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 <stdint.h>
  23 #include "config.h"
  24 #include "libavutil/attributes.h"
  25 #include "libavutil/cpu.h"
  26 #include "libavutil/x86/asm.h"
  27 #include "libavcodec/snow_dwt.h"
  28
  29 #if HAVE_INLINE_ASM
  30
  31 static void ff_snow_horizontal_compose97i_sse2(IDWTELEM *b, IDWTELEM *temp, int width){
  32     const int w2= (width+1)>>1;
  33     const int w_l= (width>>1);
  34     const int w_r= w2 - 1;
  35     int i;
  36
  37     { // Lift 0
  38         IDWTELEM * const ref = b + w2 - 1;
  39         IDWTELEM b_0 = b[0]; //By allowing the first entry in b[0] to be calculated twice
  40         // (the first time erroneously), we allow the SSE2 code to run an extra pass.
  41         // The savings in code and time are well worth having to store this value and
  42         // calculate b[0] correctly afterwards.
  43
  44         i = 0;
  45         __asm__ volatile(
  46             "pcmpeqd   %%xmm7, %%xmm7         \n\t"
  47             "pcmpeqd   %%xmm3, %%xmm3         \n\t"
  48             "psllw         $1, %%xmm3         \n\t"
  49             "paddw     %%xmm7, %%xmm3         \n\t"
  50             "psllw        $13, %%xmm3         \n\t"
  51         ::);
  52         for(; i<w_l-15; i+=16){
  53             __asm__ volatile(
  54                 "movdqu   (%1), %%xmm1        \n\t"
  55                 "movdqu 16(%1), %%xmm5        \n\t"
  56                 "movdqu  2(%1), %%xmm2        \n\t"
  57                 "movdqu 18(%1), %%xmm6        \n\t"
  58                 "paddw  %%xmm1, %%xmm2        \n\t"
  59                 "paddw  %%xmm5, %%xmm6        \n\t"
  60                 "paddw  %%xmm7, %%xmm2        \n\t"
  61                 "paddw  %%xmm7, %%xmm6        \n\t"
  62                 "pmulhw %%xmm3, %%xmm2        \n\t"
  63                 "pmulhw %%xmm3, %%xmm6        \n\t"
  64                 "paddw    (%0), %%xmm2        \n\t"
  65                 "paddw  16(%0), %%xmm6        \n\t"
  66                 "movdqa %%xmm2, (%0)          \n\t"
  67                 "movdqa %%xmm6, 16(%0)        \n\t"
  68                 :: "r"(&b[i]), "r"(&ref[i])
  69                 : "memory"
  70             );
  71         }
  72         snow_horizontal_compose_lift_lead_out(i, b, b, ref, width, w_l, 0, W_DM, W_DO, W_DS);
  73         b[0] = b_0 - ((W_DM * 2 * ref[1]+W_DO)>>W_DS);
  74     }
  75
  76     { // Lift 1
  77         IDWTELEM * const dst = b+w2;
  78
  79         i = 0;
  80         for(; (((x86_reg)&dst[i]) & 0x1F) && i<w_r; i++){
  81             dst[i] = dst[i] - (b[i] + b[i + 1]);
  82         }
  83         for(; i<w_r-15; i+=16){
  84             __asm__ volatile(
  85                 "movdqu   (%1), %%xmm1        \n\t"
  86                 "movdqu 16(%1), %%xmm5        \n\t"
  87                 "movdqu  2(%1), %%xmm2        \n\t"
  88                 "movdqu 18(%1), %%xmm6        \n\t"
  89                 "paddw  %%xmm1, %%xmm2        \n\t"
  90                 "paddw  %%xmm5, %%xmm6        \n\t"
  91                 "movdqa   (%0), %%xmm0        \n\t"
  92                 "movdqa 16(%0), %%xmm4        \n\t"
  93                 "psubw  %%xmm2, %%xmm0        \n\t"
  94                 "psubw  %%xmm6, %%xmm4        \n\t"
  95                 "movdqa %%xmm0, (%0)          \n\t"
  96                 "movdqa %%xmm4, 16(%0)        \n\t"
  97                 :: "r"(&dst[i]), "r"(&b[i])
  98                 : "memory"
  99             );
 100         }
 101         snow_horizontal_compose_lift_lead_out(i, dst, dst, b, width, w_r, 1, W_CM, W_CO, W_CS);
 102     }
 103
 104     { // Lift 2
 105         IDWTELEM * const ref = b+w2 - 1;
 106         IDWTELEM b_0 = b[0];
 107
 108         i = 0;
 109         __asm__ volatile(
 110             "psllw         $15, %%xmm7        \n\t"
 111             "pcmpeqw    %%xmm6, %%xmm6        \n\t"
 112             "psrlw         $13, %%xmm6        \n\t"
 113             "paddw      %%xmm7, %%xmm6        \n\t"
 114         ::);
 115         for(; i<w_l-15; i+=16){
 116             __asm__ volatile(
 117                 "movdqu   (%1), %%xmm0        \n\t"
 118                 "movdqu 16(%1), %%xmm4        \n\t"
 119                 "movdqu  2(%1), %%xmm1        \n\t"
 120                 "movdqu 18(%1), %%xmm5        \n\t" //FIXME try aligned reads and shifts
 121                 "paddw  %%xmm6, %%xmm0        \n\t"
 122                 "paddw  %%xmm6, %%xmm4        \n\t"
 123                 "paddw  %%xmm7, %%xmm1        \n\t"
 124                 "paddw  %%xmm7, %%xmm5        \n\t"
 125                 "pavgw  %%xmm1, %%xmm0        \n\t"
 126                 "pavgw  %%xmm5, %%xmm4        \n\t"
 127                 "psubw  %%xmm7, %%xmm0        \n\t"
 128                 "psubw  %%xmm7, %%xmm4        \n\t"
 129                 "psraw      $1, %%xmm0        \n\t"
 130                 "psraw      $1, %%xmm4        \n\t"
 131                 "movdqa   (%0), %%xmm1        \n\t"
 132                 "movdqa 16(%0), %%xmm5        \n\t"
 133                 "paddw  %%xmm1, %%xmm0        \n\t"
 134                 "paddw  %%xmm5, %%xmm4        \n\t"
 135                 "psraw      $2, %%xmm0        \n\t"
 136                 "psraw      $2, %%xmm4        \n\t"
 137                 "paddw  %%xmm1, %%xmm0        \n\t"
 138                 "paddw  %%xmm5, %%xmm4        \n\t"
 139                 "movdqa %%xmm0, (%0)          \n\t"
 140                 "movdqa %%xmm4, 16(%0)        \n\t"
 141                 :: "r"(&b[i]), "r"(&ref[i])
 142                 : "memory"
 143             );
 144         }
 145         snow_horizontal_compose_liftS_lead_out(i, b, b, ref, width, w_l);
 146         b[0] = b_0 + ((2 * ref[1] + W_BO-1 + 4 * b_0) >> W_BS);
 147     }
 148
 149     { // Lift 3
 150         IDWTELEM * const src = b+w2;
 151
 152         i = 0;
 153         for(; (((x86_reg)&temp[i]) & 0x1F) && i<w_r; i++){
 154             temp[i] = src[i] - ((-W_AM*(b[i] + b[i+1]))>>W_AS);
 155         }
 156         for(; i<w_r-7; i+=8){
 157             __asm__ volatile(
 158                 "movdqu  2(%1), %%xmm2        \n\t"
 159                 "movdqu 18(%1), %%xmm6        \n\t"
 160                 "paddw    (%1), %%xmm2        \n\t"
 161                 "paddw  16(%1), %%xmm6        \n\t"
 162                 "movdqu   (%0), %%xmm0        \n\t"
 163                 "movdqu 16(%0), %%xmm4        \n\t"
 164                 "paddw  %%xmm2, %%xmm0        \n\t"
 165                 "paddw  %%xmm6, %%xmm4        \n\t"
 166                 "psraw      $1, %%xmm2        \n\t"
 167                 "psraw      $1, %%xmm6        \n\t"
 168                 "paddw  %%xmm0, %%xmm2        \n\t"
 169                 "paddw  %%xmm4, %%xmm6        \n\t"
 170                 "movdqa %%xmm2, (%2)          \n\t"
 171                 "movdqa %%xmm6, 16(%2)        \n\t"
 172                 :: "r"(&src[i]), "r"(&b[i]), "r"(&temp[i])
 173                  : "memory"
 174                );
 175         }
 176         snow_horizontal_compose_lift_lead_out(i, temp, src, b, width, w_r, 1, -W_AM, W_AO+1, W_AS);
 177     }
 178
 179     {
 180         snow_interleave_line_header(&i, width, b, temp);
 181
 182         for (; (i & 0x3E) != 0x3E; i-=2){
 183             b[i+1] = temp[i>>1];
 184             b[i] = b[i>>1];
 185         }
 186         for (i-=62; i>=0; i-=64){
 187             __asm__ volatile(
 188                 "movdqa      (%1), %%xmm0       \n\t"
 189                 "movdqa    16(%1), %%xmm2       \n\t"
 190                 "movdqa    32(%1), %%xmm4       \n\t"
 191                 "movdqa    48(%1), %%xmm6       \n\t"
 192                 "movdqa      (%1), %%xmm1       \n\t"
 193                 "movdqa    16(%1), %%xmm3       \n\t"
 194                 "movdqa    32(%1), %%xmm5       \n\t"
 195                 "movdqa    48(%1), %%xmm7       \n\t"
 196                 "punpcklwd   (%2), %%xmm0       \n\t"
 197                 "punpcklwd 16(%2), %%xmm2       \n\t"
 198                 "punpcklwd 32(%2), %%xmm4       \n\t"
 199                 "punpcklwd 48(%2), %%xmm6       \n\t"
 200                 "movdqa    %%xmm0, (%0)         \n\t"
 201                 "movdqa    %%xmm2, 32(%0)       \n\t"
 202                 "movdqa    %%xmm4, 64(%0)       \n\t"
 203                 "movdqa    %%xmm6, 96(%0)       \n\t"
 204                 "punpckhwd   (%2), %%xmm1       \n\t"
 205                 "punpckhwd 16(%2), %%xmm3       \n\t"
 206                 "punpckhwd 32(%2), %%xmm5       \n\t"
 207                 "punpckhwd 48(%2), %%xmm7       \n\t"
 208                 "movdqa    %%xmm1, 16(%0)       \n\t"
 209                 "movdqa    %%xmm3, 48(%0)       \n\t"
 210                 "movdqa    %%xmm5, 80(%0)       \n\t"
 211                 "movdqa    %%xmm7, 112(%0)      \n\t"
 212                 :: "r"(&(b)[i]), "r"(&(b)[i>>1]), "r"(&(temp)[i>>1])
 213                  : "memory"
 214                );
 215         }
 216     }
 217 }
 218
 219 static void ff_snow_horizontal_compose97i_mmx(IDWTELEM *b, IDWTELEM *temp, int width){
 220     const int w2= (width+1)>>1;
 221     const int w_l= (width>>1);
 222     const int w_r= w2 - 1;
 223     int i;
 224
 225     { // Lift 0
 226         IDWTELEM * const ref = b + w2 - 1;
 227
 228         i = 1;
 229         b[0] = b[0] - ((W_DM * 2 * ref[1]+W_DO)>>W_DS);
 230         __asm__ volatile(
 231             "pcmpeqw    %%mm7, %%mm7         \n\t"
 232             "pcmpeqw    %%mm3, %%mm3         \n\t"
 233             "psllw         $1, %%mm3         \n\t"
 234             "paddw      %%mm7, %%mm3         \n\t"
 235             "psllw        $13, %%mm3         \n\t"
 236            ::);
 237         for(; i<w_l-7; i+=8){
 238             __asm__ volatile(
 239                 "movq     (%1), %%mm2        \n\t"
 240                 "movq    8(%1), %%mm6        \n\t"
 241                 "paddw   2(%1), %%mm2        \n\t"
 242                 "paddw  10(%1), %%mm6        \n\t"
 243                 "paddw   %%mm7, %%mm2        \n\t"
 244                 "paddw   %%mm7, %%mm6        \n\t"
 245                 "pmulhw  %%mm3, %%mm2        \n\t"
 246                 "pmulhw  %%mm3, %%mm6        \n\t"
 247                 "paddw    (%0), %%mm2        \n\t"
 248                 "paddw   8(%0), %%mm6        \n\t"
 249                 "movq    %%mm2, (%0)         \n\t"
 250                 "movq    %%mm6, 8(%0)        \n\t"
 251                 :: "r"(&b[i]), "r"(&ref[i])
 252                  : "memory"
 253                );
 254         }
 255         snow_horizontal_compose_lift_lead_out(i, b, b, ref, width, w_l, 0, W_DM, W_DO, W_DS);
 256     }
 257
 258     { // Lift 1
 259         IDWTELEM * const dst = b+w2;
 260
 261         i = 0;
 262         for(; i<w_r-7; i+=8){
 263             __asm__ volatile(
 264                 "movq     (%1), %%mm2        \n\t"
 265                 "movq    8(%1), %%mm6        \n\t"
 266                 "paddw   2(%1), %%mm2        \n\t"
 267                 "paddw  10(%1), %%mm6        \n\t"
 268                 "movq     (%0), %%mm0        \n\t"
 269                 "movq    8(%0), %%mm4        \n\t"
 270                 "psubw   %%mm2, %%mm0        \n\t"
 271                 "psubw   %%mm6, %%mm4        \n\t"
 272                 "movq    %%mm0, (%0)         \n\t"
 273                 "movq    %%mm4, 8(%0)        \n\t"
 274                 :: "r"(&dst[i]), "r"(&b[i])
 275                  : "memory"
 276                );
 277         }
 278         snow_horizontal_compose_lift_lead_out(i, dst, dst, b, width, w_r, 1, W_CM, W_CO, W_CS);
 279     }
 280
 281     { // Lift 2
 282         IDWTELEM * const ref = b+w2 - 1;
 283
 284         i = 1;
 285         b[0] = b[0] + (((2 * ref[1] + W_BO) + 4 * b[0]) >> W_BS);
 286         __asm__ volatile(
 287             "psllw         $15, %%mm7        \n\t"
 288             "pcmpeqw     %%mm6, %%mm6        \n\t"
 289             "psrlw         $13, %%mm6        \n\t"
 290             "paddw       %%mm7, %%mm6        \n\t"
 291            ::);
 292         for(; i<w_l-7; i+=8){
 293             __asm__ volatile(
 294                 "movq     (%1), %%mm0        \n\t"
 295                 "movq    8(%1), %%mm4        \n\t"
 296                 "movq    2(%1), %%mm1        \n\t"
 297                 "movq   10(%1), %%mm5        \n\t"
 298                 "paddw   %%mm6, %%mm0        \n\t"
 299                 "paddw   %%mm6, %%mm4        \n\t"
 300                 "paddw   %%mm7, %%mm1        \n\t"
 301                 "paddw   %%mm7, %%mm5        \n\t"
 302                 "pavgw   %%mm1, %%mm0        \n\t"
 303                 "pavgw   %%mm5, %%mm4        \n\t"
 304                 "psubw   %%mm7, %%mm0        \n\t"
 305                 "psubw   %%mm7, %%mm4        \n\t"
 306                 "psraw      $1, %%mm0        \n\t"
 307                 "psraw      $1, %%mm4        \n\t"
 308                 "movq     (%0), %%mm1        \n\t"
 309                 "movq    8(%0), %%mm5        \n\t"
 310                 "paddw   %%mm1, %%mm0        \n\t"
 311                 "paddw   %%mm5, %%mm4        \n\t"
 312                 "psraw      $2, %%mm0        \n\t"
 313                 "psraw      $2, %%mm4        \n\t"
 314                 "paddw   %%mm1, %%mm0        \n\t"
 315                 "paddw   %%mm5, %%mm4        \n\t"
 316                 "movq    %%mm0, (%0)         \n\t"
 317                 "movq    %%mm4, 8(%0)        \n\t"
 318                 :: "r"(&b[i]), "r"(&ref[i])
 319                  : "memory"
 320                );
 321         }
 322         snow_horizontal_compose_liftS_lead_out(i, b, b, ref, width, w_l);
 323     }
 324
 325     { // Lift 3
 326         IDWTELEM * const src = b+w2;
 327         i = 0;
 328
 329         for(; i<w_r-7; i+=8){
 330             __asm__ volatile(
 331                 "movq    2(%1), %%mm2        \n\t"
 332                 "movq   10(%1), %%mm6        \n\t"
 333                 "paddw    (%1), %%mm2        \n\t"
 334                 "paddw   8(%1), %%mm6        \n\t"
 335                 "movq     (%0), %%mm0        \n\t"
 336                 "movq    8(%0), %%mm4        \n\t"
 337                 "paddw   %%mm2, %%mm0        \n\t"
 338                 "paddw   %%mm6, %%mm4        \n\t"
 339                 "psraw      $1, %%mm2        \n\t"
 340                 "psraw      $1, %%mm6        \n\t"
 341                 "paddw   %%mm0, %%mm2        \n\t"
 342                 "paddw   %%mm4, %%mm6        \n\t"
 343                 "movq    %%mm2, (%2)         \n\t"
 344                 "movq    %%mm6, 8(%2)        \n\t"
 345                 :: "r"(&src[i]), "r"(&b[i]), "r"(&temp[i])
 346                  : "memory"
 347                );
 348         }
 349         snow_horizontal_compose_lift_lead_out(i, temp, src, b, width, w_r, 1, -W_AM, W_AO+1, W_AS);
 350     }
 351
 352     {
 353         snow_interleave_line_header(&i, width, b, temp);
 354
 355         for (; (i & 0x1E) != 0x1E; i-=2){
 356             b[i+1] = temp[i>>1];
 357             b[i] = b[i>>1];
 358         }
 359         for (i-=30; i>=0; i-=32){
 360             __asm__ volatile(
 361                 "movq        (%1), %%mm0       \n\t"
 362                 "movq       8(%1), %%mm2       \n\t"
 363                 "movq      16(%1), %%mm4       \n\t"
 364                 "movq      24(%1), %%mm6       \n\t"
 365                 "movq        (%1), %%mm1       \n\t"
 366                 "movq       8(%1), %%mm3       \n\t"
 367                 "movq      16(%1), %%mm5       \n\t"
 368                 "movq      24(%1), %%mm7       \n\t"
 369                 "punpcklwd   (%2), %%mm0       \n\t"
 370                 "punpcklwd  8(%2), %%mm2       \n\t"
 371                 "punpcklwd 16(%2), %%mm4       \n\t"
 372                 "punpcklwd 24(%2), %%mm6       \n\t"
 373                 "movq       %%mm0, (%0)        \n\t"
 374                 "movq       %%mm2, 16(%0)      \n\t"
 375                 "movq       %%mm4, 32(%0)      \n\t"
 376                 "movq       %%mm6, 48(%0)      \n\t"
 377                 "punpckhwd   (%2), %%mm1       \n\t"
 378                 "punpckhwd  8(%2), %%mm3       \n\t"
 379                 "punpckhwd 16(%2), %%mm5       \n\t"
 380                 "punpckhwd 24(%2), %%mm7       \n\t"
 381                 "movq       %%mm1, 8(%0)       \n\t"
 382                 "movq       %%mm3, 24(%0)      \n\t"
 383                 "movq       %%mm5, 40(%0)      \n\t"
 384                 "movq       %%mm7, 56(%0)      \n\t"
 385                 :: "r"(&b[i]), "r"(&b[i>>1]), "r"(&temp[i>>1])
 386                  : "memory"
 387                );
 388         }
 389     }
 390 }
 391
 392 #if HAVE_7REGS
 393 #define snow_vertical_compose_sse2_load_add(op,r,t0,t1,t2,t3)\
 394         ""op" ("r",%%"FF_REG_d"), %%"t0"      \n\t"\
 395         ""op" 16("r",%%"FF_REG_d"), %%"t1"    \n\t"\
 396         ""op" 32("r",%%"FF_REG_d"), %%"t2"    \n\t"\
 397         ""op" 48("r",%%"FF_REG_d"), %%"t3"    \n\t"
 398
 399 #define snow_vertical_compose_sse2_load(r,t0,t1,t2,t3)\
 400         snow_vertical_compose_sse2_load_add("movdqa",r,t0,t1,t2,t3)
 401
 402 #define snow_vertical_compose_sse2_add(r,t0,t1,t2,t3)\
 403         snow_vertical_compose_sse2_load_add("paddw",r,t0,t1,t2,t3)
 404
 405 #define snow_vertical_compose_r2r_sub(s0,s1,s2,s3,t0,t1,t2,t3)\
 406         "psubw %%"s0", %%"t0" \n\t"\
 407         "psubw %%"s1", %%"t1" \n\t"\
 408         "psubw %%"s2", %%"t2" \n\t"\
 409         "psubw %%"s3", %%"t3" \n\t"
 410
 411 #define snow_vertical_compose_sse2_store(w,s0,s1,s2,s3)\
 412         "movdqa %%"s0", ("w",%%"FF_REG_d")    \n\t"\
 413         "movdqa %%"s1", 16("w",%%"FF_REG_d")  \n\t"\
 414         "movdqa %%"s2", 32("w",%%"FF_REG_d")  \n\t"\
 415         "movdqa %%"s3", 48("w",%%"FF_REG_d")  \n\t"
 416
 417 #define snow_vertical_compose_sra(n,t0,t1,t2,t3)\
 418         "psraw $"n", %%"t0" \n\t"\
 419         "psraw $"n", %%"t1" \n\t"\
 420         "psraw $"n", %%"t2" \n\t"\
 421         "psraw $"n", %%"t3" \n\t"
 422
 423 #define snow_vertical_compose_r2r_add(s0,s1,s2,s3,t0,t1,t2,t3)\
 424         "paddw %%"s0", %%"t0" \n\t"\
 425         "paddw %%"s1", %%"t1" \n\t"\
 426         "paddw %%"s2", %%"t2" \n\t"\
 427         "paddw %%"s3", %%"t3" \n\t"
 428
 429 #define snow_vertical_compose_r2r_pmulhw(s0,s1,s2,s3,t0,t1,t2,t3)\
 430         "pmulhw %%"s0", %%"t0" \n\t"\
 431         "pmulhw %%"s1", %%"t1" \n\t"\
 432         "pmulhw %%"s2", %%"t2" \n\t"\
 433         "pmulhw %%"s3", %%"t3" \n\t"
 434
 435 #define snow_vertical_compose_sse2_move(s0,s1,s2,s3,t0,t1,t2,t3)\
 436         "movdqa %%"s0", %%"t0" \n\t"\
 437         "movdqa %%"s1", %%"t1" \n\t"\
 438         "movdqa %%"s2", %%"t2" \n\t"\
 439         "movdqa %%"s3", %%"t3" \n\t"
 440
 441 static void ff_snow_vertical_compose97i_sse2(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, IDWTELEM *b5, int width){
 442     x86_reg i = width;
 443
 444     while(i & 0x1F)
 445     {
 446         i--;
 447         b4[i] -= (W_DM*(b3[i] + b5[i])+W_DO)>>W_DS;
 448         b3[i] -= (W_CM*(b2[i] + b4[i])+W_CO)>>W_CS;
 449         b2[i] += (W_BM*(b1[i] + b3[i])+4*b2[i]+W_BO)>>W_BS;
 450         b1[i] += (W_AM*(b0[i] + b2[i])+W_AO)>>W_AS;
 451     }
 452     i+=i;
 453
 454          __asm__ volatile (
 455         "jmp 2f                                      \n\t"
 456         "1:                                          \n\t"
 457         snow_vertical_compose_sse2_load("%4","xmm0","xmm2","xmm4","xmm6")
 458         snow_vertical_compose_sse2_add("%6","xmm0","xmm2","xmm4","xmm6")
 459
 460
 461         "pcmpeqw    %%xmm0, %%xmm0                   \n\t"
 462         "pcmpeqw    %%xmm2, %%xmm2                   \n\t"
 463         "paddw      %%xmm2, %%xmm2                   \n\t"
 464         "paddw      %%xmm0, %%xmm2                   \n\t"
 465         "psllw         $13, %%xmm2                   \n\t"
 466         snow_vertical_compose_r2r_add("xmm0","xmm0","xmm0","xmm0","xmm1","xmm3","xmm5","xmm7")
 467         snow_vertical_compose_r2r_pmulhw("xmm2","xmm2","xmm2","xmm2","xmm1","xmm3","xmm5","xmm7")
 468         snow_vertical_compose_sse2_add("%5","xmm1","xmm3","xmm5","xmm7")
 469         snow_vertical_compose_sse2_store("%5","xmm1","xmm3","xmm5","xmm7")
 470         snow_vertical_compose_sse2_load("%4","xmm0","xmm2","xmm4","xmm6")
 471         snow_vertical_compose_sse2_add("%3","xmm1","xmm3","xmm5","xmm7")
 472         snow_vertical_compose_r2r_sub("xmm1","xmm3","xmm5","xmm7","xmm0","xmm2","xmm4","xmm6")
 473         snow_vertical_compose_sse2_store("%4","xmm0","xmm2","xmm4","xmm6")
 474
 475         "pcmpeqw %%xmm7, %%xmm7                      \n\t"
 476         "pcmpeqw %%xmm5, %%xmm5                      \n\t"
 477         "psllw $15, %%xmm7                           \n\t"
 478         "psrlw $13, %%xmm5                           \n\t"
 479         "paddw %%xmm7, %%xmm5                        \n\t"
 480         snow_vertical_compose_r2r_add("xmm5","xmm5","xmm5","xmm5","xmm0","xmm2","xmm4","xmm6")
 481         "movq   (%2,%%"FF_REG_d"), %%xmm1            \n\t"
 482         "movq  8(%2,%%"FF_REG_d"), %%xmm3            \n\t"
 483         "paddw %%xmm7, %%xmm1                        \n\t"
 484         "paddw %%xmm7, %%xmm3                        \n\t"
 485         "pavgw %%xmm1, %%xmm0                        \n\t"
 486         "pavgw %%xmm3, %%xmm2                        \n\t"
 487         "movq 16(%2,%%"FF_REG_d"), %%xmm1            \n\t"
 488         "movq 24(%2,%%"FF_REG_d"), %%xmm3            \n\t"
 489         "paddw %%xmm7, %%xmm1                        \n\t"
 490         "paddw %%xmm7, %%xmm3                        \n\t"
 491         "pavgw %%xmm1, %%xmm4                        \n\t"
 492         "pavgw %%xmm3, %%xmm6                        \n\t"
 493         snow_vertical_compose_r2r_sub("xmm7","xmm7","xmm7","xmm7","xmm0","xmm2","xmm4","xmm6")
 494         snow_vertical_compose_sra("1","xmm0","xmm2","xmm4","xmm6")
 495         snow_vertical_compose_sse2_add("%3","xmm0","xmm2","xmm4","xmm6")
 496
 497         snow_vertical_compose_sra("2","xmm0","xmm2","xmm4","xmm6")
 498         snow_vertical_compose_sse2_add("%3","xmm0","xmm2","xmm4","xmm6")
 499         snow_vertical_compose_sse2_store("%3","xmm0","xmm2","xmm4","xmm6")
 500         snow_vertical_compose_sse2_add("%1","xmm0","xmm2","xmm4","xmm6")
 501         snow_vertical_compose_sse2_move("xmm0","xmm2","xmm4","xmm6","xmm1","xmm3","xmm5","xmm7")
 502         snow_vertical_compose_sra("1","xmm0","xmm2","xmm4","xmm6")
 503         snow_vertical_compose_r2r_add("xmm1","xmm3","xmm5","xmm7","xmm0","xmm2","xmm4","xmm6")
 504         snow_vertical_compose_sse2_add("%2","xmm0","xmm2","xmm4","xmm6")
 505         snow_vertical_compose_sse2_store("%2","xmm0","xmm2","xmm4","xmm6")
 506
 507         "2:                                          \n\t"
 508         "sub $64, %%"FF_REG_d"                       \n\t"
 509         "jge 1b                                      \n\t"
 510         :"+d"(i)
 511         :"r"(b0),"r"(b1),"r"(b2),"r"(b3),"r"(b4),"r"(b5));
 512 }
 513
 514 #define snow_vertical_compose_mmx_load_add(op,r,t0,t1,t2,t3)\
 515         ""op" ("r",%%"FF_REG_d"), %%"t0"   \n\t"\
 516         ""op" 8("r",%%"FF_REG_d"), %%"t1"  \n\t"\
 517         ""op" 16("r",%%"FF_REG_d"), %%"t2" \n\t"\
 518         ""op" 24("r",%%"FF_REG_d"), %%"t3" \n\t"
 519
 520 #define snow_vertical_compose_mmx_load(r,t0,t1,t2,t3)\
 521         snow_vertical_compose_mmx_load_add("movq",r,t0,t1,t2,t3)
 522
 523 #define snow_vertical_compose_mmx_add(r,t0,t1,t2,t3)\
 524         snow_vertical_compose_mmx_load_add("paddw",r,t0,t1,t2,t3)
 525
 526 #define snow_vertical_compose_mmx_store(w,s0,s1,s2,s3)\
 527         "movq %%"s0", ("w",%%"FF_REG_d")   \n\t"\
 528         "movq %%"s1", 8("w",%%"FF_REG_d")  \n\t"\
 529         "movq %%"s2", 16("w",%%"FF_REG_d") \n\t"\
 530         "movq %%"s3", 24("w",%%"FF_REG_d") \n\t"
 531
 532 #define snow_vertical_compose_mmx_move(s0,s1,s2,s3,t0,t1,t2,t3)\
 533         "movq %%"s0", %%"t0" \n\t"\
 534         "movq %%"s1", %%"t1" \n\t"\
 535         "movq %%"s2", %%"t2" \n\t"\
 536         "movq %%"s3", %%"t3" \n\t"
 537
 538
 539 static void ff_snow_vertical_compose97i_mmx(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, IDWTELEM *b5, int width){
 540     x86_reg i = width;
 541     while(i & 15)
 542     {
 543         i--;
 544         b4[i] -= (W_DM*(b3[i] + b5[i])+W_DO)>>W_DS;
 545         b3[i] -= (W_CM*(b2[i] + b4[i])+W_CO)>>W_CS;
 546         b2[i] += (W_BM*(b1[i] + b3[i])+4*b2[i]+W_BO)>>W_BS;
 547         b1[i] += (W_AM*(b0[i] + b2[i])+W_AO)>>W_AS;
 548     }
 549     i+=i;
 550     __asm__ volatile(
 551         "jmp 2f                                      \n\t"
 552         "1:                                          \n\t"
 553
 554         snow_vertical_compose_mmx_load("%4","mm1","mm3","mm5","mm7")
 555         snow_vertical_compose_mmx_add("%6","mm1","mm3","mm5","mm7")
 556         "pcmpeqw    %%mm0, %%mm0                     \n\t"
 557         "pcmpeqw    %%mm2, %%mm2                     \n\t"
 558         "paddw      %%mm2, %%mm2                     \n\t"
 559         "paddw      %%mm0, %%mm2                     \n\t"
 560         "psllw        $13, %%mm2                     \n\t"
 561         snow_vertical_compose_r2r_add("mm0","mm0","mm0","mm0","mm1","mm3","mm5","mm7")
 562         snow_vertical_compose_r2r_pmulhw("mm2","mm2","mm2","mm2","mm1","mm3","mm5","mm7")
 563         snow_vertical_compose_mmx_add("%5","mm1","mm3","mm5","mm7")
 564         snow_vertical_compose_mmx_store("%5","mm1","mm3","mm5","mm7")
 565         snow_vertical_compose_mmx_load("%4","mm0","mm2","mm4","mm6")
 566         snow_vertical_compose_mmx_add("%3","mm1","mm3","mm5","mm7")
 567         snow_vertical_compose_r2r_sub("mm1","mm3","mm5","mm7","mm0","mm2","mm4","mm6")
 568         snow_vertical_compose_mmx_store("%4","mm0","mm2","mm4","mm6")
 569         "pcmpeqw %%mm7, %%mm7                        \n\t"
 570         "pcmpeqw %%mm5, %%mm5                        \n\t"
 571         "psllw $15, %%mm7                            \n\t"
 572         "psrlw $13, %%mm5                            \n\t"
 573         "paddw %%mm7, %%mm5                          \n\t"
 574         snow_vertical_compose_r2r_add("mm5","mm5","mm5","mm5","mm0","mm2","mm4","mm6")
 575         "movq   (%2,%%"FF_REG_d"), %%mm1             \n\t"
 576         "movq  8(%2,%%"FF_REG_d"), %%mm3             \n\t"
 577         "paddw %%mm7, %%mm1                          \n\t"
 578         "paddw %%mm7, %%mm3                          \n\t"
 579         "pavgw %%mm1, %%mm0                          \n\t"
 580         "pavgw %%mm3, %%mm2                          \n\t"
 581         "movq 16(%2,%%"FF_REG_d"), %%mm1             \n\t"
 582         "movq 24(%2,%%"FF_REG_d"), %%mm3             \n\t"
 583         "paddw %%mm7, %%mm1                          \n\t"
 584         "paddw %%mm7, %%mm3                          \n\t"
 585         "pavgw %%mm1, %%mm4                          \n\t"
 586         "pavgw %%mm3, %%mm6                          \n\t"
 587         snow_vertical_compose_r2r_sub("mm7","mm7","mm7","mm7","mm0","mm2","mm4","mm6")
 588         snow_vertical_compose_sra("1","mm0","mm2","mm4","mm6")
 589         snow_vertical_compose_mmx_add("%3","mm0","mm2","mm4","mm6")
 590
 591         snow_vertical_compose_sra("2","mm0","mm2","mm4","mm6")
 592         snow_vertical_compose_mmx_add("%3","mm0","mm2","mm4","mm6")
 593         snow_vertical_compose_mmx_store("%3","mm0","mm2","mm4","mm6")
 594         snow_vertical_compose_mmx_add("%1","mm0","mm2","mm4","mm6")
 595         snow_vertical_compose_mmx_move("mm0","mm2","mm4","mm6","mm1","mm3","mm5","mm7")
 596         snow_vertical_compose_sra("1","mm0","mm2","mm4","mm6")
 597         snow_vertical_compose_r2r_add("mm1","mm3","mm5","mm7","mm0","mm2","mm4","mm6")
 598         snow_vertical_compose_mmx_add("%2","mm0","mm2","mm4","mm6")
 599         snow_vertical_compose_mmx_store("%2","mm0","mm2","mm4","mm6")
 600
 601         "2:                                          \n\t"
 602         "sub $32, %%"FF_REG_d"                       \n\t"
 603         "jge 1b                                      \n\t"
 604         :"+d"(i)
 605         :"r"(b0),"r"(b1),"r"(b2),"r"(b3),"r"(b4),"r"(b5));
 606 }
 607 #endif //HAVE_7REGS
 608
 609 #if HAVE_6REGS
 610 #define snow_inner_add_yblock_sse2_header \
 611     IDWTELEM * * dst_array = sb->line + src_y;\
 612     x86_reg tmp;\
 613     __asm__ volatile(\
 614              "mov  %7, %%"FF_REG_c"          \n\t"\
 615              "mov  %6, %2                    \n\t"\
 616              "mov  %4, %%"FF_REG_S"          \n\t"\
 617              "pxor %%xmm7, %%xmm7            \n\t" /* 0 */\
 618              "pcmpeqd %%xmm3, %%xmm3         \n\t"\
 619              "psllw $15, %%xmm3              \n\t"\
 620              "psrlw $12, %%xmm3              \n\t" /* FRAC_BITS >> 1 */\
 621              "1:                             \n\t"\
 622              "mov %1, %%"FF_REG_D"           \n\t"\
 623              "mov (%%"FF_REG_D"), %%"FF_REG_D" \n\t"\
 624              "add %3, %%"FF_REG_D"           \n\t"
 625
 626 #define snow_inner_add_yblock_sse2_start_8(out_reg1, out_reg2, ptr_offset, s_offset)\
 627              "mov "FF_PTR_SIZE"*"ptr_offset"(%%"FF_REG_a"), %%"FF_REG_d"; \n\t"\
 628              "movq (%%"FF_REG_d"), %%"out_reg1"                           \n\t"\
 629              "movq (%%"FF_REG_d", %%"FF_REG_c"), %%"out_reg2"             \n\t"\
 630              "punpcklbw %%xmm7, %%"out_reg1" \n\t"\
 631              "punpcklbw %%xmm7, %%"out_reg2" \n\t"\
 632              "movq "s_offset"(%%"FF_REG_S"), %%xmm0    \n\t"\
 633              "movq "s_offset"+16(%%"FF_REG_S"), %%xmm4 \n\t"\
 634              "punpcklbw %%xmm7, %%xmm0       \n\t"\
 635              "punpcklbw %%xmm7, %%xmm4       \n\t"\
 636              "pmullw %%xmm0, %%"out_reg1"    \n\t"\
 637              "pmullw %%xmm4, %%"out_reg2"    \n\t"
 638
 639 #define snow_inner_add_yblock_sse2_start_16(out_reg1, out_reg2, ptr_offset, s_offset)\
 640              "mov "FF_PTR_SIZE"*"ptr_offset"(%%"FF_REG_a"), %%"FF_REG_d"; \n\t"\
 641              "movq (%%"FF_REG_d"), %%"out_reg1"                           \n\t"\
 642              "movq 8(%%"FF_REG_d"), %%"out_reg2"                          \n\t"\
 643              "punpcklbw %%xmm7, %%"out_reg1" \n\t"\
 644              "punpcklbw %%xmm7, %%"out_reg2" \n\t"\
 645              "movq "s_offset"(%%"FF_REG_S"), %%xmm0   \n\t"\
 646              "movq "s_offset"+8(%%"FF_REG_S"), %%xmm4 \n\t"\
 647              "punpcklbw %%xmm7, %%xmm0       \n\t"\
 648              "punpcklbw %%xmm7, %%xmm4       \n\t"\
 649              "pmullw %%xmm0, %%"out_reg1"    \n\t"\
 650              "pmullw %%xmm4, %%"out_reg2"    \n\t"
 651
 652 #define snow_inner_add_yblock_sse2_accum_8(ptr_offset, s_offset) \
 653              snow_inner_add_yblock_sse2_start_8("xmm2", "xmm6", ptr_offset, s_offset)\
 654              "paddusw %%xmm2, %%xmm1         \n\t"\
 655              "paddusw %%xmm6, %%xmm5         \n\t"
 656
 657 #define snow_inner_add_yblock_sse2_accum_16(ptr_offset, s_offset) \
 658              snow_inner_add_yblock_sse2_start_16("xmm2", "xmm6", ptr_offset, s_offset)\
 659              "paddusw %%xmm2, %%xmm1         \n\t"\
 660              "paddusw %%xmm6, %%xmm5         \n\t"
 661
 662 #define snow_inner_add_yblock_sse2_end_common1\
 663              "add $32, %%"FF_REG_S"                            \n\t"\
 664              "add %%"FF_REG_c", %0                             \n\t"\
 665              "add %%"FF_REG_c", "FF_PTR_SIZE"*3(%%"FF_REG_a"); \n\t"\
 666              "add %%"FF_REG_c", "FF_PTR_SIZE"*2(%%"FF_REG_a"); \n\t"\
 667              "add %%"FF_REG_c", "FF_PTR_SIZE"*1(%%"FF_REG_a"); \n\t"\
 668              "add %%"FF_REG_c", (%%"FF_REG_a")                 \n\t"
 669
 670 #define snow_inner_add_yblock_sse2_end_common2\
 671              "jnz 1b                         \n\t"\
 672              :"+m"(dst8),"+m"(dst_array),"=&r"(tmp)\
 673              :\
 674              "rm"((x86_reg)(src_x<<1)),"m"(obmc),"a"(block),"m"(b_h),"m"(src_stride):\
 675              XMM_CLOBBERS("%xmm0", "%xmm1", "%xmm2", "%xmm3", "%xmm4", "%xmm5", "%xmm6", "%xmm7", )\
 676              "%"FF_REG_c"","%"FF_REG_S"","%"FF_REG_D"","%"FF_REG_d"");
 677
 678 #define snow_inner_add_yblock_sse2_end_8\
 679              "sal $1, %%"FF_REG_c"                \n\t"\
 680              "add"FF_OPSIZE" $"FF_PTR_SIZE"*2, %1 \n\t"\
 681              snow_inner_add_yblock_sse2_end_common1\
 682              "sar $1, %%"FF_REG_c"           \n\t"\
 683              "sub $2, %2                     \n\t"\
 684              snow_inner_add_yblock_sse2_end_common2
 685
 686 #define snow_inner_add_yblock_sse2_end_16\
 687              "add"FF_OPSIZE" $"FF_PTR_SIZE"*1, %1 \n\t"\
 688              snow_inner_add_yblock_sse2_end_common1\
 689              "dec %2                         \n\t"\
 690              snow_inner_add_yblock_sse2_end_common2
 691
 692 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,
 693                       int src_x, int src_y, x86_reg src_stride, slice_buffer * sb, int add, uint8_t * dst8){
 694 snow_inner_add_yblock_sse2_header
 695 snow_inner_add_yblock_sse2_start_8("xmm1", "xmm5", "3", "0")
 696 snow_inner_add_yblock_sse2_accum_8("2", "8")
 697 snow_inner_add_yblock_sse2_accum_8("1", "128")
 698 snow_inner_add_yblock_sse2_accum_8("0", "136")
 699
 700              "mov %0, %%"FF_REG_d"           \n\t"
 701              "movdqa (%%"FF_REG_D"), %%xmm0  \n\t"
 702              "movdqa %%xmm1, %%xmm2          \n\t"
 703
 704              "punpckhwd %%xmm7, %%xmm1       \n\t"
 705              "punpcklwd %%xmm7, %%xmm2       \n\t"
 706              "paddd %%xmm2, %%xmm0           \n\t"
 707              "movdqa 16(%%"FF_REG_D"), %%xmm2\n\t"
 708              "paddd %%xmm1, %%xmm2           \n\t"
 709              "paddd %%xmm3, %%xmm0           \n\t"
 710              "paddd %%xmm3, %%xmm2           \n\t"
 711
 712              "mov %1, %%"FF_REG_D"           \n\t"
 713              "mov "FF_PTR_SIZE"(%%"FF_REG_D"), %%"FF_REG_D"; \n\t"
 714              "add %3, %%"FF_REG_D"           \n\t"
 715
 716              "movdqa (%%"FF_REG_D"), %%xmm4  \n\t"
 717              "movdqa %%xmm5, %%xmm6          \n\t"
 718              "punpckhwd %%xmm7, %%xmm5       \n\t"
 719              "punpcklwd %%xmm7, %%xmm6       \n\t"
 720              "paddd %%xmm6, %%xmm4           \n\t"
 721              "movdqa 16(%%"FF_REG_D"), %%xmm6\n\t"
 722              "paddd %%xmm5, %%xmm6           \n\t"
 723              "paddd %%xmm3, %%xmm4           \n\t"
 724              "paddd %%xmm3, %%xmm6           \n\t"
 725
 726              "psrad $8, %%xmm0               \n\t" /* FRAC_BITS. */
 727              "psrad $8, %%xmm2               \n\t" /* FRAC_BITS. */
 728              "packssdw %%xmm2, %%xmm0        \n\t"
 729              "packuswb %%xmm7, %%xmm0        \n\t"
 730              "movq %%xmm0, (%%"FF_REG_d")    \n\t"
 731
 732              "psrad $8, %%xmm4               \n\t" /* FRAC_BITS. */
 733              "psrad $8, %%xmm6               \n\t" /* FRAC_BITS. */
 734              "packssdw %%xmm6, %%xmm4        \n\t"
 735              "packuswb %%xmm7, %%xmm4        \n\t"
 736              "movq %%xmm4, (%%"FF_REG_d",%%"FF_REG_c"); \n\t"
 737 snow_inner_add_yblock_sse2_end_8
 738 }
 739
 740 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,
 741                       int src_x, int src_y, x86_reg src_stride, slice_buffer * sb, int add, uint8_t * dst8){
 742 snow_inner_add_yblock_sse2_header
 743 snow_inner_add_yblock_sse2_start_16("xmm1", "xmm5", "3", "0")
 744 snow_inner_add_yblock_sse2_accum_16("2", "16")
 745 snow_inner_add_yblock_sse2_accum_16("1", "512")
 746 snow_inner_add_yblock_sse2_accum_16("0", "528")
 747
 748              "mov %0, %%"FF_REG_d"           \n\t"
 749              "psrlw $4, %%xmm1               \n\t"
 750              "psrlw $4, %%xmm5               \n\t"
 751              "paddw   (%%"FF_REG_D"), %%xmm1 \n\t"
 752              "paddw 16(%%"FF_REG_D"), %%xmm5 \n\t"
 753              "paddw %%xmm3, %%xmm1           \n\t"
 754              "paddw %%xmm3, %%xmm5           \n\t"
 755              "psraw $4, %%xmm1               \n\t" /* FRAC_BITS. */
 756              "psraw $4, %%xmm5               \n\t" /* FRAC_BITS. */
 757              "packuswb %%xmm5, %%xmm1        \n\t"
 758
 759              "movdqu %%xmm1, (%%"FF_REG_d")  \n\t"
 760
 761 snow_inner_add_yblock_sse2_end_16
 762 }
 763
 764 #define snow_inner_add_yblock_mmx_header \
 765     IDWTELEM * * dst_array = sb->line + src_y;\
 766     x86_reg tmp;\
 767     __asm__ volatile(\
 768              "mov  %7, %%"FF_REG_c"          \n\t"\
 769              "mov  %6, %2                    \n\t"\
 770              "mov  %4, %%"FF_REG_S"          \n\t"\
 771              "pxor %%mm7, %%mm7              \n\t" /* 0 */\
 772              "pcmpeqd %%mm3, %%mm3           \n\t"\
 773              "psllw $15, %%mm3               \n\t"\
 774              "psrlw $12, %%mm3               \n\t" /* FRAC_BITS >> 1 */\
 775              "1:                             \n\t"\
 776              "mov %1, %%"FF_REG_D"           \n\t"\
 777              "mov (%%"FF_REG_D"), %%"FF_REG_D" \n\t"\
 778              "add %3, %%"FF_REG_D"           \n\t"
 779
 780 #define snow_inner_add_yblock_mmx_start(out_reg1, out_reg2, ptr_offset, s_offset, d_offset)\
 781              "mov "FF_PTR_SIZE"*"ptr_offset"(%%"FF_REG_a"), %%"FF_REG_d"; \n\t"\
 782              "movd "d_offset"(%%"FF_REG_d"), %%"out_reg1"                 \n\t"\
 783              "movd "d_offset"+4(%%"FF_REG_d"), %%"out_reg2"               \n\t"\
 784              "punpcklbw %%mm7, %%"out_reg1" \n\t"\
 785              "punpcklbw %%mm7, %%"out_reg2" \n\t"\
 786              "movd "s_offset"(%%"FF_REG_S"), %%mm0   \n\t"\
 787              "movd "s_offset"+4(%%"FF_REG_S"), %%mm4 \n\t"\
 788              "punpcklbw %%mm7, %%mm0       \n\t"\
 789              "punpcklbw %%mm7, %%mm4       \n\t"\
 790              "pmullw %%mm0, %%"out_reg1"   \n\t"\
 791              "pmullw %%mm4, %%"out_reg2"   \n\t"
 792
 793 #define snow_inner_add_yblock_mmx_accum(ptr_offset, s_offset, d_offset) \
 794              snow_inner_add_yblock_mmx_start("mm2", "mm6", ptr_offset, s_offset, d_offset)\
 795              "paddusw %%mm2, %%mm1         \n\t"\
 796              "paddusw %%mm6, %%mm5         \n\t"
 797
 798 #define snow_inner_add_yblock_mmx_mix(read_offset, write_offset)\
 799              "mov %0, %%"FF_REG_d"           \n\t"\
 800              "psrlw $4, %%mm1                \n\t"\
 801              "psrlw $4, %%mm5                \n\t"\
 802              "paddw "read_offset"(%%"FF_REG_D"), %%mm1   \n\t"\
 803              "paddw "read_offset"+8(%%"FF_REG_D"), %%mm5 \n\t"\
 804              "paddw %%mm3, %%mm1             \n\t"\
 805              "paddw %%mm3, %%mm5             \n\t"\
 806              "psraw $4, %%mm1                \n\t"\
 807              "psraw $4, %%mm5                \n\t"\
 808              "packuswb %%mm5, %%mm1          \n\t"\
 809              "movq %%mm1, "write_offset"(%%"FF_REG_d") \n\t"
 810
 811 #define snow_inner_add_yblock_mmx_end(s_step)\
 812              "add $"s_step", %%"FF_REG_S"                      \n\t"\
 813              "add %%"FF_REG_c", "FF_PTR_SIZE"*3(%%"FF_REG_a"); \n\t"\
 814              "add %%"FF_REG_c", "FF_PTR_SIZE"*2(%%"FF_REG_a"); \n\t"\
 815              "add %%"FF_REG_c", "FF_PTR_SIZE"*1(%%"FF_REG_a"); \n\t"\
 816              "add %%"FF_REG_c", (%%"FF_REG_a")                 \n\t"\
 817              "add"FF_OPSIZE " $"FF_PTR_SIZE"*1, %1             \n\t"\
 818              "add %%"FF_REG_c", %0                             \n\t"\
 819              "dec %2                         \n\t"\
 820              "jnz 1b                         \n\t"\
 821              :"+m"(dst8),"+m"(dst_array),"=&r"(tmp)\
 822              :\
 823              "rm"((x86_reg)(src_x<<1)),"m"(obmc),"a"(block),"m"(b_h),"m"(src_stride):\
 824              "%"FF_REG_c"","%"FF_REG_S"","%"FF_REG_D"","%"FF_REG_d"");
 825
 826 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,
 827                       int src_x, int src_y, x86_reg src_stride, slice_buffer * sb, int add, uint8_t * dst8){
 828 snow_inner_add_yblock_mmx_header
 829 snow_inner_add_yblock_mmx_start("mm1", "mm5", "3", "0", "0")
 830 snow_inner_add_yblock_mmx_accum("2", "8", "0")
 831 snow_inner_add_yblock_mmx_accum("1", "128", "0")
 832 snow_inner_add_yblock_mmx_accum("0", "136", "0")
 833 snow_inner_add_yblock_mmx_mix("0", "0")
 834 snow_inner_add_yblock_mmx_end("16")
 835 }
 836
 837 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,
 838                       int src_x, int src_y, x86_reg src_stride, slice_buffer * sb, int add, uint8_t * dst8){
 839 snow_inner_add_yblock_mmx_header
 840 snow_inner_add_yblock_mmx_start("mm1", "mm5", "3", "0", "0")
 841 snow_inner_add_yblock_mmx_accum("2", "16", "0")
 842 snow_inner_add_yblock_mmx_accum("1", "512", "0")
 843 snow_inner_add_yblock_mmx_accum("0", "528", "0")
 844 snow_inner_add_yblock_mmx_mix("0", "0")
 845
 846 snow_inner_add_yblock_mmx_start("mm1", "mm5", "3", "8", "8")
 847 snow_inner_add_yblock_mmx_accum("2", "24", "8")
 848 snow_inner_add_yblock_mmx_accum("1", "520", "8")
 849 snow_inner_add_yblock_mmx_accum("0", "536", "8")
 850 snow_inner_add_yblock_mmx_mix("16", "8")
 851 snow_inner_add_yblock_mmx_end("32")
 852 }
 853
 854 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,
 855                            int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8){
 856
 857     if (b_w == 16)
 858         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);
 859     else if (b_w == 8 && obmc_stride == 16) {
 860         if (!(b_h & 1))
 861             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);
 862         else
 863             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);
 864     } else
 865          ff_snow_inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
 866 }
 867
 868 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,
 869                           int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8){
 870     if (b_w == 16)
 871         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);
 872     else if (b_w == 8 && obmc_stride == 16)
 873         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);
 874     else
 875         ff_snow_inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
 876 }
 877 #endif /* HAVE_6REGS */
 878
 879 #endif /* HAVE_INLINE_ASM */
 880
 881 av_cold void ff_dwt_init_x86(SnowDWTContext *c)
 882 {
 883 #if HAVE_INLINE_ASM
 884     int mm_flags = av_get_cpu_flags();
 885
 886     if (mm_flags & AV_CPU_FLAG_MMX) {
 887         if(mm_flags & AV_CPU_FLAG_SSE2 & 0){
 888             c->horizontal_compose97i = ff_snow_horizontal_compose97i_sse2;
 889 #if HAVE_7REGS
 890             c->vertical_compose97i = ff_snow_vertical_compose97i_sse2;
 891 #endif
 892 #if HAVE_6REGS
 893             c->inner_add_yblock = ff_snow_inner_add_yblock_sse2;
 894 #endif
 895         }
 896         else{
 897             if (mm_flags & AV_CPU_FLAG_MMXEXT) {
 898             c->horizontal_compose97i = ff_snow_horizontal_compose97i_mmx;
 899 #if HAVE_7REGS
 900             c->vertical_compose97i = ff_snow_vertical_compose97i_mmx;
 901 #endif
 902             }
 903 #if HAVE_6REGS
 904             c->inner_add_yblock = ff_snow_inner_add_yblock_mmx;
 905 #endif
 906         }
 907     }
 908 #endif /* HAVE_INLINE_ASM */
 909 }