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/attributes.h"
  26 #include "libavutil/cpu.h"
  27 #include "libavutil/x86/asm.h"
  28 #include "libavcodec/dsputil.h"
  29 #include "libavcodec/h264dsp.h"
  30 #include "libavcodec/mpegvideo.h"
  31 #include "libavcodec/simple_idct.h"
  32 #include "libavcodec/videodsp.h"
  33 #include "dsputil_mmx.h"
  34 #include "idct_xvid.h"
  35 #include "diracdsp_mmx.h"
  36
  37 //#undef NDEBUG
  38 //#include <assert.h>
  39
  40 /* pixel operations */
  41 DECLARE_ALIGNED(8,  const uint64_t, ff_bone) = 0x0101010101010101ULL;
  42 DECLARE_ALIGNED(8,  const uint64_t, ff_wtwo) = 0x0002000200020002ULL;
  43
  44 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_1)    = { 0x0001000100010001ULL, 0x0001000100010001ULL };
  45 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_2)    = { 0x0002000200020002ULL, 0x0002000200020002ULL };
  46 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_3)    = { 0x0003000300030003ULL, 0x0003000300030003ULL };
  47 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_4)    = { 0x0004000400040004ULL, 0x0004000400040004ULL };
  48 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_5)    = { 0x0005000500050005ULL, 0x0005000500050005ULL };
  49 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_8)    = { 0x0008000800080008ULL, 0x0008000800080008ULL };
  50 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_9)    = { 0x0009000900090009ULL, 0x0009000900090009ULL };
  51 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_15)   =   0x000F000F000F000FULL;
  52 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_16)   = { 0x0010001000100010ULL, 0x0010001000100010ULL };
  53 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_17)   = { 0x0011001100110011ULL, 0x0011001100110011ULL };
  54 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_18)   = { 0x0012001200120012ULL, 0x0012001200120012ULL };
  55 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_20)   =   0x0014001400140014ULL;
  56 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_27)   = { 0x001B001B001B001BULL, 0x001B001B001B001BULL };
  57 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_28)   = { 0x001C001C001C001CULL, 0x001C001C001C001CULL };
  58 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_32)   = { 0x0020002000200020ULL, 0x0020002000200020ULL };
  59 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_42)   =   0x002A002A002A002AULL;
  60 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_53)   =   0x0035003500350035ULL;
  61 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_63)   = { 0x003F003F003F003FULL, 0x003F003F003F003FULL };
  62 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_64)   = { 0x0040004000400040ULL, 0x0040004000400040ULL };
  63 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_96)   =   0x0060006000600060ULL;
  64 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_128)  =   0x0080008000800080ULL;
  65 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_255)  =   0x00ff00ff00ff00ffULL;
  66 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_512)  = { 0x0200020002000200ULL, 0x0200020002000200ULL };
  67 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_1019) = { 0x03FB03FB03FB03FBULL, 0x03FB03FB03FB03FBULL };
  68
  69 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_0)    = { 0x0000000000000000ULL, 0x0000000000000000ULL };
  70 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_1)    = { 0x0101010101010101ULL, 0x0101010101010101ULL };
  71 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_3)    = { 0x0303030303030303ULL, 0x0303030303030303ULL };
  72 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_4)    = { 0x0404040404040404ULL, 0x0404040404040404ULL };
  73 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_7)    =   0x0707070707070707ULL;
  74 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_1F)   =   0x1F1F1F1F1F1F1F1FULL;
  75 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_3F)   =   0x3F3F3F3F3F3F3F3FULL;
  76 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_80)   = { 0x8080808080808080ULL, 0x8080808080808080ULL };
  77 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_81)   =   0x8181818181818181ULL;
  78 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_A1)   = { 0xA1A1A1A1A1A1A1A1ULL, 0xA1A1A1A1A1A1A1A1ULL };
  79 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_F8)   = { 0xF8F8F8F8F8F8F8F8ULL, 0xF8F8F8F8F8F8F8F8ULL };
  80 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_FC)   =   0xFCFCFCFCFCFCFCFCULL;
  81 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_FE)   = { 0xFEFEFEFEFEFEFEFEULL, 0xFEFEFEFEFEFEFEFEULL };
  82
  83 DECLARE_ALIGNED(16, const double, ff_pd_1)[2] = { 1.0, 1.0 };
  84 DECLARE_ALIGNED(16, const double, ff_pd_2)[2] = { 2.0, 2.0 };
  85
  86
  87 #if HAVE_YASM
  88 void ff_put_pixels8_x2_mmxext(uint8_t *block, const uint8_t *pixels,
  89                               ptrdiff_t line_size, int h);
  90 void ff_put_pixels8_x2_3dnow(uint8_t *block, const uint8_t *pixels,
  91                              ptrdiff_t line_size, int h);
  92 void ff_put_pixels8_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
  93                               int dstStride, int src1Stride, int h);
  94 void ff_put_no_rnd_pixels8_l2_mmxext(uint8_t *dst, uint8_t *src1,
  95                                      uint8_t *src2, int dstStride,
  96                                      int src1Stride, int h);
  97 void ff_avg_pixels8_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
  98                               int dstStride, int src1Stride, int h);
  99 void ff_put_pixels16_x2_mmxext(uint8_t *block, const uint8_t *pixels,
 100                                ptrdiff_t line_size, int h);
 101 void ff_put_pixels16_x2_3dnow(uint8_t *block, const uint8_t *pixels,
 102                               ptrdiff_t line_size, int h);
 103 void ff_put_pixels16_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
 104                                int dstStride, int src1Stride, int h);
 105 void ff_avg_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_pixels16_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
 108                                       int dstStride, int src1Stride, int h);
 109 void ff_put_no_rnd_pixels8_x2_mmxext(uint8_t *block, const uint8_t *pixels,
 110                                      ptrdiff_t line_size, int h);
 111 void ff_put_no_rnd_pixels8_x2_3dnow(uint8_t *block, const uint8_t *pixels,
 112                                     ptrdiff_t line_size, int h);
 113 void ff_put_no_rnd_pixels8_x2_exact_mmxext(uint8_t *block,
 114                                            const uint8_t *pixels,
 115                                            ptrdiff_t line_size, int h);
 116 void ff_put_no_rnd_pixels8_x2_exact_3dnow(uint8_t *block,
 117                                           const uint8_t *pixels,
 118                                           ptrdiff_t line_size, int h);
 119 void ff_put_pixels8_y2_mmxext(uint8_t *block, const uint8_t *pixels,
 120                               ptrdiff_t line_size, int h);
 121 void ff_put_pixels8_y2_3dnow(uint8_t *block, const uint8_t *pixels,
 122                              ptrdiff_t line_size, int h);
 123 void ff_put_no_rnd_pixels8_y2_mmxext(uint8_t *block, const uint8_t *pixels,
 124                                      ptrdiff_t line_size, int h);
 125 void ff_put_no_rnd_pixels8_y2_3dnow(uint8_t *block, const uint8_t *pixels,
 126                                     ptrdiff_t line_size, int h);
 127 void ff_put_no_rnd_pixels8_y2_exact_mmxext(uint8_t *block,
 128                                            const uint8_t *pixels,
 129                                            ptrdiff_t line_size, int h);
 130 void ff_put_no_rnd_pixels8_y2_exact_3dnow(uint8_t *block,
 131                                           const uint8_t *pixels,
 132                                           ptrdiff_t line_size, int h);
 133 void ff_avg_pixels8_mmxext(uint8_t *block, const uint8_t *pixels,
 134                            ptrdiff_t line_size, int h);
 135 void ff_avg_pixels8_3dnow(uint8_t *block, const uint8_t *pixels,
 136                           ptrdiff_t line_size, int h);
 137 void ff_avg_pixels8_x2_mmxext(uint8_t *block, const uint8_t *pixels,
 138                               ptrdiff_t line_size, int h);
 139 void ff_avg_pixels8_x2_3dnow(uint8_t *block, const uint8_t *pixels,
 140                              ptrdiff_t line_size, int h);
 141 void ff_avg_pixels8_y2_mmxext(uint8_t *block, const uint8_t *pixels,
 142                               ptrdiff_t line_size, int h);
 143 void ff_avg_pixels8_y2_3dnow(uint8_t *block, const uint8_t *pixels,
 144                              ptrdiff_t line_size, int h);
 145 void ff_avg_pixels8_xy2_mmxext(uint8_t *block, const uint8_t *pixels,
 146                                ptrdiff_t line_size, int h);
 147 void ff_avg_pixels8_xy2_3dnow(uint8_t *block, const uint8_t *pixels,
 148                               ptrdiff_t line_size, int h);
 149
 150 void ff_put_pixels8_mmxext(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h);
 151 static void ff_put_pixels16_mmxext(uint8_t *block, const uint8_t *pixels,
 152                                    ptrdiff_t line_size, int h)
 153 {
 154     ff_put_pixels8_mmxext(block,     pixels,     line_size, h);
 155     ff_put_pixels8_mmxext(block + 8, pixels + 8, line_size, h);
 156 }
 157
 158 void ff_put_mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 159                                          int dstStride, int srcStride, int h);
 160 void ff_avg_mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 161                                          int dstStride, int srcStride, int h);
 162 void ff_put_no_rnd_mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 163                                                  int dstStride, int srcStride,
 164                                                  int h);
 165 void ff_put_mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 166                                         int dstStride, int srcStride, int h);
 167 void ff_avg_mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 168                                         int dstStride, int srcStride, int h);
 169 void ff_put_no_rnd_mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 170                                                 int dstStride, int srcStride,
 171                                                 int h);
 172 void ff_put_mpeg4_qpel16_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 173                                          int dstStride, int srcStride);
 174 void ff_avg_mpeg4_qpel16_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 175                                          int dstStride, int srcStride);
 176 void ff_put_no_rnd_mpeg4_qpel16_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 177                                                  int dstStride, int srcStride);
 178 void ff_put_mpeg4_qpel8_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 179                                         int dstStride, int srcStride);
 180 void ff_avg_mpeg4_qpel8_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 181                                         int dstStride, int srcStride);
 182 void ff_put_no_rnd_mpeg4_qpel8_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 183                                                 int dstStride, int srcStride);
 184 #define ff_put_no_rnd_pixels16_mmxext ff_put_pixels16_mmxext
 185 #define ff_put_no_rnd_pixels8_mmxext ff_put_pixels8_mmxext
 186 #endif /* HAVE_YASM */
 187
 188
 189 #if HAVE_INLINE_ASM
 190
 191 #define JUMPALIGN()     __asm__ volatile (".p2align 3"::)
 192 #define MOVQ_ZERO(regd) __asm__ volatile ("pxor %%"#regd", %%"#regd ::)
 193
 194 #define MOVQ_BFE(regd)                                  \
 195     __asm__ volatile (                                  \
 196         "pcmpeqd %%"#regd", %%"#regd"   \n\t"           \
 197         "paddb   %%"#regd", %%"#regd"   \n\t" ::)
 198
 199 #ifndef PIC
 200 #define MOVQ_BONE(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_bone))
 201 #define MOVQ_WTWO(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_wtwo))
 202 #else
 203 // for shared library it's better to use this way for accessing constants
 204 // pcmpeqd -> -1
 205 #define MOVQ_BONE(regd)                                 \
 206     __asm__ volatile (                                  \
 207         "pcmpeqd  %%"#regd", %%"#regd"  \n\t"           \
 208         "psrlw          $15, %%"#regd"  \n\t"           \
 209         "packuswb %%"#regd", %%"#regd"  \n\t" ::)
 210
 211 #define MOVQ_WTWO(regd)                                 \
 212     __asm__ volatile (                                  \
 213         "pcmpeqd %%"#regd", %%"#regd"   \n\t"           \
 214         "psrlw         $15, %%"#regd"   \n\t"           \
 215         "psllw          $1, %%"#regd"   \n\t"::)
 216
 217 #endif
 218
 219 // using regr as temporary and for the output result
 220 // first argument is unmodifed and second is trashed
 221 // regfe is supposed to contain 0xfefefefefefefefe
 222 #define PAVGB_MMX_NO_RND(rega, regb, regr, regfe)                \
 223     "movq   "#rega", "#regr"            \n\t"                    \
 224     "pand   "#regb", "#regr"            \n\t"                    \
 225     "pxor   "#rega", "#regb"            \n\t"                    \
 226     "pand  "#regfe", "#regb"            \n\t"                    \
 227     "psrlq       $1, "#regb"            \n\t"                    \
 228     "paddb  "#regb", "#regr"            \n\t"
 229
 230 #define PAVGB_MMX(rega, regb, regr, regfe)                       \
 231     "movq   "#rega", "#regr"            \n\t"                    \
 232     "por    "#regb", "#regr"            \n\t"                    \
 233     "pxor   "#rega", "#regb"            \n\t"                    \
 234     "pand  "#regfe", "#regb"            \n\t"                    \
 235     "psrlq       $1, "#regb"            \n\t"                    \
 236     "psubb  "#regb", "#regr"            \n\t"
 237
 238 // mm6 is supposed to contain 0xfefefefefefefefe
 239 #define PAVGBP_MMX_NO_RND(rega, regb, regr,  regc, regd, regp)   \
 240     "movq  "#rega", "#regr"             \n\t"                    \
 241     "movq  "#regc", "#regp"             \n\t"                    \
 242     "pand  "#regb", "#regr"             \n\t"                    \
 243     "pand  "#regd", "#regp"             \n\t"                    \
 244     "pxor  "#rega", "#regb"             \n\t"                    \
 245     "pxor  "#regc", "#regd"             \n\t"                    \
 246     "pand    %%mm6, "#regb"             \n\t"                    \
 247     "pand    %%mm6, "#regd"             \n\t"                    \
 248     "psrlq      $1, "#regb"             \n\t"                    \
 249     "psrlq      $1, "#regd"             \n\t"                    \
 250     "paddb "#regb", "#regr"             \n\t"                    \
 251     "paddb "#regd", "#regp"             \n\t"
 252
 253 #define PAVGBP_MMX(rega, regb, regr, regc, regd, regp)           \
 254     "movq  "#rega", "#regr"             \n\t"                    \
 255     "movq  "#regc", "#regp"             \n\t"                    \
 256     "por   "#regb", "#regr"             \n\t"                    \
 257     "por   "#regd", "#regp"             \n\t"                    \
 258     "pxor  "#rega", "#regb"             \n\t"                    \
 259     "pxor  "#regc", "#regd"             \n\t"                    \
 260     "pand    %%mm6, "#regb"             \n\t"                    \
 261     "pand    %%mm6, "#regd"             \n\t"                    \
 262     "psrlq      $1, "#regd"             \n\t"                    \
 263     "psrlq      $1, "#regb"             \n\t"                    \
 264     "psubb "#regb", "#regr"             \n\t"                    \
 265     "psubb "#regd", "#regp"             \n\t"
 266
 267 /***********************************/
 268 /* MMX no rounding */
 269 #define NO_RND 1
 270 #define DEF(x, y) x ## _no_rnd_ ## y ## _mmx
 271 #define SET_RND  MOVQ_WONE
 272 #define PAVGBP(a, b, c, d, e, f)        PAVGBP_MMX_NO_RND(a, b, c, d, e, f)
 273 #define PAVGB(a, b, c, e)               PAVGB_MMX_NO_RND(a, b, c, e)
 274 #define OP_AVG(a, b, c, e)              PAVGB_MMX(a, b, c, e)
 275
 276 #include "dsputil_rnd_template.c"
 277
 278 #undef DEF
 279 #undef SET_RND
 280 #undef PAVGBP
 281 #undef PAVGB
 282 #undef NO_RND
 283 /***********************************/
 284 /* MMX rounding */
 285
 286 #define DEF(x, y) x ## _ ## y ## _mmx
 287 #define SET_RND  MOVQ_WTWO
 288 #define PAVGBP(a, b, c, d, e, f)        PAVGBP_MMX(a, b, c, d, e, f)
 289 #define PAVGB(a, b, c, e)               PAVGB_MMX(a, b, c, e)
 290
 291 #include "dsputil_rnd_template.c"
 292
 293 #undef DEF
 294 #undef SET_RND
 295 #undef PAVGBP
 296 #undef PAVGB
 297 #undef OP_AVG
 298
 299 #endif /* HAVE_INLINE_ASM */
 300
 301
 302 #if HAVE_YASM
 303 #define ff_put_pixels8_mmx ff_put_pixels8_mmxext
 304
 305 /***********************************/
 306 /* 3Dnow specific */
 307
 308 #define DEF(x) x ## _3dnow
 309
 310 #include "dsputil_avg_template.c"
 311
 312 #undef DEF
 313
 314 /***********************************/
 315 /* MMXEXT specific */
 316
 317 #define DEF(x) x ## _mmxext
 318
 319 #include "dsputil_avg_template.c"
 320
 321 #undef DEF
 322
 323 #endif /* HAVE_YASM */
 324
 325
 326 #if HAVE_INLINE_ASM
 327 #define put_no_rnd_pixels16_mmx put_pixels16_mmx
 328 #define put_no_rnd_pixels8_mmx put_pixels8_mmx
 329 #define put_pixels16_mmxext put_pixels16_mmx
 330 #define put_pixels8_mmxext put_pixels8_mmx
 331 #define put_pixels4_mmxext put_pixels4_mmx
 332 #define put_no_rnd_pixels16_mmxext put_no_rnd_pixels16_mmx
 333 #define put_no_rnd_pixels8_mmxext put_no_rnd_pixels8_mmx
 334
 335 /***********************************/
 336 /* standard MMX */
 337
 338 void ff_put_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
 339                                int line_size)
 340 {
 341     const int16_t *p;
 342     uint8_t *pix;
 343
 344     /* read the pixels */
 345     p   = block;
 346     pix = pixels;
 347     /* unrolled loop */
 348     __asm__ volatile (
 349         "movq      (%3), %%mm0          \n\t"
 350         "movq     8(%3), %%mm1          \n\t"
 351         "movq    16(%3), %%mm2          \n\t"
 352         "movq    24(%3), %%mm3          \n\t"
 353         "movq    32(%3), %%mm4          \n\t"
 354         "movq    40(%3), %%mm5          \n\t"
 355         "movq    48(%3), %%mm6          \n\t"
 356         "movq    56(%3), %%mm7          \n\t"
 357         "packuswb %%mm1, %%mm0          \n\t"
 358         "packuswb %%mm3, %%mm2          \n\t"
 359         "packuswb %%mm5, %%mm4          \n\t"
 360         "packuswb %%mm7, %%mm6          \n\t"
 361         "movq     %%mm0, (%0)           \n\t"
 362         "movq     %%mm2, (%0, %1)       \n\t"
 363         "movq     %%mm4, (%0, %1, 2)    \n\t"
 364         "movq     %%mm6, (%0, %2)       \n\t"
 365         :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3),
 366            "r"(p)
 367         : "memory");
 368     pix += line_size * 4;
 369     p   += 32;
 370
 371     // if here would be an exact copy of the code above
 372     // compiler would generate some very strange code
 373     // thus using "r"
 374     __asm__ volatile (
 375         "movq       (%3), %%mm0         \n\t"
 376         "movq      8(%3), %%mm1         \n\t"
 377         "movq     16(%3), %%mm2         \n\t"
 378         "movq     24(%3), %%mm3         \n\t"
 379         "movq     32(%3), %%mm4         \n\t"
 380         "movq     40(%3), %%mm5         \n\t"
 381         "movq     48(%3), %%mm6         \n\t"
 382         "movq     56(%3), %%mm7         \n\t"
 383         "packuswb  %%mm1, %%mm0         \n\t"
 384         "packuswb  %%mm3, %%mm2         \n\t"
 385         "packuswb  %%mm5, %%mm4         \n\t"
 386         "packuswb  %%mm7, %%mm6         \n\t"
 387         "movq      %%mm0, (%0)          \n\t"
 388         "movq      %%mm2, (%0, %1)      \n\t"
 389         "movq      %%mm4, (%0, %1, 2)   \n\t"
 390         "movq      %%mm6, (%0, %2)      \n\t"
 391         :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3), "r"(p)
 392         : "memory");
 393 }
 394
 395 #define put_signed_pixels_clamped_mmx_half(off)             \
 396     "movq          "#off"(%2), %%mm1        \n\t"           \
 397     "movq     16 + "#off"(%2), %%mm2        \n\t"           \
 398     "movq     32 + "#off"(%2), %%mm3        \n\t"           \
 399     "movq     48 + "#off"(%2), %%mm4        \n\t"           \
 400     "packsswb  8 + "#off"(%2), %%mm1        \n\t"           \
 401     "packsswb 24 + "#off"(%2), %%mm2        \n\t"           \
 402     "packsswb 40 + "#off"(%2), %%mm3        \n\t"           \
 403     "packsswb 56 + "#off"(%2), %%mm4        \n\t"           \
 404     "paddb              %%mm0, %%mm1        \n\t"           \
 405     "paddb              %%mm0, %%mm2        \n\t"           \
 406     "paddb              %%mm0, %%mm3        \n\t"           \
 407     "paddb              %%mm0, %%mm4        \n\t"           \
 408     "movq               %%mm1, (%0)         \n\t"           \
 409     "movq               %%mm2, (%0, %3)     \n\t"           \
 410     "movq               %%mm3, (%0, %3, 2)  \n\t"           \
 411     "movq               %%mm4, (%0, %1)     \n\t"
 412
 413 void ff_put_signed_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
 414                                       int line_size)
 415 {
 416     x86_reg line_skip = line_size;
 417     x86_reg line_skip3;
 418
 419     __asm__ volatile (
 420         "movq "MANGLE(ff_pb_80)", %%mm0     \n\t"
 421         "lea         (%3, %3, 2), %1        \n\t"
 422         put_signed_pixels_clamped_mmx_half(0)
 423         "lea         (%0, %3, 4), %0        \n\t"
 424         put_signed_pixels_clamped_mmx_half(64)
 425         : "+&r"(pixels), "=&r"(line_skip3)
 426         : "r"(block), "r"(line_skip)
 427         : "memory");
 428 }
 429
 430 void ff_add_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
 431                                int line_size)
 432 {
 433     const int16_t *p;
 434     uint8_t *pix;
 435     int i;
 436
 437     /* read the pixels */
 438     p   = block;
 439     pix = pixels;
 440     MOVQ_ZERO(mm7);
 441     i = 4;
 442     do {
 443         __asm__ volatile (
 444             "movq        (%2), %%mm0    \n\t"
 445             "movq       8(%2), %%mm1    \n\t"
 446             "movq      16(%2), %%mm2    \n\t"
 447             "movq      24(%2), %%mm3    \n\t"
 448             "movq          %0, %%mm4    \n\t"
 449             "movq          %1, %%mm6    \n\t"
 450             "movq       %%mm4, %%mm5    \n\t"
 451             "punpcklbw  %%mm7, %%mm4    \n\t"
 452             "punpckhbw  %%mm7, %%mm5    \n\t"
 453             "paddsw     %%mm4, %%mm0    \n\t"
 454             "paddsw     %%mm5, %%mm1    \n\t"
 455             "movq       %%mm6, %%mm5    \n\t"
 456             "punpcklbw  %%mm7, %%mm6    \n\t"
 457             "punpckhbw  %%mm7, %%mm5    \n\t"
 458             "paddsw     %%mm6, %%mm2    \n\t"
 459             "paddsw     %%mm5, %%mm3    \n\t"
 460             "packuswb   %%mm1, %%mm0    \n\t"
 461             "packuswb   %%mm3, %%mm2    \n\t"
 462             "movq       %%mm0, %0       \n\t"
 463             "movq       %%mm2, %1       \n\t"
 464             : "+m"(*pix), "+m"(*(pix + line_size))
 465             : "r"(p)
 466             : "memory");
 467         pix += line_size * 2;
 468         p   += 16;
 469     } while (--i);
 470 }
 471
 472 static void put_pixels8_mmx(uint8_t *block, const uint8_t *pixels,
 473                             ptrdiff_t line_size, int h)
 474 {
 475     __asm__ volatile (
 476         "lea   (%3, %3), %%"REG_a"      \n\t"
 477         ".p2align     3                 \n\t"
 478         "1:                             \n\t"
 479         "movq  (%1    ), %%mm0          \n\t"
 480         "movq  (%1, %3), %%mm1          \n\t"
 481         "movq     %%mm0, (%2)           \n\t"
 482         "movq     %%mm1, (%2, %3)       \n\t"
 483         "add  %%"REG_a", %1             \n\t"
 484         "add  %%"REG_a", %2             \n\t"
 485         "movq  (%1    ), %%mm0          \n\t"
 486         "movq  (%1, %3), %%mm1          \n\t"
 487         "movq     %%mm0, (%2)           \n\t"
 488         "movq     %%mm1, (%2, %3)       \n\t"
 489         "add  %%"REG_a", %1             \n\t"
 490         "add  %%"REG_a", %2             \n\t"
 491         "subl        $4, %0             \n\t"
 492         "jnz         1b                 \n\t"
 493         : "+g"(h), "+r"(pixels),  "+r"(block)
 494         : "r"((x86_reg)line_size)
 495         : "%"REG_a, "memory"
 496         );
 497 }
 498
 499 static void put_pixels16_mmx(uint8_t *block, const uint8_t *pixels,
 500                              ptrdiff_t line_size, int h)
 501 {
 502     __asm__ volatile (
 503         "lea   (%3, %3), %%"REG_a"      \n\t"
 504         ".p2align     3                 \n\t"
 505         "1:                             \n\t"
 506         "movq  (%1    ), %%mm0          \n\t"
 507         "movq 8(%1    ), %%mm4          \n\t"
 508         "movq  (%1, %3), %%mm1          \n\t"
 509         "movq 8(%1, %3), %%mm5          \n\t"
 510         "movq     %%mm0,  (%2)          \n\t"
 511         "movq     %%mm4, 8(%2)          \n\t"
 512         "movq     %%mm1,  (%2, %3)      \n\t"
 513         "movq     %%mm5, 8(%2, %3)      \n\t"
 514         "add  %%"REG_a", %1             \n\t"
 515         "add  %%"REG_a", %2             \n\t"
 516         "movq  (%1    ), %%mm0          \n\t"
 517         "movq 8(%1    ), %%mm4          \n\t"
 518         "movq  (%1, %3), %%mm1          \n\t"
 519         "movq 8(%1, %3), %%mm5          \n\t"
 520         "movq     %%mm0,  (%2)          \n\t"
 521         "movq     %%mm4, 8(%2)          \n\t"
 522         "movq     %%mm1,  (%2, %3)      \n\t"
 523         "movq     %%mm5, 8(%2, %3)      \n\t"
 524         "add  %%"REG_a", %1             \n\t"
 525         "add  %%"REG_a", %2             \n\t"
 526         "subl        $4, %0             \n\t"
 527         "jnz         1b                 \n\t"
 528         : "+g"(h), "+r"(pixels),  "+r"(block)
 529         : "r"((x86_reg)line_size)
 530         : "%"REG_a, "memory"
 531         );
 532 }
 533
 534 #define CLEAR_BLOCKS(name, n)                           \
 535 static void name(int16_t *blocks)                       \
 536 {                                                       \
 537     __asm__ volatile (                                  \
 538         "pxor %%mm7, %%mm7              \n\t"           \
 539         "mov     %1,        %%"REG_a"   \n\t"           \
 540         "1:                             \n\t"           \
 541         "movq %%mm7,   (%0, %%"REG_a")  \n\t"           \
 542         "movq %%mm7,  8(%0, %%"REG_a")  \n\t"           \
 543         "movq %%mm7, 16(%0, %%"REG_a")  \n\t"           \
 544         "movq %%mm7, 24(%0, %%"REG_a")  \n\t"           \
 545         "add    $32, %%"REG_a"          \n\t"           \
 546         "js      1b                     \n\t"           \
 547         :: "r"(((uint8_t *)blocks) + 128 * n),          \
 548            "i"(-128 * n)                                \
 549         : "%"REG_a                                      \
 550         );                                              \
 551 }
 552 CLEAR_BLOCKS(clear_blocks_mmx, 6)
 553 CLEAR_BLOCKS(clear_block_mmx, 1)
 554
 555 static void clear_block_sse(int16_t *block)
 556 {
 557     __asm__ volatile (
 558         "xorps  %%xmm0, %%xmm0          \n"
 559         "movaps %%xmm0,    (%0)         \n"
 560         "movaps %%xmm0,  16(%0)         \n"
 561         "movaps %%xmm0,  32(%0)         \n"
 562         "movaps %%xmm0,  48(%0)         \n"
 563         "movaps %%xmm0,  64(%0)         \n"
 564         "movaps %%xmm0,  80(%0)         \n"
 565         "movaps %%xmm0,  96(%0)         \n"
 566         "movaps %%xmm0, 112(%0)         \n"
 567         :: "r"(block)
 568         : "memory"
 569     );
 570 }
 571
 572 static void clear_blocks_sse(int16_t *blocks)
 573 {
 574     __asm__ volatile (
 575         "xorps  %%xmm0, %%xmm0              \n"
 576         "mov        %1,         %%"REG_a"   \n"
 577         "1:                                 \n"
 578         "movaps %%xmm0,    (%0, %%"REG_a")  \n"
 579         "movaps %%xmm0,  16(%0, %%"REG_a")  \n"
 580         "movaps %%xmm0,  32(%0, %%"REG_a")  \n"
 581         "movaps %%xmm0,  48(%0, %%"REG_a")  \n"
 582         "movaps %%xmm0,  64(%0, %%"REG_a")  \n"
 583         "movaps %%xmm0,  80(%0, %%"REG_a")  \n"
 584         "movaps %%xmm0,  96(%0, %%"REG_a")  \n"
 585         "movaps %%xmm0, 112(%0, %%"REG_a")  \n"
 586         "add      $128,         %%"REG_a"   \n"
 587         "js         1b                      \n"
 588         :: "r"(((uint8_t *)blocks) + 128 * 6),
 589            "i"(-128 * 6)
 590         : "%"REG_a
 591     );
 592 }
 593
 594 static void add_bytes_mmx(uint8_t *dst, uint8_t *src, int w)
 595 {
 596     x86_reg i = 0;
 597     __asm__ volatile (
 598         "jmp          2f                \n\t"
 599         "1:                             \n\t"
 600         "movq   (%1, %0), %%mm0         \n\t"
 601         "movq   (%2, %0), %%mm1         \n\t"
 602         "paddb     %%mm0, %%mm1         \n\t"
 603         "movq      %%mm1, (%2, %0)      \n\t"
 604         "movq  8(%1, %0), %%mm0         \n\t"
 605         "movq  8(%2, %0), %%mm1         \n\t"
 606         "paddb     %%mm0, %%mm1         \n\t"
 607         "movq      %%mm1, 8(%2, %0)     \n\t"
 608         "add         $16, %0            \n\t"
 609         "2:                             \n\t"
 610         "cmp          %3, %0            \n\t"
 611         "js           1b                \n\t"
 612         : "+r"(i)
 613         : "r"(src), "r"(dst), "r"((x86_reg)w - 15)
 614     );
 615     for ( ; i < w; i++)
 616         dst[i + 0] += src[i + 0];
 617 }
 618
 619 #if HAVE_7REGS
 620 static void add_hfyu_median_prediction_cmov(uint8_t *dst, const uint8_t *top,
 621                                             const uint8_t *diff, int w,
 622                                             int *left, int *left_top)
 623 {
 624     x86_reg w2 = -w;
 625     x86_reg x;
 626     int l  = *left     & 0xff;
 627     int tl = *left_top & 0xff;
 628     int t;
 629     __asm__ volatile (
 630         "mov          %7, %3            \n"
 631         "1:                             \n"
 632         "movzbl (%3, %4), %2            \n"
 633         "mov          %2, %k3           \n"
 634         "sub         %b1, %b3           \n"
 635         "add         %b0, %b3           \n"
 636         "mov          %2, %1            \n"
 637         "cmp          %0, %2            \n"
 638         "cmovg        %0, %2            \n"
 639         "cmovg        %1, %0            \n"
 640         "cmp         %k3, %0            \n"
 641         "cmovg       %k3, %0            \n"
 642         "mov          %7, %3            \n"
 643         "cmp          %2, %0            \n"
 644         "cmovl        %2, %0            \n"
 645         "add    (%6, %4), %b0           \n"
 646         "mov         %b0, (%5, %4)      \n"
 647         "inc          %4                \n"
 648         "jl           1b                \n"
 649         : "+&q"(l), "+&q"(tl), "=&r"(t), "=&q"(x), "+&r"(w2)
 650         : "r"(dst + w), "r"(diff + w), "rm"(top + w)
 651     );
 652     *left     = l;
 653     *left_top = tl;
 654 }
 655 #endif
 656 #endif /* HAVE_INLINE_ASM */
 657
 658 void ff_h263_v_loop_filter_mmx(uint8_t *src, int stride, int qscale);
 659 void ff_h263_h_loop_filter_mmx(uint8_t *src, int stride, int qscale);
 660
 661 #if HAVE_INLINE_ASM
 662 /* Draw the edges of width 'w' of an image of size width, height
 663  * this MMX version can only handle w == 8 || w == 16. */
 664 static void draw_edges_mmx(uint8_t *buf, int wrap, int width, int height,
 665                            int w, int h, int sides)
 666 {
 667     uint8_t *ptr, *last_line;
 668     int i;
 669
 670     last_line = buf + (height - 1) * wrap;
 671     /* left and right */
 672     ptr = buf;
 673     if (w == 8) {
 674         __asm__ volatile (
 675             "1:                             \n\t"
 676             "movd            (%0), %%mm0    \n\t"
 677             "punpcklbw      %%mm0, %%mm0    \n\t"
 678             "punpcklwd      %%mm0, %%mm0    \n\t"
 679             "punpckldq      %%mm0, %%mm0    \n\t"
 680             "movq           %%mm0, -8(%0)   \n\t"
 681             "movq      -8(%0, %2), %%mm1    \n\t"
 682             "punpckhbw      %%mm1, %%mm1    \n\t"
 683             "punpckhwd      %%mm1, %%mm1    \n\t"
 684             "punpckhdq      %%mm1, %%mm1    \n\t"
 685             "movq           %%mm1, (%0, %2) \n\t"
 686             "add               %1, %0       \n\t"
 687             "cmp               %3, %0       \n\t"
 688             "jb                1b           \n\t"
 689             : "+r"(ptr)
 690             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
 691             );
 692     } else if(w==16){
 693         __asm__ volatile (
 694             "1:                                 \n\t"
 695             "movd            (%0), %%mm0        \n\t"
 696             "punpcklbw      %%mm0, %%mm0        \n\t"
 697             "punpcklwd      %%mm0, %%mm0        \n\t"
 698             "punpckldq      %%mm0, %%mm0        \n\t"
 699             "movq           %%mm0, -8(%0)       \n\t"
 700             "movq           %%mm0, -16(%0)      \n\t"
 701             "movq      -8(%0, %2), %%mm1        \n\t"
 702             "punpckhbw      %%mm1, %%mm1        \n\t"
 703             "punpckhwd      %%mm1, %%mm1        \n\t"
 704             "punpckhdq      %%mm1, %%mm1        \n\t"
 705             "movq           %%mm1,  (%0, %2)    \n\t"
 706             "movq           %%mm1, 8(%0, %2)    \n\t"
 707             "add               %1, %0           \n\t"
 708             "cmp               %3, %0           \n\t"
 709             "jb                1b               \n\t"
 710             : "+r"(ptr)
 711             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
 712             );
 713     } else {
 714         av_assert1(w == 4);
 715         __asm__ volatile (
 716             "1:                             \n\t"
 717             "movd            (%0), %%mm0    \n\t"
 718             "punpcklbw      %%mm0, %%mm0    \n\t"
 719             "punpcklwd      %%mm0, %%mm0    \n\t"
 720             "movd           %%mm0, -4(%0)   \n\t"
 721             "movd      -4(%0, %2), %%mm1    \n\t"
 722             "punpcklbw      %%mm1, %%mm1    \n\t"
 723             "punpckhwd      %%mm1, %%mm1    \n\t"
 724             "punpckhdq      %%mm1, %%mm1    \n\t"
 725             "movd           %%mm1, (%0, %2) \n\t"
 726             "add               %1, %0       \n\t"
 727             "cmp               %3, %0       \n\t"
 728             "jb                1b           \n\t"
 729             : "+r"(ptr)
 730             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
 731             );
 732     }
 733
 734     /* top and bottom (and hopefully also the corners) */
 735     if (sides & EDGE_TOP) {
 736         for (i = 0; i < h; i += 4) {
 737             ptr = buf - (i + 1) * wrap - w;
 738             __asm__ volatile (
 739                 "1:                             \n\t"
 740                 "movq (%1, %0), %%mm0           \n\t"
 741                 "movq    %%mm0, (%0)            \n\t"
 742                 "movq    %%mm0, (%0, %2)        \n\t"
 743                 "movq    %%mm0, (%0, %2, 2)     \n\t"
 744                 "movq    %%mm0, (%0, %3)        \n\t"
 745                 "add        $8, %0              \n\t"
 746                 "cmp        %4, %0              \n\t"
 747                 "jb         1b                  \n\t"
 748                 : "+r"(ptr)
 749                 : "r"((x86_reg)buf - (x86_reg)ptr - w), "r"((x86_reg) -wrap),
 750                   "r"((x86_reg) -wrap * 3), "r"(ptr + width + 2 * w)
 751                 );
 752         }
 753     }
 754
 755     if (sides & EDGE_BOTTOM) {
 756         for (i = 0; i < h; i += 4) {
 757             ptr = last_line + (i + 1) * wrap - w;
 758             __asm__ volatile (
 759                 "1:                             \n\t"
 760                 "movq (%1, %0), %%mm0           \n\t"
 761                 "movq    %%mm0, (%0)            \n\t"
 762                 "movq    %%mm0, (%0, %2)        \n\t"
 763                 "movq    %%mm0, (%0, %2, 2)     \n\t"
 764                 "movq    %%mm0, (%0, %3)        \n\t"
 765                 "add        $8, %0              \n\t"
 766                 "cmp        %4, %0              \n\t"
 767                 "jb         1b                  \n\t"
 768                 : "+r"(ptr)
 769                 : "r"((x86_reg)last_line - (x86_reg)ptr - w),
 770                   "r"((x86_reg)wrap), "r"((x86_reg)wrap * 3),
 771                   "r"(ptr + width + 2 * w)
 772                 );
 773         }
 774     }
 775 }
 776 #endif /* HAVE_INLINE_ASM */
 777
 778
 779 #if HAVE_YASM
 780 #define QPEL_OP(OPNAME, ROUNDER, RND, MMX)                              \
 781 static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src,   \
 782                                           int stride)                   \
 783 {                                                                       \
 784     ff_ ## OPNAME ## pixels8_ ## MMX(dst, src, stride, 8);              \
 785 }                                                                       \
 786                                                                         \
 787 static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src,    \
 788                                          int stride)                    \
 789 {                                                                       \
 790     uint64_t temp[8];                                                   \
 791     uint8_t * const half = (uint8_t*)temp;                              \
 792     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8,        \
 793                                                    stride, 8);          \
 794     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src, half,                 \
 795                                         stride, stride, 8);             \
 796 }                                                                       \
 797                                                                         \
 798 static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src,    \
 799                                          int stride)                    \
 800 {                                                                       \
 801     ff_ ## OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride,    \
 802                                                    stride, 8);          \
 803 }                                                                       \
 804                                                                         \
 805 static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src,    \
 806                                          int stride)                    \
 807 {                                                                       \
 808     uint64_t temp[8];                                                   \
 809     uint8_t * const half = (uint8_t*)temp;                              \
 810     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8,        \
 811                                                    stride, 8);          \
 812     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src + 1, half, stride,     \
 813                                         stride, 8);                     \
 814 }                                                                       \
 815                                                                         \
 816 static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src,    \
 817                                          int stride)                    \
 818 {                                                                       \
 819     uint64_t temp[8];                                                   \
 820     uint8_t * const half = (uint8_t*)temp;                              \
 821     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src,           \
 822                                                    8, stride);          \
 823     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src, half,                 \
 824                                         stride, stride, 8);             \
 825 }                                                                       \
 826                                                                         \
 827 static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src,    \
 828                                          int stride)                    \
 829 {                                                                       \
 830     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src,            \
 831                                                    stride, stride);     \
 832 }                                                                       \
 833                                                                         \
 834 static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src,    \
 835                                          int stride)                    \
 836 {                                                                       \
 837     uint64_t temp[8];                                                   \
 838     uint8_t * const half = (uint8_t*)temp;                              \
 839     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src,           \
 840                                                    8, stride);          \
 841     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src + stride, half, stride,\
 842                                         stride, 8);                     \
 843 }                                                                       \
 844                                                                         \
 845 static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src,    \
 846                                          int stride)                    \
 847 {                                                                       \
 848     uint64_t half[8 + 9];                                               \
 849     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
 850     uint8_t * const halfHV = ((uint8_t*)half);                          \
 851     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 852                                                    stride, 9);          \
 853     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8,           \
 854                                         stride, 9);                     \
 855     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
 856     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV,             \
 857                                         stride, 8, 8);                  \
 858 }                                                                       \
 859                                                                         \
 860 static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src,    \
 861                                          int stride)                    \
 862 {                                                                       \
 863     uint64_t half[8 + 9];                                               \
 864     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
 865     uint8_t * const halfHV = ((uint8_t*)half);                          \
 866     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 867                                                    stride, 9);          \
 868     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,       \
 869                                         stride, 9);                     \
 870     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
 871     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV,             \
 872                                         stride, 8, 8);                  \
 873 }                                                                       \
 874                                                                         \
 875 static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src,    \
 876                                          int stride)                    \
 877 {                                                                       \
 878     uint64_t half[8 + 9];                                               \
 879     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
 880     uint8_t * const halfHV = ((uint8_t*)half);                          \
 881     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 882                                                    stride, 9);          \
 883     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8,           \
 884                                         stride, 9);                     \
 885     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
 886     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV,         \
 887                                         stride, 8, 8);                  \
 888 }                                                                       \
 889                                                                         \
 890 static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src,    \
 891                                          int stride)                    \
 892 {                                                                       \
 893     uint64_t half[8 + 9];                                               \
 894     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
 895     uint8_t * const halfHV = ((uint8_t*)half);                          \
 896     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 897                                                    stride, 9);          \
 898     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,       \
 899                                         stride, 9);                     \
 900     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
 901     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV,         \
 902                                         stride, 8, 8);                  \
 903 }                                                                       \
 904                                                                         \
 905 static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src,    \
 906                                          int stride)                    \
 907 {                                                                       \
 908     uint64_t half[8 + 9];                                               \
 909     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
 910     uint8_t * const halfHV = ((uint8_t*)half);                          \
 911     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 912                                                    stride, 9);          \
 913     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
 914     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV,             \
 915                                         stride, 8, 8);                  \
 916 }                                                                       \
 917                                                                         \
 918 static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src,    \
 919                                          int stride)                    \
 920 {                                                                       \
 921     uint64_t half[8 + 9];                                               \
 922     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
 923     uint8_t * const halfHV = ((uint8_t*)half);                          \
 924     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 925                                                    stride, 9);          \
 926     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
 927     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV,         \
 928                                         stride, 8, 8);                  \
 929 }                                                                       \
 930                                                                         \
 931 static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src,    \
 932                                          int stride)                    \
 933 {                                                                       \
 934     uint64_t half[8 + 9];                                               \
 935     uint8_t * const halfH = ((uint8_t*)half);                           \
 936     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 937                                                    stride, 9);          \
 938     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH,              \
 939                                         8, stride, 9);                  \
 940     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH,          \
 941                                                    stride, 8);          \
 942 }                                                                       \
 943                                                                         \
 944 static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src,    \
 945                                          int stride)                    \
 946 {                                                                       \
 947     uint64_t half[8 + 9];                                               \
 948     uint8_t * const halfH = ((uint8_t*)half);                           \
 949     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 950                                                    stride, 9);          \
 951     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,       \
 952                                         stride, 9);                     \
 953     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH,          \
 954                                                    stride, 8);          \
 955 }                                                                       \
 956                                                                         \
 957 static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src,    \
 958                                          int stride)                    \
 959 {                                                                       \
 960     uint64_t half[9];                                                   \
 961     uint8_t * const halfH = ((uint8_t*)half);                           \
 962     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 963                                                    stride, 9);          \
 964     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH,          \
 965                                                    stride, 8);          \
 966 }                                                                       \
 967                                                                         \
 968 static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src,  \
 969                                            int stride)                  \
 970 {                                                                       \
 971     ff_ ## OPNAME ## pixels16_ ## MMX(dst, src, stride, 16);            \
 972 }                                                                       \
 973                                                                         \
 974 static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src,   \
 975                                           int stride)                   \
 976 {                                                                       \
 977     uint64_t temp[32];                                                  \
 978     uint8_t * const half = (uint8_t*)temp;                              \
 979     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16,      \
 980                                                     stride, 16);        \
 981     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride,        \
 982                                          stride, 16);                   \
 983 }                                                                       \
 984                                                                         \
 985 static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src,   \
 986                                           int stride)                   \
 987 {                                                                       \
 988     ff_ ## OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src,           \
 989                                                     stride, stride, 16);\
 990 }                                                                       \
 991                                                                         \
 992 static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src,   \
 993                                           int stride)                   \
 994 {                                                                       \
 995     uint64_t temp[32];                                                  \
 996     uint8_t * const half = (uint8_t*)temp;                              \
 997     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16,      \
 998                                                     stride, 16);        \
 999     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src + 1, half,            \
1000                                          stride, stride, 16);           \
1001 }                                                                       \
1002                                                                         \
1003 static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src,   \
1004                                           int stride)                   \
1005 {                                                                       \
1006     uint64_t temp[32];                                                  \
1007     uint8_t * const half = (uint8_t*)temp;                              \
1008     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16,      \
1009                                                     stride);            \
1010     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride,        \
1011                                          stride, 16);                   \
1012 }                                                                       \
1013                                                                         \
1014 static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src,   \
1015                                           int stride)                   \
1016 {                                                                       \
1017     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src,           \
1018                                                     stride, stride);    \
1019 }                                                                       \
1020                                                                         \
1021 static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src,   \
1022                                           int stride)                   \
1023 {                                                                       \
1024     uint64_t temp[32];                                                  \
1025     uint8_t * const half = (uint8_t*)temp;                              \
1026     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16,      \
1027                                                     stride);            \
1028     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half,         \
1029                                          stride, stride, 16);           \
1030 }                                                                       \
1031                                                                         \
1032 static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src,   \
1033                                           int stride)                   \
1034 {                                                                       \
1035     uint64_t half[16 * 2 + 17 * 2];                                     \
1036     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1037     uint8_t * const halfHV = ((uint8_t*)half);                          \
1038     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1039                                                     stride, 17);        \
1040     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,         \
1041                                          stride, 17);                   \
1042     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1043                                                     16, 16);            \
1044     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV,            \
1045                                          stride, 16, 16);               \
1046 }                                                                       \
1047                                                                         \
1048 static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src,   \
1049                                           int stride)                   \
1050 {                                                                       \
1051     uint64_t half[16 * 2 + 17 * 2];                                     \
1052     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1053     uint8_t * const halfHV = ((uint8_t*)half);                          \
1054     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1055                                                     stride, 17);        \
1056     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,     \
1057                                          stride, 17);                   \
1058     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1059                                                     16, 16);            \
1060     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV,            \
1061                                          stride, 16, 16);               \
1062 }                                                                       \
1063                                                                         \
1064 static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src,   \
1065                                           int stride)                   \
1066 {                                                                       \
1067     uint64_t half[16 * 2 + 17 * 2];                                     \
1068     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1069     uint8_t * const halfHV = ((uint8_t*)half);                          \
1070     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1071                                                     stride, 17);        \
1072     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,         \
1073                                          stride, 17);                   \
1074     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1075                                                     16, 16);            \
1076     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV,       \
1077                                          stride, 16, 16);               \
1078 }                                                                       \
1079                                                                         \
1080 static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src,   \
1081                                           int stride)                   \
1082 {                                                                       \
1083     uint64_t half[16 * 2 + 17 * 2];                                     \
1084     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1085     uint8_t * const halfHV = ((uint8_t*)half);                          \
1086     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1087                                                     stride, 17);        \
1088     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,     \
1089                                          stride, 17);                   \
1090     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1091                                                     16, 16);            \
1092     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV,       \
1093                                          stride, 16, 16);               \
1094 }                                                                       \
1095                                                                         \
1096 static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src,   \
1097                                           int stride)                   \
1098 {                                                                       \
1099     uint64_t half[16 * 2 + 17 * 2];                                     \
1100     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1101     uint8_t * const halfHV = ((uint8_t*)half);                          \
1102     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1103                                                     stride, 17);        \
1104     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1105                                                     16, 16);            \
1106     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV,            \
1107                                          stride, 16, 16);               \
1108 }                                                                       \
1109                                                                         \
1110 static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src,   \
1111                                           int stride)                   \
1112 {                                                                       \
1113     uint64_t half[16 * 2 + 17 * 2];                                     \
1114     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1115     uint8_t * const halfHV = ((uint8_t*)half);                          \
1116     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1117                                                     stride, 17);        \
1118     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1119                                                     16, 16);            \
1120     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV,       \
1121                                          stride, 16, 16);               \
1122 }                                                                       \
1123                                                                         \
1124 static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src,   \
1125                                           int stride)                   \
1126 {                                                                       \
1127     uint64_t half[17 * 2];                                              \
1128     uint8_t * const halfH = ((uint8_t*)half);                           \
1129     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1130                                                     stride, 17);        \
1131     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,         \
1132                                          stride, 17);                   \
1133     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH,         \
1134                                                     stride, 16);        \
1135 }                                                                       \
1136                                                                         \
1137 static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src,   \
1138                                           int stride)                   \
1139 {                                                                       \
1140     uint64_t half[17 * 2];                                              \
1141     uint8_t * const halfH = ((uint8_t*)half);                           \
1142     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1143                                                     stride, 17);        \
1144     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,     \
1145                                          stride, 17);                   \
1146     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH,         \
1147                                                     stride, 16);        \
1148 }                                                                       \
1149                                                                         \
1150 static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src,   \
1151                                           int stride)                   \
1152 {                                                                       \
1153     uint64_t half[17 * 2];                                              \
1154     uint8_t * const halfH = ((uint8_t*)half);                           \
1155     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1156                                                     stride, 17);        \
1157     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH,         \
1158                                                     stride, 16);        \
1159 }
1160
1161 QPEL_OP(put_,          ff_pw_16, _,        mmxext)
1162 QPEL_OP(avg_,          ff_pw_16, _,        mmxext)
1163 QPEL_OP(put_no_rnd_,   ff_pw_15, _no_rnd_, mmxext)
1164 #endif /* HAVE_YASM */
1165
1166
1167 #if HAVE_INLINE_ASM
1168 void ff_put_rv40_qpel8_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1169 {
1170   put_pixels8_xy2_mmx(dst, src, stride, 8);
1171 }
1172 void ff_put_rv40_qpel16_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1173 {
1174   put_pixels16_xy2_mmx(dst, src, stride, 16);
1175 }
1176 void ff_avg_rv40_qpel8_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1177 {
1178   avg_pixels8_xy2_mmx(dst, src, stride, 8);
1179 }
1180 void ff_avg_rv40_qpel16_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1181 {
1182   avg_pixels16_xy2_mmx(dst, src, stride, 16);
1183 }
1184
1185 typedef void emulated_edge_mc_func(uint8_t *dst, const uint8_t *src,
1186                                    ptrdiff_t linesize, int block_w, int block_h,
1187                                    int src_x, int src_y, int w, int h);
1188
1189 static av_always_inline void gmc(uint8_t *dst, uint8_t *src,
1190                                  int stride, int h, int ox, int oy,
1191                                  int dxx, int dxy, int dyx, int dyy,
1192                                  int shift, int r, int width, int height,
1193                                  emulated_edge_mc_func *emu_edge_fn)
1194 {
1195     const int w    = 8;
1196     const int ix   = ox  >> (16 + shift);
1197     const int iy   = oy  >> (16 + shift);
1198     const int oxs  = ox  >> 4;
1199     const int oys  = oy  >> 4;
1200     const int dxxs = dxx >> 4;
1201     const int dxys = dxy >> 4;
1202     const int dyxs = dyx >> 4;
1203     const int dyys = dyy >> 4;
1204     const uint16_t r4[4]   = { r, r, r, r };
1205     const uint16_t dxy4[4] = { dxys, dxys, dxys, dxys };
1206     const uint16_t dyy4[4] = { dyys, dyys, dyys, dyys };
1207     const uint64_t shift2 = 2 * shift;
1208 #define MAX_STRIDE 4096U
1209 #define MAX_H 8U
1210     uint8_t edge_buf[(MAX_H + 1) * MAX_STRIDE];
1211     int x, y;
1212
1213     const int dxw = (dxx - (1 << (16 + shift))) * (w - 1);
1214     const int dyh = (dyy - (1 << (16 + shift))) * (h - 1);
1215     const int dxh = dxy * (h - 1);
1216     const int dyw = dyx * (w - 1);
1217     int need_emu =  (unsigned)ix >= width  - w ||
1218                     (unsigned)iy >= height - h;
1219
1220     if ( // non-constant fullpel offset (3% of blocks)
1221         ((ox ^ (ox + dxw)) | (ox ^ (ox + dxh)) | (ox ^ (ox + dxw + dxh)) |
1222          (oy ^ (oy + dyw)) | (oy ^ (oy + dyh)) | (oy ^ (oy + dyw + dyh))) >> (16 + shift)
1223         // uses more than 16 bits of subpel mv (only at huge resolution)
1224         || (dxx | dxy | dyx | dyy) & 15
1225         || (need_emu && (h > MAX_H || stride > MAX_STRIDE))) {
1226         // FIXME could still use mmx for some of the rows
1227         ff_gmc_c(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy,
1228                  shift, r, width, height);
1229         return;
1230     }
1231
1232     src += ix + iy * stride;
1233     if (need_emu) {
1234         emu_edge_fn(edge_buf, src, stride, w + 1, h + 1, ix, iy, width, height);
1235         src = edge_buf;
1236     }
1237
1238     __asm__ volatile (
1239         "movd         %0, %%mm6         \n\t"
1240         "pxor      %%mm7, %%mm7         \n\t"
1241         "punpcklwd %%mm6, %%mm6         \n\t"
1242         "punpcklwd %%mm6, %%mm6         \n\t"
1243         :: "r"(1<<shift)
1244     );
1245
1246     for (x = 0; x < w; x += 4) {
1247         uint16_t dx4[4] = { oxs - dxys + dxxs * (x + 0),
1248                             oxs - dxys + dxxs * (x + 1),
1249                             oxs - dxys + dxxs * (x + 2),
1250                             oxs - dxys + dxxs * (x + 3) };
1251         uint16_t dy4[4] = { oys - dyys + dyxs * (x + 0),
1252                             oys - dyys + dyxs * (x + 1),
1253                             oys - dyys + dyxs * (x + 2),
1254                             oys - dyys + dyxs * (x + 3) };
1255
1256         for (y = 0; y < h; y++) {
1257             __asm__ volatile (
1258                 "movq      %0, %%mm4    \n\t"
1259                 "movq      %1, %%mm5    \n\t"
1260                 "paddw     %2, %%mm4    \n\t"
1261                 "paddw     %3, %%mm5    \n\t"
1262                 "movq   %%mm4, %0       \n\t"
1263                 "movq   %%mm5, %1       \n\t"
1264                 "psrlw    $12, %%mm4    \n\t"
1265                 "psrlw    $12, %%mm5    \n\t"
1266                 : "+m"(*dx4), "+m"(*dy4)
1267                 : "m"(*dxy4), "m"(*dyy4)
1268             );
1269
1270             __asm__ volatile (
1271                 "movq      %%mm6, %%mm2 \n\t"
1272                 "movq      %%mm6, %%mm1 \n\t"
1273                 "psubw     %%mm4, %%mm2 \n\t"
1274                 "psubw     %%mm5, %%mm1 \n\t"
1275                 "movq      %%mm2, %%mm0 \n\t"
1276                 "movq      %%mm4, %%mm3 \n\t"
1277                 "pmullw    %%mm1, %%mm0 \n\t" // (s - dx) * (s - dy)
1278                 "pmullw    %%mm5, %%mm3 \n\t" // dx * dy
1279                 "pmullw    %%mm5, %%mm2 \n\t" // (s - dx) * dy
1280                 "pmullw    %%mm4, %%mm1 \n\t" // dx * (s - dy)
1281
1282                 "movd         %4, %%mm5 \n\t"
1283                 "movd         %3, %%mm4 \n\t"
1284                 "punpcklbw %%mm7, %%mm5 \n\t"
1285                 "punpcklbw %%mm7, %%mm4 \n\t"
1286                 "pmullw    %%mm5, %%mm3 \n\t" // src[1, 1] * dx * dy
1287                 "pmullw    %%mm4, %%mm2 \n\t" // src[0, 1] * (s - dx) * dy
1288
1289                 "movd         %2, %%mm5 \n\t"
1290                 "movd         %1, %%mm4 \n\t"
1291                 "punpcklbw %%mm7, %%mm5 \n\t"
1292                 "punpcklbw %%mm7, %%mm4 \n\t"
1293                 "pmullw    %%mm5, %%mm1 \n\t" // src[1, 0] * dx * (s - dy)
1294                 "pmullw    %%mm4, %%mm0 \n\t" // src[0, 0] * (s - dx) * (s - dy)
1295                 "paddw        %5, %%mm1 \n\t"
1296                 "paddw     %%mm3, %%mm2 \n\t"
1297                 "paddw     %%mm1, %%mm0 \n\t"
1298                 "paddw     %%mm2, %%mm0 \n\t"
1299
1300                 "psrlw        %6, %%mm0 \n\t"
1301                 "packuswb  %%mm0, %%mm0 \n\t"
1302                 "movd      %%mm0, %0    \n\t"
1303
1304                 : "=m"(dst[x + y * stride])
1305                 : "m"(src[0]), "m"(src[1]),
1306                   "m"(src[stride]), "m"(src[stride + 1]),
1307                   "m"(*r4), "m"(shift2)
1308             );
1309             src += stride;
1310         }
1311         src += 4 - h * stride;
1312     }
1313 }
1314
1315 #if CONFIG_VIDEODSP
1316 #if HAVE_YASM
1317 #if ARCH_X86_32
1318 static void gmc_mmx(uint8_t *dst, uint8_t *src,
1319                     int stride, int h, int ox, int oy,
1320                     int dxx, int dxy, int dyx, int dyy,
1321                     int shift, int r, int width, int height)
1322 {
1323     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,
1324         width, height, &ff_emulated_edge_mc_8);
1325 }
1326 #endif
1327 static void gmc_sse(uint8_t *dst, uint8_t *src,
1328                     int stride, int h, int ox, int oy,
1329                     int dxx, int dxy, int dyx, int dyy,
1330                     int shift, int r, int width, int height)
1331 {
1332     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,
1333         width, height, &ff_emulated_edge_mc_8);
1334 }
1335 #else
1336 static void gmc_mmx(uint8_t *dst, uint8_t *src,
1337                     int stride, int h, int ox, int oy,
1338                     int dxx, int dxy, int dyx, int dyy,
1339                     int shift, int r, int width, int height)
1340 {
1341     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,
1342         width, height, &ff_emulated_edge_mc_8);
1343 }
1344 #endif
1345 #endif
1346
1347 #endif /* HAVE_INLINE_ASM */
1348
1349 void ff_put_pixels16_sse2(uint8_t *block, const uint8_t *pixels,
1350                           ptrdiff_t line_size, int h);
1351 void ff_avg_pixels16_sse2(uint8_t *block, const uint8_t *pixels,
1352                           ptrdiff_t line_size, int h);
1353
1354 #if HAVE_INLINE_ASM
1355
1356 /* CAVS-specific */
1357 void ff_put_cavs_qpel8_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1358 {
1359     put_pixels8_mmx(dst, src, stride, 8);
1360 }
1361
1362 void ff_avg_cavs_qpel8_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1363 {
1364     avg_pixels8_mmx(dst, src, stride, 8);
1365 }
1366
1367 void ff_put_cavs_qpel16_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1368 {
1369     put_pixels16_mmx(dst, src, stride, 16);
1370 }
1371
1372 void ff_avg_cavs_qpel16_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1373 {
1374     avg_pixels16_mmx(dst, src, stride, 16);
1375 }
1376 #endif /* HAVE_INLINE_ASM */
1377
1378 #if HAVE_YASM
1379 /* VC-1-specific */
1380 void ff_put_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src,
1381                                int stride, int rnd)
1382 {
1383     ff_put_pixels8_mmx(dst, src, stride, 8);
1384 }
1385
1386 void ff_avg_vc1_mspel_mc00_mmxext(uint8_t *dst, const uint8_t *src,
1387                                   int stride, int rnd)
1388 {
1389     ff_avg_pixels8_mmxext(dst, src, stride, 8);
1390 }
1391 #endif /* HAVE_YASM */
1392
1393 #if CONFIG_DIRAC_DECODER
1394 #define DIRAC_PIXOP(OPNAME2, OPNAME, EXT)\
1395 void ff_ ## OPNAME2 ## _dirac_pixels8_ ## EXT(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
1396 {\
1397     if (h&3)\
1398         ff_ ## OPNAME2 ## _dirac_pixels8_c(dst, src, stride, h);\
1399     else\
1400         OPNAME ## _pixels8_ ## EXT(dst, src[0], stride, h);\
1401 }\
1402 void ff_ ## OPNAME2 ## _dirac_pixels16_ ## EXT(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
1403 {\
1404     if (h&3)\
1405         ff_ ## OPNAME2 ## _dirac_pixels16_c(dst, src, stride, h);\
1406     else\
1407         OPNAME ## _pixels16_ ## EXT(dst, src[0], stride, h);\
1408 }\
1409 void ff_ ## OPNAME2 ## _dirac_pixels32_ ## EXT(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
1410 {\
1411     if (h&3) {\
1412         ff_ ## OPNAME2 ## _dirac_pixels32_c(dst, src, stride, h);\
1413     } else {\
1414         OPNAME ## _pixels16_ ## EXT(dst   , src[0]   , stride, h);\
1415         OPNAME ## _pixels16_ ## EXT(dst+16, src[0]+16, stride, h);\
1416     }\
1417 }
1418
1419 #if HAVE_MMX_INLINE
1420 DIRAC_PIXOP(put, put, mmx)
1421 DIRAC_PIXOP(avg, avg, mmx)
1422 #endif
1423
1424 #if HAVE_YASM
1425 DIRAC_PIXOP(avg, ff_avg, mmxext)
1426
1427 void ff_put_dirac_pixels16_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h)
1428 {
1429     if (h&3)
1430         ff_put_dirac_pixels16_c(dst, src, stride, h);
1431     else
1432     ff_put_pixels16_sse2(dst, src[0], stride, h);
1433 }
1434 void ff_avg_dirac_pixels16_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h)
1435 {
1436     if (h&3)
1437         ff_avg_dirac_pixels16_c(dst, src, stride, h);
1438     else
1439     ff_avg_pixels16_sse2(dst, src[0], stride, h);
1440 }
1441 void ff_put_dirac_pixels32_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h)
1442 {
1443     if (h&3) {
1444         ff_put_dirac_pixels32_c(dst, src, stride, h);
1445     } else {
1446     ff_put_pixels16_sse2(dst   , src[0]   , stride, h);
1447     ff_put_pixels16_sse2(dst+16, src[0]+16, stride, h);
1448     }
1449 }
1450 void ff_avg_dirac_pixels32_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h)
1451 {
1452     if (h&3) {
1453         ff_avg_dirac_pixels32_c(dst, src, stride, h);
1454     } else {
1455     ff_avg_pixels16_sse2(dst   , src[0]   , stride, h);
1456     ff_avg_pixels16_sse2(dst+16, src[0]+16, stride, h);
1457     }
1458 }
1459 #endif
1460 #endif
1461
1462 /* XXX: Those functions should be suppressed ASAP when all IDCTs are
1463  * converted. */
1464 #if CONFIG_GPL
1465 static void ff_libmpeg2mmx_idct_put(uint8_t *dest, int line_size,
1466                                     int16_t *block)
1467 {
1468     ff_mmx_idct(block);
1469     ff_put_pixels_clamped_mmx(block, dest, line_size);
1470 }
1471
1472 static void ff_libmpeg2mmx_idct_add(uint8_t *dest, int line_size,
1473                                     int16_t *block)
1474 {
1475     ff_mmx_idct(block);
1476     ff_add_pixels_clamped_mmx(block, dest, line_size);
1477 }
1478
1479 static void ff_libmpeg2mmx2_idct_put(uint8_t *dest, int line_size,
1480                                      int16_t *block)
1481 {
1482     ff_mmxext_idct(block);
1483     ff_put_pixels_clamped_mmx(block, dest, line_size);
1484 }
1485
1486 static void ff_libmpeg2mmx2_idct_add(uint8_t *dest, int line_size,
1487                                      int16_t *block)
1488 {
1489     ff_mmxext_idct(block);
1490     ff_add_pixels_clamped_mmx(block, dest, line_size);
1491 }
1492 #endif
1493
1494 #if HAVE_INLINE_ASM
1495 static void vector_clipf_sse(float *dst, const float *src,
1496                              float min, float max, int len)
1497 {
1498     x86_reg i = (len - 16) * 4;
1499     __asm__ volatile (
1500         "movss          %3, %%xmm4      \n\t"
1501         "movss          %4, %%xmm5      \n\t"
1502         "shufps $0, %%xmm4, %%xmm4      \n\t"
1503         "shufps $0, %%xmm5, %%xmm5      \n\t"
1504         "1:                             \n\t"
1505         "movaps   (%2, %0), %%xmm0      \n\t" // 3/1 on intel
1506         "movaps 16(%2, %0), %%xmm1      \n\t"
1507         "movaps 32(%2, %0), %%xmm2      \n\t"
1508         "movaps 48(%2, %0), %%xmm3      \n\t"
1509         "maxps      %%xmm4, %%xmm0      \n\t"
1510         "maxps      %%xmm4, %%xmm1      \n\t"
1511         "maxps      %%xmm4, %%xmm2      \n\t"
1512         "maxps      %%xmm4, %%xmm3      \n\t"
1513         "minps      %%xmm5, %%xmm0      \n\t"
1514         "minps      %%xmm5, %%xmm1      \n\t"
1515         "minps      %%xmm5, %%xmm2      \n\t"
1516         "minps      %%xmm5, %%xmm3      \n\t"
1517         "movaps     %%xmm0,   (%1, %0)  \n\t"
1518         "movaps     %%xmm1, 16(%1, %0)  \n\t"
1519         "movaps     %%xmm2, 32(%1, %0)  \n\t"
1520         "movaps     %%xmm3, 48(%1, %0)  \n\t"
1521         "sub           $64, %0          \n\t"
1522         "jge            1b              \n\t"
1523         : "+&r"(i)
1524         : "r"(dst), "r"(src), "m"(min), "m"(max)
1525         : "memory"
1526     );
1527 }
1528
1529 #endif /* HAVE_INLINE_ASM */
1530
1531 int32_t ff_scalarproduct_int16_mmxext(const int16_t *v1, const int16_t *v2,
1532                                       int order);
1533 int32_t ff_scalarproduct_int16_sse2(const int16_t *v1, const int16_t *v2,
1534                                     int order);
1535 int32_t ff_scalarproduct_and_madd_int16_mmxext(int16_t *v1, const int16_t *v2,
1536                                                const int16_t *v3,
1537                                                int order, int mul);
1538 int32_t ff_scalarproduct_and_madd_int16_sse2(int16_t *v1, const int16_t *v2,
1539                                              const int16_t *v3,
1540                                              int order, int mul);
1541 int32_t ff_scalarproduct_and_madd_int16_ssse3(int16_t *v1, const int16_t *v2,
1542                                               const int16_t *v3,
1543                                               int order, int mul);
1544
1545 void ff_apply_window_int16_round_mmxext(int16_t *output, const int16_t *input,
1546                                         const int16_t *window, unsigned int len);
1547 void ff_apply_window_int16_round_sse2(int16_t *output, const int16_t *input,
1548                                       const int16_t *window, unsigned int len);
1549 void ff_apply_window_int16_mmxext(int16_t *output, const int16_t *input,
1550                                   const int16_t *window, unsigned int len);
1551 void ff_apply_window_int16_sse2(int16_t *output, const int16_t *input,
1552                                 const int16_t *window, unsigned int len);
1553 void ff_apply_window_int16_ssse3(int16_t *output, const int16_t *input,
1554                                  const int16_t *window, unsigned int len);
1555 void ff_apply_window_int16_ssse3_atom(int16_t *output, const int16_t *input,
1556                                       const int16_t *window, unsigned int len);
1557
1558 void ff_bswap32_buf_ssse3(uint32_t *dst, const uint32_t *src, int w);
1559 void ff_bswap32_buf_sse2(uint32_t *dst, const uint32_t *src, int w);
1560
1561 void ff_add_hfyu_median_prediction_mmxext(uint8_t *dst, const uint8_t *top,
1562                                           const uint8_t *diff, int w,
1563                                           int *left, int *left_top);
1564 int  ff_add_hfyu_left_prediction_ssse3(uint8_t *dst, const uint8_t *src,
1565                                        int w, int left);
1566 int  ff_add_hfyu_left_prediction_sse4(uint8_t *dst, const uint8_t *src,
1567                                       int w, int left);
1568
1569 void ff_vector_clip_int32_mmx     (int32_t *dst, const int32_t *src,
1570                                    int32_t min, int32_t max, unsigned int len);
1571 void ff_vector_clip_int32_sse2    (int32_t *dst, const int32_t *src,
1572                                    int32_t min, int32_t max, unsigned int len);
1573 void ff_vector_clip_int32_int_sse2(int32_t *dst, const int32_t *src,
1574                                    int32_t min, int32_t max, unsigned int len);
1575 void ff_vector_clip_int32_sse4    (int32_t *dst, const int32_t *src,
1576                                    int32_t min, int32_t max, unsigned int len);
1577
1578 #define SET_QPEL_FUNCS(PFX, IDX, SIZE, CPU, PREFIX)                          \
1579     do {                                                                     \
1580     c->PFX ## _pixels_tab[IDX][ 0] = PREFIX ## PFX ## SIZE ## _mc00_ ## CPU; \
1581     c->PFX ## _pixels_tab[IDX][ 1] = PREFIX ## PFX ## SIZE ## _mc10_ ## CPU; \
1582     c->PFX ## _pixels_tab[IDX][ 2] = PREFIX ## PFX ## SIZE ## _mc20_ ## CPU; \
1583     c->PFX ## _pixels_tab[IDX][ 3] = PREFIX ## PFX ## SIZE ## _mc30_ ## CPU; \
1584     c->PFX ## _pixels_tab[IDX][ 4] = PREFIX ## PFX ## SIZE ## _mc01_ ## CPU; \
1585     c->PFX ## _pixels_tab[IDX][ 5] = PREFIX ## PFX ## SIZE ## _mc11_ ## CPU; \
1586     c->PFX ## _pixels_tab[IDX][ 6] = PREFIX ## PFX ## SIZE ## _mc21_ ## CPU; \
1587     c->PFX ## _pixels_tab[IDX][ 7] = PREFIX ## PFX ## SIZE ## _mc31_ ## CPU; \
1588     c->PFX ## _pixels_tab[IDX][ 8] = PREFIX ## PFX ## SIZE ## _mc02_ ## CPU; \
1589     c->PFX ## _pixels_tab[IDX][ 9] = PREFIX ## PFX ## SIZE ## _mc12_ ## CPU; \
1590     c->PFX ## _pixels_tab[IDX][10] = PREFIX ## PFX ## SIZE ## _mc22_ ## CPU; \
1591     c->PFX ## _pixels_tab[IDX][11] = PREFIX ## PFX ## SIZE ## _mc32_ ## CPU; \
1592     c->PFX ## _pixels_tab[IDX][12] = PREFIX ## PFX ## SIZE ## _mc03_ ## CPU; \
1593     c->PFX ## _pixels_tab[IDX][13] = PREFIX ## PFX ## SIZE ## _mc13_ ## CPU; \
1594     c->PFX ## _pixels_tab[IDX][14] = PREFIX ## PFX ## SIZE ## _mc23_ ## CPU; \
1595     c->PFX ## _pixels_tab[IDX][15] = PREFIX ## PFX ## SIZE ## _mc33_ ## CPU; \
1596     } while (0)
1597
1598 #define SET_HPEL_FUNCS(PFX, IDX, SIZE, CPU)                                     \
1599     do {                                                                        \
1600         c->PFX ## _pixels_tab IDX [0] = PFX ## _pixels ## SIZE ## _     ## CPU; \
1601         c->PFX ## _pixels_tab IDX [1] = PFX ## _pixels ## SIZE ## _x2_  ## CPU; \
1602         c->PFX ## _pixels_tab IDX [2] = PFX ## _pixels ## SIZE ## _y2_  ## CPU; \
1603         c->PFX ## _pixels_tab IDX [3] = PFX ## _pixels ## SIZE ## _xy2_ ## CPU; \
1604     } while (0)
1605
1606 static av_cold void dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx,
1607                                      int mm_flags)
1608 {
1609     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
1610
1611 #if HAVE_INLINE_ASM
1612     c->put_pixels_clamped        = ff_put_pixels_clamped_mmx;
1613     c->put_signed_pixels_clamped = ff_put_signed_pixels_clamped_mmx;
1614     c->add_pixels_clamped        = ff_add_pixels_clamped_mmx;
1615
1616     if (!high_bit_depth) {
1617         c->clear_block  = clear_block_mmx;
1618         c->clear_blocks = clear_blocks_mmx;
1619         c->draw_edges   = draw_edges_mmx;
1620
1621         SET_HPEL_FUNCS(put,        [0], 16, mmx);
1622         SET_HPEL_FUNCS(put_no_rnd, [0], 16, mmx);
1623         SET_HPEL_FUNCS(avg,        [0], 16, mmx);
1624         SET_HPEL_FUNCS(avg_no_rnd,    , 16, mmx);
1625         SET_HPEL_FUNCS(put,        [1],  8, mmx);
1626         SET_HPEL_FUNCS(put_no_rnd, [1],  8, mmx);
1627         SET_HPEL_FUNCS(avg,        [1],  8, mmx);
1628     }
1629
1630 #if CONFIG_VIDEODSP && (ARCH_X86_32 || !HAVE_YASM)
1631     c->gmc = gmc_mmx;
1632 #endif
1633
1634     c->add_bytes = add_bytes_mmx;
1635 #endif /* HAVE_INLINE_ASM */
1636
1637 #if HAVE_YASM
1638     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
1639         c->h263_v_loop_filter = ff_h263_v_loop_filter_mmx;
1640         c->h263_h_loop_filter = ff_h263_h_loop_filter_mmx;
1641     }
1642
1643     c->vector_clip_int32 = ff_vector_clip_int32_mmx;
1644 #endif
1645
1646 }
1647
1648 static av_cold void dsputil_init_mmxext(DSPContext *c, AVCodecContext *avctx,
1649                                         int mm_flags)
1650 {
1651     const int bit_depth      = avctx->bits_per_raw_sample;
1652     const int high_bit_depth = bit_depth > 8;
1653
1654 #if HAVE_YASM
1655     SET_QPEL_FUNCS(avg_qpel,        0, 16, mmxext, );
1656     SET_QPEL_FUNCS(avg_qpel,        1,  8, mmxext, );
1657
1658     SET_QPEL_FUNCS(put_qpel,        0, 16, mmxext, );
1659     SET_QPEL_FUNCS(put_qpel,        1,  8, mmxext, );
1660     SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, mmxext, );
1661     SET_QPEL_FUNCS(put_no_rnd_qpel, 1,  8, mmxext, );
1662
1663     if (!high_bit_depth) {
1664         c->put_pixels_tab[0][1] = ff_put_pixels16_x2_mmxext;
1665         c->put_pixels_tab[0][2] = ff_put_pixels16_y2_mmxext;
1666
1667         c->avg_pixels_tab[0][0] = ff_avg_pixels16_mmxext;
1668         c->avg_pixels_tab[0][1] = ff_avg_pixels16_x2_mmxext;
1669         c->avg_pixels_tab[0][2] = ff_avg_pixels16_y2_mmxext;
1670
1671         c->put_pixels_tab[1][1] = ff_put_pixels8_x2_mmxext;
1672         c->put_pixels_tab[1][2] = ff_put_pixels8_y2_mmxext;
1673
1674         c->avg_pixels_tab[1][0] = ff_avg_pixels8_mmxext;
1675         c->avg_pixels_tab[1][1] = ff_avg_pixels8_x2_mmxext;
1676         c->avg_pixels_tab[1][2] = ff_avg_pixels8_y2_mmxext;
1677     }
1678
1679     if (!(avctx->flags & CODEC_FLAG_BITEXACT)) {
1680         if (!high_bit_depth) {
1681             c->put_no_rnd_pixels_tab[0][1] = ff_put_no_rnd_pixels16_x2_mmxext;
1682             c->put_no_rnd_pixels_tab[0][2] = ff_put_no_rnd_pixels16_y2_mmxext;
1683             c->put_no_rnd_pixels_tab[1][1] = ff_put_no_rnd_pixels8_x2_mmxext;
1684             c->put_no_rnd_pixels_tab[1][2] = ff_put_no_rnd_pixels8_y2_mmxext;
1685
1686             c->avg_pixels_tab[0][3] = ff_avg_pixels16_xy2_mmxext;
1687             c->avg_pixels_tab[1][3] = ff_avg_pixels8_xy2_mmxext;
1688         }
1689     }
1690 #endif /* HAVE_YASM */
1691
1692 #if HAVE_MMXEXT_EXTERNAL
1693     if (CONFIG_VP3_DECODER && (avctx->codec_id == AV_CODEC_ID_VP3 ||
1694                                avctx->codec_id == AV_CODEC_ID_THEORA)) {
1695         c->put_no_rnd_pixels_tab[1][1] = ff_put_no_rnd_pixels8_x2_exact_mmxext;
1696         c->put_no_rnd_pixels_tab[1][2] = ff_put_no_rnd_pixels8_y2_exact_mmxext;
1697     }
1698
1699     /* slower than cmov version on AMD */
1700     if (!(mm_flags & AV_CPU_FLAG_3DNOW))
1701         c->add_hfyu_median_prediction = ff_add_hfyu_median_prediction_mmxext;
1702
1703     c->scalarproduct_int16          = ff_scalarproduct_int16_mmxext;
1704     c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_mmxext;
1705
1706     if (avctx->flags & CODEC_FLAG_BITEXACT) {
1707         c->apply_window_int16 = ff_apply_window_int16_mmxext;
1708     } else {
1709         c->apply_window_int16 = ff_apply_window_int16_round_mmxext;
1710     }
1711 #endif /* HAVE_MMXEXT_EXTERNAL */
1712 }
1713
1714 static av_cold void dsputil_init_3dnow(DSPContext *c, AVCodecContext *avctx,
1715                                        int mm_flags)
1716 {
1717     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
1718
1719 #if HAVE_YASM
1720     if (!high_bit_depth) {
1721         c->put_pixels_tab[0][1] = ff_put_pixels16_x2_3dnow;
1722         c->put_pixels_tab[0][2] = ff_put_pixels16_y2_3dnow;
1723
1724         c->avg_pixels_tab[0][0] = ff_avg_pixels16_3dnow;
1725         c->avg_pixels_tab[0][1] = ff_avg_pixels16_x2_3dnow;
1726         c->avg_pixels_tab[0][2] = ff_avg_pixels16_y2_3dnow;
1727
1728         c->put_pixels_tab[1][1] = ff_put_pixels8_x2_3dnow;
1729         c->put_pixels_tab[1][2] = ff_put_pixels8_y2_3dnow;
1730
1731         c->avg_pixels_tab[1][0] = ff_avg_pixels8_3dnow;
1732         c->avg_pixels_tab[1][1] = ff_avg_pixels8_x2_3dnow;
1733         c->avg_pixels_tab[1][2] = ff_avg_pixels8_y2_3dnow;
1734
1735         if (!(avctx->flags & CODEC_FLAG_BITEXACT)){
1736             c->put_no_rnd_pixels_tab[0][1] = ff_put_no_rnd_pixels16_x2_3dnow;
1737             c->put_no_rnd_pixels_tab[0][2] = ff_put_no_rnd_pixels16_y2_3dnow;
1738             c->put_no_rnd_pixels_tab[1][1] = ff_put_no_rnd_pixels8_x2_3dnow;
1739             c->put_no_rnd_pixels_tab[1][2] = ff_put_no_rnd_pixels8_y2_3dnow;
1740
1741             c->avg_pixels_tab[0][3] = ff_avg_pixels16_xy2_3dnow;
1742             c->avg_pixels_tab[1][3] = ff_avg_pixels8_xy2_3dnow;
1743         }
1744     }
1745
1746     if (CONFIG_VP3_DECODER && (avctx->codec_id == AV_CODEC_ID_VP3 ||
1747                                avctx->codec_id == AV_CODEC_ID_THEORA)) {
1748         c->put_no_rnd_pixels_tab[1][1] = ff_put_no_rnd_pixels8_x2_exact_3dnow;
1749         c->put_no_rnd_pixels_tab[1][2] = ff_put_no_rnd_pixels8_y2_exact_3dnow;
1750     }
1751 #endif /* HAVE_YASM */
1752 }
1753
1754 static av_cold void dsputil_init_sse(DSPContext *c, AVCodecContext *avctx,
1755                                      int mm_flags)
1756 {
1757     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
1758
1759 #if HAVE_INLINE_ASM
1760     if (!high_bit_depth) {
1761         if (!(CONFIG_MPEG_XVMC_DECODER && avctx->xvmc_acceleration > 1)) {
1762             /* XvMCCreateBlocks() may not allocate 16-byte aligned blocks */
1763             c->clear_block  = clear_block_sse;
1764             c->clear_blocks = clear_blocks_sse;
1765         }
1766     }
1767
1768     c->vector_clipf = vector_clipf_sse;
1769 #endif /* HAVE_INLINE_ASM */
1770
1771 #if HAVE_YASM
1772 #if HAVE_INLINE_ASM && CONFIG_VIDEODSP
1773     c->gmc = gmc_sse;
1774 #endif
1775 #endif /* HAVE_YASM */
1776 }
1777
1778 static av_cold void dsputil_init_sse2(DSPContext *c, AVCodecContext *avctx,
1779                                       int mm_flags)
1780 {
1781     const int bit_depth      = avctx->bits_per_raw_sample;
1782     const int high_bit_depth = bit_depth > 8;
1783
1784 #if HAVE_SSE2_INLINE
1785     if (!high_bit_depth && avctx->idct_algo == FF_IDCT_XVIDMMX) {
1786         c->idct_put              = ff_idct_xvid_sse2_put;
1787         c->idct_add              = ff_idct_xvid_sse2_add;
1788         c->idct                  = ff_idct_xvid_sse2;
1789         c->idct_permutation_type = FF_SSE2_IDCT_PERM;
1790     }
1791 #endif /* HAVE_SSE2_INLINE */
1792
1793 #if HAVE_SSE2_EXTERNAL
1794     if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {
1795         // these functions are slower than mmx on AMD, but faster on Intel
1796         if (!high_bit_depth) {
1797             c->put_pixels_tab[0][0]        = ff_put_pixels16_sse2;
1798             c->put_no_rnd_pixels_tab[0][0] = ff_put_pixels16_sse2;
1799             c->avg_pixels_tab[0][0]        = ff_avg_pixels16_sse2;
1800         }
1801     }
1802
1803     c->scalarproduct_int16          = ff_scalarproduct_int16_sse2;
1804     c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_sse2;
1805     if (mm_flags & AV_CPU_FLAG_ATOM) {
1806         c->vector_clip_int32 = ff_vector_clip_int32_int_sse2;
1807     } else {
1808         c->vector_clip_int32 = ff_vector_clip_int32_sse2;
1809     }
1810     if (avctx->flags & CODEC_FLAG_BITEXACT) {
1811         c->apply_window_int16 = ff_apply_window_int16_sse2;
1812     } else if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {
1813         c->apply_window_int16 = ff_apply_window_int16_round_sse2;
1814     }
1815     c->bswap_buf = ff_bswap32_buf_sse2;
1816 #endif /* HAVE_SSE2_EXTERNAL */
1817 }
1818
1819 static av_cold void dsputil_init_ssse3(DSPContext *c, AVCodecContext *avctx,
1820                                        int mm_flags)
1821 {
1822 #if HAVE_SSSE3_EXTERNAL
1823     c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_ssse3;
1824     if (mm_flags & AV_CPU_FLAG_SSE4) // not really sse4, just slow on Conroe
1825         c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_sse4;
1826
1827     if (mm_flags & AV_CPU_FLAG_ATOM)
1828         c->apply_window_int16 = ff_apply_window_int16_ssse3_atom;
1829     else
1830         c->apply_window_int16 = ff_apply_window_int16_ssse3;
1831     if (!(mm_flags & (AV_CPU_FLAG_SSE42|AV_CPU_FLAG_3DNOW))) // cachesplit
1832         c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_ssse3;
1833     c->bswap_buf = ff_bswap32_buf_ssse3;
1834 #endif /* HAVE_SSSE3_EXTERNAL */
1835 }
1836
1837 static av_cold void dsputil_init_sse4(DSPContext *c, AVCodecContext *avctx,
1838                                       int mm_flags)
1839 {
1840 #if HAVE_SSE4_EXTERNAL
1841     c->vector_clip_int32 = ff_vector_clip_int32_sse4;
1842 #endif /* HAVE_SSE4_EXTERNAL */
1843 }
1844
1845 av_cold void ff_dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx)
1846 {
1847     int mm_flags = av_get_cpu_flags();
1848
1849 #if HAVE_7REGS && HAVE_INLINE_ASM
1850     if (mm_flags & AV_CPU_FLAG_CMOV)
1851         c->add_hfyu_median_prediction = add_hfyu_median_prediction_cmov;
1852 #endif
1853
1854     if (mm_flags & AV_CPU_FLAG_MMX) {
1855 #if HAVE_INLINE_ASM
1856         const int idct_algo = avctx->idct_algo;
1857
1858         if (avctx->lowres == 0 && avctx->bits_per_raw_sample <= 8) {
1859             if (idct_algo == FF_IDCT_AUTO || idct_algo == FF_IDCT_SIMPLEMMX) {
1860                 c->idct_put              = ff_simple_idct_put_mmx;
1861                 c->idct_add              = ff_simple_idct_add_mmx;
1862                 c->idct                  = ff_simple_idct_mmx;
1863                 c->idct_permutation_type = FF_SIMPLE_IDCT_PERM;
1864 #if CONFIG_GPL
1865             } else if (idct_algo == FF_IDCT_LIBMPEG2MMX) {
1866                 if (mm_flags & AV_CPU_FLAG_MMX2) {
1867                     c->idct_put = ff_libmpeg2mmx2_idct_put;
1868                     c->idct_add = ff_libmpeg2mmx2_idct_add;
1869                     c->idct     = ff_mmxext_idct;
1870                 } else {
1871                     c->idct_put = ff_libmpeg2mmx_idct_put;
1872                     c->idct_add = ff_libmpeg2mmx_idct_add;
1873                     c->idct     = ff_mmx_idct;
1874                 }
1875                 c->idct_permutation_type = FF_LIBMPEG2_IDCT_PERM;
1876 #endif
1877             } else if (idct_algo == FF_IDCT_XVIDMMX) {
1878                 if (mm_flags & AV_CPU_FLAG_SSE2) {
1879                     c->idct_put              = ff_idct_xvid_sse2_put;
1880                     c->idct_add              = ff_idct_xvid_sse2_add;
1881                     c->idct                  = ff_idct_xvid_sse2;
1882                     c->idct_permutation_type = FF_SSE2_IDCT_PERM;
1883                 } else if (mm_flags & AV_CPU_FLAG_MMXEXT) {
1884                     c->idct_put              = ff_idct_xvid_mmxext_put;
1885                     c->idct_add              = ff_idct_xvid_mmxext_add;
1886                     c->idct                  = ff_idct_xvid_mmxext;
1887                 } else {
1888                     c->idct_put              = ff_idct_xvid_mmx_put;
1889                     c->idct_add              = ff_idct_xvid_mmx_add;
1890                     c->idct                  = ff_idct_xvid_mmx;
1891                 }
1892             }
1893         }
1894 #endif /* HAVE_INLINE_ASM */
1895
1896         dsputil_init_mmx(c, avctx, mm_flags);
1897     }
1898
1899     if (mm_flags & AV_CPU_FLAG_MMXEXT)
1900         dsputil_init_mmxext(c, avctx, mm_flags);
1901
1902     if (mm_flags & AV_CPU_FLAG_3DNOW)
1903         dsputil_init_3dnow(c, avctx, mm_flags);
1904
1905     if (mm_flags & AV_CPU_FLAG_SSE)
1906         dsputil_init_sse(c, avctx, mm_flags);
1907
1908     if (mm_flags & AV_CPU_FLAG_SSE2)
1909         dsputil_init_sse2(c, avctx, mm_flags);
1910
1911     if (mm_flags & AV_CPU_FLAG_SSSE3)
1912         dsputil_init_ssse3(c, avctx, mm_flags);
1913
1914     if (mm_flags & AV_CPU_FLAG_SSE4)
1915         dsputil_init_sse4(c, avctx, mm_flags);
1916
1917     if (CONFIG_ENCODERS)
1918         ff_dsputilenc_init_mmx(c, avctx);
1919 }