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/attributes.h"
  26 #include "libavutil/cpu.h"
  27 #include "libavutil/x86/asm.h"
  28 #include "libavcodec/dsputil.h"
  29 #include "libavcodec/h264dsp.h"
  30 #include "libavcodec/mpegvideo.h"
  31 #include "libavcodec/simple_idct.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_pw_15)   =   0x000F000F000F000FULL;
  40 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_17)   = { 0x0011001100110011ULL, 0x0011001100110011ULL };
  41 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_20)   =   0x0014001400140014ULL;
  42 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_42)   =   0x002A002A002A002AULL;
  43 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_53)   =   0x0035003500350035ULL;
  44 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_96)   =   0x0060006000600060ULL;
  45 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_128)  =   0x0080008000800080ULL;
  46 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_255)  =   0x00ff00ff00ff00ffULL;
  47 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_512)  = { 0x0200020002000200ULL, 0x0200020002000200ULL };
  48 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_1019) = { 0x03FB03FB03FB03FBULL, 0x03FB03FB03FB03FBULL };
  49
  50 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_3F)   =   0x3F3F3F3F3F3F3F3FULL;
  51 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_FC)   =   0xFCFCFCFCFCFCFCFCULL;
  52
  53 DECLARE_ALIGNED(16, const double, ff_pd_1)[2] = { 1.0, 1.0 };
  54 DECLARE_ALIGNED(16, const double, ff_pd_2)[2] = { 2.0, 2.0 };
  55
  56
  57 #if HAVE_YASM
  58 void ff_put_pixels8_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
  59                               int dstStride, int src1Stride, int h);
  60 void ff_put_no_rnd_pixels8_l2_mmxext(uint8_t *dst, uint8_t *src1,
  61                                      uint8_t *src2, int dstStride,
  62                                      int src1Stride, int h);
  63 void ff_avg_pixels8_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
  64                               int dstStride, int src1Stride, int h);
  65 void ff_put_pixels16_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
  66                                int dstStride, int src1Stride, int h);
  67 void ff_avg_pixels16_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
  68                                int dstStride, int src1Stride, int h);
  69 void ff_put_no_rnd_pixels16_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
  70                                       int dstStride, int src1Stride, int h);
  71
  72 static void ff_put_pixels16_mmxext(uint8_t *block, const uint8_t *pixels,
  73                                    ptrdiff_t line_size, int h)
  74 {
  75     ff_put_pixels8_mmxext(block,     pixels,     line_size, h);
  76     ff_put_pixels8_mmxext(block + 8, pixels + 8, line_size, h);
  77 }
  78
  79 void ff_put_mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
  80                                          int dstStride, int srcStride, int h);
  81 void ff_avg_mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
  82                                          int dstStride, int srcStride, int h);
  83 void ff_put_no_rnd_mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
  84                                                  int dstStride, int srcStride,
  85                                                  int h);
  86 void ff_put_mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
  87                                         int dstStride, int srcStride, int h);
  88 void ff_avg_mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
  89                                         int dstStride, int srcStride, int h);
  90 void ff_put_no_rnd_mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
  91                                                 int dstStride, int srcStride,
  92                                                 int h);
  93 void ff_put_mpeg4_qpel16_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
  94                                          int dstStride, int srcStride);
  95 void ff_avg_mpeg4_qpel16_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
  96                                          int dstStride, int srcStride);
  97 void ff_put_no_rnd_mpeg4_qpel16_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
  98                                                  int dstStride, int srcStride);
  99 void ff_put_mpeg4_qpel8_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 100                                         int dstStride, int srcStride);
 101 void ff_avg_mpeg4_qpel8_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 102                                         int dstStride, int srcStride);
 103 void ff_put_no_rnd_mpeg4_qpel8_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 104                                                 int dstStride, int srcStride);
 105 #define ff_put_no_rnd_pixels16_mmxext ff_put_pixels16_mmxext
 106 #define ff_put_no_rnd_pixels8_mmxext ff_put_pixels8_mmxext
 107 #endif /* HAVE_YASM */
 108
 109
 110 #if HAVE_INLINE_ASM
 111
 112 #define JUMPALIGN()     __asm__ volatile (".p2align 3"::)
 113 #define MOVQ_ZERO(regd) __asm__ volatile ("pxor %%"#regd", %%"#regd ::)
 114
 115 #define MOVQ_BFE(regd)                                  \
 116     __asm__ volatile (                                  \
 117         "pcmpeqd %%"#regd", %%"#regd"   \n\t"           \
 118         "paddb   %%"#regd", %%"#regd"   \n\t" ::)
 119
 120 #ifndef PIC
 121 #define MOVQ_BONE(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_bone))
 122 #define MOVQ_WTWO(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_wtwo))
 123 #else
 124 // for shared library it's better to use this way for accessing constants
 125 // pcmpeqd -> -1
 126 #define MOVQ_BONE(regd)                                 \
 127     __asm__ volatile (                                  \
 128         "pcmpeqd  %%"#regd", %%"#regd"  \n\t"           \
 129         "psrlw          $15, %%"#regd"  \n\t"           \
 130         "packuswb %%"#regd", %%"#regd"  \n\t" ::)
 131
 132 #define MOVQ_WTWO(regd)                                 \
 133     __asm__ volatile (                                  \
 134         "pcmpeqd %%"#regd", %%"#regd"   \n\t"           \
 135         "psrlw         $15, %%"#regd"   \n\t"           \
 136         "psllw          $1, %%"#regd"   \n\t"::)
 137
 138 #endif
 139
 140 // using regr as temporary and for the output result
 141 // first argument is unmodifed and second is trashed
 142 // regfe is supposed to contain 0xfefefefefefefefe
 143 #define PAVGB_MMX(rega, regb, regr, regfe)                       \
 144     "movq   "#rega", "#regr"            \n\t"                    \
 145     "por    "#regb", "#regr"            \n\t"                    \
 146     "pxor   "#rega", "#regb"            \n\t"                    \
 147     "pand  "#regfe", "#regb"            \n\t"                    \
 148     "psrlq       $1, "#regb"            \n\t"                    \
 149     "psubb  "#regb", "#regr"            \n\t"
 150
 151 // mm6 is supposed to contain 0xfefefefefefefefe
 152 #define PAVGBP_MMX(rega, regb, regr, regc, regd, regp)           \
 153     "movq  "#rega", "#regr"             \n\t"                    \
 154     "movq  "#regc", "#regp"             \n\t"                    \
 155     "por   "#regb", "#regr"             \n\t"                    \
 156     "por   "#regd", "#regp"             \n\t"                    \
 157     "pxor  "#rega", "#regb"             \n\t"                    \
 158     "pxor  "#regc", "#regd"             \n\t"                    \
 159     "pand    %%mm6, "#regb"             \n\t"                    \
 160     "pand    %%mm6, "#regd"             \n\t"                    \
 161     "psrlq      $1, "#regd"             \n\t"                    \
 162     "psrlq      $1, "#regb"             \n\t"                    \
 163     "psubb "#regb", "#regr"             \n\t"                    \
 164     "psubb "#regd", "#regp"             \n\t"
 165
 166 /***********************************/
 167 /* MMX rounding */
 168
 169 #define DEF(x, y) x ## _ ## y ## _mmx
 170 #define SET_RND  MOVQ_WTWO
 171 #define PAVGBP(a, b, c, d, e, f)        PAVGBP_MMX(a, b, c, d, e, f)
 172 #define PAVGB(a, b, c, e)               PAVGB_MMX(a, b, c, e)
 173 #define OP_AVG(a, b, c, e)              PAVGB_MMX(a, b, c, e)
 174
 175 #include "dsputil_rnd_template.c"
 176
 177 #undef DEF
 178 #undef SET_RND
 179 #undef PAVGBP
 180 #undef PAVGB
 181 #undef OP_AVG
 182
 183 #endif /* HAVE_INLINE_ASM */
 184
 185
 186 #if HAVE_YASM
 187
 188 /***********************************/
 189 /* MMXEXT specific */
 190
 191 //FIXME the following could be optimized too ...
 192 static void ff_avg_pixels16_mmxext(uint8_t *block, const uint8_t *pixels,
 193                                    int line_size, int h)
 194 {
 195     ff_avg_pixels8_mmxext(block,     pixels,     line_size, h);
 196     ff_avg_pixels8_mmxext(block + 8, pixels + 8, line_size, h);
 197 }
 198
 199 #endif /* HAVE_YASM */
 200
 201
 202 #if HAVE_INLINE_ASM
 203 /***********************************/
 204 /* standard MMX */
 205
 206 void ff_put_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
 207                                int line_size)
 208 {
 209     const int16_t *p;
 210     uint8_t *pix;
 211
 212     /* read the pixels */
 213     p   = block;
 214     pix = pixels;
 215     /* unrolled loop */
 216     __asm__ volatile (
 217         "movq      (%3), %%mm0          \n\t"
 218         "movq     8(%3), %%mm1          \n\t"
 219         "movq    16(%3), %%mm2          \n\t"
 220         "movq    24(%3), %%mm3          \n\t"
 221         "movq    32(%3), %%mm4          \n\t"
 222         "movq    40(%3), %%mm5          \n\t"
 223         "movq    48(%3), %%mm6          \n\t"
 224         "movq    56(%3), %%mm7          \n\t"
 225         "packuswb %%mm1, %%mm0          \n\t"
 226         "packuswb %%mm3, %%mm2          \n\t"
 227         "packuswb %%mm5, %%mm4          \n\t"
 228         "packuswb %%mm7, %%mm6          \n\t"
 229         "movq     %%mm0, (%0)           \n\t"
 230         "movq     %%mm2, (%0, %1)       \n\t"
 231         "movq     %%mm4, (%0, %1, 2)    \n\t"
 232         "movq     %%mm6, (%0, %2)       \n\t"
 233         :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3),
 234            "r"(p)
 235         : "memory");
 236     pix += line_size * 4;
 237     p   += 32;
 238
 239     // if here would be an exact copy of the code above
 240     // compiler would generate some very strange code
 241     // thus using "r"
 242     __asm__ volatile (
 243         "movq       (%3), %%mm0         \n\t"
 244         "movq      8(%3), %%mm1         \n\t"
 245         "movq     16(%3), %%mm2         \n\t"
 246         "movq     24(%3), %%mm3         \n\t"
 247         "movq     32(%3), %%mm4         \n\t"
 248         "movq     40(%3), %%mm5         \n\t"
 249         "movq     48(%3), %%mm6         \n\t"
 250         "movq     56(%3), %%mm7         \n\t"
 251         "packuswb  %%mm1, %%mm0         \n\t"
 252         "packuswb  %%mm3, %%mm2         \n\t"
 253         "packuswb  %%mm5, %%mm4         \n\t"
 254         "packuswb  %%mm7, %%mm6         \n\t"
 255         "movq      %%mm0, (%0)          \n\t"
 256         "movq      %%mm2, (%0, %1)      \n\t"
 257         "movq      %%mm4, (%0, %1, 2)   \n\t"
 258         "movq      %%mm6, (%0, %2)      \n\t"
 259         :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3), "r"(p)
 260         : "memory");
 261 }
 262
 263 #define put_signed_pixels_clamped_mmx_half(off)             \
 264     "movq          "#off"(%2), %%mm1        \n\t"           \
 265     "movq     16 + "#off"(%2), %%mm2        \n\t"           \
 266     "movq     32 + "#off"(%2), %%mm3        \n\t"           \
 267     "movq     48 + "#off"(%2), %%mm4        \n\t"           \
 268     "packsswb  8 + "#off"(%2), %%mm1        \n\t"           \
 269     "packsswb 24 + "#off"(%2), %%mm2        \n\t"           \
 270     "packsswb 40 + "#off"(%2), %%mm3        \n\t"           \
 271     "packsswb 56 + "#off"(%2), %%mm4        \n\t"           \
 272     "paddb              %%mm0, %%mm1        \n\t"           \
 273     "paddb              %%mm0, %%mm2        \n\t"           \
 274     "paddb              %%mm0, %%mm3        \n\t"           \
 275     "paddb              %%mm0, %%mm4        \n\t"           \
 276     "movq               %%mm1, (%0)         \n\t"           \
 277     "movq               %%mm2, (%0, %3)     \n\t"           \
 278     "movq               %%mm3, (%0, %3, 2)  \n\t"           \
 279     "movq               %%mm4, (%0, %1)     \n\t"
 280
 281 void ff_put_signed_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
 282                                       int line_size)
 283 {
 284     x86_reg line_skip = line_size;
 285     x86_reg line_skip3;
 286
 287     __asm__ volatile (
 288         "movq "MANGLE(ff_pb_80)", %%mm0     \n\t"
 289         "lea         (%3, %3, 2), %1        \n\t"
 290         put_signed_pixels_clamped_mmx_half(0)
 291         "lea         (%0, %3, 4), %0        \n\t"
 292         put_signed_pixels_clamped_mmx_half(64)
 293         : "+&r"(pixels), "=&r"(line_skip3)
 294         : "r"(block), "r"(line_skip)
 295         : "memory");
 296 }
 297
 298 void ff_add_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
 299                                int line_size)
 300 {
 301     const int16_t *p;
 302     uint8_t *pix;
 303     int i;
 304
 305     /* read the pixels */
 306     p   = block;
 307     pix = pixels;
 308     MOVQ_ZERO(mm7);
 309     i = 4;
 310     do {
 311         __asm__ volatile (
 312             "movq        (%2), %%mm0    \n\t"
 313             "movq       8(%2), %%mm1    \n\t"
 314             "movq      16(%2), %%mm2    \n\t"
 315             "movq      24(%2), %%mm3    \n\t"
 316             "movq          %0, %%mm4    \n\t"
 317             "movq          %1, %%mm6    \n\t"
 318             "movq       %%mm4, %%mm5    \n\t"
 319             "punpcklbw  %%mm7, %%mm4    \n\t"
 320             "punpckhbw  %%mm7, %%mm5    \n\t"
 321             "paddsw     %%mm4, %%mm0    \n\t"
 322             "paddsw     %%mm5, %%mm1    \n\t"
 323             "movq       %%mm6, %%mm5    \n\t"
 324             "punpcklbw  %%mm7, %%mm6    \n\t"
 325             "punpckhbw  %%mm7, %%mm5    \n\t"
 326             "paddsw     %%mm6, %%mm2    \n\t"
 327             "paddsw     %%mm5, %%mm3    \n\t"
 328             "packuswb   %%mm1, %%mm0    \n\t"
 329             "packuswb   %%mm3, %%mm2    \n\t"
 330             "movq       %%mm0, %0       \n\t"
 331             "movq       %%mm2, %1       \n\t"
 332             : "+m"(*pix), "+m"(*(pix + line_size))
 333             : "r"(p)
 334             : "memory");
 335         pix += line_size * 2;
 336         p   += 16;
 337     } while (--i);
 338 }
 339
 340 static void put_pixels8_mmx(uint8_t *block, const uint8_t *pixels,
 341                             ptrdiff_t line_size, int h)
 342 {
 343     __asm__ volatile (
 344         "lea   (%3, %3), %%"REG_a"      \n\t"
 345         ".p2align     3                 \n\t"
 346         "1:                             \n\t"
 347         "movq  (%1    ), %%mm0          \n\t"
 348         "movq  (%1, %3), %%mm1          \n\t"
 349         "movq     %%mm0, (%2)           \n\t"
 350         "movq     %%mm1, (%2, %3)       \n\t"
 351         "add  %%"REG_a", %1             \n\t"
 352         "add  %%"REG_a", %2             \n\t"
 353         "movq  (%1    ), %%mm0          \n\t"
 354         "movq  (%1, %3), %%mm1          \n\t"
 355         "movq     %%mm0, (%2)           \n\t"
 356         "movq     %%mm1, (%2, %3)       \n\t"
 357         "add  %%"REG_a", %1             \n\t"
 358         "add  %%"REG_a", %2             \n\t"
 359         "subl        $4, %0             \n\t"
 360         "jnz         1b                 \n\t"
 361         : "+g"(h), "+r"(pixels),  "+r"(block)
 362         : "r"((x86_reg)line_size)
 363         : "%"REG_a, "memory"
 364         );
 365 }
 366
 367 static void put_pixels16_mmx(uint8_t *block, const uint8_t *pixels,
 368                              ptrdiff_t line_size, int h)
 369 {
 370     __asm__ volatile (
 371         "lea   (%3, %3), %%"REG_a"      \n\t"
 372         ".p2align     3                 \n\t"
 373         "1:                             \n\t"
 374         "movq  (%1    ), %%mm0          \n\t"
 375         "movq 8(%1    ), %%mm4          \n\t"
 376         "movq  (%1, %3), %%mm1          \n\t"
 377         "movq 8(%1, %3), %%mm5          \n\t"
 378         "movq     %%mm0,  (%2)          \n\t"
 379         "movq     %%mm4, 8(%2)          \n\t"
 380         "movq     %%mm1,  (%2, %3)      \n\t"
 381         "movq     %%mm5, 8(%2, %3)      \n\t"
 382         "add  %%"REG_a", %1             \n\t"
 383         "add  %%"REG_a", %2             \n\t"
 384         "movq  (%1    ), %%mm0          \n\t"
 385         "movq 8(%1    ), %%mm4          \n\t"
 386         "movq  (%1, %3), %%mm1          \n\t"
 387         "movq 8(%1, %3), %%mm5          \n\t"
 388         "movq     %%mm0,  (%2)          \n\t"
 389         "movq     %%mm4, 8(%2)          \n\t"
 390         "movq     %%mm1,  (%2, %3)      \n\t"
 391         "movq     %%mm5, 8(%2, %3)      \n\t"
 392         "add  %%"REG_a", %1             \n\t"
 393         "add  %%"REG_a", %2             \n\t"
 394         "subl        $4, %0             \n\t"
 395         "jnz         1b                 \n\t"
 396         : "+g"(h), "+r"(pixels),  "+r"(block)
 397         : "r"((x86_reg)line_size)
 398         : "%"REG_a, "memory"
 399         );
 400 }
 401
 402 #define CLEAR_BLOCKS(name, n)                           \
 403 static void name(int16_t *blocks)                       \
 404 {                                                       \
 405     __asm__ volatile (                                  \
 406         "pxor %%mm7, %%mm7              \n\t"           \
 407         "mov     %1,        %%"REG_a"   \n\t"           \
 408         "1:                             \n\t"           \
 409         "movq %%mm7,   (%0, %%"REG_a")  \n\t"           \
 410         "movq %%mm7,  8(%0, %%"REG_a")  \n\t"           \
 411         "movq %%mm7, 16(%0, %%"REG_a")  \n\t"           \
 412         "movq %%mm7, 24(%0, %%"REG_a")  \n\t"           \
 413         "add    $32, %%"REG_a"          \n\t"           \
 414         "js      1b                     \n\t"           \
 415         :: "r"(((uint8_t *)blocks) + 128 * n),          \
 416            "i"(-128 * n)                                \
 417         : "%"REG_a                                      \
 418         );                                              \
 419 }
 420 CLEAR_BLOCKS(clear_blocks_mmx, 6)
 421 CLEAR_BLOCKS(clear_block_mmx, 1)
 422
 423 static void clear_block_sse(int16_t *block)
 424 {
 425     __asm__ volatile (
 426         "xorps  %%xmm0, %%xmm0          \n"
 427         "movaps %%xmm0,    (%0)         \n"
 428         "movaps %%xmm0,  16(%0)         \n"
 429         "movaps %%xmm0,  32(%0)         \n"
 430         "movaps %%xmm0,  48(%0)         \n"
 431         "movaps %%xmm0,  64(%0)         \n"
 432         "movaps %%xmm0,  80(%0)         \n"
 433         "movaps %%xmm0,  96(%0)         \n"
 434         "movaps %%xmm0, 112(%0)         \n"
 435         :: "r"(block)
 436         : "memory"
 437     );
 438 }
 439
 440 static void clear_blocks_sse(int16_t *blocks)
 441 {
 442     __asm__ volatile (
 443         "xorps  %%xmm0, %%xmm0              \n"
 444         "mov        %1,         %%"REG_a"   \n"
 445         "1:                                 \n"
 446         "movaps %%xmm0,    (%0, %%"REG_a")  \n"
 447         "movaps %%xmm0,  16(%0, %%"REG_a")  \n"
 448         "movaps %%xmm0,  32(%0, %%"REG_a")  \n"
 449         "movaps %%xmm0,  48(%0, %%"REG_a")  \n"
 450         "movaps %%xmm0,  64(%0, %%"REG_a")  \n"
 451         "movaps %%xmm0,  80(%0, %%"REG_a")  \n"
 452         "movaps %%xmm0,  96(%0, %%"REG_a")  \n"
 453         "movaps %%xmm0, 112(%0, %%"REG_a")  \n"
 454         "add      $128,         %%"REG_a"   \n"
 455         "js         1b                      \n"
 456         :: "r"(((uint8_t *)blocks) + 128 * 6),
 457            "i"(-128 * 6)
 458         : "%"REG_a
 459     );
 460 }
 461
 462 static void add_bytes_mmx(uint8_t *dst, uint8_t *src, int w)
 463 {
 464     x86_reg i = 0;
 465     __asm__ volatile (
 466         "jmp          2f                \n\t"
 467         "1:                             \n\t"
 468         "movq   (%1, %0), %%mm0         \n\t"
 469         "movq   (%2, %0), %%mm1         \n\t"
 470         "paddb     %%mm0, %%mm1         \n\t"
 471         "movq      %%mm1, (%2, %0)      \n\t"
 472         "movq  8(%1, %0), %%mm0         \n\t"
 473         "movq  8(%2, %0), %%mm1         \n\t"
 474         "paddb     %%mm0, %%mm1         \n\t"
 475         "movq      %%mm1, 8(%2, %0)     \n\t"
 476         "add         $16, %0            \n\t"
 477         "2:                             \n\t"
 478         "cmp          %3, %0            \n\t"
 479         "js           1b                \n\t"
 480         : "+r"(i)
 481         : "r"(src), "r"(dst), "r"((x86_reg)w - 15)
 482     );
 483     for ( ; i < w; i++)
 484         dst[i + 0] += src[i + 0];
 485 }
 486
 487 #if HAVE_7REGS
 488 static void add_hfyu_median_prediction_cmov(uint8_t *dst, const uint8_t *top,
 489                                             const uint8_t *diff, int w,
 490                                             int *left, int *left_top)
 491 {
 492     x86_reg w2 = -w;
 493     x86_reg x;
 494     int l  = *left     & 0xff;
 495     int tl = *left_top & 0xff;
 496     int t;
 497     __asm__ volatile (
 498         "mov          %7, %3            \n"
 499         "1:                             \n"
 500         "movzbl (%3, %4), %2            \n"
 501         "mov          %2, %k3           \n"
 502         "sub         %b1, %b3           \n"
 503         "add         %b0, %b3           \n"
 504         "mov          %2, %1            \n"
 505         "cmp          %0, %2            \n"
 506         "cmovg        %0, %2            \n"
 507         "cmovg        %1, %0            \n"
 508         "cmp         %k3, %0            \n"
 509         "cmovg       %k3, %0            \n"
 510         "mov          %7, %3            \n"
 511         "cmp          %2, %0            \n"
 512         "cmovl        %2, %0            \n"
 513         "add    (%6, %4), %b0           \n"
 514         "mov         %b0, (%5, %4)      \n"
 515         "inc          %4                \n"
 516         "jl           1b                \n"
 517         : "+&q"(l), "+&q"(tl), "=&r"(t), "=&q"(x), "+&r"(w2)
 518         : "r"(dst + w), "r"(diff + w), "rm"(top + w)
 519     );
 520     *left     = l;
 521     *left_top = tl;
 522 }
 523 #endif
 524
 525 /* Draw the edges of width 'w' of an image of size width, height
 526  * this MMX version can only handle w == 8 || w == 16. */
 527 static void draw_edges_mmx(uint8_t *buf, int wrap, int width, int height,
 528                            int w, int h, int sides)
 529 {
 530     uint8_t *ptr, *last_line;
 531     int i;
 532
 533     last_line = buf + (height - 1) * wrap;
 534     /* left and right */
 535     ptr = buf;
 536     if (w == 8) {
 537         __asm__ volatile (
 538             "1:                             \n\t"
 539             "movd            (%0), %%mm0    \n\t"
 540             "punpcklbw      %%mm0, %%mm0    \n\t"
 541             "punpcklwd      %%mm0, %%mm0    \n\t"
 542             "punpckldq      %%mm0, %%mm0    \n\t"
 543             "movq           %%mm0, -8(%0)   \n\t"
 544             "movq      -8(%0, %2), %%mm1    \n\t"
 545             "punpckhbw      %%mm1, %%mm1    \n\t"
 546             "punpckhwd      %%mm1, %%mm1    \n\t"
 547             "punpckhdq      %%mm1, %%mm1    \n\t"
 548             "movq           %%mm1, (%0, %2) \n\t"
 549             "add               %1, %0       \n\t"
 550             "cmp               %3, %0       \n\t"
 551             "jb                1b           \n\t"
 552             : "+r"(ptr)
 553             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
 554             );
 555     } else {
 556         __asm__ volatile (
 557             "1:                                 \n\t"
 558             "movd            (%0), %%mm0        \n\t"
 559             "punpcklbw      %%mm0, %%mm0        \n\t"
 560             "punpcklwd      %%mm0, %%mm0        \n\t"
 561             "punpckldq      %%mm0, %%mm0        \n\t"
 562             "movq           %%mm0, -8(%0)       \n\t"
 563             "movq           %%mm0, -16(%0)      \n\t"
 564             "movq      -8(%0, %2), %%mm1        \n\t"
 565             "punpckhbw      %%mm1, %%mm1        \n\t"
 566             "punpckhwd      %%mm1, %%mm1        \n\t"
 567             "punpckhdq      %%mm1, %%mm1        \n\t"
 568             "movq           %%mm1,  (%0, %2)    \n\t"
 569             "movq           %%mm1, 8(%0, %2)    \n\t"
 570             "add               %1, %0           \n\t"
 571             "cmp               %3, %0           \n\t"
 572             "jb                1b               \n\t"
 573             : "+r"(ptr)
 574             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
 575             );
 576     }
 577
 578     /* top and bottom (and hopefully also the corners) */
 579     if (sides & EDGE_TOP) {
 580         for (i = 0; i < h; i += 4) {
 581             ptr = buf - (i + 1) * wrap - w;
 582             __asm__ volatile (
 583                 "1:                             \n\t"
 584                 "movq (%1, %0), %%mm0           \n\t"
 585                 "movq    %%mm0, (%0)            \n\t"
 586                 "movq    %%mm0, (%0, %2)        \n\t"
 587                 "movq    %%mm0, (%0, %2, 2)     \n\t"
 588                 "movq    %%mm0, (%0, %3)        \n\t"
 589                 "add        $8, %0              \n\t"
 590                 "cmp        %4, %0              \n\t"
 591                 "jb         1b                  \n\t"
 592                 : "+r"(ptr)
 593                 : "r"((x86_reg)buf - (x86_reg)ptr - w), "r"((x86_reg) -wrap),
 594                   "r"((x86_reg) -wrap * 3), "r"(ptr + width + 2 * w)
 595                 );
 596         }
 597     }
 598
 599     if (sides & EDGE_BOTTOM) {
 600         for (i = 0; i < h; i += 4) {
 601             ptr = last_line + (i + 1) * wrap - w;
 602             __asm__ volatile (
 603                 "1:                             \n\t"
 604                 "movq (%1, %0), %%mm0           \n\t"
 605                 "movq    %%mm0, (%0)            \n\t"
 606                 "movq    %%mm0, (%0, %2)        \n\t"
 607                 "movq    %%mm0, (%0, %2, 2)     \n\t"
 608                 "movq    %%mm0, (%0, %3)        \n\t"
 609                 "add        $8, %0              \n\t"
 610                 "cmp        %4, %0              \n\t"
 611                 "jb         1b                  \n\t"
 612                 : "+r"(ptr)
 613                 : "r"((x86_reg)last_line - (x86_reg)ptr - w),
 614                   "r"((x86_reg)wrap), "r"((x86_reg)wrap * 3),
 615                   "r"(ptr + width + 2 * w)
 616                 );
 617         }
 618     }
 619 }
 620 #endif /* HAVE_INLINE_ASM */
 621
 622
 623 #if HAVE_YASM
 624 #define QPEL_OP(OPNAME, ROUNDER, RND, MMX)                              \
 625 static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src,   \
 626                                           ptrdiff_t stride)             \
 627 {                                                                       \
 628     ff_ ## OPNAME ## pixels8_ ## MMX(dst, src, stride, 8);              \
 629 }                                                                       \
 630                                                                         \
 631 static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src,    \
 632                                          ptrdiff_t stride)              \
 633 {                                                                       \
 634     uint64_t temp[8];                                                   \
 635     uint8_t * const half = (uint8_t*)temp;                              \
 636     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8,        \
 637                                                    stride, 8);          \
 638     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src, half,                 \
 639                                         stride, stride, 8);             \
 640 }                                                                       \
 641                                                                         \
 642 static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src,    \
 643                                          ptrdiff_t stride)              \
 644 {                                                                       \
 645     ff_ ## OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride,    \
 646                                                    stride, 8);          \
 647 }                                                                       \
 648                                                                         \
 649 static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src,    \
 650                                          ptrdiff_t stride)              \
 651 {                                                                       \
 652     uint64_t temp[8];                                                   \
 653     uint8_t * const half = (uint8_t*)temp;                              \
 654     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8,        \
 655                                                    stride, 8);          \
 656     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src + 1, half, stride,     \
 657                                         stride, 8);                     \
 658 }                                                                       \
 659                                                                         \
 660 static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src,    \
 661                                          ptrdiff_t stride)              \
 662 {                                                                       \
 663     uint64_t temp[8];                                                   \
 664     uint8_t * const half = (uint8_t*)temp;                              \
 665     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src,           \
 666                                                    8, stride);          \
 667     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src, half,                 \
 668                                         stride, stride, 8);             \
 669 }                                                                       \
 670                                                                         \
 671 static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src,    \
 672                                          ptrdiff_t stride)              \
 673 {                                                                       \
 674     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src,            \
 675                                                    stride, stride);     \
 676 }                                                                       \
 677                                                                         \
 678 static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src,    \
 679                                          ptrdiff_t stride)              \
 680 {                                                                       \
 681     uint64_t temp[8];                                                   \
 682     uint8_t * const half = (uint8_t*)temp;                              \
 683     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src,           \
 684                                                    8, stride);          \
 685     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src + stride, half, stride,\
 686                                         stride, 8);                     \
 687 }                                                                       \
 688                                                                         \
 689 static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src,    \
 690                                          ptrdiff_t stride)              \
 691 {                                                                       \
 692     uint64_t half[8 + 9];                                               \
 693     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
 694     uint8_t * const halfHV = ((uint8_t*)half);                          \
 695     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 696                                                    stride, 9);          \
 697     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8,           \
 698                                         stride, 9);                     \
 699     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
 700     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV,             \
 701                                         stride, 8, 8);                  \
 702 }                                                                       \
 703                                                                         \
 704 static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src,    \
 705                                          ptrdiff_t stride)              \
 706 {                                                                       \
 707     uint64_t half[8 + 9];                                               \
 708     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
 709     uint8_t * const halfHV = ((uint8_t*)half);                          \
 710     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 711                                                    stride, 9);          \
 712     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,       \
 713                                         stride, 9);                     \
 714     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
 715     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV,             \
 716                                         stride, 8, 8);                  \
 717 }                                                                       \
 718                                                                         \
 719 static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src,    \
 720                                          ptrdiff_t stride)              \
 721 {                                                                       \
 722     uint64_t half[8 + 9];                                               \
 723     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
 724     uint8_t * const halfHV = ((uint8_t*)half);                          \
 725     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 726                                                    stride, 9);          \
 727     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8,           \
 728                                         stride, 9);                     \
 729     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
 730     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV,         \
 731                                         stride, 8, 8);                  \
 732 }                                                                       \
 733                                                                         \
 734 static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src,    \
 735                                          ptrdiff_t stride)              \
 736 {                                                                       \
 737     uint64_t half[8 + 9];                                               \
 738     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
 739     uint8_t * const halfHV = ((uint8_t*)half);                          \
 740     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 741                                                    stride, 9);          \
 742     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,       \
 743                                         stride, 9);                     \
 744     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
 745     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV,         \
 746                                         stride, 8, 8);                  \
 747 }                                                                       \
 748                                                                         \
 749 static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src,    \
 750                                          ptrdiff_t stride)              \
 751 {                                                                       \
 752     uint64_t half[8 + 9];                                               \
 753     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
 754     uint8_t * const halfHV = ((uint8_t*)half);                          \
 755     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 756                                                    stride, 9);          \
 757     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
 758     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV,             \
 759                                         stride, 8, 8);                  \
 760 }                                                                       \
 761                                                                         \
 762 static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src,    \
 763                                          ptrdiff_t stride)              \
 764 {                                                                       \
 765     uint64_t half[8 + 9];                                               \
 766     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
 767     uint8_t * const halfHV = ((uint8_t*)half);                          \
 768     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 769                                                    stride, 9);          \
 770     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
 771     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV,         \
 772                                         stride, 8, 8);                  \
 773 }                                                                       \
 774                                                                         \
 775 static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src,    \
 776                                          ptrdiff_t stride)              \
 777 {                                                                       \
 778     uint64_t half[8 + 9];                                               \
 779     uint8_t * const halfH = ((uint8_t*)half);                           \
 780     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 781                                                    stride, 9);          \
 782     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH,              \
 783                                         8, stride, 9);                  \
 784     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH,          \
 785                                                    stride, 8);          \
 786 }                                                                       \
 787                                                                         \
 788 static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src,    \
 789                                          ptrdiff_t stride)              \
 790 {                                                                       \
 791     uint64_t half[8 + 9];                                               \
 792     uint8_t * const halfH = ((uint8_t*)half);                           \
 793     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 794                                                    stride, 9);          \
 795     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,       \
 796                                         stride, 9);                     \
 797     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH,          \
 798                                                    stride, 8);          \
 799 }                                                                       \
 800                                                                         \
 801 static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src,    \
 802                                          ptrdiff_t stride)              \
 803 {                                                                       \
 804     uint64_t half[9];                                                   \
 805     uint8_t * const halfH = ((uint8_t*)half);                           \
 806     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 807                                                    stride, 9);          \
 808     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH,          \
 809                                                    stride, 8);          \
 810 }                                                                       \
 811                                                                         \
 812 static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src,  \
 813                                            ptrdiff_t stride)            \
 814 {                                                                       \
 815     ff_ ## OPNAME ## pixels16_ ## MMX(dst, src, stride, 16);            \
 816 }                                                                       \
 817                                                                         \
 818 static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src,   \
 819                                           ptrdiff_t stride)             \
 820 {                                                                       \
 821     uint64_t temp[32];                                                  \
 822     uint8_t * const half = (uint8_t*)temp;                              \
 823     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16,      \
 824                                                     stride, 16);        \
 825     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride,        \
 826                                          stride, 16);                   \
 827 }                                                                       \
 828                                                                         \
 829 static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src,   \
 830                                           ptrdiff_t stride)             \
 831 {                                                                       \
 832     ff_ ## OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src,           \
 833                                                     stride, stride, 16);\
 834 }                                                                       \
 835                                                                         \
 836 static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src,   \
 837                                           ptrdiff_t stride)             \
 838 {                                                                       \
 839     uint64_t temp[32];                                                  \
 840     uint8_t * const half = (uint8_t*)temp;                              \
 841     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16,      \
 842                                                     stride, 16);        \
 843     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src + 1, half,            \
 844                                          stride, stride, 16);           \
 845 }                                                                       \
 846                                                                         \
 847 static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src,   \
 848                                           ptrdiff_t stride)             \
 849 {                                                                       \
 850     uint64_t temp[32];                                                  \
 851     uint8_t * const half = (uint8_t*)temp;                              \
 852     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16,      \
 853                                                     stride);            \
 854     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride,        \
 855                                          stride, 16);                   \
 856 }                                                                       \
 857                                                                         \
 858 static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src,   \
 859                                           ptrdiff_t stride)             \
 860 {                                                                       \
 861     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src,           \
 862                                                     stride, stride);    \
 863 }                                                                       \
 864                                                                         \
 865 static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src,   \
 866                                           ptrdiff_t stride)             \
 867 {                                                                       \
 868     uint64_t temp[32];                                                  \
 869     uint8_t * const half = (uint8_t*)temp;                              \
 870     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16,      \
 871                                                     stride);            \
 872     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half,         \
 873                                          stride, stride, 16);           \
 874 }                                                                       \
 875                                                                         \
 876 static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src,   \
 877                                           ptrdiff_t stride)             \
 878 {                                                                       \
 879     uint64_t half[16 * 2 + 17 * 2];                                     \
 880     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
 881     uint8_t * const halfHV = ((uint8_t*)half);                          \
 882     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
 883                                                     stride, 17);        \
 884     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,         \
 885                                          stride, 17);                   \
 886     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
 887                                                     16, 16);            \
 888     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV,            \
 889                                          stride, 16, 16);               \
 890 }                                                                       \
 891                                                                         \
 892 static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src,   \
 893                                           ptrdiff_t stride)             \
 894 {                                                                       \
 895     uint64_t half[16 * 2 + 17 * 2];                                     \
 896     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
 897     uint8_t * const halfHV = ((uint8_t*)half);                          \
 898     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
 899                                                     stride, 17);        \
 900     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,     \
 901                                          stride, 17);                   \
 902     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
 903                                                     16, 16);            \
 904     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV,            \
 905                                          stride, 16, 16);               \
 906 }                                                                       \
 907                                                                         \
 908 static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src,   \
 909                                           ptrdiff_t stride)             \
 910 {                                                                       \
 911     uint64_t half[16 * 2 + 17 * 2];                                     \
 912     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
 913     uint8_t * const halfHV = ((uint8_t*)half);                          \
 914     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
 915                                                     stride, 17);        \
 916     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,         \
 917                                          stride, 17);                   \
 918     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
 919                                                     16, 16);            \
 920     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV,       \
 921                                          stride, 16, 16);               \
 922 }                                                                       \
 923                                                                         \
 924 static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src,   \
 925                                           ptrdiff_t stride)             \
 926 {                                                                       \
 927     uint64_t half[16 * 2 + 17 * 2];                                     \
 928     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
 929     uint8_t * const halfHV = ((uint8_t*)half);                          \
 930     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
 931                                                     stride, 17);        \
 932     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,     \
 933                                          stride, 17);                   \
 934     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
 935                                                     16, 16);            \
 936     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV,       \
 937                                          stride, 16, 16);               \
 938 }                                                                       \
 939                                                                         \
 940 static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src,   \
 941                                           ptrdiff_t stride)             \
 942 {                                                                       \
 943     uint64_t half[16 * 2 + 17 * 2];                                     \
 944     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
 945     uint8_t * const halfHV = ((uint8_t*)half);                          \
 946     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
 947                                                     stride, 17);        \
 948     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
 949                                                     16, 16);            \
 950     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV,            \
 951                                          stride, 16, 16);               \
 952 }                                                                       \
 953                                                                         \
 954 static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src,   \
 955                                           ptrdiff_t stride)             \
 956 {                                                                       \
 957     uint64_t half[16 * 2 + 17 * 2];                                     \
 958     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
 959     uint8_t * const halfHV = ((uint8_t*)half);                          \
 960     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
 961                                                     stride, 17);        \
 962     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
 963                                                     16, 16);            \
 964     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV,       \
 965                                          stride, 16, 16);               \
 966 }                                                                       \
 967                                                                         \
 968 static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src,   \
 969                                           ptrdiff_t stride)             \
 970 {                                                                       \
 971     uint64_t half[17 * 2];                                              \
 972     uint8_t * const halfH = ((uint8_t*)half);                           \
 973     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
 974                                                     stride, 17);        \
 975     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,         \
 976                                          stride, 17);                   \
 977     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH,         \
 978                                                     stride, 16);        \
 979 }                                                                       \
 980                                                                         \
 981 static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src,   \
 982                                           ptrdiff_t stride)             \
 983 {                                                                       \
 984     uint64_t half[17 * 2];                                              \
 985     uint8_t * const halfH = ((uint8_t*)half);                           \
 986     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
 987                                                     stride, 17);        \
 988     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,     \
 989                                          stride, 17);                   \
 990     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH,         \
 991                                                     stride, 16);        \
 992 }                                                                       \
 993                                                                         \
 994 static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src,   \
 995                                           ptrdiff_t stride)             \
 996 {                                                                       \
 997     uint64_t half[17 * 2];                                              \
 998     uint8_t * const halfH = ((uint8_t*)half);                           \
 999     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1000                                                     stride, 17);        \
1001     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH,         \
1002                                                     stride, 16);        \
1003 }
1004
1005 QPEL_OP(put_,          ff_pw_16, _,        mmxext)
1006 QPEL_OP(avg_,          ff_pw_16, _,        mmxext)
1007 QPEL_OP(put_no_rnd_,   ff_pw_15, _no_rnd_, mmxext)
1008 #endif /* HAVE_YASM */
1009
1010
1011 #if HAVE_INLINE_ASM
1012 void ff_put_rv40_qpel8_mc33_mmx(uint8_t *dst, uint8_t *src, ptrdiff_t stride)
1013 {
1014   put_pixels8_xy2_mmx(dst, src, stride, 8);
1015 }
1016 void ff_put_rv40_qpel16_mc33_mmx(uint8_t *dst, uint8_t *src, ptrdiff_t stride)
1017 {
1018   put_pixels16_xy2_mmx(dst, src, stride, 16);
1019 }
1020 void ff_avg_rv40_qpel8_mc33_mmx(uint8_t *dst, uint8_t *src, ptrdiff_t stride)
1021 {
1022   avg_pixels8_xy2_mmx(dst, src, stride, 8);
1023 }
1024 void ff_avg_rv40_qpel16_mc33_mmx(uint8_t *dst, uint8_t *src, ptrdiff_t stride)
1025 {
1026   avg_pixels16_xy2_mmx(dst, src, stride, 16);
1027 }
1028
1029 static void gmc_mmx(uint8_t *dst, uint8_t *src,
1030                     int stride, int h, int ox, int oy,
1031                     int dxx, int dxy, int dyx, int dyy,
1032                     int shift, int r, int width, int height)
1033 {
1034     const int w    = 8;
1035     const int ix   = ox  >> (16 + shift);
1036     const int iy   = oy  >> (16 + shift);
1037     const int oxs  = ox  >> 4;
1038     const int oys  = oy  >> 4;
1039     const int dxxs = dxx >> 4;
1040     const int dxys = dxy >> 4;
1041     const int dyxs = dyx >> 4;
1042     const int dyys = dyy >> 4;
1043     const uint16_t r4[4]   = { r, r, r, r };
1044     const uint16_t dxy4[4] = { dxys, dxys, dxys, dxys };
1045     const uint16_t dyy4[4] = { dyys, dyys, dyys, dyys };
1046     const uint64_t shift2 = 2 * shift;
1047     int x, y;
1048
1049     const int dxw = (dxx - (1 << (16 + shift))) * (w - 1);
1050     const int dyh = (dyy - (1 << (16 + shift))) * (h - 1);
1051     const int dxh = dxy * (h - 1);
1052     const int dyw = dyx * (w - 1);
1053     if ( // non-constant fullpel offset (3% of blocks)
1054         ((ox ^ (ox + dxw)) | (ox ^ (ox + dxh)) | (ox ^ (ox + dxw + dxh)) |
1055          (oy ^ (oy + dyw)) | (oy ^ (oy + dyh)) | (oy ^ (oy + dyw + dyh))) >> (16 + shift)
1056         // uses more than 16 bits of subpel mv (only at huge resolution)
1057         || (dxx | dxy | dyx | dyy) & 15 ||
1058         (unsigned)ix >= width  - w ||
1059         (unsigned)iy >= height - h) {
1060         // FIXME could still use mmx for some of the rows
1061         ff_gmc_c(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy,
1062                  shift, r, width, height);
1063         return;
1064     }
1065
1066     src += ix + iy * stride;
1067
1068     __asm__ volatile (
1069         "movd         %0, %%mm6         \n\t"
1070         "pxor      %%mm7, %%mm7         \n\t"
1071         "punpcklwd %%mm6, %%mm6         \n\t"
1072         "punpcklwd %%mm6, %%mm6         \n\t"
1073         :: "r"(1<<shift)
1074     );
1075
1076     for (x = 0; x < w; x += 4) {
1077         uint16_t dx4[4] = { oxs - dxys + dxxs * (x + 0),
1078                             oxs - dxys + dxxs * (x + 1),
1079                             oxs - dxys + dxxs * (x + 2),
1080                             oxs - dxys + dxxs * (x + 3) };
1081         uint16_t dy4[4] = { oys - dyys + dyxs * (x + 0),
1082                             oys - dyys + dyxs * (x + 1),
1083                             oys - dyys + dyxs * (x + 2),
1084                             oys - dyys + dyxs * (x + 3) };
1085
1086         for (y = 0; y < h; y++) {
1087             __asm__ volatile (
1088                 "movq      %0, %%mm4    \n\t"
1089                 "movq      %1, %%mm5    \n\t"
1090                 "paddw     %2, %%mm4    \n\t"
1091                 "paddw     %3, %%mm5    \n\t"
1092                 "movq   %%mm4, %0       \n\t"
1093                 "movq   %%mm5, %1       \n\t"
1094                 "psrlw    $12, %%mm4    \n\t"
1095                 "psrlw    $12, %%mm5    \n\t"
1096                 : "+m"(*dx4), "+m"(*dy4)
1097                 : "m"(*dxy4), "m"(*dyy4)
1098             );
1099
1100             __asm__ volatile (
1101                 "movq      %%mm6, %%mm2 \n\t"
1102                 "movq      %%mm6, %%mm1 \n\t"
1103                 "psubw     %%mm4, %%mm2 \n\t"
1104                 "psubw     %%mm5, %%mm1 \n\t"
1105                 "movq      %%mm2, %%mm0 \n\t"
1106                 "movq      %%mm4, %%mm3 \n\t"
1107                 "pmullw    %%mm1, %%mm0 \n\t" // (s - dx) * (s - dy)
1108                 "pmullw    %%mm5, %%mm3 \n\t" // dx * dy
1109                 "pmullw    %%mm5, %%mm2 \n\t" // (s - dx) * dy
1110                 "pmullw    %%mm4, %%mm1 \n\t" // dx * (s - dy)
1111
1112                 "movd         %4, %%mm5 \n\t"
1113                 "movd         %3, %%mm4 \n\t"
1114                 "punpcklbw %%mm7, %%mm5 \n\t"
1115                 "punpcklbw %%mm7, %%mm4 \n\t"
1116                 "pmullw    %%mm5, %%mm3 \n\t" // src[1, 1] * dx * dy
1117                 "pmullw    %%mm4, %%mm2 \n\t" // src[0, 1] * (s - dx) * dy
1118
1119                 "movd         %2, %%mm5 \n\t"
1120                 "movd         %1, %%mm4 \n\t"
1121                 "punpcklbw %%mm7, %%mm5 \n\t"
1122                 "punpcklbw %%mm7, %%mm4 \n\t"
1123                 "pmullw    %%mm5, %%mm1 \n\t" // src[1, 0] * dx * (s - dy)
1124                 "pmullw    %%mm4, %%mm0 \n\t" // src[0, 0] * (s - dx) * (s - dy)
1125                 "paddw        %5, %%mm1 \n\t"
1126                 "paddw     %%mm3, %%mm2 \n\t"
1127                 "paddw     %%mm1, %%mm0 \n\t"
1128                 "paddw     %%mm2, %%mm0 \n\t"
1129
1130                 "psrlw        %6, %%mm0 \n\t"
1131                 "packuswb  %%mm0, %%mm0 \n\t"
1132                 "movd      %%mm0, %0    \n\t"
1133
1134                 : "=m"(dst[x + y * stride])
1135                 : "m"(src[0]), "m"(src[1]),
1136                   "m"(src[stride]), "m"(src[stride + 1]),
1137                   "m"(*r4), "m"(shift2)
1138             );
1139             src += stride;
1140         }
1141         src += 4 - h * stride;
1142     }
1143 }
1144
1145 /* CAVS-specific */
1146 void ff_put_cavs_qpel8_mc00_mmxext(uint8_t *dst, uint8_t *src, ptrdiff_t stride)
1147 {
1148     put_pixels8_mmx(dst, src, stride, 8);
1149 }
1150
1151 void ff_avg_cavs_qpel8_mc00_mmxext(uint8_t *dst, uint8_t *src, ptrdiff_t stride)
1152 {
1153     avg_pixels8_mmx(dst, src, stride, 8);
1154 }
1155
1156 void ff_put_cavs_qpel16_mc00_mmxext(uint8_t *dst, uint8_t *src, ptrdiff_t stride)
1157 {
1158     put_pixels16_mmx(dst, src, stride, 16);
1159 }
1160
1161 void ff_avg_cavs_qpel16_mc00_mmxext(uint8_t *dst, uint8_t *src, ptrdiff_t stride)
1162 {
1163     avg_pixels16_mmx(dst, src, stride, 16);
1164 }
1165
1166 /* VC-1-specific */
1167 void ff_put_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src,
1168                                ptrdiff_t stride, int rnd)
1169 {
1170     put_pixels8_mmx(dst, src, stride, 8);
1171 }
1172
1173 static void vector_clipf_sse(float *dst, const float *src,
1174                              float min, float max, int len)
1175 {
1176     x86_reg i = (len - 16) * 4;
1177     __asm__ volatile (
1178         "movss          %3, %%xmm4      \n\t"
1179         "movss          %4, %%xmm5      \n\t"
1180         "shufps $0, %%xmm4, %%xmm4      \n\t"
1181         "shufps $0, %%xmm5, %%xmm5      \n\t"
1182         "1:                             \n\t"
1183         "movaps   (%2, %0), %%xmm0      \n\t" // 3/1 on intel
1184         "movaps 16(%2, %0), %%xmm1      \n\t"
1185         "movaps 32(%2, %0), %%xmm2      \n\t"
1186         "movaps 48(%2, %0), %%xmm3      \n\t"
1187         "maxps      %%xmm4, %%xmm0      \n\t"
1188         "maxps      %%xmm4, %%xmm1      \n\t"
1189         "maxps      %%xmm4, %%xmm2      \n\t"
1190         "maxps      %%xmm4, %%xmm3      \n\t"
1191         "minps      %%xmm5, %%xmm0      \n\t"
1192         "minps      %%xmm5, %%xmm1      \n\t"
1193         "minps      %%xmm5, %%xmm2      \n\t"
1194         "minps      %%xmm5, %%xmm3      \n\t"
1195         "movaps     %%xmm0,   (%1, %0)  \n\t"
1196         "movaps     %%xmm1, 16(%1, %0)  \n\t"
1197         "movaps     %%xmm2, 32(%1, %0)  \n\t"
1198         "movaps     %%xmm3, 48(%1, %0)  \n\t"
1199         "sub           $64, %0          \n\t"
1200         "jge            1b              \n\t"
1201         : "+&r"(i)
1202         : "r"(dst), "r"(src), "m"(min), "m"(max)
1203         : "memory"
1204     );
1205 }
1206
1207 #endif /* HAVE_INLINE_ASM */
1208
1209 void ff_h263_v_loop_filter_mmx(uint8_t *src, int stride, int qscale);
1210 void ff_h263_h_loop_filter_mmx(uint8_t *src, int stride, int qscale);
1211
1212 int32_t ff_scalarproduct_int16_mmxext(const int16_t *v1, const int16_t *v2,
1213                                       int order);
1214 int32_t ff_scalarproduct_int16_sse2(const int16_t *v1, const int16_t *v2,
1215                                     int order);
1216 int32_t ff_scalarproduct_and_madd_int16_mmxext(int16_t *v1, const int16_t *v2,
1217                                                const int16_t *v3,
1218                                                int order, int mul);
1219 int32_t ff_scalarproduct_and_madd_int16_sse2(int16_t *v1, const int16_t *v2,
1220                                              const int16_t *v3,
1221                                              int order, int mul);
1222 int32_t ff_scalarproduct_and_madd_int16_ssse3(int16_t *v1, const int16_t *v2,
1223                                               const int16_t *v3,
1224                                               int order, int mul);
1225
1226 void ff_apply_window_int16_round_mmxext(int16_t *output, const int16_t *input,
1227                                         const int16_t *window, unsigned int len);
1228 void ff_apply_window_int16_round_sse2(int16_t *output, const int16_t *input,
1229                                       const int16_t *window, unsigned int len);
1230 void ff_apply_window_int16_mmxext(int16_t *output, const int16_t *input,
1231                                   const int16_t *window, unsigned int len);
1232 void ff_apply_window_int16_sse2(int16_t *output, const int16_t *input,
1233                                 const int16_t *window, unsigned int len);
1234 void ff_apply_window_int16_ssse3(int16_t *output, const int16_t *input,
1235                                  const int16_t *window, unsigned int len);
1236 void ff_apply_window_int16_ssse3_atom(int16_t *output, const int16_t *input,
1237                                       const int16_t *window, unsigned int len);
1238
1239 void ff_bswap32_buf_ssse3(uint32_t *dst, const uint32_t *src, int w);
1240 void ff_bswap32_buf_sse2(uint32_t *dst, const uint32_t *src, int w);
1241
1242 void ff_add_hfyu_median_prediction_mmxext(uint8_t *dst, const uint8_t *top,
1243                                           const uint8_t *diff, int w,
1244                                           int *left, int *left_top);
1245 int  ff_add_hfyu_left_prediction_ssse3(uint8_t *dst, const uint8_t *src,
1246                                        int w, int left);
1247 int  ff_add_hfyu_left_prediction_sse4(uint8_t *dst, const uint8_t *src,
1248                                       int w, int left);
1249
1250 void ff_vector_clip_int32_mmx     (int32_t *dst, const int32_t *src,
1251                                    int32_t min, int32_t max, unsigned int len);
1252 void ff_vector_clip_int32_sse2    (int32_t *dst, const int32_t *src,
1253                                    int32_t min, int32_t max, unsigned int len);
1254 void ff_vector_clip_int32_int_sse2(int32_t *dst, const int32_t *src,
1255                                    int32_t min, int32_t max, unsigned int len);
1256 void ff_vector_clip_int32_sse4    (int32_t *dst, const int32_t *src,
1257                                    int32_t min, int32_t max, unsigned int len);
1258
1259 #define SET_QPEL_FUNCS(PFX, IDX, SIZE, CPU, PREFIX)                          \
1260     do {                                                                     \
1261     c->PFX ## _pixels_tab[IDX][ 0] = PREFIX ## PFX ## SIZE ## _mc00_ ## CPU; \
1262     c->PFX ## _pixels_tab[IDX][ 1] = PREFIX ## PFX ## SIZE ## _mc10_ ## CPU; \
1263     c->PFX ## _pixels_tab[IDX][ 2] = PREFIX ## PFX ## SIZE ## _mc20_ ## CPU; \
1264     c->PFX ## _pixels_tab[IDX][ 3] = PREFIX ## PFX ## SIZE ## _mc30_ ## CPU; \
1265     c->PFX ## _pixels_tab[IDX][ 4] = PREFIX ## PFX ## SIZE ## _mc01_ ## CPU; \
1266     c->PFX ## _pixels_tab[IDX][ 5] = PREFIX ## PFX ## SIZE ## _mc11_ ## CPU; \
1267     c->PFX ## _pixels_tab[IDX][ 6] = PREFIX ## PFX ## SIZE ## _mc21_ ## CPU; \
1268     c->PFX ## _pixels_tab[IDX][ 7] = PREFIX ## PFX ## SIZE ## _mc31_ ## CPU; \
1269     c->PFX ## _pixels_tab[IDX][ 8] = PREFIX ## PFX ## SIZE ## _mc02_ ## CPU; \
1270     c->PFX ## _pixels_tab[IDX][ 9] = PREFIX ## PFX ## SIZE ## _mc12_ ## CPU; \
1271     c->PFX ## _pixels_tab[IDX][10] = PREFIX ## PFX ## SIZE ## _mc22_ ## CPU; \
1272     c->PFX ## _pixels_tab[IDX][11] = PREFIX ## PFX ## SIZE ## _mc32_ ## CPU; \
1273     c->PFX ## _pixels_tab[IDX][12] = PREFIX ## PFX ## SIZE ## _mc03_ ## CPU; \
1274     c->PFX ## _pixels_tab[IDX][13] = PREFIX ## PFX ## SIZE ## _mc13_ ## CPU; \
1275     c->PFX ## _pixels_tab[IDX][14] = PREFIX ## PFX ## SIZE ## _mc23_ ## CPU; \
1276     c->PFX ## _pixels_tab[IDX][15] = PREFIX ## PFX ## SIZE ## _mc33_ ## CPU; \
1277     } while (0)
1278
1279 static av_cold void dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx,
1280                                      int mm_flags)
1281 {
1282     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
1283
1284 #if HAVE_INLINE_ASM
1285     c->put_pixels_clamped        = ff_put_pixels_clamped_mmx;
1286     c->put_signed_pixels_clamped = ff_put_signed_pixels_clamped_mmx;
1287     c->add_pixels_clamped        = ff_add_pixels_clamped_mmx;
1288
1289     if (!high_bit_depth) {
1290         c->clear_block  = clear_block_mmx;
1291         c->clear_blocks = clear_blocks_mmx;
1292         c->draw_edges   = draw_edges_mmx;
1293
1294         switch (avctx->idct_algo) {
1295         case FF_IDCT_AUTO:
1296         case FF_IDCT_SIMPLEMMX:
1297             c->idct_put              = ff_simple_idct_put_mmx;
1298             c->idct_add              = ff_simple_idct_add_mmx;
1299             c->idct                  = ff_simple_idct_mmx;
1300             c->idct_permutation_type = FF_SIMPLE_IDCT_PERM;
1301             break;
1302         case FF_IDCT_XVIDMMX:
1303             c->idct_put              = ff_idct_xvid_mmx_put;
1304             c->idct_add              = ff_idct_xvid_mmx_add;
1305             c->idct                  = ff_idct_xvid_mmx;
1306             break;
1307         }
1308     }
1309
1310     c->gmc = gmc_mmx;
1311
1312     c->add_bytes = add_bytes_mmx;
1313 #endif /* HAVE_INLINE_ASM */
1314
1315 #if HAVE_YASM
1316     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
1317         c->h263_v_loop_filter = ff_h263_v_loop_filter_mmx;
1318         c->h263_h_loop_filter = ff_h263_h_loop_filter_mmx;
1319     }
1320
1321     c->vector_clip_int32 = ff_vector_clip_int32_mmx;
1322 #endif
1323
1324 }
1325
1326 static av_cold void dsputil_init_mmxext(DSPContext *c, AVCodecContext *avctx,
1327                                         int mm_flags)
1328 {
1329     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
1330
1331 #if HAVE_YASM
1332     SET_QPEL_FUNCS(avg_qpel,        0, 16, mmxext, );
1333     SET_QPEL_FUNCS(avg_qpel,        1,  8, mmxext, );
1334
1335     SET_QPEL_FUNCS(put_qpel,        0, 16, mmxext, );
1336     SET_QPEL_FUNCS(put_qpel,        1,  8, mmxext, );
1337     SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, mmxext, );
1338     SET_QPEL_FUNCS(put_no_rnd_qpel, 1,  8, mmxext, );
1339 #endif /* HAVE_YASM */
1340
1341 #if HAVE_INLINE_ASM
1342     if (!high_bit_depth && avctx->idct_algo == FF_IDCT_XVIDMMX) {
1343         c->idct_put = ff_idct_xvid_mmxext_put;
1344         c->idct_add = ff_idct_xvid_mmxext_add;
1345         c->idct     = ff_idct_xvid_mmxext;
1346     }
1347 #endif /* HAVE_INLINE_ASM */
1348
1349 #if HAVE_MMXEXT_EXTERNAL
1350     /* slower than cmov version on AMD */
1351     if (!(mm_flags & AV_CPU_FLAG_3DNOW))
1352         c->add_hfyu_median_prediction = ff_add_hfyu_median_prediction_mmxext;
1353
1354     c->scalarproduct_int16          = ff_scalarproduct_int16_mmxext;
1355     c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_mmxext;
1356
1357     if (avctx->flags & CODEC_FLAG_BITEXACT) {
1358         c->apply_window_int16 = ff_apply_window_int16_mmxext;
1359     } else {
1360         c->apply_window_int16 = ff_apply_window_int16_round_mmxext;
1361     }
1362 #endif /* HAVE_MMXEXT_EXTERNAL */
1363 }
1364
1365 static av_cold void dsputil_init_sse(DSPContext *c, AVCodecContext *avctx,
1366                                      int mm_flags)
1367 {
1368 #if HAVE_INLINE_ASM
1369     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
1370
1371     if (!high_bit_depth) {
1372         if (!(CONFIG_MPEG_XVMC_DECODER && avctx->xvmc_acceleration > 1)) {
1373             /* XvMCCreateBlocks() may not allocate 16-byte aligned blocks */
1374             c->clear_block  = clear_block_sse;
1375             c->clear_blocks = clear_blocks_sse;
1376         }
1377     }
1378
1379     c->vector_clipf = vector_clipf_sse;
1380 #endif /* HAVE_INLINE_ASM */
1381 }
1382
1383 static av_cold void dsputil_init_sse2(DSPContext *c, AVCodecContext *avctx,
1384                                       int mm_flags)
1385 {
1386     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
1387
1388 #if HAVE_SSE2_INLINE
1389     if (!high_bit_depth && avctx->idct_algo == FF_IDCT_XVIDMMX) {
1390         c->idct_put              = ff_idct_xvid_sse2_put;
1391         c->idct_add              = ff_idct_xvid_sse2_add;
1392         c->idct                  = ff_idct_xvid_sse2;
1393         c->idct_permutation_type = FF_SSE2_IDCT_PERM;
1394     }
1395 #endif /* HAVE_SSE2_INLINE */
1396
1397 #if HAVE_SSE2_EXTERNAL
1398     c->scalarproduct_int16          = ff_scalarproduct_int16_sse2;
1399     c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_sse2;
1400     if (mm_flags & AV_CPU_FLAG_ATOM) {
1401         c->vector_clip_int32 = ff_vector_clip_int32_int_sse2;
1402     } else {
1403         c->vector_clip_int32 = ff_vector_clip_int32_sse2;
1404     }
1405     if (avctx->flags & CODEC_FLAG_BITEXACT) {
1406         c->apply_window_int16 = ff_apply_window_int16_sse2;
1407     } else if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {
1408         c->apply_window_int16 = ff_apply_window_int16_round_sse2;
1409     }
1410     c->bswap_buf = ff_bswap32_buf_sse2;
1411 #endif /* HAVE_SSE2_EXTERNAL */
1412 }
1413
1414 static av_cold void dsputil_init_ssse3(DSPContext *c, AVCodecContext *avctx,
1415                                        int mm_flags)
1416 {
1417 #if HAVE_SSSE3_EXTERNAL
1418     c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_ssse3;
1419     if (mm_flags & AV_CPU_FLAG_SSE4) // not really sse4, just slow on Conroe
1420         c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_sse4;
1421
1422     if (mm_flags & AV_CPU_FLAG_ATOM)
1423         c->apply_window_int16 = ff_apply_window_int16_ssse3_atom;
1424     else
1425         c->apply_window_int16 = ff_apply_window_int16_ssse3;
1426     if (!(mm_flags & (AV_CPU_FLAG_SSE42|AV_CPU_FLAG_3DNOW))) // cachesplit
1427         c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_ssse3;
1428     c->bswap_buf = ff_bswap32_buf_ssse3;
1429 #endif /* HAVE_SSSE3_EXTERNAL */
1430 }
1431
1432 static av_cold void dsputil_init_sse4(DSPContext *c, AVCodecContext *avctx,
1433                                       int mm_flags)
1434 {
1435 #if HAVE_SSE4_EXTERNAL
1436     c->vector_clip_int32 = ff_vector_clip_int32_sse4;
1437 #endif /* HAVE_SSE4_EXTERNAL */
1438 }
1439
1440 av_cold void ff_dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx)
1441 {
1442     int mm_flags = av_get_cpu_flags();
1443
1444 #if HAVE_7REGS && HAVE_INLINE_ASM
1445     if (mm_flags & AV_CPU_FLAG_CMOV)
1446         c->add_hfyu_median_prediction = add_hfyu_median_prediction_cmov;
1447 #endif
1448
1449     if (mm_flags & AV_CPU_FLAG_MMX)
1450         dsputil_init_mmx(c, avctx, mm_flags);
1451
1452     if (mm_flags & AV_CPU_FLAG_MMXEXT)
1453         dsputil_init_mmxext(c, avctx, mm_flags);
1454
1455     if (mm_flags & AV_CPU_FLAG_SSE)
1456         dsputil_init_sse(c, avctx, mm_flags);
1457
1458     if (mm_flags & AV_CPU_FLAG_SSE2)
1459         dsputil_init_sse2(c, avctx, mm_flags);
1460
1461     if (mm_flags & AV_CPU_FLAG_SSSE3)
1462         dsputil_init_ssse3(c, avctx, mm_flags);
1463
1464     if (mm_flags & AV_CPU_FLAG_SSE4)
1465         dsputil_init_sse4(c, avctx, mm_flags);
1466
1467     if (CONFIG_ENCODERS)
1468         ff_dsputilenc_init_mmx(c, avctx);
1469 }