libvo/aclib.c

   1 #include "config.h"
   2 #ifdef USE_FASTMEMCPY
   3
   4 /*
   5   aclib - advanced C library ;)
   6   This file contains functions which improve and expand standard C-library
   7   see aclib_template.c ... this file only contains runtime cpu detection and config options stuff
   8   runtime cpu detection by michael niedermayer (michaelni@gmx.at) is under GPL
   9 */
  10 #include <stddef.h>
  11 #include <string.h>
  12 #include "cpudetect.h"
  13 #include "fastmemcpy.h"
  14 #undef memcpy
  15
  16 #define BLOCK_SIZE 4096
  17 #define CONFUSION_FACTOR 0
  18 //Feel free to fine-tune the above 2, it might be possible to get some speedup with them :)
  19
  20 //#define STATISTICS
  21 #ifdef ARCH_X86
  22 #define CAN_COMPILE_X86_ASM
  23 #endif
  24
  25 //Note: we have MMX, MMX2, 3DNOW version there is no 3DNOW+MMX2 one
  26 //Plain C versions
  27 //#if !defined (HAVE_MMX) || defined (RUNTIME_CPUDETECT)
  28 //#define COMPILE_C
  29 //#endif
  30
  31 #ifdef CAN_COMPILE_X86_ASM
  32
  33 #if (defined (HAVE_MMX) && !defined (HAVE_3DNOW) && !defined (HAVE_MMX2)) || defined (RUNTIME_CPUDETECT)
  34 #define COMPILE_MMX
  35 #endif
  36
  37 #if (defined (HAVE_MMX2) && !defined (HAVE_SSE2)) || defined (RUNTIME_CPUDETECT)
  38 #define COMPILE_MMX2
  39 #endif
  40
  41 #if (defined (HAVE_3DNOW) && !defined (HAVE_MMX2)) || defined (RUNTIME_CPUDETECT)
  42 #define COMPILE_3DNOW
  43 #endif
  44
  45 #if defined (HAVE_SSE2) || defined (RUNTIME_CPUDETECT)
  46 #define COMPILE_SSE
  47 #endif
  48
  49 #undef HAVE_MMX
  50 #undef HAVE_MMX2
  51 #undef HAVE_3DNOW
  52 #undef HAVE_SSE
  53 #undef HAVE_SSE2
  54 /*
  55 #ifdef COMPILE_C
  56 #undef HAVE_MMX
  57 #undef HAVE_MMX2
  58 #undef HAVE_3DNOW
  59 #undef ARCH_X86
  60 #define RENAME(a) a ## _C
  61 #include "aclib_template.c"
  62 #endif
  63 */
  64 //MMX versions
  65 #ifdef COMPILE_MMX
  66 #undef RENAME
  67 #define HAVE_MMX
  68 #undef HAVE_MMX2
  69 #undef HAVE_3DNOW
  70 #undef HAVE_SSE
  71 #undef HAVE_SSE2
  72 #define RENAME(a) a ## _MMX
  73 #include "aclib_template.c"
  74 #endif
  75
  76 //MMX2 versions
  77 #ifdef COMPILE_MMX2
  78 #undef RENAME
  79 #define HAVE_MMX
  80 #define HAVE_MMX2
  81 #undef HAVE_3DNOW
  82 #undef HAVE_SSE
  83 #undef HAVE_SSE2
  84 #define RENAME(a) a ## _MMX2
  85 #include "aclib_template.c"
  86 #endif
  87
  88 //3DNOW versions
  89 #ifdef COMPILE_3DNOW
  90 #undef RENAME
  91 #define HAVE_MMX
  92 #undef HAVE_MMX2
  93 #define HAVE_3DNOW
  94 #undef HAVE_SSE
  95 #undef HAVE_SSE2
  96 #define RENAME(a) a ## _3DNow
  97 #include "aclib_template.c"
  98 #endif
  99
 100 //SSE versions (only used on SSE2 cpus)
 101 #ifdef COMPILE_SSE
 102 #undef RENAME
 103 #define HAVE_MMX
 104 #define HAVE_MMX2
 105 #undef HAVE_3DNOW
 106 #define HAVE_SSE
 107 #define HAVE_SSE2
 108 #define RENAME(a) a ## _SSE
 109 #include "aclib_template.c"
 110 #endif
 111
 112 #endif // CAN_COMPILE_X86_ASM
 113
 114
 115 void * fast_memcpy(void * to, const void * from, size_t len)
 116 {
 117 #ifdef RUNTIME_CPUDETECT
 118 #ifdef CAN_COMPILE_X86_ASM
 119         // ordered per speed fasterst first
 120         if(gCpuCaps.hasSSE2)
 121                 fast_memcpy_SSE(to, from, len);
 122         else if(gCpuCaps.hasMMX2)
 123                 fast_memcpy_MMX2(to, from, len);
 124         else if(gCpuCaps.has3DNow)
 125                 fast_memcpy_3DNow(to, from, len);
 126         else if(gCpuCaps.hasMMX)
 127                 fast_memcpy_MMX(to, from, len);
 128         else
 129 #endif //CAN_COMPILE_X86_ASM
 130                 memcpy(to, from, len); // prior to mmx we use the standart memcpy
 131 #else
 132 #ifdef HAVE_SSE2
 133                 fast_memcpy_SSE(to, from, len);
 134 #elif defined (HAVE_MMX2)
 135                 fast_memcpy_MMX2(to, from, len);
 136 #elif defined (HAVE_3DNOW)
 137                 fast_memcpy_3DNow(to, from, len);
 138 #elif defined (HAVE_MMX)
 139                 fast_memcpy_MMX(to, from, len);
 140 #else
 141                 memcpy(to, from, len); // prior to mmx we use the standart memcpy
 142 #endif
 143
 144 #endif //!RUNTIME_CPUDETECT
 145         return to;
 146 }
 147
 148 #undef  mem2agpcpy
 149 void * mem2agpcpy(void * to, const void * from, size_t len)
 150 {
 151 #ifdef RUNTIME_CPUDETECT
 152 #ifdef CAN_COMPILE_X86_ASM
 153         // ordered per speed fasterst first
 154         if(gCpuCaps.hasSSE2)
 155                 mem2agpcpy_SSE(to, from, len);
 156         else if(gCpuCaps.hasMMX2)
 157                 mem2agpcpy_MMX2(to, from, len);
 158         else if(gCpuCaps.has3DNow)
 159                 mem2agpcpy_3DNow(to, from, len);
 160         else if(gCpuCaps.hasMMX)
 161                 mem2agpcpy_MMX(to, from, len);
 162         else
 163 #endif //CAN_COMPILE_X86_ASM
 164                 memcpy(to, from, len); // prior to mmx we use the standart memcpy
 165 #else
 166 #ifdef HAVE_SSE2
 167                 mem2agpcpy_SSE(to, from, len);
 168 #elif defined (HAVE_MMX2)
 169                 mem2agpcpy_MMX2(to, from, len);
 170 #elif defined (HAVE_3DNOW)
 171                 mem2agpcpy_3DNow(to, from, len);
 172 #elif defined (HAVE_MMX)
 173                 mem2agpcpy_MMX(to, from, len);
 174 #else
 175                 memcpy(to, from, len); // prior to mmx we use the standart memcpy
 176 #endif
 177
 178 #endif //!RUNTIME_CPUDETECT
 179         return to;
 180 }
 181
 182 #endif /* use fastmemcpy */
 183