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 "dsputil_mmx.h"
  33 #include "idct_xvid.h"
  34 #include "diracdsp_mmx.h"
  35
  36 //#undef NDEBUG
  37 //#include <assert.h>
  38
  39 /* pixel operations */
  40 DECLARE_ALIGNED(8,  const uint64_t, ff_bone) = 0x0101010101010101ULL;
  41 DECLARE_ALIGNED(8,  const uint64_t, ff_wtwo) = 0x0002000200020002ULL;
  42
  43 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_1)    = { 0x0001000100010001ULL, 0x0001000100010001ULL };
  44 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_2)    = { 0x0002000200020002ULL, 0x0002000200020002ULL };
  45 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_3)    = { 0x0003000300030003ULL, 0x0003000300030003ULL };
  46 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_4)    = { 0x0004000400040004ULL, 0x0004000400040004ULL };
  47 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_5)    = { 0x0005000500050005ULL, 0x0005000500050005ULL };
  48 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_8)    = { 0x0008000800080008ULL, 0x0008000800080008ULL };
  49 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_9)    = { 0x0009000900090009ULL, 0x0009000900090009ULL };
  50 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_15)   =   0x000F000F000F000FULL;
  51 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_16)   = { 0x0010001000100010ULL, 0x0010001000100010ULL };
  52 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_17)   = { 0x0011001100110011ULL, 0x0011001100110011ULL };
  53 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_18)   = { 0x0012001200120012ULL, 0x0012001200120012ULL };
  54 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_20)   =   0x0014001400140014ULL;
  55 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_27)   = { 0x001B001B001B001BULL, 0x001B001B001B001BULL };
  56 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_28)   = { 0x001C001C001C001CULL, 0x001C001C001C001CULL };
  57 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_32)   = { 0x0020002000200020ULL, 0x0020002000200020ULL };
  58 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_42)   =   0x002A002A002A002AULL;
  59 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_53)   =   0x0035003500350035ULL;
  60 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_63)   = { 0x003F003F003F003FULL, 0x003F003F003F003FULL };
  61 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_64)   = { 0x0040004000400040ULL, 0x0040004000400040ULL };
  62 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_96)   =   0x0060006000600060ULL;
  63 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_128)  =   0x0080008000800080ULL;
  64 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_255)  =   0x00ff00ff00ff00ffULL;
  65 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_512)  = { 0x0200020002000200ULL, 0x0200020002000200ULL };
  66 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_1019) = { 0x03FB03FB03FB03FBULL, 0x03FB03FB03FB03FBULL };
  67
  68 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_0)    = { 0x0000000000000000ULL, 0x0000000000000000ULL };
  69 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_1)    = { 0x0101010101010101ULL, 0x0101010101010101ULL };
  70 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_3)    = { 0x0303030303030303ULL, 0x0303030303030303ULL };
  71 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_4)    = { 0x0404040404040404ULL, 0x0404040404040404ULL };
  72 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_7)    =   0x0707070707070707ULL;
  73 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_1F)   =   0x1F1F1F1F1F1F1F1FULL;
  74 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_3F)   =   0x3F3F3F3F3F3F3F3FULL;
  75 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_80)   = { 0x8080808080808080ULL, 0x8080808080808080ULL };
  76 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_81)   =   0x8181818181818181ULL;
  77 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_A1)   = { 0xA1A1A1A1A1A1A1A1ULL, 0xA1A1A1A1A1A1A1A1ULL };
  78 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_F8)   = { 0xF8F8F8F8F8F8F8F8ULL, 0xF8F8F8F8F8F8F8F8ULL };
  79 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_FC)   =   0xFCFCFCFCFCFCFCFCULL;
  80 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_FE)   = { 0xFEFEFEFEFEFEFEFEULL, 0xFEFEFEFEFEFEFEFEULL };
  81
  82 DECLARE_ALIGNED(16, const double, ff_pd_1)[2] = { 1.0, 1.0 };
  83 DECLARE_ALIGNED(16, const double, ff_pd_2)[2] = { 2.0, 2.0 };
  84
  85
  86 #if HAVE_YASM
  87 void ff_put_pixels8_x2_mmxext(uint8_t *block, const uint8_t *pixels,
  88                               ptrdiff_t line_size, int h);
  89 void ff_put_pixels8_x2_3dnow(uint8_t *block, const uint8_t *pixels,
  90                              ptrdiff_t line_size, int h);
  91 void ff_put_pixels8_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
  92                               int dstStride, int src1Stride, int h);
  93 void ff_put_no_rnd_pixels8_l2_mmxext(uint8_t *dst, uint8_t *src1,
  94                                      uint8_t *src2, int dstStride,
  95                                      int src1Stride, int h);
  96 void ff_avg_pixels8_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
  97                               int dstStride, int src1Stride, int h);
  98 void ff_put_pixels16_x2_mmxext(uint8_t *block, const uint8_t *pixels,
  99                                ptrdiff_t line_size, int h);
 100 void ff_put_pixels16_x2_3dnow(uint8_t *block, const uint8_t *pixels,
 101                               ptrdiff_t line_size, int h);
 102 void ff_put_pixels16_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
 103                                int dstStride, int src1Stride, int h);
 104 void ff_avg_pixels16_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
 105                                int dstStride, int src1Stride, int h);
 106 void ff_put_no_rnd_pixels16_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
 107                                       int dstStride, int src1Stride, int h);
 108 void ff_put_no_rnd_pixels8_x2_mmxext(uint8_t *block, const uint8_t *pixels,
 109                                      ptrdiff_t line_size, int h);
 110 void ff_put_no_rnd_pixels8_x2_3dnow(uint8_t *block, const uint8_t *pixels,
 111                                     ptrdiff_t line_size, int h);
 112 void ff_put_no_rnd_pixels8_x2_exact_mmxext(uint8_t *block,
 113                                            const uint8_t *pixels,
 114                                            ptrdiff_t line_size, int h);
 115 void ff_put_no_rnd_pixels8_x2_exact_3dnow(uint8_t *block,
 116                                           const uint8_t *pixels,
 117                                           ptrdiff_t line_size, int h);
 118 void ff_put_pixels8_y2_mmxext(uint8_t *block, const uint8_t *pixels,
 119                               ptrdiff_t line_size, int h);
 120 void ff_put_pixels8_y2_3dnow(uint8_t *block, const uint8_t *pixels,
 121                              ptrdiff_t line_size, int h);
 122 void ff_put_no_rnd_pixels8_y2_mmxext(uint8_t *block, const uint8_t *pixels,
 123                                      ptrdiff_t line_size, int h);
 124 void ff_put_no_rnd_pixels8_y2_3dnow(uint8_t *block, const uint8_t *pixels,
 125                                     ptrdiff_t line_size, int h);
 126 void ff_put_no_rnd_pixels8_y2_exact_mmxext(uint8_t *block,
 127                                            const uint8_t *pixels,
 128                                            ptrdiff_t line_size, int h);
 129 void ff_put_no_rnd_pixels8_y2_exact_3dnow(uint8_t *block,
 130                                           const uint8_t *pixels,
 131                                           ptrdiff_t line_size, int h);
 132 void ff_avg_pixels8_mmxext(uint8_t *block, const uint8_t *pixels,
 133                            ptrdiff_t line_size, int h);
 134 void ff_avg_pixels8_3dnow(uint8_t *block, const uint8_t *pixels,
 135                           ptrdiff_t line_size, int h);
 136 void ff_avg_pixels8_x2_mmxext(uint8_t *block, const uint8_t *pixels,
 137                               ptrdiff_t line_size, int h);
 138 void ff_avg_pixels8_x2_3dnow(uint8_t *block, const uint8_t *pixels,
 139                              ptrdiff_t line_size, int h);
 140 void ff_avg_pixels8_y2_mmxext(uint8_t *block, const uint8_t *pixels,
 141                               ptrdiff_t line_size, int h);
 142 void ff_avg_pixels8_y2_3dnow(uint8_t *block, const uint8_t *pixels,
 143                              ptrdiff_t line_size, int h);
 144 void ff_avg_pixels8_xy2_mmxext(uint8_t *block, const uint8_t *pixels,
 145                                ptrdiff_t line_size, int h);
 146 void ff_avg_pixels8_xy2_3dnow(uint8_t *block, const uint8_t *pixels,
 147                               ptrdiff_t line_size, int h);
 148
 149 void ff_put_pixels8_mmxext(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h);
 150 static void ff_put_pixels16_mmxext(uint8_t *block, const uint8_t *pixels,
 151                                    int line_size, int h)
 152 {
 153     ff_put_pixels8_mmxext(block,     pixels,     line_size, h);
 154     ff_put_pixels8_mmxext(block + 8, pixels + 8, line_size, h);
 155 }
 156
 157 void ff_put_mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 158                                          int dstStride, int srcStride, int h);
 159 void ff_avg_mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 160                                          int dstStride, int srcStride, int h);
 161 void ff_put_no_rnd_mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 162                                                  int dstStride, int srcStride,
 163                                                  int h);
 164 void ff_put_mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 165                                         int dstStride, int srcStride, int h);
 166 void ff_avg_mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 167                                         int dstStride, int srcStride, int h);
 168 void ff_put_no_rnd_mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 169                                                 int dstStride, int srcStride,
 170                                                 int h);
 171 void ff_put_mpeg4_qpel16_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 172                                          int dstStride, int srcStride);
 173 void ff_avg_mpeg4_qpel16_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 174                                          int dstStride, int srcStride);
 175 void ff_put_no_rnd_mpeg4_qpel16_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 176                                                  int dstStride, int srcStride);
 177 void ff_put_mpeg4_qpel8_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 178                                         int dstStride, int srcStride);
 179 void ff_avg_mpeg4_qpel8_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 180                                         int dstStride, int srcStride);
 181 void ff_put_no_rnd_mpeg4_qpel8_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 182                                                 int dstStride, int srcStride);
 183 #define ff_put_no_rnd_pixels16_mmxext ff_put_pixels16_mmxext
 184 #define ff_put_no_rnd_pixels8_mmxext ff_put_pixels8_mmxext
 185 #endif /* HAVE_YASM */
 186
 187
 188 #if HAVE_INLINE_ASM
 189
 190 #define JUMPALIGN()     __asm__ volatile (".p2align 3"::)
 191 #define MOVQ_ZERO(regd) __asm__ volatile ("pxor %%"#regd", %%"#regd ::)
 192
 193 #define MOVQ_BFE(regd)                                  \
 194     __asm__ volatile (                                  \
 195         "pcmpeqd %%"#regd", %%"#regd"   \n\t"           \
 196         "paddb   %%"#regd", %%"#regd"   \n\t" ::)
 197
 198 #ifndef PIC
 199 #define MOVQ_BONE(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_bone))
 200 #define MOVQ_WTWO(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_wtwo))
 201 #else
 202 // for shared library it's better to use this way for accessing constants
 203 // pcmpeqd -> -1
 204 #define MOVQ_BONE(regd)                                 \
 205     __asm__ volatile (                                  \
 206         "pcmpeqd  %%"#regd", %%"#regd"  \n\t"           \
 207         "psrlw          $15, %%"#regd"  \n\t"           \
 208         "packuswb %%"#regd", %%"#regd"  \n\t" ::)
 209
 210 #define MOVQ_WTWO(regd)                                 \
 211     __asm__ volatile (                                  \
 212         "pcmpeqd %%"#regd", %%"#regd"   \n\t"           \
 213         "psrlw         $15, %%"#regd"   \n\t"           \
 214         "psllw          $1, %%"#regd"   \n\t"::)
 215
 216 #endif
 217
 218 // using regr as temporary and for the output result
 219 // first argument is unmodifed and second is trashed
 220 // regfe is supposed to contain 0xfefefefefefefefe
 221 #define PAVGB_MMX_NO_RND(rega, regb, regr, regfe)                \
 222     "movq   "#rega", "#regr"            \n\t"                    \
 223     "pand   "#regb", "#regr"            \n\t"                    \
 224     "pxor   "#rega", "#regb"            \n\t"                    \
 225     "pand  "#regfe", "#regb"            \n\t"                    \
 226     "psrlq       $1, "#regb"            \n\t"                    \
 227     "paddb  "#regb", "#regr"            \n\t"
 228
 229 #define PAVGB_MMX(rega, regb, regr, regfe)                       \
 230     "movq   "#rega", "#regr"            \n\t"                    \
 231     "por    "#regb", "#regr"            \n\t"                    \
 232     "pxor   "#rega", "#regb"            \n\t"                    \
 233     "pand  "#regfe", "#regb"            \n\t"                    \
 234     "psrlq       $1, "#regb"            \n\t"                    \
 235     "psubb  "#regb", "#regr"            \n\t"
 236
 237 // mm6 is supposed to contain 0xfefefefefefefefe
 238 #define PAVGBP_MMX_NO_RND(rega, regb, regr,  regc, regd, regp)   \
 239     "movq  "#rega", "#regr"             \n\t"                    \
 240     "movq  "#regc", "#regp"             \n\t"                    \
 241     "pand  "#regb", "#regr"             \n\t"                    \
 242     "pand  "#regd", "#regp"             \n\t"                    \
 243     "pxor  "#rega", "#regb"             \n\t"                    \
 244     "pxor  "#regc", "#regd"             \n\t"                    \
 245     "pand    %%mm6, "#regb"             \n\t"                    \
 246     "pand    %%mm6, "#regd"             \n\t"                    \
 247     "psrlq      $1, "#regb"             \n\t"                    \
 248     "psrlq      $1, "#regd"             \n\t"                    \
 249     "paddb "#regb", "#regr"             \n\t"                    \
 250     "paddb "#regd", "#regp"             \n\t"
 251
 252 #define PAVGBP_MMX(rega, regb, regr, regc, regd, regp)           \
 253     "movq  "#rega", "#regr"             \n\t"                    \
 254     "movq  "#regc", "#regp"             \n\t"                    \
 255     "por   "#regb", "#regr"             \n\t"                    \
 256     "por   "#regd", "#regp"             \n\t"                    \
 257     "pxor  "#rega", "#regb"             \n\t"                    \
 258     "pxor  "#regc", "#regd"             \n\t"                    \
 259     "pand    %%mm6, "#regb"             \n\t"                    \
 260     "pand    %%mm6, "#regd"             \n\t"                    \
 261     "psrlq      $1, "#regd"             \n\t"                    \
 262     "psrlq      $1, "#regb"             \n\t"                    \
 263     "psubb "#regb", "#regr"             \n\t"                    \
 264     "psubb "#regd", "#regp"             \n\t"
 265
 266 /***********************************/
 267 /* MMX no rounding */
 268 #define NO_RND 1
 269 #define DEF(x, y) x ## _no_rnd_ ## y ## _mmx
 270 #define SET_RND  MOVQ_WONE
 271 #define PAVGBP(a, b, c, d, e, f)        PAVGBP_MMX_NO_RND(a, b, c, d, e, f)
 272 #define PAVGB(a, b, c, e)               PAVGB_MMX_NO_RND(a, b, c, e)
 273 #define OP_AVG(a, b, c, e)              PAVGB_MMX(a, b, c, e)
 274
 275 #include "dsputil_rnd_template.c"
 276
 277 #undef DEF
 278 #undef SET_RND
 279 #undef PAVGBP
 280 #undef PAVGB
 281 #undef NO_RND
 282 /***********************************/
 283 /* MMX rounding */
 284
 285 #define DEF(x, y) x ## _ ## y ## _mmx
 286 #define SET_RND  MOVQ_WTWO
 287 #define PAVGBP(a, b, c, d, e, f)        PAVGBP_MMX(a, b, c, d, e, f)
 288 #define PAVGB(a, b, c, e)               PAVGB_MMX(a, b, c, e)
 289
 290 #include "dsputil_rnd_template.c"
 291
 292 #undef DEF
 293 #undef SET_RND
 294 #undef PAVGBP
 295 #undef PAVGB
 296 #undef OP_AVG
 297
 298 #endif /* HAVE_INLINE_ASM */
 299
 300
 301 #if HAVE_YASM
 302 #define ff_put_pixels8_mmx ff_put_pixels8_mmxext
 303
 304 /***********************************/
 305 /* 3Dnow specific */
 306
 307 #define DEF(x) x ## _3dnow
 308
 309 #include "dsputil_avg_template.c"
 310
 311 #undef DEF
 312
 313 /***********************************/
 314 /* MMXEXT specific */
 315
 316 #define DEF(x) x ## _mmxext
 317
 318 #include "dsputil_avg_template.c"
 319
 320 #undef DEF
 321
 322 #endif /* HAVE_YASM */
 323
 324
 325 #if HAVE_INLINE_ASM
 326 #define put_no_rnd_pixels16_mmx put_pixels16_mmx
 327 #define put_no_rnd_pixels8_mmx put_pixels8_mmx
 328 #define put_pixels16_mmxext put_pixels16_mmx
 329 #define put_pixels8_mmxext put_pixels8_mmx
 330 #define put_pixels4_mmxext put_pixels4_mmx
 331 #define put_no_rnd_pixels16_mmxext put_no_rnd_pixels16_mmx
 332 #define put_no_rnd_pixels8_mmxext put_no_rnd_pixels8_mmx
 333
 334 /***********************************/
 335 /* standard MMX */
 336
 337 void ff_put_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
 338                                int line_size)
 339 {
 340     const int16_t *p;
 341     uint8_t *pix;
 342
 343     /* read the pixels */
 344     p   = block;
 345     pix = pixels;
 346     /* unrolled loop */
 347     __asm__ volatile (
 348         "movq      (%3), %%mm0          \n\t"
 349         "movq     8(%3), %%mm1          \n\t"
 350         "movq    16(%3), %%mm2          \n\t"
 351         "movq    24(%3), %%mm3          \n\t"
 352         "movq    32(%3), %%mm4          \n\t"
 353         "movq    40(%3), %%mm5          \n\t"
 354         "movq    48(%3), %%mm6          \n\t"
 355         "movq    56(%3), %%mm7          \n\t"
 356         "packuswb %%mm1, %%mm0          \n\t"
 357         "packuswb %%mm3, %%mm2          \n\t"
 358         "packuswb %%mm5, %%mm4          \n\t"
 359         "packuswb %%mm7, %%mm6          \n\t"
 360         "movq     %%mm0, (%0)           \n\t"
 361         "movq     %%mm2, (%0, %1)       \n\t"
 362         "movq     %%mm4, (%0, %1, 2)    \n\t"
 363         "movq     %%mm6, (%0, %2)       \n\t"
 364         :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3),
 365            "r"(p)
 366         : "memory");
 367     pix += line_size * 4;
 368     p   += 32;
 369
 370     // if here would be an exact copy of the code above
 371     // compiler would generate some very strange code
 372     // thus using "r"
 373     __asm__ volatile (
 374         "movq       (%3), %%mm0         \n\t"
 375         "movq      8(%3), %%mm1         \n\t"
 376         "movq     16(%3), %%mm2         \n\t"
 377         "movq     24(%3), %%mm3         \n\t"
 378         "movq     32(%3), %%mm4         \n\t"
 379         "movq     40(%3), %%mm5         \n\t"
 380         "movq     48(%3), %%mm6         \n\t"
 381         "movq     56(%3), %%mm7         \n\t"
 382         "packuswb  %%mm1, %%mm0         \n\t"
 383         "packuswb  %%mm3, %%mm2         \n\t"
 384         "packuswb  %%mm5, %%mm4         \n\t"
 385         "packuswb  %%mm7, %%mm6         \n\t"
 386         "movq      %%mm0, (%0)          \n\t"
 387         "movq      %%mm2, (%0, %1)      \n\t"
 388         "movq      %%mm4, (%0, %1, 2)   \n\t"
 389         "movq      %%mm6, (%0, %2)      \n\t"
 390         :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3), "r"(p)
 391         : "memory");
 392 }
 393
 394 #define put_signed_pixels_clamped_mmx_half(off)             \
 395     "movq          "#off"(%2), %%mm1        \n\t"           \
 396     "movq     16 + "#off"(%2), %%mm2        \n\t"           \
 397     "movq     32 + "#off"(%2), %%mm3        \n\t"           \
 398     "movq     48 + "#off"(%2), %%mm4        \n\t"           \
 399     "packsswb  8 + "#off"(%2), %%mm1        \n\t"           \
 400     "packsswb 24 + "#off"(%2), %%mm2        \n\t"           \
 401     "packsswb 40 + "#off"(%2), %%mm3        \n\t"           \
 402     "packsswb 56 + "#off"(%2), %%mm4        \n\t"           \
 403     "paddb              %%mm0, %%mm1        \n\t"           \
 404     "paddb              %%mm0, %%mm2        \n\t"           \
 405     "paddb              %%mm0, %%mm3        \n\t"           \
 406     "paddb              %%mm0, %%mm4        \n\t"           \
 407     "movq               %%mm1, (%0)         \n\t"           \
 408     "movq               %%mm2, (%0, %3)     \n\t"           \
 409     "movq               %%mm3, (%0, %3, 2)  \n\t"           \
 410     "movq               %%mm4, (%0, %1)     \n\t"
 411
 412 void ff_put_signed_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
 413                                       int line_size)
 414 {
 415     x86_reg line_skip = line_size;
 416     x86_reg line_skip3;
 417
 418     __asm__ volatile (
 419         "movq "MANGLE(ff_pb_80)", %%mm0     \n\t"
 420         "lea         (%3, %3, 2), %1        \n\t"
 421         put_signed_pixels_clamped_mmx_half(0)
 422         "lea         (%0, %3, 4), %0        \n\t"
 423         put_signed_pixels_clamped_mmx_half(64)
 424         : "+&r"(pixels), "=&r"(line_skip3)
 425         : "r"(block), "r"(line_skip)
 426         : "memory");
 427 }
 428
 429 void ff_add_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
 430                                int line_size)
 431 {
 432     const int16_t *p;
 433     uint8_t *pix;
 434     int i;
 435
 436     /* read the pixels */
 437     p   = block;
 438     pix = pixels;
 439     MOVQ_ZERO(mm7);
 440     i = 4;
 441     do {
 442         __asm__ volatile (
 443             "movq        (%2), %%mm0    \n\t"
 444             "movq       8(%2), %%mm1    \n\t"
 445             "movq      16(%2), %%mm2    \n\t"
 446             "movq      24(%2), %%mm3    \n\t"
 447             "movq          %0, %%mm4    \n\t"
 448             "movq          %1, %%mm6    \n\t"
 449             "movq       %%mm4, %%mm5    \n\t"
 450             "punpcklbw  %%mm7, %%mm4    \n\t"
 451             "punpckhbw  %%mm7, %%mm5    \n\t"
 452             "paddsw     %%mm4, %%mm0    \n\t"
 453             "paddsw     %%mm5, %%mm1    \n\t"
 454             "movq       %%mm6, %%mm5    \n\t"
 455             "punpcklbw  %%mm7, %%mm6    \n\t"
 456             "punpckhbw  %%mm7, %%mm5    \n\t"
 457             "paddsw     %%mm6, %%mm2    \n\t"
 458             "paddsw     %%mm5, %%mm3    \n\t"
 459             "packuswb   %%mm1, %%mm0    \n\t"
 460             "packuswb   %%mm3, %%mm2    \n\t"
 461             "movq       %%mm0, %0       \n\t"
 462             "movq       %%mm2, %1       \n\t"
 463             : "+m"(*pix), "+m"(*(pix + line_size))
 464             : "r"(p)
 465             : "memory");
 466         pix += line_size * 2;
 467         p   += 16;
 468     } while (--i);
 469 }
 470
 471 static void put_pixels8_mmx(uint8_t *block, const uint8_t *pixels,
 472                             ptrdiff_t line_size, int h)
 473 {
 474     __asm__ volatile (
 475         "lea   (%3, %3), %%"REG_a"      \n\t"
 476         ".p2align     3                 \n\t"
 477         "1:                             \n\t"
 478         "movq  (%1    ), %%mm0          \n\t"
 479         "movq  (%1, %3), %%mm1          \n\t"
 480         "movq     %%mm0, (%2)           \n\t"
 481         "movq     %%mm1, (%2, %3)       \n\t"
 482         "add  %%"REG_a", %1             \n\t"
 483         "add  %%"REG_a", %2             \n\t"
 484         "movq  (%1    ), %%mm0          \n\t"
 485         "movq  (%1, %3), %%mm1          \n\t"
 486         "movq     %%mm0, (%2)           \n\t"
 487         "movq     %%mm1, (%2, %3)       \n\t"
 488         "add  %%"REG_a", %1             \n\t"
 489         "add  %%"REG_a", %2             \n\t"
 490         "subl        $4, %0             \n\t"
 491         "jnz         1b                 \n\t"
 492         : "+g"(h), "+r"(pixels),  "+r"(block)
 493         : "r"((x86_reg)line_size)
 494         : "%"REG_a, "memory"
 495         );
 496 }
 497
 498 static void put_pixels16_mmx(uint8_t *block, const uint8_t *pixels,
 499                              ptrdiff_t line_size, int h)
 500 {
 501     __asm__ volatile (
 502         "lea   (%3, %3), %%"REG_a"      \n\t"
 503         ".p2align     3                 \n\t"
 504         "1:                             \n\t"
 505         "movq  (%1    ), %%mm0          \n\t"
 506         "movq 8(%1    ), %%mm4          \n\t"
 507         "movq  (%1, %3), %%mm1          \n\t"
 508         "movq 8(%1, %3), %%mm5          \n\t"
 509         "movq     %%mm0,  (%2)          \n\t"
 510         "movq     %%mm4, 8(%2)          \n\t"
 511         "movq     %%mm1,  (%2, %3)      \n\t"
 512         "movq     %%mm5, 8(%2, %3)      \n\t"
 513         "add  %%"REG_a", %1             \n\t"
 514         "add  %%"REG_a", %2             \n\t"
 515         "movq  (%1    ), %%mm0          \n\t"
 516         "movq 8(%1    ), %%mm4          \n\t"
 517         "movq  (%1, %3), %%mm1          \n\t"
 518         "movq 8(%1, %3), %%mm5          \n\t"
 519         "movq     %%mm0,  (%2)          \n\t"
 520         "movq     %%mm4, 8(%2)          \n\t"
 521         "movq     %%mm1,  (%2, %3)      \n\t"
 522         "movq     %%mm5, 8(%2, %3)      \n\t"
 523         "add  %%"REG_a", %1             \n\t"
 524         "add  %%"REG_a", %2             \n\t"
 525         "subl        $4, %0             \n\t"
 526         "jnz         1b                 \n\t"
 527         : "+g"(h), "+r"(pixels),  "+r"(block)
 528         : "r"((x86_reg)line_size)
 529         : "%"REG_a, "memory"
 530         );
 531 }
 532
 533 #define CLEAR_BLOCKS(name, n)                           \
 534 static void name(int16_t *blocks)                       \
 535 {                                                       \
 536     __asm__ volatile (                                  \
 537         "pxor %%mm7, %%mm7              \n\t"           \
 538         "mov     %1,        %%"REG_a"   \n\t"           \
 539         "1:                             \n\t"           \
 540         "movq %%mm7,   (%0, %%"REG_a")  \n\t"           \
 541         "movq %%mm7,  8(%0, %%"REG_a")  \n\t"           \
 542         "movq %%mm7, 16(%0, %%"REG_a")  \n\t"           \
 543         "movq %%mm7, 24(%0, %%"REG_a")  \n\t"           \
 544         "add    $32, %%"REG_a"          \n\t"           \
 545         "js      1b                     \n\t"           \
 546         :: "r"(((uint8_t *)blocks) + 128 * n),          \
 547            "i"(-128 * n)                                \
 548         : "%"REG_a                                      \
 549         );                                              \
 550 }
 551 CLEAR_BLOCKS(clear_blocks_mmx, 6)
 552 CLEAR_BLOCKS(clear_block_mmx, 1)
 553
 554 static void clear_block_sse(int16_t *block)
 555 {
 556     __asm__ volatile (
 557         "xorps  %%xmm0, %%xmm0          \n"
 558         "movaps %%xmm0,    (%0)         \n"
 559         "movaps %%xmm0,  16(%0)         \n"
 560         "movaps %%xmm0,  32(%0)         \n"
 561         "movaps %%xmm0,  48(%0)         \n"
 562         "movaps %%xmm0,  64(%0)         \n"
 563         "movaps %%xmm0,  80(%0)         \n"
 564         "movaps %%xmm0,  96(%0)         \n"
 565         "movaps %%xmm0, 112(%0)         \n"
 566         :: "r"(block)
 567         : "memory"
 568     );
 569 }
 570
 571 static void clear_blocks_sse(int16_t *blocks)
 572 {
 573     __asm__ volatile (
 574         "xorps  %%xmm0, %%xmm0              \n"
 575         "mov        %1,         %%"REG_a"   \n"
 576         "1:                                 \n"
 577         "movaps %%xmm0,    (%0, %%"REG_a")  \n"
 578         "movaps %%xmm0,  16(%0, %%"REG_a")  \n"
 579         "movaps %%xmm0,  32(%0, %%"REG_a")  \n"
 580         "movaps %%xmm0,  48(%0, %%"REG_a")  \n"
 581         "movaps %%xmm0,  64(%0, %%"REG_a")  \n"
 582         "movaps %%xmm0,  80(%0, %%"REG_a")  \n"
 583         "movaps %%xmm0,  96(%0, %%"REG_a")  \n"
 584         "movaps %%xmm0, 112(%0, %%"REG_a")  \n"
 585         "add      $128,         %%"REG_a"   \n"
 586         "js         1b                      \n"
 587         :: "r"(((uint8_t *)blocks) + 128 * 6),
 588            "i"(-128 * 6)
 589         : "%"REG_a
 590     );
 591 }
 592
 593 static void add_bytes_mmx(uint8_t *dst, uint8_t *src, int w)
 594 {
 595     x86_reg i = 0;
 596     __asm__ volatile (
 597         "jmp          2f                \n\t"
 598         "1:                             \n\t"
 599         "movq   (%1, %0), %%mm0         \n\t"
 600         "movq   (%2, %0), %%mm1         \n\t"
 601         "paddb     %%mm0, %%mm1         \n\t"
 602         "movq      %%mm1, (%2, %0)      \n\t"
 603         "movq  8(%1, %0), %%mm0         \n\t"
 604         "movq  8(%2, %0), %%mm1         \n\t"
 605         "paddb     %%mm0, %%mm1         \n\t"
 606         "movq      %%mm1, 8(%2, %0)     \n\t"
 607         "add         $16, %0            \n\t"
 608         "2:                             \n\t"
 609         "cmp          %3, %0            \n\t"
 610         "js           1b                \n\t"
 611         : "+r"(i)
 612         : "r"(src), "r"(dst), "r"((x86_reg)w - 15)
 613     );
 614     for ( ; i < w; i++)
 615         dst[i + 0] += src[i + 0];
 616 }
 617
 618 #if HAVE_7REGS
 619 static void add_hfyu_median_prediction_cmov(uint8_t *dst, const uint8_t *top,
 620                                             const uint8_t *diff, int w,
 621                                             int *left, int *left_top)
 622 {
 623     x86_reg w2 = -w;
 624     x86_reg x;
 625     int l  = *left     & 0xff;
 626     int tl = *left_top & 0xff;
 627     int t;
 628     __asm__ volatile (
 629         "mov          %7, %3            \n"
 630         "1:                             \n"
 631         "movzbl (%3, %4), %2            \n"
 632         "mov          %2, %k3           \n"
 633         "sub         %b1, %b3           \n"
 634         "add         %b0, %b3           \n"
 635         "mov          %2, %1            \n"
 636         "cmp          %0, %2            \n"
 637         "cmovg        %0, %2            \n"
 638         "cmovg        %1, %0            \n"
 639         "cmp         %k3, %0            \n"
 640         "cmovg       %k3, %0            \n"
 641         "mov          %7, %3            \n"
 642         "cmp          %2, %0            \n"
 643         "cmovl        %2, %0            \n"
 644         "add    (%6, %4), %b0           \n"
 645         "mov         %b0, (%5, %4)      \n"
 646         "inc          %4                \n"
 647         "jl           1b                \n"
 648         : "+&q"(l), "+&q"(tl), "=&r"(t), "=&q"(x), "+&r"(w2)
 649         : "r"(dst + w), "r"(diff + w), "rm"(top + w)
 650     );
 651     *left     = l;
 652     *left_top = tl;
 653 }
 654 #endif
 655 #endif /* HAVE_INLINE_ASM */
 656
 657 void ff_h263_v_loop_filter_mmx(uint8_t *src, int stride, int qscale);
 658 void ff_h263_h_loop_filter_mmx(uint8_t *src, int stride, int qscale);
 659
 660 #if HAVE_INLINE_ASM
 661 /* Draw the edges of width 'w' of an image of size width, height
 662  * this MMX version can only handle w == 8 || w == 16. */
 663 static void draw_edges_mmx(uint8_t *buf, int wrap, int width, int height,
 664                            int w, int h, int sides)
 665 {
 666     uint8_t *ptr, *last_line;
 667     int i;
 668
 669     last_line = buf + (height - 1) * wrap;
 670     /* left and right */
 671     ptr = buf;
 672     if (w == 8) {
 673         __asm__ volatile (
 674             "1:                             \n\t"
 675             "movd            (%0), %%mm0    \n\t"
 676             "punpcklbw      %%mm0, %%mm0    \n\t"
 677             "punpcklwd      %%mm0, %%mm0    \n\t"
 678             "punpckldq      %%mm0, %%mm0    \n\t"
 679             "movq           %%mm0, -8(%0)   \n\t"
 680             "movq      -8(%0, %2), %%mm1    \n\t"
 681             "punpckhbw      %%mm1, %%mm1    \n\t"
 682             "punpckhwd      %%mm1, %%mm1    \n\t"
 683             "punpckhdq      %%mm1, %%mm1    \n\t"
 684             "movq           %%mm1, (%0, %2) \n\t"
 685             "add               %1, %0       \n\t"
 686             "cmp               %3, %0       \n\t"
 687             "jb                1b           \n\t"
 688             : "+r"(ptr)
 689             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
 690             );
 691     } else if(w==16){
 692         __asm__ volatile (
 693             "1:                                 \n\t"
 694             "movd            (%0), %%mm0        \n\t"
 695             "punpcklbw      %%mm0, %%mm0        \n\t"
 696             "punpcklwd      %%mm0, %%mm0        \n\t"
 697             "punpckldq      %%mm0, %%mm0        \n\t"
 698             "movq           %%mm0, -8(%0)       \n\t"
 699             "movq           %%mm0, -16(%0)      \n\t"
 700             "movq      -8(%0, %2), %%mm1        \n\t"
 701             "punpckhbw      %%mm1, %%mm1        \n\t"
 702             "punpckhwd      %%mm1, %%mm1        \n\t"
 703             "punpckhdq      %%mm1, %%mm1        \n\t"
 704             "movq           %%mm1,  (%0, %2)    \n\t"
 705             "movq           %%mm1, 8(%0, %2)    \n\t"
 706             "add               %1, %0           \n\t"
 707             "cmp               %3, %0           \n\t"
 708             "jb                1b               \n\t"
 709             : "+r"(ptr)
 710             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
 711             );
 712     } else {
 713         av_assert1(w == 4);
 714         __asm__ volatile (
 715             "1:                             \n\t"
 716             "movd            (%0), %%mm0    \n\t"
 717             "punpcklbw      %%mm0, %%mm0    \n\t"
 718             "punpcklwd      %%mm0, %%mm0    \n\t"
 719             "movd           %%mm0, -4(%0)   \n\t"
 720             "movd      -4(%0, %2), %%mm1    \n\t"
 721             "punpcklbw      %%mm1, %%mm1    \n\t"
 722             "punpckhwd      %%mm1, %%mm1    \n\t"
 723             "punpckhdq      %%mm1, %%mm1    \n\t"
 724             "movd           %%mm1, (%0, %2) \n\t"
 725             "add               %1, %0       \n\t"
 726             "cmp               %3, %0       \n\t"
 727             "jb                1b           \n\t"
 728             : "+r"(ptr)
 729             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
 730             );
 731     }
 732
 733     /* top and bottom (and hopefully also the corners) */
 734     if (sides & EDGE_TOP) {
 735         for (i = 0; i < h; i += 4) {
 736             ptr = buf - (i + 1) * wrap - w;
 737             __asm__ volatile (
 738                 "1:                             \n\t"
 739                 "movq (%1, %0), %%mm0           \n\t"
 740                 "movq    %%mm0, (%0)            \n\t"
 741                 "movq    %%mm0, (%0, %2)        \n\t"
 742                 "movq    %%mm0, (%0, %2, 2)     \n\t"
 743                 "movq    %%mm0, (%0, %3)        \n\t"
 744                 "add        $8, %0              \n\t"
 745                 "cmp        %4, %0              \n\t"
 746                 "jb         1b                  \n\t"
 747                 : "+r"(ptr)
 748                 : "r"((x86_reg)buf - (x86_reg)ptr - w), "r"((x86_reg) -wrap),
 749                   "r"((x86_reg) -wrap * 3), "r"(ptr + width + 2 * w)
 750                 );
 751         }
 752     }
 753
 754     if (sides & EDGE_BOTTOM) {
 755         for (i = 0; i < h; i += 4) {
 756             ptr = last_line + (i + 1) * wrap - w;
 757             __asm__ volatile (
 758                 "1:                             \n\t"
 759                 "movq (%1, %0), %%mm0           \n\t"
 760                 "movq    %%mm0, (%0)            \n\t"
 761                 "movq    %%mm0, (%0, %2)        \n\t"
 762                 "movq    %%mm0, (%0, %2, 2)     \n\t"
 763                 "movq    %%mm0, (%0, %3)        \n\t"
 764                 "add        $8, %0              \n\t"
 765                 "cmp        %4, %0              \n\t"
 766                 "jb         1b                  \n\t"
 767                 : "+r"(ptr)
 768                 : "r"((x86_reg)last_line - (x86_reg)ptr - w),
 769                   "r"((x86_reg)wrap), "r"((x86_reg)wrap * 3),
 770                   "r"(ptr + width + 2 * w)
 771                 );
 772         }
 773     }
 774 }
 775 #endif /* HAVE_INLINE_ASM */
 776
 777
 778 #if HAVE_YASM
 779 #define QPEL_OP(OPNAME, ROUNDER, RND, OP, MMX)                          \
 780 static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src,   \
 781                                           int stride)                   \
 782 {                                                                       \
 783     ff_ ## OPNAME ## pixels8_ ## MMX(dst, src, stride, 8);              \
 784 }                                                                       \
 785                                                                         \
 786 static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src,    \
 787                                          int stride)                    \
 788 {                                                                       \
 789     uint64_t temp[8];                                                   \
 790     uint8_t * const half = (uint8_t*)temp;                              \
 791     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8,        \
 792                                                    stride, 8);          \
 793     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src, half,                 \
 794                                         stride, stride, 8);             \
 795 }                                                                       \
 796                                                                         \
 797 static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src,    \
 798                                          int stride)                    \
 799 {                                                                       \
 800     ff_ ## OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride,    \
 801                                                    stride, 8);          \
 802 }                                                                       \
 803                                                                         \
 804 static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src,    \
 805                                          int stride)                    \
 806 {                                                                       \
 807     uint64_t temp[8];                                                   \
 808     uint8_t * const half = (uint8_t*)temp;                              \
 809     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8,        \
 810                                                    stride, 8);          \
 811     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src + 1, half, stride,     \
 812                                         stride, 8);                     \
 813 }                                                                       \
 814                                                                         \
 815 static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src,    \
 816                                          int stride)                    \
 817 {                                                                       \
 818     uint64_t temp[8];                                                   \
 819     uint8_t * const half = (uint8_t*)temp;                              \
 820     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src,           \
 821                                                    8, stride);          \
 822     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src, half,                 \
 823                                         stride, stride, 8);             \
 824 }                                                                       \
 825                                                                         \
 826 static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src,    \
 827                                          int stride)                    \
 828 {                                                                       \
 829     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src,            \
 830                                                    stride, stride);     \
 831 }                                                                       \
 832                                                                         \
 833 static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src,    \
 834                                          int stride)                    \
 835 {                                                                       \
 836     uint64_t temp[8];                                                   \
 837     uint8_t * const half = (uint8_t*)temp;                              \
 838     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src,           \
 839                                                    8, stride);          \
 840     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src + stride, half, stride,\
 841                                         stride, 8);                     \
 842 }                                                                       \
 843                                                                         \
 844 static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src,    \
 845                                          int stride)                    \
 846 {                                                                       \
 847     uint64_t half[8 + 9];                                               \
 848     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
 849     uint8_t * const halfHV = ((uint8_t*)half);                          \
 850     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 851                                                    stride, 9);          \
 852     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8,           \
 853                                         stride, 9);                     \
 854     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
 855     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV,             \
 856                                         stride, 8, 8);                  \
 857 }                                                                       \
 858                                                                         \
 859 static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src,    \
 860                                          int stride)                    \
 861 {                                                                       \
 862     uint64_t half[8 + 9];                                               \
 863     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
 864     uint8_t * const halfHV = ((uint8_t*)half);                          \
 865     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 866                                                    stride, 9);          \
 867     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,       \
 868                                         stride, 9);                     \
 869     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
 870     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV,             \
 871                                         stride, 8, 8);                  \
 872 }                                                                       \
 873                                                                         \
 874 static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src,    \
 875                                          int stride)                    \
 876 {                                                                       \
 877     uint64_t half[8 + 9];                                               \
 878     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
 879     uint8_t * const halfHV = ((uint8_t*)half);                          \
 880     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 881                                                    stride, 9);          \
 882     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8,           \
 883                                         stride, 9);                     \
 884     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
 885     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV,         \
 886                                         stride, 8, 8);                  \
 887 }                                                                       \
 888                                                                         \
 889 static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src,    \
 890                                          int stride)                    \
 891 {                                                                       \
 892     uint64_t half[8 + 9];                                               \
 893     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
 894     uint8_t * const halfHV = ((uint8_t*)half);                          \
 895     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 896                                                    stride, 9);          \
 897     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,       \
 898                                         stride, 9);                     \
 899     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
 900     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV,         \
 901                                         stride, 8, 8);                  \
 902 }                                                                       \
 903                                                                         \
 904 static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src,    \
 905                                          int stride)                    \
 906 {                                                                       \
 907     uint64_t half[8 + 9];                                               \
 908     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
 909     uint8_t * const halfHV = ((uint8_t*)half);                          \
 910     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 911                                                    stride, 9);          \
 912     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
 913     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV,             \
 914                                         stride, 8, 8);                  \
 915 }                                                                       \
 916                                                                         \
 917 static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src,    \
 918                                          int stride)                    \
 919 {                                                                       \
 920     uint64_t half[8 + 9];                                               \
 921     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
 922     uint8_t * const halfHV = ((uint8_t*)half);                          \
 923     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 924                                                    stride, 9);          \
 925     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
 926     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV,         \
 927                                         stride, 8, 8);                  \
 928 }                                                                       \
 929                                                                         \
 930 static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src,    \
 931                                          int stride)                    \
 932 {                                                                       \
 933     uint64_t half[8 + 9];                                               \
 934     uint8_t * const halfH = ((uint8_t*)half);                           \
 935     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 936                                                    stride, 9);          \
 937     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH,              \
 938                                         8, stride, 9);                  \
 939     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH,          \
 940                                                    stride, 8);          \
 941 }                                                                       \
 942                                                                         \
 943 static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src,    \
 944                                          int stride)                    \
 945 {                                                                       \
 946     uint64_t half[8 + 9];                                               \
 947     uint8_t * const halfH = ((uint8_t*)half);                           \
 948     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 949                                                    stride, 9);          \
 950     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,       \
 951                                         stride, 9);                     \
 952     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH,          \
 953                                                    stride, 8);          \
 954 }                                                                       \
 955                                                                         \
 956 static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src,    \
 957                                          int stride)                    \
 958 {                                                                       \
 959     uint64_t half[9];                                                   \
 960     uint8_t * const halfH = ((uint8_t*)half);                           \
 961     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 962                                                    stride, 9);          \
 963     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH,          \
 964                                                    stride, 8);          \
 965 }                                                                       \
 966                                                                         \
 967 static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src,  \
 968                                            int stride)                  \
 969 {                                                                       \
 970     ff_ ## OPNAME ## pixels16_ ## MMX(dst, src, stride, 16);            \
 971 }                                                                       \
 972                                                                         \
 973 static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src,   \
 974                                           int stride)                   \
 975 {                                                                       \
 976     uint64_t temp[32];                                                  \
 977     uint8_t * const half = (uint8_t*)temp;                              \
 978     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16,      \
 979                                                     stride, 16);        \
 980     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride,        \
 981                                          stride, 16);                   \
 982 }                                                                       \
 983                                                                         \
 984 static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src,   \
 985                                           int stride)                   \
 986 {                                                                       \
 987     ff_ ## OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src,           \
 988                                                     stride, stride, 16);\
 989 }                                                                       \
 990                                                                         \
 991 static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src,   \
 992                                           int stride)                   \
 993 {                                                                       \
 994     uint64_t temp[32];                                                  \
 995     uint8_t * const half = (uint8_t*)temp;                              \
 996     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16,      \
 997                                                     stride, 16);        \
 998     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src + 1, half,            \
 999                                          stride, stride, 16);           \
1000 }                                                                       \
1001                                                                         \
1002 static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src,   \
1003                                           int stride)                   \
1004 {                                                                       \
1005     uint64_t temp[32];                                                  \
1006     uint8_t * const half = (uint8_t*)temp;                              \
1007     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16,      \
1008                                                     stride);            \
1009     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride,        \
1010                                          stride, 16);                   \
1011 }                                                                       \
1012                                                                         \
1013 static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src,   \
1014                                           int stride)                   \
1015 {                                                                       \
1016     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src,           \
1017                                                     stride, stride);    \
1018 }                                                                       \
1019                                                                         \
1020 static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src,   \
1021                                           int stride)                   \
1022 {                                                                       \
1023     uint64_t temp[32];                                                  \
1024     uint8_t * const half = (uint8_t*)temp;                              \
1025     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16,      \
1026                                                     stride);            \
1027     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half,         \
1028                                          stride, stride, 16);           \
1029 }                                                                       \
1030                                                                         \
1031 static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src,   \
1032                                           int stride)                   \
1033 {                                                                       \
1034     uint64_t half[16 * 2 + 17 * 2];                                     \
1035     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1036     uint8_t * const halfHV = ((uint8_t*)half);                          \
1037     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1038                                                     stride, 17);        \
1039     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,         \
1040                                          stride, 17);                   \
1041     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1042                                                     16, 16);            \
1043     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV,            \
1044                                          stride, 16, 16);               \
1045 }                                                                       \
1046                                                                         \
1047 static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src,   \
1048                                           int stride)                   \
1049 {                                                                       \
1050     uint64_t half[16 * 2 + 17 * 2];                                     \
1051     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1052     uint8_t * const halfHV = ((uint8_t*)half);                          \
1053     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1054                                                     stride, 17);        \
1055     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,     \
1056                                          stride, 17);                   \
1057     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1058                                                     16, 16);            \
1059     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV,            \
1060                                          stride, 16, 16);               \
1061 }                                                                       \
1062                                                                         \
1063 static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src,   \
1064                                           int stride)                   \
1065 {                                                                       \
1066     uint64_t half[16 * 2 + 17 * 2];                                     \
1067     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1068     uint8_t * const halfHV = ((uint8_t*)half);                          \
1069     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1070                                                     stride, 17);        \
1071     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,         \
1072                                          stride, 17);                   \
1073     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1074                                                     16, 16);            \
1075     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV,       \
1076                                          stride, 16, 16);               \
1077 }                                                                       \
1078                                                                         \
1079 static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src,   \
1080                                           int stride)                   \
1081 {                                                                       \
1082     uint64_t half[16 * 2 + 17 * 2];                                     \
1083     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1084     uint8_t * const halfHV = ((uint8_t*)half);                          \
1085     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1086                                                     stride, 17);        \
1087     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,     \
1088                                          stride, 17);                   \
1089     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1090                                                     16, 16);            \
1091     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV,       \
1092                                          stride, 16, 16);               \
1093 }                                                                       \
1094                                                                         \
1095 static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src,   \
1096                                           int stride)                   \
1097 {                                                                       \
1098     uint64_t half[16 * 2 + 17 * 2];                                     \
1099     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1100     uint8_t * const halfHV = ((uint8_t*)half);                          \
1101     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1102                                                     stride, 17);        \
1103     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1104                                                     16, 16);            \
1105     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV,            \
1106                                          stride, 16, 16);               \
1107 }                                                                       \
1108                                                                         \
1109 static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src,   \
1110                                           int stride)                   \
1111 {                                                                       \
1112     uint64_t half[16 * 2 + 17 * 2];                                     \
1113     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1114     uint8_t * const halfHV = ((uint8_t*)half);                          \
1115     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1116                                                     stride, 17);        \
1117     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1118                                                     16, 16);            \
1119     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV,       \
1120                                          stride, 16, 16);               \
1121 }                                                                       \
1122                                                                         \
1123 static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src,   \
1124                                           int stride)                   \
1125 {                                                                       \
1126     uint64_t half[17 * 2];                                              \
1127     uint8_t * const halfH = ((uint8_t*)half);                           \
1128     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1129                                                     stride, 17);        \
1130     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,         \
1131                                          stride, 17);                   \
1132     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH,         \
1133                                                     stride, 16);        \
1134 }                                                                       \
1135                                                                         \
1136 static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src,   \
1137                                           int stride)                   \
1138 {                                                                       \
1139     uint64_t half[17 * 2];                                              \
1140     uint8_t * const halfH = ((uint8_t*)half);                           \
1141     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1142                                                     stride, 17);        \
1143     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,     \
1144                                          stride, 17);                   \
1145     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH,         \
1146                                                     stride, 16);        \
1147 }                                                                       \
1148                                                                         \
1149 static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src,   \
1150                                           int stride)                   \
1151 {                                                                       \
1152     uint64_t half[17 * 2];                                              \
1153     uint8_t * const halfH = ((uint8_t*)half);                           \
1154     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1155                                                     stride, 17);        \
1156     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH,         \
1157                                                     stride, 16);        \
1158 }
1159
1160 #define PUT_OP(a, b, temp, size)                \
1161     "mov"#size"        "#a", "#b"       \n\t"
1162
1163 #define AVG_MMXEXT_OP(a, b, temp, size)         \
1164     "mov"#size"        "#b", "#temp"    \n\t"   \
1165     "pavgb          "#temp", "#a"       \n\t"   \
1166     "mov"#size"        "#a", "#b"       \n\t"
1167
1168 QPEL_OP(put_,          ff_pw_16, _,        PUT_OP,        mmxext)
1169 QPEL_OP(avg_,          ff_pw_16, _,        AVG_MMXEXT_OP, mmxext)
1170 QPEL_OP(put_no_rnd_,   ff_pw_15, _no_rnd_, PUT_OP,        mmxext)
1171 #endif /* HAVE_YASM */
1172
1173
1174 #if HAVE_INLINE_ASM
1175 void ff_put_rv40_qpel8_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1176 {
1177   put_pixels8_xy2_mmx(dst, src, stride, 8);
1178 }
1179 void ff_put_rv40_qpel16_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1180 {
1181   put_pixels16_xy2_mmx(dst, src, stride, 16);
1182 }
1183 void ff_avg_rv40_qpel8_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1184 {
1185   avg_pixels8_xy2_mmx(dst, src, stride, 8);
1186 }
1187 void ff_avg_rv40_qpel16_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1188 {
1189   avg_pixels16_xy2_mmx(dst, src, stride, 16);
1190 }
1191
1192 typedef void emulated_edge_mc_func(uint8_t *dst, const uint8_t *src,
1193                                    ptrdiff_t linesize, int block_w, int block_h,
1194                                    int src_x, int src_y, int w, int h);
1195
1196 static av_always_inline void gmc(uint8_t *dst, uint8_t *src,
1197                                  int stride, int h, int ox, int oy,
1198                                  int dxx, int dxy, int dyx, int dyy,
1199                                  int shift, int r, int width, int height,
1200                                  emulated_edge_mc_func *emu_edge_fn)
1201 {
1202     const int w    = 8;
1203     const int ix   = ox  >> (16 + shift);
1204     const int iy   = oy  >> (16 + shift);
1205     const int oxs  = ox  >> 4;
1206     const int oys  = oy  >> 4;
1207     const int dxxs = dxx >> 4;
1208     const int dxys = dxy >> 4;
1209     const int dyxs = dyx >> 4;
1210     const int dyys = dyy >> 4;
1211     const uint16_t r4[4]   = { r, r, r, r };
1212     const uint16_t dxy4[4] = { dxys, dxys, dxys, dxys };
1213     const uint16_t dyy4[4] = { dyys, dyys, dyys, dyys };
1214     const uint64_t shift2 = 2 * shift;
1215 #define MAX_STRIDE 4096U
1216 #define MAX_H 8U
1217     uint8_t edge_buf[(MAX_H + 1) * MAX_STRIDE];
1218     int x, y;
1219
1220     const int dxw = (dxx - (1 << (16 + shift))) * (w - 1);
1221     const int dyh = (dyy - (1 << (16 + shift))) * (h - 1);
1222     const int dxh = dxy * (h - 1);
1223     const int dyw = dyx * (w - 1);
1224     int need_emu =  (unsigned)ix >= width  - w ||
1225                     (unsigned)iy >= height - h;
1226
1227     if ( // non-constant fullpel offset (3% of blocks)
1228         ((ox ^ (ox + dxw)) | (ox ^ (ox + dxh)) | (ox ^ (ox + dxw + dxh)) |
1229          (oy ^ (oy + dyw)) | (oy ^ (oy + dyh)) | (oy ^ (oy + dyw + dyh))) >> (16 + shift)
1230         // uses more than 16 bits of subpel mv (only at huge resolution)
1231         || (dxx | dxy | dyx | dyy) & 15
1232         || (need_emu && (h > MAX_H || stride > MAX_STRIDE))) {
1233         // FIXME could still use mmx for some of the rows
1234         ff_gmc_c(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy,
1235                  shift, r, width, height);
1236         return;
1237     }
1238
1239     src += ix + iy * stride;
1240     if (need_emu) {
1241         emu_edge_fn(edge_buf, src, stride, w + 1, h + 1, ix, iy, width, height);
1242         src = edge_buf;
1243     }
1244
1245     __asm__ volatile (
1246         "movd         %0, %%mm6         \n\t"
1247         "pxor      %%mm7, %%mm7         \n\t"
1248         "punpcklwd %%mm6, %%mm6         \n\t"
1249         "punpcklwd %%mm6, %%mm6         \n\t"
1250         :: "r"(1<<shift)
1251     );
1252
1253     for (x = 0; x < w; x += 4) {
1254         uint16_t dx4[4] = { oxs - dxys + dxxs * (x + 0),
1255                             oxs - dxys + dxxs * (x + 1),
1256                             oxs - dxys + dxxs * (x + 2),
1257                             oxs - dxys + dxxs * (x + 3) };
1258         uint16_t dy4[4] = { oys - dyys + dyxs * (x + 0),
1259                             oys - dyys + dyxs * (x + 1),
1260                             oys - dyys + dyxs * (x + 2),
1261                             oys - dyys + dyxs * (x + 3) };
1262
1263         for (y = 0; y < h; y++) {
1264             __asm__ volatile (
1265                 "movq      %0, %%mm4    \n\t"
1266                 "movq      %1, %%mm5    \n\t"
1267                 "paddw     %2, %%mm4    \n\t"
1268                 "paddw     %3, %%mm5    \n\t"
1269                 "movq   %%mm4, %0       \n\t"
1270                 "movq   %%mm5, %1       \n\t"
1271                 "psrlw    $12, %%mm4    \n\t"
1272                 "psrlw    $12, %%mm5    \n\t"
1273                 : "+m"(*dx4), "+m"(*dy4)
1274                 : "m"(*dxy4), "m"(*dyy4)
1275             );
1276
1277             __asm__ volatile (
1278                 "movq      %%mm6, %%mm2 \n\t"
1279                 "movq      %%mm6, %%mm1 \n\t"
1280                 "psubw     %%mm4, %%mm2 \n\t"
1281                 "psubw     %%mm5, %%mm1 \n\t"
1282                 "movq      %%mm2, %%mm0 \n\t"
1283                 "movq      %%mm4, %%mm3 \n\t"
1284                 "pmullw    %%mm1, %%mm0 \n\t" // (s - dx) * (s - dy)
1285                 "pmullw    %%mm5, %%mm3 \n\t" // dx * dy
1286                 "pmullw    %%mm5, %%mm2 \n\t" // (s - dx) * dy
1287                 "pmullw    %%mm4, %%mm1 \n\t" // dx * (s - dy)
1288
1289                 "movd         %4, %%mm5 \n\t"
1290                 "movd         %3, %%mm4 \n\t"
1291                 "punpcklbw %%mm7, %%mm5 \n\t"
1292                 "punpcklbw %%mm7, %%mm4 \n\t"
1293                 "pmullw    %%mm5, %%mm3 \n\t" // src[1, 1] * dx * dy
1294                 "pmullw    %%mm4, %%mm2 \n\t" // src[0, 1] * (s - dx) * dy
1295
1296                 "movd         %2, %%mm5 \n\t"
1297                 "movd         %1, %%mm4 \n\t"
1298                 "punpcklbw %%mm7, %%mm5 \n\t"
1299                 "punpcklbw %%mm7, %%mm4 \n\t"
1300                 "pmullw    %%mm5, %%mm1 \n\t" // src[1, 0] * dx * (s - dy)
1301                 "pmullw    %%mm4, %%mm0 \n\t" // src[0, 0] * (s - dx) * (s - dy)
1302                 "paddw        %5, %%mm1 \n\t"
1303                 "paddw     %%mm3, %%mm2 \n\t"
1304                 "paddw     %%mm1, %%mm0 \n\t"
1305                 "paddw     %%mm2, %%mm0 \n\t"
1306
1307                 "psrlw        %6, %%mm0 \n\t"
1308                 "packuswb  %%mm0, %%mm0 \n\t"
1309                 "movd      %%mm0, %0    \n\t"
1310
1311                 : "=m"(dst[x + y * stride])
1312                 : "m"(src[0]), "m"(src[1]),
1313                   "m"(src[stride]), "m"(src[stride + 1]),
1314                   "m"(*r4), "m"(shift2)
1315             );
1316             src += stride;
1317         }
1318         src += 4 - h * stride;
1319     }
1320 }
1321
1322 #if CONFIG_VIDEODSP
1323 #if HAVE_YASM
1324 #if ARCH_X86_32
1325 static void gmc_mmx(uint8_t *dst, uint8_t *src,
1326                     int stride, int h, int ox, int oy,
1327                     int dxx, int dxy, int dyx, int dyy,
1328                     int shift, int r, int width, int height)
1329 {
1330     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,
1331         width, height, &ff_emulated_edge_mc_8);
1332 }
1333 #endif
1334 static void gmc_sse(uint8_t *dst, uint8_t *src,
1335                     int stride, int h, int ox, int oy,
1336                     int dxx, int dxy, int dyx, int dyy,
1337                     int shift, int r, int width, int height)
1338 {
1339     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,
1340         width, height, &ff_emulated_edge_mc_8);
1341 }
1342 #else
1343 static void gmc_mmx(uint8_t *dst, uint8_t *src,
1344                     int stride, int h, int ox, int oy,
1345                     int dxx, int dxy, int dyx, int dyy,
1346                     int shift, int r, int width, int height)
1347 {
1348     gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,
1349         width, height, &ff_emulated_edge_mc_8);
1350 }
1351 #endif
1352 #endif
1353
1354 #endif /* HAVE_INLINE_ASM */
1355
1356 void ff_put_pixels16_sse2(uint8_t *block, const uint8_t *pixels,
1357                           ptrdiff_t line_size, int h);
1358 void ff_avg_pixels16_sse2(uint8_t *block, const uint8_t *pixels,
1359                           ptrdiff_t line_size, int h);
1360
1361 #if HAVE_INLINE_ASM
1362
1363 /* CAVS-specific */
1364 void ff_put_cavs_qpel8_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1365 {
1366     put_pixels8_mmx(dst, src, stride, 8);
1367 }
1368
1369 void ff_avg_cavs_qpel8_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1370 {
1371     avg_pixels8_mmx(dst, src, stride, 8);
1372 }
1373
1374 void ff_put_cavs_qpel16_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1375 {
1376     put_pixels16_mmx(dst, src, stride, 16);
1377 }
1378
1379 void ff_avg_cavs_qpel16_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1380 {
1381     avg_pixels16_mmx(dst, src, stride, 16);
1382 }
1383 #endif /* HAVE_INLINE_ASM */
1384
1385 #if HAVE_YASM
1386 /* VC-1-specific */
1387 void ff_put_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src,
1388                                int stride, int rnd)
1389 {
1390     ff_put_pixels8_mmx(dst, src, stride, 8);
1391 }
1392
1393 void ff_avg_vc1_mspel_mc00_mmxext(uint8_t *dst, const uint8_t *src,
1394                                   int stride, int rnd)
1395 {
1396     ff_avg_pixels8_mmxext(dst, src, stride, 8);
1397 }
1398 #endif /* HAVE_YASM */
1399
1400 #if CONFIG_DIRAC_DECODER
1401 #define DIRAC_PIXOP(OPNAME2, OPNAME, EXT)\
1402 void ff_ ## OPNAME2 ## _dirac_pixels8_ ## EXT(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
1403 {\
1404     if (h&3)\
1405         ff_ ## OPNAME2 ## _dirac_pixels8_c(dst, src, stride, h);\
1406     else\
1407         OPNAME ## _pixels8_ ## EXT(dst, src[0], stride, h);\
1408 }\
1409 void ff_ ## OPNAME2 ## _dirac_pixels16_ ## EXT(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
1410 {\
1411     if (h&3)\
1412         ff_ ## OPNAME2 ## _dirac_pixels16_c(dst, src, stride, h);\
1413     else\
1414         OPNAME ## _pixels16_ ## EXT(dst, src[0], stride, h);\
1415 }\
1416 void ff_ ## OPNAME2 ## _dirac_pixels32_ ## EXT(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
1417 {\
1418     if (h&3) {\
1419         ff_ ## OPNAME2 ## _dirac_pixels32_c(dst, src, stride, h);\
1420     } else {\
1421         OPNAME ## _pixels16_ ## EXT(dst   , src[0]   , stride, h);\
1422         OPNAME ## _pixels16_ ## EXT(dst+16, src[0]+16, stride, h);\
1423     }\
1424 }
1425
1426 #if HAVE_MMX_INLINE
1427 DIRAC_PIXOP(put, put, mmx)
1428 DIRAC_PIXOP(avg, avg, mmx)
1429 #endif
1430
1431 #if HAVE_YASM
1432 DIRAC_PIXOP(avg, ff_avg, mmxext)
1433
1434 void ff_put_dirac_pixels16_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h)
1435 {
1436     if (h&3)
1437         ff_put_dirac_pixels16_c(dst, src, stride, h);
1438     else
1439     ff_put_pixels16_sse2(dst, src[0], stride, h);
1440 }
1441 void ff_avg_dirac_pixels16_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h)
1442 {
1443     if (h&3)
1444         ff_avg_dirac_pixels16_c(dst, src, stride, h);
1445     else
1446     ff_avg_pixels16_sse2(dst, src[0], stride, h);
1447 }
1448 void ff_put_dirac_pixels32_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h)
1449 {
1450     if (h&3) {
1451         ff_put_dirac_pixels32_c(dst, src, stride, h);
1452     } else {
1453     ff_put_pixels16_sse2(dst   , src[0]   , stride, h);
1454     ff_put_pixels16_sse2(dst+16, src[0]+16, stride, h);
1455     }
1456 }
1457 void ff_avg_dirac_pixels32_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h)
1458 {
1459     if (h&3) {
1460         ff_avg_dirac_pixels32_c(dst, src, stride, h);
1461     } else {
1462     ff_avg_pixels16_sse2(dst   , src[0]   , stride, h);
1463     ff_avg_pixels16_sse2(dst+16, src[0]+16, stride, h);
1464     }
1465 }
1466 #endif
1467 #endif
1468
1469 /* XXX: Those functions should be suppressed ASAP when all IDCTs are
1470  * converted. */
1471 #if CONFIG_GPL
1472 static void ff_libmpeg2mmx_idct_put(uint8_t *dest, int line_size,
1473                                     int16_t *block)
1474 {
1475     ff_mmx_idct(block);
1476     ff_put_pixels_clamped_mmx(block, dest, line_size);
1477 }
1478
1479 static void ff_libmpeg2mmx_idct_add(uint8_t *dest, int line_size,
1480                                     int16_t *block)
1481 {
1482     ff_mmx_idct(block);
1483     ff_add_pixels_clamped_mmx(block, dest, line_size);
1484 }
1485
1486 static void ff_libmpeg2mmx2_idct_put(uint8_t *dest, int line_size,
1487                                      int16_t *block)
1488 {
1489     ff_mmxext_idct(block);
1490     ff_put_pixels_clamped_mmx(block, dest, line_size);
1491 }
1492
1493 static void ff_libmpeg2mmx2_idct_add(uint8_t *dest, int line_size,
1494                                      int16_t *block)
1495 {
1496     ff_mmxext_idct(block);
1497     ff_add_pixels_clamped_mmx(block, dest, line_size);
1498 }
1499 #endif
1500
1501 #if HAVE_INLINE_ASM
1502 static void vector_clipf_sse(float *dst, const float *src,
1503                              float min, float max, int len)
1504 {
1505     x86_reg i = (len - 16) * 4;
1506     __asm__ volatile (
1507         "movss          %3, %%xmm4      \n\t"
1508         "movss          %4, %%xmm5      \n\t"
1509         "shufps $0, %%xmm4, %%xmm4      \n\t"
1510         "shufps $0, %%xmm5, %%xmm5      \n\t"
1511         "1:                             \n\t"
1512         "movaps   (%2, %0), %%xmm0      \n\t" // 3/1 on intel
1513         "movaps 16(%2, %0), %%xmm1      \n\t"
1514         "movaps 32(%2, %0), %%xmm2      \n\t"
1515         "movaps 48(%2, %0), %%xmm3      \n\t"
1516         "maxps      %%xmm4, %%xmm0      \n\t"
1517         "maxps      %%xmm4, %%xmm1      \n\t"
1518         "maxps      %%xmm4, %%xmm2      \n\t"
1519         "maxps      %%xmm4, %%xmm3      \n\t"
1520         "minps      %%xmm5, %%xmm0      \n\t"
1521         "minps      %%xmm5, %%xmm1      \n\t"
1522         "minps      %%xmm5, %%xmm2      \n\t"
1523         "minps      %%xmm5, %%xmm3      \n\t"
1524         "movaps     %%xmm0,   (%1, %0)  \n\t"
1525         "movaps     %%xmm1, 16(%1, %0)  \n\t"
1526         "movaps     %%xmm2, 32(%1, %0)  \n\t"
1527         "movaps     %%xmm3, 48(%1, %0)  \n\t"
1528         "sub           $64, %0          \n\t"
1529         "jge            1b              \n\t"
1530         : "+&r"(i)
1531         : "r"(dst), "r"(src), "m"(min), "m"(max)
1532         : "memory"
1533     );
1534 }
1535
1536 #endif /* HAVE_INLINE_ASM */
1537
1538 int32_t ff_scalarproduct_int16_mmxext(const int16_t *v1, const int16_t *v2,
1539                                       int order);
1540 int32_t ff_scalarproduct_int16_sse2(const int16_t *v1, const int16_t *v2,
1541                                     int order);
1542 int32_t ff_scalarproduct_and_madd_int16_mmxext(int16_t *v1, const int16_t *v2,
1543                                                const int16_t *v3,
1544                                                int order, int mul);
1545 int32_t ff_scalarproduct_and_madd_int16_sse2(int16_t *v1, const int16_t *v2,
1546                                              const int16_t *v3,
1547                                              int order, int mul);
1548 int32_t ff_scalarproduct_and_madd_int16_ssse3(int16_t *v1, const int16_t *v2,
1549                                               const int16_t *v3,
1550                                               int order, int mul);
1551
1552 void ff_apply_window_int16_round_mmxext(int16_t *output, const int16_t *input,
1553                                         const int16_t *window, unsigned int len);
1554 void ff_apply_window_int16_round_sse2(int16_t *output, const int16_t *input,
1555                                       const int16_t *window, unsigned int len);
1556 void ff_apply_window_int16_mmxext(int16_t *output, const int16_t *input,
1557                                   const int16_t *window, unsigned int len);
1558 void ff_apply_window_int16_sse2(int16_t *output, const int16_t *input,
1559                                 const int16_t *window, unsigned int len);
1560 void ff_apply_window_int16_ssse3(int16_t *output, const int16_t *input,
1561                                  const int16_t *window, unsigned int len);
1562 void ff_apply_window_int16_ssse3_atom(int16_t *output, const int16_t *input,
1563                                       const int16_t *window, unsigned int len);
1564
1565 void ff_bswap32_buf_ssse3(uint32_t *dst, const uint32_t *src, int w);
1566 void ff_bswap32_buf_sse2(uint32_t *dst, const uint32_t *src, int w);
1567
1568 void ff_add_hfyu_median_prediction_mmxext(uint8_t *dst, const uint8_t *top,
1569                                           const uint8_t *diff, int w,
1570                                           int *left, int *left_top);
1571 int  ff_add_hfyu_left_prediction_ssse3(uint8_t *dst, const uint8_t *src,
1572                                        int w, int left);
1573 int  ff_add_hfyu_left_prediction_sse4(uint8_t *dst, const uint8_t *src,
1574                                       int w, int left);
1575
1576 void ff_vector_clip_int32_mmx     (int32_t *dst, const int32_t *src,
1577                                    int32_t min, int32_t max, unsigned int len);
1578 void ff_vector_clip_int32_sse2    (int32_t *dst, const int32_t *src,
1579                                    int32_t min, int32_t max, unsigned int len);
1580 void ff_vector_clip_int32_int_sse2(int32_t *dst, const int32_t *src,
1581                                    int32_t min, int32_t max, unsigned int len);
1582 void ff_vector_clip_int32_sse4    (int32_t *dst, const int32_t *src,
1583                                    int32_t min, int32_t max, unsigned int len);
1584
1585 #define SET_QPEL_FUNCS(PFX, IDX, SIZE, CPU, PREFIX)                          \
1586     do {                                                                     \
1587     c->PFX ## _pixels_tab[IDX][ 0] = PREFIX ## PFX ## SIZE ## _mc00_ ## CPU; \
1588     c->PFX ## _pixels_tab[IDX][ 1] = PREFIX ## PFX ## SIZE ## _mc10_ ## CPU; \
1589     c->PFX ## _pixels_tab[IDX][ 2] = PREFIX ## PFX ## SIZE ## _mc20_ ## CPU; \
1590     c->PFX ## _pixels_tab[IDX][ 3] = PREFIX ## PFX ## SIZE ## _mc30_ ## CPU; \
1591     c->PFX ## _pixels_tab[IDX][ 4] = PREFIX ## PFX ## SIZE ## _mc01_ ## CPU; \
1592     c->PFX ## _pixels_tab[IDX][ 5] = PREFIX ## PFX ## SIZE ## _mc11_ ## CPU; \
1593     c->PFX ## _pixels_tab[IDX][ 6] = PREFIX ## PFX ## SIZE ## _mc21_ ## CPU; \
1594     c->PFX ## _pixels_tab[IDX][ 7] = PREFIX ## PFX ## SIZE ## _mc31_ ## CPU; \
1595     c->PFX ## _pixels_tab[IDX][ 8] = PREFIX ## PFX ## SIZE ## _mc02_ ## CPU; \
1596     c->PFX ## _pixels_tab[IDX][ 9] = PREFIX ## PFX ## SIZE ## _mc12_ ## CPU; \
1597     c->PFX ## _pixels_tab[IDX][10] = PREFIX ## PFX ## SIZE ## _mc22_ ## CPU; \
1598     c->PFX ## _pixels_tab[IDX][11] = PREFIX ## PFX ## SIZE ## _mc32_ ## CPU; \
1599     c->PFX ## _pixels_tab[IDX][12] = PREFIX ## PFX ## SIZE ## _mc03_ ## CPU; \
1600     c->PFX ## _pixels_tab[IDX][13] = PREFIX ## PFX ## SIZE ## _mc13_ ## CPU; \
1601     c->PFX ## _pixels_tab[IDX][14] = PREFIX ## PFX ## SIZE ## _mc23_ ## CPU; \
1602     c->PFX ## _pixels_tab[IDX][15] = PREFIX ## PFX ## SIZE ## _mc33_ ## CPU; \
1603     } while (0)
1604
1605 #define SET_HPEL_FUNCS(PFX, IDX, SIZE, CPU)                                     \
1606     do {                                                                        \
1607         c->PFX ## _pixels_tab IDX [0] = PFX ## _pixels ## SIZE ## _     ## CPU; \
1608         c->PFX ## _pixels_tab IDX [1] = PFX ## _pixels ## SIZE ## _x2_  ## CPU; \
1609         c->PFX ## _pixels_tab IDX [2] = PFX ## _pixels ## SIZE ## _y2_  ## CPU; \
1610         c->PFX ## _pixels_tab IDX [3] = PFX ## _pixels ## SIZE ## _xy2_ ## CPU; \
1611     } while (0)
1612
1613 static av_cold void dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx,
1614                                      int mm_flags)
1615 {
1616     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
1617
1618 #if HAVE_INLINE_ASM
1619     c->put_pixels_clamped        = ff_put_pixels_clamped_mmx;
1620     c->put_signed_pixels_clamped = ff_put_signed_pixels_clamped_mmx;
1621     c->add_pixels_clamped        = ff_add_pixels_clamped_mmx;
1622
1623     if (!high_bit_depth) {
1624         c->clear_block  = clear_block_mmx;
1625         c->clear_blocks = clear_blocks_mmx;
1626         c->draw_edges   = draw_edges_mmx;
1627
1628         SET_HPEL_FUNCS(put,        [0], 16, mmx);
1629         SET_HPEL_FUNCS(put_no_rnd, [0], 16, mmx);
1630         SET_HPEL_FUNCS(avg,        [0], 16, mmx);
1631         SET_HPEL_FUNCS(avg_no_rnd,    , 16, mmx);
1632         SET_HPEL_FUNCS(put,        [1],  8, mmx);
1633         SET_HPEL_FUNCS(put_no_rnd, [1],  8, mmx);
1634         SET_HPEL_FUNCS(avg,        [1],  8, mmx);
1635     }
1636
1637 #if CONFIG_VIDEODSP && (ARCH_X86_32 || !HAVE_YASM)
1638     c->gmc = gmc_mmx;
1639 #endif
1640
1641     c->add_bytes = add_bytes_mmx;
1642 #endif /* HAVE_INLINE_ASM */
1643
1644 #if HAVE_YASM
1645     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
1646         c->h263_v_loop_filter = ff_h263_v_loop_filter_mmx;
1647         c->h263_h_loop_filter = ff_h263_h_loop_filter_mmx;
1648     }
1649
1650     c->vector_clip_int32 = ff_vector_clip_int32_mmx;
1651 #endif
1652
1653 }
1654
1655 static av_cold void dsputil_init_mmxext(DSPContext *c, AVCodecContext *avctx,
1656                                         int mm_flags)
1657 {
1658     const int bit_depth      = avctx->bits_per_raw_sample;
1659     const int high_bit_depth = bit_depth > 8;
1660
1661 #if HAVE_YASM
1662     SET_QPEL_FUNCS(avg_qpel,        0, 16, mmxext, );
1663     SET_QPEL_FUNCS(avg_qpel,        1,  8, mmxext, );
1664
1665     SET_QPEL_FUNCS(put_qpel,        0, 16, mmxext, );
1666     SET_QPEL_FUNCS(put_qpel,        1,  8, mmxext, );
1667     SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, mmxext, );
1668     SET_QPEL_FUNCS(put_no_rnd_qpel, 1,  8, mmxext, );
1669
1670     if (!high_bit_depth) {
1671         c->put_pixels_tab[0][1] = ff_put_pixels16_x2_mmxext;
1672         c->put_pixels_tab[0][2] = ff_put_pixels16_y2_mmxext;
1673
1674         c->avg_pixels_tab[0][0] = ff_avg_pixels16_mmxext;
1675         c->avg_pixels_tab[0][1] = ff_avg_pixels16_x2_mmxext;
1676         c->avg_pixels_tab[0][2] = ff_avg_pixels16_y2_mmxext;
1677
1678         c->put_pixels_tab[1][1] = ff_put_pixels8_x2_mmxext;
1679         c->put_pixels_tab[1][2] = ff_put_pixels8_y2_mmxext;
1680
1681         c->avg_pixels_tab[1][0] = ff_avg_pixels8_mmxext;
1682         c->avg_pixels_tab[1][1] = ff_avg_pixels8_x2_mmxext;
1683         c->avg_pixels_tab[1][2] = ff_avg_pixels8_y2_mmxext;
1684     }
1685
1686     if (!(avctx->flags & CODEC_FLAG_BITEXACT)) {
1687         if (!high_bit_depth) {
1688             c->put_no_rnd_pixels_tab[0][1] = ff_put_no_rnd_pixels16_x2_mmxext;
1689             c->put_no_rnd_pixels_tab[0][2] = ff_put_no_rnd_pixels16_y2_mmxext;
1690             c->put_no_rnd_pixels_tab[1][1] = ff_put_no_rnd_pixels8_x2_mmxext;
1691             c->put_no_rnd_pixels_tab[1][2] = ff_put_no_rnd_pixels8_y2_mmxext;
1692
1693             c->avg_pixels_tab[0][3] = ff_avg_pixels16_xy2_mmxext;
1694             c->avg_pixels_tab[1][3] = ff_avg_pixels8_xy2_mmxext;
1695         }
1696     }
1697 #endif /* HAVE_YASM */
1698
1699 #if HAVE_MMXEXT_EXTERNAL
1700     if (CONFIG_VP3_DECODER && (avctx->codec_id == AV_CODEC_ID_VP3 ||
1701                                avctx->codec_id == AV_CODEC_ID_THEORA)) {
1702         c->put_no_rnd_pixels_tab[1][1] = ff_put_no_rnd_pixels8_x2_exact_mmxext;
1703         c->put_no_rnd_pixels_tab[1][2] = ff_put_no_rnd_pixels8_y2_exact_mmxext;
1704     }
1705
1706     /* slower than cmov version on AMD */
1707     if (!(mm_flags & AV_CPU_FLAG_3DNOW))
1708         c->add_hfyu_median_prediction = ff_add_hfyu_median_prediction_mmxext;
1709
1710     c->scalarproduct_int16          = ff_scalarproduct_int16_mmxext;
1711     c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_mmxext;
1712
1713     if (avctx->flags & CODEC_FLAG_BITEXACT) {
1714         c->apply_window_int16 = ff_apply_window_int16_mmxext;
1715     } else {
1716         c->apply_window_int16 = ff_apply_window_int16_round_mmxext;
1717     }
1718 #endif /* HAVE_MMXEXT_EXTERNAL */
1719 }
1720
1721 static av_cold void dsputil_init_3dnow(DSPContext *c, AVCodecContext *avctx,
1722                                        int mm_flags)
1723 {
1724     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
1725
1726 #if HAVE_YASM
1727     if (!high_bit_depth) {
1728         c->put_pixels_tab[0][1] = ff_put_pixels16_x2_3dnow;
1729         c->put_pixels_tab[0][2] = ff_put_pixels16_y2_3dnow;
1730
1731         c->avg_pixels_tab[0][0] = ff_avg_pixels16_3dnow;
1732         c->avg_pixels_tab[0][1] = ff_avg_pixels16_x2_3dnow;
1733         c->avg_pixels_tab[0][2] = ff_avg_pixels16_y2_3dnow;
1734
1735         c->put_pixels_tab[1][1] = ff_put_pixels8_x2_3dnow;
1736         c->put_pixels_tab[1][2] = ff_put_pixels8_y2_3dnow;
1737
1738         c->avg_pixels_tab[1][0] = ff_avg_pixels8_3dnow;
1739         c->avg_pixels_tab[1][1] = ff_avg_pixels8_x2_3dnow;
1740         c->avg_pixels_tab[1][2] = ff_avg_pixels8_y2_3dnow;
1741
1742         if (!(avctx->flags & CODEC_FLAG_BITEXACT)){
1743             c->put_no_rnd_pixels_tab[0][1] = ff_put_no_rnd_pixels16_x2_3dnow;
1744             c->put_no_rnd_pixels_tab[0][2] = ff_put_no_rnd_pixels16_y2_3dnow;
1745             c->put_no_rnd_pixels_tab[1][1] = ff_put_no_rnd_pixels8_x2_3dnow;
1746             c->put_no_rnd_pixels_tab[1][2] = ff_put_no_rnd_pixels8_y2_3dnow;
1747
1748             c->avg_pixels_tab[0][3] = ff_avg_pixels16_xy2_3dnow;
1749             c->avg_pixels_tab[1][3] = ff_avg_pixels8_xy2_3dnow;
1750         }
1751     }
1752
1753     if (CONFIG_VP3_DECODER && (avctx->codec_id == AV_CODEC_ID_VP3 ||
1754                                avctx->codec_id == AV_CODEC_ID_THEORA)) {
1755         c->put_no_rnd_pixels_tab[1][1] = ff_put_no_rnd_pixels8_x2_exact_3dnow;
1756         c->put_no_rnd_pixels_tab[1][2] = ff_put_no_rnd_pixels8_y2_exact_3dnow;
1757     }
1758 #endif /* HAVE_YASM */
1759 }
1760
1761 static av_cold void dsputil_init_sse(DSPContext *c, AVCodecContext *avctx,
1762                                      int mm_flags)
1763 {
1764     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
1765
1766 #if HAVE_INLINE_ASM
1767     if (!high_bit_depth) {
1768         if (!(CONFIG_MPEG_XVMC_DECODER && avctx->xvmc_acceleration > 1)) {
1769             /* XvMCCreateBlocks() may not allocate 16-byte aligned blocks */
1770             c->clear_block  = clear_block_sse;
1771             c->clear_blocks = clear_blocks_sse;
1772         }
1773     }
1774
1775     c->vector_clipf = vector_clipf_sse;
1776 #endif /* HAVE_INLINE_ASM */
1777
1778 #if HAVE_YASM
1779 #if HAVE_INLINE_ASM && CONFIG_VIDEODSP
1780     c->gmc = gmc_sse;
1781 #endif
1782 #endif /* HAVE_YASM */
1783 }
1784
1785 static av_cold void dsputil_init_sse2(DSPContext *c, AVCodecContext *avctx,
1786                                       int mm_flags)
1787 {
1788     const int bit_depth      = avctx->bits_per_raw_sample;
1789     const int high_bit_depth = bit_depth > 8;
1790
1791 #if HAVE_SSE2_INLINE
1792     if (!high_bit_depth && avctx->idct_algo == FF_IDCT_XVIDMMX) {
1793         c->idct_put              = ff_idct_xvid_sse2_put;
1794         c->idct_add              = ff_idct_xvid_sse2_add;
1795         c->idct                  = ff_idct_xvid_sse2;
1796         c->idct_permutation_type = FF_SSE2_IDCT_PERM;
1797     }
1798 #endif /* HAVE_SSE2_INLINE */
1799
1800 #if HAVE_SSE2_EXTERNAL
1801     if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {
1802         // these functions are slower than mmx on AMD, but faster on Intel
1803         if (!high_bit_depth) {
1804             c->put_pixels_tab[0][0]        = ff_put_pixels16_sse2;
1805             c->put_no_rnd_pixels_tab[0][0] = ff_put_pixels16_sse2;
1806             c->avg_pixels_tab[0][0]        = ff_avg_pixels16_sse2;
1807         }
1808     }
1809
1810     c->scalarproduct_int16          = ff_scalarproduct_int16_sse2;
1811     c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_sse2;
1812     if (mm_flags & AV_CPU_FLAG_ATOM) {
1813         c->vector_clip_int32 = ff_vector_clip_int32_int_sse2;
1814     } else {
1815         c->vector_clip_int32 = ff_vector_clip_int32_sse2;
1816     }
1817     if (avctx->flags & CODEC_FLAG_BITEXACT) {
1818         c->apply_window_int16 = ff_apply_window_int16_sse2;
1819     } else if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {
1820         c->apply_window_int16 = ff_apply_window_int16_round_sse2;
1821     }
1822     c->bswap_buf = ff_bswap32_buf_sse2;
1823 #endif /* HAVE_SSE2_EXTERNAL */
1824 }
1825
1826 static av_cold void dsputil_init_ssse3(DSPContext *c, AVCodecContext *avctx,
1827                                        int mm_flags)
1828 {
1829 #if HAVE_SSSE3_EXTERNAL
1830     c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_ssse3;
1831     if (mm_flags & AV_CPU_FLAG_SSE4) // not really sse4, just slow on Conroe
1832         c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_sse4;
1833
1834     if (mm_flags & AV_CPU_FLAG_ATOM)
1835         c->apply_window_int16 = ff_apply_window_int16_ssse3_atom;
1836     else
1837         c->apply_window_int16 = ff_apply_window_int16_ssse3;
1838     if (!(mm_flags & (AV_CPU_FLAG_SSE42|AV_CPU_FLAG_3DNOW))) // cachesplit
1839         c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_ssse3;
1840     c->bswap_buf = ff_bswap32_buf_ssse3;
1841 #endif /* HAVE_SSSE3_EXTERNAL */
1842 }
1843
1844 static av_cold void dsputil_init_sse4(DSPContext *c, AVCodecContext *avctx,
1845                                       int mm_flags)
1846 {
1847 #if HAVE_SSE4_EXTERNAL
1848     c->vector_clip_int32 = ff_vector_clip_int32_sse4;
1849 #endif /* HAVE_SSE4_EXTERNAL */
1850 }
1851
1852 av_cold void ff_dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx)
1853 {
1854     int mm_flags = av_get_cpu_flags();
1855
1856 #if HAVE_7REGS && HAVE_INLINE_ASM
1857     if (mm_flags & AV_CPU_FLAG_CMOV)
1858         c->add_hfyu_median_prediction = add_hfyu_median_prediction_cmov;
1859 #endif
1860
1861     if (mm_flags & AV_CPU_FLAG_MMX) {
1862 #if HAVE_INLINE_ASM
1863         const int idct_algo = avctx->idct_algo;
1864
1865         if (avctx->lowres == 0 && avctx->bits_per_raw_sample <= 8) {
1866             if (idct_algo == FF_IDCT_AUTO || idct_algo == FF_IDCT_SIMPLEMMX) {
1867                 c->idct_put              = ff_simple_idct_put_mmx;
1868                 c->idct_add              = ff_simple_idct_add_mmx;
1869                 c->idct                  = ff_simple_idct_mmx;
1870                 c->idct_permutation_type = FF_SIMPLE_IDCT_PERM;
1871 #if CONFIG_GPL
1872             } else if (idct_algo == FF_IDCT_LIBMPEG2MMX) {
1873                 if (mm_flags & AV_CPU_FLAG_MMX2) {
1874                     c->idct_put = ff_libmpeg2mmx2_idct_put;
1875                     c->idct_add = ff_libmpeg2mmx2_idct_add;
1876                     c->idct     = ff_mmxext_idct;
1877                 } else {
1878                     c->idct_put = ff_libmpeg2mmx_idct_put;
1879                     c->idct_add = ff_libmpeg2mmx_idct_add;
1880                     c->idct     = ff_mmx_idct;
1881                 }
1882                 c->idct_permutation_type = FF_LIBMPEG2_IDCT_PERM;
1883 #endif
1884             } else if (idct_algo == FF_IDCT_XVIDMMX) {
1885                 if (mm_flags & AV_CPU_FLAG_SSE2) {
1886                     c->idct_put              = ff_idct_xvid_sse2_put;
1887                     c->idct_add              = ff_idct_xvid_sse2_add;
1888                     c->idct                  = ff_idct_xvid_sse2;
1889                     c->idct_permutation_type = FF_SSE2_IDCT_PERM;
1890                 } else if (mm_flags & AV_CPU_FLAG_MMXEXT) {
1891                     c->idct_put              = ff_idct_xvid_mmxext_put;
1892                     c->idct_add              = ff_idct_xvid_mmxext_add;
1893                     c->idct                  = ff_idct_xvid_mmxext;
1894                 } else {
1895                     c->idct_put              = ff_idct_xvid_mmx_put;
1896                     c->idct_add              = ff_idct_xvid_mmx_add;
1897                     c->idct                  = ff_idct_xvid_mmx;
1898                 }
1899             }
1900         }
1901 #endif /* HAVE_INLINE_ASM */
1902
1903         dsputil_init_mmx(c, avctx, mm_flags);
1904     }
1905
1906     if (mm_flags & AV_CPU_FLAG_MMXEXT)
1907         dsputil_init_mmxext(c, avctx, mm_flags);
1908
1909     if (mm_flags & AV_CPU_FLAG_3DNOW)
1910         dsputil_init_3dnow(c, avctx, mm_flags);
1911
1912     if (mm_flags & AV_CPU_FLAG_SSE)
1913         dsputil_init_sse(c, avctx, mm_flags);
1914
1915     if (mm_flags & AV_CPU_FLAG_SSE2)
1916         dsputil_init_sse2(c, avctx, mm_flags);
1917
1918     if (mm_flags & AV_CPU_FLAG_SSSE3)
1919         dsputil_init_ssse3(c, avctx, mm_flags);
1920
1921     if (mm_flags & AV_CPU_FLAG_SSE4)
1922         dsputil_init_sse4(c, avctx, mm_flags);
1923
1924     if (CONFIG_ENCODERS)
1925         ff_dsputilenc_init_mmx(c, avctx);
1926 }