git.sesse.net Git - ffmpeg/blob - libavcodec/h264.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 / MPEG4 part10 codec.
  25  * @author Michael Niedermayer <michaelni@gmx.at>
  26  */
  27
  28 #define UNCHECKED_BITSTREAM_READER 1
  29
  30 #include "libavutil/imgutils.h"
  31 #include "libavutil/opt.h"
  32 #include "internal.h"
  33 #include "cabac.h"
  34 #include "cabac_functions.h"
  35 #include "dsputil.h"
  36 #include "avcodec.h"
  37 #include "mpegvideo.h"
  38 #include "h264.h"
  39 #include "h264data.h"
  40 #include "h264_mvpred.h"
  41 #include "golomb.h"
  42 #include "mathops.h"
  43 #include "rectangle.h"
  44 #include "thread.h"
  45 #include "vdpau_internal.h"
  46 #include "libavutil/avassert.h"
  47
  48 // #undef NDEBUG
  49 #include <assert.h>
  50
  51 static const uint8_t rem6[QP_MAX_NUM + 1] = {
  52     0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2,
  53     3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5,
  54     0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
  55 };
  56
  57 static const uint8_t div6[QP_MAX_NUM + 1] = {
  58     0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3,  3,  3,
  59     3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6,  6,  6,
  60     7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 10, 10, 10, 10,
  61 };
  62
  63 static const enum PixelFormat hwaccel_pixfmt_list_h264_jpeg_420[] = {
  64     PIX_FMT_DXVA2_VLD,
  65     PIX_FMT_VAAPI_VLD,
  66     PIX_FMT_VDA_VLD,
  67     PIX_FMT_YUVJ420P,
  68     PIX_FMT_NONE
  69 };
  70
  71 /**
  72  * Check if the top & left blocks are available if needed and
  73  * change the dc mode so it only uses the available blocks.
  74  */
  75 int ff_h264_check_intra4x4_pred_mode(H264Context *h)
  76 {
  77     MpegEncContext *const s     = &h->s;
  78     static const int8_t top[12] = {
  79         -1, 0, LEFT_DC_PRED, -1, -1, -1, -1, -1, 0
  80     };
  81     static const int8_t left[12] = {
  82         0, -1, TOP_DC_PRED, 0, -1, -1, -1, 0, -1, DC_128_PRED
  83     };
  84     int i;
  85
  86     if (!(h->top_samples_available & 0x8000)) {
  87         for (i = 0; i < 4; i++) {
  88             int status = top[h->intra4x4_pred_mode_cache[scan8[0] + i]];
  89             if (status < 0) {
  90                 av_log(h->s.avctx, AV_LOG_ERROR,
  91                        "top block unavailable for requested intra4x4 mode %d at %d %d\n",
  92                        status, s->mb_x, s->mb_y);
  93                 return -1;
  94             } else if (status) {
  95                 h->intra4x4_pred_mode_cache[scan8[0] + i] = status;
  96             }
  97         }
  98     }
  99
 100     if ((h->left_samples_available & 0x8888) != 0x8888) {
 101         static const int mask[4] = { 0x8000, 0x2000, 0x80, 0x20 };
 102         for (i = 0; i < 4; i++)
 103             if (!(h->left_samples_available & mask[i])) {
 104                 int status = left[h->intra4x4_pred_mode_cache[scan8[0] + 8 * i]];
 105                 if (status < 0) {
 106                     av_log(h->s.avctx, AV_LOG_ERROR,
 107                            "left block unavailable for requested intra4x4 mode %d at %d %d\n",
 108                            status, s->mb_x, s->mb_y);
 109                     return -1;
 110                 } else if (status) {
 111                     h->intra4x4_pred_mode_cache[scan8[0] + 8 * i] = status;
 112                 }
 113             }
 114     }
 115
 116     return 0;
 117 } // FIXME cleanup like ff_h264_check_intra_pred_mode
 118
 119 /**
 120  * Check if the top & left blocks are available if needed and
 121  * change the dc mode so it only uses the available blocks.
 122  */
 123 int ff_h264_check_intra_pred_mode(H264Context *h, int mode, int is_chroma)
 124 {
 125     MpegEncContext *const s     = &h->s;
 126     static const int8_t top[7]  = { LEFT_DC_PRED8x8, 1, -1, -1 };
 127     static const int8_t left[7] = { TOP_DC_PRED8x8, -1, 2, -1, DC_128_PRED8x8 };
 128
 129     if (mode > 6U) {
 130         av_log(h->s.avctx, AV_LOG_ERROR,
 131                "out of range intra chroma pred mode at %d %d\n",
 132                s->mb_x, s->mb_y);
 133         return -1;
 134     }
 135
 136     if (!(h->top_samples_available & 0x8000)) {
 137         mode = top[mode];
 138         if (mode < 0) {
 139             av_log(h->s.avctx, AV_LOG_ERROR,
 140                    "top block unavailable for requested intra mode at %d %d\n",
 141                    s->mb_x, s->mb_y);
 142             return -1;
 143         }
 144     }
 145
 146     if ((h->left_samples_available & 0x8080) != 0x8080) {
 147         mode = left[mode];
 148         if (is_chroma && (h->left_samples_available & 0x8080)) {
 149             // mad cow disease mode, aka MBAFF + constrained_intra_pred
 150             mode = ALZHEIMER_DC_L0T_PRED8x8 +
 151                    (!(h->left_samples_available & 0x8000)) +
 152                    2 * (mode == DC_128_PRED8x8);
 153         }
 154         if (mode < 0) {
 155             av_log(h->s.avctx, AV_LOG_ERROR,
 156                    "left block unavailable for requested intra mode at %d %d\n",
 157                    s->mb_x, s->mb_y);
 158             return -1;
 159         }
 160     }
 161
 162     return mode;
 163 }
 164
 165 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src,
 166                                   int *dst_length, int *consumed, int length)
 167 {
 168     int i, si, di;
 169     uint8_t *dst;
 170     int bufidx;
 171
 172     // src[0]&0x80; // forbidden bit
 173     h->nal_ref_idc   = src[0] >> 5;
 174     h->nal_unit_type = src[0] & 0x1F;
 175
 176     src++;
 177     length--;
 178
 179 #if HAVE_FAST_UNALIGNED
 180 #if HAVE_FAST_64BIT
 181 #define RS 7
 182     for (i = 0; i + 1 < length; i += 9) {
 183         if (!((~AV_RN64A(src + i) &
 184                (AV_RN64A(src + i) - 0x0100010001000101ULL)) &
 185               0x8000800080008080ULL))
 186 #else
 187 #define RS 3
 188     for (i = 0; i + 1 < length; i += 5) {
 189         if (!((~AV_RN32A(src + i) &
 190                (AV_RN32A(src + i) - 0x01000101U)) &
 191               0x80008080U))
 192 #endif
 193             continue;
 194         if (i > 0 && !src[i])
 195             i--;
 196         while (src[i])
 197             i++;
 198 #else
 199 #define RS 0
 200     for (i = 0; i + 1 < length; i += 2) {
 201         if (src[i])
 202             continue;
 203         if (i > 0 && src[i - 1] == 0)
 204             i--;
 205 #endif
 206         if (i + 2 < length && src[i + 1] == 0 && src[i + 2] <= 3) {
 207             if (src[i + 2] != 3) {
 208                 /* startcode, so we must be past the end */
 209                 length = i;
 210             }
 211             break;
 212         }
 213         i -= RS;
 214     }
 215
 216     // use second escape buffer for inter data
 217     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0;
 218
 219     si = h->rbsp_buffer_size[bufidx];
 220     av_fast_padded_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+MAX_MBPAIR_SIZE);
 221     dst = h->rbsp_buffer[bufidx];
 222
 223     if (dst == NULL)
 224         return NULL;
 225
 226     if(i>=length-1){ //no escaped 0
 227         *dst_length= length;
 228         *consumed= length+1; //+1 for the header
 229         if(h->s.avctx->flags2 & CODEC_FLAG2_FAST){
 230             return src;
 231         }else{
 232             memcpy(dst, src, length);
 233             return dst;
 234         }
 235     }
 236
 237     // printf("decoding esc\n");
 238     memcpy(dst, src, i);
 239     si = di = i;
 240     while (si + 2 < length) {
 241         // remove escapes (very rare 1:2^22)
 242         if (src[si + 2] > 3) {
 243             dst[di++] = src[si++];
 244             dst[di++] = src[si++];
 245         } else if (src[si] == 0 && src[si + 1] == 0) {
 246             if (src[si + 2] == 3) { // escape
 247                 dst[di++]  = 0;
 248                 dst[di++]  = 0;
 249                 si        += 3;
 250                 continue;
 251             } else // next start code
 252                 goto nsc;
 253         }
 254
 255         dst[di++] = src[si++];
 256     }
 257     while (si < length)
 258         dst[di++] = src[si++];
 259 nsc:
 260
 261     memset(dst + di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
 262
 263     *dst_length = di;
 264     *consumed   = si + 1; // +1 for the header
 265     /* FIXME store exact number of bits in the getbitcontext
 266      * (it is needed for decoding) */
 267     return dst;
 268 }
 269
 270 /**
 271  * Identify the exact end of the bitstream
 272  * @return the length of the trailing, or 0 if damaged
 273  */
 274 static int ff_h264_decode_rbsp_trailing(H264Context *h, const uint8_t *src)
 275 {
 276     int v = *src;
 277     int r;
 278
 279     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
 280
 281     for (r = 1; r < 9; r++) {
 282         if (v & 1)
 283             return r;
 284         v >>= 1;
 285     }
 286     return 0;
 287 }
 288
 289 static inline int get_lowest_part_list_y(H264Context *h, Picture *pic, int n,
 290                                          int height, int y_offset, int list)
 291 {
 292     int raw_my        = h->mv_cache[list][scan8[n]][1];
 293     int filter_height = (raw_my & 3) ? 2 : 0;
 294     int full_my       = (raw_my >> 2) + y_offset;
 295     int top           = full_my - filter_height;
 296     int bottom        = full_my + filter_height + height;
 297
 298     return FFMAX(abs(top), bottom);
 299 }
 300
 301 static inline void get_lowest_part_y(H264Context *h, int refs[2][48], int n,
 302                                      int height, int y_offset, int list0,
 303                                      int list1, int *nrefs)
 304 {
 305     MpegEncContext *const s = &h->s;
 306     int my;
 307
 308     y_offset += 16 * (s->mb_y >> MB_FIELD);
 309
 310     if (list0) {
 311         int ref_n    = h->ref_cache[0][scan8[n]];
 312         Picture *ref = &h->ref_list[0][ref_n];
 313
 314         // Error resilience puts the current picture in the ref list.
 315         // Don't try to wait on these as it will cause a deadlock.
 316         // Fields can wait on each other, though.
 317         if (ref->f.thread_opaque   != s->current_picture.f.thread_opaque ||
 318             (ref->f.reference & 3) != s->picture_structure) {
 319             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 0);
 320             if (refs[0][ref_n] < 0)
 321                 nrefs[0] += 1;
 322             refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
 323         }
 324     }
 325
 326     if (list1) {
 327         int ref_n    = h->ref_cache[1][scan8[n]];
 328         Picture *ref = &h->ref_list[1][ref_n];
 329
 330         if (ref->f.thread_opaque   != s->current_picture.f.thread_opaque ||
 331             (ref->f.reference & 3) != s->picture_structure) {
 332             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 1);
 333             if (refs[1][ref_n] < 0)
 334                 nrefs[1] += 1;
 335             refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
 336         }
 337     }
 338 }
 339
 340 /**
 341  * Wait until all reference frames are available for MC operations.
 342  *
 343  * @param h the H264 context
 344  */
 345 static void await_references(H264Context *h)
 346 {
 347     MpegEncContext *const s = &h->s;
 348     const int mb_xy   = h->mb_xy;
 349     const int mb_type = s->current_picture.f.mb_type[mb_xy];
 350     int refs[2][48];
 351     int nrefs[2] = { 0 };
 352     int ref, list;
 353
 354     memset(refs, -1, sizeof(refs));
 355
 356     if (IS_16X16(mb_type)) {
 357         get_lowest_part_y(h, refs, 0, 16, 0,
 358                           IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 359     } else if (IS_16X8(mb_type)) {
 360         get_lowest_part_y(h, refs, 0, 8, 0,
 361                           IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 362         get_lowest_part_y(h, refs, 8, 8, 8,
 363                           IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
 364     } else if (IS_8X16(mb_type)) {
 365         get_lowest_part_y(h, refs, 0, 16, 0,
 366                           IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 367         get_lowest_part_y(h, refs, 4, 16, 0,
 368                           IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
 369     } else {
 370         int i;
 371
 372         assert(IS_8X8(mb_type));
 373
 374         for (i = 0; i < 4; i++) {
 375             const int sub_mb_type = h->sub_mb_type[i];
 376             const int n           = 4 * i;
 377             int y_offset          = (i & 2) << 2;
 378
 379             if (IS_SUB_8X8(sub_mb_type)) {
 380                 get_lowest_part_y(h, refs, n, 8, y_offset,
 381                                   IS_DIR(sub_mb_type, 0, 0),
 382                                   IS_DIR(sub_mb_type, 0, 1),
 383                                   nrefs);
 384             } else if (IS_SUB_8X4(sub_mb_type)) {
 385                 get_lowest_part_y(h, refs, n, 4, y_offset,
 386                                   IS_DIR(sub_mb_type, 0, 0),
 387                                   IS_DIR(sub_mb_type, 0, 1),
 388                                   nrefs);
 389                 get_lowest_part_y(h, refs, n + 2, 4, y_offset + 4,
 390                                   IS_DIR(sub_mb_type, 0, 0),
 391                                   IS_DIR(sub_mb_type, 0, 1),
 392                                   nrefs);
 393             } else if (IS_SUB_4X8(sub_mb_type)) {
 394                 get_lowest_part_y(h, refs, n, 8, y_offset,
 395                                   IS_DIR(sub_mb_type, 0, 0),
 396                                   IS_DIR(sub_mb_type, 0, 1),
 397                                   nrefs);
 398                 get_lowest_part_y(h, refs, n + 1, 8, y_offset,
 399                                   IS_DIR(sub_mb_type, 0, 0),
 400                                   IS_DIR(sub_mb_type, 0, 1),
 401                                   nrefs);
 402             } else {
 403                 int j;
 404                 assert(IS_SUB_4X4(sub_mb_type));
 405                 for (j = 0; j < 4; j++) {
 406                     int sub_y_offset = y_offset + 2 * (j & 2);
 407                     get_lowest_part_y(h, refs, n + j, 4, sub_y_offset,
 408                                       IS_DIR(sub_mb_type, 0, 0),
 409                                       IS_DIR(sub_mb_type, 0, 1),
 410                                       nrefs);
 411                 }
 412             }
 413         }
 414     }
 415
 416     for (list = h->list_count - 1; list >= 0; list--)
 417         for (ref = 0; ref < 48 && nrefs[list]; ref++) {
 418             int row = refs[list][ref];
 419             if (row >= 0) {
 420                 Picture *ref_pic      = &h->ref_list[list][ref];
 421                 int ref_field         = ref_pic->f.reference - 1;
 422                 int ref_field_picture = ref_pic->field_picture;
 423                 int pic_height        = 16 * s->mb_height >> ref_field_picture;
 424
 425                 row <<= MB_MBAFF;
 426                 nrefs[list]--;
 427
 428                 if (!FIELD_PICTURE && ref_field_picture) { // frame referencing two fields
 429                     ff_thread_await_progress(&ref_pic->f,
 430                                              FFMIN((row >> 1) - !(row & 1),
 431                                                    pic_height - 1),
 432                                              1);
 433                     ff_thread_await_progress(&ref_pic->f,
 434                                              FFMIN((row >> 1), pic_height - 1),
 435                                              0);
 436                 } else if (FIELD_PICTURE && !ref_field_picture) { // field referencing one field of a frame
 437                     ff_thread_await_progress(&ref_pic->f,
 438                                              FFMIN(row * 2 + ref_field,
 439                                                    pic_height - 1),
 440                                              0);
 441                 } else if (FIELD_PICTURE) {
 442                     ff_thread_await_progress(&ref_pic->f,
 443                                              FFMIN(row, pic_height - 1),
 444                                              ref_field);
 445                 } else {
 446                     ff_thread_await_progress(&ref_pic->f,
 447                                              FFMIN(row, pic_height - 1),
 448                                              0);
 449                 }
 450             }
 451         }
 452 }
 453
 454 static av_always_inline void mc_dir_part(H264Context *h, Picture *pic,
 455                                          int n, int square, int height,
 456                                          int delta, int list,
 457                                          uint8_t *dest_y, uint8_t *dest_cb,
 458                                          uint8_t *dest_cr,
 459                                          int src_x_offset, int src_y_offset,
 460                                          qpel_mc_func *qpix_op,
 461                                          h264_chroma_mc_func chroma_op,
 462                                          int pixel_shift, int chroma_idc)
 463 {
 464     MpegEncContext *const s = &h->s;
 465     const int mx      = h->mv_cache[list][scan8[n]][0] + src_x_offset * 8;
 466     int my            = h->mv_cache[list][scan8[n]][1] + src_y_offset * 8;
 467     const int luma_xy = (mx & 3) + ((my & 3) << 2);
 468     int offset        = ((mx >> 2) << pixel_shift) + (my >> 2) * h->mb_linesize;
 469     uint8_t *src_y    = pic->f.data[0] + offset;
 470     uint8_t *src_cb, *src_cr;
 471     int extra_width  = h->emu_edge_width;
 472     int extra_height = h->emu_edge_height;
 473     int emu = 0;
 474     const int full_mx    = mx >> 2;
 475     const int full_my    = my >> 2;
 476     const int pic_width  = 16 * s->mb_width;
 477     const int pic_height = 16 * s->mb_height >> MB_FIELD;
 478     int ysh;
 479
 480     if (mx & 7)
 481         extra_width -= 3;
 482     if (my & 7)
 483         extra_height -= 3;
 484
 485     if (full_mx                <          0 - extra_width  ||
 486         full_my                <          0 - extra_height ||
 487         full_mx + 16 /*FIXME*/ > pic_width  + extra_width  ||
 488         full_my + 16 /*FIXME*/ > pic_height + extra_height) {
 489         s->dsp.emulated_edge_mc(s->edge_emu_buffer,
 490                                 src_y - (2 << pixel_shift) - 2 * h->mb_linesize,
 491                                 h->mb_linesize,
 492                                 16 + 5, 16 + 5 /*FIXME*/, full_mx - 2,
 493                                 full_my - 2, pic_width, pic_height);
 494         src_y = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
 495         emu   = 1;
 496     }
 497
 498     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); // FIXME try variable height perhaps?
 499     if (!square)
 500         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
 501
 502     if (CONFIG_GRAY && s->flags & CODEC_FLAG_GRAY)
 503         return;
 504
 505     if (chroma_idc == 3 /* yuv444 */) {
 506         src_cb = pic->f.data[1] + offset;
 507         if (emu) {
 508             s->dsp.emulated_edge_mc(s->edge_emu_buffer,
 509                                     src_cb - (2 << pixel_shift) - 2 * h->mb_linesize,
 510                                     h->mb_linesize,
 511                                     16 + 5, 16 + 5 /*FIXME*/,
 512                                     full_mx - 2, full_my - 2,
 513                                     pic_width, pic_height);
 514             src_cb = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
 515         }
 516         qpix_op[luma_xy](dest_cb, src_cb, h->mb_linesize); // FIXME try variable height perhaps?
 517         if (!square)
 518             qpix_op[luma_xy](dest_cb + delta, src_cb + delta, h->mb_linesize);
 519
 520         src_cr = pic->f.data[2] + offset;
 521         if (emu) {
 522             s->dsp.emulated_edge_mc(s->edge_emu_buffer,
 523                                     src_cr - (2 << pixel_shift) - 2 * h->mb_linesize,
 524                                     h->mb_linesize,
 525                                     16 + 5, 16 + 5 /*FIXME*/,
 526                                     full_mx - 2, full_my - 2,
 527                                     pic_width, pic_height);
 528             src_cr = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
 529         }
 530         qpix_op[luma_xy](dest_cr, src_cr, h->mb_linesize); // FIXME try variable height perhaps?
 531         if (!square)
 532             qpix_op[luma_xy](dest_cr + delta, src_cr + delta, h->mb_linesize);
 533         return;
 534     }
 535
 536     ysh = 3 - (chroma_idc == 2 /* yuv422 */);
 537     if (chroma_idc == 1 /* yuv420 */ && MB_FIELD) {
 538         // chroma offset when predicting from a field of opposite parity
 539         my  += 2 * ((s->mb_y & 1) - (pic->f.reference - 1));
 540         emu |= (my >> 3) < 0 || (my >> 3) + 8 >= (pic_height >> 1);
 541     }
 542
 543     src_cb = pic->f.data[1] + ((mx >> 3) << pixel_shift) +
 544              (my >> ysh) * h->mb_uvlinesize;
 545     src_cr = pic->f.data[2] + ((mx >> 3) << pixel_shift) +
 546              (my >> ysh) * h->mb_uvlinesize;
 547
 548     if (emu) {
 549         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize,
 550                                 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
 551                                 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
 552         src_cb = s->edge_emu_buffer;
 553     }
 554     chroma_op(dest_cb, src_cb, h->mb_uvlinesize,
 555               height >> (chroma_idc == 1 /* yuv420 */),
 556               mx & 7, (my << (chroma_idc == 2 /* yuv422 */)) & 7);
 557
 558     if (emu) {
 559         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize,
 560                                 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
 561                                 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
 562         src_cr = s->edge_emu_buffer;
 563     }
 564     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, height >> (chroma_idc == 1 /* yuv420 */),
 565               mx & 7, (my << (chroma_idc == 2 /* yuv422 */)) & 7);
 566 }
 567
 568 static av_always_inline void mc_part_std(H264Context *h, int n, int square,
 569                                          int height, int delta,
 570                                          uint8_t *dest_y, uint8_t *dest_cb,
 571                                          uint8_t *dest_cr,
 572                                          int x_offset, int y_offset,
 573                                          qpel_mc_func *qpix_put,
 574                                          h264_chroma_mc_func chroma_put,
 575                                          qpel_mc_func *qpix_avg,
 576                                          h264_chroma_mc_func chroma_avg,
 577                                          int list0, int list1,
 578                                          int pixel_shift, int chroma_idc)
 579 {
 580     MpegEncContext *const s       = &h->s;
 581     qpel_mc_func *qpix_op         = qpix_put;
 582     h264_chroma_mc_func chroma_op = chroma_put;
 583
 584     dest_y += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
 585     if (chroma_idc == 3 /* yuv444 */) {
 586         dest_cb += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
 587         dest_cr += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
 588     } else if (chroma_idc == 2 /* yuv422 */) {
 589         dest_cb += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
 590         dest_cr += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
 591     } else { /* yuv420 */
 592         dest_cb += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
 593         dest_cr += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
 594     }
 595     x_offset += 8 * s->mb_x;
 596     y_offset += 8 * (s->mb_y >> MB_FIELD);
 597
 598     if (list0) {
 599         Picture *ref = &h->ref_list[0][h->ref_cache[0][scan8[n]]];
 600         mc_dir_part(h, ref, n, square, height, delta, 0,
 601                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
 602                     qpix_op, chroma_op, pixel_shift, chroma_idc);
 603
 604         qpix_op   = qpix_avg;
 605         chroma_op = chroma_avg;
 606     }
 607
 608     if (list1) {
 609         Picture *ref = &h->ref_list[1][h->ref_cache[1][scan8[n]]];
 610         mc_dir_part(h, ref, n, square, height, delta, 1,
 611                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
 612                     qpix_op, chroma_op, pixel_shift, chroma_idc);
 613     }
 614 }
 615
 616 static av_always_inline void mc_part_weighted(H264Context *h, int n, int square,
 617                                               int height, int delta,
 618                                               uint8_t *dest_y, uint8_t *dest_cb,
 619                                               uint8_t *dest_cr,
 620                                               int x_offset, int y_offset,
 621                                               qpel_mc_func *qpix_put,
 622                                               h264_chroma_mc_func chroma_put,
 623                                               h264_weight_func luma_weight_op,
 624                                               h264_weight_func chroma_weight_op,
 625                                               h264_biweight_func luma_weight_avg,
 626                                               h264_biweight_func chroma_weight_avg,
 627                                               int list0, int list1,
 628                                               int pixel_shift, int chroma_idc)
 629 {
 630     MpegEncContext *const s = &h->s;
 631     int chroma_height;
 632
 633     dest_y += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
 634     if (chroma_idc == 3 /* yuv444 */) {
 635         chroma_height     = height;
 636         chroma_weight_avg = luma_weight_avg;
 637         chroma_weight_op  = luma_weight_op;
 638         dest_cb += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
 639         dest_cr += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
 640     } else if (chroma_idc == 2 /* yuv422 */) {
 641         chroma_height = height;
 642         dest_cb      += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
 643         dest_cr      += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
 644     } else { /* yuv420 */
 645         chroma_height = height >> 1;
 646         dest_cb      += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
 647         dest_cr      += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
 648     }
 649     x_offset += 8 * s->mb_x;
 650     y_offset += 8 * (s->mb_y >> MB_FIELD);
 651
 652     if (list0 && list1) {
 653         /* don't optimize for luma-only case, since B-frames usually
 654          * use implicit weights => chroma too. */
 655         uint8_t *tmp_cb = s->obmc_scratchpad;
 656         uint8_t *tmp_cr = s->obmc_scratchpad + (16 << pixel_shift);
 657         uint8_t *tmp_y  = s->obmc_scratchpad + 16 * h->mb_uvlinesize;
 658         int refn0       = h->ref_cache[0][scan8[n]];
 659         int refn1       = h->ref_cache[1][scan8[n]];
 660
 661         mc_dir_part(h, &h->ref_list[0][refn0], n, square, height, delta, 0,
 662                     dest_y, dest_cb, dest_cr,
 663                     x_offset, y_offset, qpix_put, chroma_put,
 664                     pixel_shift, chroma_idc);
 665         mc_dir_part(h, &h->ref_list[1][refn1], n, square, height, delta, 1,
 666                     tmp_y, tmp_cb, tmp_cr,
 667                     x_offset, y_offset, qpix_put, chroma_put,
 668                     pixel_shift, chroma_idc);
 669
 670         if (h->use_weight == 2) {
 671             int weight0 = h->implicit_weight[refn0][refn1][s->mb_y & 1];
 672             int weight1 = 64 - weight0;
 673             luma_weight_avg(dest_y, tmp_y, h->mb_linesize,
 674                             height, 5, weight0, weight1, 0);
 675             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize,
 676                               chroma_height, 5, weight0, weight1, 0);
 677             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize,
 678                               chroma_height, 5, weight0, weight1, 0);
 679         } else {
 680             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, height,
 681                             h->luma_log2_weight_denom,
 682                             h->luma_weight[refn0][0][0],
 683                             h->luma_weight[refn1][1][0],
 684                             h->luma_weight[refn0][0][1] +
 685                             h->luma_weight[refn1][1][1]);
 686             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, chroma_height,
 687                               h->chroma_log2_weight_denom,
 688                               h->chroma_weight[refn0][0][0][0],
 689                               h->chroma_weight[refn1][1][0][0],
 690                               h->chroma_weight[refn0][0][0][1] +
 691                               h->chroma_weight[refn1][1][0][1]);
 692             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, chroma_height,
 693                               h->chroma_log2_weight_denom,
 694                               h->chroma_weight[refn0][0][1][0],
 695                               h->chroma_weight[refn1][1][1][0],
 696                               h->chroma_weight[refn0][0][1][1] +
 697                               h->chroma_weight[refn1][1][1][1]);
 698         }
 699     } else {
 700         int list     = list1 ? 1 : 0;
 701         int refn     = h->ref_cache[list][scan8[n]];
 702         Picture *ref = &h->ref_list[list][refn];
 703         mc_dir_part(h, ref, n, square, height, delta, list,
 704                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
 705                     qpix_put, chroma_put, pixel_shift, chroma_idc);
 706
 707         luma_weight_op(dest_y, h->mb_linesize, height,
 708                        h->luma_log2_weight_denom,
 709                        h->luma_weight[refn][list][0],
 710                        h->luma_weight[refn][list][1]);
 711         if (h->use_weight_chroma) {
 712             chroma_weight_op(dest_cb, h->mb_uvlinesize, chroma_height,
 713                              h->chroma_log2_weight_denom,
 714                              h->chroma_weight[refn][list][0][0],
 715                              h->chroma_weight[refn][list][0][1]);
 716             chroma_weight_op(dest_cr, h->mb_uvlinesize, chroma_height,
 717                              h->chroma_log2_weight_denom,
 718                              h->chroma_weight[refn][list][1][0],
 719                              h->chroma_weight[refn][list][1][1]);
 720         }
 721     }
 722 }
 723
 724 static av_always_inline void mc_part(H264Context *h, int n, int square,
 725                                      int height, int delta,
 726                                      uint8_t *dest_y, uint8_t *dest_cb,
 727                                      uint8_t *dest_cr,
 728                                      int x_offset, int y_offset,
 729                                      qpel_mc_func *qpix_put,
 730                                      h264_chroma_mc_func chroma_put,
 731                                      qpel_mc_func *qpix_avg,
 732                                      h264_chroma_mc_func chroma_avg,
 733                                      h264_weight_func *weight_op,
 734                                      h264_biweight_func *weight_avg,
 735                                      int list0, int list1,
 736                                      int pixel_shift, int chroma_idc)
 737 {
 738     if ((h->use_weight == 2 && list0 && list1 &&
 739          (h->implicit_weight[h->ref_cache[0][scan8[n]]][h->ref_cache[1][scan8[n]]][h->s.mb_y & 1] != 32)) ||
 740         h->use_weight == 1)
 741         mc_part_weighted(h, n, square, height, delta, dest_y, dest_cb, dest_cr,
 742                          x_offset, y_offset, qpix_put, chroma_put,
 743                          weight_op[0], weight_op[1], weight_avg[0],
 744                          weight_avg[1], list0, list1, pixel_shift, chroma_idc);
 745     else
 746         mc_part_std(h, n, square, height, delta, dest_y, dest_cb, dest_cr,
 747                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg,
 748                     chroma_avg, list0, list1, pixel_shift, chroma_idc);
 749 }
 750
 751 static av_always_inline void prefetch_motion(H264Context *h, int list,
 752                                              int pixel_shift, int chroma_idc)
 753 {
 754     /* fetch pixels for estimated mv 4 macroblocks ahead
 755      * optimized for 64byte cache lines */
 756     MpegEncContext *const s = &h->s;
 757     const int refn = h->ref_cache[list][scan8[0]];
 758     if (refn >= 0) {
 759         const int mx  = (h->mv_cache[list][scan8[0]][0] >> 2) + 16 * s->mb_x + 8;
 760         const int my  = (h->mv_cache[list][scan8[0]][1] >> 2) + 16 * s->mb_y;
 761         uint8_t **src = h->ref_list[list][refn].f.data;
 762         int off       = (mx << pixel_shift) +
 763                         (my + (s->mb_x & 3) * 4) * h->mb_linesize +
 764                         (64 << pixel_shift);
 765         s->dsp.prefetch(src[0] + off, s->linesize, 4);
 766         if (chroma_idc == 3 /* yuv444 */) {
 767             s->dsp.prefetch(src[1] + off, s->linesize, 4);
 768             s->dsp.prefetch(src[2] + off, s->linesize, 4);
 769         } else {
 770             off= (((mx>>1)+64)<<pixel_shift) + ((my>>1) + (s->mb_x&7))*s->uvlinesize;
 771             s->dsp.prefetch(src[1] + off, src[2] - src[1], 2);
 772         }
 773     }
 774 }
 775
 776 static av_always_inline void hl_motion(H264Context *h, uint8_t *dest_y,
 777                                        uint8_t *dest_cb, uint8_t *dest_cr,
 778                                        qpel_mc_func(*qpix_put)[16],
 779                                        h264_chroma_mc_func(*chroma_put),
 780                                        qpel_mc_func(*qpix_avg)[16],
 781                                        h264_chroma_mc_func(*chroma_avg),
 782                                        h264_weight_func *weight_op,
 783                                        h264_biweight_func *weight_avg,
 784                                        int pixel_shift, int chroma_idc)
 785 {
 786     MpegEncContext *const s = &h->s;
 787     const int mb_xy   = h->mb_xy;
 788     const int mb_type = s->current_picture.f.mb_type[mb_xy];
 789
 790     assert(IS_INTER(mb_type));
 791
 792     if (HAVE_THREADS && (s->avctx->active_thread_type & FF_THREAD_FRAME))
 793         await_references(h);
 794     prefetch_motion(h, 0, pixel_shift, chroma_idc);
 795
 796     if (IS_16X16(mb_type)) {
 797         mc_part(h, 0, 1, 16, 0, dest_y, dest_cb, dest_cr, 0, 0,
 798                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
 799                 weight_op, weight_avg,
 800                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 801                 pixel_shift, chroma_idc);
 802     } else if (IS_16X8(mb_type)) {
 803         mc_part(h, 0, 0, 8, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 0,
 804                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
 805                 weight_op, weight_avg,
 806                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 807                 pixel_shift, chroma_idc);
 808         mc_part(h, 8, 0, 8, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 4,
 809                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
 810                 weight_op, weight_avg,
 811                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
 812                 pixel_shift, chroma_idc);
 813     } else if (IS_8X16(mb_type)) {
 814         mc_part(h, 0, 0, 16, 8 * h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
 815                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 816                 &weight_op[1], &weight_avg[1],
 817                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 818                 pixel_shift, chroma_idc);
 819         mc_part(h, 4, 0, 16, 8 * h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
 820                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 821                 &weight_op[1], &weight_avg[1],
 822                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
 823                 pixel_shift, chroma_idc);
 824     } else {
 825         int i;
 826
 827         assert(IS_8X8(mb_type));
 828
 829         for (i = 0; i < 4; i++) {
 830             const int sub_mb_type = h->sub_mb_type[i];
 831             const int n  = 4 * i;
 832             int x_offset = (i & 1) << 2;
 833             int y_offset = (i & 2) << 1;
 834
 835             if (IS_SUB_8X8(sub_mb_type)) {
 836                 mc_part(h, n, 1, 8, 0, dest_y, dest_cb, dest_cr,
 837                         x_offset, y_offset,
 838                         qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 839                         &weight_op[1], &weight_avg[1],
 840                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 841                         pixel_shift, chroma_idc);
 842             } else if (IS_SUB_8X4(sub_mb_type)) {
 843                 mc_part(h, n, 0, 4, 4 << pixel_shift, dest_y, dest_cb, dest_cr,
 844                         x_offset, y_offset,
 845                         qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
 846                         &weight_op[1], &weight_avg[1],
 847                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 848                         pixel_shift, chroma_idc);
 849                 mc_part(h, n + 2, 0, 4, 4 << pixel_shift,
 850                         dest_y, dest_cb, dest_cr, x_offset, y_offset + 2,
 851                         qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
 852                         &weight_op[1], &weight_avg[1],
 853                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 854                         pixel_shift, chroma_idc);
 855             } else if (IS_SUB_4X8(sub_mb_type)) {
 856                 mc_part(h, n, 0, 8, 4 * h->mb_linesize,
 857                         dest_y, dest_cb, dest_cr, x_offset, y_offset,
 858                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 859                         &weight_op[2], &weight_avg[2],
 860                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 861                         pixel_shift, chroma_idc);
 862                 mc_part(h, n + 1, 0, 8, 4 * h->mb_linesize,
 863                         dest_y, dest_cb, dest_cr, x_offset + 2, y_offset,
 864                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 865                         &weight_op[2], &weight_avg[2],
 866                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 867                         pixel_shift, chroma_idc);
 868             } else {
 869                 int j;
 870                 assert(IS_SUB_4X4(sub_mb_type));
 871                 for (j = 0; j < 4; j++) {
 872                     int sub_x_offset = x_offset + 2 * (j & 1);
 873                     int sub_y_offset = y_offset + (j & 2);
 874                     mc_part(h, n + j, 1, 4, 0,
 875                             dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
 876                             qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 877                             &weight_op[2], &weight_avg[2],
 878                             IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 879                             pixel_shift, chroma_idc);
 880                 }
 881             }
 882         }
 883     }
 884
 885     prefetch_motion(h, 1, pixel_shift, chroma_idc);
 886 }
 887
 888 static av_always_inline void hl_motion_420(H264Context *h, uint8_t *dest_y,
 889                                            uint8_t *dest_cb, uint8_t *dest_cr,
 890                                            qpel_mc_func(*qpix_put)[16],
 891                                            h264_chroma_mc_func(*chroma_put),
 892                                            qpel_mc_func(*qpix_avg)[16],
 893                                            h264_chroma_mc_func(*chroma_avg),
 894                                            h264_weight_func *weight_op,
 895                                            h264_biweight_func *weight_avg,
 896                                            int pixel_shift)
 897 {
 898     hl_motion(h, dest_y, dest_cb, dest_cr, qpix_put, chroma_put,
 899               qpix_avg, chroma_avg, weight_op, weight_avg, pixel_shift, 1);
 900 }
 901
 902 static av_always_inline void hl_motion_422(H264Context *h, uint8_t *dest_y,
 903                                            uint8_t *dest_cb, uint8_t *dest_cr,
 904                                            qpel_mc_func(*qpix_put)[16],
 905                                            h264_chroma_mc_func(*chroma_put),
 906                                            qpel_mc_func(*qpix_avg)[16],
 907                                            h264_chroma_mc_func(*chroma_avg),
 908                                            h264_weight_func *weight_op,
 909                                            h264_biweight_func *weight_avg,
 910                                            int pixel_shift)
 911 {
 912     hl_motion(h, dest_y, dest_cb, dest_cr, qpix_put, chroma_put,
 913               qpix_avg, chroma_avg, weight_op, weight_avg, pixel_shift, 2);
 914 }
 915
 916 static void free_tables(H264Context *h, int free_rbsp)
 917 {
 918     int i;
 919     H264Context *hx;
 920
 921     av_freep(&h->intra4x4_pred_mode);
 922     av_freep(&h->chroma_pred_mode_table);
 923     av_freep(&h->cbp_table);
 924     av_freep(&h->mvd_table[0]);
 925     av_freep(&h->mvd_table[1]);
 926     av_freep(&h->direct_table);
 927     av_freep(&h->non_zero_count);
 928     av_freep(&h->slice_table_base);
 929     h->slice_table = NULL;
 930     av_freep(&h->list_counts);
 931
 932     av_freep(&h->mb2b_xy);
 933     av_freep(&h->mb2br_xy);
 934
 935     for (i = 0; i < MAX_THREADS; i++) {
 936         hx = h->thread_context[i];
 937         if (!hx)
 938             continue;
 939         av_freep(&hx->top_borders[1]);
 940         av_freep(&hx->top_borders[0]);
 941         av_freep(&hx->s.obmc_scratchpad);
 942         if (free_rbsp) {
 943             av_freep(&hx->rbsp_buffer[1]);
 944             av_freep(&hx->rbsp_buffer[0]);
 945             hx->rbsp_buffer_size[0] = 0;
 946             hx->rbsp_buffer_size[1] = 0;
 947         }
 948         if (i)
 949             av_freep(&h->thread_context[i]);
 950     }
 951 }
 952
 953 static void init_dequant8_coeff_table(H264Context *h)
 954 {
 955     int i, j, q, x;
 956     const int max_qp = 51 + 6 * (h->sps.bit_depth_luma - 8);
 957
 958     for (i = 0; i < 6; i++) {
 959         h->dequant8_coeff[i] = h->dequant8_buffer[i];
 960         for (j = 0; j < i; j++)
 961             if (!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i],
 962                         64 * sizeof(uint8_t))) {
 963                 h->dequant8_coeff[i] = h->dequant8_buffer[j];
 964                 break;
 965             }
 966         if (j < i)
 967             continue;
 968
 969         for (q = 0; q < max_qp + 1; q++) {
 970             int shift = div6[q];
 971             int idx   = rem6[q];
 972             for (x = 0; x < 64; x++)
 973                 h->dequant8_coeff[i][q][(x >> 3) | ((x & 7) << 3)] =
 974                     ((uint32_t)dequant8_coeff_init[idx][dequant8_coeff_init_scan[((x >> 1) & 12) | (x & 3)]] *
 975                      h->pps.scaling_matrix8[i][x]) << shift;
 976         }
 977     }
 978 }
 979
 980 static void init_dequant4_coeff_table(H264Context *h)
 981 {
 982     int i, j, q, x;
 983     const int max_qp = 51 + 6 * (h->sps.bit_depth_luma - 8);
 984     for (i = 0; i < 6; i++) {
 985         h->dequant4_coeff[i] = h->dequant4_buffer[i];
 986         for (j = 0; j < i; j++)
 987             if (!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i],
 988                         16 * sizeof(uint8_t))) {
 989                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
 990                 break;
 991             }
 992         if (j < i)
 993             continue;
 994
 995         for (q = 0; q < max_qp + 1; q++) {
 996             int shift = div6[q] + 2;
 997             int idx   = rem6[q];
 998             for (x = 0; x < 16; x++)
 999                 h->dequant4_coeff[i][q][(x >> 2) | ((x << 2) & 0xF)] =
1000                     ((uint32_t)dequant4_coeff_init[idx][(x & 1) + ((x >> 2) & 1)] *
1001                      h->pps.scaling_matrix4[i][x]) << shift;
1002         }
1003     }
1004 }
1005
1006 static void init_dequant_tables(H264Context *h)
1007 {
1008     int i, x;
1009     init_dequant4_coeff_table(h);
1010     if (h->pps.transform_8x8_mode)
1011         init_dequant8_coeff_table(h);
1012     if (h->sps.transform_bypass) {
1013         for (i = 0; i < 6; i++)
1014             for (x = 0; x < 16; x++)
1015                 h->dequant4_coeff[i][0][x] = 1 << 6;
1016         if (h->pps.transform_8x8_mode)
1017             for (i = 0; i < 6; i++)
1018                 for (x = 0; x < 64; x++)
1019                     h->dequant8_coeff[i][0][x] = 1 << 6;
1020     }
1021 }
1022
1023 int ff_h264_alloc_tables(H264Context *h)
1024 {
1025     MpegEncContext *const s = &h->s;
1026     const int big_mb_num    = s->mb_stride * (s->mb_height + 1);
1027     const int row_mb_num    = 2*s->mb_stride*FFMAX(s->avctx->thread_count, 1);
1028     int x, y;
1029
1030     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode,
1031                       row_mb_num * 8 * sizeof(uint8_t), fail)
1032     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count,
1033                       big_mb_num * 48 * sizeof(uint8_t), fail)
1034     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base,
1035                       (big_mb_num + s->mb_stride) * sizeof(*h->slice_table_base), fail)
1036     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table,
1037                       big_mb_num * sizeof(uint16_t), fail)
1038     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table,
1039                       big_mb_num * sizeof(uint8_t), fail)
1040     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0],
1041                       16 * row_mb_num * sizeof(uint8_t), fail);
1042     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1],
1043                       16 * row_mb_num * sizeof(uint8_t), fail);
1044     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table,
1045                       4 * big_mb_num * sizeof(uint8_t), fail);
1046     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts,
1047                       big_mb_num * sizeof(uint8_t), fail)
1048
1049     memset(h->slice_table_base, -1,
1050            (big_mb_num + s->mb_stride) * sizeof(*h->slice_table_base));
1051     h->slice_table = h->slice_table_base + s->mb_stride * 2 + 1;
1052
1053     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy,
1054                       big_mb_num * sizeof(uint32_t), fail);
1055     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy,
1056                       big_mb_num * sizeof(uint32_t), fail);
1057     for (y = 0; y < s->mb_height; y++)
1058         for (x = 0; x < s->mb_width; x++) {
1059             const int mb_xy = x + y * s->mb_stride;
1060             const int b_xy  = 4 * x + 4 * y * h->b_stride;
1061
1062             h->mb2b_xy[mb_xy]  = b_xy;
1063             h->mb2br_xy[mb_xy] = 8 * (FMO ? mb_xy : (mb_xy % (2 * s->mb_stride)));
1064         }
1065
1066     s->obmc_scratchpad = NULL;
1067
1068     if (!h->dequant4_coeff[0])
1069         init_dequant_tables(h);
1070
1071     return 0;
1072
1073 fail:
1074     free_tables(h, 1);
1075     return -1;
1076 }
1077
1078 /**
1079  * Mimic alloc_tables(), but for every context thread.
1080  */
1081 static void clone_tables(H264Context *dst, H264Context *src, int i)
1082 {
1083     MpegEncContext *const s     = &src->s;
1084     dst->intra4x4_pred_mode     = src->intra4x4_pred_mode + i * 8 * 2 * s->mb_stride;
1085     dst->non_zero_count         = src->non_zero_count;
1086     dst->slice_table            = src->slice_table;
1087     dst->cbp_table              = src->cbp_table;
1088     dst->mb2b_xy                = src->mb2b_xy;
1089     dst->mb2br_xy               = src->mb2br_xy;
1090     dst->chroma_pred_mode_table = src->chroma_pred_mode_table;
1091     dst->mvd_table[0]           = src->mvd_table[0] + i * 8 * 2 * s->mb_stride;
1092     dst->mvd_table[1]           = src->mvd_table[1] + i * 8 * 2 * s->mb_stride;
1093     dst->direct_table           = src->direct_table;
1094     dst->list_counts            = src->list_counts;
1095     dst->s.obmc_scratchpad      = NULL;
1096     ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma,
1097                       src->sps.chroma_format_idc);
1098 }
1099
1100 /**
1101  * Init context
1102  * Allocate buffers which are not shared amongst multiple threads.
1103  */
1104 static int context_init(H264Context *h)
1105 {
1106     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0],
1107                       h->s.mb_width * 16 * 3 * sizeof(uint8_t) * 2, fail)
1108     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1],
1109                       h->s.mb_width * 16 * 3 * sizeof(uint8_t) * 2, fail)
1110
1111     h->ref_cache[0][scan8[5]  + 1] =
1112     h->ref_cache[0][scan8[7]  + 1] =
1113     h->ref_cache[0][scan8[13] + 1] =
1114     h->ref_cache[1][scan8[5]  + 1] =
1115     h->ref_cache[1][scan8[7]  + 1] =
1116     h->ref_cache[1][scan8[13] + 1] = PART_NOT_AVAILABLE;
1117
1118     return 0;
1119
1120 fail:
1121     return -1; // free_tables will clean up for us
1122 }
1123
1124 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
1125
1126 static av_cold void common_init(H264Context *h)
1127 {
1128     MpegEncContext *const s = &h->s;
1129
1130     s->width    = s->avctx->width;
1131     s->height   = s->avctx->height;
1132     s->codec_id = s->avctx->codec->id;
1133
1134     s->avctx->bits_per_raw_sample = 8;
1135     h->cur_chroma_format_idc = 1;
1136
1137     ff_h264dsp_init(&h->h264dsp,
1138                     s->avctx->bits_per_raw_sample, h->cur_chroma_format_idc);
1139     ff_h264_pred_init(&h->hpc, s->codec_id,
1140                       s->avctx->bits_per_raw_sample, h->cur_chroma_format_idc);
1141
1142     h->dequant_coeff_pps = -1;
1143     s->unrestricted_mv   = 1;
1144
1145     s->dsp.dct_bits = 16;
1146     /* needed so that IDCT permutation is known early */
1147     ff_dsputil_init(&s->dsp, s->avctx);
1148
1149     memset(h->pps.scaling_matrix4, 16, 6 * 16 * sizeof(uint8_t));
1150     memset(h->pps.scaling_matrix8, 16, 2 * 64 * sizeof(uint8_t));
1151 }
1152
1153 int ff_h264_decode_extradata(H264Context *h, const uint8_t *buf, int size)
1154 {
1155     AVCodecContext *avctx = h->s.avctx;
1156
1157     if (!buf || size <= 0)
1158         return -1;
1159
1160     if (buf[0] == 1) {
1161         int i, cnt, nalsize;
1162         const unsigned char *p = buf;
1163
1164         h->is_avc = 1;
1165
1166         if (size < 7) {
1167             av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
1168             return -1;
1169         }
1170         /* sps and pps in the avcC always have length coded with 2 bytes,
1171          * so put a fake nal_length_size = 2 while parsing them */
1172         h->nal_length_size = 2;
1173         // Decode sps from avcC
1174         cnt = *(p + 5) & 0x1f; // Number of sps
1175         p  += 6;
1176         for (i = 0; i < cnt; i++) {
1177             nalsize = AV_RB16(p) + 2;
1178             if(nalsize > size - (p-buf))
1179                 return -1;
1180             if (decode_nal_units(h, p, nalsize) < 0) {
1181                 av_log(avctx, AV_LOG_ERROR,
1182                        "Decoding sps %d from avcC failed\n", i);
1183                 return -1;
1184             }
1185             p += nalsize;
1186         }
1187         // Decode pps from avcC
1188         cnt = *(p++); // Number of pps
1189         for (i = 0; i < cnt; i++) {
1190             nalsize = AV_RB16(p) + 2;
1191             if(nalsize > size - (p-buf))
1192                 return -1;
1193             if (decode_nal_units(h, p, nalsize) < 0) {
1194                 av_log(avctx, AV_LOG_ERROR,
1195                        "Decoding pps %d from avcC failed\n", i);
1196                 return -1;
1197             }
1198             p += nalsize;
1199         }
1200         // Now store right nal length size, that will be used to parse all other nals
1201         h->nal_length_size = (buf[4] & 0x03) + 1;
1202     } else {
1203         h->is_avc = 0;
1204         if (decode_nal_units(h, buf, size) < 0)
1205             return -1;
1206     }
1207     return size;
1208 }
1209
1210 av_cold int ff_h264_decode_init(AVCodecContext *avctx)
1211 {
1212     H264Context *h = avctx->priv_data;
1213     MpegEncContext *const s = &h->s;
1214     int i;
1215
1216     ff_MPV_decode_defaults(s);
1217
1218     s->avctx = avctx;
1219     common_init(h);
1220
1221     s->out_format      = FMT_H264;
1222     s->workaround_bugs = avctx->workaround_bugs;
1223
1224     /* set defaults */
1225     // s->decode_mb = ff_h263_decode_mb;
1226     s->quarter_sample = 1;
1227     if (!avctx->has_b_frames)
1228         s->low_delay = 1;
1229
1230     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1231
1232     ff_h264_decode_init_vlc();
1233
1234     h->pixel_shift = 0;
1235     h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1236
1237     h->thread_context[0] = h;
1238     h->outputed_poc      = h->next_outputed_poc = INT_MIN;
1239     for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
1240         h->last_pocs[i] = INT_MIN;
1241     h->prev_poc_msb = 1 << 16;
1242     h->prev_frame_num = -1;
1243     h->x264_build   = -1;
1244     ff_h264_reset_sei(h);
1245     if (avctx->codec_id == CODEC_ID_H264) {
1246         if (avctx->ticks_per_frame == 1)
1247             s->avctx->time_base.den *= 2;
1248         avctx->ticks_per_frame = 2;
1249     }
1250
1251     if (avctx->extradata_size > 0 && avctx->extradata &&
1252         ff_h264_decode_extradata(h, avctx->extradata, avctx->extradata_size) < 0)
1253         return -1;
1254
1255     if (h->sps.bitstream_restriction_flag &&
1256         s->avctx->has_b_frames < h->sps.num_reorder_frames) {
1257         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1258         s->low_delay           = 0;
1259     }
1260
1261     return 0;
1262 }
1263
1264 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b) + (size))))
1265
1266 static void copy_picture_range(Picture **to, Picture **from, int count,
1267                                MpegEncContext *new_base,
1268                                MpegEncContext *old_base)
1269 {
1270     int i;
1271
1272     for (i = 0; i < count; i++) {
1273         assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1274                 IN_RANGE(from[i], old_base->picture,
1275                          sizeof(Picture) * old_base->picture_count) ||
1276                 !from[i]));
1277         to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1278     }
1279 }
1280
1281 static void copy_parameter_set(void **to, void **from, int count, int size)
1282 {
1283     int i;
1284
1285     for (i = 0; i < count; i++) {
1286         if (to[i] && !from[i])
1287             av_freep(&to[i]);
1288         else if (from[i] && !to[i])
1289             to[i] = av_malloc(size);
1290
1291         if (from[i])
1292             memcpy(to[i], from[i], size);
1293     }
1294 }
1295
1296 static int decode_init_thread_copy(AVCodecContext *avctx)
1297 {
1298     H264Context *h = avctx->priv_data;
1299
1300     if (!avctx->internal->is_copy)
1301         return 0;
1302     memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1303     memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1304
1305     return 0;
1306 }
1307
1308 #define copy_fields(to, from, start_field, end_field)                   \
1309     memcpy(&to->start_field, &from->start_field,                        \
1310            (char *)&to->end_field - (char *)&to->start_field)
1311
1312 static int decode_update_thread_context(AVCodecContext *dst,
1313                                         const AVCodecContext *src)
1314 {
1315     H264Context *h = dst->priv_data, *h1 = src->priv_data;
1316     MpegEncContext *const s = &h->s, *const s1 = &h1->s;
1317     int inited = s->context_initialized, err;
1318     int i;
1319
1320     if (dst == src)
1321         return 0;
1322
1323     err = ff_mpeg_update_thread_context(dst, src);
1324     if (err)
1325         return err;
1326
1327     // FIXME handle width/height changing
1328     if (!inited) {
1329         for (i = 0; i < MAX_SPS_COUNT; i++)
1330             av_freep(h->sps_buffers + i);
1331
1332         for (i = 0; i < MAX_PPS_COUNT; i++)
1333             av_freep(h->pps_buffers + i);
1334
1335         // copy all fields after MpegEnc
1336         memcpy(&h->s + 1, &h1->s + 1,
1337                sizeof(H264Context) - sizeof(MpegEncContext));
1338         memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1339         memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1340
1341         if (s1->context_initialized) {
1342         if (ff_h264_alloc_tables(h) < 0) {
1343             av_log(dst, AV_LOG_ERROR, "Could not allocate memory for h264\n");
1344             return AVERROR(ENOMEM);
1345         }
1346         context_init(h);
1347
1348         /* frame_start may not be called for the next thread (if it's decoding
1349          * a bottom field) so this has to be allocated here */
1350         h->s.obmc_scratchpad = av_malloc(16 * 6 * s->linesize);
1351         }
1352
1353         for (i = 0; i < 2; i++) {
1354             h->rbsp_buffer[i]      = NULL;
1355             h->rbsp_buffer_size[i] = 0;
1356         }
1357
1358         h->thread_context[0] = h;
1359
1360         s->dsp.clear_blocks(h->mb);
1361         s->dsp.clear_blocks(h->mb + (24 * 16 << h->pixel_shift));
1362     }
1363
1364     // extradata/NAL handling
1365     h->is_avc = h1->is_avc;
1366
1367     // SPS/PPS
1368     copy_parameter_set((void **)h->sps_buffers, (void **)h1->sps_buffers,
1369                        MAX_SPS_COUNT, sizeof(SPS));
1370     h->sps = h1->sps;
1371     copy_parameter_set((void **)h->pps_buffers, (void **)h1->pps_buffers,
1372                        MAX_PPS_COUNT, sizeof(PPS));
1373     h->pps = h1->pps;
1374
1375     // Dequantization matrices
1376     // FIXME these are big - can they be only copied when PPS changes?
1377     copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1378
1379     for (i = 0; i < 6; i++)
1380         h->dequant4_coeff[i] = h->dequant4_buffer[0] +
1381                                (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1382
1383     for (i = 0; i < 6; i++)
1384         h->dequant8_coeff[i] = h->dequant8_buffer[0] +
1385                                (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1386
1387     h->dequant_coeff_pps = h1->dequant_coeff_pps;
1388
1389     // POC timing
1390     copy_fields(h, h1, poc_lsb, redundant_pic_count);
1391
1392     // reference lists
1393     copy_fields(h, h1, ref_count, list_count);
1394     copy_fields(h, h1, ref_list, intra_gb);
1395     copy_fields(h, h1, short_ref, cabac_init_idc);
1396
1397     copy_picture_range(h->short_ref, h1->short_ref, 32, s, s1);
1398     copy_picture_range(h->long_ref, h1->long_ref, 32, s, s1);
1399     copy_picture_range(h->delayed_pic, h1->delayed_pic,
1400                        MAX_DELAYED_PIC_COUNT + 2, s, s1);
1401
1402     h->last_slice_type = h1->last_slice_type;
1403     h->sync            = h1->sync;
1404
1405     if (!s->current_picture_ptr)
1406         return 0;
1407
1408     if (!s->dropable) {
1409         err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1410         h->prev_poc_msb = h->poc_msb;
1411         h->prev_poc_lsb = h->poc_lsb;
1412     }
1413     h->prev_frame_num_offset = h->frame_num_offset;
1414     h->prev_frame_num        = h->frame_num;
1415     h->outputed_poc          = h->next_outputed_poc;
1416
1417     return err;
1418 }
1419
1420 int ff_h264_frame_start(H264Context *h)
1421 {
1422     MpegEncContext *const s = &h->s;
1423     int i;
1424     const int pixel_shift = h->pixel_shift;
1425
1426     if (ff_MPV_frame_start(s, s->avctx) < 0)
1427         return -1;
1428     ff_er_frame_start(s);
1429     /*
1430      * ff_MPV_frame_start uses pict_type to derive key_frame.
1431      * This is incorrect for H.264; IDR markings must be used.
1432      * Zero here; IDR markings per slice in frame or fields are ORed in later.
1433      * See decode_nal_units().
1434      */
1435     s->current_picture_ptr->f.key_frame = 0;
1436     s->current_picture_ptr->sync        = 0;
1437     s->current_picture_ptr->mmco_reset  = 0;
1438
1439     assert(s->linesize && s->uvlinesize);
1440
1441     for (i = 0; i < 16; i++) {
1442         h->block_offset[i]           = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * s->linesize * ((scan8[i] - scan8[0]) >> 3);
1443         h->block_offset[48 + i]      = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * s->linesize * ((scan8[i] - scan8[0]) >> 3);
1444     }
1445     for (i = 0; i < 16; i++) {
1446         h->block_offset[16 + i]      =
1447         h->block_offset[32 + i]      = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * s->uvlinesize * ((scan8[i] - scan8[0]) >> 3);
1448         h->block_offset[48 + 16 + i] =
1449         h->block_offset[48 + 32 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * s->uvlinesize * ((scan8[i] - scan8[0]) >> 3);
1450     }
1451
1452     /* can't be in alloc_tables because linesize isn't known there.
1453      * FIXME: redo bipred weight to not require extra buffer? */
1454     for (i = 0; i < s->slice_context_count; i++)
1455         if (h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1456             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16 * 6 * s->linesize);
1457
1458     /* Some macroblocks can be accessed before they're available in case
1459      * of lost slices, MBAFF or threading. */
1460     memset(h->slice_table, -1,
1461            (s->mb_height * s->mb_stride - 1) * sizeof(*h->slice_table));
1462
1463     // s->decode = (s->flags & CODEC_FLAG_PSNR) || !s->encoding ||
1464     //             s->current_picture.f.reference /* || h->contains_intra */ || 1;
1465
1466     /* We mark the current picture as non-reference after allocating it, so
1467      * that if we break out due to an error it can be released automatically
1468      * in the next ff_MPV_frame_start().
1469      * SVQ3 as well as most other codecs have only last/next/current and thus
1470      * get released even with set reference, besides SVQ3 and others do not
1471      * mark frames as reference later "naturally". */
1472     if (s->codec_id != CODEC_ID_SVQ3)
1473         s->current_picture_ptr->f.reference = 0;
1474
1475     s->current_picture_ptr->field_poc[0]     =
1476         s->current_picture_ptr->field_poc[1] = INT_MAX;
1477
1478     h->next_output_pic = NULL;
1479
1480     assert(s->current_picture_ptr->long_ref == 0);
1481
1482     return 0;
1483 }
1484
1485 /**
1486  * Run setup operations that must be run after slice header decoding.
1487  * This includes finding the next displayed frame.
1488  *
1489  * @param h h264 master context
1490  * @param setup_finished enough NALs have been read that we can call
1491  * ff_thread_finish_setup()
1492  */
1493 static void decode_postinit(H264Context *h, int setup_finished)
1494 {
1495     MpegEncContext *const s = &h->s;
1496     Picture *out = s->current_picture_ptr;
1497     Picture *cur = s->current_picture_ptr;
1498     int i, pics, out_of_order, out_idx;
1499
1500     s->current_picture_ptr->f.qscale_type = FF_QSCALE_TYPE_H264;
1501     s->current_picture_ptr->f.pict_type   = s->pict_type;
1502
1503     if (h->next_output_pic)
1504         return;
1505
1506     if (cur->field_poc[0] == INT_MAX || cur->field_poc[1] == INT_MAX) {
1507         /* FIXME: if we have two PAFF fields in one packet, we can't start
1508          * the next thread here. If we have one field per packet, we can.
1509          * The check in decode_nal_units() is not good enough to find this
1510          * yet, so we assume the worst for now. */
1511         // if (setup_finished)
1512         //    ff_thread_finish_setup(s->avctx);
1513         return;
1514     }
1515
1516     cur->f.interlaced_frame = 0;
1517     cur->f.repeat_pict      = 0;
1518
1519     /* Signal interlacing information externally. */
1520     /* Prioritize picture timing SEI information over used
1521      * decoding process if it exists. */
1522
1523     if (h->sps.pic_struct_present_flag) {
1524         switch (h->sei_pic_struct) {
1525         case SEI_PIC_STRUCT_FRAME:
1526             break;
1527         case SEI_PIC_STRUCT_TOP_FIELD:
1528         case SEI_PIC_STRUCT_BOTTOM_FIELD:
1529             cur->f.interlaced_frame = 1;
1530             break;
1531         case SEI_PIC_STRUCT_TOP_BOTTOM:
1532         case SEI_PIC_STRUCT_BOTTOM_TOP:
1533             if (FIELD_OR_MBAFF_PICTURE)
1534                 cur->f.interlaced_frame = 1;
1535             else
1536                 // try to flag soft telecine progressive
1537                 cur->f.interlaced_frame = h->prev_interlaced_frame;
1538             break;
1539         case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1540         case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1541             /* Signal the possibility of telecined film externally
1542              * (pic_struct 5,6). From these hints, let the applications
1543              * decide if they apply deinterlacing. */
1544             cur->f.repeat_pict = 1;
1545             break;
1546         case SEI_PIC_STRUCT_FRAME_DOUBLING:
1547             // Force progressive here, doubling interlaced frame is a bad idea.
1548             cur->f.repeat_pict = 2;
1549             break;
1550         case SEI_PIC_STRUCT_FRAME_TRIPLING:
1551             cur->f.repeat_pict = 4;
1552             break;
1553         }
1554
1555         if ((h->sei_ct_type & 3) &&
1556             h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1557             cur->f.interlaced_frame = (h->sei_ct_type & (1 << 1)) != 0;
1558     } else {
1559         /* Derive interlacing flag from used decoding process. */
1560         cur->f.interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1561     }
1562     h->prev_interlaced_frame = cur->f.interlaced_frame;
1563
1564     if (cur->field_poc[0] != cur->field_poc[1]) {
1565         /* Derive top_field_first from field pocs. */
1566         cur->f.top_field_first = cur->field_poc[0] < cur->field_poc[1];
1567     } else {
1568         if (cur->f.interlaced_frame || h->sps.pic_struct_present_flag) {
1569             /* Use picture timing SEI information. Even if it is a
1570              * information of a past frame, better than nothing. */
1571             if (h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM ||
1572                 h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1573                 cur->f.top_field_first = 1;
1574             else
1575                 cur->f.top_field_first = 0;
1576         } else {
1577             /* Most likely progressive */
1578             cur->f.top_field_first = 0;
1579         }
1580     }
1581
1582     cur->mmco_reset = h->mmco_reset;
1583     h->mmco_reset = 0;
1584     // FIXME do something with unavailable reference frames
1585
1586     /* Sort B-frames into display order */
1587
1588     if (h->sps.bitstream_restriction_flag &&
1589         s->avctx->has_b_frames < h->sps.num_reorder_frames) {
1590         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1591         s->low_delay           = 0;
1592     }
1593
1594     if (s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT &&
1595         !h->sps.bitstream_restriction_flag) {
1596         s->avctx->has_b_frames = MAX_DELAYED_PIC_COUNT - 1;
1597         s->low_delay           = 0;
1598     }
1599
1600     for (i = 0; 1; i++) {
1601         if(i == MAX_DELAYED_PIC_COUNT || cur->poc < h->last_pocs[i]){
1602             if(i)
1603                 h->last_pocs[i-1] = cur->poc;
1604             break;
1605         } else if(i) {
1606             h->last_pocs[i-1]= h->last_pocs[i];
1607         }
1608     }
1609     out_of_order = MAX_DELAYED_PIC_COUNT - i;
1610     if(   cur->f.pict_type == AV_PICTURE_TYPE_B
1611        || (h->last_pocs[MAX_DELAYED_PIC_COUNT-2] > INT_MIN && h->last_pocs[MAX_DELAYED_PIC_COUNT-1] - h->last_pocs[MAX_DELAYED_PIC_COUNT-2] > 2))
1612         out_of_order = FFMAX(out_of_order, 1);
1613     if(s->avctx->has_b_frames < out_of_order && !h->sps.bitstream_restriction_flag){
1614         av_log(s->avctx, AV_LOG_WARNING, "Increasing reorder buffer to %d\n", out_of_order);
1615         s->avctx->has_b_frames = out_of_order;
1616         s->low_delay = 0;
1617     }
1618
1619     pics = 0;
1620     while (h->delayed_pic[pics])
1621         pics++;
1622
1623     av_assert0(pics <= MAX_DELAYED_PIC_COUNT);
1624
1625     h->delayed_pic[pics++] = cur;
1626     if (cur->f.reference == 0)
1627         cur->f.reference = DELAYED_PIC_REF;
1628
1629     out = h->delayed_pic[0];
1630     out_idx = 0;
1631     for (i = 1; h->delayed_pic[i] &&
1632                 !h->delayed_pic[i]->f.key_frame &&
1633                 !h->delayed_pic[i]->mmco_reset;
1634          i++)
1635         if (h->delayed_pic[i]->poc < out->poc) {
1636             out     = h->delayed_pic[i];
1637             out_idx = i;
1638         }
1639     if (s->avctx->has_b_frames == 0 &&
1640         (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset))
1641         h->next_outputed_poc = INT_MIN;
1642     out_of_order = out->poc < h->next_outputed_poc;
1643
1644     if (out_of_order || pics > s->avctx->has_b_frames) {
1645         out->f.reference &= ~DELAYED_PIC_REF;
1646         // for frame threading, the owner must be the second field's thread or
1647         // else the first thread can release the picture and reuse it unsafely
1648         out->owner2       = s;
1649         for (i = out_idx; h->delayed_pic[i]; i++)
1650             h->delayed_pic[i] = h->delayed_pic[i + 1];
1651     }
1652     if (!out_of_order && pics > s->avctx->has_b_frames) {
1653         h->next_output_pic = out;
1654         if (out_idx == 0 && h->delayed_pic[0] && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset)) {
1655             h->next_outputed_poc = INT_MIN;
1656         } else
1657             h->next_outputed_poc = out->poc;
1658     } else {
1659         av_log(s->avctx, AV_LOG_DEBUG, "no picture %s\n", out_of_order ? "ooo" : "");
1660     }
1661
1662     if (h->next_output_pic && h->next_output_pic->sync) {
1663         h->sync |= 2;
1664     }
1665
1666     if (setup_finished)
1667         ff_thread_finish_setup(s->avctx);
1668 }
1669
1670 static av_always_inline void backup_mb_border(H264Context *h, uint8_t *src_y,
1671                                               uint8_t *src_cb, uint8_t *src_cr,
1672                                               int linesize, int uvlinesize,
1673                                               int simple)
1674 {
1675     MpegEncContext *const s = &h->s;
1676     uint8_t *top_border;
1677     int top_idx = 1;
1678     const int pixel_shift = h->pixel_shift;
1679     int chroma444 = CHROMA444;
1680     int chroma422 = CHROMA422;
1681
1682     src_y  -= linesize;
1683     src_cb -= uvlinesize;
1684     src_cr -= uvlinesize;
1685
1686     if (!simple && FRAME_MBAFF) {
1687         if (s->mb_y & 1) {
1688             if (!MB_MBAFF) {
1689                 top_border = h->top_borders[0][s->mb_x];
1690                 AV_COPY128(top_border, src_y + 15 * linesize);
1691                 if (pixel_shift)
1692                     AV_COPY128(top_border + 16, src_y + 15 * linesize + 16);
1693                 if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1694                     if (chroma444) {
1695                         if (pixel_shift) {
1696                             AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
1697                             AV_COPY128(top_border + 48, src_cb + 15 * uvlinesize + 16);
1698                             AV_COPY128(top_border + 64, src_cr + 15 * uvlinesize);
1699                             AV_COPY128(top_border + 80, src_cr + 15 * uvlinesize + 16);
1700                         } else {
1701                             AV_COPY128(top_border + 16, src_cb + 15 * uvlinesize);
1702                             AV_COPY128(top_border + 32, src_cr + 15 * uvlinesize);
1703                         }
1704                     } else if (chroma422) {
1705                         if (pixel_shift) {
1706                             AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
1707                             AV_COPY128(top_border + 48, src_cr + 15 * uvlinesize);
1708                         } else {
1709                             AV_COPY64(top_border + 16, src_cb + 15 * uvlinesize);
1710                             AV_COPY64(top_border + 24, src_cr + 15 * uvlinesize);
1711                         }
1712                     } else {
1713                         if (pixel_shift) {
1714                             AV_COPY128(top_border + 32, src_cb + 7 * uvlinesize);
1715                             AV_COPY128(top_border + 48, src_cr + 7 * uvlinesize);
1716                         } else {
1717                             AV_COPY64(top_border + 16, src_cb + 7 * uvlinesize);
1718                             AV_COPY64(top_border + 24, src_cr + 7 * uvlinesize);
1719                         }
1720                     }
1721                 }
1722             }
1723         } else if (MB_MBAFF) {
1724             top_idx = 0;
1725         } else
1726             return;
1727     }
1728
1729     top_border = h->top_borders[top_idx][s->mb_x];
1730     /* There are two lines saved, the line above the the top macroblock
1731      * of a pair, and the line above the bottom macroblock. */
1732     AV_COPY128(top_border, src_y + 16 * linesize);
1733     if (pixel_shift)
1734         AV_COPY128(top_border + 16, src_y + 16 * linesize + 16);
1735
1736     if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1737         if (chroma444) {
1738             if (pixel_shift) {
1739                 AV_COPY128(top_border + 32, src_cb + 16 * linesize);
1740                 AV_COPY128(top_border + 48, src_cb + 16 * linesize + 16);
1741                 AV_COPY128(top_border + 64, src_cr + 16 * linesize);
1742                 AV_COPY128(top_border + 80, src_cr + 16 * linesize + 16);
1743             } else {
1744                 AV_COPY128(top_border + 16, src_cb + 16 * linesize);
1745                 AV_COPY128(top_border + 32, src_cr + 16 * linesize);
1746             }
1747         } else if (chroma422) {
1748             if (pixel_shift) {
1749                 AV_COPY128(top_border + 32, src_cb + 16 * uvlinesize);
1750                 AV_COPY128(top_border + 48, src_cr + 16 * uvlinesize);
1751             } else {
1752                 AV_COPY64(top_border + 16, src_cb + 16 * uvlinesize);
1753                 AV_COPY64(top_border + 24, src_cr + 16 * uvlinesize);
1754             }
1755         } else {
1756             if (pixel_shift) {
1757                 AV_COPY128(top_border + 32, src_cb + 8 * uvlinesize);
1758                 AV_COPY128(top_border + 48, src_cr + 8 * uvlinesize);
1759             } else {
1760                 AV_COPY64(top_border + 16, src_cb + 8 * uvlinesize);
1761                 AV_COPY64(top_border + 24, src_cr + 8 * uvlinesize);
1762             }
1763         }
1764     }
1765 }
1766
1767 static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1768                                             uint8_t *src_cb, uint8_t *src_cr,
1769                                             int linesize, int uvlinesize,
1770                                             int xchg, int chroma444,
1771                                             int simple, int pixel_shift)
1772 {
1773     MpegEncContext *const s = &h->s;
1774     int deblock_topleft;
1775     int deblock_top;
1776     int top_idx = 1;
1777     uint8_t *top_border_m1;
1778     uint8_t *top_border;
1779
1780     if (!simple && FRAME_MBAFF) {
1781         if (s->mb_y & 1) {
1782             if (!MB_MBAFF)
1783                 return;
1784         } else {
1785             top_idx = MB_MBAFF ? 0 : 1;
1786         }
1787     }
1788
1789     if (h->deblocking_filter == 2) {
1790         deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1791         deblock_top     = h->top_type;
1792     } else {
1793         deblock_topleft = (s->mb_x > 0);
1794         deblock_top     = (s->mb_y > !!MB_FIELD);
1795     }
1796
1797     src_y  -= linesize   + 1 + pixel_shift;
1798     src_cb -= uvlinesize + 1 + pixel_shift;
1799     src_cr -= uvlinesize + 1 + pixel_shift;
1800
1801     top_border_m1 = h->top_borders[top_idx][s->mb_x - 1];
1802     top_border    = h->top_borders[top_idx][s->mb_x];
1803
1804 #define XCHG(a, b, xchg)                        \
1805     if (pixel_shift) {                          \
1806         if (xchg) {                             \
1807             AV_SWAP64(b + 0, a + 0);            \
1808             AV_SWAP64(b + 8, a + 8);            \
1809         } else {                                \
1810             AV_COPY128(b, a);                   \
1811         }                                       \
1812     } else if (xchg)                            \
1813         AV_SWAP64(b, a);                        \
1814     else                                        \
1815         AV_COPY64(b, a);
1816
1817     if (deblock_top) {
1818         if (deblock_topleft) {
1819             XCHG(top_border_m1 + (8 << pixel_shift),
1820                  src_y - (7 << pixel_shift), 1);
1821         }
1822         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1823         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1824         if (s->mb_x + 1 < s->mb_width) {
1825             XCHG(h->top_borders[top_idx][s->mb_x + 1],
1826                  src_y + (17 << pixel_shift), 1);
1827         }
1828     }
1829     if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1830         if (chroma444) {
1831             if (deblock_topleft) {
1832                 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1833                 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1834             }
1835             XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1836             XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1837             XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1838             XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1839             if (s->mb_x + 1 < s->mb_width) {
1840                 XCHG(h->top_borders[top_idx][s->mb_x + 1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1841                 XCHG(h->top_borders[top_idx][s->mb_x + 1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1842             }
1843         } else {
1844             if (deblock_top) {
1845                 if (deblock_topleft) {
1846                     XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1847                     XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1848                 }
1849                 XCHG(top_border + (16 << pixel_shift), src_cb + 1 + pixel_shift, 1);
1850                 XCHG(top_border + (24 << pixel_shift), src_cr + 1 + pixel_shift, 1);
1851             }
1852         }
1853     }
1854 }
1855
1856 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth,
1857                                         int index)
1858 {
1859     if (high_bit_depth) {
1860         return AV_RN32A(((int32_t *)mb) + index);
1861     } else
1862         return AV_RN16A(mb + index);
1863 }
1864
1865 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth,
1866                                          int index, int value)
1867 {
1868     if (high_bit_depth) {
1869         AV_WN32A(((int32_t *)mb) + index, value);
1870     } else
1871         AV_WN16A(mb + index, value);
1872 }
1873
1874 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h,
1875                                                        int mb_type, int is_h264,
1876                                                        int simple,
1877                                                        int transform_bypass,
1878                                                        int pixel_shift,
1879                                                        int *block_offset,
1880                                                        int linesize,
1881                                                        uint8_t *dest_y, int p)
1882 {
1883     MpegEncContext *const s = &h->s;
1884     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1885     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1886     int i;
1887     int qscale = p == 0 ? s->qscale : h->chroma_qp[p - 1];
1888     block_offset += 16 * p;
1889     if (IS_INTRA4x4(mb_type)) {
1890         if (simple || !s->encoding) {
1891             if (IS_8x8DCT(mb_type)) {
1892                 if (transform_bypass) {
1893                     idct_dc_add  =
1894                     idct_add     = s->dsp.add_pixels8;
1895                 } else {
1896                     idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1897                     idct_add    = h->h264dsp.h264_idct8_add;
1898                 }
1899                 for (i = 0; i < 16; i += 4) {
1900                     uint8_t *const ptr = dest_y + block_offset[i];
1901                     const int dir      = h->intra4x4_pred_mode_cache[scan8[i]];
1902                     if (transform_bypass && h->sps.profile_idc == 244 && dir <= 1) {
1903                         h->hpc.pred8x8l_add[dir](ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1904                     } else {
1905                         const int nnz = h->non_zero_count_cache[scan8[i + p * 16]];
1906                         h->hpc.pred8x8l[dir](ptr, (h->topleft_samples_available << i) & 0x8000,
1907                                              (h->topright_samples_available << i) & 0x4000, linesize);
1908                         if (nnz) {
1909                             if (nnz == 1 && dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
1910                                 idct_dc_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1911                             else
1912                                 idct_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1913                         }
1914                     }
1915                 }
1916             } else {
1917                 if (transform_bypass) {
1918                     idct_dc_add  =
1919                         idct_add = s->dsp.add_pixels4;
1920                 } else {
1921                     idct_dc_add = h->h264dsp.h264_idct_dc_add;
1922                     idct_add    = h->h264dsp.h264_idct_add;
1923                 }
1924                 for (i = 0; i < 16; i++) {
1925                     uint8_t *const ptr = dest_y + block_offset[i];
1926                     const int dir      = h->intra4x4_pred_mode_cache[scan8[i]];
1927
1928                     if (transform_bypass && h->sps.profile_idc == 244 && dir <= 1) {
1929                         h->hpc.pred4x4_add[dir](ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1930                     } else {
1931                         uint8_t *topright;
1932                         int nnz, tr;
1933                         uint64_t tr_high;
1934                         if (dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED) {
1935                             const int topright_avail = (h->topright_samples_available << i) & 0x8000;
1936                             assert(s->mb_y || linesize <= block_offset[i]);
1937                             if (!topright_avail) {
1938                                 if (pixel_shift) {
1939                                     tr_high  = ((uint16_t *)ptr)[3 - linesize / 2] * 0x0001000100010001ULL;
1940                                     topright = (uint8_t *)&tr_high;
1941                                 } else {
1942                                     tr       = ptr[3 - linesize] * 0x01010101u;
1943                                     topright = (uint8_t *)&tr;
1944                                 }
1945                             } else
1946                                 topright = ptr + (4 << pixel_shift) - linesize;
1947                         } else
1948                             topright = NULL;
1949
1950                         h->hpc.pred4x4[dir](ptr, topright, linesize);
1951                         nnz = h->non_zero_count_cache[scan8[i + p * 16]];
1952                         if (nnz) {
1953                             if (is_h264) {
1954                                 if (nnz == 1 && dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
1955                                     idct_dc_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1956                                 else
1957                                     idct_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1958                             } else if (CONFIG_SVQ3_DECODER)
1959                                 ff_svq3_add_idct_c(ptr, h->mb + i * 16 + p * 256, linesize, qscale, 0);
1960                         }
1961                     }
1962                 }
1963             }
1964         }
1965     } else {
1966         h->hpc.pred16x16[h->intra16x16_pred_mode](dest_y, linesize);
1967         if (is_h264) {
1968             if (h->non_zero_count_cache[scan8[LUMA_DC_BLOCK_INDEX + p]]) {
1969                 if (!transform_bypass)
1970                     h->h264dsp.h264_luma_dc_dequant_idct(h->mb + (p * 256 << pixel_shift),
1971                                                          h->mb_luma_dc[p],
1972                                                          h->dequant4_coeff[p][qscale][0]);
1973                 else {
1974                     static const uint8_t dc_mapping[16] = {
1975                          0 * 16,  1 * 16,  4 * 16,  5 * 16,
1976                          2 * 16,  3 * 16,  6 * 16,  7 * 16,
1977                          8 * 16,  9 * 16, 12 * 16, 13 * 16,
1978                         10 * 16, 11 * 16, 14 * 16, 15 * 16 };
1979                     for (i = 0; i < 16; i++)
1980                         dctcoef_set(h->mb + (p * 256 << pixel_shift),
1981                                     pixel_shift, dc_mapping[i],
1982                                     dctcoef_get(h->mb_luma_dc[p],
1983                                                 pixel_shift, i));
1984                 }
1985             }
1986         } else if (CONFIG_SVQ3_DECODER)
1987             ff_svq3_luma_dc_dequant_idct_c(h->mb + p * 256,
1988                                            h->mb_luma_dc[p], qscale);
1989     }
1990 }
1991
1992 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type,
1993                                                     int is_h264, int simple,
1994                                                     int transform_bypass,
1995                                                     int pixel_shift,
1996                                                     int *block_offset,
1997                                                     int linesize,
1998                                                     uint8_t *dest_y, int p)
1999 {
2000     MpegEncContext *const s = &h->s;
2001     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2002     int i;
2003     block_offset += 16 * p;
2004     if (!IS_INTRA4x4(mb_type)) {
2005         if (is_h264) {
2006             if (IS_INTRA16x16(mb_type)) {
2007                 if (transform_bypass) {
2008                     if (h->sps.profile_idc == 244 &&
2009                         (h->intra16x16_pred_mode == VERT_PRED8x8 ||
2010                          h->intra16x16_pred_mode == HOR_PRED8x8)) {
2011                         h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset,
2012                                                                       h->mb + (p * 256 << pixel_shift),
2013                                                                       linesize);
2014                     } else {
2015                         for (i = 0; i < 16; i++)
2016                             if (h->non_zero_count_cache[scan8[i + p * 16]] ||
2017                                 dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
2018                                 s->dsp.add_pixels4(dest_y + block_offset[i],
2019                                                    h->mb + (i * 16 + p * 256 << pixel_shift),
2020                                                    linesize);
2021                     }
2022                 } else {
2023                     h->h264dsp.h264_idct_add16intra(dest_y, block_offset,
2024                                                     h->mb + (p * 256 << pixel_shift),
2025                                                     linesize,
2026                                                     h->non_zero_count_cache + p * 5 * 8);
2027                 }
2028             } else if (h->cbp & 15) {
2029                 if (transform_bypass) {
2030                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
2031                     idct_add = IS_8x8DCT(mb_type) ? s->dsp.add_pixels8
2032                                                   : s->dsp.add_pixels4;
2033                     for (i = 0; i < 16; i += di)
2034                         if (h->non_zero_count_cache[scan8[i + p * 16]])
2035                             idct_add(dest_y + block_offset[i],
2036                                      h->mb + (i * 16 + p * 256 << pixel_shift),
2037                                      linesize);
2038                 } else {
2039                     if (IS_8x8DCT(mb_type))
2040                         h->h264dsp.h264_idct8_add4(dest_y, block_offset,
2041                                                    h->mb + (p * 256 << pixel_shift),
2042                                                    linesize,
2043                                                    h->non_zero_count_cache + p * 5 * 8);
2044                     else
2045                         h->h264dsp.h264_idct_add16(dest_y, block_offset,
2046                                                    h->mb + (p * 256 << pixel_shift),
2047                                                    linesize,
2048                                                    h->non_zero_count_cache + p * 5 * 8);
2049                 }
2050             }
2051         } else if (CONFIG_SVQ3_DECODER) {
2052             for (i = 0; i < 16; i++)
2053                 if (h->non_zero_count_cache[scan8[i + p * 16]] || h->mb[i * 16 + p * 256]) {
2054                     // FIXME benchmark weird rule, & below
2055                     uint8_t *const ptr = dest_y + block_offset[i];
2056                     ff_svq3_add_idct_c(ptr, h->mb + i * 16 + p * 256, linesize,
2057                                        s->qscale, IS_INTRA(mb_type) ? 1 : 0);
2058                 }
2059         }
2060     }
2061 }
2062
2063 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple,
2064                                                    int pixel_shift)
2065 {
2066     MpegEncContext *const s = &h->s;
2067     const int mb_x    = s->mb_x;
2068     const int mb_y    = s->mb_y;
2069     const int mb_xy   = h->mb_xy;
2070     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2071     uint8_t *dest_y, *dest_cb, *dest_cr;
2072     int linesize, uvlinesize /*dct_offset*/;
2073     int i, j;
2074     int *block_offset = &h->block_offset[0];
2075     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
2076     /* is_h264 should always be true if SVQ3 is disabled. */
2077     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
2078     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2079     const int block_h   = 16 >> s->chroma_y_shift;
2080     const int chroma422 = CHROMA422;
2081
2082     dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift)     + mb_y * s->linesize)  * 16;
2083     dest_cb = s->current_picture.f.data[1] +  (mb_x << pixel_shift) * 8 + mb_y * s->uvlinesize * block_h;
2084     dest_cr = s->current_picture.f.data[2] +  (mb_x << pixel_shift) * 8 + mb_y * s->uvlinesize * block_h;
2085
2086     s->dsp.prefetch(dest_y  + (s->mb_x & 3) * 4 * s->linesize   + (64 << pixel_shift), s->linesize,       4);
2087     s->dsp.prefetch(dest_cb + (s->mb_x & 7)     * s->uvlinesize + (64 << pixel_shift), dest_cr - dest_cb, 2);
2088
2089     h->list_counts[mb_xy] = h->list_count;
2090
2091     if (!simple && MB_FIELD) {
2092         linesize     = h->mb_linesize = s->linesize * 2;
2093         uvlinesize   = h->mb_uvlinesize = s->uvlinesize * 2;
2094         block_offset = &h->block_offset[48];
2095         if (mb_y & 1) { // FIXME move out of this function?
2096             dest_y  -= s->linesize * 15;
2097             dest_cb -= s->uvlinesize * (block_h - 1);
2098             dest_cr -= s->uvlinesize * (block_h - 1);
2099         }
2100         if (FRAME_MBAFF) {
2101             int list;
2102             for (list = 0; list < h->list_count; list++) {
2103                 if (!USES_LIST(mb_type, list))
2104                     continue;
2105                 if (IS_16X16(mb_type)) {
2106                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2107                     fill_rectangle(ref, 4, 4, 8, (16 + *ref) ^ (s->mb_y & 1), 1);
2108                 } else {
2109                     for (i = 0; i < 16; i += 4) {
2110                         int ref = h->ref_cache[list][scan8[i]];
2111                         if (ref >= 0)
2112                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2,
2113                                            8, (16 + ref) ^ (s->mb_y & 1), 1);
2114                     }
2115                 }
2116             }
2117         }
2118     } else {
2119         linesize   = h->mb_linesize   = s->linesize;
2120         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
2121         // dct_offset = s->linesize * 16;
2122     }
2123
2124     if (!simple && IS_INTRA_PCM(mb_type)) {
2125         const int bit_depth = h->sps.bit_depth_luma;
2126         if (pixel_shift) {
2127             int j;
2128             GetBitContext gb;
2129             init_get_bits(&gb, (uint8_t *)h->mb, 384 * bit_depth);
2130
2131             for (i = 0; i < 16; i++) {
2132                 uint16_t *tmp_y = (uint16_t *)(dest_y + i * linesize);
2133                 for (j = 0; j < 16; j++)
2134                     tmp_y[j] = get_bits(&gb, bit_depth);
2135             }
2136             if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
2137                 if (!h->sps.chroma_format_idc) {
2138                     for (i = 0; i < block_h; i++) {
2139                         uint16_t *tmp_cb = (uint16_t *)(dest_cb + i * uvlinesize);
2140                         uint16_t *tmp_cr = (uint16_t *)(dest_cr + i * uvlinesize);
2141                         for (j = 0; j < 8; j++) {
2142                             tmp_cb[j] = tmp_cr[j] = 1 << (bit_depth - 1);
2143                         }
2144                     }
2145                 } else {
2146                     for (i = 0; i < block_h; i++) {
2147                         uint16_t *tmp_cb = (uint16_t *)(dest_cb + i * uvlinesize);
2148                         for (j = 0; j < 8; j++)
2149                             tmp_cb[j] = get_bits(&gb, bit_depth);
2150                     }
2151                     for (i = 0; i < block_h; i++) {
2152                         uint16_t *tmp_cr = (uint16_t *)(dest_cr + i * uvlinesize);
2153                         for (j = 0; j < 8; j++)
2154                             tmp_cr[j] = get_bits(&gb, bit_depth);
2155                     }
2156                 }
2157             }
2158         } else {
2159             for (i = 0; i < 16; i++)
2160                 memcpy(dest_y + i * linesize, h->mb + i * 8, 16);
2161             if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
2162                 if (!h->sps.chroma_format_idc) {
2163                     for (i = 0; i < 8; i++) {
2164                         memset(dest_cb + i*uvlinesize, 1 << (bit_depth - 1), 8);
2165                         memset(dest_cr + i*uvlinesize, 1 << (bit_depth - 1), 8);
2166                     }
2167                 } else {
2168                     for (i = 0; i < block_h; i++) {
2169                         memcpy(dest_cb + i * uvlinesize, h->mb + 128 + i * 4, 8);
2170                         memcpy(dest_cr + i * uvlinesize, h->mb + 160 + i * 4, 8);
2171                     }
2172                 }
2173             }
2174         }
2175     } else {
2176         if (IS_INTRA(mb_type)) {
2177             if (h->deblocking_filter)
2178                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize,
2179                                uvlinesize, 1, 0, simple, pixel_shift);
2180
2181             if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
2182                 h->hpc.pred8x8[h->chroma_pred_mode](dest_cb, uvlinesize);
2183                 h->hpc.pred8x8[h->chroma_pred_mode](dest_cr, uvlinesize);
2184             }
2185
2186             hl_decode_mb_predict_luma(h, mb_type, is_h264, simple,
2187                                       transform_bypass, pixel_shift,
2188                                       block_offset, linesize, dest_y, 0);
2189
2190             if (h->deblocking_filter)
2191                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize,
2192                                uvlinesize, 0, 0, simple, pixel_shift);
2193         } else if (is_h264) {
2194             if (chroma422) {
2195                 hl_motion_422(h, dest_y, dest_cb, dest_cr,
2196                               s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2197                               s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2198                               h->h264dsp.weight_h264_pixels_tab,
2199                               h->h264dsp.biweight_h264_pixels_tab,
2200                               pixel_shift);
2201             } else {
2202                 hl_motion_420(h, dest_y, dest_cb, dest_cr,
2203                               s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2204                               s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2205                               h->h264dsp.weight_h264_pixels_tab,
2206                               h->h264dsp.biweight_h264_pixels_tab,
2207                               pixel_shift);
2208             }
2209         }
2210
2211         hl_decode_mb_idct_luma(h, mb_type, is_h264, simple, transform_bypass,
2212                                pixel_shift, block_offset, linesize, dest_y, 0);
2213
2214         if ((simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) &&
2215             (h->cbp & 0x30)) {
2216             uint8_t *dest[2] = { dest_cb, dest_cr };
2217             if (transform_bypass) {
2218                 if (IS_INTRA(mb_type) && h->sps.profile_idc == 244 &&
2219                     (h->chroma_pred_mode == VERT_PRED8x8 ||
2220                      h->chroma_pred_mode == HOR_PRED8x8)) {
2221                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0],
2222                                                             block_offset + 16,
2223                                                             h->mb + (16 * 16 * 1 << pixel_shift),
2224                                                             uvlinesize);
2225                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1],
2226                                                             block_offset + 32,
2227                                                             h->mb + (16 * 16 * 2 << pixel_shift),
2228                                                             uvlinesize);
2229                 } else {
2230                     idct_add = s->dsp.add_pixels4;
2231                     for (j = 1; j < 3; j++) {
2232                         for (i = j * 16; i < j * 16 + 4; i++)
2233                             if (h->non_zero_count_cache[scan8[i]] ||
2234                                 dctcoef_get(h->mb, pixel_shift, i * 16))
2235                                 idct_add(dest[j - 1] + block_offset[i],
2236                                          h->mb + (i * 16 << pixel_shift),
2237                                          uvlinesize);
2238                         if (chroma422) {
2239                             for (i = j * 16 + 4; i < j * 16 + 8; i++)
2240                                 if (h->non_zero_count_cache[scan8[i + 4]] ||
2241                                     dctcoef_get(h->mb, pixel_shift, i * 16))
2242                                     idct_add(dest[j - 1] + block_offset[i + 4],
2243                                              h->mb + (i * 16 << pixel_shift),
2244                                              uvlinesize);
2245                         }
2246                     }
2247                 }
2248             } else {
2249                 if (is_h264) {
2250                     int qp[2];
2251                     if (chroma422) {
2252                         qp[0] = h->chroma_qp[0] + 3;
2253                         qp[1] = h->chroma_qp[1] + 3;
2254                     } else {
2255                         qp[0] = h->chroma_qp[0];
2256                         qp[1] = h->chroma_qp[1];
2257                     }
2258                     if (h->non_zero_count_cache[scan8[CHROMA_DC_BLOCK_INDEX + 0]])
2259                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16 * 16 * 1 << pixel_shift),
2260                                                                h->dequant4_coeff[IS_INTRA(mb_type) ? 1 : 4][qp[0]][0]);
2261                     if (h->non_zero_count_cache[scan8[CHROMA_DC_BLOCK_INDEX + 1]])
2262                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16 * 16 * 2 << pixel_shift),
2263                                                                h->dequant4_coeff[IS_INTRA(mb_type) ? 2 : 5][qp[1]][0]);
2264                     h->h264dsp.h264_idct_add8(dest, block_offset,
2265                                               h->mb, uvlinesize,
2266                                               h->non_zero_count_cache);
2267                 } else if (CONFIG_SVQ3_DECODER) {
2268                     h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + 16 * 16 * 1,
2269                                                            h->dequant4_coeff[IS_INTRA(mb_type) ? 1 : 4][h->chroma_qp[0]][0]);
2270                     h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + 16 * 16 * 2,
2271                                                            h->dequant4_coeff[IS_INTRA(mb_type) ? 2 : 5][h->chroma_qp[1]][0]);
2272                     for (j = 1; j < 3; j++) {
2273                         for (i = j * 16; i < j * 16 + 4; i++)
2274                             if (h->non_zero_count_cache[scan8[i]] || h->mb[i * 16]) {
2275                                 uint8_t *const ptr = dest[j - 1] + block_offset[i];
2276                                 ff_svq3_add_idct_c(ptr, h->mb + i * 16,
2277                                                    uvlinesize,
2278                                                    ff_h264_chroma_qp[0][s->qscale + 12] - 12, 2);
2279                             }
2280                     }
2281                 }
2282             }
2283         }
2284     }
2285     if (h->cbp || IS_INTRA(mb_type)) {
2286         s->dsp.clear_blocks(h->mb);
2287         s->dsp.clear_blocks(h->mb + (24 * 16 << pixel_shift));
2288     }
2289 }
2290
2291 static av_always_inline void hl_decode_mb_444_internal(H264Context *h,
2292                                                        int simple,
2293                                                        int pixel_shift)
2294 {
2295     MpegEncContext *const s = &h->s;
2296     const int mb_x    = s->mb_x;
2297     const int mb_y    = s->mb_y;
2298     const int mb_xy   = h->mb_xy;
2299     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2300     uint8_t *dest[3];
2301     int linesize;
2302     int i, j, p;
2303     int *block_offset = &h->block_offset[0];
2304     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
2305     const int plane_count      = (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) ? 3 : 1;
2306
2307     for (p = 0; p < plane_count; p++) {
2308         dest[p] = s->current_picture.f.data[p] +
2309                   ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
2310         s->dsp.prefetch(dest[p] + (s->mb_x & 3) * 4 * s->linesize + (64 << pixel_shift),
2311                         s->linesize, 4);
2312     }
2313
2314     h->list_counts[mb_xy] = h->list_count;
2315
2316     if (!simple && MB_FIELD) {
2317         linesize     = h->mb_linesize = h->mb_uvlinesize = s->linesize * 2;
2318         block_offset = &h->block_offset[48];
2319         if (mb_y & 1) // FIXME move out of this function?
2320             for (p = 0; p < 3; p++)
2321                 dest[p] -= s->linesize * 15;
2322         if (FRAME_MBAFF) {
2323             int list;
2324             for (list = 0; list < h->list_count; list++) {
2325                 if (!USES_LIST(mb_type, list))
2326                     continue;
2327                 if (IS_16X16(mb_type)) {
2328                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2329                     fill_rectangle(ref, 4, 4, 8, (16 + *ref) ^ (s->mb_y & 1), 1);
2330                 } else {
2331                     for (i = 0; i < 16; i += 4) {
2332                         int ref = h->ref_cache[list][scan8[i]];
2333                         if (ref >= 0)
2334                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2,
2335                                            8, (16 + ref) ^ (s->mb_y & 1), 1);
2336                     }
2337                 }
2338             }
2339         }
2340     } else {
2341         linesize = h->mb_linesize = h->mb_uvlinesize = s->linesize;
2342     }
2343
2344     if (!simple && IS_INTRA_PCM(mb_type)) {
2345         if (pixel_shift) {
2346             const int bit_depth = h->sps.bit_depth_luma;
2347             GetBitContext gb;
2348             init_get_bits(&gb, (uint8_t *)h->mb, 768 * bit_depth);
2349
2350             for (p = 0; p < plane_count; p++)
2351                 for (i = 0; i < 16; i++) {
2352                     uint16_t *tmp = (uint16_t *)(dest[p] + i * linesize);
2353                     for (j = 0; j < 16; j++)
2354                         tmp[j] = get_bits(&gb, bit_depth);
2355                 }
2356         } else {
2357             for (p = 0; p < plane_count; p++)
2358                 for (i = 0; i < 16; i++)
2359                     memcpy(dest[p] + i * linesize, h->mb + p * 128 + i * 8, 16);
2360         }
2361     } else {
2362         if (IS_INTRA(mb_type)) {
2363             if (h->deblocking_filter)
2364                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize,
2365                                linesize, 1, 1, simple, pixel_shift);
2366
2367             for (p = 0; p < plane_count; p++)
2368                 hl_decode_mb_predict_luma(h, mb_type, 1, simple,
2369                                           transform_bypass, pixel_shift,
2370                                           block_offset, linesize, dest[p], p);
2371
2372             if (h->deblocking_filter)
2373                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize,
2374                                linesize, 0, 1, simple, pixel_shift);
2375         } else {
2376             hl_motion(h, dest[0], dest[1], dest[2],
2377                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2378                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2379                       h->h264dsp.weight_h264_pixels_tab,
2380                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 3);
2381         }
2382
2383         for (p = 0; p < plane_count; p++)
2384             hl_decode_mb_idct_luma(h, mb_type, 1, simple, transform_bypass,
2385                                    pixel_shift, block_offset, linesize,
2386                                    dest[p], p);
2387     }
2388     if (h->cbp || IS_INTRA(mb_type)) {
2389         s->dsp.clear_blocks(h->mb);
2390         s->dsp.clear_blocks(h->mb + (24 * 16 << pixel_shift));
2391     }
2392 }
2393
2394 /**
2395  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2396  */
2397 #define hl_decode_mb_simple(sh, bits)                          \
2398 static void hl_decode_mb_simple_ ## bits(H264Context *h)       \
2399 {                                                              \
2400     hl_decode_mb_internal(h, 1, sh);                           \
2401 }
2402
2403 hl_decode_mb_simple(0, 8)
2404 hl_decode_mb_simple(1, 16)
2405
2406 /**
2407  * Process a macroblock; this handles edge cases, such as interlacing.
2408  */
2409 static void av_noinline hl_decode_mb_complex(H264Context *h)
2410 {
2411     hl_decode_mb_internal(h, 0, h->pixel_shift);
2412 }
2413
2414 static void av_noinline hl_decode_mb_444_complex(H264Context *h)
2415 {
2416     hl_decode_mb_444_internal(h, 0, h->pixel_shift);
2417 }
2418
2419 static void av_noinline hl_decode_mb_444_simple(H264Context *h)
2420 {
2421     hl_decode_mb_444_internal(h, 1, 0);
2422 }
2423
2424 void ff_h264_hl_decode_mb(H264Context *h)
2425 {
2426     MpegEncContext *const s = &h->s;
2427     const int mb_xy   = h->mb_xy;
2428     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2429     int is_complex    = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2430
2431     if (CHROMA444) {
2432         if (is_complex || h->pixel_shift)
2433             hl_decode_mb_444_complex(h);
2434         else
2435             hl_decode_mb_444_simple(h);
2436     } else if (is_complex) {
2437         hl_decode_mb_complex(h);
2438     } else if (h->pixel_shift) {
2439         hl_decode_mb_simple_16(h);
2440     } else
2441         hl_decode_mb_simple_8(h);
2442 }
2443
2444 static int pred_weight_table(H264Context *h)
2445 {
2446     MpegEncContext *const s = &h->s;
2447     int list, i;
2448     int luma_def, chroma_def;
2449
2450     h->use_weight             = 0;
2451     h->use_weight_chroma      = 0;
2452     h->luma_log2_weight_denom = get_ue_golomb(&s->gb);
2453     if (h->sps.chroma_format_idc)
2454         h->chroma_log2_weight_denom = get_ue_golomb(&s->gb);
2455     luma_def   = 1 << h->luma_log2_weight_denom;
2456     chroma_def = 1 << h->chroma_log2_weight_denom;
2457
2458     for (list = 0; list < 2; list++) {
2459         h->luma_weight_flag[list]   = 0;
2460         h->chroma_weight_flag[list] = 0;
2461         for (i = 0; i < h->ref_count[list]; i++) {
2462             int luma_weight_flag, chroma_weight_flag;
2463
2464             luma_weight_flag = get_bits1(&s->gb);
2465             if (luma_weight_flag) {
2466                 h->luma_weight[i][list][0] = get_se_golomb(&s->gb);
2467                 h->luma_weight[i][list][1] = get_se_golomb(&s->gb);
2468                 if (h->luma_weight[i][list][0] != luma_def ||
2469                     h->luma_weight[i][list][1] != 0) {
2470                     h->use_weight             = 1;
2471                     h->luma_weight_flag[list] = 1;
2472                 }
2473             } else {
2474                 h->luma_weight[i][list][0] = luma_def;
2475                 h->luma_weight[i][list][1] = 0;
2476             }
2477
2478             if (h->sps.chroma_format_idc) {
2479                 chroma_weight_flag = get_bits1(&s->gb);
2480                 if (chroma_weight_flag) {
2481                     int j;
2482                     for (j = 0; j < 2; j++) {
2483                         h->chroma_weight[i][list][j][0] = get_se_golomb(&s->gb);
2484                         h->chroma_weight[i][list][j][1] = get_se_golomb(&s->gb);
2485                         if (h->chroma_weight[i][list][j][0] != chroma_def ||
2486                             h->chroma_weight[i][list][j][1] != 0) {
2487                             h->use_weight_chroma = 1;
2488                             h->chroma_weight_flag[list] = 1;
2489                         }
2490                     }
2491                 } else {
2492                     int j;
2493                     for (j = 0; j < 2; j++) {
2494                         h->chroma_weight[i][list][j][0] = chroma_def;
2495                         h->chroma_weight[i][list][j][1] = 0;
2496                     }
2497                 }
2498             }
2499         }
2500         if (h->slice_type_nos != AV_PICTURE_TYPE_B)
2501             break;
2502     }
2503     h->use_weight = h->use_weight || h->use_weight_chroma;
2504     return 0;
2505 }
2506
2507 /**
2508  * Initialize implicit_weight table.
2509  * @param field  0/1 initialize the weight for interlaced MBAFF
2510  *                -1 initializes the rest
2511  */
2512 static void implicit_weight_table(H264Context *h, int field)
2513 {
2514     MpegEncContext *const s = &h->s;
2515     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2516
2517     for (i = 0; i < 2; i++) {
2518         h->luma_weight_flag[i]   = 0;
2519         h->chroma_weight_flag[i] = 0;
2520     }
2521
2522     if (field < 0) {
2523         if (s->picture_structure == PICT_FRAME) {
2524             cur_poc = s->current_picture_ptr->poc;
2525         } else {
2526             cur_poc = s->current_picture_ptr->field_poc[s->picture_structure - 1];
2527         }
2528         if (h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF &&
2529             h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2 * cur_poc) {
2530             h->use_weight = 0;
2531             h->use_weight_chroma = 0;
2532             return;
2533         }
2534         ref_start  = 0;
2535         ref_count0 = h->ref_count[0];
2536         ref_count1 = h->ref_count[1];
2537     } else {
2538         cur_poc    = s->current_picture_ptr->field_poc[field];
2539         ref_start  = 16;
2540         ref_count0 = 16 + 2 * h->ref_count[0];
2541         ref_count1 = 16 + 2 * h->ref_count[1];
2542     }
2543
2544     h->use_weight               = 2;
2545     h->use_weight_chroma        = 2;
2546     h->luma_log2_weight_denom   = 5;
2547     h->chroma_log2_weight_denom = 5;
2548
2549     for (ref0 = ref_start; ref0 < ref_count0; ref0++) {
2550         int poc0 = h->ref_list[0][ref0].poc;
2551         for (ref1 = ref_start; ref1 < ref_count1; ref1++) {
2552             int w = 32;
2553             if (!h->ref_list[0][ref0].long_ref && !h->ref_list[1][ref1].long_ref) {
2554                 int poc1 = h->ref_list[1][ref1].poc;
2555                 int td   = av_clip(poc1 - poc0, -128, 127);
2556                 if (td) {
2557                     int tb = av_clip(cur_poc - poc0, -128, 127);
2558                     int tx = (16384 + (FFABS(td) >> 1)) / td;
2559                     int dist_scale_factor = (tb * tx + 32) >> 8;
2560                     if (dist_scale_factor >= -64 && dist_scale_factor <= 128)
2561                         w = 64 - dist_scale_factor;
2562                 }
2563             }
2564             if (field < 0) {
2565                 h->implicit_weight[ref0][ref1][0] =
2566                 h->implicit_weight[ref0][ref1][1] = w;
2567             } else {
2568                 h->implicit_weight[ref0][ref1][field] = w;
2569             }
2570         }
2571     }
2572 }
2573
2574 /**
2575  * instantaneous decoder refresh.
2576  */
2577 static void idr(H264Context *h)
2578 {
2579     int i;
2580     ff_h264_remove_all_refs(h);
2581     h->prev_frame_num        = 0;
2582     h->prev_frame_num_offset = 0;
2583     h->prev_poc_msb          = 1<<16;
2584     h->prev_poc_lsb          = 0;
2585     for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
2586         h->last_pocs[i] = INT_MIN;
2587 }
2588
2589 /* forget old pics after a seek */
2590 static void flush_dpb(AVCodecContext *avctx)
2591 {
2592     H264Context *h = avctx->priv_data;
2593     int i;
2594     for (i=0; i<=MAX_DELAYED_PIC_COUNT; i++) {
2595         if (h->delayed_pic[i])
2596             h->delayed_pic[i]->f.reference = 0;
2597         h->delayed_pic[i] = NULL;
2598     }
2599     h->outputed_poc = h->next_outputed_poc = INT_MIN;
2600     h->prev_interlaced_frame = 1;
2601     idr(h);
2602     h->prev_frame_num = -1;
2603     if (h->s.current_picture_ptr)
2604         h->s.current_picture_ptr->f.reference = 0;
2605     h->s.first_field = 0;
2606     ff_h264_reset_sei(h);
2607     ff_mpeg_flush(avctx);
2608     h->recovery_frame= -1;
2609     h->sync= 0;
2610 }
2611
2612 static int init_poc(H264Context *h)
2613 {
2614     MpegEncContext *const s = &h->s;
2615     const int max_frame_num = 1 << h->sps.log2_max_frame_num;
2616     int field_poc[2];
2617     Picture *cur = s->current_picture_ptr;
2618
2619     h->frame_num_offset = h->prev_frame_num_offset;
2620     if (h->frame_num < h->prev_frame_num)
2621         h->frame_num_offset += max_frame_num;
2622
2623     if (h->sps.poc_type == 0) {
2624         const int max_poc_lsb = 1 << h->sps.log2_max_poc_lsb;
2625
2626         if (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb / 2)
2627             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2628         else if (h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb / 2)
2629             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2630         else
2631             h->poc_msb = h->prev_poc_msb;
2632         // printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
2633         field_poc[0] =
2634         field_poc[1] = h->poc_msb + h->poc_lsb;
2635         if (s->picture_structure == PICT_FRAME)
2636             field_poc[1] += h->delta_poc_bottom;
2637     } else if (h->sps.poc_type == 1) {
2638         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2639         int i;
2640
2641         if (h->sps.poc_cycle_length != 0)
2642             abs_frame_num = h->frame_num_offset + h->frame_num;
2643         else
2644             abs_frame_num = 0;
2645
2646         if (h->nal_ref_idc == 0 && abs_frame_num > 0)
2647             abs_frame_num--;
2648
2649         expected_delta_per_poc_cycle = 0;
2650         for (i = 0; i < h->sps.poc_cycle_length; i++)
2651             // FIXME integrate during sps parse
2652             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[i];
2653
2654         if (abs_frame_num > 0) {
2655             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2656             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2657
2658             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2659             for (i = 0; i <= frame_num_in_poc_cycle; i++)
2660                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[i];
2661         } else
2662             expectedpoc = 0;
2663
2664         if (h->nal_ref_idc == 0)
2665             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2666
2667         field_poc[0] = expectedpoc + h->delta_poc[0];
2668         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2669
2670         if (s->picture_structure == PICT_FRAME)
2671             field_poc[1] += h->delta_poc[1];
2672     } else {
2673         int poc = 2 * (h->frame_num_offset + h->frame_num);
2674
2675         if (!h->nal_ref_idc)
2676             poc--;
2677
2678         field_poc[0] = poc;
2679         field_poc[1] = poc;
2680     }
2681
2682     if (s->picture_structure != PICT_BOTTOM_FIELD)
2683         s->current_picture_ptr->field_poc[0] = field_poc[0];
2684     if (s->picture_structure != PICT_TOP_FIELD)
2685         s->current_picture_ptr->field_poc[1] = field_poc[1];
2686     cur->poc = FFMIN(cur->field_poc[0], cur->field_poc[1]);
2687
2688     return 0;
2689 }
2690
2691 /**
2692  * initialize scan tables
2693  */
2694 static void init_scan_tables(H264Context *h)
2695 {
2696     int i;
2697     for (i = 0; i < 16; i++) {
2698 #define T(x) (x >> 2) | ((x << 2) & 0xF)
2699         h->zigzag_scan[i] = T(zigzag_scan[i]);
2700         h->field_scan[i]  = T(field_scan[i]);
2701 #undef T
2702     }
2703     for (i = 0; i < 64; i++) {
2704 #define T(x) (x >> 3) | ((x & 7) << 3)
2705         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
2706         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2707         h->field_scan8x8[i]        = T(field_scan8x8[i]);
2708         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
2709 #undef T
2710     }
2711     if (h->sps.transform_bypass) { // FIXME same ugly
2712         h->zigzag_scan_q0          = zigzag_scan;
2713         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
2714         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2715         h->field_scan_q0           = field_scan;
2716         h->field_scan8x8_q0        = field_scan8x8;
2717         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
2718     } else {
2719         h->zigzag_scan_q0          = h->zigzag_scan;
2720         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
2721         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2722         h->field_scan_q0           = h->field_scan;
2723         h->field_scan8x8_q0        = h->field_scan8x8;
2724         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
2725     }
2726 }
2727
2728 static int field_end(H264Context *h, int in_setup)
2729 {
2730     MpegEncContext *const s     = &h->s;
2731     AVCodecContext *const avctx = s->avctx;
2732     int err = 0;
2733     s->mb_y = 0;
2734
2735     if (!in_setup && !s->dropable)
2736         ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX,
2737                                   s->picture_structure == PICT_BOTTOM_FIELD);
2738
2739     if (CONFIG_H264_VDPAU_DECODER &&
2740         s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
2741         ff_vdpau_h264_set_reference_frames(s);
2742
2743     if (in_setup || !(avctx->active_thread_type & FF_THREAD_FRAME)) {
2744         if (!s->dropable) {
2745             err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2746             h->prev_poc_msb = h->poc_msb;
2747             h->prev_poc_lsb = h->poc_lsb;
2748         }
2749         h->prev_frame_num_offset = h->frame_num_offset;
2750         h->prev_frame_num        = h->frame_num;
2751         h->outputed_poc          = h->next_outputed_poc;
2752     }
2753
2754     if (avctx->hwaccel) {
2755         if (avctx->hwaccel->end_frame(avctx) < 0)
2756             av_log(avctx, AV_LOG_ERROR,
2757                    "hardware accelerator failed to decode picture\n");
2758     }
2759
2760     if (CONFIG_H264_VDPAU_DECODER &&
2761         s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
2762         ff_vdpau_h264_picture_complete(s);
2763
2764     /*
2765      * FIXME: Error handling code does not seem to support interlaced
2766      * when slices span multiple rows
2767      * The ff_er_add_slice calls don't work right for bottom
2768      * fields; they cause massive erroneous error concealing
2769      * Error marking covers both fields (top and bottom).
2770      * This causes a mismatched s->error_count
2771      * and a bad error table. Further, the error count goes to
2772      * INT_MAX when called for bottom field, because mb_y is
2773      * past end by one (callers fault) and resync_mb_y != 0
2774      * causes problems for the first MB line, too.
2775      */
2776     if (!FIELD_PICTURE)
2777         ff_er_frame_end(s);
2778
2779     ff_MPV_frame_end(s);
2780
2781     h->current_slice = 0;
2782
2783     return err;
2784 }
2785
2786 /**
2787  * Replicate H264 "master" context to thread contexts.
2788  */
2789 static void clone_slice(H264Context *dst, H264Context *src)
2790 {
2791     memcpy(dst->block_offset, src->block_offset, sizeof(dst->block_offset));
2792     dst->s.current_picture_ptr = src->s.current_picture_ptr;
2793     dst->s.current_picture     = src->s.current_picture;
2794     dst->s.linesize            = src->s.linesize;
2795     dst->s.uvlinesize          = src->s.uvlinesize;
2796     dst->s.first_field         = src->s.first_field;
2797
2798     dst->prev_poc_msb          = src->prev_poc_msb;
2799     dst->prev_poc_lsb          = src->prev_poc_lsb;
2800     dst->prev_frame_num_offset = src->prev_frame_num_offset;
2801     dst->prev_frame_num        = src->prev_frame_num;
2802     dst->short_ref_count       = src->short_ref_count;
2803
2804     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
2805     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
2806     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2807     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
2808
2809     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
2810     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
2811 }
2812
2813 /**
2814  * Compute profile from profile_idc and constraint_set?_flags.
2815  *
2816  * @param sps SPS
2817  *
2818  * @return profile as defined by FF_PROFILE_H264_*
2819  */
2820 int ff_h264_get_profile(SPS *sps)
2821 {
2822     int profile = sps->profile_idc;
2823
2824     switch (sps->profile_idc) {
2825     case FF_PROFILE_H264_BASELINE:
2826         // constraint_set1_flag set to 1
2827         profile |= (sps->constraint_set_flags & 1 << 1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2828         break;
2829     case FF_PROFILE_H264_HIGH_10:
2830     case FF_PROFILE_H264_HIGH_422:
2831     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2832         // constraint_set3_flag set to 1
2833         profile |= (sps->constraint_set_flags & 1 << 3) ? FF_PROFILE_H264_INTRA : 0;
2834         break;
2835     }
2836
2837     return profile;
2838 }
2839
2840 /**
2841  * Decode a slice header.
2842  * This will also call ff_MPV_common_init() and frame_start() as needed.
2843  *
2844  * @param h h264context
2845  * @param h0 h264 master context (differs from 'h' when doing sliced based
2846  *           parallel decoding)
2847  *
2848  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2849  */
2850 static int decode_slice_header(H264Context *h, H264Context *h0)
2851 {
2852     MpegEncContext *const s  = &h->s;
2853     MpegEncContext *const s0 = &h0->s;
2854     unsigned int first_mb_in_slice;
2855     unsigned int pps_id;
2856     int num_ref_idx_active_override_flag;
2857     unsigned int slice_type, tmp, i, j;
2858     int default_ref_list_done = 0;
2859     int last_pic_structure, last_pic_dropable;
2860     int must_reinit;
2861
2862     /* FIXME: 2tap qpel isn't implemented for high bit depth. */
2863     if ((s->avctx->flags2 & CODEC_FLAG2_FAST) &&
2864         !h->nal_ref_idc && !h->pixel_shift) {
2865         s->me.qpel_put = s->dsp.put_2tap_qpel_pixels_tab;
2866         s->me.qpel_avg = s->dsp.avg_2tap_qpel_pixels_tab;
2867     } else {
2868         s->me.qpel_put = s->dsp.put_h264_qpel_pixels_tab;
2869         s->me.qpel_avg = s->dsp.avg_h264_qpel_pixels_tab;
2870     }
2871
2872     first_mb_in_slice = get_ue_golomb_long(&s->gb);
2873
2874     if (first_mb_in_slice == 0) { // FIXME better field boundary detection
2875         if (h0->current_slice && FIELD_PICTURE) {
2876             field_end(h, 1);
2877         }
2878
2879         h0->current_slice = 0;
2880         if (!s0->first_field) {
2881             if (s->current_picture_ptr && !s->dropable &&
2882                 s->current_picture_ptr->owner2 == s) {
2883                 ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX,
2884                                           s->picture_structure == PICT_BOTTOM_FIELD);
2885             }
2886             s->current_picture_ptr = NULL;
2887         }
2888     }
2889
2890     slice_type = get_ue_golomb_31(&s->gb);
2891     if (slice_type > 9) {
2892         av_log(h->s.avctx, AV_LOG_ERROR,
2893                "slice type too large (%d) at %d %d\n",
2894                h->slice_type, s->mb_x, s->mb_y);
2895         return -1;
2896     }
2897     if (slice_type > 4) {
2898         slice_type -= 5;
2899         h->slice_type_fixed = 1;
2900     } else
2901         h->slice_type_fixed = 0;
2902
2903     slice_type = golomb_to_pict_type[slice_type];
2904     if (slice_type == AV_PICTURE_TYPE_I ||
2905         (h0->current_slice != 0 && slice_type == h0->last_slice_type)) {
2906         default_ref_list_done = 1;
2907     }
2908     h->slice_type     = slice_type;
2909     h->slice_type_nos = slice_type & 3;
2910
2911     // to make a few old functions happy, it's wrong though
2912     s->pict_type = h->slice_type;
2913
2914     pps_id = get_ue_golomb(&s->gb);
2915     if (pps_id >= MAX_PPS_COUNT) {
2916         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id %d out of range\n", pps_id);
2917         return -1;
2918     }
2919     if (!h0->pps_buffers[pps_id]) {
2920         av_log(h->s.avctx, AV_LOG_ERROR,
2921                "non-existing PPS %u referenced\n",
2922                pps_id);
2923         return -1;
2924     }
2925     h->pps = *h0->pps_buffers[pps_id];
2926
2927     if (!h0->sps_buffers[h->pps.sps_id]) {
2928         av_log(h->s.avctx, AV_LOG_ERROR,
2929                "non-existing SPS %u referenced\n",
2930                h->pps.sps_id);
2931         return -1;
2932     }
2933     h->sps = *h0->sps_buffers[h->pps.sps_id];
2934
2935     s->avctx->profile = ff_h264_get_profile(&h->sps);
2936     s->avctx->level   = h->sps.level_idc;
2937     s->avctx->refs    = h->sps.ref_frame_count;
2938
2939     must_reinit = (s->context_initialized &&
2940                     (   16*h->sps.mb_width != s->avctx->coded_width
2941                      || 16*h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag) != s->avctx->coded_height
2942                      || s->avctx->bits_per_raw_sample != h->sps.bit_depth_luma
2943                      || h->cur_chroma_format_idc != h->sps.chroma_format_idc
2944                      || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio)));
2945
2946     if(must_reinit && (h != h0 || (s->avctx->active_thread_type & FF_THREAD_FRAME))) {
2947         av_log_missing_feature(s->avctx,
2948                                 "Width/height/bit depth/chroma idc changing with threads is", 0);
2949         return -1;   // width / height changed during parallelized decoding
2950     }
2951
2952     s->mb_width  = h->sps.mb_width;
2953     s->mb_height = h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2954
2955     h->b_stride = s->mb_width * 4;
2956
2957     s->chroma_y_shift = h->sps.chroma_format_idc <= 1; // 400 uses yuv420p
2958
2959     s->width  = 16 * s->mb_width;
2960     s->height = 16 * s->mb_height;
2961
2962
2963     if(must_reinit) {
2964         free_tables(h, 0);
2965         flush_dpb(s->avctx);
2966         ff_MPV_common_end(s);
2967         h->list_count = 0;
2968     }
2969     if (!s->context_initialized) {
2970         if (h != h0) {
2971             av_log(h->s.avctx, AV_LOG_ERROR,
2972                    "Cannot (re-)initialize context during parallel decoding.\n");
2973             return -1;
2974         }
2975         avcodec_set_dimensions(s->avctx, s->width, s->height);
2976         s->avctx->width  -= (2>>CHROMA444)*FFMIN(h->sps.crop_right, (8<<CHROMA444)-1);
2977         s->avctx->height -= (1<<s->chroma_y_shift)*FFMIN(h->sps.crop_bottom, (16>>s->chroma_y_shift)-1) * (2 - h->sps.frame_mbs_only_flag);
2978         s->avctx->sample_aspect_ratio = h->sps.sar;
2979         av_assert0(s->avctx->sample_aspect_ratio.den);
2980
2981         if (s->avctx->bits_per_raw_sample != h->sps.bit_depth_luma ||
2982             h->cur_chroma_format_idc != h->sps.chroma_format_idc) {
2983             if (h->sps.bit_depth_luma >= 8 && h->sps.bit_depth_luma <= 10 &&
2984                 (h->sps.bit_depth_luma != 9 || !CHROMA422)) {
2985                 s->avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
2986                 h->cur_chroma_format_idc = h->sps.chroma_format_idc;
2987                 h->pixel_shift = h->sps.bit_depth_luma > 8;
2988
2989                 ff_h264dsp_init(&h->h264dsp, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
2990                 ff_h264_pred_init(&h->hpc, s->codec_id, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
2991                 s->dsp.dct_bits = h->sps.bit_depth_luma > 8 ? 32 : 16;
2992                 ff_dsputil_init(&s->dsp, s->avctx);
2993             } else {
2994                 av_log(s->avctx, AV_LOG_ERROR, "Unsupported bit depth: %d chroma_idc: %d\n",
2995                        h->sps.bit_depth_luma, h->sps.chroma_format_idc);
2996                 return -1;
2997             }
2998         }
2999
3000         if (h->sps.video_signal_type_present_flag) {
3001             s->avctx->color_range = h->sps.full_range>0 ? AVCOL_RANGE_JPEG
3002                                                       : AVCOL_RANGE_MPEG;
3003             if (h->sps.colour_description_present_flag) {
3004                 s->avctx->color_primaries = h->sps.color_primaries;
3005                 s->avctx->color_trc       = h->sps.color_trc;
3006                 s->avctx->colorspace      = h->sps.colorspace;
3007             }
3008         }
3009
3010         if (h->sps.timing_info_present_flag) {
3011             int64_t den = h->sps.time_scale;
3012             if (h->x264_build < 44U)
3013                 den *= 2;
3014             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
3015                       h->sps.num_units_in_tick, den, 1 << 30);
3016         }
3017
3018         switch (h->sps.bit_depth_luma) {
3019         case 9:
3020             if (CHROMA444) {
3021                 if (s->avctx->colorspace == AVCOL_SPC_RGB) {
3022                     s->avctx->pix_fmt = PIX_FMT_GBRP9;
3023                 } else
3024                     s->avctx->pix_fmt = PIX_FMT_YUV444P9;
3025             } else if (CHROMA422)
3026                 s->avctx->pix_fmt = PIX_FMT_YUV422P9;
3027             else
3028                 s->avctx->pix_fmt = PIX_FMT_YUV420P9;
3029             break;
3030         case 10:
3031             if (CHROMA444) {
3032                 if (s->avctx->colorspace == AVCOL_SPC_RGB) {
3033                     s->avctx->pix_fmt = PIX_FMT_GBRP10;
3034                 } else
3035                     s->avctx->pix_fmt = PIX_FMT_YUV444P10;
3036             } else if (CHROMA422)
3037                 s->avctx->pix_fmt = PIX_FMT_YUV422P10;
3038             else
3039                 s->avctx->pix_fmt = PIX_FMT_YUV420P10;
3040             break;
3041         case 8:
3042             if (CHROMA444) {
3043                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ444P
3044                                                                                   : PIX_FMT_YUV444P;
3045                     if (s->avctx->colorspace == AVCOL_SPC_RGB) {
3046                         s->avctx->pix_fmt = PIX_FMT_GBR24P;
3047                         av_log(h->s.avctx, AV_LOG_DEBUG, "Detected GBR colorspace.\n");
3048                     } else if (s->avctx->colorspace == AVCOL_SPC_YCGCO) {
3049                         av_log(h->s.avctx, AV_LOG_WARNING, "Detected unsupported YCgCo colorspace.\n");
3050                     }
3051             } else if (CHROMA422) {
3052                 s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ422P
3053                                                                               : PIX_FMT_YUV422P;
3054             } else {
3055                 s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
3056                                                          s->avctx->codec->pix_fmts ?
3057                                                          s->avctx->codec->pix_fmts :
3058                                                          s->avctx->color_range == AVCOL_RANGE_JPEG ?
3059                                                          hwaccel_pixfmt_list_h264_jpeg_420 :
3060                                                          ff_hwaccel_pixfmt_list_420);
3061             }
3062             break;
3063         default:
3064             av_log(s->avctx, AV_LOG_ERROR,
3065                    "Unsupported bit depth: %d\n", h->sps.bit_depth_luma);
3066             return AVERROR_INVALIDDATA;
3067         }
3068
3069         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id,
3070                                             s->avctx->pix_fmt);
3071
3072         if (ff_MPV_common_init(s) < 0) {
3073             av_log(h->s.avctx, AV_LOG_ERROR, "ff_MPV_common_init() failed.\n");
3074             return -1;
3075         }
3076         s->first_field = 0;
3077         h->prev_interlaced_frame = 1;
3078
3079         init_scan_tables(h);
3080         if (ff_h264_alloc_tables(h) < 0) {
3081             av_log(h->s.avctx, AV_LOG_ERROR,
3082                    "Could not allocate memory for h264\n");
3083             return AVERROR(ENOMEM);
3084         }
3085
3086         if (!HAVE_THREADS || !(s->avctx->active_thread_type & FF_THREAD_SLICE)) {
3087             if (context_init(h) < 0) {
3088                 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
3089                 return -1;
3090             }
3091         } else {
3092             for (i = 1; i < s->slice_context_count; i++) {
3093                 H264Context *c;
3094                 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
3095                 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
3096                 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
3097                 c->h264dsp     = h->h264dsp;
3098                 c->sps         = h->sps;
3099                 c->pps         = h->pps;
3100                 c->pixel_shift = h->pixel_shift;
3101                 c->cur_chroma_format_idc = h->cur_chroma_format_idc;
3102                 init_scan_tables(c);
3103                 clone_tables(c, h, i);
3104             }
3105
3106             for (i = 0; i < s->slice_context_count; i++)
3107                 if (context_init(h->thread_context[i]) < 0) {
3108                     av_log(h->s.avctx, AV_LOG_ERROR,
3109                            "context_init() failed.\n");
3110                     return -1;
3111                 }
3112         }
3113     }
3114
3115     if (h == h0 && h->dequant_coeff_pps != pps_id) {
3116         h->dequant_coeff_pps = pps_id;
3117         init_dequant_tables(h);
3118     }
3119
3120     h->frame_num = get_bits(&s->gb, h->sps.log2_max_frame_num);
3121
3122     h->mb_mbaff        = 0;
3123     h->mb_aff_frame    = 0;
3124     last_pic_structure = s0->picture_structure;
3125     last_pic_dropable  = s->dropable;
3126     s->dropable        = h->nal_ref_idc == 0;
3127     if (h->sps.frame_mbs_only_flag) {
3128         s->picture_structure = PICT_FRAME;
3129     } else {
3130         if (!h->sps.direct_8x8_inference_flag && slice_type == AV_PICTURE_TYPE_B) {
3131             av_log(h->s.avctx, AV_LOG_ERROR, "This stream was generated by a broken encoder, invalid 8x8 inference\n");
3132             return -1;
3133         }
3134         if (get_bits1(&s->gb)) { // field_pic_flag
3135             s->picture_structure = PICT_TOP_FIELD + get_bits1(&s->gb); // bottom_field_flag
3136         } else {
3137             s->picture_structure = PICT_FRAME;
3138             h->mb_aff_frame      = h->sps.mb_aff;
3139         }
3140     }
3141     h->mb_field_decoding_flag = s->picture_structure != PICT_FRAME;
3142
3143     if (h0->current_slice != 0) {
3144         if (last_pic_structure != s->picture_structure ||
3145             last_pic_dropable  != s->dropable) {
3146             av_log(h->s.avctx, AV_LOG_ERROR,
3147                    "Changing field mode (%d -> %d) between slices is not allowed\n",
3148                    last_pic_structure, s->picture_structure);
3149             s->picture_structure = last_pic_structure;
3150             s->dropable          = last_pic_dropable;
3151             return AVERROR_INVALIDDATA;
3152         }
3153     } else {
3154         /* Shorten frame num gaps so we don't have to allocate reference
3155          * frames just to throw them away */
3156         if (h->frame_num != h->prev_frame_num && h->prev_frame_num >= 0) {
3157             int unwrap_prev_frame_num = h->prev_frame_num;
3158             int max_frame_num         = 1 << h->sps.log2_max_frame_num;
3159
3160             if (unwrap_prev_frame_num > h->frame_num)
3161                 unwrap_prev_frame_num -= max_frame_num;
3162
3163             if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
3164                 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
3165                 if (unwrap_prev_frame_num < 0)
3166                     unwrap_prev_frame_num += max_frame_num;
3167
3168                 h->prev_frame_num = unwrap_prev_frame_num;
3169             }
3170         }
3171
3172         /* See if we have a decoded first field looking for a pair...
3173          * Here, we're using that to see if we should mark previously
3174          * decode frames as "finished".
3175          * We have to do that before the "dummy" in-between frame allocation,
3176          * since that can modify s->current_picture_ptr. */
3177         if (s0->first_field) {
3178             assert(s0->current_picture_ptr);
3179             assert(s0->current_picture_ptr->f.data[0]);
3180             assert(s0->current_picture_ptr->f.reference != DELAYED_PIC_REF);
3181
3182             /* Mark old field/frame as completed */
3183             if (!last_pic_dropable && s0->current_picture_ptr->owner2 == s0) {
3184                 ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
3185                                           last_pic_structure == PICT_BOTTOM_FIELD);
3186             }
3187
3188             /* figure out if we have a complementary field pair */
3189             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
3190                 /* Previous field is unmatched. Don't display it, but let it
3191                  * remain for reference if marked as such. */
3192                 if (!last_pic_dropable && last_pic_structure != PICT_FRAME) {
3193                     ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
3194                                               last_pic_structure == PICT_TOP_FIELD);
3195                 }
3196             } else {
3197                 if (s0->current_picture_ptr->frame_num != h->frame_num) {
3198                     /* This and previous field were reference, but had
3199                      * different frame_nums. Consider this field first in
3200                      * pair. Throw away previous field except for reference
3201                      * purposes. */
3202                     if (!last_pic_dropable && last_pic_structure != PICT_FRAME) {
3203                         ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
3204                                                   last_pic_structure == PICT_TOP_FIELD);
3205                     }
3206                 } else {
3207                     /* Second field in complementary pair */
3208                     if (!((last_pic_structure   == PICT_TOP_FIELD &&
3209                            s->picture_structure == PICT_BOTTOM_FIELD) ||
3210                           (last_pic_structure   == PICT_BOTTOM_FIELD &&
3211                            s->picture_structure == PICT_TOP_FIELD))) {
3212                         av_log(s->avctx, AV_LOG_ERROR,
3213                                "Invalid field mode combination %d/%d\n",
3214                                last_pic_structure, s->picture_structure);
3215                         s->picture_structure = last_pic_structure;
3216                         s->dropable          = last_pic_dropable;
3217                         return AVERROR_INVALIDDATA;
3218                     } else if (last_pic_dropable != s->dropable) {
3219                         av_log(s->avctx, AV_LOG_ERROR,
3220                                "Cannot combine reference and non-reference fields in the same frame\n");
3221                         av_log_ask_for_sample(s->avctx, NULL);
3222                         s->picture_structure = last_pic_structure;
3223                         s->dropable          = last_pic_dropable;
3224                         return AVERROR_INVALIDDATA;
3225                     }
3226
3227                     /* Take ownership of this buffer. Note that if another thread owned
3228                      * the first field of this buffer, we're not operating on that pointer,
3229                      * so the original thread is still responsible for reporting progress
3230                      * on that first field (or if that was us, we just did that above).
3231                      * By taking ownership, we assign responsibility to ourselves to
3232                      * report progress on the second field. */
3233                     s0->current_picture_ptr->owner2 = s0;
3234                 }
3235             }
3236         }
3237
3238         while (h->frame_num != h->prev_frame_num && h->prev_frame_num >= 0 &&
3239                h->frame_num != (h->prev_frame_num + 1) % (1 << h->sps.log2_max_frame_num)) {
3240             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
3241             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n",
3242                    h->frame_num, h->prev_frame_num);
3243             if (ff_h264_frame_start(h) < 0)
3244                 return -1;
3245             h->prev_frame_num++;
3246             h->prev_frame_num %= 1 << h->sps.log2_max_frame_num;
3247             s->current_picture_ptr->frame_num = h->prev_frame_num;
3248             ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX, 0);
3249             ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX, 1);
3250             ff_generate_sliding_window_mmcos(h);
3251             if (ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index) < 0 &&
3252                 (s->avctx->err_recognition & AV_EF_EXPLODE))
3253                 return AVERROR_INVALIDDATA;
3254             /* Error concealment: if a ref is missing, copy the previous ref in its place.
3255              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
3256              * about there being no actual duplicates.
3257              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
3258              * concealing a lost frame, this probably isn't noticeable by comparison, but it should
3259              * be fixed. */
3260             if (h->short_ref_count) {
3261                 if (prev) {
3262                     av_image_copy(h->short_ref[0]->f.data, h->short_ref[0]->f.linesize,
3263                                   (const uint8_t **)prev->f.data, prev->f.linesize,
3264                                   s->avctx->pix_fmt, s->mb_width * 16, s->mb_height * 16);
3265                     h->short_ref[0]->poc = prev->poc + 2;
3266                 }
3267                 h->short_ref[0]->frame_num = h->prev_frame_num;
3268             }
3269         }
3270
3271         /* See if we have a decoded first field looking for a pair...
3272          * We're using that to see whether to continue decoding in that
3273          * frame, or to allocate a new one. */
3274         if (s0->first_field) {
3275             assert(s0->current_picture_ptr);
3276             assert(s0->current_picture_ptr->f.data[0]);
3277             assert(s0->current_picture_ptr->f.reference != DELAYED_PIC_REF);
3278
3279             /* figure out if we have a complementary field pair */
3280             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
3281                 /* Previous field is unmatched. Don't display it, but let it
3282                  * remain for reference if marked as such. */
3283                 s0->current_picture_ptr = NULL;
3284                 s0->first_field         = FIELD_PICTURE;
3285             } else {
3286                 if (s0->current_picture_ptr->frame_num != h->frame_num) {
3287                     ff_thread_report_progress((AVFrame*)s0->current_picture_ptr, INT_MAX,
3288                                               s0->picture_structure==PICT_BOTTOM_FIELD);
3289                     /* This and the previous field had different frame_nums.
3290                      * Consider this field first in pair. Throw away previous
3291                      * one except for reference purposes. */
3292                     s0->first_field         = 1;
3293                     s0->current_picture_ptr = NULL;
3294                 } else {
3295                     /* Second field in complementary pair */
3296                     s0->first_field = 0;
3297                 }
3298             }
3299         } else {
3300             /* Frame or first field in a potentially complementary pair */
3301             assert(!s0->current_picture_ptr);
3302             s0->first_field = FIELD_PICTURE;
3303         }
3304
3305         if (!FIELD_PICTURE || s0->first_field) {
3306             if (ff_h264_frame_start(h) < 0) {
3307                 s0->first_field = 0;
3308                 return -1;
3309             }
3310         } else {
3311             ff_release_unused_pictures(s, 0);
3312         }
3313     }
3314     if (h != h0)
3315         clone_slice(h, h0);
3316
3317     s->current_picture_ptr->frame_num = h->frame_num; // FIXME frame_num cleanup
3318
3319     assert(s->mb_num == s->mb_width * s->mb_height);
3320     if (first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
3321         first_mb_in_slice >= s->mb_num) {
3322         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
3323         return -1;
3324     }
3325     s->resync_mb_x = s->mb_x =  first_mb_in_slice % s->mb_width;
3326     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
3327     if (s->picture_structure == PICT_BOTTOM_FIELD)
3328         s->resync_mb_y = s->mb_y = s->mb_y + 1;
3329     assert(s->mb_y < s->mb_height);
3330
3331     if (s->picture_structure == PICT_FRAME) {
3332         h->curr_pic_num = h->frame_num;
3333         h->max_pic_num  = 1 << h->sps.log2_max_frame_num;
3334     } else {
3335         h->curr_pic_num = 2 * h->frame_num + 1;
3336         h->max_pic_num  = 1 << (h->sps.log2_max_frame_num + 1);
3337     }
3338
3339     if (h->nal_unit_type == NAL_IDR_SLICE)
3340         get_ue_golomb(&s->gb); /* idr_pic_id */
3341
3342     if (h->sps.poc_type == 0) {
3343         h->poc_lsb = get_bits(&s->gb, h->sps.log2_max_poc_lsb);
3344
3345         if (h->pps.pic_order_present == 1 && s->picture_structure == PICT_FRAME)
3346             h->delta_poc_bottom = get_se_golomb(&s->gb);
3347     }
3348
3349     if (h->sps.poc_type == 1 && !h->sps.delta_pic_order_always_zero_flag) {
3350         h->delta_poc[0] = get_se_golomb(&s->gb);
3351
3352         if (h->pps.pic_order_present == 1 && s->picture_structure == PICT_FRAME)
3353             h->delta_poc[1] = get_se_golomb(&s->gb);
3354     }
3355
3356     init_poc(h);
3357
3358     if (h->pps.redundant_pic_cnt_present)
3359         h->redundant_pic_count = get_ue_golomb(&s->gb);
3360
3361     // set defaults, might be overridden a few lines later
3362     h->ref_count[0] = h->pps.ref_count[0];
3363     h->ref_count[1] = h->pps.ref_count[1];
3364
3365     if (h->slice_type_nos != AV_PICTURE_TYPE_I) {
3366         unsigned max = s->picture_structure == PICT_FRAME ? 15 : 31;
3367
3368         if (h->slice_type_nos == AV_PICTURE_TYPE_B)
3369             h->direct_spatial_mv_pred = get_bits1(&s->gb);
3370         num_ref_idx_active_override_flag = get_bits1(&s->gb);
3371
3372         if (num_ref_idx_active_override_flag) {
3373             h->ref_count[0] = get_ue_golomb(&s->gb) + 1;
3374             if (h->slice_type_nos == AV_PICTURE_TYPE_B)
3375                 h->ref_count[1] = get_ue_golomb(&s->gb) + 1;
3376         }
3377
3378         if (h->ref_count[0]-1 > max || h->ref_count[1]-1 > max){
3379             av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
3380             h->ref_count[0] = h->ref_count[1] = 1;
3381             return AVERROR_INVALIDDATA;
3382         }
3383
3384         if (h->slice_type_nos == AV_PICTURE_TYPE_B)
3385             h->list_count = 2;
3386         else
3387             h->list_count = 1;
3388     } else
3389         h->ref_count[1]= h->ref_count[0]= h->list_count= 0;
3390
3391     if (!default_ref_list_done)
3392         ff_h264_fill_default_ref_list(h);
3393
3394     if (h->slice_type_nos != AV_PICTURE_TYPE_I &&
3395         ff_h264_decode_ref_pic_list_reordering(h) < 0) {
3396         h->ref_count[1] = h->ref_count[0] = 0;
3397         return -1;
3398     }
3399
3400     if (h->slice_type_nos != AV_PICTURE_TYPE_I) {
3401         s->last_picture_ptr = &h->ref_list[0][0];
3402         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
3403     }
3404     if (h->slice_type_nos == AV_PICTURE_TYPE_B) {
3405         s->next_picture_ptr = &h->ref_list[1][0];
3406         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
3407     }
3408
3409     if ((h->pps.weighted_pred && h->slice_type_nos == AV_PICTURE_TYPE_P) ||
3410         (h->pps.weighted_bipred_idc == 1 &&
3411          h->slice_type_nos == AV_PICTURE_TYPE_B))
3412         pred_weight_table(h);
3413     else if (h->pps.weighted_bipred_idc == 2 &&
3414              h->slice_type_nos == AV_PICTURE_TYPE_B) {
3415         implicit_weight_table(h, -1);
3416     } else {
3417         h->use_weight = 0;
3418         for (i = 0; i < 2; i++) {
3419             h->luma_weight_flag[i]   = 0;
3420             h->chroma_weight_flag[i] = 0;
3421         }
3422     }
3423
3424     if (h->nal_ref_idc && ff_h264_decode_ref_pic_marking(h0, &s->gb) < 0 &&
3425         (s->avctx->err_recognition & AV_EF_EXPLODE))
3426         return AVERROR_INVALIDDATA;
3427
3428     if (FRAME_MBAFF) {
3429         ff_h264_fill_mbaff_ref_list(h);
3430
3431         if (h->pps.weighted_bipred_idc == 2 && h->slice_type_nos == AV_PICTURE_TYPE_B) {
3432             implicit_weight_table(h, 0);
3433             implicit_weight_table(h, 1);
3434         }
3435     }
3436
3437     if (h->slice_type_nos == AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
3438         ff_h264_direct_dist_scale_factor(h);
3439     ff_h264_direct_ref_list_init(h);
3440
3441     if (h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac) {
3442         tmp = get_ue_golomb_31(&s->gb);
3443         if (tmp > 2) {
3444             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
3445             return -1;
3446         }
3447         h->cabac_init_idc = tmp;
3448     }
3449
3450     h->last_qscale_diff = 0;
3451     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
3452     if (tmp > 51 + 6 * (h->sps.bit_depth_luma - 8)) {
3453         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
3454         return -1;
3455     }
3456     s->qscale       = tmp;
3457     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3458     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3459     // FIXME qscale / qp ... stuff
3460     if (h->slice_type == AV_PICTURE_TYPE_SP)
3461         get_bits1(&s->gb); /* sp_for_switch_flag */
3462     if (h->slice_type == AV_PICTURE_TYPE_SP ||
3463         h->slice_type == AV_PICTURE_TYPE_SI)
3464         get_se_golomb(&s->gb); /* slice_qs_delta */
3465
3466     h->deblocking_filter     = 1;
3467     h->slice_alpha_c0_offset = 52;
3468     h->slice_beta_offset     = 52;
3469     if (h->pps.deblocking_filter_parameters_present) {
3470         tmp = get_ue_golomb_31(&s->gb);
3471         if (tmp > 2) {
3472             av_log(s->avctx, AV_LOG_ERROR,
3473                    "deblocking_filter_idc %u out of range\n", tmp);
3474             return -1;
3475         }
3476         h->deblocking_filter = tmp;
3477         if (h->deblocking_filter < 2)
3478             h->deblocking_filter ^= 1;  // 1<->0
3479
3480         if (h->deblocking_filter) {
3481             h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
3482             h->slice_beta_offset     += get_se_golomb(&s->gb) << 1;
3483             if (h->slice_alpha_c0_offset > 104U ||
3484                 h->slice_beta_offset     > 104U) {
3485                 av_log(s->avctx, AV_LOG_ERROR,
3486                        "deblocking filter parameters %d %d out of range\n",
3487                        h->slice_alpha_c0_offset, h->slice_beta_offset);
3488                 return -1;
3489             }
3490         }
3491     }
3492
3493     if (s->avctx->skip_loop_filter >= AVDISCARD_ALL ||
3494         (s->avctx->skip_loop_filter >= AVDISCARD_NONKEY &&
3495          h->slice_type_nos != AV_PICTURE_TYPE_I) ||
3496         (s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  &&
3497          h->slice_type_nos == AV_PICTURE_TYPE_B) ||
3498         (s->avctx->skip_loop_filter >= AVDISCARD_NONREF &&
3499          h->nal_ref_idc == 0))
3500         h->deblocking_filter = 0;
3501
3502     if (h->deblocking_filter == 1 && h0->max_contexts > 1) {
3503         if (s->avctx->flags2 & CODEC_FLAG2_FAST) {
3504             /* Cheat slightly for speed:
3505              * Do not bother to deblock across slices. */
3506             h->deblocking_filter = 2;
3507         } else {
3508             h0->max_contexts = 1;
3509             if (!h0->single_decode_warning) {
3510                 av_log(s->avctx, AV_LOG_INFO,
3511                        "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
3512                 h0->single_decode_warning = 1;
3513             }
3514             if (h != h0) {
3515                 av_log(h->s.avctx, AV_LOG_ERROR,
3516                        "Deblocking switched inside frame.\n");
3517                 return 1;
3518             }
3519         }
3520     }
3521     h->qp_thresh = 15 + 52 -
3522                    FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset) -
3523                    FFMAX3(0,
3524                           h->pps.chroma_qp_index_offset[0],
3525                           h->pps.chroma_qp_index_offset[1]) +
3526                    6 * (h->sps.bit_depth_luma - 8);
3527
3528     h0->last_slice_type = slice_type;
3529     h->slice_num = ++h0->current_slice;
3530
3531     if (h->slice_num)
3532         h0->slice_row[(h->slice_num-1)&(MAX_SLICES-1)]= s->resync_mb_y;
3533     if (   h0->slice_row[h->slice_num&(MAX_SLICES-1)] + 3 >= s->resync_mb_y
3534         && h0->slice_row[h->slice_num&(MAX_SLICES-1)] <= s->resync_mb_y
3535         && h->slice_num >= MAX_SLICES) {
3536         //in case of ASO this check needs to be updated depending on how we decide to assign slice numbers in this case
3537         av_log(s->avctx, AV_LOG_WARNING, "Possibly too many slices (%d >= %d), increase MAX_SLICES and recompile if there are artifacts\n", h->slice_num, MAX_SLICES);
3538     }
3539
3540     for (j = 0; j < 2; j++) {
3541         int id_list[16];
3542         int *ref2frm = h->ref2frm[h->slice_num & (MAX_SLICES - 1)][j];
3543         for (i = 0; i < 16; i++) {
3544             id_list[i] = 60;
3545             if (h->ref_list[j][i].f.data[0]) {
3546                 int k;
3547                 uint8_t *base = h->ref_list[j][i].f.base[0];
3548                 for (k = 0; k < h->short_ref_count; k++)
3549                     if (h->short_ref[k]->f.base[0] == base) {
3550                         id_list[i] = k;
3551                         break;
3552                     }
3553                 for (k = 0; k < h->long_ref_count; k++)
3554                     if (h->long_ref[k] && h->long_ref[k]->f.base[0] == base) {
3555                         id_list[i] = h->short_ref_count + k;
3556                         break;
3557                     }
3558             }
3559         }
3560
3561         ref2frm[0]     =
3562             ref2frm[1] = -1;
3563         for (i = 0; i < 16; i++)
3564             ref2frm[i + 2] = 4 * id_list[i] +
3565                              (h->ref_list[j][i].f.reference & 3);
3566         ref2frm[18 + 0]     =
3567             ref2frm[18 + 1] = -1;
3568         for (i = 16; i < 48; i++)
3569             ref2frm[i + 4] = 4 * id_list[(i - 16) >> 1] +
3570                              (h->ref_list[j][i].f.reference & 3);
3571     }
3572
3573     // FIXME: fix draw_edges + PAFF + frame threads
3574     h->emu_edge_width  = (s->flags & CODEC_FLAG_EMU_EDGE ||
3575                           (!h->sps.frame_mbs_only_flag &&
3576                            s->avctx->active_thread_type))
3577                          ? 0 : 16;
3578     h->emu_edge_height = (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
3579
3580     if (s->avctx->debug & FF_DEBUG_PICT_INFO) {
3581         av_log(h->s.avctx, AV_LOG_DEBUG,
3582                "slice:%d %s mb:%d %c%s%s pps:%u frame:%d poc:%d/%d ref:%d/%d qp:%d loop:%d:%d:%d weight:%d%s %s\n",
3583                h->slice_num,
3584                (s->picture_structure == PICT_FRAME ? "F" : s->picture_structure == PICT_TOP_FIELD ? "T" : "B"),
3585                first_mb_in_slice,
3586                av_get_picture_type_char(h->slice_type),
3587                h->slice_type_fixed ? " fix" : "",
3588                h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
3589                pps_id, h->frame_num,
3590                s->current_picture_ptr->field_poc[0],
3591                s->current_picture_ptr->field_poc[1],
3592                h->ref_count[0], h->ref_count[1],
3593                s->qscale,
3594                h->deblocking_filter,
3595                h->slice_alpha_c0_offset / 2 - 26, h->slice_beta_offset / 2 - 26,
3596                h->use_weight,
3597                h->use_weight == 1 && h->use_weight_chroma ? "c" : "",
3598                h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : "");
3599     }
3600
3601     return 0;
3602 }
3603
3604 int ff_h264_get_slice_type(const H264Context *h)
3605 {
3606     switch (h->slice_type) {
3607     case AV_PICTURE_TYPE_P:
3608         return 0;
3609     case AV_PICTURE_TYPE_B:
3610         return 1;
3611     case AV_PICTURE_TYPE_I:
3612         return 2;
3613     case AV_PICTURE_TYPE_SP:
3614         return 3;
3615     case AV_PICTURE_TYPE_SI:
3616         return 4;
3617     default:
3618         return -1;
3619     }
3620 }
3621
3622 static av_always_inline void fill_filter_caches_inter(H264Context *h,
3623                                                       MpegEncContext *const s,
3624                                                       int mb_type, int top_xy,
3625                                                       int left_xy[LEFT_MBS],
3626                                                       int top_type,
3627                                                       int left_type[LEFT_MBS],
3628                                                       int mb_xy, int list)
3629 {
3630     int b_stride = h->b_stride;
3631     int16_t(*mv_dst)[2] = &h->mv_cache[list][scan8[0]];
3632     int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
3633     if (IS_INTER(mb_type) || IS_DIRECT(mb_type)) {
3634         if (USES_LIST(top_type, list)) {
3635             const int b_xy  = h->mb2b_xy[top_xy] + 3 * b_stride;
3636             const int b8_xy = 4 * top_xy + 2;
3637             int (*ref2frm)[64] = h->ref2frm[h->slice_table[top_xy] & (MAX_SLICES - 1)][0] + (MB_MBAFF ? 20 : 2);
3638             AV_COPY128(mv_dst - 1 * 8, s->current_picture.f.motion_val[list][b_xy + 0]);
3639             ref_cache[0 - 1 * 8] =
3640             ref_cache[1 - 1 * 8] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 0]];
3641             ref_cache[2 - 1 * 8] =
3642             ref_cache[3 - 1 * 8] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 1]];
3643         } else {
3644             AV_ZERO128(mv_dst - 1 * 8);
3645             AV_WN32A(&ref_cache[0 - 1 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3646         }
3647
3648         if (!IS_INTERLACED(mb_type ^ left_type[LTOP])) {
3649             if (USES_LIST(left_type[LTOP], list)) {
3650                 const int b_xy  = h->mb2b_xy[left_xy[LTOP]] + 3;
3651                 const int b8_xy = 4 * left_xy[LTOP] + 1;
3652                 int (*ref2frm)[64] = h->ref2frm[h->slice_table[left_xy[LTOP]] & (MAX_SLICES - 1)][0] + (MB_MBAFF ? 20 : 2);
3653                 AV_COPY32(mv_dst - 1 +  0, s->current_picture.f.motion_val[list][b_xy + b_stride * 0]);
3654                 AV_COPY32(mv_dst - 1 +  8, s->current_picture.f.motion_val[list][b_xy + b_stride * 1]);
3655                 AV_COPY32(mv_dst - 1 + 16, s->current_picture.f.motion_val[list][b_xy + b_stride * 2]);
3656                 AV_COPY32(mv_dst - 1 + 24, s->current_picture.f.motion_val[list][b_xy + b_stride * 3]);
3657                 ref_cache[-1 +  0] =
3658                 ref_cache[-1 +  8] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2 * 0]];
3659                 ref_cache[-1 + 16] =
3660                 ref_cache[-1 + 24] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2 * 1]];
3661             } else {
3662                 AV_ZERO32(mv_dst - 1 +  0);
3663                 AV_ZERO32(mv_dst - 1 +  8);
3664                 AV_ZERO32(mv_dst - 1 + 16);
3665                 AV_ZERO32(mv_dst - 1 + 24);
3666                 ref_cache[-1 +  0] =
3667                 ref_cache[-1 +  8] =
3668                 ref_cache[-1 + 16] =
3669                 ref_cache[-1 + 24] = LIST_NOT_USED;
3670             }
3671         }
3672     }
3673
3674     if (!USES_LIST(mb_type, list)) {
3675         fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0, 0), 4);
3676         AV_WN32A(&ref_cache[0 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3677         AV_WN32A(&ref_cache[1 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3678         AV_WN32A(&ref_cache[2 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3679         AV_WN32A(&ref_cache[3 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3680         return;
3681     }
3682
3683     {
3684         int8_t *ref = &s->current_picture.f.ref_index[list][4 * mb_xy];
3685         int (*ref2frm)[64] = h->ref2frm[h->slice_num & (MAX_SLICES - 1)][0] + (MB_MBAFF ? 20 : 2);
3686         uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]], ref2frm[list][ref[1]]) & 0x00FF00FF) * 0x0101;
3687         uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]], ref2frm[list][ref[3]]) & 0x00FF00FF) * 0x0101;
3688         AV_WN32A(&ref_cache[0 * 8], ref01);
3689         AV_WN32A(&ref_cache[1 * 8], ref01);
3690         AV_WN32A(&ref_cache[2 * 8], ref23);
3691         AV_WN32A(&ref_cache[3 * 8], ref23);
3692     }
3693
3694     {
3695         int16_t(*mv_src)[2] = &s->current_picture.f.motion_val[list][4 * s->mb_x + 4 * s->mb_y * b_stride];
3696         AV_COPY128(mv_dst + 8 * 0, mv_src + 0 * b_stride);
3697         AV_COPY128(mv_dst + 8 * 1, mv_src + 1 * b_stride);
3698         AV_COPY128(mv_dst + 8 * 2, mv_src + 2 * b_stride);
3699         AV_COPY128(mv_dst + 8 * 3, mv_src + 3 * b_stride);
3700     }
3701 }
3702
3703 /**
3704  *
3705  * @return non zero if the loop filter can be skipped
3706  */
3707 static int fill_filter_caches(H264Context *h, int mb_type)
3708 {
3709     MpegEncContext *const s = &h->s;
3710     const int mb_xy = h->mb_xy;
3711     int top_xy, left_xy[LEFT_MBS];
3712     int top_type, left_type[LEFT_MBS];
3713     uint8_t *nnz;
3714     uint8_t *nnz_cache;
3715
3716     top_xy = mb_xy - (s->mb_stride << MB_FIELD);
3717
3718     /* Wow, what a mess, why didn't they simplify the interlacing & intra
3719      * stuff, I can't imagine that these complex rules are worth it. */
3720
3721     left_xy[LBOT] = left_xy[LTOP] = mb_xy - 1;
3722     if (FRAME_MBAFF) {
3723         const int left_mb_field_flag = IS_INTERLACED(s->current_picture.f.mb_type[mb_xy - 1]);
3724         const int curr_mb_field_flag = IS_INTERLACED(mb_type);
3725         if (s->mb_y & 1) {
3726             if (left_mb_field_flag != curr_mb_field_flag)
3727                 left_xy[LTOP] -= s->mb_stride;
3728         } else {
3729             if (curr_mb_field_flag)
3730                 top_xy += s->mb_stride &
3731                     (((s->current_picture.f.mb_type[top_xy] >> 7) & 1) - 1);
3732             if (left_mb_field_flag != curr_mb_field_flag)
3733                 left_xy[LBOT] += s->mb_stride;
3734         }
3735     }
3736
3737     h->top_mb_xy        = top_xy;
3738     h->left_mb_xy[LTOP] = left_xy[LTOP];
3739     h->left_mb_xy[LBOT] = left_xy[LBOT];
3740     {
3741         /* For sufficiently low qp, filtering wouldn't do anything.
3742          * This is a conservative estimate: could also check beta_offset
3743          * and more accurate chroma_qp. */
3744         int qp_thresh = h->qp_thresh; // FIXME strictly we should store qp_thresh for each mb of a slice
3745         int qp        = s->current_picture.f.qscale_table[mb_xy];
3746         if (qp <= qp_thresh &&
3747             (left_xy[LTOP] < 0 ||
3748              ((qp + s->current_picture.f.qscale_table[left_xy[LTOP]] + 1) >> 1) <= qp_thresh) &&
3749             (top_xy < 0 ||
3750              ((qp + s->current_picture.f.qscale_table[top_xy] + 1) >> 1) <= qp_thresh)) {
3751             if (!FRAME_MBAFF)
3752                 return 1;
3753             if ((left_xy[LTOP] < 0 ||
3754                  ((qp + s->current_picture.f.qscale_table[left_xy[LBOT]] + 1) >> 1) <= qp_thresh) &&
3755                 (top_xy < s->mb_stride ||
3756                  ((qp + s->current_picture.f.qscale_table[top_xy - s->mb_stride] + 1) >> 1) <= qp_thresh))
3757                 return 1;
3758         }
3759     }
3760
3761     top_type        = s->current_picture.f.mb_type[top_xy];
3762     left_type[LTOP] = s->current_picture.f.mb_type[left_xy[LTOP]];
3763     left_type[LBOT] = s->current_picture.f.mb_type[left_xy[LBOT]];
3764     if (h->deblocking_filter == 2) {
3765         if (h->slice_table[top_xy] != h->slice_num)
3766             top_type = 0;
3767         if (h->slice_table[left_xy[LBOT]] != h->slice_num)
3768             left_type[LTOP] = left_type[LBOT] = 0;
3769     } else {
3770         if (h->slice_table[top_xy] == 0xFFFF)
3771             top_type = 0;
3772         if (h->slice_table[left_xy[LBOT]] == 0xFFFF)
3773             left_type[LTOP] = left_type[LBOT] = 0;
3774     }
3775     h->top_type        = top_type;
3776     h->left_type[LTOP] = left_type[LTOP];
3777     h->left_type[LBOT] = left_type[LBOT];
3778
3779     if (IS_INTRA(mb_type))
3780         return 0;
3781
3782     fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy,
3783                              top_type, left_type, mb_xy, 0);
3784     if (h->list_count == 2)
3785         fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy,
3786                                  top_type, left_type, mb_xy, 1);
3787
3788     nnz       = h->non_zero_count[mb_xy];
3789     nnz_cache = h->non_zero_count_cache;
3790     AV_COPY32(&nnz_cache[4 + 8 * 1], &nnz[0]);
3791     AV_COPY32(&nnz_cache[4 + 8 * 2], &nnz[4]);
3792     AV_COPY32(&nnz_cache[4 + 8 * 3], &nnz[8]);
3793     AV_COPY32(&nnz_cache[4 + 8 * 4], &nnz[12]);
3794     h->cbp = h->cbp_table[mb_xy];
3795
3796     if (top_type) {
3797         nnz = h->non_zero_count[top_xy];
3798         AV_COPY32(&nnz_cache[4 + 8 * 0], &nnz[3 * 4]);
3799     }
3800
3801     if (left_type[LTOP]) {
3802         nnz = h->non_zero_count[left_xy[LTOP]];
3803         nnz_cache[3 + 8 * 1] = nnz[3 + 0 * 4];
3804         nnz_cache[3 + 8 * 2] = nnz[3 + 1 * 4];
3805         nnz_cache[3 + 8 * 3] = nnz[3 + 2 * 4];
3806         nnz_cache[3 + 8 * 4] = nnz[3 + 3 * 4];
3807     }
3808
3809     /* CAVLC 8x8dct requires NNZ values for residual decoding that differ
3810      * from what the loop filter needs */
3811     if (!CABAC && h->pps.transform_8x8_mode) {
3812         if (IS_8x8DCT(top_type)) {
3813             nnz_cache[4 + 8 * 0]     =
3814                 nnz_cache[5 + 8 * 0] = (h->cbp_table[top_xy] & 0x4000) >> 12;
3815             nnz_cache[6 + 8 * 0]     =
3816                 nnz_cache[7 + 8 * 0] = (h->cbp_table[top_xy] & 0x8000) >> 12;
3817         }
3818         if (IS_8x8DCT(left_type[LTOP])) {
3819             nnz_cache[3 + 8 * 1]     =
3820                 nnz_cache[3 + 8 * 2] = (h->cbp_table[left_xy[LTOP]] & 0x2000) >> 12; // FIXME check MBAFF
3821         }
3822         if (IS_8x8DCT(left_type[LBOT])) {
3823             nnz_cache[3 + 8 * 3]     =
3824                 nnz_cache[3 + 8 * 4] = (h->cbp_table[left_xy[LBOT]] & 0x8000) >> 12; // FIXME check MBAFF
3825         }
3826
3827         if (IS_8x8DCT(mb_type)) {
3828             nnz_cache[scan8[0]] =
3829             nnz_cache[scan8[1]] =
3830             nnz_cache[scan8[2]] =
3831             nnz_cache[scan8[3]] = (h->cbp & 0x1000) >> 12;
3832
3833             nnz_cache[scan8[0 + 4]] =
3834             nnz_cache[scan8[1 + 4]] =
3835             nnz_cache[scan8[2 + 4]] =
3836             nnz_cache[scan8[3 + 4]] = (h->cbp & 0x2000) >> 12;
3837
3838             nnz_cache[scan8[0 + 8]] =
3839             nnz_cache[scan8[1 + 8]] =
3840             nnz_cache[scan8[2 + 8]] =
3841             nnz_cache[scan8[3 + 8]] = (h->cbp & 0x4000) >> 12;
3842
3843             nnz_cache[scan8[0 + 12]] =
3844             nnz_cache[scan8[1 + 12]] =
3845             nnz_cache[scan8[2 + 12]] =
3846             nnz_cache[scan8[3 + 12]] = (h->cbp & 0x8000) >> 12;
3847         }
3848     }
3849
3850     return 0;
3851 }
3852
3853 static void loop_filter(H264Context *h, int start_x, int end_x)
3854 {
3855     MpegEncContext *const s = &h->s;
3856     uint8_t *dest_y, *dest_cb, *dest_cr;
3857     int linesize, uvlinesize, mb_x, mb_y;
3858     const int end_mb_y       = s->mb_y + FRAME_MBAFF;
3859     const int old_slice_type = h->slice_type;
3860     const int pixel_shift    = h->pixel_shift;
3861     const int block_h        = 16 >> s->chroma_y_shift;
3862
3863     if (h->deblocking_filter) {
3864         for (mb_x = start_x; mb_x < end_x; mb_x++)
3865             for (mb_y = end_mb_y - FRAME_MBAFF; mb_y <= end_mb_y; mb_y++) {
3866                 int mb_xy, mb_type;
3867                 mb_xy         = h->mb_xy = mb_x + mb_y * s->mb_stride;
3868                 h->slice_num  = h->slice_table[mb_xy];
3869                 mb_type       = s->current_picture.f.mb_type[mb_xy];
3870                 h->list_count = h->list_counts[mb_xy];
3871
3872                 if (FRAME_MBAFF)
3873                     h->mb_mbaff               =
3874                     h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
3875
3876                 s->mb_x = mb_x;
3877                 s->mb_y = mb_y;
3878                 dest_y  = s->current_picture.f.data[0] +
3879                           ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
3880                 dest_cb = s->current_picture.f.data[1] +
3881                           (mb_x << pixel_shift) * (8 << CHROMA444) +
3882                           mb_y * s->uvlinesize * block_h;
3883                 dest_cr = s->current_picture.f.data[2] +
3884                           (mb_x << pixel_shift) * (8 << CHROMA444) +
3885                           mb_y * s->uvlinesize * block_h;
3886                 // FIXME simplify above
3887
3888                 if (MB_FIELD) {
3889                     linesize   = h->mb_linesize   = s->linesize   * 2;
3890                     uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
3891                     if (mb_y & 1) { // FIXME move out of this function?
3892                         dest_y  -= s->linesize   * 15;
3893                         dest_cb -= s->uvlinesize * (block_h - 1);
3894                         dest_cr -= s->uvlinesize * (block_h - 1);
3895                     }
3896                 } else {
3897                     linesize   = h->mb_linesize   = s->linesize;
3898                     uvlinesize = h->mb_uvlinesize = s->uvlinesize;
3899                 }
3900                 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize,
3901                                  uvlinesize, 0);
3902                 if (fill_filter_caches(h, mb_type))
3903                     continue;
3904                 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.f.qscale_table[mb_xy]);
3905                 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.f.qscale_table[mb_xy]);
3906
3907                 if (FRAME_MBAFF) {
3908                     ff_h264_filter_mb(h, mb_x, mb_y, dest_y, dest_cb, dest_cr,
3909                                       linesize, uvlinesize);
3910                 } else {
3911                     ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb,
3912                                            dest_cr, linesize, uvlinesize);
3913                 }
3914             }
3915     }
3916     h->slice_type   = old_slice_type;
3917     s->mb_x         = end_x;
3918     s->mb_y         = end_mb_y - FRAME_MBAFF;
3919     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3920     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3921 }
3922
3923 static void predict_field_decoding_flag(H264Context *h)
3924 {
3925     MpegEncContext *const s = &h->s;
3926     const int mb_xy = s->mb_x + s->mb_y * s->mb_stride;
3927     int mb_type     = (h->slice_table[mb_xy - 1] == h->slice_num) ?
3928                       s->current_picture.f.mb_type[mb_xy - 1] :
3929                       (h->slice_table[mb_xy - s->mb_stride] == h->slice_num) ?
3930                       s->current_picture.f.mb_type[mb_xy - s->mb_stride] : 0;
3931     h->mb_mbaff     = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
3932 }
3933
3934 /**
3935  * Draw edges and report progress for the last MB row.
3936  */
3937 static void decode_finish_row(H264Context *h)
3938 {
3939     MpegEncContext *const s = &h->s;
3940     int top            = 16 * (s->mb_y      >> FIELD_PICTURE);
3941     int pic_height     = 16 *  s->mb_height >> FIELD_PICTURE;
3942     int height         =  16      << FRAME_MBAFF;
3943     int deblock_border = (16 + 4) << FRAME_MBAFF;
3944
3945     if (h->deblocking_filter) {
3946         if ((top + height) >= pic_height)
3947             height += deblock_border;
3948         top -= deblock_border;
3949     }
3950
3951     if (top >= pic_height || (top + height) < h->emu_edge_height)
3952         return;
3953
3954     height = FFMIN(height, pic_height - top);
3955     if (top < h->emu_edge_height) {
3956         height = top + height;
3957         top    = 0;
3958     }
3959
3960     ff_draw_horiz_band(s, top, height);
3961
3962     if (s->dropable)
3963         return;
3964
3965     ff_thread_report_progress(&s->current_picture_ptr->f, top + height - 1,
3966                               s->picture_structure == PICT_BOTTOM_FIELD);
3967 }
3968
3969 static int decode_slice(struct AVCodecContext *avctx, void *arg)
3970 {
3971     H264Context *h = *(void **)arg;
3972     MpegEncContext *const s = &h->s;
3973     const int part_mask     = s->partitioned_frame ? (ER_AC_END | ER_AC_ERROR)
3974                                                    : 0x7F;
3975     int lf_x_start = s->mb_x;
3976
3977     s->mb_skip_run = -1;
3978
3979     h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME ||
3980                     s->codec_id != CODEC_ID_H264 ||
3981                     (CONFIG_GRAY && (s->flags & CODEC_FLAG_GRAY));
3982
3983     if (h->pps.cabac) {
3984         /* realign */
3985         align_get_bits(&s->gb);
3986
3987         /* init cabac */
3988         ff_init_cabac_states(&h->cabac);
3989         ff_init_cabac_decoder(&h->cabac,
3990                               s->gb.buffer + get_bits_count(&s->gb) / 8,
3991                               (get_bits_left(&s->gb) + 7) / 8);
3992
3993         ff_h264_init_cabac_states(h);
3994
3995         for (;;) {
3996             // START_TIMER
3997             int ret = ff_h264_decode_mb_cabac(h);
3998             int eos;
3999             // STOP_TIMER("decode_mb_cabac")
4000
4001             if (ret >= 0)
4002                 ff_h264_hl_decode_mb(h);
4003
4004             // FIXME optimal? or let mb_decode decode 16x32 ?
4005             if (ret >= 0 && FRAME_MBAFF) {
4006                 s->mb_y++;
4007
4008                 ret = ff_h264_decode_mb_cabac(h);
4009
4010                 if (ret >= 0)
4011                     ff_h264_hl_decode_mb(h);
4012                 s->mb_y--;
4013             }
4014             eos = get_cabac_terminate(&h->cabac);
4015
4016             if ((s->workaround_bugs & FF_BUG_TRUNCATED) &&
4017                 h->cabac.bytestream > h->cabac.bytestream_end + 2) {
4018                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x - 1,
4019                                 s->mb_y, ER_MB_END & part_mask);
4020                 if (s->mb_x >= lf_x_start)
4021                     loop_filter(h, lf_x_start, s->mb_x + 1);
4022                 return 0;
4023             }
4024             if (ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 2) {
4025                 av_log(h->s.avctx, AV_LOG_ERROR,
4026                        "error while decoding MB %d %d, bytestream (%td)\n",
4027                        s->mb_x, s->mb_y,
4028                        h->cabac.bytestream_end - h->cabac.bytestream);
4029                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x,
4030                                 s->mb_y, ER_MB_ERROR & part_mask);
4031                 return -1;
4032             }
4033
4034             if (++s->mb_x >= s->mb_width) {
4035                 loop_filter(h, lf_x_start, s->mb_x);
4036                 s->mb_x = lf_x_start = 0;
4037                 decode_finish_row(h);
4038                 ++s->mb_y;
4039                 if (FIELD_OR_MBAFF_PICTURE) {
4040                     ++s->mb_y;
4041                     if (FRAME_MBAFF && s->mb_y < s->mb_height)
4042                         predict_field_decoding_flag(h);
4043                 }
4044             }
4045
4046             if (eos || s->mb_y >= s->mb_height) {
4047                 tprintf(s->avctx, "slice end %d %d\n",
4048                         get_bits_count(&s->gb), s->gb.size_in_bits);
4049                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x - 1,
4050                                 s->mb_y, ER_MB_END & part_mask);
4051                 if (s->mb_x > lf_x_start)
4052                     loop_filter(h, lf_x_start, s->mb_x);
4053                 return 0;
4054             }
4055         }
4056     } else {
4057         for (;;) {
4058             int ret = ff_h264_decode_mb_cavlc(h);
4059
4060             if (ret >= 0)
4061                 ff_h264_hl_decode_mb(h);
4062
4063             // FIXME optimal? or let mb_decode decode 16x32 ?
4064             if (ret >= 0 && FRAME_MBAFF) {
4065                 s->mb_y++;
4066                 ret = ff_h264_decode_mb_cavlc(h);
4067
4068                 if (ret >= 0)
4069                     ff_h264_hl_decode_mb(h);
4070                 s->mb_y--;
4071             }
4072
4073             if (ret < 0) {
4074                 av_log(h->s.avctx, AV_LOG_ERROR,
4075                        "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
4076                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x,
4077                                 s->mb_y, ER_MB_ERROR & part_mask);
4078                 return -1;
4079             }
4080
4081             if (++s->mb_x >= s->mb_width) {
4082                 loop_filter(h, lf_x_start, s->mb_x);
4083                 s->mb_x = lf_x_start = 0;
4084                 decode_finish_row(h);
4085                 ++s->mb_y;
4086                 if (FIELD_OR_MBAFF_PICTURE) {
4087                     ++s->mb_y;
4088                     if (FRAME_MBAFF && s->mb_y < s->mb_height)
4089                         predict_field_decoding_flag(h);
4090                 }
4091                 if (s->mb_y >= s->mb_height) {
4092                     tprintf(s->avctx, "slice end %d %d\n",
4093                             get_bits_count(&s->gb), s->gb.size_in_bits);
4094
4095                     if (   get_bits_left(&s->gb) == 0
4096                         || get_bits_left(&s->gb) > 0 && !(s->avctx->err_recognition & AV_EF_AGGRESSIVE)) {
4097                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y,
4098                                         s->mb_x - 1, s->mb_y,
4099                                         ER_MB_END & part_mask);
4100
4101                         return 0;
4102                     } else {
4103                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y,
4104                                         s->mb_x, s->mb_y,
4105                                         ER_MB_END & part_mask);
4106
4107                         return -1;
4108                     }
4109                 }
4110             }
4111
4112             if (get_bits_left(&s->gb) <= 0 && s->mb_skip_run <= 0) {
4113                 tprintf(s->avctx, "slice end %d %d\n",
4114                         get_bits_count(&s->gb), s->gb.size_in_bits);
4115                 if (get_bits_left(&s->gb) == 0) {
4116                     ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y,
4117                                     s->mb_x - 1, s->mb_y,
4118                                     ER_MB_END & part_mask);
4119                     if (s->mb_x > lf_x_start)
4120                         loop_filter(h, lf_x_start, s->mb_x);
4121
4122                     return 0;
4123                 } else {
4124                     ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x,
4125                                     s->mb_y, ER_MB_ERROR & part_mask);
4126
4127                     return -1;
4128                 }
4129             }
4130         }
4131     }
4132 }
4133
4134 /**
4135  * Call decode_slice() for each context.
4136  *
4137  * @param h h264 master context
4138  * @param context_count number of contexts to execute
4139  */
4140 static int execute_decode_slices(H264Context *h, int context_count)
4141 {
4142     MpegEncContext *const s     = &h->s;
4143     AVCodecContext *const avctx = s->avctx;
4144     H264Context *hx;
4145     int i;
4146
4147     if (s->avctx->hwaccel ||
4148         s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
4149         return 0;
4150     if (context_count == 1) {
4151         return decode_slice(avctx, &h);
4152     } else {
4153         for (i = 1; i < context_count; i++) {
4154             hx                    = h->thread_context[i];
4155             hx->s.err_recognition = avctx->err_recognition;
4156             hx->s.error_count     = 0;
4157             hx->x264_build        = h->x264_build;
4158         }
4159
4160         avctx->execute(avctx, decode_slice, h->thread_context,
4161                        NULL, context_count, sizeof(void *));
4162
4163         /* pull back stuff from slices to master context */
4164         hx                   = h->thread_context[context_count - 1];
4165         s->mb_x              = hx->s.mb_x;
4166         s->mb_y              = hx->s.mb_y;
4167         s->dropable          = hx->s.dropable;
4168         s->picture_structure = hx->s.picture_structure;
4169         for (i = 1; i < context_count; i++)
4170             h->s.error_count += h->thread_context[i]->s.error_count;
4171     }
4172
4173     return 0;
4174 }
4175
4176 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size)
4177 {
4178     MpegEncContext *const s     = &h->s;
4179     AVCodecContext *const avctx = s->avctx;
4180     H264Context *hx; ///< thread context
4181     int buf_index;
4182     int context_count;
4183     int next_avc;
4184     int pass = !(avctx->active_thread_type & FF_THREAD_FRAME);
4185     int nals_needed = 0; ///< number of NALs that need decoding before the next frame thread starts
4186     int nal_index;
4187
4188     h->nal_unit_type= 0;
4189
4190     if(!s->slice_context_count)
4191          s->slice_context_count= 1;
4192     h->max_contexts = s->slice_context_count;
4193     if (!(s->flags2 & CODEC_FLAG2_CHUNKS)) {
4194         h->current_slice = 0;
4195         if (!s->first_field)
4196             s->current_picture_ptr = NULL;
4197         ff_h264_reset_sei(h);
4198     }
4199
4200     for (; pass <= 1; pass++) {
4201         buf_index     = 0;
4202         context_count = 0;
4203         next_avc      = h->is_avc ? 0 : buf_size;
4204         nal_index     = 0;
4205         for (;;) {
4206             int consumed;
4207             int dst_length;
4208             int bit_length;
4209             const uint8_t *ptr;
4210             int i, nalsize = 0;
4211             int err;
4212
4213             if (buf_index >= next_avc) {
4214                 if (buf_index >= buf_size - h->nal_length_size)
4215                     break;
4216                 nalsize = 0;
4217                 for (i = 0; i < h->nal_length_size; i++)
4218                     nalsize = (nalsize << 8) | buf[buf_index++];
4219                 if (nalsize <= 0 || nalsize > buf_size - buf_index) {
4220                     av_log(h->s.avctx, AV_LOG_ERROR,
4221                            "AVC: nal size %d\n", nalsize);
4222                     break;
4223                 }
4224                 next_avc = buf_index + nalsize;
4225             } else {
4226                 // start code prefix search
4227                 for (; buf_index + 3 < next_avc; buf_index++)
4228                     // This should always succeed in the first iteration.
4229                     if (buf[buf_index]     == 0 &&
4230                         buf[buf_index + 1] == 0 &&
4231                         buf[buf_index + 2] == 1)
4232                         break;
4233
4234                 if (buf_index + 3 >= buf_size)
4235                     break;
4236
4237                 buf_index += 3;
4238                 if (buf_index >= next_avc)
4239                     continue;
4240             }
4241
4242             hx = h->thread_context[context_count];
4243
4244             ptr = ff_h264_decode_nal(hx, buf + buf_index, &dst_length,
4245                                      &consumed, next_avc - buf_index);
4246             if (ptr == NULL || dst_length < 0) {
4247                 buf_index = -1;
4248                 goto end;
4249             }
4250             i = buf_index + consumed;
4251             if ((s->workaround_bugs & FF_BUG_AUTODETECT) && i + 3 < next_avc &&
4252                 buf[i]     == 0x00 && buf[i + 1] == 0x00 &&
4253                 buf[i + 2] == 0x01 && buf[i + 3] == 0xE0)
4254                 s->workaround_bugs |= FF_BUG_TRUNCATED;
4255
4256             if (!(s->workaround_bugs & FF_BUG_TRUNCATED))
4257                 while(dst_length > 0 && ptr[dst_length - 1] == 0)
4258                     dst_length--;
4259             bit_length = !dst_length ? 0
4260                                      : (8 * dst_length -
4261                                         ff_h264_decode_rbsp_trailing(h, ptr + dst_length - 1));
4262
4263             if (s->avctx->debug & FF_DEBUG_STARTCODE)
4264                 av_log(h->s.avctx, AV_LOG_DEBUG, "NAL %d/%d at %d/%d length %d pass %d\n", hx->nal_unit_type, hx->nal_ref_idc, buf_index, buf_size, dst_length, pass);
4265
4266             if (h->is_avc && (nalsize != consumed) && nalsize)
4267                 av_log(h->s.avctx, AV_LOG_DEBUG,
4268                        "AVC: Consumed only %d bytes instead of %d\n",
4269                        consumed, nalsize);
4270
4271             buf_index += consumed;
4272             nal_index++;
4273
4274             if (pass == 0) {
4275                 /* packets can sometimes contain multiple PPS/SPS,
4276                  * e.g. two PAFF field pictures in one packet, or a demuxer
4277                  * which splits NALs strangely if so, when frame threading we
4278                  * can't start the next thread until we've read all of them */
4279                 switch (hx->nal_unit_type) {
4280                 case NAL_SPS:
4281                 case NAL_PPS:
4282                     nals_needed = nal_index;
4283                     break;
4284                 case NAL_IDR_SLICE:
4285                 case NAL_SLICE:
4286                     init_get_bits(&hx->s.gb, ptr, bit_length);
4287                     if (!get_ue_golomb(&hx->s.gb))
4288                         nals_needed = nal_index;
4289                 }
4290                 continue;
4291             }
4292
4293             // FIXME do not discard SEI id
4294             if (avctx->skip_frame >= AVDISCARD_NONREF && h->nal_ref_idc == 0)
4295                 continue;
4296
4297 again:
4298             err = 0;
4299             switch (hx->nal_unit_type) {
4300             case NAL_IDR_SLICE:
4301                 if (h->nal_unit_type != NAL_IDR_SLICE) {
4302                     av_log(h->s.avctx, AV_LOG_ERROR,
4303                            "Invalid mix of idr and non-idr slices\n");
4304                     buf_index = -1;
4305                     goto end;
4306                 }
4307                 idr(h); // FIXME ensure we don't lose some frames if there is reordering
4308             case NAL_SLICE:
4309                 init_get_bits(&hx->s.gb, ptr, bit_length);
4310                 hx->intra_gb_ptr        =
4311                     hx->inter_gb_ptr    = &hx->s.gb;
4312                 hx->s.data_partitioning = 0;
4313
4314                 if ((err = decode_slice_header(hx, h)))
4315                     break;
4316
4317                 if (   h->sei_recovery_frame_cnt >= 0
4318                     && (   h->recovery_frame<0
4319                         || ((h->recovery_frame - h->frame_num) & ((1 << h->sps.log2_max_frame_num)-1)) > h->sei_recovery_frame_cnt)) {
4320                     h->recovery_frame = (h->frame_num + h->sei_recovery_frame_cnt) %
4321                                         (1 << h->sps.log2_max_frame_num);
4322                 }
4323
4324                 s->current_picture_ptr->f.key_frame |=
4325                         (hx->nal_unit_type == NAL_IDR_SLICE);
4326
4327                 if (h->recovery_frame == h->frame_num) {
4328                     s->current_picture_ptr->sync |= 1;
4329                     h->recovery_frame = -1;
4330                 }
4331
4332                 h->sync |= !!s->current_picture_ptr->f.key_frame;
4333                 h->sync |= 3*!!(s->flags2 & CODEC_FLAG2_SHOW_ALL);
4334                 s->current_picture_ptr->sync |= h->sync;
4335
4336                 if (h->current_slice == 1) {
4337                     if (!(s->flags2 & CODEC_FLAG2_CHUNKS))
4338                         decode_postinit(h, nal_index >= nals_needed);
4339
4340                     if (s->avctx->hwaccel &&
4341                         s->avctx->hwaccel->start_frame(s->avctx, NULL, 0) < 0)
4342                         return -1;
4343                     if (CONFIG_H264_VDPAU_DECODER &&
4344                         s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
4345                         ff_vdpau_h264_picture_start(s);
4346                 }
4347
4348                 if (hx->redundant_pic_count == 0 &&
4349                     (avctx->skip_frame < AVDISCARD_NONREF ||
4350                      hx->nal_ref_idc) &&
4351                     (avctx->skip_frame < AVDISCARD_BIDIR  ||
4352                      hx->slice_type_nos != AV_PICTURE_TYPE_B) &&
4353                     (avctx->skip_frame < AVDISCARD_NONKEY ||
4354                      hx->slice_type_nos == AV_PICTURE_TYPE_I) &&
4355                     avctx->skip_frame < AVDISCARD_ALL) {
4356                     if (avctx->hwaccel) {
4357                         if (avctx->hwaccel->decode_slice(avctx,
4358                                                          &buf[buf_index - consumed],
4359                                                          consumed) < 0)
4360                             return -1;
4361                     } else if (CONFIG_H264_VDPAU_DECODER &&
4362                                s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU) {
4363                         static const uint8_t start_code[] = {
4364                             0x00, 0x00, 0x01 };
4365                         ff_vdpau_add_data_chunk(s, start_code,
4366                                                 sizeof(start_code));
4367                         ff_vdpau_add_data_chunk(s, &buf[buf_index - consumed],
4368                                                 consumed);
4369                     } else
4370                         context_count++;
4371                 }
4372                 break;
4373             case NAL_DPA:
4374                 init_get_bits(&hx->s.gb, ptr, bit_length);
4375                 hx->intra_gb_ptr =
4376                 hx->inter_gb_ptr = NULL;
4377
4378                 if ((err = decode_slice_header(hx, h)) < 0)
4379                     break;
4380
4381                 hx->s.data_partitioning = 1;
4382                 break;
4383             case NAL_DPB:
4384                 init_get_bits(&hx->intra_gb, ptr, bit_length);
4385                 hx->intra_gb_ptr = &hx->intra_gb;
4386                 break;
4387             case NAL_DPC:
4388                 init_get_bits(&hx->inter_gb, ptr, bit_length);
4389                 hx->inter_gb_ptr = &hx->inter_gb;
4390
4391                 if (hx->redundant_pic_count == 0 &&
4392                     hx->intra_gb_ptr &&
4393                     hx->s.data_partitioning &&
4394                     s->context_initialized &&
4395                     (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc) &&
4396                     (avctx->skip_frame < AVDISCARD_BIDIR  ||
4397                      hx->slice_type_nos != AV_PICTURE_TYPE_B) &&
4398                     (avctx->skip_frame < AVDISCARD_NONKEY ||
4399                      hx->slice_type_nos == AV_PICTURE_TYPE_I) &&
4400                     avctx->skip_frame < AVDISCARD_ALL)
4401                     context_count++;
4402                 break;
4403             case NAL_SEI:
4404                 init_get_bits(&s->gb, ptr, bit_length);
4405                 ff_h264_decode_sei(h);
4406                 break;
4407             case NAL_SPS:
4408                 init_get_bits(&s->gb, ptr, bit_length);
4409                 if (ff_h264_decode_seq_parameter_set(h) < 0 && (h->is_avc ? (nalsize != consumed) && nalsize : 1)) {
4410                     av_log(h->s.avctx, AV_LOG_DEBUG,
4411                            "SPS decoding failure, trying alternative mode\n");
4412                     if (h->is_avc)
4413                         av_assert0(next_avc - buf_index + consumed == nalsize);
4414                     init_get_bits(&s->gb, &buf[buf_index + 1 - consumed],
4415                                   8*(next_avc - buf_index + consumed - 1));
4416                     ff_h264_decode_seq_parameter_set(h);
4417                 }
4418
4419                 if (s->flags & CODEC_FLAG_LOW_DELAY ||
4420                     (h->sps.bitstream_restriction_flag &&
4421                      !h->sps.num_reorder_frames))
4422                     s->low_delay = 1;
4423                 if (avctx->has_b_frames < 2)
4424                     avctx->has_b_frames = !s->low_delay;
4425                 break;
4426             case NAL_PPS:
4427                 init_get_bits(&s->gb, ptr, bit_length);
4428                 ff_h264_decode_picture_parameter_set(h, bit_length);
4429                 break;
4430             case NAL_AUD:
4431             case NAL_END_SEQUENCE:
4432             case NAL_END_STREAM:
4433             case NAL_FILLER_DATA:
4434             case NAL_SPS_EXT:
4435             case NAL_AUXILIARY_SLICE:
4436                 break;
4437             default:
4438                 av_log(avctx, AV_LOG_DEBUG, "Unknown NAL code: %d (%d bits)\n",
4439                        hx->nal_unit_type, bit_length);
4440             }
4441
4442             if (context_count == h->max_contexts) {
4443                 execute_decode_slices(h, context_count);
4444                 context_count = 0;
4445             }
4446
4447             if (err < 0)
4448                 av_log(h->s.avctx, AV_LOG_ERROR, "decode_slice_header error\n");
4449             else if (err == 1) {
4450                 /* Slice could not be decoded in parallel mode, copy down
4451                  * NAL unit stuff to context 0 and restart. Note that
4452                  * rbsp_buffer is not transferred, but since we no longer
4453                  * run in parallel mode this should not be an issue. */
4454                 h->nal_unit_type = hx->nal_unit_type;
4455                 h->nal_ref_idc   = hx->nal_ref_idc;
4456                 hx               = h;
4457                 goto again;
4458             }
4459         }
4460     }
4461     if (context_count)
4462         execute_decode_slices(h, context_count);
4463
4464 end:
4465     /* clean up */
4466     if (s->current_picture_ptr && s->current_picture_ptr->owner2 == s &&
4467         !s->dropable) {
4468         ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX,
4469                                   s->picture_structure == PICT_BOTTOM_FIELD);
4470     }
4471
4472     return buf_index;
4473 }
4474
4475 /**
4476  * Return the number of bytes consumed for building the current frame.
4477  */
4478 static int get_consumed_bytes(MpegEncContext *s, int pos, int buf_size)
4479 {
4480     if (pos == 0)
4481         pos = 1;          // avoid infinite loops (i doubt that is needed but ...)
4482     if (pos + 10 > buf_size)
4483         pos = buf_size;                   // oops ;)
4484
4485     return pos;
4486 }
4487
4488 static int decode_frame(AVCodecContext *avctx, void *data,
4489                         int *data_size, AVPacket *avpkt)
4490 {
4491     const uint8_t *buf = avpkt->data;
4492     int buf_size       = avpkt->size;
4493     H264Context *h     = avctx->priv_data;
4494     MpegEncContext *s  = &h->s;
4495     AVFrame *pict      = data;
4496     int buf_index      = 0;
4497     Picture *out;
4498     int i, out_idx;
4499
4500     s->flags  = avctx->flags;
4501     s->flags2 = avctx->flags2;
4502
4503     /* end of stream, output what is still in the buffers */
4504     if (buf_size == 0) {
4505  out:
4506
4507         s->current_picture_ptr = NULL;
4508
4509         // FIXME factorize this with the output code below
4510         out     = h->delayed_pic[0];
4511         out_idx = 0;
4512         for (i = 1;
4513              h->delayed_pic[i] &&
4514              !h->delayed_pic[i]->f.key_frame &&
4515              !h->delayed_pic[i]->mmco_reset;
4516              i++)
4517             if (h->delayed_pic[i]->poc < out->poc) {
4518                 out     = h->delayed_pic[i];
4519                 out_idx = i;
4520             }
4521
4522         for (i = out_idx; h->delayed_pic[i]; i++)
4523             h->delayed_pic[i] = h->delayed_pic[i + 1];
4524
4525         if (out) {
4526             *data_size = sizeof(AVFrame);
4527             *pict      = out->f;
4528         }
4529
4530         return buf_index;
4531     }
4532     if(h->is_avc && buf_size >= 9 && buf[0]==1 && buf[2]==0 && (buf[4]&0xFC)==0xFC && (buf[5]&0x1F) && buf[8]==0x67){
4533         int cnt= buf[5]&0x1f;
4534         const uint8_t *p= buf+6;
4535         while(cnt--){
4536             int nalsize= AV_RB16(p) + 2;
4537             if(nalsize > buf_size - (p-buf) || p[2]!=0x67)
4538                 goto not_extra;
4539             p += nalsize;
4540         }
4541         cnt = *(p++);
4542         if(!cnt)
4543             goto not_extra;
4544         while(cnt--){
4545             int nalsize= AV_RB16(p) + 2;
4546             if(nalsize > buf_size - (p-buf) || p[2]!=0x68)
4547                 goto not_extra;
4548             p += nalsize;
4549         }
4550
4551         return ff_h264_decode_extradata(h, buf, buf_size);
4552     }
4553 not_extra:
4554
4555     buf_index = decode_nal_units(h, buf, buf_size);
4556     if (buf_index < 0)
4557         return -1;
4558
4559     if (!s->current_picture_ptr && h->nal_unit_type == NAL_END_SEQUENCE) {
4560         av_assert0(buf_index <= buf_size);
4561         goto out;
4562     }
4563
4564     if (!(s->flags2 & CODEC_FLAG2_CHUNKS) && !s->current_picture_ptr) {
4565         if (avctx->skip_frame >= AVDISCARD_NONREF ||
4566             buf_size >= 4 && !memcmp("Q264", buf, 4))
4567             return buf_size;
4568         av_log(avctx, AV_LOG_ERROR, "no frame!\n");
4569         return -1;
4570     }
4571
4572     if (!(s->flags2 & CODEC_FLAG2_CHUNKS) ||
4573         (s->mb_y >= s->mb_height && s->mb_height)) {
4574         if (s->flags2 & CODEC_FLAG2_CHUNKS)
4575             decode_postinit(h, 1);
4576
4577         field_end(h, 0);
4578
4579         /* Wait for second field. */
4580         *data_size = 0;
4581         if (h->next_output_pic && (h->next_output_pic->sync || h->sync>1)) {
4582             *data_size = sizeof(AVFrame);
4583             *pict      = h->next_output_pic->f;
4584         }
4585     }
4586
4587     assert(pict->data[0] || !*data_size);
4588     ff_print_debug_info(s, pict);
4589     // printf("out %d\n", (int)pict->data[0]);
4590
4591     return get_consumed_bytes(s, buf_index, buf_size);
4592 }
4593
4594 av_cold void ff_h264_free_context(H264Context *h)
4595 {
4596     int i;
4597
4598     free_tables(h, 1); // FIXME cleanup init stuff perhaps
4599
4600     for (i = 0; i < MAX_SPS_COUNT; i++)
4601         av_freep(h->sps_buffers + i);
4602
4603     for (i = 0; i < MAX_PPS_COUNT; i++)
4604         av_freep(h->pps_buffers + i);
4605 }
4606
4607 av_cold int ff_h264_decode_end(AVCodecContext *avctx)
4608 {
4609     H264Context *h    = avctx->priv_data;
4610     MpegEncContext *s = &h->s;
4611
4612     ff_h264_remove_all_refs(h);
4613     ff_h264_free_context(h);
4614
4615     ff_MPV_common_end(s);
4616
4617     // memset(h, 0, sizeof(H264Context));
4618
4619     return 0;
4620 }
4621
4622 static const AVProfile profiles[] = {
4623     { FF_PROFILE_H264_BASELINE,             "Baseline"              },
4624     { FF_PROFILE_H264_CONSTRAINED_BASELINE, "Constrained Baseline"  },
4625     { FF_PROFILE_H264_MAIN,                 "Main"                  },
4626     { FF_PROFILE_H264_EXTENDED,             "Extended"              },
4627     { FF_PROFILE_H264_HIGH,                 "High"                  },
4628     { FF_PROFILE_H264_HIGH_10,              "High 10"               },
4629     { FF_PROFILE_H264_HIGH_10_INTRA,        "High 10 Intra"         },
4630     { FF_PROFILE_H264_HIGH_422,             "High 4:2:2"            },
4631     { FF_PROFILE_H264_HIGH_422_INTRA,       "High 4:2:2 Intra"      },
4632     { FF_PROFILE_H264_HIGH_444,             "High 4:4:4"            },
4633     { FF_PROFILE_H264_HIGH_444_PREDICTIVE,  "High 4:4:4 Predictive" },
4634     { FF_PROFILE_H264_HIGH_444_INTRA,       "High 4:4:4 Intra"      },
4635     { FF_PROFILE_H264_CAVLC_444,            "CAVLC 4:4:4"           },
4636     { FF_PROFILE_UNKNOWN },
4637 };
4638
4639 static const AVOption h264_options[] = {
4640     {"is_avc", "is avc", offsetof(H264Context, is_avc), FF_OPT_TYPE_INT, {.dbl = 0}, 0, 1, 0},
4641     {"nal_length_size", "nal_length_size", offsetof(H264Context, nal_length_size), FF_OPT_TYPE_INT, {.dbl = 0}, 0, 4, 0},
4642     {NULL}
4643 };
4644
4645 static const AVClass h264_class = {
4646     "H264 Decoder",
4647     av_default_item_name,
4648     h264_options,
4649     LIBAVUTIL_VERSION_INT,
4650 };
4651
4652 static const AVClass h264_vdpau_class = {
4653     "H264 VDPAU Decoder",
4654     av_default_item_name,
4655     h264_options,
4656     LIBAVUTIL_VERSION_INT,
4657 };
4658
4659 AVCodec ff_h264_decoder = {
4660     .name                  = "h264",
4661     .type                  = AVMEDIA_TYPE_VIDEO,
4662     .id                    = CODEC_ID_H264,
4663     .priv_data_size        = sizeof(H264Context),
4664     .init                  = ff_h264_decode_init,
4665     .close                 = ff_h264_decode_end,
4666     .decode                = decode_frame,
4667     .capabilities          = /*CODEC_CAP_DRAW_HORIZ_BAND |*/ CODEC_CAP_DR1 |
4668                              CODEC_CAP_DELAY | CODEC_CAP_SLICE_THREADS |
4669                              CODEC_CAP_FRAME_THREADS,
4670     .flush                 = flush_dpb,
4671     .long_name             = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10"),
4672     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(decode_init_thread_copy),
4673     .update_thread_context = ONLY_IF_THREADS_ENABLED(decode_update_thread_context),
4674     .profiles              = NULL_IF_CONFIG_SMALL(profiles),
4675     .priv_class            = &h264_class,
4676 };
4677
4678 #if CONFIG_H264_VDPAU_DECODER
4679 AVCodec ff_h264_vdpau_decoder = {
4680     .name           = "h264_vdpau",
4681     .type           = AVMEDIA_TYPE_VIDEO,
4682     .id             = CODEC_ID_H264,
4683     .priv_data_size = sizeof(H264Context),
4684     .init           = ff_h264_decode_init,
4685     .close          = ff_h264_decode_end,
4686     .decode         = decode_frame,
4687     .capabilities   = CODEC_CAP_DR1 | CODEC_CAP_DELAY | CODEC_CAP_HWACCEL_VDPAU,
4688     .flush          = flush_dpb,
4689     .long_name      = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10 (VDPAU acceleration)"),
4690     .pix_fmts       = (const enum PixelFormat[]) { PIX_FMT_VDPAU_H264,
4691                                                    PIX_FMT_NONE},
4692     .profiles       = NULL_IF_CONFIG_SMALL(profiles),
4693     .priv_class     = &h264_vdpau_class,
4694 };
4695 #endif