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 uint64_t, ff_pdw_80000000)[2] =
  43     { 0x8000000080000000ULL, 0x8000000080000000ULL };
  44
  45 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_1)    = { 0x0001000100010001ULL, 0x0001000100010001ULL };
  46 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_2)    = { 0x0002000200020002ULL, 0x0002000200020002ULL };
  47 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_3)    = { 0x0003000300030003ULL, 0x0003000300030003ULL };
  48 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_4)    = { 0x0004000400040004ULL, 0x0004000400040004ULL };
  49 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_5)    = { 0x0005000500050005ULL, 0x0005000500050005ULL };
  50 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_8)    = { 0x0008000800080008ULL, 0x0008000800080008ULL };
  51 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_9)    = { 0x0009000900090009ULL, 0x0009000900090009ULL };
  52 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_15)   =   0x000F000F000F000FULL;
  53 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_16)   = { 0x0010001000100010ULL, 0x0010001000100010ULL };
  54 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_17)   = { 0x0011001100110011ULL, 0x0011001100110011ULL };
  55 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_18)   = { 0x0012001200120012ULL, 0x0012001200120012ULL };
  56 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_20)   =   0x0014001400140014ULL;
  57 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_27)   = { 0x001B001B001B001BULL, 0x001B001B001B001BULL };
  58 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_28)   = { 0x001C001C001C001CULL, 0x001C001C001C001CULL };
  59 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_32)   = { 0x0020002000200020ULL, 0x0020002000200020ULL };
  60 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_42)   =   0x002A002A002A002AULL;
  61 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_53)   =   0x0035003500350035ULL;
  62 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_63)   = { 0x003F003F003F003FULL, 0x003F003F003F003FULL };
  63 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_64)   = { 0x0040004000400040ULL, 0x0040004000400040ULL };
  64 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_96)   =   0x0060006000600060ULL;
  65 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_128)  =   0x0080008000800080ULL;
  66 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_255)  =   0x00ff00ff00ff00ffULL;
  67 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_512)  = { 0x0200020002000200ULL, 0x0200020002000200ULL };
  68 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_1019) = { 0x03FB03FB03FB03FBULL, 0x03FB03FB03FB03FBULL };
  69
  70 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_0)    = { 0x0000000000000000ULL, 0x0000000000000000ULL };
  71 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_1)    = { 0x0101010101010101ULL, 0x0101010101010101ULL };
  72 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_3)    = { 0x0303030303030303ULL, 0x0303030303030303ULL };
  73 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_4)    = { 0x0404040404040404ULL, 0x0404040404040404ULL };
  74 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_7)    =   0x0707070707070707ULL;
  75 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_1F)   =   0x1F1F1F1F1F1F1F1FULL;
  76 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_3F)   =   0x3F3F3F3F3F3F3F3FULL;
  77 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_80)   = { 0x8080808080808080ULL, 0x8080808080808080ULL };
  78 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_81)   =   0x8181818181818181ULL;
  79 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_A1)   = { 0xA1A1A1A1A1A1A1A1ULL, 0xA1A1A1A1A1A1A1A1ULL };
  80 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_F8)   = { 0xF8F8F8F8F8F8F8F8ULL, 0xF8F8F8F8F8F8F8F8ULL };
  81 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_FC)   =   0xFCFCFCFCFCFCFCFCULL;
  82 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_FE)   = { 0xFEFEFEFEFEFEFEFEULL, 0xFEFEFEFEFEFEFEFEULL };
  83
  84 DECLARE_ALIGNED(16, const double, ff_pd_1)[2] = { 1.0, 1.0 };
  85 DECLARE_ALIGNED(16, const double, ff_pd_2)[2] = { 2.0, 2.0 };
  86
  87 #if HAVE_INLINE_ASM
  88
  89 #define JUMPALIGN()     __asm__ volatile (".p2align 3"::)
  90 #define MOVQ_ZERO(regd) __asm__ volatile ("pxor %%"#regd", %%"#regd ::)
  91
  92 #define MOVQ_BFE(regd)                                  \
  93     __asm__ volatile (                                  \
  94         "pcmpeqd %%"#regd", %%"#regd"   \n\t"           \
  95         "paddb   %%"#regd", %%"#regd"   \n\t" ::)
  96
  97 #ifndef PIC
  98 #define MOVQ_BONE(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_bone))
  99 #define MOVQ_WTWO(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_wtwo))
 100 #else
 101 // for shared library it's better to use this way for accessing constants
 102 // pcmpeqd -> -1
 103 #define MOVQ_BONE(regd)                                 \
 104     __asm__ volatile (                                  \
 105         "pcmpeqd  %%"#regd", %%"#regd"  \n\t"           \
 106         "psrlw          $15, %%"#regd"  \n\t"           \
 107         "packuswb %%"#regd", %%"#regd"  \n\t" ::)
 108
 109 #define MOVQ_WTWO(regd)                                 \
 110     __asm__ volatile (                                  \
 111         "pcmpeqd %%"#regd", %%"#regd"   \n\t"           \
 112         "psrlw         $15, %%"#regd"   \n\t"           \
 113         "psllw          $1, %%"#regd"   \n\t"::)
 114
 115 #endif
 116
 117 // using regr as temporary and for the output result
 118 // first argument is unmodifed and second is trashed
 119 // regfe is supposed to contain 0xfefefefefefefefe
 120 #define PAVGB_MMX_NO_RND(rega, regb, regr, regfe)                \
 121     "movq   "#rega", "#regr"            \n\t"                    \
 122     "pand   "#regb", "#regr"            \n\t"                    \
 123     "pxor   "#rega", "#regb"            \n\t"                    \
 124     "pand  "#regfe", "#regb"            \n\t"                    \
 125     "psrlq       $1, "#regb"            \n\t"                    \
 126     "paddb  "#regb", "#regr"            \n\t"
 127
 128 #define PAVGB_MMX(rega, regb, regr, regfe)                       \
 129     "movq   "#rega", "#regr"            \n\t"                    \
 130     "por    "#regb", "#regr"            \n\t"                    \
 131     "pxor   "#rega", "#regb"            \n\t"                    \
 132     "pand  "#regfe", "#regb"            \n\t"                    \
 133     "psrlq       $1, "#regb"            \n\t"                    \
 134     "psubb  "#regb", "#regr"            \n\t"
 135
 136 // mm6 is supposed to contain 0xfefefefefefefefe
 137 #define PAVGBP_MMX_NO_RND(rega, regb, regr,  regc, regd, regp)   \
 138     "movq  "#rega", "#regr"             \n\t"                    \
 139     "movq  "#regc", "#regp"             \n\t"                    \
 140     "pand  "#regb", "#regr"             \n\t"                    \
 141     "pand  "#regd", "#regp"             \n\t"                    \
 142     "pxor  "#rega", "#regb"             \n\t"                    \
 143     "pxor  "#regc", "#regd"             \n\t"                    \
 144     "pand    %%mm6, "#regb"             \n\t"                    \
 145     "pand    %%mm6, "#regd"             \n\t"                    \
 146     "psrlq      $1, "#regb"             \n\t"                    \
 147     "psrlq      $1, "#regd"             \n\t"                    \
 148     "paddb "#regb", "#regr"             \n\t"                    \
 149     "paddb "#regd", "#regp"             \n\t"
 150
 151 #define PAVGBP_MMX(rega, regb, regr, regc, regd, regp)           \
 152     "movq  "#rega", "#regr"             \n\t"                    \
 153     "movq  "#regc", "#regp"             \n\t"                    \
 154     "por   "#regb", "#regr"             \n\t"                    \
 155     "por   "#regd", "#regp"             \n\t"                    \
 156     "pxor  "#rega", "#regb"             \n\t"                    \
 157     "pxor  "#regc", "#regd"             \n\t"                    \
 158     "pand    %%mm6, "#regb"             \n\t"                    \
 159     "pand    %%mm6, "#regd"             \n\t"                    \
 160     "psrlq      $1, "#regd"             \n\t"                    \
 161     "psrlq      $1, "#regb"             \n\t"                    \
 162     "psubb "#regb", "#regr"             \n\t"                    \
 163     "psubb "#regd", "#regp"             \n\t"
 164
 165 /***********************************/
 166 /* MMX no rounding */
 167 #define DEF(x, y) x ## _no_rnd_ ## y ## _mmx
 168 #define SET_RND  MOVQ_WONE
 169 #define PAVGBP(a, b, c, d, e, f)        PAVGBP_MMX_NO_RND(a, b, c, d, e, f)
 170 #define PAVGB(a, b, c, e)               PAVGB_MMX_NO_RND(a, b, c, e)
 171 #define OP_AVG(a, b, c, e)              PAVGB_MMX(a, b, c, e)
 172
 173 #include "dsputil_rnd_template.c"
 174
 175 #undef DEF
 176 #undef SET_RND
 177 #undef PAVGBP
 178 #undef PAVGB
 179 /***********************************/
 180 /* MMX rounding */
 181
 182 #define DEF(x, y) x ## _ ## y ## _mmx
 183 #define SET_RND  MOVQ_WTWO
 184 #define PAVGBP(a, b, c, d, e, f)        PAVGBP_MMX(a, b, c, d, e, f)
 185 #define PAVGB(a, b, c, e)               PAVGB_MMX(a, b, c, e)
 186
 187 #include "dsputil_rnd_template.c"
 188
 189 #undef DEF
 190 #undef SET_RND
 191 #undef PAVGBP
 192 #undef PAVGB
 193 #undef OP_AVG
 194
 195 /***********************************/
 196 /* 3Dnow specific */
 197
 198 #define DEF(x) x ## _3dnow
 199 #define PAVGB "pavgusb"
 200 #define SKIP_FOR_3DNOW
 201
 202 #include "dsputil_avg_template.c"
 203
 204 #undef DEF
 205 #undef PAVGB
 206 #undef SKIP_FOR_3DNOW
 207
 208 /***********************************/
 209 /* MMXEXT specific */
 210
 211 #define DEF(x) x ## _mmxext
 212
 213 /* Introduced only in MMXEXT set */
 214 #define PAVGB "pavgb"
 215
 216 #include "dsputil_avg_template.c"
 217
 218 #undef DEF
 219 #undef PAVGB
 220
 221 #define put_no_rnd_pixels16_mmx put_pixels16_mmx
 222 #define put_no_rnd_pixels8_mmx put_pixels8_mmx
 223 #define put_pixels16_mmxext put_pixels16_mmx
 224 #define put_pixels8_mmxext put_pixels8_mmx
 225 #define put_pixels4_mmxext put_pixels4_mmx
 226 #define put_no_rnd_pixels16_mmxext put_no_rnd_pixels16_mmx
 227 #define put_no_rnd_pixels8_mmxext put_no_rnd_pixels8_mmx
 228
 229 /***********************************/
 230 /* standard MMX */
 231
 232 void ff_put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels,
 233                                int line_size)
 234 {
 235     const DCTELEM *p;
 236     uint8_t *pix;
 237
 238     /* read the pixels */
 239     p   = block;
 240     pix = pixels;
 241     /* unrolled loop */
 242     __asm__ volatile (
 243         "movq      (%3), %%mm0          \n\t"
 244         "movq     8(%3), %%mm1          \n\t"
 245         "movq    16(%3), %%mm2          \n\t"
 246         "movq    24(%3), %%mm3          \n\t"
 247         "movq    32(%3), %%mm4          \n\t"
 248         "movq    40(%3), %%mm5          \n\t"
 249         "movq    48(%3), %%mm6          \n\t"
 250         "movq    56(%3), %%mm7          \n\t"
 251         "packuswb %%mm1, %%mm0          \n\t"
 252         "packuswb %%mm3, %%mm2          \n\t"
 253         "packuswb %%mm5, %%mm4          \n\t"
 254         "packuswb %%mm7, %%mm6          \n\t"
 255         "movq     %%mm0, (%0)           \n\t"
 256         "movq     %%mm2, (%0, %1)       \n\t"
 257         "movq     %%mm4, (%0, %1, 2)    \n\t"
 258         "movq     %%mm6, (%0, %2)       \n\t"
 259         :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3),
 260            "r"(p)
 261         : "memory");
 262     pix += line_size * 4;
 263     p   += 32;
 264
 265     // if here would be an exact copy of the code above
 266     // compiler would generate some very strange code
 267     // thus using "r"
 268     __asm__ volatile (
 269         "movq       (%3), %%mm0         \n\t"
 270         "movq      8(%3), %%mm1         \n\t"
 271         "movq     16(%3), %%mm2         \n\t"
 272         "movq     24(%3), %%mm3         \n\t"
 273         "movq     32(%3), %%mm4         \n\t"
 274         "movq     40(%3), %%mm5         \n\t"
 275         "movq     48(%3), %%mm6         \n\t"
 276         "movq     56(%3), %%mm7         \n\t"
 277         "packuswb  %%mm1, %%mm0         \n\t"
 278         "packuswb  %%mm3, %%mm2         \n\t"
 279         "packuswb  %%mm5, %%mm4         \n\t"
 280         "packuswb  %%mm7, %%mm6         \n\t"
 281         "movq      %%mm0, (%0)          \n\t"
 282         "movq      %%mm2, (%0, %1)      \n\t"
 283         "movq      %%mm4, (%0, %1, 2)   \n\t"
 284         "movq      %%mm6, (%0, %2)      \n\t"
 285         :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3), "r"(p)
 286         : "memory");
 287 }
 288
 289 #define put_signed_pixels_clamped_mmx_half(off)             \
 290     "movq          "#off"(%2), %%mm1        \n\t"           \
 291     "movq     16 + "#off"(%2), %%mm2        \n\t"           \
 292     "movq     32 + "#off"(%2), %%mm3        \n\t"           \
 293     "movq     48 + "#off"(%2), %%mm4        \n\t"           \
 294     "packsswb  8 + "#off"(%2), %%mm1        \n\t"           \
 295     "packsswb 24 + "#off"(%2), %%mm2        \n\t"           \
 296     "packsswb 40 + "#off"(%2), %%mm3        \n\t"           \
 297     "packsswb 56 + "#off"(%2), %%mm4        \n\t"           \
 298     "paddb              %%mm0, %%mm1        \n\t"           \
 299     "paddb              %%mm0, %%mm2        \n\t"           \
 300     "paddb              %%mm0, %%mm3        \n\t"           \
 301     "paddb              %%mm0, %%mm4        \n\t"           \
 302     "movq               %%mm1, (%0)         \n\t"           \
 303     "movq               %%mm2, (%0, %3)     \n\t"           \
 304     "movq               %%mm3, (%0, %3, 2)  \n\t"           \
 305     "movq               %%mm4, (%0, %1)     \n\t"
 306
 307 void ff_put_signed_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels,
 308                                       int line_size)
 309 {
 310     x86_reg line_skip = line_size;
 311     x86_reg line_skip3;
 312
 313     __asm__ volatile (
 314         "movq "MANGLE(ff_pb_80)", %%mm0     \n\t"
 315         "lea         (%3, %3, 2), %1        \n\t"
 316         put_signed_pixels_clamped_mmx_half(0)
 317         "lea         (%0, %3, 4), %0        \n\t"
 318         put_signed_pixels_clamped_mmx_half(64)
 319         : "+&r"(pixels), "=&r"(line_skip3)
 320         : "r"(block), "r"(line_skip)
 321         : "memory");
 322 }
 323
 324 void ff_add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels,
 325                                int line_size)
 326 {
 327     const DCTELEM *p;
 328     uint8_t *pix;
 329     int i;
 330
 331     /* read the pixels */
 332     p   = block;
 333     pix = pixels;
 334     MOVQ_ZERO(mm7);
 335     i = 4;
 336     do {
 337         __asm__ volatile (
 338             "movq        (%2), %%mm0    \n\t"
 339             "movq       8(%2), %%mm1    \n\t"
 340             "movq      16(%2), %%mm2    \n\t"
 341             "movq      24(%2), %%mm3    \n\t"
 342             "movq          %0, %%mm4    \n\t"
 343             "movq          %1, %%mm6    \n\t"
 344             "movq       %%mm4, %%mm5    \n\t"
 345             "punpcklbw  %%mm7, %%mm4    \n\t"
 346             "punpckhbw  %%mm7, %%mm5    \n\t"
 347             "paddsw     %%mm4, %%mm0    \n\t"
 348             "paddsw     %%mm5, %%mm1    \n\t"
 349             "movq       %%mm6, %%mm5    \n\t"
 350             "punpcklbw  %%mm7, %%mm6    \n\t"
 351             "punpckhbw  %%mm7, %%mm5    \n\t"
 352             "paddsw     %%mm6, %%mm2    \n\t"
 353             "paddsw     %%mm5, %%mm3    \n\t"
 354             "packuswb   %%mm1, %%mm0    \n\t"
 355             "packuswb   %%mm3, %%mm2    \n\t"
 356             "movq       %%mm0, %0       \n\t"
 357             "movq       %%mm2, %1       \n\t"
 358             : "+m"(*pix), "+m"(*(pix + line_size))
 359             : "r"(p)
 360             : "memory");
 361         pix += line_size * 2;
 362         p   += 16;
 363     } while (--i);
 364 }
 365
 366 static void put_pixels8_mmx(uint8_t *block, const uint8_t *pixels,
 367                             int line_size, int h)
 368 {
 369     __asm__ volatile (
 370         "lea   (%3, %3), %%"REG_a"      \n\t"
 371         ".p2align     3                 \n\t"
 372         "1:                             \n\t"
 373         "movq  (%1    ), %%mm0          \n\t"
 374         "movq  (%1, %3), %%mm1          \n\t"
 375         "movq     %%mm0, (%2)           \n\t"
 376         "movq     %%mm1, (%2, %3)       \n\t"
 377         "add  %%"REG_a", %1             \n\t"
 378         "add  %%"REG_a", %2             \n\t"
 379         "movq  (%1    ), %%mm0          \n\t"
 380         "movq  (%1, %3), %%mm1          \n\t"
 381         "movq     %%mm0, (%2)           \n\t"
 382         "movq     %%mm1, (%2, %3)       \n\t"
 383         "add  %%"REG_a", %1             \n\t"
 384         "add  %%"REG_a", %2             \n\t"
 385         "subl        $4, %0             \n\t"
 386         "jnz         1b                 \n\t"
 387         : "+g"(h), "+r"(pixels),  "+r"(block)
 388         : "r"((x86_reg)line_size)
 389         : "%"REG_a, "memory"
 390         );
 391 }
 392
 393 static void put_pixels16_mmx(uint8_t *block, const uint8_t *pixels,
 394                              int line_size, int h)
 395 {
 396     __asm__ volatile (
 397         "lea   (%3, %3), %%"REG_a"      \n\t"
 398         ".p2align     3                 \n\t"
 399         "1:                             \n\t"
 400         "movq  (%1    ), %%mm0          \n\t"
 401         "movq 8(%1    ), %%mm4          \n\t"
 402         "movq  (%1, %3), %%mm1          \n\t"
 403         "movq 8(%1, %3), %%mm5          \n\t"
 404         "movq     %%mm0,  (%2)          \n\t"
 405         "movq     %%mm4, 8(%2)          \n\t"
 406         "movq     %%mm1,  (%2, %3)      \n\t"
 407         "movq     %%mm5, 8(%2, %3)      \n\t"
 408         "add  %%"REG_a", %1             \n\t"
 409         "add  %%"REG_a", %2             \n\t"
 410         "movq  (%1    ), %%mm0          \n\t"
 411         "movq 8(%1    ), %%mm4          \n\t"
 412         "movq  (%1, %3), %%mm1          \n\t"
 413         "movq 8(%1, %3), %%mm5          \n\t"
 414         "movq     %%mm0,  (%2)          \n\t"
 415         "movq     %%mm4, 8(%2)          \n\t"
 416         "movq     %%mm1,  (%2, %3)      \n\t"
 417         "movq     %%mm5, 8(%2, %3)      \n\t"
 418         "add  %%"REG_a", %1             \n\t"
 419         "add  %%"REG_a", %2             \n\t"
 420         "subl        $4, %0             \n\t"
 421         "jnz         1b                 \n\t"
 422         : "+g"(h), "+r"(pixels),  "+r"(block)
 423         : "r"((x86_reg)line_size)
 424         : "%"REG_a, "memory"
 425         );
 426 }
 427
 428 #define CLEAR_BLOCKS(name, n)                           \
 429 static void name(DCTELEM *blocks)                       \
 430 {                                                       \
 431     __asm__ volatile (                                  \
 432         "pxor %%mm7, %%mm7              \n\t"           \
 433         "mov     %1,        %%"REG_a"   \n\t"           \
 434         "1:                             \n\t"           \
 435         "movq %%mm7,   (%0, %%"REG_a")  \n\t"           \
 436         "movq %%mm7,  8(%0, %%"REG_a")  \n\t"           \
 437         "movq %%mm7, 16(%0, %%"REG_a")  \n\t"           \
 438         "movq %%mm7, 24(%0, %%"REG_a")  \n\t"           \
 439         "add    $32, %%"REG_a"          \n\t"           \
 440         "js      1b                     \n\t"           \
 441         :: "r"(((uint8_t *)blocks) + 128 * n),          \
 442            "i"(-128 * n)                                \
 443         : "%"REG_a                                      \
 444         );                                              \
 445 }
 446 CLEAR_BLOCKS(clear_blocks_mmx, 6)
 447 CLEAR_BLOCKS(clear_block_mmx, 1)
 448
 449 static void clear_block_sse(DCTELEM *block)
 450 {
 451     __asm__ volatile (
 452         "xorps  %%xmm0, %%xmm0          \n"
 453         "movaps %%xmm0,    (%0)         \n"
 454         "movaps %%xmm0,  16(%0)         \n"
 455         "movaps %%xmm0,  32(%0)         \n"
 456         "movaps %%xmm0,  48(%0)         \n"
 457         "movaps %%xmm0,  64(%0)         \n"
 458         "movaps %%xmm0,  80(%0)         \n"
 459         "movaps %%xmm0,  96(%0)         \n"
 460         "movaps %%xmm0, 112(%0)         \n"
 461         :: "r"(block)
 462         : "memory"
 463     );
 464 }
 465
 466 static void clear_blocks_sse(DCTELEM *blocks)
 467 {
 468     __asm__ volatile (
 469         "xorps  %%xmm0, %%xmm0              \n"
 470         "mov        %1,         %%"REG_a"   \n"
 471         "1:                                 \n"
 472         "movaps %%xmm0,    (%0, %%"REG_a")  \n"
 473         "movaps %%xmm0,  16(%0, %%"REG_a")  \n"
 474         "movaps %%xmm0,  32(%0, %%"REG_a")  \n"
 475         "movaps %%xmm0,  48(%0, %%"REG_a")  \n"
 476         "movaps %%xmm0,  64(%0, %%"REG_a")  \n"
 477         "movaps %%xmm0,  80(%0, %%"REG_a")  \n"
 478         "movaps %%xmm0,  96(%0, %%"REG_a")  \n"
 479         "movaps %%xmm0, 112(%0, %%"REG_a")  \n"
 480         "add      $128,         %%"REG_a"   \n"
 481         "js         1b                      \n"
 482         :: "r"(((uint8_t *)blocks) + 128 * 6),
 483            "i"(-128 * 6)
 484         : "%"REG_a
 485     );
 486 }
 487
 488 static void add_bytes_mmx(uint8_t *dst, uint8_t *src, int w)
 489 {
 490     x86_reg i = 0;
 491     __asm__ volatile (
 492         "jmp          2f                \n\t"
 493         "1:                             \n\t"
 494         "movq   (%1, %0), %%mm0         \n\t"
 495         "movq   (%2, %0), %%mm1         \n\t"
 496         "paddb     %%mm0, %%mm1         \n\t"
 497         "movq      %%mm1, (%2, %0)      \n\t"
 498         "movq  8(%1, %0), %%mm0         \n\t"
 499         "movq  8(%2, %0), %%mm1         \n\t"
 500         "paddb     %%mm0, %%mm1         \n\t"
 501         "movq      %%mm1, 8(%2, %0)     \n\t"
 502         "add         $16, %0            \n\t"
 503         "2:                             \n\t"
 504         "cmp          %3, %0            \n\t"
 505         "js           1b                \n\t"
 506         : "+r"(i)
 507         : "r"(src), "r"(dst), "r"((x86_reg)w - 15)
 508     );
 509     for ( ; i < w; i++)
 510         dst[i + 0] += src[i + 0];
 511 }
 512
 513 #if HAVE_7REGS
 514 static void add_hfyu_median_prediction_cmov(uint8_t *dst, const uint8_t *top,
 515                                             const uint8_t *diff, int w,
 516                                             int *left, int *left_top)
 517 {
 518     x86_reg w2 = -w;
 519     x86_reg x;
 520     int l  = *left     & 0xff;
 521     int tl = *left_top & 0xff;
 522     int t;
 523     __asm__ volatile (
 524         "mov          %7, %3            \n"
 525         "1:                             \n"
 526         "movzbl (%3, %4), %2            \n"
 527         "mov          %2, %k3           \n"
 528         "sub         %b1, %b3           \n"
 529         "add         %b0, %b3           \n"
 530         "mov          %2, %1            \n"
 531         "cmp          %0, %2            \n"
 532         "cmovg        %0, %2            \n"
 533         "cmovg        %1, %0            \n"
 534         "cmp         %k3, %0            \n"
 535         "cmovg       %k3, %0            \n"
 536         "mov          %7, %3            \n"
 537         "cmp          %2, %0            \n"
 538         "cmovl        %2, %0            \n"
 539         "add    (%6, %4), %b0           \n"
 540         "mov         %b0, (%5, %4)      \n"
 541         "inc          %4                \n"
 542         "jl           1b                \n"
 543         : "+&q"(l), "+&q"(tl), "=&r"(t), "=&q"(x), "+&r"(w2)
 544         : "r"(dst + w), "r"(diff + w), "rm"(top + w)
 545     );
 546     *left     = l;
 547     *left_top = tl;
 548 }
 549 #endif
 550
 551 static inline void transpose4x4(uint8_t *dst, uint8_t *src, x86_reg dst_stride, x86_reg src_stride){
 552     __asm__ volatile( //FIXME could save 1 instruction if done as 8x4 ...
 553         "movd  (%1), %%mm0              \n\t"
 554         "add   %3, %1                   \n\t"
 555         "movd  (%1), %%mm1              \n\t"
 556         "movd  (%1,%3,1), %%mm2         \n\t"
 557         "movd  (%1,%3,2), %%mm3         \n\t"
 558         "punpcklbw %%mm1, %%mm0         \n\t"
 559         "punpcklbw %%mm3, %%mm2         \n\t"
 560         "movq %%mm0, %%mm1              \n\t"
 561         "punpcklwd %%mm2, %%mm0         \n\t"
 562         "punpckhwd %%mm2, %%mm1         \n\t"
 563         "movd  %%mm0, (%0)              \n\t"
 564         "add   %2, %0                   \n\t"
 565         "punpckhdq %%mm0, %%mm0         \n\t"
 566         "movd  %%mm0, (%0)              \n\t"
 567         "movd  %%mm1, (%0,%2,1)         \n\t"
 568         "punpckhdq %%mm1, %%mm1         \n\t"
 569         "movd  %%mm1, (%0,%2,2)         \n\t"
 570
 571         :  "+&r" (dst),
 572            "+&r" (src)
 573         :  "r" (dst_stride),
 574            "r" (src_stride)
 575         :  "memory"
 576     );
 577 }
 578
 579 #define H263_LOOP_FILTER                        \
 580     "pxor      %%mm7, %%mm7             \n\t"   \
 581     "movq         %0, %%mm0             \n\t"   \
 582     "movq         %0, %%mm1             \n\t"   \
 583     "movq         %3, %%mm2             \n\t"   \
 584     "movq         %3, %%mm3             \n\t"   \
 585     "punpcklbw %%mm7, %%mm0             \n\t"   \
 586     "punpckhbw %%mm7, %%mm1             \n\t"   \
 587     "punpcklbw %%mm7, %%mm2             \n\t"   \
 588     "punpckhbw %%mm7, %%mm3             \n\t"   \
 589     "psubw     %%mm2, %%mm0             \n\t"   \
 590     "psubw     %%mm3, %%mm1             \n\t"   \
 591     "movq         %1, %%mm2             \n\t"   \
 592     "movq         %1, %%mm3             \n\t"   \
 593     "movq         %2, %%mm4             \n\t"   \
 594     "movq         %2, %%mm5             \n\t"   \
 595     "punpcklbw %%mm7, %%mm2             \n\t"   \
 596     "punpckhbw %%mm7, %%mm3             \n\t"   \
 597     "punpcklbw %%mm7, %%mm4             \n\t"   \
 598     "punpckhbw %%mm7, %%mm5             \n\t"   \
 599     "psubw     %%mm2, %%mm4             \n\t"   \
 600     "psubw     %%mm3, %%mm5             \n\t"   \
 601     "psllw        $2, %%mm4             \n\t"   \
 602     "psllw        $2, %%mm5             \n\t"   \
 603     "paddw     %%mm0, %%mm4             \n\t"   \
 604     "paddw     %%mm1, %%mm5             \n\t"   \
 605     "pxor      %%mm6, %%mm6             \n\t"   \
 606     "pcmpgtw   %%mm4, %%mm6             \n\t"   \
 607     "pcmpgtw   %%mm5, %%mm7             \n\t"   \
 608     "pxor      %%mm6, %%mm4             \n\t"   \
 609     "pxor      %%mm7, %%mm5             \n\t"   \
 610     "psubw     %%mm6, %%mm4             \n\t"   \
 611     "psubw     %%mm7, %%mm5             \n\t"   \
 612     "psrlw        $3, %%mm4             \n\t"   \
 613     "psrlw        $3, %%mm5             \n\t"   \
 614     "packuswb  %%mm5, %%mm4             \n\t"   \
 615     "packsswb  %%mm7, %%mm6             \n\t"   \
 616     "pxor      %%mm7, %%mm7             \n\t"   \
 617     "movd         %4, %%mm2             \n\t"   \
 618     "punpcklbw %%mm2, %%mm2             \n\t"   \
 619     "punpcklbw %%mm2, %%mm2             \n\t"   \
 620     "punpcklbw %%mm2, %%mm2             \n\t"   \
 621     "psubusb   %%mm4, %%mm2             \n\t"   \
 622     "movq      %%mm2, %%mm3             \n\t"   \
 623     "psubusb   %%mm4, %%mm3             \n\t"   \
 624     "psubb     %%mm3, %%mm2             \n\t"   \
 625     "movq         %1, %%mm3             \n\t"   \
 626     "movq         %2, %%mm4             \n\t"   \
 627     "pxor      %%mm6, %%mm3             \n\t"   \
 628     "pxor      %%mm6, %%mm4             \n\t"   \
 629     "paddusb   %%mm2, %%mm3             \n\t"   \
 630     "psubusb   %%mm2, %%mm4             \n\t"   \
 631     "pxor      %%mm6, %%mm3             \n\t"   \
 632     "pxor      %%mm6, %%mm4             \n\t"   \
 633     "paddusb   %%mm2, %%mm2             \n\t"   \
 634     "packsswb  %%mm1, %%mm0             \n\t"   \
 635     "pcmpgtb   %%mm0, %%mm7             \n\t"   \
 636     "pxor      %%mm7, %%mm0             \n\t"   \
 637     "psubb     %%mm7, %%mm0             \n\t"   \
 638     "movq      %%mm0, %%mm1             \n\t"   \
 639     "psubusb   %%mm2, %%mm0             \n\t"   \
 640     "psubb     %%mm0, %%mm1             \n\t"   \
 641     "pand         %5, %%mm1             \n\t"   \
 642     "psrlw        $2, %%mm1             \n\t"   \
 643     "pxor      %%mm7, %%mm1             \n\t"   \
 644     "psubb     %%mm7, %%mm1             \n\t"   \
 645     "movq         %0, %%mm5             \n\t"   \
 646     "movq         %3, %%mm6             \n\t"   \
 647     "psubb     %%mm1, %%mm5             \n\t"   \
 648     "paddb     %%mm1, %%mm6             \n\t"
 649
 650 static void h263_v_loop_filter_mmx(uint8_t *src, int stride, int qscale)
 651 {
 652     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
 653         const int strength = ff_h263_loop_filter_strength[qscale];
 654
 655         __asm__ volatile (
 656             H263_LOOP_FILTER
 657
 658             "movq %%mm3, %1             \n\t"
 659             "movq %%mm4, %2             \n\t"
 660             "movq %%mm5, %0             \n\t"
 661             "movq %%mm6, %3             \n\t"
 662             : "+m"(*(uint64_t*)(src - 2 * stride)),
 663               "+m"(*(uint64_t*)(src - 1 * stride)),
 664               "+m"(*(uint64_t*)(src + 0 * stride)),
 665               "+m"(*(uint64_t*)(src + 1 * stride))
 666             : "g"(2 * strength), "m"(ff_pb_FC)
 667             );
 668     }
 669 }
 670
 671 static void h263_h_loop_filter_mmx(uint8_t *src, int stride, int qscale)
 672 {
 673     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
 674         const int strength = ff_h263_loop_filter_strength[qscale];
 675         DECLARE_ALIGNED(8, uint64_t, temp)[4];
 676         uint8_t *btemp = (uint8_t*)temp;
 677
 678         src -= 2;
 679
 680         transpose4x4(btemp,     src,              8, stride);
 681         transpose4x4(btemp + 4, src + 4 * stride, 8, stride);
 682         __asm__ volatile (
 683             H263_LOOP_FILTER // 5 3 4 6
 684
 685             : "+m"(temp[0]),
 686               "+m"(temp[1]),
 687               "+m"(temp[2]),
 688               "+m"(temp[3])
 689             : "g"(2 * strength), "m"(ff_pb_FC)
 690             );
 691
 692         __asm__ volatile (
 693             "movq      %%mm5, %%mm1         \n\t"
 694             "movq      %%mm4, %%mm0         \n\t"
 695             "punpcklbw %%mm3, %%mm5         \n\t"
 696             "punpcklbw %%mm6, %%mm4         \n\t"
 697             "punpckhbw %%mm3, %%mm1         \n\t"
 698             "punpckhbw %%mm6, %%mm0         \n\t"
 699             "movq      %%mm5, %%mm3         \n\t"
 700             "movq      %%mm1, %%mm6         \n\t"
 701             "punpcklwd %%mm4, %%mm5         \n\t"
 702             "punpcklwd %%mm0, %%mm1         \n\t"
 703             "punpckhwd %%mm4, %%mm3         \n\t"
 704             "punpckhwd %%mm0, %%mm6         \n\t"
 705             "movd      %%mm5, (%0)          \n\t"
 706             "punpckhdq %%mm5, %%mm5         \n\t"
 707             "movd      %%mm5, (%0, %2)      \n\t"
 708             "movd      %%mm3, (%0, %2, 2)   \n\t"
 709             "punpckhdq %%mm3, %%mm3         \n\t"
 710             "movd      %%mm3, (%0, %3)      \n\t"
 711             "movd      %%mm1, (%1)          \n\t"
 712             "punpckhdq %%mm1, %%mm1         \n\t"
 713             "movd      %%mm1, (%1, %2)      \n\t"
 714             "movd      %%mm6, (%1, %2, 2)   \n\t"
 715             "punpckhdq %%mm6, %%mm6         \n\t"
 716             "movd      %%mm6, (%1, %3)      \n\t"
 717             :: "r"(src),
 718                "r"(src + 4 * stride),
 719                "r"((x86_reg)stride),
 720                "r"((x86_reg)(3 * stride))
 721             );
 722     }
 723 }
 724
 725 /* Draw the edges of width 'w' of an image of size width, height
 726  * this MMX version can only handle w == 8 || w == 16. */
 727 static void draw_edges_mmx(uint8_t *buf, int wrap, int width, int height,
 728                            int w, int h, int sides)
 729 {
 730     uint8_t *ptr, *last_line;
 731     int i;
 732
 733     last_line = buf + (height - 1) * wrap;
 734     /* left and right */
 735     ptr = buf;
 736     if (w == 8) {
 737         __asm__ volatile (
 738             "1:                             \n\t"
 739             "movd            (%0), %%mm0    \n\t"
 740             "punpcklbw      %%mm0, %%mm0    \n\t"
 741             "punpcklwd      %%mm0, %%mm0    \n\t"
 742             "punpckldq      %%mm0, %%mm0    \n\t"
 743             "movq           %%mm0, -8(%0)   \n\t"
 744             "movq      -8(%0, %2), %%mm1    \n\t"
 745             "punpckhbw      %%mm1, %%mm1    \n\t"
 746             "punpckhwd      %%mm1, %%mm1    \n\t"
 747             "punpckhdq      %%mm1, %%mm1    \n\t"
 748             "movq           %%mm1, (%0, %2) \n\t"
 749             "add               %1, %0       \n\t"
 750             "cmp               %3, %0       \n\t"
 751             "jb                1b           \n\t"
 752             : "+r"(ptr)
 753             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
 754             );
 755     } else if(w==16){
 756         __asm__ volatile (
 757             "1:                                 \n\t"
 758             "movd            (%0), %%mm0        \n\t"
 759             "punpcklbw      %%mm0, %%mm0        \n\t"
 760             "punpcklwd      %%mm0, %%mm0        \n\t"
 761             "punpckldq      %%mm0, %%mm0        \n\t"
 762             "movq           %%mm0, -8(%0)       \n\t"
 763             "movq           %%mm0, -16(%0)      \n\t"
 764             "movq      -8(%0, %2), %%mm1        \n\t"
 765             "punpckhbw      %%mm1, %%mm1        \n\t"
 766             "punpckhwd      %%mm1, %%mm1        \n\t"
 767             "punpckhdq      %%mm1, %%mm1        \n\t"
 768             "movq           %%mm1,  (%0, %2)    \n\t"
 769             "movq           %%mm1, 8(%0, %2)    \n\t"
 770             "add               %1, %0           \n\t"
 771             "cmp               %3, %0           \n\t"
 772             "jb                1b               \n\t"
 773             : "+r"(ptr)
 774             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
 775             );
 776     } else {
 777         av_assert1(w == 4);
 778         __asm__ volatile (
 779             "1:                             \n\t"
 780             "movd            (%0), %%mm0    \n\t"
 781             "punpcklbw      %%mm0, %%mm0    \n\t"
 782             "punpcklwd      %%mm0, %%mm0    \n\t"
 783             "movd           %%mm0, -4(%0)   \n\t"
 784             "movd      -4(%0, %2), %%mm1    \n\t"
 785             "punpcklbw      %%mm1, %%mm1    \n\t"
 786             "punpckhwd      %%mm1, %%mm1    \n\t"
 787             "punpckhdq      %%mm1, %%mm1    \n\t"
 788             "movd           %%mm1, (%0, %2) \n\t"
 789             "add               %1, %0       \n\t"
 790             "cmp               %3, %0       \n\t"
 791             "jb                1b           \n\t"
 792             : "+r"(ptr)
 793             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
 794             );
 795     }
 796
 797     /* top and bottom (and hopefully also the corners) */
 798     if (sides & EDGE_TOP) {
 799         for (i = 0; i < h; i += 4) {
 800             ptr = buf - (i + 1) * wrap - w;
 801             __asm__ volatile (
 802                 "1:                             \n\t"
 803                 "movq (%1, %0), %%mm0           \n\t"
 804                 "movq    %%mm0, (%0)            \n\t"
 805                 "movq    %%mm0, (%0, %2)        \n\t"
 806                 "movq    %%mm0, (%0, %2, 2)     \n\t"
 807                 "movq    %%mm0, (%0, %3)        \n\t"
 808                 "add        $8, %0              \n\t"
 809                 "cmp        %4, %0              \n\t"
 810                 "jb         1b                  \n\t"
 811                 : "+r"(ptr)
 812                 : "r"((x86_reg)buf - (x86_reg)ptr - w), "r"((x86_reg) -wrap),
 813                   "r"((x86_reg) -wrap * 3), "r"(ptr + width + 2 * w)
 814                 );
 815         }
 816     }
 817
 818     if (sides & EDGE_BOTTOM) {
 819         for (i = 0; i < h; i += 4) {
 820             ptr = last_line + (i + 1) * wrap - w;
 821             __asm__ volatile (
 822                 "1:                             \n\t"
 823                 "movq (%1, %0), %%mm0           \n\t"
 824                 "movq    %%mm0, (%0)            \n\t"
 825                 "movq    %%mm0, (%0, %2)        \n\t"
 826                 "movq    %%mm0, (%0, %2, 2)     \n\t"
 827                 "movq    %%mm0, (%0, %3)        \n\t"
 828                 "add        $8, %0              \n\t"
 829                 "cmp        %4, %0              \n\t"
 830                 "jb         1b                  \n\t"
 831                 : "+r"(ptr)
 832                 : "r"((x86_reg)last_line - (x86_reg)ptr - w),
 833                   "r"((x86_reg)wrap), "r"((x86_reg)wrap * 3),
 834                   "r"(ptr + width + 2 * w)
 835                 );
 836         }
 837     }
 838 }
 839
 840 #define QPEL_V_LOW(m3, m4, m5, m6, pw_20, pw_3, rnd,                      \
 841                    in0, in1, in2, in7, out, OP)                           \
 842     "paddw               "#m4", "#m3"   \n\t" /* x1 */                    \
 843     "movq   "MANGLE(ff_pw_20)", %%mm4   \n\t" /* 20 */                    \
 844     "pmullw              "#m3", %%mm4   \n\t" /* 20x1 */                  \
 845     "movq               "#in7", "#m3"   \n\t" /* d */                     \
 846     "movq               "#in0", %%mm5   \n\t" /* D */                     \
 847     "paddw               "#m3", %%mm5   \n\t" /* x4 */                    \
 848     "psubw               %%mm5, %%mm4   \n\t" /* 20x1 - x4 */             \
 849     "movq               "#in1", %%mm5   \n\t" /* C */                     \
 850     "movq               "#in2", %%mm6   \n\t" /* B */                     \
 851     "paddw               "#m6", %%mm5   \n\t" /* x3 */                    \
 852     "paddw               "#m5", %%mm6   \n\t" /* x2 */                    \
 853     "paddw               %%mm6, %%mm6   \n\t" /* 2x2 */                   \
 854     "psubw               %%mm6, %%mm5   \n\t" /* -2x2 + x3 */             \
 855     "pmullw  "MANGLE(ff_pw_3)", %%mm5   \n\t" /* -6x2 + 3x3 */            \
 856     "paddw              "#rnd", %%mm4   \n\t" /* x2 */                    \
 857     "paddw               %%mm4, %%mm5   \n\t" /* 20x1 - 6x2 + 3x3 - x4 */ \
 858     "psraw                  $5, %%mm5   \n\t"                             \
 859     "packuswb            %%mm5, %%mm5   \n\t"                             \
 860     OP(%%mm5, out, %%mm7, d)
 861
 862 #define QPEL_BASE(OPNAME, ROUNDER, RND, OP_MMXEXT)                        \
 863 static void OPNAME ## mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst,         \
 864                                                     uint8_t *src,         \
 865                                                     int dstStride,        \
 866                                                     int srcStride,        \
 867                                                     int h)                \
 868 {                                                                         \
 869     uint64_t temp;                                                        \
 870                                                                           \
 871     __asm__ volatile (                                                    \
 872         "pxor      %%mm7, %%mm7             \n\t"                         \
 873         "1:                                 \n\t"                         \
 874         "movq       (%0), %%mm0             \n\t" /* ABCDEFGH */          \
 875         "movq      %%mm0, %%mm1             \n\t" /* ABCDEFGH */          \
 876         "movq      %%mm0, %%mm2             \n\t" /* ABCDEFGH */          \
 877         "punpcklbw %%mm7, %%mm0             \n\t" /* 0A0B0C0D */          \
 878         "punpckhbw %%mm7, %%mm1             \n\t" /* 0E0F0G0H */          \
 879         "pshufw    $0x90, %%mm0, %%mm5      \n\t" /* 0A0A0B0C */          \
 880         "pshufw    $0x41, %%mm0, %%mm6      \n\t" /* 0B0A0A0B */          \
 881         "movq      %%mm2, %%mm3             \n\t" /* ABCDEFGH */          \
 882         "movq      %%mm2, %%mm4             \n\t" /* ABCDEFGH */          \
 883         "psllq        $8, %%mm2             \n\t" /* 0ABCDEFG */          \
 884         "psllq       $16, %%mm3             \n\t" /* 00ABCDEF */          \
 885         "psllq       $24, %%mm4             \n\t" /* 000ABCDE */          \
 886         "punpckhbw %%mm7, %%mm2             \n\t" /* 0D0E0F0G */          \
 887         "punpckhbw %%mm7, %%mm3             \n\t" /* 0C0D0E0F */          \
 888         "punpckhbw %%mm7, %%mm4             \n\t" /* 0B0C0D0E */          \
 889         "paddw     %%mm3, %%mm5             \n\t" /* b */                 \
 890         "paddw     %%mm2, %%mm6             \n\t" /* c */                 \
 891         "paddw     %%mm5, %%mm5             \n\t" /* 2b */                \
 892         "psubw     %%mm5, %%mm6             \n\t" /* c - 2b */            \
 893         "pshufw    $0x06, %%mm0, %%mm5      \n\t" /* 0C0B0A0A */          \
 894         "pmullw "MANGLE(ff_pw_3)", %%mm6    \n\t" /* 3c - 6b */           \
 895         "paddw     %%mm4, %%mm0             \n\t" /* a */                 \
 896         "paddw     %%mm1, %%mm5             \n\t" /* d */                 \
 897         "pmullw "MANGLE(ff_pw_20)", %%mm0   \n\t" /* 20a */               \
 898         "psubw     %%mm5, %%mm0             \n\t" /* 20a - d */           \
 899         "paddw        %6, %%mm6             \n\t"                         \
 900         "paddw     %%mm6, %%mm0             \n\t" /* 20a - 6b + 3c - d */ \
 901         "psraw        $5, %%mm0             \n\t"                         \
 902         "movq      %%mm0, %5                \n\t"                         \
 903         /* mm1 = EFGH, mm2 = DEFG, mm3 = CDEF, mm4 = BCDE, mm7 = 0 */     \
 904                                                                           \
 905         "movq      5(%0), %%mm0             \n\t" /* FGHIJKLM */          \
 906         "movq      %%mm0, %%mm5             \n\t" /* FGHIJKLM */          \
 907         "movq      %%mm0, %%mm6             \n\t" /* FGHIJKLM */          \
 908         "psrlq        $8, %%mm0             \n\t" /* GHIJKLM0 */          \
 909         "psrlq       $16, %%mm5             \n\t" /* HIJKLM00 */          \
 910         "punpcklbw %%mm7, %%mm0             \n\t" /* 0G0H0I0J */          \
 911         "punpcklbw %%mm7, %%mm5             \n\t" /* 0H0I0J0K */          \
 912         "paddw     %%mm0, %%mm2             \n\t" /* b */                 \
 913         "paddw     %%mm5, %%mm3             \n\t" /* c */                 \
 914         "paddw     %%mm2, %%mm2             \n\t" /* 2b */                \
 915         "psubw     %%mm2, %%mm3             \n\t" /* c - 2b */            \
 916         "movq      %%mm6, %%mm2             \n\t" /* FGHIJKLM */          \
 917         "psrlq       $24, %%mm6             \n\t" /* IJKLM000 */          \
 918         "punpcklbw %%mm7, %%mm2             \n\t" /* 0F0G0H0I */          \
 919         "punpcklbw %%mm7, %%mm6             \n\t" /* 0I0J0K0L */          \
 920         "pmullw "MANGLE(ff_pw_3)", %%mm3    \n\t" /* 3c - 6b */           \
 921         "paddw     %%mm2, %%mm1             \n\t" /* a */                 \
 922         "paddw     %%mm6, %%mm4             \n\t" /* d */                 \
 923         "pmullw "MANGLE(ff_pw_20)", %%mm1   \n\t" /* 20a */               \
 924         "psubw     %%mm4, %%mm3             \n\t" /* - 6b +3c - d */      \
 925         "paddw        %6, %%mm1             \n\t"                         \
 926         "paddw     %%mm1, %%mm3             \n\t" /* 20a - 6b +3c - d */  \
 927         "psraw        $5, %%mm3             \n\t"                         \
 928         "movq         %5, %%mm1             \n\t"                         \
 929         "packuswb  %%mm3, %%mm1             \n\t"                         \
 930         OP_MMXEXT(%%mm1, (%1), %%mm4, q)                                  \
 931         /* mm0 = GHIJ, mm2 = FGHI, mm5 = HIJK, mm6 = IJKL, mm7 = 0 */     \
 932                                                                           \
 933         "movq      9(%0), %%mm1             \n\t" /* JKLMNOPQ */          \
 934         "movq      %%mm1, %%mm4             \n\t" /* JKLMNOPQ */          \
 935         "movq      %%mm1, %%mm3             \n\t" /* JKLMNOPQ */          \
 936         "psrlq        $8, %%mm1             \n\t" /* KLMNOPQ0 */          \
 937         "psrlq       $16, %%mm4             \n\t" /* LMNOPQ00 */          \
 938         "punpcklbw %%mm7, %%mm1             \n\t" /* 0K0L0M0N */          \
 939         "punpcklbw %%mm7, %%mm4             \n\t" /* 0L0M0N0O */          \
 940         "paddw     %%mm1, %%mm5             \n\t" /* b */                 \
 941         "paddw     %%mm4, %%mm0             \n\t" /* c */                 \
 942         "paddw     %%mm5, %%mm5             \n\t" /* 2b */                \
 943         "psubw     %%mm5, %%mm0             \n\t" /* c - 2b */            \
 944         "movq      %%mm3, %%mm5             \n\t" /* JKLMNOPQ */          \
 945         "psrlq       $24, %%mm3             \n\t" /* MNOPQ000 */          \
 946         "pmullw "MANGLE(ff_pw_3)", %%mm0    \n\t" /* 3c - 6b */           \
 947         "punpcklbw %%mm7, %%mm3             \n\t" /* 0M0N0O0P */          \
 948         "paddw     %%mm3, %%mm2             \n\t" /* d */                 \
 949         "psubw     %%mm2, %%mm0             \n\t" /* -6b + 3c - d */      \
 950         "movq      %%mm5, %%mm2             \n\t" /* JKLMNOPQ */          \
 951         "punpcklbw %%mm7, %%mm2             \n\t" /* 0J0K0L0M */          \
 952         "punpckhbw %%mm7, %%mm5             \n\t" /* 0N0O0P0Q */          \
 953         "paddw     %%mm2, %%mm6             \n\t" /* a */                 \
 954         "pmullw "MANGLE(ff_pw_20)", %%mm6   \n\t" /* 20a */               \
 955         "paddw        %6, %%mm0             \n\t"                         \
 956         "paddw     %%mm6, %%mm0             \n\t" /* 20a - 6b + 3c - d */ \
 957         "psraw        $5, %%mm0             \n\t"                         \
 958         /* mm1 = KLMN, mm2 = JKLM, mm3 = MNOP, */                         \
 959         /* mm4 = LMNO, mm5 = NOPQ mm7 = 0 */                              \
 960                                                                           \
 961         "paddw    %%mm5, %%mm3              \n\t" /* a */                 \
 962         "pshufw   $0xF9, %%mm5, %%mm6       \n\t" /* 0O0P0Q0Q */          \
 963         "paddw    %%mm4, %%mm6              \n\t" /* b */                 \
 964         "pshufw   $0xBE, %%mm5, %%mm4       \n\t" /* 0P0Q0Q0P */          \
 965         "pshufw   $0x6F, %%mm5, %%mm5       \n\t" /* 0Q0Q0P0O */          \
 966         "paddw    %%mm1, %%mm4              \n\t" /* c */                 \
 967         "paddw    %%mm2, %%mm5              \n\t" /* d */                 \
 968         "paddw    %%mm6, %%mm6              \n\t" /* 2b */                \
 969         "psubw    %%mm6, %%mm4              \n\t" /* c - 2b */            \
 970         "pmullw "MANGLE(ff_pw_20)", %%mm3   \n\t" /* 20a */               \
 971         "pmullw  "MANGLE(ff_pw_3)", %%mm4   \n\t" /* 3c - 6b */           \
 972         "psubw    %%mm5, %%mm3              \n\t" /* -6b + 3c - d */      \
 973         "paddw       %6, %%mm4              \n\t"                         \
 974         "paddw    %%mm3, %%mm4              \n\t" /* 20a - 6b + 3c - d */ \
 975         "psraw       $5, %%mm4              \n\t"                         \
 976         "packuswb %%mm4, %%mm0              \n\t"                         \
 977         OP_MMXEXT(%%mm0, 8(%1), %%mm4, q)                                 \
 978                                                                           \
 979         "add         %3, %0                 \n\t"                         \
 980         "add         %4, %1                 \n\t"                         \
 981         "decl        %2                     \n\t"                         \
 982         "jnz         1b                     \n\t"                         \
 983         : "+a"(src), "+c"(dst), "+D"(h)                                   \
 984         : "d"((x86_reg)srcStride), "S"((x86_reg)dstStride),               \
 985           /* "m"(ff_pw_20), "m"(ff_pw_3), */ "m"(temp), "m"(ROUNDER)      \
 986         : "memory"                                                        \
 987         );                                                                \
 988 }                                                                         \
 989                                                                           \
 990 static void OPNAME ## mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst,          \
 991                                                    uint8_t *src,          \
 992                                                    int dstStride,         \
 993                                                    int srcStride,         \
 994                                                    int h)                 \
 995 {                                                                         \
 996     __asm__ volatile (                                                    \
 997         "pxor      %%mm7, %%mm7             \n\t"                         \
 998         "1:                                 \n\t"                         \
 999         "movq       (%0), %%mm0             \n\t" /* ABCDEFGH */          \
1000         "movq      %%mm0, %%mm1             \n\t" /* ABCDEFGH */          \
1001         "movq      %%mm0, %%mm2             \n\t" /* ABCDEFGH */          \
1002         "punpcklbw %%mm7, %%mm0             \n\t" /* 0A0B0C0D */          \
1003         "punpckhbw %%mm7, %%mm1             \n\t" /* 0E0F0G0H */          \
1004         "pshufw    $0x90, %%mm0, %%mm5      \n\t" /* 0A0A0B0C */          \
1005         "pshufw    $0x41, %%mm0, %%mm6      \n\t" /* 0B0A0A0B */          \
1006         "movq      %%mm2, %%mm3             \n\t" /* ABCDEFGH */          \
1007         "movq      %%mm2, %%mm4             \n\t" /* ABCDEFGH */          \
1008         "psllq        $8, %%mm2             \n\t" /* 0ABCDEFG */          \
1009         "psllq       $16, %%mm3             \n\t" /* 00ABCDEF */          \
1010         "psllq       $24, %%mm4             \n\t" /* 000ABCDE */          \
1011         "punpckhbw %%mm7, %%mm2             \n\t" /* 0D0E0F0G */          \
1012         "punpckhbw %%mm7, %%mm3             \n\t" /* 0C0D0E0F */          \
1013         "punpckhbw %%mm7, %%mm4             \n\t" /* 0B0C0D0E */          \
1014         "paddw     %%mm3, %%mm5             \n\t" /* b */                 \
1015         "paddw     %%mm2, %%mm6             \n\t" /* c */                 \
1016         "paddw     %%mm5, %%mm5             \n\t" /* 2b */                \
1017         "psubw     %%mm5, %%mm6             \n\t" /* c - 2b */            \
1018         "pshufw    $0x06, %%mm0, %%mm5      \n\t" /* 0C0B0A0A */          \
1019         "pmullw "MANGLE(ff_pw_3)", %%mm6    \n\t" /* 3c - 6b */           \
1020         "paddw     %%mm4, %%mm0             \n\t" /* a */                 \
1021         "paddw     %%mm1, %%mm5             \n\t" /* d */                 \
1022         "pmullw "MANGLE(ff_pw_20)", %%mm0   \n\t" /* 20a */               \
1023         "psubw     %%mm5, %%mm0             \n\t" /* 20a - d */           \
1024         "paddw        %5, %%mm6             \n\t"                         \
1025         "paddw     %%mm6, %%mm0             \n\t" /* 20a - 6b + 3c - d */ \
1026         "psraw        $5, %%mm0             \n\t"                         \
1027         /* mm1 = EFGH, mm2 = DEFG, mm3 = CDEF, mm4 = BCDE, mm7 = 0 */     \
1028                                                                           \
1029         "movd      5(%0), %%mm5             \n\t" /* FGHI */              \
1030         "punpcklbw %%mm7, %%mm5             \n\t" /* 0F0G0H0I */          \
1031         "pshufw    $0xF9, %%mm5, %%mm6      \n\t" /* 0G0H0I0I */          \
1032         "paddw     %%mm5, %%mm1             \n\t" /* a */                 \
1033         "paddw     %%mm6, %%mm2             \n\t" /* b */                 \
1034         "pshufw    $0xBE, %%mm5, %%mm6      \n\t" /* 0H0I0I0H */          \
1035         "pshufw    $0x6F, %%mm5, %%mm5      \n\t" /* 0I0I0H0G */          \
1036         "paddw     %%mm6, %%mm3             \n\t" /* c */                 \
1037         "paddw     %%mm5, %%mm4             \n\t" /* d */                 \
1038         "paddw     %%mm2, %%mm2             \n\t" /* 2b */                \
1039         "psubw     %%mm2, %%mm3             \n\t" /* c - 2b */            \
1040         "pmullw "MANGLE(ff_pw_20)", %%mm1   \n\t" /* 20a */               \
1041         "pmullw  "MANGLE(ff_pw_3)", %%mm3   \n\t" /* 3c - 6b */           \
1042         "psubw     %%mm4, %%mm3             \n\t" /* -6b + 3c - d */      \
1043         "paddw        %5, %%mm1             \n\t"                         \
1044         "paddw     %%mm1, %%mm3             \n\t" /* 20a - 6b + 3c - d */ \
1045         "psraw        $5, %%mm3             \n\t"                         \
1046         "packuswb  %%mm3, %%mm0             \n\t"                         \
1047         OP_MMXEXT(%%mm0, (%1), %%mm4, q)                                  \
1048                                                                           \
1049         "add          %3, %0                \n\t"                         \
1050         "add          %4, %1                \n\t"                         \
1051         "decl         %2                    \n\t"                         \
1052         "jnz          1b                    \n\t"                         \
1053         : "+a"(src), "+c"(dst), "+d"(h)                                   \
1054         : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride),               \
1055           /* "m"(ff_pw_20), "m"(ff_pw_3), */ "m"(ROUNDER)                 \
1056         : "memory"                                                        \
1057         );                                                                \
1058 }
1059
1060 #define QPEL_OP(OPNAME, ROUNDER, RND, OP, MMX)                          \
1061 static void OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(uint8_t *dst,      \
1062                                                      uint8_t *src,      \
1063                                                      int dstStride,     \
1064                                                      int srcStride)     \
1065 {                                                                       \
1066     uint64_t temp[17 * 4];                                              \
1067     uint64_t *temp_ptr = temp;                                          \
1068     int count = 17;                                                     \
1069                                                                         \
1070     /* FIXME unroll */                                                  \
1071     __asm__ volatile (                                                  \
1072         "pxor      %%mm7, %%mm7             \n\t"                       \
1073         "1:                                 \n\t"                       \
1074         "movq       (%0), %%mm0             \n\t"                       \
1075         "movq       (%0), %%mm1             \n\t"                       \
1076         "movq      8(%0), %%mm2             \n\t"                       \
1077         "movq      8(%0), %%mm3             \n\t"                       \
1078         "punpcklbw %%mm7, %%mm0             \n\t"                       \
1079         "punpckhbw %%mm7, %%mm1             \n\t"                       \
1080         "punpcklbw %%mm7, %%mm2             \n\t"                       \
1081         "punpckhbw %%mm7, %%mm3             \n\t"                       \
1082         "movq      %%mm0, (%1)              \n\t"                       \
1083         "movq      %%mm1, 17 * 8(%1)        \n\t"                       \
1084         "movq      %%mm2, 2 * 17 * 8(%1)    \n\t"                       \
1085         "movq      %%mm3, 3 * 17 * 8(%1)    \n\t"                       \
1086         "add          $8, %1                \n\t"                       \
1087         "add          %3, %0                \n\t"                       \
1088         "decl         %2                    \n\t"                       \
1089         "jnz          1b                    \n\t"                       \
1090         : "+r"(src), "+r"(temp_ptr), "+r"(count)                        \
1091         : "r"((x86_reg)srcStride)                                       \
1092         : "memory"                                                      \
1093         );                                                              \
1094                                                                         \
1095     temp_ptr = temp;                                                    \
1096     count    = 4;                                                       \
1097                                                                         \
1098     /* FIXME reorder for speed */                                       \
1099     __asm__ volatile (                                                  \
1100         /* "pxor  %%mm7, %%mm7            \n\t" */                      \
1101         "1:                             \n\t"                           \
1102         "movq    (%0), %%mm0            \n\t"                           \
1103         "movq   8(%0), %%mm1            \n\t"                           \
1104         "movq  16(%0), %%mm2            \n\t"                           \
1105         "movq  24(%0), %%mm3            \n\t"                           \
1106         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0),   8(%0),    (%0),  32(%0), (%1),     OP) \
1107         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5,  8(%0),    (%0),    (%0),  40(%0), (%1, %3), OP) \
1108         "add       %4, %1               \n\t"                           \
1109         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5,   (%0),    (%0),   8(%0),  48(%0), (%1),     OP) \
1110                                                                         \
1111         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5,   (%0),   8(%0),  16(%0),  56(%0), (%1, %3), OP) \
1112         "add       %4, %1               \n\t"                           \
1113         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5,  8(%0),  16(%0),  24(%0),  64(%0), (%1),     OP) \
1114         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0),  24(%0),  32(%0),  72(%0), (%1, %3), OP) \
1115         "add       %4, %1               \n\t"                           \
1116         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0),  32(%0),  40(%0),  80(%0), (%1),     OP) \
1117         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0),  40(%0),  48(%0),  88(%0), (%1, %3), OP) \
1118         "add       %4, %1               \n\t"                           \
1119         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 40(%0),  48(%0),  56(%0),  96(%0), (%1),     OP) \
1120         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 48(%0),  56(%0),  64(%0), 104(%0), (%1, %3), OP) \
1121         "add       %4, %1               \n\t"                           \
1122         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 56(%0),  64(%0),  72(%0), 112(%0), (%1),     OP) \
1123         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 64(%0),  72(%0),  80(%0), 120(%0), (%1, %3), OP) \
1124         "add       %4, %1               \n\t"                           \
1125         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 72(%0),  80(%0),  88(%0), 128(%0), (%1),     OP) \
1126                                                                         \
1127         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 80(%0),  88(%0),  96(%0), 128(%0), (%1, %3), OP) \
1128         "add       %4, %1               \n\t"                           \
1129         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 88(%0),  96(%0), 104(%0), 120(%0), (%1),     OP) \
1130         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 96(%0), 104(%0), 112(%0), 112(%0), (%1, %3), OP) \
1131                                                                         \
1132         "add     $136, %0               \n\t"                           \
1133         "add       %6, %1               \n\t"                           \
1134         "decl      %2                   \n\t"                           \
1135         "jnz       1b                   \n\t"                           \
1136                                                                         \
1137         : "+r"(temp_ptr), "+r"(dst), "+g"(count)                        \
1138         : "r"((x86_reg)dstStride), "r"(2 * (x86_reg)dstStride),         \
1139           /* "m"(ff_pw_20), "m"(ff_pw_3), */ "m"(ROUNDER),              \
1140           "g"(4 - 14 * (x86_reg)dstStride)                              \
1141         : "memory"                                                      \
1142         );                                                              \
1143 }                                                                       \
1144                                                                         \
1145 static void OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(uint8_t *dst,       \
1146                                                     uint8_t *src,       \
1147                                                     int dstStride,      \
1148                                                     int srcStride)      \
1149 {                                                                       \
1150     uint64_t temp[9 * 2];                                               \
1151     uint64_t *temp_ptr = temp;                                          \
1152     int count = 9;                                                      \
1153                                                                         \
1154     /* FIXME unroll */                                                  \
1155     __asm__ volatile (                                                  \
1156         "pxor      %%mm7, %%mm7         \n\t"                           \
1157         "1:                             \n\t"                           \
1158         "movq       (%0), %%mm0         \n\t"                           \
1159         "movq       (%0), %%mm1         \n\t"                           \
1160         "punpcklbw %%mm7, %%mm0         \n\t"                           \
1161         "punpckhbw %%mm7, %%mm1         \n\t"                           \
1162         "movq      %%mm0, (%1)          \n\t"                           \
1163         "movq      %%mm1, 9*8(%1)       \n\t"                           \
1164         "add          $8, %1            \n\t"                           \
1165         "add          %3, %0            \n\t"                           \
1166         "decl         %2                \n\t"                           \
1167         "jnz          1b                \n\t"                           \
1168         : "+r"(src), "+r"(temp_ptr), "+r"(count)                        \
1169         : "r"((x86_reg)srcStride)                                       \
1170         : "memory"                                                      \
1171         );                                                              \
1172                                                                         \
1173     temp_ptr = temp;                                                    \
1174     count    = 2;                                                       \
1175                                                                         \
1176     /* FIXME reorder for speed */                                       \
1177     __asm__ volatile (                                                  \
1178         /* "pxor  %%mm7, %%mm7            \n\t" */                      \
1179         "1:                             \n\t"                           \
1180         "movq    (%0), %%mm0            \n\t"                           \
1181         "movq   8(%0), %%mm1            \n\t"                           \
1182         "movq  16(%0), %%mm2            \n\t"                           \
1183         "movq  24(%0), %%mm3            \n\t"                           \
1184         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0),  8(%0),   (%0), 32(%0), (%1), OP)     \
1185         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5,  8(%0),   (%0),   (%0), 40(%0), (%1, %3), OP) \
1186         "add       %4, %1               \n\t"                           \
1187         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5,   (%0),   (%0),  8(%0), 48(%0), (%1), OP)     \
1188                                                                         \
1189         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5,   (%0),  8(%0), 16(%0), 56(%0), (%1, %3), OP) \
1190         "add       %4, %1               \n\t"                           \
1191         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5,  8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)     \
1192                                                                         \
1193         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 64(%0), (%1, %3), OP) \
1194         "add       %4, %1               \n\t"                           \
1195         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 56(%0), (%1), OP)     \
1196         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 48(%0), (%1, %3), OP) \
1197                                                                         \
1198         "add      $72, %0               \n\t"                           \
1199         "add       %6, %1               \n\t"                           \
1200         "decl      %2                   \n\t"                           \
1201         "jnz       1b                   \n\t"                           \
1202                                                                         \
1203         : "+r"(temp_ptr), "+r"(dst), "+g"(count)                        \
1204         : "r"((x86_reg)dstStride), "r"(2 * (x86_reg)dstStride),         \
1205           /* "m"(ff_pw_20), "m"(ff_pw_3), */ "m"(ROUNDER),              \
1206           "g"(4 - 6 * (x86_reg)dstStride)                               \
1207         : "memory"                                                      \
1208         );                                                              \
1209 }                                                                       \
1210                                                                         \
1211 static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src,   \
1212                                           int stride)                   \
1213 {                                                                       \
1214     OPNAME ## pixels8_ ## MMX(dst, src, stride, 8);                     \
1215 }                                                                       \
1216                                                                         \
1217 static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src,    \
1218                                          int stride)                    \
1219 {                                                                       \
1220     uint64_t temp[8];                                                   \
1221     uint8_t * const half = (uint8_t*)temp;                              \
1222     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8,           \
1223                                                 stride, 8);             \
1224     OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);    \
1225 }                                                                       \
1226                                                                         \
1227 static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src,    \
1228                                          int stride)                    \
1229 {                                                                       \
1230     OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride,           \
1231                                             stride, 8);                 \
1232 }                                                                       \
1233                                                                         \
1234 static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src,    \
1235                                          int stride)                    \
1236 {                                                                       \
1237     uint64_t temp[8];                                                   \
1238     uint8_t * const half = (uint8_t*)temp;                              \
1239     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8,           \
1240                                                 stride, 8);             \
1241     OPNAME ## pixels8_l2_ ## MMX(dst, src + 1, half, stride,            \
1242                                  stride, 8);                            \
1243 }                                                                       \
1244                                                                         \
1245 static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src,    \
1246                                          int stride)                    \
1247 {                                                                       \
1248     uint64_t temp[8];                                                   \
1249     uint8_t * const half = (uint8_t*)temp;                              \
1250     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);  \
1251     OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);    \
1252 }                                                                       \
1253                                                                         \
1254 static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src,    \
1255                                          int stride)                    \
1256 {                                                                       \
1257     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src, stride, stride);  \
1258 }                                                                       \
1259                                                                         \
1260 static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src,    \
1261                                          int stride)                    \
1262 {                                                                       \
1263     uint64_t temp[8];                                                   \
1264     uint8_t * const half = (uint8_t*)temp;                              \
1265     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);  \
1266     OPNAME ## pixels8_l2_ ## MMX(dst, src + stride, half, stride,       \
1267                                  stride, 8);                            \
1268 }                                                                       \
1269                                                                         \
1270 static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src,    \
1271                                          int stride)                    \
1272 {                                                                       \
1273     uint64_t half[8 + 9];                                               \
1274     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1275     uint8_t * const halfHV = ((uint8_t*)half);                          \
1276     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1277                                                 stride, 9);             \
1278     put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);  \
1279     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1280     OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);     \
1281 }                                                                       \
1282                                                                         \
1283 static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src,    \
1284                                          int stride)                    \
1285 {                                                                       \
1286     uint64_t half[8 + 9];                                               \
1287     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1288     uint8_t * const halfHV = ((uint8_t*)half);                          \
1289     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1290                                                 stride, 9);             \
1291     put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,          \
1292                                      stride, 9);                        \
1293     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1294     OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);     \
1295 }                                                                       \
1296                                                                         \
1297 static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src,    \
1298                                          int stride)                    \
1299 {                                                                       \
1300     uint64_t half[8 + 9];                                               \
1301     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1302     uint8_t * const halfHV = ((uint8_t*)half);                          \
1303     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1304                                                 stride, 9);             \
1305     put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);  \
1306     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1307     OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV, stride, 8, 8); \
1308 }                                                                       \
1309                                                                         \
1310 static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src,    \
1311                                          int stride)                    \
1312 {                                                                       \
1313     uint64_t half[8 + 9];                                               \
1314     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1315     uint8_t * const halfHV = ((uint8_t*)half);                          \
1316     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1317                                                 stride, 9);             \
1318     put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,          \
1319                                      stride, 9);                        \
1320     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1321     OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV, stride, 8, 8); \
1322 }                                                                       \
1323                                                                         \
1324 static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src,    \
1325                                          int stride)                    \
1326 {                                                                       \
1327     uint64_t half[8 + 9];                                               \
1328     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1329     uint8_t * const halfHV = ((uint8_t*)half);                          \
1330     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1331                                                 stride, 9);             \
1332     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1333     OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);     \
1334 }                                                                       \
1335                                                                         \
1336 static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src,    \
1337                                          int stride)                    \
1338 {                                                                       \
1339     uint64_t half[8 + 9];                                               \
1340     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1341     uint8_t * const halfHV = ((uint8_t*)half);                          \
1342     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1343                                                 stride, 9);             \
1344     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1345     OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV, stride, 8, 8); \
1346 }                                                                       \
1347                                                                         \
1348 static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src,    \
1349                                          int stride)                    \
1350 {                                                                       \
1351     uint64_t half[8 + 9];                                               \
1352     uint8_t * const halfH = ((uint8_t*)half);                           \
1353     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1354                                                 stride, 9);             \
1355     put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);  \
1356     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);     \
1357 }                                                                       \
1358                                                                         \
1359 static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src,    \
1360                                          int stride)                    \
1361 {                                                                       \
1362     uint64_t half[8 + 9];                                               \
1363     uint8_t * const halfH = ((uint8_t*)half);                           \
1364     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1365                                                 stride, 9);             \
1366     put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,          \
1367                                      stride, 9);                        \
1368     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);     \
1369 }                                                                       \
1370                                                                         \
1371 static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src,    \
1372                                          int stride)                    \
1373 {                                                                       \
1374     uint64_t half[9];                                                   \
1375     uint8_t * const halfH = ((uint8_t*)half);                           \
1376     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1377                                                 stride, 9);             \
1378     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);     \
1379 }                                                                       \
1380                                                                         \
1381 static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src,  \
1382                                            int stride)                  \
1383 {                                                                       \
1384     OPNAME ## pixels16_ ## MMX(dst, src, stride, 16);                   \
1385 }                                                                       \
1386                                                                         \
1387 static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src,   \
1388                                           int stride)                   \
1389 {                                                                       \
1390     uint64_t temp[32];                                                  \
1391     uint8_t * const half = (uint8_t*)temp;                              \
1392     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16,         \
1393                                                  stride, 16);           \
1394     OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);  \
1395 }                                                                       \
1396                                                                         \
1397 static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src,   \
1398                                           int stride)                   \
1399 {                                                                       \
1400     OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src,                  \
1401                                              stride, stride, 16);       \
1402 }                                                                       \
1403                                                                         \
1404 static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src,   \
1405                                           int stride)                   \
1406 {                                                                       \
1407     uint64_t temp[32];                                                  \
1408     uint8_t * const half = (uint8_t*)temp;                              \
1409     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16,         \
1410                                                  stride, 16);           \
1411     OPNAME ## pixels16_l2_ ## MMX(dst, src + 1, half,                   \
1412                                   stride, stride, 16);                  \
1413 }                                                                       \
1414                                                                         \
1415 static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src,   \
1416                                           int stride)                   \
1417 {                                                                       \
1418     uint64_t temp[32];                                                  \
1419     uint8_t * const half = (uint8_t*)temp;                              \
1420     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16,         \
1421                                                  stride);               \
1422     OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);  \
1423 }                                                                       \
1424                                                                         \
1425 static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src,   \
1426                                           int stride)                   \
1427 {                                                                       \
1428     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src, stride, stride); \
1429 }                                                                       \
1430                                                                         \
1431 static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src,   \
1432                                           int stride)                   \
1433 {                                                                       \
1434     uint64_t temp[32];                                                  \
1435     uint8_t * const half = (uint8_t*)temp;                              \
1436     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16,         \
1437                                                  stride);               \
1438     OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half,                \
1439                                   stride, stride, 16);                  \
1440 }                                                                       \
1441                                                                         \
1442 static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src,   \
1443                                           int stride)                   \
1444 {                                                                       \
1445     uint64_t half[16 * 2 + 17 * 2];                                     \
1446     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1447     uint8_t * const halfHV = ((uint8_t*)half);                          \
1448     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1449                                                  stride, 17);           \
1450     put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,            \
1451                                       stride, 17);                      \
1452     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1453                                                  16, 16);               \
1454     OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);  \
1455 }                                                                       \
1456                                                                         \
1457 static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src,   \
1458                                           int stride)                   \
1459 {                                                                       \
1460     uint64_t half[16 * 2 + 17 * 2];                                     \
1461     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1462     uint8_t * const halfHV = ((uint8_t*)half);                          \
1463     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1464                                                  stride, 17);           \
1465     put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,        \
1466                                       stride, 17);                      \
1467     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1468                                                  16, 16);               \
1469     OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);  \
1470 }                                                                       \
1471                                                                         \
1472 static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src,   \
1473                                           int stride)                   \
1474 {                                                                       \
1475     uint64_t half[16 * 2 + 17 * 2];                                     \
1476     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1477     uint8_t * const halfHV = ((uint8_t*)half);                          \
1478     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1479                                                  stride, 17);           \
1480     put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,            \
1481                                       stride, 17);                      \
1482     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1483                                                  16, 16);               \
1484     OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV, stride,      \
1485                                   16, 16);                              \
1486 }                                                                       \
1487                                                                         \
1488 static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src,   \
1489                                           int stride)                   \
1490 {                                                                       \
1491     uint64_t half[16 * 2 + 17 * 2];                                     \
1492     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1493     uint8_t * const halfHV = ((uint8_t*)half);                          \
1494     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1495                                                  stride, 17);           \
1496     put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,        \
1497                                       stride, 17);                      \
1498     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1499                                                  16, 16);               \
1500     OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV, stride,      \
1501                                   16, 16);                              \
1502 }                                                                       \
1503                                                                         \
1504 static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src,   \
1505                                           int stride)                   \
1506 {                                                                       \
1507     uint64_t half[16 * 2 + 17 * 2];                                     \
1508     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1509     uint8_t * const halfHV = ((uint8_t*)half);                          \
1510     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1511                                                  stride, 17);           \
1512     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1513                                                  16, 16);               \
1514     OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);  \
1515 }                                                                       \
1516                                                                         \
1517 static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src,   \
1518                                           int stride)                   \
1519 {                                                                       \
1520     uint64_t half[16 * 2 + 17 * 2];                                     \
1521     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1522     uint8_t * const halfHV = ((uint8_t*)half);                          \
1523     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1524                                                  stride, 17);           \
1525     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1526                                                  16, 16);               \
1527     OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV, stride,      \
1528                                   16, 16);                              \
1529 }                                                                       \
1530                                                                         \
1531 static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src,   \
1532                                           int stride)                   \
1533 {                                                                       \
1534     uint64_t half[17 * 2];                                              \
1535     uint8_t * const halfH = ((uint8_t*)half);                           \
1536     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1537                                                  stride, 17);           \
1538     put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,            \
1539                                       stride, 17);                      \
1540     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);   \
1541 }                                                                       \
1542                                                                         \
1543 static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src,   \
1544                                           int stride)                   \
1545 {                                                                       \
1546     uint64_t half[17 * 2];                                              \
1547     uint8_t * const halfH = ((uint8_t*)half);                           \
1548     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1549                                                  stride, 17);           \
1550     put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,        \
1551                                       stride, 17);                      \
1552     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);   \
1553 }                                                                       \
1554                                                                         \
1555 static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src,   \
1556                                           int stride)                   \
1557 {                                                                       \
1558     uint64_t half[17 * 2];                                              \
1559     uint8_t * const halfH = ((uint8_t*)half);                           \
1560     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1561                                                  stride, 17);           \
1562     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);   \
1563 }
1564
1565 #define PUT_OP(a, b, temp, size)                \
1566     "mov"#size"        "#a", "#b"       \n\t"
1567
1568 #define AVG_MMXEXT_OP(a, b, temp, size)         \
1569     "mov"#size"        "#b", "#temp"    \n\t"   \
1570     "pavgb          "#temp", "#a"       \n\t"   \
1571     "mov"#size"        "#a", "#b"       \n\t"
1572
1573 QPEL_BASE(put_,        ff_pw_16, _,        PUT_OP)
1574 QPEL_BASE(avg_,        ff_pw_16, _,        AVG_MMXEXT_OP)
1575 QPEL_BASE(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP)
1576 QPEL_OP(put_,          ff_pw_16, _,        PUT_OP,        mmxext)
1577 QPEL_OP(avg_,          ff_pw_16, _,        AVG_MMXEXT_OP, mmxext)
1578 QPEL_OP(put_no_rnd_,   ff_pw_15, _no_rnd_, PUT_OP,        mmxext)
1579
1580 void ff_put_rv40_qpel8_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1581 {
1582   put_pixels8_xy2_mmx(dst, src, stride, 8);
1583 }
1584 void ff_put_rv40_qpel16_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1585 {
1586   put_pixels16_xy2_mmx(dst, src, stride, 16);
1587 }
1588 void ff_avg_rv40_qpel8_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1589 {
1590   avg_pixels8_xy2_mmx(dst, src, stride, 8);
1591 }
1592 void ff_avg_rv40_qpel16_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1593 {
1594   avg_pixels16_xy2_mmx(dst, src, stride, 16);
1595 }
1596
1597 typedef void emulated_edge_mc_func(uint8_t *dst, const uint8_t *src,
1598                                    ptrdiff_t linesize, int block_w, int block_h,
1599                                    int src_x, int src_y, int w, int h);
1600
1601 static av_always_inline void gmc(uint8_t *dst, uint8_t *src,
1602                                  int stride, int h, int ox, int oy,
1603                                  int dxx, int dxy, int dyx, int dyy,
1604                                  int shift, int r, int width, int height,
1605                                  emulated_edge_mc_func *emu_edge_fn)
1606 {
1607     const int w    = 8;
1608     const int ix   = ox  >> (16 + shift);
1609     const int iy   = oy  >> (16 + shift);
1610     const int oxs  = ox  >> 4;
1611     const int oys  = oy  >> 4;
1612     const int dxxs = dxx >> 4;
1613     const int dxys = dxy >> 4;
1614     const int dyxs = dyx >> 4;
1615     const int dyys = dyy >> 4;
1616     const uint16_t r4[4]   = { r, r, r, r };
1617     const uint16_t dxy4[4] = { dxys, dxys, dxys, dxys };
1618     const uint16_t dyy4[4] = { dyys, dyys, dyys, dyys };
1619     const uint64_t shift2 = 2 * shift;
1620 #define MAX_STRIDE 4096U
1621 #define MAX_H 8U
1622     uint8_t edge_buf[(MAX_H + 1) * MAX_STRIDE];
1623     int x, y;
1624
1625     const int dxw = (dxx - (1 << (16 + shift))) * (w - 1);
1626     const int dyh = (dyy - (1 << (16 + shift))) * (h - 1);
1627     const int dxh = dxy * (h - 1);
1628     const int dyw = dyx * (w - 1);
1629     int need_emu =  (unsigned)ix >= width  - w ||
1630                     (unsigned)iy >= height - h;
1631
1632     if ( // non-constant fullpel offset (3% of blocks)
1633         ((ox ^ (ox + dxw)) | (ox ^ (ox + dxh)) | (ox ^ (ox + dxw + dxh)) |
1634          (oy ^ (oy + dyw)) | (oy ^ (oy + dyh)) | (oy ^ (oy + dyw + dyh))) >> (16 + shift)
1635         // uses more than 16 bits of subpel mv (only at huge resolution)
1636         || (dxx | dxy | dyx | dyy) & 15
1637         || (need_emu && (h > MAX_H || stride > MAX_STRIDE))) {
1638         // FIXME could still use mmx for some of the rows
1639         ff_gmc_c(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy,
1640                  shift, r, width, height);
1641         return;
1642     }
1643
1644     src += ix + iy * stride;
1645     if (need_emu) {
1646         emu_edge_fn(edge_buf, src, stride, w + 1, h + 1, ix, iy, width, height);
1647         src = edge_buf;
1648     }
1649
1650     __asm__ volatile (
1651         "movd         %0, %%mm6         \n\t"
1652         "pxor      %%mm7, %%mm7         \n\t"
1653         "punpcklwd %%mm6, %%mm6         \n\t"
1654         "punpcklwd %%mm6, %%mm6         \n\t"
1655         :: "r"(1<<shift)
1656     );
1657
1658     for (x = 0; x < w; x += 4) {
1659         uint16_t dx4[4] = { oxs - dxys + dxxs * (x + 0),
1660                             oxs - dxys + dxxs * (x + 1),
1661                             oxs - dxys + dxxs * (x + 2),
1662                             oxs - dxys + dxxs * (x + 3) };
1663         uint16_t dy4[4] = { oys - dyys + dyxs * (x + 0),
1664                             oys - dyys + dyxs * (x + 1),
1665                             oys - dyys + dyxs * (x + 2),
1666                             oys - dyys + dyxs * (x + 3) };
1667
1668         for (y = 0; y < h; y++) {
1669             __asm__ volatile (
1670                 "movq      %0, %%mm4    \n\t"
1671                 "movq      %1, %%mm5    \n\t"
1672                 "paddw     %2, %%mm4    \n\t"
1673                 "paddw     %3, %%mm5    \n\t"
1674                 "movq   %%mm4, %0       \n\t"
1675                 "movq   %%mm5, %1       \n\t"
1676                 "psrlw    $12, %%mm4    \n\t"
1677                 "psrlw    $12, %%mm5    \n\t"
1678                 : "+m"(*dx4), "+m"(*dy4)
1679                 : "m"(*dxy4), "m"(*dyy4)
1680             );
1681
1682             __asm__ volatile (
1683                 "movq      %%mm6, %%mm2 \n\t"
1684                 "movq      %%mm6, %%mm1 \n\t"
1685                 "psubw     %%mm4, %%mm2 \n\t"
1686                 "psubw     %%mm5, %%mm1 \n\t"
1687                 "movq      %%mm2, %%mm0 \n\t"
1688                 "movq      %%mm4, %%mm3 \n\t"
1689                 "pmullw    %%mm1, %%mm0 \n\t" // (s - dx) * (s - dy)
1690                 "pmullw    %%mm5, %%mm3 \n\t" // dx * dy
1691                 "pmullw    %%mm5, %%mm2 \n\t" // (s - dx) * dy
1692                 "pmullw    %%mm4, %%mm1 \n\t" // dx * (s - dy)
1693
1694                 "movd         %4, %%mm5 \n\t"
1695                 "movd         %3, %%mm4 \n\t"
1696                 "punpcklbw %%mm7, %%mm5 \n\t"
1697                 "punpcklbw %%mm7, %%mm4 \n\t"
1698                 "pmullw    %%mm5, %%mm3 \n\t" // src[1, 1] * dx * dy
1699                 "pmullw    %%mm4, %%mm2 \n\t" // src[0, 1] * (s - dx) * dy
1700
1701                 "movd         %2, %%mm5 \n\t"
1702                 "movd         %1, %%mm4 \n\t"
1703                 "punpcklbw %%mm7, %%mm5 \n\t"
1704                 "punpcklbw %%mm7, %%mm4 \n\t"
1705                 "pmullw    %%mm5, %%mm1 \n\t" // src[1, 0] * dx * (s - dy)
1706                 "pmullw    %%mm4, %%mm0 \n\t" // src[0, 0] * (s - dx) * (s - dy)
1707                 "paddw        %5, %%mm1 \n\t"
1708                 "paddw     %%mm3, %%mm2 \n\t"
1709                 "paddw     %%mm1, %%mm0 \n\t"
1710                 "paddw     %%mm2, %%mm0 \n\t"
1711
1712                 "psrlw        %6, %%mm0 \n\t"
1713                 "packuswb  %%mm0, %%mm0 \n\t"
1714                 "movd      %%mm0, %0    \n\t"
1715
1716                 : "=m"(dst[x + y * stride])
1717                 : "m"(src[0]), "m"(src[1]),
1718                   "m"(src[stride]), "m"(src[stride + 1]),
1719                   "m"(*r4), "m"(shift2)
1720             );
1721             src += stride;
1722         }
1723         src += 4 - h * stride;
1724     }
1725 }
1726
1727 #if CONFIG_VIDEODSP
1728 #if HAVE_YASM
1729 #if ARCH_X86_32
1730 static void gmc_mmx(uint8_t *dst, uint8_t *src,
1731                     int stride, int h, int ox, int oy,
1732                     int dxx, int dxy, int dyx, int dyy,
1733                     int shift, int r, int width, int height)
1734 {
1735     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,
1736         width, height, &ff_emulated_edge_mc_8);
1737 }
1738 #endif
1739 static void gmc_sse(uint8_t *dst, uint8_t *src,
1740                     int stride, int h, int ox, int oy,
1741                     int dxx, int dxy, int dyx, int dyy,
1742                     int shift, int r, int width, int height)
1743 {
1744     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,
1745         width, height, &ff_emulated_edge_mc_8);
1746 }
1747 #else
1748 static void gmc_mmx(uint8_t *dst, uint8_t *src,
1749                     int stride, int h, int ox, int oy,
1750                     int dxx, int dxy, int dyx, int dyy,
1751                     int shift, int r, int width, int height)
1752 {
1753     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,
1754         width, height, &ff_emulated_edge_mc_8);
1755 }
1756 #endif
1757 #endif
1758
1759 #endif /* HAVE_INLINE_ASM */
1760
1761 #include "h264_qpel.c"
1762
1763 void ff_put_h264_chroma_mc8_rnd_mmx  (uint8_t *dst, uint8_t *src,
1764                                       int stride, int h, int x, int y);
1765 void ff_avg_h264_chroma_mc8_rnd_mmxext(uint8_t *dst, uint8_t *src,
1766                                        int stride, int h, int x, int y);
1767 void ff_avg_h264_chroma_mc8_rnd_3dnow(uint8_t *dst, uint8_t *src,
1768                                       int stride, int h, int x, int y);
1769
1770 void ff_put_h264_chroma_mc4_mmx      (uint8_t *dst, uint8_t *src,
1771                                       int stride, int h, int x, int y);
1772 void ff_avg_h264_chroma_mc4_mmxext   (uint8_t *dst, uint8_t *src,
1773                                       int stride, int h, int x, int y);
1774 void ff_avg_h264_chroma_mc4_3dnow    (uint8_t *dst, uint8_t *src,
1775                                       int stride, int h, int x, int y);
1776
1777 void ff_put_h264_chroma_mc2_mmxext   (uint8_t *dst, uint8_t *src,
1778                                       int stride, int h, int x, int y);
1779 void ff_avg_h264_chroma_mc2_mmxext   (uint8_t *dst, uint8_t *src,
1780                                       int stride, int h, int x, int y);
1781
1782 void ff_put_h264_chroma_mc8_rnd_ssse3(uint8_t *dst, uint8_t *src,
1783                                       int stride, int h, int x, int y);
1784 void ff_put_h264_chroma_mc4_ssse3    (uint8_t *dst, uint8_t *src,
1785                                       int stride, int h, int x, int y);
1786
1787 void ff_avg_h264_chroma_mc8_rnd_ssse3(uint8_t *dst, uint8_t *src,
1788                                       int stride, int h, int x, int y);
1789 void ff_avg_h264_chroma_mc4_ssse3    (uint8_t *dst, uint8_t *src,
1790                                       int stride, int h, int x, int y);
1791
1792 #define CHROMA_MC(OP, NUM, DEPTH, OPT)                                  \
1793 void ff_ ## OP ## _h264_chroma_mc ## NUM ## _ ## DEPTH ## _ ## OPT      \
1794                                       (uint8_t *dst, uint8_t *src,      \
1795                                        int stride, int h, int x, int y);
1796
1797 CHROMA_MC(put, 2, 10, mmxext)
1798 CHROMA_MC(avg, 2, 10, mmxext)
1799 CHROMA_MC(put, 4, 10, mmxext)
1800 CHROMA_MC(avg, 4, 10, mmxext)
1801 CHROMA_MC(put, 8, 10, sse2)
1802 CHROMA_MC(avg, 8, 10, sse2)
1803 CHROMA_MC(put, 8, 10, avx)
1804 CHROMA_MC(avg, 8, 10, avx)
1805
1806 #if HAVE_INLINE_ASM
1807
1808 /* CAVS-specific */
1809 void ff_put_cavs_qpel8_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1810 {
1811     put_pixels8_mmx(dst, src, stride, 8);
1812 }
1813
1814 void ff_avg_cavs_qpel8_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1815 {
1816     avg_pixels8_mmx(dst, src, stride, 8);
1817 }
1818
1819 void ff_put_cavs_qpel16_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1820 {
1821     put_pixels16_mmx(dst, src, stride, 16);
1822 }
1823
1824 void ff_avg_cavs_qpel16_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1825 {
1826     avg_pixels16_mmx(dst, src, stride, 16);
1827 }
1828
1829 /* VC-1-specific */
1830 void ff_put_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src,
1831                                int stride, int rnd)
1832 {
1833     put_pixels8_mmx(dst, src, stride, 8);
1834 }
1835
1836 void ff_avg_vc1_mspel_mc00_mmxext(uint8_t *dst, const uint8_t *src,
1837                                   int stride, int rnd)
1838 {
1839     avg_pixels8_mmxext(dst, src, stride, 8);
1840 }
1841
1842 #if CONFIG_DIRAC_DECODER
1843 #define DIRAC_PIXOP(OPNAME, EXT)\
1844 void ff_ ## OPNAME ## _dirac_pixels8_ ## EXT(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
1845 {\
1846     OPNAME ## _pixels8_ ## EXT(dst, src[0], stride, h);\
1847 }\
1848 void ff_ ## OPNAME ## _dirac_pixels16_ ## EXT(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
1849 {\
1850     OPNAME ## _pixels16_ ## EXT(dst, src[0], stride, h);\
1851 }\
1852 void ff_ ## OPNAME ## _dirac_pixels32_ ## EXT(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
1853 {\
1854     OPNAME ## _pixels16_ ## EXT(dst   , src[0]   , stride, h);\
1855     OPNAME ## _pixels16_ ## EXT(dst+16, src[0]+16, stride, h);\
1856 }
1857
1858 DIRAC_PIXOP(put, mmx)
1859 DIRAC_PIXOP(avg, mmx)
1860 DIRAC_PIXOP(avg, mmxext)
1861
1862 #if HAVE_YASM
1863 void ff_put_dirac_pixels16_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h)
1864 {
1865     ff_put_pixels16_sse2(dst, src[0], stride, h);
1866 }
1867 void ff_avg_dirac_pixels16_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h)
1868 {
1869     ff_avg_pixels16_sse2(dst, src[0], stride, h);
1870 }
1871 void ff_put_dirac_pixels32_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h)
1872 {
1873     ff_put_pixels16_sse2(dst   , src[0]   , stride, h);
1874     ff_put_pixels16_sse2(dst+16, src[0]+16, stride, h);
1875 }
1876 void ff_avg_dirac_pixels32_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h)
1877 {
1878     ff_avg_pixels16_sse2(dst   , src[0]   , stride, h);
1879     ff_avg_pixels16_sse2(dst+16, src[0]+16, stride, h);
1880 }
1881 #endif
1882 #endif
1883
1884 /* XXX: Those functions should be suppressed ASAP when all IDCTs are
1885  * converted. */
1886 #if CONFIG_GPL
1887 static void ff_libmpeg2mmx_idct_put(uint8_t *dest, int line_size,
1888                                     DCTELEM *block)
1889 {
1890     ff_mmx_idct(block);
1891     ff_put_pixels_clamped_mmx(block, dest, line_size);
1892 }
1893
1894 static void ff_libmpeg2mmx_idct_add(uint8_t *dest, int line_size,
1895                                     DCTELEM *block)
1896 {
1897     ff_mmx_idct(block);
1898     ff_add_pixels_clamped_mmx(block, dest, line_size);
1899 }
1900
1901 static void ff_libmpeg2mmx2_idct_put(uint8_t *dest, int line_size,
1902                                      DCTELEM *block)
1903 {
1904     ff_mmxext_idct(block);
1905     ff_put_pixels_clamped_mmx(block, dest, line_size);
1906 }
1907
1908 static void ff_libmpeg2mmx2_idct_add(uint8_t *dest, int line_size,
1909                                      DCTELEM *block)
1910 {
1911     ff_mmxext_idct(block);
1912     ff_add_pixels_clamped_mmx(block, dest, line_size);
1913 }
1914 #endif
1915
1916 static void vector_clipf_sse(float *dst, const float *src,
1917                              float min, float max, int len)
1918 {
1919     x86_reg i = (len - 16) * 4;
1920     __asm__ volatile (
1921         "movss          %3, %%xmm4      \n\t"
1922         "movss          %4, %%xmm5      \n\t"
1923         "shufps $0, %%xmm4, %%xmm4      \n\t"
1924         "shufps $0, %%xmm5, %%xmm5      \n\t"
1925         "1:                             \n\t"
1926         "movaps   (%2, %0), %%xmm0      \n\t" // 3/1 on intel
1927         "movaps 16(%2, %0), %%xmm1      \n\t"
1928         "movaps 32(%2, %0), %%xmm2      \n\t"
1929         "movaps 48(%2, %0), %%xmm3      \n\t"
1930         "maxps      %%xmm4, %%xmm0      \n\t"
1931         "maxps      %%xmm4, %%xmm1      \n\t"
1932         "maxps      %%xmm4, %%xmm2      \n\t"
1933         "maxps      %%xmm4, %%xmm3      \n\t"
1934         "minps      %%xmm5, %%xmm0      \n\t"
1935         "minps      %%xmm5, %%xmm1      \n\t"
1936         "minps      %%xmm5, %%xmm2      \n\t"
1937         "minps      %%xmm5, %%xmm3      \n\t"
1938         "movaps     %%xmm0,   (%1, %0)  \n\t"
1939         "movaps     %%xmm1, 16(%1, %0)  \n\t"
1940         "movaps     %%xmm2, 32(%1, %0)  \n\t"
1941         "movaps     %%xmm3, 48(%1, %0)  \n\t"
1942         "sub           $64, %0          \n\t"
1943         "jge            1b              \n\t"
1944         : "+&r"(i)
1945         : "r"(dst), "r"(src), "m"(min), "m"(max)
1946         : "memory"
1947     );
1948 }
1949
1950 #endif /* HAVE_INLINE_ASM */
1951
1952 int32_t ff_scalarproduct_int16_mmxext(const int16_t *v1, const int16_t *v2,
1953                                       int order);
1954 int32_t ff_scalarproduct_int16_sse2(const int16_t *v1, const int16_t *v2,
1955                                     int order);
1956 int32_t ff_scalarproduct_and_madd_int16_mmxext(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_sse2(int16_t *v1, const int16_t *v2,
1960                                              const int16_t *v3,
1961                                              int order, int mul);
1962 int32_t ff_scalarproduct_and_madd_int16_ssse3(int16_t *v1, const int16_t *v2,
1963                                               const int16_t *v3,
1964                                               int order, int mul);
1965
1966 void ff_apply_window_int16_round_mmxext(int16_t *output, const int16_t *input,
1967                                         const int16_t *window, unsigned int len);
1968 void ff_apply_window_int16_round_sse2(int16_t *output, const int16_t *input,
1969                                       const int16_t *window, unsigned int len);
1970 void ff_apply_window_int16_mmxext(int16_t *output, const int16_t *input,
1971                                   const int16_t *window, unsigned int len);
1972 void ff_apply_window_int16_sse2(int16_t *output, const int16_t *input,
1973                                 const int16_t *window, unsigned int len);
1974 void ff_apply_window_int16_ssse3(int16_t *output, const int16_t *input,
1975                                  const int16_t *window, unsigned int len);
1976 void ff_apply_window_int16_ssse3_atom(int16_t *output, const int16_t *input,
1977                                       const int16_t *window, unsigned int len);
1978
1979 void ff_bswap32_buf_ssse3(uint32_t *dst, const uint32_t *src, int w);
1980 void ff_bswap32_buf_sse2(uint32_t *dst, const uint32_t *src, int w);
1981
1982 void ff_add_hfyu_median_prediction_mmxext(uint8_t *dst, const uint8_t *top,
1983                                           const uint8_t *diff, int w,
1984                                           int *left, int *left_top);
1985 int  ff_add_hfyu_left_prediction_ssse3(uint8_t *dst, const uint8_t *src,
1986                                        int w, int left);
1987 int  ff_add_hfyu_left_prediction_sse4(uint8_t *dst, const uint8_t *src,
1988                                       int w, int left);
1989
1990 void ff_vector_clip_int32_mmx     (int32_t *dst, const int32_t *src,
1991                                    int32_t min, int32_t max, unsigned int len);
1992 void ff_vector_clip_int32_sse2    (int32_t *dst, const int32_t *src,
1993                                    int32_t min, int32_t max, unsigned int len);
1994 void ff_vector_clip_int32_int_sse2(int32_t *dst, const int32_t *src,
1995                                    int32_t min, int32_t max, unsigned int len);
1996 void ff_vector_clip_int32_sse4    (int32_t *dst, const int32_t *src,
1997                                    int32_t min, int32_t max, unsigned int len);
1998
1999 #define SET_QPEL_FUNCS(PFX, IDX, SIZE, CPU, PREFIX)                          \
2000     do {                                                                     \
2001     c->PFX ## _pixels_tab[IDX][ 0] = PREFIX ## PFX ## SIZE ## _mc00_ ## CPU; \
2002     c->PFX ## _pixels_tab[IDX][ 1] = PREFIX ## PFX ## SIZE ## _mc10_ ## CPU; \
2003     c->PFX ## _pixels_tab[IDX][ 2] = PREFIX ## PFX ## SIZE ## _mc20_ ## CPU; \
2004     c->PFX ## _pixels_tab[IDX][ 3] = PREFIX ## PFX ## SIZE ## _mc30_ ## CPU; \
2005     c->PFX ## _pixels_tab[IDX][ 4] = PREFIX ## PFX ## SIZE ## _mc01_ ## CPU; \
2006     c->PFX ## _pixels_tab[IDX][ 5] = PREFIX ## PFX ## SIZE ## _mc11_ ## CPU; \
2007     c->PFX ## _pixels_tab[IDX][ 6] = PREFIX ## PFX ## SIZE ## _mc21_ ## CPU; \
2008     c->PFX ## _pixels_tab[IDX][ 7] = PREFIX ## PFX ## SIZE ## _mc31_ ## CPU; \
2009     c->PFX ## _pixels_tab[IDX][ 8] = PREFIX ## PFX ## SIZE ## _mc02_ ## CPU; \
2010     c->PFX ## _pixels_tab[IDX][ 9] = PREFIX ## PFX ## SIZE ## _mc12_ ## CPU; \
2011     c->PFX ## _pixels_tab[IDX][10] = PREFIX ## PFX ## SIZE ## _mc22_ ## CPU; \
2012     c->PFX ## _pixels_tab[IDX][11] = PREFIX ## PFX ## SIZE ## _mc32_ ## CPU; \
2013     c->PFX ## _pixels_tab[IDX][12] = PREFIX ## PFX ## SIZE ## _mc03_ ## CPU; \
2014     c->PFX ## _pixels_tab[IDX][13] = PREFIX ## PFX ## SIZE ## _mc13_ ## CPU; \
2015     c->PFX ## _pixels_tab[IDX][14] = PREFIX ## PFX ## SIZE ## _mc23_ ## CPU; \
2016     c->PFX ## _pixels_tab[IDX][15] = PREFIX ## PFX ## SIZE ## _mc33_ ## CPU; \
2017     } while (0)
2018
2019 #define SET_HPEL_FUNCS(PFX, IDX, SIZE, CPU)                                     \
2020     do {                                                                        \
2021         c->PFX ## _pixels_tab[IDX][0] = PFX ## _pixels ## SIZE ## _     ## CPU; \
2022         c->PFX ## _pixels_tab[IDX][1] = PFX ## _pixels ## SIZE ## _x2_  ## CPU; \
2023         c->PFX ## _pixels_tab[IDX][2] = PFX ## _pixels ## SIZE ## _y2_  ## CPU; \
2024         c->PFX ## _pixels_tab[IDX][3] = PFX ## _pixels ## SIZE ## _xy2_ ## CPU; \
2025     } while (0)
2026
2027 #define H264_QPEL_FUNCS(x, y, CPU)                                                            \
2028     do {                                                                                      \
2029         c->put_h264_qpel_pixels_tab[0][x + y * 4] = put_h264_qpel16_mc ## x ## y ## _ ## CPU; \
2030         c->put_h264_qpel_pixels_tab[1][x + y * 4] = put_h264_qpel8_mc  ## x ## y ## _ ## CPU; \
2031         c->avg_h264_qpel_pixels_tab[0][x + y * 4] = avg_h264_qpel16_mc ## x ## y ## _ ## CPU; \
2032         c->avg_h264_qpel_pixels_tab[1][x + y * 4] = avg_h264_qpel8_mc  ## x ## y ## _ ## CPU; \
2033     } while (0)
2034
2035 #define H264_QPEL_FUNCS_10(x, y, CPU)                                                               \
2036     do {                                                                                            \
2037         c->put_h264_qpel_pixels_tab[0][x + y * 4] = ff_put_h264_qpel16_mc ## x ## y ## _10_ ## CPU; \
2038         c->put_h264_qpel_pixels_tab[1][x + y * 4] = ff_put_h264_qpel8_mc  ## x ## y ## _10_ ## CPU; \
2039         c->avg_h264_qpel_pixels_tab[0][x + y * 4] = ff_avg_h264_qpel16_mc ## x ## y ## _10_ ## CPU; \
2040         c->avg_h264_qpel_pixels_tab[1][x + y * 4] = ff_avg_h264_qpel8_mc  ## x ## y ## _10_ ## CPU; \
2041     } while (0)
2042
2043 static void dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx, int mm_flags)
2044 {
2045     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
2046
2047 #if HAVE_INLINE_ASM
2048     c->put_pixels_clamped        = ff_put_pixels_clamped_mmx;
2049     c->put_signed_pixels_clamped = ff_put_signed_pixels_clamped_mmx;
2050     c->add_pixels_clamped        = ff_add_pixels_clamped_mmx;
2051
2052     if (!high_bit_depth) {
2053         c->clear_block  = clear_block_mmx;
2054         c->clear_blocks = clear_blocks_mmx;
2055         c->draw_edges   = draw_edges_mmx;
2056
2057         SET_HPEL_FUNCS(put,        0, 16, mmx);
2058         SET_HPEL_FUNCS(put_no_rnd, 0, 16, mmx);
2059         SET_HPEL_FUNCS(avg,        0, 16, mmx);
2060         SET_HPEL_FUNCS(avg_no_rnd, 0, 16, mmx);
2061         SET_HPEL_FUNCS(put,        1,  8, mmx);
2062         SET_HPEL_FUNCS(put_no_rnd, 1,  8, mmx);
2063         SET_HPEL_FUNCS(avg,        1,  8, mmx);
2064         SET_HPEL_FUNCS(avg_no_rnd, 1,  8, mmx);
2065     }
2066
2067 #if ARCH_X86_32 || !HAVE_YASM
2068     c->gmc = gmc_mmx;
2069 #endif
2070
2071     c->add_bytes = add_bytes_mmx;
2072
2073     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
2074         c->h263_v_loop_filter = h263_v_loop_filter_mmx;
2075         c->h263_h_loop_filter = h263_h_loop_filter_mmx;
2076     }
2077 #endif /* HAVE_INLINE_ASM */
2078
2079 #if HAVE_YASM
2080     if (!high_bit_depth && CONFIG_H264CHROMA) {
2081         c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_rnd_mmx;
2082         c->put_h264_chroma_pixels_tab[1] = ff_put_h264_chroma_mc4_mmx;
2083     }
2084
2085     c->vector_clip_int32 = ff_vector_clip_int32_mmx;
2086 #endif
2087
2088 }
2089
2090 static void dsputil_init_mmxext(DSPContext *c, AVCodecContext *avctx,
2091                                 int mm_flags)
2092 {
2093     const int bit_depth      = avctx->bits_per_raw_sample;
2094     const int high_bit_depth = bit_depth > 8;
2095
2096 #if HAVE_INLINE_ASM
2097     SET_QPEL_FUNCS(avg_qpel,        0, 16, mmxext, );
2098     SET_QPEL_FUNCS(avg_qpel,        1,  8, mmxext, );
2099
2100     SET_QPEL_FUNCS(put_qpel,        0, 16, mmxext, );
2101     SET_QPEL_FUNCS(put_qpel,        1,  8, mmxext, );
2102     SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, mmxext, );
2103     SET_QPEL_FUNCS(put_no_rnd_qpel, 1,  8, mmxext, );
2104
2105     if (!high_bit_depth) {
2106         c->put_pixels_tab[0][1] = put_pixels16_x2_mmxext;
2107         c->put_pixels_tab[0][2] = put_pixels16_y2_mmxext;
2108
2109         c->avg_pixels_tab[0][0] = avg_pixels16_mmxext;
2110         c->avg_pixels_tab[0][1] = avg_pixels16_x2_mmxext;
2111         c->avg_pixels_tab[0][2] = avg_pixels16_y2_mmxext;
2112
2113         c->put_pixels_tab[1][1] = put_pixels8_x2_mmxext;
2114         c->put_pixels_tab[1][2] = put_pixels8_y2_mmxext;
2115
2116         c->avg_pixels_tab[1][0] = avg_pixels8_mmxext;
2117         c->avg_pixels_tab[1][1] = avg_pixels8_x2_mmxext;
2118         c->avg_pixels_tab[1][2] = avg_pixels8_y2_mmxext;
2119     }
2120
2121     if (!(avctx->flags & CODEC_FLAG_BITEXACT)) {
2122         if (!high_bit_depth) {
2123             c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmxext;
2124             c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmxext;
2125             c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmxext;
2126             c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmxext;
2127
2128             c->avg_pixels_tab[0][3] = avg_pixels16_xy2_mmxext;
2129             c->avg_pixels_tab[1][3] = avg_pixels8_xy2_mmxext;
2130         }
2131     }
2132
2133     if (CONFIG_VP3_DECODER && (avctx->codec_id == AV_CODEC_ID_VP3 ||
2134                                avctx->codec_id == AV_CODEC_ID_THEORA)) {
2135         c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_exact_mmxext;
2136         c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_exact_mmxext;
2137     }
2138 #endif /* HAVE_INLINE_ASM */
2139
2140 #if HAVE_MMXEXT_EXTERNAL
2141     if (CONFIG_H264QPEL) {
2142         if (!high_bit_depth) {
2143             SET_QPEL_FUNCS(put_h264_qpel, 0, 16, mmxext, );
2144             SET_QPEL_FUNCS(put_h264_qpel, 1,  8, mmxext, );
2145             SET_QPEL_FUNCS(put_h264_qpel, 2,  4, mmxext, );
2146             SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, mmxext, );
2147             SET_QPEL_FUNCS(avg_h264_qpel, 1,  8, mmxext, );
2148             SET_QPEL_FUNCS(avg_h264_qpel, 2,  4, mmxext, );
2149         } else if (bit_depth == 10) {
2150 #if !ARCH_X86_64
2151             SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, 10_mmxext, ff_);
2152             SET_QPEL_FUNCS(put_h264_qpel, 0, 16, 10_mmxext, ff_);
2153             SET_QPEL_FUNCS(put_h264_qpel, 1,  8, 10_mmxext, ff_);
2154             SET_QPEL_FUNCS(avg_h264_qpel, 1,  8, 10_mmxext, ff_);
2155 #endif
2156             SET_QPEL_FUNCS(put_h264_qpel, 2, 4,  10_mmxext, ff_);
2157             SET_QPEL_FUNCS(avg_h264_qpel, 2, 4,  10_mmxext, ff_);
2158         }
2159     }
2160
2161     if (!high_bit_depth && CONFIG_H264CHROMA) {
2162         c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_rnd_mmxext;
2163         c->avg_h264_chroma_pixels_tab[1] = ff_avg_h264_chroma_mc4_mmxext;
2164         c->avg_h264_chroma_pixels_tab[2] = ff_avg_h264_chroma_mc2_mmxext;
2165         c->put_h264_chroma_pixels_tab[2] = ff_put_h264_chroma_mc2_mmxext;
2166     }
2167     if (bit_depth == 10 && CONFIG_H264CHROMA) {
2168         c->put_h264_chroma_pixels_tab[2] = ff_put_h264_chroma_mc2_10_mmxext;
2169         c->avg_h264_chroma_pixels_tab[2] = ff_avg_h264_chroma_mc2_10_mmxext;
2170         c->put_h264_chroma_pixels_tab[1] = ff_put_h264_chroma_mc4_10_mmxext;
2171         c->avg_h264_chroma_pixels_tab[1] = ff_avg_h264_chroma_mc4_10_mmxext;
2172     }
2173
2174     /* slower than cmov version on AMD */
2175     if (!(mm_flags & AV_CPU_FLAG_3DNOW))
2176         c->add_hfyu_median_prediction = ff_add_hfyu_median_prediction_mmxext;
2177
2178     c->scalarproduct_int16          = ff_scalarproduct_int16_mmxext;
2179     c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_mmxext;
2180
2181     if (avctx->flags & CODEC_FLAG_BITEXACT) {
2182         c->apply_window_int16 = ff_apply_window_int16_mmxext;
2183     } else {
2184         c->apply_window_int16 = ff_apply_window_int16_round_mmxext;
2185     }
2186 #endif /* HAVE_MMXEXT_EXTERNAL */
2187 }
2188
2189 static void dsputil_init_3dnow(DSPContext *c, AVCodecContext *avctx,
2190                                int mm_flags)
2191 {
2192     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
2193
2194 #if HAVE_INLINE_ASM
2195     if (!high_bit_depth) {
2196         c->put_pixels_tab[0][1] = put_pixels16_x2_3dnow;
2197         c->put_pixels_tab[0][2] = put_pixels16_y2_3dnow;
2198
2199         c->avg_pixels_tab[0][0] = avg_pixels16_3dnow;
2200         c->avg_pixels_tab[0][1] = avg_pixels16_x2_3dnow;
2201         c->avg_pixels_tab[0][2] = avg_pixels16_y2_3dnow;
2202
2203         c->put_pixels_tab[1][1] = put_pixels8_x2_3dnow;
2204         c->put_pixels_tab[1][2] = put_pixels8_y2_3dnow;
2205
2206         c->avg_pixels_tab[1][0] = avg_pixels8_3dnow;
2207         c->avg_pixels_tab[1][1] = avg_pixels8_x2_3dnow;
2208         c->avg_pixels_tab[1][2] = avg_pixels8_y2_3dnow;
2209
2210         if (!(avctx->flags & CODEC_FLAG_BITEXACT)){
2211             c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_3dnow;
2212             c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_3dnow;
2213             c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_3dnow;
2214             c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_3dnow;
2215
2216             c->avg_pixels_tab[0][3] = avg_pixels16_xy2_3dnow;
2217             c->avg_pixels_tab[1][3] = avg_pixels8_xy2_3dnow;
2218         }
2219     }
2220
2221     if (CONFIG_VP3_DECODER && (avctx->codec_id == AV_CODEC_ID_VP3 ||
2222                                avctx->codec_id == AV_CODEC_ID_THEORA)) {
2223         c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_exact_3dnow;
2224         c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_exact_3dnow;
2225     }
2226 #endif /* HAVE_INLINE_ASM */
2227
2228 #if HAVE_YASM
2229     if (!high_bit_depth && CONFIG_H264CHROMA) {
2230         c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_rnd_3dnow;
2231         c->avg_h264_chroma_pixels_tab[1] = ff_avg_h264_chroma_mc4_3dnow;
2232     }
2233 #endif /* HAVE_YASM */
2234 }
2235
2236 static void dsputil_init_sse(DSPContext *c, AVCodecContext *avctx, int mm_flags)
2237 {
2238     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
2239
2240 #if HAVE_INLINE_ASM
2241     if (!high_bit_depth) {
2242         if (!(CONFIG_MPEG_XVMC_DECODER && avctx->xvmc_acceleration > 1)) {
2243             /* XvMCCreateBlocks() may not allocate 16-byte aligned blocks */
2244             c->clear_block  = clear_block_sse;
2245             c->clear_blocks = clear_blocks_sse;
2246         }
2247     }
2248
2249     c->vector_clipf = vector_clipf_sse;
2250 #endif /* HAVE_INLINE_ASM */
2251
2252 #if HAVE_YASM
2253 #if HAVE_INLINE_ASM && CONFIG_VIDEODSP
2254     c->gmc = gmc_sse;
2255 #endif
2256 #endif /* HAVE_YASM */
2257 }
2258
2259 static void dsputil_init_sse2(DSPContext *c, AVCodecContext *avctx,
2260                               int mm_flags)
2261 {
2262     const int bit_depth      = avctx->bits_per_raw_sample;
2263     const int high_bit_depth = bit_depth > 8;
2264
2265 #if HAVE_SSE2_INLINE
2266     if (!high_bit_depth && avctx->idct_algo == FF_IDCT_XVIDMMX) {
2267         c->idct_put              = ff_idct_xvid_sse2_put;
2268         c->idct_add              = ff_idct_xvid_sse2_add;
2269         c->idct                  = ff_idct_xvid_sse2;
2270         c->idct_permutation_type = FF_SSE2_IDCT_PERM;
2271     }
2272 #endif /* HAVE_SSE2_INLINE */
2273
2274 #if HAVE_SSE2_EXTERNAL
2275     if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {
2276         // these functions are slower than mmx on AMD, but faster on Intel
2277         if (!high_bit_depth) {
2278             c->put_pixels_tab[0][0]        = ff_put_pixels16_sse2;
2279             c->put_no_rnd_pixels_tab[0][0] = ff_put_pixels16_sse2;
2280             c->avg_pixels_tab[0][0]        = ff_avg_pixels16_sse2;
2281             if (CONFIG_H264QPEL)
2282                 H264_QPEL_FUNCS(0, 0, sse2);
2283         }
2284     }
2285
2286     if (!high_bit_depth && CONFIG_H264QPEL) {
2287         H264_QPEL_FUNCS(0, 1, sse2);
2288         H264_QPEL_FUNCS(0, 2, sse2);
2289         H264_QPEL_FUNCS(0, 3, sse2);
2290         H264_QPEL_FUNCS(1, 1, sse2);
2291         H264_QPEL_FUNCS(1, 2, sse2);
2292         H264_QPEL_FUNCS(1, 3, sse2);
2293         H264_QPEL_FUNCS(2, 1, sse2);
2294         H264_QPEL_FUNCS(2, 2, sse2);
2295         H264_QPEL_FUNCS(2, 3, sse2);
2296         H264_QPEL_FUNCS(3, 1, sse2);
2297         H264_QPEL_FUNCS(3, 2, sse2);
2298         H264_QPEL_FUNCS(3, 3, sse2);
2299     }
2300
2301     if (bit_depth == 10) {
2302         if (CONFIG_H264QPEL) {
2303             SET_QPEL_FUNCS(put_h264_qpel, 0, 16, 10_sse2, ff_);
2304             SET_QPEL_FUNCS(put_h264_qpel, 1,  8, 10_sse2, ff_);
2305             SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, 10_sse2, ff_);
2306             SET_QPEL_FUNCS(avg_h264_qpel, 1,  8, 10_sse2, ff_);
2307             H264_QPEL_FUNCS_10(1, 0, sse2_cache64);
2308             H264_QPEL_FUNCS_10(2, 0, sse2_cache64);
2309             H264_QPEL_FUNCS_10(3, 0, sse2_cache64);
2310         }
2311         if (CONFIG_H264CHROMA) {
2312             c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_10_sse2;
2313             c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_10_sse2;
2314         }
2315     }
2316
2317     c->scalarproduct_int16          = ff_scalarproduct_int16_sse2;
2318     c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_sse2;
2319     if (mm_flags & AV_CPU_FLAG_ATOM) {
2320         c->vector_clip_int32 = ff_vector_clip_int32_int_sse2;
2321     } else {
2322         c->vector_clip_int32 = ff_vector_clip_int32_sse2;
2323     }
2324     if (avctx->flags & CODEC_FLAG_BITEXACT) {
2325         c->apply_window_int16 = ff_apply_window_int16_sse2;
2326     } else if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {
2327         c->apply_window_int16 = ff_apply_window_int16_round_sse2;
2328     }
2329     c->bswap_buf = ff_bswap32_buf_sse2;
2330 #endif /* HAVE_SSE2_EXTERNAL */
2331 }
2332
2333 static void dsputil_init_ssse3(DSPContext *c, AVCodecContext *avctx,
2334                                int mm_flags)
2335 {
2336 #if HAVE_SSSE3_EXTERNAL
2337     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
2338     const int bit_depth      = avctx->bits_per_raw_sample;
2339
2340     if (!high_bit_depth && CONFIG_H264QPEL) {
2341         H264_QPEL_FUNCS(1, 0, ssse3);
2342         H264_QPEL_FUNCS(1, 1, ssse3);
2343         H264_QPEL_FUNCS(1, 2, ssse3);
2344         H264_QPEL_FUNCS(1, 3, ssse3);
2345         H264_QPEL_FUNCS(2, 0, ssse3);
2346         H264_QPEL_FUNCS(2, 1, ssse3);
2347         H264_QPEL_FUNCS(2, 2, ssse3);
2348         H264_QPEL_FUNCS(2, 3, ssse3);
2349         H264_QPEL_FUNCS(3, 0, ssse3);
2350         H264_QPEL_FUNCS(3, 1, ssse3);
2351         H264_QPEL_FUNCS(3, 2, ssse3);
2352         H264_QPEL_FUNCS(3, 3, ssse3);
2353     }
2354     if (bit_depth == 10 && CONFIG_H264QPEL) {
2355         H264_QPEL_FUNCS_10(1, 0, ssse3_cache64);
2356         H264_QPEL_FUNCS_10(2, 0, ssse3_cache64);
2357         H264_QPEL_FUNCS_10(3, 0, ssse3_cache64);
2358     }
2359     if (!high_bit_depth && CONFIG_H264CHROMA) {
2360         c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_rnd_ssse3;
2361         c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_rnd_ssse3;
2362         c->put_h264_chroma_pixels_tab[1] = ff_put_h264_chroma_mc4_ssse3;
2363         c->avg_h264_chroma_pixels_tab[1] = ff_avg_h264_chroma_mc4_ssse3;
2364     }
2365     c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_ssse3;
2366     if (mm_flags & AV_CPU_FLAG_SSE4) // not really sse4, just slow on Conroe
2367         c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_sse4;
2368
2369     if (mm_flags & AV_CPU_FLAG_ATOM)
2370         c->apply_window_int16 = ff_apply_window_int16_ssse3_atom;
2371     else
2372         c->apply_window_int16 = ff_apply_window_int16_ssse3;
2373     if (!(mm_flags & (AV_CPU_FLAG_SSE42|AV_CPU_FLAG_3DNOW))) // cachesplit
2374         c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_ssse3;
2375     c->bswap_buf = ff_bswap32_buf_ssse3;
2376 #endif /* HAVE_SSSE3_EXTERNAL */
2377 }
2378
2379 static void dsputil_init_sse4(DSPContext *c, AVCodecContext *avctx,
2380                               int mm_flags)
2381 {
2382 #if HAVE_SSE4_EXTERNAL
2383     c->vector_clip_int32 = ff_vector_clip_int32_sse4;
2384 #endif /* HAVE_SSE4_EXTERNAL */
2385 }
2386
2387 static void dsputil_init_avx(DSPContext *c, AVCodecContext *avctx, int mm_flags)
2388 {
2389 #if HAVE_AVX_EXTERNAL
2390     const int bit_depth = avctx->bits_per_raw_sample;
2391
2392     if (bit_depth == 10) {
2393         // AVX implies !cache64.
2394         // TODO: Port cache(32|64) detection from x264.
2395         if (CONFIG_H264QPEL) {
2396             H264_QPEL_FUNCS_10(1, 0, sse2);
2397             H264_QPEL_FUNCS_10(2, 0, sse2);
2398             H264_QPEL_FUNCS_10(3, 0, sse2);
2399         }
2400
2401         if (CONFIG_H264CHROMA) {
2402             c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_10_avx;
2403             c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_10_avx;
2404         }
2405     }
2406 #endif /* HAVE_AVX_EXTERNAL */
2407 }
2408
2409 void ff_dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx)
2410 {
2411     int mm_flags = av_get_cpu_flags();
2412
2413 #if HAVE_7REGS && HAVE_INLINE_ASM
2414     if (mm_flags & AV_CPU_FLAG_CMOV)
2415         c->add_hfyu_median_prediction = add_hfyu_median_prediction_cmov;
2416 #endif
2417
2418     if (mm_flags & AV_CPU_FLAG_MMX) {
2419 #if HAVE_INLINE_ASM
2420         const int idct_algo = avctx->idct_algo;
2421
2422         if (avctx->lowres == 0 && avctx->bits_per_raw_sample <= 8) {
2423             if (idct_algo == FF_IDCT_AUTO || idct_algo == FF_IDCT_SIMPLEMMX) {
2424                 c->idct_put              = ff_simple_idct_put_mmx;
2425                 c->idct_add              = ff_simple_idct_add_mmx;
2426                 c->idct                  = ff_simple_idct_mmx;
2427                 c->idct_permutation_type = FF_SIMPLE_IDCT_PERM;
2428 #if CONFIG_GPL
2429             } else if (idct_algo == FF_IDCT_LIBMPEG2MMX) {
2430                 if (mm_flags & AV_CPU_FLAG_MMX2) {
2431                     c->idct_put = ff_libmpeg2mmx2_idct_put;
2432                     c->idct_add = ff_libmpeg2mmx2_idct_add;
2433                     c->idct     = ff_mmxext_idct;
2434                 } else {
2435                     c->idct_put = ff_libmpeg2mmx_idct_put;
2436                     c->idct_add = ff_libmpeg2mmx_idct_add;
2437                     c->idct     = ff_mmx_idct;
2438                 }
2439                 c->idct_permutation_type = FF_LIBMPEG2_IDCT_PERM;
2440 #endif
2441             } else if (idct_algo == FF_IDCT_XVIDMMX) {
2442                 if (mm_flags & AV_CPU_FLAG_SSE2) {
2443                     c->idct_put              = ff_idct_xvid_sse2_put;
2444                     c->idct_add              = ff_idct_xvid_sse2_add;
2445                     c->idct                  = ff_idct_xvid_sse2;
2446                     c->idct_permutation_type = FF_SSE2_IDCT_PERM;
2447                 } else if (mm_flags & AV_CPU_FLAG_MMXEXT) {
2448                     c->idct_put              = ff_idct_xvid_mmxext_put;
2449                     c->idct_add              = ff_idct_xvid_mmxext_add;
2450                     c->idct                  = ff_idct_xvid_mmxext;
2451                 } else {
2452                     c->idct_put              = ff_idct_xvid_mmx_put;
2453                     c->idct_add              = ff_idct_xvid_mmx_add;
2454                     c->idct                  = ff_idct_xvid_mmx;
2455                 }
2456             }
2457         }
2458 #endif /* HAVE_INLINE_ASM */
2459
2460         dsputil_init_mmx(c, avctx, mm_flags);
2461     }
2462
2463     if (mm_flags & AV_CPU_FLAG_MMXEXT)
2464         dsputil_init_mmxext(c, avctx, mm_flags);
2465
2466     if (mm_flags & AV_CPU_FLAG_3DNOW)
2467         dsputil_init_3dnow(c, avctx, mm_flags);
2468
2469     if (mm_flags & AV_CPU_FLAG_SSE)
2470         dsputil_init_sse(c, avctx, mm_flags);
2471
2472     if (mm_flags & AV_CPU_FLAG_SSE2)
2473         dsputil_init_sse2(c, avctx, mm_flags);
2474
2475     if (mm_flags & AV_CPU_FLAG_SSSE3)
2476         dsputil_init_ssse3(c, avctx, mm_flags);
2477
2478     if (mm_flags & AV_CPU_FLAG_SSE4)
2479         dsputil_init_sse4(c, avctx, mm_flags);
2480
2481     if (mm_flags & AV_CPU_FLAG_AVX)
2482         dsputil_init_avx(c, avctx, mm_flags);
2483
2484     if (CONFIG_ENCODERS)
2485         ff_dsputilenc_init_mmx(c, avctx);
2486 }