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