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 Libav.
   5  *
   6  * Libav 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  * Libav 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 Libav; 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 void ff_h264_luma_dc_dequant_idct_mmx (DCTELEM *output, DCTELEM *input, int qmul);
  63 void ff_h264_luma_dc_dequant_idct_sse2(DCTELEM *output, DCTELEM *input, int qmul);
  64
  65 /***********************************/
  66 /* deblocking */
  67
  68 #define h264_loop_filter_strength_iteration_mmx2(bS, nz, ref, mv, bidir, edges, step, mask_mv, dir, d_idx, mask_dir) \
  69     do { \
  70         x86_reg b_idx; \
  71         mask_mv <<= 3; \
  72         for( b_idx=0; b_idx<edges; b_idx+=step ) { \
  73             if (!mask_dir) \
  74             __asm__ volatile( \
  75                     "pxor %%mm0, %%mm0 \n\t" \
  76                     :: \
  77             ); \
  78             if(!(mask_mv & b_idx)) { \
  79                 if(bidir) { \
  80                     __asm__ volatile( \
  81                         "movd         %a3(%0,%2), %%mm2 \n" \
  82                         "punpckldq    %a4(%0,%2), %%mm2 \n" /* { ref0[bn], ref1[bn] } */ \
  83                         "pshufw $0x44, 12(%0,%2), %%mm0 \n" /* { ref0[b], ref0[b] } */ \
  84                         "pshufw $0x44, 52(%0,%2), %%mm1 \n" /* { ref1[b], ref1[b] } */ \
  85                         "pshufw $0x4E, %%mm2, %%mm3 \n" \
  86                         "psubb         %%mm2, %%mm0 \n" /* { ref0[b]!=ref0[bn], ref0[b]!=ref1[bn] } */ \
  87                         "psubb         %%mm3, %%mm1 \n" /* { ref1[b]!=ref1[bn], ref1[b]!=ref0[bn] } */ \
  88  \
  89                         "por           %%mm1, %%mm0 \n" \
  90                         "movq   %a5(%1,%2,4), %%mm1 \n" \
  91                         "movq   %a6(%1,%2,4), %%mm2 \n" \
  92                         "movq          %%mm1, %%mm3 \n" \
  93                         "movq          %%mm2, %%mm4 \n" \
  94                         "psubw   48(%1,%2,4), %%mm1 \n" \
  95                         "psubw   56(%1,%2,4), %%mm2 \n" \
  96                         "psubw  208(%1,%2,4), %%mm3 \n" \
  97                         "psubw  216(%1,%2,4), %%mm4 \n" \
  98                         "packsswb      %%mm2, %%mm1 \n" \
  99                         "packsswb      %%mm4, %%mm3 \n" \
 100                         "paddb         %%mm6, %%mm1 \n" \
 101                         "paddb         %%mm6, %%mm3 \n" \
 102                         "psubusb       %%mm5, %%mm1 \n" /* abs(mv[b] - mv[bn]) >= limit */ \
 103                         "psubusb       %%mm5, %%mm3 \n" \
 104                         "packsswb      %%mm3, %%mm1 \n" \
 105  \
 106                         "por           %%mm1, %%mm0 \n" \
 107                         "movq   %a7(%1,%2,4), %%mm1 \n" \
 108                         "movq   %a8(%1,%2,4), %%mm2 \n" \
 109                         "movq          %%mm1, %%mm3 \n" \
 110                         "movq          %%mm2, %%mm4 \n" \
 111                         "psubw   48(%1,%2,4), %%mm1 \n" \
 112                         "psubw   56(%1,%2,4), %%mm2 \n" \
 113                         "psubw  208(%1,%2,4), %%mm3 \n" \
 114                         "psubw  216(%1,%2,4), %%mm4 \n" \
 115                         "packsswb      %%mm2, %%mm1 \n" \
 116                         "packsswb      %%mm4, %%mm3 \n" \
 117                         "paddb         %%mm6, %%mm1 \n" \
 118                         "paddb         %%mm6, %%mm3 \n" \
 119                         "psubusb       %%mm5, %%mm1 \n" /* abs(mv[b] - mv[bn]) >= limit */ \
 120                         "psubusb       %%mm5, %%mm3 \n" \
 121                         "packsswb      %%mm3, %%mm1 \n" \
 122  \
 123                         "pshufw $0x4E, %%mm1, %%mm1 \n" \
 124                         "por           %%mm1, %%mm0 \n" \
 125                         "pshufw $0x4E, %%mm0, %%mm1 \n" \
 126                         "pminub        %%mm1, %%mm0 \n" \
 127                         ::"r"(ref), \
 128                           "r"(mv), \
 129                           "r"(b_idx), \
 130                           "i"(d_idx+12), \
 131                           "i"(d_idx+52), \
 132                           "i"(d_idx*4+48), \
 133                           "i"(d_idx*4+56), \
 134                           "i"(d_idx*4+208), \
 135                           "i"(d_idx*4+216) \
 136                     ); \
 137                 } else { \
 138                     __asm__ volatile( \
 139                         "movd   12(%0,%2), %%mm0 \n" \
 140                         "psubb %a3(%0,%2), %%mm0 \n" /* ref[b] != ref[bn] */ \
 141                         "movq   48(%1,%2,4), %%mm1 \n" \
 142                         "movq   56(%1,%2,4), %%mm2 \n" \
 143                         "psubw %a4(%1,%2,4), %%mm1 \n" \
 144                         "psubw %a5(%1,%2,4), %%mm2 \n" \
 145                         "packsswb   %%mm2, %%mm1 \n" \
 146                         "paddb      %%mm6, %%mm1 \n" \
 147                         "psubusb    %%mm5, %%mm1 \n" /* abs(mv[b] - mv[bn]) >= limit */ \
 148                         "packsswb   %%mm1, %%mm1 \n" \
 149                         "por        %%mm1, %%mm0 \n" \
 150                         ::"r"(ref), \
 151                           "r"(mv), \
 152                           "r"(b_idx), \
 153                           "i"(d_idx+12), \
 154                           "i"(d_idx*4+48), \
 155                           "i"(d_idx*4+56) \
 156                     ); \
 157                 } \
 158             } \
 159             __asm__ volatile( \
 160                 "movd 12(%0,%1), %%mm1 \n" \
 161                 "por %a2(%0,%1), %%mm1 \n" /* nnz[b] || nnz[bn] */ \
 162                 ::"r"(nnz), \
 163                   "r"(b_idx), \
 164                   "i"(d_idx+12) \
 165             ); \
 166             __asm__ volatile( \
 167                 "pminub    %%mm7, %%mm1 \n" \
 168                 "pminub    %%mm7, %%mm0 \n" \
 169                 "psllw        $1, %%mm1 \n" \
 170                 "pxor      %%mm2, %%mm2 \n" \
 171                 "pmaxub    %%mm0, %%mm1 \n" \
 172                 "punpcklbw %%mm2, %%mm1 \n" \
 173                 "movq      %%mm1, %a1(%0,%2) \n" \
 174                 ::"r"(bS), \
 175                   "i"(32*dir), \
 176                   "r"(b_idx) \
 177                 :"memory" \
 178             ); \
 179         } \
 180     } while (0)
 181
 182 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],
 183                                             int bidir, int edges, int step, int mask_mv0, int mask_mv1, int field ) {
 184     __asm__ volatile(
 185         "movq %0, %%mm7 \n"
 186         "movq %1, %%mm6 \n"
 187         ::"m"(ff_pb_1), "m"(ff_pb_3)
 188     );
 189     if(field)
 190         __asm__ volatile(
 191             "movq %0, %%mm6 \n"
 192             ::"m"(ff_pb_3_1)
 193         );
 194     __asm__ volatile(
 195         "movq  %%mm6, %%mm5 \n"
 196         "paddb %%mm5, %%mm5 \n"
 197     :);
 198
 199     // could do a special case for dir==0 && edges==1, but it only reduces the
 200     // average filter time by 1.2%
 201     step  <<= 3;
 202     edges <<= 3;
 203     h264_loop_filter_strength_iteration_mmx2(bS, nnz, ref, mv, bidir, edges, step, mask_mv1, 1, -8,  0);
 204     h264_loop_filter_strength_iteration_mmx2(bS, nnz, ref, mv, bidir,    32,    8, mask_mv0, 0, -1, -1);
 205
 206     __asm__ volatile(
 207         "movq   (%0), %%mm0 \n\t"
 208         "movq  8(%0), %%mm1 \n\t"
 209         "movq 16(%0), %%mm2 \n\t"
 210         "movq 24(%0), %%mm3 \n\t"
 211         TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4)
 212         "movq %%mm0,   (%0) \n\t"
 213         "movq %%mm3,  8(%0) \n\t"
 214         "movq %%mm4, 16(%0) \n\t"
 215         "movq %%mm2, 24(%0) \n\t"
 216         ::"r"(bS[0])
 217         :"memory"
 218     );
 219 }
 220
 221 #define LF_FUNC(DIR, TYPE, OPT) \
 222 void ff_x264_deblock_ ## DIR ## _ ## TYPE ## _ ## OPT (uint8_t *pix, int stride, \
 223                                                int alpha, int beta, int8_t *tc0);
 224 #define LF_IFUNC(DIR, TYPE, OPT) \
 225 void ff_x264_deblock_ ## DIR ## _ ## TYPE ## _ ## OPT (uint8_t *pix, int stride, \
 226                                                int alpha, int beta);
 227
 228 LF_FUNC (h,  chroma,       mmxext)
 229 LF_IFUNC(h,  chroma_intra, mmxext)
 230 LF_FUNC (v,  chroma,       mmxext)
 231 LF_IFUNC(v,  chroma_intra, mmxext)
 232
 233 LF_FUNC (h,  luma,         mmxext)
 234 LF_IFUNC(h,  luma_intra,   mmxext)
 235 #if HAVE_YASM && ARCH_X86_32
 236 LF_FUNC (v8, luma,         mmxext)
 237 static void ff_x264_deblock_v_luma_mmxext(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0)
 238 {
 239     if((tc0[0] & tc0[1]) >= 0)
 240         ff_x264_deblock_v8_luma_mmxext(pix+0, stride, alpha, beta, tc0);
 241     if((tc0[2] & tc0[3]) >= 0)
 242         ff_x264_deblock_v8_luma_mmxext(pix+8, stride, alpha, beta, tc0+2);
 243 }
 244 LF_IFUNC(v8, luma_intra,   mmxext)
 245 static void ff_x264_deblock_v_luma_intra_mmxext(uint8_t *pix, int stride, int alpha, int beta)
 246 {
 247     ff_x264_deblock_v8_luma_intra_mmxext(pix+0, stride, alpha, beta);
 248     ff_x264_deblock_v8_luma_intra_mmxext(pix+8, stride, alpha, beta);
 249 }
 250 #endif
 251
 252 LF_FUNC (h,  luma,         sse2)
 253 LF_IFUNC(h,  luma_intra,   sse2)
 254 LF_FUNC (v,  luma,         sse2)
 255 LF_IFUNC(v,  luma_intra,   sse2)
 256
 257 /***********************************/
 258 /* weighted prediction */
 259
 260 #define H264_WEIGHT(W, H, OPT) \
 261 void ff_h264_weight_ ## W ## x ## H ## _ ## OPT(uint8_t *dst, \
 262     int stride, int log2_denom, int weight, int offset);
 263
 264 #define H264_BIWEIGHT(W, H, OPT) \
 265 void ff_h264_biweight_ ## W ## x ## H ## _ ## OPT(uint8_t *dst, \
 266     uint8_t *src, int stride, int log2_denom, int weightd, \
 267     int weights, int offset);
 268
 269 #define H264_BIWEIGHT_MMX(W,H) \
 270 H264_WEIGHT  (W, H, mmx2) \
 271 H264_BIWEIGHT(W, H, mmx2)
 272
 273 #define H264_BIWEIGHT_MMX_SSE(W,H) \
 274 H264_BIWEIGHT_MMX(W, H) \
 275 H264_WEIGHT      (W, H, sse2) \
 276 H264_BIWEIGHT    (W, H, sse2) \
 277 H264_BIWEIGHT    (W, H, ssse3)
 278
 279 H264_BIWEIGHT_MMX_SSE(16, 16)
 280 H264_BIWEIGHT_MMX_SSE(16,  8)
 281 H264_BIWEIGHT_MMX_SSE( 8, 16)
 282 H264_BIWEIGHT_MMX_SSE( 8,  8)
 283 H264_BIWEIGHT_MMX_SSE( 8,  4)
 284 H264_BIWEIGHT_MMX    ( 4,  8)
 285 H264_BIWEIGHT_MMX    ( 4,  4)
 286 H264_BIWEIGHT_MMX    ( 4,  2)
 287
 288 void ff_h264dsp_init_x86(H264DSPContext *c, const int bit_depth)
 289 {
 290     int mm_flags = av_get_cpu_flags();
 291
 292     if (bit_depth == 8) {
 293     if (mm_flags & AV_CPU_FLAG_MMX2) {
 294         c->h264_loop_filter_strength= h264_loop_filter_strength_mmx2;
 295     }
 296 #if HAVE_YASM
 297     if (mm_flags & AV_CPU_FLAG_MMX) {
 298         c->h264_idct_dc_add=
 299         c->h264_idct_add= ff_h264_idct_add_mmx;
 300         c->h264_idct8_dc_add=
 301         c->h264_idct8_add= ff_h264_idct8_add_mmx;
 302
 303         c->h264_idct_add16     = ff_h264_idct_add16_mmx;
 304         c->h264_idct8_add4     = ff_h264_idct8_add4_mmx;
 305         c->h264_idct_add8      = ff_h264_idct_add8_mmx;
 306         c->h264_idct_add16intra= ff_h264_idct_add16intra_mmx;
 307         c->h264_luma_dc_dequant_idct= ff_h264_luma_dc_dequant_idct_mmx;
 308
 309         if (mm_flags & AV_CPU_FLAG_MMX2) {
 310             c->h264_idct_dc_add= ff_h264_idct_dc_add_mmx2;
 311             c->h264_idct8_dc_add= ff_h264_idct8_dc_add_mmx2;
 312             c->h264_idct_add16     = ff_h264_idct_add16_mmx2;
 313             c->h264_idct8_add4     = ff_h264_idct8_add4_mmx2;
 314             c->h264_idct_add8      = ff_h264_idct_add8_mmx2;
 315             c->h264_idct_add16intra= ff_h264_idct_add16intra_mmx2;
 316
 317             c->h264_v_loop_filter_chroma= ff_x264_deblock_v_chroma_mmxext;
 318             c->h264_h_loop_filter_chroma= ff_x264_deblock_h_chroma_mmxext;
 319             c->h264_v_loop_filter_chroma_intra= ff_x264_deblock_v_chroma_intra_mmxext;
 320             c->h264_h_loop_filter_chroma_intra= ff_x264_deblock_h_chroma_intra_mmxext;
 321 #if ARCH_X86_32
 322             c->h264_v_loop_filter_luma= ff_x264_deblock_v_luma_mmxext;
 323             c->h264_h_loop_filter_luma= ff_x264_deblock_h_luma_mmxext;
 324             c->h264_v_loop_filter_luma_intra = ff_x264_deblock_v_luma_intra_mmxext;
 325             c->h264_h_loop_filter_luma_intra = ff_x264_deblock_h_luma_intra_mmxext;
 326 #endif
 327             c->weight_h264_pixels_tab[0]= ff_h264_weight_16x16_mmx2;
 328             c->weight_h264_pixels_tab[1]= ff_h264_weight_16x8_mmx2;
 329             c->weight_h264_pixels_tab[2]= ff_h264_weight_8x16_mmx2;
 330             c->weight_h264_pixels_tab[3]= ff_h264_weight_8x8_mmx2;
 331             c->weight_h264_pixels_tab[4]= ff_h264_weight_8x4_mmx2;
 332             c->weight_h264_pixels_tab[5]= ff_h264_weight_4x8_mmx2;
 333             c->weight_h264_pixels_tab[6]= ff_h264_weight_4x4_mmx2;
 334             c->weight_h264_pixels_tab[7]= ff_h264_weight_4x2_mmx2;
 335
 336             c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_mmx2;
 337             c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_mmx2;
 338             c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_mmx2;
 339             c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_mmx2;
 340             c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_mmx2;
 341             c->biweight_h264_pixels_tab[5]= ff_h264_biweight_4x8_mmx2;
 342             c->biweight_h264_pixels_tab[6]= ff_h264_biweight_4x4_mmx2;
 343             c->biweight_h264_pixels_tab[7]= ff_h264_biweight_4x2_mmx2;
 344
 345             if (mm_flags&AV_CPU_FLAG_SSE2) {
 346                 c->h264_idct8_add = ff_h264_idct8_add_sse2;
 347                 c->h264_idct8_add4= ff_h264_idct8_add4_sse2;
 348                 c->h264_luma_dc_dequant_idct= ff_h264_luma_dc_dequant_idct_sse2;
 349
 350                 c->weight_h264_pixels_tab[0]= ff_h264_weight_16x16_sse2;
 351                 c->weight_h264_pixels_tab[1]= ff_h264_weight_16x8_sse2;
 352                 c->weight_h264_pixels_tab[2]= ff_h264_weight_8x16_sse2;
 353                 c->weight_h264_pixels_tab[3]= ff_h264_weight_8x8_sse2;
 354                 c->weight_h264_pixels_tab[4]= ff_h264_weight_8x4_sse2;
 355
 356                 c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_sse2;
 357                 c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_sse2;
 358                 c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_sse2;
 359                 c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_sse2;
 360                 c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_sse2;
 361
 362 #if HAVE_ALIGNED_STACK
 363                 c->h264_v_loop_filter_luma = ff_x264_deblock_v_luma_sse2;
 364                 c->h264_h_loop_filter_luma = ff_x264_deblock_h_luma_sse2;
 365                 c->h264_v_loop_filter_luma_intra = ff_x264_deblock_v_luma_intra_sse2;
 366                 c->h264_h_loop_filter_luma_intra = ff_x264_deblock_h_luma_intra_sse2;
 367 #endif
 368
 369                 c->h264_idct_add16 = ff_h264_idct_add16_sse2;
 370                 c->h264_idct_add8  = ff_h264_idct_add8_sse2;
 371                 c->h264_idct_add16intra = ff_h264_idct_add16intra_sse2;
 372             }
 373             if (mm_flags&AV_CPU_FLAG_SSSE3) {
 374                 c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_ssse3;
 375                 c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_ssse3;
 376                 c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_ssse3;
 377                 c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_ssse3;
 378                 c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_ssse3;
 379             }
 380         }
 381     }
 382     }
 383 #endif
 384 }