git.sesse.net Git - ffmpeg/blob - libavcodec/i386/mpegvideo_mmx.c

   1 /*
   2  * The simplest mpeg encoder (well, it was the simplest!)
   3  * Copyright (c) 2000,2001 Fabrice Bellard.
   4  *
   5  * This library is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU Lesser General Public
   7  * License as published by the Free Software Foundation; either
   8  * version 2 of the License, or (at your option) any later version.
   9  *
  10  * This library is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * Lesser General Public License for more details.
  14  *
  15  * You should have received a copy of the GNU Lesser General Public
  16  * License along with this library; if not, write to the Free Software
  17  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  18  *
  19  * Optimized for ia32 cpus by Nick Kurshev <nickols_k@mail.ru>
  20  * h263, mpeg1, mpeg2 dequantizer & draw_edges by Michael Niedermayer <michaelni@gmx.at>
  21  */
  22
  23 #include "../dsputil.h"
  24 #include "../mpegvideo.h"
  25 #include "../avcodec.h"
  26 #include "x86_cpu.h"
  27
  28 extern uint8_t zigzag_direct_noperm[64];
  29 extern uint16_t inv_zigzag_direct16[64];
  30
  31 static const unsigned long long int mm_wabs __attribute__ ((aligned(8))) = 0xffffffffffffffffULL;
  32 static const unsigned long long int mm_wone __attribute__ ((aligned(8))) = 0x0001000100010001ULL;
  33
  34
  35 static void dct_unquantize_h263_intra_mmx(MpegEncContext *s,
  36                                   DCTELEM *block, int n, int qscale)
  37 {
  38     long level, qmul, qadd, nCoeffs;
  39
  40     qmul = qscale << 1;
  41
  42     assert(s->block_last_index[n]>=0 || s->h263_aic);
  43
  44     if (!s->h263_aic) {
  45         if (n < 4)
  46             level = block[0] * s->y_dc_scale;
  47         else
  48             level = block[0] * s->c_dc_scale;
  49         qadd = (qscale - 1) | 1;
  50     }else{
  51         qadd = 0;
  52         level= block[0];
  53     }
  54     if(s->ac_pred)
  55         nCoeffs=63;
  56     else
  57         nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ];
  58 //printf("%d %d  ", qmul, qadd);
  59 asm volatile(
  60                 "movd %1, %%mm6                 \n\t" //qmul
  61                 "packssdw %%mm6, %%mm6          \n\t"
  62                 "packssdw %%mm6, %%mm6          \n\t"
  63                 "movd %2, %%mm5                 \n\t" //qadd
  64                 "pxor %%mm7, %%mm7              \n\t"
  65                 "packssdw %%mm5, %%mm5          \n\t"
  66                 "packssdw %%mm5, %%mm5          \n\t"
  67                 "psubw %%mm5, %%mm7             \n\t"
  68                 "pxor %%mm4, %%mm4              \n\t"
  69                 ".balign 16                     \n\t"
  70                 "1:                             \n\t"
  71                 "movq (%0, %3), %%mm0           \n\t"
  72                 "movq 8(%0, %3), %%mm1          \n\t"
  73
  74                 "pmullw %%mm6, %%mm0            \n\t"
  75                 "pmullw %%mm6, %%mm1            \n\t"
  76
  77                 "movq (%0, %3), %%mm2           \n\t"
  78                 "movq 8(%0, %3), %%mm3          \n\t"
  79
  80                 "pcmpgtw %%mm4, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
  81                 "pcmpgtw %%mm4, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
  82
  83                 "pxor %%mm2, %%mm0              \n\t"
  84                 "pxor %%mm3, %%mm1              \n\t"
  85
  86                 "paddw %%mm7, %%mm0             \n\t"
  87                 "paddw %%mm7, %%mm1             \n\t"
  88
  89                 "pxor %%mm0, %%mm2              \n\t"
  90                 "pxor %%mm1, %%mm3              \n\t"
  91
  92                 "pcmpeqw %%mm7, %%mm0           \n\t" // block[i] == 0 ? -1 : 0
  93                 "pcmpeqw %%mm7, %%mm1           \n\t" // block[i] == 0 ? -1 : 0
  94
  95                 "pandn %%mm2, %%mm0             \n\t"
  96                 "pandn %%mm3, %%mm1             \n\t"
  97
  98                 "movq %%mm0, (%0, %3)           \n\t"
  99                 "movq %%mm1, 8(%0, %3)          \n\t"
 100
 101                 "add $16, %3                    \n\t"
 102                 "jng 1b                         \n\t"
 103                 ::"r" (block+nCoeffs), "g"(qmul), "g" (qadd), "r" (2*(-nCoeffs))
 104                 : "memory"
 105         );
 106         block[0]= level;
 107 }
 108
 109
 110 static void dct_unquantize_h263_inter_mmx(MpegEncContext *s,
 111                                   DCTELEM *block, int n, int qscale)
 112 {
 113     long qmul, qadd, nCoeffs;
 114
 115     qmul = qscale << 1;
 116     qadd = (qscale - 1) | 1;
 117
 118     assert(s->block_last_index[n]>=0 || s->h263_aic);
 119
 120     nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ];
 121 //printf("%d %d  ", qmul, qadd);
 122 asm volatile(
 123                 "movd %1, %%mm6                 \n\t" //qmul
 124                 "packssdw %%mm6, %%mm6          \n\t"
 125                 "packssdw %%mm6, %%mm6          \n\t"
 126                 "movd %2, %%mm5                 \n\t" //qadd
 127                 "pxor %%mm7, %%mm7              \n\t"
 128                 "packssdw %%mm5, %%mm5          \n\t"
 129                 "packssdw %%mm5, %%mm5          \n\t"
 130                 "psubw %%mm5, %%mm7             \n\t"
 131                 "pxor %%mm4, %%mm4              \n\t"
 132                 ".balign 16                     \n\t"
 133                 "1:                             \n\t"
 134                 "movq (%0, %3), %%mm0           \n\t"
 135                 "movq 8(%0, %3), %%mm1          \n\t"
 136
 137                 "pmullw %%mm6, %%mm0            \n\t"
 138                 "pmullw %%mm6, %%mm1            \n\t"
 139
 140                 "movq (%0, %3), %%mm2           \n\t"
 141                 "movq 8(%0, %3), %%mm3          \n\t"
 142
 143                 "pcmpgtw %%mm4, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
 144                 "pcmpgtw %%mm4, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
 145
 146                 "pxor %%mm2, %%mm0              \n\t"
 147                 "pxor %%mm3, %%mm1              \n\t"
 148
 149                 "paddw %%mm7, %%mm0             \n\t"
 150                 "paddw %%mm7, %%mm1             \n\t"
 151
 152                 "pxor %%mm0, %%mm2              \n\t"
 153                 "pxor %%mm1, %%mm3              \n\t"
 154
 155                 "pcmpeqw %%mm7, %%mm0           \n\t" // block[i] == 0 ? -1 : 0
 156                 "pcmpeqw %%mm7, %%mm1           \n\t" // block[i] == 0 ? -1 : 0
 157
 158                 "pandn %%mm2, %%mm0             \n\t"
 159                 "pandn %%mm3, %%mm1             \n\t"
 160
 161                 "movq %%mm0, (%0, %3)           \n\t"
 162                 "movq %%mm1, 8(%0, %3)          \n\t"
 163
 164                 "add $16, %3                    \n\t"
 165                 "jng 1b                         \n\t"
 166                 ::"r" (block+nCoeffs), "g"(qmul), "g" (qadd), "r" (2*(-nCoeffs))
 167                 : "memory"
 168         );
 169 }
 170
 171
 172 /*
 173   NK:
 174   Note: looking at PARANOID:
 175   "enable all paranoid tests for rounding, overflows, etc..."
 176
 177 #ifdef PARANOID
 178                 if (level < -2048 || level > 2047)
 179                     fprintf(stderr, "unquant error %d %d\n", i, level);
 180 #endif
 181   We can suppose that result of two multiplications can't be greate of 0xFFFF
 182   i.e. is 16-bit, so we use here only PMULLW instruction and can avoid
 183   a complex multiplication.
 184 =====================================================
 185  Full formula for multiplication of 2 integer numbers
 186  which are represent as high:low words:
 187  input: value1 = high1:low1
 188         value2 = high2:low2
 189  output: value3 = value1*value2
 190  value3=high3:low3 (on overflow: modulus 2^32 wrap-around)
 191  this mean that for 0x123456 * 0x123456 correct result is 0x766cb0ce4
 192  but this algorithm will compute only 0x66cb0ce4
 193  this limited by 16-bit size of operands
 194  ---------------------------------
 195  tlow1 = high1*low2
 196  tlow2 = high2*low1
 197  tlow1 = tlow1 + tlow2
 198  high3:low3 = low1*low2
 199  high3 += tlow1
 200 */
 201 static void dct_unquantize_mpeg1_intra_mmx(MpegEncContext *s,
 202                                      DCTELEM *block, int n, int qscale)
 203 {
 204     long nCoeffs;
 205     const uint16_t *quant_matrix;
 206     int block0;
 207
 208     assert(s->block_last_index[n]>=0);
 209
 210     nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]+1;
 211
 212     if (n < 4)
 213         block0 = block[0] * s->y_dc_scale;
 214     else
 215         block0 = block[0] * s->c_dc_scale;
 216     /* XXX: only mpeg1 */
 217     quant_matrix = s->intra_matrix;
 218 asm volatile(
 219                 "pcmpeqw %%mm7, %%mm7           \n\t"
 220                 "psrlw $15, %%mm7               \n\t"
 221                 "movd %2, %%mm6                 \n\t"
 222                 "packssdw %%mm6, %%mm6          \n\t"
 223                 "packssdw %%mm6, %%mm6          \n\t"
 224                 "mov %3, %%"REG_a"              \n\t"
 225                 ".balign 16                     \n\t"
 226                 "1:                             \n\t"
 227                 "movq (%0, %%"REG_a"), %%mm0    \n\t"
 228                 "movq 8(%0, %%"REG_a"), %%mm1   \n\t"
 229                 "movq (%1, %%"REG_a"), %%mm4    \n\t"
 230                 "movq 8(%1, %%"REG_a"), %%mm5   \n\t"
 231                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
 232                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
 233                 "pxor %%mm2, %%mm2              \n\t"
 234                 "pxor %%mm3, %%mm3              \n\t"
 235                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
 236                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
 237                 "pxor %%mm2, %%mm0              \n\t"
 238                 "pxor %%mm3, %%mm1              \n\t"
 239                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
 240                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
 241                 "pmullw %%mm4, %%mm0            \n\t" // abs(block[i])*q
 242                 "pmullw %%mm5, %%mm1            \n\t" // abs(block[i])*q
 243                 "pxor %%mm4, %%mm4              \n\t"
 244                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
 245                 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
 246                 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
 247                 "psraw $3, %%mm0                \n\t"
 248                 "psraw $3, %%mm1                \n\t"
 249                 "psubw %%mm7, %%mm0             \n\t"
 250                 "psubw %%mm7, %%mm1             \n\t"
 251                 "por %%mm7, %%mm0               \n\t"
 252                 "por %%mm7, %%mm1               \n\t"
 253                 "pxor %%mm2, %%mm0              \n\t"
 254                 "pxor %%mm3, %%mm1              \n\t"
 255                 "psubw %%mm2, %%mm0             \n\t"
 256                 "psubw %%mm3, %%mm1             \n\t"
 257                 "pandn %%mm0, %%mm4             \n\t"
 258                 "pandn %%mm1, %%mm5             \n\t"
 259                 "movq %%mm4, (%0, %%"REG_a")    \n\t"
 260                 "movq %%mm5, 8(%0, %%"REG_a")   \n\t"
 261
 262                 "add $16, %%"REG_a"             \n\t"
 263                 "js 1b                          \n\t"
 264                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "g" (-2*nCoeffs)
 265                 : "%"REG_a, "memory"
 266         );
 267     block[0]= block0;
 268 }
 269
 270 static void dct_unquantize_mpeg1_inter_mmx(MpegEncContext *s,
 271                                      DCTELEM *block, int n, int qscale)
 272 {
 273     long nCoeffs;
 274     const uint16_t *quant_matrix;
 275
 276     assert(s->block_last_index[n]>=0);
 277
 278     nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]+1;
 279
 280         quant_matrix = s->inter_matrix;
 281 asm volatile(
 282                 "pcmpeqw %%mm7, %%mm7           \n\t"
 283                 "psrlw $15, %%mm7               \n\t"
 284                 "movd %2, %%mm6                 \n\t"
 285                 "packssdw %%mm6, %%mm6          \n\t"
 286                 "packssdw %%mm6, %%mm6          \n\t"
 287                 "mov %3, %%"REG_a"              \n\t"
 288                 ".balign 16                     \n\t"
 289                 "1:                             \n\t"
 290                 "movq (%0, %%"REG_a"), %%mm0    \n\t"
 291                 "movq 8(%0, %%"REG_a"), %%mm1   \n\t"
 292                 "movq (%1, %%"REG_a"), %%mm4    \n\t"
 293                 "movq 8(%1, %%"REG_a"), %%mm5   \n\t"
 294                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
 295                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
 296                 "pxor %%mm2, %%mm2              \n\t"
 297                 "pxor %%mm3, %%mm3              \n\t"
 298                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
 299                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
 300                 "pxor %%mm2, %%mm0              \n\t"
 301                 "pxor %%mm3, %%mm1              \n\t"
 302                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
 303                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
 304                 "paddw %%mm0, %%mm0             \n\t" // abs(block[i])*2
 305                 "paddw %%mm1, %%mm1             \n\t" // abs(block[i])*2
 306                 "paddw %%mm7, %%mm0             \n\t" // abs(block[i])*2 + 1
 307                 "paddw %%mm7, %%mm1             \n\t" // abs(block[i])*2 + 1
 308                 "pmullw %%mm4, %%mm0            \n\t" // (abs(block[i])*2 + 1)*q
 309                 "pmullw %%mm5, %%mm1            \n\t" // (abs(block[i])*2 + 1)*q
 310                 "pxor %%mm4, %%mm4              \n\t"
 311                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
 312                 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
 313                 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
 314                 "psraw $4, %%mm0                \n\t"
 315                 "psraw $4, %%mm1                \n\t"
 316                 "psubw %%mm7, %%mm0             \n\t"
 317                 "psubw %%mm7, %%mm1             \n\t"
 318                 "por %%mm7, %%mm0               \n\t"
 319                 "por %%mm7, %%mm1               \n\t"
 320                 "pxor %%mm2, %%mm0              \n\t"
 321                 "pxor %%mm3, %%mm1              \n\t"
 322                 "psubw %%mm2, %%mm0             \n\t"
 323                 "psubw %%mm3, %%mm1             \n\t"
 324                 "pandn %%mm0, %%mm4             \n\t"
 325                 "pandn %%mm1, %%mm5             \n\t"
 326                 "movq %%mm4, (%0, %%"REG_a")    \n\t"
 327                 "movq %%mm5, 8(%0, %%"REG_a")   \n\t"
 328
 329                 "add $16, %%"REG_a"             \n\t"
 330                 "js 1b                          \n\t"
 331                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "g" (-2*nCoeffs)
 332                 : "%"REG_a, "memory"
 333         );
 334 }
 335
 336 static void dct_unquantize_mpeg2_intra_mmx(MpegEncContext *s,
 337                                      DCTELEM *block, int n, int qscale)
 338 {
 339     long nCoeffs;
 340     const uint16_t *quant_matrix;
 341     int block0;
 342
 343     assert(s->block_last_index[n]>=0);
 344
 345     if(s->alternate_scan) nCoeffs= 63; //FIXME
 346     else nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
 347
 348     if (n < 4)
 349         block0 = block[0] * s->y_dc_scale;
 350     else
 351         block0 = block[0] * s->c_dc_scale;
 352     quant_matrix = s->intra_matrix;
 353 asm volatile(
 354                 "pcmpeqw %%mm7, %%mm7           \n\t"
 355                 "psrlw $15, %%mm7               \n\t"
 356                 "movd %2, %%mm6                 \n\t"
 357                 "packssdw %%mm6, %%mm6          \n\t"
 358                 "packssdw %%mm6, %%mm6          \n\t"
 359                 "mov %3, %%"REG_a"              \n\t"
 360                 ".balign 16                     \n\t"
 361                 "1:                             \n\t"
 362                 "movq (%0, %%"REG_a"), %%mm0    \n\t"
 363                 "movq 8(%0, %%"REG_a"), %%mm1   \n\t"
 364                 "movq (%1, %%"REG_a"), %%mm4    \n\t"
 365                 "movq 8(%1, %%"REG_a"), %%mm5   \n\t"
 366                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
 367                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
 368                 "pxor %%mm2, %%mm2              \n\t"
 369                 "pxor %%mm3, %%mm3              \n\t"
 370                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
 371                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
 372                 "pxor %%mm2, %%mm0              \n\t"
 373                 "pxor %%mm3, %%mm1              \n\t"
 374                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
 375                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
 376                 "pmullw %%mm4, %%mm0            \n\t" // abs(block[i])*q
 377                 "pmullw %%mm5, %%mm1            \n\t" // abs(block[i])*q
 378                 "pxor %%mm4, %%mm4              \n\t"
 379                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
 380                 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
 381                 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
 382                 "psraw $3, %%mm0                \n\t"
 383                 "psraw $3, %%mm1                \n\t"
 384                 "pxor %%mm2, %%mm0              \n\t"
 385                 "pxor %%mm3, %%mm1              \n\t"
 386                 "psubw %%mm2, %%mm0             \n\t"
 387                 "psubw %%mm3, %%mm1             \n\t"
 388                 "pandn %%mm0, %%mm4             \n\t"
 389                 "pandn %%mm1, %%mm5             \n\t"
 390                 "movq %%mm4, (%0, %%"REG_a")    \n\t"
 391                 "movq %%mm5, 8(%0, %%"REG_a")   \n\t"
 392
 393                 "add $16, %%"REG_a"             \n\t"
 394                 "jng 1b                         \n\t"
 395                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "g" (-2*nCoeffs)
 396                 : "%"REG_a, "memory"
 397         );
 398     block[0]= block0;
 399         //Note, we dont do mismatch control for intra as errors cannot accumulate
 400 }
 401
 402 static void dct_unquantize_mpeg2_inter_mmx(MpegEncContext *s,
 403                                      DCTELEM *block, int n, int qscale)
 404 {
 405     long nCoeffs;
 406     const uint16_t *quant_matrix;
 407
 408     assert(s->block_last_index[n]>=0);
 409
 410     if(s->alternate_scan) nCoeffs= 63; //FIXME
 411     else nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
 412
 413         quant_matrix = s->inter_matrix;
 414 asm volatile(
 415                 "pcmpeqw %%mm7, %%mm7           \n\t"
 416                 "psrlq $48, %%mm7               \n\t"
 417                 "movd %2, %%mm6                 \n\t"
 418                 "packssdw %%mm6, %%mm6          \n\t"
 419                 "packssdw %%mm6, %%mm6          \n\t"
 420                 "mov %3, %%"REG_a"              \n\t"
 421                 ".balign 16                     \n\t"
 422                 "1:                             \n\t"
 423                 "movq (%0, %%"REG_a"), %%mm0    \n\t"
 424                 "movq 8(%0, %%"REG_a"), %%mm1   \n\t"
 425                 "movq (%1, %%"REG_a"), %%mm4    \n\t"
 426                 "movq 8(%1, %%"REG_a"), %%mm5   \n\t"
 427                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
 428                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
 429                 "pxor %%mm2, %%mm2              \n\t"
 430                 "pxor %%mm3, %%mm3              \n\t"
 431                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
 432                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
 433                 "pxor %%mm2, %%mm0              \n\t"
 434                 "pxor %%mm3, %%mm1              \n\t"
 435                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
 436                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
 437                 "paddw %%mm0, %%mm0             \n\t" // abs(block[i])*2
 438                 "paddw %%mm1, %%mm1             \n\t" // abs(block[i])*2
 439                 "pmullw %%mm4, %%mm0            \n\t" // abs(block[i])*2*q
 440                 "pmullw %%mm5, %%mm1            \n\t" // abs(block[i])*2*q
 441                 "paddw %%mm4, %%mm0             \n\t" // (abs(block[i])*2 + 1)*q
 442                 "paddw %%mm5, %%mm1             \n\t" // (abs(block[i])*2 + 1)*q
 443                 "pxor %%mm4, %%mm4              \n\t"
 444                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
 445                 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
 446                 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
 447                 "psrlw $4, %%mm0                \n\t"
 448                 "psrlw $4, %%mm1                \n\t"
 449                 "pxor %%mm2, %%mm0              \n\t"
 450                 "pxor %%mm3, %%mm1              \n\t"
 451                 "psubw %%mm2, %%mm0             \n\t"
 452                 "psubw %%mm3, %%mm1             \n\t"
 453                 "pandn %%mm0, %%mm4             \n\t"
 454                 "pandn %%mm1, %%mm5             \n\t"
 455                 "pxor %%mm4, %%mm7              \n\t"
 456                 "pxor %%mm5, %%mm7              \n\t"
 457                 "movq %%mm4, (%0, %%"REG_a")    \n\t"
 458                 "movq %%mm5, 8(%0, %%"REG_a")   \n\t"
 459
 460                 "add $16, %%"REG_a"             \n\t"
 461                 "jng 1b                         \n\t"
 462                 "movd 124(%0, %3), %%mm0        \n\t"
 463                 "movq %%mm7, %%mm6              \n\t"
 464                 "psrlq $32, %%mm7               \n\t"
 465                 "pxor %%mm6, %%mm7              \n\t"
 466                 "movq %%mm7, %%mm6              \n\t"
 467                 "psrlq $16, %%mm7               \n\t"
 468                 "pxor %%mm6, %%mm7              \n\t"
 469                 "pslld $31, %%mm7               \n\t"
 470                 "psrlq $15, %%mm7               \n\t"
 471                 "pxor %%mm7, %%mm0              \n\t"
 472                 "movd %%mm0, 124(%0, %3)        \n\t"
 473
 474                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "r" (-2*nCoeffs)
 475                 : "%"REG_a, "memory"
 476         );
 477 }
 478
 479 /* draw the edges of width 'w' of an image of size width, height
 480    this mmx version can only handle w==8 || w==16 */
 481 static void draw_edges_mmx(uint8_t *buf, int wrap, int width, int height, int w)
 482 {
 483     uint8_t *ptr, *last_line;
 484     int i;
 485
 486     last_line = buf + (height - 1) * wrap;
 487     /* left and right */
 488     ptr = buf;
 489     if(w==8)
 490     {
 491         asm volatile(
 492                 "1:                             \n\t"
 493                 "movd (%0), %%mm0               \n\t"
 494                 "punpcklbw %%mm0, %%mm0         \n\t"
 495                 "punpcklwd %%mm0, %%mm0         \n\t"
 496                 "punpckldq %%mm0, %%mm0         \n\t"
 497                 "movq %%mm0, -8(%0)             \n\t"
 498                 "movq -8(%0, %2), %%mm1         \n\t"
 499                 "punpckhbw %%mm1, %%mm1         \n\t"
 500                 "punpckhwd %%mm1, %%mm1         \n\t"
 501                 "punpckhdq %%mm1, %%mm1         \n\t"
 502                 "movq %%mm1, (%0, %2)           \n\t"
 503                 "add %1, %0                     \n\t"
 504                 "cmp %3, %0                     \n\t"
 505                 " jb 1b                         \n\t"
 506                 : "+r" (ptr)
 507                 : "r" ((long)wrap), "r" ((long)width), "r" (ptr + wrap*height)
 508         );
 509     }
 510     else
 511     {
 512         asm volatile(
 513                 "1:                             \n\t"
 514                 "movd (%0), %%mm0               \n\t"
 515                 "punpcklbw %%mm0, %%mm0         \n\t"
 516                 "punpcklwd %%mm0, %%mm0         \n\t"
 517                 "punpckldq %%mm0, %%mm0         \n\t"
 518                 "movq %%mm0, -8(%0)             \n\t"
 519                 "movq %%mm0, -16(%0)            \n\t"
 520                 "movq -8(%0, %2), %%mm1         \n\t"
 521                 "punpckhbw %%mm1, %%mm1         \n\t"
 522                 "punpckhwd %%mm1, %%mm1         \n\t"
 523                 "punpckhdq %%mm1, %%mm1         \n\t"
 524                 "movq %%mm1, (%0, %2)           \n\t"
 525                 "movq %%mm1, 8(%0, %2)          \n\t"
 526                 "add %1, %0                     \n\t"
 527                 "cmp %3, %0                     \n\t"
 528                 " jb 1b                         \n\t"
 529                 : "+r" (ptr)
 530                 : "r" ((long)wrap), "r" ((long)width), "r" (ptr + wrap*height)
 531         );
 532     }
 533
 534     for(i=0;i<w;i+=4) {
 535         /* top and bottom (and hopefully also the corners) */
 536         ptr= buf - (i + 1) * wrap - w;
 537         asm volatile(
 538                 "1:                             \n\t"
 539                 "movq (%1, %0), %%mm0           \n\t"
 540                 "movq %%mm0, (%0)               \n\t"
 541                 "movq %%mm0, (%0, %2)           \n\t"
 542                 "movq %%mm0, (%0, %2, 2)        \n\t"
 543                 "movq %%mm0, (%0, %3)           \n\t"
 544                 "add $8, %0                     \n\t"
 545                 "cmp %4, %0                     \n\t"
 546                 " jb 1b                         \n\t"
 547                 : "+r" (ptr)
 548                 : "r" ((long)buf - (long)ptr - w), "r" ((long)-wrap), "r" ((long)-wrap*3), "r" (ptr+width+2*w)
 549         );
 550         ptr= last_line + (i + 1) * wrap - w;
 551         asm volatile(
 552                 "1:                             \n\t"
 553                 "movq (%1, %0), %%mm0           \n\t"
 554                 "movq %%mm0, (%0)               \n\t"
 555                 "movq %%mm0, (%0, %2)           \n\t"
 556                 "movq %%mm0, (%0, %2, 2)        \n\t"
 557                 "movq %%mm0, (%0, %3)           \n\t"
 558                 "add $8, %0                     \n\t"
 559                 "cmp %4, %0                     \n\t"
 560                 " jb 1b                         \n\t"
 561                 : "+r" (ptr)
 562                 : "r" ((long)last_line - (long)ptr - w), "r" ((long)wrap), "r" ((long)wrap*3), "r" (ptr+width+2*w)
 563         );
 564     }
 565 }
 566
 567 static void  denoise_dct_mmx(MpegEncContext *s, DCTELEM *block){
 568     const int intra= s->mb_intra;
 569     int *sum= s->dct_error_sum[intra];
 570     uint16_t *offset= s->dct_offset[intra];
 571
 572     s->dct_count[intra]++;
 573
 574     asm volatile(
 575         "pxor %%mm7, %%mm7                      \n\t"
 576         "1:                                     \n\t"
 577         "pxor %%mm0, %%mm0                      \n\t"
 578         "pxor %%mm1, %%mm1                      \n\t"
 579         "movq (%0), %%mm2                       \n\t"
 580         "movq 8(%0), %%mm3                      \n\t"
 581         "pcmpgtw %%mm2, %%mm0                   \n\t"
 582         "pcmpgtw %%mm3, %%mm1                   \n\t"
 583         "pxor %%mm0, %%mm2                      \n\t"
 584         "pxor %%mm1, %%mm3                      \n\t"
 585         "psubw %%mm0, %%mm2                     \n\t"
 586         "psubw %%mm1, %%mm3                     \n\t"
 587         "movq %%mm2, %%mm4                      \n\t"
 588         "movq %%mm3, %%mm5                      \n\t"
 589         "psubusw (%2), %%mm2                    \n\t"
 590         "psubusw 8(%2), %%mm3                   \n\t"
 591         "pxor %%mm0, %%mm2                      \n\t"
 592         "pxor %%mm1, %%mm3                      \n\t"
 593         "psubw %%mm0, %%mm2                     \n\t"
 594         "psubw %%mm1, %%mm3                     \n\t"
 595         "movq %%mm2, (%0)                       \n\t"
 596         "movq %%mm3, 8(%0)                      \n\t"
 597         "movq %%mm4, %%mm2                      \n\t"
 598         "movq %%mm5, %%mm3                      \n\t"
 599         "punpcklwd %%mm7, %%mm4                 \n\t"
 600         "punpckhwd %%mm7, %%mm2                 \n\t"
 601         "punpcklwd %%mm7, %%mm5                 \n\t"
 602         "punpckhwd %%mm7, %%mm3                 \n\t"
 603         "paddd (%1), %%mm4                      \n\t"
 604         "paddd 8(%1), %%mm2                     \n\t"
 605         "paddd 16(%1), %%mm5                    \n\t"
 606         "paddd 24(%1), %%mm3                    \n\t"
 607         "movq %%mm4, (%1)                       \n\t"
 608         "movq %%mm2, 8(%1)                      \n\t"
 609         "movq %%mm5, 16(%1)                     \n\t"
 610         "movq %%mm3, 24(%1)                     \n\t"
 611         "add $16, %0                            \n\t"
 612         "add $32, %1                            \n\t"
 613         "add $16, %2                            \n\t"
 614         "cmp %3, %0                             \n\t"
 615             " jb 1b                             \n\t"
 616         : "+r" (block), "+r" (sum), "+r" (offset)
 617         : "r"(block+64)
 618     );
 619 }
 620
 621 static void  denoise_dct_sse2(MpegEncContext *s, DCTELEM *block){
 622     const int intra= s->mb_intra;
 623     int *sum= s->dct_error_sum[intra];
 624     uint16_t *offset= s->dct_offset[intra];
 625
 626     s->dct_count[intra]++;
 627
 628     asm volatile(
 629         "pxor %%xmm7, %%xmm7                    \n\t"
 630         "1:                                     \n\t"
 631         "pxor %%xmm0, %%xmm0                    \n\t"
 632         "pxor %%xmm1, %%xmm1                    \n\t"
 633         "movdqa (%0), %%xmm2                    \n\t"
 634         "movdqa 16(%0), %%xmm3                  \n\t"
 635         "pcmpgtw %%xmm2, %%xmm0                 \n\t"
 636         "pcmpgtw %%xmm3, %%xmm1                 \n\t"
 637         "pxor %%xmm0, %%xmm2                    \n\t"
 638         "pxor %%xmm1, %%xmm3                    \n\t"
 639         "psubw %%xmm0, %%xmm2                   \n\t"
 640         "psubw %%xmm1, %%xmm3                   \n\t"
 641         "movdqa %%xmm2, %%xmm4                  \n\t"
 642         "movdqa %%xmm3, %%xmm5                  \n\t"
 643         "psubusw (%2), %%xmm2                   \n\t"
 644         "psubusw 16(%2), %%xmm3                 \n\t"
 645         "pxor %%xmm0, %%xmm2                    \n\t"
 646         "pxor %%xmm1, %%xmm3                    \n\t"
 647         "psubw %%xmm0, %%xmm2                   \n\t"
 648         "psubw %%xmm1, %%xmm3                   \n\t"
 649         "movdqa %%xmm2, (%0)                    \n\t"
 650         "movdqa %%xmm3, 16(%0)                  \n\t"
 651         "movdqa %%xmm4, %%xmm6                  \n\t"
 652         "movdqa %%xmm5, %%xmm0                  \n\t"
 653         "punpcklwd %%xmm7, %%xmm4               \n\t"
 654         "punpckhwd %%xmm7, %%xmm6               \n\t"
 655         "punpcklwd %%xmm7, %%xmm5               \n\t"
 656         "punpckhwd %%xmm7, %%xmm0               \n\t"
 657         "paddd (%1), %%xmm4                     \n\t"
 658         "paddd 16(%1), %%xmm6                   \n\t"
 659         "paddd 32(%1), %%xmm5                   \n\t"
 660         "paddd 48(%1), %%xmm0                   \n\t"
 661         "movdqa %%xmm4, (%1)                    \n\t"
 662         "movdqa %%xmm6, 16(%1)                  \n\t"
 663         "movdqa %%xmm5, 32(%1)                  \n\t"
 664         "movdqa %%xmm0, 48(%1)                  \n\t"
 665         "add $32, %0                            \n\t"
 666         "add $64, %1                            \n\t"
 667         "add $32, %2                            \n\t"
 668         "cmp %3, %0                             \n\t"
 669             " jb 1b                             \n\t"
 670         : "+r" (block), "+r" (sum), "+r" (offset)
 671         : "r"(block+64)
 672     );
 673 }
 674
 675 #undef HAVE_MMX2
 676 #define RENAME(a) a ## _MMX
 677 #define RENAMEl(a) a ## _mmx
 678 #include "mpegvideo_mmx_template.c"
 679
 680 #define HAVE_MMX2
 681 #undef RENAME
 682 #undef RENAMEl
 683 #define RENAME(a) a ## _MMX2
 684 #define RENAMEl(a) a ## _mmx2
 685 #include "mpegvideo_mmx_template.c"
 686
 687 #undef RENAME
 688 #undef RENAMEl
 689 #define RENAME(a) a ## _SSE2
 690 #define RENAMEl(a) a ## _sse2
 691 #include "mpegvideo_mmx_template.c"
 692
 693 void MPV_common_init_mmx(MpegEncContext *s)
 694 {
 695     if (mm_flags & MM_MMX) {
 696         const int dct_algo = s->avctx->dct_algo;
 697
 698         s->dct_unquantize_h263_intra = dct_unquantize_h263_intra_mmx;
 699         s->dct_unquantize_h263_inter = dct_unquantize_h263_inter_mmx;
 700         s->dct_unquantize_mpeg1_intra = dct_unquantize_mpeg1_intra_mmx;
 701         s->dct_unquantize_mpeg1_inter = dct_unquantize_mpeg1_inter_mmx;
 702         if(!(s->flags & CODEC_FLAG_BITEXACT))
 703             s->dct_unquantize_mpeg2_intra = dct_unquantize_mpeg2_intra_mmx;
 704         s->dct_unquantize_mpeg2_inter = dct_unquantize_mpeg2_inter_mmx;
 705
 706         draw_edges = draw_edges_mmx;
 707
 708         if (mm_flags & MM_SSE2) {
 709             s->denoise_dct= denoise_dct_sse2;
 710         } else {
 711                 s->denoise_dct= denoise_dct_mmx;
 712         }
 713
 714         if(dct_algo==FF_DCT_AUTO || dct_algo==FF_DCT_MMX){
 715             if(mm_flags & MM_SSE2){
 716                 s->dct_quantize= dct_quantize_SSE2;
 717             } else if(mm_flags & MM_MMXEXT){
 718                 s->dct_quantize= dct_quantize_MMX2;
 719             } else {
 720                 s->dct_quantize= dct_quantize_MMX;
 721             }
 722         }
 723     }
 724 }