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->vdsp.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->vdsp.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->vdsp.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->vdsp.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->vdsp.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 = h->bipred_scratchpad;
656 uint8_t *tmp_cr = h->bipred_scratchpad + (16 << pixel_shift);
657 uint8_t *tmp_y = h->bipred_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->vdsp.prefetch(src[0] + off, s->linesize, 4);
739 if (chroma_idc == 3 /* yuv444 */) {
740 s->vdsp.prefetch(src[1] + off, s->linesize, 4);
741 s->vdsp.prefetch(src[2] + off, s->linesize, 4);
743 off = ((mx >> 1) << pixel_shift) +
744 ((my >> 1) + (s->mb_x & 7)) * s->uvlinesize +
746 s->vdsp.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->bipred_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 if (!h->dequant4_coeff[0])
902 init_dequant_tables(h);
912 * Mimic alloc_tables(), but for every context thread.
914 static void clone_tables(H264Context *dst, H264Context *src, int i)
916 MpegEncContext *const s = &src->s;
917 dst->intra4x4_pred_mode = src->intra4x4_pred_mode + i * 8 * 2 * s->mb_stride;
918 dst->non_zero_count = src->non_zero_count;
919 dst->slice_table = src->slice_table;
920 dst->cbp_table = src->cbp_table;
921 dst->mb2b_xy = src->mb2b_xy;
922 dst->mb2br_xy = src->mb2br_xy;
923 dst->chroma_pred_mode_table = src->chroma_pred_mode_table;
924 dst->mvd_table[0] = src->mvd_table[0] + i * 8 * 2 * s->mb_stride;
925 dst->mvd_table[1] = src->mvd_table[1] + i * 8 * 2 * s->mb_stride;
926 dst->direct_table = src->direct_table;
927 dst->list_counts = src->list_counts;
928 dst->bipred_scratchpad = NULL;
929 ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma,
930 src->sps.chroma_format_idc);
935 * Allocate buffers which are not shared amongst multiple threads.
937 static int context_init(H264Context *h)
939 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0],
940 h->s.mb_width * 16 * 3 * sizeof(uint8_t) * 2, fail)
941 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1],
942 h->s.mb_width * 16 * 3 * sizeof(uint8_t) * 2, fail)
944 h->ref_cache[0][scan8[5] + 1] =
945 h->ref_cache[0][scan8[7] + 1] =
946 h->ref_cache[0][scan8[13] + 1] =
947 h->ref_cache[1][scan8[5] + 1] =
948 h->ref_cache[1][scan8[7] + 1] =
949 h->ref_cache[1][scan8[13] + 1] = PART_NOT_AVAILABLE;
954 return -1; // free_tables will clean up for us
957 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size,
958 int parse_extradata);
960 static av_cold void common_init(H264Context *h)
962 MpegEncContext *const s = &h->s;
964 s->width = s->avctx->width;
965 s->height = s->avctx->height;
966 s->codec_id = s->avctx->codec->id;
968 ff_h264dsp_init(&h->h264dsp, 8, 1);
969 ff_h264_pred_init(&h->hpc, s->codec_id, 8, 1);
971 h->dequant_coeff_pps = -1;
972 s->unrestricted_mv = 1;
974 /* needed so that IDCT permutation is known early */
975 ff_dsputil_init(&s->dsp, s->avctx);
976 ff_videodsp_init(&s->vdsp, 8);
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, 1) < 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, 1) < 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, 1) < 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));
1130 h->s.context_initialized = 0;
1135 #define copy_fields(to, from, start_field, end_field) \
1136 memcpy(&to->start_field, &from->start_field, \
1137 (char *)&to->end_field - (char *)&to->start_field)
1139 static int h264_slice_header_init(H264Context *, int);
1141 static int h264_set_parameter_from_sps(H264Context *h);
1143 static int decode_update_thread_context(AVCodecContext *dst,
1144 const AVCodecContext *src)
1146 H264Context *h = dst->priv_data, *h1 = src->priv_data;
1147 MpegEncContext *const s = &h->s, *const s1 = &h1->s;
1148 int inited = s->context_initialized, err;
1151 if (dst == src || !s1->context_initialized)
1155 (s->width != s1->width ||
1156 s->height != s1->height ||
1157 s->mb_width != s1->mb_width ||
1158 s->mb_height != s1->mb_height ||
1159 h->sps.bit_depth_luma != h1->sps.bit_depth_luma ||
1160 h->sps.chroma_format_idc != h1->sps.chroma_format_idc ||
1161 h->sps.colorspace != h1->sps.colorspace)) {
1163 av_freep(&h->bipred_scratchpad);
1165 s->width = s1->width;
1166 s->height = s1->height;
1167 s->mb_height = s1->mb_height;
1168 h->b_stride = h1->b_stride;
1170 if ((err = h264_slice_header_init(h, 1)) < 0) {
1171 av_log(h->s.avctx, AV_LOG_ERROR, "h264_slice_header_init() failed");
1174 h->context_reinitialized = 1;
1176 /* update linesize on resize for h264. The h264 decoder doesn't
1177 * necessarily call ff_MPV_frame_start in the new thread */
1178 s->linesize = s1->linesize;
1179 s->uvlinesize = s1->uvlinesize;
1181 /* copy block_offset since frame_start may not be called */
1182 memcpy(h->block_offset, h1->block_offset, sizeof(h->block_offset));
1183 h264_set_parameter_from_sps(h);
1186 err = ff_mpeg_update_thread_context(dst, src);
1191 for (i = 0; i < MAX_SPS_COUNT; i++)
1192 av_freep(h->sps_buffers + i);
1194 for (i = 0; i < MAX_PPS_COUNT; i++)
1195 av_freep(h->pps_buffers + i);
1197 // copy all fields after MpegEnc
1198 memcpy(&h->s + 1, &h1->s + 1,
1199 sizeof(H264Context) - sizeof(MpegEncContext));
1200 memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1201 memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1202 if (ff_h264_alloc_tables(h) < 0) {
1203 av_log(dst, AV_LOG_ERROR, "Could not allocate memory for h264\n");
1204 return AVERROR(ENOMEM);
1208 for (i = 0; i < 2; i++) {
1209 h->rbsp_buffer[i] = NULL;
1210 h->rbsp_buffer_size[i] = 0;
1212 h->bipred_scratchpad = NULL;
1214 h->thread_context[0] = h;
1216 s->dsp.clear_blocks(h->mb);
1217 s->dsp.clear_blocks(h->mb + (24 * 16 << h->pixel_shift));
1220 /* frame_start may not be called for the next thread (if it's decoding
1221 * a bottom field) so this has to be allocated here */
1222 if (!h->bipred_scratchpad)
1223 h->bipred_scratchpad = av_malloc(16 * 6 * s->linesize);
1225 // extradata/NAL handling
1226 h->is_avc = h1->is_avc;
1229 copy_parameter_set((void **)h->sps_buffers, (void **)h1->sps_buffers,
1230 MAX_SPS_COUNT, sizeof(SPS));
1232 copy_parameter_set((void **)h->pps_buffers, (void **)h1->pps_buffers,
1233 MAX_PPS_COUNT, sizeof(PPS));
1236 // Dequantization matrices
1237 // FIXME these are big - can they be only copied when PPS changes?
1238 copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1240 for (i = 0; i < 6; i++)
1241 h->dequant4_coeff[i] = h->dequant4_buffer[0] +
1242 (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1244 for (i = 0; i < 6; i++)
1245 h->dequant8_coeff[i] = h->dequant8_buffer[0] +
1246 (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1248 h->dequant_coeff_pps = h1->dequant_coeff_pps;
1251 copy_fields(h, h1, poc_lsb, redundant_pic_count);
1254 copy_fields(h, h1, ref_count, list_count);
1255 copy_fields(h, h1, ref_list, intra_gb);
1256 copy_fields(h, h1, short_ref, cabac_init_idc);
1258 copy_picture_range(h->short_ref, h1->short_ref, 32, s, s1);
1259 copy_picture_range(h->long_ref, h1->long_ref, 32, s, s1);
1260 copy_picture_range(h->delayed_pic, h1->delayed_pic,
1261 MAX_DELAYED_PIC_COUNT + 2, s, s1);
1263 h->last_slice_type = h1->last_slice_type;
1265 if (!s->current_picture_ptr)
1268 if (!s->droppable) {
1269 err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1270 h->prev_poc_msb = h->poc_msb;
1271 h->prev_poc_lsb = h->poc_lsb;
1273 h->prev_frame_num_offset = h->frame_num_offset;
1274 h->prev_frame_num = h->frame_num;
1275 h->outputed_poc = h->next_outputed_poc;
1280 int ff_h264_frame_start(H264Context *h)
1282 MpegEncContext *const s = &h->s;
1284 const int pixel_shift = h->pixel_shift;
1286 if (ff_MPV_frame_start(s, s->avctx) < 0)
1288 ff_er_frame_start(s);
1290 * ff_MPV_frame_start uses pict_type to derive key_frame.
1291 * This is incorrect for H.264; IDR markings must be used.
1292 * Zero here; IDR markings per slice in frame or fields are ORed in later.
1293 * See decode_nal_units().
1295 s->current_picture_ptr->f.key_frame = 0;
1296 s->current_picture_ptr->mmco_reset = 0;
1298 assert(s->linesize && s->uvlinesize);
1300 for (i = 0; i < 16; i++) {
1301 h->block_offset[i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * s->linesize * ((scan8[i] - scan8[0]) >> 3);
1302 h->block_offset[48 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * s->linesize * ((scan8[i] - scan8[0]) >> 3);
1304 for (i = 0; i < 16; i++) {
1305 h->block_offset[16 + i] =
1306 h->block_offset[32 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * s->uvlinesize * ((scan8[i] - scan8[0]) >> 3);
1307 h->block_offset[48 + 16 + i] =
1308 h->block_offset[48 + 32 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * s->uvlinesize * ((scan8[i] - scan8[0]) >> 3);
1311 /* can't be in alloc_tables because linesize isn't known there.
1312 * FIXME: redo bipred weight to not require extra buffer? */
1313 for (i = 0; i < s->slice_context_count; i++)
1314 if (h->thread_context[i] && !h->thread_context[i]->bipred_scratchpad)
1315 h->thread_context[i]->bipred_scratchpad = av_malloc(16 * 6 * s->linesize);
1317 /* Some macroblocks can be accessed before they're available in case
1318 * of lost slices, MBAFF or threading. */
1319 memset(h->slice_table, -1,
1320 (s->mb_height * s->mb_stride - 1) * sizeof(*h->slice_table));
1322 // s->decode = (s->flags & CODEC_FLAG_PSNR) || !s->encoding ||
1323 // s->current_picture.f.reference /* || h->contains_intra */ || 1;
1325 /* We mark the current picture as non-reference after allocating it, so
1326 * that if we break out due to an error it can be released automatically
1327 * in the next ff_MPV_frame_start().
1328 * SVQ3 as well as most other codecs have only last/next/current and thus
1329 * get released even with set reference, besides SVQ3 and others do not
1330 * mark frames as reference later "naturally". */
1331 if (s->codec_id != AV_CODEC_ID_SVQ3)
1332 s->current_picture_ptr->f.reference = 0;
1334 s->current_picture_ptr->field_poc[0] =
1335 s->current_picture_ptr->field_poc[1] = INT_MAX;
1337 h->next_output_pic = NULL;
1339 assert(s->current_picture_ptr->long_ref == 0);
1345 * Run setup operations that must be run after slice header decoding.
1346 * This includes finding the next displayed frame.
1348 * @param h h264 master context
1349 * @param setup_finished enough NALs have been read that we can call
1350 * ff_thread_finish_setup()
1352 static void decode_postinit(H264Context *h, int setup_finished)
1354 MpegEncContext *const s = &h->s;
1355 Picture *out = s->current_picture_ptr;
1356 Picture *cur = s->current_picture_ptr;
1357 int i, pics, out_of_order, out_idx;
1358 int invalid = 0, cnt = 0;
1360 s->current_picture_ptr->f.qscale_type = FF_QSCALE_TYPE_H264;
1361 s->current_picture_ptr->f.pict_type = s->pict_type;
1363 if (h->next_output_pic)
1366 if (cur->field_poc[0] == INT_MAX || cur->field_poc[1] == INT_MAX) {
1367 /* FIXME: if we have two PAFF fields in one packet, we can't start
1368 * the next thread here. If we have one field per packet, we can.
1369 * The check in decode_nal_units() is not good enough to find this
1370 * yet, so we assume the worst for now. */
1371 // if (setup_finished)
1372 // ff_thread_finish_setup(s->avctx);
1376 cur->f.interlaced_frame = 0;
1377 cur->f.repeat_pict = 0;
1379 /* Signal interlacing information externally. */
1380 /* Prioritize picture timing SEI information over used
1381 * decoding process if it exists. */
1383 if (h->sps.pic_struct_present_flag) {
1384 switch (h->sei_pic_struct) {
1385 case SEI_PIC_STRUCT_FRAME:
1387 case SEI_PIC_STRUCT_TOP_FIELD:
1388 case SEI_PIC_STRUCT_BOTTOM_FIELD:
1389 cur->f.interlaced_frame = 1;
1391 case SEI_PIC_STRUCT_TOP_BOTTOM:
1392 case SEI_PIC_STRUCT_BOTTOM_TOP:
1393 if (FIELD_OR_MBAFF_PICTURE)
1394 cur->f.interlaced_frame = 1;
1396 // try to flag soft telecine progressive
1397 cur->f.interlaced_frame = h->prev_interlaced_frame;
1399 case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1400 case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1401 /* Signal the possibility of telecined film externally
1402 * (pic_struct 5,6). From these hints, let the applications
1403 * decide if they apply deinterlacing. */
1404 cur->f.repeat_pict = 1;
1406 case SEI_PIC_STRUCT_FRAME_DOUBLING:
1407 // Force progressive here, doubling interlaced frame is a bad idea.
1408 cur->f.repeat_pict = 2;
1410 case SEI_PIC_STRUCT_FRAME_TRIPLING:
1411 cur->f.repeat_pict = 4;
1415 if ((h->sei_ct_type & 3) &&
1416 h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1417 cur->f.interlaced_frame = (h->sei_ct_type & (1 << 1)) != 0;
1419 /* Derive interlacing flag from used decoding process. */
1420 cur->f.interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1422 h->prev_interlaced_frame = cur->f.interlaced_frame;
1424 if (cur->field_poc[0] != cur->field_poc[1]) {
1425 /* Derive top_field_first from field pocs. */
1426 cur->f.top_field_first = cur->field_poc[0] < cur->field_poc[1];
1428 if (cur->f.interlaced_frame || h->sps.pic_struct_present_flag) {
1429 /* Use picture timing SEI information. Even if it is a
1430 * information of a past frame, better than nothing. */
1431 if (h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM ||
1432 h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1433 cur->f.top_field_first = 1;
1435 cur->f.top_field_first = 0;
1437 /* Most likely progressive */
1438 cur->f.top_field_first = 0;
1442 // FIXME do something with unavailable reference frames
1444 /* Sort B-frames into display order */
1446 if (h->sps.bitstream_restriction_flag &&
1447 s->avctx->has_b_frames < h->sps.num_reorder_frames) {
1448 s->avctx->has_b_frames = h->sps.num_reorder_frames;
1452 if (s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT &&
1453 !h->sps.bitstream_restriction_flag) {
1454 s->avctx->has_b_frames = MAX_DELAYED_PIC_COUNT - 1;
1459 while (h->delayed_pic[pics])
1462 assert(pics <= MAX_DELAYED_PIC_COUNT);
1464 h->delayed_pic[pics++] = cur;
1465 if (cur->f.reference == 0)
1466 cur->f.reference = DELAYED_PIC_REF;
1468 /* Frame reordering. This code takes pictures from coding order and sorts
1469 * them by their incremental POC value into display order. It supports POC
1470 * gaps, MMCO reset codes and random resets.
1471 * A "display group" can start either with a IDR frame (f.key_frame = 1),
1472 * and/or can be closed down with a MMCO reset code. In sequences where
1473 * there is no delay, we can't detect that (since the frame was already
1474 * output to the user), so we also set h->mmco_reset to detect the MMCO
1476 * FIXME: if we detect insufficient delays (as per s->avctx->has_b_frames),
1477 * we increase the delay between input and output. All frames affected by
1478 * the lag (e.g. those that should have been output before another frame
1479 * that we already returned to the user) will be dropped. This is a bug
1480 * that we will fix later. */
1481 for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++) {
1482 cnt += out->poc < h->last_pocs[i];
1483 invalid += out->poc == INT_MIN;
1485 if (!h->mmco_reset && !cur->f.key_frame &&
1486 cnt + invalid == MAX_DELAYED_PIC_COUNT && cnt > 0) {
1489 h->delayed_pic[pics - 2]->mmco_reset = 2;
1491 if (h->mmco_reset || cur->f.key_frame) {
1492 for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
1493 h->last_pocs[i] = INT_MIN;
1495 invalid = MAX_DELAYED_PIC_COUNT;
1497 out = h->delayed_pic[0];
1499 for (i = 1; i < MAX_DELAYED_PIC_COUNT &&
1500 h->delayed_pic[i] &&
1501 !h->delayed_pic[i - 1]->mmco_reset &&
1502 !h->delayed_pic[i]->f.key_frame;
1504 if (h->delayed_pic[i]->poc < out->poc) {
1505 out = h->delayed_pic[i];
1508 if (s->avctx->has_b_frames == 0 &&
1509 (h->delayed_pic[0]->f.key_frame || h->mmco_reset))
1510 h->next_outputed_poc = INT_MIN;
1511 out_of_order = !out->f.key_frame && !h->mmco_reset &&
1512 (out->poc < h->next_outputed_poc);
1514 if (h->sps.bitstream_restriction_flag &&
1515 s->avctx->has_b_frames >= h->sps.num_reorder_frames) {
1516 } else if (out_of_order && pics - 1 == s->avctx->has_b_frames &&
1517 s->avctx->has_b_frames < MAX_DELAYED_PIC_COUNT) {
1518 if (invalid + cnt < MAX_DELAYED_PIC_COUNT) {
1519 s->avctx->has_b_frames = FFMAX(s->avctx->has_b_frames, cnt);
1522 } else if (s->low_delay &&
1523 ((h->next_outputed_poc != INT_MIN &&
1524 out->poc > h->next_outputed_poc + 2) ||
1525 cur->f.pict_type == AV_PICTURE_TYPE_B)) {
1527 s->avctx->has_b_frames++;
1530 if (pics > s->avctx->has_b_frames) {
1531 out->f.reference &= ~DELAYED_PIC_REF;
1532 // for frame threading, the owner must be the second field's thread or
1533 // else the first thread can release the picture and reuse it unsafely
1535 for (i = out_idx; h->delayed_pic[i]; i++)
1536 h->delayed_pic[i] = h->delayed_pic[i + 1];
1538 memmove(h->last_pocs, &h->last_pocs[1],
1539 sizeof(*h->last_pocs) * (MAX_DELAYED_PIC_COUNT - 1));
1540 h->last_pocs[MAX_DELAYED_PIC_COUNT - 1] = cur->poc;
1541 if (!out_of_order && pics > s->avctx->has_b_frames) {
1542 h->next_output_pic = out;
1543 if (out->mmco_reset) {
1545 h->next_outputed_poc = out->poc;
1546 h->delayed_pic[out_idx - 1]->mmco_reset = out->mmco_reset;
1548 h->next_outputed_poc = INT_MIN;
1551 if (out_idx == 0 && pics > 1 && h->delayed_pic[0]->f.key_frame) {
1552 h->next_outputed_poc = INT_MIN;
1554 h->next_outputed_poc = out->poc;
1559 av_log(s->avctx, AV_LOG_DEBUG, "no picture\n");
1563 ff_thread_finish_setup(s->avctx);
1566 static av_always_inline void backup_mb_border(H264Context *h, uint8_t *src_y,
1567 uint8_t *src_cb, uint8_t *src_cr,
1568 int linesize, int uvlinesize,
1571 MpegEncContext *const s = &h->s;
1572 uint8_t *top_border;
1574 const int pixel_shift = h->pixel_shift;
1575 int chroma444 = CHROMA444;
1576 int chroma422 = CHROMA422;
1579 src_cb -= uvlinesize;
1580 src_cr -= uvlinesize;
1582 if (!simple && FRAME_MBAFF) {
1585 top_border = h->top_borders[0][s->mb_x];
1586 AV_COPY128(top_border, src_y + 15 * linesize);
1588 AV_COPY128(top_border + 16, src_y + 15 * linesize + 16);
1589 if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1592 AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
1593 AV_COPY128(top_border + 48, src_cb + 15 * uvlinesize + 16);
1594 AV_COPY128(top_border + 64, src_cr + 15 * uvlinesize);
1595 AV_COPY128(top_border + 80, src_cr + 15 * uvlinesize + 16);
1597 AV_COPY128(top_border + 16, src_cb + 15 * uvlinesize);
1598 AV_COPY128(top_border + 32, src_cr + 15 * uvlinesize);
1600 } else if (chroma422) {
1602 AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
1603 AV_COPY128(top_border + 48, src_cr + 15 * uvlinesize);
1605 AV_COPY64(top_border + 16, src_cb + 15 * uvlinesize);
1606 AV_COPY64(top_border + 24, src_cr + 15 * uvlinesize);
1610 AV_COPY128(top_border + 32, src_cb + 7 * uvlinesize);
1611 AV_COPY128(top_border + 48, src_cr + 7 * uvlinesize);
1613 AV_COPY64(top_border + 16, src_cb + 7 * uvlinesize);
1614 AV_COPY64(top_border + 24, src_cr + 7 * uvlinesize);
1619 } else if (MB_MBAFF) {
1625 top_border = h->top_borders[top_idx][s->mb_x];
1626 /* There are two lines saved, the line above the top macroblock
1627 * of a pair, and the line above the bottom macroblock. */
1628 AV_COPY128(top_border, src_y + 16 * linesize);
1630 AV_COPY128(top_border + 16, src_y + 16 * linesize + 16);
1632 if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1635 AV_COPY128(top_border + 32, src_cb + 16 * linesize);
1636 AV_COPY128(top_border + 48, src_cb + 16 * linesize + 16);
1637 AV_COPY128(top_border + 64, src_cr + 16 * linesize);
1638 AV_COPY128(top_border + 80, src_cr + 16 * linesize + 16);
1640 AV_COPY128(top_border + 16, src_cb + 16 * linesize);
1641 AV_COPY128(top_border + 32, src_cr + 16 * linesize);
1643 } else if (chroma422) {
1645 AV_COPY128(top_border + 32, src_cb + 16 * uvlinesize);
1646 AV_COPY128(top_border + 48, src_cr + 16 * uvlinesize);
1648 AV_COPY64(top_border + 16, src_cb + 16 * uvlinesize);
1649 AV_COPY64(top_border + 24, src_cr + 16 * uvlinesize);
1653 AV_COPY128(top_border + 32, src_cb + 8 * uvlinesize);
1654 AV_COPY128(top_border + 48, src_cr + 8 * uvlinesize);
1656 AV_COPY64(top_border + 16, src_cb + 8 * uvlinesize);
1657 AV_COPY64(top_border + 24, src_cr + 8 * uvlinesize);
1663 static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1664 uint8_t *src_cb, uint8_t *src_cr,
1665 int linesize, int uvlinesize,
1666 int xchg, int chroma444,
1667 int simple, int pixel_shift)
1669 MpegEncContext *const s = &h->s;
1670 int deblock_topleft;
1673 uint8_t *top_border_m1;
1674 uint8_t *top_border;
1676 if (!simple && FRAME_MBAFF) {
1681 top_idx = MB_MBAFF ? 0 : 1;
1685 if (h->deblocking_filter == 2) {
1686 deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1687 deblock_top = h->top_type;
1689 deblock_topleft = (s->mb_x > 0);
1690 deblock_top = (s->mb_y > !!MB_FIELD);
1693 src_y -= linesize + 1 + pixel_shift;
1694 src_cb -= uvlinesize + 1 + pixel_shift;
1695 src_cr -= uvlinesize + 1 + pixel_shift;
1697 top_border_m1 = h->top_borders[top_idx][s->mb_x - 1];
1698 top_border = h->top_borders[top_idx][s->mb_x];
1700 #define XCHG(a, b, xchg) \
1701 if (pixel_shift) { \
1703 AV_SWAP64(b + 0, a + 0); \
1704 AV_SWAP64(b + 8, a + 8); \
1714 if (deblock_topleft) {
1715 XCHG(top_border_m1 + (8 << pixel_shift),
1716 src_y - (7 << pixel_shift), 1);
1718 XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1719 XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1720 if (s->mb_x + 1 < s->mb_width) {
1721 XCHG(h->top_borders[top_idx][s->mb_x + 1],
1722 src_y + (17 << pixel_shift), 1);
1725 if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1727 if (deblock_topleft) {
1728 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1729 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1731 XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1732 XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1733 XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1734 XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1735 if (s->mb_x + 1 < s->mb_width) {
1736 XCHG(h->top_borders[top_idx][s->mb_x + 1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1737 XCHG(h->top_borders[top_idx][s->mb_x + 1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1741 if (deblock_topleft) {
1742 XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1743 XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1745 XCHG(top_border + (16 << pixel_shift), src_cb + 1 + pixel_shift, 1);
1746 XCHG(top_border + (24 << pixel_shift), src_cr + 1 + pixel_shift, 1);
1752 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth,
1755 if (high_bit_depth) {
1756 return AV_RN32A(((int32_t *)mb) + index);
1758 return AV_RN16A(mb + index);
1761 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth,
1762 int index, int value)
1764 if (high_bit_depth) {
1765 AV_WN32A(((int32_t *)mb) + index, value);
1767 AV_WN16A(mb + index, value);
1770 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h,
1771 int mb_type, int is_h264,
1773 int transform_bypass,
1777 uint8_t *dest_y, int p)
1779 MpegEncContext *const s = &h->s;
1780 void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1781 void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1783 int qscale = p == 0 ? s->qscale : h->chroma_qp[p - 1];
1784 block_offset += 16 * p;
1785 if (IS_INTRA4x4(mb_type)) {
1786 if (simple || !s->encoding) {
1787 if (IS_8x8DCT(mb_type)) {
1788 if (transform_bypass) {
1790 idct_add = s->dsp.add_pixels8;
1792 idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1793 idct_add = h->h264dsp.h264_idct8_add;
1795 for (i = 0; i < 16; i += 4) {
1796 uint8_t *const ptr = dest_y + block_offset[i];
1797 const int dir = h->intra4x4_pred_mode_cache[scan8[i]];
1798 if (transform_bypass && h->sps.profile_idc == 244 && dir <= 1) {
1799 h->hpc.pred8x8l_add[dir](ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1801 const int nnz = h->non_zero_count_cache[scan8[i + p * 16]];
1802 h->hpc.pred8x8l[dir](ptr, (h->topleft_samples_available << i) & 0x8000,
1803 (h->topright_samples_available << i) & 0x4000, linesize);
1805 if (nnz == 1 && dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
1806 idct_dc_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1808 idct_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1813 if (transform_bypass) {
1815 idct_add = s->dsp.add_pixels4;
1817 idct_dc_add = h->h264dsp.h264_idct_dc_add;
1818 idct_add = h->h264dsp.h264_idct_add;
1820 for (i = 0; i < 16; i++) {
1821 uint8_t *const ptr = dest_y + block_offset[i];
1822 const int dir = h->intra4x4_pred_mode_cache[scan8[i]];
1824 if (transform_bypass && h->sps.profile_idc == 244 && dir <= 1) {
1825 h->hpc.pred4x4_add[dir](ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1830 if (dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED) {
1831 const int topright_avail = (h->topright_samples_available << i) & 0x8000;
1832 assert(s->mb_y || linesize <= block_offset[i]);
1833 if (!topright_avail) {
1835 tr_high = ((uint16_t *)ptr)[3 - linesize / 2] * 0x0001000100010001ULL;
1836 topright = (uint8_t *)&tr_high;
1838 tr = ptr[3 - linesize] * 0x01010101u;
1839 topright = (uint8_t *)&tr;
1842 topright = ptr + (4 << pixel_shift) - linesize;
1846 h->hpc.pred4x4[dir](ptr, topright, linesize);
1847 nnz = h->non_zero_count_cache[scan8[i + p * 16]];
1850 if (nnz == 1 && dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
1851 idct_dc_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1853 idct_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1854 } else if (CONFIG_SVQ3_DECODER)
1855 ff_svq3_add_idct_c(ptr, h->mb + i * 16 + p * 256, linesize, qscale, 0);
1862 h->hpc.pred16x16[h->intra16x16_pred_mode](dest_y, linesize);
1864 if (h->non_zero_count_cache[scan8[LUMA_DC_BLOCK_INDEX + p]]) {
1865 if (!transform_bypass)
1866 h->h264dsp.h264_luma_dc_dequant_idct(h->mb + (p * 256 << pixel_shift),
1868 h->dequant4_coeff[p][qscale][0]);
1870 static const uint8_t dc_mapping[16] = {
1871 0 * 16, 1 * 16, 4 * 16, 5 * 16,
1872 2 * 16, 3 * 16, 6 * 16, 7 * 16,
1873 8 * 16, 9 * 16, 12 * 16, 13 * 16,
1874 10 * 16, 11 * 16, 14 * 16, 15 * 16 };
1875 for (i = 0; i < 16; i++)
1876 dctcoef_set(h->mb + (p * 256 << pixel_shift),
1877 pixel_shift, dc_mapping[i],
1878 dctcoef_get(h->mb_luma_dc[p],
1882 } else if (CONFIG_SVQ3_DECODER)
1883 ff_svq3_luma_dc_dequant_idct_c(h->mb + p * 256,
1884 h->mb_luma_dc[p], qscale);
1888 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type,
1889 int is_h264, int simple,
1890 int transform_bypass,
1894 uint8_t *dest_y, int p)
1896 MpegEncContext *const s = &h->s;
1897 void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1899 block_offset += 16 * p;
1900 if (!IS_INTRA4x4(mb_type)) {
1902 if (IS_INTRA16x16(mb_type)) {
1903 if (transform_bypass) {
1904 if (h->sps.profile_idc == 244 &&
1905 (h->intra16x16_pred_mode == VERT_PRED8x8 ||
1906 h->intra16x16_pred_mode == HOR_PRED8x8)) {
1907 h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset,
1908 h->mb + (p * 256 << pixel_shift),
1911 for (i = 0; i < 16; i++)
1912 if (h->non_zero_count_cache[scan8[i + p * 16]] ||
1913 dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
1914 s->dsp.add_pixels4(dest_y + block_offset[i],
1915 h->mb + (i * 16 + p * 256 << pixel_shift),
1919 h->h264dsp.h264_idct_add16intra(dest_y, block_offset,
1920 h->mb + (p * 256 << pixel_shift),
1922 h->non_zero_count_cache + p * 5 * 8);
1924 } else if (h->cbp & 15) {
1925 if (transform_bypass) {
1926 const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1927 idct_add = IS_8x8DCT(mb_type) ? s->dsp.add_pixels8
1928 : s->dsp.add_pixels4;
1929 for (i = 0; i < 16; i += di)
1930 if (h->non_zero_count_cache[scan8[i + p * 16]])
1931 idct_add(dest_y + block_offset[i],
1932 h->mb + (i * 16 + p * 256 << pixel_shift),
1935 if (IS_8x8DCT(mb_type))
1936 h->h264dsp.h264_idct8_add4(dest_y, block_offset,
1937 h->mb + (p * 256 << pixel_shift),
1939 h->non_zero_count_cache + p * 5 * 8);
1941 h->h264dsp.h264_idct_add16(dest_y, block_offset,
1942 h->mb + (p * 256 << pixel_shift),
1944 h->non_zero_count_cache + p * 5 * 8);
1947 } else if (CONFIG_SVQ3_DECODER) {
1948 for (i = 0; i < 16; i++)
1949 if (h->non_zero_count_cache[scan8[i + p * 16]] || h->mb[i * 16 + p * 256]) {
1950 // FIXME benchmark weird rule, & below
1951 uint8_t *const ptr = dest_y + block_offset[i];
1952 ff_svq3_add_idct_c(ptr, h->mb + i * 16 + p * 256, linesize,
1953 s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1961 #include "h264_mb_template.c"
1965 #include "h264_mb_template.c"
1969 #include "h264_mb_template.c"
1971 void ff_h264_hl_decode_mb(H264Context *h)
1973 MpegEncContext *const s = &h->s;
1974 const int mb_xy = h->mb_xy;
1975 const int mb_type = s->current_picture.f.mb_type[mb_xy];
1976 int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
1979 if (is_complex || h->pixel_shift)
1980 hl_decode_mb_444_complex(h);
1982 hl_decode_mb_444_simple_8(h);
1983 } else if (is_complex) {
1984 hl_decode_mb_complex(h);
1985 } else if (h->pixel_shift) {
1986 hl_decode_mb_simple_16(h);
1988 hl_decode_mb_simple_8(h);
1991 static int pred_weight_table(H264Context *h)
1993 MpegEncContext *const s = &h->s;
1995 int luma_def, chroma_def;
1998 h->use_weight_chroma = 0;
1999 h->luma_log2_weight_denom = get_ue_golomb(&s->gb);
2000 if (h->sps.chroma_format_idc)
2001 h->chroma_log2_weight_denom = get_ue_golomb(&s->gb);
2002 luma_def = 1 << h->luma_log2_weight_denom;
2003 chroma_def = 1 << h->chroma_log2_weight_denom;
2005 for (list = 0; list < 2; list++) {
2006 h->luma_weight_flag[list] = 0;
2007 h->chroma_weight_flag[list] = 0;
2008 for (i = 0; i < h->ref_count[list]; i++) {
2009 int luma_weight_flag, chroma_weight_flag;
2011 luma_weight_flag = get_bits1(&s->gb);
2012 if (luma_weight_flag) {
2013 h->luma_weight[i][list][0] = get_se_golomb(&s->gb);
2014 h->luma_weight[i][list][1] = get_se_golomb(&s->gb);
2015 if (h->luma_weight[i][list][0] != luma_def ||
2016 h->luma_weight[i][list][1] != 0) {
2018 h->luma_weight_flag[list] = 1;
2021 h->luma_weight[i][list][0] = luma_def;
2022 h->luma_weight[i][list][1] = 0;
2025 if (h->sps.chroma_format_idc) {
2026 chroma_weight_flag = get_bits1(&s->gb);
2027 if (chroma_weight_flag) {
2029 for (j = 0; j < 2; j++) {
2030 h->chroma_weight[i][list][j][0] = get_se_golomb(&s->gb);
2031 h->chroma_weight[i][list][j][1] = get_se_golomb(&s->gb);
2032 if (h->chroma_weight[i][list][j][0] != chroma_def ||
2033 h->chroma_weight[i][list][j][1] != 0) {
2034 h->use_weight_chroma = 1;
2035 h->chroma_weight_flag[list] = 1;
2040 for (j = 0; j < 2; j++) {
2041 h->chroma_weight[i][list][j][0] = chroma_def;
2042 h->chroma_weight[i][list][j][1] = 0;
2047 if (h->slice_type_nos != AV_PICTURE_TYPE_B)
2050 h->use_weight = h->use_weight || h->use_weight_chroma;
2055 * Initialize implicit_weight table.
2056 * @param field 0/1 initialize the weight for interlaced MBAFF
2057 * -1 initializes the rest
2059 static void implicit_weight_table(H264Context *h, int field)
2061 MpegEncContext *const s = &h->s;
2062 int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2064 for (i = 0; i < 2; i++) {
2065 h->luma_weight_flag[i] = 0;
2066 h->chroma_weight_flag[i] = 0;
2070 if (s->picture_structure == PICT_FRAME) {
2071 cur_poc = s->current_picture_ptr->poc;
2073 cur_poc = s->current_picture_ptr->field_poc[s->picture_structure - 1];
2075 if (h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF &&
2076 h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2 * cur_poc) {
2078 h->use_weight_chroma = 0;
2082 ref_count0 = h->ref_count[0];
2083 ref_count1 = h->ref_count[1];
2085 cur_poc = s->current_picture_ptr->field_poc[field];
2087 ref_count0 = 16 + 2 * h->ref_count[0];
2088 ref_count1 = 16 + 2 * h->ref_count[1];
2092 h->use_weight_chroma = 2;
2093 h->luma_log2_weight_denom = 5;
2094 h->chroma_log2_weight_denom = 5;
2096 for (ref0 = ref_start; ref0 < ref_count0; ref0++) {
2097 int poc0 = h->ref_list[0][ref0].poc;
2098 for (ref1 = ref_start; ref1 < ref_count1; ref1++) {
2100 if (!h->ref_list[0][ref0].long_ref && !h->ref_list[1][ref1].long_ref) {
2101 int poc1 = h->ref_list[1][ref1].poc;
2102 int td = av_clip(poc1 - poc0, -128, 127);
2104 int tb = av_clip(cur_poc - poc0, -128, 127);
2105 int tx = (16384 + (FFABS(td) >> 1)) / td;
2106 int dist_scale_factor = (tb * tx + 32) >> 8;
2107 if (dist_scale_factor >= -64 && dist_scale_factor <= 128)
2108 w = 64 - dist_scale_factor;
2112 h->implicit_weight[ref0][ref1][0] =
2113 h->implicit_weight[ref0][ref1][1] = w;
2115 h->implicit_weight[ref0][ref1][field] = w;
2122 * instantaneous decoder refresh.
2124 static void idr(H264Context *h)
2126 ff_h264_remove_all_refs(h);
2127 h->prev_frame_num = 0;
2128 h->prev_frame_num_offset = 0;
2130 h->prev_poc_lsb = 0;
2133 /* forget old pics after a seek */
2134 static void flush_change(H264Context *h)
2137 for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
2138 h->last_pocs[i] = INT_MIN;
2139 h->outputed_poc = h->next_outputed_poc = INT_MIN;
2140 h->prev_interlaced_frame = 1;
2142 if (h->s.current_picture_ptr)
2143 h->s.current_picture_ptr->f.reference = 0;
2144 h->s.first_field = 0;
2145 memset(h->ref_list[0], 0, sizeof(h->ref_list[0]));
2146 memset(h->ref_list[1], 0, sizeof(h->ref_list[1]));
2147 memset(h->default_ref_list[0], 0, sizeof(h->default_ref_list[0]));
2148 memset(h->default_ref_list[1], 0, sizeof(h->default_ref_list[1]));
2149 ff_h264_reset_sei(h);
2152 /* forget old pics after a seek */
2153 static void flush_dpb(AVCodecContext *avctx)
2155 H264Context *h = avctx->priv_data;
2158 for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++) {
2159 if (h->delayed_pic[i])
2160 h->delayed_pic[i]->f.reference = 0;
2161 h->delayed_pic[i] = NULL;
2165 ff_mpeg_flush(avctx);
2168 static int init_poc(H264Context *h)
2170 MpegEncContext *const s = &h->s;
2171 const int max_frame_num = 1 << h->sps.log2_max_frame_num;
2173 Picture *cur = s->current_picture_ptr;
2175 h->frame_num_offset = h->prev_frame_num_offset;
2176 if (h->frame_num < h->prev_frame_num)
2177 h->frame_num_offset += max_frame_num;
2179 if (h->sps.poc_type == 0) {
2180 const int max_poc_lsb = 1 << h->sps.log2_max_poc_lsb;
2182 if (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb / 2)
2183 h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2184 else if (h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb / 2)
2185 h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2187 h->poc_msb = h->prev_poc_msb;
2189 field_poc[1] = h->poc_msb + h->poc_lsb;
2190 if (s->picture_structure == PICT_FRAME)
2191 field_poc[1] += h->delta_poc_bottom;
2192 } else if (h->sps.poc_type == 1) {
2193 int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2196 if (h->sps.poc_cycle_length != 0)
2197 abs_frame_num = h->frame_num_offset + h->frame_num;
2201 if (h->nal_ref_idc == 0 && abs_frame_num > 0)
2204 expected_delta_per_poc_cycle = 0;
2205 for (i = 0; i < h->sps.poc_cycle_length; i++)
2206 // FIXME integrate during sps parse
2207 expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[i];
2209 if (abs_frame_num > 0) {
2210 int poc_cycle_cnt = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2211 int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2213 expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2214 for (i = 0; i <= frame_num_in_poc_cycle; i++)
2215 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[i];
2219 if (h->nal_ref_idc == 0)
2220 expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2222 field_poc[0] = expectedpoc + h->delta_poc[0];
2223 field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2225 if (s->picture_structure == PICT_FRAME)
2226 field_poc[1] += h->delta_poc[1];
2228 int poc = 2 * (h->frame_num_offset + h->frame_num);
2230 if (!h->nal_ref_idc)
2237 if (s->picture_structure != PICT_BOTTOM_FIELD)
2238 s->current_picture_ptr->field_poc[0] = field_poc[0];
2239 if (s->picture_structure != PICT_TOP_FIELD)
2240 s->current_picture_ptr->field_poc[1] = field_poc[1];
2241 cur->poc = FFMIN(cur->field_poc[0], cur->field_poc[1]);
2247 * initialize scan tables
2249 static void init_scan_tables(H264Context *h)
2252 for (i = 0; i < 16; i++) {
2253 #define T(x) (x >> 2) | ((x << 2) & 0xF)
2254 h->zigzag_scan[i] = T(zigzag_scan[i]);
2255 h->field_scan[i] = T(field_scan[i]);
2258 for (i = 0; i < 64; i++) {
2259 #define T(x) (x >> 3) | ((x & 7) << 3)
2260 h->zigzag_scan8x8[i] = T(ff_zigzag_direct[i]);
2261 h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2262 h->field_scan8x8[i] = T(field_scan8x8[i]);
2263 h->field_scan8x8_cavlc[i] = T(field_scan8x8_cavlc[i]);
2266 if (h->sps.transform_bypass) { // FIXME same ugly
2267 h->zigzag_scan_q0 = zigzag_scan;
2268 h->zigzag_scan8x8_q0 = ff_zigzag_direct;
2269 h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2270 h->field_scan_q0 = field_scan;
2271 h->field_scan8x8_q0 = field_scan8x8;
2272 h->field_scan8x8_cavlc_q0 = field_scan8x8_cavlc;
2274 h->zigzag_scan_q0 = h->zigzag_scan;
2275 h->zigzag_scan8x8_q0 = h->zigzag_scan8x8;
2276 h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2277 h->field_scan_q0 = h->field_scan;
2278 h->field_scan8x8_q0 = h->field_scan8x8;
2279 h->field_scan8x8_cavlc_q0 = h->field_scan8x8_cavlc;
2283 static int field_end(H264Context *h, int in_setup)
2285 MpegEncContext *const s = &h->s;
2286 AVCodecContext *const avctx = s->avctx;
2290 if (!in_setup && !s->droppable)
2291 ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX,
2292 s->picture_structure == PICT_BOTTOM_FIELD);
2294 if (CONFIG_H264_VDPAU_DECODER &&
2295 s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
2296 ff_vdpau_h264_set_reference_frames(s);
2298 if (in_setup || !(avctx->active_thread_type & FF_THREAD_FRAME)) {
2299 if (!s->droppable) {
2300 err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2301 h->prev_poc_msb = h->poc_msb;
2302 h->prev_poc_lsb = h->poc_lsb;
2304 h->prev_frame_num_offset = h->frame_num_offset;
2305 h->prev_frame_num = h->frame_num;
2306 h->outputed_poc = h->next_outputed_poc;
2309 if (avctx->hwaccel) {
2310 if (avctx->hwaccel->end_frame(avctx) < 0)
2311 av_log(avctx, AV_LOG_ERROR,
2312 "hardware accelerator failed to decode picture\n");
2315 if (CONFIG_H264_VDPAU_DECODER &&
2316 s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
2317 ff_vdpau_h264_picture_complete(s);
2320 * FIXME: Error handling code does not seem to support interlaced
2321 * when slices span multiple rows
2322 * The ff_er_add_slice calls don't work right for bottom
2323 * fields; they cause massive erroneous error concealing
2324 * Error marking covers both fields (top and bottom).
2325 * This causes a mismatched s->error_count
2326 * and a bad error table. Further, the error count goes to
2327 * INT_MAX when called for bottom field, because mb_y is
2328 * past end by one (callers fault) and resync_mb_y != 0
2329 * causes problems for the first MB line, too.
2334 ff_MPV_frame_end(s);
2336 h->current_slice = 0;
2342 * Replicate H264 "master" context to thread contexts.
2344 static int clone_slice(H264Context *dst, H264Context *src)
2348 memcpy(dst->block_offset, src->block_offset, sizeof(dst->block_offset));
2349 dst->s.current_picture_ptr = src->s.current_picture_ptr;
2350 dst->s.current_picture = src->s.current_picture;
2351 dst->s.linesize = src->s.linesize;
2352 dst->s.uvlinesize = src->s.uvlinesize;
2353 dst->s.first_field = src->s.first_field;
2355 if (!dst->s.edge_emu_buffer &&
2356 (ret = ff_mpv_frame_size_alloc(&dst->s, dst->s.linesize))) {
2357 av_log(dst->s.avctx, AV_LOG_ERROR,
2358 "Failed to allocate scratch buffers\n");
2362 dst->prev_poc_msb = src->prev_poc_msb;
2363 dst->prev_poc_lsb = src->prev_poc_lsb;
2364 dst->prev_frame_num_offset = src->prev_frame_num_offset;
2365 dst->prev_frame_num = src->prev_frame_num;
2366 dst->short_ref_count = src->short_ref_count;
2368 memcpy(dst->short_ref, src->short_ref, sizeof(dst->short_ref));
2369 memcpy(dst->long_ref, src->long_ref, sizeof(dst->long_ref));
2370 memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2371 memcpy(dst->ref_list, src->ref_list, sizeof(dst->ref_list));
2373 memcpy(dst->dequant4_coeff, src->dequant4_coeff, sizeof(src->dequant4_coeff));
2374 memcpy(dst->dequant8_coeff, src->dequant8_coeff, sizeof(src->dequant8_coeff));
2380 * Compute profile from profile_idc and constraint_set?_flags.
2384 * @return profile as defined by FF_PROFILE_H264_*
2386 int ff_h264_get_profile(SPS *sps)
2388 int profile = sps->profile_idc;
2390 switch (sps->profile_idc) {
2391 case FF_PROFILE_H264_BASELINE:
2392 // constraint_set1_flag set to 1
2393 profile |= (sps->constraint_set_flags & 1 << 1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2395 case FF_PROFILE_H264_HIGH_10:
2396 case FF_PROFILE_H264_HIGH_422:
2397 case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2398 // constraint_set3_flag set to 1
2399 profile |= (sps->constraint_set_flags & 1 << 3) ? FF_PROFILE_H264_INTRA : 0;
2406 static int h264_set_parameter_from_sps(H264Context *h)
2408 MpegEncContext *s = &h->s;
2410 if (s->flags & CODEC_FLAG_LOW_DELAY ||
2411 (h->sps.bitstream_restriction_flag &&
2412 !h->sps.num_reorder_frames)) {
2413 if (s->avctx->has_b_frames > 1 || h->delayed_pic[0])
2414 av_log(h->s.avctx, AV_LOG_WARNING, "Delayed frames seen. "
2415 "Reenabling low delay requires a codec flush.\n");
2420 if (s->avctx->has_b_frames < 2)
2421 s->avctx->has_b_frames = !s->low_delay;
2423 if (s->avctx->bits_per_raw_sample != h->sps.bit_depth_luma ||
2424 h->cur_chroma_format_idc != h->sps.chroma_format_idc) {
2425 if (s->avctx->codec &&
2426 s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU &&
2427 (h->sps.bit_depth_luma != 8 || h->sps.chroma_format_idc > 1)) {
2428 av_log(s->avctx, AV_LOG_ERROR,
2429 "VDPAU decoding does not support video colorspace.\n");
2430 return AVERROR_INVALIDDATA;
2432 if (h->sps.bit_depth_luma >= 8 && h->sps.bit_depth_luma <= 10) {
2433 s->avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
2434 h->cur_chroma_format_idc = h->sps.chroma_format_idc;
2435 h->pixel_shift = h->sps.bit_depth_luma > 8;
2437 ff_h264dsp_init(&h->h264dsp, h->sps.bit_depth_luma,
2438 h->sps.chroma_format_idc);
2439 ff_h264_pred_init(&h->hpc, s->codec_id, h->sps.bit_depth_luma,
2440 h->sps.chroma_format_idc);
2441 s->dsp.dct_bits = h->sps.bit_depth_luma > 8 ? 32 : 16;
2442 ff_dsputil_init(&s->dsp, s->avctx);
2443 ff_videodsp_init(&s->vdsp, h->sps.bit_depth_luma);
2445 av_log(s->avctx, AV_LOG_ERROR, "Unsupported bit depth: %d\n",
2446 h->sps.bit_depth_luma);
2447 return AVERROR_INVALIDDATA;
2453 static enum PixelFormat get_pixel_format(H264Context *h)
2455 MpegEncContext *const s = &h->s;
2456 switch (h->sps.bit_depth_luma) {
2459 if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2460 return AV_PIX_FMT_GBRP9;
2462 return AV_PIX_FMT_YUV444P9;
2463 } else if (CHROMA422)
2464 return AV_PIX_FMT_YUV422P9;
2466 return AV_PIX_FMT_YUV420P9;
2470 if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2471 return AV_PIX_FMT_GBRP10;
2473 return AV_PIX_FMT_YUV444P10;
2474 } else if (CHROMA422)
2475 return AV_PIX_FMT_YUV422P10;
2477 return AV_PIX_FMT_YUV420P10;
2481 if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2482 return AV_PIX_FMT_GBRP;
2484 return s->avctx->color_range == AVCOL_RANGE_JPEG ? AV_PIX_FMT_YUVJ444P
2485 : AV_PIX_FMT_YUV444P;
2486 } else if (CHROMA422) {
2487 return s->avctx->color_range == AVCOL_RANGE_JPEG ? AV_PIX_FMT_YUVJ422P
2488 : AV_PIX_FMT_YUV422P;
2490 return s->avctx->get_format(s->avctx, s->avctx->codec->pix_fmts ?
2491 s->avctx->codec->pix_fmts :
2492 s->avctx->color_range == AVCOL_RANGE_JPEG ?
2493 hwaccel_pixfmt_list_h264_jpeg_420 :
2494 ff_hwaccel_pixfmt_list_420);
2498 av_log(s->avctx, AV_LOG_ERROR,
2499 "Unsupported bit depth: %d\n", h->sps.bit_depth_luma);
2500 return AVERROR_INVALIDDATA;
2504 static int h264_slice_header_init(H264Context *h, int reinit)
2506 MpegEncContext *const s = &h->s;
2509 avcodec_set_dimensions(s->avctx, s->width, s->height);
2510 s->avctx->sample_aspect_ratio = h->sps.sar;
2511 av_assert0(s->avctx->sample_aspect_ratio.den);
2513 if (h->sps.timing_info_present_flag) {
2514 int64_t den = h->sps.time_scale;
2515 if (h->x264_build < 44U)
2517 av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
2518 h->sps.num_units_in_tick, den, 1 << 30);
2521 s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id, s->avctx->pix_fmt);
2525 if ((ret = ff_MPV_common_frame_size_change(s)) < 0) {
2526 av_log(h->s.avctx, AV_LOG_ERROR, "ff_MPV_common_frame_size_change() failed.\n");
2530 if ((ret = ff_MPV_common_init(s)) < 0) {
2531 av_log(h->s.avctx, AV_LOG_ERROR, "ff_MPV_common_init() failed.\n");
2536 h->prev_interlaced_frame = 1;
2538 init_scan_tables(h);
2539 if (ff_h264_alloc_tables(h) < 0) {
2540 av_log(h->s.avctx, AV_LOG_ERROR,
2541 "Could not allocate memory for h264\n");
2542 return AVERROR(ENOMEM);
2545 if (!HAVE_THREADS || !(s->avctx->active_thread_type & FF_THREAD_SLICE)) {
2546 if (context_init(h) < 0) {
2547 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2551 for (i = 1; i < s->slice_context_count; i++) {
2553 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
2554 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
2555 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
2556 c->h264dsp = h->h264dsp;
2559 c->pixel_shift = h->pixel_shift;
2560 init_scan_tables(c);
2561 clone_tables(c, h, i);
2564 for (i = 0; i < s->slice_context_count; i++)
2565 if (context_init(h->thread_context[i]) < 0) {
2566 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2575 * Decode a slice header.
2576 * This will also call ff_MPV_common_init() and frame_start() as needed.
2578 * @param h h264context
2579 * @param h0 h264 master context (differs from 'h' when doing sliced based
2580 * parallel decoding)
2582 * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2584 static int decode_slice_header(H264Context *h, H264Context *h0)
2586 MpegEncContext *const s = &h->s;
2587 MpegEncContext *const s0 = &h0->s;
2588 unsigned int first_mb_in_slice;
2589 unsigned int pps_id;
2590 int num_ref_idx_active_override_flag, max_refs, ret;
2591 unsigned int slice_type, tmp, i, j;
2592 int default_ref_list_done = 0;
2593 int last_pic_structure, last_pic_droppable;
2594 int needs_reinit = 0;
2596 /* FIXME: 2tap qpel isn't implemented for high bit depth. */
2597 if ((s->avctx->flags2 & CODEC_FLAG2_FAST) &&
2598 !h->nal_ref_idc && !h->pixel_shift) {
2599 s->me.qpel_put = s->dsp.put_2tap_qpel_pixels_tab;
2600 s->me.qpel_avg = s->dsp.avg_2tap_qpel_pixels_tab;
2602 s->me.qpel_put = s->dsp.put_h264_qpel_pixels_tab;
2603 s->me.qpel_avg = s->dsp.avg_h264_qpel_pixels_tab;
2606 first_mb_in_slice = get_ue_golomb(&s->gb);
2608 if (first_mb_in_slice == 0) { // FIXME better field boundary detection
2609 if (h0->current_slice && FIELD_PICTURE) {
2613 h0->current_slice = 0;
2614 if (!s0->first_field) {
2615 if (s->current_picture_ptr && !s->droppable &&
2616 s->current_picture_ptr->owner2 == s) {
2617 ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX,
2618 s->picture_structure == PICT_BOTTOM_FIELD);
2620 s->current_picture_ptr = NULL;
2624 slice_type = get_ue_golomb_31(&s->gb);
2625 if (slice_type > 9) {
2626 av_log(h->s.avctx, AV_LOG_ERROR,
2627 "slice type too large (%d) at %d %d\n",
2628 h->slice_type, s->mb_x, s->mb_y);
2631 if (slice_type > 4) {
2633 h->slice_type_fixed = 1;
2635 h->slice_type_fixed = 0;
2637 slice_type = golomb_to_pict_type[slice_type];
2638 if (slice_type == AV_PICTURE_TYPE_I ||
2639 (h0->current_slice != 0 && slice_type == h0->last_slice_type)) {
2640 default_ref_list_done = 1;
2642 h->slice_type = slice_type;
2643 h->slice_type_nos = slice_type & 3;
2645 // to make a few old functions happy, it's wrong though
2646 s->pict_type = h->slice_type;
2648 pps_id = get_ue_golomb(&s->gb);
2649 if (pps_id >= MAX_PPS_COUNT) {
2650 av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
2653 if (!h0->pps_buffers[pps_id]) {
2654 av_log(h->s.avctx, AV_LOG_ERROR,
2655 "non-existing PPS %u referenced\n",
2659 h->pps = *h0->pps_buffers[pps_id];
2661 if (!h0->sps_buffers[h->pps.sps_id]) {
2662 av_log(h->s.avctx, AV_LOG_ERROR,
2663 "non-existing SPS %u referenced\n",
2668 if (h->pps.sps_id != h->current_sps_id ||
2669 h->context_reinitialized ||
2670 h0->sps_buffers[h->pps.sps_id]->new) {
2671 SPS *new_sps = h0->sps_buffers[h->pps.sps_id];
2673 h0->sps_buffers[h->pps.sps_id]->new = 0;
2675 if (h->sps.chroma_format_idc != new_sps->chroma_format_idc ||
2676 h->sps.bit_depth_luma != new_sps->bit_depth_luma)
2679 h->current_sps_id = h->pps.sps_id;
2680 h->sps = *h0->sps_buffers[h->pps.sps_id];
2682 if ((ret = h264_set_parameter_from_sps(h)) < 0)
2686 s->avctx->profile = ff_h264_get_profile(&h->sps);
2687 s->avctx->level = h->sps.level_idc;
2688 s->avctx->refs = h->sps.ref_frame_count;
2690 if (s->mb_width != h->sps.mb_width ||
2691 s->mb_height != h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag))
2694 s->mb_width = h->sps.mb_width;
2695 s->mb_height = h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2697 h->b_stride = s->mb_width * 4;
2699 s->chroma_y_shift = h->sps.chroma_format_idc <= 1; // 400 uses yuv420p
2701 s->width = 16 * s->mb_width - (2 >> CHROMA444) * FFMIN(h->sps.crop_right, (8 << CHROMA444) - 1);
2702 if (h->sps.frame_mbs_only_flag)
2703 s->height = 16 * s->mb_height - (1 << s->chroma_y_shift) * FFMIN(h->sps.crop_bottom, (16 >> s->chroma_y_shift) - 1);
2705 s->height = 16 * s->mb_height - (2 << s->chroma_y_shift) * FFMIN(h->sps.crop_bottom, (16 >> s->chroma_y_shift) - 1);
2707 if (FFALIGN(s->avctx->width, 16) == s->width &&
2708 FFALIGN(s->avctx->height, 16) == s->height) {
2709 s->width = s->avctx->width;
2710 s->height = s->avctx->height;
2713 if (h->sps.video_signal_type_present_flag) {
2714 s->avctx->color_range = h->sps.full_range ? AVCOL_RANGE_JPEG
2716 if (h->sps.colour_description_present_flag) {
2717 if (s->avctx->colorspace != h->sps.colorspace)
2719 s->avctx->color_primaries = h->sps.color_primaries;
2720 s->avctx->color_trc = h->sps.color_trc;
2721 s->avctx->colorspace = h->sps.colorspace;
2725 if (s->context_initialized &&
2726 (s->width != s->avctx->width ||
2727 s->height != s->avctx->height ||
2729 av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio))) {
2732 av_log(s->avctx, AV_LOG_ERROR, "changing width/height on "
2733 "slice %d\n", h0->current_slice + 1);
2734 return AVERROR_INVALIDDATA;
2739 if ((ret = get_pixel_format(h)) < 0)
2741 s->avctx->pix_fmt = ret;
2743 av_log(h->s.avctx, AV_LOG_INFO, "Reinit context to %dx%d, "
2744 "pix_fmt: %d\n", s->width, s->height, s->avctx->pix_fmt);
2746 if ((ret = h264_slice_header_init(h, 1)) < 0) {
2747 av_log(h->s.avctx, AV_LOG_ERROR,
2748 "h264_slice_header_init() failed\n");
2751 h->context_reinitialized = 1;
2753 if (!s->context_initialized) {
2755 av_log(h->s.avctx, AV_LOG_ERROR,
2756 "Cannot (re-)initialize context during parallel decoding.\n");
2760 if ((ret = get_pixel_format(h)) < 0)
2762 s->avctx->pix_fmt = ret;
2764 if ((ret = h264_slice_header_init(h, 0)) < 0) {
2765 av_log(h->s.avctx, AV_LOG_ERROR,
2766 "h264_slice_header_init() failed\n");
2771 if (h == h0 && h->dequant_coeff_pps != pps_id) {
2772 h->dequant_coeff_pps = pps_id;
2773 init_dequant_tables(h);
2776 h->frame_num = get_bits(&s->gb, h->sps.log2_max_frame_num);
2779 h->mb_aff_frame = 0;
2780 last_pic_structure = s0->picture_structure;
2781 last_pic_droppable = s0->droppable;
2782 s->droppable = h->nal_ref_idc == 0;
2783 if (h->sps.frame_mbs_only_flag) {
2784 s->picture_structure = PICT_FRAME;
2786 if (get_bits1(&s->gb)) { // field_pic_flag
2787 s->picture_structure = PICT_TOP_FIELD + get_bits1(&s->gb); // bottom_field_flag
2789 s->picture_structure = PICT_FRAME;
2790 h->mb_aff_frame = h->sps.mb_aff;
2793 h->mb_field_decoding_flag = s->picture_structure != PICT_FRAME;
2795 if (h0->current_slice != 0) {
2796 if (last_pic_structure != s->picture_structure ||
2797 last_pic_droppable != s->droppable) {
2798 av_log(h->s.avctx, AV_LOG_ERROR,
2799 "Changing field mode (%d -> %d) between slices is not allowed\n",
2800 last_pic_structure, s->picture_structure);
2801 s->picture_structure = last_pic_structure;
2802 s->droppable = last_pic_droppable;
2803 return AVERROR_INVALIDDATA;
2804 } else if (!s0->current_picture_ptr) {
2805 av_log(s->avctx, AV_LOG_ERROR,
2806 "unset current_picture_ptr on %d. slice\n",
2807 h0->current_slice + 1);
2808 return AVERROR_INVALIDDATA;
2811 /* Shorten frame num gaps so we don't have to allocate reference
2812 * frames just to throw them away */
2813 if (h->frame_num != h->prev_frame_num) {
2814 int unwrap_prev_frame_num = h->prev_frame_num;
2815 int max_frame_num = 1 << h->sps.log2_max_frame_num;
2817 if (unwrap_prev_frame_num > h->frame_num)
2818 unwrap_prev_frame_num -= max_frame_num;
2820 if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
2821 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
2822 if (unwrap_prev_frame_num < 0)
2823 unwrap_prev_frame_num += max_frame_num;
2825 h->prev_frame_num = unwrap_prev_frame_num;
2829 /* See if we have a decoded first field looking for a pair...
2830 * Here, we're using that to see if we should mark previously
2831 * decode frames as "finished".
2832 * We have to do that before the "dummy" in-between frame allocation,
2833 * since that can modify s->current_picture_ptr. */
2834 if (s0->first_field) {
2835 assert(s0->current_picture_ptr);
2836 assert(s0->current_picture_ptr->f.data[0]);
2837 assert(s0->current_picture_ptr->f.reference != DELAYED_PIC_REF);
2839 /* Mark old field/frame as completed */
2840 if (!last_pic_droppable && s0->current_picture_ptr->owner2 == s0) {
2841 ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
2842 last_pic_structure == PICT_BOTTOM_FIELD);
2845 /* figure out if we have a complementary field pair */
2846 if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2847 /* Previous field is unmatched. Don't display it, but let it
2848 * remain for reference if marked as such. */
2849 if (!last_pic_droppable && last_pic_structure != PICT_FRAME) {
2850 ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
2851 last_pic_structure == PICT_TOP_FIELD);
2854 if (s0->current_picture_ptr->frame_num != h->frame_num) {
2855 /* This and previous field were reference, but had
2856 * different frame_nums. Consider this field first in
2857 * pair. Throw away previous field except for reference
2859 if (!last_pic_droppable && last_pic_structure != PICT_FRAME) {
2860 ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
2861 last_pic_structure == PICT_TOP_FIELD);
2864 /* Second field in complementary pair */
2865 if (!((last_pic_structure == PICT_TOP_FIELD &&
2866 s->picture_structure == PICT_BOTTOM_FIELD) ||
2867 (last_pic_structure == PICT_BOTTOM_FIELD &&
2868 s->picture_structure == PICT_TOP_FIELD))) {
2869 av_log(s->avctx, AV_LOG_ERROR,
2870 "Invalid field mode combination %d/%d\n",
2871 last_pic_structure, s->picture_structure);
2872 s->picture_structure = last_pic_structure;
2873 s->droppable = last_pic_droppable;
2874 return AVERROR_INVALIDDATA;
2875 } else if (last_pic_droppable != s->droppable) {
2876 av_log(s->avctx, AV_LOG_ERROR,
2877 "Cannot combine reference and non-reference fields in the same frame\n");
2878 av_log_ask_for_sample(s->avctx, NULL);
2879 s->picture_structure = last_pic_structure;
2880 s->droppable = last_pic_droppable;
2881 return AVERROR_PATCHWELCOME;
2884 /* Take ownership of this buffer. Note that if another thread owned
2885 * the first field of this buffer, we're not operating on that pointer,
2886 * so the original thread is still responsible for reporting progress
2887 * on that first field (or if that was us, we just did that above).
2888 * By taking ownership, we assign responsibility to ourselves to
2889 * report progress on the second field. */
2890 s0->current_picture_ptr->owner2 = s0;
2895 while (h->frame_num != h->prev_frame_num &&
2896 h->frame_num != (h->prev_frame_num + 1) % (1 << h->sps.log2_max_frame_num)) {
2897 Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
2898 av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n",
2899 h->frame_num, h->prev_frame_num);
2900 if (ff_h264_frame_start(h) < 0)
2902 h->prev_frame_num++;
2903 h->prev_frame_num %= 1 << h->sps.log2_max_frame_num;
2904 s->current_picture_ptr->frame_num = h->prev_frame_num;
2905 ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX, 0);
2906 ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX, 1);
2907 ff_generate_sliding_window_mmcos(h);
2908 if (ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index) < 0 &&
2909 (s->avctx->err_recognition & AV_EF_EXPLODE))
2910 return AVERROR_INVALIDDATA;
2911 /* Error concealment: if a ref is missing, copy the previous ref in its place.
2912 * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
2913 * about there being no actual duplicates.
2914 * FIXME: this doesn't copy padding for out-of-frame motion vectors. Given we're
2915 * concealing a lost frame, this probably isn't noticeable by comparison, but it should
2917 if (h->short_ref_count) {
2919 av_image_copy(h->short_ref[0]->f.data, h->short_ref[0]->f.linesize,
2920 (const uint8_t **)prev->f.data, prev->f.linesize,
2921 s->avctx->pix_fmt, s->mb_width * 16, s->mb_height * 16);
2922 h->short_ref[0]->poc = prev->poc + 2;
2924 h->short_ref[0]->frame_num = h->prev_frame_num;
2928 /* See if we have a decoded first field looking for a pair...
2929 * We're using that to see whether to continue decoding in that
2930 * frame, or to allocate a new one. */
2931 if (s0->first_field) {
2932 assert(s0->current_picture_ptr);
2933 assert(s0->current_picture_ptr->f.data[0]);
2934 assert(s0->current_picture_ptr->f.reference != DELAYED_PIC_REF);
2936 /* figure out if we have a complementary field pair */
2937 if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2938 /* Previous field is unmatched. Don't display it, but let it
2939 * remain for reference if marked as such. */
2940 s0->current_picture_ptr = NULL;
2941 s0->first_field = FIELD_PICTURE;
2943 if (s0->current_picture_ptr->frame_num != h->frame_num) {
2944 /* This and the previous field had different frame_nums.
2945 * Consider this field first in pair. Throw away previous
2946 * one except for reference purposes. */
2947 s0->first_field = 1;
2948 s0->current_picture_ptr = NULL;
2950 /* Second field in complementary pair */
2951 s0->first_field = 0;
2955 /* Frame or first field in a potentially complementary pair */
2956 s0->first_field = FIELD_PICTURE;
2959 if (!FIELD_PICTURE || s0->first_field) {
2960 if (ff_h264_frame_start(h) < 0) {
2961 s0->first_field = 0;
2965 ff_release_unused_pictures(s, 0);
2968 if (h != h0 && (ret = clone_slice(h, h0)) < 0)
2971 s->current_picture_ptr->frame_num = h->frame_num; // FIXME frame_num cleanup
2973 assert(s->mb_num == s->mb_width * s->mb_height);
2974 if (first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
2975 first_mb_in_slice >= s->mb_num) {
2976 av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
2979 s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
2980 s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
2981 if (s->picture_structure == PICT_BOTTOM_FIELD)
2982 s->resync_mb_y = s->mb_y = s->mb_y + 1;
2983 assert(s->mb_y < s->mb_height);
2985 if (s->picture_structure == PICT_FRAME) {
2986 h->curr_pic_num = h->frame_num;
2987 h->max_pic_num = 1 << h->sps.log2_max_frame_num;
2989 h->curr_pic_num = 2 * h->frame_num + 1;
2990 h->max_pic_num = 1 << (h->sps.log2_max_frame_num + 1);
2993 if (h->nal_unit_type == NAL_IDR_SLICE)
2994 get_ue_golomb(&s->gb); /* idr_pic_id */
2996 if (h->sps.poc_type == 0) {
2997 h->poc_lsb = get_bits(&s->gb, h->sps.log2_max_poc_lsb);
2999 if (h->pps.pic_order_present == 1 && s->picture_structure == PICT_FRAME)
3000 h->delta_poc_bottom = get_se_golomb(&s->gb);
3003 if (h->sps.poc_type == 1 && !h->sps.delta_pic_order_always_zero_flag) {
3004 h->delta_poc[0] = get_se_golomb(&s->gb);
3006 if (h->pps.pic_order_present == 1 && s->picture_structure == PICT_FRAME)
3007 h->delta_poc[1] = get_se_golomb(&s->gb);
3012 if (h->pps.redundant_pic_cnt_present)
3013 h->redundant_pic_count = get_ue_golomb(&s->gb);
3015 // set defaults, might be overridden a few lines later
3016 h->ref_count[0] = h->pps.ref_count[0];
3017 h->ref_count[1] = h->pps.ref_count[1];
3019 if (h->slice_type_nos != AV_PICTURE_TYPE_I) {
3020 if (h->slice_type_nos == AV_PICTURE_TYPE_B)
3021 h->direct_spatial_mv_pred = get_bits1(&s->gb);
3022 num_ref_idx_active_override_flag = get_bits1(&s->gb);
3024 if (num_ref_idx_active_override_flag) {
3025 h->ref_count[0] = get_ue_golomb(&s->gb) + 1;
3026 if (h->ref_count[0] < 1)
3027 return AVERROR_INVALIDDATA;
3028 if (h->slice_type_nos == AV_PICTURE_TYPE_B) {
3029 h->ref_count[1] = get_ue_golomb(&s->gb) + 1;
3030 if (h->ref_count[1] < 1)
3031 return AVERROR_INVALIDDATA;
3035 if (h->slice_type_nos == AV_PICTURE_TYPE_B)
3042 max_refs = s->picture_structure == PICT_FRAME ? 16 : 32;
3044 if (h->ref_count[0] > max_refs || h->ref_count[1] > max_refs) {
3045 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
3046 h->ref_count[0] = h->ref_count[1] = 1;
3047 return AVERROR_INVALIDDATA;
3050 if (!default_ref_list_done)
3051 ff_h264_fill_default_ref_list(h);
3053 if (h->slice_type_nos != AV_PICTURE_TYPE_I &&
3054 ff_h264_decode_ref_pic_list_reordering(h) < 0) {
3055 h->ref_count[1] = h->ref_count[0] = 0;
3059 if (h->slice_type_nos != AV_PICTURE_TYPE_I) {
3060 s->last_picture_ptr = &h->ref_list[0][0];
3061 s->last_picture_ptr->owner2 = s;
3062 ff_copy_picture(&s->last_picture, s->last_picture_ptr);
3064 if (h->slice_type_nos == AV_PICTURE_TYPE_B) {
3065 s->next_picture_ptr = &h->ref_list[1][0];
3066 s->next_picture_ptr->owner2 = s;
3067 ff_copy_picture(&s->next_picture, s->next_picture_ptr);
3070 if ((h->pps.weighted_pred && h->slice_type_nos == AV_PICTURE_TYPE_P) ||
3071 (h->pps.weighted_bipred_idc == 1 &&
3072 h->slice_type_nos == AV_PICTURE_TYPE_B))
3073 pred_weight_table(h);
3074 else if (h->pps.weighted_bipred_idc == 2 &&
3075 h->slice_type_nos == AV_PICTURE_TYPE_B) {
3076 implicit_weight_table(h, -1);
3079 for (i = 0; i < 2; i++) {
3080 h->luma_weight_flag[i] = 0;
3081 h->chroma_weight_flag[i] = 0;
3085 if (h->nal_ref_idc && ff_h264_decode_ref_pic_marking(h0, &s->gb) < 0 &&
3086 (s->avctx->err_recognition & AV_EF_EXPLODE))
3087 return AVERROR_INVALIDDATA;
3090 ff_h264_fill_mbaff_ref_list(h);
3092 if (h->pps.weighted_bipred_idc == 2 && h->slice_type_nos == AV_PICTURE_TYPE_B) {
3093 implicit_weight_table(h, 0);
3094 implicit_weight_table(h, 1);
3098 if (h->slice_type_nos == AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
3099 ff_h264_direct_dist_scale_factor(h);
3100 ff_h264_direct_ref_list_init(h);
3102 if (h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac) {
3103 tmp = get_ue_golomb_31(&s->gb);
3105 av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
3108 h->cabac_init_idc = tmp;
3111 h->last_qscale_diff = 0;
3112 tmp = h->pps.init_qp + get_se_golomb(&s->gb);
3113 if (tmp > 51 + 6 * (h->sps.bit_depth_luma - 8)) {
3114 av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
3118 h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3119 h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3120 // FIXME qscale / qp ... stuff
3121 if (h->slice_type == AV_PICTURE_TYPE_SP)
3122 get_bits1(&s->gb); /* sp_for_switch_flag */
3123 if (h->slice_type == AV_PICTURE_TYPE_SP ||
3124 h->slice_type == AV_PICTURE_TYPE_SI)
3125 get_se_golomb(&s->gb); /* slice_qs_delta */
3127 h->deblocking_filter = 1;
3128 h->slice_alpha_c0_offset = 52;
3129 h->slice_beta_offset = 52;
3130 if (h->pps.deblocking_filter_parameters_present) {
3131 tmp = get_ue_golomb_31(&s->gb);
3133 av_log(s->avctx, AV_LOG_ERROR,
3134 "deblocking_filter_idc %u out of range\n", tmp);
3137 h->deblocking_filter = tmp;
3138 if (h->deblocking_filter < 2)
3139 h->deblocking_filter ^= 1; // 1<->0
3141 if (h->deblocking_filter) {
3142 h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
3143 h->slice_beta_offset += get_se_golomb(&s->gb) << 1;
3144 if (h->slice_alpha_c0_offset > 104U ||
3145 h->slice_beta_offset > 104U) {
3146 av_log(s->avctx, AV_LOG_ERROR,
3147 "deblocking filter parameters %d %d out of range\n",
3148 h->slice_alpha_c0_offset, h->slice_beta_offset);
3154 if (s->avctx->skip_loop_filter >= AVDISCARD_ALL ||
3155 (s->avctx->skip_loop_filter >= AVDISCARD_NONKEY &&
3156 h->slice_type_nos != AV_PICTURE_TYPE_I) ||
3157 (s->avctx->skip_loop_filter >= AVDISCARD_BIDIR &&
3158 h->slice_type_nos == AV_PICTURE_TYPE_B) ||
3159 (s->avctx->skip_loop_filter >= AVDISCARD_NONREF &&
3160 h->nal_ref_idc == 0))
3161 h->deblocking_filter = 0;
3163 if (h->deblocking_filter == 1 && h0->max_contexts > 1) {
3164 if (s->avctx->flags2 & CODEC_FLAG2_FAST) {
3165 /* Cheat slightly for speed:
3166 * Do not bother to deblock across slices. */
3167 h->deblocking_filter = 2;
3169 h0->max_contexts = 1;
3170 if (!h0->single_decode_warning) {
3171 av_log(s->avctx, AV_LOG_INFO,
3172 "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
3173 h0->single_decode_warning = 1;
3176 av_log(h->s.avctx, AV_LOG_ERROR,
3177 "Deblocking switched inside frame.\n");
3182 h->qp_thresh = 15 + 52 -
3183 FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset) -
3185 h->pps.chroma_qp_index_offset[0],
3186 h->pps.chroma_qp_index_offset[1]) +
3187 6 * (h->sps.bit_depth_luma - 8);
3189 h0->last_slice_type = slice_type;
3190 h->slice_num = ++h0->current_slice;
3191 if (h->slice_num >= MAX_SLICES) {
3192 av_log(s->avctx, AV_LOG_ERROR,
3193 "Too many slices, increase MAX_SLICES and recompile\n");
3196 for (j = 0; j < 2; j++) {
3198 int *ref2frm = h->ref2frm[h->slice_num & (MAX_SLICES - 1)][j];
3199 for (i = 0; i < 16; i++) {
3201 if (h->ref_list[j][i].f.data[0]) {
3203 uint8_t *base = h->ref_list[j][i].f.base[0];
3204 for (k = 0; k < h->short_ref_count; k++)
3205 if (h->short_ref[k]->f.base[0] == base) {
3209 for (k = 0; k < h->long_ref_count; k++)
3210 if (h->long_ref[k] && h->long_ref[k]->f.base[0] == base) {
3211 id_list[i] = h->short_ref_count + k;
3219 for (i = 0; i < 16; i++)
3220 ref2frm[i + 2] = 4 * id_list[i] +
3221 (h->ref_list[j][i].f.reference & 3);
3223 ref2frm[18 + 1] = -1;
3224 for (i = 16; i < 48; i++)
3225 ref2frm[i + 4] = 4 * id_list[(i - 16) >> 1] +
3226 (h->ref_list[j][i].f.reference & 3);
3229 // FIXME: fix draw_edges + PAFF + frame threads
3230 h->emu_edge_width = (s->flags & CODEC_FLAG_EMU_EDGE ||
3231 (!h->sps.frame_mbs_only_flag &&
3232 s->avctx->active_thread_type))
3234 h->emu_edge_height = (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
3236 if (s->avctx->debug & FF_DEBUG_PICT_INFO) {
3237 av_log(h->s.avctx, AV_LOG_DEBUG,
3238 "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",
3240 (s->picture_structure == PICT_FRAME ? "F" : s->picture_structure == PICT_TOP_FIELD ? "T" : "B"),
3242 av_get_picture_type_char(h->slice_type),
3243 h->slice_type_fixed ? " fix" : "",
3244 h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
3245 pps_id, h->frame_num,
3246 s->current_picture_ptr->field_poc[0],
3247 s->current_picture_ptr->field_poc[1],
3248 h->ref_count[0], h->ref_count[1],
3250 h->deblocking_filter,
3251 h->slice_alpha_c0_offset / 2 - 26, h->slice_beta_offset / 2 - 26,
3253 h->use_weight == 1 && h->use_weight_chroma ? "c" : "",
3254 h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : "");
3260 int ff_h264_get_slice_type(const H264Context *h)
3262 switch (h->slice_type) {
3263 case AV_PICTURE_TYPE_P:
3265 case AV_PICTURE_TYPE_B:
3267 case AV_PICTURE_TYPE_I:
3269 case AV_PICTURE_TYPE_SP:
3271 case AV_PICTURE_TYPE_SI:
3278 static av_always_inline void fill_filter_caches_inter(H264Context *h,
3279 MpegEncContext *const s,
3280 int mb_type, int top_xy,
3281 int left_xy[LEFT_MBS],
3283 int left_type[LEFT_MBS],
3284 int mb_xy, int list)
3286 int b_stride = h->b_stride;
3287 int16_t(*mv_dst)[2] = &h->mv_cache[list][scan8[0]];
3288 int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
3289 if (IS_INTER(mb_type) || IS_DIRECT(mb_type)) {
3290 if (USES_LIST(top_type, list)) {
3291 const int b_xy = h->mb2b_xy[top_xy] + 3 * b_stride;
3292 const int b8_xy = 4 * top_xy + 2;
3293 int (*ref2frm)[64] = h->ref2frm[h->slice_table[top_xy] & (MAX_SLICES - 1)][0] + (MB_MBAFF ? 20 : 2);
3294 AV_COPY128(mv_dst - 1 * 8, s->current_picture.f.motion_val[list][b_xy + 0]);
3295 ref_cache[0 - 1 * 8] =
3296 ref_cache[1 - 1 * 8] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 0]];
3297 ref_cache[2 - 1 * 8] =
3298 ref_cache[3 - 1 * 8] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 1]];
3300 AV_ZERO128(mv_dst - 1 * 8);
3301 AV_WN32A(&ref_cache[0 - 1 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3304 if (!IS_INTERLACED(mb_type ^ left_type[LTOP])) {
3305 if (USES_LIST(left_type[LTOP], list)) {
3306 const int b_xy = h->mb2b_xy[left_xy[LTOP]] + 3;
3307 const int b8_xy = 4 * left_xy[LTOP] + 1;
3308 int (*ref2frm)[64] = h->ref2frm[h->slice_table[left_xy[LTOP]] & (MAX_SLICES - 1)][0] + (MB_MBAFF ? 20 : 2);
3309 AV_COPY32(mv_dst - 1 + 0, s->current_picture.f.motion_val[list][b_xy + b_stride * 0]);
3310 AV_COPY32(mv_dst - 1 + 8, s->current_picture.f.motion_val[list][b_xy + b_stride * 1]);
3311 AV_COPY32(mv_dst - 1 + 16, s->current_picture.f.motion_val[list][b_xy + b_stride * 2]);
3312 AV_COPY32(mv_dst - 1 + 24, s->current_picture.f.motion_val[list][b_xy + b_stride * 3]);
3314 ref_cache[-1 + 8] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2 * 0]];
3315 ref_cache[-1 + 16] =
3316 ref_cache[-1 + 24] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2 * 1]];
3318 AV_ZERO32(mv_dst - 1 + 0);
3319 AV_ZERO32(mv_dst - 1 + 8);
3320 AV_ZERO32(mv_dst - 1 + 16);
3321 AV_ZERO32(mv_dst - 1 + 24);
3324 ref_cache[-1 + 16] =
3325 ref_cache[-1 + 24] = LIST_NOT_USED;
3330 if (!USES_LIST(mb_type, list)) {
3331 fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0, 0), 4);
3332 AV_WN32A(&ref_cache[0 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3333 AV_WN32A(&ref_cache[1 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3334 AV_WN32A(&ref_cache[2 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3335 AV_WN32A(&ref_cache[3 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3340 int8_t *ref = &s->current_picture.f.ref_index[list][4 * mb_xy];
3341 int (*ref2frm)[64] = h->ref2frm[h->slice_num & (MAX_SLICES - 1)][0] + (MB_MBAFF ? 20 : 2);
3342 uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]], ref2frm[list][ref[1]]) & 0x00FF00FF) * 0x0101;
3343 uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]], ref2frm[list][ref[3]]) & 0x00FF00FF) * 0x0101;
3344 AV_WN32A(&ref_cache[0 * 8], ref01);
3345 AV_WN32A(&ref_cache[1 * 8], ref01);
3346 AV_WN32A(&ref_cache[2 * 8], ref23);
3347 AV_WN32A(&ref_cache[3 * 8], ref23);
3351 int16_t(*mv_src)[2] = &s->current_picture.f.motion_val[list][4 * s->mb_x + 4 * s->mb_y * b_stride];
3352 AV_COPY128(mv_dst + 8 * 0, mv_src + 0 * b_stride);
3353 AV_COPY128(mv_dst + 8 * 1, mv_src + 1 * b_stride);
3354 AV_COPY128(mv_dst + 8 * 2, mv_src + 2 * b_stride);
3355 AV_COPY128(mv_dst + 8 * 3, mv_src + 3 * b_stride);
3361 * @return non zero if the loop filter can be skipped
3363 static int fill_filter_caches(H264Context *h, int mb_type)
3365 MpegEncContext *const s = &h->s;
3366 const int mb_xy = h->mb_xy;
3367 int top_xy, left_xy[LEFT_MBS];
3368 int top_type, left_type[LEFT_MBS];
3372 top_xy = mb_xy - (s->mb_stride << MB_FIELD);
3374 /* Wow, what a mess, why didn't they simplify the interlacing & intra
3375 * stuff, I can't imagine that these complex rules are worth it. */
3377 left_xy[LBOT] = left_xy[LTOP] = mb_xy - 1;
3379 const int left_mb_field_flag = IS_INTERLACED(s->current_picture.f.mb_type[mb_xy - 1]);
3380 const int curr_mb_field_flag = IS_INTERLACED(mb_type);
3382 if (left_mb_field_flag != curr_mb_field_flag)
3383 left_xy[LTOP] -= s->mb_stride;
3385 if (curr_mb_field_flag)
3386 top_xy += s->mb_stride &
3387 (((s->current_picture.f.mb_type[top_xy] >> 7) & 1) - 1);
3388 if (left_mb_field_flag != curr_mb_field_flag)
3389 left_xy[LBOT] += s->mb_stride;
3393 h->top_mb_xy = top_xy;
3394 h->left_mb_xy[LTOP] = left_xy[LTOP];
3395 h->left_mb_xy[LBOT] = left_xy[LBOT];
3397 /* For sufficiently low qp, filtering wouldn't do anything.
3398 * This is a conservative estimate: could also check beta_offset
3399 * and more accurate chroma_qp. */
3400 int qp_thresh = h->qp_thresh; // FIXME strictly we should store qp_thresh for each mb of a slice
3401 int qp = s->current_picture.f.qscale_table[mb_xy];
3402 if (qp <= qp_thresh &&
3403 (left_xy[LTOP] < 0 ||
3404 ((qp + s->current_picture.f.qscale_table[left_xy[LTOP]] + 1) >> 1) <= qp_thresh) &&
3406 ((qp + s->current_picture.f.qscale_table[top_xy] + 1) >> 1) <= qp_thresh)) {
3409 if ((left_xy[LTOP] < 0 ||
3410 ((qp + s->current_picture.f.qscale_table[left_xy[LBOT]] + 1) >> 1) <= qp_thresh) &&
3411 (top_xy < s->mb_stride ||
3412 ((qp + s->current_picture.f.qscale_table[top_xy - s->mb_stride] + 1) >> 1) <= qp_thresh))
3417 top_type = s->current_picture.f.mb_type[top_xy];
3418 left_type[LTOP] = s->current_picture.f.mb_type[left_xy[LTOP]];
3419 left_type[LBOT] = s->current_picture.f.mb_type[left_xy[LBOT]];
3420 if (h->deblocking_filter == 2) {
3421 if (h->slice_table[top_xy] != h->slice_num)
3423 if (h->slice_table[left_xy[LBOT]] != h->slice_num)
3424 left_type[LTOP] = left_type[LBOT] = 0;
3426 if (h->slice_table[top_xy] == 0xFFFF)
3428 if (h->slice_table[left_xy[LBOT]] == 0xFFFF)
3429 left_type[LTOP] = left_type[LBOT] = 0;
3431 h->top_type = top_type;
3432 h->left_type[LTOP] = left_type[LTOP];
3433 h->left_type[LBOT] = left_type[LBOT];
3435 if (IS_INTRA(mb_type))
3438 fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy,
3439 top_type, left_type, mb_xy, 0);
3440 if (h->list_count == 2)
3441 fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy,
3442 top_type, left_type, mb_xy, 1);
3444 nnz = h->non_zero_count[mb_xy];
3445 nnz_cache = h->non_zero_count_cache;
3446 AV_COPY32(&nnz_cache[4 + 8 * 1], &nnz[0]);
3447 AV_COPY32(&nnz_cache[4 + 8 * 2], &nnz[4]);
3448 AV_COPY32(&nnz_cache[4 + 8 * 3], &nnz[8]);
3449 AV_COPY32(&nnz_cache[4 + 8 * 4], &nnz[12]);
3450 h->cbp = h->cbp_table[mb_xy];
3453 nnz = h->non_zero_count[top_xy];
3454 AV_COPY32(&nnz_cache[4 + 8 * 0], &nnz[3 * 4]);
3457 if (left_type[LTOP]) {
3458 nnz = h->non_zero_count[left_xy[LTOP]];
3459 nnz_cache[3 + 8 * 1] = nnz[3 + 0 * 4];
3460 nnz_cache[3 + 8 * 2] = nnz[3 + 1 * 4];
3461 nnz_cache[3 + 8 * 3] = nnz[3 + 2 * 4];
3462 nnz_cache[3 + 8 * 4] = nnz[3 + 3 * 4];
3465 /* CAVLC 8x8dct requires NNZ values for residual decoding that differ
3466 * from what the loop filter needs */
3467 if (!CABAC && h->pps.transform_8x8_mode) {
3468 if (IS_8x8DCT(top_type)) {
3469 nnz_cache[4 + 8 * 0] =
3470 nnz_cache[5 + 8 * 0] = (h->cbp_table[top_xy] & 0x4000) >> 12;
3471 nnz_cache[6 + 8 * 0] =
3472 nnz_cache[7 + 8 * 0] = (h->cbp_table[top_xy] & 0x8000) >> 12;
3474 if (IS_8x8DCT(left_type[LTOP])) {
3475 nnz_cache[3 + 8 * 1] =
3476 nnz_cache[3 + 8 * 2] = (h->cbp_table[left_xy[LTOP]] & 0x2000) >> 12; // FIXME check MBAFF
3478 if (IS_8x8DCT(left_type[LBOT])) {
3479 nnz_cache[3 + 8 * 3] =
3480 nnz_cache[3 + 8 * 4] = (h->cbp_table[left_xy[LBOT]] & 0x8000) >> 12; // FIXME check MBAFF
3483 if (IS_8x8DCT(mb_type)) {
3484 nnz_cache[scan8[0]] =
3485 nnz_cache[scan8[1]] =
3486 nnz_cache[scan8[2]] =
3487 nnz_cache[scan8[3]] = (h->cbp & 0x1000) >> 12;
3489 nnz_cache[scan8[0 + 4]] =
3490 nnz_cache[scan8[1 + 4]] =
3491 nnz_cache[scan8[2 + 4]] =
3492 nnz_cache[scan8[3 + 4]] = (h->cbp & 0x2000) >> 12;
3494 nnz_cache[scan8[0 + 8]] =
3495 nnz_cache[scan8[1 + 8]] =
3496 nnz_cache[scan8[2 + 8]] =
3497 nnz_cache[scan8[3 + 8]] = (h->cbp & 0x4000) >> 12;
3499 nnz_cache[scan8[0 + 12]] =
3500 nnz_cache[scan8[1 + 12]] =
3501 nnz_cache[scan8[2 + 12]] =
3502 nnz_cache[scan8[3 + 12]] = (h->cbp & 0x8000) >> 12;
3509 static void loop_filter(H264Context *h, int start_x, int end_x)
3511 MpegEncContext *const s = &h->s;
3512 uint8_t *dest_y, *dest_cb, *dest_cr;
3513 int linesize, uvlinesize, mb_x, mb_y;
3514 const int end_mb_y = s->mb_y + FRAME_MBAFF;
3515 const int old_slice_type = h->slice_type;
3516 const int pixel_shift = h->pixel_shift;
3517 const int block_h = 16 >> s->chroma_y_shift;
3519 if (h->deblocking_filter) {
3520 for (mb_x = start_x; mb_x < end_x; mb_x++)
3521 for (mb_y = end_mb_y - FRAME_MBAFF; mb_y <= end_mb_y; mb_y++) {
3523 mb_xy = h->mb_xy = mb_x + mb_y * s->mb_stride;
3524 h->slice_num = h->slice_table[mb_xy];
3525 mb_type = s->current_picture.f.mb_type[mb_xy];
3526 h->list_count = h->list_counts[mb_xy];
3530 h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
3534 dest_y = s->current_picture.f.data[0] +
3535 ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
3536 dest_cb = s->current_picture.f.data[1] +
3537 (mb_x << pixel_shift) * (8 << CHROMA444) +
3538 mb_y * s->uvlinesize * block_h;
3539 dest_cr = s->current_picture.f.data[2] +
3540 (mb_x << pixel_shift) * (8 << CHROMA444) +
3541 mb_y * s->uvlinesize * block_h;
3542 // FIXME simplify above
3545 linesize = h->mb_linesize = s->linesize * 2;
3546 uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
3547 if (mb_y & 1) { // FIXME move out of this function?
3548 dest_y -= s->linesize * 15;
3549 dest_cb -= s->uvlinesize * (block_h - 1);
3550 dest_cr -= s->uvlinesize * (block_h - 1);
3553 linesize = h->mb_linesize = s->linesize;
3554 uvlinesize = h->mb_uvlinesize = s->uvlinesize;
3556 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize,
3558 if (fill_filter_caches(h, mb_type))
3560 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.f.qscale_table[mb_xy]);
3561 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.f.qscale_table[mb_xy]);
3564 ff_h264_filter_mb(h, mb_x, mb_y, dest_y, dest_cb, dest_cr,
3565 linesize, uvlinesize);
3567 ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb,
3568 dest_cr, linesize, uvlinesize);
3572 h->slice_type = old_slice_type;
3574 s->mb_y = end_mb_y - FRAME_MBAFF;
3575 h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3576 h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3579 static void predict_field_decoding_flag(H264Context *h)
3581 MpegEncContext *const s = &h->s;
3582 const int mb_xy = s->mb_x + s->mb_y * s->mb_stride;
3583 int mb_type = (h->slice_table[mb_xy - 1] == h->slice_num) ?
3584 s->current_picture.f.mb_type[mb_xy - 1] :
3585 (h->slice_table[mb_xy - s->mb_stride] == h->slice_num) ?
3586 s->current_picture.f.mb_type[mb_xy - s->mb_stride] : 0;
3587 h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
3591 * Draw edges and report progress for the last MB row.
3593 static void decode_finish_row(H264Context *h)
3595 MpegEncContext *const s = &h->s;
3596 int top = 16 * (s->mb_y >> FIELD_PICTURE);
3597 int pic_height = 16 * s->mb_height >> FIELD_PICTURE;
3598 int height = 16 << FRAME_MBAFF;
3599 int deblock_border = (16 + 4) << FRAME_MBAFF;
3601 if (h->deblocking_filter) {
3602 if ((top + height) >= pic_height)
3603 height += deblock_border;
3604 top -= deblock_border;
3607 if (top >= pic_height || (top + height) < h->emu_edge_height)
3610 height = FFMIN(height, pic_height - top);
3611 if (top < h->emu_edge_height) {
3612 height = top + height;
3616 ff_draw_horiz_band(s, top, height);
3621 ff_thread_report_progress(&s->current_picture_ptr->f, top + height - 1,
3622 s->picture_structure == PICT_BOTTOM_FIELD);
3625 static int decode_slice(struct AVCodecContext *avctx, void *arg)
3627 H264Context *h = *(void **)arg;
3628 MpegEncContext *const s = &h->s;
3629 const int part_mask = s->partitioned_frame ? (ER_AC_END | ER_AC_ERROR)
3631 int lf_x_start = s->mb_x;
3633 s->mb_skip_run = -1;
3635 h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME ||
3636 s->codec_id != AV_CODEC_ID_H264 ||
3637 (CONFIG_GRAY && (s->flags & CODEC_FLAG_GRAY));
3641 align_get_bits(&s->gb);
3644 ff_init_cabac_states(&h->cabac);
3645 ff_init_cabac_decoder(&h->cabac,
3646 s->gb.buffer + get_bits_count(&s->gb) / 8,
3647 (get_bits_left(&s->gb) + 7) / 8);
3649 ff_h264_init_cabac_states(h);
3653 int ret = ff_h264_decode_mb_cabac(h);
3655 // STOP_TIMER("decode_mb_cabac")
3658 ff_h264_hl_decode_mb(h);
3660 // FIXME optimal? or let mb_decode decode 16x32 ?
3661 if (ret >= 0 && FRAME_MBAFF) {
3664 ret = ff_h264_decode_mb_cabac(h);
3667 ff_h264_hl_decode_mb(h);
3670 eos = get_cabac_terminate(&h->cabac);
3672 if ((s->workaround_bugs & FF_BUG_TRUNCATED) &&
3673 h->cabac.bytestream > h->cabac.bytestream_end + 2) {
3674 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x - 1,
3675 s->mb_y, ER_MB_END & part_mask);
3676 if (s->mb_x >= lf_x_start)
3677 loop_filter(h, lf_x_start, s->mb_x + 1);
3680 if (ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 2) {
3681 av_log(h->s.avctx, AV_LOG_ERROR,
3682 "error while decoding MB %d %d, bytestream (%td)\n",
3684 h->cabac.bytestream_end - h->cabac.bytestream);
3685 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x,
3686 s->mb_y, ER_MB_ERROR & part_mask);
3690 if (++s->mb_x >= s->mb_width) {
3691 loop_filter(h, lf_x_start, s->mb_x);
3692 s->mb_x = lf_x_start = 0;
3693 decode_finish_row(h);
3695 if (FIELD_OR_MBAFF_PICTURE) {
3697 if (FRAME_MBAFF && s->mb_y < s->mb_height)
3698 predict_field_decoding_flag(h);
3702 if (eos || s->mb_y >= s->mb_height) {
3703 tprintf(s->avctx, "slice end %d %d\n",
3704 get_bits_count(&s->gb), s->gb.size_in_bits);
3705 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x - 1,
3706 s->mb_y, ER_MB_END & part_mask);
3707 if (s->mb_x > lf_x_start)
3708 loop_filter(h, lf_x_start, s->mb_x);
3714 int ret = ff_h264_decode_mb_cavlc(h);
3717 ff_h264_hl_decode_mb(h);
3719 // FIXME optimal? or let mb_decode decode 16x32 ?
3720 if (ret >= 0 && FRAME_MBAFF) {
3722 ret = ff_h264_decode_mb_cavlc(h);
3725 ff_h264_hl_decode_mb(h);
3730 av_log(h->s.avctx, AV_LOG_ERROR,
3731 "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
3732 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x,
3733 s->mb_y, ER_MB_ERROR & part_mask);
3737 if (++s->mb_x >= s->mb_width) {
3738 loop_filter(h, lf_x_start, s->mb_x);
3739 s->mb_x = lf_x_start = 0;
3740 decode_finish_row(h);
3742 if (FIELD_OR_MBAFF_PICTURE) {
3744 if (FRAME_MBAFF && s->mb_y < s->mb_height)
3745 predict_field_decoding_flag(h);
3747 if (s->mb_y >= s->mb_height) {
3748 tprintf(s->avctx, "slice end %d %d\n",
3749 get_bits_count(&s->gb), s->gb.size_in_bits);
3751 if (get_bits_left(&s->gb) == 0) {
3752 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y,
3753 s->mb_x - 1, s->mb_y,
3754 ER_MB_END & part_mask);
3758 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y,
3759 s->mb_x - 1, s->mb_y,
3760 ER_MB_END & part_mask);
3767 if (get_bits_left(&s->gb) <= 0 && s->mb_skip_run <= 0) {
3768 tprintf(s->avctx, "slice end %d %d\n",
3769 get_bits_count(&s->gb), s->gb.size_in_bits);
3770 if (get_bits_left(&s->gb) == 0) {
3771 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y,
3772 s->mb_x - 1, s->mb_y,
3773 ER_MB_END & part_mask);
3774 if (s->mb_x > lf_x_start)
3775 loop_filter(h, lf_x_start, s->mb_x);
3779 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x,
3780 s->mb_y, ER_MB_ERROR & part_mask);
3790 * Call decode_slice() for each context.
3792 * @param h h264 master context
3793 * @param context_count number of contexts to execute
3795 static int execute_decode_slices(H264Context *h, int context_count)
3797 MpegEncContext *const s = &h->s;
3798 AVCodecContext *const avctx = s->avctx;
3802 if (s->avctx->hwaccel ||
3803 s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
3805 if (context_count == 1) {
3806 return decode_slice(avctx, &h);
3808 for (i = 1; i < context_count; i++) {
3809 hx = h->thread_context[i];
3810 hx->s.err_recognition = avctx->err_recognition;
3811 hx->s.error_count = 0;
3814 avctx->execute(avctx, decode_slice, h->thread_context,
3815 NULL, context_count, sizeof(void *));
3817 /* pull back stuff from slices to master context */
3818 hx = h->thread_context[context_count - 1];
3819 s->mb_x = hx->s.mb_x;
3820 s->mb_y = hx->s.mb_y;
3821 s->droppable = hx->s.droppable;
3822 s->picture_structure = hx->s.picture_structure;
3823 for (i = 1; i < context_count; i++)
3824 h->s.error_count += h->thread_context[i]->s.error_count;
3830 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size,
3831 int parse_extradata)
3833 MpegEncContext *const s = &h->s;
3834 AVCodecContext *const avctx = s->avctx;
3835 H264Context *hx; ///< thread context
3839 int pass = !(avctx->active_thread_type & FF_THREAD_FRAME);
3840 int nals_needed = 0; ///< number of NALs that need decoding before the next frame thread starts
3843 h->max_contexts = s->slice_context_count;
3844 if (!(s->flags2 & CODEC_FLAG2_CHUNKS)) {
3845 h->current_slice = 0;
3846 if (!s->first_field)
3847 s->current_picture_ptr = NULL;
3848 ff_h264_reset_sei(h);
3851 for (; pass <= 1; pass++) {
3854 next_avc = h->is_avc ? 0 : buf_size;
3864 if (buf_index >= next_avc) {
3865 if (buf_index >= buf_size - h->nal_length_size)
3868 for (i = 0; i < h->nal_length_size; i++)
3869 nalsize = (nalsize << 8) | buf[buf_index++];
3870 if (nalsize <= 0 || nalsize > buf_size - buf_index) {
3871 av_log(h->s.avctx, AV_LOG_ERROR,
3872 "AVC: nal size %d\n", nalsize);
3875 next_avc = buf_index + nalsize;
3877 // start code prefix search
3878 for (; buf_index + 3 < next_avc; buf_index++)
3879 // This should always succeed in the first iteration.
3880 if (buf[buf_index] == 0 &&
3881 buf[buf_index + 1] == 0 &&
3882 buf[buf_index + 2] == 1)
3885 if (buf_index + 3 >= buf_size) {
3886 buf_index = buf_size;
3891 if (buf_index >= next_avc)
3895 hx = h->thread_context[context_count];
3897 ptr = ff_h264_decode_nal(hx, buf + buf_index, &dst_length,
3898 &consumed, next_avc - buf_index);
3899 if (ptr == NULL || dst_length < 0) {
3903 i = buf_index + consumed;
3904 if ((s->workaround_bugs & FF_BUG_AUTODETECT) && i + 3 < next_avc &&
3905 buf[i] == 0x00 && buf[i + 1] == 0x00 &&
3906 buf[i + 2] == 0x01 && buf[i + 3] == 0xE0)
3907 s->workaround_bugs |= FF_BUG_TRUNCATED;
3909 if (!(s->workaround_bugs & FF_BUG_TRUNCATED))
3910 while (ptr[dst_length - 1] == 0 && dst_length > 0)
3912 bit_length = !dst_length ? 0
3914 decode_rbsp_trailing(h, ptr + dst_length - 1));
3916 if (s->avctx->debug & FF_DEBUG_STARTCODE)
3917 av_log(h->s.avctx, AV_LOG_DEBUG,
3918 "NAL %d at %d/%d length %d\n",
3919 hx->nal_unit_type, buf_index, buf_size, dst_length);
3921 if (h->is_avc && (nalsize != consumed) && nalsize)
3922 av_log(h->s.avctx, AV_LOG_DEBUG,
3923 "AVC: Consumed only %d bytes instead of %d\n",
3926 buf_index += consumed;
3930 /* packets can sometimes contain multiple PPS/SPS,
3931 * e.g. two PAFF field pictures in one packet, or a demuxer
3932 * which splits NALs strangely if so, when frame threading we
3933 * can't start the next thread until we've read all of them */
3934 switch (hx->nal_unit_type) {
3937 nals_needed = nal_index;
3942 init_get_bits(&hx->s.gb, ptr, bit_length);
3943 if (!get_ue_golomb(&hx->s.gb))
3944 nals_needed = nal_index;
3949 // FIXME do not discard SEI id
3950 if (avctx->skip_frame >= AVDISCARD_NONREF && h->nal_ref_idc == 0)
3954 /* Ignore every NAL unit type except PPS and SPS during extradata
3955 * parsing. Decoding slices is not possible in codec init
3957 if (parse_extradata && HAVE_THREADS &&
3958 (s->avctx->active_thread_type & FF_THREAD_FRAME) &&
3959 (hx->nal_unit_type != NAL_PPS &&
3960 hx->nal_unit_type != NAL_SPS)) {
3961 av_log(avctx, AV_LOG_INFO, "Ignoring NAL unit %d during "
3962 "extradata parsing\n", hx->nal_unit_type);
3963 hx->nal_unit_type = NAL_FF_IGNORE;
3966 switch (hx->nal_unit_type) {
3968 if (h->nal_unit_type != NAL_IDR_SLICE) {
3969 av_log(h->s.avctx, AV_LOG_ERROR,
3970 "Invalid mix of idr and non-idr slices\n");
3974 idr(h); // FIXME ensure we don't lose some frames if there is reordering
3976 init_get_bits(&hx->s.gb, ptr, bit_length);
3978 hx->inter_gb_ptr = &hx->s.gb;
3979 hx->s.data_partitioning = 0;
3981 if ((err = decode_slice_header(hx, h)))
3984 s->current_picture_ptr->f.key_frame |=
3985 (hx->nal_unit_type == NAL_IDR_SLICE) ||
3986 (h->sei_recovery_frame_cnt >= 0);
3988 if (h->current_slice == 1) {
3989 if (!(s->flags2 & CODEC_FLAG2_CHUNKS))
3990 decode_postinit(h, nal_index >= nals_needed);
3992 if (s->avctx->hwaccel &&
3993 s->avctx->hwaccel->start_frame(s->avctx, NULL, 0) < 0)
3995 if (CONFIG_H264_VDPAU_DECODER &&
3996 s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
3997 ff_vdpau_h264_picture_start(s);
4000 if (hx->redundant_pic_count == 0 &&
4001 (avctx->skip_frame < AVDISCARD_NONREF ||
4003 (avctx->skip_frame < AVDISCARD_BIDIR ||
4004 hx->slice_type_nos != AV_PICTURE_TYPE_B) &&
4005 (avctx->skip_frame < AVDISCARD_NONKEY ||
4006 hx->slice_type_nos == AV_PICTURE_TYPE_I) &&
4007 avctx->skip_frame < AVDISCARD_ALL) {
4008 if (avctx->hwaccel) {
4009 if (avctx->hwaccel->decode_slice(avctx,
4010 &buf[buf_index - consumed],
4013 } else if (CONFIG_H264_VDPAU_DECODER &&
4014 s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU) {
4015 static const uint8_t start_code[] = {
4017 ff_vdpau_add_data_chunk(s, start_code,
4018 sizeof(start_code));
4019 ff_vdpau_add_data_chunk(s, &buf[buf_index - consumed],
4026 init_get_bits(&hx->s.gb, ptr, bit_length);
4028 hx->inter_gb_ptr = NULL;
4030 if ((err = decode_slice_header(hx, h)) < 0)
4033 hx->s.data_partitioning = 1;
4036 init_get_bits(&hx->intra_gb, ptr, bit_length);
4037 hx->intra_gb_ptr = &hx->intra_gb;
4040 init_get_bits(&hx->inter_gb, ptr, bit_length);
4041 hx->inter_gb_ptr = &hx->inter_gb;
4043 if (hx->redundant_pic_count == 0 &&
4045 hx->s.data_partitioning &&
4046 s->current_picture_ptr &&
4047 s->context_initialized &&
4048 (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc) &&
4049 (avctx->skip_frame < AVDISCARD_BIDIR ||
4050 hx->slice_type_nos != AV_PICTURE_TYPE_B) &&
4051 (avctx->skip_frame < AVDISCARD_NONKEY ||
4052 hx->slice_type_nos == AV_PICTURE_TYPE_I) &&
4053 avctx->skip_frame < AVDISCARD_ALL)
4057 init_get_bits(&s->gb, ptr, bit_length);
4058 ff_h264_decode_sei(h);
4061 init_get_bits(&s->gb, ptr, bit_length);
4062 if (ff_h264_decode_seq_parameter_set(h) < 0 &&
4063 h->is_avc && (nalsize != consumed) && nalsize) {
4064 av_log(h->s.avctx, AV_LOG_DEBUG,
4065 "SPS decoding failure, trying again with the complete NAL\n");
4066 init_get_bits(&s->gb, buf + buf_index + 1 - consumed,
4068 ff_h264_decode_seq_parameter_set(h);
4071 if (h264_set_parameter_from_sps(h) < 0) {
4077 init_get_bits(&s->gb, ptr, bit_length);
4078 ff_h264_decode_picture_parameter_set(h, bit_length);
4081 case NAL_END_SEQUENCE:
4082 case NAL_END_STREAM:
4083 case NAL_FILLER_DATA:
4085 case NAL_AUXILIARY_SLICE:
4090 av_log(avctx, AV_LOG_DEBUG, "Unknown NAL code: %d (%d bits)\n",
4091 hx->nal_unit_type, bit_length);
4094 if (context_count == h->max_contexts) {
4095 execute_decode_slices(h, context_count);
4100 av_log(h->s.avctx, AV_LOG_ERROR, "decode_slice_header error\n");
4101 else if (err == 1) {
4102 /* Slice could not be decoded in parallel mode, copy down
4103 * NAL unit stuff to context 0 and restart. Note that
4104 * rbsp_buffer is not transferred, but since we no longer
4105 * run in parallel mode this should not be an issue. */
4106 h->nal_unit_type = hx->nal_unit_type;
4107 h->nal_ref_idc = hx->nal_ref_idc;
4114 execute_decode_slices(h, context_count);
4118 if (s->current_picture_ptr && s->current_picture_ptr->owner2 == s &&
4120 ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX,
4121 s->picture_structure == PICT_BOTTOM_FIELD);
4128 * Return the number of bytes consumed for building the current frame.
4130 static int get_consumed_bytes(MpegEncContext *s, int pos, int buf_size)
4133 pos = 1; // avoid infinite loops (i doubt that is needed but ...)
4134 if (pos + 10 > buf_size)
4135 pos = buf_size; // oops ;)
4140 static int decode_frame(AVCodecContext *avctx, void *data,
4141 int *got_frame, AVPacket *avpkt)
4143 const uint8_t *buf = avpkt->data;
4144 int buf_size = avpkt->size;
4145 H264Context *h = avctx->priv_data;
4146 MpegEncContext *s = &h->s;
4147 AVFrame *pict = data;
4150 s->flags = avctx->flags;
4151 s->flags2 = avctx->flags2;
4153 /* end of stream, output what is still in the buffers */
4155 if (buf_size == 0) {
4159 s->current_picture_ptr = NULL;
4161 // FIXME factorize this with the output code below
4162 out = h->delayed_pic[0];
4165 h->delayed_pic[i] &&
4166 !h->delayed_pic[i]->f.key_frame &&
4167 !h->delayed_pic[i]->mmco_reset;
4169 if (h->delayed_pic[i]->poc < out->poc) {
4170 out = h->delayed_pic[i];
4174 for (i = out_idx; h->delayed_pic[i]; i++)
4175 h->delayed_pic[i] = h->delayed_pic[i + 1];
4185 buf_index = decode_nal_units(h, buf, buf_size, 0);
4189 if (!s->current_picture_ptr && h->nal_unit_type == NAL_END_SEQUENCE) {
4194 if (!(s->flags2 & CODEC_FLAG2_CHUNKS) && !s->current_picture_ptr) {
4195 if (avctx->skip_frame >= AVDISCARD_NONREF)
4197 av_log(avctx, AV_LOG_ERROR, "no frame!\n");
4201 if (!(s->flags2 & CODEC_FLAG2_CHUNKS) ||
4202 (s->mb_y >= s->mb_height && s->mb_height)) {
4203 if (s->flags2 & CODEC_FLAG2_CHUNKS)
4204 decode_postinit(h, 1);
4207 h->context_reinitialized = 0;
4209 if (!h->next_output_pic) {
4210 /* Wait for second field. */
4214 *pict = h->next_output_pic->f;
4218 assert(pict->data[0] || !*got_frame);
4219 ff_print_debug_info(s, pict);
4221 return get_consumed_bytes(s, buf_index, buf_size);
4224 av_cold void ff_h264_free_context(H264Context *h)
4228 free_tables(h, 1); // FIXME cleanup init stuff perhaps
4230 for (i = 0; i < MAX_SPS_COUNT; i++)
4231 av_freep(h->sps_buffers + i);
4233 for (i = 0; i < MAX_PPS_COUNT; i++)
4234 av_freep(h->pps_buffers + i);
4237 static av_cold int h264_decode_end(AVCodecContext *avctx)
4239 H264Context *h = avctx->priv_data;
4240 MpegEncContext *s = &h->s;
4242 ff_h264_free_context(h);
4244 ff_MPV_common_end(s);
4246 // memset(h, 0, sizeof(H264Context));
4251 static const AVProfile profiles[] = {
4252 { FF_PROFILE_H264_BASELINE, "Baseline" },
4253 { FF_PROFILE_H264_CONSTRAINED_BASELINE, "Constrained Baseline" },
4254 { FF_PROFILE_H264_MAIN, "Main" },
4255 { FF_PROFILE_H264_EXTENDED, "Extended" },
4256 { FF_PROFILE_H264_HIGH, "High" },
4257 { FF_PROFILE_H264_HIGH_10, "High 10" },
4258 { FF_PROFILE_H264_HIGH_10_INTRA, "High 10 Intra" },
4259 { FF_PROFILE_H264_HIGH_422, "High 4:2:2" },
4260 { FF_PROFILE_H264_HIGH_422_INTRA, "High 4:2:2 Intra" },
4261 { FF_PROFILE_H264_HIGH_444, "High 4:4:4" },
4262 { FF_PROFILE_H264_HIGH_444_PREDICTIVE, "High 4:4:4 Predictive" },
4263 { FF_PROFILE_H264_HIGH_444_INTRA, "High 4:4:4 Intra" },
4264 { FF_PROFILE_H264_CAVLC_444, "CAVLC 4:4:4" },
4265 { FF_PROFILE_UNKNOWN },
4268 AVCodec ff_h264_decoder = {
4270 .type = AVMEDIA_TYPE_VIDEO,
4271 .id = AV_CODEC_ID_H264,
4272 .priv_data_size = sizeof(H264Context),
4273 .init = ff_h264_decode_init,
4274 .close = h264_decode_end,
4275 .decode = decode_frame,
4276 .capabilities = /*CODEC_CAP_DRAW_HORIZ_BAND |*/ CODEC_CAP_DR1 |
4277 CODEC_CAP_DELAY | CODEC_CAP_SLICE_THREADS |
4278 CODEC_CAP_FRAME_THREADS,
4280 .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10"),
4281 .init_thread_copy = ONLY_IF_THREADS_ENABLED(decode_init_thread_copy),
4282 .update_thread_context = ONLY_IF_THREADS_ENABLED(decode_update_thread_context),
4283 .profiles = NULL_IF_CONFIG_SMALL(profiles),
4286 #if CONFIG_H264_VDPAU_DECODER
4287 AVCodec ff_h264_vdpau_decoder = {
4288 .name = "h264_vdpau",
4289 .type = AVMEDIA_TYPE_VIDEO,
4290 .id = AV_CODEC_ID_H264,
4291 .priv_data_size = sizeof(H264Context),
4292 .init = ff_h264_decode_init,
4293 .close = h264_decode_end,
4294 .decode = decode_frame,
4295 .capabilities = CODEC_CAP_DR1 | CODEC_CAP_DELAY | CODEC_CAP_HWACCEL_VDPAU,
4297 .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10 (VDPAU acceleration)"),
4298 .pix_fmts = (const enum AVPixelFormat[]) { AV_PIX_FMT_VDPAU_H264,
4300 .profiles = NULL_IF_CONFIG_SMALL(profiles),