benchmarks/malloc_aligned.hpp

   1 //  functions for aligned memory allocation
   2 //  Copyright (C) 2009 Tim Blechmann
   3 //
   4 //  This program is free software; you can redistribute it and/or modify
   5 //  it under the terms of the GNU General Public License as published by
   6 //  the Free Software Foundation; either version 2 of the License, or
   7 //  (at your option) any later version.
   8 //
   9 //  This program is distributed in the hope that it will be useful,
  10 //  but WITHOUT ANY WARRANTY; without even the implied warranty of
  11 //  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12 //  GNU General Public License for more details.
  13 //
  14 //  You should have received a copy of the GNU General Public License
  15 //  along with this program; see the file COPYING.  If not, write to
  16 //  the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  17 //  Boston, MA 02111-1307, USA.
  18
  19 #ifndef UTILITIES_MALLOC_ALIGNED_HPP
  20 #define UTILITIES_MALLOC_ALIGNED_HPP
  21
  22 #include <cstdlib>
  23 #include <cstring>
  24
  25 #include <boost/noncopyable.hpp>
  26
  27 #ifdef __SSE2__
  28 #include <xmmintrin.h>
  29 #elif defined(HAVE_TBB)
  30 #include <tbb/cache_aligned_allocator.h>
  31 #endif /* HAVE_TBB */
  32
  33
  34 namespace nova
  35 {
  36
  37 #if _POSIX_C_SOURCE >= 200112L || _XOPEN_SOURCE >= 600
  38 /* we have posix_memalign */
  39
  40 /* memory alignment constraints:
  41  *
  42  * - 16 byte for SSE operations
  43  * - the cache lines size of modern x86 cpus is 64 bytes (pentium-m, pentium 4, core, k8)
  44  */
  45 const int malloc_memory_alignment = 64;
  46
  47 inline void* malloc_aligned(std::size_t nbytes)
  48 {
  49     void * ret;
  50     int status = posix_memalign(&ret, malloc_memory_alignment, nbytes);
  51     if (!status)
  52         return ret;
  53     else
  54         return 0;
  55 }
  56
  57 inline void free_aligned(void *ptr)
  58 {
  59     free(ptr);
  60 }
  61
  62 #elif defined(__APPLE__)
  63
  64 const int malloc_memory_alignment = 64;
  65
  66 /* apple's malloc implementation returns 16-byte aligned chunks */
  67 inline void* malloc_aligned(std::size_t nbytes)
  68 {
  69     return malloc(nbytes);
  70 }
  71
  72 inline void free_aligned(void *ptr)
  73 {
  74     free(ptr);
  75 }
  76
  77
  78 #elif defined(__SSE2__)
  79
  80 const int malloc_memory_alignment = 64;
  81
  82 inline void* malloc_aligned(std::size_t nbytes)
  83 {
  84     return _mm_malloc(nbytes, malloc_memory_alignment);
  85 }
  86
  87 inline void free_aligned(void *ptr)
  88 {
  89     _mm_free(ptr);
  90 }
  91
  92 #elif defined(HAVE_TBB)
  93
  94 inline void* malloc_aligned(std::size_t nbytes)
  95 {
  96     tbb::cache_aligned_allocator<void*> ca_alloc;
  97     return static_cast<void*>(ca_alloc.allocate(nbytes));
  98 }
  99
 100 inline void free_aligned(void *ptr)
 101 {
 102     tbb::cache_aligned_allocator<void*> ca_alloc;
 103     ca_alloc.deallocate(static_cast<void**>(ptr), 0);
 104 }
 105
 106 #else
 107
 108 /* on other systems, we use the aligned memory allocation taken
 109  * from thomas grill's implementation for pd */
 110 #define VECTORALIGNMENT 128
 111 inline void* malloc_aligned(std::size_t nbytes)
 112 {
 113     void* vec = malloc(nbytes+ (VECTORALIGNMENT/8-1) + sizeof(void *));
 114
 115     if (vec != NULL)
 116     {
 117         /* get alignment of first possible signal vector byte */
 118         long alignment = ((long)vec+sizeof(void *))&(VECTORALIGNMENT/8-1);
 119         /* calculate aligned pointer */
 120         void *ret = (unsigned char *)vec+sizeof(void *)+(alignment == 0?0:VECTORALIGNMENT/8-alignment);
 121         /* save original memory location */
 122         *(void **)((unsigned char *)ret-sizeof(void *)) = vec;
 123         return ret;
 124     }
 125     else
 126         return 0;
 127 }
 128
 129 inline void free_aligned(void *ptr)
 130 {
 131     /* get original memory location */
 132     void *ori = *(void **)((unsigned char *)ptr-sizeof(void *));
 133     free(ori);
 134 }
 135
 136 #undef VECTORALIGNMENT
 137
 138 #endif
 139
 140 inline void * calloc_aligned(std::size_t nbytes)
 141 {
 142     void * ret = malloc_aligned(nbytes);
 143     if (ret)
 144         std::memset(ret, 0, nbytes);
 145     return ret;
 146 }
 147
 148 template <typename T>
 149 T* malloc_aligned(std::size_t n)
 150 {
 151     return static_cast<T*>(malloc_aligned(n * sizeof(T)));
 152 }
 153
 154 template <typename T>
 155 T* calloc_aligned(std::size_t n)
 156 {
 157     return static_cast<T*>(calloc_aligned(n * sizeof(T)));
 158 }
 159
 160
 161 /** aligned allocator. uses malloc_aligned and free_aligned internally
 162  *  */
 163 template <class T>
 164 class aligned_allocator
 165 {
 166 public:
 167     typedef std::size_t size_type;
 168     typedef std::ptrdiff_t difference_type;
 169     typedef T*        pointer;
 170     typedef const T*  const_pointer;
 171     typedef T&        reference;
 172     typedef const T&  const_reference;
 173     typedef T         value_type;
 174
 175     template <class U> struct rebind
 176     {
 177         typedef aligned_allocator<U> other;
 178     };
 179
 180     pointer address(reference x) const
 181     {
 182         return &x;
 183     }
 184
 185     const_pointer address(const_reference x) const
 186     {
 187         return &x;
 188     }
 189
 190     pointer allocate(size_type n,
 191                      const_pointer hint = 0)
 192     {
 193         pointer ret = malloc_aligned<T>(n);
 194         if (ret == 0)
 195             throw std::bad_alloc();
 196         return ret;
 197     }
 198
 199     void deallocate(pointer p, size_type n)
 200     {
 201         return free_aligned(p);
 202     }
 203
 204     size_type max_size() const throw()
 205     {
 206         return size_type(-1) / sizeof(T);
 207     }
 208
 209     void construct(pointer p, const T& val)
 210     {
 211         ::new(p) T(val);
 212     }
 213
 214     void destroy(pointer p)
 215     {
 216         p->~T();
 217     }
 218 };
 219
 220
 221 template<typename T, typename U>
 222 bool operator==( aligned_allocator<T> const& left, aligned_allocator<U> const& right )
 223 {
 224     return !(left != right);
 225 }
 226
 227 template<typename T, typename U>
 228 bool operator!=( aligned_allocator<T> const& left, aligned_allocator<U> const& right )
 229 {
 230     return true;
 231 }
 232
 233
 234 /** smart-pointer, freeing the managed pointer via free_aligned */
 235 template<class T, bool managed = true>
 236 class aligned_storage_ptr
 237 {
 238 public:
 239     explicit aligned_storage_ptr(T * p = 0):
 240         ptr(p)
 241     {}
 242
 243     ~aligned_storage_ptr(void)
 244     {
 245         if (managed && ptr)
 246             free_aligned(ptr);
 247     }
 248
 249     void reset(T * p = 0)
 250     {
 251         if (managed && ptr)
 252             free_aligned(ptr);
 253         ptr = p;
 254     }
 255
 256     T & operator*() const
 257     {
 258         return *ptr;
 259     }
 260
 261     T * operator->() const
 262     {
 263         return ptr;
 264     }
 265
 266     T * get() const
 267     {
 268         return ptr;
 269     }
 270
 271     aligned_storage_ptr & operator=(T * p)
 272     {
 273         reset(p);
 274         return *this;
 275     }
 276
 277     operator bool() const
 278     {
 279         return bool(ptr);
 280     }
 281
 282     void swap(aligned_storage_ptr & b)
 283     {
 284         T * p = ptr;
 285         ptr = b.ptr;
 286         b.ptr = p;
 287     }
 288
 289 private:
 290     T * ptr;
 291 };
 292
 293 } /* namespace nova */
 294
 295 #endif /* UTILITIES_MALLOC_ALIGNED_HPP */