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

   1 /*
   2  * MMX optimized DSP utils
   3  * Copyright (c) 2000, 2001 Fabrice Bellard
   4  * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
   5  *
   6  * MMX optimization by Nick Kurshev <nickols_k@mail.ru>
   7  *
   8  * This file is part of Libav.
   9  *
  10  * Libav is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2.1 of the License, or (at your option) any later version.
  14  *
  15  * Libav is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * Lesser General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU Lesser General Public
  21  * License along with Libav; if not, write to the Free Software
  22  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  23  */
  24
  25 #include "libavutil/attributes.h"
  26 #include "libavutil/cpu.h"
  27 #include "libavutil/x86/asm.h"
  28 #include "libavcodec/dsputil.h"
  29 #include "libavcodec/h264dsp.h"
  30 #include "libavcodec/mpegvideo.h"
  31 #include "libavcodec/simple_idct.h"
  32 #include "dsputil_mmx.h"
  33 #include "idct_xvid.h"
  34
  35 //#undef NDEBUG
  36 //#include <assert.h>
  37
  38 /* pixel operations */
  39 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_15)   =   0x000F000F000F000FULL;
  40 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_17)   = { 0x0011001100110011ULL, 0x0011001100110011ULL };
  41 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_20)   =   0x0014001400140014ULL;
  42 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_42)   =   0x002A002A002A002AULL;
  43 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_53)   =   0x0035003500350035ULL;
  44 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_96)   =   0x0060006000600060ULL;
  45 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_128)  =   0x0080008000800080ULL;
  46 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_255)  =   0x00ff00ff00ff00ffULL;
  47 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_512)  = { 0x0200020002000200ULL, 0x0200020002000200ULL };
  48 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_1019) = { 0x03FB03FB03FB03FBULL, 0x03FB03FB03FB03FBULL };
  49
  50 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_7)    =   0x0707070707070707ULL;
  51 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_1F)   =   0x1F1F1F1F1F1F1F1FULL;
  52 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_3F)   =   0x3F3F3F3F3F3F3F3FULL;
  53 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_81)   =   0x8181818181818181ULL;
  54 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_FC)   =   0xFCFCFCFCFCFCFCFCULL;
  55
  56 DECLARE_ALIGNED(16, const double, ff_pd_1)[2] = { 1.0, 1.0 };
  57 DECLARE_ALIGNED(16, const double, ff_pd_2)[2] = { 2.0, 2.0 };
  58
  59
  60 #if HAVE_YASM
  61 void ff_put_pixels8_x2_mmxext(uint8_t *block, const uint8_t *pixels,
  62                               ptrdiff_t line_size, int h);
  63 void ff_put_pixels8_x2_3dnow(uint8_t *block, const uint8_t *pixels,
  64                              ptrdiff_t line_size, int h);
  65 void ff_put_pixels8_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
  66                               int dstStride, int src1Stride, int h);
  67 void ff_put_no_rnd_pixels8_l2_mmxext(uint8_t *dst, uint8_t *src1,
  68                                      uint8_t *src2, int dstStride,
  69                                      int src1Stride, int h);
  70 void ff_avg_pixels8_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
  71                               int dstStride, int src1Stride, int h);
  72 void ff_put_pixels16_x2_mmxext(uint8_t *block, const uint8_t *pixels,
  73                                ptrdiff_t line_size, int h);
  74 void ff_put_pixels16_x2_3dnow(uint8_t *block, const uint8_t *pixels,
  75                               ptrdiff_t line_size, int h);
  76 void ff_put_pixels16_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
  77                                int dstStride, int src1Stride, int h);
  78 void ff_avg_pixels16_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
  79                                int dstStride, int src1Stride, int h);
  80 void ff_put_no_rnd_pixels16_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
  81                                       int dstStride, int src1Stride, int h);
  82 void ff_put_no_rnd_pixels8_x2_mmxext(uint8_t *block, const uint8_t *pixels,
  83                                      ptrdiff_t line_size, int h);
  84 void ff_put_no_rnd_pixels8_x2_3dnow(uint8_t *block, const uint8_t *pixels,
  85                                     ptrdiff_t line_size, int h);
  86 void ff_put_no_rnd_pixels8_x2_exact_mmxext(uint8_t *block,
  87                                            const uint8_t *pixels,
  88                                            ptrdiff_t line_size, int h);
  89 void ff_put_no_rnd_pixels8_x2_exact_3dnow(uint8_t *block,
  90                                           const uint8_t *pixels,
  91                                           ptrdiff_t line_size, int h);
  92 void ff_put_pixels8_y2_mmxext(uint8_t *block, const uint8_t *pixels,
  93                               ptrdiff_t line_size, int h);
  94 void ff_put_pixels8_y2_3dnow(uint8_t *block, const uint8_t *pixels,
  95                              ptrdiff_t line_size, int h);
  96 void ff_put_no_rnd_pixels8_y2_mmxext(uint8_t *block, const uint8_t *pixels,
  97                                      ptrdiff_t line_size, int h);
  98 void ff_put_no_rnd_pixels8_y2_3dnow(uint8_t *block, const uint8_t *pixels,
  99                                     ptrdiff_t line_size, int h);
 100 void ff_put_no_rnd_pixels8_y2_exact_mmxext(uint8_t *block,
 101                                            const uint8_t *pixels,
 102                                            ptrdiff_t line_size, int h);
 103 void ff_put_no_rnd_pixels8_y2_exact_3dnow(uint8_t *block,
 104                                           const uint8_t *pixels,
 105                                           ptrdiff_t line_size, int h);
 106 void ff_avg_pixels8_3dnow(uint8_t *block, const uint8_t *pixels,
 107                           ptrdiff_t line_size, int h);
 108 void ff_avg_pixels8_x2_mmxext(uint8_t *block, const uint8_t *pixels,
 109                               ptrdiff_t line_size, int h);
 110 void ff_avg_pixels8_x2_3dnow(uint8_t *block, const uint8_t *pixels,
 111                              ptrdiff_t line_size, int h);
 112 void ff_avg_pixels8_y2_mmxext(uint8_t *block, const uint8_t *pixels,
 113                               ptrdiff_t line_size, int h);
 114 void ff_avg_pixels8_y2_3dnow(uint8_t *block, const uint8_t *pixels,
 115                              ptrdiff_t line_size, int h);
 116 void ff_avg_pixels8_xy2_mmxext(uint8_t *block, const uint8_t *pixels,
 117                                ptrdiff_t line_size, int h);
 118 void ff_avg_pixels8_xy2_3dnow(uint8_t *block, const uint8_t *pixels,
 119                               ptrdiff_t line_size, int h);
 120
 121 void ff_put_pixels8_mmxext(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h);
 122 static void ff_put_pixels16_mmxext(uint8_t *block, const uint8_t *pixels,
 123                                    ptrdiff_t line_size, int h)
 124 {
 125     ff_put_pixels8_mmxext(block,     pixels,     line_size, h);
 126     ff_put_pixels8_mmxext(block + 8, pixels + 8, line_size, h);
 127 }
 128
 129 void ff_put_mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 130                                          int dstStride, int srcStride, int h);
 131 void ff_avg_mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 132                                          int dstStride, int srcStride, int h);
 133 void ff_put_no_rnd_mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 134                                                  int dstStride, int srcStride,
 135                                                  int h);
 136 void ff_put_mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 137                                         int dstStride, int srcStride, int h);
 138 void ff_avg_mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 139                                         int dstStride, int srcStride, int h);
 140 void ff_put_no_rnd_mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 141                                                 int dstStride, int srcStride,
 142                                                 int h);
 143 void ff_put_mpeg4_qpel16_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 144                                          int dstStride, int srcStride);
 145 void ff_avg_mpeg4_qpel16_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 146                                          int dstStride, int srcStride);
 147 void ff_put_no_rnd_mpeg4_qpel16_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 148                                                  int dstStride, int srcStride);
 149 void ff_put_mpeg4_qpel8_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 150                                         int dstStride, int srcStride);
 151 void ff_avg_mpeg4_qpel8_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 152                                         int dstStride, int srcStride);
 153 void ff_put_no_rnd_mpeg4_qpel8_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
 154                                                 int dstStride, int srcStride);
 155 #define ff_put_no_rnd_pixels16_mmxext ff_put_pixels16_mmxext
 156 #define ff_put_no_rnd_pixels8_mmxext ff_put_pixels8_mmxext
 157 #endif /* HAVE_YASM */
 158
 159
 160 #if HAVE_INLINE_ASM
 161
 162 #define JUMPALIGN()     __asm__ volatile (".p2align 3"::)
 163 #define MOVQ_ZERO(regd) __asm__ volatile ("pxor %%"#regd", %%"#regd ::)
 164
 165 #define MOVQ_BFE(regd)                                  \
 166     __asm__ volatile (                                  \
 167         "pcmpeqd %%"#regd", %%"#regd"   \n\t"           \
 168         "paddb   %%"#regd", %%"#regd"   \n\t" ::)
 169
 170 #ifndef PIC
 171 #define MOVQ_BONE(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_bone))
 172 #define MOVQ_WTWO(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_wtwo))
 173 #else
 174 // for shared library it's better to use this way for accessing constants
 175 // pcmpeqd -> -1
 176 #define MOVQ_BONE(regd)                                 \
 177     __asm__ volatile (                                  \
 178         "pcmpeqd  %%"#regd", %%"#regd"  \n\t"           \
 179         "psrlw          $15, %%"#regd"  \n\t"           \
 180         "packuswb %%"#regd", %%"#regd"  \n\t" ::)
 181
 182 #define MOVQ_WTWO(regd)                                 \
 183     __asm__ volatile (                                  \
 184         "pcmpeqd %%"#regd", %%"#regd"   \n\t"           \
 185         "psrlw         $15, %%"#regd"   \n\t"           \
 186         "psllw          $1, %%"#regd"   \n\t"::)
 187
 188 #endif
 189
 190 // using regr as temporary and for the output result
 191 // first argument is unmodifed and second is trashed
 192 // regfe is supposed to contain 0xfefefefefefefefe
 193 #define PAVGB_MMX_NO_RND(rega, regb, regr, regfe)                \
 194     "movq   "#rega", "#regr"            \n\t"                    \
 195     "pand   "#regb", "#regr"            \n\t"                    \
 196     "pxor   "#rega", "#regb"            \n\t"                    \
 197     "pand  "#regfe", "#regb"            \n\t"                    \
 198     "psrlq       $1, "#regb"            \n\t"                    \
 199     "paddb  "#regb", "#regr"            \n\t"
 200
 201 #define PAVGB_MMX(rega, regb, regr, regfe)                       \
 202     "movq   "#rega", "#regr"            \n\t"                    \
 203     "por    "#regb", "#regr"            \n\t"                    \
 204     "pxor   "#rega", "#regb"            \n\t"                    \
 205     "pand  "#regfe", "#regb"            \n\t"                    \
 206     "psrlq       $1, "#regb"            \n\t"                    \
 207     "psubb  "#regb", "#regr"            \n\t"
 208
 209 // mm6 is supposed to contain 0xfefefefefefefefe
 210 #define PAVGBP_MMX_NO_RND(rega, regb, regr,  regc, regd, regp)   \
 211     "movq  "#rega", "#regr"             \n\t"                    \
 212     "movq  "#regc", "#regp"             \n\t"                    \
 213     "pand  "#regb", "#regr"             \n\t"                    \
 214     "pand  "#regd", "#regp"             \n\t"                    \
 215     "pxor  "#rega", "#regb"             \n\t"                    \
 216     "pxor  "#regc", "#regd"             \n\t"                    \
 217     "pand    %%mm6, "#regb"             \n\t"                    \
 218     "pand    %%mm6, "#regd"             \n\t"                    \
 219     "psrlq      $1, "#regb"             \n\t"                    \
 220     "psrlq      $1, "#regd"             \n\t"                    \
 221     "paddb "#regb", "#regr"             \n\t"                    \
 222     "paddb "#regd", "#regp"             \n\t"
 223
 224 #define PAVGBP_MMX(rega, regb, regr, regc, regd, regp)           \
 225     "movq  "#rega", "#regr"             \n\t"                    \
 226     "movq  "#regc", "#regp"             \n\t"                    \
 227     "por   "#regb", "#regr"             \n\t"                    \
 228     "por   "#regd", "#regp"             \n\t"                    \
 229     "pxor  "#rega", "#regb"             \n\t"                    \
 230     "pxor  "#regc", "#regd"             \n\t"                    \
 231     "pand    %%mm6, "#regb"             \n\t"                    \
 232     "pand    %%mm6, "#regd"             \n\t"                    \
 233     "psrlq      $1, "#regd"             \n\t"                    \
 234     "psrlq      $1, "#regb"             \n\t"                    \
 235     "psubb "#regb", "#regr"             \n\t"                    \
 236     "psubb "#regd", "#regp"             \n\t"
 237
 238 /***********************************/
 239 /* MMX no rounding */
 240 #define NO_RND 1
 241 #define DEF(x, y) x ## _no_rnd_ ## y ## _mmx
 242 #define SET_RND  MOVQ_WONE
 243 #define PAVGBP(a, b, c, d, e, f)        PAVGBP_MMX_NO_RND(a, b, c, d, e, f)
 244 #define PAVGB(a, b, c, e)               PAVGB_MMX_NO_RND(a, b, c, e)
 245 #define OP_AVG(a, b, c, e)              PAVGB_MMX(a, b, c, e)
 246
 247 #include "dsputil_rnd_template.c"
 248
 249 #undef DEF
 250 #undef SET_RND
 251 #undef PAVGBP
 252 #undef PAVGB
 253 #undef NO_RND
 254 /***********************************/
 255 /* MMX rounding */
 256
 257 #define DEF(x, y) x ## _ ## y ## _mmx
 258 #define SET_RND  MOVQ_WTWO
 259 #define PAVGBP(a, b, c, d, e, f)        PAVGBP_MMX(a, b, c, d, e, f)
 260 #define PAVGB(a, b, c, e)               PAVGB_MMX(a, b, c, e)
 261
 262 #include "dsputil_rnd_template.c"
 263
 264 #undef DEF
 265 #undef SET_RND
 266 #undef PAVGBP
 267 #undef PAVGB
 268 #undef OP_AVG
 269
 270 #endif /* HAVE_INLINE_ASM */
 271
 272
 273 #if HAVE_YASM
 274
 275 /***********************************/
 276 /* 3Dnow specific */
 277
 278 #define DEF(x) x ## _3dnow
 279
 280 #include "dsputil_avg_template.c"
 281
 282 #undef DEF
 283
 284 /***********************************/
 285 /* MMXEXT specific */
 286
 287 #define DEF(x) x ## _mmxext
 288
 289 #include "dsputil_avg_template.c"
 290
 291 #undef DEF
 292
 293 #endif /* HAVE_YASM */
 294
 295
 296 #if HAVE_INLINE_ASM
 297 #define put_no_rnd_pixels16_mmx put_pixels16_mmx
 298 #define put_no_rnd_pixels8_mmx put_pixels8_mmx
 299
 300 /***********************************/
 301 /* standard MMX */
 302
 303 void ff_put_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
 304                                int line_size)
 305 {
 306     const int16_t *p;
 307     uint8_t *pix;
 308
 309     /* read the pixels */
 310     p   = block;
 311     pix = pixels;
 312     /* unrolled loop */
 313     __asm__ volatile (
 314         "movq      (%3), %%mm0          \n\t"
 315         "movq     8(%3), %%mm1          \n\t"
 316         "movq    16(%3), %%mm2          \n\t"
 317         "movq    24(%3), %%mm3          \n\t"
 318         "movq    32(%3), %%mm4          \n\t"
 319         "movq    40(%3), %%mm5          \n\t"
 320         "movq    48(%3), %%mm6          \n\t"
 321         "movq    56(%3), %%mm7          \n\t"
 322         "packuswb %%mm1, %%mm0          \n\t"
 323         "packuswb %%mm3, %%mm2          \n\t"
 324         "packuswb %%mm5, %%mm4          \n\t"
 325         "packuswb %%mm7, %%mm6          \n\t"
 326         "movq     %%mm0, (%0)           \n\t"
 327         "movq     %%mm2, (%0, %1)       \n\t"
 328         "movq     %%mm4, (%0, %1, 2)    \n\t"
 329         "movq     %%mm6, (%0, %2)       \n\t"
 330         :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3),
 331            "r"(p)
 332         : "memory");
 333     pix += line_size * 4;
 334     p   += 32;
 335
 336     // if here would be an exact copy of the code above
 337     // compiler would generate some very strange code
 338     // thus using "r"
 339     __asm__ volatile (
 340         "movq       (%3), %%mm0         \n\t"
 341         "movq      8(%3), %%mm1         \n\t"
 342         "movq     16(%3), %%mm2         \n\t"
 343         "movq     24(%3), %%mm3         \n\t"
 344         "movq     32(%3), %%mm4         \n\t"
 345         "movq     40(%3), %%mm5         \n\t"
 346         "movq     48(%3), %%mm6         \n\t"
 347         "movq     56(%3), %%mm7         \n\t"
 348         "packuswb  %%mm1, %%mm0         \n\t"
 349         "packuswb  %%mm3, %%mm2         \n\t"
 350         "packuswb  %%mm5, %%mm4         \n\t"
 351         "packuswb  %%mm7, %%mm6         \n\t"
 352         "movq      %%mm0, (%0)          \n\t"
 353         "movq      %%mm2, (%0, %1)      \n\t"
 354         "movq      %%mm4, (%0, %1, 2)   \n\t"
 355         "movq      %%mm6, (%0, %2)      \n\t"
 356         :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3), "r"(p)
 357         : "memory");
 358 }
 359
 360 #define put_signed_pixels_clamped_mmx_half(off)             \
 361     "movq          "#off"(%2), %%mm1        \n\t"           \
 362     "movq     16 + "#off"(%2), %%mm2        \n\t"           \
 363     "movq     32 + "#off"(%2), %%mm3        \n\t"           \
 364     "movq     48 + "#off"(%2), %%mm4        \n\t"           \
 365     "packsswb  8 + "#off"(%2), %%mm1        \n\t"           \
 366     "packsswb 24 + "#off"(%2), %%mm2        \n\t"           \
 367     "packsswb 40 + "#off"(%2), %%mm3        \n\t"           \
 368     "packsswb 56 + "#off"(%2), %%mm4        \n\t"           \
 369     "paddb              %%mm0, %%mm1        \n\t"           \
 370     "paddb              %%mm0, %%mm2        \n\t"           \
 371     "paddb              %%mm0, %%mm3        \n\t"           \
 372     "paddb              %%mm0, %%mm4        \n\t"           \
 373     "movq               %%mm1, (%0)         \n\t"           \
 374     "movq               %%mm2, (%0, %3)     \n\t"           \
 375     "movq               %%mm3, (%0, %3, 2)  \n\t"           \
 376     "movq               %%mm4, (%0, %1)     \n\t"
 377
 378 void ff_put_signed_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
 379                                       int line_size)
 380 {
 381     x86_reg line_skip = line_size;
 382     x86_reg line_skip3;
 383
 384     __asm__ volatile (
 385         "movq "MANGLE(ff_pb_80)", %%mm0     \n\t"
 386         "lea         (%3, %3, 2), %1        \n\t"
 387         put_signed_pixels_clamped_mmx_half(0)
 388         "lea         (%0, %3, 4), %0        \n\t"
 389         put_signed_pixels_clamped_mmx_half(64)
 390         : "+&r"(pixels), "=&r"(line_skip3)
 391         : "r"(block), "r"(line_skip)
 392         : "memory");
 393 }
 394
 395 void ff_add_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
 396                                int line_size)
 397 {
 398     const int16_t *p;
 399     uint8_t *pix;
 400     int i;
 401
 402     /* read the pixels */
 403     p   = block;
 404     pix = pixels;
 405     MOVQ_ZERO(mm7);
 406     i = 4;
 407     do {
 408         __asm__ volatile (
 409             "movq        (%2), %%mm0    \n\t"
 410             "movq       8(%2), %%mm1    \n\t"
 411             "movq      16(%2), %%mm2    \n\t"
 412             "movq      24(%2), %%mm3    \n\t"
 413             "movq          %0, %%mm4    \n\t"
 414             "movq          %1, %%mm6    \n\t"
 415             "movq       %%mm4, %%mm5    \n\t"
 416             "punpcklbw  %%mm7, %%mm4    \n\t"
 417             "punpckhbw  %%mm7, %%mm5    \n\t"
 418             "paddsw     %%mm4, %%mm0    \n\t"
 419             "paddsw     %%mm5, %%mm1    \n\t"
 420             "movq       %%mm6, %%mm5    \n\t"
 421             "punpcklbw  %%mm7, %%mm6    \n\t"
 422             "punpckhbw  %%mm7, %%mm5    \n\t"
 423             "paddsw     %%mm6, %%mm2    \n\t"
 424             "paddsw     %%mm5, %%mm3    \n\t"
 425             "packuswb   %%mm1, %%mm0    \n\t"
 426             "packuswb   %%mm3, %%mm2    \n\t"
 427             "movq       %%mm0, %0       \n\t"
 428             "movq       %%mm2, %1       \n\t"
 429             : "+m"(*pix), "+m"(*(pix + line_size))
 430             : "r"(p)
 431             : "memory");
 432         pix += line_size * 2;
 433         p   += 16;
 434     } while (--i);
 435 }
 436
 437 static void put_pixels8_mmx(uint8_t *block, const uint8_t *pixels,
 438                             ptrdiff_t line_size, int h)
 439 {
 440     __asm__ volatile (
 441         "lea   (%3, %3), %%"REG_a"      \n\t"
 442         ".p2align     3                 \n\t"
 443         "1:                             \n\t"
 444         "movq  (%1    ), %%mm0          \n\t"
 445         "movq  (%1, %3), %%mm1          \n\t"
 446         "movq     %%mm0, (%2)           \n\t"
 447         "movq     %%mm1, (%2, %3)       \n\t"
 448         "add  %%"REG_a", %1             \n\t"
 449         "add  %%"REG_a", %2             \n\t"
 450         "movq  (%1    ), %%mm0          \n\t"
 451         "movq  (%1, %3), %%mm1          \n\t"
 452         "movq     %%mm0, (%2)           \n\t"
 453         "movq     %%mm1, (%2, %3)       \n\t"
 454         "add  %%"REG_a", %1             \n\t"
 455         "add  %%"REG_a", %2             \n\t"
 456         "subl        $4, %0             \n\t"
 457         "jnz         1b                 \n\t"
 458         : "+g"(h), "+r"(pixels),  "+r"(block)
 459         : "r"((x86_reg)line_size)
 460         : "%"REG_a, "memory"
 461         );
 462 }
 463
 464 static void put_pixels16_mmx(uint8_t *block, const uint8_t *pixels,
 465                              ptrdiff_t line_size, int h)
 466 {
 467     __asm__ volatile (
 468         "lea   (%3, %3), %%"REG_a"      \n\t"
 469         ".p2align     3                 \n\t"
 470         "1:                             \n\t"
 471         "movq  (%1    ), %%mm0          \n\t"
 472         "movq 8(%1    ), %%mm4          \n\t"
 473         "movq  (%1, %3), %%mm1          \n\t"
 474         "movq 8(%1, %3), %%mm5          \n\t"
 475         "movq     %%mm0,  (%2)          \n\t"
 476         "movq     %%mm4, 8(%2)          \n\t"
 477         "movq     %%mm1,  (%2, %3)      \n\t"
 478         "movq     %%mm5, 8(%2, %3)      \n\t"
 479         "add  %%"REG_a", %1             \n\t"
 480         "add  %%"REG_a", %2             \n\t"
 481         "movq  (%1    ), %%mm0          \n\t"
 482         "movq 8(%1    ), %%mm4          \n\t"
 483         "movq  (%1, %3), %%mm1          \n\t"
 484         "movq 8(%1, %3), %%mm5          \n\t"
 485         "movq     %%mm0,  (%2)          \n\t"
 486         "movq     %%mm4, 8(%2)          \n\t"
 487         "movq     %%mm1,  (%2, %3)      \n\t"
 488         "movq     %%mm5, 8(%2, %3)      \n\t"
 489         "add  %%"REG_a", %1             \n\t"
 490         "add  %%"REG_a", %2             \n\t"
 491         "subl        $4, %0             \n\t"
 492         "jnz         1b                 \n\t"
 493         : "+g"(h), "+r"(pixels),  "+r"(block)
 494         : "r"((x86_reg)line_size)
 495         : "%"REG_a, "memory"
 496         );
 497 }
 498
 499 #define CLEAR_BLOCKS(name, n)                           \
 500 static void name(int16_t *blocks)                       \
 501 {                                                       \
 502     __asm__ volatile (                                  \
 503         "pxor %%mm7, %%mm7              \n\t"           \
 504         "mov     %1,        %%"REG_a"   \n\t"           \
 505         "1:                             \n\t"           \
 506         "movq %%mm7,   (%0, %%"REG_a")  \n\t"           \
 507         "movq %%mm7,  8(%0, %%"REG_a")  \n\t"           \
 508         "movq %%mm7, 16(%0, %%"REG_a")  \n\t"           \
 509         "movq %%mm7, 24(%0, %%"REG_a")  \n\t"           \
 510         "add    $32, %%"REG_a"          \n\t"           \
 511         "js      1b                     \n\t"           \
 512         :: "r"(((uint8_t *)blocks) + 128 * n),          \
 513            "i"(-128 * n)                                \
 514         : "%"REG_a                                      \
 515         );                                              \
 516 }
 517 CLEAR_BLOCKS(clear_blocks_mmx, 6)
 518 CLEAR_BLOCKS(clear_block_mmx, 1)
 519
 520 static void clear_block_sse(int16_t *block)
 521 {
 522     __asm__ volatile (
 523         "xorps  %%xmm0, %%xmm0          \n"
 524         "movaps %%xmm0,    (%0)         \n"
 525         "movaps %%xmm0,  16(%0)         \n"
 526         "movaps %%xmm0,  32(%0)         \n"
 527         "movaps %%xmm0,  48(%0)         \n"
 528         "movaps %%xmm0,  64(%0)         \n"
 529         "movaps %%xmm0,  80(%0)         \n"
 530         "movaps %%xmm0,  96(%0)         \n"
 531         "movaps %%xmm0, 112(%0)         \n"
 532         :: "r"(block)
 533         : "memory"
 534     );
 535 }
 536
 537 static void clear_blocks_sse(int16_t *blocks)
 538 {
 539     __asm__ volatile (
 540         "xorps  %%xmm0, %%xmm0              \n"
 541         "mov        %1,         %%"REG_a"   \n"
 542         "1:                                 \n"
 543         "movaps %%xmm0,    (%0, %%"REG_a")  \n"
 544         "movaps %%xmm0,  16(%0, %%"REG_a")  \n"
 545         "movaps %%xmm0,  32(%0, %%"REG_a")  \n"
 546         "movaps %%xmm0,  48(%0, %%"REG_a")  \n"
 547         "movaps %%xmm0,  64(%0, %%"REG_a")  \n"
 548         "movaps %%xmm0,  80(%0, %%"REG_a")  \n"
 549         "movaps %%xmm0,  96(%0, %%"REG_a")  \n"
 550         "movaps %%xmm0, 112(%0, %%"REG_a")  \n"
 551         "add      $128,         %%"REG_a"   \n"
 552         "js         1b                      \n"
 553         :: "r"(((uint8_t *)blocks) + 128 * 6),
 554            "i"(-128 * 6)
 555         : "%"REG_a
 556     );
 557 }
 558
 559 static void add_bytes_mmx(uint8_t *dst, uint8_t *src, int w)
 560 {
 561     x86_reg i = 0;
 562     __asm__ volatile (
 563         "jmp          2f                \n\t"
 564         "1:                             \n\t"
 565         "movq   (%1, %0), %%mm0         \n\t"
 566         "movq   (%2, %0), %%mm1         \n\t"
 567         "paddb     %%mm0, %%mm1         \n\t"
 568         "movq      %%mm1, (%2, %0)      \n\t"
 569         "movq  8(%1, %0), %%mm0         \n\t"
 570         "movq  8(%2, %0), %%mm1         \n\t"
 571         "paddb     %%mm0, %%mm1         \n\t"
 572         "movq      %%mm1, 8(%2, %0)     \n\t"
 573         "add         $16, %0            \n\t"
 574         "2:                             \n\t"
 575         "cmp          %3, %0            \n\t"
 576         "js           1b                \n\t"
 577         : "+r"(i)
 578         : "r"(src), "r"(dst), "r"((x86_reg)w - 15)
 579     );
 580     for ( ; i < w; i++)
 581         dst[i + 0] += src[i + 0];
 582 }
 583
 584 #if HAVE_7REGS
 585 static void add_hfyu_median_prediction_cmov(uint8_t *dst, const uint8_t *top,
 586                                             const uint8_t *diff, int w,
 587                                             int *left, int *left_top)
 588 {
 589     x86_reg w2 = -w;
 590     x86_reg x;
 591     int l  = *left     & 0xff;
 592     int tl = *left_top & 0xff;
 593     int t;
 594     __asm__ volatile (
 595         "mov          %7, %3            \n"
 596         "1:                             \n"
 597         "movzbl (%3, %4), %2            \n"
 598         "mov          %2, %k3           \n"
 599         "sub         %b1, %b3           \n"
 600         "add         %b0, %b3           \n"
 601         "mov          %2, %1            \n"
 602         "cmp          %0, %2            \n"
 603         "cmovg        %0, %2            \n"
 604         "cmovg        %1, %0            \n"
 605         "cmp         %k3, %0            \n"
 606         "cmovg       %k3, %0            \n"
 607         "mov          %7, %3            \n"
 608         "cmp          %2, %0            \n"
 609         "cmovl        %2, %0            \n"
 610         "add    (%6, %4), %b0           \n"
 611         "mov         %b0, (%5, %4)      \n"
 612         "inc          %4                \n"
 613         "jl           1b                \n"
 614         : "+&q"(l), "+&q"(tl), "=&r"(t), "=&q"(x), "+&r"(w2)
 615         : "r"(dst + w), "r"(diff + w), "rm"(top + w)
 616     );
 617     *left     = l;
 618     *left_top = tl;
 619 }
 620 #endif
 621 #endif /* HAVE_INLINE_ASM */
 622
 623 void ff_h263_v_loop_filter_mmx(uint8_t *src, int stride, int qscale);
 624 void ff_h263_h_loop_filter_mmx(uint8_t *src, int stride, int qscale);
 625
 626 #if HAVE_INLINE_ASM
 627 /* Draw the edges of width 'w' of an image of size width, height
 628  * this MMX version can only handle w == 8 || w == 16. */
 629 static void draw_edges_mmx(uint8_t *buf, int wrap, int width, int height,
 630                            int w, int h, int sides)
 631 {
 632     uint8_t *ptr, *last_line;
 633     int i;
 634
 635     last_line = buf + (height - 1) * wrap;
 636     /* left and right */
 637     ptr = buf;
 638     if (w == 8) {
 639         __asm__ volatile (
 640             "1:                             \n\t"
 641             "movd            (%0), %%mm0    \n\t"
 642             "punpcklbw      %%mm0, %%mm0    \n\t"
 643             "punpcklwd      %%mm0, %%mm0    \n\t"
 644             "punpckldq      %%mm0, %%mm0    \n\t"
 645             "movq           %%mm0, -8(%0)   \n\t"
 646             "movq      -8(%0, %2), %%mm1    \n\t"
 647             "punpckhbw      %%mm1, %%mm1    \n\t"
 648             "punpckhwd      %%mm1, %%mm1    \n\t"
 649             "punpckhdq      %%mm1, %%mm1    \n\t"
 650             "movq           %%mm1, (%0, %2) \n\t"
 651             "add               %1, %0       \n\t"
 652             "cmp               %3, %0       \n\t"
 653             "jb                1b           \n\t"
 654             : "+r"(ptr)
 655             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
 656             );
 657     } else {
 658         __asm__ volatile (
 659             "1:                                 \n\t"
 660             "movd            (%0), %%mm0        \n\t"
 661             "punpcklbw      %%mm0, %%mm0        \n\t"
 662             "punpcklwd      %%mm0, %%mm0        \n\t"
 663             "punpckldq      %%mm0, %%mm0        \n\t"
 664             "movq           %%mm0, -8(%0)       \n\t"
 665             "movq           %%mm0, -16(%0)      \n\t"
 666             "movq      -8(%0, %2), %%mm1        \n\t"
 667             "punpckhbw      %%mm1, %%mm1        \n\t"
 668             "punpckhwd      %%mm1, %%mm1        \n\t"
 669             "punpckhdq      %%mm1, %%mm1        \n\t"
 670             "movq           %%mm1,  (%0, %2)    \n\t"
 671             "movq           %%mm1, 8(%0, %2)    \n\t"
 672             "add               %1, %0           \n\t"
 673             "cmp               %3, %0           \n\t"
 674             "jb                1b               \n\t"
 675             : "+r"(ptr)
 676             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
 677             );
 678     }
 679
 680     /* top and bottom (and hopefully also the corners) */
 681     if (sides & EDGE_TOP) {
 682         for (i = 0; i < h; i += 4) {
 683             ptr = buf - (i + 1) * wrap - w;
 684             __asm__ volatile (
 685                 "1:                             \n\t"
 686                 "movq (%1, %0), %%mm0           \n\t"
 687                 "movq    %%mm0, (%0)            \n\t"
 688                 "movq    %%mm0, (%0, %2)        \n\t"
 689                 "movq    %%mm0, (%0, %2, 2)     \n\t"
 690                 "movq    %%mm0, (%0, %3)        \n\t"
 691                 "add        $8, %0              \n\t"
 692                 "cmp        %4, %0              \n\t"
 693                 "jb         1b                  \n\t"
 694                 : "+r"(ptr)
 695                 : "r"((x86_reg)buf - (x86_reg)ptr - w), "r"((x86_reg) -wrap),
 696                   "r"((x86_reg) -wrap * 3), "r"(ptr + width + 2 * w)
 697                 );
 698         }
 699     }
 700
 701     if (sides & EDGE_BOTTOM) {
 702         for (i = 0; i < h; i += 4) {
 703             ptr = last_line + (i + 1) * wrap - w;
 704             __asm__ volatile (
 705                 "1:                             \n\t"
 706                 "movq (%1, %0), %%mm0           \n\t"
 707                 "movq    %%mm0, (%0)            \n\t"
 708                 "movq    %%mm0, (%0, %2)        \n\t"
 709                 "movq    %%mm0, (%0, %2, 2)     \n\t"
 710                 "movq    %%mm0, (%0, %3)        \n\t"
 711                 "add        $8, %0              \n\t"
 712                 "cmp        %4, %0              \n\t"
 713                 "jb         1b                  \n\t"
 714                 : "+r"(ptr)
 715                 : "r"((x86_reg)last_line - (x86_reg)ptr - w),
 716                   "r"((x86_reg)wrap), "r"((x86_reg)wrap * 3),
 717                   "r"(ptr + width + 2 * w)
 718                 );
 719         }
 720     }
 721 }
 722 #endif /* HAVE_INLINE_ASM */
 723
 724
 725 #if HAVE_YASM
 726 #define QPEL_OP(OPNAME, ROUNDER, RND, MMX)                              \
 727 static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src,   \
 728                                           ptrdiff_t stride)             \
 729 {                                                                       \
 730     ff_ ## OPNAME ## pixels8_ ## MMX(dst, src, stride, 8);              \
 731 }                                                                       \
 732                                                                         \
 733 static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src,    \
 734                                          ptrdiff_t stride)              \
 735 {                                                                       \
 736     uint64_t temp[8];                                                   \
 737     uint8_t * const half = (uint8_t*)temp;                              \
 738     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8,        \
 739                                                    stride, 8);          \
 740     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src, half,                 \
 741                                         stride, stride, 8);             \
 742 }                                                                       \
 743                                                                         \
 744 static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src,    \
 745                                          ptrdiff_t stride)              \
 746 {                                                                       \
 747     ff_ ## OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride,    \
 748                                                    stride, 8);          \
 749 }                                                                       \
 750                                                                         \
 751 static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src,    \
 752                                          ptrdiff_t stride)              \
 753 {                                                                       \
 754     uint64_t temp[8];                                                   \
 755     uint8_t * const half = (uint8_t*)temp;                              \
 756     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8,        \
 757                                                    stride, 8);          \
 758     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src + 1, half, stride,     \
 759                                         stride, 8);                     \
 760 }                                                                       \
 761                                                                         \
 762 static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src,    \
 763                                          ptrdiff_t stride)              \
 764 {                                                                       \
 765     uint64_t temp[8];                                                   \
 766     uint8_t * const half = (uint8_t*)temp;                              \
 767     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src,           \
 768                                                    8, stride);          \
 769     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src, half,                 \
 770                                         stride, stride, 8);             \
 771 }                                                                       \
 772                                                                         \
 773 static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src,    \
 774                                          ptrdiff_t stride)              \
 775 {                                                                       \
 776     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src,            \
 777                                                    stride, stride);     \
 778 }                                                                       \
 779                                                                         \
 780 static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src,    \
 781                                          ptrdiff_t stride)              \
 782 {                                                                       \
 783     uint64_t temp[8];                                                   \
 784     uint8_t * const half = (uint8_t*)temp;                              \
 785     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src,           \
 786                                                    8, stride);          \
 787     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src + stride, half, stride,\
 788                                         stride, 8);                     \
 789 }                                                                       \
 790                                                                         \
 791 static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src,    \
 792                                          ptrdiff_t stride)              \
 793 {                                                                       \
 794     uint64_t half[8 + 9];                                               \
 795     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
 796     uint8_t * const halfHV = ((uint8_t*)half);                          \
 797     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 798                                                    stride, 9);          \
 799     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8,           \
 800                                         stride, 9);                     \
 801     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
 802     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV,             \
 803                                         stride, 8, 8);                  \
 804 }                                                                       \
 805                                                                         \
 806 static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src,    \
 807                                          ptrdiff_t stride)              \
 808 {                                                                       \
 809     uint64_t half[8 + 9];                                               \
 810     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
 811     uint8_t * const halfHV = ((uint8_t*)half);                          \
 812     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 813                                                    stride, 9);          \
 814     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,       \
 815                                         stride, 9);                     \
 816     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
 817     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV,             \
 818                                         stride, 8, 8);                  \
 819 }                                                                       \
 820                                                                         \
 821 static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src,    \
 822                                          ptrdiff_t stride)              \
 823 {                                                                       \
 824     uint64_t half[8 + 9];                                               \
 825     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
 826     uint8_t * const halfHV = ((uint8_t*)half);                          \
 827     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 828                                                    stride, 9);          \
 829     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8,           \
 830                                         stride, 9);                     \
 831     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
 832     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV,         \
 833                                         stride, 8, 8);                  \
 834 }                                                                       \
 835                                                                         \
 836 static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src,    \
 837                                          ptrdiff_t stride)              \
 838 {                                                                       \
 839     uint64_t half[8 + 9];                                               \
 840     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
 841     uint8_t * const halfHV = ((uint8_t*)half);                          \
 842     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 843                                                    stride, 9);          \
 844     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,       \
 845                                         stride, 9);                     \
 846     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
 847     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV,         \
 848                                         stride, 8, 8);                  \
 849 }                                                                       \
 850                                                                         \
 851 static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src,    \
 852                                          ptrdiff_t stride)              \
 853 {                                                                       \
 854     uint64_t half[8 + 9];                                               \
 855     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
 856     uint8_t * const halfHV = ((uint8_t*)half);                          \
 857     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 858                                                    stride, 9);          \
 859     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
 860     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV,             \
 861                                         stride, 8, 8);                  \
 862 }                                                                       \
 863                                                                         \
 864 static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src,    \
 865                                          ptrdiff_t stride)              \
 866 {                                                                       \
 867     uint64_t half[8 + 9];                                               \
 868     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
 869     uint8_t * const halfHV = ((uint8_t*)half);                          \
 870     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 871                                                    stride, 9);          \
 872     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
 873     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV,         \
 874                                         stride, 8, 8);                  \
 875 }                                                                       \
 876                                                                         \
 877 static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src,    \
 878                                          ptrdiff_t stride)              \
 879 {                                                                       \
 880     uint64_t half[8 + 9];                                               \
 881     uint8_t * const halfH = ((uint8_t*)half);                           \
 882     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 883                                                    stride, 9);          \
 884     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH,              \
 885                                         8, stride, 9);                  \
 886     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH,          \
 887                                                    stride, 8);          \
 888 }                                                                       \
 889                                                                         \
 890 static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src,    \
 891                                          ptrdiff_t stride)              \
 892 {                                                                       \
 893     uint64_t half[8 + 9];                                               \
 894     uint8_t * const halfH = ((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_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH,          \
 900                                                    stride, 8);          \
 901 }                                                                       \
 902                                                                         \
 903 static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src,    \
 904                                          ptrdiff_t stride)              \
 905 {                                                                       \
 906     uint64_t half[9];                                                   \
 907     uint8_t * const halfH = ((uint8_t*)half);                           \
 908     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
 909                                                    stride, 9);          \
 910     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH,          \
 911                                                    stride, 8);          \
 912 }                                                                       \
 913                                                                         \
 914 static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src,  \
 915                                            ptrdiff_t stride)            \
 916 {                                                                       \
 917     ff_ ## OPNAME ## pixels16_ ## MMX(dst, src, stride, 16);            \
 918 }                                                                       \
 919                                                                         \
 920 static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src,   \
 921                                           ptrdiff_t stride)             \
 922 {                                                                       \
 923     uint64_t temp[32];                                                  \
 924     uint8_t * const half = (uint8_t*)temp;                              \
 925     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16,      \
 926                                                     stride, 16);        \
 927     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride,        \
 928                                          stride, 16);                   \
 929 }                                                                       \
 930                                                                         \
 931 static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src,   \
 932                                           ptrdiff_t stride)             \
 933 {                                                                       \
 934     ff_ ## OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src,           \
 935                                                     stride, stride, 16);\
 936 }                                                                       \
 937                                                                         \
 938 static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src,   \
 939                                           ptrdiff_t stride)             \
 940 {                                                                       \
 941     uint64_t temp[32];                                                  \
 942     uint8_t * const half = (uint8_t*)temp;                              \
 943     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16,      \
 944                                                     stride, 16);        \
 945     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src + 1, half,            \
 946                                          stride, stride, 16);           \
 947 }                                                                       \
 948                                                                         \
 949 static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src,   \
 950                                           ptrdiff_t stride)             \
 951 {                                                                       \
 952     uint64_t temp[32];                                                  \
 953     uint8_t * const half = (uint8_t*)temp;                              \
 954     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16,      \
 955                                                     stride);            \
 956     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride,        \
 957                                          stride, 16);                   \
 958 }                                                                       \
 959                                                                         \
 960 static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src,   \
 961                                           ptrdiff_t stride)             \
 962 {                                                                       \
 963     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src,           \
 964                                                     stride, stride);    \
 965 }                                                                       \
 966                                                                         \
 967 static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src,   \
 968                                           ptrdiff_t stride)             \
 969 {                                                                       \
 970     uint64_t temp[32];                                                  \
 971     uint8_t * const half = (uint8_t*)temp;                              \
 972     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16,      \
 973                                                     stride);            \
 974     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half,         \
 975                                          stride, stride, 16);           \
 976 }                                                                       \
 977                                                                         \
 978 static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src,   \
 979                                           ptrdiff_t stride)             \
 980 {                                                                       \
 981     uint64_t half[16 * 2 + 17 * 2];                                     \
 982     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
 983     uint8_t * const halfHV = ((uint8_t*)half);                          \
 984     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
 985                                                     stride, 17);        \
 986     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,         \
 987                                          stride, 17);                   \
 988     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
 989                                                     16, 16);            \
 990     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV,            \
 991                                          stride, 16, 16);               \
 992 }                                                                       \
 993                                                                         \
 994 static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src,   \
 995                                           ptrdiff_t stride)             \
 996 {                                                                       \
 997     uint64_t half[16 * 2 + 17 * 2];                                     \
 998     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
 999     uint8_t * const halfHV = ((uint8_t*)half);                          \
1000     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1001                                                     stride, 17);        \
1002     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,     \
1003                                          stride, 17);                   \
1004     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1005                                                     16, 16);            \
1006     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV,            \
1007                                          stride, 16, 16);               \
1008 }                                                                       \
1009                                                                         \
1010 static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src,   \
1011                                           ptrdiff_t stride)             \
1012 {                                                                       \
1013     uint64_t half[16 * 2 + 17 * 2];                                     \
1014     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1015     uint8_t * const halfHV = ((uint8_t*)half);                          \
1016     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1017                                                     stride, 17);        \
1018     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,         \
1019                                          stride, 17);                   \
1020     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1021                                                     16, 16);            \
1022     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV,       \
1023                                          stride, 16, 16);               \
1024 }                                                                       \
1025                                                                         \
1026 static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src,   \
1027                                           ptrdiff_t stride)             \
1028 {                                                                       \
1029     uint64_t half[16 * 2 + 17 * 2];                                     \
1030     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1031     uint8_t * const halfHV = ((uint8_t*)half);                          \
1032     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1033                                                     stride, 17);        \
1034     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,     \
1035                                          stride, 17);                   \
1036     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1037                                                     16, 16);            \
1038     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV,       \
1039                                          stride, 16, 16);               \
1040 }                                                                       \
1041                                                                         \
1042 static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src,   \
1043                                           ptrdiff_t stride)             \
1044 {                                                                       \
1045     uint64_t half[16 * 2 + 17 * 2];                                     \
1046     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1047     uint8_t * const halfHV = ((uint8_t*)half);                          \
1048     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1049                                                     stride, 17);        \
1050     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1051                                                     16, 16);            \
1052     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV,            \
1053                                          stride, 16, 16);               \
1054 }                                                                       \
1055                                                                         \
1056 static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src,   \
1057                                           ptrdiff_t stride)             \
1058 {                                                                       \
1059     uint64_t half[16 * 2 + 17 * 2];                                     \
1060     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1061     uint8_t * const halfHV = ((uint8_t*)half);                          \
1062     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1063                                                     stride, 17);        \
1064     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1065                                                     16, 16);            \
1066     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV,       \
1067                                          stride, 16, 16);               \
1068 }                                                                       \
1069                                                                         \
1070 static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src,   \
1071                                           ptrdiff_t stride)             \
1072 {                                                                       \
1073     uint64_t half[17 * 2];                                              \
1074     uint8_t * const halfH = ((uint8_t*)half);                           \
1075     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1076                                                     stride, 17);        \
1077     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,         \
1078                                          stride, 17);                   \
1079     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH,         \
1080                                                     stride, 16);        \
1081 }                                                                       \
1082                                                                         \
1083 static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src,   \
1084                                           ptrdiff_t stride)             \
1085 {                                                                       \
1086     uint64_t half[17 * 2];                                              \
1087     uint8_t * const halfH = ((uint8_t*)half);                           \
1088     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1089                                                     stride, 17);        \
1090     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,     \
1091                                          stride, 17);                   \
1092     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH,         \
1093                                                     stride, 16);        \
1094 }                                                                       \
1095                                                                         \
1096 static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src,   \
1097                                           ptrdiff_t stride)             \
1098 {                                                                       \
1099     uint64_t half[17 * 2];                                              \
1100     uint8_t * const halfH = ((uint8_t*)half);                           \
1101     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1102                                                     stride, 17);        \
1103     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH,         \
1104                                                     stride, 16);        \
1105 }
1106
1107 QPEL_OP(put_,          ff_pw_16, _,        mmxext)
1108 QPEL_OP(avg_,          ff_pw_16, _,        mmxext)
1109 QPEL_OP(put_no_rnd_,   ff_pw_15, _no_rnd_, mmxext)
1110 #endif /* HAVE_YASM */
1111
1112
1113 #if HAVE_INLINE_ASM
1114 void ff_put_rv40_qpel8_mc33_mmx(uint8_t *dst, uint8_t *src, ptrdiff_t stride)
1115 {
1116   put_pixels8_xy2_mmx(dst, src, stride, 8);
1117 }
1118 void ff_put_rv40_qpel16_mc33_mmx(uint8_t *dst, uint8_t *src, ptrdiff_t stride)
1119 {
1120   put_pixels16_xy2_mmx(dst, src, stride, 16);
1121 }
1122 void ff_avg_rv40_qpel8_mc33_mmx(uint8_t *dst, uint8_t *src, ptrdiff_t stride)
1123 {
1124   avg_pixels8_xy2_mmx(dst, src, stride, 8);
1125 }
1126 void ff_avg_rv40_qpel16_mc33_mmx(uint8_t *dst, uint8_t *src, ptrdiff_t stride)
1127 {
1128   avg_pixels16_xy2_mmx(dst, src, stride, 16);
1129 }
1130
1131 static void gmc_mmx(uint8_t *dst, uint8_t *src,
1132                     int stride, int h, int ox, int oy,
1133                     int dxx, int dxy, int dyx, int dyy,
1134                     int shift, int r, int width, int height)
1135 {
1136     const int w    = 8;
1137     const int ix   = ox  >> (16 + shift);
1138     const int iy   = oy  >> (16 + shift);
1139     const int oxs  = ox  >> 4;
1140     const int oys  = oy  >> 4;
1141     const int dxxs = dxx >> 4;
1142     const int dxys = dxy >> 4;
1143     const int dyxs = dyx >> 4;
1144     const int dyys = dyy >> 4;
1145     const uint16_t r4[4]   = { r, r, r, r };
1146     const uint16_t dxy4[4] = { dxys, dxys, dxys, dxys };
1147     const uint16_t dyy4[4] = { dyys, dyys, dyys, dyys };
1148     const uint64_t shift2 = 2 * shift;
1149     int x, y;
1150
1151     const int dxw = (dxx - (1 << (16 + shift))) * (w - 1);
1152     const int dyh = (dyy - (1 << (16 + shift))) * (h - 1);
1153     const int dxh = dxy * (h - 1);
1154     const int dyw = dyx * (w - 1);
1155     if ( // non-constant fullpel offset (3% of blocks)
1156         ((ox ^ (ox + dxw)) | (ox ^ (ox + dxh)) | (ox ^ (ox + dxw + dxh)) |
1157          (oy ^ (oy + dyw)) | (oy ^ (oy + dyh)) | (oy ^ (oy + dyw + dyh))) >> (16 + shift)
1158         // uses more than 16 bits of subpel mv (only at huge resolution)
1159         || (dxx | dxy | dyx | dyy) & 15 ||
1160         (unsigned)ix >= width  - w ||
1161         (unsigned)iy >= height - h) {
1162         // FIXME could still use mmx for some of the rows
1163         ff_gmc_c(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy,
1164                  shift, r, width, height);
1165         return;
1166     }
1167
1168     src += ix + iy * stride;
1169
1170     __asm__ volatile (
1171         "movd         %0, %%mm6         \n\t"
1172         "pxor      %%mm7, %%mm7         \n\t"
1173         "punpcklwd %%mm6, %%mm6         \n\t"
1174         "punpcklwd %%mm6, %%mm6         \n\t"
1175         :: "r"(1<<shift)
1176     );
1177
1178     for (x = 0; x < w; x += 4) {
1179         uint16_t dx4[4] = { oxs - dxys + dxxs * (x + 0),
1180                             oxs - dxys + dxxs * (x + 1),
1181                             oxs - dxys + dxxs * (x + 2),
1182                             oxs - dxys + dxxs * (x + 3) };
1183         uint16_t dy4[4] = { oys - dyys + dyxs * (x + 0),
1184                             oys - dyys + dyxs * (x + 1),
1185                             oys - dyys + dyxs * (x + 2),
1186                             oys - dyys + dyxs * (x + 3) };
1187
1188         for (y = 0; y < h; y++) {
1189             __asm__ volatile (
1190                 "movq      %0, %%mm4    \n\t"
1191                 "movq      %1, %%mm5    \n\t"
1192                 "paddw     %2, %%mm4    \n\t"
1193                 "paddw     %3, %%mm5    \n\t"
1194                 "movq   %%mm4, %0       \n\t"
1195                 "movq   %%mm5, %1       \n\t"
1196                 "psrlw    $12, %%mm4    \n\t"
1197                 "psrlw    $12, %%mm5    \n\t"
1198                 : "+m"(*dx4), "+m"(*dy4)
1199                 : "m"(*dxy4), "m"(*dyy4)
1200             );
1201
1202             __asm__ volatile (
1203                 "movq      %%mm6, %%mm2 \n\t"
1204                 "movq      %%mm6, %%mm1 \n\t"
1205                 "psubw     %%mm4, %%mm2 \n\t"
1206                 "psubw     %%mm5, %%mm1 \n\t"
1207                 "movq      %%mm2, %%mm0 \n\t"
1208                 "movq      %%mm4, %%mm3 \n\t"
1209                 "pmullw    %%mm1, %%mm0 \n\t" // (s - dx) * (s - dy)
1210                 "pmullw    %%mm5, %%mm3 \n\t" // dx * dy
1211                 "pmullw    %%mm5, %%mm2 \n\t" // (s - dx) * dy
1212                 "pmullw    %%mm4, %%mm1 \n\t" // dx * (s - dy)
1213
1214                 "movd         %4, %%mm5 \n\t"
1215                 "movd         %3, %%mm4 \n\t"
1216                 "punpcklbw %%mm7, %%mm5 \n\t"
1217                 "punpcklbw %%mm7, %%mm4 \n\t"
1218                 "pmullw    %%mm5, %%mm3 \n\t" // src[1, 1] * dx * dy
1219                 "pmullw    %%mm4, %%mm2 \n\t" // src[0, 1] * (s - dx) * dy
1220
1221                 "movd         %2, %%mm5 \n\t"
1222                 "movd         %1, %%mm4 \n\t"
1223                 "punpcklbw %%mm7, %%mm5 \n\t"
1224                 "punpcklbw %%mm7, %%mm4 \n\t"
1225                 "pmullw    %%mm5, %%mm1 \n\t" // src[1, 0] * dx * (s - dy)
1226                 "pmullw    %%mm4, %%mm0 \n\t" // src[0, 0] * (s - dx) * (s - dy)
1227                 "paddw        %5, %%mm1 \n\t"
1228                 "paddw     %%mm3, %%mm2 \n\t"
1229                 "paddw     %%mm1, %%mm0 \n\t"
1230                 "paddw     %%mm2, %%mm0 \n\t"
1231
1232                 "psrlw        %6, %%mm0 \n\t"
1233                 "packuswb  %%mm0, %%mm0 \n\t"
1234                 "movd      %%mm0, %0    \n\t"
1235
1236                 : "=m"(dst[x + y * stride])
1237                 : "m"(src[0]), "m"(src[1]),
1238                   "m"(src[stride]), "m"(src[stride + 1]),
1239                   "m"(*r4), "m"(shift2)
1240             );
1241             src += stride;
1242         }
1243         src += 4 - h * stride;
1244     }
1245 }
1246 #endif /* HAVE_INLINE_ASM */
1247
1248 void ff_put_pixels16_sse2(uint8_t *block, const uint8_t *pixels,
1249                           ptrdiff_t line_size, int h);
1250 void ff_avg_pixels16_sse2(uint8_t *block, const uint8_t *pixels,
1251                           ptrdiff_t line_size, int h);
1252
1253 #if HAVE_INLINE_ASM
1254
1255 /* CAVS-specific */
1256 void ff_put_cavs_qpel8_mc00_mmxext(uint8_t *dst, uint8_t *src, ptrdiff_t stride)
1257 {
1258     put_pixels8_mmx(dst, src, stride, 8);
1259 }
1260
1261 void ff_avg_cavs_qpel8_mc00_mmxext(uint8_t *dst, uint8_t *src, ptrdiff_t stride)
1262 {
1263     avg_pixels8_mmx(dst, src, stride, 8);
1264 }
1265
1266 void ff_put_cavs_qpel16_mc00_mmxext(uint8_t *dst, uint8_t *src, ptrdiff_t stride)
1267 {
1268     put_pixels16_mmx(dst, src, stride, 16);
1269 }
1270
1271 void ff_avg_cavs_qpel16_mc00_mmxext(uint8_t *dst, uint8_t *src, ptrdiff_t stride)
1272 {
1273     avg_pixels16_mmx(dst, src, stride, 16);
1274 }
1275
1276 /* VC-1-specific */
1277 void ff_put_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src,
1278                                ptrdiff_t stride, int rnd)
1279 {
1280     put_pixels8_mmx(dst, src, stride, 8);
1281 }
1282
1283 static void vector_clipf_sse(float *dst, const float *src,
1284                              float min, float max, int len)
1285 {
1286     x86_reg i = (len - 16) * 4;
1287     __asm__ volatile (
1288         "movss          %3, %%xmm4      \n\t"
1289         "movss          %4, %%xmm5      \n\t"
1290         "shufps $0, %%xmm4, %%xmm4      \n\t"
1291         "shufps $0, %%xmm5, %%xmm5      \n\t"
1292         "1:                             \n\t"
1293         "movaps   (%2, %0), %%xmm0      \n\t" // 3/1 on intel
1294         "movaps 16(%2, %0), %%xmm1      \n\t"
1295         "movaps 32(%2, %0), %%xmm2      \n\t"
1296         "movaps 48(%2, %0), %%xmm3      \n\t"
1297         "maxps      %%xmm4, %%xmm0      \n\t"
1298         "maxps      %%xmm4, %%xmm1      \n\t"
1299         "maxps      %%xmm4, %%xmm2      \n\t"
1300         "maxps      %%xmm4, %%xmm3      \n\t"
1301         "minps      %%xmm5, %%xmm0      \n\t"
1302         "minps      %%xmm5, %%xmm1      \n\t"
1303         "minps      %%xmm5, %%xmm2      \n\t"
1304         "minps      %%xmm5, %%xmm3      \n\t"
1305         "movaps     %%xmm0,   (%1, %0)  \n\t"
1306         "movaps     %%xmm1, 16(%1, %0)  \n\t"
1307         "movaps     %%xmm2, 32(%1, %0)  \n\t"
1308         "movaps     %%xmm3, 48(%1, %0)  \n\t"
1309         "sub           $64, %0          \n\t"
1310         "jge            1b              \n\t"
1311         : "+&r"(i)
1312         : "r"(dst), "r"(src), "m"(min), "m"(max)
1313         : "memory"
1314     );
1315 }
1316
1317 #endif /* HAVE_INLINE_ASM */
1318
1319 int32_t ff_scalarproduct_int16_mmxext(const int16_t *v1, const int16_t *v2,
1320                                       int order);
1321 int32_t ff_scalarproduct_int16_sse2(const int16_t *v1, const int16_t *v2,
1322                                     int order);
1323 int32_t ff_scalarproduct_and_madd_int16_mmxext(int16_t *v1, const int16_t *v2,
1324                                                const int16_t *v3,
1325                                                int order, int mul);
1326 int32_t ff_scalarproduct_and_madd_int16_sse2(int16_t *v1, const int16_t *v2,
1327                                              const int16_t *v3,
1328                                              int order, int mul);
1329 int32_t ff_scalarproduct_and_madd_int16_ssse3(int16_t *v1, const int16_t *v2,
1330                                               const int16_t *v3,
1331                                               int order, int mul);
1332
1333 void ff_apply_window_int16_round_mmxext(int16_t *output, const int16_t *input,
1334                                         const int16_t *window, unsigned int len);
1335 void ff_apply_window_int16_round_sse2(int16_t *output, const int16_t *input,
1336                                       const int16_t *window, unsigned int len);
1337 void ff_apply_window_int16_mmxext(int16_t *output, const int16_t *input,
1338                                   const int16_t *window, unsigned int len);
1339 void ff_apply_window_int16_sse2(int16_t *output, const int16_t *input,
1340                                 const int16_t *window, unsigned int len);
1341 void ff_apply_window_int16_ssse3(int16_t *output, const int16_t *input,
1342                                  const int16_t *window, unsigned int len);
1343 void ff_apply_window_int16_ssse3_atom(int16_t *output, const int16_t *input,
1344                                       const int16_t *window, unsigned int len);
1345
1346 void ff_bswap32_buf_ssse3(uint32_t *dst, const uint32_t *src, int w);
1347 void ff_bswap32_buf_sse2(uint32_t *dst, const uint32_t *src, int w);
1348
1349 void ff_add_hfyu_median_prediction_mmxext(uint8_t *dst, const uint8_t *top,
1350                                           const uint8_t *diff, int w,
1351                                           int *left, int *left_top);
1352 int  ff_add_hfyu_left_prediction_ssse3(uint8_t *dst, const uint8_t *src,
1353                                        int w, int left);
1354 int  ff_add_hfyu_left_prediction_sse4(uint8_t *dst, const uint8_t *src,
1355                                       int w, int left);
1356
1357 void ff_vector_clip_int32_mmx     (int32_t *dst, const int32_t *src,
1358                                    int32_t min, int32_t max, unsigned int len);
1359 void ff_vector_clip_int32_sse2    (int32_t *dst, const int32_t *src,
1360                                    int32_t min, int32_t max, unsigned int len);
1361 void ff_vector_clip_int32_int_sse2(int32_t *dst, const int32_t *src,
1362                                    int32_t min, int32_t max, unsigned int len);
1363 void ff_vector_clip_int32_sse4    (int32_t *dst, const int32_t *src,
1364                                    int32_t min, int32_t max, unsigned int len);
1365
1366 #define SET_QPEL_FUNCS(PFX, IDX, SIZE, CPU, PREFIX)                          \
1367     do {                                                                     \
1368     c->PFX ## _pixels_tab[IDX][ 0] = PREFIX ## PFX ## SIZE ## _mc00_ ## CPU; \
1369     c->PFX ## _pixels_tab[IDX][ 1] = PREFIX ## PFX ## SIZE ## _mc10_ ## CPU; \
1370     c->PFX ## _pixels_tab[IDX][ 2] = PREFIX ## PFX ## SIZE ## _mc20_ ## CPU; \
1371     c->PFX ## _pixels_tab[IDX][ 3] = PREFIX ## PFX ## SIZE ## _mc30_ ## CPU; \
1372     c->PFX ## _pixels_tab[IDX][ 4] = PREFIX ## PFX ## SIZE ## _mc01_ ## CPU; \
1373     c->PFX ## _pixels_tab[IDX][ 5] = PREFIX ## PFX ## SIZE ## _mc11_ ## CPU; \
1374     c->PFX ## _pixels_tab[IDX][ 6] = PREFIX ## PFX ## SIZE ## _mc21_ ## CPU; \
1375     c->PFX ## _pixels_tab[IDX][ 7] = PREFIX ## PFX ## SIZE ## _mc31_ ## CPU; \
1376     c->PFX ## _pixels_tab[IDX][ 8] = PREFIX ## PFX ## SIZE ## _mc02_ ## CPU; \
1377     c->PFX ## _pixels_tab[IDX][ 9] = PREFIX ## PFX ## SIZE ## _mc12_ ## CPU; \
1378     c->PFX ## _pixels_tab[IDX][10] = PREFIX ## PFX ## SIZE ## _mc22_ ## CPU; \
1379     c->PFX ## _pixels_tab[IDX][11] = PREFIX ## PFX ## SIZE ## _mc32_ ## CPU; \
1380     c->PFX ## _pixels_tab[IDX][12] = PREFIX ## PFX ## SIZE ## _mc03_ ## CPU; \
1381     c->PFX ## _pixels_tab[IDX][13] = PREFIX ## PFX ## SIZE ## _mc13_ ## CPU; \
1382     c->PFX ## _pixels_tab[IDX][14] = PREFIX ## PFX ## SIZE ## _mc23_ ## CPU; \
1383     c->PFX ## _pixels_tab[IDX][15] = PREFIX ## PFX ## SIZE ## _mc33_ ## CPU; \
1384     } while (0)
1385
1386 #define SET_HPEL_FUNCS(PFX, IDX, SIZE, CPU)                                     \
1387     do {                                                                        \
1388         c->PFX ## _pixels_tab IDX [0] = PFX ## _pixels ## SIZE ## _     ## CPU; \
1389         c->PFX ## _pixels_tab IDX [1] = PFX ## _pixels ## SIZE ## _x2_  ## CPU; \
1390         c->PFX ## _pixels_tab IDX [2] = PFX ## _pixels ## SIZE ## _y2_  ## CPU; \
1391         c->PFX ## _pixels_tab IDX [3] = PFX ## _pixels ## SIZE ## _xy2_ ## CPU; \
1392     } while (0)
1393
1394 static av_cold void dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx,
1395                                      int mm_flags)
1396 {
1397     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
1398
1399 #if HAVE_INLINE_ASM
1400     c->put_pixels_clamped        = ff_put_pixels_clamped_mmx;
1401     c->put_signed_pixels_clamped = ff_put_signed_pixels_clamped_mmx;
1402     c->add_pixels_clamped        = ff_add_pixels_clamped_mmx;
1403
1404     if (!high_bit_depth) {
1405         c->clear_block  = clear_block_mmx;
1406         c->clear_blocks = clear_blocks_mmx;
1407         c->draw_edges   = draw_edges_mmx;
1408
1409         SET_HPEL_FUNCS(put,        [0], 16, mmx);
1410         SET_HPEL_FUNCS(put_no_rnd, [0], 16, mmx);
1411         SET_HPEL_FUNCS(avg,        [0], 16, mmx);
1412         SET_HPEL_FUNCS(avg_no_rnd,    , 16, mmx);
1413         SET_HPEL_FUNCS(put,        [1],  8, mmx);
1414         SET_HPEL_FUNCS(put_no_rnd, [1],  8, mmx);
1415         SET_HPEL_FUNCS(avg,        [1],  8, mmx);
1416
1417         switch (avctx->idct_algo) {
1418         case FF_IDCT_AUTO:
1419         case FF_IDCT_SIMPLEMMX:
1420             c->idct_put              = ff_simple_idct_put_mmx;
1421             c->idct_add              = ff_simple_idct_add_mmx;
1422             c->idct                  = ff_simple_idct_mmx;
1423             c->idct_permutation_type = FF_SIMPLE_IDCT_PERM;
1424             break;
1425         case FF_IDCT_XVIDMMX:
1426             c->idct_put              = ff_idct_xvid_mmx_put;
1427             c->idct_add              = ff_idct_xvid_mmx_add;
1428             c->idct                  = ff_idct_xvid_mmx;
1429             break;
1430         }
1431     }
1432
1433     c->gmc = gmc_mmx;
1434
1435     c->add_bytes = add_bytes_mmx;
1436 #endif /* HAVE_INLINE_ASM */
1437
1438 #if HAVE_YASM
1439     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
1440         c->h263_v_loop_filter = ff_h263_v_loop_filter_mmx;
1441         c->h263_h_loop_filter = ff_h263_h_loop_filter_mmx;
1442     }
1443
1444     c->vector_clip_int32 = ff_vector_clip_int32_mmx;
1445 #endif
1446
1447 }
1448
1449 static av_cold void dsputil_init_mmxext(DSPContext *c, AVCodecContext *avctx,
1450                                         int mm_flags)
1451 {
1452     const int bit_depth      = avctx->bits_per_raw_sample;
1453     const int high_bit_depth = bit_depth > 8;
1454
1455 #if HAVE_YASM
1456     SET_QPEL_FUNCS(avg_qpel,        0, 16, mmxext, );
1457     SET_QPEL_FUNCS(avg_qpel,        1,  8, mmxext, );
1458
1459     SET_QPEL_FUNCS(put_qpel,        0, 16, mmxext, );
1460     SET_QPEL_FUNCS(put_qpel,        1,  8, mmxext, );
1461     SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, mmxext, );
1462     SET_QPEL_FUNCS(put_no_rnd_qpel, 1,  8, mmxext, );
1463
1464     if (!high_bit_depth) {
1465         c->put_pixels_tab[0][1] = ff_put_pixels16_x2_mmxext;
1466         c->put_pixels_tab[0][2] = ff_put_pixels16_y2_mmxext;
1467
1468         c->avg_pixels_tab[0][0] = ff_avg_pixels16_mmxext;
1469         c->avg_pixels_tab[0][1] = ff_avg_pixels16_x2_mmxext;
1470         c->avg_pixels_tab[0][2] = ff_avg_pixels16_y2_mmxext;
1471
1472         c->put_pixels_tab[1][1] = ff_put_pixels8_x2_mmxext;
1473         c->put_pixels_tab[1][2] = ff_put_pixels8_y2_mmxext;
1474
1475         c->avg_pixels_tab[1][0] = ff_avg_pixels8_mmxext;
1476         c->avg_pixels_tab[1][1] = ff_avg_pixels8_x2_mmxext;
1477         c->avg_pixels_tab[1][2] = ff_avg_pixels8_y2_mmxext;
1478     }
1479
1480     if (!(avctx->flags & CODEC_FLAG_BITEXACT)) {
1481         if (!high_bit_depth) {
1482             c->put_no_rnd_pixels_tab[0][1] = ff_put_no_rnd_pixels16_x2_mmxext;
1483             c->put_no_rnd_pixels_tab[0][2] = ff_put_no_rnd_pixels16_y2_mmxext;
1484             c->put_no_rnd_pixels_tab[1][1] = ff_put_no_rnd_pixels8_x2_mmxext;
1485             c->put_no_rnd_pixels_tab[1][2] = ff_put_no_rnd_pixels8_y2_mmxext;
1486
1487             c->avg_pixels_tab[0][3] = ff_avg_pixels16_xy2_mmxext;
1488             c->avg_pixels_tab[1][3] = ff_avg_pixels8_xy2_mmxext;
1489         }
1490     }
1491 #endif /* HAVE_YASM */
1492
1493 #if HAVE_INLINE_ASM
1494     if (!high_bit_depth && avctx->idct_algo == FF_IDCT_XVIDMMX) {
1495         c->idct_put = ff_idct_xvid_mmxext_put;
1496         c->idct_add = ff_idct_xvid_mmxext_add;
1497         c->idct     = ff_idct_xvid_mmxext;
1498     }
1499 #endif /* HAVE_INLINE_ASM */
1500
1501 #if HAVE_MMXEXT_EXTERNAL
1502     if (CONFIG_VP3_DECODER && (avctx->codec_id == AV_CODEC_ID_VP3 ||
1503                                avctx->codec_id == AV_CODEC_ID_THEORA)) {
1504         c->put_no_rnd_pixels_tab[1][1] = ff_put_no_rnd_pixels8_x2_exact_mmxext;
1505         c->put_no_rnd_pixels_tab[1][2] = ff_put_no_rnd_pixels8_y2_exact_mmxext;
1506     }
1507
1508     /* slower than cmov version on AMD */
1509     if (!(mm_flags & AV_CPU_FLAG_3DNOW))
1510         c->add_hfyu_median_prediction = ff_add_hfyu_median_prediction_mmxext;
1511
1512     c->scalarproduct_int16          = ff_scalarproduct_int16_mmxext;
1513     c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_mmxext;
1514
1515     if (avctx->flags & CODEC_FLAG_BITEXACT) {
1516         c->apply_window_int16 = ff_apply_window_int16_mmxext;
1517     } else {
1518         c->apply_window_int16 = ff_apply_window_int16_round_mmxext;
1519     }
1520 #endif /* HAVE_MMXEXT_EXTERNAL */
1521 }
1522
1523 static av_cold void dsputil_init_3dnow(DSPContext *c, AVCodecContext *avctx,
1524                                        int mm_flags)
1525 {
1526     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
1527
1528 #if HAVE_YASM
1529     if (!high_bit_depth) {
1530         c->put_pixels_tab[0][1] = ff_put_pixels16_x2_3dnow;
1531         c->put_pixels_tab[0][2] = ff_put_pixels16_y2_3dnow;
1532
1533         c->avg_pixels_tab[0][0] = ff_avg_pixels16_3dnow;
1534         c->avg_pixels_tab[0][1] = ff_avg_pixels16_x2_3dnow;
1535         c->avg_pixels_tab[0][2] = ff_avg_pixels16_y2_3dnow;
1536
1537         c->put_pixels_tab[1][1] = ff_put_pixels8_x2_3dnow;
1538         c->put_pixels_tab[1][2] = ff_put_pixels8_y2_3dnow;
1539
1540         c->avg_pixels_tab[1][0] = ff_avg_pixels8_3dnow;
1541         c->avg_pixels_tab[1][1] = ff_avg_pixels8_x2_3dnow;
1542         c->avg_pixels_tab[1][2] = ff_avg_pixels8_y2_3dnow;
1543
1544         if (!(avctx->flags & CODEC_FLAG_BITEXACT)){
1545             c->put_no_rnd_pixels_tab[0][1] = ff_put_no_rnd_pixels16_x2_3dnow;
1546             c->put_no_rnd_pixels_tab[0][2] = ff_put_no_rnd_pixels16_y2_3dnow;
1547             c->put_no_rnd_pixels_tab[1][1] = ff_put_no_rnd_pixels8_x2_3dnow;
1548             c->put_no_rnd_pixels_tab[1][2] = ff_put_no_rnd_pixels8_y2_3dnow;
1549
1550             c->avg_pixels_tab[0][3] = ff_avg_pixels16_xy2_3dnow;
1551             c->avg_pixels_tab[1][3] = ff_avg_pixels8_xy2_3dnow;
1552         }
1553     }
1554
1555     if (CONFIG_VP3_DECODER && (avctx->codec_id == AV_CODEC_ID_VP3 ||
1556                                avctx->codec_id == AV_CODEC_ID_THEORA)) {
1557         c->put_no_rnd_pixels_tab[1][1] = ff_put_no_rnd_pixels8_x2_exact_3dnow;
1558         c->put_no_rnd_pixels_tab[1][2] = ff_put_no_rnd_pixels8_y2_exact_3dnow;
1559     }
1560 #endif /* HAVE_YASM */
1561 }
1562
1563 static av_cold void dsputil_init_sse(DSPContext *c, AVCodecContext *avctx,
1564                                      int mm_flags)
1565 {
1566     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
1567
1568 #if HAVE_INLINE_ASM
1569     if (!high_bit_depth) {
1570         if (!(CONFIG_MPEG_XVMC_DECODER && avctx->xvmc_acceleration > 1)) {
1571             /* XvMCCreateBlocks() may not allocate 16-byte aligned blocks */
1572             c->clear_block  = clear_block_sse;
1573             c->clear_blocks = clear_blocks_sse;
1574         }
1575     }
1576
1577     c->vector_clipf = vector_clipf_sse;
1578 #endif /* HAVE_INLINE_ASM */
1579 }
1580
1581 static av_cold void dsputil_init_sse2(DSPContext *c, AVCodecContext *avctx,
1582                                       int mm_flags)
1583 {
1584     const int bit_depth      = avctx->bits_per_raw_sample;
1585     const int high_bit_depth = bit_depth > 8;
1586
1587 #if HAVE_SSE2_INLINE
1588     if (!high_bit_depth && avctx->idct_algo == FF_IDCT_XVIDMMX) {
1589         c->idct_put              = ff_idct_xvid_sse2_put;
1590         c->idct_add              = ff_idct_xvid_sse2_add;
1591         c->idct                  = ff_idct_xvid_sse2;
1592         c->idct_permutation_type = FF_SSE2_IDCT_PERM;
1593     }
1594 #endif /* HAVE_SSE2_INLINE */
1595
1596 #if HAVE_SSE2_EXTERNAL
1597     if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {
1598         // these functions are slower than mmx on AMD, but faster on Intel
1599         if (!high_bit_depth) {
1600             c->put_pixels_tab[0][0]        = ff_put_pixels16_sse2;
1601             c->put_no_rnd_pixels_tab[0][0] = ff_put_pixels16_sse2;
1602             c->avg_pixels_tab[0][0]        = ff_avg_pixels16_sse2;
1603         }
1604     }
1605
1606     c->scalarproduct_int16          = ff_scalarproduct_int16_sse2;
1607     c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_sse2;
1608     if (mm_flags & AV_CPU_FLAG_ATOM) {
1609         c->vector_clip_int32 = ff_vector_clip_int32_int_sse2;
1610     } else {
1611         c->vector_clip_int32 = ff_vector_clip_int32_sse2;
1612     }
1613     if (avctx->flags & CODEC_FLAG_BITEXACT) {
1614         c->apply_window_int16 = ff_apply_window_int16_sse2;
1615     } else if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {
1616         c->apply_window_int16 = ff_apply_window_int16_round_sse2;
1617     }
1618     c->bswap_buf = ff_bswap32_buf_sse2;
1619 #endif /* HAVE_SSE2_EXTERNAL */
1620 }
1621
1622 static av_cold void dsputil_init_ssse3(DSPContext *c, AVCodecContext *avctx,
1623                                        int mm_flags)
1624 {
1625 #if HAVE_SSSE3_EXTERNAL
1626     c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_ssse3;
1627     if (mm_flags & AV_CPU_FLAG_SSE4) // not really sse4, just slow on Conroe
1628         c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_sse4;
1629
1630     if (mm_flags & AV_CPU_FLAG_ATOM)
1631         c->apply_window_int16 = ff_apply_window_int16_ssse3_atom;
1632     else
1633         c->apply_window_int16 = ff_apply_window_int16_ssse3;
1634     if (!(mm_flags & (AV_CPU_FLAG_SSE42|AV_CPU_FLAG_3DNOW))) // cachesplit
1635         c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_ssse3;
1636     c->bswap_buf = ff_bswap32_buf_ssse3;
1637 #endif /* HAVE_SSSE3_EXTERNAL */
1638 }
1639
1640 static av_cold void dsputil_init_sse4(DSPContext *c, AVCodecContext *avctx,
1641                                       int mm_flags)
1642 {
1643 #if HAVE_SSE4_EXTERNAL
1644     c->vector_clip_int32 = ff_vector_clip_int32_sse4;
1645 #endif /* HAVE_SSE4_EXTERNAL */
1646 }
1647
1648 av_cold void ff_dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx)
1649 {
1650     int mm_flags = av_get_cpu_flags();
1651
1652 #if HAVE_7REGS && HAVE_INLINE_ASM
1653     if (mm_flags & AV_CPU_FLAG_CMOV)
1654         c->add_hfyu_median_prediction = add_hfyu_median_prediction_cmov;
1655 #endif
1656
1657     if (mm_flags & AV_CPU_FLAG_MMX)
1658         dsputil_init_mmx(c, avctx, mm_flags);
1659
1660     if (mm_flags & AV_CPU_FLAG_MMXEXT)
1661         dsputil_init_mmxext(c, avctx, mm_flags);
1662
1663     if (mm_flags & AV_CPU_FLAG_3DNOW)
1664         dsputil_init_3dnow(c, avctx, mm_flags);
1665
1666     if (mm_flags & AV_CPU_FLAG_SSE)
1667         dsputil_init_sse(c, avctx, mm_flags);
1668
1669     if (mm_flags & AV_CPU_FLAG_SSE2)
1670         dsputil_init_sse2(c, avctx, mm_flags);
1671
1672     if (mm_flags & AV_CPU_FLAG_SSSE3)
1673         dsputil_init_ssse3(c, avctx, mm_flags);
1674
1675     if (mm_flags & AV_CPU_FLAG_SSE4)
1676         dsputil_init_sse4(c, avctx, mm_flags);
1677
1678     if (CONFIG_ENCODERS)
1679         ff_dsputilenc_init_mmx(c, avctx);
1680 }