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/mem_internal.h"
  28 #include "libavutil/x86/asm.h"
  29 #include "libavutil/x86/cpu.h"
  30 #include "libavcodec/me_cmp.h"
  31 #include "libavcodec/mpegvideo.h"
  32
  33 int ff_sum_abs_dctelem_mmx(int16_t *block);
  34 int ff_sum_abs_dctelem_mmxext(int16_t *block);
  35 int ff_sum_abs_dctelem_sse2(int16_t *block);
  36 int ff_sum_abs_dctelem_ssse3(int16_t *block);
  37 int ff_sse8_mmx(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  38                 ptrdiff_t stride, int h);
  39 int ff_sse16_mmx(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  40                  ptrdiff_t stride, int h);
  41 int ff_sse16_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  42                   ptrdiff_t stride, int h);
  43 int ff_hf_noise8_mmx(uint8_t *pix1, ptrdiff_t stride, int h);
  44 int ff_hf_noise16_mmx(uint8_t *pix1, ptrdiff_t stride, int h);
  45 int ff_sad8_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  46                    ptrdiff_t stride, int h);
  47 int ff_sad16_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  48                     ptrdiff_t stride, int h);
  49 int ff_sad16_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  50                   ptrdiff_t stride, int h);
  51 int ff_sad8_x2_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  52                       ptrdiff_t stride, int h);
  53 int ff_sad16_x2_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  54                        ptrdiff_t stride, int h);
  55 int ff_sad16_x2_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  56                      ptrdiff_t stride, int h);
  57 int ff_sad8_y2_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  58                       ptrdiff_t stride, int h);
  59 int ff_sad16_y2_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  60                        ptrdiff_t stride, int h);
  61 int ff_sad16_y2_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  62                      ptrdiff_t stride, int h);
  63 int ff_sad8_approx_xy2_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  64                               ptrdiff_t stride, int h);
  65 int ff_sad16_approx_xy2_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  66                                ptrdiff_t stride, int h);
  67 int ff_sad16_approx_xy2_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  68                              ptrdiff_t stride, int h);
  69 int ff_vsad_intra8_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  70                           ptrdiff_t stride, int h);
  71 int ff_vsad_intra16_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  72                            ptrdiff_t stride, int h);
  73 int ff_vsad_intra16_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  74                          ptrdiff_t stride, int h);
  75 int ff_vsad8_approx_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  76                     ptrdiff_t stride, int h);
  77 int ff_vsad16_approx_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  78                      ptrdiff_t stride, int h);
  79 int ff_vsad16_approx_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  80                    ptrdiff_t stride, int h);
  81
  82 #define hadamard_func(cpu)                                                    \
  83     int ff_hadamard8_diff_ ## cpu(MpegEncContext *s, uint8_t *src1,           \
  84                                   uint8_t *src2, ptrdiff_t stride, int h);    \
  85     int ff_hadamard8_diff16_ ## cpu(MpegEncContext *s, uint8_t *src1,         \
  86                                     uint8_t *src2, ptrdiff_t stride, int h);
  87
  88 hadamard_func(mmx)
  89 hadamard_func(mmxext)
  90 hadamard_func(sse2)
  91 hadamard_func(ssse3)
  92
  93 #if HAVE_X86ASM
  94 static int nsse16_mmx(MpegEncContext *c, uint8_t *pix1, uint8_t *pix2,
  95                       ptrdiff_t stride, int h)
  96 {
  97     int score1, score2;
  98
  99     if (c)
 100         score1 = c->mecc.sse[0](c, pix1, pix2, stride, h);
 101     else
 102         score1 = ff_sse16_mmx(c, pix1, pix2, stride, h);
 103     score2 = ff_hf_noise16_mmx(pix1, stride, h) + ff_hf_noise8_mmx(pix1+8, stride, h)
 104            - ff_hf_noise16_mmx(pix2, stride, h) - ff_hf_noise8_mmx(pix2+8, stride, h);
 105
 106     if (c)
 107         return score1 + FFABS(score2) * c->avctx->nsse_weight;
 108     else
 109         return score1 + FFABS(score2) * 8;
 110 }
 111
 112 static int nsse8_mmx(MpegEncContext *c, uint8_t *pix1, uint8_t *pix2,
 113                      ptrdiff_t stride, int h)
 114 {
 115     int score1 = ff_sse8_mmx(c, pix1, pix2, stride, h);
 116     int score2 = ff_hf_noise8_mmx(pix1, stride, h) -
 117                  ff_hf_noise8_mmx(pix2, stride, h);
 118
 119     if (c)
 120         return score1 + FFABS(score2) * c->avctx->nsse_weight;
 121     else
 122         return score1 + FFABS(score2) * 8;
 123 }
 124
 125 #endif /* HAVE_X86ASM */
 126
 127 #if HAVE_INLINE_ASM
 128
 129 static int vsad_intra16_mmx(MpegEncContext *v, uint8_t *pix, uint8_t *dummy,
 130                             ptrdiff_t stride, int h)
 131 {
 132     int tmp;
 133
 134     av_assert2((((int) pix) & 7) == 0);
 135     av_assert2((stride & 7) == 0);
 136
 137 #define SUM(in0, in1, out0, out1)               \
 138     "movq (%0), %%mm2\n"                        \
 139     "movq 8(%0), %%mm3\n"                       \
 140     "add %2,%0\n"                               \
 141     "movq %%mm2, " #out0 "\n"                   \
 142     "movq %%mm3, " #out1 "\n"                   \
 143     "psubusb " #in0 ", %%mm2\n"                 \
 144     "psubusb " #in1 ", %%mm3\n"                 \
 145     "psubusb " #out0 ", " #in0 "\n"             \
 146     "psubusb " #out1 ", " #in1 "\n"             \
 147     "por %%mm2, " #in0 "\n"                     \
 148     "por %%mm3, " #in1 "\n"                     \
 149     "movq " #in0 ", %%mm2\n"                    \
 150     "movq " #in1 ", %%mm3\n"                    \
 151     "punpcklbw %%mm7, " #in0 "\n"               \
 152     "punpcklbw %%mm7, " #in1 "\n"               \
 153     "punpckhbw %%mm7, %%mm2\n"                  \
 154     "punpckhbw %%mm7, %%mm3\n"                  \
 155     "paddw " #in1 ", " #in0 "\n"                \
 156     "paddw %%mm3, %%mm2\n"                      \
 157     "paddw %%mm2, " #in0 "\n"                   \
 158     "paddw " #in0 ", %%mm6\n"
 159
 160
 161     __asm__ volatile (
 162         "movl    %3, %%ecx\n"
 163         "pxor %%mm6, %%mm6\n"
 164         "pxor %%mm7, %%mm7\n"
 165         "movq  (%0), %%mm0\n"
 166         "movq 8(%0), %%mm1\n"
 167         "add %2, %0\n"
 168         "jmp 2f\n"
 169         "1:\n"
 170
 171         SUM(%%mm4, %%mm5, %%mm0, %%mm1)
 172         "2:\n"
 173         SUM(%%mm0, %%mm1, %%mm4, %%mm5)
 174
 175         "subl $2, %%ecx\n"
 176         "jnz 1b\n"
 177
 178         "movq  %%mm6, %%mm0\n"
 179         "psrlq $32,   %%mm6\n"
 180         "paddw %%mm6, %%mm0\n"
 181         "movq  %%mm0, %%mm6\n"
 182         "psrlq $16,   %%mm0\n"
 183         "paddw %%mm6, %%mm0\n"
 184         "movd  %%mm0, %1\n"
 185         : "+r" (pix), "=r" (tmp)
 186         : "r" (stride), "m" (h)
 187         : "%ecx");
 188
 189     return tmp & 0xFFFF;
 190 }
 191 #undef SUM
 192
 193 static int vsad16_mmx(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
 194                       ptrdiff_t stride, int h)
 195 {
 196     int tmp;
 197
 198     av_assert2((((int) pix1) & 7) == 0);
 199     av_assert2((((int) pix2) & 7) == 0);
 200     av_assert2((stride & 7) == 0);
 201
 202 #define SUM(in0, in1, out0, out1)       \
 203     "movq (%0), %%mm2\n"                \
 204     "movq (%1), " #out0 "\n"            \
 205     "movq 8(%0), %%mm3\n"               \
 206     "movq 8(%1), " #out1 "\n"           \
 207     "add %3, %0\n"                      \
 208     "add %3, %1\n"                      \
 209     "psubb " #out0 ", %%mm2\n"          \
 210     "psubb " #out1 ", %%mm3\n"          \
 211     "pxor %%mm7, %%mm2\n"               \
 212     "pxor %%mm7, %%mm3\n"               \
 213     "movq %%mm2, " #out0 "\n"           \
 214     "movq %%mm3, " #out1 "\n"           \
 215     "psubusb " #in0 ", %%mm2\n"         \
 216     "psubusb " #in1 ", %%mm3\n"         \
 217     "psubusb " #out0 ", " #in0 "\n"     \
 218     "psubusb " #out1 ", " #in1 "\n"     \
 219     "por %%mm2, " #in0 "\n"             \
 220     "por %%mm3, " #in1 "\n"             \
 221     "movq " #in0 ", %%mm2\n"            \
 222     "movq " #in1 ", %%mm3\n"            \
 223     "punpcklbw %%mm7, " #in0 "\n"       \
 224     "punpcklbw %%mm7, " #in1 "\n"       \
 225     "punpckhbw %%mm7, %%mm2\n"          \
 226     "punpckhbw %%mm7, %%mm3\n"          \
 227     "paddw " #in1 ", " #in0 "\n"        \
 228     "paddw %%mm3, %%mm2\n"              \
 229     "paddw %%mm2, " #in0 "\n"           \
 230     "paddw " #in0 ", %%mm6\n"
 231
 232
 233     __asm__ volatile (
 234         "movl %4, %%ecx\n"
 235         "pxor %%mm6, %%mm6\n"
 236         "pcmpeqw %%mm7, %%mm7\n"
 237         "psllw $15, %%mm7\n"
 238         "packsswb %%mm7, %%mm7\n"
 239         "movq (%0), %%mm0\n"
 240         "movq (%1), %%mm2\n"
 241         "movq 8(%0), %%mm1\n"
 242         "movq 8(%1), %%mm3\n"
 243         "add %3, %0\n"
 244         "add %3, %1\n"
 245         "psubb %%mm2, %%mm0\n"
 246         "psubb %%mm3, %%mm1\n"
 247         "pxor %%mm7, %%mm0\n"
 248         "pxor %%mm7, %%mm1\n"
 249         "jmp 2f\n"
 250         "1:\n"
 251
 252         SUM(%%mm4, %%mm5, %%mm0, %%mm1)
 253         "2:\n"
 254         SUM(%%mm0, %%mm1, %%mm4, %%mm5)
 255
 256         "subl $2, %%ecx\n"
 257         "jnz 1b\n"
 258
 259         "movq %%mm6, %%mm0\n"
 260         "psrlq $32, %%mm6\n"
 261         "paddw %%mm6, %%mm0\n"
 262         "movq %%mm0, %%mm6\n"
 263         "psrlq $16, %%mm0\n"
 264         "paddw %%mm6, %%mm0\n"
 265         "movd %%mm0, %2\n"
 266         : "+r" (pix1), "+r" (pix2), "=r" (tmp)
 267         : "r" (stride), "m" (h)
 268         : "%ecx");
 269
 270     return tmp & 0x7FFF;
 271 }
 272 #undef SUM
 273
 274 DECLARE_ASM_CONST(8, uint64_t, round_tab)[3] = {
 275     0x0000000000000000ULL,
 276     0x0001000100010001ULL,
 277     0x0002000200020002ULL,
 278 };
 279
 280 static inline void sad8_1_mmx(uint8_t *blk1, uint8_t *blk2,
 281                               ptrdiff_t stride, int h)
 282 {
 283     x86_reg len = -stride * h;
 284     __asm__ volatile (
 285         ".p2align 4                     \n\t"
 286         "1:                             \n\t"
 287         "movq (%1, %%"FF_REG_a"), %%mm0 \n\t"
 288         "movq (%2, %%"FF_REG_a"), %%mm2 \n\t"
 289         "movq (%2, %%"FF_REG_a"), %%mm4 \n\t"
 290         "add %3, %%"FF_REG_a"           \n\t"
 291         "psubusb %%mm0, %%mm2           \n\t"
 292         "psubusb %%mm4, %%mm0           \n\t"
 293         "movq (%1, %%"FF_REG_a"), %%mm1 \n\t"
 294         "movq (%2, %%"FF_REG_a"), %%mm3 \n\t"
 295         "movq (%2, %%"FF_REG_a"), %%mm5 \n\t"
 296         "psubusb %%mm1, %%mm3           \n\t"
 297         "psubusb %%mm5, %%mm1           \n\t"
 298         "por %%mm2, %%mm0               \n\t"
 299         "por %%mm1, %%mm3               \n\t"
 300         "movq %%mm0, %%mm1              \n\t"
 301         "movq %%mm3, %%mm2              \n\t"
 302         "punpcklbw %%mm7, %%mm0         \n\t"
 303         "punpckhbw %%mm7, %%mm1         \n\t"
 304         "punpcklbw %%mm7, %%mm3         \n\t"
 305         "punpckhbw %%mm7, %%mm2         \n\t"
 306         "paddw %%mm1, %%mm0             \n\t"
 307         "paddw %%mm3, %%mm2             \n\t"
 308         "paddw %%mm2, %%mm0             \n\t"
 309         "paddw %%mm0, %%mm6             \n\t"
 310         "add %3, %%"FF_REG_a"           \n\t"
 311         " js 1b                         \n\t"
 312         : "+a" (len)
 313         : "r" (blk1 - len), "r" (blk2 - len), "r" (stride));
 314 }
 315
 316 static inline void sad8_2_mmx(uint8_t *blk1a, uint8_t *blk1b, uint8_t *blk2,
 317                               ptrdiff_t stride, int h)
 318 {
 319     x86_reg len = -stride * h;
 320     __asm__ volatile (
 321         ".p2align 4                     \n\t"
 322         "1:                             \n\t"
 323         "movq (%1, %%"FF_REG_a"), %%mm0 \n\t"
 324         "movq (%2, %%"FF_REG_a"), %%mm1 \n\t"
 325         "movq (%1, %%"FF_REG_a"), %%mm2 \n\t"
 326         "movq (%2, %%"FF_REG_a"), %%mm3 \n\t"
 327         "punpcklbw %%mm7, %%mm0         \n\t"
 328         "punpcklbw %%mm7, %%mm1         \n\t"
 329         "punpckhbw %%mm7, %%mm2         \n\t"
 330         "punpckhbw %%mm7, %%mm3         \n\t"
 331         "paddw %%mm0, %%mm1             \n\t"
 332         "paddw %%mm2, %%mm3             \n\t"
 333         "movq (%3, %%"FF_REG_a"), %%mm4 \n\t"
 334         "movq (%3, %%"FF_REG_a"), %%mm2 \n\t"
 335         "paddw %%mm5, %%mm1             \n\t"
 336         "paddw %%mm5, %%mm3             \n\t"
 337         "psrlw $1, %%mm1                \n\t"
 338         "psrlw $1, %%mm3                \n\t"
 339         "packuswb %%mm3, %%mm1          \n\t"
 340         "psubusb %%mm1, %%mm4           \n\t"
 341         "psubusb %%mm2, %%mm1           \n\t"
 342         "por %%mm4, %%mm1               \n\t"
 343         "movq %%mm1, %%mm0              \n\t"
 344         "punpcklbw %%mm7, %%mm0         \n\t"
 345         "punpckhbw %%mm7, %%mm1         \n\t"
 346         "paddw %%mm1, %%mm0             \n\t"
 347         "paddw %%mm0, %%mm6             \n\t"
 348         "add %4, %%"FF_REG_a"           \n\t"
 349         " js 1b                         \n\t"
 350         : "+a" (len)
 351         : "r" (blk1a - len), "r" (blk1b - len), "r" (blk2 - len),
 352           "r" (stride));
 353 }
 354
 355 static inline void sad8_4_mmx(uint8_t *blk1, uint8_t *blk2,
 356                               ptrdiff_t stride, int h)
 357 {
 358     x86_reg len = -stride * h;
 359     __asm__ volatile (
 360         "movq  (%1, %%"FF_REG_a"), %%mm0\n\t"
 361         "movq 1(%1, %%"FF_REG_a"), %%mm2\n\t"
 362         "movq %%mm0, %%mm1              \n\t"
 363         "movq %%mm2, %%mm3              \n\t"
 364         "punpcklbw %%mm7, %%mm0         \n\t"
 365         "punpckhbw %%mm7, %%mm1         \n\t"
 366         "punpcklbw %%mm7, %%mm2         \n\t"
 367         "punpckhbw %%mm7, %%mm3         \n\t"
 368         "paddw %%mm2, %%mm0             \n\t"
 369         "paddw %%mm3, %%mm1             \n\t"
 370         ".p2align 4                     \n\t"
 371         "1:                             \n\t"
 372         "movq  (%2, %%"FF_REG_a"), %%mm2\n\t"
 373         "movq 1(%2, %%"FF_REG_a"), %%mm4\n\t"
 374         "movq %%mm2, %%mm3              \n\t"
 375         "movq %%mm4, %%mm5              \n\t"
 376         "punpcklbw %%mm7, %%mm2         \n\t"
 377         "punpckhbw %%mm7, %%mm3         \n\t"
 378         "punpcklbw %%mm7, %%mm4         \n\t"
 379         "punpckhbw %%mm7, %%mm5         \n\t"
 380         "paddw %%mm4, %%mm2             \n\t"
 381         "paddw %%mm5, %%mm3             \n\t"
 382         "movq %5, %%mm5                 \n\t"
 383         "paddw %%mm2, %%mm0             \n\t"
 384         "paddw %%mm3, %%mm1             \n\t"
 385         "paddw %%mm5, %%mm0             \n\t"
 386         "paddw %%mm5, %%mm1             \n\t"
 387         "movq (%3, %%"FF_REG_a"), %%mm4 \n\t"
 388         "movq (%3, %%"FF_REG_a"), %%mm5 \n\t"
 389         "psrlw $2, %%mm0                \n\t"
 390         "psrlw $2, %%mm1                \n\t"
 391         "packuswb %%mm1, %%mm0          \n\t"
 392         "psubusb %%mm0, %%mm4           \n\t"
 393         "psubusb %%mm5, %%mm0           \n\t"
 394         "por %%mm4, %%mm0               \n\t"
 395         "movq %%mm0, %%mm4              \n\t"
 396         "punpcklbw %%mm7, %%mm0         \n\t"
 397         "punpckhbw %%mm7, %%mm4         \n\t"
 398         "paddw %%mm0, %%mm6             \n\t"
 399         "paddw %%mm4, %%mm6             \n\t"
 400         "movq  %%mm2, %%mm0             \n\t"
 401         "movq  %%mm3, %%mm1             \n\t"
 402         "add %4, %%"FF_REG_a"           \n\t"
 403         " js 1b                         \n\t"
 404         : "+a" (len)
 405         : "r" (blk1 - len), "r" (blk1 - len + stride), "r" (blk2 - len),
 406           "r" (stride), "m" (round_tab[2]));
 407 }
 408
 409 static inline int sum_mmx(void)
 410 {
 411     int ret;
 412     __asm__ volatile (
 413         "movq %%mm6, %%mm0              \n\t"
 414         "psrlq $32, %%mm6               \n\t"
 415         "paddw %%mm0, %%mm6             \n\t"
 416         "movq %%mm6, %%mm0              \n\t"
 417         "psrlq $16, %%mm6               \n\t"
 418         "paddw %%mm0, %%mm6             \n\t"
 419         "movd %%mm6, %0                 \n\t"
 420         : "=r" (ret));
 421     return ret & 0xFFFF;
 422 }
 423
 424 static inline void sad8_x2a_mmx(uint8_t *blk1, uint8_t *blk2,
 425                                 ptrdiff_t stride, int h)
 426 {
 427     sad8_2_mmx(blk1, blk1 + 1, blk2, stride, h);
 428 }
 429
 430 static inline void sad8_y2a_mmx(uint8_t *blk1, uint8_t *blk2,
 431                                 ptrdiff_t stride, int h)
 432 {
 433     sad8_2_mmx(blk1, blk1 + stride, blk2, stride, h);
 434 }
 435
 436 #define PIX_SAD(suf)                                                    \
 437 static int sad8_ ## suf(MpegEncContext *v, uint8_t *blk2,               \
 438                         uint8_t *blk1, ptrdiff_t stride, int h)         \
 439 {                                                                       \
 440     av_assert2(h == 8);                                                     \
 441     __asm__ volatile (                                                  \
 442         "pxor %%mm7, %%mm7     \n\t"                                    \
 443         "pxor %%mm6, %%mm6     \n\t"                                    \
 444         :);                                                             \
 445                                                                         \
 446     sad8_1_ ## suf(blk1, blk2, stride, 8);                              \
 447                                                                         \
 448     return sum_ ## suf();                                               \
 449 }                                                                       \
 450                                                                         \
 451 static int sad8_x2_ ## suf(MpegEncContext *v, uint8_t *blk2,            \
 452                            uint8_t *blk1, ptrdiff_t stride, int h)      \
 453 {                                                                       \
 454     av_assert2(h == 8);                                                     \
 455     __asm__ volatile (                                                  \
 456         "pxor %%mm7, %%mm7     \n\t"                                    \
 457         "pxor %%mm6, %%mm6     \n\t"                                    \
 458         "movq %0, %%mm5        \n\t"                                    \
 459         :: "m" (round_tab[1]));                                         \
 460                                                                         \
 461     sad8_x2a_ ## suf(blk1, blk2, stride, 8);                            \
 462                                                                         \
 463     return sum_ ## suf();                                               \
 464 }                                                                       \
 465                                                                         \
 466 static int sad8_y2_ ## suf(MpegEncContext *v, uint8_t *blk2,            \
 467                            uint8_t *blk1, ptrdiff_t stride, int h)      \
 468 {                                                                       \
 469     av_assert2(h == 8);                                                     \
 470     __asm__ volatile (                                                  \
 471         "pxor %%mm7, %%mm7     \n\t"                                    \
 472         "pxor %%mm6, %%mm6     \n\t"                                    \
 473         "movq %0, %%mm5        \n\t"                                    \
 474         :: "m" (round_tab[1]));                                         \
 475                                                                         \
 476     sad8_y2a_ ## suf(blk1, blk2, stride, 8);                            \
 477                                                                         \
 478     return sum_ ## suf();                                               \
 479 }                                                                       \
 480                                                                         \
 481 static int sad8_xy2_ ## suf(MpegEncContext *v, uint8_t *blk2,           \
 482                             uint8_t *blk1, ptrdiff_t stride, int h)     \
 483 {                                                                       \
 484     av_assert2(h == 8);                                                     \
 485     __asm__ volatile (                                                  \
 486         "pxor %%mm7, %%mm7     \n\t"                                    \
 487         "pxor %%mm6, %%mm6     \n\t"                                    \
 488         ::);                                                            \
 489                                                                         \
 490     sad8_4_ ## suf(blk1, blk2, stride, 8);                              \
 491                                                                         \
 492     return sum_ ## suf();                                               \
 493 }                                                                       \
 494                                                                         \
 495 static int sad16_ ## suf(MpegEncContext *v, uint8_t *blk2,              \
 496                          uint8_t *blk1, ptrdiff_t stride, int h)        \
 497 {                                                                       \
 498     __asm__ volatile (                                                  \
 499         "pxor %%mm7, %%mm7     \n\t"                                    \
 500         "pxor %%mm6, %%mm6     \n\t"                                    \
 501         :);                                                             \
 502                                                                         \
 503     sad8_1_ ## suf(blk1,     blk2,     stride, h);                      \
 504     sad8_1_ ## suf(blk1 + 8, blk2 + 8, stride, h);                      \
 505                                                                         \
 506     return sum_ ## suf();                                               \
 507 }                                                                       \
 508                                                                         \
 509 static int sad16_x2_ ## suf(MpegEncContext *v, uint8_t *blk2,           \
 510                             uint8_t *blk1, ptrdiff_t stride, int h)     \
 511 {                                                                       \
 512     __asm__ volatile (                                                  \
 513         "pxor %%mm7, %%mm7     \n\t"                                    \
 514         "pxor %%mm6, %%mm6     \n\t"                                    \
 515         "movq %0, %%mm5        \n\t"                                    \
 516         :: "m" (round_tab[1]));                                         \
 517                                                                         \
 518     sad8_x2a_ ## suf(blk1,     blk2,     stride, h);                    \
 519     sad8_x2a_ ## suf(blk1 + 8, blk2 + 8, stride, h);                    \
 520                                                                         \
 521     return sum_ ## suf();                                               \
 522 }                                                                       \
 523                                                                         \
 524 static int sad16_y2_ ## suf(MpegEncContext *v, uint8_t *blk2,           \
 525                             uint8_t *blk1, ptrdiff_t stride, int h)     \
 526 {                                                                       \
 527     __asm__ volatile (                                                  \
 528         "pxor %%mm7, %%mm7     \n\t"                                    \
 529         "pxor %%mm6, %%mm6     \n\t"                                    \
 530         "movq %0, %%mm5        \n\t"                                    \
 531         :: "m" (round_tab[1]));                                         \
 532                                                                         \
 533     sad8_y2a_ ## suf(blk1,     blk2,     stride, h);                    \
 534     sad8_y2a_ ## suf(blk1 + 8, blk2 + 8, stride, h);                    \
 535                                                                         \
 536     return sum_ ## suf();                                               \
 537 }                                                                       \
 538                                                                         \
 539 static int sad16_xy2_ ## suf(MpegEncContext *v, uint8_t *blk2,          \
 540                              uint8_t *blk1, ptrdiff_t stride, int h)    \
 541 {                                                                       \
 542     __asm__ volatile (                                                  \
 543         "pxor %%mm7, %%mm7     \n\t"                                    \
 544         "pxor %%mm6, %%mm6     \n\t"                                    \
 545         ::);                                                            \
 546                                                                         \
 547     sad8_4_ ## suf(blk1,     blk2,     stride, h);                      \
 548     sad8_4_ ## suf(blk1 + 8, blk2 + 8, stride, h);                      \
 549                                                                         \
 550     return sum_ ## suf();                                               \
 551 }                                                                       \
 552
 553 PIX_SAD(mmx)
 554
 555 #endif /* HAVE_INLINE_ASM */
 556
 557 av_cold void ff_me_cmp_init_x86(MECmpContext *c, AVCodecContext *avctx)
 558 {
 559     int cpu_flags = av_get_cpu_flags();
 560
 561 #if HAVE_INLINE_ASM
 562     if (INLINE_MMX(cpu_flags)) {
 563         c->pix_abs[0][0] = sad16_mmx;
 564         c->pix_abs[0][1] = sad16_x2_mmx;
 565         c->pix_abs[0][2] = sad16_y2_mmx;
 566         c->pix_abs[0][3] = sad16_xy2_mmx;
 567         c->pix_abs[1][0] = sad8_mmx;
 568         c->pix_abs[1][1] = sad8_x2_mmx;
 569         c->pix_abs[1][2] = sad8_y2_mmx;
 570         c->pix_abs[1][3] = sad8_xy2_mmx;
 571
 572         c->sad[0] = sad16_mmx;
 573         c->sad[1] = sad8_mmx;
 574
 575         c->vsad[4] = vsad_intra16_mmx;
 576
 577         if (!(avctx->flags & AV_CODEC_FLAG_BITEXACT)) {
 578             c->vsad[0] = vsad16_mmx;
 579         }
 580     }
 581
 582 #endif /* HAVE_INLINE_ASM */
 583
 584     if (EXTERNAL_MMX(cpu_flags)) {
 585         c->hadamard8_diff[0] = ff_hadamard8_diff16_mmx;
 586         c->hadamard8_diff[1] = ff_hadamard8_diff_mmx;
 587         c->sum_abs_dctelem   = ff_sum_abs_dctelem_mmx;
 588         c->sse[0]            = ff_sse16_mmx;
 589         c->sse[1]            = ff_sse8_mmx;
 590 #if HAVE_X86ASM
 591         c->nsse[0]           = nsse16_mmx;
 592         c->nsse[1]           = nsse8_mmx;
 593 #endif
 594     }
 595
 596     if (EXTERNAL_MMXEXT(cpu_flags)) {
 597         c->hadamard8_diff[0] = ff_hadamard8_diff16_mmxext;
 598         c->hadamard8_diff[1] = ff_hadamard8_diff_mmxext;
 599         c->sum_abs_dctelem   = ff_sum_abs_dctelem_mmxext;
 600
 601         c->sad[0] = ff_sad16_mmxext;
 602         c->sad[1] = ff_sad8_mmxext;
 603
 604         c->pix_abs[0][0] = ff_sad16_mmxext;
 605         c->pix_abs[0][1] = ff_sad16_x2_mmxext;
 606         c->pix_abs[0][2] = ff_sad16_y2_mmxext;
 607         c->pix_abs[1][0] = ff_sad8_mmxext;
 608         c->pix_abs[1][1] = ff_sad8_x2_mmxext;
 609         c->pix_abs[1][2] = ff_sad8_y2_mmxext;
 610
 611         c->vsad[4] = ff_vsad_intra16_mmxext;
 612         c->vsad[5] = ff_vsad_intra8_mmxext;
 613
 614         if (!(avctx->flags & AV_CODEC_FLAG_BITEXACT)) {
 615             c->pix_abs[0][3] = ff_sad16_approx_xy2_mmxext;
 616             c->pix_abs[1][3] = ff_sad8_approx_xy2_mmxext;
 617
 618             c->vsad[0] = ff_vsad16_approx_mmxext;
 619             c->vsad[1] = ff_vsad8_approx_mmxext;
 620         }
 621     }
 622
 623     if (EXTERNAL_SSE2(cpu_flags)) {
 624         c->sse[0] = ff_sse16_sse2;
 625         c->sum_abs_dctelem   = ff_sum_abs_dctelem_sse2;
 626
 627 #if HAVE_ALIGNED_STACK
 628         c->hadamard8_diff[0] = ff_hadamard8_diff16_sse2;
 629         c->hadamard8_diff[1] = ff_hadamard8_diff_sse2;
 630 #endif
 631         if (!(cpu_flags & AV_CPU_FLAG_SSE2SLOW) && avctx->codec_id != AV_CODEC_ID_SNOW) {
 632             c->sad[0]        = ff_sad16_sse2;
 633             c->pix_abs[0][0] = ff_sad16_sse2;
 634             c->pix_abs[0][1] = ff_sad16_x2_sse2;
 635             c->pix_abs[0][2] = ff_sad16_y2_sse2;
 636
 637             c->vsad[4]       = ff_vsad_intra16_sse2;
 638             if (!(avctx->flags & AV_CODEC_FLAG_BITEXACT)) {
 639                 c->pix_abs[0][3] = ff_sad16_approx_xy2_sse2;
 640                 c->vsad[0]       = ff_vsad16_approx_sse2;
 641             }
 642         }
 643     }
 644
 645     if (EXTERNAL_SSSE3(cpu_flags)) {
 646         c->sum_abs_dctelem   = ff_sum_abs_dctelem_ssse3;
 647 #if HAVE_ALIGNED_STACK
 648         c->hadamard8_diff[0] = ff_hadamard8_diff16_ssse3;
 649         c->hadamard8_diff[1] = ff_hadamard8_diff_ssse3;
 650 #endif
 651     }
 652 }