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 FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
24 * H.264 / AVC / MPEG4 part10 codec.
25 * @author Michael Niedermayer <michaelni@gmx.at>
28 #define UNCHECKED_BITSTREAM_READER 1
30 #include "libavutil/imgutils.h"
31 #include "libavutil/opt.h"
34 #include "cabac_functions.h"
37 #include "mpegvideo.h"
40 #include "h264_mvpred.h"
43 #include "rectangle.h"
45 #include "vdpau_internal.h"
46 #include "libavutil/avassert.h"
51 const uint16_t ff_h264_mb_sizes[4] = { 256, 384, 512, 768 };
53 static const uint8_t rem6[QP_MAX_NUM + 1] = {
54 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2,
55 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5,
56 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
59 static const uint8_t div6[QP_MAX_NUM + 1] = {
60 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3, 3, 3,
61 3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6,
62 7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 10, 10, 10, 10,
65 static const enum PixelFormat hwaccel_pixfmt_list_h264_jpeg_420[] = {
74 * Check if the top & left blocks are available if needed and
75 * change the dc mode so it only uses the available blocks.
77 int ff_h264_check_intra4x4_pred_mode(H264Context *h)
79 MpegEncContext *const s = &h->s;
80 static const int8_t top[12] = {
81 -1, 0, LEFT_DC_PRED, -1, -1, -1, -1, -1, 0
83 static const int8_t left[12] = {
84 0, -1, TOP_DC_PRED, 0, -1, -1, -1, 0, -1, DC_128_PRED
88 if (!(h->top_samples_available & 0x8000)) {
89 for (i = 0; i < 4; i++) {
90 int status = top[h->intra4x4_pred_mode_cache[scan8[0] + i]];
92 av_log(h->s.avctx, AV_LOG_ERROR,
93 "top block unavailable for requested intra4x4 mode %d at %d %d\n",
94 status, s->mb_x, s->mb_y);
97 h->intra4x4_pred_mode_cache[scan8[0] + i] = status;
102 if ((h->left_samples_available & 0x8888) != 0x8888) {
103 static const int mask[4] = { 0x8000, 0x2000, 0x80, 0x20 };
104 for (i = 0; i < 4; i++)
105 if (!(h->left_samples_available & mask[i])) {
106 int status = left[h->intra4x4_pred_mode_cache[scan8[0] + 8 * i]];
108 av_log(h->s.avctx, AV_LOG_ERROR,
109 "left block unavailable for requested intra4x4 mode %d at %d %d\n",
110 status, s->mb_x, s->mb_y);
113 h->intra4x4_pred_mode_cache[scan8[0] + 8 * i] = status;
119 } // FIXME cleanup like ff_h264_check_intra_pred_mode
122 * Check if the top & left blocks are available if needed and
123 * change the dc mode so it only uses the available blocks.
125 int ff_h264_check_intra_pred_mode(H264Context *h, int mode, int is_chroma)
127 MpegEncContext *const s = &h->s;
128 static const int8_t top[7] = { LEFT_DC_PRED8x8, 1, -1, -1 };
129 static const int8_t left[7] = { TOP_DC_PRED8x8, -1, 2, -1, DC_128_PRED8x8 };
132 av_log(h->s.avctx, AV_LOG_ERROR,
133 "out of range intra chroma pred mode at %d %d\n",
138 if (!(h->top_samples_available & 0x8000)) {
141 av_log(h->s.avctx, AV_LOG_ERROR,
142 "top block unavailable for requested intra mode at %d %d\n",
148 if ((h->left_samples_available & 0x8080) != 0x8080) {
150 if (is_chroma && (h->left_samples_available & 0x8080)) {
151 // mad cow disease mode, aka MBAFF + constrained_intra_pred
152 mode = ALZHEIMER_DC_L0T_PRED8x8 +
153 (!(h->left_samples_available & 0x8000)) +
154 2 * (mode == DC_128_PRED8x8);
157 av_log(h->s.avctx, AV_LOG_ERROR,
158 "left block unavailable for requested intra mode at %d %d\n",
167 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src,
168 int *dst_length, int *consumed, int length)
174 // src[0]&0x80; // forbidden bit
175 h->nal_ref_idc = src[0] >> 5;
176 h->nal_unit_type = src[0] & 0x1F;
181 #if HAVE_FAST_UNALIGNED
184 for (i = 0; i + 1 < length; i += 9) {
185 if (!((~AV_RN64A(src + i) &
186 (AV_RN64A(src + i) - 0x0100010001000101ULL)) &
187 0x8000800080008080ULL))
190 for (i = 0; i + 1 < length; i += 5) {
191 if (!((~AV_RN32A(src + i) &
192 (AV_RN32A(src + i) - 0x01000101U)) &
196 if (i > 0 && !src[i])
202 for (i = 0; i + 1 < length; i += 2) {
205 if (i > 0 && src[i - 1] == 0)
208 if (i + 2 < length && src[i + 1] == 0 && src[i + 2] <= 3) {
209 if (src[i + 2] != 3) {
210 /* startcode, so we must be past the end */
218 // use second escape buffer for inter data
219 bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0;
221 si = h->rbsp_buffer_size[bufidx];
222 av_fast_padded_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+MAX_MBPAIR_SIZE);
223 dst = h->rbsp_buffer[bufidx];
228 if(i>=length-1){ //no escaped 0
230 *consumed= length+1; //+1 for the header
231 if(h->s.avctx->flags2 & CODEC_FLAG2_FAST){
234 memcpy(dst, src, length);
239 // printf("decoding esc\n");
242 while (si + 2 < length) {
243 // remove escapes (very rare 1:2^22)
244 if (src[si + 2] > 3) {
245 dst[di++] = src[si++];
246 dst[di++] = src[si++];
247 } else if (src[si] == 0 && src[si + 1] == 0) {
248 if (src[si + 2] == 3) { // escape
253 } else // next start code
257 dst[di++] = src[si++];
260 dst[di++] = src[si++];
263 memset(dst + di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
266 *consumed = si + 1; // +1 for the header
267 /* FIXME store exact number of bits in the getbitcontext
268 * (it is needed for decoding) */
273 * Identify the exact end of the bitstream
274 * @return the length of the trailing, or 0 if damaged
276 static int decode_rbsp_trailing(H264Context *h, const uint8_t *src)
281 tprintf(h->s.avctx, "rbsp trailing %X\n", v);
283 for (r = 1; r < 9; r++) {
291 static inline int get_lowest_part_list_y(H264Context *h, Picture *pic, int n,
292 int height, int y_offset, int list)
294 int raw_my = h->mv_cache[list][scan8[n]][1];
295 int filter_height = (raw_my & 3) ? 2 : 0;
296 int full_my = (raw_my >> 2) + y_offset;
297 int top = full_my - filter_height;
298 int bottom = full_my + filter_height + height;
300 return FFMAX(abs(top), bottom);
303 static inline void get_lowest_part_y(H264Context *h, int refs[2][48], int n,
304 int height, int y_offset, int list0,
305 int list1, int *nrefs)
307 MpegEncContext *const s = &h->s;
310 y_offset += 16 * (s->mb_y >> MB_FIELD);
313 int ref_n = h->ref_cache[0][scan8[n]];
314 Picture *ref = &h->ref_list[0][ref_n];
316 // Error resilience puts the current picture in the ref list.
317 // Don't try to wait on these as it will cause a deadlock.
318 // Fields can wait on each other, though.
319 if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
320 (ref->f.reference & 3) != s->picture_structure) {
321 my = get_lowest_part_list_y(h, ref, n, height, y_offset, 0);
322 if (refs[0][ref_n] < 0)
324 refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
329 int ref_n = h->ref_cache[1][scan8[n]];
330 Picture *ref = &h->ref_list[1][ref_n];
332 if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
333 (ref->f.reference & 3) != s->picture_structure) {
334 my = get_lowest_part_list_y(h, ref, n, height, y_offset, 1);
335 if (refs[1][ref_n] < 0)
337 refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
343 * Wait until all reference frames are available for MC operations.
345 * @param h the H264 context
347 static void await_references(H264Context *h)
349 MpegEncContext *const s = &h->s;
350 const int mb_xy = h->mb_xy;
351 const int mb_type = s->current_picture.f.mb_type[mb_xy];
353 int nrefs[2] = { 0 };
356 memset(refs, -1, sizeof(refs));
358 if (IS_16X16(mb_type)) {
359 get_lowest_part_y(h, refs, 0, 16, 0,
360 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
361 } else if (IS_16X8(mb_type)) {
362 get_lowest_part_y(h, refs, 0, 8, 0,
363 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
364 get_lowest_part_y(h, refs, 8, 8, 8,
365 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
366 } else if (IS_8X16(mb_type)) {
367 get_lowest_part_y(h, refs, 0, 16, 0,
368 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
369 get_lowest_part_y(h, refs, 4, 16, 0,
370 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
374 assert(IS_8X8(mb_type));
376 for (i = 0; i < 4; i++) {
377 const int sub_mb_type = h->sub_mb_type[i];
379 int y_offset = (i & 2) << 2;
381 if (IS_SUB_8X8(sub_mb_type)) {
382 get_lowest_part_y(h, refs, n, 8, y_offset,
383 IS_DIR(sub_mb_type, 0, 0),
384 IS_DIR(sub_mb_type, 0, 1),
386 } else if (IS_SUB_8X4(sub_mb_type)) {
387 get_lowest_part_y(h, refs, n, 4, y_offset,
388 IS_DIR(sub_mb_type, 0, 0),
389 IS_DIR(sub_mb_type, 0, 1),
391 get_lowest_part_y(h, refs, n + 2, 4, y_offset + 4,
392 IS_DIR(sub_mb_type, 0, 0),
393 IS_DIR(sub_mb_type, 0, 1),
395 } else if (IS_SUB_4X8(sub_mb_type)) {
396 get_lowest_part_y(h, refs, n, 8, y_offset,
397 IS_DIR(sub_mb_type, 0, 0),
398 IS_DIR(sub_mb_type, 0, 1),
400 get_lowest_part_y(h, refs, n + 1, 8, y_offset,
401 IS_DIR(sub_mb_type, 0, 0),
402 IS_DIR(sub_mb_type, 0, 1),
406 assert(IS_SUB_4X4(sub_mb_type));
407 for (j = 0; j < 4; j++) {
408 int sub_y_offset = y_offset + 2 * (j & 2);
409 get_lowest_part_y(h, refs, n + j, 4, sub_y_offset,
410 IS_DIR(sub_mb_type, 0, 0),
411 IS_DIR(sub_mb_type, 0, 1),
418 for (list = h->list_count - 1; list >= 0; list--)
419 for (ref = 0; ref < 48 && nrefs[list]; ref++) {
420 int row = refs[list][ref];
422 Picture *ref_pic = &h->ref_list[list][ref];
423 int ref_field = ref_pic->f.reference - 1;
424 int ref_field_picture = ref_pic->field_picture;
425 int pic_height = 16 * s->mb_height >> ref_field_picture;
430 if (!FIELD_PICTURE && ref_field_picture) { // frame referencing two fields
431 ff_thread_await_progress(&ref_pic->f,
432 FFMIN((row >> 1) - !(row & 1),
435 ff_thread_await_progress(&ref_pic->f,
436 FFMIN((row >> 1), pic_height - 1),
438 } else if (FIELD_PICTURE && !ref_field_picture) { // field referencing one field of a frame
439 ff_thread_await_progress(&ref_pic->f,
440 FFMIN(row * 2 + ref_field,
443 } else if (FIELD_PICTURE) {
444 ff_thread_await_progress(&ref_pic->f,
445 FFMIN(row, pic_height - 1),
448 ff_thread_await_progress(&ref_pic->f,
449 FFMIN(row, pic_height - 1),
456 static av_always_inline void mc_dir_part(H264Context *h, Picture *pic,
457 int n, int square, int height,
459 uint8_t *dest_y, uint8_t *dest_cb,
461 int src_x_offset, int src_y_offset,
462 qpel_mc_func *qpix_op,
463 h264_chroma_mc_func chroma_op,
464 int pixel_shift, int chroma_idc)
466 MpegEncContext *const s = &h->s;
467 const int mx = h->mv_cache[list][scan8[n]][0] + src_x_offset * 8;
468 int my = h->mv_cache[list][scan8[n]][1] + src_y_offset * 8;
469 const int luma_xy = (mx & 3) + ((my & 3) << 2);
470 int offset = ((mx >> 2) << pixel_shift) + (my >> 2) * h->mb_linesize;
471 uint8_t *src_y = pic->f.data[0] + offset;
472 uint8_t *src_cb, *src_cr;
473 int extra_width = h->emu_edge_width;
474 int extra_height = h->emu_edge_height;
476 const int full_mx = mx >> 2;
477 const int full_my = my >> 2;
478 const int pic_width = 16 * s->mb_width;
479 const int pic_height = 16 * s->mb_height >> MB_FIELD;
487 if (full_mx < 0 - extra_width ||
488 full_my < 0 - extra_height ||
489 full_mx + 16 /*FIXME*/ > pic_width + extra_width ||
490 full_my + 16 /*FIXME*/ > pic_height + extra_height) {
491 s->dsp.emulated_edge_mc(s->edge_emu_buffer,
492 src_y - (2 << pixel_shift) - 2 * h->mb_linesize,
494 16 + 5, 16 + 5 /*FIXME*/, full_mx - 2,
495 full_my - 2, pic_width, pic_height);
496 src_y = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
500 qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); // FIXME try variable height perhaps?
502 qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
504 if (CONFIG_GRAY && s->flags & CODEC_FLAG_GRAY)
507 if (chroma_idc == 3 /* yuv444 */) {
508 src_cb = pic->f.data[1] + offset;
510 s->dsp.emulated_edge_mc(s->edge_emu_buffer,
511 src_cb - (2 << pixel_shift) - 2 * h->mb_linesize,
513 16 + 5, 16 + 5 /*FIXME*/,
514 full_mx - 2, full_my - 2,
515 pic_width, pic_height);
516 src_cb = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
518 qpix_op[luma_xy](dest_cb, src_cb, h->mb_linesize); // FIXME try variable height perhaps?
520 qpix_op[luma_xy](dest_cb + delta, src_cb + delta, h->mb_linesize);
522 src_cr = pic->f.data[2] + offset;
524 s->dsp.emulated_edge_mc(s->edge_emu_buffer,
525 src_cr - (2 << pixel_shift) - 2 * h->mb_linesize,
527 16 + 5, 16 + 5 /*FIXME*/,
528 full_mx - 2, full_my - 2,
529 pic_width, pic_height);
530 src_cr = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
532 qpix_op[luma_xy](dest_cr, src_cr, h->mb_linesize); // FIXME try variable height perhaps?
534 qpix_op[luma_xy](dest_cr + delta, src_cr + delta, h->mb_linesize);
538 ysh = 3 - (chroma_idc == 2 /* yuv422 */);
539 if (chroma_idc == 1 /* yuv420 */ && MB_FIELD) {
540 // chroma offset when predicting from a field of opposite parity
541 my += 2 * ((s->mb_y & 1) - (pic->f.reference - 1));
542 emu |= (my >> 3) < 0 || (my >> 3) + 8 >= (pic_height >> 1);
545 src_cb = pic->f.data[1] + ((mx >> 3) << pixel_shift) +
546 (my >> ysh) * h->mb_uvlinesize;
547 src_cr = pic->f.data[2] + ((mx >> 3) << pixel_shift) +
548 (my >> ysh) * h->mb_uvlinesize;
551 s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize,
552 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
553 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
554 src_cb = s->edge_emu_buffer;
556 chroma_op(dest_cb, src_cb, h->mb_uvlinesize,
557 height >> (chroma_idc == 1 /* yuv420 */),
558 mx & 7, (my << (chroma_idc == 2 /* yuv422 */)) & 7);
561 s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize,
562 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
563 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
564 src_cr = s->edge_emu_buffer;
566 chroma_op(dest_cr, src_cr, h->mb_uvlinesize, height >> (chroma_idc == 1 /* yuv420 */),
567 mx & 7, (my << (chroma_idc == 2 /* yuv422 */)) & 7);
570 static av_always_inline void mc_part_std(H264Context *h, int n, int square,
571 int height, int delta,
572 uint8_t *dest_y, uint8_t *dest_cb,
574 int x_offset, int y_offset,
575 qpel_mc_func *qpix_put,
576 h264_chroma_mc_func chroma_put,
577 qpel_mc_func *qpix_avg,
578 h264_chroma_mc_func chroma_avg,
579 int list0, int list1,
580 int pixel_shift, int chroma_idc)
582 MpegEncContext *const s = &h->s;
583 qpel_mc_func *qpix_op = qpix_put;
584 h264_chroma_mc_func chroma_op = chroma_put;
586 dest_y += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
587 if (chroma_idc == 3 /* yuv444 */) {
588 dest_cb += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
589 dest_cr += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
590 } else if (chroma_idc == 2 /* yuv422 */) {
591 dest_cb += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
592 dest_cr += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
593 } else { /* yuv420 */
594 dest_cb += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
595 dest_cr += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
597 x_offset += 8 * s->mb_x;
598 y_offset += 8 * (s->mb_y >> MB_FIELD);
601 Picture *ref = &h->ref_list[0][h->ref_cache[0][scan8[n]]];
602 mc_dir_part(h, ref, n, square, height, delta, 0,
603 dest_y, dest_cb, dest_cr, x_offset, y_offset,
604 qpix_op, chroma_op, pixel_shift, chroma_idc);
607 chroma_op = chroma_avg;
611 Picture *ref = &h->ref_list[1][h->ref_cache[1][scan8[n]]];
612 mc_dir_part(h, ref, n, square, height, delta, 1,
613 dest_y, dest_cb, dest_cr, x_offset, y_offset,
614 qpix_op, chroma_op, pixel_shift, chroma_idc);
618 static av_always_inline void mc_part_weighted(H264Context *h, int n, int square,
619 int height, int delta,
620 uint8_t *dest_y, uint8_t *dest_cb,
622 int x_offset, int y_offset,
623 qpel_mc_func *qpix_put,
624 h264_chroma_mc_func chroma_put,
625 h264_weight_func luma_weight_op,
626 h264_weight_func chroma_weight_op,
627 h264_biweight_func luma_weight_avg,
628 h264_biweight_func chroma_weight_avg,
629 int list0, int list1,
630 int pixel_shift, int chroma_idc)
632 MpegEncContext *const s = &h->s;
635 dest_y += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
636 if (chroma_idc == 3 /* yuv444 */) {
637 chroma_height = height;
638 chroma_weight_avg = luma_weight_avg;
639 chroma_weight_op = luma_weight_op;
640 dest_cb += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
641 dest_cr += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
642 } else if (chroma_idc == 2 /* yuv422 */) {
643 chroma_height = height;
644 dest_cb += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
645 dest_cr += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
646 } else { /* yuv420 */
647 chroma_height = height >> 1;
648 dest_cb += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
649 dest_cr += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
651 x_offset += 8 * s->mb_x;
652 y_offset += 8 * (s->mb_y >> MB_FIELD);
654 if (list0 && list1) {
655 /* don't optimize for luma-only case, since B-frames usually
656 * use implicit weights => chroma too. */
657 uint8_t *tmp_cb = s->obmc_scratchpad;
658 uint8_t *tmp_cr = s->obmc_scratchpad + (16 << pixel_shift);
659 uint8_t *tmp_y = s->obmc_scratchpad + 16 * h->mb_uvlinesize;
660 int refn0 = h->ref_cache[0][scan8[n]];
661 int refn1 = h->ref_cache[1][scan8[n]];
663 mc_dir_part(h, &h->ref_list[0][refn0], n, square, height, delta, 0,
664 dest_y, dest_cb, dest_cr,
665 x_offset, y_offset, qpix_put, chroma_put,
666 pixel_shift, chroma_idc);
667 mc_dir_part(h, &h->ref_list[1][refn1], n, square, height, delta, 1,
668 tmp_y, tmp_cb, tmp_cr,
669 x_offset, y_offset, qpix_put, chroma_put,
670 pixel_shift, chroma_idc);
672 if (h->use_weight == 2) {
673 int weight0 = h->implicit_weight[refn0][refn1][s->mb_y & 1];
674 int weight1 = 64 - weight0;
675 luma_weight_avg(dest_y, tmp_y, h->mb_linesize,
676 height, 5, weight0, weight1, 0);
677 chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize,
678 chroma_height, 5, weight0, weight1, 0);
679 chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize,
680 chroma_height, 5, weight0, weight1, 0);
682 luma_weight_avg(dest_y, tmp_y, h->mb_linesize, height,
683 h->luma_log2_weight_denom,
684 h->luma_weight[refn0][0][0],
685 h->luma_weight[refn1][1][0],
686 h->luma_weight[refn0][0][1] +
687 h->luma_weight[refn1][1][1]);
688 chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, chroma_height,
689 h->chroma_log2_weight_denom,
690 h->chroma_weight[refn0][0][0][0],
691 h->chroma_weight[refn1][1][0][0],
692 h->chroma_weight[refn0][0][0][1] +
693 h->chroma_weight[refn1][1][0][1]);
694 chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, chroma_height,
695 h->chroma_log2_weight_denom,
696 h->chroma_weight[refn0][0][1][0],
697 h->chroma_weight[refn1][1][1][0],
698 h->chroma_weight[refn0][0][1][1] +
699 h->chroma_weight[refn1][1][1][1]);
702 int list = list1 ? 1 : 0;
703 int refn = h->ref_cache[list][scan8[n]];
704 Picture *ref = &h->ref_list[list][refn];
705 mc_dir_part(h, ref, n, square, height, delta, list,
706 dest_y, dest_cb, dest_cr, x_offset, y_offset,
707 qpix_put, chroma_put, pixel_shift, chroma_idc);
709 luma_weight_op(dest_y, h->mb_linesize, height,
710 h->luma_log2_weight_denom,
711 h->luma_weight[refn][list][0],
712 h->luma_weight[refn][list][1]);
713 if (h->use_weight_chroma) {
714 chroma_weight_op(dest_cb, h->mb_uvlinesize, chroma_height,
715 h->chroma_log2_weight_denom,
716 h->chroma_weight[refn][list][0][0],
717 h->chroma_weight[refn][list][0][1]);
718 chroma_weight_op(dest_cr, h->mb_uvlinesize, chroma_height,
719 h->chroma_log2_weight_denom,
720 h->chroma_weight[refn][list][1][0],
721 h->chroma_weight[refn][list][1][1]);
726 static av_always_inline void mc_part(H264Context *h, int n, int square,
727 int height, int delta,
728 uint8_t *dest_y, uint8_t *dest_cb,
730 int x_offset, int y_offset,
731 qpel_mc_func *qpix_put,
732 h264_chroma_mc_func chroma_put,
733 qpel_mc_func *qpix_avg,
734 h264_chroma_mc_func chroma_avg,
735 h264_weight_func *weight_op,
736 h264_biweight_func *weight_avg,
737 int list0, int list1,
738 int pixel_shift, int chroma_idc)
740 if ((h->use_weight == 2 && list0 && list1 &&
741 (h->implicit_weight[h->ref_cache[0][scan8[n]]][h->ref_cache[1][scan8[n]]][h->s.mb_y & 1] != 32)) ||
743 mc_part_weighted(h, n, square, height, delta, dest_y, dest_cb, dest_cr,
744 x_offset, y_offset, qpix_put, chroma_put,
745 weight_op[0], weight_op[1], weight_avg[0],
746 weight_avg[1], list0, list1, pixel_shift, chroma_idc);
748 mc_part_std(h, n, square, height, delta, dest_y, dest_cb, dest_cr,
749 x_offset, y_offset, qpix_put, chroma_put, qpix_avg,
750 chroma_avg, list0, list1, pixel_shift, chroma_idc);
753 static av_always_inline void prefetch_motion(H264Context *h, int list,
754 int pixel_shift, int chroma_idc)
756 /* fetch pixels for estimated mv 4 macroblocks ahead
757 * optimized for 64byte cache lines */
758 MpegEncContext *const s = &h->s;
759 const int refn = h->ref_cache[list][scan8[0]];
761 const int mx = (h->mv_cache[list][scan8[0]][0] >> 2) + 16 * s->mb_x + 8;
762 const int my = (h->mv_cache[list][scan8[0]][1] >> 2) + 16 * s->mb_y;
763 uint8_t **src = h->ref_list[list][refn].f.data;
764 int off = (mx << pixel_shift) +
765 (my + (s->mb_x & 3) * 4) * h->mb_linesize +
767 s->dsp.prefetch(src[0] + off, s->linesize, 4);
768 if (chroma_idc == 3 /* yuv444 */) {
769 s->dsp.prefetch(src[1] + off, s->linesize, 4);
770 s->dsp.prefetch(src[2] + off, s->linesize, 4);
772 off= (((mx>>1)+64)<<pixel_shift) + ((my>>1) + (s->mb_x&7))*s->uvlinesize;
773 s->dsp.prefetch(src[1] + off, src[2] - src[1], 2);
778 static av_always_inline void hl_motion(H264Context *h, uint8_t *dest_y,
779 uint8_t *dest_cb, uint8_t *dest_cr,
780 qpel_mc_func(*qpix_put)[16],
781 h264_chroma_mc_func(*chroma_put),
782 qpel_mc_func(*qpix_avg)[16],
783 h264_chroma_mc_func(*chroma_avg),
784 h264_weight_func *weight_op,
785 h264_biweight_func *weight_avg,
786 int pixel_shift, int chroma_idc)
788 MpegEncContext *const s = &h->s;
789 const int mb_xy = h->mb_xy;
790 const int mb_type = s->current_picture.f.mb_type[mb_xy];
792 assert(IS_INTER(mb_type));
794 if (HAVE_THREADS && (s->avctx->active_thread_type & FF_THREAD_FRAME))
796 prefetch_motion(h, 0, pixel_shift, chroma_idc);
798 if (IS_16X16(mb_type)) {
799 mc_part(h, 0, 1, 16, 0, dest_y, dest_cb, dest_cr, 0, 0,
800 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
801 weight_op, weight_avg,
802 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
803 pixel_shift, chroma_idc);
804 } else if (IS_16X8(mb_type)) {
805 mc_part(h, 0, 0, 8, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 0,
806 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
807 weight_op, weight_avg,
808 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
809 pixel_shift, chroma_idc);
810 mc_part(h, 8, 0, 8, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 4,
811 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
812 weight_op, weight_avg,
813 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
814 pixel_shift, chroma_idc);
815 } else if (IS_8X16(mb_type)) {
816 mc_part(h, 0, 0, 16, 8 * h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
817 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
818 &weight_op[1], &weight_avg[1],
819 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
820 pixel_shift, chroma_idc);
821 mc_part(h, 4, 0, 16, 8 * h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
822 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
823 &weight_op[1], &weight_avg[1],
824 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
825 pixel_shift, chroma_idc);
829 assert(IS_8X8(mb_type));
831 for (i = 0; i < 4; i++) {
832 const int sub_mb_type = h->sub_mb_type[i];
834 int x_offset = (i & 1) << 2;
835 int y_offset = (i & 2) << 1;
837 if (IS_SUB_8X8(sub_mb_type)) {
838 mc_part(h, n, 1, 8, 0, dest_y, dest_cb, dest_cr,
840 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
841 &weight_op[1], &weight_avg[1],
842 IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
843 pixel_shift, chroma_idc);
844 } else if (IS_SUB_8X4(sub_mb_type)) {
845 mc_part(h, n, 0, 4, 4 << pixel_shift, dest_y, dest_cb, dest_cr,
847 qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
848 &weight_op[1], &weight_avg[1],
849 IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
850 pixel_shift, chroma_idc);
851 mc_part(h, n + 2, 0, 4, 4 << pixel_shift,
852 dest_y, dest_cb, dest_cr, x_offset, y_offset + 2,
853 qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
854 &weight_op[1], &weight_avg[1],
855 IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
856 pixel_shift, chroma_idc);
857 } else if (IS_SUB_4X8(sub_mb_type)) {
858 mc_part(h, n, 0, 8, 4 * h->mb_linesize,
859 dest_y, dest_cb, dest_cr, x_offset, y_offset,
860 qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
861 &weight_op[2], &weight_avg[2],
862 IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
863 pixel_shift, chroma_idc);
864 mc_part(h, n + 1, 0, 8, 4 * h->mb_linesize,
865 dest_y, dest_cb, dest_cr, x_offset + 2, y_offset,
866 qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
867 &weight_op[2], &weight_avg[2],
868 IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
869 pixel_shift, chroma_idc);
872 assert(IS_SUB_4X4(sub_mb_type));
873 for (j = 0; j < 4; j++) {
874 int sub_x_offset = x_offset + 2 * (j & 1);
875 int sub_y_offset = y_offset + (j & 2);
876 mc_part(h, n + j, 1, 4, 0,
877 dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
878 qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
879 &weight_op[2], &weight_avg[2],
880 IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
881 pixel_shift, chroma_idc);
887 prefetch_motion(h, 1, pixel_shift, chroma_idc);
890 static av_always_inline void hl_motion_420(H264Context *h, uint8_t *dest_y,
891 uint8_t *dest_cb, uint8_t *dest_cr,
892 qpel_mc_func(*qpix_put)[16],
893 h264_chroma_mc_func(*chroma_put),
894 qpel_mc_func(*qpix_avg)[16],
895 h264_chroma_mc_func(*chroma_avg),
896 h264_weight_func *weight_op,
897 h264_biweight_func *weight_avg,
900 hl_motion(h, dest_y, dest_cb, dest_cr, qpix_put, chroma_put,
901 qpix_avg, chroma_avg, weight_op, weight_avg, pixel_shift, 1);
904 static av_always_inline void hl_motion_422(H264Context *h, uint8_t *dest_y,
905 uint8_t *dest_cb, uint8_t *dest_cr,
906 qpel_mc_func(*qpix_put)[16],
907 h264_chroma_mc_func(*chroma_put),
908 qpel_mc_func(*qpix_avg)[16],
909 h264_chroma_mc_func(*chroma_avg),
910 h264_weight_func *weight_op,
911 h264_biweight_func *weight_avg,
914 hl_motion(h, dest_y, dest_cb, dest_cr, qpix_put, chroma_put,
915 qpix_avg, chroma_avg, weight_op, weight_avg, pixel_shift, 2);
918 static void free_tables(H264Context *h, int free_rbsp)
923 av_freep(&h->intra4x4_pred_mode);
924 av_freep(&h->chroma_pred_mode_table);
925 av_freep(&h->cbp_table);
926 av_freep(&h->mvd_table[0]);
927 av_freep(&h->mvd_table[1]);
928 av_freep(&h->direct_table);
929 av_freep(&h->non_zero_count);
930 av_freep(&h->slice_table_base);
931 h->slice_table = NULL;
932 av_freep(&h->list_counts);
934 av_freep(&h->mb2b_xy);
935 av_freep(&h->mb2br_xy);
937 for (i = 0; i < MAX_THREADS; i++) {
938 hx = h->thread_context[i];
941 av_freep(&hx->top_borders[1]);
942 av_freep(&hx->top_borders[0]);
943 av_freep(&hx->s.obmc_scratchpad);
945 av_freep(&hx->rbsp_buffer[1]);
946 av_freep(&hx->rbsp_buffer[0]);
947 hx->rbsp_buffer_size[0] = 0;
948 hx->rbsp_buffer_size[1] = 0;
951 av_freep(&h->thread_context[i]);
955 static void init_dequant8_coeff_table(H264Context *h)
958 const int max_qp = 51 + 6 * (h->sps.bit_depth_luma - 8);
960 for (i = 0; i < 6; i++) {
961 h->dequant8_coeff[i] = h->dequant8_buffer[i];
962 for (j = 0; j < i; j++)
963 if (!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i],
964 64 * sizeof(uint8_t))) {
965 h->dequant8_coeff[i] = h->dequant8_buffer[j];
971 for (q = 0; q < max_qp + 1; q++) {
974 for (x = 0; x < 64; x++)
975 h->dequant8_coeff[i][q][(x >> 3) | ((x & 7) << 3)] =
976 ((uint32_t)dequant8_coeff_init[idx][dequant8_coeff_init_scan[((x >> 1) & 12) | (x & 3)]] *
977 h->pps.scaling_matrix8[i][x]) << shift;
982 static void init_dequant4_coeff_table(H264Context *h)
985 const int max_qp = 51 + 6 * (h->sps.bit_depth_luma - 8);
986 for (i = 0; i < 6; i++) {
987 h->dequant4_coeff[i] = h->dequant4_buffer[i];
988 for (j = 0; j < i; j++)
989 if (!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i],
990 16 * sizeof(uint8_t))) {
991 h->dequant4_coeff[i] = h->dequant4_buffer[j];
997 for (q = 0; q < max_qp + 1; q++) {
998 int shift = div6[q] + 2;
1000 for (x = 0; x < 16; x++)
1001 h->dequant4_coeff[i][q][(x >> 2) | ((x << 2) & 0xF)] =
1002 ((uint32_t)dequant4_coeff_init[idx][(x & 1) + ((x >> 2) & 1)] *
1003 h->pps.scaling_matrix4[i][x]) << shift;
1008 static void init_dequant_tables(H264Context *h)
1011 init_dequant4_coeff_table(h);
1012 if (h->pps.transform_8x8_mode)
1013 init_dequant8_coeff_table(h);
1014 if (h->sps.transform_bypass) {
1015 for (i = 0; i < 6; i++)
1016 for (x = 0; x < 16; x++)
1017 h->dequant4_coeff[i][0][x] = 1 << 6;
1018 if (h->pps.transform_8x8_mode)
1019 for (i = 0; i < 6; i++)
1020 for (x = 0; x < 64; x++)
1021 h->dequant8_coeff[i][0][x] = 1 << 6;
1025 int ff_h264_alloc_tables(H264Context *h)
1027 MpegEncContext *const s = &h->s;
1028 const int big_mb_num = s->mb_stride * (s->mb_height + 1);
1029 const int row_mb_num = 2*s->mb_stride*FFMAX(s->avctx->thread_count, 1);
1032 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode,
1033 row_mb_num * 8 * sizeof(uint8_t), fail)
1034 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count,
1035 big_mb_num * 48 * sizeof(uint8_t), fail)
1036 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base,
1037 (big_mb_num + s->mb_stride) * sizeof(*h->slice_table_base), fail)
1038 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table,
1039 big_mb_num * sizeof(uint16_t), fail)
1040 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table,
1041 big_mb_num * sizeof(uint8_t), fail)
1042 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0],
1043 16 * row_mb_num * sizeof(uint8_t), fail);
1044 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1],
1045 16 * row_mb_num * sizeof(uint8_t), fail);
1046 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table,
1047 4 * big_mb_num * sizeof(uint8_t), fail);
1048 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts,
1049 big_mb_num * sizeof(uint8_t), fail)
1051 memset(h->slice_table_base, -1,
1052 (big_mb_num + s->mb_stride) * sizeof(*h->slice_table_base));
1053 h->slice_table = h->slice_table_base + s->mb_stride * 2 + 1;
1055 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy,
1056 big_mb_num * sizeof(uint32_t), fail);
1057 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy,
1058 big_mb_num * sizeof(uint32_t), fail);
1059 for (y = 0; y < s->mb_height; y++)
1060 for (x = 0; x < s->mb_width; x++) {
1061 const int mb_xy = x + y * s->mb_stride;
1062 const int b_xy = 4 * x + 4 * y * h->b_stride;
1064 h->mb2b_xy[mb_xy] = b_xy;
1065 h->mb2br_xy[mb_xy] = 8 * (FMO ? mb_xy : (mb_xy % (2 * s->mb_stride)));
1068 s->obmc_scratchpad = NULL;
1070 if (!h->dequant4_coeff[0])
1071 init_dequant_tables(h);
1081 * Mimic alloc_tables(), but for every context thread.
1083 static void clone_tables(H264Context *dst, H264Context *src, int i)
1085 MpegEncContext *const s = &src->s;
1086 dst->intra4x4_pred_mode = src->intra4x4_pred_mode + i * 8 * 2 * s->mb_stride;
1087 dst->non_zero_count = src->non_zero_count;
1088 dst->slice_table = src->slice_table;
1089 dst->cbp_table = src->cbp_table;
1090 dst->mb2b_xy = src->mb2b_xy;
1091 dst->mb2br_xy = src->mb2br_xy;
1092 dst->chroma_pred_mode_table = src->chroma_pred_mode_table;
1093 dst->mvd_table[0] = src->mvd_table[0] + i * 8 * 2 * s->mb_stride;
1094 dst->mvd_table[1] = src->mvd_table[1] + i * 8 * 2 * s->mb_stride;
1095 dst->direct_table = src->direct_table;
1096 dst->list_counts = src->list_counts;
1097 dst->s.obmc_scratchpad = NULL;
1098 ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma,
1099 src->sps.chroma_format_idc);
1104 * Allocate buffers which are not shared amongst multiple threads.
1106 static int context_init(H264Context *h)
1108 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0],
1109 h->s.mb_width * 16 * 3 * sizeof(uint8_t) * 2, fail)
1110 FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1],
1111 h->s.mb_width * 16 * 3 * sizeof(uint8_t) * 2, fail)
1113 h->ref_cache[0][scan8[5] + 1] =
1114 h->ref_cache[0][scan8[7] + 1] =
1115 h->ref_cache[0][scan8[13] + 1] =
1116 h->ref_cache[1][scan8[5] + 1] =
1117 h->ref_cache[1][scan8[7] + 1] =
1118 h->ref_cache[1][scan8[13] + 1] = PART_NOT_AVAILABLE;
1123 return -1; // free_tables will clean up for us
1126 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
1128 static av_cold void common_init(H264Context *h)
1130 MpegEncContext *const s = &h->s;
1132 s->width = s->avctx->width;
1133 s->height = s->avctx->height;
1134 s->codec_id = s->avctx->codec->id;
1136 s->avctx->bits_per_raw_sample = 8;
1137 h->cur_chroma_format_idc = 1;
1139 ff_h264dsp_init(&h->h264dsp,
1140 s->avctx->bits_per_raw_sample, h->cur_chroma_format_idc);
1141 ff_h264_pred_init(&h->hpc, s->codec_id,
1142 s->avctx->bits_per_raw_sample, h->cur_chroma_format_idc);
1144 h->dequant_coeff_pps = -1;
1145 s->unrestricted_mv = 1;
1147 s->dsp.dct_bits = 16;
1148 /* needed so that IDCT permutation is known early */
1149 ff_dsputil_init(&s->dsp, s->avctx);
1151 memset(h->pps.scaling_matrix4, 16, 6 * 16 * sizeof(uint8_t));
1152 memset(h->pps.scaling_matrix8, 16, 2 * 64 * sizeof(uint8_t));
1155 int ff_h264_decode_extradata(H264Context *h, const uint8_t *buf, int size)
1157 AVCodecContext *avctx = h->s.avctx;
1159 if (!buf || size <= 0)
1163 int i, cnt, nalsize;
1164 const unsigned char *p = buf;
1169 av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
1172 /* sps and pps in the avcC always have length coded with 2 bytes,
1173 * so put a fake nal_length_size = 2 while parsing them */
1174 h->nal_length_size = 2;
1175 // Decode sps from avcC
1176 cnt = *(p + 5) & 0x1f; // Number of sps
1178 for (i = 0; i < cnt; i++) {
1179 nalsize = AV_RB16(p) + 2;
1180 if(nalsize > size - (p-buf))
1182 if (decode_nal_units(h, p, nalsize) < 0) {
1183 av_log(avctx, AV_LOG_ERROR,
1184 "Decoding sps %d from avcC failed\n", i);
1189 // Decode pps from avcC
1190 cnt = *(p++); // Number of pps
1191 for (i = 0; i < cnt; i++) {
1192 nalsize = AV_RB16(p) + 2;
1193 if(nalsize > size - (p-buf))
1195 if (decode_nal_units(h, p, nalsize) < 0) {
1196 av_log(avctx, AV_LOG_ERROR,
1197 "Decoding pps %d from avcC failed\n", i);
1202 // Now store right nal length size, that will be used to parse all other nals
1203 h->nal_length_size = (buf[4] & 0x03) + 1;
1206 if (decode_nal_units(h, buf, size) < 0)
1212 av_cold int ff_h264_decode_init(AVCodecContext *avctx)
1214 H264Context *h = avctx->priv_data;
1215 MpegEncContext *const s = &h->s;
1218 ff_MPV_decode_defaults(s);
1223 s->out_format = FMT_H264;
1224 s->workaround_bugs = avctx->workaround_bugs;
1227 // s->decode_mb = ff_h263_decode_mb;
1228 s->quarter_sample = 1;
1229 if (!avctx->has_b_frames)
1232 avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1234 ff_h264_decode_init_vlc();
1237 h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1239 h->thread_context[0] = h;
1240 h->outputed_poc = h->next_outputed_poc = INT_MIN;
1241 for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
1242 h->last_pocs[i] = INT_MIN;
1243 h->prev_poc_msb = 1 << 16;
1244 h->prev_frame_num = -1;
1246 ff_h264_reset_sei(h);
1247 if (avctx->codec_id == CODEC_ID_H264) {
1248 if (avctx->ticks_per_frame == 1)
1249 s->avctx->time_base.den *= 2;
1250 avctx->ticks_per_frame = 2;
1253 if (avctx->extradata_size > 0 && avctx->extradata &&
1254 ff_h264_decode_extradata(h, avctx->extradata, avctx->extradata_size) < 0)
1257 if (h->sps.bitstream_restriction_flag &&
1258 s->avctx->has_b_frames < h->sps.num_reorder_frames) {
1259 s->avctx->has_b_frames = h->sps.num_reorder_frames;
1266 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b) + (size))))
1268 static void copy_picture_range(Picture **to, Picture **from, int count,
1269 MpegEncContext *new_base,
1270 MpegEncContext *old_base)
1274 for (i = 0; i < count; i++) {
1275 assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1276 IN_RANGE(from[i], old_base->picture,
1277 sizeof(Picture) * old_base->picture_count) ||
1279 to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1283 static void copy_parameter_set(void **to, void **from, int count, int size)
1287 for (i = 0; i < count; i++) {
1288 if (to[i] && !from[i])
1290 else if (from[i] && !to[i])
1291 to[i] = av_malloc(size);
1294 memcpy(to[i], from[i], size);
1298 static int decode_init_thread_copy(AVCodecContext *avctx)
1300 H264Context *h = avctx->priv_data;
1302 if (!avctx->internal->is_copy)
1304 memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1305 memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1310 #define copy_fields(to, from, start_field, end_field) \
1311 memcpy(&to->start_field, &from->start_field, \
1312 (char *)&to->end_field - (char *)&to->start_field)
1314 static int decode_update_thread_context(AVCodecContext *dst,
1315 const AVCodecContext *src)
1317 H264Context *h = dst->priv_data, *h1 = src->priv_data;
1318 MpegEncContext *const s = &h->s, *const s1 = &h1->s;
1319 int inited = s->context_initialized, err;
1325 err = ff_mpeg_update_thread_context(dst, src);
1329 // FIXME handle width/height changing
1331 for (i = 0; i < MAX_SPS_COUNT; i++)
1332 av_freep(h->sps_buffers + i);
1334 for (i = 0; i < MAX_PPS_COUNT; i++)
1335 av_freep(h->pps_buffers + i);
1337 // copy all fields after MpegEnc
1338 memcpy(&h->s + 1, &h1->s + 1,
1339 sizeof(H264Context) - sizeof(MpegEncContext));
1340 memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1341 memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1343 if (s1->context_initialized) {
1344 if (ff_h264_alloc_tables(h) < 0) {
1345 av_log(dst, AV_LOG_ERROR, "Could not allocate memory for h264\n");
1346 return AVERROR(ENOMEM);
1350 /* frame_start may not be called for the next thread (if it's decoding
1351 * a bottom field) so this has to be allocated here */
1352 h->s.obmc_scratchpad = av_malloc(16 * 6 * s->linesize);
1355 for (i = 0; i < 2; i++) {
1356 h->rbsp_buffer[i] = NULL;
1357 h->rbsp_buffer_size[i] = 0;
1360 h->thread_context[0] = h;
1362 s->dsp.clear_blocks(h->mb);
1363 s->dsp.clear_blocks(h->mb + (24 * 16 << h->pixel_shift));
1366 // extradata/NAL handling
1367 h->is_avc = h1->is_avc;
1370 copy_parameter_set((void **)h->sps_buffers, (void **)h1->sps_buffers,
1371 MAX_SPS_COUNT, sizeof(SPS));
1373 copy_parameter_set((void **)h->pps_buffers, (void **)h1->pps_buffers,
1374 MAX_PPS_COUNT, sizeof(PPS));
1377 // Dequantization matrices
1378 // FIXME these are big - can they be only copied when PPS changes?
1379 copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1381 for (i = 0; i < 6; i++)
1382 h->dequant4_coeff[i] = h->dequant4_buffer[0] +
1383 (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1385 for (i = 0; i < 6; i++)
1386 h->dequant8_coeff[i] = h->dequant8_buffer[0] +
1387 (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1389 h->dequant_coeff_pps = h1->dequant_coeff_pps;
1392 copy_fields(h, h1, poc_lsb, redundant_pic_count);
1395 copy_fields(h, h1, ref_count, list_count);
1396 copy_fields(h, h1, ref_list, intra_gb);
1397 copy_fields(h, h1, short_ref, cabac_init_idc);
1399 copy_picture_range(h->short_ref, h1->short_ref, 32, s, s1);
1400 copy_picture_range(h->long_ref, h1->long_ref, 32, s, s1);
1401 copy_picture_range(h->delayed_pic, h1->delayed_pic,
1402 MAX_DELAYED_PIC_COUNT + 2, s, s1);
1404 h->last_slice_type = h1->last_slice_type;
1407 if (!s->current_picture_ptr)
1411 err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1412 h->prev_poc_msb = h->poc_msb;
1413 h->prev_poc_lsb = h->poc_lsb;
1415 h->prev_frame_num_offset = h->frame_num_offset;
1416 h->prev_frame_num = h->frame_num;
1417 h->outputed_poc = h->next_outputed_poc;
1422 int ff_h264_frame_start(H264Context *h)
1424 MpegEncContext *const s = &h->s;
1426 const int pixel_shift = h->pixel_shift;
1428 if (ff_MPV_frame_start(s, s->avctx) < 0)
1430 ff_er_frame_start(s);
1432 * ff_MPV_frame_start uses pict_type to derive key_frame.
1433 * This is incorrect for H.264; IDR markings must be used.
1434 * Zero here; IDR markings per slice in frame or fields are ORed in later.
1435 * See decode_nal_units().
1437 s->current_picture_ptr->f.key_frame = 0;
1438 s->current_picture_ptr->sync = 0;
1439 s->current_picture_ptr->mmco_reset = 0;
1441 assert(s->linesize && s->uvlinesize);
1443 for (i = 0; i < 16; i++) {
1444 h->block_offset[i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * s->linesize * ((scan8[i] - scan8[0]) >> 3);
1445 h->block_offset[48 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * s->linesize * ((scan8[i] - scan8[0]) >> 3);
1447 for (i = 0; i < 16; i++) {
1448 h->block_offset[16 + i] =
1449 h->block_offset[32 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * s->uvlinesize * ((scan8[i] - scan8[0]) >> 3);
1450 h->block_offset[48 + 16 + i] =
1451 h->block_offset[48 + 32 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * s->uvlinesize * ((scan8[i] - scan8[0]) >> 3);
1454 /* can't be in alloc_tables because linesize isn't known there.
1455 * FIXME: redo bipred weight to not require extra buffer? */
1456 for (i = 0; i < s->slice_context_count; i++)
1457 if (h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1458 h->thread_context[i]->s.obmc_scratchpad = av_malloc(16 * 6 * s->linesize);
1460 /* Some macroblocks can be accessed before they're available in case
1461 * of lost slices, MBAFF or threading. */
1462 memset(h->slice_table, -1,
1463 (s->mb_height * s->mb_stride - 1) * sizeof(*h->slice_table));
1465 // s->decode = (s->flags & CODEC_FLAG_PSNR) || !s->encoding ||
1466 // s->current_picture.f.reference /* || h->contains_intra */ || 1;
1468 /* We mark the current picture as non-reference after allocating it, so
1469 * that if we break out due to an error it can be released automatically
1470 * in the next ff_MPV_frame_start().
1471 * SVQ3 as well as most other codecs have only last/next/current and thus
1472 * get released even with set reference, besides SVQ3 and others do not
1473 * mark frames as reference later "naturally". */
1474 if (s->codec_id != CODEC_ID_SVQ3)
1475 s->current_picture_ptr->f.reference = 0;
1477 s->current_picture_ptr->field_poc[0] =
1478 s->current_picture_ptr->field_poc[1] = INT_MAX;
1480 h->next_output_pic = NULL;
1482 assert(s->current_picture_ptr->long_ref == 0);
1488 * Run setup operations that must be run after slice header decoding.
1489 * This includes finding the next displayed frame.
1491 * @param h h264 master context
1492 * @param setup_finished enough NALs have been read that we can call
1493 * ff_thread_finish_setup()
1495 static void decode_postinit(H264Context *h, int setup_finished)
1497 MpegEncContext *const s = &h->s;
1498 Picture *out = s->current_picture_ptr;
1499 Picture *cur = s->current_picture_ptr;
1500 int i, pics, out_of_order, out_idx;
1502 s->current_picture_ptr->f.qscale_type = FF_QSCALE_TYPE_H264;
1503 s->current_picture_ptr->f.pict_type = s->pict_type;
1505 if (h->next_output_pic)
1508 if (cur->field_poc[0] == INT_MAX || cur->field_poc[1] == INT_MAX) {
1509 /* FIXME: if we have two PAFF fields in one packet, we can't start
1510 * the next thread here. If we have one field per packet, we can.
1511 * The check in decode_nal_units() is not good enough to find this
1512 * yet, so we assume the worst for now. */
1513 // if (setup_finished)
1514 // ff_thread_finish_setup(s->avctx);
1518 cur->f.interlaced_frame = 0;
1519 cur->f.repeat_pict = 0;
1521 /* Signal interlacing information externally. */
1522 /* Prioritize picture timing SEI information over used
1523 * decoding process if it exists. */
1525 if (h->sps.pic_struct_present_flag) {
1526 switch (h->sei_pic_struct) {
1527 case SEI_PIC_STRUCT_FRAME:
1529 case SEI_PIC_STRUCT_TOP_FIELD:
1530 case SEI_PIC_STRUCT_BOTTOM_FIELD:
1531 cur->f.interlaced_frame = 1;
1533 case SEI_PIC_STRUCT_TOP_BOTTOM:
1534 case SEI_PIC_STRUCT_BOTTOM_TOP:
1535 if (FIELD_OR_MBAFF_PICTURE)
1536 cur->f.interlaced_frame = 1;
1538 // try to flag soft telecine progressive
1539 cur->f.interlaced_frame = h->prev_interlaced_frame;
1541 case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1542 case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1543 /* Signal the possibility of telecined film externally
1544 * (pic_struct 5,6). From these hints, let the applications
1545 * decide if they apply deinterlacing. */
1546 cur->f.repeat_pict = 1;
1548 case SEI_PIC_STRUCT_FRAME_DOUBLING:
1549 // Force progressive here, doubling interlaced frame is a bad idea.
1550 cur->f.repeat_pict = 2;
1552 case SEI_PIC_STRUCT_FRAME_TRIPLING:
1553 cur->f.repeat_pict = 4;
1557 if ((h->sei_ct_type & 3) &&
1558 h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1559 cur->f.interlaced_frame = (h->sei_ct_type & (1 << 1)) != 0;
1561 /* Derive interlacing flag from used decoding process. */
1562 cur->f.interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1564 h->prev_interlaced_frame = cur->f.interlaced_frame;
1566 if (cur->field_poc[0] != cur->field_poc[1]) {
1567 /* Derive top_field_first from field pocs. */
1568 cur->f.top_field_first = cur->field_poc[0] < cur->field_poc[1];
1570 if (cur->f.interlaced_frame || h->sps.pic_struct_present_flag) {
1571 /* Use picture timing SEI information. Even if it is a
1572 * information of a past frame, better than nothing. */
1573 if (h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM ||
1574 h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1575 cur->f.top_field_first = 1;
1577 cur->f.top_field_first = 0;
1579 /* Most likely progressive */
1580 cur->f.top_field_first = 0;
1584 cur->mmco_reset = h->mmco_reset;
1586 // FIXME do something with unavailable reference frames
1588 /* Sort B-frames into display order */
1590 if (h->sps.bitstream_restriction_flag &&
1591 s->avctx->has_b_frames < h->sps.num_reorder_frames) {
1592 s->avctx->has_b_frames = h->sps.num_reorder_frames;
1596 if (s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT &&
1597 !h->sps.bitstream_restriction_flag) {
1598 s->avctx->has_b_frames = MAX_DELAYED_PIC_COUNT - 1;
1602 for (i = 0; 1; i++) {
1603 if(i == MAX_DELAYED_PIC_COUNT || cur->poc < h->last_pocs[i]){
1605 h->last_pocs[i-1] = cur->poc;
1608 h->last_pocs[i-1]= h->last_pocs[i];
1611 out_of_order = MAX_DELAYED_PIC_COUNT - i;
1612 if( cur->f.pict_type == AV_PICTURE_TYPE_B
1613 || (h->last_pocs[MAX_DELAYED_PIC_COUNT-2] > INT_MIN && h->last_pocs[MAX_DELAYED_PIC_COUNT-1] - h->last_pocs[MAX_DELAYED_PIC_COUNT-2] > 2))
1614 out_of_order = FFMAX(out_of_order, 1);
1615 if(s->avctx->has_b_frames < out_of_order && !h->sps.bitstream_restriction_flag){
1616 av_log(s->avctx, AV_LOG_WARNING, "Increasing reorder buffer to %d\n", out_of_order);
1617 s->avctx->has_b_frames = out_of_order;
1622 while (h->delayed_pic[pics])
1625 av_assert0(pics <= MAX_DELAYED_PIC_COUNT);
1627 h->delayed_pic[pics++] = cur;
1628 if (cur->f.reference == 0)
1629 cur->f.reference = DELAYED_PIC_REF;
1631 out = h->delayed_pic[0];
1633 for (i = 1; h->delayed_pic[i] &&
1634 !h->delayed_pic[i]->f.key_frame &&
1635 !h->delayed_pic[i]->mmco_reset;
1637 if (h->delayed_pic[i]->poc < out->poc) {
1638 out = h->delayed_pic[i];
1641 if (s->avctx->has_b_frames == 0 &&
1642 (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset))
1643 h->next_outputed_poc = INT_MIN;
1644 out_of_order = out->poc < h->next_outputed_poc;
1646 if (out_of_order || pics > s->avctx->has_b_frames) {
1647 out->f.reference &= ~DELAYED_PIC_REF;
1648 // for frame threading, the owner must be the second field's thread or
1649 // else the first thread can release the picture and reuse it unsafely
1651 for (i = out_idx; h->delayed_pic[i]; i++)
1652 h->delayed_pic[i] = h->delayed_pic[i + 1];
1654 if (!out_of_order && pics > s->avctx->has_b_frames) {
1655 h->next_output_pic = out;
1656 if (out_idx == 0 && h->delayed_pic[0] && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset)) {
1657 h->next_outputed_poc = INT_MIN;
1659 h->next_outputed_poc = out->poc;
1661 av_log(s->avctx, AV_LOG_DEBUG, "no picture %s\n", out_of_order ? "ooo" : "");
1664 if (h->next_output_pic && h->next_output_pic->sync) {
1669 ff_thread_finish_setup(s->avctx);
1672 static av_always_inline void backup_mb_border(H264Context *h, uint8_t *src_y,
1673 uint8_t *src_cb, uint8_t *src_cr,
1674 int linesize, int uvlinesize,
1677 MpegEncContext *const s = &h->s;
1678 uint8_t *top_border;
1680 const int pixel_shift = h->pixel_shift;
1681 int chroma444 = CHROMA444;
1682 int chroma422 = CHROMA422;
1685 src_cb -= uvlinesize;
1686 src_cr -= uvlinesize;
1688 if (!simple && FRAME_MBAFF) {
1691 top_border = h->top_borders[0][s->mb_x];
1692 AV_COPY128(top_border, src_y + 15 * linesize);
1694 AV_COPY128(top_border + 16, src_y + 15 * linesize + 16);
1695 if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1698 AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
1699 AV_COPY128(top_border + 48, src_cb + 15 * uvlinesize + 16);
1700 AV_COPY128(top_border + 64, src_cr + 15 * uvlinesize);
1701 AV_COPY128(top_border + 80, src_cr + 15 * uvlinesize + 16);
1703 AV_COPY128(top_border + 16, src_cb + 15 * uvlinesize);
1704 AV_COPY128(top_border + 32, src_cr + 15 * uvlinesize);
1706 } else if (chroma422) {
1708 AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
1709 AV_COPY128(top_border + 48, src_cr + 15 * uvlinesize);
1711 AV_COPY64(top_border + 16, src_cb + 15 * uvlinesize);
1712 AV_COPY64(top_border + 24, src_cr + 15 * uvlinesize);
1716 AV_COPY128(top_border + 32, src_cb + 7 * uvlinesize);
1717 AV_COPY128(top_border + 48, src_cr + 7 * uvlinesize);
1719 AV_COPY64(top_border + 16, src_cb + 7 * uvlinesize);
1720 AV_COPY64(top_border + 24, src_cr + 7 * uvlinesize);
1725 } else if (MB_MBAFF) {
1731 top_border = h->top_borders[top_idx][s->mb_x];
1732 /* There are two lines saved, the line above the the top macroblock
1733 * of a pair, and the line above the bottom macroblock. */
1734 AV_COPY128(top_border, src_y + 16 * linesize);
1736 AV_COPY128(top_border + 16, src_y + 16 * linesize + 16);
1738 if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1741 AV_COPY128(top_border + 32, src_cb + 16 * linesize);
1742 AV_COPY128(top_border + 48, src_cb + 16 * linesize + 16);
1743 AV_COPY128(top_border + 64, src_cr + 16 * linesize);
1744 AV_COPY128(top_border + 80, src_cr + 16 * linesize + 16);
1746 AV_COPY128(top_border + 16, src_cb + 16 * linesize);
1747 AV_COPY128(top_border + 32, src_cr + 16 * linesize);
1749 } else if (chroma422) {
1751 AV_COPY128(top_border + 32, src_cb + 16 * uvlinesize);
1752 AV_COPY128(top_border + 48, src_cr + 16 * uvlinesize);
1754 AV_COPY64(top_border + 16, src_cb + 16 * uvlinesize);
1755 AV_COPY64(top_border + 24, src_cr + 16 * uvlinesize);
1759 AV_COPY128(top_border + 32, src_cb + 8 * uvlinesize);
1760 AV_COPY128(top_border + 48, src_cr + 8 * uvlinesize);
1762 AV_COPY64(top_border + 16, src_cb + 8 * uvlinesize);
1763 AV_COPY64(top_border + 24, src_cr + 8 * uvlinesize);
1769 static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1770 uint8_t *src_cb, uint8_t *src_cr,
1771 int linesize, int uvlinesize,
1772 int xchg, int chroma444,
1773 int simple, int pixel_shift)
1775 MpegEncContext *const s = &h->s;
1776 int deblock_topleft;
1779 uint8_t *top_border_m1;
1780 uint8_t *top_border;
1782 if (!simple && FRAME_MBAFF) {
1787 top_idx = MB_MBAFF ? 0 : 1;
1791 if (h->deblocking_filter == 2) {
1792 deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1793 deblock_top = h->top_type;
1795 deblock_topleft = (s->mb_x > 0);
1796 deblock_top = (s->mb_y > !!MB_FIELD);
1799 src_y -= linesize + 1 + pixel_shift;
1800 src_cb -= uvlinesize + 1 + pixel_shift;
1801 src_cr -= uvlinesize + 1 + pixel_shift;
1803 top_border_m1 = h->top_borders[top_idx][s->mb_x - 1];
1804 top_border = h->top_borders[top_idx][s->mb_x];
1806 #define XCHG(a, b, xchg) \
1807 if (pixel_shift) { \
1809 AV_SWAP64(b + 0, a + 0); \
1810 AV_SWAP64(b + 8, a + 8); \
1820 if (deblock_topleft) {
1821 XCHG(top_border_m1 + (8 << pixel_shift),
1822 src_y - (7 << pixel_shift), 1);
1824 XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1825 XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1826 if (s->mb_x + 1 < s->mb_width) {
1827 XCHG(h->top_borders[top_idx][s->mb_x + 1],
1828 src_y + (17 << pixel_shift), 1);
1831 if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1833 if (deblock_topleft) {
1834 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1835 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1837 XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1838 XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1839 XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1840 XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1841 if (s->mb_x + 1 < s->mb_width) {
1842 XCHG(h->top_borders[top_idx][s->mb_x + 1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1843 XCHG(h->top_borders[top_idx][s->mb_x + 1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1847 if (deblock_topleft) {
1848 XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1849 XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1851 XCHG(top_border + (16 << pixel_shift), src_cb + 1 + pixel_shift, 1);
1852 XCHG(top_border + (24 << pixel_shift), src_cr + 1 + pixel_shift, 1);
1858 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth,
1861 if (high_bit_depth) {
1862 return AV_RN32A(((int32_t *)mb) + index);
1864 return AV_RN16A(mb + index);
1867 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth,
1868 int index, int value)
1870 if (high_bit_depth) {
1871 AV_WN32A(((int32_t *)mb) + index, value);
1873 AV_WN16A(mb + index, value);
1876 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h,
1877 int mb_type, int is_h264,
1879 int transform_bypass,
1883 uint8_t *dest_y, int p)
1885 MpegEncContext *const s = &h->s;
1886 void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1887 void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1889 int qscale = p == 0 ? s->qscale : h->chroma_qp[p - 1];
1890 block_offset += 16 * p;
1891 if (IS_INTRA4x4(mb_type)) {
1892 if (simple || !s->encoding) {
1893 if (IS_8x8DCT(mb_type)) {
1894 if (transform_bypass) {
1896 idct_add = s->dsp.add_pixels8;
1898 idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1899 idct_add = h->h264dsp.h264_idct8_add;
1901 for (i = 0; i < 16; i += 4) {
1902 uint8_t *const ptr = dest_y + block_offset[i];
1903 const int dir = h->intra4x4_pred_mode_cache[scan8[i]];
1904 if (transform_bypass && h->sps.profile_idc == 244 && dir <= 1) {
1905 h->hpc.pred8x8l_add[dir](ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1907 const int nnz = h->non_zero_count_cache[scan8[i + p * 16]];
1908 h->hpc.pred8x8l[dir](ptr, (h->topleft_samples_available << i) & 0x8000,
1909 (h->topright_samples_available << i) & 0x4000, linesize);
1911 if (nnz == 1 && dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
1912 idct_dc_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1914 idct_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1919 if (transform_bypass) {
1921 idct_add = s->dsp.add_pixels4;
1923 idct_dc_add = h->h264dsp.h264_idct_dc_add;
1924 idct_add = h->h264dsp.h264_idct_add;
1926 for (i = 0; i < 16; i++) {
1927 uint8_t *const ptr = dest_y + block_offset[i];
1928 const int dir = h->intra4x4_pred_mode_cache[scan8[i]];
1930 if (transform_bypass && h->sps.profile_idc == 244 && dir <= 1) {
1931 h->hpc.pred4x4_add[dir](ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1936 if (dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED) {
1937 const int topright_avail = (h->topright_samples_available << i) & 0x8000;
1938 assert(s->mb_y || linesize <= block_offset[i]);
1939 if (!topright_avail) {
1941 tr_high = ((uint16_t *)ptr)[3 - linesize / 2] * 0x0001000100010001ULL;
1942 topright = (uint8_t *)&tr_high;
1944 tr = ptr[3 - linesize] * 0x01010101u;
1945 topright = (uint8_t *)&tr;
1948 topright = ptr + (4 << pixel_shift) - linesize;
1952 h->hpc.pred4x4[dir](ptr, topright, linesize);
1953 nnz = h->non_zero_count_cache[scan8[i + p * 16]];
1956 if (nnz == 1 && dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
1957 idct_dc_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1959 idct_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1960 } else if (CONFIG_SVQ3_DECODER)
1961 ff_svq3_add_idct_c(ptr, h->mb + i * 16 + p * 256, linesize, qscale, 0);
1968 h->hpc.pred16x16[h->intra16x16_pred_mode](dest_y, linesize);
1970 if (h->non_zero_count_cache[scan8[LUMA_DC_BLOCK_INDEX + p]]) {
1971 if (!transform_bypass)
1972 h->h264dsp.h264_luma_dc_dequant_idct(h->mb + (p * 256 << pixel_shift),
1974 h->dequant4_coeff[p][qscale][0]);
1976 static const uint8_t dc_mapping[16] = {
1977 0 * 16, 1 * 16, 4 * 16, 5 * 16,
1978 2 * 16, 3 * 16, 6 * 16, 7 * 16,
1979 8 * 16, 9 * 16, 12 * 16, 13 * 16,
1980 10 * 16, 11 * 16, 14 * 16, 15 * 16 };
1981 for (i = 0; i < 16; i++)
1982 dctcoef_set(h->mb + (p * 256 << pixel_shift),
1983 pixel_shift, dc_mapping[i],
1984 dctcoef_get(h->mb_luma_dc[p],
1988 } else if (CONFIG_SVQ3_DECODER)
1989 ff_svq3_luma_dc_dequant_idct_c(h->mb + p * 256,
1990 h->mb_luma_dc[p], qscale);
1994 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type,
1995 int is_h264, int simple,
1996 int transform_bypass,
2000 uint8_t *dest_y, int p)
2002 MpegEncContext *const s = &h->s;
2003 void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2005 block_offset += 16 * p;
2006 if (!IS_INTRA4x4(mb_type)) {
2008 if (IS_INTRA16x16(mb_type)) {
2009 if (transform_bypass) {
2010 if (h->sps.profile_idc == 244 &&
2011 (h->intra16x16_pred_mode == VERT_PRED8x8 ||
2012 h->intra16x16_pred_mode == HOR_PRED8x8)) {
2013 h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset,
2014 h->mb + (p * 256 << pixel_shift),
2017 for (i = 0; i < 16; i++)
2018 if (h->non_zero_count_cache[scan8[i + p * 16]] ||
2019 dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
2020 s->dsp.add_pixels4(dest_y + block_offset[i],
2021 h->mb + (i * 16 + p * 256 << pixel_shift),
2025 h->h264dsp.h264_idct_add16intra(dest_y, block_offset,
2026 h->mb + (p * 256 << pixel_shift),
2028 h->non_zero_count_cache + p * 5 * 8);
2030 } else if (h->cbp & 15) {
2031 if (transform_bypass) {
2032 const int di = IS_8x8DCT(mb_type) ? 4 : 1;
2033 idct_add = IS_8x8DCT(mb_type) ? s->dsp.add_pixels8
2034 : s->dsp.add_pixels4;
2035 for (i = 0; i < 16; i += di)
2036 if (h->non_zero_count_cache[scan8[i + p * 16]])
2037 idct_add(dest_y + block_offset[i],
2038 h->mb + (i * 16 + p * 256 << pixel_shift),
2041 if (IS_8x8DCT(mb_type))
2042 h->h264dsp.h264_idct8_add4(dest_y, block_offset,
2043 h->mb + (p * 256 << pixel_shift),
2045 h->non_zero_count_cache + p * 5 * 8);
2047 h->h264dsp.h264_idct_add16(dest_y, block_offset,
2048 h->mb + (p * 256 << pixel_shift),
2050 h->non_zero_count_cache + p * 5 * 8);
2053 } else if (CONFIG_SVQ3_DECODER) {
2054 for (i = 0; i < 16; i++)
2055 if (h->non_zero_count_cache[scan8[i + p * 16]] || h->mb[i * 16 + p * 256]) {
2056 // FIXME benchmark weird rule, & below
2057 uint8_t *const ptr = dest_y + block_offset[i];
2058 ff_svq3_add_idct_c(ptr, h->mb + i * 16 + p * 256, linesize,
2059 s->qscale, IS_INTRA(mb_type) ? 1 : 0);
2065 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple,
2068 MpegEncContext *const s = &h->s;
2069 const int mb_x = s->mb_x;
2070 const int mb_y = s->mb_y;
2071 const int mb_xy = h->mb_xy;
2072 const int mb_type = s->current_picture.f.mb_type[mb_xy];
2073 uint8_t *dest_y, *dest_cb, *dest_cr;
2074 int linesize, uvlinesize /*dct_offset*/;
2076 int *block_offset = &h->block_offset[0];
2077 const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
2078 /* is_h264 should always be true if SVQ3 is disabled. */
2079 const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
2080 void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2081 const int block_h = 16 >> s->chroma_y_shift;
2082 const int chroma422 = CHROMA422;
2084 dest_y = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
2085 dest_cb = s->current_picture.f.data[1] + (mb_x << pixel_shift) * 8 + mb_y * s->uvlinesize * block_h;
2086 dest_cr = s->current_picture.f.data[2] + (mb_x << pixel_shift) * 8 + mb_y * s->uvlinesize * block_h;
2088 s->dsp.prefetch(dest_y + (s->mb_x & 3) * 4 * s->linesize + (64 << pixel_shift), s->linesize, 4);
2089 s->dsp.prefetch(dest_cb + (s->mb_x & 7) * s->uvlinesize + (64 << pixel_shift), dest_cr - dest_cb, 2);
2091 h->list_counts[mb_xy] = h->list_count;
2093 if (!simple && MB_FIELD) {
2094 linesize = h->mb_linesize = s->linesize * 2;
2095 uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
2096 block_offset = &h->block_offset[48];
2097 if (mb_y & 1) { // FIXME move out of this function?
2098 dest_y -= s->linesize * 15;
2099 dest_cb -= s->uvlinesize * (block_h - 1);
2100 dest_cr -= s->uvlinesize * (block_h - 1);
2104 for (list = 0; list < h->list_count; list++) {
2105 if (!USES_LIST(mb_type, list))
2107 if (IS_16X16(mb_type)) {
2108 int8_t *ref = &h->ref_cache[list][scan8[0]];
2109 fill_rectangle(ref, 4, 4, 8, (16 + *ref) ^ (s->mb_y & 1), 1);
2111 for (i = 0; i < 16; i += 4) {
2112 int ref = h->ref_cache[list][scan8[i]];
2114 fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2,
2115 8, (16 + ref) ^ (s->mb_y & 1), 1);
2121 linesize = h->mb_linesize = s->linesize;
2122 uvlinesize = h->mb_uvlinesize = s->uvlinesize;
2123 // dct_offset = s->linesize * 16;
2126 if (!simple && IS_INTRA_PCM(mb_type)) {
2127 const int bit_depth = h->sps.bit_depth_luma;
2131 init_get_bits(&gb, (uint8_t *)h->mb,
2132 ff_h264_mb_sizes[h->sps.chroma_format_idc] * bit_depth);
2134 for (i = 0; i < 16; i++) {
2135 uint16_t *tmp_y = (uint16_t *)(dest_y + i * linesize);
2136 for (j = 0; j < 16; j++)
2137 tmp_y[j] = get_bits(&gb, bit_depth);
2139 if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
2140 if (!h->sps.chroma_format_idc) {
2141 for (i = 0; i < block_h; i++) {
2142 uint16_t *tmp_cb = (uint16_t *)(dest_cb + i * uvlinesize);
2143 uint16_t *tmp_cr = (uint16_t *)(dest_cr + i * uvlinesize);
2144 for (j = 0; j < 8; j++) {
2145 tmp_cb[j] = tmp_cr[j] = 1 << (bit_depth - 1);
2149 for (i = 0; i < block_h; i++) {
2150 uint16_t *tmp_cb = (uint16_t *)(dest_cb + i * uvlinesize);
2151 for (j = 0; j < 8; j++)
2152 tmp_cb[j] = get_bits(&gb, bit_depth);
2154 for (i = 0; i < block_h; i++) {
2155 uint16_t *tmp_cr = (uint16_t *)(dest_cr + i * uvlinesize);
2156 for (j = 0; j < 8; j++)
2157 tmp_cr[j] = get_bits(&gb, bit_depth);
2162 for (i = 0; i < 16; i++)
2163 memcpy(dest_y + i * linesize, (uint8_t *)h->mb + i * 16, 16);
2164 if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
2165 if (!h->sps.chroma_format_idc) {
2166 for (i = 0; i < 8; i++) {
2167 memset(dest_cb + i*uvlinesize, 1 << (bit_depth - 1), 8);
2168 memset(dest_cr + i*uvlinesize, 1 << (bit_depth - 1), 8);
2171 uint8_t *src_cb = (uint8_t *)h->mb + 256;
2172 uint8_t *src_cr = (uint8_t *)h->mb + 256 + block_h * 8;
2173 for (i = 0; i < block_h; i++) {
2174 memcpy(dest_cb + i * uvlinesize, src_cb + i * 8, 8);
2175 memcpy(dest_cr + i * uvlinesize, src_cr + i * 8, 8);
2181 if (IS_INTRA(mb_type)) {
2182 if (h->deblocking_filter)
2183 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize,
2184 uvlinesize, 1, 0, simple, pixel_shift);
2186 if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
2188 h->hpc.pred8x8[h->chroma_pred_mode](dest_cb, uvlinesize);
2189 h->hpc.pred8x8[h->chroma_pred_mode](dest_cr, uvlinesize);
2193 hl_decode_mb_predict_luma(h, mb_type, is_h264, simple,
2194 transform_bypass, pixel_shift,
2195 block_offset, linesize, dest_y, 0);
2197 if (h->deblocking_filter)
2198 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize,
2199 uvlinesize, 0, 0, simple, pixel_shift);
2200 } else if (is_h264) {
2202 hl_motion_422(h, dest_y, dest_cb, dest_cr,
2203 s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2204 s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2205 h->h264dsp.weight_h264_pixels_tab,
2206 h->h264dsp.biweight_h264_pixels_tab,
2209 hl_motion_420(h, dest_y, dest_cb, dest_cr,
2210 s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2211 s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2212 h->h264dsp.weight_h264_pixels_tab,
2213 h->h264dsp.biweight_h264_pixels_tab,
2218 hl_decode_mb_idct_luma(h, mb_type, is_h264, simple, transform_bypass,
2219 pixel_shift, block_offset, linesize, dest_y, 0);
2221 if ((simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) &&
2223 uint8_t *dest[2] = { dest_cb, dest_cr };
2224 if (transform_bypass) {
2225 if (IS_INTRA(mb_type) && h->sps.profile_idc == 244 &&
2226 (h->chroma_pred_mode == VERT_PRED8x8 ||
2227 h->chroma_pred_mode == HOR_PRED8x8)) {
2228 h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0],
2230 h->mb + (16 * 16 * 1 << pixel_shift),
2232 h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1],
2234 h->mb + (16 * 16 * 2 << pixel_shift),
2237 idct_add = s->dsp.add_pixels4;
2238 for (j = 1; j < 3; j++) {
2239 for (i = j * 16; i < j * 16 + 4; i++)
2240 if (h->non_zero_count_cache[scan8[i]] ||
2241 dctcoef_get(h->mb, pixel_shift, i * 16))
2242 idct_add(dest[j - 1] + block_offset[i],
2243 h->mb + (i * 16 << pixel_shift),
2246 for (i = j * 16 + 4; i < j * 16 + 8; i++)
2247 if (h->non_zero_count_cache[scan8[i + 4]] ||
2248 dctcoef_get(h->mb, pixel_shift, i * 16))
2249 idct_add(dest[j - 1] + block_offset[i + 4],
2250 h->mb + (i * 16 << pixel_shift),
2259 qp[0] = h->chroma_qp[0] + 3;
2260 qp[1] = h->chroma_qp[1] + 3;
2262 qp[0] = h->chroma_qp[0];
2263 qp[1] = h->chroma_qp[1];
2265 if (h->non_zero_count_cache[scan8[CHROMA_DC_BLOCK_INDEX + 0]])
2266 h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16 * 16 * 1 << pixel_shift),
2267 h->dequant4_coeff[IS_INTRA(mb_type) ? 1 : 4][qp[0]][0]);
2268 if (h->non_zero_count_cache[scan8[CHROMA_DC_BLOCK_INDEX + 1]])
2269 h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16 * 16 * 2 << pixel_shift),
2270 h->dequant4_coeff[IS_INTRA(mb_type) ? 2 : 5][qp[1]][0]);
2271 h->h264dsp.h264_idct_add8(dest, block_offset,
2273 h->non_zero_count_cache);
2274 } else if (CONFIG_SVQ3_DECODER) {
2275 h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + 16 * 16 * 1,
2276 h->dequant4_coeff[IS_INTRA(mb_type) ? 1 : 4][h->chroma_qp[0]][0]);
2277 h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + 16 * 16 * 2,
2278 h->dequant4_coeff[IS_INTRA(mb_type) ? 2 : 5][h->chroma_qp[1]][0]);
2279 for (j = 1; j < 3; j++) {
2280 for (i = j * 16; i < j * 16 + 4; i++)
2281 if (h->non_zero_count_cache[scan8[i]] || h->mb[i * 16]) {
2282 uint8_t *const ptr = dest[j - 1] + block_offset[i];
2283 ff_svq3_add_idct_c(ptr, h->mb + i * 16,
2285 ff_h264_chroma_qp[0][s->qscale + 12] - 12, 2);
2292 if (h->cbp || IS_INTRA(mb_type)) {
2293 s->dsp.clear_blocks(h->mb);
2294 s->dsp.clear_blocks(h->mb + (24 * 16 << pixel_shift));
2298 static av_always_inline void hl_decode_mb_444_internal(H264Context *h,
2302 MpegEncContext *const s = &h->s;
2303 const int mb_x = s->mb_x;
2304 const int mb_y = s->mb_y;
2305 const int mb_xy = h->mb_xy;
2306 const int mb_type = s->current_picture.f.mb_type[mb_xy];
2310 int *block_offset = &h->block_offset[0];
2311 const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
2312 const int plane_count = (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) ? 3 : 1;
2314 for (p = 0; p < plane_count; p++) {
2315 dest[p] = s->current_picture.f.data[p] +
2316 ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
2317 s->dsp.prefetch(dest[p] + (s->mb_x & 3) * 4 * s->linesize + (64 << pixel_shift),
2321 h->list_counts[mb_xy] = h->list_count;
2323 if (!simple && MB_FIELD) {
2324 linesize = h->mb_linesize = h->mb_uvlinesize = s->linesize * 2;
2325 block_offset = &h->block_offset[48];
2326 if (mb_y & 1) // FIXME move out of this function?
2327 for (p = 0; p < 3; p++)
2328 dest[p] -= s->linesize * 15;
2331 for (list = 0; list < h->list_count; list++) {
2332 if (!USES_LIST(mb_type, list))
2334 if (IS_16X16(mb_type)) {
2335 int8_t *ref = &h->ref_cache[list][scan8[0]];
2336 fill_rectangle(ref, 4, 4, 8, (16 + *ref) ^ (s->mb_y & 1), 1);
2338 for (i = 0; i < 16; i += 4) {
2339 int ref = h->ref_cache[list][scan8[i]];
2341 fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2,
2342 8, (16 + ref) ^ (s->mb_y & 1), 1);
2348 linesize = h->mb_linesize = h->mb_uvlinesize = s->linesize;
2351 if (!simple && IS_INTRA_PCM(mb_type)) {
2353 const int bit_depth = h->sps.bit_depth_luma;
2355 init_get_bits(&gb, (uint8_t *)h->mb, 768 * bit_depth);
2357 for (p = 0; p < plane_count; p++)
2358 for (i = 0; i < 16; i++) {
2359 uint16_t *tmp = (uint16_t *)(dest[p] + i * linesize);
2360 for (j = 0; j < 16; j++)
2361 tmp[j] = get_bits(&gb, bit_depth);
2364 for (p = 0; p < plane_count; p++)
2365 for (i = 0; i < 16; i++)
2366 memcpy(dest[p] + i * linesize,
2367 (uint8_t *)h->mb + p * 256 + i * 16, 16);
2370 if (IS_INTRA(mb_type)) {
2371 if (h->deblocking_filter)
2372 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize,
2373 linesize, 1, 1, simple, pixel_shift);
2375 for (p = 0; p < plane_count; p++)
2376 hl_decode_mb_predict_luma(h, mb_type, 1, simple,
2377 transform_bypass, pixel_shift,
2378 block_offset, linesize, dest[p], p);
2380 if (h->deblocking_filter)
2381 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize,
2382 linesize, 0, 1, simple, pixel_shift);
2384 hl_motion(h, dest[0], dest[1], dest[2],
2385 s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2386 s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2387 h->h264dsp.weight_h264_pixels_tab,
2388 h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 3);
2391 for (p = 0; p < plane_count; p++)
2392 hl_decode_mb_idct_luma(h, mb_type, 1, simple, transform_bypass,
2393 pixel_shift, block_offset, linesize,
2396 if (h->cbp || IS_INTRA(mb_type)) {
2397 s->dsp.clear_blocks(h->mb);
2398 s->dsp.clear_blocks(h->mb + (24 * 16 << pixel_shift));
2403 * Process a macroblock; this case avoids checks for expensive uncommon cases.
2405 #define hl_decode_mb_simple(sh, bits) \
2406 static void hl_decode_mb_simple_ ## bits(H264Context *h) \
2408 hl_decode_mb_internal(h, 1, sh); \
2411 hl_decode_mb_simple(0, 8)
2412 hl_decode_mb_simple(1, 16)
2415 * Process a macroblock; this handles edge cases, such as interlacing.
2417 static av_noinline void hl_decode_mb_complex(H264Context *h)
2419 hl_decode_mb_internal(h, 0, h->pixel_shift);
2422 static av_noinline void hl_decode_mb_444_complex(H264Context *h)
2424 hl_decode_mb_444_internal(h, 0, h->pixel_shift);
2427 static av_noinline void hl_decode_mb_444_simple(H264Context *h)
2429 hl_decode_mb_444_internal(h, 1, 0);
2432 void ff_h264_hl_decode_mb(H264Context *h)
2434 MpegEncContext *const s = &h->s;
2435 const int mb_xy = h->mb_xy;
2436 const int mb_type = s->current_picture.f.mb_type[mb_xy];
2437 int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2440 if (is_complex || h->pixel_shift)
2441 hl_decode_mb_444_complex(h);
2443 hl_decode_mb_444_simple(h);
2444 } else if (is_complex) {
2445 hl_decode_mb_complex(h);
2446 } else if (h->pixel_shift) {
2447 hl_decode_mb_simple_16(h);
2449 hl_decode_mb_simple_8(h);
2452 static int pred_weight_table(H264Context *h)
2454 MpegEncContext *const s = &h->s;
2456 int luma_def, chroma_def;
2459 h->use_weight_chroma = 0;
2460 h->luma_log2_weight_denom = get_ue_golomb(&s->gb);
2461 if (h->sps.chroma_format_idc)
2462 h->chroma_log2_weight_denom = get_ue_golomb(&s->gb);
2463 luma_def = 1 << h->luma_log2_weight_denom;
2464 chroma_def = 1 << h->chroma_log2_weight_denom;
2466 for (list = 0; list < 2; list++) {
2467 h->luma_weight_flag[list] = 0;
2468 h->chroma_weight_flag[list] = 0;
2469 for (i = 0; i < h->ref_count[list]; i++) {
2470 int luma_weight_flag, chroma_weight_flag;
2472 luma_weight_flag = get_bits1(&s->gb);
2473 if (luma_weight_flag) {
2474 h->luma_weight[i][list][0] = get_se_golomb(&s->gb);
2475 h->luma_weight[i][list][1] = get_se_golomb(&s->gb);
2476 if (h->luma_weight[i][list][0] != luma_def ||
2477 h->luma_weight[i][list][1] != 0) {
2479 h->luma_weight_flag[list] = 1;
2482 h->luma_weight[i][list][0] = luma_def;
2483 h->luma_weight[i][list][1] = 0;
2486 if (h->sps.chroma_format_idc) {
2487 chroma_weight_flag = get_bits1(&s->gb);
2488 if (chroma_weight_flag) {
2490 for (j = 0; j < 2; j++) {
2491 h->chroma_weight[i][list][j][0] = get_se_golomb(&s->gb);
2492 h->chroma_weight[i][list][j][1] = get_se_golomb(&s->gb);
2493 if (h->chroma_weight[i][list][j][0] != chroma_def ||
2494 h->chroma_weight[i][list][j][1] != 0) {
2495 h->use_weight_chroma = 1;
2496 h->chroma_weight_flag[list] = 1;
2501 for (j = 0; j < 2; j++) {
2502 h->chroma_weight[i][list][j][0] = chroma_def;
2503 h->chroma_weight[i][list][j][1] = 0;
2508 if (h->slice_type_nos != AV_PICTURE_TYPE_B)
2511 h->use_weight = h->use_weight || h->use_weight_chroma;
2516 * Initialize implicit_weight table.
2517 * @param field 0/1 initialize the weight for interlaced MBAFF
2518 * -1 initializes the rest
2520 static void implicit_weight_table(H264Context *h, int field)
2522 MpegEncContext *const s = &h->s;
2523 int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2525 for (i = 0; i < 2; i++) {
2526 h->luma_weight_flag[i] = 0;
2527 h->chroma_weight_flag[i] = 0;
2531 if (s->picture_structure == PICT_FRAME) {
2532 cur_poc = s->current_picture_ptr->poc;
2534 cur_poc = s->current_picture_ptr->field_poc[s->picture_structure - 1];
2536 if (h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF &&
2537 h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2 * cur_poc) {
2539 h->use_weight_chroma = 0;
2543 ref_count0 = h->ref_count[0];
2544 ref_count1 = h->ref_count[1];
2546 cur_poc = s->current_picture_ptr->field_poc[field];
2548 ref_count0 = 16 + 2 * h->ref_count[0];
2549 ref_count1 = 16 + 2 * h->ref_count[1];
2553 h->use_weight_chroma = 2;
2554 h->luma_log2_weight_denom = 5;
2555 h->chroma_log2_weight_denom = 5;
2557 for (ref0 = ref_start; ref0 < ref_count0; ref0++) {
2558 int poc0 = h->ref_list[0][ref0].poc;
2559 for (ref1 = ref_start; ref1 < ref_count1; ref1++) {
2561 if (!h->ref_list[0][ref0].long_ref && !h->ref_list[1][ref1].long_ref) {
2562 int poc1 = h->ref_list[1][ref1].poc;
2563 int td = av_clip(poc1 - poc0, -128, 127);
2565 int tb = av_clip(cur_poc - poc0, -128, 127);
2566 int tx = (16384 + (FFABS(td) >> 1)) / td;
2567 int dist_scale_factor = (tb * tx + 32) >> 8;
2568 if (dist_scale_factor >= -64 && dist_scale_factor <= 128)
2569 w = 64 - dist_scale_factor;
2573 h->implicit_weight[ref0][ref1][0] =
2574 h->implicit_weight[ref0][ref1][1] = w;
2576 h->implicit_weight[ref0][ref1][field] = w;
2583 * instantaneous decoder refresh.
2585 static void idr(H264Context *h)
2588 ff_h264_remove_all_refs(h);
2589 h->prev_frame_num = 0;
2590 h->prev_frame_num_offset = 0;
2591 h->prev_poc_msb = 1<<16;
2592 h->prev_poc_lsb = 0;
2593 for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
2594 h->last_pocs[i] = INT_MIN;
2597 /* forget old pics after a seek */
2598 static void flush_dpb(AVCodecContext *avctx)
2600 H264Context *h = avctx->priv_data;
2602 for (i=0; i<=MAX_DELAYED_PIC_COUNT; i++) {
2603 if (h->delayed_pic[i])
2604 h->delayed_pic[i]->f.reference = 0;
2605 h->delayed_pic[i] = NULL;
2607 h->outputed_poc = h->next_outputed_poc = INT_MIN;
2608 h->prev_interlaced_frame = 1;
2610 h->prev_frame_num = -1;
2611 if (h->s.current_picture_ptr)
2612 h->s.current_picture_ptr->f.reference = 0;
2613 h->s.first_field = 0;
2614 ff_h264_reset_sei(h);
2615 ff_mpeg_flush(avctx);
2616 h->recovery_frame= -1;
2620 static int init_poc(H264Context *h)
2622 MpegEncContext *const s = &h->s;
2623 const int max_frame_num = 1 << h->sps.log2_max_frame_num;
2625 Picture *cur = s->current_picture_ptr;
2627 h->frame_num_offset = h->prev_frame_num_offset;
2628 if (h->frame_num < h->prev_frame_num)
2629 h->frame_num_offset += max_frame_num;
2631 if (h->sps.poc_type == 0) {
2632 const int max_poc_lsb = 1 << h->sps.log2_max_poc_lsb;
2634 if (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb / 2)
2635 h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2636 else if (h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb / 2)
2637 h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2639 h->poc_msb = h->prev_poc_msb;
2640 // printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
2642 field_poc[1] = h->poc_msb + h->poc_lsb;
2643 if (s->picture_structure == PICT_FRAME)
2644 field_poc[1] += h->delta_poc_bottom;
2645 } else if (h->sps.poc_type == 1) {
2646 int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2649 if (h->sps.poc_cycle_length != 0)
2650 abs_frame_num = h->frame_num_offset + h->frame_num;
2654 if (h->nal_ref_idc == 0 && abs_frame_num > 0)
2657 expected_delta_per_poc_cycle = 0;
2658 for (i = 0; i < h->sps.poc_cycle_length; i++)
2659 // FIXME integrate during sps parse
2660 expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[i];
2662 if (abs_frame_num > 0) {
2663 int poc_cycle_cnt = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2664 int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2666 expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2667 for (i = 0; i <= frame_num_in_poc_cycle; i++)
2668 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[i];
2672 if (h->nal_ref_idc == 0)
2673 expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2675 field_poc[0] = expectedpoc + h->delta_poc[0];
2676 field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2678 if (s->picture_structure == PICT_FRAME)
2679 field_poc[1] += h->delta_poc[1];
2681 int poc = 2 * (h->frame_num_offset + h->frame_num);
2683 if (!h->nal_ref_idc)
2690 if (s->picture_structure != PICT_BOTTOM_FIELD)
2691 s->current_picture_ptr->field_poc[0] = field_poc[0];
2692 if (s->picture_structure != PICT_TOP_FIELD)
2693 s->current_picture_ptr->field_poc[1] = field_poc[1];
2694 cur->poc = FFMIN(cur->field_poc[0], cur->field_poc[1]);
2700 * initialize scan tables
2702 static void init_scan_tables(H264Context *h)
2705 for (i = 0; i < 16; i++) {
2706 #define T(x) (x >> 2) | ((x << 2) & 0xF)
2707 h->zigzag_scan[i] = T(zigzag_scan[i]);
2708 h->field_scan[i] = T(field_scan[i]);
2711 for (i = 0; i < 64; i++) {
2712 #define T(x) (x >> 3) | ((x & 7) << 3)
2713 h->zigzag_scan8x8[i] = T(ff_zigzag_direct[i]);
2714 h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2715 h->field_scan8x8[i] = T(field_scan8x8[i]);
2716 h->field_scan8x8_cavlc[i] = T(field_scan8x8_cavlc[i]);
2719 if (h->sps.transform_bypass) { // FIXME same ugly
2720 h->zigzag_scan_q0 = zigzag_scan;
2721 h->zigzag_scan8x8_q0 = ff_zigzag_direct;
2722 h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2723 h->field_scan_q0 = field_scan;
2724 h->field_scan8x8_q0 = field_scan8x8;
2725 h->field_scan8x8_cavlc_q0 = field_scan8x8_cavlc;
2727 h->zigzag_scan_q0 = h->zigzag_scan;
2728 h->zigzag_scan8x8_q0 = h->zigzag_scan8x8;
2729 h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2730 h->field_scan_q0 = h->field_scan;
2731 h->field_scan8x8_q0 = h->field_scan8x8;
2732 h->field_scan8x8_cavlc_q0 = h->field_scan8x8_cavlc;
2736 static int field_end(H264Context *h, int in_setup)
2738 MpegEncContext *const s = &h->s;
2739 AVCodecContext *const avctx = s->avctx;
2743 if (!in_setup && !s->dropable)
2744 ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX,
2745 s->picture_structure == PICT_BOTTOM_FIELD);
2747 if (CONFIG_H264_VDPAU_DECODER &&
2748 s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
2749 ff_vdpau_h264_set_reference_frames(s);
2751 if (in_setup || !(avctx->active_thread_type & FF_THREAD_FRAME)) {
2753 err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2754 h->prev_poc_msb = h->poc_msb;
2755 h->prev_poc_lsb = h->poc_lsb;
2757 h->prev_frame_num_offset = h->frame_num_offset;
2758 h->prev_frame_num = h->frame_num;
2759 h->outputed_poc = h->next_outputed_poc;
2762 if (avctx->hwaccel) {
2763 if (avctx->hwaccel->end_frame(avctx) < 0)
2764 av_log(avctx, AV_LOG_ERROR,
2765 "hardware accelerator failed to decode picture\n");
2768 if (CONFIG_H264_VDPAU_DECODER &&
2769 s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
2770 ff_vdpau_h264_picture_complete(s);
2773 * FIXME: Error handling code does not seem to support interlaced
2774 * when slices span multiple rows
2775 * The ff_er_add_slice calls don't work right for bottom
2776 * fields; they cause massive erroneous error concealing
2777 * Error marking covers both fields (top and bottom).
2778 * This causes a mismatched s->error_count
2779 * and a bad error table. Further, the error count goes to
2780 * INT_MAX when called for bottom field, because mb_y is
2781 * past end by one (callers fault) and resync_mb_y != 0
2782 * causes problems for the first MB line, too.
2787 ff_MPV_frame_end(s);
2789 h->current_slice = 0;
2795 * Replicate H264 "master" context to thread contexts.
2797 static void clone_slice(H264Context *dst, H264Context *src)
2799 memcpy(dst->block_offset, src->block_offset, sizeof(dst->block_offset));
2800 dst->s.current_picture_ptr = src->s.current_picture_ptr;
2801 dst->s.current_picture = src->s.current_picture;
2802 dst->s.linesize = src->s.linesize;
2803 dst->s.uvlinesize = src->s.uvlinesize;
2804 dst->s.first_field = src->s.first_field;
2806 dst->prev_poc_msb = src->prev_poc_msb;
2807 dst->prev_poc_lsb = src->prev_poc_lsb;
2808 dst->prev_frame_num_offset = src->prev_frame_num_offset;
2809 dst->prev_frame_num = src->prev_frame_num;
2810 dst->short_ref_count = src->short_ref_count;
2812 memcpy(dst->short_ref, src->short_ref, sizeof(dst->short_ref));
2813 memcpy(dst->long_ref, src->long_ref, sizeof(dst->long_ref));
2814 memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2815 memcpy(dst->ref_list, src->ref_list, sizeof(dst->ref_list));
2817 memcpy(dst->dequant4_coeff, src->dequant4_coeff, sizeof(src->dequant4_coeff));
2818 memcpy(dst->dequant8_coeff, src->dequant8_coeff, sizeof(src->dequant8_coeff));
2822 * Compute profile from profile_idc and constraint_set?_flags.
2826 * @return profile as defined by FF_PROFILE_H264_*
2828 int ff_h264_get_profile(SPS *sps)
2830 int profile = sps->profile_idc;
2832 switch (sps->profile_idc) {
2833 case FF_PROFILE_H264_BASELINE:
2834 // constraint_set1_flag set to 1
2835 profile |= (sps->constraint_set_flags & 1 << 1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2837 case FF_PROFILE_H264_HIGH_10:
2838 case FF_PROFILE_H264_HIGH_422:
2839 case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2840 // constraint_set3_flag set to 1
2841 profile |= (sps->constraint_set_flags & 1 << 3) ? FF_PROFILE_H264_INTRA : 0;
2849 * Decode a slice header.
2850 * This will also call ff_MPV_common_init() and frame_start() as needed.
2852 * @param h h264context
2853 * @param h0 h264 master context (differs from 'h' when doing sliced based
2854 * parallel decoding)
2856 * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2858 static int decode_slice_header(H264Context *h, H264Context *h0)
2860 MpegEncContext *const s = &h->s;
2861 MpegEncContext *const s0 = &h0->s;
2862 unsigned int first_mb_in_slice;
2863 unsigned int pps_id;
2864 int num_ref_idx_active_override_flag;
2865 unsigned int slice_type, tmp, i, j;
2866 int default_ref_list_done = 0;
2867 int last_pic_structure, last_pic_dropable;
2870 /* FIXME: 2tap qpel isn't implemented for high bit depth. */
2871 if ((s->avctx->flags2 & CODEC_FLAG2_FAST) &&
2872 !h->nal_ref_idc && !h->pixel_shift) {
2873 s->me.qpel_put = s->dsp.put_2tap_qpel_pixels_tab;
2874 s->me.qpel_avg = s->dsp.avg_2tap_qpel_pixels_tab;
2876 s->me.qpel_put = s->dsp.put_h264_qpel_pixels_tab;
2877 s->me.qpel_avg = s->dsp.avg_h264_qpel_pixels_tab;
2880 first_mb_in_slice = get_ue_golomb_long(&s->gb);
2882 if (first_mb_in_slice == 0) { // FIXME better field boundary detection
2883 if (h0->current_slice && FIELD_PICTURE) {
2887 h0->current_slice = 0;
2888 if (!s0->first_field) {
2889 if (s->current_picture_ptr && !s->dropable &&
2890 s->current_picture_ptr->owner2 == s) {
2891 ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX,
2892 s->picture_structure == PICT_BOTTOM_FIELD);
2894 s->current_picture_ptr = NULL;
2898 slice_type = get_ue_golomb_31(&s->gb);
2899 if (slice_type > 9) {
2900 av_log(h->s.avctx, AV_LOG_ERROR,
2901 "slice type too large (%d) at %d %d\n",
2902 h->slice_type, s->mb_x, s->mb_y);
2905 if (slice_type > 4) {
2907 h->slice_type_fixed = 1;
2909 h->slice_type_fixed = 0;
2911 slice_type = golomb_to_pict_type[slice_type];
2912 if (slice_type == AV_PICTURE_TYPE_I ||
2913 (h0->current_slice != 0 && slice_type == h0->last_slice_type)) {
2914 default_ref_list_done = 1;
2916 h->slice_type = slice_type;
2917 h->slice_type_nos = slice_type & 3;
2919 // to make a few old functions happy, it's wrong though
2920 s->pict_type = h->slice_type;
2922 pps_id = get_ue_golomb(&s->gb);
2923 if (pps_id >= MAX_PPS_COUNT) {
2924 av_log(h->s.avctx, AV_LOG_ERROR, "pps_id %d out of range\n", pps_id);
2927 if (!h0->pps_buffers[pps_id]) {
2928 av_log(h->s.avctx, AV_LOG_ERROR,
2929 "non-existing PPS %u referenced\n",
2933 h->pps = *h0->pps_buffers[pps_id];
2935 if (!h0->sps_buffers[h->pps.sps_id]) {
2936 av_log(h->s.avctx, AV_LOG_ERROR,
2937 "non-existing SPS %u referenced\n",
2941 h->sps = *h0->sps_buffers[h->pps.sps_id];
2943 s->avctx->profile = ff_h264_get_profile(&h->sps);
2944 s->avctx->level = h->sps.level_idc;
2945 s->avctx->refs = h->sps.ref_frame_count;
2947 must_reinit = (s->context_initialized &&
2948 ( 16*h->sps.mb_width != s->avctx->coded_width
2949 || 16*h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag) != s->avctx->coded_height
2950 || s->avctx->bits_per_raw_sample != h->sps.bit_depth_luma
2951 || h->cur_chroma_format_idc != h->sps.chroma_format_idc
2952 || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio)));
2954 if(must_reinit && (h != h0 || (s->avctx->active_thread_type & FF_THREAD_FRAME))) {
2955 av_log_missing_feature(s->avctx,
2956 "Width/height/bit depth/chroma idc changing with threads is", 0);
2957 return -1; // width / height changed during parallelized decoding
2960 s->mb_width = h->sps.mb_width;
2961 s->mb_height = h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2963 h->b_stride = s->mb_width * 4;
2965 s->chroma_y_shift = h->sps.chroma_format_idc <= 1; // 400 uses yuv420p
2967 s->width = 16 * s->mb_width;
2968 s->height = 16 * s->mb_height;
2973 flush_dpb(s->avctx);
2974 ff_MPV_common_end(s);
2977 if (!s->context_initialized) {
2979 av_log(h->s.avctx, AV_LOG_ERROR,
2980 "Cannot (re-)initialize context during parallel decoding.\n");
2983 avcodec_set_dimensions(s->avctx, s->width, s->height);
2984 s->avctx->width -= (2>>CHROMA444)*FFMIN(h->sps.crop_right, (8<<CHROMA444)-1);
2985 s->avctx->height -= (1<<s->chroma_y_shift)*FFMIN(h->sps.crop_bottom, (16>>s->chroma_y_shift)-1) * (2 - h->sps.frame_mbs_only_flag);
2986 s->avctx->sample_aspect_ratio = h->sps.sar;
2987 av_assert0(s->avctx->sample_aspect_ratio.den);
2989 if (s->avctx->bits_per_raw_sample != h->sps.bit_depth_luma ||
2990 h->cur_chroma_format_idc != h->sps.chroma_format_idc) {
2991 if (h->sps.bit_depth_luma >= 8 && h->sps.bit_depth_luma <= 10 &&
2992 (h->sps.bit_depth_luma != 9 || !CHROMA422)) {
2993 s->avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
2994 h->cur_chroma_format_idc = h->sps.chroma_format_idc;
2995 h->pixel_shift = h->sps.bit_depth_luma > 8;
2997 ff_h264dsp_init(&h->h264dsp, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
2998 ff_h264_pred_init(&h->hpc, s->codec_id, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
2999 s->dsp.dct_bits = h->sps.bit_depth_luma > 8 ? 32 : 16;
3000 ff_dsputil_init(&s->dsp, s->avctx);
3002 av_log(s->avctx, AV_LOG_ERROR, "Unsupported bit depth: %d chroma_idc: %d\n",
3003 h->sps.bit_depth_luma, h->sps.chroma_format_idc);
3008 if (h->sps.video_signal_type_present_flag) {
3009 s->avctx->color_range = h->sps.full_range>0 ? AVCOL_RANGE_JPEG
3011 if (h->sps.colour_description_present_flag) {
3012 s->avctx->color_primaries = h->sps.color_primaries;
3013 s->avctx->color_trc = h->sps.color_trc;
3014 s->avctx->colorspace = h->sps.colorspace;
3018 if (h->sps.timing_info_present_flag) {
3019 int64_t den = h->sps.time_scale;
3020 if (h->x264_build < 44U)
3022 av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
3023 h->sps.num_units_in_tick, den, 1 << 30);
3026 switch (h->sps.bit_depth_luma) {
3029 if (s->avctx->colorspace == AVCOL_SPC_RGB) {
3030 s->avctx->pix_fmt = PIX_FMT_GBRP9;
3032 s->avctx->pix_fmt = PIX_FMT_YUV444P9;
3033 } else if (CHROMA422)
3034 s->avctx->pix_fmt = PIX_FMT_YUV422P9;
3036 s->avctx->pix_fmt = PIX_FMT_YUV420P9;
3040 if (s->avctx->colorspace == AVCOL_SPC_RGB) {
3041 s->avctx->pix_fmt = PIX_FMT_GBRP10;
3043 s->avctx->pix_fmt = PIX_FMT_YUV444P10;
3044 } else if (CHROMA422)
3045 s->avctx->pix_fmt = PIX_FMT_YUV422P10;
3047 s->avctx->pix_fmt = PIX_FMT_YUV420P10;
3051 s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ444P
3053 if (s->avctx->colorspace == AVCOL_SPC_RGB) {
3054 s->avctx->pix_fmt = PIX_FMT_GBR24P;
3055 av_log(h->s.avctx, AV_LOG_DEBUG, "Detected GBR colorspace.\n");
3056 } else if (s->avctx->colorspace == AVCOL_SPC_YCGCO) {
3057 av_log(h->s.avctx, AV_LOG_WARNING, "Detected unsupported YCgCo colorspace.\n");
3059 } else if (CHROMA422) {
3060 s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ422P
3063 s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
3064 s->avctx->codec->pix_fmts ?
3065 s->avctx->codec->pix_fmts :
3066 s->avctx->color_range == AVCOL_RANGE_JPEG ?
3067 hwaccel_pixfmt_list_h264_jpeg_420 :
3068 ff_hwaccel_pixfmt_list_420);
3072 av_log(s->avctx, AV_LOG_ERROR,
3073 "Unsupported bit depth: %d\n", h->sps.bit_depth_luma);
3074 return AVERROR_INVALIDDATA;
3077 s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id,
3080 if (ff_MPV_common_init(s) < 0) {
3081 av_log(h->s.avctx, AV_LOG_ERROR, "ff_MPV_common_init() failed.\n");
3085 h->prev_interlaced_frame = 1;
3087 init_scan_tables(h);
3088 if (ff_h264_alloc_tables(h) < 0) {
3089 av_log(h->s.avctx, AV_LOG_ERROR,
3090 "Could not allocate memory for h264\n");
3091 return AVERROR(ENOMEM);
3094 if (!HAVE_THREADS || !(s->avctx->active_thread_type & FF_THREAD_SLICE)) {
3095 if (context_init(h) < 0) {
3096 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
3100 for (i = 1; i < s->slice_context_count; i++) {
3102 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
3103 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
3104 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
3105 c->h264dsp = h->h264dsp;
3108 c->pixel_shift = h->pixel_shift;
3109 c->cur_chroma_format_idc = h->cur_chroma_format_idc;
3110 init_scan_tables(c);
3111 clone_tables(c, h, i);
3114 for (i = 0; i < s->slice_context_count; i++)
3115 if (context_init(h->thread_context[i]) < 0) {
3116 av_log(h->s.avctx, AV_LOG_ERROR,
3117 "context_init() failed.\n");
3123 if (h == h0 && h->dequant_coeff_pps != pps_id) {
3124 h->dequant_coeff_pps = pps_id;
3125 init_dequant_tables(h);
3128 h->frame_num = get_bits(&s->gb, h->sps.log2_max_frame_num);
3131 h->mb_aff_frame = 0;
3132 last_pic_structure = s0->picture_structure;
3133 last_pic_dropable = s->dropable;
3134 s->dropable = h->nal_ref_idc == 0;
3135 if (h->sps.frame_mbs_only_flag) {
3136 s->picture_structure = PICT_FRAME;
3138 if (!h->sps.direct_8x8_inference_flag && slice_type == AV_PICTURE_TYPE_B) {
3139 av_log(h->s.avctx, AV_LOG_ERROR, "This stream was generated by a broken encoder, invalid 8x8 inference\n");
3142 if (get_bits1(&s->gb)) { // field_pic_flag
3143 s->picture_structure = PICT_TOP_FIELD + get_bits1(&s->gb); // bottom_field_flag
3145 s->picture_structure = PICT_FRAME;
3146 h->mb_aff_frame = h->sps.mb_aff;
3149 h->mb_field_decoding_flag = s->picture_structure != PICT_FRAME;
3151 if (h0->current_slice != 0) {
3152 if (last_pic_structure != s->picture_structure ||
3153 last_pic_dropable != s->dropable) {
3154 av_log(h->s.avctx, AV_LOG_ERROR,
3155 "Changing field mode (%d -> %d) between slices is not allowed\n",
3156 last_pic_structure, s->picture_structure);
3157 s->picture_structure = last_pic_structure;
3158 s->dropable = last_pic_dropable;
3159 return AVERROR_INVALIDDATA;
3162 /* Shorten frame num gaps so we don't have to allocate reference
3163 * frames just to throw them away */
3164 if (h->frame_num != h->prev_frame_num && h->prev_frame_num >= 0) {
3165 int unwrap_prev_frame_num = h->prev_frame_num;
3166 int max_frame_num = 1 << h->sps.log2_max_frame_num;
3168 if (unwrap_prev_frame_num > h->frame_num)
3169 unwrap_prev_frame_num -= max_frame_num;
3171 if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
3172 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
3173 if (unwrap_prev_frame_num < 0)
3174 unwrap_prev_frame_num += max_frame_num;
3176 h->prev_frame_num = unwrap_prev_frame_num;
3180 /* See if we have a decoded first field looking for a pair...
3181 * Here, we're using that to see if we should mark previously
3182 * decode frames as "finished".
3183 * We have to do that before the "dummy" in-between frame allocation,
3184 * since that can modify s->current_picture_ptr. */
3185 if (s0->first_field) {
3186 assert(s0->current_picture_ptr);
3187 assert(s0->current_picture_ptr->f.data[0]);
3188 assert(s0->current_picture_ptr->f.reference != DELAYED_PIC_REF);
3190 /* Mark old field/frame as completed */
3191 if (!last_pic_dropable && s0->current_picture_ptr->owner2 == s0) {
3192 ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
3193 last_pic_structure == PICT_BOTTOM_FIELD);
3196 /* figure out if we have a complementary field pair */
3197 if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
3198 /* Previous field is unmatched. Don't display it, but let it
3199 * remain for reference if marked as such. */
3200 if (!last_pic_dropable && last_pic_structure != PICT_FRAME) {
3201 ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
3202 last_pic_structure == PICT_TOP_FIELD);
3205 if (s0->current_picture_ptr->frame_num != h->frame_num) {
3206 /* This and previous field were reference, but had
3207 * different frame_nums. Consider this field first in
3208 * pair. Throw away previous field except for reference
3210 if (!last_pic_dropable && last_pic_structure != PICT_FRAME) {
3211 ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
3212 last_pic_structure == PICT_TOP_FIELD);
3215 /* Second field in complementary pair */
3216 if (!((last_pic_structure == PICT_TOP_FIELD &&
3217 s->picture_structure == PICT_BOTTOM_FIELD) ||
3218 (last_pic_structure == PICT_BOTTOM_FIELD &&
3219 s->picture_structure == PICT_TOP_FIELD))) {
3220 av_log(s->avctx, AV_LOG_ERROR,
3221 "Invalid field mode combination %d/%d\n",
3222 last_pic_structure, s->picture_structure);
3223 s->picture_structure = last_pic_structure;
3224 s->dropable = last_pic_dropable;
3225 return AVERROR_INVALIDDATA;
3226 } else if (last_pic_dropable != s->dropable) {
3227 av_log(s->avctx, AV_LOG_ERROR,
3228 "Cannot combine reference and non-reference fields in the same frame\n");
3229 av_log_ask_for_sample(s->avctx, NULL);
3230 s->picture_structure = last_pic_structure;
3231 s->dropable = last_pic_dropable;
3232 return AVERROR_INVALIDDATA;
3235 /* Take ownership of this buffer. Note that if another thread owned
3236 * the first field of this buffer, we're not operating on that pointer,
3237 * so the original thread is still responsible for reporting progress
3238 * on that first field (or if that was us, we just did that above).
3239 * By taking ownership, we assign responsibility to ourselves to
3240 * report progress on the second field. */
3241 s0->current_picture_ptr->owner2 = s0;
3246 while (h->frame_num != h->prev_frame_num && h->prev_frame_num >= 0 &&
3247 h->frame_num != (h->prev_frame_num + 1) % (1 << h->sps.log2_max_frame_num)) {
3248 Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
3249 av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n",
3250 h->frame_num, h->prev_frame_num);
3251 if (ff_h264_frame_start(h) < 0)
3253 h->prev_frame_num++;
3254 h->prev_frame_num %= 1 << h->sps.log2_max_frame_num;
3255 s->current_picture_ptr->frame_num = h->prev_frame_num;
3256 ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX, 0);
3257 ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX, 1);
3258 ff_generate_sliding_window_mmcos(h);
3259 if (ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index) < 0 &&
3260 (s->avctx->err_recognition & AV_EF_EXPLODE))
3261 return AVERROR_INVALIDDATA;
3262 /* Error concealment: if a ref is missing, copy the previous ref in its place.
3263 * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
3264 * about there being no actual duplicates.
3265 * FIXME: this doesn't copy padding for out-of-frame motion vectors. Given we're
3266 * concealing a lost frame, this probably isn't noticeable by comparison, but it should
3268 if (h->short_ref_count) {
3270 av_image_copy(h->short_ref[0]->f.data, h->short_ref[0]->f.linesize,
3271 (const uint8_t **)prev->f.data, prev->f.linesize,
3272 s->avctx->pix_fmt, s->mb_width * 16, s->mb_height * 16);
3273 h->short_ref[0]->poc = prev->poc + 2;
3275 h->short_ref[0]->frame_num = h->prev_frame_num;
3279 /* See if we have a decoded first field looking for a pair...
3280 * We're using that to see whether to continue decoding in that
3281 * frame, or to allocate a new one. */
3282 if (s0->first_field) {
3283 assert(s0->current_picture_ptr);
3284 assert(s0->current_picture_ptr->f.data[0]);
3285 assert(s0->current_picture_ptr->f.reference != DELAYED_PIC_REF);
3287 /* figure out if we have a complementary field pair */
3288 if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
3289 /* Previous field is unmatched. Don't display it, but let it
3290 * remain for reference if marked as such. */
3291 s0->current_picture_ptr = NULL;
3292 s0->first_field = FIELD_PICTURE;
3294 if (s0->current_picture_ptr->frame_num != h->frame_num) {
3295 ff_thread_report_progress((AVFrame*)s0->current_picture_ptr, INT_MAX,
3296 s0->picture_structure==PICT_BOTTOM_FIELD);
3297 /* This and the previous field had different frame_nums.
3298 * Consider this field first in pair. Throw away previous
3299 * one except for reference purposes. */
3300 s0->first_field = 1;
3301 s0->current_picture_ptr = NULL;
3303 /* Second field in complementary pair */
3304 s0->first_field = 0;
3308 /* Frame or first field in a potentially complementary pair */
3309 assert(!s0->current_picture_ptr);
3310 s0->first_field = FIELD_PICTURE;
3313 if (!FIELD_PICTURE || s0->first_field) {
3314 if (ff_h264_frame_start(h) < 0) {
3315 s0->first_field = 0;
3319 ff_release_unused_pictures(s, 0);
3325 s->current_picture_ptr->frame_num = h->frame_num; // FIXME frame_num cleanup
3327 assert(s->mb_num == s->mb_width * s->mb_height);
3328 if (first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
3329 first_mb_in_slice >= s->mb_num) {
3330 av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
3333 s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
3334 s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
3335 if (s->picture_structure == PICT_BOTTOM_FIELD)
3336 s->resync_mb_y = s->mb_y = s->mb_y + 1;
3337 assert(s->mb_y < s->mb_height);
3339 if (s->picture_structure == PICT_FRAME) {
3340 h->curr_pic_num = h->frame_num;
3341 h->max_pic_num = 1 << h->sps.log2_max_frame_num;
3343 h->curr_pic_num = 2 * h->frame_num + 1;
3344 h->max_pic_num = 1 << (h->sps.log2_max_frame_num + 1);
3347 if (h->nal_unit_type == NAL_IDR_SLICE)
3348 get_ue_golomb(&s->gb); /* idr_pic_id */
3350 if (h->sps.poc_type == 0) {
3351 h->poc_lsb = get_bits(&s->gb, h->sps.log2_max_poc_lsb);
3353 if (h->pps.pic_order_present == 1 && s->picture_structure == PICT_FRAME)
3354 h->delta_poc_bottom = get_se_golomb(&s->gb);
3357 if (h->sps.poc_type == 1 && !h->sps.delta_pic_order_always_zero_flag) {
3358 h->delta_poc[0] = get_se_golomb(&s->gb);
3360 if (h->pps.pic_order_present == 1 && s->picture_structure == PICT_FRAME)
3361 h->delta_poc[1] = get_se_golomb(&s->gb);
3366 if (h->pps.redundant_pic_cnt_present)
3367 h->redundant_pic_count = get_ue_golomb(&s->gb);
3369 // set defaults, might be overridden a few lines later
3370 h->ref_count[0] = h->pps.ref_count[0];
3371 h->ref_count[1] = h->pps.ref_count[1];
3373 if (h->slice_type_nos != AV_PICTURE_TYPE_I) {
3374 unsigned max = s->picture_structure == PICT_FRAME ? 15 : 31;
3376 if (h->slice_type_nos == AV_PICTURE_TYPE_B)
3377 h->direct_spatial_mv_pred = get_bits1(&s->gb);
3378 num_ref_idx_active_override_flag = get_bits1(&s->gb);
3380 if (num_ref_idx_active_override_flag) {
3381 h->ref_count[0] = get_ue_golomb(&s->gb) + 1;
3382 if (h->slice_type_nos == AV_PICTURE_TYPE_B)
3383 h->ref_count[1] = get_ue_golomb(&s->gb) + 1;
3386 if (h->ref_count[0]-1 > max || h->ref_count[1]-1 > max){
3387 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
3388 h->ref_count[0] = h->ref_count[1] = 1;
3389 return AVERROR_INVALIDDATA;
3392 if (h->slice_type_nos == AV_PICTURE_TYPE_B)
3397 h->ref_count[1]= h->ref_count[0]= h->list_count= 0;
3399 if (!default_ref_list_done)
3400 ff_h264_fill_default_ref_list(h);
3402 if (h->slice_type_nos != AV_PICTURE_TYPE_I &&
3403 ff_h264_decode_ref_pic_list_reordering(h) < 0) {
3404 h->ref_count[1] = h->ref_count[0] = 0;
3408 if (h->slice_type_nos != AV_PICTURE_TYPE_I) {
3409 s->last_picture_ptr = &h->ref_list[0][0];
3410 ff_copy_picture(&s->last_picture, s->last_picture_ptr);
3412 if (h->slice_type_nos == AV_PICTURE_TYPE_B) {
3413 s->next_picture_ptr = &h->ref_list[1][0];
3414 ff_copy_picture(&s->next_picture, s->next_picture_ptr);
3417 if ((h->pps.weighted_pred && h->slice_type_nos == AV_PICTURE_TYPE_P) ||
3418 (h->pps.weighted_bipred_idc == 1 &&
3419 h->slice_type_nos == AV_PICTURE_TYPE_B))
3420 pred_weight_table(h);
3421 else if (h->pps.weighted_bipred_idc == 2 &&
3422 h->slice_type_nos == AV_PICTURE_TYPE_B) {
3423 implicit_weight_table(h, -1);
3426 for (i = 0; i < 2; i++) {
3427 h->luma_weight_flag[i] = 0;
3428 h->chroma_weight_flag[i] = 0;
3432 if (h->nal_ref_idc && ff_h264_decode_ref_pic_marking(h0, &s->gb) < 0 &&
3433 (s->avctx->err_recognition & AV_EF_EXPLODE))
3434 return AVERROR_INVALIDDATA;
3437 ff_h264_fill_mbaff_ref_list(h);
3439 if (h->pps.weighted_bipred_idc == 2 && h->slice_type_nos == AV_PICTURE_TYPE_B) {
3440 implicit_weight_table(h, 0);
3441 implicit_weight_table(h, 1);
3445 if (h->slice_type_nos == AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
3446 ff_h264_direct_dist_scale_factor(h);
3447 ff_h264_direct_ref_list_init(h);
3449 if (h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac) {
3450 tmp = get_ue_golomb_31(&s->gb);
3452 av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
3455 h->cabac_init_idc = tmp;
3458 h->last_qscale_diff = 0;
3459 tmp = h->pps.init_qp + get_se_golomb(&s->gb);
3460 if (tmp > 51 + 6 * (h->sps.bit_depth_luma - 8)) {
3461 av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
3465 h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3466 h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3467 // FIXME qscale / qp ... stuff
3468 if (h->slice_type == AV_PICTURE_TYPE_SP)
3469 get_bits1(&s->gb); /* sp_for_switch_flag */
3470 if (h->slice_type == AV_PICTURE_TYPE_SP ||
3471 h->slice_type == AV_PICTURE_TYPE_SI)
3472 get_se_golomb(&s->gb); /* slice_qs_delta */
3474 h->deblocking_filter = 1;
3475 h->slice_alpha_c0_offset = 52;
3476 h->slice_beta_offset = 52;
3477 if (h->pps.deblocking_filter_parameters_present) {
3478 tmp = get_ue_golomb_31(&s->gb);
3480 av_log(s->avctx, AV_LOG_ERROR,
3481 "deblocking_filter_idc %u out of range\n", tmp);
3484 h->deblocking_filter = tmp;
3485 if (h->deblocking_filter < 2)
3486 h->deblocking_filter ^= 1; // 1<->0
3488 if (h->deblocking_filter) {
3489 h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
3490 h->slice_beta_offset += get_se_golomb(&s->gb) << 1;
3491 if (h->slice_alpha_c0_offset > 104U ||
3492 h->slice_beta_offset > 104U) {
3493 av_log(s->avctx, AV_LOG_ERROR,
3494 "deblocking filter parameters %d %d out of range\n",
3495 h->slice_alpha_c0_offset, h->slice_beta_offset);
3501 if (s->avctx->skip_loop_filter >= AVDISCARD_ALL ||
3502 (s->avctx->skip_loop_filter >= AVDISCARD_NONKEY &&
3503 h->slice_type_nos != AV_PICTURE_TYPE_I) ||
3504 (s->avctx->skip_loop_filter >= AVDISCARD_BIDIR &&
3505 h->slice_type_nos == AV_PICTURE_TYPE_B) ||
3506 (s->avctx->skip_loop_filter >= AVDISCARD_NONREF &&
3507 h->nal_ref_idc == 0))
3508 h->deblocking_filter = 0;
3510 if (h->deblocking_filter == 1 && h0->max_contexts > 1) {
3511 if (s->avctx->flags2 & CODEC_FLAG2_FAST) {
3512 /* Cheat slightly for speed:
3513 * Do not bother to deblock across slices. */
3514 h->deblocking_filter = 2;
3516 h0->max_contexts = 1;
3517 if (!h0->single_decode_warning) {
3518 av_log(s->avctx, AV_LOG_INFO,
3519 "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
3520 h0->single_decode_warning = 1;
3523 av_log(h->s.avctx, AV_LOG_ERROR,
3524 "Deblocking switched inside frame.\n");
3529 h->qp_thresh = 15 + 52 -
3530 FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset) -
3532 h->pps.chroma_qp_index_offset[0],
3533 h->pps.chroma_qp_index_offset[1]) +
3534 6 * (h->sps.bit_depth_luma - 8);
3536 h0->last_slice_type = slice_type;
3537 h->slice_num = ++h0->current_slice;
3540 h0->slice_row[(h->slice_num-1)&(MAX_SLICES-1)]= s->resync_mb_y;
3541 if ( h0->slice_row[h->slice_num&(MAX_SLICES-1)] + 3 >= s->resync_mb_y
3542 && h0->slice_row[h->slice_num&(MAX_SLICES-1)] <= s->resync_mb_y
3543 && h->slice_num >= MAX_SLICES) {
3544 //in case of ASO this check needs to be updated depending on how we decide to assign slice numbers in this case
3545 av_log(s->avctx, AV_LOG_WARNING, "Possibly too many slices (%d >= %d), increase MAX_SLICES and recompile if there are artifacts\n", h->slice_num, MAX_SLICES);
3548 for (j = 0; j < 2; j++) {
3550 int *ref2frm = h->ref2frm[h->slice_num & (MAX_SLICES - 1)][j];
3551 for (i = 0; i < 16; i++) {
3553 if (h->ref_list[j][i].f.data[0]) {
3555 uint8_t *base = h->ref_list[j][i].f.base[0];
3556 for (k = 0; k < h->short_ref_count; k++)
3557 if (h->short_ref[k]->f.base[0] == base) {
3561 for (k = 0; k < h->long_ref_count; k++)
3562 if (h->long_ref[k] && h->long_ref[k]->f.base[0] == base) {
3563 id_list[i] = h->short_ref_count + k;
3571 for (i = 0; i < 16; i++)
3572 ref2frm[i + 2] = 4 * id_list[i] +
3573 (h->ref_list[j][i].f.reference & 3);
3575 ref2frm[18 + 1] = -1;
3576 for (i = 16; i < 48; i++)
3577 ref2frm[i + 4] = 4 * id_list[(i - 16) >> 1] +
3578 (h->ref_list[j][i].f.reference & 3);
3581 // FIXME: fix draw_edges + PAFF + frame threads
3582 h->emu_edge_width = (s->flags & CODEC_FLAG_EMU_EDGE ||
3583 (!h->sps.frame_mbs_only_flag &&
3584 s->avctx->active_thread_type))
3586 h->emu_edge_height = (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
3588 if (s->avctx->debug & FF_DEBUG_PICT_INFO) {
3589 av_log(h->s.avctx, AV_LOG_DEBUG,
3590 "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",
3592 (s->picture_structure == PICT_FRAME ? "F" : s->picture_structure == PICT_TOP_FIELD ? "T" : "B"),
3594 av_get_picture_type_char(h->slice_type),
3595 h->slice_type_fixed ? " fix" : "",
3596 h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
3597 pps_id, h->frame_num,
3598 s->current_picture_ptr->field_poc[0],
3599 s->current_picture_ptr->field_poc[1],
3600 h->ref_count[0], h->ref_count[1],
3602 h->deblocking_filter,
3603 h->slice_alpha_c0_offset / 2 - 26, h->slice_beta_offset / 2 - 26,
3605 h->use_weight == 1 && h->use_weight_chroma ? "c" : "",
3606 h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : "");
3612 int ff_h264_get_slice_type(const H264Context *h)
3614 switch (h->slice_type) {
3615 case AV_PICTURE_TYPE_P:
3617 case AV_PICTURE_TYPE_B:
3619 case AV_PICTURE_TYPE_I:
3621 case AV_PICTURE_TYPE_SP:
3623 case AV_PICTURE_TYPE_SI:
3630 static av_always_inline void fill_filter_caches_inter(H264Context *h,
3631 MpegEncContext *const s,
3632 int mb_type, int top_xy,
3633 int left_xy[LEFT_MBS],
3635 int left_type[LEFT_MBS],
3636 int mb_xy, int list)
3638 int b_stride = h->b_stride;
3639 int16_t(*mv_dst)[2] = &h->mv_cache[list][scan8[0]];
3640 int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
3641 if (IS_INTER(mb_type) || IS_DIRECT(mb_type)) {
3642 if (USES_LIST(top_type, list)) {
3643 const int b_xy = h->mb2b_xy[top_xy] + 3 * b_stride;
3644 const int b8_xy = 4 * top_xy + 2;
3645 int (*ref2frm)[64] = (void*)(h->ref2frm[h->slice_table[top_xy] & (MAX_SLICES - 1)][0] + (MB_MBAFF ? 20 : 2));
3646 AV_COPY128(mv_dst - 1 * 8, s->current_picture.f.motion_val[list][b_xy + 0]);
3647 ref_cache[0 - 1 * 8] =
3648 ref_cache[1 - 1 * 8] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 0]];
3649 ref_cache[2 - 1 * 8] =
3650 ref_cache[3 - 1 * 8] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 1]];
3652 AV_ZERO128(mv_dst - 1 * 8);
3653 AV_WN32A(&ref_cache[0 - 1 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3656 if (!IS_INTERLACED(mb_type ^ left_type[LTOP])) {
3657 if (USES_LIST(left_type[LTOP], list)) {
3658 const int b_xy = h->mb2b_xy[left_xy[LTOP]] + 3;
3659 const int b8_xy = 4 * left_xy[LTOP] + 1;
3660 int (*ref2frm)[64] =(void*)( h->ref2frm[h->slice_table[left_xy[LTOP]] & (MAX_SLICES - 1)][0] + (MB_MBAFF ? 20 : 2));
3661 AV_COPY32(mv_dst - 1 + 0, s->current_picture.f.motion_val[list][b_xy + b_stride * 0]);
3662 AV_COPY32(mv_dst - 1 + 8, s->current_picture.f.motion_val[list][b_xy + b_stride * 1]);
3663 AV_COPY32(mv_dst - 1 + 16, s->current_picture.f.motion_val[list][b_xy + b_stride * 2]);
3664 AV_COPY32(mv_dst - 1 + 24, s->current_picture.f.motion_val[list][b_xy + b_stride * 3]);
3666 ref_cache[-1 + 8] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2 * 0]];
3667 ref_cache[-1 + 16] =
3668 ref_cache[-1 + 24] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2 * 1]];
3670 AV_ZERO32(mv_dst - 1 + 0);
3671 AV_ZERO32(mv_dst - 1 + 8);
3672 AV_ZERO32(mv_dst - 1 + 16);
3673 AV_ZERO32(mv_dst - 1 + 24);
3676 ref_cache[-1 + 16] =
3677 ref_cache[-1 + 24] = LIST_NOT_USED;
3682 if (!USES_LIST(mb_type, list)) {
3683 fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0, 0), 4);
3684 AV_WN32A(&ref_cache[0 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3685 AV_WN32A(&ref_cache[1 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3686 AV_WN32A(&ref_cache[2 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3687 AV_WN32A(&ref_cache[3 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3692 int8_t *ref = &s->current_picture.f.ref_index[list][4 * mb_xy];
3693 int (*ref2frm)[64] = (void*)(h->ref2frm[h->slice_num & (MAX_SLICES - 1)][0] + (MB_MBAFF ? 20 : 2));
3694 uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]], ref2frm[list][ref[1]]) & 0x00FF00FF) * 0x0101;
3695 uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]], ref2frm[list][ref[3]]) & 0x00FF00FF) * 0x0101;
3696 AV_WN32A(&ref_cache[0 * 8], ref01);
3697 AV_WN32A(&ref_cache[1 * 8], ref01);
3698 AV_WN32A(&ref_cache[2 * 8], ref23);
3699 AV_WN32A(&ref_cache[3 * 8], ref23);
3703 int16_t(*mv_src)[2] = &s->current_picture.f.motion_val[list][4 * s->mb_x + 4 * s->mb_y * b_stride];
3704 AV_COPY128(mv_dst + 8 * 0, mv_src + 0 * b_stride);
3705 AV_COPY128(mv_dst + 8 * 1, mv_src + 1 * b_stride);
3706 AV_COPY128(mv_dst + 8 * 2, mv_src + 2 * b_stride);
3707 AV_COPY128(mv_dst + 8 * 3, mv_src + 3 * b_stride);
3713 * @return non zero if the loop filter can be skipped
3715 static int fill_filter_caches(H264Context *h, int mb_type)
3717 MpegEncContext *const s = &h->s;
3718 const int mb_xy = h->mb_xy;
3719 int top_xy, left_xy[LEFT_MBS];
3720 int top_type, left_type[LEFT_MBS];
3724 top_xy = mb_xy - (s->mb_stride << MB_FIELD);
3726 /* Wow, what a mess, why didn't they simplify the interlacing & intra
3727 * stuff, I can't imagine that these complex rules are worth it. */
3729 left_xy[LBOT] = left_xy[LTOP] = mb_xy - 1;
3731 const int left_mb_field_flag = IS_INTERLACED(s->current_picture.f.mb_type[mb_xy - 1]);
3732 const int curr_mb_field_flag = IS_INTERLACED(mb_type);
3734 if (left_mb_field_flag != curr_mb_field_flag)
3735 left_xy[LTOP] -= s->mb_stride;
3737 if (curr_mb_field_flag)
3738 top_xy += s->mb_stride &
3739 (((s->current_picture.f.mb_type[top_xy] >> 7) & 1) - 1);
3740 if (left_mb_field_flag != curr_mb_field_flag)
3741 left_xy[LBOT] += s->mb_stride;
3745 h->top_mb_xy = top_xy;
3746 h->left_mb_xy[LTOP] = left_xy[LTOP];
3747 h->left_mb_xy[LBOT] = left_xy[LBOT];
3749 /* For sufficiently low qp, filtering wouldn't do anything.
3750 * This is a conservative estimate: could also check beta_offset
3751 * and more accurate chroma_qp. */
3752 int qp_thresh = h->qp_thresh; // FIXME strictly we should store qp_thresh for each mb of a slice
3753 int qp = s->current_picture.f.qscale_table[mb_xy];
3754 if (qp <= qp_thresh &&
3755 (left_xy[LTOP] < 0 ||
3756 ((qp + s->current_picture.f.qscale_table[left_xy[LTOP]] + 1) >> 1) <= qp_thresh) &&
3758 ((qp + s->current_picture.f.qscale_table[top_xy] + 1) >> 1) <= qp_thresh)) {
3761 if ((left_xy[LTOP] < 0 ||
3762 ((qp + s->current_picture.f.qscale_table[left_xy[LBOT]] + 1) >> 1) <= qp_thresh) &&
3763 (top_xy < s->mb_stride ||
3764 ((qp + s->current_picture.f.qscale_table[top_xy - s->mb_stride] + 1) >> 1) <= qp_thresh))
3769 top_type = s->current_picture.f.mb_type[top_xy];
3770 left_type[LTOP] = s->current_picture.f.mb_type[left_xy[LTOP]];
3771 left_type[LBOT] = s->current_picture.f.mb_type[left_xy[LBOT]];
3772 if (h->deblocking_filter == 2) {
3773 if (h->slice_table[top_xy] != h->slice_num)
3775 if (h->slice_table[left_xy[LBOT]] != h->slice_num)
3776 left_type[LTOP] = left_type[LBOT] = 0;
3778 if (h->slice_table[top_xy] == 0xFFFF)
3780 if (h->slice_table[left_xy[LBOT]] == 0xFFFF)
3781 left_type[LTOP] = left_type[LBOT] = 0;
3783 h->top_type = top_type;
3784 h->left_type[LTOP] = left_type[LTOP];
3785 h->left_type[LBOT] = left_type[LBOT];
3787 if (IS_INTRA(mb_type))
3790 fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy,
3791 top_type, left_type, mb_xy, 0);
3792 if (h->list_count == 2)
3793 fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy,
3794 top_type, left_type, mb_xy, 1);
3796 nnz = h->non_zero_count[mb_xy];
3797 nnz_cache = h->non_zero_count_cache;
3798 AV_COPY32(&nnz_cache[4 + 8 * 1], &nnz[0]);
3799 AV_COPY32(&nnz_cache[4 + 8 * 2], &nnz[4]);
3800 AV_COPY32(&nnz_cache[4 + 8 * 3], &nnz[8]);
3801 AV_COPY32(&nnz_cache[4 + 8 * 4], &nnz[12]);
3802 h->cbp = h->cbp_table[mb_xy];
3805 nnz = h->non_zero_count[top_xy];
3806 AV_COPY32(&nnz_cache[4 + 8 * 0], &nnz[3 * 4]);
3809 if (left_type[LTOP]) {
3810 nnz = h->non_zero_count[left_xy[LTOP]];
3811 nnz_cache[3 + 8 * 1] = nnz[3 + 0 * 4];
3812 nnz_cache[3 + 8 * 2] = nnz[3 + 1 * 4];
3813 nnz_cache[3 + 8 * 3] = nnz[3 + 2 * 4];
3814 nnz_cache[3 + 8 * 4] = nnz[3 + 3 * 4];
3817 /* CAVLC 8x8dct requires NNZ values for residual decoding that differ
3818 * from what the loop filter needs */
3819 if (!CABAC && h->pps.transform_8x8_mode) {
3820 if (IS_8x8DCT(top_type)) {
3821 nnz_cache[4 + 8 * 0] =
3822 nnz_cache[5 + 8 * 0] = (h->cbp_table[top_xy] & 0x4000) >> 12;
3823 nnz_cache[6 + 8 * 0] =
3824 nnz_cache[7 + 8 * 0] = (h->cbp_table[top_xy] & 0x8000) >> 12;
3826 if (IS_8x8DCT(left_type[LTOP])) {
3827 nnz_cache[3 + 8 * 1] =
3828 nnz_cache[3 + 8 * 2] = (h->cbp_table[left_xy[LTOP]] & 0x2000) >> 12; // FIXME check MBAFF
3830 if (IS_8x8DCT(left_type[LBOT])) {
3831 nnz_cache[3 + 8 * 3] =
3832 nnz_cache[3 + 8 * 4] = (h->cbp_table[left_xy[LBOT]] & 0x8000) >> 12; // FIXME check MBAFF
3835 if (IS_8x8DCT(mb_type)) {
3836 nnz_cache[scan8[0]] =
3837 nnz_cache[scan8[1]] =
3838 nnz_cache[scan8[2]] =
3839 nnz_cache[scan8[3]] = (h->cbp & 0x1000) >> 12;
3841 nnz_cache[scan8[0 + 4]] =
3842 nnz_cache[scan8[1 + 4]] =
3843 nnz_cache[scan8[2 + 4]] =
3844 nnz_cache[scan8[3 + 4]] = (h->cbp & 0x2000) >> 12;
3846 nnz_cache[scan8[0 + 8]] =
3847 nnz_cache[scan8[1 + 8]] =
3848 nnz_cache[scan8[2 + 8]] =
3849 nnz_cache[scan8[3 + 8]] = (h->cbp & 0x4000) >> 12;
3851 nnz_cache[scan8[0 + 12]] =
3852 nnz_cache[scan8[1 + 12]] =
3853 nnz_cache[scan8[2 + 12]] =
3854 nnz_cache[scan8[3 + 12]] = (h->cbp & 0x8000) >> 12;
3861 static void loop_filter(H264Context *h, int start_x, int end_x)
3863 MpegEncContext *const s = &h->s;
3864 uint8_t *dest_y, *dest_cb, *dest_cr;
3865 int linesize, uvlinesize, mb_x, mb_y;
3866 const int end_mb_y = s->mb_y + FRAME_MBAFF;
3867 const int old_slice_type = h->slice_type;
3868 const int pixel_shift = h->pixel_shift;
3869 const int block_h = 16 >> s->chroma_y_shift;
3871 if (h->deblocking_filter) {
3872 for (mb_x = start_x; mb_x < end_x; mb_x++)
3873 for (mb_y = end_mb_y - FRAME_MBAFF; mb_y <= end_mb_y; mb_y++) {
3875 mb_xy = h->mb_xy = mb_x + mb_y * s->mb_stride;
3876 h->slice_num = h->slice_table[mb_xy];
3877 mb_type = s->current_picture.f.mb_type[mb_xy];
3878 h->list_count = h->list_counts[mb_xy];
3882 h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
3886 dest_y = s->current_picture.f.data[0] +
3887 ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
3888 dest_cb = s->current_picture.f.data[1] +
3889 (mb_x << pixel_shift) * (8 << CHROMA444) +
3890 mb_y * s->uvlinesize * block_h;
3891 dest_cr = s->current_picture.f.data[2] +
3892 (mb_x << pixel_shift) * (8 << CHROMA444) +
3893 mb_y * s->uvlinesize * block_h;
3894 // FIXME simplify above
3897 linesize = h->mb_linesize = s->linesize * 2;
3898 uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
3899 if (mb_y & 1) { // FIXME move out of this function?
3900 dest_y -= s->linesize * 15;
3901 dest_cb -= s->uvlinesize * (block_h - 1);
3902 dest_cr -= s->uvlinesize * (block_h - 1);
3905 linesize = h->mb_linesize = s->linesize;
3906 uvlinesize = h->mb_uvlinesize = s->uvlinesize;
3908 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize,
3910 if (fill_filter_caches(h, mb_type))
3912 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.f.qscale_table[mb_xy]);
3913 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.f.qscale_table[mb_xy]);
3916 ff_h264_filter_mb(h, mb_x, mb_y, dest_y, dest_cb, dest_cr,
3917 linesize, uvlinesize);
3919 ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb,
3920 dest_cr, linesize, uvlinesize);
3924 h->slice_type = old_slice_type;
3926 s->mb_y = end_mb_y - FRAME_MBAFF;
3927 h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3928 h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3931 static void predict_field_decoding_flag(H264Context *h)
3933 MpegEncContext *const s = &h->s;
3934 const int mb_xy = s->mb_x + s->mb_y * s->mb_stride;
3935 int mb_type = (h->slice_table[mb_xy - 1] == h->slice_num) ?
3936 s->current_picture.f.mb_type[mb_xy - 1] :
3937 (h->slice_table[mb_xy - s->mb_stride] == h->slice_num) ?
3938 s->current_picture.f.mb_type[mb_xy - s->mb_stride] : 0;
3939 h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
3943 * Draw edges and report progress for the last MB row.
3945 static void decode_finish_row(H264Context *h)
3947 MpegEncContext *const s = &h->s;
3948 int top = 16 * (s->mb_y >> FIELD_PICTURE);
3949 int pic_height = 16 * s->mb_height >> FIELD_PICTURE;
3950 int height = 16 << FRAME_MBAFF;
3951 int deblock_border = (16 + 4) << FRAME_MBAFF;
3953 if (h->deblocking_filter) {
3954 if ((top + height) >= pic_height)
3955 height += deblock_border;
3956 top -= deblock_border;
3959 if (top >= pic_height || (top + height) < h->emu_edge_height)
3962 height = FFMIN(height, pic_height - top);
3963 if (top < h->emu_edge_height) {
3964 height = top + height;
3968 ff_draw_horiz_band(s, top, height);
3973 ff_thread_report_progress(&s->current_picture_ptr->f, top + height - 1,
3974 s->picture_structure == PICT_BOTTOM_FIELD);
3977 static int decode_slice(struct AVCodecContext *avctx, void *arg)
3979 H264Context *h = *(void **)arg;
3980 MpegEncContext *const s = &h->s;
3981 const int part_mask = s->partitioned_frame ? (ER_AC_END | ER_AC_ERROR)
3983 int lf_x_start = s->mb_x;
3985 s->mb_skip_run = -1;
3987 h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME ||
3988 s->codec_id != CODEC_ID_H264 ||
3989 (CONFIG_GRAY && (s->flags & CODEC_FLAG_GRAY));
3993 align_get_bits(&s->gb);
3996 ff_init_cabac_states(&h->cabac);
3997 ff_init_cabac_decoder(&h->cabac,
3998 s->gb.buffer + get_bits_count(&s->gb) / 8,
3999 (get_bits_left(&s->gb) + 7) / 8);
4001 ff_h264_init_cabac_states(h);
4005 int ret = ff_h264_decode_mb_cabac(h);
4007 // STOP_TIMER("decode_mb_cabac")
4010 ff_h264_hl_decode_mb(h);
4012 // FIXME optimal? or let mb_decode decode 16x32 ?
4013 if (ret >= 0 && FRAME_MBAFF) {
4016 ret = ff_h264_decode_mb_cabac(h);
4019 ff_h264_hl_decode_mb(h);
4022 eos = get_cabac_terminate(&h->cabac);
4024 if ((s->workaround_bugs & FF_BUG_TRUNCATED) &&
4025 h->cabac.bytestream > h->cabac.bytestream_end + 2) {
4026 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x - 1,
4027 s->mb_y, ER_MB_END & part_mask);
4028 if (s->mb_x >= lf_x_start)
4029 loop_filter(h, lf_x_start, s->mb_x + 1);
4032 if (ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 2) {
4033 av_log(h->s.avctx, AV_LOG_ERROR,
4034 "error while decoding MB %d %d, bytestream (%td)\n",
4036 h->cabac.bytestream_end - h->cabac.bytestream);
4037 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x,
4038 s->mb_y, ER_MB_ERROR & part_mask);
4042 if (++s->mb_x >= s->mb_width) {
4043 loop_filter(h, lf_x_start, s->mb_x);
4044 s->mb_x = lf_x_start = 0;
4045 decode_finish_row(h);
4047 if (FIELD_OR_MBAFF_PICTURE) {
4049 if (FRAME_MBAFF && s->mb_y < s->mb_height)
4050 predict_field_decoding_flag(h);
4054 if (eos || s->mb_y >= s->mb_height) {
4055 tprintf(s->avctx, "slice end %d %d\n",
4056 get_bits_count(&s->gb), s->gb.size_in_bits);
4057 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x - 1,
4058 s->mb_y, ER_MB_END & part_mask);
4059 if (s->mb_x > lf_x_start)
4060 loop_filter(h, lf_x_start, s->mb_x);
4066 int ret = ff_h264_decode_mb_cavlc(h);
4069 ff_h264_hl_decode_mb(h);
4071 // FIXME optimal? or let mb_decode decode 16x32 ?
4072 if (ret >= 0 && FRAME_MBAFF) {
4074 ret = ff_h264_decode_mb_cavlc(h);
4077 ff_h264_hl_decode_mb(h);
4082 av_log(h->s.avctx, AV_LOG_ERROR,
4083 "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
4084 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x,
4085 s->mb_y, ER_MB_ERROR & part_mask);
4089 if (++s->mb_x >= s->mb_width) {
4090 loop_filter(h, lf_x_start, s->mb_x);
4091 s->mb_x = lf_x_start = 0;
4092 decode_finish_row(h);
4094 if (FIELD_OR_MBAFF_PICTURE) {
4096 if (FRAME_MBAFF && s->mb_y < s->mb_height)
4097 predict_field_decoding_flag(h);
4099 if (s->mb_y >= s->mb_height) {
4100 tprintf(s->avctx, "slice end %d %d\n",
4101 get_bits_count(&s->gb), s->gb.size_in_bits);
4103 if ( get_bits_left(&s->gb) == 0
4104 || get_bits_left(&s->gb) > 0 && !(s->avctx->err_recognition & AV_EF_AGGRESSIVE)) {
4105 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y,
4106 s->mb_x - 1, s->mb_y,
4107 ER_MB_END & part_mask);
4111 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y,
4113 ER_MB_END & part_mask);
4120 if (get_bits_left(&s->gb) <= 0 && s->mb_skip_run <= 0) {
4121 tprintf(s->avctx, "slice end %d %d\n",
4122 get_bits_count(&s->gb), s->gb.size_in_bits);
4123 if (get_bits_left(&s->gb) == 0) {
4124 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y,
4125 s->mb_x - 1, s->mb_y,
4126 ER_MB_END & part_mask);
4127 if (s->mb_x > lf_x_start)
4128 loop_filter(h, lf_x_start, s->mb_x);
4132 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x,
4133 s->mb_y, ER_MB_ERROR & part_mask);
4143 * Call decode_slice() for each context.
4145 * @param h h264 master context
4146 * @param context_count number of contexts to execute
4148 static int execute_decode_slices(H264Context *h, int context_count)
4150 MpegEncContext *const s = &h->s;
4151 AVCodecContext *const avctx = s->avctx;
4155 if (s->avctx->hwaccel ||
4156 s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
4158 if (context_count == 1) {
4159 return decode_slice(avctx, &h);
4161 for (i = 1; i < context_count; i++) {
4162 hx = h->thread_context[i];
4163 hx->s.err_recognition = avctx->err_recognition;
4164 hx->s.error_count = 0;
4165 hx->x264_build = h->x264_build;
4168 avctx->execute(avctx, decode_slice, h->thread_context,
4169 NULL, context_count, sizeof(void *));
4171 /* pull back stuff from slices to master context */
4172 hx = h->thread_context[context_count - 1];
4173 s->mb_x = hx->s.mb_x;
4174 s->mb_y = hx->s.mb_y;
4175 s->dropable = hx->s.dropable;
4176 s->picture_structure = hx->s.picture_structure;
4177 for (i = 1; i < context_count; i++)
4178 h->s.error_count += h->thread_context[i]->s.error_count;
4184 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size)
4186 MpegEncContext *const s = &h->s;
4187 AVCodecContext *const avctx = s->avctx;
4188 H264Context *hx; ///< thread context
4192 int pass = !(avctx->active_thread_type & FF_THREAD_FRAME);
4193 int nals_needed = 0; ///< number of NALs that need decoding before the next frame thread starts
4196 h->nal_unit_type= 0;
4198 if(!s->slice_context_count)
4199 s->slice_context_count= 1;
4200 h->max_contexts = s->slice_context_count;
4201 if (!(s->flags2 & CODEC_FLAG2_CHUNKS)) {
4202 h->current_slice = 0;
4203 if (!s->first_field)
4204 s->current_picture_ptr = NULL;
4205 ff_h264_reset_sei(h);
4208 for (; pass <= 1; pass++) {
4211 next_avc = h->is_avc ? 0 : buf_size;
4221 if (buf_index >= next_avc) {
4222 if (buf_index >= buf_size - h->nal_length_size)
4225 for (i = 0; i < h->nal_length_size; i++)
4226 nalsize = (nalsize << 8) | buf[buf_index++];
4227 if (nalsize <= 0 || nalsize > buf_size - buf_index) {
4228 av_log(h->s.avctx, AV_LOG_ERROR,
4229 "AVC: nal size %d\n", nalsize);
4232 next_avc = buf_index + nalsize;
4234 // start code prefix search
4235 for (; buf_index + 3 < next_avc; buf_index++)
4236 // This should always succeed in the first iteration.
4237 if (buf[buf_index] == 0 &&
4238 buf[buf_index + 1] == 0 &&
4239 buf[buf_index + 2] == 1)
4242 if (buf_index + 3 >= buf_size)
4246 if (buf_index >= next_avc)
4250 hx = h->thread_context[context_count];
4252 ptr = ff_h264_decode_nal(hx, buf + buf_index, &dst_length,
4253 &consumed, next_avc - buf_index);
4254 if (ptr == NULL || dst_length < 0) {
4258 i = buf_index + consumed;
4259 if ((s->workaround_bugs & FF_BUG_AUTODETECT) && i + 3 < next_avc &&
4260 buf[i] == 0x00 && buf[i + 1] == 0x00 &&
4261 buf[i + 2] == 0x01 && buf[i + 3] == 0xE0)
4262 s->workaround_bugs |= FF_BUG_TRUNCATED;
4264 if (!(s->workaround_bugs & FF_BUG_TRUNCATED))
4265 while(dst_length > 0 && ptr[dst_length - 1] == 0)
4267 bit_length = !dst_length ? 0
4269 decode_rbsp_trailing(h, ptr + dst_length - 1));
4271 if (s->avctx->debug & FF_DEBUG_STARTCODE)
4272 av_log(h->s.avctx, AV_LOG_DEBUG, "NAL %d/%d at %d/%d length %d pass %d\n", hx->nal_unit_type, hx->nal_ref_idc, buf_index, buf_size, dst_length, pass);
4274 if (h->is_avc && (nalsize != consumed) && nalsize)
4275 av_log(h->s.avctx, AV_LOG_DEBUG,
4276 "AVC: Consumed only %d bytes instead of %d\n",
4279 buf_index += consumed;
4283 /* packets can sometimes contain multiple PPS/SPS,
4284 * e.g. two PAFF field pictures in one packet, or a demuxer
4285 * which splits NALs strangely if so, when frame threading we
4286 * can't start the next thread until we've read all of them */
4287 switch (hx->nal_unit_type) {
4290 nals_needed = nal_index;
4294 init_get_bits(&hx->s.gb, ptr, bit_length);
4295 if (!get_ue_golomb(&hx->s.gb))
4296 nals_needed = nal_index;
4301 // FIXME do not discard SEI id
4302 if (avctx->skip_frame >= AVDISCARD_NONREF && h->nal_ref_idc == 0)
4307 switch (hx->nal_unit_type) {
4309 if (h->nal_unit_type != NAL_IDR_SLICE) {
4310 av_log(h->s.avctx, AV_LOG_ERROR,
4311 "Invalid mix of idr and non-idr slices\n");
4315 idr(h); // FIXME ensure we don't lose some frames if there is reordering
4317 init_get_bits(&hx->s.gb, ptr, bit_length);
4319 hx->inter_gb_ptr = &hx->s.gb;
4320 hx->s.data_partitioning = 0;
4322 if ((err = decode_slice_header(hx, h)))
4325 if ( h->sei_recovery_frame_cnt >= 0
4326 && ( h->recovery_frame<0
4327 || ((h->recovery_frame - h->frame_num) & ((1 << h->sps.log2_max_frame_num)-1)) > h->sei_recovery_frame_cnt)) {
4328 h->recovery_frame = (h->frame_num + h->sei_recovery_frame_cnt) %
4329 (1 << h->sps.log2_max_frame_num);
4332 s->current_picture_ptr->f.key_frame |=
4333 (hx->nal_unit_type == NAL_IDR_SLICE);
4335 if (h->recovery_frame == h->frame_num) {
4336 s->current_picture_ptr->sync |= 1;
4337 h->recovery_frame = -1;
4340 h->sync |= !!s->current_picture_ptr->f.key_frame;
4341 h->sync |= 3*!!(s->flags2 & CODEC_FLAG2_SHOW_ALL);
4342 s->current_picture_ptr->sync |= h->sync;
4344 if (h->current_slice == 1) {
4345 if (!(s->flags2 & CODEC_FLAG2_CHUNKS))
4346 decode_postinit(h, nal_index >= nals_needed);
4348 if (s->avctx->hwaccel &&
4349 s->avctx->hwaccel->start_frame(s->avctx, NULL, 0) < 0)
4351 if (CONFIG_H264_VDPAU_DECODER &&
4352 s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
4353 ff_vdpau_h264_picture_start(s);
4356 if (hx->redundant_pic_count == 0 &&
4357 (avctx->skip_frame < AVDISCARD_NONREF ||
4359 (avctx->skip_frame < AVDISCARD_BIDIR ||
4360 hx->slice_type_nos != AV_PICTURE_TYPE_B) &&
4361 (avctx->skip_frame < AVDISCARD_NONKEY ||
4362 hx->slice_type_nos == AV_PICTURE_TYPE_I) &&
4363 avctx->skip_frame < AVDISCARD_ALL) {
4364 if (avctx->hwaccel) {
4365 if (avctx->hwaccel->decode_slice(avctx,
4366 &buf[buf_index - consumed],
4369 } else if (CONFIG_H264_VDPAU_DECODER &&
4370 s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU) {
4371 static const uint8_t start_code[] = {
4373 ff_vdpau_add_data_chunk(s, start_code,
4374 sizeof(start_code));
4375 ff_vdpau_add_data_chunk(s, &buf[buf_index - consumed],
4382 init_get_bits(&hx->s.gb, ptr, bit_length);
4384 hx->inter_gb_ptr = NULL;
4386 if ((err = decode_slice_header(hx, h)) < 0)
4389 hx->s.data_partitioning = 1;
4392 init_get_bits(&hx->intra_gb, ptr, bit_length);
4393 hx->intra_gb_ptr = &hx->intra_gb;
4396 init_get_bits(&hx->inter_gb, ptr, bit_length);
4397 hx->inter_gb_ptr = &hx->inter_gb;
4399 if (hx->redundant_pic_count == 0 &&
4401 hx->s.data_partitioning &&
4402 s->context_initialized &&
4403 (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc) &&
4404 (avctx->skip_frame < AVDISCARD_BIDIR ||
4405 hx->slice_type_nos != AV_PICTURE_TYPE_B) &&
4406 (avctx->skip_frame < AVDISCARD_NONKEY ||
4407 hx->slice_type_nos == AV_PICTURE_TYPE_I) &&
4408 avctx->skip_frame < AVDISCARD_ALL)
4412 init_get_bits(&s->gb, ptr, bit_length);
4413 ff_h264_decode_sei(h);
4416 init_get_bits(&s->gb, ptr, bit_length);
4417 if (ff_h264_decode_seq_parameter_set(h) < 0 && (h->is_avc ? (nalsize != consumed) && nalsize : 1)) {
4418 av_log(h->s.avctx, AV_LOG_DEBUG,
4419 "SPS decoding failure, trying alternative mode\n");
4421 av_assert0(next_avc - buf_index + consumed == nalsize);
4422 init_get_bits(&s->gb, &buf[buf_index + 1 - consumed],
4423 8*(next_avc - buf_index + consumed - 1));
4424 ff_h264_decode_seq_parameter_set(h);
4427 if (s->flags & CODEC_FLAG_LOW_DELAY ||
4428 (h->sps.bitstream_restriction_flag &&
4429 !h->sps.num_reorder_frames))
4431 if (avctx->has_b_frames < 2)
4432 avctx->has_b_frames = !s->low_delay;
4435 init_get_bits(&s->gb, ptr, bit_length);
4436 ff_h264_decode_picture_parameter_set(h, bit_length);
4439 case NAL_END_SEQUENCE:
4440 case NAL_END_STREAM:
4441 case NAL_FILLER_DATA:
4443 case NAL_AUXILIARY_SLICE:
4446 av_log(avctx, AV_LOG_DEBUG, "Unknown NAL code: %d (%d bits)\n",
4447 hx->nal_unit_type, bit_length);
4450 if (context_count == h->max_contexts) {
4451 execute_decode_slices(h, context_count);
4456 av_log(h->s.avctx, AV_LOG_ERROR, "decode_slice_header error\n");
4457 else if (err == 1) {
4458 /* Slice could not be decoded in parallel mode, copy down
4459 * NAL unit stuff to context 0 and restart. Note that
4460 * rbsp_buffer is not transferred, but since we no longer
4461 * run in parallel mode this should not be an issue. */
4462 h->nal_unit_type = hx->nal_unit_type;
4463 h->nal_ref_idc = hx->nal_ref_idc;
4470 execute_decode_slices(h, context_count);
4474 if (s->current_picture_ptr && s->current_picture_ptr->owner2 == s &&
4476 ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX,
4477 s->picture_structure == PICT_BOTTOM_FIELD);
4484 * Return the number of bytes consumed for building the current frame.
4486 static int get_consumed_bytes(MpegEncContext *s, int pos, int buf_size)
4489 pos = 1; // avoid infinite loops (i doubt that is needed but ...)
4490 if (pos + 10 > buf_size)
4491 pos = buf_size; // oops ;)
4496 static int decode_frame(AVCodecContext *avctx, void *data,
4497 int *data_size, AVPacket *avpkt)
4499 const uint8_t *buf = avpkt->data;
4500 int buf_size = avpkt->size;
4501 H264Context *h = avctx->priv_data;
4502 MpegEncContext *s = &h->s;
4503 AVFrame *pict = data;
4508 s->flags = avctx->flags;
4509 s->flags2 = avctx->flags2;
4511 /* end of stream, output what is still in the buffers */
4512 if (buf_size == 0) {
4515 s->current_picture_ptr = NULL;
4517 // FIXME factorize this with the output code below
4518 out = h->delayed_pic[0];
4521 h->delayed_pic[i] &&
4522 !h->delayed_pic[i]->f.key_frame &&
4523 !h->delayed_pic[i]->mmco_reset;
4525 if (h->delayed_pic[i]->poc < out->poc) {
4526 out = h->delayed_pic[i];
4530 for (i = out_idx; h->delayed_pic[i]; i++)
4531 h->delayed_pic[i] = h->delayed_pic[i + 1];
4534 *data_size = sizeof(AVFrame);
4540 if(h->is_avc && buf_size >= 9 && buf[0]==1 && buf[2]==0 && (buf[4]&0xFC)==0xFC && (buf[5]&0x1F) && buf[8]==0x67){
4541 int cnt= buf[5]&0x1f;
4542 const uint8_t *p= buf+6;
4544 int nalsize= AV_RB16(p) + 2;
4545 if(nalsize > buf_size - (p-buf) || p[2]!=0x67)
4553 int nalsize= AV_RB16(p) + 2;
4554 if(nalsize > buf_size - (p-buf) || p[2]!=0x68)
4559 return ff_h264_decode_extradata(h, buf, buf_size);
4563 buf_index = decode_nal_units(h, buf, buf_size);
4567 if (!s->current_picture_ptr && h->nal_unit_type == NAL_END_SEQUENCE) {
4568 av_assert0(buf_index <= buf_size);
4572 if (!(s->flags2 & CODEC_FLAG2_CHUNKS) && !s->current_picture_ptr) {
4573 if (avctx->skip_frame >= AVDISCARD_NONREF ||
4574 buf_size >= 4 && !memcmp("Q264", buf, 4))
4576 av_log(avctx, AV_LOG_ERROR, "no frame!\n");
4580 if (!(s->flags2 & CODEC_FLAG2_CHUNKS) ||
4581 (s->mb_y >= s->mb_height && s->mb_height)) {
4582 if (s->flags2 & CODEC_FLAG2_CHUNKS)
4583 decode_postinit(h, 1);
4587 /* Wait for second field. */
4589 if (h->next_output_pic && (h->next_output_pic->sync || h->sync>1)) {
4590 *data_size = sizeof(AVFrame);
4591 *pict = h->next_output_pic->f;
4595 assert(pict->data[0] || !*data_size);
4596 ff_print_debug_info(s, pict);
4597 // printf("out %d\n", (int)pict->data[0]);
4599 return get_consumed_bytes(s, buf_index, buf_size);
4602 av_cold void ff_h264_free_context(H264Context *h)
4606 free_tables(h, 1); // FIXME cleanup init stuff perhaps
4608 for (i = 0; i < MAX_SPS_COUNT; i++)
4609 av_freep(h->sps_buffers + i);
4611 for (i = 0; i < MAX_PPS_COUNT; i++)
4612 av_freep(h->pps_buffers + i);
4615 static av_cold int h264_decode_end(AVCodecContext *avctx)
4617 H264Context *h = avctx->priv_data;
4618 MpegEncContext *s = &h->s;
4620 ff_h264_remove_all_refs(h);
4621 ff_h264_free_context(h);
4623 ff_MPV_common_end(s);
4625 // memset(h, 0, sizeof(H264Context));
4630 static const AVProfile profiles[] = {
4631 { FF_PROFILE_H264_BASELINE, "Baseline" },
4632 { FF_PROFILE_H264_CONSTRAINED_BASELINE, "Constrained Baseline" },
4633 { FF_PROFILE_H264_MAIN, "Main" },
4634 { FF_PROFILE_H264_EXTENDED, "Extended" },
4635 { FF_PROFILE_H264_HIGH, "High" },
4636 { FF_PROFILE_H264_HIGH_10, "High 10" },
4637 { FF_PROFILE_H264_HIGH_10_INTRA, "High 10 Intra" },
4638 { FF_PROFILE_H264_HIGH_422, "High 4:2:2" },
4639 { FF_PROFILE_H264_HIGH_422_INTRA, "High 4:2:2 Intra" },
4640 { FF_PROFILE_H264_HIGH_444, "High 4:4:4" },
4641 { FF_PROFILE_H264_HIGH_444_PREDICTIVE, "High 4:4:4 Predictive" },
4642 { FF_PROFILE_H264_HIGH_444_INTRA, "High 4:4:4 Intra" },
4643 { FF_PROFILE_H264_CAVLC_444, "CAVLC 4:4:4" },
4644 { FF_PROFILE_UNKNOWN },
4647 static const AVOption h264_options[] = {
4648 {"is_avc", "is avc", offsetof(H264Context, is_avc), FF_OPT_TYPE_INT, {.dbl = 0}, 0, 1, 0},
4649 {"nal_length_size", "nal_length_size", offsetof(H264Context, nal_length_size), FF_OPT_TYPE_INT, {.dbl = 0}, 0, 4, 0},
4653 static const AVClass h264_class = {
4655 av_default_item_name,
4657 LIBAVUTIL_VERSION_INT,
4660 static const AVClass h264_vdpau_class = {
4661 "H264 VDPAU Decoder",
4662 av_default_item_name,
4664 LIBAVUTIL_VERSION_INT,
4667 AVCodec ff_h264_decoder = {
4669 .type = AVMEDIA_TYPE_VIDEO,
4670 .id = CODEC_ID_H264,
4671 .priv_data_size = sizeof(H264Context),
4672 .init = ff_h264_decode_init,
4673 .close = h264_decode_end,
4674 .decode = decode_frame,
4675 .capabilities = /*CODEC_CAP_DRAW_HORIZ_BAND |*/ CODEC_CAP_DR1 |
4676 CODEC_CAP_DELAY | CODEC_CAP_SLICE_THREADS |
4677 CODEC_CAP_FRAME_THREADS,
4679 .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10"),
4680 .init_thread_copy = ONLY_IF_THREADS_ENABLED(decode_init_thread_copy),
4681 .update_thread_context = ONLY_IF_THREADS_ENABLED(decode_update_thread_context),
4682 .profiles = NULL_IF_CONFIG_SMALL(profiles),
4683 .priv_class = &h264_class,
4686 #if CONFIG_H264_VDPAU_DECODER
4687 AVCodec ff_h264_vdpau_decoder = {
4688 .name = "h264_vdpau",
4689 .type = AVMEDIA_TYPE_VIDEO,
4690 .id = CODEC_ID_H264,
4691 .priv_data_size = sizeof(H264Context),
4692 .init = ff_h264_decode_init,
4693 .close = h264_decode_end,
4694 .decode = decode_frame,
4695 .capabilities = CODEC_CAP_DR1 | CODEC_CAP_DELAY | CODEC_CAP_HWACCEL_VDPAU,
4697 .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10 (VDPAU acceleration)"),
4698 .pix_fmts = (const enum PixelFormat[]) { PIX_FMT_VDPAU_H264,
4700 .profiles = NULL_IF_CONFIG_SMALL(profiles),
4701 .priv_class = &h264_vdpau_class,