libavcodec/i386/motion_est_mmx.c

   1 /*
   2  * MMX optimized motion estimation
   3  * Copyright (c) 2001 Fabrice Bellard.
   4  *
   5  * This library is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU Lesser General Public
   7  * License as published by the Free Software Foundation; either
   8  * version 2 of the License, or (at your option) any later version.
   9  *
  10  * This library is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * Lesser General Public License for more details.
  14  *
  15  * You should have received a copy of the GNU Lesser General Public
  16  * License along with this library; if not, write to the Free Software
  17  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  18  *
  19  * mostly by Michael Niedermayer <michaelni@gmx.at>
  20  */
  21 #include "../dsputil.h"
  22
  23 static const __attribute__ ((aligned(8))) uint64_t round_tab[3]={
  24 0x0000000000000000ULL,
  25 0x0001000100010001ULL,
  26 0x0002000200020002ULL,
  27 };
  28
  29 static __attribute__ ((aligned(8), unused)) uint64_t bone= 0x0101010101010101LL;
  30
  31 static inline void sad8_1_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
  32 {
  33     int len= -(stride*h);
  34     asm volatile(
  35         ".balign 16                     \n\t"
  36         "1:                             \n\t"
  37         "movq (%1, %%eax), %%mm0        \n\t"
  38         "movq (%2, %%eax), %%mm2        \n\t"
  39         "movq (%2, %%eax), %%mm4        \n\t"
  40         "addl %3, %%eax                 \n\t"
  41         "psubusb %%mm0, %%mm2           \n\t"
  42         "psubusb %%mm4, %%mm0           \n\t"
  43         "movq (%1, %%eax), %%mm1        \n\t"
  44         "movq (%2, %%eax), %%mm3        \n\t"
  45         "movq (%2, %%eax), %%mm5        \n\t"
  46         "psubusb %%mm1, %%mm3           \n\t"
  47         "psubusb %%mm5, %%mm1           \n\t"
  48         "por %%mm2, %%mm0               \n\t"
  49         "por %%mm1, %%mm3               \n\t"
  50         "movq %%mm0, %%mm1              \n\t"
  51         "movq %%mm3, %%mm2              \n\t"
  52         "punpcklbw %%mm7, %%mm0         \n\t"
  53         "punpckhbw %%mm7, %%mm1         \n\t"
  54         "punpcklbw %%mm7, %%mm3         \n\t"
  55         "punpckhbw %%mm7, %%mm2         \n\t"
  56         "paddw %%mm1, %%mm0             \n\t"
  57         "paddw %%mm3, %%mm2             \n\t"
  58         "paddw %%mm2, %%mm0             \n\t"
  59         "paddw %%mm0, %%mm6             \n\t"
  60         "addl %3, %%eax                 \n\t"
  61         " js 1b                         \n\t"
  62         : "+a" (len)
  63         : "r" (blk1 - len), "r" (blk2 - len), "r" (stride)
  64     );
  65 }
  66
  67 static inline void sad8_1_mmx2(uint8_t *blk1, uint8_t *blk2, int stride, int h)
  68 {
  69     int len= -(stride*h);
  70     asm volatile(
  71         ".balign 16                     \n\t"
  72         "1:                             \n\t"
  73         "movq (%1, %%eax), %%mm0        \n\t"
  74         "movq (%2, %%eax), %%mm2        \n\t"
  75         "psadbw %%mm2, %%mm0            \n\t"
  76         "addl %3, %%eax                 \n\t"
  77         "movq (%1, %%eax), %%mm1        \n\t"
  78         "movq (%2, %%eax), %%mm3        \n\t"
  79         "psadbw %%mm1, %%mm3            \n\t"
  80         "paddw %%mm3, %%mm0             \n\t"
  81         "paddw %%mm0, %%mm6             \n\t"
  82         "addl %3, %%eax                 \n\t"
  83         " js 1b                         \n\t"
  84         : "+a" (len)
  85         : "r" (blk1 - len), "r" (blk2 - len), "r" (stride)
  86     );
  87 }
  88
  89 static inline void sad8_2_mmx2(uint8_t *blk1a, uint8_t *blk1b, uint8_t *blk2, int stride, int h)
  90 {
  91     int len= -(stride*h);
  92     asm volatile(
  93         ".balign 16                     \n\t"
  94         "1:                             \n\t"
  95         "movq (%1, %%eax), %%mm0        \n\t"
  96         "movq (%2, %%eax), %%mm2        \n\t"
  97         "pavgb %%mm2, %%mm0             \n\t"
  98         "movq (%3, %%eax), %%mm2        \n\t"
  99         "psadbw %%mm2, %%mm0            \n\t"
 100         "addl %4, %%eax                 \n\t"
 101         "movq (%1, %%eax), %%mm1        \n\t"
 102         "movq (%2, %%eax), %%mm3        \n\t"
 103         "pavgb %%mm1, %%mm3             \n\t"
 104         "movq (%3, %%eax), %%mm1        \n\t"
 105         "psadbw %%mm1, %%mm3            \n\t"
 106         "paddw %%mm3, %%mm0             \n\t"
 107         "paddw %%mm0, %%mm6             \n\t"
 108         "addl %4, %%eax                 \n\t"
 109         " js 1b                         \n\t"
 110         : "+a" (len)
 111         : "r" (blk1a - len), "r" (blk1b -len), "r" (blk2 - len), "r" (stride)
 112     );
 113 }
 114
 115 static inline void sad8_4_mmx2(uint8_t *blk1, uint8_t *blk2, int stride, int h)
 116 { //FIXME reuse src
 117     int len= -(stride*h);
 118     asm volatile(
 119         ".balign 16                     \n\t"
 120         "movq "MANGLE(bone)", %%mm5     \n\t"
 121         "1:                             \n\t"
 122         "movq (%1, %%eax), %%mm0        \n\t"
 123         "movq (%2, %%eax), %%mm2        \n\t"
 124         "movq 1(%1, %%eax), %%mm1       \n\t"
 125         "movq 1(%2, %%eax), %%mm3       \n\t"
 126         "pavgb %%mm2, %%mm0             \n\t"
 127         "pavgb %%mm1, %%mm3             \n\t"
 128         "psubusb %%mm5, %%mm3           \n\t"
 129         "pavgb %%mm3, %%mm0             \n\t"
 130         "movq (%3, %%eax), %%mm2        \n\t"
 131         "psadbw %%mm2, %%mm0            \n\t"
 132         "addl %4, %%eax                 \n\t"
 133         "movq (%1, %%eax), %%mm1        \n\t"
 134         "movq (%2, %%eax), %%mm3        \n\t"
 135         "movq 1(%1, %%eax), %%mm2       \n\t"
 136         "movq 1(%2, %%eax), %%mm4       \n\t"
 137         "pavgb %%mm3, %%mm1             \n\t"
 138         "pavgb %%mm4, %%mm2             \n\t"
 139         "psubusb %%mm5, %%mm2           \n\t"
 140         "pavgb %%mm1, %%mm2             \n\t"
 141         "movq (%3, %%eax), %%mm1        \n\t"
 142         "psadbw %%mm1, %%mm2            \n\t"
 143         "paddw %%mm2, %%mm0             \n\t"
 144         "paddw %%mm0, %%mm6             \n\t"
 145         "addl %4, %%eax                 \n\t"
 146         " js 1b                         \n\t"
 147         : "+a" (len)
 148         : "r" (blk1 - len), "r" (blk1 - len + stride), "r" (blk2 - len), "r" (stride)
 149     );
 150 }
 151
 152 static inline void sad8_2_mmx(uint8_t *blk1a, uint8_t *blk1b, uint8_t *blk2, int stride, int h)
 153 {
 154     int len= -(stride*h);
 155     asm volatile(
 156         ".balign 16                     \n\t"
 157         "1:                             \n\t"
 158         "movq (%1, %%eax), %%mm0        \n\t"
 159         "movq (%2, %%eax), %%mm1        \n\t"
 160         "movq (%1, %%eax), %%mm2        \n\t"
 161         "movq (%2, %%eax), %%mm3        \n\t"
 162         "punpcklbw %%mm7, %%mm0         \n\t"
 163         "punpcklbw %%mm7, %%mm1         \n\t"
 164         "punpckhbw %%mm7, %%mm2         \n\t"
 165         "punpckhbw %%mm7, %%mm3         \n\t"
 166         "paddw %%mm0, %%mm1             \n\t"
 167         "paddw %%mm2, %%mm3             \n\t"
 168         "movq (%3, %%eax), %%mm4        \n\t"
 169         "movq (%3, %%eax), %%mm2        \n\t"
 170         "paddw %%mm5, %%mm1             \n\t"
 171         "paddw %%mm5, %%mm3             \n\t"
 172         "psrlw $1, %%mm1                \n\t"
 173         "psrlw $1, %%mm3                \n\t"
 174         "packuswb %%mm3, %%mm1          \n\t"
 175         "psubusb %%mm1, %%mm4           \n\t"
 176         "psubusb %%mm2, %%mm1           \n\t"
 177         "por %%mm4, %%mm1               \n\t"
 178         "movq %%mm1, %%mm0              \n\t"
 179         "punpcklbw %%mm7, %%mm0         \n\t"
 180         "punpckhbw %%mm7, %%mm1         \n\t"
 181         "paddw %%mm1, %%mm0             \n\t"
 182         "paddw %%mm0, %%mm6             \n\t"
 183         "addl %4, %%eax                 \n\t"
 184         " js 1b                         \n\t"
 185         : "+a" (len)
 186         : "r" (blk1a - len), "r" (blk1b -len), "r" (blk2 - len), "r" (stride)
 187     );
 188 }
 189
 190 static inline void sad8_4_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
 191 {
 192     int len= -(stride*h);
 193     asm volatile(
 194         ".balign 16                     \n\t"
 195         "1:                             \n\t"
 196         "movq (%1, %%eax), %%mm0        \n\t"
 197         "movq (%2, %%eax), %%mm1        \n\t"
 198         "movq %%mm0, %%mm4              \n\t"
 199         "movq %%mm1, %%mm2              \n\t"
 200         "punpcklbw %%mm7, %%mm0         \n\t"
 201         "punpcklbw %%mm7, %%mm1         \n\t"
 202         "punpckhbw %%mm7, %%mm4         \n\t"
 203         "punpckhbw %%mm7, %%mm2         \n\t"
 204         "paddw %%mm1, %%mm0             \n\t"
 205         "paddw %%mm2, %%mm4             \n\t"
 206         "movq 1(%1, %%eax), %%mm2       \n\t"
 207         "movq 1(%2, %%eax), %%mm3       \n\t"
 208         "movq %%mm2, %%mm1              \n\t"
 209         "punpcklbw %%mm7, %%mm2         \n\t"
 210         "punpckhbw %%mm7, %%mm1         \n\t"
 211         "paddw %%mm0, %%mm2             \n\t"
 212         "paddw %%mm4, %%mm1             \n\t"
 213         "movq %%mm3, %%mm4              \n\t"
 214         "punpcklbw %%mm7, %%mm3         \n\t"
 215         "punpckhbw %%mm7, %%mm4         \n\t"
 216         "paddw %%mm3, %%mm2             \n\t"
 217         "paddw %%mm4, %%mm1             \n\t"
 218         "movq (%3, %%eax), %%mm3        \n\t"
 219         "movq (%3, %%eax), %%mm4        \n\t"
 220         "paddw %%mm5, %%mm2             \n\t"
 221         "paddw %%mm5, %%mm1             \n\t"
 222         "psrlw $2, %%mm2                \n\t"
 223         "psrlw $2, %%mm1                \n\t"
 224         "packuswb %%mm1, %%mm2          \n\t"
 225         "psubusb %%mm2, %%mm3           \n\t"
 226         "psubusb %%mm4, %%mm2           \n\t"
 227         "por %%mm3, %%mm2               \n\t"
 228         "movq %%mm2, %%mm0              \n\t"
 229         "punpcklbw %%mm7, %%mm0         \n\t"
 230         "punpckhbw %%mm7, %%mm2         \n\t"
 231         "paddw %%mm2, %%mm0             \n\t"
 232         "paddw %%mm0, %%mm6             \n\t"
 233         "addl %4, %%eax                 \n\t"
 234         " js 1b                         \n\t"
 235         : "+a" (len)
 236         : "r" (blk1 - len), "r" (blk1 -len + stride), "r" (blk2 - len), "r" (stride)
 237     );
 238 }
 239
 240 static inline int sum_mmx(void)
 241 {
 242     int ret;
 243     asm volatile(
 244         "movq %%mm6, %%mm0              \n\t"
 245         "psrlq $32, %%mm6               \n\t"
 246         "paddw %%mm0, %%mm6             \n\t"
 247         "movq %%mm6, %%mm0              \n\t"
 248         "psrlq $16, %%mm6               \n\t"
 249         "paddw %%mm0, %%mm6             \n\t"
 250         "movd %%mm6, %0                 \n\t"
 251         : "=r" (ret)
 252     );
 253     return ret&0xFFFF;
 254 }
 255
 256 static inline int sum_mmx2(void)
 257 {
 258     int ret;
 259     asm volatile(
 260         "movd %%mm6, %0                 \n\t"
 261         : "=r" (ret)
 262     );
 263     return ret;
 264 }
 265
 266
 267 #define PIX_SAD(suf)\
 268 static int sad8_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
 269 {\
 270     assert(h==8);\
 271     asm volatile("pxor %%mm7, %%mm7             \n\t"\
 272                  "pxor %%mm6, %%mm6             \n\t":);\
 273 \
 274     sad8_1_ ## suf(blk1, blk2, stride, 8);\
 275 \
 276     return sum_ ## suf();\
 277 }\
 278 static int sad8_x2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
 279 {\
 280     assert(h==8);\
 281     asm volatile("pxor %%mm7, %%mm7             \n\t"\
 282                  "pxor %%mm6, %%mm6             \n\t"\
 283                  "movq %0, %%mm5                \n\t"\
 284                  :: "m"(round_tab[1]) \
 285                  );\
 286 \
 287     sad8_2_ ## suf(blk1, blk1+1, blk2, stride, 8);\
 288 \
 289     return sum_ ## suf();\
 290 }\
 291 \
 292 static int sad8_y2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
 293 {\
 294     assert(h==8);\
 295     asm volatile("pxor %%mm7, %%mm7             \n\t"\
 296                  "pxor %%mm6, %%mm6             \n\t"\
 297                  "movq %0, %%mm5                \n\t"\
 298                  :: "m"(round_tab[1]) \
 299                  );\
 300 \
 301     sad8_2_ ## suf(blk1, blk1+stride, blk2, stride, 8);\
 302 \
 303     return sum_ ## suf();\
 304 }\
 305 \
 306 static int sad8_xy2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
 307 {\
 308     assert(h==8);\
 309     asm volatile("pxor %%mm7, %%mm7             \n\t"\
 310                  "pxor %%mm6, %%mm6             \n\t"\
 311                  "movq %0, %%mm5                \n\t"\
 312                  :: "m"(round_tab[2]) \
 313                  );\
 314 \
 315     sad8_4_ ## suf(blk1, blk2, stride, 8);\
 316 \
 317     return sum_ ## suf();\
 318 }\
 319 \
 320 static int sad16_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
 321 {\
 322     asm volatile("pxor %%mm7, %%mm7             \n\t"\
 323                  "pxor %%mm6, %%mm6             \n\t":);\
 324 \
 325     sad8_1_ ## suf(blk1  , blk2  , stride, h);\
 326     sad8_1_ ## suf(blk1+8, blk2+8, stride, h);\
 327 \
 328     return sum_ ## suf();\
 329 }\
 330 static int sad16_x2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
 331 {\
 332     asm volatile("pxor %%mm7, %%mm7             \n\t"\
 333                  "pxor %%mm6, %%mm6             \n\t"\
 334                  "movq %0, %%mm5                \n\t"\
 335                  :: "m"(round_tab[1]) \
 336                  );\
 337 \
 338     sad8_2_ ## suf(blk1  , blk1+1, blk2  , stride, h);\
 339     sad8_2_ ## suf(blk1+8, blk1+9, blk2+8, stride, h);\
 340 \
 341     return sum_ ## suf();\
 342 }\
 343 static int sad16_y2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
 344 {\
 345     asm volatile("pxor %%mm7, %%mm7             \n\t"\
 346                  "pxor %%mm6, %%mm6             \n\t"\
 347                  "movq %0, %%mm5                \n\t"\
 348                  :: "m"(round_tab[1]) \
 349                  );\
 350 \
 351     sad8_2_ ## suf(blk1  , blk1+stride,  blk2  , stride, h);\
 352     sad8_2_ ## suf(blk1+8, blk1+stride+8,blk2+8, stride, h);\
 353 \
 354     return sum_ ## suf();\
 355 }\
 356 static int sad16_xy2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
 357 {\
 358     asm volatile("pxor %%mm7, %%mm7             \n\t"\
 359                  "pxor %%mm6, %%mm6             \n\t"\
 360                  "movq %0, %%mm5                \n\t"\
 361                  :: "m"(round_tab[2]) \
 362                  );\
 363 \
 364     sad8_4_ ## suf(blk1  , blk2  , stride, h);\
 365     sad8_4_ ## suf(blk1+8, blk2+8, stride, h);\
 366 \
 367     return sum_ ## suf();\
 368 }\
 369
 370 PIX_SAD(mmx)
 371 PIX_SAD(mmx2)
 372
 373 void dsputil_init_pix_mmx(DSPContext* c, AVCodecContext *avctx)
 374 {
 375     if (mm_flags & MM_MMX) {
 376         c->pix_abs[0][0] = sad16_mmx;
 377         c->pix_abs[0][1] = sad16_x2_mmx;
 378         c->pix_abs[0][2] = sad16_y2_mmx;
 379         c->pix_abs[0][3] = sad16_xy2_mmx;
 380         c->pix_abs[1][0] = sad8_mmx;
 381         c->pix_abs[1][1] = sad8_x2_mmx;
 382         c->pix_abs[1][2] = sad8_y2_mmx;
 383         c->pix_abs[1][3] = sad8_xy2_mmx;
 384
 385         c->sad[0]= sad16_mmx;
 386         c->sad[1]= sad8_mmx;
 387     }
 388     if (mm_flags & MM_MMXEXT) {
 389         c->pix_abs[0][0] = sad16_mmx2;
 390         c->pix_abs[1][0] = sad8_mmx2;
 391
 392         c->sad[0]= sad16_mmx2;
 393         c->sad[1]= sad8_mmx2;
 394
 395         if(!(avctx->flags & CODEC_FLAG_BITEXACT)){
 396             c->pix_abs[0][1] = sad16_x2_mmx2;
 397             c->pix_abs[0][2] = sad16_y2_mmx2;
 398             c->pix_abs[0][3] = sad16_xy2_mmx2;
 399             c->pix_abs[1][1] = sad8_x2_mmx2;
 400             c->pix_abs[1][2] = sad8_y2_mmx2;
 401             c->pix_abs[1][3] = sad8_xy2_mmx2;
 402         }
 403     }
 404 }