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

   1 /*
   2  * MMX and SSE2 optimized snow DSP utils
   3  * Copyright (c) 2005-2006 Robert Edele <yartrebo@earthlink.net>
   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/x86_cpu.h"
  23 #include "libavcodec/avcodec.h"
  24 #include "libavcodec/snow.h"
  25
  26 void ff_snow_horizontal_compose97i_sse2(IDWTELEM *b, int width){
  27     const int w2= (width+1)>>1;
  28     DECLARE_ALIGNED_16(IDWTELEM, temp)[width>>1];
  29     const int w_l= (width>>1);
  30     const int w_r= w2 - 1;
  31     int i;
  32
  33     { // Lift 0
  34         IDWTELEM * const ref = b + w2 - 1;
  35         IDWTELEM b_0 = b[0]; //By allowing the first entry in b[0] to be calculated twice
  36         // (the first time erroneously), we allow the SSE2 code to run an extra pass.
  37         // The savings in code and time are well worth having to store this value and
  38         // calculate b[0] correctly afterwards.
  39
  40         i = 0;
  41         __asm__ volatile(
  42             "pcmpeqd   %%xmm7, %%xmm7         \n\t"
  43             "pcmpeqd   %%xmm3, %%xmm3         \n\t"
  44             "psllw         $1, %%xmm3         \n\t"
  45             "paddw     %%xmm7, %%xmm3         \n\t"
  46             "psllw        $13, %%xmm3         \n\t"
  47         ::);
  48         for(; i<w_l-15; i+=16){
  49             __asm__ volatile(
  50                 "movdqu   (%1), %%xmm1        \n\t"
  51                 "movdqu 16(%1), %%xmm5        \n\t"
  52                 "movdqu  2(%1), %%xmm2        \n\t"
  53                 "movdqu 18(%1), %%xmm6        \n\t"
  54                 "paddw  %%xmm1, %%xmm2        \n\t"
  55                 "paddw  %%xmm5, %%xmm6        \n\t"
  56                 "paddw  %%xmm7, %%xmm2        \n\t"
  57                 "paddw  %%xmm7, %%xmm6        \n\t"
  58                 "pmulhw %%xmm3, %%xmm2        \n\t"
  59                 "pmulhw %%xmm3, %%xmm6        \n\t"
  60                 "paddw    (%0), %%xmm2        \n\t"
  61                 "paddw  16(%0), %%xmm6        \n\t"
  62                 "movdqa %%xmm2, (%0)          \n\t"
  63                 "movdqa %%xmm6, 16(%0)        \n\t"
  64                 :: "r"(&b[i]), "r"(&ref[i])
  65                 : "memory"
  66             );
  67         }
  68         snow_horizontal_compose_lift_lead_out(i, b, b, ref, width, w_l, 0, W_DM, W_DO, W_DS);
  69         b[0] = b_0 - ((W_DM * 2 * ref[1]+W_DO)>>W_DS);
  70     }
  71
  72     { // Lift 1
  73         IDWTELEM * const dst = b+w2;
  74
  75         i = 0;
  76         for(; (((x86_reg)&dst[i]) & 0x1F) && i<w_r; i++){
  77             dst[i] = dst[i] - (b[i] + b[i + 1]);
  78         }
  79         for(; i<w_r-15; i+=16){
  80             __asm__ volatile(
  81                 "movdqu   (%1), %%xmm1        \n\t"
  82                 "movdqu 16(%1), %%xmm5        \n\t"
  83                 "movdqu  2(%1), %%xmm2        \n\t"
  84                 "movdqu 18(%1), %%xmm6        \n\t"
  85                 "paddw  %%xmm1, %%xmm2        \n\t"
  86                 "paddw  %%xmm5, %%xmm6        \n\t"
  87                 "movdqa   (%0), %%xmm0        \n\t"
  88                 "movdqa 16(%0), %%xmm4        \n\t"
  89                 "psubw  %%xmm2, %%xmm0        \n\t"
  90                 "psubw  %%xmm6, %%xmm4        \n\t"
  91                 "movdqa %%xmm0, (%0)          \n\t"
  92                 "movdqa %%xmm4, 16(%0)        \n\t"
  93                 :: "r"(&dst[i]), "r"(&b[i])
  94                 : "memory"
  95             );
  96         }
  97         snow_horizontal_compose_lift_lead_out(i, dst, dst, b, width, w_r, 1, W_CM, W_CO, W_CS);
  98     }
  99
 100     { // Lift 2
 101         IDWTELEM * const ref = b+w2 - 1;
 102         IDWTELEM b_0 = b[0];
 103
 104         i = 0;
 105         __asm__ volatile(
 106             "psllw         $15, %%xmm7        \n\t"
 107             "pcmpeqw    %%xmm6, %%xmm6        \n\t"
 108             "psrlw         $13, %%xmm6        \n\t"
 109             "paddw      %%xmm7, %%xmm6        \n\t"
 110         ::);
 111         for(; i<w_l-15; i+=16){
 112             __asm__ volatile(
 113                 "movdqu   (%1), %%xmm0        \n\t"
 114                 "movdqu 16(%1), %%xmm4        \n\t"
 115                 "movdqu  2(%1), %%xmm1        \n\t"
 116                 "movdqu 18(%1), %%xmm5        \n\t" //FIXME try aligned reads and shifts
 117                 "paddw  %%xmm6, %%xmm0        \n\t"
 118                 "paddw  %%xmm6, %%xmm4        \n\t"
 119                 "paddw  %%xmm7, %%xmm1        \n\t"
 120                 "paddw  %%xmm7, %%xmm5        \n\t"
 121                 "pavgw  %%xmm1, %%xmm0        \n\t"
 122                 "pavgw  %%xmm5, %%xmm4        \n\t"
 123                 "psubw  %%xmm7, %%xmm0        \n\t"
 124                 "psubw  %%xmm7, %%xmm4        \n\t"
 125                 "psraw      $1, %%xmm0        \n\t"
 126                 "psraw      $1, %%xmm4        \n\t"
 127                 "movdqa   (%0), %%xmm1        \n\t"
 128                 "movdqa 16(%0), %%xmm5        \n\t"
 129                 "paddw  %%xmm1, %%xmm0        \n\t"
 130                 "paddw  %%xmm5, %%xmm4        \n\t"
 131                 "psraw      $2, %%xmm0        \n\t"
 132                 "psraw      $2, %%xmm4        \n\t"
 133                 "paddw  %%xmm1, %%xmm0        \n\t"
 134                 "paddw  %%xmm5, %%xmm4        \n\t"
 135                 "movdqa %%xmm0, (%0)          \n\t"
 136                 "movdqa %%xmm4, 16(%0)        \n\t"
 137                 :: "r"(&b[i]), "r"(&ref[i])
 138                 : "memory"
 139             );
 140         }
 141         snow_horizontal_compose_liftS_lead_out(i, b, b, ref, width, w_l);
 142         b[0] = b_0 + ((2 * ref[1] + W_BO-1 + 4 * b_0) >> W_BS);
 143     }
 144
 145     { // Lift 3
 146         IDWTELEM * const src = b+w2;
 147
 148         i = 0;
 149         for(; (((x86_reg)&temp[i]) & 0x1F) && i<w_r; i++){
 150             temp[i] = src[i] - ((-W_AM*(b[i] + b[i+1]))>>W_AS);
 151         }
 152         for(; i<w_r-7; i+=8){
 153             __asm__ volatile(
 154                 "movdqu  2(%1), %%xmm2        \n\t"
 155                 "movdqu 18(%1), %%xmm6        \n\t"
 156                 "paddw    (%1), %%xmm2        \n\t"
 157                 "paddw  16(%1), %%xmm6        \n\t"
 158                 "movdqu   (%0), %%xmm0        \n\t"
 159                 "movdqu 16(%0), %%xmm4        \n\t"
 160                 "paddw  %%xmm2, %%xmm0        \n\t"
 161                 "paddw  %%xmm6, %%xmm4        \n\t"
 162                 "psraw      $1, %%xmm2        \n\t"
 163                 "psraw      $1, %%xmm6        \n\t"
 164                 "paddw  %%xmm0, %%xmm2        \n\t"
 165                 "paddw  %%xmm4, %%xmm6        \n\t"
 166                 "movdqa %%xmm2, (%2)          \n\t"
 167                 "movdqa %%xmm6, 16(%2)        \n\t"
 168                 :: "r"(&src[i]), "r"(&b[i]), "r"(&temp[i])
 169                  : "memory"
 170                );
 171         }
 172         snow_horizontal_compose_lift_lead_out(i, temp, src, b, width, w_r, 1, -W_AM, W_AO+1, W_AS);
 173     }
 174
 175     {
 176         snow_interleave_line_header(&i, width, b, temp);
 177
 178         for (; (i & 0x3E) != 0x3E; i-=2){
 179             b[i+1] = temp[i>>1];
 180             b[i] = b[i>>1];
 181         }
 182         for (i-=62; i>=0; i-=64){
 183             __asm__ volatile(
 184                 "movdqa      (%1), %%xmm0       \n\t"
 185                 "movdqa    16(%1), %%xmm2       \n\t"
 186                 "movdqa    32(%1), %%xmm4       \n\t"
 187                 "movdqa    48(%1), %%xmm6       \n\t"
 188                 "movdqa      (%1), %%xmm1       \n\t"
 189                 "movdqa    16(%1), %%xmm3       \n\t"
 190                 "movdqa    32(%1), %%xmm5       \n\t"
 191                 "movdqa    48(%1), %%xmm7       \n\t"
 192                 "punpcklwd   (%2), %%xmm0       \n\t"
 193                 "punpcklwd 16(%2), %%xmm2       \n\t"
 194                 "punpcklwd 32(%2), %%xmm4       \n\t"
 195                 "punpcklwd 48(%2), %%xmm6       \n\t"
 196                 "movdqa    %%xmm0, (%0)         \n\t"
 197                 "movdqa    %%xmm2, 32(%0)       \n\t"
 198                 "movdqa    %%xmm4, 64(%0)       \n\t"
 199                 "movdqa    %%xmm6, 96(%0)       \n\t"
 200                 "punpckhwd   (%2), %%xmm1       \n\t"
 201                 "punpckhwd 16(%2), %%xmm3       \n\t"
 202                 "punpckhwd 32(%2), %%xmm5       \n\t"
 203                 "punpckhwd 48(%2), %%xmm7       \n\t"
 204                 "movdqa    %%xmm1, 16(%0)       \n\t"
 205                 "movdqa    %%xmm3, 48(%0)       \n\t"
 206                 "movdqa    %%xmm5, 80(%0)       \n\t"
 207                 "movdqa    %%xmm7, 112(%0)      \n\t"
 208                 :: "r"(&(b)[i]), "r"(&(b)[i>>1]), "r"(&(temp)[i>>1])
 209                  : "memory"
 210                );
 211         }
 212     }
 213 }
 214
 215 void ff_snow_horizontal_compose97i_mmx(IDWTELEM *b, int width){
 216     const int w2= (width+1)>>1;
 217     IDWTELEM temp[width >> 1];
 218     const int w_l= (width>>1);
 219     const int w_r= w2 - 1;
 220     int i;
 221
 222     { // Lift 0
 223         IDWTELEM * const ref = b + w2 - 1;
 224
 225         i = 1;
 226         b[0] = b[0] - ((W_DM * 2 * ref[1]+W_DO)>>W_DS);
 227         __asm__ volatile(
 228             "pcmpeqw    %%mm7, %%mm7         \n\t"
 229             "pcmpeqw    %%mm3, %%mm3         \n\t"
 230             "psllw         $1, %%mm3         \n\t"
 231             "paddw      %%mm7, %%mm3         \n\t"
 232             "psllw        $13, %%mm3         \n\t"
 233            ::);
 234         for(; i<w_l-7; i+=8){
 235             __asm__ volatile(
 236                 "movq     (%1), %%mm2        \n\t"
 237                 "movq    8(%1), %%mm6        \n\t"
 238                 "paddw   2(%1), %%mm2        \n\t"
 239                 "paddw  10(%1), %%mm6        \n\t"
 240                 "paddw   %%mm7, %%mm2        \n\t"
 241                 "paddw   %%mm7, %%mm6        \n\t"
 242                 "pmulhw  %%mm3, %%mm2        \n\t"
 243                 "pmulhw  %%mm3, %%mm6        \n\t"
 244                 "paddw    (%0), %%mm2        \n\t"
 245                 "paddw   8(%0), %%mm6        \n\t"
 246                 "movq    %%mm2, (%0)         \n\t"
 247                 "movq    %%mm6, 8(%0)        \n\t"
 248                 :: "r"(&b[i]), "r"(&ref[i])
 249                  : "memory"
 250                );
 251         }
 252         snow_horizontal_compose_lift_lead_out(i, b, b, ref, width, w_l, 0, W_DM, W_DO, W_DS);
 253     }
 254
 255     { // Lift 1
 256         IDWTELEM * const dst = b+w2;
 257
 258         i = 0;
 259         for(; i<w_r-7; i+=8){
 260             __asm__ volatile(
 261                 "movq     (%1), %%mm2        \n\t"
 262                 "movq    8(%1), %%mm6        \n\t"
 263                 "paddw   2(%1), %%mm2        \n\t"
 264                 "paddw  10(%1), %%mm6        \n\t"
 265                 "movq     (%0), %%mm0        \n\t"
 266                 "movq    8(%0), %%mm4        \n\t"
 267                 "psubw   %%mm2, %%mm0        \n\t"
 268                 "psubw   %%mm6, %%mm4        \n\t"
 269                 "movq    %%mm0, (%0)         \n\t"
 270                 "movq    %%mm4, 8(%0)        \n\t"
 271                 :: "r"(&dst[i]), "r"(&b[i])
 272                  : "memory"
 273                );
 274         }
 275         snow_horizontal_compose_lift_lead_out(i, dst, dst, b, width, w_r, 1, W_CM, W_CO, W_CS);
 276     }
 277
 278     { // Lift 2
 279         IDWTELEM * const ref = b+w2 - 1;
 280
 281         i = 1;
 282         b[0] = b[0] + (((2 * ref[1] + W_BO) + 4 * b[0]) >> W_BS);
 283         __asm__ volatile(
 284             "psllw         $15, %%mm7        \n\t"
 285             "pcmpeqw     %%mm6, %%mm6        \n\t"
 286             "psrlw         $13, %%mm6        \n\t"
 287             "paddw       %%mm7, %%mm6        \n\t"
 288            ::);
 289         for(; i<w_l-7; i+=8){
 290             __asm__ volatile(
 291                 "movq     (%1), %%mm0        \n\t"
 292                 "movq    8(%1), %%mm4        \n\t"
 293                 "movq    2(%1), %%mm1        \n\t"
 294                 "movq   10(%1), %%mm5        \n\t"
 295                 "paddw   %%mm6, %%mm0        \n\t"
 296                 "paddw   %%mm6, %%mm4        \n\t"
 297                 "paddw   %%mm7, %%mm1        \n\t"
 298                 "paddw   %%mm7, %%mm5        \n\t"
 299                 "pavgw   %%mm1, %%mm0        \n\t"
 300                 "pavgw   %%mm5, %%mm4        \n\t"
 301                 "psubw   %%mm7, %%mm0        \n\t"
 302                 "psubw   %%mm7, %%mm4        \n\t"
 303                 "psraw      $1, %%mm0        \n\t"
 304                 "psraw      $1, %%mm4        \n\t"
 305                 "movq     (%0), %%mm1        \n\t"
 306                 "movq    8(%0), %%mm5        \n\t"
 307                 "paddw   %%mm1, %%mm0        \n\t"
 308                 "paddw   %%mm5, %%mm4        \n\t"
 309                 "psraw      $2, %%mm0        \n\t"
 310                 "psraw      $2, %%mm4        \n\t"
 311                 "paddw   %%mm1, %%mm0        \n\t"
 312                 "paddw   %%mm5, %%mm4        \n\t"
 313                 "movq    %%mm0, (%0)         \n\t"
 314                 "movq    %%mm4, 8(%0)        \n\t"
 315                 :: "r"(&b[i]), "r"(&ref[i])
 316                  : "memory"
 317                );
 318         }
 319         snow_horizontal_compose_liftS_lead_out(i, b, b, ref, width, w_l);
 320     }
 321
 322     { // Lift 3
 323         IDWTELEM * const src = b+w2;
 324         i = 0;
 325
 326         for(; i<w_r-7; i+=8){
 327             __asm__ volatile(
 328                 "movq    2(%1), %%mm2        \n\t"
 329                 "movq   10(%1), %%mm6        \n\t"
 330                 "paddw    (%1), %%mm2        \n\t"
 331                 "paddw   8(%1), %%mm6        \n\t"
 332                 "movq     (%0), %%mm0        \n\t"
 333                 "movq    8(%0), %%mm4        \n\t"
 334                 "paddw   %%mm2, %%mm0        \n\t"
 335                 "paddw   %%mm6, %%mm4        \n\t"
 336                 "psraw      $1, %%mm2        \n\t"
 337                 "psraw      $1, %%mm6        \n\t"
 338                 "paddw   %%mm0, %%mm2        \n\t"
 339                 "paddw   %%mm4, %%mm6        \n\t"
 340                 "movq    %%mm2, (%2)         \n\t"
 341                 "movq    %%mm6, 8(%2)        \n\t"
 342                 :: "r"(&src[i]), "r"(&b[i]), "r"(&temp[i])
 343                  : "memory"
 344                );
 345         }
 346         snow_horizontal_compose_lift_lead_out(i, temp, src, b, width, w_r, 1, -W_AM, W_AO+1, W_AS);
 347     }
 348
 349     {
 350         snow_interleave_line_header(&i, width, b, temp);
 351
 352         for (; (i & 0x1E) != 0x1E; i-=2){
 353             b[i+1] = temp[i>>1];
 354             b[i] = b[i>>1];
 355         }
 356         for (i-=30; i>=0; i-=32){
 357             __asm__ volatile(
 358                 "movq        (%1), %%mm0       \n\t"
 359                 "movq       8(%1), %%mm2       \n\t"
 360                 "movq      16(%1), %%mm4       \n\t"
 361                 "movq      24(%1), %%mm6       \n\t"
 362                 "movq        (%1), %%mm1       \n\t"
 363                 "movq       8(%1), %%mm3       \n\t"
 364                 "movq      16(%1), %%mm5       \n\t"
 365                 "movq      24(%1), %%mm7       \n\t"
 366                 "punpcklwd   (%2), %%mm0       \n\t"
 367                 "punpcklwd  8(%2), %%mm2       \n\t"
 368                 "punpcklwd 16(%2), %%mm4       \n\t"
 369                 "punpcklwd 24(%2), %%mm6       \n\t"
 370                 "movq       %%mm0, (%0)        \n\t"
 371                 "movq       %%mm2, 16(%0)      \n\t"
 372                 "movq       %%mm4, 32(%0)      \n\t"
 373                 "movq       %%mm6, 48(%0)      \n\t"
 374                 "punpckhwd   (%2), %%mm1       \n\t"
 375                 "punpckhwd  8(%2), %%mm3       \n\t"
 376                 "punpckhwd 16(%2), %%mm5       \n\t"
 377                 "punpckhwd 24(%2), %%mm7       \n\t"
 378                 "movq       %%mm1, 8(%0)       \n\t"
 379                 "movq       %%mm3, 24(%0)      \n\t"
 380                 "movq       %%mm5, 40(%0)      \n\t"
 381                 "movq       %%mm7, 56(%0)      \n\t"
 382                 :: "r"(&b[i]), "r"(&b[i>>1]), "r"(&temp[i>>1])
 383                  : "memory"
 384                );
 385         }
 386     }
 387 }
 388
 389 #if HAVE_7REGS
 390 #define snow_vertical_compose_sse2_load_add(op,r,t0,t1,t2,t3)\
 391         ""op" ("r",%%"REG_d"), %%"t0"      \n\t"\
 392         ""op" 16("r",%%"REG_d"), %%"t1"    \n\t"\
 393         ""op" 32("r",%%"REG_d"), %%"t2"    \n\t"\
 394         ""op" 48("r",%%"REG_d"), %%"t3"    \n\t"
 395
 396 #define snow_vertical_compose_sse2_load(r,t0,t1,t2,t3)\
 397         snow_vertical_compose_sse2_load_add("movdqa",r,t0,t1,t2,t3)
 398
 399 #define snow_vertical_compose_sse2_add(r,t0,t1,t2,t3)\
 400         snow_vertical_compose_sse2_load_add("paddw",r,t0,t1,t2,t3)
 401
 402 #define snow_vertical_compose_r2r_sub(s0,s1,s2,s3,t0,t1,t2,t3)\
 403         "psubw %%"s0", %%"t0" \n\t"\
 404         "psubw %%"s1", %%"t1" \n\t"\
 405         "psubw %%"s2", %%"t2" \n\t"\
 406         "psubw %%"s3", %%"t3" \n\t"
 407
 408 #define snow_vertical_compose_sse2_store(w,s0,s1,s2,s3)\
 409         "movdqa %%"s0", ("w",%%"REG_d")      \n\t"\
 410         "movdqa %%"s1", 16("w",%%"REG_d")    \n\t"\
 411         "movdqa %%"s2", 32("w",%%"REG_d")    \n\t"\
 412         "movdqa %%"s3", 48("w",%%"REG_d")    \n\t"
 413
 414 #define snow_vertical_compose_sra(n,t0,t1,t2,t3)\
 415         "psraw $"n", %%"t0" \n\t"\
 416         "psraw $"n", %%"t1" \n\t"\
 417         "psraw $"n", %%"t2" \n\t"\
 418         "psraw $"n", %%"t3" \n\t"
 419
 420 #define snow_vertical_compose_r2r_add(s0,s1,s2,s3,t0,t1,t2,t3)\
 421         "paddw %%"s0", %%"t0" \n\t"\
 422         "paddw %%"s1", %%"t1" \n\t"\
 423         "paddw %%"s2", %%"t2" \n\t"\
 424         "paddw %%"s3", %%"t3" \n\t"
 425
 426 #define snow_vertical_compose_r2r_pmulhw(s0,s1,s2,s3,t0,t1,t2,t3)\
 427         "pmulhw %%"s0", %%"t0" \n\t"\
 428         "pmulhw %%"s1", %%"t1" \n\t"\
 429         "pmulhw %%"s2", %%"t2" \n\t"\
 430         "pmulhw %%"s3", %%"t3" \n\t"
 431
 432 #define snow_vertical_compose_sse2_move(s0,s1,s2,s3,t0,t1,t2,t3)\
 433         "movdqa %%"s0", %%"t0" \n\t"\
 434         "movdqa %%"s1", %%"t1" \n\t"\
 435         "movdqa %%"s2", %%"t2" \n\t"\
 436         "movdqa %%"s3", %%"t3" \n\t"
 437
 438 void ff_snow_vertical_compose97i_sse2(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, IDWTELEM *b5, int width){
 439     x86_reg i = width;
 440
 441     while(i & 0x1F)
 442     {
 443         i--;
 444         b4[i] -= (W_DM*(b3[i] + b5[i])+W_DO)>>W_DS;
 445         b3[i] -= (W_CM*(b2[i] + b4[i])+W_CO)>>W_CS;
 446         b2[i] += (W_BM*(b1[i] + b3[i])+4*b2[i]+W_BO)>>W_BS;
 447         b1[i] += (W_AM*(b0[i] + b2[i])+W_AO)>>W_AS;
 448     }
 449     i+=i;
 450
 451          __asm__ volatile (
 452         "jmp 2f                                      \n\t"
 453         "1:                                          \n\t"
 454         snow_vertical_compose_sse2_load("%4","xmm0","xmm2","xmm4","xmm6")
 455         snow_vertical_compose_sse2_add("%6","xmm0","xmm2","xmm4","xmm6")
 456
 457
 458         "pcmpeqw    %%xmm0, %%xmm0                   \n\t"
 459         "pcmpeqw    %%xmm2, %%xmm2                   \n\t"
 460         "paddw      %%xmm2, %%xmm2                   \n\t"
 461         "paddw      %%xmm0, %%xmm2                   \n\t"
 462         "psllw         $13, %%xmm2                   \n\t"
 463         snow_vertical_compose_r2r_add("xmm0","xmm0","xmm0","xmm0","xmm1","xmm3","xmm5","xmm7")
 464         snow_vertical_compose_r2r_pmulhw("xmm2","xmm2","xmm2","xmm2","xmm1","xmm3","xmm5","xmm7")
 465         snow_vertical_compose_sse2_add("%5","xmm1","xmm3","xmm5","xmm7")
 466         snow_vertical_compose_sse2_store("%5","xmm1","xmm3","xmm5","xmm7")
 467         snow_vertical_compose_sse2_load("%4","xmm0","xmm2","xmm4","xmm6")
 468         snow_vertical_compose_sse2_add("%3","xmm1","xmm3","xmm5","xmm7")
 469         snow_vertical_compose_r2r_sub("xmm1","xmm3","xmm5","xmm7","xmm0","xmm2","xmm4","xmm6")
 470         snow_vertical_compose_sse2_store("%4","xmm0","xmm2","xmm4","xmm6")
 471
 472         "pcmpeqw %%xmm7, %%xmm7                      \n\t"
 473         "pcmpeqw %%xmm5, %%xmm5                      \n\t"
 474         "psllw $15, %%xmm7                           \n\t"
 475         "psrlw $13, %%xmm5                           \n\t"
 476         "paddw %%xmm7, %%xmm5                        \n\t"
 477         snow_vertical_compose_r2r_add("xmm5","xmm5","xmm5","xmm5","xmm0","xmm2","xmm4","xmm6")
 478         "movq   (%2,%%"REG_d"), %%xmm1        \n\t"
 479         "movq  8(%2,%%"REG_d"), %%xmm3        \n\t"
 480         "paddw %%xmm7, %%xmm1                        \n\t"
 481         "paddw %%xmm7, %%xmm3                        \n\t"
 482         "pavgw %%xmm1, %%xmm0                        \n\t"
 483         "pavgw %%xmm3, %%xmm2                        \n\t"
 484         "movq 16(%2,%%"REG_d"), %%xmm1        \n\t"
 485         "movq 24(%2,%%"REG_d"), %%xmm3        \n\t"
 486         "paddw %%xmm7, %%xmm1                        \n\t"
 487         "paddw %%xmm7, %%xmm3                        \n\t"
 488         "pavgw %%xmm1, %%xmm4                        \n\t"
 489         "pavgw %%xmm3, %%xmm6                        \n\t"
 490         snow_vertical_compose_r2r_sub("xmm7","xmm7","xmm7","xmm7","xmm0","xmm2","xmm4","xmm6")
 491         snow_vertical_compose_sra("1","xmm0","xmm2","xmm4","xmm6")
 492         snow_vertical_compose_sse2_add("%3","xmm0","xmm2","xmm4","xmm6")
 493
 494         snow_vertical_compose_sra("2","xmm0","xmm2","xmm4","xmm6")
 495         snow_vertical_compose_sse2_add("%3","xmm0","xmm2","xmm4","xmm6")
 496         snow_vertical_compose_sse2_store("%3","xmm0","xmm2","xmm4","xmm6")
 497         snow_vertical_compose_sse2_add("%1","xmm0","xmm2","xmm4","xmm6")
 498         snow_vertical_compose_sse2_move("xmm0","xmm2","xmm4","xmm6","xmm1","xmm3","xmm5","xmm7")
 499         snow_vertical_compose_sra("1","xmm0","xmm2","xmm4","xmm6")
 500         snow_vertical_compose_r2r_add("xmm1","xmm3","xmm5","xmm7","xmm0","xmm2","xmm4","xmm6")
 501         snow_vertical_compose_sse2_add("%2","xmm0","xmm2","xmm4","xmm6")
 502         snow_vertical_compose_sse2_store("%2","xmm0","xmm2","xmm4","xmm6")
 503
 504         "2:                                          \n\t"
 505         "sub $64, %%"REG_d"                          \n\t"
 506         "jge 1b                                      \n\t"
 507         :"+d"(i)
 508         :"r"(b0),"r"(b1),"r"(b2),"r"(b3),"r"(b4),"r"(b5));
 509 }
 510
 511 #define snow_vertical_compose_mmx_load_add(op,r,t0,t1,t2,t3)\
 512         ""op" ("r",%%"REG_d"), %%"t0"   \n\t"\
 513         ""op" 8("r",%%"REG_d"), %%"t1"  \n\t"\
 514         ""op" 16("r",%%"REG_d"), %%"t2" \n\t"\
 515         ""op" 24("r",%%"REG_d"), %%"t3" \n\t"
 516
 517 #define snow_vertical_compose_mmx_load(r,t0,t1,t2,t3)\
 518         snow_vertical_compose_mmx_load_add("movq",r,t0,t1,t2,t3)
 519
 520 #define snow_vertical_compose_mmx_add(r,t0,t1,t2,t3)\
 521         snow_vertical_compose_mmx_load_add("paddw",r,t0,t1,t2,t3)
 522
 523 #define snow_vertical_compose_mmx_store(w,s0,s1,s2,s3)\
 524         "movq %%"s0", ("w",%%"REG_d")   \n\t"\
 525         "movq %%"s1", 8("w",%%"REG_d")  \n\t"\
 526         "movq %%"s2", 16("w",%%"REG_d") \n\t"\
 527         "movq %%"s3", 24("w",%%"REG_d") \n\t"
 528
 529 #define snow_vertical_compose_mmx_move(s0,s1,s2,s3,t0,t1,t2,t3)\
 530         "movq %%"s0", %%"t0" \n\t"\
 531         "movq %%"s1", %%"t1" \n\t"\
 532         "movq %%"s2", %%"t2" \n\t"\
 533         "movq %%"s3", %%"t3" \n\t"
 534
 535
 536 void ff_snow_vertical_compose97i_mmx(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, IDWTELEM *b5, int width){
 537     x86_reg i = width;
 538     while(i & 15)
 539     {
 540         i--;
 541         b4[i] -= (W_DM*(b3[i] + b5[i])+W_DO)>>W_DS;
 542         b3[i] -= (W_CM*(b2[i] + b4[i])+W_CO)>>W_CS;
 543         b2[i] += (W_BM*(b1[i] + b3[i])+4*b2[i]+W_BO)>>W_BS;
 544         b1[i] += (W_AM*(b0[i] + b2[i])+W_AO)>>W_AS;
 545     }
 546     i+=i;
 547     __asm__ volatile(
 548         "jmp 2f                                      \n\t"
 549         "1:                                          \n\t"
 550
 551         snow_vertical_compose_mmx_load("%4","mm1","mm3","mm5","mm7")
 552         snow_vertical_compose_mmx_add("%6","mm1","mm3","mm5","mm7")
 553         "pcmpeqw    %%mm0, %%mm0                     \n\t"
 554         "pcmpeqw    %%mm2, %%mm2                     \n\t"
 555         "paddw      %%mm2, %%mm2                     \n\t"
 556         "paddw      %%mm0, %%mm2                     \n\t"
 557         "psllw        $13, %%mm2                     \n\t"
 558         snow_vertical_compose_r2r_add("mm0","mm0","mm0","mm0","mm1","mm3","mm5","mm7")
 559         snow_vertical_compose_r2r_pmulhw("mm2","mm2","mm2","mm2","mm1","mm3","mm5","mm7")
 560         snow_vertical_compose_mmx_add("%5","mm1","mm3","mm5","mm7")
 561         snow_vertical_compose_mmx_store("%5","mm1","mm3","mm5","mm7")
 562         snow_vertical_compose_mmx_load("%4","mm0","mm2","mm4","mm6")
 563         snow_vertical_compose_mmx_add("%3","mm1","mm3","mm5","mm7")
 564         snow_vertical_compose_r2r_sub("mm1","mm3","mm5","mm7","mm0","mm2","mm4","mm6")
 565         snow_vertical_compose_mmx_store("%4","mm0","mm2","mm4","mm6")
 566         "pcmpeqw %%mm7, %%mm7                        \n\t"
 567         "pcmpeqw %%mm5, %%mm5                        \n\t"
 568         "psllw $15, %%mm7                            \n\t"
 569         "psrlw $13, %%mm5                            \n\t"
 570         "paddw %%mm7, %%mm5                          \n\t"
 571         snow_vertical_compose_r2r_add("mm5","mm5","mm5","mm5","mm0","mm2","mm4","mm6")
 572         "movq   (%2,%%"REG_d"), %%mm1         \n\t"
 573         "movq  8(%2,%%"REG_d"), %%mm3         \n\t"
 574         "paddw %%mm7, %%mm1                          \n\t"
 575         "paddw %%mm7, %%mm3                          \n\t"
 576         "pavgw %%mm1, %%mm0                          \n\t"
 577         "pavgw %%mm3, %%mm2                          \n\t"
 578         "movq 16(%2,%%"REG_d"), %%mm1         \n\t"
 579         "movq 24(%2,%%"REG_d"), %%mm3         \n\t"
 580         "paddw %%mm7, %%mm1                          \n\t"
 581         "paddw %%mm7, %%mm3                          \n\t"
 582         "pavgw %%mm1, %%mm4                          \n\t"
 583         "pavgw %%mm3, %%mm6                          \n\t"
 584         snow_vertical_compose_r2r_sub("mm7","mm7","mm7","mm7","mm0","mm2","mm4","mm6")
 585         snow_vertical_compose_sra("1","mm0","mm2","mm4","mm6")
 586         snow_vertical_compose_mmx_add("%3","mm0","mm2","mm4","mm6")
 587
 588         snow_vertical_compose_sra("2","mm0","mm2","mm4","mm6")
 589         snow_vertical_compose_mmx_add("%3","mm0","mm2","mm4","mm6")
 590         snow_vertical_compose_mmx_store("%3","mm0","mm2","mm4","mm6")
 591         snow_vertical_compose_mmx_add("%1","mm0","mm2","mm4","mm6")
 592         snow_vertical_compose_mmx_move("mm0","mm2","mm4","mm6","mm1","mm3","mm5","mm7")
 593         snow_vertical_compose_sra("1","mm0","mm2","mm4","mm6")
 594         snow_vertical_compose_r2r_add("mm1","mm3","mm5","mm7","mm0","mm2","mm4","mm6")
 595         snow_vertical_compose_mmx_add("%2","mm0","mm2","mm4","mm6")
 596         snow_vertical_compose_mmx_store("%2","mm0","mm2","mm4","mm6")
 597
 598         "2:                                          \n\t"
 599         "sub $32, %%"REG_d"                          \n\t"
 600         "jge 1b                                      \n\t"
 601         :"+d"(i)
 602         :"r"(b0),"r"(b1),"r"(b2),"r"(b3),"r"(b4),"r"(b5));
 603 }
 604 #endif //HAVE_7REGS
 605
 606 #define snow_inner_add_yblock_sse2_header \
 607     IDWTELEM * * dst_array = sb->line + src_y;\
 608     x86_reg tmp;\
 609     __asm__ volatile(\
 610              "mov  %7, %%"REG_c"             \n\t"\
 611              "mov  %6, %2                    \n\t"\
 612              "mov  %4, %%"REG_S"             \n\t"\
 613              "pxor %%xmm7, %%xmm7            \n\t" /* 0 */\
 614              "pcmpeqd %%xmm3, %%xmm3         \n\t"\
 615              "psllw $15, %%xmm3              \n\t"\
 616              "psrlw $12, %%xmm3              \n\t" /* FRAC_BITS >> 1 */\
 617              "1:                             \n\t"\
 618              "mov %1, %%"REG_D"              \n\t"\
 619              "mov (%%"REG_D"), %%"REG_D"     \n\t"\
 620              "add %3, %%"REG_D"              \n\t"
 621
 622 #define snow_inner_add_yblock_sse2_start_8(out_reg1, out_reg2, ptr_offset, s_offset)\
 623              "mov "PTR_SIZE"*"ptr_offset"(%%"REG_a"), %%"REG_d"; \n\t"\
 624              "movq (%%"REG_d"), %%"out_reg1" \n\t"\
 625              "movq (%%"REG_d", %%"REG_c"), %%"out_reg2" \n\t"\
 626              "punpcklbw %%xmm7, %%"out_reg1" \n\t"\
 627              "punpcklbw %%xmm7, %%"out_reg2" \n\t"\
 628              "movq "s_offset"(%%"REG_S"), %%xmm0 \n\t"\
 629              "movq "s_offset"+16(%%"REG_S"), %%xmm4 \n\t"\
 630              "punpcklbw %%xmm7, %%xmm0       \n\t"\
 631              "punpcklbw %%xmm7, %%xmm4       \n\t"\
 632              "pmullw %%xmm0, %%"out_reg1"    \n\t"\
 633              "pmullw %%xmm4, %%"out_reg2"    \n\t"
 634
 635 #define snow_inner_add_yblock_sse2_start_16(out_reg1, out_reg2, ptr_offset, s_offset)\
 636              "mov "PTR_SIZE"*"ptr_offset"(%%"REG_a"), %%"REG_d"; \n\t"\
 637              "movq (%%"REG_d"), %%"out_reg1" \n\t"\
 638              "movq 8(%%"REG_d"), %%"out_reg2" \n\t"\
 639              "punpcklbw %%xmm7, %%"out_reg1" \n\t"\
 640              "punpcklbw %%xmm7, %%"out_reg2" \n\t"\
 641              "movq "s_offset"(%%"REG_S"), %%xmm0 \n\t"\
 642              "movq "s_offset"+8(%%"REG_S"), %%xmm4 \n\t"\
 643              "punpcklbw %%xmm7, %%xmm0       \n\t"\
 644              "punpcklbw %%xmm7, %%xmm4       \n\t"\
 645              "pmullw %%xmm0, %%"out_reg1"    \n\t"\
 646              "pmullw %%xmm4, %%"out_reg2"    \n\t"
 647
 648 #define snow_inner_add_yblock_sse2_accum_8(ptr_offset, s_offset) \
 649              snow_inner_add_yblock_sse2_start_8("xmm2", "xmm6", ptr_offset, s_offset)\
 650              "paddusw %%xmm2, %%xmm1         \n\t"\
 651              "paddusw %%xmm6, %%xmm5         \n\t"
 652
 653 #define snow_inner_add_yblock_sse2_accum_16(ptr_offset, s_offset) \
 654              snow_inner_add_yblock_sse2_start_16("xmm2", "xmm6", ptr_offset, s_offset)\
 655              "paddusw %%xmm2, %%xmm1         \n\t"\
 656              "paddusw %%xmm6, %%xmm5         \n\t"
 657
 658 #define snow_inner_add_yblock_sse2_end_common1\
 659              "add $32, %%"REG_S"             \n\t"\
 660              "add %%"REG_c", %0              \n\t"\
 661              "add %%"REG_c", "PTR_SIZE"*3(%%"REG_a");\n\t"\
 662              "add %%"REG_c", "PTR_SIZE"*2(%%"REG_a");\n\t"\
 663              "add %%"REG_c", "PTR_SIZE"*1(%%"REG_a");\n\t"\
 664              "add %%"REG_c", (%%"REG_a")     \n\t"
 665
 666 #define snow_inner_add_yblock_sse2_end_common2\
 667              "jnz 1b                         \n\t"\
 668              :"+m"(dst8),"+m"(dst_array),"=&r"(tmp)\
 669              :\
 670              "rm"((x86_reg)(src_x<<1)),"m"(obmc),"a"(block),"m"(b_h),"m"(src_stride):\
 671              "%"REG_c"","%"REG_S"","%"REG_D"","%"REG_d"");
 672
 673 #define snow_inner_add_yblock_sse2_end_8\
 674              "sal $1, %%"REG_c"              \n\t"\
 675              "add $"PTR_SIZE"*2, %1          \n\t"\
 676              snow_inner_add_yblock_sse2_end_common1\
 677              "sar $1, %%"REG_c"              \n\t"\
 678              "sub $2, %2                     \n\t"\
 679              snow_inner_add_yblock_sse2_end_common2
 680
 681 #define snow_inner_add_yblock_sse2_end_16\
 682              "add $"PTR_SIZE"*1, %1          \n\t"\
 683              snow_inner_add_yblock_sse2_end_common1\
 684              "dec %2                         \n\t"\
 685              snow_inner_add_yblock_sse2_end_common2
 686
 687 static void inner_add_yblock_bw_8_obmc_16_bh_even_sse2(const uint8_t *obmc, const x86_reg obmc_stride, uint8_t * * block, int b_w, x86_reg b_h,
 688                       int src_x, int src_y, x86_reg src_stride, slice_buffer * sb, int add, uint8_t * dst8){
 689 snow_inner_add_yblock_sse2_header
 690 snow_inner_add_yblock_sse2_start_8("xmm1", "xmm5", "3", "0")
 691 snow_inner_add_yblock_sse2_accum_8("2", "8")
 692 snow_inner_add_yblock_sse2_accum_8("1", "128")
 693 snow_inner_add_yblock_sse2_accum_8("0", "136")
 694
 695              "mov %0, %%"REG_d"              \n\t"
 696              "movdqa (%%"REG_D"), %%xmm0     \n\t"
 697              "movdqa %%xmm1, %%xmm2          \n\t"
 698
 699              "punpckhwd %%xmm7, %%xmm1       \n\t"
 700              "punpcklwd %%xmm7, %%xmm2       \n\t"
 701              "paddd %%xmm2, %%xmm0           \n\t"
 702              "movdqa 16(%%"REG_D"), %%xmm2   \n\t"
 703              "paddd %%xmm1, %%xmm2           \n\t"
 704              "paddd %%xmm3, %%xmm0           \n\t"
 705              "paddd %%xmm3, %%xmm2           \n\t"
 706
 707              "mov %1, %%"REG_D"              \n\t"
 708              "mov "PTR_SIZE"(%%"REG_D"), %%"REG_D";\n\t"
 709              "add %3, %%"REG_D"              \n\t"
 710
 711              "movdqa (%%"REG_D"), %%xmm4     \n\t"
 712              "movdqa %%xmm5, %%xmm6          \n\t"
 713              "punpckhwd %%xmm7, %%xmm5       \n\t"
 714              "punpcklwd %%xmm7, %%xmm6       \n\t"
 715              "paddd %%xmm6, %%xmm4           \n\t"
 716              "movdqa 16(%%"REG_D"), %%xmm6   \n\t"
 717              "paddd %%xmm5, %%xmm6           \n\t"
 718              "paddd %%xmm3, %%xmm4           \n\t"
 719              "paddd %%xmm3, %%xmm6           \n\t"
 720
 721              "psrad $8, %%xmm0               \n\t" /* FRAC_BITS. */
 722              "psrad $8, %%xmm2               \n\t" /* FRAC_BITS. */
 723              "packssdw %%xmm2, %%xmm0        \n\t"
 724              "packuswb %%xmm7, %%xmm0        \n\t"
 725              "movq %%xmm0, (%%"REG_d")       \n\t"
 726
 727              "psrad $8, %%xmm4               \n\t" /* FRAC_BITS. */
 728              "psrad $8, %%xmm6               \n\t" /* FRAC_BITS. */
 729              "packssdw %%xmm6, %%xmm4        \n\t"
 730              "packuswb %%xmm7, %%xmm4        \n\t"
 731              "movq %%xmm4, (%%"REG_d",%%"REG_c");\n\t"
 732 snow_inner_add_yblock_sse2_end_8
 733 }
 734
 735 static void inner_add_yblock_bw_16_obmc_32_sse2(const uint8_t *obmc, const x86_reg obmc_stride, uint8_t * * block, int b_w, x86_reg b_h,
 736                       int src_x, int src_y, x86_reg src_stride, slice_buffer * sb, int add, uint8_t * dst8){
 737 snow_inner_add_yblock_sse2_header
 738 snow_inner_add_yblock_sse2_start_16("xmm1", "xmm5", "3", "0")
 739 snow_inner_add_yblock_sse2_accum_16("2", "16")
 740 snow_inner_add_yblock_sse2_accum_16("1", "512")
 741 snow_inner_add_yblock_sse2_accum_16("0", "528")
 742
 743              "mov %0, %%"REG_d"              \n\t"
 744              "psrlw $4, %%xmm1               \n\t"
 745              "psrlw $4, %%xmm5               \n\t"
 746              "paddw   (%%"REG_D"), %%xmm1    \n\t"
 747              "paddw 16(%%"REG_D"), %%xmm5    \n\t"
 748              "paddw %%xmm3, %%xmm1           \n\t"
 749              "paddw %%xmm3, %%xmm5           \n\t"
 750              "psraw $4, %%xmm1               \n\t" /* FRAC_BITS. */
 751              "psraw $4, %%xmm5               \n\t" /* FRAC_BITS. */
 752              "packuswb %%xmm5, %%xmm1        \n\t"
 753
 754              "movdqu %%xmm1, (%%"REG_d")       \n\t"
 755
 756 snow_inner_add_yblock_sse2_end_16
 757 }
 758
 759 #define snow_inner_add_yblock_mmx_header \
 760     IDWTELEM * * dst_array = sb->line + src_y;\
 761     x86_reg tmp;\
 762     __asm__ volatile(\
 763              "mov  %7, %%"REG_c"             \n\t"\
 764              "mov  %6, %2                    \n\t"\
 765              "mov  %4, %%"REG_S"             \n\t"\
 766              "pxor %%mm7, %%mm7              \n\t" /* 0 */\
 767              "pcmpeqd %%mm3, %%mm3           \n\t"\
 768              "psllw $15, %%mm3               \n\t"\
 769              "psrlw $12, %%mm3               \n\t" /* FRAC_BITS >> 1 */\
 770              "1:                             \n\t"\
 771              "mov %1, %%"REG_D"              \n\t"\
 772              "mov (%%"REG_D"), %%"REG_D"     \n\t"\
 773              "add %3, %%"REG_D"              \n\t"
 774
 775 #define snow_inner_add_yblock_mmx_start(out_reg1, out_reg2, ptr_offset, s_offset, d_offset)\
 776              "mov "PTR_SIZE"*"ptr_offset"(%%"REG_a"), %%"REG_d"; \n\t"\
 777              "movd "d_offset"(%%"REG_d"), %%"out_reg1" \n\t"\
 778              "movd "d_offset"+4(%%"REG_d"), %%"out_reg2" \n\t"\
 779              "punpcklbw %%mm7, %%"out_reg1" \n\t"\
 780              "punpcklbw %%mm7, %%"out_reg2" \n\t"\
 781              "movd "s_offset"(%%"REG_S"), %%mm0 \n\t"\
 782              "movd "s_offset"+4(%%"REG_S"), %%mm4 \n\t"\
 783              "punpcklbw %%mm7, %%mm0       \n\t"\
 784              "punpcklbw %%mm7, %%mm4       \n\t"\
 785              "pmullw %%mm0, %%"out_reg1"    \n\t"\
 786              "pmullw %%mm4, %%"out_reg2"    \n\t"
 787
 788 #define snow_inner_add_yblock_mmx_accum(ptr_offset, s_offset, d_offset) \
 789              snow_inner_add_yblock_mmx_start("mm2", "mm6", ptr_offset, s_offset, d_offset)\
 790              "paddusw %%mm2, %%mm1         \n\t"\
 791              "paddusw %%mm6, %%mm5         \n\t"
 792
 793 #define snow_inner_add_yblock_mmx_mix(read_offset, write_offset)\
 794              "mov %0, %%"REG_d"              \n\t"\
 795              "psrlw $4, %%mm1                \n\t"\
 796              "psrlw $4, %%mm5                \n\t"\
 797              "paddw "read_offset"(%%"REG_D"), %%mm1 \n\t"\
 798              "paddw "read_offset"+8(%%"REG_D"), %%mm5 \n\t"\
 799              "paddw %%mm3, %%mm1             \n\t"\
 800              "paddw %%mm3, %%mm5             \n\t"\
 801              "psraw $4, %%mm1                \n\t"\
 802              "psraw $4, %%mm5                \n\t"\
 803              "packuswb %%mm5, %%mm1          \n\t"\
 804              "movq %%mm1, "write_offset"(%%"REG_d") \n\t"
 805
 806 #define snow_inner_add_yblock_mmx_end(s_step)\
 807              "add $"s_step", %%"REG_S"             \n\t"\
 808              "add %%"REG_c", "PTR_SIZE"*3(%%"REG_a");\n\t"\
 809              "add %%"REG_c", "PTR_SIZE"*2(%%"REG_a");\n\t"\
 810              "add %%"REG_c", "PTR_SIZE"*1(%%"REG_a");\n\t"\
 811              "add %%"REG_c", (%%"REG_a")     \n\t"\
 812              "add $"PTR_SIZE"*1, %1          \n\t"\
 813              "add %%"REG_c", %0              \n\t"\
 814              "dec %2                         \n\t"\
 815              "jnz 1b                         \n\t"\
 816              :"+m"(dst8),"+m"(dst_array),"=&r"(tmp)\
 817              :\
 818              "rm"((x86_reg)(src_x<<1)),"m"(obmc),"a"(block),"m"(b_h),"m"(src_stride):\
 819              "%"REG_c"","%"REG_S"","%"REG_D"","%"REG_d"");
 820
 821 static void inner_add_yblock_bw_8_obmc_16_mmx(const uint8_t *obmc, const x86_reg obmc_stride, uint8_t * * block, int b_w, x86_reg b_h,
 822                       int src_x, int src_y, x86_reg src_stride, slice_buffer * sb, int add, uint8_t * dst8){
 823 snow_inner_add_yblock_mmx_header
 824 snow_inner_add_yblock_mmx_start("mm1", "mm5", "3", "0", "0")
 825 snow_inner_add_yblock_mmx_accum("2", "8", "0")
 826 snow_inner_add_yblock_mmx_accum("1", "128", "0")
 827 snow_inner_add_yblock_mmx_accum("0", "136", "0")
 828 snow_inner_add_yblock_mmx_mix("0", "0")
 829 snow_inner_add_yblock_mmx_end("16")
 830 }
 831
 832 static void inner_add_yblock_bw_16_obmc_32_mmx(const uint8_t *obmc, const x86_reg obmc_stride, uint8_t * * block, int b_w, x86_reg b_h,
 833                       int src_x, int src_y, x86_reg src_stride, slice_buffer * sb, int add, uint8_t * dst8){
 834 snow_inner_add_yblock_mmx_header
 835 snow_inner_add_yblock_mmx_start("mm1", "mm5", "3", "0", "0")
 836 snow_inner_add_yblock_mmx_accum("2", "16", "0")
 837 snow_inner_add_yblock_mmx_accum("1", "512", "0")
 838 snow_inner_add_yblock_mmx_accum("0", "528", "0")
 839 snow_inner_add_yblock_mmx_mix("0", "0")
 840
 841 snow_inner_add_yblock_mmx_start("mm1", "mm5", "3", "8", "8")
 842 snow_inner_add_yblock_mmx_accum("2", "24", "8")
 843 snow_inner_add_yblock_mmx_accum("1", "520", "8")
 844 snow_inner_add_yblock_mmx_accum("0", "536", "8")
 845 snow_inner_add_yblock_mmx_mix("16", "8")
 846 snow_inner_add_yblock_mmx_end("32")
 847 }
 848
 849 void ff_snow_inner_add_yblock_sse2(const uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
 850                            int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8){
 851
 852     if (b_w == 16)
 853         inner_add_yblock_bw_16_obmc_32_sse2(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
 854     else if (b_w == 8 && obmc_stride == 16) {
 855         if (!(b_h & 1))
 856             inner_add_yblock_bw_8_obmc_16_bh_even_sse2(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
 857         else
 858             inner_add_yblock_bw_8_obmc_16_mmx(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
 859     } else
 860          ff_snow_inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
 861 }
 862
 863 void ff_snow_inner_add_yblock_mmx(const uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
 864                           int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8){
 865     if (b_w == 16)
 866         inner_add_yblock_bw_16_obmc_32_mmx(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
 867     else if (b_w == 8 && obmc_stride == 16)
 868         inner_add_yblock_bw_8_obmc_16_mmx(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
 869     else
 870         ff_snow_inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
 871 }