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/attributes.h"
  23 #include "libavutil/cpu.h"
  24 #include "libavutil/x86/asm.h"
  25 #include "libavcodec/avcodec.h"
  26 #include "libavcodec/dsputil.h"
  27 #include "libavcodec/mpegvideo.h"
  28 #include "dsputil_mmx.h"
  29
  30 #if HAVE_INLINE_ASM
  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->inter_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     assert(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
 169 /*
 170   We can suppose that result of two multiplications can't be greater than 0xFFFF
 171   i.e. is 16-bit, so we use here only PMULLW instruction and can avoid
 172   a complex multiplication.
 173 =====================================================
 174  Full formula for multiplication of 2 integer numbers
 175  which are represent as high:low words:
 176  input: value1 = high1:low1
 177         value2 = high2:low2
 178  output: value3 = value1*value2
 179  value3=high3:low3 (on overflow: modulus 2^32 wrap-around)
 180  this mean that for 0x123456 * 0x123456 correct result is 0x766cb0ce4
 181  but this algorithm will compute only 0x66cb0ce4
 182  this limited by 16-bit size of operands
 183  ---------------------------------
 184  tlow1 = high1*low2
 185  tlow2 = high2*low1
 186  tlow1 = tlow1 + tlow2
 187  high3:low3 = low1*low2
 188  high3 += tlow1
 189 */
 190 static void dct_unquantize_mpeg1_intra_mmx(MpegEncContext *s,
 191                                      int16_t *block, int n, int qscale)
 192 {
 193     x86_reg nCoeffs;
 194     const uint16_t *quant_matrix;
 195     int block0;
 196
 197     av_assert2(s->block_last_index[n]>=0);
 198
 199     nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]+1;
 200
 201     if (n < 4)
 202         block0 = block[0] * s->y_dc_scale;
 203     else
 204         block0 = block[0] * s->c_dc_scale;
 205     /* XXX: only mpeg1 */
 206     quant_matrix = s->intra_matrix;
 207 __asm__ volatile(
 208                 "pcmpeqw %%mm7, %%mm7           \n\t"
 209                 "psrlw $15, %%mm7               \n\t"
 210                 "movd %2, %%mm6                 \n\t"
 211                 "packssdw %%mm6, %%mm6          \n\t"
 212                 "packssdw %%mm6, %%mm6          \n\t"
 213                 "mov %3, %%"REG_a"              \n\t"
 214                 ".p2align 4                     \n\t"
 215                 "1:                             \n\t"
 216                 "movq (%0, %%"REG_a"), %%mm0    \n\t"
 217                 "movq 8(%0, %%"REG_a"), %%mm1   \n\t"
 218                 "movq (%1, %%"REG_a"), %%mm4    \n\t"
 219                 "movq 8(%1, %%"REG_a"), %%mm5   \n\t"
 220                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
 221                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
 222                 "pxor %%mm2, %%mm2              \n\t"
 223                 "pxor %%mm3, %%mm3              \n\t"
 224                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
 225                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
 226                 "pxor %%mm2, %%mm0              \n\t"
 227                 "pxor %%mm3, %%mm1              \n\t"
 228                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
 229                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
 230                 "pmullw %%mm4, %%mm0            \n\t" // abs(block[i])*q
 231                 "pmullw %%mm5, %%mm1            \n\t" // abs(block[i])*q
 232                 "pxor %%mm4, %%mm4              \n\t"
 233                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
 234                 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
 235                 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
 236                 "psraw $3, %%mm0                \n\t"
 237                 "psraw $3, %%mm1                \n\t"
 238                 "psubw %%mm7, %%mm0             \n\t"
 239                 "psubw %%mm7, %%mm1             \n\t"
 240                 "por %%mm7, %%mm0               \n\t"
 241                 "por %%mm7, %%mm1               \n\t"
 242                 "pxor %%mm2, %%mm0              \n\t"
 243                 "pxor %%mm3, %%mm1              \n\t"
 244                 "psubw %%mm2, %%mm0             \n\t"
 245                 "psubw %%mm3, %%mm1             \n\t"
 246                 "pandn %%mm0, %%mm4             \n\t"
 247                 "pandn %%mm1, %%mm5             \n\t"
 248                 "movq %%mm4, (%0, %%"REG_a")    \n\t"
 249                 "movq %%mm5, 8(%0, %%"REG_a")   \n\t"
 250
 251                 "add $16, %%"REG_a"             \n\t"
 252                 "js 1b                          \n\t"
 253                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "g" (-2*nCoeffs)
 254                 : "%"REG_a, "memory"
 255         );
 256     block[0]= block0;
 257 }
 258
 259 static void dct_unquantize_mpeg1_inter_mmx(MpegEncContext *s,
 260                                      int16_t *block, int n, int qscale)
 261 {
 262     x86_reg nCoeffs;
 263     const uint16_t *quant_matrix;
 264
 265     av_assert2(s->block_last_index[n]>=0);
 266
 267     nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]+1;
 268
 269         quant_matrix = s->inter_matrix;
 270 __asm__ volatile(
 271                 "pcmpeqw %%mm7, %%mm7           \n\t"
 272                 "psrlw $15, %%mm7               \n\t"
 273                 "movd %2, %%mm6                 \n\t"
 274                 "packssdw %%mm6, %%mm6          \n\t"
 275                 "packssdw %%mm6, %%mm6          \n\t"
 276                 "mov %3, %%"REG_a"              \n\t"
 277                 ".p2align 4                     \n\t"
 278                 "1:                             \n\t"
 279                 "movq (%0, %%"REG_a"), %%mm0    \n\t"
 280                 "movq 8(%0, %%"REG_a"), %%mm1   \n\t"
 281                 "movq (%1, %%"REG_a"), %%mm4    \n\t"
 282                 "movq 8(%1, %%"REG_a"), %%mm5   \n\t"
 283                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
 284                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
 285                 "pxor %%mm2, %%mm2              \n\t"
 286                 "pxor %%mm3, %%mm3              \n\t"
 287                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
 288                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
 289                 "pxor %%mm2, %%mm0              \n\t"
 290                 "pxor %%mm3, %%mm1              \n\t"
 291                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
 292                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
 293                 "paddw %%mm0, %%mm0             \n\t" // abs(block[i])*2
 294                 "paddw %%mm1, %%mm1             \n\t" // abs(block[i])*2
 295                 "paddw %%mm7, %%mm0             \n\t" // abs(block[i])*2 + 1
 296                 "paddw %%mm7, %%mm1             \n\t" // abs(block[i])*2 + 1
 297                 "pmullw %%mm4, %%mm0            \n\t" // (abs(block[i])*2 + 1)*q
 298                 "pmullw %%mm5, %%mm1            \n\t" // (abs(block[i])*2 + 1)*q
 299                 "pxor %%mm4, %%mm4              \n\t"
 300                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
 301                 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
 302                 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
 303                 "psraw $4, %%mm0                \n\t"
 304                 "psraw $4, %%mm1                \n\t"
 305                 "psubw %%mm7, %%mm0             \n\t"
 306                 "psubw %%mm7, %%mm1             \n\t"
 307                 "por %%mm7, %%mm0               \n\t"
 308                 "por %%mm7, %%mm1               \n\t"
 309                 "pxor %%mm2, %%mm0              \n\t"
 310                 "pxor %%mm3, %%mm1              \n\t"
 311                 "psubw %%mm2, %%mm0             \n\t"
 312                 "psubw %%mm3, %%mm1             \n\t"
 313                 "pandn %%mm0, %%mm4             \n\t"
 314                 "pandn %%mm1, %%mm5             \n\t"
 315                 "movq %%mm4, (%0, %%"REG_a")    \n\t"
 316                 "movq %%mm5, 8(%0, %%"REG_a")   \n\t"
 317
 318                 "add $16, %%"REG_a"             \n\t"
 319                 "js 1b                          \n\t"
 320                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "g" (-2*nCoeffs)
 321                 : "%"REG_a, "memory"
 322         );
 323 }
 324
 325 static void dct_unquantize_mpeg2_intra_mmx(MpegEncContext *s,
 326                                      int16_t *block, int n, int qscale)
 327 {
 328     x86_reg nCoeffs;
 329     const uint16_t *quant_matrix;
 330     int block0;
 331
 332     av_assert2(s->block_last_index[n]>=0);
 333
 334     if(s->alternate_scan) nCoeffs= 63; //FIXME
 335     else nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
 336
 337     if (n < 4)
 338         block0 = block[0] * s->y_dc_scale;
 339     else
 340         block0 = block[0] * s->c_dc_scale;
 341     quant_matrix = s->intra_matrix;
 342 __asm__ volatile(
 343                 "pcmpeqw %%mm7, %%mm7           \n\t"
 344                 "psrlw $15, %%mm7               \n\t"
 345                 "movd %2, %%mm6                 \n\t"
 346                 "packssdw %%mm6, %%mm6          \n\t"
 347                 "packssdw %%mm6, %%mm6          \n\t"
 348                 "mov %3, %%"REG_a"              \n\t"
 349                 ".p2align 4                     \n\t"
 350                 "1:                             \n\t"
 351                 "movq (%0, %%"REG_a"), %%mm0    \n\t"
 352                 "movq 8(%0, %%"REG_a"), %%mm1   \n\t"
 353                 "movq (%1, %%"REG_a"), %%mm4    \n\t"
 354                 "movq 8(%1, %%"REG_a"), %%mm5   \n\t"
 355                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
 356                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
 357                 "pxor %%mm2, %%mm2              \n\t"
 358                 "pxor %%mm3, %%mm3              \n\t"
 359                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
 360                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
 361                 "pxor %%mm2, %%mm0              \n\t"
 362                 "pxor %%mm3, %%mm1              \n\t"
 363                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
 364                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
 365                 "pmullw %%mm4, %%mm0            \n\t" // abs(block[i])*q
 366                 "pmullw %%mm5, %%mm1            \n\t" // abs(block[i])*q
 367                 "pxor %%mm4, %%mm4              \n\t"
 368                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
 369                 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
 370                 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
 371                 "psraw $3, %%mm0                \n\t"
 372                 "psraw $3, %%mm1                \n\t"
 373                 "pxor %%mm2, %%mm0              \n\t"
 374                 "pxor %%mm3, %%mm1              \n\t"
 375                 "psubw %%mm2, %%mm0             \n\t"
 376                 "psubw %%mm3, %%mm1             \n\t"
 377                 "pandn %%mm0, %%mm4             \n\t"
 378                 "pandn %%mm1, %%mm5             \n\t"
 379                 "movq %%mm4, (%0, %%"REG_a")    \n\t"
 380                 "movq %%mm5, 8(%0, %%"REG_a")   \n\t"
 381
 382                 "add $16, %%"REG_a"             \n\t"
 383                 "jng 1b                         \n\t"
 384                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "g" (-2*nCoeffs)
 385                 : "%"REG_a, "memory"
 386         );
 387     block[0]= block0;
 388         //Note, we do not do mismatch control for intra as errors cannot accumulate
 389 }
 390
 391 static void dct_unquantize_mpeg2_inter_mmx(MpegEncContext *s,
 392                                      int16_t *block, int n, int qscale)
 393 {
 394     x86_reg nCoeffs;
 395     const uint16_t *quant_matrix;
 396
 397     av_assert2(s->block_last_index[n]>=0);
 398
 399     if(s->alternate_scan) nCoeffs= 63; //FIXME
 400     else nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
 401
 402         quant_matrix = s->inter_matrix;
 403 __asm__ volatile(
 404                 "pcmpeqw %%mm7, %%mm7           \n\t"
 405                 "psrlq $48, %%mm7               \n\t"
 406                 "movd %2, %%mm6                 \n\t"
 407                 "packssdw %%mm6, %%mm6          \n\t"
 408                 "packssdw %%mm6, %%mm6          \n\t"
 409                 "mov %3, %%"REG_a"              \n\t"
 410                 ".p2align 4                     \n\t"
 411                 "1:                             \n\t"
 412                 "movq (%0, %%"REG_a"), %%mm0    \n\t"
 413                 "movq 8(%0, %%"REG_a"), %%mm1   \n\t"
 414                 "movq (%1, %%"REG_a"), %%mm4    \n\t"
 415                 "movq 8(%1, %%"REG_a"), %%mm5   \n\t"
 416                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
 417                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
 418                 "pxor %%mm2, %%mm2              \n\t"
 419                 "pxor %%mm3, %%mm3              \n\t"
 420                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
 421                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
 422                 "pxor %%mm2, %%mm0              \n\t"
 423                 "pxor %%mm3, %%mm1              \n\t"
 424                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
 425                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
 426                 "paddw %%mm0, %%mm0             \n\t" // abs(block[i])*2
 427                 "paddw %%mm1, %%mm1             \n\t" // abs(block[i])*2
 428                 "pmullw %%mm4, %%mm0            \n\t" // abs(block[i])*2*q
 429                 "pmullw %%mm5, %%mm1            \n\t" // abs(block[i])*2*q
 430                 "paddw %%mm4, %%mm0             \n\t" // (abs(block[i])*2 + 1)*q
 431                 "paddw %%mm5, %%mm1             \n\t" // (abs(block[i])*2 + 1)*q
 432                 "pxor %%mm4, %%mm4              \n\t"
 433                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
 434                 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
 435                 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
 436                 "psrlw $4, %%mm0                \n\t"
 437                 "psrlw $4, %%mm1                \n\t"
 438                 "pxor %%mm2, %%mm0              \n\t"
 439                 "pxor %%mm3, %%mm1              \n\t"
 440                 "psubw %%mm2, %%mm0             \n\t"
 441                 "psubw %%mm3, %%mm1             \n\t"
 442                 "pandn %%mm0, %%mm4             \n\t"
 443                 "pandn %%mm1, %%mm5             \n\t"
 444                 "pxor %%mm4, %%mm7              \n\t"
 445                 "pxor %%mm5, %%mm7              \n\t"
 446                 "movq %%mm4, (%0, %%"REG_a")    \n\t"
 447                 "movq %%mm5, 8(%0, %%"REG_a")   \n\t"
 448
 449                 "add $16, %%"REG_a"             \n\t"
 450                 "jng 1b                         \n\t"
 451                 "movd 124(%0, %3), %%mm0        \n\t"
 452                 "movq %%mm7, %%mm6              \n\t"
 453                 "psrlq $32, %%mm7               \n\t"
 454                 "pxor %%mm6, %%mm7              \n\t"
 455                 "movq %%mm7, %%mm6              \n\t"
 456                 "psrlq $16, %%mm7               \n\t"
 457                 "pxor %%mm6, %%mm7              \n\t"
 458                 "pslld $31, %%mm7               \n\t"
 459                 "psrlq $15, %%mm7               \n\t"
 460                 "pxor %%mm7, %%mm0              \n\t"
 461                 "movd %%mm0, 124(%0, %3)        \n\t"
 462
 463                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "r" (-2*nCoeffs)
 464                 : "%"REG_a, "memory"
 465         );
 466 }
 467
 468 static void  denoise_dct_mmx(MpegEncContext *s, int16_t *block){
 469     const int intra= s->mb_intra;
 470     int *sum= s->dct_error_sum[intra];
 471     uint16_t *offset= s->dct_offset[intra];
 472
 473     s->dct_count[intra]++;
 474
 475     __asm__ volatile(
 476         "pxor %%mm7, %%mm7                      \n\t"
 477         "1:                                     \n\t"
 478         "pxor %%mm0, %%mm0                      \n\t"
 479         "pxor %%mm1, %%mm1                      \n\t"
 480         "movq (%0), %%mm2                       \n\t"
 481         "movq 8(%0), %%mm3                      \n\t"
 482         "pcmpgtw %%mm2, %%mm0                   \n\t"
 483         "pcmpgtw %%mm3, %%mm1                   \n\t"
 484         "pxor %%mm0, %%mm2                      \n\t"
 485         "pxor %%mm1, %%mm3                      \n\t"
 486         "psubw %%mm0, %%mm2                     \n\t"
 487         "psubw %%mm1, %%mm3                     \n\t"
 488         "movq %%mm2, %%mm4                      \n\t"
 489         "movq %%mm3, %%mm5                      \n\t"
 490         "psubusw (%2), %%mm2                    \n\t"
 491         "psubusw 8(%2), %%mm3                   \n\t"
 492         "pxor %%mm0, %%mm2                      \n\t"
 493         "pxor %%mm1, %%mm3                      \n\t"
 494         "psubw %%mm0, %%mm2                     \n\t"
 495         "psubw %%mm1, %%mm3                     \n\t"
 496         "movq %%mm2, (%0)                       \n\t"
 497         "movq %%mm3, 8(%0)                      \n\t"
 498         "movq %%mm4, %%mm2                      \n\t"
 499         "movq %%mm5, %%mm3                      \n\t"
 500         "punpcklwd %%mm7, %%mm4                 \n\t"
 501         "punpckhwd %%mm7, %%mm2                 \n\t"
 502         "punpcklwd %%mm7, %%mm5                 \n\t"
 503         "punpckhwd %%mm7, %%mm3                 \n\t"
 504         "paddd (%1), %%mm4                      \n\t"
 505         "paddd 8(%1), %%mm2                     \n\t"
 506         "paddd 16(%1), %%mm5                    \n\t"
 507         "paddd 24(%1), %%mm3                    \n\t"
 508         "movq %%mm4, (%1)                       \n\t"
 509         "movq %%mm2, 8(%1)                      \n\t"
 510         "movq %%mm5, 16(%1)                     \n\t"
 511         "movq %%mm3, 24(%1)                     \n\t"
 512         "add $16, %0                            \n\t"
 513         "add $32, %1                            \n\t"
 514         "add $16, %2                            \n\t"
 515         "cmp %3, %0                             \n\t"
 516             " jb 1b                             \n\t"
 517         : "+r" (block), "+r" (sum), "+r" (offset)
 518         : "r"(block+64)
 519     );
 520 }
 521
 522 static void  denoise_dct_sse2(MpegEncContext *s, int16_t *block){
 523     const int intra= s->mb_intra;
 524     int *sum= s->dct_error_sum[intra];
 525     uint16_t *offset= s->dct_offset[intra];
 526
 527     s->dct_count[intra]++;
 528
 529     __asm__ volatile(
 530         "pxor %%xmm7, %%xmm7                    \n\t"
 531         "1:                                     \n\t"
 532         "pxor %%xmm0, %%xmm0                    \n\t"
 533         "pxor %%xmm1, %%xmm1                    \n\t"
 534         "movdqa (%0), %%xmm2                    \n\t"
 535         "movdqa 16(%0), %%xmm3                  \n\t"
 536         "pcmpgtw %%xmm2, %%xmm0                 \n\t"
 537         "pcmpgtw %%xmm3, %%xmm1                 \n\t"
 538         "pxor %%xmm0, %%xmm2                    \n\t"
 539         "pxor %%xmm1, %%xmm3                    \n\t"
 540         "psubw %%xmm0, %%xmm2                   \n\t"
 541         "psubw %%xmm1, %%xmm3                   \n\t"
 542         "movdqa %%xmm2, %%xmm4                  \n\t"
 543         "movdqa %%xmm3, %%xmm5                  \n\t"
 544         "psubusw (%2), %%xmm2                   \n\t"
 545         "psubusw 16(%2), %%xmm3                 \n\t"
 546         "pxor %%xmm0, %%xmm2                    \n\t"
 547         "pxor %%xmm1, %%xmm3                    \n\t"
 548         "psubw %%xmm0, %%xmm2                   \n\t"
 549         "psubw %%xmm1, %%xmm3                   \n\t"
 550         "movdqa %%xmm2, (%0)                    \n\t"
 551         "movdqa %%xmm3, 16(%0)                  \n\t"
 552         "movdqa %%xmm4, %%xmm6                  \n\t"
 553         "movdqa %%xmm5, %%xmm0                  \n\t"
 554         "punpcklwd %%xmm7, %%xmm4               \n\t"
 555         "punpckhwd %%xmm7, %%xmm6               \n\t"
 556         "punpcklwd %%xmm7, %%xmm5               \n\t"
 557         "punpckhwd %%xmm7, %%xmm0               \n\t"
 558         "paddd (%1), %%xmm4                     \n\t"
 559         "paddd 16(%1), %%xmm6                   \n\t"
 560         "paddd 32(%1), %%xmm5                   \n\t"
 561         "paddd 48(%1), %%xmm0                   \n\t"
 562         "movdqa %%xmm4, (%1)                    \n\t"
 563         "movdqa %%xmm6, 16(%1)                  \n\t"
 564         "movdqa %%xmm5, 32(%1)                  \n\t"
 565         "movdqa %%xmm0, 48(%1)                  \n\t"
 566         "add $32, %0                            \n\t"
 567         "add $64, %1                            \n\t"
 568         "add $32, %2                            \n\t"
 569         "cmp %3, %0                             \n\t"
 570             " jb 1b                             \n\t"
 571         : "+r" (block), "+r" (sum), "+r" (offset)
 572         : "r"(block+64)
 573           XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
 574                             "%xmm4", "%xmm5", "%xmm6", "%xmm7")
 575     );
 576 }
 577
 578 #endif /* HAVE_INLINE_ASM */
 579
 580 av_cold void ff_MPV_common_init_x86(MpegEncContext *s)
 581 {
 582 #if HAVE_INLINE_ASM
 583     int mm_flags = av_get_cpu_flags();
 584
 585     if (mm_flags & AV_CPU_FLAG_MMX) {
 586         s->dct_unquantize_h263_intra = dct_unquantize_h263_intra_mmx;
 587         s->dct_unquantize_h263_inter = dct_unquantize_h263_inter_mmx;
 588         s->dct_unquantize_mpeg1_intra = dct_unquantize_mpeg1_intra_mmx;
 589         s->dct_unquantize_mpeg1_inter = dct_unquantize_mpeg1_inter_mmx;
 590         if(!(s->flags & CODEC_FLAG_BITEXACT))
 591             s->dct_unquantize_mpeg2_intra = dct_unquantize_mpeg2_intra_mmx;
 592         s->dct_unquantize_mpeg2_inter = dct_unquantize_mpeg2_inter_mmx;
 593
 594         if (mm_flags & AV_CPU_FLAG_SSE2) {
 595             s->denoise_dct= denoise_dct_sse2;
 596         } else {
 597                 s->denoise_dct= denoise_dct_mmx;
 598         }
 599     }
 600 #endif /* HAVE_INLINE_ASM */
 601 }