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 const uint16_t ff_h264_mb_sizes[4] = { 256, 384, 512, 768 };
  52
  53 static const uint8_t rem6[QP_MAX_NUM + 1] = {
  54     0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2,
  55     3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5,
  56     0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
  57 };
  58
  59 static const uint8_t div6[QP_MAX_NUM + 1] = {
  60     0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3,  3,  3,
  61     3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6,  6,  6,
  62     7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 10, 10, 10, 10,
  63 };
  64
  65 static const enum PixelFormat hwaccel_pixfmt_list_h264_jpeg_420[] = {
  66     PIX_FMT_DXVA2_VLD,
  67     PIX_FMT_VAAPI_VLD,
  68     PIX_FMT_VDA_VLD,
  69     PIX_FMT_YUVJ420P,
  70     PIX_FMT_NONE
  71 };
  72
  73 /**
  74  * Check if the top & left blocks are available if needed and
  75  * change the dc mode so it only uses the available blocks.
  76  */
  77 int ff_h264_check_intra4x4_pred_mode(H264Context *h)
  78 {
  79     MpegEncContext *const s     = &h->s;
  80     static const int8_t top[12] = {
  81         -1, 0, LEFT_DC_PRED, -1, -1, -1, -1, -1, 0
  82     };
  83     static const int8_t left[12] = {
  84         0, -1, TOP_DC_PRED, 0, -1, -1, -1, 0, -1, DC_128_PRED
  85     };
  86     int i;
  87
  88     if (!(h->top_samples_available & 0x8000)) {
  89         for (i = 0; i < 4; i++) {
  90             int status = top[h->intra4x4_pred_mode_cache[scan8[0] + i]];
  91             if (status < 0) {
  92                 av_log(h->s.avctx, AV_LOG_ERROR,
  93                        "top block unavailable for requested intra4x4 mode %d at %d %d\n",
  94                        status, s->mb_x, s->mb_y);
  95                 return -1;
  96             } else if (status) {
  97                 h->intra4x4_pred_mode_cache[scan8[0] + i] = status;
  98             }
  99         }
 100     }
 101
 102     if ((h->left_samples_available & 0x8888) != 0x8888) {
 103         static const int mask[4] = { 0x8000, 0x2000, 0x80, 0x20 };
 104         for (i = 0; i < 4; i++)
 105             if (!(h->left_samples_available & mask[i])) {
 106                 int status = left[h->intra4x4_pred_mode_cache[scan8[0] + 8 * i]];
 107                 if (status < 0) {
 108                     av_log(h->s.avctx, AV_LOG_ERROR,
 109                            "left block unavailable for requested intra4x4 mode %d at %d %d\n",
 110                            status, s->mb_x, s->mb_y);
 111                     return -1;
 112                 } else if (status) {
 113                     h->intra4x4_pred_mode_cache[scan8[0] + 8 * i] = status;
 114                 }
 115             }
 116     }
 117
 118     return 0;
 119 } // FIXME cleanup like ff_h264_check_intra_pred_mode
 120
 121 /**
 122  * Check if the top & left blocks are available if needed and
 123  * change the dc mode so it only uses the available blocks.
 124  */
 125 int ff_h264_check_intra_pred_mode(H264Context *h, int mode, int is_chroma)
 126 {
 127     MpegEncContext *const s     = &h->s;
 128     static const int8_t top[7]  = { LEFT_DC_PRED8x8, 1, -1, -1 };
 129     static const int8_t left[7] = { TOP_DC_PRED8x8, -1, 2, -1, DC_128_PRED8x8 };
 130
 131     if (mode > 6U) {
 132         av_log(h->s.avctx, AV_LOG_ERROR,
 133                "out of range intra chroma pred mode at %d %d\n",
 134                s->mb_x, s->mb_y);
 135         return -1;
 136     }
 137
 138     if (!(h->top_samples_available & 0x8000)) {
 139         mode = top[mode];
 140         if (mode < 0) {
 141             av_log(h->s.avctx, AV_LOG_ERROR,
 142                    "top block unavailable for requested intra mode at %d %d\n",
 143                    s->mb_x, s->mb_y);
 144             return -1;
 145         }
 146     }
 147
 148     if ((h->left_samples_available & 0x8080) != 0x8080) {
 149         mode = left[mode];
 150         if (is_chroma && (h->left_samples_available & 0x8080)) {
 151             // mad cow disease mode, aka MBAFF + constrained_intra_pred
 152             mode = ALZHEIMER_DC_L0T_PRED8x8 +
 153                    (!(h->left_samples_available & 0x8000)) +
 154                    2 * (mode == DC_128_PRED8x8);
 155         }
 156         if (mode < 0) {
 157             av_log(h->s.avctx, AV_LOG_ERROR,
 158                    "left block unavailable for requested intra mode at %d %d\n",
 159                    s->mb_x, s->mb_y);
 160             return -1;
 161         }
 162     }
 163
 164     return mode;
 165 }
 166
 167 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src,
 168                                   int *dst_length, int *consumed, int length)
 169 {
 170     int i, si, di;
 171     uint8_t *dst;
 172     int bufidx;
 173
 174     // src[0]&0x80; // forbidden bit
 175     h->nal_ref_idc   = src[0] >> 5;
 176     h->nal_unit_type = src[0] & 0x1F;
 177
 178     src++;
 179     length--;
 180
 181 #if HAVE_FAST_UNALIGNED
 182 #if HAVE_FAST_64BIT
 183 #define RS 7
 184     for (i = 0; i + 1 < length; i += 9) {
 185         if (!((~AV_RN64A(src + i) &
 186                (AV_RN64A(src + i) - 0x0100010001000101ULL)) &
 187               0x8000800080008080ULL))
 188 #else
 189 #define RS 3
 190     for (i = 0; i + 1 < length; i += 5) {
 191         if (!((~AV_RN32A(src + i) &
 192                (AV_RN32A(src + i) - 0x01000101U)) &
 193               0x80008080U))
 194 #endif
 195             continue;
 196         if (i > 0 && !src[i])
 197             i--;
 198         while (src[i])
 199             i++;
 200 #else
 201 #define RS 0
 202     for (i = 0; i + 1 < length; i += 2) {
 203         if (src[i])
 204             continue;
 205         if (i > 0 && src[i - 1] == 0)
 206             i--;
 207 #endif
 208         if (i + 2 < length && src[i + 1] == 0 && src[i + 2] <= 3) {
 209             if (src[i + 2] != 3) {
 210                 /* startcode, so we must be past the end */
 211                 length = i;
 212             }
 213             break;
 214         }
 215         i -= RS;
 216     }
 217
 218     // use second escape buffer for inter data
 219     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0;
 220
 221     si = h->rbsp_buffer_size[bufidx];
 222     av_fast_padded_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+MAX_MBPAIR_SIZE);
 223     dst = h->rbsp_buffer[bufidx];
 224
 225     if (dst == NULL)
 226         return NULL;
 227
 228     if(i>=length-1){ //no escaped 0
 229         *dst_length= length;
 230         *consumed= length+1; //+1 for the header
 231         if(h->s.avctx->flags2 & CODEC_FLAG2_FAST){
 232             return src;
 233         }else{
 234             memcpy(dst, src, length);
 235             return dst;
 236         }
 237     }
 238
 239     // printf("decoding esc\n");
 240     memcpy(dst, src, i);
 241     si = di = i;
 242     while (si + 2 < length) {
 243         // remove escapes (very rare 1:2^22)
 244         if (src[si + 2] > 3) {
 245             dst[di++] = src[si++];
 246             dst[di++] = src[si++];
 247         } else if (src[si] == 0 && src[si + 1] == 0) {
 248             if (src[si + 2] == 3) { // escape
 249                 dst[di++]  = 0;
 250                 dst[di++]  = 0;
 251                 si        += 3;
 252                 continue;
 253             } else // next start code
 254                 goto nsc;
 255         }
 256
 257         dst[di++] = src[si++];
 258     }
 259     while (si < length)
 260         dst[di++] = src[si++];
 261 nsc:
 262
 263     memset(dst + di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
 264
 265     *dst_length = di;
 266     *consumed   = si + 1; // +1 for the header
 267     /* FIXME store exact number of bits in the getbitcontext
 268      * (it is needed for decoding) */
 269     return dst;
 270 }
 271
 272 /**
 273  * Identify the exact end of the bitstream
 274  * @return the length of the trailing, or 0 if damaged
 275  */
 276 static int decode_rbsp_trailing(H264Context *h, const uint8_t *src)
 277 {
 278     int v = *src;
 279     int r;
 280
 281     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
 282
 283     for (r = 1; r < 9; r++) {
 284         if (v & 1)
 285             return r;
 286         v >>= 1;
 287     }
 288     return 0;
 289 }
 290
 291 static inline int get_lowest_part_list_y(H264Context *h, Picture *pic, int n,
 292                                          int height, int y_offset, int list)
 293 {
 294     int raw_my        = h->mv_cache[list][scan8[n]][1];
 295     int filter_height = (raw_my & 3) ? 2 : 0;
 296     int full_my       = (raw_my >> 2) + y_offset;
 297     int top           = full_my - filter_height;
 298     int bottom        = full_my + filter_height + height;
 299
 300     return FFMAX(abs(top), bottom);
 301 }
 302
 303 static inline void get_lowest_part_y(H264Context *h, int refs[2][48], int n,
 304                                      int height, int y_offset, int list0,
 305                                      int list1, int *nrefs)
 306 {
 307     MpegEncContext *const s = &h->s;
 308     int my;
 309
 310     y_offset += 16 * (s->mb_y >> MB_FIELD);
 311
 312     if (list0) {
 313         int ref_n    = h->ref_cache[0][scan8[n]];
 314         Picture *ref = &h->ref_list[0][ref_n];
 315
 316         // Error resilience puts the current picture in the ref list.
 317         // Don't try to wait on these as it will cause a deadlock.
 318         // Fields can wait on each other, though.
 319         if (ref->f.thread_opaque   != s->current_picture.f.thread_opaque ||
 320             (ref->f.reference & 3) != s->picture_structure) {
 321             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 0);
 322             if (refs[0][ref_n] < 0)
 323                 nrefs[0] += 1;
 324             refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
 325         }
 326     }
 327
 328     if (list1) {
 329         int ref_n    = h->ref_cache[1][scan8[n]];
 330         Picture *ref = &h->ref_list[1][ref_n];
 331
 332         if (ref->f.thread_opaque   != s->current_picture.f.thread_opaque ||
 333             (ref->f.reference & 3) != s->picture_structure) {
 334             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 1);
 335             if (refs[1][ref_n] < 0)
 336                 nrefs[1] += 1;
 337             refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
 338         }
 339     }
 340 }
 341
 342 /**
 343  * Wait until all reference frames are available for MC operations.
 344  *
 345  * @param h the H264 context
 346  */
 347 static void await_references(H264Context *h)
 348 {
 349     MpegEncContext *const s = &h->s;
 350     const int mb_xy   = h->mb_xy;
 351     const int mb_type = s->current_picture.f.mb_type[mb_xy];
 352     int refs[2][48];
 353     int nrefs[2] = { 0 };
 354     int ref, list;
 355
 356     memset(refs, -1, sizeof(refs));
 357
 358     if (IS_16X16(mb_type)) {
 359         get_lowest_part_y(h, refs, 0, 16, 0,
 360                           IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 361     } else if (IS_16X8(mb_type)) {
 362         get_lowest_part_y(h, refs, 0, 8, 0,
 363                           IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 364         get_lowest_part_y(h, refs, 8, 8, 8,
 365                           IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
 366     } else if (IS_8X16(mb_type)) {
 367         get_lowest_part_y(h, refs, 0, 16, 0,
 368                           IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 369         get_lowest_part_y(h, refs, 4, 16, 0,
 370                           IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
 371     } else {
 372         int i;
 373
 374         assert(IS_8X8(mb_type));
 375
 376         for (i = 0; i < 4; i++) {
 377             const int sub_mb_type = h->sub_mb_type[i];
 378             const int n           = 4 * i;
 379             int y_offset          = (i & 2) << 2;
 380
 381             if (IS_SUB_8X8(sub_mb_type)) {
 382                 get_lowest_part_y(h, refs, n, 8, y_offset,
 383                                   IS_DIR(sub_mb_type, 0, 0),
 384                                   IS_DIR(sub_mb_type, 0, 1),
 385                                   nrefs);
 386             } else if (IS_SUB_8X4(sub_mb_type)) {
 387                 get_lowest_part_y(h, refs, n, 4, y_offset,
 388                                   IS_DIR(sub_mb_type, 0, 0),
 389                                   IS_DIR(sub_mb_type, 0, 1),
 390                                   nrefs);
 391                 get_lowest_part_y(h, refs, n + 2, 4, y_offset + 4,
 392                                   IS_DIR(sub_mb_type, 0, 0),
 393                                   IS_DIR(sub_mb_type, 0, 1),
 394                                   nrefs);
 395             } else if (IS_SUB_4X8(sub_mb_type)) {
 396                 get_lowest_part_y(h, refs, n, 8, y_offset,
 397                                   IS_DIR(sub_mb_type, 0, 0),
 398                                   IS_DIR(sub_mb_type, 0, 1),
 399                                   nrefs);
 400                 get_lowest_part_y(h, refs, n + 1, 8, y_offset,
 401                                   IS_DIR(sub_mb_type, 0, 0),
 402                                   IS_DIR(sub_mb_type, 0, 1),
 403                                   nrefs);
 404             } else {
 405                 int j;
 406                 assert(IS_SUB_4X4(sub_mb_type));
 407                 for (j = 0; j < 4; j++) {
 408                     int sub_y_offset = y_offset + 2 * (j & 2);
 409                     get_lowest_part_y(h, refs, n + j, 4, sub_y_offset,
 410                                       IS_DIR(sub_mb_type, 0, 0),
 411                                       IS_DIR(sub_mb_type, 0, 1),
 412                                       nrefs);
 413                 }
 414             }
 415         }
 416     }
 417
 418     for (list = h->list_count - 1; list >= 0; list--)
 419         for (ref = 0; ref < 48 && nrefs[list]; ref++) {
 420             int row = refs[list][ref];
 421             if (row >= 0) {
 422                 Picture *ref_pic      = &h->ref_list[list][ref];
 423                 int ref_field         = ref_pic->f.reference - 1;
 424                 int ref_field_picture = ref_pic->field_picture;
 425                 int pic_height        = 16 * s->mb_height >> ref_field_picture;
 426
 427                 row <<= MB_MBAFF;
 428                 nrefs[list]--;
 429
 430                 if (!FIELD_PICTURE && ref_field_picture) { // frame referencing two fields
 431                     ff_thread_await_progress(&ref_pic->f,
 432                                              FFMIN((row >> 1) - !(row & 1),
 433                                                    pic_height - 1),
 434                                              1);
 435                     ff_thread_await_progress(&ref_pic->f,
 436                                              FFMIN((row >> 1), pic_height - 1),
 437                                              0);
 438                 } else if (FIELD_PICTURE && !ref_field_picture) { // field referencing one field of a frame
 439                     ff_thread_await_progress(&ref_pic->f,
 440                                              FFMIN(row * 2 + ref_field,
 441                                                    pic_height - 1),
 442                                              0);
 443                 } else if (FIELD_PICTURE) {
 444                     ff_thread_await_progress(&ref_pic->f,
 445                                              FFMIN(row, pic_height - 1),
 446                                              ref_field);
 447                 } else {
 448                     ff_thread_await_progress(&ref_pic->f,
 449                                              FFMIN(row, pic_height - 1),
 450                                              0);
 451                 }
 452             }
 453         }
 454 }
 455
 456 static av_always_inline void mc_dir_part(H264Context *h, Picture *pic,
 457                                          int n, int square, int height,
 458                                          int delta, int list,
 459                                          uint8_t *dest_y, uint8_t *dest_cb,
 460                                          uint8_t *dest_cr,
 461                                          int src_x_offset, int src_y_offset,
 462                                          qpel_mc_func *qpix_op,
 463                                          h264_chroma_mc_func chroma_op,
 464                                          int pixel_shift, int chroma_idc)
 465 {
 466     MpegEncContext *const s = &h->s;
 467     const int mx      = h->mv_cache[list][scan8[n]][0] + src_x_offset * 8;
 468     int my            = h->mv_cache[list][scan8[n]][1] + src_y_offset * 8;
 469     const int luma_xy = (mx & 3) + ((my & 3) << 2);
 470     int offset        = ((mx >> 2) << pixel_shift) + (my >> 2) * h->mb_linesize;
 471     uint8_t *src_y    = pic->f.data[0] + offset;
 472     uint8_t *src_cb, *src_cr;
 473     int extra_width  = h->emu_edge_width;
 474     int extra_height = h->emu_edge_height;
 475     int emu = 0;
 476     const int full_mx    = mx >> 2;
 477     const int full_my    = my >> 2;
 478     const int pic_width  = 16 * s->mb_width;
 479     const int pic_height = 16 * s->mb_height >> MB_FIELD;
 480     int ysh;
 481
 482     if (mx & 7)
 483         extra_width -= 3;
 484     if (my & 7)
 485         extra_height -= 3;
 486
 487     if (full_mx                <          0 - extra_width  ||
 488         full_my                <          0 - extra_height ||
 489         full_mx + 16 /*FIXME*/ > pic_width  + extra_width  ||
 490         full_my + 16 /*FIXME*/ > pic_height + extra_height) {
 491         s->dsp.emulated_edge_mc(s->edge_emu_buffer,
 492                                 src_y - (2 << pixel_shift) - 2 * h->mb_linesize,
 493                                 h->mb_linesize,
 494                                 16 + 5, 16 + 5 /*FIXME*/, full_mx - 2,
 495                                 full_my - 2, pic_width, pic_height);
 496         src_y = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
 497         emu   = 1;
 498     }
 499
 500     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); // FIXME try variable height perhaps?
 501     if (!square)
 502         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
 503
 504     if (CONFIG_GRAY && s->flags & CODEC_FLAG_GRAY)
 505         return;
 506
 507     if (chroma_idc == 3 /* yuv444 */) {
 508         src_cb = pic->f.data[1] + offset;
 509         if (emu) {
 510             s->dsp.emulated_edge_mc(s->edge_emu_buffer,
 511                                     src_cb - (2 << pixel_shift) - 2 * h->mb_linesize,
 512                                     h->mb_linesize,
 513                                     16 + 5, 16 + 5 /*FIXME*/,
 514                                     full_mx - 2, full_my - 2,
 515                                     pic_width, pic_height);
 516             src_cb = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
 517         }
 518         qpix_op[luma_xy](dest_cb, src_cb, h->mb_linesize); // FIXME try variable height perhaps?
 519         if (!square)
 520             qpix_op[luma_xy](dest_cb + delta, src_cb + delta, h->mb_linesize);
 521
 522         src_cr = pic->f.data[2] + offset;
 523         if (emu) {
 524             s->dsp.emulated_edge_mc(s->edge_emu_buffer,
 525                                     src_cr - (2 << pixel_shift) - 2 * h->mb_linesize,
 526                                     h->mb_linesize,
 527                                     16 + 5, 16 + 5 /*FIXME*/,
 528                                     full_mx - 2, full_my - 2,
 529                                     pic_width, pic_height);
 530             src_cr = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
 531         }
 532         qpix_op[luma_xy](dest_cr, src_cr, h->mb_linesize); // FIXME try variable height perhaps?
 533         if (!square)
 534             qpix_op[luma_xy](dest_cr + delta, src_cr + delta, h->mb_linesize);
 535         return;
 536     }
 537
 538     ysh = 3 - (chroma_idc == 2 /* yuv422 */);
 539     if (chroma_idc == 1 /* yuv420 */ && MB_FIELD) {
 540         // chroma offset when predicting from a field of opposite parity
 541         my  += 2 * ((s->mb_y & 1) - (pic->f.reference - 1));
 542         emu |= (my >> 3) < 0 || (my >> 3) + 8 >= (pic_height >> 1);
 543     }
 544
 545     src_cb = pic->f.data[1] + ((mx >> 3) << pixel_shift) +
 546              (my >> ysh) * h->mb_uvlinesize;
 547     src_cr = pic->f.data[2] + ((mx >> 3) << pixel_shift) +
 548              (my >> ysh) * h->mb_uvlinesize;
 549
 550     if (emu) {
 551         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize,
 552                                 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
 553                                 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
 554         src_cb = s->edge_emu_buffer;
 555     }
 556     chroma_op(dest_cb, src_cb, h->mb_uvlinesize,
 557               height >> (chroma_idc == 1 /* yuv420 */),
 558               mx & 7, (my << (chroma_idc == 2 /* yuv422 */)) & 7);
 559
 560     if (emu) {
 561         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize,
 562                                 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
 563                                 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
 564         src_cr = s->edge_emu_buffer;
 565     }
 566     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, height >> (chroma_idc == 1 /* yuv420 */),
 567               mx & 7, (my << (chroma_idc == 2 /* yuv422 */)) & 7);
 568 }
 569
 570 static av_always_inline void mc_part_std(H264Context *h, int n, int square,
 571                                          int height, int delta,
 572                                          uint8_t *dest_y, uint8_t *dest_cb,
 573                                          uint8_t *dest_cr,
 574                                          int x_offset, int y_offset,
 575                                          qpel_mc_func *qpix_put,
 576                                          h264_chroma_mc_func chroma_put,
 577                                          qpel_mc_func *qpix_avg,
 578                                          h264_chroma_mc_func chroma_avg,
 579                                          int list0, int list1,
 580                                          int pixel_shift, int chroma_idc)
 581 {
 582     MpegEncContext *const s       = &h->s;
 583     qpel_mc_func *qpix_op         = qpix_put;
 584     h264_chroma_mc_func chroma_op = chroma_put;
 585
 586     dest_y += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
 587     if (chroma_idc == 3 /* yuv444 */) {
 588         dest_cb += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
 589         dest_cr += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
 590     } else if (chroma_idc == 2 /* yuv422 */) {
 591         dest_cb += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
 592         dest_cr += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
 593     } else { /* yuv420 */
 594         dest_cb += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
 595         dest_cr += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
 596     }
 597     x_offset += 8 * s->mb_x;
 598     y_offset += 8 * (s->mb_y >> MB_FIELD);
 599
 600     if (list0) {
 601         Picture *ref = &h->ref_list[0][h->ref_cache[0][scan8[n]]];
 602         mc_dir_part(h, ref, n, square, height, delta, 0,
 603                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
 604                     qpix_op, chroma_op, pixel_shift, chroma_idc);
 605
 606         qpix_op   = qpix_avg;
 607         chroma_op = chroma_avg;
 608     }
 609
 610     if (list1) {
 611         Picture *ref = &h->ref_list[1][h->ref_cache[1][scan8[n]]];
 612         mc_dir_part(h, ref, n, square, height, delta, 1,
 613                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
 614                     qpix_op, chroma_op, pixel_shift, chroma_idc);
 615     }
 616 }
 617
 618 static av_always_inline void mc_part_weighted(H264Context *h, int n, int square,
 619                                               int height, int delta,
 620                                               uint8_t *dest_y, uint8_t *dest_cb,
 621                                               uint8_t *dest_cr,
 622                                               int x_offset, int y_offset,
 623                                               qpel_mc_func *qpix_put,
 624                                               h264_chroma_mc_func chroma_put,
 625                                               h264_weight_func luma_weight_op,
 626                                               h264_weight_func chroma_weight_op,
 627                                               h264_biweight_func luma_weight_avg,
 628                                               h264_biweight_func chroma_weight_avg,
 629                                               int list0, int list1,
 630                                               int pixel_shift, int chroma_idc)
 631 {
 632     MpegEncContext *const s = &h->s;
 633     int chroma_height;
 634
 635     dest_y += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
 636     if (chroma_idc == 3 /* yuv444 */) {
 637         chroma_height     = height;
 638         chroma_weight_avg = luma_weight_avg;
 639         chroma_weight_op  = luma_weight_op;
 640         dest_cb += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
 641         dest_cr += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
 642     } else if (chroma_idc == 2 /* yuv422 */) {
 643         chroma_height = height;
 644         dest_cb      += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
 645         dest_cr      += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
 646     } else { /* yuv420 */
 647         chroma_height = height >> 1;
 648         dest_cb      += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
 649         dest_cr      += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
 650     }
 651     x_offset += 8 * s->mb_x;
 652     y_offset += 8 * (s->mb_y >> MB_FIELD);
 653
 654     if (list0 && list1) {
 655         /* don't optimize for luma-only case, since B-frames usually
 656          * use implicit weights => chroma too. */
 657         uint8_t *tmp_cb = s->obmc_scratchpad;
 658         uint8_t *tmp_cr = s->obmc_scratchpad + (16 << pixel_shift);
 659         uint8_t *tmp_y  = s->obmc_scratchpad + 16 * h->mb_uvlinesize;
 660         int refn0       = h->ref_cache[0][scan8[n]];
 661         int refn1       = h->ref_cache[1][scan8[n]];
 662
 663         mc_dir_part(h, &h->ref_list[0][refn0], n, square, height, delta, 0,
 664                     dest_y, dest_cb, dest_cr,
 665                     x_offset, y_offset, qpix_put, chroma_put,
 666                     pixel_shift, chroma_idc);
 667         mc_dir_part(h, &h->ref_list[1][refn1], n, square, height, delta, 1,
 668                     tmp_y, tmp_cb, tmp_cr,
 669                     x_offset, y_offset, qpix_put, chroma_put,
 670                     pixel_shift, chroma_idc);
 671
 672         if (h->use_weight == 2) {
 673             int weight0 = h->implicit_weight[refn0][refn1][s->mb_y & 1];
 674             int weight1 = 64 - weight0;
 675             luma_weight_avg(dest_y, tmp_y, h->mb_linesize,
 676                             height, 5, weight0, weight1, 0);
 677             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize,
 678                               chroma_height, 5, weight0, weight1, 0);
 679             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize,
 680                               chroma_height, 5, weight0, weight1, 0);
 681         } else {
 682             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, height,
 683                             h->luma_log2_weight_denom,
 684                             h->luma_weight[refn0][0][0],
 685                             h->luma_weight[refn1][1][0],
 686                             h->luma_weight[refn0][0][1] +
 687                             h->luma_weight[refn1][1][1]);
 688             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, chroma_height,
 689                               h->chroma_log2_weight_denom,
 690                               h->chroma_weight[refn0][0][0][0],
 691                               h->chroma_weight[refn1][1][0][0],
 692                               h->chroma_weight[refn0][0][0][1] +
 693                               h->chroma_weight[refn1][1][0][1]);
 694             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, chroma_height,
 695                               h->chroma_log2_weight_denom,
 696                               h->chroma_weight[refn0][0][1][0],
 697                               h->chroma_weight[refn1][1][1][0],
 698                               h->chroma_weight[refn0][0][1][1] +
 699                               h->chroma_weight[refn1][1][1][1]);
 700         }
 701     } else {
 702         int list     = list1 ? 1 : 0;
 703         int refn     = h->ref_cache[list][scan8[n]];
 704         Picture *ref = &h->ref_list[list][refn];
 705         mc_dir_part(h, ref, n, square, height, delta, list,
 706                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
 707                     qpix_put, chroma_put, pixel_shift, chroma_idc);
 708
 709         luma_weight_op(dest_y, h->mb_linesize, height,
 710                        h->luma_log2_weight_denom,
 711                        h->luma_weight[refn][list][0],
 712                        h->luma_weight[refn][list][1]);
 713         if (h->use_weight_chroma) {
 714             chroma_weight_op(dest_cb, h->mb_uvlinesize, chroma_height,
 715                              h->chroma_log2_weight_denom,
 716                              h->chroma_weight[refn][list][0][0],
 717                              h->chroma_weight[refn][list][0][1]);
 718             chroma_weight_op(dest_cr, h->mb_uvlinesize, chroma_height,
 719                              h->chroma_log2_weight_denom,
 720                              h->chroma_weight[refn][list][1][0],
 721                              h->chroma_weight[refn][list][1][1]);
 722         }
 723     }
 724 }
 725
 726 static av_always_inline void mc_part(H264Context *h, int n, int square,
 727                                      int height, int delta,
 728                                      uint8_t *dest_y, uint8_t *dest_cb,
 729                                      uint8_t *dest_cr,
 730                                      int x_offset, int y_offset,
 731                                      qpel_mc_func *qpix_put,
 732                                      h264_chroma_mc_func chroma_put,
 733                                      qpel_mc_func *qpix_avg,
 734                                      h264_chroma_mc_func chroma_avg,
 735                                      h264_weight_func *weight_op,
 736                                      h264_biweight_func *weight_avg,
 737                                      int list0, int list1,
 738                                      int pixel_shift, int chroma_idc)
 739 {
 740     if ((h->use_weight == 2 && list0 && list1 &&
 741          (h->implicit_weight[h->ref_cache[0][scan8[n]]][h->ref_cache[1][scan8[n]]][h->s.mb_y & 1] != 32)) ||
 742         h->use_weight == 1)
 743         mc_part_weighted(h, n, square, height, delta, dest_y, dest_cb, dest_cr,
 744                          x_offset, y_offset, qpix_put, chroma_put,
 745                          weight_op[0], weight_op[1], weight_avg[0],
 746                          weight_avg[1], list0, list1, pixel_shift, chroma_idc);
 747     else
 748         mc_part_std(h, n, square, height, delta, dest_y, dest_cb, dest_cr,
 749                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg,
 750                     chroma_avg, list0, list1, pixel_shift, chroma_idc);
 751 }
 752
 753 static av_always_inline void prefetch_motion(H264Context *h, int list,
 754                                              int pixel_shift, int chroma_idc)
 755 {
 756     /* fetch pixels for estimated mv 4 macroblocks ahead
 757      * optimized for 64byte cache lines */
 758     MpegEncContext *const s = &h->s;
 759     const int refn = h->ref_cache[list][scan8[0]];
 760     if (refn >= 0) {
 761         const int mx  = (h->mv_cache[list][scan8[0]][0] >> 2) + 16 * s->mb_x + 8;
 762         const int my  = (h->mv_cache[list][scan8[0]][1] >> 2) + 16 * s->mb_y;
 763         uint8_t **src = h->ref_list[list][refn].f.data;
 764         int off       = (mx << pixel_shift) +
 765                         (my + (s->mb_x & 3) * 4) * h->mb_linesize +
 766                         (64 << pixel_shift);
 767         s->dsp.prefetch(src[0] + off, s->linesize, 4);
 768         if (chroma_idc == 3 /* yuv444 */) {
 769             s->dsp.prefetch(src[1] + off, s->linesize, 4);
 770             s->dsp.prefetch(src[2] + off, s->linesize, 4);
 771         } else {
 772             off= (((mx>>1)+64)<<pixel_shift) + ((my>>1) + (s->mb_x&7))*s->uvlinesize;
 773             s->dsp.prefetch(src[1] + off, src[2] - src[1], 2);
 774         }
 775     }
 776 }
 777
 778 static av_always_inline void hl_motion(H264Context *h, uint8_t *dest_y,
 779                                        uint8_t *dest_cb, uint8_t *dest_cr,
 780                                        qpel_mc_func(*qpix_put)[16],
 781                                        h264_chroma_mc_func(*chroma_put),
 782                                        qpel_mc_func(*qpix_avg)[16],
 783                                        h264_chroma_mc_func(*chroma_avg),
 784                                        h264_weight_func *weight_op,
 785                                        h264_biweight_func *weight_avg,
 786                                        int pixel_shift, int chroma_idc)
 787 {
 788     MpegEncContext *const s = &h->s;
 789     const int mb_xy   = h->mb_xy;
 790     const int mb_type = s->current_picture.f.mb_type[mb_xy];
 791
 792     assert(IS_INTER(mb_type));
 793
 794     if (HAVE_THREADS && (s->avctx->active_thread_type & FF_THREAD_FRAME))
 795         await_references(h);
 796     prefetch_motion(h, 0, pixel_shift, chroma_idc);
 797
 798     if (IS_16X16(mb_type)) {
 799         mc_part(h, 0, 1, 16, 0, dest_y, dest_cb, dest_cr, 0, 0,
 800                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
 801                 weight_op, weight_avg,
 802                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 803                 pixel_shift, chroma_idc);
 804     } else if (IS_16X8(mb_type)) {
 805         mc_part(h, 0, 0, 8, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 0,
 806                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
 807                 weight_op, weight_avg,
 808                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 809                 pixel_shift, chroma_idc);
 810         mc_part(h, 8, 0, 8, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 4,
 811                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
 812                 weight_op, weight_avg,
 813                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
 814                 pixel_shift, chroma_idc);
 815     } else if (IS_8X16(mb_type)) {
 816         mc_part(h, 0, 0, 16, 8 * h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
 817                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 818                 &weight_op[1], &weight_avg[1],
 819                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 820                 pixel_shift, chroma_idc);
 821         mc_part(h, 4, 0, 16, 8 * h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
 822                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 823                 &weight_op[1], &weight_avg[1],
 824                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
 825                 pixel_shift, chroma_idc);
 826     } else {
 827         int i;
 828
 829         assert(IS_8X8(mb_type));
 830
 831         for (i = 0; i < 4; i++) {
 832             const int sub_mb_type = h->sub_mb_type[i];
 833             const int n  = 4 * i;
 834             int x_offset = (i & 1) << 2;
 835             int y_offset = (i & 2) << 1;
 836
 837             if (IS_SUB_8X8(sub_mb_type)) {
 838                 mc_part(h, n, 1, 8, 0, dest_y, dest_cb, dest_cr,
 839                         x_offset, y_offset,
 840                         qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 841                         &weight_op[1], &weight_avg[1],
 842                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 843                         pixel_shift, chroma_idc);
 844             } else if (IS_SUB_8X4(sub_mb_type)) {
 845                 mc_part(h, n, 0, 4, 4 << pixel_shift, dest_y, dest_cb, dest_cr,
 846                         x_offset, y_offset,
 847                         qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
 848                         &weight_op[1], &weight_avg[1],
 849                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 850                         pixel_shift, chroma_idc);
 851                 mc_part(h, n + 2, 0, 4, 4 << pixel_shift,
 852                         dest_y, dest_cb, dest_cr, x_offset, y_offset + 2,
 853                         qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
 854                         &weight_op[1], &weight_avg[1],
 855                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 856                         pixel_shift, chroma_idc);
 857             } else if (IS_SUB_4X8(sub_mb_type)) {
 858                 mc_part(h, n, 0, 8, 4 * h->mb_linesize,
 859                         dest_y, dest_cb, dest_cr, x_offset, y_offset,
 860                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 861                         &weight_op[2], &weight_avg[2],
 862                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 863                         pixel_shift, chroma_idc);
 864                 mc_part(h, n + 1, 0, 8, 4 * h->mb_linesize,
 865                         dest_y, dest_cb, dest_cr, x_offset + 2, y_offset,
 866                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 867                         &weight_op[2], &weight_avg[2],
 868                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 869                         pixel_shift, chroma_idc);
 870             } else {
 871                 int j;
 872                 assert(IS_SUB_4X4(sub_mb_type));
 873                 for (j = 0; j < 4; j++) {
 874                     int sub_x_offset = x_offset + 2 * (j & 1);
 875                     int sub_y_offset = y_offset + (j & 2);
 876                     mc_part(h, n + j, 1, 4, 0,
 877                             dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
 878                             qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 879                             &weight_op[2], &weight_avg[2],
 880                             IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 881                             pixel_shift, chroma_idc);
 882                 }
 883             }
 884         }
 885     }
 886
 887     prefetch_motion(h, 1, pixel_shift, chroma_idc);
 888 }
 889
 890 static av_always_inline void hl_motion_420(H264Context *h, uint8_t *dest_y,
 891                                            uint8_t *dest_cb, uint8_t *dest_cr,
 892                                            qpel_mc_func(*qpix_put)[16],
 893                                            h264_chroma_mc_func(*chroma_put),
 894                                            qpel_mc_func(*qpix_avg)[16],
 895                                            h264_chroma_mc_func(*chroma_avg),
 896                                            h264_weight_func *weight_op,
 897                                            h264_biweight_func *weight_avg,
 898                                            int pixel_shift)
 899 {
 900     hl_motion(h, dest_y, dest_cb, dest_cr, qpix_put, chroma_put,
 901               qpix_avg, chroma_avg, weight_op, weight_avg, pixel_shift, 1);
 902 }
 903
 904 static av_always_inline void hl_motion_422(H264Context *h, uint8_t *dest_y,
 905                                            uint8_t *dest_cb, uint8_t *dest_cr,
 906                                            qpel_mc_func(*qpix_put)[16],
 907                                            h264_chroma_mc_func(*chroma_put),
 908                                            qpel_mc_func(*qpix_avg)[16],
 909                                            h264_chroma_mc_func(*chroma_avg),
 910                                            h264_weight_func *weight_op,
 911                                            h264_biweight_func *weight_avg,
 912                                            int pixel_shift)
 913 {
 914     hl_motion(h, dest_y, dest_cb, dest_cr, qpix_put, chroma_put,
 915               qpix_avg, chroma_avg, weight_op, weight_avg, pixel_shift, 2);
 916 }
 917
 918 static void free_tables(H264Context *h, int free_rbsp)
 919 {
 920     int i;
 921     H264Context *hx;
 922
 923     av_freep(&h->intra4x4_pred_mode);
 924     av_freep(&h->chroma_pred_mode_table);
 925     av_freep(&h->cbp_table);
 926     av_freep(&h->mvd_table[0]);
 927     av_freep(&h->mvd_table[1]);
 928     av_freep(&h->direct_table);
 929     av_freep(&h->non_zero_count);
 930     av_freep(&h->slice_table_base);
 931     h->slice_table = NULL;
 932     av_freep(&h->list_counts);
 933
 934     av_freep(&h->mb2b_xy);
 935     av_freep(&h->mb2br_xy);
 936
 937     for (i = 0; i < MAX_THREADS; i++) {
 938         hx = h->thread_context[i];
 939         if (!hx)
 940             continue;
 941         av_freep(&hx->top_borders[1]);
 942         av_freep(&hx->top_borders[0]);
 943         av_freep(&hx->s.obmc_scratchpad);
 944         if (free_rbsp) {
 945             av_freep(&hx->rbsp_buffer[1]);
 946             av_freep(&hx->rbsp_buffer[0]);
 947             hx->rbsp_buffer_size[0] = 0;
 948             hx->rbsp_buffer_size[1] = 0;
 949         }
 950         if (i)
 951             av_freep(&h->thread_context[i]);
 952     }
 953 }
 954
 955 static void init_dequant8_coeff_table(H264Context *h)
 956 {
 957     int i, j, q, x;
 958     const int max_qp = 51 + 6 * (h->sps.bit_depth_luma - 8);
 959
 960     for (i = 0; i < 6; i++) {
 961         h->dequant8_coeff[i] = h->dequant8_buffer[i];
 962         for (j = 0; j < i; j++)
 963             if (!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i],
 964                         64 * sizeof(uint8_t))) {
 965                 h->dequant8_coeff[i] = h->dequant8_buffer[j];
 966                 break;
 967             }
 968         if (j < i)
 969             continue;
 970
 971         for (q = 0; q < max_qp + 1; q++) {
 972             int shift = div6[q];
 973             int idx   = rem6[q];
 974             for (x = 0; x < 64; x++)
 975                 h->dequant8_coeff[i][q][(x >> 3) | ((x & 7) << 3)] =
 976                     ((uint32_t)dequant8_coeff_init[idx][dequant8_coeff_init_scan[((x >> 1) & 12) | (x & 3)]] *
 977                      h->pps.scaling_matrix8[i][x]) << shift;
 978         }
 979     }
 980 }
 981
 982 static void init_dequant4_coeff_table(H264Context *h)
 983 {
 984     int i, j, q, x;
 985     const int max_qp = 51 + 6 * (h->sps.bit_depth_luma - 8);
 986     for (i = 0; i < 6; i++) {
 987         h->dequant4_coeff[i] = h->dequant4_buffer[i];
 988         for (j = 0; j < i; j++)
 989             if (!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i],
 990                         16 * sizeof(uint8_t))) {
 991                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
 992                 break;
 993             }
 994         if (j < i)
 995             continue;
 996
 997         for (q = 0; q < max_qp + 1; q++) {
 998             int shift = div6[q] + 2;
 999             int idx   = rem6[q];
1000             for (x = 0; x < 16; x++)
1001                 h->dequant4_coeff[i][q][(x >> 2) | ((x << 2) & 0xF)] =
1002                     ((uint32_t)dequant4_coeff_init[idx][(x & 1) + ((x >> 2) & 1)] *
1003                      h->pps.scaling_matrix4[i][x]) << shift;
1004         }
1005     }
1006 }
1007
1008 static void init_dequant_tables(H264Context *h)
1009 {
1010     int i, x;
1011     init_dequant4_coeff_table(h);
1012     if (h->pps.transform_8x8_mode)
1013         init_dequant8_coeff_table(h);
1014     if (h->sps.transform_bypass) {
1015         for (i = 0; i < 6; i++)
1016             for (x = 0; x < 16; x++)
1017                 h->dequant4_coeff[i][0][x] = 1 << 6;
1018         if (h->pps.transform_8x8_mode)
1019             for (i = 0; i < 6; i++)
1020                 for (x = 0; x < 64; x++)
1021                     h->dequant8_coeff[i][0][x] = 1 << 6;
1022     }
1023 }
1024
1025 int ff_h264_alloc_tables(H264Context *h)
1026 {
1027     MpegEncContext *const s = &h->s;
1028     const int big_mb_num    = s->mb_stride * (s->mb_height + 1);
1029     const int row_mb_num    = 2*s->mb_stride*FFMAX(s->avctx->thread_count, 1);
1030     int x, y;
1031
1032     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode,
1033                       row_mb_num * 8 * sizeof(uint8_t), fail)
1034     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count,
1035                       big_mb_num * 48 * sizeof(uint8_t), fail)
1036     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base,
1037                       (big_mb_num + s->mb_stride) * sizeof(*h->slice_table_base), fail)
1038     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table,
1039                       big_mb_num * sizeof(uint16_t), fail)
1040     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table,
1041                       big_mb_num * sizeof(uint8_t), fail)
1042     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0],
1043                       16 * row_mb_num * sizeof(uint8_t), fail);
1044     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1],
1045                       16 * row_mb_num * sizeof(uint8_t), fail);
1046     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table,
1047                       4 * big_mb_num * sizeof(uint8_t), fail);
1048     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts,
1049                       big_mb_num * sizeof(uint8_t), fail)
1050
1051     memset(h->slice_table_base, -1,
1052            (big_mb_num + s->mb_stride) * sizeof(*h->slice_table_base));
1053     h->slice_table = h->slice_table_base + s->mb_stride * 2 + 1;
1054
1055     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy,
1056                       big_mb_num * sizeof(uint32_t), fail);
1057     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy,
1058                       big_mb_num * sizeof(uint32_t), fail);
1059     for (y = 0; y < s->mb_height; y++)
1060         for (x = 0; x < s->mb_width; x++) {
1061             const int mb_xy = x + y * s->mb_stride;
1062             const int b_xy  = 4 * x + 4 * y * h->b_stride;
1063
1064             h->mb2b_xy[mb_xy]  = b_xy;
1065             h->mb2br_xy[mb_xy] = 8 * (FMO ? mb_xy : (mb_xy % (2 * s->mb_stride)));
1066         }
1067
1068     s->obmc_scratchpad = NULL;
1069
1070     if (!h->dequant4_coeff[0])
1071         init_dequant_tables(h);
1072
1073     return 0;
1074
1075 fail:
1076     free_tables(h, 1);
1077     return -1;
1078 }
1079
1080 /**
1081  * Mimic alloc_tables(), but for every context thread.
1082  */
1083 static void clone_tables(H264Context *dst, H264Context *src, int i)
1084 {
1085     MpegEncContext *const s     = &src->s;
1086     dst->intra4x4_pred_mode     = src->intra4x4_pred_mode + i * 8 * 2 * s->mb_stride;
1087     dst->non_zero_count         = src->non_zero_count;
1088     dst->slice_table            = src->slice_table;
1089     dst->cbp_table              = src->cbp_table;
1090     dst->mb2b_xy                = src->mb2b_xy;
1091     dst->mb2br_xy               = src->mb2br_xy;
1092     dst->chroma_pred_mode_table = src->chroma_pred_mode_table;
1093     dst->mvd_table[0]           = src->mvd_table[0] + i * 8 * 2 * s->mb_stride;
1094     dst->mvd_table[1]           = src->mvd_table[1] + i * 8 * 2 * s->mb_stride;
1095     dst->direct_table           = src->direct_table;
1096     dst->list_counts            = src->list_counts;
1097     dst->s.obmc_scratchpad      = NULL;
1098     ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma,
1099                       src->sps.chroma_format_idc);
1100 }
1101
1102 /**
1103  * Init context
1104  * Allocate buffers which are not shared amongst multiple threads.
1105  */
1106 static int context_init(H264Context *h)
1107 {
1108     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0],
1109                       h->s.mb_width * 16 * 3 * sizeof(uint8_t) * 2, fail)
1110     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1],
1111                       h->s.mb_width * 16 * 3 * sizeof(uint8_t) * 2, fail)
1112
1113     h->ref_cache[0][scan8[5]  + 1] =
1114     h->ref_cache[0][scan8[7]  + 1] =
1115     h->ref_cache[0][scan8[13] + 1] =
1116     h->ref_cache[1][scan8[5]  + 1] =
1117     h->ref_cache[1][scan8[7]  + 1] =
1118     h->ref_cache[1][scan8[13] + 1] = PART_NOT_AVAILABLE;
1119
1120     return 0;
1121
1122 fail:
1123     return -1; // free_tables will clean up for us
1124 }
1125
1126 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
1127
1128 static av_cold void common_init(H264Context *h)
1129 {
1130     MpegEncContext *const s = &h->s;
1131
1132     s->width    = s->avctx->width;
1133     s->height   = s->avctx->height;
1134     s->codec_id = s->avctx->codec->id;
1135
1136     s->avctx->bits_per_raw_sample = 8;
1137     h->cur_chroma_format_idc = 1;
1138
1139     ff_h264dsp_init(&h->h264dsp,
1140                     s->avctx->bits_per_raw_sample, h->cur_chroma_format_idc);
1141     ff_h264_pred_init(&h->hpc, s->codec_id,
1142                       s->avctx->bits_per_raw_sample, h->cur_chroma_format_idc);
1143
1144     h->dequant_coeff_pps = -1;
1145     s->unrestricted_mv   = 1;
1146
1147     s->dsp.dct_bits = 16;
1148     /* needed so that IDCT permutation is known early */
1149     ff_dsputil_init(&s->dsp, s->avctx);
1150
1151     memset(h->pps.scaling_matrix4, 16, 6 * 16 * sizeof(uint8_t));
1152     memset(h->pps.scaling_matrix8, 16, 2 * 64 * sizeof(uint8_t));
1153 }
1154
1155 int ff_h264_decode_extradata(H264Context *h, const uint8_t *buf, int size)
1156 {
1157     AVCodecContext *avctx = h->s.avctx;
1158
1159     if (!buf || size <= 0)
1160         return -1;
1161
1162     if (buf[0] == 1) {
1163         int i, cnt, nalsize;
1164         const unsigned char *p = buf;
1165
1166         h->is_avc = 1;
1167
1168         if (size < 7) {
1169             av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
1170             return -1;
1171         }
1172         /* sps and pps in the avcC always have length coded with 2 bytes,
1173          * so put a fake nal_length_size = 2 while parsing them */
1174         h->nal_length_size = 2;
1175         // Decode sps from avcC
1176         cnt = *(p + 5) & 0x1f; // Number of sps
1177         p  += 6;
1178         for (i = 0; i < cnt; i++) {
1179             nalsize = AV_RB16(p) + 2;
1180             if(nalsize > size - (p-buf))
1181                 return -1;
1182             if (decode_nal_units(h, p, nalsize) < 0) {
1183                 av_log(avctx, AV_LOG_ERROR,
1184                        "Decoding sps %d from avcC failed\n", i);
1185                 return -1;
1186             }
1187             p += nalsize;
1188         }
1189         // Decode pps from avcC
1190         cnt = *(p++); // Number of pps
1191         for (i = 0; i < cnt; i++) {
1192             nalsize = AV_RB16(p) + 2;
1193             if(nalsize > size - (p-buf))
1194                 return -1;
1195             if (decode_nal_units(h, p, nalsize) < 0) {
1196                 av_log(avctx, AV_LOG_ERROR,
1197                        "Decoding pps %d from avcC failed\n", i);
1198                 return -1;
1199             }
1200             p += nalsize;
1201         }
1202         // Now store right nal length size, that will be used to parse all other nals
1203         h->nal_length_size = (buf[4] & 0x03) + 1;
1204     } else {
1205         h->is_avc = 0;
1206         if (decode_nal_units(h, buf, size) < 0)
1207             return -1;
1208     }
1209     return size;
1210 }
1211
1212 av_cold int ff_h264_decode_init(AVCodecContext *avctx)
1213 {
1214     H264Context *h = avctx->priv_data;
1215     MpegEncContext *const s = &h->s;
1216     int i;
1217
1218     ff_MPV_decode_defaults(s);
1219
1220     s->avctx = avctx;
1221     common_init(h);
1222
1223     s->out_format      = FMT_H264;
1224     s->workaround_bugs = avctx->workaround_bugs;
1225
1226     /* set defaults */
1227     // s->decode_mb = ff_h263_decode_mb;
1228     s->quarter_sample = 1;
1229     if (!avctx->has_b_frames)
1230         s->low_delay = 1;
1231
1232     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1233
1234     ff_h264_decode_init_vlc();
1235
1236     h->pixel_shift = 0;
1237     h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1238
1239     h->thread_context[0] = h;
1240     h->outputed_poc      = h->next_outputed_poc = INT_MIN;
1241     for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
1242         h->last_pocs[i] = INT_MIN;
1243     h->prev_poc_msb = 1 << 16;
1244     h->prev_frame_num = -1;
1245     h->x264_build   = -1;
1246     ff_h264_reset_sei(h);
1247     if (avctx->codec_id == CODEC_ID_H264) {
1248         if (avctx->ticks_per_frame == 1)
1249             s->avctx->time_base.den *= 2;
1250         avctx->ticks_per_frame = 2;
1251     }
1252
1253     if (avctx->extradata_size > 0 && avctx->extradata &&
1254         ff_h264_decode_extradata(h, avctx->extradata, avctx->extradata_size) < 0) {
1255         ff_h264_free_context(h);
1256         return -1;
1257     }
1258
1259     if (h->sps.bitstream_restriction_flag &&
1260         s->avctx->has_b_frames < h->sps.num_reorder_frames) {
1261         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1262         s->low_delay           = 0;
1263     }
1264
1265     return 0;
1266 }
1267
1268 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b) + (size))))
1269
1270 static void copy_picture_range(Picture **to, Picture **from, int count,
1271                                MpegEncContext *new_base,
1272                                MpegEncContext *old_base)
1273 {
1274     int i;
1275
1276     for (i = 0; i < count; i++) {
1277         assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1278                 IN_RANGE(from[i], old_base->picture,
1279                          sizeof(Picture) * old_base->picture_count) ||
1280                 !from[i]));
1281         to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1282     }
1283 }
1284
1285 static void copy_parameter_set(void **to, void **from, int count, int size)
1286 {
1287     int i;
1288
1289     for (i = 0; i < count; i++) {
1290         if (to[i] && !from[i])
1291             av_freep(&to[i]);
1292         else if (from[i] && !to[i])
1293             to[i] = av_malloc(size);
1294
1295         if (from[i])
1296             memcpy(to[i], from[i], size);
1297     }
1298 }
1299
1300 static int decode_init_thread_copy(AVCodecContext *avctx)
1301 {
1302     H264Context *h = avctx->priv_data;
1303
1304     if (!avctx->internal->is_copy)
1305         return 0;
1306     memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1307     memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1308
1309     return 0;
1310 }
1311
1312 #define copy_fields(to, from, start_field, end_field)                   \
1313     memcpy(&to->start_field, &from->start_field,                        \
1314            (char *)&to->end_field - (char *)&to->start_field)
1315
1316 static int decode_update_thread_context(AVCodecContext *dst,
1317                                         const AVCodecContext *src)
1318 {
1319     H264Context *h = dst->priv_data, *h1 = src->priv_data;
1320     MpegEncContext *const s = &h->s, *const s1 = &h1->s;
1321     int inited = s->context_initialized, err;
1322     int i;
1323
1324     if (dst == src)
1325         return 0;
1326
1327     err = ff_mpeg_update_thread_context(dst, src);
1328     if (err)
1329         return err;
1330
1331     // FIXME handle width/height changing
1332     if (!inited) {
1333         for (i = 0; i < MAX_SPS_COUNT; i++)
1334             av_freep(h->sps_buffers + i);
1335
1336         for (i = 0; i < MAX_PPS_COUNT; i++)
1337             av_freep(h->pps_buffers + i);
1338
1339         // copy all fields after MpegEnc
1340         memcpy(&h->s + 1, &h1->s + 1,
1341                sizeof(H264Context) - sizeof(MpegEncContext));
1342         memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1343         memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1344
1345         if (s1->context_initialized) {
1346         if (ff_h264_alloc_tables(h) < 0) {
1347             av_log(dst, AV_LOG_ERROR, "Could not allocate memory for h264\n");
1348             return AVERROR(ENOMEM);
1349         }
1350         context_init(h);
1351
1352         /* frame_start may not be called for the next thread (if it's decoding
1353          * a bottom field) so this has to be allocated here */
1354         h->s.obmc_scratchpad = av_malloc(16 * 6 * s->linesize);
1355         }
1356
1357         for (i = 0; i < 2; i++) {
1358             h->rbsp_buffer[i]      = NULL;
1359             h->rbsp_buffer_size[i] = 0;
1360         }
1361
1362         h->thread_context[0] = h;
1363
1364         s->dsp.clear_blocks(h->mb);
1365         s->dsp.clear_blocks(h->mb + (24 * 16 << h->pixel_shift));
1366     }
1367
1368     // extradata/NAL handling
1369     h->is_avc = h1->is_avc;
1370
1371     // SPS/PPS
1372     copy_parameter_set((void **)h->sps_buffers, (void **)h1->sps_buffers,
1373                        MAX_SPS_COUNT, sizeof(SPS));
1374     h->sps = h1->sps;
1375     copy_parameter_set((void **)h->pps_buffers, (void **)h1->pps_buffers,
1376                        MAX_PPS_COUNT, sizeof(PPS));
1377     h->pps = h1->pps;
1378
1379     // Dequantization matrices
1380     // FIXME these are big - can they be only copied when PPS changes?
1381     copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1382
1383     for (i = 0; i < 6; i++)
1384         h->dequant4_coeff[i] = h->dequant4_buffer[0] +
1385                                (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1386
1387     for (i = 0; i < 6; i++)
1388         h->dequant8_coeff[i] = h->dequant8_buffer[0] +
1389                                (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1390
1391     h->dequant_coeff_pps = h1->dequant_coeff_pps;
1392
1393     // POC timing
1394     copy_fields(h, h1, poc_lsb, redundant_pic_count);
1395
1396     // reference lists
1397     copy_fields(h, h1, ref_count, list_count);
1398     copy_fields(h, h1, ref_list, intra_gb);
1399     copy_fields(h, h1, short_ref, cabac_init_idc);
1400
1401     copy_picture_range(h->short_ref, h1->short_ref, 32, s, s1);
1402     copy_picture_range(h->long_ref, h1->long_ref, 32, s, s1);
1403     copy_picture_range(h->delayed_pic, h1->delayed_pic,
1404                        MAX_DELAYED_PIC_COUNT + 2, s, s1);
1405
1406     h->last_slice_type = h1->last_slice_type;
1407     h->sync            = h1->sync;
1408
1409     if (!s->current_picture_ptr)
1410         return 0;
1411
1412     if (!s->dropable) {
1413         err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1414         h->prev_poc_msb = h->poc_msb;
1415         h->prev_poc_lsb = h->poc_lsb;
1416     }
1417     h->prev_frame_num_offset = h->frame_num_offset;
1418     h->prev_frame_num        = h->frame_num;
1419     h->outputed_poc          = h->next_outputed_poc;
1420
1421     return err;
1422 }
1423
1424 int ff_h264_frame_start(H264Context *h)
1425 {
1426     MpegEncContext *const s = &h->s;
1427     int i;
1428     const int pixel_shift = h->pixel_shift;
1429
1430     if (ff_MPV_frame_start(s, s->avctx) < 0)
1431         return -1;
1432     ff_er_frame_start(s);
1433     /*
1434      * ff_MPV_frame_start uses pict_type to derive key_frame.
1435      * This is incorrect for H.264; IDR markings must be used.
1436      * Zero here; IDR markings per slice in frame or fields are ORed in later.
1437      * See decode_nal_units().
1438      */
1439     s->current_picture_ptr->f.key_frame = 0;
1440     s->current_picture_ptr->sync        = 0;
1441     s->current_picture_ptr->mmco_reset  = 0;
1442
1443     assert(s->linesize && s->uvlinesize);
1444
1445     for (i = 0; i < 16; i++) {
1446         h->block_offset[i]           = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * s->linesize * ((scan8[i] - scan8[0]) >> 3);
1447         h->block_offset[48 + i]      = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * s->linesize * ((scan8[i] - scan8[0]) >> 3);
1448     }
1449     for (i = 0; i < 16; i++) {
1450         h->block_offset[16 + i]      =
1451         h->block_offset[32 + i]      = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * s->uvlinesize * ((scan8[i] - scan8[0]) >> 3);
1452         h->block_offset[48 + 16 + i] =
1453         h->block_offset[48 + 32 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * s->uvlinesize * ((scan8[i] - scan8[0]) >> 3);
1454     }
1455
1456     /* can't be in alloc_tables because linesize isn't known there.
1457      * FIXME: redo bipred weight to not require extra buffer? */
1458     for (i = 0; i < s->slice_context_count; i++)
1459         if (h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1460             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16 * 6 * s->linesize);
1461
1462     /* Some macroblocks can be accessed before they're available in case
1463      * of lost slices, MBAFF or threading. */
1464     memset(h->slice_table, -1,
1465            (s->mb_height * s->mb_stride - 1) * sizeof(*h->slice_table));
1466
1467     // s->decode = (s->flags & CODEC_FLAG_PSNR) || !s->encoding ||
1468     //             s->current_picture.f.reference /* || h->contains_intra */ || 1;
1469
1470     /* We mark the current picture as non-reference after allocating it, so
1471      * that if we break out due to an error it can be released automatically
1472      * in the next ff_MPV_frame_start().
1473      * SVQ3 as well as most other codecs have only last/next/current and thus
1474      * get released even with set reference, besides SVQ3 and others do not
1475      * mark frames as reference later "naturally". */
1476     if (s->codec_id != CODEC_ID_SVQ3)
1477         s->current_picture_ptr->f.reference = 0;
1478
1479     s->current_picture_ptr->field_poc[0]     =
1480         s->current_picture_ptr->field_poc[1] = INT_MAX;
1481
1482     h->next_output_pic = NULL;
1483
1484     assert(s->current_picture_ptr->long_ref == 0);
1485
1486     return 0;
1487 }
1488
1489 /**
1490  * Run setup operations that must be run after slice header decoding.
1491  * This includes finding the next displayed frame.
1492  *
1493  * @param h h264 master context
1494  * @param setup_finished enough NALs have been read that we can call
1495  * ff_thread_finish_setup()
1496  */
1497 static void decode_postinit(H264Context *h, int setup_finished)
1498 {
1499     MpegEncContext *const s = &h->s;
1500     Picture *out = s->current_picture_ptr;
1501     Picture *cur = s->current_picture_ptr;
1502     int i, pics, out_of_order, out_idx;
1503
1504     s->current_picture_ptr->f.qscale_type = FF_QSCALE_TYPE_H264;
1505     s->current_picture_ptr->f.pict_type   = s->pict_type;
1506
1507     if (h->next_output_pic)
1508         return;
1509
1510     if (cur->field_poc[0] == INT_MAX || cur->field_poc[1] == INT_MAX) {
1511         /* FIXME: if we have two PAFF fields in one packet, we can't start
1512          * the next thread here. If we have one field per packet, we can.
1513          * The check in decode_nal_units() is not good enough to find this
1514          * yet, so we assume the worst for now. */
1515         // if (setup_finished)
1516         //    ff_thread_finish_setup(s->avctx);
1517         return;
1518     }
1519
1520     cur->f.interlaced_frame = 0;
1521     cur->f.repeat_pict      = 0;
1522
1523     /* Signal interlacing information externally. */
1524     /* Prioritize picture timing SEI information over used
1525      * decoding process if it exists. */
1526
1527     if (h->sps.pic_struct_present_flag) {
1528         switch (h->sei_pic_struct) {
1529         case SEI_PIC_STRUCT_FRAME:
1530             break;
1531         case SEI_PIC_STRUCT_TOP_FIELD:
1532         case SEI_PIC_STRUCT_BOTTOM_FIELD:
1533             cur->f.interlaced_frame = 1;
1534             break;
1535         case SEI_PIC_STRUCT_TOP_BOTTOM:
1536         case SEI_PIC_STRUCT_BOTTOM_TOP:
1537             if (FIELD_OR_MBAFF_PICTURE)
1538                 cur->f.interlaced_frame = 1;
1539             else
1540                 // try to flag soft telecine progressive
1541                 cur->f.interlaced_frame = h->prev_interlaced_frame;
1542             break;
1543         case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1544         case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1545             /* Signal the possibility of telecined film externally
1546              * (pic_struct 5,6). From these hints, let the applications
1547              * decide if they apply deinterlacing. */
1548             cur->f.repeat_pict = 1;
1549             break;
1550         case SEI_PIC_STRUCT_FRAME_DOUBLING:
1551             // Force progressive here, doubling interlaced frame is a bad idea.
1552             cur->f.repeat_pict = 2;
1553             break;
1554         case SEI_PIC_STRUCT_FRAME_TRIPLING:
1555             cur->f.repeat_pict = 4;
1556             break;
1557         }
1558
1559         if ((h->sei_ct_type & 3) &&
1560             h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1561             cur->f.interlaced_frame = (h->sei_ct_type & (1 << 1)) != 0;
1562     } else {
1563         /* Derive interlacing flag from used decoding process. */
1564         cur->f.interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1565     }
1566     h->prev_interlaced_frame = cur->f.interlaced_frame;
1567
1568     if (cur->field_poc[0] != cur->field_poc[1]) {
1569         /* Derive top_field_first from field pocs. */
1570         cur->f.top_field_first = cur->field_poc[0] < cur->field_poc[1];
1571     } else {
1572         if (cur->f.interlaced_frame || h->sps.pic_struct_present_flag) {
1573             /* Use picture timing SEI information. Even if it is a
1574              * information of a past frame, better than nothing. */
1575             if (h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM ||
1576                 h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1577                 cur->f.top_field_first = 1;
1578             else
1579                 cur->f.top_field_first = 0;
1580         } else {
1581             /* Most likely progressive */
1582             cur->f.top_field_first = 0;
1583         }
1584     }
1585
1586     cur->mmco_reset = h->mmco_reset;
1587     h->mmco_reset = 0;
1588     // FIXME do something with unavailable reference frames
1589
1590     /* Sort B-frames into display order */
1591
1592     if (h->sps.bitstream_restriction_flag &&
1593         s->avctx->has_b_frames < h->sps.num_reorder_frames) {
1594         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1595         s->low_delay           = 0;
1596     }
1597
1598     if (s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT &&
1599         !h->sps.bitstream_restriction_flag) {
1600         s->avctx->has_b_frames = MAX_DELAYED_PIC_COUNT - 1;
1601         s->low_delay           = 0;
1602     }
1603
1604     for (i = 0; 1; i++) {
1605         if(i == MAX_DELAYED_PIC_COUNT || cur->poc < h->last_pocs[i]){
1606             if(i)
1607                 h->last_pocs[i-1] = cur->poc;
1608             break;
1609         } else if(i) {
1610             h->last_pocs[i-1]= h->last_pocs[i];
1611         }
1612     }
1613     out_of_order = MAX_DELAYED_PIC_COUNT - i;
1614     if(   cur->f.pict_type == AV_PICTURE_TYPE_B
1615        || (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))
1616         out_of_order = FFMAX(out_of_order, 1);
1617     if(s->avctx->has_b_frames < out_of_order && !h->sps.bitstream_restriction_flag){
1618         av_log(s->avctx, AV_LOG_VERBOSE, "Increasing reorder buffer to %d\n", out_of_order);
1619         s->avctx->has_b_frames = out_of_order;
1620         s->low_delay = 0;
1621     }
1622
1623     pics = 0;
1624     while (h->delayed_pic[pics])
1625         pics++;
1626
1627     av_assert0(pics <= MAX_DELAYED_PIC_COUNT);
1628
1629     h->delayed_pic[pics++] = cur;
1630     if (cur->f.reference == 0)
1631         cur->f.reference = DELAYED_PIC_REF;
1632
1633     out = h->delayed_pic[0];
1634     out_idx = 0;
1635     for (i = 1; h->delayed_pic[i] &&
1636                 !h->delayed_pic[i]->f.key_frame &&
1637                 !h->delayed_pic[i]->mmco_reset;
1638          i++)
1639         if (h->delayed_pic[i]->poc < out->poc) {
1640             out     = h->delayed_pic[i];
1641             out_idx = i;
1642         }
1643     if (s->avctx->has_b_frames == 0 &&
1644         (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset))
1645         h->next_outputed_poc = INT_MIN;
1646     out_of_order = out->poc < h->next_outputed_poc;
1647
1648     if (out_of_order || pics > s->avctx->has_b_frames) {
1649         out->f.reference &= ~DELAYED_PIC_REF;
1650         // for frame threading, the owner must be the second field's thread or
1651         // else the first thread can release the picture and reuse it unsafely
1652         out->owner2       = s;
1653         for (i = out_idx; h->delayed_pic[i]; i++)
1654             h->delayed_pic[i] = h->delayed_pic[i + 1];
1655     }
1656     if (!out_of_order && pics > s->avctx->has_b_frames) {
1657         h->next_output_pic = out;
1658         if (out_idx == 0 && h->delayed_pic[0] && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset)) {
1659             h->next_outputed_poc = INT_MIN;
1660         } else
1661             h->next_outputed_poc = out->poc;
1662     } else {
1663         av_log(s->avctx, AV_LOG_DEBUG, "no picture %s\n", out_of_order ? "ooo" : "");
1664     }
1665
1666     if (h->next_output_pic && h->next_output_pic->sync) {
1667         h->sync |= 2;
1668     }
1669
1670     if (setup_finished)
1671         ff_thread_finish_setup(s->avctx);
1672 }
1673
1674 static av_always_inline void backup_mb_border(H264Context *h, uint8_t *src_y,
1675                                               uint8_t *src_cb, uint8_t *src_cr,
1676                                               int linesize, int uvlinesize,
1677                                               int simple)
1678 {
1679     MpegEncContext *const s = &h->s;
1680     uint8_t *top_border;
1681     int top_idx = 1;
1682     const int pixel_shift = h->pixel_shift;
1683     int chroma444 = CHROMA444;
1684     int chroma422 = CHROMA422;
1685
1686     src_y  -= linesize;
1687     src_cb -= uvlinesize;
1688     src_cr -= uvlinesize;
1689
1690     if (!simple && FRAME_MBAFF) {
1691         if (s->mb_y & 1) {
1692             if (!MB_MBAFF) {
1693                 top_border = h->top_borders[0][s->mb_x];
1694                 AV_COPY128(top_border, src_y + 15 * linesize);
1695                 if (pixel_shift)
1696                     AV_COPY128(top_border + 16, src_y + 15 * linesize + 16);
1697                 if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1698                     if (chroma444) {
1699                         if (pixel_shift) {
1700                             AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
1701                             AV_COPY128(top_border + 48, src_cb + 15 * uvlinesize + 16);
1702                             AV_COPY128(top_border + 64, src_cr + 15 * uvlinesize);
1703                             AV_COPY128(top_border + 80, src_cr + 15 * uvlinesize + 16);
1704                         } else {
1705                             AV_COPY128(top_border + 16, src_cb + 15 * uvlinesize);
1706                             AV_COPY128(top_border + 32, src_cr + 15 * uvlinesize);
1707                         }
1708                     } else if (chroma422) {
1709                         if (pixel_shift) {
1710                             AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
1711                             AV_COPY128(top_border + 48, src_cr + 15 * uvlinesize);
1712                         } else {
1713                             AV_COPY64(top_border + 16, src_cb + 15 * uvlinesize);
1714                             AV_COPY64(top_border + 24, src_cr + 15 * uvlinesize);
1715                         }
1716                     } else {
1717                         if (pixel_shift) {
1718                             AV_COPY128(top_border + 32, src_cb + 7 * uvlinesize);
1719                             AV_COPY128(top_border + 48, src_cr + 7 * uvlinesize);
1720                         } else {
1721                             AV_COPY64(top_border + 16, src_cb + 7 * uvlinesize);
1722                             AV_COPY64(top_border + 24, src_cr + 7 * uvlinesize);
1723                         }
1724                     }
1725                 }
1726             }
1727         } else if (MB_MBAFF) {
1728             top_idx = 0;
1729         } else
1730             return;
1731     }
1732
1733     top_border = h->top_borders[top_idx][s->mb_x];
1734     /* There are two lines saved, the line above the the top macroblock
1735      * of a pair, and the line above the bottom macroblock. */
1736     AV_COPY128(top_border, src_y + 16 * linesize);
1737     if (pixel_shift)
1738         AV_COPY128(top_border + 16, src_y + 16 * linesize + 16);
1739
1740     if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1741         if (chroma444) {
1742             if (pixel_shift) {
1743                 AV_COPY128(top_border + 32, src_cb + 16 * linesize);
1744                 AV_COPY128(top_border + 48, src_cb + 16 * linesize + 16);
1745                 AV_COPY128(top_border + 64, src_cr + 16 * linesize);
1746                 AV_COPY128(top_border + 80, src_cr + 16 * linesize + 16);
1747             } else {
1748                 AV_COPY128(top_border + 16, src_cb + 16 * linesize);
1749                 AV_COPY128(top_border + 32, src_cr + 16 * linesize);
1750             }
1751         } else if (chroma422) {
1752             if (pixel_shift) {
1753                 AV_COPY128(top_border + 32, src_cb + 16 * uvlinesize);
1754                 AV_COPY128(top_border + 48, src_cr + 16 * uvlinesize);
1755             } else {
1756                 AV_COPY64(top_border + 16, src_cb + 16 * uvlinesize);
1757                 AV_COPY64(top_border + 24, src_cr + 16 * uvlinesize);
1758             }
1759         } else {
1760             if (pixel_shift) {
1761                 AV_COPY128(top_border + 32, src_cb + 8 * uvlinesize);
1762                 AV_COPY128(top_border + 48, src_cr + 8 * uvlinesize);
1763             } else {
1764                 AV_COPY64(top_border + 16, src_cb + 8 * uvlinesize);
1765                 AV_COPY64(top_border + 24, src_cr + 8 * uvlinesize);
1766             }
1767         }
1768     }
1769 }
1770
1771 static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1772                                             uint8_t *src_cb, uint8_t *src_cr,
1773                                             int linesize, int uvlinesize,
1774                                             int xchg, int chroma444,
1775                                             int simple, int pixel_shift)
1776 {
1777     MpegEncContext *const s = &h->s;
1778     int deblock_topleft;
1779     int deblock_top;
1780     int top_idx = 1;
1781     uint8_t *top_border_m1;
1782     uint8_t *top_border;
1783
1784     if (!simple && FRAME_MBAFF) {
1785         if (s->mb_y & 1) {
1786             if (!MB_MBAFF)
1787                 return;
1788         } else {
1789             top_idx = MB_MBAFF ? 0 : 1;
1790         }
1791     }
1792
1793     if (h->deblocking_filter == 2) {
1794         deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1795         deblock_top     = h->top_type;
1796     } else {
1797         deblock_topleft = (s->mb_x > 0);
1798         deblock_top     = (s->mb_y > !!MB_FIELD);
1799     }
1800
1801     src_y  -= linesize   + 1 + pixel_shift;
1802     src_cb -= uvlinesize + 1 + pixel_shift;
1803     src_cr -= uvlinesize + 1 + pixel_shift;
1804
1805     top_border_m1 = h->top_borders[top_idx][s->mb_x - 1];
1806     top_border    = h->top_borders[top_idx][s->mb_x];
1807
1808 #define XCHG(a, b, xchg)                        \
1809     if (pixel_shift) {                          \
1810         if (xchg) {                             \
1811             AV_SWAP64(b + 0, a + 0);            \
1812             AV_SWAP64(b + 8, a + 8);            \
1813         } else {                                \
1814             AV_COPY128(b, a);                   \
1815         }                                       \
1816     } else if (xchg)                            \
1817         AV_SWAP64(b, a);                        \
1818     else                                        \
1819         AV_COPY64(b, a);
1820
1821     if (deblock_top) {
1822         if (deblock_topleft) {
1823             XCHG(top_border_m1 + (8 << pixel_shift),
1824                  src_y - (7 << pixel_shift), 1);
1825         }
1826         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1827         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1828         if (s->mb_x + 1 < s->mb_width) {
1829             XCHG(h->top_borders[top_idx][s->mb_x + 1],
1830                  src_y + (17 << pixel_shift), 1);
1831         }
1832     }
1833     if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1834         if (chroma444) {
1835             if (deblock_topleft) {
1836                 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1837                 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1838             }
1839             XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1840             XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1841             XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1842             XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1843             if (s->mb_x + 1 < s->mb_width) {
1844                 XCHG(h->top_borders[top_idx][s->mb_x + 1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1845                 XCHG(h->top_borders[top_idx][s->mb_x + 1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1846             }
1847         } else {
1848             if (deblock_top) {
1849                 if (deblock_topleft) {
1850                     XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1851                     XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1852                 }
1853                 XCHG(top_border + (16 << pixel_shift), src_cb + 1 + pixel_shift, 1);
1854                 XCHG(top_border + (24 << pixel_shift), src_cr + 1 + pixel_shift, 1);
1855             }
1856         }
1857     }
1858 }
1859
1860 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth,
1861                                         int index)
1862 {
1863     if (high_bit_depth) {
1864         return AV_RN32A(((int32_t *)mb) + index);
1865     } else
1866         return AV_RN16A(mb + index);
1867 }
1868
1869 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth,
1870                                          int index, int value)
1871 {
1872     if (high_bit_depth) {
1873         AV_WN32A(((int32_t *)mb) + index, value);
1874     } else
1875         AV_WN16A(mb + index, value);
1876 }
1877
1878 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h,
1879                                                        int mb_type, int is_h264,
1880                                                        int simple,
1881                                                        int transform_bypass,
1882                                                        int pixel_shift,
1883                                                        int *block_offset,
1884                                                        int linesize,
1885                                                        uint8_t *dest_y, int p)
1886 {
1887     MpegEncContext *const s = &h->s;
1888     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1889     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1890     int i;
1891     int qscale = p == 0 ? s->qscale : h->chroma_qp[p - 1];
1892     block_offset += 16 * p;
1893     if (IS_INTRA4x4(mb_type)) {
1894         if (simple || !s->encoding) {
1895             if (IS_8x8DCT(mb_type)) {
1896                 if (transform_bypass) {
1897                     idct_dc_add  =
1898                     idct_add     = s->dsp.add_pixels8;
1899                 } else {
1900                     idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1901                     idct_add    = h->h264dsp.h264_idct8_add;
1902                 }
1903                 for (i = 0; i < 16; i += 4) {
1904                     uint8_t *const ptr = dest_y + block_offset[i];
1905                     const int dir      = h->intra4x4_pred_mode_cache[scan8[i]];
1906                     if (transform_bypass && h->sps.profile_idc == 244 && dir <= 1) {
1907                         h->hpc.pred8x8l_add[dir](ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1908                     } else {
1909                         const int nnz = h->non_zero_count_cache[scan8[i + p * 16]];
1910                         h->hpc.pred8x8l[dir](ptr, (h->topleft_samples_available << i) & 0x8000,
1911                                              (h->topright_samples_available << i) & 0x4000, linesize);
1912                         if (nnz) {
1913                             if (nnz == 1 && dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
1914                                 idct_dc_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1915                             else
1916                                 idct_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1917                         }
1918                     }
1919                 }
1920             } else {
1921                 if (transform_bypass) {
1922                     idct_dc_add  =
1923                         idct_add = s->dsp.add_pixels4;
1924                 } else {
1925                     idct_dc_add = h->h264dsp.h264_idct_dc_add;
1926                     idct_add    = h->h264dsp.h264_idct_add;
1927                 }
1928                 for (i = 0; i < 16; i++) {
1929                     uint8_t *const ptr = dest_y + block_offset[i];
1930                     const int dir      = h->intra4x4_pred_mode_cache[scan8[i]];
1931
1932                     if (transform_bypass && h->sps.profile_idc == 244 && dir <= 1) {
1933                         h->hpc.pred4x4_add[dir](ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1934                     } else {
1935                         uint8_t *topright;
1936                         int nnz, tr;
1937                         uint64_t tr_high;
1938                         if (dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED) {
1939                             const int topright_avail = (h->topright_samples_available << i) & 0x8000;
1940                             assert(s->mb_y || linesize <= block_offset[i]);
1941                             if (!topright_avail) {
1942                                 if (pixel_shift) {
1943                                     tr_high  = ((uint16_t *)ptr)[3 - linesize / 2] * 0x0001000100010001ULL;
1944                                     topright = (uint8_t *)&tr_high;
1945                                 } else {
1946                                     tr       = ptr[3 - linesize] * 0x01010101u;
1947                                     topright = (uint8_t *)&tr;
1948                                 }
1949                             } else
1950                                 topright = ptr + (4 << pixel_shift) - linesize;
1951                         } else
1952                             topright = NULL;
1953
1954                         h->hpc.pred4x4[dir](ptr, topright, linesize);
1955                         nnz = h->non_zero_count_cache[scan8[i + p * 16]];
1956                         if (nnz) {
1957                             if (is_h264) {
1958                                 if (nnz == 1 && dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
1959                                     idct_dc_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1960                                 else
1961                                     idct_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1962                             } else if (CONFIG_SVQ3_DECODER)
1963                                 ff_svq3_add_idct_c(ptr, h->mb + i * 16 + p * 256, linesize, qscale, 0);
1964                         }
1965                     }
1966                 }
1967             }
1968         }
1969     } else {
1970         h->hpc.pred16x16[h->intra16x16_pred_mode](dest_y, linesize);
1971         if (is_h264) {
1972             if (h->non_zero_count_cache[scan8[LUMA_DC_BLOCK_INDEX + p]]) {
1973                 if (!transform_bypass)
1974                     h->h264dsp.h264_luma_dc_dequant_idct(h->mb + (p * 256 << pixel_shift),
1975                                                          h->mb_luma_dc[p],
1976                                                          h->dequant4_coeff[p][qscale][0]);
1977                 else {
1978                     static const uint8_t dc_mapping[16] = {
1979                          0 * 16,  1 * 16,  4 * 16,  5 * 16,
1980                          2 * 16,  3 * 16,  6 * 16,  7 * 16,
1981                          8 * 16,  9 * 16, 12 * 16, 13 * 16,
1982                         10 * 16, 11 * 16, 14 * 16, 15 * 16 };
1983                     for (i = 0; i < 16; i++)
1984                         dctcoef_set(h->mb + (p * 256 << pixel_shift),
1985                                     pixel_shift, dc_mapping[i],
1986                                     dctcoef_get(h->mb_luma_dc[p],
1987                                                 pixel_shift, i));
1988                 }
1989             }
1990         } else if (CONFIG_SVQ3_DECODER)
1991             ff_svq3_luma_dc_dequant_idct_c(h->mb + p * 256,
1992                                            h->mb_luma_dc[p], qscale);
1993     }
1994 }
1995
1996 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type,
1997                                                     int is_h264, int simple,
1998                                                     int transform_bypass,
1999                                                     int pixel_shift,
2000                                                     int *block_offset,
2001                                                     int linesize,
2002                                                     uint8_t *dest_y, int p)
2003 {
2004     MpegEncContext *const s = &h->s;
2005     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2006     int i;
2007     block_offset += 16 * p;
2008     if (!IS_INTRA4x4(mb_type)) {
2009         if (is_h264) {
2010             if (IS_INTRA16x16(mb_type)) {
2011                 if (transform_bypass) {
2012                     if (h->sps.profile_idc == 244 &&
2013                         (h->intra16x16_pred_mode == VERT_PRED8x8 ||
2014                          h->intra16x16_pred_mode == HOR_PRED8x8)) {
2015                         h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset,
2016                                                                       h->mb + (p * 256 << pixel_shift),
2017                                                                       linesize);
2018                     } else {
2019                         for (i = 0; i < 16; i++)
2020                             if (h->non_zero_count_cache[scan8[i + p * 16]] ||
2021                                 dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
2022                                 s->dsp.add_pixels4(dest_y + block_offset[i],
2023                                                    h->mb + (i * 16 + p * 256 << pixel_shift),
2024                                                    linesize);
2025                     }
2026                 } else {
2027                     h->h264dsp.h264_idct_add16intra(dest_y, block_offset,
2028                                                     h->mb + (p * 256 << pixel_shift),
2029                                                     linesize,
2030                                                     h->non_zero_count_cache + p * 5 * 8);
2031                 }
2032             } else if (h->cbp & 15) {
2033                 if (transform_bypass) {
2034                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
2035                     idct_add = IS_8x8DCT(mb_type) ? s->dsp.add_pixels8
2036                                                   : s->dsp.add_pixels4;
2037                     for (i = 0; i < 16; i += di)
2038                         if (h->non_zero_count_cache[scan8[i + p * 16]])
2039                             idct_add(dest_y + block_offset[i],
2040                                      h->mb + (i * 16 + p * 256 << pixel_shift),
2041                                      linesize);
2042                 } else {
2043                     if (IS_8x8DCT(mb_type))
2044                         h->h264dsp.h264_idct8_add4(dest_y, block_offset,
2045                                                    h->mb + (p * 256 << pixel_shift),
2046                                                    linesize,
2047                                                    h->non_zero_count_cache + p * 5 * 8);
2048                     else
2049                         h->h264dsp.h264_idct_add16(dest_y, block_offset,
2050                                                    h->mb + (p * 256 << pixel_shift),
2051                                                    linesize,
2052                                                    h->non_zero_count_cache + p * 5 * 8);
2053                 }
2054             }
2055         } else if (CONFIG_SVQ3_DECODER) {
2056             for (i = 0; i < 16; i++)
2057                 if (h->non_zero_count_cache[scan8[i + p * 16]] || h->mb[i * 16 + p * 256]) {
2058                     // FIXME benchmark weird rule, & below
2059                     uint8_t *const ptr = dest_y + block_offset[i];
2060                     ff_svq3_add_idct_c(ptr, h->mb + i * 16 + p * 256, linesize,
2061                                        s->qscale, IS_INTRA(mb_type) ? 1 : 0);
2062                 }
2063         }
2064     }
2065 }
2066
2067 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple,
2068                                                    int pixel_shift)
2069 {
2070     MpegEncContext *const s = &h->s;
2071     const int mb_x    = s->mb_x;
2072     const int mb_y    = s->mb_y;
2073     const int mb_xy   = h->mb_xy;
2074     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2075     uint8_t *dest_y, *dest_cb, *dest_cr;
2076     int linesize, uvlinesize /*dct_offset*/;
2077     int i, j;
2078     int *block_offset = &h->block_offset[0];
2079     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
2080     /* is_h264 should always be true if SVQ3 is disabled. */
2081     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
2082     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2083     const int block_h   = 16 >> s->chroma_y_shift;
2084     const int chroma422 = CHROMA422;
2085
2086     dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift)     + mb_y * s->linesize)  * 16;
2087     dest_cb = s->current_picture.f.data[1] +  (mb_x << pixel_shift) * 8 + mb_y * s->uvlinesize * block_h;
2088     dest_cr = s->current_picture.f.data[2] +  (mb_x << pixel_shift) * 8 + mb_y * s->uvlinesize * block_h;
2089
2090     s->dsp.prefetch(dest_y  + (s->mb_x & 3) * 4 * s->linesize   + (64 << pixel_shift), s->linesize,       4);
2091     s->dsp.prefetch(dest_cb + (s->mb_x & 7)     * s->uvlinesize + (64 << pixel_shift), dest_cr - dest_cb, 2);
2092
2093     h->list_counts[mb_xy] = h->list_count;
2094
2095     if (!simple && MB_FIELD) {
2096         linesize     = h->mb_linesize = s->linesize * 2;
2097         uvlinesize   = h->mb_uvlinesize = s->uvlinesize * 2;
2098         block_offset = &h->block_offset[48];
2099         if (mb_y & 1) { // FIXME move out of this function?
2100             dest_y  -= s->linesize * 15;
2101             dest_cb -= s->uvlinesize * (block_h - 1);
2102             dest_cr -= s->uvlinesize * (block_h - 1);
2103         }
2104         if (FRAME_MBAFF) {
2105             int list;
2106             for (list = 0; list < h->list_count; list++) {
2107                 if (!USES_LIST(mb_type, list))
2108                     continue;
2109                 if (IS_16X16(mb_type)) {
2110                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2111                     fill_rectangle(ref, 4, 4, 8, (16 + *ref) ^ (s->mb_y & 1), 1);
2112                 } else {
2113                     for (i = 0; i < 16; i += 4) {
2114                         int ref = h->ref_cache[list][scan8[i]];
2115                         if (ref >= 0)
2116                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2,
2117                                            8, (16 + ref) ^ (s->mb_y & 1), 1);
2118                     }
2119                 }
2120             }
2121         }
2122     } else {
2123         linesize   = h->mb_linesize   = s->linesize;
2124         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
2125         // dct_offset = s->linesize * 16;
2126     }
2127
2128     if (!simple && IS_INTRA_PCM(mb_type)) {
2129         const int bit_depth = h->sps.bit_depth_luma;
2130         if (pixel_shift) {
2131             int j;
2132             GetBitContext gb;
2133             init_get_bits(&gb, (uint8_t *)h->mb,
2134                           ff_h264_mb_sizes[h->sps.chroma_format_idc] * bit_depth);
2135
2136             for (i = 0; i < 16; i++) {
2137                 uint16_t *tmp_y = (uint16_t *)(dest_y + i * linesize);
2138                 for (j = 0; j < 16; j++)
2139                     tmp_y[j] = get_bits(&gb, bit_depth);
2140             }
2141             if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
2142                 if (!h->sps.chroma_format_idc) {
2143                     for (i = 0; i < block_h; i++) {
2144                         uint16_t *tmp_cb = (uint16_t *)(dest_cb + i * uvlinesize);
2145                         uint16_t *tmp_cr = (uint16_t *)(dest_cr + i * uvlinesize);
2146                         for (j = 0; j < 8; j++) {
2147                             tmp_cb[j] = tmp_cr[j] = 1 << (bit_depth - 1);
2148                         }
2149                     }
2150                 } else {
2151                     for (i = 0; i < block_h; i++) {
2152                         uint16_t *tmp_cb = (uint16_t *)(dest_cb + i * uvlinesize);
2153                         for (j = 0; j < 8; j++)
2154                             tmp_cb[j] = get_bits(&gb, bit_depth);
2155                     }
2156                     for (i = 0; i < block_h; i++) {
2157                         uint16_t *tmp_cr = (uint16_t *)(dest_cr + i * uvlinesize);
2158                         for (j = 0; j < 8; j++)
2159                             tmp_cr[j] = get_bits(&gb, bit_depth);
2160                     }
2161                 }
2162             }
2163         } else {
2164             for (i = 0; i < 16; i++)
2165                 memcpy(dest_y + i * linesize, (uint8_t *)h->mb + i * 16, 16);
2166             if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
2167                 if (!h->sps.chroma_format_idc) {
2168                     for (i = 0; i < 8; i++) {
2169                         memset(dest_cb + i*uvlinesize, 1 << (bit_depth - 1), 8);
2170                         memset(dest_cr + i*uvlinesize, 1 << (bit_depth - 1), 8);
2171                     }
2172                 } else {
2173                     uint8_t *src_cb = (uint8_t *)h->mb + 256;
2174                     uint8_t *src_cr = (uint8_t *)h->mb + 256 + block_h * 8;
2175                     for (i = 0; i < block_h; i++) {
2176                         memcpy(dest_cb + i * uvlinesize, src_cb + i * 8, 8);
2177                         memcpy(dest_cr + i * uvlinesize, src_cr + i * 8, 8);
2178                     }
2179                 }
2180             }
2181         }
2182     } else {
2183         if (IS_INTRA(mb_type)) {
2184             if (h->deblocking_filter)
2185                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize,
2186                                uvlinesize, 1, 0, simple, pixel_shift);
2187
2188             if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
2189                 if (CHROMA) {
2190                     h->hpc.pred8x8[h->chroma_pred_mode](dest_cb, uvlinesize);
2191                     h->hpc.pred8x8[h->chroma_pred_mode](dest_cr, uvlinesize);
2192                 }
2193             }
2194
2195             hl_decode_mb_predict_luma(h, mb_type, is_h264, simple,
2196                                       transform_bypass, pixel_shift,
2197                                       block_offset, linesize, dest_y, 0);
2198
2199             if (h->deblocking_filter)
2200                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize,
2201                                uvlinesize, 0, 0, simple, pixel_shift);
2202         } else if (is_h264) {
2203             if (chroma422) {
2204                 hl_motion_422(h, dest_y, dest_cb, dest_cr,
2205                               s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2206                               s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2207                               h->h264dsp.weight_h264_pixels_tab,
2208                               h->h264dsp.biweight_h264_pixels_tab,
2209                               pixel_shift);
2210             } else {
2211                 hl_motion_420(h, dest_y, dest_cb, dest_cr,
2212                               s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2213                               s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2214                               h->h264dsp.weight_h264_pixels_tab,
2215                               h->h264dsp.biweight_h264_pixels_tab,
2216                               pixel_shift);
2217             }
2218         }
2219
2220         hl_decode_mb_idct_luma(h, mb_type, is_h264, simple, transform_bypass,
2221                                pixel_shift, block_offset, linesize, dest_y, 0);
2222
2223         if ((simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) &&
2224             (h->cbp & 0x30)) {
2225             uint8_t *dest[2] = { dest_cb, dest_cr };
2226             if (transform_bypass) {
2227                 if (IS_INTRA(mb_type) && h->sps.profile_idc == 244 &&
2228                     (h->chroma_pred_mode == VERT_PRED8x8 ||
2229                      h->chroma_pred_mode == HOR_PRED8x8)) {
2230                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0],
2231                                                             block_offset + 16,
2232                                                             h->mb + (16 * 16 * 1 << pixel_shift),
2233                                                             uvlinesize);
2234                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1],
2235                                                             block_offset + 32,
2236                                                             h->mb + (16 * 16 * 2 << pixel_shift),
2237                                                             uvlinesize);
2238                 } else {
2239                     idct_add = s->dsp.add_pixels4;
2240                     for (j = 1; j < 3; j++) {
2241                         for (i = j * 16; i < j * 16 + 4; i++)
2242                             if (h->non_zero_count_cache[scan8[i]] ||
2243                                 dctcoef_get(h->mb, pixel_shift, i * 16))
2244                                 idct_add(dest[j - 1] + block_offset[i],
2245                                          h->mb + (i * 16 << pixel_shift),
2246                                          uvlinesize);
2247                         if (chroma422) {
2248                             for (i = j * 16 + 4; i < j * 16 + 8; i++)
2249                                 if (h->non_zero_count_cache[scan8[i + 4]] ||
2250                                     dctcoef_get(h->mb, pixel_shift, i * 16))
2251                                     idct_add(dest[j - 1] + block_offset[i + 4],
2252                                              h->mb + (i * 16 << pixel_shift),
2253                                              uvlinesize);
2254                         }
2255                     }
2256                 }
2257             } else {
2258                 if (is_h264) {
2259                     int qp[2];
2260                     if (chroma422) {
2261                         qp[0] = h->chroma_qp[0] + 3;
2262                         qp[1] = h->chroma_qp[1] + 3;
2263                     } else {
2264                         qp[0] = h->chroma_qp[0];
2265                         qp[1] = h->chroma_qp[1];
2266                     }
2267                     if (h->non_zero_count_cache[scan8[CHROMA_DC_BLOCK_INDEX + 0]])
2268                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16 * 16 * 1 << pixel_shift),
2269                                                                h->dequant4_coeff[IS_INTRA(mb_type) ? 1 : 4][qp[0]][0]);
2270                     if (h->non_zero_count_cache[scan8[CHROMA_DC_BLOCK_INDEX + 1]])
2271                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16 * 16 * 2 << pixel_shift),
2272                                                                h->dequant4_coeff[IS_INTRA(mb_type) ? 2 : 5][qp[1]][0]);
2273                     h->h264dsp.h264_idct_add8(dest, block_offset,
2274                                               h->mb, uvlinesize,
2275                                               h->non_zero_count_cache);
2276                 } else if (CONFIG_SVQ3_DECODER) {
2277                     h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + 16 * 16 * 1,
2278                                                            h->dequant4_coeff[IS_INTRA(mb_type) ? 1 : 4][h->chroma_qp[0]][0]);
2279                     h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + 16 * 16 * 2,
2280                                                            h->dequant4_coeff[IS_INTRA(mb_type) ? 2 : 5][h->chroma_qp[1]][0]);
2281                     for (j = 1; j < 3; j++) {
2282                         for (i = j * 16; i < j * 16 + 4; i++)
2283                             if (h->non_zero_count_cache[scan8[i]] || h->mb[i * 16]) {
2284                                 uint8_t *const ptr = dest[j - 1] + block_offset[i];
2285                                 ff_svq3_add_idct_c(ptr, h->mb + i * 16,
2286                                                    uvlinesize,
2287                                                    ff_h264_chroma_qp[0][s->qscale + 12] - 12, 2);
2288                             }
2289                     }
2290                 }
2291             }
2292         }
2293     }
2294     if (h->cbp || IS_INTRA(mb_type)) {
2295         s->dsp.clear_blocks(h->mb);
2296         s->dsp.clear_blocks(h->mb + (24 * 16 << pixel_shift));
2297     }
2298 }
2299
2300 static av_always_inline void hl_decode_mb_444_internal(H264Context *h,
2301                                                        int simple,
2302                                                        int pixel_shift)
2303 {
2304     MpegEncContext *const s = &h->s;
2305     const int mb_x    = s->mb_x;
2306     const int mb_y    = s->mb_y;
2307     const int mb_xy   = h->mb_xy;
2308     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2309     uint8_t *dest[3];
2310     int linesize;
2311     int i, j, p;
2312     int *block_offset = &h->block_offset[0];
2313     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
2314     const int plane_count      = (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) ? 3 : 1;
2315
2316     for (p = 0; p < plane_count; p++) {
2317         dest[p] = s->current_picture.f.data[p] +
2318                   ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
2319         s->dsp.prefetch(dest[p] + (s->mb_x & 3) * 4 * s->linesize + (64 << pixel_shift),
2320                         s->linesize, 4);
2321     }
2322
2323     h->list_counts[mb_xy] = h->list_count;
2324
2325     if (!simple && MB_FIELD) {
2326         linesize     = h->mb_linesize = h->mb_uvlinesize = s->linesize * 2;
2327         block_offset = &h->block_offset[48];
2328         if (mb_y & 1) // FIXME move out of this function?
2329             for (p = 0; p < 3; p++)
2330                 dest[p] -= s->linesize * 15;
2331         if (FRAME_MBAFF) {
2332             int list;
2333             for (list = 0; list < h->list_count; list++) {
2334                 if (!USES_LIST(mb_type, list))
2335                     continue;
2336                 if (IS_16X16(mb_type)) {
2337                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2338                     fill_rectangle(ref, 4, 4, 8, (16 + *ref) ^ (s->mb_y & 1), 1);
2339                 } else {
2340                     for (i = 0; i < 16; i += 4) {
2341                         int ref = h->ref_cache[list][scan8[i]];
2342                         if (ref >= 0)
2343                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2,
2344                                            8, (16 + ref) ^ (s->mb_y & 1), 1);
2345                     }
2346                 }
2347             }
2348         }
2349     } else {
2350         linesize = h->mb_linesize = h->mb_uvlinesize = s->linesize;
2351     }
2352
2353     if (!simple && IS_INTRA_PCM(mb_type)) {
2354         if (pixel_shift) {
2355             const int bit_depth = h->sps.bit_depth_luma;
2356             GetBitContext gb;
2357             init_get_bits(&gb, (uint8_t *)h->mb, 768 * bit_depth);
2358
2359             for (p = 0; p < plane_count; p++)
2360                 for (i = 0; i < 16; i++) {
2361                     uint16_t *tmp = (uint16_t *)(dest[p] + i * linesize);
2362                     for (j = 0; j < 16; j++)
2363                         tmp[j] = get_bits(&gb, bit_depth);
2364                 }
2365         } else {
2366             for (p = 0; p < plane_count; p++)
2367                 for (i = 0; i < 16; i++)
2368                     memcpy(dest[p] + i * linesize,
2369                            (uint8_t *)h->mb + p * 256 + i * 16, 16);
2370         }
2371     } else {
2372         if (IS_INTRA(mb_type)) {
2373             if (h->deblocking_filter)
2374                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize,
2375                                linesize, 1, 1, simple, pixel_shift);
2376
2377             for (p = 0; p < plane_count; p++)
2378                 hl_decode_mb_predict_luma(h, mb_type, 1, simple,
2379                                           transform_bypass, pixel_shift,
2380                                           block_offset, linesize, dest[p], p);
2381
2382             if (h->deblocking_filter)
2383                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize,
2384                                linesize, 0, 1, simple, pixel_shift);
2385         } else {
2386             hl_motion(h, dest[0], dest[1], dest[2],
2387                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2388                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2389                       h->h264dsp.weight_h264_pixels_tab,
2390                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 3);
2391         }
2392
2393         for (p = 0; p < plane_count; p++)
2394             hl_decode_mb_idct_luma(h, mb_type, 1, simple, transform_bypass,
2395                                    pixel_shift, block_offset, linesize,
2396                                    dest[p], p);
2397     }
2398     if (h->cbp || IS_INTRA(mb_type)) {
2399         s->dsp.clear_blocks(h->mb);
2400         s->dsp.clear_blocks(h->mb + (24 * 16 << pixel_shift));
2401     }
2402 }
2403
2404 /**
2405  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2406  */
2407 #define hl_decode_mb_simple(sh, bits)                          \
2408 static void hl_decode_mb_simple_ ## bits(H264Context *h)       \
2409 {                                                              \
2410     hl_decode_mb_internal(h, 1, sh);                           \
2411 }
2412
2413 hl_decode_mb_simple(0, 8)
2414 hl_decode_mb_simple(1, 16)
2415
2416 /**
2417  * Process a macroblock; this handles edge cases, such as interlacing.
2418  */
2419 static av_noinline void hl_decode_mb_complex(H264Context *h)
2420 {
2421     hl_decode_mb_internal(h, 0, h->pixel_shift);
2422 }
2423
2424 static av_noinline void hl_decode_mb_444_complex(H264Context *h)
2425 {
2426     hl_decode_mb_444_internal(h, 0, h->pixel_shift);
2427 }
2428
2429 static av_noinline void hl_decode_mb_444_simple(H264Context *h)
2430 {
2431     hl_decode_mb_444_internal(h, 1, 0);
2432 }
2433
2434 void ff_h264_hl_decode_mb(H264Context *h)
2435 {
2436     MpegEncContext *const s = &h->s;
2437     const int mb_xy   = h->mb_xy;
2438     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2439     int is_complex    = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2440
2441     if (CHROMA444) {
2442         if (is_complex || h->pixel_shift)
2443             hl_decode_mb_444_complex(h);
2444         else
2445             hl_decode_mb_444_simple(h);
2446     } else if (is_complex) {
2447         hl_decode_mb_complex(h);
2448     } else if (h->pixel_shift) {
2449         hl_decode_mb_simple_16(h);
2450     } else
2451         hl_decode_mb_simple_8(h);
2452 }
2453
2454 static int pred_weight_table(H264Context *h)
2455 {
2456     MpegEncContext *const s = &h->s;
2457     int list, i;
2458     int luma_def, chroma_def;
2459
2460     h->use_weight             = 0;
2461     h->use_weight_chroma      = 0;
2462     h->luma_log2_weight_denom = get_ue_golomb(&s->gb);
2463     if (h->sps.chroma_format_idc)
2464         h->chroma_log2_weight_denom = get_ue_golomb(&s->gb);
2465     luma_def   = 1 << h->luma_log2_weight_denom;
2466     chroma_def = 1 << h->chroma_log2_weight_denom;
2467
2468     for (list = 0; list < 2; list++) {
2469         h->luma_weight_flag[list]   = 0;
2470         h->chroma_weight_flag[list] = 0;
2471         for (i = 0; i < h->ref_count[list]; i++) {
2472             int luma_weight_flag, chroma_weight_flag;
2473
2474             luma_weight_flag = get_bits1(&s->gb);
2475             if (luma_weight_flag) {
2476                 h->luma_weight[i][list][0] = get_se_golomb(&s->gb);
2477                 h->luma_weight[i][list][1] = get_se_golomb(&s->gb);
2478                 if (h->luma_weight[i][list][0] != luma_def ||
2479                     h->luma_weight[i][list][1] != 0) {
2480                     h->use_weight             = 1;
2481                     h->luma_weight_flag[list] = 1;
2482                 }
2483             } else {
2484                 h->luma_weight[i][list][0] = luma_def;
2485                 h->luma_weight[i][list][1] = 0;
2486             }
2487
2488             if (h->sps.chroma_format_idc) {
2489                 chroma_weight_flag = get_bits1(&s->gb);
2490                 if (chroma_weight_flag) {
2491                     int j;
2492                     for (j = 0; j < 2; j++) {
2493                         h->chroma_weight[i][list][j][0] = get_se_golomb(&s->gb);
2494                         h->chroma_weight[i][list][j][1] = get_se_golomb(&s->gb);
2495                         if (h->chroma_weight[i][list][j][0] != chroma_def ||
2496                             h->chroma_weight[i][list][j][1] != 0) {
2497                             h->use_weight_chroma = 1;
2498                             h->chroma_weight_flag[list] = 1;
2499                         }
2500                     }
2501                 } else {
2502                     int j;
2503                     for (j = 0; j < 2; j++) {
2504                         h->chroma_weight[i][list][j][0] = chroma_def;
2505                         h->chroma_weight[i][list][j][1] = 0;
2506                     }
2507                 }
2508             }
2509         }
2510         if (h->slice_type_nos != AV_PICTURE_TYPE_B)
2511             break;
2512     }
2513     h->use_weight = h->use_weight || h->use_weight_chroma;
2514     return 0;
2515 }
2516
2517 /**
2518  * Initialize implicit_weight table.
2519  * @param field  0/1 initialize the weight for interlaced MBAFF
2520  *                -1 initializes the rest
2521  */
2522 static void implicit_weight_table(H264Context *h, int field)
2523 {
2524     MpegEncContext *const s = &h->s;
2525     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2526
2527     for (i = 0; i < 2; i++) {
2528         h->luma_weight_flag[i]   = 0;
2529         h->chroma_weight_flag[i] = 0;
2530     }
2531
2532     if (field < 0) {
2533         if (s->picture_structure == PICT_FRAME) {
2534             cur_poc = s->current_picture_ptr->poc;
2535         } else {
2536             cur_poc = s->current_picture_ptr->field_poc[s->picture_structure - 1];
2537         }
2538         if (h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF &&
2539             h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2 * cur_poc) {
2540             h->use_weight = 0;
2541             h->use_weight_chroma = 0;
2542             return;
2543         }
2544         ref_start  = 0;
2545         ref_count0 = h->ref_count[0];
2546         ref_count1 = h->ref_count[1];
2547     } else {
2548         cur_poc    = s->current_picture_ptr->field_poc[field];
2549         ref_start  = 16;
2550         ref_count0 = 16 + 2 * h->ref_count[0];
2551         ref_count1 = 16 + 2 * h->ref_count[1];
2552     }
2553
2554     h->use_weight               = 2;
2555     h->use_weight_chroma        = 2;
2556     h->luma_log2_weight_denom   = 5;
2557     h->chroma_log2_weight_denom = 5;
2558
2559     for (ref0 = ref_start; ref0 < ref_count0; ref0++) {
2560         int poc0 = h->ref_list[0][ref0].poc;
2561         for (ref1 = ref_start; ref1 < ref_count1; ref1++) {
2562             int w = 32;
2563             if (!h->ref_list[0][ref0].long_ref && !h->ref_list[1][ref1].long_ref) {
2564                 int poc1 = h->ref_list[1][ref1].poc;
2565                 int td   = av_clip(poc1 - poc0, -128, 127);
2566                 if (td) {
2567                     int tb = av_clip(cur_poc - poc0, -128, 127);
2568                     int tx = (16384 + (FFABS(td) >> 1)) / td;
2569                     int dist_scale_factor = (tb * tx + 32) >> 8;
2570                     if (dist_scale_factor >= -64 && dist_scale_factor <= 128)
2571                         w = 64 - dist_scale_factor;
2572                 }
2573             }
2574             if (field < 0) {
2575                 h->implicit_weight[ref0][ref1][0] =
2576                 h->implicit_weight[ref0][ref1][1] = w;
2577             } else {
2578                 h->implicit_weight[ref0][ref1][field] = w;
2579             }
2580         }
2581     }
2582 }
2583
2584 /**
2585  * instantaneous decoder refresh.
2586  */
2587 static void idr(H264Context *h)
2588 {
2589     int i;
2590     ff_h264_remove_all_refs(h);
2591     h->prev_frame_num        = 0;
2592     h->prev_frame_num_offset = 0;
2593     h->prev_poc_msb          = 1<<16;
2594     h->prev_poc_lsb          = 0;
2595     for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
2596         h->last_pocs[i] = INT_MIN;
2597 }
2598
2599 /* forget old pics after a seek */
2600 static void flush_dpb(AVCodecContext *avctx)
2601 {
2602     H264Context *h = avctx->priv_data;
2603     int i;
2604     for (i=0; i<=MAX_DELAYED_PIC_COUNT; i++) {
2605         if (h->delayed_pic[i])
2606             h->delayed_pic[i]->f.reference = 0;
2607         h->delayed_pic[i] = NULL;
2608     }
2609     h->outputed_poc = h->next_outputed_poc = INT_MIN;
2610     h->prev_interlaced_frame = 1;
2611     idr(h);
2612     h->prev_frame_num = -1;
2613     if (h->s.current_picture_ptr)
2614         h->s.current_picture_ptr->f.reference = 0;
2615     h->s.first_field = 0;
2616     ff_h264_reset_sei(h);
2617     ff_mpeg_flush(avctx);
2618     h->recovery_frame= -1;
2619     h->sync= 0;
2620 }
2621
2622 static int init_poc(H264Context *h)
2623 {
2624     MpegEncContext *const s = &h->s;
2625     const int max_frame_num = 1 << h->sps.log2_max_frame_num;
2626     int field_poc[2];
2627     Picture *cur = s->current_picture_ptr;
2628
2629     h->frame_num_offset = h->prev_frame_num_offset;
2630     if (h->frame_num < h->prev_frame_num)
2631         h->frame_num_offset += max_frame_num;
2632
2633     if (h->sps.poc_type == 0) {
2634         const int max_poc_lsb = 1 << h->sps.log2_max_poc_lsb;
2635
2636         if (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb / 2)
2637             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2638         else if (h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb / 2)
2639             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2640         else
2641             h->poc_msb = h->prev_poc_msb;
2642         // printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
2643         field_poc[0] =
2644         field_poc[1] = h->poc_msb + h->poc_lsb;
2645         if (s->picture_structure == PICT_FRAME)
2646             field_poc[1] += h->delta_poc_bottom;
2647     } else if (h->sps.poc_type == 1) {
2648         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2649         int i;
2650
2651         if (h->sps.poc_cycle_length != 0)
2652             abs_frame_num = h->frame_num_offset + h->frame_num;
2653         else
2654             abs_frame_num = 0;
2655
2656         if (h->nal_ref_idc == 0 && abs_frame_num > 0)
2657             abs_frame_num--;
2658
2659         expected_delta_per_poc_cycle = 0;
2660         for (i = 0; i < h->sps.poc_cycle_length; i++)
2661             // FIXME integrate during sps parse
2662             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[i];
2663
2664         if (abs_frame_num > 0) {
2665             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2666             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2667
2668             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2669             for (i = 0; i <= frame_num_in_poc_cycle; i++)
2670                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[i];
2671         } else
2672             expectedpoc = 0;
2673
2674         if (h->nal_ref_idc == 0)
2675             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2676
2677         field_poc[0] = expectedpoc + h->delta_poc[0];
2678         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2679
2680         if (s->picture_structure == PICT_FRAME)
2681             field_poc[1] += h->delta_poc[1];
2682     } else {
2683         int poc = 2 * (h->frame_num_offset + h->frame_num);
2684
2685         if (!h->nal_ref_idc)
2686             poc--;
2687
2688         field_poc[0] = poc;
2689         field_poc[1] = poc;
2690     }
2691
2692     if (s->picture_structure != PICT_BOTTOM_FIELD)
2693         s->current_picture_ptr->field_poc[0] = field_poc[0];
2694     if (s->picture_structure != PICT_TOP_FIELD)
2695         s->current_picture_ptr->field_poc[1] = field_poc[1];
2696     cur->poc = FFMIN(cur->field_poc[0], cur->field_poc[1]);
2697
2698     return 0;
2699 }
2700
2701 /**
2702  * initialize scan tables
2703  */
2704 static void init_scan_tables(H264Context *h)
2705 {
2706     int i;
2707     for (i = 0; i < 16; i++) {
2708 #define T(x) (x >> 2) | ((x << 2) & 0xF)
2709         h->zigzag_scan[i] = T(zigzag_scan[i]);
2710         h->field_scan[i]  = T(field_scan[i]);
2711 #undef T
2712     }
2713     for (i = 0; i < 64; i++) {
2714 #define T(x) (x >> 3) | ((x & 7) << 3)
2715         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
2716         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2717         h->field_scan8x8[i]        = T(field_scan8x8[i]);
2718         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
2719 #undef T
2720     }
2721     if (h->sps.transform_bypass) { // FIXME same ugly
2722         memcpy(h->zigzag_scan_q0          , zigzag_scan             , sizeof(h->zigzag_scan_q0         ));
2723         memcpy(h->zigzag_scan8x8_q0       , ff_zigzag_direct        , sizeof(h->zigzag_scan8x8_q0      ));
2724         memcpy(h->zigzag_scan8x8_cavlc_q0 , zigzag_scan8x8_cavlc    , sizeof(h->zigzag_scan8x8_cavlc_q0));
2725         memcpy(h->field_scan_q0           , field_scan              , sizeof(h->field_scan_q0          ));
2726         memcpy(h->field_scan8x8_q0        , field_scan8x8           , sizeof(h->field_scan8x8_q0       ));
2727         memcpy(h->field_scan8x8_cavlc_q0  , field_scan8x8_cavlc     , sizeof(h->field_scan8x8_cavlc_q0 ));
2728     } else {
2729         memcpy(h->zigzag_scan_q0          , h->zigzag_scan          , sizeof(h->zigzag_scan_q0         ));
2730         memcpy(h->zigzag_scan8x8_q0       , h->zigzag_scan8x8       , sizeof(h->zigzag_scan8x8_q0      ));
2731         memcpy(h->zigzag_scan8x8_cavlc_q0 , h->zigzag_scan8x8_cavlc , sizeof(h->zigzag_scan8x8_cavlc_q0));
2732         memcpy(h->field_scan_q0           , h->field_scan           , sizeof(h->field_scan_q0          ));
2733         memcpy(h->field_scan8x8_q0        , h->field_scan8x8        , sizeof(h->field_scan8x8_q0       ));
2734         memcpy(h->field_scan8x8_cavlc_q0  , h->field_scan8x8_cavlc  , sizeof(h->field_scan8x8_cavlc_q0 ));
2735     }
2736 }
2737
2738 static int field_end(H264Context *h, int in_setup)
2739 {
2740     MpegEncContext *const s     = &h->s;
2741     AVCodecContext *const avctx = s->avctx;
2742     int err = 0;
2743     s->mb_y = 0;
2744
2745     if (!in_setup && !s->dropable)
2746         ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX,
2747                                   s->picture_structure == PICT_BOTTOM_FIELD);
2748
2749     if (CONFIG_H264_VDPAU_DECODER &&
2750         s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
2751         ff_vdpau_h264_set_reference_frames(s);
2752
2753     if (in_setup || !(avctx->active_thread_type & FF_THREAD_FRAME)) {
2754         if (!s->dropable) {
2755             err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2756             h->prev_poc_msb = h->poc_msb;
2757             h->prev_poc_lsb = h->poc_lsb;
2758         }
2759         h->prev_frame_num_offset = h->frame_num_offset;
2760         h->prev_frame_num        = h->frame_num;
2761         h->outputed_poc          = h->next_outputed_poc;
2762     }
2763
2764     if (avctx->hwaccel) {
2765         if (avctx->hwaccel->end_frame(avctx) < 0)
2766             av_log(avctx, AV_LOG_ERROR,
2767                    "hardware accelerator failed to decode picture\n");
2768     }
2769
2770     if (CONFIG_H264_VDPAU_DECODER &&
2771         s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
2772         ff_vdpau_h264_picture_complete(s);
2773
2774     /*
2775      * FIXME: Error handling code does not seem to support interlaced
2776      * when slices span multiple rows
2777      * The ff_er_add_slice calls don't work right for bottom
2778      * fields; they cause massive erroneous error concealing
2779      * Error marking covers both fields (top and bottom).
2780      * This causes a mismatched s->error_count
2781      * and a bad error table. Further, the error count goes to
2782      * INT_MAX when called for bottom field, because mb_y is
2783      * past end by one (callers fault) and resync_mb_y != 0
2784      * causes problems for the first MB line, too.
2785      */
2786     if (!FIELD_PICTURE)
2787         ff_er_frame_end(s);
2788
2789     ff_MPV_frame_end(s);
2790
2791     h->current_slice = 0;
2792
2793     return err;
2794 }
2795
2796 /**
2797  * Replicate H264 "master" context to thread contexts.
2798  */
2799 static void clone_slice(H264Context *dst, H264Context *src)
2800 {
2801     memcpy(dst->block_offset, src->block_offset, sizeof(dst->block_offset));
2802     dst->s.current_picture_ptr = src->s.current_picture_ptr;
2803     dst->s.current_picture     = src->s.current_picture;
2804     dst->s.linesize            = src->s.linesize;
2805     dst->s.uvlinesize          = src->s.uvlinesize;
2806     dst->s.first_field         = src->s.first_field;
2807
2808     dst->prev_poc_msb          = src->prev_poc_msb;
2809     dst->prev_poc_lsb          = src->prev_poc_lsb;
2810     dst->prev_frame_num_offset = src->prev_frame_num_offset;
2811     dst->prev_frame_num        = src->prev_frame_num;
2812     dst->short_ref_count       = src->short_ref_count;
2813
2814     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
2815     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
2816     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2817     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
2818
2819     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
2820     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
2821 }
2822
2823 /**
2824  * Compute profile from profile_idc and constraint_set?_flags.
2825  *
2826  * @param sps SPS
2827  *
2828  * @return profile as defined by FF_PROFILE_H264_*
2829  */
2830 int ff_h264_get_profile(SPS *sps)
2831 {
2832     int profile = sps->profile_idc;
2833
2834     switch (sps->profile_idc) {
2835     case FF_PROFILE_H264_BASELINE:
2836         // constraint_set1_flag set to 1
2837         profile |= (sps->constraint_set_flags & 1 << 1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2838         break;
2839     case FF_PROFILE_H264_HIGH_10:
2840     case FF_PROFILE_H264_HIGH_422:
2841     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2842         // constraint_set3_flag set to 1
2843         profile |= (sps->constraint_set_flags & 1 << 3) ? FF_PROFILE_H264_INTRA : 0;
2844         break;
2845     }
2846
2847     return profile;
2848 }
2849
2850 /**
2851  * Decode a slice header.
2852  * This will also call ff_MPV_common_init() and frame_start() as needed.
2853  *
2854  * @param h h264context
2855  * @param h0 h264 master context (differs from 'h' when doing sliced based
2856  *           parallel decoding)
2857  *
2858  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2859  */
2860 static int decode_slice_header(H264Context *h, H264Context *h0)
2861 {
2862     MpegEncContext *const s  = &h->s;
2863     MpegEncContext *const s0 = &h0->s;
2864     unsigned int first_mb_in_slice;
2865     unsigned int pps_id;
2866     int num_ref_idx_active_override_flag;
2867     unsigned int slice_type, tmp, i, j;
2868     int default_ref_list_done = 0;
2869     int last_pic_structure, last_pic_dropable;
2870     int must_reinit;
2871
2872     /* FIXME: 2tap qpel isn't implemented for high bit depth. */
2873     if ((s->avctx->flags2 & CODEC_FLAG2_FAST) &&
2874         !h->nal_ref_idc && !h->pixel_shift) {
2875         s->me.qpel_put = s->dsp.put_2tap_qpel_pixels_tab;
2876         s->me.qpel_avg = s->dsp.avg_2tap_qpel_pixels_tab;
2877     } else {
2878         s->me.qpel_put = s->dsp.put_h264_qpel_pixels_tab;
2879         s->me.qpel_avg = s->dsp.avg_h264_qpel_pixels_tab;
2880     }
2881
2882     first_mb_in_slice = get_ue_golomb_long(&s->gb);
2883
2884     if (first_mb_in_slice == 0) { // FIXME better field boundary detection
2885         if (h0->current_slice && FIELD_PICTURE) {
2886             field_end(h, 1);
2887         }
2888
2889         h0->current_slice = 0;
2890         if (!s0->first_field) {
2891             if (s->current_picture_ptr && !s->dropable &&
2892                 s->current_picture_ptr->owner2 == s) {
2893                 ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX,
2894                                           s->picture_structure == PICT_BOTTOM_FIELD);
2895             }
2896             s->current_picture_ptr = NULL;
2897         }
2898     }
2899
2900     slice_type = get_ue_golomb_31(&s->gb);
2901     if (slice_type > 9) {
2902         av_log(h->s.avctx, AV_LOG_ERROR,
2903                "slice type too large (%d) at %d %d\n",
2904                h->slice_type, s->mb_x, s->mb_y);
2905         return -1;
2906     }
2907     if (slice_type > 4) {
2908         slice_type -= 5;
2909         h->slice_type_fixed = 1;
2910     } else
2911         h->slice_type_fixed = 0;
2912
2913     slice_type = golomb_to_pict_type[slice_type];
2914     if (slice_type == AV_PICTURE_TYPE_I ||
2915         (h0->current_slice != 0 && slice_type == h0->last_slice_type)) {
2916         default_ref_list_done = 1;
2917     }
2918     h->slice_type     = slice_type;
2919     h->slice_type_nos = slice_type & 3;
2920
2921     // to make a few old functions happy, it's wrong though
2922     s->pict_type = h->slice_type;
2923
2924     pps_id = get_ue_golomb(&s->gb);
2925     if (pps_id >= MAX_PPS_COUNT) {
2926         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id %d out of range\n", pps_id);
2927         return -1;
2928     }
2929     if (!h0->pps_buffers[pps_id]) {
2930         av_log(h->s.avctx, AV_LOG_ERROR,
2931                "non-existing PPS %u referenced\n",
2932                pps_id);
2933         return -1;
2934     }
2935     h->pps = *h0->pps_buffers[pps_id];
2936
2937     if (!h0->sps_buffers[h->pps.sps_id]) {
2938         av_log(h->s.avctx, AV_LOG_ERROR,
2939                "non-existing SPS %u referenced\n",
2940                h->pps.sps_id);
2941         return -1;
2942     }
2943     h->sps = *h0->sps_buffers[h->pps.sps_id];
2944
2945     s->avctx->profile = ff_h264_get_profile(&h->sps);
2946     s->avctx->level   = h->sps.level_idc;
2947     s->avctx->refs    = h->sps.ref_frame_count;
2948
2949     must_reinit = (s->context_initialized &&
2950                     (   16*h->sps.mb_width != s->avctx->coded_width
2951                      || 16*h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag) != s->avctx->coded_height
2952                      || s->avctx->bits_per_raw_sample != h->sps.bit_depth_luma
2953                      || h->cur_chroma_format_idc != h->sps.chroma_format_idc
2954                      || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio)));
2955
2956     if(must_reinit && (h != h0 || (s->avctx->active_thread_type & FF_THREAD_FRAME))) {
2957         av_log_missing_feature(s->avctx,
2958                                 "Width/height/bit depth/chroma idc changing with threads is", 0);
2959         return AVERROR_PATCHWELCOME;   // width / height changed during parallelized decoding
2960     }
2961
2962     s->mb_width  = h->sps.mb_width;
2963     s->mb_height = h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2964
2965     h->b_stride = s->mb_width * 4;
2966
2967     s->chroma_y_shift = h->sps.chroma_format_idc <= 1; // 400 uses yuv420p
2968
2969     s->width  = 16 * s->mb_width;
2970     s->height = 16 * s->mb_height;
2971
2972     if(must_reinit) {
2973         free_tables(h, 0);
2974         flush_dpb(s->avctx);
2975         ff_MPV_common_end(s);
2976         h->list_count = 0;
2977         h->current_slice = 0;
2978     }
2979     if (!s->context_initialized) {
2980         if (h != h0) {
2981             av_log(h->s.avctx, AV_LOG_ERROR,
2982                    "Cannot (re-)initialize context during parallel decoding.\n");
2983             return -1;
2984         }
2985         if(   FFALIGN(s->avctx->width , 16                                 ) == s->width
2986            && FFALIGN(s->avctx->height, 16*(2 - h->sps.frame_mbs_only_flag)) == s->height
2987            && !h->sps.crop_right && !h->sps.crop_bottom
2988            && (s->avctx->width != s->width || s->avctx->height && s->height)
2989         ) {
2990             av_log(h->s.avctx, AV_LOG_DEBUG, "Using externally provided dimensions\n");
2991             s->avctx->coded_width  = s->width;
2992             s->avctx->coded_height = s->height;
2993         } else{
2994             avcodec_set_dimensions(s->avctx, s->width, s->height);
2995             s->avctx->width  -= (2>>CHROMA444)*FFMIN(h->sps.crop_right, (8<<CHROMA444)-1);
2996             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);
2997         }
2998         s->avctx->sample_aspect_ratio = h->sps.sar;
2999         av_assert0(s->avctx->sample_aspect_ratio.den);
3000
3001         if (s->avctx->bits_per_raw_sample != h->sps.bit_depth_luma ||
3002             h->cur_chroma_format_idc != h->sps.chroma_format_idc) {
3003             if (h->sps.bit_depth_luma >= 8 && h->sps.bit_depth_luma <= 10 &&
3004                 (h->sps.bit_depth_luma != 9 || !CHROMA422)) {
3005                 s->avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
3006                 h->cur_chroma_format_idc = h->sps.chroma_format_idc;
3007                 h->pixel_shift = h->sps.bit_depth_luma > 8;
3008
3009                 ff_h264dsp_init(&h->h264dsp, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
3010                 ff_h264_pred_init(&h->hpc, s->codec_id, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
3011                 s->dsp.dct_bits = h->sps.bit_depth_luma > 8 ? 32 : 16;
3012                 ff_dsputil_init(&s->dsp, s->avctx);
3013             } else {
3014                 av_log(s->avctx, AV_LOG_ERROR, "Unsupported bit depth: %d chroma_idc: %d\n",
3015                        h->sps.bit_depth_luma, h->sps.chroma_format_idc);
3016                 return -1;
3017             }
3018         }
3019
3020         if (h->sps.video_signal_type_present_flag) {
3021             s->avctx->color_range = h->sps.full_range>0 ? AVCOL_RANGE_JPEG
3022                                                       : AVCOL_RANGE_MPEG;
3023             if (h->sps.colour_description_present_flag) {
3024                 s->avctx->color_primaries = h->sps.color_primaries;
3025                 s->avctx->color_trc       = h->sps.color_trc;
3026                 s->avctx->colorspace      = h->sps.colorspace;
3027             }
3028         }
3029
3030         if (h->sps.timing_info_present_flag) {
3031             int64_t den = h->sps.time_scale;
3032             if (h->x264_build < 44U)
3033                 den *= 2;
3034             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
3035                       h->sps.num_units_in_tick, den, 1 << 30);
3036         }
3037
3038         switch (h->sps.bit_depth_luma) {
3039         case 9:
3040             if (CHROMA444) {
3041                 if (s->avctx->colorspace == AVCOL_SPC_RGB) {
3042                     s->avctx->pix_fmt = PIX_FMT_GBRP9;
3043                 } else
3044                     s->avctx->pix_fmt = PIX_FMT_YUV444P9;
3045             } else if (CHROMA422)
3046                 s->avctx->pix_fmt = PIX_FMT_YUV422P9;
3047             else
3048                 s->avctx->pix_fmt = PIX_FMT_YUV420P9;
3049             break;
3050         case 10:
3051             if (CHROMA444) {
3052                 if (s->avctx->colorspace == AVCOL_SPC_RGB) {
3053                     s->avctx->pix_fmt = PIX_FMT_GBRP10;
3054                 } else
3055                     s->avctx->pix_fmt = PIX_FMT_YUV444P10;
3056             } else if (CHROMA422)
3057                 s->avctx->pix_fmt = PIX_FMT_YUV422P10;
3058             else
3059                 s->avctx->pix_fmt = PIX_FMT_YUV420P10;
3060             break;
3061         case 8:
3062             if (CHROMA444) {
3063                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ444P
3064                                                                                   : PIX_FMT_YUV444P;
3065                     if (s->avctx->colorspace == AVCOL_SPC_RGB) {
3066                         s->avctx->pix_fmt = PIX_FMT_GBR24P;
3067                         av_log(h->s.avctx, AV_LOG_DEBUG, "Detected GBR colorspace.\n");
3068                     } else if (s->avctx->colorspace == AVCOL_SPC_YCGCO) {
3069                         av_log(h->s.avctx, AV_LOG_WARNING, "Detected unsupported YCgCo colorspace.\n");
3070                     }
3071             } else if (CHROMA422) {
3072                 s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ422P
3073                                                                               : PIX_FMT_YUV422P;
3074             } else {
3075                 s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
3076                                                          s->avctx->codec->pix_fmts ?
3077                                                          s->avctx->codec->pix_fmts :
3078                                                          s->avctx->color_range == AVCOL_RANGE_JPEG ?
3079                                                          hwaccel_pixfmt_list_h264_jpeg_420 :
3080                                                          ff_hwaccel_pixfmt_list_420);
3081             }
3082             break;
3083         default:
3084             av_log(s->avctx, AV_LOG_ERROR,
3085                    "Unsupported bit depth: %d\n", h->sps.bit_depth_luma);
3086             return AVERROR_INVALIDDATA;
3087         }
3088
3089         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id,
3090                                             s->avctx->pix_fmt);
3091
3092         if (ff_MPV_common_init(s) < 0) {
3093             av_log(h->s.avctx, AV_LOG_ERROR, "ff_MPV_common_init() failed.\n");
3094             return -1;
3095         }
3096         s->first_field = 0;
3097         h->prev_interlaced_frame = 1;
3098
3099         init_scan_tables(h);
3100         if (ff_h264_alloc_tables(h) < 0) {
3101             av_log(h->s.avctx, AV_LOG_ERROR,
3102                    "Could not allocate memory for h264\n");
3103             return AVERROR(ENOMEM);
3104         }
3105
3106         if (!HAVE_THREADS || !(s->avctx->active_thread_type & FF_THREAD_SLICE)) {
3107             if (context_init(h) < 0) {
3108                 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
3109                 return -1;
3110             }
3111         } else {
3112             for (i = 1; i < s->slice_context_count; i++) {
3113                 H264Context *c;
3114                 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
3115                 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
3116                 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
3117                 c->h264dsp     = h->h264dsp;
3118                 c->sps         = h->sps;
3119                 c->pps         = h->pps;
3120                 c->pixel_shift = h->pixel_shift;
3121                 c->cur_chroma_format_idc = h->cur_chroma_format_idc;
3122                 init_scan_tables(c);
3123                 clone_tables(c, h, i);
3124             }
3125
3126             for (i = 0; i < s->slice_context_count; i++)
3127                 if (context_init(h->thread_context[i]) < 0) {
3128                     av_log(h->s.avctx, AV_LOG_ERROR,
3129                            "context_init() failed.\n");
3130                     return -1;
3131                 }
3132         }
3133     }
3134
3135     if (h == h0 && h->dequant_coeff_pps != pps_id) {
3136         h->dequant_coeff_pps = pps_id;
3137         init_dequant_tables(h);
3138     }
3139
3140     h->frame_num = get_bits(&s->gb, h->sps.log2_max_frame_num);
3141
3142     h->mb_mbaff        = 0;
3143     h->mb_aff_frame    = 0;
3144     last_pic_structure = s0->picture_structure;
3145     last_pic_dropable  = s->dropable;
3146     s->dropable        = h->nal_ref_idc == 0;
3147     if (h->sps.frame_mbs_only_flag) {
3148         s->picture_structure = PICT_FRAME;
3149     } else {
3150         if (!h->sps.direct_8x8_inference_flag && slice_type == AV_PICTURE_TYPE_B) {
3151             av_log(h->s.avctx, AV_LOG_ERROR, "This stream was generated by a broken encoder, invalid 8x8 inference\n");
3152             return -1;
3153         }
3154         if (get_bits1(&s->gb)) { // field_pic_flag
3155             s->picture_structure = PICT_TOP_FIELD + get_bits1(&s->gb); // bottom_field_flag
3156         } else {
3157             s->picture_structure = PICT_FRAME;
3158             h->mb_aff_frame      = h->sps.mb_aff;
3159         }
3160     }
3161     h->mb_field_decoding_flag = s->picture_structure != PICT_FRAME;
3162
3163     if (h0->current_slice != 0) {
3164         if (last_pic_structure != s->picture_structure ||
3165             last_pic_dropable  != s->dropable) {
3166             av_log(h->s.avctx, AV_LOG_ERROR,
3167                    "Changing field mode (%d -> %d) between slices is not allowed\n",
3168                    last_pic_structure, s->picture_structure);
3169             s->picture_structure = last_pic_structure;
3170             s->dropable          = last_pic_dropable;
3171             return AVERROR_INVALIDDATA;
3172         }
3173     } else {
3174         /* Shorten frame num gaps so we don't have to allocate reference
3175          * frames just to throw them away */
3176         if (h->frame_num != h->prev_frame_num && h->prev_frame_num >= 0) {
3177             int unwrap_prev_frame_num = h->prev_frame_num;
3178             int max_frame_num         = 1 << h->sps.log2_max_frame_num;
3179
3180             if (unwrap_prev_frame_num > h->frame_num)
3181                 unwrap_prev_frame_num -= max_frame_num;
3182
3183             if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
3184                 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
3185                 if (unwrap_prev_frame_num < 0)
3186                     unwrap_prev_frame_num += max_frame_num;
3187
3188                 h->prev_frame_num = unwrap_prev_frame_num;
3189             }
3190         }
3191
3192         /* See if we have a decoded first field looking for a pair...
3193          * Here, we're using that to see if we should mark previously
3194          * decode frames as "finished".
3195          * We have to do that before the "dummy" in-between frame allocation,
3196          * since that can modify s->current_picture_ptr. */
3197         if (s0->first_field) {
3198             assert(s0->current_picture_ptr);
3199             assert(s0->current_picture_ptr->f.data[0]);
3200             assert(s0->current_picture_ptr->f.reference != DELAYED_PIC_REF);
3201
3202             /* Mark old field/frame as completed */
3203             if (!last_pic_dropable && s0->current_picture_ptr->owner2 == s0) {
3204                 ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
3205                                           last_pic_structure == PICT_BOTTOM_FIELD);
3206             }
3207
3208             /* figure out if we have a complementary field pair */
3209             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
3210                 /* Previous field is unmatched. Don't display it, but let it
3211                  * remain for reference if marked as such. */
3212                 if (!last_pic_dropable && last_pic_structure != PICT_FRAME) {
3213                     ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
3214                                               last_pic_structure == PICT_TOP_FIELD);
3215                 }
3216             } else {
3217                 if (s0->current_picture_ptr->frame_num != h->frame_num) {
3218                     /* This and previous field were reference, but had
3219                      * different frame_nums. Consider this field first in
3220                      * pair. Throw away previous field except for reference
3221                      * purposes. */
3222                     if (!last_pic_dropable && last_pic_structure != PICT_FRAME) {
3223                         ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
3224                                                   last_pic_structure == PICT_TOP_FIELD);
3225                     }
3226                 } else {
3227                     /* Second field in complementary pair */
3228                     if (!((last_pic_structure   == PICT_TOP_FIELD &&
3229                            s->picture_structure == PICT_BOTTOM_FIELD) ||
3230                           (last_pic_structure   == PICT_BOTTOM_FIELD &&
3231                            s->picture_structure == PICT_TOP_FIELD))) {
3232                         av_log(s->avctx, AV_LOG_ERROR,
3233                                "Invalid field mode combination %d/%d\n",
3234                                last_pic_structure, s->picture_structure);
3235                         s->picture_structure = last_pic_structure;
3236                         s->dropable          = last_pic_dropable;
3237                         return AVERROR_INVALIDDATA;
3238                     } else if (last_pic_dropable != s->dropable) {
3239                         av_log(s->avctx, AV_LOG_ERROR,
3240                                "Cannot combine reference and non-reference fields in the same frame\n");
3241                         av_log_ask_for_sample(s->avctx, NULL);
3242                         s->picture_structure = last_pic_structure;
3243                         s->dropable          = last_pic_dropable;
3244                         return AVERROR_INVALIDDATA;
3245                     }
3246
3247                     /* Take ownership of this buffer. Note that if another thread owned
3248                      * the first field of this buffer, we're not operating on that pointer,
3249                      * so the original thread is still responsible for reporting progress
3250                      * on that first field (or if that was us, we just did that above).
3251                      * By taking ownership, we assign responsibility to ourselves to
3252                      * report progress on the second field. */
3253                     s0->current_picture_ptr->owner2 = s0;
3254                 }
3255             }
3256         }
3257
3258         while (h->frame_num != h->prev_frame_num && h->prev_frame_num >= 0 &&
3259                h->frame_num != (h->prev_frame_num + 1) % (1 << h->sps.log2_max_frame_num)) {
3260             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
3261             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n",
3262                    h->frame_num, h->prev_frame_num);
3263             if (ff_h264_frame_start(h) < 0)
3264                 return -1;
3265             h->prev_frame_num++;
3266             h->prev_frame_num %= 1 << h->sps.log2_max_frame_num;
3267             s->current_picture_ptr->frame_num = h->prev_frame_num;
3268             ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX, 0);
3269             ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX, 1);
3270             ff_generate_sliding_window_mmcos(h);
3271             if (ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index) < 0 &&
3272                 (s->avctx->err_recognition & AV_EF_EXPLODE))
3273                 return AVERROR_INVALIDDATA;
3274             /* Error concealment: if a ref is missing, copy the previous ref in its place.
3275              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
3276              * about there being no actual duplicates.
3277              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
3278              * concealing a lost frame, this probably isn't noticeable by comparison, but it should
3279              * be fixed. */
3280             if (h->short_ref_count) {
3281                 if (prev) {
3282                     av_image_copy(h->short_ref[0]->f.data, h->short_ref[0]->f.linesize,
3283                                   (const uint8_t **)prev->f.data, prev->f.linesize,
3284                                   s->avctx->pix_fmt, s->mb_width * 16, s->mb_height * 16);
3285                     h->short_ref[0]->poc = prev->poc + 2;
3286                 }
3287                 h->short_ref[0]->frame_num = h->prev_frame_num;
3288             }
3289         }
3290
3291         /* See if we have a decoded first field looking for a pair...
3292          * We're using that to see whether to continue decoding in that
3293          * frame, or to allocate a new one. */
3294         if (s0->first_field) {
3295             assert(s0->current_picture_ptr);
3296             assert(s0->current_picture_ptr->f.data[0]);
3297             assert(s0->current_picture_ptr->f.reference != DELAYED_PIC_REF);
3298
3299             /* figure out if we have a complementary field pair */
3300             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
3301                 /* Previous field is unmatched. Don't display it, but let it
3302                  * remain for reference if marked as such. */
3303                 s0->current_picture_ptr = NULL;
3304                 s0->first_field         = FIELD_PICTURE;
3305             } else {
3306                 if (s0->current_picture_ptr->frame_num != h->frame_num) {
3307                     ff_thread_report_progress((AVFrame*)s0->current_picture_ptr, INT_MAX,
3308                                               s0->picture_structure==PICT_BOTTOM_FIELD);
3309                     /* This and the previous field had different frame_nums.
3310                      * Consider this field first in pair. Throw away previous
3311                      * one except for reference purposes. */
3312                     s0->first_field         = 1;
3313                     s0->current_picture_ptr = NULL;
3314                 } else {
3315                     /* Second field in complementary pair */
3316                     s0->first_field = 0;
3317                 }
3318             }
3319         } else {
3320             /* Frame or first field in a potentially complementary pair */
3321             assert(!s0->current_picture_ptr);
3322             s0->first_field = FIELD_PICTURE;
3323         }
3324
3325         if (!FIELD_PICTURE || s0->first_field) {
3326             if (ff_h264_frame_start(h) < 0) {
3327                 s0->first_field = 0;
3328                 return -1;
3329             }
3330         } else {
3331             ff_release_unused_pictures(s, 0);
3332         }
3333     }
3334     if (h != h0)
3335         clone_slice(h, h0);
3336
3337     s->current_picture_ptr->frame_num = h->frame_num; // FIXME frame_num cleanup
3338
3339     assert(s->mb_num == s->mb_width * s->mb_height);
3340     if (first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
3341         first_mb_in_slice >= s->mb_num) {
3342         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
3343         return -1;
3344     }
3345     s->resync_mb_x = s->mb_x =  first_mb_in_slice % s->mb_width;
3346     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
3347     if (s->picture_structure == PICT_BOTTOM_FIELD)
3348         s->resync_mb_y = s->mb_y = s->mb_y + 1;
3349     assert(s->mb_y < s->mb_height);
3350
3351     if (s->picture_structure == PICT_FRAME) {
3352         h->curr_pic_num = h->frame_num;
3353         h->max_pic_num  = 1 << h->sps.log2_max_frame_num;
3354     } else {
3355         h->curr_pic_num = 2 * h->frame_num + 1;
3356         h->max_pic_num  = 1 << (h->sps.log2_max_frame_num + 1);
3357     }
3358
3359     if (h->nal_unit_type == NAL_IDR_SLICE)
3360         get_ue_golomb(&s->gb); /* idr_pic_id */
3361
3362     if (h->sps.poc_type == 0) {
3363         h->poc_lsb = get_bits(&s->gb, h->sps.log2_max_poc_lsb);
3364
3365         if (h->pps.pic_order_present == 1 && s->picture_structure == PICT_FRAME)
3366             h->delta_poc_bottom = get_se_golomb(&s->gb);
3367     }
3368
3369     if (h->sps.poc_type == 1 && !h->sps.delta_pic_order_always_zero_flag) {
3370         h->delta_poc[0] = get_se_golomb(&s->gb);
3371
3372         if (h->pps.pic_order_present == 1 && s->picture_structure == PICT_FRAME)
3373             h->delta_poc[1] = get_se_golomb(&s->gb);
3374     }
3375
3376     init_poc(h);
3377
3378     if (h->pps.redundant_pic_cnt_present)
3379         h->redundant_pic_count = get_ue_golomb(&s->gb);
3380
3381     // set defaults, might be overridden a few lines later
3382     h->ref_count[0] = h->pps.ref_count[0];
3383     h->ref_count[1] = h->pps.ref_count[1];
3384
3385     if (h->slice_type_nos != AV_PICTURE_TYPE_I) {
3386         unsigned max = s->picture_structure == PICT_FRAME ? 15 : 31;
3387
3388         if (h->slice_type_nos == AV_PICTURE_TYPE_B)
3389             h->direct_spatial_mv_pred = get_bits1(&s->gb);
3390         num_ref_idx_active_override_flag = get_bits1(&s->gb);
3391
3392         if (num_ref_idx_active_override_flag) {
3393             h->ref_count[0] = get_ue_golomb(&s->gb) + 1;
3394             if (h->slice_type_nos == AV_PICTURE_TYPE_B)
3395                 h->ref_count[1] = get_ue_golomb(&s->gb) + 1;
3396         }
3397
3398         if (h->ref_count[0]-1 > max || h->ref_count[1]-1 > max){
3399             av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
3400             h->ref_count[0] = h->ref_count[1] = 1;
3401             return AVERROR_INVALIDDATA;
3402         }
3403
3404         if (h->slice_type_nos == AV_PICTURE_TYPE_B)
3405             h->list_count = 2;
3406         else
3407             h->list_count = 1;
3408     } else
3409         h->ref_count[1]= h->ref_count[0]= h->list_count= 0;
3410
3411     if (!default_ref_list_done)
3412         ff_h264_fill_default_ref_list(h);
3413
3414     if (h->slice_type_nos != AV_PICTURE_TYPE_I &&
3415         ff_h264_decode_ref_pic_list_reordering(h) < 0) {
3416         h->ref_count[1] = h->ref_count[0] = 0;
3417         return -1;
3418     }
3419
3420     if (h->slice_type_nos != AV_PICTURE_TYPE_I) {
3421         s->last_picture_ptr = &h->ref_list[0][0];
3422         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
3423     }
3424     if (h->slice_type_nos == AV_PICTURE_TYPE_B) {
3425         s->next_picture_ptr = &h->ref_list[1][0];
3426         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
3427     }
3428
3429     if ((h->pps.weighted_pred && h->slice_type_nos == AV_PICTURE_TYPE_P) ||
3430         (h->pps.weighted_bipred_idc == 1 &&
3431          h->slice_type_nos == AV_PICTURE_TYPE_B))
3432         pred_weight_table(h);
3433     else if (h->pps.weighted_bipred_idc == 2 &&
3434              h->slice_type_nos == AV_PICTURE_TYPE_B) {
3435         implicit_weight_table(h, -1);
3436     } else {
3437         h->use_weight = 0;
3438         for (i = 0; i < 2; i++) {
3439             h->luma_weight_flag[i]   = 0;
3440             h->chroma_weight_flag[i] = 0;
3441         }
3442     }
3443
3444     if (h->nal_ref_idc && ff_h264_decode_ref_pic_marking(h0, &s->gb) < 0 &&
3445         (s->avctx->err_recognition & AV_EF_EXPLODE))
3446         return AVERROR_INVALIDDATA;
3447
3448     if (FRAME_MBAFF) {
3449         ff_h264_fill_mbaff_ref_list(h);
3450
3451         if (h->pps.weighted_bipred_idc == 2 && h->slice_type_nos == AV_PICTURE_TYPE_B) {
3452             implicit_weight_table(h, 0);
3453             implicit_weight_table(h, 1);
3454         }
3455     }
3456
3457     if (h->slice_type_nos == AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
3458         ff_h264_direct_dist_scale_factor(h);
3459     ff_h264_direct_ref_list_init(h);
3460
3461     if (h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac) {
3462         tmp = get_ue_golomb_31(&s->gb);
3463         if (tmp > 2) {
3464             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
3465             return -1;
3466         }
3467         h->cabac_init_idc = tmp;
3468     }
3469
3470     h->last_qscale_diff = 0;
3471     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
3472     if (tmp > 51 + 6 * (h->sps.bit_depth_luma - 8)) {
3473         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
3474         return -1;
3475     }
3476     s->qscale       = tmp;
3477     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3478     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3479     // FIXME qscale / qp ... stuff
3480     if (h->slice_type == AV_PICTURE_TYPE_SP)
3481         get_bits1(&s->gb); /* sp_for_switch_flag */
3482     if (h->slice_type == AV_PICTURE_TYPE_SP ||
3483         h->slice_type == AV_PICTURE_TYPE_SI)
3484         get_se_golomb(&s->gb); /* slice_qs_delta */
3485
3486     h->deblocking_filter     = 1;
3487     h->slice_alpha_c0_offset = 52;
3488     h->slice_beta_offset     = 52;
3489     if (h->pps.deblocking_filter_parameters_present) {
3490         tmp = get_ue_golomb_31(&s->gb);
3491         if (tmp > 2) {
3492             av_log(s->avctx, AV_LOG_ERROR,
3493                    "deblocking_filter_idc %u out of range\n", tmp);
3494             return -1;
3495         }
3496         h->deblocking_filter = tmp;
3497         if (h->deblocking_filter < 2)
3498             h->deblocking_filter ^= 1;  // 1<->0
3499
3500         if (h->deblocking_filter) {
3501             h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
3502             h->slice_beta_offset     += get_se_golomb(&s->gb) << 1;
3503             if (h->slice_alpha_c0_offset > 104U ||
3504                 h->slice_beta_offset     > 104U) {
3505                 av_log(s->avctx, AV_LOG_ERROR,
3506                        "deblocking filter parameters %d %d out of range\n",
3507                        h->slice_alpha_c0_offset, h->slice_beta_offset);
3508                 return -1;
3509             }
3510         }
3511     }
3512
3513     if (s->avctx->skip_loop_filter >= AVDISCARD_ALL ||
3514         (s->avctx->skip_loop_filter >= AVDISCARD_NONKEY &&
3515          h->slice_type_nos != AV_PICTURE_TYPE_I) ||
3516         (s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  &&
3517          h->slice_type_nos == AV_PICTURE_TYPE_B) ||
3518         (s->avctx->skip_loop_filter >= AVDISCARD_NONREF &&
3519          h->nal_ref_idc == 0))
3520         h->deblocking_filter = 0;
3521
3522     if (h->deblocking_filter == 1 && h0->max_contexts > 1) {
3523         if (s->avctx->flags2 & CODEC_FLAG2_FAST) {
3524             /* Cheat slightly for speed:
3525              * Do not bother to deblock across slices. */
3526             h->deblocking_filter = 2;
3527         } else {
3528             h0->max_contexts = 1;
3529             if (!h0->single_decode_warning) {
3530                 av_log(s->avctx, AV_LOG_INFO,
3531                        "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
3532                 h0->single_decode_warning = 1;
3533             }
3534             if (h != h0) {
3535                 av_log(h->s.avctx, AV_LOG_ERROR,
3536                        "Deblocking switched inside frame.\n");
3537                 return 1;
3538             }
3539         }
3540     }
3541     h->qp_thresh = 15 + 52 -
3542                    FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset) -
3543                    FFMAX3(0,
3544                           h->pps.chroma_qp_index_offset[0],
3545                           h->pps.chroma_qp_index_offset[1]) +
3546                    6 * (h->sps.bit_depth_luma - 8);
3547
3548     h0->last_slice_type = slice_type;
3549     h->slice_num = ++h0->current_slice;
3550
3551     if (h->slice_num)
3552         h0->slice_row[(h->slice_num-1)&(MAX_SLICES-1)]= s->resync_mb_y;
3553     if (   h0->slice_row[h->slice_num&(MAX_SLICES-1)] + 3 >= s->resync_mb_y
3554         && h0->slice_row[h->slice_num&(MAX_SLICES-1)] <= s->resync_mb_y
3555         && h->slice_num >= MAX_SLICES) {
3556         //in case of ASO this check needs to be updated depending on how we decide to assign slice numbers in this case
3557         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);
3558     }
3559
3560     for (j = 0; j < 2; j++) {
3561         int id_list[16];
3562         int *ref2frm = h->ref2frm[h->slice_num & (MAX_SLICES - 1)][j];
3563         for (i = 0; i < 16; i++) {
3564             id_list[i] = 60;
3565             if (h->ref_list[j][i].f.data[0]) {
3566                 int k;
3567                 uint8_t *base = h->ref_list[j][i].f.base[0];
3568                 for (k = 0; k < h->short_ref_count; k++)
3569                     if (h->short_ref[k]->f.base[0] == base) {
3570                         id_list[i] = k;
3571                         break;
3572                     }
3573                 for (k = 0; k < h->long_ref_count; k++)
3574                     if (h->long_ref[k] && h->long_ref[k]->f.base[0] == base) {
3575                         id_list[i] = h->short_ref_count + k;
3576                         break;
3577                     }
3578             }
3579         }
3580
3581         ref2frm[0]     =
3582             ref2frm[1] = -1;
3583         for (i = 0; i < 16; i++)
3584             ref2frm[i + 2] = 4 * id_list[i] +
3585                              (h->ref_list[j][i].f.reference & 3);
3586         ref2frm[18 + 0]     =
3587             ref2frm[18 + 1] = -1;
3588         for (i = 16; i < 48; i++)
3589             ref2frm[i + 4] = 4 * id_list[(i - 16) >> 1] +
3590                              (h->ref_list[j][i].f.reference & 3);
3591     }
3592
3593     // FIXME: fix draw_edges + PAFF + frame threads
3594     h->emu_edge_width  = (s->flags & CODEC_FLAG_EMU_EDGE ||
3595                           (!h->sps.frame_mbs_only_flag &&
3596                            s->avctx->active_thread_type))
3597                          ? 0 : 16;
3598     h->emu_edge_height = (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
3599
3600     if (s->avctx->debug & FF_DEBUG_PICT_INFO) {
3601         av_log(h->s.avctx, AV_LOG_DEBUG,
3602                "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",
3603                h->slice_num,
3604                (s->picture_structure == PICT_FRAME ? "F" : s->picture_structure == PICT_TOP_FIELD ? "T" : "B"),
3605                first_mb_in_slice,
3606                av_get_picture_type_char(h->slice_type),
3607                h->slice_type_fixed ? " fix" : "",
3608                h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
3609                pps_id, h->frame_num,
3610                s->current_picture_ptr->field_poc[0],
3611                s->current_picture_ptr->field_poc[1],
3612                h->ref_count[0], h->ref_count[1],
3613                s->qscale,
3614                h->deblocking_filter,
3615                h->slice_alpha_c0_offset / 2 - 26, h->slice_beta_offset / 2 - 26,
3616                h->use_weight,
3617                h->use_weight == 1 && h->use_weight_chroma ? "c" : "",
3618                h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : "");
3619     }
3620
3621     return 0;
3622 }
3623
3624 int ff_h264_get_slice_type(const H264Context *h)
3625 {
3626     switch (h->slice_type) {
3627     case AV_PICTURE_TYPE_P:
3628         return 0;
3629     case AV_PICTURE_TYPE_B:
3630         return 1;
3631     case AV_PICTURE_TYPE_I:
3632         return 2;
3633     case AV_PICTURE_TYPE_SP:
3634         return 3;
3635     case AV_PICTURE_TYPE_SI:
3636         return 4;
3637     default:
3638         return -1;
3639     }
3640 }
3641
3642 static av_always_inline void fill_filter_caches_inter(H264Context *h,
3643                                                       MpegEncContext *const s,
3644                                                       int mb_type, int top_xy,
3645                                                       int left_xy[LEFT_MBS],
3646                                                       int top_type,
3647                                                       int left_type[LEFT_MBS],
3648                                                       int mb_xy, int list)
3649 {
3650     int b_stride = h->b_stride;
3651     int16_t(*mv_dst)[2] = &h->mv_cache[list][scan8[0]];
3652     int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
3653     if (IS_INTER(mb_type) || IS_DIRECT(mb_type)) {
3654         if (USES_LIST(top_type, list)) {
3655             const int b_xy  = h->mb2b_xy[top_xy] + 3 * b_stride;
3656             const int b8_xy = 4 * top_xy + 2;
3657             int (*ref2frm)[64] = (void*)(h->ref2frm[h->slice_table[top_xy] & (MAX_SLICES - 1)][0] + (MB_MBAFF ? 20 : 2));
3658             AV_COPY128(mv_dst - 1 * 8, s->current_picture.f.motion_val[list][b_xy + 0]);
3659             ref_cache[0 - 1 * 8] =
3660             ref_cache[1 - 1 * 8] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 0]];
3661             ref_cache[2 - 1 * 8] =
3662             ref_cache[3 - 1 * 8] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 1]];
3663         } else {
3664             AV_ZERO128(mv_dst - 1 * 8);
3665             AV_WN32A(&ref_cache[0 - 1 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3666         }
3667
3668         if (!IS_INTERLACED(mb_type ^ left_type[LTOP])) {
3669             if (USES_LIST(left_type[LTOP], list)) {
3670                 const int b_xy  = h->mb2b_xy[left_xy[LTOP]] + 3;
3671                 const int b8_xy = 4 * left_xy[LTOP] + 1;
3672                 int (*ref2frm)[64] =(void*)( h->ref2frm[h->slice_table[left_xy[LTOP]] & (MAX_SLICES - 1)][0] + (MB_MBAFF ? 20 : 2));
3673                 AV_COPY32(mv_dst - 1 +  0, s->current_picture.f.motion_val[list][b_xy + b_stride * 0]);
3674                 AV_COPY32(mv_dst - 1 +  8, s->current_picture.f.motion_val[list][b_xy + b_stride * 1]);
3675                 AV_COPY32(mv_dst - 1 + 16, s->current_picture.f.motion_val[list][b_xy + b_stride * 2]);
3676                 AV_COPY32(mv_dst - 1 + 24, s->current_picture.f.motion_val[list][b_xy + b_stride * 3]);
3677                 ref_cache[-1 +  0] =
3678                 ref_cache[-1 +  8] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2 * 0]];
3679                 ref_cache[-1 + 16] =
3680                 ref_cache[-1 + 24] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2 * 1]];
3681             } else {
3682                 AV_ZERO32(mv_dst - 1 +  0);
3683                 AV_ZERO32(mv_dst - 1 +  8);
3684                 AV_ZERO32(mv_dst - 1 + 16);
3685                 AV_ZERO32(mv_dst - 1 + 24);
3686                 ref_cache[-1 +  0] =
3687                 ref_cache[-1 +  8] =
3688                 ref_cache[-1 + 16] =
3689                 ref_cache[-1 + 24] = LIST_NOT_USED;
3690             }
3691         }
3692     }
3693
3694     if (!USES_LIST(mb_type, list)) {
3695         fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0, 0), 4);
3696         AV_WN32A(&ref_cache[0 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3697         AV_WN32A(&ref_cache[1 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3698         AV_WN32A(&ref_cache[2 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3699         AV_WN32A(&ref_cache[3 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3700         return;
3701     }
3702
3703     {
3704         int8_t *ref = &s->current_picture.f.ref_index[list][4 * mb_xy];
3705         int (*ref2frm)[64] = (void*)(h->ref2frm[h->slice_num & (MAX_SLICES - 1)][0] + (MB_MBAFF ? 20 : 2));
3706         uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]], ref2frm[list][ref[1]]) & 0x00FF00FF) * 0x0101;
3707         uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]], ref2frm[list][ref[3]]) & 0x00FF00FF) * 0x0101;
3708         AV_WN32A(&ref_cache[0 * 8], ref01);
3709         AV_WN32A(&ref_cache[1 * 8], ref01);
3710         AV_WN32A(&ref_cache[2 * 8], ref23);
3711         AV_WN32A(&ref_cache[3 * 8], ref23);
3712     }
3713
3714     {
3715         int16_t(*mv_src)[2] = &s->current_picture.f.motion_val[list][4 * s->mb_x + 4 * s->mb_y * b_stride];
3716         AV_COPY128(mv_dst + 8 * 0, mv_src + 0 * b_stride);
3717         AV_COPY128(mv_dst + 8 * 1, mv_src + 1 * b_stride);
3718         AV_COPY128(mv_dst + 8 * 2, mv_src + 2 * b_stride);
3719         AV_COPY128(mv_dst + 8 * 3, mv_src + 3 * b_stride);
3720     }
3721 }
3722
3723 /**
3724  *
3725  * @return non zero if the loop filter can be skipped
3726  */
3727 static int fill_filter_caches(H264Context *h, int mb_type)
3728 {
3729     MpegEncContext *const s = &h->s;
3730     const int mb_xy = h->mb_xy;
3731     int top_xy, left_xy[LEFT_MBS];
3732     int top_type, left_type[LEFT_MBS];
3733     uint8_t *nnz;
3734     uint8_t *nnz_cache;
3735
3736     top_xy = mb_xy - (s->mb_stride << MB_FIELD);
3737
3738     /* Wow, what a mess, why didn't they simplify the interlacing & intra
3739      * stuff, I can't imagine that these complex rules are worth it. */
3740
3741     left_xy[LBOT] = left_xy[LTOP] = mb_xy - 1;
3742     if (FRAME_MBAFF) {
3743         const int left_mb_field_flag = IS_INTERLACED(s->current_picture.f.mb_type[mb_xy - 1]);
3744         const int curr_mb_field_flag = IS_INTERLACED(mb_type);
3745         if (s->mb_y & 1) {
3746             if (left_mb_field_flag != curr_mb_field_flag)
3747                 left_xy[LTOP] -= s->mb_stride;
3748         } else {
3749             if (curr_mb_field_flag)
3750                 top_xy += s->mb_stride &
3751                     (((s->current_picture.f.mb_type[top_xy] >> 7) & 1) - 1);
3752             if (left_mb_field_flag != curr_mb_field_flag)
3753                 left_xy[LBOT] += s->mb_stride;
3754         }
3755     }
3756
3757     h->top_mb_xy        = top_xy;
3758     h->left_mb_xy[LTOP] = left_xy[LTOP];
3759     h->left_mb_xy[LBOT] = left_xy[LBOT];
3760     {
3761         /* For sufficiently low qp, filtering wouldn't do anything.
3762          * This is a conservative estimate: could also check beta_offset
3763          * and more accurate chroma_qp. */
3764         int qp_thresh = h->qp_thresh; // FIXME strictly we should store qp_thresh for each mb of a slice
3765         int qp        = s->current_picture.f.qscale_table[mb_xy];
3766         if (qp <= qp_thresh &&
3767             (left_xy[LTOP] < 0 ||
3768              ((qp + s->current_picture.f.qscale_table[left_xy[LTOP]] + 1) >> 1) <= qp_thresh) &&
3769             (top_xy < 0 ||
3770              ((qp + s->current_picture.f.qscale_table[top_xy] + 1) >> 1) <= qp_thresh)) {
3771             if (!FRAME_MBAFF)
3772                 return 1;
3773             if ((left_xy[LTOP] < 0 ||
3774                  ((qp + s->current_picture.f.qscale_table[left_xy[LBOT]] + 1) >> 1) <= qp_thresh) &&
3775                 (top_xy < s->mb_stride ||
3776                  ((qp + s->current_picture.f.qscale_table[top_xy - s->mb_stride] + 1) >> 1) <= qp_thresh))
3777                 return 1;
3778         }
3779     }
3780
3781     top_type        = s->current_picture.f.mb_type[top_xy];
3782     left_type[LTOP] = s->current_picture.f.mb_type[left_xy[LTOP]];
3783     left_type[LBOT] = s->current_picture.f.mb_type[left_xy[LBOT]];
3784     if (h->deblocking_filter == 2) {
3785         if (h->slice_table[top_xy] != h->slice_num)
3786             top_type = 0;
3787         if (h->slice_table[left_xy[LBOT]] != h->slice_num)
3788             left_type[LTOP] = left_type[LBOT] = 0;
3789     } else {
3790         if (h->slice_table[top_xy] == 0xFFFF)
3791             top_type = 0;
3792         if (h->slice_table[left_xy[LBOT]] == 0xFFFF)
3793             left_type[LTOP] = left_type[LBOT] = 0;
3794     }
3795     h->top_type        = top_type;
3796     h->left_type[LTOP] = left_type[LTOP];
3797     h->left_type[LBOT] = left_type[LBOT];
3798
3799     if (IS_INTRA(mb_type))
3800         return 0;
3801
3802     fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy,
3803                              top_type, left_type, mb_xy, 0);
3804     if (h->list_count == 2)
3805         fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy,
3806                                  top_type, left_type, mb_xy, 1);
3807
3808     nnz       = h->non_zero_count[mb_xy];
3809     nnz_cache = h->non_zero_count_cache;
3810     AV_COPY32(&nnz_cache[4 + 8 * 1], &nnz[0]);
3811     AV_COPY32(&nnz_cache[4 + 8 * 2], &nnz[4]);
3812     AV_COPY32(&nnz_cache[4 + 8 * 3], &nnz[8]);
3813     AV_COPY32(&nnz_cache[4 + 8 * 4], &nnz[12]);
3814     h->cbp = h->cbp_table[mb_xy];
3815
3816     if (top_type) {
3817         nnz = h->non_zero_count[top_xy];
3818         AV_COPY32(&nnz_cache[4 + 8 * 0], &nnz[3 * 4]);
3819     }
3820
3821     if (left_type[LTOP]) {
3822         nnz = h->non_zero_count[left_xy[LTOP]];
3823         nnz_cache[3 + 8 * 1] = nnz[3 + 0 * 4];
3824         nnz_cache[3 + 8 * 2] = nnz[3 + 1 * 4];
3825         nnz_cache[3 + 8 * 3] = nnz[3 + 2 * 4];
3826         nnz_cache[3 + 8 * 4] = nnz[3 + 3 * 4];
3827     }
3828
3829     /* CAVLC 8x8dct requires NNZ values for residual decoding that differ
3830      * from what the loop filter needs */
3831     if (!CABAC && h->pps.transform_8x8_mode) {
3832         if (IS_8x8DCT(top_type)) {
3833             nnz_cache[4 + 8 * 0]     =
3834                 nnz_cache[5 + 8 * 0] = (h->cbp_table[top_xy] & 0x4000) >> 12;
3835             nnz_cache[6 + 8 * 0]     =
3836                 nnz_cache[7 + 8 * 0] = (h->cbp_table[top_xy] & 0x8000) >> 12;
3837         }
3838         if (IS_8x8DCT(left_type[LTOP])) {
3839             nnz_cache[3 + 8 * 1]     =
3840                 nnz_cache[3 + 8 * 2] = (h->cbp_table[left_xy[LTOP]] & 0x2000) >> 12; // FIXME check MBAFF
3841         }
3842         if (IS_8x8DCT(left_type[LBOT])) {
3843             nnz_cache[3 + 8 * 3]     =
3844                 nnz_cache[3 + 8 * 4] = (h->cbp_table[left_xy[LBOT]] & 0x8000) >> 12; // FIXME check MBAFF
3845         }
3846
3847         if (IS_8x8DCT(mb_type)) {
3848             nnz_cache[scan8[0]] =
3849             nnz_cache[scan8[1]] =
3850             nnz_cache[scan8[2]] =
3851             nnz_cache[scan8[3]] = (h->cbp & 0x1000) >> 12;
3852
3853             nnz_cache[scan8[0 + 4]] =
3854             nnz_cache[scan8[1 + 4]] =
3855             nnz_cache[scan8[2 + 4]] =
3856             nnz_cache[scan8[3 + 4]] = (h->cbp & 0x2000) >> 12;
3857
3858             nnz_cache[scan8[0 + 8]] =
3859             nnz_cache[scan8[1 + 8]] =
3860             nnz_cache[scan8[2 + 8]] =
3861             nnz_cache[scan8[3 + 8]] = (h->cbp & 0x4000) >> 12;
3862
3863             nnz_cache[scan8[0 + 12]] =
3864             nnz_cache[scan8[1 + 12]] =
3865             nnz_cache[scan8[2 + 12]] =
3866             nnz_cache[scan8[3 + 12]] = (h->cbp & 0x8000) >> 12;
3867         }
3868     }
3869
3870     return 0;
3871 }
3872
3873 static void loop_filter(H264Context *h, int start_x, int end_x)
3874 {
3875     MpegEncContext *const s = &h->s;
3876     uint8_t *dest_y, *dest_cb, *dest_cr;
3877     int linesize, uvlinesize, mb_x, mb_y;
3878     const int end_mb_y       = s->mb_y + FRAME_MBAFF;
3879     const int old_slice_type = h->slice_type;
3880     const int pixel_shift    = h->pixel_shift;
3881     const int block_h        = 16 >> s->chroma_y_shift;
3882
3883     if (h->deblocking_filter) {
3884         for (mb_x = start_x; mb_x < end_x; mb_x++)
3885             for (mb_y = end_mb_y - FRAME_MBAFF; mb_y <= end_mb_y; mb_y++) {
3886                 int mb_xy, mb_type;
3887                 mb_xy         = h->mb_xy = mb_x + mb_y * s->mb_stride;
3888                 h->slice_num  = h->slice_table[mb_xy];
3889                 mb_type       = s->current_picture.f.mb_type[mb_xy];
3890                 h->list_count = h->list_counts[mb_xy];
3891
3892                 if (FRAME_MBAFF)
3893                     h->mb_mbaff               =
3894                     h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
3895
3896                 s->mb_x = mb_x;
3897                 s->mb_y = mb_y;
3898                 dest_y  = s->current_picture.f.data[0] +
3899                           ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
3900                 dest_cb = s->current_picture.f.data[1] +
3901                           (mb_x << pixel_shift) * (8 << CHROMA444) +
3902                           mb_y * s->uvlinesize * block_h;
3903                 dest_cr = s->current_picture.f.data[2] +
3904                           (mb_x << pixel_shift) * (8 << CHROMA444) +
3905                           mb_y * s->uvlinesize * block_h;
3906                 // FIXME simplify above
3907
3908                 if (MB_FIELD) {
3909                     linesize   = h->mb_linesize   = s->linesize   * 2;
3910                     uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
3911                     if (mb_y & 1) { // FIXME move out of this function?
3912                         dest_y  -= s->linesize   * 15;
3913                         dest_cb -= s->uvlinesize * (block_h - 1);
3914                         dest_cr -= s->uvlinesize * (block_h - 1);
3915                     }
3916                 } else {
3917                     linesize   = h->mb_linesize   = s->linesize;
3918                     uvlinesize = h->mb_uvlinesize = s->uvlinesize;
3919                 }
3920                 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize,
3921                                  uvlinesize, 0);
3922                 if (fill_filter_caches(h, mb_type))
3923                     continue;
3924                 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.f.qscale_table[mb_xy]);
3925                 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.f.qscale_table[mb_xy]);
3926
3927                 if (FRAME_MBAFF) {
3928                     ff_h264_filter_mb(h, mb_x, mb_y, dest_y, dest_cb, dest_cr,
3929                                       linesize, uvlinesize);
3930                 } else {
3931                     ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb,
3932                                            dest_cr, linesize, uvlinesize);
3933                 }
3934             }
3935     }
3936     h->slice_type   = old_slice_type;
3937     s->mb_x         = end_x;
3938     s->mb_y         = end_mb_y - FRAME_MBAFF;
3939     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3940     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3941 }
3942
3943 static void predict_field_decoding_flag(H264Context *h)
3944 {
3945     MpegEncContext *const s = &h->s;
3946     const int mb_xy = s->mb_x + s->mb_y * s->mb_stride;
3947     int mb_type     = (h->slice_table[mb_xy - 1] == h->slice_num) ?
3948                       s->current_picture.f.mb_type[mb_xy - 1] :
3949                       (h->slice_table[mb_xy - s->mb_stride] == h->slice_num) ?
3950                       s->current_picture.f.mb_type[mb_xy - s->mb_stride] : 0;
3951     h->mb_mbaff     = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
3952 }
3953
3954 /**
3955  * Draw edges and report progress for the last MB row.
3956  */
3957 static void decode_finish_row(H264Context *h)
3958 {
3959     MpegEncContext *const s = &h->s;
3960     int top            = 16 * (s->mb_y      >> FIELD_PICTURE);
3961     int pic_height     = 16 *  s->mb_height >> FIELD_PICTURE;
3962     int height         =  16      << FRAME_MBAFF;
3963     int deblock_border = (16 + 4) << FRAME_MBAFF;
3964
3965     if (h->deblocking_filter) {
3966         if ((top + height) >= pic_height)
3967             height += deblock_border;
3968         top -= deblock_border;
3969     }
3970
3971     if (top >= pic_height || (top + height) < h->emu_edge_height)
3972         return;
3973
3974     height = FFMIN(height, pic_height - top);
3975     if (top < h->emu_edge_height) {
3976         height = top + height;
3977         top    = 0;
3978     }
3979
3980     ff_draw_horiz_band(s, top, height);
3981
3982     if (s->dropable)
3983         return;
3984
3985     ff_thread_report_progress(&s->current_picture_ptr->f, top + height - 1,
3986                               s->picture_structure == PICT_BOTTOM_FIELD);
3987 }
3988
3989 static int decode_slice(struct AVCodecContext *avctx, void *arg)
3990 {
3991     H264Context *h = *(void **)arg;
3992     MpegEncContext *const s = &h->s;
3993     const int part_mask     = s->partitioned_frame ? (ER_AC_END | ER_AC_ERROR)
3994                                                    : 0x7F;
3995     int lf_x_start = s->mb_x;
3996
3997     s->mb_skip_run = -1;
3998
3999     h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME ||
4000                     s->codec_id != CODEC_ID_H264 ||
4001                     (CONFIG_GRAY && (s->flags & CODEC_FLAG_GRAY));
4002
4003     if (h->pps.cabac) {
4004         /* realign */
4005         align_get_bits(&s->gb);
4006
4007         /* init cabac */
4008         ff_init_cabac_states(&h->cabac);
4009         ff_init_cabac_decoder(&h->cabac,
4010                               s->gb.buffer + get_bits_count(&s->gb) / 8,
4011                               (get_bits_left(&s->gb) + 7) / 8);
4012
4013         ff_h264_init_cabac_states(h);
4014
4015         for (;;) {
4016             // START_TIMER
4017             int ret = ff_h264_decode_mb_cabac(h);
4018             int eos;
4019             // STOP_TIMER("decode_mb_cabac")
4020
4021             if (ret >= 0)
4022                 ff_h264_hl_decode_mb(h);
4023
4024             // FIXME optimal? or let mb_decode decode 16x32 ?
4025             if (ret >= 0 && FRAME_MBAFF) {
4026                 s->mb_y++;
4027
4028                 ret = ff_h264_decode_mb_cabac(h);
4029
4030                 if (ret >= 0)
4031                     ff_h264_hl_decode_mb(h);
4032                 s->mb_y--;
4033             }
4034             eos = get_cabac_terminate(&h->cabac);
4035
4036             if ((s->workaround_bugs & FF_BUG_TRUNCATED) &&
4037                 h->cabac.bytestream > h->cabac.bytestream_end + 2) {
4038                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x - 1,
4039                                 s->mb_y, ER_MB_END & part_mask);
4040                 if (s->mb_x >= lf_x_start)
4041                     loop_filter(h, lf_x_start, s->mb_x + 1);
4042                 return 0;
4043             }
4044             if (h->cabac.bytestream > h->cabac.bytestream_end + 2 )
4045                 av_log(h->s.avctx, AV_LOG_DEBUG, "bytestream overread %td\n", h->cabac.bytestream_end - h->cabac.bytestream);
4046             if (ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 4) {
4047                 av_log(h->s.avctx, AV_LOG_ERROR,
4048                        "error while decoding MB %d %d, bytestream (%td)\n",
4049                        s->mb_x, s->mb_y,
4050                        h->cabac.bytestream_end - h->cabac.bytestream);
4051                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x,
4052                                 s->mb_y, ER_MB_ERROR & part_mask);
4053                 return -1;
4054             }
4055
4056             if (++s->mb_x >= s->mb_width) {
4057                 loop_filter(h, lf_x_start, s->mb_x);
4058                 s->mb_x = lf_x_start = 0;
4059                 decode_finish_row(h);
4060                 ++s->mb_y;
4061                 if (FIELD_OR_MBAFF_PICTURE) {
4062                     ++s->mb_y;
4063                     if (FRAME_MBAFF && s->mb_y < s->mb_height)
4064                         predict_field_decoding_flag(h);
4065                 }
4066             }
4067
4068             if (eos || s->mb_y >= s->mb_height) {
4069                 tprintf(s->avctx, "slice end %d %d\n",
4070                         get_bits_count(&s->gb), s->gb.size_in_bits);
4071                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x - 1,
4072                                 s->mb_y, ER_MB_END & part_mask);
4073                 if (s->mb_x > lf_x_start)
4074                     loop_filter(h, lf_x_start, s->mb_x);
4075                 return 0;
4076             }
4077         }
4078     } else {
4079         for (;;) {
4080             int ret = ff_h264_decode_mb_cavlc(h);
4081
4082             if (ret >= 0)
4083                 ff_h264_hl_decode_mb(h);
4084
4085             // FIXME optimal? or let mb_decode decode 16x32 ?
4086             if (ret >= 0 && FRAME_MBAFF) {
4087                 s->mb_y++;
4088                 ret = ff_h264_decode_mb_cavlc(h);
4089
4090                 if (ret >= 0)
4091                     ff_h264_hl_decode_mb(h);
4092                 s->mb_y--;
4093             }
4094
4095             if (ret < 0) {
4096                 av_log(h->s.avctx, AV_LOG_ERROR,
4097                        "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
4098                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x,
4099                                 s->mb_y, ER_MB_ERROR & part_mask);
4100                 return -1;
4101             }
4102
4103             if (++s->mb_x >= s->mb_width) {
4104                 loop_filter(h, lf_x_start, s->mb_x);
4105                 s->mb_x = lf_x_start = 0;
4106                 decode_finish_row(h);
4107                 ++s->mb_y;
4108                 if (FIELD_OR_MBAFF_PICTURE) {
4109                     ++s->mb_y;
4110                     if (FRAME_MBAFF && s->mb_y < s->mb_height)
4111                         predict_field_decoding_flag(h);
4112                 }
4113                 if (s->mb_y >= s->mb_height) {
4114                     tprintf(s->avctx, "slice end %d %d\n",
4115                             get_bits_count(&s->gb), s->gb.size_in_bits);
4116
4117                     if (   get_bits_left(&s->gb) == 0
4118                         || get_bits_left(&s->gb) > 0 && !(s->avctx->err_recognition & AV_EF_AGGRESSIVE)) {
4119                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y,
4120                                         s->mb_x - 1, s->mb_y,
4121                                         ER_MB_END & part_mask);
4122
4123                         return 0;
4124                     } else {
4125                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y,
4126                                         s->mb_x, s->mb_y,
4127                                         ER_MB_END & part_mask);
4128
4129                         return -1;
4130                     }
4131                 }
4132             }
4133
4134             if (get_bits_left(&s->gb) <= 0 && s->mb_skip_run <= 0) {
4135                 tprintf(s->avctx, "slice end %d %d\n",
4136                         get_bits_count(&s->gb), s->gb.size_in_bits);
4137                 if (get_bits_left(&s->gb) == 0) {
4138                     ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y,
4139                                     s->mb_x - 1, s->mb_y,
4140                                     ER_MB_END & part_mask);
4141                     if (s->mb_x > lf_x_start)
4142                         loop_filter(h, lf_x_start, s->mb_x);
4143
4144                     return 0;
4145                 } else {
4146                     ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x,
4147                                     s->mb_y, ER_MB_ERROR & part_mask);
4148
4149                     return -1;
4150                 }
4151             }
4152         }
4153     }
4154 }
4155
4156 /**
4157  * Call decode_slice() for each context.
4158  *
4159  * @param h h264 master context
4160  * @param context_count number of contexts to execute
4161  */
4162 static int execute_decode_slices(H264Context *h, int context_count)
4163 {
4164     MpegEncContext *const s     = &h->s;
4165     AVCodecContext *const avctx = s->avctx;
4166     H264Context *hx;
4167     int i;
4168
4169     if (s->avctx->hwaccel ||
4170         s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
4171         return 0;
4172     if (context_count == 1) {
4173         return decode_slice(avctx, &h);
4174     } else {
4175         for (i = 1; i < context_count; i++) {
4176             hx                    = h->thread_context[i];
4177             hx->s.err_recognition = avctx->err_recognition;
4178             hx->s.error_count     = 0;
4179             hx->x264_build        = h->x264_build;
4180         }
4181
4182         avctx->execute(avctx, decode_slice, h->thread_context,
4183                        NULL, context_count, sizeof(void *));
4184
4185         /* pull back stuff from slices to master context */
4186         hx                   = h->thread_context[context_count - 1];
4187         s->mb_x              = hx->s.mb_x;
4188         s->mb_y              = hx->s.mb_y;
4189         s->dropable          = hx->s.dropable;
4190         s->picture_structure = hx->s.picture_structure;
4191         for (i = 1; i < context_count; i++)
4192             h->s.error_count += h->thread_context[i]->s.error_count;
4193     }
4194
4195     return 0;
4196 }
4197
4198 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size)
4199 {
4200     MpegEncContext *const s     = &h->s;
4201     AVCodecContext *const avctx = s->avctx;
4202     H264Context *hx; ///< thread context
4203     int buf_index;
4204     int context_count;
4205     int next_avc;
4206     int pass = !(avctx->active_thread_type & FF_THREAD_FRAME);
4207     int nals_needed = 0; ///< number of NALs that need decoding before the next frame thread starts
4208     int nal_index;
4209
4210     h->nal_unit_type= 0;
4211
4212     if(!s->slice_context_count)
4213          s->slice_context_count= 1;
4214     h->max_contexts = s->slice_context_count;
4215     if (!(s->flags2 & CODEC_FLAG2_CHUNKS)) {
4216         h->current_slice = 0;
4217         if (!s->first_field)
4218             s->current_picture_ptr = NULL;
4219         ff_h264_reset_sei(h);
4220     }
4221
4222     for (; pass <= 1; pass++) {
4223         buf_index     = 0;
4224         context_count = 0;
4225         next_avc      = h->is_avc ? 0 : buf_size;
4226         nal_index     = 0;
4227         for (;;) {
4228             int consumed;
4229             int dst_length;
4230             int bit_length;
4231             const uint8_t *ptr;
4232             int i, nalsize = 0;
4233             int err;
4234
4235             if (buf_index >= next_avc) {
4236                 if (buf_index >= buf_size - h->nal_length_size)
4237                     break;
4238                 nalsize = 0;
4239                 for (i = 0; i < h->nal_length_size; i++)
4240                     nalsize = (nalsize << 8) | buf[buf_index++];
4241                 if (nalsize <= 0 || nalsize > buf_size - buf_index) {
4242                     av_log(h->s.avctx, AV_LOG_ERROR,
4243                            "AVC: nal size %d\n", nalsize);
4244                     break;
4245                 }
4246                 next_avc = buf_index + nalsize;
4247             } else {
4248                 // start code prefix search
4249                 for (; buf_index + 3 < next_avc; buf_index++)
4250                     // This should always succeed in the first iteration.
4251                     if (buf[buf_index]     == 0 &&
4252                         buf[buf_index + 1] == 0 &&
4253                         buf[buf_index + 2] == 1)
4254                         break;
4255
4256                 if (buf_index + 3 >= buf_size)
4257                     break;
4258
4259                 buf_index += 3;
4260                 if (buf_index >= next_avc)
4261                     continue;
4262             }
4263
4264             hx = h->thread_context[context_count];
4265
4266             ptr = ff_h264_decode_nal(hx, buf + buf_index, &dst_length,
4267                                      &consumed, next_avc - buf_index);
4268             if (ptr == NULL || dst_length < 0) {
4269                 buf_index = -1;
4270                 goto end;
4271             }
4272             i = buf_index + consumed;
4273             if ((s->workaround_bugs & FF_BUG_AUTODETECT) && i + 3 < next_avc &&
4274                 buf[i]     == 0x00 && buf[i + 1] == 0x00 &&
4275                 buf[i + 2] == 0x01 && buf[i + 3] == 0xE0)
4276                 s->workaround_bugs |= FF_BUG_TRUNCATED;
4277
4278             if (!(s->workaround_bugs & FF_BUG_TRUNCATED))
4279                 while(dst_length > 0 && ptr[dst_length - 1] == 0)
4280                     dst_length--;
4281             bit_length = !dst_length ? 0
4282                                      : (8 * dst_length -
4283                                         decode_rbsp_trailing(h, ptr + dst_length - 1));
4284
4285             if (s->avctx->debug & FF_DEBUG_STARTCODE)
4286                 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);
4287
4288             if (h->is_avc && (nalsize != consumed) && nalsize)
4289                 av_log(h->s.avctx, AV_LOG_DEBUG,
4290                        "AVC: Consumed only %d bytes instead of %d\n",
4291                        consumed, nalsize);
4292
4293             buf_index += consumed;
4294             nal_index++;
4295
4296             if (pass == 0) {
4297                 /* packets can sometimes contain multiple PPS/SPS,
4298                  * e.g. two PAFF field pictures in one packet, or a demuxer
4299                  * which splits NALs strangely if so, when frame threading we
4300                  * can't start the next thread until we've read all of them */
4301                 switch (hx->nal_unit_type) {
4302                 case NAL_SPS:
4303                 case NAL_PPS:
4304                     nals_needed = nal_index;
4305                     break;
4306                 case NAL_IDR_SLICE:
4307                 case NAL_SLICE:
4308                     init_get_bits(&hx->s.gb, ptr, bit_length);
4309                     if (!get_ue_golomb(&hx->s.gb))
4310                         nals_needed = nal_index;
4311                 }
4312                 continue;
4313             }
4314
4315             // FIXME do not discard SEI id
4316             if (avctx->skip_frame >= AVDISCARD_NONREF && h->nal_ref_idc == 0)
4317                 continue;
4318
4319 again:
4320             err = 0;
4321             switch (hx->nal_unit_type) {
4322             case NAL_IDR_SLICE:
4323                 if (h->nal_unit_type != NAL_IDR_SLICE) {
4324                     av_log(h->s.avctx, AV_LOG_ERROR,
4325                            "Invalid mix of idr and non-idr slices\n");
4326                     buf_index = -1;
4327                     goto end;
4328                 }
4329                 idr(h); // FIXME ensure we don't lose some frames if there is reordering
4330             case NAL_SLICE:
4331                 init_get_bits(&hx->s.gb, ptr, bit_length);
4332                 hx->intra_gb_ptr        =
4333                     hx->inter_gb_ptr    = &hx->s.gb;
4334                 hx->s.data_partitioning = 0;
4335
4336                 if ((err = decode_slice_header(hx, h)))
4337                     break;
4338
4339                 if (   h->sei_recovery_frame_cnt >= 0
4340                     && (   h->recovery_frame<0
4341                         || ((h->recovery_frame - h->frame_num) & ((1 << h->sps.log2_max_frame_num)-1)) > h->sei_recovery_frame_cnt)) {
4342                     h->recovery_frame = (h->frame_num + h->sei_recovery_frame_cnt) %
4343                                         (1 << h->sps.log2_max_frame_num);
4344                 }
4345
4346                 s->current_picture_ptr->f.key_frame |=
4347                         (hx->nal_unit_type == NAL_IDR_SLICE);
4348
4349                 if (h->recovery_frame == h->frame_num) {
4350                     s->current_picture_ptr->sync |= 1;
4351                     h->recovery_frame = -1;
4352                 }
4353
4354                 h->sync |= !!s->current_picture_ptr->f.key_frame;
4355                 h->sync |= 3*!!(s->flags2 & CODEC_FLAG2_SHOW_ALL);
4356                 s->current_picture_ptr->sync |= h->sync;
4357
4358                 if (h->current_slice == 1) {
4359                     if (!(s->flags2 & CODEC_FLAG2_CHUNKS))
4360                         decode_postinit(h, nal_index >= nals_needed);
4361
4362                     if (s->avctx->hwaccel &&
4363                         s->avctx->hwaccel->start_frame(s->avctx, NULL, 0) < 0)
4364                         return -1;
4365                     if (CONFIG_H264_VDPAU_DECODER &&
4366                         s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
4367                         ff_vdpau_h264_picture_start(s);
4368                 }
4369
4370                 if (hx->redundant_pic_count == 0 &&
4371                     (avctx->skip_frame < AVDISCARD_NONREF ||
4372                      hx->nal_ref_idc) &&
4373                     (avctx->skip_frame < AVDISCARD_BIDIR  ||
4374                      hx->slice_type_nos != AV_PICTURE_TYPE_B) &&
4375                     (avctx->skip_frame < AVDISCARD_NONKEY ||
4376                      hx->slice_type_nos == AV_PICTURE_TYPE_I) &&
4377                     avctx->skip_frame < AVDISCARD_ALL) {
4378                     if (avctx->hwaccel) {
4379                         if (avctx->hwaccel->decode_slice(avctx,
4380                                                          &buf[buf_index - consumed],
4381                                                          consumed) < 0)
4382                             return -1;
4383                     } else if (CONFIG_H264_VDPAU_DECODER &&
4384                                s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU) {
4385                         static const uint8_t start_code[] = {
4386                             0x00, 0x00, 0x01 };
4387                         ff_vdpau_add_data_chunk(s, start_code,
4388                                                 sizeof(start_code));
4389                         ff_vdpau_add_data_chunk(s, &buf[buf_index - consumed],
4390                                                 consumed);
4391                     } else
4392                         context_count++;
4393                 }
4394                 break;
4395             case NAL_DPA:
4396                 init_get_bits(&hx->s.gb, ptr, bit_length);
4397                 hx->intra_gb_ptr =
4398                 hx->inter_gb_ptr = NULL;
4399
4400                 if ((err = decode_slice_header(hx, h)) < 0)
4401                     break;
4402
4403                 hx->s.data_partitioning = 1;
4404                 break;
4405             case NAL_DPB:
4406                 init_get_bits(&hx->intra_gb, ptr, bit_length);
4407                 hx->intra_gb_ptr = &hx->intra_gb;
4408                 break;
4409             case NAL_DPC:
4410                 init_get_bits(&hx->inter_gb, ptr, bit_length);
4411                 hx->inter_gb_ptr = &hx->inter_gb;
4412
4413                 av_log(h->s.avctx, AV_LOG_ERROR, "Partitioned H.264 support is incomplete\n");
4414                 return AVERROR_PATCHWELCOME;
4415
4416                 if (hx->redundant_pic_count == 0 &&
4417                     hx->intra_gb_ptr &&
4418                     hx->s.data_partitioning &&
4419                     s->context_initialized &&
4420                     (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc) &&
4421                     (avctx->skip_frame < AVDISCARD_BIDIR  ||
4422                      hx->slice_type_nos != AV_PICTURE_TYPE_B) &&
4423                     (avctx->skip_frame < AVDISCARD_NONKEY ||
4424                      hx->slice_type_nos == AV_PICTURE_TYPE_I) &&
4425                     avctx->skip_frame < AVDISCARD_ALL)
4426                     context_count++;
4427                 break;
4428             case NAL_SEI:
4429                 init_get_bits(&s->gb, ptr, bit_length);
4430                 ff_h264_decode_sei(h);
4431                 break;
4432             case NAL_SPS:
4433                 init_get_bits(&s->gb, ptr, bit_length);
4434                 if (ff_h264_decode_seq_parameter_set(h) < 0 && (h->is_avc ? (nalsize != consumed) && nalsize : 1)) {
4435                     av_log(h->s.avctx, AV_LOG_DEBUG,
4436                            "SPS decoding failure, trying alternative mode\n");
4437                     if (h->is_avc)
4438                         av_assert0(next_avc - buf_index + consumed == nalsize);
4439                     init_get_bits(&s->gb, &buf[buf_index + 1 - consumed],
4440                                   8*(next_avc - buf_index + consumed - 1));
4441                     ff_h264_decode_seq_parameter_set(h);
4442                 }
4443
4444                 if (s->flags & CODEC_FLAG_LOW_DELAY ||
4445                     (h->sps.bitstream_restriction_flag &&
4446                      !h->sps.num_reorder_frames))
4447                     s->low_delay = 1;
4448                 if (avctx->has_b_frames < 2)
4449                     avctx->has_b_frames = !s->low_delay;
4450                 break;
4451             case NAL_PPS:
4452                 init_get_bits(&s->gb, ptr, bit_length);
4453                 ff_h264_decode_picture_parameter_set(h, bit_length);
4454                 break;
4455             case NAL_AUD:
4456             case NAL_END_SEQUENCE:
4457             case NAL_END_STREAM:
4458             case NAL_FILLER_DATA:
4459             case NAL_SPS_EXT:
4460             case NAL_AUXILIARY_SLICE:
4461                 break;
4462             default:
4463                 av_log(avctx, AV_LOG_DEBUG, "Unknown NAL code: %d (%d bits)\n",
4464                        hx->nal_unit_type, bit_length);
4465             }
4466
4467             if (context_count == h->max_contexts) {
4468                 execute_decode_slices(h, context_count);
4469                 context_count = 0;
4470             }
4471
4472             if (err < 0)
4473                 av_log(h->s.avctx, AV_LOG_ERROR, "decode_slice_header error\n");
4474             else if (err == 1) {
4475                 /* Slice could not be decoded in parallel mode, copy down
4476                  * NAL unit stuff to context 0 and restart. Note that
4477                  * rbsp_buffer is not transferred, but since we no longer
4478                  * run in parallel mode this should not be an issue. */
4479                 h->nal_unit_type = hx->nal_unit_type;
4480                 h->nal_ref_idc   = hx->nal_ref_idc;
4481                 hx               = h;
4482                 goto again;
4483             }
4484         }
4485     }
4486     if (context_count)
4487         execute_decode_slices(h, context_count);
4488
4489 end:
4490     /* clean up */
4491     if (s->current_picture_ptr && s->current_picture_ptr->owner2 == s &&
4492         !s->dropable) {
4493         ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX,
4494                                   s->picture_structure == PICT_BOTTOM_FIELD);
4495     }
4496
4497     return buf_index;
4498 }
4499
4500 /**
4501  * Return the number of bytes consumed for building the current frame.
4502  */
4503 static int get_consumed_bytes(MpegEncContext *s, int pos, int buf_size)
4504 {
4505     if (pos == 0)
4506         pos = 1;          // avoid infinite loops (i doubt that is needed but ...)
4507     if (pos + 10 > buf_size)
4508         pos = buf_size;                   // oops ;)
4509
4510     return pos;
4511 }
4512
4513 static int decode_frame(AVCodecContext *avctx, void *data,
4514                         int *data_size, AVPacket *avpkt)
4515 {
4516     const uint8_t *buf = avpkt->data;
4517     int buf_size       = avpkt->size;
4518     H264Context *h     = avctx->priv_data;
4519     MpegEncContext *s  = &h->s;
4520     AVFrame *pict      = data;
4521     int buf_index      = 0;
4522     Picture *out;
4523     int i, out_idx;
4524
4525     s->flags  = avctx->flags;
4526     s->flags2 = avctx->flags2;
4527
4528     /* end of stream, output what is still in the buffers */
4529     if (buf_size == 0) {
4530  out:
4531
4532         s->current_picture_ptr = NULL;
4533
4534         // FIXME factorize this with the output code below
4535         out     = h->delayed_pic[0];
4536         out_idx = 0;
4537         for (i = 1;
4538              h->delayed_pic[i] &&
4539              !h->delayed_pic[i]->f.key_frame &&
4540              !h->delayed_pic[i]->mmco_reset;
4541              i++)
4542             if (h->delayed_pic[i]->poc < out->poc) {
4543                 out     = h->delayed_pic[i];
4544                 out_idx = i;
4545             }
4546
4547         for (i = out_idx; h->delayed_pic[i]; i++)
4548             h->delayed_pic[i] = h->delayed_pic[i + 1];
4549
4550         if (out) {
4551             *data_size = sizeof(AVFrame);
4552             *pict      = out->f;
4553         }
4554
4555         return buf_index;
4556     }
4557     if(h->is_avc && buf_size >= 9 && buf[0]==1 && buf[2]==0 && (buf[4]&0xFC)==0xFC && (buf[5]&0x1F) && buf[8]==0x67){
4558         int cnt= buf[5]&0x1f;
4559         const uint8_t *p= buf+6;
4560         while(cnt--){
4561             int nalsize= AV_RB16(p) + 2;
4562             if(nalsize > buf_size - (p-buf) || p[2]!=0x67)
4563                 goto not_extra;
4564             p += nalsize;
4565         }
4566         cnt = *(p++);
4567         if(!cnt)
4568             goto not_extra;
4569         while(cnt--){
4570             int nalsize= AV_RB16(p) + 2;
4571             if(nalsize > buf_size - (p-buf) || p[2]!=0x68)
4572                 goto not_extra;
4573             p += nalsize;
4574         }
4575
4576         return ff_h264_decode_extradata(h, buf, buf_size);
4577     }
4578 not_extra:
4579
4580     buf_index = decode_nal_units(h, buf, buf_size);
4581     if (buf_index < 0)
4582         return -1;
4583
4584     if (!s->current_picture_ptr && h->nal_unit_type == NAL_END_SEQUENCE) {
4585         av_assert0(buf_index <= buf_size);
4586         goto out;
4587     }
4588
4589     if (!(s->flags2 & CODEC_FLAG2_CHUNKS) && !s->current_picture_ptr) {
4590         if (avctx->skip_frame >= AVDISCARD_NONREF ||
4591             buf_size >= 4 && !memcmp("Q264", buf, 4))
4592             return buf_size;
4593         av_log(avctx, AV_LOG_ERROR, "no frame!\n");
4594         return -1;
4595     }
4596
4597     if (!(s->flags2 & CODEC_FLAG2_CHUNKS) ||
4598         (s->mb_y >= s->mb_height && s->mb_height)) {
4599         if (s->flags2 & CODEC_FLAG2_CHUNKS)
4600             decode_postinit(h, 1);
4601
4602         field_end(h, 0);
4603
4604         /* Wait for second field. */
4605         *data_size = 0;
4606         if (h->next_output_pic && (h->next_output_pic->sync || h->sync>1)) {
4607             *data_size = sizeof(AVFrame);
4608             *pict      = h->next_output_pic->f;
4609         }
4610     }
4611
4612     assert(pict->data[0] || !*data_size);
4613     ff_print_debug_info(s, pict);
4614     // printf("out %d\n", (int)pict->data[0]);
4615
4616     return get_consumed_bytes(s, buf_index, buf_size);
4617 }
4618
4619 av_cold void ff_h264_free_context(H264Context *h)
4620 {
4621     int i;
4622
4623     free_tables(h, 1); // FIXME cleanup init stuff perhaps
4624
4625     for (i = 0; i < MAX_SPS_COUNT; i++)
4626         av_freep(h->sps_buffers + i);
4627
4628     for (i = 0; i < MAX_PPS_COUNT; i++)
4629         av_freep(h->pps_buffers + i);
4630 }
4631
4632 static av_cold int h264_decode_end(AVCodecContext *avctx)
4633 {
4634     H264Context *h    = avctx->priv_data;
4635     MpegEncContext *s = &h->s;
4636
4637     ff_h264_remove_all_refs(h);
4638     ff_h264_free_context(h);
4639
4640     ff_MPV_common_end(s);
4641
4642     // memset(h, 0, sizeof(H264Context));
4643
4644     return 0;
4645 }
4646
4647 static const AVProfile profiles[] = {
4648     { FF_PROFILE_H264_BASELINE,             "Baseline"              },
4649     { FF_PROFILE_H264_CONSTRAINED_BASELINE, "Constrained Baseline"  },
4650     { FF_PROFILE_H264_MAIN,                 "Main"                  },
4651     { FF_PROFILE_H264_EXTENDED,             "Extended"              },
4652     { FF_PROFILE_H264_HIGH,                 "High"                  },
4653     { FF_PROFILE_H264_HIGH_10,              "High 10"               },
4654     { FF_PROFILE_H264_HIGH_10_INTRA,        "High 10 Intra"         },
4655     { FF_PROFILE_H264_HIGH_422,             "High 4:2:2"            },
4656     { FF_PROFILE_H264_HIGH_422_INTRA,       "High 4:2:2 Intra"      },
4657     { FF_PROFILE_H264_HIGH_444,             "High 4:4:4"            },
4658     { FF_PROFILE_H264_HIGH_444_PREDICTIVE,  "High 4:4:4 Predictive" },
4659     { FF_PROFILE_H264_HIGH_444_INTRA,       "High 4:4:4 Intra"      },
4660     { FF_PROFILE_H264_CAVLC_444,            "CAVLC 4:4:4"           },
4661     { FF_PROFILE_UNKNOWN },
4662 };
4663
4664 static const AVOption h264_options[] = {
4665     {"is_avc", "is avc", offsetof(H264Context, is_avc), FF_OPT_TYPE_INT, {.dbl = 0}, 0, 1, 0},
4666     {"nal_length_size", "nal_length_size", offsetof(H264Context, nal_length_size), FF_OPT_TYPE_INT, {.dbl = 0}, 0, 4, 0},
4667     {NULL}
4668 };
4669
4670 static const AVClass h264_class = {
4671     "H264 Decoder",
4672     av_default_item_name,
4673     h264_options,
4674     LIBAVUTIL_VERSION_INT,
4675 };
4676
4677 static const AVClass h264_vdpau_class = {
4678     "H264 VDPAU Decoder",
4679     av_default_item_name,
4680     h264_options,
4681     LIBAVUTIL_VERSION_INT,
4682 };
4683
4684 AVCodec ff_h264_decoder = {
4685     .name                  = "h264",
4686     .type                  = AVMEDIA_TYPE_VIDEO,
4687     .id                    = CODEC_ID_H264,
4688     .priv_data_size        = sizeof(H264Context),
4689     .init                  = ff_h264_decode_init,
4690     .close                 = h264_decode_end,
4691     .decode                = decode_frame,
4692     .capabilities          = /*CODEC_CAP_DRAW_HORIZ_BAND |*/ CODEC_CAP_DR1 |
4693                              CODEC_CAP_DELAY | CODEC_CAP_SLICE_THREADS |
4694                              CODEC_CAP_FRAME_THREADS,
4695     .flush                 = flush_dpb,
4696     .long_name             = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10"),
4697     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(decode_init_thread_copy),
4698     .update_thread_context = ONLY_IF_THREADS_ENABLED(decode_update_thread_context),
4699     .profiles              = NULL_IF_CONFIG_SMALL(profiles),
4700     .priv_class            = &h264_class,
4701 };
4702
4703 #if CONFIG_H264_VDPAU_DECODER
4704 AVCodec ff_h264_vdpau_decoder = {
4705     .name           = "h264_vdpau",
4706     .type           = AVMEDIA_TYPE_VIDEO,
4707     .id             = CODEC_ID_H264,
4708     .priv_data_size = sizeof(H264Context),
4709     .init           = ff_h264_decode_init,
4710     .close          = h264_decode_end,
4711     .decode         = decode_frame,
4712     .capabilities   = CODEC_CAP_DR1 | CODEC_CAP_DELAY | CODEC_CAP_HWACCEL_VDPAU,
4713     .flush          = flush_dpb,
4714     .long_name      = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10 (VDPAU acceleration)"),
4715     .pix_fmts       = (const enum PixelFormat[]) { PIX_FMT_VDPAU_H264,
4716                                                    PIX_FMT_NONE},
4717     .profiles       = NULL_IF_CONFIG_SMALL(profiles),
4718     .priv_class     = &h264_vdpau_class,
4719 };
4720 #endif