git.sesse.net Git - ffmpeg/blob - libavcodec/x86/h264dsp_mmx.c

   1 /*
   2  * Copyright (c) 2004-2005 Michael Niedermayer, Loren Merritt
   3  *
   4  * This file is part of FFmpeg.
   5  *
   6  * FFmpeg is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2.1 of the License, or (at your option) any later version.
  10  *
  11  * FFmpeg is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with FFmpeg; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  19  */
  20
  21 #include "libavutil/cpu.h"
  22 #include "libavutil/x86_cpu.h"
  23 #include "libavcodec/h264dsp.h"
  24 #include "dsputil_mmx.h"
  25
  26 DECLARE_ALIGNED(8, static const uint64_t, ff_pb_3_1  ) = 0x0103010301030103ULL;
  27
  28 /***********************************/
  29 /* IDCT */
  30
  31 void ff_h264_idct_add_mmx     (uint8_t *dst, int16_t *block, int stride);
  32 void ff_h264_idct8_add_mmx    (uint8_t *dst, int16_t *block, int stride);
  33 void ff_h264_idct8_add_sse2   (uint8_t *dst, int16_t *block, int stride);
  34 void ff_h264_idct_dc_add_mmx2 (uint8_t *dst, int16_t *block, int stride);
  35 void ff_h264_idct8_dc_add_mmx2(uint8_t *dst, int16_t *block, int stride);
  36
  37 void ff_h264_idct_add16_mmx      (uint8_t *dst, const int *block_offset,
  38                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  39 void ff_h264_idct8_add4_mmx      (uint8_t *dst, const int *block_offset,
  40                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  41 void ff_h264_idct_add16_mmx2     (uint8_t *dst, const int *block_offset,
  42                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  43 void ff_h264_idct_add16intra_mmx (uint8_t *dst, const int *block_offset,
  44                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  45 void ff_h264_idct_add16intra_mmx2(uint8_t *dst, const int *block_offset,
  46                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  47 void ff_h264_idct8_add4_mmx2     (uint8_t *dst, const int *block_offset,
  48                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  49 void ff_h264_idct8_add4_sse2     (uint8_t *dst, const int *block_offset,
  50                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  51 void ff_h264_idct_add8_mmx       (uint8_t **dest, const int *block_offset,
  52                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  53 void ff_h264_idct_add8_mmx2      (uint8_t **dest, const int *block_offset,
  54                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  55
  56 void ff_h264_idct_add16_sse2     (uint8_t *dst, const int *block_offset, DCTELEM *block,
  57                                   int stride, const uint8_t nnzc[6*8]);
  58 void ff_h264_idct_add16intra_sse2(uint8_t *dst, const int *block_offset, DCTELEM *block,
  59                                   int stride, const uint8_t nnzc[6*8]);
  60 void ff_h264_idct_add8_sse2      (uint8_t **dest, const int *block_offset, DCTELEM *block,
  61                                   int stride, const uint8_t nnzc[6*8]);
  62
  63 /***********************************/
  64 /* deblocking */
  65
  66 #define h264_loop_filter_strength_iteration_mmx2(bS, nz, ref, mv, bidir, edges, step, mask_mv, dir, d_idx, mask_dir) \
  67     do { \
  68         x86_reg b_idx; \
  69         mask_mv <<= 3; \
  70         for( b_idx=0; b_idx<edges; b_idx+=step ) { \
  71             if (!mask_dir) \
  72             __asm__ volatile( \
  73                     "pxor %%mm0, %%mm0 \n\t" \
  74                     :: \
  75             ); \
  76             if(!(mask_mv & b_idx)) { \
  77                 if(bidir) { \
  78                     __asm__ volatile( \
  79                         "movd         %a3(%0,%2), %%mm2 \n" \
  80                         "punpckldq    %a4(%0,%2), %%mm2 \n" /* { ref0[bn], ref1[bn] } */ \
  81                         "pshufw $0x44, 12(%0,%2), %%mm0 \n" /* { ref0[b], ref0[b] } */ \
  82                         "pshufw $0x44, 52(%0,%2), %%mm1 \n" /* { ref1[b], ref1[b] } */ \
  83                         "pshufw $0x4E, %%mm2, %%mm3 \n" \
  84                         "psubb         %%mm2, %%mm0 \n" /* { ref0[b]!=ref0[bn], ref0[b]!=ref1[bn] } */ \
  85                         "psubb         %%mm3, %%mm1 \n" /* { ref1[b]!=ref1[bn], ref1[b]!=ref0[bn] } */ \
  86  \
  87                         "por           %%mm1, %%mm0 \n" \
  88                         "movq   %a5(%1,%2,4), %%mm1 \n" \
  89                         "movq   %a6(%1,%2,4), %%mm2 \n" \
  90                         "movq          %%mm1, %%mm3 \n" \
  91                         "movq          %%mm2, %%mm4 \n" \
  92                         "psubw   48(%1,%2,4), %%mm1 \n" \
  93                         "psubw   56(%1,%2,4), %%mm2 \n" \
  94                         "psubw  208(%1,%2,4), %%mm3 \n" \
  95                         "psubw  216(%1,%2,4), %%mm4 \n" \
  96                         "packsswb      %%mm2, %%mm1 \n" \
  97                         "packsswb      %%mm4, %%mm3 \n" \
  98                         "paddb         %%mm6, %%mm1 \n" \
  99                         "paddb         %%mm6, %%mm3 \n" \
 100                         "psubusb       %%mm5, %%mm1 \n" /* abs(mv[b] - mv[bn]) >= limit */ \
 101                         "psubusb       %%mm5, %%mm3 \n" \
 102                         "packsswb      %%mm3, %%mm1 \n" \
 103  \
 104                         "por           %%mm1, %%mm0 \n" \
 105                         "movq   %a7(%1,%2,4), %%mm1 \n" \
 106                         "movq   %a8(%1,%2,4), %%mm2 \n" \
 107                         "movq          %%mm1, %%mm3 \n" \
 108                         "movq          %%mm2, %%mm4 \n" \
 109                         "psubw   48(%1,%2,4), %%mm1 \n" \
 110                         "psubw   56(%1,%2,4), %%mm2 \n" \
 111                         "psubw  208(%1,%2,4), %%mm3 \n" \
 112                         "psubw  216(%1,%2,4), %%mm4 \n" \
 113                         "packsswb      %%mm2, %%mm1 \n" \
 114                         "packsswb      %%mm4, %%mm3 \n" \
 115                         "paddb         %%mm6, %%mm1 \n" \
 116                         "paddb         %%mm6, %%mm3 \n" \
 117                         "psubusb       %%mm5, %%mm1 \n" /* abs(mv[b] - mv[bn]) >= limit */ \
 118                         "psubusb       %%mm5, %%mm3 \n" \
 119                         "packsswb      %%mm3, %%mm1 \n" \
 120  \
 121                         "pshufw $0x4E, %%mm1, %%mm1 \n" \
 122                         "por           %%mm1, %%mm0 \n" \
 123                         "pshufw $0x4E, %%mm0, %%mm1 \n" \
 124                         "pminub        %%mm1, %%mm0 \n" \
 125                         ::"r"(ref), \
 126                           "r"(mv), \
 127                           "r"(b_idx), \
 128                           "i"(d_idx+12), \
 129                           "i"(d_idx+52), \
 130                           "i"(d_idx*4+48), \
 131                           "i"(d_idx*4+56), \
 132                           "i"(d_idx*4+208), \
 133                           "i"(d_idx*4+216) \
 134                     ); \
 135                 } else { \
 136                     __asm__ volatile( \
 137                         "movd   12(%0,%2), %%mm0 \n" \
 138                         "psubb %a3(%0,%2), %%mm0 \n" /* ref[b] != ref[bn] */ \
 139                         "movq   48(%1,%2,4), %%mm1 \n" \
 140                         "movq   56(%1,%2,4), %%mm2 \n" \
 141                         "psubw %a4(%1,%2,4), %%mm1 \n" \
 142                         "psubw %a5(%1,%2,4), %%mm2 \n" \
 143                         "packsswb   %%mm2, %%mm1 \n" \
 144                         "paddb      %%mm6, %%mm1 \n" \
 145                         "psubusb    %%mm5, %%mm1 \n" /* abs(mv[b] - mv[bn]) >= limit */ \
 146                         "packsswb   %%mm1, %%mm1 \n" \
 147                         "por        %%mm1, %%mm0 \n" \
 148                         ::"r"(ref), \
 149                           "r"(mv), \
 150                           "r"(b_idx), \
 151                           "i"(d_idx+12), \
 152                           "i"(d_idx*4+48), \
 153                           "i"(d_idx*4+56) \
 154                     ); \
 155                 } \
 156             } \
 157             __asm__ volatile( \
 158                 "movd 12(%0,%1), %%mm1 \n" \
 159                 "por %a2(%0,%1), %%mm1 \n" /* nnz[b] || nnz[bn] */ \
 160                 ::"r"(nnz), \
 161                   "r"(b_idx), \
 162                   "i"(d_idx+12) \
 163             ); \
 164             __asm__ volatile( \
 165                 "pminub    %%mm7, %%mm1 \n" \
 166                 "pminub    %%mm7, %%mm0 \n" \
 167                 "psllw        $1, %%mm1 \n" \
 168                 "pxor      %%mm2, %%mm2 \n" \
 169                 "pmaxub    %%mm0, %%mm1 \n" \
 170                 "punpcklbw %%mm2, %%mm1 \n" \
 171                 "movq      %%mm1, %a1(%0,%2) \n" \
 172                 ::"r"(bS), \
 173                   "i"(32*dir), \
 174                   "r"(b_idx) \
 175                 :"memory" \
 176             ); \
 177         } \
 178     } while (0)
 179
 180 static void h264_loop_filter_strength_mmx2( int16_t bS[2][4][4], uint8_t nnz[40], int8_t ref[2][40], int16_t mv[2][40][2],
 181                                             int bidir, int edges, int step, int mask_mv0, int mask_mv1, int field ) {
 182     __asm__ volatile(
 183         "movq %0, %%mm7 \n"
 184         "movq %1, %%mm6 \n"
 185         ::"m"(ff_pb_1), "m"(ff_pb_3)
 186     );
 187     if(field)
 188         __asm__ volatile(
 189             "movq %0, %%mm6 \n"
 190             ::"m"(ff_pb_3_1)
 191         );
 192     __asm__ volatile(
 193         "movq  %%mm6, %%mm5 \n"
 194         "paddb %%mm5, %%mm5 \n"
 195     :);
 196
 197     // could do a special case for dir==0 && edges==1, but it only reduces the
 198     // average filter time by 1.2%
 199     step  <<= 3;
 200     edges <<= 3;
 201     h264_loop_filter_strength_iteration_mmx2(bS, nnz, ref, mv, bidir, edges, step, mask_mv1, 1, -8,  0);
 202     h264_loop_filter_strength_iteration_mmx2(bS, nnz, ref, mv, bidir,    32,    8, mask_mv0, 0, -1, -1);
 203
 204     __asm__ volatile(
 205         "movq   (%0), %%mm0 \n\t"
 206         "movq  8(%0), %%mm1 \n\t"
 207         "movq 16(%0), %%mm2 \n\t"
 208         "movq 24(%0), %%mm3 \n\t"
 209         TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4)
 210         "movq %%mm0,   (%0) \n\t"
 211         "movq %%mm3,  8(%0) \n\t"
 212         "movq %%mm4, 16(%0) \n\t"
 213         "movq %%mm2, 24(%0) \n\t"
 214         ::"r"(bS[0])
 215         :"memory"
 216     );
 217 }
 218
 219 #define LF_FUNC(DIR, TYPE, OPT) \
 220 void ff_x264_deblock_ ## DIR ## _ ## TYPE ## _ ## OPT (uint8_t *pix, int stride, \
 221                                                int alpha, int beta, int8_t *tc0);
 222 #define LF_IFUNC(DIR, TYPE, OPT) \
 223 void ff_x264_deblock_ ## DIR ## _ ## TYPE ## _ ## OPT (uint8_t *pix, int stride, \
 224                                                int alpha, int beta);
 225
 226 LF_FUNC (h,  chroma,       mmxext)
 227 LF_IFUNC(h,  chroma_intra, mmxext)
 228 LF_FUNC (v,  chroma,       mmxext)
 229 LF_IFUNC(v,  chroma_intra, mmxext)
 230
 231 LF_FUNC (h,  luma,         mmxext)
 232 LF_IFUNC(h,  luma_intra,   mmxext)
 233 #if HAVE_YASM && ARCH_X86_32
 234 LF_FUNC (v8, luma,         mmxext)
 235 static void ff_x264_deblock_v_luma_mmxext(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0)
 236 {
 237     if((tc0[0] & tc0[1]) >= 0)
 238         ff_x264_deblock_v8_luma_mmxext(pix+0, stride, alpha, beta, tc0);
 239     if((tc0[2] & tc0[3]) >= 0)
 240         ff_x264_deblock_v8_luma_mmxext(pix+8, stride, alpha, beta, tc0+2);
 241 }
 242 LF_IFUNC(v8, luma_intra,   mmxext)
 243 static void ff_x264_deblock_v_luma_intra_mmxext(uint8_t *pix, int stride, int alpha, int beta)
 244 {
 245     ff_x264_deblock_v8_luma_intra_mmxext(pix+0, stride, alpha, beta);
 246     ff_x264_deblock_v8_luma_intra_mmxext(pix+8, stride, alpha, beta);
 247 }
 248 #endif
 249
 250 LF_FUNC (h,  luma,         sse2)
 251 LF_IFUNC(h,  luma_intra,   sse2)
 252 LF_FUNC (v,  luma,         sse2)
 253 LF_IFUNC(v,  luma_intra,   sse2)
 254
 255 /***********************************/
 256 /* weighted prediction */
 257
 258 #define H264_WEIGHT(W, H, OPT) \
 259 void ff_h264_weight_ ## W ## x ## H ## _ ## OPT(uint8_t *dst, \
 260     int stride, int log2_denom, int weight, int offset);
 261
 262 #define H264_BIWEIGHT(W, H, OPT) \
 263 void ff_h264_biweight_ ## W ## x ## H ## _ ## OPT(uint8_t *dst, \
 264     uint8_t *src, int stride, int log2_denom, int weightd, \
 265     int weights, int offset);
 266
 267 #define H264_BIWEIGHT_MMX(W,H) \
 268 H264_WEIGHT  (W, H, mmx2) \
 269 H264_BIWEIGHT(W, H, mmx2)
 270
 271 #define H264_BIWEIGHT_MMX_SSE(W,H) \
 272 H264_BIWEIGHT_MMX(W, H) \
 273 H264_WEIGHT      (W, H, sse2) \
 274 H264_BIWEIGHT    (W, H, sse2) \
 275 H264_BIWEIGHT    (W, H, ssse3)
 276
 277 H264_BIWEIGHT_MMX_SSE(16, 16)
 278 H264_BIWEIGHT_MMX_SSE(16,  8)
 279 H264_BIWEIGHT_MMX_SSE( 8, 16)
 280 H264_BIWEIGHT_MMX_SSE( 8,  8)
 281 H264_BIWEIGHT_MMX_SSE( 8,  4)
 282 H264_BIWEIGHT_MMX    ( 4,  8)
 283 H264_BIWEIGHT_MMX    ( 4,  4)
 284 H264_BIWEIGHT_MMX    ( 4,  2)
 285
 286 void ff_h264dsp_init_x86(H264DSPContext *c)
 287 {
 288     int mm_flags = av_get_cpu_flags();
 289
 290     if (mm_flags & AV_CPU_FLAG_MMX2) {
 291         c->h264_loop_filter_strength= h264_loop_filter_strength_mmx2;
 292     }
 293 #if HAVE_YASM
 294     if (mm_flags & AV_CPU_FLAG_MMX) {
 295         c->h264_idct_dc_add=
 296         c->h264_idct_add= ff_h264_idct_add_mmx;
 297         c->h264_idct8_dc_add=
 298         c->h264_idct8_add= ff_h264_idct8_add_mmx;
 299
 300         c->h264_idct_add16     = ff_h264_idct_add16_mmx;
 301         c->h264_idct8_add4     = ff_h264_idct8_add4_mmx;
 302         c->h264_idct_add8      = ff_h264_idct_add8_mmx;
 303         c->h264_idct_add16intra= ff_h264_idct_add16intra_mmx;
 304
 305         if (mm_flags & AV_CPU_FLAG_MMX2) {
 306             c->h264_idct_dc_add= ff_h264_idct_dc_add_mmx2;
 307             c->h264_idct8_dc_add= ff_h264_idct8_dc_add_mmx2;
 308             c->h264_idct_add16     = ff_h264_idct_add16_mmx2;
 309             c->h264_idct8_add4     = ff_h264_idct8_add4_mmx2;
 310             c->h264_idct_add8      = ff_h264_idct_add8_mmx2;
 311             c->h264_idct_add16intra= ff_h264_idct_add16intra_mmx2;
 312
 313             c->h264_v_loop_filter_chroma= ff_x264_deblock_v_chroma_mmxext;
 314             c->h264_h_loop_filter_chroma= ff_x264_deblock_h_chroma_mmxext;
 315             c->h264_v_loop_filter_chroma_intra= ff_x264_deblock_v_chroma_intra_mmxext;
 316             c->h264_h_loop_filter_chroma_intra= ff_x264_deblock_h_chroma_intra_mmxext;
 317 #if ARCH_X86_32
 318             c->h264_v_loop_filter_luma= ff_x264_deblock_v_luma_mmxext;
 319             c->h264_h_loop_filter_luma= ff_x264_deblock_h_luma_mmxext;
 320             c->h264_v_loop_filter_luma_intra = ff_x264_deblock_v_luma_intra_mmxext;
 321             c->h264_h_loop_filter_luma_intra = ff_x264_deblock_h_luma_intra_mmxext;
 322 #endif
 323             c->weight_h264_pixels_tab[0]= ff_h264_weight_16x16_mmx2;
 324             c->weight_h264_pixels_tab[1]= ff_h264_weight_16x8_mmx2;
 325             c->weight_h264_pixels_tab[2]= ff_h264_weight_8x16_mmx2;
 326             c->weight_h264_pixels_tab[3]= ff_h264_weight_8x8_mmx2;
 327             c->weight_h264_pixels_tab[4]= ff_h264_weight_8x4_mmx2;
 328             c->weight_h264_pixels_tab[5]= ff_h264_weight_4x8_mmx2;
 329             c->weight_h264_pixels_tab[6]= ff_h264_weight_4x4_mmx2;
 330             c->weight_h264_pixels_tab[7]= ff_h264_weight_4x2_mmx2;
 331
 332             c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_mmx2;
 333             c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_mmx2;
 334             c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_mmx2;
 335             c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_mmx2;
 336             c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_mmx2;
 337             c->biweight_h264_pixels_tab[5]= ff_h264_biweight_4x8_mmx2;
 338             c->biweight_h264_pixels_tab[6]= ff_h264_biweight_4x4_mmx2;
 339             c->biweight_h264_pixels_tab[7]= ff_h264_biweight_4x2_mmx2;
 340
 341             if (mm_flags&AV_CPU_FLAG_SSE2) {
 342                 c->h264_idct8_add = ff_h264_idct8_add_sse2;
 343                 c->h264_idct8_add4= ff_h264_idct8_add4_sse2;
 344
 345                 c->weight_h264_pixels_tab[0]= ff_h264_weight_16x16_sse2;
 346                 c->weight_h264_pixels_tab[1]= ff_h264_weight_16x8_sse2;
 347                 c->weight_h264_pixels_tab[2]= ff_h264_weight_8x16_sse2;
 348                 c->weight_h264_pixels_tab[3]= ff_h264_weight_8x8_sse2;
 349                 c->weight_h264_pixels_tab[4]= ff_h264_weight_8x4_sse2;
 350
 351                 c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_sse2;
 352                 c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_sse2;
 353                 c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_sse2;
 354                 c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_sse2;
 355                 c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_sse2;
 356
 357 #if HAVE_ALIGNED_STACK
 358                 c->h264_v_loop_filter_luma = ff_x264_deblock_v_luma_sse2;
 359                 c->h264_h_loop_filter_luma = ff_x264_deblock_h_luma_sse2;
 360                 c->h264_v_loop_filter_luma_intra = ff_x264_deblock_v_luma_intra_sse2;
 361                 c->h264_h_loop_filter_luma_intra = ff_x264_deblock_h_luma_intra_sse2;
 362 #endif
 363
 364                 c->h264_idct_add16 = ff_h264_idct_add16_sse2;
 365                 c->h264_idct_add8  = ff_h264_idct_add8_sse2;
 366                 c->h264_idct_add16intra = ff_h264_idct_add16intra_sse2;
 367             }
 368             if (mm_flags&AV_CPU_FLAG_SSSE3) {
 369                 c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_ssse3;
 370                 c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_ssse3;
 371                 c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_ssse3;
 372                 c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_ssse3;
 373                 c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_ssse3;
 374             }
 375         }
 376     }
 377 #endif
 378 }