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, H264SliceContext *sl,
367 H264Context *src, int i)
369 sl->intra4x4_pred_mode = src->intra4x4_pred_mode + i * 8 * 2 * src->mb_stride;
371 dst->non_zero_count = src->non_zero_count;
372 dst->slice_table = src->slice_table;
373 dst->cbp_table = src->cbp_table;
374 dst->mb2b_xy = src->mb2b_xy;
375 dst->mb2br_xy = src->mb2br_xy;
376 dst->chroma_pred_mode_table = src->chroma_pred_mode_table;
377 dst->mvd_table[0] = src->mvd_table[0] + i * 8 * 2 * src->mb_stride;
378 dst->mvd_table[1] = src->mvd_table[1] + i * 8 * 2 * src->mb_stride;
379 dst->direct_table = src->direct_table;
380 dst->list_counts = src->list_counts;
382 dst->cur_pic_ptr = src->cur_pic_ptr;
383 dst->cur_pic = src->cur_pic;
384 dst->bipred_scratchpad = NULL;
385 dst->edge_emu_buffer = NULL;
386 ff_h264_pred_init(&dst->hpc, src->avctx->codec_id, src->sps.bit_depth_luma,
387 src->sps.chroma_format_idc);
390 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b) + (size))))
392 #define REBASE_PICTURE(pic, new_ctx, old_ctx) \
393 ((pic && pic >= old_ctx->DPB && \
394 pic < old_ctx->DPB + H264_MAX_PICTURE_COUNT) ? \
395 &new_ctx->DPB[pic - old_ctx->DPB] : NULL)
397 static void copy_picture_range(H264Picture **to, H264Picture **from, int count,
398 H264Context *new_base,
399 H264Context *old_base)
403 for (i = 0; i < count; i++) {
404 assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
405 IN_RANGE(from[i], old_base->DPB,
406 sizeof(H264Picture) * H264_MAX_PICTURE_COUNT) ||
408 to[i] = REBASE_PICTURE(from[i], new_base, old_base);
412 static int copy_parameter_set(void **to, void **from, int count, int size)
416 for (i = 0; i < count; i++) {
417 if (to[i] && !from[i]) {
419 } else if (from[i] && !to[i]) {
420 to[i] = av_malloc(size);
422 return AVERROR(ENOMEM);
426 memcpy(to[i], from[i], size);
432 #define copy_fields(to, from, start_field, end_field) \
433 memcpy(&to->start_field, &from->start_field, \
434 (char *)&to->end_field - (char *)&to->start_field)
436 static int h264_slice_header_init(H264Context *h, int reinit);
438 int ff_h264_update_thread_context(AVCodecContext *dst,
439 const AVCodecContext *src)
441 H264Context *h = dst->priv_data, *h1 = src->priv_data;
442 int inited = h->context_initialized, err = 0;
443 int context_reinitialized = 0;
446 if (dst == src || !h1->context_initialized)
450 (h->width != h1->width ||
451 h->height != h1->height ||
452 h->mb_width != h1->mb_width ||
453 h->mb_height != h1->mb_height ||
454 h->sps.bit_depth_luma != h1->sps.bit_depth_luma ||
455 h->sps.chroma_format_idc != h1->sps.chroma_format_idc ||
456 h->sps.colorspace != h1->sps.colorspace)) {
458 /* set bits_per_raw_sample to the previous value. the check for changed
459 * bit depth in h264_set_parameter_from_sps() uses it and sets it to
460 * the current value */
461 h->avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
463 av_freep(&h->bipred_scratchpad);
465 h->width = h1->width;
466 h->height = h1->height;
467 h->mb_height = h1->mb_height;
468 h->mb_width = h1->mb_width;
469 h->mb_num = h1->mb_num;
470 h->mb_stride = h1->mb_stride;
471 h->b_stride = h1->b_stride;
473 if ((err = h264_slice_header_init(h, 1)) < 0) {
474 av_log(h->avctx, AV_LOG_ERROR, "h264_slice_header_init() failed");
477 context_reinitialized = 1;
479 /* update linesize on resize. The decoder doesn't
480 * necessarily call h264_frame_start in the new thread */
481 h->linesize = h1->linesize;
482 h->uvlinesize = h1->uvlinesize;
484 /* copy block_offset since frame_start may not be called */
485 memcpy(h->block_offset, h1->block_offset, sizeof(h->block_offset));
489 H264SliceContext *orig_slice_ctx = h->slice_ctx;
491 for (i = 0; i < MAX_SPS_COUNT; i++)
492 av_freep(h->sps_buffers + i);
494 for (i = 0; i < MAX_PPS_COUNT; i++)
495 av_freep(h->pps_buffers + i);
497 memcpy(h, h1, sizeof(*h1));
498 memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
499 memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
500 memset(&h->er, 0, sizeof(h->er));
501 h->context_initialized = 0;
503 memset(&h->cur_pic, 0, sizeof(h->cur_pic));
504 av_frame_unref(&h->cur_pic.f);
505 h->cur_pic.tf.f = &h->cur_pic.f;
507 h->slice_ctx = orig_slice_ctx;
509 memset(&h->slice_ctx[0].mb, 0, sizeof(h->slice_ctx[0].mb));
510 memset(&h->slice_ctx[0].mb_luma_dc, 0, sizeof(h->slice_ctx[0].mb_luma_dc));
511 memset(&h->slice_ctx[0].mb_padding, 0, sizeof(h->slice_ctx[0].mb_padding));
515 h->qscale_table_pool = NULL;
516 h->mb_type_pool = NULL;
517 h->ref_index_pool = NULL;
518 h->motion_val_pool = NULL;
520 ret = ff_h264_alloc_tables(h);
522 av_log(dst, AV_LOG_ERROR, "Could not allocate memory\n");
525 ret = ff_h264_context_init(h);
527 av_log(dst, AV_LOG_ERROR, "context_init() failed.\n");
531 for (i = 0; i < 2; i++) {
532 h->rbsp_buffer[i] = NULL;
533 h->rbsp_buffer_size[i] = 0;
535 h->bipred_scratchpad = NULL;
536 h->edge_emu_buffer = NULL;
538 h->thread_context[0] = h;
540 h->context_initialized = 1;
543 h->avctx->coded_height = h1->avctx->coded_height;
544 h->avctx->coded_width = h1->avctx->coded_width;
545 h->avctx->width = h1->avctx->width;
546 h->avctx->height = h1->avctx->height;
547 h->coded_picture_number = h1->coded_picture_number;
548 h->first_field = h1->first_field;
549 h->picture_structure = h1->picture_structure;
550 h->droppable = h1->droppable;
551 h->low_delay = h1->low_delay;
553 for (i = 0; i < H264_MAX_PICTURE_COUNT; i++) {
554 ff_h264_unref_picture(h, &h->DPB[i]);
555 if (h1->DPB[i].f.buf[0] &&
556 (ret = ff_h264_ref_picture(h, &h->DPB[i], &h1->DPB[i])) < 0)
560 h->cur_pic_ptr = REBASE_PICTURE(h1->cur_pic_ptr, h, h1);
561 ff_h264_unref_picture(h, &h->cur_pic);
562 if (h1->cur_pic.f.buf[0]) {
563 ret = ff_h264_ref_picture(h, &h->cur_pic, &h1->cur_pic);
568 h->workaround_bugs = h1->workaround_bugs;
569 h->low_delay = h1->low_delay;
570 h->droppable = h1->droppable;
572 /* frame_start may not be called for the next thread (if it's decoding
573 * a bottom field) so this has to be allocated here */
574 err = alloc_scratch_buffers(h, h1->linesize);
578 // extradata/NAL handling
579 h->is_avc = h1->is_avc;
582 if ((ret = copy_parameter_set((void **)h->sps_buffers,
583 (void **)h1->sps_buffers,
584 MAX_SPS_COUNT, sizeof(SPS))) < 0)
587 if ((ret = copy_parameter_set((void **)h->pps_buffers,
588 (void **)h1->pps_buffers,
589 MAX_PPS_COUNT, sizeof(PPS))) < 0)
593 // Dequantization matrices
594 // FIXME these are big - can they be only copied when PPS changes?
595 copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
597 for (i = 0; i < 6; i++)
598 h->dequant4_coeff[i] = h->dequant4_buffer[0] +
599 (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
601 for (i = 0; i < 6; i++)
602 h->dequant8_coeff[i] = h->dequant8_buffer[0] +
603 (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
605 h->dequant_coeff_pps = h1->dequant_coeff_pps;
608 copy_fields(h, h1, poc_lsb, redundant_pic_count);
611 copy_fields(h, h1, short_ref, cabac_init_idc);
613 copy_picture_range(h->short_ref, h1->short_ref, 32, h, h1);
614 copy_picture_range(h->long_ref, h1->long_ref, 32, h, h1);
615 copy_picture_range(h->delayed_pic, h1->delayed_pic,
616 MAX_DELAYED_PIC_COUNT + 2, h, h1);
618 h->last_slice_type = h1->last_slice_type;
620 if (context_reinitialized)
621 ff_h264_set_parameter_from_sps(h);
627 err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
628 h->prev_poc_msb = h->poc_msb;
629 h->prev_poc_lsb = h->poc_lsb;
631 h->prev_frame_num_offset = h->frame_num_offset;
632 h->prev_frame_num = h->frame_num;
633 h->outputed_poc = h->next_outputed_poc;
635 h->recovery_frame = h1->recovery_frame;
636 h->frame_recovered = h1->frame_recovered;
641 static int h264_frame_start(H264Context *h)
645 const int pixel_shift = h->pixel_shift;
647 release_unused_pictures(h, 1);
648 h->cur_pic_ptr = NULL;
650 i = find_unused_picture(h);
652 av_log(h->avctx, AV_LOG_ERROR, "no frame buffer available\n");
657 pic->reference = h->droppable ? 0 : h->picture_structure;
658 pic->f.coded_picture_number = h->coded_picture_number++;
659 pic->field_picture = h->picture_structure != PICT_FRAME;
661 * Zero key_frame here; IDR markings per slice in frame or fields are ORed
663 * See decode_nal_units().
665 pic->f.key_frame = 0;
669 if ((ret = alloc_picture(h, pic)) < 0)
672 h->cur_pic_ptr = pic;
673 ff_h264_unref_picture(h, &h->cur_pic);
674 if ((ret = ff_h264_ref_picture(h, &h->cur_pic, h->cur_pic_ptr)) < 0)
677 if (CONFIG_ERROR_RESILIENCE)
678 ff_er_frame_start(&h->er);
680 assert(h->linesize && h->uvlinesize);
682 for (i = 0; i < 16; i++) {
683 h->block_offset[i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * h->linesize * ((scan8[i] - scan8[0]) >> 3);
684 h->block_offset[48 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * h->linesize * ((scan8[i] - scan8[0]) >> 3);
686 for (i = 0; i < 16; i++) {
687 h->block_offset[16 + i] =
688 h->block_offset[32 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * h->uvlinesize * ((scan8[i] - scan8[0]) >> 3);
689 h->block_offset[48 + 16 + i] =
690 h->block_offset[48 + 32 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * h->uvlinesize * ((scan8[i] - scan8[0]) >> 3);
693 /* can't be in alloc_tables because linesize isn't known there.
694 * FIXME: redo bipred weight to not require extra buffer? */
695 for (i = 0; i < h->slice_context_count; i++)
696 if (h->thread_context[i]) {
697 ret = alloc_scratch_buffers(h->thread_context[i], h->linesize);
702 /* Some macroblocks can be accessed before they're available in case
703 * of lost slices, MBAFF or threading. */
704 memset(h->slice_table, -1,
705 (h->mb_height * h->mb_stride - 1) * sizeof(*h->slice_table));
707 /* We mark the current picture as non-reference after allocating it, so
708 * that if we break out due to an error it can be released automatically
709 * in the next ff_mpv_frame_start().
711 h->cur_pic_ptr->reference = 0;
713 h->cur_pic_ptr->field_poc[0] = h->cur_pic_ptr->field_poc[1] = INT_MAX;
715 h->next_output_pic = NULL;
717 assert(h->cur_pic_ptr->long_ref == 0);
722 static av_always_inline void backup_mb_border(H264Context *h, uint8_t *src_y,
723 uint8_t *src_cb, uint8_t *src_cr,
724 int linesize, int uvlinesize,
729 const int pixel_shift = h->pixel_shift;
730 int chroma444 = CHROMA444(h);
731 int chroma422 = CHROMA422(h);
734 src_cb -= uvlinesize;
735 src_cr -= uvlinesize;
737 if (!simple && FRAME_MBAFF(h)) {
740 top_border = h->top_borders[0][h->mb_x];
741 AV_COPY128(top_border, src_y + 15 * linesize);
743 AV_COPY128(top_border + 16, src_y + 15 * linesize + 16);
744 if (simple || !CONFIG_GRAY || !(h->flags & CODEC_FLAG_GRAY)) {
747 AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
748 AV_COPY128(top_border + 48, src_cb + 15 * uvlinesize + 16);
749 AV_COPY128(top_border + 64, src_cr + 15 * uvlinesize);
750 AV_COPY128(top_border + 80, src_cr + 15 * uvlinesize + 16);
752 AV_COPY128(top_border + 16, src_cb + 15 * uvlinesize);
753 AV_COPY128(top_border + 32, src_cr + 15 * uvlinesize);
755 } else if (chroma422) {
757 AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
758 AV_COPY128(top_border + 48, src_cr + 15 * uvlinesize);
760 AV_COPY64(top_border + 16, src_cb + 15 * uvlinesize);
761 AV_COPY64(top_border + 24, src_cr + 15 * uvlinesize);
765 AV_COPY128(top_border + 32, src_cb + 7 * uvlinesize);
766 AV_COPY128(top_border + 48, src_cr + 7 * uvlinesize);
768 AV_COPY64(top_border + 16, src_cb + 7 * uvlinesize);
769 AV_COPY64(top_border + 24, src_cr + 7 * uvlinesize);
774 } else if (MB_MBAFF(h)) {
780 top_border = h->top_borders[top_idx][h->mb_x];
781 /* There are two lines saved, the line above the top macroblock
782 * of a pair, and the line above the bottom macroblock. */
783 AV_COPY128(top_border, src_y + 16 * linesize);
785 AV_COPY128(top_border + 16, src_y + 16 * linesize + 16);
787 if (simple || !CONFIG_GRAY || !(h->flags & CODEC_FLAG_GRAY)) {
790 AV_COPY128(top_border + 32, src_cb + 16 * linesize);
791 AV_COPY128(top_border + 48, src_cb + 16 * linesize + 16);
792 AV_COPY128(top_border + 64, src_cr + 16 * linesize);
793 AV_COPY128(top_border + 80, src_cr + 16 * linesize + 16);
795 AV_COPY128(top_border + 16, src_cb + 16 * linesize);
796 AV_COPY128(top_border + 32, src_cr + 16 * linesize);
798 } else if (chroma422) {
800 AV_COPY128(top_border + 32, src_cb + 16 * uvlinesize);
801 AV_COPY128(top_border + 48, src_cr + 16 * uvlinesize);
803 AV_COPY64(top_border + 16, src_cb + 16 * uvlinesize);
804 AV_COPY64(top_border + 24, src_cr + 16 * uvlinesize);
808 AV_COPY128(top_border + 32, src_cb + 8 * uvlinesize);
809 AV_COPY128(top_border + 48, src_cr + 8 * uvlinesize);
811 AV_COPY64(top_border + 16, src_cb + 8 * uvlinesize);
812 AV_COPY64(top_border + 24, src_cr + 8 * uvlinesize);
819 * Initialize implicit_weight table.
820 * @param field 0/1 initialize the weight for interlaced MBAFF
821 * -1 initializes the rest
823 static void implicit_weight_table(H264Context *h, H264SliceContext *sl, int field)
825 int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
827 for (i = 0; i < 2; i++) {
828 sl->luma_weight_flag[i] = 0;
829 sl->chroma_weight_flag[i] = 0;
833 if (h->picture_structure == PICT_FRAME) {
834 cur_poc = h->cur_pic_ptr->poc;
836 cur_poc = h->cur_pic_ptr->field_poc[h->picture_structure - 1];
838 if (sl->ref_count[0] == 1 && sl->ref_count[1] == 1 && !FRAME_MBAFF(h) &&
839 sl->ref_list[0][0].poc + sl->ref_list[1][0].poc == 2 * cur_poc) {
841 sl->use_weight_chroma = 0;
845 ref_count0 = sl->ref_count[0];
846 ref_count1 = sl->ref_count[1];
848 cur_poc = h->cur_pic_ptr->field_poc[field];
850 ref_count0 = 16 + 2 * sl->ref_count[0];
851 ref_count1 = 16 + 2 * sl->ref_count[1];
855 sl->use_weight_chroma = 2;
856 sl->luma_log2_weight_denom = 5;
857 sl->chroma_log2_weight_denom = 5;
859 for (ref0 = ref_start; ref0 < ref_count0; ref0++) {
860 int poc0 = sl->ref_list[0][ref0].poc;
861 for (ref1 = ref_start; ref1 < ref_count1; ref1++) {
863 if (!sl->ref_list[0][ref0].long_ref && !sl->ref_list[1][ref1].long_ref) {
864 int poc1 = sl->ref_list[1][ref1].poc;
865 int td = av_clip_int8(poc1 - poc0);
867 int tb = av_clip_int8(cur_poc - poc0);
868 int tx = (16384 + (FFABS(td) >> 1)) / td;
869 int dist_scale_factor = (tb * tx + 32) >> 8;
870 if (dist_scale_factor >= -64 && dist_scale_factor <= 128)
871 w = 64 - dist_scale_factor;
875 sl->implicit_weight[ref0][ref1][0] =
876 sl->implicit_weight[ref0][ref1][1] = w;
878 sl->implicit_weight[ref0][ref1][field] = w;
885 * initialize scan tables
887 static void init_scan_tables(H264Context *h)
890 for (i = 0; i < 16; i++) {
891 #define TRANSPOSE(x) (x >> 2) | ((x << 2) & 0xF)
892 h->zigzag_scan[i] = TRANSPOSE(zigzag_scan[i]);
893 h->field_scan[i] = TRANSPOSE(field_scan[i]);
896 for (i = 0; i < 64; i++) {
897 #define TRANSPOSE(x) (x >> 3) | ((x & 7) << 3)
898 h->zigzag_scan8x8[i] = TRANSPOSE(ff_zigzag_direct[i]);
899 h->zigzag_scan8x8_cavlc[i] = TRANSPOSE(zigzag_scan8x8_cavlc[i]);
900 h->field_scan8x8[i] = TRANSPOSE(field_scan8x8[i]);
901 h->field_scan8x8_cavlc[i] = TRANSPOSE(field_scan8x8_cavlc[i]);
904 if (h->sps.transform_bypass) { // FIXME same ugly
905 h->zigzag_scan_q0 = zigzag_scan;
906 h->zigzag_scan8x8_q0 = ff_zigzag_direct;
907 h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
908 h->field_scan_q0 = field_scan;
909 h->field_scan8x8_q0 = field_scan8x8;
910 h->field_scan8x8_cavlc_q0 = field_scan8x8_cavlc;
912 h->zigzag_scan_q0 = h->zigzag_scan;
913 h->zigzag_scan8x8_q0 = h->zigzag_scan8x8;
914 h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
915 h->field_scan_q0 = h->field_scan;
916 h->field_scan8x8_q0 = h->field_scan8x8;
917 h->field_scan8x8_cavlc_q0 = h->field_scan8x8_cavlc;
922 * Replicate H264 "master" context to thread contexts.
924 static int clone_slice(H264Context *dst, H264Context *src)
926 memcpy(dst->block_offset, src->block_offset, sizeof(dst->block_offset));
927 dst->cur_pic_ptr = src->cur_pic_ptr;
928 dst->cur_pic = src->cur_pic;
929 dst->linesize = src->linesize;
930 dst->uvlinesize = src->uvlinesize;
931 dst->first_field = src->first_field;
933 dst->prev_poc_msb = src->prev_poc_msb;
934 dst->prev_poc_lsb = src->prev_poc_lsb;
935 dst->prev_frame_num_offset = src->prev_frame_num_offset;
936 dst->prev_frame_num = src->prev_frame_num;
937 dst->short_ref_count = src->short_ref_count;
939 memcpy(dst->short_ref, src->short_ref, sizeof(dst->short_ref));
940 memcpy(dst->long_ref, src->long_ref, sizeof(dst->long_ref));
941 memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
943 memcpy(dst->dequant4_coeff, src->dequant4_coeff, sizeof(src->dequant4_coeff));
944 memcpy(dst->dequant8_coeff, src->dequant8_coeff, sizeof(src->dequant8_coeff));
949 static enum AVPixelFormat get_pixel_format(H264Context *h)
951 #define HWACCEL_MAX (CONFIG_H264_DXVA2_HWACCEL + \
952 CONFIG_H264_VAAPI_HWACCEL + \
953 (CONFIG_H264_VDA_HWACCEL * 2) + \
954 CONFIG_H264_VDPAU_HWACCEL)
955 enum AVPixelFormat pix_fmts[HWACCEL_MAX + 2], *fmt = pix_fmts;
956 const enum AVPixelFormat *choices = pix_fmts;
958 switch (h->sps.bit_depth_luma) {
961 if (h->avctx->colorspace == AVCOL_SPC_RGB) {
962 *fmt++ = AV_PIX_FMT_GBRP9;
964 *fmt++ = AV_PIX_FMT_YUV444P9;
965 } else if (CHROMA422(h))
966 *fmt++ = AV_PIX_FMT_YUV422P9;
968 *fmt++ = AV_PIX_FMT_YUV420P9;
972 if (h->avctx->colorspace == AVCOL_SPC_RGB) {
973 *fmt++ = AV_PIX_FMT_GBRP10;
975 *fmt++ = AV_PIX_FMT_YUV444P10;
976 } else if (CHROMA422(h))
977 *fmt++ = AV_PIX_FMT_YUV422P10;
979 *fmt++ = AV_PIX_FMT_YUV420P10;
982 #if CONFIG_H264_VDPAU_HWACCEL
983 *fmt++ = AV_PIX_FMT_VDPAU;
986 if (h->avctx->colorspace == AVCOL_SPC_RGB)
987 *fmt++ = AV_PIX_FMT_GBRP;
988 else if (h->avctx->color_range == AVCOL_RANGE_JPEG)
989 *fmt++ = AV_PIX_FMT_YUVJ444P;
991 *fmt++ = AV_PIX_FMT_YUV444P;
992 } else if (CHROMA422(h)) {
993 if (h->avctx->color_range == AVCOL_RANGE_JPEG)
994 *fmt++ = AV_PIX_FMT_YUVJ422P;
996 *fmt++ = AV_PIX_FMT_YUV422P;
998 #if CONFIG_H264_DXVA2_HWACCEL
999 *fmt++ = AV_PIX_FMT_DXVA2_VLD;
1001 #if CONFIG_H264_VAAPI_HWACCEL
1002 *fmt++ = AV_PIX_FMT_VAAPI_VLD;
1004 #if CONFIG_H264_VDA_HWACCEL
1005 *fmt++ = AV_PIX_FMT_VDA_VLD;
1006 *fmt++ = AV_PIX_FMT_VDA;
1008 if (h->avctx->codec->pix_fmts)
1009 choices = h->avctx->codec->pix_fmts;
1010 else if (h->avctx->color_range == AVCOL_RANGE_JPEG)
1011 *fmt++ = AV_PIX_FMT_YUVJ420P;
1013 *fmt++ = AV_PIX_FMT_YUV420P;
1017 av_log(h->avctx, AV_LOG_ERROR,
1018 "Unsupported bit depth %d\n", h->sps.bit_depth_luma);
1019 return AVERROR_INVALIDDATA;
1022 *fmt = AV_PIX_FMT_NONE;
1024 return ff_get_format(h->avctx, choices);
1027 /* export coded and cropped frame dimensions to AVCodecContext */
1028 static int init_dimensions(H264Context *h)
1030 int width = h->width - (h->sps.crop_right + h->sps.crop_left);
1031 int height = h->height - (h->sps.crop_top + h->sps.crop_bottom);
1032 int crop_present = h->sps.crop_left || h->sps.crop_top ||
1033 h->sps.crop_right || h->sps.crop_bottom;
1035 /* handle container cropping */
1036 if (!crop_present &&
1037 FFALIGN(h->avctx->width, 16) == h->width &&
1038 FFALIGN(h->avctx->height, 16) == h->height) {
1039 width = h->avctx->width;
1040 height = h->avctx->height;
1043 if (width <= 0 || height <= 0) {
1044 av_log(h->avctx, AV_LOG_ERROR, "Invalid cropped dimensions: %dx%d.\n",
1046 if (h->avctx->err_recognition & AV_EF_EXPLODE)
1047 return AVERROR_INVALIDDATA;
1049 av_log(h->avctx, AV_LOG_WARNING, "Ignoring cropping information.\n");
1050 h->sps.crop_bottom =
1060 h->avctx->coded_width = h->width;
1061 h->avctx->coded_height = h->height;
1062 h->avctx->width = width;
1063 h->avctx->height = height;
1068 static int h264_slice_header_init(H264Context *h, int reinit)
1070 int nb_slices = (HAVE_THREADS &&
1071 h->avctx->active_thread_type & FF_THREAD_SLICE) ?
1072 h->avctx->thread_count : 1;
1075 ff_set_sar(h->avctx, h->sps.sar);
1076 av_pix_fmt_get_chroma_sub_sample(h->avctx->pix_fmt,
1077 &h->chroma_x_shift, &h->chroma_y_shift);
1079 if (h->sps.timing_info_present_flag) {
1080 int64_t den = h->sps.time_scale;
1081 if (h->x264_build < 44U)
1083 av_reduce(&h->avctx->framerate.den, &h->avctx->framerate.num,
1084 h->sps.num_units_in_tick, den, 1 << 30);
1088 ff_h264_free_tables(h, 0);
1090 h->prev_interlaced_frame = 1;
1092 init_scan_tables(h);
1093 ret = ff_h264_alloc_tables(h);
1095 av_log(h->avctx, AV_LOG_ERROR, "Could not allocate memory\n");
1099 if (nb_slices > H264_MAX_THREADS || (nb_slices > h->mb_height && h->mb_height)) {
1102 max_slices = FFMIN(H264_MAX_THREADS, h->mb_height);
1104 max_slices = H264_MAX_THREADS;
1105 av_log(h->avctx, AV_LOG_WARNING, "too many threads/slices %d,"
1106 " reducing to %d\n", nb_slices, max_slices);
1107 nb_slices = max_slices;
1109 h->slice_context_count = nb_slices;
1111 if (!HAVE_THREADS || !(h->avctx->active_thread_type & FF_THREAD_SLICE)) {
1112 ret = ff_h264_context_init(h);
1114 av_log(h->avctx, AV_LOG_ERROR, "context_init() failed.\n");
1118 for (i = 1; i < h->slice_context_count; i++) {
1120 c = h->thread_context[i] = av_mallocz(sizeof(H264Context));
1122 return AVERROR(ENOMEM);
1123 c->avctx = h->avctx;
1125 c->h264dsp = h->h264dsp;
1126 c->h264qpel = h->h264qpel;
1127 c->h264chroma = h->h264chroma;
1130 c->pixel_shift = h->pixel_shift;
1131 c->width = h->width;
1132 c->height = h->height;
1133 c->linesize = h->linesize;
1134 c->uvlinesize = h->uvlinesize;
1135 c->chroma_x_shift = h->chroma_x_shift;
1136 c->chroma_y_shift = h->chroma_y_shift;
1137 c->droppable = h->droppable;
1138 c->low_delay = h->low_delay;
1139 c->mb_width = h->mb_width;
1140 c->mb_height = h->mb_height;
1141 c->mb_stride = h->mb_stride;
1142 c->mb_num = h->mb_num;
1143 c->flags = h->flags;
1144 c->workaround_bugs = h->workaround_bugs;
1145 c->pict_type = h->pict_type;
1147 h->slice_ctx[i].h264 = c;
1149 init_scan_tables(c);
1150 clone_tables(c, &h->slice_ctx[i], h, i);
1151 c->context_initialized = 1;
1154 for (i = 0; i < h->slice_context_count; i++)
1155 if ((ret = ff_h264_context_init(h->thread_context[i])) < 0) {
1156 av_log(h->avctx, AV_LOG_ERROR, "context_init() failed.\n");
1161 h->context_initialized = 1;
1167 * Decode a slice header.
1168 * This will (re)intialize the decoder and call h264_frame_start() as needed.
1170 * @param h h264context
1171 * @param h0 h264 master context (differs from 'h' when doing sliced based
1172 * parallel decoding)
1174 * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
1176 int ff_h264_decode_slice_header(H264Context *h, H264SliceContext *sl, H264Context *h0)
1178 unsigned int first_mb_in_slice;
1179 unsigned int pps_id;
1181 unsigned int slice_type, tmp, i, j;
1182 int default_ref_list_done = 0;
1183 int last_pic_structure, last_pic_droppable;
1184 int needs_reinit = 0;
1185 int field_pic_flag, bottom_field_flag;
1187 h->qpel_put = h->h264qpel.put_h264_qpel_pixels_tab;
1188 h->qpel_avg = h->h264qpel.avg_h264_qpel_pixels_tab;
1190 first_mb_in_slice = get_ue_golomb(&h->gb);
1192 if (first_mb_in_slice == 0) { // FIXME better field boundary detection
1193 if (h0->current_slice && h->cur_pic_ptr && FIELD_PICTURE(h)) {
1194 ff_h264_field_end(h, sl, 1);
1197 h0->current_slice = 0;
1198 if (!h0->first_field) {
1199 if (h->cur_pic_ptr && !h->droppable) {
1200 ff_thread_report_progress(&h->cur_pic_ptr->tf, INT_MAX,
1201 h->picture_structure == PICT_BOTTOM_FIELD);
1203 h->cur_pic_ptr = NULL;
1207 slice_type = get_ue_golomb_31(&h->gb);
1208 if (slice_type > 9) {
1209 av_log(h->avctx, AV_LOG_ERROR,
1210 "slice type %d too large at %d %d\n",
1211 slice_type, h->mb_x, h->mb_y);
1212 return AVERROR_INVALIDDATA;
1214 if (slice_type > 4) {
1216 sl->slice_type_fixed = 1;
1218 sl->slice_type_fixed = 0;
1220 slice_type = golomb_to_pict_type[slice_type];
1221 if (slice_type == AV_PICTURE_TYPE_I ||
1222 (h0->current_slice != 0 && slice_type == h0->last_slice_type)) {
1223 default_ref_list_done = 1;
1225 sl->slice_type = slice_type;
1226 sl->slice_type_nos = slice_type & 3;
1228 if (h->nal_unit_type == NAL_IDR_SLICE &&
1229 sl->slice_type_nos != AV_PICTURE_TYPE_I) {
1230 av_log(h->avctx, AV_LOG_ERROR, "A non-intra slice in an IDR NAL unit.\n");
1231 return AVERROR_INVALIDDATA;
1234 // to make a few old functions happy, it's wrong though
1235 h->pict_type = sl->slice_type;
1237 pps_id = get_ue_golomb(&h->gb);
1238 if (pps_id >= MAX_PPS_COUNT) {
1239 av_log(h->avctx, AV_LOG_ERROR, "pps_id %u out of range\n", pps_id);
1240 return AVERROR_INVALIDDATA;
1242 if (!h0->pps_buffers[pps_id]) {
1243 av_log(h->avctx, AV_LOG_ERROR,
1244 "non-existing PPS %u referenced\n",
1246 return AVERROR_INVALIDDATA;
1248 h->pps = *h0->pps_buffers[pps_id];
1250 if (!h0->sps_buffers[h->pps.sps_id]) {
1251 av_log(h->avctx, AV_LOG_ERROR,
1252 "non-existing SPS %u referenced\n",
1254 return AVERROR_INVALIDDATA;
1257 if (h->pps.sps_id != h->sps.sps_id ||
1258 h0->sps_buffers[h->pps.sps_id]->new) {
1259 h0->sps_buffers[h->pps.sps_id]->new = 0;
1261 h->sps = *h0->sps_buffers[h->pps.sps_id];
1263 if (h->bit_depth_luma != h->sps.bit_depth_luma ||
1264 h->chroma_format_idc != h->sps.chroma_format_idc) {
1265 h->bit_depth_luma = h->sps.bit_depth_luma;
1266 h->chroma_format_idc = h->sps.chroma_format_idc;
1269 if ((ret = ff_h264_set_parameter_from_sps(h)) < 0)
1273 h->avctx->profile = ff_h264_get_profile(&h->sps);
1274 h->avctx->level = h->sps.level_idc;
1275 h->avctx->refs = h->sps.ref_frame_count;
1277 if (h->mb_width != h->sps.mb_width ||
1278 h->mb_height != h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag))
1281 h->mb_width = h->sps.mb_width;
1282 h->mb_height = h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
1283 h->mb_num = h->mb_width * h->mb_height;
1284 h->mb_stride = h->mb_width + 1;
1286 h->b_stride = h->mb_width * 4;
1288 h->chroma_y_shift = h->sps.chroma_format_idc <= 1; // 400 uses yuv420p
1290 h->width = 16 * h->mb_width;
1291 h->height = 16 * h->mb_height;
1293 ret = init_dimensions(h);
1297 if (h->sps.video_signal_type_present_flag) {
1298 h->avctx->color_range = h->sps.full_range ? AVCOL_RANGE_JPEG
1300 if (h->sps.colour_description_present_flag) {
1301 if (h->avctx->colorspace != h->sps.colorspace)
1303 h->avctx->color_primaries = h->sps.color_primaries;
1304 h->avctx->color_trc = h->sps.color_trc;
1305 h->avctx->colorspace = h->sps.colorspace;
1309 if (h->context_initialized && needs_reinit) {
1311 av_log(h->avctx, AV_LOG_ERROR,
1312 "changing width %d -> %d / height %d -> %d on "
1314 h->width, h->avctx->coded_width,
1315 h->height, h->avctx->coded_height,
1316 h0->current_slice + 1);
1317 return AVERROR_INVALIDDATA;
1320 ff_h264_flush_change(h);
1322 if ((ret = get_pixel_format(h)) < 0)
1324 h->avctx->pix_fmt = ret;
1326 av_log(h->avctx, AV_LOG_INFO, "Reinit context to %dx%d, "
1327 "pix_fmt: %d\n", h->width, h->height, h->avctx->pix_fmt);
1329 if ((ret = h264_slice_header_init(h, 1)) < 0) {
1330 av_log(h->avctx, AV_LOG_ERROR,
1331 "h264_slice_header_init() failed\n");
1335 if (!h->context_initialized) {
1337 av_log(h->avctx, AV_LOG_ERROR,
1338 "Cannot (re-)initialize context during parallel decoding.\n");
1339 return AVERROR_PATCHWELCOME;
1342 if ((ret = get_pixel_format(h)) < 0)
1344 h->avctx->pix_fmt = ret;
1346 if ((ret = h264_slice_header_init(h, 0)) < 0) {
1347 av_log(h->avctx, AV_LOG_ERROR,
1348 "h264_slice_header_init() failed\n");
1353 if (h == h0 && h->dequant_coeff_pps != pps_id) {
1354 h->dequant_coeff_pps = pps_id;
1355 h264_init_dequant_tables(h);
1358 h->frame_num = get_bits(&h->gb, h->sps.log2_max_frame_num);
1361 h->mb_aff_frame = 0;
1362 last_pic_structure = h0->picture_structure;
1363 last_pic_droppable = h0->droppable;
1364 h->droppable = h->nal_ref_idc == 0;
1365 if (h->sps.frame_mbs_only_flag) {
1366 h->picture_structure = PICT_FRAME;
1368 field_pic_flag = get_bits1(&h->gb);
1369 if (field_pic_flag) {
1370 bottom_field_flag = get_bits1(&h->gb);
1371 h->picture_structure = PICT_TOP_FIELD + bottom_field_flag;
1373 h->picture_structure = PICT_FRAME;
1374 h->mb_aff_frame = h->sps.mb_aff;
1377 h->mb_field_decoding_flag = h->picture_structure != PICT_FRAME;
1379 if (h0->current_slice != 0) {
1380 if (last_pic_structure != h->picture_structure ||
1381 last_pic_droppable != h->droppable) {
1382 av_log(h->avctx, AV_LOG_ERROR,
1383 "Changing field mode (%d -> %d) between slices is not allowed\n",
1384 last_pic_structure, h->picture_structure);
1385 h->picture_structure = last_pic_structure;
1386 h->droppable = last_pic_droppable;
1387 return AVERROR_INVALIDDATA;
1388 } else if (!h0->cur_pic_ptr) {
1389 av_log(h->avctx, AV_LOG_ERROR,
1390 "unset cur_pic_ptr on slice %d\n",
1391 h0->current_slice + 1);
1392 return AVERROR_INVALIDDATA;
1395 /* Shorten frame num gaps so we don't have to allocate reference
1396 * frames just to throw them away */
1397 if (h->frame_num != h->prev_frame_num) {
1398 int unwrap_prev_frame_num = h->prev_frame_num;
1399 int max_frame_num = 1 << h->sps.log2_max_frame_num;
1401 if (unwrap_prev_frame_num > h->frame_num)
1402 unwrap_prev_frame_num -= max_frame_num;
1404 if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
1405 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
1406 if (unwrap_prev_frame_num < 0)
1407 unwrap_prev_frame_num += max_frame_num;
1409 h->prev_frame_num = unwrap_prev_frame_num;
1413 /* See if we have a decoded first field looking for a pair...
1414 * Here, we're using that to see if we should mark previously
1415 * decode frames as "finished".
1416 * We have to do that before the "dummy" in-between frame allocation,
1417 * since that can modify s->current_picture_ptr. */
1418 if (h0->first_field) {
1419 assert(h0->cur_pic_ptr);
1420 assert(h0->cur_pic_ptr->f.buf[0]);
1421 assert(h0->cur_pic_ptr->reference != DELAYED_PIC_REF);
1423 /* figure out if we have a complementary field pair */
1424 if (!FIELD_PICTURE(h) || h->picture_structure == last_pic_structure) {
1425 /* Previous field is unmatched. Don't display it, but let it
1426 * remain for reference if marked as such. */
1427 if (!last_pic_droppable && last_pic_structure != PICT_FRAME) {
1428 ff_thread_report_progress(&h0->cur_pic_ptr->tf, INT_MAX,
1429 last_pic_structure == PICT_TOP_FIELD);
1432 if (h0->cur_pic_ptr->frame_num != h->frame_num) {
1433 /* This and previous field were reference, but had
1434 * different frame_nums. Consider this field first in
1435 * pair. Throw away previous field except for reference
1437 if (!last_pic_droppable && last_pic_structure != PICT_FRAME) {
1438 ff_thread_report_progress(&h0->cur_pic_ptr->tf, INT_MAX,
1439 last_pic_structure == PICT_TOP_FIELD);
1442 /* Second field in complementary pair */
1443 if (!((last_pic_structure == PICT_TOP_FIELD &&
1444 h->picture_structure == PICT_BOTTOM_FIELD) ||
1445 (last_pic_structure == PICT_BOTTOM_FIELD &&
1446 h->picture_structure == PICT_TOP_FIELD))) {
1447 av_log(h->avctx, AV_LOG_ERROR,
1448 "Invalid field mode combination %d/%d\n",
1449 last_pic_structure, h->picture_structure);
1450 h->picture_structure = last_pic_structure;
1451 h->droppable = last_pic_droppable;
1452 return AVERROR_INVALIDDATA;
1453 } else if (last_pic_droppable != h->droppable) {
1454 avpriv_request_sample(h->avctx,
1455 "Found reference and non-reference fields in the same frame, which");
1456 h->picture_structure = last_pic_structure;
1457 h->droppable = last_pic_droppable;
1458 return AVERROR_PATCHWELCOME;
1464 while (h->frame_num != h->prev_frame_num &&
1465 h->frame_num != (h->prev_frame_num + 1) % (1 << h->sps.log2_max_frame_num)) {
1466 H264Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
1467 av_log(h->avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n",
1468 h->frame_num, h->prev_frame_num);
1469 ret = h264_frame_start(h);
1471 h0->first_field = 0;
1475 h->prev_frame_num++;
1476 h->prev_frame_num %= 1 << h->sps.log2_max_frame_num;
1477 h->cur_pic_ptr->frame_num = h->prev_frame_num;
1478 ff_thread_report_progress(&h->cur_pic_ptr->tf, INT_MAX, 0);
1479 ff_thread_report_progress(&h->cur_pic_ptr->tf, INT_MAX, 1);
1480 ret = ff_generate_sliding_window_mmcos(h, 1);
1481 if (ret < 0 && (h->avctx->err_recognition & AV_EF_EXPLODE))
1483 ret = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1484 if (ret < 0 && (h->avctx->err_recognition & AV_EF_EXPLODE))
1486 /* Error concealment: If a ref is missing, copy the previous ref
1488 * FIXME: Avoiding a memcpy would be nice, but ref handling makes
1489 * many assumptions about there being no actual duplicates.
1490 * FIXME: This does not copy padding for out-of-frame motion
1491 * vectors. Given we are concealing a lost frame, this probably
1492 * is not noticeable by comparison, but it should be fixed. */
1493 if (h->short_ref_count) {
1495 av_image_copy(h->short_ref[0]->f.data,
1496 h->short_ref[0]->f.linesize,
1497 (const uint8_t **)prev->f.data,
1502 h->short_ref[0]->poc = prev->poc + 2;
1504 h->short_ref[0]->frame_num = h->prev_frame_num;
1508 /* See if we have a decoded first field looking for a pair...
1509 * We're using that to see whether to continue decoding in that
1510 * frame, or to allocate a new one. */
1511 if (h0->first_field) {
1512 assert(h0->cur_pic_ptr);
1513 assert(h0->cur_pic_ptr->f.buf[0]);
1514 assert(h0->cur_pic_ptr->reference != DELAYED_PIC_REF);
1516 /* figure out if we have a complementary field pair */
1517 if (!FIELD_PICTURE(h) || h->picture_structure == last_pic_structure) {
1518 /* Previous field is unmatched. Don't display it, but let it
1519 * remain for reference if marked as such. */
1520 h0->cur_pic_ptr = NULL;
1521 h0->first_field = FIELD_PICTURE(h);
1523 if (h0->cur_pic_ptr->frame_num != h->frame_num) {
1524 /* This and the previous field had different frame_nums.
1525 * Consider this field first in pair. Throw away previous
1526 * one except for reference purposes. */
1527 h0->first_field = 1;
1528 h0->cur_pic_ptr = NULL;
1530 /* Second field in complementary pair */
1531 h0->first_field = 0;
1535 /* Frame or first field in a potentially complementary pair */
1536 h0->first_field = FIELD_PICTURE(h);
1539 if (!FIELD_PICTURE(h) || h0->first_field) {
1540 if (h264_frame_start(h) < 0) {
1541 h0->first_field = 0;
1542 return AVERROR_INVALIDDATA;
1545 release_unused_pictures(h, 0);
1548 if (h != h0 && (ret = clone_slice(h, h0)) < 0)
1551 h->cur_pic_ptr->frame_num = h->frame_num; // FIXME frame_num cleanup
1553 assert(h->mb_num == h->mb_width * h->mb_height);
1554 if (first_mb_in_slice << FIELD_OR_MBAFF_PICTURE(h) >= h->mb_num ||
1555 first_mb_in_slice >= h->mb_num) {
1556 av_log(h->avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
1557 return AVERROR_INVALIDDATA;
1559 h->resync_mb_x = h->mb_x = first_mb_in_slice % h->mb_width;
1560 h->resync_mb_y = h->mb_y = (first_mb_in_slice / h->mb_width) <<
1561 FIELD_OR_MBAFF_PICTURE(h);
1562 if (h->picture_structure == PICT_BOTTOM_FIELD)
1563 h->resync_mb_y = h->mb_y = h->mb_y + 1;
1564 assert(h->mb_y < h->mb_height);
1566 if (h->picture_structure == PICT_FRAME) {
1567 h->curr_pic_num = h->frame_num;
1568 h->max_pic_num = 1 << h->sps.log2_max_frame_num;
1570 h->curr_pic_num = 2 * h->frame_num + 1;
1571 h->max_pic_num = 1 << (h->sps.log2_max_frame_num + 1);
1574 if (h->nal_unit_type == NAL_IDR_SLICE)
1575 get_ue_golomb(&h->gb); /* idr_pic_id */
1577 if (h->sps.poc_type == 0) {
1578 h->poc_lsb = get_bits(&h->gb, h->sps.log2_max_poc_lsb);
1580 if (h->pps.pic_order_present == 1 && h->picture_structure == PICT_FRAME)
1581 h->delta_poc_bottom = get_se_golomb(&h->gb);
1584 if (h->sps.poc_type == 1 && !h->sps.delta_pic_order_always_zero_flag) {
1585 h->delta_poc[0] = get_se_golomb(&h->gb);
1587 if (h->pps.pic_order_present == 1 && h->picture_structure == PICT_FRAME)
1588 h->delta_poc[1] = get_se_golomb(&h->gb);
1591 ff_init_poc(h, h->cur_pic_ptr->field_poc, &h->cur_pic_ptr->poc);
1593 if (h->pps.redundant_pic_cnt_present)
1594 h->redundant_pic_count = get_ue_golomb(&h->gb);
1596 ret = ff_set_ref_count(h, sl);
1600 default_ref_list_done = 0;
1602 if (!default_ref_list_done)
1603 ff_h264_fill_default_ref_list(h, sl);
1605 if (sl->slice_type_nos != AV_PICTURE_TYPE_I) {
1606 ret = ff_h264_decode_ref_pic_list_reordering(h, sl);
1608 sl->ref_count[1] = sl->ref_count[0] = 0;
1613 if ((h->pps.weighted_pred && sl->slice_type_nos == AV_PICTURE_TYPE_P) ||
1614 (h->pps.weighted_bipred_idc == 1 &&
1615 sl->slice_type_nos == AV_PICTURE_TYPE_B))
1616 ff_pred_weight_table(h, sl);
1617 else if (h->pps.weighted_bipred_idc == 2 &&
1618 sl->slice_type_nos == AV_PICTURE_TYPE_B) {
1619 implicit_weight_table(h, sl, -1);
1622 for (i = 0; i < 2; i++) {
1623 sl->luma_weight_flag[i] = 0;
1624 sl->chroma_weight_flag[i] = 0;
1628 // If frame-mt is enabled, only update mmco tables for the first slice
1629 // in a field. Subsequent slices can temporarily clobber h->mmco_index
1630 // or h->mmco, which will cause ref list mix-ups and decoding errors
1631 // further down the line. This may break decoding if the first slice is
1632 // corrupt, thus we only do this if frame-mt is enabled.
1633 if (h->nal_ref_idc) {
1634 ret = ff_h264_decode_ref_pic_marking(h0, &h->gb,
1635 !(h->avctx->active_thread_type & FF_THREAD_FRAME) ||
1636 h0->current_slice == 0);
1637 if (ret < 0 && (h->avctx->err_recognition & AV_EF_EXPLODE))
1638 return AVERROR_INVALIDDATA;
1641 if (FRAME_MBAFF(h)) {
1642 ff_h264_fill_mbaff_ref_list(h, sl);
1644 if (h->pps.weighted_bipred_idc == 2 && sl->slice_type_nos == AV_PICTURE_TYPE_B) {
1645 implicit_weight_table(h, sl, 0);
1646 implicit_weight_table(h, sl, 1);
1650 if (sl->slice_type_nos == AV_PICTURE_TYPE_B && !sl->direct_spatial_mv_pred)
1651 ff_h264_direct_dist_scale_factor(h, sl);
1652 ff_h264_direct_ref_list_init(h, sl);
1654 if (sl->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac) {
1655 tmp = get_ue_golomb_31(&h->gb);
1657 av_log(h->avctx, AV_LOG_ERROR, "cabac_init_idc %u overflow\n", tmp);
1658 return AVERROR_INVALIDDATA;
1660 h->cabac_init_idc = tmp;
1663 h->last_qscale_diff = 0;
1664 tmp = h->pps.init_qp + get_se_golomb(&h->gb);
1665 if (tmp > 51 + 6 * (h->sps.bit_depth_luma - 8)) {
1666 av_log(h->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
1667 return AVERROR_INVALIDDATA;
1670 sl->chroma_qp[0] = get_chroma_qp(h, 0, sl->qscale);
1671 sl->chroma_qp[1] = get_chroma_qp(h, 1, sl->qscale);
1672 // FIXME qscale / qp ... stuff
1673 if (sl->slice_type == AV_PICTURE_TYPE_SP)
1674 get_bits1(&h->gb); /* sp_for_switch_flag */
1675 if (sl->slice_type == AV_PICTURE_TYPE_SP ||
1676 sl->slice_type == AV_PICTURE_TYPE_SI)
1677 get_se_golomb(&h->gb); /* slice_qs_delta */
1679 h->deblocking_filter = 1;
1680 h->slice_alpha_c0_offset = 0;
1681 h->slice_beta_offset = 0;
1682 if (h->pps.deblocking_filter_parameters_present) {
1683 tmp = get_ue_golomb_31(&h->gb);
1685 av_log(h->avctx, AV_LOG_ERROR,
1686 "deblocking_filter_idc %u out of range\n", tmp);
1687 return AVERROR_INVALIDDATA;
1689 h->deblocking_filter = tmp;
1690 if (h->deblocking_filter < 2)
1691 h->deblocking_filter ^= 1; // 1<->0
1693 if (h->deblocking_filter) {
1694 h->slice_alpha_c0_offset = get_se_golomb(&h->gb) * 2;
1695 h->slice_beta_offset = get_se_golomb(&h->gb) * 2;
1696 if (h->slice_alpha_c0_offset > 12 ||
1697 h->slice_alpha_c0_offset < -12 ||
1698 h->slice_beta_offset > 12 ||
1699 h->slice_beta_offset < -12) {
1700 av_log(h->avctx, AV_LOG_ERROR,
1701 "deblocking filter parameters %d %d out of range\n",
1702 h->slice_alpha_c0_offset, h->slice_beta_offset);
1703 return AVERROR_INVALIDDATA;
1708 if (h->avctx->skip_loop_filter >= AVDISCARD_ALL ||
1709 (h->avctx->skip_loop_filter >= AVDISCARD_NONKEY &&
1710 sl->slice_type_nos != AV_PICTURE_TYPE_I) ||
1711 (h->avctx->skip_loop_filter >= AVDISCARD_BIDIR &&
1712 sl->slice_type_nos == AV_PICTURE_TYPE_B) ||
1713 (h->avctx->skip_loop_filter >= AVDISCARD_NONREF &&
1714 h->nal_ref_idc == 0))
1715 h->deblocking_filter = 0;
1717 if (h->deblocking_filter == 1 && h0->max_contexts > 1) {
1718 if (h->avctx->flags2 & CODEC_FLAG2_FAST) {
1719 /* Cheat slightly for speed:
1720 * Do not bother to deblock across slices. */
1721 h->deblocking_filter = 2;
1723 h0->max_contexts = 1;
1724 if (!h0->single_decode_warning) {
1725 av_log(h->avctx, AV_LOG_INFO,
1726 "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
1727 h0->single_decode_warning = 1;
1730 av_log(h->avctx, AV_LOG_ERROR,
1731 "Deblocking switched inside frame.\n");
1736 sl->qp_thresh = 15 -
1737 FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset) -
1739 h->pps.chroma_qp_index_offset[0],
1740 h->pps.chroma_qp_index_offset[1]) +
1741 6 * (h->sps.bit_depth_luma - 8);
1743 h0->last_slice_type = slice_type;
1744 sl->slice_num = ++h0->current_slice;
1745 if (sl->slice_num >= MAX_SLICES) {
1746 av_log(h->avctx, AV_LOG_ERROR,
1747 "Too many slices, increase MAX_SLICES and recompile\n");
1750 for (j = 0; j < 2; j++) {
1752 int *ref2frm = sl->ref2frm[sl->slice_num & (MAX_SLICES - 1)][j];
1753 for (i = 0; i < 16; i++) {
1755 if (j < sl->list_count && i < sl->ref_count[j] &&
1756 sl->ref_list[j][i].f.buf[0]) {
1758 AVBuffer *buf = sl->ref_list[j][i].f.buf[0]->buffer;
1759 for (k = 0; k < h->short_ref_count; k++)
1760 if (h->short_ref[k]->f.buf[0]->buffer == buf) {
1764 for (k = 0; k < h->long_ref_count; k++)
1765 if (h->long_ref[k] && h->long_ref[k]->f.buf[0]->buffer == buf) {
1766 id_list[i] = h->short_ref_count + k;
1774 for (i = 0; i < 16; i++)
1775 ref2frm[i + 2] = 4 * id_list[i] + (sl->ref_list[j][i].reference & 3);
1777 ref2frm[18 + 1] = -1;
1778 for (i = 16; i < 48; i++)
1779 ref2frm[i + 4] = 4 * id_list[(i - 16) >> 1] +
1780 (sl->ref_list[j][i].reference & 3);
1783 if (h->avctx->debug & FF_DEBUG_PICT_INFO) {
1784 av_log(h->avctx, AV_LOG_DEBUG,
1785 "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",
1787 (h->picture_structure == PICT_FRAME ? "F" : h->picture_structure == PICT_TOP_FIELD ? "T" : "B"),
1789 av_get_picture_type_char(sl->slice_type),
1790 sl->slice_type_fixed ? " fix" : "",
1791 h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
1792 pps_id, h->frame_num,
1793 h->cur_pic_ptr->field_poc[0],
1794 h->cur_pic_ptr->field_poc[1],
1795 sl->ref_count[0], sl->ref_count[1],
1797 h->deblocking_filter,
1798 h->slice_alpha_c0_offset, h->slice_beta_offset,
1800 sl->use_weight == 1 && sl->use_weight_chroma ? "c" : "",
1801 sl->slice_type == AV_PICTURE_TYPE_B ? (sl->direct_spatial_mv_pred ? "SPAT" : "TEMP") : "");
1807 int ff_h264_get_slice_type(const H264SliceContext *sl)
1809 switch (sl->slice_type) {
1810 case AV_PICTURE_TYPE_P:
1812 case AV_PICTURE_TYPE_B:
1814 case AV_PICTURE_TYPE_I:
1816 case AV_PICTURE_TYPE_SP:
1818 case AV_PICTURE_TYPE_SI:
1821 return AVERROR_INVALIDDATA;
1825 static av_always_inline void fill_filter_caches_inter(H264Context *h,
1826 H264SliceContext *sl,
1827 int mb_type, int top_xy,
1828 int left_xy[LEFT_MBS],
1830 int left_type[LEFT_MBS],
1831 int mb_xy, int list)
1833 int b_stride = h->b_stride;
1834 int16_t(*mv_dst)[2] = &sl->mv_cache[list][scan8[0]];
1835 int8_t *ref_cache = &sl->ref_cache[list][scan8[0]];
1836 if (IS_INTER(mb_type) || IS_DIRECT(mb_type)) {
1837 if (USES_LIST(top_type, list)) {
1838 const int b_xy = h->mb2b_xy[top_xy] + 3 * b_stride;
1839 const int b8_xy = 4 * top_xy + 2;
1840 int (*ref2frm)[64] = sl->ref2frm[h->slice_table[top_xy] & (MAX_SLICES - 1)][0] + (MB_MBAFF(h) ? 20 : 2);
1841 AV_COPY128(mv_dst - 1 * 8, h->cur_pic.motion_val[list][b_xy + 0]);
1842 ref_cache[0 - 1 * 8] =
1843 ref_cache[1 - 1 * 8] = ref2frm[list][h->cur_pic.ref_index[list][b8_xy + 0]];
1844 ref_cache[2 - 1 * 8] =
1845 ref_cache[3 - 1 * 8] = ref2frm[list][h->cur_pic.ref_index[list][b8_xy + 1]];
1847 AV_ZERO128(mv_dst - 1 * 8);
1848 AV_WN32A(&ref_cache[0 - 1 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
1851 if (!IS_INTERLACED(mb_type ^ left_type[LTOP])) {
1852 if (USES_LIST(left_type[LTOP], list)) {
1853 const int b_xy = h->mb2b_xy[left_xy[LTOP]] + 3;
1854 const int b8_xy = 4 * left_xy[LTOP] + 1;
1855 int (*ref2frm)[64] = sl->ref2frm[h->slice_table[left_xy[LTOP]] & (MAX_SLICES - 1)][0] + (MB_MBAFF(h) ? 20 : 2);
1856 AV_COPY32(mv_dst - 1 + 0, h->cur_pic.motion_val[list][b_xy + b_stride * 0]);
1857 AV_COPY32(mv_dst - 1 + 8, h->cur_pic.motion_val[list][b_xy + b_stride * 1]);
1858 AV_COPY32(mv_dst - 1 + 16, h->cur_pic.motion_val[list][b_xy + b_stride * 2]);
1859 AV_COPY32(mv_dst - 1 + 24, h->cur_pic.motion_val[list][b_xy + b_stride * 3]);
1861 ref_cache[-1 + 8] = ref2frm[list][h->cur_pic.ref_index[list][b8_xy + 2 * 0]];
1862 ref_cache[-1 + 16] =
1863 ref_cache[-1 + 24] = ref2frm[list][h->cur_pic.ref_index[list][b8_xy + 2 * 1]];
1865 AV_ZERO32(mv_dst - 1 + 0);
1866 AV_ZERO32(mv_dst - 1 + 8);
1867 AV_ZERO32(mv_dst - 1 + 16);
1868 AV_ZERO32(mv_dst - 1 + 24);
1871 ref_cache[-1 + 16] =
1872 ref_cache[-1 + 24] = LIST_NOT_USED;
1877 if (!USES_LIST(mb_type, list)) {
1878 fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0, 0), 4);
1879 AV_WN32A(&ref_cache[0 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
1880 AV_WN32A(&ref_cache[1 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
1881 AV_WN32A(&ref_cache[2 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
1882 AV_WN32A(&ref_cache[3 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
1887 int8_t *ref = &h->cur_pic.ref_index[list][4 * mb_xy];
1888 int (*ref2frm)[64] = sl->ref2frm[sl->slice_num & (MAX_SLICES - 1)][0] + (MB_MBAFF(h) ? 20 : 2);
1889 uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]], ref2frm[list][ref[1]]) & 0x00FF00FF) * 0x0101;
1890 uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]], ref2frm[list][ref[3]]) & 0x00FF00FF) * 0x0101;
1891 AV_WN32A(&ref_cache[0 * 8], ref01);
1892 AV_WN32A(&ref_cache[1 * 8], ref01);
1893 AV_WN32A(&ref_cache[2 * 8], ref23);
1894 AV_WN32A(&ref_cache[3 * 8], ref23);
1898 int16_t(*mv_src)[2] = &h->cur_pic.motion_val[list][4 * h->mb_x + 4 * h->mb_y * b_stride];
1899 AV_COPY128(mv_dst + 8 * 0, mv_src + 0 * b_stride);
1900 AV_COPY128(mv_dst + 8 * 1, mv_src + 1 * b_stride);
1901 AV_COPY128(mv_dst + 8 * 2, mv_src + 2 * b_stride);
1902 AV_COPY128(mv_dst + 8 * 3, mv_src + 3 * b_stride);
1908 * @return non zero if the loop filter can be skipped
1910 static int fill_filter_caches(H264Context *h, H264SliceContext *sl, int mb_type)
1912 const int mb_xy = h->mb_xy;
1913 int top_xy, left_xy[LEFT_MBS];
1914 int top_type, left_type[LEFT_MBS];
1918 top_xy = mb_xy - (h->mb_stride << MB_FIELD(h));
1920 /* Wow, what a mess, why didn't they simplify the interlacing & intra
1921 * stuff, I can't imagine that these complex rules are worth it. */
1923 left_xy[LBOT] = left_xy[LTOP] = mb_xy - 1;
1924 if (FRAME_MBAFF(h)) {
1925 const int left_mb_field_flag = IS_INTERLACED(h->cur_pic.mb_type[mb_xy - 1]);
1926 const int curr_mb_field_flag = IS_INTERLACED(mb_type);
1928 if (left_mb_field_flag != curr_mb_field_flag)
1929 left_xy[LTOP] -= h->mb_stride;
1931 if (curr_mb_field_flag)
1932 top_xy += h->mb_stride &
1933 (((h->cur_pic.mb_type[top_xy] >> 7) & 1) - 1);
1934 if (left_mb_field_flag != curr_mb_field_flag)
1935 left_xy[LBOT] += h->mb_stride;
1939 sl->top_mb_xy = top_xy;
1940 sl->left_mb_xy[LTOP] = left_xy[LTOP];
1941 sl->left_mb_xy[LBOT] = left_xy[LBOT];
1943 /* For sufficiently low qp, filtering wouldn't do anything.
1944 * This is a conservative estimate: could also check beta_offset
1945 * and more accurate chroma_qp. */
1946 int qp_thresh = sl->qp_thresh; // FIXME strictly we should store qp_thresh for each mb of a slice
1947 int qp = h->cur_pic.qscale_table[mb_xy];
1948 if (qp <= qp_thresh &&
1949 (left_xy[LTOP] < 0 ||
1950 ((qp + h->cur_pic.qscale_table[left_xy[LTOP]] + 1) >> 1) <= qp_thresh) &&
1952 ((qp + h->cur_pic.qscale_table[top_xy] + 1) >> 1) <= qp_thresh)) {
1953 if (!FRAME_MBAFF(h))
1955 if ((left_xy[LTOP] < 0 ||
1956 ((qp + h->cur_pic.qscale_table[left_xy[LBOT]] + 1) >> 1) <= qp_thresh) &&
1957 (top_xy < h->mb_stride ||
1958 ((qp + h->cur_pic.qscale_table[top_xy - h->mb_stride] + 1) >> 1) <= qp_thresh))
1963 top_type = h->cur_pic.mb_type[top_xy];
1964 left_type[LTOP] = h->cur_pic.mb_type[left_xy[LTOP]];
1965 left_type[LBOT] = h->cur_pic.mb_type[left_xy[LBOT]];
1966 if (h->deblocking_filter == 2) {
1967 if (h->slice_table[top_xy] != sl->slice_num)
1969 if (h->slice_table[left_xy[LBOT]] != sl->slice_num)
1970 left_type[LTOP] = left_type[LBOT] = 0;
1972 if (h->slice_table[top_xy] == 0xFFFF)
1974 if (h->slice_table[left_xy[LBOT]] == 0xFFFF)
1975 left_type[LTOP] = left_type[LBOT] = 0;
1977 sl->top_type = top_type;
1978 sl->left_type[LTOP] = left_type[LTOP];
1979 sl->left_type[LBOT] = left_type[LBOT];
1981 if (IS_INTRA(mb_type))
1984 fill_filter_caches_inter(h, sl, mb_type, top_xy, left_xy,
1985 top_type, left_type, mb_xy, 0);
1986 if (sl->list_count == 2)
1987 fill_filter_caches_inter(h, sl, mb_type, top_xy, left_xy,
1988 top_type, left_type, mb_xy, 1);
1990 nnz = h->non_zero_count[mb_xy];
1991 nnz_cache = sl->non_zero_count_cache;
1992 AV_COPY32(&nnz_cache[4 + 8 * 1], &nnz[0]);
1993 AV_COPY32(&nnz_cache[4 + 8 * 2], &nnz[4]);
1994 AV_COPY32(&nnz_cache[4 + 8 * 3], &nnz[8]);
1995 AV_COPY32(&nnz_cache[4 + 8 * 4], &nnz[12]);
1996 h->cbp = h->cbp_table[mb_xy];
1999 nnz = h->non_zero_count[top_xy];
2000 AV_COPY32(&nnz_cache[4 + 8 * 0], &nnz[3 * 4]);
2003 if (left_type[LTOP]) {
2004 nnz = h->non_zero_count[left_xy[LTOP]];
2005 nnz_cache[3 + 8 * 1] = nnz[3 + 0 * 4];
2006 nnz_cache[3 + 8 * 2] = nnz[3 + 1 * 4];
2007 nnz_cache[3 + 8 * 3] = nnz[3 + 2 * 4];
2008 nnz_cache[3 + 8 * 4] = nnz[3 + 3 * 4];
2011 /* CAVLC 8x8dct requires NNZ values for residual decoding that differ
2012 * from what the loop filter needs */
2013 if (!CABAC(h) && h->pps.transform_8x8_mode) {
2014 if (IS_8x8DCT(top_type)) {
2015 nnz_cache[4 + 8 * 0] =
2016 nnz_cache[5 + 8 * 0] = (h->cbp_table[top_xy] & 0x4000) >> 12;
2017 nnz_cache[6 + 8 * 0] =
2018 nnz_cache[7 + 8 * 0] = (h->cbp_table[top_xy] & 0x8000) >> 12;
2020 if (IS_8x8DCT(left_type[LTOP])) {
2021 nnz_cache[3 + 8 * 1] =
2022 nnz_cache[3 + 8 * 2] = (h->cbp_table[left_xy[LTOP]] & 0x2000) >> 12; // FIXME check MBAFF
2024 if (IS_8x8DCT(left_type[LBOT])) {
2025 nnz_cache[3 + 8 * 3] =
2026 nnz_cache[3 + 8 * 4] = (h->cbp_table[left_xy[LBOT]] & 0x8000) >> 12; // FIXME check MBAFF
2029 if (IS_8x8DCT(mb_type)) {
2030 nnz_cache[scan8[0]] =
2031 nnz_cache[scan8[1]] =
2032 nnz_cache[scan8[2]] =
2033 nnz_cache[scan8[3]] = (h->cbp & 0x1000) >> 12;
2035 nnz_cache[scan8[0 + 4]] =
2036 nnz_cache[scan8[1 + 4]] =
2037 nnz_cache[scan8[2 + 4]] =
2038 nnz_cache[scan8[3 + 4]] = (h->cbp & 0x2000) >> 12;
2040 nnz_cache[scan8[0 + 8]] =
2041 nnz_cache[scan8[1 + 8]] =
2042 nnz_cache[scan8[2 + 8]] =
2043 nnz_cache[scan8[3 + 8]] = (h->cbp & 0x4000) >> 12;
2045 nnz_cache[scan8[0 + 12]] =
2046 nnz_cache[scan8[1 + 12]] =
2047 nnz_cache[scan8[2 + 12]] =
2048 nnz_cache[scan8[3 + 12]] = (h->cbp & 0x8000) >> 12;
2055 static void loop_filter(H264Context *h, H264SliceContext *sl, int start_x, int end_x)
2057 uint8_t *dest_y, *dest_cb, *dest_cr;
2058 int linesize, uvlinesize, mb_x, mb_y;
2059 const int end_mb_y = h->mb_y + FRAME_MBAFF(h);
2060 const int old_slice_type = sl->slice_type;
2061 const int pixel_shift = h->pixel_shift;
2062 const int block_h = 16 >> h->chroma_y_shift;
2064 if (h->deblocking_filter) {
2065 for (mb_x = start_x; mb_x < end_x; mb_x++)
2066 for (mb_y = end_mb_y - FRAME_MBAFF(h); mb_y <= end_mb_y; mb_y++) {
2068 mb_xy = h->mb_xy = mb_x + mb_y * h->mb_stride;
2069 sl->slice_num = h->slice_table[mb_xy];
2070 mb_type = h->cur_pic.mb_type[mb_xy];
2071 sl->list_count = h->list_counts[mb_xy];
2075 h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
2079 dest_y = h->cur_pic.f.data[0] +
2080 ((mb_x << pixel_shift) + mb_y * h->linesize) * 16;
2081 dest_cb = h->cur_pic.f.data[1] +
2082 (mb_x << pixel_shift) * (8 << CHROMA444(h)) +
2083 mb_y * h->uvlinesize * block_h;
2084 dest_cr = h->cur_pic.f.data[2] +
2085 (mb_x << pixel_shift) * (8 << CHROMA444(h)) +
2086 mb_y * h->uvlinesize * block_h;
2087 // FIXME simplify above
2090 linesize = sl->mb_linesize = h->linesize * 2;
2091 uvlinesize = sl->mb_uvlinesize = h->uvlinesize * 2;
2092 if (mb_y & 1) { // FIXME move out of this function?
2093 dest_y -= h->linesize * 15;
2094 dest_cb -= h->uvlinesize * (block_h - 1);
2095 dest_cr -= h->uvlinesize * (block_h - 1);
2098 linesize = sl->mb_linesize = h->linesize;
2099 uvlinesize = sl->mb_uvlinesize = h->uvlinesize;
2101 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize,
2103 if (fill_filter_caches(h, sl, mb_type))
2105 sl->chroma_qp[0] = get_chroma_qp(h, 0, h->cur_pic.qscale_table[mb_xy]);
2106 sl->chroma_qp[1] = get_chroma_qp(h, 1, h->cur_pic.qscale_table[mb_xy]);
2108 if (FRAME_MBAFF(h)) {
2109 ff_h264_filter_mb(h, sl, mb_x, mb_y, dest_y, dest_cb, dest_cr,
2110 linesize, uvlinesize);
2112 ff_h264_filter_mb_fast(h, sl, mb_x, mb_y, dest_y, dest_cb,
2113 dest_cr, linesize, uvlinesize);
2117 sl->slice_type = old_slice_type;
2119 h->mb_y = end_mb_y - FRAME_MBAFF(h);
2120 sl->chroma_qp[0] = get_chroma_qp(h, 0, sl->qscale);
2121 sl->chroma_qp[1] = get_chroma_qp(h, 1, sl->qscale);
2124 static void predict_field_decoding_flag(H264Context *h, H264SliceContext *sl)
2126 const int mb_xy = h->mb_x + h->mb_y * h->mb_stride;
2127 int mb_type = (h->slice_table[mb_xy - 1] == sl->slice_num) ?
2128 h->cur_pic.mb_type[mb_xy - 1] :
2129 (h->slice_table[mb_xy - h->mb_stride] == sl->slice_num) ?
2130 h->cur_pic.mb_type[mb_xy - h->mb_stride] : 0;
2131 h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
2135 * Draw edges and report progress for the last MB row.
2137 static void decode_finish_row(H264Context *h, H264SliceContext *sl)
2139 int top = 16 * (h->mb_y >> FIELD_PICTURE(h));
2140 int pic_height = 16 * h->mb_height >> FIELD_PICTURE(h);
2141 int height = 16 << FRAME_MBAFF(h);
2142 int deblock_border = (16 + 4) << FRAME_MBAFF(h);
2144 if (h->deblocking_filter) {
2145 if ((top + height) >= pic_height)
2146 height += deblock_border;
2147 top -= deblock_border;
2150 if (top >= pic_height || (top + height) < 0)
2153 height = FFMIN(height, pic_height - top);
2155 height = top + height;
2159 ff_h264_draw_horiz_band(h, sl, top, height);
2164 ff_thread_report_progress(&h->cur_pic_ptr->tf, top + height - 1,
2165 h->picture_structure == PICT_BOTTOM_FIELD);
2168 static void er_add_slice(H264Context *h, H264SliceContext *sl,
2169 int startx, int starty,
2170 int endx, int endy, int status)
2172 #if CONFIG_ERROR_RESILIENCE
2173 ERContext *er = &h->er;
2175 er->ref_count = sl->ref_count[0];
2176 ff_er_add_slice(er, startx, starty, endx, endy, status);
2180 static int decode_slice(struct AVCodecContext *avctx, void *arg)
2182 H264SliceContext *sl = arg;
2183 H264Context *h = sl->h264;
2184 int lf_x_start = h->mb_x;
2186 h->mb_skip_run = -1;
2188 h->is_complex = FRAME_MBAFF(h) || h->picture_structure != PICT_FRAME ||
2189 avctx->codec_id != AV_CODEC_ID_H264 ||
2190 (CONFIG_GRAY && (h->flags & CODEC_FLAG_GRAY));
2194 align_get_bits(&h->gb);
2197 ff_init_cabac_decoder(&h->cabac,
2198 h->gb.buffer + get_bits_count(&h->gb) / 8,
2199 (get_bits_left(&h->gb) + 7) / 8);
2201 ff_h264_init_cabac_states(h, sl);
2205 int ret = ff_h264_decode_mb_cabac(h, sl);
2207 // STOP_TIMER("decode_mb_cabac")
2210 ff_h264_hl_decode_mb(h, sl);
2212 // FIXME optimal? or let mb_decode decode 16x32 ?
2213 if (ret >= 0 && FRAME_MBAFF(h)) {
2216 ret = ff_h264_decode_mb_cabac(h, sl);
2219 ff_h264_hl_decode_mb(h, sl);
2222 eos = get_cabac_terminate(&h->cabac);
2224 if ((h->workaround_bugs & FF_BUG_TRUNCATED) &&
2225 h->cabac.bytestream > h->cabac.bytestream_end + 2) {
2226 er_add_slice(h, sl, h->resync_mb_x, h->resync_mb_y, h->mb_x - 1,
2227 h->mb_y, ER_MB_END);
2228 if (h->mb_x >= lf_x_start)
2229 loop_filter(h, sl, lf_x_start, h->mb_x + 1);
2232 if (ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 2) {
2233 av_log(h->avctx, AV_LOG_ERROR,
2234 "error while decoding MB %d %d, bytestream %td\n",
2236 h->cabac.bytestream_end - h->cabac.bytestream);
2237 er_add_slice(h, sl, h->resync_mb_x, h->resync_mb_y, h->mb_x,
2238 h->mb_y, ER_MB_ERROR);
2239 return AVERROR_INVALIDDATA;
2242 if (++h->mb_x >= h->mb_width) {
2243 loop_filter(h, sl, lf_x_start, h->mb_x);
2244 h->mb_x = lf_x_start = 0;
2245 decode_finish_row(h, sl);
2247 if (FIELD_OR_MBAFF_PICTURE(h)) {
2249 if (FRAME_MBAFF(h) && h->mb_y < h->mb_height)
2250 predict_field_decoding_flag(h, sl);
2254 if (eos || h->mb_y >= h->mb_height) {
2255 tprintf(h->avctx, "slice end %d %d\n",
2256 get_bits_count(&h->gb), h->gb.size_in_bits);
2257 er_add_slice(h, sl, h->resync_mb_x, h->resync_mb_y, h->mb_x - 1,
2258 h->mb_y, ER_MB_END);
2259 if (h->mb_x > lf_x_start)
2260 loop_filter(h, sl, lf_x_start, h->mb_x);
2266 int ret = ff_h264_decode_mb_cavlc(h, sl);
2269 ff_h264_hl_decode_mb(h, sl);
2271 // FIXME optimal? or let mb_decode decode 16x32 ?
2272 if (ret >= 0 && FRAME_MBAFF(h)) {
2274 ret = ff_h264_decode_mb_cavlc(h, sl);
2277 ff_h264_hl_decode_mb(h, sl);
2282 av_log(h->avctx, AV_LOG_ERROR,
2283 "error while decoding MB %d %d\n", h->mb_x, h->mb_y);
2284 er_add_slice(h, sl, h->resync_mb_x, h->resync_mb_y, h->mb_x,
2285 h->mb_y, ER_MB_ERROR);
2289 if (++h->mb_x >= h->mb_width) {
2290 loop_filter(h, sl, lf_x_start, h->mb_x);
2291 h->mb_x = lf_x_start = 0;
2292 decode_finish_row(h, sl);
2294 if (FIELD_OR_MBAFF_PICTURE(h)) {
2296 if (FRAME_MBAFF(h) && h->mb_y < h->mb_height)
2297 predict_field_decoding_flag(h, sl);
2299 if (h->mb_y >= h->mb_height) {
2300 tprintf(h->avctx, "slice end %d %d\n",
2301 get_bits_count(&h->gb), h->gb.size_in_bits);
2303 if (get_bits_left(&h->gb) == 0) {
2304 er_add_slice(h, sl, h->resync_mb_x, h->resync_mb_y,
2305 h->mb_x - 1, h->mb_y, ER_MB_END);
2309 er_add_slice(h, sl, h->resync_mb_x, h->resync_mb_y,
2310 h->mb_x - 1, h->mb_y, ER_MB_END);
2312 return AVERROR_INVALIDDATA;
2317 if (get_bits_left(&h->gb) <= 0 && h->mb_skip_run <= 0) {
2318 tprintf(h->avctx, "slice end %d %d\n",
2319 get_bits_count(&h->gb), h->gb.size_in_bits);
2321 if (get_bits_left(&h->gb) == 0) {
2322 er_add_slice(h, sl, h->resync_mb_x, h->resync_mb_y,
2323 h->mb_x - 1, h->mb_y, ER_MB_END);
2324 if (h->mb_x > lf_x_start)
2325 loop_filter(h, sl, lf_x_start, h->mb_x);
2329 er_add_slice(h, sl, h->resync_mb_x, h->resync_mb_y, h->mb_x,
2330 h->mb_y, ER_MB_ERROR);
2332 return AVERROR_INVALIDDATA;
2340 * Call decode_slice() for each context.
2342 * @param h h264 master context
2343 * @param context_count number of contexts to execute
2345 int ff_h264_execute_decode_slices(H264Context *h, unsigned context_count)
2347 AVCodecContext *const avctx = h->avctx;
2351 if (h->avctx->hwaccel)
2353 if (context_count == 1) {
2354 return decode_slice(avctx, &h->slice_ctx[0]);
2356 for (i = 1; i < context_count; i++) {
2357 hx = h->thread_context[i];
2358 hx->er.error_count = 0;
2361 avctx->execute(avctx, decode_slice, h->slice_ctx,
2362 NULL, context_count, sizeof(h->slice_ctx[0]));
2364 /* pull back stuff from slices to master context */
2365 hx = h->thread_context[context_count - 1];
2368 h->droppable = hx->droppable;
2369 h->picture_structure = hx->picture_structure;
2370 for (i = 1; i < context_count; i++)
2371 h->er.error_count += h->thread_context[i]->er.error_count;