, 2048 KB, 8 ways, 32 byte lines
   1054       case 0x86: l2 = 512; break;   // code and data L2 cache, 512 KB, 4 ways, 64 byte lines
   1055       case 0x87: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines
   1056       case 0x88: l3 = 2048; break;   // code and data L3 cache, 2048 KB, 4 ways, 64 byte lines (IA-64)
   1057       case 0x89: l3 = 4096; break;   // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines (IA-64)
   1058       case 0x8A: l3 = 8192; break;   // code and data L3 cache, 8192 KB, 4 ways, 64 byte lines (IA-64)
   1059       case 0x8D: l3 = 3072; break;   // code and data L3 cache, 3072 KB, 12 ways, 128 byte lines (IA-64)
   1060 
   1061       default: break;
   1062     }
   1063   }
   1064   if(check_for_p2_core2 && l2 == l3)
   1065     l3 = 0;
   1066   l1 *= 1024;
   1067   l2 *= 1024;
   1068   l3 *= 1024;
   1069 }
   1070 
   1071 inline void queryCacheSizes_intel(int& l1, int& l2, int& l3, int max_std_funcs)
   1072 {
   1073   if(max_std_funcs>=4)
   1074     queryCacheSizes_intel_direct(l1,l2,l3);
   1075   else if(max_std_funcs>=2)
   1076     queryCacheSizes_intel_codes(l1,l2,l3);
   1077   else
   1078     l1 = l2 = l3 = 0;
   1079 }
   1080 
   1081 inline void queryCacheSizes_amd(int& l1, int& l2, int& l3)
   1082 {
   1083   int abcd[4];
   1084   abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
   1085   
   1086   // First query the max supported function.
   1087   EIGEN_CPUID(abcd,0x80000000,0);
   1088   if(static_cast<numext::uint32_t>(abcd[0]) >= static_cast<numext::uint32_t>(0x80000006))
   1089   {
   1090     EIGEN_CPUID(abcd,0x80000005,0);
   1091     l1 = (abcd[2] >> 24) * 1024; // C[31:24] = L1 size in KB
   1092     abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
   1093     EIGEN_CPUID(abcd,0x80000006,0);
   1094     l2 = (abcd[2] >> 16) * 1024; // C[31;16] = l2 cache size in KB
   1095     l3 = ((abcd[3] & 0xFFFC000) >> 18) * 512 * 1024; // D[31;18] = l3 cache size in 512KB
   1096   }
   1097   else
   1098   {
   1099     l1 = l2 = l3 = 0;
   1100   }
   1101 }
   1102 #endif
   1103 
   1104 /** \internal
   1105  * Queries and returns the cache sizes in Bytes of the L1, L2, and L3 data caches respectively */
   1106 inline void queryCacheSizes(int& l1, int& l2, int& l3)
   1107 {
   1108   #ifdef EIGEN_CPUID
   1109   int abcd[4];
   1110   const int GenuineIntel[] = {0x756e6547, 0x49656e69, 0x6c65746e};
   1111   const int AuthenticAMD[] = {0x68747541, 0x69746e65, 0x444d4163};
   1112   const int AMDisbetter_[] = {0x69444d41, 0x74656273, 0x21726574}; // "AMDisbetter!"
   1113 
   1114   // identify the CPU vendor
   1115   EIGEN_CPUID(abcd,0x0,0);
   1116   int max_std_funcs = abcd[0];
   1117   if(cpuid_is_vendor(abcd,GenuineIntel))
   1118     queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
   1119   else if(cpuid_is_vendor(abcd,AuthenticAMD) || cpuid_is_vendor(abcd,AMDisbetter_))
   1120     queryCacheSizes_amd(l1,l2,l3);
   1121   else
   1122     // by default let's use Intel's API
   1123     queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
   1124 
   1125   // here is the list of other vendors:
   1126 //   ||cpuid_is_vendor(abcd,"VIA VIA VIA ")
   1127 //   ||cpuid_is_vendor(abcd,"CyrixInstead")
   1128 //   ||cpuid_is_vendor(abcd,"CentaurHauls")
   1129 //   ||cpuid_is_vendor(abcd,"GenuineTMx86")
   1130 //   ||cpuid_is_vendor(abcd,"TransmetaCPU")
   1131 //   ||cpuid_is_vendor(abcd,"RiseRiseRise")
   1132 //   ||cpuid_is_vendor(abcd,"Geode by NSC")
   1133 //   ||cpuid_is_vendor(abcd,"SiS SiS SiS ")
   1134 //   ||cpuid_is_vendor(abcd,"UMC UMC UMC ")
   1135 //   ||cpuid_is_vendor(abcd,"NexGenDriven")
   1136   #else
   1137   l1 = l2 = l3 = -1;
   1138   #endif
   1139 }
   1140 
   1141 /** \internal
   1142  * \returns the size in Bytes of the L1 data cache */
   1143 inline int queryL1CacheSize()
   1144 {
   1145   int l1(-1), l2, l3;
   1146   queryCacheSizes(l1,l2,l3);
   1147   return l1;
   1148 }
   1149 
   1150 /** \internal
   1151  * \returns the size in Bytes of the L2 or L3 cache if this later is present */
   1152 inline int queryTopLevelCacheSize()
   1153 {
   1154   int l1, l2(-1), l3(-1);
   1155   queryCacheSizes(l1,l2,l3);
   1156   return (std::max)(l2,l3);
   1157 }
   1158 
   1159 } // end namespace internal
   1160 
   1161 } // end namespace Eigen
   1162 
   1163 #endif // EIGEN_MEMORY_H