cmdutils: fix -fdebug without -v xy
[ffmpeg.git] / libavcodec / h264.c
1 /*
2  * H.26L/H.264/AVC/JVT/14496-10/... decoder
3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21
22 /**
23  * @file
24  * H.264 / AVC / MPEG4 part10 codec.
25  * @author Michael Niedermayer <michaelni@gmx.at>
26  */
27
28 #define UNCHECKED_BITSTREAM_READER 1
29
30 #include "libavutil/imgutils.h"
31 #include "libavutil/opt.h"
32 #include "internal.h"
33 #include "cabac.h"
34 #include "cabac_functions.h"
35 #include "dsputil.h"
36 #include "avcodec.h"
37 #include "mpegvideo.h"
38 #include "h264.h"
39 #include "h264data.h"
40 #include "h264_mvpred.h"
41 #include "golomb.h"
42 #include "mathops.h"
43 #include "rectangle.h"
44 #include "thread.h"
45 #include "vdpau_internal.h"
46 #include "libavutil/avassert.h"
47
48 // #undef NDEBUG
49 #include <assert.h>
50
51 static const uint8_t rem6[QP_MAX_NUM + 1] = {
52     0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2,
53     3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5,
54     0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
55 };
56
57 static const uint8_t div6[QP_MAX_NUM + 1] = {
58     0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3,  3,  3,
59     3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6,  6,  6,
60     7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 10, 10, 10, 10,
61 };
62
63 static const enum PixelFormat hwaccel_pixfmt_list_h264_jpeg_420[] = {
64     PIX_FMT_DXVA2_VLD,
65     PIX_FMT_VAAPI_VLD,
66     PIX_FMT_VDA_VLD,
67     PIX_FMT_YUVJ420P,
68     PIX_FMT_NONE
69 };
70
71 /**
72  * Check if the top & left blocks are available if needed and
73  * change the dc mode so it only uses the available blocks.
74  */
75 int ff_h264_check_intra4x4_pred_mode(H264Context *h)
76 {
77     MpegEncContext *const s     = &h->s;
78     static const int8_t top[12] = {
79         -1, 0, LEFT_DC_PRED, -1, -1, -1, -1, -1, 0
80     };
81     static const int8_t left[12] = {
82         0, -1, TOP_DC_PRED, 0, -1, -1, -1, 0, -1, DC_128_PRED
83     };
84     int i;
85
86     if (!(h->top_samples_available & 0x8000)) {
87         for (i = 0; i < 4; i++) {
88             int status = top[h->intra4x4_pred_mode_cache[scan8[0] + i]];
89             if (status < 0) {
90                 av_log(h->s.avctx, AV_LOG_ERROR,
91                        "top block unavailable for requested intra4x4 mode %d at %d %d\n",
92                        status, s->mb_x, s->mb_y);
93                 return -1;
94             } else if (status) {
95                 h->intra4x4_pred_mode_cache[scan8[0] + i] = status;
96             }
97         }
98     }
99
100     if ((h->left_samples_available & 0x8888) != 0x8888) {
101         static const int mask[4] = { 0x8000, 0x2000, 0x80, 0x20 };
102         for (i = 0; i < 4; i++)
103             if (!(h->left_samples_available & mask[i])) {
104                 int status = left[h->intra4x4_pred_mode_cache[scan8[0] + 8 * i]];
105                 if (status < 0) {
106                     av_log(h->s.avctx, AV_LOG_ERROR,
107                            "left block unavailable for requested intra4x4 mode %d at %d %d\n",
108                            status, s->mb_x, s->mb_y);
109                     return -1;
110                 } else if (status) {
111                     h->intra4x4_pred_mode_cache[scan8[0] + 8 * i] = status;
112                 }
113             }
114     }
115
116     return 0;
117 } // FIXME cleanup like ff_h264_check_intra_pred_mode
118
119 /**
120  * Check if the top & left blocks are available if needed and
121  * change the dc mode so it only uses the available blocks.
122  */
123 int ff_h264_check_intra_pred_mode(H264Context *h, int mode, int is_chroma)
124 {
125     MpegEncContext *const s     = &h->s;
126     static const int8_t top[7]  = { LEFT_DC_PRED8x8, 1, -1, -1 };
127     static const int8_t left[7] = { TOP_DC_PRED8x8, -1, 2, -1, DC_128_PRED8x8 };
128
129     if (mode > 6U) {
130         av_log(h->s.avctx, AV_LOG_ERROR,
131                "out of range intra chroma pred mode at %d %d\n",
132                s->mb_x, s->mb_y);
133         return -1;
134     }
135
136     if (!(h->top_samples_available & 0x8000)) {
137         mode = top[mode];
138         if (mode < 0) {
139             av_log(h->s.avctx, AV_LOG_ERROR,
140                    "top block unavailable for requested intra mode at %d %d\n",
141                    s->mb_x, s->mb_y);
142             return -1;
143         }
144     }
145
146     if ((h->left_samples_available & 0x8080) != 0x8080) {
147         mode = left[mode];
148         if (is_chroma && (h->left_samples_available & 0x8080)) {
149             // mad cow disease mode, aka MBAFF + constrained_intra_pred
150             mode = ALZHEIMER_DC_L0T_PRED8x8 +
151                    (!(h->left_samples_available & 0x8000)) +
152                    2 * (mode == DC_128_PRED8x8);
153         }
154         if (mode < 0) {
155             av_log(h->s.avctx, AV_LOG_ERROR,
156                    "left block unavailable for requested intra mode at %d %d\n",
157                    s->mb_x, s->mb_y);
158             return -1;
159         }
160     }
161
162     return mode;
163 }
164
165 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src,
166                                   int *dst_length, int *consumed, int length)
167 {
168     int i, si, di;
169     uint8_t *dst;
170     int bufidx;
171
172     // src[0]&0x80; // forbidden bit
173     h->nal_ref_idc   = src[0] >> 5;
174     h->nal_unit_type = src[0] & 0x1F;
175
176     src++;
177     length--;
178
179 #if HAVE_FAST_UNALIGNED
180 #if HAVE_FAST_64BIT
181 #define RS 7
182     for (i = 0; i + 1 < length; i += 9) {
183         if (!((~AV_RN64A(src + i) &
184                (AV_RN64A(src + i) - 0x0100010001000101ULL)) &
185               0x8000800080008080ULL))
186 #else
187 #define RS 3
188     for (i = 0; i + 1 < length; i += 5) {
189         if (!((~AV_RN32A(src + i) &
190                (AV_RN32A(src + i) - 0x01000101U)) &
191               0x80008080U))
192 #endif
193             continue;
194         if (i > 0 && !src[i])
195             i--;
196         while (src[i])
197             i++;
198 #else
199 #define RS 0
200     for (i = 0; i + 1 < length; i += 2) {
201         if (src[i])
202             continue;
203         if (i > 0 && src[i - 1] == 0)
204             i--;
205 #endif
206         if (i + 2 < length && src[i + 1] == 0 && src[i + 2] <= 3) {
207             if (src[i + 2] != 3) {
208                 /* startcode, so we must be past the end */
209                 length = i;
210             }
211             break;
212         }
213         i -= RS;
214     }
215
216     // use second escape buffer for inter data
217     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0;
218
219     si = h->rbsp_buffer_size[bufidx];
220     av_fast_padded_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+MAX_MBPAIR_SIZE);
221     dst = h->rbsp_buffer[bufidx];
222
223     if (dst == NULL)
224         return NULL;
225
226     if(i>=length-1){ //no escaped 0
227         *dst_length= length;
228         *consumed= length+1; //+1 for the header
229         if(h->s.avctx->flags2 & CODEC_FLAG2_FAST){
230             return src;
231         }else{
232             memcpy(dst, src, length);
233             return dst;
234         }
235     }
236
237     // printf("decoding esc\n");
238     memcpy(dst, src, i);
239     si = di = i;
240     while (si + 2 < length) {
241         // remove escapes (very rare 1:2^22)
242         if (src[si + 2] > 3) {
243             dst[di++] = src[si++];
244             dst[di++] = src[si++];
245         } else if (src[si] == 0 && src[si + 1] == 0) {
246             if (src[si + 2] == 3) { // escape
247                 dst[di++]  = 0;
248                 dst[di++]  = 0;
249                 si        += 3;
250                 continue;
251             } else // next start code
252                 goto nsc;
253         }
254
255         dst[di++] = src[si++];
256     }
257     while (si < length)
258         dst[di++] = src[si++];
259 nsc:
260
261     memset(dst + di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
262
263     *dst_length = di;
264     *consumed   = si + 1; // +1 for the header
265     /* FIXME store exact number of bits in the getbitcontext
266      * (it is needed for decoding) */
267     return dst;
268 }
269
270 /**
271  * Identify the exact end of the bitstream
272  * @return the length of the trailing, or 0 if damaged
273  */
274 static int decode_rbsp_trailing(H264Context *h, const uint8_t *src)
275 {
276     int v = *src;
277     int r;
278
279     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
280
281     for (r = 1; r < 9; r++) {
282         if (v & 1)
283             return r;
284         v >>= 1;
285     }
286     return 0;
287 }
288
289 static inline int get_lowest_part_list_y(H264Context *h, Picture *pic, int n,
290                                          int height, int y_offset, int list)
291 {
292     int raw_my        = h->mv_cache[list][scan8[n]][1];
293     int filter_height = (raw_my & 3) ? 2 : 0;
294     int full_my       = (raw_my >> 2) + y_offset;
295     int top           = full_my - filter_height;
296     int bottom        = full_my + filter_height + height;
297
298     return FFMAX(abs(top), bottom);
299 }
300
301 static inline void get_lowest_part_y(H264Context *h, int refs[2][48], int n,
302                                      int height, int y_offset, int list0,
303                                      int list1, int *nrefs)
304 {
305     MpegEncContext *const s = &h->s;
306     int my;
307
308     y_offset += 16 * (s->mb_y >> MB_FIELD);
309
310     if (list0) {
311         int ref_n    = h->ref_cache[0][scan8[n]];
312         Picture *ref = &h->ref_list[0][ref_n];
313
314         // Error resilience puts the current picture in the ref list.
315         // Don't try to wait on these as it will cause a deadlock.
316         // Fields can wait on each other, though.
317         if (ref->f.thread_opaque   != s->current_picture.f.thread_opaque ||
318             (ref->f.reference & 3) != s->picture_structure) {
319             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 0);
320             if (refs[0][ref_n] < 0)
321                 nrefs[0] += 1;
322             refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
323         }
324     }
325
326     if (list1) {
327         int ref_n    = h->ref_cache[1][scan8[n]];
328         Picture *ref = &h->ref_list[1][ref_n];
329
330         if (ref->f.thread_opaque   != s->current_picture.f.thread_opaque ||
331             (ref->f.reference & 3) != s->picture_structure) {
332             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 1);
333             if (refs[1][ref_n] < 0)
334                 nrefs[1] += 1;
335             refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
336         }
337     }
338 }
339
340 /**
341  * Wait until all reference frames are available for MC operations.
342  *
343  * @param h the H264 context
344  */
345 static void await_references(H264Context *h)
346 {
347     MpegEncContext *const s = &h->s;
348     const int mb_xy   = h->mb_xy;
349     const int mb_type = s->current_picture.f.mb_type[mb_xy];
350     int refs[2][48];
351     int nrefs[2] = { 0 };
352     int ref, list;
353
354     memset(refs, -1, sizeof(refs));
355
356     if (IS_16X16(mb_type)) {
357         get_lowest_part_y(h, refs, 0, 16, 0,
358                           IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
359     } else if (IS_16X8(mb_type)) {
360         get_lowest_part_y(h, refs, 0, 8, 0,
361                           IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
362         get_lowest_part_y(h, refs, 8, 8, 8,
363                           IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
364     } else if (IS_8X16(mb_type)) {
365         get_lowest_part_y(h, refs, 0, 16, 0,
366                           IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
367         get_lowest_part_y(h, refs, 4, 16, 0,
368                           IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
369     } else {
370         int i;
371
372         assert(IS_8X8(mb_type));
373
374         for (i = 0; i < 4; i++) {
375             const int sub_mb_type = h->sub_mb_type[i];
376             const int n           = 4 * i;
377             int y_offset          = (i & 2) << 2;
378
379             if (IS_SUB_8X8(sub_mb_type)) {
380                 get_lowest_part_y(h, refs, n, 8, y_offset,
381                                   IS_DIR(sub_mb_type, 0, 0),
382                                   IS_DIR(sub_mb_type, 0, 1),
383                                   nrefs);
384             } else if (IS_SUB_8X4(sub_mb_type)) {
385                 get_lowest_part_y(h, refs, n, 4, y_offset,
386                                   IS_DIR(sub_mb_type, 0, 0),
387                                   IS_DIR(sub_mb_type, 0, 1),
388                                   nrefs);
389                 get_lowest_part_y(h, refs, n + 2, 4, y_offset + 4,
390                                   IS_DIR(sub_mb_type, 0, 0),
391                                   IS_DIR(sub_mb_type, 0, 1),
392                                   nrefs);
393             } else if (IS_SUB_4X8(sub_mb_type)) {
394                 get_lowest_part_y(h, refs, n, 8, y_offset,
395                                   IS_DIR(sub_mb_type, 0, 0),
396                                   IS_DIR(sub_mb_type, 0, 1),
397                                   nrefs);
398                 get_lowest_part_y(h, refs, n + 1, 8, y_offset,
399                                   IS_DIR(sub_mb_type, 0, 0),
400                                   IS_DIR(sub_mb_type, 0, 1),
401                                   nrefs);
402             } else {
403                 int j;
404                 assert(IS_SUB_4X4(sub_mb_type));
405                 for (j = 0; j < 4; j++) {
406                     int sub_y_offset = y_offset + 2 * (j & 2);
407                     get_lowest_part_y(h, refs, n + j, 4, sub_y_offset,
408                                       IS_DIR(sub_mb_type, 0, 0),
409                                       IS_DIR(sub_mb_type, 0, 1),
410                                       nrefs);
411                 }
412             }
413         }
414     }
415
416     for (list = h->list_count - 1; list >= 0; list--)
417         for (ref = 0; ref < 48 && nrefs[list]; ref++) {
418             int row = refs[list][ref];
419             if (row >= 0) {
420                 Picture *ref_pic      = &h->ref_list[list][ref];
421                 int ref_field         = ref_pic->f.reference - 1;
422                 int ref_field_picture = ref_pic->field_picture;
423                 int pic_height        = 16 * s->mb_height >> ref_field_picture;
424
425                 row <<= MB_MBAFF;
426                 nrefs[list]--;
427
428                 if (!FIELD_PICTURE && ref_field_picture) { // frame referencing two fields
429                     ff_thread_await_progress(&ref_pic->f,
430                                              FFMIN((row >> 1) - !(row & 1),
431                                                    pic_height - 1),
432                                              1);
433                     ff_thread_await_progress(&ref_pic->f,
434                                              FFMIN((row >> 1), pic_height - 1),
435                                              0);
436                 } else if (FIELD_PICTURE && !ref_field_picture) { // field referencing one field of a frame
437                     ff_thread_await_progress(&ref_pic->f,
438                                              FFMIN(row * 2 + ref_field,
439                                                    pic_height - 1),
440                                              0);
441                 } else if (FIELD_PICTURE) {
442                     ff_thread_await_progress(&ref_pic->f,
443                                              FFMIN(row, pic_height - 1),
444                                              ref_field);
445                 } else {
446                     ff_thread_await_progress(&ref_pic->f,
447                                              FFMIN(row, pic_height - 1),
448                                              0);
449                 }
450             }
451         }
452 }
453
454 static av_always_inline void mc_dir_part(H264Context *h, Picture *pic,
455                                          int n, int square, int height,
456                                          int delta, int list,
457                                          uint8_t *dest_y, uint8_t *dest_cb,
458                                          uint8_t *dest_cr,
459                                          int src_x_offset, int src_y_offset,
460                                          qpel_mc_func *qpix_op,
461                                          h264_chroma_mc_func chroma_op,
462                                          int pixel_shift, int chroma_idc)
463 {
464     MpegEncContext *const s = &h->s;
465     const int mx      = h->mv_cache[list][scan8[n]][0] + src_x_offset * 8;
466     int my            = h->mv_cache[list][scan8[n]][1] + src_y_offset * 8;
467     const int luma_xy = (mx & 3) + ((my & 3) << 2);
468     int offset        = ((mx >> 2) << pixel_shift) + (my >> 2) * h->mb_linesize;
469     uint8_t *src_y    = pic->f.data[0] + offset;
470     uint8_t *src_cb, *src_cr;
471     int extra_width  = h->emu_edge_width;
472     int extra_height = h->emu_edge_height;
473     int emu = 0;
474     const int full_mx    = mx >> 2;
475     const int full_my    = my >> 2;
476     const int pic_width  = 16 * s->mb_width;
477     const int pic_height = 16 * s->mb_height >> MB_FIELD;
478     int ysh;
479
480     if (mx & 7)
481         extra_width -= 3;
482     if (my & 7)
483         extra_height -= 3;
484
485     if (full_mx                <          0 - extra_width  ||
486         full_my                <          0 - extra_height ||
487         full_mx + 16 /*FIXME*/ > pic_width  + extra_width  ||
488         full_my + 16 /*FIXME*/ > pic_height + extra_height) {
489         s->dsp.emulated_edge_mc(s->edge_emu_buffer,
490                                 src_y - (2 << pixel_shift) - 2 * h->mb_linesize,
491                                 h->mb_linesize,
492                                 16 + 5, 16 + 5 /*FIXME*/, full_mx - 2,
493                                 full_my - 2, pic_width, pic_height);
494         src_y = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
495         emu   = 1;
496     }
497
498     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); // FIXME try variable height perhaps?
499     if (!square)
500         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
501
502     if (CONFIG_GRAY && s->flags & CODEC_FLAG_GRAY)
503         return;
504
505     if (chroma_idc == 3 /* yuv444 */) {
506         src_cb = pic->f.data[1] + offset;
507         if (emu) {
508             s->dsp.emulated_edge_mc(s->edge_emu_buffer,
509                                     src_cb - (2 << pixel_shift) - 2 * h->mb_linesize,
510                                     h->mb_linesize,
511                                     16 + 5, 16 + 5 /*FIXME*/,
512                                     full_mx - 2, full_my - 2,
513                                     pic_width, pic_height);
514             src_cb = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
515         }
516         qpix_op[luma_xy](dest_cb, src_cb, h->mb_linesize); // FIXME try variable height perhaps?
517         if (!square)
518             qpix_op[luma_xy](dest_cb + delta, src_cb + delta, h->mb_linesize);
519
520         src_cr = pic->f.data[2] + offset;
521         if (emu) {
522             s->dsp.emulated_edge_mc(s->edge_emu_buffer,
523                                     src_cr - (2 << pixel_shift) - 2 * h->mb_linesize,
524                                     h->mb_linesize,
525                                     16 + 5, 16 + 5 /*FIXME*/,
526                                     full_mx - 2, full_my - 2,
527                                     pic_width, pic_height);
528             src_cr = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
529         }
530         qpix_op[luma_xy](dest_cr, src_cr, h->mb_linesize); // FIXME try variable height perhaps?
531         if (!square)
532             qpix_op[luma_xy](dest_cr + delta, src_cr + delta, h->mb_linesize);
533         return;
534     }
535
536     ysh = 3 - (chroma_idc == 2 /* yuv422 */);
537     if (chroma_idc == 1 /* yuv420 */ && MB_FIELD) {
538         // chroma offset when predicting from a field of opposite parity
539         my  += 2 * ((s->mb_y & 1) - (pic->f.reference - 1));
540         emu |= (my >> 3) < 0 || (my >> 3) + 8 >= (pic_height >> 1);
541     }
542
543     src_cb = pic->f.data[1] + ((mx >> 3) << pixel_shift) +
544              (my >> ysh) * h->mb_uvlinesize;
545     src_cr = pic->f.data[2] + ((mx >> 3) << pixel_shift) +
546              (my >> ysh) * h->mb_uvlinesize;
547
548     if (emu) {
549         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize,
550                                 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
551                                 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
552         src_cb = s->edge_emu_buffer;
553     }
554     chroma_op(dest_cb, src_cb, h->mb_uvlinesize,
555               height >> (chroma_idc == 1 /* yuv420 */),
556               mx & 7, (my << (chroma_idc == 2 /* yuv422 */)) & 7);
557
558     if (emu) {
559         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize,
560                                 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
561                                 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
562         src_cr = s->edge_emu_buffer;
563     }
564     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, height >> (chroma_idc == 1 /* yuv420 */),
565               mx & 7, (my << (chroma_idc == 2 /* yuv422 */)) & 7);
566 }
567
568 static av_always_inline void mc_part_std(H264Context *h, int n, int square,
569                                          int height, int delta,
570                                          uint8_t *dest_y, uint8_t *dest_cb,
571                                          uint8_t *dest_cr,
572                                          int x_offset, int y_offset,
573                                          qpel_mc_func *qpix_put,
574                                          h264_chroma_mc_func chroma_put,
575                                          qpel_mc_func *qpix_avg,
576                                          h264_chroma_mc_func chroma_avg,
577                                          int list0, int list1,
578                                          int pixel_shift, int chroma_idc)
579 {
580     MpegEncContext *const s       = &h->s;
581     qpel_mc_func *qpix_op         = qpix_put;
582     h264_chroma_mc_func chroma_op = chroma_put;
583
584     dest_y += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
585     if (chroma_idc == 3 /* yuv444 */) {
586         dest_cb += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
587         dest_cr += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
588     } else if (chroma_idc == 2 /* yuv422 */) {
589         dest_cb += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
590         dest_cr += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
591     } else { /* yuv420 */
592         dest_cb += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
593         dest_cr += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
594     }
595     x_offset += 8 * s->mb_x;
596     y_offset += 8 * (s->mb_y >> MB_FIELD);
597
598     if (list0) {
599         Picture *ref = &h->ref_list[0][h->ref_cache[0][scan8[n]]];
600         mc_dir_part(h, ref, n, square, height, delta, 0,
601                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
602                     qpix_op, chroma_op, pixel_shift, chroma_idc);
603
604         qpix_op   = qpix_avg;
605         chroma_op = chroma_avg;
606     }
607
608     if (list1) {
609         Picture *ref = &h->ref_list[1][h->ref_cache[1][scan8[n]]];
610         mc_dir_part(h, ref, n, square, height, delta, 1,
611                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
612                     qpix_op, chroma_op, pixel_shift, chroma_idc);
613     }
614 }
615
616 static av_always_inline void mc_part_weighted(H264Context *h, int n, int square,
617                                               int height, int delta,
618                                               uint8_t *dest_y, uint8_t *dest_cb,
619                                               uint8_t *dest_cr,
620                                               int x_offset, int y_offset,
621                                               qpel_mc_func *qpix_put,
622                                               h264_chroma_mc_func chroma_put,
623                                               h264_weight_func luma_weight_op,
624                                               h264_weight_func chroma_weight_op,
625                                               h264_biweight_func luma_weight_avg,
626                                               h264_biweight_func chroma_weight_avg,
627                                               int list0, int list1,
628                                               int pixel_shift, int chroma_idc)
629 {
630     MpegEncContext *const s = &h->s;
631     int chroma_height;
632
633     dest_y += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
634     if (chroma_idc == 3 /* yuv444 */) {
635         chroma_height     = height;
636         chroma_weight_avg = luma_weight_avg;
637         chroma_weight_op  = luma_weight_op;
638         dest_cb += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
639         dest_cr += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
640     } else if (chroma_idc == 2 /* yuv422 */) {
641         chroma_height = height;
642         dest_cb      += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
643         dest_cr      += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
644     } else { /* yuv420 */
645         chroma_height = height >> 1;
646         dest_cb      += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
647         dest_cr      += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
648     }
649     x_offset += 8 * s->mb_x;
650     y_offset += 8 * (s->mb_y >> MB_FIELD);
651
652     if (list0 && list1) {
653         /* don't optimize for luma-only case, since B-frames usually
654          * use implicit weights => chroma too. */
655         uint8_t *tmp_cb = s->obmc_scratchpad;
656         uint8_t *tmp_cr = s->obmc_scratchpad + (16 << pixel_shift);
657         uint8_t *tmp_y  = s->obmc_scratchpad + 16 * h->mb_uvlinesize;
658         int refn0       = h->ref_cache[0][scan8[n]];
659         int refn1       = h->ref_cache[1][scan8[n]];
660
661         mc_dir_part(h, &h->ref_list[0][refn0], n, square, height, delta, 0,
662                     dest_y, dest_cb, dest_cr,
663                     x_offset, y_offset, qpix_put, chroma_put,
664                     pixel_shift, chroma_idc);
665         mc_dir_part(h, &h->ref_list[1][refn1], n, square, height, delta, 1,
666                     tmp_y, tmp_cb, tmp_cr,
667                     x_offset, y_offset, qpix_put, chroma_put,
668                     pixel_shift, chroma_idc);
669
670         if (h->use_weight == 2) {
671             int weight0 = h->implicit_weight[refn0][refn1][s->mb_y & 1];
672             int weight1 = 64 - weight0;
673             luma_weight_avg(dest_y, tmp_y, h->mb_linesize,
674                             height, 5, weight0, weight1, 0);
675             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize,
676                               chroma_height, 5, weight0, weight1, 0);
677             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize,
678                               chroma_height, 5, weight0, weight1, 0);
679         } else {
680             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, height,
681                             h->luma_log2_weight_denom,
682                             h->luma_weight[refn0][0][0],
683                             h->luma_weight[refn1][1][0],
684                             h->luma_weight[refn0][0][1] +
685                             h->luma_weight[refn1][1][1]);
686             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, chroma_height,
687                               h->chroma_log2_weight_denom,
688                               h->chroma_weight[refn0][0][0][0],
689                               h->chroma_weight[refn1][1][0][0],
690                               h->chroma_weight[refn0][0][0][1] +
691                               h->chroma_weight[refn1][1][0][1]);
692             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, chroma_height,
693                               h->chroma_log2_weight_denom,
694                               h->chroma_weight[refn0][0][1][0],
695                               h->chroma_weight[refn1][1][1][0],
696                               h->chroma_weight[refn0][0][1][1] +
697                               h->chroma_weight[refn1][1][1][1]);
698         }
699     } else {
700         int list     = list1 ? 1 : 0;
701         int refn     = h->ref_cache[list][scan8[n]];
702         Picture *ref = &h->ref_list[list][refn];
703         mc_dir_part(h, ref, n, square, height, delta, list,
704                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
705                     qpix_put, chroma_put, pixel_shift, chroma_idc);
706
707         luma_weight_op(dest_y, h->mb_linesize, height,
708                        h->luma_log2_weight_denom,
709                        h->luma_weight[refn][list][0],
710                        h->luma_weight[refn][list][1]);
711         if (h->use_weight_chroma) {
712             chroma_weight_op(dest_cb, h->mb_uvlinesize, chroma_height,
713                              h->chroma_log2_weight_denom,
714                              h->chroma_weight[refn][list][0][0],
715                              h->chroma_weight[refn][list][0][1]);
716             chroma_weight_op(dest_cr, h->mb_uvlinesize, chroma_height,
717                              h->chroma_log2_weight_denom,
718                              h->chroma_weight[refn][list][1][0],
719                              h->chroma_weight[refn][list][1][1]);
720         }
721     }
722 }
723
724 static av_always_inline void mc_part(H264Context *h, int n, int square,
725                                      int height, int delta,
726                                      uint8_t *dest_y, uint8_t *dest_cb,
727                                      uint8_t *dest_cr,
728                                      int x_offset, int y_offset,
729                                      qpel_mc_func *qpix_put,
730                                      h264_chroma_mc_func chroma_put,
731                                      qpel_mc_func *qpix_avg,
732                                      h264_chroma_mc_func chroma_avg,
733                                      h264_weight_func *weight_op,
734                                      h264_biweight_func *weight_avg,
735                                      int list0, int list1,
736                                      int pixel_shift, int chroma_idc)
737 {
738     if ((h->use_weight == 2 && list0 && list1 &&
739          (h->implicit_weight[h->ref_cache[0][scan8[n]]][h->ref_cache[1][scan8[n]]][h->s.mb_y & 1] != 32)) ||
740         h->use_weight == 1)
741         mc_part_weighted(h, n, square, height, delta, dest_y, dest_cb, dest_cr,
742                          x_offset, y_offset, qpix_put, chroma_put,
743                          weight_op[0], weight_op[1], weight_avg[0],
744                          weight_avg[1], list0, list1, pixel_shift, chroma_idc);
745     else
746         mc_part_std(h, n, square, height, delta, dest_y, dest_cb, dest_cr,
747                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg,
748                     chroma_avg, list0, list1, pixel_shift, chroma_idc);
749 }
750
751 static av_always_inline void prefetch_motion(H264Context *h, int list,
752                                              int pixel_shift, int chroma_idc)
753 {
754     /* fetch pixels for estimated mv 4 macroblocks ahead
755      * optimized for 64byte cache lines */
756     MpegEncContext *const s = &h->s;
757     const int refn = h->ref_cache[list][scan8[0]];
758     if (refn >= 0) {
759         const int mx  = (h->mv_cache[list][scan8[0]][0] >> 2) + 16 * s->mb_x + 8;
760         const int my  = (h->mv_cache[list][scan8[0]][1] >> 2) + 16 * s->mb_y;
761         uint8_t **src = h->ref_list[list][refn].f.data;
762         int off       = (mx << pixel_shift) +
763                         (my + (s->mb_x & 3) * 4) * h->mb_linesize +
764                         (64 << pixel_shift);
765         s->dsp.prefetch(src[0] + off, s->linesize, 4);
766         if (chroma_idc == 3 /* yuv444 */) {
767             s->dsp.prefetch(src[1] + off, s->linesize, 4);
768             s->dsp.prefetch(src[2] + off, s->linesize, 4);
769         } else {
770             off= (((mx>>1)+64)<<pixel_shift) + ((my>>1) + (s->mb_x&7))*s->uvlinesize;
771             s->dsp.prefetch(src[1] + off, src[2] - src[1], 2);
772         }
773     }
774 }
775
776 static av_always_inline void hl_motion(H264Context *h, uint8_t *dest_y,
777                                        uint8_t *dest_cb, uint8_t *dest_cr,
778                                        qpel_mc_func(*qpix_put)[16],
779                                        h264_chroma_mc_func(*chroma_put),
780                                        qpel_mc_func(*qpix_avg)[16],
781                                        h264_chroma_mc_func(*chroma_avg),
782                                        h264_weight_func *weight_op,
783                                        h264_biweight_func *weight_avg,
784                                        int pixel_shift, int chroma_idc)
785 {
786     MpegEncContext *const s = &h->s;
787     const int mb_xy   = h->mb_xy;
788     const int mb_type = s->current_picture.f.mb_type[mb_xy];
789
790     assert(IS_INTER(mb_type));
791
792     if (HAVE_THREADS && (s->avctx->active_thread_type & FF_THREAD_FRAME))
793         await_references(h);
794     prefetch_motion(h, 0, pixel_shift, chroma_idc);
795
796     if (IS_16X16(mb_type)) {
797         mc_part(h, 0, 1, 16, 0, dest_y, dest_cb, dest_cr, 0, 0,
798                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
799                 weight_op, weight_avg,
800                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
801                 pixel_shift, chroma_idc);
802     } else if (IS_16X8(mb_type)) {
803         mc_part(h, 0, 0, 8, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 0,
804                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
805                 weight_op, weight_avg,
806                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
807                 pixel_shift, chroma_idc);
808         mc_part(h, 8, 0, 8, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 4,
809                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
810                 weight_op, weight_avg,
811                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
812                 pixel_shift, chroma_idc);
813     } else if (IS_8X16(mb_type)) {
814         mc_part(h, 0, 0, 16, 8 * h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
815                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
816                 &weight_op[1], &weight_avg[1],
817                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
818                 pixel_shift, chroma_idc);
819         mc_part(h, 4, 0, 16, 8 * h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
820                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
821                 &weight_op[1], &weight_avg[1],
822                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
823                 pixel_shift, chroma_idc);
824     } else {
825         int i;
826
827         assert(IS_8X8(mb_type));
828
829         for (i = 0; i < 4; i++) {
830             const int sub_mb_type = h->sub_mb_type[i];
831             const int n  = 4 * i;
832             int x_offset = (i & 1) << 2;
833             int y_offset = (i & 2) << 1;
834
835             if (IS_SUB_8X8(sub_mb_type)) {
836                 mc_part(h, n, 1, 8, 0, dest_y, dest_cb, dest_cr,
837                         x_offset, y_offset,
838                         qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
839                         &weight_op[1], &weight_avg[1],
840                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
841                         pixel_shift, chroma_idc);
842             } else if (IS_SUB_8X4(sub_mb_type)) {
843                 mc_part(h, n, 0, 4, 4 << pixel_shift, dest_y, dest_cb, dest_cr,
844                         x_offset, y_offset,
845                         qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
846                         &weight_op[1], &weight_avg[1],
847                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
848                         pixel_shift, chroma_idc);
849                 mc_part(h, n + 2, 0, 4, 4 << pixel_shift,
850                         dest_y, dest_cb, dest_cr, x_offset, y_offset + 2,
851                         qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
852                         &weight_op[1], &weight_avg[1],
853                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
854                         pixel_shift, chroma_idc);
855             } else if (IS_SUB_4X8(sub_mb_type)) {
856                 mc_part(h, n, 0, 8, 4 * h->mb_linesize,
857                         dest_y, dest_cb, dest_cr, x_offset, y_offset,
858                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
859                         &weight_op[2], &weight_avg[2],
860                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
861                         pixel_shift, chroma_idc);
862                 mc_part(h, n + 1, 0, 8, 4 * h->mb_linesize,
863                         dest_y, dest_cb, dest_cr, x_offset + 2, y_offset,
864                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
865                         &weight_op[2], &weight_avg[2],
866                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
867                         pixel_shift, chroma_idc);
868             } else {
869                 int j;
870                 assert(IS_SUB_4X4(sub_mb_type));
871                 for (j = 0; j < 4; j++) {
872                     int sub_x_offset = x_offset + 2 * (j & 1);
873                     int sub_y_offset = y_offset + (j & 2);
874                     mc_part(h, n + j, 1, 4, 0,
875                             dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
876                             qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
877                             &weight_op[2], &weight_avg[2],
878                             IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
879                             pixel_shift, chroma_idc);
880                 }
881             }
882         }
883     }
884
885     prefetch_motion(h, 1, pixel_shift, chroma_idc);
886 }
887
888 static av_always_inline void hl_motion_420(H264Context *h, uint8_t *dest_y,
889                                            uint8_t *dest_cb, uint8_t *dest_cr,
890                                            qpel_mc_func(*qpix_put)[16],
891                                            h264_chroma_mc_func(*chroma_put),
892                                            qpel_mc_func(*qpix_avg)[16],
893                                            h264_chroma_mc_func(*chroma_avg),
894                                            h264_weight_func *weight_op,
895                                            h264_biweight_func *weight_avg,
896                                            int pixel_shift)
897 {
898     hl_motion(h, dest_y, dest_cb, dest_cr, qpix_put, chroma_put,
899               qpix_avg, chroma_avg, weight_op, weight_avg, pixel_shift, 1);
900 }
901
902 static av_always_inline void hl_motion_422(H264Context *h, uint8_t *dest_y,
903                                            uint8_t *dest_cb, uint8_t *dest_cr,
904                                            qpel_mc_func(*qpix_put)[16],
905                                            h264_chroma_mc_func(*chroma_put),
906                                            qpel_mc_func(*qpix_avg)[16],
907                                            h264_chroma_mc_func(*chroma_avg),
908                                            h264_weight_func *weight_op,
909                                            h264_biweight_func *weight_avg,
910                                            int pixel_shift)
911 {
912     hl_motion(h, dest_y, dest_cb, dest_cr, qpix_put, chroma_put,
913               qpix_avg, chroma_avg, weight_op, weight_avg, pixel_shift, 2);
914 }
915
916 static void free_tables(H264Context *h, int free_rbsp)
917 {
918     int i;
919     H264Context *hx;
920
921     av_freep(&h->intra4x4_pred_mode);
922     av_freep(&h->chroma_pred_mode_table);
923     av_freep(&h->cbp_table);
924     av_freep(&h->mvd_table[0]);
925     av_freep(&h->mvd_table[1]);
926     av_freep(&h->direct_table);
927     av_freep(&h->non_zero_count);
928     av_freep(&h->slice_table_base);
929     h->slice_table = NULL;
930     av_freep(&h->list_counts);
931
932     av_freep(&h->mb2b_xy);
933     av_freep(&h->mb2br_xy);
934
935     for (i = 0; i < MAX_THREADS; i++) {
936         hx = h->thread_context[i];
937         if (!hx)
938             continue;
939         av_freep(&hx->top_borders[1]);
940         av_freep(&hx->top_borders[0]);
941         av_freep(&hx->s.obmc_scratchpad);
942         if (free_rbsp) {
943             av_freep(&hx->rbsp_buffer[1]);
944             av_freep(&hx->rbsp_buffer[0]);
945             hx->rbsp_buffer_size[0] = 0;
946             hx->rbsp_buffer_size[1] = 0;
947         }
948         if (i)
949             av_freep(&h->thread_context[i]);
950     }
951 }
952
953 static void init_dequant8_coeff_table(H264Context *h)
954 {
955     int i, j, q, x;
956     const int max_qp = 51 + 6 * (h->sps.bit_depth_luma - 8);
957
958     for (i = 0; i < 6; i++) {
959         h->dequant8_coeff[i] = h->dequant8_buffer[i];
960         for (j = 0; j < i; j++)
961             if (!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i],
962                         64 * sizeof(uint8_t))) {
963                 h->dequant8_coeff[i] = h->dequant8_buffer[j];
964                 break;
965             }
966         if (j < i)
967             continue;
968
969         for (q = 0; q < max_qp + 1; q++) {
970             int shift = div6[q];
971             int idx   = rem6[q];
972             for (x = 0; x < 64; x++)
973                 h->dequant8_coeff[i][q][(x >> 3) | ((x & 7) << 3)] =
974                     ((uint32_t)dequant8_coeff_init[idx][dequant8_coeff_init_scan[((x >> 1) & 12) | (x & 3)]] *
975                      h->pps.scaling_matrix8[i][x]) << shift;
976         }
977     }
978 }
979
980 static void init_dequant4_coeff_table(H264Context *h)
981 {
982     int i, j, q, x;
983     const int max_qp = 51 + 6 * (h->sps.bit_depth_luma - 8);
984     for (i = 0; i < 6; i++) {
985         h->dequant4_coeff[i] = h->dequant4_buffer[i];
986         for (j = 0; j < i; j++)
987             if (!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i],
988                         16 * sizeof(uint8_t))) {
989                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
990                 break;
991             }
992         if (j < i)
993             continue;
994
995         for (q = 0; q < max_qp + 1; q++) {
996             int shift = div6[q] + 2;
997             int idx   = rem6[q];
998             for (x = 0; x < 16; x++)
999                 h->dequant4_coeff[i][q][(x >> 2) | ((x << 2) & 0xF)] =
1000                     ((uint32_t)dequant4_coeff_init[idx][(x & 1) + ((x >> 2) & 1)] *
1001                      h->pps.scaling_matrix4[i][x]) << shift;
1002         }
1003     }
1004 }
1005
1006 static void init_dequant_tables(H264Context *h)
1007 {
1008     int i, x;
1009     init_dequant4_coeff_table(h);
1010     if (h->pps.transform_8x8_mode)
1011         init_dequant8_coeff_table(h);
1012     if (h->sps.transform_bypass) {
1013         for (i = 0; i < 6; i++)
1014             for (x = 0; x < 16; x++)
1015                 h->dequant4_coeff[i][0][x] = 1 << 6;
1016         if (h->pps.transform_8x8_mode)
1017             for (i = 0; i < 6; i++)
1018                 for (x = 0; x < 64; x++)
1019                     h->dequant8_coeff[i][0][x] = 1 << 6;
1020     }
1021 }
1022
1023 int ff_h264_alloc_tables(H264Context *h)
1024 {
1025     MpegEncContext *const s = &h->s;
1026     const int big_mb_num    = s->mb_stride * (s->mb_height + 1);
1027     const int row_mb_num    = 2*s->mb_stride*FFMAX(s->avctx->thread_count, 1);
1028     int x, y;
1029
1030     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode,
1031                       row_mb_num * 8 * sizeof(uint8_t), fail)
1032     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count,
1033                       big_mb_num * 48 * sizeof(uint8_t), fail)
1034     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base,
1035                       (big_mb_num + s->mb_stride) * sizeof(*h->slice_table_base), fail)
1036     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table,
1037                       big_mb_num * sizeof(uint16_t), fail)
1038     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table,
1039                       big_mb_num * sizeof(uint8_t), fail)
1040     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0],
1041                       16 * row_mb_num * sizeof(uint8_t), fail);
1042     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1],
1043                       16 * row_mb_num * sizeof(uint8_t), fail);
1044     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table,
1045                       4 * big_mb_num * sizeof(uint8_t), fail);
1046     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts,
1047                       big_mb_num * sizeof(uint8_t), fail)
1048
1049     memset(h->slice_table_base, -1,
1050            (big_mb_num + s->mb_stride) * sizeof(*h->slice_table_base));
1051     h->slice_table = h->slice_table_base + s->mb_stride * 2 + 1;
1052
1053     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy,
1054                       big_mb_num * sizeof(uint32_t), fail);
1055     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy,
1056                       big_mb_num * sizeof(uint32_t), fail);
1057     for (y = 0; y < s->mb_height; y++)
1058         for (x = 0; x < s->mb_width; x++) {
1059             const int mb_xy = x + y * s->mb_stride;
1060             const int b_xy  = 4 * x + 4 * y * h->b_stride;
1061
1062             h->mb2b_xy[mb_xy]  = b_xy;
1063             h->mb2br_xy[mb_xy] = 8 * (FMO ? mb_xy : (mb_xy % (2 * s->mb_stride)));
1064         }
1065
1066     s->obmc_scratchpad = NULL;
1067
1068     if (!h->dequant4_coeff[0])
1069         init_dequant_tables(h);
1070
1071     return 0;
1072
1073 fail:
1074     free_tables(h, 1);
1075     return -1;
1076 }
1077
1078 /**
1079  * Mimic alloc_tables(), but for every context thread.
1080  */
1081 static void clone_tables(H264Context *dst, H264Context *src, int i)
1082 {
1083     MpegEncContext *const s     = &src->s;
1084     dst->intra4x4_pred_mode     = src->intra4x4_pred_mode + i * 8 * 2 * s->mb_stride;
1085     dst->non_zero_count         = src->non_zero_count;
1086     dst->slice_table            = src->slice_table;
1087     dst->cbp_table              = src->cbp_table;
1088     dst->mb2b_xy                = src->mb2b_xy;
1089     dst->mb2br_xy               = src->mb2br_xy;
1090     dst->chroma_pred_mode_table = src->chroma_pred_mode_table;
1091     dst->mvd_table[0]           = src->mvd_table[0] + i * 8 * 2 * s->mb_stride;
1092     dst->mvd_table[1]           = src->mvd_table[1] + i * 8 * 2 * s->mb_stride;
1093     dst->direct_table           = src->direct_table;
1094     dst->list_counts            = src->list_counts;
1095     dst->s.obmc_scratchpad      = NULL;
1096     ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma,
1097                       src->sps.chroma_format_idc);
1098 }
1099
1100 /**
1101  * Init context
1102  * Allocate buffers which are not shared amongst multiple threads.
1103  */
1104 static int context_init(H264Context *h)
1105 {
1106     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0],
1107                       h->s.mb_width * 16 * 3 * sizeof(uint8_t) * 2, fail)
1108     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1],
1109                       h->s.mb_width * 16 * 3 * sizeof(uint8_t) * 2, fail)
1110
1111     h->ref_cache[0][scan8[5]  + 1] =
1112     h->ref_cache[0][scan8[7]  + 1] =
1113     h->ref_cache[0][scan8[13] + 1] =
1114     h->ref_cache[1][scan8[5]  + 1] =
1115     h->ref_cache[1][scan8[7]  + 1] =
1116     h->ref_cache[1][scan8[13] + 1] = PART_NOT_AVAILABLE;
1117
1118     return 0;
1119
1120 fail:
1121     return -1; // free_tables will clean up for us
1122 }
1123
1124 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
1125
1126 static av_cold void common_init(H264Context *h)
1127 {
1128     MpegEncContext *const s = &h->s;
1129
1130     s->width    = s->avctx->width;
1131     s->height   = s->avctx->height;
1132     s->codec_id = s->avctx->codec->id;
1133
1134     s->avctx->bits_per_raw_sample = 8;
1135     h->cur_chroma_format_idc = 1;
1136
1137     ff_h264dsp_init(&h->h264dsp,
1138                     s->avctx->bits_per_raw_sample, h->cur_chroma_format_idc);
1139     ff_h264_pred_init(&h->hpc, s->codec_id,
1140                       s->avctx->bits_per_raw_sample, h->cur_chroma_format_idc);
1141
1142     h->dequant_coeff_pps = -1;
1143     s->unrestricted_mv   = 1;
1144
1145     s->dsp.dct_bits = 16;
1146     /* needed so that IDCT permutation is known early */
1147     ff_dsputil_init(&s->dsp, s->avctx);
1148
1149     memset(h->pps.scaling_matrix4, 16, 6 * 16 * sizeof(uint8_t));
1150     memset(h->pps.scaling_matrix8, 16, 2 * 64 * sizeof(uint8_t));
1151 }
1152
1153 int ff_h264_decode_extradata(H264Context *h, const uint8_t *buf, int size)
1154 {
1155     AVCodecContext *avctx = h->s.avctx;
1156
1157     if (!buf || size <= 0)
1158         return -1;
1159
1160     if (buf[0] == 1) {
1161         int i, cnt, nalsize;
1162         const unsigned char *p = buf;
1163
1164         h->is_avc = 1;
1165
1166         if (size < 7) {
1167             av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
1168             return -1;
1169         }
1170         /* sps and pps in the avcC always have length coded with 2 bytes,
1171          * so put a fake nal_length_size = 2 while parsing them */
1172         h->nal_length_size = 2;
1173         // Decode sps from avcC
1174         cnt = *(p + 5) & 0x1f; // Number of sps
1175         p  += 6;
1176         for (i = 0; i < cnt; i++) {
1177             nalsize = AV_RB16(p) + 2;
1178             if(nalsize > size - (p-buf))
1179                 return -1;
1180             if (decode_nal_units(h, p, nalsize) < 0) {
1181                 av_log(avctx, AV_LOG_ERROR,
1182                        "Decoding sps %d from avcC failed\n", i);
1183                 return -1;
1184             }
1185             p += nalsize;
1186         }
1187         // Decode pps from avcC
1188         cnt = *(p++); // Number of pps
1189         for (i = 0; i < cnt; i++) {
1190             nalsize = AV_RB16(p) + 2;
1191             if(nalsize > size - (p-buf))
1192                 return -1;
1193             if (decode_nal_units(h, p, nalsize) < 0) {
1194                 av_log(avctx, AV_LOG_ERROR,
1195                        "Decoding pps %d from avcC failed\n", i);
1196                 return -1;
1197             }
1198             p += nalsize;
1199         }
1200         // Now store right nal length size, that will be used to parse all other nals
1201         h->nal_length_size = (buf[4] & 0x03) + 1;
1202     } else {
1203         h->is_avc = 0;
1204         if (decode_nal_units(h, buf, size) < 0)
1205             return -1;
1206     }
1207     return size;
1208 }
1209
1210 av_cold int ff_h264_decode_init(AVCodecContext *avctx)
1211 {
1212     H264Context *h = avctx->priv_data;
1213     MpegEncContext *const s = &h->s;
1214     int i;
1215
1216     ff_MPV_decode_defaults(s);
1217
1218     s->avctx = avctx;
1219     common_init(h);
1220
1221     s->out_format      = FMT_H264;
1222     s->workaround_bugs = avctx->workaround_bugs;
1223
1224     /* set defaults */
1225     // s->decode_mb = ff_h263_decode_mb;
1226     s->quarter_sample = 1;
1227     if (!avctx->has_b_frames)
1228         s->low_delay = 1;
1229
1230     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1231
1232     ff_h264_decode_init_vlc();
1233
1234     h->pixel_shift = 0;
1235     h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1236
1237     h->thread_context[0] = h;
1238     h->outputed_poc      = h->next_outputed_poc = INT_MIN;
1239     for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
1240         h->last_pocs[i] = INT_MIN;
1241     h->prev_poc_msb = 1 << 16;
1242     h->prev_frame_num = -1;
1243     h->x264_build   = -1;
1244     ff_h264_reset_sei(h);
1245     if (avctx->codec_id == CODEC_ID_H264) {
1246         if (avctx->ticks_per_frame == 1)
1247             s->avctx->time_base.den *= 2;
1248         avctx->ticks_per_frame = 2;
1249     }
1250
1251     if (avctx->extradata_size > 0 && avctx->extradata &&
1252         ff_h264_decode_extradata(h, avctx->extradata, avctx->extradata_size) < 0)
1253         return -1;
1254
1255     if (h->sps.bitstream_restriction_flag &&
1256         s->avctx->has_b_frames < h->sps.num_reorder_frames) {
1257         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1258         s->low_delay           = 0;
1259     }
1260
1261     return 0;
1262 }
1263
1264 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b) + (size))))
1265
1266 static void copy_picture_range(Picture **to, Picture **from, int count,
1267                                MpegEncContext *new_base,
1268                                MpegEncContext *old_base)
1269 {
1270     int i;
1271
1272     for (i = 0; i < count; i++) {
1273         assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1274                 IN_RANGE(from[i], old_base->picture,
1275                          sizeof(Picture) * old_base->picture_count) ||
1276                 !from[i]));
1277         to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1278     }
1279 }
1280
1281 static void copy_parameter_set(void **to, void **from, int count, int size)
1282 {
1283     int i;
1284
1285     for (i = 0; i < count; i++) {
1286         if (to[i] && !from[i])
1287             av_freep(&to[i]);
1288         else if (from[i] && !to[i])
1289             to[i] = av_malloc(size);
1290
1291         if (from[i])
1292             memcpy(to[i], from[i], size);
1293     }
1294 }
1295
1296 static int decode_init_thread_copy(AVCodecContext *avctx)
1297 {
1298     H264Context *h = avctx->priv_data;
1299
1300     if (!avctx->internal->is_copy)
1301         return 0;
1302     memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1303     memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1304
1305     return 0;
1306 }
1307
1308 #define copy_fields(to, from, start_field, end_field)                   \
1309     memcpy(&to->start_field, &from->start_field,                        \
1310            (char *)&to->end_field - (char *)&to->start_field)
1311
1312 static int decode_update_thread_context(AVCodecContext *dst,
1313                                         const AVCodecContext *src)
1314 {
1315     H264Context *h = dst->priv_data, *h1 = src->priv_data;
1316     MpegEncContext *const s = &h->s, *const s1 = &h1->s;
1317     int inited = s->context_initialized, err;
1318     int i;
1319
1320     if (dst == src)
1321         return 0;
1322
1323     err = ff_mpeg_update_thread_context(dst, src);
1324     if (err)
1325         return err;
1326
1327     // FIXME handle width/height changing
1328     if (!inited) {
1329         for (i = 0; i < MAX_SPS_COUNT; i++)
1330             av_freep(h->sps_buffers + i);
1331
1332         for (i = 0; i < MAX_PPS_COUNT; i++)
1333             av_freep(h->pps_buffers + i);
1334
1335         // copy all fields after MpegEnc
1336         memcpy(&h->s + 1, &h1->s + 1,
1337                sizeof(H264Context) - sizeof(MpegEncContext));
1338         memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1339         memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1340
1341         if (s1->context_initialized) {
1342         if (ff_h264_alloc_tables(h) < 0) {
1343             av_log(dst, AV_LOG_ERROR, "Could not allocate memory for h264\n");
1344             return AVERROR(ENOMEM);
1345         }
1346         context_init(h);
1347
1348         /* frame_start may not be called for the next thread (if it's decoding
1349          * a bottom field) so this has to be allocated here */
1350         h->s.obmc_scratchpad = av_malloc(16 * 6 * s->linesize);
1351         }
1352
1353         for (i = 0; i < 2; i++) {
1354             h->rbsp_buffer[i]      = NULL;
1355             h->rbsp_buffer_size[i] = 0;
1356         }
1357
1358         h->thread_context[0] = h;
1359
1360         s->dsp.clear_blocks(h->mb);
1361         s->dsp.clear_blocks(h->mb + (24 * 16 << h->pixel_shift));
1362     }
1363
1364     // extradata/NAL handling
1365     h->is_avc = h1->is_avc;
1366
1367     // SPS/PPS
1368     copy_parameter_set((void **)h->sps_buffers, (void **)h1->sps_buffers,
1369                        MAX_SPS_COUNT, sizeof(SPS));
1370     h->sps = h1->sps;
1371     copy_parameter_set((void **)h->pps_buffers, (void **)h1->pps_buffers,
1372                        MAX_PPS_COUNT, sizeof(PPS));
1373     h->pps = h1->pps;
1374
1375     // Dequantization matrices
1376     // FIXME these are big - can they be only copied when PPS changes?
1377     copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1378
1379     for (i = 0; i < 6; i++)
1380         h->dequant4_coeff[i] = h->dequant4_buffer[0] +
1381                                (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1382
1383     for (i = 0; i < 6; i++)
1384         h->dequant8_coeff[i] = h->dequant8_buffer[0] +
1385                                (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1386
1387     h->dequant_coeff_pps = h1->dequant_coeff_pps;
1388
1389     // POC timing
1390     copy_fields(h, h1, poc_lsb, redundant_pic_count);
1391
1392     // reference lists
1393     copy_fields(h, h1, ref_count, list_count);
1394     copy_fields(h, h1, ref_list, intra_gb);
1395     copy_fields(h, h1, short_ref, cabac_init_idc);
1396
1397     copy_picture_range(h->short_ref, h1->short_ref, 32, s, s1);
1398     copy_picture_range(h->long_ref, h1->long_ref, 32, s, s1);
1399     copy_picture_range(h->delayed_pic, h1->delayed_pic,
1400                        MAX_DELAYED_PIC_COUNT + 2, s, s1);
1401
1402     h->last_slice_type = h1->last_slice_type;
1403     h->sync            = h1->sync;
1404
1405     if (!s->current_picture_ptr)
1406         return 0;
1407
1408     if (!s->dropable) {
1409         err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1410         h->prev_poc_msb = h->poc_msb;
1411         h->prev_poc_lsb = h->poc_lsb;
1412     }
1413     h->prev_frame_num_offset = h->frame_num_offset;
1414     h->prev_frame_num        = h->frame_num;
1415     h->outputed_poc          = h->next_outputed_poc;
1416
1417     return err;
1418 }
1419
1420 int ff_h264_frame_start(H264Context *h)
1421 {
1422     MpegEncContext *const s = &h->s;
1423     int i;
1424     const int pixel_shift = h->pixel_shift;
1425
1426     if (ff_MPV_frame_start(s, s->avctx) < 0)
1427         return -1;
1428     ff_er_frame_start(s);
1429     /*
1430      * ff_MPV_frame_start uses pict_type to derive key_frame.
1431      * This is incorrect for H.264; IDR markings must be used.
1432      * Zero here; IDR markings per slice in frame or fields are ORed in later.
1433      * See decode_nal_units().
1434      */
1435     s->current_picture_ptr->f.key_frame = 0;
1436     s->current_picture_ptr->sync        = 0;
1437     s->current_picture_ptr->mmco_reset  = 0;
1438
1439     assert(s->linesize && s->uvlinesize);
1440
1441     for (i = 0; i < 16; i++) {
1442         h->block_offset[i]           = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * s->linesize * ((scan8[i] - scan8[0]) >> 3);
1443         h->block_offset[48 + i]      = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * s->linesize * ((scan8[i] - scan8[0]) >> 3);
1444     }
1445     for (i = 0; i < 16; i++) {
1446         h->block_offset[16 + i]      =
1447         h->block_offset[32 + i]      = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * s->uvlinesize * ((scan8[i] - scan8[0]) >> 3);
1448         h->block_offset[48 + 16 + i] =
1449         h->block_offset[48 + 32 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * s->uvlinesize * ((scan8[i] - scan8[0]) >> 3);
1450     }
1451
1452     /* can't be in alloc_tables because linesize isn't known there.
1453      * FIXME: redo bipred weight to not require extra buffer? */
1454     for (i = 0; i < s->slice_context_count; i++)
1455         if (h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1456             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16 * 6 * s->linesize);
1457
1458     /* Some macroblocks can be accessed before they're available in case
1459      * of lost slices, MBAFF or threading. */
1460     memset(h->slice_table, -1,
1461            (s->mb_height * s->mb_stride - 1) * sizeof(*h->slice_table));
1462
1463     // s->decode = (s->flags & CODEC_FLAG_PSNR) || !s->encoding ||
1464     //             s->current_picture.f.reference /* || h->contains_intra */ || 1;
1465
1466     /* We mark the current picture as non-reference after allocating it, so
1467      * that if we break out due to an error it can be released automatically
1468      * in the next ff_MPV_frame_start().
1469      * SVQ3 as well as most other codecs have only last/next/current and thus
1470      * get released even with set reference, besides SVQ3 and others do not
1471      * mark frames as reference later "naturally". */
1472     if (s->codec_id != CODEC_ID_SVQ3)
1473         s->current_picture_ptr->f.reference = 0;
1474
1475     s->current_picture_ptr->field_poc[0]     =
1476         s->current_picture_ptr->field_poc[1] = INT_MAX;
1477
1478     h->next_output_pic = NULL;
1479
1480     assert(s->current_picture_ptr->long_ref == 0);
1481
1482     return 0;
1483 }
1484
1485 /**
1486  * Run setup operations that must be run after slice header decoding.
1487  * This includes finding the next displayed frame.
1488  *
1489  * @param h h264 master context
1490  * @param setup_finished enough NALs have been read that we can call
1491  * ff_thread_finish_setup()
1492  */
1493 static void decode_postinit(H264Context *h, int setup_finished)
1494 {
1495     MpegEncContext *const s = &h->s;
1496     Picture *out = s->current_picture_ptr;
1497     Picture *cur = s->current_picture_ptr;
1498     int i, pics, out_of_order, out_idx;
1499
1500     s->current_picture_ptr->f.qscale_type = FF_QSCALE_TYPE_H264;
1501     s->current_picture_ptr->f.pict_type   = s->pict_type;
1502
1503     if (h->next_output_pic)
1504         return;
1505
1506     if (cur->field_poc[0] == INT_MAX || cur->field_poc[1] == INT_MAX) {
1507         /* FIXME: if we have two PAFF fields in one packet, we can't start
1508          * the next thread here. If we have one field per packet, we can.
1509          * The check in decode_nal_units() is not good enough to find this
1510          * yet, so we assume the worst for now. */
1511         // if (setup_finished)
1512         //    ff_thread_finish_setup(s->avctx);
1513         return;
1514     }
1515
1516     cur->f.interlaced_frame = 0;
1517     cur->f.repeat_pict      = 0;
1518
1519     /* Signal interlacing information externally. */
1520     /* Prioritize picture timing SEI information over used
1521      * decoding process if it exists. */
1522
1523     if (h->sps.pic_struct_present_flag) {
1524         switch (h->sei_pic_struct) {
1525         case SEI_PIC_STRUCT_FRAME:
1526             break;
1527         case SEI_PIC_STRUCT_TOP_FIELD:
1528         case SEI_PIC_STRUCT_BOTTOM_FIELD:
1529             cur->f.interlaced_frame = 1;
1530             break;
1531         case SEI_PIC_STRUCT_TOP_BOTTOM:
1532         case SEI_PIC_STRUCT_BOTTOM_TOP:
1533             if (FIELD_OR_MBAFF_PICTURE)
1534                 cur->f.interlaced_frame = 1;
1535             else
1536                 // try to flag soft telecine progressive
1537                 cur->f.interlaced_frame = h->prev_interlaced_frame;
1538             break;
1539         case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1540         case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1541             /* Signal the possibility of telecined film externally
1542              * (pic_struct 5,6). From these hints, let the applications
1543              * decide if they apply deinterlacing. */
1544             cur->f.repeat_pict = 1;
1545             break;
1546         case SEI_PIC_STRUCT_FRAME_DOUBLING:
1547             // Force progressive here, doubling interlaced frame is a bad idea.
1548             cur->f.repeat_pict = 2;
1549             break;
1550         case SEI_PIC_STRUCT_FRAME_TRIPLING:
1551             cur->f.repeat_pict = 4;
1552             break;
1553         }
1554
1555         if ((h->sei_ct_type & 3) &&
1556             h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1557             cur->f.interlaced_frame = (h->sei_ct_type & (1 << 1)) != 0;
1558     } else {
1559         /* Derive interlacing flag from used decoding process. */
1560         cur->f.interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1561     }
1562     h->prev_interlaced_frame = cur->f.interlaced_frame;
1563
1564     if (cur->field_poc[0] != cur->field_poc[1]) {
1565         /* Derive top_field_first from field pocs. */
1566         cur->f.top_field_first = cur->field_poc[0] < cur->field_poc[1];
1567     } else {
1568         if (cur->f.interlaced_frame || h->sps.pic_struct_present_flag) {
1569             /* Use picture timing SEI information. Even if it is a
1570              * information of a past frame, better than nothing. */
1571             if (h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM ||
1572                 h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1573                 cur->f.top_field_first = 1;
1574             else
1575                 cur->f.top_field_first = 0;
1576         } else {
1577             /* Most likely progressive */
1578             cur->f.top_field_first = 0;
1579         }
1580     }
1581
1582     cur->mmco_reset = h->mmco_reset;
1583     h->mmco_reset = 0;
1584     // FIXME do something with unavailable reference frames
1585
1586     /* Sort B-frames into display order */
1587
1588     if (h->sps.bitstream_restriction_flag &&
1589         s->avctx->has_b_frames < h->sps.num_reorder_frames) {
1590         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1591         s->low_delay           = 0;
1592     }
1593
1594     if (s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT &&
1595         !h->sps.bitstream_restriction_flag) {
1596         s->avctx->has_b_frames = MAX_DELAYED_PIC_COUNT - 1;
1597         s->low_delay           = 0;
1598     }
1599
1600     for (i = 0; 1; i++) {
1601         if(i == MAX_DELAYED_PIC_COUNT || cur->poc < h->last_pocs[i]){
1602             if(i)
1603                 h->last_pocs[i-1] = cur->poc;
1604             break;
1605         } else if(i) {
1606             h->last_pocs[i-1]= h->last_pocs[i];
1607         }
1608     }
1609     out_of_order = MAX_DELAYED_PIC_COUNT - i;
1610     if(   cur->f.pict_type == AV_PICTURE_TYPE_B
1611        || (h->last_pocs[MAX_DELAYED_PIC_COUNT-2] > INT_MIN && h->last_pocs[MAX_DELAYED_PIC_COUNT-1] - h->last_pocs[MAX_DELAYED_PIC_COUNT-2] > 2))
1612         out_of_order = FFMAX(out_of_order, 1);
1613     if(s->avctx->has_b_frames < out_of_order && !h->sps.bitstream_restriction_flag){
1614         av_log(s->avctx, AV_LOG_WARNING, "Increasing reorder buffer to %d\n", out_of_order);
1615         s->avctx->has_b_frames = out_of_order;
1616         s->low_delay = 0;
1617     }
1618
1619     pics = 0;
1620     while (h->delayed_pic[pics])
1621         pics++;
1622
1623     av_assert0(pics <= MAX_DELAYED_PIC_COUNT);
1624
1625     h->delayed_pic[pics++] = cur;
1626     if (cur->f.reference == 0)
1627         cur->f.reference = DELAYED_PIC_REF;
1628
1629     out = h->delayed_pic[0];
1630     out_idx = 0;
1631     for (i = 1; h->delayed_pic[i] &&
1632                 !h->delayed_pic[i]->f.key_frame &&
1633                 !h->delayed_pic[i]->mmco_reset;
1634          i++)
1635         if (h->delayed_pic[i]->poc < out->poc) {
1636             out     = h->delayed_pic[i];
1637             out_idx = i;
1638         }
1639     if (s->avctx->has_b_frames == 0 &&
1640         (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset))
1641         h->next_outputed_poc = INT_MIN;
1642     out_of_order = out->poc < h->next_outputed_poc;
1643
1644     if (out_of_order || pics > s->avctx->has_b_frames) {
1645         out->f.reference &= ~DELAYED_PIC_REF;
1646         // for frame threading, the owner must be the second field's thread or
1647         // else the first thread can release the picture and reuse it unsafely
1648         out->owner2       = s;
1649         for (i = out_idx; h->delayed_pic[i]; i++)
1650             h->delayed_pic[i] = h->delayed_pic[i + 1];
1651     }
1652     if (!out_of_order && pics > s->avctx->has_b_frames) {
1653         h->next_output_pic = out;
1654         if (out_idx == 0 && h->delayed_pic[0] && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset)) {
1655             h->next_outputed_poc = INT_MIN;
1656         } else
1657             h->next_outputed_poc = out->poc;
1658     } else {
1659         av_log(s->avctx, AV_LOG_DEBUG, "no picture %s\n", out_of_order ? "ooo" : "");
1660     }
1661
1662     if (h->next_output_pic && h->next_output_pic->sync) {
1663         h->sync |= 2;
1664     }
1665
1666     if (setup_finished)
1667         ff_thread_finish_setup(s->avctx);
1668 }
1669
1670 static av_always_inline void backup_mb_border(H264Context *h, uint8_t *src_y,
1671                                               uint8_t *src_cb, uint8_t *src_cr,
1672                                               int linesize, int uvlinesize,
1673                                               int simple)
1674 {
1675     MpegEncContext *const s = &h->s;
1676     uint8_t *top_border;
1677     int top_idx = 1;
1678     const int pixel_shift = h->pixel_shift;
1679     int chroma444 = CHROMA444;
1680     int chroma422 = CHROMA422;
1681
1682     src_y  -= linesize;
1683     src_cb -= uvlinesize;
1684     src_cr -= uvlinesize;
1685
1686     if (!simple && FRAME_MBAFF) {
1687         if (s->mb_y & 1) {
1688             if (!MB_MBAFF) {
1689                 top_border = h->top_borders[0][s->mb_x];
1690                 AV_COPY128(top_border, src_y + 15 * linesize);
1691                 if (pixel_shift)
1692                     AV_COPY128(top_border + 16, src_y + 15 * linesize + 16);
1693                 if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1694                     if (chroma444) {
1695                         if (pixel_shift) {
1696                             AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
1697                             AV_COPY128(top_border + 48, src_cb + 15 * uvlinesize + 16);
1698                             AV_COPY128(top_border + 64, src_cr + 15 * uvlinesize);
1699                             AV_COPY128(top_border + 80, src_cr + 15 * uvlinesize + 16);
1700                         } else {
1701                             AV_COPY128(top_border + 16, src_cb + 15 * uvlinesize);
1702                             AV_COPY128(top_border + 32, src_cr + 15 * uvlinesize);
1703                         }
1704                     } else if (chroma422) {
1705                         if (pixel_shift) {
1706                             AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
1707                             AV_COPY128(top_border + 48, src_cr + 15 * uvlinesize);
1708                         } else {
1709                             AV_COPY64(top_border + 16, src_cb + 15 * uvlinesize);
1710                             AV_COPY64(top_border + 24, src_cr + 15 * uvlinesize);
1711                         }
1712                     } else {
1713                         if (pixel_shift) {
1714                             AV_COPY128(top_border + 32, src_cb + 7 * uvlinesize);
1715                             AV_COPY128(top_border + 48, src_cr + 7 * uvlinesize);
1716                         } else {
1717                             AV_COPY64(top_border + 16, src_cb + 7 * uvlinesize);
1718                             AV_COPY64(top_border + 24, src_cr + 7 * uvlinesize);
1719                         }
1720                     }
1721                 }
1722             }
1723         } else if (MB_MBAFF) {
1724             top_idx = 0;
1725         } else
1726             return;
1727     }
1728
1729     top_border = h->top_borders[top_idx][s->mb_x];
1730     /* There are two lines saved, the line above the the top macroblock
1731      * of a pair, and the line above the bottom macroblock. */
1732     AV_COPY128(top_border, src_y + 16 * linesize);
1733     if (pixel_shift)
1734         AV_COPY128(top_border + 16, src_y + 16 * linesize + 16);
1735
1736     if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1737         if (chroma444) {
1738             if (pixel_shift) {
1739                 AV_COPY128(top_border + 32, src_cb + 16 * linesize);
1740                 AV_COPY128(top_border + 48, src_cb + 16 * linesize + 16);
1741                 AV_COPY128(top_border + 64, src_cr + 16 * linesize);
1742                 AV_COPY128(top_border + 80, src_cr + 16 * linesize + 16);
1743             } else {
1744                 AV_COPY128(top_border + 16, src_cb + 16 * linesize);
1745                 AV_COPY128(top_border + 32, src_cr + 16 * linesize);
1746             }
1747         } else if (chroma422) {
1748             if (pixel_shift) {
1749                 AV_COPY128(top_border + 32, src_cb + 16 * uvlinesize);
1750                 AV_COPY128(top_border + 48, src_cr + 16 * uvlinesize);
1751             } else {
1752                 AV_COPY64(top_border + 16, src_cb + 16 * uvlinesize);
1753                 AV_COPY64(top_border + 24, src_cr + 16 * uvlinesize);
1754             }
1755         } else {
1756             if (pixel_shift) {
1757                 AV_COPY128(top_border + 32, src_cb + 8 * uvlinesize);
1758                 AV_COPY128(top_border + 48, src_cr + 8 * uvlinesize);
1759             } else {
1760                 AV_COPY64(top_border + 16, src_cb + 8 * uvlinesize);
1761                 AV_COPY64(top_border + 24, src_cr + 8 * uvlinesize);
1762             }
1763         }
1764     }
1765 }
1766
1767 static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1768                                             uint8_t *src_cb, uint8_t *src_cr,
1769                                             int linesize, int uvlinesize,
1770                                             int xchg, int chroma444,
1771                                             int simple, int pixel_shift)
1772 {
1773     MpegEncContext *const s = &h->s;
1774     int deblock_topleft;
1775     int deblock_top;
1776     int top_idx = 1;
1777     uint8_t *top_border_m1;
1778     uint8_t *top_border;
1779
1780     if (!simple && FRAME_MBAFF) {
1781         if (s->mb_y & 1) {
1782             if (!MB_MBAFF)
1783                 return;
1784         } else {
1785             top_idx = MB_MBAFF ? 0 : 1;
1786         }
1787     }
1788
1789     if (h->deblocking_filter == 2) {
1790         deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1791         deblock_top     = h->top_type;
1792     } else {
1793         deblock_topleft = (s->mb_x > 0);
1794         deblock_top     = (s->mb_y > !!MB_FIELD);
1795     }
1796
1797     src_y  -= linesize   + 1 + pixel_shift;
1798     src_cb -= uvlinesize + 1 + pixel_shift;
1799     src_cr -= uvlinesize + 1 + pixel_shift;
1800
1801     top_border_m1 = h->top_borders[top_idx][s->mb_x - 1];
1802     top_border    = h->top_borders[top_idx][s->mb_x];
1803
1804 #define XCHG(a, b, xchg)                        \
1805     if (pixel_shift) {                          \
1806         if (xchg) {                             \
1807             AV_SWAP64(b + 0, a + 0);            \
1808             AV_SWAP64(b + 8, a + 8);            \
1809         } else {                                \
1810             AV_COPY128(b, a);                   \
1811         }                                       \
1812     } else if (xchg)                            \
1813         AV_SWAP64(b, a);                        \
1814     else                                        \
1815         AV_COPY64(b, a);
1816
1817     if (deblock_top) {
1818         if (deblock_topleft) {
1819             XCHG(top_border_m1 + (8 << pixel_shift),
1820                  src_y - (7 << pixel_shift), 1);
1821         }
1822         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1823         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1824         if (s->mb_x + 1 < s->mb_width) {
1825             XCHG(h->top_borders[top_idx][s->mb_x + 1],
1826                  src_y + (17 << pixel_shift), 1);
1827         }
1828     }
1829     if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1830         if (chroma444) {
1831             if (deblock_topleft) {
1832                 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1833                 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1834             }
1835             XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1836             XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1837             XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1838             XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1839             if (s->mb_x + 1 < s->mb_width) {
1840                 XCHG(h->top_borders[top_idx][s->mb_x + 1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1841                 XCHG(h->top_borders[top_idx][s->mb_x + 1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1842             }
1843         } else {
1844             if (deblock_top) {
1845                 if (deblock_topleft) {
1846                     XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1847                     XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1848                 }
1849                 XCHG(top_border + (16 << pixel_shift), src_cb + 1 + pixel_shift, 1);
1850                 XCHG(top_border + (24 << pixel_shift), src_cr + 1 + pixel_shift, 1);
1851             }
1852         }
1853     }
1854 }
1855
1856 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth,
1857                                         int index)
1858 {
1859     if (high_bit_depth) {
1860         return AV_RN32A(((int32_t *)mb) + index);
1861     } else
1862         return AV_RN16A(mb + index);
1863 }
1864
1865 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth,
1866                                          int index, int value)
1867 {
1868     if (high_bit_depth) {
1869         AV_WN32A(((int32_t *)mb) + index, value);
1870     } else
1871         AV_WN16A(mb + index, value);
1872 }
1873
1874 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h,
1875                                                        int mb_type, int is_h264,
1876                                                        int simple,
1877                                                        int transform_bypass,
1878                                                        int pixel_shift,
1879                                                        int *block_offset,
1880                                                        int linesize,
1881                                                        uint8_t *dest_y, int p)
1882 {
1883     MpegEncContext *const s = &h->s;
1884     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1885     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1886     int i;
1887     int qscale = p == 0 ? s->qscale : h->chroma_qp[p - 1];
1888     block_offset += 16 * p;
1889     if (IS_INTRA4x4(mb_type)) {
1890         if (simple || !s->encoding) {
1891             if (IS_8x8DCT(mb_type)) {
1892                 if (transform_bypass) {
1893                     idct_dc_add  =
1894                     idct_add     = s->dsp.add_pixels8;
1895                 } else {
1896                     idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1897                     idct_add    = h->h264dsp.h264_idct8_add;
1898                 }
1899                 for (i = 0; i < 16; i += 4) {
1900                     uint8_t *const ptr = dest_y + block_offset[i];
1901                     const int dir      = h->intra4x4_pred_mode_cache[scan8[i]];
1902                     if (transform_bypass && h->sps.profile_idc == 244 && dir <= 1) {
1903                         h->hpc.pred8x8l_add[dir](ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1904                     } else {
1905                         const int nnz = h->non_zero_count_cache[scan8[i + p * 16]];
1906                         h->hpc.pred8x8l[dir](ptr, (h->topleft_samples_available << i) & 0x8000,
1907                                              (h->topright_samples_available << i) & 0x4000, linesize);
1908                         if (nnz) {
1909                             if (nnz == 1 && dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
1910                                 idct_dc_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1911                             else
1912                                 idct_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1913                         }
1914                     }
1915                 }
1916             } else {
1917                 if (transform_bypass) {
1918                     idct_dc_add  =
1919                         idct_add = s->dsp.add_pixels4;
1920                 } else {
1921                     idct_dc_add = h->h264dsp.h264_idct_dc_add;
1922                     idct_add    = h->h264dsp.h264_idct_add;
1923                 }
1924                 for (i = 0; i < 16; i++) {
1925                     uint8_t *const ptr = dest_y + block_offset[i];
1926                     const int dir      = h->intra4x4_pred_mode_cache[scan8[i]];
1927
1928                     if (transform_bypass && h->sps.profile_idc == 244 && dir <= 1) {
1929                         h->hpc.pred4x4_add[dir](ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1930                     } else {
1931                         uint8_t *topright;
1932                         int nnz, tr;
1933                         uint64_t tr_high;
1934                         if (dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED) {
1935                             const int topright_avail = (h->topright_samples_available << i) & 0x8000;
1936                             assert(s->mb_y || linesize <= block_offset[i]);
1937                             if (!topright_avail) {
1938                                 if (pixel_shift) {
1939                                     tr_high  = ((uint16_t *)ptr)[3 - linesize / 2] * 0x0001000100010001ULL;
1940                                     topright = (uint8_t *)&tr_high;
1941                                 } else {
1942                                     tr       = ptr[3 - linesize] * 0x01010101u;
1943                                     topright = (uint8_t *)&tr;
1944                                 }
1945                             } else
1946                                 topright = ptr + (4 << pixel_shift) - linesize;
1947                         } else
1948                             topright = NULL;
1949
1950                         h->hpc.pred4x4[dir](ptr, topright, linesize);
1951                         nnz = h->non_zero_count_cache[scan8[i + p * 16]];
1952                         if (nnz) {
1953                             if (is_h264) {
1954                                 if (nnz == 1 && dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
1955                                     idct_dc_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1956                                 else
1957                                     idct_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1958                             } else if (CONFIG_SVQ3_DECODER)
1959                                 ff_svq3_add_idct_c(ptr, h->mb + i * 16 + p * 256, linesize, qscale, 0);
1960                         }
1961                     }
1962                 }
1963             }
1964         }
1965     } else {
1966         h->hpc.pred16x16[h->intra16x16_pred_mode](dest_y, linesize);
1967         if (is_h264) {
1968             if (h->non_zero_count_cache[scan8[LUMA_DC_BLOCK_INDEX + p]]) {
1969                 if (!transform_bypass)
1970                     h->h264dsp.h264_luma_dc_dequant_idct(h->mb + (p * 256 << pixel_shift),
1971                                                          h->mb_luma_dc[p],
1972                                                          h->dequant4_coeff[p][qscale][0]);
1973                 else {
1974                     static const uint8_t dc_mapping[16] = {
1975                          0 * 16,  1 * 16,  4 * 16,  5 * 16,
1976                          2 * 16,  3 * 16,  6 * 16,  7 * 16,
1977                          8 * 16,  9 * 16, 12 * 16, 13 * 16,
1978                         10 * 16, 11 * 16, 14 * 16, 15 * 16 };
1979                     for (i = 0; i < 16; i++)
1980                         dctcoef_set(h->mb + (p * 256 << pixel_shift),
1981                                     pixel_shift, dc_mapping[i],
1982                                     dctcoef_get(h->mb_luma_dc[p],
1983                                                 pixel_shift, i));
1984                 }
1985             }
1986         } else if (CONFIG_SVQ3_DECODER)
1987             ff_svq3_luma_dc_dequant_idct_c(h->mb + p * 256,
1988                                            h->mb_luma_dc[p], qscale);
1989     }
1990 }
1991
1992 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type,
1993                                                     int is_h264, int simple,
1994                                                     int transform_bypass,
1995                                                     int pixel_shift,
1996                                                     int *block_offset,
1997                                                     int linesize,
1998                                                     uint8_t *dest_y, int p)
1999 {
2000     MpegEncContext *const s = &h->s;
2001     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2002     int i;
2003     block_offset += 16 * p;
2004     if (!IS_INTRA4x4(mb_type)) {
2005         if (is_h264) {
2006             if (IS_INTRA16x16(mb_type)) {
2007                 if (transform_bypass) {
2008                     if (h->sps.profile_idc == 244 &&
2009                         (h->intra16x16_pred_mode == VERT_PRED8x8 ||
2010                          h->intra16x16_pred_mode == HOR_PRED8x8)) {
2011                         h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset,
2012                                                                       h->mb + (p * 256 << pixel_shift),
2013                                                                       linesize);
2014                     } else {
2015                         for (i = 0; i < 16; i++)
2016                             if (h->non_zero_count_cache[scan8[i + p * 16]] ||
2017                                 dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
2018                                 s->dsp.add_pixels4(dest_y + block_offset[i],
2019                                                    h->mb + (i * 16 + p * 256 << pixel_shift),
2020                                                    linesize);
2021                     }
2022                 } else {
2023                     h->h264dsp.h264_idct_add16intra(dest_y, block_offset,
2024                                                     h->mb + (p * 256 << pixel_shift),
2025                                                     linesize,
2026                                                     h->non_zero_count_cache + p * 5 * 8);
2027                 }
2028             } else if (h->cbp & 15) {
2029                 if (transform_bypass) {
2030                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
2031                     idct_add = IS_8x8DCT(mb_type) ? s->dsp.add_pixels8
2032                                                   : s->dsp.add_pixels4;
2033                     for (i = 0; i < 16; i += di)
2034                         if (h->non_zero_count_cache[scan8[i + p * 16]])
2035                             idct_add(dest_y + block_offset[i],
2036                                      h->mb + (i * 16 + p * 256 << pixel_shift),
2037                                      linesize);
2038                 } else {
2039                     if (IS_8x8DCT(mb_type))
2040                         h->h264dsp.h264_idct8_add4(dest_y, block_offset,
2041                                                    h->mb + (p * 256 << pixel_shift),
2042                                                    linesize,
2043                                                    h->non_zero_count_cache + p * 5 * 8);
2044                     else
2045                         h->h264dsp.h264_idct_add16(dest_y, block_offset,
2046                                                    h->mb + (p * 256 << pixel_shift),
2047                                                    linesize,
2048                                                    h->non_zero_count_cache + p * 5 * 8);
2049                 }
2050             }
2051         } else if (CONFIG_SVQ3_DECODER) {
2052             for (i = 0; i < 16; i++)
2053                 if (h->non_zero_count_cache[scan8[i + p * 16]] || h->mb[i * 16 + p * 256]) {
2054                     // FIXME benchmark weird rule, & below
2055                     uint8_t *const ptr = dest_y + block_offset[i];
2056                     ff_svq3_add_idct_c(ptr, h->mb + i * 16 + p * 256, linesize,
2057                                        s->qscale, IS_INTRA(mb_type) ? 1 : 0);
2058                 }
2059         }
2060     }
2061 }
2062
2063 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple,
2064                                                    int pixel_shift)
2065 {
2066     MpegEncContext *const s = &h->s;
2067     const int mb_x    = s->mb_x;
2068     const int mb_y    = s->mb_y;
2069     const int mb_xy   = h->mb_xy;
2070     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2071     uint8_t *dest_y, *dest_cb, *dest_cr;
2072     int linesize, uvlinesize /*dct_offset*/;
2073     int i, j;
2074     int *block_offset = &h->block_offset[0];
2075     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
2076     /* is_h264 should always be true if SVQ3 is disabled. */
2077     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
2078     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2079     const int block_h   = 16 >> s->chroma_y_shift;
2080     const int chroma422 = CHROMA422;
2081
2082     dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift)     + mb_y * s->linesize)  * 16;
2083     dest_cb = s->current_picture.f.data[1] +  (mb_x << pixel_shift) * 8 + mb_y * s->uvlinesize * block_h;
2084     dest_cr = s->current_picture.f.data[2] +  (mb_x << pixel_shift) * 8 + mb_y * s->uvlinesize * block_h;
2085
2086     s->dsp.prefetch(dest_y  + (s->mb_x & 3) * 4 * s->linesize   + (64 << pixel_shift), s->linesize,       4);
2087     s->dsp.prefetch(dest_cb + (s->mb_x & 7)     * s->uvlinesize + (64 << pixel_shift), dest_cr - dest_cb, 2);
2088
2089     h->list_counts[mb_xy] = h->list_count;
2090
2091     if (!simple && MB_FIELD) {
2092         linesize     = h->mb_linesize = s->linesize * 2;
2093         uvlinesize   = h->mb_uvlinesize = s->uvlinesize * 2;
2094         block_offset = &h->block_offset[48];
2095         if (mb_y & 1) { // FIXME move out of this function?
2096             dest_y  -= s->linesize * 15;
2097             dest_cb -= s->uvlinesize * (block_h - 1);
2098             dest_cr -= s->uvlinesize * (block_h - 1);
2099         }
2100         if (FRAME_MBAFF) {
2101             int list;
2102             for (list = 0; list < h->list_count; list++) {
2103                 if (!USES_LIST(mb_type, list))
2104                     continue;
2105                 if (IS_16X16(mb_type)) {
2106                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2107                     fill_rectangle(ref, 4, 4, 8, (16 + *ref) ^ (s->mb_y & 1), 1);
2108                 } else {
2109                     for (i = 0; i < 16; i += 4) {
2110                         int ref = h->ref_cache[list][scan8[i]];
2111                         if (ref >= 0)
2112                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2,
2113                                            8, (16 + ref) ^ (s->mb_y & 1), 1);
2114                     }
2115                 }
2116             }
2117         }
2118     } else {
2119         linesize   = h->mb_linesize   = s->linesize;
2120         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
2121         // dct_offset = s->linesize * 16;
2122     }
2123
2124     if (!simple && IS_INTRA_PCM(mb_type)) {
2125         const int bit_depth = h->sps.bit_depth_luma;
2126         if (pixel_shift) {
2127             int j;
2128             GetBitContext gb;
2129             init_get_bits(&gb, (uint8_t *)h->mb, 384 * bit_depth);
2130
2131             for (i = 0; i < 16; i++) {
2132                 uint16_t *tmp_y = (uint16_t *)(dest_y + i * linesize);
2133                 for (j = 0; j < 16; j++)
2134                     tmp_y[j] = get_bits(&gb, bit_depth);
2135             }
2136             if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
2137                 if (!h->sps.chroma_format_idc) {
2138                     for (i = 0; i < block_h; i++) {
2139                         uint16_t *tmp_cb = (uint16_t *)(dest_cb + i * uvlinesize);
2140                         uint16_t *tmp_cr = (uint16_t *)(dest_cr + i * uvlinesize);
2141                         for (j = 0; j < 8; j++) {
2142                             tmp_cb[j] = tmp_cr[j] = 1 << (bit_depth - 1);
2143                         }
2144                     }
2145                 } else {
2146                     for (i = 0; i < block_h; i++) {
2147                         uint16_t *tmp_cb = (uint16_t *)(dest_cb + i * uvlinesize);
2148                         for (j = 0; j < 8; j++)
2149                             tmp_cb[j] = get_bits(&gb, bit_depth);
2150                     }
2151                     for (i = 0; i < block_h; i++) {
2152                         uint16_t *tmp_cr = (uint16_t *)(dest_cr + i * uvlinesize);
2153                         for (j = 0; j < 8; j++)
2154                             tmp_cr[j] = get_bits(&gb, bit_depth);
2155                     }
2156                 }
2157             }
2158         } else {
2159             for (i = 0; i < 16; i++)
2160                 memcpy(dest_y + i * linesize, h->mb + i * 8, 16);
2161             if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
2162                 if (!h->sps.chroma_format_idc) {
2163                     for (i = 0; i < 8; i++) {
2164                         memset(dest_cb + i*uvlinesize, 1 << (bit_depth - 1), 8);
2165                         memset(dest_cr + i*uvlinesize, 1 << (bit_depth - 1), 8);
2166                     }
2167                 } else {
2168                     for (i = 0; i < block_h; i++) {
2169                         memcpy(dest_cb + i * uvlinesize, h->mb + 128 + i * 4, 8);
2170                         memcpy(dest_cr + i * uvlinesize, h->mb + 160 + i * 4, 8);
2171                     }
2172                 }
2173             }
2174         }
2175     } else {
2176         if (IS_INTRA(mb_type)) {
2177             if (h->deblocking_filter)
2178                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize,
2179                                uvlinesize, 1, 0, simple, pixel_shift);
2180
2181             if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
2182                 if (CHROMA) {
2183                     h->hpc.pred8x8[h->chroma_pred_mode](dest_cb, uvlinesize);
2184                     h->hpc.pred8x8[h->chroma_pred_mode](dest_cr, uvlinesize);
2185                 }
2186             }
2187
2188             hl_decode_mb_predict_luma(h, mb_type, is_h264, simple,
2189                                       transform_bypass, pixel_shift,
2190                                       block_offset, linesize, dest_y, 0);
2191
2192             if (h->deblocking_filter)
2193                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize,
2194                                uvlinesize, 0, 0, simple, pixel_shift);
2195         } else if (is_h264) {
2196             if (chroma422) {
2197                 hl_motion_422(h, dest_y, dest_cb, dest_cr,
2198                               s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2199                               s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2200                               h->h264dsp.weight_h264_pixels_tab,
2201                               h->h264dsp.biweight_h264_pixels_tab,
2202                               pixel_shift);
2203             } else {
2204                 hl_motion_420(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             }
2211         }
2212
2213         hl_decode_mb_idct_luma(h, mb_type, is_h264, simple, transform_bypass,
2214                                pixel_shift, block_offset, linesize, dest_y, 0);
2215
2216         if ((simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) &&
2217             (h->cbp & 0x30)) {
2218             uint8_t *dest[2] = { dest_cb, dest_cr };
2219             if (transform_bypass) {
2220                 if (IS_INTRA(mb_type) && h->sps.profile_idc == 244 &&
2221                     (h->chroma_pred_mode == VERT_PRED8x8 ||
2222                      h->chroma_pred_mode == HOR_PRED8x8)) {
2223                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0],
2224                                                             block_offset + 16,
2225                                                             h->mb + (16 * 16 * 1 << pixel_shift),
2226                                                             uvlinesize);
2227                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1],
2228                                                             block_offset + 32,
2229                                                             h->mb + (16 * 16 * 2 << pixel_shift),
2230                                                             uvlinesize);
2231                 } else {
2232                     idct_add = s->dsp.add_pixels4;
2233                     for (j = 1; j < 3; j++) {
2234                         for (i = j * 16; i < j * 16 + 4; i++)
2235                             if (h->non_zero_count_cache[scan8[i]] ||
2236                                 dctcoef_get(h->mb, pixel_shift, i * 16))
2237                                 idct_add(dest[j - 1] + block_offset[i],
2238                                          h->mb + (i * 16 << pixel_shift),
2239                                          uvlinesize);
2240                         if (chroma422) {
2241                             for (i = j * 16 + 4; i < j * 16 + 8; i++)
2242                                 if (h->non_zero_count_cache[scan8[i + 4]] ||
2243                                     dctcoef_get(h->mb, pixel_shift, i * 16))
2244                                     idct_add(dest[j - 1] + block_offset[i + 4],
2245                                              h->mb + (i * 16 << pixel_shift),
2246                                              uvlinesize);
2247                         }
2248                     }
2249                 }
2250             } else {
2251                 if (is_h264) {
2252                     int qp[2];
2253                     if (chroma422) {
2254                         qp[0] = h->chroma_qp[0] + 3;
2255                         qp[1] = h->chroma_qp[1] + 3;
2256                     } else {
2257                         qp[0] = h->chroma_qp[0];
2258                         qp[1] = h->chroma_qp[1];
2259                     }
2260                     if (h->non_zero_count_cache[scan8[CHROMA_DC_BLOCK_INDEX + 0]])
2261                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16 * 16 * 1 << pixel_shift),
2262                                                                h->dequant4_coeff[IS_INTRA(mb_type) ? 1 : 4][qp[0]][0]);
2263                     if (h->non_zero_count_cache[scan8[CHROMA_DC_BLOCK_INDEX + 1]])
2264                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16 * 16 * 2 << pixel_shift),
2265                                                                h->dequant4_coeff[IS_INTRA(mb_type) ? 2 : 5][qp[1]][0]);
2266                     h->h264dsp.h264_idct_add8(dest, block_offset,
2267                                               h->mb, uvlinesize,
2268                                               h->non_zero_count_cache);
2269                 } else if (CONFIG_SVQ3_DECODER) {
2270                     h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + 16 * 16 * 1,
2271                                                            h->dequant4_coeff[IS_INTRA(mb_type) ? 1 : 4][h->chroma_qp[0]][0]);
2272                     h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + 16 * 16 * 2,
2273                                                            h->dequant4_coeff[IS_INTRA(mb_type) ? 2 : 5][h->chroma_qp[1]][0]);
2274                     for (j = 1; j < 3; j++) {
2275                         for (i = j * 16; i < j * 16 + 4; i++)
2276                             if (h->non_zero_count_cache[scan8[i]] || h->mb[i * 16]) {
2277                                 uint8_t *const ptr = dest[j - 1] + block_offset[i];
2278                                 ff_svq3_add_idct_c(ptr, h->mb + i * 16,
2279                                                    uvlinesize,
2280                                                    ff_h264_chroma_qp[0][s->qscale + 12] - 12, 2);
2281                             }
2282                     }
2283                 }
2284             }
2285         }
2286     }
2287     if (h->cbp || IS_INTRA(mb_type)) {
2288         s->dsp.clear_blocks(h->mb);
2289         s->dsp.clear_blocks(h->mb + (24 * 16 << pixel_shift));
2290     }
2291 }
2292
2293 static av_always_inline void hl_decode_mb_444_internal(H264Context *h,
2294                                                        int simple,
2295                                                        int pixel_shift)
2296 {
2297     MpegEncContext *const s = &h->s;
2298     const int mb_x    = s->mb_x;
2299     const int mb_y    = s->mb_y;
2300     const int mb_xy   = h->mb_xy;
2301     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2302     uint8_t *dest[3];
2303     int linesize;
2304     int i, j, p;
2305     int *block_offset = &h->block_offset[0];
2306     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
2307     const int plane_count      = (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) ? 3 : 1;
2308
2309     for (p = 0; p < plane_count; p++) {
2310         dest[p] = s->current_picture.f.data[p] +
2311                   ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
2312         s->dsp.prefetch(dest[p] + (s->mb_x & 3) * 4 * s->linesize + (64 << pixel_shift),
2313                         s->linesize, 4);
2314     }
2315
2316     h->list_counts[mb_xy] = h->list_count;
2317
2318     if (!simple && MB_FIELD) {
2319         linesize     = h->mb_linesize = h->mb_uvlinesize = s->linesize * 2;
2320         block_offset = &h->block_offset[48];
2321         if (mb_y & 1) // FIXME move out of this function?
2322             for (p = 0; p < 3; p++)
2323                 dest[p] -= s->linesize * 15;
2324         if (FRAME_MBAFF) {
2325             int list;
2326             for (list = 0; list < h->list_count; list++) {
2327                 if (!USES_LIST(mb_type, list))
2328                     continue;
2329                 if (IS_16X16(mb_type)) {
2330                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2331                     fill_rectangle(ref, 4, 4, 8, (16 + *ref) ^ (s->mb_y & 1), 1);
2332                 } else {
2333                     for (i = 0; i < 16; i += 4) {
2334                         int ref = h->ref_cache[list][scan8[i]];
2335                         if (ref >= 0)
2336                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2,
2337                                            8, (16 + ref) ^ (s->mb_y & 1), 1);
2338                     }
2339                 }
2340             }
2341         }
2342     } else {
2343         linesize = h->mb_linesize = h->mb_uvlinesize = s->linesize;
2344     }
2345
2346     if (!simple && IS_INTRA_PCM(mb_type)) {
2347         if (pixel_shift) {
2348             const int bit_depth = h->sps.bit_depth_luma;
2349             GetBitContext gb;
2350             init_get_bits(&gb, (uint8_t *)h->mb, 768 * bit_depth);
2351
2352             for (p = 0; p < plane_count; p++)
2353                 for (i = 0; i < 16; i++) {
2354                     uint16_t *tmp = (uint16_t *)(dest[p] + i * linesize);
2355                     for (j = 0; j < 16; j++)
2356                         tmp[j] = get_bits(&gb, bit_depth);
2357                 }
2358         } else {
2359             for (p = 0; p < plane_count; p++)
2360                 for (i = 0; i < 16; i++)
2361                     memcpy(dest[p] + i * linesize, h->mb + p * 128 + i * 8, 16);
2362         }
2363     } else {
2364         if (IS_INTRA(mb_type)) {
2365             if (h->deblocking_filter)
2366                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize,
2367                                linesize, 1, 1, simple, pixel_shift);
2368
2369             for (p = 0; p < plane_count; p++)
2370                 hl_decode_mb_predict_luma(h, mb_type, 1, simple,
2371                                           transform_bypass, pixel_shift,
2372                                           block_offset, linesize, dest[p], p);
2373
2374             if (h->deblocking_filter)
2375                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize,
2376                                linesize, 0, 1, simple, pixel_shift);
2377         } else {
2378             hl_motion(h, dest[0], dest[1], dest[2],
2379                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2380                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2381                       h->h264dsp.weight_h264_pixels_tab,
2382                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 3);
2383         }
2384
2385         for (p = 0; p < plane_count; p++)
2386             hl_decode_mb_idct_luma(h, mb_type, 1, simple, transform_bypass,
2387                                    pixel_shift, block_offset, linesize,
2388                                    dest[p], p);
2389     }
2390     if (h->cbp || IS_INTRA(mb_type)) {
2391         s->dsp.clear_blocks(h->mb);
2392         s->dsp.clear_blocks(h->mb + (24 * 16 << pixel_shift));
2393     }
2394 }
2395
2396 /**
2397  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2398  */
2399 #define hl_decode_mb_simple(sh, bits)                          \
2400 static void hl_decode_mb_simple_ ## bits(H264Context *h)       \
2401 {                                                              \
2402     hl_decode_mb_internal(h, 1, sh);                           \
2403 }
2404
2405 hl_decode_mb_simple(0, 8)
2406 hl_decode_mb_simple(1, 16)
2407
2408 /**
2409  * Process a macroblock; this handles edge cases, such as interlacing.
2410  */
2411 static av_noinline void hl_decode_mb_complex(H264Context *h)
2412 {
2413     hl_decode_mb_internal(h, 0, h->pixel_shift);
2414 }
2415
2416 static av_noinline void hl_decode_mb_444_complex(H264Context *h)
2417 {
2418     hl_decode_mb_444_internal(h, 0, h->pixel_shift);
2419 }
2420
2421 static av_noinline void hl_decode_mb_444_simple(H264Context *h)
2422 {
2423     hl_decode_mb_444_internal(h, 1, 0);
2424 }
2425
2426 void ff_h264_hl_decode_mb(H264Context *h)
2427 {
2428     MpegEncContext *const s = &h->s;
2429     const int mb_xy   = h->mb_xy;
2430     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2431     int is_complex    = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2432
2433     if (CHROMA444) {
2434         if (is_complex || h->pixel_shift)
2435             hl_decode_mb_444_complex(h);
2436         else
2437             hl_decode_mb_444_simple(h);
2438     } else if (is_complex) {
2439         hl_decode_mb_complex(h);
2440     } else if (h->pixel_shift) {
2441         hl_decode_mb_simple_16(h);
2442     } else
2443         hl_decode_mb_simple_8(h);
2444 }
2445
2446 static int pred_weight_table(H264Context *h)
2447 {
2448     MpegEncContext *const s = &h->s;
2449     int list, i;
2450     int luma_def, chroma_def;
2451
2452     h->use_weight             = 0;
2453     h->use_weight_chroma      = 0;
2454     h->luma_log2_weight_denom = get_ue_golomb(&s->gb);
2455     if (h->sps.chroma_format_idc)
2456         h->chroma_log2_weight_denom = get_ue_golomb(&s->gb);
2457     luma_def   = 1 << h->luma_log2_weight_denom;
2458     chroma_def = 1 << h->chroma_log2_weight_denom;
2459
2460     for (list = 0; list < 2; list++) {
2461         h->luma_weight_flag[list]   = 0;
2462         h->chroma_weight_flag[list] = 0;
2463         for (i = 0; i < h->ref_count[list]; i++) {
2464             int luma_weight_flag, chroma_weight_flag;
2465
2466             luma_weight_flag = get_bits1(&s->gb);
2467             if (luma_weight_flag) {
2468                 h->luma_weight[i][list][0] = get_se_golomb(&s->gb);
2469                 h->luma_weight[i][list][1] = get_se_golomb(&s->gb);
2470                 if (h->luma_weight[i][list][0] != luma_def ||
2471                     h->luma_weight[i][list][1] != 0) {
2472                     h->use_weight             = 1;
2473                     h->luma_weight_flag[list] = 1;
2474                 }
2475             } else {
2476                 h->luma_weight[i][list][0] = luma_def;
2477                 h->luma_weight[i][list][1] = 0;
2478             }
2479
2480             if (h->sps.chroma_format_idc) {
2481                 chroma_weight_flag = get_bits1(&s->gb);
2482                 if (chroma_weight_flag) {
2483                     int j;
2484                     for (j = 0; j < 2; j++) {
2485                         h->chroma_weight[i][list][j][0] = get_se_golomb(&s->gb);
2486                         h->chroma_weight[i][list][j][1] = get_se_golomb(&s->gb);
2487                         if (h->chroma_weight[i][list][j][0] != chroma_def ||
2488                             h->chroma_weight[i][list][j][1] != 0) {
2489                             h->use_weight_chroma = 1;
2490                             h->chroma_weight_flag[list] = 1;
2491                         }
2492                     }
2493                 } else {
2494                     int j;
2495                     for (j = 0; j < 2; j++) {
2496                         h->chroma_weight[i][list][j][0] = chroma_def;
2497                         h->chroma_weight[i][list][j][1] = 0;
2498                     }
2499                 }
2500             }
2501         }
2502         if (h->slice_type_nos != AV_PICTURE_TYPE_B)
2503             break;
2504     }
2505     h->use_weight = h->use_weight || h->use_weight_chroma;
2506     return 0;
2507 }
2508
2509 /**
2510  * Initialize implicit_weight table.
2511  * @param field  0/1 initialize the weight for interlaced MBAFF
2512  *                -1 initializes the rest
2513  */
2514 static void implicit_weight_table(H264Context *h, int field)
2515 {
2516     MpegEncContext *const s = &h->s;
2517     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2518
2519     for (i = 0; i < 2; i++) {
2520         h->luma_weight_flag[i]   = 0;
2521         h->chroma_weight_flag[i] = 0;
2522     }
2523
2524     if (field < 0) {
2525         if (s->picture_structure == PICT_FRAME) {
2526             cur_poc = s->current_picture_ptr->poc;
2527         } else {
2528             cur_poc = s->current_picture_ptr->field_poc[s->picture_structure - 1];
2529         }
2530         if (h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF &&
2531             h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2 * cur_poc) {
2532             h->use_weight = 0;
2533             h->use_weight_chroma = 0;
2534             return;
2535         }
2536         ref_start  = 0;
2537         ref_count0 = h->ref_count[0];
2538         ref_count1 = h->ref_count[1];
2539     } else {
2540         cur_poc    = s->current_picture_ptr->field_poc[field];
2541         ref_start  = 16;
2542         ref_count0 = 16 + 2 * h->ref_count[0];
2543         ref_count1 = 16 + 2 * h->ref_count[1];
2544     }
2545
2546     h->use_weight               = 2;
2547     h->use_weight_chroma        = 2;
2548     h->luma_log2_weight_denom   = 5;
2549     h->chroma_log2_weight_denom = 5;
2550
2551     for (ref0 = ref_start; ref0 < ref_count0; ref0++) {
2552         int poc0 = h->ref_list[0][ref0].poc;
2553         for (ref1 = ref_start; ref1 < ref_count1; ref1++) {
2554             int w = 32;
2555             if (!h->ref_list[0][ref0].long_ref && !h->ref_list[1][ref1].long_ref) {
2556                 int poc1 = h->ref_list[1][ref1].poc;
2557                 int td   = av_clip(poc1 - poc0, -128, 127);
2558                 if (td) {
2559                     int tb = av_clip(cur_poc - poc0, -128, 127);
2560                     int tx = (16384 + (FFABS(td) >> 1)) / td;
2561                     int dist_scale_factor = (tb * tx + 32) >> 8;
2562                     if (dist_scale_factor >= -64 && dist_scale_factor <= 128)
2563                         w = 64 - dist_scale_factor;
2564                 }
2565             }
2566             if (field < 0) {
2567                 h->implicit_weight[ref0][ref1][0] =
2568                 h->implicit_weight[ref0][ref1][1] = w;
2569             } else {
2570                 h->implicit_weight[ref0][ref1][field] = w;
2571             }
2572         }
2573     }
2574 }
2575
2576 /**
2577  * instantaneous decoder refresh.
2578  */
2579 static void idr(H264Context *h)
2580 {
2581     int i;
2582     ff_h264_remove_all_refs(h);
2583     h->prev_frame_num        = 0;
2584     h->prev_frame_num_offset = 0;
2585     h->prev_poc_msb          = 1<<16;
2586     h->prev_poc_lsb          = 0;
2587     for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
2588         h->last_pocs[i] = INT_MIN;
2589 }
2590
2591 /* forget old pics after a seek */
2592 static void flush_dpb(AVCodecContext *avctx)
2593 {
2594     H264Context *h = avctx->priv_data;
2595     int i;
2596     for (i=0; i<=MAX_DELAYED_PIC_COUNT; i++) {
2597         if (h->delayed_pic[i])
2598             h->delayed_pic[i]->f.reference = 0;
2599         h->delayed_pic[i] = NULL;
2600     }
2601     h->outputed_poc = h->next_outputed_poc = INT_MIN;
2602     h->prev_interlaced_frame = 1;
2603     idr(h);
2604     h->prev_frame_num = -1;
2605     if (h->s.current_picture_ptr)
2606         h->s.current_picture_ptr->f.reference = 0;
2607     h->s.first_field = 0;
2608     ff_h264_reset_sei(h);
2609     ff_mpeg_flush(avctx);
2610     h->recovery_frame= -1;
2611     h->sync= 0;
2612 }
2613
2614 static int init_poc(H264Context *h)
2615 {
2616     MpegEncContext *const s = &h->s;
2617     const int max_frame_num = 1 << h->sps.log2_max_frame_num;
2618     int field_poc[2];
2619     Picture *cur = s->current_picture_ptr;
2620
2621     h->frame_num_offset = h->prev_frame_num_offset;
2622     if (h->frame_num < h->prev_frame_num)
2623         h->frame_num_offset += max_frame_num;
2624
2625     if (h->sps.poc_type == 0) {
2626         const int max_poc_lsb = 1 << h->sps.log2_max_poc_lsb;
2627
2628         if (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb / 2)
2629             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2630         else if (h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb / 2)
2631             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2632         else
2633             h->poc_msb = h->prev_poc_msb;
2634         // printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
2635         field_poc[0] =
2636         field_poc[1] = h->poc_msb + h->poc_lsb;
2637         if (s->picture_structure == PICT_FRAME)
2638             field_poc[1] += h->delta_poc_bottom;
2639     } else if (h->sps.poc_type == 1) {
2640         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2641         int i;
2642
2643         if (h->sps.poc_cycle_length != 0)
2644             abs_frame_num = h->frame_num_offset + h->frame_num;
2645         else
2646             abs_frame_num = 0;
2647
2648         if (h->nal_ref_idc == 0 && abs_frame_num > 0)
2649             abs_frame_num--;
2650
2651         expected_delta_per_poc_cycle = 0;
2652         for (i = 0; i < h->sps.poc_cycle_length; i++)
2653             // FIXME integrate during sps parse
2654             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[i];
2655
2656         if (abs_frame_num > 0) {
2657             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2658             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2659
2660             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2661             for (i = 0; i <= frame_num_in_poc_cycle; i++)
2662                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[i];
2663         } else
2664             expectedpoc = 0;
2665
2666         if (h->nal_ref_idc == 0)
2667             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2668
2669         field_poc[0] = expectedpoc + h->delta_poc[0];
2670         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2671
2672         if (s->picture_structure == PICT_FRAME)
2673             field_poc[1] += h->delta_poc[1];
2674     } else {
2675         int poc = 2 * (h->frame_num_offset + h->frame_num);
2676
2677         if (!h->nal_ref_idc)
2678             poc--;
2679
2680         field_poc[0] = poc;
2681         field_poc[1] = poc;
2682     }
2683
2684     if (s->picture_structure != PICT_BOTTOM_FIELD)
2685         s->current_picture_ptr->field_poc[0] = field_poc[0];
2686     if (s->picture_structure != PICT_TOP_FIELD)
2687         s->current_picture_ptr->field_poc[1] = field_poc[1];
2688     cur->poc = FFMIN(cur->field_poc[0], cur->field_poc[1]);
2689
2690     return 0;
2691 }
2692
2693 /**
2694  * initialize scan tables
2695  */
2696 static void init_scan_tables(H264Context *h)
2697 {
2698     int i;
2699     for (i = 0; i < 16; i++) {
2700 #define T(x) (x >> 2) | ((x << 2) & 0xF)
2701         h->zigzag_scan[i] = T(zigzag_scan[i]);
2702         h->field_scan[i]  = T(field_scan[i]);
2703 #undef T
2704     }
2705     for (i = 0; i < 64; i++) {
2706 #define T(x) (x >> 3) | ((x & 7) << 3)
2707         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
2708         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2709         h->field_scan8x8[i]        = T(field_scan8x8[i]);
2710         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
2711 #undef T
2712     }
2713     if (h->sps.transform_bypass) { // FIXME same ugly
2714         h->zigzag_scan_q0          = zigzag_scan;
2715         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
2716         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2717         h->field_scan_q0           = field_scan;
2718         h->field_scan8x8_q0        = field_scan8x8;
2719         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
2720     } else {
2721         h->zigzag_scan_q0          = h->zigzag_scan;
2722         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
2723         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2724         h->field_scan_q0           = h->field_scan;
2725         h->field_scan8x8_q0        = h->field_scan8x8;
2726         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
2727     }
2728 }
2729
2730 static int field_end(H264Context *h, int in_setup)
2731 {
2732     MpegEncContext *const s     = &h->s;
2733     AVCodecContext *const avctx = s->avctx;
2734     int err = 0;
2735     s->mb_y = 0;
2736
2737     if (!in_setup && !s->dropable)
2738         ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX,
2739                                   s->picture_structure == PICT_BOTTOM_FIELD);
2740
2741     if (CONFIG_H264_VDPAU_DECODER &&
2742         s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
2743         ff_vdpau_h264_set_reference_frames(s);
2744
2745     if (in_setup || !(avctx->active_thread_type & FF_THREAD_FRAME)) {
2746         if (!s->dropable) {
2747             err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2748             h->prev_poc_msb = h->poc_msb;
2749             h->prev_poc_lsb = h->poc_lsb;
2750         }
2751         h->prev_frame_num_offset = h->frame_num_offset;
2752         h->prev_frame_num        = h->frame_num;
2753         h->outputed_poc          = h->next_outputed_poc;
2754     }
2755
2756     if (avctx->hwaccel) {
2757         if (avctx->hwaccel->end_frame(avctx) < 0)
2758             av_log(avctx, AV_LOG_ERROR,
2759                    "hardware accelerator failed to decode picture\n");
2760     }
2761
2762     if (CONFIG_H264_VDPAU_DECODER &&
2763         s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
2764         ff_vdpau_h264_picture_complete(s);
2765
2766     /*
2767      * FIXME: Error handling code does not seem to support interlaced
2768      * when slices span multiple rows
2769      * The ff_er_add_slice calls don't work right for bottom
2770      * fields; they cause massive erroneous error concealing
2771      * Error marking covers both fields (top and bottom).
2772      * This causes a mismatched s->error_count
2773      * and a bad error table. Further, the error count goes to
2774      * INT_MAX when called for bottom field, because mb_y is
2775      * past end by one (callers fault) and resync_mb_y != 0
2776      * causes problems for the first MB line, too.
2777      */
2778     if (!FIELD_PICTURE)
2779         ff_er_frame_end(s);
2780
2781     ff_MPV_frame_end(s);
2782
2783     h->current_slice = 0;
2784
2785     return err;
2786 }
2787
2788 /**
2789  * Replicate H264 "master" context to thread contexts.
2790  */
2791 static void clone_slice(H264Context *dst, H264Context *src)
2792 {
2793     memcpy(dst->block_offset, src->block_offset, sizeof(dst->block_offset));
2794     dst->s.current_picture_ptr = src->s.current_picture_ptr;
2795     dst->s.current_picture     = src->s.current_picture;
2796     dst->s.linesize            = src->s.linesize;
2797     dst->s.uvlinesize          = src->s.uvlinesize;
2798     dst->s.first_field         = src->s.first_field;
2799
2800     dst->prev_poc_msb          = src->prev_poc_msb;
2801     dst->prev_poc_lsb          = src->prev_poc_lsb;
2802     dst->prev_frame_num_offset = src->prev_frame_num_offset;
2803     dst->prev_frame_num        = src->prev_frame_num;
2804     dst->short_ref_count       = src->short_ref_count;
2805
2806     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
2807     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
2808     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2809     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
2810
2811     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
2812     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
2813 }
2814
2815 /**
2816  * Compute profile from profile_idc and constraint_set?_flags.
2817  *
2818  * @param sps SPS
2819  *
2820  * @return profile as defined by FF_PROFILE_H264_*
2821  */
2822 int ff_h264_get_profile(SPS *sps)
2823 {
2824     int profile = sps->profile_idc;
2825
2826     switch (sps->profile_idc) {
2827     case FF_PROFILE_H264_BASELINE:
2828         // constraint_set1_flag set to 1
2829         profile |= (sps->constraint_set_flags & 1 << 1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2830         break;
2831     case FF_PROFILE_H264_HIGH_10:
2832     case FF_PROFILE_H264_HIGH_422:
2833     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2834         // constraint_set3_flag set to 1
2835         profile |= (sps->constraint_set_flags & 1 << 3) ? FF_PROFILE_H264_INTRA : 0;
2836         break;
2837     }
2838
2839     return profile;
2840 }
2841
2842 /**
2843  * Decode a slice header.
2844  * This will also call ff_MPV_common_init() and frame_start() as needed.
2845  *
2846  * @param h h264context
2847  * @param h0 h264 master context (differs from 'h' when doing sliced based
2848  *           parallel decoding)
2849  *
2850  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2851  */
2852 static int decode_slice_header(H264Context *h, H264Context *h0)
2853 {
2854     MpegEncContext *const s  = &h->s;
2855     MpegEncContext *const s0 = &h0->s;
2856     unsigned int first_mb_in_slice;
2857     unsigned int pps_id;
2858     int num_ref_idx_active_override_flag;
2859     unsigned int slice_type, tmp, i, j;
2860     int default_ref_list_done = 0;
2861     int last_pic_structure, last_pic_dropable;
2862     int must_reinit;
2863
2864     /* FIXME: 2tap qpel isn't implemented for high bit depth. */
2865     if ((s->avctx->flags2 & CODEC_FLAG2_FAST) &&
2866         !h->nal_ref_idc && !h->pixel_shift) {
2867         s->me.qpel_put = s->dsp.put_2tap_qpel_pixels_tab;
2868         s->me.qpel_avg = s->dsp.avg_2tap_qpel_pixels_tab;
2869     } else {
2870         s->me.qpel_put = s->dsp.put_h264_qpel_pixels_tab;
2871         s->me.qpel_avg = s->dsp.avg_h264_qpel_pixels_tab;
2872     }
2873
2874     first_mb_in_slice = get_ue_golomb_long(&s->gb);
2875
2876     if (first_mb_in_slice == 0) { // FIXME better field boundary detection
2877         if (h0->current_slice && FIELD_PICTURE) {
2878             field_end(h, 1);
2879         }
2880
2881         h0->current_slice = 0;
2882         if (!s0->first_field) {
2883             if (s->current_picture_ptr && !s->dropable &&
2884                 s->current_picture_ptr->owner2 == s) {
2885                 ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX,
2886                                           s->picture_structure == PICT_BOTTOM_FIELD);
2887             }
2888             s->current_picture_ptr = NULL;
2889         }
2890     }
2891
2892     slice_type = get_ue_golomb_31(&s->gb);
2893     if (slice_type > 9) {
2894         av_log(h->s.avctx, AV_LOG_ERROR,
2895                "slice type too large (%d) at %d %d\n",
2896                h->slice_type, s->mb_x, s->mb_y);
2897         return -1;
2898     }
2899     if (slice_type > 4) {
2900         slice_type -= 5;
2901         h->slice_type_fixed = 1;
2902     } else
2903         h->slice_type_fixed = 0;
2904
2905     slice_type = golomb_to_pict_type[slice_type];
2906     if (slice_type == AV_PICTURE_TYPE_I ||
2907         (h0->current_slice != 0 && slice_type == h0->last_slice_type)) {
2908         default_ref_list_done = 1;
2909     }
2910     h->slice_type     = slice_type;
2911     h->slice_type_nos = slice_type & 3;
2912
2913     // to make a few old functions happy, it's wrong though
2914     s->pict_type = h->slice_type;
2915
2916     pps_id = get_ue_golomb(&s->gb);
2917     if (pps_id >= MAX_PPS_COUNT) {
2918         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id %d out of range\n", pps_id);
2919         return -1;
2920     }
2921     if (!h0->pps_buffers[pps_id]) {
2922         av_log(h->s.avctx, AV_LOG_ERROR,
2923                "non-existing PPS %u referenced\n",
2924                pps_id);
2925         return -1;
2926     }
2927     h->pps = *h0->pps_buffers[pps_id];
2928
2929     if (!h0->sps_buffers[h->pps.sps_id]) {
2930         av_log(h->s.avctx, AV_LOG_ERROR,
2931                "non-existing SPS %u referenced\n",
2932                h->pps.sps_id);
2933         return -1;
2934     }
2935     h->sps = *h0->sps_buffers[h->pps.sps_id];
2936
2937     s->avctx->profile = ff_h264_get_profile(&h->sps);
2938     s->avctx->level   = h->sps.level_idc;
2939     s->avctx->refs    = h->sps.ref_frame_count;
2940
2941     must_reinit = (s->context_initialized &&
2942                     (   16*h->sps.mb_width != s->avctx->coded_width
2943                      || 16*h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag) != s->avctx->coded_height
2944                      || s->avctx->bits_per_raw_sample != h->sps.bit_depth_luma
2945                      || h->cur_chroma_format_idc != h->sps.chroma_format_idc
2946                      || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio)));
2947
2948     if(must_reinit && (h != h0 || (s->avctx->active_thread_type & FF_THREAD_FRAME))) {
2949         av_log_missing_feature(s->avctx,
2950                                 "Width/height/bit depth/chroma idc changing with threads is", 0);
2951         return -1;   // width / height changed during parallelized decoding
2952     }
2953
2954     s->mb_width  = h->sps.mb_width;
2955     s->mb_height = h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2956
2957     h->b_stride = s->mb_width * 4;
2958
2959     s->chroma_y_shift = h->sps.chroma_format_idc <= 1; // 400 uses yuv420p
2960
2961     s->width  = 16 * s->mb_width;
2962     s->height = 16 * s->mb_height;
2963
2964
2965     if(must_reinit) {
2966         free_tables(h, 0);
2967         flush_dpb(s->avctx);
2968         ff_MPV_common_end(s);
2969         h->list_count = 0;
2970     }
2971     if (!s->context_initialized) {
2972         if (h != h0) {
2973             av_log(h->s.avctx, AV_LOG_ERROR,
2974                    "Cannot (re-)initialize context during parallel decoding.\n");
2975             return -1;
2976         }
2977         avcodec_set_dimensions(s->avctx, s->width, s->height);
2978         s->avctx->width  -= (2>>CHROMA444)*FFMIN(h->sps.crop_right, (8<<CHROMA444)-1);
2979         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);
2980         s->avctx->sample_aspect_ratio = h->sps.sar;
2981         av_assert0(s->avctx->sample_aspect_ratio.den);
2982
2983         if (s->avctx->bits_per_raw_sample != h->sps.bit_depth_luma ||
2984             h->cur_chroma_format_idc != h->sps.chroma_format_idc) {
2985             if (h->sps.bit_depth_luma >= 8 && h->sps.bit_depth_luma <= 10 &&
2986                 (h->sps.bit_depth_luma != 9 || !CHROMA422)) {
2987                 s->avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
2988                 h->cur_chroma_format_idc = h->sps.chroma_format_idc;
2989                 h->pixel_shift = h->sps.bit_depth_luma > 8;
2990
2991                 ff_h264dsp_init(&h->h264dsp, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
2992                 ff_h264_pred_init(&h->hpc, s->codec_id, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
2993                 s->dsp.dct_bits = h->sps.bit_depth_luma > 8 ? 32 : 16;
2994                 ff_dsputil_init(&s->dsp, s->avctx);
2995             } else {
2996                 av_log(s->avctx, AV_LOG_ERROR, "Unsupported bit depth: %d chroma_idc: %d\n",
2997                        h->sps.bit_depth_luma, h->sps.chroma_format_idc);
2998                 return -1;
2999             }
3000         }
3001
3002         if (h->sps.video_signal_type_present_flag) {
3003             s->avctx->color_range = h->sps.full_range>0 ? AVCOL_RANGE_JPEG
3004                                                       : AVCOL_RANGE_MPEG;
3005             if (h->sps.colour_description_present_flag) {
3006                 s->avctx->color_primaries = h->sps.color_primaries;
3007                 s->avctx->color_trc       = h->sps.color_trc;
3008                 s->avctx->colorspace      = h->sps.colorspace;
3009             }
3010         }
3011
3012         if (h->sps.timing_info_present_flag) {
3013             int64_t den = h->sps.time_scale;
3014             if (h->x264_build < 44U)
3015                 den *= 2;
3016             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
3017                       h->sps.num_units_in_tick, den, 1 << 30);
3018         }
3019
3020         switch (h->sps.bit_depth_luma) {
3021         case 9:
3022             if (CHROMA444) {
3023                 if (s->avctx->colorspace == AVCOL_SPC_RGB) {
3024                     s->avctx->pix_fmt = PIX_FMT_GBRP9;
3025                 } else
3026                     s->avctx->pix_fmt = PIX_FMT_YUV444P9;
3027             } else if (CHROMA422)
3028                 s->avctx->pix_fmt = PIX_FMT_YUV422P9;
3029             else
3030                 s->avctx->pix_fmt = PIX_FMT_YUV420P9;
3031             break;
3032         case 10:
3033             if (CHROMA444) {
3034                 if (s->avctx->colorspace == AVCOL_SPC_RGB) {
3035                     s->avctx->pix_fmt = PIX_FMT_GBRP10;
3036                 } else
3037                     s->avctx->pix_fmt = PIX_FMT_YUV444P10;
3038             } else if (CHROMA422)
3039                 s->avctx->pix_fmt = PIX_FMT_YUV422P10;
3040             else
3041                 s->avctx->pix_fmt = PIX_FMT_YUV420P10;
3042             break;
3043         case 8:
3044             if (CHROMA444) {
3045                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ444P
3046                                                                                   : PIX_FMT_YUV444P;
3047                     if (s->avctx->colorspace == AVCOL_SPC_RGB) {
3048                         s->avctx->pix_fmt = PIX_FMT_GBR24P;
3049                         av_log(h->s.avctx, AV_LOG_DEBUG, "Detected GBR colorspace.\n");
3050                     } else if (s->avctx->colorspace == AVCOL_SPC_YCGCO) {
3051                         av_log(h->s.avctx, AV_LOG_WARNING, "Detected unsupported YCgCo colorspace.\n");
3052                     }
3053             } else if (CHROMA422) {
3054                 s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ422P
3055                                                                               : PIX_FMT_YUV422P;
3056             } else {
3057                 s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
3058                                                          s->avctx->codec->pix_fmts ?
3059                                                          s->avctx->codec->pix_fmts :
3060                                                          s->avctx->color_range == AVCOL_RANGE_JPEG ?
3061                                                          hwaccel_pixfmt_list_h264_jpeg_420 :
3062                                                          ff_hwaccel_pixfmt_list_420);
3063             }
3064             break;
3065         default:
3066             av_log(s->avctx, AV_LOG_ERROR,
3067                    "Unsupported bit depth: %d\n", h->sps.bit_depth_luma);
3068             return AVERROR_INVALIDDATA;
3069         }
3070
3071         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id,
3072                                             s->avctx->pix_fmt);
3073
3074         if (ff_MPV_common_init(s) < 0) {
3075             av_log(h->s.avctx, AV_LOG_ERROR, "ff_MPV_common_init() failed.\n");
3076             return -1;
3077         }
3078         s->first_field = 0;
3079         h->prev_interlaced_frame = 1;
3080
3081         init_scan_tables(h);
3082         if (ff_h264_alloc_tables(h) < 0) {
3083             av_log(h->s.avctx, AV_LOG_ERROR,
3084                    "Could not allocate memory for h264\n");
3085             return AVERROR(ENOMEM);
3086         }
3087
3088         if (!HAVE_THREADS || !(s->avctx->active_thread_type & FF_THREAD_SLICE)) {
3089             if (context_init(h) < 0) {
3090                 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
3091                 return -1;
3092             }
3093         } else {
3094             for (i = 1; i < s->slice_context_count; i++) {
3095                 H264Context *c;
3096                 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
3097                 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
3098                 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
3099                 c->h264dsp     = h->h264dsp;
3100                 c->sps         = h->sps;
3101                 c->pps         = h->pps;
3102                 c->pixel_shift = h->pixel_shift;
3103                 c->cur_chroma_format_idc = h->cur_chroma_format_idc;
3104                 init_scan_tables(c);
3105                 clone_tables(c, h, i);
3106             }
3107
3108             for (i = 0; i < s->slice_context_count; i++)
3109                 if (context_init(h->thread_context[i]) < 0) {
3110                     av_log(h->s.avctx, AV_LOG_ERROR,
3111                            "context_init() failed.\n");
3112                     return -1;
3113                 }
3114         }
3115     }
3116
3117     if (h == h0 && h->dequant_coeff_pps != pps_id) {
3118         h->dequant_coeff_pps = pps_id;
3119         init_dequant_tables(h);
3120     }
3121
3122     h->frame_num = get_bits(&s->gb, h->sps.log2_max_frame_num);
3123
3124     h->mb_mbaff        = 0;
3125     h->mb_aff_frame    = 0;
3126     last_pic_structure = s0->picture_structure;
3127     last_pic_dropable  = s->dropable;
3128     s->dropable        = h->nal_ref_idc == 0;
3129     if (h->sps.frame_mbs_only_flag) {
3130         s->picture_structure = PICT_FRAME;
3131     } else {
3132         if (!h->sps.direct_8x8_inference_flag && slice_type == AV_PICTURE_TYPE_B) {
3133             av_log(h->s.avctx, AV_LOG_ERROR, "This stream was generated by a broken encoder, invalid 8x8 inference\n");
3134             return -1;
3135         }
3136         if (get_bits1(&s->gb)) { // field_pic_flag
3137             s->picture_structure = PICT_TOP_FIELD + get_bits1(&s->gb); // bottom_field_flag
3138         } else {
3139             s->picture_structure = PICT_FRAME;
3140             h->mb_aff_frame      = h->sps.mb_aff;
3141         }
3142     }
3143     h->mb_field_decoding_flag = s->picture_structure != PICT_FRAME;
3144
3145     if (h0->current_slice != 0) {
3146         if (last_pic_structure != s->picture_structure ||
3147             last_pic_dropable  != s->dropable) {
3148             av_log(h->s.avctx, AV_LOG_ERROR,
3149                    "Changing field mode (%d -> %d) between slices is not allowed\n",
3150                    last_pic_structure, s->picture_structure);
3151             s->picture_structure = last_pic_structure;
3152             s->dropable          = last_pic_dropable;
3153             return AVERROR_INVALIDDATA;
3154         }
3155     } else {
3156         /* Shorten frame num gaps so we don't have to allocate reference
3157          * frames just to throw them away */
3158         if (h->frame_num != h->prev_frame_num && h->prev_frame_num >= 0) {
3159             int unwrap_prev_frame_num = h->prev_frame_num;
3160             int max_frame_num         = 1 << h->sps.log2_max_frame_num;
3161
3162             if (unwrap_prev_frame_num > h->frame_num)
3163                 unwrap_prev_frame_num -= max_frame_num;
3164
3165             if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
3166                 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
3167                 if (unwrap_prev_frame_num < 0)
3168                     unwrap_prev_frame_num += max_frame_num;
3169
3170                 h->prev_frame_num = unwrap_prev_frame_num;
3171             }
3172         }
3173
3174         /* See if we have a decoded first field looking for a pair...
3175          * Here, we're using that to see if we should mark previously
3176          * decode frames as "finished".
3177          * We have to do that before the "dummy" in-between frame allocation,
3178          * since that can modify s->current_picture_ptr. */
3179         if (s0->first_field) {
3180             assert(s0->current_picture_ptr);
3181             assert(s0->current_picture_ptr->f.data[0]);
3182             assert(s0->current_picture_ptr->f.reference != DELAYED_PIC_REF);
3183
3184             /* Mark old field/frame as completed */
3185             if (!last_pic_dropable && s0->current_picture_ptr->owner2 == s0) {
3186                 ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
3187                                           last_pic_structure == PICT_BOTTOM_FIELD);
3188             }
3189
3190             /* figure out if we have a complementary field pair */
3191             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
3192                 /* Previous field is unmatched. Don't display it, but let it
3193                  * remain for reference if marked as such. */
3194                 if (!last_pic_dropable && last_pic_structure != PICT_FRAME) {
3195                     ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
3196                                               last_pic_structure == PICT_TOP_FIELD);
3197                 }
3198             } else {
3199                 if (s0->current_picture_ptr->frame_num != h->frame_num) {
3200                     /* This and previous field were reference, but had
3201                      * different frame_nums. Consider this field first in
3202                      * pair. Throw away previous field except for reference
3203                      * purposes. */
3204                     if (!last_pic_dropable && last_pic_structure != PICT_FRAME) {
3205                         ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
3206                                                   last_pic_structure == PICT_TOP_FIELD);
3207                     }
3208                 } else {
3209                     /* Second field in complementary pair */
3210                     if (!((last_pic_structure   == PICT_TOP_FIELD &&
3211                            s->picture_structure == PICT_BOTTOM_FIELD) ||
3212                           (last_pic_structure   == PICT_BOTTOM_FIELD &&
3213                            s->picture_structure == PICT_TOP_FIELD))) {
3214                         av_log(s->avctx, AV_LOG_ERROR,
3215                                "Invalid field mode combination %d/%d\n",
3216                                last_pic_structure, s->picture_structure);
3217                         s->picture_structure = last_pic_structure;
3218                         s->dropable          = last_pic_dropable;
3219                         return AVERROR_INVALIDDATA;
3220                     } else if (last_pic_dropable != s->dropable) {
3221                         av_log(s->avctx, AV_LOG_ERROR,
3222                                "Cannot combine reference and non-reference fields in the same frame\n");
3223                         av_log_ask_for_sample(s->avctx, NULL);
3224                         s->picture_structure = last_pic_structure;
3225                         s->dropable          = last_pic_dropable;
3226                         return AVERROR_INVALIDDATA;
3227                     }
3228
3229                     /* Take ownership of this buffer. Note that if another thread owned
3230                      * the first field of this buffer, we're not operating on that pointer,
3231                      * so the original thread is still responsible for reporting progress
3232                      * on that first field (or if that was us, we just did that above).
3233                      * By taking ownership, we assign responsibility to ourselves to
3234                      * report progress on the second field. */
3235                     s0->current_picture_ptr->owner2 = s0;
3236                 }
3237             }
3238         }
3239
3240         while (h->frame_num != h->prev_frame_num && h->prev_frame_num >= 0 &&
3241                h->frame_num != (h->prev_frame_num + 1) % (1 << h->sps.log2_max_frame_num)) {
3242             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
3243             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n",
3244                    h->frame_num, h->prev_frame_num);
3245             if (ff_h264_frame_start(h) < 0)
3246                 return -1;
3247             h->prev_frame_num++;
3248             h->prev_frame_num %= 1 << h->sps.log2_max_frame_num;
3249             s->current_picture_ptr->frame_num = h->prev_frame_num;
3250             ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX, 0);
3251             ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX, 1);
3252             ff_generate_sliding_window_mmcos(h);
3253             if (ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index) < 0 &&
3254                 (s->avctx->err_recognition & AV_EF_EXPLODE))
3255                 return AVERROR_INVALIDDATA;
3256             /* Error concealment: if a ref is missing, copy the previous ref in its place.
3257              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
3258              * about there being no actual duplicates.
3259              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
3260              * concealing a lost frame, this probably isn't noticeable by comparison, but it should
3261              * be fixed. */
3262             if (h->short_ref_count) {
3263                 if (prev) {
3264                     av_image_copy(h->short_ref[0]->f.data, h->short_ref[0]->f.linesize,
3265                                   (const uint8_t **)prev->f.data, prev->f.linesize,
3266                                   s->avctx->pix_fmt, s->mb_width * 16, s->mb_height * 16);
3267                     h->short_ref[0]->poc = prev->poc + 2;
3268                 }
3269                 h->short_ref[0]->frame_num = h->prev_frame_num;
3270             }
3271         }