theora-old/lib/x86_32/dsp_mmxext.c

   1 /********************************************************************
   2  *                                                                  *
   3  * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
   4  * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
   5  * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
   6  * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
   7  *                                                                  *
   8  * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2003                *
   9  * by the Xiph.Org Foundation http://www.xiph.org/                  *
  10  *                                                                  *
  11  ********************************************************************
  12
  13   function:
  14   last mod: $Id$
  15
  16  ********************************************************************/
  17
  18 #include <stdlib.h>
  19
  20 #include "codec_internal.h"
  21 #include "dsp.h"
  22
  23 #if defined(USE_ASM)
  24
  25 #define SAD_MMXEXT_LOOP \
  26  "  movq (%1), %%mm0             \n\t"  /* take 8 bytes */ \
  27  "  movq (%2), %%mm1             \n\t" \
  28  "  psadbw %%mm1, %%mm0          \n\t" \
  29  "  add %3, %1                   \n\t"  /* Inc pointer into the new data */ \
  30  "  paddw %%mm0, %%mm7           \n\t"  /* accumulate difference... */ \
  31  "  add %4, %2                   \n\t"  /* Inc pointer into ref data */
  32
  33
  34 static ogg_uint32_t sad8x8__mmxext (unsigned char *ptr1, ogg_uint32_t stride1,
  35                             unsigned char *ptr2, ogg_uint32_t stride2)
  36 {
  37   ogg_uint32_t  DiffVal;
  38
  39   __asm__ __volatile__ (
  40     "  .p2align 4                   \n\t"
  41     "  pxor %%mm7, %%mm7            \n\t"       /* mm7 contains the result */
  42
  43     SAD_MMXEXT_LOOP
  44     SAD_MMXEXT_LOOP
  45     SAD_MMXEXT_LOOP
  46     SAD_MMXEXT_LOOP
  47     SAD_MMXEXT_LOOP
  48     SAD_MMXEXT_LOOP
  49     SAD_MMXEXT_LOOP
  50
  51     "  movq (%1), %%mm0             \n\t"       /* take 8 bytes */
  52     "  movq (%2), %%mm1             \n\t"
  53     "  psadbw %%mm1, %%mm0          \n\t"
  54     "  paddw %%mm0, %%mm7           \n\t"       /* accumulate difference... */
  55     "  movd %%mm7, %0               \n\t"
  56
  57      : "=r" (DiffVal),
  58        "+r" (ptr1),
  59        "+r" (ptr2)
  60      : "r" (stride1),
  61        "r" (stride2)
  62      : "memory"
  63   );
  64
  65   return DiffVal;
  66 }
  67
  68 #define SAD_TRES_LOOP \
  69   "  movq (%1), %%mm0             \n\t" /* take 8 bytes */ \
  70   "  movq (%2), %%mm1             \n\t" \
  71   "  psadbw %%mm1, %%mm0          \n\t" \
  72   "  add %3, %1                   \n\t" /* Inc pointer into the new data */ \
  73   "  paddw %%mm0, %%mm7           \n\t" /* accumulate difference... */ \
  74   "  add %4, %2                   \n\t" /* Inc pointer into ref data */
  75
  76
  77 static ogg_uint32_t sad8x8_thres__mmxext (unsigned char *ptr1, ogg_uint32_t stride1,
  78                                   unsigned char *ptr2, ogg_uint32_t stride2,
  79                                   ogg_uint32_t thres)
  80 {
  81   ogg_uint32_t  DiffVal;
  82
  83   __asm__ __volatile__ (
  84     "  .p2align 4                   \n\t"
  85     "  pxor %%mm7, %%mm7            \n\t"       /* mm7 contains the result */
  86
  87     SAD_TRES_LOOP
  88     SAD_TRES_LOOP
  89     SAD_TRES_LOOP
  90     SAD_TRES_LOOP
  91     SAD_TRES_LOOP
  92     SAD_TRES_LOOP
  93     SAD_TRES_LOOP
  94     SAD_TRES_LOOP
  95
  96     "  movd %%mm7, %0               \n\t"
  97
  98      : "=r" (DiffVal),
  99        "+r" (ptr1),
 100        "+r" (ptr2)
 101      : "r" (stride1),
 102        "r" (stride2)
 103      : "memory"
 104   );
 105
 106   return DiffVal;
 107 }
 108
 109 #define SAD_XY2_TRES \
 110   "  movq (%1), %%mm0             \n\t" /* take 8 bytes */ \
 111   "  movq (%2), %%mm1             \n\t" \
 112   "  movq (%3), %%mm2             \n\t" \
 113   "  pavgb %%mm2, %%mm1           \n\t" \
 114   "  psadbw %%mm1, %%mm0          \n\t" \
 115  \
 116   "  add %4, %1                   \n\t" /* Inc pointer into the new data */ \
 117   "  paddw %%mm0, %%mm7           \n\t" /* accumulate difference... */ \
 118   "  add %5, %2                   \n\t" /* Inc pointer into ref data */ \
 119   "  add %5, %3                   \n\t" /* Inc pointer into ref data */
 120
 121
 122 static ogg_uint32_t sad8x8_xy2_thres__mmxext (unsigned char *SrcData, ogg_uint32_t SrcStride,
 123                                       unsigned char *RefDataPtr1,
 124                                       unsigned char *RefDataPtr2, ogg_uint32_t RefStride,
 125                                       ogg_uint32_t thres)
 126 {
 127   ogg_uint32_t  DiffVal;
 128
 129   __asm__ __volatile__ (
 130     "  .p2align 4                   \n\t"
 131     "  pxor %%mm7, %%mm7            \n\t"       /* mm7 contains the result */
 132     SAD_XY2_TRES
 133     SAD_XY2_TRES
 134     SAD_XY2_TRES
 135     SAD_XY2_TRES
 136     SAD_XY2_TRES
 137     SAD_XY2_TRES
 138     SAD_XY2_TRES
 139     SAD_XY2_TRES
 140
 141     "  movd %%mm7, %0               \n\t"
 142      : "=m" (DiffVal),
 143        "+r" (SrcData),
 144        "+r" (RefDataPtr1),
 145        "+r" (RefDataPtr2)
 146      : "m" (SrcStride),
 147        "m" (RefStride)
 148      : "memory"
 149   );
 150
 151   return DiffVal;
 152 }
 153
 154 static ogg_uint32_t row_sad8__mmxext (unsigned char *Src1, unsigned char *Src2)
 155 {
 156   ogg_uint32_t MaxSad;
 157
 158   __asm__ __volatile__ (
 159     "  .p2align 4                   \n\t"
 160
 161     "  movd        (%1), %%mm0      \n\t"
 162     "  movd        (%2), %%mm1      \n\t"
 163     "  psadbw      %%mm0, %%mm1     \n\t"
 164     "  movd        4(%1), %%mm2     \n\t"
 165     "  movd        4(%2), %%mm3     \n\t"
 166     "  psadbw      %%mm2, %%mm3     \n\t"
 167
 168     "  pmaxsw      %%mm1, %%mm3     \n\t"
 169     "  movd        %%mm3, %0        \n\t"
 170     "  andl        $0xffff, %0      \n\t"
 171
 172      : "=m" (MaxSad),
 173        "+r" (Src1),
 174        "+r" (Src2)
 175      :
 176      : "memory"
 177   );
 178
 179   return MaxSad;
 180 }
 181
 182 static ogg_uint32_t col_sad8x8__mmxext (unsigned char *Src1, unsigned char *Src2,
 183                                     ogg_uint32_t stride)
 184 {
 185   ogg_uint32_t MaxSad;
 186
 187   __asm__ __volatile__ (
 188     "  .p2align 4                   \n\t"
 189
 190     "  pxor        %%mm3, %%mm3     \n\t"       /* zero out mm3 for unpack */
 191     "  pxor        %%mm4, %%mm4     \n\t"       /* mm4 low sum */
 192     "  pxor        %%mm5, %%mm5     \n\t"       /* mm5 high sum */
 193     "  pxor        %%mm6, %%mm6     \n\t"       /* mm6 low sum */
 194     "  pxor        %%mm7, %%mm7     \n\t"       /* mm7 high sum */
 195     "  mov         $4, %%edi        \n\t"       /* 4 rows */
 196     "1:                             \n\t"
 197     "  movq        (%1), %%mm0      \n\t"       /* take 8 bytes */
 198     "  movq        (%2), %%mm1      \n\t"       /* take 8 bytes */
 199
 200     "  movq        %%mm0, %%mm2     \n\t"
 201     "  psubusb     %%mm1, %%mm0     \n\t"       /* A - B */
 202     "  psubusb     %%mm2, %%mm1     \n\t"       /* B - A */
 203     "  por         %%mm1, %%mm0     \n\t"       /* and or gives abs difference */
 204     "  movq        %%mm0, %%mm1     \n\t"
 205
 206     "  punpcklbw   %%mm3, %%mm0     \n\t"       /* unpack to higher precision for accumulation */
 207     "  paddw       %%mm0, %%mm4     \n\t"       /* accumulate difference... */
 208     "  punpckhbw   %%mm3, %%mm1     \n\t"       /* unpack high four bytes to higher precision */
 209     "  paddw       %%mm1, %%mm5     \n\t"       /* accumulate difference... */
 210     "  add         %3, %1           \n\t"       /* Inc pointer into the new data */
 211     "  add         %3, %2           \n\t"       /* Inc pointer into the new data */
 212
 213     "  dec         %%edi            \n\t"
 214     "  jnz 1b                       \n\t"
 215
 216     "  mov         $4, %%edi        \n\t"       /* 4 rows */
 217     "2:                             \n\t"
 218     "  movq        (%1), %%mm0      \n\t"       /* take 8 bytes */
 219     "  movq        (%2), %%mm1      \n\t"       /* take 8 bytes */
 220
 221     "  movq        %%mm0, %%mm2     \n\t"
 222     "  psubusb     %%mm1, %%mm0     \n\t"       /* A - B */
 223     "  psubusb     %%mm2, %%mm1     \n\t"       /* B - A */
 224     "  por         %%mm1, %%mm0     \n\t"       /* and or gives abs difference */
 225     "  movq        %%mm0, %%mm1     \n\t"
 226
 227     "  punpcklbw   %%mm3, %%mm0     \n\t"       /* unpack to higher precision for accumulation */
 228     "  paddw       %%mm0, %%mm6     \n\t"       /* accumulate difference... */
 229     "  punpckhbw   %%mm3, %%mm1     \n\t"       /* unpack high four bytes to higher precision */
 230     "  paddw       %%mm1, %%mm7     \n\t"       /* accumulate difference... */
 231     "  add         %3, %1           \n\t"       /* Inc pointer into the new data */
 232     "  add         %3, %2           \n\t"       /* Inc pointer into the new data */
 233
 234     "  dec         %%edi            \n\t"
 235     "  jnz 2b                       \n\t"
 236
 237     "  pmaxsw      %%mm6, %%mm7     \n\t"
 238     "  pmaxsw      %%mm4, %%mm5     \n\t"
 239     "  pmaxsw      %%mm5, %%mm7     \n\t"
 240     "  movq        %%mm7, %%mm6     \n\t"
 241     "  psrlq       $32, %%mm6       \n\t"
 242     "  pmaxsw      %%mm6, %%mm7     \n\t"
 243     "  movq        %%mm7, %%mm6     \n\t"
 244     "  psrlq       $16, %%mm6       \n\t"
 245     "  pmaxsw      %%mm6, %%mm7     \n\t"
 246     "  movd        %%mm7, %0        \n\t"
 247     "  andl        $0xffff, %0      \n\t"
 248
 249      : "=r" (MaxSad),
 250        "+r" (Src1),
 251        "+r" (Src2)
 252      : "r" (stride)
 253      : "memory", "edi"
 254   );
 255
 256   return MaxSad;
 257 }
 258
 259 static ogg_uint32_t inter8x8_err_xy2__mmxext (unsigned char *SrcData, ogg_uint32_t SrcStride,
 260                                      unsigned char *RefDataPtr1,
 261                                      unsigned char *RefDataPtr2, ogg_uint32_t RefStride)
 262 {
 263   ogg_uint32_t XSum;
 264   ogg_uint32_t XXSum;
 265
 266   __asm__ __volatile__ (
 267     "  .p2align 4                   \n\t"
 268
 269     "  pxor        %%mm4, %%mm4     \n\t"
 270     "  pxor        %%mm5, %%mm5     \n\t"
 271     "  pxor        %%mm6, %%mm6     \n\t"
 272     "  pxor        %%mm7, %%mm7     \n\t"
 273     "  mov         $8, %%edi        \n\t"
 274     "1:                             \n\t"
 275     "  movq        (%2), %%mm0      \n\t"       /* take 8 bytes */
 276
 277     "  movq        (%3), %%mm2      \n\t"
 278     "  movq        (%4), %%mm1      \n\t"       /* take average of mm2 and mm1 */
 279     "  pavgb       %%mm2, %%mm1     \n\t"
 280
 281     "  movq        %%mm0, %%mm2     \n\t"
 282     "  movq        %%mm1, %%mm3     \n\t"
 283
 284     "  punpcklbw   %%mm6, %%mm0     \n\t"
 285     "  punpcklbw   %%mm4, %%mm1     \n\t"
 286     "  punpckhbw   %%mm6, %%mm2     \n\t"
 287     "  punpckhbw   %%mm4, %%mm3     \n\t"
 288
 289     "  psubsw      %%mm1, %%mm0     \n\t"
 290     "  psubsw      %%mm3, %%mm2     \n\t"
 291
 292     "  paddw       %%mm0, %%mm5     \n\t"
 293     "  paddw       %%mm2, %%mm5     \n\t"
 294
 295     "  pmaddwd     %%mm0, %%mm0     \n\t"
 296     "  pmaddwd     %%mm2, %%mm2     \n\t"
 297
 298     "  paddd       %%mm0, %%mm7     \n\t"
 299     "  paddd       %%mm2, %%mm7     \n\t"
 300
 301     "  add         %5, %2           \n\t"       /* Inc pointer into src data */
 302     "  add         %6, %3           \n\t"       /* Inc pointer into ref data */
 303     "  add         %6, %4           \n\t"       /* Inc pointer into ref data */
 304
 305     "  dec         %%edi            \n\t"
 306     "  jnz 1b                       \n\t"
 307
 308     "  movq        %%mm5, %%mm0     \n\t"
 309     "  psrlq       $32, %%mm5       \n\t"
 310     "  paddw       %%mm0, %%mm5     \n\t"
 311     "  movq        %%mm5, %%mm0     \n\t"
 312     "  psrlq       $16, %%mm5       \n\t"
 313     "  paddw       %%mm0, %%mm5     \n\t"
 314     "  movd        %%mm5, %%edi     \n\t"
 315     "  movsx       %%di, %%edi      \n\t"
 316     "  movl        %%edi, %0        \n\t"
 317
 318     "  movq        %%mm7, %%mm0     \n\t"
 319     "  psrlq       $32, %%mm7       \n\t"
 320     "  paddd       %%mm0, %%mm7     \n\t"
 321     "  movd        %%mm7, %1        \n\t"
 322
 323      : "=m" (XSum),
 324        "=m" (XXSum),
 325        "+r" (SrcData),
 326        "+r" (RefDataPtr1),
 327        "+r" (RefDataPtr2)
 328      : "m" (SrcStride),
 329        "m" (RefStride)
 330      : "edi", "memory"
 331   );
 332
 333   /* Compute and return population variance as mis-match metric. */
 334   return (( (XXSum<<6) - XSum*XSum ));
 335 }
 336
 337 void dsp_mmxext_init(DspFunctions *funcs)
 338 {
 339   TH_DEBUG("enabling accelerated x86_32 mmxext dsp functions.\n");
 340   funcs->row_sad8 = row_sad8__mmxext;
 341   funcs->col_sad8x8 = col_sad8x8__mmxext;
 342   funcs->sad8x8 = sad8x8__mmxext;
 343   funcs->sad8x8_thres = sad8x8_thres__mmxext;
 344   funcs->sad8x8_xy2_thres = sad8x8_xy2_thres__mmxext;
 345   funcs->inter8x8_err_xy2 = inter8x8_err_xy2__mmxext;
 346 }
 347
 348 #endif /* USE_ASM */