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

   1 /*
   2  * Optimized for ia32 CPUs by Nick Kurshev <nickols_k@mail.ru>
   3  * h263, mpeg1, mpeg2 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/cpu.h"
  23 #include "libavutil/x86/asm.h"
  24 #include "libavcodec/avcodec.h"
  25 #include "libavcodec/dsputil.h"
  26 #include "libavcodec/mpegvideo.h"
  27 #include "dsputil_mmx.h"
  28
  29 #if HAVE_INLINE_ASM
  30
  31 static void dct_unquantize_h263_intra_mmx(MpegEncContext *s,
  32                                   int16_t *block, int n, int qscale)
  33 {
  34     x86_reg level, qmul, qadd, nCoeffs;
  35
  36     qmul = qscale << 1;
  37
  38     av_assert2(s->block_last_index[n]>=0 || s->h263_aic);
  39
  40     if (!s->h263_aic) {
  41         if (n < 4)
  42             level = block[0] * s->y_dc_scale;
  43         else
  44             level = block[0] * s->c_dc_scale;
  45         qadd = (qscale - 1) | 1;
  46     }else{
  47         qadd = 0;
  48         level= block[0];
  49     }
  50     if(s->ac_pred)
  51         nCoeffs=63;
  52     else
  53         nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ];
  54
  55 __asm__ volatile(
  56                 "movd %1, %%mm6                 \n\t" //qmul
  57                 "packssdw %%mm6, %%mm6          \n\t"
  58                 "packssdw %%mm6, %%mm6          \n\t"
  59                 "movd %2, %%mm5                 \n\t" //qadd
  60                 "pxor %%mm7, %%mm7              \n\t"
  61                 "packssdw %%mm5, %%mm5          \n\t"
  62                 "packssdw %%mm5, %%mm5          \n\t"
  63                 "psubw %%mm5, %%mm7             \n\t"
  64                 "pxor %%mm4, %%mm4              \n\t"
  65                 ".p2align 4                     \n\t"
  66                 "1:                             \n\t"
  67                 "movq (%0, %3), %%mm0           \n\t"
  68                 "movq 8(%0, %3), %%mm1          \n\t"
  69
  70                 "pmullw %%mm6, %%mm0            \n\t"
  71                 "pmullw %%mm6, %%mm1            \n\t"
  72
  73                 "movq (%0, %3), %%mm2           \n\t"
  74                 "movq 8(%0, %3), %%mm3          \n\t"
  75
  76                 "pcmpgtw %%mm4, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
  77                 "pcmpgtw %%mm4, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
  78
  79                 "pxor %%mm2, %%mm0              \n\t"
  80                 "pxor %%mm3, %%mm1              \n\t"
  81
  82                 "paddw %%mm7, %%mm0             \n\t"
  83                 "paddw %%mm7, %%mm1             \n\t"
  84
  85                 "pxor %%mm0, %%mm2              \n\t"
  86                 "pxor %%mm1, %%mm3              \n\t"
  87
  88                 "pcmpeqw %%mm7, %%mm0           \n\t" // block[i] == 0 ? -1 : 0
  89                 "pcmpeqw %%mm7, %%mm1           \n\t" // block[i] == 0 ? -1 : 0
  90
  91                 "pandn %%mm2, %%mm0             \n\t"
  92                 "pandn %%mm3, %%mm1             \n\t"
  93
  94                 "movq %%mm0, (%0, %3)           \n\t"
  95                 "movq %%mm1, 8(%0, %3)          \n\t"
  96
  97                 "add $16, %3                    \n\t"
  98                 "jng 1b                         \n\t"
  99                 ::"r" (block+nCoeffs), "rm"(qmul), "rm" (qadd), "r" (2*(-nCoeffs))
 100                 : "memory"
 101         );
 102         block[0]= level;
 103 }
 104
 105
 106 static void dct_unquantize_h263_inter_mmx(MpegEncContext *s,
 107                                   int16_t *block, int n, int qscale)
 108 {
 109     x86_reg qmul, qadd, nCoeffs;
 110
 111     qmul = qscale << 1;
 112     qadd = (qscale - 1) | 1;
 113
 114     assert(s->block_last_index[n]>=0 || s->h263_aic);
 115
 116     nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ];
 117
 118 __asm__ volatile(
 119                 "movd %1, %%mm6                 \n\t" //qmul
 120                 "packssdw %%mm6, %%mm6          \n\t"
 121                 "packssdw %%mm6, %%mm6          \n\t"
 122                 "movd %2, %%mm5                 \n\t" //qadd
 123                 "pxor %%mm7, %%mm7              \n\t"
 124                 "packssdw %%mm5, %%mm5          \n\t"
 125                 "packssdw %%mm5, %%mm5          \n\t"
 126                 "psubw %%mm5, %%mm7             \n\t"
 127                 "pxor %%mm4, %%mm4              \n\t"
 128                 ".p2align 4                     \n\t"
 129                 "1:                             \n\t"
 130                 "movq (%0, %3), %%mm0           \n\t"
 131                 "movq 8(%0, %3), %%mm1          \n\t"
 132
 133                 "pmullw %%mm6, %%mm0            \n\t"
 134                 "pmullw %%mm6, %%mm1            \n\t"
 135
 136                 "movq (%0, %3), %%mm2           \n\t"
 137                 "movq 8(%0, %3), %%mm3          \n\t"
 138
 139                 "pcmpgtw %%mm4, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
 140                 "pcmpgtw %%mm4, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
 141
 142                 "pxor %%mm2, %%mm0              \n\t"
 143                 "pxor %%mm3, %%mm1              \n\t"
 144
 145                 "paddw %%mm7, %%mm0             \n\t"
 146                 "paddw %%mm7, %%mm1             \n\t"
 147
 148                 "pxor %%mm0, %%mm2              \n\t"
 149                 "pxor %%mm1, %%mm3              \n\t"
 150
 151                 "pcmpeqw %%mm7, %%mm0           \n\t" // block[i] == 0 ? -1 : 0
 152                 "pcmpeqw %%mm7, %%mm1           \n\t" // block[i] == 0 ? -1 : 0
 153
 154                 "pandn %%mm2, %%mm0             \n\t"
 155                 "pandn %%mm3, %%mm1             \n\t"
 156
 157                 "movq %%mm0, (%0, %3)           \n\t"
 158                 "movq %%mm1, 8(%0, %3)          \n\t"
 159
 160                 "add $16, %3                    \n\t"
 161                 "jng 1b                         \n\t"
 162                 ::"r" (block+nCoeffs), "rm"(qmul), "rm" (qadd), "r" (2*(-nCoeffs))
 163                 : "memory"
 164         );
 165 }
 166
 167
 168 /*
 169   We can suppose that result of two multiplications can't be greater than 0xFFFF
 170   i.e. is 16-bit, so we use here only PMULLW instruction and can avoid
 171   a complex multiplication.
 172 =====================================================
 173  Full formula for multiplication of 2 integer numbers
 174  which are represent as high:low words:
 175  input: value1 = high1:low1
 176         value2 = high2:low2
 177  output: value3 = value1*value2
 178  value3=high3:low3 (on overflow: modulus 2^32 wrap-around)
 179  this mean that for 0x123456 * 0x123456 correct result is 0x766cb0ce4
 180  but this algorithm will compute only 0x66cb0ce4
 181  this limited by 16-bit size of operands
 182  ---------------------------------
 183  tlow1 = high1*low2
 184  tlow2 = high2*low1
 185  tlow1 = tlow1 + tlow2
 186  high3:low3 = low1*low2
 187  high3 += tlow1
 188 */
 189 static void dct_unquantize_mpeg1_intra_mmx(MpegEncContext *s,
 190                                      int16_t *block, int n, int qscale)
 191 {
 192     x86_reg nCoeffs;
 193     const uint16_t *quant_matrix;
 194     int block0;
 195
 196     av_assert2(s->block_last_index[n]>=0);
 197
 198     nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]+1;
 199
 200     if (n < 4)
 201         block0 = block[0] * s->y_dc_scale;
 202     else
 203         block0 = block[0] * s->c_dc_scale;
 204     /* XXX: only mpeg1 */
 205     quant_matrix = s->intra_matrix;
 206 __asm__ volatile(
 207                 "pcmpeqw %%mm7, %%mm7           \n\t"
 208                 "psrlw $15, %%mm7               \n\t"
 209                 "movd %2, %%mm6                 \n\t"
 210                 "packssdw %%mm6, %%mm6          \n\t"
 211                 "packssdw %%mm6, %%mm6          \n\t"
 212                 "mov %3, %%"REG_a"              \n\t"
 213                 ".p2align 4                     \n\t"
 214                 "1:                             \n\t"
 215                 "movq (%0, %%"REG_a"), %%mm0    \n\t"
 216                 "movq 8(%0, %%"REG_a"), %%mm1   \n\t"
 217                 "movq (%1, %%"REG_a"), %%mm4    \n\t"
 218                 "movq 8(%1, %%"REG_a"), %%mm5   \n\t"
 219                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
 220                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
 221                 "pxor %%mm2, %%mm2              \n\t"
 222                 "pxor %%mm3, %%mm3              \n\t"
 223                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
 224                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
 225                 "pxor %%mm2, %%mm0              \n\t"
 226                 "pxor %%mm3, %%mm1              \n\t"
 227                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
 228                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
 229                 "pmullw %%mm4, %%mm0            \n\t" // abs(block[i])*q
 230                 "pmullw %%mm5, %%mm1            \n\t" // abs(block[i])*q
 231                 "pxor %%mm4, %%mm4              \n\t"
 232                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
 233                 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
 234                 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
 235                 "psraw $3, %%mm0                \n\t"
 236                 "psraw $3, %%mm1                \n\t"
 237                 "psubw %%mm7, %%mm0             \n\t"
 238                 "psubw %%mm7, %%mm1             \n\t"
 239                 "por %%mm7, %%mm0               \n\t"
 240                 "por %%mm7, %%mm1               \n\t"
 241                 "pxor %%mm2, %%mm0              \n\t"
 242                 "pxor %%mm3, %%mm1              \n\t"
 243                 "psubw %%mm2, %%mm0             \n\t"
 244                 "psubw %%mm3, %%mm1             \n\t"
 245                 "pandn %%mm0, %%mm4             \n\t"
 246                 "pandn %%mm1, %%mm5             \n\t"
 247                 "movq %%mm4, (%0, %%"REG_a")    \n\t"
 248                 "movq %%mm5, 8(%0, %%"REG_a")   \n\t"
 249
 250                 "add $16, %%"REG_a"             \n\t"
 251                 "js 1b                          \n\t"
 252                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "g" (-2*nCoeffs)
 253                 : "%"REG_a, "memory"
 254         );
 255     block[0]= block0;
 256 }
 257
 258 static void dct_unquantize_mpeg1_inter_mmx(MpegEncContext *s,
 259                                      int16_t *block, int n, int qscale)
 260 {
 261     x86_reg nCoeffs;
 262     const uint16_t *quant_matrix;
 263
 264     av_assert2(s->block_last_index[n]>=0);
 265
 266     nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]+1;
 267
 268         quant_matrix = s->inter_matrix;
 269 __asm__ volatile(
 270                 "pcmpeqw %%mm7, %%mm7           \n\t"
 271                 "psrlw $15, %%mm7               \n\t"
 272                 "movd %2, %%mm6                 \n\t"
 273                 "packssdw %%mm6, %%mm6          \n\t"
 274                 "packssdw %%mm6, %%mm6          \n\t"
 275                 "mov %3, %%"REG_a"              \n\t"
 276                 ".p2align 4                     \n\t"
 277                 "1:                             \n\t"
 278                 "movq (%0, %%"REG_a"), %%mm0    \n\t"
 279                 "movq 8(%0, %%"REG_a"), %%mm1   \n\t"
 280                 "movq (%1, %%"REG_a"), %%mm4    \n\t"
 281                 "movq 8(%1, %%"REG_a"), %%mm5   \n\t"
 282                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
 283                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
 284                 "pxor %%mm2, %%mm2              \n\t"
 285                 "pxor %%mm3, %%mm3              \n\t"
 286                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
 287                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
 288                 "pxor %%mm2, %%mm0              \n\t"
 289                 "pxor %%mm3, %%mm1              \n\t"
 290                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
 291                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
 292                 "paddw %%mm0, %%mm0             \n\t" // abs(block[i])*2
 293                 "paddw %%mm1, %%mm1             \n\t" // abs(block[i])*2
 294                 "paddw %%mm7, %%mm0             \n\t" // abs(block[i])*2 + 1
 295                 "paddw %%mm7, %%mm1             \n\t" // abs(block[i])*2 + 1
 296                 "pmullw %%mm4, %%mm0            \n\t" // (abs(block[i])*2 + 1)*q
 297                 "pmullw %%mm5, %%mm1            \n\t" // (abs(block[i])*2 + 1)*q
 298                 "pxor %%mm4, %%mm4              \n\t"
 299                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
 300                 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
 301                 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
 302                 "psraw $4, %%mm0                \n\t"
 303                 "psraw $4, %%mm1                \n\t"
 304                 "psubw %%mm7, %%mm0             \n\t"
 305                 "psubw %%mm7, %%mm1             \n\t"
 306                 "por %%mm7, %%mm0               \n\t"
 307                 "por %%mm7, %%mm1               \n\t"
 308                 "pxor %%mm2, %%mm0              \n\t"
 309                 "pxor %%mm3, %%mm1              \n\t"
 310                 "psubw %%mm2, %%mm0             \n\t"
 311                 "psubw %%mm3, %%mm1             \n\t"
 312                 "pandn %%mm0, %%mm4             \n\t"
 313                 "pandn %%mm1, %%mm5             \n\t"
 314                 "movq %%mm4, (%0, %%"REG_a")    \n\t"
 315                 "movq %%mm5, 8(%0, %%"REG_a")   \n\t"
 316
 317                 "add $16, %%"REG_a"             \n\t"
 318                 "js 1b                          \n\t"
 319                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "g" (-2*nCoeffs)
 320                 : "%"REG_a, "memory"
 321         );
 322 }
 323
 324 static void dct_unquantize_mpeg2_intra_mmx(MpegEncContext *s,
 325                                      int16_t *block, int n, int qscale)
 326 {
 327     x86_reg nCoeffs;
 328     const uint16_t *quant_matrix;
 329     int block0;
 330
 331     av_assert2(s->block_last_index[n]>=0);
 332
 333     if(s->alternate_scan) nCoeffs= 63; //FIXME
 334     else nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
 335
 336     if (n < 4)
 337         block0 = block[0] * s->y_dc_scale;
 338     else
 339         block0 = block[0] * s->c_dc_scale;
 340     quant_matrix = s->intra_matrix;
 341 __asm__ volatile(
 342                 "pcmpeqw %%mm7, %%mm7           \n\t"
 343                 "psrlw $15, %%mm7               \n\t"
 344                 "movd %2, %%mm6                 \n\t"
 345                 "packssdw %%mm6, %%mm6          \n\t"
 346                 "packssdw %%mm6, %%mm6          \n\t"
 347                 "mov %3, %%"REG_a"              \n\t"
 348                 ".p2align 4                     \n\t"
 349                 "1:                             \n\t"
 350                 "movq (%0, %%"REG_a"), %%mm0    \n\t"
 351                 "movq 8(%0, %%"REG_a"), %%mm1   \n\t"
 352                 "movq (%1, %%"REG_a"), %%mm4    \n\t"
 353                 "movq 8(%1, %%"REG_a"), %%mm5   \n\t"
 354                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
 355                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
 356                 "pxor %%mm2, %%mm2              \n\t"
 357                 "pxor %%mm3, %%mm3              \n\t"
 358                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
 359                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
 360                 "pxor %%mm2, %%mm0              \n\t"
 361                 "pxor %%mm3, %%mm1              \n\t"
 362                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
 363                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
 364                 "pmullw %%mm4, %%mm0            \n\t" // abs(block[i])*q
 365                 "pmullw %%mm5, %%mm1            \n\t" // abs(block[i])*q
 366                 "pxor %%mm4, %%mm4              \n\t"
 367                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
 368                 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
 369                 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
 370                 "psraw $3, %%mm0                \n\t"
 371                 "psraw $3, %%mm1                \n\t"
 372                 "pxor %%mm2, %%mm0              \n\t"
 373                 "pxor %%mm3, %%mm1              \n\t"
 374                 "psubw %%mm2, %%mm0             \n\t"
 375                 "psubw %%mm3, %%mm1             \n\t"
 376                 "pandn %%mm0, %%mm4             \n\t"
 377                 "pandn %%mm1, %%mm5             \n\t"
 378                 "movq %%mm4, (%0, %%"REG_a")    \n\t"
 379                 "movq %%mm5, 8(%0, %%"REG_a")   \n\t"
 380
 381                 "add $16, %%"REG_a"             \n\t"
 382                 "jng 1b                         \n\t"
 383                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "g" (-2*nCoeffs)
 384                 : "%"REG_a, "memory"
 385         );
 386     block[0]= block0;
 387         //Note, we do not do mismatch control for intra as errors cannot accumulate
 388 }
 389
 390 static void dct_unquantize_mpeg2_inter_mmx(MpegEncContext *s,
 391                                      int16_t *block, int n, int qscale)
 392 {
 393     x86_reg nCoeffs;
 394     const uint16_t *quant_matrix;
 395
 396     av_assert2(s->block_last_index[n]>=0);
 397
 398     if(s->alternate_scan) nCoeffs= 63; //FIXME
 399     else nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
 400
 401         quant_matrix = s->inter_matrix;
 402 __asm__ volatile(
 403                 "pcmpeqw %%mm7, %%mm7           \n\t"
 404                 "psrlq $48, %%mm7               \n\t"
 405                 "movd %2, %%mm6                 \n\t"
 406                 "packssdw %%mm6, %%mm6          \n\t"
 407                 "packssdw %%mm6, %%mm6          \n\t"
 408                 "mov %3, %%"REG_a"              \n\t"
 409                 ".p2align 4                     \n\t"
 410                 "1:                             \n\t"
 411                 "movq (%0, %%"REG_a"), %%mm0    \n\t"
 412                 "movq 8(%0, %%"REG_a"), %%mm1   \n\t"
 413                 "movq (%1, %%"REG_a"), %%mm4    \n\t"
 414                 "movq 8(%1, %%"REG_a"), %%mm5   \n\t"
 415                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
 416                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
 417                 "pxor %%mm2, %%mm2              \n\t"
 418                 "pxor %%mm3, %%mm3              \n\t"
 419                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
 420                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
 421                 "pxor %%mm2, %%mm0              \n\t"
 422                 "pxor %%mm3, %%mm1              \n\t"
 423                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
 424                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
 425                 "paddw %%mm0, %%mm0             \n\t" // abs(block[i])*2
 426                 "paddw %%mm1, %%mm1             \n\t" // abs(block[i])*2
 427                 "pmullw %%mm4, %%mm0            \n\t" // abs(block[i])*2*q
 428                 "pmullw %%mm5, %%mm1            \n\t" // abs(block[i])*2*q
 429                 "paddw %%mm4, %%mm0             \n\t" // (abs(block[i])*2 + 1)*q
 430                 "paddw %%mm5, %%mm1             \n\t" // (abs(block[i])*2 + 1)*q
 431                 "pxor %%mm4, %%mm4              \n\t"
 432                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
 433                 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
 434                 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
 435                 "psrlw $4, %%mm0                \n\t"
 436                 "psrlw $4, %%mm1                \n\t"
 437                 "pxor %%mm2, %%mm0              \n\t"
 438                 "pxor %%mm3, %%mm1              \n\t"
 439                 "psubw %%mm2, %%mm0             \n\t"
 440                 "psubw %%mm3, %%mm1             \n\t"
 441                 "pandn %%mm0, %%mm4             \n\t"
 442                 "pandn %%mm1, %%mm5             \n\t"
 443                 "pxor %%mm4, %%mm7              \n\t"
 444                 "pxor %%mm5, %%mm7              \n\t"
 445                 "movq %%mm4, (%0, %%"REG_a")    \n\t"
 446                 "movq %%mm5, 8(%0, %%"REG_a")   \n\t"
 447
 448                 "add $16, %%"REG_a"             \n\t"
 449                 "jng 1b                         \n\t"
 450                 "movd 124(%0, %3), %%mm0        \n\t"
 451                 "movq %%mm7, %%mm6              \n\t"
 452                 "psrlq $32, %%mm7               \n\t"
 453                 "pxor %%mm6, %%mm7              \n\t"
 454                 "movq %%mm7, %%mm6              \n\t"
 455                 "psrlq $16, %%mm7               \n\t"
 456                 "pxor %%mm6, %%mm7              \n\t"
 457                 "pslld $31, %%mm7               \n\t"
 458                 "psrlq $15, %%mm7               \n\t"
 459                 "pxor %%mm7, %%mm0              \n\t"
 460                 "movd %%mm0, 124(%0, %3)        \n\t"
 461
 462                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "r" (-2*nCoeffs)
 463                 : "%"REG_a, "memory"
 464         );
 465 }
 466
 467 static void  denoise_dct_mmx(MpegEncContext *s, int16_t *block){
 468     const int intra= s->mb_intra;
 469     int *sum= s->dct_error_sum[intra];
 470     uint16_t *offset= s->dct_offset[intra];
 471
 472     s->dct_count[intra]++;
 473
 474     __asm__ volatile(
 475         "pxor %%mm7, %%mm7                      \n\t"
 476         "1:                                     \n\t"
 477         "pxor %%mm0, %%mm0                      \n\t"
 478         "pxor %%mm1, %%mm1                      \n\t"
 479         "movq (%0), %%mm2                       \n\t"
 480         "movq 8(%0), %%mm3                      \n\t"
 481         "pcmpgtw %%mm2, %%mm0                   \n\t"
 482         "pcmpgtw %%mm3, %%mm1                   \n\t"
 483         "pxor %%mm0, %%mm2                      \n\t"
 484         "pxor %%mm1, %%mm3                      \n\t"
 485         "psubw %%mm0, %%mm2                     \n\t"
 486         "psubw %%mm1, %%mm3                     \n\t"
 487         "movq %%mm2, %%mm4                      \n\t"
 488         "movq %%mm3, %%mm5                      \n\t"
 489         "psubusw (%2), %%mm2                    \n\t"
 490         "psubusw 8(%2), %%mm3                   \n\t"
 491         "pxor %%mm0, %%mm2                      \n\t"
 492         "pxor %%mm1, %%mm3                      \n\t"
 493         "psubw %%mm0, %%mm2                     \n\t"
 494         "psubw %%mm1, %%mm3                     \n\t"
 495         "movq %%mm2, (%0)                       \n\t"
 496         "movq %%mm3, 8(%0)                      \n\t"
 497         "movq %%mm4, %%mm2                      \n\t"
 498         "movq %%mm5, %%mm3                      \n\t"
 499         "punpcklwd %%mm7, %%mm4                 \n\t"
 500         "punpckhwd %%mm7, %%mm2                 \n\t"
 501         "punpcklwd %%mm7, %%mm5                 \n\t"
 502         "punpckhwd %%mm7, %%mm3                 \n\t"
 503         "paddd (%1), %%mm4                      \n\t"
 504         "paddd 8(%1), %%mm2                     \n\t"
 505         "paddd 16(%1), %%mm5                    \n\t"
 506         "paddd 24(%1), %%mm3                    \n\t"
 507         "movq %%mm4, (%1)                       \n\t"
 508         "movq %%mm2, 8(%1)                      \n\t"
 509         "movq %%mm5, 16(%1)                     \n\t"
 510         "movq %%mm3, 24(%1)                     \n\t"
 511         "add $16, %0                            \n\t"
 512         "add $32, %1                            \n\t"
 513         "add $16, %2                            \n\t"
 514         "cmp %3, %0                             \n\t"
 515             " jb 1b                             \n\t"
 516         : "+r" (block), "+r" (sum), "+r" (offset)
 517         : "r"(block+64)
 518     );
 519 }
 520
 521 static void  denoise_dct_sse2(MpegEncContext *s, int16_t *block){
 522     const int intra= s->mb_intra;
 523     int *sum= s->dct_error_sum[intra];
 524     uint16_t *offset= s->dct_offset[intra];
 525
 526     s->dct_count[intra]++;
 527
 528     __asm__ volatile(
 529         "pxor %%xmm7, %%xmm7                    \n\t"
 530         "1:                                     \n\t"
 531         "pxor %%xmm0, %%xmm0                    \n\t"
 532         "pxor %%xmm1, %%xmm1                    \n\t"
 533         "movdqa (%0), %%xmm2                    \n\t"
 534         "movdqa 16(%0), %%xmm3                  \n\t"
 535         "pcmpgtw %%xmm2, %%xmm0                 \n\t"
 536         "pcmpgtw %%xmm3, %%xmm1                 \n\t"
 537         "pxor %%xmm0, %%xmm2                    \n\t"
 538         "pxor %%xmm1, %%xmm3                    \n\t"
 539         "psubw %%xmm0, %%xmm2                   \n\t"
 540         "psubw %%xmm1, %%xmm3                   \n\t"
 541         "movdqa %%xmm2, %%xmm4                  \n\t"
 542         "movdqa %%xmm3, %%xmm5                  \n\t"
 543         "psubusw (%2), %%xmm2                   \n\t"
 544         "psubusw 16(%2), %%xmm3                 \n\t"
 545         "pxor %%xmm0, %%xmm2                    \n\t"
 546         "pxor %%xmm1, %%xmm3                    \n\t"
 547         "psubw %%xmm0, %%xmm2                   \n\t"
 548         "psubw %%xmm1, %%xmm3                   \n\t"
 549         "movdqa %%xmm2, (%0)                    \n\t"
 550         "movdqa %%xmm3, 16(%0)                  \n\t"
 551         "movdqa %%xmm4, %%xmm6                  \n\t"
 552         "movdqa %%xmm5, %%xmm0                  \n\t"
 553         "punpcklwd %%xmm7, %%xmm4               \n\t"
 554         "punpckhwd %%xmm7, %%xmm6               \n\t"
 555         "punpcklwd %%xmm7, %%xmm5               \n\t"
 556         "punpckhwd %%xmm7, %%xmm0               \n\t"
 557         "paddd (%1), %%xmm4                     \n\t"
 558         "paddd 16(%1), %%xmm6                   \n\t"
 559         "paddd 32(%1), %%xmm5                   \n\t"
 560         "paddd 48(%1), %%xmm0                   \n\t"
 561         "movdqa %%xmm4, (%1)                    \n\t"
 562         "movdqa %%xmm6, 16(%1)                  \n\t"
 563         "movdqa %%xmm5, 32(%1)                  \n\t"
 564         "movdqa %%xmm0, 48(%1)                  \n\t"
 565         "add $32, %0                            \n\t"
 566         "add $64, %1                            \n\t"
 567         "add $32, %2                            \n\t"
 568         "cmp %3, %0                             \n\t"
 569             " jb 1b                             \n\t"
 570         : "+r" (block), "+r" (sum), "+r" (offset)
 571         : "r"(block+64)
 572           XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
 573                             "%xmm4", "%xmm5", "%xmm6", "%xmm7")
 574     );
 575 }
 576
 577 #endif /* HAVE_INLINE_ASM */
 578
 579 void ff_MPV_common_init_x86(MpegEncContext *s)
 580 {
 581 #if HAVE_INLINE_ASM
 582     int mm_flags = av_get_cpu_flags();
 583
 584     if (mm_flags & AV_CPU_FLAG_MMX) {
 585         s->dct_unquantize_h263_intra = dct_unquantize_h263_intra_mmx;
 586         s->dct_unquantize_h263_inter = dct_unquantize_h263_inter_mmx;
 587         s->dct_unquantize_mpeg1_intra = dct_unquantize_mpeg1_intra_mmx;
 588         s->dct_unquantize_mpeg1_inter = dct_unquantize_mpeg1_inter_mmx;
 589         if(!(s->flags & CODEC_FLAG_BITEXACT))
 590             s->dct_unquantize_mpeg2_intra = dct_unquantize_mpeg2_intra_mmx;
 591         s->dct_unquantize_mpeg2_inter = dct_unquantize_mpeg2_inter_mmx;
 592
 593         if (mm_flags & AV_CPU_FLAG_SSE2) {
 594             s->denoise_dct= denoise_dct_sse2;
 595         } else {
 596                 s->denoise_dct= denoise_dct_mmx;
 597         }
 598     }
 599 #endif /* HAVE_INLINE_ASM */
 600 }