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

   1 /*
   2  * Optimized for ia32 CPUs by Nick Kurshev <nickols_k@mail.ru>
   3  * H.263, MPEG-1, MPEG-2 dequantizer & draw_edges by Michael Niedermayer <michaelni@gmx.at>
   4  *
   5  * This file is part of FFmpeg.
   6  *
   7  * FFmpeg is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; either
  10  * version 2.1 of the License, or (at your option) any later version.
  11  *
  12  * FFmpeg is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with FFmpeg; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 #include "libavutil/attributes.h"
  23 #include "libavutil/cpu.h"
  24 #include "libavutil/x86/asm.h"
  25 #include "libavutil/x86/cpu.h"
  26 #include "libavcodec/avcodec.h"
  27 #include "libavcodec/mpegvideo.h"
  28 #include "libavcodec/mpegvideodata.h"
  29
  30 #if HAVE_MMX_INLINE
  31
  32 static void dct_unquantize_h263_intra_mmx(MpegEncContext *s,
  33                                   int16_t *block, int n, int qscale)
  34 {
  35     x86_reg level, qmul, qadd, nCoeffs;
  36
  37     qmul = qscale << 1;
  38
  39     av_assert2(s->block_last_index[n]>=0 || s->h263_aic);
  40
  41     if (!s->h263_aic) {
  42         if (n < 4)
  43             level = block[0] * s->y_dc_scale;
  44         else
  45             level = block[0] * s->c_dc_scale;
  46         qadd = (qscale - 1) | 1;
  47     }else{
  48         qadd = 0;
  49         level= block[0];
  50     }
  51     if(s->ac_pred)
  52         nCoeffs=63;
  53     else
  54         nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
  55
  56 __asm__ volatile(
  57                 "movd %1, %%mm6                 \n\t" //qmul
  58                 "packssdw %%mm6, %%mm6          \n\t"
  59                 "packssdw %%mm6, %%mm6          \n\t"
  60                 "movd %2, %%mm5                 \n\t" //qadd
  61                 "pxor %%mm7, %%mm7              \n\t"
  62                 "packssdw %%mm5, %%mm5          \n\t"
  63                 "packssdw %%mm5, %%mm5          \n\t"
  64                 "psubw %%mm5, %%mm7             \n\t"
  65                 "pxor %%mm4, %%mm4              \n\t"
  66                 ".p2align 4                     \n\t"
  67                 "1:                             \n\t"
  68                 "movq (%0, %3), %%mm0           \n\t"
  69                 "movq 8(%0, %3), %%mm1          \n\t"
  70
  71                 "pmullw %%mm6, %%mm0            \n\t"
  72                 "pmullw %%mm6, %%mm1            \n\t"
  73
  74                 "movq (%0, %3), %%mm2           \n\t"
  75                 "movq 8(%0, %3), %%mm3          \n\t"
  76
  77                 "pcmpgtw %%mm4, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
  78                 "pcmpgtw %%mm4, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
  79
  80                 "pxor %%mm2, %%mm0              \n\t"
  81                 "pxor %%mm3, %%mm1              \n\t"
  82
  83                 "paddw %%mm7, %%mm0             \n\t"
  84                 "paddw %%mm7, %%mm1             \n\t"
  85
  86                 "pxor %%mm0, %%mm2              \n\t"
  87                 "pxor %%mm1, %%mm3              \n\t"
  88
  89                 "pcmpeqw %%mm7, %%mm0           \n\t" // block[i] == 0 ? -1 : 0
  90                 "pcmpeqw %%mm7, %%mm1           \n\t" // block[i] == 0 ? -1 : 0
  91
  92                 "pandn %%mm2, %%mm0             \n\t"
  93                 "pandn %%mm3, %%mm1             \n\t"
  94
  95                 "movq %%mm0, (%0, %3)           \n\t"
  96                 "movq %%mm1, 8(%0, %3)          \n\t"
  97
  98                 "add $16, %3                    \n\t"
  99                 "jng 1b                         \n\t"
 100                 ::"r" (block+nCoeffs), "rm"(qmul), "rm" (qadd), "r" (2*(-nCoeffs))
 101                 : "memory"
 102         );
 103         block[0]= level;
 104 }
 105
 106
 107 static void dct_unquantize_h263_inter_mmx(MpegEncContext *s,
 108                                   int16_t *block, int n, int qscale)
 109 {
 110     x86_reg qmul, qadd, nCoeffs;
 111
 112     qmul = qscale << 1;
 113     qadd = (qscale - 1) | 1;
 114
 115     av_assert2(s->block_last_index[n]>=0 || s->h263_aic);
 116
 117     nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ];
 118
 119 __asm__ volatile(
 120                 "movd %1, %%mm6                 \n\t" //qmul
 121                 "packssdw %%mm6, %%mm6          \n\t"
 122                 "packssdw %%mm6, %%mm6          \n\t"
 123                 "movd %2, %%mm5                 \n\t" //qadd
 124                 "pxor %%mm7, %%mm7              \n\t"
 125                 "packssdw %%mm5, %%mm5          \n\t"
 126                 "packssdw %%mm5, %%mm5          \n\t"
 127                 "psubw %%mm5, %%mm7             \n\t"
 128                 "pxor %%mm4, %%mm4              \n\t"
 129                 ".p2align 4                     \n\t"
 130                 "1:                             \n\t"
 131                 "movq (%0, %3), %%mm0           \n\t"
 132                 "movq 8(%0, %3), %%mm1          \n\t"
 133
 134                 "pmullw %%mm6, %%mm0            \n\t"
 135                 "pmullw %%mm6, %%mm1            \n\t"
 136
 137                 "movq (%0, %3), %%mm2           \n\t"
 138                 "movq 8(%0, %3), %%mm3          \n\t"
 139
 140                 "pcmpgtw %%mm4, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
 141                 "pcmpgtw %%mm4, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
 142
 143                 "pxor %%mm2, %%mm0              \n\t"
 144                 "pxor %%mm3, %%mm1              \n\t"
 145
 146                 "paddw %%mm7, %%mm0             \n\t"
 147                 "paddw %%mm7, %%mm1             \n\t"
 148
 149                 "pxor %%mm0, %%mm2              \n\t"
 150                 "pxor %%mm1, %%mm3              \n\t"
 151
 152                 "pcmpeqw %%mm7, %%mm0           \n\t" // block[i] == 0 ? -1 : 0
 153                 "pcmpeqw %%mm7, %%mm1           \n\t" // block[i] == 0 ? -1 : 0
 154
 155                 "pandn %%mm2, %%mm0             \n\t"
 156                 "pandn %%mm3, %%mm1             \n\t"
 157
 158                 "movq %%mm0, (%0, %3)           \n\t"
 159                 "movq %%mm1, 8(%0, %3)          \n\t"
 160
 161                 "add $16, %3                    \n\t"
 162                 "jng 1b                         \n\t"
 163                 ::"r" (block+nCoeffs), "rm"(qmul), "rm" (qadd), "r" (2*(-nCoeffs))
 164                 : "memory"
 165         );
 166 }
 167
 168 static void dct_unquantize_mpeg1_intra_mmx(MpegEncContext *s,
 169                                      int16_t *block, int n, int qscale)
 170 {
 171     x86_reg nCoeffs;
 172     const uint16_t *quant_matrix;
 173     int block0;
 174
 175     av_assert2(s->block_last_index[n]>=0);
 176
 177     nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]+1;
 178
 179     if (n < 4)
 180         block0 = block[0] * s->y_dc_scale;
 181     else
 182         block0 = block[0] * s->c_dc_scale;
 183     /* XXX: only MPEG-1 */
 184     quant_matrix = s->intra_matrix;
 185 __asm__ volatile(
 186                 "pcmpeqw %%mm7, %%mm7           \n\t"
 187                 "psrlw $15, %%mm7               \n\t"
 188                 "movd %2, %%mm6                 \n\t"
 189                 "packssdw %%mm6, %%mm6          \n\t"
 190                 "packssdw %%mm6, %%mm6          \n\t"
 191                 "mov %3, %%"FF_REG_a"           \n\t"
 192                 ".p2align 4                     \n\t"
 193                 "1:                             \n\t"
 194                 "movq (%0, %%"FF_REG_a"), %%mm0 \n\t"
 195                 "movq 8(%0, %%"FF_REG_a"), %%mm1\n\t"
 196                 "movq (%1, %%"FF_REG_a"), %%mm4 \n\t"
 197                 "movq 8(%1, %%"FF_REG_a"), %%mm5\n\t"
 198                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
 199                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
 200                 "pxor %%mm2, %%mm2              \n\t"
 201                 "pxor %%mm3, %%mm3              \n\t"
 202                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
 203                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
 204                 "pxor %%mm2, %%mm0              \n\t"
 205                 "pxor %%mm3, %%mm1              \n\t"
 206                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
 207                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
 208                 "pmullw %%mm4, %%mm0            \n\t" // abs(block[i])*q
 209                 "pmullw %%mm5, %%mm1            \n\t" // abs(block[i])*q
 210                 "pxor %%mm4, %%mm4              \n\t"
 211                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
 212                 "pcmpeqw (%0, %%"FF_REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
 213                 "pcmpeqw 8(%0, %%"FF_REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
 214                 "psraw $3, %%mm0                \n\t"
 215                 "psraw $3, %%mm1                \n\t"
 216                 "psubw %%mm7, %%mm0             \n\t"
 217                 "psubw %%mm7, %%mm1             \n\t"
 218                 "por %%mm7, %%mm0               \n\t"
 219                 "por %%mm7, %%mm1               \n\t"
 220                 "pxor %%mm2, %%mm0              \n\t"
 221                 "pxor %%mm3, %%mm1              \n\t"
 222                 "psubw %%mm2, %%mm0             \n\t"
 223                 "psubw %%mm3, %%mm1             \n\t"
 224                 "pandn %%mm0, %%mm4             \n\t"
 225                 "pandn %%mm1, %%mm5             \n\t"
 226                 "movq %%mm4, (%0, %%"FF_REG_a") \n\t"
 227                 "movq %%mm5, 8(%0, %%"FF_REG_a")\n\t"
 228
 229                 "add $16, %%"FF_REG_a"          \n\t"
 230                 "js 1b                          \n\t"
 231                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "g" (-2*nCoeffs)
 232                 : "%"FF_REG_a, "memory"
 233         );
 234     block[0]= block0;
 235 }
 236
 237 static void dct_unquantize_mpeg1_inter_mmx(MpegEncContext *s,
 238                                      int16_t *block, int n, int qscale)
 239 {
 240     x86_reg nCoeffs;
 241     const uint16_t *quant_matrix;
 242
 243     av_assert2(s->block_last_index[n]>=0);
 244
 245     nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]+1;
 246
 247         quant_matrix = s->inter_matrix;
 248 __asm__ volatile(
 249                 "pcmpeqw %%mm7, %%mm7           \n\t"
 250                 "psrlw $15, %%mm7               \n\t"
 251                 "movd %2, %%mm6                 \n\t"
 252                 "packssdw %%mm6, %%mm6          \n\t"
 253                 "packssdw %%mm6, %%mm6          \n\t"
 254                 "mov %3, %%"FF_REG_a"           \n\t"
 255                 ".p2align 4                     \n\t"
 256                 "1:                             \n\t"
 257                 "movq (%0, %%"FF_REG_a"), %%mm0 \n\t"
 258                 "movq 8(%0, %%"FF_REG_a"), %%mm1\n\t"
 259                 "movq (%1, %%"FF_REG_a"), %%mm4 \n\t"
 260                 "movq 8(%1, %%"FF_REG_a"), %%mm5\n\t"
 261                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
 262                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
 263                 "pxor %%mm2, %%mm2              \n\t"
 264                 "pxor %%mm3, %%mm3              \n\t"
 265                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
 266                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
 267                 "pxor %%mm2, %%mm0              \n\t"
 268                 "pxor %%mm3, %%mm1              \n\t"
 269                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
 270                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
 271                 "paddw %%mm0, %%mm0             \n\t" // abs(block[i])*2
 272                 "paddw %%mm1, %%mm1             \n\t" // abs(block[i])*2
 273                 "paddw %%mm7, %%mm0             \n\t" // abs(block[i])*2 + 1
 274                 "paddw %%mm7, %%mm1             \n\t" // abs(block[i])*2 + 1
 275                 "pmullw %%mm4, %%mm0            \n\t" // (abs(block[i])*2 + 1)*q
 276                 "pmullw %%mm5, %%mm1            \n\t" // (abs(block[i])*2 + 1)*q
 277                 "pxor %%mm4, %%mm4              \n\t"
 278                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
 279                 "pcmpeqw (%0, %%"FF_REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
 280                 "pcmpeqw 8(%0, %%"FF_REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
 281                 "psraw $4, %%mm0                \n\t"
 282                 "psraw $4, %%mm1                \n\t"
 283                 "psubw %%mm7, %%mm0             \n\t"
 284                 "psubw %%mm7, %%mm1             \n\t"
 285                 "por %%mm7, %%mm0               \n\t"
 286                 "por %%mm7, %%mm1               \n\t"
 287                 "pxor %%mm2, %%mm0              \n\t"
 288                 "pxor %%mm3, %%mm1              \n\t"
 289                 "psubw %%mm2, %%mm0             \n\t"
 290                 "psubw %%mm3, %%mm1             \n\t"
 291                 "pandn %%mm0, %%mm4             \n\t"
 292                 "pandn %%mm1, %%mm5             \n\t"
 293                 "movq %%mm4, (%0, %%"FF_REG_a") \n\t"
 294                 "movq %%mm5, 8(%0, %%"FF_REG_a")\n\t"
 295
 296                 "add $16, %%"FF_REG_a"          \n\t"
 297                 "js 1b                          \n\t"
 298                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "g" (-2*nCoeffs)
 299                 : "%"FF_REG_a, "memory"
 300         );
 301 }
 302
 303 static void dct_unquantize_mpeg2_intra_mmx(MpegEncContext *s,
 304                                      int16_t *block, int n, int qscale)
 305 {
 306     x86_reg nCoeffs;
 307     const uint16_t *quant_matrix;
 308     int block0;
 309
 310     av_assert2(s->block_last_index[n]>=0);
 311
 312     if (s->q_scale_type) qscale = ff_mpeg2_non_linear_qscale[qscale];
 313     else                 qscale <<= 1;
 314
 315     if(s->alternate_scan) nCoeffs= 63; //FIXME
 316     else nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
 317
 318     if (n < 4)
 319         block0 = block[0] * s->y_dc_scale;
 320     else
 321         block0 = block[0] * s->c_dc_scale;
 322     quant_matrix = s->intra_matrix;
 323 __asm__ volatile(
 324                 "pcmpeqw %%mm7, %%mm7           \n\t"
 325                 "psrlw $15, %%mm7               \n\t"
 326                 "movd %2, %%mm6                 \n\t"
 327                 "packssdw %%mm6, %%mm6          \n\t"
 328                 "packssdw %%mm6, %%mm6          \n\t"
 329                 "mov %3, %%"FF_REG_a"           \n\t"
 330                 ".p2align 4                     \n\t"
 331                 "1:                             \n\t"
 332                 "movq (%0, %%"FF_REG_a"), %%mm0 \n\t"
 333                 "movq 8(%0, %%"FF_REG_a"), %%mm1\n\t"
 334                 "movq (%1, %%"FF_REG_a"), %%mm4 \n\t"
 335                 "movq 8(%1, %%"FF_REG_a"), %%mm5\n\t"
 336                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
 337                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
 338                 "pxor %%mm2, %%mm2              \n\t"
 339                 "pxor %%mm3, %%mm3              \n\t"
 340                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
 341                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
 342                 "pxor %%mm2, %%mm0              \n\t"
 343                 "pxor %%mm3, %%mm1              \n\t"
 344                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
 345                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
 346                 "pmullw %%mm4, %%mm0            \n\t" // abs(block[i])*q
 347                 "pmullw %%mm5, %%mm1            \n\t" // abs(block[i])*q
 348                 "pxor %%mm4, %%mm4              \n\t"
 349                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
 350                 "pcmpeqw (%0, %%"FF_REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
 351                 "pcmpeqw 8(%0, %%"FF_REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
 352                 "psraw $4, %%mm0                \n\t"
 353                 "psraw $4, %%mm1                \n\t"
 354                 "pxor %%mm2, %%mm0              \n\t"
 355                 "pxor %%mm3, %%mm1              \n\t"
 356                 "psubw %%mm2, %%mm0             \n\t"
 357                 "psubw %%mm3, %%mm1             \n\t"
 358                 "pandn %%mm0, %%mm4             \n\t"
 359                 "pandn %%mm1, %%mm5             \n\t"
 360                 "movq %%mm4, (%0, %%"FF_REG_a") \n\t"
 361                 "movq %%mm5, 8(%0, %%"FF_REG_a")\n\t"
 362
 363                 "add $16, %%"FF_REG_a"          \n\t"
 364                 "jng 1b                         \n\t"
 365                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "g" (-2*nCoeffs)
 366                 : "%"FF_REG_a, "memory"
 367         );
 368     block[0]= block0;
 369         //Note, we do not do mismatch control for intra as errors cannot accumulate
 370 }
 371
 372 static void dct_unquantize_mpeg2_inter_mmx(MpegEncContext *s,
 373                                      int16_t *block, int n, int qscale)
 374 {
 375     x86_reg nCoeffs;
 376     const uint16_t *quant_matrix;
 377
 378     av_assert2(s->block_last_index[n]>=0);
 379
 380     if (s->q_scale_type) qscale = ff_mpeg2_non_linear_qscale[qscale];
 381     else                 qscale <<= 1;
 382
 383     if(s->alternate_scan) nCoeffs= 63; //FIXME
 384     else nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
 385
 386         quant_matrix = s->inter_matrix;
 387 __asm__ volatile(
 388                 "pcmpeqw %%mm7, %%mm7           \n\t"
 389                 "psrlq $48, %%mm7               \n\t"
 390                 "movd %2, %%mm6                 \n\t"
 391                 "packssdw %%mm6, %%mm6          \n\t"
 392                 "packssdw %%mm6, %%mm6          \n\t"
 393                 "mov %3, %%"FF_REG_a"           \n\t"
 394                 ".p2align 4                     \n\t"
 395                 "1:                             \n\t"
 396                 "movq (%0, %%"FF_REG_a"), %%mm0 \n\t"
 397                 "movq 8(%0, %%"FF_REG_a"), %%mm1\n\t"
 398                 "movq (%1, %%"FF_REG_a"), %%mm4 \n\t"
 399                 "movq 8(%1, %%"FF_REG_a"), %%mm5\n\t"
 400                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
 401                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
 402                 "pxor %%mm2, %%mm2              \n\t"
 403                 "pxor %%mm3, %%mm3              \n\t"
 404                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
 405                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
 406                 "pxor %%mm2, %%mm0              \n\t"
 407                 "pxor %%mm3, %%mm1              \n\t"
 408                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
 409                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
 410                 "paddw %%mm0, %%mm0             \n\t" // abs(block[i])*2
 411                 "paddw %%mm1, %%mm1             \n\t" // abs(block[i])*2
 412                 "pmullw %%mm4, %%mm0            \n\t" // abs(block[i])*2*q
 413                 "pmullw %%mm5, %%mm1            \n\t" // abs(block[i])*2*q
 414                 "paddw %%mm4, %%mm0             \n\t" // (abs(block[i])*2 + 1)*q
 415                 "paddw %%mm5, %%mm1             \n\t" // (abs(block[i])*2 + 1)*q
 416                 "pxor %%mm4, %%mm4              \n\t"
 417                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
 418                 "pcmpeqw (%0, %%"FF_REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
 419                 "pcmpeqw 8(%0, %%"FF_REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
 420                 "psrlw $5, %%mm0                \n\t"
 421                 "psrlw $5, %%mm1                \n\t"
 422                 "pxor %%mm2, %%mm0              \n\t"
 423                 "pxor %%mm3, %%mm1              \n\t"
 424                 "psubw %%mm2, %%mm0             \n\t"
 425                 "psubw %%mm3, %%mm1             \n\t"
 426                 "pandn %%mm0, %%mm4             \n\t"
 427                 "pandn %%mm1, %%mm5             \n\t"
 428                 "pxor %%mm4, %%mm7              \n\t"
 429                 "pxor %%mm5, %%mm7              \n\t"
 430                 "movq %%mm4, (%0, %%"FF_REG_a") \n\t"
 431                 "movq %%mm5, 8(%0, %%"FF_REG_a")\n\t"
 432
 433                 "add $16, %%"FF_REG_a"          \n\t"
 434                 "jng 1b                         \n\t"
 435                 "movd 124(%0, %3), %%mm0        \n\t"
 436                 "movq %%mm7, %%mm6              \n\t"
 437                 "psrlq $32, %%mm7               \n\t"
 438                 "pxor %%mm6, %%mm7              \n\t"
 439                 "movq %%mm7, %%mm6              \n\t"
 440                 "psrlq $16, %%mm7               \n\t"
 441                 "pxor %%mm6, %%mm7              \n\t"
 442                 "pslld $31, %%mm7               \n\t"
 443                 "psrlq $15, %%mm7               \n\t"
 444                 "pxor %%mm7, %%mm0              \n\t"
 445                 "movd %%mm0, 124(%0, %3)        \n\t"
 446
 447                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "r" (-2*nCoeffs)
 448                 : "%"FF_REG_a, "memory"
 449         );
 450 }
 451
 452 #endif /* HAVE_MMX_INLINE */
 453
 454 av_cold void ff_mpv_common_init_x86(MpegEncContext *s)
 455 {
 456 #if HAVE_MMX_INLINE
 457     int cpu_flags = av_get_cpu_flags();
 458
 459     if (INLINE_MMX(cpu_flags)) {
 460         s->dct_unquantize_h263_intra = dct_unquantize_h263_intra_mmx;
 461         s->dct_unquantize_h263_inter = dct_unquantize_h263_inter_mmx;
 462         s->dct_unquantize_mpeg1_intra = dct_unquantize_mpeg1_intra_mmx;
 463         s->dct_unquantize_mpeg1_inter = dct_unquantize_mpeg1_inter_mmx;
 464         if (!(s->avctx->flags & AV_CODEC_FLAG_BITEXACT))
 465             s->dct_unquantize_mpeg2_intra = dct_unquantize_mpeg2_intra_mmx;
 466         s->dct_unquantize_mpeg2_inter = dct_unquantize_mpeg2_inter_mmx;
 467     }
 468 #endif /* HAVE_MMX_INLINE */
 469 }