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

   1 /*
   2  * MMX optimized DSP utils
   3  * Copyright (c) 2000, 2001 Fabrice Bellard
   4  * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
   5  *
   6  * This file is part of FFmpeg.
   7  *
   8  * FFmpeg is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU Lesser General Public
  10  * License as published by the Free Software Foundation; either
  11  * version 2.1 of the License, or (at your option) any later version.
  12  *
  13  * FFmpeg is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * Lesser General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU Lesser General Public
  19  * License along with FFmpeg; if not, write to the Free Software
  20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  21  *
  22  * MMX optimization by Nick Kurshev <nickols_k@mail.ru>
  23  */
  24
  25 #include "libavutil/cpu.h"
  26 #include "libavutil/x86/asm.h"
  27 #include "libavcodec/dsputil.h"
  28 #include "libavcodec/h264dsp.h"
  29 #include "libavcodec/mpegvideo.h"
  30 #include "libavcodec/simple_idct.h"
  31 #include "dsputil_mmx.h"
  32 #include "idct_xvid.h"
  33 #include "diracdsp_mmx.h"
  34
  35 //#undef NDEBUG
  36 //#include <assert.h>
  37
  38 /* pixel operations */
  39 DECLARE_ALIGNED(8,  const uint64_t, ff_bone) = 0x0101010101010101ULL;
  40 DECLARE_ALIGNED(8,  const uint64_t, ff_wtwo) = 0x0002000200020002ULL;
  41
  42 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_1)    = { 0x0001000100010001ULL, 0x0001000100010001ULL };
  43 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_2)    = { 0x0002000200020002ULL, 0x0002000200020002ULL };
  44 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_3)    = { 0x0003000300030003ULL, 0x0003000300030003ULL };
  45 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_4)    = { 0x0004000400040004ULL, 0x0004000400040004ULL };
  46 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_5)    = { 0x0005000500050005ULL, 0x0005000500050005ULL };
  47 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_8)    = { 0x0008000800080008ULL, 0x0008000800080008ULL };
  48 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_9)    = { 0x0009000900090009ULL, 0x0009000900090009ULL };
  49 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_15)   =   0x000F000F000F000FULL;
  50 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_16)   = { 0x0010001000100010ULL, 0x0010001000100010ULL };
  51 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_17)   = { 0x0011001100110011ULL, 0x0011001100110011ULL };
  52 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_18)   = { 0x0012001200120012ULL, 0x0012001200120012ULL };
  53 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_20)   =   0x0014001400140014ULL;
  54 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_27)   = { 0x001B001B001B001BULL, 0x001B001B001B001BULL };
  55 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_28)   = { 0x001C001C001C001CULL, 0x001C001C001C001CULL };
  56 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_32)   = { 0x0020002000200020ULL, 0x0020002000200020ULL };
  57 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_42)   =   0x002A002A002A002AULL;
  58 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_53)   =   0x0035003500350035ULL;
  59 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_63)   = { 0x003F003F003F003FULL, 0x003F003F003F003FULL };
  60 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_64)   = { 0x0040004000400040ULL, 0x0040004000400040ULL };
  61 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_96)   =   0x0060006000600060ULL;
  62 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_128)  =   0x0080008000800080ULL;
  63 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_255)  =   0x00ff00ff00ff00ffULL;
  64 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_512)  = { 0x0200020002000200ULL, 0x0200020002000200ULL };
  65 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_1019) = { 0x03FB03FB03FB03FBULL, 0x03FB03FB03FB03FBULL };
  66
  67 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_0)    = { 0x0000000000000000ULL, 0x0000000000000000ULL };
  68 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_1)    = { 0x0101010101010101ULL, 0x0101010101010101ULL };
  69 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_3)    = { 0x0303030303030303ULL, 0x0303030303030303ULL };
  70 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_4)    = { 0x0404040404040404ULL, 0x0404040404040404ULL };
  71 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_7)    =   0x0707070707070707ULL;
  72 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_1F)   =   0x1F1F1F1F1F1F1F1FULL;
  73 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_3F)   =   0x3F3F3F3F3F3F3F3FULL;
  74 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_80)   = { 0x8080808080808080ULL, 0x8080808080808080ULL };
  75 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_81)   =   0x8181818181818181ULL;
  76 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_A1)   = { 0xA1A1A1A1A1A1A1A1ULL, 0xA1A1A1A1A1A1A1A1ULL };
  77 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_F8)   = { 0xF8F8F8F8F8F8F8F8ULL, 0xF8F8F8F8F8F8F8F8ULL };
  78 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_FC)   =   0xFCFCFCFCFCFCFCFCULL;
  79 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_FE)   = { 0xFEFEFEFEFEFEFEFEULL, 0xFEFEFEFEFEFEFEFEULL };
  80
  81 DECLARE_ALIGNED(16, const double, ff_pd_1)[2] = { 1.0, 1.0 };
  82 DECLARE_ALIGNED(16, const double, ff_pd_2)[2] = { 2.0, 2.0 };
  83
  84
  85 #if HAVE_YASM
  86 void ff_put_pixels8_x2_mmxext(uint8_t *block, const uint8_t *pixels,
  87                               int line_size, int h);
  88 void ff_put_pixels8_x2_3dnow(uint8_t *block, const uint8_t *pixels,
  89                              int line_size, int h);
  90 void ff_put_pixels8_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
  91                               int dstStride, int src1Stride, int h);
  92 void ff_put_no_rnd_pixels8_l2_mmxext(uint8_t *dst, uint8_t *src1,
  93                                      uint8_t *src2, int dstStride,
  94                                      int src1Stride, int h);
  95 void ff_avg_pixels8_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
  96                               int dstStride, int src1Stride, int h);
  97 void ff_put_pixels16_x2_mmxext(uint8_t *block, const uint8_t *pixels,
  98                                int line_size, int h);
  99 void ff_put_pixels16_x2_3dnow(uint8_t *block, const uint8_t *pixels,
 100                               int line_size, int h);
 101 void ff_put_pixels16_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
 102                                int dstStride, int src1Stride, int h);
 103 void ff_avg_pixels16_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
 104                                int dstStride, int src1Stride, int h);
 105 void ff_put_no_rnd_pixels16_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
 106                                       int dstStride, int src1Stride, int h);
 107 void ff_put_no_rnd_pixels8_x2_mmxext(uint8_t *block, const uint8_t *pixels,
 108                                      int line_size, int h);
 109 void ff_put_no_rnd_pixels8_x2_3dnow(uint8_t *block, const uint8_t *pixels,
 110                                     int line_size, int h);
 111 void ff_put_no_rnd_pixels8_x2_exact_mmxext(uint8_t *block,
 112                                            const uint8_t *pixels,
 113                                            int line_size, int h);
 114 void ff_put_no_rnd_pixels8_x2_exact_3dnow(uint8_t *block,
 115                                           const uint8_t *pixels,
 116                                           int line_size, int h);
 117 void ff_put_pixels8_y2_mmxext(uint8_t *block, const uint8_t *pixels,
 118                               int line_size, int h);
 119 void ff_put_pixels8_y2_3dnow(uint8_t *block, const uint8_t *pixels,
 120                              int line_size, int h);
 121 void ff_put_no_rnd_pixels8_y2_mmxext(uint8_t *block, const uint8_t *pixels,
 122                                      int line_size, int h);
 123 void ff_put_no_rnd_pixels8_y2_3dnow(uint8_t *block, const uint8_t *pixels,
 124                                     int line_size, int h);
 125 void ff_put_no_rnd_pixels8_y2_exact_mmxext(uint8_t *block,
 126                                            const uint8_t *pixels,
 127                                            int line_size, int h);
 128 void ff_put_no_rnd_pixels8_y2_exact_3dnow(uint8_t *block,
 129                                           const uint8_t *pixels,
 130                                           int line_size, int h);
 131 void ff_avg_pixels8_mmxext(uint8_t *block, const uint8_t *pixels,
 132                            int line_size, int h);
 133 void ff_avg_pixels8_3dnow(uint8_t *block, const uint8_t *pixels,
 134                           int line_size, int h);
 135 void ff_avg_pixels8_x2_mmxext(uint8_t *block, const uint8_t *pixels,
 136                               int line_size, int h);
 137 void ff_avg_pixels8_x2_3dnow(uint8_t *block, const uint8_t *pixels,
 138                              int line_size, int h);
 139 void ff_avg_pixels8_y2_mmxext(uint8_t *block, const uint8_t *pixels,
 140                               int line_size, int h);
 141 void ff_avg_pixels8_y2_3dnow(uint8_t *block, const uint8_t *pixels,
 142                              int line_size, int h);
 143 void ff_avg_pixels8_xy2_mmxext(uint8_t *block, const uint8_t *pixels,
 144                                int line_size, int h);
 145 void ff_avg_pixels8_xy2_3dnow(uint8_t *block, const uint8_t *pixels,
 146                               int line_size, int h);
 147
 148 void ff_put_pixels8_mmxext(uint8_t *block, const uint8_t *pixels, int line_size, int h);
 149 static void ff_put_pixels16_mmxext(uint8_t *block, const uint8_t *pixels,
 150                                    int line_size, int h)
 151 {
 152     ff_put_pixels8_mmxext(block,     pixels,     line_size, h);
 153     ff_put_pixels8_mmxext(block + 8, pixels + 8, line_size, h);
 154 }
 155
 156 void ff_put_mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 157                                          int dstStride, int srcStride, int h);
 158 void ff_avg_mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 159                                          int dstStride, int srcStride, int h);
 160 void ff_put_no_rnd_mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 161                                                  int dstStride, int srcStride,
 162                                                  int h);
 163 void ff_put_mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 164                                         int dstStride, int srcStride, int h);
 165 void ff_avg_mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 166                                         int dstStride, int srcStride, int h);
 167 void ff_put_no_rnd_mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 168                                                 int dstStride, int srcStride,
 169                                                 int h);
 170 void ff_put_mpeg4_qpel16_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 171                                          int dstStride, int srcStride);
 172 void ff_avg_mpeg4_qpel16_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 173                                          int dstStride, int srcStride);
 174 void ff_put_no_rnd_mpeg4_qpel16_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 175                                                  int dstStride, int srcStride);
 176 void ff_put_mpeg4_qpel8_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 177                                         int dstStride, int srcStride);
 178 void ff_avg_mpeg4_qpel8_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 179                                         int dstStride, int srcStride);
 180 void ff_put_no_rnd_mpeg4_qpel8_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 181                                                 int dstStride, int srcStride);
 182 #define ff_put_no_rnd_pixels16_mmxext ff_put_pixels16_mmxext
 183 #define ff_put_no_rnd_pixels8_mmxext ff_put_pixels8_mmxext
 184 #endif /* HAVE_YASM */
 185
 186
 187 #if HAVE_INLINE_ASM
 188
 189 #define JUMPALIGN()     __asm__ volatile (".p2align 3"::)
 190 #define MOVQ_ZERO(regd) __asm__ volatile ("pxor %%"#regd", %%"#regd ::)
 191
 192 #define MOVQ_BFE(regd)                                  \
 193     __asm__ volatile (                                  \
 194         "pcmpeqd %%"#regd", %%"#regd"   \n\t"           \
 195         "paddb   %%"#regd", %%"#regd"   \n\t" ::)
 196
 197 #ifndef PIC
 198 #define MOVQ_BONE(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_bone))
 199 #define MOVQ_WTWO(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_wtwo))
 200 #else
 201 // for shared library it's better to use this way for accessing constants
 202 // pcmpeqd -> -1
 203 #define MOVQ_BONE(regd)                                 \
 204     __asm__ volatile (                                  \
 205         "pcmpeqd  %%"#regd", %%"#regd"  \n\t"           \
 206         "psrlw          $15, %%"#regd"  \n\t"           \
 207         "packuswb %%"#regd", %%"#regd"  \n\t" ::)
 208
 209 #define MOVQ_WTWO(regd)                                 \
 210     __asm__ volatile (                                  \
 211         "pcmpeqd %%"#regd", %%"#regd"   \n\t"           \
 212         "psrlw         $15, %%"#regd"   \n\t"           \
 213         "psllw          $1, %%"#regd"   \n\t"::)
 214
 215 #endif
 216
 217 // using regr as temporary and for the output result
 218 // first argument is unmodifed and second is trashed
 219 // regfe is supposed to contain 0xfefefefefefefefe
 220 #define PAVGB_MMX_NO_RND(rega, regb, regr, regfe)                \
 221     "movq   "#rega", "#regr"            \n\t"                    \
 222     "pand   "#regb", "#regr"            \n\t"                    \
 223     "pxor   "#rega", "#regb"            \n\t"                    \
 224     "pand  "#regfe", "#regb"            \n\t"                    \
 225     "psrlq       $1, "#regb"            \n\t"                    \
 226     "paddb  "#regb", "#regr"            \n\t"
 227
 228 #define PAVGB_MMX(rega, regb, regr, regfe)                       \
 229     "movq   "#rega", "#regr"            \n\t"                    \
 230     "por    "#regb", "#regr"            \n\t"                    \
 231     "pxor   "#rega", "#regb"            \n\t"                    \
 232     "pand  "#regfe", "#regb"            \n\t"                    \
 233     "psrlq       $1, "#regb"            \n\t"                    \
 234     "psubb  "#regb", "#regr"            \n\t"
 235
 236 // mm6 is supposed to contain 0xfefefefefefefefe
 237 #define PAVGBP_MMX_NO_RND(rega, regb, regr,  regc, regd, regp)   \
 238     "movq  "#rega", "#regr"             \n\t"                    \
 239     "movq  "#regc", "#regp"             \n\t"                    \
 240     "pand  "#regb", "#regr"             \n\t"                    \
 241     "pand  "#regd", "#regp"             \n\t"                    \
 242     "pxor  "#rega", "#regb"             \n\t"                    \
 243     "pxor  "#regc", "#regd"             \n\t"                    \
 244     "pand    %%mm6, "#regb"             \n\t"                    \
 245     "pand    %%mm6, "#regd"             \n\t"                    \
 246     "psrlq      $1, "#regb"             \n\t"                    \
 247     "psrlq      $1, "#regd"             \n\t"                    \
 248     "paddb "#regb", "#regr"             \n\t"                    \
 249     "paddb "#regd", "#regp"             \n\t"
 250
 251 #define PAVGBP_MMX(rega, regb, regr, regc, regd, regp)           \
 252     "movq  "#rega", "#regr"             \n\t"                    \
 253     "movq  "#regc", "#regp"             \n\t"                    \
 254     "por   "#regb", "#regr"             \n\t"                    \
 255     "por   "#regd", "#regp"             \n\t"                    \
 256     "pxor  "#rega", "#regb"             \n\t"                    \
 257     "pxor  "#regc", "#regd"             \n\t"                    \
 258     "pand    %%mm6, "#regb"             \n\t"                    \
 259     "pand    %%mm6, "#regd"             \n\t"                    \
 260     "psrlq      $1, "#regd"             \n\t"                    \
 261     "psrlq      $1, "#regb"             \n\t"                    \
 262     "psubb "#regb", "#regr"             \n\t"                    \
 263     "psubb "#regd", "#regp"             \n\t"
 264
 265 /***********************************/
 266 /* MMX no rounding */
 267 #define NO_RND 1
 268 #define DEF(x, y) x ## _no_rnd_ ## y ## _mmx
 269 #define SET_RND  MOVQ_WONE
 270 #define PAVGBP(a, b, c, d, e, f)        PAVGBP_MMX_NO_RND(a, b, c, d, e, f)
 271 #define PAVGB(a, b, c, e)               PAVGB_MMX_NO_RND(a, b, c, e)
 272 #define OP_AVG(a, b, c, e)              PAVGB_MMX(a, b, c, e)
 273
 274 #include "dsputil_rnd_template.c"
 275
 276 #undef DEF
 277 #undef SET_RND
 278 #undef PAVGBP
 279 #undef PAVGB
 280 #undef NO_RND
 281 /***********************************/
 282 /* MMX rounding */
 283
 284 #define DEF(x, y) x ## _ ## y ## _mmx
 285 #define SET_RND  MOVQ_WTWO
 286 #define PAVGBP(a, b, c, d, e, f)        PAVGBP_MMX(a, b, c, d, e, f)
 287 #define PAVGB(a, b, c, e)               PAVGB_MMX(a, b, c, e)
 288
 289 #include "dsputil_rnd_template.c"
 290
 291 #undef DEF
 292 #undef SET_RND
 293 #undef PAVGBP
 294 #undef PAVGB
 295 #undef OP_AVG
 296
 297 #endif /* HAVE_INLINE_ASM */
 298
 299
 300 #if HAVE_YASM
 301 #define ff_put_pixels8_mmx ff_put_pixels8_mmxext
 302
 303 /***********************************/
 304 /* 3Dnow specific */
 305
 306 #define DEF(x) x ## _3dnow
 307
 308 #include "dsputil_avg_template.c"
 309
 310 #undef DEF
 311
 312 /***********************************/
 313 /* MMXEXT specific */
 314
 315 #define DEF(x) x ## _mmxext
 316
 317 #include "dsputil_avg_template.c"
 318
 319 #undef DEF
 320
 321 #endif /* HAVE_YASM */
 322
 323
 324 #if HAVE_INLINE_ASM
 325 #define put_no_rnd_pixels16_mmx put_pixels16_mmx
 326 #define put_no_rnd_pixels8_mmx put_pixels8_mmx
 327 #define put_pixels16_mmxext put_pixels16_mmx
 328 #define put_pixels8_mmxext put_pixels8_mmx
 329 #define put_pixels4_mmxext put_pixels4_mmx
 330 #define put_no_rnd_pixels16_mmxext put_no_rnd_pixels16_mmx
 331 #define put_no_rnd_pixels8_mmxext put_no_rnd_pixels8_mmx
 332
 333 /***********************************/
 334 /* standard MMX */
 335
 336 void ff_put_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
 337                                int line_size)
 338 {
 339     const int16_t *p;
 340     uint8_t *pix;
 341
 342     /* read the pixels */
 343     p   = block;
 344     pix = pixels;
 345     /* unrolled loop */
 346     __asm__ volatile (
 347         "movq      (%3), %%mm0          \n\t"
 348         "movq     8(%3), %%mm1          \n\t"
 349         "movq    16(%3), %%mm2          \n\t"
 350         "movq    24(%3), %%mm3          \n\t"
 351         "movq    32(%3), %%mm4          \n\t"
 352         "movq    40(%3), %%mm5          \n\t"
 353         "movq    48(%3), %%mm6          \n\t"
 354         "movq    56(%3), %%mm7          \n\t"
 355         "packuswb %%mm1, %%mm0          \n\t"
 356         "packuswb %%mm3, %%mm2          \n\t"
 357         "packuswb %%mm5, %%mm4          \n\t"
 358         "packuswb %%mm7, %%mm6          \n\t"
 359         "movq     %%mm0, (%0)           \n\t"
 360         "movq     %%mm2, (%0, %1)       \n\t"
 361         "movq     %%mm4, (%0, %1, 2)    \n\t"
 362         "movq     %%mm6, (%0, %2)       \n\t"
 363         :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3),
 364            "r"(p)
 365         : "memory");
 366     pix += line_size * 4;
 367     p   += 32;
 368
 369     // if here would be an exact copy of the code above
 370     // compiler would generate some very strange code
 371     // thus using "r"
 372     __asm__ volatile (
 373         "movq       (%3), %%mm0         \n\t"
 374         "movq      8(%3), %%mm1         \n\t"
 375         "movq     16(%3), %%mm2         \n\t"
 376         "movq     24(%3), %%mm3         \n\t"
 377         "movq     32(%3), %%mm4         \n\t"
 378         "movq     40(%3), %%mm5         \n\t"
 379         "movq     48(%3), %%mm6         \n\t"
 380         "movq     56(%3), %%mm7         \n\t"
 381         "packuswb  %%mm1, %%mm0         \n\t"
 382         "packuswb  %%mm3, %%mm2         \n\t"
 383         "packuswb  %%mm5, %%mm4         \n\t"
 384         "packuswb  %%mm7, %%mm6         \n\t"
 385         "movq      %%mm0, (%0)          \n\t"
 386         "movq      %%mm2, (%0, %1)      \n\t"
 387         "movq      %%mm4, (%0, %1, 2)   \n\t"
 388         "movq      %%mm6, (%0, %2)      \n\t"
 389         :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3), "r"(p)
 390         : "memory");
 391 }
 392
 393 #define put_signed_pixels_clamped_mmx_half(off)             \
 394     "movq          "#off"(%2), %%mm1        \n\t"           \
 395     "movq     16 + "#off"(%2), %%mm2        \n\t"           \
 396     "movq     32 + "#off"(%2), %%mm3        \n\t"           \
 397     "movq     48 + "#off"(%2), %%mm4        \n\t"           \
 398     "packsswb  8 + "#off"(%2), %%mm1        \n\t"           \
 399     "packsswb 24 + "#off"(%2), %%mm2        \n\t"           \
 400     "packsswb 40 + "#off"(%2), %%mm3        \n\t"           \
 401     "packsswb 56 + "#off"(%2), %%mm4        \n\t"           \
 402     "paddb              %%mm0, %%mm1        \n\t"           \
 403     "paddb              %%mm0, %%mm2        \n\t"           \
 404     "paddb              %%mm0, %%mm3        \n\t"           \
 405     "paddb              %%mm0, %%mm4        \n\t"           \
 406     "movq               %%mm1, (%0)         \n\t"           \
 407     "movq               %%mm2, (%0, %3)     \n\t"           \
 408     "movq               %%mm3, (%0, %3, 2)  \n\t"           \
 409     "movq               %%mm4, (%0, %1)     \n\t"
 410
 411 void ff_put_signed_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
 412                                       int line_size)
 413 {
 414     x86_reg line_skip = line_size;
 415     x86_reg line_skip3;
 416
 417     __asm__ volatile (
 418         "movq "MANGLE(ff_pb_80)", %%mm0     \n\t"
 419         "lea         (%3, %3, 2), %1        \n\t"
 420         put_signed_pixels_clamped_mmx_half(0)
 421         "lea         (%0, %3, 4), %0        \n\t"
 422         put_signed_pixels_clamped_mmx_half(64)
 423         : "+&r"(pixels), "=&r"(line_skip3)
 424         : "r"(block), "r"(line_skip)
 425         : "memory");
 426 }
 427
 428 void ff_add_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
 429                                int line_size)
 430 {
 431     const int16_t *p;
 432     uint8_t *pix;
 433     int i;
 434
 435     /* read the pixels */
 436     p   = block;
 437     pix = pixels;
 438     MOVQ_ZERO(mm7);
 439     i = 4;
 440     do {
 441         __asm__ volatile (
 442             "movq        (%2), %%mm0    \n\t"
 443             "movq       8(%2), %%mm1    \n\t"
 444             "movq      16(%2), %%mm2    \n\t"
 445             "movq      24(%2), %%mm3    \n\t"
 446             "movq          %0, %%mm4    \n\t"
 447             "movq          %1, %%mm6    \n\t"
 448             "movq       %%mm4, %%mm5    \n\t"
 449             "punpcklbw  %%mm7, %%mm4    \n\t"
 450             "punpckhbw  %%mm7, %%mm5    \n\t"
 451             "paddsw     %%mm4, %%mm0    \n\t"
 452             "paddsw     %%mm5, %%mm1    \n\t"
 453             "movq       %%mm6, %%mm5    \n\t"
 454             "punpcklbw  %%mm7, %%mm6    \n\t"
 455             "punpckhbw  %%mm7, %%mm5    \n\t"
 456             "paddsw     %%mm6, %%mm2    \n\t"
 457             "paddsw     %%mm5, %%mm3    \n\t"
 458             "packuswb   %%mm1, %%mm0    \n\t"
 459             "packuswb   %%mm3, %%mm2    \n\t"
 460             "movq       %%mm0, %0       \n\t"
 461             "movq       %%mm2, %1       \n\t"
 462             : "+m"(*pix), "+m"(*(pix + line_size))
 463             : "r"(p)
 464             : "memory");
 465         pix += line_size * 2;
 466         p   += 16;
 467     } while (--i);
 468 }
 469
 470 static void put_pixels8_mmx(uint8_t *block, const uint8_t *pixels,
 471                             int line_size, int h)
 472 {
 473     __asm__ volatile (
 474         "lea   (%3, %3), %%"REG_a"      \n\t"
 475         ".p2align     3                 \n\t"
 476         "1:                             \n\t"
 477         "movq  (%1    ), %%mm0          \n\t"
 478         "movq  (%1, %3), %%mm1          \n\t"
 479         "movq     %%mm0, (%2)           \n\t"
 480         "movq     %%mm1, (%2, %3)       \n\t"
 481         "add  %%"REG_a", %1             \n\t"
 482         "add  %%"REG_a", %2             \n\t"
 483         "movq  (%1    ), %%mm0          \n\t"
 484         "movq  (%1, %3), %%mm1          \n\t"
 485         "movq     %%mm0, (%2)           \n\t"
 486         "movq     %%mm1, (%2, %3)       \n\t"
 487         "add  %%"REG_a", %1             \n\t"
 488         "add  %%"REG_a", %2             \n\t"
 489         "subl        $4, %0             \n\t"
 490         "jnz         1b                 \n\t"
 491         : "+g"(h), "+r"(pixels),  "+r"(block)
 492         : "r"((x86_reg)line_size)
 493         : "%"REG_a, "memory"
 494         );
 495 }
 496
 497 static void put_pixels16_mmx(uint8_t *block, const uint8_t *pixels,
 498                              int line_size, int h)
 499 {
 500     __asm__ volatile (
 501         "lea   (%3, %3), %%"REG_a"      \n\t"
 502         ".p2align     3                 \n\t"
 503         "1:                             \n\t"
 504         "movq  (%1    ), %%mm0          \n\t"
 505         "movq 8(%1    ), %%mm4          \n\t"
 506         "movq  (%1, %3), %%mm1          \n\t"
 507         "movq 8(%1, %3), %%mm5          \n\t"
 508         "movq     %%mm0,  (%2)          \n\t"
 509         "movq     %%mm4, 8(%2)          \n\t"
 510         "movq     %%mm1,  (%2, %3)      \n\t"
 511         "movq     %%mm5, 8(%2, %3)      \n\t"
 512         "add  %%"REG_a", %1             \n\t"
 513         "add  %%"REG_a", %2             \n\t"
 514         "movq  (%1    ), %%mm0          \n\t"
 515         "movq 8(%1    ), %%mm4          \n\t"
 516         "movq  (%1, %3), %%mm1          \n\t"
 517         "movq 8(%1, %3), %%mm5          \n\t"
 518         "movq     %%mm0,  (%2)          \n\t"
 519         "movq     %%mm4, 8(%2)          \n\t"
 520         "movq     %%mm1,  (%2, %3)      \n\t"
 521         "movq     %%mm5, 8(%2, %3)      \n\t"
 522         "add  %%"REG_a", %1             \n\t"
 523         "add  %%"REG_a", %2             \n\t"
 524         "subl        $4, %0             \n\t"
 525         "jnz         1b                 \n\t"
 526         : "+g"(h), "+r"(pixels),  "+r"(block)
 527         : "r"((x86_reg)line_size)
 528         : "%"REG_a, "memory"
 529         );
 530 }
 531
 532 #define CLEAR_BLOCKS(name, n)                           \
 533 static void name(int16_t *blocks)                       \
 534 {                                                       \
 535     __asm__ volatile (                                  \
 536         "pxor %%mm7, %%mm7              \n\t"           \
 537         "mov     %1,        %%"REG_a"   \n\t"           \
 538         "1:                             \n\t"           \
 539         "movq %%mm7,   (%0, %%"REG_a")  \n\t"           \
 540         "movq %%mm7,  8(%0, %%"REG_a")  \n\t"           \
 541         "movq %%mm7, 16(%0, %%"REG_a")  \n\t"           \
 542         "movq %%mm7, 24(%0, %%"REG_a")  \n\t"           \
 543         "add    $32, %%"REG_a"          \n\t"           \
 544         "js      1b                     \n\t"           \
 545         :: "r"(((uint8_t *)blocks) + 128 * n),          \
 546            "i"(-128 * n)                                \
 547         : "%"REG_a                                      \
 548         );                                              \
 549 }
 550 CLEAR_BLOCKS(clear_blocks_mmx, 6)
 551 CLEAR_BLOCKS(clear_block_mmx, 1)
 552
 553 static void clear_block_sse(int16_t *block)
 554 {
 555     __asm__ volatile (
 556         "xorps  %%xmm0, %%xmm0          \n"
 557         "movaps %%xmm0,    (%0)         \n"
 558         "movaps %%xmm0,  16(%0)         \n"
 559         "movaps %%xmm0,  32(%0)         \n"
 560         "movaps %%xmm0,  48(%0)         \n"
 561         "movaps %%xmm0,  64(%0)         \n"
 562         "movaps %%xmm0,  80(%0)         \n"
 563         "movaps %%xmm0,  96(%0)         \n"
 564         "movaps %%xmm0, 112(%0)         \n"
 565         :: "r"(block)
 566         : "memory"
 567     );
 568 }
 569
 570 static void clear_blocks_sse(int16_t *blocks)
 571 {
 572     __asm__ volatile (
 573         "xorps  %%xmm0, %%xmm0              \n"
 574         "mov        %1,         %%"REG_a"   \n"
 575         "1:                                 \n"
 576         "movaps %%xmm0,    (%0, %%"REG_a")  \n"
 577         "movaps %%xmm0,  16(%0, %%"REG_a")  \n"
 578         "movaps %%xmm0,  32(%0, %%"REG_a")  \n"
 579         "movaps %%xmm0,  48(%0, %%"REG_a")  \n"
 580         "movaps %%xmm0,  64(%0, %%"REG_a")  \n"
 581         "movaps %%xmm0,  80(%0, %%"REG_a")  \n"
 582         "movaps %%xmm0,  96(%0, %%"REG_a")  \n"
 583         "movaps %%xmm0, 112(%0, %%"REG_a")  \n"
 584         "add      $128,         %%"REG_a"   \n"
 585         "js         1b                      \n"
 586         :: "r"(((uint8_t *)blocks) + 128 * 6),
 587            "i"(-128 * 6)
 588         : "%"REG_a
 589     );
 590 }
 591
 592 static void add_bytes_mmx(uint8_t *dst, uint8_t *src, int w)
 593 {
 594     x86_reg i = 0;
 595     __asm__ volatile (
 596         "jmp          2f                \n\t"
 597         "1:                             \n\t"
 598         "movq   (%1, %0), %%mm0         \n\t"
 599         "movq   (%2, %0), %%mm1         \n\t"
 600         "paddb     %%mm0, %%mm1         \n\t"
 601         "movq      %%mm1, (%2, %0)      \n\t"
 602         "movq  8(%1, %0), %%mm0         \n\t"
 603         "movq  8(%2, %0), %%mm1         \n\t"
 604         "paddb     %%mm0, %%mm1         \n\t"
 605         "movq      %%mm1, 8(%2, %0)     \n\t"
 606         "add         $16, %0            \n\t"
 607         "2:                             \n\t"
 608         "cmp          %3, %0            \n\t"
 609         "js           1b                \n\t"
 610         : "+r"(i)
 611         : "r"(src), "r"(dst), "r"((x86_reg)w - 15)
 612     );
 613     for ( ; i < w; i++)
 614         dst[i + 0] += src[i + 0];
 615 }
 616
 617 #if HAVE_7REGS
 618 static void add_hfyu_median_prediction_cmov(uint8_t *dst, const uint8_t *top,
 619                                             const uint8_t *diff, int w,
 620                                             int *left, int *left_top)
 621 {
 622     x86_reg w2 = -w;
 623     x86_reg x;
 624     int l  = *left     & 0xff;
 625     int tl = *left_top & 0xff;
 626     int t;
 627     __asm__ volatile (
 628         "mov          %7, %3            \n"
 629         "1:                             \n"
 630         "movzbl (%3, %4), %2            \n"
 631         "mov          %2, %k3           \n"
 632         "sub         %b1, %b3           \n"
 633         "add         %b0, %b3           \n"
 634         "mov          %2, %1            \n"
 635         "cmp          %0, %2            \n"
 636         "cmovg        %0, %2            \n"
 637         "cmovg        %1, %0            \n"
 638         "cmp         %k3, %0            \n"
 639         "cmovg       %k3, %0            \n"
 640         "mov          %7, %3            \n"
 641         "cmp          %2, %0            \n"
 642         "cmovl        %2, %0            \n"
 643         "add    (%6, %4), %b0           \n"
 644         "mov         %b0, (%5, %4)      \n"
 645         "inc          %4                \n"
 646         "jl           1b                \n"
 647         : "+&q"(l), "+&q"(tl), "=&r"(t), "=&q"(x), "+&r"(w2)
 648         : "r"(dst + w), "r"(diff + w), "rm"(top + w)
 649     );
 650     *left     = l;
 651     *left_top = tl;
 652 }
 653 #endif
 654
 655 static inline void transpose4x4(uint8_t *dst, uint8_t *src, x86_reg dst_stride, x86_reg src_stride){
 656     __asm__ volatile( //FIXME could save 1 instruction if done as 8x4 ...
 657         "movd  (%1), %%mm0              \n\t"
 658         "add   %3, %1                   \n\t"
 659         "movd  (%1), %%mm1              \n\t"
 660         "movd  (%1,%3,1), %%mm2         \n\t"
 661         "movd  (%1,%3,2), %%mm3         \n\t"
 662         "punpcklbw %%mm1, %%mm0         \n\t"
 663         "punpcklbw %%mm3, %%mm2         \n\t"
 664         "movq %%mm0, %%mm1              \n\t"
 665         "punpcklwd %%mm2, %%mm0         \n\t"
 666         "punpckhwd %%mm2, %%mm1         \n\t"
 667         "movd  %%mm0, (%0)              \n\t"
 668         "add   %2, %0                   \n\t"
 669         "punpckhdq %%mm0, %%mm0         \n\t"
 670         "movd  %%mm0, (%0)              \n\t"
 671         "movd  %%mm1, (%0,%2,1)         \n\t"
 672         "punpckhdq %%mm1, %%mm1         \n\t"
 673         "movd  %%mm1, (%0,%2,2)         \n\t"
 674
 675         :  "+&r" (dst),
 676            "+&r" (src)
 677         :  "r" (dst_stride),
 678            "r" (src_stride)
 679         :  "memory"
 680     );
 681 }
 682
 683 #define H263_LOOP_FILTER                        \
 684     "pxor      %%mm7, %%mm7             \n\t"   \
 685     "movq         %0, %%mm0             \n\t"   \
 686     "movq         %0, %%mm1             \n\t"   \
 687     "movq         %3, %%mm2             \n\t"   \
 688     "movq         %3, %%mm3             \n\t"   \
 689     "punpcklbw %%mm7, %%mm0             \n\t"   \
 690     "punpckhbw %%mm7, %%mm1             \n\t"   \
 691     "punpcklbw %%mm7, %%mm2             \n\t"   \
 692     "punpckhbw %%mm7, %%mm3             \n\t"   \
 693     "psubw     %%mm2, %%mm0             \n\t"   \
 694     "psubw     %%mm3, %%mm1             \n\t"   \
 695     "movq         %1, %%mm2             \n\t"   \
 696     "movq         %1, %%mm3             \n\t"   \
 697     "movq         %2, %%mm4             \n\t"   \
 698     "movq         %2, %%mm5             \n\t"   \
 699     "punpcklbw %%mm7, %%mm2             \n\t"   \
 700     "punpckhbw %%mm7, %%mm3             \n\t"   \
 701     "punpcklbw %%mm7, %%mm4             \n\t"   \
 702     "punpckhbw %%mm7, %%mm5             \n\t"   \
 703     "psubw     %%mm2, %%mm4             \n\t"   \
 704     "psubw     %%mm3, %%mm5             \n\t"   \
 705     "psllw        $2, %%mm4             \n\t"   \
 706     "psllw        $2, %%mm5             \n\t"   \
 707     "paddw     %%mm0, %%mm4             \n\t"   \
 708     "paddw     %%mm1, %%mm5             \n\t"   \
 709     "pxor      %%mm6, %%mm6             \n\t"   \
 710     "pcmpgtw   %%mm4, %%mm6             \n\t"   \
 711     "pcmpgtw   %%mm5, %%mm7             \n\t"   \
 712     "pxor      %%mm6, %%mm4             \n\t"   \
 713     "pxor      %%mm7, %%mm5             \n\t"   \
 714     "psubw     %%mm6, %%mm4             \n\t"   \
 715     "psubw     %%mm7, %%mm5             \n\t"   \
 716     "psrlw        $3, %%mm4             \n\t"   \
 717     "psrlw        $3, %%mm5             \n\t"   \
 718     "packuswb  %%mm5, %%mm4             \n\t"   \
 719     "packsswb  %%mm7, %%mm6             \n\t"   \
 720     "pxor      %%mm7, %%mm7             \n\t"   \
 721     "movd         %4, %%mm2             \n\t"   \
 722     "punpcklbw %%mm2, %%mm2             \n\t"   \
 723     "punpcklbw %%mm2, %%mm2             \n\t"   \
 724     "punpcklbw %%mm2, %%mm2             \n\t"   \
 725     "psubusb   %%mm4, %%mm2             \n\t"   \
 726     "movq      %%mm2, %%mm3             \n\t"   \
 727     "psubusb   %%mm4, %%mm3             \n\t"   \
 728     "psubb     %%mm3, %%mm2             \n\t"   \
 729     "movq         %1, %%mm3             \n\t"   \
 730     "movq         %2, %%mm4             \n\t"   \
 731     "pxor      %%mm6, %%mm3             \n\t"   \
 732     "pxor      %%mm6, %%mm4             \n\t"   \
 733     "paddusb   %%mm2, %%mm3             \n\t"   \
 734     "psubusb   %%mm2, %%mm4             \n\t"   \
 735     "pxor      %%mm6, %%mm3             \n\t"   \
 736     "pxor      %%mm6, %%mm4             \n\t"   \
 737     "paddusb   %%mm2, %%mm2             \n\t"   \
 738     "packsswb  %%mm1, %%mm0             \n\t"   \
 739     "pcmpgtb   %%mm0, %%mm7             \n\t"   \
 740     "pxor      %%mm7, %%mm0             \n\t"   \
 741     "psubb     %%mm7, %%mm0             \n\t"   \
 742     "movq      %%mm0, %%mm1             \n\t"   \
 743     "psubusb   %%mm2, %%mm0             \n\t"   \
 744     "psubb     %%mm0, %%mm1             \n\t"   \
 745     "pand         %5, %%mm1             \n\t"   \
 746     "psrlw        $2, %%mm1             \n\t"   \
 747     "pxor      %%mm7, %%mm1             \n\t"   \
 748     "psubb     %%mm7, %%mm1             \n\t"   \
 749     "movq         %0, %%mm5             \n\t"   \
 750     "movq         %3, %%mm6             \n\t"   \
 751     "psubb     %%mm1, %%mm5             \n\t"   \
 752     "paddb     %%mm1, %%mm6             \n\t"
 753
 754 static void h263_v_loop_filter_mmx(uint8_t *src, int stride, int qscale)
 755 {
 756     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
 757         const int strength = ff_h263_loop_filter_strength[qscale];
 758
 759         __asm__ volatile (
 760             H263_LOOP_FILTER
 761
 762             "movq %%mm3, %1             \n\t"
 763             "movq %%mm4, %2             \n\t"
 764             "movq %%mm5, %0             \n\t"
 765             "movq %%mm6, %3             \n\t"
 766             : "+m"(*(uint64_t*)(src - 2 * stride)),
 767               "+m"(*(uint64_t*)(src - 1 * stride)),
 768               "+m"(*(uint64_t*)(src + 0 * stride)),
 769               "+m"(*(uint64_t*)(src + 1 * stride))
 770             : "g"(2 * strength), "m"(ff_pb_FC)
 771             );
 772     }
 773 }
 774
 775 static void h263_h_loop_filter_mmx(uint8_t *src, int stride, int qscale)
 776 {
 777     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
 778         const int strength = ff_h263_loop_filter_strength[qscale];
 779         DECLARE_ALIGNED(8, uint64_t, temp)[4];
 780         uint8_t *btemp = (uint8_t*)temp;
 781
 782         src -= 2;
 783
 784         transpose4x4(btemp,     src,              8, stride);
 785         transpose4x4(btemp + 4, src + 4 * stride, 8, stride);
 786         __asm__ volatile (
 787             H263_LOOP_FILTER // 5 3 4 6
 788
 789             : "+m"(temp[0]),
 790               "+m"(temp[1]),
 791               "+m"(temp[2]),
 792               "+m"(temp[3])
 793             : "g"(2 * strength), "m"(ff_pb_FC)
 794             );
 795
 796         __asm__ volatile (
 797             "movq      %%mm5, %%mm1         \n\t"
 798             "movq      %%mm4, %%mm0         \n\t"
 799             "punpcklbw %%mm3, %%mm5         \n\t"
 800             "punpcklbw %%mm6, %%mm4         \n\t"
 801             "punpckhbw %%mm3, %%mm1         \n\t"
 802             "punpckhbw %%mm6, %%mm0         \n\t"
 803             "movq      %%mm5, %%mm3         \n\t"
 804             "movq      %%mm1, %%mm6         \n\t"
 805             "punpcklwd %%mm4, %%mm5         \n\t"
 806             "punpcklwd %%mm0, %%mm1         \n\t"
 807             "punpckhwd %%mm4, %%mm3         \n\t"
 808             "punpckhwd %%mm0, %%mm6         \n\t"
 809             "movd      %%mm5, (%0)          \n\t"
 810             "punpckhdq %%mm5, %%mm5         \n\t"
 811             "movd      %%mm5, (%0, %2)      \n\t"
 812             "movd      %%mm3, (%0, %2, 2)   \n\t"
 813             "punpckhdq %%mm3, %%mm3         \n\t"
 814             "movd      %%mm3, (%0, %3)      \n\t"
 815             "movd      %%mm1, (%1)          \n\t"
 816             "punpckhdq %%mm1, %%mm1         \n\t"
 817             "movd      %%mm1, (%1, %2)      \n\t"
 818             "movd      %%mm6, (%1, %2, 2)   \n\t"
 819             "punpckhdq %%mm6, %%mm6         \n\t"
 820             "movd      %%mm6, (%1, %3)      \n\t"
 821             :: "r"(src),
 822                "r"(src + 4 * stride),
 823                "r"((x86_reg)stride),
 824                "r"((x86_reg)(3 * stride))
 825             );
 826     }
 827 }
 828
 829 /* Draw the edges of width 'w' of an image of size width, height
 830  * this MMX version can only handle w == 8 || w == 16. */
 831 static void draw_edges_mmx(uint8_t *buf, int wrap, int width, int height,
 832                            int w, int h, int sides)
 833 {
 834     uint8_t *ptr, *last_line;
 835     int i;
 836
 837     last_line = buf + (height - 1) * wrap;
 838     /* left and right */
 839     ptr = buf;
 840     if (w == 8) {
 841         __asm__ volatile (
 842             "1:                             \n\t"
 843             "movd            (%0), %%mm0    \n\t"
 844             "punpcklbw      %%mm0, %%mm0    \n\t"
 845             "punpcklwd      %%mm0, %%mm0    \n\t"
 846             "punpckldq      %%mm0, %%mm0    \n\t"
 847             "movq           %%mm0, -8(%0)   \n\t"
 848             "movq      -8(%0, %2), %%mm1    \n\t"
 849             "punpckhbw      %%mm1, %%mm1    \n\t"
 850             "punpckhwd      %%mm1, %%mm1    \n\t"
 851             "punpckhdq      %%mm1, %%mm1    \n\t"
 852             "movq           %%mm1, (%0, %2) \n\t"
 853             "add               %1, %0       \n\t"
 854             "cmp               %3, %0       \n\t"
 855             "jb                1b           \n\t"
 856             : "+r"(ptr)
 857             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
 858             );
 859     } else if(w==16){
 860         __asm__ volatile (
 861             "1:                                 \n\t"
 862             "movd            (%0), %%mm0        \n\t"
 863             "punpcklbw      %%mm0, %%mm0        \n\t"
 864             "punpcklwd      %%mm0, %%mm0        \n\t"
 865             "punpckldq      %%mm0, %%mm0        \n\t"
 866             "movq           %%mm0, -8(%0)       \n\t"
 867             "movq           %%mm0, -16(%0)      \n\t"
 868             "movq      -8(%0, %2), %%mm1        \n\t"
 869             "punpckhbw      %%mm1, %%mm1        \n\t"
 870             "punpckhwd      %%mm1, %%mm1        \n\t"
 871             "punpckhdq      %%mm1, %%mm1        \n\t"
 872             "movq           %%mm1,  (%0, %2)    \n\t"
 873             "movq           %%mm1, 8(%0, %2)    \n\t"
 874             "add               %1, %0           \n\t"
 875             "cmp               %3, %0           \n\t"
 876             "jb                1b               \n\t"
 877             : "+r"(ptr)
 878             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
 879             );
 880     } else {
 881         av_assert1(w == 4);
 882         __asm__ volatile (
 883             "1:                             \n\t"
 884             "movd            (%0), %%mm0    \n\t"
 885             "punpcklbw      %%mm0, %%mm0    \n\t"
 886             "punpcklwd      %%mm0, %%mm0    \n\t"
 887             "movd           %%mm0, -4(%0)   \n\t"
 888             "movd      -4(%0, %2), %%mm1    \n\t"
 889             "punpcklbw      %%mm1, %%mm1    \n\t"
 890             "punpckhwd      %%mm1, %%mm1    \n\t"
 891             "punpckhdq      %%mm1, %%mm1    \n\t"
 892             "movd           %%mm1, (%0, %2) \n\t"
 893             "add               %1, %0       \n\t"
 894             "cmp               %3, %0       \n\t"
 895             "jb                1b           \n\t"
 896             : "+r"(ptr)
 897             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
 898             );
 899     }
 900
 901     /* top and bottom (and hopefully also the corners) */
 902     if (sides & EDGE_TOP) {
 903         for (i = 0; i < h; i += 4) {
 904             ptr = buf - (i + 1) * wrap - w;
 905             __asm__ volatile (
 906                 "1:                             \n\t"
 907                 "movq (%1, %0), %%mm0           \n\t"
 908                 "movq    %%mm0, (%0)            \n\t"
 909                 "movq    %%mm0, (%0, %2)        \n\t"
 910                 "movq    %%mm0, (%0, %2, 2)     \n\t"
 911                 "movq    %%mm0, (%0, %3)        \n\t"
 912                 "add        $8, %0              \n\t"
 913                 "cmp        %4, %0              \n\t"
 914                 "jb         1b                  \n\t"
 915                 : "+r"(ptr)
 916                 : "r"((x86_reg)buf - (x86_reg)ptr - w), "r"((x86_reg) -wrap),
 917                   "r"((x86_reg) -wrap * 3), "r"(ptr + width + 2 * w)
 918                 );
 919         }
 920     }
 921
 922     if (sides & EDGE_BOTTOM) {
 923         for (i = 0; i < h; i += 4) {
 924             ptr = last_line + (i + 1) * wrap - w;
 925             __asm__ volatile (
 926                 "1:                             \n\t"
 927                 "movq (%1, %0), %%mm0           \n\t"
 928                 "movq    %%mm0, (%0)            \n\t"
 929                 "movq    %%mm0, (%0, %2)        \n\t"
 930                 "movq    %%mm0, (%0, %2, 2)     \n\t"
 931                 "movq    %%mm0, (%0, %3)        \n\t"
 932                 "add        $8, %0              \n\t"
 933                 "cmp        %4, %0              \n\t"
 934                 "jb         1b                  \n\t"
 935                 : "+r"(ptr)
 936                 : "r"((x86_reg)last_line - (x86_reg)ptr - w),
 937                   "r"((x86_reg)wrap), "r"((x86_reg)wrap * 3),
 938                   "r"(ptr + width + 2 * w)
 939                 );
 940         }
 941     }
 942 }
 943 #endif /* HAVE_INLINE_ASM */
 944
 945
 946 #if HAVE_YASM
 947 #define QPEL_OP(OPNAME, ROUNDER, RND, OP, MMX)                          \
 948 static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src,   \
 949                                           int stride)                   \
 950 {                                                                       \
 951     ff_ ## OPNAME ## pixels8_ ## MMX(dst, src, stride, 8);              \
 952 }                                                                       \
 953                                                                         \
 954 static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src,    \
 955                                          int stride)                    \
 956 {                                                                       \
 957     uint64_t temp[8];                                                   \
 958     uint8_t * const half = (uint8_t*)temp;                              \
 959     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8,        \
 960                                                    stride, 8);          \
 961     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src, half,                 \
 962                                         stride, stride, 8);             \
 963 }                                                                       \
 964                                                                         \
 965 static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src,    \
 966                                          int stride)                    \
 967 {                                                                       \
 968     ff_ ## OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride,    \
 969                                                    stride, 8);          \
 970 }                                                                       \
 971                                                                         \
 972 static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src,    \
 973                                          int stride)                    \
 974 {                                                                       \
 975     uint64_t temp[8];                                                   \
 976     uint8_t * const half = (uint8_t*)temp;                              \
 977     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8,        \
 978                                                    stride, 8);          \
 979     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src + 1, half, stride,     \
 980                                         stride, 8);                     \
 981 }                                                                       \
 982                                                                         \
 983 static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src,    \
 984                                          int stride)                    \
 985 {                                                                       \
 986     uint64_t temp[8];                                                   \
 987     uint8_t * const half = (uint8_t*)temp;                              \
 988     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src,           \
 989                                                    8, stride);          \
 990     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src, half,                 \
 991                                         stride, stride, 8);             \
 992 }                                                                       \
 993                                                                         \
 994 static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src,    \
 995                                          int stride)                    \
 996 {                                                                       \
 997     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src,            \
 998                                                    stride, stride);     \
 999 }                                                                       \
1000                                                                         \
1001 static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src,    \
1002                                          int stride)                    \
1003 {                                                                       \
1004     uint64_t temp[8];                                                   \
1005     uint8_t * const half = (uint8_t*)temp;                              \
1006     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src,           \
1007                                                    8, stride);          \
1008     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src + stride, half, stride,\
1009                                         stride, 8);                     \
1010 }                                                                       \
1011                                                                         \
1012 static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src,    \
1013                                          int stride)                    \
1014 {                                                                       \
1015     uint64_t half[8 + 9];                                               \
1016     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1017     uint8_t * const halfHV = ((uint8_t*)half);                          \
1018     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
1019                                                    stride, 9);          \
1020     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8,           \
1021                                         stride, 9);                     \
1022     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
1023     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV,             \
1024                                         stride, 8, 8);                  \
1025 }                                                                       \
1026                                                                         \
1027 static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src,    \
1028                                          int stride)                    \
1029 {                                                                       \
1030     uint64_t half[8 + 9];                                               \
1031     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1032     uint8_t * const halfHV = ((uint8_t*)half);                          \
1033     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
1034                                                    stride, 9);          \
1035     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,       \
1036                                         stride, 9);                     \
1037     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
1038     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV,             \
1039                                         stride, 8, 8);                  \
1040 }                                                                       \
1041                                                                         \
1042 static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src,    \
1043                                          int stride)                    \
1044 {                                                                       \
1045     uint64_t half[8 + 9];                                               \
1046     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1047     uint8_t * const halfHV = ((uint8_t*)half);                          \
1048     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
1049                                                    stride, 9);          \
1050     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8,           \
1051                                         stride, 9);                     \
1052     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
1053     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV,         \
1054                                         stride, 8, 8);                  \
1055 }                                                                       \
1056                                                                         \
1057 static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src,    \
1058                                          int stride)                    \
1059 {                                                                       \
1060     uint64_t half[8 + 9];                                               \
1061     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1062     uint8_t * const halfHV = ((uint8_t*)half);                          \
1063     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
1064                                                    stride, 9);          \
1065     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,       \
1066                                         stride, 9);                     \
1067     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
1068     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV,         \
1069                                         stride, 8, 8);                  \
1070 }                                                                       \
1071                                                                         \
1072 static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src,    \
1073                                          int stride)                    \
1074 {                                                                       \
1075     uint64_t half[8 + 9];                                               \
1076     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1077     uint8_t * const halfHV = ((uint8_t*)half);                          \
1078     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
1079                                                    stride, 9);          \
1080     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
1081     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV,             \
1082                                         stride, 8, 8);                  \
1083 }                                                                       \
1084                                                                         \
1085 static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src,    \
1086                                          int stride)                    \
1087 {                                                                       \
1088     uint64_t half[8 + 9];                                               \
1089     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
1090     uint8_t * const halfHV = ((uint8_t*)half);                          \
1091     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
1092                                                    stride, 9);          \
1093     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
1094     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV,         \
1095                                         stride, 8, 8);                  \
1096 }                                                                       \
1097                                                                         \
1098 static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src,    \
1099                                          int stride)                    \
1100 {                                                                       \
1101     uint64_t half[8 + 9];                                               \
1102     uint8_t * const halfH = ((uint8_t*)half);                           \
1103     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
1104                                                    stride, 9);          \
1105     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH,              \
1106                                         8, stride, 9);                  \
1107     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH,          \
1108                                                    stride, 8);          \
1109 }                                                                       \
1110                                                                         \
1111 static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src,    \
1112                                          int stride)                    \
1113 {                                                                       \
1114     uint64_t half[8 + 9];                                               \
1115     uint8_t * const halfH = ((uint8_t*)half);                           \
1116     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
1117                                                    stride, 9);          \
1118     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,       \
1119                                         stride, 9);                     \
1120     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH,          \
1121                                                    stride, 8);          \
1122 }                                                                       \
1123                                                                         \
1124 static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src,    \
1125                                          int stride)                    \
1126 {                                                                       \
1127     uint64_t half[9];                                                   \
1128     uint8_t * const halfH = ((uint8_t*)half);                           \
1129     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
1130                                                    stride, 9);          \
1131     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH,          \
1132                                                    stride, 8);          \
1133 }                                                                       \
1134                                                                         \
1135 static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src,  \
1136                                            int stride)                  \
1137 {                                                                       \
1138     ff_ ## OPNAME ## pixels16_ ## MMX(dst, src, stride, 16);            \
1139 }                                                                       \
1140                                                                         \
1141 static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src,   \
1142                                           int stride)                   \
1143 {                                                                       \
1144     uint64_t temp[32];                                                  \
1145     uint8_t * const half = (uint8_t*)temp;                              \
1146     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16,      \
1147                                                     stride, 16);        \
1148     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride,        \
1149                                          stride, 16);                   \
1150 }                                                                       \
1151                                                                         \
1152 static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src,   \
1153                                           int stride)                   \
1154 {                                                                       \
1155     ff_ ## OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src,           \
1156                                                     stride, stride, 16);\
1157 }                                                                       \
1158                                                                         \
1159 static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src,   \
1160                                           int stride)                   \
1161 {                                                                       \
1162     uint64_t temp[32];                                                  \
1163     uint8_t * const half = (uint8_t*)temp;                              \
1164     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16,      \
1165                                                     stride, 16);        \
1166     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src + 1, half,            \
1167                                          stride, stride, 16);           \
1168 }                                                                       \
1169                                                                         \
1170 static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src,   \
1171                                           int stride)                   \
1172 {                                                                       \
1173     uint64_t temp[32];                                                  \
1174     uint8_t * const half = (uint8_t*)temp;                              \
1175     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16,      \
1176                                                     stride);            \
1177     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride,        \
1178                                          stride, 16);                   \
1179 }                                                                       \
1180                                                                         \
1181 static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src,   \
1182                                           int stride)                   \
1183 {                                                                       \
1184     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src,           \
1185                                                     stride, stride);    \
1186 }                                                                       \
1187                                                                         \
1188 static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src,   \
1189                                           int stride)                   \
1190 {                                                                       \
1191     uint64_t temp[32];                                                  \
1192     uint8_t * const half = (uint8_t*)temp;                              \
1193     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16,      \
1194                                                     stride);            \
1195     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half,         \
1196                                          stride, stride, 16);           \
1197 }                                                                       \
1198                                                                         \
1199 static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src,   \
1200                                           int stride)                   \
1201 {                                                                       \
1202     uint64_t half[16 * 2 + 17 * 2];                                     \
1203     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1204     uint8_t * const halfHV = ((uint8_t*)half);                          \
1205     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1206                                                     stride, 17);        \
1207     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,         \
1208                                          stride, 17);                   \
1209     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1210                                                     16, 16);            \
1211     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV,            \
1212                                          stride, 16, 16);               \
1213 }                                                                       \
1214                                                                         \
1215 static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src,   \
1216                                           int stride)                   \
1217 {                                                                       \
1218     uint64_t half[16 * 2 + 17 * 2];                                     \
1219     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1220     uint8_t * const halfHV = ((uint8_t*)half);                          \
1221     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1222                                                     stride, 17);        \
1223     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,     \
1224                                          stride, 17);                   \
1225     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1226                                                     16, 16);            \
1227     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV,            \
1228                                          stride, 16, 16);               \
1229 }                                                                       \
1230                                                                         \
1231 static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src,   \
1232                                           int stride)                   \
1233 {                                                                       \
1234     uint64_t half[16 * 2 + 17 * 2];                                     \
1235     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1236     uint8_t * const halfHV = ((uint8_t*)half);                          \
1237     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1238                                                     stride, 17);        \
1239     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,         \
1240                                          stride, 17);                   \
1241     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1242                                                     16, 16);            \
1243     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV,       \
1244                                          stride, 16, 16);               \
1245 }                                                                       \
1246                                                                         \
1247 static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src,   \
1248                                           int stride)                   \
1249 {                                                                       \
1250     uint64_t half[16 * 2 + 17 * 2];                                     \
1251     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1252     uint8_t * const halfHV = ((uint8_t*)half);                          \
1253     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1254                                                     stride, 17);        \
1255     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,     \
1256                                          stride, 17);                   \
1257     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1258                                                     16, 16);            \
1259     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV,       \
1260                                          stride, 16, 16);               \
1261 }                                                                       \
1262                                                                         \
1263 static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src,   \
1264                                           int stride)                   \
1265 {                                                                       \
1266     uint64_t half[16 * 2 + 17 * 2];                                     \
1267     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1268     uint8_t * const halfHV = ((uint8_t*)half);                          \
1269     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1270                                                     stride, 17);        \
1271     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1272                                                     16, 16);            \
1273     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV,            \
1274                                          stride, 16, 16);               \
1275 }                                                                       \
1276                                                                         \
1277 static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src,   \
1278                                           int stride)                   \
1279 {                                                                       \
1280     uint64_t half[16 * 2 + 17 * 2];                                     \
1281     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1282     uint8_t * const halfHV = ((uint8_t*)half);                          \
1283     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1284                                                     stride, 17);        \
1285     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1286                                                     16, 16);            \
1287     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV,       \
1288                                          stride, 16, 16);               \
1289 }                                                                       \
1290                                                                         \
1291 static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src,   \
1292                                           int stride)                   \
1293 {                                                                       \
1294     uint64_t half[17 * 2];                                              \
1295     uint8_t * const halfH = ((uint8_t*)half);                           \
1296     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1297                                                     stride, 17);        \
1298     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,         \
1299                                          stride, 17);                   \
1300     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH,         \
1301                                                     stride, 16);        \
1302 }                                                                       \
1303                                                                         \
1304 static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src,   \
1305                                           int stride)                   \
1306 {                                                                       \
1307     uint64_t half[17 * 2];                                              \
1308     uint8_t * const halfH = ((uint8_t*)half);                           \
1309     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1310                                                     stride, 17);        \
1311     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,     \
1312                                          stride, 17);                   \
1313     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH,         \
1314                                                     stride, 16);        \
1315 }                                                                       \
1316                                                                         \
1317 static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src,   \
1318                                           int stride)                   \
1319 {                                                                       \
1320     uint64_t half[17 * 2];                                              \
1321     uint8_t * const halfH = ((uint8_t*)half);                           \
1322     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1323                                                     stride, 17);        \
1324     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH,         \
1325                                                     stride, 16);        \
1326 }
1327
1328 #define PUT_OP(a, b, temp, size)                \
1329     "mov"#size"        "#a", "#b"       \n\t"
1330
1331 #define AVG_MMXEXT_OP(a, b, temp, size)         \
1332     "mov"#size"        "#b", "#temp"    \n\t"   \
1333     "pavgb          "#temp", "#a"       \n\t"   \
1334     "mov"#size"        "#a", "#b"       \n\t"
1335
1336 QPEL_OP(put_,          ff_pw_16, _,        PUT_OP,        mmxext)
1337 QPEL_OP(avg_,          ff_pw_16, _,        AVG_MMXEXT_OP, mmxext)
1338 QPEL_OP(put_no_rnd_,   ff_pw_15, _no_rnd_, PUT_OP,        mmxext)
1339 #endif /* HAVE_YASM */
1340
1341
1342 #if HAVE_INLINE_ASM
1343 void ff_put_rv40_qpel8_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1344 {
1345   put_pixels8_xy2_mmx(dst, src, stride, 8);
1346 }
1347 void ff_put_rv40_qpel16_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1348 {
1349   put_pixels16_xy2_mmx(dst, src, stride, 16);
1350 }
1351 void ff_avg_rv40_qpel8_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1352 {
1353   avg_pixels8_xy2_mmx(dst, src, stride, 8);
1354 }
1355 void ff_avg_rv40_qpel16_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1356 {
1357   avg_pixels16_xy2_mmx(dst, src, stride, 16);
1358 }
1359
1360 typedef void emulated_edge_mc_func(uint8_t *dst, const uint8_t *src,
1361                                    ptrdiff_t linesize, int block_w, int block_h,
1362                                    int src_x, int src_y, int w, int h);
1363
1364 static av_always_inline void gmc(uint8_t *dst, uint8_t *src,
1365                                  int stride, int h, int ox, int oy,
1366                                  int dxx, int dxy, int dyx, int dyy,
1367                                  int shift, int r, int width, int height,
1368                                  emulated_edge_mc_func *emu_edge_fn)
1369 {
1370     const int w    = 8;
1371     const int ix   = ox  >> (16 + shift);
1372     const int iy   = oy  >> (16 + shift);
1373     const int oxs  = ox  >> 4;
1374     const int oys  = oy  >> 4;
1375     const int dxxs = dxx >> 4;
1376     const int dxys = dxy >> 4;
1377     const int dyxs = dyx >> 4;
1378     const int dyys = dyy >> 4;
1379     const uint16_t r4[4]   = { r, r, r, r };
1380     const uint16_t dxy4[4] = { dxys, dxys, dxys, dxys };
1381     const uint16_t dyy4[4] = { dyys, dyys, dyys, dyys };
1382     const uint64_t shift2 = 2 * shift;
1383 #define MAX_STRIDE 4096U
1384 #define MAX_H 8U
1385     uint8_t edge_buf[(MAX_H + 1) * MAX_STRIDE];
1386     int x, y;
1387
1388     const int dxw = (dxx - (1 << (16 + shift))) * (w - 1);
1389     const int dyh = (dyy - (1 << (16 + shift))) * (h - 1);
1390     const int dxh = dxy * (h - 1);
1391     const int dyw = dyx * (w - 1);
1392     int need_emu =  (unsigned)ix >= width  - w ||
1393                     (unsigned)iy >= height - h;
1394
1395     if ( // non-constant fullpel offset (3% of blocks)
1396         ((ox ^ (ox + dxw)) | (ox ^ (ox + dxh)) | (ox ^ (ox + dxw + dxh)) |
1397          (oy ^ (oy + dyw)) | (oy ^ (oy + dyh)) | (oy ^ (oy + dyw + dyh))) >> (16 + shift)
1398         // uses more than 16 bits of subpel mv (only at huge resolution)
1399         || (dxx | dxy | dyx | dyy) & 15
1400         || (need_emu && (h > MAX_H || stride > MAX_STRIDE))) {
1401         // FIXME could still use mmx for some of the rows
1402         ff_gmc_c(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy,
1403                  shift, r, width, height);
1404         return;
1405     }
1406
1407     src += ix + iy * stride;
1408     if (need_emu) {
1409         emu_edge_fn(edge_buf, src, stride, w + 1, h + 1, ix, iy, width, height);
1410         src = edge_buf;
1411     }
1412
1413     __asm__ volatile (
1414         "movd         %0, %%mm6         \n\t"
1415         "pxor      %%mm7, %%mm7         \n\t"
1416         "punpcklwd %%mm6, %%mm6         \n\t"
1417         "punpcklwd %%mm6, %%mm6         \n\t"
1418         :: "r"(1<<shift)
1419     );
1420
1421     for (x = 0; x < w; x += 4) {
1422         uint16_t dx4[4] = { oxs - dxys + dxxs * (x + 0),
1423                             oxs - dxys + dxxs * (x + 1),
1424                             oxs - dxys + dxxs * (x + 2),
1425                             oxs - dxys + dxxs * (x + 3) };
1426         uint16_t dy4[4] = { oys - dyys + dyxs * (x + 0),
1427                             oys - dyys + dyxs * (x + 1),
1428                             oys - dyys + dyxs * (x + 2),
1429                             oys - dyys + dyxs * (x + 3) };
1430
1431         for (y = 0; y < h; y++) {
1432             __asm__ volatile (
1433                 "movq      %0, %%mm4    \n\t"
1434                 "movq      %1, %%mm5    \n\t"
1435                 "paddw     %2, %%mm4    \n\t"
1436                 "paddw     %3, %%mm5    \n\t"
1437                 "movq   %%mm4, %0       \n\t"
1438                 "movq   %%mm5, %1       \n\t"
1439                 "psrlw    $12, %%mm4    \n\t"
1440                 "psrlw    $12, %%mm5    \n\t"
1441                 : "+m"(*dx4), "+m"(*dy4)
1442                 : "m"(*dxy4), "m"(*dyy4)
1443             );
1444
1445             __asm__ volatile (
1446                 "movq      %%mm6, %%mm2 \n\t"
1447                 "movq      %%mm6, %%mm1 \n\t"
1448                 "psubw     %%mm4, %%mm2 \n\t"
1449                 "psubw     %%mm5, %%mm1 \n\t"
1450                 "movq      %%mm2, %%mm0 \n\t"
1451                 "movq      %%mm4, %%mm3 \n\t"
1452                 "pmullw    %%mm1, %%mm0 \n\t" // (s - dx) * (s - dy)
1453                 "pmullw    %%mm5, %%mm3 \n\t" // dx * dy
1454                 "pmullw    %%mm5, %%mm2 \n\t" // (s - dx) * dy
1455                 "pmullw    %%mm4, %%mm1 \n\t" // dx * (s - dy)
1456
1457                 "movd         %4, %%mm5 \n\t"
1458                 "movd         %3, %%mm4 \n\t"
1459                 "punpcklbw %%mm7, %%mm5 \n\t"
1460                 "punpcklbw %%mm7, %%mm4 \n\t"
1461                 "pmullw    %%mm5, %%mm3 \n\t" // src[1, 1] * dx * dy
1462                 "pmullw    %%mm4, %%mm2 \n\t" // src[0, 1] * (s - dx) * dy
1463
1464                 "movd         %2, %%mm5 \n\t"
1465                 "movd         %1, %%mm4 \n\t"
1466                 "punpcklbw %%mm7, %%mm5 \n\t"
1467                 "punpcklbw %%mm7, %%mm4 \n\t"
1468                 "pmullw    %%mm5, %%mm1 \n\t" // src[1, 0] * dx * (s - dy)
1469                 "pmullw    %%mm4, %%mm0 \n\t" // src[0, 0] * (s - dx) * (s - dy)
1470                 "paddw        %5, %%mm1 \n\t"
1471                 "paddw     %%mm3, %%mm2 \n\t"
1472                 "paddw     %%mm1, %%mm0 \n\t"
1473                 "paddw     %%mm2, %%mm0 \n\t"
1474
1475                 "psrlw        %6, %%mm0 \n\t"
1476                 "packuswb  %%mm0, %%mm0 \n\t"
1477                 "movd      %%mm0, %0    \n\t"
1478
1479                 : "=m"(dst[x + y * stride])
1480                 : "m"(src[0]), "m"(src[1]),
1481                   "m"(src[stride]), "m"(src[stride + 1]),
1482                   "m"(*r4), "m"(shift2)
1483             );
1484             src += stride;
1485         }
1486         src += 4 - h * stride;
1487     }
1488 }
1489
1490 #if CONFIG_VIDEODSP
1491 #if HAVE_YASM
1492 #if ARCH_X86_32
1493 static void gmc_mmx(uint8_t *dst, uint8_t *src,
1494                     int stride, int h, int ox, int oy,
1495                     int dxx, int dxy, int dyx, int dyy,
1496                     int shift, int r, int width, int height)
1497 {
1498     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,
1499         width, height, &ff_emulated_edge_mc_8);
1500 }
1501 #endif
1502 static void gmc_sse(uint8_t *dst, uint8_t *src,
1503                     int stride, int h, int ox, int oy,
1504                     int dxx, int dxy, int dyx, int dyy,
1505                     int shift, int r, int width, int height)
1506 {
1507     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,
1508         width, height, &ff_emulated_edge_mc_8);
1509 }
1510 #else
1511 static void gmc_mmx(uint8_t *dst, uint8_t *src,
1512                     int stride, int h, int ox, int oy,
1513                     int dxx, int dxy, int dyx, int dyy,
1514                     int shift, int r, int width, int height)
1515 {
1516     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,
1517         width, height, &ff_emulated_edge_mc_8);
1518 }
1519 #endif
1520 #endif
1521
1522 #endif /* HAVE_INLINE_ASM */
1523
1524 void ff_put_pixels16_sse2(uint8_t *block, const uint8_t *pixels,
1525                           int line_size, int h);
1526 void ff_avg_pixels16_sse2(uint8_t *block, const uint8_t *pixels,
1527                           int line_size, int h);
1528
1529 void ff_put_h264_chroma_mc8_rnd_mmx  (uint8_t *dst, uint8_t *src,
1530                                       int stride, int h, int x, int y);
1531 void ff_avg_h264_chroma_mc8_rnd_mmxext(uint8_t *dst, uint8_t *src,
1532                                        int stride, int h, int x, int y);
1533 void ff_avg_h264_chroma_mc8_rnd_3dnow(uint8_t *dst, uint8_t *src,
1534                                       int stride, int h, int x, int y);
1535
1536 void ff_put_h264_chroma_mc4_mmx      (uint8_t *dst, uint8_t *src,
1537                                       int stride, int h, int x, int y);
1538 void ff_avg_h264_chroma_mc4_mmxext   (uint8_t *dst, uint8_t *src,
1539                                       int stride, int h, int x, int y);
1540 void ff_avg_h264_chroma_mc4_3dnow    (uint8_t *dst, uint8_t *src,
1541                                       int stride, int h, int x, int y);
1542
1543 void ff_put_h264_chroma_mc2_mmxext   (uint8_t *dst, uint8_t *src,
1544                                       int stride, int h, int x, int y);
1545 void ff_avg_h264_chroma_mc2_mmxext   (uint8_t *dst, uint8_t *src,
1546                                       int stride, int h, int x, int y);
1547
1548 void ff_put_h264_chroma_mc8_rnd_ssse3(uint8_t *dst, uint8_t *src,
1549                                       int stride, int h, int x, int y);
1550 void ff_put_h264_chroma_mc4_ssse3    (uint8_t *dst, uint8_t *src,
1551                                       int stride, int h, int x, int y);
1552
1553 void ff_avg_h264_chroma_mc8_rnd_ssse3(uint8_t *dst, uint8_t *src,
1554                                       int stride, int h, int x, int y);
1555 void ff_avg_h264_chroma_mc4_ssse3    (uint8_t *dst, uint8_t *src,
1556                                       int stride, int h, int x, int y);
1557
1558 #define CHROMA_MC(OP, NUM, DEPTH, OPT)                                  \
1559 void ff_ ## OP ## _h264_chroma_mc ## NUM ## _ ## DEPTH ## _ ## OPT      \
1560                                       (uint8_t *dst, uint8_t *src,      \
1561                                        int stride, int h, int x, int y);
1562
1563 CHROMA_MC(put, 2, 10, mmxext)
1564 CHROMA_MC(avg, 2, 10, mmxext)
1565 CHROMA_MC(put, 4, 10, mmxext)
1566 CHROMA_MC(avg, 4, 10, mmxext)
1567 CHROMA_MC(put, 8, 10, sse2)
1568 CHROMA_MC(avg, 8, 10, sse2)
1569 CHROMA_MC(put, 8, 10, avx)
1570 CHROMA_MC(avg, 8, 10, avx)
1571
1572 #if HAVE_INLINE_ASM
1573
1574 /* CAVS-specific */
1575 void ff_put_cavs_qpel8_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1576 {
1577     put_pixels8_mmx(dst, src, stride, 8);
1578 }
1579
1580 void ff_avg_cavs_qpel8_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1581 {
1582     avg_pixels8_mmx(dst, src, stride, 8);
1583 }
1584
1585 void ff_put_cavs_qpel16_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1586 {
1587     put_pixels16_mmx(dst, src, stride, 16);
1588 }
1589
1590 void ff_avg_cavs_qpel16_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1591 {
1592     avg_pixels16_mmx(dst, src, stride, 16);
1593 }
1594 #endif /* HAVE_INLINE_ASM */
1595
1596 #if HAVE_YASM
1597 /* VC-1-specific */
1598 void ff_put_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src,
1599                                int stride, int rnd)
1600 {
1601     ff_put_pixels8_mmx(dst, src, stride, 8);
1602 }
1603
1604 void ff_avg_vc1_mspel_mc00_mmxext(uint8_t *dst, const uint8_t *src,
1605                                   int stride, int rnd)
1606 {
1607     ff_avg_pixels8_mmxext(dst, src, stride, 8);
1608 }
1609 #endif /* HAVE_YASM */
1610
1611 #if CONFIG_DIRAC_DECODER
1612 #define DIRAC_PIXOP(OPNAME2, OPNAME, EXT)\
1613 void ff_ ## OPNAME2 ## _dirac_pixels8_ ## EXT(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
1614 {\
1615     if (h&3)\
1616         ff_ ## OPNAME2 ## _dirac_pixels8_c(dst, src, stride, h);\
1617     else\
1618         OPNAME ## _pixels8_ ## EXT(dst, src[0], stride, h);\
1619 }\
1620 void ff_ ## OPNAME2 ## _dirac_pixels16_ ## EXT(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
1621 {\
1622     if (h&3)\
1623         ff_ ## OPNAME2 ## _dirac_pixels16_c(dst, src, stride, h);\
1624     else\
1625         OPNAME ## _pixels16_ ## EXT(dst, src[0], stride, h);\
1626 }\
1627 void ff_ ## OPNAME2 ## _dirac_pixels32_ ## EXT(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
1628 {\
1629     if (h&3) {\
1630         ff_ ## OPNAME2 ## _dirac_pixels32_c(dst, src, stride, h);\
1631     } else {\
1632         OPNAME ## _pixels16_ ## EXT(dst   , src[0]   , stride, h);\
1633         OPNAME ## _pixels16_ ## EXT(dst+16, src[0]+16, stride, h);\
1634     }\
1635 }
1636
1637 #if HAVE_MMX_INLINE
1638 DIRAC_PIXOP(put, put, mmx)
1639 DIRAC_PIXOP(avg, avg, mmx)
1640 #endif
1641
1642 #if HAVE_YASM
1643 DIRAC_PIXOP(avg, ff_avg, mmxext)
1644
1645 void ff_put_dirac_pixels16_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h)
1646 {
1647     if (h&3)
1648         ff_put_dirac_pixels16_c(dst, src, stride, h);
1649     else
1650     ff_put_pixels16_sse2(dst, src[0], stride, h);
1651 }
1652 void ff_avg_dirac_pixels16_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h)
1653 {
1654     if (h&3)
1655         ff_avg_dirac_pixels16_c(dst, src, stride, h);
1656     else
1657     ff_avg_pixels16_sse2(dst, src[0], stride, h);
1658 }
1659 void ff_put_dirac_pixels32_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h)
1660 {
1661     if (h&3) {
1662         ff_put_dirac_pixels32_c(dst, src, stride, h);
1663     } else {
1664     ff_put_pixels16_sse2(dst   , src[0]   , stride, h);
1665     ff_put_pixels16_sse2(dst+16, src[0]+16, stride, h);
1666     }
1667 }
1668 void ff_avg_dirac_pixels32_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h)
1669 {
1670     if (h&3) {
1671         ff_avg_dirac_pixels32_c(dst, src, stride, h);
1672     } else {
1673     ff_avg_pixels16_sse2(dst   , src[0]   , stride, h);
1674     ff_avg_pixels16_sse2(dst+16, src[0]+16, stride, h);
1675     }
1676 }
1677 #endif
1678 #endif
1679
1680 /* XXX: Those functions should be suppressed ASAP when all IDCTs are
1681  * converted. */
1682 #if CONFIG_GPL
1683 static void ff_libmpeg2mmx_idct_put(uint8_t *dest, int line_size,
1684                                     int16_t *block)
1685 {
1686     ff_mmx_idct(block);
1687     ff_put_pixels_clamped_mmx(block, dest, line_size);
1688 }
1689
1690 static void ff_libmpeg2mmx_idct_add(uint8_t *dest, int line_size,
1691                                     int16_t *block)
1692 {
1693     ff_mmx_idct(block);
1694     ff_add_pixels_clamped_mmx(block, dest, line_size);
1695 }
1696
1697 static void ff_libmpeg2mmx2_idct_put(uint8_t *dest, int line_size,
1698                                      int16_t *block)
1699 {
1700     ff_mmxext_idct(block);
1701     ff_put_pixels_clamped_mmx(block, dest, line_size);
1702 }
1703
1704 static void ff_libmpeg2mmx2_idct_add(uint8_t *dest, int line_size,
1705                                      int16_t *block)
1706 {
1707     ff_mmxext_idct(block);
1708     ff_add_pixels_clamped_mmx(block, dest, line_size);
1709 }
1710 #endif
1711
1712 #if HAVE_INLINE_ASM
1713 static void vector_clipf_sse(float *dst, const float *src,
1714                              float min, float max, int len)
1715 {
1716     x86_reg i = (len - 16) * 4;
1717     __asm__ volatile (
1718         "movss          %3, %%xmm4      \n\t"
1719         "movss          %4, %%xmm5      \n\t"
1720         "shufps $0, %%xmm4, %%xmm4      \n\t"
1721         "shufps $0, %%xmm5, %%xmm5      \n\t"
1722         "1:                             \n\t"
1723         "movaps   (%2, %0), %%xmm0      \n\t" // 3/1 on intel
1724         "movaps 16(%2, %0), %%xmm1      \n\t"
1725         "movaps 32(%2, %0), %%xmm2      \n\t"
1726         "movaps 48(%2, %0), %%xmm3      \n\t"
1727         "maxps      %%xmm4, %%xmm0      \n\t"
1728         "maxps      %%xmm4, %%xmm1      \n\t"
1729         "maxps      %%xmm4, %%xmm2      \n\t"
1730         "maxps      %%xmm4, %%xmm3      \n\t"
1731         "minps      %%xmm5, %%xmm0      \n\t"
1732         "minps      %%xmm5, %%xmm1      \n\t"
1733         "minps      %%xmm5, %%xmm2      \n\t"
1734         "minps      %%xmm5, %%xmm3      \n\t"
1735         "movaps     %%xmm0,   (%1, %0)  \n\t"
1736         "movaps     %%xmm1, 16(%1, %0)  \n\t"
1737         "movaps     %%xmm2, 32(%1, %0)  \n\t"
1738         "movaps     %%xmm3, 48(%1, %0)  \n\t"
1739         "sub           $64, %0          \n\t"
1740         "jge            1b              \n\t"
1741         : "+&r"(i)
1742         : "r"(dst), "r"(src), "m"(min), "m"(max)
1743         : "memory"
1744     );
1745 }
1746
1747 #endif /* HAVE_INLINE_ASM */
1748
1749 int32_t ff_scalarproduct_int16_mmxext(const int16_t *v1, const int16_t *v2,
1750                                       int order);
1751 int32_t ff_scalarproduct_int16_sse2(const int16_t *v1, const int16_t *v2,
1752                                     int order);
1753 int32_t ff_scalarproduct_and_madd_int16_mmxext(int16_t *v1, const int16_t *v2,
1754                                                const int16_t *v3,
1755                                                int order, int mul);
1756 int32_t ff_scalarproduct_and_madd_int16_sse2(int16_t *v1, const int16_t *v2,
1757                                              const int16_t *v3,
1758                                              int order, int mul);
1759 int32_t ff_scalarproduct_and_madd_int16_ssse3(int16_t *v1, const int16_t *v2,
1760                                               const int16_t *v3,
1761                                               int order, int mul);
1762
1763 void ff_apply_window_int16_round_mmxext(int16_t *output, const int16_t *input,
1764                                         const int16_t *window, unsigned int len);
1765 void ff_apply_window_int16_round_sse2(int16_t *output, const int16_t *input,
1766                                       const int16_t *window, unsigned int len);
1767 void ff_apply_window_int16_mmxext(int16_t *output, const int16_t *input,
1768                                   const int16_t *window, unsigned int len);
1769 void ff_apply_window_int16_sse2(int16_t *output, const int16_t *input,
1770                                 const int16_t *window, unsigned int len);
1771 void ff_apply_window_int16_ssse3(int16_t *output, const int16_t *input,
1772                                  const int16_t *window, unsigned int len);
1773 void ff_apply_window_int16_ssse3_atom(int16_t *output, const int16_t *input,
1774                                       const int16_t *window, unsigned int len);
1775
1776 void ff_bswap32_buf_ssse3(uint32_t *dst, const uint32_t *src, int w);
1777 void ff_bswap32_buf_sse2(uint32_t *dst, const uint32_t *src, int w);
1778
1779 void ff_add_hfyu_median_prediction_mmxext(uint8_t *dst, const uint8_t *top,
1780                                           const uint8_t *diff, int w,
1781                                           int *left, int *left_top);
1782 int  ff_add_hfyu_left_prediction_ssse3(uint8_t *dst, const uint8_t *src,
1783                                        int w, int left);
1784 int  ff_add_hfyu_left_prediction_sse4(uint8_t *dst, const uint8_t *src,
1785                                       int w, int left);
1786
1787 void ff_vector_clip_int32_mmx     (int32_t *dst, const int32_t *src,
1788                                    int32_t min, int32_t max, unsigned int len);
1789 void ff_vector_clip_int32_sse2    (int32_t *dst, const int32_t *src,
1790                                    int32_t min, int32_t max, unsigned int len);
1791 void ff_vector_clip_int32_int_sse2(int32_t *dst, const int32_t *src,
1792                                    int32_t min, int32_t max, unsigned int len);
1793 void ff_vector_clip_int32_sse4    (int32_t *dst, const int32_t *src,
1794                                    int32_t min, int32_t max, unsigned int len);
1795
1796 #define SET_QPEL_FUNCS(PFX, IDX, SIZE, CPU, PREFIX)                          \
1797     do {                                                                     \
1798     c->PFX ## _pixels_tab[IDX][ 0] = PREFIX ## PFX ## SIZE ## _mc00_ ## CPU; \
1799     c->PFX ## _pixels_tab[IDX][ 1] = PREFIX ## PFX ## SIZE ## _mc10_ ## CPU; \
1800     c->PFX ## _pixels_tab[IDX][ 2] = PREFIX ## PFX ## SIZE ## _mc20_ ## CPU; \
1801     c->PFX ## _pixels_tab[IDX][ 3] = PREFIX ## PFX ## SIZE ## _mc30_ ## CPU; \
1802     c->PFX ## _pixels_tab[IDX][ 4] = PREFIX ## PFX ## SIZE ## _mc01_ ## CPU; \
1803     c->PFX ## _pixels_tab[IDX][ 5] = PREFIX ## PFX ## SIZE ## _mc11_ ## CPU; \
1804     c->PFX ## _pixels_tab[IDX][ 6] = PREFIX ## PFX ## SIZE ## _mc21_ ## CPU; \
1805     c->PFX ## _pixels_tab[IDX][ 7] = PREFIX ## PFX ## SIZE ## _mc31_ ## CPU; \
1806     c->PFX ## _pixels_tab[IDX][ 8] = PREFIX ## PFX ## SIZE ## _mc02_ ## CPU; \
1807     c->PFX ## _pixels_tab[IDX][ 9] = PREFIX ## PFX ## SIZE ## _mc12_ ## CPU; \
1808     c->PFX ## _pixels_tab[IDX][10] = PREFIX ## PFX ## SIZE ## _mc22_ ## CPU; \
1809     c->PFX ## _pixels_tab[IDX][11] = PREFIX ## PFX ## SIZE ## _mc32_ ## CPU; \
1810     c->PFX ## _pixels_tab[IDX][12] = PREFIX ## PFX ## SIZE ## _mc03_ ## CPU; \
1811     c->PFX ## _pixels_tab[IDX][13] = PREFIX ## PFX ## SIZE ## _mc13_ ## CPU; \
1812     c->PFX ## _pixels_tab[IDX][14] = PREFIX ## PFX ## SIZE ## _mc23_ ## CPU; \
1813     c->PFX ## _pixels_tab[IDX][15] = PREFIX ## PFX ## SIZE ## _mc33_ ## CPU; \
1814     } while (0)
1815
1816 #define SET_HPEL_FUNCS(PFX, IDX, SIZE, CPU)                                     \
1817     do {                                                                        \
1818         c->PFX ## _pixels_tab IDX [0] = PFX ## _pixels ## SIZE ## _     ## CPU; \
1819         c->PFX ## _pixels_tab IDX [1] = PFX ## _pixels ## SIZE ## _x2_  ## CPU; \
1820         c->PFX ## _pixels_tab IDX [2] = PFX ## _pixels ## SIZE ## _y2_  ## CPU; \
1821         c->PFX ## _pixels_tab IDX [3] = PFX ## _pixels ## SIZE ## _xy2_ ## CPU; \
1822     } while (0)
1823
1824 static void dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx, int mm_flags)
1825 {
1826     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
1827
1828 #if HAVE_INLINE_ASM
1829     c->put_pixels_clamped        = ff_put_pixels_clamped_mmx;
1830     c->put_signed_pixels_clamped = ff_put_signed_pixels_clamped_mmx;
1831     c->add_pixels_clamped        = ff_add_pixels_clamped_mmx;
1832
1833     if (!high_bit_depth) {
1834         c->clear_block  = clear_block_mmx;
1835         c->clear_blocks = clear_blocks_mmx;
1836         c->draw_edges   = draw_edges_mmx;
1837
1838         SET_HPEL_FUNCS(put,        [0], 16, mmx);
1839         SET_HPEL_FUNCS(put_no_rnd, [0], 16, mmx);
1840         SET_HPEL_FUNCS(avg,        [0], 16, mmx);
1841         SET_HPEL_FUNCS(avg_no_rnd,    , 16, mmx);
1842         SET_HPEL_FUNCS(put,        [1],  8, mmx);
1843         SET_HPEL_FUNCS(put_no_rnd, [1],  8, mmx);
1844         SET_HPEL_FUNCS(avg,        [1],  8, mmx);
1845     }
1846
1847 #if CONFIG_VIDEODSP && (ARCH_X86_32 || !HAVE_YASM)
1848     c->gmc = gmc_mmx;
1849 #endif
1850
1851     c->add_bytes = add_bytes_mmx;
1852
1853     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
1854         c->h263_v_loop_filter = h263_v_loop_filter_mmx;
1855         c->h263_h_loop_filter = h263_h_loop_filter_mmx;
1856     }
1857 #endif /* HAVE_INLINE_ASM */
1858
1859 #if HAVE_YASM
1860     if (!high_bit_depth && CONFIG_H264CHROMA) {
1861         c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_rnd_mmx;
1862         c->put_h264_chroma_pixels_tab[1] = ff_put_h264_chroma_mc4_mmx;
1863     }
1864
1865     c->vector_clip_int32 = ff_vector_clip_int32_mmx;
1866 #endif
1867
1868 }
1869
1870 static void dsputil_init_mmxext(DSPContext *c, AVCodecContext *avctx,
1871                                 int mm_flags)
1872 {
1873     const int bit_depth      = avctx->bits_per_raw_sample;
1874     const int high_bit_depth = bit_depth > 8;
1875
1876 #if HAVE_YASM
1877     SET_QPEL_FUNCS(avg_qpel,        0, 16, mmxext, );
1878     SET_QPEL_FUNCS(avg_qpel,        1,  8, mmxext, );
1879
1880     SET_QPEL_FUNCS(put_qpel,        0, 16, mmxext, );
1881     SET_QPEL_FUNCS(put_qpel,        1,  8, mmxext, );
1882     SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, mmxext, );
1883     SET_QPEL_FUNCS(put_no_rnd_qpel, 1,  8, mmxext, );
1884
1885     if (!high_bit_depth) {
1886         c->put_pixels_tab[0][1] = ff_put_pixels16_x2_mmxext;
1887         c->put_pixels_tab[0][2] = ff_put_pixels16_y2_mmxext;
1888
1889         c->avg_pixels_tab[0][0] = ff_avg_pixels16_mmxext;
1890         c->avg_pixels_tab[0][1] = ff_avg_pixels16_x2_mmxext;
1891         c->avg_pixels_tab[0][2] = ff_avg_pixels16_y2_mmxext;
1892
1893         c->put_pixels_tab[1][1] = ff_put_pixels8_x2_mmxext;
1894         c->put_pixels_tab[1][2] = ff_put_pixels8_y2_mmxext;
1895
1896         c->avg_pixels_tab[1][0] = ff_avg_pixels8_mmxext;
1897         c->avg_pixels_tab[1][1] = ff_avg_pixels8_x2_mmxext;
1898         c->avg_pixels_tab[1][2] = ff_avg_pixels8_y2_mmxext;
1899     }
1900
1901     if (!(avctx->flags & CODEC_FLAG_BITEXACT)) {
1902         if (!high_bit_depth) {
1903             c->put_no_rnd_pixels_tab[0][1] = ff_put_no_rnd_pixels16_x2_mmxext;
1904             c->put_no_rnd_pixels_tab[0][2] = ff_put_no_rnd_pixels16_y2_mmxext;
1905             c->put_no_rnd_pixels_tab[1][1] = ff_put_no_rnd_pixels8_x2_mmxext;
1906             c->put_no_rnd_pixels_tab[1][2] = ff_put_no_rnd_pixels8_y2_mmxext;
1907
1908             c->avg_pixels_tab[0][3] = ff_avg_pixels16_xy2_mmxext;
1909             c->avg_pixels_tab[1][3] = ff_avg_pixels8_xy2_mmxext;
1910         }
1911     }
1912 #endif /* HAVE_YASM */
1913
1914 #if HAVE_MMXEXT_EXTERNAL
1915     if (CONFIG_VP3_DECODER && (avctx->codec_id == AV_CODEC_ID_VP3 ||
1916                                avctx->codec_id == AV_CODEC_ID_THEORA)) {
1917         c->put_no_rnd_pixels_tab[1][1] = ff_put_no_rnd_pixels8_x2_exact_mmxext;
1918         c->put_no_rnd_pixels_tab[1][2] = ff_put_no_rnd_pixels8_y2_exact_mmxext;
1919     }
1920
1921     if (!high_bit_depth && CONFIG_H264CHROMA) {
1922         c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_rnd_mmxext;
1923         c->avg_h264_chroma_pixels_tab[1] = ff_avg_h264_chroma_mc4_mmxext;
1924         c->avg_h264_chroma_pixels_tab[2] = ff_avg_h264_chroma_mc2_mmxext;
1925         c->put_h264_chroma_pixels_tab[2] = ff_put_h264_chroma_mc2_mmxext;
1926     }
1927     if (bit_depth == 10 && CONFIG_H264CHROMA) {
1928         c->put_h264_chroma_pixels_tab[2] = ff_put_h264_chroma_mc2_10_mmxext;
1929         c->avg_h264_chroma_pixels_tab[2] = ff_avg_h264_chroma_mc2_10_mmxext;
1930         c->put_h264_chroma_pixels_tab[1] = ff_put_h264_chroma_mc4_10_mmxext;
1931         c->avg_h264_chroma_pixels_tab[1] = ff_avg_h264_chroma_mc4_10_mmxext;
1932     }
1933
1934     /* slower than cmov version on AMD */
1935     if (!(mm_flags & AV_CPU_FLAG_3DNOW))
1936         c->add_hfyu_median_prediction = ff_add_hfyu_median_prediction_mmxext;
1937
1938     c->scalarproduct_int16          = ff_scalarproduct_int16_mmxext;
1939     c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_mmxext;
1940
1941     if (avctx->flags & CODEC_FLAG_BITEXACT) {
1942         c->apply_window_int16 = ff_apply_window_int16_mmxext;
1943     } else {
1944         c->apply_window_int16 = ff_apply_window_int16_round_mmxext;
1945     }
1946 #endif /* HAVE_MMXEXT_EXTERNAL */
1947 }
1948
1949 static void dsputil_init_3dnow(DSPContext *c, AVCodecContext *avctx,
1950                                int mm_flags)
1951 {
1952     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
1953
1954 #if HAVE_YASM
1955     if (!high_bit_depth) {
1956         c->put_pixels_tab[0][1] = ff_put_pixels16_x2_3dnow;
1957         c->put_pixels_tab[0][2] = ff_put_pixels16_y2_3dnow;
1958
1959         c->avg_pixels_tab[0][0] = ff_avg_pixels16_3dnow;
1960         c->avg_pixels_tab[0][1] = ff_avg_pixels16_x2_3dnow;
1961         c->avg_pixels_tab[0][2] = ff_avg_pixels16_y2_3dnow;
1962
1963         c->put_pixels_tab[1][1] = ff_put_pixels8_x2_3dnow;
1964         c->put_pixels_tab[1][2] = ff_put_pixels8_y2_3dnow;
1965
1966         c->avg_pixels_tab[1][0] = ff_avg_pixels8_3dnow;
1967         c->avg_pixels_tab[1][1] = ff_avg_pixels8_x2_3dnow;
1968         c->avg_pixels_tab[1][2] = ff_avg_pixels8_y2_3dnow;
1969
1970         if (!(avctx->flags & CODEC_FLAG_BITEXACT)){
1971             c->put_no_rnd_pixels_tab[0][1] = ff_put_no_rnd_pixels16_x2_3dnow;
1972             c->put_no_rnd_pixels_tab[0][2] = ff_put_no_rnd_pixels16_y2_3dnow;
1973             c->put_no_rnd_pixels_tab[1][1] = ff_put_no_rnd_pixels8_x2_3dnow;
1974             c->put_no_rnd_pixels_tab[1][2] = ff_put_no_rnd_pixels8_y2_3dnow;
1975
1976             c->avg_pixels_tab[0][3] = ff_avg_pixels16_xy2_3dnow;
1977             c->avg_pixels_tab[1][3] = ff_avg_pixels8_xy2_3dnow;
1978         }
1979     }
1980
1981     if (CONFIG_VP3_DECODER && (avctx->codec_id == AV_CODEC_ID_VP3 ||
1982                                avctx->codec_id == AV_CODEC_ID_THEORA)) {
1983         c->put_no_rnd_pixels_tab[1][1] = ff_put_no_rnd_pixels8_x2_exact_3dnow;
1984         c->put_no_rnd_pixels_tab[1][2] = ff_put_no_rnd_pixels8_y2_exact_3dnow;
1985     }
1986
1987     if (!high_bit_depth && CONFIG_H264CHROMA) {
1988         c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_rnd_3dnow;
1989         c->avg_h264_chroma_pixels_tab[1] = ff_avg_h264_chroma_mc4_3dnow;
1990     }
1991 #endif /* HAVE_YASM */
1992 }
1993
1994 static void dsputil_init_sse(DSPContext *c, AVCodecContext *avctx, int mm_flags)
1995 {
1996     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
1997
1998 #if HAVE_INLINE_ASM
1999     if (!high_bit_depth) {
2000         if (!(CONFIG_MPEG_XVMC_DECODER && avctx->xvmc_acceleration > 1)) {
2001             /* XvMCCreateBlocks() may not allocate 16-byte aligned blocks */
2002             c->clear_block  = clear_block_sse;
2003             c->clear_blocks = clear_blocks_sse;
2004         }
2005     }
2006
2007     c->vector_clipf = vector_clipf_sse;
2008 #endif /* HAVE_INLINE_ASM */
2009
2010 #if HAVE_YASM
2011 #if HAVE_INLINE_ASM && CONFIG_VIDEODSP
2012     c->gmc = gmc_sse;
2013 #endif
2014 #endif /* HAVE_YASM */
2015 }
2016
2017 static void dsputil_init_sse2(DSPContext *c, AVCodecContext *avctx,
2018                               int mm_flags)
2019 {
2020     const int bit_depth      = avctx->bits_per_raw_sample;
2021     const int high_bit_depth = bit_depth > 8;
2022
2023 #if HAVE_SSE2_INLINE
2024     if (!high_bit_depth && avctx->idct_algo == FF_IDCT_XVIDMMX) {
2025         c->idct_put              = ff_idct_xvid_sse2_put;
2026         c->idct_add              = ff_idct_xvid_sse2_add;
2027         c->idct                  = ff_idct_xvid_sse2;
2028         c->idct_permutation_type = FF_SSE2_IDCT_PERM;
2029     }
2030 #endif /* HAVE_SSE2_INLINE */
2031
2032 #if HAVE_SSE2_EXTERNAL
2033     if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {
2034         // these functions are slower than mmx on AMD, but faster on Intel
2035         if (!high_bit_depth) {
2036             c->put_pixels_tab[0][0]        = ff_put_pixels16_sse2;
2037             c->put_no_rnd_pixels_tab[0][0] = ff_put_pixels16_sse2;
2038             c->avg_pixels_tab[0][0]        = ff_avg_pixels16_sse2;
2039         }
2040     }
2041
2042     if (bit_depth == 10) {
2043         if (CONFIG_H264CHROMA) {
2044             c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_10_sse2;
2045             c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_10_sse2;
2046         }
2047     }
2048
2049     c->scalarproduct_int16          = ff_scalarproduct_int16_sse2;
2050     c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_sse2;
2051     if (mm_flags & AV_CPU_FLAG_ATOM) {
2052         c->vector_clip_int32 = ff_vector_clip_int32_int_sse2;
2053     } else {
2054         c->vector_clip_int32 = ff_vector_clip_int32_sse2;
2055     }
2056     if (avctx->flags & CODEC_FLAG_BITEXACT) {
2057         c->apply_window_int16 = ff_apply_window_int16_sse2;
2058     } else if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {
2059         c->apply_window_int16 = ff_apply_window_int16_round_sse2;
2060     }
2061     c->bswap_buf = ff_bswap32_buf_sse2;
2062 #endif /* HAVE_SSE2_EXTERNAL */
2063 }
2064
2065 static void dsputil_init_ssse3(DSPContext *c, AVCodecContext *avctx,
2066                                int mm_flags)
2067 {
2068 #if HAVE_SSSE3_EXTERNAL
2069     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
2070
2071     if (!high_bit_depth && CONFIG_H264CHROMA) {
2072         c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_rnd_ssse3;
2073         c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_rnd_ssse3;
2074         c->put_h264_chroma_pixels_tab[1] = ff_put_h264_chroma_mc4_ssse3;
2075         c->avg_h264_chroma_pixels_tab[1] = ff_avg_h264_chroma_mc4_ssse3;
2076     }
2077     c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_ssse3;
2078     if (mm_flags & AV_CPU_FLAG_SSE4) // not really sse4, just slow on Conroe
2079         c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_sse4;
2080
2081     if (mm_flags & AV_CPU_FLAG_ATOM)
2082         c->apply_window_int16 = ff_apply_window_int16_ssse3_atom;
2083     else
2084         c->apply_window_int16 = ff_apply_window_int16_ssse3;
2085     if (!(mm_flags & (AV_CPU_FLAG_SSE42|AV_CPU_FLAG_3DNOW))) // cachesplit
2086         c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_ssse3;
2087     c->bswap_buf = ff_bswap32_buf_ssse3;
2088 #endif /* HAVE_SSSE3_EXTERNAL */
2089 }
2090
2091 static void dsputil_init_sse4(DSPContext *c, AVCodecContext *avctx,
2092                               int mm_flags)
2093 {
2094 #if HAVE_SSE4_EXTERNAL
2095     c->vector_clip_int32 = ff_vector_clip_int32_sse4;
2096 #endif /* HAVE_SSE4_EXTERNAL */
2097 }
2098
2099 static void dsputil_init_avx(DSPContext *c, AVCodecContext *avctx, int mm_flags)
2100 {
2101 #if HAVE_AVX_EXTERNAL
2102     const int bit_depth = avctx->bits_per_raw_sample;
2103
2104     if (bit_depth == 10) {
2105         if (CONFIG_H264CHROMA) {
2106             c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_10_avx;
2107             c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_10_avx;
2108         }
2109     }
2110 #endif /* HAVE_AVX_EXTERNAL */
2111 }
2112
2113 void ff_dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx)
2114 {
2115     int mm_flags = av_get_cpu_flags();
2116
2117 #if HAVE_7REGS && HAVE_INLINE_ASM
2118     if (mm_flags & AV_CPU_FLAG_CMOV)
2119         c->add_hfyu_median_prediction = add_hfyu_median_prediction_cmov;
2120 #endif
2121
2122     if (mm_flags & AV_CPU_FLAG_MMX) {
2123 #if HAVE_INLINE_ASM
2124         const int idct_algo = avctx->idct_algo;
2125
2126         if (avctx->lowres == 0 && avctx->bits_per_raw_sample <= 8) {
2127             if (idct_algo == FF_IDCT_AUTO || idct_algo == FF_IDCT_SIMPLEMMX) {
2128                 c->idct_put              = ff_simple_idct_put_mmx;
2129                 c->idct_add              = ff_simple_idct_add_mmx;
2130                 c->idct                  = ff_simple_idct_mmx;
2131                 c->idct_permutation_type = FF_SIMPLE_IDCT_PERM;
2132 #if CONFIG_GPL
2133             } else if (idct_algo == FF_IDCT_LIBMPEG2MMX) {
2134                 if (mm_flags & AV_CPU_FLAG_MMX2) {
2135                     c->idct_put = ff_libmpeg2mmx2_idct_put;
2136                     c->idct_add = ff_libmpeg2mmx2_idct_add;
2137                     c->idct     = ff_mmxext_idct;
2138                 } else {
2139                     c->idct_put = ff_libmpeg2mmx_idct_put;
2140                     c->idct_add = ff_libmpeg2mmx_idct_add;
2141                     c->idct     = ff_mmx_idct;
2142                 }
2143                 c->idct_permutation_type = FF_LIBMPEG2_IDCT_PERM;
2144 #endif
2145             } else if (idct_algo == FF_IDCT_XVIDMMX) {
2146                 if (mm_flags & AV_CPU_FLAG_SSE2) {
2147                     c->idct_put              = ff_idct_xvid_sse2_put;
2148                     c->idct_add              = ff_idct_xvid_sse2_add;
2149                     c->idct                  = ff_idct_xvid_sse2;
2150                     c->idct_permutation_type = FF_SSE2_IDCT_PERM;
2151                 } else if (mm_flags & AV_CPU_FLAG_MMXEXT) {
2152                     c->idct_put              = ff_idct_xvid_mmxext_put;
2153                     c->idct_add              = ff_idct_xvid_mmxext_add;
2154                     c->idct                  = ff_idct_xvid_mmxext;
2155                 } else {
2156                     c->idct_put              = ff_idct_xvid_mmx_put;
2157                     c->idct_add              = ff_idct_xvid_mmx_add;
2158                     c->idct                  = ff_idct_xvid_mmx;
2159                 }
2160             }
2161         }
2162 #endif /* HAVE_INLINE_ASM */
2163
2164         dsputil_init_mmx(c, avctx, mm_flags);
2165     }
2166
2167     if (mm_flags & AV_CPU_FLAG_MMXEXT)
2168         dsputil_init_mmxext(c, avctx, mm_flags);
2169
2170     if (mm_flags & AV_CPU_FLAG_3DNOW)
2171         dsputil_init_3dnow(c, avctx, mm_flags);
2172
2173     if (mm_flags & AV_CPU_FLAG_SSE)
2174         dsputil_init_sse(c, avctx, mm_flags);
2175
2176     if (mm_flags & AV_CPU_FLAG_SSE2)
2177         dsputil_init_sse2(c, avctx, mm_flags);
2178
2179     if (mm_flags & AV_CPU_FLAG_SSSE3)
2180         dsputil_init_ssse3(c, avctx, mm_flags);
2181
2182     if (mm_flags & AV_CPU_FLAG_SSE4)
2183         dsputil_init_sse4(c, avctx, mm_flags);
2184
2185     if (mm_flags & AV_CPU_FLAG_AVX)
2186         dsputil_init_avx(c, avctx, mm_flags);
2187
2188     if (CONFIG_ENCODERS)
2189         ff_dsputilenc_init_mmx(c, avctx);
2190 }