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

   1 /*
   2  * MMX optimized DSP utils
   3  * Copyright (c) 2000, 2001 Fabrice Bellard
   4  * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
   5  *
   6  * This file is part of FFmpeg.
   7  *
   8  * FFmpeg is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU Lesser General Public
  10  * License as published by the Free Software Foundation; either
  11  * version 2.1 of the License, or (at your option) any later version.
  12  *
  13  * FFmpeg is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * Lesser General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU Lesser General Public
  19  * License along with FFmpeg; if not, write to the Free Software
  20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  21  *
  22  * MMX optimization by Nick Kurshev <nickols_k@mail.ru>
  23  */
  24
  25 #include "libavutil/cpu.h"
  26 #include "libavutil/x86/asm.h"
  27 #include "libavcodec/dsputil.h"
  28 #include "libavcodec/h264dsp.h"
  29 #include "libavcodec/mpegvideo.h"
  30 #include "libavcodec/simple_idct.h"
  31 #include "dsputil_mmx.h"
  32 #include "idct_xvid.h"
  33 #include "diracdsp_mmx.h"
  34
  35 //#undef NDEBUG
  36 //#include <assert.h>
  37
  38 /* pixel operations */
  39 DECLARE_ALIGNED(8,  const uint64_t, ff_bone) = 0x0101010101010101ULL;
  40 DECLARE_ALIGNED(8,  const uint64_t, ff_wtwo) = 0x0002000200020002ULL;
  41
  42 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_1)    = { 0x0001000100010001ULL, 0x0001000100010001ULL };
  43 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_2)    = { 0x0002000200020002ULL, 0x0002000200020002ULL };
  44 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_3)    = { 0x0003000300030003ULL, 0x0003000300030003ULL };
  45 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_4)    = { 0x0004000400040004ULL, 0x0004000400040004ULL };
  46 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_5)    = { 0x0005000500050005ULL, 0x0005000500050005ULL };
  47 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_8)    = { 0x0008000800080008ULL, 0x0008000800080008ULL };
  48 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_9)    = { 0x0009000900090009ULL, 0x0009000900090009ULL };
  49 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_15)   =   0x000F000F000F000FULL;
  50 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_16)   = { 0x0010001000100010ULL, 0x0010001000100010ULL };
  51 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_17)   = { 0x0011001100110011ULL, 0x0011001100110011ULL };
  52 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_18)   = { 0x0012001200120012ULL, 0x0012001200120012ULL };
  53 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_20)   =   0x0014001400140014ULL;
  54 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_27)   = { 0x001B001B001B001BULL, 0x001B001B001B001BULL };
  55 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_28)   = { 0x001C001C001C001CULL, 0x001C001C001C001CULL };
  56 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_32)   = { 0x0020002000200020ULL, 0x0020002000200020ULL };
  57 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_42)   =   0x002A002A002A002AULL;
  58 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_53)   =   0x0035003500350035ULL;
  59 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_63)   = { 0x003F003F003F003FULL, 0x003F003F003F003FULL };
  60 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_64)   = { 0x0040004000400040ULL, 0x0040004000400040ULL };
  61 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_96)   =   0x0060006000600060ULL;
  62 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_128)  =   0x0080008000800080ULL;
  63 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_255)  =   0x00ff00ff00ff00ffULL;
  64 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_512)  = { 0x0200020002000200ULL, 0x0200020002000200ULL };
  65 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_1019) = { 0x03FB03FB03FB03FBULL, 0x03FB03FB03FB03FBULL };
  66
  67 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_0)    = { 0x0000000000000000ULL, 0x0000000000000000ULL };
  68 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_1)    = { 0x0101010101010101ULL, 0x0101010101010101ULL };
  69 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_3)    = { 0x0303030303030303ULL, 0x0303030303030303ULL };
  70 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_4)    = { 0x0404040404040404ULL, 0x0404040404040404ULL };
  71 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_7)    =   0x0707070707070707ULL;
  72 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_1F)   =   0x1F1F1F1F1F1F1F1FULL;
  73 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_3F)   =   0x3F3F3F3F3F3F3F3FULL;
  74 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_80)   = { 0x8080808080808080ULL, 0x8080808080808080ULL };
  75 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_81)   =   0x8181818181818181ULL;
  76 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_A1)   = { 0xA1A1A1A1A1A1A1A1ULL, 0xA1A1A1A1A1A1A1A1ULL };
  77 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_F8)   = { 0xF8F8F8F8F8F8F8F8ULL, 0xF8F8F8F8F8F8F8F8ULL };
  78 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_FC)   =   0xFCFCFCFCFCFCFCFCULL;
  79 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_FE)   = { 0xFEFEFEFEFEFEFEFEULL, 0xFEFEFEFEFEFEFEFEULL };
  80
  81 DECLARE_ALIGNED(16, const double, ff_pd_1)[2] = { 1.0, 1.0 };
  82 DECLARE_ALIGNED(16, const double, ff_pd_2)[2] = { 2.0, 2.0 };
  83
  84 #if HAVE_INLINE_ASM
  85
  86 #define JUMPALIGN()     __asm__ volatile (".p2align 3"::)
  87 #define MOVQ_ZERO(regd) __asm__ volatile ("pxor %%"#regd", %%"#regd ::)
  88
  89 #define MOVQ_BFE(regd)                                  \
  90     __asm__ volatile (                                  \
  91         "pcmpeqd %%"#regd", %%"#regd"   \n\t"           \
  92         "paddb   %%"#regd", %%"#regd"   \n\t" ::)
  93
  94 #ifndef PIC
  95 #define MOVQ_BONE(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_bone))
  96 #define MOVQ_WTWO(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_wtwo))
  97 #else
  98 // for shared library it's better to use this way for accessing constants
  99 // pcmpeqd -> -1
 100 #define MOVQ_BONE(regd)                                 \
 101     __asm__ volatile (                                  \
 102         "pcmpeqd  %%"#regd", %%"#regd"  \n\t"           \
 103         "psrlw          $15, %%"#regd"  \n\t"           \
 104         "packuswb %%"#regd", %%"#regd"  \n\t" ::)
 105
 106 #define MOVQ_WTWO(regd)                                 \
 107     __asm__ volatile (                                  \
 108         "pcmpeqd %%"#regd", %%"#regd"   \n\t"           \
 109         "psrlw         $15, %%"#regd"   \n\t"           \
 110         "psllw          $1, %%"#regd"   \n\t"::)
 111
 112 #endif
 113
 114 // using regr as temporary and for the output result
 115 // first argument is unmodifed and second is trashed
 116 // regfe is supposed to contain 0xfefefefefefefefe
 117 #define PAVGB_MMX_NO_RND(rega, regb, regr, regfe)                \
 118     "movq   "#rega", "#regr"            \n\t"                    \
 119     "pand   "#regb", "#regr"            \n\t"                    \
 120     "pxor   "#rega", "#regb"            \n\t"                    \
 121     "pand  "#regfe", "#regb"            \n\t"                    \
 122     "psrlq       $1, "#regb"            \n\t"                    \
 123     "paddb  "#regb", "#regr"            \n\t"
 124
 125 #define PAVGB_MMX(rega, regb, regr, regfe)                       \
 126     "movq   "#rega", "#regr"            \n\t"                    \
 127     "por    "#regb", "#regr"            \n\t"                    \
 128     "pxor   "#rega", "#regb"            \n\t"                    \
 129     "pand  "#regfe", "#regb"            \n\t"                    \
 130     "psrlq       $1, "#regb"            \n\t"                    \
 131     "psubb  "#regb", "#regr"            \n\t"
 132
 133 // mm6 is supposed to contain 0xfefefefefefefefe
 134 #define PAVGBP_MMX_NO_RND(rega, regb, regr,  regc, regd, regp)   \
 135     "movq  "#rega", "#regr"             \n\t"                    \
 136     "movq  "#regc", "#regp"             \n\t"                    \
 137     "pand  "#regb", "#regr"             \n\t"                    \
 138     "pand  "#regd", "#regp"             \n\t"                    \
 139     "pxor  "#rega", "#regb"             \n\t"                    \
 140     "pxor  "#regc", "#regd"             \n\t"                    \
 141     "pand    %%mm6, "#regb"             \n\t"                    \
 142     "pand    %%mm6, "#regd"             \n\t"                    \
 143     "psrlq      $1, "#regb"             \n\t"                    \
 144     "psrlq      $1, "#regd"             \n\t"                    \
 145     "paddb "#regb", "#regr"             \n\t"                    \
 146     "paddb "#regd", "#regp"             \n\t"
 147
 148 #define PAVGBP_MMX(rega, regb, regr, regc, regd, regp)           \
 149     "movq  "#rega", "#regr"             \n\t"                    \
 150     "movq  "#regc", "#regp"             \n\t"                    \
 151     "por   "#regb", "#regr"             \n\t"                    \
 152     "por   "#regd", "#regp"             \n\t"                    \
 153     "pxor  "#rega", "#regb"             \n\t"                    \
 154     "pxor  "#regc", "#regd"             \n\t"                    \
 155     "pand    %%mm6, "#regb"             \n\t"                    \
 156     "pand    %%mm6, "#regd"             \n\t"                    \
 157     "psrlq      $1, "#regd"             \n\t"                    \
 158     "psrlq      $1, "#regb"             \n\t"                    \
 159     "psubb "#regb", "#regr"             \n\t"                    \
 160     "psubb "#regd", "#regp"             \n\t"
 161
 162 /***********************************/
 163 /* MMX no rounding */
 164 #define DEF(x, y) x ## _no_rnd_ ## y ## _mmx
 165 #define SET_RND  MOVQ_WONE
 166 #define PAVGBP(a, b, c, d, e, f)        PAVGBP_MMX_NO_RND(a, b, c, d, e, f)
 167 #define PAVGB(a, b, c, e)               PAVGB_MMX_NO_RND(a, b, c, e)
 168 #define OP_AVG(a, b, c, e)              PAVGB_MMX(a, b, c, e)
 169
 170 #include "dsputil_rnd_template.c"
 171
 172 #undef DEF
 173 #undef SET_RND
 174 #undef PAVGBP
 175 #undef PAVGB
 176 /***********************************/
 177 /* MMX rounding */
 178
 179 #define DEF(x, y) x ## _ ## y ## _mmx
 180 #define SET_RND  MOVQ_WTWO
 181 #define PAVGBP(a, b, c, d, e, f)        PAVGBP_MMX(a, b, c, d, e, f)
 182 #define PAVGB(a, b, c, e)               PAVGB_MMX(a, b, c, e)
 183
 184 #include "dsputil_rnd_template.c"
 185
 186 #undef DEF
 187 #undef SET_RND
 188 #undef PAVGBP
 189 #undef PAVGB
 190 #undef OP_AVG
 191
 192 /***********************************/
 193 /* 3Dnow specific */
 194
 195 #define DEF(x) x ## _3dnow
 196 #define PAVGB "pavgusb"
 197 #define SKIP_FOR_3DNOW
 198
 199 #include "dsputil_avg_template.c"
 200
 201 #undef DEF
 202 #undef PAVGB
 203 #undef SKIP_FOR_3DNOW
 204
 205 /***********************************/
 206 /* MMXEXT specific */
 207
 208 #define DEF(x) x ## _mmxext
 209
 210 /* Introduced only in MMXEXT set */
 211 #define PAVGB "pavgb"
 212
 213 #include "dsputil_avg_template.c"
 214
 215 #undef DEF
 216 #undef PAVGB
 217
 218 #define put_no_rnd_pixels16_mmx put_pixels16_mmx
 219 #define put_no_rnd_pixels8_mmx put_pixels8_mmx
 220 #define put_pixels16_mmxext put_pixels16_mmx
 221 #define put_pixels8_mmxext put_pixels8_mmx
 222 #define put_pixels4_mmxext put_pixels4_mmx
 223 #define put_no_rnd_pixels16_mmxext put_no_rnd_pixels16_mmx
 224 #define put_no_rnd_pixels8_mmxext put_no_rnd_pixels8_mmx
 225
 226 /***********************************/
 227 /* standard MMX */
 228
 229 void ff_put_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
 230                                int line_size)
 231 {
 232     const int16_t *p;
 233     uint8_t *pix;
 234
 235     /* read the pixels */
 236     p   = block;
 237     pix = pixels;
 238     /* unrolled loop */
 239     __asm__ volatile (
 240         "movq      (%3), %%mm0          \n\t"
 241         "movq     8(%3), %%mm1          \n\t"
 242         "movq    16(%3), %%mm2          \n\t"
 243         "movq    24(%3), %%mm3          \n\t"
 244         "movq    32(%3), %%mm4          \n\t"
 245         "movq    40(%3), %%mm5          \n\t"
 246         "movq    48(%3), %%mm6          \n\t"
 247         "movq    56(%3), %%mm7          \n\t"
 248         "packuswb %%mm1, %%mm0          \n\t"
 249         "packuswb %%mm3, %%mm2          \n\t"
 250         "packuswb %%mm5, %%mm4          \n\t"
 251         "packuswb %%mm7, %%mm6          \n\t"
 252         "movq     %%mm0, (%0)           \n\t"
 253         "movq     %%mm2, (%0, %1)       \n\t"
 254         "movq     %%mm4, (%0, %1, 2)    \n\t"
 255         "movq     %%mm6, (%0, %2)       \n\t"
 256         :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3),
 257            "r"(p)
 258         : "memory");
 259     pix += line_size * 4;
 260     p   += 32;
 261
 262     // if here would be an exact copy of the code above
 263     // compiler would generate some very strange code
 264     // thus using "r"
 265     __asm__ volatile (
 266         "movq       (%3), %%mm0         \n\t"
 267         "movq      8(%3), %%mm1         \n\t"
 268         "movq     16(%3), %%mm2         \n\t"
 269         "movq     24(%3), %%mm3         \n\t"
 270         "movq     32(%3), %%mm4         \n\t"
 271         "movq     40(%3), %%mm5         \n\t"
 272         "movq     48(%3), %%mm6         \n\t"
 273         "movq     56(%3), %%mm7         \n\t"
 274         "packuswb  %%mm1, %%mm0         \n\t"
 275         "packuswb  %%mm3, %%mm2         \n\t"
 276         "packuswb  %%mm5, %%mm4         \n\t"
 277         "packuswb  %%mm7, %%mm6         \n\t"
 278         "movq      %%mm0, (%0)          \n\t"
 279         "movq      %%mm2, (%0, %1)      \n\t"
 280         "movq      %%mm4, (%0, %1, 2)   \n\t"
 281         "movq      %%mm6, (%0, %2)      \n\t"
 282         :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3), "r"(p)
 283         : "memory");
 284 }
 285
 286 #define put_signed_pixels_clamped_mmx_half(off)             \
 287     "movq          "#off"(%2), %%mm1        \n\t"           \
 288     "movq     16 + "#off"(%2), %%mm2        \n\t"           \
 289     "movq     32 + "#off"(%2), %%mm3        \n\t"           \
 290     "movq     48 + "#off"(%2), %%mm4        \n\t"           \
 291     "packsswb  8 + "#off"(%2), %%mm1        \n\t"           \
 292     "packsswb 24 + "#off"(%2), %%mm2        \n\t"           \
 293     "packsswb 40 + "#off"(%2), %%mm3        \n\t"           \
 294     "packsswb 56 + "#off"(%2), %%mm4        \n\t"           \
 295     "paddb              %%mm0, %%mm1        \n\t"           \
 296     "paddb              %%mm0, %%mm2        \n\t"           \
 297     "paddb              %%mm0, %%mm3        \n\t"           \
 298     "paddb              %%mm0, %%mm4        \n\t"           \
 299     "movq               %%mm1, (%0)         \n\t"           \
 300     "movq               %%mm2, (%0, %3)     \n\t"           \
 301     "movq               %%mm3, (%0, %3, 2)  \n\t"           \
 302     "movq               %%mm4, (%0, %1)     \n\t"
 303
 304 void ff_put_signed_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
 305                                       int line_size)
 306 {
 307     x86_reg line_skip = line_size;
 308     x86_reg line_skip3;
 309
 310     __asm__ volatile (
 311         "movq "MANGLE(ff_pb_80)", %%mm0     \n\t"
 312         "lea         (%3, %3, 2), %1        \n\t"
 313         put_signed_pixels_clamped_mmx_half(0)
 314         "lea         (%0, %3, 4), %0        \n\t"
 315         put_signed_pixels_clamped_mmx_half(64)
 316         : "+&r"(pixels), "=&r"(line_skip3)
 317         : "r"(block), "r"(line_skip)
 318         : "memory");
 319 }
 320
 321 void ff_add_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
 322                                int line_size)
 323 {
 324     const int16_t *p;
 325     uint8_t *pix;
 326     int i;
 327
 328     /* read the pixels */
 329     p   = block;
 330     pix = pixels;
 331     MOVQ_ZERO(mm7);
 332     i = 4;
 333     do {
 334         __asm__ volatile (
 335             "movq        (%2), %%mm0    \n\t"
 336             "movq       8(%2), %%mm1    \n\t"
 337             "movq      16(%2), %%mm2    \n\t"
 338             "movq      24(%2), %%mm3    \n\t"
 339             "movq          %0, %%mm4    \n\t"
 340             "movq          %1, %%mm6    \n\t"
 341             "movq       %%mm4, %%mm5    \n\t"
 342             "punpcklbw  %%mm7, %%mm4    \n\t"
 343             "punpckhbw  %%mm7, %%mm5    \n\t"
 344             "paddsw     %%mm4, %%mm0    \n\t"
 345             "paddsw     %%mm5, %%mm1    \n\t"
 346             "movq       %%mm6, %%mm5    \n\t"
 347             "punpcklbw  %%mm7, %%mm6    \n\t"
 348             "punpckhbw  %%mm7, %%mm5    \n\t"
 349             "paddsw     %%mm6, %%mm2    \n\t"
 350             "paddsw     %%mm5, %%mm3    \n\t"
 351             "packuswb   %%mm1, %%mm0    \n\t"
 352             "packuswb   %%mm3, %%mm2    \n\t"
 353             "movq       %%mm0, %0       \n\t"
 354             "movq       %%mm2, %1       \n\t"
 355             : "+m"(*pix), "+m"(*(pix + line_size))
 356             : "r"(p)
 357             : "memory");
 358         pix += line_size * 2;
 359         p   += 16;
 360     } while (--i);
 361 }
 362
 363 static void put_pixels8_mmx(uint8_t *block, const uint8_t *pixels,
 364                             int line_size, int h)
 365 {
 366     __asm__ volatile (
 367         "lea   (%3, %3), %%"REG_a"      \n\t"
 368         ".p2align     3                 \n\t"
 369         "1:                             \n\t"
 370         "movq  (%1    ), %%mm0          \n\t"
 371         "movq  (%1, %3), %%mm1          \n\t"
 372         "movq     %%mm0, (%2)           \n\t"
 373         "movq     %%mm1, (%2, %3)       \n\t"
 374         "add  %%"REG_a", %1             \n\t"
 375         "add  %%"REG_a", %2             \n\t"
 376         "movq  (%1    ), %%mm0          \n\t"
 377         "movq  (%1, %3), %%mm1          \n\t"
 378         "movq     %%mm0, (%2)           \n\t"
 379         "movq     %%mm1, (%2, %3)       \n\t"
 380         "add  %%"REG_a", %1             \n\t"
 381         "add  %%"REG_a", %2             \n\t"
 382         "subl        $4, %0             \n\t"
 383         "jnz         1b                 \n\t"
 384         : "+g"(h), "+r"(pixels),  "+r"(block)
 385         : "r"((x86_reg)line_size)
 386         : "%"REG_a, "memory"
 387         );
 388 }
 389
 390 static void put_pixels16_mmx(uint8_t *block, const uint8_t *pixels,
 391                              int line_size, int h)
 392 {
 393     __asm__ volatile (
 394         "lea   (%3, %3), %%"REG_a"      \n\t"
 395         ".p2align     3                 \n\t"
 396         "1:                             \n\t"
 397         "movq  (%1    ), %%mm0          \n\t"
 398         "movq 8(%1    ), %%mm4          \n\t"
 399         "movq  (%1, %3), %%mm1          \n\t"
 400         "movq 8(%1, %3), %%mm5          \n\t"
 401         "movq     %%mm0,  (%2)          \n\t"
 402         "movq     %%mm4, 8(%2)          \n\t"
 403         "movq     %%mm1,  (%2, %3)      \n\t"
 404         "movq     %%mm5, 8(%2, %3)      \n\t"
 405         "add  %%"REG_a", %1             \n\t"
 406         "add  %%"REG_a", %2             \n\t"
 407         "movq  (%1    ), %%mm0          \n\t"
 408         "movq 8(%1    ), %%mm4          \n\t"
 409         "movq  (%1, %3), %%mm1          \n\t"
 410         "movq 8(%1, %3), %%mm5          \n\t"
 411         "movq     %%mm0,  (%2)          \n\t"
 412         "movq     %%mm4, 8(%2)          \n\t"
 413         "movq     %%mm1,  (%2, %3)      \n\t"
 414         "movq     %%mm5, 8(%2, %3)      \n\t"
 415         "add  %%"REG_a", %1             \n\t"
 416         "add  %%"REG_a", %2             \n\t"
 417         "subl        $4, %0             \n\t"
 418         "jnz         1b                 \n\t"
 419         : "+g"(h), "+r"(pixels),  "+r"(block)
 420         : "r"((x86_reg)line_size)
 421         : "%"REG_a, "memory"
 422         );
 423 }
 424
 425 #define CLEAR_BLOCKS(name, n)                           \
 426 static void name(int16_t *blocks)                       \
 427 {                                                       \
 428     __asm__ volatile (                                  \
 429         "pxor %%mm7, %%mm7              \n\t"           \
 430         "mov     %1,        %%"REG_a"   \n\t"           \
 431         "1:                             \n\t"           \
 432         "movq %%mm7,   (%0, %%"REG_a")  \n\t"           \
 433         "movq %%mm7,  8(%0, %%"REG_a")  \n\t"           \
 434         "movq %%mm7, 16(%0, %%"REG_a")  \n\t"           \
 435         "movq %%mm7, 24(%0, %%"REG_a")  \n\t"           \
 436         "add    $32, %%"REG_a"          \n\t"           \
 437         "js      1b                     \n\t"           \
 438         :: "r"(((uint8_t *)blocks) + 128 * n),          \
 439            "i"(-128 * n)                                \
 440         : "%"REG_a                                      \
 441         );                                              \
 442 }
 443 CLEAR_BLOCKS(clear_blocks_mmx, 6)
 444 CLEAR_BLOCKS(clear_block_mmx, 1)
 445
 446 static void clear_block_sse(int16_t *block)
 447 {
 448     __asm__ volatile (
 449         "xorps  %%xmm0, %%xmm0          \n"
 450         "movaps %%xmm0,    (%0)         \n"
 451         "movaps %%xmm0,  16(%0)         \n"
 452         "movaps %%xmm0,  32(%0)         \n"
 453         "movaps %%xmm0,  48(%0)         \n"
 454         "movaps %%xmm0,  64(%0)         \n"
 455         "movaps %%xmm0,  80(%0)         \n"
 456         "movaps %%xmm0,  96(%0)         \n"
 457         "movaps %%xmm0, 112(%0)         \n"
 458         :: "r"(block)
 459         : "memory"
 460     );
 461 }
 462
 463 static void clear_blocks_sse(int16_t *blocks)
 464 {
 465     __asm__ volatile (
 466         "xorps  %%xmm0, %%xmm0              \n"
 467         "mov        %1,         %%"REG_a"   \n"
 468         "1:                                 \n"
 469         "movaps %%xmm0,    (%0, %%"REG_a")  \n"
 470         "movaps %%xmm0,  16(%0, %%"REG_a")  \n"
 471         "movaps %%xmm0,  32(%0, %%"REG_a")  \n"
 472         "movaps %%xmm0,  48(%0, %%"REG_a")  \n"
 473         "movaps %%xmm0,  64(%0, %%"REG_a")  \n"
 474         "movaps %%xmm0,  80(%0, %%"REG_a")  \n"
 475         "movaps %%xmm0,  96(%0, %%"REG_a")  \n"
 476         "movaps %%xmm0, 112(%0, %%"REG_a")  \n"
 477         "add      $128,         %%"REG_a"   \n"
 478         "js         1b                      \n"
 479         :: "r"(((uint8_t *)blocks) + 128 * 6),
 480            "i"(-128 * 6)
 481         : "%"REG_a
 482     );
 483 }
 484
 485 static void add_bytes_mmx(uint8_t *dst, uint8_t *src, int w)
 486 {
 487     x86_reg i = 0;
 488     __asm__ volatile (
 489         "jmp          2f                \n\t"
 490         "1:                             \n\t"
 491         "movq   (%1, %0), %%mm0         \n\t"
 492         "movq   (%2, %0), %%mm1         \n\t"
 493         "paddb     %%mm0, %%mm1         \n\t"
 494         "movq      %%mm1, (%2, %0)      \n\t"
 495         "movq  8(%1, %0), %%mm0         \n\t"
 496         "movq  8(%2, %0), %%mm1         \n\t"
 497         "paddb     %%mm0, %%mm1         \n\t"
 498         "movq      %%mm1, 8(%2, %0)     \n\t"
 499         "add         $16, %0            \n\t"
 500         "2:                             \n\t"
 501         "cmp          %3, %0            \n\t"
 502         "js           1b                \n\t"
 503         : "+r"(i)
 504         : "r"(src), "r"(dst), "r"((x86_reg)w - 15)
 505     );
 506     for ( ; i < w; i++)
 507         dst[i + 0] += src[i + 0];
 508 }
 509
 510 #if HAVE_7REGS
 511 static void add_hfyu_median_prediction_cmov(uint8_t *dst, const uint8_t *top,
 512                                             const uint8_t *diff, int w,
 513                                             int *left, int *left_top)
 514 {
 515     x86_reg w2 = -w;
 516     x86_reg x;
 517     int l  = *left     & 0xff;
 518     int tl = *left_top & 0xff;
 519     int t;
 520     __asm__ volatile (
 521         "mov          %7, %3            \n"
 522         "1:                             \n"
 523         "movzbl (%3, %4), %2            \n"
 524         "mov          %2, %k3           \n"
 525         "sub         %b1, %b3           \n"
 526         "add         %b0, %b3           \n"
 527         "mov          %2, %1            \n"
 528         "cmp          %0, %2            \n"
 529         "cmovg        %0, %2            \n"
 530         "cmovg        %1, %0            \n"
 531         "cmp         %k3, %0            \n"
 532         "cmovg       %k3, %0            \n"
 533         "mov          %7, %3            \n"
 534         "cmp          %2, %0            \n"
 535         "cmovl        %2, %0            \n"
 536         "add    (%6, %4), %b0           \n"
 537         "mov         %b0, (%5, %4)      \n"
 538         "inc          %4                \n"
 539         "jl           1b                \n"
 540         : "+&q"(l), "+&q"(tl), "=&r"(t), "=&q"(x), "+&r"(w2)
 541         : "r"(dst + w), "r"(diff + w), "rm"(top + w)
 542     );
 543     *left     = l;
 544     *left_top = tl;
 545 }
 546 #endif
 547
 548 static inline void transpose4x4(uint8_t *dst, uint8_t *src, x86_reg dst_stride, x86_reg src_stride){
 549     __asm__ volatile( //FIXME could save 1 instruction if done as 8x4 ...
 550         "movd  (%1), %%mm0              \n\t"
 551         "add   %3, %1                   \n\t"
 552         "movd  (%1), %%mm1              \n\t"
 553         "movd  (%1,%3,1), %%mm2         \n\t"
 554         "movd  (%1,%3,2), %%mm3         \n\t"
 555         "punpcklbw %%mm1, %%mm0         \n\t"
 556         "punpcklbw %%mm3, %%mm2         \n\t"
 557         "movq %%mm0, %%mm1              \n\t"
 558         "punpcklwd %%mm2, %%mm0         \n\t"
 559         "punpckhwd %%mm2, %%mm1         \n\t"
 560         "movd  %%mm0, (%0)              \n\t"
 561         "add   %2, %0                   \n\t"
 562         "punpckhdq %%mm0, %%mm0         \n\t"
 563         "movd  %%mm0, (%0)              \n\t"
 564         "movd  %%mm1, (%0,%2,1)         \n\t"
 565         "punpckhdq %%mm1, %%mm1         \n\t"
 566         "movd  %%mm1, (%0,%2,2)         \n\t"
 567
 568         :  "+&r" (dst),
 569            "+&r" (src)
 570         :  "r" (dst_stride),
 571            "r" (src_stride)
 572         :  "memory"
 573     );
 574 }
 575
 576 #define H263_LOOP_FILTER                        \
 577     "pxor      %%mm7, %%mm7             \n\t"   \
 578     "movq         %0, %%mm0             \n\t"   \
 579     "movq         %0, %%mm1             \n\t"   \
 580     "movq         %3, %%mm2             \n\t"   \
 581     "movq         %3, %%mm3             \n\t"   \
 582     "punpcklbw %%mm7, %%mm0             \n\t"   \
 583     "punpckhbw %%mm7, %%mm1             \n\t"   \
 584     "punpcklbw %%mm7, %%mm2             \n\t"   \
 585     "punpckhbw %%mm7, %%mm3             \n\t"   \
 586     "psubw     %%mm2, %%mm0             \n\t"   \
 587     "psubw     %%mm3, %%mm1             \n\t"   \
 588     "movq         %1, %%mm2             \n\t"   \
 589     "movq         %1, %%mm3             \n\t"   \
 590     "movq         %2, %%mm4             \n\t"   \
 591     "movq         %2, %%mm5             \n\t"   \
 592     "punpcklbw %%mm7, %%mm2             \n\t"   \
 593     "punpckhbw %%mm7, %%mm3             \n\t"   \
 594     "punpcklbw %%mm7, %%mm4             \n\t"   \
 595     "punpckhbw %%mm7, %%mm5             \n\t"   \
 596     "psubw     %%mm2, %%mm4             \n\t"   \
 597     "psubw     %%mm3, %%mm5             \n\t"   \
 598     "psllw        $2, %%mm4             \n\t"   \
 599     "psllw        $2, %%mm5             \n\t"   \
 600     "paddw     %%mm0, %%mm4             \n\t"   \
 601     "paddw     %%mm1, %%mm5             \n\t"   \
 602     "pxor      %%mm6, %%mm6             \n\t"   \
 603     "pcmpgtw   %%mm4, %%mm6             \n\t"   \
 604     "pcmpgtw   %%mm5, %%mm7             \n\t"   \
 605     "pxor      %%mm6, %%mm4             \n\t"   \
 606     "pxor      %%mm7, %%mm5             \n\t"   \
 607     "psubw     %%mm6, %%mm4             \n\t"   \
 608     "psubw     %%mm7, %%mm5             \n\t"   \
 609     "psrlw        $3, %%mm4             \n\t"   \
 610     "psrlw        $3, %%mm5             \n\t"   \
 611     "packuswb  %%mm5, %%mm4             \n\t"   \
 612     "packsswb  %%mm7, %%mm6             \n\t"   \
 613     "pxor      %%mm7, %%mm7             \n\t"   \
 614     "movd         %4, %%mm2             \n\t"   \
 615     "punpcklbw %%mm2, %%mm2             \n\t"   \
 616     "punpcklbw %%mm2, %%mm2             \n\t"   \
 617     "punpcklbw %%mm2, %%mm2             \n\t"   \
 618     "psubusb   %%mm4, %%mm2             \n\t"   \
 619     "movq      %%mm2, %%mm3             \n\t"   \
 620     "psubusb   %%mm4, %%mm3             \n\t"   \
 621     "psubb     %%mm3, %%mm2             \n\t"   \
 622     "movq         %1, %%mm3             \n\t"   \
 623     "movq         %2, %%mm4             \n\t"   \
 624     "pxor      %%mm6, %%mm3             \n\t"   \
 625     "pxor      %%mm6, %%mm4             \n\t"   \
 626     "paddusb   %%mm2, %%mm3             \n\t"   \
 627     "psubusb   %%mm2, %%mm4             \n\t"   \
 628     "pxor      %%mm6, %%mm3             \n\t"   \
 629     "pxor      %%mm6, %%mm4             \n\t"   \
 630     "paddusb   %%mm2, %%mm2             \n\t"   \
 631     "packsswb  %%mm1, %%mm0             \n\t"   \
 632     "pcmpgtb   %%mm0, %%mm7             \n\t"   \
 633     "pxor      %%mm7, %%mm0             \n\t"   \
 634     "psubb     %%mm7, %%mm0             \n\t"   \
 635     "movq      %%mm0, %%mm1             \n\t"   \
 636     "psubusb   %%mm2, %%mm0             \n\t"   \
 637     "psubb     %%mm0, %%mm1             \n\t"   \
 638     "pand         %5, %%mm1             \n\t"   \
 639     "psrlw        $2, %%mm1             \n\t"   \
 640     "pxor      %%mm7, %%mm1             \n\t"   \
 641     "psubb     %%mm7, %%mm1             \n\t"   \
 642     "movq         %0, %%mm5             \n\t"   \
 643     "movq         %3, %%mm6             \n\t"   \
 644     "psubb     %%mm1, %%mm5             \n\t"   \
 645     "paddb     %%mm1, %%mm6             \n\t"
 646
 647 static void h263_v_loop_filter_mmx(uint8_t *src, int stride, int qscale)
 648 {
 649     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
 650         const int strength = ff_h263_loop_filter_strength[qscale];
 651
 652         __asm__ volatile (
 653             H263_LOOP_FILTER
 654
 655             "movq %%mm3, %1             \n\t"
 656             "movq %%mm4, %2             \n\t"
 657             "movq %%mm5, %0             \n\t"
 658             "movq %%mm6, %3             \n\t"
 659             : "+m"(*(uint64_t*)(src - 2 * stride)),
 660               "+m"(*(uint64_t*)(src - 1 * stride)),
 661               "+m"(*(uint64_t*)(src + 0 * stride)),
 662               "+m"(*(uint64_t*)(src + 1 * stride))
 663             : "g"(2 * strength), "m"(ff_pb_FC)
 664             );
 665     }
 666 }
 667
 668 static void h263_h_loop_filter_mmx(uint8_t *src, int stride, int qscale)
 669 {
 670     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
 671         const int strength = ff_h263_loop_filter_strength[qscale];
 672         DECLARE_ALIGNED(8, uint64_t, temp)[4];
 673         uint8_t *btemp = (uint8_t*)temp;
 674
 675         src -= 2;
 676
 677         transpose4x4(btemp,     src,              8, stride);
 678         transpose4x4(btemp + 4, src + 4 * stride, 8, stride);
 679         __asm__ volatile (
 680             H263_LOOP_FILTER // 5 3 4 6
 681
 682             : "+m"(temp[0]),
 683               "+m"(temp[1]),
 684               "+m"(temp[2]),
 685               "+m"(temp[3])
 686             : "g"(2 * strength), "m"(ff_pb_FC)
 687             );
 688
 689         __asm__ volatile (
 690             "movq      %%mm5, %%mm1         \n\t"
 691             "movq      %%mm4, %%mm0         \n\t"
 692             "punpcklbw %%mm3, %%mm5         \n\t"
 693             "punpcklbw %%mm6, %%mm4         \n\t"
 694             "punpckhbw %%mm3, %%mm1         \n\t"
 695             "punpckhbw %%mm6, %%mm0         \n\t"
 696             "movq      %%mm5, %%mm3         \n\t"
 697             "movq      %%mm1, %%mm6         \n\t"
 698             "punpcklwd %%mm4, %%mm5         \n\t"
 699             "punpcklwd %%mm0, %%mm1         \n\t"
 700             "punpckhwd %%mm4, %%mm3         \n\t"
 701             "punpckhwd %%mm0, %%mm6         \n\t"
 702             "movd      %%mm5, (%0)          \n\t"
 703             "punpckhdq %%mm5, %%mm5         \n\t"
 704             "movd      %%mm5, (%0, %2)      \n\t"
 705             "movd      %%mm3, (%0, %2, 2)   \n\t"
 706             "punpckhdq %%mm3, %%mm3         \n\t"
 707             "movd      %%mm3, (%0, %3)      \n\t"
 708             "movd      %%mm1, (%1)          \n\t"
 709             "punpckhdq %%mm1, %%mm1         \n\t"
 710             "movd      %%mm1, (%1, %2)      \n\t"
 711             "movd      %%mm6, (%1, %2, 2)   \n\t"
 712             "punpckhdq %%mm6, %%mm6         \n\t"
 713             "movd      %%mm6, (%1, %3)      \n\t"
 714             :: "r"(src),
 715                "r"(src + 4 * stride),
 716                "r"((x86_reg)stride),
 717                "r"((x86_reg)(3 * stride))
 718             );
 719     }
 720 }
 721
 722 /* Draw the edges of width 'w' of an image of size width, height
 723  * this MMX version can only handle w == 8 || w == 16. */
 724 static void draw_edges_mmx(uint8_t *buf, int wrap, int width, int height,
 725                            int w, int h, int sides)
 726 {
 727     uint8_t *ptr, *last_line;
 728     int i;
 729
 730     last_line = buf + (height - 1) * wrap;
 731     /* left and right */
 732     ptr = buf;
 733     if (w == 8) {
 734         __asm__ volatile (
 735             "1:                             \n\t"
 736             "movd            (%0), %%mm0    \n\t"
 737             "punpcklbw      %%mm0, %%mm0    \n\t"
 738             "punpcklwd      %%mm0, %%mm0    \n\t"
 739             "punpckldq      %%mm0, %%mm0    \n\t"
 740             "movq           %%mm0, -8(%0)   \n\t"
 741             "movq      -8(%0, %2), %%mm1    \n\t"
 742             "punpckhbw      %%mm1, %%mm1    \n\t"
 743             "punpckhwd      %%mm1, %%mm1    \n\t"
 744             "punpckhdq      %%mm1, %%mm1    \n\t"
 745             "movq           %%mm1, (%0, %2) \n\t"
 746             "add               %1, %0       \n\t"
 747             "cmp               %3, %0       \n\t"
 748             "jb                1b           \n\t"
 749             : "+r"(ptr)
 750             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
 751             );
 752     } else if(w==16){
 753         __asm__ volatile (
 754             "1:                                 \n\t"
 755             "movd            (%0), %%mm0        \n\t"
 756             "punpcklbw      %%mm0, %%mm0        \n\t"
 757             "punpcklwd      %%mm0, %%mm0        \n\t"
 758             "punpckldq      %%mm0, %%mm0        \n\t"
 759             "movq           %%mm0, -8(%0)       \n\t"
 760             "movq           %%mm0, -16(%0)      \n\t"
 761             "movq      -8(%0, %2), %%mm1        \n\t"
 762             "punpckhbw      %%mm1, %%mm1        \n\t"
 763             "punpckhwd      %%mm1, %%mm1        \n\t"
 764             "punpckhdq      %%mm1, %%mm1        \n\t"
 765             "movq           %%mm1,  (%0, %2)    \n\t"
 766             "movq           %%mm1, 8(%0, %2)    \n\t"
 767             "add               %1, %0           \n\t"
 768             "cmp               %3, %0           \n\t"
 769             "jb                1b               \n\t"
 770             : "+r"(ptr)
 771             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
 772             );
 773     } else {
 774         av_assert1(w == 4);
 775         __asm__ volatile (
 776             "1:                             \n\t"
 777             "movd            (%0), %%mm0    \n\t"
 778             "punpcklbw      %%mm0, %%mm0    \n\t"
 779             "punpcklwd      %%mm0, %%mm0    \n\t"
 780             "movd           %%mm0, -4(%0)   \n\t"
 781             "movd      -4(%0, %2), %%mm1    \n\t"
 782             "punpcklbw      %%mm1, %%mm1    \n\t"
 783             "punpckhwd      %%mm1, %%mm1    \n\t"
 784             "punpckhdq      %%mm1, %%mm1    \n\t"
 785             "movd           %%mm1, (%0, %2) \n\t"
 786             "add               %1, %0       \n\t"
 787             "cmp               %3, %0       \n\t"
 788             "jb                1b           \n\t"
 789             : "+r"(ptr)
 790             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
 791             );
 792     }
 793
 794     /* top and bottom (and hopefully also the corners) */
 795     if (sides & EDGE_TOP) {
 796         for (i = 0; i < h; i += 4) {
 797             ptr = buf - (i + 1) * wrap - w;
 798             __asm__ volatile (
 799                 "1:                             \n\t"
 800                 "movq (%1, %0), %%mm0           \n\t"
 801                 "movq    %%mm0, (%0)            \n\t"
 802                 "movq    %%mm0, (%0, %2)        \n\t"
 803                 "movq    %%mm0, (%0, %2, 2)     \n\t"
 804                 "movq    %%mm0, (%0, %3)        \n\t"
 805                 "add        $8, %0              \n\t"
 806                 "cmp        %4, %0              \n\t"
 807                 "jb         1b                  \n\t"
 808                 : "+r"(ptr)
 809                 : "r"((x86_reg)buf - (x86_reg)ptr - w), "r"((x86_reg) -wrap),
 810                   "r"((x86_reg) -wrap * 3), "r"(ptr + width + 2 * w)
 811                 );
 812         }
 813     }
 814
 815     if (sides & EDGE_BOTTOM) {
 816         for (i = 0; i < h; i += 4) {
 817             ptr = last_line + (i + 1) * wrap - w;
 818             __asm__ volatile (
 819                 "1:                             \n\t"
 820                 "movq (%1, %0), %%mm0           \n\t"
 821                 "movq    %%mm0, (%0)            \n\t"
 822                 "movq    %%mm0, (%0, %2)        \n\t"
 823                 "movq    %%mm0, (%0, %2, 2)     \n\t"
 824                 "movq    %%mm0, (%0, %3)        \n\t"
 825                 "add        $8, %0              \n\t"
 826                 "cmp        %4, %0              \n\t"
 827                 "jb         1b                  \n\t"
 828                 : "+r"(ptr)
 829                 : "r"((x86_reg)last_line - (x86_reg)ptr - w),
 830                   "r"((x86_reg)wrap), "r"((x86_reg)wrap * 3),
 831                   "r"(ptr + width + 2 * w)
 832                 );
 833         }
 834     }
 835 }
 836
 837 #define QPEL_V_LOW(m3, m4, m5, m6, pw_20, pw_3, rnd,                      \
 838                    in0, in1, in2, in7, out, OP)                           \
 839     "paddw               "#m4", "#m3"   \n\t" /* x1 */                    \
 840     "movq   "MANGLE(ff_pw_20)", %%mm4   \n\t" /* 20 */                    \
 841     "pmullw              "#m3", %%mm4   \n\t" /* 20x1 */                  \
 842     "movq               "#in7", "#m3"   \n\t" /* d */                     \
 843     "movq               "#in0", %%mm5   \n\t" /* D */                     \
 844     "paddw               "#m3", %%mm5   \n\t" /* x4 */                    \
 845     "psubw               %%mm5, %%mm4   \n\t" /* 20x1 - x4 */             \
 846     "movq               "#in1", %%mm5   \n\t" /* C */                     \
 847     "movq               "#in2", %%mm6   \n\t" /* B */                     \
 848     "paddw               "#m6", %%mm5   \n\t" /* x3 */                    \
 849     "paddw               "#m5", %%mm6   \n\t" /* x2 */                    \
 850     "paddw               %%mm6, %%mm6   \n\t" /* 2x2 */                   \
 851     "psubw               %%mm6, %%mm5   \n\t" /* -2x2 + x3 */             \
 852     "pmullw  "MANGLE(ff_pw_3)", %%mm5   \n\t" /* -6x2 + 3x3 */            \
 853     "paddw              "#rnd", %%mm4   \n\t" /* x2 */                    \
 854     "paddw               %%mm4, %%mm5   \n\t" /* 20x1 - 6x2 + 3x3 - x4 */ \
 855     "psraw                  $5, %%mm5   \n\t"                             \
 856     "packuswb            %%mm5, %%mm5   \n\t"                             \
 857     OP(%%mm5, out, %%mm7, d)
 858
 859 #define QPEL_BASE(OPNAME, ROUNDER, RND, OP_MMXEXT)                        \
 860 static void OPNAME ## mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst,         \
 861                                                     uint8_t *src,         \
 862                                                     int dstStride,        \
 863                                                     int srcStride,        \
 864                                                     int h)                \
 865 {                                                                         \
 866     uint64_t temp;                                                        \
 867                                                                           \
 868     __asm__ volatile (                                                    \
 869         "pxor      %%mm7, %%mm7             \n\t"                         \
 870         "1:                                 \n\t"                         \
 871         "movq       (%0), %%mm0             \n\t" /* ABCDEFGH */          \
 872         "movq      %%mm0, %%mm1             \n\t" /* ABCDEFGH */          \
 873         "movq      %%mm0, %%mm2             \n\t" /* ABCDEFGH */          \
 874         "punpcklbw %%mm7, %%mm0             \n\t" /* 0A0B0C0D */          \
 875         "punpckhbw %%mm7, %%mm1             \n\t" /* 0E0F0G0H */          \
 876         "pshufw    $0x90, %%mm0, %%mm5      \n\t" /* 0A0A0B0C */          \
 877         "pshufw    $0x41, %%mm0, %%mm6      \n\t" /* 0B0A0A0B */          \
 878         "movq      %%mm2, %%mm3             \n\t" /* ABCDEFGH */          \
 879         "movq      %%mm2, %%mm4             \n\t" /* ABCDEFGH */          \
 880         "psllq        $8, %%mm2             \n\t" /* 0ABCDEFG */          \
 881         "psllq       $16, %%mm3             \n\t" /* 00ABCDEF */          \
 882         "psllq       $24, %%mm4             \n\t" /* 000ABCDE */          \
 883         "punpckhbw %%mm7, %%mm2             \n\t" /* 0D0E0F0G */          \
 884         "punpckhbw %%mm7, %%mm3             \n\t" /* 0C0D0E0F */          \
 885         "punpckhbw %%mm7, %%mm4             \n\t" /* 0B0C0D0E */          \
 886         "paddw     %%mm3, %%mm5             \n\t" /* b */                 \
 887         "paddw     %%mm2, %%mm6             \n\t" /* c */                 \
 888         "paddw     %%mm5, %%mm5             \n\t" /* 2b */                \
 889         "psubw     %%mm5, %%mm6             \n\t" /* c - 2b */            \
 890         "pshufw    $0x06, %%mm0, %%mm5      \n\t" /* 0C0B0A0A */          \
 891         "pmullw "MANGLE(ff_pw_3)", %%mm6    \n\t" /* 3c - 6b */           \
 892         "paddw     %%mm4, %%mm0             \n\t" /* a */                 \
 893         "paddw     %%mm1, %%mm5             \n\t" /* d */                 \
 894         "pmullw "MANGLE(ff_pw_20)", %%mm0   \n\t" /* 20a */               \
 895         "psubw     %%mm5, %%mm0             \n\t" /* 20a - d */           \
 896         "paddw        %6, %%mm6             \n\t"                         \
 897         "paddw     %%mm6, %%mm0             \n\t" /* 20a - 6b + 3c - d */ \
 898         "psraw        $5, %%mm0             \n\t"                         \
 899         "movq      %%mm0, %5                \n\t"                         \
 900         /* mm1 = EFGH, mm2 = DEFG, mm3 = CDEF, mm4 = BCDE, mm7 = 0 */     \
 901                                                                           \
 902         "movq      5(%0), %%mm0             \n\t" /* FGHIJKLM */          \
 903         "movq      %%mm0, %%mm5             \n\t" /* FGHIJKLM */          \
 904         "movq      %%mm0, %%mm6             \n\t" /* FGHIJKLM */          \
 905         "psrlq        $8, %%mm0             \n\t" /* GHIJKLM0 */          \
 906         "psrlq       $16, %%mm5             \n\t" /* HIJKLM00 */          \
 907         "punpcklbw %%mm7, %%mm0             \n\t" /* 0G0H0I0J */          \
 908         "punpcklbw %%mm7, %%mm5             \n\t" /* 0H0I0J0K */          \
 909         "paddw     %%mm0, %%mm2             \n\t" /* b */                 \
 910         "paddw     %%mm5, %%mm3             \n\t" /* c */                 \
 911         "paddw     %%mm2, %%mm2             \n\t" /* 2b */                \
 912         "psubw     %%mm2, %%mm3             \n\t" /* c - 2b */            \
 913         "movq      %%mm6, %%mm2             \n\t" /* FGHIJKLM */          \
 914         "psrlq       $24, %%mm6             \n\t" /* IJKLM000 */          \
 915         "punpcklbw %%mm7, %%mm2             \n\t" /* 0F0G0H0I */          \
 916         "punpcklbw %%mm7, %%mm6             \n\t" /* 0I0J0K0L */          \
 917         "pmullw "MANGLE(ff_pw_3)", %%mm3    \n\t" /* 3c - 6b */           \
 918         "paddw     %%mm2, %%mm1             \n\t" /* a */                 \
 919         "paddw     %%mm6, %%mm4             \n\t" /* d */                 \
 920         "pmullw "MANGLE(ff_pw_20)", %%mm1   \n\t" /* 20a */               \
 921         "psubw     %%mm4, %%mm3             \n\t" /* - 6b +3c - d */      \
 922         "paddw        %6, %%mm1             \n\t"                         \
 923         "paddw     %%mm1, %%mm3             \n\t" /* 20a - 6b +3c - d */  \
 924         "psraw        $5, %%mm3             \n\t"                         \
 925         "movq         %5, %%mm1             \n\t"                         \
 926         "packuswb  %%mm3, %%mm1             \n\t"                         \
 927         OP_MMXEXT(%%mm1, (%1), %%mm4, q)                                  \
 928         /* mm0 = GHIJ, mm2 = FGHI, mm5 = HIJK, mm6 = IJKL, mm7 = 0 */     \
 929                                                                           \
 930         "movq      9(%0), %%mm1             \n\t" /* JKLMNOPQ */          \
 931         "movq      %%mm1, %%mm4             \n\t" /* JKLMNOPQ */          \
 932         "movq      %%mm1, %%mm3             \n\t" /* JKLMNOPQ */          \
 933         "psrlq        $8, %%mm1             \n\t" /* KLMNOPQ0 */          \
 934         "psrlq       $16, %%mm4             \n\t" /* LMNOPQ00 */          \
 935         "punpcklbw %%mm7, %%mm1             \n\t" /* 0K0L0M0N */          \
 936         "punpcklbw %%mm7, %%mm4             \n\t" /* 0L0M0N0O */          \
 937         "paddw     %%mm1, %%mm5             \n\t" /* b */                 \
 938         "paddw     %%mm4, %%mm0             \n\t" /* c */                 \
 939         "paddw     %%mm5, %%mm5             \n\t" /* 2b */                \
 940         "psubw     %%mm5, %%mm0             \n\t" /* c - 2b */            \
 941         "movq      %%mm3, %%mm5             \n\t" /* JKLMNOPQ */          \
 942         "psrlq       $24, %%mm3             \n\t" /* MNOPQ000 */          \
 943         "pmullw "MANGLE(ff_pw_3)", %%mm0    \n\t" /* 3c - 6b */           \
 944         "punpcklbw %%mm7, %%mm3             \n\t" /* 0M0N0O0P */          \
 945         "paddw     %%mm3, %%mm2             \n\t" /* d */                 \
 946         "psubw     %%mm2, %%mm0             \n\t" /* -6b + 3c - d */      \
 947         "movq      %%mm5, %%mm2             \n\t" /* JKLMNOPQ */          \
 948         "punpcklbw %%mm7, %%mm2             \n\t" /* 0J0K0L0M */          \
 949         "punpckhbw %%mm7, %%mm5             \n\t" /* 0N0O0P0Q */          \
 950         "paddw     %%mm2, %%mm6             \n\t" /* a */                 \
 951         "pmullw "MANGLE(ff_pw_20)", %%mm6   \n\t" /* 20a */               \
 952         "paddw        %6, %%mm0             \n\t"                         \
 953         "paddw     %%mm6, %%mm0             \n\t" /* 20a - 6b + 3c - d */ \
 954         "psraw        $5, %%mm0             \n\t"                         \
 955         /* mm1 = KLMN, mm2 = JKLM, mm3 = MNOP, */                         \
 956         /* mm4 = LMNO, mm5 = NOPQ mm7 = 0 */                              \
 957                                                                           \
 958         "paddw    %%mm5, %%mm3              \n\t" /* a */                 \
 959         "pshufw   $0xF9, %%mm5, %%mm6       \n\t" /* 0O0P0Q0Q */          \
 960         "paddw    %%mm4, %%mm6              \n\t" /* b */                 \
 961         "pshufw   $0xBE, %%mm5, %%mm4       \n\t" /* 0P0Q0Q0P */          \
 962         "pshufw   $0x6F, %%mm5, %%mm5       \n\t" /* 0Q0Q0P0O */          \
 963         "paddw    %%mm1, %%mm4              \n\t" /* c */                 \
 964         "paddw    %%mm2, %%mm5              \n\t" /* d */                 \
 965         "paddw    %%mm6, %%mm6              \n\t" /* 2b */                \
 966         "psubw    %%mm6, %%mm4              \n\t" /* c - 2b */            \
 967         "pmullw "MANGLE(ff_pw_20)", %%mm3   \n\t" /* 20a */               \
 968         "pmullw  "MANGLE(ff_pw_3)", %%mm4   \n\t" /* 3c - 6b */           \
 969         "psubw    %%mm5, %%mm3              \n\t" /* -6b + 3c - d */      \
 970         "paddw       %6, %%mm4              \n\t"                         \
 971         "paddw    %%mm3, %%mm4              \n\t" /* 20a - 6b + 3c - d */ \
 972         "psraw       $5, %%mm4              \n\t"                         \
 973         "packuswb %%mm4, %%mm0              \n\t"                         \
 974         OP_MMXEXT(%%mm0, 8(%1), %%mm4, q)                                 \
 975                                                                           \
 976         "add         %3, %0                 \n\t"                         \
 977         "add         %4, %1                 \n\t"                         \
 978         "decl        %2                     \n\t"                         \
 979         "jnz         1b                     \n\t"                         \
 980         : "+a"(src), "+c"(dst), "+D"(h)                                   \
 981         : "d"((x86_reg)srcStride), "S"((x86_reg)dstStride),               \
 982           /* "m"(ff_pw_20), "m"(ff_pw_3), */ "m"(temp), "m"(ROUNDER)      \
 983         : "memory"                                                        \
 984         );                                                                \
 985 }                                                                         \
 986                                                                           \
 987 static void OPNAME ## mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst,          \
 988                                                    uint8_t *src,          \
 989                                                    int dstStride,         \
 990                                                    int srcStride,         \
 991                                                    int h)                 \
 992 {                                                                         \
 993     __asm__ volatile (                                                    \
 994         "pxor      %%mm7, %%mm7             \n\t"                         \
 995         "1:                                 \n\t"                         \
 996         "movq       (%0), %%mm0             \n\t" /* ABCDEFGH */          \
 997         "movq      %%mm0, %%mm1             \n\t" /* ABCDEFGH */          \
 998         "movq      %%mm0, %%mm2             \n\t" /* ABCDEFGH */          \
 999         "punpcklbw %%mm7, %%mm0             \n\t" /* 0A0B0C0D */          \
1000         "punpckhbw %%mm7, %%mm1             \n\t" /* 0E0F0G0H */          \
1001         "pshufw    $0x90, %%mm0, %%mm5      \n\t" /* 0A0A0B0C */          \
1002         "pshufw    $0x41, %%mm0, %%mm6      \n\t" /* 0B0A0A0B */          \
1003         "movq      %%mm2, %%mm3             \n\t" /* ABCDEFGH */          \
1004         "movq      %%mm2, %%mm4             \n\t" /* ABCDEFGH */          \
1005         "psllq        $8, %%mm2             \n\t" /* 0ABCDEFG */          \
1006         "psllq       $16, %%mm3             \n\t" /* 00ABCDEF */          \
1007         "psllq       $24, %%mm4             \n\t" /* 000ABCDE */          \
1008         "punpckhbw %%mm7, %%mm2             \n\t" /* 0D0E0F0G */          \
1009         "punpckhbw %%mm7, %%mm3             \n\t" /* 0C0D0E0F */          \
1010         "punpckhbw %%mm7, %%mm4             \n\t" /* 0B0C0D0E */          \
1011         "paddw     %%mm3, %%mm5             \n\t" /* b */                 \
1012         "paddw     %%mm2, %%mm6             \n\t" /* c */                 \
1013         "paddw     %%mm5, %%mm5             \n\t" /* 2b */                \
1014         "psubw     %%mm5, %%mm6             \n\t" /* c - 2b */            \
1015         "pshufw    $0x06, %%mm0, %%mm5      \n\t" /* 0C0B0A0A */          \
1016         "pmullw "MANGLE(ff_pw_3)", %%mm6    \n\t" /* 3c - 6b */           \
1017         "paddw     %%mm4, %%mm0             \n\t" /* a */                 \
1018         "paddw     %%mm1, %%mm5             \n\t" /* d */                 \
1019         "pmullw "MANGLE(ff_pw_20)", %%mm0   \n\t" /* 20a */               \
1020         "psubw     %%mm5, %%mm0             \n\t" /* 20a - d */           \
1021         "paddw        %5, %%mm6             \n\t"                         \
1022         "paddw     %%mm6, %%mm0             \n\t" /* 20a - 6b + 3c - d */ \
1023         "psraw        $5, %%mm0             \n\t"                         \
1024         /* mm1 = EFGH, mm2 = DEFG, mm3 = CDEF, mm4 = BCDE, mm7 = 0 */     \
1025                                                                           \
1026         "movd      5(%0), %%mm5             \n\t" /* FGHI */              \
1027         "punpcklbw %%mm7, %%mm5             \n\t" /* 0F0G0H0I */          \
1028         "pshufw    $0xF9, %%mm5, %%mm6      \n\t" /* 0G0H0I0I */          \
1029         "paddw     %%mm5, %%mm1             \n\t" /* a */                 \
1030         "paddw     %%mm6, %%mm2             \n\t" /* b */                 \
1031         "pshufw    $0xBE, %%mm5, %%mm6      \n\t" /* 0H0I0I0H */          \
1032         "pshufw    $0x6F, %%mm5, %%mm5      \n\t" /* 0I0I0H0G */          \
1033         "paddw     %%mm6, %%mm3             \n\t" /* c */                 \
1034         "paddw     %%mm5, %%mm4             \n\t" /* d */                 \
1035         "paddw     %%mm2, %%mm2             \n\t" /* 2b */                \
1036         "psubw     %%mm2, %%mm3             \n\t" /* c - 2b */            \
1037         "pmullw "MANGLE(ff_pw_20)", %%mm1   \n\t" /* 20a */               \
1038         "pmullw  "MANGLE(ff_pw_3)", %%mm3   \n\t" /* 3c - 6b */           \
1039         "psubw     %%mm4, %%mm3             \n\t" /* -6b + 3c - d */      \
1040         "paddw        %5, %%mm1             \n\t"                         \
1041         "paddw     %%mm1, %%mm3             \n\t" /* 20a - 6b + 3c - d */ \
1042         "psraw        $5, %%mm3             \n\t"                         \
1043         "packuswb  %%mm3, %%mm0             \n\t"                         \
1044         OP_MMXEXT(%%mm0, (%1), %%mm4, q)                                  \
1045                                                                           \
1046         "add          %3, %0                \n\t"                         \
1047         "add          %4, %1                \n\t"                         \
1048         "decl         %2                    \n\t"                         \
1049         "jnz          1b                    \n\t"                         \
1050         : "+a"(src), "+c"(dst), "+d"(h)                                   \
1051         : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride),               \
1052           /* "m"(ff_pw_20), "m"(ff_pw_3), */ "m"(ROUNDER)                 \
1053         : "memory"                                                        \
1054         );                                                                \
1055 }
1056
1057 #define QPEL_OP(OPNAME, ROUNDER, RND, OP, MMX)                          \
1058 static void OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(uint8_t *dst,      \
1059                                                      uint8_t *src,      \
1060                                                      int dstStride,     \
1061                                                      int srcStride)     \
1062 {                                                                       \
1063     uint64_t temp[17 * 4];                                              \
1064     uint64_t *temp_ptr = temp;                                          \
1065     int count = 17;                                                     \
1066                                                                         \
1067     /* FIXME unroll */                                                  \
1068     __asm__ volatile (                                                  \
1069         "pxor      %%mm7, %%mm7             \n\t"                       \
1070         "1:                                 \n\t"                       \
1071         "movq       (%0), %%mm0             \n\t"                       \
1072         "movq       (%0), %%mm1             \n\t"                       \
1073         "movq      8(%0), %%mm2             \n\t"                       \
1074         "movq      8(%0), %%mm3             \n\t"                       \
1075         "punpcklbw %%mm7, %%mm0             \n\t"                       \
1076         "punpckhbw %%mm7, %%mm1             \n\t"                       \
1077         "punpcklbw %%mm7, %%mm2             \n\t"                       \
1078         "punpckhbw %%mm7, %%mm3             \n\t"                       \
1079         "movq      %%mm0, (%1)              \n\t"                       \
1080         "movq      %%mm1, 17 * 8(%1)        \n\t"                       \
1081         "movq      %%mm2, 2 * 17 * 8(%1)    \n\t"                       \
1082         "movq      %%mm3, 3 * 17 * 8(%1)    \n\t"                       \
1083         "add          $8, %1                \n\t"                       \
1084         "add          %3, %0                \n\t"                       \
1085         "decl         %2                    \n\t"                       \
1086         "jnz          1b                    \n\t"                       \
1087         : "+r"(src), "+r"(temp_ptr), "+r"(count)                        \
1088         : "r"((x86_reg)srcStride)                                       \
1089         : "memory"                                                      \
1090         );                                                              \
1091                                                                         \
1092     temp_ptr = temp;                                                    \
1093     count    = 4;                                                       \
1094                                                                         \
1095     /* FIXME reorder for speed */                                       \
1096     __asm__ volatile (                                                  \
1097         /* "pxor  %%mm7, %%mm7            \n\t" */                      \
1098         "1:                             \n\t"                           \
1099         "movq    (%0), %%mm0            \n\t"                           \
1100         "movq   8(%0), %%mm1            \n\t"                           \
1101         "movq  16(%0), %%mm2            \n\t"                           \
1102         "movq  24(%0), %%mm3            \n\t"                           \
1103         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0),   8(%0),    (%0),  32(%0), (%1),     OP) \
1104         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5,  8(%0),    (%0),    (%0),  40(%0), (%1, %3), OP) \
1105         "add       %4, %1               \n\t"                           \
1106         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5,   (%0),    (%0),   8(%0),  48(%0), (%1),     OP) \
1107                                                                         \
1108         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5,   (%0),   8(%0),  16(%0),  56(%0), (%1, %3), OP) \
1109         "add       %4, %1               \n\t"                           \
1110         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5,  8(%0),  16(%0),  24(%0),  64(%0), (%1),     OP) \
1111         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0),  24(%0),  32(%0),  72(%0), (%1, %3), OP) \
1112         "add       %4, %1               \n\t"                           \
1113         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0),  32(%0),  40(%0),  80(%0), (%1),     OP) \
1114         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0),  40(%0),  48(%0),  88(%0), (%1, %3), OP) \
1115         "add       %4, %1               \n\t"                           \
1116         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 40(%0),  48(%0),  56(%0),  96(%0), (%1),     OP) \
1117         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 48(%0),  56(%0),  64(%0), 104(%0), (%1, %3), OP) \
1118         "add       %4, %1               \n\t"                           \
1119         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 56(%0),  64(%0),  72(%0), 112(%0), (%1),     OP) \
1120         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 64(%0),  72(%0),  80(%0), 120(%0), (%1, %3), OP) \
1121         "add       %4, %1               \n\t"                           \
1122         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 72(%0),  80(%0),  88(%0), 128(%0), (%1),     OP) \
1123                                                                         \
1124         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 80(%0),  88(%0),  96(%0), 128(%0), (%1, %3), OP) \
1125         "add       %4, %1               \n\t"                           \
1126         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 88(%0),  96(%0), 104(%0), 120(%0), (%1),     OP) \
1127         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 96(%0), 104(%0), 112(%0), 112(%0), (%1, %3), OP) \
1128                                                                         \
1129         "add     $136, %0               \n\t"                           \
1130         "add       %6, %1               \n\t"                           \
1131         "decl      %2                   \n\t"                           \
1132         "jnz       1b                   \n\t"                           \
1133                                                                         \
1134         : "+r"(temp_ptr), "+r"(dst), "+g"(count)                        \
1135         : "r"((x86_reg)dstStride), "r"(2 * (x86_reg)dstStride),         \
1136           /* "m"(ff_pw_20), "m"(ff_pw_3), */ "m"(ROUNDER),              \
1137           "g"(4 - 14 * (x86_reg)dstStride)                              \
1138         : "memory"                                                      \
1139         );                                                              \
1140 }                                                                       \
1141                                                                         \
1142 static void OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(uint8_t *dst,       \
1143                                                     uint8_t *src,       \
1144                                                     int dstStride,      \
1145                                                     int srcStride)      \
1146 {                                                                       \
1147     uint64_t temp[9 * 2];                                               \
1148     uint64_t *temp_ptr = temp;                                          \
1149     int count = 9;                                                      \
1150                                                                         \
1151     /* FIXME unroll */                                                  \
1152     __asm__ volatile (                                                  \
1153         "pxor      %%mm7, %%mm7         \n\t"                           \
1154         "1:                             \n\t"                           \
1155         "movq       (%0), %%mm0         \n\t"                           \
1156         "movq       (%0), %%mm1         \n\t"                           \
1157         "punpcklbw %%mm7, %%mm0         \n\t"                           \
1158         "punpckhbw %%mm7, %%mm1         \n\t"                           \
1159         "movq      %%mm0, (%1)          \n\t"                           \
1160         "movq      %%mm1, 9*8(%1)       \n\t"                           \
1161         "add          $8, %1            \n\t"                           \
1162         "add          %3, %0            \n\t"                           \
1163         "decl         %2                \n\t"                           \
1164         "jnz          1b                \n\t"                           \
1165         : "+r"(src), "+r"(temp_ptr), "+r"(count)                        \
1166         : "r"((x86_reg)srcStride)                                       \
1167         : "memory"                                                      \
1168         );                                                              \
1169                                                                         \
1170     temp_ptr = temp;                                                    \
1171     count    = 2;                                                       \
1172                                                                         \
1173     /* FIXME reorder for speed */                                       \
1174     __asm__ volatile (                                                  \
1175         /* "pxor  %%mm7, %%mm7            \n\t" */                      \
1176         "1:                             \n\t"                           \
1177         "movq    (%0), %%mm0            \n\t"                           \
1178         "movq   8(%0), %%mm1            \n\t"                           \
1179         "movq  16(%0), %%mm2            \n\t"                           \
1180         "movq  24(%0), %%mm3            \n\t"                           \
1181         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0),  8(%0),   (%0), 32(%0), (%1), OP)     \
1182         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5,  8(%0),   (%0),   (%0), 40(%0), (%1, %3), OP) \
1183         "add       %4, %1               \n\t"                           \
1184         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5,   (%0),   (%0),  8(%0), 48(%0), (%1), OP)     \
1185                                                                         \
1186         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5,   (%0),  8(%0), 16(%0), 56(%0), (%1, %3), OP) \
1187         "add       %4, %1               \n\t"                           \
1188         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5,  8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)     \
1189                                                                         \
1190         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 64(%0), (%1, %3), OP) \
1191         "add       %4, %1               \n\t"                           \
1192         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 56(%0), (%1), OP)     \
1193         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 48(%0), (%1, %3), OP) \
1194                                                                         \
1195         "add      $72, %0               \n\t"                           \
1196         "add       %6, %1               \n\t"                           \
1197         "decl      %2                   \n\t"                           \
1198         "jnz       1b                   \n\t"                           \
1199                                                                         \
1200         : "+r"(temp_ptr), "+r"(dst), "+g"(count)                        \
1201         : "r"((x86_reg)dstStride), "r"(2 * (x86_reg)dstStride),         \
1202           /* "m"(ff_pw_20), "m"(ff_pw_3), */ "m"(ROUNDER),              \
1203           "g"(4 - 6 * (x86_reg)dstStride)                               \
1204         : "memory"                                                      \
1205         );                                                              \
1206 }                                                                       \
1207                                                                         \
1208 static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src,   \
1209                                           int stride)                   \
1210 {                                                                       \
1211     OPNAME ## pixels8_ ## MMX(dst, src, stride, 8);                     \
1212 }                                                                       \
1213                                                                         \
1214 static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src,    \
1215                                          int stride)                    \
1216 {                                                                       \
1217     uint64_t temp[8];                                                   \
1218     uint8_t * const half = (uint8_t*)temp;                              \
1219     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8,           \
1220                                                 stride, 8);             \
1221     OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);    \
1222 }                                                                       \
1223                                                                         \
1224 static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src,    \
1225                                          int stride)                    \
1226 {                                                                       \
1227     OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride,           \
1228                                             stride, 8);                 \
1229 }                                                                       \
1230                                                                         \
1231 static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src,    \
1232                                          int stride)                    \
1233 {                                                                       \
1234     uint64_t temp[8];                                                   \
1235     uint8_t * const half = (uint8_t*)temp;                              \
1236     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8,           \
1237                                                 stride, 8);             \
1238     OPNAME ## pixels8_l2_ ## MMX(dst, src + 1, half, stride,            \
1239                                  stride, 8);                            \
1240 }                                                                       \
1241                                                                         \
1242 static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src,    \
1243                                          int stride)                    \
1244 {                                                                       \
1245     uint64_t temp[8];                                                   \
1246     uint8_t * const half = (uint8_t*)temp;                              \
1247     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);  \
1248     OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);    \
1249 }                                                                       \
1250                                                                         \
1251 static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src,    \
1252                                          int stride)                    \
1253 {                                                                       \
1254     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src, stride, stride);  \
1255 }                                                                       \
1256                                                                         \
1257 static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src,    \
1258                                          int stride)                    \
1259 {                                                                       \
1260     uint64_t temp[8];                                                   \
1261     uint8_t * const half = (uint8_t*)temp;                              \
1262     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);  \
1263     OPNAME ## pixels8_l2_ ## MMX(dst, src + stride, half, stride,       \
1264                                  stride, 8);                            \
1265 }                                                                       \
1266                                                                         \
1267 static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src,    \
1268                                          int stride)                    \
1269 {                                                                       \
1270     uint64_t half[8 + 9];                                               \
1271     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1272     uint8_t * const halfHV = ((uint8_t*)half);                          \
1273     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1274                                                 stride, 9);             \
1275     put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);  \
1276     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1277     OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);     \
1278 }                                                                       \
1279                                                                         \
1280 static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src,    \
1281                                          int stride)                    \
1282 {                                                                       \
1283     uint64_t half[8 + 9];                                               \
1284     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1285     uint8_t * const halfHV = ((uint8_t*)half);                          \
1286     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1287                                                 stride, 9);             \
1288     put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,          \
1289                                      stride, 9);                        \
1290     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1291     OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);     \
1292 }                                                                       \
1293                                                                         \
1294 static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src,    \
1295                                          int stride)                    \
1296 {                                                                       \
1297     uint64_t half[8 + 9];                                               \
1298     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1299     uint8_t * const halfHV = ((uint8_t*)half);                          \
1300     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1301                                                 stride, 9);             \
1302     put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);  \
1303     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1304     OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV, stride, 8, 8); \
1305 }                                                                       \
1306                                                                         \
1307 static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src,    \
1308                                          int stride)                    \
1309 {                                                                       \
1310     uint64_t half[8 + 9];                                               \
1311     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1312     uint8_t * const halfHV = ((uint8_t*)half);                          \
1313     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1314                                                 stride, 9);             \
1315     put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,          \
1316                                      stride, 9);                        \
1317     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1318     OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV, stride, 8, 8); \
1319 }                                                                       \
1320                                                                         \
1321 static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src,    \
1322                                          int stride)                    \
1323 {                                                                       \
1324     uint64_t half[8 + 9];                                               \
1325     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1326     uint8_t * const halfHV = ((uint8_t*)half);                          \
1327     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1328                                                 stride, 9);             \
1329     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1330     OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);     \
1331 }                                                                       \
1332                                                                         \
1333 static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src,    \
1334                                          int stride)                    \
1335 {                                                                       \
1336     uint64_t half[8 + 9];                                               \
1337     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1338     uint8_t * const halfHV = ((uint8_t*)half);                          \
1339     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1340                                                 stride, 9);             \
1341     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1342     OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV, stride, 8, 8); \
1343 }                                                                       \
1344                                                                         \
1345 static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src,    \
1346                                          int stride)                    \
1347 {                                                                       \
1348     uint64_t half[8 + 9];                                               \
1349     uint8_t * const halfH = ((uint8_t*)half);                           \
1350     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1351                                                 stride, 9);             \
1352     put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);  \
1353     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);     \
1354 }                                                                       \
1355                                                                         \
1356 static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src,    \
1357                                          int stride)                    \
1358 {                                                                       \
1359     uint64_t half[8 + 9];                                               \
1360     uint8_t * const halfH = ((uint8_t*)half);                           \
1361     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1362                                                 stride, 9);             \
1363     put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,          \
1364                                      stride, 9);                        \
1365     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);     \
1366 }                                                                       \
1367                                                                         \
1368 static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src,    \
1369                                          int stride)                    \
1370 {                                                                       \
1371     uint64_t half[9];                                                   \
1372     uint8_t * const halfH = ((uint8_t*)half);                           \
1373     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1374                                                 stride, 9);             \
1375     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);     \
1376 }                                                                       \
1377                                                                         \
1378 static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src,  \
1379                                            int stride)                  \
1380 {                                                                       \
1381     OPNAME ## pixels16_ ## MMX(dst, src, stride, 16);                   \
1382 }                                                                       \
1383                                                                         \
1384 static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src,   \
1385                                           int stride)                   \
1386 {                                                                       \
1387     uint64_t temp[32];                                                  \
1388     uint8_t * const half = (uint8_t*)temp;                              \
1389     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16,         \
1390                                                  stride, 16);           \
1391     OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);  \
1392 }                                                                       \
1393                                                                         \
1394 static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src,   \
1395                                           int stride)                   \
1396 {                                                                       \
1397     OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src,                  \
1398                                              stride, stride, 16);       \
1399 }                                                                       \
1400                                                                         \
1401 static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src,   \
1402                                           int stride)                   \
1403 {                                                                       \
1404     uint64_t temp[32];                                                  \
1405     uint8_t * const half = (uint8_t*)temp;                              \
1406     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16,         \
1407                                                  stride, 16);           \
1408     OPNAME ## pixels16_l2_ ## MMX(dst, src + 1, half,                   \
1409                                   stride, stride, 16);                  \
1410 }                                                                       \
1411                                                                         \
1412 static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src,   \
1413                                           int stride)                   \
1414 {                                                                       \
1415     uint64_t temp[32];                                                  \
1416     uint8_t * const half = (uint8_t*)temp;                              \
1417     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16,         \
1418                                                  stride);               \
1419     OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);  \
1420 }                                                                       \
1421                                                                         \
1422 static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src,   \
1423                                           int stride)                   \
1424 {                                                                       \
1425     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src, stride, stride); \
1426 }                                                                       \
1427                                                                         \
1428 static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src,   \
1429                                           int stride)                   \
1430 {                                                                       \
1431     uint64_t temp[32];                                                  \
1432     uint8_t * const half = (uint8_t*)temp;                              \
1433     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16,         \
1434                                                  stride);               \
1435     OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half,                \
1436                                   stride, stride, 16);                  \
1437 }                                                                       \
1438                                                                         \
1439 static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src,   \
1440                                           int stride)                   \
1441 {                                                                       \
1442     uint64_t half[16 * 2 + 17 * 2];                                     \
1443     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1444     uint8_t * const halfHV = ((uint8_t*)half);                          \
1445     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1446                                                  stride, 17);           \
1447     put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,            \
1448                                       stride, 17);                      \
1449     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1450                                                  16, 16);               \
1451     OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);  \
1452 }                                                                       \
1453                                                                         \
1454 static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src,   \
1455                                           int stride)                   \
1456 {                                                                       \
1457     uint64_t half[16 * 2 + 17 * 2];                                     \
1458     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1459     uint8_t * const halfHV = ((uint8_t*)half);                          \
1460     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1461                                                  stride, 17);           \
1462     put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,        \
1463                                       stride, 17);                      \
1464     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1465                                                  16, 16);               \
1466     OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);  \
1467 }                                                                       \
1468                                                                         \
1469 static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src,   \
1470                                           int stride)                   \
1471 {                                                                       \
1472     uint64_t half[16 * 2 + 17 * 2];                                     \
1473     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1474     uint8_t * const halfHV = ((uint8_t*)half);                          \
1475     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1476                                                  stride, 17);           \
1477     put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,            \
1478                                       stride, 17);                      \
1479     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1480                                                  16, 16);               \
1481     OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV, stride,      \
1482                                   16, 16);                              \
1483 }                                                                       \
1484                                                                         \
1485 static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src,   \
1486                                           int stride)                   \
1487 {                                                                       \
1488     uint64_t half[16 * 2 + 17 * 2];                                     \
1489     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1490     uint8_t * const halfHV = ((uint8_t*)half);                          \
1491     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1492                                                  stride, 17);           \
1493     put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,        \
1494                                       stride, 17);                      \
1495     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1496                                                  16, 16);               \
1497     OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV, stride,      \
1498                                   16, 16);                              \
1499 }                                                                       \
1500                                                                         \
1501 static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src,   \
1502                                           int stride)                   \
1503 {                                                                       \
1504     uint64_t half[16 * 2 + 17 * 2];                                     \
1505     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1506     uint8_t * const halfHV = ((uint8_t*)half);                          \
1507     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1508                                                  stride, 17);           \
1509     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1510                                                  16, 16);               \
1511     OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);  \
1512 }                                                                       \
1513                                                                         \
1514 static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src,   \
1515                                           int stride)                   \
1516 {                                                                       \
1517     uint64_t half[16 * 2 + 17 * 2];                                     \
1518     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1519     uint8_t * const halfHV = ((uint8_t*)half);                          \
1520     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1521                                                  stride, 17);           \
1522     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1523                                                  16, 16);               \
1524     OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV, stride,      \
1525                                   16, 16);                              \
1526 }                                                                       \
1527                                                                         \
1528 static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src,   \
1529                                           int stride)                   \
1530 {                                                                       \
1531     uint64_t half[17 * 2];                                              \
1532     uint8_t * const halfH = ((uint8_t*)half);                           \
1533     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1534                                                  stride, 17);           \
1535     put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,            \
1536                                       stride, 17);                      \
1537     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);   \
1538 }                                                                       \
1539                                                                         \
1540 static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src,   \
1541                                           int stride)                   \
1542 {                                                                       \
1543     uint64_t half[17 * 2];                                              \
1544     uint8_t * const halfH = ((uint8_t*)half);                           \
1545     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1546                                                  stride, 17);           \
1547     put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,        \
1548                                       stride, 17);                      \
1549     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);   \
1550 }                                                                       \
1551                                                                         \
1552 static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src,   \
1553                                           int stride)                   \
1554 {                                                                       \
1555     uint64_t half[17 * 2];                                              \
1556     uint8_t * const halfH = ((uint8_t*)half);                           \
1557     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1558                                                  stride, 17);           \
1559     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);   \
1560 }
1561
1562 #define PUT_OP(a, b, temp, size)                \
1563     "mov"#size"        "#a", "#b"       \n\t"
1564
1565 #define AVG_MMXEXT_OP(a, b, temp, size)         \
1566     "mov"#size"        "#b", "#temp"    \n\t"   \
1567     "pavgb          "#temp", "#a"       \n\t"   \
1568     "mov"#size"        "#a", "#b"       \n\t"
1569
1570 QPEL_BASE(put_,        ff_pw_16, _,        PUT_OP)
1571 QPEL_BASE(avg_,        ff_pw_16, _,        AVG_MMXEXT_OP)
1572 QPEL_BASE(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP)
1573 QPEL_OP(put_,          ff_pw_16, _,        PUT_OP,        mmxext)
1574 QPEL_OP(avg_,          ff_pw_16, _,        AVG_MMXEXT_OP, mmxext)
1575 QPEL_OP(put_no_rnd_,   ff_pw_15, _no_rnd_, PUT_OP,        mmxext)
1576
1577 void ff_put_rv40_qpel8_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1578 {
1579   put_pixels8_xy2_mmx(dst, src, stride, 8);
1580 }
1581 void ff_put_rv40_qpel16_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1582 {
1583   put_pixels16_xy2_mmx(dst, src, stride, 16);
1584 }
1585 void ff_avg_rv40_qpel8_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1586 {
1587   avg_pixels8_xy2_mmx(dst, src, stride, 8);
1588 }
1589 void ff_avg_rv40_qpel16_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1590 {
1591   avg_pixels16_xy2_mmx(dst, src, stride, 16);
1592 }
1593
1594 typedef void emulated_edge_mc_func(uint8_t *dst, const uint8_t *src,
1595                                    ptrdiff_t linesize, int block_w, int block_h,
1596                                    int src_x, int src_y, int w, int h);
1597
1598 static av_always_inline void gmc(uint8_t *dst, uint8_t *src,
1599                                  int stride, int h, int ox, int oy,
1600                                  int dxx, int dxy, int dyx, int dyy,
1601                                  int shift, int r, int width, int height,
1602                                  emulated_edge_mc_func *emu_edge_fn)
1603 {
1604     const int w    = 8;
1605     const int ix   = ox  >> (16 + shift);
1606     const int iy   = oy  >> (16 + shift);
1607     const int oxs  = ox  >> 4;
1608     const int oys  = oy  >> 4;
1609     const int dxxs = dxx >> 4;
1610     const int dxys = dxy >> 4;
1611     const int dyxs = dyx >> 4;
1612     const int dyys = dyy >> 4;
1613     const uint16_t r4[4]   = { r, r, r, r };
1614     const uint16_t dxy4[4] = { dxys, dxys, dxys, dxys };
1615     const uint16_t dyy4[4] = { dyys, dyys, dyys, dyys };
1616     const uint64_t shift2 = 2 * shift;
1617 #define MAX_STRIDE 4096U
1618 #define MAX_H 8U
1619     uint8_t edge_buf[(MAX_H + 1) * MAX_STRIDE];
1620     int x, y;
1621
1622     const int dxw = (dxx - (1 << (16 + shift))) * (w - 1);
1623     const int dyh = (dyy - (1 << (16 + shift))) * (h - 1);
1624     const int dxh = dxy * (h - 1);
1625     const int dyw = dyx * (w - 1);
1626     int need_emu =  (unsigned)ix >= width  - w ||
1627                     (unsigned)iy >= height - h;
1628
1629     if ( // non-constant fullpel offset (3% of blocks)
1630         ((ox ^ (ox + dxw)) | (ox ^ (ox + dxh)) | (ox ^ (ox + dxw + dxh)) |
1631          (oy ^ (oy + dyw)) | (oy ^ (oy + dyh)) | (oy ^ (oy + dyw + dyh))) >> (16 + shift)
1632         // uses more than 16 bits of subpel mv (only at huge resolution)
1633         || (dxx | dxy | dyx | dyy) & 15
1634         || (need_emu && (h > MAX_H || stride > MAX_STRIDE))) {
1635         // FIXME could still use mmx for some of the rows
1636         ff_gmc_c(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy,
1637                  shift, r, width, height);
1638         return;
1639     }
1640
1641     src += ix + iy * stride;
1642     if (need_emu) {
1643         emu_edge_fn(edge_buf, src, stride, w + 1, h + 1, ix, iy, width, height);
1644         src = edge_buf;
1645     }
1646
1647     __asm__ volatile (
1648         "movd         %0, %%mm6         \n\t"
1649         "pxor      %%mm7, %%mm7         \n\t"
1650         "punpcklwd %%mm6, %%mm6         \n\t"
1651         "punpcklwd %%mm6, %%mm6         \n\t"
1652         :: "r"(1<<shift)
1653     );
1654
1655     for (x = 0; x < w; x += 4) {
1656         uint16_t dx4[4] = { oxs - dxys + dxxs * (x + 0),
1657                             oxs - dxys + dxxs * (x + 1),
1658                             oxs - dxys + dxxs * (x + 2),
1659                             oxs - dxys + dxxs * (x + 3) };
1660         uint16_t dy4[4] = { oys - dyys + dyxs * (x + 0),
1661                             oys - dyys + dyxs * (x + 1),
1662                             oys - dyys + dyxs * (x + 2),
1663                             oys - dyys + dyxs * (x + 3) };
1664
1665         for (y = 0; y < h; y++) {
1666             __asm__ volatile (
1667                 "movq      %0, %%mm4    \n\t"
1668                 "movq      %1, %%mm5    \n\t"
1669                 "paddw     %2, %%mm4    \n\t"
1670                 "paddw     %3, %%mm5    \n\t"
1671                 "movq   %%mm4, %0       \n\t"
1672                 "movq   %%mm5, %1       \n\t"
1673                 "psrlw    $12, %%mm4    \n\t"
1674                 "psrlw    $12, %%mm5    \n\t"
1675                 : "+m"(*dx4), "+m"(*dy4)
1676                 : "m"(*dxy4), "m"(*dyy4)
1677             );
1678
1679             __asm__ volatile (
1680                 "movq      %%mm6, %%mm2 \n\t"
1681                 "movq      %%mm6, %%mm1 \n\t"
1682                 "psubw     %%mm4, %%mm2 \n\t"
1683                 "psubw     %%mm5, %%mm1 \n\t"
1684                 "movq      %%mm2, %%mm0 \n\t"
1685                 "movq      %%mm4, %%mm3 \n\t"
1686                 "pmullw    %%mm1, %%mm0 \n\t" // (s - dx) * (s - dy)
1687                 "pmullw    %%mm5, %%mm3 \n\t" // dx * dy
1688                 "pmullw    %%mm5, %%mm2 \n\t" // (s - dx) * dy
1689                 "pmullw    %%mm4, %%mm1 \n\t" // dx * (s - dy)
1690
1691                 "movd         %4, %%mm5 \n\t"
1692                 "movd         %3, %%mm4 \n\t"
1693                 "punpcklbw %%mm7, %%mm5 \n\t"
1694                 "punpcklbw %%mm7, %%mm4 \n\t"
1695                 "pmullw    %%mm5, %%mm3 \n\t" // src[1, 1] * dx * dy
1696                 "pmullw    %%mm4, %%mm2 \n\t" // src[0, 1] * (s - dx) * dy
1697
1698                 "movd         %2, %%mm5 \n\t"
1699                 "movd         %1, %%mm4 \n\t"
1700                 "punpcklbw %%mm7, %%mm5 \n\t"
1701                 "punpcklbw %%mm7, %%mm4 \n\t"
1702                 "pmullw    %%mm5, %%mm1 \n\t" // src[1, 0] * dx * (s - dy)
1703                 "pmullw    %%mm4, %%mm0 \n\t" // src[0, 0] * (s - dx) * (s - dy)
1704                 "paddw        %5, %%mm1 \n\t"
1705                 "paddw     %%mm3, %%mm2 \n\t"
1706                 "paddw     %%mm1, %%mm0 \n\t"
1707                 "paddw     %%mm2, %%mm0 \n\t"
1708
1709                 "psrlw        %6, %%mm0 \n\t"
1710                 "packuswb  %%mm0, %%mm0 \n\t"
1711                 "movd      %%mm0, %0    \n\t"
1712
1713                 : "=m"(dst[x + y * stride])
1714                 : "m"(src[0]), "m"(src[1]),
1715                   "m"(src[stride]), "m"(src[stride + 1]),
1716                   "m"(*r4), "m"(shift2)
1717             );
1718             src += stride;
1719         }
1720         src += 4 - h * stride;
1721     }
1722 }
1723
1724 #if CONFIG_VIDEODSP
1725 #if HAVE_YASM
1726 #if ARCH_X86_32
1727 static void gmc_mmx(uint8_t *dst, uint8_t *src,
1728                     int stride, int h, int ox, int oy,
1729                     int dxx, int dxy, int dyx, int dyy,
1730                     int shift, int r, int width, int height)
1731 {
1732     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,
1733         width, height, &ff_emulated_edge_mc_8);
1734 }
1735 #endif
1736 static void gmc_sse(uint8_t *dst, uint8_t *src,
1737                     int stride, int h, int ox, int oy,
1738                     int dxx, int dxy, int dyx, int dyy,
1739                     int shift, int r, int width, int height)
1740 {
1741     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,
1742         width, height, &ff_emulated_edge_mc_8);
1743 }
1744 #else
1745 static void gmc_mmx(uint8_t *dst, uint8_t *src,
1746                     int stride, int h, int ox, int oy,
1747                     int dxx, int dxy, int dyx, int dyy,
1748                     int shift, int r, int width, int height)
1749 {
1750     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,
1751         width, height, &ff_emulated_edge_mc_8);
1752 }
1753 #endif
1754 #endif
1755
1756 #endif /* HAVE_INLINE_ASM */
1757
1758 #include "h264_qpel.c"
1759
1760 void ff_put_h264_chroma_mc8_rnd_mmx  (uint8_t *dst, uint8_t *src,
1761                                       int stride, int h, int x, int y);
1762 void ff_avg_h264_chroma_mc8_rnd_mmxext(uint8_t *dst, uint8_t *src,
1763                                        int stride, int h, int x, int y);
1764 void ff_avg_h264_chroma_mc8_rnd_3dnow(uint8_t *dst, uint8_t *src,
1765                                       int stride, int h, int x, int y);
1766
1767 void ff_put_h264_chroma_mc4_mmx      (uint8_t *dst, uint8_t *src,
1768                                       int stride, int h, int x, int y);
1769 void ff_avg_h264_chroma_mc4_mmxext   (uint8_t *dst, uint8_t *src,
1770                                       int stride, int h, int x, int y);
1771 void ff_avg_h264_chroma_mc4_3dnow    (uint8_t *dst, uint8_t *src,
1772                                       int stride, int h, int x, int y);
1773
1774 void ff_put_h264_chroma_mc2_mmxext   (uint8_t *dst, uint8_t *src,
1775                                       int stride, int h, int x, int y);
1776 void ff_avg_h264_chroma_mc2_mmxext   (uint8_t *dst, uint8_t *src,
1777                                       int stride, int h, int x, int y);
1778
1779 void ff_put_h264_chroma_mc8_rnd_ssse3(uint8_t *dst, uint8_t *src,
1780                                       int stride, int h, int x, int y);
1781 void ff_put_h264_chroma_mc4_ssse3    (uint8_t *dst, uint8_t *src,
1782                                       int stride, int h, int x, int y);
1783
1784 void ff_avg_h264_chroma_mc8_rnd_ssse3(uint8_t *dst, uint8_t *src,
1785                                       int stride, int h, int x, int y);
1786 void ff_avg_h264_chroma_mc4_ssse3    (uint8_t *dst, uint8_t *src,
1787                                       int stride, int h, int x, int y);
1788
1789 #define CHROMA_MC(OP, NUM, DEPTH, OPT)                                  \
1790 void ff_ ## OP ## _h264_chroma_mc ## NUM ## _ ## DEPTH ## _ ## OPT      \
1791                                       (uint8_t *dst, uint8_t *src,      \
1792                                        int stride, int h, int x, int y);
1793
1794 CHROMA_MC(put, 2, 10, mmxext)
1795 CHROMA_MC(avg, 2, 10, mmxext)
1796 CHROMA_MC(put, 4, 10, mmxext)
1797 CHROMA_MC(avg, 4, 10, mmxext)
1798 CHROMA_MC(put, 8, 10, sse2)
1799 CHROMA_MC(avg, 8, 10, sse2)
1800 CHROMA_MC(put, 8, 10, avx)
1801 CHROMA_MC(avg, 8, 10, avx)
1802
1803 #if HAVE_INLINE_ASM
1804
1805 /* CAVS-specific */
1806 void ff_put_cavs_qpel8_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1807 {
1808     put_pixels8_mmx(dst, src, stride, 8);
1809 }
1810
1811 void ff_avg_cavs_qpel8_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1812 {
1813     avg_pixels8_mmx(dst, src, stride, 8);
1814 }
1815
1816 void ff_put_cavs_qpel16_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1817 {
1818     put_pixels16_mmx(dst, src, stride, 16);
1819 }
1820
1821 void ff_avg_cavs_qpel16_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1822 {
1823     avg_pixels16_mmx(dst, src, stride, 16);
1824 }
1825
1826 /* VC-1-specific */
1827 void ff_put_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src,
1828                                int stride, int rnd)
1829 {
1830     put_pixels8_mmx(dst, src, stride, 8);
1831 }
1832
1833 void ff_avg_vc1_mspel_mc00_mmxext(uint8_t *dst, const uint8_t *src,
1834                                   int stride, int rnd)
1835 {
1836     avg_pixels8_mmxext(dst, src, stride, 8);
1837 }
1838
1839 #if CONFIG_DIRAC_DECODER
1840 #define DIRAC_PIXOP(OPNAME, EXT)\
1841 void ff_ ## OPNAME ## _dirac_pixels8_ ## EXT(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
1842 {\
1843     OPNAME ## _pixels8_ ## EXT(dst, src[0], stride, h);\
1844 }\
1845 void ff_ ## OPNAME ## _dirac_pixels16_ ## EXT(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
1846 {\
1847     OPNAME ## _pixels16_ ## EXT(dst, src[0], stride, h);\
1848 }\
1849 void ff_ ## OPNAME ## _dirac_pixels32_ ## EXT(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
1850 {\
1851     OPNAME ## _pixels16_ ## EXT(dst   , src[0]   , stride, h);\
1852     OPNAME ## _pixels16_ ## EXT(dst+16, src[0]+16, stride, h);\
1853 }
1854
1855 DIRAC_PIXOP(put, mmx)
1856 DIRAC_PIXOP(avg, mmx)
1857 DIRAC_PIXOP(avg, mmxext)
1858
1859 #if HAVE_YASM
1860 void ff_put_dirac_pixels16_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h)
1861 {
1862     ff_put_pixels16_sse2(dst, src[0], stride, h);
1863 }
1864 void ff_avg_dirac_pixels16_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h)
1865 {
1866     ff_avg_pixels16_sse2(dst, src[0], stride, h);
1867 }
1868 void ff_put_dirac_pixels32_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h)
1869 {
1870     ff_put_pixels16_sse2(dst   , src[0]   , stride, h);
1871     ff_put_pixels16_sse2(dst+16, src[0]+16, stride, h);
1872 }
1873 void ff_avg_dirac_pixels32_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h)
1874 {
1875     ff_avg_pixels16_sse2(dst   , src[0]   , stride, h);
1876     ff_avg_pixels16_sse2(dst+16, src[0]+16, stride, h);
1877 }
1878 #endif
1879 #endif
1880
1881 /* XXX: Those functions should be suppressed ASAP when all IDCTs are
1882  * converted. */
1883 #if CONFIG_GPL
1884 static void ff_libmpeg2mmx_idct_put(uint8_t *dest, int line_size,
1885                                     int16_t *block)
1886 {
1887     ff_mmx_idct(block);
1888     ff_put_pixels_clamped_mmx(block, dest, line_size);
1889 }
1890
1891 static void ff_libmpeg2mmx_idct_add(uint8_t *dest, int line_size,
1892                                     int16_t *block)
1893 {
1894     ff_mmx_idct(block);
1895     ff_add_pixels_clamped_mmx(block, dest, line_size);
1896 }
1897
1898 static void ff_libmpeg2mmx2_idct_put(uint8_t *dest, int line_size,
1899                                      int16_t *block)
1900 {
1901     ff_mmxext_idct(block);
1902     ff_put_pixels_clamped_mmx(block, dest, line_size);
1903 }
1904
1905 static void ff_libmpeg2mmx2_idct_add(uint8_t *dest, int line_size,
1906                                      int16_t *block)
1907 {
1908     ff_mmxext_idct(block);
1909     ff_add_pixels_clamped_mmx(block, dest, line_size);
1910 }
1911 #endif
1912
1913 static void vector_clipf_sse(float *dst, const float *src,
1914                              float min, float max, int len)
1915 {
1916     x86_reg i = (len - 16) * 4;
1917     __asm__ volatile (
1918         "movss          %3, %%xmm4      \n\t"
1919         "movss          %4, %%xmm5      \n\t"
1920         "shufps $0, %%xmm4, %%xmm4      \n\t"
1921         "shufps $0, %%xmm5, %%xmm5      \n\t"
1922         "1:                             \n\t"
1923         "movaps   (%2, %0), %%xmm0      \n\t" // 3/1 on intel
1924         "movaps 16(%2, %0), %%xmm1      \n\t"
1925         "movaps 32(%2, %0), %%xmm2      \n\t"
1926         "movaps 48(%2, %0), %%xmm3      \n\t"
1927         "maxps      %%xmm4, %%xmm0      \n\t"
1928         "maxps      %%xmm4, %%xmm1      \n\t"
1929         "maxps      %%xmm4, %%xmm2      \n\t"
1930         "maxps      %%xmm4, %%xmm3      \n\t"
1931         "minps      %%xmm5, %%xmm0      \n\t"
1932         "minps      %%xmm5, %%xmm1      \n\t"
1933         "minps      %%xmm5, %%xmm2      \n\t"
1934         "minps      %%xmm5, %%xmm3      \n\t"
1935         "movaps     %%xmm0,   (%1, %0)  \n\t"
1936         "movaps     %%xmm1, 16(%1, %0)  \n\t"
1937         "movaps     %%xmm2, 32(%1, %0)  \n\t"
1938         "movaps     %%xmm3, 48(%1, %0)  \n\t"
1939         "sub           $64, %0          \n\t"
1940         "jge            1b              \n\t"
1941         : "+&r"(i)
1942         : "r"(dst), "r"(src), "m"(min), "m"(max)
1943         : "memory"
1944     );
1945 }
1946
1947 #endif /* HAVE_INLINE_ASM */
1948
1949 int32_t ff_scalarproduct_int16_mmxext(const int16_t *v1, const int16_t *v2,
1950                                       int order);
1951 int32_t ff_scalarproduct_int16_sse2(const int16_t *v1, const int16_t *v2,
1952                                     int order);
1953 int32_t ff_scalarproduct_and_madd_int16_mmxext(int16_t *v1, const int16_t *v2,
1954                                                const int16_t *v3,
1955                                                int order, int mul);
1956 int32_t ff_scalarproduct_and_madd_int16_sse2(int16_t *v1, const int16_t *v2,
1957                                              const int16_t *v3,
1958                                              int order, int mul);
1959 int32_t ff_scalarproduct_and_madd_int16_ssse3(int16_t *v1, const int16_t *v2,
1960                                               const int16_t *v3,
1961                                               int order, int mul);
1962
1963 void ff_apply_window_int16_round_mmxext(int16_t *output, const int16_t *input,
1964                                         const int16_t *window, unsigned int len);
1965 void ff_apply_window_int16_round_sse2(int16_t *output, const int16_t *input,
1966                                       const int16_t *window, unsigned int len);
1967 void ff_apply_window_int16_mmxext(int16_t *output, const int16_t *input,
1968                                   const int16_t *window, unsigned int len);
1969 void ff_apply_window_int16_sse2(int16_t *output, const int16_t *input,
1970                                 const int16_t *window, unsigned int len);
1971 void ff_apply_window_int16_ssse3(int16_t *output, const int16_t *input,
1972                                  const int16_t *window, unsigned int len);
1973 void ff_apply_window_int16_ssse3_atom(int16_t *output, const int16_t *input,
1974                                       const int16_t *window, unsigned int len);
1975
1976 void ff_bswap32_buf_ssse3(uint32_t *dst, const uint32_t *src, int w);
1977 void ff_bswap32_buf_sse2(uint32_t *dst, const uint32_t *src, int w);
1978
1979 void ff_add_hfyu_median_prediction_mmxext(uint8_t *dst, const uint8_t *top,
1980                                           const uint8_t *diff, int w,
1981                                           int *left, int *left_top);
1982 int  ff_add_hfyu_left_prediction_ssse3(uint8_t *dst, const uint8_t *src,
1983                                        int w, int left);
1984 int  ff_add_hfyu_left_prediction_sse4(uint8_t *dst, const uint8_t *src,
1985                                       int w, int left);
1986
1987 void ff_vector_clip_int32_mmx     (int32_t *dst, const int32_t *src,
1988                                    int32_t min, int32_t max, unsigned int len);
1989 void ff_vector_clip_int32_sse2    (int32_t *dst, const int32_t *src,
1990                                    int32_t min, int32_t max, unsigned int len);
1991 void ff_vector_clip_int32_int_sse2(int32_t *dst, const int32_t *src,
1992                                    int32_t min, int32_t max, unsigned int len);
1993 void ff_vector_clip_int32_sse4    (int32_t *dst, const int32_t *src,
1994                                    int32_t min, int32_t max, unsigned int len);
1995
1996 #define SET_QPEL_FUNCS(PFX, IDX, SIZE, CPU, PREFIX)                          \
1997     do {                                                                     \
1998     c->PFX ## _pixels_tab[IDX][ 0] = PREFIX ## PFX ## SIZE ## _mc00_ ## CPU; \
1999     c->PFX ## _pixels_tab[IDX][ 1] = PREFIX ## PFX ## SIZE ## _mc10_ ## CPU; \
2000     c->PFX ## _pixels_tab[IDX][ 2] = PREFIX ## PFX ## SIZE ## _mc20_ ## CPU; \
2001     c->PFX ## _pixels_tab[IDX][ 3] = PREFIX ## PFX ## SIZE ## _mc30_ ## CPU; \
2002     c->PFX ## _pixels_tab[IDX][ 4] = PREFIX ## PFX ## SIZE ## _mc01_ ## CPU; \
2003     c->PFX ## _pixels_tab[IDX][ 5] = PREFIX ## PFX ## SIZE ## _mc11_ ## CPU; \
2004     c->PFX ## _pixels_tab[IDX][ 6] = PREFIX ## PFX ## SIZE ## _mc21_ ## CPU; \
2005     c->PFX ## _pixels_tab[IDX][ 7] = PREFIX ## PFX ## SIZE ## _mc31_ ## CPU; \
2006     c->PFX ## _pixels_tab[IDX][ 8] = PREFIX ## PFX ## SIZE ## _mc02_ ## CPU; \
2007     c->PFX ## _pixels_tab[IDX][ 9] = PREFIX ## PFX ## SIZE ## _mc12_ ## CPU; \
2008     c->PFX ## _pixels_tab[IDX][10] = PREFIX ## PFX ## SIZE ## _mc22_ ## CPU; \
2009     c->PFX ## _pixels_tab[IDX][11] = PREFIX ## PFX ## SIZE ## _mc32_ ## CPU; \
2010     c->PFX ## _pixels_tab[IDX][12] = PREFIX ## PFX ## SIZE ## _mc03_ ## CPU; \
2011     c->PFX ## _pixels_tab[IDX][13] = PREFIX ## PFX ## SIZE ## _mc13_ ## CPU; \
2012     c->PFX ## _pixels_tab[IDX][14] = PREFIX ## PFX ## SIZE ## _mc23_ ## CPU; \
2013     c->PFX ## _pixels_tab[IDX][15] = PREFIX ## PFX ## SIZE ## _mc33_ ## CPU; \
2014     } while (0)
2015
2016 #define SET_HPEL_FUNCS(PFX, IDX, SIZE, CPU)                                     \
2017     do {                                                                        \
2018         c->PFX ## _pixels_tab[IDX][0] = PFX ## _pixels ## SIZE ## _     ## CPU; \
2019         c->PFX ## _pixels_tab[IDX][1] = PFX ## _pixels ## SIZE ## _x2_  ## CPU; \
2020         c->PFX ## _pixels_tab[IDX][2] = PFX ## _pixels ## SIZE ## _y2_  ## CPU; \
2021         c->PFX ## _pixels_tab[IDX][3] = PFX ## _pixels ## SIZE ## _xy2_ ## CPU; \
2022     } while (0)
2023
2024 #define H264_QPEL_FUNCS(x, y, CPU)                                                            \
2025     do {                                                                                      \
2026         c->put_h264_qpel_pixels_tab[0][x + y * 4] = put_h264_qpel16_mc ## x ## y ## _ ## CPU; \
2027         c->put_h264_qpel_pixels_tab[1][x + y * 4] = put_h264_qpel8_mc  ## x ## y ## _ ## CPU; \
2028         c->avg_h264_qpel_pixels_tab[0][x + y * 4] = avg_h264_qpel16_mc ## x ## y ## _ ## CPU; \
2029         c->avg_h264_qpel_pixels_tab[1][x + y * 4] = avg_h264_qpel8_mc  ## x ## y ## _ ## CPU; \
2030     } while (0)
2031
2032 #define H264_QPEL_FUNCS_10(x, y, CPU)                                                               \
2033     do {                                                                                            \
2034         c->put_h264_qpel_pixels_tab[0][x + y * 4] = ff_put_h264_qpel16_mc ## x ## y ## _10_ ## CPU; \
2035         c->put_h264_qpel_pixels_tab[1][x + y * 4] = ff_put_h264_qpel8_mc  ## x ## y ## _10_ ## CPU; \
2036         c->avg_h264_qpel_pixels_tab[0][x + y * 4] = ff_avg_h264_qpel16_mc ## x ## y ## _10_ ## CPU; \
2037         c->avg_h264_qpel_pixels_tab[1][x + y * 4] = ff_avg_h264_qpel8_mc  ## x ## y ## _10_ ## CPU; \
2038     } while (0)
2039
2040 static void dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx, int mm_flags)
2041 {
2042     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
2043
2044 #if HAVE_INLINE_ASM
2045     c->put_pixels_clamped        = ff_put_pixels_clamped_mmx;
2046     c->put_signed_pixels_clamped = ff_put_signed_pixels_clamped_mmx;
2047     c->add_pixels_clamped        = ff_add_pixels_clamped_mmx;
2048
2049     if (!high_bit_depth) {
2050         c->clear_block  = clear_block_mmx;
2051         c->clear_blocks = clear_blocks_mmx;
2052         c->draw_edges   = draw_edges_mmx;
2053
2054         SET_HPEL_FUNCS(put,        0, 16, mmx);
2055         SET_HPEL_FUNCS(put_no_rnd, 0, 16, mmx);
2056         SET_HPEL_FUNCS(avg,        0, 16, mmx);
2057         SET_HPEL_FUNCS(avg_no_rnd, 0, 16, mmx);
2058         SET_HPEL_FUNCS(put,        1,  8, mmx);
2059         SET_HPEL_FUNCS(put_no_rnd, 1,  8, mmx);
2060         SET_HPEL_FUNCS(avg,        1,  8, mmx);
2061         SET_HPEL_FUNCS(avg_no_rnd, 1,  8, mmx);
2062     }
2063
2064 #if ARCH_X86_32 || !HAVE_YASM
2065     c->gmc = gmc_mmx;
2066 #endif
2067
2068     c->add_bytes = add_bytes_mmx;
2069
2070     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
2071         c->h263_v_loop_filter = h263_v_loop_filter_mmx;
2072         c->h263_h_loop_filter = h263_h_loop_filter_mmx;
2073     }
2074 #endif /* HAVE_INLINE_ASM */
2075
2076 #if HAVE_YASM
2077     if (!high_bit_depth && CONFIG_H264CHROMA) {
2078         c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_rnd_mmx;
2079         c->put_h264_chroma_pixels_tab[1] = ff_put_h264_chroma_mc4_mmx;
2080     }
2081
2082     c->vector_clip_int32 = ff_vector_clip_int32_mmx;
2083 #endif
2084
2085 }
2086
2087 static void dsputil_init_mmxext(DSPContext *c, AVCodecContext *avctx,
2088                                 int mm_flags)
2089 {
2090     const int bit_depth      = avctx->bits_per_raw_sample;
2091     const int high_bit_depth = bit_depth > 8;
2092
2093 #if HAVE_INLINE_ASM
2094     SET_QPEL_FUNCS(avg_qpel,        0, 16, mmxext, );
2095     SET_QPEL_FUNCS(avg_qpel,        1,  8, mmxext, );
2096
2097     SET_QPEL_FUNCS(put_qpel,        0, 16, mmxext, );
2098     SET_QPEL_FUNCS(put_qpel,        1,  8, mmxext, );
2099     SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, mmxext, );
2100     SET_QPEL_FUNCS(put_no_rnd_qpel, 1,  8, mmxext, );
2101
2102     if (!high_bit_depth) {
2103         c->put_pixels_tab[0][1] = put_pixels16_x2_mmxext;
2104         c->put_pixels_tab[0][2] = put_pixels16_y2_mmxext;
2105
2106         c->avg_pixels_tab[0][0] = avg_pixels16_mmxext;
2107         c->avg_pixels_tab[0][1] = avg_pixels16_x2_mmxext;
2108         c->avg_pixels_tab[0][2] = avg_pixels16_y2_mmxext;
2109
2110         c->put_pixels_tab[1][1] = put_pixels8_x2_mmxext;
2111         c->put_pixels_tab[1][2] = put_pixels8_y2_mmxext;
2112
2113         c->avg_pixels_tab[1][0] = avg_pixels8_mmxext;
2114         c->avg_pixels_tab[1][1] = avg_pixels8_x2_mmxext;
2115         c->avg_pixels_tab[1][2] = avg_pixels8_y2_mmxext;
2116     }
2117
2118     if (!(avctx->flags & CODEC_FLAG_BITEXACT)) {
2119         if (!high_bit_depth) {
2120             c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmxext;
2121             c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmxext;
2122             c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmxext;
2123             c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmxext;
2124
2125             c->avg_pixels_tab[0][3] = avg_pixels16_xy2_mmxext;
2126             c->avg_pixels_tab[1][3] = avg_pixels8_xy2_mmxext;
2127         }
2128     }
2129
2130     if (CONFIG_VP3_DECODER && (avctx->codec_id == AV_CODEC_ID_VP3 ||
2131                                avctx->codec_id == AV_CODEC_ID_THEORA)) {
2132         c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_exact_mmxext;
2133         c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_exact_mmxext;
2134     }
2135 #endif /* HAVE_INLINE_ASM */
2136
2137 #if HAVE_MMXEXT_EXTERNAL
2138     if (CONFIG_H264QPEL) {
2139         if (!high_bit_depth) {
2140             SET_QPEL_FUNCS(put_h264_qpel, 0, 16, mmxext, );
2141             SET_QPEL_FUNCS(put_h264_qpel, 1,  8, mmxext, );
2142             SET_QPEL_FUNCS(put_h264_qpel, 2,  4, mmxext, );
2143             SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, mmxext, );
2144             SET_QPEL_FUNCS(avg_h264_qpel, 1,  8, mmxext, );
2145             SET_QPEL_FUNCS(avg_h264_qpel, 2,  4, mmxext, );
2146         } else if (bit_depth == 10) {
2147 #if !ARCH_X86_64
2148             SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, 10_mmxext, ff_);
2149             SET_QPEL_FUNCS(put_h264_qpel, 0, 16, 10_mmxext, ff_);
2150             SET_QPEL_FUNCS(put_h264_qpel, 1,  8, 10_mmxext, ff_);
2151             SET_QPEL_FUNCS(avg_h264_qpel, 1,  8, 10_mmxext, ff_);
2152 #endif
2153             SET_QPEL_FUNCS(put_h264_qpel, 2, 4,  10_mmxext, ff_);
2154             SET_QPEL_FUNCS(avg_h264_qpel, 2, 4,  10_mmxext, ff_);
2155         }
2156     }
2157
2158     if (!high_bit_depth && CONFIG_H264CHROMA) {
2159         c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_rnd_mmxext;
2160         c->avg_h264_chroma_pixels_tab[1] = ff_avg_h264_chroma_mc4_mmxext;
2161         c->avg_h264_chroma_pixels_tab[2] = ff_avg_h264_chroma_mc2_mmxext;
2162         c->put_h264_chroma_pixels_tab[2] = ff_put_h264_chroma_mc2_mmxext;
2163     }
2164     if (bit_depth == 10 && CONFIG_H264CHROMA) {
2165         c->put_h264_chroma_pixels_tab[2] = ff_put_h264_chroma_mc2_10_mmxext;
2166         c->avg_h264_chroma_pixels_tab[2] = ff_avg_h264_chroma_mc2_10_mmxext;
2167         c->put_h264_chroma_pixels_tab[1] = ff_put_h264_chroma_mc4_10_mmxext;
2168         c->avg_h264_chroma_pixels_tab[1] = ff_avg_h264_chroma_mc4_10_mmxext;
2169     }
2170
2171     /* slower than cmov version on AMD */
2172     if (!(mm_flags & AV_CPU_FLAG_3DNOW))
2173         c->add_hfyu_median_prediction = ff_add_hfyu_median_prediction_mmxext;
2174
2175     c->scalarproduct_int16          = ff_scalarproduct_int16_mmxext;
2176     c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_mmxext;
2177
2178     if (avctx->flags & CODEC_FLAG_BITEXACT) {
2179         c->apply_window_int16 = ff_apply_window_int16_mmxext;
2180     } else {
2181         c->apply_window_int16 = ff_apply_window_int16_round_mmxext;
2182     }
2183 #endif /* HAVE_MMXEXT_EXTERNAL */
2184 }
2185
2186 static void dsputil_init_3dnow(DSPContext *c, AVCodecContext *avctx,
2187                                int mm_flags)
2188 {
2189     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
2190
2191 #if HAVE_INLINE_ASM
2192     if (!high_bit_depth) {
2193         c->put_pixels_tab[0][1] = put_pixels16_x2_3dnow;
2194         c->put_pixels_tab[0][2] = put_pixels16_y2_3dnow;
2195
2196         c->avg_pixels_tab[0][0] = avg_pixels16_3dnow;
2197         c->avg_pixels_tab[0][1] = avg_pixels16_x2_3dnow;
2198         c->avg_pixels_tab[0][2] = avg_pixels16_y2_3dnow;
2199
2200         c->put_pixels_tab[1][1] = put_pixels8_x2_3dnow;
2201         c->put_pixels_tab[1][2] = put_pixels8_y2_3dnow;
2202
2203         c->avg_pixels_tab[1][0] = avg_pixels8_3dnow;
2204         c->avg_pixels_tab[1][1] = avg_pixels8_x2_3dnow;
2205         c->avg_pixels_tab[1][2] = avg_pixels8_y2_3dnow;
2206
2207         if (!(avctx->flags & CODEC_FLAG_BITEXACT)){
2208             c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_3dnow;
2209             c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_3dnow;
2210             c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_3dnow;
2211             c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_3dnow;
2212
2213             c->avg_pixels_tab[0][3] = avg_pixels16_xy2_3dnow;
2214             c->avg_pixels_tab[1][3] = avg_pixels8_xy2_3dnow;
2215         }
2216     }
2217
2218     if (CONFIG_VP3_DECODER && (avctx->codec_id == AV_CODEC_ID_VP3 ||
2219                                avctx->codec_id == AV_CODEC_ID_THEORA)) {
2220         c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_exact_3dnow;
2221         c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_exact_3dnow;
2222     }
2223 #endif /* HAVE_INLINE_ASM */
2224
2225 #if HAVE_YASM
2226     if (!high_bit_depth && CONFIG_H264CHROMA) {
2227         c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_rnd_3dnow;
2228         c->avg_h264_chroma_pixels_tab[1] = ff_avg_h264_chroma_mc4_3dnow;
2229     }
2230 #endif /* HAVE_YASM */
2231 }
2232
2233 static void dsputil_init_sse(DSPContext *c, AVCodecContext *avctx, int mm_flags)
2234 {
2235     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
2236
2237 #if HAVE_INLINE_ASM
2238     if (!high_bit_depth) {
2239         if (!(CONFIG_MPEG_XVMC_DECODER && avctx->xvmc_acceleration > 1)) {
2240             /* XvMCCreateBlocks() may not allocate 16-byte aligned blocks */
2241             c->clear_block  = clear_block_sse;
2242             c->clear_blocks = clear_blocks_sse;
2243         }
2244     }
2245
2246     c->vector_clipf = vector_clipf_sse;
2247 #endif /* HAVE_INLINE_ASM */
2248
2249 #if HAVE_YASM
2250 #if HAVE_INLINE_ASM && CONFIG_VIDEODSP
2251     c->gmc = gmc_sse;
2252 #endif
2253 #endif /* HAVE_YASM */
2254 }
2255
2256 static void dsputil_init_sse2(DSPContext *c, AVCodecContext *avctx,
2257                               int mm_flags)
2258 {
2259     const int bit_depth      = avctx->bits_per_raw_sample;
2260     const int high_bit_depth = bit_depth > 8;
2261
2262 #if HAVE_SSE2_INLINE
2263     if (!high_bit_depth && avctx->idct_algo == FF_IDCT_XVIDMMX) {
2264         c->idct_put              = ff_idct_xvid_sse2_put;
2265         c->idct_add              = ff_idct_xvid_sse2_add;
2266         c->idct                  = ff_idct_xvid_sse2;
2267         c->idct_permutation_type = FF_SSE2_IDCT_PERM;
2268     }
2269 #endif /* HAVE_SSE2_INLINE */
2270
2271 #if HAVE_SSE2_EXTERNAL
2272     if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {
2273         // these functions are slower than mmx on AMD, but faster on Intel
2274         if (!high_bit_depth) {
2275             c->put_pixels_tab[0][0]        = ff_put_pixels16_sse2;
2276             c->put_no_rnd_pixels_tab[0][0] = ff_put_pixels16_sse2;
2277             c->avg_pixels_tab[0][0]        = ff_avg_pixels16_sse2;
2278             if (CONFIG_H264QPEL)
2279                 H264_QPEL_FUNCS(0, 0, sse2);
2280         }
2281     }
2282
2283     if (!high_bit_depth && CONFIG_H264QPEL) {
2284         H264_QPEL_FUNCS(0, 1, sse2);
2285         H264_QPEL_FUNCS(0, 2, sse2);
2286         H264_QPEL_FUNCS(0, 3, sse2);
2287         H264_QPEL_FUNCS(1, 1, sse2);
2288         H264_QPEL_FUNCS(1, 2, sse2);
2289         H264_QPEL_FUNCS(1, 3, sse2);
2290         H264_QPEL_FUNCS(2, 1, sse2);
2291         H264_QPEL_FUNCS(2, 2, sse2);
2292         H264_QPEL_FUNCS(2, 3, sse2);
2293         H264_QPEL_FUNCS(3, 1, sse2);
2294         H264_QPEL_FUNCS(3, 2, sse2);
2295         H264_QPEL_FUNCS(3, 3, sse2);
2296     }
2297
2298     if (bit_depth == 10) {
2299         if (CONFIG_H264QPEL) {
2300             SET_QPEL_FUNCS(put_h264_qpel, 0, 16, 10_sse2, ff_);
2301             SET_QPEL_FUNCS(put_h264_qpel, 1,  8, 10_sse2, ff_);
2302             SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, 10_sse2, ff_);
2303             SET_QPEL_FUNCS(avg_h264_qpel, 1,  8, 10_sse2, ff_);
2304             H264_QPEL_FUNCS_10(1, 0, sse2_cache64);
2305             H264_QPEL_FUNCS_10(2, 0, sse2_cache64);
2306             H264_QPEL_FUNCS_10(3, 0, sse2_cache64);
2307         }
2308         if (CONFIG_H264CHROMA) {
2309             c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_10_sse2;
2310             c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_10_sse2;
2311         }
2312     }
2313
2314     c->scalarproduct_int16          = ff_scalarproduct_int16_sse2;
2315     c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_sse2;
2316     if (mm_flags & AV_CPU_FLAG_ATOM) {
2317         c->vector_clip_int32 = ff_vector_clip_int32_int_sse2;
2318     } else {
2319         c->vector_clip_int32 = ff_vector_clip_int32_sse2;
2320     }
2321     if (avctx->flags & CODEC_FLAG_BITEXACT) {
2322         c->apply_window_int16 = ff_apply_window_int16_sse2;
2323     } else if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {
2324         c->apply_window_int16 = ff_apply_window_int16_round_sse2;
2325     }
2326     c->bswap_buf = ff_bswap32_buf_sse2;
2327 #endif /* HAVE_SSE2_EXTERNAL */
2328 }
2329
2330 static void dsputil_init_ssse3(DSPContext *c, AVCodecContext *avctx,
2331                                int mm_flags)
2332 {
2333 #if HAVE_SSSE3_EXTERNAL
2334     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
2335     const int bit_depth      = avctx->bits_per_raw_sample;
2336
2337     if (!high_bit_depth && CONFIG_H264QPEL) {
2338         H264_QPEL_FUNCS(1, 0, ssse3);
2339         H264_QPEL_FUNCS(1, 1, ssse3);
2340         H264_QPEL_FUNCS(1, 2, ssse3);
2341         H264_QPEL_FUNCS(1, 3, ssse3);
2342         H264_QPEL_FUNCS(2, 0, ssse3);
2343         H264_QPEL_FUNCS(2, 1, ssse3);
2344         H264_QPEL_FUNCS(2, 2, ssse3);
2345         H264_QPEL_FUNCS(2, 3, ssse3);
2346         H264_QPEL_FUNCS(3, 0, ssse3);
2347         H264_QPEL_FUNCS(3, 1, ssse3);
2348         H264_QPEL_FUNCS(3, 2, ssse3);
2349         H264_QPEL_FUNCS(3, 3, ssse3);
2350     }
2351     if (bit_depth == 10 && CONFIG_H264QPEL) {
2352         H264_QPEL_FUNCS_10(1, 0, ssse3_cache64);
2353         H264_QPEL_FUNCS_10(2, 0, ssse3_cache64);
2354         H264_QPEL_FUNCS_10(3, 0, ssse3_cache64);
2355     }
2356     if (!high_bit_depth && CONFIG_H264CHROMA) {
2357         c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_rnd_ssse3;
2358         c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_rnd_ssse3;
2359         c->put_h264_chroma_pixels_tab[1] = ff_put_h264_chroma_mc4_ssse3;
2360         c->avg_h264_chroma_pixels_tab[1] = ff_avg_h264_chroma_mc4_ssse3;
2361     }
2362     c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_ssse3;
2363     if (mm_flags & AV_CPU_FLAG_SSE4) // not really sse4, just slow on Conroe
2364         c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_sse4;
2365
2366     if (mm_flags & AV_CPU_FLAG_ATOM)
2367         c->apply_window_int16 = ff_apply_window_int16_ssse3_atom;
2368     else
2369         c->apply_window_int16 = ff_apply_window_int16_ssse3;
2370     if (!(mm_flags & (AV_CPU_FLAG_SSE42|AV_CPU_FLAG_3DNOW))) // cachesplit
2371         c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_ssse3;
2372     c->bswap_buf = ff_bswap32_buf_ssse3;
2373 #endif /* HAVE_SSSE3_EXTERNAL */
2374 }
2375
2376 static void dsputil_init_sse4(DSPContext *c, AVCodecContext *avctx,
2377                               int mm_flags)
2378 {
2379 #if HAVE_SSE4_EXTERNAL
2380     c->vector_clip_int32 = ff_vector_clip_int32_sse4;
2381 #endif /* HAVE_SSE4_EXTERNAL */
2382 }
2383
2384 static void dsputil_init_avx(DSPContext *c, AVCodecContext *avctx, int mm_flags)
2385 {
2386 #if HAVE_AVX_EXTERNAL
2387     const int bit_depth = avctx->bits_per_raw_sample;
2388
2389     if (bit_depth == 10) {
2390         // AVX implies !cache64.
2391         // TODO: Port cache(32|64) detection from x264.
2392         if (CONFIG_H264QPEL) {
2393             H264_QPEL_FUNCS_10(1, 0, sse2);
2394             H264_QPEL_FUNCS_10(2, 0, sse2);
2395             H264_QPEL_FUNCS_10(3, 0, sse2);
2396         }
2397
2398         if (CONFIG_H264CHROMA) {
2399             c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_10_avx;
2400             c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_10_avx;
2401         }
2402     }
2403 #endif /* HAVE_AVX_EXTERNAL */
2404 }
2405
2406 void ff_dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx)
2407 {
2408     int mm_flags = av_get_cpu_flags();
2409
2410 #if HAVE_7REGS && HAVE_INLINE_ASM
2411     if (mm_flags & AV_CPU_FLAG_CMOV)
2412         c->add_hfyu_median_prediction = add_hfyu_median_prediction_cmov;
2413 #endif
2414
2415     if (mm_flags & AV_CPU_FLAG_MMX) {
2416 #if HAVE_INLINE_ASM
2417         const int idct_algo = avctx->idct_algo;
2418
2419         if (avctx->lowres == 0 && avctx->bits_per_raw_sample <= 8) {
2420             if (idct_algo == FF_IDCT_AUTO || idct_algo == FF_IDCT_SIMPLEMMX) {
2421                 c->idct_put              = ff_simple_idct_put_mmx;
2422                 c->idct_add              = ff_simple_idct_add_mmx;
2423                 c->idct                  = ff_simple_idct_mmx;
2424                 c->idct_permutation_type = FF_SIMPLE_IDCT_PERM;
2425 #if CONFIG_GPL
2426             } else if (idct_algo == FF_IDCT_LIBMPEG2MMX) {
2427                 if (mm_flags & AV_CPU_FLAG_MMX2) {
2428                     c->idct_put = ff_libmpeg2mmx2_idct_put;
2429                     c->idct_add = ff_libmpeg2mmx2_idct_add;
2430                     c->idct     = ff_mmxext_idct;
2431                 } else {
2432                     c->idct_put = ff_libmpeg2mmx_idct_put;
2433                     c->idct_add = ff_libmpeg2mmx_idct_add;
2434                     c->idct     = ff_mmx_idct;
2435                 }
2436                 c->idct_permutation_type = FF_LIBMPEG2_IDCT_PERM;
2437 #endif
2438             } else if (idct_algo == FF_IDCT_XVIDMMX) {
2439                 if (mm_flags & AV_CPU_FLAG_SSE2) {
2440                     c->idct_put              = ff_idct_xvid_sse2_put;
2441                     c->idct_add              = ff_idct_xvid_sse2_add;
2442                     c->idct                  = ff_idct_xvid_sse2;
2443                     c->idct_permutation_type = FF_SSE2_IDCT_PERM;
2444                 } else if (mm_flags & AV_CPU_FLAG_MMXEXT) {
2445                     c->idct_put              = ff_idct_xvid_mmxext_put;
2446                     c->idct_add              = ff_idct_xvid_mmxext_add;
2447                     c->idct                  = ff_idct_xvid_mmxext;
2448                 } else {
2449                     c->idct_put              = ff_idct_xvid_mmx_put;
2450                     c->idct_add              = ff_idct_xvid_mmx_add;
2451                     c->idct                  = ff_idct_xvid_mmx;
2452                 }
2453             }
2454         }
2455 #endif /* HAVE_INLINE_ASM */
2456
2457         dsputil_init_mmx(c, avctx, mm_flags);
2458     }
2459
2460     if (mm_flags & AV_CPU_FLAG_MMXEXT)
2461         dsputil_init_mmxext(c, avctx, mm_flags);
2462
2463     if (mm_flags & AV_CPU_FLAG_3DNOW)
2464         dsputil_init_3dnow(c, avctx, mm_flags);
2465
2466     if (mm_flags & AV_CPU_FLAG_SSE)
2467         dsputil_init_sse(c, avctx, mm_flags);
2468
2469     if (mm_flags & AV_CPU_FLAG_SSE2)
2470         dsputil_init_sse2(c, avctx, mm_flags);
2471
2472     if (mm_flags & AV_CPU_FLAG_SSSE3)
2473         dsputil_init_ssse3(c, avctx, mm_flags);
2474
2475     if (mm_flags & AV_CPU_FLAG_SSE4)
2476         dsputil_init_sse4(c, avctx, mm_flags);
2477
2478     if (mm_flags & AV_CPU_FLAG_AVX)
2479         dsputil_init_avx(c, avctx, mm_flags);
2480
2481     if (CONFIG_ENCODERS)
2482         ff_dsputilenc_init_mmx(c, avctx);
2483 }