libavcodec/h264_mb.c

   1 /*
   2  * H.26L/H.264/AVC/JVT/14496-10/... decoder
   3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
   4  *
   5  * This file is part of FFmpeg.
   6  *
   7  * FFmpeg is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; either
  10  * version 2.1 of the License, or (at your option) any later version.
  11  *
  12  * FFmpeg is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with FFmpeg; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 /**
  23  * @file
  24  * H.264 / AVC / MPEG-4 part10 macroblock decoding
  25  */
  26
  27 #include <stdint.h>
  28
  29 #include "config.h"
  30
  31 #include "libavutil/common.h"
  32 #include "libavutil/intreadwrite.h"
  33 #include "avcodec.h"
  34 #include "h264dec.h"
  35 #include "h264_ps.h"
  36 #include "qpeldsp.h"
  37 #include "rectangle.h"
  38 #include "threadframe.h"
  39
  40 static inline int get_lowest_part_list_y(H264SliceContext *sl,
  41                                          int n, int height, int y_offset, int list)
  42 {
  43     int raw_my             = sl->mv_cache[list][scan8[n]][1];
  44     int filter_height_down = (raw_my & 3) ? 3 : 0;
  45     int full_my            = (raw_my >> 2) + y_offset;
  46     int bottom             = full_my + filter_height_down + height;
  47
  48     av_assert2(height >= 0);
  49
  50     return FFMAX(0, bottom);
  51 }
  52
  53 static inline void get_lowest_part_y(const H264Context *h, H264SliceContext *sl,
  54                                      int16_t refs[2][48], int n,
  55                                      int height, int y_offset, int list0,
  56                                      int list1, int *nrefs)
  57 {
  58     int my;
  59
  60     y_offset += 16 * (sl->mb_y >> MB_FIELD(sl));
  61
  62     if (list0) {
  63         int ref_n = sl->ref_cache[0][scan8[n]];
  64         H264Ref *ref = &sl->ref_list[0][ref_n];
  65
  66         // Error resilience puts the current picture in the ref list.
  67         // Don't try to wait on these as it will cause a deadlock.
  68         // Fields can wait on each other, though.
  69         if (ref->parent->tf.progress != h->cur_pic.tf.progress ||
  70             (ref->reference & 3) != h->picture_structure) {
  71             my = get_lowest_part_list_y(sl, n, height, y_offset, 0);
  72             if (refs[0][ref_n] < 0)
  73                 nrefs[0] += 1;
  74             refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
  75         }
  76     }
  77
  78     if (list1) {
  79         int ref_n    = sl->ref_cache[1][scan8[n]];
  80         H264Ref *ref = &sl->ref_list[1][ref_n];
  81
  82         if (ref->parent->tf.progress != h->cur_pic.tf.progress ||
  83             (ref->reference & 3) != h->picture_structure) {
  84             my = get_lowest_part_list_y(sl, n, height, y_offset, 1);
  85             if (refs[1][ref_n] < 0)
  86                 nrefs[1] += 1;
  87             refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
  88         }
  89     }
  90 }
  91
  92 /**
  93  * Wait until all reference frames are available for MC operations.
  94  *
  95  * @param h the H.264 context
  96  */
  97 static void await_references(const H264Context *h, H264SliceContext *sl)
  98 {
  99     const int mb_xy   = sl->mb_xy;
 100     const int mb_type = h->cur_pic.mb_type[mb_xy];
 101     int16_t refs[2][48];
 102     int nrefs[2] = { 0 };
 103     int ref, list;
 104
 105     memset(refs, -1, sizeof(refs));
 106
 107     if (IS_16X16(mb_type)) {
 108         get_lowest_part_y(h, sl, refs, 0, 16, 0,
 109                           IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 110     } else if (IS_16X8(mb_type)) {
 111         get_lowest_part_y(h, sl, refs, 0, 8, 0,
 112                           IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 113         get_lowest_part_y(h, sl, refs, 8, 8, 8,
 114                           IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
 115     } else if (IS_8X16(mb_type)) {
 116         get_lowest_part_y(h, sl, refs, 0, 16, 0,
 117                           IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 118         get_lowest_part_y(h, sl, refs, 4, 16, 0,
 119                           IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
 120     } else {
 121         int i;
 122
 123         av_assert2(IS_8X8(mb_type));
 124
 125         for (i = 0; i < 4; i++) {
 126             const int sub_mb_type = sl->sub_mb_type[i];
 127             const int n           = 4 * i;
 128             int y_offset          = (i & 2) << 2;
 129
 130             if (IS_SUB_8X8(sub_mb_type)) {
 131                 get_lowest_part_y(h, sl, refs, n, 8, y_offset,
 132                                   IS_DIR(sub_mb_type, 0, 0),
 133                                   IS_DIR(sub_mb_type, 0, 1),
 134                                   nrefs);
 135             } else if (IS_SUB_8X4(sub_mb_type)) {
 136                 get_lowest_part_y(h, sl, refs, n, 4, y_offset,
 137                                   IS_DIR(sub_mb_type, 0, 0),
 138                                   IS_DIR(sub_mb_type, 0, 1),
 139                                   nrefs);
 140                 get_lowest_part_y(h, sl, refs, n + 2, 4, y_offset + 4,
 141                                   IS_DIR(sub_mb_type, 0, 0),
 142                                   IS_DIR(sub_mb_type, 0, 1),
 143                                   nrefs);
 144             } else if (IS_SUB_4X8(sub_mb_type)) {
 145                 get_lowest_part_y(h, sl, refs, n, 8, y_offset,
 146                                   IS_DIR(sub_mb_type, 0, 0),
 147                                   IS_DIR(sub_mb_type, 0, 1),
 148                                   nrefs);
 149                 get_lowest_part_y(h, sl, refs, n + 1, 8, y_offset,
 150                                   IS_DIR(sub_mb_type, 0, 0),
 151                                   IS_DIR(sub_mb_type, 0, 1),
 152                                   nrefs);
 153             } else {
 154                 int j;
 155                 av_assert2(IS_SUB_4X4(sub_mb_type));
 156                 for (j = 0; j < 4; j++) {
 157                     int sub_y_offset = y_offset + 2 * (j & 2);
 158                     get_lowest_part_y(h, sl, refs, n + j, 4, sub_y_offset,
 159                                       IS_DIR(sub_mb_type, 0, 0),
 160                                       IS_DIR(sub_mb_type, 0, 1),
 161                                       nrefs);
 162                 }
 163             }
 164         }
 165     }
 166
 167     for (list = sl->list_count - 1; list >= 0; list--)
 168         for (ref = 0; ref < 48 && nrefs[list]; ref++) {
 169             int row = refs[list][ref];
 170             if (row >= 0) {
 171                 H264Ref *ref_pic  = &sl->ref_list[list][ref];
 172                 int ref_field         = ref_pic->reference - 1;
 173                 int ref_field_picture = ref_pic->parent->field_picture;
 174                 int pic_height        = 16 * h->mb_height >> ref_field_picture;
 175
 176                 row <<= MB_MBAFF(sl);
 177                 nrefs[list]--;
 178
 179                 if (!FIELD_PICTURE(h) && ref_field_picture) { // frame referencing two fields
 180                     av_assert2((ref_pic->parent->reference & 3) == 3);
 181                     ff_thread_await_progress(&ref_pic->parent->tf,
 182                                              FFMIN((row >> 1) - !(row & 1),
 183                                                    pic_height - 1),
 184                                              1);
 185                     ff_thread_await_progress(&ref_pic->parent->tf,
 186                                              FFMIN((row >> 1), pic_height - 1),
 187                                              0);
 188                 } else if (FIELD_PICTURE(h) && !ref_field_picture) { // field referencing one field of a frame
 189                     ff_thread_await_progress(&ref_pic->parent->tf,
 190                                              FFMIN(row * 2 + ref_field,
 191                                                    pic_height - 1),
 192                                              0);
 193                 } else if (FIELD_PICTURE(h)) {
 194                     ff_thread_await_progress(&ref_pic->parent->tf,
 195                                              FFMIN(row, pic_height - 1),
 196                                              ref_field);
 197                 } else {
 198                     ff_thread_await_progress(&ref_pic->parent->tf,
 199                                              FFMIN(row, pic_height - 1),
 200                                              0);
 201                 }
 202             }
 203         }
 204 }
 205
 206 static av_always_inline void mc_dir_part(const H264Context *h, H264SliceContext *sl,
 207                                          H264Ref *pic,
 208                                          int n, int square, int height,
 209                                          int delta, int list,
 210                                          uint8_t *dest_y, uint8_t *dest_cb,
 211                                          uint8_t *dest_cr,
 212                                          int src_x_offset, int src_y_offset,
 213                                          const qpel_mc_func *qpix_op,
 214                                          h264_chroma_mc_func chroma_op,
 215                                          int pixel_shift, int chroma_idc)
 216 {
 217     const int mx      = sl->mv_cache[list][scan8[n]][0] + src_x_offset * 8;
 218     int my            = sl->mv_cache[list][scan8[n]][1] + src_y_offset * 8;
 219     const int luma_xy = (mx & 3) + ((my & 3) << 2);
 220     ptrdiff_t offset  = (mx >> 2) * (1 << pixel_shift) + (my >> 2) * sl->mb_linesize;
 221     uint8_t *src_y    = pic->data[0] + offset;
 222     uint8_t *src_cb, *src_cr;
 223     int extra_width  = 0;
 224     int extra_height = 0;
 225     int emu = 0;
 226     const int full_mx    = mx >> 2;
 227     const int full_my    = my >> 2;
 228     const int pic_width  = 16 * h->mb_width;
 229     const int pic_height = 16 * h->mb_height >> MB_FIELD(sl);
 230     int ysh;
 231
 232     if (mx & 7)
 233         extra_width -= 3;
 234     if (my & 7)
 235         extra_height -= 3;
 236
 237     if (full_mx                <          0 - extra_width  ||
 238         full_my                <          0 - extra_height ||
 239         full_mx + 16 /*FIXME*/ > pic_width  + extra_width  ||
 240         full_my + 16 /*FIXME*/ > pic_height + extra_height) {
 241         h->vdsp.emulated_edge_mc(sl->edge_emu_buffer,
 242                                  src_y - (2 << pixel_shift) - 2 * sl->mb_linesize,
 243                                  sl->mb_linesize, sl->mb_linesize,
 244                                  16 + 5, 16 + 5 /*FIXME*/, full_mx - 2,
 245                                  full_my - 2, pic_width, pic_height);
 246         src_y = sl->edge_emu_buffer + (2 << pixel_shift) + 2 * sl->mb_linesize;
 247         emu   = 1;
 248     }
 249
 250     qpix_op[luma_xy](dest_y, src_y, sl->mb_linesize); // FIXME try variable height perhaps?
 251     if (!square)
 252         qpix_op[luma_xy](dest_y + delta, src_y + delta, sl->mb_linesize);
 253
 254     if (CONFIG_GRAY && h->flags & AV_CODEC_FLAG_GRAY)
 255         return;
 256
 257     if (chroma_idc == 3 /* yuv444 */) {
 258         src_cb = pic->data[1] + offset;
 259         if (emu) {
 260             h->vdsp.emulated_edge_mc(sl->edge_emu_buffer,
 261                                      src_cb - (2 << pixel_shift) - 2 * sl->mb_linesize,
 262                                      sl->mb_linesize, sl->mb_linesize,
 263                                      16 + 5, 16 + 5 /*FIXME*/,
 264                                      full_mx - 2, full_my - 2,
 265                                      pic_width, pic_height);
 266             src_cb = sl->edge_emu_buffer + (2 << pixel_shift) + 2 * sl->mb_linesize;
 267         }
 268         qpix_op[luma_xy](dest_cb, src_cb, sl->mb_linesize); // FIXME try variable height perhaps?
 269         if (!square)
 270             qpix_op[luma_xy](dest_cb + delta, src_cb + delta, sl->mb_linesize);
 271
 272         src_cr = pic->data[2] + offset;
 273         if (emu) {
 274             h->vdsp.emulated_edge_mc(sl->edge_emu_buffer,
 275                                      src_cr - (2 << pixel_shift) - 2 * sl->mb_linesize,
 276                                      sl->mb_linesize, sl->mb_linesize,
 277                                      16 + 5, 16 + 5 /*FIXME*/,
 278                                      full_mx - 2, full_my - 2,
 279                                      pic_width, pic_height);
 280             src_cr = sl->edge_emu_buffer + (2 << pixel_shift) + 2 * sl->mb_linesize;
 281         }
 282         qpix_op[luma_xy](dest_cr, src_cr, sl->mb_linesize); // FIXME try variable height perhaps?
 283         if (!square)
 284             qpix_op[luma_xy](dest_cr + delta, src_cr + delta, sl->mb_linesize);
 285         return;
 286     }
 287
 288     ysh = 3 - (chroma_idc == 2 /* yuv422 */);
 289     if (chroma_idc == 1 /* yuv420 */ && MB_FIELD(sl)) {
 290         // chroma offset when predicting from a field of opposite parity
 291         my  += 2 * ((sl->mb_y & 1) - (pic->reference - 1));
 292         emu |= (my >> 3) < 0 || (my >> 3) + 8 >= (pic_height >> 1);
 293     }
 294
 295     src_cb = pic->data[1] + ((mx >> 3) * (1 << pixel_shift)) +
 296              (my >> ysh) * sl->mb_uvlinesize;
 297     src_cr = pic->data[2] + ((mx >> 3) * (1 << pixel_shift)) +
 298              (my >> ysh) * sl->mb_uvlinesize;
 299
 300     if (emu) {
 301         h->vdsp.emulated_edge_mc(sl->edge_emu_buffer, src_cb,
 302                                  sl->mb_uvlinesize, sl->mb_uvlinesize,
 303                                  9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
 304                                  pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
 305         src_cb = sl->edge_emu_buffer;
 306     }
 307     chroma_op(dest_cb, src_cb, sl->mb_uvlinesize,
 308               height >> (chroma_idc == 1 /* yuv420 */),
 309               mx & 7, ((unsigned)my << (chroma_idc == 2 /* yuv422 */)) & 7);
 310
 311     if (emu) {
 312         h->vdsp.emulated_edge_mc(sl->edge_emu_buffer, src_cr,
 313                                  sl->mb_uvlinesize, sl->mb_uvlinesize,
 314                                  9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
 315                                  pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
 316         src_cr = sl->edge_emu_buffer;
 317     }
 318     chroma_op(dest_cr, src_cr, sl->mb_uvlinesize, height >> (chroma_idc == 1 /* yuv420 */),
 319               mx & 7, ((unsigned)my << (chroma_idc == 2 /* yuv422 */)) & 7);
 320 }
 321
 322 static av_always_inline void mc_part_std(const H264Context *h, H264SliceContext *sl,
 323                                          int n, int square,
 324                                          int height, int delta,
 325                                          uint8_t *dest_y, uint8_t *dest_cb,
 326                                          uint8_t *dest_cr,
 327                                          int x_offset, int y_offset,
 328                                          const qpel_mc_func *qpix_put,
 329                                          h264_chroma_mc_func chroma_put,
 330                                          const qpel_mc_func *qpix_avg,
 331                                          h264_chroma_mc_func chroma_avg,
 332                                          int list0, int list1,
 333                                          int pixel_shift, int chroma_idc)
 334 {
 335     const qpel_mc_func *qpix_op   = qpix_put;
 336     h264_chroma_mc_func chroma_op = chroma_put;
 337
 338     dest_y += (2 * x_offset << pixel_shift) + 2 * y_offset * sl->mb_linesize;
 339     if (chroma_idc == 3 /* yuv444 */) {
 340         dest_cb += (2 * x_offset << pixel_shift) + 2 * y_offset * sl->mb_linesize;
 341         dest_cr += (2 * x_offset << pixel_shift) + 2 * y_offset * sl->mb_linesize;
 342     } else if (chroma_idc == 2 /* yuv422 */) {
 343         dest_cb += (x_offset << pixel_shift) + 2 * y_offset * sl->mb_uvlinesize;
 344         dest_cr += (x_offset << pixel_shift) + 2 * y_offset * sl->mb_uvlinesize;
 345     } else { /* yuv420 */
 346         dest_cb += (x_offset << pixel_shift) + y_offset * sl->mb_uvlinesize;
 347         dest_cr += (x_offset << pixel_shift) + y_offset * sl->mb_uvlinesize;
 348     }
 349     x_offset += 8 * sl->mb_x;
 350     y_offset += 8 * (sl->mb_y >> MB_FIELD(sl));
 351
 352     if (list0) {
 353         H264Ref *ref = &sl->ref_list[0][sl->ref_cache[0][scan8[n]]];
 354         mc_dir_part(h, sl, ref, n, square, height, delta, 0,
 355                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
 356                     qpix_op, chroma_op, pixel_shift, chroma_idc);
 357
 358         qpix_op   = qpix_avg;
 359         chroma_op = chroma_avg;
 360     }
 361
 362     if (list1) {
 363         H264Ref *ref = &sl->ref_list[1][sl->ref_cache[1][scan8[n]]];
 364         mc_dir_part(h, sl, ref, n, square, height, delta, 1,
 365                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
 366                     qpix_op, chroma_op, pixel_shift, chroma_idc);
 367     }
 368 }
 369
 370 static av_always_inline void mc_part_weighted(const H264Context *h, H264SliceContext *sl,
 371                                               int n, int square,
 372                                               int height, int delta,
 373                                               uint8_t *dest_y, uint8_t *dest_cb,
 374                                               uint8_t *dest_cr,
 375                                               int x_offset, int y_offset,
 376                                               const qpel_mc_func *qpix_put,
 377                                               h264_chroma_mc_func chroma_put,
 378                                               h264_weight_func luma_weight_op,
 379                                               h264_weight_func chroma_weight_op,
 380                                               h264_biweight_func luma_weight_avg,
 381                                               h264_biweight_func chroma_weight_avg,
 382                                               int list0, int list1,
 383                                               int pixel_shift, int chroma_idc)
 384 {
 385     int chroma_height;
 386
 387     dest_y += (2 * x_offset << pixel_shift) + 2 * y_offset * sl->mb_linesize;
 388     if (chroma_idc == 3 /* yuv444 */) {
 389         chroma_height     = height;
 390         chroma_weight_avg = luma_weight_avg;
 391         chroma_weight_op  = luma_weight_op;
 392         dest_cb += (2 * x_offset << pixel_shift) + 2 * y_offset * sl->mb_linesize;
 393         dest_cr += (2 * x_offset << pixel_shift) + 2 * y_offset * sl->mb_linesize;
 394     } else if (chroma_idc == 2 /* yuv422 */) {
 395         chroma_height = height;
 396         dest_cb      += (x_offset << pixel_shift) + 2 * y_offset * sl->mb_uvlinesize;
 397         dest_cr      += (x_offset << pixel_shift) + 2 * y_offset * sl->mb_uvlinesize;
 398     } else { /* yuv420 */
 399         chroma_height = height >> 1;
 400         dest_cb      += (x_offset << pixel_shift) + y_offset * sl->mb_uvlinesize;
 401         dest_cr      += (x_offset << pixel_shift) + y_offset * sl->mb_uvlinesize;
 402     }
 403     x_offset += 8 * sl->mb_x;
 404     y_offset += 8 * (sl->mb_y >> MB_FIELD(sl));
 405
 406     if (list0 && list1) {
 407         /* don't optimize for luma-only case, since B-frames usually
 408          * use implicit weights => chroma too. */
 409         uint8_t *tmp_cb = sl->bipred_scratchpad;
 410         uint8_t *tmp_cr = sl->bipred_scratchpad + (16 << pixel_shift);
 411         uint8_t *tmp_y  = sl->bipred_scratchpad + 16 * sl->mb_uvlinesize;
 412         int refn0       = sl->ref_cache[0][scan8[n]];
 413         int refn1       = sl->ref_cache[1][scan8[n]];
 414
 415         mc_dir_part(h, sl, &sl->ref_list[0][refn0], n, square, height, delta, 0,
 416                     dest_y, dest_cb, dest_cr,
 417                     x_offset, y_offset, qpix_put, chroma_put,
 418                     pixel_shift, chroma_idc);
 419         mc_dir_part(h, sl, &sl->ref_list[1][refn1], n, square, height, delta, 1,
 420                     tmp_y, tmp_cb, tmp_cr,
 421                     x_offset, y_offset, qpix_put, chroma_put,
 422                     pixel_shift, chroma_idc);
 423
 424         if (sl->pwt.use_weight == 2) {
 425             int weight0 = sl->pwt.implicit_weight[refn0][refn1][sl->mb_y & 1];
 426             int weight1 = 64 - weight0;
 427             luma_weight_avg(dest_y, tmp_y, sl->mb_linesize,
 428                             height, 5, weight0, weight1, 0);
 429             if (!CONFIG_GRAY || !(h->flags & AV_CODEC_FLAG_GRAY)) {
 430                 chroma_weight_avg(dest_cb, tmp_cb, sl->mb_uvlinesize,
 431                                   chroma_height, 5, weight0, weight1, 0);
 432                 chroma_weight_avg(dest_cr, tmp_cr, sl->mb_uvlinesize,
 433                                   chroma_height, 5, weight0, weight1, 0);
 434             }
 435         } else {
 436             luma_weight_avg(dest_y, tmp_y, sl->mb_linesize, height,
 437                             sl->pwt.luma_log2_weight_denom,
 438                             sl->pwt.luma_weight[refn0][0][0],
 439                             sl->pwt.luma_weight[refn1][1][0],
 440                             sl->pwt.luma_weight[refn0][0][1] +
 441                             sl->pwt.luma_weight[refn1][1][1]);
 442             if (!CONFIG_GRAY || !(h->flags & AV_CODEC_FLAG_GRAY)) {
 443                 chroma_weight_avg(dest_cb, tmp_cb, sl->mb_uvlinesize, chroma_height,
 444                                   sl->pwt.chroma_log2_weight_denom,
 445                                   sl->pwt.chroma_weight[refn0][0][0][0],
 446                                   sl->pwt.chroma_weight[refn1][1][0][0],
 447                                   sl->pwt.chroma_weight[refn0][0][0][1] +
 448                                   sl->pwt.chroma_weight[refn1][1][0][1]);
 449                 chroma_weight_avg(dest_cr, tmp_cr, sl->mb_uvlinesize, chroma_height,
 450                                   sl->pwt.chroma_log2_weight_denom,
 451                                   sl->pwt.chroma_weight[refn0][0][1][0],
 452                                   sl->pwt.chroma_weight[refn1][1][1][0],
 453                                   sl->pwt.chroma_weight[refn0][0][1][1] +
 454                                   sl->pwt.chroma_weight[refn1][1][1][1]);
 455             }
 456         }
 457     } else {
 458         int list     = list1 ? 1 : 0;
 459         int refn     = sl->ref_cache[list][scan8[n]];
 460         H264Ref *ref = &sl->ref_list[list][refn];
 461         mc_dir_part(h, sl, ref, n, square, height, delta, list,
 462                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
 463                     qpix_put, chroma_put, pixel_shift, chroma_idc);
 464
 465         luma_weight_op(dest_y, sl->mb_linesize, height,
 466                        sl->pwt.luma_log2_weight_denom,
 467                        sl->pwt.luma_weight[refn][list][0],
 468                        sl->pwt.luma_weight[refn][list][1]);
 469         if (!CONFIG_GRAY || !(h->flags & AV_CODEC_FLAG_GRAY)) {
 470             if (sl->pwt.use_weight_chroma) {
 471                 chroma_weight_op(dest_cb, sl->mb_uvlinesize, chroma_height,
 472                                  sl->pwt.chroma_log2_weight_denom,
 473                                  sl->pwt.chroma_weight[refn][list][0][0],
 474                                  sl->pwt.chroma_weight[refn][list][0][1]);
 475                 chroma_weight_op(dest_cr, sl->mb_uvlinesize, chroma_height,
 476                                  sl->pwt.chroma_log2_weight_denom,
 477                                  sl->pwt.chroma_weight[refn][list][1][0],
 478                                  sl->pwt.chroma_weight[refn][list][1][1]);
 479             }
 480         }
 481     }
 482 }
 483
 484 static av_always_inline void prefetch_motion(const H264Context *h, H264SliceContext *sl,
 485                                              int list, int pixel_shift,
 486                                              int chroma_idc)
 487 {
 488     /* fetch pixels for estimated mv 4 macroblocks ahead
 489      * optimized for 64byte cache lines */
 490     const int refn = sl->ref_cache[list][scan8[0]];
 491     if (refn >= 0) {
 492         const int mx  = (sl->mv_cache[list][scan8[0]][0] >> 2) + 16 * sl->mb_x + 8;
 493         const int my  = (sl->mv_cache[list][scan8[0]][1] >> 2) + 16 * sl->mb_y;
 494         uint8_t **src = sl->ref_list[list][refn].data;
 495         int off       =  mx * (1<< pixel_shift) +
 496                         (my + (sl->mb_x & 3) * 4) * sl->mb_linesize +
 497                         (64 << pixel_shift);
 498         h->vdsp.prefetch(src[0] + off, sl->linesize, 4);
 499         if (chroma_idc == 3 /* yuv444 */) {
 500             h->vdsp.prefetch(src[1] + off, sl->linesize, 4);
 501             h->vdsp.prefetch(src[2] + off, sl->linesize, 4);
 502         } else {
 503             off= ((mx>>1)+64) * (1<<pixel_shift) + ((my>>1) + (sl->mb_x&7))*sl->uvlinesize;
 504             h->vdsp.prefetch(src[1] + off, src[2] - src[1], 2);
 505         }
 506     }
 507 }
 508
 509 static av_always_inline void xchg_mb_border(const H264Context *h, H264SliceContext *sl,
 510                                             uint8_t *src_y,
 511                                             uint8_t *src_cb, uint8_t *src_cr,
 512                                             int linesize, int uvlinesize,
 513                                             int xchg, int chroma444,
 514                                             int simple, int pixel_shift)
 515 {
 516     int deblock_topleft;
 517     int deblock_top;
 518     int top_idx = 1;
 519     uint8_t *top_border_m1;
 520     uint8_t *top_border;
 521
 522     if (!simple && FRAME_MBAFF(h)) {
 523         if (sl->mb_y & 1) {
 524             if (!MB_MBAFF(sl))
 525                 return;
 526         } else {
 527             top_idx = MB_MBAFF(sl) ? 0 : 1;
 528         }
 529     }
 530
 531     if (sl->deblocking_filter == 2) {
 532         deblock_topleft = h->slice_table[sl->mb_xy - 1 - h->mb_stride] == sl->slice_num;
 533         deblock_top     = sl->top_type;
 534     } else {
 535         deblock_topleft = (sl->mb_x > 0);
 536         deblock_top     = (sl->mb_y > !!MB_FIELD(sl));
 537     }
 538
 539     src_y  -= linesize   + 1 + pixel_shift;
 540     src_cb -= uvlinesize + 1 + pixel_shift;
 541     src_cr -= uvlinesize + 1 + pixel_shift;
 542
 543     top_border_m1 = sl->top_borders[top_idx][sl->mb_x - 1];
 544     top_border    = sl->top_borders[top_idx][sl->mb_x];
 545
 546 #define XCHG(a, b, xchg)                        \
 547     if (pixel_shift) {                          \
 548         if (xchg) {                             \
 549             AV_SWAP64(b + 0, a + 0);            \
 550             AV_SWAP64(b + 8, a + 8);            \
 551         } else {                                \
 552             AV_COPY128(b, a);                   \
 553         }                                       \
 554     } else if (xchg)                            \
 555         AV_SWAP64(b, a);                        \
 556     else                                        \
 557         AV_COPY64(b, a);
 558
 559     if (deblock_top) {
 560         if (deblock_topleft) {
 561             XCHG(top_border_m1 + (8 << pixel_shift),
 562                  src_y - (7 << pixel_shift), 1);
 563         }
 564         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
 565         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
 566         if (sl->mb_x + 1 < h->mb_width) {
 567             XCHG(sl->top_borders[top_idx][sl->mb_x + 1],
 568                  src_y + (17 << pixel_shift), 1);
 569         }
 570         if (simple || !CONFIG_GRAY || !(h->flags & AV_CODEC_FLAG_GRAY)) {
 571             if (chroma444) {
 572                 if (deblock_topleft) {
 573                     XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
 574                     XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
 575                 }
 576                 XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
 577                 XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
 578                 XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
 579                 XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
 580                 if (sl->mb_x + 1 < h->mb_width) {
 581                     XCHG(sl->top_borders[top_idx][sl->mb_x + 1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
 582                     XCHG(sl->top_borders[top_idx][sl->mb_x + 1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
 583                 }
 584             } else {
 585                 if (deblock_topleft) {
 586                     XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
 587                     XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
 588                 }
 589                 XCHG(top_border + (16 << pixel_shift), src_cb + 1 + pixel_shift, 1);
 590                 XCHG(top_border + (24 << pixel_shift), src_cr + 1 + pixel_shift, 1);
 591             }
 592         }
 593     }
 594 }
 595
 596 static av_always_inline int dctcoef_get(int16_t *mb, int high_bit_depth,
 597                                         int index)
 598 {
 599     if (high_bit_depth) {
 600         return AV_RN32A(((int32_t *)mb) + index);
 601     } else
 602         return AV_RN16A(mb + index);
 603 }
 604
 605 static av_always_inline void dctcoef_set(int16_t *mb, int high_bit_depth,
 606                                          int index, int value)
 607 {
 608     if (high_bit_depth) {
 609         AV_WN32A(((int32_t *)mb) + index, value);
 610     } else
 611         AV_WN16A(mb + index, value);
 612 }
 613
 614 static av_always_inline void hl_decode_mb_predict_luma(const H264Context *h,
 615                                                        H264SliceContext *sl,
 616                                                        int mb_type, int simple,
 617                                                        int transform_bypass,
 618                                                        int pixel_shift,
 619                                                        const int *block_offset,
 620                                                        int linesize,
 621                                                        uint8_t *dest_y, int p)
 622 {
 623     void (*idct_add)(uint8_t *dst, int16_t *block, int stride);
 624     void (*idct_dc_add)(uint8_t *dst, int16_t *block, int stride);
 625     int i;
 626     int qscale = p == 0 ? sl->qscale : sl->chroma_qp[p - 1];
 627     block_offset += 16 * p;
 628     if (IS_INTRA4x4(mb_type)) {
 629         if (IS_8x8DCT(mb_type)) {
 630             if (transform_bypass) {
 631                 idct_dc_add =
 632                 idct_add    = h->h264dsp.h264_add_pixels8_clear;
 633             } else {
 634                 idct_dc_add = h->h264dsp.h264_idct8_dc_add;
 635                 idct_add    = h->h264dsp.h264_idct8_add;
 636             }
 637             for (i = 0; i < 16; i += 4) {
 638                 uint8_t *const ptr = dest_y + block_offset[i];
 639                 const int dir      = sl->intra4x4_pred_mode_cache[scan8[i]];
 640                 if (transform_bypass && h->ps.sps->profile_idc == 244 && dir <= 1) {
 641                     if (h->x264_build < 151U) {
 642                         h->hpc.pred8x8l_add[dir](ptr, sl->mb + (i * 16 + p * 256 << pixel_shift), linesize);
 643                     } else
 644                         h->hpc.pred8x8l_filter_add[dir](ptr, sl->mb + (i * 16 + p * 256 << pixel_shift),
 645                                                         (sl-> topleft_samples_available << i) & 0x8000,
 646                                                         (sl->topright_samples_available << i) & 0x4000, linesize);
 647                 } else {
 648                     const int nnz = sl->non_zero_count_cache[scan8[i + p * 16]];
 649                     h->hpc.pred8x8l[dir](ptr, (sl->topleft_samples_available << i) & 0x8000,
 650                                          (sl->topright_samples_available << i) & 0x4000, linesize);
 651                     if (nnz) {
 652                         if (nnz == 1 && dctcoef_get(sl->mb, pixel_shift, i * 16 + p * 256))
 653                             idct_dc_add(ptr, sl->mb + (i * 16 + p * 256 << pixel_shift), linesize);
 654                         else
 655                             idct_add(ptr, sl->mb + (i * 16 + p * 256 << pixel_shift), linesize);
 656                     }
 657                 }
 658             }
 659         } else {
 660             if (transform_bypass) {
 661                 idct_dc_add  =
 662                 idct_add     = h->h264dsp.h264_add_pixels4_clear;
 663             } else {
 664                 idct_dc_add = h->h264dsp.h264_idct_dc_add;
 665                 idct_add    = h->h264dsp.h264_idct_add;
 666             }
 667             for (i = 0; i < 16; i++) {
 668                 uint8_t *const ptr = dest_y + block_offset[i];
 669                 const int dir      = sl->intra4x4_pred_mode_cache[scan8[i]];
 670
 671                 if (transform_bypass && h->ps.sps->profile_idc == 244 && dir <= 1) {
 672                     h->hpc.pred4x4_add[dir](ptr, sl->mb + (i * 16 + p * 256 << pixel_shift), linesize);
 673                 } else {
 674                     uint8_t *topright;
 675                     int nnz, tr;
 676                     uint64_t tr_high;
 677                     if (dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED) {
 678                         const int topright_avail = (sl->topright_samples_available << i) & 0x8000;
 679                         av_assert2(sl->mb_y || linesize <= block_offset[i]);
 680                         if (!topright_avail) {
 681                             if (pixel_shift) {
 682                                 tr_high  = ((uint16_t *)ptr)[3 - linesize / 2] * 0x0001000100010001ULL;
 683                                 topright = (uint8_t *)&tr_high;
 684                             } else {
 685                                 tr       = ptr[3 - linesize] * 0x01010101u;
 686                                 topright = (uint8_t *)&tr;
 687                             }
 688                         } else
 689                             topright = ptr + (4 << pixel_shift) - linesize;
 690                     } else
 691                         topright = NULL;
 692
 693                     h->hpc.pred4x4[dir](ptr, topright, linesize);
 694                     nnz = sl->non_zero_count_cache[scan8[i + p * 16]];
 695                     if (nnz) {
 696                         if (nnz == 1 && dctcoef_get(sl->mb, pixel_shift, i * 16 + p * 256))
 697                             idct_dc_add(ptr, sl->mb + (i * 16 + p * 256 << pixel_shift), linesize);
 698                         else
 699                             idct_add(ptr, sl->mb + (i * 16 + p * 256 << pixel_shift), linesize);
 700                     }
 701                 }
 702             }
 703         }
 704     } else {
 705         h->hpc.pred16x16[sl->intra16x16_pred_mode](dest_y, linesize);
 706         if (sl->non_zero_count_cache[scan8[LUMA_DC_BLOCK_INDEX + p]]) {
 707             if (!transform_bypass)
 708                 h->h264dsp.h264_luma_dc_dequant_idct(sl->mb + (p * 256 << pixel_shift),
 709                                                      sl->mb_luma_dc[p],
 710                                                      h->ps.pps->dequant4_coeff[p][qscale][0]);
 711             else {
 712                 static const uint8_t dc_mapping[16] = {
 713                      0 * 16,  1 * 16,  4 * 16,  5 * 16,
 714                      2 * 16,  3 * 16,  6 * 16,  7 * 16,
 715                      8 * 16,  9 * 16, 12 * 16, 13 * 16,
 716                     10 * 16, 11 * 16, 14 * 16, 15 * 16
 717                 };
 718                 for (i = 0; i < 16; i++)
 719                     dctcoef_set(sl->mb + (p * 256 << pixel_shift),
 720                                 pixel_shift, dc_mapping[i],
 721                                 dctcoef_get(sl->mb_luma_dc[p],
 722                                             pixel_shift, i));
 723             }
 724         }
 725     }
 726 }
 727
 728 static av_always_inline void hl_decode_mb_idct_luma(const H264Context *h, H264SliceContext *sl,
 729                                                     int mb_type, int simple,
 730                                                     int transform_bypass,
 731                                                     int pixel_shift,
 732                                                     const int *block_offset,
 733                                                     int linesize,
 734                                                     uint8_t *dest_y, int p)
 735 {
 736     void (*idct_add)(uint8_t *dst, int16_t *block, int stride);
 737     int i;
 738     block_offset += 16 * p;
 739     if (!IS_INTRA4x4(mb_type)) {
 740         if (IS_INTRA16x16(mb_type)) {
 741             if (transform_bypass) {
 742                 if (h->ps.sps->profile_idc == 244 &&
 743                     (sl->intra16x16_pred_mode == VERT_PRED8x8 ||
 744                      sl->intra16x16_pred_mode == HOR_PRED8x8)) {
 745                     h->hpc.pred16x16_add[sl->intra16x16_pred_mode](dest_y, block_offset,
 746                                                                    sl->mb + (p * 256 << pixel_shift),
 747                                                                    linesize);
 748                 } else {
 749                     for (i = 0; i < 16; i++)
 750                         if (sl->non_zero_count_cache[scan8[i + p * 16]] ||
 751                             dctcoef_get(sl->mb, pixel_shift, i * 16 + p * 256))
 752                             h->h264dsp.h264_add_pixels4_clear(dest_y + block_offset[i],
 753                                                               sl->mb + (i * 16 + p * 256 << pixel_shift),
 754                                                               linesize);
 755                 }
 756             } else {
 757                 h->h264dsp.h264_idct_add16intra(dest_y, block_offset,
 758                                                 sl->mb + (p * 256 << pixel_shift),
 759                                                 linesize,
 760                                                 sl->non_zero_count_cache + p * 5 * 8);
 761             }
 762         } else if (sl->cbp & 15) {
 763             if (transform_bypass) {
 764                 const int di = IS_8x8DCT(mb_type) ? 4 : 1;
 765                 idct_add = IS_8x8DCT(mb_type) ? h->h264dsp.h264_add_pixels8_clear
 766                     : h->h264dsp.h264_add_pixels4_clear;
 767                 for (i = 0; i < 16; i += di)
 768                     if (sl->non_zero_count_cache[scan8[i + p * 16]])
 769                         idct_add(dest_y + block_offset[i],
 770                                  sl->mb + (i * 16 + p * 256 << pixel_shift),
 771                                  linesize);
 772             } else {
 773                 if (IS_8x8DCT(mb_type))
 774                     h->h264dsp.h264_idct8_add4(dest_y, block_offset,
 775                                                sl->mb + (p * 256 << pixel_shift),
 776                                                linesize,
 777                                                sl->non_zero_count_cache + p * 5 * 8);
 778                 else
 779                     h->h264dsp.h264_idct_add16(dest_y, block_offset,
 780                                                sl->mb + (p * 256 << pixel_shift),
 781                                                linesize,
 782                                                sl->non_zero_count_cache + p * 5 * 8);
 783             }
 784         }
 785     }
 786 }
 787
 788 #define BITS   8
 789 #define SIMPLE 1
 790 #include "h264_mb_template.c"
 791
 792 #undef  BITS
 793 #define BITS   16
 794 #include "h264_mb_template.c"
 795
 796 #undef  SIMPLE
 797 #define SIMPLE 0
 798 #include "h264_mb_template.c"
 799
 800 void ff_h264_hl_decode_mb(const H264Context *h, H264SliceContext *sl)
 801 {
 802     const int mb_xy   = sl->mb_xy;
 803     const int mb_type = h->cur_pic.mb_type[mb_xy];
 804     int is_complex    = CONFIG_SMALL || sl->is_complex ||
 805                         IS_INTRA_PCM(mb_type) || sl->qscale == 0;
 806
 807     if (CHROMA444(h)) {
 808         if (is_complex || h->pixel_shift)
 809             hl_decode_mb_444_complex(h, sl);
 810         else
 811             hl_decode_mb_444_simple_8(h, sl);
 812     } else if (is_complex) {
 813         hl_decode_mb_complex(h, sl);
 814     } else if (h->pixel_shift) {
 815         hl_decode_mb_simple_16(h, sl);
 816     } else
 817         hl_decode_mb_simple_8(h, sl);
 818 }