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_up = (raw_my & 3) ? 2 : 0;
294 int filter_height_down = (raw_my & 3) ? 3 : 0;
295 int full_my = (raw_my >> 2) + y_offset;
296 int top = full_my - filter_height_up;
297 int bottom = full_my + filter_height_down + height;
299 return FFMAX(abs(top), bottom);
302 static inline void get_lowest_part_y(H264Context *h, int refs[2][48], int n,
303 int height, int y_offset, int list0,
304 int list1, int *nrefs)
306 MpegEncContext *const s = &h->s;
309 y_offset += 16 * (s->mb_y >> MB_FIELD);
312 int ref_n = h->ref_cache[0][scan8[n]];
313 Picture *ref = &h->ref_list[0][ref_n];
315 // Error resilience puts the current picture in the ref list.
316 // Don't try to wait on these as it will cause a deadlock.
317 // Fields can wait on each other, though.
318 if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
319 (ref->f.reference & 3) != s->picture_structure) {
320 my = get_lowest_part_list_y(h, ref, n, height, y_offset, 0);
321 if (refs[0][ref_n] < 0)
323 refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
328 int ref_n = h->ref_cache[1][scan8[n]];
329 Picture *ref = &h->ref_list[1][ref_n];
331 if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
332 (ref->f.reference & 3) != s->picture_structure) {
333 my = get_lowest_part_list_y(h, ref, n, height, y_offset, 1);
334 if (refs[1][ref_n] < 0)
336 refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
342 * Wait until all reference frames are available for MC operations.
344 * @param h the H264 context
346 static void await_references(H264Context *h)
348 MpegEncContext *const s = &h->s;
349 const int mb_xy = h->mb_xy;
350 const int mb_type = s->current_picture.f.mb_type[mb_xy];
352 int nrefs[2] = { 0 };
355 memset(refs, -1, sizeof(refs));
357 if (IS_16X16(mb_type)) {
358 get_lowest_part_y(h, refs, 0, 16, 0,
359 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
360 } else if (IS_16X8(mb_type)) {
361 get_lowest_part_y(h, refs, 0, 8, 0,
362 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
363 get_lowest_part_y(h, refs, 8, 8, 8,
364 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
365 } else if (IS_8X16(mb_type)) {
366 get_lowest_part_y(h, refs, 0, 16, 0,
367 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
368 get_lowest_part_y(h, refs, 4, 16, 0,
369 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
373 assert(IS_8X8(mb_type));
375 for (i = 0; i < 4; i++) {
376 const int sub_mb_type = h->sub_mb_type[i];
378 int y_offset = (i & 2) << 2;
380 if (IS_SUB_8X8(sub_mb_type)) {
381 get_lowest_part_y(h, refs, n, 8, y_offset,
382 IS_DIR(sub_mb_type, 0, 0),
383 IS_DIR(sub_mb_type, 0, 1),
385 } else if (IS_SUB_8X4(sub_mb_type)) {
386 get_lowest_part_y(h, refs, n, 4, y_offset,
387 IS_DIR(sub_mb_type, 0, 0),
388 IS_DIR(sub_mb_type, 0, 1),
390 get_lowest_part_y(h, refs, n + 2, 4, y_offset + 4,
391 IS_DIR(sub_mb_type, 0, 0),
392 IS_DIR(sub_mb_type, 0, 1),
394 } else if (IS_SUB_4X8(sub_mb_type)) {
395 get_lowest_part_y(h, refs, n, 8, y_offset,
396 IS_DIR(sub_mb_type, 0, 0),
397 IS_DIR(sub_mb_type, 0, 1),
399 get_lowest_part_y(h, refs, n + 1, 8, y_offset,
400 IS_DIR(sub_mb_type, 0, 0),
401 IS_DIR(sub_mb_type, 0, 1),
405 assert(IS_SUB_4X4(sub_mb_type));
406 for (j = 0; j < 4; j++) {
407 int sub_y_offset = y_offset + 2 * (j & 2);
408 get_lowest_part_y(h, refs, n + j, 4, sub_y_offset,
409 IS_DIR(sub_mb_type, 0, 0),
410 IS_DIR(sub_mb_type, 0, 1),
417 for (list = h->list_count - 1; list >= 0; list--)
418 for (ref = 0; ref < 48 && nrefs[list]; ref++) {
419 int row = refs[list][ref];
421 Picture *ref_pic = &h->ref_list[list][ref];
422 int ref_field = ref_pic->f.reference - 1;
423 int ref_field_picture = ref_pic->field_picture;
424 int pic_height = 16 * s->mb_height >> ref_field_picture;
429 if (!FIELD_PICTURE && ref_field_picture) { // frame referencing two fields
430 ff_thread_await_progress(&ref_pic->f,
431 FFMIN((row >> 1) - !(row & 1),
434 ff_thread_await_progress(&ref_pic->f,
435 FFMIN((row >> 1), pic_height - 1),
437 } else if (FIELD_PICTURE && !ref_field_picture) { // field referencing one field of a frame
438 ff_thread_await_progress(&ref_pic->f,
439 FFMIN(row * 2 + ref_field,
442 } else if (FIELD_PICTURE) {
443 ff_thread_await_progress(&ref_pic->f,
444 FFMIN(row, pic_height - 1),
447 ff_thread_await_progress(&ref_pic->f,
448 FFMIN(row, pic_height - 1),
455 static av_always_inline void mc_dir_part(H264Context *h, Picture *pic,
456 int n, int square, int height,
458 uint8_t *dest_y, uint8_t *dest_cb,
460 int src_x_offset, int src_y_offset,
461 qpel_mc_func *qpix_op,
462 h264_chroma_mc_func chroma_op,
463 int pixel_shift, int chroma_idc)
465 MpegEncContext *const s = &h->s;
466 const int mx = h->mv_cache[list][scan8[n]][0] + src_x_offset * 8;
467 int my = h->mv_cache[list][scan8[n]][1] + src_y_offset * 8;
468 const int luma_xy = (mx & 3) + ((my & 3) << 2);
469 int offset = ((mx >> 2) << pixel_shift) + (my >> 2) * h->mb_linesize;
470 uint8_t *src_y = pic->f.data[0] + offset;
471 uint8_t *src_cb, *src_cr;
472 int extra_width = h->emu_edge_width;
473 int extra_height = h->emu_edge_height;
475 const int full_mx = mx >> 2;
476 const int full_my = my >> 2;
477 const int pic_width = 16 * s->mb_width;
478 const int pic_height = 16 * s->mb_height >> MB_FIELD;
486 if (full_mx < 0 - extra_width ||
487 full_my < 0 - extra_height ||
488 full_mx + 16 /*FIXME*/ > pic_width + extra_width ||
489 full_my + 16 /*FIXME*/ > pic_height + extra_height) {
490 s->vdsp.emulated_edge_mc(s->edge_emu_buffer,
491 src_y - (2 << pixel_shift) - 2 * h->mb_linesize,
493 16 + 5, 16 + 5 /*FIXME*/, full_mx - 2,
494 full_my - 2, pic_width, pic_height);
495 src_y = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
499 qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); // FIXME try variable height perhaps?
501 qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
503 if (CONFIG_GRAY && s->flags & CODEC_FLAG_GRAY)
506 if (chroma_idc == 3 /* yuv444 */) {
507 src_cb = pic->f.data[1] + offset;
509 s->vdsp.emulated_edge_mc(s->edge_emu_buffer,
510 src_cb - (2 << pixel_shift) - 2 * h->mb_linesize,
512 16 + 5, 16 + 5 /*FIXME*/,
513 full_mx - 2, full_my - 2,
514 pic_width, pic_height);
515 src_cb = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
517 qpix_op[luma_xy](dest_cb, src_cb, h->mb_linesize); // FIXME try variable height perhaps?
519 qpix_op[luma_xy](dest_cb + delta, src_cb + delta, h->mb_linesize);
521 src_cr = pic->f.data[2] + offset;
523 s->vdsp.emulated_edge_mc(s->edge_emu_buffer,
524 src_cr - (2 << pixel_shift) - 2 * h->mb_linesize,
526 16 + 5, 16 + 5 /*FIXME*/,
527 full_mx - 2, full_my - 2,
528 pic_width, pic_height);
529 src_cr = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
531 qpix_op[luma_xy](dest_cr, src_cr, h->mb_linesize); // FIXME try variable height perhaps?
533 qpix_op[luma_xy](dest_cr + delta, src_cr + delta, h->mb_linesize);
537 ysh = 3 - (chroma_idc == 2 /* yuv422 */);
538 if (chroma_idc == 1 /* yuv420 */ && MB_FIELD) {
539 // chroma offset when predicting from a field of opposite parity
540 my += 2 * ((s->mb_y & 1) - (pic->f.reference - 1));
541 emu |= (my >> 3) < 0 || (my >> 3) + 8 >= (pic_height >> 1);
544 src_cb = pic->f.data[1] + ((mx >> 3) << pixel_shift) +
545 (my >> ysh) * h->mb_uvlinesize;
546 src_cr = pic->f.data[2] + ((mx >> 3) << pixel_shift) +
547 (my >> ysh) * h->mb_uvlinesize;
550 s->vdsp.emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize,
551 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
552 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
553 src_cb = s->edge_emu_buffer;
555 chroma_op(dest_cb, src_cb, h->mb_uvlinesize,
556 height >> (chroma_idc == 1 /* yuv420 */),
557 mx & 7, (my << (chroma_idc == 2 /* yuv422 */)) & 7);
560 s->vdsp.emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize,
561 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
562 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
563 src_cr = s->edge_emu_buffer;
565 chroma_op(dest_cr, src_cr, h->mb_uvlinesize, height >> (chroma_idc == 1 /* yuv420 */),
566 mx & 7, (my << (chroma_idc == 2 /* yuv422 */)) & 7);
569 static av_always_inline void mc_part_std(H264Context *h, int n, int square,
570 int height, int delta,
571 uint8_t *dest_y, uint8_t *dest_cb,
573 int x_offset, int y_offset,
574 qpel_mc_func *qpix_put,
575 h264_chroma_mc_func chroma_put,
576 qpel_mc_func *qpix_avg,
577 h264_chroma_mc_func chroma_avg,
578 int list0, int list1,
579 int pixel_shift, int chroma_idc)
581 MpegEncContext *const s = &h->s;
582 qpel_mc_func *qpix_op = qpix_put;
583 h264_chroma_mc_func chroma_op = chroma_put;
585 dest_y += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
586 if (chroma_idc == 3 /* yuv444 */) {
587 dest_cb += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
588 dest_cr += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
589 } else if (chroma_idc == 2 /* yuv422 */) {
590 dest_cb += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
591 dest_cr += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
592 } else { /* yuv420 */
593 dest_cb += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
594 dest_cr += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
596 x_offset += 8 * s->mb_x;
597 y_offset += 8 * (s->mb_y >> MB_FIELD);
600 Picture *ref = &h->ref_list[0][h->ref_cache[0][scan8[n]]];
601 mc_dir_part(h, ref, n, square, height, delta, 0,
602 dest_y, dest_cb, dest_cr, x_offset, y_offset,
603 qpix_op, chroma_op, pixel_shift, chroma_idc);
606 chroma_op = chroma_avg;
610 Picture *ref = &h->ref_list[1][h->ref_cache[1][scan8[n]]];
611 mc_dir_part(h, ref, n, square, height, delta, 1,
612 dest_y, dest_cb, dest_cr, x_offset, y_offset,
613 qpix_op, chroma_op, pixel_shift, chroma_idc);
617 static av_always_inline void mc_part_weighted(H264Context *h, int n, int square,
618 int height, int delta,
619 uint8_t *dest_y, uint8_t *dest_cb,
621 int x_offset, int y_offset,
622 qpel_mc_func *qpix_put,
623 h264_chroma_mc_func chroma_put,
624 h264_weight_func luma_weight_op,
625 h264_weight_func chroma_weight_op,
626 h264_biweight_func luma_weight_avg,
627 h264_biweight_func chroma_weight_avg,
628 int list0, int list1,
629 int pixel_shift, int chroma_idc)
631 MpegEncContext *const s = &h->s;
634 dest_y += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
635 if (chroma_idc == 3 /* yuv444 */) {
636 chroma_height = height;
637 chroma_weight_avg = luma_weight_avg;
638 chroma_weight_op = luma_weight_op;
639 dest_cb += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
640 dest_cr += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
641 } else if (chroma_idc == 2 /* yuv422 */) {
642 chroma_height = height;
643 dest_cb += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
644 dest_cr += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
645 } else { /* yuv420 */
646 chroma_height = height >> 1;
647 dest_cb += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
648 dest_cr += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
650 x_offset += 8 * s->mb_x;
651 y_offset += 8 * (s->mb_y >> MB_FIELD);
653 if (list0 && list1) {
654 /* don't optimize for luma-only case, since B-frames usually
655 * use implicit weights => chroma too. */
656 uint8_t *tmp_cb = h->bipred_scratchpad;
657 uint8_t *tmp_cr = h->bipred_scratchpad + (16 << pixel_shift);
658 uint8_t *tmp_y = h->bipred_scratchpad + 16 * h->mb_uvlinesize;
659 int refn0 = h->ref_cache[0][scan8[n]];
660 int refn1 = h->ref_cache[1][scan8[n]];
662 mc_dir_part(h, &h->ref_list[0][refn0], n, square, height, delta, 0,
663 dest_y, dest_cb, dest_cr,
664 x_offset, y_offset, qpix_put, chroma_put,
665 pixel_shift, chroma_idc);
666 mc_dir_part(h, &h->ref_list[1][refn1], n, square, height, delta, 1,
667 tmp_y, tmp_cb, tmp_cr,
668 x_offset, y_offset, qpix_put, chroma_put,
669 pixel_shift, chroma_idc);
671 if (h->use_weight == 2) {
672 int weight0 = h->implicit_weight[refn0][refn1][s->mb_y & 1];
673 int weight1 = 64 - weight0;
674 luma_weight_avg(dest_y, tmp_y, h->mb_linesize,
675 height, 5, weight0, weight1, 0);
676 chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize,
677 chroma_height, 5, weight0, weight1, 0);
678 chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize,
679 chroma_height, 5, weight0, weight1, 0);
681 luma_weight_avg(dest_y, tmp_y, h->mb_linesize, height,
682 h->luma_log2_weight_denom,
683 h->luma_weight[refn0][0][0],
684 h->luma_weight[refn1][1][0],
685 h->luma_weight[refn0][0][1] +
686 h->luma_weight[refn1][1][1]);
687 chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, chroma_height,
688 h->chroma_log2_weight_denom,
689 h->chroma_weight[refn0][0][0][0],
690 h->chroma_weight[refn1][1][0][0],
691 h->chroma_weight[refn0][0][0][1] +
692 h->chroma_weight[refn1][1][0][1]);
693 chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, chroma_height,
694 h->chroma_log2_weight_denom,
695 h->chroma_weight[refn0][0][1][0],
696 h->chroma_weight[refn1][1][1][0],
697 h->chroma_weight[refn0][0][1][1] +
698 h->chroma_weight[refn1][1][1][1]);
701 int list = list1 ? 1 : 0;
702 int refn = h->ref_cache[list][scan8[n]];
703 Picture *ref = &h->ref_list[list][refn];
704 mc_dir_part(h, ref, n, square, height, delta, list,
705 dest_y, dest_cb, dest_cr, x_offset, y_offset,
706 qpix_put, chroma_put, pixel_shift, chroma_idc);
708 luma_weight_op(dest_y, h->mb_linesize, height,
709 h->luma_log2_weight_denom,
710 h->luma_weight[refn][list][0],
711 h->luma_weight[refn][list][1]);
712 if (h->use_weight_chroma) {
713 chroma_weight_op(dest_cb, h->mb_uvlinesize, chroma_height,
714 h->chroma_log2_weight_denom,
715 h->chroma_weight[refn][list][0][0],
716 h->chroma_weight[refn][list][0][1]);
717 chroma_weight_op(dest_cr, h->mb_uvlinesize, chroma_height,
718 h->chroma_log2_weight_denom,
719 h->chroma_weight[refn][list][1][0],
720 h->chroma_weight[refn][list][1][1]);
725 static av_always_inline void prefetch_motion(H264Context *h, int list,
726 int pixel_shift, int chroma_idc)
728 /* fetch pixels for estimated mv 4 macroblocks ahead
729 * optimized for 64byte cache lines */
730 MpegEncContext *const s = &h->s;
731 const int refn = h->ref_cache[list][scan8[0]];
733 const int mx = (h->mv_cache[list][scan8[0]][0] >> 2) + 16 * s->mb_x + 8;
734 const int my = (h->mv_cache[list][scan8[0]][1] >> 2) + 16 * s->mb_y;
735 uint8_t **src = h->ref_list[list][refn].f.data;
736 int off = (mx << pixel_shift) +
737 (my + (s->mb_x & 3) * 4) * h->mb_linesize +
739 s->vdsp.prefetch(src[0] + off, s->linesize, 4);
740 if (chroma_idc == 3 /* yuv444 */) {
741 s->vdsp.prefetch(src[1] + off, s->linesize, 4);
742 s->vdsp.prefetch(src[2] + off, s->linesize, 4);
744 off = ((mx >> 1) << pixel_shift) +
745 ((my >> 1) + (s->mb_x & 7)) * s->uvlinesize +
747 s->vdsp.prefetch(src[1] + off, src[2] - src[1], 2);
752 static void free_tables(H264Context *h, int free_rbsp)
757 av_freep(&h->intra4x4_pred_mode);
758 av_freep(&h->chroma_pred_mode_table);
759 av_freep(&h->cbp_table);
760 av_freep(&h->mvd_table[0]);
761 av_freep(&h->mvd_table[1]);
762 av_freep(&h->direct_table);
763 av_freep(&h->non_zero_count);
764 av_freep(&h->slice_table_base);
765 h->slice_table = NULL;
766 av_freep(&h->list_counts);
768 av_freep(&h->mb2b_xy);
769 av_freep(&h->mb2br_xy);
771 for (i = 0; i < MAX_THREADS; i++) {
772 hx = h->thread_context[i];
775 av_freep(&hx->top_borders[1]);
776 av_freep(&hx->top_borders[0]);
777 av_freep(&hx->bipred_scratchpad);
779 av_freep(&hx->rbsp_buffer[1]);
780 av_freep(&hx->rbsp_buffer[0]);
781 hx->rbsp_buffer_size[0] = 0;
782 hx->rbsp_buffer_size[1] = 0;
785 av_freep(&h->thread_context[i]);
789 static void init_dequant8_coeff_table(H264Context *h)
792 const int max_qp = 51 + 6 * (h->sps.bit_depth_luma - 8);
794 for (i = 0; i < 6; i++) {
795 h->dequant8_coeff[i] = h->dequant8_buffer[i];
796 for (j = 0; j < i; j++)
797 if (!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i],
798 64 * sizeof(uint8_t))) {
799 h->dequant8_coeff[i] = h->dequant8_buffer[j];
805 for (q = 0; q < max_qp + 1; q++) {
808 for (x = 0; x < 64; x++)
809 h->dequant8_coeff[i][q][(x >> 3) | ((x & 7) << 3)] =
810 ((uint32_t)dequant8_coeff_init[idx][dequant8_coeff_init_scan[((x >> 1) & 12) | (x & 3)]] *
811 h->pps.scaling_matrix8[i][x]) << shift;
816 static void init_dequant4_coeff_table(H264Context *h)
819 const int max_qp = 51 + 6 * (h->sps.bit_depth_luma - 8);
820 for (i = 0; i < 6; i++) {
821 h->dequant4_coeff[i] = h->dequant4_buffer[i];
822 for (j = 0; j < i; j++)
823 if (!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i],
824 16 * sizeof(uint8_t))) {
825 h->dequant4_coeff[i] = h->dequant4_buffer[j];
831 for (q = 0; q < max_qp + 1; q++) {
832 int shift = div6[q] + 2;
834 for (x = 0; x < 16; x++)
835 h->dequant4_coeff[i][q][(x >> 2) | ((x << 2) & 0xF)] =
836 ((uint32_t)dequant4_coeff_init[idx][(x & 1) + ((x >> 2) & 1)] *
837 h->pps.scaling_matrix4[i][x]) << shift;
842 static void init_dequant_tables(H264Context *h)
845 init_dequant4_coeff_table(h);
846 if (h->pps.transform_8x8_mode)
847 init_dequant8_coeff_table(h);
848 if (h->sps.transform_bypass) {
849 for (i = 0; i < 6; i++)
850 for (x = 0; x < 16; x++)
851 h->dequant4_coeff[i][0][x] = 1 << 6;
852 if (h->pps.transform_8x8_mode)
853 for (i = 0; i < 6; i++)
854 for (x = 0; x < 64; x++)
855 h->dequant8_coeff[i][0][x] = 1 << 6;
859 int ff_h264_alloc_tables(H264Context *h)
861 MpegEncContext *const s = &h->s;
862 const int big_mb_num = s->mb_stride * (s->mb_height + 1);
863 const int row_mb_num = s->mb_stride * 2 * s->avctx->thread_count;
866 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode,
867 row_mb_num * 8 * sizeof(uint8_t), fail)
868 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count,
869 big_mb_num * 48 * sizeof(uint8_t), fail)
870 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base,
871 (big_mb_num + s->mb_stride) * sizeof(*h->slice_table_base), fail)
872 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table,
873 big_mb_num * sizeof(uint16_t), fail)
874 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table,
875 big_mb_num * sizeof(uint8_t), fail)
876 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0],
877 16 * row_mb_num * sizeof(uint8_t), fail);
878 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1],
879 16 * row_mb_num * sizeof(uint8_t), fail);
880 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table,
881 4 * big_mb_num * sizeof(uint8_t), fail);
882 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts,
883 big_mb_num * sizeof(uint8_t), fail)
885 memset(h->slice_table_base, -1,
886 (big_mb_num + s->mb_stride) * sizeof(*h->slice_table_base));
887 h->slice_table = h->slice_table_base + s->mb_stride * 2 + 1;
889 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy,
890 big_mb_num * sizeof(uint32_t), fail);
891 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy,
892 big_mb_num * sizeof(uint32_t), fail);
893 for (y = 0; y < s->mb_height; y++)
894 for (x = 0; x < s->mb_width; x++) {
895 const int mb_xy = x + y * s->mb_stride;
896 const int b_xy = 4 * x + 4 * y * h->b_stride;
898 h->mb2b_xy[mb_xy] = b_xy;
899 h->mb2br_xy[mb_xy] = 8 * (FMO ? mb_xy : (mb_xy % (2 * s->mb_stride)));
902 if (!h->dequant4_coeff[0])
903 init_dequant_tables(h);
913 * Mimic alloc_tables(), but for every context thread.
915 static void clone_tables(H264Context *dst, H264Context *src, int i)
917 MpegEncContext *const s = &src->s;
918 dst->intra4x4_pred_mode = src->intra4x4_pred_mode + i * 8 * 2 * s->mb_stride;
919 dst->non_zero_count = src->non_zero_count;
920 dst->slice_table = src->slice_table;
921 dst->cbp_table = src->cbp_table;
922 dst->mb2b_xy = src->mb2b_xy;
923 dst->mb2br_xy = src->mb2br_xy;
924 dst->chroma_pred_mode_table = src->chroma_pred_mode_table;
925 dst->mvd_table[0] = src->mvd_table[0] + i * 8 * 2 * s->mb_stride;
926 dst->mvd_table[1] = src->mvd_table[1] + i * 8 * 2 * s->mb_stride;
927 dst->direct_table = src->direct_table;
928 dst->list_counts = src->list_counts;
929 dst->bipred_scratchpad = NULL;
930 ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma,
931 src->sps.chroma_format_idc);
936 * Allocate buffers which are not shared amongst multiple threads.
938 static int context_init(H264Context *h)
940 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0],
941 h->s.mb_width * 16 * 3 * sizeof(uint8_t) * 2, fail)
942 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1],
943 h->s.mb_width * 16 * 3 * sizeof(uint8_t) * 2, fail)
945 h->ref_cache[0][scan8[5] + 1] =
946 h->ref_cache[0][scan8[7] + 1] =
947 h->ref_cache[0][scan8[13] + 1] =
948 h->ref_cache[1][scan8[5] + 1] =
949 h->ref_cache[1][scan8[7] + 1] =
950 h->ref_cache[1][scan8[13] + 1] = PART_NOT_AVAILABLE;
955 return -1; // free_tables will clean up for us
958 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size,
959 int parse_extradata);
961 static av_cold void common_init(H264Context *h)
963 MpegEncContext *const s = &h->s;
965 s->width = s->avctx->width;
966 s->height = s->avctx->height;
967 s->codec_id = s->avctx->codec->id;
969 ff_h264dsp_init(&h->h264dsp, 8, 1);
970 ff_h264_pred_init(&h->hpc, s->codec_id, 8, 1);
972 h->dequant_coeff_pps = -1;
973 s->unrestricted_mv = 1;
975 /* needed so that IDCT permutation is known early */
976 ff_dsputil_init(&s->dsp, s->avctx);
977 ff_videodsp_init(&s->vdsp, 8);
979 memset(h->pps.scaling_matrix4, 16, 6 * 16 * sizeof(uint8_t));
980 memset(h->pps.scaling_matrix8, 16, 2 * 64 * sizeof(uint8_t));
983 int ff_h264_decode_extradata(H264Context *h)
985 AVCodecContext *avctx = h->s.avctx;
987 if (avctx->extradata[0] == 1) {
989 unsigned char *p = avctx->extradata;
993 if (avctx->extradata_size < 7) {
994 av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
997 /* sps and pps in the avcC always have length coded with 2 bytes,
998 * so put a fake nal_length_size = 2 while parsing them */
999 h->nal_length_size = 2;
1000 // Decode sps from avcC
1001 cnt = *(p + 5) & 0x1f; // Number of sps
1003 for (i = 0; i < cnt; i++) {
1004 nalsize = AV_RB16(p) + 2;
1005 if (p - avctx->extradata + nalsize > avctx->extradata_size)
1007 if (decode_nal_units(h, p, nalsize, 1) < 0) {
1008 av_log(avctx, AV_LOG_ERROR,
1009 "Decoding sps %d from avcC failed\n", i);
1014 // Decode pps from avcC
1015 cnt = *(p++); // Number of pps
1016 for (i = 0; i < cnt; i++) {
1017 nalsize = AV_RB16(p) + 2;
1018 if (p - avctx->extradata + nalsize > avctx->extradata_size)
1020 if (decode_nal_units(h, p, nalsize, 1) < 0) {
1021 av_log(avctx, AV_LOG_ERROR,
1022 "Decoding pps %d from avcC failed\n", i);
1027 // Now store right nal length size, that will be used to parse all other nals
1028 h->nal_length_size = (avctx->extradata[4] & 0x03) + 1;
1031 if (decode_nal_units(h, avctx->extradata, avctx->extradata_size, 1) < 0)
1037 av_cold int ff_h264_decode_init(AVCodecContext *avctx)
1039 H264Context *h = avctx->priv_data;
1040 MpegEncContext *const s = &h->s;
1043 ff_MPV_decode_defaults(s);
1048 s->out_format = FMT_H264;
1049 s->workaround_bugs = avctx->workaround_bugs;
1052 // s->decode_mb = ff_h263_decode_mb;
1053 s->quarter_sample = 1;
1054 if (!avctx->has_b_frames)
1057 avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1059 ff_h264_decode_init_vlc();
1062 h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1064 h->thread_context[0] = h;
1065 h->outputed_poc = h->next_outputed_poc = INT_MIN;
1066 for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
1067 h->last_pocs[i] = INT_MIN;
1068 h->prev_poc_msb = 1 << 16;
1070 ff_h264_reset_sei(h);
1071 if (avctx->codec_id == AV_CODEC_ID_H264) {
1072 if (avctx->ticks_per_frame == 1)
1073 s->avctx->time_base.den *= 2;
1074 avctx->ticks_per_frame = 2;
1077 if (avctx->extradata_size > 0 && avctx->extradata &&
1078 ff_h264_decode_extradata(h))
1081 if (h->sps.bitstream_restriction_flag &&
1082 s->avctx->has_b_frames < h->sps.num_reorder_frames) {
1083 s->avctx->has_b_frames = h->sps.num_reorder_frames;
1090 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b) + (size))))
1092 static void copy_picture_range(Picture **to, Picture **from, int count,
1093 MpegEncContext *new_base,
1094 MpegEncContext *old_base)
1098 for (i = 0; i < count; i++) {
1099 assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1100 IN_RANGE(from[i], old_base->picture,
1101 sizeof(Picture) * old_base->picture_count) ||
1103 to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1107 static void copy_parameter_set(void **to, void **from, int count, int size)
1111 for (i = 0; i < count; i++) {
1112 if (to[i] && !from[i])
1114 else if (from[i] && !to[i])
1115 to[i] = av_malloc(size);
1118 memcpy(to[i], from[i], size);
1122 static int decode_init_thread_copy(AVCodecContext *avctx)
1124 H264Context *h = avctx->priv_data;
1126 if (!avctx->internal->is_copy)
1128 memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1129 memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1131 h->s.context_initialized = 0;
1136 #define copy_fields(to, from, start_field, end_field) \
1137 memcpy(&to->start_field, &from->start_field, \
1138 (char *)&to->end_field - (char *)&to->start_field)
1140 static int h264_slice_header_init(H264Context *, int);
1142 static int h264_set_parameter_from_sps(H264Context *h);
1144 static int decode_update_thread_context(AVCodecContext *dst,
1145 const AVCodecContext *src)
1147 H264Context *h = dst->priv_data, *h1 = src->priv_data;
1148 MpegEncContext *const s = &h->s, *const s1 = &h1->s;
1149 int inited = s->context_initialized, err;
1152 if (dst == src || !s1->context_initialized)
1156 (s->width != s1->width ||
1157 s->height != s1->height ||
1158 s->mb_width != s1->mb_width ||
1159 s->mb_height != s1->mb_height ||
1160 h->sps.bit_depth_luma != h1->sps.bit_depth_luma ||
1161 h->sps.chroma_format_idc != h1->sps.chroma_format_idc ||
1162 h->sps.colorspace != h1->sps.colorspace)) {
1164 av_freep(&h->bipred_scratchpad);
1166 s->width = s1->width;
1167 s->height = s1->height;
1168 s->mb_height = s1->mb_height;
1169 h->b_stride = h1->b_stride;
1171 if ((err = h264_slice_header_init(h, 1)) < 0) {
1172 av_log(h->s.avctx, AV_LOG_ERROR, "h264_slice_header_init() failed");
1175 h->context_reinitialized = 1;
1177 /* update linesize on resize for h264. The h264 decoder doesn't
1178 * necessarily call ff_MPV_frame_start in the new thread */
1179 s->linesize = s1->linesize;
1180 s->uvlinesize = s1->uvlinesize;
1182 /* copy block_offset since frame_start may not be called */
1183 memcpy(h->block_offset, h1->block_offset, sizeof(h->block_offset));
1184 h264_set_parameter_from_sps(h);
1187 err = ff_mpeg_update_thread_context(dst, src);
1192 for (i = 0; i < MAX_SPS_COUNT; i++)
1193 av_freep(h->sps_buffers + i);
1195 for (i = 0; i < MAX_PPS_COUNT; i++)
1196 av_freep(h->pps_buffers + i);
1198 // copy all fields after MpegEnc
1199 memcpy(&h->s + 1, &h1->s + 1,
1200 sizeof(H264Context) - sizeof(MpegEncContext));
1201 memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1202 memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1203 if (ff_h264_alloc_tables(h) < 0) {
1204 av_log(dst, AV_LOG_ERROR, "Could not allocate memory for h264\n");
1205 return AVERROR(ENOMEM);
1209 for (i = 0; i < 2; i++) {
1210 h->rbsp_buffer[i] = NULL;
1211 h->rbsp_buffer_size[i] = 0;
1213 h->bipred_scratchpad = NULL;
1215 h->thread_context[0] = h;
1217 s->dsp.clear_blocks(h->mb);
1218 s->dsp.clear_blocks(h->mb + (24 * 16 << h->pixel_shift));
1221 /* frame_start may not be called for the next thread (if it's decoding
1222 * a bottom field) so this has to be allocated here */
1223 if (!h->bipred_scratchpad)
1224 h->bipred_scratchpad = av_malloc(16 * 6 * s->linesize);
1226 // extradata/NAL handling
1227 h->is_avc = h1->is_avc;
1230 copy_parameter_set((void **)h->sps_buffers, (void **)h1->sps_buffers,
1231 MAX_SPS_COUNT, sizeof(SPS));
1233 copy_parameter_set((void **)h->pps_buffers, (void **)h1->pps_buffers,
1234 MAX_PPS_COUNT, sizeof(PPS));
1237 // Dequantization matrices
1238 // FIXME these are big - can they be only copied when PPS changes?
1239 copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1241 for (i = 0; i < 6; i++)
1242 h->dequant4_coeff[i] = h->dequant4_buffer[0] +
1243 (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1245 for (i = 0; i < 6; i++)
1246 h->dequant8_coeff[i] = h->dequant8_buffer[0] +
1247 (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1249 h->dequant_coeff_pps = h1->dequant_coeff_pps;
1252 copy_fields(h, h1, poc_lsb, redundant_pic_count);
1255 copy_fields(h, h1, ref_count, list_count);
1256 copy_fields(h, h1, ref2frm, intra_gb);
1257 copy_fields(h, h1, short_ref, cabac_init_idc);
1259 copy_picture_range(h->short_ref, h1->short_ref, 32, s, s1);
1260 copy_picture_range(h->long_ref, h1->long_ref, 32, s, s1);
1261 copy_picture_range(h->delayed_pic, h1->delayed_pic,
1262 MAX_DELAYED_PIC_COUNT + 2, s, s1);
1264 h->last_slice_type = h1->last_slice_type;
1266 if (!s->current_picture_ptr)
1269 if (!s->droppable) {
1270 err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1271 h->prev_poc_msb = h->poc_msb;
1272 h->prev_poc_lsb = h->poc_lsb;
1274 h->prev_frame_num_offset = h->frame_num_offset;
1275 h->prev_frame_num = h->frame_num;
1276 h->outputed_poc = h->next_outputed_poc;
1281 int ff_h264_frame_start(H264Context *h)
1283 MpegEncContext *const s = &h->s;
1285 const int pixel_shift = h->pixel_shift;
1287 if (ff_MPV_frame_start(s, s->avctx) < 0)
1289 ff_er_frame_start(s);
1291 * ff_MPV_frame_start uses pict_type to derive key_frame.
1292 * This is incorrect for H.264; IDR markings must be used.
1293 * Zero here; IDR markings per slice in frame or fields are ORed in later.
1294 * See decode_nal_units().
1296 s->current_picture_ptr->f.key_frame = 0;
1297 s->current_picture_ptr->mmco_reset = 0;
1299 assert(s->linesize && s->uvlinesize);
1301 for (i = 0; i < 16; i++) {
1302 h->block_offset[i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * s->linesize * ((scan8[i] - scan8[0]) >> 3);
1303 h->block_offset[48 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * s->linesize * ((scan8[i] - scan8[0]) >> 3);
1305 for (i = 0; i < 16; i++) {
1306 h->block_offset[16 + i] =
1307 h->block_offset[32 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * s->uvlinesize * ((scan8[i] - scan8[0]) >> 3);
1308 h->block_offset[48 + 16 + i] =
1309 h->block_offset[48 + 32 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * s->uvlinesize * ((scan8[i] - scan8[0]) >> 3);
1312 /* can't be in alloc_tables because linesize isn't known there.
1313 * FIXME: redo bipred weight to not require extra buffer? */
1314 for (i = 0; i < s->slice_context_count; i++)
1315 if (h->thread_context[i] && !h->thread_context[i]->bipred_scratchpad)
1316 h->thread_context[i]->bipred_scratchpad = av_malloc(16 * 6 * s->linesize);
1318 /* Some macroblocks can be accessed before they're available in case
1319 * of lost slices, MBAFF or threading. */
1320 memset(h->slice_table, -1,
1321 (s->mb_height * s->mb_stride - 1) * sizeof(*h->slice_table));
1323 // s->decode = (s->flags & CODEC_FLAG_PSNR) || !s->encoding ||
1324 // s->current_picture.f.reference /* || h->contains_intra */ || 1;
1326 /* We mark the current picture as non-reference after allocating it, so
1327 * that if we break out due to an error it can be released automatically
1328 * in the next ff_MPV_frame_start().
1329 * SVQ3 as well as most other codecs have only last/next/current and thus
1330 * get released even with set reference, besides SVQ3 and others do not
1331 * mark frames as reference later "naturally". */
1332 if (s->codec_id != AV_CODEC_ID_SVQ3)
1333 s->current_picture_ptr->f.reference = 0;
1335 s->current_picture_ptr->field_poc[0] =
1336 s->current_picture_ptr->field_poc[1] = INT_MAX;
1338 h->next_output_pic = NULL;
1340 assert(s->current_picture_ptr->long_ref == 0);
1346 * Run setup operations that must be run after slice header decoding.
1347 * This includes finding the next displayed frame.
1349 * @param h h264 master context
1350 * @param setup_finished enough NALs have been read that we can call
1351 * ff_thread_finish_setup()
1353 static void decode_postinit(H264Context *h, int setup_finished)
1355 MpegEncContext *const s = &h->s;
1356 Picture *out = s->current_picture_ptr;
1357 Picture *cur = s->current_picture_ptr;
1358 int i, pics, out_of_order, out_idx;
1359 int invalid = 0, cnt = 0;
1361 s->current_picture_ptr->f.qscale_type = FF_QSCALE_TYPE_H264;
1362 s->current_picture_ptr->f.pict_type = s->pict_type;
1364 if (h->next_output_pic)
1367 if (cur->field_poc[0] == INT_MAX || cur->field_poc[1] == INT_MAX) {
1368 /* FIXME: if we have two PAFF fields in one packet, we can't start
1369 * the next thread here. If we have one field per packet, we can.
1370 * The check in decode_nal_units() is not good enough to find this
1371 * yet, so we assume the worst for now. */
1372 // if (setup_finished)
1373 // ff_thread_finish_setup(s->avctx);
1377 cur->f.interlaced_frame = 0;
1378 cur->f.repeat_pict = 0;
1380 /* Signal interlacing information externally. */
1381 /* Prioritize picture timing SEI information over used
1382 * decoding process if it exists. */
1384 if (h->sps.pic_struct_present_flag) {
1385 switch (h->sei_pic_struct) {
1386 case SEI_PIC_STRUCT_FRAME:
1388 case SEI_PIC_STRUCT_TOP_FIELD:
1389 case SEI_PIC_STRUCT_BOTTOM_FIELD:
1390 cur->f.interlaced_frame = 1;
1392 case SEI_PIC_STRUCT_TOP_BOTTOM:
1393 case SEI_PIC_STRUCT_BOTTOM_TOP:
1394 if (FIELD_OR_MBAFF_PICTURE)
1395 cur->f.interlaced_frame = 1;
1397 // try to flag soft telecine progressive
1398 cur->f.interlaced_frame = h->prev_interlaced_frame;
1400 case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1401 case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1402 /* Signal the possibility of telecined film externally
1403 * (pic_struct 5,6). From these hints, let the applications
1404 * decide if they apply deinterlacing. */
1405 cur->f.repeat_pict = 1;
1407 case SEI_PIC_STRUCT_FRAME_DOUBLING:
1408 // Force progressive here, doubling interlaced frame is a bad idea.
1409 cur->f.repeat_pict = 2;
1411 case SEI_PIC_STRUCT_FRAME_TRIPLING:
1412 cur->f.repeat_pict = 4;
1416 if ((h->sei_ct_type & 3) &&
1417 h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1418 cur->f.interlaced_frame = (h->sei_ct_type & (1 << 1)) != 0;
1420 /* Derive interlacing flag from used decoding process. */
1421 cur->f.interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1423 h->prev_interlaced_frame = cur->f.interlaced_frame;
1425 if (cur->field_poc[0] != cur->field_poc[1]) {
1426 /* Derive top_field_first from field pocs. */
1427 cur->f.top_field_first = cur->field_poc[0] < cur->field_poc[1];
1429 if (cur->f.interlaced_frame || h->sps.pic_struct_present_flag) {
1430 /* Use picture timing SEI information. Even if it is a
1431 * information of a past frame, better than nothing. */
1432 if (h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM ||
1433 h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1434 cur->f.top_field_first = 1;
1436 cur->f.top_field_first = 0;
1438 /* Most likely progressive */
1439 cur->f.top_field_first = 0;
1443 // FIXME do something with unavailable reference frames
1445 /* Sort B-frames into display order */
1447 if (h->sps.bitstream_restriction_flag &&
1448 s->avctx->has_b_frames < h->sps.num_reorder_frames) {
1449 s->avctx->has_b_frames = h->sps.num_reorder_frames;
1453 if (s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT &&
1454 !h->sps.bitstream_restriction_flag) {
1455 s->avctx->has_b_frames = MAX_DELAYED_PIC_COUNT - 1;
1460 while (h->delayed_pic[pics])
1463 assert(pics <= MAX_DELAYED_PIC_COUNT);
1465 h->delayed_pic[pics++] = cur;
1466 if (cur->f.reference == 0)
1467 cur->f.reference = DELAYED_PIC_REF;
1469 /* Frame reordering. This code takes pictures from coding order and sorts
1470 * them by their incremental POC value into display order. It supports POC
1471 * gaps, MMCO reset codes and random resets.
1472 * A "display group" can start either with a IDR frame (f.key_frame = 1),
1473 * and/or can be closed down with a MMCO reset code. In sequences where
1474 * there is no delay, we can't detect that (since the frame was already
1475 * output to the user), so we also set h->mmco_reset to detect the MMCO
1477 * FIXME: if we detect insufficient delays (as per s->avctx->has_b_frames),
1478 * we increase the delay between input and output. All frames affected by
1479 * the lag (e.g. those that should have been output before another frame
1480 * that we already returned to the user) will be dropped. This is a bug
1481 * that we will fix later. */
1482 for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++) {
1483 cnt += out->poc < h->last_pocs[i];
1484 invalid += out->poc == INT_MIN;
1486 if (!h->mmco_reset && !cur->f.key_frame &&
1487 cnt + invalid == MAX_DELAYED_PIC_COUNT && cnt > 0) {
1490 h->delayed_pic[pics - 2]->mmco_reset = 2;
1492 if (h->mmco_reset || cur->f.key_frame) {
1493 for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
1494 h->last_pocs[i] = INT_MIN;
1496 invalid = MAX_DELAYED_PIC_COUNT;
1498 out = h->delayed_pic[0];
1500 for (i = 1; i < MAX_DELAYED_PIC_COUNT &&
1501 h->delayed_pic[i] &&
1502 !h->delayed_pic[i - 1]->mmco_reset &&
1503 !h->delayed_pic[i]->f.key_frame;
1505 if (h->delayed_pic[i]->poc < out->poc) {
1506 out = h->delayed_pic[i];
1509 if (s->avctx->has_b_frames == 0 &&
1510 (h->delayed_pic[0]->f.key_frame || h->mmco_reset))
1511 h->next_outputed_poc = INT_MIN;
1512 out_of_order = !out->f.key_frame && !h->mmco_reset &&
1513 (out->poc < h->next_outputed_poc);
1515 if (h->sps.bitstream_restriction_flag &&
1516 s->avctx->has_b_frames >= h->sps.num_reorder_frames) {
1517 } else if (out_of_order && pics - 1 == s->avctx->has_b_frames &&
1518 s->avctx->has_b_frames < MAX_DELAYED_PIC_COUNT) {
1519 if (invalid + cnt < MAX_DELAYED_PIC_COUNT) {
1520 s->avctx->has_b_frames = FFMAX(s->avctx->has_b_frames, cnt);
1523 } else if (s->low_delay &&
1524 ((h->next_outputed_poc != INT_MIN &&
1525 out->poc > h->next_outputed_poc + 2) ||
1526 cur->f.pict_type == AV_PICTURE_TYPE_B)) {
1528 s->avctx->has_b_frames++;
1531 if (pics > s->avctx->has_b_frames) {
1532 out->f.reference &= ~DELAYED_PIC_REF;
1533 // for frame threading, the owner must be the second field's thread or
1534 // else the first thread can release the picture and reuse it unsafely
1536 for (i = out_idx; h->delayed_pic[i]; i++)
1537 h->delayed_pic[i] = h->delayed_pic[i + 1];
1539 memmove(h->last_pocs, &h->last_pocs[1],
1540 sizeof(*h->last_pocs) * (MAX_DELAYED_PIC_COUNT - 1));
1541 h->last_pocs[MAX_DELAYED_PIC_COUNT - 1] = cur->poc;
1542 if (!out_of_order && pics > s->avctx->has_b_frames) {
1543 h->next_output_pic = out;
1544 if (out->mmco_reset) {
1546 h->next_outputed_poc = out->poc;
1547 h->delayed_pic[out_idx - 1]->mmco_reset = out->mmco_reset;
1549 h->next_outputed_poc = INT_MIN;
1552 if (out_idx == 0 && pics > 1 && h->delayed_pic[0]->f.key_frame) {
1553 h->next_outputed_poc = INT_MIN;
1555 h->next_outputed_poc = out->poc;
1560 av_log(s->avctx, AV_LOG_DEBUG, "no picture\n");
1564 ff_thread_finish_setup(s->avctx);
1567 static av_always_inline void backup_mb_border(H264Context *h, uint8_t *src_y,
1568 uint8_t *src_cb, uint8_t *src_cr,
1569 int linesize, int uvlinesize,
1572 MpegEncContext *const s = &h->s;
1573 uint8_t *top_border;
1575 const int pixel_shift = h->pixel_shift;
1576 int chroma444 = CHROMA444;
1577 int chroma422 = CHROMA422;
1580 src_cb -= uvlinesize;
1581 src_cr -= uvlinesize;
1583 if (!simple && FRAME_MBAFF) {
1586 top_border = h->top_borders[0][s->mb_x];
1587 AV_COPY128(top_border, src_y + 15 * linesize);
1589 AV_COPY128(top_border + 16, src_y + 15 * linesize + 16);
1590 if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1593 AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
1594 AV_COPY128(top_border + 48, src_cb + 15 * uvlinesize + 16);
1595 AV_COPY128(top_border + 64, src_cr + 15 * uvlinesize);
1596 AV_COPY128(top_border + 80, src_cr + 15 * uvlinesize + 16);
1598 AV_COPY128(top_border + 16, src_cb + 15 * uvlinesize);
1599 AV_COPY128(top_border + 32, src_cr + 15 * uvlinesize);
1601 } else if (chroma422) {
1603 AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
1604 AV_COPY128(top_border + 48, src_cr + 15 * uvlinesize);
1606 AV_COPY64(top_border + 16, src_cb + 15 * uvlinesize);
1607 AV_COPY64(top_border + 24, src_cr + 15 * uvlinesize);
1611 AV_COPY128(top_border + 32, src_cb + 7 * uvlinesize);
1612 AV_COPY128(top_border + 48, src_cr + 7 * uvlinesize);
1614 AV_COPY64(top_border + 16, src_cb + 7 * uvlinesize);
1615 AV_COPY64(top_border + 24, src_cr + 7 * uvlinesize);
1620 } else if (MB_MBAFF) {
1626 top_border = h->top_borders[top_idx][s->mb_x];
1627 /* There are two lines saved, the line above the top macroblock
1628 * of a pair, and the line above the bottom macroblock. */
1629 AV_COPY128(top_border, src_y + 16 * linesize);
1631 AV_COPY128(top_border + 16, src_y + 16 * linesize + 16);
1633 if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1636 AV_COPY128(top_border + 32, src_cb + 16 * linesize);
1637 AV_COPY128(top_border + 48, src_cb + 16 * linesize + 16);
1638 AV_COPY128(top_border + 64, src_cr + 16 * linesize);
1639 AV_COPY128(top_border + 80, src_cr + 16 * linesize + 16);
1641 AV_COPY128(top_border + 16, src_cb + 16 * linesize);
1642 AV_COPY128(top_border + 32, src_cr + 16 * linesize);
1644 } else if (chroma422) {
1646 AV_COPY128(top_border + 32, src_cb + 16 * uvlinesize);
1647 AV_COPY128(top_border + 48, src_cr + 16 * uvlinesize);
1649 AV_COPY64(top_border + 16, src_cb + 16 * uvlinesize);
1650 AV_COPY64(top_border + 24, src_cr + 16 * uvlinesize);
1654 AV_COPY128(top_border + 32, src_cb + 8 * uvlinesize);
1655 AV_COPY128(top_border + 48, src_cr + 8 * uvlinesize);
1657 AV_COPY64(top_border + 16, src_cb + 8 * uvlinesize);
1658 AV_COPY64(top_border + 24, src_cr + 8 * uvlinesize);
1664 static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1665 uint8_t *src_cb, uint8_t *src_cr,
1666 int linesize, int uvlinesize,
1667 int xchg, int chroma444,
1668 int simple, int pixel_shift)
1670 MpegEncContext *const s = &h->s;
1671 int deblock_topleft;
1674 uint8_t *top_border_m1;
1675 uint8_t *top_border;
1677 if (!simple && FRAME_MBAFF) {
1682 top_idx = MB_MBAFF ? 0 : 1;
1686 if (h->deblocking_filter == 2) {
1687 deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1688 deblock_top = h->top_type;
1690 deblock_topleft = (s->mb_x > 0);
1691 deblock_top = (s->mb_y > !!MB_FIELD);
1694 src_y -= linesize + 1 + pixel_shift;
1695 src_cb -= uvlinesize + 1 + pixel_shift;
1696 src_cr -= uvlinesize + 1 + pixel_shift;
1698 top_border_m1 = h->top_borders[top_idx][s->mb_x - 1];
1699 top_border = h->top_borders[top_idx][s->mb_x];
1701 #define XCHG(a, b, xchg) \
1702 if (pixel_shift) { \
1704 AV_SWAP64(b + 0, a + 0); \
1705 AV_SWAP64(b + 8, a + 8); \
1715 if (deblock_topleft) {
1716 XCHG(top_border_m1 + (8 << pixel_shift),
1717 src_y - (7 << pixel_shift), 1);
1719 XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1720 XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1721 if (s->mb_x + 1 < s->mb_width) {
1722 XCHG(h->top_borders[top_idx][s->mb_x + 1],
1723 src_y + (17 << pixel_shift), 1);
1726 if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1728 if (deblock_topleft) {
1729 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1730 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1732 XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1733 XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1734 XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1735 XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1736 if (s->mb_x + 1 < s->mb_width) {
1737 XCHG(h->top_borders[top_idx][s->mb_x + 1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1738 XCHG(h->top_borders[top_idx][s->mb_x + 1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1742 if (deblock_topleft) {
1743 XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1744 XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1746 XCHG(top_border + (16 << pixel_shift), src_cb + 1 + pixel_shift, 1);
1747 XCHG(top_border + (24 << pixel_shift), src_cr + 1 + pixel_shift, 1);
1753 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth,
1756 if (high_bit_depth) {
1757 return AV_RN32A(((int32_t *)mb) + index);
1759 return AV_RN16A(mb + index);
1762 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth,
1763 int index, int value)
1765 if (high_bit_depth) {
1766 AV_WN32A(((int32_t *)mb) + index, value);
1768 AV_WN16A(mb + index, value);
1771 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h,
1772 int mb_type, int is_h264,
1774 int transform_bypass,
1778 uint8_t *dest_y, int p)
1780 MpegEncContext *const s = &h->s;
1781 void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1782 void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1784 int qscale = p == 0 ? s->qscale : h->chroma_qp[p - 1];
1785 block_offset += 16 * p;
1786 if (IS_INTRA4x4(mb_type)) {
1787 if (simple || !s->encoding) {
1788 if (IS_8x8DCT(mb_type)) {
1789 if (transform_bypass) {
1791 idct_add = s->dsp.add_pixels8;
1793 idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1794 idct_add = h->h264dsp.h264_idct8_add;
1796 for (i = 0; i < 16; i += 4) {
1797 uint8_t *const ptr = dest_y + block_offset[i];
1798 const int dir = h->intra4x4_pred_mode_cache[scan8[i]];
1799 if (transform_bypass && h->sps.profile_idc == 244 && dir <= 1) {
1800 h->hpc.pred8x8l_add[dir](ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1802 const int nnz = h->non_zero_count_cache[scan8[i + p * 16]];
1803 h->hpc.pred8x8l[dir](ptr, (h->topleft_samples_available << i) & 0x8000,
1804 (h->topright_samples_available << i) & 0x4000, linesize);
1806 if (nnz == 1 && dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
1807 idct_dc_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1809 idct_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1814 if (transform_bypass) {
1816 idct_add = s->dsp.add_pixels4;
1818 idct_dc_add = h->h264dsp.h264_idct_dc_add;
1819 idct_add = h->h264dsp.h264_idct_add;
1821 for (i = 0; i < 16; i++) {
1822 uint8_t *const ptr = dest_y + block_offset[i];
1823 const int dir = h->intra4x4_pred_mode_cache[scan8[i]];
1825 if (transform_bypass && h->sps.profile_idc == 244 && dir <= 1) {
1826 h->hpc.pred4x4_add[dir](ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1831 if (dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED) {
1832 const int topright_avail = (h->topright_samples_available << i) & 0x8000;
1833 assert(s->mb_y || linesize <= block_offset[i]);
1834 if (!topright_avail) {
1836 tr_high = ((uint16_t *)ptr)[3 - linesize / 2] * 0x0001000100010001ULL;
1837 topright = (uint8_t *)&tr_high;
1839 tr = ptr[3 - linesize] * 0x01010101u;
1840 topright = (uint8_t *)&tr;
1843 topright = ptr + (4 << pixel_shift) - linesize;
1847 h->hpc.pred4x4[dir](ptr, topright, linesize);
1848 nnz = h->non_zero_count_cache[scan8[i + p * 16]];
1851 if (nnz == 1 && dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
1852 idct_dc_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1854 idct_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1855 } else if (CONFIG_SVQ3_DECODER)
1856 ff_svq3_add_idct_c(ptr, h->mb + i * 16 + p * 256, linesize, qscale, 0);
1863 h->hpc.pred16x16[h->intra16x16_pred_mode](dest_y, linesize);
1865 if (h->non_zero_count_cache[scan8[LUMA_DC_BLOCK_INDEX + p]]) {
1866 if (!transform_bypass)
1867 h->h264dsp.h264_luma_dc_dequant_idct(h->mb + (p * 256 << pixel_shift),
1869 h->dequant4_coeff[p][qscale][0]);
1871 static const uint8_t dc_mapping[16] = {
1872 0 * 16, 1 * 16, 4 * 16, 5 * 16,
1873 2 * 16, 3 * 16, 6 * 16, 7 * 16,
1874 8 * 16, 9 * 16, 12 * 16, 13 * 16,
1875 10 * 16, 11 * 16, 14 * 16, 15 * 16 };
1876 for (i = 0; i < 16; i++)
1877 dctcoef_set(h->mb + (p * 256 << pixel_shift),
1878 pixel_shift, dc_mapping[i],
1879 dctcoef_get(h->mb_luma_dc[p],
1883 } else if (CONFIG_SVQ3_DECODER)
1884 ff_svq3_luma_dc_dequant_idct_c(h->mb + p * 256,
1885 h->mb_luma_dc[p], qscale);
1889 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type,
1890 int is_h264, int simple,
1891 int transform_bypass,
1895 uint8_t *dest_y, int p)
1897 MpegEncContext *const s = &h->s;
1898 void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1900 block_offset += 16 * p;
1901 if (!IS_INTRA4x4(mb_type)) {
1903 if (IS_INTRA16x16(mb_type)) {
1904 if (transform_bypass) {
1905 if (h->sps.profile_idc == 244 &&
1906 (h->intra16x16_pred_mode == VERT_PRED8x8 ||
1907 h->intra16x16_pred_mode == HOR_PRED8x8)) {
1908 h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset,
1909 h->mb + (p * 256 << pixel_shift),
1912 for (i = 0; i < 16; i++)
1913 if (h->non_zero_count_cache[scan8[i + p * 16]] ||
1914 dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
1915 s->dsp.add_pixels4(dest_y + block_offset[i],
1916 h->mb + (i * 16 + p * 256 << pixel_shift),
1920 h->h264dsp.h264_idct_add16intra(dest_y, block_offset,
1921 h->mb + (p * 256 << pixel_shift),
1923 h->non_zero_count_cache + p * 5 * 8);
1925 } else if (h->cbp & 15) {
1926 if (transform_bypass) {
1927 const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1928 idct_add = IS_8x8DCT(mb_type) ? s->dsp.add_pixels8
1929 : s->dsp.add_pixels4;
1930 for (i = 0; i < 16; i += di)
1931 if (h->non_zero_count_cache[scan8[i + p * 16]])
1932 idct_add(dest_y + block_offset[i],
1933 h->mb + (i * 16 + p * 256 << pixel_shift),
1936 if (IS_8x8DCT(mb_type))
1937 h->h264dsp.h264_idct8_add4(dest_y, block_offset,
1938 h->mb + (p * 256 << pixel_shift),
1940 h->non_zero_count_cache + p * 5 * 8);
1942 h->h264dsp.h264_idct_add16(dest_y, block_offset,
1943 h->mb + (p * 256 << pixel_shift),
1945 h->non_zero_count_cache + p * 5 * 8);
1948 } else if (CONFIG_SVQ3_DECODER) {
1949 for (i = 0; i < 16; i++)
1950 if (h->non_zero_count_cache[scan8[i + p * 16]] || h->mb[i * 16 + p * 256]) {
1951 // FIXME benchmark weird rule, & below
1952 uint8_t *const ptr = dest_y + block_offset[i];
1953 ff_svq3_add_idct_c(ptr, h->mb + i * 16 + p * 256, linesize,
1954 s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1962 #include "h264_mb_template.c"
1966 #include "h264_mb_template.c"
1970 #include "h264_mb_template.c"
1972 void ff_h264_hl_decode_mb(H264Context *h)
1974 MpegEncContext *const s = &h->s;
1975 const int mb_xy = h->mb_xy;
1976 const int mb_type = s->current_picture.f.mb_type[mb_xy];
1977 int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
1980 if (is_complex || h->pixel_shift)
1981 hl_decode_mb_444_complex(h);
1983 hl_decode_mb_444_simple_8(h);
1984 } else if (is_complex) {
1985 hl_decode_mb_complex(h);
1986 } else if (h->pixel_shift) {
1987 hl_decode_mb_simple_16(h);
1989 hl_decode_mb_simple_8(h);
1992 static int pred_weight_table(H264Context *h)
1994 MpegEncContext *const s = &h->s;
1996 int luma_def, chroma_def;
1999 h->use_weight_chroma = 0;
2000 h->luma_log2_weight_denom = get_ue_golomb(&s->gb);
2001 if (h->sps.chroma_format_idc)
2002 h->chroma_log2_weight_denom = get_ue_golomb(&s->gb);
2003 luma_def = 1 << h->luma_log2_weight_denom;
2004 chroma_def = 1 << h->chroma_log2_weight_denom;
2006 for (list = 0; list < 2; list++) {
2007 h->luma_weight_flag[list] = 0;
2008 h->chroma_weight_flag[list] = 0;
2009 for (i = 0; i < h->ref_count[list]; i++) {
2010 int luma_weight_flag, chroma_weight_flag;
2012 luma_weight_flag = get_bits1(&s->gb);
2013 if (luma_weight_flag) {
2014 h->luma_weight[i][list][0] = get_se_golomb(&s->gb);
2015 h->luma_weight[i][list][1] = get_se_golomb(&s->gb);
2016 if (h->luma_weight[i][list][0] != luma_def ||
2017 h->luma_weight[i][list][1] != 0) {
2019 h->luma_weight_flag[list] = 1;
2022 h->luma_weight[i][list][0] = luma_def;
2023 h->luma_weight[i][list][1] = 0;
2026 if (h->sps.chroma_format_idc) {
2027 chroma_weight_flag = get_bits1(&s->gb);
2028 if (chroma_weight_flag) {
2030 for (j = 0; j < 2; j++) {
2031 h->chroma_weight[i][list][j][0] = get_se_golomb(&s->gb);
2032 h->chroma_weight[i][list][j][1] = get_se_golomb(&s->gb);
2033 if (h->chroma_weight[i][list][j][0] != chroma_def ||
2034 h->chroma_weight[i][list][j][1] != 0) {
2035 h->use_weight_chroma = 1;
2036 h->chroma_weight_flag[list] = 1;
2041 for (j = 0; j < 2; j++) {
2042 h->chroma_weight[i][list][j][0] = chroma_def;
2043 h->chroma_weight[i][list][j][1] = 0;
2048 if (h->slice_type_nos != AV_PICTURE_TYPE_B)
2051 h->use_weight = h->use_weight || h->use_weight_chroma;
2056 * Initialize implicit_weight table.
2057 * @param field 0/1 initialize the weight for interlaced MBAFF
2058 * -1 initializes the rest
2060 static void implicit_weight_table(H264Context *h, int field)
2062 MpegEncContext *const s = &h->s;
2063 int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2065 for (i = 0; i < 2; i++) {
2066 h->luma_weight_flag[i] = 0;
2067 h->chroma_weight_flag[i] = 0;
2071 if (s->picture_structure == PICT_FRAME) {
2072 cur_poc = s->current_picture_ptr->poc;
2074 cur_poc = s->current_picture_ptr->field_poc[s->picture_structure - 1];
2076 if (h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF &&
2077 h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2 * cur_poc) {
2079 h->use_weight_chroma = 0;
2083 ref_count0 = h->ref_count[0];
2084 ref_count1 = h->ref_count[1];
2086 cur_poc = s->current_picture_ptr->field_poc[field];
2088 ref_count0 = 16 + 2 * h->ref_count[0];
2089 ref_count1 = 16 + 2 * h->ref_count[1];
2093 h->use_weight_chroma = 2;
2094 h->luma_log2_weight_denom = 5;
2095 h->chroma_log2_weight_denom = 5;
2097 for (ref0 = ref_start; ref0 < ref_count0; ref0++) {
2098 int poc0 = h->ref_list[0][ref0].poc;
2099 for (ref1 = ref_start; ref1 < ref_count1; ref1++) {
2101 if (!h->ref_list[0][ref0].long_ref && !h->ref_list[1][ref1].long_ref) {
2102 int poc1 = h->ref_list[1][ref1].poc;
2103 int td = av_clip(poc1 - poc0, -128, 127);
2105 int tb = av_clip(cur_poc - poc0, -128, 127);
2106 int tx = (16384 + (FFABS(td) >> 1)) / td;
2107 int dist_scale_factor = (tb * tx + 32) >> 8;
2108 if (dist_scale_factor >= -64 && dist_scale_factor <= 128)
2109 w = 64 - dist_scale_factor;
2113 h->implicit_weight[ref0][ref1][0] =
2114 h->implicit_weight[ref0][ref1][1] = w;
2116 h->implicit_weight[ref0][ref1][field] = w;
2123 * instantaneous decoder refresh.
2125 static void idr(H264Context *h)
2127 ff_h264_remove_all_refs(h);
2128 h->prev_frame_num = 0;
2129 h->prev_frame_num_offset = 0;
2131 h->prev_poc_lsb = 0;
2134 /* forget old pics after a seek */
2135 static void flush_change(H264Context *h)
2138 for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
2139 h->last_pocs[i] = INT_MIN;
2140 h->outputed_poc = h->next_outputed_poc = INT_MIN;
2141 h->prev_interlaced_frame = 1;
2143 if (h->s.current_picture_ptr)
2144 h->s.current_picture_ptr->f.reference = 0;
2145 h->s.first_field = 0;
2146 memset(h->ref_list[0], 0, sizeof(h->ref_list[0]));
2147 memset(h->ref_list[1], 0, sizeof(h->ref_list[1]));
2148 memset(h->default_ref_list[0], 0, sizeof(h->default_ref_list[0]));
2149 memset(h->default_ref_list[1], 0, sizeof(h->default_ref_list[1]));
2150 ff_h264_reset_sei(h);
2153 /* forget old pics after a seek */
2154 static void flush_dpb(AVCodecContext *avctx)
2156 H264Context *h = avctx->priv_data;
2159 for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++) {
2160 if (h->delayed_pic[i])
2161 h->delayed_pic[i]->f.reference = 0;
2162 h->delayed_pic[i] = NULL;
2166 ff_mpeg_flush(avctx);
2169 static int init_poc(H264Context *h)
2171 MpegEncContext *const s = &h->s;
2172 const int max_frame_num = 1 << h->sps.log2_max_frame_num;
2174 Picture *cur = s->current_picture_ptr;
2176 h->frame_num_offset = h->prev_frame_num_offset;
2177 if (h->frame_num < h->prev_frame_num)
2178 h->frame_num_offset += max_frame_num;
2180 if (h->sps.poc_type == 0) {
2181 const int max_poc_lsb = 1 << h->sps.log2_max_poc_lsb;
2183 if (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb / 2)
2184 h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2185 else if (h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb / 2)
2186 h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2188 h->poc_msb = h->prev_poc_msb;
2190 field_poc[1] = h->poc_msb + h->poc_lsb;
2191 if (s->picture_structure == PICT_FRAME)
2192 field_poc[1] += h->delta_poc_bottom;
2193 } else if (h->sps.poc_type == 1) {
2194 int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2197 if (h->sps.poc_cycle_length != 0)
2198 abs_frame_num = h->frame_num_offset + h->frame_num;
2202 if (h->nal_ref_idc == 0 && abs_frame_num > 0)
2205 expected_delta_per_poc_cycle = 0;
2206 for (i = 0; i < h->sps.poc_cycle_length; i++)
2207 // FIXME integrate during sps parse
2208 expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[i];
2210 if (abs_frame_num > 0) {
2211 int poc_cycle_cnt = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2212 int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2214 expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2215 for (i = 0; i <= frame_num_in_poc_cycle; i++)
2216 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[i];
2220 if (h->nal_ref_idc == 0)
2221 expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2223 field_poc[0] = expectedpoc + h->delta_poc[0];
2224 field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2226 if (s->picture_structure == PICT_FRAME)
2227 field_poc[1] += h->delta_poc[1];
2229 int poc = 2 * (h->frame_num_offset + h->frame_num);
2231 if (!h->nal_ref_idc)
2238 if (s->picture_structure != PICT_BOTTOM_FIELD)
2239 s->current_picture_ptr->field_poc[0] = field_poc[0];
2240 if (s->picture_structure != PICT_TOP_FIELD)
2241 s->current_picture_ptr->field_poc[1] = field_poc[1];
2242 cur->poc = FFMIN(cur->field_poc[0], cur->field_poc[1]);
2248 * initialize scan tables
2250 static void init_scan_tables(H264Context *h)
2253 for (i = 0; i < 16; i++) {
2254 #define T(x) (x >> 2) | ((x << 2) & 0xF)
2255 h->zigzag_scan[i] = T(zigzag_scan[i]);
2256 h->field_scan[i] = T(field_scan[i]);
2259 for (i = 0; i < 64; i++) {
2260 #define T(x) (x >> 3) | ((x & 7) << 3)
2261 h->zigzag_scan8x8[i] = T(ff_zigzag_direct[i]);
2262 h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2263 h->field_scan8x8[i] = T(field_scan8x8[i]);
2264 h->field_scan8x8_cavlc[i] = T(field_scan8x8_cavlc[i]);
2267 if (h->sps.transform_bypass) { // FIXME same ugly
2268 h->zigzag_scan_q0 = zigzag_scan;
2269 h->zigzag_scan8x8_q0 = ff_zigzag_direct;
2270 h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2271 h->field_scan_q0 = field_scan;
2272 h->field_scan8x8_q0 = field_scan8x8;
2273 h->field_scan8x8_cavlc_q0 = field_scan8x8_cavlc;
2275 h->zigzag_scan_q0 = h->zigzag_scan;
2276 h->zigzag_scan8x8_q0 = h->zigzag_scan8x8;
2277 h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2278 h->field_scan_q0 = h->field_scan;
2279 h->field_scan8x8_q0 = h->field_scan8x8;
2280 h->field_scan8x8_cavlc_q0 = h->field_scan8x8_cavlc;
2284 static int field_end(H264Context *h, int in_setup)
2286 MpegEncContext *const s = &h->s;
2287 AVCodecContext *const avctx = s->avctx;
2291 if (!in_setup && !s->droppable)
2292 ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX,
2293 s->picture_structure == PICT_BOTTOM_FIELD);
2295 if (CONFIG_H264_VDPAU_DECODER &&
2296 s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
2297 ff_vdpau_h264_set_reference_frames(s);
2299 if (in_setup || !(avctx->active_thread_type & FF_THREAD_FRAME)) {
2300 if (!s->droppable) {
2301 err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2302 h->prev_poc_msb = h->poc_msb;
2303 h->prev_poc_lsb = h->poc_lsb;
2305 h->prev_frame_num_offset = h->frame_num_offset;
2306 h->prev_frame_num = h->frame_num;
2307 h->outputed_poc = h->next_outputed_poc;
2310 if (avctx->hwaccel) {
2311 if (avctx->hwaccel->end_frame(avctx) < 0)
2312 av_log(avctx, AV_LOG_ERROR,
2313 "hardware accelerator failed to decode picture\n");
2316 if (CONFIG_H264_VDPAU_DECODER &&
2317 s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
2318 ff_vdpau_h264_picture_complete(s);
2321 * FIXME: Error handling code does not seem to support interlaced
2322 * when slices span multiple rows
2323 * The ff_er_add_slice calls don't work right for bottom
2324 * fields; they cause massive erroneous error concealing
2325 * Error marking covers both fields (top and bottom).
2326 * This causes a mismatched s->error_count
2327 * and a bad error table. Further, the error count goes to
2328 * INT_MAX when called for bottom field, because mb_y is
2329 * past end by one (callers fault) and resync_mb_y != 0
2330 * causes problems for the first MB line, too.
2335 ff_MPV_frame_end(s);
2337 h->current_slice = 0;
2343 * Replicate H264 "master" context to thread contexts.
2345 static int clone_slice(H264Context *dst, H264Context *src)
2349 memcpy(dst->block_offset, src->block_offset, sizeof(dst->block_offset));
2350 dst->s.current_picture_ptr = src->s.current_picture_ptr;
2351 dst->s.current_picture = src->s.current_picture;
2352 dst->s.linesize = src->s.linesize;
2353 dst->s.uvlinesize = src->s.uvlinesize;
2354 dst->s.first_field = src->s.first_field;
2356 if (!dst->s.edge_emu_buffer &&
2357 (ret = ff_mpv_frame_size_alloc(&dst->s, dst->s.linesize))) {
2358 av_log(dst->s.avctx, AV_LOG_ERROR,
2359 "Failed to allocate scratch buffers\n");
2363 dst->prev_poc_msb = src->prev_poc_msb;
2364 dst->prev_poc_lsb = src->prev_poc_lsb;
2365 dst->prev_frame_num_offset = src->prev_frame_num_offset;
2366 dst->prev_frame_num = src->prev_frame_num;
2367 dst->short_ref_count = src->short_ref_count;
2369 memcpy(dst->short_ref, src->short_ref, sizeof(dst->short_ref));
2370 memcpy(dst->long_ref, src->long_ref, sizeof(dst->long_ref));
2371 memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_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 s->me.qpel_put = s->dsp.put_h264_qpel_pixels_tab;
2597 s->me.qpel_avg = s->dsp.avg_h264_qpel_pixels_tab;
2599 first_mb_in_slice = get_ue_golomb(&s->gb);
2601 if (first_mb_in_slice == 0) { // FIXME better field boundary detection
2602 if (h0->current_slice && FIELD_PICTURE) {
2606 h0->current_slice = 0;
2607 if (!s0->first_field) {
2608 if (s->current_picture_ptr && !s->droppable &&
2609 s->current_picture_ptr->owner2 == s) {
2610 ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX,
2611 s->picture_structure == PICT_BOTTOM_FIELD);
2613 s->current_picture_ptr = NULL;
2617 slice_type = get_ue_golomb_31(&s->gb);
2618 if (slice_type > 9) {
2619 av_log(h->s.avctx, AV_LOG_ERROR,
2620 "slice type too large (%d) at %d %d\n",
2621 h->slice_type, s->mb_x, s->mb_y);
2624 if (slice_type > 4) {
2626 h->slice_type_fixed = 1;
2628 h->slice_type_fixed = 0;
2630 slice_type = golomb_to_pict_type[slice_type];
2631 if (slice_type == AV_PICTURE_TYPE_I ||
2632 (h0->current_slice != 0 && slice_type == h0->last_slice_type)) {
2633 default_ref_list_done = 1;
2635 h->slice_type = slice_type;
2636 h->slice_type_nos = slice_type & 3;
2638 // to make a few old functions happy, it's wrong though
2639 s->pict_type = h->slice_type;
2641 pps_id = get_ue_golomb(&s->gb);
2642 if (pps_id >= MAX_PPS_COUNT) {
2643 av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
2646 if (!h0->pps_buffers[pps_id]) {
2647 av_log(h->s.avctx, AV_LOG_ERROR,
2648 "non-existing PPS %u referenced\n",
2652 h->pps = *h0->pps_buffers[pps_id];
2654 if (!h0->sps_buffers[h->pps.sps_id]) {
2655 av_log(h->s.avctx, AV_LOG_ERROR,
2656 "non-existing SPS %u referenced\n",
2661 if (h->pps.sps_id != h->current_sps_id ||
2662 h->context_reinitialized ||
2663 h0->sps_buffers[h->pps.sps_id]->new) {
2664 SPS *new_sps = h0->sps_buffers[h->pps.sps_id];
2666 h0->sps_buffers[h->pps.sps_id]->new = 0;
2668 if (h->sps.chroma_format_idc != new_sps->chroma_format_idc ||
2669 h->sps.bit_depth_luma != new_sps->bit_depth_luma)
2672 h->current_sps_id = h->pps.sps_id;
2673 h->sps = *h0->sps_buffers[h->pps.sps_id];
2675 if ((ret = h264_set_parameter_from_sps(h)) < 0)
2679 s->avctx->profile = ff_h264_get_profile(&h->sps);
2680 s->avctx->level = h->sps.level_idc;
2681 s->avctx->refs = h->sps.ref_frame_count;
2683 if (s->mb_width != h->sps.mb_width ||
2684 s->mb_height != h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag))
2687 s->mb_width = h->sps.mb_width;
2688 s->mb_height = h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2690 h->b_stride = s->mb_width * 4;
2692 s->chroma_y_shift = h->sps.chroma_format_idc <= 1; // 400 uses yuv420p
2694 s->width = 16 * s->mb_width - (2 >> CHROMA444) * FFMIN(h->sps.crop_right, (8 << CHROMA444) - 1);
2695 if (h->sps.frame_mbs_only_flag)
2696 s->height = 16 * s->mb_height - (1 << s->chroma_y_shift) * FFMIN(h->sps.crop_bottom, (16 >> s->chroma_y_shift) - 1);
2698 s->height = 16 * s->mb_height - (2 << s->chroma_y_shift) * FFMIN(h->sps.crop_bottom, (16 >> s->chroma_y_shift) - 1);
2700 if (FFALIGN(s->avctx->width, 16) == s->width &&
2701 FFALIGN(s->avctx->height, 16) == s->height) {
2702 s->width = s->avctx->width;
2703 s->height = s->avctx->height;
2706 if (h->sps.video_signal_type_present_flag) {
2707 s->avctx->color_range = h->sps.full_range ? AVCOL_RANGE_JPEG
2709 if (h->sps.colour_description_present_flag) {
2710 if (s->avctx->colorspace != h->sps.colorspace)
2712 s->avctx->color_primaries = h->sps.color_primaries;
2713 s->avctx->color_trc = h->sps.color_trc;
2714 s->avctx->colorspace = h->sps.colorspace;
2718 if (s->context_initialized &&
2719 (s->width != s->avctx->width ||
2720 s->height != s->avctx->height ||
2722 av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio))) {
2725 av_log(s->avctx, AV_LOG_ERROR, "changing width/height on "
2726 "slice %d\n", h0->current_slice + 1);
2727 return AVERROR_INVALIDDATA;
2732 if ((ret = get_pixel_format(h)) < 0)
2734 s->avctx->pix_fmt = ret;
2736 av_log(h->s.avctx, AV_LOG_INFO, "Reinit context to %dx%d, "
2737 "pix_fmt: %d\n", s->width, s->height, s->avctx->pix_fmt);
2739 if ((ret = h264_slice_header_init(h, 1)) < 0) {
2740 av_log(h->s.avctx, AV_LOG_ERROR,
2741 "h264_slice_header_init() failed\n");
2744 h->context_reinitialized = 1;
2746 if (!s->context_initialized) {
2748 av_log(h->s.avctx, AV_LOG_ERROR,
2749 "Cannot (re-)initialize context during parallel decoding.\n");
2753 if ((ret = get_pixel_format(h)) < 0)
2755 s->avctx->pix_fmt = ret;
2757 if ((ret = h264_slice_header_init(h, 0)) < 0) {
2758 av_log(h->s.avctx, AV_LOG_ERROR,
2759 "h264_slice_header_init() failed\n");
2764 if (h == h0 && h->dequant_coeff_pps != pps_id) {
2765 h->dequant_coeff_pps = pps_id;
2766 init_dequant_tables(h);
2769 h->frame_num = get_bits(&s->gb, h->sps.log2_max_frame_num);
2772 h->mb_aff_frame = 0;
2773 last_pic_structure = s0->picture_structure;
2774 last_pic_droppable = s0->droppable;
2775 s->droppable = h->nal_ref_idc == 0;
2776 if (h->sps.frame_mbs_only_flag) {
2777 s->picture_structure = PICT_FRAME;
2779 if (get_bits1(&s->gb)) { // field_pic_flag
2780 s->picture_structure = PICT_TOP_FIELD + get_bits1(&s->gb); // bottom_field_flag
2782 s->picture_structure = PICT_FRAME;
2783 h->mb_aff_frame = h->sps.mb_aff;
2786 h->mb_field_decoding_flag = s->picture_structure != PICT_FRAME;
2788 if (h0->current_slice != 0) {
2789 if (last_pic_structure != s->picture_structure ||
2790 last_pic_droppable != s->droppable) {
2791 av_log(h->s.avctx, AV_LOG_ERROR,
2792 "Changing field mode (%d -> %d) between slices is not allowed\n",
2793 last_pic_structure, s->picture_structure);
2794 s->picture_structure = last_pic_structure;
2795 s->droppable = last_pic_droppable;
2796 return AVERROR_INVALIDDATA;
2797 } else if (!s0->current_picture_ptr) {
2798 av_log(s->avctx, AV_LOG_ERROR,
2799 "unset current_picture_ptr on %d. slice\n",
2800 h0->current_slice + 1);
2801 return AVERROR_INVALIDDATA;
2804 /* Shorten frame num gaps so we don't have to allocate reference
2805 * frames just to throw them away */
2806 if (h->frame_num != h->prev_frame_num) {
2807 int unwrap_prev_frame_num = h->prev_frame_num;
2808 int max_frame_num = 1 << h->sps.log2_max_frame_num;
2810 if (unwrap_prev_frame_num > h->frame_num)
2811 unwrap_prev_frame_num -= max_frame_num;
2813 if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
2814 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
2815 if (unwrap_prev_frame_num < 0)
2816 unwrap_prev_frame_num += max_frame_num;
2818 h->prev_frame_num = unwrap_prev_frame_num;
2822 /* See if we have a decoded first field looking for a pair...
2823 * Here, we're using that to see if we should mark previously
2824 * decode frames as "finished".
2825 * We have to do that before the "dummy" in-between frame allocation,
2826 * since that can modify s->current_picture_ptr. */
2827 if (s0->first_field) {
2828 assert(s0->current_picture_ptr);
2829 assert(s0->current_picture_ptr->f.data[0]);
2830 assert(s0->current_picture_ptr->f.reference != DELAYED_PIC_REF);
2832 /* Mark old field/frame as completed */
2833 if (!last_pic_droppable && s0->current_picture_ptr->owner2 == s0) {
2834 ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
2835 last_pic_structure == PICT_BOTTOM_FIELD);
2838 /* figure out if we have a complementary field pair */
2839 if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2840 /* Previous field is unmatched. Don't display it, but let it
2841 * remain for reference if marked as such. */
2842 if (!last_pic_droppable && last_pic_structure != PICT_FRAME) {
2843 ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
2844 last_pic_structure == PICT_TOP_FIELD);
2847 if (s0->current_picture_ptr->frame_num != h->frame_num) {
2848 /* This and previous field were reference, but had
2849 * different frame_nums. Consider this field first in
2850 * pair. Throw away previous field except for reference
2852 if (!last_pic_droppable && last_pic_structure != PICT_FRAME) {
2853 ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
2854 last_pic_structure == PICT_TOP_FIELD);
2857 /* Second field in complementary pair */
2858 if (!((last_pic_structure == PICT_TOP_FIELD &&
2859 s->picture_structure == PICT_BOTTOM_FIELD) ||
2860 (last_pic_structure == PICT_BOTTOM_FIELD &&
2861 s->picture_structure == PICT_TOP_FIELD))) {
2862 av_log(s->avctx, AV_LOG_ERROR,
2863 "Invalid field mode combination %d/%d\n",
2864 last_pic_structure, s->picture_structure);
2865 s->picture_structure = last_pic_structure;
2866 s->droppable = last_pic_droppable;
2867 return AVERROR_INVALIDDATA;
2868 } else if (last_pic_droppable != s->droppable) {
2869 av_log(s->avctx, AV_LOG_ERROR,
2870 "Cannot combine reference and non-reference fields in the same frame\n");
2871 av_log_ask_for_sample(s->avctx, NULL);
2872 s->picture_structure = last_pic_structure;
2873 s->droppable = last_pic_droppable;
2874 return AVERROR_PATCHWELCOME;
2877 /* Take ownership of this buffer. Note that if another thread owned
2878 * the first field of this buffer, we're not operating on that pointer,
2879 * so the original thread is still responsible for reporting progress
2880 * on that first field (or if that was us, we just did that above).
2881 * By taking ownership, we assign responsibility to ourselves to
2882 * report progress on the second field. */
2883 s0->current_picture_ptr->owner2 = s0;
2888 while (h->frame_num != h->prev_frame_num &&
2889 h->frame_num != (h->prev_frame_num + 1) % (1 << h->sps.log2_max_frame_num)) {
2890 Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
2891 av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n",
2892 h->frame_num, h->prev_frame_num);
2893 if (ff_h264_frame_start(h) < 0)
2895 h->prev_frame_num++;
2896 h->prev_frame_num %= 1 << h->sps.log2_max_frame_num;
2897 s->current_picture_ptr->frame_num = h->prev_frame_num;
2898 ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX, 0);
2899 ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX, 1);
2900 if ((ret = ff_generate_sliding_window_mmcos(h, 1)) < 0 &&
2901 s->avctx->err_recognition & AV_EF_EXPLODE)
2903 if (ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index) < 0 &&
2904 (s->avctx->err_recognition & AV_EF_EXPLODE))
2905 return AVERROR_INVALIDDATA;
2906 /* Error concealment: if a ref is missing, copy the previous ref in its place.
2907 * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
2908 * about there being no actual duplicates.
2909 * FIXME: this doesn't copy padding for out-of-frame motion vectors. Given we're
2910 * concealing a lost frame, this probably isn't noticeable by comparison, but it should
2912 if (h->short_ref_count) {
2914 av_image_copy(h->short_ref[0]->f.data, h->short_ref[0]->f.linesize,
2915 (const uint8_t **)prev->f.data, prev->f.linesize,
2916 s->avctx->pix_fmt, s->mb_width * 16, s->mb_height * 16);
2917 h->short_ref[0]->poc = prev->poc + 2;
2919 h->short_ref[0]->frame_num = h->prev_frame_num;
2923 /* See if we have a decoded first field looking for a pair...
2924 * We're using that to see whether to continue decoding in that
2925 * frame, or to allocate a new one. */
2926 if (s0->first_field) {
2927 assert(s0->current_picture_ptr);
2928 assert(s0->current_picture_ptr->f.data[0]);
2929 assert(s0->current_picture_ptr->f.reference != DELAYED_PIC_REF);
2931 /* figure out if we have a complementary field pair */
2932 if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2933 /* Previous field is unmatched. Don't display it, but let it
2934 * remain for reference if marked as such. */
2935 s0->current_picture_ptr = NULL;
2936 s0->first_field = FIELD_PICTURE;
2938 if (s0->current_picture_ptr->frame_num != h->frame_num) {
2939 /* This and the previous field had different frame_nums.
2940 * Consider this field first in pair. Throw away previous
2941 * one except for reference purposes. */
2942 s0->first_field = 1;
2943 s0->current_picture_ptr = NULL;
2945 /* Second field in complementary pair */
2946 s0->first_field = 0;
2950 /* Frame or first field in a potentially complementary pair */
2951 s0->first_field = FIELD_PICTURE;
2954 if (!FIELD_PICTURE || s0->first_field) {
2955 if (ff_h264_frame_start(h) < 0) {
2956 s0->first_field = 0;
2960 ff_release_unused_pictures(s, 0);
2963 if (h != h0 && (ret = clone_slice(h, h0)) < 0)
2966 s->current_picture_ptr->frame_num = h->frame_num; // FIXME frame_num cleanup
2968 assert(s->mb_num == s->mb_width * s->mb_height);
2969 if (first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
2970 first_mb_in_slice >= s->mb_num) {
2971 av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
2974 s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
2975 s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
2976 if (s->picture_structure == PICT_BOTTOM_FIELD)
2977 s->resync_mb_y = s->mb_y = s->mb_y + 1;
2978 assert(s->mb_y < s->mb_height);
2980 if (s->picture_structure == PICT_FRAME) {
2981 h->curr_pic_num = h->frame_num;
2982 h->max_pic_num = 1 << h->sps.log2_max_frame_num;
2984 h->curr_pic_num = 2 * h->frame_num + 1;
2985 h->max_pic_num = 1 << (h->sps.log2_max_frame_num + 1);
2988 if (h->nal_unit_type == NAL_IDR_SLICE)
2989 get_ue_golomb(&s->gb); /* idr_pic_id */
2991 if (h->sps.poc_type == 0) {
2992 h->poc_lsb = get_bits(&s->gb, h->sps.log2_max_poc_lsb);
2994 if (h->pps.pic_order_present == 1 && s->picture_structure == PICT_FRAME)
2995 h->delta_poc_bottom = get_se_golomb(&s->gb);
2998 if (h->sps.poc_type == 1 && !h->sps.delta_pic_order_always_zero_flag) {
2999 h->delta_poc[0] = get_se_golomb(&s->gb);
3001 if (h->pps.pic_order_present == 1 && s->picture_structure == PICT_FRAME)
3002 h->delta_poc[1] = get_se_golomb(&s->gb);
3007 if (h->pps.redundant_pic_cnt_present)
3008 h->redundant_pic_count = get_ue_golomb(&s->gb);
3010 // set defaults, might be overridden a few lines later
3011 h->ref_count[0] = h->pps.ref_count[0];
3012 h->ref_count[1] = h->pps.ref_count[1];
3014 if (h->slice_type_nos != AV_PICTURE_TYPE_I) {
3015 if (h->slice_type_nos == AV_PICTURE_TYPE_B)
3016 h->direct_spatial_mv_pred = get_bits1(&s->gb);
3017 num_ref_idx_active_override_flag = get_bits1(&s->gb);
3019 if (num_ref_idx_active_override_flag) {
3020 h->ref_count[0] = get_ue_golomb(&s->gb) + 1;
3021 if (h->ref_count[0] < 1)
3022 return AVERROR_INVALIDDATA;
3023 if (h->slice_type_nos == AV_PICTURE_TYPE_B) {
3024 h->ref_count[1] = get_ue_golomb(&s->gb) + 1;
3025 if (h->ref_count[1] < 1)
3026 return AVERROR_INVALIDDATA;
3030 if (h->slice_type_nos == AV_PICTURE_TYPE_B)
3037 max_refs = s->picture_structure == PICT_FRAME ? 16 : 32;
3039 if (h->ref_count[0] > max_refs || h->ref_count[1] > max_refs) {
3040 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
3041 h->ref_count[0] = h->ref_count[1] = 1;
3042 return AVERROR_INVALIDDATA;
3045 if (!default_ref_list_done)
3046 ff_h264_fill_default_ref_list(h);
3048 if (h->slice_type_nos != AV_PICTURE_TYPE_I &&
3049 ff_h264_decode_ref_pic_list_reordering(h) < 0) {
3050 h->ref_count[1] = h->ref_count[0] = 0;
3054 if (h->slice_type_nos != AV_PICTURE_TYPE_I) {
3055 s->last_picture_ptr = &h->ref_list[0][0];
3056 s->last_picture_ptr->owner2 = s;
3057 ff_copy_picture(&s->last_picture, s->last_picture_ptr);
3059 if (h->slice_type_nos == AV_PICTURE_TYPE_B) {
3060 s->next_picture_ptr = &h->ref_list[1][0];
3061 s->next_picture_ptr->owner2 = s;
3062 ff_copy_picture(&s->next_picture, s->next_picture_ptr);
3065 if ((h->pps.weighted_pred && h->slice_type_nos == AV_PICTURE_TYPE_P) ||
3066 (h->pps.weighted_bipred_idc == 1 &&
3067 h->slice_type_nos == AV_PICTURE_TYPE_B))
3068 pred_weight_table(h);
3069 else if (h->pps.weighted_bipred_idc == 2 &&
3070 h->slice_type_nos == AV_PICTURE_TYPE_B) {
3071 implicit_weight_table(h, -1);
3074 for (i = 0; i < 2; i++) {
3075 h->luma_weight_flag[i] = 0;
3076 h->chroma_weight_flag[i] = 0;
3080 // If frame-mt is enabled, only update mmco tables for the first slice
3081 // in a field. Subsequent slices can temporarily clobber h->mmco_index
3082 // or h->mmco, which will cause ref list mix-ups and decoding errors
3083 // further down the line. This may break decoding if the first slice is
3084 // corrupt, thus we only do this if frame-mt is enabled.
3085 if (h->nal_ref_idc &&
3086 ff_h264_decode_ref_pic_marking(h0, &s->gb,
3087 !(s->avctx->active_thread_type & FF_THREAD_FRAME) ||
3088 h0->current_slice == 0) < 0 &&
3089 (s->avctx->err_recognition & AV_EF_EXPLODE))
3090 return AVERROR_INVALIDDATA;
3093 ff_h264_fill_mbaff_ref_list(h);
3095 if (h->pps.weighted_bipred_idc == 2 && h->slice_type_nos == AV_PICTURE_TYPE_B) {
3096 implicit_weight_table(h, 0);
3097 implicit_weight_table(h, 1);
3101 if (h->slice_type_nos == AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
3102 ff_h264_direct_dist_scale_factor(h);
3103 ff_h264_direct_ref_list_init(h);
3105 if (h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac) {
3106 tmp = get_ue_golomb_31(&s->gb);
3108 av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
3111 h->cabac_init_idc = tmp;
3114 h->last_qscale_diff = 0;
3115 tmp = h->pps.init_qp + get_se_golomb(&s->gb);
3116 if (tmp > 51 + 6 * (h->sps.bit_depth_luma - 8)) {
3117 av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
3121 h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3122 h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3123 // FIXME qscale / qp ... stuff
3124 if (h->slice_type == AV_PICTURE_TYPE_SP)
3125 get_bits1(&s->gb); /* sp_for_switch_flag */
3126 if (h->slice_type == AV_PICTURE_TYPE_SP ||
3127 h->slice_type == AV_PICTURE_TYPE_SI)
3128 get_se_golomb(&s->gb); /* slice_qs_delta */
3130 h->deblocking_filter = 1;
3131 h->slice_alpha_c0_offset = 52;
3132 h->slice_beta_offset = 52;
3133 if (h->pps.deblocking_filter_parameters_present) {
3134 tmp = get_ue_golomb_31(&s->gb);
3136 av_log(s->avctx, AV_LOG_ERROR,
3137 "deblocking_filter_idc %u out of range\n", tmp);
3140 h->deblocking_filter = tmp;
3141 if (h->deblocking_filter < 2)
3142 h->deblocking_filter ^= 1; // 1<->0
3144 if (h->deblocking_filter) {
3145 h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
3146 h->slice_beta_offset += get_se_golomb(&s->gb) << 1;
3147 if (h->slice_alpha_c0_offset > 104U ||
3148 h->slice_beta_offset > 104U) {
3149 av_log(s->avctx, AV_LOG_ERROR,
3150 "deblocking filter parameters %d %d out of range\n",
3151 h->slice_alpha_c0_offset, h->slice_beta_offset);
3157 if (s->avctx->skip_loop_filter >= AVDISCARD_ALL ||
3158 (s->avctx->skip_loop_filter >= AVDISCARD_NONKEY &&
3159 h->slice_type_nos != AV_PICTURE_TYPE_I) ||
3160 (s->avctx->skip_loop_filter >= AVDISCARD_BIDIR &&
3161 h->slice_type_nos == AV_PICTURE_TYPE_B) ||
3162 (s->avctx->skip_loop_filter >= AVDISCARD_NONREF &&
3163 h->nal_ref_idc == 0))
3164 h->deblocking_filter = 0;
3166 if (h->deblocking_filter == 1 && h0->max_contexts > 1) {
3167 if (s->avctx->flags2 & CODEC_FLAG2_FAST) {
3168 /* Cheat slightly for speed:
3169 * Do not bother to deblock across slices. */
3170 h->deblocking_filter = 2;
3172 h0->max_contexts = 1;
3173 if (!h0->single_decode_warning) {
3174 av_log(s->avctx, AV_LOG_INFO,
3175 "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
3176 h0->single_decode_warning = 1;
3179 av_log(h->s.avctx, AV_LOG_ERROR,
3180 "Deblocking switched inside frame.\n");
3185 h->qp_thresh = 15 + 52 -
3186 FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset) -
3188 h->pps.chroma_qp_index_offset[0],
3189 h->pps.chroma_qp_index_offset[1]) +
3190 6 * (h->sps.bit_depth_luma - 8);
3192 h0->last_slice_type = slice_type;
3193 h->slice_num = ++h0->current_slice;
3194 if (h->slice_num >= MAX_SLICES) {
3195 av_log(s->avctx, AV_LOG_ERROR,
3196 "Too many slices, increase MAX_SLICES and recompile\n");
3199 for (j = 0; j < 2; j++) {
3201 int *ref2frm = h->ref2frm[h->slice_num & (MAX_SLICES - 1)][j];
3202 for (i = 0; i < 16; i++) {
3204 if (h->ref_list[j][i].f.data[0]) {
3206 uint8_t *base = h->ref_list[j][i].f.base[0];
3207 for (k = 0; k < h->short_ref_count; k++)
3208 if (h->short_ref[k]->f.base[0] == base) {
3212 for (k = 0; k < h->long_ref_count; k++)
3213 if (h->long_ref[k] && h->long_ref[k]->f.base[0] == base) {
3214 id_list[i] = h->short_ref_count + k;
3222 for (i = 0; i < 16; i++)
3223 ref2frm[i + 2] = 4 * id_list[i] +
3224 (h->ref_list[j][i].f.reference & 3);
3226 ref2frm[18 + 1] = -1;
3227 for (i = 16; i < 48; i++)
3228 ref2frm[i + 4] = 4 * id_list[(i - 16) >> 1] +
3229 (h->ref_list[j][i].f.reference & 3);
3232 // FIXME: fix draw_edges + PAFF + frame threads
3233 h->emu_edge_width = (s->flags & CODEC_FLAG_EMU_EDGE ||
3234 (!h->sps.frame_mbs_only_flag &&
3235 s->avctx->active_thread_type))
3237 h->emu_edge_height = (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
3239 if (s->avctx->debug & FF_DEBUG_PICT_INFO) {
3240 av_log(h->s.avctx, AV_LOG_DEBUG,
3241 "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",
3243 (s->picture_structure == PICT_FRAME ? "F" : s->picture_structure == PICT_TOP_FIELD ? "T" : "B"),
3245 av_get_picture_type_char(h->slice_type),
3246 h->slice_type_fixed ? " fix" : "",
3247 h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
3248 pps_id, h->frame_num,
3249 s->current_picture_ptr->field_poc[0],
3250 s->current_picture_ptr->field_poc[1],
3251 h->ref_count[0], h->ref_count[1],
3253 h->deblocking_filter,
3254 h->slice_alpha_c0_offset / 2 - 26, h->slice_beta_offset / 2 - 26,
3256 h->use_weight == 1 && h->use_weight_chroma ? "c" : "",
3257 h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : "");
3263 int ff_h264_get_slice_type(const H264Context *h)
3265 switch (h->slice_type) {
3266 case AV_PICTURE_TYPE_P:
3268 case AV_PICTURE_TYPE_B:
3270 case AV_PICTURE_TYPE_I:
3272 case AV_PICTURE_TYPE_SP:
3274 case AV_PICTURE_TYPE_SI:
3281 static av_always_inline void fill_filter_caches_inter(H264Context *h,
3282 MpegEncContext *const s,
3283 int mb_type, int top_xy,
3284 int left_xy[LEFT_MBS],
3286 int left_type[LEFT_MBS],
3287 int mb_xy, int list)
3289 int b_stride = h->b_stride;
3290 int16_t(*mv_dst)[2] = &h->mv_cache[list][scan8[0]];
3291 int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
3292 if (IS_INTER(mb_type) || IS_DIRECT(mb_type)) {
3293 if (USES_LIST(top_type, list)) {
3294 const int b_xy = h->mb2b_xy[top_xy] + 3 * b_stride;
3295 const int b8_xy = 4 * top_xy + 2;
3296 int (*ref2frm)[64] = h->ref2frm[h->slice_table[top_xy] & (MAX_SLICES - 1)][0] + (MB_MBAFF ? 20 : 2);
3297 AV_COPY128(mv_dst - 1 * 8, s->current_picture.f.motion_val[list][b_xy + 0]);
3298 ref_cache[0 - 1 * 8] =
3299 ref_cache[1 - 1 * 8] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 0]];
3300 ref_cache[2 - 1 * 8] =
3301 ref_cache[3 - 1 * 8] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 1]];
3303 AV_ZERO128(mv_dst - 1 * 8);
3304 AV_WN32A(&ref_cache[0 - 1 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3307 if (!IS_INTERLACED(mb_type ^ left_type[LTOP])) {
3308 if (USES_LIST(left_type[LTOP], list)) {
3309 const int b_xy = h->mb2b_xy[left_xy[LTOP]] + 3;
3310 const int b8_xy = 4 * left_xy[LTOP] + 1;
3311 int (*ref2frm)[64] = h->ref2frm[h->slice_table[left_xy[LTOP]] & (MAX_SLICES - 1)][0] + (MB_MBAFF ? 20 : 2);
3312 AV_COPY32(mv_dst - 1 + 0, s->current_picture.f.motion_val[list][b_xy + b_stride * 0]);
3313 AV_COPY32(mv_dst - 1 + 8, s->current_picture.f.motion_val[list][b_xy + b_stride * 1]);
3314 AV_COPY32(mv_dst - 1 + 16, s->current_picture.f.motion_val[list][b_xy + b_stride * 2]);
3315 AV_COPY32(mv_dst - 1 + 24, s->current_picture.f.motion_val[list][b_xy + b_stride * 3]);
3317 ref_cache[-1 + 8] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2 * 0]];
3318 ref_cache[-1 + 16] =
3319 ref_cache[-1 + 24] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2 * 1]];
3321 AV_ZERO32(mv_dst - 1 + 0);
3322 AV_ZERO32(mv_dst - 1 + 8);
3323 AV_ZERO32(mv_dst - 1 + 16);
3324 AV_ZERO32(mv_dst - 1 + 24);
3327 ref_cache[-1 + 16] =
3328 ref_cache[-1 + 24] = LIST_NOT_USED;
3333 if (!USES_LIST(mb_type, list)) {
3334 fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0, 0), 4);
3335 AV_WN32A(&ref_cache[0 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3336 AV_WN32A(&ref_cache[1 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3337 AV_WN32A(&ref_cache[2 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3338 AV_WN32A(&ref_cache[3 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3343 int8_t *ref = &s->current_picture.f.ref_index[list][4 * mb_xy];
3344 int (*ref2frm)[64] = h->ref2frm[h->slice_num & (MAX_SLICES - 1)][0] + (MB_MBAFF ? 20 : 2);
3345 uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]], ref2frm[list][ref[1]]) & 0x00FF00FF) * 0x0101;
3346 uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]], ref2frm[list][ref[3]]) & 0x00FF00FF) * 0x0101;
3347 AV_WN32A(&ref_cache[0 * 8], ref01);
3348 AV_WN32A(&ref_cache[1 * 8], ref01);
3349 AV_WN32A(&ref_cache[2 * 8], ref23);
3350 AV_WN32A(&ref_cache[3 * 8], ref23);
3354 int16_t(*mv_src)[2] = &s->current_picture.f.motion_val[list][4 * s->mb_x + 4 * s->mb_y * b_stride];
3355 AV_COPY128(mv_dst + 8 * 0, mv_src + 0 * b_stride);
3356 AV_COPY128(mv_dst + 8 * 1, mv_src + 1 * b_stride);
3357 AV_COPY128(mv_dst + 8 * 2, mv_src + 2 * b_stride);
3358 AV_COPY128(mv_dst + 8 * 3, mv_src + 3 * b_stride);
3364 * @return non zero if the loop filter can be skipped
3366 static int fill_filter_caches(H264Context *h, int mb_type)
3368 MpegEncContext *const s = &h->s;
3369 const int mb_xy = h->mb_xy;
3370 int top_xy, left_xy[LEFT_MBS];
3371 int top_type, left_type[LEFT_MBS];
3375 top_xy = mb_xy - (s->mb_stride << MB_FIELD);
3377 /* Wow, what a mess, why didn't they simplify the interlacing & intra
3378 * stuff, I can't imagine that these complex rules are worth it. */
3380 left_xy[LBOT] = left_xy[LTOP] = mb_xy - 1;
3382 const int left_mb_field_flag = IS_INTERLACED(s->current_picture.f.mb_type[mb_xy - 1]);
3383 const int curr_mb_field_flag = IS_INTERLACED(mb_type);
3385 if (left_mb_field_flag != curr_mb_field_flag)
3386 left_xy[LTOP] -= s->mb_stride;
3388 if (curr_mb_field_flag)
3389 top_xy += s->mb_stride &
3390 (((s->current_picture.f.mb_type[top_xy] >> 7) & 1) - 1);
3391 if (left_mb_field_flag != curr_mb_field_flag)
3392 left_xy[LBOT] += s->mb_stride;
3396 h->top_mb_xy = top_xy;
3397 h->left_mb_xy[LTOP] = left_xy[LTOP];
3398 h->left_mb_xy[LBOT] = left_xy[LBOT];
3400 /* For sufficiently low qp, filtering wouldn't do anything.
3401 * This is a conservative estimate: could also check beta_offset
3402 * and more accurate chroma_qp. */
3403 int qp_thresh = h->qp_thresh; // FIXME strictly we should store qp_thresh for each mb of a slice
3404 int qp = s->current_picture.f.qscale_table[mb_xy];
3405 if (qp <= qp_thresh &&
3406 (left_xy[LTOP] < 0 ||
3407 ((qp + s->current_picture.f.qscale_table[left_xy[LTOP]] + 1) >> 1) <= qp_thresh) &&
3409 ((qp + s->current_picture.f.qscale_table[top_xy] + 1) >> 1) <= qp_thresh)) {
3412 if ((left_xy[LTOP] < 0 ||
3413 ((qp + s->current_picture.f.qscale_table[left_xy[LBOT]] + 1) >> 1) <= qp_thresh) &&
3414 (top_xy < s->mb_stride ||
3415 ((qp + s->current_picture.f.qscale_table[top_xy - s->mb_stride] + 1) >> 1) <= qp_thresh))
3420 top_type = s->current_picture.f.mb_type[top_xy];
3421 left_type[LTOP] = s->current_picture.f.mb_type[left_xy[LTOP]];
3422 left_type[LBOT] = s->current_picture.f.mb_type[left_xy[LBOT]];
3423 if (h->deblocking_filter == 2) {
3424 if (h->slice_table[top_xy] != h->slice_num)
3426 if (h->slice_table[left_xy[LBOT]] != h->slice_num)
3427 left_type[LTOP] = left_type[LBOT] = 0;
3429 if (h->slice_table[top_xy] == 0xFFFF)
3431 if (h->slice_table[left_xy[LBOT]] == 0xFFFF)
3432 left_type[LTOP] = left_type[LBOT] = 0;
3434 h->top_type = top_type;
3435 h->left_type[LTOP] = left_type[LTOP];
3436 h->left_type[LBOT] = left_type[LBOT];
3438 if (IS_INTRA(mb_type))
3441 fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy,
3442 top_type, left_type, mb_xy, 0);
3443 if (h->list_count == 2)
3444 fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy,
3445 top_type, left_type, mb_xy, 1);
3447 nnz = h->non_zero_count[mb_xy];
3448 nnz_cache = h->non_zero_count_cache;
3449 AV_COPY32(&nnz_cache[4 + 8 * 1], &nnz[0]);
3450 AV_COPY32(&nnz_cache[4 + 8 * 2], &nnz[4]);
3451 AV_COPY32(&nnz_cache[4 + 8 * 3], &nnz[8]);
3452 AV_COPY32(&nnz_cache[4 + 8 * 4], &nnz[12]);
3453 h->cbp = h->cbp_table[mb_xy];
3456 nnz = h->non_zero_count[top_xy];
3457 AV_COPY32(&nnz_cache[4 + 8 * 0], &nnz[3 * 4]);
3460 if (left_type[LTOP]) {
3461 nnz = h->non_zero_count[left_xy[LTOP]];
3462 nnz_cache[3 + 8 * 1] = nnz[3 + 0 * 4];
3463 nnz_cache[3 + 8 * 2] = nnz[3 + 1 * 4];
3464 nnz_cache[3 + 8 * 3] = nnz[3 + 2 * 4];
3465 nnz_cache[3 + 8 * 4] = nnz[3 + 3 * 4];
3468 /* CAVLC 8x8dct requires NNZ values for residual decoding that differ
3469 * from what the loop filter needs */
3470 if (!CABAC && h->pps.transform_8x8_mode) {
3471 if (IS_8x8DCT(top_type)) {
3472 nnz_cache[4 + 8 * 0] =
3473 nnz_cache[5 + 8 * 0] = (h->cbp_table[top_xy] & 0x4000) >> 12;
3474 nnz_cache[6 + 8 * 0] =
3475 nnz_cache[7 + 8 * 0] = (h->cbp_table[top_xy] & 0x8000) >> 12;
3477 if (IS_8x8DCT(left_type[LTOP])) {
3478 nnz_cache[3 + 8 * 1] =
3479 nnz_cache[3 + 8 * 2] = (h->cbp_table[left_xy[LTOP]] & 0x2000) >> 12; // FIXME check MBAFF
3481 if (IS_8x8DCT(left_type[LBOT])) {
3482 nnz_cache[3 + 8 * 3] =
3483 nnz_cache[3 + 8 * 4] = (h->cbp_table[left_xy[LBOT]] & 0x8000) >> 12; // FIXME check MBAFF
3486 if (IS_8x8DCT(mb_type)) {
3487 nnz_cache[scan8[0]] =
3488 nnz_cache[scan8[1]] =
3489 nnz_cache[scan8[2]] =
3490 nnz_cache[scan8[3]] = (h->cbp & 0x1000) >> 12;
3492 nnz_cache[scan8[0 + 4]] =
3493 nnz_cache[scan8[1 + 4]] =
3494 nnz_cache[scan8[2 + 4]] =
3495 nnz_cache[scan8[3 + 4]] = (h->cbp & 0x2000) >> 12;
3497 nnz_cache[scan8[0 + 8]] =
3498 nnz_cache[scan8[1 + 8]] =
3499 nnz_cache[scan8[2 + 8]] =
3500 nnz_cache[scan8[3 + 8]] = (h->cbp & 0x4000) >> 12;
3502 nnz_cache[scan8[0 + 12]] =
3503 nnz_cache[scan8[1 + 12]] =
3504 nnz_cache[scan8[2 + 12]] =
3505 nnz_cache[scan8[3 + 12]] = (h->cbp & 0x8000) >> 12;
3512 static void loop_filter(H264Context *h, int start_x, int end_x)
3514 MpegEncContext *const s = &h->s;
3515 uint8_t *dest_y, *dest_cb, *dest_cr;
3516 int linesize, uvlinesize, mb_x, mb_y;
3517 const int end_mb_y = s->mb_y + FRAME_MBAFF;
3518 const int old_slice_type = h->slice_type;
3519 const int pixel_shift = h->pixel_shift;
3520 const int block_h = 16 >> s->chroma_y_shift;
3522 if (h->deblocking_filter) {
3523 for (mb_x = start_x; mb_x < end_x; mb_x++)
3524 for (mb_y = end_mb_y - FRAME_MBAFF; mb_y <= end_mb_y; mb_y++) {
3526 mb_xy = h->mb_xy = mb_x + mb_y * s->mb_stride;
3527 h->slice_num = h->slice_table[mb_xy];
3528 mb_type = s->current_picture.f.mb_type[mb_xy];
3529 h->list_count = h->list_counts[mb_xy];
3533 h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
3537 dest_y = s->current_picture.f.data[0] +
3538 ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
3539 dest_cb = s->current_picture.f.data[1] +
3540 (mb_x << pixel_shift) * (8 << CHROMA444) +
3541 mb_y * s->uvlinesize * block_h;
3542 dest_cr = s->current_picture.f.data[2] +
3543 (mb_x << pixel_shift) * (8 << CHROMA444) +
3544 mb_y * s->uvlinesize * block_h;
3545 // FIXME simplify above
3548 linesize = h->mb_linesize = s->linesize * 2;
3549 uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
3550 if (mb_y & 1) { // FIXME move out of this function?
3551 dest_y -= s->linesize * 15;
3552 dest_cb -= s->uvlinesize * (block_h - 1);
3553 dest_cr -= s->uvlinesize * (block_h - 1);
3556 linesize = h->mb_linesize = s->linesize;
3557 uvlinesize = h->mb_uvlinesize = s->uvlinesize;
3559 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize,
3561 if (fill_filter_caches(h, mb_type))
3563 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.f.qscale_table[mb_xy]);
3564 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.f.qscale_table[mb_xy]);
3567 ff_h264_filter_mb(h, mb_x, mb_y, dest_y, dest_cb, dest_cr,
3568 linesize, uvlinesize);
3570 ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb,
3571 dest_cr, linesize, uvlinesize);
3575 h->slice_type = old_slice_type;
3577 s->mb_y = end_mb_y - FRAME_MBAFF;
3578 h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3579 h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3582 static void predict_field_decoding_flag(H264Context *h)
3584 MpegEncContext *const s = &h->s;
3585 const int mb_xy = s->mb_x + s->mb_y * s->mb_stride;
3586 int mb_type = (h->slice_table[mb_xy - 1] == h->slice_num) ?
3587 s->current_picture.f.mb_type[mb_xy - 1] :
3588 (h->slice_table[mb_xy - s->mb_stride] == h->slice_num) ?
3589 s->current_picture.f.mb_type[mb_xy - s->mb_stride] : 0;
3590 h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
3594 * Draw edges and report progress for the last MB row.
3596 static void decode_finish_row(H264Context *h)
3598 MpegEncContext *const s = &h->s;
3599 int top = 16 * (s->mb_y >> FIELD_PICTURE);
3600 int pic_height = 16 * s->mb_height >> FIELD_PICTURE;
3601 int height = 16 << FRAME_MBAFF;
3602 int deblock_border = (16 + 4) << FRAME_MBAFF;
3604 if (h->deblocking_filter) {
3605 if ((top + height) >= pic_height)
3606 height += deblock_border;
3607 top -= deblock_border;
3610 if (top >= pic_height || (top + height) < h->emu_edge_height)
3613 height = FFMIN(height, pic_height - top);
3614 if (top < h->emu_edge_height) {
3615 height = top + height;
3619 ff_draw_horiz_band(s, top, height);
3624 ff_thread_report_progress(&s->current_picture_ptr->f, top + height - 1,
3625 s->picture_structure == PICT_BOTTOM_FIELD);
3628 static int decode_slice(struct AVCodecContext *avctx, void *arg)
3630 H264Context *h = *(void **)arg;
3631 MpegEncContext *const s = &h->s;
3632 const int part_mask = s->partitioned_frame ? (ER_AC_END | ER_AC_ERROR)
3634 int lf_x_start = s->mb_x;
3636 s->mb_skip_run = -1;
3638 h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME ||
3639 s->codec_id != AV_CODEC_ID_H264 ||
3640 (CONFIG_GRAY && (s->flags & CODEC_FLAG_GRAY));
3644 align_get_bits(&s->gb);
3647 ff_init_cabac_states(&h->cabac);
3648 ff_init_cabac_decoder(&h->cabac,
3649 s->gb.buffer + get_bits_count(&s->gb) / 8,
3650 (get_bits_left(&s->gb) + 7) / 8);
3652 ff_h264_init_cabac_states(h);
3656 int ret = ff_h264_decode_mb_cabac(h);
3658 // STOP_TIMER("decode_mb_cabac")
3661 ff_h264_hl_decode_mb(h);
3663 // FIXME optimal? or let mb_decode decode 16x32 ?
3664 if (ret >= 0 && FRAME_MBAFF) {
3667 ret = ff_h264_decode_mb_cabac(h);
3670 ff_h264_hl_decode_mb(h);
3673 eos = get_cabac_terminate(&h->cabac);
3675 if ((s->workaround_bugs & FF_BUG_TRUNCATED) &&
3676 h->cabac.bytestream > h->cabac.bytestream_end + 2) {
3677 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x - 1,
3678 s->mb_y, ER_MB_END & part_mask);
3679 if (s->mb_x >= lf_x_start)
3680 loop_filter(h, lf_x_start, s->mb_x + 1);
3683 if (ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 2) {
3684 av_log(h->s.avctx, AV_LOG_ERROR,
3685 "error while decoding MB %d %d, bytestream (%td)\n",
3687 h->cabac.bytestream_end - h->cabac.bytestream);
3688 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x,
3689 s->mb_y, ER_MB_ERROR & part_mask);
3693 if (++s->mb_x >= s->mb_width) {
3694 loop_filter(h, lf_x_start, s->mb_x);
3695 s->mb_x = lf_x_start = 0;
3696 decode_finish_row(h);
3698 if (FIELD_OR_MBAFF_PICTURE) {
3700 if (FRAME_MBAFF && s->mb_y < s->mb_height)
3701 predict_field_decoding_flag(h);
3705 if (eos || s->mb_y >= s->mb_height) {
3706 tprintf(s->avctx, "slice end %d %d\n",
3707 get_bits_count(&s->gb), s->gb.size_in_bits);
3708 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x - 1,
3709 s->mb_y, ER_MB_END & part_mask);
3710 if (s->mb_x > lf_x_start)
3711 loop_filter(h, lf_x_start, s->mb_x);
3717 int ret = ff_h264_decode_mb_cavlc(h);
3720 ff_h264_hl_decode_mb(h);
3722 // FIXME optimal? or let mb_decode decode 16x32 ?
3723 if (ret >= 0 && FRAME_MBAFF) {
3725 ret = ff_h264_decode_mb_cavlc(h);
3728 ff_h264_hl_decode_mb(h);
3733 av_log(h->s.avctx, AV_LOG_ERROR,
3734 "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
3735 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x,
3736 s->mb_y, ER_MB_ERROR & part_mask);
3740 if (++s->mb_x >= s->mb_width) {
3741 loop_filter(h, lf_x_start, s->mb_x);
3742 s->mb_x = lf_x_start = 0;
3743 decode_finish_row(h);
3745 if (FIELD_OR_MBAFF_PICTURE) {
3747 if (FRAME_MBAFF && s->mb_y < s->mb_height)
3748 predict_field_decoding_flag(h);
3750 if (s->mb_y >= s->mb_height) {
3751 tprintf(s->avctx, "slice end %d %d\n",
3752 get_bits_count(&s->gb), s->gb.size_in_bits);
3754 if (get_bits_left(&s->gb) == 0) {
3755 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y,
3756 s->mb_x - 1, s->mb_y,
3757 ER_MB_END & part_mask);
3761 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y,
3762 s->mb_x - 1, s->mb_y,
3763 ER_MB_END & part_mask);
3770 if (get_bits_left(&s->gb) <= 0 && s->mb_skip_run <= 0) {
3771 tprintf(s->avctx, "slice end %d %d\n",
3772 get_bits_count(&s->gb), s->gb.size_in_bits);
3773 if (get_bits_left(&s->gb) == 0) {
3774 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y,
3775 s->mb_x - 1, s->mb_y,
3776 ER_MB_END & part_mask);
3777 if (s->mb_x > lf_x_start)
3778 loop_filter(h, lf_x_start, s->mb_x);
3782 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x,
3783 s->mb_y, ER_MB_ERROR & part_mask);
3793 * Call decode_slice() for each context.
3795 * @param h h264 master context
3796 * @param context_count number of contexts to execute
3798 static int execute_decode_slices(H264Context *h, int context_count)
3800 MpegEncContext *const s = &h->s;
3801 AVCodecContext *const avctx = s->avctx;
3805 if (s->avctx->hwaccel ||
3806 s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
3808 if (context_count == 1) {
3809 return decode_slice(avctx, &h);
3811 for (i = 1; i < context_count; i++) {
3812 hx = h->thread_context[i];
3813 hx->s.err_recognition = avctx->err_recognition;
3814 hx->s.error_count = 0;
3817 avctx->execute(avctx, decode_slice, h->thread_context,
3818 NULL, context_count, sizeof(void *));
3820 /* pull back stuff from slices to master context */
3821 hx = h->thread_context[context_count - 1];
3822 s->mb_x = hx->s.mb_x;
3823 s->mb_y = hx->s.mb_y;
3824 s->droppable = hx->s.droppable;
3825 s->picture_structure = hx->s.picture_structure;
3826 for (i = 1; i < context_count; i++)
3827 h->s.error_count += h->thread_context[i]->s.error_count;
3833 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size,
3834 int parse_extradata)
3836 MpegEncContext *const s = &h->s;
3837 AVCodecContext *const avctx = s->avctx;
3838 H264Context *hx; ///< thread context
3842 int pass = !(avctx->active_thread_type & FF_THREAD_FRAME);
3843 int nals_needed = 0; ///< number of NALs that need decoding before the next frame thread starts
3846 h->max_contexts = s->slice_context_count;
3847 if (!(s->flags2 & CODEC_FLAG2_CHUNKS)) {
3848 h->current_slice = 0;
3849 if (!s->first_field)
3850 s->current_picture_ptr = NULL;
3851 ff_h264_reset_sei(h);
3854 for (; pass <= 1; pass++) {
3857 next_avc = h->is_avc ? 0 : buf_size;
3867 if (buf_index >= next_avc) {
3868 if (buf_index >= buf_size - h->nal_length_size)
3871 for (i = 0; i < h->nal_length_size; i++)
3872 nalsize = (nalsize << 8) | buf[buf_index++];
3873 if (nalsize <= 0 || nalsize > buf_size - buf_index) {
3874 av_log(h->s.avctx, AV_LOG_ERROR,
3875 "AVC: nal size %d\n", nalsize);
3878 next_avc = buf_index + nalsize;
3880 // start code prefix search
3881 for (; buf_index + 3 < next_avc; buf_index++)
3882 // This should always succeed in the first iteration.
3883 if (buf[buf_index] == 0 &&
3884 buf[buf_index + 1] == 0 &&
3885 buf[buf_index + 2] == 1)
3888 if (buf_index + 3 >= buf_size) {
3889 buf_index = buf_size;
3894 if (buf_index >= next_avc)
3898 hx = h->thread_context[context_count];
3900 ptr = ff_h264_decode_nal(hx, buf + buf_index, &dst_length,
3901 &consumed, next_avc - buf_index);
3902 if (ptr == NULL || dst_length < 0) {
3906 i = buf_index + consumed;
3907 if ((s->workaround_bugs & FF_BUG_AUTODETECT) && i + 3 < next_avc &&
3908 buf[i] == 0x00 && buf[i + 1] == 0x00 &&
3909 buf[i + 2] == 0x01 && buf[i + 3] == 0xE0)
3910 s->workaround_bugs |= FF_BUG_TRUNCATED;
3912 if (!(s->workaround_bugs & FF_BUG_TRUNCATED))
3913 while (ptr[dst_length - 1] == 0 && dst_length > 0)
3915 bit_length = !dst_length ? 0
3917 decode_rbsp_trailing(h, ptr + dst_length - 1));
3919 if (s->avctx->debug & FF_DEBUG_STARTCODE)
3920 av_log(h->s.avctx, AV_LOG_DEBUG,
3921 "NAL %d at %d/%d length %d\n",
3922 hx->nal_unit_type, buf_index, buf_size, dst_length);
3924 if (h->is_avc && (nalsize != consumed) && nalsize)
3925 av_log(h->s.avctx, AV_LOG_DEBUG,
3926 "AVC: Consumed only %d bytes instead of %d\n",
3929 buf_index += consumed;
3933 /* packets can sometimes contain multiple PPS/SPS,
3934 * e.g. two PAFF field pictures in one packet, or a demuxer
3935 * which splits NALs strangely if so, when frame threading we
3936 * can't start the next thread until we've read all of them */
3937 switch (hx->nal_unit_type) {
3940 nals_needed = nal_index;
3945 init_get_bits(&hx->s.gb, ptr, bit_length);
3946 if (!get_ue_golomb(&hx->s.gb))
3947 nals_needed = nal_index;
3952 // FIXME do not discard SEI id
3953 if (avctx->skip_frame >= AVDISCARD_NONREF && h->nal_ref_idc == 0)
3957 /* Ignore every NAL unit type except PPS and SPS during extradata
3958 * parsing. Decoding slices is not possible in codec init
3960 if (parse_extradata && HAVE_THREADS &&
3961 (s->avctx->active_thread_type & FF_THREAD_FRAME) &&
3962 (hx->nal_unit_type != NAL_PPS &&
3963 hx->nal_unit_type != NAL_SPS)) {
3964 av_log(avctx, AV_LOG_INFO, "Ignoring NAL unit %d during "
3965 "extradata parsing\n", hx->nal_unit_type);
3966 hx->nal_unit_type = NAL_FF_IGNORE;
3969 switch (hx->nal_unit_type) {
3971 if (h->nal_unit_type != NAL_IDR_SLICE) {
3972 av_log(h->s.avctx, AV_LOG_ERROR,
3973 "Invalid mix of idr and non-idr slices\n");
3977 idr(h); // FIXME ensure we don't lose some frames if there is reordering
3979 init_get_bits(&hx->s.gb, ptr, bit_length);
3981 hx->inter_gb_ptr = &hx->s.gb;
3982 hx->s.data_partitioning = 0;
3984 if ((err = decode_slice_header(hx, h)))
3987 s->current_picture_ptr->f.key_frame |=
3988 (hx->nal_unit_type == NAL_IDR_SLICE) ||
3989 (h->sei_recovery_frame_cnt >= 0);
3991 if (h->current_slice == 1) {
3992 if (!(s->flags2 & CODEC_FLAG2_CHUNKS))
3993 decode_postinit(h, nal_index >= nals_needed);
3995 if (s->avctx->hwaccel &&
3996 s->avctx->hwaccel->start_frame(s->avctx, NULL, 0) < 0)
3998 if (CONFIG_H264_VDPAU_DECODER &&
3999 s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
4000 ff_vdpau_h264_picture_start(s);
4003 if (hx->redundant_pic_count == 0 &&
4004 (avctx->skip_frame < AVDISCARD_NONREF ||
4006 (avctx->skip_frame < AVDISCARD_BIDIR ||
4007 hx->slice_type_nos != AV_PICTURE_TYPE_B) &&
4008 (avctx->skip_frame < AVDISCARD_NONKEY ||
4009 hx->slice_type_nos == AV_PICTURE_TYPE_I) &&
4010 avctx->skip_frame < AVDISCARD_ALL) {
4011 if (avctx->hwaccel) {
4012 if (avctx->hwaccel->decode_slice(avctx,
4013 &buf[buf_index - consumed],
4016 } else if (CONFIG_H264_VDPAU_DECODER &&
4017 s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU) {
4018 static const uint8_t start_code[] = {
4020 ff_vdpau_add_data_chunk(s, start_code,
4021 sizeof(start_code));
4022 ff_vdpau_add_data_chunk(s, &buf[buf_index - consumed],
4029 init_get_bits(&hx->s.gb, ptr, bit_length);
4031 hx->inter_gb_ptr = NULL;
4033 if ((err = decode_slice_header(hx, h)) < 0)
4036 hx->s.data_partitioning = 1;
4039 init_get_bits(&hx->intra_gb, ptr, bit_length);
4040 hx->intra_gb_ptr = &hx->intra_gb;
4043 init_get_bits(&hx->inter_gb, ptr, bit_length);
4044 hx->inter_gb_ptr = &hx->inter_gb;
4046 if (hx->redundant_pic_count == 0 &&
4048 hx->s.data_partitioning &&
4049 s->current_picture_ptr &&
4050 s->context_initialized &&
4051 (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc) &&
4052 (avctx->skip_frame < AVDISCARD_BIDIR ||
4053 hx->slice_type_nos != AV_PICTURE_TYPE_B) &&
4054 (avctx->skip_frame < AVDISCARD_NONKEY ||
4055 hx->slice_type_nos == AV_PICTURE_TYPE_I) &&
4056 avctx->skip_frame < AVDISCARD_ALL)
4060 init_get_bits(&s->gb, ptr, bit_length);
4061 ff_h264_decode_sei(h);
4064 init_get_bits(&s->gb, ptr, bit_length);
4065 if (ff_h264_decode_seq_parameter_set(h) < 0 &&
4066 h->is_avc && (nalsize != consumed) && nalsize) {
4067 av_log(h->s.avctx, AV_LOG_DEBUG,
4068 "SPS decoding failure, trying again with the complete NAL\n");
4069 init_get_bits(&s->gb, buf + buf_index + 1 - consumed,
4071 ff_h264_decode_seq_parameter_set(h);
4074 if (h264_set_parameter_from_sps(h) < 0) {
4080 init_get_bits(&s->gb, ptr, bit_length);
4081 ff_h264_decode_picture_parameter_set(h, bit_length);
4084 case NAL_END_SEQUENCE:
4085 case NAL_END_STREAM:
4086 case NAL_FILLER_DATA:
4088 case NAL_AUXILIARY_SLICE:
4093 av_log(avctx, AV_LOG_DEBUG, "Unknown NAL code: %d (%d bits)\n",
4094 hx->nal_unit_type, bit_length);
4097 if (context_count == h->max_contexts) {
4098 execute_decode_slices(h, context_count);
4103 av_log(h->s.avctx, AV_LOG_ERROR, "decode_slice_header error\n");
4104 else if (err == 1) {
4105 /* Slice could not be decoded in parallel mode, copy down
4106 * NAL unit stuff to context 0 and restart. Note that
4107 * rbsp_buffer is not transferred, but since we no longer
4108 * run in parallel mode this should not be an issue. */
4109 h->nal_unit_type = hx->nal_unit_type;
4110 h->nal_ref_idc = hx->nal_ref_idc;
4117 execute_decode_slices(h, context_count);
4121 if (s->current_picture_ptr && s->current_picture_ptr->owner2 == s &&
4123 ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX,
4124 s->picture_structure == PICT_BOTTOM_FIELD);
4131 * Return the number of bytes consumed for building the current frame.
4133 static int get_consumed_bytes(MpegEncContext *s, int pos, int buf_size)
4136 pos = 1; // avoid infinite loops (i doubt that is needed but ...)
4137 if (pos + 10 > buf_size)
4138 pos = buf_size; // oops ;)
4143 static int decode_frame(AVCodecContext *avctx, void *data,
4144 int *got_frame, AVPacket *avpkt)
4146 const uint8_t *buf = avpkt->data;
4147 int buf_size = avpkt->size;
4148 H264Context *h = avctx->priv_data;
4149 MpegEncContext *s = &h->s;
4150 AVFrame *pict = data;
4153 s->flags = avctx->flags;
4154 s->flags2 = avctx->flags2;
4156 /* end of stream, output what is still in the buffers */
4158 if (buf_size == 0) {
4162 s->current_picture_ptr = NULL;
4164 // FIXME factorize this with the output code below
4165 out = h->delayed_pic[0];
4168 h->delayed_pic[i] &&
4169 !h->delayed_pic[i]->f.key_frame &&
4170 !h->delayed_pic[i]->mmco_reset;
4172 if (h->delayed_pic[i]->poc < out->poc) {
4173 out = h->delayed_pic[i];
4177 for (i = out_idx; h->delayed_pic[i]; i++)
4178 h->delayed_pic[i] = h->delayed_pic[i + 1];
4188 buf_index = decode_nal_units(h, buf, buf_size, 0);
4192 if (!s->current_picture_ptr && h->nal_unit_type == NAL_END_SEQUENCE) {
4197 if (!(s->flags2 & CODEC_FLAG2_CHUNKS) && !s->current_picture_ptr) {
4198 if (avctx->skip_frame >= AVDISCARD_NONREF)
4200 av_log(avctx, AV_LOG_ERROR, "no frame!\n");
4204 if (!(s->flags2 & CODEC_FLAG2_CHUNKS) ||
4205 (s->mb_y >= s->mb_height && s->mb_height)) {
4206 if (s->flags2 & CODEC_FLAG2_CHUNKS)
4207 decode_postinit(h, 1);
4210 h->context_reinitialized = 0;
4212 if (!h->next_output_pic) {
4213 /* Wait for second field. */
4217 *pict = h->next_output_pic->f;
4221 assert(pict->data[0] || !*got_frame);
4222 ff_print_debug_info(s, pict);
4224 return get_consumed_bytes(s, buf_index, buf_size);
4227 av_cold void ff_h264_free_context(H264Context *h)
4231 free_tables(h, 1); // FIXME cleanup init stuff perhaps
4233 for (i = 0; i < MAX_SPS_COUNT; i++)
4234 av_freep(h->sps_buffers + i);
4236 for (i = 0; i < MAX_PPS_COUNT; i++)
4237 av_freep(h->pps_buffers + i);
4240 static av_cold int h264_decode_end(AVCodecContext *avctx)
4242 H264Context *h = avctx->priv_data;
4243 MpegEncContext *s = &h->s;
4245 ff_h264_free_context(h);
4247 ff_MPV_common_end(s);
4249 // memset(h, 0, sizeof(H264Context));
4254 static const AVProfile profiles[] = {
4255 { FF_PROFILE_H264_BASELINE, "Baseline" },
4256 { FF_PROFILE_H264_CONSTRAINED_BASELINE, "Constrained Baseline" },
4257 { FF_PROFILE_H264_MAIN, "Main" },
4258 { FF_PROFILE_H264_EXTENDED, "Extended" },
4259 { FF_PROFILE_H264_HIGH, "High" },
4260 { FF_PROFILE_H264_HIGH_10, "High 10" },
4261 { FF_PROFILE_H264_HIGH_10_INTRA, "High 10 Intra" },
4262 { FF_PROFILE_H264_HIGH_422, "High 4:2:2" },
4263 { FF_PROFILE_H264_HIGH_422_INTRA, "High 4:2:2 Intra" },
4264 { FF_PROFILE_H264_HIGH_444, "High 4:4:4" },
4265 { FF_PROFILE_H264_HIGH_444_PREDICTIVE, "High 4:4:4 Predictive" },
4266 { FF_PROFILE_H264_HIGH_444_INTRA, "High 4:4:4 Intra" },
4267 { FF_PROFILE_H264_CAVLC_444, "CAVLC 4:4:4" },
4268 { FF_PROFILE_UNKNOWN },
4271 AVCodec ff_h264_decoder = {
4273 .type = AVMEDIA_TYPE_VIDEO,
4274 .id = AV_CODEC_ID_H264,
4275 .priv_data_size = sizeof(H264Context),
4276 .init = ff_h264_decode_init,
4277 .close = h264_decode_end,
4278 .decode = decode_frame,
4279 .capabilities = /*CODEC_CAP_DRAW_HORIZ_BAND |*/ CODEC_CAP_DR1 |
4280 CODEC_CAP_DELAY | CODEC_CAP_SLICE_THREADS |
4281 CODEC_CAP_FRAME_THREADS,
4283 .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10"),
4284 .init_thread_copy = ONLY_IF_THREADS_ENABLED(decode_init_thread_copy),
4285 .update_thread_context = ONLY_IF_THREADS_ENABLED(decode_update_thread_context),
4286 .profiles = NULL_IF_CONFIG_SMALL(profiles),
4289 #if CONFIG_H264_VDPAU_DECODER
4290 AVCodec ff_h264_vdpau_decoder = {
4291 .name = "h264_vdpau",
4292 .type = AVMEDIA_TYPE_VIDEO,
4293 .id = AV_CODEC_ID_H264,
4294 .priv_data_size = sizeof(H264Context),
4295 .init = ff_h264_decode_init,
4296 .close = h264_decode_end,
4297 .decode = decode_frame,
4298 .capabilities = CODEC_CAP_DR1 | CODEC_CAP_DELAY | CODEC_CAP_HWACCEL_VDPAU,
4300 .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10 (VDPAU acceleration)"),
4301 .pix_fmts = (const enum AVPixelFormat[]) { AV_PIX_FMT_VDPAU_H264,
4303 .profiles = NULL_IF_CONFIG_SMALL(profiles),