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., 59 Temple Place, Suite 330, Boston, MA  02111-1307  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 "../simple_idct.h"
  27
  28 /* Input permutation for the simple_idct_mmx */
  29 static UINT8 simple_mmx_permutation[64]={
  30         0x00, 0x08, 0x04, 0x09, 0x01, 0x0C, 0x05, 0x0D,
  31         0x10, 0x18, 0x14, 0x19, 0x11, 0x1C, 0x15, 0x1D,
  32         0x20, 0x28, 0x24, 0x29, 0x21, 0x2C, 0x25, 0x2D,
  33         0x12, 0x1A, 0x16, 0x1B, 0x13, 0x1E, 0x17, 0x1F,
  34         0x02, 0x0A, 0x06, 0x0B, 0x03, 0x0E, 0x07, 0x0F,
  35         0x30, 0x38, 0x34, 0x39, 0x31, 0x3C, 0x35, 0x3D,
  36         0x22, 0x2A, 0x26, 0x2B, 0x23, 0x2E, 0x27, 0x2F,
  37         0x32, 0x3A, 0x36, 0x3B, 0x33, 0x3E, 0x37, 0x3F,
  38 };
  39
  40 extern UINT8 zigzag_direct_noperm[64];
  41 extern UINT16 inv_zigzag_direct16[64];
  42 extern UINT32 inverse[256];
  43
  44 static const unsigned long long int mm_wabs __attribute__ ((aligned(8))) = 0xffffffffffffffffULL;
  45 static const unsigned long long int mm_wone __attribute__ ((aligned(8))) = 0x0001000100010001ULL;
  46
  47
  48 static void dct_unquantize_h263_mmx(MpegEncContext *s,
  49                                   DCTELEM *block, int n, int qscale)
  50 {
  51     int level, qmul, qadd, nCoeffs;
  52
  53     qmul = qscale << 1;
  54     qadd = (qscale - 1) | 1;
  55
  56     assert(s->block_last_index[n]>=0);
  57
  58     if (s->mb_intra) {
  59         if (!s->h263_aic) {
  60             if (n < 4)
  61                 level = block[0] * s->y_dc_scale;
  62             else
  63                 level = block[0] * s->c_dc_scale;
  64         }else{
  65             qadd = 0;
  66             level= block[0];
  67         }
  68         nCoeffs=63;
  69     } else {
  70         nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ];
  71     }
  72 //printf("%d %d  ", qmul, qadd);
  73 asm volatile(
  74                 "movd %1, %%mm6                 \n\t" //qmul
  75                 "packssdw %%mm6, %%mm6          \n\t"
  76                 "packssdw %%mm6, %%mm6          \n\t"
  77                 "movd %2, %%mm5                 \n\t" //qadd
  78                 "pxor %%mm7, %%mm7              \n\t"
  79                 "packssdw %%mm5, %%mm5          \n\t"
  80                 "packssdw %%mm5, %%mm5          \n\t"
  81                 "psubw %%mm5, %%mm7             \n\t"
  82                 "pxor %%mm4, %%mm4              \n\t"
  83                 ".balign 16\n\t"
  84                 "1:                             \n\t"
  85                 "movq (%0, %3), %%mm0           \n\t"
  86                 "movq 8(%0, %3), %%mm1          \n\t"
  87
  88                 "pmullw %%mm6, %%mm0            \n\t"
  89                 "pmullw %%mm6, %%mm1            \n\t"
  90
  91                 "movq (%0, %3), %%mm2           \n\t"
  92                 "movq 8(%0, %3), %%mm3          \n\t"
  93
  94                 "pcmpgtw %%mm4, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
  95                 "pcmpgtw %%mm4, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
  96
  97                 "pxor %%mm2, %%mm0              \n\t"
  98                 "pxor %%mm3, %%mm1              \n\t"
  99
 100                 "paddw %%mm7, %%mm0             \n\t"
 101                 "paddw %%mm7, %%mm1             \n\t"
 102
 103                 "pxor %%mm0, %%mm2              \n\t"
 104                 "pxor %%mm1, %%mm3              \n\t"
 105
 106                 "pcmpeqw %%mm7, %%mm0           \n\t" // block[i] == 0 ? -1 : 0
 107                 "pcmpeqw %%mm7, %%mm1           \n\t" // block[i] == 0 ? -1 : 0
 108
 109                 "pandn %%mm2, %%mm0             \n\t"
 110                 "pandn %%mm3, %%mm1             \n\t"
 111
 112                 "movq %%mm0, (%0, %3)           \n\t"
 113                 "movq %%mm1, 8(%0, %3)          \n\t"
 114
 115                 "addl $16, %3                   \n\t"
 116                 "jng 1b                         \n\t"
 117                 ::"r" (block+nCoeffs), "g"(qmul), "g" (qadd), "r" (2*(-nCoeffs))
 118                 : "memory"
 119         );
 120         if(s->mb_intra)
 121             block[0]= level;
 122 }
 123
 124
 125 /*
 126   NK:
 127   Note: looking at PARANOID:
 128   "enable all paranoid tests for rounding, overflows, etc..."
 129
 130 #ifdef PARANOID
 131                 if (level < -2048 || level > 2047)
 132                     fprintf(stderr, "unquant error %d %d\n", i, level);
 133 #endif
 134   We can suppose that result of two multiplications can't be greate of 0xFFFF
 135   i.e. is 16-bit, so we use here only PMULLW instruction and can avoid
 136   a complex multiplication.
 137 =====================================================
 138  Full formula for multiplication of 2 integer numbers
 139  which are represent as high:low words:
 140  input: value1 = high1:low1
 141         value2 = high2:low2
 142  output: value3 = value1*value2
 143  value3=high3:low3 (on overflow: modulus 2^32 wrap-around)
 144  this mean that for 0x123456 * 0x123456 correct result is 0x766cb0ce4
 145  but this algorithm will compute only 0x66cb0ce4
 146  this limited by 16-bit size of operands
 147  ---------------------------------
 148  tlow1 = high1*low2
 149  tlow2 = high2*low1
 150  tlow1 = tlow1 + tlow2
 151  high3:low3 = low1*low2
 152  high3 += tlow1
 153 */
 154 static void dct_unquantize_mpeg1_mmx(MpegEncContext *s,
 155                                      DCTELEM *block, int n, int qscale)
 156 {
 157     int nCoeffs;
 158     const UINT16 *quant_matrix;
 159
 160     assert(s->block_last_index[n]>=0);
 161
 162     nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]+1;
 163
 164     if (s->mb_intra) {
 165         int block0;
 166         if (n < 4)
 167             block0 = block[0] * s->y_dc_scale;
 168         else
 169             block0 = block[0] * s->c_dc_scale;
 170         /* XXX: only mpeg1 */
 171         quant_matrix = s->intra_matrix;
 172 asm volatile(
 173                 "pcmpeqw %%mm7, %%mm7           \n\t"
 174                 "psrlw $15, %%mm7               \n\t"
 175                 "movd %2, %%mm6                 \n\t"
 176                 "packssdw %%mm6, %%mm6          \n\t"
 177                 "packssdw %%mm6, %%mm6          \n\t"
 178                 "movl %3, %%eax                 \n\t"
 179                 ".balign 16\n\t"
 180                 "1:                             \n\t"
 181                 "movq (%0, %%eax), %%mm0        \n\t"
 182                 "movq 8(%0, %%eax), %%mm1       \n\t"
 183                 "movq (%1, %%eax), %%mm4        \n\t"
 184                 "movq 8(%1, %%eax), %%mm5       \n\t"
 185                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
 186                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
 187                 "pxor %%mm2, %%mm2              \n\t"
 188                 "pxor %%mm3, %%mm3              \n\t"
 189                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
 190                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
 191                 "pxor %%mm2, %%mm0              \n\t"
 192                 "pxor %%mm3, %%mm1              \n\t"
 193                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
 194                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
 195                 "pmullw %%mm4, %%mm0            \n\t" // abs(block[i])*q
 196                 "pmullw %%mm5, %%mm1            \n\t" // abs(block[i])*q
 197                 "pxor %%mm4, %%mm4              \n\t"
 198                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
 199                 "pcmpeqw (%0, %%eax), %%mm4     \n\t" // block[i] == 0 ? -1 : 0
 200                 "pcmpeqw 8(%0, %%eax), %%mm5    \n\t" // block[i] == 0 ? -1 : 0
 201                 "psraw $3, %%mm0                \n\t"
 202                 "psraw $3, %%mm1                \n\t"
 203                 "psubw %%mm7, %%mm0             \n\t"
 204                 "psubw %%mm7, %%mm1             \n\t"
 205                 "por %%mm7, %%mm0               \n\t"
 206                 "por %%mm7, %%mm1               \n\t"
 207                 "pxor %%mm2, %%mm0              \n\t"
 208                 "pxor %%mm3, %%mm1              \n\t"
 209                 "psubw %%mm2, %%mm0             \n\t"
 210                 "psubw %%mm3, %%mm1             \n\t"
 211                 "pandn %%mm0, %%mm4             \n\t"
 212                 "pandn %%mm1, %%mm5             \n\t"
 213                 "movq %%mm4, (%0, %%eax)        \n\t"
 214                 "movq %%mm5, 8(%0, %%eax)       \n\t"
 215
 216                 "addl $16, %%eax                \n\t"
 217                 "js 1b                          \n\t"
 218                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "g" (-2*nCoeffs)
 219                 : "%eax", "memory"
 220         );
 221         block[0]= block0;
 222
 223         } else {
 224         quant_matrix = s->inter_matrix;
 225 asm volatile(
 226                 "pcmpeqw %%mm7, %%mm7           \n\t"
 227                 "psrlw $15, %%mm7               \n\t"
 228                 "movd %2, %%mm6                 \n\t"
 229                 "packssdw %%mm6, %%mm6          \n\t"
 230                 "packssdw %%mm6, %%mm6          \n\t"
 231                 "movl %3, %%eax                 \n\t"
 232                 ".balign 16\n\t"
 233                 "1:                             \n\t"
 234                 "movq (%0, %%eax), %%mm0        \n\t"
 235                 "movq 8(%0, %%eax), %%mm1       \n\t"
 236                 "movq (%1, %%eax), %%mm4        \n\t"
 237                 "movq 8(%1, %%eax), %%mm5       \n\t"
 238                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
 239                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
 240                 "pxor %%mm2, %%mm2              \n\t"
 241                 "pxor %%mm3, %%mm3              \n\t"
 242                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
 243                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
 244                 "pxor %%mm2, %%mm0              \n\t"
 245                 "pxor %%mm3, %%mm1              \n\t"
 246                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
 247                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
 248                 "paddw %%mm0, %%mm0             \n\t" // abs(block[i])*2
 249                 "paddw %%mm1, %%mm1             \n\t" // abs(block[i])*2
 250                 "paddw %%mm7, %%mm0             \n\t" // abs(block[i])*2 + 1
 251                 "paddw %%mm7, %%mm1             \n\t" // abs(block[i])*2 + 1
 252                 "pmullw %%mm4, %%mm0            \n\t" // (abs(block[i])*2 + 1)*q
 253                 "pmullw %%mm5, %%mm1            \n\t" // (abs(block[i])*2 + 1)*q
 254                 "pxor %%mm4, %%mm4              \n\t"
 255                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
 256                 "pcmpeqw (%0, %%eax), %%mm4     \n\t" // block[i] == 0 ? -1 : 0
 257                 "pcmpeqw 8(%0, %%eax), %%mm5    \n\t" // block[i] == 0 ? -1 : 0
 258                 "psraw $4, %%mm0                \n\t"
 259                 "psraw $4, %%mm1                \n\t"
 260                 "psubw %%mm7, %%mm0             \n\t"
 261                 "psubw %%mm7, %%mm1             \n\t"
 262                 "por %%mm7, %%mm0               \n\t"
 263                 "por %%mm7, %%mm1               \n\t"
 264                 "pxor %%mm2, %%mm0              \n\t"
 265                 "pxor %%mm3, %%mm1              \n\t"
 266                 "psubw %%mm2, %%mm0             \n\t"
 267                 "psubw %%mm3, %%mm1             \n\t"
 268                 "pandn %%mm0, %%mm4             \n\t"
 269                 "pandn %%mm1, %%mm5             \n\t"
 270                 "movq %%mm4, (%0, %%eax)        \n\t"
 271                 "movq %%mm5, 8(%0, %%eax)       \n\t"
 272
 273                 "addl $16, %%eax                \n\t"
 274                 "js 1b                          \n\t"
 275                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "g" (-2*nCoeffs)
 276                 : "%eax", "memory"
 277         );
 278     }
 279
 280 }
 281
 282 static void dct_unquantize_mpeg2_mmx(MpegEncContext *s,
 283                                      DCTELEM *block, int n, int qscale)
 284 {
 285     int nCoeffs;
 286     const UINT16 *quant_matrix;
 287
 288     assert(s->block_last_index[n]>=0);
 289
 290     if(s->alternate_scan) nCoeffs= 63; //FIXME
 291     else nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
 292
 293     if (s->mb_intra) {
 294         int block0;
 295         if (n < 4)
 296             block0 = block[0] * s->y_dc_scale;
 297         else
 298             block0 = block[0] * s->c_dc_scale;
 299         quant_matrix = s->intra_matrix;
 300 asm volatile(
 301                 "pcmpeqw %%mm7, %%mm7           \n\t"
 302                 "psrlw $15, %%mm7               \n\t"
 303                 "movd %2, %%mm6                 \n\t"
 304                 "packssdw %%mm6, %%mm6          \n\t"
 305                 "packssdw %%mm6, %%mm6          \n\t"
 306                 "movl %3, %%eax                 \n\t"
 307                 ".balign 16\n\t"
 308                 "1:                             \n\t"
 309                 "movq (%0, %%eax), %%mm0        \n\t"
 310                 "movq 8(%0, %%eax), %%mm1       \n\t"
 311                 "movq (%1, %%eax), %%mm4        \n\t"
 312                 "movq 8(%1, %%eax), %%mm5       \n\t"
 313                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
 314                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
 315                 "pxor %%mm2, %%mm2              \n\t"
 316                 "pxor %%mm3, %%mm3              \n\t"
 317                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
 318                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
 319                 "pxor %%mm2, %%mm0              \n\t"
 320                 "pxor %%mm3, %%mm1              \n\t"
 321                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
 322                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
 323                 "pmullw %%mm4, %%mm0            \n\t" // abs(block[i])*q
 324                 "pmullw %%mm5, %%mm1            \n\t" // abs(block[i])*q
 325                 "pxor %%mm4, %%mm4              \n\t"
 326                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
 327                 "pcmpeqw (%0, %%eax), %%mm4     \n\t" // block[i] == 0 ? -1 : 0
 328                 "pcmpeqw 8(%0, %%eax), %%mm5    \n\t" // block[i] == 0 ? -1 : 0
 329                 "psraw $3, %%mm0                \n\t"
 330                 "psraw $3, %%mm1                \n\t"
 331                 "pxor %%mm2, %%mm0              \n\t"
 332                 "pxor %%mm3, %%mm1              \n\t"
 333                 "psubw %%mm2, %%mm0             \n\t"
 334                 "psubw %%mm3, %%mm1             \n\t"
 335                 "pandn %%mm0, %%mm4             \n\t"
 336                 "pandn %%mm1, %%mm5             \n\t"
 337                 "movq %%mm4, (%0, %%eax)        \n\t"
 338                 "movq %%mm5, 8(%0, %%eax)       \n\t"
 339
 340                 "addl $16, %%eax                \n\t"
 341                 "jng 1b                         \n\t"
 342                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "g" (-2*nCoeffs)
 343                 : "%eax", "memory"
 344         );
 345         block[0]= block0;
 346         //Note, we dont do mismatch control for intra as errors cannot accumulate
 347
 348     } else {
 349         quant_matrix = s->inter_matrix;
 350 asm volatile(
 351                 "pcmpeqw %%mm7, %%mm7           \n\t"
 352                 "psrlq $48, %%mm7               \n\t"
 353                 "movd %2, %%mm6                 \n\t"
 354                 "packssdw %%mm6, %%mm6          \n\t"
 355                 "packssdw %%mm6, %%mm6          \n\t"
 356                 "movl %3, %%eax                 \n\t"
 357                 ".balign 16\n\t"
 358                 "1:                             \n\t"
 359                 "movq (%0, %%eax), %%mm0        \n\t"
 360                 "movq 8(%0, %%eax), %%mm1       \n\t"
 361                 "movq (%1, %%eax), %%mm4        \n\t"
 362                 "movq 8(%1, %%eax), %%mm5       \n\t"
 363                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
 364                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
 365                 "pxor %%mm2, %%mm2              \n\t"
 366                 "pxor %%mm3, %%mm3              \n\t"
 367                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
 368                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
 369                 "pxor %%mm2, %%mm0              \n\t"
 370                 "pxor %%mm3, %%mm1              \n\t"
 371                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
 372                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
 373                 "paddw %%mm0, %%mm0             \n\t" // abs(block[i])*2
 374                 "paddw %%mm1, %%mm1             \n\t" // abs(block[i])*2
 375                 "pmullw %%mm4, %%mm0            \n\t" // abs(block[i])*2*q
 376                 "pmullw %%mm5, %%mm1            \n\t" // abs(block[i])*2*q
 377                 "paddw %%mm4, %%mm0             \n\t" // (abs(block[i])*2 + 1)*q
 378                 "paddw %%mm5, %%mm1             \n\t" // (abs(block[i])*2 + 1)*q
 379                 "pxor %%mm4, %%mm4              \n\t"
 380                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
 381                 "pcmpeqw (%0, %%eax), %%mm4     \n\t" // block[i] == 0 ? -1 : 0
 382                 "pcmpeqw 8(%0, %%eax), %%mm5    \n\t" // block[i] == 0 ? -1 : 0
 383                 "psrlw $4, %%mm0                \n\t"
 384                 "psrlw $4, %%mm1                \n\t"
 385                 "pxor %%mm2, %%mm0              \n\t"
 386                 "pxor %%mm3, %%mm1              \n\t"
 387                 "psubw %%mm2, %%mm0             \n\t"
 388                 "psubw %%mm3, %%mm1             \n\t"
 389                 "pandn %%mm0, %%mm4             \n\t"
 390                 "pandn %%mm1, %%mm5             \n\t"
 391                 "pxor %%mm4, %%mm7              \n\t"
 392                 "pxor %%mm5, %%mm7              \n\t"
 393                 "movq %%mm4, (%0, %%eax)        \n\t"
 394                 "movq %%mm5, 8(%0, %%eax)       \n\t"
 395
 396                 "addl $16, %%eax                \n\t"
 397                 "jng 1b                         \n\t"
 398                 "movd 124(%0, %3), %%mm0        \n\t"
 399                 "movq %%mm7, %%mm6              \n\t"
 400                 "psrlq $32, %%mm7               \n\t"
 401                 "pxor %%mm6, %%mm7              \n\t"
 402                 "movq %%mm7, %%mm6              \n\t"
 403                 "psrlq $16, %%mm7               \n\t"
 404                 "pxor %%mm6, %%mm7              \n\t"
 405                 "pslld $31, %%mm7               \n\t"
 406                 "psrlq $15, %%mm7               \n\t"
 407                 "pxor %%mm7, %%mm0              \n\t"
 408                 "movd %%mm0, 124(%0, %3)        \n\t"
 409
 410                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "r" (-2*nCoeffs)
 411                 : "%eax", "memory"
 412         );
 413     }
 414 }
 415
 416 /* draw the edges of width 'w' of an image of size width, height
 417    this mmx version can only handle w==8 || w==16 */
 418 static void draw_edges_mmx(UINT8 *buf, int wrap, int width, int height, int w)
 419 {
 420     UINT8 *ptr, *last_line;
 421     int i;
 422
 423     last_line = buf + (height - 1) * wrap;
 424     /* left and right */
 425     ptr = buf;
 426     if(w==8)
 427     {
 428         asm volatile(
 429                 "1:                             \n\t"
 430                 "movd (%0), %%mm0               \n\t"
 431                 "punpcklbw %%mm0, %%mm0         \n\t"
 432                 "punpcklwd %%mm0, %%mm0         \n\t"
 433                 "punpckldq %%mm0, %%mm0         \n\t"
 434                 "movq %%mm0, -8(%0)             \n\t"
 435                 "movq -8(%0, %2), %%mm1         \n\t"
 436                 "punpckhbw %%mm1, %%mm1         \n\t"
 437                 "punpckhwd %%mm1, %%mm1         \n\t"
 438                 "punpckhdq %%mm1, %%mm1         \n\t"
 439                 "movq %%mm1, (%0, %2)           \n\t"
 440                 "addl %1, %0                    \n\t"
 441                 "cmpl %3, %0                    \n\t"
 442                 " jb 1b                         \n\t"
 443                 : "+r" (ptr)
 444                 : "r" (wrap), "r" (width), "r" (ptr + wrap*height)
 445         );
 446     }
 447     else
 448     {
 449         asm volatile(
 450                 "1:                             \n\t"
 451                 "movd (%0), %%mm0               \n\t"
 452                 "punpcklbw %%mm0, %%mm0         \n\t"
 453                 "punpcklwd %%mm0, %%mm0         \n\t"
 454                 "punpckldq %%mm0, %%mm0         \n\t"
 455                 "movq %%mm0, -8(%0)             \n\t"
 456                 "movq %%mm0, -16(%0)            \n\t"
 457                 "movq -8(%0, %2), %%mm1         \n\t"
 458                 "punpckhbw %%mm1, %%mm1         \n\t"
 459                 "punpckhwd %%mm1, %%mm1         \n\t"
 460                 "punpckhdq %%mm1, %%mm1         \n\t"
 461                 "movq %%mm1, (%0, %2)           \n\t"
 462                 "movq %%mm1, 8(%0, %2)          \n\t"
 463                 "addl %1, %0                    \n\t"
 464                 "cmpl %3, %0                    \n\t"
 465                 " jb 1b                         \n\t"
 466                 : "+r" (ptr)
 467                 : "r" (wrap), "r" (width), "r" (ptr + wrap*height)
 468         );
 469     }
 470
 471     for(i=0;i<w;i+=4) {
 472         /* top and bottom (and hopefully also the corners) */
 473         ptr= buf - (i + 1) * wrap - w;
 474         asm volatile(
 475                 "1:                             \n\t"
 476                 "movq (%1, %0), %%mm0           \n\t"
 477                 "movq %%mm0, (%0)               \n\t"
 478                 "movq %%mm0, (%0, %2)           \n\t"
 479                 "movq %%mm0, (%0, %2, 2)        \n\t"
 480                 "movq %%mm0, (%0, %3)           \n\t"
 481                 "addl $8, %0                    \n\t"
 482                 "cmpl %4, %0                    \n\t"
 483                 " jb 1b                         \n\t"
 484                 : "+r" (ptr)
 485                 : "r" ((int)buf - (int)ptr - w), "r" (-wrap), "r" (-wrap*3), "r" (ptr+width+2*w)
 486         );
 487         ptr= last_line + (i + 1) * wrap - w;
 488         asm volatile(
 489                 "1:                             \n\t"
 490                 "movq (%1, %0), %%mm0           \n\t"
 491                 "movq %%mm0, (%0)               \n\t"
 492                 "movq %%mm0, (%0, %2)           \n\t"
 493                 "movq %%mm0, (%0, %2, 2)        \n\t"
 494                 "movq %%mm0, (%0, %3)           \n\t"
 495                 "addl $8, %0                    \n\t"
 496                 "cmpl %4, %0                    \n\t"
 497                 " jb 1b                         \n\t"
 498                 : "+r" (ptr)
 499                 : "r" ((int)last_line - (int)ptr - w), "r" (wrap), "r" (wrap*3), "r" (ptr+width+2*w)
 500         );
 501     }
 502 }
 503
 504 #undef HAVE_MMX2
 505 #define RENAME(a) a ## _MMX
 506 #include "mpegvideo_mmx_template.c"
 507
 508 #define HAVE_MMX2
 509 #undef RENAME
 510 #define RENAME(a) a ## _MMX2
 511 #include "mpegvideo_mmx_template.c"
 512
 513 /* external functions, from idct_mmx.c */
 514 void ff_mmx_idct(DCTELEM *block);
 515 void ff_mmxext_idct(DCTELEM *block);
 516
 517 /* XXX: those functions should be suppressed ASAP when all IDCTs are
 518    converted */
 519 static void ff_libmpeg2mmx_idct_put(UINT8 *dest, int line_size, DCTELEM *block)
 520 {
 521     ff_mmx_idct (block);
 522     put_pixels_clamped(block, dest, line_size);
 523 }
 524 static void ff_libmpeg2mmx_idct_add(UINT8 *dest, int line_size, DCTELEM *block)
 525 {
 526     ff_mmx_idct (block);
 527     add_pixels_clamped(block, dest, line_size);
 528 }
 529 static void ff_libmpeg2mmx2_idct_put(UINT8 *dest, int line_size, DCTELEM *block)
 530 {
 531     ff_mmxext_idct (block);
 532     put_pixels_clamped(block, dest, line_size);
 533 }
 534 static void ff_libmpeg2mmx2_idct_add(UINT8 *dest, int line_size, DCTELEM *block)
 535 {
 536     ff_mmxext_idct (block);
 537     add_pixels_clamped(block, dest, line_size);
 538 }
 539
 540 void MPV_common_init_mmx(MpegEncContext *s)
 541 {
 542     if (mm_flags & MM_MMX) {
 543         int i;
 544         const int dct_algo = s->avctx->dct_algo;
 545         const int idct_algo= s->avctx->idct_algo;
 546
 547         s->dct_unquantize_h263 = dct_unquantize_h263_mmx;
 548         s->dct_unquantize_mpeg1 = dct_unquantize_mpeg1_mmx;
 549         s->dct_unquantize_mpeg2 = dct_unquantize_mpeg2_mmx;
 550
 551         draw_edges = draw_edges_mmx;
 552
 553         if(dct_algo==FF_DCT_AUTO || dct_algo==FF_DCT_MMX){
 554             s->fdct = ff_fdct_mmx;
 555
 556             if(mm_flags & MM_MMXEXT){
 557                 s->dct_quantize= dct_quantize_MMX2;
 558             } else {
 559                 s->dct_quantize= dct_quantize_MMX;
 560             }
 561         }
 562
 563         if(idct_algo==FF_IDCT_AUTO || idct_algo==FF_IDCT_SIMPLEMMX){
 564             s->idct_put= ff_simple_idct_put_mmx;
 565             s->idct_add= ff_simple_idct_add_mmx;
 566             for(i=0; i<64; i++)
 567                 s->idct_permutation[i]= simple_mmx_permutation[i];
 568         }else if(idct_algo==FF_IDCT_LIBMPEG2MMX){
 569             if(mm_flags & MM_MMXEXT){
 570                 s->idct_put= ff_libmpeg2mmx2_idct_put;
 571                 s->idct_add= ff_libmpeg2mmx2_idct_add;
 572             }else{
 573                 s->idct_put= ff_libmpeg2mmx_idct_put;
 574                 s->idct_add= ff_libmpeg2mmx_idct_add;
 575             }
 576             for(i=0; i<64; i++)
 577                 s->idct_permutation[i]= (i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2);
 578         }
 579     }
 580 }