Merge remote-tracking branch 'qatar/master'
[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 const uint16_t ff_h264_mb_sizes[4] = { 256, 384, 512, 768 };
52
53 static const uint8_t rem6[QP_MAX_NUM + 1] = {
54     0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2,
55     3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5,
56     0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
57 };
58
59 static const uint8_t div6[QP_MAX_NUM + 1] = {
60     0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3,  3,  3,
61     3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6,  6,  6,
62     7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 10, 10, 10, 10,
63 };
64
65 static const enum PixelFormat hwaccel_pixfmt_list_h264_jpeg_420[] = {
66     PIX_FMT_DXVA2_VLD,
67     PIX_FMT_VAAPI_VLD,
68     PIX_FMT_VDA_VLD,
69     PIX_FMT_YUVJ420P,
70     PIX_FMT_NONE
71 };
72
73 /**
74  * Check if the top & left blocks are available if needed and
75  * change the dc mode so it only uses the available blocks.
76  */
77 int ff_h264_check_intra4x4_pred_mode(H264Context *h)
78 {
79     MpegEncContext *const s     = &h->s;
80     static const int8_t top[12] = {
81         -1, 0, LEFT_DC_PRED, -1, -1, -1, -1, -1, 0
82     };
83     static const int8_t left[12] = {
84         0, -1, TOP_DC_PRED, 0, -1, -1, -1, 0, -1, DC_128_PRED
85     };
86     int i;
87
88     if (!(h->top_samples_available & 0x8000)) {
89         for (i = 0; i < 4; i++) {
90             int status = top[h->intra4x4_pred_mode_cache[scan8[0] + i]];
91             if (status < 0) {
92                 av_log(h->s.avctx, AV_LOG_ERROR,
93                        "top block unavailable for requested intra4x4 mode %d at %d %d\n",
94                        status, s->mb_x, s->mb_y);
95                 return -1;
96             } else if (status) {
97                 h->intra4x4_pred_mode_cache[scan8[0] + i] = status;
98             }
99         }
100     }
101
102     if ((h->left_samples_available & 0x8888) != 0x8888) {
103         static const int mask[4] = { 0x8000, 0x2000, 0x80, 0x20 };
104         for (i = 0; i < 4; i++)
105             if (!(h->left_samples_available & mask[i])) {
106                 int status = left[h->intra4x4_pred_mode_cache[scan8[0] + 8 * i]];
107                 if (status < 0) {
108                     av_log(h->s.avctx, AV_LOG_ERROR,
109                            "left block unavailable for requested intra4x4 mode %d at %d %d\n",
110                            status, s->mb_x, s->mb_y);
111                     return -1;
112                 } else if (status) {
113                     h->intra4x4_pred_mode_cache[scan8[0] + 8 * i] = status;
114                 }
115             }
116     }
117
118     return 0;
119 } // FIXME cleanup like ff_h264_check_intra_pred_mode
120
121 /**
122  * Check if the top & left blocks are available if needed and
123  * change the dc mode so it only uses the available blocks.
124  */
125 int ff_h264_check_intra_pred_mode(H264Context *h, int mode, int is_chroma)
126 {
127     MpegEncContext *const s     = &h->s;
128     static const int8_t top[7]  = { LEFT_DC_PRED8x8, 1, -1, -1 };
129     static const int8_t left[7] = { TOP_DC_PRED8x8, -1, 2, -1, DC_128_PRED8x8 };
130
131     if (mode > 6U) {
132         av_log(h->s.avctx, AV_LOG_ERROR,
133                "out of range intra chroma pred mode at %d %d\n",
134                s->mb_x, s->mb_y);
135         return -1;
136     }
137
138     if (!(h->top_samples_available & 0x8000)) {
139         mode = top[mode];
140         if (mode < 0) {
141             av_log(h->s.avctx, AV_LOG_ERROR,
142                    "top block unavailable for requested intra mode at %d %d\n",
143                    s->mb_x, s->mb_y);
144             return -1;
145         }
146     }
147
148     if ((h->left_samples_available & 0x8080) != 0x8080) {
149         mode = left[mode];
150         if (is_chroma && (h->left_samples_available & 0x8080)) {
151             // mad cow disease mode, aka MBAFF + constrained_intra_pred
152             mode = ALZHEIMER_DC_L0T_PRED8x8 +
153                    (!(h->left_samples_available & 0x8000)) +
154                    2 * (mode == DC_128_PRED8x8);
155         }
156         if (mode < 0) {
157             av_log(h->s.avctx, AV_LOG_ERROR,
158                    "left block unavailable for requested intra mode at %d %d\n",
159                    s->mb_x, s->mb_y);
160             return -1;
161         }
162     }
163
164     return mode;
165 }
166
167 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src,
168                                   int *dst_length, int *consumed, int length)
169 {
170     int i, si, di;
171     uint8_t *dst;
172     int bufidx;
173
174     // src[0]&0x80; // forbidden bit
175     h->nal_ref_idc   = src[0] >> 5;
176     h->nal_unit_type = src[0] & 0x1F;
177
178     src++;
179     length--;
180
181 #if HAVE_FAST_UNALIGNED
182 #if HAVE_FAST_64BIT
183 #define RS 7
184     for (i = 0; i + 1 < length; i += 9) {
185         if (!((~AV_RN64A(src + i) &
186                (AV_RN64A(src + i) - 0x0100010001000101ULL)) &
187               0x8000800080008080ULL))
188 #else
189 #define RS 3
190     for (i = 0; i + 1 < length; i += 5) {
191         if (!((~AV_RN32A(src + i) &
192                (AV_RN32A(src + i) - 0x01000101U)) &
193               0x80008080U))
194 #endif
195             continue;
196         if (i > 0 && !src[i])
197             i--;
198         while (src[i])
199             i++;
200 #else
201 #define RS 0
202     for (i = 0; i + 1 < length; i += 2) {
203         if (src[i])
204             continue;
205         if (i > 0 && src[i - 1] == 0)
206             i--;
207 #endif
208         if (i + 2 < length && src[i + 1] == 0 && src[i + 2] <= 3) {
209             if (src[i + 2] != 3) {
210                 /* startcode, so we must be past the end */
211                 length = i;
212             }
213             break;
214         }
215         i -= RS;
216     }
217
218     // use second escape buffer for inter data
219     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0;
220
221     si = h->rbsp_buffer_size[bufidx];
222     av_fast_padded_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+MAX_MBPAIR_SIZE);
223     dst = h->rbsp_buffer[bufidx];
224
225     if (dst == NULL)
226         return NULL;
227
228     if(i>=length-1){ //no escaped 0
229         *dst_length= length;
230         *consumed= length+1; //+1 for the header
231         if(h->s.avctx->flags2 & CODEC_FLAG2_FAST){
232             return src;
233         }else{
234             memcpy(dst, src, length);
235             return dst;
236         }
237     }
238
239     // printf("decoding esc\n");
240     memcpy(dst, src, i);
241     si = di = i;
242     while (si + 2 < length) {
243         // remove escapes (very rare 1:2^22)
244         if (src[si + 2] > 3) {
245             dst[di++] = src[si++];
246             dst[di++] = src[si++];
247         } else if (src[si] == 0 && src[si + 1] == 0) {
248             if (src[si + 2] == 3) { // escape
249                 dst[di++]  = 0;
250                 dst[di++]  = 0;
251                 si        += 3;
252                 continue;
253             } else // next start code
254                 goto nsc;
255         }
256
257         dst[di++] = src[si++];
258     }
259     while (si < length)
260         dst[di++] = src[si++];
261 nsc:
262
263     memset(dst + di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
264
265     *dst_length = di;
266     *consumed   = si + 1; // +1 for the header
267     /* FIXME store exact number of bits in the getbitcontext
268      * (it is needed for decoding) */
269     return dst;
270 }
271
272 /**
273  * Identify the exact end of the bitstream
274  * @return the length of the trailing, or 0 if damaged
275  */
276 static int decode_rbsp_trailing(H264Context *h, const uint8_t *src)
277 {
278     int v = *src;
279     int r;
280
281     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
282
283     for (r = 1; r < 9; r++) {
284         if (v & 1)
285             return r;
286         v >>= 1;
287     }
288     return 0;
289 }
290
291 static inline int get_lowest_part_list_y(H264Context *h, Picture *pic, int n,
292                                          int height, int y_offset, int list)
293 {
294     int raw_my        = h->mv_cache[list][scan8[n]][1];
295     int filter_height = (raw_my & 3) ? 2 : 0;
296     int full_my       = (raw_my >> 2) + y_offset;
297     int top           = full_my - filter_height;
298     int bottom        = full_my + filter_height + height;
299
300     return FFMAX(abs(top), bottom);
301 }
302
303 static inline void get_lowest_part_y(H264Context *h, int refs[2][48], int n,
304                                      int height, int y_offset, int list0,
305                                      int list1, int *nrefs)
306 {
307     MpegEncContext *const s = &h->s;
308     int my;
309
310     y_offset += 16 * (s->mb_y >> MB_FIELD);
311
312     if (list0) {
313         int ref_n    = h->ref_cache[0][scan8[n]];
314         Picture *ref = &h->ref_list[0][ref_n];
315
316         // Error resilience puts the current picture in the ref list.
317         // Don't try to wait on these as it will cause a deadlock.
318         // Fields can wait on each other, though.
319         if (ref->f.thread_opaque   != s->current_picture.f.thread_opaque ||
320             (ref->f.reference & 3) != s->picture_structure) {
321             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 0);
322             if (refs[0][ref_n] < 0)
323                 nrefs[0] += 1;
324             refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
325         }
326     }
327
328     if (list1) {
329         int ref_n    = h->ref_cache[1][scan8[n]];
330         Picture *ref = &h->ref_list[1][ref_n];
331
332         if (ref->f.thread_opaque   != s->current_picture.f.thread_opaque ||
333             (ref->f.reference & 3) != s->picture_structure) {
334             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 1);
335             if (refs[1][ref_n] < 0)
336                 nrefs[1] += 1;
337             refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
338         }
339     }
340 }
341
342 /**
343  * Wait until all reference frames are available for MC operations.
344  *
345  * @param h the H264 context
346  */
347 static void await_references(H264Context *h)
348 {
349     MpegEncContext *const s = &h->s;
350     const int mb_xy   = h->mb_xy;
351     const int mb_type = s->current_picture.f.mb_type[mb_xy];
352     int refs[2][48];
353     int nrefs[2] = { 0 };
354     int ref, list;
355
356     memset(refs, -1, sizeof(refs));
357
358     if (IS_16X16(mb_type)) {
359         get_lowest_part_y(h, refs, 0, 16, 0,
360                           IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
361     } else if (IS_16X8(mb_type)) {
362         get_lowest_part_y(h, refs, 0, 8, 0,
363                           IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
364         get_lowest_part_y(h, refs, 8, 8, 8,
365                           IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
366     } else if (IS_8X16(mb_type)) {
367         get_lowest_part_y(h, refs, 0, 16, 0,
368                           IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
369         get_lowest_part_y(h, refs, 4, 16, 0,
370                           IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
371     } else {
372         int i;
373
374         assert(IS_8X8(mb_type));
375
376         for (i = 0; i < 4; i++) {
377             const int sub_mb_type = h->sub_mb_type[i];
378             const int n           = 4 * i;
379             int y_offset          = (i & 2) << 2;
380
381             if (IS_SUB_8X8(sub_mb_type)) {
382                 get_lowest_part_y(h, refs, n, 8, y_offset,
383                                   IS_DIR(sub_mb_type, 0, 0),
384                                   IS_DIR(sub_mb_type, 0, 1),
385                                   nrefs);
386             } else if (IS_SUB_8X4(sub_mb_type)) {
387                 get_lowest_part_y(h, refs, n, 4, y_offset,
388                                   IS_DIR(sub_mb_type, 0, 0),
389                                   IS_DIR(sub_mb_type, 0, 1),
390                                   nrefs);
391                 get_lowest_part_y(h, refs, n + 2, 4, y_offset + 4,
392                                   IS_DIR(sub_mb_type, 0, 0),
393                                   IS_DIR(sub_mb_type, 0, 1),
394                                   nrefs);
395             } else if (IS_SUB_4X8(sub_mb_type)) {
396                 get_lowest_part_y(h, refs, n, 8, y_offset,
397                                   IS_DIR(sub_mb_type, 0, 0),
398                                   IS_DIR(sub_mb_type, 0, 1),
399                                   nrefs);
400                 get_lowest_part_y(h, refs, n + 1, 8, y_offset,
401                                   IS_DIR(sub_mb_type, 0, 0),
402                                   IS_DIR(sub_mb_type, 0, 1),
403                                   nrefs);
404             } else {
405                 int j;
406                 assert(IS_SUB_4X4(sub_mb_type));
407                 for (j = 0; j < 4; j++) {
408                     int sub_y_offset = y_offset + 2 * (j & 2);
409                     get_lowest_part_y(h, refs, n + j, 4, sub_y_offset,
410                                       IS_DIR(sub_mb_type, 0, 0),
411                                       IS_DIR(sub_mb_type, 0, 1),
412                                       nrefs);
413                 }
414             }
415         }
416     }
417
418     for (list = h->list_count - 1; list >= 0; list--)
419         for (ref = 0; ref < 48 && nrefs[list]; ref++) {
420             int row = refs[list][ref];
421             if (row >= 0) {
422                 Picture *ref_pic      = &h->ref_list[list][ref];
423                 int ref_field         = ref_pic->f.reference - 1;
424                 int ref_field_picture = ref_pic->field_picture;
425                 int pic_height        = 16 * s->mb_height >> ref_field_picture;
426
427                 row <<= MB_MBAFF;
428                 nrefs[list]--;
429
430                 if (!FIELD_PICTURE && ref_field_picture) { // frame referencing two fields
431                     ff_thread_await_progress(&ref_pic->f,
432                                              FFMIN((row >> 1) - !(row & 1),
433                                                    pic_height - 1),
434                                              1);
435                     ff_thread_await_progress(&ref_pic->f,
436                                              FFMIN((row >> 1), pic_height - 1),
437                                              0);
438                 } else if (FIELD_PICTURE && !ref_field_picture) { // field referencing one field of a frame
439                     ff_thread_await_progress(&ref_pic->f,
440                                              FFMIN(row * 2 + ref_field,
441                                                    pic_height - 1),
442                                              0);
443                 } else if (FIELD_PICTURE) {
444                     ff_thread_await_progress(&ref_pic->f,
445                                              FFMIN(row, pic_height - 1),
446                                              ref_field);
447                 } else {
448                     ff_thread_await_progress(&ref_pic->f,
449                                              FFMIN(row, pic_height - 1),
450                                              0);
451                 }
452             }
453         }
454 }
455
456 static av_always_inline void mc_dir_part(H264Context *h, Picture *pic,
457                                          int n, int square, int height,
458                                          int delta, int list,
459                                          uint8_t *dest_y, uint8_t *dest_cb,
460                                          uint8_t *dest_cr,
461                                          int src_x_offset, int src_y_offset,
462                                          qpel_mc_func *qpix_op,
463                                          h264_chroma_mc_func chroma_op,
464                                          int pixel_shift, int chroma_idc)
465 {
466     MpegEncContext *const s = &h->s;
467     const int mx      = h->mv_cache[list][scan8[n]][0] + src_x_offset * 8;
468     int my            = h->mv_cache[list][scan8[n]][1] + src_y_offset * 8;
469     const int luma_xy = (mx & 3) + ((my & 3) << 2);
470     int offset        = ((mx >> 2) << pixel_shift) + (my >> 2) * h->mb_linesize;
471     uint8_t *src_y    = pic->f.data[0] + offset;
472     uint8_t *src_cb, *src_cr;
473     int extra_width  = h->emu_edge_width;
474     int extra_height = h->emu_edge_height;
475     int emu = 0;
476     const int full_mx    = mx >> 2;
477     const int full_my    = my >> 2;
478     const int pic_width  = 16 * s->mb_width;
479     const int pic_height = 16 * s->mb_height >> MB_FIELD;
480     int ysh;
481
482     if (mx & 7)
483         extra_width -= 3;
484     if (my & 7)
485         extra_height -= 3;
486
487     if (full_mx                <          0 - extra_width  ||
488         full_my                <          0 - extra_height ||
489         full_mx + 16 /*FIXME*/ > pic_width  + extra_width  ||
490         full_my + 16 /*FIXME*/ > pic_height + extra_height) {
491         s->dsp.emulated_edge_mc(s->edge_emu_buffer,
492                                 src_y - (2 << pixel_shift) - 2 * h->mb_linesize,
493                                 h->mb_linesize,
494                                 16 + 5, 16 + 5 /*FIXME*/, full_mx - 2,
495                                 full_my - 2, pic_width, pic_height);
496         src_y = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
497         emu   = 1;
498     }
499
500     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); // FIXME try variable height perhaps?
501     if (!square)
502         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
503
504     if (CONFIG_GRAY && s->flags & CODEC_FLAG_GRAY)
505         return;
506
507     if (chroma_idc == 3 /* yuv444 */) {
508         src_cb = pic->f.data[1] + offset;
509         if (emu) {
510             s->dsp.emulated_edge_mc(s->edge_emu_buffer,
511                                     src_cb - (2 << pixel_shift) - 2 * h->mb_linesize,
512                                     h->mb_linesize,
513                                     16 + 5, 16 + 5 /*FIXME*/,
514                                     full_mx - 2, full_my - 2,
515                                     pic_width, pic_height);
516             src_cb = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
517         }
518         qpix_op[luma_xy](dest_cb, src_cb, h->mb_linesize); // FIXME try variable height perhaps?
519         if (!square)
520             qpix_op[luma_xy](dest_cb + delta, src_cb + delta, h->mb_linesize);
521
522         src_cr = pic->f.data[2] + offset;
523         if (emu) {
524             s->dsp.emulated_edge_mc(s->edge_emu_buffer,
525                                     src_cr - (2 << pixel_shift) - 2 * h->mb_linesize,
526                                     h->mb_linesize,
527                                     16 + 5, 16 + 5 /*FIXME*/,
528                                     full_mx - 2, full_my - 2,
529                                     pic_width, pic_height);
530             src_cr = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
531         }
532         qpix_op[luma_xy](dest_cr, src_cr, h->mb_linesize); // FIXME try variable height perhaps?
533         if (!square)
534             qpix_op[luma_xy](dest_cr + delta, src_cr + delta, h->mb_linesize);
535         return;
536     }
537
538     ysh = 3 - (chroma_idc == 2 /* yuv422 */);
539     if (chroma_idc == 1 /* yuv420 */ && MB_FIELD) {
540         // chroma offset when predicting from a field of opposite parity
541         my  += 2 * ((s->mb_y & 1) - (pic->f.reference - 1));
542         emu |= (my >> 3) < 0 || (my >> 3) + 8 >= (pic_height >> 1);
543     }
544
545     src_cb = pic->f.data[1] + ((mx >> 3) << pixel_shift) +
546              (my >> ysh) * h->mb_uvlinesize;
547     src_cr = pic->f.data[2] + ((mx >> 3) << pixel_shift) +
548              (my >> ysh) * h->mb_uvlinesize;
549
550     if (emu) {
551         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize,
552                                 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
553                                 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
554         src_cb = s->edge_emu_buffer;
555     }
556     chroma_op(dest_cb, src_cb, h->mb_uvlinesize,
557               height >> (chroma_idc == 1 /* yuv420 */),
558               mx & 7, (my << (chroma_idc == 2 /* yuv422 */)) & 7);
559
560     if (emu) {
561         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize,
562                                 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
563                                 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
564         src_cr = s->edge_emu_buffer;
565     }
566     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, height >> (chroma_idc == 1 /* yuv420 */),
567               mx & 7, (my << (chroma_idc == 2 /* yuv422 */)) & 7);
568 }
569
570 static av_always_inline void mc_part_std(H264Context *h, int n, int square,
571                                          int height, int delta,
572                                          uint8_t *dest_y, uint8_t *dest_cb,
573                                          uint8_t *dest_cr,
574                                          int x_offset, int y_offset,
575                                          qpel_mc_func *qpix_put,
576                                          h264_chroma_mc_func chroma_put,
577                                          qpel_mc_func *qpix_avg,
578                                          h264_chroma_mc_func chroma_avg,
579                                          int list0, int list1,
580                                          int pixel_shift, int chroma_idc)
581 {
582     MpegEncContext *const s       = &h->s;
583     qpel_mc_func *qpix_op         = qpix_put;
584     h264_chroma_mc_func chroma_op = chroma_put;
585
586     dest_y += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
587     if (chroma_idc == 3 /* yuv444 */) {
588         dest_cb += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
589         dest_cr += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
590     } else if (chroma_idc == 2 /* yuv422 */) {
591         dest_cb += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
592         dest_cr += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
593     } else { /* yuv420 */
594         dest_cb += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
595         dest_cr += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
596     }
597     x_offset += 8 * s->mb_x;
598     y_offset += 8 * (s->mb_y >> MB_FIELD);
599
600     if (list0) {
601         Picture *ref = &h->ref_list[0][h->ref_cache[0][scan8[n]]];
602         mc_dir_part(h, ref, n, square, height, delta, 0,
603                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
604                     qpix_op, chroma_op, pixel_shift, chroma_idc);
605
606         qpix_op   = qpix_avg;
607         chroma_op = chroma_avg;
608     }
609
610     if (list1) {
611         Picture *ref = &h->ref_list[1][h->ref_cache[1][scan8[n]]];
612         mc_dir_part(h, ref, n, square, height, delta, 1,
613                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
614                     qpix_op, chroma_op, pixel_shift, chroma_idc);
615     }
616 }
617
618 static av_always_inline void mc_part_weighted(H264Context *h, int n, int square,
619                                               int height, int delta,
620                                               uint8_t *dest_y, uint8_t *dest_cb,
621                                               uint8_t *dest_cr,
622                                               int x_offset, int y_offset,
623                                               qpel_mc_func *qpix_put,
624                                               h264_chroma_mc_func chroma_put,
625                                               h264_weight_func luma_weight_op,
626                                               h264_weight_func chroma_weight_op,
627                                               h264_biweight_func luma_weight_avg,
628                                               h264_biweight_func chroma_weight_avg,
629                                               int list0, int list1,
630                                               int pixel_shift, int chroma_idc)
631 {
632     MpegEncContext *const s = &h->s;
633     int chroma_height;
634
635     dest_y += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
636     if (chroma_idc == 3 /* yuv444 */) {
637         chroma_height     = height;
638         chroma_weight_avg = luma_weight_avg;
639         chroma_weight_op  = luma_weight_op;
640         dest_cb += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
641         dest_cr += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
642     } else if (chroma_idc == 2 /* yuv422 */) {
643         chroma_height = height;
644         dest_cb      += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
645         dest_cr      += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
646     } else { /* yuv420 */
647         chroma_height = height >> 1;
648         dest_cb      += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
649         dest_cr      += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
650     }
651     x_offset += 8 * s->mb_x;
652     y_offset += 8 * (s->mb_y >> MB_FIELD);
653
654     if (list0 && list1) {
655         /* don't optimize for luma-only case, since B-frames usually
656          * use implicit weights => chroma too. */
657         uint8_t *tmp_cb = s->obmc_scratchpad;
658         uint8_t *tmp_cr = s->obmc_scratchpad + (16 << pixel_shift);
659         uint8_t *tmp_y  = s->obmc_scratchpad + 16 * h->mb_uvlinesize;
660         int refn0       = h->ref_cache[0][scan8[n]];
661         int refn1       = h->ref_cache[1][scan8[n]];
662
663         mc_dir_part(h, &h->ref_list[0][refn0], n, square, height, delta, 0,
664                     dest_y, dest_cb, dest_cr,
665                     x_offset, y_offset, qpix_put, chroma_put,
666                     pixel_shift, chroma_idc);
667         mc_dir_part(h, &h->ref_list[1][refn1], n, square, height, delta, 1,
668                     tmp_y, tmp_cb, tmp_cr,
669                     x_offset, y_offset, qpix_put, chroma_put,
670                     pixel_shift, chroma_idc);
671
672         if (h->use_weight == 2) {
673             int weight0 = h->implicit_weight[refn0][refn1][s->mb_y & 1];
674             int weight1 = 64 - weight0;
675             luma_weight_avg(dest_y, tmp_y, h->mb_linesize,
676                             height, 5, weight0, weight1, 0);
677             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize,
678                               chroma_height, 5, weight0, weight1, 0);
679             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize,
680                               chroma_height, 5, weight0, weight1, 0);
681         } else {
682             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, height,
683                             h->luma_log2_weight_denom,
684                             h->luma_weight[refn0][0][0],
685                             h->luma_weight[refn1][1][0],
686                             h->luma_weight[refn0][0][1] +
687                             h->luma_weight[refn1][1][1]);
688             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, chroma_height,
689                               h->chroma_log2_weight_denom,
690                               h->chroma_weight[refn0][0][0][0],
691                               h->chroma_weight[refn1][1][0][0],
692                               h->chroma_weight[refn0][0][0][1] +
693                               h->chroma_weight[refn1][1][0][1]);
694             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, chroma_height,
695                               h->chroma_log2_weight_denom,
696                               h->chroma_weight[refn0][0][1][0],
697                               h->chroma_weight[refn1][1][1][0],
698                               h->chroma_weight[refn0][0][1][1] +
699                               h->chroma_weight[refn1][1][1][1]);
700         }
701     } else {
702         int list     = list1 ? 1 : 0;
703         int refn     = h->ref_cache[list][scan8[n]];
704         Picture *ref = &h->ref_list[list][refn];
705         mc_dir_part(h, ref, n, square, height, delta, list,
706                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
707                     qpix_put, chroma_put, pixel_shift, chroma_idc);
708
709         luma_weight_op(dest_y, h->mb_linesize, height,
710                        h->luma_log2_weight_denom,
711                        h->luma_weight[refn][list][0],
712                        h->luma_weight[refn][list][1]);
713         if (h->use_weight_chroma) {
714             chroma_weight_op(dest_cb, h->mb_uvlinesize, chroma_height,
715                              h->chroma_log2_weight_denom,
716                              h->chroma_weight[refn][list][0][0],
717                              h->chroma_weight[refn][list][0][1]);
718             chroma_weight_op(dest_cr, h->mb_uvlinesize, chroma_height,
719                              h->chroma_log2_weight_denom,
720                              h->chroma_weight[refn][list][1][0],
721                              h->chroma_weight[refn][list][1][1]);
722         }
723     }
724 }
725
726 static av_always_inline void mc_part(H264Context *h, int n, int square,
727                                      int height, int delta,
728                                      uint8_t *dest_y, uint8_t *dest_cb,
729                                      uint8_t *dest_cr,
730                                      int x_offset, int y_offset,
731                                      qpel_mc_func *qpix_put,
732                                      h264_chroma_mc_func chroma_put,
733                                      qpel_mc_func *qpix_avg,
734                                      h264_chroma_mc_func chroma_avg,
735                                      h264_weight_func *weight_op,
736                                      h264_biweight_func *weight_avg,
737                                      int list0, int list1,
738                                      int pixel_shift, int chroma_idc)
739 {
740     if ((h->use_weight == 2 && list0 && list1 &&
741          (h->implicit_weight[h->ref_cache[0][scan8[n]]][h->ref_cache[1][scan8[n]]][h->s.mb_y & 1] != 32)) ||
742         h->use_weight == 1)
743         mc_part_weighted(h, n, square, height, delta, dest_y, dest_cb, dest_cr,
744                          x_offset, y_offset, qpix_put, chroma_put,
745                          weight_op[0], weight_op[1], weight_avg[0],
746                          weight_avg[1], list0, list1, pixel_shift, chroma_idc);
747     else
748         mc_part_std(h, n, square, height, delta, dest_y, dest_cb, dest_cr,
749                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg,
750                     chroma_avg, list0, list1, pixel_shift, chroma_idc);
751 }
752
753 static av_always_inline void prefetch_motion(H264Context *h, int list,
754                                              int pixel_shift, int chroma_idc)
755 {
756     /* fetch pixels for estimated mv 4 macroblocks ahead
757      * optimized for 64byte cache lines */
758     MpegEncContext *const s = &h->s;
759     const int refn = h->ref_cache[list][scan8[0]];
760     if (refn >= 0) {
761         const int mx  = (h->mv_cache[list][scan8[0]][0] >> 2) + 16 * s->mb_x + 8;
762         const int my  = (h->mv_cache[list][scan8[0]][1] >> 2) + 16 * s->mb_y;
763         uint8_t **src = h->ref_list[list][refn].f.data;
764         int off       = (mx << pixel_shift) +
765                         (my + (s->mb_x & 3) * 4) * h->mb_linesize +
766                         (64 << pixel_shift);
767         s->dsp.prefetch(src[0] + off, s->linesize, 4);
768         if (chroma_idc == 3 /* yuv444 */) {
769             s->dsp.prefetch(src[1] + off, s->linesize, 4);
770             s->dsp.prefetch(src[2] + off, s->linesize, 4);
771         } else {
772             off= (((mx>>1)+64)<<pixel_shift) + ((my>>1) + (s->mb_x&7))*s->uvlinesize;
773             s->dsp.prefetch(src[1] + off, src[2] - src[1], 2);
774         }
775     }
776 }
777
778 static av_always_inline void hl_motion(H264Context *h, uint8_t *dest_y,
779                                        uint8_t *dest_cb, uint8_t *dest_cr,
780                                        qpel_mc_func(*qpix_put)[16],
781                                        h264_chroma_mc_func(*chroma_put),
782                                        qpel_mc_func(*qpix_avg)[16],
783                                        h264_chroma_mc_func(*chroma_avg),
784                                        h264_weight_func *weight_op,
785                                        h264_biweight_func *weight_avg,
786                                        int pixel_shift, int chroma_idc)
787 {
788     MpegEncContext *const s = &h->s;
789     const int mb_xy   = h->mb_xy;
790     const int mb_type = s->current_picture.f.mb_type[mb_xy];
791
792     assert(IS_INTER(mb_type));
793
794     if (HAVE_THREADS && (s->avctx->active_thread_type & FF_THREAD_FRAME))
795         await_references(h);
796     prefetch_motion(h, 0, pixel_shift, chroma_idc);
797
798     if (IS_16X16(mb_type)) {
799         mc_part(h, 0, 1, 16, 0, dest_y, dest_cb, dest_cr, 0, 0,
800                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
801                 weight_op, weight_avg,
802                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
803                 pixel_shift, chroma_idc);
804     } else if (IS_16X8(mb_type)) {
805         mc_part(h, 0, 0, 8, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 0,
806                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
807                 weight_op, weight_avg,
808                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
809                 pixel_shift, chroma_idc);
810         mc_part(h, 8, 0, 8, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 4,
811                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
812                 weight_op, weight_avg,
813                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
814                 pixel_shift, chroma_idc);
815     } else if (IS_8X16(mb_type)) {
816         mc_part(h, 0, 0, 16, 8 * h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
817                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
818                 &weight_op[1], &weight_avg[1],
819                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
820                 pixel_shift, chroma_idc);
821         mc_part(h, 4, 0, 16, 8 * h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
822                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
823                 &weight_op[1], &weight_avg[1],
824                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
825                 pixel_shift, chroma_idc);
826     } else {
827         int i;
828
829         assert(IS_8X8(mb_type));
830
831         for (i = 0; i < 4; i++) {
832             const int sub_mb_type = h->sub_mb_type[i];
833             const int n  = 4 * i;
834             int x_offset = (i & 1) << 2;
835             int y_offset = (i & 2) << 1;
836
837             if (IS_SUB_8X8(sub_mb_type)) {
838                 mc_part(h, n, 1, 8, 0, dest_y, dest_cb, dest_cr,
839                         x_offset, y_offset,
840                         qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
841                         &weight_op[1], &weight_avg[1],
842                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
843                         pixel_shift, chroma_idc);
844             } else if (IS_SUB_8X4(sub_mb_type)) {
845                 mc_part(h, n, 0, 4, 4 << pixel_shift, dest_y, dest_cb, dest_cr,
846                         x_offset, y_offset,
847                         qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
848                         &weight_op[1], &weight_avg[1],
849                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
850                         pixel_shift, chroma_idc);
851                 mc_part(h, n + 2, 0, 4, 4 << pixel_shift,
852                         dest_y, dest_cb, dest_cr, x_offset, y_offset + 2,
853                         qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
854                         &weight_op[1], &weight_avg[1],
855                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
856                         pixel_shift, chroma_idc);
857             } else if (IS_SUB_4X8(sub_mb_type)) {
858                 mc_part(h, n, 0, 8, 4 * h->mb_linesize,
859                         dest_y, dest_cb, dest_cr, x_offset, y_offset,
860                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
861                         &weight_op[2], &weight_avg[2],
862                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
863                         pixel_shift, chroma_idc);
864                 mc_part(h, n + 1, 0, 8, 4 * h->mb_linesize,
865                         dest_y, dest_cb, dest_cr, x_offset + 2, y_offset,
866                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
867                         &weight_op[2], &weight_avg[2],
868                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
869                         pixel_shift, chroma_idc);
870             } else {
871                 int j;
872                 assert(IS_SUB_4X4(sub_mb_type));
873                 for (j = 0; j < 4; j++) {
874                     int sub_x_offset = x_offset + 2 * (j & 1);
875                     int sub_y_offset = y_offset + (j & 2);
876                     mc_part(h, n + j, 1, 4, 0,
877                             dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
878                             qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
879                             &weight_op[2], &weight_avg[2],
880                             IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
881                             pixel_shift, chroma_idc);
882                 }
883             }
884         }
885     }
886
887     prefetch_motion(h, 1, pixel_shift, chroma_idc);
888 }
889
890 static av_always_inline void hl_motion_420(H264Context *h, uint8_t *dest_y,
891                                            uint8_t *dest_cb, uint8_t *dest_cr,
892                                            qpel_mc_func(*qpix_put)[16],
893                                            h264_chroma_mc_func(*chroma_put),
894                                            qpel_mc_func(*qpix_avg)[16],
895                                            h264_chroma_mc_func(*chroma_avg),
896                                            h264_weight_func *weight_op,
897                                            h264_biweight_func *weight_avg,
898                                            int pixel_shift)
899 {
900     hl_motion(h, dest_y, dest_cb, dest_cr, qpix_put, chroma_put,
901               qpix_avg, chroma_avg, weight_op, weight_avg, pixel_shift, 1);
902 }
903
904 static av_always_inline void hl_motion_422(H264Context *h, uint8_t *dest_y,
905                                            uint8_t *dest_cb, uint8_t *dest_cr,
906                                            qpel_mc_func(*qpix_put)[16],
907                                            h264_chroma_mc_func(*chroma_put),
908                                            qpel_mc_func(*qpix_avg)[16],
909                                            h264_chroma_mc_func(*chroma_avg),
910                                            h264_weight_func *weight_op,
911                                            h264_biweight_func *weight_avg,
912                                            int pixel_shift)
913 {
914     hl_motion(h, dest_y, dest_cb, dest_cr, qpix_put, chroma_put,
915               qpix_avg, chroma_avg, weight_op, weight_avg, pixel_shift, 2);
916 }
917
918 static void free_tables(H264Context *h, int free_rbsp)
919 {
920     int i;
921     H264Context *hx;
922
923     av_freep(&h->intra4x4_pred_mode);
924     av_freep(&h->chroma_pred_mode_table);
925     av_freep(&h->cbp_table);
926     av_freep(&h->mvd_table[0]);
927     av_freep(&h->mvd_table[1]);
928     av_freep(&h->direct_table);
929     av_freep(&h->non_zero_count);
930     av_freep(&h->slice_table_base);
931     h->slice_table = NULL;
932     av_freep(&h->list_counts);
933
934     av_freep(&h->mb2b_xy);
935     av_freep(&h->mb2br_xy);
936
937     for (i = 0; i < MAX_THREADS; i++) {
938         hx = h->thread_context[i];
939         if (!hx)
940             continue;
941         av_freep(&hx->top_borders[1]);
942         av_freep(&hx->top_borders[0]);
943         av_freep(&hx->s.obmc_scratchpad);
944         if (free_rbsp) {
945             av_freep(&hx->rbsp_buffer[1]);
946             av_freep(&hx->rbsp_buffer[0]);
947             hx->rbsp_buffer_size[0] = 0;
948             hx->rbsp_buffer_size[1] = 0;
949         }
950         if (i)
951             av_freep(&h->thread_context[i]);
952     }
953 }
954
955 static void init_dequant8_coeff_table(H264Context *h)
956 {
957     int i, j, q, x;
958     const int max_qp = 51 + 6 * (h->sps.bit_depth_luma - 8);
959
960     for (i = 0; i < 6; i++) {
961         h->dequant8_coeff[i] = h->dequant8_buffer[i];
962         for (j = 0; j < i; j++)
963             if (!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i],
964                         64 * sizeof(uint8_t))) {
965                 h->dequant8_coeff[i] = h->dequant8_buffer[j];
966                 break;
967             }
968         if (j < i)
969             continue;
970
971         for (q = 0; q < max_qp + 1; q++) {
972             int shift = div6[q];
973             int idx   = rem6[q];
974             for (x = 0; x < 64; x++)
975                 h->dequant8_coeff[i][q][(x >> 3) | ((x & 7) << 3)] =
976                     ((uint32_t)dequant8_coeff_init[idx][dequant8_coeff_init_scan[((x >> 1) & 12) | (x & 3)]] *
977                      h->pps.scaling_matrix8[i][x]) << shift;
978         }
979     }
980 }
981
982 static void init_dequant4_coeff_table(H264Context *h)
983 {
984     int i, j, q, x;
985     const int max_qp = 51 + 6 * (h->sps.bit_depth_luma - 8);
986     for (i = 0; i < 6; i++) {
987         h->dequant4_coeff[i] = h->dequant4_buffer[i];
988         for (j = 0; j < i; j++)
989             if (!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i],
990                         16 * sizeof(uint8_t))) {
991                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
992                 break;
993             }
994         if (j < i)
995             continue;
996
997         for (q = 0; q < max_qp + 1; q++) {
998             int shift = div6[q] + 2;
999             int idx   = rem6[q];
1000             for (x = 0; x < 16; x++)
1001                 h->dequant4_coeff[i][q][(x >> 2) | ((x << 2) & 0xF)] =
1002                     ((uint32_t)dequant4_coeff_init[idx][(x & 1) + ((x >> 2) & 1)] *
1003                      h->pps.scaling_matrix4[i][x]) << shift;
1004         }
1005     }
1006 }
1007
1008 static void init_dequant_tables(H264Context *h)
1009 {
1010     int i, x;
1011     init_dequant4_coeff_table(h);
1012     if (h->pps.transform_8x8_mode)
1013         init_dequant8_coeff_table(h);
1014     if (h->sps.transform_bypass) {
1015         for (i = 0; i < 6; i++)
1016             for (x = 0; x < 16; x++)
1017                 h->dequant4_coeff[i][0][x] = 1 << 6;
1018         if (h->pps.transform_8x8_mode)
1019             for (i = 0; i < 6; i++)
1020                 for (x = 0; x < 64; x++)
1021                     h->dequant8_coeff[i][0][x] = 1 << 6;
1022     }
1023 }
1024
1025 int ff_h264_alloc_tables(H264Context *h)
1026 {
1027     MpegEncContext *const s = &h->s;
1028     const int big_mb_num    = s->mb_stride * (s->mb_height + 1);
1029     const int row_mb_num    = 2*s->mb_stride*FFMAX(s->avctx->thread_count, 1);
1030     int x, y;
1031
1032     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode,
1033                       row_mb_num * 8 * sizeof(uint8_t), fail)
1034     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count,
1035                       big_mb_num * 48 * sizeof(uint8_t), fail)
1036     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base,
1037                       (big_mb_num + s->mb_stride) * sizeof(*h->slice_table_base), fail)
1038     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table,
1039                       big_mb_num * sizeof(uint16_t), fail)
1040     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table,
1041                       big_mb_num * sizeof(uint8_t), fail)
1042     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0],
1043                       16 * row_mb_num * sizeof(uint8_t), fail);
1044     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1],
1045                       16 * row_mb_num * sizeof(uint8_t), fail);
1046     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table,
1047                       4 * big_mb_num * sizeof(uint8_t), fail);
1048     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts,
1049                       big_mb_num * sizeof(uint8_t), fail)
1050
1051     memset(h->slice_table_base, -1,
1052            (big_mb_num + s->mb_stride) * sizeof(*h->slice_table_base));
1053     h->slice_table = h->slice_table_base + s->mb_stride * 2 + 1;
1054
1055     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy,
1056                       big_mb_num * sizeof(uint32_t), fail);
1057     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy,
1058                       big_mb_num * sizeof(uint32_t), fail);
1059     for (y = 0; y < s->mb_height; y++)
1060         for (x = 0; x < s->mb_width; x++) {
1061             const int mb_xy = x + y * s->mb_stride;
1062             const int b_xy  = 4 * x + 4 * y * h->b_stride;
1063
1064             h->mb2b_xy[mb_xy]  = b_xy;
1065             h->mb2br_xy[mb_xy] = 8 * (FMO ? mb_xy : (mb_xy % (2 * s->mb_stride)));
1066         }
1067
1068     s->obmc_scratchpad = NULL;
1069
1070     if (!h->dequant4_coeff[0])
1071         init_dequant_tables(h);
1072
1073     return 0;
1074
1075 fail:
1076     free_tables(h, 1);
1077     return -1;
1078 }
1079
1080 /**
1081  * Mimic alloc_tables(), but for every context thread.
1082  */
1083 static void clone_tables(H264Context *dst, H264Context *src, int i)
1084 {
1085     MpegEncContext *const s     = &src->s;
1086     dst->intra4x4_pred_mode     = src->intra4x4_pred_mode + i * 8 * 2 * s->mb_stride;
1087     dst->non_zero_count         = src->non_zero_count;
1088     dst->slice_table            = src->slice_table;
1089     dst->cbp_table              = src->cbp_table;
1090     dst->mb2b_xy                = src->mb2b_xy;
1091     dst->mb2br_xy               = src->mb2br_xy;
1092     dst->chroma_pred_mode_table = src->chroma_pred_mode_table;
1093     dst->mvd_table[0]           = src->mvd_table[0] + i * 8 * 2 * s->mb_stride;
1094     dst->mvd_table[1]           = src->mvd_table[1] + i * 8 * 2 * s->mb_stride;
1095     dst->direct_table           = src->direct_table;
1096     dst->list_counts            = src->list_counts;
1097     dst->s.obmc_scratchpad      = NULL;
1098     ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma,
1099                       src->sps.chroma_format_idc);
1100 }
1101
1102 /**
1103  * Init context
1104  * Allocate buffers which are not shared amongst multiple threads.
1105  */
1106 static int context_init(H264Context *h)
1107 {
1108     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0],
1109                       h->s.mb_width * 16 * 3 * sizeof(uint8_t) * 2, fail)
1110     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1],
1111                       h->s.mb_width * 16 * 3 * sizeof(uint8_t) * 2, fail)
1112
1113     h->ref_cache[0][scan8[5]  + 1] =
1114     h->ref_cache[0][scan8[7]  + 1] =
1115     h->ref_cache[0][scan8[13] + 1] =
1116     h->ref_cache[1][scan8[5]  + 1] =
1117     h->ref_cache[1][scan8[7]  + 1] =
1118     h->ref_cache[1][scan8[13] + 1] = PART_NOT_AVAILABLE;
1119
1120     return 0;
1121
1122 fail:
1123     return -1; // free_tables will clean up for us
1124 }
1125
1126 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
1127
1128 static av_cold void common_init(H264Context *h)
1129 {
1130     MpegEncContext *const s = &h->s;
1131
1132     s->width    = s->avctx->width;
1133     s->height   = s->avctx->height;
1134     s->codec_id = s->avctx->codec->id;
1135
1136     s->avctx->bits_per_raw_sample = 8;
1137     h->cur_chroma_format_idc = 1;
1138
1139     ff_h264dsp_init(&h->h264dsp,
1140                     s->avctx->bits_per_raw_sample, h->cur_chroma_format_idc);
1141     ff_h264_pred_init(&h->hpc, s->codec_id,
1142                       s->avctx->bits_per_raw_sample, h->cur_chroma_format_idc);
1143
1144     h->dequant_coeff_pps = -1;
1145     s->unrestricted_mv   = 1;
1146
1147     s->dsp.dct_bits = 16;
1148     /* needed so that IDCT permutation is known early */
1149     ff_dsputil_init(&s->dsp, s->avctx);
1150
1151     memset(h->pps.scaling_matrix4, 16, 6 * 16 * sizeof(uint8_t));
1152     memset(h->pps.scaling_matrix8, 16, 2 * 64 * sizeof(uint8_t));
1153 }
1154
1155 int ff_h264_decode_extradata(H264Context *h, const uint8_t *buf, int size)
1156 {
1157     AVCodecContext *avctx = h->s.avctx;
1158
1159     if (!buf || size <= 0)
1160         return -1;
1161
1162     if (buf[0] == 1) {
1163         int i, cnt, nalsize;
1164         const unsigned char *p = buf;
1165
1166         h->is_avc = 1;
1167
1168         if (size < 7) {
1169             av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
1170             return -1;
1171         }
1172         /* sps and pps in the avcC always have length coded with 2 bytes,
1173          * so put a fake nal_length_size = 2 while parsing them */
1174         h->nal_length_size = 2;
1175         // Decode sps from avcC
1176         cnt = *(p + 5) & 0x1f; // Number of sps
1177         p  += 6;
1178         for (i = 0; i < cnt; i++) {
1179             nalsize = AV_RB16(p) + 2;
1180             if(nalsize > size - (p-buf))
1181                 return -1;
1182             if (decode_nal_units(h, p, nalsize) < 0) {
1183                 av_log(avctx, AV_LOG_ERROR,
1184                        "Decoding sps %d from avcC failed\n", i);
1185                 return -1;
1186             }
1187             p += nalsize;
1188         }
1189         // Decode pps from avcC
1190         cnt = *(p++); // Number of pps
1191         for (i = 0; i < cnt; i++) {
1192             nalsize = AV_RB16(p) + 2;
1193             if(nalsize > size - (p-buf))
1194                 return -1;
1195             if (decode_nal_units(h, p, nalsize) < 0) {
1196                 av_log(avctx, AV_LOG_ERROR,
1197                        "Decoding pps %d from avcC failed\n", i);
1198                 return -1;
1199             }
1200             p += nalsize;
1201         }
1202         // Now store right nal length size, that will be used to parse all other nals
1203         h->nal_length_size = (buf[4] & 0x03) + 1;
1204     } else {
1205         h->is_avc = 0;
1206         if (decode_nal_units(h, buf, size) < 0)
1207             return -1;
1208     }
1209     return size;
1210 }
1211
1212 av_cold int ff_h264_decode_init(AVCodecContext *avctx)
1213 {
1214     H264Context *h = avctx->priv_data;
1215     MpegEncContext *const s = &h->s;
1216     int i;
1217
1218     ff_MPV_decode_defaults(s);
1219
1220     s->avctx = avctx;
1221     common_init(h);
1222
1223     s->out_format      = FMT_H264;
1224     s->workaround_bugs = avctx->workaround_bugs;
1225
1226     /* set defaults */
1227     // s->decode_mb = ff_h263_decode_mb;
1228     s->quarter_sample = 1;
1229     if (!avctx->has_b_frames)
1230         s->low_delay = 1;
1231
1232     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1233
1234     ff_h264_decode_init_vlc();
1235
1236     h->pixel_shift = 0;
1237     h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1238
1239     h->thread_context[0] = h;
1240     h->outputed_poc      = h->next_outputed_poc = INT_MIN;
1241     for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
1242         h->last_pocs[i] = INT_MIN;
1243     h->prev_poc_msb = 1 << 16;
1244     h->prev_frame_num = -1;
1245     h->x264_build   = -1;
1246     ff_h264_reset_sei(h);
1247     if (avctx->codec_id == CODEC_ID_H264) {
1248         if (avctx->ticks_per_frame == 1)
1249             s->avctx->time_base.den *= 2;
1250         avctx->ticks_per_frame = 2;
1251     }
1252
1253     if (avctx->extradata_size > 0 && avctx->extradata &&
1254         ff_h264_decode_extradata(h, avctx->extradata, avctx->extradata_size) < 0)
1255         return -1;
1256
1257     if (h->sps.bitstream_restriction_flag &&
1258         s->avctx->has_b_frames < h->sps.num_reorder_frames) {
1259         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1260         s->low_delay           = 0;
1261     }
1262
1263     return 0;
1264 }
1265
1266 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b) + (size))))
1267
1268 static void copy_picture_range(Picture **to, Picture **from, int count,
1269                                MpegEncContext *new_base,
1270                                MpegEncContext *old_base)
1271 {
1272     int i;
1273
1274     for (i = 0; i < count; i++) {
1275         assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1276                 IN_RANGE(from[i], old_base->picture,
1277                          sizeof(Picture) * old_base->picture_count) ||
1278                 !from[i]));
1279         to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1280     }
1281 }
1282
1283 static void copy_parameter_set(void **to, void **from, int count, int size)
1284 {
1285     int i;
1286
1287     for (i = 0; i < count; i++) {
1288         if (to[i] && !from[i])
1289             av_freep(&to[i]);
1290         else if (from[i] && !to[i])
1291             to[i] = av_malloc(size);
1292
1293         if (from[i])
1294             memcpy(to[i], from[i], size);
1295     }
1296 }
1297
1298 static int decode_init_thread_copy(AVCodecContext *avctx)
1299 {
1300     H264Context *h = avctx->priv_data;
1301
1302     if (!avctx->internal->is_copy)
1303         return 0;
1304     memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1305     memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1306
1307     return 0;
1308 }
1309
1310 #define copy_fields(to, from, start_field, end_field)                   \
1311     memcpy(&to->start_field, &from->start_field,                        \
1312            (char *)&to->end_field - (char *)&to->start_field)
1313
1314 static int decode_update_thread_context(AVCodecContext *dst,
1315                                         const AVCodecContext *src)
1316 {
1317     H264Context *h = dst->priv_data, *h1 = src->priv_data;
1318     MpegEncContext *const s = &h->s, *const s1 = &h1->s;
1319     int inited = s->context_initialized, err;
1320     int i;
1321
1322     if (dst == src)
1323         return 0;
1324
1325     err = ff_mpeg_update_thread_context(dst, src);
1326     if (err)
1327         return err;
1328
1329     // FIXME handle width/height changing
1330     if (!inited) {
1331         for (i = 0; i < MAX_SPS_COUNT; i++)
1332             av_freep(h->sps_buffers + i);
1333
1334         for (i = 0; i < MAX_PPS_COUNT; i++)
1335             av_freep(h->pps_buffers + i);
1336
1337         // copy all fields after MpegEnc
1338         memcpy(&h->s + 1, &h1->s + 1,
1339                sizeof(H264Context) - sizeof(MpegEncContext));
1340         memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1341         memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1342
1343         if (s1->context_initialized) {
1344         if (ff_h264_alloc_tables(h) < 0) {
1345             av_log(dst, AV_LOG_ERROR, "Could not allocate memory for h264\n");
1346             return AVERROR(ENOMEM);
1347         }
1348         context_init(h);
1349
1350         /* frame_start may not be called for the next thread (if it's decoding
1351          * a bottom field) so this has to be allocated here */
1352         h->s.obmc_scratchpad = av_malloc(16 * 6 * s->linesize);
1353         }
1354
1355         for (i = 0; i < 2; i++) {
1356             h->rbsp_buffer[i]      = NULL;
1357             h->rbsp_buffer_size[i] = 0;
1358         }
1359
1360         h->thread_context[0] = h;
1361
1362         s->dsp.clear_blocks(h->mb);
1363         s->dsp.clear_blocks(h->mb + (24 * 16 << h->pixel_shift));
1364     }
1365
1366     // extradata/NAL handling
1367     h->is_avc = h1->is_avc;
1368
1369     // SPS/PPS
1370     copy_parameter_set((void **)h->sps_buffers, (void **)h1->sps_buffers,
1371                        MAX_SPS_COUNT, sizeof(SPS));
1372     h->sps = h1->sps;
1373     copy_parameter_set((void **)h->pps_buffers, (void **)h1->pps_buffers,
1374                        MAX_PPS_COUNT, sizeof(PPS));
1375     h->pps = h1->pps;
1376
1377     // Dequantization matrices
1378     // FIXME these are big - can they be only copied when PPS changes?
1379     copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1380
1381     for (i = 0; i < 6; i++)
1382         h->dequant4_coeff[i] = h->dequant4_buffer[0] +
1383                                (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1384
1385     for (i = 0; i < 6; i++)
1386         h->dequant8_coeff[i] = h->dequant8_buffer[0] +
1387                                (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1388
1389     h->dequant_coeff_pps = h1->dequant_coeff_pps;
1390
1391     // POC timing
1392     copy_fields(h, h1, poc_lsb, redundant_pic_count);
1393
1394     // reference lists
1395     copy_fields(h, h1, ref_count, list_count);
1396     copy_fields(h, h1, ref_list, intra_gb);
1397     copy_fields(h, h1, short_ref, cabac_init_idc);
1398
1399     copy_picture_range(h->short_ref, h1->short_ref, 32, s, s1);
1400     copy_picture_range(h->long_ref, h1->long_ref, 32, s, s1);
1401     copy_picture_range(h->delayed_pic, h1->delayed_pic,
1402                        MAX_DELAYED_PIC_COUNT + 2, s, s1);
1403
1404     h->last_slice_type = h1->last_slice_type;
1405     h->sync            = h1->sync;
1406
1407     if (!s->current_picture_ptr)
1408         return 0;
1409
1410     if (!s->dropable) {
1411         err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1412         h->prev_poc_msb = h->poc_msb;
1413         h->prev_poc_lsb = h->poc_lsb;
1414     }
1415     h->prev_frame_num_offset = h->frame_num_offset;
1416     h->prev_frame_num        = h->frame_num;
1417     h->outputed_poc          = h->next_outputed_poc;
1418
1419     return err;
1420 }
1421
1422 int ff_h264_frame_start(H264Context *h)
1423 {
1424     MpegEncContext *const s = &h->s;
1425     int i;
1426     const int pixel_shift = h->pixel_shift;
1427
1428     if (ff_MPV_frame_start(s, s->avctx) < 0)
1429         return -1;
1430     ff_er_frame_start(s);
1431     /*
1432      * ff_MPV_frame_start uses pict_type to derive key_frame.
1433      * This is incorrect for H.264; IDR markings must be used.
1434      * Zero here; IDR markings per slice in frame or fields are ORed in later.
1435      * See decode_nal_units().
1436      */
1437     s->current_picture_ptr->f.key_frame = 0;
1438     s->current_picture_ptr->sync        = 0;
1439     s->current_picture_ptr->mmco_reset  = 0;
1440
1441     assert(s->linesize && s->uvlinesize);
1442
1443     for (i = 0; i < 16; i++) {
1444         h->block_offset[i]           = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * s->linesize * ((scan8[i] - scan8[0]) >> 3);
1445         h->block_offset[48 + i]      = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * s->linesize * ((scan8[i] - scan8[0]) >> 3);
1446     }
1447     for (i = 0; i < 16; i++) {
1448         h->block_offset[16 + i]      =
1449         h->block_offset[32 + i]      = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * s->uvlinesize * ((scan8[i] - scan8[0]) >> 3);
1450         h->block_offset[48 + 16 + i] =
1451         h->block_offset[48 + 32 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * s->uvlinesize * ((scan8[i] - scan8[0]) >> 3);
1452     }
1453
1454     /* can't be in alloc_tables because linesize isn't known there.
1455      * FIXME: redo bipred weight to not require extra buffer? */
1456     for (i = 0; i < s->slice_context_count; i++)
1457         if (h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1458             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16 * 6 * s->linesize);
1459
1460     /* Some macroblocks can be accessed before they're available in case
1461      * of lost slices, MBAFF or threading. */
1462     memset(h->slice_table, -1,
1463            (s->mb_height * s->mb_stride - 1) * sizeof(*h->slice_table));
1464
1465     // s->decode = (s->flags & CODEC_FLAG_PSNR) || !s->encoding ||
1466     //             s->current_picture.f.reference /* || h->contains_intra */ || 1;
1467
1468     /* We mark the current picture as non-reference after allocating it, so
1469      * that if we break out due to an error it can be released automatically
1470      * in the next ff_MPV_frame_start().
1471      * SVQ3 as well as most other codecs have only last/next/current and thus
1472      * get released even with set reference, besides SVQ3 and others do not
1473      * mark frames as reference later "naturally". */
1474     if (s->codec_id != CODEC_ID_SVQ3)
1475         s->current_picture_ptr->f.reference = 0;
1476
1477     s->current_picture_ptr->field_poc[0]     =
1478         s->current_picture_ptr->field_poc[1] = INT_MAX;
1479
1480     h->next_output_pic = NULL;
1481
1482     assert(s->current_picture_ptr->long_ref == 0);
1483
1484     return 0;
1485 }
1486
1487 /**
1488  * Run setup operations that must be run after slice header decoding.
1489  * This includes finding the next displayed frame.
1490  *
1491  * @param h h264 master context
1492  * @param setup_finished enough NALs have been read that we can call
1493  * ff_thread_finish_setup()
1494  */
1495 static void decode_postinit(H264Context *h, int setup_finished)
1496 {
1497     MpegEncContext *const s = &h->s;
1498     Picture *out = s->current_picture_ptr;
1499     Picture *cur = s->current_picture_ptr;
1500     int i, pics, out_of_order, out_idx;
1501
1502     s->current_picture_ptr->f.qscale_type = FF_QSCALE_TYPE_H264;
1503     s->current_picture_ptr->f.pict_type   = s->pict_type;
1504
1505     if (h->next_output_pic)
1506         return;
1507
1508     if (cur->field_poc[0] == INT_MAX || cur->field_poc[1] == INT_MAX) {
1509         /* FIXME: if we have two PAFF fields in one packet, we can't start
1510          * the next thread here. If we have one field per packet, we can.
1511          * The check in decode_nal_units() is not good enough to find this
1512          * yet, so we assume the worst for now. */
1513         // if (setup_finished)
1514         //    ff_thread_finish_setup(s->avctx);
1515         return;
1516     }
1517
1518     cur->f.interlaced_frame = 0;
1519     cur->f.repeat_pict      = 0;
1520
1521     /* Signal interlacing information externally. */
1522     /* Prioritize picture timing SEI information over used
1523      * decoding process if it exists. */
1524
1525     if (h->sps.pic_struct_present_flag) {
1526         switch (h->sei_pic_struct) {
1527         case SEI_PIC_STRUCT_FRAME:
1528             break;
1529         case SEI_PIC_STRUCT_TOP_FIELD:
1530         case SEI_PIC_STRUCT_BOTTOM_FIELD:
1531             cur->f.interlaced_frame = 1;
1532             break;
1533         case SEI_PIC_STRUCT_TOP_BOTTOM:
1534         case SEI_PIC_STRUCT_BOTTOM_TOP:
1535             if (FIELD_OR_MBAFF_PICTURE)
1536                 cur->f.interlaced_frame = 1;
1537             else
1538                 // try to flag soft telecine progressive
1539                 cur->f.interlaced_frame = h->prev_interlaced_frame;
1540             break;
1541         case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1542         case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1543             /* Signal the possibility of telecined film externally
1544              * (pic_struct 5,6). From these hints, let the applications
1545              * decide if they apply deinterlacing. */
1546             cur->f.repeat_pict = 1;
1547             break;
1548         case SEI_PIC_STRUCT_FRAME_DOUBLING:
1549             // Force progressive here, doubling interlaced frame is a bad idea.
1550             cur->f.repeat_pict = 2;
1551             break;
1552         case SEI_PIC_STRUCT_FRAME_TRIPLING:
1553             cur->f.repeat_pict = 4;
1554             break;
1555         }
1556
1557         if ((h->sei_ct_type & 3) &&
1558             h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1559             cur->f.interlaced_frame = (h->sei_ct_type & (1 << 1)) != 0;
1560     } else {
1561         /* Derive interlacing flag from used decoding process. */
1562         cur->f.interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1563     }
1564     h->prev_interlaced_frame = cur->f.interlaced_frame;
1565
1566     if (cur->field_poc[0] != cur->field_poc[1]) {
1567         /* Derive top_field_first from field pocs. */
1568         cur->f.top_field_first = cur->field_poc[0] < cur->field_poc[1];
1569     } else {
1570         if (cur->f.interlaced_frame || h->sps.pic_struct_present_flag) {
1571             /* Use picture timing SEI information. Even if it is a
1572              * information of a past frame, better than nothing. */
1573             if (h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM ||
1574                 h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1575                 cur->f.top_field_first = 1;
1576             else
1577                 cur->f.top_field_first = 0;
1578         } else {
1579             /* Most likely progressive */
1580             cur->f.top_field_first = 0;
1581         }
1582     }
1583
1584     cur->mmco_reset = h->mmco_reset;
1585     h->mmco_reset = 0;
1586     // FIXME do something with unavailable reference frames
1587
1588     /* Sort B-frames into display order */
1589
1590     if (h->sps.bitstream_restriction_flag &&
1591         s->avctx->has_b_frames < h->sps.num_reorder_frames) {
1592         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1593         s->low_delay           = 0;
1594     }
1595
1596     if (s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT &&
1597         !h->sps.bitstream_restriction_flag) {
1598         s->avctx->has_b_frames = MAX_DELAYED_PIC_COUNT - 1;
1599         s->low_delay           = 0;
1600     }
1601
1602     for (i = 0; 1; i++) {
1603         if(i == MAX_DELAYED_PIC_COUNT || cur->poc < h->last_pocs[i]){
1604             if(i)
1605                 h->last_pocs[i-1] = cur->poc;
1606             break;
1607         } else if(i) {
1608             h->last_pocs[i-1]= h->last_pocs[i];
1609         }
1610     }
1611     out_of_order = MAX_DELAYED_PIC_COUNT - i;
1612     if(   cur->f.pict_type == AV_PICTURE_TYPE_B
1613        || (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))
1614         out_of_order = FFMAX(out_of_order, 1);
1615     if(s->avctx->has_b_frames < out_of_order && !h->sps.bitstream_restriction_flag){
1616         av_log(s->avctx, AV_LOG_WARNING, "Increasing reorder buffer to %d\n", out_of_order);
1617         s->avctx->has_b_frames = out_of_order;
1618         s->low_delay = 0;
1619     }
1620
1621     pics = 0;
1622     while (h->delayed_pic[pics])
1623         pics++;
1624
1625     av_assert0(pics <= MAX_DELAYED_PIC_COUNT);
1626
1627     h->delayed_pic[pics++] = cur;
1628     if (cur->f.reference == 0)
1629         cur->f.reference = DELAYED_PIC_REF;
1630
1631     out = h->delayed_pic[0];
1632     out_idx = 0;
1633     for (i = 1; h->delayed_pic[i] &&
1634                 !h->delayed_pic[i]->f.key_frame &&
1635                 !h->delayed_pic[i]->mmco_reset;
1636          i++)
1637         if (h->delayed_pic[i]->poc < out->poc) {
1638             out     = h->delayed_pic[i];
1639             out_idx = i;
1640         }
1641     if (s->avctx->has_b_frames == 0 &&
1642         (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset))
1643         h->next_outputed_poc = INT_MIN;
1644     out_of_order = out->poc < h->next_outputed_poc;
1645
1646     if (out_of_order || pics > s->avctx->has_b_frames) {
1647         out->f.reference &= ~DELAYED_PIC_REF;
1648         // for frame threading, the owner must be the second field's thread or
1649         // else the first thread can release the picture and reuse it unsafely
1650         out->owner2       = s;
1651         for (i = out_idx; h->delayed_pic[i]; i++)
1652             h->delayed_pic[i] = h->delayed_pic[i + 1];
1653     }
1654     if (!out_of_order && pics > s->avctx->has_b_frames) {
1655         h->next_output_pic = out;
1656         if (out_idx == 0 && h->delayed_pic[0] && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset)) {
1657             h->next_outputed_poc = INT_MIN;
1658         } else
1659             h->next_outputed_poc = out->poc;
1660     } else {
1661         av_log(s->avctx, AV_LOG_DEBUG, "no picture %s\n", out_of_order ? "ooo" : "");
1662     }
1663
1664     if (h->next_output_pic && h->next_output_pic->sync) {
1665         h->sync |= 2;
1666     }
1667
1668     if (setup_finished)
1669         ff_thread_finish_setup(s->avctx);
1670 }
1671
1672 static av_always_inline void backup_mb_border(H264Context *h, uint8_t *src_y,
1673                                               uint8_t *src_cb, uint8_t *src_cr,
1674                                               int linesize, int uvlinesize,
1675                                               int simple)
1676 {
1677     MpegEncContext *const s = &h->s;
1678     uint8_t *top_border;
1679     int top_idx = 1;
1680     const int pixel_shift = h->pixel_shift;
1681     int chroma444 = CHROMA444;
1682     int chroma422 = CHROMA422;
1683
1684     src_y  -= linesize;
1685     src_cb -= uvlinesize;
1686     src_cr -= uvlinesize;
1687
1688     if (!simple && FRAME_MBAFF) {
1689         if (s->mb_y & 1) {
1690             if (!MB_MBAFF) {
1691                 top_border = h->top_borders[0][s->mb_x];
1692                 AV_COPY128(top_border, src_y + 15 * linesize);
1693                 if (pixel_shift)
1694                     AV_COPY128(top_border + 16, src_y + 15 * linesize + 16);
1695                 if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1696                     if (chroma444) {
1697                         if (pixel_shift) {
1698                             AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
1699                             AV_COPY128(top_border + 48, src_cb + 15 * uvlinesize + 16);
1700                             AV_COPY128(top_border + 64, src_cr + 15 * uvlinesize);
1701                             AV_COPY128(top_border + 80, src_cr + 15 * uvlinesize + 16);
1702                         } else {
1703                             AV_COPY128(top_border + 16, src_cb + 15 * uvlinesize);
1704                             AV_COPY128(top_border + 32, src_cr + 15 * uvlinesize);
1705                         }
1706                     } else if (chroma422) {
1707                         if (pixel_shift) {
1708                             AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
1709                             AV_COPY128(top_border + 48, src_cr + 15 * uvlinesize);
1710                         } else {
1711                             AV_COPY64(top_border + 16, src_cb + 15 * uvlinesize);
1712                             AV_COPY64(top_border + 24, src_cr + 15 * uvlinesize);
1713                         }
1714                     } else {
1715                         if (pixel_shift) {
1716                             AV_COPY128(top_border + 32, src_cb + 7 * uvlinesize);
1717                             AV_COPY128(top_border + 48, src_cr + 7 * uvlinesize);
1718                         } else {
1719                             AV_COPY64(top_border + 16, src_cb + 7 * uvlinesize);
1720                             AV_COPY64(top_border + 24, src_cr + 7 * uvlinesize);
1721                         }
1722                     }
1723                 }
1724             }
1725         } else if (MB_MBAFF) {
1726             top_idx = 0;
1727         } else
1728             return;
1729     }
1730
1731     top_border = h->top_borders[top_idx][s->mb_x];
1732     /* There are two lines saved, the line above the the top macroblock
1733      * of a pair, and the line above the bottom macroblock. */
1734     AV_COPY128(top_border, src_y + 16 * linesize);
1735     if (pixel_shift)
1736         AV_COPY128(top_border + 16, src_y + 16 * linesize + 16);
1737
1738     if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1739         if (chroma444) {
1740             if (pixel_shift) {
1741                 AV_COPY128(top_border + 32, src_cb + 16 * linesize);
1742                 AV_COPY128(top_border + 48, src_cb + 16 * linesize + 16);
1743                 AV_COPY128(top_border + 64, src_cr + 16 * linesize);
1744                 AV_COPY128(top_border + 80, src_cr + 16 * linesize + 16);
1745             } else {
1746                 AV_COPY128(top_border + 16, src_cb + 16 * linesize);
1747                 AV_COPY128(top_border + 32, src_cr + 16 * linesize);
1748             }
1749         } else if (chroma422) {
1750             if (pixel_shift) {
1751                 AV_COPY128(top_border + 32, src_cb + 16 * uvlinesize);
1752                 AV_COPY128(top_border + 48, src_cr + 16 * uvlinesize);
1753             } else {
1754                 AV_COPY64(top_border + 16, src_cb + 16 * uvlinesize);
1755                 AV_COPY64(top_border + 24, src_cr + 16 * uvlinesize);
1756             }
1757         } else {
1758             if (pixel_shift) {
1759                 AV_COPY128(top_border + 32, src_cb + 8 * uvlinesize);
1760                 AV_COPY128(top_border + 48, src_cr + 8 * uvlinesize);
1761             } else {
1762                 AV_COPY64(top_border + 16, src_cb + 8 * uvlinesize);
1763                 AV_COPY64(top_border + 24, src_cr + 8 * uvlinesize);
1764             }
1765         }
1766     }
1767 }
1768
1769 static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1770                                             uint8_t *src_cb, uint8_t *src_cr,
1771                                             int linesize, int uvlinesize,
1772                                             int xchg, int chroma444,
1773                                             int simple, int pixel_shift)
1774 {
1775     MpegEncContext *const s = &h->s;
1776     int deblock_topleft;
1777     int deblock_top;
1778     int top_idx = 1;
1779     uint8_t *top_border_m1;
1780     uint8_t *top_border;
1781
1782     if (!simple && FRAME_MBAFF) {
1783         if (s->mb_y & 1) {
1784             if (!MB_MBAFF)
1785                 return;
1786         } else {
1787             top_idx = MB_MBAFF ? 0 : 1;
1788         }
1789     }
1790
1791     if (h->deblocking_filter == 2) {
1792         deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1793         deblock_top     = h->top_type;
1794     } else {
1795         deblock_topleft = (s->mb_x > 0);
1796         deblock_top     = (s->mb_y > !!MB_FIELD);
1797     }
1798
1799     src_y  -= linesize   + 1 + pixel_shift;
1800     src_cb -= uvlinesize + 1 + pixel_shift;
1801     src_cr -= uvlinesize + 1 + pixel_shift;
1802
1803     top_border_m1 = h->top_borders[top_idx][s->mb_x - 1];
1804     top_border    = h->top_borders[top_idx][s->mb_x];
1805
1806 #define XCHG(a, b, xchg)                        \
1807     if (pixel_shift) {                          \
1808         if (xchg) {                             \
1809             AV_SWAP64(b + 0, a + 0);            \
1810             AV_SWAP64(b + 8, a + 8);            \
1811         } else {                                \
1812             AV_COPY128(b, a);                   \
1813         }                                       \
1814     } else if (xchg)                            \
1815         AV_SWAP64(b, a);                        \
1816     else                                        \
1817         AV_COPY64(b, a);
1818
1819     if (deblock_top) {
1820         if (deblock_topleft) {
1821             XCHG(top_border_m1 + (8 << pixel_shift),
1822                  src_y - (7 << pixel_shift), 1);
1823         }
1824         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1825         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1826         if (s->mb_x + 1 < s->mb_width) {
1827             XCHG(h->top_borders[top_idx][s->mb_x + 1],
1828                  src_y + (17 << pixel_shift), 1);
1829         }
1830     }
1831     if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1832         if (chroma444) {
1833             if (deblock_topleft) {
1834                 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1835                 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1836             }
1837             XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1838             XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1839             XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1840             XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1841             if (s->mb_x + 1 < s->mb_width) {
1842                 XCHG(h->top_borders[top_idx][s->mb_x + 1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1843                 XCHG(h->top_borders[top_idx][s->mb_x + 1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1844             }
1845         } else {
1846             if (deblock_top) {
1847                 if (deblock_topleft) {
1848                     XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1849                     XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1850                 }
1851                 XCHG(top_border + (16 << pixel_shift), src_cb + 1 + pixel_shift, 1);
1852                 XCHG(top_border + (24 << pixel_shift), src_cr + 1 + pixel_shift, 1);
1853             }
1854         }
1855     }
1856 }
1857
1858 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth,
1859                                         int index)
1860 {
1861     if (high_bit_depth) {
1862         return AV_RN32A(((int32_t *)mb) + index);
1863     } else
1864         return AV_RN16A(mb + index);
1865 }
1866
1867 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth,
1868                                          int index, int value)
1869 {
1870     if (high_bit_depth) {
1871         AV_WN32A(((int32_t *)mb) + index, value);
1872     } else
1873         AV_WN16A(mb + index, value);
1874 }
1875
1876 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h,
1877                                                        int mb_type, int is_h264,
1878                                                        int simple,
1879                                                        int transform_bypass,
1880                                                        int pixel_shift,
1881                                                        int *block_offset,
1882                                                        int linesize,
1883                                                        uint8_t *dest_y, int p)
1884 {
1885     MpegEncContext *const s = &h->s;
1886     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1887     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1888     int i;
1889     int qscale = p == 0 ? s->qscale : h->chroma_qp[p - 1];
1890     block_offset += 16 * p;
1891     if (IS_INTRA4x4(mb_type)) {
1892         if (simple || !s->encoding) {
1893             if (IS_8x8DCT(mb_type)) {
1894                 if (transform_bypass) {
1895                     idct_dc_add  =
1896                     idct_add     = s->dsp.add_pixels8;
1897                 } else {
1898                     idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1899                     idct_add    = h->h264dsp.h264_idct8_add;
1900                 }
1901                 for (i = 0; i < 16; i += 4) {
1902                     uint8_t *const ptr = dest_y + block_offset[i];
1903                     const int dir      = h->intra4x4_pred_mode_cache[scan8[i]];
1904                     if (transform_bypass && h->sps.profile_idc == 244 && dir <= 1) {
1905                         h->hpc.pred8x8l_add[dir](ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1906                     } else {
1907                         const int nnz = h->non_zero_count_cache[scan8[i + p * 16]];
1908                         h->hpc.pred8x8l[dir](ptr, (h->topleft_samples_available << i) & 0x8000,
1909                                              (h->topright_samples_available << i) & 0x4000, linesize);
1910                         if (nnz) {
1911                             if (nnz == 1 && dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
1912                                 idct_dc_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1913                             else
1914                                 idct_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1915                         }
1916                     }
1917                 }
1918             } else {
1919                 if (transform_bypass) {
1920                     idct_dc_add  =
1921                         idct_add = s->dsp.add_pixels4;
1922                 } else {
1923                     idct_dc_add = h->h264dsp.h264_idct_dc_add;
1924                     idct_add    = h->h264dsp.h264_idct_add;
1925                 }
1926                 for (i = 0; i < 16; i++) {
1927                     uint8_t *const ptr = dest_y + block_offset[i];
1928                     const int dir      = h->intra4x4_pred_mode_cache[scan8[i]];
1929
1930                     if (transform_bypass && h->sps.profile_idc == 244 && dir <= 1) {
1931                         h->hpc.pred4x4_add[dir](ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1932                     } else {
1933                         uint8_t *topright;
1934                         int nnz, tr;
1935                         uint64_t tr_high;
1936                         if (dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED) {
1937                             const int topright_avail = (h->topright_samples_available << i) & 0x8000;
1938                             assert(s->mb_y || linesize <= block_offset[i]);
1939                             if (!topright_avail) {
1940                                 if (pixel_shift) {
1941                                     tr_high  = ((uint16_t *)ptr)[3 - linesize / 2] * 0x0001000100010001ULL;
1942                                     topright = (uint8_t *)&tr_high;
1943                                 } else {
1944                                     tr       = ptr[3 - linesize] * 0x01010101u;
1945                                     topright = (uint8_t *)&tr;
1946                                 }
1947                             } else
1948                                 topright = ptr + (4 << pixel_shift) - linesize;
1949                         } else
1950                             topright = NULL;
1951
1952                         h->hpc.pred4x4[dir](ptr, topright, linesize);
1953                         nnz = h->non_zero_count_cache[scan8[i + p * 16]];
1954                         if (nnz) {
1955                             if (is_h264) {
1956                                 if (nnz == 1 && dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
1957                                     idct_dc_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1958                                 else
1959                                     idct_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1960                             } else if (CONFIG_SVQ3_DECODER)
1961                                 ff_svq3_add_idct_c(ptr, h->mb + i * 16 + p * 256, linesize, qscale, 0);
1962                         }
1963                     }
1964                 }
1965             }
1966         }
1967     } else {
1968         h->hpc.pred16x16[h->intra16x16_pred_mode](dest_y, linesize);
1969         if (is_h264) {
1970             if (h->non_zero_count_cache[scan8[LUMA_DC_BLOCK_INDEX + p]]) {
1971                 if (!transform_bypass)
1972                     h->h264dsp.h264_luma_dc_dequant_idct(h->mb + (p * 256 << pixel_shift),
1973                                                          h->mb_luma_dc[p],
1974                                                          h->dequant4_coeff[p][qscale][0]);
1975                 else {
1976                     static const uint8_t dc_mapping[16] = {
1977                          0 * 16,  1 * 16,  4 * 16,  5 * 16,
1978                          2 * 16,  3 * 16,  6 * 16,  7 * 16,
1979                          8 * 16,  9 * 16, 12 * 16, 13 * 16,
1980                         10 * 16, 11 * 16, 14 * 16, 15 * 16 };
1981                     for (i = 0; i < 16; i++)
1982                         dctcoef_set(h->mb + (p * 256 << pixel_shift),
1983                                     pixel_shift, dc_mapping[i],
1984                                     dctcoef_get(h->mb_luma_dc[p],
1985                                                 pixel_shift, i));
1986                 }
1987             }
1988         } else if (CONFIG_SVQ3_DECODER)
1989             ff_svq3_luma_dc_dequant_idct_c(h->mb + p * 256,
1990                                            h->mb_luma_dc[p], qscale);
1991     }
1992 }
1993
1994 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type,
1995                                                     int is_h264, int simple,
1996                                                     int transform_bypass,
1997                                                     int pixel_shift,
1998                                                     int *block_offset,
1999                                                     int linesize,
2000                                                     uint8_t *dest_y, int p)
2001 {
2002     MpegEncContext *const s = &h->s;
2003     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2004     int i;
2005     block_offset += 16 * p;
2006     if (!IS_INTRA4x4(mb_type)) {
2007         if (is_h264) {
2008             if (IS_INTRA16x16(mb_type)) {
2009                 if (transform_bypass) {
2010                     if (h->sps.profile_idc == 244 &&
2011                         (h->intra16x16_pred_mode == VERT_PRED8x8 ||
2012                          h->intra16x16_pred_mode == HOR_PRED8x8)) {
2013                         h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset,
2014                                                                       h->mb + (p * 256 << pixel_shift),
2015                                                                       linesize);
2016                     } else {
2017                         for (i = 0; i < 16; i++)
2018                             if (h->non_zero_count_cache[scan8[i + p * 16]] ||
2019                                 dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
2020                                 s->dsp.add_pixels4(dest_y + block_offset[i],
2021                                                    h->mb + (i * 16 + p * 256 << pixel_shift),
2022                                                    linesize);
2023                     }
2024                 } else {
2025                     h->h264dsp.h264_idct_add16intra(dest_y, block_offset,
2026                                                     h->mb + (p * 256 << pixel_shift),
2027                                                     linesize,
2028                                                     h->non_zero_count_cache + p * 5 * 8);
2029                 }
2030             } else if (h->cbp & 15) {
2031                 if (transform_bypass) {
2032                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
2033                     idct_add = IS_8x8DCT(mb_type) ? s->dsp.add_pixels8
2034                                                   : s->dsp.add_pixels4;
2035                     for (i = 0; i < 16; i += di)
2036                         if (h->non_zero_count_cache[scan8[i + p * 16]])
2037                             idct_add(dest_y + block_offset[i],
2038                                      h->mb + (i * 16 + p * 256 << pixel_shift),
2039                                      linesize);
2040                 } else {
2041                     if (IS_8x8DCT(mb_type))
2042                         h->h264dsp.h264_idct8_add4(dest_y, block_offset,
2043                                                    h->mb + (p * 256 << pixel_shift),
2044                                                    linesize,
2045                                                    h->non_zero_count_cache + p * 5 * 8);
2046                     else
2047                         h->h264dsp.h264_idct_add16(dest_y, block_offset,
2048                                                    h->mb + (p * 256 << pixel_shift),
2049                                                    linesize,
2050                                                    h->non_zero_count_cache + p * 5 * 8);
2051                 }
2052             }
2053         } else if (CONFIG_SVQ3_DECODER) {
2054             for (i = 0; i < 16; i++)
2055                 if (h->non_zero_count_cache[scan8[i + p * 16]] || h->mb[i * 16 + p * 256]) {
2056                     // FIXME benchmark weird rule, & below
2057                     uint8_t *const ptr = dest_y + block_offset[i];
2058                     ff_svq3_add_idct_c(ptr, h->mb + i * 16 + p * 256, linesize,
2059                                        s->qscale, IS_INTRA(mb_type) ? 1 : 0);
2060                 }
2061         }
2062     }
2063 }
2064
2065 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple,
2066                                                    int pixel_shift)
2067 {
2068     MpegEncContext *const s = &h->s;
2069     const int mb_x    = s->mb_x;
2070     const int mb_y    = s->mb_y;
2071     const int mb_xy   = h->mb_xy;
2072     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2073     uint8_t *dest_y, *dest_cb, *dest_cr;
2074     int linesize, uvlinesize /*dct_offset*/;
2075     int i, j;
2076     int *block_offset = &h->block_offset[0];
2077     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
2078     /* is_h264 should always be true if SVQ3 is disabled. */
2079     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
2080     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2081     const int block_h   = 16 >> s->chroma_y_shift;
2082     const int chroma422 = CHROMA422;
2083
2084     dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift)     + mb_y * s->linesize)  * 16;
2085     dest_cb = s->current_picture.f.data[1] +  (mb_x << pixel_shift) * 8 + mb_y * s->uvlinesize * block_h;
2086     dest_cr = s->current_picture.f.data[2] +  (mb_x << pixel_shift) * 8 + mb_y * s->uvlinesize * block_h;
2087
2088     s->dsp.prefetch(dest_y  + (s->mb_x & 3) * 4 * s->linesize   + (64 << pixel_shift), s->linesize,       4);
2089     s->dsp.prefetch(dest_cb + (s->mb_x & 7)     * s->uvlinesize + (64 << pixel_shift), dest_cr - dest_cb, 2);
2090
2091     h->list_counts[mb_xy] = h->list_count;
2092
2093     if (!simple && MB_FIELD) {
2094         linesize     = h->mb_linesize = s->linesize * 2;
2095         uvlinesize   = h->mb_uvlinesize = s->uvlinesize * 2;
2096         block_offset = &h->block_offset[48];
2097         if (mb_y & 1) { // FIXME move out of this function?
2098             dest_y  -= s->linesize * 15;
2099             dest_cb -= s->uvlinesize * (block_h - 1);
2100             dest_cr -= s->uvlinesize * (block_h - 1);
2101         }
2102         if (FRAME_MBAFF) {
2103             int list;
2104             for (list = 0; list < h->list_count; list++) {
2105                 if (!USES_LIST(mb_type, list))
2106                     continue;
2107                 if (IS_16X16(mb_type)) {
2108                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2109                     fill_rectangle(ref, 4, 4, 8, (16 + *ref) ^ (s->mb_y & 1), 1);
2110                 } else {
2111                     for (i = 0; i < 16; i += 4) {
2112                         int ref = h->ref_cache[list][scan8[i]];
2113                         if (ref >= 0)
2114                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2,
2115                                            8, (16 + ref) ^ (s->mb_y & 1), 1);
2116                     }
2117                 }
2118             }
2119         }
2120     } else {
2121         linesize   = h->mb_linesize   = s->linesize;
2122         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
2123         // dct_offset = s->linesize * 16;
2124     }
2125
2126     if (!simple && IS_INTRA_PCM(mb_type)) {
2127         const int bit_depth = h->sps.bit_depth_luma;
2128         if (pixel_shift) {
2129             int j;
2130             GetBitContext gb;
2131             init_get_bits(&gb, (uint8_t *)h->mb,
2132                           ff_h264_mb_sizes[h->sps.chroma_format_idc] * bit_depth);
2133
2134             for (i = 0; i < 16; i++) {
2135                 uint16_t *tmp_y = (uint16_t *)(dest_y + i * linesize);
2136                 for (j = 0; j < 16; j++)
2137                     tmp_y[j] = get_bits(&gb, bit_depth);
2138             }
2139             if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
2140                 if (!h->sps.chroma_format_idc) {
2141                     for (i = 0; i < block_h; i++) {
2142                         uint16_t *tmp_cb = (uint16_t *)(dest_cb + i * uvlinesize);
2143                         uint16_t *tmp_cr = (uint16_t *)(dest_cr + i * uvlinesize);
2144                         for (j = 0; j < 8; j++) {
2145                             tmp_cb[j] = tmp_cr[j] = 1 << (bit_depth - 1);
2146                         }
2147                     }
2148                 } else {
2149                     for (i = 0; i < block_h; i++) {
2150                         uint16_t *tmp_cb = (uint16_t *)(dest_cb + i * uvlinesize);
2151                         for (j = 0; j < 8; j++)
2152                             tmp_cb[j] = get_bits(&gb, bit_depth);
2153                     }
2154                     for (i = 0; i < block_h; i++) {
2155                         uint16_t *tmp_cr = (uint16_t *)(dest_cr + i * uvlinesize);
2156                         for (j = 0; j < 8; j++)
2157                             tmp_cr[j] = get_bits(&gb, bit_depth);
2158                     }
2159                 }
2160             }
2161         } else {
2162             for (i = 0; i < 16; i++)
2163                 memcpy(dest_y + i * linesize, (uint8_t *)h->mb + i * 16, 16);
2164             if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
2165                 if (!h->sps.chroma_format_idc) {
2166                     for (i = 0; i < 8; i++) {
2167                         memset(dest_cb + i*uvlinesize, 1 << (bit_depth - 1), 8);
2168                         memset(dest_cr + i*uvlinesize, 1 << (bit_depth - 1), 8);
2169                     }
2170                 } else {
2171                     uint8_t *src_cb = (uint8_t *)h->mb + 256;
2172                     uint8_t *src_cr = (uint8_t *)h->mb + 256 + block_h * 8;
2173                     for (i = 0; i < block_h; i++) {
2174                         memcpy(dest_cb + i * uvlinesize, src_cb + i * 8, 8);
2175                         memcpy(dest_cr + i * uvlinesize, src_cr + i * 8, 8);
2176                     }
2177                 }
2178             }
2179         }
2180     } else {
2181         if (IS_INTRA(mb_type)) {
2182             if (h->deblocking_filter)
2183                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize,
2184                                uvlinesize, 1, 0, simple, pixel_shift);
2185
2186             if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
2187                 if (CHROMA) {
2188                     h->hpc.pred8x8[h->chroma_pred_mode](dest_cb, uvlinesize);
2189                     h->hpc.pred8x8[h->chroma_pred_mode](dest_cr, uvlinesize);
2190                 }
2191             }
2192
2193             hl_decode_mb_predict_luma(h, mb_type, is_h264, simple,
2194                                       transform_bypass, pixel_shift,
2195                                       block_offset, linesize, dest_y, 0);
2196
2197             if (h->deblocking_filter)
2198                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize,
2199                                uvlinesize, 0, 0, simple, pixel_shift);
2200         } else if (is_h264) {
2201             if (chroma422) {
2202                 hl_motion_422(h, dest_y, dest_cb, dest_cr,
2203                               s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2204                               s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2205                               h->h264dsp.weight_h264_pixels_tab,
2206                               h->h264dsp.biweight_h264_pixels_tab,
2207                               pixel_shift);
2208             } else {
2209                 hl_motion_420(h, dest_y, dest_cb, dest_cr,
2210                               s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2211                               s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2212                               h->h264dsp.weight_h264_pixels_tab,
2213                               h->h264dsp.biweight_h264_pixels_tab,
2214                               pixel_shift);
2215             }
2216         }
2217
2218         hl_decode_mb_idct_luma(h, mb_type, is_h264, simple, transform_bypass,
2219                                pixel_shift, block_offset, linesize, dest_y, 0);
2220
2221         if ((simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) &&
2222             (h->cbp & 0x30)) {
2223             uint8_t *dest[2] = { dest_cb, dest_cr };
2224             if (transform_bypass) {
2225                 if (IS_INTRA(mb_type) && h->sps.profile_idc == 244 &&
2226                     (h->chroma_pred_mode == VERT_PRED8x8 ||
2227                      h->chroma_pred_mode == HOR_PRED8x8)) {
2228                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0],
2229                                                             block_offset + 16,
2230                                                             h->mb + (16 * 16 * 1 << pixel_shift),
2231                                                             uvlinesize);
2232                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1],
2233                                                             block_offset + 32,
2234                                                             h->mb + (16 * 16 * 2 << pixel_shift),
2235                                                             uvlinesize);
2236                 } else {
2237                     idct_add = s->dsp.add_pixels4;
2238                     for (j = 1; j < 3; j++) {
2239                         for (i = j * 16; i < j * 16 + 4; i++)
2240                             if (h->non_zero_count_cache[scan8[i]] ||
2241                                 dctcoef_get(h->mb, pixel_shift, i * 16))
2242                                 idct_add(dest[j - 1] + block_offset[i],
2243                                          h->mb + (i * 16 << pixel_shift),
2244                                          uvlinesize);
2245                         if (chroma422) {
2246                             for (i = j * 16 + 4; i < j * 16 + 8; i++)
2247                                 if (h->non_zero_count_cache[scan8[i + 4]] ||
2248                                     dctcoef_get(h->mb, pixel_shift, i * 16))
2249                                     idct_add(dest[j - 1] + block_offset[i + 4],
2250                                              h->mb + (i * 16 << pixel_shift),
2251                                              uvlinesize);
2252                         }
2253                     }
2254                 }
2255             } else {
2256                 if (is_h264) {
2257                     int qp[2];
2258                     if (chroma422) {
2259                         qp[0] = h->chroma_qp[0] + 3;
2260                         qp[1] = h->chroma_qp[1] + 3;
2261                     } else {
2262                         qp[0] = h->chroma_qp[0];
2263                         qp[1] = h->chroma_qp[1];
2264                     }
2265                     if (h->non_zero_count_cache[scan8[CHROMA_DC_BLOCK_INDEX + 0]])
2266                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16 * 16 * 1 << pixel_shift),
2267                                                                h->dequant4_coeff[IS_INTRA(mb_type) ? 1 : 4][qp[0]][0]);
2268                     if (h->non_zero_count_cache[scan8[CHROMA_DC_BLOCK_INDEX + 1]])
2269                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16 * 16 * 2 << pixel_shift),
2270                                                                h->dequant4_coeff[IS_INTRA(mb_type) ? 2 : 5][qp[1]][0]);
2271                     h->h264dsp.h264_idct_add8(dest, block_offset,
2272                                               h->mb, uvlinesize,
2273                                               h->non_zero_count_cache);
2274                 } else if (CONFIG_SVQ3_DECODER) {
2275                     h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + 16 * 16 * 1,
2276                                                            h->dequant4_coeff[IS_INTRA(mb_type) ? 1 : 4][h->chroma_qp[0]][0]);
2277                     h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + 16 * 16 * 2,
2278                                                            h->dequant4_coeff[IS_INTRA(mb_type) ? 2 : 5][h->chroma_qp[1]][0]);
2279                     for (j = 1; j < 3; j++) {
2280                         for (i = j * 16; i < j * 16 + 4; i++)
2281                             if (h->non_zero_count_cache[scan8[i]] || h->mb[i * 16]) {
2282                                 uint8_t *const ptr = dest[j - 1] + block_offset[i];
2283                                 ff_svq3_add_idct_c(ptr, h->mb + i * 16,
2284                                                    uvlinesize,
2285                                                    ff_h264_chroma_qp[0][s->qscale + 12] - 12, 2);
2286                             }
2287                     }
2288                 }
2289             }
2290         }
2291     }
2292     if (h->cbp || IS_INTRA(mb_type)) {
2293         s->dsp.clear_blocks(h->mb);
2294         s->dsp.clear_blocks(h->mb + (24 * 16 << pixel_shift));
2295     }
2296 }
2297
2298 static av_always_inline void hl_decode_mb_444_internal(H264Context *h,
2299                                                        int simple,
2300                                                        int pixel_shift)
2301 {
2302     MpegEncContext *const s = &h->s;
2303     const int mb_x    = s->mb_x;
2304     const int mb_y    = s->mb_y;
2305     const int mb_xy   = h->mb_xy;
2306     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2307     uint8_t *dest[3];
2308     int linesize;
2309     int i, j, p;
2310     int *block_offset = &h->block_offset[0];
2311     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
2312     const int plane_count      = (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) ? 3 : 1;
2313
2314     for (p = 0; p < plane_count; p++) {
2315         dest[p] = s->current_picture.f.data[p] +
2316                   ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
2317         s->dsp.prefetch(dest[p] + (s->mb_x & 3) * 4 * s->linesize + (64 << pixel_shift),
2318                         s->linesize, 4);
2319     }
2320
2321     h->list_counts[mb_xy] = h->list_count;
2322
2323     if (!simple && MB_FIELD) {
2324         linesize     = h->mb_linesize = h->mb_uvlinesize = s->linesize * 2;
2325         block_offset = &h->block_offset[48];
2326         if (mb_y & 1) // FIXME move out of this function?
2327             for (p = 0; p < 3; p++)
2328                 dest[p] -= s->linesize * 15;
2329         if (FRAME_MBAFF) {
2330             int list;
2331             for (list = 0; list < h->list_count; list++) {
2332                 if (!USES_LIST(mb_type, list))
2333                     continue;
2334                 if (IS_16X16(mb_type)) {
2335                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2336                     fill_rectangle(ref, 4, 4, 8, (16 + *ref) ^ (s->mb_y & 1), 1);
2337                 } else {
2338                     for (i = 0; i < 16; i += 4) {
2339                         int ref = h->ref_cache[list][scan8[i]];
2340                         if (ref >= 0)
2341                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2,
2342                                            8, (16 + ref) ^ (s->mb_y & 1), 1);
2343                     }
2344                 }
2345             }
2346         }
2347     } else {
2348         linesize = h->mb_linesize = h->mb_uvlinesize = s->linesize;
2349     }
2350
2351     if (!simple && IS_INTRA_PCM(mb_type)) {
2352         if (pixel_shift) {
2353             const int bit_depth = h->sps.bit_depth_luma;
2354             GetBitContext gb;
2355             init_get_bits(&gb, (uint8_t *)h->mb, 768 * bit_depth);
2356
2357             for (p = 0; p < plane_count; p++)
2358                 for (i = 0; i < 16; i++) {
2359                     uint16_t *tmp = (uint16_t *)(dest[p] + i * linesize);
2360                     for (j = 0; j < 16; j++)
2361                         tmp[j] = get_bits(&gb, bit_depth);
2362                 }
2363         } else {
2364             for (p = 0; p < plane_count; p++)
2365                 for (i = 0; i < 16; i++)
2366                     memcpy(dest[p] + i * linesize,
2367                            (uint8_t *)h->mb + p * 256 + i * 16, 16);
2368         }
2369     } else {
2370         if (IS_INTRA(mb_type)) {
2371             if (h->deblocking_filter)
2372                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize,
2373                                linesize, 1, 1, simple, pixel_shift);
2374
2375             for (p = 0; p < plane_count; p++)
2376                 hl_decode_mb_predict_luma(h, mb_type, 1, simple,
2377                                           transform_bypass, pixel_shift,
2378                                           block_offset, linesize, dest[p], p);
2379
2380             if (h->deblocking_filter)
2381                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize,
2382                                linesize, 0, 1, simple, pixel_shift);
2383         } else {
2384             hl_motion(h, dest[0], dest[1], dest[2],
2385                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2386                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2387                       h->h264dsp.weight_h264_pixels_tab,
2388                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 3);
2389         }
2390
2391         for (p = 0; p < plane_count; p++)
2392             hl_decode_mb_idct_luma(h, mb_type, 1, simple, transform_bypass,
2393                                    pixel_shift, block_offset, linesize,
2394                                    dest[p], p);
2395     }
2396     if (h->cbp || IS_INTRA(mb_type)) {
2397         s->dsp.clear_blocks(h->mb);
2398         s->dsp.clear_blocks(h->mb + (24 * 16 << pixel_shift));
2399     }
2400 }
2401
2402 /**
2403  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2404  */
2405 #define hl_decode_mb_simple(sh, bits)                          \
2406 static void hl_decode_mb_simple_ ## bits(H264Context *h)       \
2407 {                                                              \
2408     hl_decode_mb_internal(h, 1, sh);                           \
2409 }
2410
2411 hl_decode_mb_simple(0, 8)
2412 hl_decode_mb_simple(1, 16)
2413
2414 /**
2415  * Process a macroblock; this handles edge cases, such as interlacing.
2416  */
2417 static av_noinline void hl_decode_mb_complex(H264Context *h)
2418 {
2419     hl_decode_mb_internal(h, 0, h->pixel_shift);
2420 }
2421
2422 static av_noinline void hl_decode_mb_444_complex(H264Context *h)
2423 {
2424     hl_decode_mb_444_internal(h, 0, h->pixel_shift);
2425 }
2426
2427 static av_noinline void hl_decode_mb_444_simple(H264Context *h)
2428 {
2429     hl_decode_mb_444_internal(h, 1, 0);
2430 }
2431
2432 void ff_h264_hl_decode_mb(H264Context *h)
2433 {
2434     MpegEncContext *const s = &h->s;
2435     const int mb_xy   = h->mb_xy;
2436     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2437     int is_complex    = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2438
2439     if (CHROMA444) {
2440         if (is_complex || h->pixel_shift)
2441             hl_decode_mb_444_complex(h);
2442         else
2443             hl_decode_mb_444_simple(h);
2444     } else if (is_complex) {
2445         hl_decode_mb_complex(h);
2446     } else if (h->pixel_shift) {
2447         hl_decode_mb_simple_16(h);
2448     } else
2449         hl_decode_mb_simple_8(h);
2450 }
2451
2452 static int pred_weight_table(H264Context *h)
2453 {
2454     MpegEncContext *const s = &h->s;
2455     int list, i;
2456     int luma_def, chroma_def;
2457
2458     h->use_weight             = 0;
2459     h->use_weight_chroma      = 0;
2460     h->luma_log2_weight_denom = get_ue_golomb(&s->gb);
2461     if (h->sps.chroma_format_idc)
2462         h->chroma_log2_weight_denom = get_ue_golomb(&s->gb);
2463     luma_def   = 1 << h->luma_log2_weight_denom;
2464     chroma_def = 1 << h->chroma_log2_weight_denom;
2465
2466     for (list = 0; list < 2; list++) {
2467         h->luma_weight_flag[list]   = 0;
2468         h->chroma_weight_flag[list] = 0;
2469         for (i = 0; i < h->ref_count[list]; i++) {
2470             int luma_weight_flag, chroma_weight_flag;
2471
2472             luma_weight_flag = get_bits1(&s->gb);
2473             if (luma_weight_flag) {
2474                 h->luma_weight[i][list][0] = get_se_golomb(&s->gb);
2475                 h->luma_weight[i][list][1] = get_se_golomb(&s->gb);
2476                 if (h->luma_weight[i][list][0] != luma_def ||
2477                     h->luma_weight[i][list][1] != 0) {
2478                     h->use_weight             = 1;
2479                     h->luma_weight_flag[list] = 1;
2480                 }
2481             } else {
2482                 h->luma_weight[i][list][0] = luma_def;
2483                 h->luma_weight[i][list][1] = 0;
2484             }
2485
2486             if (h->sps.chroma_format_idc) {
2487                 chroma_weight_flag = get_bits1(&s->gb);
2488                 if (chroma_weight_flag) {
2489                     int j;
2490                     for (j = 0; j < 2; j++) {
2491                         h->chroma_weight[i][list][j][0] = get_se_golomb(&s->gb);
2492                         h->chroma_weight[i][list][j][1] = get_se_golomb(&s->gb);
2493                         if (h->chroma_weight[i][list][j][0] != chroma_def ||
2494                             h->chroma_weight[i][list][j][1] != 0) {
2495                             h->use_weight_chroma = 1;
2496                             h->chroma_weight_flag[list] = 1;
2497                         }
2498                     }
2499                 } else {
2500                     int j;
2501                     for (j = 0; j < 2; j++) {
2502                         h->chroma_weight[i][list][j][0] = chroma_def;
2503                         h->chroma_weight[i][list][j][1] = 0;
2504                     }
2505                 }
2506             }
2507         }
2508         if (h->slice_type_nos != AV_PICTURE_TYPE_B)
2509             break;
2510     }
2511     h->use_weight = h->use_weight || h->use_weight_chroma;
2512     return 0;
2513 }
2514
2515 /**
2516  * Initialize implicit_weight table.
2517  * @param field  0/1 initialize the weight for interlaced MBAFF
2518  *                -1 initializes the rest
2519  */
2520 static void implicit_weight_table(H264Context *h, int field)
2521 {
2522     MpegEncContext *const s = &h->s;
2523     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2524
2525     for (i = 0; i < 2; i++) {
2526         h->luma_weight_flag[i]   = 0;
2527         h->chroma_weight_flag[i] = 0;
2528     }
2529
2530     if (field < 0) {
2531         if (s->picture_structure == PICT_FRAME) {
2532             cur_poc = s->current_picture_ptr->poc;
2533         } else {
2534             cur_poc = s->current_picture_ptr->field_poc[s->picture_structure - 1];
2535         }
2536         if (h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF &&
2537             h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2 * cur_poc) {
2538             h->use_weight = 0;
2539             h->use_weight_chroma = 0;
2540             return;
2541         }
2542         ref_start  = 0;
2543         ref_count0 = h->ref_count[0];
2544         ref_count1 = h->ref_count[1];
2545     } else {
2546         cur_poc    = s->current_picture_ptr->field_poc[field];
2547         ref_start  = 16;
2548         ref_count0 = 16 + 2 * h->ref_count[0];
2549         ref_count1 = 16 + 2 * h->ref_count[1];
2550     }
2551
2552     h->use_weight               = 2;
2553     h->use_weight_chroma        = 2;
2554     h->luma_log2_weight_denom   = 5;
2555     h->chroma_log2_weight_denom = 5;
2556
2557     for (ref0 = ref_start; ref0 < ref_count0; ref0++) {
2558         int poc0 = h->ref_list[0][ref0].poc;
2559         for (ref1 = ref_start; ref1 < ref_count1; ref1++) {
2560             int w = 32;
2561             if (!h->ref_list[0][ref0].long_ref && !h->ref_list[1][ref1].long_ref) {
2562                 int poc1 = h->ref_list[1][ref1].poc;
2563                 int td   = av_clip(poc1 - poc0, -128, 127);
2564                 if (td) {
2565                     int tb = av_clip(cur_poc - poc0, -128, 127);
2566                     int tx = (16384 + (FFABS(td) >> 1)) / td;
2567                     int dist_scale_factor = (tb * tx + 32) >> 8;
2568                     if (dist_scale_factor >= -64 && dist_scale_factor <= 128)
2569                         w = 64 - dist_scale_factor;
2570                 }
2571             }
2572             if (field < 0) {
2573                 h->implicit_weight[ref0][ref1][0] =
2574                 h->implicit_weight[ref0][ref1][1] = w;
2575             } else {
2576                 h->implicit_weight[ref0][ref1][field] = w;
2577             }
2578         }
2579     }
2580 }
2581
2582 /**
2583  * instantaneous decoder refresh.
2584  */
2585 static void idr(H264Context *h)
2586 {
2587     int i;
2588     ff_h264_remove_all_refs(h);
2589     h->prev_frame_num        = 0;
2590     h->prev_frame_num_offset = 0;
2591     h->prev_poc_msb          = 1<<16;
2592     h->prev_poc_lsb          = 0;
2593     for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
2594         h->last_pocs[i] = INT_MIN;
2595 }
2596
2597 /* forget old pics after a seek */
2598 static void flush_dpb(AVCodecContext *avctx)
2599 {
2600     H264Context *h = avctx->priv_data;
2601     int i;
2602     for (i=0; i<=MAX_DELAYED_PIC_COUNT; i++) {
2603         if (h->delayed_pic[i])
2604             h->delayed_pic[i]->f.reference = 0;
2605         h->delayed_pic[i] = NULL;
2606     }
2607     h->outputed_poc = h->next_outputed_poc = INT_MIN;
2608     h->prev_interlaced_frame = 1;
2609     idr(h);
2610     h->prev_frame_num = -1;
2611     if (h->s.current_picture_ptr)
2612         h->s.current_picture_ptr->f.reference = 0;
2613     h->s.first_field = 0;
2614     ff_h264_reset_sei(h);
2615     ff_mpeg_flush(avctx);
2616     h->recovery_frame= -1;
2617     h->sync= 0;
2618 }
2619
2620 static int init_poc(H264Context *h)
2621 {
2622     MpegEncContext *const s = &h->s;
2623     const int max_frame_num = 1 << h->sps.log2_max_frame_num;
2624     int field_poc[2];
2625     Picture *cur = s->current_picture_ptr;
2626
2627     h->frame_num_offset = h->prev_frame_num_offset;
2628     if (h->frame_num < h->prev_frame_num)
2629         h->frame_num_offset += max_frame_num;
2630
2631     if (h->sps.poc_type == 0) {
2632         const int max_poc_lsb = 1 << h->sps.log2_max_poc_lsb;
2633
2634         if (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb / 2)
2635             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2636         else if (h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb / 2)
2637             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2638         else
2639             h->poc_msb = h->prev_poc_msb;
2640         // printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
2641         field_poc[0] =
2642         field_poc[1] = h->poc_msb + h->poc_lsb;
2643         if (s->picture_structure == PICT_FRAME)
2644             field_poc[1] += h->delta_poc_bottom;
2645     } else if (h->sps.poc_type == 1) {
2646         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2647         int i;
2648
2649         if (h->sps.poc_cycle_length != 0)
2650             abs_frame_num = h->frame_num_offset + h->frame_num;
2651         else
2652             abs_frame_num = 0;
2653
2654         if (h->nal_ref_idc == 0 && abs_frame_num > 0)
2655             abs_frame_num--;
2656
2657         expected_delta_per_poc_cycle = 0;
2658         for (i = 0; i < h->sps.poc_cycle_length; i++)
2659             // FIXME integrate during sps parse
2660             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[i];
2661
2662         if (abs_frame_num > 0) {
2663             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2664             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2665
2666             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2667             for (i = 0; i <= frame_num_in_poc_cycle; i++)
2668                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[i];
2669         } else
2670             expectedpoc = 0;
2671
2672         if (h->nal_ref_idc == 0)
2673             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2674
2675         field_poc[0] = expectedpoc + h->delta_poc[0];
2676         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2677
2678         if (s->picture_structure == PICT_FRAME)
2679             field_poc[1] += h->delta_poc[1];
2680     } else {
2681         int poc = 2 * (h->frame_num_offset + h->frame_num);
2682
2683         if (!h->nal_ref_idc)
2684             poc--;
2685
2686         field_poc[0] = poc;
2687         field_poc[1] = poc;
2688     }
2689
2690     if (s->picture_structure != PICT_BOTTOM_FIELD)
2691         s->current_picture_ptr->field_poc[0] = field_poc[0];
2692     if (s->picture_structure != PICT_TOP_FIELD)
2693         s->current_picture_ptr->field_poc[1] = field_poc[1];
2694     cur->poc = FFMIN(cur->field_poc[0], cur->field_poc[1]);
2695
2696     return 0;
2697 }
2698
2699 /**
2700  * initialize scan tables
2701  */
2702 static void init_scan_tables(H264Context *h)
2703 {
2704     int i;
2705     for (i = 0; i < 16; i++) {
2706 #define T(x) (x >> 2) | ((x << 2) & 0xF)
2707         h->zigzag_scan[i] = T(zigzag_scan[i]);
2708         h->field_scan[i]  = T(field_scan[i]);
2709 #undef T
2710     }
2711     for (i = 0; i < 64; i++) {
2712 #define T(x) (x >> 3) | ((x & 7) << 3)
2713         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
2714         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2715         h->field_scan8x8[i]        = T(field_scan8x8[i]);
2716         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
2717 #undef T
2718     }
2719     if (h->sps.transform_bypass) { // FIXME same ugly
2720         h->zigzag_scan_q0          = zigzag_scan;
2721         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
2722         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2723         h->field_scan_q0           = field_scan;
2724         h->field_scan8x8_q0        = field_scan8x8;
2725         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
2726     } else {
2727         h->zigzag_scan_q0          = h->zigzag_scan;
2728         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
2729         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2730         h->field_scan_q0           = h->field_scan;
2731         h->field_scan8x8_q0        = h->field_scan8x8;
2732         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
2733     }
2734 }
2735
2736 static int field_end(H264Context *h, int in_setup)
2737 {
2738     MpegEncContext *const s     = &h->s;
2739     AVCodecContext *const avctx = s->avctx;
2740     int err = 0;
2741     s->mb_y = 0;
2742
2743     if (!in_setup && !s->dropable)
2744         ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX,
2745                                   s->picture_structure == PICT_BOTTOM_FIELD);
2746
2747     if (CONFIG_H264_VDPAU_DECODER &&
2748         s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
2749         ff_vdpau_h264_set_reference_frames(s);
2750
2751     if (in_setup || !(avctx->active_thread_type & FF_THREAD_FRAME)) {
2752         if (!s->dropable) {
2753             err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2754             h->prev_poc_msb = h->poc_msb;
2755             h->prev_poc_lsb = h->poc_lsb;
2756         }
2757         h->prev_frame_num_offset = h->frame_num_offset;
2758         h->prev_frame_num        = h->frame_num;
2759         h->outputed_poc          = h->next_outputed_poc;
2760     }
2761
2762     if (avctx->hwaccel) {
2763         if (avctx->hwaccel->end_frame(avctx) < 0)
2764             av_log(avctx, AV_LOG_ERROR,
2765                    "hardware accelerator failed to decode picture\n");
2766     }
2767
2768     if (CONFIG_H264_VDPAU_DECODER &&
2769         s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
2770         ff_vdpau_h264_picture_complete(s);
2771
2772     /*
2773      * FIXME: Error handling code does not seem to support interlaced
2774      * when slices span multiple rows
2775      * The ff_er_add_slice calls don't work right for bottom
2776      * fields; they cause massive erroneous error concealing
2777      * Error marking covers both fields (top and bottom).
2778      * This causes a mismatched s->error_count
2779      * and a bad error table. Further, the error count goes to
2780      * INT_MAX when called for bottom field, because mb_y is
2781      * past end by one (callers fault) and resync_mb_y != 0
2782      * causes problems for the first MB line, too.
2783      */
2784     if (!FIELD_PICTURE)
2785         ff_er_frame_end(s);
2786
2787     ff_MPV_frame_end(s);
2788
2789     h->current_slice = 0;
2790
2791     return err;
2792 }
2793
2794 /**
2795  * Replicate H264 "master" context to thread contexts.
2796  */
2797 static void clone_slice(H264Context *dst, H264Context *src)
2798 {
2799     memcpy(dst->block_offset, src->block_offset, sizeof(dst->block_offset));
2800     dst->s.current_picture_ptr = src->s.current_picture_ptr;
2801     dst->s.current_picture     = src->s.current_picture;
2802     dst->s.linesize            = src->s.linesize;
2803     dst->s.uvlinesize          = src->s.uvlinesize;
2804     dst->s.first_field         = src->s.first_field;
2805
2806     dst->prev_poc_msb          = src->prev_poc_msb;
2807     dst->prev_poc_lsb          = src->prev_poc_lsb;
2808     dst->prev_frame_num_offset = src->prev_frame_num_offset;
2809     dst->prev_frame_num        = src->prev_frame_num;
2810     dst->short_ref_count       = src->short_ref_count;
2811
2812     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
2813     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
2814     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2815     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
2816
2817     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
2818     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
2819 }
2820
2821 /**
2822  * Compute profile from profile_idc and constraint_set?_flags.
2823  *
2824  * @param sps SPS
2825  *
2826  * @return profile as defined by FF_PROFILE_H264_*
2827  */
2828 int ff_h264_get_profile(SPS *sps)
2829 {
2830     int profile = sps->profile_idc;
2831
2832     switch (sps->profile_idc) {
2833     case FF_PROFILE_H264_BASELINE:
2834         // constraint_set1_flag set to 1
2835         profile |= (sps->constraint_set_flags & 1 << 1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2836         break;
2837     case FF_PROFILE_H264_HIGH_10:
2838     case FF_PROFILE_H264_HIGH_422:
2839     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2840         // constraint_set3_flag set to 1
2841         profile |= (sps->constraint_set_flags & 1 << 3) ? FF_PROFILE_H264_INTRA : 0;
2842         break;
2843     }
2844
2845     return profile;
2846 }
2847
2848 /**
2849  * Decode a slice header.
2850  * This will also call ff_MPV_common_init() and frame_start() as needed.
2851  *
2852  * @param h h264context
2853  * @param h0 h264 master context (differs from 'h' when doing sliced based
2854  *           parallel decoding)
2855  *
2856  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2857  */
2858 static int decode_slice_header(H264Context *h, H264Context *h0)
2859 {
2860     MpegEncContext *const s  = &h->s;
2861     MpegEncContext *const s0 = &h0->s;
2862     unsigned int first_mb_in_slice;
2863     unsigned int pps_id;
2864     int num_ref_idx_active_override_flag;
2865     unsigned int slice_type, tmp, i, j;
2866     int default_ref_list_done = 0;
2867     int last_pic_structure, last_pic_dropable;
2868     int must_reinit;
2869
2870     /* FIXME: 2tap qpel isn't implemented for high bit depth. */
2871     if ((s->avctx->flags2 & CODEC_FLAG2_FAST) &&
2872         !h->nal_ref_idc && !h->pixel_shift) {
2873         s->me.qpel_put = s->dsp.put_2tap_qpel_pixels_tab;
2874         s->me.qpel_avg = s->dsp.avg_2tap_qpel_pixels_tab;
2875     } else {
2876         s->me.qpel_put = s->dsp.put_h264_qpel_pixels_tab;
2877         s->me.qpel_avg = s->dsp.avg_h264_qpel_pixels_tab;
2878     }
2879
2880     first_mb_in_slice = get_ue_golomb_long(&s->gb);
2881
2882     if (first_mb_in_slice == 0) { // FIXME better field boundary detection
2883         if (h0->current_slice && FIELD_PICTURE) {
2884             field_end(h, 1);
2885         }
2886
2887         h0->current_slice = 0;
2888         if (!s0->first_field) {
2889             if (s->current_picture_ptr && !s->dropable &&
2890                 s->current_picture_ptr->owner2 == s) {
2891                 ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX,
2892                                           s->picture_structure == PICT_BOTTOM_FIELD);
2893             }
2894             s->current_picture_ptr = NULL;
2895         }
2896     }
2897
2898     slice_type = get_ue_golomb_31(&s->gb);
2899     if (slice_type > 9) {
2900         av_log(h->s.avctx, AV_LOG_ERROR,
2901                "slice type too large (%d) at %d %d\n",
2902                h->slice_type, s->mb_x, s->mb_y);
2903         return -1;
2904     }
2905     if (slice_type > 4) {
2906         slice_type -= 5;
2907         h->slice_type_fixed = 1;
2908     } else
2909         h->slice_type_fixed = 0;
2910
2911     slice_type = golomb_to_pict_type[slice_type];
2912     if (slice_type == AV_PICTURE_TYPE_I ||
2913         (h0->current_slice != 0 && slice_type == h0->last_slice_type)) {
2914         default_ref_list_done = 1;
2915     }
2916     h->slice_type     = slice_type;
2917     h->slice_type_nos = slice_type & 3;
2918
2919     // to make a few old functions happy, it's wrong though
2920     s->pict_type = h->slice_type;
2921
2922     pps_id = get_ue_golomb(&s->gb);
2923     if (pps_id >= MAX_PPS_COUNT) {
2924         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id %d out of range\n", pps_id);
2925         return -1;
2926     }
2927     if (!h0->pps_buffers[pps_id]) {
2928         av_log(h->s.avctx, AV_LOG_ERROR,
2929                "non-existing PPS %u referenced\n",
2930                pps_id);
2931         return -1;
2932     }
2933     h->pps = *h0->pps_buffers[pps_id];
2934
2935     if (!h0->sps_buffers[h->pps.sps_id]) {
2936         av_log(h->s.avctx, AV_LOG_ERROR,
2937                "non-existing SPS %u referenced\n",
2938                h->pps.sps_id);
2939         return -1;
2940     }
2941     h->sps = *h0->sps_buffers[h->pps.sps_id];
2942
2943     s->avctx->profile = ff_h264_get_profile(&h->sps);
2944     s->avctx->level   = h->sps.level_idc;
2945     s->avctx->refs    = h->sps.ref_frame_count;
2946
2947     must_reinit = (s->context_initialized &&
2948                     (   16*h->sps.mb_width != s->avctx->coded_width
2949                      || 16*h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag) != s->avctx->coded_height
2950                      || s->avctx->bits_per_raw_sample != h->sps.bit_depth_luma
2951                      || h->cur_chroma_format_idc != h->sps.chroma_format_idc
2952                      || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio)));
2953
2954     if(must_reinit && (h != h0 || (s->avctx->active_thread_type & FF_THREAD_FRAME))) {
2955         av_log_missing_feature(s->avctx,
2956                                 "Width/height/bit depth/chroma idc changing with threads is", 0);
2957         return -1;   // width / height changed during parallelized decoding
2958     }
2959
2960     s->mb_width  = h->sps.mb_width;
2961     s->mb_height = h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2962
2963     h->b_stride = s->mb_width * 4;
2964
2965     s->chroma_y_shift = h->sps.chroma_format_idc <= 1; // 400 uses yuv420p
2966
2967     s->width  = 16 * s->mb_width;
2968     s->height = 16 * s->mb_height;
2969
2970
2971     if(must_reinit) {
2972         free_tables(h, 0);
2973         flush_dpb(s->avctx);
2974         ff_MPV_common_end(s);
2975         h->list_count = 0;
2976     }
2977     if (!s->context_initialized) {
2978         if (h != h0) {
2979             av_log(h->s.avctx, AV_LOG_ERROR,
2980                    "Cannot (re-)initialize context during parallel decoding.\n");
2981             return -1;
2982         }
2983         avcodec_set_dimensions(s->avctx, s->width, s->height);
2984         s->avctx->width  -= (2>>CHROMA444)*FFMIN(h->sps.crop_right, (8<<CHROMA444)-1);
2985         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);
2986         s->avctx->sample_aspect_ratio = h->sps.sar;
2987         av_assert0(s->avctx->sample_aspect_ratio.den);
2988
2989         if (s->avctx->bits_per_raw_sample != h->sps.bit_depth_luma ||
2990             h->cur_chroma_format_idc != h->sps.chroma_format_idc) {
2991             if (h->sps.bit_depth_luma >= 8 && h->sps.bit_depth_luma <= 10 &&
2992                 (h->sps.bit_depth_luma != 9 || !CHROMA422)) {
2993                 s->avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
2994                 h->cur_chroma_format_idc = h->sps.chroma_format_idc;
2995                 h->pixel_shift = h->sps.bit_depth_luma > 8;
2996
2997                 ff_h264dsp_init(&h->h264dsp, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
2998                 ff_h264_pred_init(&h->hpc, s->codec_id, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
2999                 s->dsp.dct_bits = h->sps.bit_depth_luma > 8 ? 32 : 16;
3000                 ff_dsputil_init(&s->dsp, s->avctx);
3001             } else {
3002                 av_log(s->avctx, AV_LOG_ERROR, "Unsupported bit depth: %d chroma_idc: %d\n",
3003                        h->sps.bit_depth_luma, h->sps.chroma_format_idc);
3004                 return -1;
3005             }
3006         }
3007
3008         if (h->sps.video_signal_type_present_flag) {
3009             s->avctx->color_range = h->sps.full_range>0 ? AVCOL_RANGE_JPEG
3010                                                       : AVCOL_RANGE_MPEG;
3011             if (h->sps.colour_description_present_flag) {
3012                 s->avctx->color_primaries = h->sps.color_primaries;
3013                 s->avctx->color_trc       = h->sps.color_trc;
3014                 s->avctx->colorspace      = h->sps.colorspace;
3015             }
3016         }
3017
3018         if (h->sps.timing_info_present_flag) {
3019             int64_t den = h->sps.time_scale;
3020             if (h->x264_build < 44U)
3021                 den *= 2;
3022             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
3023                       h->sps.num_units_in_tick, den, 1 << 30);
3024         }
3025
3026         switch (h->sps.bit_depth_luma) {
3027         case 9:
3028             if (CHROMA444) {
3029                 if (s->avctx->colorspace == AVCOL_SPC_RGB) {
3030                     s->avctx->pix_fmt = PIX_FMT_GBRP9;
3031                 } else
3032                     s->avctx->pix_fmt = PIX_FMT_YUV444P9;
3033             } else if (CHROMA422)
3034                 s->avctx->pix_fmt = PIX_FMT_YUV422P9;
3035             else
3036                 s->avctx->pix_fmt = PIX_FMT_YUV420P9;
3037             break;
3038         case 10:
3039             if (CHROMA444) {
3040                 if (s->avctx->colorspace == AVCOL_SPC_RGB) {
3041                     s->avctx->pix_fmt = PIX_FMT_GBRP10;
3042                 } else
3043                     s->avctx->pix_fmt = PIX_FMT_YUV444P10;
3044             } else if (CHROMA422)
3045                 s->avctx->pix_fmt = PIX_FMT_YUV422P10;
3046             else
3047                 s->avctx->pix_fmt = PIX_FMT_YUV420P10;
3048             break;
3049         case 8:
3050             if (CHROMA444) {
3051                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ444P
3052                                                                                   : PIX_FMT_YUV444P;
3053                     if (s->avctx->colorspace == AVCOL_SPC_RGB) {
3054                         s->avctx->pix_fmt = PIX_FMT_GBR24P;
3055                         av_log(h->s.avctx, AV_LOG_DEBUG, "Detected GBR colorspace.\n");
3056                     } else if (s->avctx->colorspace == AVCOL_SPC_YCGCO) {
3057                         av_log(h->s.avctx, AV_LOG_WARNING, "Detected unsupported YCgCo colorspace.\n");
3058                     }
3059             } else if (CHROMA422) {
3060                 s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ422P
3061                                                                               : PIX_FMT_YUV422P;
3062             } else {
3063                 s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
3064                                                          s->avctx->codec->pix_fmts ?
3065                                                          s->avctx->codec->pix_fmts :
3066                                                          s->avctx->color_range == AVCOL_RANGE_JPEG ?
3067                                                          hwaccel_pixfmt_list_h264_jpeg_420 :
3068                                                          ff_hwaccel_pixfmt_list_420);
3069             }
3070             break;
3071         default:
3072             av_log(s->avctx, AV_LOG_ERROR,
3073                    "Unsupported bit depth: %d\n", h->sps.bit_depth_luma);
3074             return AVERROR_INVALIDDATA;
3075         }
3076
3077         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id,
3078                                             s->avctx->pix_fmt);
3079
3080         if (ff_MPV_common_init(s) < 0) {
3081             av_log(h->s.avctx, AV_LOG_ERROR, "ff_MPV_common_init() failed.\n");
3082             return -1;
3083         }
3084         s->first_field = 0;
3085         h->prev_interlaced_frame = 1;
3086
3087         init_scan_tables(h);
3088         if (ff_h264_alloc_tables(h) < 0) {
3089             av_log(h->s.avctx, AV_LOG_ERROR,
3090                    "Could not allocate memory for h264\n");
3091             return AVERROR(ENOMEM);
3092         }
3093
3094         if (!HAVE_THREADS || !(s->avctx->active_thread_type & FF_THREAD_SLICE)) {
3095             if (context_init(h) < 0) {
3096                 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
3097                 return -1;
3098             }
3099         } else {
3100             for (i = 1; i < s->slice_context_count; i++) {
3101                 H264Context *c;
3102                 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
3103                 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
3104                 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
3105                 c->h264dsp     = h->h264dsp;
3106                 c->sps         = h->sps;
3107                 c->pps         = h->pps;
3108                 c->pixel_shift = h->pixel_shift;
3109                 c->cur_chroma_format_idc = h->cur_chroma_format_idc;
3110                 init_scan_tables(c);
3111                 clone_tables(c, h, i);
3112             }
3113
3114             for (i = 0; i < s->slice_context_count; i++)
3115                 if (context_init(h->thread_context[i]) < 0) {
3116                     av_log(h->s.avctx, AV_LOG_ERROR,
3117                            "context_init() failed.\n");
3118                     return -1;
3119                 }
3120         }
3121     }
3122
3123     if (h == h0 && h->dequant_coeff_pps != pps_id) {
3124         h->dequant_coeff_pps = pps_id;
3125         init_dequant_tables(h);
3126     }
3127
3128     h->frame_num = get_bits(&s->gb, h->sps.log2_max_frame_num);
3129
3130     h->mb_mbaff        = 0;
3131     h->mb_aff_frame    = 0;
3132     last_pic_structure = s0->picture_structure;
3133     last_pic_dropable  = s->dropable;
3134     s->dropable        = h->nal_ref_idc == 0;
3135     if (h->sps.frame_mbs_only_flag) {
3136         s->picture_structure = PICT_FRAME;
3137     } else {
3138         if (!h->sps.direct_8x8_inference_flag && slice_type == AV_PICTURE_TYPE_B) {
3139             av_log(h->s.avctx, AV_LOG_ERROR, "This stream was generated by a broken encoder, invalid 8x8 inference\n");
3140             return -1;
3141         }
3142         if (get_bits1(&s->gb)) { // field_pic_flag
3143             s->picture_structure = PICT_TOP_FIELD + get_bits1(&s->gb); // bottom_field_flag
3144         } else {
3145             s->picture_structure = PICT_FRAME;
3146             h->mb_aff_frame      = h->sps.mb_aff;
3147         }
3148     }
3149     h->mb_field_decoding_flag = s->picture_structure != PICT_FRAME;
3150
3151     if (h0->current_slice != 0) {
3152         if (last_pic_structure != s->picture_structure ||
3153             last_pic_dropable  != s->dropable) {
3154             av_log(h->s.avctx, AV_LOG_ERROR,
3155                    "Changing field mode (%d -> %d) between slices is not allowed\n",
3156                    last_pic_structure, s->picture_structure);
3157             s->picture_structure = last_pic_structure;
3158             s->dropable          = last_pic_dropable;
3159             return AVERROR_INVALIDDATA;
3160         }
3161     } else {
3162         /* Shorten frame num gaps so we don't have to allocate reference
3163          * frames just to throw them away */
3164         if (h->frame_num != h->prev_frame_num && h->prev_frame_num >= 0) {
3165             int unwrap_prev_frame_num = h->prev_frame_num;
3166             int max_frame_num         = 1 << h->sps.log2_max_frame_num;
3167
3168             if (unwrap_prev_frame_num > h->frame_num)
3169                 unwrap_prev_frame_num -= max_frame_num;
3170
3171             if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
3172                 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
3173                 if (unwrap_prev_frame_num < 0)
3174                     unwrap_prev_frame_num += max_frame_num;
3175
3176                 h->prev_frame_num = unwrap_prev_frame_num;
3177             }
3178         }
3179
3180         /* See if we have a decoded first field looking for a pair...
3181          * Here, we're using that to see if we should mark previously
3182          * decode frames as "finished".
3183          * We have to do that before the "dummy" in-between frame allocation,
3184          * since that can modify s->current_picture_ptr. */
3185         if (s0->first_field) {
3186             assert(s0->current_picture_ptr);
3187             assert(s0->current_picture_ptr->f.data[0]);
3188             assert(s0->current_picture_ptr->f.reference != DELAYED_PIC_REF);
3189
3190             /* Mark old field/frame as completed */
3191             if (!last_pic_dropable && s0->current_picture_ptr->owner2 == s0) {
3192                 ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
3193                                           last_pic_structure == PICT_BOTTOM_FIELD);
3194             }
3195
3196             /* figure out if we have a complementary field pair */
3197             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
3198                 /* Previous field is unmatched. Don't display it, but let it
3199                  * remain for reference if marked as such. */
3200                 if (!last_pic_dropable && last_pic_structure != PICT_FRAME) {
3201                     ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
3202                                               last_pic_structure == PICT_TOP_FIELD);
3203                 }
3204             } else {
3205                 if (s0->current_picture_ptr->frame_num != h->frame_num) {
3206                     /* This and previous field were reference, but had
3207                      * different frame_nums. Consider this field first in
3208                      * pair. Throw away previous field except for reference
3209                      * purposes. */
3210                     if (!last_pic_dropable && last_pic_structure != PICT_FRAME) {
3211                         ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
3212                                                   last_pic_structure == PICT_TOP_FIELD);
3213                     }
3214                 } else {
3215                     /* Second field in complementary pair */
3216                     if (!((last_pic_structure   == PICT_TOP_FIELD &&
3217                            s->picture_structure == PICT_BOTTOM_FIELD) ||
3218                           (last_pic_structure   == PICT_BOTTOM_FIELD &&
3219                            s->picture_structure == PICT_TOP_FIELD))) {
3220                         av_log(s->avctx, AV_LOG_ERROR,
3221                                "Invalid field mode combination %d/%d\n",
3222                                last_pic_structure, s->picture_structure);
3223                         s->picture_structure = last_pic_structure;
3224                         s->dropable          = last_pic_dropable;
3225                         return AVERROR_INVALIDDATA;
3226                     } else if (last_pic_dropable != s->dropable) {
3227                         av_log(s->avctx, AV_LOG_ERROR,
3228                                "Cannot combine reference and non-reference fields in the same frame\n");
3229                         av_log_ask_for_sample(s->avctx, NULL);
3230                         s->picture_structure = last_pic_structure;
3231                         s->dropable          = last_pic_dropable;
3232                         return AVERROR_INVALIDDATA;
3233                     }
3234
3235                     /* Take ownership of this buffer. Note that if another thread owned
3236                      * the first field of this buffer, we're not operating on that pointer,
3237                      * so the original thread is still responsible for reporting progress
3238                      * on that first field (or if that was us, we just did that above).
3239                      * By taking ownership, we assign responsibility to ourselves to
3240                      * report progress on the second field. */
3241                     s0->current_picture_ptr->owner2 = s0;
3242                 }
3243             }
3244         }
3245
3246         while (h->frame_num != h->prev_frame_num && h->prev_frame_num >= 0 &&
3247                h->frame_num != (h->prev_frame_num + 1) % (1 << h->sps.log2_max_frame_num)) {
3248             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
3249             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n",
3250                    h->frame_num, h->prev_frame_num);
3251             if (ff_h264_frame_start(h) < 0)
3252                 return -1;
3253             h->prev_frame_num++;
3254             h->prev_frame_num %= 1 << h->sps.log2_max_frame_num;
3255             s->current_picture_ptr->frame_num = h->prev_frame_num;
3256             ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX, 0);
3257             ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX, 1);
3258             ff_generate_sliding_window_mmcos(h);
3259             if (ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index) < 0 &&
3260                 (s->avctx->err_recognition & AV_EF_EXPLODE))
3261                 return AVERROR_INVALIDDATA;
3262             /* Error concealment: if a ref is missing, copy the previous ref in its place.
3263              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
3264              * about there being no actual duplicates.
3265              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
3266              * concealing a lost frame, this probably isn't noticeable by comparison, but it should
3267              * be fixed. */
3268             if (h->short_ref_count) {
3269                 if (prev) {
3270                     av_image_copy(h->short_ref[0]->f.data, h->short_ref[0]->f.linesize,
3271                                   (const uint8_t **)prev->f.data, prev->f.linesize,