2 * H.26L/H.264/AVC/JVT/14496-10/... decoder
3 * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
5 * This file is part of Libav.
7 * Libav is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * Libav is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with Libav; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
24 * H.264 / AVC / MPEG4 part10 codec.
25 * @author Michael Niedermayer <michaelni@gmx.at>
28 #include "libavutil/avassert.h"
29 #include "libavutil/imgutils.h"
30 #include "libavutil/timer.h"
33 #include "cabac_functions.h"
34 #include "error_resilience.h"
38 #include "h264chroma.h"
39 #include "h264_mvpred.h"
42 #include "mpegutils.h"
43 #include "rectangle.h"
47 static const uint8_t rem6[QP_MAX_NUM + 1] = {
48 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2,
49 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5,
50 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
53 static const uint8_t div6[QP_MAX_NUM + 1] = {
54 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3, 3, 3,
55 3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6,
56 7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 10, 10, 10, 10,
59 static const uint8_t field_scan[16] = {
60 0 + 0 * 4, 0 + 1 * 4, 1 + 0 * 4, 0 + 2 * 4,
61 0 + 3 * 4, 1 + 1 * 4, 1 + 2 * 4, 1 + 3 * 4,
62 2 + 0 * 4, 2 + 1 * 4, 2 + 2 * 4, 2 + 3 * 4,
63 3 + 0 * 4, 3 + 1 * 4, 3 + 2 * 4, 3 + 3 * 4,
66 static const uint8_t field_scan8x8[64] = {
67 0 + 0 * 8, 0 + 1 * 8, 0 + 2 * 8, 1 + 0 * 8,
68 1 + 1 * 8, 0 + 3 * 8, 0 + 4 * 8, 1 + 2 * 8,
69 2 + 0 * 8, 1 + 3 * 8, 0 + 5 * 8, 0 + 6 * 8,
70 0 + 7 * 8, 1 + 4 * 8, 2 + 1 * 8, 3 + 0 * 8,
71 2 + 2 * 8, 1 + 5 * 8, 1 + 6 * 8, 1 + 7 * 8,
72 2 + 3 * 8, 3 + 1 * 8, 4 + 0 * 8, 3 + 2 * 8,
73 2 + 4 * 8, 2 + 5 * 8, 2 + 6 * 8, 2 + 7 * 8,
74 3 + 3 * 8, 4 + 1 * 8, 5 + 0 * 8, 4 + 2 * 8,
75 3 + 4 * 8, 3 + 5 * 8, 3 + 6 * 8, 3 + 7 * 8,
76 4 + 3 * 8, 5 + 1 * 8, 6 + 0 * 8, 5 + 2 * 8,
77 4 + 4 * 8, 4 + 5 * 8, 4 + 6 * 8, 4 + 7 * 8,
78 5 + 3 * 8, 6 + 1 * 8, 6 + 2 * 8, 5 + 4 * 8,
79 5 + 5 * 8, 5 + 6 * 8, 5 + 7 * 8, 6 + 3 * 8,
80 7 + 0 * 8, 7 + 1 * 8, 6 + 4 * 8, 6 + 5 * 8,
81 6 + 6 * 8, 6 + 7 * 8, 7 + 2 * 8, 7 + 3 * 8,
82 7 + 4 * 8, 7 + 5 * 8, 7 + 6 * 8, 7 + 7 * 8,
85 static const uint8_t field_scan8x8_cavlc[64] = {
86 0 + 0 * 8, 1 + 1 * 8, 2 + 0 * 8, 0 + 7 * 8,
87 2 + 2 * 8, 2 + 3 * 8, 2 + 4 * 8, 3 + 3 * 8,
88 3 + 4 * 8, 4 + 3 * 8, 4 + 4 * 8, 5 + 3 * 8,
89 5 + 5 * 8, 7 + 0 * 8, 6 + 6 * 8, 7 + 4 * 8,
90 0 + 1 * 8, 0 + 3 * 8, 1 + 3 * 8, 1 + 4 * 8,
91 1 + 5 * 8, 3 + 1 * 8, 2 + 5 * 8, 4 + 1 * 8,
92 3 + 5 * 8, 5 + 1 * 8, 4 + 5 * 8, 6 + 1 * 8,
93 5 + 6 * 8, 7 + 1 * 8, 6 + 7 * 8, 7 + 5 * 8,
94 0 + 2 * 8, 0 + 4 * 8, 0 + 5 * 8, 2 + 1 * 8,
95 1 + 6 * 8, 4 + 0 * 8, 2 + 6 * 8, 5 + 0 * 8,
96 3 + 6 * 8, 6 + 0 * 8, 4 + 6 * 8, 6 + 2 * 8,
97 5 + 7 * 8, 6 + 4 * 8, 7 + 2 * 8, 7 + 6 * 8,
98 1 + 0 * 8, 1 + 2 * 8, 0 + 6 * 8, 3 + 0 * 8,
99 1 + 7 * 8, 3 + 2 * 8, 2 + 7 * 8, 4 + 2 * 8,
100 3 + 7 * 8, 5 + 2 * 8, 4 + 7 * 8, 5 + 4 * 8,
101 6 + 3 * 8, 6 + 5 * 8, 7 + 3 * 8, 7 + 7 * 8,
104 // zigzag_scan8x8_cavlc[i] = zigzag_scan8x8[(i/4) + 16*(i%4)]
105 static const uint8_t zigzag_scan8x8_cavlc[64] = {
106 0 + 0 * 8, 1 + 1 * 8, 1 + 2 * 8, 2 + 2 * 8,
107 4 + 1 * 8, 0 + 5 * 8, 3 + 3 * 8, 7 + 0 * 8,
108 3 + 4 * 8, 1 + 7 * 8, 5 + 3 * 8, 6 + 3 * 8,
109 2 + 7 * 8, 6 + 4 * 8, 5 + 6 * 8, 7 + 5 * 8,
110 1 + 0 * 8, 2 + 0 * 8, 0 + 3 * 8, 3 + 1 * 8,
111 3 + 2 * 8, 0 + 6 * 8, 4 + 2 * 8, 6 + 1 * 8,
112 2 + 5 * 8, 2 + 6 * 8, 6 + 2 * 8, 5 + 4 * 8,
113 3 + 7 * 8, 7 + 3 * 8, 4 + 7 * 8, 7 + 6 * 8,
114 0 + 1 * 8, 3 + 0 * 8, 0 + 4 * 8, 4 + 0 * 8,
115 2 + 3 * 8, 1 + 5 * 8, 5 + 1 * 8, 5 + 2 * 8,
116 1 + 6 * 8, 3 + 5 * 8, 7 + 1 * 8, 4 + 5 * 8,
117 4 + 6 * 8, 7 + 4 * 8, 5 + 7 * 8, 6 + 7 * 8,
118 0 + 2 * 8, 2 + 1 * 8, 1 + 3 * 8, 5 + 0 * 8,
119 1 + 4 * 8, 2 + 4 * 8, 6 + 0 * 8, 4 + 3 * 8,
120 0 + 7 * 8, 4 + 4 * 8, 7 + 2 * 8, 3 + 6 * 8,
121 5 + 5 * 8, 6 + 5 * 8, 6 + 6 * 8, 7 + 7 * 8,
124 static const uint8_t dequant4_coeff_init[6][3] = {
133 static const uint8_t dequant8_coeff_init_scan[16] = {
134 0, 3, 4, 3, 3, 1, 5, 1, 4, 5, 2, 5, 3, 1, 5, 1
137 static const uint8_t dequant8_coeff_init[6][6] = {
138 { 20, 18, 32, 19, 25, 24 },
139 { 22, 19, 35, 21, 28, 26 },
140 { 26, 23, 42, 24, 33, 31 },
141 { 28, 25, 45, 26, 35, 33 },
142 { 32, 28, 51, 30, 40, 38 },
143 { 36, 32, 58, 34, 46, 43 },
147 static void release_unused_pictures(H264Context *h, int remove_current)
151 /* release non reference frames */
152 for (i = 0; i < H264_MAX_PICTURE_COUNT; i++) {
153 if (h->DPB[i].f.buf[0] && !h->DPB[i].reference &&
154 (remove_current || &h->DPB[i] != h->cur_pic_ptr)) {
155 ff_h264_unref_picture(h, &h->DPB[i]);
160 static int alloc_scratch_buffers(H264Context *h, int linesize)
162 int alloc_size = FFALIGN(FFABS(linesize) + 32, 32);
164 if (h->bipred_scratchpad)
167 h->bipred_scratchpad = av_malloc(16 * 6 * alloc_size);
168 // edge emu needs blocksize + filter length - 1
169 // (= 21x21 for h264)
170 h->edge_emu_buffer = av_mallocz(alloc_size * 2 * 21);
172 if (!h->bipred_scratchpad || !h->edge_emu_buffer) {
173 av_freep(&h->bipred_scratchpad);
174 av_freep(&h->edge_emu_buffer);
175 return AVERROR(ENOMEM);
181 static int init_table_pools(H264Context *h)
183 const int big_mb_num = h->mb_stride * (h->mb_height + 1) + 1;
184 const int mb_array_size = h->mb_stride * h->mb_height;
185 const int b4_stride = h->mb_width * 4 + 1;
186 const int b4_array_size = b4_stride * h->mb_height * 4;
188 h->qscale_table_pool = av_buffer_pool_init(big_mb_num + h->mb_stride,
190 h->mb_type_pool = av_buffer_pool_init((big_mb_num + h->mb_stride) *
191 sizeof(uint32_t), av_buffer_allocz);
192 h->motion_val_pool = av_buffer_pool_init(2 * (b4_array_size + 4) *
193 sizeof(int16_t), av_buffer_allocz);
194 h->ref_index_pool = av_buffer_pool_init(4 * mb_array_size, av_buffer_allocz);
196 if (!h->qscale_table_pool || !h->mb_type_pool || !h->motion_val_pool ||
197 !h->ref_index_pool) {
198 av_buffer_pool_uninit(&h->qscale_table_pool);
199 av_buffer_pool_uninit(&h->mb_type_pool);
200 av_buffer_pool_uninit(&h->motion_val_pool);
201 av_buffer_pool_uninit(&h->ref_index_pool);
202 return AVERROR(ENOMEM);
208 static int alloc_picture(H264Context *h, H264Picture *pic)
212 av_assert0(!pic->f.data[0]);
215 ret = ff_thread_get_buffer(h->avctx, &pic->tf, pic->reference ?
216 AV_GET_BUFFER_FLAG_REF : 0);
220 h->linesize = pic->f.linesize[0];
221 h->uvlinesize = pic->f.linesize[1];
223 if (h->avctx->hwaccel) {
224 const AVHWAccel *hwaccel = h->avctx->hwaccel;
225 av_assert0(!pic->hwaccel_picture_private);
226 if (hwaccel->frame_priv_data_size) {
227 pic->hwaccel_priv_buf = av_buffer_allocz(hwaccel->frame_priv_data_size);
228 if (!pic->hwaccel_priv_buf)
229 return AVERROR(ENOMEM);
230 pic->hwaccel_picture_private = pic->hwaccel_priv_buf->data;
234 if (!h->qscale_table_pool) {
235 ret = init_table_pools(h);
240 pic->qscale_table_buf = av_buffer_pool_get(h->qscale_table_pool);
241 pic->mb_type_buf = av_buffer_pool_get(h->mb_type_pool);
242 if (!pic->qscale_table_buf || !pic->mb_type_buf)
245 pic->mb_type = (uint32_t*)pic->mb_type_buf->data + 2 * h->mb_stride + 1;
246 pic->qscale_table = pic->qscale_table_buf->data + 2 * h->mb_stride + 1;
248 for (i = 0; i < 2; i++) {
249 pic->motion_val_buf[i] = av_buffer_pool_get(h->motion_val_pool);
250 pic->ref_index_buf[i] = av_buffer_pool_get(h->ref_index_pool);
251 if (!pic->motion_val_buf[i] || !pic->ref_index_buf[i])
254 pic->motion_val[i] = (int16_t (*)[2])pic->motion_val_buf[i]->data + 4;
255 pic->ref_index[i] = pic->ref_index_buf[i]->data;
260 ff_h264_unref_picture(h, pic);
261 return (ret < 0) ? ret : AVERROR(ENOMEM);
264 static inline int pic_is_unused(H264Context *h, H264Picture *pic)
268 if (pic->needs_realloc && !(pic->reference & DELAYED_PIC_REF))
273 static int find_unused_picture(H264Context *h)
277 for (i = 0; i < H264_MAX_PICTURE_COUNT; i++) {
278 if (pic_is_unused(h, &h->DPB[i]))
281 if (i == H264_MAX_PICTURE_COUNT)
282 return AVERROR_INVALIDDATA;
284 if (h->DPB[i].needs_realloc) {
285 h->DPB[i].needs_realloc = 0;
286 ff_h264_unref_picture(h, &h->DPB[i]);
293 static void init_dequant8_coeff_table(H264Context *h)
296 const int max_qp = 51 + 6 * (h->sps.bit_depth_luma - 8);
298 for (i = 0; i < 6; i++) {
299 h->dequant8_coeff[i] = h->dequant8_buffer[i];
300 for (j = 0; j < i; j++)
301 if (!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i],
302 64 * sizeof(uint8_t))) {
303 h->dequant8_coeff[i] = h->dequant8_buffer[j];
309 for (q = 0; q < max_qp + 1; q++) {
312 for (x = 0; x < 64; x++)
313 h->dequant8_coeff[i][q][(x >> 3) | ((x & 7) << 3)] =
314 ((uint32_t)dequant8_coeff_init[idx][dequant8_coeff_init_scan[((x >> 1) & 12) | (x & 3)]] *
315 h->pps.scaling_matrix8[i][x]) << shift;
320 static void init_dequant4_coeff_table(H264Context *h)
323 const int max_qp = 51 + 6 * (h->sps.bit_depth_luma - 8);
324 for (i = 0; i < 6; i++) {
325 h->dequant4_coeff[i] = h->dequant4_buffer[i];
326 for (j = 0; j < i; j++)
327 if (!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i],
328 16 * sizeof(uint8_t))) {
329 h->dequant4_coeff[i] = h->dequant4_buffer[j];
335 for (q = 0; q < max_qp + 1; q++) {
336 int shift = div6[q] + 2;
338 for (x = 0; x < 16; x++)
339 h->dequant4_coeff[i][q][(x >> 2) | ((x << 2) & 0xF)] =
340 ((uint32_t)dequant4_coeff_init[idx][(x & 1) + ((x >> 2) & 1)] *
341 h->pps.scaling_matrix4[i][x]) << shift;
346 void h264_init_dequant_tables(H264Context *h)
349 init_dequant4_coeff_table(h);
350 if (h->pps.transform_8x8_mode)
351 init_dequant8_coeff_table(h);
352 if (h->sps.transform_bypass) {
353 for (i = 0; i < 6; i++)
354 for (x = 0; x < 16; x++)
355 h->dequant4_coeff[i][0][x] = 1 << 6;
356 if (h->pps.transform_8x8_mode)
357 for (i = 0; i < 6; i++)
358 for (x = 0; x < 64; x++)
359 h->dequant8_coeff[i][0][x] = 1 << 6;
364 * Mimic alloc_tables(), but for every context thread.
366 static void clone_tables(H264Context *dst, H264Context *src, int i)
368 dst->intra4x4_pred_mode = src->intra4x4_pred_mode + i * 8 * 2 * src->mb_stride;
369 dst->non_zero_count = src->non_zero_count;
370 dst->slice_table = src->slice_table;
371 dst->cbp_table = src->cbp_table;
372 dst->mb2b_xy = src->mb2b_xy;
373 dst->mb2br_xy = src->mb2br_xy;
374 dst->chroma_pred_mode_table = src->chroma_pred_mode_table;
375 dst->mvd_table[0] = src->mvd_table[0] + i * 8 * 2 * src->mb_stride;
376 dst->mvd_table[1] = src->mvd_table[1] + i * 8 * 2 * src->mb_stride;
377 dst->direct_table = src->direct_table;
378 dst->list_counts = src->list_counts;
380 dst->cur_pic_ptr = src->cur_pic_ptr;
381 dst->cur_pic = src->cur_pic;
382 dst->bipred_scratchpad = NULL;
383 dst->edge_emu_buffer = NULL;
384 ff_h264_pred_init(&dst->hpc, src->avctx->codec_id, src->sps.bit_depth_luma,
385 src->sps.chroma_format_idc);
388 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b) + (size))))
390 #define REBASE_PICTURE(pic, new_ctx, old_ctx) \
391 ((pic && pic >= old_ctx->DPB && \
392 pic < old_ctx->DPB + H264_MAX_PICTURE_COUNT) ? \
393 &new_ctx->DPB[pic - old_ctx->DPB] : NULL)
395 static void copy_picture_range(H264Picture **to, H264Picture **from, int count,
396 H264Context *new_base,
397 H264Context *old_base)
401 for (i = 0; i < count; i++) {
402 assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
403 IN_RANGE(from[i], old_base->DPB,
404 sizeof(H264Picture) * H264_MAX_PICTURE_COUNT) ||
406 to[i] = REBASE_PICTURE(from[i], new_base, old_base);
410 static int copy_parameter_set(void **to, void **from, int count, int size)
414 for (i = 0; i < count; i++) {
415 if (to[i] && !from[i]) {
417 } else if (from[i] && !to[i]) {
418 to[i] = av_malloc(size);
420 return AVERROR(ENOMEM);
424 memcpy(to[i], from[i], size);
430 #define copy_fields(to, from, start_field, end_field) \
431 memcpy(&to->start_field, &from->start_field, \
432 (char *)&to->end_field - (char *)&to->start_field)
434 static int h264_slice_header_init(H264Context *h, int reinit);
436 int ff_h264_update_thread_context(AVCodecContext *dst,
437 const AVCodecContext *src)
439 H264Context *h = dst->priv_data, *h1 = src->priv_data;
440 int inited = h->context_initialized, err = 0;
441 int context_reinitialized = 0;
444 if (dst == src || !h1->context_initialized)
448 (h->width != h1->width ||
449 h->height != h1->height ||
450 h->mb_width != h1->mb_width ||
451 h->mb_height != h1->mb_height ||
452 h->sps.bit_depth_luma != h1->sps.bit_depth_luma ||
453 h->sps.chroma_format_idc != h1->sps.chroma_format_idc ||
454 h->sps.colorspace != h1->sps.colorspace)) {
456 /* set bits_per_raw_sample to the previous value. the check for changed
457 * bit depth in h264_set_parameter_from_sps() uses it and sets it to
458 * the current value */
459 h->avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
461 av_freep(&h->bipred_scratchpad);
463 h->width = h1->width;
464 h->height = h1->height;
465 h->mb_height = h1->mb_height;
466 h->mb_width = h1->mb_width;
467 h->mb_num = h1->mb_num;
468 h->mb_stride = h1->mb_stride;
469 h->b_stride = h1->b_stride;
471 if ((err = h264_slice_header_init(h, 1)) < 0) {
472 av_log(h->avctx, AV_LOG_ERROR, "h264_slice_header_init() failed");
475 context_reinitialized = 1;
477 /* update linesize on resize. The decoder doesn't
478 * necessarily call h264_frame_start in the new thread */
479 h->linesize = h1->linesize;
480 h->uvlinesize = h1->uvlinesize;
482 /* copy block_offset since frame_start may not be called */
483 memcpy(h->block_offset, h1->block_offset, sizeof(h->block_offset));
487 H264SliceContext *orig_slice_ctx = h->slice_ctx;
489 for (i = 0; i < MAX_SPS_COUNT; i++)
490 av_freep(h->sps_buffers + i);
492 for (i = 0; i < MAX_PPS_COUNT; i++)
493 av_freep(h->pps_buffers + i);
495 memcpy(h, h1, sizeof(*h1));
496 memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
497 memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
498 memset(&h->er, 0, sizeof(h->er));
499 memset(&h->mb, 0, sizeof(h->mb));
500 memset(&h->mb_luma_dc, 0, sizeof(h->mb_luma_dc));
501 memset(&h->mb_padding, 0, sizeof(h->mb_padding));
502 h->context_initialized = 0;
504 memset(&h->cur_pic, 0, sizeof(h->cur_pic));
505 av_frame_unref(&h->cur_pic.f);
506 h->cur_pic.tf.f = &h->cur_pic.f;
508 h->slice_ctx = orig_slice_ctx;
512 h->qscale_table_pool = NULL;
513 h->mb_type_pool = NULL;
514 h->ref_index_pool = NULL;
515 h->motion_val_pool = NULL;
517 ret = ff_h264_alloc_tables(h);
519 av_log(dst, AV_LOG_ERROR, "Could not allocate memory\n");
522 ret = ff_h264_context_init(h);
524 av_log(dst, AV_LOG_ERROR, "context_init() failed.\n");
528 for (i = 0; i < 2; i++) {
529 h->rbsp_buffer[i] = NULL;
530 h->rbsp_buffer_size[i] = 0;
532 h->bipred_scratchpad = NULL;
533 h->edge_emu_buffer = NULL;
535 h->thread_context[0] = h;
537 h->context_initialized = 1;
540 h->avctx->coded_height = h1->avctx->coded_height;
541 h->avctx->coded_width = h1->avctx->coded_width;
542 h->avctx->width = h1->avctx->width;
543 h->avctx->height = h1->avctx->height;
544 h->coded_picture_number = h1->coded_picture_number;
545 h->first_field = h1->first_field;
546 h->picture_structure = h1->picture_structure;
547 h->qscale = h1->qscale;
548 h->droppable = h1->droppable;
549 h->low_delay = h1->low_delay;
551 for (i = 0; i < H264_MAX_PICTURE_COUNT; i++) {
552 ff_h264_unref_picture(h, &h->DPB[i]);
553 if (h1->DPB[i].f.buf[0] &&
554 (ret = ff_h264_ref_picture(h, &h->DPB[i], &h1->DPB[i])) < 0)
558 h->cur_pic_ptr = REBASE_PICTURE(h1->cur_pic_ptr, h, h1);
559 ff_h264_unref_picture(h, &h->cur_pic);
560 if (h1->cur_pic.f.buf[0]) {
561 ret = ff_h264_ref_picture(h, &h->cur_pic, &h1->cur_pic);
566 h->workaround_bugs = h1->workaround_bugs;
567 h->low_delay = h1->low_delay;
568 h->droppable = h1->droppable;
570 /* frame_start may not be called for the next thread (if it's decoding
571 * a bottom field) so this has to be allocated here */
572 err = alloc_scratch_buffers(h, h1->linesize);
576 // extradata/NAL handling
577 h->is_avc = h1->is_avc;
580 if ((ret = copy_parameter_set((void **)h->sps_buffers,
581 (void **)h1->sps_buffers,
582 MAX_SPS_COUNT, sizeof(SPS))) < 0)
585 if ((ret = copy_parameter_set((void **)h->pps_buffers,
586 (void **)h1->pps_buffers,
587 MAX_PPS_COUNT, sizeof(PPS))) < 0)
591 // Dequantization matrices
592 // FIXME these are big - can they be only copied when PPS changes?
593 copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
595 for (i = 0; i < 6; i++)
596 h->dequant4_coeff[i] = h->dequant4_buffer[0] +
597 (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
599 for (i = 0; i < 6; i++)
600 h->dequant8_coeff[i] = h->dequant8_buffer[0] +
601 (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
603 h->dequant_coeff_pps = h1->dequant_coeff_pps;
606 copy_fields(h, h1, poc_lsb, redundant_pic_count);
609 copy_fields(h, h1, short_ref, cabac_init_idc);
611 copy_picture_range(h->short_ref, h1->short_ref, 32, h, h1);
612 copy_picture_range(h->long_ref, h1->long_ref, 32, h, h1);
613 copy_picture_range(h->delayed_pic, h1->delayed_pic,
614 MAX_DELAYED_PIC_COUNT + 2, h, h1);
616 h->last_slice_type = h1->last_slice_type;
618 if (context_reinitialized)
619 ff_h264_set_parameter_from_sps(h);
625 err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
626 h->prev_poc_msb = h->poc_msb;
627 h->prev_poc_lsb = h->poc_lsb;
629 h->prev_frame_num_offset = h->frame_num_offset;
630 h->prev_frame_num = h->frame_num;
631 h->outputed_poc = h->next_outputed_poc;
633 h->recovery_frame = h1->recovery_frame;
634 h->frame_recovered = h1->frame_recovered;
639 static int h264_frame_start(H264Context *h)
643 const int pixel_shift = h->pixel_shift;
645 release_unused_pictures(h, 1);
646 h->cur_pic_ptr = NULL;
648 i = find_unused_picture(h);
650 av_log(h->avctx, AV_LOG_ERROR, "no frame buffer available\n");
655 pic->reference = h->droppable ? 0 : h->picture_structure;
656 pic->f.coded_picture_number = h->coded_picture_number++;
657 pic->field_picture = h->picture_structure != PICT_FRAME;
659 * Zero key_frame here; IDR markings per slice in frame or fields are ORed
661 * See decode_nal_units().
663 pic->f.key_frame = 0;
667 if ((ret = alloc_picture(h, pic)) < 0)
670 h->cur_pic_ptr = pic;
671 ff_h264_unref_picture(h, &h->cur_pic);
672 if ((ret = ff_h264_ref_picture(h, &h->cur_pic, h->cur_pic_ptr)) < 0)
675 if (CONFIG_ERROR_RESILIENCE)
676 ff_er_frame_start(&h->er);
678 assert(h->linesize && h->uvlinesize);
680 for (i = 0; i < 16; i++) {
681 h->block_offset[i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * h->linesize * ((scan8[i] - scan8[0]) >> 3);
682 h->block_offset[48 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * h->linesize * ((scan8[i] - scan8[0]) >> 3);
684 for (i = 0; i < 16; i++) {
685 h->block_offset[16 + i] =
686 h->block_offset[32 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * h->uvlinesize * ((scan8[i] - scan8[0]) >> 3);
687 h->block_offset[48 + 16 + i] =
688 h->block_offset[48 + 32 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * h->uvlinesize * ((scan8[i] - scan8[0]) >> 3);
691 /* can't be in alloc_tables because linesize isn't known there.
692 * FIXME: redo bipred weight to not require extra buffer? */
693 for (i = 0; i < h->slice_context_count; i++)
694 if (h->thread_context[i]) {
695 ret = alloc_scratch_buffers(h->thread_context[i], h->linesize);
700 /* Some macroblocks can be accessed before they're available in case
701 * of lost slices, MBAFF or threading. */
702 memset(h->slice_table, -1,
703 (h->mb_height * h->mb_stride - 1) * sizeof(*h->slice_table));
705 /* We mark the current picture as non-reference after allocating it, so
706 * that if we break out due to an error it can be released automatically
707 * in the next ff_mpv_frame_start().
709 h->cur_pic_ptr->reference = 0;
711 h->cur_pic_ptr->field_poc[0] = h->cur_pic_ptr->field_poc[1] = INT_MAX;
713 h->next_output_pic = NULL;
715 assert(h->cur_pic_ptr->long_ref == 0);
720 static av_always_inline void backup_mb_border(H264Context *h, uint8_t *src_y,
721 uint8_t *src_cb, uint8_t *src_cr,
722 int linesize, int uvlinesize,
727 const int pixel_shift = h->pixel_shift;
728 int chroma444 = CHROMA444(h);
729 int chroma422 = CHROMA422(h);
732 src_cb -= uvlinesize;
733 src_cr -= uvlinesize;
735 if (!simple && FRAME_MBAFF(h)) {
738 top_border = h->top_borders[0][h->mb_x];
739 AV_COPY128(top_border, src_y + 15 * linesize);
741 AV_COPY128(top_border + 16, src_y + 15 * linesize + 16);
742 if (simple || !CONFIG_GRAY || !(h->flags & CODEC_FLAG_GRAY)) {
745 AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
746 AV_COPY128(top_border + 48, src_cb + 15 * uvlinesize + 16);
747 AV_COPY128(top_border + 64, src_cr + 15 * uvlinesize);
748 AV_COPY128(top_border + 80, src_cr + 15 * uvlinesize + 16);
750 AV_COPY128(top_border + 16, src_cb + 15 * uvlinesize);
751 AV_COPY128(top_border + 32, src_cr + 15 * uvlinesize);
753 } else if (chroma422) {
755 AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
756 AV_COPY128(top_border + 48, src_cr + 15 * uvlinesize);
758 AV_COPY64(top_border + 16, src_cb + 15 * uvlinesize);
759 AV_COPY64(top_border + 24, src_cr + 15 * uvlinesize);
763 AV_COPY128(top_border + 32, src_cb + 7 * uvlinesize);
764 AV_COPY128(top_border + 48, src_cr + 7 * uvlinesize);
766 AV_COPY64(top_border + 16, src_cb + 7 * uvlinesize);
767 AV_COPY64(top_border + 24, src_cr + 7 * uvlinesize);
772 } else if (MB_MBAFF(h)) {
778 top_border = h->top_borders[top_idx][h->mb_x];
779 /* There are two lines saved, the line above the top macroblock
780 * of a pair, and the line above the bottom macroblock. */
781 AV_COPY128(top_border, src_y + 16 * linesize);
783 AV_COPY128(top_border + 16, src_y + 16 * linesize + 16);
785 if (simple || !CONFIG_GRAY || !(h->flags & CODEC_FLAG_GRAY)) {
788 AV_COPY128(top_border + 32, src_cb + 16 * linesize);
789 AV_COPY128(top_border + 48, src_cb + 16 * linesize + 16);
790 AV_COPY128(top_border + 64, src_cr + 16 * linesize);
791 AV_COPY128(top_border + 80, src_cr + 16 * linesize + 16);
793 AV_COPY128(top_border + 16, src_cb + 16 * linesize);
794 AV_COPY128(top_border + 32, src_cr + 16 * linesize);
796 } else if (chroma422) {
798 AV_COPY128(top_border + 32, src_cb + 16 * uvlinesize);
799 AV_COPY128(top_border + 48, src_cr + 16 * uvlinesize);
801 AV_COPY64(top_border + 16, src_cb + 16 * uvlinesize);
802 AV_COPY64(top_border + 24, src_cr + 16 * uvlinesize);
806 AV_COPY128(top_border + 32, src_cb + 8 * uvlinesize);
807 AV_COPY128(top_border + 48, src_cr + 8 * uvlinesize);
809 AV_COPY64(top_border + 16, src_cb + 8 * uvlinesize);
810 AV_COPY64(top_border + 24, src_cr + 8 * uvlinesize);
817 * Initialize implicit_weight table.
818 * @param field 0/1 initialize the weight for interlaced MBAFF
819 * -1 initializes the rest
821 static void implicit_weight_table(H264Context *h, H264SliceContext *sl, int field)
823 int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
825 for (i = 0; i < 2; i++) {
826 sl->luma_weight_flag[i] = 0;
827 sl->chroma_weight_flag[i] = 0;
831 if (h->picture_structure == PICT_FRAME) {
832 cur_poc = h->cur_pic_ptr->poc;
834 cur_poc = h->cur_pic_ptr->field_poc[h->picture_structure - 1];
836 if (h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF(h) &&
837 h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2 * cur_poc) {
839 sl->use_weight_chroma = 0;
843 ref_count0 = h->ref_count[0];
844 ref_count1 = h->ref_count[1];
846 cur_poc = h->cur_pic_ptr->field_poc[field];
848 ref_count0 = 16 + 2 * h->ref_count[0];
849 ref_count1 = 16 + 2 * h->ref_count[1];
853 sl->use_weight_chroma = 2;
854 sl->luma_log2_weight_denom = 5;
855 sl->chroma_log2_weight_denom = 5;
857 for (ref0 = ref_start; ref0 < ref_count0; ref0++) {
858 int poc0 = h->ref_list[0][ref0].poc;
859 for (ref1 = ref_start; ref1 < ref_count1; ref1++) {
861 if (!h->ref_list[0][ref0].long_ref && !h->ref_list[1][ref1].long_ref) {
862 int poc1 = h->ref_list[1][ref1].poc;
863 int td = av_clip_int8(poc1 - poc0);
865 int tb = av_clip_int8(cur_poc - poc0);
866 int tx = (16384 + (FFABS(td) >> 1)) / td;
867 int dist_scale_factor = (tb * tx + 32) >> 8;
868 if (dist_scale_factor >= -64 && dist_scale_factor <= 128)
869 w = 64 - dist_scale_factor;
873 sl->implicit_weight[ref0][ref1][0] =
874 sl->implicit_weight[ref0][ref1][1] = w;
876 sl->implicit_weight[ref0][ref1][field] = w;
883 * initialize scan tables
885 static void init_scan_tables(H264Context *h)
888 for (i = 0; i < 16; i++) {
889 #define TRANSPOSE(x) (x >> 2) | ((x << 2) & 0xF)
890 h->zigzag_scan[i] = TRANSPOSE(zigzag_scan[i]);
891 h->field_scan[i] = TRANSPOSE(field_scan[i]);
894 for (i = 0; i < 64; i++) {
895 #define TRANSPOSE(x) (x >> 3) | ((x & 7) << 3)
896 h->zigzag_scan8x8[i] = TRANSPOSE(ff_zigzag_direct[i]);
897 h->zigzag_scan8x8_cavlc[i] = TRANSPOSE(zigzag_scan8x8_cavlc[i]);
898 h->field_scan8x8[i] = TRANSPOSE(field_scan8x8[i]);
899 h->field_scan8x8_cavlc[i] = TRANSPOSE(field_scan8x8_cavlc[i]);
902 if (h->sps.transform_bypass) { // FIXME same ugly
903 h->zigzag_scan_q0 = zigzag_scan;
904 h->zigzag_scan8x8_q0 = ff_zigzag_direct;
905 h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
906 h->field_scan_q0 = field_scan;
907 h->field_scan8x8_q0 = field_scan8x8;
908 h->field_scan8x8_cavlc_q0 = field_scan8x8_cavlc;
910 h->zigzag_scan_q0 = h->zigzag_scan;
911 h->zigzag_scan8x8_q0 = h->zigzag_scan8x8;
912 h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
913 h->field_scan_q0 = h->field_scan;
914 h->field_scan8x8_q0 = h->field_scan8x8;
915 h->field_scan8x8_cavlc_q0 = h->field_scan8x8_cavlc;
920 * Replicate H264 "master" context to thread contexts.
922 static int clone_slice(H264Context *dst, H264Context *src)
924 memcpy(dst->block_offset, src->block_offset, sizeof(dst->block_offset));
925 dst->cur_pic_ptr = src->cur_pic_ptr;
926 dst->cur_pic = src->cur_pic;
927 dst->linesize = src->linesize;
928 dst->uvlinesize = src->uvlinesize;
929 dst->first_field = src->first_field;
931 dst->prev_poc_msb = src->prev_poc_msb;
932 dst->prev_poc_lsb = src->prev_poc_lsb;
933 dst->prev_frame_num_offset = src->prev_frame_num_offset;
934 dst->prev_frame_num = src->prev_frame_num;
935 dst->short_ref_count = src->short_ref_count;
937 memcpy(dst->short_ref, src->short_ref, sizeof(dst->short_ref));
938 memcpy(dst->long_ref, src->long_ref, sizeof(dst->long_ref));
939 memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
941 memcpy(dst->dequant4_coeff, src->dequant4_coeff, sizeof(src->dequant4_coeff));
942 memcpy(dst->dequant8_coeff, src->dequant8_coeff, sizeof(src->dequant8_coeff));
947 static enum AVPixelFormat get_pixel_format(H264Context *h)
949 #define HWACCEL_MAX (CONFIG_H264_DXVA2_HWACCEL + \
950 CONFIG_H264_VAAPI_HWACCEL + \
951 (CONFIG_H264_VDA_HWACCEL * 2) + \
952 CONFIG_H264_VDPAU_HWACCEL)
953 enum AVPixelFormat pix_fmts[HWACCEL_MAX + 2], *fmt = pix_fmts;
954 const enum AVPixelFormat *choices = pix_fmts;
956 switch (h->sps.bit_depth_luma) {
959 if (h->avctx->colorspace == AVCOL_SPC_RGB) {
960 *fmt++ = AV_PIX_FMT_GBRP9;
962 *fmt++ = AV_PIX_FMT_YUV444P9;
963 } else if (CHROMA422(h))
964 *fmt++ = AV_PIX_FMT_YUV422P9;
966 *fmt++ = AV_PIX_FMT_YUV420P9;
970 if (h->avctx->colorspace == AVCOL_SPC_RGB) {
971 *fmt++ = AV_PIX_FMT_GBRP10;
973 *fmt++ = AV_PIX_FMT_YUV444P10;
974 } else if (CHROMA422(h))
975 *fmt++ = AV_PIX_FMT_YUV422P10;
977 *fmt++ = AV_PIX_FMT_YUV420P10;
980 #if CONFIG_H264_VDPAU_HWACCEL
981 *fmt++ = AV_PIX_FMT_VDPAU;
984 if (h->avctx->colorspace == AVCOL_SPC_RGB)
985 *fmt++ = AV_PIX_FMT_GBRP;
986 else if (h->avctx->color_range == AVCOL_RANGE_JPEG)
987 *fmt++ = AV_PIX_FMT_YUVJ444P;
989 *fmt++ = AV_PIX_FMT_YUV444P;
990 } else if (CHROMA422(h)) {
991 if (h->avctx->color_range == AVCOL_RANGE_JPEG)
992 *fmt++ = AV_PIX_FMT_YUVJ422P;
994 *fmt++ = AV_PIX_FMT_YUV422P;
996 #if CONFIG_H264_DXVA2_HWACCEL
997 *fmt++ = AV_PIX_FMT_DXVA2_VLD;
999 #if CONFIG_H264_VAAPI_HWACCEL
1000 *fmt++ = AV_PIX_FMT_VAAPI_VLD;
1002 #if CONFIG_H264_VDA_HWACCEL
1003 *fmt++ = AV_PIX_FMT_VDA_VLD;
1004 *fmt++ = AV_PIX_FMT_VDA;
1006 if (h->avctx->codec->pix_fmts)
1007 choices = h->avctx->codec->pix_fmts;
1008 else if (h->avctx->color_range == AVCOL_RANGE_JPEG)
1009 *fmt++ = AV_PIX_FMT_YUVJ420P;
1011 *fmt++ = AV_PIX_FMT_YUV420P;
1015 av_log(h->avctx, AV_LOG_ERROR,
1016 "Unsupported bit depth %d\n", h->sps.bit_depth_luma);
1017 return AVERROR_INVALIDDATA;
1020 *fmt = AV_PIX_FMT_NONE;
1022 return ff_get_format(h->avctx, choices);
1025 /* export coded and cropped frame dimensions to AVCodecContext */
1026 static int init_dimensions(H264Context *h)
1028 int width = h->width - (h->sps.crop_right + h->sps.crop_left);
1029 int height = h->height - (h->sps.crop_top + h->sps.crop_bottom);
1030 int crop_present = h->sps.crop_left || h->sps.crop_top ||
1031 h->sps.crop_right || h->sps.crop_bottom;
1033 /* handle container cropping */
1034 if (!crop_present &&
1035 FFALIGN(h->avctx->width, 16) == h->width &&
1036 FFALIGN(h->avctx->height, 16) == h->height) {
1037 width = h->avctx->width;
1038 height = h->avctx->height;
1041 if (width <= 0 || height <= 0) {
1042 av_log(h->avctx, AV_LOG_ERROR, "Invalid cropped dimensions: %dx%d.\n",
1044 if (h->avctx->err_recognition & AV_EF_EXPLODE)
1045 return AVERROR_INVALIDDATA;
1047 av_log(h->avctx, AV_LOG_WARNING, "Ignoring cropping information.\n");
1048 h->sps.crop_bottom =
1058 h->avctx->coded_width = h->width;
1059 h->avctx->coded_height = h->height;
1060 h->avctx->width = width;
1061 h->avctx->height = height;
1066 static int h264_slice_header_init(H264Context *h, int reinit)
1068 int nb_slices = (HAVE_THREADS &&
1069 h->avctx->active_thread_type & FF_THREAD_SLICE) ?
1070 h->avctx->thread_count : 1;
1073 ff_set_sar(h->avctx, h->sps.sar);
1074 av_pix_fmt_get_chroma_sub_sample(h->avctx->pix_fmt,
1075 &h->chroma_x_shift, &h->chroma_y_shift);
1077 if (h->sps.timing_info_present_flag) {
1078 int64_t den = h->sps.time_scale;
1079 if (h->x264_build < 44U)
1081 av_reduce(&h->avctx->framerate.den, &h->avctx->framerate.num,
1082 h->sps.num_units_in_tick, den, 1 << 30);
1086 ff_h264_free_tables(h, 0);
1088 h->prev_interlaced_frame = 1;
1090 init_scan_tables(h);
1091 ret = ff_h264_alloc_tables(h);
1093 av_log(h->avctx, AV_LOG_ERROR, "Could not allocate memory\n");
1097 if (nb_slices > H264_MAX_THREADS || (nb_slices > h->mb_height && h->mb_height)) {
1100 max_slices = FFMIN(H264_MAX_THREADS, h->mb_height);
1102 max_slices = H264_MAX_THREADS;
1103 av_log(h->avctx, AV_LOG_WARNING, "too many threads/slices %d,"
1104 " reducing to %d\n", nb_slices, max_slices);
1105 nb_slices = max_slices;
1107 h->slice_context_count = nb_slices;
1109 if (!HAVE_THREADS || !(h->avctx->active_thread_type & FF_THREAD_SLICE)) {
1110 ret = ff_h264_context_init(h);
1112 av_log(h->avctx, AV_LOG_ERROR, "context_init() failed.\n");
1116 for (i = 1; i < h->slice_context_count; i++) {
1118 c = h->thread_context[i] = av_mallocz(sizeof(H264Context));
1120 return AVERROR(ENOMEM);
1121 c->avctx = h->avctx;
1123 c->h264dsp = h->h264dsp;
1124 c->h264qpel = h->h264qpel;
1125 c->h264chroma = h->h264chroma;
1128 c->pixel_shift = h->pixel_shift;
1129 c->width = h->width;
1130 c->height = h->height;
1131 c->linesize = h->linesize;
1132 c->uvlinesize = h->uvlinesize;
1133 c->chroma_x_shift = h->chroma_x_shift;
1134 c->chroma_y_shift = h->chroma_y_shift;
1135 c->qscale = h->qscale;
1136 c->droppable = h->droppable;
1137 c->low_delay = h->low_delay;
1138 c->mb_width = h->mb_width;
1139 c->mb_height = h->mb_height;
1140 c->mb_stride = h->mb_stride;
1141 c->mb_num = h->mb_num;
1142 c->flags = h->flags;
1143 c->workaround_bugs = h->workaround_bugs;
1144 c->pict_type = h->pict_type;
1146 h->slice_ctx[i].h264 = c;
1148 init_scan_tables(c);
1149 clone_tables(c, h, i);
1150 c->context_initialized = 1;
1153 for (i = 0; i < h->slice_context_count; i++)
1154 if ((ret = ff_h264_context_init(h->thread_context[i])) < 0) {
1155 av_log(h->avctx, AV_LOG_ERROR, "context_init() failed.\n");
1160 h->context_initialized = 1;
1166 * Decode a slice header.
1167 * This will (re)intialize the decoder and call h264_frame_start() as needed.
1169 * @param h h264context
1170 * @param h0 h264 master context (differs from 'h' when doing sliced based
1171 * parallel decoding)
1173 * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
1175 int ff_h264_decode_slice_header(H264Context *h, H264SliceContext *sl, H264Context *h0)
1177 unsigned int first_mb_in_slice;
1178 unsigned int pps_id;
1180 unsigned int slice_type, tmp, i, j;
1181 int default_ref_list_done = 0;
1182 int last_pic_structure, last_pic_droppable;
1183 int needs_reinit = 0;
1184 int field_pic_flag, bottom_field_flag;
1186 h->qpel_put = h->h264qpel.put_h264_qpel_pixels_tab;
1187 h->qpel_avg = h->h264qpel.avg_h264_qpel_pixels_tab;
1189 first_mb_in_slice = get_ue_golomb(&h->gb);
1191 if (first_mb_in_slice == 0) { // FIXME better field boundary detection
1192 if (h0->current_slice && h->cur_pic_ptr && FIELD_PICTURE(h)) {
1193 ff_h264_field_end(h, 1);
1196 h0->current_slice = 0;
1197 if (!h0->first_field) {
1198 if (h->cur_pic_ptr && !h->droppable) {
1199 ff_thread_report_progress(&h->cur_pic_ptr->tf, INT_MAX,
1200 h->picture_structure == PICT_BOTTOM_FIELD);
1202 h->cur_pic_ptr = NULL;
1206 slice_type = get_ue_golomb_31(&h->gb);
1207 if (slice_type > 9) {
1208 av_log(h->avctx, AV_LOG_ERROR,
1209 "slice type %d too large at %d %d\n",
1210 slice_type, h->mb_x, h->mb_y);
1211 return AVERROR_INVALIDDATA;
1213 if (slice_type > 4) {
1215 h->slice_type_fixed = 1;
1217 h->slice_type_fixed = 0;
1219 slice_type = golomb_to_pict_type[slice_type];
1220 if (slice_type == AV_PICTURE_TYPE_I ||
1221 (h0->current_slice != 0 && slice_type == h0->last_slice_type)) {
1222 default_ref_list_done = 1;
1224 h->slice_type = slice_type;
1225 h->slice_type_nos = slice_type & 3;
1227 if (h->nal_unit_type == NAL_IDR_SLICE &&
1228 h->slice_type_nos != AV_PICTURE_TYPE_I) {
1229 av_log(h->avctx, AV_LOG_ERROR, "A non-intra slice in an IDR NAL unit.\n");
1230 return AVERROR_INVALIDDATA;
1233 // to make a few old functions happy, it's wrong though
1234 h->pict_type = h->slice_type;
1236 pps_id = get_ue_golomb(&h->gb);
1237 if (pps_id >= MAX_PPS_COUNT) {
1238 av_log(h->avctx, AV_LOG_ERROR, "pps_id %u out of range\n", pps_id);
1239 return AVERROR_INVALIDDATA;
1241 if (!h0->pps_buffers[pps_id]) {
1242 av_log(h->avctx, AV_LOG_ERROR,
1243 "non-existing PPS %u referenced\n",
1245 return AVERROR_INVALIDDATA;
1247 h->pps = *h0->pps_buffers[pps_id];
1249 if (!h0->sps_buffers[h->pps.sps_id]) {
1250 av_log(h->avctx, AV_LOG_ERROR,
1251 "non-existing SPS %u referenced\n",
1253 return AVERROR_INVALIDDATA;
1256 if (h->pps.sps_id != h->sps.sps_id ||
1257 h0->sps_buffers[h->pps.sps_id]->new) {
1258 h0->sps_buffers[h->pps.sps_id]->new = 0;
1260 h->sps = *h0->sps_buffers[h->pps.sps_id];
1262 if (h->bit_depth_luma != h->sps.bit_depth_luma ||
1263 h->chroma_format_idc != h->sps.chroma_format_idc) {
1264 h->bit_depth_luma = h->sps.bit_depth_luma;
1265 h->chroma_format_idc = h->sps.chroma_format_idc;
1268 if ((ret = ff_h264_set_parameter_from_sps(h)) < 0)
1272 h->avctx->profile = ff_h264_get_profile(&h->sps);
1273 h->avctx->level = h->sps.level_idc;
1274 h->avctx->refs = h->sps.ref_frame_count;
1276 if (h->mb_width != h->sps.mb_width ||
1277 h->mb_height != h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag))
1280 h->mb_width = h->sps.mb_width;
1281 h->mb_height = h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
1282 h->mb_num = h->mb_width * h->mb_height;
1283 h->mb_stride = h->mb_width + 1;
1285 h->b_stride = h->mb_width * 4;
1287 h->chroma_y_shift = h->sps.chroma_format_idc <= 1; // 400 uses yuv420p
1289 h->width = 16 * h->mb_width;
1290 h->height = 16 * h->mb_height;
1292 ret = init_dimensions(h);
1296 if (h->sps.video_signal_type_present_flag) {
1297 h->avctx->color_range = h->sps.full_range ? AVCOL_RANGE_JPEG
1299 if (h->sps.colour_description_present_flag) {
1300 if (h->avctx->colorspace != h->sps.colorspace)
1302 h->avctx->color_primaries = h->sps.color_primaries;
1303 h->avctx->color_trc = h->sps.color_trc;
1304 h->avctx->colorspace = h->sps.colorspace;
1308 if (h->context_initialized && needs_reinit) {
1310 av_log(h->avctx, AV_LOG_ERROR,
1311 "changing width %d -> %d / height %d -> %d on "
1313 h->width, h->avctx->coded_width,
1314 h->height, h->avctx->coded_height,
1315 h0->current_slice + 1);
1316 return AVERROR_INVALIDDATA;
1319 ff_h264_flush_change(h);
1321 if ((ret = get_pixel_format(h)) < 0)
1323 h->avctx->pix_fmt = ret;
1325 av_log(h->avctx, AV_LOG_INFO, "Reinit context to %dx%d, "
1326 "pix_fmt: %d\n", h->width, h->height, h->avctx->pix_fmt);
1328 if ((ret = h264_slice_header_init(h, 1)) < 0) {
1329 av_log(h->avctx, AV_LOG_ERROR,
1330 "h264_slice_header_init() failed\n");
1334 if (!h->context_initialized) {
1336 av_log(h->avctx, AV_LOG_ERROR,
1337 "Cannot (re-)initialize context during parallel decoding.\n");
1338 return AVERROR_PATCHWELCOME;
1341 if ((ret = get_pixel_format(h)) < 0)
1343 h->avctx->pix_fmt = ret;
1345 if ((ret = h264_slice_header_init(h, 0)) < 0) {
1346 av_log(h->avctx, AV_LOG_ERROR,
1347 "h264_slice_header_init() failed\n");
1352 if (h == h0 && h->dequant_coeff_pps != pps_id) {
1353 h->dequant_coeff_pps = pps_id;
1354 h264_init_dequant_tables(h);
1357 h->frame_num = get_bits(&h->gb, h->sps.log2_max_frame_num);
1360 h->mb_aff_frame = 0;
1361 last_pic_structure = h0->picture_structure;
1362 last_pic_droppable = h0->droppable;
1363 h->droppable = h->nal_ref_idc == 0;
1364 if (h->sps.frame_mbs_only_flag) {
1365 h->picture_structure = PICT_FRAME;
1367 field_pic_flag = get_bits1(&h->gb);
1368 if (field_pic_flag) {
1369 bottom_field_flag = get_bits1(&h->gb);
1370 h->picture_structure = PICT_TOP_FIELD + bottom_field_flag;
1372 h->picture_structure = PICT_FRAME;
1373 h->mb_aff_frame = h->sps.mb_aff;
1376 h->mb_field_decoding_flag = h->picture_structure != PICT_FRAME;
1378 if (h0->current_slice != 0) {
1379 if (last_pic_structure != h->picture_structure ||
1380 last_pic_droppable != h->droppable) {
1381 av_log(h->avctx, AV_LOG_ERROR,
1382 "Changing field mode (%d -> %d) between slices is not allowed\n",
1383 last_pic_structure, h->picture_structure);
1384 h->picture_structure = last_pic_structure;
1385 h->droppable = last_pic_droppable;
1386 return AVERROR_INVALIDDATA;
1387 } else if (!h0->cur_pic_ptr) {
1388 av_log(h->avctx, AV_LOG_ERROR,
1389 "unset cur_pic_ptr on slice %d\n",
1390 h0->current_slice + 1);
1391 return AVERROR_INVALIDDATA;
1394 /* Shorten frame num gaps so we don't have to allocate reference
1395 * frames just to throw them away */
1396 if (h->frame_num != h->prev_frame_num) {
1397 int unwrap_prev_frame_num = h->prev_frame_num;
1398 int max_frame_num = 1 << h->sps.log2_max_frame_num;
1400 if (unwrap_prev_frame_num > h->frame_num)
1401 unwrap_prev_frame_num -= max_frame_num;
1403 if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
1404 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
1405 if (unwrap_prev_frame_num < 0)
1406 unwrap_prev_frame_num += max_frame_num;
1408 h->prev_frame_num = unwrap_prev_frame_num;
1412 /* See if we have a decoded first field looking for a pair...
1413 * Here, we're using that to see if we should mark previously
1414 * decode frames as "finished".
1415 * We have to do that before the "dummy" in-between frame allocation,
1416 * since that can modify s->current_picture_ptr. */
1417 if (h0->first_field) {
1418 assert(h0->cur_pic_ptr);
1419 assert(h0->cur_pic_ptr->f.buf[0]);
1420 assert(h0->cur_pic_ptr->reference != DELAYED_PIC_REF);
1422 /* figure out if we have a complementary field pair */
1423 if (!FIELD_PICTURE(h) || h->picture_structure == last_pic_structure) {
1424 /* Previous field is unmatched. Don't display it, but let it
1425 * remain for reference if marked as such. */
1426 if (!last_pic_droppable && last_pic_structure != PICT_FRAME) {
1427 ff_thread_report_progress(&h0->cur_pic_ptr->tf, INT_MAX,
1428 last_pic_structure == PICT_TOP_FIELD);
1431 if (h0->cur_pic_ptr->frame_num != h->frame_num) {
1432 /* This and previous field were reference, but had
1433 * different frame_nums. Consider this field first in
1434 * pair. Throw away previous field except for reference
1436 if (!last_pic_droppable && last_pic_structure != PICT_FRAME) {
1437 ff_thread_report_progress(&h0->cur_pic_ptr->tf, INT_MAX,
1438 last_pic_structure == PICT_TOP_FIELD);
1441 /* Second field in complementary pair */
1442 if (!((last_pic_structure == PICT_TOP_FIELD &&
1443 h->picture_structure == PICT_BOTTOM_FIELD) ||
1444 (last_pic_structure == PICT_BOTTOM_FIELD &&
1445 h->picture_structure == PICT_TOP_FIELD))) {
1446 av_log(h->avctx, AV_LOG_ERROR,
1447 "Invalid field mode combination %d/%d\n",
1448 last_pic_structure, h->picture_structure);
1449 h->picture_structure = last_pic_structure;
1450 h->droppable = last_pic_droppable;
1451 return AVERROR_INVALIDDATA;
1452 } else if (last_pic_droppable != h->droppable) {
1453 avpriv_request_sample(h->avctx,
1454 "Found reference and non-reference fields in the same frame, which");
1455 h->picture_structure = last_pic_structure;
1456 h->droppable = last_pic_droppable;
1457 return AVERROR_PATCHWELCOME;
1463 while (h->frame_num != h->prev_frame_num &&
1464 h->frame_num != (h->prev_frame_num + 1) % (1 << h->sps.log2_max_frame_num)) {
1465 H264Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
1466 av_log(h->avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n",
1467 h->frame_num, h->prev_frame_num);
1468 ret = h264_frame_start(h);
1470 h0->first_field = 0;
1474 h->prev_frame_num++;
1475 h->prev_frame_num %= 1 << h->sps.log2_max_frame_num;
1476 h->cur_pic_ptr->frame_num = h->prev_frame_num;
1477 ff_thread_report_progress(&h->cur_pic_ptr->tf, INT_MAX, 0);
1478 ff_thread_report_progress(&h->cur_pic_ptr->tf, INT_MAX, 1);
1479 ret = ff_generate_sliding_window_mmcos(h, 1);
1480 if (ret < 0 && (h->avctx->err_recognition & AV_EF_EXPLODE))
1482 ret = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1483 if (ret < 0 && (h->avctx->err_recognition & AV_EF_EXPLODE))
1485 /* Error concealment: If a ref is missing, copy the previous ref
1487 * FIXME: Avoiding a memcpy would be nice, but ref handling makes
1488 * many assumptions about there being no actual duplicates.
1489 * FIXME: This does not copy padding for out-of-frame motion
1490 * vectors. Given we are concealing a lost frame, this probably
1491 * is not noticeable by comparison, but it should be fixed. */
1492 if (h->short_ref_count) {
1494 av_image_copy(h->short_ref[0]->f.data,
1495 h->short_ref[0]->f.linesize,
1496 (const uint8_t **)prev->f.data,
1501 h->short_ref[0]->poc = prev->poc + 2;
1503 h->short_ref[0]->frame_num = h->prev_frame_num;
1507 /* See if we have a decoded first field looking for a pair...
1508 * We're using that to see whether to continue decoding in that
1509 * frame, or to allocate a new one. */
1510 if (h0->first_field) {
1511 assert(h0->cur_pic_ptr);
1512 assert(h0->cur_pic_ptr->f.buf[0]);
1513 assert(h0->cur_pic_ptr->reference != DELAYED_PIC_REF);
1515 /* figure out if we have a complementary field pair */
1516 if (!FIELD_PICTURE(h) || h->picture_structure == last_pic_structure) {
1517 /* Previous field is unmatched. Don't display it, but let it
1518 * remain for reference if marked as such. */
1519 h0->cur_pic_ptr = NULL;
1520 h0->first_field = FIELD_PICTURE(h);
1522 if (h0->cur_pic_ptr->frame_num != h->frame_num) {
1523 /* This and the previous field had different frame_nums.
1524 * Consider this field first in pair. Throw away previous
1525 * one except for reference purposes. */
1526 h0->first_field = 1;
1527 h0->cur_pic_ptr = NULL;
1529 /* Second field in complementary pair */
1530 h0->first_field = 0;
1534 /* Frame or first field in a potentially complementary pair */
1535 h0->first_field = FIELD_PICTURE(h);
1538 if (!FIELD_PICTURE(h) || h0->first_field) {
1539 if (h264_frame_start(h) < 0) {
1540 h0->first_field = 0;
1541 return AVERROR_INVALIDDATA;
1544 release_unused_pictures(h, 0);
1547 if (h != h0 && (ret = clone_slice(h, h0)) < 0)
1550 h->cur_pic_ptr->frame_num = h->frame_num; // FIXME frame_num cleanup
1552 assert(h->mb_num == h->mb_width * h->mb_height);
1553 if (first_mb_in_slice << FIELD_OR_MBAFF_PICTURE(h) >= h->mb_num ||
1554 first_mb_in_slice >= h->mb_num) {
1555 av_log(h->avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
1556 return AVERROR_INVALIDDATA;
1558 h->resync_mb_x = h->mb_x = first_mb_in_slice % h->mb_width;
1559 h->resync_mb_y = h->mb_y = (first_mb_in_slice / h->mb_width) <<
1560 FIELD_OR_MBAFF_PICTURE(h);
1561 if (h->picture_structure == PICT_BOTTOM_FIELD)
1562 h->resync_mb_y = h->mb_y = h->mb_y + 1;
1563 assert(h->mb_y < h->mb_height);
1565 if (h->picture_structure == PICT_FRAME) {
1566 h->curr_pic_num = h->frame_num;
1567 h->max_pic_num = 1 << h->sps.log2_max_frame_num;
1569 h->curr_pic_num = 2 * h->frame_num + 1;
1570 h->max_pic_num = 1 << (h->sps.log2_max_frame_num + 1);
1573 if (h->nal_unit_type == NAL_IDR_SLICE)
1574 get_ue_golomb(&h->gb); /* idr_pic_id */
1576 if (h->sps.poc_type == 0) {
1577 h->poc_lsb = get_bits(&h->gb, h->sps.log2_max_poc_lsb);
1579 if (h->pps.pic_order_present == 1 && h->picture_structure == PICT_FRAME)
1580 h->delta_poc_bottom = get_se_golomb(&h->gb);
1583 if (h->sps.poc_type == 1 && !h->sps.delta_pic_order_always_zero_flag) {
1584 h->delta_poc[0] = get_se_golomb(&h->gb);
1586 if (h->pps.pic_order_present == 1 && h->picture_structure == PICT_FRAME)
1587 h->delta_poc[1] = get_se_golomb(&h->gb);
1590 ff_init_poc(h, h->cur_pic_ptr->field_poc, &h->cur_pic_ptr->poc);
1592 if (h->pps.redundant_pic_cnt_present)
1593 h->redundant_pic_count = get_ue_golomb(&h->gb);
1595 ret = ff_set_ref_count(h);
1599 default_ref_list_done = 0;
1601 if (!default_ref_list_done)
1602 ff_h264_fill_default_ref_list(h);
1604 if (h->slice_type_nos != AV_PICTURE_TYPE_I) {
1605 ret = ff_h264_decode_ref_pic_list_reordering(h);
1607 h->ref_count[1] = h->ref_count[0] = 0;
1612 if ((h->pps.weighted_pred && h->slice_type_nos == AV_PICTURE_TYPE_P) ||
1613 (h->pps.weighted_bipred_idc == 1 &&
1614 h->slice_type_nos == AV_PICTURE_TYPE_B))
1615 ff_pred_weight_table(h, sl);
1616 else if (h->pps.weighted_bipred_idc == 2 &&
1617 h->slice_type_nos == AV_PICTURE_TYPE_B) {
1618 implicit_weight_table(h, sl, -1);
1621 for (i = 0; i < 2; i++) {
1622 sl->luma_weight_flag[i] = 0;
1623 sl->chroma_weight_flag[i] = 0;
1627 // If frame-mt is enabled, only update mmco tables for the first slice
1628 // in a field. Subsequent slices can temporarily clobber h->mmco_index
1629 // or h->mmco, which will cause ref list mix-ups and decoding errors
1630 // further down the line. This may break decoding if the first slice is
1631 // corrupt, thus we only do this if frame-mt is enabled.
1632 if (h->nal_ref_idc) {
1633 ret = ff_h264_decode_ref_pic_marking(h0, &h->gb,
1634 !(h->avctx->active_thread_type & FF_THREAD_FRAME) ||
1635 h0->current_slice == 0);
1636 if (ret < 0 && (h->avctx->err_recognition & AV_EF_EXPLODE))
1637 return AVERROR_INVALIDDATA;
1640 if (FRAME_MBAFF(h)) {
1641 ff_h264_fill_mbaff_ref_list(h, sl);
1643 if (h->pps.weighted_bipred_idc == 2 && h->slice_type_nos == AV_PICTURE_TYPE_B) {
1644 implicit_weight_table(h, sl, 0);
1645 implicit_weight_table(h, sl, 1);
1649 if (h->slice_type_nos == AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
1650 ff_h264_direct_dist_scale_factor(h);
1651 ff_h264_direct_ref_list_init(h);
1653 if (h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac) {
1654 tmp = get_ue_golomb_31(&h->gb);
1656 av_log(h->avctx, AV_LOG_ERROR, "cabac_init_idc %u overflow\n", tmp);
1657 return AVERROR_INVALIDDATA;
1659 h->cabac_init_idc = tmp;
1662 h->last_qscale_diff = 0;
1663 tmp = h->pps.init_qp + get_se_golomb(&h->gb);
1664 if (tmp > 51 + 6 * (h->sps.bit_depth_luma - 8)) {
1665 av_log(h->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
1666 return AVERROR_INVALIDDATA;
1669 h->chroma_qp[0] = get_chroma_qp(h, 0, h->qscale);
1670 h->chroma_qp[1] = get_chroma_qp(h, 1, h->qscale);
1671 // FIXME qscale / qp ... stuff
1672 if (h->slice_type == AV_PICTURE_TYPE_SP)
1673 get_bits1(&h->gb); /* sp_for_switch_flag */
1674 if (h->slice_type == AV_PICTURE_TYPE_SP ||
1675 h->slice_type == AV_PICTURE_TYPE_SI)
1676 get_se_golomb(&h->gb); /* slice_qs_delta */
1678 h->deblocking_filter = 1;
1679 h->slice_alpha_c0_offset = 0;
1680 h->slice_beta_offset = 0;
1681 if (h->pps.deblocking_filter_parameters_present) {
1682 tmp = get_ue_golomb_31(&h->gb);
1684 av_log(h->avctx, AV_LOG_ERROR,
1685 "deblocking_filter_idc %u out of range\n", tmp);
1686 return AVERROR_INVALIDDATA;
1688 h->deblocking_filter = tmp;
1689 if (h->deblocking_filter < 2)
1690 h->deblocking_filter ^= 1; // 1<->0
1692 if (h->deblocking_filter) {
1693 h->slice_alpha_c0_offset = get_se_golomb(&h->gb) * 2;
1694 h->slice_beta_offset = get_se_golomb(&h->gb) * 2;
1695 if (h->slice_alpha_c0_offset > 12 ||
1696 h->slice_alpha_c0_offset < -12 ||
1697 h->slice_beta_offset > 12 ||
1698 h->slice_beta_offset < -12) {
1699 av_log(h->avctx, AV_LOG_ERROR,
1700 "deblocking filter parameters %d %d out of range\n",
1701 h->slice_alpha_c0_offset, h->slice_beta_offset);
1702 return AVERROR_INVALIDDATA;
1707 if (h->avctx->skip_loop_filter >= AVDISCARD_ALL ||
1708 (h->avctx->skip_loop_filter >= AVDISCARD_NONKEY &&
1709 h->slice_type_nos != AV_PICTURE_TYPE_I) ||
1710 (h->avctx->skip_loop_filter >= AVDISCARD_BIDIR &&
1711 h->slice_type_nos == AV_PICTURE_TYPE_B) ||
1712 (h->avctx->skip_loop_filter >= AVDISCARD_NONREF &&
1713 h->nal_ref_idc == 0))
1714 h->deblocking_filter = 0;
1716 if (h->deblocking_filter == 1 && h0->max_contexts > 1) {
1717 if (h->avctx->flags2 & CODEC_FLAG2_FAST) {
1718 /* Cheat slightly for speed:
1719 * Do not bother to deblock across slices. */
1720 h->deblocking_filter = 2;
1722 h0->max_contexts = 1;
1723 if (!h0->single_decode_warning) {
1724 av_log(h->avctx, AV_LOG_INFO,
1725 "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
1726 h0->single_decode_warning = 1;
1729 av_log(h->avctx, AV_LOG_ERROR,
1730 "Deblocking switched inside frame.\n");
1736 FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset) -
1738 h->pps.chroma_qp_index_offset[0],
1739 h->pps.chroma_qp_index_offset[1]) +
1740 6 * (h->sps.bit_depth_luma - 8);
1742 h0->last_slice_type = slice_type;
1743 h->slice_num = ++h0->current_slice;
1744 if (h->slice_num >= MAX_SLICES) {
1745 av_log(h->avctx, AV_LOG_ERROR,
1746 "Too many slices, increase MAX_SLICES and recompile\n");
1749 for (j = 0; j < 2; j++) {
1751 int *ref2frm = h->ref2frm[h->slice_num & (MAX_SLICES - 1)][j];
1752 for (i = 0; i < 16; i++) {
1754 if (j < h->list_count && i < h->ref_count[j] &&
1755 h->ref_list[j][i].f.buf[0]) {
1757 AVBuffer *buf = h->ref_list[j][i].f.buf[0]->buffer;
1758 for (k = 0; k < h->short_ref_count; k++)
1759 if (h->short_ref[k]->f.buf[0]->buffer == buf) {
1763 for (k = 0; k < h->long_ref_count; k++)
1764 if (h->long_ref[k] && h->long_ref[k]->f.buf[0]->buffer == buf) {
1765 id_list[i] = h->short_ref_count + k;
1773 for (i = 0; i < 16; i++)
1774 ref2frm[i + 2] = 4 * id_list[i] + (h->ref_list[j][i].reference & 3);
1776 ref2frm[18 + 1] = -1;
1777 for (i = 16; i < 48; i++)
1778 ref2frm[i + 4] = 4 * id_list[(i - 16) >> 1] +
1779 (h->ref_list[j][i].reference & 3);
1782 if (h->avctx->debug & FF_DEBUG_PICT_INFO) {
1783 av_log(h->avctx, AV_LOG_DEBUG,
1784 "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",
1786 (h->picture_structure == PICT_FRAME ? "F" : h->picture_structure == PICT_TOP_FIELD ? "T" : "B"),
1788 av_get_picture_type_char(h->slice_type),
1789 h->slice_type_fixed ? " fix" : "",
1790 h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
1791 pps_id, h->frame_num,
1792 h->cur_pic_ptr->field_poc[0],
1793 h->cur_pic_ptr->field_poc[1],
1794 h->ref_count[0], h->ref_count[1],
1796 h->deblocking_filter,
1797 h->slice_alpha_c0_offset, h->slice_beta_offset,
1799 sl->use_weight == 1 && sl->use_weight_chroma ? "c" : "",
1800 h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : "");
1806 int ff_h264_get_slice_type(const H264Context *h)
1808 switch (h->slice_type) {
1809 case AV_PICTURE_TYPE_P:
1811 case AV_PICTURE_TYPE_B:
1813 case AV_PICTURE_TYPE_I:
1815 case AV_PICTURE_TYPE_SP:
1817 case AV_PICTURE_TYPE_SI:
1820 return AVERROR_INVALIDDATA;
1824 static av_always_inline void fill_filter_caches_inter(H264Context *h,
1825 int mb_type, int top_xy,
1826 int left_xy[LEFT_MBS],
1828 int left_type[LEFT_MBS],
1829 int mb_xy, int list)
1831 int b_stride = h->b_stride;
1832 int16_t(*mv_dst)[2] = &h->mv_cache[list][scan8[0]];
1833 int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
1834 if (IS_INTER(mb_type) || IS_DIRECT(mb_type)) {
1835 if (USES_LIST(top_type, list)) {
1836 const int b_xy = h->mb2b_xy[top_xy] + 3 * b_stride;
1837 const int b8_xy = 4 * top_xy + 2;
1838 int (*ref2frm)[64] = h->ref2frm[h->slice_table[top_xy] & (MAX_SLICES - 1)][0] + (MB_MBAFF(h) ? 20 : 2);
1839 AV_COPY128(mv_dst - 1 * 8, h->cur_pic.motion_val[list][b_xy + 0]);
1840 ref_cache[0 - 1 * 8] =
1841 ref_cache[1 - 1 * 8] = ref2frm[list][h->cur_pic.ref_index[list][b8_xy + 0]];
1842 ref_cache[2 - 1 * 8] =
1843 ref_cache[3 - 1 * 8] = ref2frm[list][h->cur_pic.ref_index[list][b8_xy + 1]];
1845 AV_ZERO128(mv_dst - 1 * 8);
1846 AV_WN32A(&ref_cache[0 - 1 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
1849 if (!IS_INTERLACED(mb_type ^ left_type[LTOP])) {
1850 if (USES_LIST(left_type[LTOP], list)) {
1851 const int b_xy = h->mb2b_xy[left_xy[LTOP]] + 3;
1852 const int b8_xy = 4 * left_xy[LTOP] + 1;
1853 int (*ref2frm)[64] = h->ref2frm[h->slice_table[left_xy[LTOP]] & (MAX_SLICES - 1)][0] + (MB_MBAFF(h) ? 20 : 2);
1854 AV_COPY32(mv_dst - 1 + 0, h->cur_pic.motion_val[list][b_xy + b_stride * 0]);
1855 AV_COPY32(mv_dst - 1 + 8, h->cur_pic.motion_val[list][b_xy + b_stride * 1]);
1856 AV_COPY32(mv_dst - 1 + 16, h->cur_pic.motion_val[list][b_xy + b_stride * 2]);
1857 AV_COPY32(mv_dst - 1 + 24, h->cur_pic.motion_val[list][b_xy + b_stride * 3]);
1859 ref_cache[-1 + 8] = ref2frm[list][h->cur_pic.ref_index[list][b8_xy + 2 * 0]];
1860 ref_cache[-1 + 16] =
1861 ref_cache[-1 + 24] = ref2frm[list][h->cur_pic.ref_index[list][b8_xy + 2 * 1]];
1863 AV_ZERO32(mv_dst - 1 + 0);
1864 AV_ZERO32(mv_dst - 1 + 8);
1865 AV_ZERO32(mv_dst - 1 + 16);
1866 AV_ZERO32(mv_dst - 1 + 24);
1869 ref_cache[-1 + 16] =
1870 ref_cache[-1 + 24] = LIST_NOT_USED;
1875 if (!USES_LIST(mb_type, list)) {
1876 fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0, 0), 4);
1877 AV_WN32A(&ref_cache[0 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
1878 AV_WN32A(&ref_cache[1 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
1879 AV_WN32A(&ref_cache[2 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
1880 AV_WN32A(&ref_cache[3 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
1885 int8_t *ref = &h->cur_pic.ref_index[list][4 * mb_xy];
1886 int (*ref2frm)[64] = h->ref2frm[h->slice_num & (MAX_SLICES - 1)][0] + (MB_MBAFF(h) ? 20 : 2);
1887 uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]], ref2frm[list][ref[1]]) & 0x00FF00FF) * 0x0101;
1888 uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]], ref2frm[list][ref[3]]) & 0x00FF00FF) * 0x0101;
1889 AV_WN32A(&ref_cache[0 * 8], ref01);
1890 AV_WN32A(&ref_cache[1 * 8], ref01);
1891 AV_WN32A(&ref_cache[2 * 8], ref23);
1892 AV_WN32A(&ref_cache[3 * 8], ref23);
1896 int16_t(*mv_src)[2] = &h->cur_pic.motion_val[list][4 * h->mb_x + 4 * h->mb_y * b_stride];
1897 AV_COPY128(mv_dst + 8 * 0, mv_src + 0 * b_stride);
1898 AV_COPY128(mv_dst + 8 * 1, mv_src + 1 * b_stride);
1899 AV_COPY128(mv_dst + 8 * 2, mv_src + 2 * b_stride);
1900 AV_COPY128(mv_dst + 8 * 3, mv_src + 3 * b_stride);
1906 * @return non zero if the loop filter can be skipped
1908 static int fill_filter_caches(H264Context *h, int mb_type)
1910 const int mb_xy = h->mb_xy;
1911 int top_xy, left_xy[LEFT_MBS];
1912 int top_type, left_type[LEFT_MBS];
1916 top_xy = mb_xy - (h->mb_stride << MB_FIELD(h));
1918 /* Wow, what a mess, why didn't they simplify the interlacing & intra
1919 * stuff, I can't imagine that these complex rules are worth it. */
1921 left_xy[LBOT] = left_xy[LTOP] = mb_xy - 1;
1922 if (FRAME_MBAFF(h)) {
1923 const int left_mb_field_flag = IS_INTERLACED(h->cur_pic.mb_type[mb_xy - 1]);
1924 const int curr_mb_field_flag = IS_INTERLACED(mb_type);
1926 if (left_mb_field_flag != curr_mb_field_flag)
1927 left_xy[LTOP] -= h->mb_stride;
1929 if (curr_mb_field_flag)
1930 top_xy += h->mb_stride &
1931 (((h->cur_pic.mb_type[top_xy] >> 7) & 1) - 1);
1932 if (left_mb_field_flag != curr_mb_field_flag)
1933 left_xy[LBOT] += h->mb_stride;
1937 h->top_mb_xy = top_xy;
1938 h->left_mb_xy[LTOP] = left_xy[LTOP];
1939 h->left_mb_xy[LBOT] = left_xy[LBOT];
1941 /* For sufficiently low qp, filtering wouldn't do anything.
1942 * This is a conservative estimate: could also check beta_offset
1943 * and more accurate chroma_qp. */
1944 int qp_thresh = h->qp_thresh; // FIXME strictly we should store qp_thresh for each mb of a slice
1945 int qp = h->cur_pic.qscale_table[mb_xy];
1946 if (qp <= qp_thresh &&
1947 (left_xy[LTOP] < 0 ||
1948 ((qp + h->cur_pic.qscale_table[left_xy[LTOP]] + 1) >> 1) <= qp_thresh) &&
1950 ((qp + h->cur_pic.qscale_table[top_xy] + 1) >> 1) <= qp_thresh)) {
1951 if (!FRAME_MBAFF(h))
1953 if ((left_xy[LTOP] < 0 ||
1954 ((qp + h->cur_pic.qscale_table[left_xy[LBOT]] + 1) >> 1) <= qp_thresh) &&
1955 (top_xy < h->mb_stride ||
1956 ((qp + h->cur_pic.qscale_table[top_xy - h->mb_stride] + 1) >> 1) <= qp_thresh))
1961 top_type = h->cur_pic.mb_type[top_xy];
1962 left_type[LTOP] = h->cur_pic.mb_type[left_xy[LTOP]];
1963 left_type[LBOT] = h->cur_pic.mb_type[left_xy[LBOT]];
1964 if (h->deblocking_filter == 2) {
1965 if (h->slice_table[top_xy] != h->slice_num)
1967 if (h->slice_table[left_xy[LBOT]] != h->slice_num)
1968 left_type[LTOP] = left_type[LBOT] = 0;
1970 if (h->slice_table[top_xy] == 0xFFFF)
1972 if (h->slice_table[left_xy[LBOT]] == 0xFFFF)
1973 left_type[LTOP] = left_type[LBOT] = 0;
1975 h->top_type = top_type;
1976 h->left_type[LTOP] = left_type[LTOP];
1977 h->left_type[LBOT] = left_type[LBOT];
1979 if (IS_INTRA(mb_type))
1982 fill_filter_caches_inter(h, mb_type, top_xy, left_xy,
1983 top_type, left_type, mb_xy, 0);
1984 if (h->list_count == 2)
1985 fill_filter_caches_inter(h, mb_type, top_xy, left_xy,
1986 top_type, left_type, mb_xy, 1);
1988 nnz = h->non_zero_count[mb_xy];
1989 nnz_cache = h->non_zero_count_cache;
1990 AV_COPY32(&nnz_cache[4 + 8 * 1], &nnz[0]);
1991 AV_COPY32(&nnz_cache[4 + 8 * 2], &nnz[4]);
1992 AV_COPY32(&nnz_cache[4 + 8 * 3], &nnz[8]);
1993 AV_COPY32(&nnz_cache[4 + 8 * 4], &nnz[12]);
1994 h->cbp = h->cbp_table[mb_xy];
1997 nnz = h->non_zero_count[top_xy];
1998 AV_COPY32(&nnz_cache[4 + 8 * 0], &nnz[3 * 4]);
2001 if (left_type[LTOP]) {
2002 nnz = h->non_zero_count[left_xy[LTOP]];
2003 nnz_cache[3 + 8 * 1] = nnz[3 + 0 * 4];
2004 nnz_cache[3 + 8 * 2] = nnz[3 + 1 * 4];
2005 nnz_cache[3 + 8 * 3] = nnz[3 + 2 * 4];
2006 nnz_cache[3 + 8 * 4] = nnz[3 + 3 * 4];
2009 /* CAVLC 8x8dct requires NNZ values for residual decoding that differ
2010 * from what the loop filter needs */
2011 if (!CABAC(h) && h->pps.transform_8x8_mode) {
2012 if (IS_8x8DCT(top_type)) {
2013 nnz_cache[4 + 8 * 0] =
2014 nnz_cache[5 + 8 * 0] = (h->cbp_table[top_xy] & 0x4000) >> 12;
2015 nnz_cache[6 + 8 * 0] =
2016 nnz_cache[7 + 8 * 0] = (h->cbp_table[top_xy] & 0x8000) >> 12;
2018 if (IS_8x8DCT(left_type[LTOP])) {
2019 nnz_cache[3 + 8 * 1] =
2020 nnz_cache[3 + 8 * 2] = (h->cbp_table[left_xy[LTOP]] & 0x2000) >> 12; // FIXME check MBAFF
2022 if (IS_8x8DCT(left_type[LBOT])) {
2023 nnz_cache[3 + 8 * 3] =
2024 nnz_cache[3 + 8 * 4] = (h->cbp_table[left_xy[LBOT]] & 0x8000) >> 12; // FIXME check MBAFF
2027 if (IS_8x8DCT(mb_type)) {
2028 nnz_cache[scan8[0]] =
2029 nnz_cache[scan8[1]] =
2030 nnz_cache[scan8[2]] =
2031 nnz_cache[scan8[3]] = (h->cbp & 0x1000) >> 12;
2033 nnz_cache[scan8[0 + 4]] =
2034 nnz_cache[scan8[1 + 4]] =
2035 nnz_cache[scan8[2 + 4]] =
2036 nnz_cache[scan8[3 + 4]] = (h->cbp & 0x2000) >> 12;
2038 nnz_cache[scan8[0 + 8]] =
2039 nnz_cache[scan8[1 + 8]] =
2040 nnz_cache[scan8[2 + 8]] =
2041 nnz_cache[scan8[3 + 8]] = (h->cbp & 0x4000) >> 12;
2043 nnz_cache[scan8[0 + 12]] =
2044 nnz_cache[scan8[1 + 12]] =
2045 nnz_cache[scan8[2 + 12]] =
2046 nnz_cache[scan8[3 + 12]] = (h->cbp & 0x8000) >> 12;
2053 static void loop_filter(H264Context *h, int start_x, int end_x)
2055 uint8_t *dest_y, *dest_cb, *dest_cr;
2056 int linesize, uvlinesize, mb_x, mb_y;
2057 const int end_mb_y = h->mb_y + FRAME_MBAFF(h);
2058 const int old_slice_type = h->slice_type;
2059 const int pixel_shift = h->pixel_shift;
2060 const int block_h = 16 >> h->chroma_y_shift;
2062 if (h->deblocking_filter) {
2063 for (mb_x = start_x; mb_x < end_x; mb_x++)
2064 for (mb_y = end_mb_y - FRAME_MBAFF(h); mb_y <= end_mb_y; mb_y++) {
2066 mb_xy = h->mb_xy = mb_x + mb_y * h->mb_stride;
2067 h->slice_num = h->slice_table[mb_xy];
2068 mb_type = h->cur_pic.mb_type[mb_xy];
2069 h->list_count = h->list_counts[mb_xy];
2073 h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
2077 dest_y = h->cur_pic.f.data[0] +
2078 ((mb_x << pixel_shift) + mb_y * h->linesize) * 16;
2079 dest_cb = h->cur_pic.f.data[1] +
2080 (mb_x << pixel_shift) * (8 << CHROMA444(h)) +
2081 mb_y * h->uvlinesize * block_h;
2082 dest_cr = h->cur_pic.f.data[2] +
2083 (mb_x << pixel_shift) * (8 << CHROMA444(h)) +
2084 mb_y * h->uvlinesize * block_h;
2085 // FIXME simplify above
2088 linesize = h->mb_linesize = h->linesize * 2;
2089 uvlinesize = h->mb_uvlinesize = h->uvlinesize * 2;
2090 if (mb_y & 1) { // FIXME move out of this function?
2091 dest_y -= h->linesize * 15;
2092 dest_cb -= h->uvlinesize * (block_h - 1);
2093 dest_cr -= h->uvlinesize * (block_h - 1);
2096 linesize = h->mb_linesize = h->linesize;
2097 uvlinesize = h->mb_uvlinesize = h->uvlinesize;
2099 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize,
2101 if (fill_filter_caches(h, mb_type))
2103 h->chroma_qp[0] = get_chroma_qp(h, 0, h->cur_pic.qscale_table[mb_xy]);
2104 h->chroma_qp[1] = get_chroma_qp(h, 1, h->cur_pic.qscale_table[mb_xy]);
2106 if (FRAME_MBAFF(h)) {
2107 ff_h264_filter_mb(h, mb_x, mb_y, dest_y, dest_cb, dest_cr,
2108 linesize, uvlinesize);
2110 ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb,
2111 dest_cr, linesize, uvlinesize);
2115 h->slice_type = old_slice_type;
2117 h->mb_y = end_mb_y - FRAME_MBAFF(h);
2118 h->chroma_qp[0] = get_chroma_qp(h, 0, h->qscale);
2119 h->chroma_qp[1] = get_chroma_qp(h, 1, h->qscale);
2122 static void predict_field_decoding_flag(H264Context *h)
2124 const int mb_xy = h->mb_x + h->mb_y * h->mb_stride;
2125 int mb_type = (h->slice_table[mb_xy - 1] == h->slice_num) ?
2126 h->cur_pic.mb_type[mb_xy - 1] :
2127 (h->slice_table[mb_xy - h->mb_stride] == h->slice_num) ?
2128 h->cur_pic.mb_type[mb_xy - h->mb_stride] : 0;
2129 h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
2133 * Draw edges and report progress for the last MB row.
2135 static void decode_finish_row(H264Context *h)
2137 int top = 16 * (h->mb_y >> FIELD_PICTURE(h));
2138 int pic_height = 16 * h->mb_height >> FIELD_PICTURE(h);
2139 int height = 16 << FRAME_MBAFF(h);
2140 int deblock_border = (16 + 4) << FRAME_MBAFF(h);
2142 if (h->deblocking_filter) {
2143 if ((top + height) >= pic_height)
2144 height += deblock_border;
2145 top -= deblock_border;
2148 if (top >= pic_height || (top + height) < 0)
2151 height = FFMIN(height, pic_height - top);
2153 height = top + height;
2157 ff_h264_draw_horiz_band(h, top, height);
2162 ff_thread_report_progress(&h->cur_pic_ptr->tf, top + height - 1,
2163 h->picture_structure == PICT_BOTTOM_FIELD);
2166 static void er_add_slice(H264Context *h, int startx, int starty,
2167 int endx, int endy, int status)
2169 #if CONFIG_ERROR_RESILIENCE
2170 ERContext *er = &h->er;
2172 er->ref_count = h->ref_count[0];
2173 ff_er_add_slice(er, startx, starty, endx, endy, status);
2177 static int decode_slice(struct AVCodecContext *avctx, void *arg)
2179 H264SliceContext *sl = arg;
2180 H264Context *h = sl->h264;
2181 int lf_x_start = h->mb_x;
2183 h->mb_skip_run = -1;
2185 h->is_complex = FRAME_MBAFF(h) || h->picture_structure != PICT_FRAME ||
2186 avctx->codec_id != AV_CODEC_ID_H264 ||
2187 (CONFIG_GRAY && (h->flags & CODEC_FLAG_GRAY));
2191 align_get_bits(&h->gb);
2194 ff_init_cabac_decoder(&h->cabac,
2195 h->gb.buffer + get_bits_count(&h->gb) / 8,
2196 (get_bits_left(&h->gb) + 7) / 8);
2198 ff_h264_init_cabac_states(h);
2202 int ret = ff_h264_decode_mb_cabac(h);
2204 // STOP_TIMER("decode_mb_cabac")
2207 ff_h264_hl_decode_mb(h, sl);
2209 // FIXME optimal? or let mb_decode decode 16x32 ?
2210 if (ret >= 0 && FRAME_MBAFF(h)) {
2213 ret = ff_h264_decode_mb_cabac(h);
2216 ff_h264_hl_decode_mb(h, sl);
2219 eos = get_cabac_terminate(&h->cabac);
2221 if ((h->workaround_bugs & FF_BUG_TRUNCATED) &&
2222 h->cabac.bytestream > h->cabac.bytestream_end + 2) {
2223 er_add_slice(h, h->resync_mb_x, h->resync_mb_y, h->mb_x - 1,
2224 h->mb_y, ER_MB_END);
2225 if (h->mb_x >= lf_x_start)
2226 loop_filter(h, lf_x_start, h->mb_x + 1);
2229 if (ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 2) {
2230 av_log(h->avctx, AV_LOG_ERROR,
2231 "error while decoding MB %d %d, bytestream %td\n",
2233 h->cabac.bytestream_end - h->cabac.bytestream);
2234 er_add_slice(h, h->resync_mb_x, h->resync_mb_y, h->mb_x,
2235 h->mb_y, ER_MB_ERROR);
2236 return AVERROR_INVALIDDATA;
2239 if (++h->mb_x >= h->mb_width) {
2240 loop_filter(h, lf_x_start, h->mb_x);
2241 h->mb_x = lf_x_start = 0;
2242 decode_finish_row(h);
2244 if (FIELD_OR_MBAFF_PICTURE(h)) {
2246 if (FRAME_MBAFF(h) && h->mb_y < h->mb_height)
2247 predict_field_decoding_flag(h);
2251 if (eos || h->mb_y >= h->mb_height) {
2252 tprintf(h->avctx, "slice end %d %d\n",
2253 get_bits_count(&h->gb), h->gb.size_in_bits);
2254 er_add_slice(h, h->resync_mb_x, h->resync_mb_y, h->mb_x - 1,
2255 h->mb_y, ER_MB_END);
2256 if (h->mb_x > lf_x_start)
2257 loop_filter(h, lf_x_start, h->mb_x);
2263 int ret = ff_h264_decode_mb_cavlc(h);
2266 ff_h264_hl_decode_mb(h, sl);
2268 // FIXME optimal? or let mb_decode decode 16x32 ?
2269 if (ret >= 0 && FRAME_MBAFF(h)) {
2271 ret = ff_h264_decode_mb_cavlc(h);
2274 ff_h264_hl_decode_mb(h, sl);
2279 av_log(h->avctx, AV_LOG_ERROR,
2280 "error while decoding MB %d %d\n", h->mb_x, h->mb_y);
2281 er_add_slice(h, h->resync_mb_x, h->resync_mb_y, h->mb_x,
2282 h->mb_y, ER_MB_ERROR);
2286 if (++h->mb_x >= h->mb_width) {
2287 loop_filter(h, lf_x_start, h->mb_x);
2288 h->mb_x = lf_x_start = 0;
2289 decode_finish_row(h);
2291 if (FIELD_OR_MBAFF_PICTURE(h)) {
2293 if (FRAME_MBAFF(h) && h->mb_y < h->mb_height)
2294 predict_field_decoding_flag(h);
2296 if (h->mb_y >= h->mb_height) {
2297 tprintf(h->avctx, "slice end %d %d\n",
2298 get_bits_count(&h->gb), h->gb.size_in_bits);
2300 if (get_bits_left(&h->gb) == 0) {
2301 er_add_slice(h, h->resync_mb_x, h->resync_mb_y,
2302 h->mb_x - 1, h->mb_y, ER_MB_END);
2306 er_add_slice(h, h->resync_mb_x, h->resync_mb_y,
2307 h->mb_x - 1, h->mb_y, ER_MB_END);
2309 return AVERROR_INVALIDDATA;
2314 if (get_bits_left(&h->gb) <= 0 && h->mb_skip_run <= 0) {
2315 tprintf(h->avctx, "slice end %d %d\n",
2316 get_bits_count(&h->gb), h->gb.size_in_bits);
2318 if (get_bits_left(&h->gb) == 0) {
2319 er_add_slice(h, h->resync_mb_x, h->resync_mb_y,
2320 h->mb_x - 1, h->mb_y, ER_MB_END);
2321 if (h->mb_x > lf_x_start)
2322 loop_filter(h, lf_x_start, h->mb_x);
2326 er_add_slice(h, h->resync_mb_x, h->resync_mb_y, h->mb_x,
2327 h->mb_y, ER_MB_ERROR);
2329 return AVERROR_INVALIDDATA;
2337 * Call decode_slice() for each context.
2339 * @param h h264 master context
2340 * @param context_count number of contexts to execute
2342 int ff_h264_execute_decode_slices(H264Context *h, unsigned context_count)
2344 AVCodecContext *const avctx = h->avctx;
2348 if (h->avctx->hwaccel)
2350 if (context_count == 1) {
2351 return decode_slice(avctx, &h->slice_ctx[0]);
2353 for (i = 1; i < context_count; i++) {
2354 hx = h->thread_context[i];
2355 hx->er.error_count = 0;
2358 avctx->execute(avctx, decode_slice, h->slice_ctx,
2359 NULL, context_count, sizeof(h->slice_ctx[0]));
2361 /* pull back stuff from slices to master context */
2362 hx = h->thread_context[context_count - 1];
2365 h->droppable = hx->droppable;
2366 h->picture_structure = hx->picture_structure;
2367 for (i = 1; i < context_count; i++)
2368 h->er.error_count += h->thread_context[i]->er.error_count;