git.sesse.net Git - ffmpeg/blob - libavcodec/x86/snowdsp.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/cpu.h"
  23 #include "libavutil/x86/asm.h"
  24 #include "libavcodec/avcodec.h"
  25 #include "libavcodec/snow.h"
  26 #include "libavcodec/snow_dwt.h"
  27
  28 #if HAVE_INLINE_ASM
  29
  30 static void ff_snow_horizontal_compose97i_sse2(IDWTELEM *b, IDWTELEM *temp, int width){
  31     const int w2= (width+1)>>1;
  32     const int w_l= (width>>1);
  33     const int w_r= w2 - 1;
  34     int i;
  35
  36     { // Lift 0
  37         IDWTELEM * const ref = b + w2 - 1;
  38         IDWTELEM b_0 = b[0]; //By allowing the first entry in b[0] to be calculated twice
  39         // (the first time erroneously), we allow the SSE2 code to run an extra pass.
  40         // The savings in code and time are well worth having to store this value and
  41         // calculate b[0] correctly afterwards.
  42
  43         i = 0;
  44         __asm__ volatile(
  45             "pcmpeqd   %%xmm7, %%xmm7         \n\t"
  46             "pcmpeqd   %%xmm3, %%xmm3         \n\t"
  47             "psllw         $1, %%xmm3         \n\t"
  48             "paddw     %%xmm7, %%xmm3         \n\t"
  49             "psllw        $13, %%xmm3         \n\t"
  50         ::);
  51         for(; i<w_l-15; i+=16){
  52             __asm__ volatile(
  53                 "movdqu   (%1), %%xmm1        \n\t"
  54                 "movdqu 16(%1), %%xmm5        \n\t"
  55                 "movdqu  2(%1), %%xmm2        \n\t"
  56                 "movdqu 18(%1), %%xmm6        \n\t"
  57                 "paddw  %%xmm1, %%xmm2        \n\t"
  58                 "paddw  %%xmm5, %%xmm6        \n\t"
  59                 "paddw  %%xmm7, %%xmm2        \n\t"
  60                 "paddw  %%xmm7, %%xmm6        \n\t"
  61                 "pmulhw %%xmm3, %%xmm2        \n\t"
  62                 "pmulhw %%xmm3, %%xmm6        \n\t"
  63                 "paddw    (%0), %%xmm2        \n\t"
  64                 "paddw  16(%0), %%xmm6        \n\t"
  65                 "movdqa %%xmm2, (%0)          \n\t"
  66                 "movdqa %%xmm6, 16(%0)        \n\t"
  67                 :: "r"(&b[i]), "r"(&ref[i])
  68                 : "memory"
  69             );
  70         }
  71         snow_horizontal_compose_lift_lead_out(i, b, b, ref, width, w_l, 0, W_DM, W_DO, W_DS);
  72         b[0] = b_0 - ((W_DM * 2 * ref[1]+W_DO)>>W_DS);
  73     }
  74
  75     { // Lift 1
  76         IDWTELEM * const dst = b+w2;
  77
  78         i = 0;
  79         for(; (((x86_reg)&dst[i]) & 0x1F) && i<w_r; i++){
  80             dst[i] = dst[i] - (b[i] + b[i + 1]);
  81         }
  82         for(; i<w_r-15; i+=16){
  83             __asm__ volatile(
  84                 "movdqu   (%1), %%xmm1        \n\t"
  85                 "movdqu 16(%1), %%xmm5        \n\t"
  86                 "movdqu  2(%1), %%xmm2        \n\t"
  87                 "movdqu 18(%1), %%xmm6        \n\t"
  88                 "paddw  %%xmm1, %%xmm2        \n\t"
  89                 "paddw  %%xmm5, %%xmm6        \n\t"
  90                 "movdqa   (%0), %%xmm0        \n\t"
  91                 "movdqa 16(%0), %%xmm4        \n\t"
  92                 "psubw  %%xmm2, %%xmm0        \n\t"
  93                 "psubw  %%xmm6, %%xmm4        \n\t"
  94                 "movdqa %%xmm0, (%0)          \n\t"
  95                 "movdqa %%xmm4, 16(%0)        \n\t"
  96                 :: "r"(&dst[i]), "r"(&b[i])
  97                 : "memory"
  98             );
  99         }
 100         snow_horizontal_compose_lift_lead_out(i, dst, dst, b, width, w_r, 1, W_CM, W_CO, W_CS);
 101     }
 102
 103     { // Lift 2
 104         IDWTELEM * const ref = b+w2 - 1;
 105         IDWTELEM b_0 = b[0];
 106
 107         i = 0;
 108         __asm__ volatile(
 109             "psllw         $15, %%xmm7        \n\t"
 110             "pcmpeqw    %%xmm6, %%xmm6        \n\t"
 111             "psrlw         $13, %%xmm6        \n\t"
 112             "paddw      %%xmm7, %%xmm6        \n\t"
 113         ::);
 114         for(; i<w_l-15; i+=16){
 115             __asm__ volatile(
 116                 "movdqu   (%1), %%xmm0        \n\t"
 117                 "movdqu 16(%1), %%xmm4        \n\t"
 118                 "movdqu  2(%1), %%xmm1        \n\t"
 119                 "movdqu 18(%1), %%xmm5        \n\t" //FIXME try aligned reads and shifts
 120                 "paddw  %%xmm6, %%xmm0        \n\t"
 121                 "paddw  %%xmm6, %%xmm4        \n\t"
 122                 "paddw  %%xmm7, %%xmm1        \n\t"
 123                 "paddw  %%xmm7, %%xmm5        \n\t"
 124                 "pavgw  %%xmm1, %%xmm0        \n\t"
 125                 "pavgw  %%xmm5, %%xmm4        \n\t"
 126                 "psubw  %%xmm7, %%xmm0        \n\t"
 127                 "psubw  %%xmm7, %%xmm4        \n\t"
 128                 "psraw      $1, %%xmm0        \n\t"
 129                 "psraw      $1, %%xmm4        \n\t"
 130                 "movdqa   (%0), %%xmm1        \n\t"
 131                 "movdqa 16(%0), %%xmm5        \n\t"
 132                 "paddw  %%xmm1, %%xmm0        \n\t"
 133                 "paddw  %%xmm5, %%xmm4        \n\t"
 134                 "psraw      $2, %%xmm0        \n\t"
 135                 "psraw      $2, %%xmm4        \n\t"
 136                 "paddw  %%xmm1, %%xmm0        \n\t"
 137                 "paddw  %%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-1 + 4 * b_0) >> W_BS);
 146     }
 147
 148     { // Lift 3
 149         IDWTELEM * const src = b+w2;
 150
 151         i = 0;
 152         for(; (((x86_reg)&temp[i]) & 0x1F) && 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  2(%1), %%xmm2        \n\t"
 158                 "movdqu 18(%1), %%xmm6        \n\t"
 159                 "paddw    (%1), %%xmm2        \n\t"
 160                 "paddw  16(%1), %%xmm6        \n\t"
 161                 "movdqu   (%0), %%xmm0        \n\t"
 162                 "movdqu 16(%0), %%xmm4        \n\t"
 163                 "paddw  %%xmm2, %%xmm0        \n\t"
 164                 "paddw  %%xmm6, %%xmm4        \n\t"
 165                 "psraw      $1, %%xmm2        \n\t"
 166                 "psraw      $1, %%xmm6        \n\t"
 167                 "paddw  %%xmm0, %%xmm2        \n\t"
 168                 "paddw  %%xmm4, %%xmm6        \n\t"
 169                 "movdqa %%xmm2, (%2)          \n\t"
 170                 "movdqa %%xmm6, 16(%2)        \n\t"
 171                 :: "r"(&src[i]), "r"(&b[i]), "r"(&temp[i])
 172                  : "memory"
 173                );
 174         }
 175         snow_horizontal_compose_lift_lead_out(i, temp, src, b, width, w_r, 1, -W_AM, W_AO+1, W_AS);
 176     }
 177
 178     {
 179         snow_interleave_line_header(&i, width, b, temp);
 180
 181         for (; (i & 0x3E) != 0x3E; i-=2){
 182             b[i+1] = temp[i>>1];
 183             b[i] = b[i>>1];
 184         }
 185         for (i-=62; i>=0; i-=64){
 186             __asm__ volatile(
 187                 "movdqa      (%1), %%xmm0       \n\t"
 188                 "movdqa    16(%1), %%xmm2       \n\t"
 189                 "movdqa    32(%1), %%xmm4       \n\t"
 190                 "movdqa    48(%1), %%xmm6       \n\t"
 191                 "movdqa      (%1), %%xmm1       \n\t"
 192                 "movdqa    16(%1), %%xmm3       \n\t"
 193                 "movdqa    32(%1), %%xmm5       \n\t"
 194                 "movdqa    48(%1), %%xmm7       \n\t"
 195                 "punpcklwd   (%2), %%xmm0       \n\t"
 196                 "punpcklwd 16(%2), %%xmm2       \n\t"
 197                 "punpcklwd 32(%2), %%xmm4       \n\t"
 198                 "punpcklwd 48(%2), %%xmm6       \n\t"
 199                 "movdqa    %%xmm0, (%0)         \n\t"
 200                 "movdqa    %%xmm2, 32(%0)       \n\t"
 201                 "movdqa    %%xmm4, 64(%0)       \n\t"
 202                 "movdqa    %%xmm6, 96(%0)       \n\t"
 203                 "punpckhwd   (%2), %%xmm1       \n\t"
 204                 "punpckhwd 16(%2), %%xmm3       \n\t"
 205                 "punpckhwd 32(%2), %%xmm5       \n\t"
 206                 "punpckhwd 48(%2), %%xmm7       \n\t"
 207                 "movdqa    %%xmm1, 16(%0)       \n\t"
 208                 "movdqa    %%xmm3, 48(%0)       \n\t"
 209                 "movdqa    %%xmm5, 80(%0)       \n\t"
 210                 "movdqa    %%xmm7, 112(%0)      \n\t"
 211                 :: "r"(&(b)[i]), "r"(&(b)[i>>1]), "r"(&(temp)[i>>1])
 212                  : "memory"
 213                );
 214         }
 215     }
 216 }
 217
 218 static void ff_snow_horizontal_compose97i_mmx(IDWTELEM *b, IDWTELEM *temp, int width){
 219     const int w2= (width+1)>>1;
 220     const int w_l= (width>>1);
 221     const int w_r= w2 - 1;
 222     int i;
 223
 224     { // Lift 0
 225         IDWTELEM * const ref = b + w2 - 1;
 226
 227         i = 1;
 228         b[0] = b[0] - ((W_DM * 2 * ref[1]+W_DO)>>W_DS);
 229         __asm__ volatile(
 230             "pcmpeqw    %%mm7, %%mm7         \n\t"
 231             "pcmpeqw    %%mm3, %%mm3         \n\t"
 232             "psllw         $1, %%mm3         \n\t"
 233             "paddw      %%mm7, %%mm3         \n\t"
 234             "psllw        $13, %%mm3         \n\t"
 235            ::);
 236         for(; i<w_l-7; i+=8){
 237             __asm__ volatile(
 238                 "movq     (%1), %%mm2        \n\t"
 239                 "movq    8(%1), %%mm6        \n\t"
 240                 "paddw   2(%1), %%mm2        \n\t"
 241                 "paddw  10(%1), %%mm6        \n\t"
 242                 "paddw   %%mm7, %%mm2        \n\t"
 243                 "paddw   %%mm7, %%mm6        \n\t"
 244                 "pmulhw  %%mm3, %%mm2        \n\t"
 245                 "pmulhw  %%mm3, %%mm6        \n\t"
 246                 "paddw    (%0), %%mm2        \n\t"
 247                 "paddw   8(%0), %%mm6        \n\t"
 248                 "movq    %%mm2, (%0)         \n\t"
 249                 "movq    %%mm6, 8(%0)        \n\t"
 250                 :: "r"(&b[i]), "r"(&ref[i])
 251                  : "memory"
 252                );
 253         }
 254         snow_horizontal_compose_lift_lead_out(i, b, b, ref, width, w_l, 0, W_DM, W_DO, W_DS);
 255     }
 256
 257     { // Lift 1
 258         IDWTELEM * const dst = b+w2;
 259
 260         i = 0;
 261         for(; i<w_r-7; i+=8){
 262             __asm__ volatile(
 263                 "movq     (%1), %%mm2        \n\t"
 264                 "movq    8(%1), %%mm6        \n\t"
 265                 "paddw   2(%1), %%mm2        \n\t"
 266                 "paddw  10(%1), %%mm6        \n\t"
 267                 "movq     (%0), %%mm0        \n\t"
 268                 "movq    8(%0), %%mm4        \n\t"
 269                 "psubw   %%mm2, %%mm0        \n\t"
 270                 "psubw   %%mm6, %%mm4        \n\t"
 271                 "movq    %%mm0, (%0)         \n\t"
 272                 "movq    %%mm4, 8(%0)        \n\t"
 273                 :: "r"(&dst[i]), "r"(&b[i])
 274                  : "memory"
 275                );
 276         }
 277         snow_horizontal_compose_lift_lead_out(i, dst, dst, b, width, w_r, 1, W_CM, W_CO, W_CS);
 278     }
 279
 280     { // Lift 2
 281         IDWTELEM * const ref = b+w2 - 1;
 282
 283         i = 1;
 284         b[0] = b[0] + (((2 * ref[1] + W_BO) + 4 * b[0]) >> W_BS);
 285         __asm__ volatile(
 286             "psllw         $15, %%mm7        \n\t"
 287             "pcmpeqw     %%mm6, %%mm6        \n\t"
 288             "psrlw         $13, %%mm6        \n\t"
 289             "paddw       %%mm7, %%mm6        \n\t"
 290            ::);
 291         for(; i<w_l-7; i+=8){
 292             __asm__ volatile(
 293                 "movq     (%1), %%mm0        \n\t"
 294                 "movq    8(%1), %%mm4        \n\t"
 295                 "movq    2(%1), %%mm1        \n\t"
 296                 "movq   10(%1), %%mm5        \n\t"
 297                 "paddw   %%mm6, %%mm0        \n\t"
 298                 "paddw   %%mm6, %%mm4        \n\t"
 299                 "paddw   %%mm7, %%mm1        \n\t"
 300                 "paddw   %%mm7, %%mm5        \n\t"
 301                 "pavgw   %%mm1, %%mm0        \n\t"
 302                 "pavgw   %%mm5, %%mm4        \n\t"
 303                 "psubw   %%mm7, %%mm0        \n\t"
 304                 "psubw   %%mm7, %%mm4        \n\t"
 305                 "psraw      $1, %%mm0        \n\t"
 306                 "psraw      $1, %%mm4        \n\t"
 307                 "movq     (%0), %%mm1        \n\t"
 308                 "movq    8(%0), %%mm5        \n\t"
 309                 "paddw   %%mm1, %%mm0        \n\t"
 310                 "paddw   %%mm5, %%mm4        \n\t"
 311                 "psraw      $2, %%mm0        \n\t"
 312                 "psraw      $2, %%mm4        \n\t"
 313                 "paddw   %%mm1, %%mm0        \n\t"
 314                 "paddw   %%mm5, %%mm4        \n\t"
 315                 "movq    %%mm0, (%0)         \n\t"
 316                 "movq    %%mm4, 8(%0)        \n\t"
 317                 :: "r"(&b[i]), "r"(&ref[i])
 318                  : "memory"
 319                );
 320         }
 321         snow_horizontal_compose_liftS_lead_out(i, b, b, ref, width, w_l);
 322     }
 323
 324     { // Lift 3
 325         IDWTELEM * const src = b+w2;
 326         i = 0;
 327
 328         for(; i<w_r-7; i+=8){
 329             __asm__ volatile(
 330                 "movq    2(%1), %%mm2        \n\t"
 331                 "movq   10(%1), %%mm6        \n\t"
 332                 "paddw    (%1), %%mm2        \n\t"
 333                 "paddw   8(%1), %%mm6        \n\t"
 334                 "movq     (%0), %%mm0        \n\t"
 335                 "movq    8(%0), %%mm4        \n\t"
 336                 "paddw   %%mm2, %%mm0        \n\t"
 337                 "paddw   %%mm6, %%mm4        \n\t"
 338                 "psraw      $1, %%mm2        \n\t"
 339                 "psraw      $1, %%mm6        \n\t"
 340                 "paddw   %%mm0, %%mm2        \n\t"
 341                 "paddw   %%mm4, %%mm6        \n\t"
 342                 "movq    %%mm2, (%2)         \n\t"
 343                 "movq    %%mm6, 8(%2)        \n\t"
 344                 :: "r"(&src[i]), "r"(&b[i]), "r"(&temp[i])
 345                  : "memory"
 346                );
 347         }
 348         snow_horizontal_compose_lift_lead_out(i, temp, src, b, width, w_r, 1, -W_AM, W_AO+1, W_AS);
 349     }
 350
 351     {
 352         snow_interleave_line_header(&i, width, b, temp);
 353
 354         for (; (i & 0x1E) != 0x1E; i-=2){
 355             b[i+1] = temp[i>>1];
 356             b[i] = b[i>>1];
 357         }
 358         for (i-=30; i>=0; i-=32){
 359             __asm__ volatile(
 360                 "movq        (%1), %%mm0       \n\t"
 361                 "movq       8(%1), %%mm2       \n\t"
 362                 "movq      16(%1), %%mm4       \n\t"
 363                 "movq      24(%1), %%mm6       \n\t"
 364                 "movq        (%1), %%mm1       \n\t"
 365                 "movq       8(%1), %%mm3       \n\t"
 366                 "movq      16(%1), %%mm5       \n\t"
 367                 "movq      24(%1), %%mm7       \n\t"
 368                 "punpcklwd   (%2), %%mm0       \n\t"
 369                 "punpcklwd  8(%2), %%mm2       \n\t"
 370                 "punpcklwd 16(%2), %%mm4       \n\t"
 371                 "punpcklwd 24(%2), %%mm6       \n\t"
 372                 "movq       %%mm0, (%0)        \n\t"
 373                 "movq       %%mm2, 16(%0)      \n\t"
 374                 "movq       %%mm4, 32(%0)      \n\t"
 375                 "movq       %%mm6, 48(%0)      \n\t"
 376                 "punpckhwd   (%2), %%mm1       \n\t"
 377                 "punpckhwd  8(%2), %%mm3       \n\t"
 378                 "punpckhwd 16(%2), %%mm5       \n\t"
 379                 "punpckhwd 24(%2), %%mm7       \n\t"
 380                 "movq       %%mm1, 8(%0)       \n\t"
 381                 "movq       %%mm3, 24(%0)      \n\t"
 382                 "movq       %%mm5, 40(%0)      \n\t"
 383                 "movq       %%mm7, 56(%0)      \n\t"
 384                 :: "r"(&b[i]), "r"(&b[i>>1]), "r"(&temp[i>>1])
 385                  : "memory"
 386                );
 387         }
 388     }
 389 }
 390
 391 #if HAVE_7REGS
 392 #define snow_vertical_compose_sse2_load_add(op,r,t0,t1,t2,t3)\
 393         ""op" ("r",%%"FF_REG_d"), %%"t0"      \n\t"\
 394         ""op" 16("r",%%"FF_REG_d"), %%"t1"    \n\t"\
 395         ""op" 32("r",%%"FF_REG_d"), %%"t2"    \n\t"\
 396         ""op" 48("r",%%"FF_REG_d"), %%"t3"    \n\t"
 397
 398 #define snow_vertical_compose_sse2_load(r,t0,t1,t2,t3)\
 399         snow_vertical_compose_sse2_load_add("movdqa",r,t0,t1,t2,t3)
 400
 401 #define snow_vertical_compose_sse2_add(r,t0,t1,t2,t3)\
 402         snow_vertical_compose_sse2_load_add("paddw",r,t0,t1,t2,t3)
 403
 404 #define snow_vertical_compose_r2r_sub(s0,s1,s2,s3,t0,t1,t2,t3)\
 405         "psubw %%"s0", %%"t0" \n\t"\
 406         "psubw %%"s1", %%"t1" \n\t"\
 407         "psubw %%"s2", %%"t2" \n\t"\
 408         "psubw %%"s3", %%"t3" \n\t"
 409
 410 #define snow_vertical_compose_sse2_store(w,s0,s1,s2,s3)\
 411         "movdqa %%"s0", ("w",%%"FF_REG_d")    \n\t"\
 412         "movdqa %%"s1", 16("w",%%"FF_REG_d")  \n\t"\
 413         "movdqa %%"s2", 32("w",%%"FF_REG_d")  \n\t"\
 414         "movdqa %%"s3", 48("w",%%"FF_REG_d")  \n\t"
 415
 416 #define snow_vertical_compose_sra(n,t0,t1,t2,t3)\
 417         "psraw $"n", %%"t0" \n\t"\
 418         "psraw $"n", %%"t1" \n\t"\
 419         "psraw $"n", %%"t2" \n\t"\
 420         "psraw $"n", %%"t3" \n\t"
 421
 422 #define snow_vertical_compose_r2r_add(s0,s1,s2,s3,t0,t1,t2,t3)\
 423         "paddw %%"s0", %%"t0" \n\t"\
 424         "paddw %%"s1", %%"t1" \n\t"\
 425         "paddw %%"s2", %%"t2" \n\t"\
 426         "paddw %%"s3", %%"t3" \n\t"
 427
 428 #define snow_vertical_compose_r2r_pmulhw(s0,s1,s2,s3,t0,t1,t2,t3)\
 429         "pmulhw %%"s0", %%"t0" \n\t"\
 430         "pmulhw %%"s1", %%"t1" \n\t"\
 431         "pmulhw %%"s2", %%"t2" \n\t"\
 432         "pmulhw %%"s3", %%"t3" \n\t"
 433
 434 #define snow_vertical_compose_sse2_move(s0,s1,s2,s3,t0,t1,t2,t3)\
 435         "movdqa %%"s0", %%"t0" \n\t"\
 436         "movdqa %%"s1", %%"t1" \n\t"\
 437         "movdqa %%"s2", %%"t2" \n\t"\
 438         "movdqa %%"s3", %%"t3" \n\t"
 439
 440 static void ff_snow_vertical_compose97i_sse2(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, IDWTELEM *b5, int width){
 441     x86_reg i = width;
 442
 443     while(i & 0x1F)
 444     {
 445         i--;
 446         b4[i] -= (W_DM*(b3[i] + b5[i])+W_DO)>>W_DS;
 447         b3[i] -= (W_CM*(b2[i] + b4[i])+W_CO)>>W_CS;
 448         b2[i] += (W_BM*(b1[i] + b3[i])+4*b2[i]+W_BO)>>W_BS;
 449         b1[i] += (W_AM*(b0[i] + b2[i])+W_AO)>>W_AS;
 450     }
 451     i+=i;
 452
 453          __asm__ volatile (
 454         "jmp 2f                                      \n\t"
 455         "1:                                          \n\t"
 456         snow_vertical_compose_sse2_load("%4","xmm0","xmm2","xmm4","xmm6")
 457         snow_vertical_compose_sse2_add("%6","xmm0","xmm2","xmm4","xmm6")
 458
 459
 460         "pcmpeqw    %%xmm0, %%xmm0                   \n\t"
 461         "pcmpeqw    %%xmm2, %%xmm2                   \n\t"
 462         "paddw      %%xmm2, %%xmm2                   \n\t"
 463         "paddw      %%xmm0, %%xmm2                   \n\t"
 464         "psllw         $13, %%xmm2                   \n\t"
 465         snow_vertical_compose_r2r_add("xmm0","xmm0","xmm0","xmm0","xmm1","xmm3","xmm5","xmm7")
 466         snow_vertical_compose_r2r_pmulhw("xmm2","xmm2","xmm2","xmm2","xmm1","xmm3","xmm5","xmm7")
 467         snow_vertical_compose_sse2_add("%5","xmm1","xmm3","xmm5","xmm7")
 468         snow_vertical_compose_sse2_store("%5","xmm1","xmm3","xmm5","xmm7")
 469         snow_vertical_compose_sse2_load("%4","xmm0","xmm2","xmm4","xmm6")
 470         snow_vertical_compose_sse2_add("%3","xmm1","xmm3","xmm5","xmm7")
 471         snow_vertical_compose_r2r_sub("xmm1","xmm3","xmm5","xmm7","xmm0","xmm2","xmm4","xmm6")
 472         snow_vertical_compose_sse2_store("%4","xmm0","xmm2","xmm4","xmm6")
 473
 474         "pcmpeqw %%xmm7, %%xmm7                      \n\t"
 475         "pcmpeqw %%xmm5, %%xmm5                      \n\t"
 476         "psllw $15, %%xmm7                           \n\t"
 477         "psrlw $13, %%xmm5                           \n\t"
 478         "paddw %%xmm7, %%xmm5                        \n\t"
 479         snow_vertical_compose_r2r_add("xmm5","xmm5","xmm5","xmm5","xmm0","xmm2","xmm4","xmm6")
 480         "movq   (%2,%%"FF_REG_d"), %%xmm1            \n\t"
 481         "movq  8(%2,%%"FF_REG_d"), %%xmm3            \n\t"
 482         "paddw %%xmm7, %%xmm1                        \n\t"
 483         "paddw %%xmm7, %%xmm3                        \n\t"
 484         "pavgw %%xmm1, %%xmm0                        \n\t"
 485         "pavgw %%xmm3, %%xmm2                        \n\t"
 486         "movq 16(%2,%%"FF_REG_d"), %%xmm1            \n\t"
 487         "movq 24(%2,%%"FF_REG_d"), %%xmm3            \n\t"
 488         "paddw %%xmm7, %%xmm1                        \n\t"
 489         "paddw %%xmm7, %%xmm3                        \n\t"
 490         "pavgw %%xmm1, %%xmm4                        \n\t"
 491         "pavgw %%xmm3, %%xmm6                        \n\t"
 492         snow_vertical_compose_r2r_sub("xmm7","xmm7","xmm7","xmm7","xmm0","xmm2","xmm4","xmm6")
 493         snow_vertical_compose_sra("1","xmm0","xmm2","xmm4","xmm6")
 494         snow_vertical_compose_sse2_add("%3","xmm0","xmm2","xmm4","xmm6")
 495
 496         snow_vertical_compose_sra("2","xmm0","xmm2","xmm4","xmm6")
 497         snow_vertical_compose_sse2_add("%3","xmm0","xmm2","xmm4","xmm6")
 498         snow_vertical_compose_sse2_store("%3","xmm0","xmm2","xmm4","xmm6")
 499         snow_vertical_compose_sse2_add("%1","xmm0","xmm2","xmm4","xmm6")
 500         snow_vertical_compose_sse2_move("xmm0","xmm2","xmm4","xmm6","xmm1","xmm3","xmm5","xmm7")
 501         snow_vertical_compose_sra("1","xmm0","xmm2","xmm4","xmm6")
 502         snow_vertical_compose_r2r_add("xmm1","xmm3","xmm5","xmm7","xmm0","xmm2","xmm4","xmm6")
 503         snow_vertical_compose_sse2_add("%2","xmm0","xmm2","xmm4","xmm6")
 504         snow_vertical_compose_sse2_store("%2","xmm0","xmm2","xmm4","xmm6")
 505
 506         "2:                                          \n\t"
 507         "sub $64, %%"FF_REG_d"                       \n\t"
 508         "jge 1b                                      \n\t"
 509         :"+d"(i)
 510         :"r"(b0),"r"(b1),"r"(b2),"r"(b3),"r"(b4),"r"(b5));
 511 }
 512
 513 #define snow_vertical_compose_mmx_load_add(op,r,t0,t1,t2,t3)\
 514         ""op" ("r",%%"FF_REG_d"), %%"t0"   \n\t"\
 515         ""op" 8("r",%%"FF_REG_d"), %%"t1"  \n\t"\
 516         ""op" 16("r",%%"FF_REG_d"), %%"t2" \n\t"\
 517         ""op" 24("r",%%"FF_REG_d"), %%"t3" \n\t"
 518
 519 #define snow_vertical_compose_mmx_load(r,t0,t1,t2,t3)\
 520         snow_vertical_compose_mmx_load_add("movq",r,t0,t1,t2,t3)
 521
 522 #define snow_vertical_compose_mmx_add(r,t0,t1,t2,t3)\
 523         snow_vertical_compose_mmx_load_add("paddw",r,t0,t1,t2,t3)
 524
 525 #define snow_vertical_compose_mmx_store(w,s0,s1,s2,s3)\
 526         "movq %%"s0", ("w",%%"FF_REG_d")   \n\t"\
 527         "movq %%"s1", 8("w",%%"FF_REG_d")  \n\t"\
 528         "movq %%"s2", 16("w",%%"FF_REG_d") \n\t"\
 529         "movq %%"s3", 24("w",%%"FF_REG_d") \n\t"
 530
 531 #define snow_vertical_compose_mmx_move(s0,s1,s2,s3,t0,t1,t2,t3)\
 532         "movq %%"s0", %%"t0" \n\t"\
 533         "movq %%"s1", %%"t1" \n\t"\
 534         "movq %%"s2", %%"t2" \n\t"\
 535         "movq %%"s3", %%"t3" \n\t"
 536
 537
 538 static void ff_snow_vertical_compose97i_mmx(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, IDWTELEM *b5, int width){
 539     x86_reg i = width;
 540     while(i & 15)
 541     {
 542         i--;
 543         b4[i] -= (W_DM*(b3[i] + b5[i])+W_DO)>>W_DS;
 544         b3[i] -= (W_CM*(b2[i] + b4[i])+W_CO)>>W_CS;
 545         b2[i] += (W_BM*(b1[i] + b3[i])+4*b2[i]+W_BO)>>W_BS;
 546         b1[i] += (W_AM*(b0[i] + b2[i])+W_AO)>>W_AS;
 547     }
 548     i+=i;
 549     __asm__ volatile(
 550         "jmp 2f                                      \n\t"
 551         "1:                                          \n\t"
 552
 553         snow_vertical_compose_mmx_load("%4","mm1","mm3","mm5","mm7")
 554         snow_vertical_compose_mmx_add("%6","mm1","mm3","mm5","mm7")
 555         "pcmpeqw    %%mm0, %%mm0                     \n\t"
 556         "pcmpeqw    %%mm2, %%mm2                     \n\t"
 557         "paddw      %%mm2, %%mm2                     \n\t"
 558         "paddw      %%mm0, %%mm2                     \n\t"
 559         "psllw        $13, %%mm2                     \n\t"
 560         snow_vertical_compose_r2r_add("mm0","mm0","mm0","mm0","mm1","mm3","mm5","mm7")
 561         snow_vertical_compose_r2r_pmulhw("mm2","mm2","mm2","mm2","mm1","mm3","mm5","mm7")
 562         snow_vertical_compose_mmx_add("%5","mm1","mm3","mm5","mm7")
 563         snow_vertical_compose_mmx_store("%5","mm1","mm3","mm5","mm7")
 564         snow_vertical_compose_mmx_load("%4","mm0","mm2","mm4","mm6")
 565         snow_vertical_compose_mmx_add("%3","mm1","mm3","mm5","mm7")
 566         snow_vertical_compose_r2r_sub("mm1","mm3","mm5","mm7","mm0","mm2","mm4","mm6")
 567         snow_vertical_compose_mmx_store("%4","mm0","mm2","mm4","mm6")
 568         "pcmpeqw %%mm7, %%mm7                        \n\t"
 569         "pcmpeqw %%mm5, %%mm5                        \n\t"
 570         "psllw $15, %%mm7                            \n\t"
 571         "psrlw $13, %%mm5                            \n\t"
 572         "paddw %%mm7, %%mm5                          \n\t"
 573         snow_vertical_compose_r2r_add("mm5","mm5","mm5","mm5","mm0","mm2","mm4","mm6")
 574         "movq   (%2,%%"FF_REG_d"), %%mm1             \n\t"
 575         "movq  8(%2,%%"FF_REG_d"), %%mm3             \n\t"
 576         "paddw %%mm7, %%mm1                          \n\t"
 577         "paddw %%mm7, %%mm3                          \n\t"
 578         "pavgw %%mm1, %%mm0                          \n\t"
 579         "pavgw %%mm3, %%mm2                          \n\t"
 580         "movq 16(%2,%%"FF_REG_d"), %%mm1             \n\t"
 581         "movq 24(%2,%%"FF_REG_d"), %%mm3             \n\t"
 582         "paddw %%mm7, %%mm1                          \n\t"
 583         "paddw %%mm7, %%mm3                          \n\t"
 584         "pavgw %%mm1, %%mm4                          \n\t"
 585         "pavgw %%mm3, %%mm6                          \n\t"
 586         snow_vertical_compose_r2r_sub("mm7","mm7","mm7","mm7","mm0","mm2","mm4","mm6")
 587         snow_vertical_compose_sra("1","mm0","mm2","mm4","mm6")
 588         snow_vertical_compose_mmx_add("%3","mm0","mm2","mm4","mm6")
 589
 590         snow_vertical_compose_sra("2","mm0","mm2","mm4","mm6")
 591         snow_vertical_compose_mmx_add("%3","mm0","mm2","mm4","mm6")
 592         snow_vertical_compose_mmx_store("%3","mm0","mm2","mm4","mm6")
 593         snow_vertical_compose_mmx_add("%1","mm0","mm2","mm4","mm6")
 594         snow_vertical_compose_mmx_move("mm0","mm2","mm4","mm6","mm1","mm3","mm5","mm7")
 595         snow_vertical_compose_sra("1","mm0","mm2","mm4","mm6")
 596         snow_vertical_compose_r2r_add("mm1","mm3","mm5","mm7","mm0","mm2","mm4","mm6")
 597         snow_vertical_compose_mmx_add("%2","mm0","mm2","mm4","mm6")
 598         snow_vertical_compose_mmx_store("%2","mm0","mm2","mm4","mm6")
 599
 600         "2:                                          \n\t"
 601         "sub $32, %%"FF_REG_d"                       \n\t"
 602         "jge 1b                                      \n\t"
 603         :"+d"(i)
 604         :"r"(b0),"r"(b1),"r"(b2),"r"(b3),"r"(b4),"r"(b5));
 605 }
 606 #endif //HAVE_7REGS
 607
 608 #if HAVE_6REGS
 609 #define snow_inner_add_yblock_sse2_header \
 610     IDWTELEM * * dst_array = sb->line + src_y;\
 611     x86_reg tmp;\
 612     __asm__ volatile(\
 613              "mov  %7, %%"FF_REG_c"          \n\t"\
 614              "mov  %6, %2                    \n\t"\
 615              "mov  %4, %%"FF_REG_S"          \n\t"\
 616              "pxor %%xmm7, %%xmm7            \n\t" /* 0 */\
 617              "pcmpeqd %%xmm3, %%xmm3         \n\t"\
 618              "psllw $15, %%xmm3              \n\t"\
 619              "psrlw $12, %%xmm3              \n\t" /* FRAC_BITS >> 1 */\
 620              "1:                             \n\t"\
 621              "mov %1, %%"FF_REG_D"           \n\t"\
 622              "mov (%%"FF_REG_D"), %%"FF_REG_D" \n\t"\
 623              "add %3, %%"FF_REG_D"           \n\t"
 624
 625 #define snow_inner_add_yblock_sse2_start_8(out_reg1, out_reg2, ptr_offset, s_offset)\
 626              "mov "FF_PTR_SIZE"*"ptr_offset"(%%"FF_REG_a"), %%"FF_REG_d"; \n\t"\
 627              "movq (%%"FF_REG_d"), %%"out_reg1"                           \n\t"\
 628              "movq (%%"FF_REG_d", %%"FF_REG_c"), %%"out_reg2"             \n\t"\
 629              "punpcklbw %%xmm7, %%"out_reg1" \n\t"\
 630              "punpcklbw %%xmm7, %%"out_reg2" \n\t"\
 631              "movq "s_offset"(%%"FF_REG_S"), %%xmm0    \n\t"\
 632              "movq "s_offset"+16(%%"FF_REG_S"), %%xmm4 \n\t"\
 633              "punpcklbw %%xmm7, %%xmm0       \n\t"\
 634              "punpcklbw %%xmm7, %%xmm4       \n\t"\
 635              "pmullw %%xmm0, %%"out_reg1"    \n\t"\
 636              "pmullw %%xmm4, %%"out_reg2"    \n\t"
 637
 638 #define snow_inner_add_yblock_sse2_start_16(out_reg1, out_reg2, ptr_offset, s_offset)\
 639              "mov "FF_PTR_SIZE"*"ptr_offset"(%%"FF_REG_a"), %%"FF_REG_d"; \n\t"\
 640              "movq (%%"FF_REG_d"), %%"out_reg1"                           \n\t"\
 641              "movq 8(%%"FF_REG_d"), %%"out_reg2"                          \n\t"\
 642              "punpcklbw %%xmm7, %%"out_reg1" \n\t"\
 643              "punpcklbw %%xmm7, %%"out_reg2" \n\t"\
 644              "movq "s_offset"(%%"FF_REG_S"), %%xmm0   \n\t"\
 645              "movq "s_offset"+8(%%"FF_REG_S"), %%xmm4 \n\t"\
 646              "punpcklbw %%xmm7, %%xmm0       \n\t"\
 647              "punpcklbw %%xmm7, %%xmm4       \n\t"\
 648              "pmullw %%xmm0, %%"out_reg1"    \n\t"\
 649              "pmullw %%xmm4, %%"out_reg2"    \n\t"
 650
 651 #define snow_inner_add_yblock_sse2_accum_8(ptr_offset, s_offset) \
 652              snow_inner_add_yblock_sse2_start_8("xmm2", "xmm6", ptr_offset, s_offset)\
 653              "paddusw %%xmm2, %%xmm1         \n\t"\
 654              "paddusw %%xmm6, %%xmm5         \n\t"
 655
 656 #define snow_inner_add_yblock_sse2_accum_16(ptr_offset, s_offset) \
 657              snow_inner_add_yblock_sse2_start_16("xmm2", "xmm6", ptr_offset, s_offset)\
 658              "paddusw %%xmm2, %%xmm1         \n\t"\
 659              "paddusw %%xmm6, %%xmm5         \n\t"
 660
 661 #define snow_inner_add_yblock_sse2_end_common1\
 662              "add $32, %%"FF_REG_S"                            \n\t"\
 663              "add %%"FF_REG_c", %0                             \n\t"\
 664              "add %%"FF_REG_c", "FF_PTR_SIZE"*3(%%"FF_REG_a"); \n\t"\
 665              "add %%"FF_REG_c", "FF_PTR_SIZE"*2(%%"FF_REG_a"); \n\t"\
 666              "add %%"FF_REG_c", "FF_PTR_SIZE"*1(%%"FF_REG_a"); \n\t"\
 667              "add %%"FF_REG_c", (%%"FF_REG_a")                 \n\t"
 668
 669 #define snow_inner_add_yblock_sse2_end_common2\
 670              "jnz 1b                         \n\t"\
 671              :"+m"(dst8),"+m"(dst_array),"=&r"(tmp)\
 672              :\
 673              "rm"((x86_reg)(src_x<<1)),"m"(obmc),"a"(block),"m"(b_h),"m"(src_stride):\
 674              XMM_CLOBBERS("%xmm0", "%xmm1", "%xmm2", "%xmm3", "%xmm4", "%xmm5", "%xmm6", "%xmm7", )\
 675              "%"FF_REG_c"","%"FF_REG_S"","%"FF_REG_D"","%"FF_REG_d"");
 676
 677 #define snow_inner_add_yblock_sse2_end_8\
 678              "sal $1, %%"FF_REG_c"                \n\t"\
 679              "add"FF_OPSIZE" $"FF_PTR_SIZE"*2, %1 \n\t"\
 680              snow_inner_add_yblock_sse2_end_common1\
 681              "sar $1, %%"FF_REG_c"           \n\t"\
 682              "sub $2, %2                     \n\t"\
 683              snow_inner_add_yblock_sse2_end_common2
 684
 685 #define snow_inner_add_yblock_sse2_end_16\
 686              "add"FF_OPSIZE" $"FF_PTR_SIZE"*1, %1 \n\t"\
 687              snow_inner_add_yblock_sse2_end_common1\
 688              "dec %2                         \n\t"\
 689              snow_inner_add_yblock_sse2_end_common2
 690
 691 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,
 692                       int src_x, int src_y, x86_reg src_stride, slice_buffer * sb, int add, uint8_t * dst8){
 693 snow_inner_add_yblock_sse2_header
 694 snow_inner_add_yblock_sse2_start_8("xmm1", "xmm5", "3", "0")
 695 snow_inner_add_yblock_sse2_accum_8("2", "8")
 696 snow_inner_add_yblock_sse2_accum_8("1", "128")
 697 snow_inner_add_yblock_sse2_accum_8("0", "136")
 698
 699              "mov %0, %%"FF_REG_d"           \n\t"
 700              "movdqa (%%"FF_REG_D"), %%xmm0  \n\t"
 701              "movdqa %%xmm1, %%xmm2          \n\t"
 702
 703              "punpckhwd %%xmm7, %%xmm1       \n\t"
 704              "punpcklwd %%xmm7, %%xmm2       \n\t"
 705              "paddd %%xmm2, %%xmm0           \n\t"
 706              "movdqa 16(%%"FF_REG_D"), %%xmm2\n\t"
 707              "paddd %%xmm1, %%xmm2           \n\t"
 708              "paddd %%xmm3, %%xmm0           \n\t"
 709              "paddd %%xmm3, %%xmm2           \n\t"
 710
 711              "mov %1, %%"FF_REG_D"           \n\t"
 712              "mov "FF_PTR_SIZE"(%%"FF_REG_D"), %%"FF_REG_D"; \n\t"
 713              "add %3, %%"FF_REG_D"           \n\t"
 714
 715              "movdqa (%%"FF_REG_D"), %%xmm4  \n\t"
 716              "movdqa %%xmm5, %%xmm6          \n\t"
 717              "punpckhwd %%xmm7, %%xmm5       \n\t"
 718              "punpcklwd %%xmm7, %%xmm6       \n\t"
 719              "paddd %%xmm6, %%xmm4           \n\t"
 720              "movdqa 16(%%"FF_REG_D"), %%xmm6\n\t"
 721              "paddd %%xmm5, %%xmm6           \n\t"
 722              "paddd %%xmm3, %%xmm4           \n\t"
 723              "paddd %%xmm3, %%xmm6           \n\t"
 724
 725              "psrad $8, %%xmm0               \n\t" /* FRAC_BITS. */
 726              "psrad $8, %%xmm2               \n\t" /* FRAC_BITS. */
 727              "packssdw %%xmm2, %%xmm0        \n\t"
 728              "packuswb %%xmm7, %%xmm0        \n\t"
 729              "movq %%xmm0, (%%"FF_REG_d")    \n\t"
 730
 731              "psrad $8, %%xmm4               \n\t" /* FRAC_BITS. */
 732              "psrad $8, %%xmm6               \n\t" /* FRAC_BITS. */
 733              "packssdw %%xmm6, %%xmm4        \n\t"
 734              "packuswb %%xmm7, %%xmm4        \n\t"
 735              "movq %%xmm4, (%%"FF_REG_d",%%"FF_REG_c"); \n\t"
 736 snow_inner_add_yblock_sse2_end_8
 737 }
 738
 739 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,
 740                       int src_x, int src_y, x86_reg src_stride, slice_buffer * sb, int add, uint8_t * dst8){
 741 snow_inner_add_yblock_sse2_header
 742 snow_inner_add_yblock_sse2_start_16("xmm1", "xmm5", "3", "0")
 743 snow_inner_add_yblock_sse2_accum_16("2", "16")
 744 snow_inner_add_yblock_sse2_accum_16("1", "512")
 745 snow_inner_add_yblock_sse2_accum_16("0", "528")
 746
 747              "mov %0, %%"FF_REG_d"           \n\t"
 748              "psrlw $4, %%xmm1               \n\t"
 749              "psrlw $4, %%xmm5               \n\t"
 750              "paddw   (%%"FF_REG_D"), %%xmm1 \n\t"
 751              "paddw 16(%%"FF_REG_D"), %%xmm5 \n\t"
 752              "paddw %%xmm3, %%xmm1           \n\t"
 753              "paddw %%xmm3, %%xmm5           \n\t"
 754              "psraw $4, %%xmm1               \n\t" /* FRAC_BITS. */
 755              "psraw $4, %%xmm5               \n\t" /* FRAC_BITS. */
 756              "packuswb %%xmm5, %%xmm1        \n\t"
 757
 758              "movdqu %%xmm1, (%%"FF_REG_d")  \n\t"
 759
 760 snow_inner_add_yblock_sse2_end_16
 761 }
 762
 763 #define snow_inner_add_yblock_mmx_header \
 764     IDWTELEM * * dst_array = sb->line + src_y;\
 765     x86_reg tmp;\
 766     __asm__ volatile(\
 767              "mov  %7, %%"FF_REG_c"          \n\t"\
 768              "mov  %6, %2                    \n\t"\
 769              "mov  %4, %%"FF_REG_S"          \n\t"\
 770              "pxor %%mm7, %%mm7              \n\t" /* 0 */\
 771              "pcmpeqd %%mm3, %%mm3           \n\t"\
 772              "psllw $15, %%mm3               \n\t"\
 773              "psrlw $12, %%mm3               \n\t" /* FRAC_BITS >> 1 */\
 774              "1:                             \n\t"\
 775              "mov %1, %%"FF_REG_D"           \n\t"\
 776              "mov (%%"FF_REG_D"), %%"FF_REG_D" \n\t"\
 777              "add %3, %%"FF_REG_D"           \n\t"
 778
 779 #define snow_inner_add_yblock_mmx_start(out_reg1, out_reg2, ptr_offset, s_offset, d_offset)\
 780              "mov "FF_PTR_SIZE"*"ptr_offset"(%%"FF_REG_a"), %%"FF_REG_d"; \n\t"\
 781              "movd "d_offset"(%%"FF_REG_d"), %%"out_reg1"                 \n\t"\
 782              "movd "d_offset"+4(%%"FF_REG_d"), %%"out_reg2"               \n\t"\
 783              "punpcklbw %%mm7, %%"out_reg1" \n\t"\
 784              "punpcklbw %%mm7, %%"out_reg2" \n\t"\
 785              "movd "s_offset"(%%"FF_REG_S"), %%mm0   \n\t"\
 786              "movd "s_offset"+4(%%"FF_REG_S"), %%mm4 \n\t"\
 787              "punpcklbw %%mm7, %%mm0       \n\t"\
 788              "punpcklbw %%mm7, %%mm4       \n\t"\
 789              "pmullw %%mm0, %%"out_reg1"   \n\t"\
 790              "pmullw %%mm4, %%"out_reg2"   \n\t"
 791
 792 #define snow_inner_add_yblock_mmx_accum(ptr_offset, s_offset, d_offset) \
 793              snow_inner_add_yblock_mmx_start("mm2", "mm6", ptr_offset, s_offset, d_offset)\
 794              "paddusw %%mm2, %%mm1         \n\t"\
 795              "paddusw %%mm6, %%mm5         \n\t"
 796
 797 #define snow_inner_add_yblock_mmx_mix(read_offset, write_offset)\
 798              "mov %0, %%"FF_REG_d"           \n\t"\
 799              "psrlw $4, %%mm1                \n\t"\
 800              "psrlw $4, %%mm5                \n\t"\
 801              "paddw "read_offset"(%%"FF_REG_D"), %%mm1   \n\t"\
 802              "paddw "read_offset"+8(%%"FF_REG_D"), %%mm5 \n\t"\
 803              "paddw %%mm3, %%mm1             \n\t"\
 804              "paddw %%mm3, %%mm5             \n\t"\
 805              "psraw $4, %%mm1                \n\t"\
 806              "psraw $4, %%mm5                \n\t"\
 807              "packuswb %%mm5, %%mm1          \n\t"\
 808              "movq %%mm1, "write_offset"(%%"FF_REG_d") \n\t"
 809
 810 #define snow_inner_add_yblock_mmx_end(s_step)\
 811              "add $"s_step", %%"FF_REG_S"                      \n\t"\
 812              "add %%"FF_REG_c", "FF_PTR_SIZE"*3(%%"FF_REG_a"); \n\t"\
 813              "add %%"FF_REG_c", "FF_PTR_SIZE"*2(%%"FF_REG_a"); \n\t"\
 814              "add %%"FF_REG_c", "FF_PTR_SIZE"*1(%%"FF_REG_a"); \n\t"\
 815              "add %%"FF_REG_c", (%%"FF_REG_a")                 \n\t"\
 816              "add"FF_OPSIZE " $"FF_PTR_SIZE"*1, %1             \n\t"\
 817              "add %%"FF_REG_c", %0                             \n\t"\
 818              "dec %2                         \n\t"\
 819              "jnz 1b                         \n\t"\
 820              :"+m"(dst8),"+m"(dst_array),"=&r"(tmp)\
 821              :\
 822              "rm"((x86_reg)(src_x<<1)),"m"(obmc),"a"(block),"m"(b_h),"m"(src_stride):\
 823              "%"FF_REG_c"","%"FF_REG_S"","%"FF_REG_D"","%"FF_REG_d"");
 824
 825 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,
 826                       int src_x, int src_y, x86_reg src_stride, slice_buffer * sb, int add, uint8_t * dst8){
 827 snow_inner_add_yblock_mmx_header
 828 snow_inner_add_yblock_mmx_start("mm1", "mm5", "3", "0", "0")
 829 snow_inner_add_yblock_mmx_accum("2", "8", "0")
 830 snow_inner_add_yblock_mmx_accum("1", "128", "0")
 831 snow_inner_add_yblock_mmx_accum("0", "136", "0")
 832 snow_inner_add_yblock_mmx_mix("0", "0")
 833 snow_inner_add_yblock_mmx_end("16")
 834 }
 835
 836 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,
 837                       int src_x, int src_y, x86_reg src_stride, slice_buffer * sb, int add, uint8_t * dst8){
 838 snow_inner_add_yblock_mmx_header
 839 snow_inner_add_yblock_mmx_start("mm1", "mm5", "3", "0", "0")
 840 snow_inner_add_yblock_mmx_accum("2", "16", "0")
 841 snow_inner_add_yblock_mmx_accum("1", "512", "0")
 842 snow_inner_add_yblock_mmx_accum("0", "528", "0")
 843 snow_inner_add_yblock_mmx_mix("0", "0")
 844
 845 snow_inner_add_yblock_mmx_start("mm1", "mm5", "3", "8", "8")
 846 snow_inner_add_yblock_mmx_accum("2", "24", "8")
 847 snow_inner_add_yblock_mmx_accum("1", "520", "8")
 848 snow_inner_add_yblock_mmx_accum("0", "536", "8")
 849 snow_inner_add_yblock_mmx_mix("16", "8")
 850 snow_inner_add_yblock_mmx_end("32")
 851 }
 852
 853 static void ff_snow_inner_add_yblock_sse2(const uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
 854                            int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8){
 855
 856     if (b_w == 16)
 857         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);
 858     else if (b_w == 8 && obmc_stride == 16) {
 859         if (!(b_h & 1))
 860             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);
 861         else
 862             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);
 863     } else
 864          ff_snow_inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
 865 }
 866
 867 static void ff_snow_inner_add_yblock_mmx(const uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
 868                           int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8){
 869     if (b_w == 16)
 870         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);
 871     else if (b_w == 8 && obmc_stride == 16)
 872         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);
 873     else
 874         ff_snow_inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
 875 }
 876 #endif /* HAVE_6REGS */
 877
 878 #endif /* HAVE_INLINE_ASM */
 879
 880 av_cold void ff_dwt_init_x86(SnowDWTContext *c)
 881 {
 882 #if HAVE_INLINE_ASM
 883     int mm_flags = av_get_cpu_flags();
 884
 885     if (mm_flags & AV_CPU_FLAG_MMX) {
 886         if(mm_flags & AV_CPU_FLAG_SSE2 & 0){
 887             c->horizontal_compose97i = ff_snow_horizontal_compose97i_sse2;
 888 #if HAVE_7REGS
 889             c->vertical_compose97i = ff_snow_vertical_compose97i_sse2;
 890 #endif
 891 #if HAVE_6REGS
 892             c->inner_add_yblock = ff_snow_inner_add_yblock_sse2;
 893 #endif
 894         }
 895         else{
 896             if (mm_flags & AV_CPU_FLAG_MMXEXT) {
 897             c->horizontal_compose97i = ff_snow_horizontal_compose97i_mmx;
 898 #if HAVE_7REGS
 899             c->vertical_compose97i = ff_snow_vertical_compose97i_mmx;
 900 #endif
 901             }
 902 #if HAVE_6REGS
 903             c->inner_add_yblock = ff_snow_inner_add_yblock_mmx;
 904 #endif
 905         }
 906     }
 907 #endif /* HAVE_INLINE_ASM */
 908 }