git.sesse.net Git - ffmpeg/blob - libavcodec/i386/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 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
  20 #include "../avcodec.h"
  21 #include "../snow.h"
  22 #include "x86_cpu.h"
  23
  24 void ff_snow_horizontal_compose97i_sse2(DWTELEM *b, int width){
  25     const int w2= (width+1)>>1;
  26     // SSE2 code runs faster with pointers aligned on a 32-byte boundary.
  27     DWTELEM temp_buf[(width>>1) + 4];
  28     DWTELEM * const temp = temp_buf + 4 - (((int)temp_buf & 0xF) >> 2);
  29     const int w_l= (width>>1);
  30     const int w_r= w2 - 1;
  31     int i;
  32
  33     { // Lift 0
  34         DWTELEM * const ref = b + w2 - 1;
  35         DWTELEM 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             "pslld        $31, %%xmm7         \n\t"
  44             "psrld        $29, %%xmm7         \n\t"
  45         ::);
  46         for(; i<w_l-7; i+=8){
  47             asm volatile(
  48                 "movdqu   (%1), %%xmm1        \n\t"
  49                 "movdqu 16(%1), %%xmm5        \n\t"
  50                 "movdqu  4(%1), %%xmm2        \n\t"
  51                 "movdqu 20(%1), %%xmm6        \n\t"
  52                 "paddd  %%xmm1, %%xmm2        \n\t"
  53                 "paddd  %%xmm5, %%xmm6        \n\t"
  54                 "movdqa %%xmm2, %%xmm0        \n\t"
  55                 "movdqa %%xmm6, %%xmm4        \n\t"
  56                 "paddd  %%xmm2, %%xmm2        \n\t"
  57                 "paddd  %%xmm6, %%xmm6        \n\t"
  58                 "paddd  %%xmm0, %%xmm2        \n\t"
  59                 "paddd  %%xmm4, %%xmm6        \n\t"
  60                 "paddd  %%xmm7, %%xmm2        \n\t"
  61                 "paddd  %%xmm7, %%xmm6        \n\t"
  62                 "psrad      $3, %%xmm2        \n\t"
  63                 "psrad      $3, %%xmm6        \n\t"
  64                 "movdqa   (%0), %%xmm0        \n\t"
  65                 "movdqa 16(%0), %%xmm4        \n\t"
  66                 "psubd  %%xmm2, %%xmm0        \n\t"
  67                 "psubd  %%xmm6, %%xmm4        \n\t"
  68                 "movdqa %%xmm0, (%0)          \n\t"
  69                 "movdqa %%xmm4, 16(%0)        \n\t"
  70                 :: "r"(&b[i]), "r"(&ref[i])
  71                 : "memory"
  72             );
  73         }
  74         snow_horizontal_compose_lift_lead_out(i, b, b, ref, width, w_l, 0, W_DM, W_DO, W_DS);
  75         b[0] = b_0 - ((W_DM * 2 * ref[1]+W_DO)>>W_DS);
  76     }
  77
  78     { // Lift 1
  79         DWTELEM * const dst = b+w2;
  80
  81         i = 0;
  82         for(; (((long)&dst[i]) & 0xF) && i<w_r; i++){
  83             dst[i] = dst[i] - (b[i] + b[i + 1]);
  84         }
  85         for(; i<w_r-7; i+=8){
  86             asm volatile(
  87                 "movdqu   (%1), %%xmm1        \n\t"
  88                 "movdqu 16(%1), %%xmm5        \n\t"
  89                 "movdqu  4(%1), %%xmm2        \n\t"
  90                 "movdqu 20(%1), %%xmm6        \n\t"
  91                 "paddd  %%xmm1, %%xmm2        \n\t"
  92                 "paddd  %%xmm5, %%xmm6        \n\t"
  93                 "movdqa   (%0), %%xmm0        \n\t"
  94                 "movdqa 16(%0), %%xmm4        \n\t"
  95                 "psubd  %%xmm2, %%xmm0        \n\t"
  96                 "psubd  %%xmm6, %%xmm4        \n\t"
  97                 "movdqa %%xmm0, (%0)          \n\t"
  98                 "movdqa %%xmm4, 16(%0)        \n\t"
  99                 :: "r"(&dst[i]), "r"(&b[i])
 100                 : "memory"
 101             );
 102         }
 103         snow_horizontal_compose_lift_lead_out(i, dst, dst, b, width, w_r, 1, W_CM, W_CO, W_CS);
 104     }
 105
 106     { // Lift 2
 107         DWTELEM * const ref = b+w2 - 1;
 108         DWTELEM b_0 = b[0];
 109
 110         i = 0;
 111         asm volatile(
 112             "pslld          $1, %%xmm7       \n\t" /* xmm7 already holds a '4' from 2 lifts ago. */
 113         ::);
 114         for(; i<w_l-7; i+=8){
 115             asm volatile(
 116                 "movdqu   (%1), %%xmm1        \n\t"
 117                 "movdqu 16(%1), %%xmm5        \n\t"
 118                 "movdqu  4(%1), %%xmm0        \n\t"
 119                 "movdqu 20(%1), %%xmm4        \n\t"
 120                 "paddd  %%xmm1, %%xmm0        \n\t"
 121                 "paddd  %%xmm5, %%xmm4        \n\t"
 122                 "movdqa %%xmm7, %%xmm1        \n\t"
 123                 "movdqa %%xmm7, %%xmm5        \n\t"
 124                 "psubd  %%xmm0, %%xmm1        \n\t"
 125                 "psubd  %%xmm4, %%xmm5        \n\t"
 126                 "movdqa   (%0), %%xmm0        \n\t"
 127                 "movdqa 16(%0), %%xmm4        \n\t"
 128                 "pslld      $2, %%xmm0        \n\t"
 129                 "pslld      $2, %%xmm4        \n\t"
 130                 "psubd  %%xmm0, %%xmm1        \n\t"
 131                 "psubd  %%xmm4, %%xmm5        \n\t"
 132                 "psrad      $4, %%xmm1        \n\t"
 133                 "psrad      $4, %%xmm5        \n\t"
 134                 "movdqa   (%0), %%xmm0        \n\t"
 135                 "movdqa 16(%0), %%xmm4        \n\t"
 136                 "psubd  %%xmm1, %%xmm0        \n\t"
 137                 "psubd  %%xmm5, %%xmm4        \n\t"
 138                 "movdqa %%xmm0, (%0)          \n\t"
 139                 "movdqa %%xmm4, 16(%0)        \n\t"
 140                 :: "r"(&b[i]), "r"(&ref[i])
 141                 : "memory"
 142             );
 143         }
 144         snow_horizontal_compose_liftS_lead_out(i, b, b, ref, width, w_l);
 145         b[0] = b_0 - (((-2 * ref[1] + W_BO) - 4 * b_0) >> W_BS);
 146     }
 147
 148     { // Lift 3
 149         DWTELEM * const src = b+w2;
 150
 151         i = 0;
 152         for(; (((long)&temp[i]) & 0xF) && i<w_r; i++){
 153             temp[i] = src[i] - ((-W_AM*(b[i] + b[i+1]))>>W_AS);
 154         }
 155         for(; i<w_r-7; i+=8){
 156             asm volatile(
 157                 "movdqu  4(%1), %%xmm2        \n\t"
 158                 "movdqu 20(%1), %%xmm6        \n\t"
 159                 "paddd    (%1), %%xmm2        \n\t"
 160                 "paddd  16(%1), %%xmm6        \n\t"
 161                 "movdqa %%xmm2, %%xmm0        \n\t"
 162                 "movdqa %%xmm6, %%xmm4        \n\t"
 163                 "pslld      $2, %%xmm2        \n\t"
 164                 "pslld      $2, %%xmm6        \n\t"
 165                 "psubd  %%xmm2, %%xmm0        \n\t"
 166                 "psubd  %%xmm6, %%xmm4        \n\t"
 167                 "psrad      $1, %%xmm0        \n\t"
 168                 "psrad      $1, %%xmm4        \n\t"
 169                 "movdqu   (%0), %%xmm2        \n\t"
 170                 "movdqu 16(%0), %%xmm6        \n\t"
 171                 "psubd  %%xmm0, %%xmm2        \n\t"
 172                 "psubd  %%xmm4, %%xmm6        \n\t"
 173                 "movdqa %%xmm2, (%2)          \n\t"
 174                 "movdqa %%xmm6, 16(%2)        \n\t"
 175                 :: "r"(&src[i]), "r"(&b[i]), "r"(&temp[i])
 176                  : "memory"
 177                );
 178         }
 179         snow_horizontal_compose_lift_lead_out(i, temp, src, b, width, w_r, 1, -W_AM, W_AO, W_AS);
 180     }
 181
 182     {
 183         snow_interleave_line_header(&i, width, b, temp);
 184
 185         for (; (i & 0x1E) != 0x1E; i-=2){
 186             b[i+1] = temp[i>>1];
 187             b[i] = b[i>>1];
 188         }
 189         for (i-=30; i>=0; i-=32){
 190             asm volatile(
 191                 "movdqa      (%1), %%xmm0       \n\t"
 192                 "movdqa    16(%1), %%xmm2       \n\t"
 193                 "movdqa    32(%1), %%xmm4       \n\t"
 194                 "movdqa    48(%1), %%xmm6       \n\t"
 195                 "movdqa      (%1), %%xmm1       \n\t"
 196                 "movdqa    16(%1), %%xmm3       \n\t"
 197                 "movdqa    32(%1), %%xmm5       \n\t"
 198                 "movdqa    48(%1), %%xmm7       \n\t"
 199                 "punpckldq   (%2), %%xmm0       \n\t"
 200                 "punpckldq 16(%2), %%xmm2       \n\t"
 201                 "punpckldq 32(%2), %%xmm4       \n\t"
 202                 "punpckldq 48(%2), %%xmm6       \n\t"
 203                 "movdqa    %%xmm0, (%0)         \n\t"
 204                 "movdqa    %%xmm2, 32(%0)       \n\t"
 205                 "movdqa    %%xmm4, 64(%0)       \n\t"
 206                 "movdqa    %%xmm6, 96(%0)       \n\t"
 207                 "punpckhdq   (%2), %%xmm1       \n\t"
 208                 "punpckhdq 16(%2), %%xmm3       \n\t"
 209                 "punpckhdq 32(%2), %%xmm5       \n\t"
 210                 "punpckhdq 48(%2), %%xmm7       \n\t"
 211                 "movdqa    %%xmm1, 16(%0)       \n\t"
 212                 "movdqa    %%xmm3, 48(%0)       \n\t"
 213                 "movdqa    %%xmm5, 80(%0)       \n\t"
 214                 "movdqa    %%xmm7, 112(%0)      \n\t"
 215                 :: "r"(&(b)[i]), "r"(&(b)[i>>1]), "r"(&(temp)[i>>1])
 216                  : "memory"
 217                );
 218         }
 219     }
 220 }
 221
 222 void ff_snow_horizontal_compose97i_mmx(DWTELEM *b, int width){
 223     const int w2= (width+1)>>1;
 224     DWTELEM temp[width >> 1];
 225     const int w_l= (width>>1);
 226     const int w_r= w2 - 1;
 227     int i;
 228
 229     { // Lift 0
 230         DWTELEM * const ref = b + w2 - 1;
 231
 232         i = 1;
 233         b[0] = b[0] - ((W_DM * 2 * ref[1]+W_DO)>>W_DS);
 234         asm volatile(
 235             "pcmpeqd    %%mm7, %%mm7         \n\t"
 236             "pslld        $31, %%mm7         \n\t"
 237             "psrld        $29, %%mm7         \n\t"
 238            ::);
 239         for(; i<w_l-3; i+=4){
 240             asm volatile(
 241                 "movq     (%1), %%mm2        \n\t"
 242                 "movq    8(%1), %%mm6        \n\t"
 243                 "paddd   4(%1), %%mm2        \n\t"
 244                 "paddd  12(%1), %%mm6        \n\t"
 245                 "movq    %%mm2, %%mm0        \n\t"
 246                 "movq    %%mm6, %%mm4        \n\t"
 247                 "paddd   %%mm2, %%mm2        \n\t"
 248                 "paddd   %%mm6, %%mm6        \n\t"
 249                 "paddd   %%mm0, %%mm2        \n\t"
 250                 "paddd   %%mm4, %%mm6        \n\t"
 251                 "paddd   %%mm7, %%mm2        \n\t"
 252                 "paddd   %%mm7, %%mm6        \n\t"
 253                 "psrad      $3, %%mm2        \n\t"
 254                 "psrad      $3, %%mm6        \n\t"
 255                 "movq     (%0), %%mm0        \n\t"
 256                 "movq    8(%0), %%mm4        \n\t"
 257                 "psubd   %%mm2, %%mm0        \n\t"
 258                 "psubd   %%mm6, %%mm4        \n\t"
 259                 "movq    %%mm0, (%0)         \n\t"
 260                 "movq    %%mm4, 8(%0)        \n\t"
 261                 :: "r"(&b[i]), "r"(&ref[i])
 262                  : "memory"
 263                );
 264         }
 265         snow_horizontal_compose_lift_lead_out(i, b, b, ref, width, w_l, 0, W_DM, W_DO, W_DS);
 266     }
 267
 268     { // Lift 1
 269         DWTELEM * const dst = b+w2;
 270
 271         i = 0;
 272         for(; i<w_r-3; i+=4){
 273             asm volatile(
 274                 "movq     (%1), %%mm2        \n\t"
 275                 "movq    8(%1), %%mm6        \n\t"
 276                 "paddd   4(%1), %%mm2        \n\t"
 277                 "paddd  12(%1), %%mm6        \n\t"
 278                 "movq     (%0), %%mm0        \n\t"
 279                 "movq    8(%0), %%mm4        \n\t"
 280                 "psubd   %%mm2, %%mm0        \n\t"
 281                 "psubd   %%mm6, %%mm4        \n\t"
 282                 "movq    %%mm0, (%0)         \n\t"
 283                 "movq    %%mm4, 8(%0)        \n\t"
 284                 :: "r"(&dst[i]), "r"(&b[i])
 285                  : "memory"
 286                );
 287         }
 288         snow_horizontal_compose_lift_lead_out(i, dst, dst, b, width, w_r, 1, W_CM, W_CO, W_CS);
 289     }
 290
 291     { // Lift 2
 292         DWTELEM * const ref = b+w2 - 1;
 293
 294         i = 1;
 295         b[0] = b[0] - (((-2 * ref[1] + W_BO) - 4 * b[0]) >> W_BS);
 296         asm volatile(
 297             "pslld          $1, %%mm7       \n\t" /* xmm7 already holds a '4' from 2 lifts ago. */
 298            ::);
 299         for(; i<w_l-3; i+=4){
 300             asm volatile(
 301                 "movq     (%1), %%mm0        \n\t"
 302                 "movq    8(%1), %%mm4        \n\t"
 303                 "paddd   4(%1), %%mm0        \n\t"
 304                 "paddd  12(%1), %%mm4        \n\t"
 305                 "movq    %%mm7, %%mm1        \n\t"
 306                 "movq    %%mm7, %%mm5        \n\t"
 307                 "psubd   %%mm0, %%mm1        \n\t"
 308                 "psubd   %%mm4, %%mm5        \n\t"
 309                 "movq     (%0), %%mm0        \n\t"
 310                 "movq    8(%0), %%mm4        \n\t"
 311                 "pslld      $2, %%mm0        \n\t"
 312                 "pslld      $2, %%mm4        \n\t"
 313                 "psubd   %%mm0, %%mm1        \n\t"
 314                 "psubd   %%mm4, %%mm5        \n\t"
 315                 "psrad      $4, %%mm1        \n\t"
 316                 "psrad      $4, %%mm5        \n\t"
 317                 "movq     (%0), %%mm0        \n\t"
 318                 "movq    8(%0), %%mm4        \n\t"
 319                 "psubd   %%mm1, %%mm0        \n\t"
 320                 "psubd   %%mm5, %%mm4        \n\t"
 321                 "movq    %%mm0, (%0)         \n\t"
 322                 "movq    %%mm4, 8(%0)        \n\t"
 323                 :: "r"(&b[i]), "r"(&ref[i])
 324                  : "memory"
 325                );
 326         }
 327         snow_horizontal_compose_liftS_lead_out(i, b, b, ref, width, w_l);
 328     }
 329
 330     { // Lift 3
 331         DWTELEM * const src = b+w2;
 332         i = 0;
 333
 334         for(; i<w_r-3; i+=4){
 335             asm volatile(
 336                 "movq    4(%1), %%mm2        \n\t"
 337                 "movq   12(%1), %%mm6        \n\t"
 338                 "paddd    (%1), %%mm2        \n\t"
 339                 "paddd   8(%1), %%mm6        \n\t"
 340                 "movq    %%mm2, %%mm0        \n\t"
 341                 "movq    %%mm6, %%mm4        \n\t"
 342                 "pslld      $2, %%mm2        \n\t"
 343                 "pslld      $2, %%mm6        \n\t"
 344                 "psubd   %%mm2, %%mm0        \n\t"
 345                 "psubd   %%mm6, %%mm4        \n\t"
 346                 "psrad      $1, %%mm0        \n\t"
 347                 "psrad      $1, %%mm4        \n\t"
 348                 "movq     (%0), %%mm2        \n\t"
 349                 "movq    8(%0), %%mm6        \n\t"
 350                 "psubd   %%mm0, %%mm2        \n\t"
 351                 "psubd   %%mm4, %%mm6        \n\t"
 352                 "movq    %%mm2, (%2)         \n\t"
 353                 "movq    %%mm6, 8(%2)        \n\t"
 354                 :: "r"(&src[i]), "r"(&b[i]), "r"(&temp[i])
 355                  : "memory"
 356                );
 357         }
 358         snow_horizontal_compose_lift_lead_out(i, temp, src, b, width, w_r, 1, -W_AM, W_AO, W_AS);
 359     }
 360
 361     {
 362         snow_interleave_line_header(&i, width, b, temp);
 363
 364         for (; (i & 0xE) != 0xE; i-=2){
 365             b[i+1] = temp[i>>1];
 366             b[i] = b[i>>1];
 367         }
 368         for (i-=14; i>=0; i-=16){
 369             asm volatile(
 370                 "movq        (%1), %%mm0       \n\t"
 371                 "movq       8(%1), %%mm2       \n\t"
 372                 "movq      16(%1), %%mm4       \n\t"
 373                 "movq      24(%1), %%mm6       \n\t"
 374                 "movq        (%1), %%mm1       \n\t"
 375                 "movq       8(%1), %%mm3       \n\t"
 376                 "movq      16(%1), %%mm5       \n\t"
 377                 "movq      24(%1), %%mm7       \n\t"
 378                 "punpckldq   (%2), %%mm0       \n\t"
 379                 "punpckldq  8(%2), %%mm2       \n\t"
 380                 "punpckldq 16(%2), %%mm4       \n\t"
 381                 "punpckldq 24(%2), %%mm6       \n\t"
 382                 "movq       %%mm0, (%0)        \n\t"
 383                 "movq       %%mm2, 16(%0)      \n\t"
 384                 "movq       %%mm4, 32(%0)      \n\t"
 385                 "movq       %%mm6, 48(%0)      \n\t"
 386                 "punpckhdq   (%2), %%mm1       \n\t"
 387                 "punpckhdq  8(%2), %%mm3       \n\t"
 388                 "punpckhdq 16(%2), %%mm5       \n\t"
 389                 "punpckhdq 24(%2), %%mm7       \n\t"
 390                 "movq       %%mm1, 8(%0)       \n\t"
 391                 "movq       %%mm3, 24(%0)      \n\t"
 392                 "movq       %%mm5, 40(%0)      \n\t"
 393                 "movq       %%mm7, 56(%0)      \n\t"
 394                 :: "r"(&b[i]), "r"(&b[i>>1]), "r"(&temp[i>>1])
 395                  : "memory"
 396                );
 397         }
 398     }
 399 }
 400
 401 #define snow_vertical_compose_sse2_load_add(op,r,t0,t1,t2,t3)\
 402         ""op" (%%"r",%%"REG_d",4), %%"t0"      \n\t"\
 403         ""op" 16(%%"r",%%"REG_d",4), %%"t1"    \n\t"\
 404         ""op" 32(%%"r",%%"REG_d",4), %%"t2"    \n\t"\
 405         ""op" 48(%%"r",%%"REG_d",4), %%"t3"    \n\t"
 406
 407 #define snow_vertical_compose_sse2_load(r,t0,t1,t2,t3)\
 408         snow_vertical_compose_sse2_load_add("movdqa",r,t0,t1,t2,t3)
 409
 410 #define snow_vertical_compose_sse2_add(r,t0,t1,t2,t3)\
 411         snow_vertical_compose_sse2_load_add("paddd",r,t0,t1,t2,t3)
 412
 413 #define snow_vertical_compose_sse2_sub(s0,s1,s2,s3,t0,t1,t2,t3)\
 414         "psubd %%"s0", %%"t0" \n\t"\
 415         "psubd %%"s1", %%"t1" \n\t"\
 416         "psubd %%"s2", %%"t2" \n\t"\
 417         "psubd %%"s3", %%"t3" \n\t"
 418
 419 #define snow_vertical_compose_sse2_store(w,s0,s1,s2,s3)\
 420         "movdqa %%"s0", (%%"w",%%"REG_d",4)      \n\t"\
 421         "movdqa %%"s1", 16(%%"w",%%"REG_d",4)    \n\t"\
 422         "movdqa %%"s2", 32(%%"w",%%"REG_d",4)    \n\t"\
 423         "movdqa %%"s3", 48(%%"w",%%"REG_d",4)    \n\t"
 424
 425 #define snow_vertical_compose_sse2_sra(n,t0,t1,t2,t3)\
 426         "psrad $"n", %%"t0" \n\t"\
 427         "psrad $"n", %%"t1" \n\t"\
 428         "psrad $"n", %%"t2" \n\t"\
 429         "psrad $"n", %%"t3" \n\t"
 430
 431 #define snow_vertical_compose_sse2_r2r_add(s0,s1,s2,s3,t0,t1,t2,t3)\
 432         "paddd %%"s0", %%"t0" \n\t"\
 433         "paddd %%"s1", %%"t1" \n\t"\
 434         "paddd %%"s2", %%"t2" \n\t"\
 435         "paddd %%"s3", %%"t3" \n\t"
 436
 437 #define snow_vertical_compose_sse2_sll(n,t0,t1,t2,t3)\
 438         "pslld $"n", %%"t0" \n\t"\
 439         "pslld $"n", %%"t1" \n\t"\
 440         "pslld $"n", %%"t2" \n\t"\
 441         "pslld $"n", %%"t3" \n\t"
 442
 443 #define snow_vertical_compose_sse2_move(s0,s1,s2,s3,t0,t1,t2,t3)\
 444         "movdqa %%"s0", %%"t0" \n\t"\
 445         "movdqa %%"s1", %%"t1" \n\t"\
 446         "movdqa %%"s2", %%"t2" \n\t"\
 447         "movdqa %%"s3", %%"t3" \n\t"
 448
 449 void ff_snow_vertical_compose97i_sse2(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, DWTELEM *b3, DWTELEM *b4, DWTELEM *b5, int width){
 450     long i = width;
 451
 452     while(i & 0xF)
 453     {
 454         i--;
 455         b4[i] -= (W_DM*(b3[i] + b5[i])+W_DO)>>W_DS;
 456         b3[i] -= (W_CM*(b2[i] + b4[i])+W_CO)>>W_CS;
 457         b2[i] += (W_BM*(b1[i] + b3[i])+4*b2[i]+W_BO)>>W_BS;
 458         b1[i] += (W_AM*(b0[i] + b2[i])+W_AO)>>W_AS;
 459     }
 460
 461          asm volatile (
 462         "jmp 2f                                      \n\t"
 463         "1:                                          \n\t"
 464
 465         "mov %6, %%"REG_a"                           \n\t"
 466         "mov %4, %%"REG_S"                           \n\t"
 467
 468         snow_vertical_compose_sse2_load(REG_S,"xmm0","xmm2","xmm4","xmm6")
 469         snow_vertical_compose_sse2_add(REG_a,"xmm0","xmm2","xmm4","xmm6")
 470         snow_vertical_compose_sse2_move("xmm0","xmm2","xmm4","xmm6","xmm1","xmm3","xmm5","xmm7")
 471         snow_vertical_compose_sse2_sll("1","xmm0","xmm2","xmm4","xmm6")\
 472         snow_vertical_compose_sse2_r2r_add("xmm1","xmm3","xmm5","xmm7","xmm0","xmm2","xmm4","xmm6")
 473
 474         "pcmpeqd %%xmm1, %%xmm1                      \n\t"
 475         "pslld $31, %%xmm1                           \n\t"
 476         "psrld $29, %%xmm1                           \n\t"
 477         "mov %5, %%"REG_a"                           \n\t"
 478
 479         snow_vertical_compose_sse2_r2r_add("xmm1","xmm1","xmm1","xmm1","xmm0","xmm2","xmm4","xmm6")
 480         snow_vertical_compose_sse2_sra("3","xmm0","xmm2","xmm4","xmm6")
 481         snow_vertical_compose_sse2_load(REG_a,"xmm1","xmm3","xmm5","xmm7")
 482         snow_vertical_compose_sse2_sub("xmm0","xmm2","xmm4","xmm6","xmm1","xmm3","xmm5","xmm7")
 483         snow_vertical_compose_sse2_store(REG_a,"xmm1","xmm3","xmm5","xmm7")
 484         "mov %3, %%"REG_c"                           \n\t"
 485         snow_vertical_compose_sse2_load(REG_S,"xmm0","xmm2","xmm4","xmm6")
 486         snow_vertical_compose_sse2_add(REG_c,"xmm1","xmm3","xmm5","xmm7")
 487         snow_vertical_compose_sse2_sub("xmm1","xmm3","xmm5","xmm7","xmm0","xmm2","xmm4","xmm6")
 488         snow_vertical_compose_sse2_store(REG_S,"xmm0","xmm2","xmm4","xmm6")
 489         "mov %2, %%"REG_a"                           \n\t"
 490         snow_vertical_compose_sse2_load(REG_c,"xmm1","xmm3","xmm5","xmm7")
 491         snow_vertical_compose_sse2_add(REG_a,"xmm0","xmm2","xmm4","xmm6")
 492         snow_vertical_compose_sse2_sll("2","xmm1","xmm3","xmm5","xmm7")\
 493         snow_vertical_compose_sse2_r2r_add("xmm1","xmm3","xmm5","xmm7","xmm0","xmm2","xmm4","xmm6")
 494
 495         "pcmpeqd %%xmm1, %%xmm1                      \n\t"
 496         "pslld $31, %%xmm1                           \n\t"
 497         "psrld $28, %%xmm1                           \n\t"
 498         "mov %1, %%"REG_S"                           \n\t"
 499
 500         snow_vertical_compose_sse2_r2r_add("xmm1","xmm1","xmm1","xmm1","xmm0","xmm2","xmm4","xmm6")
 501         snow_vertical_compose_sse2_sra("4","xmm0","xmm2","xmm4","xmm6")
 502         snow_vertical_compose_sse2_add(REG_c,"xmm0","xmm2","xmm4","xmm6")
 503         snow_vertical_compose_sse2_store(REG_c,"xmm0","xmm2","xmm4","xmm6")
 504         snow_vertical_compose_sse2_add(REG_S,"xmm0","xmm2","xmm4","xmm6")
 505         snow_vertical_compose_sse2_move("xmm0","xmm2","xmm4","xmm6","xmm1","xmm3","xmm5","xmm7")
 506         snow_vertical_compose_sse2_sll("1","xmm0","xmm2","xmm4","xmm6")\
 507         snow_vertical_compose_sse2_r2r_add("xmm1","xmm3","xmm5","xmm7","xmm0","xmm2","xmm4","xmm6")
 508         snow_vertical_compose_sse2_sra("1","xmm0","xmm2","xmm4","xmm6")
 509         snow_vertical_compose_sse2_add(REG_a,"xmm0","xmm2","xmm4","xmm6")
 510         snow_vertical_compose_sse2_store(REG_a,"xmm0","xmm2","xmm4","xmm6")
 511
 512         "2:                                          \n\t"
 513         "sub $16, %%"REG_d"                          \n\t"
 514         "jge 1b                                      \n\t"
 515         :"+d"(i)
 516         :
 517         "m"(b0),"m"(b1),"m"(b2),"m"(b3),"m"(b4),"m"(b5):
 518         "%"REG_a"","%"REG_S"","%"REG_c"");
 519 }
 520
 521 #define snow_vertical_compose_mmx_load_add(op,r,t0,t1,t2,t3)\
 522         ""op" (%%"r",%%"REG_d",4), %%"t0"   \n\t"\
 523         ""op" 8(%%"r",%%"REG_d",4), %%"t1"  \n\t"\
 524         ""op" 16(%%"r",%%"REG_d",4), %%"t2" \n\t"\
 525         ""op" 24(%%"r",%%"REG_d",4), %%"t3" \n\t"
 526
 527 #define snow_vertical_compose_mmx_load(r,t0,t1,t2,t3)\
 528         snow_vertical_compose_mmx_load_add("movq",r,t0,t1,t2,t3)
 529
 530 #define snow_vertical_compose_mmx_add(r,t0,t1,t2,t3)\
 531         snow_vertical_compose_mmx_load_add("paddd",r,t0,t1,t2,t3)
 532
 533 #define snow_vertical_compose_mmx_sub(s0,s1,s2,s3,t0,t1,t2,t3)\
 534         snow_vertical_compose_sse2_sub(s0,s1,s2,s3,t0,t1,t2,t3)
 535
 536 #define snow_vertical_compose_mmx_store(w,s0,s1,s2,s3)\
 537         "movq %%"s0", (%%"w",%%"REG_d",4)   \n\t"\
 538         "movq %%"s1", 8(%%"w",%%"REG_d",4)  \n\t"\
 539         "movq %%"s2", 16(%%"w",%%"REG_d",4) \n\t"\
 540         "movq %%"s3", 24(%%"w",%%"REG_d",4) \n\t"
 541
 542 #define snow_vertical_compose_mmx_sra(n,t0,t1,t2,t3)\
 543         snow_vertical_compose_sse2_sra(n,t0,t1,t2,t3)
 544
 545 #define snow_vertical_compose_mmx_r2r_add(s0,s1,s2,s3,t0,t1,t2,t3)\
 546         snow_vertical_compose_sse2_r2r_add(s0,s1,s2,s3,t0,t1,t2,t3)
 547
 548 #define snow_vertical_compose_mmx_sll(n,t0,t1,t2,t3)\
 549         snow_vertical_compose_sse2_sll(n,t0,t1,t2,t3)
 550
 551 #define snow_vertical_compose_mmx_move(s0,s1,s2,s3,t0,t1,t2,t3)\
 552         "movq %%"s0", %%"t0" \n\t"\
 553         "movq %%"s1", %%"t1" \n\t"\
 554         "movq %%"s2", %%"t2" \n\t"\
 555         "movq %%"s3", %%"t3" \n\t"
 556
 557 void ff_snow_vertical_compose97i_mmx(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, DWTELEM *b3, DWTELEM *b4, DWTELEM *b5, int width){
 558     long i = width;
 559     while(i & 0x7)
 560     {
 561         i--;
 562         b4[i] -= (W_DM*(b3[i] + b5[i])+W_DO)>>W_DS;
 563         b3[i] -= (W_CM*(b2[i] + b4[i])+W_CO)>>W_CS;
 564         b2[i] += (W_BM*(b1[i] + b3[i])+4*b2[i]+W_BO)>>W_BS;
 565         b1[i] += (W_AM*(b0[i] + b2[i])+W_AO)>>W_AS;
 566     }
 567
 568     asm volatile(
 569         "jmp 2f                                      \n\t"
 570         "1:                                          \n\t"
 571
 572         "mov %6, %%"REG_a"                           \n\t"
 573         "mov %4, %%"REG_S"                           \n\t"
 574
 575         snow_vertical_compose_mmx_load(REG_S,"mm0","mm2","mm4","mm6")
 576         snow_vertical_compose_mmx_add(REG_a,"mm0","mm2","mm4","mm6")
 577         snow_vertical_compose_mmx_move("mm0","mm2","mm4","mm6","mm1","mm3","mm5","mm7")
 578         snow_vertical_compose_mmx_sll("1","mm0","mm2","mm4","mm6")
 579         snow_vertical_compose_mmx_r2r_add("mm1","mm3","mm5","mm7","mm0","mm2","mm4","mm6")
 580
 581         "pcmpeqd %%mm1, %%mm1                        \n\t"
 582         "pslld $31, %%mm1                            \n\t"
 583         "psrld $29, %%mm1                            \n\t"
 584         "mov %5, %%"REG_a"                           \n\t"
 585
 586         snow_vertical_compose_mmx_r2r_add("mm1","mm1","mm1","mm1","mm0","mm2","mm4","mm6")
 587         snow_vertical_compose_mmx_sra("3","mm0","mm2","mm4","mm6")
 588         snow_vertical_compose_mmx_load(REG_a,"mm1","mm3","mm5","mm7")
 589         snow_vertical_compose_mmx_sub("mm0","mm2","mm4","mm6","mm1","mm3","mm5","mm7")
 590         snow_vertical_compose_mmx_store(REG_a,"mm1","mm3","mm5","mm7")
 591         "mov %3, %%"REG_c"                           \n\t"
 592         snow_vertical_compose_mmx_load(REG_S,"mm0","mm2","mm4","mm6")
 593         snow_vertical_compose_mmx_add(REG_c,"mm1","mm3","mm5","mm7")
 594         snow_vertical_compose_mmx_sub("mm1","mm3","mm5","mm7","mm0","mm2","mm4","mm6")
 595         snow_vertical_compose_mmx_store(REG_S,"mm0","mm2","mm4","mm6")
 596         "mov %2, %%"REG_a"                           \n\t"
 597         snow_vertical_compose_mmx_load(REG_c,"mm1","mm3","mm5","mm7")
 598         snow_vertical_compose_mmx_add(REG_a,"mm0","mm2","mm4","mm6")
 599         snow_vertical_compose_mmx_sll("2","mm1","mm3","mm5","mm7")
 600         snow_vertical_compose_mmx_r2r_add("mm1","mm3","mm5","mm7","mm0","mm2","mm4","mm6")
 601
 602         "pcmpeqd %%mm1, %%mm1                        \n\t"
 603         "pslld $31, %%mm1                            \n\t"
 604         "psrld $28, %%mm1                            \n\t"
 605         "mov %1, %%"REG_S"                           \n\t"
 606
 607         snow_vertical_compose_mmx_r2r_add("mm1","mm1","mm1","mm1","mm0","mm2","mm4","mm6")
 608         snow_vertical_compose_mmx_sra("4","mm0","mm2","mm4","mm6")
 609         snow_vertical_compose_mmx_add(REG_c,"mm0","mm2","mm4","mm6")
 610         snow_vertical_compose_mmx_store(REG_c,"mm0","mm2","mm4","mm6")
 611         snow_vertical_compose_mmx_add(REG_S,"mm0","mm2","mm4","mm6")
 612         snow_vertical_compose_mmx_move("mm0","mm2","mm4","mm6","mm1","mm3","mm5","mm7")
 613         snow_vertical_compose_mmx_sll("1","mm0","mm2","mm4","mm6")
 614         snow_vertical_compose_mmx_r2r_add("mm1","mm3","mm5","mm7","mm0","mm2","mm4","mm6")
 615         snow_vertical_compose_mmx_sra("1","mm0","mm2","mm4","mm6")
 616         snow_vertical_compose_mmx_add(REG_a,"mm0","mm2","mm4","mm6")
 617         snow_vertical_compose_mmx_store(REG_a,"mm0","mm2","mm4","mm6")
 618
 619         "2:                                          \n\t"
 620         "sub $8, %%"REG_d"                           \n\t"
 621         "jge 1b                                      \n\t"
 622         :"+d"(i)
 623         :
 624         "m"(b0),"m"(b1),"m"(b2),"m"(b3),"m"(b4),"m"(b5):
 625         "%"REG_a"","%"REG_S"","%"REG_c"");
 626 }
 627
 628 #define snow_inner_add_yblock_sse2_header \
 629     DWTELEM * * dst_array = sb->line + src_y;\
 630     long tmp;\
 631     asm volatile(\
 632              "mov  %7, %%"REG_c"             \n\t"\
 633              "mov  %6, %2                    \n\t"\
 634              "mov  %4, %%"REG_S"             \n\t"\
 635              "pxor %%xmm7, %%xmm7            \n\t" /* 0 */\
 636              "pcmpeqd %%xmm3, %%xmm3         \n\t"\
 637              "pslld $31, %%xmm3              \n\t"\
 638              "psrld $24, %%xmm3              \n\t" /* FRAC_BITS >> 1 */\
 639              "1:                             \n\t"\
 640              "mov %1, %%"REG_D"              \n\t"\
 641              "mov (%%"REG_D"), %%"REG_D"     \n\t"\
 642              "add %3, %%"REG_D"              \n\t"
 643
 644 #define snow_inner_add_yblock_sse2_start_8(out_reg1, out_reg2, ptr_offset, s_offset)\
 645              "mov "PTR_SIZE"*"ptr_offset"(%%"REG_a"), %%"REG_d"; \n\t"\
 646              "movq (%%"REG_d"), %%"out_reg1" \n\t"\
 647              "movq (%%"REG_d", %%"REG_c"), %%"out_reg2" \n\t"\
 648              "punpcklbw %%xmm7, %%"out_reg1" \n\t"\
 649              "punpcklbw %%xmm7, %%"out_reg2" \n\t"\
 650              "movq "s_offset"(%%"REG_S"), %%xmm0 \n\t"\
 651              "movq "s_offset"+16(%%"REG_S"), %%xmm4 \n\t"\
 652              "punpcklbw %%xmm7, %%xmm0       \n\t"\
 653              "punpcklbw %%xmm7, %%xmm4       \n\t"\
 654              "pmullw %%xmm0, %%"out_reg1"    \n\t"\
 655              "pmullw %%xmm4, %%"out_reg2"    \n\t"
 656
 657 #define snow_inner_add_yblock_sse2_start_16(out_reg1, out_reg2, ptr_offset, s_offset)\
 658              "mov "PTR_SIZE"*"ptr_offset"(%%"REG_a"), %%"REG_d"; \n\t"\
 659              "movq (%%"REG_d"), %%"out_reg1" \n\t"\
 660              "movq 8(%%"REG_d"), %%"out_reg2" \n\t"\
 661              "punpcklbw %%xmm7, %%"out_reg1" \n\t"\
 662              "punpcklbw %%xmm7, %%"out_reg2" \n\t"\
 663              "movq "s_offset"(%%"REG_S"), %%xmm0 \n\t"\
 664              "movq "s_offset"+8(%%"REG_S"), %%xmm4 \n\t"\
 665              "punpcklbw %%xmm7, %%xmm0       \n\t"\
 666              "punpcklbw %%xmm7, %%xmm4       \n\t"\
 667              "pmullw %%xmm0, %%"out_reg1"    \n\t"\
 668              "pmullw %%xmm4, %%"out_reg2"    \n\t"
 669
 670 #define snow_inner_add_yblock_sse2_accum_8(ptr_offset, s_offset) \
 671              snow_inner_add_yblock_sse2_start_8("xmm2", "xmm6", ptr_offset, s_offset)\
 672              "paddusw %%xmm2, %%xmm1         \n\t"\
 673              "paddusw %%xmm6, %%xmm5         \n\t"
 674
 675 #define snow_inner_add_yblock_sse2_accum_16(ptr_offset, s_offset) \
 676              snow_inner_add_yblock_sse2_start_16("xmm2", "xmm6", ptr_offset, s_offset)\
 677              "paddusw %%xmm2, %%xmm1         \n\t"\
 678              "paddusw %%xmm6, %%xmm5         \n\t"
 679
 680 #define snow_inner_add_yblock_sse2_end_common1\
 681              "add $32, %%"REG_S"             \n\t"\
 682              "add %%"REG_c", %0              \n\t"\
 683              "add %%"REG_c", "PTR_SIZE"*3(%%"REG_a");\n\t"\
 684              "add %%"REG_c", "PTR_SIZE"*2(%%"REG_a");\n\t"\
 685              "add %%"REG_c", "PTR_SIZE"*1(%%"REG_a");\n\t"\
 686              "add %%"REG_c", (%%"REG_a")     \n\t"
 687
 688 #define snow_inner_add_yblock_sse2_end_common2\
 689              "jnz 1b                         \n\t"\
 690              :"+m"(dst8),"+m"(dst_array),"=&r"(tmp)\
 691              :\
 692              "rm"((long)(src_x<<2)),"m"(obmc),"a"(block),"m"((long)b_h),"m"((long)src_stride):\
 693              "%"REG_c"","%"REG_S"","%"REG_D"","%"REG_d"");
 694
 695 #define snow_inner_add_yblock_sse2_end_8\
 696              "sal $1, %%"REG_c"              \n\t"\
 697              "add $"PTR_SIZE"*2, %1          \n\t"\
 698              snow_inner_add_yblock_sse2_end_common1\
 699              "sar $1, %%"REG_c"              \n\t"\
 700              "sub $2, %2                     \n\t"\
 701              snow_inner_add_yblock_sse2_end_common2
 702
 703 #define snow_inner_add_yblock_sse2_end_16\
 704              "add $"PTR_SIZE"*1, %1          \n\t"\
 705              snow_inner_add_yblock_sse2_end_common1\
 706              "dec %2                         \n\t"\
 707              snow_inner_add_yblock_sse2_end_common2
 708
 709 static void inner_add_yblock_bw_8_obmc_16_bh_even_sse2(uint8_t *obmc, const long obmc_stride, uint8_t * * block, int b_w, long b_h,
 710                       int src_x, int src_y, long src_stride, slice_buffer * sb, int add, uint8_t * dst8){
 711 snow_inner_add_yblock_sse2_header
 712 snow_inner_add_yblock_sse2_start_8("xmm1", "xmm5", "3", "0")
 713 snow_inner_add_yblock_sse2_accum_8("2", "8")
 714 snow_inner_add_yblock_sse2_accum_8("1", "128")
 715 snow_inner_add_yblock_sse2_accum_8("0", "136")
 716
 717              "mov %0, %%"REG_d"              \n\t"
 718              "movdqa (%%"REG_D"), %%xmm0     \n\t"
 719              "movdqa %%xmm1, %%xmm2          \n\t"
 720
 721              "punpckhwd %%xmm7, %%xmm1       \n\t"
 722              "punpcklwd %%xmm7, %%xmm2       \n\t"
 723              "paddd %%xmm2, %%xmm0           \n\t"
 724              "movdqa 16(%%"REG_D"), %%xmm2   \n\t"
 725              "paddd %%xmm1, %%xmm2           \n\t"
 726              "paddd %%xmm3, %%xmm0           \n\t"
 727              "paddd %%xmm3, %%xmm2           \n\t"
 728
 729              "mov %1, %%"REG_D"              \n\t"
 730              "mov "PTR_SIZE"(%%"REG_D"), %%"REG_D";\n\t"
 731              "add %3, %%"REG_D"              \n\t"
 732
 733              "movdqa (%%"REG_D"), %%xmm4     \n\t"
 734              "movdqa %%xmm5, %%xmm6          \n\t"
 735              "punpckhwd %%xmm7, %%xmm5       \n\t"
 736              "punpcklwd %%xmm7, %%xmm6       \n\t"
 737              "paddd %%xmm6, %%xmm4           \n\t"
 738              "movdqa 16(%%"REG_D"), %%xmm6   \n\t"
 739              "paddd %%xmm5, %%xmm6           \n\t"
 740              "paddd %%xmm3, %%xmm4           \n\t"
 741              "paddd %%xmm3, %%xmm6           \n\t"
 742
 743              "psrad $8, %%xmm0               \n\t" /* FRAC_BITS. */
 744              "psrad $8, %%xmm2               \n\t" /* FRAC_BITS. */
 745              "packssdw %%xmm2, %%xmm0        \n\t"
 746              "packuswb %%xmm7, %%xmm0        \n\t"
 747              "movq %%xmm0, (%%"REG_d")       \n\t"
 748
 749              "psrad $8, %%xmm4               \n\t" /* FRAC_BITS. */
 750              "psrad $8, %%xmm6               \n\t" /* FRAC_BITS. */
 751              "packssdw %%xmm6, %%xmm4        \n\t"
 752              "packuswb %%xmm7, %%xmm4        \n\t"
 753              "movq %%xmm4, (%%"REG_d",%%"REG_c");\n\t"
 754 snow_inner_add_yblock_sse2_end_8
 755 }
 756
 757 static void inner_add_yblock_bw_16_obmc_32_sse2(uint8_t *obmc, const long obmc_stride, uint8_t * * block, int b_w, long b_h,
 758                       int src_x, int src_y, long src_stride, slice_buffer * sb, int add, uint8_t * dst8){
 759 snow_inner_add_yblock_sse2_header
 760 snow_inner_add_yblock_sse2_start_16("xmm1", "xmm5", "3", "0")
 761 snow_inner_add_yblock_sse2_accum_16("2", "16")
 762 snow_inner_add_yblock_sse2_accum_16("1", "512")
 763 snow_inner_add_yblock_sse2_accum_16("0", "528")
 764
 765              "mov %0, %%"REG_d"              \n\t"
 766              "movdqa %%xmm1, %%xmm0          \n\t"
 767              "movdqa %%xmm5, %%xmm4          \n\t"
 768              "punpcklwd %%xmm7, %%xmm0       \n\t"
 769              "paddd (%%"REG_D"), %%xmm0      \n\t"
 770              "punpckhwd %%xmm7, %%xmm1       \n\t"
 771              "paddd 16(%%"REG_D"), %%xmm1    \n\t"
 772              "punpcklwd %%xmm7, %%xmm4       \n\t"
 773              "paddd 32(%%"REG_D"), %%xmm4    \n\t"
 774              "punpckhwd %%xmm7, %%xmm5       \n\t"
 775              "paddd 48(%%"REG_D"), %%xmm5    \n\t"
 776              "paddd %%xmm3, %%xmm0           \n\t"
 777              "paddd %%xmm3, %%xmm1           \n\t"
 778              "paddd %%xmm3, %%xmm4           \n\t"
 779              "paddd %%xmm3, %%xmm5           \n\t"
 780              "psrad $8, %%xmm0               \n\t" /* FRAC_BITS. */
 781              "psrad $8, %%xmm1               \n\t" /* FRAC_BITS. */
 782              "psrad $8, %%xmm4               \n\t" /* FRAC_BITS. */
 783              "psrad $8, %%xmm5               \n\t" /* FRAC_BITS. */
 784
 785              "packssdw %%xmm1, %%xmm0        \n\t"
 786              "packssdw %%xmm5, %%xmm4        \n\t"
 787              "packuswb %%xmm4, %%xmm0        \n\t"
 788
 789              "movdqu %%xmm0, (%%"REG_d")       \n\t"
 790
 791 snow_inner_add_yblock_sse2_end_16
 792 }
 793
 794 #define snow_inner_add_yblock_mmx_header \
 795     DWTELEM * * dst_array = sb->line + src_y;\
 796     long tmp;\
 797     asm volatile(\
 798              "mov  %7, %%"REG_c"             \n\t"\
 799              "mov  %6, %2                    \n\t"\
 800              "mov  %4, %%"REG_S"             \n\t"\
 801              "pxor %%mm7, %%mm7              \n\t" /* 0 */\
 802              "pcmpeqd %%mm3, %%mm3           \n\t"\
 803              "pslld $31, %%mm3               \n\t"\
 804              "psrld $24, %%mm3               \n\t" /* FRAC_BITS >> 1 */\
 805              "1:                             \n\t"\
 806              "mov %1, %%"REG_D"              \n\t"\
 807              "mov (%%"REG_D"), %%"REG_D"     \n\t"\
 808              "add %3, %%"REG_D"              \n\t"
 809
 810 #define snow_inner_add_yblock_mmx_start(out_reg1, out_reg2, ptr_offset, s_offset, d_offset)\
 811              "mov "PTR_SIZE"*"ptr_offset"(%%"REG_a"), %%"REG_d"; \n\t"\
 812              "movd "d_offset"(%%"REG_d"), %%"out_reg1" \n\t"\
 813              "movd "d_offset"+4(%%"REG_d"), %%"out_reg2" \n\t"\
 814              "punpcklbw %%mm7, %%"out_reg1" \n\t"\
 815              "punpcklbw %%mm7, %%"out_reg2" \n\t"\
 816              "movd "s_offset"(%%"REG_S"), %%mm0 \n\t"\
 817              "movd "s_offset"+4(%%"REG_S"), %%mm4 \n\t"\
 818              "punpcklbw %%mm7, %%mm0       \n\t"\
 819              "punpcklbw %%mm7, %%mm4       \n\t"\
 820              "pmullw %%mm0, %%"out_reg1"    \n\t"\
 821              "pmullw %%mm4, %%"out_reg2"    \n\t"
 822
 823 #define snow_inner_add_yblock_mmx_accum(ptr_offset, s_offset, d_offset) \
 824              snow_inner_add_yblock_mmx_start("mm2", "mm6", ptr_offset, s_offset, d_offset)\
 825              "paddusw %%mm2, %%mm1         \n\t"\
 826              "paddusw %%mm6, %%mm5         \n\t"
 827
 828 #define snow_inner_add_yblock_mmx_mix(read_offset, write_offset)\
 829              "mov %0, %%"REG_d"              \n\t"\
 830              "movq %%mm1, %%mm0              \n\t"\
 831              "movq %%mm5, %%mm4              \n\t"\
 832              "punpcklwd %%mm7, %%mm0         \n\t"\
 833              "paddd "read_offset"(%%"REG_D"), %%mm0 \n\t"\
 834              "punpckhwd %%mm7, %%mm1         \n\t"\
 835              "paddd "read_offset"+8(%%"REG_D"), %%mm1 \n\t"\
 836              "punpcklwd %%mm7, %%mm4         \n\t"\
 837              "paddd "read_offset"+16(%%"REG_D"), %%mm4 \n\t"\
 838              "punpckhwd %%mm7, %%mm5         \n\t"\
 839              "paddd "read_offset"+24(%%"REG_D"), %%mm5 \n\t"\
 840              "paddd %%mm3, %%mm0             \n\t"\
 841              "paddd %%mm3, %%mm1             \n\t"\
 842              "paddd %%mm3, %%mm4             \n\t"\
 843              "paddd %%mm3, %%mm5             \n\t"\
 844              "psrad $8, %%mm0                \n\t"\
 845              "psrad $8, %%mm1                \n\t"\
 846              "psrad $8, %%mm4                \n\t"\
 847              "psrad $8, %%mm5                \n\t"\
 848 \
 849              "packssdw %%mm1, %%mm0          \n\t"\
 850              "packssdw %%mm5, %%mm4          \n\t"\
 851              "packuswb %%mm4, %%mm0          \n\t"\
 852              "movq %%mm0, "write_offset"(%%"REG_d") \n\t"
 853
 854 #define snow_inner_add_yblock_mmx_end(s_step)\
 855              "add $"s_step", %%"REG_S"             \n\t"\
 856              "add %%"REG_c", "PTR_SIZE"*3(%%"REG_a");\n\t"\
 857              "add %%"REG_c", "PTR_SIZE"*2(%%"REG_a");\n\t"\
 858              "add %%"REG_c", "PTR_SIZE"*1(%%"REG_a");\n\t"\
 859              "add %%"REG_c", (%%"REG_a")     \n\t"\
 860              "add $"PTR_SIZE"*1, %1          \n\t"\
 861              "add %%"REG_c", %0              \n\t"\
 862              "dec %2                         \n\t"\
 863              "jnz 1b                         \n\t"\
 864              :"+m"(dst8),"+m"(dst_array),"=&r"(tmp)\
 865              :\
 866              "rm"((long)(src_x<<2)),"m"(obmc),"a"(block),"m"((long)b_h),"m"((long)src_stride):\
 867              "%"REG_c"","%"REG_S"","%"REG_D"","%"REG_d"");
 868
 869 static void inner_add_yblock_bw_8_obmc_16_mmx(uint8_t *obmc, const long obmc_stride, uint8_t * * block, int b_w, long b_h,
 870                       int src_x, int src_y, long src_stride, slice_buffer * sb, int add, uint8_t * dst8){
 871 snow_inner_add_yblock_mmx_header
 872 snow_inner_add_yblock_mmx_start("mm1", "mm5", "3", "0", "0")
 873 snow_inner_add_yblock_mmx_accum("2", "8", "0")
 874 snow_inner_add_yblock_mmx_accum("1", "128", "0")
 875 snow_inner_add_yblock_mmx_accum("0", "136", "0")
 876 snow_inner_add_yblock_mmx_mix("0", "0")
 877 snow_inner_add_yblock_mmx_end("16")
 878 }
 879
 880 static void inner_add_yblock_bw_16_obmc_32_mmx(uint8_t *obmc, const long obmc_stride, uint8_t * * block, int b_w, long b_h,
 881                       int src_x, int src_y, long src_stride, slice_buffer * sb, int add, uint8_t * dst8){
 882 snow_inner_add_yblock_mmx_header
 883 snow_inner_add_yblock_mmx_start("mm1", "mm5", "3", "0", "0")
 884 snow_inner_add_yblock_mmx_accum("2", "16", "0")
 885 snow_inner_add_yblock_mmx_accum("1", "512", "0")
 886 snow_inner_add_yblock_mmx_accum("0", "528", "0")
 887 snow_inner_add_yblock_mmx_mix("0", "0")
 888
 889 snow_inner_add_yblock_mmx_start("mm1", "mm5", "3", "8", "8")
 890 snow_inner_add_yblock_mmx_accum("2", "24", "8")
 891 snow_inner_add_yblock_mmx_accum("1", "520", "8")
 892 snow_inner_add_yblock_mmx_accum("0", "536", "8")
 893 snow_inner_add_yblock_mmx_mix("32", "8")
 894 snow_inner_add_yblock_mmx_end("32")
 895 }
 896
 897 void ff_snow_inner_add_yblock_sse2(uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
 898                            int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8){
 899
 900     if (b_w == 16)
 901         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);
 902     else if (b_w == 8 && obmc_stride == 16) {
 903         if (!(b_h & 1))
 904             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);
 905         else
 906             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);
 907     } else
 908          ff_snow_inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
 909 }
 910
 911 void ff_snow_inner_add_yblock_mmx(uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
 912                           int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8){
 913     if (b_w == 16)
 914         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);
 915     else if (b_w == 8 && obmc_stride == 16)
 916         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);
 917     else
 918         ff_snow_inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
 919 }