git.sesse.net Git - ffmpeg/blob - libavcodec/x86/me_cmp_init.c

   1 /*
   2  * SIMD-optimized motion estimation
   3  * Copyright (c) 2000, 2001 Fabrice Bellard
   4  * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
   5  *
   6  * MMX optimization by Nick Kurshev <nickols_k@mail.ru>
   7  *
   8  * This file is part of FFmpeg.
   9  *
  10  * FFmpeg is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2.1 of the License, or (at your option) any later version.
  14  *
  15  * FFmpeg is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * Lesser General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU Lesser General Public
  21  * License along with FFmpeg; if not, write to the Free Software
  22  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  23  */
  24
  25 #include "libavutil/attributes.h"
  26 #include "libavutil/cpu.h"
  27 #include "libavutil/x86/asm.h"
  28 #include "libavutil/x86/cpu.h"
  29 #include "libavcodec/me_cmp.h"
  30 #include "libavcodec/mpegvideo.h"
  31
  32 int ff_sum_abs_dctelem_mmx(int16_t *block);
  33 int ff_sum_abs_dctelem_mmxext(int16_t *block);
  34 int ff_sum_abs_dctelem_sse2(int16_t *block);
  35 int ff_sum_abs_dctelem_ssse3(int16_t *block);
  36 int ff_sse8_mmx(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  37                 int line_size, int h);
  38 int ff_sse16_mmx(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  39                  int line_size, int h);
  40 int ff_sse16_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  41                   int line_size, int h);
  42 int ff_hf_noise8_mmx(uint8_t *pix1, int lsize, int h);
  43 int ff_hf_noise16_mmx(uint8_t *pix1, int lsize, int h);
  44 int ff_sad8_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  45                    int stride, int h);
  46 int ff_sad16_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  47                     int stride, int h);
  48 int ff_sad16_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  49                   int stride, int h);
  50 int ff_sad8_x2_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  51                       int stride, int h);
  52 int ff_sad16_x2_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  53                        int stride, int h);
  54 int ff_sad16_x2_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  55                      int stride, int h);
  56 int ff_sad8_y2_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  57                       int stride, int h);
  58 int ff_sad16_y2_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  59                        int stride, int h);
  60 int ff_sad16_y2_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  61                      int stride, int h);
  62 int ff_sad8_approx_xy2_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  63                               int stride, int h);
  64 int ff_sad16_approx_xy2_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  65                                int stride, int h);
  66 int ff_sad16_approx_xy2_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  67                              int stride, int h);
  68 int ff_vsad_intra8_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  69                           int line_size, int h);
  70 int ff_vsad_intra16_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  71                            int line_size, int h);
  72 int ff_vsad_intra16_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  73                          int line_size, int h);
  74 int ff_vsad8_approx_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  75                     int line_size, int h);
  76 int ff_vsad16_approx_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  77                      int line_size, int h);
  78 int ff_vsad16_approx_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  79                    int line_size, int h);
  80
  81 #define hadamard_func(cpu)                                              \
  82     int ff_hadamard8_diff_ ## cpu(MpegEncContext *s, uint8_t *src1,     \
  83                                   uint8_t *src2, int stride, int h);    \
  84     int ff_hadamard8_diff16_ ## cpu(MpegEncContext *s, uint8_t *src1,   \
  85                                     uint8_t *src2, int stride, int h);
  86
  87 hadamard_func(mmx)
  88 hadamard_func(mmxext)
  89 hadamard_func(sse2)
  90 hadamard_func(ssse3)
  91
  92 #if HAVE_YASM
  93 static int nsse16_mmx(MpegEncContext *c, uint8_t *pix1, uint8_t *pix2,
  94                       int line_size, int h)
  95 {
  96     int score1, score2;
  97
  98     if (c)
  99         score1 = c->mecc.sse[0](c, pix1, pix2, line_size, h);
 100     else
 101         score1 = ff_sse16_mmx(c, pix1, pix2, line_size, h);
 102     score2 = ff_hf_noise16_mmx(pix1, line_size, h) + ff_hf_noise8_mmx(pix1+8, line_size, h)
 103            - ff_hf_noise16_mmx(pix2, line_size, h) - ff_hf_noise8_mmx(pix2+8, line_size, h);
 104
 105     if (c)
 106         return score1 + FFABS(score2) * c->avctx->nsse_weight;
 107     else
 108         return score1 + FFABS(score2) * 8;
 109 }
 110
 111 static int nsse8_mmx(MpegEncContext *c, uint8_t *pix1, uint8_t *pix2,
 112                      int line_size, int h)
 113 {
 114     int score1 = ff_sse8_mmx(c, pix1, pix2, line_size, h);
 115     int score2 = ff_hf_noise8_mmx(pix1, line_size, h) -
 116                  ff_hf_noise8_mmx(pix2, line_size, h);
 117
 118     if (c)
 119         return score1 + FFABS(score2) * c->avctx->nsse_weight;
 120     else
 121         return score1 + FFABS(score2) * 8;
 122 }
 123
 124 #endif /* HAVE_YASM */
 125
 126 #if HAVE_INLINE_ASM
 127
 128 static int vsad_intra16_mmx(MpegEncContext *v, uint8_t *pix, uint8_t *dummy,
 129                             int line_size, int h)
 130 {
 131     int tmp;
 132
 133     av_assert2((((int) pix) & 7) == 0);
 134     av_assert2((line_size & 7) == 0);
 135
 136 #define SUM(in0, in1, out0, out1)               \
 137     "movq (%0), %%mm2\n"                        \
 138     "movq 8(%0), %%mm3\n"                       \
 139     "add %2,%0\n"                               \
 140     "movq %%mm2, " #out0 "\n"                   \
 141     "movq %%mm3, " #out1 "\n"                   \
 142     "psubusb " #in0 ", %%mm2\n"                 \
 143     "psubusb " #in1 ", %%mm3\n"                 \
 144     "psubusb " #out0 ", " #in0 "\n"             \
 145     "psubusb " #out1 ", " #in1 "\n"             \
 146     "por %%mm2, " #in0 "\n"                     \
 147     "por %%mm3, " #in1 "\n"                     \
 148     "movq " #in0 ", %%mm2\n"                    \
 149     "movq " #in1 ", %%mm3\n"                    \
 150     "punpcklbw %%mm7, " #in0 "\n"               \
 151     "punpcklbw %%mm7, " #in1 "\n"               \
 152     "punpckhbw %%mm7, %%mm2\n"                  \
 153     "punpckhbw %%mm7, %%mm3\n"                  \
 154     "paddw " #in1 ", " #in0 "\n"                \
 155     "paddw %%mm3, %%mm2\n"                      \
 156     "paddw %%mm2, " #in0 "\n"                   \
 157     "paddw " #in0 ", %%mm6\n"
 158
 159
 160     __asm__ volatile (
 161         "movl    %3, %%ecx\n"
 162         "pxor %%mm6, %%mm6\n"
 163         "pxor %%mm7, %%mm7\n"
 164         "movq  (%0), %%mm0\n"
 165         "movq 8(%0), %%mm1\n"
 166         "add %2, %0\n"
 167         "jmp 2f\n"
 168         "1:\n"
 169
 170         SUM(%%mm4, %%mm5, %%mm0, %%mm1)
 171         "2:\n"
 172         SUM(%%mm0, %%mm1, %%mm4, %%mm5)
 173
 174         "subl $2, %%ecx\n"
 175         "jnz 1b\n"
 176
 177         "movq  %%mm6, %%mm0\n"
 178         "psrlq $32,   %%mm6\n"
 179         "paddw %%mm6, %%mm0\n"
 180         "movq  %%mm0, %%mm6\n"
 181         "psrlq $16,   %%mm0\n"
 182         "paddw %%mm6, %%mm0\n"
 183         "movd  %%mm0, %1\n"
 184         : "+r" (pix), "=r" (tmp)
 185         : "r" ((x86_reg) line_size), "m" (h)
 186         : "%ecx");
 187
 188     return tmp & 0xFFFF;
 189 }
 190 #undef SUM
 191
 192 static int vsad16_mmx(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
 193                       int line_size, int h)
 194 {
 195     int tmp;
 196
 197     av_assert2((((int) pix1) & 7) == 0);
 198     av_assert2((((int) pix2) & 7) == 0);
 199     av_assert2((line_size & 7) == 0);
 200
 201 #define SUM(in0, in1, out0, out1)       \
 202     "movq (%0), %%mm2\n"                \
 203     "movq (%1), " #out0 "\n"            \
 204     "movq 8(%0), %%mm3\n"               \
 205     "movq 8(%1), " #out1 "\n"           \
 206     "add %3, %0\n"                      \
 207     "add %3, %1\n"                      \
 208     "psubb " #out0 ", %%mm2\n"          \
 209     "psubb " #out1 ", %%mm3\n"          \
 210     "pxor %%mm7, %%mm2\n"               \
 211     "pxor %%mm7, %%mm3\n"               \
 212     "movq %%mm2, " #out0 "\n"           \
 213     "movq %%mm3, " #out1 "\n"           \
 214     "psubusb " #in0 ", %%mm2\n"         \
 215     "psubusb " #in1 ", %%mm3\n"         \
 216     "psubusb " #out0 ", " #in0 "\n"     \
 217     "psubusb " #out1 ", " #in1 "\n"     \
 218     "por %%mm2, " #in0 "\n"             \
 219     "por %%mm3, " #in1 "\n"             \
 220     "movq " #in0 ", %%mm2\n"            \
 221     "movq " #in1 ", %%mm3\n"            \
 222     "punpcklbw %%mm7, " #in0 "\n"       \
 223     "punpcklbw %%mm7, " #in1 "\n"       \
 224     "punpckhbw %%mm7, %%mm2\n"          \
 225     "punpckhbw %%mm7, %%mm3\n"          \
 226     "paddw " #in1 ", " #in0 "\n"        \
 227     "paddw %%mm3, %%mm2\n"              \
 228     "paddw %%mm2, " #in0 "\n"           \
 229     "paddw " #in0 ", %%mm6\n"
 230
 231
 232     __asm__ volatile (
 233         "movl %4, %%ecx\n"
 234         "pxor %%mm6, %%mm6\n"
 235         "pcmpeqw %%mm7, %%mm7\n"
 236         "psllw $15, %%mm7\n"
 237         "packsswb %%mm7, %%mm7\n"
 238         "movq (%0), %%mm0\n"
 239         "movq (%1), %%mm2\n"
 240         "movq 8(%0), %%mm1\n"
 241         "movq 8(%1), %%mm3\n"
 242         "add %3, %0\n"
 243         "add %3, %1\n"
 244         "psubb %%mm2, %%mm0\n"
 245         "psubb %%mm3, %%mm1\n"
 246         "pxor %%mm7, %%mm0\n"
 247         "pxor %%mm7, %%mm1\n"
 248         "jmp 2f\n"
 249         "1:\n"
 250
 251         SUM(%%mm4, %%mm5, %%mm0, %%mm1)
 252         "2:\n"
 253         SUM(%%mm0, %%mm1, %%mm4, %%mm5)
 254
 255         "subl $2, %%ecx\n"
 256         "jnz 1b\n"
 257
 258         "movq %%mm6, %%mm0\n"
 259         "psrlq $32, %%mm6\n"
 260         "paddw %%mm6, %%mm0\n"
 261         "movq %%mm0, %%mm6\n"
 262         "psrlq $16, %%mm0\n"
 263         "paddw %%mm6, %%mm0\n"
 264         "movd %%mm0, %2\n"
 265         : "+r" (pix1), "+r" (pix2), "=r" (tmp)
 266         : "r" ((x86_reg) line_size), "m" (h)
 267         : "%ecx");
 268
 269     return tmp & 0x7FFF;
 270 }
 271 #undef SUM
 272
 273 DECLARE_ASM_CONST(8, uint64_t, round_tab)[3] = {
 274     0x0000000000000000ULL,
 275     0x0001000100010001ULL,
 276     0x0002000200020002ULL,
 277 };
 278
 279 static inline void sad8_1_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
 280 {
 281     x86_reg len = -(x86_reg)stride * h;
 282     __asm__ volatile (
 283         ".p2align 4                     \n\t"
 284         "1:                             \n\t"
 285         "movq (%1, %%"REG_a"), %%mm0    \n\t"
 286         "movq (%2, %%"REG_a"), %%mm2    \n\t"
 287         "movq (%2, %%"REG_a"), %%mm4    \n\t"
 288         "add %3, %%"REG_a"              \n\t"
 289         "psubusb %%mm0, %%mm2           \n\t"
 290         "psubusb %%mm4, %%mm0           \n\t"
 291         "movq (%1, %%"REG_a"), %%mm1    \n\t"
 292         "movq (%2, %%"REG_a"), %%mm3    \n\t"
 293         "movq (%2, %%"REG_a"), %%mm5    \n\t"
 294         "psubusb %%mm1, %%mm3           \n\t"
 295         "psubusb %%mm5, %%mm1           \n\t"
 296         "por %%mm2, %%mm0               \n\t"
 297         "por %%mm1, %%mm3               \n\t"
 298         "movq %%mm0, %%mm1              \n\t"
 299         "movq %%mm3, %%mm2              \n\t"
 300         "punpcklbw %%mm7, %%mm0         \n\t"
 301         "punpckhbw %%mm7, %%mm1         \n\t"
 302         "punpcklbw %%mm7, %%mm3         \n\t"
 303         "punpckhbw %%mm7, %%mm2         \n\t"
 304         "paddw %%mm1, %%mm0             \n\t"
 305         "paddw %%mm3, %%mm2             \n\t"
 306         "paddw %%mm2, %%mm0             \n\t"
 307         "paddw %%mm0, %%mm6             \n\t"
 308         "add %3, %%"REG_a"              \n\t"
 309         " js 1b                         \n\t"
 310         : "+a" (len)
 311         : "r" (blk1 - len), "r" (blk2 - len), "r" ((x86_reg) stride));
 312 }
 313
 314 static inline void sad8_2_mmx(uint8_t *blk1a, uint8_t *blk1b, uint8_t *blk2,
 315                               int stride, int h)
 316 {
 317     x86_reg len = -(x86_reg)stride * h;
 318     __asm__ volatile (
 319         ".p2align 4                     \n\t"
 320         "1:                             \n\t"
 321         "movq (%1, %%"REG_a"), %%mm0    \n\t"
 322         "movq (%2, %%"REG_a"), %%mm1    \n\t"
 323         "movq (%1, %%"REG_a"), %%mm2    \n\t"
 324         "movq (%2, %%"REG_a"), %%mm3    \n\t"
 325         "punpcklbw %%mm7, %%mm0         \n\t"
 326         "punpcklbw %%mm7, %%mm1         \n\t"
 327         "punpckhbw %%mm7, %%mm2         \n\t"
 328         "punpckhbw %%mm7, %%mm3         \n\t"
 329         "paddw %%mm0, %%mm1             \n\t"
 330         "paddw %%mm2, %%mm3             \n\t"
 331         "movq (%3, %%"REG_a"), %%mm4    \n\t"
 332         "movq (%3, %%"REG_a"), %%mm2    \n\t"
 333         "paddw %%mm5, %%mm1             \n\t"
 334         "paddw %%mm5, %%mm3             \n\t"
 335         "psrlw $1, %%mm1                \n\t"
 336         "psrlw $1, %%mm3                \n\t"
 337         "packuswb %%mm3, %%mm1          \n\t"
 338         "psubusb %%mm1, %%mm4           \n\t"
 339         "psubusb %%mm2, %%mm1           \n\t"
 340         "por %%mm4, %%mm1               \n\t"
 341         "movq %%mm1, %%mm0              \n\t"
 342         "punpcklbw %%mm7, %%mm0         \n\t"
 343         "punpckhbw %%mm7, %%mm1         \n\t"
 344         "paddw %%mm1, %%mm0             \n\t"
 345         "paddw %%mm0, %%mm6             \n\t"
 346         "add %4, %%"REG_a"              \n\t"
 347         " js 1b                         \n\t"
 348         : "+a" (len)
 349         : "r" (blk1a - len), "r" (blk1b - len), "r" (blk2 - len),
 350           "r" ((x86_reg) stride));
 351 }
 352
 353 static inline void sad8_4_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
 354 {
 355     x86_reg len = -(x86_reg)stride * h;
 356     __asm__ volatile (
 357         "movq  (%1, %%"REG_a"), %%mm0   \n\t"
 358         "movq 1(%1, %%"REG_a"), %%mm2   \n\t"
 359         "movq %%mm0, %%mm1              \n\t"
 360         "movq %%mm2, %%mm3              \n\t"
 361         "punpcklbw %%mm7, %%mm0         \n\t"
 362         "punpckhbw %%mm7, %%mm1         \n\t"
 363         "punpcklbw %%mm7, %%mm2         \n\t"
 364         "punpckhbw %%mm7, %%mm3         \n\t"
 365         "paddw %%mm2, %%mm0             \n\t"
 366         "paddw %%mm3, %%mm1             \n\t"
 367         ".p2align 4                     \n\t"
 368         "1:                             \n\t"
 369         "movq  (%2, %%"REG_a"), %%mm2   \n\t"
 370         "movq 1(%2, %%"REG_a"), %%mm4   \n\t"
 371         "movq %%mm2, %%mm3              \n\t"
 372         "movq %%mm4, %%mm5              \n\t"
 373         "punpcklbw %%mm7, %%mm2         \n\t"
 374         "punpckhbw %%mm7, %%mm3         \n\t"
 375         "punpcklbw %%mm7, %%mm4         \n\t"
 376         "punpckhbw %%mm7, %%mm5         \n\t"
 377         "paddw %%mm4, %%mm2             \n\t"
 378         "paddw %%mm5, %%mm3             \n\t"
 379         "movq %5, %%mm5                 \n\t"
 380         "paddw %%mm2, %%mm0             \n\t"
 381         "paddw %%mm3, %%mm1             \n\t"
 382         "paddw %%mm5, %%mm0             \n\t"
 383         "paddw %%mm5, %%mm1             \n\t"
 384         "movq (%3, %%"REG_a"), %%mm4    \n\t"
 385         "movq (%3, %%"REG_a"), %%mm5    \n\t"
 386         "psrlw $2, %%mm0                \n\t"
 387         "psrlw $2, %%mm1                \n\t"
 388         "packuswb %%mm1, %%mm0          \n\t"
 389         "psubusb %%mm0, %%mm4           \n\t"
 390         "psubusb %%mm5, %%mm0           \n\t"
 391         "por %%mm4, %%mm0               \n\t"
 392         "movq %%mm0, %%mm4              \n\t"
 393         "punpcklbw %%mm7, %%mm0         \n\t"
 394         "punpckhbw %%mm7, %%mm4         \n\t"
 395         "paddw %%mm0, %%mm6             \n\t"
 396         "paddw %%mm4, %%mm6             \n\t"
 397         "movq  %%mm2, %%mm0             \n\t"
 398         "movq  %%mm3, %%mm1             \n\t"
 399         "add %4, %%"REG_a"              \n\t"
 400         " js 1b                         \n\t"
 401         : "+a" (len)
 402         : "r" (blk1 - len), "r" (blk1 - len + stride), "r" (blk2 - len),
 403           "r" ((x86_reg) stride), "m" (round_tab[2]));
 404 }
 405
 406 static inline int sum_mmx(void)
 407 {
 408     int ret;
 409     __asm__ volatile (
 410         "movq %%mm6, %%mm0              \n\t"
 411         "psrlq $32, %%mm6               \n\t"
 412         "paddw %%mm0, %%mm6             \n\t"
 413         "movq %%mm6, %%mm0              \n\t"
 414         "psrlq $16, %%mm6               \n\t"
 415         "paddw %%mm0, %%mm6             \n\t"
 416         "movd %%mm6, %0                 \n\t"
 417         : "=r" (ret));
 418     return ret & 0xFFFF;
 419 }
 420
 421 static inline void sad8_x2a_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
 422 {
 423     sad8_2_mmx(blk1, blk1 + 1, blk2, stride, h);
 424 }
 425
 426 static inline void sad8_y2a_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
 427 {
 428     sad8_2_mmx(blk1, blk1 + stride, blk2, stride, h);
 429 }
 430
 431 #define PIX_SAD(suf)                                                    \
 432 static int sad8_ ## suf(MpegEncContext *v, uint8_t *blk2,               \
 433                         uint8_t *blk1, int stride, int h)               \
 434 {                                                                       \
 435     av_assert2(h == 8);                                                     \
 436     __asm__ volatile (                                                  \
 437         "pxor %%mm7, %%mm7     \n\t"                                    \
 438         "pxor %%mm6, %%mm6     \n\t"                                    \
 439         :);                                                             \
 440                                                                         \
 441     sad8_1_ ## suf(blk1, blk2, stride, 8);                              \
 442                                                                         \
 443     return sum_ ## suf();                                               \
 444 }                                                                       \
 445                                                                         \
 446 static int sad8_x2_ ## suf(MpegEncContext *v, uint8_t *blk2,            \
 447                            uint8_t *blk1, int stride, int h)            \
 448 {                                                                       \
 449     av_assert2(h == 8);                                                     \
 450     __asm__ volatile (                                                  \
 451         "pxor %%mm7, %%mm7     \n\t"                                    \
 452         "pxor %%mm6, %%mm6     \n\t"                                    \
 453         "movq %0, %%mm5        \n\t"                                    \
 454         :: "m" (round_tab[1]));                                         \
 455                                                                         \
 456     sad8_x2a_ ## suf(blk1, blk2, stride, 8);                            \
 457                                                                         \
 458     return sum_ ## suf();                                               \
 459 }                                                                       \
 460                                                                         \
 461 static int sad8_y2_ ## suf(MpegEncContext *v, uint8_t *blk2,            \
 462                            uint8_t *blk1, int stride, int h)            \
 463 {                                                                       \
 464     av_assert2(h == 8);                                                     \
 465     __asm__ volatile (                                                  \
 466         "pxor %%mm7, %%mm7     \n\t"                                    \
 467         "pxor %%mm6, %%mm6     \n\t"                                    \
 468         "movq %0, %%mm5        \n\t"                                    \
 469         :: "m" (round_tab[1]));                                         \
 470                                                                         \
 471     sad8_y2a_ ## suf(blk1, blk2, stride, 8);                            \
 472                                                                         \
 473     return sum_ ## suf();                                               \
 474 }                                                                       \
 475                                                                         \
 476 static int sad8_xy2_ ## suf(MpegEncContext *v, uint8_t *blk2,           \
 477                             uint8_t *blk1, int stride, int h)           \
 478 {                                                                       \
 479     av_assert2(h == 8);                                                     \
 480     __asm__ volatile (                                                  \
 481         "pxor %%mm7, %%mm7     \n\t"                                    \
 482         "pxor %%mm6, %%mm6     \n\t"                                    \
 483         ::);                                                            \
 484                                                                         \
 485     sad8_4_ ## suf(blk1, blk2, stride, 8);                              \
 486                                                                         \
 487     return sum_ ## suf();                                               \
 488 }                                                                       \
 489                                                                         \
 490 static int sad16_ ## suf(MpegEncContext *v, uint8_t *blk2,              \
 491                          uint8_t *blk1, int stride, int h)              \
 492 {                                                                       \
 493     __asm__ volatile (                                                  \
 494         "pxor %%mm7, %%mm7     \n\t"                                    \
 495         "pxor %%mm6, %%mm6     \n\t"                                    \
 496         :);                                                             \
 497                                                                         \
 498     sad8_1_ ## suf(blk1,     blk2,     stride, h);                      \
 499     sad8_1_ ## suf(blk1 + 8, blk2 + 8, stride, h);                      \
 500                                                                         \
 501     return sum_ ## suf();                                               \
 502 }                                                                       \
 503                                                                         \
 504 static int sad16_x2_ ## suf(MpegEncContext *v, uint8_t *blk2,           \
 505                             uint8_t *blk1, int stride, int h)           \
 506 {                                                                       \
 507     __asm__ volatile (                                                  \
 508         "pxor %%mm7, %%mm7     \n\t"                                    \
 509         "pxor %%mm6, %%mm6     \n\t"                                    \
 510         "movq %0, %%mm5        \n\t"                                    \
 511         :: "m" (round_tab[1]));                                         \
 512                                                                         \
 513     sad8_x2a_ ## suf(blk1,     blk2,     stride, h);                    \
 514     sad8_x2a_ ## suf(blk1 + 8, blk2 + 8, stride, h);                    \
 515                                                                         \
 516     return sum_ ## suf();                                               \
 517 }                                                                       \
 518                                                                         \
 519 static int sad16_y2_ ## suf(MpegEncContext *v, uint8_t *blk2,           \
 520                             uint8_t *blk1, int stride, int h)           \
 521 {                                                                       \
 522     __asm__ volatile (                                                  \
 523         "pxor %%mm7, %%mm7     \n\t"                                    \
 524         "pxor %%mm6, %%mm6     \n\t"                                    \
 525         "movq %0, %%mm5        \n\t"                                    \
 526         :: "m" (round_tab[1]));                                         \
 527                                                                         \
 528     sad8_y2a_ ## suf(blk1,     blk2,     stride, h);                    \
 529     sad8_y2a_ ## suf(blk1 + 8, blk2 + 8, stride, h);                    \
 530                                                                         \
 531     return sum_ ## suf();                                               \
 532 }                                                                       \
 533                                                                         \
 534 static int sad16_xy2_ ## suf(MpegEncContext *v, uint8_t *blk2,          \
 535                              uint8_t *blk1, int stride, int h)          \
 536 {                                                                       \
 537     __asm__ volatile (                                                  \
 538         "pxor %%mm7, %%mm7     \n\t"                                    \
 539         "pxor %%mm6, %%mm6     \n\t"                                    \
 540         ::);                                                            \
 541                                                                         \
 542     sad8_4_ ## suf(blk1,     blk2,     stride, h);                      \
 543     sad8_4_ ## suf(blk1 + 8, blk2 + 8, stride, h);                      \
 544                                                                         \
 545     return sum_ ## suf();                                               \
 546 }                                                                       \
 547
 548 PIX_SAD(mmx)
 549
 550 #endif /* HAVE_INLINE_ASM */
 551
 552 av_cold void ff_me_cmp_init_x86(MECmpContext *c, AVCodecContext *avctx)
 553 {
 554     int cpu_flags = av_get_cpu_flags();
 555
 556 #if HAVE_INLINE_ASM
 557     if (INLINE_MMX(cpu_flags)) {
 558         c->pix_abs[0][0] = sad16_mmx;
 559         c->pix_abs[0][1] = sad16_x2_mmx;
 560         c->pix_abs[0][2] = sad16_y2_mmx;
 561         c->pix_abs[0][3] = sad16_xy2_mmx;
 562         c->pix_abs[1][0] = sad8_mmx;
 563         c->pix_abs[1][1] = sad8_x2_mmx;
 564         c->pix_abs[1][2] = sad8_y2_mmx;
 565         c->pix_abs[1][3] = sad8_xy2_mmx;
 566
 567         c->sad[0] = sad16_mmx;
 568         c->sad[1] = sad8_mmx;
 569
 570         c->vsad[4] = vsad_intra16_mmx;
 571
 572         if (!(avctx->flags & CODEC_FLAG_BITEXACT)) {
 573             c->vsad[0] = vsad16_mmx;
 574         }
 575     }
 576
 577 #endif /* HAVE_INLINE_ASM */
 578
 579     if (EXTERNAL_MMX(cpu_flags)) {
 580         c->hadamard8_diff[0] = ff_hadamard8_diff16_mmx;
 581         c->hadamard8_diff[1] = ff_hadamard8_diff_mmx;
 582         c->sum_abs_dctelem   = ff_sum_abs_dctelem_mmx;
 583         c->sse[0]            = ff_sse16_mmx;
 584         c->sse[1]            = ff_sse8_mmx;
 585 #if HAVE_YASM
 586         c->nsse[0]           = nsse16_mmx;
 587         c->nsse[1]           = nsse8_mmx;
 588 #endif
 589     }
 590
 591     if (EXTERNAL_MMXEXT(cpu_flags)) {
 592         c->hadamard8_diff[0] = ff_hadamard8_diff16_mmxext;
 593         c->hadamard8_diff[1] = ff_hadamard8_diff_mmxext;
 594         c->sum_abs_dctelem   = ff_sum_abs_dctelem_mmxext;
 595
 596         c->sad[0] = ff_sad16_mmxext;
 597         c->sad[1] = ff_sad8_mmxext;
 598
 599         c->pix_abs[0][0] = ff_sad16_mmxext;
 600         c->pix_abs[0][1] = ff_sad16_x2_mmxext;
 601         c->pix_abs[0][2] = ff_sad16_y2_mmxext;
 602         c->pix_abs[1][0] = ff_sad8_mmxext;
 603         c->pix_abs[1][1] = ff_sad8_x2_mmxext;
 604         c->pix_abs[1][2] = ff_sad8_y2_mmxext;
 605
 606         c->vsad[4] = ff_vsad_intra16_mmxext;
 607         c->vsad[5] = ff_vsad_intra8_mmxext;
 608
 609         if (!(avctx->flags & CODEC_FLAG_BITEXACT)) {
 610             c->pix_abs[0][3] = ff_sad16_approx_xy2_mmxext;
 611             c->pix_abs[1][3] = ff_sad8_approx_xy2_mmxext;
 612
 613             c->vsad[0] = ff_vsad16_approx_mmxext;
 614             c->vsad[1] = ff_vsad8_approx_mmxext;
 615         }
 616     }
 617
 618     if (EXTERNAL_SSE2(cpu_flags)) {
 619         c->sse[0] = ff_sse16_sse2;
 620         c->sum_abs_dctelem   = ff_sum_abs_dctelem_sse2;
 621
 622 #if HAVE_ALIGNED_STACK
 623         c->hadamard8_diff[0] = ff_hadamard8_diff16_sse2;
 624         c->hadamard8_diff[1] = ff_hadamard8_diff_sse2;
 625 #endif
 626         if (!(cpu_flags & AV_CPU_FLAG_SSE2SLOW) && avctx->codec_id != AV_CODEC_ID_SNOW) {
 627             c->sad[0]        = ff_sad16_sse2;
 628             c->pix_abs[0][0] = ff_sad16_sse2;
 629             c->pix_abs[0][1] = ff_sad16_x2_sse2;
 630             c->pix_abs[0][2] = ff_sad16_y2_sse2;
 631
 632             c->vsad[4]       = ff_vsad_intra16_sse2;
 633             if (!(avctx->flags & CODEC_FLAG_BITEXACT)) {
 634                 c->pix_abs[0][3] = ff_sad16_approx_xy2_sse2;
 635                 c->vsad[0]       = ff_vsad16_approx_sse2;
 636             }
 637         }
 638     }
 639
 640     if (EXTERNAL_SSSE3(cpu_flags)) {
 641         c->sum_abs_dctelem   = ff_sum_abs_dctelem_ssse3;
 642 #if HAVE_ALIGNED_STACK
 643         c->hadamard8_diff[0] = ff_hadamard8_diff16_ssse3;
 644         c->hadamard8_diff[1] = ff_hadamard8_diff_ssse3;
 645 #endif
 646     }
 647 }