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  * MMX optimization by Nick Kurshev <nickols_k@mail.ru>
   7  *
   8  * This file is part of Libav.
   9  *
  10  * Libav is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2.1 of the License, or (at your option) any later version.
  14  *
  15  * Libav is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * Lesser General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU Lesser General Public
  21  * License along with Libav; if not, write to the Free Software
  22  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  23  */
  24
  25 #include "libavutil/cpu.h"
  26 #include "libavutil/x86_cpu.h"
  27 #include "libavcodec/dsputil.h"
  28 #include "libavcodec/h264dsp.h"
  29 #include "libavcodec/mpegvideo.h"
  30 #include "libavcodec/simple_idct.h"
  31 #include "libavcodec/ac3dec.h"
  32 #include "dsputil_mmx.h"
  33 #include "idct_xvid.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 #define JUMPALIGN()     __asm__ volatile (".p2align 3"::)
  88 #define MOVQ_ZERO(regd) __asm__ volatile ("pxor %%"#regd", %%"#regd ::)
  89
  90 #define MOVQ_BFE(regd)                                  \
  91     __asm__ volatile (                                  \
  92         "pcmpeqd %%"#regd", %%"#regd"   \n\t"           \
  93         "paddb   %%"#regd", %%"#regd"   \n\t" ::)
  94
  95 #ifndef PIC
  96 #define MOVQ_BONE(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_bone))
  97 #define MOVQ_WTWO(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_wtwo))
  98 #else
  99 // for shared library it's better to use this way for accessing constants
 100 // pcmpeqd -> -1
 101 #define MOVQ_BONE(regd)                                 \
 102     __asm__ volatile (                                  \
 103         "pcmpeqd  %%"#regd", %%"#regd"  \n\t"           \
 104         "psrlw          $15, %%"#regd"  \n\t"           \
 105         "packuswb %%"#regd", %%"#regd"  \n\t" ::)
 106
 107 #define MOVQ_WTWO(regd)                                 \
 108     __asm__ volatile (                                  \
 109         "pcmpeqd %%"#regd", %%"#regd"   \n\t"           \
 110         "psrlw         $15, %%"#regd"   \n\t"           \
 111         "psllw          $1, %%"#regd"   \n\t"::)
 112
 113 #endif
 114
 115 // using regr as temporary and for the output result
 116 // first argument is unmodifed and second is trashed
 117 // regfe is supposed to contain 0xfefefefefefefefe
 118 #define PAVGB_MMX_NO_RND(rega, regb, regr, regfe)                \
 119     "movq   "#rega", "#regr"            \n\t"                    \
 120     "pand   "#regb", "#regr"            \n\t"                    \
 121     "pxor   "#rega", "#regb"            \n\t"                    \
 122     "pand  "#regfe", "#regb"            \n\t"                    \
 123     "psrlq       $1, "#regb"            \n\t"                    \
 124     "paddb  "#regb", "#regr"            \n\t"
 125
 126 #define PAVGB_MMX(rega, regb, regr, regfe)                       \
 127     "movq   "#rega", "#regr"            \n\t"                    \
 128     "por    "#regb", "#regr"            \n\t"                    \
 129     "pxor   "#rega", "#regb"            \n\t"                    \
 130     "pand  "#regfe", "#regb"            \n\t"                    \
 131     "psrlq       $1, "#regb"            \n\t"                    \
 132     "psubb  "#regb", "#regr"            \n\t"
 133
 134 // mm6 is supposed to contain 0xfefefefefefefefe
 135 #define PAVGBP_MMX_NO_RND(rega, regb, regr,  regc, regd, regp)   \
 136     "movq  "#rega", "#regr"             \n\t"                    \
 137     "movq  "#regc", "#regp"             \n\t"                    \
 138     "pand  "#regb", "#regr"             \n\t"                    \
 139     "pand  "#regd", "#regp"             \n\t"                    \
 140     "pxor  "#rega", "#regb"             \n\t"                    \
 141     "pxor  "#regc", "#regd"             \n\t"                    \
 142     "pand    %%mm6, "#regb"             \n\t"                    \
 143     "pand    %%mm6, "#regd"             \n\t"                    \
 144     "psrlq      $1, "#regb"             \n\t"                    \
 145     "psrlq      $1, "#regd"             \n\t"                    \
 146     "paddb "#regb", "#regr"             \n\t"                    \
 147     "paddb "#regd", "#regp"             \n\t"
 148
 149 #define PAVGBP_MMX(rega, regb, regr, regc, regd, regp)           \
 150     "movq  "#rega", "#regr"             \n\t"                    \
 151     "movq  "#regc", "#regp"             \n\t"                    \
 152     "por   "#regb", "#regr"             \n\t"                    \
 153     "por   "#regd", "#regp"             \n\t"                    \
 154     "pxor  "#rega", "#regb"             \n\t"                    \
 155     "pxor  "#regc", "#regd"             \n\t"                    \
 156     "pand    %%mm6, "#regb"             \n\t"                    \
 157     "pand    %%mm6, "#regd"             \n\t"                    \
 158     "psrlq      $1, "#regd"             \n\t"                    \
 159     "psrlq      $1, "#regb"             \n\t"                    \
 160     "psubb "#regb", "#regr"             \n\t"                    \
 161     "psubb "#regd", "#regp"             \n\t"
 162
 163 /***********************************/
 164 /* MMX no rounding */
 165 #define DEF(x, y) x ## _no_rnd_ ## y ## _mmx
 166 #define SET_RND  MOVQ_WONE
 167 #define PAVGBP(a, b, c, d, e, f)        PAVGBP_MMX_NO_RND(a, b, c, d, e, f)
 168 #define PAVGB(a, b, c, e)               PAVGB_MMX_NO_RND(a, b, c, e)
 169 #define OP_AVG(a, b, c, e)              PAVGB_MMX(a, b, c, e)
 170
 171 #include "dsputil_mmx_rnd_template.c"
 172
 173 #undef DEF
 174 #undef SET_RND
 175 #undef PAVGBP
 176 #undef PAVGB
 177 /***********************************/
 178 /* MMX rounding */
 179
 180 #define DEF(x, y) x ## _ ## y ## _mmx
 181 #define SET_RND  MOVQ_WTWO
 182 #define PAVGBP(a, b, c, d, e, f)        PAVGBP_MMX(a, b, c, d, e, f)
 183 #define PAVGB(a, b, c, e)               PAVGB_MMX(a, b, c, e)
 184
 185 #include "dsputil_mmx_rnd_template.c"
 186
 187 #undef DEF
 188 #undef SET_RND
 189 #undef PAVGBP
 190 #undef PAVGB
 191 #undef OP_AVG
 192
 193 /***********************************/
 194 /* 3Dnow specific */
 195
 196 #define DEF(x) x ## _3dnow
 197 #define PAVGB "pavgusb"
 198 #define OP_AVG PAVGB
 199
 200 #include "dsputil_mmx_avg_template.c"
 201
 202 #undef DEF
 203 #undef PAVGB
 204 #undef OP_AVG
 205
 206 /***********************************/
 207 /* MMX2 specific */
 208
 209 #define DEF(x) x ## _mmx2
 210
 211 /* Introduced only in MMX2 set */
 212 #define PAVGB "pavgb"
 213 #define OP_AVG PAVGB
 214
 215 #include "dsputil_mmx_avg_template.c"
 216
 217 #undef DEF
 218 #undef PAVGB
 219 #undef OP_AVG
 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_mmx2 put_pixels16_mmx
 224 #define put_pixels8_mmx2 put_pixels8_mmx
 225 #define put_pixels4_mmx2 put_pixels4_mmx
 226 #define put_no_rnd_pixels16_mmx2 put_no_rnd_pixels16_mmx
 227 #define put_no_rnd_pixels8_mmx2 put_no_rnd_pixels8_mmx
 228 #define put_pixels16_3dnow put_pixels16_mmx
 229 #define put_pixels8_3dnow put_pixels8_mmx
 230 #define put_pixels4_3dnow put_pixels4_mmx
 231 #define put_no_rnd_pixels16_3dnow put_no_rnd_pixels16_mmx
 232 #define put_no_rnd_pixels8_3dnow put_no_rnd_pixels8_mmx
 233
 234 /***********************************/
 235 /* standard MMX */
 236
 237 void ff_put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels,
 238                                int line_size)
 239 {
 240     const DCTELEM *p;
 241     uint8_t *pix;
 242
 243     /* read the pixels */
 244     p   = block;
 245     pix = pixels;
 246     /* unrolled loop */
 247     __asm__ volatile (
 248         "movq        %3, %%mm0          \n\t"
 249         "movq       8%3, %%mm1          \n\t"
 250         "movq      16%3, %%mm2          \n\t"
 251         "movq      24%3, %%mm3          \n\t"
 252         "movq      32%3, %%mm4          \n\t"
 253         "movq      40%3, %%mm5          \n\t"
 254         "movq      48%3, %%mm6          \n\t"
 255         "movq      56%3, %%mm7          \n\t"
 256         "packuswb %%mm1, %%mm0          \n\t"
 257         "packuswb %%mm3, %%mm2          \n\t"
 258         "packuswb %%mm5, %%mm4          \n\t"
 259         "packuswb %%mm7, %%mm6          \n\t"
 260         "movq     %%mm0, (%0)           \n\t"
 261         "movq     %%mm2, (%0, %1)       \n\t"
 262         "movq     %%mm4, (%0, %1, 2)    \n\t"
 263         "movq     %%mm6, (%0, %2)       \n\t"
 264         :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3),
 265            "m"(*p)
 266         : "memory");
 267     pix += line_size * 4;
 268     p   += 32;
 269
 270     // if here would be an exact copy of the code above
 271     // compiler would generate some very strange code
 272     // thus using "r"
 273     __asm__ volatile (
 274         "movq       (%3), %%mm0         \n\t"
 275         "movq      8(%3), %%mm1         \n\t"
 276         "movq     16(%3), %%mm2         \n\t"
 277         "movq     24(%3), %%mm3         \n\t"
 278         "movq     32(%3), %%mm4         \n\t"
 279         "movq     40(%3), %%mm5         \n\t"
 280         "movq     48(%3), %%mm6         \n\t"
 281         "movq     56(%3), %%mm7         \n\t"
 282         "packuswb  %%mm1, %%mm0         \n\t"
 283         "packuswb  %%mm3, %%mm2         \n\t"
 284         "packuswb  %%mm5, %%mm4         \n\t"
 285         "packuswb  %%mm7, %%mm6         \n\t"
 286         "movq      %%mm0, (%0)          \n\t"
 287         "movq      %%mm2, (%0, %1)      \n\t"
 288         "movq      %%mm4, (%0, %1, 2)   \n\t"
 289         "movq      %%mm6, (%0, %2)      \n\t"
 290         :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3), "r"(p)
 291         : "memory");
 292 }
 293
 294 #define put_signed_pixels_clamped_mmx_half(off)             \
 295     "movq          "#off"(%2), %%mm1        \n\t"           \
 296     "movq     16 + "#off"(%2), %%mm2        \n\t"           \
 297     "movq     32 + "#off"(%2), %%mm3        \n\t"           \
 298     "movq     48 + "#off"(%2), %%mm4        \n\t"           \
 299     "packsswb  8 + "#off"(%2), %%mm1        \n\t"           \
 300     "packsswb 24 + "#off"(%2), %%mm2        \n\t"           \
 301     "packsswb 40 + "#off"(%2), %%mm3        \n\t"           \
 302     "packsswb 56 + "#off"(%2), %%mm4        \n\t"           \
 303     "paddb              %%mm0, %%mm1        \n\t"           \
 304     "paddb              %%mm0, %%mm2        \n\t"           \
 305     "paddb              %%mm0, %%mm3        \n\t"           \
 306     "paddb              %%mm0, %%mm4        \n\t"           \
 307     "movq               %%mm1, (%0)         \n\t"           \
 308     "movq               %%mm2, (%0, %3)     \n\t"           \
 309     "movq               %%mm3, (%0, %3, 2)  \n\t"           \
 310     "movq               %%mm4, (%0, %1)     \n\t"
 311
 312 void ff_put_signed_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels,
 313                                       int line_size)
 314 {
 315     x86_reg line_skip = line_size;
 316     x86_reg line_skip3;
 317
 318     __asm__ volatile (
 319         "movq "MANGLE(ff_pb_80)", %%mm0     \n\t"
 320         "lea         (%3, %3, 2), %1        \n\t"
 321         put_signed_pixels_clamped_mmx_half(0)
 322         "lea         (%0, %3, 4), %0        \n\t"
 323         put_signed_pixels_clamped_mmx_half(64)
 324         : "+&r"(pixels), "=&r"(line_skip3)
 325         : "r"(block), "r"(line_skip)
 326         : "memory");
 327 }
 328
 329 void ff_add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels,
 330                                int line_size)
 331 {
 332     const DCTELEM *p;
 333     uint8_t *pix;
 334     int i;
 335
 336     /* read the pixels */
 337     p   = block;
 338     pix = pixels;
 339     MOVQ_ZERO(mm7);
 340     i = 4;
 341     do {
 342         __asm__ volatile (
 343             "movq        (%2), %%mm0    \n\t"
 344             "movq       8(%2), %%mm1    \n\t"
 345             "movq      16(%2), %%mm2    \n\t"
 346             "movq      24(%2), %%mm3    \n\t"
 347             "movq          %0, %%mm4    \n\t"
 348             "movq          %1, %%mm6    \n\t"
 349             "movq       %%mm4, %%mm5    \n\t"
 350             "punpcklbw  %%mm7, %%mm4    \n\t"
 351             "punpckhbw  %%mm7, %%mm5    \n\t"
 352             "paddsw     %%mm4, %%mm0    \n\t"
 353             "paddsw     %%mm5, %%mm1    \n\t"
 354             "movq       %%mm6, %%mm5    \n\t"
 355             "punpcklbw  %%mm7, %%mm6    \n\t"
 356             "punpckhbw  %%mm7, %%mm5    \n\t"
 357             "paddsw     %%mm6, %%mm2    \n\t"
 358             "paddsw     %%mm5, %%mm3    \n\t"
 359             "packuswb   %%mm1, %%mm0    \n\t"
 360             "packuswb   %%mm3, %%mm2    \n\t"
 361             "movq       %%mm0, %0       \n\t"
 362             "movq       %%mm2, %1       \n\t"
 363             : "+m"(*pix), "+m"(*(pix + line_size))
 364             : "r"(p)
 365             : "memory");
 366         pix += line_size * 2;
 367         p   += 16;
 368     } while (--i);
 369 }
 370
 371 static void put_pixels4_mmx(uint8_t *block, const uint8_t *pixels,
 372                             int line_size, int h)
 373 {
 374     __asm__ volatile (
 375         "lea   (%3, %3), %%"REG_a"      \n\t"
 376         ".p2align     3                 \n\t"
 377         "1:                             \n\t"
 378         "movd  (%1    ), %%mm0          \n\t"
 379         "movd  (%1, %3), %%mm1          \n\t"
 380         "movd     %%mm0, (%2)           \n\t"
 381         "movd     %%mm1, (%2, %3)       \n\t"
 382         "add  %%"REG_a", %1             \n\t"
 383         "add  %%"REG_a", %2             \n\t"
 384         "movd  (%1    ), %%mm0          \n\t"
 385         "movd  (%1, %3), %%mm1          \n\t"
 386         "movd     %%mm0, (%2)           \n\t"
 387         "movd     %%mm1, (%2, %3)       \n\t"
 388         "add  %%"REG_a", %1             \n\t"
 389         "add  %%"REG_a", %2             \n\t"
 390         "subl        $4, %0             \n\t"
 391         "jnz         1b                 \n\t"
 392         : "+g"(h), "+r"(pixels),  "+r"(block)
 393         : "r"((x86_reg)line_size)
 394         : "%"REG_a, "memory"
 395         );
 396 }
 397
 398 static void put_pixels8_mmx(uint8_t *block, const uint8_t *pixels,
 399                             int line_size, int h)
 400 {
 401     __asm__ volatile (
 402         "lea   (%3, %3), %%"REG_a"      \n\t"
 403         ".p2align     3                 \n\t"
 404         "1:                             \n\t"
 405         "movq  (%1    ), %%mm0          \n\t"
 406         "movq  (%1, %3), %%mm1          \n\t"
 407         "movq     %%mm0, (%2)           \n\t"
 408         "movq     %%mm1, (%2, %3)       \n\t"
 409         "add  %%"REG_a", %1             \n\t"
 410         "add  %%"REG_a", %2             \n\t"
 411         "movq  (%1    ), %%mm0          \n\t"
 412         "movq  (%1, %3), %%mm1          \n\t"
 413         "movq     %%mm0, (%2)           \n\t"
 414         "movq     %%mm1, (%2, %3)       \n\t"
 415         "add  %%"REG_a", %1             \n\t"
 416         "add  %%"REG_a", %2             \n\t"
 417         "subl        $4, %0             \n\t"
 418         "jnz         1b                 \n\t"
 419         : "+g"(h), "+r"(pixels),  "+r"(block)
 420         : "r"((x86_reg)line_size)
 421         : "%"REG_a, "memory"
 422         );
 423 }
 424
 425 static void put_pixels16_mmx(uint8_t *block, const uint8_t *pixels,
 426                              int line_size, int h)
 427 {
 428     __asm__ volatile (
 429         "lea   (%3, %3), %%"REG_a"      \n\t"
 430         ".p2align     3                 \n\t"
 431         "1:                             \n\t"
 432         "movq  (%1    ), %%mm0          \n\t"
 433         "movq 8(%1    ), %%mm4          \n\t"
 434         "movq  (%1, %3), %%mm1          \n\t"
 435         "movq 8(%1, %3), %%mm5          \n\t"
 436         "movq     %%mm0,  (%2)          \n\t"
 437         "movq     %%mm4, 8(%2)          \n\t"
 438         "movq     %%mm1,  (%2, %3)      \n\t"
 439         "movq     %%mm5, 8(%2, %3)      \n\t"
 440         "add  %%"REG_a", %1             \n\t"
 441         "add  %%"REG_a", %2             \n\t"
 442         "movq  (%1    ), %%mm0          \n\t"
 443         "movq 8(%1    ), %%mm4          \n\t"
 444         "movq  (%1, %3), %%mm1          \n\t"
 445         "movq 8(%1, %3), %%mm5          \n\t"
 446         "movq     %%mm0,  (%2)          \n\t"
 447         "movq     %%mm4, 8(%2)          \n\t"
 448         "movq     %%mm1,  (%2, %3)      \n\t"
 449         "movq     %%mm5, 8(%2, %3)      \n\t"
 450         "add  %%"REG_a", %1             \n\t"
 451         "add  %%"REG_a", %2             \n\t"
 452         "subl        $4, %0             \n\t"
 453         "jnz         1b                 \n\t"
 454         : "+g"(h), "+r"(pixels),  "+r"(block)
 455         : "r"((x86_reg)line_size)
 456         : "%"REG_a, "memory"
 457         );
 458 }
 459
 460 static void put_pixels16_sse2(uint8_t *block, const uint8_t *pixels,
 461                               int line_size, int h)
 462 {
 463     __asm__ volatile (
 464         "1:                              \n\t"
 465         "movdqu (%1       ), %%xmm0      \n\t"
 466         "movdqu (%1, %3   ), %%xmm1      \n\t"
 467         "movdqu (%1, %3, 2), %%xmm2      \n\t"
 468         "movdqu (%1, %4   ), %%xmm3      \n\t"
 469         "lea    (%1, %3, 4), %1          \n\t"
 470         "movdqa      %%xmm0, (%2)        \n\t"
 471         "movdqa      %%xmm1, (%2, %3)    \n\t"
 472         "movdqa      %%xmm2, (%2, %3, 2) \n\t"
 473         "movdqa      %%xmm3, (%2, %4)    \n\t"
 474         "subl            $4, %0          \n\t"
 475         "lea    (%2, %3, 4), %2          \n\t"
 476         "jnz             1b              \n\t"
 477         : "+g"(h), "+r"(pixels),  "+r"(block)
 478         : "r"((x86_reg)line_size), "r"((x86_reg)3L * line_size)
 479         : "memory"
 480         );
 481 }
 482
 483 static void avg_pixels16_sse2(uint8_t *block, const uint8_t *pixels,
 484                               int line_size, int h)
 485 {
 486     __asm__ volatile (
 487         "1:                                 \n\t"
 488         "movdqu (%1       ), %%xmm0         \n\t"
 489         "movdqu (%1, %3   ), %%xmm1         \n\t"
 490         "movdqu (%1, %3, 2), %%xmm2         \n\t"
 491         "movdqu (%1, %4   ), %%xmm3         \n\t"
 492         "lea    (%1, %3, 4), %1             \n\t"
 493         "pavgb  (%2       ), %%xmm0         \n\t"
 494         "pavgb  (%2, %3   ), %%xmm1         \n\t"
 495         "pavgb  (%2, %3, 2), %%xmm2         \n\t"
 496         "pavgb     (%2, %4), %%xmm3         \n\t"
 497         "movdqa      %%xmm0, (%2)           \n\t"
 498         "movdqa      %%xmm1, (%2, %3)       \n\t"
 499         "movdqa      %%xmm2, (%2, %3, 2)    \n\t"
 500         "movdqa      %%xmm3, (%2, %4)       \n\t"
 501         "subl            $4, %0             \n\t"
 502         "lea    (%2, %3, 4), %2             \n\t"
 503         "jnz             1b                 \n\t"
 504         : "+g"(h), "+r"(pixels),  "+r"(block)
 505         : "r"((x86_reg)line_size), "r"((x86_reg)3L * line_size)
 506         : "memory"
 507         );
 508 }
 509
 510 #define CLEAR_BLOCKS(name, n)                           \
 511 static void name(DCTELEM *blocks)                       \
 512 {                                                       \
 513     __asm__ volatile (                                  \
 514         "pxor %%mm7, %%mm7              \n\t"           \
 515         "mov     %1,        %%"REG_a"   \n\t"           \
 516         "1:                             \n\t"           \
 517         "movq %%mm7,   (%0, %%"REG_a")  \n\t"           \
 518         "movq %%mm7,  8(%0, %%"REG_a")  \n\t"           \
 519         "movq %%mm7, 16(%0, %%"REG_a")  \n\t"           \
 520         "movq %%mm7, 24(%0, %%"REG_a")  \n\t"           \
 521         "add    $32, %%"REG_a"          \n\t"           \
 522         "js      1b                     \n\t"           \
 523         :: "r"(((uint8_t *)blocks) + 128 * n),          \
 524            "i"(-128 * n)                                \
 525         : "%"REG_a                                      \
 526         );                                              \
 527 }
 528 CLEAR_BLOCKS(clear_blocks_mmx, 6)
 529 CLEAR_BLOCKS(clear_block_mmx, 1)
 530
 531 static void clear_block_sse(DCTELEM *block)
 532 {
 533     __asm__ volatile (
 534         "xorps  %%xmm0, %%xmm0          \n"
 535         "movaps %%xmm0,    (%0)         \n"
 536         "movaps %%xmm0,  16(%0)         \n"
 537         "movaps %%xmm0,  32(%0)         \n"
 538         "movaps %%xmm0,  48(%0)         \n"
 539         "movaps %%xmm0,  64(%0)         \n"
 540         "movaps %%xmm0,  80(%0)         \n"
 541         "movaps %%xmm0,  96(%0)         \n"
 542         "movaps %%xmm0, 112(%0)         \n"
 543         :: "r"(block)
 544         : "memory"
 545     );
 546 }
 547
 548 static void clear_blocks_sse(DCTELEM *blocks)
 549 {
 550     __asm__ volatile (
 551         "xorps  %%xmm0, %%xmm0              \n"
 552         "mov        %1,         %%"REG_a"   \n"
 553         "1:                                 \n"
 554         "movaps %%xmm0,    (%0, %%"REG_a")  \n"
 555         "movaps %%xmm0,  16(%0, %%"REG_a")  \n"
 556         "movaps %%xmm0,  32(%0, %%"REG_a")  \n"
 557         "movaps %%xmm0,  48(%0, %%"REG_a")  \n"
 558         "movaps %%xmm0,  64(%0, %%"REG_a")  \n"
 559         "movaps %%xmm0,  80(%0, %%"REG_a")  \n"
 560         "movaps %%xmm0,  96(%0, %%"REG_a")  \n"
 561         "movaps %%xmm0, 112(%0, %%"REG_a")  \n"
 562         "add      $128,         %%"REG_a"   \n"
 563         "js         1b                      \n"
 564         :: "r"(((uint8_t *)blocks) + 128 * 6),
 565            "i"(-128 * 6)
 566         : "%"REG_a
 567     );
 568 }
 569
 570 static void add_bytes_mmx(uint8_t *dst, uint8_t *src, int w)
 571 {
 572     x86_reg i = 0;
 573     __asm__ volatile (
 574         "jmp          2f                \n\t"
 575         "1:                             \n\t"
 576         "movq   (%1, %0), %%mm0         \n\t"
 577         "movq   (%2, %0), %%mm1         \n\t"
 578         "paddb     %%mm0, %%mm1         \n\t"
 579         "movq      %%mm1, (%2, %0)      \n\t"
 580         "movq  8(%1, %0), %%mm0         \n\t"
 581         "movq  8(%2, %0), %%mm1         \n\t"
 582         "paddb     %%mm0, %%mm1         \n\t"
 583         "movq      %%mm1, 8(%2, %0)     \n\t"
 584         "add         $16, %0            \n\t"
 585         "2:                             \n\t"
 586         "cmp          %3, %0            \n\t"
 587         "js           1b                \n\t"
 588         : "+r"(i)
 589         : "r"(src), "r"(dst), "r"((x86_reg)w - 15)
 590     );
 591     for ( ; i < w; i++)
 592         dst[i + 0] += src[i + 0];
 593 }
 594
 595 #if HAVE_7REGS
 596 static void add_hfyu_median_prediction_cmov(uint8_t *dst, const uint8_t *top,
 597                                             const uint8_t *diff, int w,
 598                                             int *left, int *left_top)
 599 {
 600     x86_reg w2 = -w;
 601     x86_reg x;
 602     int l  = *left     & 0xff;
 603     int tl = *left_top & 0xff;
 604     int t;
 605     __asm__ volatile (
 606         "mov          %7, %3            \n"
 607         "1:                             \n"
 608         "movzbl (%3, %4), %2            \n"
 609         "mov          %2, %k3           \n"
 610         "sub         %b1, %b3           \n"
 611         "add         %b0, %b3           \n"
 612         "mov          %2, %1            \n"
 613         "cmp          %0, %2            \n"
 614         "cmovg        %0, %2            \n"
 615         "cmovg        %1, %0            \n"
 616         "cmp         %k3, %0            \n"
 617         "cmovg       %k3, %0            \n"
 618         "mov          %7, %3            \n"
 619         "cmp          %2, %0            \n"
 620         "cmovl        %2, %0            \n"
 621         "add    (%6, %4), %b0           \n"
 622         "mov         %b0, (%5, %4)      \n"
 623         "inc          %4                \n"
 624         "jl           1b                \n"
 625         : "+&q"(l), "+&q"(tl), "=&r"(t), "=&q"(x), "+&r"(w2)
 626         : "r"(dst + w), "r"(diff + w), "rm"(top + w)
 627     );
 628     *left     = l;
 629     *left_top = tl;
 630 }
 631 #endif
 632
 633 #define H263_LOOP_FILTER                        \
 634     "pxor      %%mm7, %%mm7             \n\t"   \
 635     "movq         %0, %%mm0             \n\t"   \
 636     "movq         %0, %%mm1             \n\t"   \
 637     "movq         %3, %%mm2             \n\t"   \
 638     "movq         %3, %%mm3             \n\t"   \
 639     "punpcklbw %%mm7, %%mm0             \n\t"   \
 640     "punpckhbw %%mm7, %%mm1             \n\t"   \
 641     "punpcklbw %%mm7, %%mm2             \n\t"   \
 642     "punpckhbw %%mm7, %%mm3             \n\t"   \
 643     "psubw     %%mm2, %%mm0             \n\t"   \
 644     "psubw     %%mm3, %%mm1             \n\t"   \
 645     "movq         %1, %%mm2             \n\t"   \
 646     "movq         %1, %%mm3             \n\t"   \
 647     "movq         %2, %%mm4             \n\t"   \
 648     "movq         %2, %%mm5             \n\t"   \
 649     "punpcklbw %%mm7, %%mm2             \n\t"   \
 650     "punpckhbw %%mm7, %%mm3             \n\t"   \
 651     "punpcklbw %%mm7, %%mm4             \n\t"   \
 652     "punpckhbw %%mm7, %%mm5             \n\t"   \
 653     "psubw     %%mm2, %%mm4             \n\t"   \
 654     "psubw     %%mm3, %%mm5             \n\t"   \
 655     "psllw        $2, %%mm4             \n\t"   \
 656     "psllw        $2, %%mm5             \n\t"   \
 657     "paddw     %%mm0, %%mm4             \n\t"   \
 658     "paddw     %%mm1, %%mm5             \n\t"   \
 659     "pxor      %%mm6, %%mm6             \n\t"   \
 660     "pcmpgtw   %%mm4, %%mm6             \n\t"   \
 661     "pcmpgtw   %%mm5, %%mm7             \n\t"   \
 662     "pxor      %%mm6, %%mm4             \n\t"   \
 663     "pxor      %%mm7, %%mm5             \n\t"   \
 664     "psubw     %%mm6, %%mm4             \n\t"   \
 665     "psubw     %%mm7, %%mm5             \n\t"   \
 666     "psrlw        $3, %%mm4             \n\t"   \
 667     "psrlw        $3, %%mm5             \n\t"   \
 668     "packuswb  %%mm5, %%mm4             \n\t"   \
 669     "packsswb  %%mm7, %%mm6             \n\t"   \
 670     "pxor      %%mm7, %%mm7             \n\t"   \
 671     "movd         %4, %%mm2             \n\t"   \
 672     "punpcklbw %%mm2, %%mm2             \n\t"   \
 673     "punpcklbw %%mm2, %%mm2             \n\t"   \
 674     "punpcklbw %%mm2, %%mm2             \n\t"   \
 675     "psubusb   %%mm4, %%mm2             \n\t"   \
 676     "movq      %%mm2, %%mm3             \n\t"   \
 677     "psubusb   %%mm4, %%mm3             \n\t"   \
 678     "psubb     %%mm3, %%mm2             \n\t"   \
 679     "movq         %1, %%mm3             \n\t"   \
 680     "movq         %2, %%mm4             \n\t"   \
 681     "pxor      %%mm6, %%mm3             \n\t"   \
 682     "pxor      %%mm6, %%mm4             \n\t"   \
 683     "paddusb   %%mm2, %%mm3             \n\t"   \
 684     "psubusb   %%mm2, %%mm4             \n\t"   \
 685     "pxor      %%mm6, %%mm3             \n\t"   \
 686     "pxor      %%mm6, %%mm4             \n\t"   \
 687     "paddusb   %%mm2, %%mm2             \n\t"   \
 688     "packsswb  %%mm1, %%mm0             \n\t"   \
 689     "pcmpgtb   %%mm0, %%mm7             \n\t"   \
 690     "pxor      %%mm7, %%mm0             \n\t"   \
 691     "psubb     %%mm7, %%mm0             \n\t"   \
 692     "movq      %%mm0, %%mm1             \n\t"   \
 693     "psubusb   %%mm2, %%mm0             \n\t"   \
 694     "psubb     %%mm0, %%mm1             \n\t"   \
 695     "pand         %5, %%mm1             \n\t"   \
 696     "psrlw        $2, %%mm1             \n\t"   \
 697     "pxor      %%mm7, %%mm1             \n\t"   \
 698     "psubb     %%mm7, %%mm1             \n\t"   \
 699     "movq         %0, %%mm5             \n\t"   \
 700     "movq         %3, %%mm6             \n\t"   \
 701     "psubb     %%mm1, %%mm5             \n\t"   \
 702     "paddb     %%mm1, %%mm6             \n\t"
 703
 704 static void h263_v_loop_filter_mmx(uint8_t *src, int stride, int qscale)
 705 {
 706     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
 707         const int strength = ff_h263_loop_filter_strength[qscale];
 708
 709         __asm__ volatile (
 710             H263_LOOP_FILTER
 711
 712             "movq %%mm3, %1             \n\t"
 713             "movq %%mm4, %2             \n\t"
 714             "movq %%mm5, %0             \n\t"
 715             "movq %%mm6, %3             \n\t"
 716             : "+m"(*(uint64_t*)(src - 2 * stride)),
 717               "+m"(*(uint64_t*)(src - 1 * stride)),
 718               "+m"(*(uint64_t*)(src + 0 * stride)),
 719               "+m"(*(uint64_t*)(src + 1 * stride))
 720             : "g"(2 * strength), "m"(ff_pb_FC)
 721             );
 722     }
 723 }
 724
 725 static void h263_h_loop_filter_mmx(uint8_t *src, int stride, int qscale)
 726 {
 727     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
 728         const int strength = ff_h263_loop_filter_strength[qscale];
 729         DECLARE_ALIGNED(8, uint64_t, temp)[4];
 730         uint8_t *btemp = (uint8_t*)temp;
 731
 732         src -= 2;
 733
 734         transpose4x4(btemp,     src,              8, stride);
 735         transpose4x4(btemp + 4, src + 4 * stride, 8, stride);
 736         __asm__ volatile (
 737             H263_LOOP_FILTER // 5 3 4 6
 738
 739             : "+m"(temp[0]),
 740               "+m"(temp[1]),
 741               "+m"(temp[2]),
 742               "+m"(temp[3])
 743             : "g"(2 * strength), "m"(ff_pb_FC)
 744             );
 745
 746         __asm__ volatile (
 747             "movq      %%mm5, %%mm1         \n\t"
 748             "movq      %%mm4, %%mm0         \n\t"
 749             "punpcklbw %%mm3, %%mm5         \n\t"
 750             "punpcklbw %%mm6, %%mm4         \n\t"
 751             "punpckhbw %%mm3, %%mm1         \n\t"
 752             "punpckhbw %%mm6, %%mm0         \n\t"
 753             "movq      %%mm5, %%mm3         \n\t"
 754             "movq      %%mm1, %%mm6         \n\t"
 755             "punpcklwd %%mm4, %%mm5         \n\t"
 756             "punpcklwd %%mm0, %%mm1         \n\t"
 757             "punpckhwd %%mm4, %%mm3         \n\t"
 758             "punpckhwd %%mm0, %%mm6         \n\t"
 759             "movd      %%mm5, (%0)          \n\t"
 760             "punpckhdq %%mm5, %%mm5         \n\t"
 761             "movd      %%mm5, (%0, %2)      \n\t"
 762             "movd      %%mm3, (%0, %2, 2)   \n\t"
 763             "punpckhdq %%mm3, %%mm3         \n\t"
 764             "movd      %%mm3, (%0, %3)      \n\t"
 765             "movd      %%mm1, (%1)          \n\t"
 766             "punpckhdq %%mm1, %%mm1         \n\t"
 767             "movd      %%mm1, (%1, %2)      \n\t"
 768             "movd      %%mm6, (%1, %2, 2)   \n\t"
 769             "punpckhdq %%mm6, %%mm6         \n\t"
 770             "movd      %%mm6, (%1, %3)      \n\t"
 771             :: "r"(src),
 772                "r"(src + 4 * stride),
 773                "r"((x86_reg)stride),
 774                "r"((x86_reg)(3 * stride))
 775             );
 776     }
 777 }
 778
 779 /* Draw the edges of width 'w' of an image of size width, height
 780  * this MMX version can only handle w == 8 || w == 16. */
 781 static void draw_edges_mmx(uint8_t *buf, int wrap, int width, int height,
 782                            int w, int h, int sides)
 783 {
 784     uint8_t *ptr, *last_line;
 785     int i;
 786
 787     last_line = buf + (height - 1) * wrap;
 788     /* left and right */
 789     ptr = buf;
 790     if (w == 8) {
 791         __asm__ volatile (
 792             "1:                             \n\t"
 793             "movd            (%0), %%mm0    \n\t"
 794             "punpcklbw      %%mm0, %%mm0    \n\t"
 795             "punpcklwd      %%mm0, %%mm0    \n\t"
 796             "punpckldq      %%mm0, %%mm0    \n\t"
 797             "movq           %%mm0, -8(%0)   \n\t"
 798             "movq      -8(%0, %2), %%mm1    \n\t"
 799             "punpckhbw      %%mm1, %%mm1    \n\t"
 800             "punpckhwd      %%mm1, %%mm1    \n\t"
 801             "punpckhdq      %%mm1, %%mm1    \n\t"
 802             "movq           %%mm1, (%0, %2) \n\t"
 803             "add               %1, %0       \n\t"
 804             "cmp               %3, %0       \n\t"
 805             "jb                1b           \n\t"
 806             : "+r"(ptr)
 807             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
 808             );
 809     } else {
 810         __asm__ volatile (
 811             "1:                                 \n\t"
 812             "movd            (%0), %%mm0        \n\t"
 813             "punpcklbw      %%mm0, %%mm0        \n\t"
 814             "punpcklwd      %%mm0, %%mm0        \n\t"
 815             "punpckldq      %%mm0, %%mm0        \n\t"
 816             "movq           %%mm0, -8(%0)       \n\t"
 817             "movq           %%mm0, -16(%0)      \n\t"
 818             "movq      -8(%0, %2), %%mm1        \n\t"
 819             "punpckhbw      %%mm1, %%mm1        \n\t"
 820             "punpckhwd      %%mm1, %%mm1        \n\t"
 821             "punpckhdq      %%mm1, %%mm1        \n\t"
 822             "movq           %%mm1,  (%0, %2)    \n\t"
 823             "movq           %%mm1, 8(%0, %2)    \n\t"
 824             "add               %1, %0           \n\t"
 825             "cmp               %3, %0           \n\t"
 826             "jb                1b               \n\t"
 827             : "+r"(ptr)
 828             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
 829             );
 830     }
 831
 832     /* top and bottom (and hopefully also the corners) */
 833     if (sides & EDGE_TOP) {
 834         for (i = 0; i < h; i += 4) {
 835             ptr = buf - (i + 1) * wrap - w;
 836             __asm__ volatile (
 837                 "1:                             \n\t"
 838                 "movq (%1, %0), %%mm0           \n\t"
 839                 "movq    %%mm0, (%0)            \n\t"
 840                 "movq    %%mm0, (%0, %2)        \n\t"
 841                 "movq    %%mm0, (%0, %2, 2)     \n\t"
 842                 "movq    %%mm0, (%0, %3)        \n\t"
 843                 "add        $8, %0              \n\t"
 844                 "cmp        %4, %0              \n\t"
 845                 "jb         1b                  \n\t"
 846                 : "+r"(ptr)
 847                 : "r"((x86_reg)buf - (x86_reg)ptr - w), "r"((x86_reg) -wrap),
 848                   "r"((x86_reg) -wrap * 3), "r"(ptr + width + 2 * w)
 849                 );
 850         }
 851     }
 852
 853     if (sides & EDGE_BOTTOM) {
 854         for (i = 0; i < h; i += 4) {
 855             ptr = last_line + (i + 1) * wrap - w;
 856             __asm__ volatile (
 857                 "1:                             \n\t"
 858                 "movq (%1, %0), %%mm0           \n\t"
 859                 "movq    %%mm0, (%0)            \n\t"
 860                 "movq    %%mm0, (%0, %2)        \n\t"
 861                 "movq    %%mm0, (%0, %2, 2)     \n\t"
 862                 "movq    %%mm0, (%0, %3)        \n\t"
 863                 "add        $8, %0              \n\t"
 864                 "cmp        %4, %0              \n\t"
 865                 "jb         1b                  \n\t"
 866                 : "+r"(ptr)
 867                 : "r"((x86_reg)last_line - (x86_reg)ptr - w),
 868                   "r"((x86_reg)wrap), "r"((x86_reg)wrap * 3),
 869                   "r"(ptr + width + 2 * w)
 870                 );
 871         }
 872     }
 873 }
 874
 875 #define QPEL_V_LOW(m3, m4, m5, m6, pw_20, pw_3, rnd,                      \
 876                    in0, in1, in2, in7, out, OP)                           \
 877     "paddw               "#m4", "#m3"   \n\t" /* x1 */                    \
 878     "movq   "MANGLE(ff_pw_20)", %%mm4   \n\t" /* 20 */                    \
 879     "pmullw              "#m3", %%mm4   \n\t" /* 20x1 */                  \
 880     "movq               "#in7", "#m3"   \n\t" /* d */                     \
 881     "movq               "#in0", %%mm5   \n\t" /* D */                     \
 882     "paddw               "#m3", %%mm5   \n\t" /* x4 */                    \
 883     "psubw               %%mm5, %%mm4   \n\t" /* 20x1 - x4 */             \
 884     "movq               "#in1", %%mm5   \n\t" /* C */                     \
 885     "movq               "#in2", %%mm6   \n\t" /* B */                     \
 886     "paddw               "#m6", %%mm5   \n\t" /* x3 */                    \
 887     "paddw               "#m5", %%mm6   \n\t" /* x2 */                    \
 888     "paddw               %%mm6, %%mm6   \n\t" /* 2x2 */                   \
 889     "psubw               %%mm6, %%mm5   \n\t" /* -2x2 + x3 */             \
 890     "pmullw  "MANGLE(ff_pw_3)", %%mm5   \n\t" /* -6x2 + 3x3 */            \
 891     "paddw              "#rnd", %%mm4   \n\t" /* x2 */                    \
 892     "paddw               %%mm4, %%mm5   \n\t" /* 20x1 - 6x2 + 3x3 - x4 */ \
 893     "psraw                  $5, %%mm5   \n\t"                             \
 894     "packuswb            %%mm5, %%mm5   \n\t"                             \
 895     OP(%%mm5, out, %%mm7, d)
 896
 897 #define QPEL_BASE(OPNAME, ROUNDER, RND, OP_MMX2, OP_3DNOW)                \
 898 static void OPNAME ## mpeg4_qpel16_h_lowpass_mmx2(uint8_t *dst,           \
 899                                                   uint8_t *src,           \
 900                                                   int dstStride,          \
 901                                                   int srcStride,          \
 902                                                   int h)                  \
 903 {                                                                         \
 904     uint64_t temp;                                                        \
 905                                                                           \
 906     __asm__ volatile (                                                    \
 907         "pxor      %%mm7, %%mm7             \n\t"                         \
 908         "1:                                 \n\t"                         \
 909         "movq       (%0), %%mm0             \n\t" /* ABCDEFGH */          \
 910         "movq      %%mm0, %%mm1             \n\t" /* ABCDEFGH */          \
 911         "movq      %%mm0, %%mm2             \n\t" /* ABCDEFGH */          \
 912         "punpcklbw %%mm7, %%mm0             \n\t" /* 0A0B0C0D */          \
 913         "punpckhbw %%mm7, %%mm1             \n\t" /* 0E0F0G0H */          \
 914         "pshufw    $0x90, %%mm0, %%mm5      \n\t" /* 0A0A0B0C */          \
 915         "pshufw    $0x41, %%mm0, %%mm6      \n\t" /* 0B0A0A0B */          \
 916         "movq      %%mm2, %%mm3             \n\t" /* ABCDEFGH */          \
 917         "movq      %%mm2, %%mm4             \n\t" /* ABCDEFGH */          \
 918         "psllq        $8, %%mm2             \n\t" /* 0ABCDEFG */          \
 919         "psllq       $16, %%mm3             \n\t" /* 00ABCDEF */          \
 920         "psllq       $24, %%mm4             \n\t" /* 000ABCDE */          \
 921         "punpckhbw %%mm7, %%mm2             \n\t" /* 0D0E0F0G */          \
 922         "punpckhbw %%mm7, %%mm3             \n\t" /* 0C0D0E0F */          \
 923         "punpckhbw %%mm7, %%mm4             \n\t" /* 0B0C0D0E */          \
 924         "paddw     %%mm3, %%mm5             \n\t" /* b */                 \
 925         "paddw     %%mm2, %%mm6             \n\t" /* c */                 \
 926         "paddw     %%mm5, %%mm5             \n\t" /* 2b */                \
 927         "psubw     %%mm5, %%mm6             \n\t" /* c - 2b */            \
 928         "pshufw    $0x06, %%mm0, %%mm5      \n\t" /* 0C0B0A0A */          \
 929         "pmullw "MANGLE(ff_pw_3)", %%mm6    \n\t" /* 3c - 6b */           \
 930         "paddw     %%mm4, %%mm0             \n\t" /* a */                 \
 931         "paddw     %%mm1, %%mm5             \n\t" /* d */                 \
 932         "pmullw "MANGLE(ff_pw_20)", %%mm0   \n\t" /* 20a */               \
 933         "psubw     %%mm5, %%mm0             \n\t" /* 20a - d */           \
 934         "paddw        %6, %%mm6             \n\t"                         \
 935         "paddw     %%mm6, %%mm0             \n\t" /* 20a - 6b + 3c - d */ \
 936         "psraw        $5, %%mm0             \n\t"                         \
 937         "movq      %%mm0, %5                \n\t"                         \
 938         /* mm1 = EFGH, mm2 = DEFG, mm3 = CDEF, mm4 = BCDE, mm7 = 0 */     \
 939                                                                           \
 940         "movq      5(%0), %%mm0             \n\t" /* FGHIJKLM */          \
 941         "movq      %%mm0, %%mm5             \n\t" /* FGHIJKLM */          \
 942         "movq      %%mm0, %%mm6             \n\t" /* FGHIJKLM */          \
 943         "psrlq        $8, %%mm0             \n\t" /* GHIJKLM0 */          \
 944         "psrlq       $16, %%mm5             \n\t" /* HIJKLM00 */          \
 945         "punpcklbw %%mm7, %%mm0             \n\t" /* 0G0H0I0J */          \
 946         "punpcklbw %%mm7, %%mm5             \n\t" /* 0H0I0J0K */          \
 947         "paddw     %%mm0, %%mm2             \n\t" /* b */                 \
 948         "paddw     %%mm5, %%mm3             \n\t" /* c */                 \
 949         "paddw     %%mm2, %%mm2             \n\t" /* 2b */                \
 950         "psubw     %%mm2, %%mm3             \n\t" /* c - 2b */            \
 951         "movq      %%mm6, %%mm2             \n\t" /* FGHIJKLM */          \
 952         "psrlq       $24, %%mm6             \n\t" /* IJKLM000 */          \
 953         "punpcklbw %%mm7, %%mm2             \n\t" /* 0F0G0H0I */          \
 954         "punpcklbw %%mm7, %%mm6             \n\t" /* 0I0J0K0L */          \
 955         "pmullw "MANGLE(ff_pw_3)", %%mm3    \n\t" /* 3c - 6b */           \
 956         "paddw     %%mm2, %%mm1             \n\t" /* a */                 \
 957         "paddw     %%mm6, %%mm4             \n\t" /* d */                 \
 958         "pmullw "MANGLE(ff_pw_20)", %%mm1   \n\t" /* 20a */               \
 959         "psubw     %%mm4, %%mm3             \n\t" /* - 6b +3c - d */      \
 960         "paddw        %6, %%mm1             \n\t"                         \
 961         "paddw     %%mm1, %%mm3             \n\t" /* 20a - 6b +3c - d */  \
 962         "psraw        $5, %%mm3             \n\t"                         \
 963         "movq         %5, %%mm1             \n\t"                         \
 964         "packuswb  %%mm3, %%mm1             \n\t"                         \
 965         OP_MMX2(%%mm1, (%1), %%mm4, q)                                    \
 966         /* mm0 = GHIJ, mm2 = FGHI, mm5 = HIJK, mm6 = IJKL, mm7 = 0 */     \
 967                                                                           \
 968         "movq      9(%0), %%mm1             \n\t" /* JKLMNOPQ */          \
 969         "movq      %%mm1, %%mm4             \n\t" /* JKLMNOPQ */          \
 970         "movq      %%mm1, %%mm3             \n\t" /* JKLMNOPQ */          \
 971         "psrlq        $8, %%mm1             \n\t" /* KLMNOPQ0 */          \
 972         "psrlq       $16, %%mm4             \n\t" /* LMNOPQ00 */          \
 973         "punpcklbw %%mm7, %%mm1             \n\t" /* 0K0L0M0N */          \
 974         "punpcklbw %%mm7, %%mm4             \n\t" /* 0L0M0N0O */          \
 975         "paddw     %%mm1, %%mm5             \n\t" /* b */                 \
 976         "paddw     %%mm4, %%mm0             \n\t" /* c */                 \
 977         "paddw     %%mm5, %%mm5             \n\t" /* 2b */                \
 978         "psubw     %%mm5, %%mm0             \n\t" /* c - 2b */            \
 979         "movq      %%mm3, %%mm5             \n\t" /* JKLMNOPQ */          \
 980         "psrlq       $24, %%mm3             \n\t" /* MNOPQ000 */          \
 981         "pmullw "MANGLE(ff_pw_3)", %%mm0    \n\t" /* 3c - 6b */           \
 982         "punpcklbw %%mm7, %%mm3             \n\t" /* 0M0N0O0P */          \
 983         "paddw     %%mm3, %%mm2             \n\t" /* d */                 \
 984         "psubw     %%mm2, %%mm0             \n\t" /* -6b + 3c - d */      \
 985         "movq      %%mm5, %%mm2             \n\t" /* JKLMNOPQ */          \
 986         "punpcklbw %%mm7, %%mm2             \n\t" /* 0J0K0L0M */          \
 987         "punpckhbw %%mm7, %%mm5             \n\t" /* 0N0O0P0Q */          \
 988         "paddw     %%mm2, %%mm6             \n\t" /* a */                 \
 989         "pmullw "MANGLE(ff_pw_20)", %%mm6   \n\t" /* 20a */               \
 990         "paddw        %6, %%mm0             \n\t"                         \
 991         "paddw     %%mm6, %%mm0             \n\t" /* 20a - 6b + 3c - d */ \
 992         "psraw        $5, %%mm0             \n\t"                         \
 993         /* mm1 = KLMN, mm2 = JKLM, mm3 = MNOP, */                         \
 994         /* mm4 = LMNO, mm5 = NOPQ mm7 = 0 */                              \
 995                                                                           \
 996         "paddw    %%mm5, %%mm3              \n\t" /* a */                 \
 997         "pshufw   $0xF9, %%mm5, %%mm6       \n\t" /* 0O0P0Q0Q */          \
 998         "paddw    %%mm4, %%mm6              \n\t" /* b */                 \
 999         "pshufw   $0xBE, %%mm5, %%mm4       \n\t" /* 0P0Q0Q0P */          \
1000         "pshufw   $0x6F, %%mm5, %%mm5       \n\t" /* 0Q0Q0P0O */          \
1001         "paddw    %%mm1, %%mm4              \n\t" /* c */                 \
1002         "paddw    %%mm2, %%mm5              \n\t" /* d */                 \
1003         "paddw    %%mm6, %%mm6              \n\t" /* 2b */                \
1004         "psubw    %%mm6, %%mm4              \n\t" /* c - 2b */            \
1005         "pmullw "MANGLE(ff_pw_20)", %%mm3   \n\t" /* 20a */               \
1006         "pmullw  "MANGLE(ff_pw_3)", %%mm4   \n\t" /* 3c - 6b */           \
1007         "psubw    %%mm5, %%mm3              \n\t" /* -6b + 3c - d */      \
1008         "paddw       %6, %%mm4              \n\t"                         \
1009         "paddw    %%mm3, %%mm4              \n\t" /* 20a - 6b + 3c - d */ \
1010         "psraw       $5, %%mm4              \n\t"                         \
1011         "packuswb %%mm4, %%mm0              \n\t"                         \
1012         OP_MMX2(%%mm0, 8(%1), %%mm4, q)                                   \
1013                                                                           \
1014         "add         %3, %0                 \n\t"                         \
1015         "add         %4, %1                 \n\t"                         \
1016         "decl        %2                     \n\t"                         \
1017         "jnz         1b                     \n\t"                         \
1018         : "+a"(src), "+c"(dst), "+D"(h)                                   \
1019         : "d"((x86_reg)srcStride), "S"((x86_reg)dstStride),               \
1020           /* "m"(ff_pw_20), "m"(ff_pw_3), */ "m"(temp), "m"(ROUNDER)      \
1021         : "memory"                                                        \
1022         );                                                                \
1023 }                                                                         \
1024                                                                           \
1025 static void OPNAME ## mpeg4_qpel16_h_lowpass_3dnow(uint8_t *dst,          \
1026                                                    uint8_t *src,          \
1027                                                    int dstStride,         \
1028                                                    int srcStride,         \
1029                                                    int h)                 \
1030 {                                                                         \
1031     int i;                                                                \
1032     int16_t temp[16];                                                     \
1033     /* quick HACK, XXX FIXME MUST be optimized */                         \
1034     for (i = 0; i < h; i++) {                                             \
1035         temp[ 0] = (src[ 0] + src[ 1]) * 20 - (src[ 0] + src[ 2]) * 6 +   \
1036                    (src[ 1] + src[ 3]) *  3 - (src[ 2] + src[ 4]);        \
1037         temp[ 1] = (src[ 1] + src[ 2]) * 20 - (src[ 0] + src[ 3]) * 6 +   \
1038                    (src[ 0] + src[ 4]) *  3 - (src[ 1] + src[ 5]);        \
1039         temp[ 2] = (src[ 2] + src[ 3]) * 20 - (src[ 1] + src[ 4]) * 6 +   \
1040                    (src[ 0] + src[ 5]) *  3 - (src[ 0] + src[ 6]);        \
1041         temp[ 3] = (src[ 3] + src[ 4]) * 20 - (src[ 2] + src[ 5]) * 6 +   \
1042                    (src[ 1] + src[ 6]) *  3 - (src[ 0] + src[ 7]);        \
1043         temp[ 4] = (src[ 4] + src[ 5]) * 20 - (src[ 3] + src[ 6]) * 6 +   \
1044                    (src[ 2] + src[ 7]) *  3 - (src[ 1] + src[ 8]);        \
1045         temp[ 5] = (src[ 5] + src[ 6]) * 20 - (src[ 4] + src[ 7]) * 6 +   \
1046                    (src[ 3] + src[ 8]) *  3 - (src[ 2] + src[ 9]);        \
1047         temp[ 6] = (src[ 6] + src[ 7]) * 20 - (src[ 5] + src[ 8]) * 6 +   \
1048                    (src[ 4] + src[ 9]) *  3 - (src[ 3] + src[10]);        \
1049         temp[ 7] = (src[ 7] + src[ 8]) * 20 - (src[ 6] + src[ 9]) * 6 +   \
1050                    (src[ 5] + src[10]) *  3 - (src[ 4] + src[11]);        \
1051         temp[ 8] = (src[ 8] + src[ 9]) * 20 - (src[ 7] + src[10]) * 6 +   \
1052                    (src[ 6] + src[11]) *  3 - (src[ 5] + src[12]);        \
1053         temp[ 9] = (src[ 9] + src[10]) * 20 - (src[ 8] + src[11]) * 6 +   \
1054                    (src[ 7] + src[12]) *  3 - (src[ 6] + src[13]);        \
1055         temp[10] = (src[10] + src[11]) * 20 - (src[ 9] + src[12]) * 6 +   \
1056                    (src[ 8] + src[13]) *  3 - (src[ 7] + src[14]);        \
1057         temp[11] = (src[11] + src[12]) * 20 - (src[10] + src[13]) * 6 +   \
1058                    (src[ 9] + src[14]) *  3 - (src[ 8] + src[15]);        \
1059         temp[12] = (src[12] + src[13]) * 20 - (src[11] + src[14]) * 6 +   \
1060                    (src[10] + src[15]) *  3 - (src[ 9] + src[16]);        \
1061         temp[13] = (src[13] + src[14]) * 20 - (src[12] + src[15]) * 6 +   \
1062                    (src[11] + src[16]) *  3 - (src[10] + src[16]);        \
1063         temp[14] = (src[14] + src[15]) * 20 - (src[13] + src[16]) * 6 +   \
1064                    (src[12] + src[16]) *  3 - (src[11] + src[15]);        \
1065         temp[15] = (src[15] + src[16]) * 20 - (src[14] + src[16]) * 6 +   \
1066                    (src[13] + src[15]) *  3 - (src[12] + src[14]);        \
1067         __asm__ volatile (                                                \
1068             "movq      (%0), %%mm0          \n\t"                         \
1069             "movq     8(%0), %%mm1          \n\t"                         \
1070             "paddw       %2, %%mm0          \n\t"                         \
1071             "paddw       %2, %%mm1          \n\t"                         \
1072             "psraw       $5, %%mm0          \n\t"                         \
1073             "psraw       $5, %%mm1          \n\t"                         \
1074             "packuswb %%mm1, %%mm0          \n\t"                         \
1075             OP_3DNOW(%%mm0, (%1), %%mm1, q)                               \
1076             "movq    16(%0), %%mm0          \n\t"                         \
1077             "movq    24(%0), %%mm1          \n\t"                         \
1078             "paddw       %2, %%mm0          \n\t"                         \
1079             "paddw       %2, %%mm1          \n\t"                         \
1080             "psraw       $5, %%mm0          \n\t"                         \
1081             "psraw       $5, %%mm1          \n\t"                         \
1082             "packuswb %%mm1, %%mm0          \n\t"                         \
1083             OP_3DNOW(%%mm0, 8(%1), %%mm1, q)                              \
1084             :: "r"(temp), "r"(dst), "m"(ROUNDER)                          \
1085             : "memory"                                                    \
1086             );                                                            \
1087         dst += dstStride;                                                 \
1088         src += srcStride;                                                 \
1089     }                                                                     \
1090 }                                                                         \
1091                                                                           \
1092 static void OPNAME ## mpeg4_qpel8_h_lowpass_mmx2(uint8_t *dst,            \
1093                                                  uint8_t *src,            \
1094                                                  int dstStride,           \
1095                                                  int srcStride,           \
1096                                                  int h)                   \
1097 {                                                                         \
1098     __asm__ volatile (                                                    \
1099         "pxor      %%mm7, %%mm7             \n\t"                         \
1100         "1:                                 \n\t"                         \
1101         "movq       (%0), %%mm0             \n\t" /* ABCDEFGH */          \
1102         "movq      %%mm0, %%mm1             \n\t" /* ABCDEFGH */          \
1103         "movq      %%mm0, %%mm2             \n\t" /* ABCDEFGH */          \
1104         "punpcklbw %%mm7, %%mm0             \n\t" /* 0A0B0C0D */          \
1105         "punpckhbw %%mm7, %%mm1             \n\t" /* 0E0F0G0H */          \
1106         "pshufw    $0x90, %%mm0, %%mm5      \n\t" /* 0A0A0B0C */          \
1107         "pshufw    $0x41, %%mm0, %%mm6      \n\t" /* 0B0A0A0B */          \
1108         "movq      %%mm2, %%mm3             \n\t" /* ABCDEFGH */          \
1109         "movq      %%mm2, %%mm4             \n\t" /* ABCDEFGH */          \
1110         "psllq        $8, %%mm2             \n\t" /* 0ABCDEFG */          \
1111         "psllq       $16, %%mm3             \n\t" /* 00ABCDEF */          \
1112         "psllq       $24, %%mm4             \n\t" /* 000ABCDE */          \
1113         "punpckhbw %%mm7, %%mm2             \n\t" /* 0D0E0F0G */          \
1114         "punpckhbw %%mm7, %%mm3             \n\t" /* 0C0D0E0F */          \
1115         "punpckhbw %%mm7, %%mm4             \n\t" /* 0B0C0D0E */          \
1116         "paddw     %%mm3, %%mm5             \n\t" /* b */                 \
1117         "paddw     %%mm2, %%mm6             \n\t" /* c */                 \
1118         "paddw     %%mm5, %%mm5             \n\t" /* 2b */                \
1119         "psubw     %%mm5, %%mm6             \n\t" /* c - 2b */            \
1120         "pshufw    $0x06, %%mm0, %%mm5      \n\t" /* 0C0B0A0A */          \
1121         "pmullw "MANGLE(ff_pw_3)", %%mm6    \n\t" /* 3c - 6b */           \
1122         "paddw     %%mm4, %%mm0             \n\t" /* a */                 \
1123         "paddw     %%mm1, %%mm5             \n\t" /* d */                 \
1124         "pmullw "MANGLE(ff_pw_20)", %%mm0   \n\t" /* 20a */               \
1125         "psubw     %%mm5, %%mm0             \n\t" /* 20a - d */           \
1126         "paddw        %5, %%mm6             \n\t"                         \
1127         "paddw     %%mm6, %%mm0             \n\t" /* 20a - 6b + 3c - d */ \
1128         "psraw        $5, %%mm0             \n\t"                         \
1129         /* mm1 = EFGH, mm2 = DEFG, mm3 = CDEF, mm4 = BCDE, mm7 = 0 */     \
1130                                                                           \
1131         "movd      5(%0), %%mm5             \n\t" /* FGHI */              \
1132         "punpcklbw %%mm7, %%mm5             \n\t" /* 0F0G0H0I */          \
1133         "pshufw    $0xF9, %%mm5, %%mm6      \n\t" /* 0G0H0I0I */          \
1134         "paddw     %%mm5, %%mm1             \n\t" /* a */                 \
1135         "paddw     %%mm6, %%mm2             \n\t" /* b */                 \
1136         "pshufw    $0xBE, %%mm5, %%mm6      \n\t" /* 0H0I0I0H */          \
1137         "pshufw    $0x6F, %%mm5, %%mm5      \n\t" /* 0I0I0H0G */          \
1138         "paddw     %%mm6, %%mm3             \n\t" /* c */                 \
1139         "paddw     %%mm5, %%mm4             \n\t" /* d */                 \
1140         "paddw     %%mm2, %%mm2             \n\t" /* 2b */                \
1141         "psubw     %%mm2, %%mm3             \n\t" /* c - 2b */            \
1142         "pmullw "MANGLE(ff_pw_20)", %%mm1   \n\t" /* 20a */               \
1143         "pmullw  "MANGLE(ff_pw_3)", %%mm3   \n\t" /* 3c - 6b */           \
1144         "psubw     %%mm4, %%mm3             \n\t" /* -6b + 3c - d */      \
1145         "paddw        %5, %%mm1             \n\t"                         \
1146         "paddw     %%mm1, %%mm3             \n\t" /* 20a - 6b + 3c - d */ \
1147         "psraw        $5, %%mm3             \n\t"                         \
1148         "packuswb  %%mm3, %%mm0             \n\t"                         \
1149         OP_MMX2(%%mm0, (%1), %%mm4, q)                                    \
1150                                                                           \
1151         "add          %3, %0                \n\t"                         \
1152         "add          %4, %1                \n\t"                         \
1153         "decl         %2                    \n\t"                         \
1154         "jnz          1b                    \n\t"                         \
1155         : "+a"(src), "+c"(dst), "+d"(h)                                   \
1156         : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride),               \
1157           /* "m"(ff_pw_20), "m"(ff_pw_3), */ "m"(ROUNDER)                 \
1158         : "memory"                                                        \
1159         );                                                                \
1160 }                                                                         \
1161                                                                           \
1162 static void OPNAME ## mpeg4_qpel8_h_lowpass_3dnow(uint8_t *dst,           \
1163                                                   uint8_t *src,           \
1164                                                   int dstStride,          \
1165                                                   int srcStride,          \
1166                                                   int h)                  \
1167 {                                                                         \
1168     int i;                                                                \
1169     int16_t temp[8];                                                      \
1170     /* quick HACK, XXX FIXME MUST be optimized */                         \
1171     for (i = 0; i < h; i++) {                                             \
1172         temp[0] = (src[0] + src[1]) * 20 - (src[0] + src[2]) * 6 +        \
1173                   (src[1] + src[3]) *  3 - (src[2] + src[4]);             \
1174         temp[1] = (src[1] + src[2]) * 20 - (src[0] + src[3]) * 6 +        \
1175                   (src[0] + src[4]) *  3 - (src[1] + src[5]);             \
1176         temp[2] = (src[2] + src[3]) * 20 - (src[1] + src[4]) * 6 +        \
1177                   (src[0] + src[5]) *  3 - (src[0] + src[6]);             \
1178         temp[3] = (src[3] + src[4]) * 20 - (src[2] + src[5]) * 6 +        \
1179                   (src[1] + src[6]) *  3 - (src[0] + src[7]);             \
1180         temp[4] = (src[4] + src[5]) * 20 - (src[3] + src[6]) * 6 +        \
1181                   (src[2] + src[7]) *  3 - (src[1] + src[8]);             \
1182         temp[5] = (src[5] + src[6]) * 20 - (src[4] + src[7]) * 6 +        \
1183                   (src[3] + src[8]) *  3 - (src[2] + src[8]);             \
1184         temp[6] = (src[6] + src[7]) * 20 - (src[5] + src[8]) * 6 +        \
1185                   (src[4] + src[8]) *  3 - (src[3] + src[7]);             \
1186         temp[7] = (src[7] + src[8]) * 20 - (src[6] + src[8]) * 6 +        \
1187                   (src[5] + src[7]) *  3 - (src[4] + src[6]);             \
1188         __asm__ volatile (                                                \
1189             "movq      (%0), %%mm0      \n\t"                             \
1190             "movq     8(%0), %%mm1      \n\t"                             \
1191             "paddw       %2, %%mm0      \n\t"                             \
1192             "paddw       %2, %%mm1      \n\t"                             \
1193             "psraw       $5, %%mm0      \n\t"                             \
1194             "psraw       $5, %%mm1      \n\t"                             \
1195             "packuswb %%mm1, %%mm0      \n\t"                             \
1196             OP_3DNOW(%%mm0, (%1), %%mm1, q)                               \
1197             :: "r"(temp), "r"(dst), "m"(ROUNDER)                          \
1198             : "memory"                                                    \
1199             );                                                            \
1200         dst += dstStride;                                                 \
1201         src += srcStride;                                                 \
1202     }                                                                     \
1203 }
1204
1205 #define QPEL_OP(OPNAME, ROUNDER, RND, OP, MMX)                          \
1206 static void OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(uint8_t *dst,      \
1207                                                      uint8_t *src,      \
1208                                                      int dstStride,     \
1209                                                      int srcStride)     \
1210 {                                                                       \
1211     uint64_t temp[17 * 4];                                              \
1212     uint64_t *temp_ptr = temp;                                          \
1213     int count = 17;                                                     \
1214                                                                         \
1215     /* FIXME unroll */                                                  \
1216     __asm__ volatile (                                                  \
1217         "pxor      %%mm7, %%mm7             \n\t"                       \
1218         "1:                                 \n\t"                       \
1219         "movq       (%0), %%mm0             \n\t"                       \
1220         "movq       (%0), %%mm1             \n\t"                       \
1221         "movq      8(%0), %%mm2             \n\t"                       \
1222         "movq      8(%0), %%mm3             \n\t"                       \
1223         "punpcklbw %%mm7, %%mm0             \n\t"                       \
1224         "punpckhbw %%mm7, %%mm1             \n\t"                       \
1225         "punpcklbw %%mm7, %%mm2             \n\t"                       \
1226         "punpckhbw %%mm7, %%mm3             \n\t"                       \
1227         "movq      %%mm0, (%1)              \n\t"                       \
1228         "movq      %%mm1, 17 * 8(%1)        \n\t"                       \
1229         "movq      %%mm2, 2 * 17 * 8(%1)    \n\t"                       \
1230         "movq      %%mm3, 3 * 17 * 8(%1)    \n\t"                       \
1231         "add          $8, %1                \n\t"                       \
1232         "add          %3, %0                \n\t"                       \
1233         "decl         %2                    \n\t"                       \
1234         "jnz          1b                    \n\t"                       \
1235         : "+r"(src), "+r"(temp_ptr), "+r"(count)                        \
1236         : "r"((x86_reg)srcStride)                                       \
1237         : "memory"                                                      \
1238         );                                                              \
1239                                                                         \
1240     temp_ptr = temp;                                                    \
1241     count    = 4;                                                       \
1242                                                                         \
1243     /* FIXME reorder for speed */                                       \
1244     __asm__ volatile (                                                  \
1245         /* "pxor  %%mm7, %%mm7            \n\t" */                      \
1246         "1:                             \n\t"                           \
1247         "movq    (%0), %%mm0            \n\t"                           \
1248         "movq   8(%0), %%mm1            \n\t"                           \
1249         "movq  16(%0), %%mm2            \n\t"                           \
1250         "movq  24(%0), %%mm3            \n\t"                           \
1251         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0),   8(%0),    (%0),  32(%0), (%1),     OP) \
1252         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5,  8(%0),    (%0),    (%0),  40(%0), (%1, %3), OP) \
1253         "add       %4, %1               \n\t"                           \
1254         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5,   (%0),    (%0),   8(%0),  48(%0), (%1),     OP) \
1255                                                                         \
1256         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5,   (%0),   8(%0),  16(%0),  56(%0), (%1, %3), OP) \
1257         "add       %4, %1               \n\t"                           \
1258         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5,  8(%0),  16(%0),  24(%0),  64(%0), (%1),     OP) \
1259         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0),  24(%0),  32(%0),  72(%0), (%1, %3), OP) \
1260         "add       %4, %1               \n\t"                           \
1261         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0),  32(%0),  40(%0),  80(%0), (%1),     OP) \
1262         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0),  40(%0),  48(%0),  88(%0), (%1, %3), OP) \
1263         "add       %4, %1               \n\t"                           \
1264         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 40(%0),  48(%0),  56(%0),  96(%0), (%1),     OP) \
1265         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 48(%0),  56(%0),  64(%0), 104(%0), (%1, %3), OP) \
1266         "add       %4, %1               \n\t"                           \
1267         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 56(%0),  64(%0),  72(%0), 112(%0), (%1),     OP) \
1268         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 64(%0),  72(%0),  80(%0), 120(%0), (%1, %3), OP) \
1269         "add       %4, %1               \n\t"                           \
1270         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 72(%0),  80(%0),  88(%0), 128(%0), (%1),     OP) \
1271                                                                         \
1272         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 80(%0),  88(%0),  96(%0), 128(%0), (%1, %3), OP) \
1273         "add       %4, %1               \n\t"                           \
1274         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 88(%0),  96(%0), 104(%0), 120(%0), (%1),     OP) \
1275         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 96(%0), 104(%0), 112(%0), 112(%0), (%1, %3), OP) \
1276                                                                         \
1277         "add     $136, %0               \n\t"                           \
1278         "add       %6, %1               \n\t"                           \
1279         "decl      %2                   \n\t"                           \
1280         "jnz       1b                   \n\t"                           \
1281                                                                         \
1282         : "+r"(temp_ptr), "+r"(dst), "+g"(count)                        \
1283         : "r"((x86_reg)dstStride), "r"(2 * (x86_reg)dstStride),         \
1284           /* "m"(ff_pw_20), "m"(ff_pw_3), */ "m"(ROUNDER),              \
1285           "g"(4 - 14 * (x86_reg)dstStride)                              \
1286         : "memory"                                                      \
1287         );                                                              \
1288 }                                                                       \
1289                                                                         \
1290 static void OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(uint8_t *dst,       \
1291                                                     uint8_t *src,       \
1292                                                     int dstStride,      \
1293                                                     int srcStride)      \
1294 {                                                                       \
1295     uint64_t temp[9 * 2];                                               \
1296     uint64_t *temp_ptr = temp;                                          \
1297     int count = 9;                                                      \
1298                                                                         \
1299     /* FIXME unroll */                                                  \
1300     __asm__ volatile (                                                  \
1301         "pxor      %%mm7, %%mm7         \n\t"                           \
1302         "1:                             \n\t"                           \
1303         "movq       (%0), %%mm0         \n\t"                           \
1304         "movq       (%0), %%mm1         \n\t"                           \
1305         "punpcklbw %%mm7, %%mm0         \n\t"                           \
1306         "punpckhbw %%mm7, %%mm1         \n\t"                           \
1307         "movq      %%mm0, (%1)          \n\t"                           \
1308         "movq      %%mm1, 9*8(%1)       \n\t"                           \
1309         "add          $8, %1            \n\t"                           \
1310         "add          %3, %0            \n\t"                           \
1311         "decl         %2                \n\t"                           \
1312         "jnz          1b                \n\t"                           \
1313         : "+r"(src), "+r"(temp_ptr), "+r"(count)                        \
1314         : "r"((x86_reg)srcStride)                                       \
1315         : "memory"                                                      \
1316         );                                                              \
1317                                                                         \
1318     temp_ptr = temp;                                                    \
1319     count    = 2;                                                       \
1320                                                                         \
1321     /* FIXME reorder for speed */                                       \
1322     __asm__ volatile (                                                  \
1323         /* "pxor  %%mm7, %%mm7            \n\t" */                      \
1324         "1:                             \n\t"                           \
1325         "movq    (%0), %%mm0            \n\t"                           \
1326         "movq   8(%0), %%mm1            \n\t"                           \
1327         "movq  16(%0), %%mm2            \n\t"                           \
1328         "movq  24(%0), %%mm3            \n\t"                           \
1329         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0),  8(%0),   (%0), 32(%0), (%1), OP)     \
1330         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5,  8(%0),   (%0),   (%0), 40(%0), (%1, %3), OP) \
1331         "add       %4, %1               \n\t"                           \
1332         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5,   (%0),   (%0),  8(%0), 48(%0), (%1), OP)     \
1333                                                                         \
1334         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5,   (%0),  8(%0), 16(%0), 56(%0), (%1, %3), OP) \
1335         "add       %4, %1               \n\t"                           \
1336         QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5,  8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)     \
1337                                                                         \
1338         QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 64(%0), (%1, %3), OP) \
1339         "add       %4, %1               \n\t"                           \
1340         QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 56(%0), (%1), OP)     \
1341         QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 48(%0), (%1, %3), OP) \
1342                                                                         \
1343         "add      $72, %0               \n\t"                           \
1344         "add       %6, %1               \n\t"                           \
1345         "decl      %2                   \n\t"                           \
1346         "jnz       1b                   \n\t"                           \
1347                                                                         \
1348         : "+r"(temp_ptr), "+r"(dst), "+g"(count)                        \
1349         : "r"((x86_reg)dstStride), "r"(2 * (x86_reg)dstStride),         \
1350           /* "m"(ff_pw_20), "m"(ff_pw_3), */ "m"(ROUNDER),              \
1351           "g"(4 - 6 * (x86_reg)dstStride)                               \
1352         : "memory"                                                      \
1353         );                                                              \
1354 }                                                                       \
1355                                                                         \
1356 static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src,   \
1357                                           int stride)                   \
1358 {                                                                       \
1359     OPNAME ## pixels8_ ## MMX(dst, src, stride, 8);                     \
1360 }                                                                       \
1361                                                                         \
1362 static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src,    \
1363                                          int stride)                    \
1364 {                                                                       \
1365     uint64_t temp[8];                                                   \
1366     uint8_t * const half = (uint8_t*)temp;                              \
1367     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8,           \
1368                                                 stride, 8);             \
1369     OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);    \
1370 }                                                                       \
1371                                                                         \
1372 static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src,    \
1373                                          int stride)                    \
1374 {                                                                       \
1375     OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride,           \
1376                                             stride, 8);                 \
1377 }                                                                       \
1378                                                                         \
1379 static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src,    \
1380                                          int stride)                    \
1381 {                                                                       \
1382     uint64_t temp[8];                                                   \
1383     uint8_t * const half = (uint8_t*)temp;                              \
1384     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8,           \
1385                                                 stride, 8);             \
1386     OPNAME ## pixels8_l2_ ## MMX(dst, src + 1, half, stride,            \
1387                                  stride, 8);                            \
1388 }                                                                       \
1389                                                                         \
1390 static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src,    \
1391                                          int stride)                    \
1392 {                                                                       \
1393     uint64_t temp[8];                                                   \
1394     uint8_t * const half = (uint8_t*)temp;                              \
1395     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);  \
1396     OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);    \
1397 }                                                                       \
1398                                                                         \
1399 static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src,    \
1400                                          int stride)                    \
1401 {                                                                       \
1402     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src, stride, stride);  \
1403 }                                                                       \
1404                                                                         \
1405 static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src,    \
1406                                          int stride)                    \
1407 {                                                                       \
1408     uint64_t temp[8];                                                   \
1409     uint8_t * const half = (uint8_t*)temp;                              \
1410     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);  \
1411     OPNAME ## pixels8_l2_ ## MMX(dst, src + stride, half, stride,       \
1412                                  stride, 8);                            \
1413 }                                                                       \
1414                                                                         \
1415 static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src,    \
1416                                          int stride)                    \
1417 {                                                                       \
1418     uint64_t half[8 + 9];                                               \
1419     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1420     uint8_t * const halfHV = ((uint8_t*)half);                          \
1421     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1422                                                 stride, 9);             \
1423     put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);  \
1424     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1425     OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);     \
1426 }                                                                       \
1427                                                                         \
1428 static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src,    \
1429                                          int stride)                    \
1430 {                                                                       \
1431     uint64_t half[8 + 9];                                               \
1432     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1433     uint8_t * const halfHV = ((uint8_t*)half);                          \
1434     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1435                                                 stride, 9);             \
1436     put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,          \
1437                                      stride, 9);                        \
1438     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1439     OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);     \
1440 }                                                                       \
1441                                                                         \
1442 static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src,    \
1443                                          int stride)                    \
1444 {                                                                       \
1445     uint64_t half[8 + 9];                                               \
1446     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1447     uint8_t * const halfHV = ((uint8_t*)half);                          \
1448     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1449                                                 stride, 9);             \
1450     put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);  \
1451     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1452     OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV, stride, 8, 8); \
1453 }                                                                       \
1454                                                                         \
1455 static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src,    \
1456                                          int stride)                    \
1457 {                                                                       \
1458     uint64_t half[8 + 9];                                               \
1459     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1460     uint8_t * const halfHV = ((uint8_t*)half);                          \
1461     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1462                                                 stride, 9);             \
1463     put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,          \
1464                                      stride, 9);                        \
1465     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1466     OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV, stride, 8, 8); \
1467 }                                                                       \
1468                                                                         \
1469 static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src,    \
1470                                          int stride)                    \
1471 {                                                                       \
1472     uint64_t half[8 + 9];                                               \
1473     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1474     uint8_t * const halfHV = ((uint8_t*)half);                          \
1475     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1476                                                 stride, 9);             \
1477     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1478     OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);     \
1479 }                                                                       \
1480                                                                         \
1481 static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src,    \
1482                                          int stride)                    \
1483 {                                                                       \
1484     uint64_t half[8 + 9];                                               \
1485     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1486     uint8_t * const halfHV = ((uint8_t*)half);                          \
1487     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1488                                                 stride, 9);             \
1489     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);   \
1490     OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV, stride, 8, 8); \
1491 }                                                                       \
1492                                                                         \
1493 static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src,    \
1494                                          int stride)                    \
1495 {                                                                       \
1496     uint64_t half[8 + 9];                                               \
1497     uint8_t * const halfH = ((uint8_t*)half);                           \
1498     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1499                                                 stride, 9);             \
1500     put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);  \
1501     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);     \
1502 }                                                                       \
1503                                                                         \
1504 static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src,    \
1505                                          int stride)                    \
1506 {                                                                       \
1507     uint64_t half[8 + 9];                                               \
1508     uint8_t * const halfH = ((uint8_t*)half);                           \
1509     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1510                                                 stride, 9);             \
1511     put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,          \
1512                                      stride, 9);                        \
1513     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);     \
1514 }                                                                       \
1515                                                                         \
1516 static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src,    \
1517                                          int stride)                    \
1518 {                                                                       \
1519     uint64_t half[9];                                                   \
1520     uint8_t * const halfH = ((uint8_t*)half);                           \
1521     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,          \
1522                                                 stride, 9);             \
1523     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);     \
1524 }                                                                       \
1525                                                                         \
1526 static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src,  \
1527                                            int stride)                  \
1528 {                                                                       \
1529     OPNAME ## pixels16_ ## MMX(dst, src, stride, 16);                   \
1530 }                                                                       \
1531                                                                         \
1532 static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src,   \
1533                                           int stride)                   \
1534 {                                                                       \
1535     uint64_t temp[32];                                                  \
1536     uint8_t * const half = (uint8_t*)temp;                              \
1537     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16,         \
1538                                                  stride, 16);           \
1539     OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);  \
1540 }                                                                       \
1541                                                                         \
1542 static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src,   \
1543                                           int stride)                   \
1544 {                                                                       \
1545     OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src,                  \
1546                                              stride, stride, 16);       \
1547 }                                                                       \
1548                                                                         \
1549 static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src,   \
1550                                           int stride)                   \
1551 {                                                                       \
1552     uint64_t temp[32];                                                  \
1553     uint8_t * const half = (uint8_t*)temp;                              \
1554     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16,         \
1555                                                  stride, 16);           \
1556     OPNAME ## pixels16_l2_ ## MMX(dst, src + 1, half,                   \
1557                                   stride, stride, 16);                  \
1558 }                                                                       \
1559                                                                         \
1560 static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src,   \
1561                                           int stride)                   \
1562 {                                                                       \
1563     uint64_t temp[32];                                                  \
1564     uint8_t * const half = (uint8_t*)temp;                              \
1565     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16,         \
1566                                                  stride);               \
1567     OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);  \
1568 }                                                                       \
1569                                                                         \
1570 static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src,   \
1571                                           int stride)                   \
1572 {                                                                       \
1573     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src, stride, stride); \
1574 }                                                                       \
1575                                                                         \
1576 static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src,   \
1577                                           int stride)                   \
1578 {                                                                       \
1579     uint64_t temp[32];                                                  \
1580     uint8_t * const half = (uint8_t*)temp;                              \
1581     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16,         \
1582                                                  stride);               \
1583     OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half,                \
1584                                   stride, stride, 16);                  \
1585 }                                                                       \
1586                                                                         \
1587 static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src,   \
1588                                           int stride)                   \
1589 {                                                                       \
1590     uint64_t half[16 * 2 + 17 * 2];                                     \
1591     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1592     uint8_t * const halfHV = ((uint8_t*)half);                          \
1593     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1594                                                  stride, 17);           \
1595     put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,            \
1596                                       stride, 17);                      \
1597     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1598                                                  16, 16);               \
1599     OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);  \
1600 }                                                                       \
1601                                                                         \
1602 static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src,   \
1603                                           int stride)                   \
1604 {                                                                       \
1605     uint64_t half[16 * 2 + 17 * 2];                                     \
1606     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1607     uint8_t * const halfHV = ((uint8_t*)half);                          \
1608     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1609                                                  stride, 17);           \
1610     put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,        \
1611                                       stride, 17);                      \
1612     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1613                                                  16, 16);               \
1614     OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);  \
1615 }                                                                       \
1616                                                                         \
1617 static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src,   \
1618                                           int stride)                   \
1619 {                                                                       \
1620     uint64_t half[16 * 2 + 17 * 2];                                     \
1621     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1622     uint8_t * const halfHV = ((uint8_t*)half);                          \
1623     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1624                                                  stride, 17);           \
1625     put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,            \
1626                                       stride, 17);                      \
1627     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1628                                                  16, 16);               \
1629     OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV, stride,      \
1630                                   16, 16);                              \
1631 }                                                                       \
1632                                                                         \
1633 static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src,   \
1634                                           int stride)                   \
1635 {                                                                       \
1636     uint64_t half[16 * 2 + 17 * 2];                                     \
1637     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1638     uint8_t * const halfHV = ((uint8_t*)half);                          \
1639     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1640                                                  stride, 17);           \
1641     put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,        \
1642                                       stride, 17);                      \
1643     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1644                                                  16, 16);               \
1645     OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV, stride,      \
1646                                   16, 16);                              \
1647 }                                                                       \
1648                                                                         \
1649 static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src,   \
1650                                           int stride)                   \
1651 {                                                                       \
1652     uint64_t half[16 * 2 + 17 * 2];                                     \
1653     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1654     uint8_t * const halfHV = ((uint8_t*)half);                          \
1655     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1656                                                  stride, 17);           \
1657     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1658                                                  16, 16);               \
1659     OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);  \
1660 }                                                                       \
1661                                                                         \
1662 static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src,   \
1663                                           int stride)                   \
1664 {                                                                       \
1665     uint64_t half[16 * 2 + 17 * 2];                                     \
1666     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1667     uint8_t * const halfHV = ((uint8_t*)half);                          \
1668     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1669                                                  stride, 17);           \
1670     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,         \
1671                                                  16, 16);               \
1672     OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV, stride,      \
1673                                   16, 16);                              \
1674 }                                                                       \
1675                                                                         \
1676 static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src,   \
1677                                           int stride)                   \
1678 {                                                                       \
1679     uint64_t half[17 * 2];                                              \
1680     uint8_t * const halfH = ((uint8_t*)half);                           \
1681     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1682                                                  stride, 17);           \
1683     put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,            \
1684                                       stride, 17);                      \
1685     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);   \
1686 }                                                                       \
1687                                                                         \
1688 static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src,   \
1689                                           int stride)                   \
1690 {                                                                       \
1691     uint64_t half[17 * 2];                                              \
1692     uint8_t * const halfH = ((uint8_t*)half);                           \
1693     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1694                                                  stride, 17);           \
1695     put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,        \
1696                                       stride, 17);                      \
1697     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);   \
1698 }                                                                       \
1699                                                                         \
1700 static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src,   \
1701                                           int stride)                   \
1702 {                                                                       \
1703     uint64_t half[17 * 2];                                              \
1704     uint8_t * const halfH = ((uint8_t*)half);                           \
1705     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,        \
1706                                                  stride, 17);           \
1707     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);   \
1708 }
1709
1710 #define PUT_OP(a, b, temp, size)                \
1711     "mov"#size"        "#a", "#b"       \n\t"
1712
1713 #define AVG_3DNOW_OP(a, b, temp, size)          \
1714     "mov"#size"        "#b", "#temp"    \n\t"   \
1715     "pavgusb        "#temp", "#a"       \n\t"   \
1716     "mov"#size"        "#a", "#b"       \n\t"
1717
1718 #define AVG_MMX2_OP(a, b, temp, size)           \
1719     "mov"#size"        "#b", "#temp"    \n\t"   \
1720     "pavgb          "#temp", "#a"       \n\t"   \
1721     "mov"#size"        "#a", "#b"       \n\t"
1722
1723 QPEL_BASE(put_,        ff_pw_16, _,        PUT_OP,       PUT_OP)
1724 QPEL_BASE(avg_,        ff_pw_16, _,        AVG_MMX2_OP,  AVG_3DNOW_OP)
1725 QPEL_BASE(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP,       PUT_OP)
1726 QPEL_OP(put_,          ff_pw_16, _,        PUT_OP,       3dnow)
1727 QPEL_OP(avg_,          ff_pw_16, _,        AVG_3DNOW_OP, 3dnow)
1728 QPEL_OP(put_no_rnd_,   ff_pw_15, _no_rnd_, PUT_OP,       3dnow)
1729 QPEL_OP(put_,          ff_pw_16, _,        PUT_OP,       mmx2)
1730 QPEL_OP(avg_,          ff_pw_16, _,        AVG_MMX2_OP,  mmx2)
1731 QPEL_OP(put_no_rnd_,   ff_pw_15, _no_rnd_, PUT_OP,       mmx2)
1732
1733 /***********************************/
1734 /* bilinear qpel: not compliant to any spec, only for -lavdopts fast */
1735
1736 #define QPEL_2TAP_XY(OPNAME, SIZE, MMX, XY, HPEL)                              \
1737 static void OPNAME ## 2tap_qpel ## SIZE ## _mc ## XY ## _ ## MMX(uint8_t *dst, \
1738                                                                  uint8_t *src, \
1739                                                                  int stride)   \
1740 {                                                                              \
1741     OPNAME ## pixels ## SIZE ## HPEL(dst, src, stride, SIZE);                  \
1742 }
1743
1744 #define QPEL_2TAP_L3(OPNAME, SIZE, MMX, XY, S0, S1, S2)                        \
1745 static void OPNAME ## 2tap_qpel ## SIZE ## _mc ## XY ## _ ## MMX(uint8_t *dst, \
1746                                                                  uint8_t *src, \
1747                                                                  int stride)   \
1748 {                                                                              \
1749     OPNAME ## 2tap_qpel ## SIZE ## _l3_ ## MMX(dst, src + S0, stride, SIZE,    \
1750                                                S1, S2);                        \
1751 }
1752
1753 #define QPEL_2TAP(OPNAME, SIZE, MMX)                                        \
1754 QPEL_2TAP_XY(OPNAME, SIZE, MMX, 20, _x2_ ## MMX)                            \
1755 QPEL_2TAP_XY(OPNAME, SIZE, MMX, 02, _y2_ ## MMX)                            \
1756 QPEL_2TAP_XY(OPNAME, SIZE, MMX, 22, _xy2_mmx)                               \
1757 static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc00_ ## MMX =    \
1758     OPNAME ## qpel ## SIZE ## _mc00_ ## MMX;                                \
1759 static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc21_ ## MMX =    \
1760     OPNAME ## 2tap_qpel ## SIZE ## _mc20_ ## MMX;                           \
1761 static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc12_ ## MMX =    \
1762     OPNAME ## 2tap_qpel ## SIZE ## _mc02_ ## MMX;                           \
1763 static void OPNAME ## 2tap_qpel ## SIZE ## _mc32_ ## MMX(uint8_t *dst,      \
1764                                                          uint8_t *src,      \
1765                                                          int stride)        \
1766 {                                                                           \
1767     OPNAME ## pixels ## SIZE ## _y2_ ## MMX(dst, src + 1, stride, SIZE);    \
1768 }                                                                           \
1769 static void OPNAME ## 2tap_qpel ## SIZE ## _mc23_ ## MMX(uint8_t *dst,      \
1770                                                          uint8_t *src,      \
1771                                                          int stride)        \
1772 {                                                                           \
1773     OPNAME ## pixels ## SIZE ## _x2_ ## MMX(dst, src + stride,              \
1774                                             stride, SIZE);                  \
1775 }                                                                           \
1776 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 10, 0,           1,       0)                \
1777 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 30, 1,          -1,       0)                \
1778 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 01, 0,           stride,  0)                \
1779 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 03, stride,     -stride,  0)                \
1780 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 11, 0,           stride,  1)                \
1781 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 31, 1,           stride, -1)                \
1782 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 13, stride,     -stride,  1)                \
1783 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 33, stride + 1, -stride, -1)                \
1784
1785 QPEL_2TAP(put_, 16, mmx2)
1786 QPEL_2TAP(avg_, 16, mmx2)
1787 QPEL_2TAP(put_,  8, mmx2)
1788 QPEL_2TAP(avg_,  8, mmx2)
1789 QPEL_2TAP(put_, 16, 3dnow)
1790 QPEL_2TAP(avg_, 16, 3dnow)
1791 QPEL_2TAP(put_,  8, 3dnow)
1792 QPEL_2TAP(avg_,  8, 3dnow)
1793
1794
1795 #if HAVE_YASM
1796 typedef void emu_edge_core_func(uint8_t *buf, const uint8_t *src,
1797                                 x86_reg linesize, x86_reg start_y,
1798                                 x86_reg end_y, x86_reg block_h,
1799                                 x86_reg start_x, x86_reg end_x,
1800                                 x86_reg block_w);
1801 extern emu_edge_core_func ff_emu_edge_core_mmx;
1802 extern emu_edge_core_func ff_emu_edge_core_sse;
1803
1804 static av_always_inline void emulated_edge_mc(uint8_t *buf, const uint8_t *src,
1805                                               int linesize,
1806                                               int block_w, int block_h,
1807                                               int src_x, int src_y,
1808                                               int w, int h,
1809                                               emu_edge_core_func *core_fn)
1810 {
1811     int start_y, start_x, end_y, end_x, src_y_add = 0;
1812
1813     if (src_y >= h) {
1814         src_y_add = h - 1 - src_y;
1815         src_y     = h - 1;
1816     } else if (src_y <= -block_h) {
1817         src_y_add = 1 - block_h - src_y;
1818         src_y     = 1 - block_h;
1819     }
1820     if (src_x >= w) {
1821         src   += w - 1 - src_x;
1822         src_x  = w - 1;
1823     } else if (src_x <= -block_w) {
1824         src   += 1 - block_w - src_x;
1825         src_x  = 1 - block_w;
1826     }
1827
1828     start_y = FFMAX(0, -src_y);
1829     start_x = FFMAX(0, -src_x);
1830     end_y   = FFMIN(block_h, h-src_y);
1831     end_x   = FFMIN(block_w, w-src_x);
1832     assert(start_x < end_x && block_w > 0);
1833     assert(start_y < end_y && block_h > 0);
1834
1835     // fill in the to-be-copied part plus all above/below
1836     src += (src_y_add + start_y) * linesize + start_x;
1837     buf += start_x;
1838     core_fn(buf, src, linesize, start_y, end_y,
1839             block_h, start_x, end_x, block_w);
1840 }
1841
1842 #if ARCH_X86_32
1843 static av_noinline void emulated_edge_mc_mmx(uint8_t *buf, const uint8_t *src,
1844                                              int linesize,
1845                                              int block_w, int block_h,
1846                                              int src_x, int src_y, int w, int h)
1847 {
1848     emulated_edge_mc(buf, src, linesize, block_w, block_h, src_x, src_y,
1849                      w, h, &ff_emu_edge_core_mmx);
1850 }
1851 #endif
1852
1853 static av_noinline void emulated_edge_mc_sse(uint8_t *buf, const uint8_t *src,
1854                                              int linesize,
1855                                              int block_w, int block_h,
1856                                              int src_x, int src_y, int w, int h)
1857 {
1858     emulated_edge_mc(buf, src, linesize, block_w, block_h, src_x, src_y,
1859                      w, h, &ff_emu_edge_core_sse);
1860 }
1861 #endif /* HAVE_YASM */
1862
1863 typedef void emulated_edge_mc_func(uint8_t *dst, const uint8_t *src,
1864                                    int linesize, int block_w, int block_h,
1865                                    int src_x, int src_y, int w, int h);
1866
1867 static av_always_inline void gmc(uint8_t *dst, uint8_t *src,
1868                                  int stride, int h, int ox, int oy,
1869                                  int dxx, int dxy, int dyx, int dyy,
1870                                  int shift, int r, int width, int height,
1871                                  emulated_edge_mc_func *emu_edge_fn)
1872 {
1873     const int w    = 8;
1874     const int ix   = ox  >> (16 + shift);
1875     const int iy   = oy  >> (16 + shift);
1876     const int oxs  = ox  >> 4;
1877     const int oys  = oy  >> 4;
1878     const int dxxs = dxx >> 4;
1879     const int dxys = dxy >> 4;
1880     const int dyxs = dyx >> 4;
1881     const int dyys = dyy >> 4;
1882     const uint16_t r4[4]   = { r, r, r, r };
1883     const uint16_t dxy4[4] = { dxys, dxys, dxys, dxys };
1884     const uint16_t dyy4[4] = { dyys, dyys, dyys, dyys };
1885     const uint64_t shift2 = 2 * shift;
1886     uint8_t edge_buf[(h + 1) * stride];
1887     int x, y;
1888
1889     const int dxw = (dxx - (1 << (16 + shift))) * (w - 1);
1890     const int dyh = (dyy - (1 << (16 + shift))) * (h - 1);
1891     const int dxh = dxy * (h - 1);
1892     const int dyw = dyx * (w - 1);
1893     if ( // non-constant fullpel offset (3% of blocks)
1894         ((ox ^ (ox + dxw)) | (ox ^ (ox + dxh)) | (ox ^ (ox + dxw + dxh)) |
1895          (oy ^ (oy + dyw)) | (oy ^ (oy + dyh)) | (oy ^ (oy + dyw + dyh))) >> (16 + shift)
1896         // uses more than 16 bits of subpel mv (only at huge resolution)
1897         || (dxx | dxy | dyx | dyy) & 15) {
1898         // FIXME could still use mmx for some of the rows
1899         ff_gmc_c(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy,
1900                  shift, r, width, height);
1901         return;
1902     }
1903
1904     src += ix + iy * stride;
1905     if ((unsigned)ix >= width  - w ||
1906         (unsigned)iy >= height - h) {
1907         emu_edge_fn(edge_buf, src, stride, w + 1, h + 1, ix, iy, width, height);
1908         src = edge_buf;
1909     }
1910
1911     __asm__ volatile (
1912         "movd         %0, %%mm6         \n\t"
1913         "pxor      %%mm7, %%mm7         \n\t"
1914         "punpcklwd %%mm6, %%mm6         \n\t"
1915         "punpcklwd %%mm6, %%mm6         \n\t"
1916         :: "r"(1<<shift)
1917     );
1918
1919     for (x = 0; x < w; x += 4) {
1920         uint16_t dx4[4] = { oxs - dxys + dxxs * (x + 0),
1921                             oxs - dxys + dxxs * (x + 1),
1922                             oxs - dxys + dxxs * (x + 2),
1923                             oxs - dxys + dxxs * (x + 3) };
1924         uint16_t dy4[4] = { oys - dyys + dyxs * (x + 0),
1925                             oys - dyys + dyxs * (x + 1),
1926                             oys - dyys + dyxs * (x + 2),
1927                             oys - dyys + dyxs * (x + 3) };
1928
1929         for (y = 0; y < h; y++) {
1930             __asm__ volatile (
1931                 "movq      %0, %%mm4    \n\t"
1932                 "movq      %1, %%mm5    \n\t"
1933                 "paddw     %2, %%mm4    \n\t"
1934                 "paddw     %3, %%mm5    \n\t"
1935                 "movq   %%mm4, %0       \n\t"
1936                 "movq   %%mm5, %1       \n\t"
1937                 "psrlw    $12, %%mm4    \n\t"
1938                 "psrlw    $12, %%mm5    \n\t"
1939                 : "+m"(*dx4), "+m"(*dy4)
1940                 : "m"(*dxy4), "m"(*dyy4)
1941             );
1942
1943             __asm__ volatile (
1944                 "movq      %%mm6, %%mm2 \n\t"
1945                 "movq      %%mm6, %%mm1 \n\t"
1946                 "psubw     %%mm4, %%mm2 \n\t"
1947                 "psubw     %%mm5, %%mm1 \n\t"
1948                 "movq      %%mm2, %%mm0 \n\t"
1949                 "movq      %%mm4, %%mm3 \n\t"
1950                 "pmullw    %%mm1, %%mm0 \n\t" // (s - dx) * (s - dy)
1951                 "pmullw    %%mm5, %%mm3 \n\t" // dx * dy
1952                 "pmullw    %%mm5, %%mm2 \n\t" // (s - dx) * dy
1953                 "pmullw    %%mm4, %%mm1 \n\t" // dx * (s - dy)
1954
1955                 "movd         %4, %%mm5 \n\t"
1956                 "movd         %3, %%mm4 \n\t"
1957                 "punpcklbw %%mm7, %%mm5 \n\t"
1958                 "punpcklbw %%mm7, %%mm4 \n\t"
1959                 "pmullw    %%mm5, %%mm3 \n\t" // src[1, 1] * dx * dy
1960                 "pmullw    %%mm4, %%mm2 \n\t" // src[0, 1] * (s - dx) * dy
1961
1962                 "movd         %2, %%mm5 \n\t"
1963                 "movd         %1, %%mm4 \n\t"
1964                 "punpcklbw %%mm7, %%mm5 \n\t"
1965                 "punpcklbw %%mm7, %%mm4 \n\t"
1966                 "pmullw    %%mm5, %%mm1 \n\t" // src[1, 0] * dx * (s - dy)
1967                 "pmullw    %%mm4, %%mm0 \n\t" // src[0, 0] * (s - dx) * (s - dy)
1968                 "paddw        %5, %%mm1 \n\t"
1969                 "paddw     %%mm3, %%mm2 \n\t"
1970                 "paddw     %%mm1, %%mm0 \n\t"
1971                 "paddw     %%mm2, %%mm0 \n\t"
1972
1973                 "psrlw        %6, %%mm0 \n\t"
1974                 "packuswb  %%mm0, %%mm0 \n\t"
1975                 "movd      %%mm0, %0    \n\t"
1976
1977                 : "=m"(dst[x + y * stride])
1978                 : "m"(src[0]), "m"(src[1]),
1979                   "m"(src[stride]), "m"(src[stride + 1]),
1980                   "m"(*r4), "m"(shift2)
1981             );
1982             src += stride;
1983         }
1984         src += 4 - h * stride;
1985     }
1986 }
1987
1988 #if HAVE_YASM
1989 #if ARCH_X86_32
1990 static void gmc_mmx(uint8_t *dst, uint8_t *src,
1991                     int stride, int h, int ox, int oy,
1992                     int dxx, int dxy, int dyx, int dyy,
1993                     int shift, int r, int width, int height)
1994 {
1995     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,
1996         width, height, &emulated_edge_mc_mmx);
1997 }
1998 #endif
1999 static void gmc_sse(uint8_t *dst, uint8_t *src,
2000                     int stride, int h, int ox, int oy,
2001                     int dxx, int dxy, int dyx, int dyy,
2002                     int shift, int r, int width, int height)
2003 {
2004     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,
2005         width, height, &emulated_edge_mc_sse);
2006 }
2007 #else
2008 static void gmc_mmx(uint8_t *dst, uint8_t *src,
2009                     int stride, int h, int ox, int oy,
2010                     int dxx, int dxy, int dyx, int dyy,
2011                     int shift, int r, int width, int height)
2012 {
2013     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,
2014         width, height, &ff_emulated_edge_mc_8);
2015 }
2016 #endif
2017
2018 #define PREFETCH(name, op)                      \
2019 static void name(void *mem, int stride, int h)  \
2020 {                                               \
2021     const uint8_t *p = mem;                     \
2022     do {                                        \
2023         __asm__ volatile (#op" %0" :: "m"(*p)); \
2024         p += stride;                            \
2025     } while (--h);                              \
2026 }
2027
2028 PREFETCH(prefetch_mmx2,  prefetcht0)
2029 PREFETCH(prefetch_3dnow, prefetch)
2030 #undef PREFETCH
2031
2032 #include "h264_qpel_mmx.c"
2033
2034 void ff_put_h264_chroma_mc8_mmx_rnd  (uint8_t *dst, uint8_t *src,
2035                                       int stride, int h, int x, int y);
2036 void ff_avg_h264_chroma_mc8_mmx2_rnd (uint8_t *dst, uint8_t *src,
2037                                       int stride, int h, int x, int y);
2038 void ff_avg_h264_chroma_mc8_3dnow_rnd(uint8_t *dst, uint8_t *src,
2039                                       int stride, int h, int x, int y);
2040
2041 void ff_put_h264_chroma_mc4_mmx      (uint8_t *dst, uint8_t *src,
2042                                       int stride, int h, int x, int y);
2043 void ff_avg_h264_chroma_mc4_mmx2     (uint8_t *dst, uint8_t *src,
2044                                       int stride, int h, int x, int y);
2045 void ff_avg_h264_chroma_mc4_3dnow    (uint8_t *dst, uint8_t *src,
2046                                       int stride, int h, int x, int y);
2047
2048 void ff_put_h264_chroma_mc2_mmx2     (uint8_t *dst, uint8_t *src,
2049                                       int stride, int h, int x, int y);
2050 void ff_avg_h264_chroma_mc2_mmx2     (uint8_t *dst, uint8_t *src,
2051                                       int stride, int h, int x, int y);
2052
2053 void ff_put_h264_chroma_mc8_ssse3_rnd(uint8_t *dst, uint8_t *src,
2054                                       int stride, int h, int x, int y);
2055 void ff_put_h264_chroma_mc4_ssse3    (uint8_t *dst, uint8_t *src,
2056                                       int stride, int h, int x, int y);
2057
2058 void ff_avg_h264_chroma_mc8_ssse3_rnd(uint8_t *dst, uint8_t *src,
2059                                       int stride, int h, int x, int y);
2060 void ff_avg_h264_chroma_mc4_ssse3    (uint8_t *dst, uint8_t *src,
2061                                       int stride, int h, int x, int y);
2062
2063 #define CHROMA_MC(OP, NUM, DEPTH, OPT)                                  \
2064 void ff_ ## OP ## _h264_chroma_mc ## NUM ## _ ## DEPTH ## _ ## OPT      \
2065                                       (uint8_t *dst, uint8_t *src,      \
2066                                        int stride, int h, int x, int y);
2067
2068 CHROMA_MC(put, 2, 10, mmxext)
2069 CHROMA_MC(avg, 2, 10, mmxext)
2070 CHROMA_MC(put, 4, 10, mmxext)
2071 CHROMA_MC(avg, 4, 10, mmxext)
2072 CHROMA_MC(put, 8, 10, sse2)
2073 CHROMA_MC(avg, 8, 10, sse2)
2074 CHROMA_MC(put, 8, 10, avx)
2075 CHROMA_MC(avg, 8, 10, avx)
2076
2077 /* CAVS-specific */
2078 void ff_put_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride)
2079 {
2080     put_pixels8_mmx(dst, src, stride, 8);
2081 }
2082
2083 void ff_avg_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride)
2084 {
2085     avg_pixels8_mmx(dst, src, stride, 8);
2086 }
2087
2088 void ff_put_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride)
2089 {
2090     put_pixels16_mmx(dst, src, stride, 16);
2091 }
2092
2093 void ff_avg_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride)
2094 {
2095     avg_pixels16_mmx(dst, src, stride, 16);
2096 }
2097
2098 /* VC-1-specific */
2099 void ff_put_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src,
2100                                int stride, int rnd)
2101 {
2102     put_pixels8_mmx(dst, src, stride, 8);
2103 }
2104
2105 void ff_avg_vc1_mspel_mc00_mmx2(uint8_t *dst, const uint8_t *src,
2106                                 int stride, int rnd)
2107 {
2108     avg_pixels8_mmx2(dst, src, stride, 8);
2109 }
2110
2111 /* XXX: Those functions should be suppressed ASAP when all IDCTs are
2112  * converted. */
2113 #if CONFIG_GPL
2114 static void ff_libmpeg2mmx_idct_put(uint8_t *dest, int line_size,
2115                                     DCTELEM *block)
2116 {
2117     ff_mmx_idct(block);
2118     ff_put_pixels_clamped_mmx(block, dest, line_size);
2119 }
2120
2121 static void ff_libmpeg2mmx_idct_add(uint8_t *dest, int line_size,
2122                                     DCTELEM *block)
2123 {
2124     ff_mmx_idct(block);
2125     ff_add_pixels_clamped_mmx(block, dest, line_size);
2126 }
2127
2128 static void ff_libmpeg2mmx2_idct_put(uint8_t *dest, int line_size,
2129                                      DCTELEM *block)
2130 {
2131     ff_mmxext_idct(block);
2132     ff_put_pixels_clamped_mmx(block, dest, line_size);
2133 }
2134
2135 static void ff_libmpeg2mmx2_idct_add(uint8_t *dest, int line_size,
2136                                      DCTELEM *block)
2137 {
2138     ff_mmxext_idct(block);
2139     ff_add_pixels_clamped_mmx(block, dest, line_size);
2140 }
2141 #endif
2142
2143 static void ff_idct_xvid_mmx_put(uint8_t *dest, int line_size, DCTELEM *block)
2144 {
2145     ff_idct_xvid_mmx(block);
2146     ff_put_pixels_clamped_mmx(block, dest, line_size);
2147 }
2148
2149 static void ff_idct_xvid_mmx_add(uint8_t *dest, int line_size, DCTELEM *block)
2150 {
2151     ff_idct_xvid_mmx(block);
2152     ff_add_pixels_clamped_mmx(block, dest, line_size);
2153 }
2154
2155 static void ff_idct_xvid_mmx2_put(uint8_t *dest, int line_size, DCTELEM *block)
2156 {
2157     ff_idct_xvid_mmx2(block);
2158     ff_put_pixels_clamped_mmx(block, dest, line_size);
2159 }
2160
2161 static void ff_idct_xvid_mmx2_add(uint8_t *dest, int line_size, DCTELEM *block)
2162 {
2163     ff_idct_xvid_mmx2(block);
2164     ff_add_pixels_clamped_mmx(block, dest, line_size);
2165 }
2166
2167 static void vorbis_inverse_coupling_3dnow(float *mag, float *ang, int blocksize)
2168 {
2169     int i;
2170     __asm__ volatile ("pxor %%mm7, %%mm7":);
2171     for (i = 0; i < blocksize; i += 2) {
2172         __asm__ volatile (
2173             "movq       %0, %%mm0   \n\t"
2174             "movq       %1, %%mm1   \n\t"
2175             "movq    %%mm0, %%mm2   \n\t"
2176             "movq    %%mm1, %%mm3   \n\t"
2177             "pfcmpge %%mm7, %%mm2   \n\t" // m <= 0.0
2178             "pfcmpge %%mm7, %%mm3   \n\t" // a <= 0.0
2179             "pslld     $31, %%mm2   \n\t" // keep only the sign bit
2180             "pxor    %%mm2, %%mm1   \n\t"
2181             "movq    %%mm3, %%mm4   \n\t"
2182             "pand    %%mm1, %%mm3   \n\t"
2183             "pandn   %%mm1, %%mm4   \n\t"
2184             "pfadd   %%mm0, %%mm3   \n\t" // a = m + ((a < 0) & (a ^ sign(m)))
2185             "pfsub   %%mm4, %%mm0   \n\t" // m = m + ((a > 0) & (a ^ sign(m)))
2186             "movq    %%mm3, %1      \n\t"
2187             "movq    %%mm0, %0      \n\t"
2188             : "+m"(mag[i]), "+m"(ang[i])
2189             :: "memory"
2190         );
2191     }
2192     __asm__ volatile ("femms");
2193 }
2194
2195 static void vorbis_inverse_coupling_sse(float *mag, float *ang, int blocksize)
2196 {
2197     int i;
2198
2199     __asm__ volatile (
2200         "movaps  %0, %%xmm5 \n\t"
2201         :: "m"(ff_pdw_80000000[0])
2202     );
2203     for (i = 0; i < blocksize; i += 4) {
2204         __asm__ volatile (
2205             "movaps      %0, %%xmm0 \n\t"
2206             "movaps      %1, %%xmm1 \n\t"
2207             "xorps   %%xmm2, %%xmm2 \n\t"
2208             "xorps   %%xmm3, %%xmm3 \n\t"
2209             "cmpleps %%xmm0, %%xmm2 \n\t" // m <= 0.0
2210             "cmpleps %%xmm1, %%xmm3 \n\t" // a <= 0.0
2211             "andps   %%xmm5, %%xmm2 \n\t" // keep only the sign bit
2212             "xorps   %%xmm2, %%xmm1 \n\t"
2213             "movaps  %%xmm3, %%xmm4 \n\t"
2214             "andps   %%xmm1, %%xmm3 \n\t"
2215             "andnps  %%xmm1, %%xmm4 \n\t"
2216             "addps   %%xmm0, %%xmm3 \n\t" // a = m + ((a < 0) & (a ^ sign(m)))
2217             "subps   %%xmm4, %%xmm0 \n\t" // m = m + ((a > 0) & (a ^ sign(m)))
2218             "movaps  %%xmm3, %1     \n\t"
2219             "movaps  %%xmm0, %0     \n\t"
2220             : "+m"(mag[i]), "+m"(ang[i])
2221             :: "memory"
2222         );
2223     }
2224 }
2225
2226 #define IF1(x) x
2227 #define IF0(x)
2228
2229 #define MIX5(mono, stereo)                                      \
2230     __asm__ volatile (                                          \
2231         "movss           0(%2), %%xmm5          \n"             \
2232         "movss           8(%2), %%xmm6          \n"             \
2233         "movss          24(%2), %%xmm7          \n"             \
2234         "shufps     $0, %%xmm5, %%xmm5          \n"             \
2235         "shufps     $0, %%xmm6, %%xmm6          \n"             \
2236         "shufps     $0, %%xmm7, %%xmm7          \n"             \
2237         "1:                                     \n"             \
2238         "movaps       (%0, %1), %%xmm0          \n"             \
2239         "movaps  0x400(%0, %1), %%xmm1          \n"             \
2240         "movaps  0x800(%0, %1), %%xmm2          \n"             \
2241         "movaps  0xc00(%0, %1), %%xmm3          \n"             \
2242         "movaps 0x1000(%0, %1), %%xmm4          \n"             \
2243         "mulps          %%xmm5, %%xmm0          \n"             \
2244         "mulps          %%xmm6, %%xmm1          \n"             \
2245         "mulps          %%xmm5, %%xmm2          \n"             \
2246         "mulps          %%xmm7, %%xmm3          \n"             \
2247         "mulps          %%xmm7, %%xmm4          \n"             \
2248  stereo("addps          %%xmm1, %%xmm0          \n")            \
2249         "addps          %%xmm1, %%xmm2          \n"             \
2250         "addps          %%xmm3, %%xmm0          \n"             \
2251         "addps          %%xmm4, %%xmm2          \n"             \
2252    mono("addps          %%xmm2, %%xmm0          \n")            \
2253         "movaps         %%xmm0, (%0, %1)        \n"             \
2254  stereo("movaps         %%xmm2, 0x400(%0, %1)   \n")            \
2255         "add               $16, %0              \n"             \
2256         "jl                 1b                  \n"             \
2257         : "+&r"(i)                                              \
2258         : "r"(samples[0] + len), "r"(matrix)                    \
2259         : XMM_CLOBBERS("%xmm0", "%xmm1", "%xmm2", "%xmm3",      \
2260                       "%xmm4", "%xmm5", "%xmm6", "%xmm7",)      \
2261          "memory"                                               \
2262     );
2263
2264 #define MIX_MISC(stereo)                                        \
2265     __asm__ volatile (                                          \
2266         "1:                                 \n"                 \
2267         "movaps     (%3, %0), %%xmm0        \n"                 \
2268  stereo("movaps       %%xmm0, %%xmm1        \n")                \
2269         "mulps        %%xmm4, %%xmm0        \n"                 \
2270  stereo("mulps        %%xmm5, %%xmm1        \n")                \
2271         "lea    1024(%3, %0), %1            \n"                 \
2272         "mov              %5, %2            \n"                 \
2273         "2:                                 \n"                 \
2274         "movaps         (%1), %%xmm2        \n"                 \
2275  stereo("movaps       %%xmm2, %%xmm3        \n")                \
2276         "mulps      (%4, %2), %%xmm2        \n"                 \
2277  stereo("mulps    16(%4, %2), %%xmm3        \n")                \
2278         "addps        %%xmm2, %%xmm0        \n"                 \
2279  stereo("addps        %%xmm3, %%xmm1        \n")                \
2280         "add           $1024, %1            \n"                 \
2281         "add             $32, %2            \n"                 \
2282         "jl               2b                \n"                 \
2283         "movaps       %%xmm0,     (%3, %0)  \n"                 \
2284  stereo("movaps       %%xmm1, 1024(%3, %0)  \n")                \
2285         "add             $16, %0            \n"                 \
2286         "jl               1b                \n"                 \
2287         : "+&r"(i), "=&r"(j), "=&r"(k)                          \
2288         : "r"(samples[0] + len), "r"(matrix_simd + in_ch),      \
2289           "g"((intptr_t) - 32 * (in_ch - 1))                    \
2290         : "memory"                                              \
2291     );
2292
2293 static void ac3_downmix_sse(float (*samples)[256], float (*matrix)[2],
2294                             int out_ch, int in_ch, int len)
2295 {
2296     int (*matrix_cmp)[2] = (int(*)[2])matrix;
2297     intptr_t i, j, k;
2298
2299     i = -len * sizeof(float);
2300     if (in_ch == 5 && out_ch == 2 &&
2301         !(matrix_cmp[0][1] | matrix_cmp[2][0]   |
2302           matrix_cmp[3][1] | matrix_cmp[4][0]   |
2303           (matrix_cmp[1][0] ^ matrix_cmp[1][1]) |
2304           (matrix_cmp[0][0] ^ matrix_cmp[2][1]))) {
2305         MIX5(IF0, IF1);
2306     } else if (in_ch == 5 && out_ch == 1 &&
2307                matrix_cmp[0][0] == matrix_cmp[2][0] &&
2308                matrix_cmp[3][0] == matrix_cmp[4][0]) {
2309         MIX5(IF1, IF0);
2310     } else {
2311         DECLARE_ALIGNED(16, float, matrix_simd)[AC3_MAX_CHANNELS][2][4];
2312         j = 2 * in_ch * sizeof(float);
2313         __asm__ volatile (
2314             "1:                                 \n"
2315             "sub             $8, %0             \n"
2316             "movss     (%2, %0), %%xmm4         \n"
2317             "movss    4(%2, %0), %%xmm5         \n"
2318             "shufps          $0, %%xmm4, %%xmm4 \n"
2319             "shufps          $0, %%xmm5, %%xmm5 \n"
2320             "movaps      %%xmm4,   (%1, %0, 4)  \n"
2321             "movaps      %%xmm5, 16(%1, %0, 4)  \n"
2322             "jg              1b                 \n"
2323             : "+&r"(j)
2324             : "r"(matrix_simd), "r"(matrix)
2325             : "memory"
2326         );
2327         if (out_ch == 2) {
2328             MIX_MISC(IF1);
2329         } else {
2330             MIX_MISC(IF0);
2331         }
2332     }
2333 }
2334
2335 static void vector_fmul_3dnow(float *dst, const float *src0, const float *src1,
2336                               int len)
2337 {
2338     x86_reg i = (len - 4) * 4;
2339     __asm__ volatile (
2340         "1:                             \n\t"
2341         "movq    (%2, %0), %%mm0        \n\t"
2342         "movq   8(%2, %0), %%mm1        \n\t"
2343         "pfmul   (%3, %0), %%mm0        \n\t"
2344         "pfmul  8(%3, %0), %%mm1        \n\t"
2345         "movq       %%mm0,  (%1, %0)    \n\t"
2346         "movq       %%mm1, 8(%1, %0)    \n\t"
2347         "sub          $16, %0           \n\t"
2348         "jge           1b               \n\t"
2349         "femms                          \n\t"
2350         : "+r"(i)
2351         : "r"(dst), "r"(src0), "r"(src1)
2352         : "memory"
2353     );
2354 }
2355
2356 static void vector_fmul_sse(float *dst, const float *src0, const float *src1,
2357                             int len)
2358 {
2359     x86_reg i = (len - 8) * 4;
2360     __asm__ volatile (
2361         "1:                             \n\t"
2362         "movaps    (%2, %0), %%xmm0     \n\t"
2363         "movaps  16(%2, %0), %%xmm1     \n\t"
2364         "mulps     (%3, %0), %%xmm0     \n\t"
2365         "mulps   16(%3, %0), %%xmm1     \n\t"
2366         "movaps      %%xmm0,   (%1, %0) \n\t"
2367         "movaps      %%xmm1, 16(%1, %0) \n\t"
2368         "sub            $32, %0         \n\t"
2369         "jge             1b             \n\t"
2370         : "+r"(i)
2371         : "r"(dst), "r"(src0), "r"(src1)
2372         : "memory"
2373     );
2374 }
2375
2376 static void vector_fmul_reverse_3dnow2(float *dst, const float *src0,
2377                                        const float *src1, int len)
2378 {
2379     x86_reg i = len * 4 - 16;
2380     __asm__ volatile (
2381         "1:                             \n\t"
2382         "pswapd     8(%1), %%mm0        \n\t"
2383         "pswapd      (%1), %%mm1        \n\t"
2384         "pfmul   (%3, %0), %%mm0        \n\t"
2385         "pfmul  8(%3, %0), %%mm1        \n\t"
2386         "movq       %%mm0,  (%2, %0)    \n\t"
2387         "movq       %%mm1, 8(%2, %0)    \n\t"
2388         "add          $16, %1           \n\t"
2389         "sub          $16, %0           \n\t"
2390         "jge           1b               \n\t"
2391         : "+r"(i), "+r"(src1)
2392         : "r"(dst), "r"(src0)
2393     );
2394     __asm__ volatile ("femms");
2395 }
2396
2397 static void vector_fmul_reverse_sse(float *dst, const float *src0,
2398                                     const float *src1, int len)
2399 {
2400     x86_reg i = len * 4 - 32;
2401     __asm__ volatile (
2402         "1:                                 \n\t"
2403         "movaps         16(%1), %%xmm0      \n\t"
2404         "movaps           (%1), %%xmm1      \n\t"
2405         "shufps  $0x1b, %%xmm0, %%xmm0      \n\t"
2406         "shufps  $0x1b, %%xmm1, %%xmm1      \n\t"
2407         "mulps        (%3, %0), %%xmm0      \n\t"
2408         "mulps      16(%3, %0), %%xmm1      \n\t"
2409         "movaps         %%xmm0,   (%2, %0)  \n\t"
2410         "movaps         %%xmm1, 16(%2, %0)  \n\t"
2411         "add               $32, %1          \n\t"
2412         "sub               $32, %0          \n\t"
2413         "jge                1b              \n\t"
2414         : "+r"(i), "+r"(src1)
2415         : "r"(dst), "r"(src0)
2416     );
2417 }
2418
2419 static void vector_fmul_add_3dnow(float *dst, const float *src0,
2420                                   const float *src1, const float *src2, int len)
2421 {
2422     x86_reg i = (len - 4) * 4;
2423     __asm__ volatile (
2424         "1:                             \n\t"
2425         "movq   (%2, %0), %%mm0         \n\t"
2426         "movq  8(%2, %0), %%mm1         \n\t"
2427         "pfmul  (%3, %0), %%mm0         \n\t"
2428         "pfmul 8(%3, %0), %%mm1         \n\t"
2429         "pfadd  (%4, %0), %%mm0         \n\t"
2430         "pfadd 8(%4, %0), %%mm1         \n\t"
2431         "movq      %%mm0,  (%1, %0)     \n\t"
2432         "movq      %%mm1, 8(%1, %0)     \n\t"
2433         "sub         $16, %0            \n\t"
2434         "jge          1b                \n\t"
2435         : "+r"(i)
2436         : "r"(dst), "r"(src0), "r"(src1), "r"(src2)
2437         : "memory"
2438     );
2439     __asm__ volatile ("femms");
2440 }
2441
2442 static void vector_fmul_add_sse(float *dst, const float *src0,
2443                                 const float *src1, const float *src2, int len)
2444 {
2445     x86_reg i = (len - 8) * 4;
2446     __asm__ volatile (
2447         "1:                             \n\t"
2448         "movaps   (%2, %0), %%xmm0      \n\t"
2449         "movaps 16(%2, %0), %%xmm1      \n\t"
2450         "mulps    (%3, %0), %%xmm0      \n\t"
2451         "mulps  16(%3, %0), %%xmm1      \n\t"
2452         "addps    (%4, %0), %%xmm0      \n\t"
2453         "addps  16(%4, %0), %%xmm1      \n\t"
2454         "movaps     %%xmm0,   (%1, %0)  \n\t"
2455         "movaps     %%xmm1, 16(%1, %0)  \n\t"
2456         "sub           $32, %0          \n\t"
2457         "jge            1b              \n\t"
2458         : "+r"(i)
2459         : "r"(dst), "r"(src0), "r"(src1), "r"(src2)
2460         : "memory"
2461     );
2462 }
2463
2464 #if HAVE_6REGS
2465 static void vector_fmul_window_3dnow2(float *dst, const float *src0,
2466                                       const float *src1, const float *win,
2467                                       int len)
2468 {
2469     x86_reg i = -len * 4;
2470     x86_reg j =  len * 4 - 8;
2471     __asm__ volatile (
2472         "1:                             \n"
2473         "pswapd (%5, %1), %%mm1         \n"
2474         "movq   (%5, %0), %%mm0         \n"
2475         "pswapd (%4, %1), %%mm5         \n"
2476         "movq   (%3, %0), %%mm4         \n"
2477         "movq      %%mm0, %%mm2         \n"
2478         "movq      %%mm1, %%mm3         \n"
2479         "pfmul     %%mm4, %%mm2         \n" // src0[len + i] * win[len + i]
2480         "pfmul     %%mm5, %%mm3         \n" // src1[j]       * win[len + j]
2481         "pfmul     %%mm4, %%mm1         \n" // src0[len + i] * win[len + j]
2482         "pfmul     %%mm5, %%mm0         \n" // src1[j]       * win[len + i]
2483         "pfadd     %%mm3, %%mm2         \n"
2484         "pfsub     %%mm0, %%mm1         \n"
2485         "pswapd    %%mm2, %%mm2         \n"
2486         "movq      %%mm1, (%2, %0)      \n"
2487         "movq      %%mm2, (%2, %1)      \n"
2488         "sub          $8, %1            \n"
2489         "add          $8, %0            \n"
2490         "jl           1b                \n"
2491         "femms                          \n"
2492         : "+r"(i), "+r"(j)
2493         : "r"(dst + len), "r"(src0 + len), "r"(src1), "r"(win + len)
2494     );
2495 }
2496
2497 static void vector_fmul_window_sse(float *dst, const float *src0,
2498                                    const float *src1, const float *win, int len)
2499 {
2500     x86_reg i = -len * 4;
2501     x86_reg j =  len * 4 - 16;
2502     __asm__ volatile (
2503         "1:                             \n"
2504         "movaps      (%5, %1), %%xmm1   \n"
2505         "movaps      (%5, %0), %%xmm0   \n"
2506         "movaps      (%4, %1), %%xmm5   \n"
2507         "movaps      (%3, %0), %%xmm4   \n"
2508         "shufps $0x1b, %%xmm1, %%xmm1   \n"
2509         "shufps $0x1b, %%xmm5, %%xmm5   \n"
2510         "movaps        %%xmm0, %%xmm2   \n"
2511         "movaps        %%xmm1, %%xmm3   \n"
2512         "mulps         %%xmm4, %%xmm2   \n" // src0[len + i] * win[len + i]
2513         "mulps         %%xmm5, %%xmm3   \n" // src1[j]       * win[len + j]
2514         "mulps         %%xmm4, %%xmm1   \n" // src0[len + i] * win[len + j]
2515         "mulps         %%xmm5, %%xmm0   \n" // src1[j]       * win[len + i]
2516         "addps         %%xmm3, %%xmm2   \n"
2517         "subps         %%xmm0, %%xmm1   \n"
2518         "shufps $0x1b, %%xmm2, %%xmm2   \n"
2519         "movaps        %%xmm1, (%2, %0) \n"
2520         "movaps        %%xmm2, (%2, %1) \n"
2521         "sub              $16, %1       \n"
2522         "add              $16, %0       \n"
2523         "jl                1b           \n"
2524         : "+r"(i), "+r"(j)
2525         : "r"(dst + len), "r"(src0 + len), "r"(src1), "r"(win + len)
2526     );
2527 }
2528 #endif /* HAVE_6REGS */
2529
2530 static void vector_clipf_sse(float *dst, const float *src,
2531                              float min, float max, int len)
2532 {
2533     x86_reg i = (len - 16) * 4;
2534     __asm__ volatile (
2535         "movss          %3, %%xmm4      \n\t"
2536         "movss          %4, %%xmm5      \n\t"
2537         "shufps $0, %%xmm4, %%xmm4      \n\t"
2538         "shufps $0, %%xmm5, %%xmm5      \n\t"
2539         "1:                             \n\t"
2540         "movaps   (%2, %0), %%xmm0      \n\t" // 3/1 on intel
2541         "movaps 16(%2, %0), %%xmm1      \n\t"
2542         "movaps 32(%2, %0), %%xmm2      \n\t"
2543         "movaps 48(%2, %0), %%xmm3      \n\t"
2544         "maxps      %%xmm4, %%xmm0      \n\t"
2545         "maxps      %%xmm4, %%xmm1      \n\t"
2546         "maxps      %%xmm4, %%xmm2      \n\t"
2547         "maxps      %%xmm4, %%xmm3      \n\t"
2548         "minps      %%xmm5, %%xmm0      \n\t"
2549         "minps      %%xmm5, %%xmm1      \n\t"
2550         "minps      %%xmm5, %%xmm2      \n\t"
2551         "minps      %%xmm5, %%xmm3      \n\t"
2552         "movaps     %%xmm0,   (%1, %0)  \n\t"
2553         "movaps     %%xmm1, 16(%1, %0)  \n\t"
2554         "movaps     %%xmm2, 32(%1, %0)  \n\t"
2555         "movaps     %%xmm3, 48(%1, %0)  \n\t"
2556         "sub           $64, %0          \n\t"
2557         "jge            1b              \n\t"
2558         : "+&r"(i)
2559         : "r"(dst), "r"(src), "m"(min), "m"(max)
2560         : "memory"
2561     );
2562 }
2563
2564 void ff_vp3_idct_mmx(int16_t *input_data);
2565 void ff_vp3_idct_put_mmx(uint8_t *dest, int line_size, DCTELEM *block);
2566 void ff_vp3_idct_add_mmx(uint8_t *dest, int line_size, DCTELEM *block);
2567
2568 void ff_vp3_idct_dc_add_mmx2(uint8_t *dest, int line_size,
2569                              const DCTELEM *block);
2570
2571 void ff_vp3_v_loop_filter_mmx2(uint8_t *src, int stride, int *bounding_values);
2572 void ff_vp3_h_loop_filter_mmx2(uint8_t *src, int stride, int *bounding_values);
2573
2574 void ff_vp3_idct_sse2(int16_t *input_data);
2575 void ff_vp3_idct_put_sse2(uint8_t *dest, int line_size, DCTELEM *block);
2576 void ff_vp3_idct_add_sse2(uint8_t *dest, int line_size, DCTELEM *block);
2577
2578 int32_t ff_scalarproduct_int16_mmx2(const int16_t *v1, const int16_t *v2,
2579                                     int order);
2580 int32_t ff_scalarproduct_int16_sse2(const int16_t *v1, const int16_t *v2,
2581                                     int order);
2582 int32_t ff_scalarproduct_and_madd_int16_mmx2(int16_t *v1, const int16_t *v2,
2583                                              const int16_t *v3,
2584                                              int order, int mul);
2585 int32_t ff_scalarproduct_and_madd_int16_sse2(int16_t *v1, const int16_t *v2,
2586                                              const int16_t *v3,
2587                                              int order, int mul);
2588 int32_t ff_scalarproduct_and_madd_int16_ssse3(int16_t *v1, const int16_t *v2,
2589                                               const int16_t *v3,
2590                                               int order, int mul);
2591
2592 void ff_apply_window_int16_mmxext    (int16_t *output, const int16_t *input,
2593                                       const int16_t *window, unsigned int len);
2594 void ff_apply_window_int16_mmxext_ba (int16_t *output, const int16_t *input,
2595                                       const int16_t *window, unsigned int len);
2596 void ff_apply_window_int16_sse2      (int16_t *output, const int16_t *input,
2597                                       const int16_t *window, unsigned int len);
2598 void ff_apply_window_int16_sse2_ba   (int16_t *output, const int16_t *input,
2599                                       const int16_t *window, unsigned int len);
2600 void ff_apply_window_int16_ssse3     (int16_t *output, const int16_t *input,
2601                                       const int16_t *window, unsigned int len);
2602 void ff_apply_window_int16_ssse3_atom(int16_t *output, const int16_t *input,
2603                                       const int16_t *window, unsigned int len);
2604
2605 void ff_bswap32_buf_ssse3(uint32_t *dst, const uint32_t *src, int w);
2606 void ff_bswap32_buf_sse2(uint32_t *dst, const uint32_t *src, int w);
2607
2608 void ff_add_hfyu_median_prediction_mmx2(uint8_t *dst, const uint8_t *top,
2609                                         const uint8_t *diff, int w,
2610                                         int *left, int *left_top);
2611 int  ff_add_hfyu_left_prediction_ssse3(uint8_t *dst, const uint8_t *src,
2612                                        int w, int left);
2613 int  ff_add_hfyu_left_prediction_sse4(uint8_t *dst, const uint8_t *src,
2614                                       int w, int left);
2615
2616 float ff_scalarproduct_float_sse(const float *v1, const float *v2, int order);
2617
2618 void ff_vector_clip_int32_mmx     (int32_t *dst, const int32_t *src,
2619                                    int32_t min, int32_t max, unsigned int len);
2620 void ff_vector_clip_int32_sse2    (int32_t *dst, const int32_t *src,
2621                                    int32_t min, int32_t max, unsigned int len);
2622 void ff_vector_clip_int32_int_sse2(int32_t *dst, const int32_t *src,
2623                                    int32_t min, int32_t max, unsigned int len);
2624 void ff_vector_clip_int32_sse4    (int32_t *dst, const int32_t *src,
2625                                    int32_t min, int32_t max, unsigned int len);
2626
2627 extern void ff_butterflies_float_interleave_sse(float *dst, const float *src0,
2628                                                 const float *src1, int len);
2629 extern void ff_butterflies_float_interleave_avx(float *dst, const float *src0,
2630                                                 const float *src1, int len);
2631
2632 #define SET_QPEL_FUNCS(PFX, IDX, SIZE, CPU, PREFIX)                          \
2633     do {                                                                     \
2634     c->PFX ## _pixels_tab[IDX][ 0] = PREFIX ## PFX ## SIZE ## _mc00_ ## CPU; \
2635     c->PFX ## _pixels_tab[IDX][ 1] = PREFIX ## PFX ## SIZE ## _mc10_ ## CPU; \
2636     c->PFX ## _pixels_tab[IDX][ 2] = PREFIX ## PFX ## SIZE ## _mc20_ ## CPU; \
2637     c->PFX ## _pixels_tab[IDX][ 3] = PREFIX ## PFX ## SIZE ## _mc30_ ## CPU; \
2638     c->PFX ## _pixels_tab[IDX][ 4] = PREFIX ## PFX ## SIZE ## _mc01_ ## CPU; \
2639     c->PFX ## _pixels_tab[IDX][ 5] = PREFIX ## PFX ## SIZE ## _mc11_ ## CPU; \
2640     c->PFX ## _pixels_tab[IDX][ 6] = PREFIX ## PFX ## SIZE ## _mc21_ ## CPU; \
2641     c->PFX ## _pixels_tab[IDX][ 7] = PREFIX ## PFX ## SIZE ## _mc31_ ## CPU; \
2642     c->PFX ## _pixels_tab[IDX][ 8] = PREFIX ## PFX ## SIZE ## _mc02_ ## CPU; \
2643     c->PFX ## _pixels_tab[IDX][ 9] = PREFIX ## PFX ## SIZE ## _mc12_ ## CPU; \
2644     c->PFX ## _pixels_tab[IDX][10] = PREFIX ## PFX ## SIZE ## _mc22_ ## CPU; \
2645     c->PFX ## _pixels_tab[IDX][11] = PREFIX ## PFX ## SIZE ## _mc32_ ## CPU; \
2646     c->PFX ## _pixels_tab[IDX][12] = PREFIX ## PFX ## SIZE ## _mc03_ ## CPU; \
2647     c->PFX ## _pixels_tab[IDX][13] = PREFIX ## PFX ## SIZE ## _mc13_ ## CPU; \
2648     c->PFX ## _pixels_tab[IDX][14] = PREFIX ## PFX ## SIZE ## _mc23_ ## CPU; \
2649     c->PFX ## _pixels_tab[IDX][15] = PREFIX ## PFX ## SIZE ## _mc33_ ## CPU; \
2650     } while (0)
2651
2652 #define SET_HPEL_FUNCS(PFX, IDX, SIZE, CPU)                                     \
2653     do {                                                                        \
2654         c->PFX ## _pixels_tab[IDX][0] = PFX ## _pixels ## SIZE ## _     ## CPU; \
2655         c->PFX ## _pixels_tab[IDX][1] = PFX ## _pixels ## SIZE ## _x2_  ## CPU; \
2656         c->PFX ## _pixels_tab[IDX][2] = PFX ## _pixels ## SIZE ## _y2_  ## CPU; \
2657         c->PFX ## _pixels_tab[IDX][3] = PFX ## _pixels ## SIZE ## _xy2_ ## CPU; \
2658     } while (0)
2659
2660 #define H264_QPEL_FUNCS(x, y, CPU)                                                            \
2661     do {                                                                                      \
2662         c->put_h264_qpel_pixels_tab[0][x + y * 4] = put_h264_qpel16_mc ## x ## y ## _ ## CPU; \
2663         c->put_h264_qpel_pixels_tab[1][x + y * 4] = put_h264_qpel8_mc  ## x ## y ## _ ## CPU; \
2664         c->avg_h264_qpel_pixels_tab[0][x + y * 4] = avg_h264_qpel16_mc ## x ## y ## _ ## CPU; \
2665         c->avg_h264_qpel_pixels_tab[1][x + y * 4] = avg_h264_qpel8_mc  ## x ## y ## _ ## CPU; \
2666     } while (0)
2667
2668 #define H264_QPEL_FUNCS_10(x, y, CPU)                                                               \
2669     do {                                                                                            \
2670         c->put_h264_qpel_pixels_tab[0][x + y * 4] = ff_put_h264_qpel16_mc ## x ## y ## _10_ ## CPU; \
2671         c->put_h264_qpel_pixels_tab[1][x + y * 4] = ff_put_h264_qpel8_mc  ## x ## y ## _10_ ## CPU; \
2672         c->avg_h264_qpel_pixels_tab[0][x + y * 4] = ff_avg_h264_qpel16_mc ## x ## y ## _10_ ## CPU; \
2673         c->avg_h264_qpel_pixels_tab[1][x + y * 4] = ff_avg_h264_qpel8_mc  ## x ## y ## _10_ ## CPU; \
2674     } while (0)
2675
2676 static void dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx, int mm_flags)
2677 {
2678     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
2679
2680     c->put_pixels_clamped        = ff_put_pixels_clamped_mmx;
2681     c->put_signed_pixels_clamped = ff_put_signed_pixels_clamped_mmx;
2682     c->add_pixels_clamped        = ff_add_pixels_clamped_mmx;
2683
2684     if (!high_bit_depth) {
2685         c->clear_block  = clear_block_mmx;
2686         c->clear_blocks = clear_blocks_mmx;
2687         c->draw_edges   = draw_edges_mmx;
2688
2689         SET_HPEL_FUNCS(put,        0, 16, mmx);
2690         SET_HPEL_FUNCS(put_no_rnd, 0, 16, mmx);
2691         SET_HPEL_FUNCS(avg,        0, 16, mmx);
2692         SET_HPEL_FUNCS(avg_no_rnd, 0, 16, mmx);
2693         SET_HPEL_FUNCS(put,        1,  8, mmx);
2694         SET_HPEL_FUNCS(put_no_rnd, 1,  8, mmx);
2695         SET_HPEL_FUNCS(avg,        1,  8, mmx);
2696         SET_HPEL_FUNCS(avg_no_rnd, 1,  8, mmx);
2697     }
2698
2699 #if ARCH_X86_32 || !HAVE_YASM
2700     c->gmc = gmc_mmx;
2701 #endif
2702 #if ARCH_X86_32 && HAVE_YASM
2703     if (!high_bit_depth)
2704         c->emulated_edge_mc = emulated_edge_mc_mmx;
2705 #endif
2706
2707     c->add_bytes = add_bytes_mmx;
2708
2709     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
2710         c->h263_v_loop_filter = h263_v_loop_filter_mmx;
2711         c->h263_h_loop_filter = h263_h_loop_filter_mmx;
2712     }
2713
2714 #if HAVE_YASM
2715     if (!high_bit_depth && CONFIG_H264CHROMA) {
2716         c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_mmx_rnd;
2717         c->put_h264_chroma_pixels_tab[1] = ff_put_h264_chroma_mc4_mmx;
2718     }
2719
2720     c->vector_clip_int32 = ff_vector_clip_int32_mmx;
2721 #endif
2722
2723 }
2724
2725 static void dsputil_init_mmx2(DSPContext *c, AVCodecContext *avctx,
2726                               int mm_flags)
2727 {
2728     const int bit_depth      = avctx->bits_per_raw_sample;
2729     const int high_bit_depth = bit_depth > 8;
2730
2731     c->prefetch = prefetch_mmx2;
2732
2733     if (!high_bit_depth) {
2734         c->put_pixels_tab[0][1] = put_pixels16_x2_mmx2;
2735         c->put_pixels_tab[0][2] = put_pixels16_y2_mmx2;
2736
2737         c->avg_pixels_tab[0][0] = avg_pixels16_mmx2;
2738         c->avg_pixels_tab[0][1] = avg_pixels16_x2_mmx2;
2739         c->avg_pixels_tab[0][2] = avg_pixels16_y2_mmx2;
2740
2741         c->put_pixels_tab[1][1] = put_pixels8_x2_mmx2;
2742         c->put_pixels_tab[1][2] = put_pixels8_y2_mmx2;
2743
2744         c->avg_pixels_tab[1][0] = avg_pixels8_mmx2;
2745         c->avg_pixels_tab[1][1] = avg_pixels8_x2_mmx2;
2746         c->avg_pixels_tab[1][2] = avg_pixels8_y2_mmx2;
2747     }
2748
2749     if (!(avctx->flags & CODEC_FLAG_BITEXACT)) {
2750         if (!high_bit_depth) {
2751             c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx2;
2752             c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx2;
2753             c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx2;
2754             c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx2;
2755
2756             c->avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx2;
2757             c->avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx2;
2758         }
2759
2760         if (CONFIG_VP3_DECODER && HAVE_YASM) {
2761             c->vp3_v_loop_filter = ff_vp3_v_loop_filter_mmx2;
2762             c->vp3_h_loop_filter = ff_vp3_h_loop_filter_mmx2;
2763         }
2764     }
2765     if (CONFIG_VP3_DECODER && HAVE_YASM)
2766         c->vp3_idct_dc_add = ff_vp3_idct_dc_add_mmx2;
2767
2768     if (CONFIG_VP3_DECODER && (avctx->codec_id == CODEC_ID_VP3 ||
2769                                avctx->codec_id == CODEC_ID_THEORA)) {
2770         c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_exact_mmx2;
2771         c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_exact_mmx2;
2772     }
2773
2774     if (CONFIG_H264QPEL) {
2775         SET_QPEL_FUNCS(put_qpel,        0, 16, mmx2, );
2776         SET_QPEL_FUNCS(put_qpel,        1,  8, mmx2, );
2777         SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, mmx2, );
2778         SET_QPEL_FUNCS(put_no_rnd_qpel, 1,  8, mmx2, );
2779         SET_QPEL_FUNCS(avg_qpel,        0, 16, mmx2, );
2780         SET_QPEL_FUNCS(avg_qpel,        1,  8, mmx2, );
2781
2782         if (!high_bit_depth) {
2783             SET_QPEL_FUNCS(put_h264_qpel, 0, 16, mmx2, );
2784             SET_QPEL_FUNCS(put_h264_qpel, 1,  8, mmx2, );
2785             SET_QPEL_FUNCS(put_h264_qpel, 2,  4, mmx2, );
2786             SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, mmx2, );
2787             SET_QPEL_FUNCS(avg_h264_qpel, 1,  8, mmx2, );
2788             SET_QPEL_FUNCS(avg_h264_qpel, 2,  4, mmx2, );
2789         } else if (bit_depth == 10) {
2790 #if HAVE_YASM
2791 #if !ARCH_X86_64
2792             SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, 10_mmxext, ff_);
2793             SET_QPEL_FUNCS(put_h264_qpel, 0, 16, 10_mmxext, ff_);
2794             SET_QPEL_FUNCS(put_h264_qpel, 1,  8, 10_mmxext, ff_);
2795             SET_QPEL_FUNCS(avg_h264_qpel, 1,  8, 10_mmxext, ff_);
2796 #endif
2797             SET_QPEL_FUNCS(put_h264_qpel, 2, 4,  10_mmxext, ff_);
2798             SET_QPEL_FUNCS(avg_h264_qpel, 2, 4,  10_mmxext, ff_);
2799 #endif
2800         }
2801
2802         SET_QPEL_FUNCS(put_2tap_qpel, 0, 16, mmx2, );
2803         SET_QPEL_FUNCS(put_2tap_qpel, 1,  8, mmx2, );
2804         SET_QPEL_FUNCS(avg_2tap_qpel, 0, 16, mmx2, );
2805         SET_QPEL_FUNCS(avg_2tap_qpel, 1,  8, mmx2, );
2806     }
2807
2808 #if HAVE_YASM
2809     if (!high_bit_depth && CONFIG_H264CHROMA) {
2810         c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_mmx2_rnd;
2811         c->avg_h264_chroma_pixels_tab[1] = ff_avg_h264_chroma_mc4_mmx2;
2812         c->avg_h264_chroma_pixels_tab[2] = ff_avg_h264_chroma_mc2_mmx2;
2813         c->put_h264_chroma_pixels_tab[2] = ff_put_h264_chroma_mc2_mmx2;
2814     }
2815     if (bit_depth == 10 && CONFIG_H264CHROMA) {
2816         c->put_h264_chroma_pixels_tab[2] = ff_put_h264_chroma_mc2_10_mmxext;
2817         c->avg_h264_chroma_pixels_tab[2] = ff_avg_h264_chroma_mc2_10_mmxext;
2818         c->put_h264_chroma_pixels_tab[1] = ff_put_h264_chroma_mc4_10_mmxext;
2819         c->avg_h264_chroma_pixels_tab[1] = ff_avg_h264_chroma_mc4_10_mmxext;
2820     }
2821
2822     c->add_hfyu_median_prediction   = ff_add_hfyu_median_prediction_mmx2;
2823
2824     c->scalarproduct_int16          = ff_scalarproduct_int16_mmx2;
2825     c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_mmx2;
2826
2827     if (avctx->flags & CODEC_FLAG_BITEXACT) {
2828         c->apply_window_int16 = ff_apply_window_int16_mmxext_ba;
2829     } else {
2830         c->apply_window_int16 = ff_apply_window_int16_mmxext;
2831     }
2832 #endif
2833 }
2834
2835 static void dsputil_init_3dnow(DSPContext *c, AVCodecContext *avctx,
2836                                int mm_flags)
2837 {
2838     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
2839
2840     c->prefetch = prefetch_3dnow;
2841
2842     if (!high_bit_depth) {
2843         c->put_pixels_tab[0][1] = put_pixels16_x2_3dnow;
2844         c->put_pixels_tab[0][2] = put_pixels16_y2_3dnow;
2845
2846         c->avg_pixels_tab[0][0] = avg_pixels16_3dnow;
2847         c->avg_pixels_tab[0][1] = avg_pixels16_x2_3dnow;
2848         c->avg_pixels_tab[0][2] = avg_pixels16_y2_3dnow;
2849
2850         c->put_pixels_tab[1][1] = put_pixels8_x2_3dnow;
2851         c->put_pixels_tab[1][2] = put_pixels8_y2_3dnow;
2852
2853         c->avg_pixels_tab[1][0] = avg_pixels8_3dnow;
2854         c->avg_pixels_tab[1][1] = avg_pixels8_x2_3dnow;
2855         c->avg_pixels_tab[1][2] = avg_pixels8_y2_3dnow;
2856
2857         if (!(avctx->flags & CODEC_FLAG_BITEXACT)){
2858             c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_3dnow;
2859             c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_3dnow;
2860             c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_3dnow;
2861             c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_3dnow;
2862
2863             c->avg_pixels_tab[0][3] = avg_pixels16_xy2_3dnow;
2864             c->avg_pixels_tab[1][3] = avg_pixels8_xy2_3dnow;
2865         }
2866     }
2867
2868     if (CONFIG_VP3_DECODER && (avctx->codec_id == CODEC_ID_VP3 ||
2869                                avctx->codec_id == CODEC_ID_THEORA)) {
2870         c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_exact_3dnow;
2871         c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_exact_3dnow;
2872     }
2873
2874     if (CONFIG_H264QPEL) {
2875         SET_QPEL_FUNCS(put_qpel,        0, 16, 3dnow, );
2876         SET_QPEL_FUNCS(put_qpel,        1,  8, 3dnow, );
2877         SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, 3dnow, );
2878         SET_QPEL_FUNCS(put_no_rnd_qpel, 1,  8, 3dnow, );
2879         SET_QPEL_FUNCS(avg_qpel,        0, 16, 3dnow, );
2880         SET_QPEL_FUNCS(avg_qpel,        1,  8, 3dnow, );
2881
2882         if (!high_bit_depth) {
2883             SET_QPEL_FUNCS(put_h264_qpel, 0, 16, 3dnow, );
2884             SET_QPEL_FUNCS(put_h264_qpel, 1,  8, 3dnow, );
2885             SET_QPEL_FUNCS(put_h264_qpel, 2,  4, 3dnow, );
2886             SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, 3dnow, );
2887             SET_QPEL_FUNCS(avg_h264_qpel, 1,  8, 3dnow, );
2888             SET_QPEL_FUNCS(avg_h264_qpel, 2,  4, 3dnow, );
2889         }
2890
2891         SET_QPEL_FUNCS(put_2tap_qpel, 0, 16, 3dnow, );
2892         SET_QPEL_FUNCS(put_2tap_qpel, 1,  8, 3dnow, );
2893         SET_QPEL_FUNCS(avg_2tap_qpel, 0, 16, 3dnow, );
2894         SET_QPEL_FUNCS(avg_2tap_qpel, 1,  8, 3dnow, );
2895     }
2896
2897 #if HAVE_YASM
2898     if (!high_bit_depth && CONFIG_H264CHROMA) {
2899         c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_3dnow_rnd;
2900         c->avg_h264_chroma_pixels_tab[1] = ff_avg_h264_chroma_mc4_3dnow;
2901     }
2902 #endif
2903
2904     c->vorbis_inverse_coupling = vorbis_inverse_coupling_3dnow;
2905     c->vector_fmul             = vector_fmul_3dnow;
2906     c->vector_fmul_add         = vector_fmul_add_3dnow;
2907
2908 #if HAVE_7REGS
2909     c->add_hfyu_median_prediction = add_hfyu_median_prediction_cmov;
2910 #endif
2911 }
2912
2913 static void dsputil_init_3dnow2(DSPContext *c, AVCodecContext *avctx,
2914                                 int mm_flags)
2915 {
2916     c->vector_fmul_reverse = vector_fmul_reverse_3dnow2;
2917 #if HAVE_6REGS
2918     c->vector_fmul_window  = vector_fmul_window_3dnow2;
2919 #endif
2920 }
2921
2922 static void dsputil_init_sse(DSPContext *c, AVCodecContext *avctx, int mm_flags)
2923 {
2924     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
2925
2926     if (!high_bit_depth) {
2927         if (!(CONFIG_MPEG_XVMC_DECODER && avctx->xvmc_acceleration > 1)) {
2928             /* XvMCCreateBlocks() may not allocate 16-byte aligned blocks */
2929             c->clear_block  = clear_block_sse;
2930             c->clear_blocks = clear_blocks_sse;
2931         }
2932     }
2933
2934     c->vorbis_inverse_coupling = vorbis_inverse_coupling_sse;
2935     c->ac3_downmix             = ac3_downmix_sse;
2936     c->vector_fmul             = vector_fmul_sse;
2937     c->vector_fmul_reverse     = vector_fmul_reverse_sse;
2938
2939     if (!(mm_flags & AV_CPU_FLAG_3DNOW))
2940         c->vector_fmul_add = vector_fmul_add_sse;
2941
2942 #if HAVE_6REGS
2943     c->vector_fmul_window = vector_fmul_window_sse;
2944 #endif
2945
2946     c->vector_clipf = vector_clipf_sse;
2947
2948 #if HAVE_YASM
2949     c->scalarproduct_float          = ff_scalarproduct_float_sse;
2950     c->butterflies_float_interleave = ff_butterflies_float_interleave_sse;
2951
2952     if (!high_bit_depth)
2953         c->emulated_edge_mc = emulated_edge_mc_sse;
2954     c->gmc = gmc_sse;
2955 #endif
2956 }
2957
2958 static void dsputil_init_sse2(DSPContext *c, AVCodecContext *avctx,
2959                               int mm_flags)
2960 {
2961     const int bit_depth      = avctx->bits_per_raw_sample;
2962     const int high_bit_depth = bit_depth > 8;
2963
2964     if (mm_flags & AV_CPU_FLAG_3DNOW) {
2965         // these functions are slower than mmx on AMD, but faster on Intel
2966         if (!high_bit_depth) {
2967             c->put_pixels_tab[0][0]        = put_pixels16_sse2;
2968             c->put_no_rnd_pixels_tab[0][0] = put_pixels16_sse2;
2969             c->avg_pixels_tab[0][0]        = avg_pixels16_sse2;
2970             if (CONFIG_H264QPEL)
2971                 H264_QPEL_FUNCS(0, 0, sse2);
2972         }
2973     }
2974
2975     if (!high_bit_depth && CONFIG_H264QPEL) {
2976         H264_QPEL_FUNCS(0, 1, sse2);
2977         H264_QPEL_FUNCS(0, 2, sse2);
2978         H264_QPEL_FUNCS(0, 3, sse2);
2979         H264_QPEL_FUNCS(1, 1, sse2);
2980         H264_QPEL_FUNCS(1, 2, sse2);
2981         H264_QPEL_FUNCS(1, 3, sse2);
2982         H264_QPEL_FUNCS(2, 1, sse2);
2983         H264_QPEL_FUNCS(2, 2, sse2);
2984         H264_QPEL_FUNCS(2, 3, sse2);
2985         H264_QPEL_FUNCS(3, 1, sse2);
2986         H264_QPEL_FUNCS(3, 2, sse2);
2987         H264_QPEL_FUNCS(3, 3, sse2);
2988     }
2989
2990 #if HAVE_YASM
2991     if (bit_depth == 10) {
2992         if (CONFIG_H264QPEL) {
2993             SET_QPEL_FUNCS(put_h264_qpel, 0, 16, 10_sse2, ff_);
2994             SET_QPEL_FUNCS(put_h264_qpel, 1,  8, 10_sse2, ff_);
2995             SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, 10_sse2, ff_);
2996             SET_QPEL_FUNCS(avg_h264_qpel, 1,  8, 10_sse2, ff_);
2997             H264_QPEL_FUNCS_10(1, 0, sse2_cache64);
2998             H264_QPEL_FUNCS_10(2, 0, sse2_cache64);
2999             H264_QPEL_FUNCS_10(3, 0, sse2_cache64);
3000         }
3001         if (CONFIG_H264CHROMA) {
3002             c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_10_sse2;
3003             c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_10_sse2;
3004         }
3005     }
3006
3007     c->scalarproduct_int16          = ff_scalarproduct_int16_sse2;
3008     c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_sse2;
3009     if (mm_flags & AV_CPU_FLAG_ATOM) {
3010         c->vector_clip_int32 = ff_vector_clip_int32_int_sse2;
3011     } else {
3012         c->vector_clip_int32 = ff_vector_clip_int32_sse2;
3013     }
3014     if (avctx->flags & CODEC_FLAG_BITEXACT) {
3015         c->apply_window_int16 = ff_apply_window_int16_sse2_ba;
3016     } else if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {
3017         c->apply_window_int16 = ff_apply_window_int16_sse2;
3018     }
3019     c->bswap_buf = ff_bswap32_buf_sse2;
3020 #endif
3021 }
3022
3023 static void dsputil_init_ssse3(DSPContext *c, AVCodecContext *avctx,
3024                                int mm_flags)
3025 {
3026 #if HAVE_SSSE3
3027     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
3028     const int bit_depth      = avctx->bits_per_raw_sample;
3029
3030     if (!high_bit_depth && CONFIG_H264QPEL) {
3031         H264_QPEL_FUNCS(1, 0, ssse3);
3032         H264_QPEL_FUNCS(1, 1, ssse3);
3033         H264_QPEL_FUNCS(1, 2, ssse3);
3034         H264_QPEL_FUNCS(1, 3, ssse3);
3035         H264_QPEL_FUNCS(2, 0, ssse3);
3036         H264_QPEL_FUNCS(2, 1, ssse3);
3037         H264_QPEL_FUNCS(2, 2, ssse3);
3038         H264_QPEL_FUNCS(2, 3, ssse3);
3039         H264_QPEL_FUNCS(3, 0, ssse3);
3040         H264_QPEL_FUNCS(3, 1, ssse3);
3041         H264_QPEL_FUNCS(3, 2, ssse3);
3042         H264_QPEL_FUNCS(3, 3, ssse3);
3043     }
3044 #if HAVE_YASM
3045     else if (bit_depth == 10 && CONFIG_H264QPEL) {
3046         H264_QPEL_FUNCS_10(1, 0, ssse3_cache64);
3047         H264_QPEL_FUNCS_10(2, 0, ssse3_cache64);
3048         H264_QPEL_FUNCS_10(3, 0, ssse3_cache64);
3049     }
3050     if (!high_bit_depth && CONFIG_H264CHROMA) {
3051         c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_ssse3_rnd;
3052         c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_ssse3_rnd;
3053         c->put_h264_chroma_pixels_tab[1] = ff_put_h264_chroma_mc4_ssse3;
3054         c->avg_h264_chroma_pixels_tab[1] = ff_avg_h264_chroma_mc4_ssse3;
3055     }
3056     c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_ssse3;
3057     if (mm_flags & AV_CPU_FLAG_SSE4) // not really sse4, just slow on Conroe
3058         c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_sse4;
3059
3060     if (mm_flags & AV_CPU_FLAG_ATOM)
3061         c->apply_window_int16 = ff_apply_window_int16_ssse3_atom;
3062     else
3063         c->apply_window_int16 = ff_apply_window_int16_ssse3;
3064     if (!(mm_flags & (AV_CPU_FLAG_SSE42|AV_CPU_FLAG_3DNOW))) // cachesplit
3065         c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_ssse3;
3066     c->bswap_buf = ff_bswap32_buf_ssse3;
3067 #endif
3068 #endif
3069 }
3070
3071 static void dsputil_init_sse4(DSPContext *c, AVCodecContext *avctx,
3072                               int mm_flags)
3073 {
3074 #if HAVE_YASM
3075     c->vector_clip_int32 = ff_vector_clip_int32_sse4;
3076 #endif
3077 }
3078
3079 static void dsputil_init_avx(DSPContext *c, AVCodecContext *avctx, int mm_flags)
3080 {
3081 #if HAVE_AVX && HAVE_YASM
3082     const int bit_depth = avctx->bits_per_raw_sample;
3083
3084     if (bit_depth == 10) {
3085         // AVX implies !cache64.
3086         // TODO: Port cache(32|64) detection from x264.
3087         if (CONFIG_H264QPEL) {
3088             H264_QPEL_FUNCS_10(1, 0, sse2);
3089             H264_QPEL_FUNCS_10(2, 0, sse2);
3090             H264_QPEL_FUNCS_10(3, 0, sse2);
3091         }
3092
3093         if (CONFIG_H264CHROMA) {
3094             c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_10_avx;
3095             c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_10_avx;
3096         }
3097     }
3098     c->butterflies_float_interleave = ff_butterflies_float_interleave_avx;
3099 #endif
3100 }
3101
3102 void ff_dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx)
3103 {
3104     int mm_flags = av_get_cpu_flags();
3105
3106 #if 0
3107     av_log(avctx, AV_LOG_INFO, "libavcodec: CPU flags:");
3108     if (mm_flags & AV_CPU_FLAG_MMX)
3109         av_log(avctx, AV_LOG_INFO, " mmx");
3110     if (mm_flags & AV_CPU_FLAG_MMX2)
3111         av_log(avctx, AV_LOG_INFO, " mmx2");
3112     if (mm_flags & AV_CPU_FLAG_3DNOW)
3113         av_log(avctx, AV_LOG_INFO, " 3dnow");
3114     if (mm_flags & AV_CPU_FLAG_SSE)
3115         av_log(avctx, AV_LOG_INFO, " sse");
3116     if (mm_flags & AV_CPU_FLAG_SSE2)
3117         av_log(avctx, AV_LOG_INFO, " sse2");
3118     av_log(avctx, AV_LOG_INFO, "\n");
3119 #endif
3120
3121     if (mm_flags & AV_CPU_FLAG_MMX) {
3122         const int idct_algo = avctx->idct_algo;
3123
3124         if (avctx->lowres == 0 && avctx->bits_per_raw_sample <= 8) {
3125             if (idct_algo == FF_IDCT_AUTO || idct_algo == FF_IDCT_SIMPLEMMX) {
3126                 c->idct_put              = ff_simple_idct_put_mmx;
3127                 c->idct_add              = ff_simple_idct_add_mmx;
3128                 c->idct                  = ff_simple_idct_mmx;
3129                 c->idct_permutation_type = FF_SIMPLE_IDCT_PERM;
3130 #if CONFIG_GPL
3131             } else if (idct_algo == FF_IDCT_LIBMPEG2MMX) {
3132                 if (mm_flags & AV_CPU_FLAG_MMX2) {
3133                     c->idct_put = ff_libmpeg2mmx2_idct_put;
3134                     c->idct_add = ff_libmpeg2mmx2_idct_add;
3135                     c->idct     = ff_mmxext_idct;
3136                 } else {
3137                     c->idct_put = ff_libmpeg2mmx_idct_put;
3138                     c->idct_add = ff_libmpeg2mmx_idct_add;
3139                     c->idct     = ff_mmx_idct;
3140                 }
3141                 c->idct_permutation_type = FF_LIBMPEG2_IDCT_PERM;
3142 #endif
3143             } else if ((CONFIG_VP3_DECODER || CONFIG_VP5_DECODER ||
3144                         CONFIG_VP6_DECODER) &&
3145                        idct_algo == FF_IDCT_VP3 && HAVE_YASM) {
3146                 if (mm_flags & AV_CPU_FLAG_SSE2) {
3147                     c->idct_put              = ff_vp3_idct_put_sse2;
3148                     c->idct_add              = ff_vp3_idct_add_sse2;
3149                     c->idct                  = ff_vp3_idct_sse2;
3150                     c->idct_permutation_type = FF_TRANSPOSE_IDCT_PERM;
3151                 } else {
3152                     c->idct_put              = ff_vp3_idct_put_mmx;
3153                     c->idct_add              = ff_vp3_idct_add_mmx;
3154                     c->idct                  = ff_vp3_idct_mmx;
3155                     c->idct_permutation_type = FF_PARTTRANS_IDCT_PERM;
3156                 }
3157             } else if (idct_algo == FF_IDCT_CAVS) {
3158                     c->idct_permutation_type = FF_TRANSPOSE_IDCT_PERM;
3159             } else if (idct_algo == FF_IDCT_XVIDMMX) {
3160                 if (mm_flags & AV_CPU_FLAG_SSE2) {
3161                     c->idct_put              = ff_idct_xvid_sse2_put;
3162                     c->idct_add              = ff_idct_xvid_sse2_add;
3163                     c->idct                  = ff_idct_xvid_sse2;
3164                     c->idct_permutation_type = FF_SSE2_IDCT_PERM;
3165                 } else if (mm_flags & AV_CPU_FLAG_MMX2) {
3166                     c->idct_put              = ff_idct_xvid_mmx2_put;
3167                     c->idct_add              = ff_idct_xvid_mmx2_add;
3168                     c->idct                  = ff_idct_xvid_mmx2;
3169                 } else {
3170                     c->idct_put              = ff_idct_xvid_mmx_put;
3171                     c->idct_add              = ff_idct_xvid_mmx_add;
3172                     c->idct                  = ff_idct_xvid_mmx;
3173                 }
3174             }
3175         }
3176
3177         dsputil_init_mmx(c, avctx, mm_flags);
3178     }
3179
3180     if (mm_flags & AV_CPU_FLAG_MMX2)
3181         dsputil_init_mmx2(c, avctx, mm_flags);
3182
3183     if (mm_flags & AV_CPU_FLAG_3DNOW && HAVE_AMD3DNOW)
3184         dsputil_init_3dnow(c, avctx, mm_flags);
3185
3186     if (mm_flags & AV_CPU_FLAG_3DNOWEXT && HAVE_AMD3DNOWEXT)
3187         dsputil_init_3dnow2(c, avctx, mm_flags);
3188
3189     if (mm_flags & AV_CPU_FLAG_SSE && HAVE_SSE)
3190         dsputil_init_sse(c, avctx, mm_flags);
3191
3192     if (mm_flags & AV_CPU_FLAG_SSE2)
3193         dsputil_init_sse2(c, avctx, mm_flags);
3194
3195     if (mm_flags & AV_CPU_FLAG_SSSE3)
3196         dsputil_init_ssse3(c, avctx, mm_flags);
3197
3198     if (mm_flags & AV_CPU_FLAG_SSE4 && HAVE_SSE)
3199         dsputil_init_sse4(c, avctx, mm_flags);
3200
3201     if (mm_flags & AV_CPU_FLAG_AVX)
3202         dsputil_init_avx(c, avctx, mm_flags);
3203
3204     if (CONFIG_ENCODERS)
3205         ff_dsputilenc_init_mmx(c, avctx);
3206 }