projects / libav.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
lavc: Rename avpriv_mpv_find_start_code after moving out from mpegvideo
[libav.git] / libavcodec / h264.c
CommitLineData
0da71265 1/*
ff3d4310 2 * H.26L/H.264/AVC/JVT/14496-10/... decoder
0da71265
MN		 3 * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
4 *
2912e87a 5 * This file is part of Libav.
b78e7197 6 *
2912e87a 7 * Libav is free software; you can redistribute it and/or
0da71265
MN		 8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
b78e7197 10 * version 2.1 of the License, or (at your option) any later version.
0da71265 11 *
2912e87a 12 * Libav is distributed in the hope that it will be useful,
0da71265
MN		 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.
16 *
17 * You should have received a copy of the GNU Lesser General Public
2912e87a 18 * License along with Libav; if not, write to the Free Software
5509bffa 19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
0da71265 20 */
115329f1 21
0da71265 22/**
ba87f080 23 * @file
0da71265
MN		 24 * H.264 / AVC / MPEG4 part10 codec.
25 * @author Michael Niedermayer <michaelni@gmx.at>
26 */
27
737eb597 28#include "libavutil/imgutils.h"
40e5d31b 29#include "internal.h"
55b9ef18
DB		 30#include "cabac.h"
31#include "cabac_functions.h"
0da71265 32#include "dsputil.h"
5f401b7b 33#include "error_resilience.h"
0da71265
MN		 34#include "avcodec.h"
35#include "mpegvideo.h"
26b4fe82 36#include "h264.h"
0da71265 37#include "h264data.h"
79dad2a9 38#include "h264chroma.h"
188d3c51 39#include "h264_mvpred.h"
0da71265 40#include "golomb.h"
199436b9 41#include "mathops.h"
626464fb 42#include "rectangle.h"
75d5156a 43#include "svq3.h"
6a9c8594 44#include "thread.h"
369122dd 45#include "vdpau_internal.h"
cfa5a81e 46#include "libavutil/avassert.h"
0da71265 47
e5d40372 48// #undef NDEBUG
0da71265
MN		 49#include <assert.h>
50
0becb078
DB		 51const uint16_t ff_h264_mb_sizes[4] = { 256, 384, 512, 768 };
52
e5d40372
DB		 53static const uint8_t rem6[QP_MAX_NUM + 1] = {
54 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2,
55 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5,
56 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
acd8d10f
PI		 57};
58
e5d40372
DB		 59static const uint8_t div6[QP_MAX_NUM + 1] = {
60 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3, 3, 3,
61 3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6,
62 7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 10, 10, 10, 10,
acd8d10f
PI		 63};
64
8d061989
RB		 65static const enum AVPixelFormat h264_hwaccel_pixfmt_list_420[] = {
66#if CONFIG_H264_DXVA2_HWACCEL
67 AV_PIX_FMT_DXVA2_VLD,
68#endif
69#if CONFIG_H264_VAAPI_HWACCEL
70 AV_PIX_FMT_VAAPI_VLD,
71#endif
72#if CONFIG_H264_VDA_HWACCEL
73 AV_PIX_FMT_VDA_VLD,
74#endif
75#if CONFIG_H264_VDPAU_HWACCEL
76 AV_PIX_FMT_VDPAU,
77#endif
78 AV_PIX_FMT_YUV420P,
79 AV_PIX_FMT_NONE
80};
81
d65522e8 82static const enum AVPixelFormat h264_hwaccel_pixfmt_list_jpeg_420[] = {
78bc4d69 83#if CONFIG_H264_DXVA2_HWACCEL
716d413c 84 AV_PIX_FMT_DXVA2_VLD,
78bc4d69
RDC		 85#endif
86#if CONFIG_H264_VAAPI_HWACCEL
716d413c 87 AV_PIX_FMT_VAAPI_VLD,
78bc4d69
RDC		 88#endif
89#if CONFIG_H264_VDA_HWACCEL
716d413c 90 AV_PIX_FMT_VDA_VLD,
78bc4d69
RDC		 91#endif
92#if CONFIG_H264_VDPAU_HWACCEL
ec0e9200 93 AV_PIX_FMT_VDPAU,
78bc4d69 94#endif
716d413c
AK		 95 AV_PIX_FMT_YUVJ420P,
96 AV_PIX_FMT_NONE
0435fb16
BC		 97};
98
54974c62
AK		 99static void h264_er_decode_mb(void *opaque, int ref, int mv_dir, int mv_type,
100 int (*mv)[2][4][2],
101 int mb_x, int mb_y, int mb_intra, int mb_skipped)
102{
103 H264Context *h = opaque;
54974c62 104
2c541554
AK		 105 h->mb_x = mb_x;
106 h->mb_y = mb_y;
107 h->mb_xy = mb_x + mb_y * h->mb_stride;
54974c62
AK		 108 memset(h->non_zero_count_cache, 0, sizeof(h->non_zero_count_cache));
109 assert(ref >= 0);
110 /* FIXME: It is possible albeit uncommon that slice references
111 * differ between slices. We take the easy approach and ignore
112 * it for now. If this turns out to have any relevance in
113 * practice then correct remapping should be added. */
114 if (ref >= h->ref_count[0])
115 ref = 0;
759001c5 116 fill_rectangle(&h->cur_pic.ref_index[0][4 * h->mb_xy],
54974c62
AK		 117 2, 2, 2, ref, 1);
118 fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1);
119 fill_rectangle(h->mv_cache[0][scan8[0]], 4, 4, 8,
2c541554 120 pack16to32((*mv)[0][0][0], (*mv)[0][0][1]), 4);
7bece9b2 121 assert(!FRAME_MBAFF(h));
54974c62
AK		 122 ff_h264_hl_decode_mb(h);
123}
124
2c541554
AK		 125void ff_h264_draw_horiz_band(H264Context *h, int y, int height)
126{
64e43869
RB		 127 AVCodecContext *avctx = h->avctx;
128 Picture *cur = &h->cur_pic;
129 Picture *last = h->ref_list[0][0].f.data[0] ? &h->ref_list[0][0] : NULL;
130 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(avctx->pix_fmt);
131 int vshift = desc->log2_chroma_h;
132 const int field_pic = h->picture_structure != PICT_FRAME;
133 if (field_pic) {
134 height <<= 1;
135 y <<= 1;
136 }
137
138 height = FFMIN(height, avctx->height - y);
139
140 if (field_pic && h->first_field && !(avctx->slice_flags & SLICE_FLAG_ALLOW_FIELD))
141 return;
142
143 if (avctx->draw_horiz_band) {
144 AVFrame *src;
145 int offset[AV_NUM_DATA_POINTERS];
146 int i;
147
148 if (cur->f.pict_type == AV_PICTURE_TYPE_B || h->low_delay ||
149 (avctx->slice_flags & SLICE_FLAG_CODED_ORDER))
150 src = &cur->f;
151 else if (last)
152 src = &last->f;
153 else
154 return;
155
156 offset[0] = y * src->linesize[0];
157 offset[1] =
158 offset[2] = (y >> vshift) * src->linesize[1];
159 for (i = 3; i < AV_NUM_DATA_POINTERS; i++)
160 offset[i] = 0;
161
162 emms_c();
163
164 avctx->draw_horiz_band(avctx, src, offset,
165 y, h->picture_structure, height);
166 }
2c541554
AK		 167}
168
759001c5 169static void unref_picture(H264Context *h, Picture *pic)
2c541554 170{
759001c5 171 int off = offsetof(Picture, tf) + sizeof(pic->tf);
2c541554
AK		 172 int i;
173
759001c5
AK		 174 if (!pic->f.data[0])
175 return;
176
177 ff_thread_release_buffer(h->avctx, &pic->tf);
178 av_buffer_unref(&pic->hwaccel_priv_buf);
2c541554 179
759001c5
AK		 180 av_buffer_unref(&pic->qscale_table_buf);
181 av_buffer_unref(&pic->mb_type_buf);
2c541554 182 for (i = 0; i < 2; i++) {
759001c5
AK		 183 av_buffer_unref(&pic->motion_val_buf[i]);
184 av_buffer_unref(&pic->ref_index_buf[i]);
2c541554 185 }
759001c5
AK		 186
187 memset((uint8_t*)pic + off, 0, sizeof(*pic) - off);
2c541554
AK		 188}
189
190static void release_unused_pictures(H264Context *h, int remove_current)
191{
192 int i;
193
194 /* release non reference frames */
759001c5
AK		 195 for (i = 0; i < MAX_PICTURE_COUNT; i++) {
196 if (h->DPB[i].f.data[0] && !h->DPB[i].reference &&
2c541554 197 (remove_current || &h->DPB[i] != h->cur_pic_ptr)) {
759001c5 198 unref_picture(h, &h->DPB[i]);
2c541554
AK		 199 }
200 }
201}
202
759001c5
AK		 203static int ref_picture(H264Context *h, Picture *dst, Picture *src)
204{
205 int ret, i;
206
207 av_assert0(!dst->f.buf[0]);
208 av_assert0(src->f.buf[0]);
209
210 src->tf.f = &src->f;
211 dst->tf.f = &dst->f;
212 ret = ff_thread_ref_frame(&dst->tf, &src->tf);
213 if (ret < 0)
214 goto fail;
215
216
217 dst->qscale_table_buf = av_buffer_ref(src->qscale_table_buf);
218 dst->mb_type_buf = av_buffer_ref(src->mb_type_buf);
219 if (!dst->qscale_table_buf || !dst->mb_type_buf)
220 goto fail;
221 dst->qscale_table = src->qscale_table;
222 dst->mb_type = src->mb_type;
223
224 for (i = 0; i < 2; i ++) {
225 dst->motion_val_buf[i] = av_buffer_ref(src->motion_val_buf[i]);
226 dst->ref_index_buf[i] = av_buffer_ref(src->ref_index_buf[i]);
227 if (!dst->motion_val_buf[i] || !dst->ref_index_buf[i])
228 goto fail;
229 dst->motion_val[i] = src->motion_val[i];
230 dst->ref_index[i] = src->ref_index[i];
231 }
232
233 if (src->hwaccel_picture_private) {
234 dst->hwaccel_priv_buf = av_buffer_ref(src->hwaccel_priv_buf);
235 if (!dst->hwaccel_priv_buf)
236 goto fail;
237 dst->hwaccel_picture_private = dst->hwaccel_priv_buf->data;
238 }
239
240 for (i = 0; i < 2; i++)
241 dst->field_poc[i] = src->field_poc[i];
242
243 memcpy(dst->ref_poc, src->ref_poc, sizeof(src->ref_poc));
244 memcpy(dst->ref_count, src->ref_count, sizeof(src->ref_count));
245
246 dst->poc = src->poc;
247 dst->frame_num = src->frame_num;
248 dst->mmco_reset = src->mmco_reset;
249 dst->pic_id = src->pic_id;
250 dst->long_ref = src->long_ref;
251 dst->mbaff = src->mbaff;
252 dst->field_picture = src->field_picture;
253 dst->needs_realloc = src->needs_realloc;
254 dst->reference = src->reference;
255
256 return 0;
257fail:
258 unref_picture(h, dst);
259 return ret;
260}
261
262
2c541554
AK		 263static int alloc_scratch_buffers(H264Context *h, int linesize)
264{
265 int alloc_size = FFALIGN(FFABS(linesize) + 32, 32);
266
267 if (h->bipred_scratchpad)
268 return 0;
269
270 h->bipred_scratchpad = av_malloc(16 * 6 * alloc_size);
271 // edge emu needs blocksize + filter length - 1
272 // (= 21x21 for h264)
273 h->edge_emu_buffer = av_mallocz(alloc_size * 2 * 21);
274 h->me.scratchpad = av_mallocz(alloc_size * 2 * 16 * 2);
275
276 if (!h->bipred_scratchpad || !h->edge_emu_buffer || !h->me.scratchpad) {
277 av_freep(&h->bipred_scratchpad);
278 av_freep(&h->edge_emu_buffer);
279 av_freep(&h->me.scratchpad);
280 return AVERROR(ENOMEM);
281 }
282
283 h->me.temp = h->me.scratchpad;
284
285 return 0;
286}
287
759001c5 288static int init_table_pools(H264Context *h)
2c541554
AK		 289{
290 const int big_mb_num = h->mb_stride * (h->mb_height + 1) + 1;
291 const int mb_array_size = h->mb_stride * h->mb_height;
292 const int b4_stride = h->mb_width * 4 + 1;
293 const int b4_array_size = b4_stride * h->mb_height * 4;
759001c5
AK		 294
295 h->qscale_table_pool = av_buffer_pool_init(big_mb_num + h->mb_stride,
296 av_buffer_allocz);
297 h->mb_type_pool = av_buffer_pool_init((big_mb_num + h->mb_stride) *
298 sizeof(uint32_t), av_buffer_allocz);
299 h->motion_val_pool = av_buffer_pool_init(2 * (b4_array_size + 4) *
300 sizeof(int16_t), av_buffer_allocz);
301 h->ref_index_pool = av_buffer_pool_init(4 * mb_array_size, av_buffer_allocz);
302
303 if (!h->qscale_table_pool || !h->mb_type_pool || !h->motion_val_pool ||
304 !h->ref_index_pool) {
305 av_buffer_pool_uninit(&h->qscale_table_pool);
306 av_buffer_pool_uninit(&h->mb_type_pool);
307 av_buffer_pool_uninit(&h->motion_val_pool);
308 av_buffer_pool_uninit(&h->ref_index_pool);
309 return AVERROR(ENOMEM);
310 }
311
312 return 0;
313}
314
315static int alloc_picture(H264Context *h, Picture *pic)
316{
2c541554
AK		 317 int i, ret = 0;
318
319 av_assert0(!pic->f.data[0]);
320
9c9ede44
HL		 321 pic->tf.f = &pic->f;
322 ret = ff_thread_get_buffer(h->avctx, &pic->tf, pic->reference ?
323 AV_GET_BUFFER_FLAG_REF : 0);
324 if (ret < 0)
325 goto fail;
326
327 h->linesize = pic->f.linesize[0];
328 h->uvlinesize = pic->f.linesize[1];
329
2c541554
AK		 330 if (h->avctx->hwaccel) {
331 const AVHWAccel *hwaccel = h->avctx->hwaccel;
759001c5 332 av_assert0(!pic->hwaccel_picture_private);
2c541554 333 if (hwaccel->priv_data_size) {
759001c5
AK		 334 pic->hwaccel_priv_buf = av_buffer_allocz(hwaccel->priv_data_size);
335 if (!pic->hwaccel_priv_buf)
2c541554 336 return AVERROR(ENOMEM);
759001c5 337 pic->hwaccel_picture_private = pic->hwaccel_priv_buf->data;
2c541554
AK		 338 }
339 }
2c541554 340
759001c5
AK		 341 if (!h->qscale_table_pool) {
342 ret = init_table_pools(h);
343 if (ret < 0)
344 goto fail;
345 }
2c541554 346
759001c5
AK		 347 pic->qscale_table_buf = av_buffer_pool_get(h->qscale_table_pool);
348 pic->mb_type_buf = av_buffer_pool_get(h->mb_type_pool);
349 if (!pic->qscale_table_buf || !pic->mb_type_buf)
350 goto fail;
2c541554 351
759001c5
AK		 352 pic->mb_type = (uint32_t*)pic->mb_type_buf->data + 2 * h->mb_stride + 1;
353 pic->qscale_table = pic->qscale_table_buf->data + 2 * h->mb_stride + 1;
354
355 for (i = 0; i < 2; i++) {
356 pic->motion_val_buf[i] = av_buffer_pool_get(h->motion_val_pool);
357 pic->ref_index_buf[i] = av_buffer_pool_get(h->ref_index_pool);
358 if (!pic->motion_val_buf[i] || !pic->ref_index_buf[i])
359 goto fail;
2c541554 360
759001c5
AK		 361 pic->motion_val[i] = (int16_t (*)[2])pic->motion_val_buf[i]->data + 4;
362 pic->ref_index[i] = pic->ref_index_buf[i]->data;
363 }
2c541554
AK		 364
365 return 0;
366fail:
759001c5 367 unref_picture(h, pic);
2c541554
AK		 368 return (ret < 0) ? ret : AVERROR(ENOMEM);
369}
370
371static inline int pic_is_unused(H264Context *h, Picture *pic)
372{
373 if (pic->f.data[0] == NULL)
374 return 1;
759001c5
AK		 375 if (pic->needs_realloc && !(pic->reference & DELAYED_PIC_REF))
376 return 1;
2c541554
AK		 377 return 0;
378}
379
380static int find_unused_picture(H264Context *h)
381{
382 int i;
383
759001c5 384 for (i = 0; i < MAX_PICTURE_COUNT; i++) {
2c541554
AK		 385 if (pic_is_unused(h, &h->DPB[i]))
386 break;
387 }
759001c5 388 if (i == MAX_PICTURE_COUNT)
2c541554
AK		 389 return AVERROR_INVALIDDATA;
390
391 if (h->DPB[i].needs_realloc) {
392 h->DPB[i].needs_realloc = 0;
759001c5 393 unref_picture(h, &h->DPB[i]);
2c541554
AK		 394 }
395
396 return i;
397}
398
0da71265 399/**
58c42af7
DB		 400 * Check if the top & left blocks are available if needed and
401 * change the dc mode so it only uses the available blocks.
0da71265 402 */
e5d40372
DB		 403int ff_h264_check_intra4x4_pred_mode(H264Context *h)
404{
e5d40372
DB		 405 static const int8_t top[12] = {
406 -1, 0, LEFT_DC_PRED, -1, -1, -1, -1, -1, 0
407 };
408 static const int8_t left[12] = {
409 0, -1, TOP_DC_PRED, 0, -1, -1, -1, 0, -1, DC_128_PRED
410 };
2bedc0e8
MN		 411 int i;
412
e5d40372
DB		 413 if (!(h->top_samples_available & 0x8000)) {
414 for (i = 0; i < 4; i++) {
415 int status = top[h->intra4x4_pred_mode_cache[scan8[0] + i]];
416 if (status < 0) {
2c541554 417 av_log(h->avctx, AV_LOG_ERROR,
e5d40372 418 "top block unavailable for requested intra4x4 mode %d at %d %d\n",
2c541554 419 status, h->mb_x, h->mb_y);
2bedc0e8 420 return -1;
e5d40372
DB		 421 } else if (status) {
422 h->intra4x4_pred_mode_cache[scan8[0] + i] = status;
2bedc0e8
MN		 423 }
424 }
425 }
426
e5d40372
DB		 427 if ((h->left_samples_available & 0x8888) != 0x8888) {
428 static const int mask[4] = { 0x8000, 0x2000, 0x80, 0x20 };
429 for (i = 0; i < 4; i++)
430 if (!(h->left_samples_available & mask[i])) {
431 int status = left[h->intra4x4_pred_mode_cache[scan8[0] + 8 * i]];
432 if (status < 0) {
2c541554 433 av_log(h->avctx, AV_LOG_ERROR,
e5d40372 434 "left block unavailable for requested intra4x4 mode %d at %d %d\n",
2c541554 435 status, h->mb_x, h->mb_y);
2bedc0e8 436 return -1;
e5d40372
DB		 437 } else if (status) {
438 h->intra4x4_pred_mode_cache[scan8[0] + 8 * i] = status;
2bedc0e8
MN		 439 }
440 }
2bedc0e8
MN		 441 }
442
443 return 0;
e5d40372 444} // FIXME cleanup like ff_h264_check_intra_pred_mode
2bedc0e8
MN		 445
446/**
58c42af7
DB		 447 * Check if the top & left blocks are available if needed and
448 * change the dc mode so it only uses the available blocks.
2bedc0e8 449 */
e5d40372
DB		 450int ff_h264_check_intra_pred_mode(H264Context *h, int mode, int is_chroma)
451{
e5d40372
DB		 452 static const int8_t top[7] = { LEFT_DC_PRED8x8, 1, -1, -1 };
453 static const int8_t left[7] = { TOP_DC_PRED8x8, -1, 2, -1, DC_128_PRED8x8 };
454
455 if (mode > 6U) {
2c541554 456 av_log(h->avctx, AV_LOG_ERROR,
e5d40372 457 "out of range intra chroma pred mode at %d %d\n",
2c541554 458 h->mb_x, h->mb_y);
7440fe83 459 return -1;
5175b937 460 }
115329f1 461
e5d40372
DB		 462 if (!(h->top_samples_available & 0x8000)) {
463 mode = top[mode];
464 if (mode < 0) {
2c541554 465 av_log(h->avctx, AV_LOG_ERROR,
e5d40372 466 "top block unavailable for requested intra mode at %d %d\n",
2c541554 467 h->mb_x, h->mb_y);
0da71265
MN		 468 return -1;
469 }
470 }
115329f1 471
e5d40372
DB		 472 if ((h->left_samples_available & 0x8080) != 0x8080) {
473 mode = left[mode];
474 if (is_chroma && (h->left_samples_available & 0x8080)) {
475 // mad cow disease mode, aka MBAFF + constrained_intra_pred
476 mode = ALZHEIMER_DC_L0T_PRED8x8 +
477 (!(h->left_samples_available & 0x8000)) +
478 2 * (mode == DC_128_PRED8x8);
d1d10e91 479 }
e5d40372 480 if (mode < 0) {
2c541554 481 av_log(h->avctx, AV_LOG_ERROR,
e5d40372 482 "left block unavailable for requested intra mode at %d %d\n",
2c541554 483 h->mb_x, h->mb_y);
0da71265 484 return -1;
115329f1 485 }
0da71265
MN		 486 }
487
488 return mode;
489}
490
e5d40372
DB		 491const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src,
492 int *dst_length, int *consumed, int length)
493{
0da71265
MN		 494 int i, si, di;
495 uint8_t *dst;
24456882 496 int bufidx;
0da71265 497
e5d40372
DB		 498 // src[0]&0x80; // forbidden bit
499 h->nal_ref_idc = src[0] >> 5;
500 h->nal_unit_type = src[0] & 0x1F;
0da71265 501
e5d40372
DB		 502 src++;
503 length--;
e08715d3 504
58db34aa
RB		 505#define STARTCODE_TEST \
506 if (i + 2 < length && src[i + 1] == 0 && src[i + 2] <= 3) { \
507 if (src[i + 2] != 3) { \
508 /* startcode, so we must be past the end */ \
509 length = i; \
510 } \
511 break; \
512 }
b250f9c6 513#if HAVE_FAST_UNALIGNED
58db34aa
RB		 514#define FIND_FIRST_ZERO \
515 if (i > 0 && !src[i]) \
516 i--; \
517 while (src[i]) \
518 i++
e5d40372 519#if HAVE_FAST_64BIT
e5d40372
DB		 520 for (i = 0; i + 1 < length; i += 9) {
521 if (!((~AV_RN64A(src + i) &
522 (AV_RN64A(src + i) - 0x0100010001000101ULL)) &
523 0x8000800080008080ULL))
58db34aa
RB		 524 continue;
525 FIND_FIRST_ZERO;
526 STARTCODE_TEST;
527 i -= 7;
528 }
e5d40372 529#else
e5d40372
DB		 530 for (i = 0; i + 1 < length; i += 5) {
531 if (!((~AV_RN32A(src + i) &
532 (AV_RN32A(src + i) - 0x01000101U)) &
533 0x80008080U))
e08715d3 534 continue;
58db34aa
RB		 535 FIND_FIRST_ZERO;
536 STARTCODE_TEST;
537 i -= 3;
538 }
539#endif
e08715d3 540#else
e5d40372
DB		 541 for (i = 0; i + 1 < length; i += 2) {
542 if (src[i])
543 continue;
544 if (i > 0 && src[i - 1] == 0)
545 i--;
58db34aa 546 STARTCODE_TEST;
0da71265 547 }
58db34aa 548#endif
0da71265 549
e5d40372
DB		 550 if (i >= length - 1) { // no escaped 0
551 *dst_length = length;
552 *consumed = length + 1; // +1 for the header
115329f1 553 return src;
0da71265
MN		 554 }
555
e5d40372
DB		 556 // use second escape buffer for inter data
557 bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0;
558 av_fast_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx],
559 length + FF_INPUT_BUFFER_PADDING_SIZE);
560 dst = h->rbsp_buffer[bufidx];
0da71265 561
e5d40372 562 if (dst == NULL)
ac658be5 563 return NULL;
ac658be5 564
593af7cd 565 memcpy(dst, src, i);
e5d40372
DB		 566 si = di = i;
567 while (si + 2 < length) {
568 // remove escapes (very rare 1:2^22)
569 if (src[si + 2] > 3) {
570 dst[di++] = src[si++];
571 dst[di++] = src[si++];
572 } else if (src[si] == 0 && src[si + 1] == 0) {
573 if (src[si + 2] == 3) { // escape
574 dst[di++] = 0;
575 dst[di++] = 0;
576 si += 3;
c8470cc1 577 continue;
e5d40372 578 } else // next start code
593af7cd 579 goto nsc;
0da71265
MN		 580 }
581
e5d40372 582 dst[di++] = src[si++];
0da71265 583 }
e5d40372
DB		 584 while (si < length)
585 dst[di++] = src[si++];
593af7cd 586nsc:
0da71265 587
e5d40372 588 memset(dst + di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
d4369630 589
e5d40372
DB		 590 *dst_length = di;
591 *consumed = si + 1; // +1 for the header
592 /* FIXME store exact number of bits in the getbitcontext
593 * (it is needed for decoding) */
0da71265
MN		 594 return dst;
595}
596
85297319
DEP		 597/**
598 * Identify the exact end of the bitstream
599 * @return the length of the trailing, or 0 if damaged
600 */
b691fd7a 601static int decode_rbsp_trailing(H264Context *h, const uint8_t *src)
e5d40372
DB		 602{
603 int v = *src;
0da71265
MN		 604 int r;
605
2c541554 606 tprintf(h->avctx, "rbsp trailing %X\n", v);
0da71265 607
e5d40372
DB		 608 for (r = 1; r < 9; r++) {
609 if (v & 1)
610 return r;
611 v >>= 1;
0da71265
MN		 612 }
613 return 0;
614}
615
e5d40372
DB		 616static inline int get_lowest_part_list_y(H264Context *h, Picture *pic, int n,
617 int height, int y_offset, int list)
618{
619 int raw_my = h->mv_cache[list][scan8[n]][1];
fb845ffd
RB		 620 int filter_height_up = (raw_my & 3) ? 2 : 0;
621 int filter_height_down = (raw_my & 3) ? 3 : 0;
e5d40372 622 int full_my = (raw_my >> 2) + y_offset;
fb845ffd
RB		 623 int top = full_my - filter_height_up;
624 int bottom = full_my + filter_height_down + height;
6a9c8594
AS		 625
626 return FFMAX(abs(top), bottom);
627}
628
e5d40372
DB		 629static inline void get_lowest_part_y(H264Context *h, int refs[2][48], int n,
630 int height, int y_offset, int list0,
631 int list1, int *nrefs)
632{
6a9c8594
AS		 633 int my;
634
da6be8fc 635 y_offset += 16 * (h->mb_y >> MB_FIELD(h));
6a9c8594 636
e5d40372
DB		 637 if (list0) {
638 int ref_n = h->ref_cache[0][scan8[n]];
639 Picture *ref = &h->ref_list[0][ref_n];
6a9c8594
AS		 640
641 // Error resilience puts the current picture in the ref list.
642 // Don't try to wait on these as it will cause a deadlock.
643 // Fields can wait on each other, though.
759001c5
AK		 644 if (ref->tf.progress->data != h->cur_pic.tf.progress->data ||
645 (ref->reference & 3) != h->picture_structure) {
6a9c8594 646 my = get_lowest_part_list_y(h, ref, n, height, y_offset, 0);
e5d40372
DB		 647 if (refs[0][ref_n] < 0)
648 nrefs[0] += 1;
6a9c8594
AS		 649 refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
650 }
651 }
652
e5d40372
DB		 653 if (list1) {
654 int ref_n = h->ref_cache[1][scan8[n]];
655 Picture *ref = &h->ref_list[1][ref_n];
6a9c8594 656
759001c5
AK		 657 if (ref->tf.progress->data != h->cur_pic.tf.progress->data ||
658 (ref->reference & 3) != h->picture_structure) {
6a9c8594 659 my = get_lowest_part_list_y(h, ref, n, height, y_offset, 1);
e5d40372
DB		 660 if (refs[1][ref_n] < 0)
661 nrefs[1] += 1;
6a9c8594
AS		 662 refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
663 }
664 }
665}
666
667/**
668 * Wait until all reference frames are available for MC operations.
669 *
670 * @param h the H264 context
671 */
e5d40372
DB		 672static void await_references(H264Context *h)
673{
e5d40372 674 const int mb_xy = h->mb_xy;
759001c5 675 const int mb_type = h->cur_pic.mb_type[mb_xy];
6a9c8594 676 int refs[2][48];
e5d40372 677 int nrefs[2] = { 0 };
6a9c8594
AS		 678 int ref, list;
679
680 memset(refs, -1, sizeof(refs));
681
e5d40372 682 if (IS_16X16(mb_type)) {
6a9c8594 683 get_lowest_part_y(h, refs, 0, 16, 0,
e5d40372
DB		 684 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
685 } else if (IS_16X8(mb_type)) {
6a9c8594 686 get_lowest_part_y(h, refs, 0, 8, 0,
e5d40372 687 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
6a9c8594 688 get_lowest_part_y(h, refs, 8, 8, 8,
e5d40372
DB		 689 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
690 } else if (IS_8X16(mb_type)) {
6a9c8594 691 get_lowest_part_y(h, refs, 0, 16, 0,
e5d40372 692 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
6a9c8594 693 get_lowest_part_y(h, refs, 4, 16, 0,
e5d40372
DB		 694 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
695 } else {
6a9c8594
AS		 696 int i;
697
698 assert(IS_8X8(mb_type));
699
e5d40372
DB		 700 for (i = 0; i < 4; i++) {
701 const int sub_mb_type = h->sub_mb_type[i];
702 const int n = 4 * i;
703 int y_offset = (i & 2) << 2;
704
705 if (IS_SUB_8X8(sub_mb_type)) {
706 get_lowest_part_y(h, refs, n, 8, y_offset,
707 IS_DIR(sub_mb_type, 0, 0),
708 IS_DIR(sub_mb_type, 0, 1),
709 nrefs);
710 } else if (IS_SUB_8X4(sub_mb_type)) {
711 get_lowest_part_y(h, refs, n, 4, y_offset,
712 IS_DIR(sub_mb_type, 0, 0),
713 IS_DIR(sub_mb_type, 0, 1),
714 nrefs);
715 get_lowest_part_y(h, refs, n + 2, 4, y_offset + 4,
716 IS_DIR(sub_mb_type, 0, 0),
717 IS_DIR(sub_mb_type, 0, 1),
718 nrefs);
719 } else if (IS_SUB_4X8(sub_mb_type)) {
720 get_lowest_part_y(h, refs, n, 8, y_offset,
721 IS_DIR(sub_mb_type, 0, 0),
722 IS_DIR(sub_mb_type, 0, 1),
723 nrefs);
724 get_lowest_part_y(h, refs, n + 1, 8, y_offset,
725 IS_DIR(sub_mb_type, 0, 0),
726 IS_DIR(sub_mb_type, 0, 1),
727 nrefs);
728 } else {
6a9c8594
AS		 729 int j;
730 assert(IS_SUB_4X4(sub_mb_type));
e5d40372
DB		 731 for (j = 0; j < 4; j++) {
732 int sub_y_offset = y_offset + 2 * (j & 2);
733 get_lowest_part_y(h, refs, n + j, 4, sub_y_offset,
734 IS_DIR(sub_mb_type, 0, 0),
735 IS_DIR(sub_mb_type, 0, 1),
736 nrefs);
6a9c8594
AS		 737 }
738 }
739 }
740 }
741
e5d40372
DB		 742 for (list = h->list_count - 1; list >= 0; list--)
743 for (ref = 0; ref < 48 && nrefs[list]; ref++) {
6a9c8594 744 int row = refs[list][ref];
e5d40372
DB		 745 if (row >= 0) {
746 Picture *ref_pic = &h->ref_list[list][ref];
759001c5 747 int ref_field = ref_pic->reference - 1;
6a9c8594 748 int ref_field_picture = ref_pic->field_picture;
2c541554 749 int pic_height = 16 * h->mb_height >> ref_field_picture;
6a9c8594 750
82313eaa 751 row <<= MB_MBAFF(h);
6a9c8594
AS		 752 nrefs[list]--;
753
7fa00653 754 if (!FIELD_PICTURE(h) && ref_field_picture) { // frame referencing two fields
759001c5 755 ff_thread_await_progress(&ref_pic->tf,
e5d40372
DB		 756 FFMIN((row >> 1) - !(row & 1),
757 pic_height - 1),
758 1);
759001c5 759 ff_thread_await_progress(&ref_pic->tf,
e5d40372
DB		 760 FFMIN((row >> 1), pic_height - 1),
761 0);
7fa00653 762 } else if (FIELD_PICTURE(h) && !ref_field_picture) { // field referencing one field of a frame
759001c5 763 ff_thread_await_progress(&ref_pic->tf,
e5d40372
DB		 764 FFMIN(row * 2 + ref_field,
765 pic_height - 1),
766 0);
7fa00653 767 } else if (FIELD_PICTURE(h)) {
759001c5 768 ff_thread_await_progress(&ref_pic->tf,
e5d40372
DB		 769 FFMIN(row, pic_height - 1),
770 ref_field);
771 } else {
759001c5 772 ff_thread_await_progress(&ref_pic->tf,
e5d40372
DB		 773 FFMIN(row, pic_height - 1),
774 0);
6a9c8594
AS		 775 }
776 }
777 }
66c6b5e2 778}
66c6b5e2 779
e5d40372
DB		 780static av_always_inline void mc_dir_part(H264Context *h, Picture *pic,
781 int n, int square, int height,
782 int delta, int list,
783 uint8_t *dest_y, uint8_t *dest_cb,
784 uint8_t *dest_cr,
785 int src_x_offset, int src_y_offset,
786 qpel_mc_func *qpix_op,
787 h264_chroma_mc_func chroma_op,
788 int pixel_shift, int chroma_idc)
05fb63f5 789{
e5d40372
DB		 790 const int mx = h->mv_cache[list][scan8[n]][0] + src_x_offset * 8;
791 int my = h->mv_cache[list][scan8[n]][1] + src_y_offset * 8;
792 const int luma_xy = (mx & 3) + ((my & 3) << 2);
793 int offset = ((mx >> 2) << pixel_shift) + (my >> 2) * h->mb_linesize;
794 uint8_t *src_y = pic->f.data[0] + offset;
795 uint8_t *src_cb, *src_cr;
fae6fd5b
RB		 796 int extra_width = 0;
797 int extra_height = 0;
e5d40372
DB		 798 int emu = 0;
799 const int full_mx = mx >> 2;
800 const int full_my = my >> 2;
2c541554 801 const int pic_width = 16 * h->mb_width;
da6be8fc 802 const int pic_height = 16 * h->mb_height >> MB_FIELD(h);
229d263c 803 int ysh;
115329f1 804
e5d40372
DB		 805 if (mx & 7)
806 extra_width -= 3;
807 if (my & 7)
808 extra_height -= 3;
809
810 if (full_mx < 0 - extra_width ||
811 full_my < 0 - extra_height ||
812 full_mx + 16 /*FIXME*/ > pic_width + extra_width ||
813 full_my + 16 /*FIXME*/ > pic_height + extra_height) {
2c541554 814 h->vdsp.emulated_edge_mc(h->edge_emu_buffer,
8c53d39e
RB		 815 src_y - (2 << pixel_shift) - 2 * h->mb_linesize,
816 h->mb_linesize,
817 16 + 5, 16 + 5 /*FIXME*/, full_mx - 2,
818 full_my - 2, pic_width, pic_height);
2c541554 819 src_y = h->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
e5d40372
DB		 820 emu = 1;
821 }
822
823 qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); // FIXME try variable height perhaps?
824 if (!square)
5d18eaad 825 qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
115329f1 826
2c541554 827 if (CONFIG_GRAY && h->flags & CODEC_FLAG_GRAY)
e5d40372 828 return;
7b442ad9 829
e5d40372 830 if (chroma_idc == 3 /* yuv444 */) {
657ccb5a 831 src_cb = pic->f.data[1] + offset;
e5d40372 832 if (emu) {
2c541554 833 h->vdsp.emulated_edge_mc(h->edge_emu_buffer,
8c53d39e
RB		 834 src_cb - (2 << pixel_shift) - 2 * h->mb_linesize,
835 h->mb_linesize,
836 16 + 5, 16 + 5 /*FIXME*/,
837 full_mx - 2, full_my - 2,
838 pic_width, pic_height);
2c541554 839 src_cb = h->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
c90b9442 840 }
e5d40372
DB		 841 qpix_op[luma_xy](dest_cb, src_cb, h->mb_linesize); // FIXME try variable height perhaps?
842 if (!square)
c90b9442 843 qpix_op[luma_xy](dest_cb + delta, src_cb + delta, h->mb_linesize);
c90b9442 844
657ccb5a 845 src_cr = pic->f.data[2] + offset;
e5d40372 846 if (emu) {
2c541554 847 h->vdsp.emulated_edge_mc(h->edge_emu_buffer,
8c53d39e
RB		 848 src_cr - (2 << pixel_shift) - 2 * h->mb_linesize,
849 h->mb_linesize,
850 16 + 5, 16 + 5 /*FIXME*/,
851 full_mx - 2, full_my - 2,
852 pic_width, pic_height);
2c541554 853 src_cr = h->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
c90b9442 854 }
e5d40372
DB		 855 qpix_op[luma_xy](dest_cr, src_cr, h->mb_linesize); // FIXME try variable height perhaps?
856 if (!square)
c90b9442 857 qpix_op[luma_xy](dest_cr + delta, src_cr + delta, h->mb_linesize);
c90b9442
JGG		 858 return;
859 }
860
05fb63f5 861 ysh = 3 - (chroma_idc == 2 /* yuv422 */);
da6be8fc 862 if (chroma_idc == 1 /* yuv420 */ && MB_FIELD(h)) {
5d18eaad 863 // chroma offset when predicting from a field of opposite parity
759001c5 864 my += 2 * ((h->mb_y & 1) - (pic->reference - 1));
e5d40372 865 emu |= (my >> 3) < 0 || (my >> 3) + 8 >= (pic_height >> 1);
5d18eaad 866 }
229d263c 867
e5d40372
DB		 868 src_cb = pic->f.data[1] + ((mx >> 3) << pixel_shift) +
869 (my >> ysh) * h->mb_uvlinesize;
870 src_cr = pic->f.data[2] + ((mx >> 3) << pixel_shift) +
871 (my >> ysh) * h->mb_uvlinesize;
5d18eaad 872
e5d40372 873 if (emu) {
2c541554 874 h->vdsp.emulated_edge_mc(h->edge_emu_buffer, src_cb, h->mb_uvlinesize,
8c53d39e
RB		 875 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
876 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
2c541554 877 src_cb = h->edge_emu_buffer;
0da71265 878 }
e5d40372
DB		 879 chroma_op(dest_cb, src_cb, h->mb_uvlinesize,
880 height >> (chroma_idc == 1 /* yuv420 */),
881 mx & 7, (my << (chroma_idc == 2 /* yuv422 */)) & 7);
0da71265 882
e5d40372 883 if (emu) {
2c541554 884 h->vdsp.emulated_edge_mc(h->edge_emu_buffer, src_cr, h->mb_uvlinesize,
8c53d39e
RB		 885 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
886 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
2c541554 887 src_cr = h->edge_emu_buffer;
0da71265 888 }
05fb63f5 889 chroma_op(dest_cr, src_cr, h->mb_uvlinesize, height >> (chroma_idc == 1 /* yuv420 */),
e5d40372 890 mx & 7, (my << (chroma_idc == 2 /* yuv422 */)) & 7);
0da71265
MN		 891}
892
e5d40372
DB		 893static av_always_inline void mc_part_std(H264Context *h, int n, int square,
894 int height, int delta,
895 uint8_t *dest_y, uint8_t *dest_cb,
896 uint8_t *dest_cr,
897 int x_offset, int y_offset,
898 qpel_mc_func *qpix_put,
899 h264_chroma_mc_func chroma_put,
900 qpel_mc_func *qpix_avg,
901 h264_chroma_mc_func chroma_avg,
902 int list0, int list1,
903 int pixel_shift, int chroma_idc)
05fb63f5 904{
e5d40372
DB		 905 qpel_mc_func *qpix_op = qpix_put;
906 h264_chroma_mc_func chroma_op = chroma_put;
115329f1 907
e5d40372 908 dest_y += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
05fb63f5 909 if (chroma_idc == 3 /* yuv444 */) {
e5d40372
DB		 910 dest_cb += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
911 dest_cr += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
05fb63f5 912 } else if (chroma_idc == 2 /* yuv422 */) {
e5d40372
DB		 913 dest_cb += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
914 dest_cr += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
915 } else { /* yuv420 */
916 dest_cb += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
917 dest_cr += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
c90b9442 918 }
2c541554 919 x_offset += 8 * h->mb_x;
da6be8fc 920 y_offset += 8 * (h->mb_y >> MB_FIELD(h));
115329f1 921
e5d40372
DB		 922 if (list0) {
923 Picture *ref = &h->ref_list[0][h->ref_cache[0][scan8[n]]];
c2d33742 924 mc_dir_part(h, ref, n, square, height, delta, 0,
e5d40372
DB		 925 dest_y, dest_cb, dest_cr, x_offset, y_offset,
926 qpix_op, chroma_op, pixel_shift, chroma_idc);
0da71265 927
e5d40372
DB		 928 qpix_op = qpix_avg;
929 chroma_op = chroma_avg;
0da71265
MN		 930 }
931
e5d40372
DB		 932 if (list1) {
933 Picture *ref = &h->ref_list[1][h->ref_cache[1][scan8[n]]];
c2d33742 934 mc_dir_part(h, ref, n, square, height, delta, 1,
e5d40372
DB		 935 dest_y, dest_cb, dest_cr, x_offset, y_offset,
936 qpix_op, chroma_op, pixel_shift, chroma_idc);
0da71265
MN		 937 }
938}
939
e5d40372
DB		 940static av_always_inline void mc_part_weighted(H264Context *h, int n, int square,
941 int height, int delta,
942 uint8_t *dest_y, uint8_t *dest_cb,
943 uint8_t *dest_cr,
944 int x_offset, int y_offset,
945 qpel_mc_func *qpix_put,
946 h264_chroma_mc_func chroma_put,
947 h264_weight_func luma_weight_op,
948 h264_weight_func chroma_weight_op,
949 h264_biweight_func luma_weight_avg,
950 h264_biweight_func chroma_weight_avg,
951 int list0, int list1,
952 int pixel_shift, int chroma_idc)
953{
c2d33742 954 int chroma_height;
9f2d1b4f 955
e5d40372 956 dest_y += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
05fb63f5 957 if (chroma_idc == 3 /* yuv444 */) {
e5d40372 958 chroma_height = height;
c90b9442 959 chroma_weight_avg = luma_weight_avg;
e5d40372
DB		 960 chroma_weight_op = luma_weight_op;
961 dest_cb += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
962 dest_cr += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
05fb63f5 963 } else if (chroma_idc == 2 /* yuv422 */) {
c2d33742 964 chroma_height = height;
e5d40372
DB		 965 dest_cb += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
966 dest_cr += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
967 } else { /* yuv420 */
c2d33742 968 chroma_height = height >> 1;
e5d40372
DB		 969 dest_cb += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
970 dest_cr += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
c90b9442 971 }
2c541554 972 x_offset += 8 * h->mb_x;
da6be8fc 973 y_offset += 8 * (h->mb_y >> MB_FIELD(h));
115329f1 974
e5d40372 975 if (list0 && list1) {
9f2d1b4f
LM		 976 /* don't optimize for luma-only case, since B-frames usually
977 * use implicit weights => chroma too. */
a394959b
JG		 978 uint8_t *tmp_cb = h->bipred_scratchpad;
979 uint8_t *tmp_cr = h->bipred_scratchpad + (16 << pixel_shift);
980 uint8_t *tmp_y = h->bipred_scratchpad + 16 * h->mb_uvlinesize;
e5d40372
DB		 981 int refn0 = h->ref_cache[0][scan8[n]];
982 int refn1 = h->ref_cache[1][scan8[n]];
9f2d1b4f 983
c2d33742 984 mc_dir_part(h, &h->ref_list[0][refn0], n, square, height, delta, 0,
9f2d1b4f 985 dest_y, dest_cb, dest_cr,
05fb63f5
RB		 986 x_offset, y_offset, qpix_put, chroma_put,
987 pixel_shift, chroma_idc);
c2d33742 988 mc_dir_part(h, &h->ref_list[1][refn1], n, square, height, delta, 1,
9f2d1b4f 989 tmp_y, tmp_cb, tmp_cr,
05fb63f5
RB		 990 x_offset, y_offset, qpix_put, chroma_put,
991 pixel_shift, chroma_idc);
9f2d1b4f 992
e5d40372 993 if (h->use_weight == 2) {
2c541554 994 int weight0 = h->implicit_weight[refn0][refn1][h->mb_y & 1];
9f2d1b4f 995 int weight1 = 64 - weight0;
e5d40372
DB		 996 luma_weight_avg(dest_y, tmp_y, h->mb_linesize,
997 height, 5, weight0, weight1, 0);
c2d33742
RB		 998 chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize,
999 chroma_height, 5, weight0, weight1, 0);
1000 chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize,
1001 chroma_height, 5, weight0, weight1, 0);
e5d40372
DB		 1002 } else {
1003 luma_weight_avg(dest_y, tmp_y, h->mb_linesize, height,
1004 h->luma_log2_weight_denom,
1005 h->luma_weight[refn0][0][0],
1006 h->luma_weight[refn1][1][0],
1007 h->luma_weight[refn0][0][1] +
1008 h->luma_weight[refn1][1][1]);
1009 chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, chroma_height,
1010 h->chroma_log2_weight_denom,
1011 h->chroma_weight[refn0][0][0][0],
1012 h->chroma_weight[refn1][1][0][0],
1013 h->chroma_weight[refn0][0][0][1] +
1014 h->chroma_weight[refn1][1][0][1]);
1015 chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, chroma_height,
1016 h->chroma_log2_weight_denom,
1017 h->chroma_weight[refn0][0][1][0],
1018 h->chroma_weight[refn1][1][1][0],
1019 h->chroma_weight[refn0][0][1][1] +
1020 h->chroma_weight[refn1][1][1][1]);
9f2d1b4f 1021 }
e5d40372
DB		 1022 } else {
1023 int list = list1 ? 1 : 0;
1024 int refn = h->ref_cache[list][scan8[n]];
1025 Picture *ref = &h->ref_list[list][refn];
c2d33742 1026 mc_dir_part(h, ref, n, square, height, delta, list,
9f2d1b4f 1027 dest_y, dest_cb, dest_cr, x_offset, y_offset,
05fb63f5 1028 qpix_put, chroma_put, pixel_shift, chroma_idc);
9f2d1b4f 1029
e5d40372
DB		 1030 luma_weight_op(dest_y, h->mb_linesize, height,
1031 h->luma_log2_weight_denom,
1032 h->luma_weight[refn][list][0],
1033 h->luma_weight[refn][list][1]);
1034 if (h->use_weight_chroma) {
1035 chroma_weight_op(dest_cb, h->mb_uvlinesize, chroma_height,
1036 h->chroma_log2_weight_denom,
1037 h->chroma_weight[refn][list][0][0],
1038 h->chroma_weight[refn][list][0][1]);
1039 chroma_weight_op(dest_cr, h->mb_uvlinesize, chroma_height,
1040 h->chroma_log2_weight_denom,
1041 h->chroma_weight[refn][list][1][0],
1042 h->chroma_weight[refn][list][1][1]);
9f2d1b4f
LM		 1043 }
1044 }
1045}
1046
e5d40372
DB		 1047static av_always_inline void prefetch_motion(H264Context *h, int list,
1048 int pixel_shift, int chroma_idc)
27209bb1 1049{
513fbd8e
LM		 1050 /* fetch pixels for estimated mv 4 macroblocks ahead
1051 * optimized for 64byte cache lines */
513fbd8e 1052 const int refn = h->ref_cache[list][scan8[0]];
e5d40372 1053 if (refn >= 0) {
2c541554
AK		 1054 const int mx = (h->mv_cache[list][scan8[0]][0] >> 2) + 16 * h->mb_x + 8;
1055 const int my = (h->mv_cache[list][scan8[0]][1] >> 2) + 16 * h->mb_y;
657ccb5a 1056 uint8_t **src = h->ref_list[list][refn].f.data;
e5d40372 1057 int off = (mx << pixel_shift) +
2c541554 1058 (my + (h->mb_x & 3) * 4) * h->mb_linesize +
e5d40372 1059 (64 << pixel_shift);
2c541554 1060 h->vdsp.prefetch(src[0] + off, h->linesize, 4);
05fb63f5 1061 if (chroma_idc == 3 /* yuv444 */) {
2c541554
AK		 1062 h->vdsp.prefetch(src[1] + off, h->linesize, 4);
1063 h->vdsp.prefetch(src[2] + off, h->linesize, 4);
e5d40372
DB		 1064 } else {
1065 off = ((mx >> 1) << pixel_shift) +
2c541554 1066 ((my >> 1) + (h->mb_x & 7)) * h->uvlinesize +
e5d40372 1067 (64 << pixel_shift);
2c541554 1068 h->vdsp.prefetch(src[1] + off, src[2] - src[1], 2);
c90b9442 1069 }
513fbd8e
LM		 1070 }
1071}
1072
e5d40372
DB		 1073static void free_tables(H264Context *h, int free_rbsp)
1074{
7978debd 1075 int i;
afebe2f7 1076 H264Context *hx;
6a9c8594 1077
0da71265 1078 av_freep(&h->intra4x4_pred_mode);
e5017ab8
LA		 1079 av_freep(&h->chroma_pred_mode_table);
1080 av_freep(&h->cbp_table);
9e528114
LA		 1081 av_freep(&h->mvd_table[0]);
1082 av_freep(&h->mvd_table[1]);
5ad984c9 1083 av_freep(&h->direct_table);
0da71265
MN		 1084 av_freep(&h->non_zero_count);
1085 av_freep(&h->slice_table_base);
e5d40372 1086 h->slice_table = NULL;
c988f975 1087 av_freep(&h->list_counts);
e5017ab8 1088
0da71265 1089 av_freep(&h->mb2b_xy);
d43c1922 1090 av_freep(&h->mb2br_xy);
9f2d1b4f 1091
759001c5
AK		 1092 av_buffer_pool_uninit(&h->qscale_table_pool);
1093 av_buffer_pool_uninit(&h->mb_type_pool);
1094 av_buffer_pool_uninit(&h->motion_val_pool);
1095 av_buffer_pool_uninit(&h->ref_index_pool);
1096
1097 if (free_rbsp && h->DPB) {
1098 for (i = 0; i < MAX_PICTURE_COUNT; i++)
1099 unref_picture(h, &h->DPB[i]);
2c541554 1100 av_freep(&h->DPB);
2c541554 1101 } else if (h->DPB) {
759001c5 1102 for (i = 0; i < MAX_PICTURE_COUNT; i++)
2c541554
AK		 1103 h->DPB[i].needs_realloc = 1;
1104 }
1105
1106 h->cur_pic_ptr = NULL;
1107
e5d40372 1108 for (i = 0; i < MAX_THREADS; i++) {
afebe2f7 1109 hx = h->thread_context[i];
e5d40372
DB		 1110 if (!hx)
1111 continue;
afebe2f7
AÖ		 1112 av_freep(&hx->top_borders[1]);
1113 av_freep(&hx->top_borders[0]);
a394959b 1114 av_freep(&hx->bipred_scratchpad);
2c541554
AK		 1115 av_freep(&hx->edge_emu_buffer);
1116 av_freep(&hx->dc_val_base);
1117 av_freep(&hx->me.scratchpad);
1118 av_freep(&hx->er.mb_index2xy);
1119 av_freep(&hx->er.error_status_table);
1120 av_freep(&hx->er.er_temp_buffer);
1121 av_freep(&hx->er.mbintra_table);
1122 av_freep(&hx->er.mbskip_table);
1123
e5d40372 1124 if (free_rbsp) {
fcb7e535
RB		 1125 av_freep(&hx->rbsp_buffer[1]);
1126 av_freep(&hx->rbsp_buffer[0]);
1127 hx->rbsp_buffer_size[0] = 0;
1128 hx->rbsp_buffer_size[1] = 0;
91078926 1129 }
e5d40372
DB		 1130 if (i)
1131 av_freep(&h->thread_context[i]);
afebe2f7 1132 }
0da71265
MN		 1133}
1134
e5d40372
DB		 1135static void init_dequant8_coeff_table(H264Context *h)
1136{
1137 int i, j, q, x;
1138 const int max_qp = 51 + 6 * (h->sps.bit_depth_luma - 8);
239ea04c 1139
e5d40372 1140 for (i = 0; i < 6; i++) {
c90b9442 1141 h->dequant8_coeff[i] = h->dequant8_buffer[i];
e5d40372
DB		 1142 for (j = 0; j < i; j++)
1143 if (!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i],
1144 64 * sizeof(uint8_t))) {
c90b9442
JGG		 1145 h->dequant8_coeff[i] = h->dequant8_buffer[j];
1146 break;
1147 }
e5d40372 1148 if (j < i)
c90b9442 1149 continue;
239ea04c 1150
e5d40372 1151 for (q = 0; q < max_qp + 1; q++) {
d9ec210b 1152 int shift = div6[q];
e5d40372
DB		 1153 int idx = rem6[q];
1154 for (x = 0; x < 64; x++)
1155 h->dequant8_coeff[i][q][(x >> 3) | ((x & 7) << 3)] =
1156 ((uint32_t)dequant8_coeff_init[idx][dequant8_coeff_init_scan[((x >> 1) & 12) | (x & 3)]] *
1157 h->pps.scaling_matrix8[i][x]) << shift;
239ea04c
LM		 1158 }
1159 }
1160}
1161
e5d40372
DB		 1162static void init_dequant4_coeff_table(H264Context *h)
1163{
1164 int i, j, q, x;
1165 const int max_qp = 51 + 6 * (h->sps.bit_depth_luma - 8);
1166 for (i = 0; i < 6; i++) {
239ea04c 1167 h->dequant4_coeff[i] = h->dequant4_buffer[i];
e5d40372
DB		 1168 for (j = 0; j < i; j++)
1169 if (!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i],
1170 16 * sizeof(uint8_t))) {
239ea04c
LM		 1171 h->dequant4_coeff[i] = h->dequant4_buffer[j];
1172 break;
1173 }
e5d40372 1174 if (j < i)
239ea04c
LM		 1175 continue;
1176
e5d40372 1177 for (q = 0; q < max_qp + 1; q++) {
d9ec210b 1178 int shift = div6[q] + 2;
e5d40372
DB		 1179 int idx = rem6[q];
1180 for (x = 0; x < 16; x++)
1181 h->dequant4_coeff[i][q][(x >> 2) | ((x << 2) & 0xF)] =
1182 ((uint32_t)dequant4_coeff_init[idx][(x & 1) + ((x >> 2) & 1)] *
1183 h->pps.scaling_matrix4[i][x]) << shift;
239ea04c
LM		 1184 }
1185 }
1186}
1187
e5d40372
DB		 1188static void init_dequant_tables(H264Context *h)
1189{
1190 int i, x;
239ea04c 1191 init_dequant4_coeff_table(h);
e5d40372 1192 if (h->pps.transform_8x8_mode)
239ea04c 1193 init_dequant8_coeff_table(h);
e5d40372
DB		 1194 if (h->sps.transform_bypass) {
1195 for (i = 0; i < 6; i++)
1196 for (x = 0; x < 16; x++)
1197 h->dequant4_coeff[i][0][x] = 1 << 6;
1198 if (h->pps.transform_8x8_mode)
1199 for (i = 0; i < 6; i++)
1200 for (x = 0; x < 64; x++)
1201 h->dequant8_coeff[i][0][x] = 1 << 6;
239ea04c
LM		 1202 }
1203}
1204
e5d40372
DB		 1205int ff_h264_alloc_tables(H264Context *h)
1206{
2c541554
AK		 1207 const int big_mb_num = h->mb_stride * (h->mb_height + 1);
1208 const int row_mb_num = h->mb_stride * 2 * h->avctx->thread_count;
1209 int x, y, i;
e5d40372 1210
2c541554 1211 FF_ALLOCZ_OR_GOTO(h->avctx, h->intra4x4_pred_mode,
e5d40372 1212 row_mb_num * 8 * sizeof(uint8_t), fail)
2c541554 1213 FF_ALLOCZ_OR_GOTO(h->avctx, h->non_zero_count,
e5d40372 1214 big_mb_num * 48 * sizeof(uint8_t), fail)
2c541554
AK		 1215 FF_ALLOCZ_OR_GOTO(h->avctx, h->slice_table_base,
1216 (big_mb_num + h->mb_stride) * sizeof(*h->slice_table_base), fail)
1217 FF_ALLOCZ_OR_GOTO(h->avctx, h->cbp_table,
e5d40372 1218 big_mb_num * sizeof(uint16_t), fail)
2c541554 1219 FF_ALLOCZ_OR_GOTO(h->avctx, h->chroma_pred_mode_table,
e5d40372 1220 big_mb_num * sizeof(uint8_t), fail)
2c541554 1221 FF_ALLOCZ_OR_GOTO(h->avctx, h->mvd_table[0],
e5d40372 1222 16 * row_mb_num * sizeof(uint8_t), fail);
2c541554 1223 FF_ALLOCZ_OR_GOTO(h->avctx, h->mvd_table[1],
e5d40372 1224 16 * row_mb_num * sizeof(uint8_t), fail);
2c541554 1225 FF_ALLOCZ_OR_GOTO(h->avctx, h->direct_table,
e5d40372 1226 4 * big_mb_num * sizeof(uint8_t), fail);
2c541554 1227 FF_ALLOCZ_OR_GOTO(h->avctx, h->list_counts,
e5d40372
DB		 1228 big_mb_num * sizeof(uint8_t), fail)
1229
1230 memset(h->slice_table_base, -1,
2c541554
AK		 1231 (big_mb_num + h->mb_stride) * sizeof(*h->slice_table_base));
1232 h->slice_table = h->slice_table_base + h->mb_stride * 2 + 1;
e5d40372 1233
2c541554 1234 FF_ALLOCZ_OR_GOTO(h->avctx, h->mb2b_xy,
e5d40372 1235 big_mb_num * sizeof(uint32_t), fail);
2c541554 1236 FF_ALLOCZ_OR_GOTO(h->avctx, h->mb2br_xy,
e5d40372 1237 big_mb_num * sizeof(uint32_t), fail);
2c541554
AK		 1238 for (y = 0; y < h->mb_height; y++)
1239 for (x = 0; x < h->mb_width; x++) {
1240 const int mb_xy = x + y * h->mb_stride;
e5d40372
DB		 1241 const int b_xy = 4 * x + 4 * y * h->b_stride;
1242
1243 h->mb2b_xy[mb_xy] = b_xy;
2c541554 1244 h->mb2br_xy[mb_xy] = 8 * (FMO ? mb_xy : (mb_xy % (2 * h->mb_stride)));
0da71265 1245 }
9f2d1b4f 1246
e5d40372 1247 if (!h->dequant4_coeff[0])
56edbd81
LM		 1248 init_dequant_tables(h);
1249
2c541554 1250 if (!h->DPB) {
759001c5 1251 h->DPB = av_mallocz_array(MAX_PICTURE_COUNT, sizeof(*h->DPB));
2c541554
AK		 1252 if (!h->DPB)
1253 return AVERROR(ENOMEM);
759001c5 1254 for (i = 0; i < MAX_PICTURE_COUNT; i++)
2c541554
AK		 1255 avcodec_get_frame_defaults(&h->DPB[i].f);
1256 avcodec_get_frame_defaults(&h->cur_pic.f);
1257 }
1258
0da71265 1259 return 0;
e5d40372 1260
0da71265 1261fail:
91078926 1262 free_tables(h, 1);
0da71265
MN		 1263 return -1;
1264}
1265
afebe2f7
AÖ		 1266/**
1267 * Mimic alloc_tables(), but for every context thread.
1268 */
e5d40372
DB		 1269static void clone_tables(H264Context *dst, H264Context *src, int i)
1270{
2c541554 1271 dst->intra4x4_pred_mode = src->intra4x4_pred_mode + i * 8 * 2 * src->mb_stride;
e5d40372
DB		 1272 dst->non_zero_count = src->non_zero_count;
1273 dst->slice_table = src->slice_table;
1274 dst->cbp_table = src->cbp_table;
1275 dst->mb2b_xy = src->mb2b_xy;
1276 dst->mb2br_xy = src->mb2br_xy;
1277 dst->chroma_pred_mode_table = src->chroma_pred_mode_table;
2c541554
AK		 1278 dst->mvd_table[0] = src->mvd_table[0] + i * 8 * 2 * src->mb_stride;
1279 dst->mvd_table[1] = src->mvd_table[1] + i * 8 * 2 * src->mb_stride;
e5d40372
DB		 1280 dst->direct_table = src->direct_table;
1281 dst->list_counts = src->list_counts;
2c541554
AK		 1282 dst->DPB = src->DPB;
1283 dst->cur_pic_ptr = src->cur_pic_ptr;
1284 dst->cur_pic = src->cur_pic;
a394959b 1285 dst->bipred_scratchpad = NULL;
2c541554
AK		 1286 dst->edge_emu_buffer = NULL;
1287 dst->me.scratchpad = NULL;
1288 ff_h264_pred_init(&dst->hpc, src->avctx->codec_id, src->sps.bit_depth_luma,
e5d40372 1289 src->sps.chroma_format_idc);
afebe2f7
AÖ		 1290}
1291
1292/**
1293 * Init context
1294 * Allocate buffers which are not shared amongst multiple threads.
1295 */
e5d40372
DB		 1296static int context_init(H264Context *h)
1297{
2c541554
AK		 1298 ERContext *er = &h->er;
1299 int mb_array_size = h->mb_height * h->mb_stride;
1300 int y_size = (2 * h->mb_width + 1) * (2 * h->mb_height + 1);
1301 int c_size = h->mb_stride * (h->mb_height + 1);
1302 int yc_size = y_size + 2 * c_size;
1303 int x, y, i;
1304
1305 FF_ALLOCZ_OR_GOTO(h->avctx, h->top_borders[0],
1306 h->mb_width * 16 * 3 * sizeof(uint8_t) * 2, fail)
1307 FF_ALLOCZ_OR_GOTO(h->avctx, h->top_borders[1],
1308 h->mb_width * 16 * 3 * sizeof(uint8_t) * 2, fail)
e5d40372
DB		 1309
1310 h->ref_cache[0][scan8[5] + 1] =
1311 h->ref_cache[0][scan8[7] + 1] =
1312 h->ref_cache[0][scan8[13] + 1] =
1313 h->ref_cache[1][scan8[5] + 1] =
1314 h->ref_cache[1][scan8[7] + 1] =
1315 h->ref_cache[1][scan8[13] + 1] = PART_NOT_AVAILABLE;
145061a1 1316
2c541554
AK		 1317 /* init ER */
1318 er->avctx = h->avctx;
1319 er->dsp = &h->dsp;
1320 er->decode_mb = h264_er_decode_mb;
1321 er->opaque = h;
1322 er->quarter_sample = 1;
1323
1324 er->mb_num = h->mb_num;
1325 er->mb_width = h->mb_width;
1326 er->mb_height = h->mb_height;
1327 er->mb_stride = h->mb_stride;
1328 er->b8_stride = h->mb_width * 2 + 1;
1329
1330 FF_ALLOCZ_OR_GOTO(h->avctx, er->mb_index2xy, (h->mb_num + 1) * sizeof(int),
1331 fail); // error ressilience code looks cleaner with this
1332 for (y = 0; y < h->mb_height; y++)
1333 for (x = 0; x < h->mb_width; x++)
1334 er->mb_index2xy[x + y * h->mb_width] = x + y * h->mb_stride;
1335
1336 er->mb_index2xy[h->mb_height * h->mb_width] = (h->mb_height - 1) *
1337 h->mb_stride + h->mb_width;
1338
1339 FF_ALLOCZ_OR_GOTO(h->avctx, er->error_status_table,
1340 mb_array_size * sizeof(uint8_t), fail);
1341
1342 FF_ALLOC_OR_GOTO(h->avctx, er->mbintra_table, mb_array_size, fail);
1343 memset(er->mbintra_table, 1, mb_array_size);
1344
1345 FF_ALLOCZ_OR_GOTO(h->avctx, er->mbskip_table, mb_array_size + 2, fail);
1346
1347 FF_ALLOC_OR_GOTO(h->avctx, er->er_temp_buffer, h->mb_height * h->mb_stride,
1348 fail);
1349
1350 FF_ALLOCZ_OR_GOTO(h->avctx, h->dc_val_base, yc_size * sizeof(int16_t), fail);
1351 er->dc_val[0] = h->dc_val_base + h->mb_width * 2 + 2;
1352 er->dc_val[1] = h->dc_val_base + y_size + h->mb_stride + 1;
1353 er->dc_val[2] = er->dc_val[1] + c_size;
1354 for (i = 0; i < yc_size; i++)
1355 h->dc_val_base[i] = 1024;
54974c62 1356
afebe2f7 1357 return 0;
e5d40372 1358
afebe2f7
AÖ		 1359fail:
1360 return -1; // free_tables will clean up for us
1361}
1362
61c6eef5
JG		 1363static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size,
1364 int parse_extradata);
9855b2e3 1365
05e95319
HC		 1366int ff_h264_decode_extradata(H264Context *h)
1367{
2c541554 1368 AVCodecContext *avctx = h->avctx;
9855b2e3 1369
e5d40372 1370 if (avctx->extradata[0] == 1) {
9855b2e3
MN		 1371 int i, cnt, nalsize;
1372 unsigned char *p = avctx->extradata;
1373
1374 h->is_avc = 1;
1375
e5d40372 1376 if (avctx->extradata_size < 7) {
9855b2e3
MN		 1377 av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
1378 return -1;
1379 }
1380 /* sps and pps in the avcC always have length coded with 2 bytes,
e5d40372 1381 * so put a fake nal_length_size = 2 while parsing them */
9855b2e3
MN		 1382 h->nal_length_size = 2;
1383 // Decode sps from avcC
e5d40372
DB		 1384 cnt = *(p + 5) & 0x1f; // Number of sps
1385 p += 6;
9855b2e3
MN		 1386 for (i = 0; i < cnt; i++) {
1387 nalsize = AV_RB16(p) + 2;
d1186ff7
LA		 1388 if (p - avctx->extradata + nalsize > avctx->extradata_size)
1389 return -1;
61c6eef5 1390 if (decode_nal_units(h, p, nalsize, 1) < 0) {
e5d40372
DB		 1391 av_log(avctx, AV_LOG_ERROR,
1392 "Decoding sps %d from avcC failed\n", i);
9855b2e3
MN		 1393 return -1;
1394 }
1395 p += nalsize;
1396 }
1397 // Decode pps from avcC
1398 cnt = *(p++); // Number of pps
1399 for (i = 0; i < cnt; i++) {
1400 nalsize = AV_RB16(p) + 2;
d1186ff7
LA		 1401 if (p - avctx->extradata + nalsize > avctx->extradata_size)
1402 return -1;
61c6eef5 1403 if (decode_nal_units(h, p, nalsize, 1) < 0) {
e5d40372
DB		 1404 av_log(avctx, AV_LOG_ERROR,
1405 "Decoding pps %d from avcC failed\n", i);
9855b2e3
MN		 1406 return -1;
1407 }
1408 p += nalsize;
1409 }
e5d40372 1410 // Now store right nal length size, that will be used to parse all other nals
0c17beba 1411 h->nal_length_size = (avctx->extradata[4] & 0x03) + 1;
9855b2e3
MN		 1412 } else {
1413 h->is_avc = 0;
61c6eef5 1414 if (decode_nal_units(h, avctx->extradata, avctx->extradata_size, 1) < 0)
9855b2e3
MN		 1415 return -1;
1416 }
05e95319
HC		 1417 return 0;
1418}
1419
e5d40372
DB		 1420av_cold int ff_h264_decode_init(AVCodecContext *avctx)
1421{
1422 H264Context *h = avctx->priv_data;
ea2bb12e 1423 int i;
05e95319 1424
2c541554 1425 h->avctx = avctx;
48d0fd2d
AK		 1426
1427 h->width = h->avctx->width;
1428 h->height = h->avctx->height;
1429
1430 h->bit_depth_luma = 8;
1431 h->chroma_format_idc = 1;
1432
1433 ff_h264dsp_init(&h->h264dsp, 8, 1);
1434 ff_h264chroma_init(&h->h264chroma, h->sps.bit_depth_chroma);
1435 ff_h264qpel_init(&h->h264qpel, 8);
1436 ff_h264_pred_init(&h->hpc, h->avctx->codec_id, 8, 1);
1437
1438 h->dequant_coeff_pps = -1;
1439
1440 /* needed so that IDCT permutation is known early */
1441 ff_dsputil_init(&h->dsp, h->avctx);
1442 ff_videodsp_init(&h->vdsp, 8);
1443
1444 memset(h->pps.scaling_matrix4, 16, 6 * 16 * sizeof(uint8_t));
1445 memset(h->pps.scaling_matrix8, 16, 2 * 64 * sizeof(uint8_t));
05e95319 1446
2c541554 1447 h->picture_structure = PICT_FRAME;
2c541554
AK		 1448 h->slice_context_count = 1;
1449 h->workaround_bugs = avctx->workaround_bugs;
1450 h->flags = avctx->flags;
05e95319 1451
e5d40372
DB		 1452 /* set defaults */
1453 // s->decode_mb = ff_h263_decode_mb;
e5d40372 1454 if (!avctx->has_b_frames)
2c541554 1455 h->low_delay = 1;
05e95319
HC		 1456
1457 avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1458
1459 ff_h264_decode_init_vlc();
1460
6e3ef511 1461 h->pixel_shift = 0;
19a0729b 1462 h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
6e3ef511 1463
05e95319 1464 h->thread_context[0] = h;
e5d40372 1465 h->outputed_poc = h->next_outputed_poc = INT_MIN;
ea2bb12e
RB		 1466 for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
1467 h->last_pocs[i] = INT_MIN;
e5d40372
DB		 1468 h->prev_poc_msb = 1 << 16;
1469 h->x264_build = -1;
05e95319 1470 ff_h264_reset_sei(h);
36ef5369 1471 if (avctx->codec_id == AV_CODEC_ID_H264) {
e5d40372 1472 if (avctx->ticks_per_frame == 1)
2c541554 1473 h->avctx->time_base.den *= 2;
05e95319
HC		 1474 avctx->ticks_per_frame = 2;
1475 }
1476
e5d40372 1477 if (avctx->extradata_size > 0 && avctx->extradata &&
05e95319
HC		 1478 ff_h264_decode_extradata(h))
1479 return -1;
1480
e5d40372 1481 if (h->sps.bitstream_restriction_flag &&
2c541554
AK		 1482 h->avctx->has_b_frames < h->sps.num_reorder_frames) {
1483 h->avctx->has_b_frames = h->sps.num_reorder_frames;
1484 h->low_delay = 0;
db8cb47d 1485 }
9855b2e3 1486
759001c5
AK		 1487 avctx->internal->allocate_progress = 1;
1488
0da71265
MN		 1489 return 0;
1490}
1491
e5d40372 1492#define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b) + (size))))
2c541554
AK		 1493#undef REBASE_PICTURE
1494#define REBASE_PICTURE(pic, new_ctx, old_ctx) \
1495 ((pic && pic >= old_ctx->DPB && \
759001c5 1496 pic < old_ctx->DPB + MAX_PICTURE_COUNT) ? \
2c541554 1497 &new_ctx->DPB[pic - old_ctx->DPB] : NULL)
e5d40372
DB		 1498
1499static void copy_picture_range(Picture **to, Picture **from, int count,
2c541554
AK		 1500 H264Context *new_base,
1501 H264Context *old_base)
6a9c8594
AS		 1502{
1503 int i;
1504
e5d40372 1505 for (i = 0; i < count; i++) {
6a9c8594 1506 assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
2c541554 1507 IN_RANGE(from[i], old_base->DPB,
759001c5 1508 sizeof(Picture) * MAX_PICTURE_COUNT) ||
6a9c8594
AS		 1509 !from[i]));
1510 to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1511 }
1512}
1513
1514static void copy_parameter_set(void **to, void **from, int count, int size)
1515{
1516 int i;
1517
e5d40372
DB		 1518 for (i = 0; i < count; i++) {
1519 if (to[i] && !from[i])
1520 av_freep(&to[i]);
1521 else if (from[i] && !to[i])
1522 to[i] = av_malloc(size);
6a9c8594 1523
e5d40372
DB		 1524 if (from[i])
1525 memcpy(to[i], from[i], size);
6a9c8594
AS		 1526 }
1527}
1528
e5d40372
DB		 1529static int decode_init_thread_copy(AVCodecContext *avctx)
1530{
1531 H264Context *h = avctx->priv_data;
6a9c8594 1532
f3a29b75
JR		 1533 if (!avctx->internal->is_copy)
1534 return 0;
6a9c8594
AS		 1535 memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1536 memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1537
2c541554 1538 h->context_initialized = 0;
0eae920c 1539
6a9c8594
AS		 1540 return 0;
1541}
1542
e5d40372
DB		 1543#define copy_fields(to, from, start_field, end_field) \
1544 memcpy(&to->start_field, &from->start_field, \
1545 (char *)&to->end_field - (char *)&to->start_field)
1546
9e696d2e
JG		 1547static int h264_slice_header_init(H264Context *, int);
1548
1549static int h264_set_parameter_from_sps(H264Context *h);
1550
e5d40372
DB		 1551static int decode_update_thread_context(AVCodecContext *dst,
1552 const AVCodecContext *src)
1553{
1554 H264Context *h = dst->priv_data, *h1 = src->priv_data;
2c541554
AK		 1555 int inited = h->context_initialized, err = 0;
1556 int context_reinitialized = 0;
759001c5 1557 int i, ret;
6a9c8594 1558
2c541554 1559 if (dst == src || !h1->context_initialized)
e5d40372 1560 return 0;
6a9c8594 1561
9e696d2e 1562 if (inited &&
2c541554
AK		 1563 (h->width != h1->width ||
1564 h->height != h1->height ||
1565 h->mb_width != h1->mb_width ||
1566 h->mb_height != h1->mb_height ||
9e696d2e
JG		 1567 h->sps.bit_depth_luma != h1->sps.bit_depth_luma ||
1568 h->sps.chroma_format_idc != h1->sps.chroma_format_idc ||
1569 h->sps.colorspace != h1->sps.colorspace)) {
1570
a157c7f2
JG		 1571 /* set bits_per_raw_sample to the previous value. the check for changed
1572 * bit depth in h264_set_parameter_from_sps() uses it and sets it to
1573 * the current value */
1574 h->avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
1575
9e696d2e
JG		 1576 av_freep(&h->bipred_scratchpad);
1577
2c541554
AK		 1578 h->width = h1->width;
1579 h->height = h1->height;
1580 h->mb_height = h1->mb_height;
1581 h->mb_width = h1->mb_width;
1582 h->mb_num = h1->mb_num;
1583 h->mb_stride = h1->mb_stride;
9e696d2e
JG		 1584 h->b_stride = h1->b_stride;
1585
1586 if ((err = h264_slice_header_init(h, 1)) < 0) {
2c541554 1587 av_log(h->avctx, AV_LOG_ERROR, "h264_slice_header_init() failed");
9e696d2e
JG		 1588 return err;
1589 }
2c541554 1590 context_reinitialized = 1;
9e696d2e 1591
2c541554 1592 /* update linesize on resize. The decoder doesn't
25408b2a 1593 * necessarily call h264_frame_start in the new thread */
2c541554
AK		 1594 h->linesize = h1->linesize;
1595 h->uvlinesize = h1->uvlinesize;
9e696d2e
JG		 1596
1597 /* copy block_offset since frame_start may not be called */
1598 memcpy(h->block_offset, h1->block_offset, sizeof(h->block_offset));
9e696d2e
JG		 1599 }
1600
e5d40372
DB		 1601 if (!inited) {
1602 for (i = 0; i < MAX_SPS_COUNT; i++)
6a9c8594
AS		 1603 av_freep(h->sps_buffers + i);
1604
e5d40372 1605 for (i = 0; i < MAX_PPS_COUNT; i++)
6a9c8594
AS		 1606 av_freep(h->pps_buffers + i);
1607
2c541554 1608 memcpy(h, h1, sizeof(*h1));
6a9c8594
AS		 1609 memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1610 memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
2c541554
AK		 1611 memset(&h->er, 0, sizeof(h->er));
1612 memset(&h->me, 0, sizeof(h->me));
1613 h->context_initialized = 0;
1614
759001c5
AK		 1615 memset(&h->cur_pic, 0, sizeof(h->cur_pic));
1616 avcodec_get_frame_defaults(&h->cur_pic.f);
1617 h->cur_pic.tf.f = &h->cur_pic.f;
2c541554
AK		 1618
1619 h->avctx = dst;
1620 h->DPB = NULL;
759001c5
AK		 1621 h->qscale_table_pool = NULL;
1622 h->mb_type_pool = NULL;
1623 h->ref_index_pool = NULL;
1624 h->motion_val_pool = NULL;
2c541554 1625
bac3ab13
DB		 1626 if (ff_h264_alloc_tables(h) < 0) {
1627 av_log(dst, AV_LOG_ERROR, "Could not allocate memory for h264\n");
1628 return AVERROR(ENOMEM);
1629 }
6a9c8594
AS		 1630 context_init(h);
1631
e5d40372
DB		 1632 for (i = 0; i < 2; i++) {
1633 h->rbsp_buffer[i] = NULL;
6a9c8594
AS		 1634 h->rbsp_buffer_size[i] = 0;
1635 }
a394959b 1636 h->bipred_scratchpad = NULL;
2c541554 1637 h->edge_emu_buffer = NULL;
6a9c8594
AS		 1638
1639 h->thread_context[0] = h;
1640
2c541554
AK		 1641 h->dsp.clear_blocks(h->mb);
1642 h->dsp.clear_blocks(h->mb + (24 * 16 << h->pixel_shift));
1643 h->context_initialized = 1;
6a9c8594
AS		 1644 }
1645
2c541554
AK		 1646 h->avctx->coded_height = h1->avctx->coded_height;
1647 h->avctx->coded_width = h1->avctx->coded_width;
1648 h->avctx->width = h1->avctx->width;
1649 h->avctx->height = h1->avctx->height;
1650 h->coded_picture_number = h1->coded_picture_number;
1651 h->first_field = h1->first_field;
1652 h->picture_structure = h1->picture_structure;
1653 h->qscale = h1->qscale;
1654 h->droppable = h1->droppable;
1655 h->data_partitioning = h1->data_partitioning;
1656 h->low_delay = h1->low_delay;
1657
759001c5
AK		 1658 for (i = 0; i < MAX_PICTURE_COUNT; i++) {
1659 unref_picture(h, &h->DPB[i]);
1660 if (h1->DPB[i].f.data[0] &&
1661 (ret = ref_picture(h, &h->DPB[i], &h1->DPB[i])) < 0)
1662 return ret;
1663 }
2c541554
AK		 1664
1665 h->cur_pic_ptr = REBASE_PICTURE(h1->cur_pic_ptr, h, h1);
759001c5
AK		 1666 unref_picture(h, &h->cur_pic);
1667 if ((ret = ref_picture(h, &h->cur_pic, &h1->cur_pic)) < 0)
1668 return ret;
2c541554
AK		 1669
1670 h->workaround_bugs = h1->workaround_bugs;
1671 h->low_delay = h1->low_delay;
1672 h->droppable = h1->droppable;
1673
a394959b
JG		 1674 /* frame_start may not be called for the next thread (if it's decoding
1675 * a bottom field) so this has to be allocated here */
2c541554
AK		 1676 err = alloc_scratch_buffers(h, h1->linesize);
1677 if (err < 0)
1678 return err;
a394959b 1679
e5d40372
DB		 1680 // extradata/NAL handling
1681 h->is_avc = h1->is_avc;
6a9c8594 1682
e5d40372
DB		 1683 // SPS/PPS
1684 copy_parameter_set((void **)h->sps_buffers, (void **)h1->sps_buffers,
1685 MAX_SPS_COUNT, sizeof(SPS));
1686 h->sps = h1->sps;
1687 copy_parameter_set((void **)h->pps_buffers, (void **)h1->pps_buffers,
1688 MAX_PPS_COUNT, sizeof(PPS));
1689 h->pps = h1->pps;
6a9c8594 1690
e5d40372
DB		 1691 // Dequantization matrices
1692 // FIXME these are big - can they be only copied when PPS changes?
6a9c8594
AS		 1693 copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1694
e5d40372
DB		 1695 for (i = 0; i < 6; i++)
1696 h->dequant4_coeff[i] = h->dequant4_buffer[0] +
1697 (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
6a9c8594 1698
e5d40372
DB		 1699 for (i = 0; i < 6; i++)
1700 h->dequant8_coeff[i] = h->dequant8_buffer[0] +
1701 (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
6a9c8594
AS		 1702
1703 h->dequant_coeff_pps = h1->dequant_coeff_pps;
1704
e5d40372 1705 // POC timing
6a9c8594
AS		 1706 copy_fields(h, h1, poc_lsb, redundant_pic_count);
1707
e5d40372 1708 // reference lists
6a9c8594
AS		 1709 copy_fields(h, h1, short_ref, cabac_init_idc);
1710
2c541554
AK		 1711 copy_picture_range(h->short_ref, h1->short_ref, 32, h, h1);
1712 copy_picture_range(h->long_ref, h1->long_ref, 32, h, h1);
e5d40372 1713 copy_picture_range(h->delayed_pic, h1->delayed_pic,
2c541554 1714 MAX_DELAYED_PIC_COUNT + 2, h, h1);
6a9c8594
AS		 1715
1716 h->last_slice_type = h1->last_slice_type;
1717
2c541554
AK		 1718 if (context_reinitialized)
1719 h264_set_parameter_from_sps(h);
1720
1721 if (!h->cur_pic_ptr)
e5d40372 1722 return 0;
6a9c8594 1723
2c541554 1724 if (!h->droppable) {
12fe7594 1725 err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
e5d40372
DB		 1726 h->prev_poc_msb = h->poc_msb;
1727 h->prev_poc_lsb = h->poc_lsb;
6a9c8594 1728 }
e5d40372
DB		 1729 h->prev_frame_num_offset = h->frame_num_offset;
1730 h->prev_frame_num = h->frame_num;
1731 h->outputed_poc = h->next_outputed_poc;
6a9c8594 1732
12fe7594 1733 return err;
6a9c8594
AS		 1734}
1735
25408b2a 1736static int h264_frame_start(H264Context *h)
e5d40372 1737{
2c541554
AK		 1738 Picture *pic;
1739 int i, ret;
6e3ef511 1740 const int pixel_shift = h->pixel_shift;
0da71265 1741
2c541554
AK		 1742 release_unused_pictures(h, 1);
1743 h->cur_pic_ptr = NULL;
1744
1745 i = find_unused_picture(h);
1746 if (i < 0) {
1747 av_log(h->avctx, AV_LOG_ERROR, "no frame buffer available\n");
1748 return i;
1749 }
1750 pic = &h->DPB[i];
1751
759001c5 1752 pic->reference = h->droppable ? 0 : h->picture_structure;
2c541554
AK		 1753 pic->f.coded_picture_number = h->coded_picture_number++;
1754 pic->field_picture = h->picture_structure != PICT_FRAME;
3a22d7fa 1755 /*
2c541554
AK		 1756 * Zero key_frame here; IDR markings per slice in frame or fields are ORed
1757 * in later.
3a22d7fa
JD		 1758 * See decode_nal_units().
1759 */
2c541554
AK		 1760 pic->f.key_frame = 0;
1761 pic->mmco_reset = 0;
1762
1763 if ((ret = alloc_picture(h, pic)) < 0)
1764 return ret;
1765
1766 h->cur_pic_ptr = pic;
759001c5
AK		 1767 unref_picture(h, &h->cur_pic);
1768 if ((ret = ref_picture(h, &h->cur_pic, h->cur_pic_ptr)) < 0)
1769 return ret;
2c541554
AK		 1770
1771 ff_er_frame_start(&h->er);
0da71265 1772
2c541554 1773 assert(h->linesize && h->uvlinesize);
0da71265 1774
e5d40372 1775 for (i = 0; i < 16; i++) {
2c541554
AK		 1776 h->block_offset[i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * h->linesize * ((scan8[i] - scan8[0]) >> 3);
1777 h->block_offset[48 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * h->linesize * ((scan8[i] - scan8[0]) >> 3);
0da71265 1778 }
e5d40372
DB		 1779 for (i = 0; i < 16; i++) {
1780 h->block_offset[16 + i] =
2c541554 1781 h->block_offset[32 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * h->uvlinesize * ((scan8[i] - scan8[0]) >> 3);
e5d40372 1782 h->block_offset[48 + 16 + i] =
2c541554 1783 h->block_offset[48 + 32 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * h->uvlinesize * ((scan8[i] - scan8[0]) >> 3);
0da71265
MN		 1784 }
1785
934b0821
LM		 1786 /* can't be in alloc_tables because linesize isn't known there.
1787 * FIXME: redo bipred weight to not require extra buffer? */
2c541554
AK		 1788 for (i = 0; i < h->slice_context_count; i++)
1789 if (h->thread_context[i]) {
1790 ret = alloc_scratch_buffers(h->thread_context[i], h->linesize);
1791 if (ret < 0)
1792 return ret;
1793 }
e5d40372
DB		 1794
1795 /* Some macroblocks can be accessed before they're available in case
1796 * of lost slices, MBAFF or threading. */
1797 memset(h->slice_table, -1,
2c541554 1798 (h->mb_height * h->mb_stride - 1) * sizeof(*h->slice_table));
e5d40372
DB		 1799
1800 // s->decode = (s->flags & CODEC_FLAG_PSNR) || !s->encoding ||
1801 // s->current_picture.f.reference /* || h->contains_intra */ || 1;
1802
1803 /* We mark the current picture as non-reference after allocating it, so
1804 * that if we break out due to an error it can be released automatically
1805 * in the next ff_MPV_frame_start().
f08fefc4
AK		 1806 */
1807 h->cur_pic_ptr->reference = 0;
357282c6 1808
2c541554 1809 h->cur_pic_ptr->field_poc[0] = h->cur_pic_ptr->field_poc[1] = INT_MAX;
6a9c8594
AS		 1810
1811 h->next_output_pic = NULL;
1812
2c541554 1813 assert(h->cur_pic_ptr->long_ref == 0);
357282c6 1814
af8aa846 1815 return 0;
0da71265
MN		 1816}
1817
6a9c8594 1818/**
e5d40372
DB		 1819 * Run setup operations that must be run after slice header decoding.
1820 * This includes finding the next displayed frame.
1821 *
1822 * @param h h264 master context
1823 * @param setup_finished enough NALs have been read that we can call
1824 * ff_thread_finish_setup()
1825 */
1826static void decode_postinit(H264Context *h, int setup_finished)
1827{
2c541554
AK		 1828 Picture *out = h->cur_pic_ptr;
1829 Picture *cur = h->cur_pic_ptr;
6a9c8594 1830 int i, pics, out_of_order, out_idx;
adedd840 1831 int invalid = 0, cnt = 0;
6a9c8594 1832
2c541554 1833 h->cur_pic_ptr->f.pict_type = h->pict_type;
6a9c8594 1834
e5d40372
DB		 1835 if (h->next_output_pic)
1836 return;
6a9c8594 1837
e5d40372
DB		 1838 if (cur->field_poc[0] == INT_MAX || cur->field_poc[1] == INT_MAX) {
1839 /* FIXME: if we have two PAFF fields in one packet, we can't start
1840 * the next thread here. If we have one field per packet, we can.
1841 * The check in decode_nal_units() is not good enough to find this
1842 * yet, so we assume the worst for now. */
1843 // if (setup_finished)
2c541554 1844 // ff_thread_finish_setup(h->avctx);
6a9c8594
AS		 1845 return;
1846 }
1847
657ccb5a
DB		 1848 cur->f.interlaced_frame = 0;
1849 cur->f.repeat_pict = 0;
6a9c8594
AS		 1850
1851 /* Signal interlacing information externally. */
e5d40372
DB		 1852 /* Prioritize picture timing SEI information over used
1853 * decoding process if it exists. */
6a9c8594 1854
e5d40372
DB		 1855 if (h->sps.pic_struct_present_flag) {
1856 switch (h->sei_pic_struct) {
6a9c8594
AS		 1857 case SEI_PIC_STRUCT_FRAME:
1858 break;
1859 case SEI_PIC_STRUCT_TOP_FIELD:
1860 case SEI_PIC_STRUCT_BOTTOM_FIELD:
657ccb5a 1861 cur->f.interlaced_frame = 1;
6a9c8594
AS		 1862 break;
1863 case SEI_PIC_STRUCT_TOP_BOTTOM:
1864 case SEI_PIC_STRUCT_BOTTOM_TOP:
a6931d8e 1865 if (FIELD_OR_MBAFF_PICTURE(h))
657ccb5a 1866 cur->f.interlaced_frame = 1;
6a9c8594
AS		 1867 else
1868 // try to flag soft telecine progressive
657ccb5a 1869 cur->f.interlaced_frame = h->prev_interlaced_frame;
6a9c8594
AS		 1870 break;
1871 case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1872 case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
e5d40372
DB		 1873 /* Signal the possibility of telecined film externally
1874 * (pic_struct 5,6). From these hints, let the applications
1875 * decide if they apply deinterlacing. */
657ccb5a 1876 cur->f.repeat_pict = 1;
6a9c8594
AS		 1877 break;
1878 case SEI_PIC_STRUCT_FRAME_DOUBLING:
657ccb5a 1879 cur->f.repeat_pict = 2;
6a9c8594
AS		 1880 break;
1881 case SEI_PIC_STRUCT_FRAME_TRIPLING:
657ccb5a 1882 cur->f.repeat_pict = 4;
6a9c8594
AS		 1883 break;
1884 }
1885
e5d40372
DB		 1886 if ((h->sei_ct_type & 3) &&
1887 h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
657ccb5a 1888 cur->f.interlaced_frame = (h->sei_ct_type & (1 << 1)) != 0;
e5d40372 1889 } else {
6a9c8594 1890 /* Derive interlacing flag from used decoding process. */
a6931d8e 1891 cur->f.interlaced_frame = FIELD_OR_MBAFF_PICTURE(h);
6a9c8594 1892 }
657ccb5a 1893 h->prev_interlaced_frame = cur->f.interlaced_frame;
6a9c8594 1894
e5d40372 1895 if (cur->field_poc[0] != cur->field_poc[1]) {
6a9c8594 1896 /* Derive top_field_first from field pocs. */
657ccb5a 1897 cur->f.top_field_first = cur->field_poc[0] < cur->field_poc[1];
e5d40372 1898 } else {
657ccb5a 1899 if (cur->f.interlaced_frame || h->sps.pic_struct_present_flag) {
e5d40372
DB		 1900 /* Use picture timing SEI information. Even if it is a
1901 * information of a past frame, better than nothing. */
1902 if (h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM ||
1903 h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
657ccb5a 1904 cur->f.top_field_first = 1;
6a9c8594 1905 else
657ccb5a 1906 cur->f.top_field_first = 0;
e5d40372 1907 } else {
6a9c8594 1908 /* Most likely progressive */
657ccb5a 1909 cur->f.top_field_first = 0;
6a9c8594
AS		 1910 }
1911 }
1912
e5d40372 1913 // FIXME do something with unavailable reference frames
6a9c8594
AS		 1914
1915 /* Sort B-frames into display order */
1916
e5d40372 1917 if (h->sps.bitstream_restriction_flag &&
2c541554
AK		 1918 h->avctx->has_b_frames < h->sps.num_reorder_frames) {
1919 h->avctx->has_b_frames = h->sps.num_reorder_frames;
1920 h->low_delay = 0;
6a9c8594
AS		 1921 }
1922
2c541554 1923 if (h->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT &&
e5d40372 1924 !h->sps.bitstream_restriction_flag) {
2c541554
AK		 1925 h->avctx->has_b_frames = MAX_DELAYED_PIC_COUNT - 1;
1926 h->low_delay = 0;
6a9c8594
AS		 1927 }
1928
1929 pics = 0;
e5d40372
DB		 1930 while (h->delayed_pic[pics])
1931 pics++;
6a9c8594
AS		 1932
1933 assert(pics <= MAX_DELAYED_PIC_COUNT);
1934
1935 h->delayed_pic[pics++] = cur;
759001c5
AK		 1936 if (cur->reference == 0)
1937 cur->reference = DELAYED_PIC_REF;
6a9c8594 1938
adedd840
RB		 1939 /* Frame reordering. This code takes pictures from coding order and sorts
1940 * them by their incremental POC value into display order. It supports POC
1941 * gaps, MMCO reset codes and random resets.
1942 * A "display group" can start either with a IDR frame (f.key_frame = 1),
1943 * and/or can be closed down with a MMCO reset code. In sequences where
1944 * there is no delay, we can't detect that (since the frame was already
1945 * output to the user), so we also set h->mmco_reset to detect the MMCO
1946 * reset code.
2c541554 1947 * FIXME: if we detect insufficient delays (as per h->avctx->has_b_frames),
adedd840
RB		 1948 * we increase the delay between input and output. All frames affected by
1949 * the lag (e.g. those that should have been output before another frame
1950 * that we already returned to the user) will be dropped. This is a bug
1951 * that we will fix later. */
1952 for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++) {
1953 cnt += out->poc < h->last_pocs[i];
1954 invalid += out->poc == INT_MIN;
1955 }
e5d40372
DB		 1956 if (!h->mmco_reset && !cur->f.key_frame &&
1957 cnt + invalid == MAX_DELAYED_PIC_COUNT && cnt > 0) {
adedd840
RB		 1958 h->mmco_reset = 2;
1959 if (pics > 1)
1960 h->delayed_pic[pics - 2]->mmco_reset = 2;
1961 }
1962 if (h->mmco_reset || cur->f.key_frame) {
1963 for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
1964 h->last_pocs[i] = INT_MIN;
1965 cnt = 0;
1966 invalid = MAX_DELAYED_PIC_COUNT;
1967 }
e5d40372 1968 out = h->delayed_pic[0];
6a9c8594 1969 out_idx = 0;
e5d40372
DB		 1970 for (i = 1; i < MAX_DELAYED_PIC_COUNT &&
1971 h->delayed_pic[i] &&
1972 !h->delayed_pic[i - 1]->mmco_reset &&
1973 !h->delayed_pic[i]->f.key_frame;
1974 i++)
1975 if (h->delayed_pic[i]->poc < out->poc) {
1976 out = h->delayed_pic[i];
6a9c8594
AS		 1977 out_idx = i;
1978 }
2c541554 1979 if (h->avctx->has_b_frames == 0 &&
e5d40372 1980 (h->delayed_pic[0]->f.key_frame || h->mmco_reset))
adedd840 1981 h->next_outputed_poc = INT_MIN;
e5d40372
DB		 1982 out_of_order = !out->f.key_frame && !h->mmco_reset &&
1983 (out->poc < h->next_outputed_poc);
6a9c8594 1984
e5d40372 1985 if (h->sps.bitstream_restriction_flag &&
2c541554
AK		 1986 h->avctx->has_b_frames >= h->sps.num_reorder_frames) {
1987 } else if (out_of_order && pics - 1 == h->avctx->has_b_frames &&
1988 h->avctx->has_b_frames < MAX_DELAYED_PIC_COUNT) {
ea2bb12e 1989 if (invalid + cnt < MAX_DELAYED_PIC_COUNT) {
2c541554 1990 h->avctx->has_b_frames = FFMAX(h->avctx->has_b_frames, cnt);
ea2bb12e 1991 }
2c541554
AK		 1992 h->low_delay = 0;
1993 } else if (h->low_delay &&
e5d40372
DB		 1994 ((h->next_outputed_poc != INT_MIN &&
1995 out->poc > h->next_outputed_poc + 2) ||
ea2bb12e 1996 cur->f.pict_type == AV_PICTURE_TYPE_B)) {
2c541554
AK		 1997 h->low_delay = 0;
1998 h->avctx->has_b_frames++;
6a9c8594
AS		 1999 }
2000
2c541554 2001 if (pics > h->avctx->has_b_frames) {
759001c5 2002 out->reference &= ~DELAYED_PIC_REF;
e5d40372
DB		 2003 // for frame threading, the owner must be the second field's thread or
2004 // else the first thread can release the picture and reuse it unsafely
e5d40372
DB		 2005 for (i = out_idx; h->delayed_pic[i]; i++)
2006 h->delayed_pic[i] = h->delayed_pic[i + 1];
2007 }
2008 memmove(h->last_pocs, &h->last_pocs[1],
2009 sizeof(*h->last_pocs) * (MAX_DELAYED_PIC_COUNT - 1));
adedd840 2010 h->last_pocs[MAX_DELAYED_PIC_COUNT - 1] = cur->poc;
2c541554 2011 if (!out_of_order && pics > h->avctx->has_b_frames) {
6a9c8594 2012 h->next_output_pic = out;
adedd840
RB		 2013 if (out->mmco_reset) {
2014 if (out_idx > 0) {
e5d40372 2015 h->next_outputed_poc = out->poc;
adedd840
RB		 2016 h->delayed_pic[out_idx - 1]->mmco_reset = out->mmco_reset;
2017 } else {
2018 h->next_outputed_poc = INT_MIN;
2019 }
2020 } else {
0b4c3232
RB		 2021 if (out_idx == 0 && pics > 1 && h->delayed_pic[0]->f.key_frame) {
2022 h->next_outputed_poc = INT_MIN;
2023 } else {
2024 h->next_outputed_poc = out->poc;
2025 }
adedd840
RB		 2026 }
2027 h->mmco_reset = 0;
e5d40372 2028 } else {
2c541554 2029 av_log(h->avctx, AV_LOG_DEBUG, "no picture\n");
6a9c8594
AS		 2030 }
2031
2032 if (setup_finished)
2c541554 2033 ff_thread_finish_setup(h->avctx);
6a9c8594
AS		 2034}
2035
76741b0e
BC		 2036static av_always_inline void backup_mb_border(H264Context *h, uint8_t *src_y,
2037 uint8_t *src_cb, uint8_t *src_cr,
e5d40372
DB		 2038 int linesize, int uvlinesize,
2039 int simple)
76741b0e 2040{
0b69d625 2041 uint8_t *top_border;
5f7f9719 2042 int top_idx = 1;
6e3ef511 2043 const int pixel_shift = h->pixel_shift;
23e85be5 2044 int chroma444 = CHROMA444(h);
e962bd08 2045 int chroma422 = CHROMA422(h);
115329f1 2046
e5d40372 2047 src_y -= linesize;
53c05b1e
MN		 2048 src_cb -= uvlinesize;
2049 src_cr -= uvlinesize;
2050
7bece9b2 2051 if (!simple && FRAME_MBAFF(h)) {
2c541554 2052 if (h->mb_y & 1) {
82313eaa 2053 if (!MB_MBAFF(h)) {
2c541554 2054 top_border = h->top_borders[0][h->mb_x];
e5d40372 2055 AV_COPY128(top_border, src_y + 15 * linesize);
6e3ef511 2056 if (pixel_shift)
e5d40372 2057 AV_COPY128(top_border + 16, src_y + 15 * linesize + 16);
2c541554 2058 if (simple || !CONFIG_GRAY || !(h->flags & CODEC_FLAG_GRAY)) {
e5d40372
DB		 2059 if (chroma444) {
2060 if (pixel_shift) {
2061 AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
2062 AV_COPY128(top_border + 48, src_cb + 15 * uvlinesize + 16);
2063 AV_COPY128(top_border + 64, src_cr + 15 * uvlinesize);
2064 AV_COPY128(top_border + 80, src_cr + 15 * uvlinesize + 16);
c90b9442 2065 } else {
e5d40372
DB		 2066 AV_COPY128(top_border + 16, src_cb + 15 * uvlinesize);
2067 AV_COPY128(top_border + 32, src_cr + 15 * uvlinesize);
c90b9442 2068 }
e5d40372 2069 } else if (chroma422) {
76741b0e 2070 if (pixel_shift) {
e5d40372
DB		 2071 AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
2072 AV_COPY128(top_border + 48, src_cr + 15 * uvlinesize);
76741b0e 2073 } else {
e5d40372
DB		 2074 AV_COPY64(top_border + 16, src_cb + 15 * uvlinesize);
2075 AV_COPY64(top_border + 24, src_cr + 15 * uvlinesize);
76741b0e 2076 }
6e3ef511 2077 } else {
c90b9442 2078 if (pixel_shift) {
e5d40372
DB		 2079 AV_COPY128(top_border + 32, src_cb + 7 * uvlinesize);
2080 AV_COPY128(top_border + 48, src_cr + 7 * uvlinesize);
c90b9442 2081 } else {
e5d40372
DB		 2082 AV_COPY64(top_border + 16, src_cb + 7 * uvlinesize);
2083 AV_COPY64(top_border + 24, src_cr + 7 * uvlinesize);
c90b9442 2084 }
6e3ef511 2085 }
5f7f9719
MN		 2086 }
2087 }
82313eaa 2088 } else if (MB_MBAFF(h)) {
c988f975 2089 top_idx = 0;
e5d40372 2090 } else
c988f975 2091 return;
5f7f9719
MN		 2092 }
2093
2c541554 2094 top_border = h->top_borders[top_idx][h->mb_x];
09f21198 2095 /* There are two lines saved, the line above the top macroblock
e5d40372
DB		 2096 * of a pair, and the line above the bottom macroblock. */
2097 AV_COPY128(top_border, src_y + 16 * linesize);
6e3ef511 2098 if (pixel_shift)
e5d40372
DB		 2099 AV_COPY128(top_border + 16, src_y + 16 * linesize + 16);
2100
2c541554 2101 if (simple || !CONFIG_GRAY || !(h->flags & CODEC_FLAG_GRAY)) {
e5d40372
DB		 2102 if (chroma444) {
2103 if (pixel_shift) {
2104 AV_COPY128(top_border + 32, src_cb + 16 * linesize);
2105 AV_COPY128(top_border + 48, src_cb + 16 * linesize + 16);
2106 AV_COPY128(top_border + 64, src_cr + 16 * linesize);
2107 AV_COPY128(top_border + 80, src_cr + 16 * linesize + 16);
c90b9442 2108 } else {
e5d40372
DB		 2109 AV_COPY128(top_border + 16, src_cb + 16 * linesize);
2110 AV_COPY128(top_border + 32, src_cr + 16 * linesize);
c90b9442 2111 }
e5d40372 2112 } else if (chroma422) {
76741b0e 2113 if (pixel_shift) {
e5d40372
DB		 2114 AV_COPY128(top_border + 32, src_cb + 16 * uvlinesize);
2115 AV_COPY128(top_border + 48, src_cr + 16 * uvlinesize);
76741b0e 2116 } else {
e5d40372
DB		 2117 AV_COPY64(top_border + 16, src_cb + 16 * uvlinesize);
2118 AV_COPY64(top_border + 24, src_cr + 16 * uvlinesize);
76741b0e 2119 }
6e3ef511 2120 } else {
c90b9442 2121 if (pixel_shift) {
e5d40372
DB		 2122 AV_COPY128(top_border + 32, src_cb + 8 * uvlinesize);
2123 AV_COPY128(top_border + 48, src_cr + 8 * uvlinesize);
c90b9442 2124 } else {
e5d40372
DB		 2125 AV_COPY64(top_border + 16, src_cb + 8 * uvlinesize);
2126 AV_COPY64(top_border + 24, src_cr + 8 * uvlinesize);
c90b9442 2127 }
6e3ef511 2128 }
53c05b1e
MN		 2129 }
2130}
2131
bbdd52ed 2132static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
e5d40372
DB		 2133 uint8_t *src_cb, uint8_t *src_cr,
2134 int linesize, int uvlinesize,
2135 int xchg, int chroma444,
2136 int simple, int pixel_shift)
2137{
4e987f82 2138 int deblock_topleft;
b69378e2 2139 int deblock_top;
5f7f9719 2140 int top_idx = 1;
1e4f1c56
AS		 2141 uint8_t *top_border_m1;
2142 uint8_t *top_border;
5f7f9719 2143
7bece9b2 2144 if (!simple && FRAME_MBAFF(h)) {
2c541554 2145 if (h->mb_y & 1) {
82313eaa 2146 if (!MB_MBAFF(h))
c988f975 2147 return;
e5d40372 2148 } else {
82313eaa 2149 top_idx = MB_MBAFF(h) ? 0 : 1;
5f7f9719 2150 }
5f7f9719 2151 }
b69378e2 2152
e5d40372 2153 if (h->deblocking_filter == 2) {
2c541554 2154 deblock_topleft = h->slice_table[h->mb_xy - 1 - h->mb_stride] == h->slice_num;
4e987f82 2155 deblock_top = h->top_type;
b69378e2 2156 } else {
2c541554 2157 deblock_topleft = (h->mb_x > 0);
da6be8fc 2158 deblock_top = (h->mb_y > !!MB_FIELD(h));
b69378e2 2159 }
53c05b1e 2160
e5d40372 2161 src_y -= linesize + 1 + pixel_shift;
6e3ef511
OA		 2162 src_cb -= uvlinesize + 1 + pixel_shift;
2163 src_cr -= uvlinesize + 1 + pixel_shift;
53c05b1e 2164
2c541554
AK		 2165 top_border_m1 = h->top_borders[top_idx][h->mb_x - 1];
2166 top_border = h->top_borders[top_idx][h->mb_x];
1e4f1c56 2167
e5d40372
DB		 2168#define XCHG(a, b, xchg) \
2169 if (pixel_shift) { \
2170 if (xchg) { \
2171 AV_SWAP64(b + 0, a + 0); \
2172 AV_SWAP64(b + 8, a + 8); \
2173 } else { \
2174 AV_COPY128(b, a); \
2175 } \
2176 } else if (xchg) \
2177 AV_SWAP64(b, a); \
2178 else \
2179 AV_COPY64(b, a);
2180
2181 if (deblock_top) {
2182 if (deblock_topleft) {
2183 XCHG(top_border_m1 + (8 << pixel_shift),
2184 src_y - (7 << pixel_shift), 1);
c988f975 2185 }
6e3ef511
OA		 2186 XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
2187 XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
2c541554
AK		 2188 if (h->mb_x + 1 < h->mb_width) {
2189 XCHG(h->top_borders[top_idx][h->mb_x + 1],
e5d40372 2190 src_y + (17 << pixel_shift), 1);
43efd19a 2191 }
53c05b1e 2192 }
2c541554 2193 if (simple || !CONFIG_GRAY || !(h->flags & CODEC_FLAG_GRAY)) {
e5d40372
DB		 2194 if (chroma444) {
2195 if (deblock_topleft) {
c90b9442
JGG		 2196 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
2197 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
2198 }
2199 XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
2200 XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
2201 XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
2202 XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
2c541554
AK		 2203 if (h->mb_x + 1 < h->mb_width) {
2204 XCHG(h->top_borders[top_idx][h->mb_x + 1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
2205 XCHG(h->top_borders[top_idx][h->mb_x + 1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
c90b9442
JGG		 2206 }
2207 } else {
e5d40372
DB		 2208 if (deblock_top) {
2209 if (deblock_topleft) {
c90b9442
JGG		 2210 XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
2211 XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
2212 }
e5d40372
DB		 2213 XCHG(top_border + (16 << pixel_shift), src_cb + 1 + pixel_shift, 1);
2214 XCHG(top_border + (24 << pixel_shift), src_cr + 1 + pixel_shift, 1);
c988f975 2215 }
53c05b1e 2216 }
53c05b1e
MN		 2217 }
2218}
2219
88bd7fdc 2220static av_always_inline int dctcoef_get(int16_t *mb, int high_bit_depth,
e5d40372
DB		 2221 int index)
2222{
6e3ef511 2223 if (high_bit_depth) {
e5d40372 2224 return AV_RN32A(((int32_t *)mb) + index);
6e3ef511
OA		 2225 } else
2226 return AV_RN16A(mb + index);
2227}
2228
88bd7fdc 2229static av_always_inline void dctcoef_set(int16_t *mb, int high_bit_depth,
e5d40372
DB		 2230 int index, int value)
2231{
6e3ef511 2232 if (high_bit_depth) {
e5d40372 2233 AV_WN32A(((int32_t *)mb) + index, value);
6e3ef511
OA		 2234 } else
2235 AV_WN16A(mb + index, value);
2236}
2237
e5d40372
DB		 2238static av_always_inline void hl_decode_mb_predict_luma(H264Context *h,
2239 int mb_type, int is_h264,
2240 int simple,
2241 int transform_bypass,
2242 int pixel_shift,
2243 int *block_offset,
2244 int linesize,
2245 uint8_t *dest_y, int p)
c90b9442 2246{
88bd7fdc
DB		 2247 void (*idct_add)(uint8_t *dst, int16_t *block, int stride);
2248 void (*idct_dc_add)(uint8_t *dst, int16_t *block, int stride);
c90b9442 2249 int i;
2c541554 2250 int qscale = p == 0 ? h->qscale : h->chroma_qp[p - 1];
e5d40372
DB		 2251 block_offset += 16 * p;
2252 if (IS_INTRA4x4(mb_type)) {
2491f9ee
AK		 2253 if (IS_8x8DCT(mb_type)) {
2254 if (transform_bypass) {
2255 idct_dc_add =
2ed00820 2256 idct_add = h->h264dsp.h264_add_pixels8;
2491f9ee
AK		 2257 } else {
2258 idct_dc_add = h->h264dsp.h264_idct8_dc_add;
2259 idct_add = h->h264dsp.h264_idct8_add;
2260 }
2261 for (i = 0; i < 16; i += 4) {
2262 uint8_t *const ptr = dest_y + block_offset[i];
2263 const int dir = h->intra4x4_pred_mode_cache[scan8[i]];
2264 if (transform_bypass && h->sps.profile_idc == 244 && dir <= 1) {
2265 h->hpc.pred8x8l_add[dir](ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
e5d40372 2266 } else {
2491f9ee
AK		 2267 const int nnz = h->non_zero_count_cache[scan8[i + p * 16]];
2268 h->hpc.pred8x8l[dir](ptr, (h->topleft_samples_available << i) & 0x8000,
2269 (h->topright_samples_available << i) & 0x4000, linesize);
2270 if (nnz) {
2271 if (nnz == 1 && dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
2272 idct_dc_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
2273 else
2274 idct_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
c90b9442
JGG		 2275 }
2276 }
2491f9ee
AK		 2277 }
2278 } else {
2279 if (transform_bypass) {
2280 idct_dc_add =
2ed00820 2281 idct_add = h->h264dsp.h264_add_pixels4;
e5d40372 2282 } else {
2491f9ee
AK		 2283 idct_dc_add = h->h264dsp.h264_idct_dc_add;
2284 idct_add = h->h264dsp.h264_idct_add;
2285 }
2286 for (i = 0; i < 16; i++) {
2287 uint8_t *const ptr = dest_y + block_offset[i];
2288 const int dir = h->intra4x4_pred_mode_cache[scan8[i]];
c90b9442 2289
2491f9ee
AK		 2290 if (transform_bypass && h->sps.profile_idc == 244 && dir <= 1) {
2291 h->hpc.pred4x4_add[dir](ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
2292 } else {
2293 uint8_t *topright;
2294 int nnz, tr;
2295 uint64_t tr_high;
2296 if (dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED) {
2297 const int topright_avail = (h->topright_samples_available << i) & 0x8000;
2c541554 2298 assert(h->mb_y || linesize <= block_offset[i]);
2491f9ee
AK		 2299 if (!topright_avail) {
2300 if (pixel_shift) {
2301 tr_high = ((uint16_t *)ptr)[3 - linesize / 2] * 0x0001000100010001ULL;
2302 topright = (uint8_t *)&tr_high;
2303 } else {
2304 tr = ptr[3 - linesize] * 0x01010101u;
2305 topright = (uint8_t *)&tr;
2306 }
e5d40372 2307 } else
2491f9ee
AK		 2308 topright = ptr + (4 << pixel_shift) - linesize;
2309 } else
2310 topright = NULL;
2311
2312 h->hpc.pred4x4[dir](ptr, topright, linesize);
2313 nnz = h->non_zero_count_cache[scan8[i + p * 16]];
2314 if (nnz) {
2315 if (is_h264) {
2316 if (nnz == 1 && dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
2317 idct_dc_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
2318 else
2319 idct_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
2320 } else if (CONFIG_SVQ3_DECODER)
2321 ff_svq3_add_idct_c(ptr, h->mb + i * 16 + p * 256, linesize, qscale, 0);
c90b9442
JGG		 2322 }
2323 }
2324 }
2325 }
e5d40372
DB		 2326 } else {
2327 h->hpc.pred16x16[h->intra16x16_pred_mode](dest_y, linesize);
2328 if (is_h264) {
2329 if (h->non_zero_count_cache[scan8[LUMA_DC_BLOCK_INDEX + p]]) {
2330 if (!transform_bypass)
2331 h->h264dsp.h264_luma_dc_dequant_idct(h->mb + (p * 256 << pixel_shift),
2332 h->mb_luma_dc[p],
2333 h->dequant4_coeff[p][qscale][0]);
2334 else {
2335 static const uint8_t dc_mapping[16] = {
2336 0 * 16, 1 * 16, 4 * 16, 5 * 16,
2337 2 * 16, 3 * 16, 6 * 16, 7 * 16,
2338 8 * 16, 9 * 16, 12 * 16, 13 * 16,
2339 10 * 16, 11 * 16, 14 * 16, 15 * 16 };
2340 for (i = 0; i < 16; i++)
2341 dctcoef_set(h->mb + (p * 256 << pixel_shift),
2342 pixel_shift, dc_mapping[i],
2343 dctcoef_get(h->mb_luma_dc[p],
2344 pixel_shift, i));
c90b9442
JGG		 2345 }
2346 }
301fb921 2347 } else if (CONFIG_SVQ3_DECODER)
e5d40372
DB		 2348 ff_svq3_luma_dc_dequant_idct_c(h->mb + p * 256,
2349 h->mb_luma_dc[p], qscale);
c90b9442
JGG		 2350 }
2351}
2352
e5d40372
DB		 2353static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type,
2354 int is_h264, int simple,
2355 int transform_bypass,
2356 int pixel_shift,
2357 int *block_offset,
2358 int linesize,
2359 uint8_t *dest_y, int p)
c90b9442 2360{
88bd7fdc 2361 void (*idct_add)(uint8_t *dst, int16_t *block, int stride);
c90b9442 2362 int i;
e5d40372
DB		 2363 block_offset += 16 * p;
2364 if (!IS_INTRA4x4(mb_type)) {
2365 if (is_h264) {
2366 if (IS_INTRA16x16(mb_type)) {
2367 if (transform_bypass) {
2368 if (h->sps.profile_idc == 244 &&
2369 (h->intra16x16_pred_mode == VERT_PRED8x8 ||
2370 h->intra16x16_pred_mode == HOR_PRED8x8)) {
2371 h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset,
2372 h->mb + (p * 256 << pixel_shift),
2373 linesize);
2374 } else {
2375 for (i = 0; i < 16; i++)
2376 if (h->non_zero_count_cache[scan8[i + p * 16]] ||
2377 dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
2ed00820
RB		 2378 h->h264dsp.h264_add_pixels4(dest_y + block_offset[i],
2379 h->mb + (i * 16 + p * 256 << pixel_shift),
2380 linesize);
c90b9442 2381 }
e5d40372
DB		 2382 } else {
2383 h->h264dsp.h264_idct_add16intra(dest_y, block_offset,
2384 h->mb + (p * 256 << pixel_shift),
2385 linesize,
2386 h->non_zero_count_cache + p * 5 * 8);
c90b9442 2387 }
e5d40372
DB		 2388 } else if (h->cbp & 15) {
2389 if (transform_bypass) {
c90b9442 2390 const int di = IS_8x8DCT(mb_type) ? 4 : 1;
2ed00820
RB		 2391 idct_add = IS_8x8DCT(mb_type) ? h->h264dsp.h264_add_pixels8
2392 : h->h264dsp.h264_add_pixels4;
e5d40372
DB		 2393 for (i = 0; i < 16; i += di)
2394 if (h->non_zero_count_cache[scan8[i + p * 16]])
2395 idct_add(dest_y + block_offset[i],
2396 h->mb + (i * 16 + p * 256 << pixel_shift),
2397 linesize);
2398 } else {
2399 if (IS_8x8DCT(mb_type))
2400 h->h264dsp.h264_idct8_add4(dest_y, block_offset,
2401 h->mb + (p * 256 << pixel_shift),
2402 linesize,
2403 h->non_zero_count_cache + p * 5 * 8);
2404 else
2405 h->h264dsp.h264_idct_add16(dest_y, block_offset,
2406 h->mb + (p * 256 << pixel_shift),
2407 linesize,
2408 h->non_zero_count_cache + p * 5 * 8);
c90b9442
JGG		 2409 }
2410 }
301fb921 2411 } else if (CONFIG_SVQ3_DECODER) {
e5d40372
DB		 2412 for (i = 0; i < 16; i++)
2413 if (h->non_zero_count_cache[scan8[i + p * 16]] || h->mb[i * 16 + p * 256]) {
2414 // FIXME benchmark weird rule, & below
2415 uint8_t *const ptr = dest_y + block_offset[i];
2416 ff_svq3_add_idct_c(ptr, h->mb + i * 16 + p * 256, linesize,
2c541554 2417 h->qscale, IS_INTRA(mb_type) ? 1 : 0);
c90b9442 2418 }
c90b9442
JGG		 2419 }
2420 }
2421}
2422
28fff0d9
MR		 2423#define BITS 8
2424#define SIMPLE 1
2425#include "h264_mb_template.c"
e5d40372 2426
28fff0d9
MR		 2427#undef BITS
2428#define BITS 16
2429#include "h264_mb_template.c"
bd91fee3 2430
28fff0d9
MR		 2431#undef SIMPLE
2432#define SIMPLE 0
2433#include "h264_mb_template.c"
c90b9442 2434
e5d40372
DB		 2435void ff_h264_hl_decode_mb(H264Context *h)
2436{
e5d40372 2437 const int mb_xy = h->mb_xy;
759001c5 2438 const int mb_type = h->cur_pic.mb_type[mb_xy];
2c541554 2439 int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || h->qscale == 0;
bd91fee3 2440
23e85be5 2441 if (CHROMA444(h)) {
e5d40372 2442 if (is_complex || h->pixel_shift)
c90b9442
JGG		 2443 hl_decode_mb_444_complex(h);
2444 else
28fff0d9 2445 hl_decode_mb_444_simple_8(h);
c90b9442 2446 } else if (is_complex) {
bd91fee3 2447 hl_decode_mb_complex(h);
6e3ef511
OA		 2448 } else if (h->pixel_shift) {
2449 hl_decode_mb_simple_16(h);
2450 } else
2451 hl_decode_mb_simple_8(h);
bd91fee3
AS		 2452}
2453
e5d40372
DB		 2454static int pred_weight_table(H264Context *h)
2455{
0da71265 2456 int list, i;
9f2d1b4f 2457 int luma_def, chroma_def;
115329f1 2458
e5d40372
DB		 2459 h->use_weight = 0;
2460 h->use_weight_chroma = 0;
2c541554 2461 h->luma_log2_weight_denom = get_ue_golomb(&h->gb);
e5d40372 2462 if (h->sps.chroma_format_idc)
2c541554 2463 h->chroma_log2_weight_denom = get_ue_golomb(&h->gb);
e5d40372
DB		 2464 luma_def = 1 << h->luma_log2_weight_denom;
2465 chroma_def = 1 << h->chroma_log2_weight_denom;
0da71265 2466
e5d40372 2467 for (list = 0; list < 2; list++) {
cb99c652
GB		 2468 h->luma_weight_flag[list] = 0;
2469 h->chroma_weight_flag[list] = 0;
e5d40372 2470 for (i = 0; i < h->ref_count[list]; i++) {
0da71265 2471 int luma_weight_flag, chroma_weight_flag;
115329f1 2472
2c541554 2473 luma_weight_flag = get_bits1(&h->gb);
e5d40372 2474 if (luma_weight_flag) {
2c541554
AK		 2475 h->luma_weight[i][list][0] = get_se_golomb(&h->gb);
2476 h->luma_weight[i][list][1] = get_se_golomb(&h->gb);
e5d40372
DB		 2477 if (h->luma_weight[i][list][0] != luma_def ||
2478 h->luma_weight[i][list][1] != 0) {
2479 h->use_weight = 1;
2480 h->luma_weight_flag[list] = 1;
cb99c652 2481 }
e5d40372
DB		 2482 } else {
2483 h->luma_weight[i][list][0] = luma_def;
2484 h->luma_weight[i][list][1] = 0;
0da71265
MN		 2485 }
2486
e5d40372 2487 if (h->sps.chroma_format_idc) {
2c541554 2488 chroma_weight_flag = get_bits1(&h->gb);
e5d40372 2489 if (chroma_weight_flag) {
fef744d4 2490 int j;
e5d40372 2491 for (j = 0; j < 2; j++) {
2c541554
AK		 2492 h->chroma_weight[i][list][j][0] = get_se_golomb(&h->gb);
2493 h->chroma_weight[i][list][j][1] = get_se_golomb(&h->gb);
e5d40372
DB		 2494 if (h->chroma_weight[i][list][j][0] != chroma_def ||
2495 h->chroma_weight[i][list][j][1] != 0) {
2496 h->use_weight_chroma = 1;
2497 h->chroma_weight_flag[list] = 1;
cb99c652 2498 }
fef744d4 2499 }
e5d40372 2500 } else {
fef744d4 2501 int j;
e5d40372
DB		 2502 for (j = 0; j < 2; j++) {
2503 h->chroma_weight[i][list][j][0] = chroma_def;
2504 h->chroma_weight[i][list][j][1] = 0;
fef744d4 2505 }
0da71265
MN		 2506 }
2507 }
2508 }
e5d40372
DB		 2509 if (h->slice_type_nos != AV_PICTURE_TYPE_B)
2510 break;
0da71265 2511 }
e5d40372 2512 h->use_weight = h->use_weight || h->use_weight_chroma;
0da71265
MN		 2513 return 0;
2514}
2515
1052b76f
MN		 2516/**
2517 * Initialize implicit_weight table.
6da88bd3 2518 * @param field 0/1 initialize the weight for interlaced MBAFF
1052b76f
MN		 2519 * -1 initializes the rest
2520 */
e5d40372
DB		 2521static void implicit_weight_table(H264Context *h, int field)
2522{
1052b76f 2523 int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
9f2d1b4f 2524
ce09f927
GB		 2525 for (i = 0; i < 2; i++) {
2526 h->luma_weight_flag[i] = 0;
2527 h->chroma_weight_flag[i] = 0;
2528 }
2529
e5d40372 2530 if (field < 0) {
2c541554
AK		 2531 if (h->picture_structure == PICT_FRAME) {
2532 cur_poc = h->cur_pic_ptr->poc;
4418aa9c 2533 } else {
2c541554 2534 cur_poc = h->cur_pic_ptr->field_poc[h->picture_structure - 1];
4418aa9c 2535 }
7bece9b2 2536 if (h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF(h) &&
e5d40372
DB		 2537 h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2 * cur_poc) {
2538 h->use_weight = 0;
2539 h->use_weight_chroma = 0;
2540 return;
2541 }
2542 ref_start = 0;
2543 ref_count0 = h->ref_count[0];
2544 ref_count1 = h->ref_count[1];
2545 } else {
2c541554 2546 cur_poc = h->cur_pic_ptr->field_poc[field];
e5d40372
DB		 2547 ref_start = 16;
2548 ref_count0 = 16 + 2 * h->ref_count[0];
2549 ref_count1 = 16 + 2 * h->ref_count[1];
1052b76f 2550 }
9f2d1b4f 2551
e5d40372
DB		 2552 h->use_weight = 2;
2553 h->use_weight_chroma = 2;
2554 h->luma_log2_weight_denom = 5;
2555 h->chroma_log2_weight_denom = 5;
9f2d1b4f 2556
e5d40372 2557 for (ref0 = ref_start; ref0 < ref_count0; ref0++) {
9f2d1b4f 2558 int poc0 = h->ref_list[0][ref0].poc;
e5d40372 2559 for (ref1 = ref_start; ref1 < ref_count1; ref1++) {
87cf70eb
JD		 2560 int w = 32;
2561 if (!h->ref_list[0][ref0].long_ref && !h->ref_list[1][ref1].long_ref) {
2562 int poc1 = h->ref_list[1][ref1].poc;
e5d40372
DB		 2563 int td = av_clip(poc1 - poc0, -128, 127);
2564 if (td) {
87cf70eb
JD		 2565 int tb = av_clip(cur_poc - poc0, -128, 127);
2566 int tx = (16384 + (FFABS(td) >> 1)) / td;
e5d40372
DB		 2567 int dist_scale_factor = (tb * tx + 32) >> 8;
2568 if (dist_scale_factor >= -64 && dist_scale_factor <= 128)
87cf70eb
JD		 2569 w = 64 - dist_scale_factor;
2570 }
1052b76f 2571 }
e5d40372
DB		 2572 if (field < 0) {
2573 h->implicit_weight[ref0][ref1][0] =
2574 h->implicit_weight[ref0][ref1][1] = w;
2575 } else {
2576 h->implicit_weight[ref0][ref1][field] = w;
72f86ec0 2577 }
9f2d1b4f
LM		 2578 }
2579 }
2580}
2581
8fd57a66 2582/**
5175b937 2583 * instantaneous decoder refresh.
0da71265 2584 */
e5d40372
DB		 2585static void idr(H264Context *h)
2586{
ea6f00c4 2587 ff_h264_remove_all_refs(h);
e5d40372
DB		 2588 h->prev_frame_num = 0;
2589 h->prev_frame_num_offset = 0;
2590 h->prev_poc_msb =
2591 h->prev_poc_lsb = 0;
0da71265
MN		 2592}
2593
7c33ad19 2594/* forget old pics after a seek */
9e696d2e 2595static void flush_change(H264Context *h)
e5d40372 2596{
7c33ad19 2597 int i;
adedd840
RB		 2598 for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
2599 h->last_pocs[i] = INT_MIN;
e5d40372 2600 h->outputed_poc = h->next_outputed_poc = INT_MIN;
b19d493f 2601 h->prev_interlaced_frame = 1;
7c33ad19 2602 idr(h);
2c541554 2603 if (h->cur_pic_ptr)
759001c5 2604 h->cur_pic_ptr->reference = 0;
2c541554 2605 h->first_field = 0;
9e696d2e
JG		 2606 memset(h->ref_list[0], 0, sizeof(h->ref_list[0]));
2607 memset(h->ref_list[1], 0, sizeof(h->ref_list[1]));
2608 memset(h->default_ref_list[0], 0, sizeof(h->default_ref_list[0]));
2609 memset(h->default_ref_list[1], 0, sizeof(h->default_ref_list[1]));
9c095463 2610 ff_h264_reset_sei(h);
9e696d2e
JG		 2611}
2612
2613/* forget old pics after a seek */
2614static void flush_dpb(AVCodecContext *avctx)
2615{
2616 H264Context *h = avctx->priv_data;
2617 int i;
2618
2619 for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++) {
2620 if (h->delayed_pic[i])
759001c5 2621 h->delayed_pic[i]->reference = 0;
9e696d2e
JG		 2622 h->delayed_pic[i] = NULL;
2623 }
2624
2625 flush_change(h);
2c541554 2626
555000c7
AK		 2627 if (h->DPB)
2628 for (i = 0; i < MAX_PICTURE_COUNT; i++)
2629 unref_picture(h, &h->DPB[i]);
2c541554 2630 h->cur_pic_ptr = NULL;
759001c5 2631 unref_picture(h, &h->cur_pic);
2c541554
AK		 2632
2633 h->mb_x = h->mb_y = 0;
2634
2635 h->parse_context.state = -1;
2636 h->parse_context.frame_start_found = 0;
2637 h->parse_context.overread = 0;
2638 h->parse_context.overread_index = 0;
2639 h->parse_context.index = 0;
2640 h->parse_context.last_index = 0;
7c33ad19
LM		 2641}
2642
e5d40372
DB		 2643static int init_poc(H264Context *h)
2644{
e5d40372 2645 const int max_frame_num = 1 << h->sps.log2_max_frame_num;
0da71265 2646 int field_poc[2];
2c541554 2647 Picture *cur = h->cur_pic_ptr;
0da71265 2648
e5d40372
DB		 2649 h->frame_num_offset = h->prev_frame_num_offset;
2650 if (h->frame_num < h->prev_frame_num)
b78a6baa 2651 h->frame_num_offset += max_frame_num;
0da71265 2652
e5d40372
DB		 2653 if (h->sps.poc_type == 0) {
2654 const int max_poc_lsb = 1 << h->sps.log2_max_poc_lsb;
0da71265 2655
e5d40372 2656 if (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb / 2)
0da71265 2657 h->poc_msb = h->prev_poc_msb + max_poc_lsb;
e5d40372 2658 else if (h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb / 2)
0da71265
MN		 2659 h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2660 else
2661 h->poc_msb = h->prev_poc_msb;
115329f1 2662 field_poc[0] =
0da71265 2663 field_poc[1] = h->poc_msb + h->poc_lsb;
2c541554 2664 if (h->picture_structure == PICT_FRAME)
0da71265 2665 field_poc[1] += h->delta_poc_bottom;
e5d40372 2666 } else if (h->sps.poc_type == 1) {
0da71265
MN		 2667 int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2668 int i;
2669
e5d40372 2670 if (h->sps.poc_cycle_length != 0)
0da71265
MN		 2671 abs_frame_num = h->frame_num_offset + h->frame_num;
2672 else
2673 abs_frame_num = 0;
2674
e5d40372 2675 if (h->nal_ref_idc == 0 && abs_frame_num > 0)
0da71265 2676 abs_frame_num--;
115329f1 2677
0da71265 2678 expected_delta_per_poc_cycle = 0;
e5d40372
DB		 2679 for (i = 0; i < h->sps.poc_cycle_length; i++)
2680 // FIXME integrate during sps parse
2681 expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[i];
0da71265 2682
e5d40372 2683 if (abs_frame_num > 0) {
0da71265
MN		 2684 int poc_cycle_cnt = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2685 int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2686
2687 expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
e5d40372
DB		 2688 for (i = 0; i <= frame_num_in_poc_cycle; i++)
2689 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[i];
0da71265
MN		 2690 } else
2691 expectedpoc = 0;
2692
e5d40372 2693 if (h->nal_ref_idc == 0)
0da71265 2694 expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
115329f1 2695
0da71265
MN		 2696 field_poc[0] = expectedpoc + h->delta_poc[0];
2697 field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2698
2c541554 2699 if (h->picture_structure == PICT_FRAME)
0da71265 2700 field_poc[1] += h->delta_poc[1];
e5d40372
DB		 2701 } else {
2702 int poc = 2 * (h->frame_num_offset + h->frame_num);
5710b371 2703
e5d40372 2704 if (!h->nal_ref_idc)
b78a6baa 2705 poc--;
5710b371 2706
e5d40372
DB		 2707 field_poc[0] = poc;
2708 field_poc[1] = poc;
0da71265 2709 }
115329f1 2710
2c541554
AK		 2711 if (h->picture_structure != PICT_BOTTOM_FIELD)
2712 h->cur_pic_ptr->field_poc[0] = field_poc[0];
2713 if (h->picture_structure != PICT_TOP_FIELD)
2714 h->cur_pic_ptr->field_poc[1] = field_poc[1];
e5d40372 2715 cur->poc = FFMIN(cur->field_poc[0], cur->field_poc[1]);
0da71265
MN		 2716
2717 return 0;
2718}
2719
b41c1db3
AÖ		 2720/**
2721 * initialize scan tables
2722 */
e5d40372
DB		 2723static void init_scan_tables(H264Context *h)
2724{
b41c1db3 2725 int i;
e5d40372
DB		 2726 for (i = 0; i < 16; i++) {
2727#define T(x) (x >> 2) | ((x << 2) & 0xF)
ca32f7f2 2728 h->zigzag_scan[i] = T(zigzag_scan[i]);
e5d40372 2729 h->field_scan[i] = T(field_scan[i]);
b41c1db3 2730#undef T
b41c1db3 2731 }
e5d40372
DB		 2732 for (i = 0; i < 64; i++) {
2733#define T(x) (x >> 3) | ((x & 7) << 3)
ca32f7f2
JGG		 2734 h->zigzag_scan8x8[i] = T(ff_zigzag_direct[i]);
2735 h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2736 h->field_scan8x8[i] = T(field_scan8x8[i]);
2737 h->field_scan8x8_cavlc[i] = T(field_scan8x8_cavlc[i]);
b41c1db3 2738#undef T
b41c1db3 2739 }
e5d40372 2740 if (h->sps.transform_bypass) { // FIXME same ugly
b41c1db3 2741 h->zigzag_scan_q0 = zigzag_scan;
45beb850 2742 h->zigzag_scan8x8_q0 = ff_zigzag_direct;
b41c1db3
AÖ		 2743 h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2744 h->field_scan_q0 = field_scan;
2745 h->field_scan8x8_q0 = field_scan8x8;
2746 h->field_scan8x8_cavlc_q0 = field_scan8x8_cavlc;
e5d40372 2747 } else {
b41c1db3
AÖ		 2748 h->zigzag_scan_q0 = h->zigzag_scan;
2749 h->zigzag_scan8x8_q0 = h->zigzag_scan8x8;
2750 h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2751 h->field_scan_q0 = h->field_scan;
2752 h->field_scan8x8_q0 = h->field_scan8x8;
2753 h->field_scan8x8_cavlc_q0 = h->field_scan8x8_cavlc;
2754 }
2755}
afebe2f7 2756
e5d40372
DB		 2757static int field_end(H264Context *h, int in_setup)
2758{
2c541554 2759 AVCodecContext *const avctx = h->avctx;
12fe7594 2760 int err = 0;
2c541554 2761 h->mb_y = 0;
256299d3 2762
2c541554 2763 if (!in_setup && !h->droppable)
759001c5 2764 ff_thread_report_progress(&h->cur_pic_ptr->tf, INT_MAX,
2c541554 2765 h->picture_structure == PICT_BOTTOM_FIELD);
256299d3 2766
e5d40372 2767 if (CONFIG_H264_VDPAU_DECODER &&
2c541554
AK		 2768 h->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
2769 ff_vdpau_h264_set_reference_frames(h);
256299d3 2770
e5d40372 2771 if (in_setup || !(avctx->active_thread_type & FF_THREAD_FRAME)) {
2c541554 2772 if (!h->droppable) {
12fe7594 2773 err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
e5d40372
DB		 2774 h->prev_poc_msb = h->poc_msb;
2775 h->prev_poc_lsb = h->poc_lsb;
6a9c8594 2776 }
e5d40372
DB		 2777 h->prev_frame_num_offset = h->frame_num_offset;
2778 h->prev_frame_num = h->frame_num;
2779 h->outputed_poc = h->next_outputed_poc;
256299d3 2780 }
256299d3
MN		 2781
2782 if (avctx->hwaccel) {
2783 if (avctx->hwaccel->end_frame(avctx) < 0)
e5d40372
DB		 2784 av_log(avctx, AV_LOG_ERROR,
2785 "hardware accelerator failed to decode picture\n");
256299d3
MN		 2786 }
2787
e5d40372 2788 if (CONFIG_H264_VDPAU_DECODER &&
2c541554
AK		 2789 h->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
2790 ff_vdpau_h264_picture_complete(h);
256299d3
MN		 2791
2792 /*
2793 * FIXME: Error handling code does not seem to support interlaced
2794 * when slices span multiple rows
2795 * The ff_er_add_slice calls don't work right for bottom
2796 * fields; they cause massive erroneous error concealing
2797 * Error marking covers both fields (top and bottom).
2798 * This causes a mismatched s->error_count
2799 * and a bad error table. Further, the error count goes to
2800 * INT_MAX when called for bottom field, because mb_y is
2801 * past end by one (callers fault) and resync_mb_y != 0
2802 * causes problems for the first MB line, too.
2803 */
7fa00653 2804 if (!FIELD_PICTURE(h)) {
2c541554
AK		 2805 h->er.cur_pic = h->cur_pic_ptr;
2806 h->er.last_pic = h->ref_count[0] ? &h->ref_list[0][0] : NULL;
2807 h->er.next_pic = h->ref_count[1] ? &h->ref_list[1][0] : NULL;
2808 ff_er_frame_end(&h->er);
2809 }
2c541554 2810 emms_c();
d225a1e2 2811
e5d40372 2812 h->current_slice = 0;
12fe7594
DB		 2813
2814 return err;
256299d3
MN		 2815}
2816
afebe2f7 2817/**
49bd8e4b 2818 * Replicate H264 "master" context to thread contexts.
afebe2f7 2819 */
f1d8763a 2820static int clone_slice(H264Context *dst, H264Context *src)
afebe2f7 2821{
e5d40372 2822 memcpy(dst->block_offset, src->block_offset, sizeof(dst->block_offset));
2c541554
AK		 2823 dst->cur_pic_ptr = src->cur_pic_ptr;
2824 dst->cur_pic = src->cur_pic;
2825 dst->linesize = src->linesize;
2826 dst->uvlinesize = src->uvlinesize;
2827 dst->first_field = src->first_field;
f1d8763a 2828
e5d40372
DB		 2829 dst->prev_poc_msb = src->prev_poc_msb;
2830 dst->prev_poc_lsb = src->prev_poc_lsb;
2831 dst->prev_frame_num_offset = src->prev_frame_num_offset;
2832 dst->prev_frame_num = src->prev_frame_num;
2833 dst->short_ref_count = src->short_ref_count;
afebe2f7
AÖ		 2834
2835 memcpy(dst->short_ref, src->short_ref, sizeof(dst->short_ref));
2836 memcpy(dst->long_ref, src->long_ref, sizeof(dst->long_ref));
2837 memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
50c21814
AÖ		 2838
2839 memcpy(dst->dequant4_coeff, src->dequant4_coeff, sizeof(src->dequant4_coeff));
2840 memcpy(dst->dequant8_coeff, src->dequant8_coeff, sizeof(src->dequant8_coeff));
f1d8763a
JG		 2841
2842 return 0;
afebe2f7
AÖ		 2843}
2844
0da71265 2845/**
58c42af7 2846 * Compute profile from profile_idc and constraint_set?_flags.
fe9a3fbe
JG		 2847 *
2848 * @param sps SPS
2849 *
2850 * @return profile as defined by FF_PROFILE_H264_*
2851 */
2852int ff_h264_get_profile(SPS *sps)
2853{
2854 int profile = sps->profile_idc;
2855
e5d40372 2856 switch (sps->profile_idc) {
fe9a3fbe
JG		 2857 case FF_PROFILE_H264_BASELINE:
2858 // constraint_set1_flag set to 1
e5d40372 2859 profile |= (sps->constraint_set_flags & 1 << 1) ? FF_PROFILE_H264_CONSTRAINED : 0;
fe9a3fbe
JG		 2860 break;
2861 case FF_PROFILE_H264_HIGH_10:
2862 case FF_PROFILE_H264_HIGH_422:
2863 case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2864 // constraint_set3_flag set to 1
e5d40372 2865 profile |= (sps->constraint_set_flags & 1 << 3) ? FF_PROFILE_H264_INTRA : 0;
fe9a3fbe
JG		 2866 break;
2867 }
2868
2869 return profile;
2870}
2871
072be3e8
JG		 2872static int h264_set_parameter_from_sps(H264Context *h)
2873{
2c541554 2874 if (h->flags & CODEC_FLAG_LOW_DELAY ||
072be3e8
JG		 2875 (h->sps.bitstream_restriction_flag &&
2876 !h->sps.num_reorder_frames)) {
2c541554
AK		 2877 if (h->avctx->has_b_frames > 1 || h->delayed_pic[0])
2878 av_log(h->avctx, AV_LOG_WARNING, "Delayed frames seen. "
072be3e8
JG		 2879 "Reenabling low delay requires a codec flush.\n");
2880 else
2c541554 2881 h->low_delay = 1;
072be3e8
JG		 2882 }
2883
2c541554
AK		 2884 if (h->avctx->has_b_frames < 2)
2885 h->avctx->has_b_frames = !h->low_delay;
072be3e8 2886
4987faee 2887 if (h->sps.bit_depth_luma != h->sps.bit_depth_chroma) {
12e25ed2
DB		 2888 avpriv_request_sample(h->avctx,
2889 "Different chroma and luma bit depth");
4987faee
LB		 2890 return AVERROR_PATCHWELCOME;
2891 }
2892
2c541554 2893 if (h->avctx->bits_per_raw_sample != h->sps.bit_depth_luma ||
072be3e8 2894 h->cur_chroma_format_idc != h->sps.chroma_format_idc) {
2c541554
AK		 2895 if (h->avctx->codec &&
2896 h->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU &&
072be3e8 2897 (h->sps.bit_depth_luma != 8 || h->sps.chroma_format_idc > 1)) {
2c541554 2898 av_log(h->avctx, AV_LOG_ERROR,
072be3e8
JG		 2899 "VDPAU decoding does not support video colorspace.\n");
2900 return AVERROR_INVALIDDATA;
2901 }
2902 if (h->sps.bit_depth_luma >= 8 && h->sps.bit_depth_luma <= 10) {
2c541554 2903 h->avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
072be3e8
JG		 2904 h->cur_chroma_format_idc = h->sps.chroma_format_idc;
2905 h->pixel_shift = h->sps.bit_depth_luma > 8;
2906
2907 ff_h264dsp_init(&h->h264dsp, h->sps.bit_depth_luma,
2908 h->sps.chroma_format_idc);
79dad2a9 2909 ff_h264chroma_init(&h->h264chroma, h->sps.bit_depth_chroma);
e9d81735 2910 ff_h264qpel_init(&h->h264qpel, h->sps.bit_depth_luma);
2c541554 2911 ff_h264_pred_init(&h->hpc, h->avctx->codec_id, h->sps.bit_depth_luma,
072be3e8 2912 h->sps.chroma_format_idc);
2c541554
AK		 2913 h->dsp.dct_bits = h->sps.bit_depth_luma > 8 ? 32 : 16;
2914 ff_dsputil_init(&h->dsp, h->avctx);
2915 ff_videodsp_init(&h->vdsp, h->sps.bit_depth_luma);
072be3e8 2916 } else {
2c541554 2917 av_log(h->avctx, AV_LOG_ERROR, "Unsupported bit depth: %d\n",
072be3e8
JG		 2918 h->sps.bit_depth_luma);
2919 return AVERROR_INVALIDDATA;
2920 }
2921 }
2922 return 0;
2923}
2924
542b83fc 2925static enum AVPixelFormat get_pixel_format(H264Context *h)
9e696d2e 2926{
9e696d2e
JG		 2927 switch (h->sps.bit_depth_luma) {
2928 case 9:
23e85be5 2929 if (CHROMA444(h)) {
2c541554 2930 if (h->avctx->colorspace == AVCOL_SPC_RGB) {
9e696d2e
JG		 2931 return AV_PIX_FMT_GBRP9;
2932 } else
2933 return AV_PIX_FMT_YUV444P9;
e962bd08 2934 } else if (CHROMA422(h))
9e696d2e
JG		 2935 return AV_PIX_FMT_YUV422P9;
2936 else
2937 return AV_PIX_FMT_YUV420P9;
2938 break;
2939 case 10:
23e85be5 2940 if (CHROMA444(h)) {
2c541554 2941 if (h->avctx->colorspace == AVCOL_SPC_RGB) {
9e696d2e
JG		 2942 return AV_PIX_FMT_GBRP10;
2943 } else
2944 return AV_PIX_FMT_YUV444P10;
e962bd08 2945 } else if (CHROMA422(h))
9e696d2e
JG		 2946 return AV_PIX_FMT_YUV422P10;
2947 else
2948 return AV_PIX_FMT_YUV420P10;
2949 break;
2950 case 8:
23e85be5 2951 if (CHROMA444(h)) {
2c541554 2952 if (h->avctx->colorspace == AVCOL_SPC_RGB) {
9e696d2e
JG		 2953 return AV_PIX_FMT_GBRP;
2954 } else
2c541554 2955 return h->avctx->color_range == AVCOL_RANGE_JPEG ? AV_PIX_FMT_YUVJ444P
9e696d2e 2956 : AV_PIX_FMT_YUV444P;
e962bd08 2957 } else if (CHROMA422(h)) {
2c541554 2958 return h->avctx->color_range == AVCOL_RANGE_JPEG ? AV_PIX_FMT_YUVJ422P
9e696d2e
JG		 2959 : AV_PIX_FMT_YUV422P;
2960 } else {
2c541554
AK		 2961 return h->avctx->get_format(h->avctx, h->avctx->codec->pix_fmts ?
2962 h->avctx->codec->pix_fmts :
2963 h->avctx->color_range == AVCOL_RANGE_JPEG ?
d65522e8 2964 h264_hwaccel_pixfmt_list_jpeg_420 :
8d061989 2965 h264_hwaccel_pixfmt_list_420);
9e696d2e
JG		 2966 }
2967 break;
2968 default:
2c541554 2969 av_log(h->avctx, AV_LOG_ERROR,
9e696d2e
JG		 2970 "Unsupported bit depth: %d\n", h->sps.bit_depth_luma);
2971 return AVERROR_INVALIDDATA;
2972 }
2973}
2974
2975static int h264_slice_header_init(H264Context *h, int reinit)
2976{
2c541554
AK		 2977 int nb_slices = (HAVE_THREADS &&
2978 h->avctx->active_thread_type & FF_THREAD_SLICE) ?
2979 h->avctx->thread_count : 1;
2980 int i;
9e696d2e 2981
2c541554
AK		 2982 avcodec_set_dimensions(h->avctx, h->width, h->height);
2983 h->avctx->sample_aspect_ratio = h->sps.sar;
2984 av_assert0(h->avctx->sample_aspect_ratio.den);
2985 av_pix_fmt_get_chroma_sub_sample(h->avctx->pix_fmt,
2986 &h->chroma_x_shift, &h->chroma_y_shift);
9e696d2e
JG		 2987
2988 if (h->sps.timing_info_present_flag) {
2989 int64_t den = h->sps.time_scale;
2990 if (h->x264_build < 44U)
2991 den *= 2;
2c541554 2992 av_reduce(&h->avctx->time_base.num, &h->avctx->time_base.den,
9e696d2e
JG		 2993 h->sps.num_units_in_tick, den, 1 << 30);
2994 }
2995
2c541554 2996 h->avctx->hwaccel = ff_find_hwaccel(h->avctx->codec->id, h->avctx->pix_fmt);
9e696d2e 2997
2c541554 2998 if (reinit)
9e696d2e 2999 free_tables(h, 0);
2c541554 3000 h->first_field = 0;
9e696d2e
JG		 3001 h->prev_interlaced_frame = 1;
3002
3003 init_scan_tables(h);
3004 if (ff_h264_alloc_tables(h) < 0) {
2c541554 3005 av_log(h->avctx, AV_LOG_ERROR,
9e696d2e
JG		 3006 "Could not allocate memory for h264\n");
3007 return AVERROR(ENOMEM);
3008 }
3009
2c541554
AK		 3010 if (nb_slices > MAX_THREADS || (nb_slices > h->mb_height && h->mb_height)) {
3011 int max_slices;
3012 if (h->mb_height)
3013 max_slices = FFMIN(MAX_THREADS, h->mb_height);
3014 else
3015 max_slices = MAX_THREADS;
3016 av_log(h->avctx, AV_LOG_WARNING, "too many threads/slices (%d),"
3017 " reducing to %d\n", nb_slices, max_slices);
3018 nb_slices = max_slices;
3019 }
3020 h->slice_context_count = nb_slices;
3021
3022 if (!HAVE_THREADS || !(h->avctx->active_thread_type & FF_THREAD_SLICE)) {
9e696d2e 3023 if (context_init(h) < 0) {
2c541554 3024 av_log(h->avctx, AV_LOG_ERROR, "context_init() failed.\n");
9e696d2e
JG		 3025 return -1;
3026 }
3027 } else {
2c541554 3028 for (i = 1; i < h->slice_context_count; i++) {
9e696d2e 3029 H264Context *c;
2c541554
AK		 3030 c = h->thread_context[i] = av_mallocz(sizeof(H264Context));
3031 c->avctx = h->avctx;
3032 c->dsp = h->dsp;
3033 c->vdsp = h->vdsp;
9e696d2e 3034 c->h264dsp = h->h264dsp;
6bdb841b 3035 c->h264qpel = h->h264qpel;
4c51fe48 3036 c->h264chroma = h->h264chroma;
9e696d2e
JG		 3037 c->sps = h->sps;
3038 c->pps = h->pps;
3039 c->pixel_shift = h->pixel_shift;
2c541554
AK		 3040 c->width = h->width;
3041 c->height = h->height;
3042 c->linesize = h->linesize;
3043 c->uvlinesize = h->uvlinesize;
3044 c->chroma_x_shift = h->chroma_x_shift;
3045 c->chroma_y_shift = h->chroma_y_shift;
3046 c->qscale = h->qscale;
3047 c->droppable = h->droppable;
3048 c->data_partitioning = h->data_partitioning;
3049 c->low_delay = h->low_delay;
3050 c->mb_width = h->mb_width;
3051 c->mb_height = h->mb_height;
3052 c->mb_stride = h->mb_stride;
3053 c->mb_num = h->mb_num;
3054 c->flags = h->flags;
3055 c->workaround_bugs = h->workaround_bugs;
3056 c->pict_type = h->pict_type;
3057
9e696d2e
JG		 3058 init_scan_tables(c);
3059 clone_tables(c, h, i);
2c541554 3060 c->context_initialized = 1;
9e696d2e
JG		 3061 }
3062
2c541554 3063 for (i = 0; i < h->slice_context_count; i++)
9e696d2e 3064 if (context_init(h->thread_context[i]) < 0) {
2c541554 3065 av_log(h->avctx, AV_LOG_ERROR, "context_init() failed.\n");
9e696d2e
JG		 3066 return -1;
3067 }
3068 }
3069
2c541554
AK		 3070 h->context_initialized = 1;
3071
9e696d2e
JG		 3072 return 0;
3073}
3074
fe9a3fbe 3075/**
58c42af7 3076 * Decode a slice header.
efd29844 3077 * This will also call ff_MPV_common_init() and frame_start() as needed.
afebe2f7
AÖ		 3078 *
3079 * @param h h264context
e5d40372
DB		 3080 * @param h0 h264 master context (differs from 'h' when doing sliced based
3081 * parallel decoding)
afebe2f7 3082 *
d9526386 3083 * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
0da71265 3084 */
e5d40372
DB		 3085static int decode_slice_header(H264Context *h, H264Context *h0)
3086{
88e7a4d1 3087 unsigned int first_mb_in_slice;
ac658be5 3088 unsigned int pps_id;
072be3e8 3089 int num_ref_idx_active_override_flag, max_refs, ret;
41f5c62f 3090 unsigned int slice_type, tmp, i, j;
0bf79634 3091 int default_ref_list_done = 0;
ba0c8981 3092 int last_pic_structure, last_pic_droppable;
9e696d2e 3093 int needs_reinit = 0;
0da71265 3094
2c541554
AK		 3095 h->me.qpel_put = h->h264qpel.put_h264_qpel_pixels_tab;
3096 h->me.qpel_avg = h->h264qpel.avg_h264_qpel_pixels_tab;
cf653d08 3097
2c541554 3098 first_mb_in_slice = get_ue_golomb(&h->gb);
0da71265 3099
e5d40372 3100 if (first_mb_in_slice == 0) { // FIXME better field boundary detection
7fa00653 3101 if (h0->current_slice && FIELD_PICTURE(h)) {
6a9c8594 3102 field_end(h, 1);
d225a1e2
MN		 3103 }
3104
afebe2f7 3105 h0->current_slice = 0;
2c541554 3106 if (!h0->first_field) {
759001c5
AK		 3107 if (h->cur_pic_ptr && !h->droppable) {
3108 ff_thread_report_progress(&h->cur_pic_ptr->tf, INT_MAX,
2c541554 3109 h->picture_structure == PICT_BOTTOM_FIELD);
1e26a48f 3110 }
2c541554 3111 h->cur_pic_ptr = NULL;
1e26a48f 3112 }
66a4b2c1
MN		 3113 }
3114
2c541554 3115 slice_type = get_ue_golomb_31(&h->gb);
e5d40372 3116 if (slice_type > 9) {
2c541554 3117 av_log(h->avctx, AV_LOG_ERROR,
e5d40372 3118 "slice type too large (%d) at %d %d\n",
2c541554 3119 h->slice_type, h->mb_x, h->mb_y);
5175b937 3120 return -1;
0da71265 3121 }
e5d40372 3122 if (slice_type > 4) {
0bf79634 3123 slice_type -= 5;
e5d40372
DB		 3124 h->slice_type_fixed = 1;
3125 } else
3126 h->slice_type_fixed = 0;
115329f1 3127
e5d40372
DB		 3128 slice_type = golomb_to_pict_type[slice_type];
3129 if (slice_type == AV_PICTURE_TYPE_I ||
3130 (h0->current_slice != 0 && slice_type == h0->last_slice_type)) {
0bf79634
LLL		 3131 default_ref_list_done = 1;
3132 }
e5d40372
DB		 3133 h->slice_type = slice_type;
3134 h->slice_type_nos = slice_type & 3;
0bf79634 3135
e5d40372 3136 // to make a few old functions happy, it's wrong though
2c541554 3137 h->pict_type = h->slice_type;
115329f1 3138
2c541554 3139 pps_id = get_ue_golomb(&h->gb);
e5d40372 3140 if (pps_id >= MAX_PPS_COUNT) {
2c541554 3141 av_log(h->avctx, AV_LOG_ERROR, "pps_id out of range\n");
0da71265
MN		 3142 return -1;
3143 }
e5d40372 3144 if (!h0->pps_buffers[pps_id]) {
2c541554 3145 av_log(h->avctx, AV_LOG_ERROR,
e5d40372
DB		 3146 "non-existing PPS %u referenced\n",
3147 pps_id);
8b92b792
MN		 3148 return -1;
3149 }
e5d40372 3150 h->pps = *h0->pps_buffers[pps_id];
8b92b792 3151
e5d40372 3152 if (!h0->sps_buffers[h-&