2 * H.26L/H.264/AVC/JVT/14496-10/... decoder
3 * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
5 * This file is part of Libav.
7 * Libav is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * Libav is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with Libav; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
24 * H.264 / AVC / MPEG4 part10 codec.
25 * @author Michael Niedermayer <michaelni@gmx.at>
28 #include "libavutil/imgutils.h"
31 #include "cabac_functions.h"
34 #include "mpegvideo.h"
37 #include "h264_mvpred.h"
40 #include "rectangle.h"
42 #include "vdpau_internal.h"
43 #include "libavutil/avassert.h"
48 const uint16_t ff_h264_mb_sizes[4] = { 256, 384, 512, 768 };
50 static const uint8_t rem6[QP_MAX_NUM + 1] = {
51 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2,
52 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5,
53 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
56 static const uint8_t div6[QP_MAX_NUM + 1] = {
57 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3, 3, 3,
58 3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6,
59 7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 10, 10, 10, 10,
62 static const enum AVPixelFormat hwaccel_pixfmt_list_h264_jpeg_420[] = {
71 * Check if the top & left blocks are available if needed and
72 * change the dc mode so it only uses the available blocks.
74 int ff_h264_check_intra4x4_pred_mode(H264Context *h)
76 MpegEncContext *const s = &h->s;
77 static const int8_t top[12] = {
78 -1, 0, LEFT_DC_PRED, -1, -1, -1, -1, -1, 0
80 static const int8_t left[12] = {
81 0, -1, TOP_DC_PRED, 0, -1, -1, -1, 0, -1, DC_128_PRED
85 if (!(h->top_samples_available & 0x8000)) {
86 for (i = 0; i < 4; i++) {
87 int status = top[h->intra4x4_pred_mode_cache[scan8[0] + i]];
89 av_log(h->s.avctx, AV_LOG_ERROR,
90 "top block unavailable for requested intra4x4 mode %d at %d %d\n",
91 status, s->mb_x, s->mb_y);
94 h->intra4x4_pred_mode_cache[scan8[0] + i] = status;
99 if ((h->left_samples_available & 0x8888) != 0x8888) {
100 static const int mask[4] = { 0x8000, 0x2000, 0x80, 0x20 };
101 for (i = 0; i < 4; i++)
102 if (!(h->left_samples_available & mask[i])) {
103 int status = left[h->intra4x4_pred_mode_cache[scan8[0] + 8 * i]];
105 av_log(h->s.avctx, AV_LOG_ERROR,
106 "left block unavailable for requested intra4x4 mode %d at %d %d\n",
107 status, s->mb_x, s->mb_y);
110 h->intra4x4_pred_mode_cache[scan8[0] + 8 * i] = status;
116 } // FIXME cleanup like ff_h264_check_intra_pred_mode
119 * Check if the top & left blocks are available if needed and
120 * change the dc mode so it only uses the available blocks.
122 int ff_h264_check_intra_pred_mode(H264Context *h, int mode, int is_chroma)
124 MpegEncContext *const s = &h->s;
125 static const int8_t top[7] = { LEFT_DC_PRED8x8, 1, -1, -1 };
126 static const int8_t left[7] = { TOP_DC_PRED8x8, -1, 2, -1, DC_128_PRED8x8 };
129 av_log(h->s.avctx, AV_LOG_ERROR,
130 "out of range intra chroma pred mode at %d %d\n",
135 if (!(h->top_samples_available & 0x8000)) {
138 av_log(h->s.avctx, AV_LOG_ERROR,
139 "top block unavailable for requested intra mode at %d %d\n",
145 if ((h->left_samples_available & 0x8080) != 0x8080) {
147 if (is_chroma && (h->left_samples_available & 0x8080)) {
148 // mad cow disease mode, aka MBAFF + constrained_intra_pred
149 mode = ALZHEIMER_DC_L0T_PRED8x8 +
150 (!(h->left_samples_available & 0x8000)) +
151 2 * (mode == DC_128_PRED8x8);
154 av_log(h->s.avctx, AV_LOG_ERROR,
155 "left block unavailable for requested intra mode at %d %d\n",
164 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src,
165 int *dst_length, int *consumed, int length)
171 // src[0]&0x80; // forbidden bit
172 h->nal_ref_idc = src[0] >> 5;
173 h->nal_unit_type = src[0] & 0x1F;
178 #define STARTCODE_TEST \
179 if (i + 2 < length && src[i + 1] == 0 && src[i + 2] <= 3) { \
180 if (src[i + 2] != 3) { \
181 /* startcode, so we must be past the end */ \
186 #if HAVE_FAST_UNALIGNED
187 #define FIND_FIRST_ZERO \
188 if (i > 0 && !src[i]) \
193 for (i = 0; i + 1 < length; i += 9) {
194 if (!((~AV_RN64A(src + i) &
195 (AV_RN64A(src + i) - 0x0100010001000101ULL)) &
196 0x8000800080008080ULL))
203 for (i = 0; i + 1 < length; i += 5) {
204 if (!((~AV_RN32A(src + i) &
205 (AV_RN32A(src + i) - 0x01000101U)) &
214 for (i = 0; i + 1 < length; i += 2) {
217 if (i > 0 && src[i - 1] == 0)
223 if (i >= length - 1) { // no escaped 0
224 *dst_length = length;
225 *consumed = length + 1; // +1 for the header
229 // use second escape buffer for inter data
230 bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0;
231 av_fast_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx],
232 length + FF_INPUT_BUFFER_PADDING_SIZE);
233 dst = h->rbsp_buffer[bufidx];
240 while (si + 2 < length) {
241 // remove escapes (very rare 1:2^22)
242 if (src[si + 2] > 3) {
243 dst[di++] = src[si++];
244 dst[di++] = src[si++];
245 } else if (src[si] == 0 && src[si + 1] == 0) {
246 if (src[si + 2] == 3) { // escape
251 } else // next start code
255 dst[di++] = src[si++];
258 dst[di++] = src[si++];
261 memset(dst + di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
264 *consumed = si + 1; // +1 for the header
265 /* FIXME store exact number of bits in the getbitcontext
266 * (it is needed for decoding) */
271 * Identify the exact end of the bitstream
272 * @return the length of the trailing, or 0 if damaged
274 static int decode_rbsp_trailing(H264Context *h, const uint8_t *src)
279 tprintf(h->s.avctx, "rbsp trailing %X\n", v);
281 for (r = 1; r < 9; r++) {
289 static inline int get_lowest_part_list_y(H264Context *h, Picture *pic, int n,
290 int height, int y_offset, int list)
292 int raw_my = h->mv_cache[list][scan8[n]][1];
293 int filter_height = (raw_my & 3) ? 2 : 0;
294 int full_my = (raw_my >> 2) + y_offset;
295 int top = full_my - filter_height;
296 int bottom = full_my + filter_height + height;
298 return FFMAX(abs(top), bottom);
301 static inline void get_lowest_part_y(H264Context *h, int refs[2][48], int n,
302 int height, int y_offset, int list0,
303 int list1, int *nrefs)
305 MpegEncContext *const s = &h->s;
308 y_offset += 16 * (s->mb_y >> MB_FIELD);
311 int ref_n = h->ref_cache[0][scan8[n]];
312 Picture *ref = &h->ref_list[0][ref_n];
314 // Error resilience puts the current picture in the ref list.
315 // Don't try to wait on these as it will cause a deadlock.
316 // Fields can wait on each other, though.
317 if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
318 (ref->f.reference & 3) != s->picture_structure) {
319 my = get_lowest_part_list_y(h, ref, n, height, y_offset, 0);
320 if (refs[0][ref_n] < 0)
322 refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
327 int ref_n = h->ref_cache[1][scan8[n]];
328 Picture *ref = &h->ref_list[1][ref_n];
330 if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
331 (ref->f.reference & 3) != s->picture_structure) {
332 my = get_lowest_part_list_y(h, ref, n, height, y_offset, 1);
333 if (refs[1][ref_n] < 0)
335 refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
341 * Wait until all reference frames are available for MC operations.
343 * @param h the H264 context
345 static void await_references(H264Context *h)
347 MpegEncContext *const s = &h->s;
348 const int mb_xy = h->mb_xy;
349 const int mb_type = s->current_picture.f.mb_type[mb_xy];
351 int nrefs[2] = { 0 };
354 memset(refs, -1, sizeof(refs));
356 if (IS_16X16(mb_type)) {
357 get_lowest_part_y(h, refs, 0, 16, 0,
358 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
359 } else if (IS_16X8(mb_type)) {
360 get_lowest_part_y(h, refs, 0, 8, 0,
361 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
362 get_lowest_part_y(h, refs, 8, 8, 8,
363 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
364 } else if (IS_8X16(mb_type)) {
365 get_lowest_part_y(h, refs, 0, 16, 0,
366 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
367 get_lowest_part_y(h, refs, 4, 16, 0,
368 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
372 assert(IS_8X8(mb_type));
374 for (i = 0; i < 4; i++) {
375 const int sub_mb_type = h->sub_mb_type[i];
377 int y_offset = (i & 2) << 2;
379 if (IS_SUB_8X8(sub_mb_type)) {
380 get_lowest_part_y(h, refs, n, 8, y_offset,
381 IS_DIR(sub_mb_type, 0, 0),
382 IS_DIR(sub_mb_type, 0, 1),
384 } else if (IS_SUB_8X4(sub_mb_type)) {
385 get_lowest_part_y(h, refs, n, 4, y_offset,
386 IS_DIR(sub_mb_type, 0, 0),
387 IS_DIR(sub_mb_type, 0, 1),
389 get_lowest_part_y(h, refs, n + 2, 4, y_offset + 4,
390 IS_DIR(sub_mb_type, 0, 0),
391 IS_DIR(sub_mb_type, 0, 1),
393 } else if (IS_SUB_4X8(sub_mb_type)) {
394 get_lowest_part_y(h, refs, n, 8, y_offset,
395 IS_DIR(sub_mb_type, 0, 0),
396 IS_DIR(sub_mb_type, 0, 1),
398 get_lowest_part_y(h, refs, n + 1, 8, y_offset,
399 IS_DIR(sub_mb_type, 0, 0),
400 IS_DIR(sub_mb_type, 0, 1),
404 assert(IS_SUB_4X4(sub_mb_type));
405 for (j = 0; j < 4; j++) {
406 int sub_y_offset = y_offset + 2 * (j & 2);
407 get_lowest_part_y(h, refs, n + j, 4, sub_y_offset,
408 IS_DIR(sub_mb_type, 0, 0),
409 IS_DIR(sub_mb_type, 0, 1),
416 for (list = h->list_count - 1; list >= 0; list--)
417 for (ref = 0; ref < 48 && nrefs[list]; ref++) {
418 int row = refs[list][ref];
420 Picture *ref_pic = &h->ref_list[list][ref];
421 int ref_field = ref_pic->f.reference - 1;
422 int ref_field_picture = ref_pic->field_picture;
423 int pic_height = 16 * s->mb_height >> ref_field_picture;
428 if (!FIELD_PICTURE && ref_field_picture) { // frame referencing two fields
429 ff_thread_await_progress(&ref_pic->f,
430 FFMIN((row >> 1) - !(row & 1),
433 ff_thread_await_progress(&ref_pic->f,
434 FFMIN((row >> 1), pic_height - 1),
436 } else if (FIELD_PICTURE && !ref_field_picture) { // field referencing one field of a frame
437 ff_thread_await_progress(&ref_pic->f,
438 FFMIN(row * 2 + ref_field,
441 } else if (FIELD_PICTURE) {
442 ff_thread_await_progress(&ref_pic->f,
443 FFMIN(row, pic_height - 1),
446 ff_thread_await_progress(&ref_pic->f,
447 FFMIN(row, pic_height - 1),
454 static av_always_inline void mc_dir_part(H264Context *h, Picture *pic,
455 int n, int square, int height,
457 uint8_t *dest_y, uint8_t *dest_cb,
459 int src_x_offset, int src_y_offset,
460 qpel_mc_func *qpix_op,
461 h264_chroma_mc_func chroma_op,
462 int pixel_shift, int chroma_idc)
464 MpegEncContext *const s = &h->s;
465 const int mx = h->mv_cache[list][scan8[n]][0] + src_x_offset * 8;
466 int my = h->mv_cache[list][scan8[n]][1] + src_y_offset * 8;
467 const int luma_xy = (mx & 3) + ((my & 3) << 2);
468 int offset = ((mx >> 2) << pixel_shift) + (my >> 2) * h->mb_linesize;
469 uint8_t *src_y = pic->f.data[0] + offset;
470 uint8_t *src_cb, *src_cr;
471 int extra_width = h->emu_edge_width;
472 int extra_height = h->emu_edge_height;
474 const int full_mx = mx >> 2;
475 const int full_my = my >> 2;
476 const int pic_width = 16 * s->mb_width;
477 const int pic_height = 16 * s->mb_height >> MB_FIELD;
485 if (full_mx < 0 - extra_width ||
486 full_my < 0 - extra_height ||
487 full_mx + 16 /*FIXME*/ > pic_width + extra_width ||
488 full_my + 16 /*FIXME*/ > pic_height + extra_height) {
489 s->dsp.emulated_edge_mc(s->edge_emu_buffer,
490 src_y - (2 << pixel_shift) - 2 * h->mb_linesize,
492 16 + 5, 16 + 5 /*FIXME*/, full_mx - 2,
493 full_my - 2, pic_width, pic_height);
494 src_y = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
498 qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); // FIXME try variable height perhaps?
500 qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
502 if (CONFIG_GRAY && s->flags & CODEC_FLAG_GRAY)
505 if (chroma_idc == 3 /* yuv444 */) {
506 src_cb = pic->f.data[1] + offset;
508 s->dsp.emulated_edge_mc(s->edge_emu_buffer,
509 src_cb - (2 << pixel_shift) - 2 * h->mb_linesize,
511 16 + 5, 16 + 5 /*FIXME*/,
512 full_mx - 2, full_my - 2,
513 pic_width, pic_height);
514 src_cb = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
516 qpix_op[luma_xy](dest_cb, src_cb, h->mb_linesize); // FIXME try variable height perhaps?
518 qpix_op[luma_xy](dest_cb + delta, src_cb + delta, h->mb_linesize);
520 src_cr = pic->f.data[2] + offset;
522 s->dsp.emulated_edge_mc(s->edge_emu_buffer,
523 src_cr - (2 << pixel_shift) - 2 * h->mb_linesize,
525 16 + 5, 16 + 5 /*FIXME*/,
526 full_mx - 2, full_my - 2,
527 pic_width, pic_height);
528 src_cr = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
530 qpix_op[luma_xy](dest_cr, src_cr, h->mb_linesize); // FIXME try variable height perhaps?
532 qpix_op[luma_xy](dest_cr + delta, src_cr + delta, h->mb_linesize);
536 ysh = 3 - (chroma_idc == 2 /* yuv422 */);
537 if (chroma_idc == 1 /* yuv420 */ && MB_FIELD) {
538 // chroma offset when predicting from a field of opposite parity
539 my += 2 * ((s->mb_y & 1) - (pic->f.reference - 1));
540 emu |= (my >> 3) < 0 || (my >> 3) + 8 >= (pic_height >> 1);
543 src_cb = pic->f.data[1] + ((mx >> 3) << pixel_shift) +
544 (my >> ysh) * h->mb_uvlinesize;
545 src_cr = pic->f.data[2] + ((mx >> 3) << pixel_shift) +
546 (my >> ysh) * h->mb_uvlinesize;
549 s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize,
550 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
551 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
552 src_cb = s->edge_emu_buffer;
554 chroma_op(dest_cb, src_cb, h->mb_uvlinesize,
555 height >> (chroma_idc == 1 /* yuv420 */),
556 mx & 7, (my << (chroma_idc == 2 /* yuv422 */)) & 7);
559 s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize,
560 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
561 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
562 src_cr = s->edge_emu_buffer;
564 chroma_op(dest_cr, src_cr, h->mb_uvlinesize, height >> (chroma_idc == 1 /* yuv420 */),
565 mx & 7, (my << (chroma_idc == 2 /* yuv422 */)) & 7);
568 static av_always_inline void mc_part_std(H264Context *h, int n, int square,
569 int height, int delta,
570 uint8_t *dest_y, uint8_t *dest_cb,
572 int x_offset, int y_offset,
573 qpel_mc_func *qpix_put,
574 h264_chroma_mc_func chroma_put,
575 qpel_mc_func *qpix_avg,
576 h264_chroma_mc_func chroma_avg,
577 int list0, int list1,
578 int pixel_shift, int chroma_idc)
580 MpegEncContext *const s = &h->s;
581 qpel_mc_func *qpix_op = qpix_put;
582 h264_chroma_mc_func chroma_op = chroma_put;
584 dest_y += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
585 if (chroma_idc == 3 /* yuv444 */) {
586 dest_cb += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
587 dest_cr += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
588 } else if (chroma_idc == 2 /* yuv422 */) {
589 dest_cb += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
590 dest_cr += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
591 } else { /* yuv420 */
592 dest_cb += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
593 dest_cr += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
595 x_offset += 8 * s->mb_x;
596 y_offset += 8 * (s->mb_y >> MB_FIELD);
599 Picture *ref = &h->ref_list[0][h->ref_cache[0][scan8[n]]];
600 mc_dir_part(h, ref, n, square, height, delta, 0,
601 dest_y, dest_cb, dest_cr, x_offset, y_offset,
602 qpix_op, chroma_op, pixel_shift, chroma_idc);
605 chroma_op = chroma_avg;
609 Picture *ref = &h->ref_list[1][h->ref_cache[1][scan8[n]]];
610 mc_dir_part(h, ref, n, square, height, delta, 1,
611 dest_y, dest_cb, dest_cr, x_offset, y_offset,
612 qpix_op, chroma_op, pixel_shift, chroma_idc);
616 static av_always_inline void mc_part_weighted(H264Context *h, int n, int square,
617 int height, int delta,
618 uint8_t *dest_y, uint8_t *dest_cb,
620 int x_offset, int y_offset,
621 qpel_mc_func *qpix_put,
622 h264_chroma_mc_func chroma_put,
623 h264_weight_func luma_weight_op,
624 h264_weight_func chroma_weight_op,
625 h264_biweight_func luma_weight_avg,
626 h264_biweight_func chroma_weight_avg,
627 int list0, int list1,
628 int pixel_shift, int chroma_idc)
630 MpegEncContext *const s = &h->s;
633 dest_y += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
634 if (chroma_idc == 3 /* yuv444 */) {
635 chroma_height = height;
636 chroma_weight_avg = luma_weight_avg;
637 chroma_weight_op = luma_weight_op;
638 dest_cb += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
639 dest_cr += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
640 } else if (chroma_idc == 2 /* yuv422 */) {
641 chroma_height = height;
642 dest_cb += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
643 dest_cr += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
644 } else { /* yuv420 */
645 chroma_height = height >> 1;
646 dest_cb += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
647 dest_cr += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
649 x_offset += 8 * s->mb_x;
650 y_offset += 8 * (s->mb_y >> MB_FIELD);
652 if (list0 && list1) {
653 /* don't optimize for luma-only case, since B-frames usually
654 * use implicit weights => chroma too. */
655 uint8_t *tmp_cb = s->obmc_scratchpad;
656 uint8_t *tmp_cr = s->obmc_scratchpad + (16 << pixel_shift);
657 uint8_t *tmp_y = s->obmc_scratchpad + 16 * h->mb_uvlinesize;
658 int refn0 = h->ref_cache[0][scan8[n]];
659 int refn1 = h->ref_cache[1][scan8[n]];
661 mc_dir_part(h, &h->ref_list[0][refn0], n, square, height, delta, 0,
662 dest_y, dest_cb, dest_cr,
663 x_offset, y_offset, qpix_put, chroma_put,
664 pixel_shift, chroma_idc);
665 mc_dir_part(h, &h->ref_list[1][refn1], n, square, height, delta, 1,
666 tmp_y, tmp_cb, tmp_cr,
667 x_offset, y_offset, qpix_put, chroma_put,
668 pixel_shift, chroma_idc);
670 if (h->use_weight == 2) {
671 int weight0 = h->implicit_weight[refn0][refn1][s->mb_y & 1];
672 int weight1 = 64 - weight0;
673 luma_weight_avg(dest_y, tmp_y, h->mb_linesize,
674 height, 5, weight0, weight1, 0);
675 chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize,
676 chroma_height, 5, weight0, weight1, 0);
677 chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize,
678 chroma_height, 5, weight0, weight1, 0);
680 luma_weight_avg(dest_y, tmp_y, h->mb_linesize, height,
681 h->luma_log2_weight_denom,
682 h->luma_weight[refn0][0][0],
683 h->luma_weight[refn1][1][0],
684 h->luma_weight[refn0][0][1] +
685 h->luma_weight[refn1][1][1]);
686 chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, chroma_height,
687 h->chroma_log2_weight_denom,
688 h->chroma_weight[refn0][0][0][0],
689 h->chroma_weight[refn1][1][0][0],
690 h->chroma_weight[refn0][0][0][1] +
691 h->chroma_weight[refn1][1][0][1]);
692 chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, chroma_height,
693 h->chroma_log2_weight_denom,
694 h->chroma_weight[refn0][0][1][0],
695 h->chroma_weight[refn1][1][1][0],
696 h->chroma_weight[refn0][0][1][1] +
697 h->chroma_weight[refn1][1][1][1]);
700 int list = list1 ? 1 : 0;
701 int refn = h->ref_cache[list][scan8[n]];
702 Picture *ref = &h->ref_list[list][refn];
703 mc_dir_part(h, ref, n, square, height, delta, list,
704 dest_y, dest_cb, dest_cr, x_offset, y_offset,
705 qpix_put, chroma_put, pixel_shift, chroma_idc);
707 luma_weight_op(dest_y, h->mb_linesize, height,
708 h->luma_log2_weight_denom,
709 h->luma_weight[refn][list][0],
710 h->luma_weight[refn][list][1]);
711 if (h->use_weight_chroma) {
712 chroma_weight_op(dest_cb, h->mb_uvlinesize, chroma_height,
713 h->chroma_log2_weight_denom,
714 h->chroma_weight[refn][list][0][0],
715 h->chroma_weight[refn][list][0][1]);
716 chroma_weight_op(dest_cr, h->mb_uvlinesize, chroma_height,
717 h->chroma_log2_weight_denom,
718 h->chroma_weight[refn][list][1][0],
719 h->chroma_weight[refn][list][1][1]);
724 static av_always_inline void prefetch_motion(H264Context *h, int list,
725 int pixel_shift, int chroma_idc)
727 /* fetch pixels for estimated mv 4 macroblocks ahead
728 * optimized for 64byte cache lines */
729 MpegEncContext *const s = &h->s;
730 const int refn = h->ref_cache[list][scan8[0]];
732 const int mx = (h->mv_cache[list][scan8[0]][0] >> 2) + 16 * s->mb_x + 8;
733 const int my = (h->mv_cache[list][scan8[0]][1] >> 2) + 16 * s->mb_y;
734 uint8_t **src = h->ref_list[list][refn].f.data;
735 int off = (mx << pixel_shift) +
736 (my + (s->mb_x & 3) * 4) * h->mb_linesize +
738 s->dsp.prefetch(src[0] + off, s->linesize, 4);
739 if (chroma_idc == 3 /* yuv444 */) {
740 s->dsp.prefetch(src[1] + off, s->linesize, 4);
741 s->dsp.prefetch(src[2] + off, s->linesize, 4);
743 off = ((mx >> 1) << pixel_shift) +
744 ((my >> 1) + (s->mb_x & 7)) * s->uvlinesize +
746 s->dsp.prefetch(src[1] + off, src[2] - src[1], 2);
751 static void free_tables(H264Context *h, int free_rbsp)
756 av_freep(&h->intra4x4_pred_mode);
757 av_freep(&h->chroma_pred_mode_table);
758 av_freep(&h->cbp_table);
759 av_freep(&h->mvd_table[0]);
760 av_freep(&h->mvd_table[1]);
761 av_freep(&h->direct_table);
762 av_freep(&h->non_zero_count);
763 av_freep(&h->slice_table_base);
764 h->slice_table = NULL;
765 av_freep(&h->list_counts);
767 av_freep(&h->mb2b_xy);
768 av_freep(&h->mb2br_xy);
770 for (i = 0; i < MAX_THREADS; i++) {
771 hx = h->thread_context[i];
774 av_freep(&hx->top_borders[1]);
775 av_freep(&hx->top_borders[0]);
776 av_freep(&hx->s.obmc_scratchpad);
778 av_freep(&hx->rbsp_buffer[1]);
779 av_freep(&hx->rbsp_buffer[0]);
780 hx->rbsp_buffer_size[0] = 0;
781 hx->rbsp_buffer_size[1] = 0;
784 av_freep(&h->thread_context[i]);
788 static void init_dequant8_coeff_table(H264Context *h)
791 const int max_qp = 51 + 6 * (h->sps.bit_depth_luma - 8);
793 for (i = 0; i < 6; i++) {
794 h->dequant8_coeff[i] = h->dequant8_buffer[i];
795 for (j = 0; j < i; j++)
796 if (!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i],
797 64 * sizeof(uint8_t))) {
798 h->dequant8_coeff[i] = h->dequant8_buffer[j];
804 for (q = 0; q < max_qp + 1; q++) {
807 for (x = 0; x < 64; x++)
808 h->dequant8_coeff[i][q][(x >> 3) | ((x & 7) << 3)] =
809 ((uint32_t)dequant8_coeff_init[idx][dequant8_coeff_init_scan[((x >> 1) & 12) | (x & 3)]] *
810 h->pps.scaling_matrix8[i][x]) << shift;
815 static void init_dequant4_coeff_table(H264Context *h)
818 const int max_qp = 51 + 6 * (h->sps.bit_depth_luma - 8);
819 for (i = 0; i < 6; i++) {
820 h->dequant4_coeff[i] = h->dequant4_buffer[i];
821 for (j = 0; j < i; j++)
822 if (!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i],
823 16 * sizeof(uint8_t))) {
824 h->dequant4_coeff[i] = h->dequant4_buffer[j];
830 for (q = 0; q < max_qp + 1; q++) {
831 int shift = div6[q] + 2;
833 for (x = 0; x < 16; x++)
834 h->dequant4_coeff[i][q][(x >> 2) | ((x << 2) & 0xF)] =
835 ((uint32_t)dequant4_coeff_init[idx][(x & 1) + ((x >> 2) & 1)] *
836 h->pps.scaling_matrix4[i][x]) << shift;
841 static void init_dequant_tables(H264Context *h)
844 init_dequant4_coeff_table(h);
845 if (h->pps.transform_8x8_mode)
846 init_dequant8_coeff_table(h);
847 if (h->sps.transform_bypass) {
848 for (i = 0; i < 6; i++)
849 for (x = 0; x < 16; x++)
850 h->dequant4_coeff[i][0][x] = 1 << 6;
851 if (h->pps.transform_8x8_mode)
852 for (i = 0; i < 6; i++)
853 for (x = 0; x < 64; x++)
854 h->dequant8_coeff[i][0][x] = 1 << 6;
858 int ff_h264_alloc_tables(H264Context *h)
860 MpegEncContext *const s = &h->s;
861 const int big_mb_num = s->mb_stride * (s->mb_height + 1);
862 const int row_mb_num = s->mb_stride * 2 * s->avctx->thread_count;
865 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode,
866 row_mb_num * 8 * sizeof(uint8_t), fail)
867 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count,
868 big_mb_num * 48 * sizeof(uint8_t), fail)
869 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base,
870 (big_mb_num + s->mb_stride) * sizeof(*h->slice_table_base), fail)
871 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table,
872 big_mb_num * sizeof(uint16_t), fail)
873 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table,
874 big_mb_num * sizeof(uint8_t), fail)
875 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0],
876 16 * row_mb_num * sizeof(uint8_t), fail);
877 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1],
878 16 * row_mb_num * sizeof(uint8_t), fail);
879 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table,
880 4 * big_mb_num * sizeof(uint8_t), fail);
881 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts,
882 big_mb_num * sizeof(uint8_t), fail)
884 memset(h->slice_table_base, -1,
885 (big_mb_num + s->mb_stride) * sizeof(*h->slice_table_base));
886 h->slice_table = h->slice_table_base + s->mb_stride * 2 + 1;
888 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy,
889 big_mb_num * sizeof(uint32_t), fail);
890 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy,
891 big_mb_num * sizeof(uint32_t), fail);
892 for (y = 0; y < s->mb_height; y++)
893 for (x = 0; x < s->mb_width; x++) {
894 const int mb_xy = x + y * s->mb_stride;
895 const int b_xy = 4 * x + 4 * y * h->b_stride;
897 h->mb2b_xy[mb_xy] = b_xy;
898 h->mb2br_xy[mb_xy] = 8 * (FMO ? mb_xy : (mb_xy % (2 * s->mb_stride)));
901 s->obmc_scratchpad = NULL;
903 if (!h->dequant4_coeff[0])
904 init_dequant_tables(h);
914 * Mimic alloc_tables(), but for every context thread.
916 static void clone_tables(H264Context *dst, H264Context *src, int i)
918 MpegEncContext *const s = &src->s;
919 dst->intra4x4_pred_mode = src->intra4x4_pred_mode + i * 8 * 2 * s->mb_stride;
920 dst->non_zero_count = src->non_zero_count;
921 dst->slice_table = src->slice_table;
922 dst->cbp_table = src->cbp_table;
923 dst->mb2b_xy = src->mb2b_xy;
924 dst->mb2br_xy = src->mb2br_xy;
925 dst->chroma_pred_mode_table = src->chroma_pred_mode_table;
926 dst->mvd_table[0] = src->mvd_table[0] + i * 8 * 2 * s->mb_stride;
927 dst->mvd_table[1] = src->mvd_table[1] + i * 8 * 2 * s->mb_stride;
928 dst->direct_table = src->direct_table;
929 dst->list_counts = src->list_counts;
930 dst->s.obmc_scratchpad = NULL;
931 ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma,
932 src->sps.chroma_format_idc);
937 * Allocate buffers which are not shared amongst multiple threads.
939 static int context_init(H264Context *h)
941 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0],
942 h->s.mb_width * 16 * 3 * sizeof(uint8_t) * 2, fail)
943 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1],
944 h->s.mb_width * 16 * 3 * sizeof(uint8_t) * 2, fail)
946 h->ref_cache[0][scan8[5] + 1] =
947 h->ref_cache[0][scan8[7] + 1] =
948 h->ref_cache[0][scan8[13] + 1] =
949 h->ref_cache[1][scan8[5] + 1] =
950 h->ref_cache[1][scan8[7] + 1] =
951 h->ref_cache[1][scan8[13] + 1] = PART_NOT_AVAILABLE;
956 return -1; // free_tables will clean up for us
959 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
961 static av_cold void common_init(H264Context *h)
963 MpegEncContext *const s = &h->s;
965 s->width = s->avctx->width;
966 s->height = s->avctx->height;
967 s->codec_id = s->avctx->codec->id;
969 ff_h264dsp_init(&h->h264dsp, 8, 1);
970 ff_h264_pred_init(&h->hpc, s->codec_id, 8, 1);
972 h->dequant_coeff_pps = -1;
973 s->unrestricted_mv = 1;
975 /* needed so that IDCT permutation is known early */
976 ff_dsputil_init(&s->dsp, s->avctx);
978 memset(h->pps.scaling_matrix4, 16, 6 * 16 * sizeof(uint8_t));
979 memset(h->pps.scaling_matrix8, 16, 2 * 64 * sizeof(uint8_t));
982 int ff_h264_decode_extradata(H264Context *h)
984 AVCodecContext *avctx = h->s.avctx;
986 if (avctx->extradata[0] == 1) {
988 unsigned char *p = avctx->extradata;
992 if (avctx->extradata_size < 7) {
993 av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
996 /* sps and pps in the avcC always have length coded with 2 bytes,
997 * so put a fake nal_length_size = 2 while parsing them */
998 h->nal_length_size = 2;
999 // Decode sps from avcC
1000 cnt = *(p + 5) & 0x1f; // Number of sps
1002 for (i = 0; i < cnt; i++) {
1003 nalsize = AV_RB16(p) + 2;
1004 if (p - avctx->extradata + nalsize > avctx->extradata_size)
1006 if (decode_nal_units(h, p, nalsize) < 0) {
1007 av_log(avctx, AV_LOG_ERROR,
1008 "Decoding sps %d from avcC failed\n", i);
1013 // Decode pps from avcC
1014 cnt = *(p++); // Number of pps
1015 for (i = 0; i < cnt; i++) {
1016 nalsize = AV_RB16(p) + 2;
1017 if (p - avctx->extradata + nalsize > avctx->extradata_size)
1019 if (decode_nal_units(h, p, nalsize) < 0) {
1020 av_log(avctx, AV_LOG_ERROR,
1021 "Decoding pps %d from avcC failed\n", i);
1026 // Now store right nal length size, that will be used to parse all other nals
1027 h->nal_length_size = (avctx->extradata[4] & 0x03) + 1;
1030 if (decode_nal_units(h, avctx->extradata, avctx->extradata_size) < 0)
1036 av_cold int ff_h264_decode_init(AVCodecContext *avctx)
1038 H264Context *h = avctx->priv_data;
1039 MpegEncContext *const s = &h->s;
1042 ff_MPV_decode_defaults(s);
1047 s->out_format = FMT_H264;
1048 s->workaround_bugs = avctx->workaround_bugs;
1051 // s->decode_mb = ff_h263_decode_mb;
1052 s->quarter_sample = 1;
1053 if (!avctx->has_b_frames)
1056 avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1058 ff_h264_decode_init_vlc();
1061 h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1063 h->thread_context[0] = h;
1064 h->outputed_poc = h->next_outputed_poc = INT_MIN;
1065 for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
1066 h->last_pocs[i] = INT_MIN;
1067 h->prev_poc_msb = 1 << 16;
1069 ff_h264_reset_sei(h);
1070 if (avctx->codec_id == AV_CODEC_ID_H264) {
1071 if (avctx->ticks_per_frame == 1)
1072 s->avctx->time_base.den *= 2;
1073 avctx->ticks_per_frame = 2;
1076 if (avctx->extradata_size > 0 && avctx->extradata &&
1077 ff_h264_decode_extradata(h))
1080 if (h->sps.bitstream_restriction_flag &&
1081 s->avctx->has_b_frames < h->sps.num_reorder_frames) {
1082 s->avctx->has_b_frames = h->sps.num_reorder_frames;
1089 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b) + (size))))
1091 static void copy_picture_range(Picture **to, Picture **from, int count,
1092 MpegEncContext *new_base,
1093 MpegEncContext *old_base)
1097 for (i = 0; i < count; i++) {
1098 assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1099 IN_RANGE(from[i], old_base->picture,
1100 sizeof(Picture) * old_base->picture_count) ||
1102 to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1106 static void copy_parameter_set(void **to, void **from, int count, int size)
1110 for (i = 0; i < count; i++) {
1111 if (to[i] && !from[i])
1113 else if (from[i] && !to[i])
1114 to[i] = av_malloc(size);
1117 memcpy(to[i], from[i], size);
1121 static int decode_init_thread_copy(AVCodecContext *avctx)
1123 H264Context *h = avctx->priv_data;
1125 if (!avctx->internal->is_copy)
1127 memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1128 memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1133 #define copy_fields(to, from, start_field, end_field) \
1134 memcpy(&to->start_field, &from->start_field, \
1135 (char *)&to->end_field - (char *)&to->start_field)
1137 static int decode_update_thread_context(AVCodecContext *dst,
1138 const AVCodecContext *src)
1140 H264Context *h = dst->priv_data, *h1 = src->priv_data;
1141 MpegEncContext *const s = &h->s, *const s1 = &h1->s;
1142 int inited = s->context_initialized, err;
1145 if (dst == src || !s1->context_initialized)
1148 err = ff_mpeg_update_thread_context(dst, src);
1152 // FIXME handle width/height changing
1154 for (i = 0; i < MAX_SPS_COUNT; i++)
1155 av_freep(h->sps_buffers + i);
1157 for (i = 0; i < MAX_PPS_COUNT; i++)
1158 av_freep(h->pps_buffers + i);
1160 // copy all fields after MpegEnc
1161 memcpy(&h->s + 1, &h1->s + 1,
1162 sizeof(H264Context) - sizeof(MpegEncContext));
1163 memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1164 memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1165 if (ff_h264_alloc_tables(h) < 0) {
1166 av_log(dst, AV_LOG_ERROR, "Could not allocate memory for h264\n");
1167 return AVERROR(ENOMEM);
1171 for (i = 0; i < 2; i++) {
1172 h->rbsp_buffer[i] = NULL;
1173 h->rbsp_buffer_size[i] = 0;
1176 h->thread_context[0] = h;
1178 /* frame_start may not be called for the next thread (if it's decoding
1179 * a bottom field) so this has to be allocated here */
1180 h->s.obmc_scratchpad = av_malloc(16 * 6 * s->linesize);
1182 s->dsp.clear_blocks(h->mb);
1183 s->dsp.clear_blocks(h->mb + (24 * 16 << h->pixel_shift));
1186 // extradata/NAL handling
1187 h->is_avc = h1->is_avc;
1190 copy_parameter_set((void **)h->sps_buffers, (void **)h1->sps_buffers,
1191 MAX_SPS_COUNT, sizeof(SPS));
1193 copy_parameter_set((void **)h->pps_buffers, (void **)h1->pps_buffers,
1194 MAX_PPS_COUNT, sizeof(PPS));
1197 // Dequantization matrices
1198 // FIXME these are big - can they be only copied when PPS changes?
1199 copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1201 for (i = 0; i < 6; i++)
1202 h->dequant4_coeff[i] = h->dequant4_buffer[0] +
1203 (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1205 for (i = 0; i < 6; i++)
1206 h->dequant8_coeff[i] = h->dequant8_buffer[0] +
1207 (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1209 h->dequant_coeff_pps = h1->dequant_coeff_pps;
1212 copy_fields(h, h1, poc_lsb, redundant_pic_count);
1215 copy_fields(h, h1, ref_count, list_count);
1216 copy_fields(h, h1, ref_list, intra_gb);
1217 copy_fields(h, h1, short_ref, cabac_init_idc);
1219 copy_picture_range(h->short_ref, h1->short_ref, 32, s, s1);
1220 copy_picture_range(h->long_ref, h1->long_ref, 32, s, s1);
1221 copy_picture_range(h->delayed_pic, h1->delayed_pic,
1222 MAX_DELAYED_PIC_COUNT + 2, s, s1);
1224 h->last_slice_type = h1->last_slice_type;
1226 if (!s->current_picture_ptr)
1230 err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1231 h->prev_poc_msb = h->poc_msb;
1232 h->prev_poc_lsb = h->poc_lsb;
1234 h->prev_frame_num_offset = h->frame_num_offset;
1235 h->prev_frame_num = h->frame_num;
1236 h->outputed_poc = h->next_outputed_poc;
1241 int ff_h264_frame_start(H264Context *h)
1243 MpegEncContext *const s = &h->s;
1245 const int pixel_shift = h->pixel_shift;
1247 if (ff_MPV_frame_start(s, s->avctx) < 0)
1249 ff_er_frame_start(s);
1251 * ff_MPV_frame_start uses pict_type to derive key_frame.
1252 * This is incorrect for H.264; IDR markings must be used.
1253 * Zero here; IDR markings per slice in frame or fields are ORed in later.
1254 * See decode_nal_units().
1256 s->current_picture_ptr->f.key_frame = 0;
1257 s->current_picture_ptr->mmco_reset = 0;
1259 assert(s->linesize && s->uvlinesize);
1261 for (i = 0; i < 16; i++) {
1262 h->block_offset[i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * s->linesize * ((scan8[i] - scan8[0]) >> 3);
1263 h->block_offset[48 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * s->linesize * ((scan8[i] - scan8[0]) >> 3);
1265 for (i = 0; i < 16; i++) {
1266 h->block_offset[16 + i] =
1267 h->block_offset[32 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * s->uvlinesize * ((scan8[i] - scan8[0]) >> 3);
1268 h->block_offset[48 + 16 + i] =
1269 h->block_offset[48 + 32 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * s->uvlinesize * ((scan8[i] - scan8[0]) >> 3);
1272 /* can't be in alloc_tables because linesize isn't known there.
1273 * FIXME: redo bipred weight to not require extra buffer? */
1274 for (i = 0; i < s->slice_context_count; i++)
1275 if (h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1276 h->thread_context[i]->s.obmc_scratchpad = av_malloc(16 * 6 * s->linesize);
1278 /* Some macroblocks can be accessed before they're available in case
1279 * of lost slices, MBAFF or threading. */
1280 memset(h->slice_table, -1,
1281 (s->mb_height * s->mb_stride - 1) * sizeof(*h->slice_table));
1283 // s->decode = (s->flags & CODEC_FLAG_PSNR) || !s->encoding ||
1284 // s->current_picture.f.reference /* || h->contains_intra */ || 1;
1286 /* We mark the current picture as non-reference after allocating it, so
1287 * that if we break out due to an error it can be released automatically
1288 * in the next ff_MPV_frame_start().
1289 * SVQ3 as well as most other codecs have only last/next/current and thus
1290 * get released even with set reference, besides SVQ3 and others do not
1291 * mark frames as reference later "naturally". */
1292 if (s->codec_id != AV_CODEC_ID_SVQ3)
1293 s->current_picture_ptr->f.reference = 0;
1295 s->current_picture_ptr->field_poc[0] =
1296 s->current_picture_ptr->field_poc[1] = INT_MAX;
1298 h->next_output_pic = NULL;
1300 assert(s->current_picture_ptr->long_ref == 0);
1306 * Run setup operations that must be run after slice header decoding.
1307 * This includes finding the next displayed frame.
1309 * @param h h264 master context
1310 * @param setup_finished enough NALs have been read that we can call
1311 * ff_thread_finish_setup()
1313 static void decode_postinit(H264Context *h, int setup_finished)
1315 MpegEncContext *const s = &h->s;
1316 Picture *out = s->current_picture_ptr;
1317 Picture *cur = s->current_picture_ptr;
1318 int i, pics, out_of_order, out_idx;
1319 int invalid = 0, cnt = 0;
1321 s->current_picture_ptr->f.qscale_type = FF_QSCALE_TYPE_H264;
1322 s->current_picture_ptr->f.pict_type = s->pict_type;
1324 if (h->next_output_pic)
1327 if (cur->field_poc[0] == INT_MAX || cur->field_poc[1] == INT_MAX) {
1328 /* FIXME: if we have two PAFF fields in one packet, we can't start
1329 * the next thread here. If we have one field per packet, we can.
1330 * The check in decode_nal_units() is not good enough to find this
1331 * yet, so we assume the worst for now. */
1332 // if (setup_finished)
1333 // ff_thread_finish_setup(s->avctx);
1337 cur->f.interlaced_frame = 0;
1338 cur->f.repeat_pict = 0;
1340 /* Signal interlacing information externally. */
1341 /* Prioritize picture timing SEI information over used
1342 * decoding process if it exists. */
1344 if (h->sps.pic_struct_present_flag) {
1345 switch (h->sei_pic_struct) {
1346 case SEI_PIC_STRUCT_FRAME:
1348 case SEI_PIC_STRUCT_TOP_FIELD:
1349 case SEI_PIC_STRUCT_BOTTOM_FIELD:
1350 cur->f.interlaced_frame = 1;
1352 case SEI_PIC_STRUCT_TOP_BOTTOM:
1353 case SEI_PIC_STRUCT_BOTTOM_TOP:
1354 if (FIELD_OR_MBAFF_PICTURE)
1355 cur->f.interlaced_frame = 1;
1357 // try to flag soft telecine progressive
1358 cur->f.interlaced_frame = h->prev_interlaced_frame;
1360 case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1361 case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1362 /* Signal the possibility of telecined film externally
1363 * (pic_struct 5,6). From these hints, let the applications
1364 * decide if they apply deinterlacing. */
1365 cur->f.repeat_pict = 1;
1367 case SEI_PIC_STRUCT_FRAME_DOUBLING:
1368 // Force progressive here, doubling interlaced frame is a bad idea.
1369 cur->f.repeat_pict = 2;
1371 case SEI_PIC_STRUCT_FRAME_TRIPLING:
1372 cur->f.repeat_pict = 4;
1376 if ((h->sei_ct_type & 3) &&
1377 h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1378 cur->f.interlaced_frame = (h->sei_ct_type & (1 << 1)) != 0;
1380 /* Derive interlacing flag from used decoding process. */
1381 cur->f.interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1383 h->prev_interlaced_frame = cur->f.interlaced_frame;
1385 if (cur->field_poc[0] != cur->field_poc[1]) {
1386 /* Derive top_field_first from field pocs. */
1387 cur->f.top_field_first = cur->field_poc[0] < cur->field_poc[1];
1389 if (cur->f.interlaced_frame || h->sps.pic_struct_present_flag) {
1390 /* Use picture timing SEI information. Even if it is a
1391 * information of a past frame, better than nothing. */
1392 if (h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM ||
1393 h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1394 cur->f.top_field_first = 1;
1396 cur->f.top_field_first = 0;
1398 /* Most likely progressive */
1399 cur->f.top_field_first = 0;
1403 // FIXME do something with unavailable reference frames
1405 /* Sort B-frames into display order */
1407 if (h->sps.bitstream_restriction_flag &&
1408 s->avctx->has_b_frames < h->sps.num_reorder_frames) {
1409 s->avctx->has_b_frames = h->sps.num_reorder_frames;
1413 if (s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT &&
1414 !h->sps.bitstream_restriction_flag) {
1415 s->avctx->has_b_frames = MAX_DELAYED_PIC_COUNT - 1;
1420 while (h->delayed_pic[pics])
1423 assert(pics <= MAX_DELAYED_PIC_COUNT);
1425 h->delayed_pic[pics++] = cur;
1426 if (cur->f.reference == 0)
1427 cur->f.reference = DELAYED_PIC_REF;
1429 /* Frame reordering. This code takes pictures from coding order and sorts
1430 * them by their incremental POC value into display order. It supports POC
1431 * gaps, MMCO reset codes and random resets.
1432 * A "display group" can start either with a IDR frame (f.key_frame = 1),
1433 * and/or can be closed down with a MMCO reset code. In sequences where
1434 * there is no delay, we can't detect that (since the frame was already
1435 * output to the user), so we also set h->mmco_reset to detect the MMCO
1437 * FIXME: if we detect insufficient delays (as per s->avctx->has_b_frames),
1438 * we increase the delay between input and output. All frames affected by
1439 * the lag (e.g. those that should have been output before another frame
1440 * that we already returned to the user) will be dropped. This is a bug
1441 * that we will fix later. */
1442 for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++) {
1443 cnt += out->poc < h->last_pocs[i];
1444 invalid += out->poc == INT_MIN;
1446 if (!h->mmco_reset && !cur->f.key_frame &&
1447 cnt + invalid == MAX_DELAYED_PIC_COUNT && cnt > 0) {
1450 h->delayed_pic[pics - 2]->mmco_reset = 2;
1452 if (h->mmco_reset || cur->f.key_frame) {
1453 for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
1454 h->last_pocs[i] = INT_MIN;
1456 invalid = MAX_DELAYED_PIC_COUNT;
1458 out = h->delayed_pic[0];
1460 for (i = 1; i < MAX_DELAYED_PIC_COUNT &&
1461 h->delayed_pic[i] &&
1462 !h->delayed_pic[i - 1]->mmco_reset &&
1463 !h->delayed_pic[i]->f.key_frame;
1465 if (h->delayed_pic[i]->poc < out->poc) {
1466 out = h->delayed_pic[i];
1469 if (s->avctx->has_b_frames == 0 &&
1470 (h->delayed_pic[0]->f.key_frame || h->mmco_reset))
1471 h->next_outputed_poc = INT_MIN;
1472 out_of_order = !out->f.key_frame && !h->mmco_reset &&
1473 (out->poc < h->next_outputed_poc);
1475 if (h->sps.bitstream_restriction_flag &&
1476 s->avctx->has_b_frames >= h->sps.num_reorder_frames) {
1477 } else if (out_of_order && pics - 1 == s->avctx->has_b_frames &&
1478 s->avctx->has_b_frames < MAX_DELAYED_PIC_COUNT) {
1479 if (invalid + cnt < MAX_DELAYED_PIC_COUNT) {
1480 s->avctx->has_b_frames = FFMAX(s->avctx->has_b_frames, cnt);
1483 } else if (s->low_delay &&
1484 ((h->next_outputed_poc != INT_MIN &&
1485 out->poc > h->next_outputed_poc + 2) ||
1486 cur->f.pict_type == AV_PICTURE_TYPE_B)) {
1488 s->avctx->has_b_frames++;
1491 if (pics > s->avctx->has_b_frames) {
1492 out->f.reference &= ~DELAYED_PIC_REF;
1493 // for frame threading, the owner must be the second field's thread or
1494 // else the first thread can release the picture and reuse it unsafely
1496 for (i = out_idx; h->delayed_pic[i]; i++)
1497 h->delayed_pic[i] = h->delayed_pic[i + 1];
1499 memmove(h->last_pocs, &h->last_pocs[1],
1500 sizeof(*h->last_pocs) * (MAX_DELAYED_PIC_COUNT - 1));
1501 h->last_pocs[MAX_DELAYED_PIC_COUNT - 1] = cur->poc;
1502 if (!out_of_order && pics > s->avctx->has_b_frames) {
1503 h->next_output_pic = out;
1504 if (out->mmco_reset) {
1506 h->next_outputed_poc = out->poc;
1507 h->delayed_pic[out_idx - 1]->mmco_reset = out->mmco_reset;
1509 h->next_outputed_poc = INT_MIN;
1512 if (out_idx == 0 && pics > 1 && h->delayed_pic[0]->f.key_frame) {
1513 h->next_outputed_poc = INT_MIN;
1515 h->next_outputed_poc = out->poc;
1520 av_log(s->avctx, AV_LOG_DEBUG, "no picture\n");
1524 ff_thread_finish_setup(s->avctx);
1527 static av_always_inline void backup_mb_border(H264Context *h, uint8_t *src_y,
1528 uint8_t *src_cb, uint8_t *src_cr,
1529 int linesize, int uvlinesize,
1532 MpegEncContext *const s = &h->s;
1533 uint8_t *top_border;
1535 const int pixel_shift = h->pixel_shift;
1536 int chroma444 = CHROMA444;
1537 int chroma422 = CHROMA422;
1540 src_cb -= uvlinesize;
1541 src_cr -= uvlinesize;
1543 if (!simple && FRAME_MBAFF) {
1546 top_border = h->top_borders[0][s->mb_x];
1547 AV_COPY128(top_border, src_y + 15 * linesize);
1549 AV_COPY128(top_border + 16, src_y + 15 * linesize + 16);
1550 if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1553 AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
1554 AV_COPY128(top_border + 48, src_cb + 15 * uvlinesize + 16);
1555 AV_COPY128(top_border + 64, src_cr + 15 * uvlinesize);
1556 AV_COPY128(top_border + 80, src_cr + 15 * uvlinesize + 16);
1558 AV_COPY128(top_border + 16, src_cb + 15 * uvlinesize);
1559 AV_COPY128(top_border + 32, src_cr + 15 * uvlinesize);
1561 } else if (chroma422) {
1563 AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
1564 AV_COPY128(top_border + 48, src_cr + 15 * uvlinesize);
1566 AV_COPY64(top_border + 16, src_cb + 15 * uvlinesize);
1567 AV_COPY64(top_border + 24, src_cr + 15 * uvlinesize);
1571 AV_COPY128(top_border + 32, src_cb + 7 * uvlinesize);
1572 AV_COPY128(top_border + 48, src_cr + 7 * uvlinesize);
1574 AV_COPY64(top_border + 16, src_cb + 7 * uvlinesize);
1575 AV_COPY64(top_border + 24, src_cr + 7 * uvlinesize);
1580 } else if (MB_MBAFF) {
1586 top_border = h->top_borders[top_idx][s->mb_x];
1587 /* There are two lines saved, the line above the top macroblock
1588 * of a pair, and the line above the bottom macroblock. */
1589 AV_COPY128(top_border, src_y + 16 * linesize);
1591 AV_COPY128(top_border + 16, src_y + 16 * linesize + 16);
1593 if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1596 AV_COPY128(top_border + 32, src_cb + 16 * linesize);
1597 AV_COPY128(top_border + 48, src_cb + 16 * linesize + 16);
1598 AV_COPY128(top_border + 64, src_cr + 16 * linesize);
1599 AV_COPY128(top_border + 80, src_cr + 16 * linesize + 16);
1601 AV_COPY128(top_border + 16, src_cb + 16 * linesize);
1602 AV_COPY128(top_border + 32, src_cr + 16 * linesize);
1604 } else if (chroma422) {
1606 AV_COPY128(top_border + 32, src_cb + 16 * uvlinesize);
1607 AV_COPY128(top_border + 48, src_cr + 16 * uvlinesize);
1609 AV_COPY64(top_border + 16, src_cb + 16 * uvlinesize);
1610 AV_COPY64(top_border + 24, src_cr + 16 * uvlinesize);
1614 AV_COPY128(top_border + 32, src_cb + 8 * uvlinesize);
1615 AV_COPY128(top_border + 48, src_cr + 8 * uvlinesize);
1617 AV_COPY64(top_border + 16, src_cb + 8 * uvlinesize);
1618 AV_COPY64(top_border + 24, src_cr + 8 * uvlinesize);
1624 static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1625 uint8_t *src_cb, uint8_t *src_cr,
1626 int linesize, int uvlinesize,
1627 int xchg, int chroma444,
1628 int simple, int pixel_shift)
1630 MpegEncContext *const s = &h->s;
1631 int deblock_topleft;
1634 uint8_t *top_border_m1;
1635 uint8_t *top_border;
1637 if (!simple && FRAME_MBAFF) {
1642 top_idx = MB_MBAFF ? 0 : 1;
1646 if (h->deblocking_filter == 2) {
1647 deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1648 deblock_top = h->top_type;
1650 deblock_topleft = (s->mb_x > 0);
1651 deblock_top = (s->mb_y > !!MB_FIELD);
1654 src_y -= linesize + 1 + pixel_shift;
1655 src_cb -= uvlinesize + 1 + pixel_shift;
1656 src_cr -= uvlinesize + 1 + pixel_shift;
1658 top_border_m1 = h->top_borders[top_idx][s->mb_x - 1];
1659 top_border = h->top_borders[top_idx][s->mb_x];
1661 #define XCHG(a, b, xchg) \
1662 if (pixel_shift) { \
1664 AV_SWAP64(b + 0, a + 0); \
1665 AV_SWAP64(b + 8, a + 8); \
1675 if (deblock_topleft) {
1676 XCHG(top_border_m1 + (8 << pixel_shift),
1677 src_y - (7 << pixel_shift), 1);
1679 XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1680 XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1681 if (s->mb_x + 1 < s->mb_width) {
1682 XCHG(h->top_borders[top_idx][s->mb_x + 1],
1683 src_y + (17 << pixel_shift), 1);
1686 if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1688 if (deblock_topleft) {
1689 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1690 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1692 XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1693 XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1694 XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1695 XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1696 if (s->mb_x + 1 < s->mb_width) {
1697 XCHG(h->top_borders[top_idx][s->mb_x + 1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1698 XCHG(h->top_borders[top_idx][s->mb_x + 1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1702 if (deblock_topleft) {
1703 XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1704 XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1706 XCHG(top_border + (16 << pixel_shift), src_cb + 1 + pixel_shift, 1);
1707 XCHG(top_border + (24 << pixel_shift), src_cr + 1 + pixel_shift, 1);
1713 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth,
1716 if (high_bit_depth) {
1717 return AV_RN32A(((int32_t *)mb) + index);
1719 return AV_RN16A(mb + index);
1722 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth,
1723 int index, int value)
1725 if (high_bit_depth) {
1726 AV_WN32A(((int32_t *)mb) + index, value);
1728 AV_WN16A(mb + index, value);
1731 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h,
1732 int mb_type, int is_h264,
1734 int transform_bypass,
1738 uint8_t *dest_y, int p)
1740 MpegEncContext *const s = &h->s;
1741 void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1742 void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1744 int qscale = p == 0 ? s->qscale : h->chroma_qp[p - 1];
1745 block_offset += 16 * p;
1746 if (IS_INTRA4x4(mb_type)) {
1747 if (simple || !s->encoding) {
1748 if (IS_8x8DCT(mb_type)) {
1749 if (transform_bypass) {
1751 idct_add = s->dsp.add_pixels8;
1753 idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1754 idct_add = h->h264dsp.h264_idct8_add;
1756 for (i = 0; i < 16; i += 4) {
1757 uint8_t *const ptr = dest_y + block_offset[i];
1758 const int dir = h->intra4x4_pred_mode_cache[scan8[i]];
1759 if (transform_bypass && h->sps.profile_idc == 244 && dir <= 1) {
1760 h->hpc.pred8x8l_add[dir](ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1762 const int nnz = h->non_zero_count_cache[scan8[i + p * 16]];
1763 h->hpc.pred8x8l[dir](ptr, (h->topleft_samples_available << i) & 0x8000,
1764 (h->topright_samples_available << i) & 0x4000, linesize);
1766 if (nnz == 1 && dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
1767 idct_dc_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1769 idct_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1774 if (transform_bypass) {
1776 idct_add = s->dsp.add_pixels4;
1778 idct_dc_add = h->h264dsp.h264_idct_dc_add;
1779 idct_add = h->h264dsp.h264_idct_add;
1781 for (i = 0; i < 16; i++) {
1782 uint8_t *const ptr = dest_y + block_offset[i];
1783 const int dir = h->intra4x4_pred_mode_cache[scan8[i]];
1785 if (transform_bypass && h->sps.profile_idc == 244 && dir <= 1) {
1786 h->hpc.pred4x4_add[dir](ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1791 if (dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED) {
1792 const int topright_avail = (h->topright_samples_available << i) & 0x8000;
1793 assert(s->mb_y || linesize <= block_offset[i]);
1794 if (!topright_avail) {
1796 tr_high = ((uint16_t *)ptr)[3 - linesize / 2] * 0x0001000100010001ULL;
1797 topright = (uint8_t *)&tr_high;
1799 tr = ptr[3 - linesize] * 0x01010101u;
1800 topright = (uint8_t *)&tr;
1803 topright = ptr + (4 << pixel_shift) - linesize;
1807 h->hpc.pred4x4[dir](ptr, topright, linesize);
1808 nnz = h->non_zero_count_cache[scan8[i + p * 16]];
1811 if (nnz == 1 && dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
1812 idct_dc_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1814 idct_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1815 } else if (CONFIG_SVQ3_DECODER)
1816 ff_svq3_add_idct_c(ptr, h->mb + i * 16 + p * 256, linesize, qscale, 0);
1823 h->hpc.pred16x16[h->intra16x16_pred_mode](dest_y, linesize);
1825 if (h->non_zero_count_cache[scan8[LUMA_DC_BLOCK_INDEX + p]]) {
1826 if (!transform_bypass)
1827 h->h264dsp.h264_luma_dc_dequant_idct(h->mb + (p * 256 << pixel_shift),
1829 h->dequant4_coeff[p][qscale][0]);
1831 static const uint8_t dc_mapping[16] = {
1832 0 * 16, 1 * 16, 4 * 16, 5 * 16,
1833 2 * 16, 3 * 16, 6 * 16, 7 * 16,
1834 8 * 16, 9 * 16, 12 * 16, 13 * 16,
1835 10 * 16, 11 * 16, 14 * 16, 15 * 16 };
1836 for (i = 0; i < 16; i++)
1837 dctcoef_set(h->mb + (p * 256 << pixel_shift),
1838 pixel_shift, dc_mapping[i],
1839 dctcoef_get(h->mb_luma_dc[p],
1843 } else if (CONFIG_SVQ3_DECODER)
1844 ff_svq3_luma_dc_dequant_idct_c(h->mb + p * 256,
1845 h->mb_luma_dc[p], qscale);
1849 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type,
1850 int is_h264, int simple,
1851 int transform_bypass,
1855 uint8_t *dest_y, int p)
1857 MpegEncContext *const s = &h->s;
1858 void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1860 block_offset += 16 * p;
1861 if (!IS_INTRA4x4(mb_type)) {
1863 if (IS_INTRA16x16(mb_type)) {
1864 if (transform_bypass) {
1865 if (h->sps.profile_idc == 244 &&
1866 (h->intra16x16_pred_mode == VERT_PRED8x8 ||
1867 h->intra16x16_pred_mode == HOR_PRED8x8)) {
1868 h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset,
1869 h->mb + (p * 256 << pixel_shift),
1872 for (i = 0; i < 16; i++)
1873 if (h->non_zero_count_cache[scan8[i + p * 16]] ||
1874 dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
1875 s->dsp.add_pixels4(dest_y + block_offset[i],
1876 h->mb + (i * 16 + p * 256 << pixel_shift),
1880 h->h264dsp.h264_idct_add16intra(dest_y, block_offset,
1881 h->mb + (p * 256 << pixel_shift),
1883 h->non_zero_count_cache + p * 5 * 8);
1885 } else if (h->cbp & 15) {
1886 if (transform_bypass) {
1887 const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1888 idct_add = IS_8x8DCT(mb_type) ? s->dsp.add_pixels8
1889 : s->dsp.add_pixels4;
1890 for (i = 0; i < 16; i += di)
1891 if (h->non_zero_count_cache[scan8[i + p * 16]])
1892 idct_add(dest_y + block_offset[i],
1893 h->mb + (i * 16 + p * 256 << pixel_shift),
1896 if (IS_8x8DCT(mb_type))
1897 h->h264dsp.h264_idct8_add4(dest_y, block_offset,
1898 h->mb + (p * 256 << pixel_shift),
1900 h->non_zero_count_cache + p * 5 * 8);
1902 h->h264dsp.h264_idct_add16(dest_y, block_offset,
1903 h->mb + (p * 256 << pixel_shift),
1905 h->non_zero_count_cache + p * 5 * 8);
1908 } else if (CONFIG_SVQ3_DECODER) {
1909 for (i = 0; i < 16; i++)
1910 if (h->non_zero_count_cache[scan8[i + p * 16]] || h->mb[i * 16 + p * 256]) {
1911 // FIXME benchmark weird rule, & below
1912 uint8_t *const ptr = dest_y + block_offset[i];
1913 ff_svq3_add_idct_c(ptr, h->mb + i * 16 + p * 256, linesize,
1914 s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1922 #include "h264_mb_template.c"
1926 #include "h264_mb_template.c"
1930 #include "h264_mb_template.c"
1932 void ff_h264_hl_decode_mb(H264Context *h)
1934 MpegEncContext *const s = &h->s;
1935 const int mb_xy = h->mb_xy;
1936 const int mb_type = s->current_picture.f.mb_type[mb_xy];
1937 int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
1940 if (is_complex || h->pixel_shift)
1941 hl_decode_mb_444_complex(h);
1943 hl_decode_mb_444_simple_8(h);
1944 } else if (is_complex) {
1945 hl_decode_mb_complex(h);
1946 } else if (h->pixel_shift) {
1947 hl_decode_mb_simple_16(h);
1949 hl_decode_mb_simple_8(h);
1952 static int pred_weight_table(H264Context *h)
1954 MpegEncContext *const s = &h->s;
1956 int luma_def, chroma_def;
1959 h->use_weight_chroma = 0;
1960 h->luma_log2_weight_denom = get_ue_golomb(&s->gb);
1961 if (h->sps.chroma_format_idc)
1962 h->chroma_log2_weight_denom = get_ue_golomb(&s->gb);
1963 luma_def = 1 << h->luma_log2_weight_denom;
1964 chroma_def = 1 << h->chroma_log2_weight_denom;
1966 for (list = 0; list < 2; list++) {
1967 h->luma_weight_flag[list] = 0;
1968 h->chroma_weight_flag[list] = 0;
1969 for (i = 0; i < h->ref_count[list]; i++) {
1970 int luma_weight_flag, chroma_weight_flag;
1972 luma_weight_flag = get_bits1(&s->gb);
1973 if (luma_weight_flag) {
1974 h->luma_weight[i][list][0] = get_se_golomb(&s->gb);
1975 h->luma_weight[i][list][1] = get_se_golomb(&s->gb);
1976 if (h->luma_weight[i][list][0] != luma_def ||
1977 h->luma_weight[i][list][1] != 0) {
1979 h->luma_weight_flag[list] = 1;
1982 h->luma_weight[i][list][0] = luma_def;
1983 h->luma_weight[i][list][1] = 0;
1986 if (h->sps.chroma_format_idc) {
1987 chroma_weight_flag = get_bits1(&s->gb);
1988 if (chroma_weight_flag) {
1990 for (j = 0; j < 2; j++) {
1991 h->chroma_weight[i][list][j][0] = get_se_golomb(&s->gb);
1992 h->chroma_weight[i][list][j][1] = get_se_golomb(&s->gb);
1993 if (h->chroma_weight[i][list][j][0] != chroma_def ||
1994 h->chroma_weight[i][list][j][1] != 0) {
1995 h->use_weight_chroma = 1;
1996 h->chroma_weight_flag[list] = 1;
2001 for (j = 0; j < 2; j++) {
2002 h->chroma_weight[i][list][j][0] = chroma_def;
2003 h->chroma_weight[i][list][j][1] = 0;
2008 if (h->slice_type_nos != AV_PICTURE_TYPE_B)
2011 h->use_weight = h->use_weight || h->use_weight_chroma;
2016 * Initialize implicit_weight table.
2017 * @param field 0/1 initialize the weight for interlaced MBAFF
2018 * -1 initializes the rest
2020 static void implicit_weight_table(H264Context *h, int field)
2022 MpegEncContext *const s = &h->s;
2023 int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2025 for (i = 0; i < 2; i++) {
2026 h->luma_weight_flag[i] = 0;
2027 h->chroma_weight_flag[i] = 0;
2031 if (s->picture_structure == PICT_FRAME) {
2032 cur_poc = s->current_picture_ptr->poc;
2034 cur_poc = s->current_picture_ptr->field_poc[s->picture_structure - 1];
2036 if (h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF &&
2037 h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2 * cur_poc) {
2039 h->use_weight_chroma = 0;
2043 ref_count0 = h->ref_count[0];
2044 ref_count1 = h->ref_count[1];
2046 cur_poc = s->current_picture_ptr->field_poc[field];
2048 ref_count0 = 16 + 2 * h->ref_count[0];
2049 ref_count1 = 16 + 2 * h->ref_count[1];
2053 h->use_weight_chroma = 2;
2054 h->luma_log2_weight_denom = 5;
2055 h->chroma_log2_weight_denom = 5;
2057 for (ref0 = ref_start; ref0 < ref_count0; ref0++) {
2058 int poc0 = h->ref_list[0][ref0].poc;
2059 for (ref1 = ref_start; ref1 < ref_count1; ref1++) {
2061 if (!h->ref_list[0][ref0].long_ref && !h->ref_list[1][ref1].long_ref) {
2062 int poc1 = h->ref_list[1][ref1].poc;
2063 int td = av_clip(poc1 - poc0, -128, 127);
2065 int tb = av_clip(cur_poc - poc0, -128, 127);
2066 int tx = (16384 + (FFABS(td) >> 1)) / td;
2067 int dist_scale_factor = (tb * tx + 32) >> 8;
2068 if (dist_scale_factor >= -64 && dist_scale_factor <= 128)
2069 w = 64 - dist_scale_factor;
2073 h->implicit_weight[ref0][ref1][0] =
2074 h->implicit_weight[ref0][ref1][1] = w;
2076 h->implicit_weight[ref0][ref1][field] = w;
2083 * instantaneous decoder refresh.
2085 static void idr(H264Context *h)
2087 ff_h264_remove_all_refs(h);
2088 h->prev_frame_num = 0;
2089 h->prev_frame_num_offset = 0;
2091 h->prev_poc_lsb = 0;
2094 /* forget old pics after a seek */
2095 static void flush_dpb(AVCodecContext *avctx)
2097 H264Context *h = avctx->priv_data;
2099 for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++) {
2100 if (h->delayed_pic[i])
2101 h->delayed_pic[i]->f.reference = 0;
2102 h->delayed_pic[i] = NULL;
2104 for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
2105 h->last_pocs[i] = INT_MIN;
2106 h->outputed_poc = h->next_outputed_poc = INT_MIN;
2107 h->prev_interlaced_frame = 1;
2109 if (h->s.current_picture_ptr)
2110 h->s.current_picture_ptr->f.reference = 0;
2111 h->s.first_field = 0;
2112 ff_h264_reset_sei(h);
2113 ff_mpeg_flush(avctx);
2116 static int init_poc(H264Context *h)
2118 MpegEncContext *const s = &h->s;
2119 const int max_frame_num = 1 << h->sps.log2_max_frame_num;
2121 Picture *cur = s->current_picture_ptr;
2123 h->frame_num_offset = h->prev_frame_num_offset;
2124 if (h->frame_num < h->prev_frame_num)
2125 h->frame_num_offset += max_frame_num;
2127 if (h->sps.poc_type == 0) {
2128 const int max_poc_lsb = 1 << h->sps.log2_max_poc_lsb;
2130 if (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb / 2)
2131 h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2132 else if (h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb / 2)
2133 h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2135 h->poc_msb = h->prev_poc_msb;
2137 field_poc[1] = h->poc_msb + h->poc_lsb;
2138 if (s->picture_structure == PICT_FRAME)
2139 field_poc[1] += h->delta_poc_bottom;
2140 } else if (h->sps.poc_type == 1) {
2141 int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2144 if (h->sps.poc_cycle_length != 0)
2145 abs_frame_num = h->frame_num_offset + h->frame_num;
2149 if (h->nal_ref_idc == 0 && abs_frame_num > 0)
2152 expected_delta_per_poc_cycle = 0;
2153 for (i = 0; i < h->sps.poc_cycle_length; i++)
2154 // FIXME integrate during sps parse
2155 expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[i];
2157 if (abs_frame_num > 0) {
2158 int poc_cycle_cnt = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2159 int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2161 expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2162 for (i = 0; i <= frame_num_in_poc_cycle; i++)
2163 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[i];
2167 if (h->nal_ref_idc == 0)
2168 expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2170 field_poc[0] = expectedpoc + h->delta_poc[0];
2171 field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2173 if (s->picture_structure == PICT_FRAME)
2174 field_poc[1] += h->delta_poc[1];
2176 int poc = 2 * (h->frame_num_offset + h->frame_num);
2178 if (!h->nal_ref_idc)
2185 if (s->picture_structure != PICT_BOTTOM_FIELD)
2186 s->current_picture_ptr->field_poc[0] = field_poc[0];
2187 if (s->picture_structure != PICT_TOP_FIELD)
2188 s->current_picture_ptr->field_poc[1] = field_poc[1];
2189 cur->poc = FFMIN(cur->field_poc[0], cur->field_poc[1]);
2195 * initialize scan tables
2197 static void init_scan_tables(H264Context *h)
2200 for (i = 0; i < 16; i++) {
2201 #define T(x) (x >> 2) | ((x << 2) & 0xF)
2202 h->zigzag_scan[i] = T(zigzag_scan[i]);
2203 h->field_scan[i] = T(field_scan[i]);
2206 for (i = 0; i < 64; i++) {
2207 #define T(x) (x >> 3) | ((x & 7) << 3)
2208 h->zigzag_scan8x8[i] = T(ff_zigzag_direct[i]);
2209 h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2210 h->field_scan8x8[i] = T(field_scan8x8[i]);
2211 h->field_scan8x8_cavlc[i] = T(field_scan8x8_cavlc[i]);
2214 if (h->sps.transform_bypass) { // FIXME same ugly
2215 h->zigzag_scan_q0 = zigzag_scan;
2216 h->zigzag_scan8x8_q0 = ff_zigzag_direct;
2217 h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2218 h->field_scan_q0 = field_scan;
2219 h->field_scan8x8_q0 = field_scan8x8;
2220 h->field_scan8x8_cavlc_q0 = field_scan8x8_cavlc;
2222 h->zigzag_scan_q0 = h->zigzag_scan;
2223 h->zigzag_scan8x8_q0 = h->zigzag_scan8x8;
2224 h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2225 h->field_scan_q0 = h->field_scan;
2226 h->field_scan8x8_q0 = h->field_scan8x8;
2227 h->field_scan8x8_cavlc_q0 = h->field_scan8x8_cavlc;
2231 static int field_end(H264Context *h, int in_setup)
2233 MpegEncContext *const s = &h->s;
2234 AVCodecContext *const avctx = s->avctx;
2238 if (!in_setup && !s->dropable)
2239 ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX,
2240 s->picture_structure == PICT_BOTTOM_FIELD);
2242 if (CONFIG_H264_VDPAU_DECODER &&
2243 s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
2244 ff_vdpau_h264_set_reference_frames(s);
2246 if (in_setup || !(avctx->active_thread_type & FF_THREAD_FRAME)) {
2248 err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2249 h->prev_poc_msb = h->poc_msb;
2250 h->prev_poc_lsb = h->poc_lsb;
2252 h->prev_frame_num_offset = h->frame_num_offset;
2253 h->prev_frame_num = h->frame_num;
2254 h->outputed_poc = h->next_outputed_poc;
2257 if (avctx->hwaccel) {
2258 if (avctx->hwaccel->end_frame(avctx) < 0)
2259 av_log(avctx, AV_LOG_ERROR,
2260 "hardware accelerator failed to decode picture\n");
2263 if (CONFIG_H264_VDPAU_DECODER &&
2264 s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
2265 ff_vdpau_h264_picture_complete(s);
2268 * FIXME: Error handling code does not seem to support interlaced
2269 * when slices span multiple rows
2270 * The ff_er_add_slice calls don't work right for bottom
2271 * fields; they cause massive erroneous error concealing
2272 * Error marking covers both fields (top and bottom).
2273 * This causes a mismatched s->error_count
2274 * and a bad error table. Further, the error count goes to
2275 * INT_MAX when called for bottom field, because mb_y is
2276 * past end by one (callers fault) and resync_mb_y != 0
2277 * causes problems for the first MB line, too.
2282 ff_MPV_frame_end(s);
2284 h->current_slice = 0;
2290 * Replicate H264 "master" context to thread contexts.
2292 static void clone_slice(H264Context *dst, H264Context *src)
2294 memcpy(dst->block_offset, src->block_offset, sizeof(dst->block_offset));
2295 dst->s.current_picture_ptr = src->s.current_picture_ptr;
2296 dst->s.current_picture = src->s.current_picture;
2297 dst->s.linesize = src->s.linesize;
2298 dst->s.uvlinesize = src->s.uvlinesize;
2299 dst->s.first_field = src->s.first_field;
2301 dst->prev_poc_msb = src->prev_poc_msb;
2302 dst->prev_poc_lsb = src->prev_poc_lsb;
2303 dst->prev_frame_num_offset = src->prev_frame_num_offset;
2304 dst->prev_frame_num = src->prev_frame_num;
2305 dst->short_ref_count = src->short_ref_count;
2307 memcpy(dst->short_ref, src->short_ref, sizeof(dst->short_ref));
2308 memcpy(dst->long_ref, src->long_ref, sizeof(dst->long_ref));
2309 memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2310 memcpy(dst->ref_list, src->ref_list, sizeof(dst->ref_list));
2312 memcpy(dst->dequant4_coeff, src->dequant4_coeff, sizeof(src->dequant4_coeff));
2313 memcpy(dst->dequant8_coeff, src->dequant8_coeff, sizeof(src->dequant8_coeff));
2317 * Compute profile from profile_idc and constraint_set?_flags.
2321 * @return profile as defined by FF_PROFILE_H264_*
2323 int ff_h264_get_profile(SPS *sps)
2325 int profile = sps->profile_idc;
2327 switch (sps->profile_idc) {
2328 case FF_PROFILE_H264_BASELINE:
2329 // constraint_set1_flag set to 1
2330 profile |= (sps->constraint_set_flags & 1 << 1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2332 case FF_PROFILE_H264_HIGH_10:
2333 case FF_PROFILE_H264_HIGH_422:
2334 case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2335 // constraint_set3_flag set to 1
2336 profile |= (sps->constraint_set_flags & 1 << 3) ? FF_PROFILE_H264_INTRA : 0;
2344 * Decode a slice header.
2345 * This will also call ff_MPV_common_init() and frame_start() as needed.
2347 * @param h h264context
2348 * @param h0 h264 master context (differs from 'h' when doing sliced based
2349 * parallel decoding)
2351 * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2353 static int decode_slice_header(H264Context *h, H264Context *h0)
2355 MpegEncContext *const s = &h->s;
2356 MpegEncContext *const s0 = &h0->s;
2357 unsigned int first_mb_in_slice;
2358 unsigned int pps_id;
2359 int num_ref_idx_active_override_flag, max_refs;
2360 unsigned int slice_type, tmp, i, j;
2361 int default_ref_list_done = 0;
2362 int last_pic_structure, last_pic_dropable;
2364 /* FIXME: 2tap qpel isn't implemented for high bit depth. */
2365 if ((s->avctx->flags2 & CODEC_FLAG2_FAST) &&
2366 !h->nal_ref_idc && !h->pixel_shift) {
2367 s->me.qpel_put = s->dsp.put_2tap_qpel_pixels_tab;
2368 s->me.qpel_avg = s->dsp.avg_2tap_qpel_pixels_tab;
2370 s->me.qpel_put = s->dsp.put_h264_qpel_pixels_tab;
2371 s->me.qpel_avg = s->dsp.avg_h264_qpel_pixels_tab;
2374 first_mb_in_slice = get_ue_golomb(&s->gb);
2376 if (first_mb_in_slice == 0) { // FIXME better field boundary detection
2377 if (h0->current_slice && FIELD_PICTURE) {
2381 h0->current_slice = 0;
2382 if (!s0->first_field) {
2383 if (s->current_picture_ptr && !s->dropable &&
2384 s->current_picture_ptr->owner2 == s) {
2385 ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX,
2386 s->picture_structure == PICT_BOTTOM_FIELD);
2388 s->current_picture_ptr = NULL;
2392 slice_type = get_ue_golomb_31(&s->gb);
2393 if (slice_type > 9) {
2394 av_log(h->s.avctx, AV_LOG_ERROR,
2395 "slice type too large (%d) at %d %d\n",
2396 h->slice_type, s->mb_x, s->mb_y);
2399 if (slice_type > 4) {
2401 h->slice_type_fixed = 1;
2403 h->slice_type_fixed = 0;
2405 slice_type = golomb_to_pict_type[slice_type];
2406 if (slice_type == AV_PICTURE_TYPE_I ||
2407 (h0->current_slice != 0 && slice_type == h0->last_slice_type)) {
2408 default_ref_list_done = 1;
2410 h->slice_type = slice_type;
2411 h->slice_type_nos = slice_type & 3;
2413 // to make a few old functions happy, it's wrong though
2414 s->pict_type = h->slice_type;
2416 pps_id = get_ue_golomb(&s->gb);
2417 if (pps_id >= MAX_PPS_COUNT) {
2418 av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
2421 if (!h0->pps_buffers[pps_id]) {
2422 av_log(h->s.avctx, AV_LOG_ERROR,
2423 "non-existing PPS %u referenced\n",
2427 h->pps = *h0->pps_buffers[pps_id];
2429 if (!h0->sps_buffers[h->pps.sps_id]) {
2430 av_log(h->s.avctx, AV_LOG_ERROR,
2431 "non-existing SPS %u referenced\n",
2435 h->sps = *h0->sps_buffers[h->pps.sps_id];
2437 s->avctx->profile = ff_h264_get_profile(&h->sps);
2438 s->avctx->level = h->sps.level_idc;
2439 s->avctx->refs = h->sps.ref_frame_count;
2441 s->mb_width = h->sps.mb_width;
2442 s->mb_height = h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2444 h->b_stride = s->mb_width * 4;
2446 s->chroma_y_shift = h->sps.chroma_format_idc <= 1; // 400 uses yuv420p
2448 s->width = 16 * s->mb_width - (2 >> CHROMA444) * FFMIN(h->sps.crop_right, (8 << CHROMA444) - 1);
2449 if (h->sps.frame_mbs_only_flag)
2450 s->height = 16 * s->mb_height - (1 << s->chroma_y_shift) * FFMIN(h->sps.crop_bottom, (16 >> s->chroma_y_shift) - 1);
2452 s->height = 16 * s->mb_height - (2 << s->chroma_y_shift) * FFMIN(h->sps.crop_bottom, (16 >> s->chroma_y_shift) - 1);
2454 if (FFALIGN(s->avctx->width, 16) == s->width &&
2455 FFALIGN(s->avctx->height, 16) == s->height) {
2456 s->width = s->avctx->width;
2457 s->height = s->avctx->height;
2460 if (s->context_initialized &&
2461 (s->width != s->avctx->width || s->height != s->avctx->height ||
2462 av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio))) {
2463 if (h != h0 || (HAVE_THREADS && h->s.avctx->active_thread_type & FF_THREAD_FRAME)) {
2464 av_log_missing_feature(s->avctx,
2465 "Width/height changing with threads", 0);
2466 return AVERROR_PATCHWELCOME; // width / height changed during parallelized decoding
2469 flush_dpb(s->avctx);
2470 ff_MPV_common_end(s);
2472 if (!s->context_initialized) {
2474 av_log(h->s.avctx, AV_LOG_ERROR,
2475 "Cannot (re-)initialize context during parallel decoding.\n");
2479 avcodec_set_dimensions(s->avctx, s->width, s->height);
2480 s->avctx->sample_aspect_ratio = h->sps.sar;
2481 av_assert0(s->avctx->sample_aspect_ratio.den);
2483 if (h->sps.video_signal_type_present_flag) {
2484 s->avctx->color_range = h->sps.full_range ? AVCOL_RANGE_JPEG
2486 if (h->sps.colour_description_present_flag) {
2487 s->avctx->color_primaries = h->sps.color_primaries;
2488 s->avctx->color_trc = h->sps.color_trc;
2489 s->avctx->colorspace = h->sps.colorspace;
2493 if (h->sps.timing_info_present_flag) {
2494 int64_t den = h->sps.time_scale;
2495 if (h->x264_build < 44U)
2497 av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
2498 h->sps.num_units_in_tick, den, 1 << 30);
2501 switch (h->sps.bit_depth_luma) {
2504 if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2505 s->avctx->pix_fmt = AV_PIX_FMT_GBRP9;
2507 s->avctx->pix_fmt = AV_PIX_FMT_YUV444P9;
2508 } else if (CHROMA422)
2509 s->avctx->pix_fmt = AV_PIX_FMT_YUV422P9;
2511 s->avctx->pix_fmt = AV_PIX_FMT_YUV420P9;
2515 if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2516 s->avctx->pix_fmt = AV_PIX_FMT_GBRP10;
2518 s->avctx->pix_fmt = AV_PIX_FMT_YUV444P10;
2519 } else if (CHROMA422)
2520 s->avctx->pix_fmt = AV_PIX_FMT_YUV422P10;
2522 s->avctx->pix_fmt = AV_PIX_FMT_YUV420P10;
2526 if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2527 s->avctx->pix_fmt = AV_PIX_FMT_GBRP;
2529 s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? AV_PIX_FMT_YUVJ444P
2530 : AV_PIX_FMT_YUV444P;
2531 } else if (CHROMA422) {
2532 s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? AV_PIX_FMT_YUVJ422P
2533 : AV_PIX_FMT_YUV422P;
2535 s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
2536 s->avctx->codec->pix_fmts ?
2537 s->avctx->codec->pix_fmts :
2538 s->avctx->color_range == AVCOL_RANGE_JPEG ?
2539 hwaccel_pixfmt_list_h264_jpeg_420 :
2540 ff_hwaccel_pixfmt_list_420);
2544 av_log(s->avctx, AV_LOG_ERROR,
2545 "Unsupported bit depth: %d\n", h->sps.bit_depth_luma);
2546 return AVERROR_INVALIDDATA;
2549 s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id,
2552 if (ff_MPV_common_init(s) < 0) {
2553 av_log(h->s.avctx, AV_LOG_ERROR, "ff_MPV_common_init() failed.\n");
2557 h->prev_interlaced_frame = 1;
2559 init_scan_tables(h);
2560 if (ff_h264_alloc_tables(h) < 0) {
2561 av_log(h->s.avctx, AV_LOG_ERROR,
2562 "Could not allocate memory for h264\n");
2563 return AVERROR(ENOMEM);
2566 if (!HAVE_THREADS || !(s->avctx->active_thread_type & FF_THREAD_SLICE)) {
2567 if (context_init(h) < 0) {
2568 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2572 for (i = 1; i < s->slice_context_count; i++) {
2574 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
2575 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
2576 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
2577 c->h264dsp = h->h264dsp;
2580 c->pixel_shift = h->pixel_shift;
2581 init_scan_tables(c);
2582 clone_tables(c, h, i);
2585 for (i = 0; i < s->slice_context_count; i++)
2586 if (context_init(h->thread_context[i]) < 0) {
2587 av_log(h->s.avctx, AV_LOG_ERROR,
2588 "context_init() failed.\n");
2594 if (h == h0 && h->dequant_coeff_pps != pps_id) {
2595 h->dequant_coeff_pps = pps_id;
2596 init_dequant_tables(h);
2599 h->frame_num = get_bits(&s->gb, h->sps.log2_max_frame_num);
2602 h->mb_aff_frame = 0;
2603 last_pic_structure = s0->picture_structure;
2604 last_pic_dropable = s->dropable;
2605 s->dropable = h->nal_ref_idc == 0;
2606 if (h->sps.frame_mbs_only_flag) {
2607 s->picture_structure = PICT_FRAME;
2609 if (get_bits1(&s->gb)) { // field_pic_flag
2610 s->picture_structure = PICT_TOP_FIELD + get_bits1(&s->gb); // bottom_field_flag
2612 s->picture_structure = PICT_FRAME;
2613 h->mb_aff_frame = h->sps.mb_aff;
2616 h->mb_field_decoding_flag = s->picture_structure != PICT_FRAME;
2618 if (h0->current_slice != 0) {
2619 if (last_pic_structure != s->picture_structure ||
2620 last_pic_dropable != s->dropable) {
2621 av_log(h->s.avctx, AV_LOG_ERROR,
2622 "Changing field mode (%d -> %d) between slices is not allowed\n",
2623 last_pic_structure, s->picture_structure);
2624 s->picture_structure = last_pic_structure;
2625 s->dropable = last_pic_dropable;
2626 return AVERROR_INVALIDDATA;
2629 /* Shorten frame num gaps so we don't have to allocate reference
2630 * frames just to throw them away */
2631 if (h->frame_num != h->prev_frame_num) {
2632 int unwrap_prev_frame_num = h->prev_frame_num;
2633 int max_frame_num = 1 << h->sps.log2_max_frame_num;
2635 if (unwrap_prev_frame_num > h->frame_num)
2636 unwrap_prev_frame_num -= max_frame_num;
2638 if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
2639 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
2640 if (unwrap_prev_frame_num < 0)
2641 unwrap_prev_frame_num += max_frame_num;
2643 h->prev_frame_num = unwrap_prev_frame_num;
2647 /* See if we have a decoded first field looking for a pair...
2648 * Here, we're using that to see if we should mark previously
2649 * decode frames as "finished".
2650 * We have to do that before the "dummy" in-between frame allocation,
2651 * since that can modify s->current_picture_ptr. */
2652 if (s0->first_field) {
2653 assert(s0->current_picture_ptr);
2654 assert(s0->current_picture_ptr->f.data[0]);
2655 assert(s0->current_picture_ptr->f.reference != DELAYED_PIC_REF);
2657 /* Mark old field/frame as completed */
2658 if (!last_pic_dropable && s0->current_picture_ptr->owner2 == s0) {
2659 ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
2660 last_pic_structure == PICT_BOTTOM_FIELD);
2663 /* figure out if we have a complementary field pair */
2664 if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2665 /* Previous field is unmatched. Don't display it, but let it
2666 * remain for reference if marked as such. */
2667 if (!last_pic_dropable && last_pic_structure != PICT_FRAME) {
2668 ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
2669 last_pic_structure == PICT_TOP_FIELD);
2672 if (s0->current_picture_ptr->frame_num != h->frame_num) {
2673 /* This and previous field were reference, but had
2674 * different frame_nums. Consider this field first in
2675 * pair. Throw away previous field except for reference
2677 if (!last_pic_dropable && last_pic_structure != PICT_FRAME) {
2678 ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
2679 last_pic_structure == PICT_TOP_FIELD);
2682 /* Second field in complementary pair */
2683 if (!((last_pic_structure == PICT_TOP_FIELD &&
2684 s->picture_structure == PICT_BOTTOM_FIELD) ||
2685 (last_pic_structure == PICT_BOTTOM_FIELD &&
2686 s->picture_structure == PICT_TOP_FIELD))) {
2687 av_log(s->avctx, AV_LOG_ERROR,
2688 "Invalid field mode combination %d/%d\n",
2689 last_pic_structure, s->picture_structure);
2690 s->picture_structure = last_pic_structure;
2691 s->dropable = last_pic_dropable;
2692 return AVERROR_INVALIDDATA;
2693 } else if (last_pic_dropable != s->dropable) {
2694 av_log(s->avctx, AV_LOG_ERROR,
2695 "Cannot combine reference and non-reference fields in the same frame\n");
2696 av_log_ask_for_sample(s->avctx, NULL);
2697 s->picture_structure = last_pic_structure;
2698 s->dropable = last_pic_dropable;
2699 return AVERROR_INVALIDDATA;
2702 /* Take ownership of this buffer. Note that if another thread owned
2703 * the first field of this buffer, we're not operating on that pointer,
2704 * so the original thread is still responsible for reporting progress
2705 * on that first field (or if that was us, we just did that above).
2706 * By taking ownership, we assign responsibility to ourselves to
2707 * report progress on the second field. */
2708 s0->current_picture_ptr->owner2 = s0;
2713 while (h->frame_num != h->prev_frame_num &&
2714 h->frame_num != (h->prev_frame_num + 1) % (1 << h->sps.log2_max_frame_num)) {
2715 Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
2716 av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n",
2717 h->frame_num, h->prev_frame_num);
2718 if (ff_h264_frame_start(h) < 0)
2720 h->prev_frame_num++;
2721 h->prev_frame_num %= 1 << h->sps.log2_max_frame_num;
2722 s->current_picture_ptr->frame_num = h->prev_frame_num;
2723 ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX, 0);
2724 ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX, 1);
2725 ff_generate_sliding_window_mmcos(h);
2726 if (ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index) < 0 &&
2727 (s->avctx->err_recognition & AV_EF_EXPLODE))
2728 return AVERROR_INVALIDDATA;
2729 /* Error concealment: if a ref is missing, copy the previous ref in its place.
2730 * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
2731 * about there being no actual duplicates.
2732 * FIXME: this doesn't copy padding for out-of-frame motion vectors. Given we're
2733 * concealing a lost frame, this probably isn't noticeable by comparison, but it should
2735 if (h->short_ref_count) {
2737 av_image_copy(h->short_ref[0]->f.data, h->short_ref[0]->f.linesize,
2738 (const uint8_t **)prev->f.data, prev->f.linesize,
2739 s->avctx->pix_fmt, s->mb_width * 16, s->mb_height * 16);
2740 h->short_ref[0]->poc = prev->poc + 2;
2742 h->short_ref[0]->frame_num = h->prev_frame_num;
2746 /* See if we have a decoded first field looking for a pair...
2747 * We're using that to see whether to continue decoding in that
2748 * frame, or to allocate a new one. */
2749 if (s0->first_field) {
2750 assert(s0->current_picture_ptr);
2751 assert(s0->current_picture_ptr->f.data[0]);
2752 assert(s0->current_picture_ptr->f.reference != DELAYED_PIC_REF);
2754 /* figure out if we have a complementary field pair */
2755 if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2756 /* Previous field is unmatched. Don't display it, but let it
2757 * remain for reference if marked as such. */
2758 s0->current_picture_ptr = NULL;
2759 s0->first_field = FIELD_PICTURE;
2761 if (s0->current_picture_ptr->frame_num != h->frame_num) {
2762 /* This and the previous field had different frame_nums.
2763 * Consider this field first in pair. Throw away previous
2764 * one except for reference purposes. */
2765 s0->first_field = 1;
2766 s0->current_picture_ptr = NULL;
2768 /* Second field in complementary pair */
2769 s0->first_field = 0;
2773 /* Frame or first field in a potentially complementary pair */
2774 s0->first_field = FIELD_PICTURE;
2777 if (!FIELD_PICTURE || s0->first_field) {
2778 if (ff_h264_frame_start(h) < 0) {
2779 s0->first_field = 0;
2783 ff_release_unused_pictures(s, 0);
2789 s->current_picture_ptr->frame_num = h->frame_num; // FIXME frame_num cleanup
2791 assert(s->mb_num == s->mb_width * s->mb_height);
2792 if (first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
2793 first_mb_in_slice >= s->mb_num) {
2794 av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
2797 s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
2798 s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
2799 if (s->picture_structure == PICT_BOTTOM_FIELD)
2800 s->resync_mb_y = s->mb_y = s->mb_y + 1;
2801 assert(s->mb_y < s->mb_height);
2803 if (s->picture_structure == PICT_FRAME) {
2804 h->curr_pic_num = h->frame_num;
2805 h->max_pic_num = 1 << h->sps.log2_max_frame_num;
2807 h->curr_pic_num = 2 * h->frame_num + 1;
2808 h->max_pic_num = 1 << (h->sps.log2_max_frame_num + 1);
2811 if (h->nal_unit_type == NAL_IDR_SLICE)
2812 get_ue_golomb(&s->gb); /* idr_pic_id */
2814 if (h->sps.poc_type == 0) {
2815 h->poc_lsb = get_bits(&s->gb, h->sps.log2_max_poc_lsb);
2817 if (h->pps.pic_order_present == 1 && s->picture_structure == PICT_FRAME)
2818 h->delta_poc_bottom = get_se_golomb(&s->gb);
2821 if (h->sps.poc_type == 1 && !h->sps.delta_pic_order_always_zero_flag) {
2822 h->delta_poc[0] = get_se_golomb(&s->gb);
2824 if (h->pps.pic_order_present == 1 && s->picture_structure == PICT_FRAME)
2825 h->delta_poc[1] = get_se_golomb(&s->gb);
2830 if (h->pps.redundant_pic_cnt_present)
2831 h->redundant_pic_count = get_ue_golomb(&s->gb);
2833 // set defaults, might be overridden a few lines later
2834 h->ref_count[0] = h->pps.ref_count[0];
2835 h->ref_count[1] = h->pps.ref_count[1];
2837 if (h->slice_type_nos != AV_PICTURE_TYPE_I) {
2838 if (h->slice_type_nos == AV_PICTURE_TYPE_B)
2839 h->direct_spatial_mv_pred = get_bits1(&s->gb);
2840 num_ref_idx_active_override_flag = get_bits1(&s->gb);
2842 if (num_ref_idx_active_override_flag) {
2843 h->ref_count[0] = get_ue_golomb(&s->gb) + 1;
2844 if (h->ref_count[0] < 1)
2845 return AVERROR_INVALIDDATA;
2846 if (h->slice_type_nos == AV_PICTURE_TYPE_B) {
2847 h->ref_count[1] = get_ue_golomb(&s->gb) + 1;
2848 if (h->ref_count[1] < 1)
2849 return AVERROR_INVALIDDATA;
2853 if (h->slice_type_nos == AV_PICTURE_TYPE_B)
2860 max_refs = s->picture_structure == PICT_FRAME ? 16 : 32;
2862 if (h->ref_count[0] > max_refs || h->ref_count[1] > max_refs) {
2863 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
2864 h->ref_count[0] = h->ref_count[1] = 1;
2865 return AVERROR_INVALIDDATA;
2868 if (!default_ref_list_done)
2869 ff_h264_fill_default_ref_list(h);
2871 if (h->slice_type_nos != AV_PICTURE_TYPE_I &&
2872 ff_h264_decode_ref_pic_list_reordering(h) < 0) {
2873 h->ref_count[1] = h->ref_count[0] = 0;
2877 if (h->slice_type_nos != AV_PICTURE_TYPE_I) {
2878 s->last_picture_ptr = &h->ref_list[0][0];
2879 s->last_picture_ptr->owner2 = s;
2880 ff_copy_picture(&s->last_picture, s->last_picture_ptr);
2882 if (h->slice_type_nos == AV_PICTURE_TYPE_B) {
2883 s->next_picture_ptr = &h->ref_list[1][0];
2884 s->next_picture_ptr->owner2 = s;
2885 ff_copy_picture(&s->next_picture, s->next_picture_ptr);
2888 if ((h->pps.weighted_pred && h->slice_type_nos == AV_PICTURE_TYPE_P) ||
2889 (h->pps.weighted_bipred_idc == 1 &&
2890 h->slice_type_nos == AV_PICTURE_TYPE_B))
2891 pred_weight_table(h);
2892 else if (h->pps.weighted_bipred_idc == 2 &&
2893 h->slice_type_nos == AV_PICTURE_TYPE_B) {
2894 implicit_weight_table(h, -1);
2897 for (i = 0; i < 2; i++) {
2898 h->luma_weight_flag[i] = 0;
2899 h->chroma_weight_flag[i] = 0;
2903 if (h->nal_ref_idc && ff_h264_decode_ref_pic_marking(h0, &s->gb) < 0 &&
2904 (s->avctx->err_recognition & AV_EF_EXPLODE))
2905 return AVERROR_INVALIDDATA;
2908 ff_h264_fill_mbaff_ref_list(h);
2910 if (h->pps.weighted_bipred_idc == 2 && h->slice_type_nos == AV_PICTURE_TYPE_B) {
2911 implicit_weight_table(h, 0);
2912 implicit_weight_table(h, 1);
2916 if (h->slice_type_nos == AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
2917 ff_h264_direct_dist_scale_factor(h);
2918 ff_h264_direct_ref_list_init(h);
2920 if (h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac) {
2921 tmp = get_ue_golomb_31(&s->gb);
2923 av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
2926 h->cabac_init_idc = tmp;
2929 h->last_qscale_diff = 0;
2930 tmp = h->pps.init_qp + get_se_golomb(&s->gb);
2931 if (tmp > 51 + 6 * (h->sps.bit_depth_luma - 8)) {
2932 av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
2936 h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
2937 h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
2938 // FIXME qscale / qp ... stuff
2939 if (h->slice_type == AV_PICTURE_TYPE_SP)
2940 get_bits1(&s->gb); /* sp_for_switch_flag */
2941 if (h->slice_type == AV_PICTURE_TYPE_SP ||
2942 h->slice_type == AV_PICTURE_TYPE_SI)
2943 get_se_golomb(&s->gb); /* slice_qs_delta */
2945 h->deblocking_filter = 1;
2946 h->slice_alpha_c0_offset = 52;
2947 h->slice_beta_offset = 52;
2948 if (h->pps.deblocking_filter_parameters_present) {
2949 tmp = get_ue_golomb_31(&s->gb);
2951 av_log(s->avctx, AV_LOG_ERROR,
2952 "deblocking_filter_idc %u out of range\n", tmp);
2955 h->deblocking_filter = tmp;
2956 if (h->deblocking_filter < 2)
2957 h->deblocking_filter ^= 1; // 1<->0
2959 if (h->deblocking_filter) {
2960 h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
2961 h->slice_beta_offset += get_se_golomb(&s->gb) << 1;
2962 if (h->slice_alpha_c0_offset > 104U ||
2963 h->slice_beta_offset > 104U) {
2964 av_log(s->avctx, AV_LOG_ERROR,
2965 "deblocking filter parameters %d %d out of range\n",
2966 h->slice_alpha_c0_offset, h->slice_beta_offset);
2972 if (s->avctx->skip_loop_filter >= AVDISCARD_ALL ||
2973 (s->avctx->skip_loop_filter >= AVDISCARD_NONKEY &&
2974 h->slice_type_nos != AV_PICTURE_TYPE_I) ||
2975 (s->avctx->skip_loop_filter >= AVDISCARD_BIDIR &&
2976 h->slice_type_nos == AV_PICTURE_TYPE_B) ||
2977 (s->avctx->skip_loop_filter >= AVDISCARD_NONREF &&
2978 h->nal_ref_idc == 0))
2979 h->deblocking_filter = 0;
2981 if (h->deblocking_filter == 1 && h0->max_contexts > 1) {
2982 if (s->avctx->flags2 & CODEC_FLAG2_FAST) {
2983 /* Cheat slightly for speed:
2984 * Do not bother to deblock across slices. */
2985 h->deblocking_filter = 2;
2987 h0->max_contexts = 1;
2988 if (!h0->single_decode_warning) {
2989 av_log(s->avctx, AV_LOG_INFO,
2990 "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
2991 h0->single_decode_warning = 1;
2994 av_log(h->s.avctx, AV_LOG_ERROR,
2995 "Deblocking switched inside frame.\n");
3000 h->qp_thresh = 15 + 52 -
3001 FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset) -
3003 h->pps.chroma_qp_index_offset[0],
3004 h->pps.chroma_qp_index_offset[1]) +
3005 6 * (h->sps.bit_depth_luma - 8);
3007 h0->last_slice_type = slice_type;
3008 h->slice_num = ++h0->current_slice;
3009 if (h->slice_num >= MAX_SLICES) {
3010 av_log(s->avctx, AV_LOG_ERROR,
3011 "Too many slices, increase MAX_SLICES and recompile\n");
3014 for (j = 0; j < 2; j++) {
3016 int *ref2frm = h->ref2frm[h->slice_num & (MAX_SLICES - 1)][j];
3017 for (i = 0; i < 16; i++) {
3019 if (h->ref_list[j][i].f.data[0]) {
3021 uint8_t *base = h->ref_list[j][i].f.base[0];
3022 for (k = 0; k < h->short_ref_count; k++)
3023 if (h->short_ref[k]->f.base[0] == base) {
3027 for (k = 0; k < h->long_ref_count; k++)
3028 if (h->long_ref[k] && h->long_ref[k]->f.base[0] == base) {
3029 id_list[i] = h->short_ref_count + k;
3037 for (i = 0; i < 16; i++)
3038 ref2frm[i + 2] = 4 * id_list[i] +
3039 (h->ref_list[j][i].f.reference & 3);
3041 ref2frm[18 + 1] = -1;
3042 for (i = 16; i < 48; i++)
3043 ref2frm[i + 4] = 4 * id_list[(i - 16) >> 1] +
3044 (h->ref_list[j][i].f.reference & 3);
3047 // FIXME: fix draw_edges + PAFF + frame threads
3048 h->emu_edge_width = (s->flags & CODEC_FLAG_EMU_EDGE ||
3049 (!h->sps.frame_mbs_only_flag &&
3050 s->avctx->active_thread_type))
3052 h->emu_edge_height = (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
3054 if (s->avctx->debug & FF_DEBUG_PICT_INFO) {
3055 av_log(h->s.avctx, AV_LOG_DEBUG,
3056 "slice:%d %s mb:%d %c%s%s pps:%u frame:%d poc:%d/%d ref:%d/%d qp:%d loop:%d:%d:%d weight:%d%s %s\n",
3058 (s->picture_structure == PICT_FRAME ? "F" : s->picture_structure == PICT_TOP_FIELD ? "T" : "B"),
3060 av_get_picture_type_char(h->slice_type),
3061 h->slice_type_fixed ? " fix" : "",
3062 h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
3063 pps_id, h->frame_num,
3064 s->current_picture_ptr->field_poc[0],
3065 s->current_picture_ptr->field_poc[1],
3066 h->ref_count[0], h->ref_count[1],
3068 h->deblocking_filter,
3069 h->slice_alpha_c0_offset / 2 - 26, h->slice_beta_offset / 2 - 26,
3071 h->use_weight == 1 && h->use_weight_chroma ? "c" : "",
3072 h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : "");
3078 int ff_h264_get_slice_type(const H264Context *h)
3080 switch (h->slice_type) {
3081 case AV_PICTURE_TYPE_P:
3083 case AV_PICTURE_TYPE_B:
3085 case AV_PICTURE_TYPE_I:
3087 case AV_PICTURE_TYPE_SP:
3089 case AV_PICTURE_TYPE_SI:
3096 static av_always_inline void fill_filter_caches_inter(H264Context *h,
3097 MpegEncContext *const s,
3098 int mb_type, int top_xy,
3099 int left_xy[LEFT_MBS],
3101 int left_type[LEFT_MBS],
3102 int mb_xy, int list)
3104 int b_stride = h->b_stride;
3105 int16_t(*mv_dst)[2] = &h->mv_cache[list][scan8[0]];
3106 int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
3107 if (IS_INTER(mb_type) || IS_DIRECT(mb_type)) {
3108 if (USES_LIST(top_type, list)) {
3109 const int b_xy = h->mb2b_xy[top_xy] + 3 * b_stride;
3110 const int b8_xy = 4 * top_xy + 2;
3111 int (*ref2frm)[64] = h->ref2frm[h->slice_table[top_xy] & (MAX_SLICES - 1)][0] + (MB_MBAFF ? 20 : 2);
3112 AV_COPY128(mv_dst - 1 * 8, s->current_picture.f.motion_val[list][b_xy + 0]);
3113 ref_cache[0 - 1 * 8] =
3114 ref_cache[1 - 1 * 8] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 0]];
3115 ref_cache[2 - 1 * 8] =
3116 ref_cache[3 - 1 * 8] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 1]];
3118 AV_ZERO128(mv_dst - 1 * 8);
3119 AV_WN32A(&ref_cache[0 - 1 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3122 if (!IS_INTERLACED(mb_type ^ left_type[LTOP])) {
3123 if (USES_LIST(left_type[LTOP], list)) {
3124 const int b_xy = h->mb2b_xy[left_xy[LTOP]] + 3;
3125 const int b8_xy = 4 * left_xy[LTOP] + 1;
3126 int (*ref2frm)[64] = h->ref2frm[h->slice_table[left_xy[LTOP]] & (MAX_SLICES - 1)][0] + (MB_MBAFF ? 20 : 2);
3127 AV_COPY32(mv_dst - 1 + 0, s->current_picture.f.motion_val[list][b_xy + b_stride * 0]);
3128 AV_COPY32(mv_dst - 1 + 8, s->current_picture.f.motion_val[list][b_xy + b_stride * 1]);
3129 AV_COPY32(mv_dst - 1 + 16, s->current_picture.f.motion_val[list][b_xy + b_stride * 2]);
3130 AV_COPY32(mv_dst - 1 + 24, s->current_picture.f.motion_val[list][b_xy + b_stride * 3]);
3132 ref_cache[-1 + 8] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2 * 0]];
3133 ref_cache[-1 + 16] =
3134 ref_cache[-1 + 24] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2 * 1]];
3136 AV_ZERO32(mv_dst - 1 + 0);
3137 AV_ZERO32(mv_dst - 1 + 8);
3138 AV_ZERO32(mv_dst - 1 + 16);
3139 AV_ZERO32(mv_dst - 1 + 24);
3142 ref_cache[-1 + 16] =
3143 ref_cache[-1 + 24] = LIST_NOT_USED;
3148 if (!USES_LIST(mb_type, list)) {
3149 fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0, 0), 4);
3150 AV_WN32A(&ref_cache[0 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3151 AV_WN32A(&ref_cache[1 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3152 AV_WN32A(&ref_cache[2 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3153 AV_WN32A(&ref_cache[3 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3158 int8_t *ref = &s->current_picture.f.ref_index[list][4 * mb_xy];
3159 int (*ref2frm)[64] = h->ref2frm[h->slice_num & (MAX_SLICES - 1)][0] + (MB_MBAFF ? 20 : 2);
3160 uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]], ref2frm[list][ref[1]]) & 0x00FF00FF) * 0x0101;
3161 uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]], ref2frm[list][ref[3]]) & 0x00FF00FF) * 0x0101;
3162 AV_WN32A(&ref_cache[0 * 8], ref01);
3163 AV_WN32A(&ref_cache[1 * 8], ref01);
3164 AV_WN32A(&ref_cache[2 * 8], ref23);
3165 AV_WN32A(&ref_cache[3 * 8], ref23);
3169 int16_t(*mv_src)[2] = &s->current_picture.f.motion_val[list][4 * s->mb_x + 4 * s->mb_y * b_stride];
3170 AV_COPY128(mv_dst + 8 * 0, mv_src + 0 * b_stride);
3171 AV_COPY128(mv_dst + 8 * 1, mv_src + 1 * b_stride);
3172 AV_COPY128(mv_dst + 8 * 2, mv_src + 2 * b_stride);
3173 AV_COPY128(mv_dst + 8 * 3, mv_src + 3 * b_stride);
3179 * @return non zero if the loop filter can be skipped
3181 static int fill_filter_caches(H264Context *h, int mb_type)
3183 MpegEncContext *const s = &h->s;
3184 const int mb_xy = h->mb_xy;
3185 int top_xy, left_xy[LEFT_MBS];
3186 int top_type, left_type[LEFT_MBS];
3190 top_xy = mb_xy - (s->mb_stride << MB_FIELD);
3192 /* Wow, what a mess, why didn't they simplify the interlacing & intra
3193 * stuff, I can't imagine that these complex rules are worth it. */
3195 left_xy[LBOT] = left_xy[LTOP] = mb_xy - 1;
3197 const int left_mb_field_flag = IS_INTERLACED(s->current_picture.f.mb_type[mb_xy - 1]);
3198 const int curr_mb_field_flag = IS_INTERLACED(mb_type);
3200 if (left_mb_field_flag != curr_mb_field_flag)
3201 left_xy[LTOP] -= s->mb_stride;
3203 if (curr_mb_field_flag)
3204 top_xy += s->mb_stride &
3205 (((s->current_picture.f.mb_type[top_xy] >> 7) & 1) - 1);
3206 if (left_mb_field_flag != curr_mb_field_flag)
3207 left_xy[LBOT] += s->mb_stride;
3211 h->top_mb_xy = top_xy;
3212 h->left_mb_xy[LTOP] = left_xy[LTOP];
3213 h->left_mb_xy[LBOT] = left_xy[LBOT];
3215 /* For sufficiently low qp, filtering wouldn't do anything.
3216 * This is a conservative estimate: could also check beta_offset
3217 * and more accurate chroma_qp. */
3218 int qp_thresh = h->qp_thresh; // FIXME strictly we should store qp_thresh for each mb of a slice
3219 int qp = s->current_picture.f.qscale_table[mb_xy];
3220 if (qp <= qp_thresh &&
3221 (left_xy[LTOP] < 0 ||
3222 ((qp + s->current_picture.f.qscale_table[left_xy[LTOP]] + 1) >> 1) <= qp_thresh) &&
3224 ((qp + s->current_picture.f.qscale_table[top_xy] + 1) >> 1) <= qp_thresh)) {
3227 if ((left_xy[LTOP] < 0 ||
3228 ((qp + s->current_picture.f.qscale_table[left_xy[LBOT]] + 1) >> 1) <= qp_thresh) &&
3229 (top_xy < s->mb_stride ||
3230 ((qp + s->current_picture.f.qscale_table[top_xy - s->mb_stride] + 1) >> 1) <= qp_thresh))
3235 top_type = s->current_picture.f.mb_type[top_xy];
3236 left_type[LTOP] = s->current_picture.f.mb_type[left_xy[LTOP]];
3237 left_type[LBOT] = s->current_picture.f.mb_type[left_xy[LBOT]];
3238 if (h->deblocking_filter == 2) {
3239 if (h->slice_table[top_xy] != h->slice_num)
3241 if (h->slice_table[left_xy[LBOT]] != h->slice_num)
3242 left_type[LTOP] = left_type[LBOT] = 0;
3244 if (h->slice_table[top_xy] == 0xFFFF)
3246 if (h->slice_table[left_xy[LBOT]] == 0xFFFF)
3247 left_type[LTOP] = left_type[LBOT] = 0;
3249 h->top_type = top_type;
3250 h->left_type[LTOP] = left_type[LTOP];
3251 h->left_type[LBOT] = left_type[LBOT];
3253 if (IS_INTRA(mb_type))
3256 fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy,
3257 top_type, left_type, mb_xy, 0);
3258 if (h->list_count == 2)
3259 fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy,
3260 top_type, left_type, mb_xy, 1);
3262 nnz = h->non_zero_count[mb_xy];
3263 nnz_cache = h->non_zero_count_cache;
3264 AV_COPY32(&nnz_cache[4 + 8 * 1], &nnz[0]);
3265 AV_COPY32(&nnz_cache[4 + 8 * 2], &nnz[4]);
3266 AV_COPY32(&nnz_cache[4 + 8 * 3], &nnz[8]);
3267 AV_COPY32(&nnz_cache[4 + 8 * 4], &nnz[12]);
3268 h->cbp = h->cbp_table[mb_xy];
3271 nnz = h->non_zero_count[top_xy];
3272 AV_COPY32(&nnz_cache[4 + 8 * 0], &nnz[3 * 4]);
3275 if (left_type[LTOP]) {
3276 nnz = h->non_zero_count[left_xy[LTOP]];
3277 nnz_cache[3 + 8 * 1] = nnz[3 + 0 * 4];
3278 nnz_cache[3 + 8 * 2] = nnz[3 + 1 * 4];
3279 nnz_cache[3 + 8 * 3] = nnz[3 + 2 * 4];
3280 nnz_cache[3 + 8 * 4] = nnz[3 + 3 * 4];
3283 /* CAVLC 8x8dct requires NNZ values for residual decoding that differ
3284 * from what the loop filter needs */
3285 if (!CABAC && h->pps.transform_8x8_mode) {
3286 if (IS_8x8DCT(top_type)) {
3287 nnz_cache[4 + 8 * 0] =
3288 nnz_cache[5 + 8 * 0] = (h->cbp_table[top_xy] & 0x4000) >> 12;
3289 nnz_cache[6 + 8 * 0] =
3290 nnz_cache[7 + 8 * 0] = (h->cbp_table[top_xy] & 0x8000) >> 12;
3292 if (IS_8x8DCT(left_type[LTOP])) {
3293 nnz_cache[3 + 8 * 1] =
3294 nnz_cache[3 + 8 * 2] = (h->cbp_table[left_xy[LTOP]] & 0x2000) >> 12; // FIXME check MBAFF
3296 if (IS_8x8DCT(left_type[LBOT])) {
3297 nnz_cache[3 + 8 * 3] =
3298 nnz_cache[3 + 8 * 4] = (h->cbp_table[left_xy[LBOT]] & 0x8000) >> 12; // FIXME check MBAFF
3301 if (IS_8x8DCT(mb_type)) {
3302 nnz_cache[scan8[0]] =
3303 nnz_cache[scan8[1]] =
3304 nnz_cache[scan8[2]] =
3305 nnz_cache[scan8[3]] = (h->cbp & 0x1000) >> 12;
3307 nnz_cache[scan8[0 + 4]] =
3308 nnz_cache[scan8[1 + 4]] =
3309 nnz_cache[scan8[2 + 4]] =
3310 nnz_cache[scan8[3 + 4]] = (h->cbp & 0x2000) >> 12;
3312 nnz_cache[scan8[0 + 8]] =
3313 nnz_cache[scan8[1 + 8]] =
3314 nnz_cache[scan8[2 + 8]] =
3315 nnz_cache[scan8[3 + 8]] = (h->cbp & 0x4000) >> 12;
3317 nnz_cache[scan8[0 + 12]] =
3318 nnz_cache[scan8[1 + 12]] =
3319 nnz_cache[scan8[2 + 12]] =
3320 nnz_cache[scan8[3 + 12]] = (h->cbp & 0x8000) >> 12;
3327 static void loop_filter(H264Context *h, int start_x, int end_x)
3329 MpegEncContext *const s = &h->s;
3330 uint8_t *dest_y, *dest_cb, *dest_cr;
3331 int linesize, uvlinesize, mb_x, mb_y;
3332 const int end_mb_y = s->mb_y + FRAME_MBAFF;
3333 const int old_slice_type = h->slice_type;
3334 const int pixel_shift = h->pixel_shift;
3335 const int block_h = 16 >> s->chroma_y_shift;
3337 if (h->deblocking_filter) {
3338 for (mb_x = start_x; mb_x < end_x; mb_x++)
3339 for (mb_y = end_mb_y - FRAME_MBAFF; mb_y <= end_mb_y; mb_y++) {
3341 mb_xy = h->mb_xy = mb_x + mb_y * s->mb_stride;
3342 h->slice_num = h->slice_table[mb_xy];
3343 mb_type = s->current_picture.f.mb_type[mb_xy];
3344 h->list_count = h->list_counts[mb_xy];
3348 h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
3352 dest_y = s->current_picture.f.data[0] +
3353 ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
3354 dest_cb = s->current_picture.f.data[1] +
3355 (mb_x << pixel_shift) * (8 << CHROMA444) +
3356 mb_y * s->uvlinesize * block_h;
3357 dest_cr = s->current_picture.f.data[2] +
3358 (mb_x << pixel_shift) * (8 << CHROMA444) +
3359 mb_y * s->uvlinesize * block_h;
3360 // FIXME simplify above
3363 linesize = h->mb_linesize = s->linesize * 2;
3364 uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
3365 if (mb_y & 1) { // FIXME move out of this function?
3366 dest_y -= s->linesize * 15;
3367 dest_cb -= s->uvlinesize * (block_h - 1);
3368 dest_cr -= s->uvlinesize * (block_h - 1);
3371 linesize = h->mb_linesize = s->linesize;
3372 uvlinesize = h->mb_uvlinesize = s->uvlinesize;
3374 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize,
3376 if (fill_filter_caches(h, mb_type))
3378 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.f.qscale_table[mb_xy]);
3379 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.f.qscale_table[mb_xy]);
3382 ff_h264_filter_mb(h, mb_x, mb_y, dest_y, dest_cb, dest_cr,
3383 linesize, uvlinesize);
3385 ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb,
3386 dest_cr, linesize, uvlinesize);
3390 h->slice_type = old_slice_type;
3392 s->mb_y = end_mb_y - FRAME_MBAFF;
3393 h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3394 h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3397 static void predict_field_decoding_flag(H264Context *h)
3399 MpegEncContext *const s = &h->s;
3400 const int mb_xy = s->mb_x + s->mb_y * s->mb_stride;
3401 int mb_type = (h->slice_table[mb_xy - 1] == h->slice_num) ?
3402 s->current_picture.f.mb_type[mb_xy - 1] :
3403 (h->slice_table[mb_xy - s->mb_stride] == h->slice_num) ?
3404 s->current_picture.f.mb_type[mb_xy - s->mb_stride] : 0;
3405 h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
3409 * Draw edges and report progress for the last MB row.
3411 static void decode_finish_row(H264Context *h)
3413 MpegEncContext *const s = &h->s;
3414 int top = 16 * (s->mb_y >> FIELD_PICTURE);
3415 int pic_height = 16 * s->mb_height >> FIELD_PICTURE;
3416 int height = 16 << FRAME_MBAFF;
3417 int deblock_border = (16 + 4) << FRAME_MBAFF;
3419 if (h->deblocking_filter) {
3420 if ((top + height) >= pic_height)
3421 height += deblock_border;
3422 top -= deblock_border;
3425 if (top >= pic_height || (top + height) < h->emu_edge_height)
3428 height = FFMIN(height, pic_height - top);
3429 if (top < h->emu_edge_height) {
3430 height = top + height;
3434 ff_draw_horiz_band(s, top, height);
3439 ff_thread_report_progress(&s->current_picture_ptr->f, top + height - 1,
3440 s->picture_structure == PICT_BOTTOM_FIELD);
3443 static int decode_slice(struct AVCodecContext *avctx, void *arg)
3445 H264Context *h = *(void **)arg;
3446 MpegEncContext *const s = &h->s;
3447 const int part_mask = s->partitioned_frame ? (ER_AC_END | ER_AC_ERROR)
3449 int lf_x_start = s->mb_x;
3451 s->mb_skip_run = -1;
3453 h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME ||
3454 s->codec_id != AV_CODEC_ID_H264 ||
3455 (CONFIG_GRAY && (s->flags & CODEC_FLAG_GRAY));
3459 align_get_bits(&s->gb);
3462 ff_init_cabac_states(&h->cabac);
3463 ff_init_cabac_decoder(&h->cabac,
3464 s->gb.buffer + get_bits_count(&s->gb) / 8,
3465 (get_bits_left(&s->gb) + 7) / 8);
3467 ff_h264_init_cabac_states(h);
3471 int ret = ff_h264_decode_mb_cabac(h);
3473 // STOP_TIMER("decode_mb_cabac")
3476 ff_h264_hl_decode_mb(h);
3478 // FIXME optimal? or let mb_decode decode 16x32 ?
3479 if (ret >= 0 && FRAME_MBAFF) {
3482 ret = ff_h264_decode_mb_cabac(h);
3485 ff_h264_hl_decode_mb(h);
3488 eos = get_cabac_terminate(&h->cabac);
3490 if ((s->workaround_bugs & FF_BUG_TRUNCATED) &&
3491 h->cabac.bytestream > h->cabac.bytestream_end + 2) {
3492 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x - 1,
3493 s->mb_y, ER_MB_END & part_mask);
3494 if (s->mb_x >= lf_x_start)
3495 loop_filter(h, lf_x_start, s->mb_x + 1);
3498 if (ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 2) {
3499 av_log(h->s.avctx, AV_LOG_ERROR,
3500 "error while decoding MB %d %d, bytestream (%td)\n",
3502 h->cabac.bytestream_end - h->cabac.bytestream);
3503 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x,
3504 s->mb_y, ER_MB_ERROR & part_mask);
3508 if (++s->mb_x >= s->mb_width) {
3509 loop_filter(h, lf_x_start, s->mb_x);
3510 s->mb_x = lf_x_start = 0;
3511 decode_finish_row(h);
3513 if (FIELD_OR_MBAFF_PICTURE) {
3515 if (FRAME_MBAFF && s->mb_y < s->mb_height)
3516 predict_field_decoding_flag(h);
3520 if (eos || s->mb_y >= s->mb_height) {
3521 tprintf(s->avctx, "slice end %d %d\n",
3522 get_bits_count(&s->gb), s->gb.size_in_bits);
3523 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x - 1,
3524 s->mb_y, ER_MB_END & part_mask);
3525 if (s->mb_x > lf_x_start)
3526 loop_filter(h, lf_x_start, s->mb_x);
3532 int ret = ff_h264_decode_mb_cavlc(h);
3535 ff_h264_hl_decode_mb(h);
3537 // FIXME optimal? or let mb_decode decode 16x32 ?
3538 if (ret >= 0 && FRAME_MBAFF) {
3540 ret = ff_h264_decode_mb_cavlc(h);
3543 ff_h264_hl_decode_mb(h);
3548 av_log(h->s.avctx, AV_LOG_ERROR,
3549 "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
3550 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x,
3551 s->mb_y, ER_MB_ERROR & part_mask);
3555 if (++s->mb_x >= s->mb_width) {
3556 loop_filter(h, lf_x_start, s->mb_x);
3557 s->mb_x = lf_x_start = 0;
3558 decode_finish_row(h);
3560 if (FIELD_OR_MBAFF_PICTURE) {
3562 if (FRAME_MBAFF && s->mb_y < s->mb_height)
3563 predict_field_decoding_flag(h);
3565 if (s->mb_y >= s->mb_height) {
3566 tprintf(s->avctx, "slice end %d %d\n",
3567 get_bits_count(&s->gb), s->gb.size_in_bits);
3569 if (get_bits_left(&s->gb) == 0) {
3570 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y,
3571 s->mb_x - 1, s->mb_y,
3572 ER_MB_END & part_mask);
3576 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y,
3577 s->mb_x - 1, s->mb_y,
3578 ER_MB_END & part_mask);
3585 if (get_bits_left(&s->gb) <= 0 && s->mb_skip_run <= 0) {
3586 tprintf(s->avctx, "slice end %d %d\n",
3587 get_bits_count(&s->gb), s->gb.size_in_bits);
3588 if (get_bits_left(&s->gb) == 0) {
3589 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y,
3590 s->mb_x - 1, s->mb_y,
3591 ER_MB_END & part_mask);
3592 if (s->mb_x > lf_x_start)
3593 loop_filter(h, lf_x_start, s->mb_x);
3597 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x,
3598 s->mb_y, ER_MB_ERROR & part_mask);
3608 * Call decode_slice() for each context.
3610 * @param h h264 master context
3611 * @param context_count number of contexts to execute
3613 static int execute_decode_slices(H264Context *h, int context_count)
3615 MpegEncContext *const s = &h->s;
3616 AVCodecContext *const avctx = s->avctx;
3620 if (s->avctx->hwaccel ||
3621 s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
3623 if (context_count == 1) {
3624 return decode_slice(avctx, &h);
3626 for (i = 1; i < context_count; i++) {
3627 hx = h->thread_context[i];
3628 hx->s.err_recognition = avctx->err_recognition;
3629 hx->s.error_count = 0;
3632 avctx->execute(avctx, decode_slice, h->thread_context,
3633 NULL, context_count, sizeof(void *));
3635 /* pull back stuff from slices to master context */
3636 hx = h->thread_context[context_count - 1];
3637 s->mb_x = hx->s.mb_x;
3638 s->mb_y = hx->s.mb_y;
3639 s->dropable = hx->s.dropable;
3640 s->picture_structure = hx->s.picture_structure;
3641 for (i = 1; i < context_count; i++)
3642 h->s.error_count += h->thread_context[i]->s.error_count;
3648 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size)
3650 MpegEncContext *const s = &h->s;
3651 AVCodecContext *const avctx = s->avctx;
3652 H264Context *hx; ///< thread context
3656 int pass = !(avctx->active_thread_type & FF_THREAD_FRAME);
3657 int nals_needed = 0; ///< number of NALs that need decoding before the next frame thread starts
3660 h->max_contexts = s->slice_context_count;
3661 if (!(s->flags2 & CODEC_FLAG2_CHUNKS)) {
3662 h->current_slice = 0;
3663 if (!s->first_field)
3664 s->current_picture_ptr = NULL;
3665 ff_h264_reset_sei(h);
3668 for (; pass <= 1; pass++) {
3671 next_avc = h->is_avc ? 0 : buf_size;
3681 if (buf_index >= next_avc) {
3682 if (buf_index >= buf_size - h->nal_length_size)
3685 for (i = 0; i < h->nal_length_size; i++)
3686 nalsize = (nalsize << 8) | buf[buf_index++];
3687 if (nalsize <= 0 || nalsize > buf_size - buf_index) {
3688 av_log(h->s.avctx, AV_LOG_ERROR,
3689 "AVC: nal size %d\n", nalsize);
3692 next_avc = buf_index + nalsize;
3694 // start code prefix search
3695 for (; buf_index + 3 < next_avc; buf_index++)
3696 // This should always succeed in the first iteration.
3697 if (buf[buf_index] == 0 &&
3698 buf[buf_index + 1] == 0 &&
3699 buf[buf_index + 2] == 1)
3702 if (buf_index + 3 >= buf_size) {
3703 buf_index = buf_size;
3708 if (buf_index >= next_avc)
3712 hx = h->thread_context[context_count];
3714 ptr = ff_h264_decode_nal(hx, buf + buf_index, &dst_length,
3715 &consumed, next_avc - buf_index);
3716 if (ptr == NULL || dst_length < 0) {
3720 i = buf_index + consumed;
3721 if ((s->workaround_bugs & FF_BUG_AUTODETECT) && i + 3 < next_avc &&
3722 buf[i] == 0x00 && buf[i + 1] == 0x00 &&
3723 buf[i + 2] == 0x01 && buf[i + 3] == 0xE0)
3724 s->workaround_bugs |= FF_BUG_TRUNCATED;
3726 if (!(s->workaround_bugs & FF_BUG_TRUNCATED))
3727 while (ptr[dst_length - 1] == 0 && dst_length > 0)
3729 bit_length = !dst_length ? 0
3731 decode_rbsp_trailing(h, ptr + dst_length - 1));
3733 if (s->avctx->debug & FF_DEBUG_STARTCODE)
3734 av_log(h->s.avctx, AV_LOG_DEBUG,
3735 "NAL %d at %d/%d length %d\n",
3736 hx->nal_unit_type, buf_index, buf_size, dst_length);
3738 if (h->is_avc && (nalsize != consumed) && nalsize)
3739 av_log(h->s.avctx, AV_LOG_DEBUG,
3740 "AVC: Consumed only %d bytes instead of %d\n",
3743 buf_index += consumed;
3747 /* packets can sometimes contain multiple PPS/SPS,
3748 * e.g. two PAFF field pictures in one packet, or a demuxer
3749 * which splits NALs strangely if so, when frame threading we
3750 * can't start the next thread until we've read all of them */
3751 switch (hx->nal_unit_type) {
3754 nals_needed = nal_index;
3758 init_get_bits(&hx->s.gb, ptr, bit_length);
3759 if (!get_ue_golomb(&hx->s.gb))
3760 nals_needed = nal_index;
3765 // FIXME do not discard SEI id
3766 if (avctx->skip_frame >= AVDISCARD_NONREF && h->nal_ref_idc == 0)
3771 switch (hx->nal_unit_type) {
3773 if (h->nal_unit_type != NAL_IDR_SLICE) {
3774 av_log(h->s.avctx, AV_LOG_ERROR,
3775 "Invalid mix of idr and non-idr slices\n");
3779 idr(h); // FIXME ensure we don't lose some frames if there is reordering
3781 init_get_bits(&hx->s.gb, ptr, bit_length);
3783 hx->inter_gb_ptr = &hx->s.gb;
3784 hx->s.data_partitioning = 0;
3786 if ((err = decode_slice_header(hx, h)))
3789 s->current_picture_ptr->f.key_frame |=
3790 (hx->nal_unit_type == NAL_IDR_SLICE) ||
3791 (h->sei_recovery_frame_cnt >= 0);
3793 if (h->current_slice == 1) {
3794 if (!(s->flags2 & CODEC_FLAG2_CHUNKS))
3795 decode_postinit(h, nal_index >= nals_needed);
3797 if (s->avctx->hwaccel &&
3798 s->avctx->hwaccel->start_frame(s->avctx, NULL, 0) < 0)
3800 if (CONFIG_H264_VDPAU_DECODER &&
3801 s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
3802 ff_vdpau_h264_picture_start(s);
3805 if (hx->redundant_pic_count == 0 &&
3806 (avctx->skip_frame < AVDISCARD_NONREF ||
3808 (avctx->skip_frame < AVDISCARD_BIDIR ||
3809 hx->slice_type_nos != AV_PICTURE_TYPE_B) &&
3810 (avctx->skip_frame < AVDISCARD_NONKEY ||
3811 hx->slice_type_nos == AV_PICTURE_TYPE_I) &&
3812 avctx->skip_frame < AVDISCARD_ALL) {
3813 if (avctx->hwaccel) {
3814 if (avctx->hwaccel->decode_slice(avctx,
3815 &buf[buf_index - consumed],
3818 } else if (CONFIG_H264_VDPAU_DECODER &&
3819 s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU) {
3820 static const uint8_t start_code[] = {
3822 ff_vdpau_add_data_chunk(s, start_code,
3823 sizeof(start_code));
3824 ff_vdpau_add_data_chunk(s, &buf[buf_index - consumed],
3831 init_get_bits(&hx->s.gb, ptr, bit_length);
3833 hx->inter_gb_ptr = NULL;
3835 if ((err = decode_slice_header(hx, h)) < 0)
3838 hx->s.data_partitioning = 1;
3841 init_get_bits(&hx->intra_gb, ptr, bit_length);
3842 hx->intra_gb_ptr = &hx->intra_gb;
3845 init_get_bits(&hx->inter_gb, ptr, bit_length);
3846 hx->inter_gb_ptr = &hx->inter_gb;
3848 if (hx->redundant_pic_count == 0 &&
3850 hx->s.data_partitioning &&
3851 s->current_picture_ptr &&
3852 s->context_initialized &&
3853 (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc) &&
3854 (avctx->skip_frame < AVDISCARD_BIDIR ||
3855 hx->slice_type_nos != AV_PICTURE_TYPE_B) &&
3856 (avctx->skip_frame < AVDISCARD_NONKEY ||
3857 hx->slice_type_nos == AV_PICTURE_TYPE_I) &&
3858 avctx->skip_frame < AVDISCARD_ALL)
3862 init_get_bits(&s->gb, ptr, bit_length);
3863 ff_h264_decode_sei(h);
3866 init_get_bits(&s->gb, ptr, bit_length);
3867 if (ff_h264_decode_seq_parameter_set(h) < 0 &&
3868 h->is_avc && (nalsize != consumed) && nalsize) {
3869 av_log(h->s.avctx, AV_LOG_DEBUG,
3870 "SPS decoding failure, trying again with the complete NAL\n");
3871 init_get_bits(&s->gb, buf + buf_index + 1 - consumed,
3873 ff_h264_decode_seq_parameter_set(h);
3876 if (s->flags & CODEC_FLAG_LOW_DELAY ||
3877 (h->sps.bitstream_restriction_flag &&
3878 !h->sps.num_reorder_frames)) {
3879 if (s->avctx->has_b_frames > 1 || h->delayed_pic[0])
3880 av_log(avctx, AV_LOG_WARNING, "Delayed frames seen "
3881 "reenabling low delay requires a codec "
3887 if (avctx->has_b_frames < 2)
3888 avctx->has_b_frames = !s->low_delay;
3890 if (avctx->bits_per_raw_sample != h->sps.bit_depth_luma ||
3891 h->cur_chroma_format_idc != h->sps.chroma_format_idc) {
3892 if (s->avctx->codec &&
3893 s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU
3894 && (h->sps.bit_depth_luma != 8 ||
3895 h->sps.chroma_format_idc > 1)) {
3896 av_log(avctx, AV_LOG_ERROR,
3897 "VDPAU decoding does not support video "
3902 if (h->sps.bit_depth_luma >= 8 && h->sps.bit_depth_luma <= 10) {
3903 avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
3904 h->cur_chroma_format_idc = h->sps.chroma_format_idc;
3905 h->pixel_shift = h->sps.bit_depth_luma > 8;
3907 ff_h264dsp_init(&h->h264dsp, h->sps.bit_depth_luma,
3908 h->sps.chroma_format_idc);
3909 ff_h264_pred_init(&h->hpc, s->codec_id,
3910 h->sps.bit_depth_luma,
3911 h->sps.chroma_format_idc);
3912 s->dsp.dct_bits = h->sps.bit_depth_luma > 8 ? 32 : 16;
3913 ff_dsputil_init(&s->dsp, s->avctx);
3915 av_log(avctx, AV_LOG_ERROR,
3916 "Unsupported bit depth: %d\n",
3917 h->sps.bit_depth_luma);
3924 init_get_bits(&s->gb, ptr, bit_length);
3925 ff_h264_decode_picture_parameter_set(h, bit_length);
3928 case NAL_END_SEQUENCE:
3929 case NAL_END_STREAM:
3930 case NAL_FILLER_DATA:
3932 case NAL_AUXILIARY_SLICE:
3935 av_log(avctx, AV_LOG_DEBUG, "Unknown NAL code: %d (%d bits)\n",
3936 hx->nal_unit_type, bit_length);
3939 if (context_count == h->max_contexts) {
3940 execute_decode_slices(h, context_count);
3945 av_log(h->s.avctx, AV_LOG_ERROR, "decode_slice_header error\n");
3946 else if (err == 1) {
3947 /* Slice could not be decoded in parallel mode, copy down
3948 * NAL unit stuff to context 0 and restart. Note that
3949 * rbsp_buffer is not transferred, but since we no longer
3950 * run in parallel mode this should not be an issue. */
3951 h->nal_unit_type = hx->nal_unit_type;
3952 h->nal_ref_idc = hx->nal_ref_idc;
3959 execute_decode_slices(h, context_count);
3963 if (s->current_picture_ptr && s->current_picture_ptr->owner2 == s &&
3965 ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX,
3966 s->picture_structure == PICT_BOTTOM_FIELD);
3973 * Return the number of bytes consumed for building the current frame.
3975 static int get_consumed_bytes(MpegEncContext *s, int pos, int buf_size)
3978 pos = 1; // avoid infinite loops (i doubt that is needed but ...)
3979 if (pos + 10 > buf_size)
3980 pos = buf_size; // oops ;)
3985 static int decode_frame(AVCodecContext *avctx, void *data,
3986 int *data_size, AVPacket *avpkt)
3988 const uint8_t *buf = avpkt->data;
3989 int buf_size = avpkt->size;
3990 H264Context *h = avctx->priv_data;
3991 MpegEncContext *s = &h->s;
3992 AVFrame *pict = data;
3995 s->flags = avctx->flags;
3996 s->flags2 = avctx->flags2;
3998 /* end of stream, output what is still in the buffers */
4000 if (buf_size == 0) {
4004 s->current_picture_ptr = NULL;
4006 // FIXME factorize this with the output code below
4007 out = h->delayed_pic[0];
4010 h->delayed_pic[i] &&
4011 !h->delayed_pic[i]->f.key_frame &&
4012 !h->delayed_pic[i]->mmco_reset;
4014 if (h->delayed_pic[i]->poc < out->poc) {
4015 out = h->delayed_pic[i];
4019 for (i = out_idx; h->delayed_pic[i]; i++)
4020 h->delayed_pic[i] = h->delayed_pic[i + 1];
4023 *data_size = sizeof(AVFrame);
4030 buf_index = decode_nal_units(h, buf, buf_size);
4034 if (!s->current_picture_ptr && h->nal_unit_type == NAL_END_SEQUENCE) {
4039 if (!(s->flags2 & CODEC_FLAG2_CHUNKS) && !s->current_picture_ptr) {
4040 if (avctx->skip_frame >= AVDISCARD_NONREF)
4042 av_log(avctx, AV_LOG_ERROR, "no frame!\n");
4046 if (!(s->flags2 & CODEC_FLAG2_CHUNKS) ||
4047 (s->mb_y >= s->mb_height && s->mb_height)) {
4048 if (s->flags2 & CODEC_FLAG2_CHUNKS)
4049 decode_postinit(h, 1);
4053 if (!h->next_output_pic) {
4054 /* Wait for second field. */
4057 *data_size = sizeof(AVFrame);
4058 *pict = h->next_output_pic->f;
4062 assert(pict->data[0] || !*data_size);
4063 ff_print_debug_info(s, pict);
4065 return get_consumed_bytes(s, buf_index, buf_size);
4068 av_cold void ff_h264_free_context(H264Context *h)
4072 free_tables(h, 1); // FIXME cleanup init stuff perhaps
4074 for (i = 0; i < MAX_SPS_COUNT; i++)
4075 av_freep(h->sps_buffers + i);
4077 for (i = 0; i < MAX_PPS_COUNT; i++)
4078 av_freep(h->pps_buffers + i);
4081 static av_cold int h264_decode_end(AVCodecContext *avctx)
4083 H264Context *h = avctx->priv_data;
4084 MpegEncContext *s = &h->s;
4086 ff_h264_free_context(h);
4088 ff_MPV_common_end(s);
4090 // memset(h, 0, sizeof(H264Context));
4095 static const AVProfile profiles[] = {
4096 { FF_PROFILE_H264_BASELINE, "Baseline" },
4097 { FF_PROFILE_H264_CONSTRAINED_BASELINE, "Constrained Baseline" },
4098 { FF_PROFILE_H264_MAIN, "Main" },
4099 { FF_PROFILE_H264_EXTENDED, "Extended" },
4100 { FF_PROFILE_H264_HIGH, "High" },
4101 { FF_PROFILE_H264_HIGH_10, "High 10" },
4102 { FF_PROFILE_H264_HIGH_10_INTRA, "High 10 Intra" },
4103 { FF_PROFILE_H264_HIGH_422, "High 4:2:2" },
4104 { FF_PROFILE_H264_HIGH_422_INTRA, "High 4:2:2 Intra" },
4105 { FF_PROFILE_H264_HIGH_444, "High 4:4:4" },
4106 { FF_PROFILE_H264_HIGH_444_PREDICTIVE, "High 4:4:4 Predictive" },
4107 { FF_PROFILE_H264_HIGH_444_INTRA, "High 4:4:4 Intra" },
4108 { FF_PROFILE_H264_CAVLC_444, "CAVLC 4:4:4" },
4109 { FF_PROFILE_UNKNOWN },
4112 AVCodec ff_h264_decoder = {
4114 .type = AVMEDIA_TYPE_VIDEO,
4115 .id = AV_CODEC_ID_H264,
4116 .priv_data_size = sizeof(H264Context),
4117 .init = ff_h264_decode_init,
4118 .close = h264_decode_end,
4119 .decode = decode_frame,
4120 .capabilities = /*CODEC_CAP_DRAW_HORIZ_BAND |*/ CODEC_CAP_DR1 |
4121 CODEC_CAP_DELAY | CODEC_CAP_SLICE_THREADS |
4122 CODEC_CAP_FRAME_THREADS,
4124 .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10"),
4125 .init_thread_copy = ONLY_IF_THREADS_ENABLED(decode_init_thread_copy),
4126 .update_thread_context = ONLY_IF_THREADS_ENABLED(decode_update_thread_context),
4127 .profiles = NULL_IF_CONFIG_SMALL(profiles),
4130 #if CONFIG_H264_VDPAU_DECODER
4131 AVCodec ff_h264_vdpau_decoder = {
4132 .name = "h264_vdpau",
4133 .type = AVMEDIA_TYPE_VIDEO,
4134 .id = AV_CODEC_ID_H264,
4135 .priv_data_size = sizeof(H264Context),
4136 .init = ff_h264_decode_init,
4137 .close = h264_decode_end,
4138 .decode = decode_frame,
4139 .capabilities = CODEC_CAP_DR1 | CODEC_CAP_DELAY | CODEC_CAP_HWACCEL_VDPAU,
4141 .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10 (VDPAU acceleration)"),
4142 .pix_fmts = (const enum AVPixelFormat[]) { AV_PIX_FMT_VDPAU_H264,
4144 .profiles = NULL_IF_CONFIG_SMALL(profiles),