You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

2879 lines
108KB

  1. /*
  2. * H.26L/H.264/AVC/JVT/14496-10/... decoder
  3. * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
  4. *
  5. * This file is part of FFmpeg.
  6. *
  7. * FFmpeg is free software; you can redistribute it and/or
  8. * modify it under the terms of the GNU Lesser General Public
  9. * License as published by the Free Software Foundation; either
  10. * version 2.1 of the License, or (at your option) any later version.
  11. *
  12. * FFmpeg is distributed in the hope that it will be useful,
  13. * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14. * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  15. * Lesser General Public License for more details.
  16. *
  17. * You should have received a copy of the GNU Lesser General Public
  18. * License along with FFmpeg; if not, write to the Free Software
  19. * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20. */
  21. /**
  22. * @file
  23. * H.264 / AVC / MPEG-4 part10 codec.
  24. * @author Michael Niedermayer <michaelni@gmx.at>
  25. */
  26. #include "libavutil/avassert.h"
  27. #include "libavutil/display.h"
  28. #include "libavutil/imgutils.h"
  29. #include "libavutil/stereo3d.h"
  30. #include "internal.h"
  31. #include "cabac.h"
  32. #include "cabac_functions.h"
  33. #include "error_resilience.h"
  34. #include "avcodec.h"
  35. #include "h264.h"
  36. #include "h264dec.h"
  37. #include "h264data.h"
  38. #include "h264chroma.h"
  39. #include "h264_mvpred.h"
  40. #include "h264_ps.h"
  41. #include "golomb.h"
  42. #include "mathops.h"
  43. #include "mpegutils.h"
  44. #include "mpegvideo.h"
  45. #include "rectangle.h"
  46. #include "thread.h"
  47. static const uint8_t field_scan[16+1] = {
  48. 0 + 0 * 4, 0 + 1 * 4, 1 + 0 * 4, 0 + 2 * 4,
  49. 0 + 3 * 4, 1 + 1 * 4, 1 + 2 * 4, 1 + 3 * 4,
  50. 2 + 0 * 4, 2 + 1 * 4, 2 + 2 * 4, 2 + 3 * 4,
  51. 3 + 0 * 4, 3 + 1 * 4, 3 + 2 * 4, 3 + 3 * 4,
  52. };
  53. static const uint8_t field_scan8x8[64+1] = {
  54. 0 + 0 * 8, 0 + 1 * 8, 0 + 2 * 8, 1 + 0 * 8,
  55. 1 + 1 * 8, 0 + 3 * 8, 0 + 4 * 8, 1 + 2 * 8,
  56. 2 + 0 * 8, 1 + 3 * 8, 0 + 5 * 8, 0 + 6 * 8,
  57. 0 + 7 * 8, 1 + 4 * 8, 2 + 1 * 8, 3 + 0 * 8,
  58. 2 + 2 * 8, 1 + 5 * 8, 1 + 6 * 8, 1 + 7 * 8,
  59. 2 + 3 * 8, 3 + 1 * 8, 4 + 0 * 8, 3 + 2 * 8,
  60. 2 + 4 * 8, 2 + 5 * 8, 2 + 6 * 8, 2 + 7 * 8,
  61. 3 + 3 * 8, 4 + 1 * 8, 5 + 0 * 8, 4 + 2 * 8,
  62. 3 + 4 * 8, 3 + 5 * 8, 3 + 6 * 8, 3 + 7 * 8,
  63. 4 + 3 * 8, 5 + 1 * 8, 6 + 0 * 8, 5 + 2 * 8,
  64. 4 + 4 * 8, 4 + 5 * 8, 4 + 6 * 8, 4 + 7 * 8,
  65. 5 + 3 * 8, 6 + 1 * 8, 6 + 2 * 8, 5 + 4 * 8,
  66. 5 + 5 * 8, 5 + 6 * 8, 5 + 7 * 8, 6 + 3 * 8,
  67. 7 + 0 * 8, 7 + 1 * 8, 6 + 4 * 8, 6 + 5 * 8,
  68. 6 + 6 * 8, 6 + 7 * 8, 7 + 2 * 8, 7 + 3 * 8,
  69. 7 + 4 * 8, 7 + 5 * 8, 7 + 6 * 8, 7 + 7 * 8,
  70. };
  71. static const uint8_t field_scan8x8_cavlc[64+1] = {
  72. 0 + 0 * 8, 1 + 1 * 8, 2 + 0 * 8, 0 + 7 * 8,
  73. 2 + 2 * 8, 2 + 3 * 8, 2 + 4 * 8, 3 + 3 * 8,
  74. 3 + 4 * 8, 4 + 3 * 8, 4 + 4 * 8, 5 + 3 * 8,
  75. 5 + 5 * 8, 7 + 0 * 8, 6 + 6 * 8, 7 + 4 * 8,
  76. 0 + 1 * 8, 0 + 3 * 8, 1 + 3 * 8, 1 + 4 * 8,
  77. 1 + 5 * 8, 3 + 1 * 8, 2 + 5 * 8, 4 + 1 * 8,
  78. 3 + 5 * 8, 5 + 1 * 8, 4 + 5 * 8, 6 + 1 * 8,
  79. 5 + 6 * 8, 7 + 1 * 8, 6 + 7 * 8, 7 + 5 * 8,
  80. 0 + 2 * 8, 0 + 4 * 8, 0 + 5 * 8, 2 + 1 * 8,
  81. 1 + 6 * 8, 4 + 0 * 8, 2 + 6 * 8, 5 + 0 * 8,
  82. 3 + 6 * 8, 6 + 0 * 8, 4 + 6 * 8, 6 + 2 * 8,
  83. 5 + 7 * 8, 6 + 4 * 8, 7 + 2 * 8, 7 + 6 * 8,
  84. 1 + 0 * 8, 1 + 2 * 8, 0 + 6 * 8, 3 + 0 * 8,
  85. 1 + 7 * 8, 3 + 2 * 8, 2 + 7 * 8, 4 + 2 * 8,
  86. 3 + 7 * 8, 5 + 2 * 8, 4 + 7 * 8, 5 + 4 * 8,
  87. 6 + 3 * 8, 6 + 5 * 8, 7 + 3 * 8, 7 + 7 * 8,
  88. };
  89. // zigzag_scan8x8_cavlc[i] = zigzag_scan8x8[(i/4) + 16*(i%4)]
  90. static const uint8_t zigzag_scan8x8_cavlc[64+1] = {
  91. 0 + 0 * 8, 1 + 1 * 8, 1 + 2 * 8, 2 + 2 * 8,
  92. 4 + 1 * 8, 0 + 5 * 8, 3 + 3 * 8, 7 + 0 * 8,
  93. 3 + 4 * 8, 1 + 7 * 8, 5 + 3 * 8, 6 + 3 * 8,
  94. 2 + 7 * 8, 6 + 4 * 8, 5 + 6 * 8, 7 + 5 * 8,
  95. 1 + 0 * 8, 2 + 0 * 8, 0 + 3 * 8, 3 + 1 * 8,
  96. 3 + 2 * 8, 0 + 6 * 8, 4 + 2 * 8, 6 + 1 * 8,
  97. 2 + 5 * 8, 2 + 6 * 8, 6 + 2 * 8, 5 + 4 * 8,
  98. 3 + 7 * 8, 7 + 3 * 8, 4 + 7 * 8, 7 + 6 * 8,
  99. 0 + 1 * 8, 3 + 0 * 8, 0 + 4 * 8, 4 + 0 * 8,
  100. 2 + 3 * 8, 1 + 5 * 8, 5 + 1 * 8, 5 + 2 * 8,
  101. 1 + 6 * 8, 3 + 5 * 8, 7 + 1 * 8, 4 + 5 * 8,
  102. 4 + 6 * 8, 7 + 4 * 8, 5 + 7 * 8, 6 + 7 * 8,
  103. 0 + 2 * 8, 2 + 1 * 8, 1 + 3 * 8, 5 + 0 * 8,
  104. 1 + 4 * 8, 2 + 4 * 8, 6 + 0 * 8, 4 + 3 * 8,
  105. 0 + 7 * 8, 4 + 4 * 8, 7 + 2 * 8, 3 + 6 * 8,
  106. 5 + 5 * 8, 6 + 5 * 8, 6 + 6 * 8, 7 + 7 * 8,
  107. };
  108. static void release_unused_pictures(H264Context *h, int remove_current)
  109. {
  110. int i;
  111. /* release non reference frames */
  112. for (i = 0; i < H264_MAX_PICTURE_COUNT; i++) {
  113. if (h->DPB[i].f->buf[0] && !h->DPB[i].reference &&
  114. (remove_current || &h->DPB[i] != h->cur_pic_ptr)) {
  115. ff_h264_unref_picture(h, &h->DPB[i]);
  116. }
  117. }
  118. }
  119. static int alloc_scratch_buffers(H264SliceContext *sl, int linesize)
  120. {
  121. const H264Context *h = sl->h264;
  122. int alloc_size = FFALIGN(FFABS(linesize) + 32, 32);
  123. av_fast_malloc(&sl->bipred_scratchpad, &sl->bipred_scratchpad_allocated, 16 * 6 * alloc_size);
  124. // edge emu needs blocksize + filter length - 1
  125. // (= 21x21 for H.264)
  126. av_fast_malloc(&sl->edge_emu_buffer, &sl->edge_emu_buffer_allocated, alloc_size * 2 * 21);
  127. av_fast_mallocz(&sl->top_borders[0], &sl->top_borders_allocated[0],
  128. h->mb_width * 16 * 3 * sizeof(uint8_t) * 2);
  129. av_fast_mallocz(&sl->top_borders[1], &sl->top_borders_allocated[1],
  130. h->mb_width * 16 * 3 * sizeof(uint8_t) * 2);
  131. if (!sl->bipred_scratchpad || !sl->edge_emu_buffer ||
  132. !sl->top_borders[0] || !sl->top_borders[1]) {
  133. av_freep(&sl->bipred_scratchpad);
  134. av_freep(&sl->edge_emu_buffer);
  135. av_freep(&sl->top_borders[0]);
  136. av_freep(&sl->top_borders[1]);
  137. sl->bipred_scratchpad_allocated = 0;
  138. sl->edge_emu_buffer_allocated = 0;
  139. sl->top_borders_allocated[0] = 0;
  140. sl->top_borders_allocated[1] = 0;
  141. return AVERROR(ENOMEM);
  142. }
  143. return 0;
  144. }
  145. static int init_table_pools(H264Context *h)
  146. {
  147. const int big_mb_num = h->mb_stride * (h->mb_height + 1) + 1;
  148. const int mb_array_size = h->mb_stride * h->mb_height;
  149. const int b4_stride = h->mb_width * 4 + 1;
  150. const int b4_array_size = b4_stride * h->mb_height * 4;
  151. h->qscale_table_pool = av_buffer_pool_init(big_mb_num + h->mb_stride,
  152. av_buffer_allocz);
  153. h->mb_type_pool = av_buffer_pool_init((big_mb_num + h->mb_stride) *
  154. sizeof(uint32_t), av_buffer_allocz);
  155. h->motion_val_pool = av_buffer_pool_init(2 * (b4_array_size + 4) *
  156. sizeof(int16_t), av_buffer_allocz);
  157. h->ref_index_pool = av_buffer_pool_init(4 * mb_array_size, av_buffer_allocz);
  158. if (!h->qscale_table_pool || !h->mb_type_pool || !h->motion_val_pool ||
  159. !h->ref_index_pool) {
  160. av_buffer_pool_uninit(&h->qscale_table_pool);
  161. av_buffer_pool_uninit(&h->mb_type_pool);
  162. av_buffer_pool_uninit(&h->motion_val_pool);
  163. av_buffer_pool_uninit(&h->ref_index_pool);
  164. return AVERROR(ENOMEM);
  165. }
  166. return 0;
  167. }
  168. static int alloc_picture(H264Context *h, H264Picture *pic)
  169. {
  170. int i, ret = 0;
  171. av_assert0(!pic->f->data[0]);
  172. pic->tf.f = pic->f;
  173. ret = ff_thread_get_buffer(h->avctx, &pic->tf, pic->reference ?
  174. AV_GET_BUFFER_FLAG_REF : 0);
  175. if (ret < 0)
  176. goto fail;
  177. if (h->avctx->hwaccel) {
  178. const AVHWAccel *hwaccel = h->avctx->hwaccel;
  179. av_assert0(!pic->hwaccel_picture_private);
  180. if (hwaccel->frame_priv_data_size) {
  181. pic->hwaccel_priv_buf = av_buffer_allocz(hwaccel->frame_priv_data_size);
  182. if (!pic->hwaccel_priv_buf)
  183. return AVERROR(ENOMEM);
  184. pic->hwaccel_picture_private = pic->hwaccel_priv_buf->data;
  185. }
  186. }
  187. if (CONFIG_GRAY && !h->avctx->hwaccel && h->flags & AV_CODEC_FLAG_GRAY && pic->f->data[2]) {
  188. int h_chroma_shift, v_chroma_shift;
  189. av_pix_fmt_get_chroma_sub_sample(pic->f->format,
  190. &h_chroma_shift, &v_chroma_shift);
  191. for(i=0; i<AV_CEIL_RSHIFT(pic->f->height, v_chroma_shift); i++) {
  192. memset(pic->f->data[1] + pic->f->linesize[1]*i,
  193. 0x80, AV_CEIL_RSHIFT(pic->f->width, h_chroma_shift));
  194. memset(pic->f->data[2] + pic->f->linesize[2]*i,
  195. 0x80, AV_CEIL_RSHIFT(pic->f->width, h_chroma_shift));
  196. }
  197. }
  198. if (!h->qscale_table_pool) {
  199. ret = init_table_pools(h);
  200. if (ret < 0)
  201. goto fail;
  202. }
  203. pic->qscale_table_buf = av_buffer_pool_get(h->qscale_table_pool);
  204. pic->mb_type_buf = av_buffer_pool_get(h->mb_type_pool);
  205. if (!pic->qscale_table_buf || !pic->mb_type_buf)
  206. goto fail;
  207. pic->mb_type = (uint32_t*)pic->mb_type_buf->data + 2 * h->mb_stride + 1;
  208. pic->qscale_table = pic->qscale_table_buf->data + 2 * h->mb_stride + 1;
  209. for (i = 0; i < 2; i++) {
  210. pic->motion_val_buf[i] = av_buffer_pool_get(h->motion_val_pool);
  211. pic->ref_index_buf[i] = av_buffer_pool_get(h->ref_index_pool);
  212. if (!pic->motion_val_buf[i] || !pic->ref_index_buf[i])
  213. goto fail;
  214. pic->motion_val[i] = (int16_t (*)[2])pic->motion_val_buf[i]->data + 4;
  215. pic->ref_index[i] = pic->ref_index_buf[i]->data;
  216. }
  217. pic->pps_buf = av_buffer_ref(h->ps.pps_ref);
  218. if (!pic->pps_buf)
  219. goto fail;
  220. pic->pps = (const PPS*)pic->pps_buf->data;
  221. pic->mb_width = h->mb_width;
  222. pic->mb_height = h->mb_height;
  223. pic->mb_stride = h->mb_stride;
  224. return 0;
  225. fail:
  226. ff_h264_unref_picture(h, pic);
  227. return (ret < 0) ? ret : AVERROR(ENOMEM);
  228. }
  229. static int find_unused_picture(H264Context *h)
  230. {
  231. int i;
  232. for (i = 0; i < H264_MAX_PICTURE_COUNT; i++) {
  233. if (!h->DPB[i].f->buf[0])
  234. return i;
  235. }
  236. return AVERROR_INVALIDDATA;
  237. }
  238. #define IN_RANGE(a, b, size) (((void*)(a) >= (void*)(b)) && ((void*)(a) < (void*)((b) + (size))))
  239. #define REBASE_PICTURE(pic, new_ctx, old_ctx) \
  240. (((pic) && (pic) >= (old_ctx)->DPB && \
  241. (pic) < (old_ctx)->DPB + H264_MAX_PICTURE_COUNT) ? \
  242. &(new_ctx)->DPB[(pic) - (old_ctx)->DPB] : NULL)
  243. static void copy_picture_range(H264Picture **to, H264Picture **from, int count,
  244. H264Context *new_base,
  245. H264Context *old_base)
  246. {
  247. int i;
  248. for (i = 0; i < count; i++) {
  249. av_assert1(!from[i] ||
  250. IN_RANGE(from[i], old_base, 1) ||
  251. IN_RANGE(from[i], old_base->DPB, H264_MAX_PICTURE_COUNT));
  252. to[i] = REBASE_PICTURE(from[i], new_base, old_base);
  253. }
  254. }
  255. static int h264_slice_header_init(H264Context *h);
  256. int ff_h264_update_thread_context(AVCodecContext *dst,
  257. const AVCodecContext *src)
  258. {
  259. H264Context *h = dst->priv_data, *h1 = src->priv_data;
  260. int inited = h->context_initialized, err = 0;
  261. int need_reinit = 0;
  262. int i, ret;
  263. if (dst == src)
  264. return 0;
  265. if (inited && !h1->ps.sps)
  266. return AVERROR_INVALIDDATA;
  267. if (inited &&
  268. (h->width != h1->width ||
  269. h->height != h1->height ||
  270. h->mb_width != h1->mb_width ||
  271. h->mb_height != h1->mb_height ||
  272. !h->ps.sps ||
  273. h->ps.sps->bit_depth_luma != h1->ps.sps->bit_depth_luma ||
  274. h->ps.sps->chroma_format_idc != h1->ps.sps->chroma_format_idc ||
  275. h->ps.sps->colorspace != h1->ps.sps->colorspace)) {
  276. need_reinit = 1;
  277. }
  278. /* copy block_offset since frame_start may not be called */
  279. memcpy(h->block_offset, h1->block_offset, sizeof(h->block_offset));
  280. // SPS/PPS
  281. for (i = 0; i < FF_ARRAY_ELEMS(h->ps.sps_list); i++) {
  282. ret = av_buffer_replace(&h->ps.sps_list[i], h1->ps.sps_list[i]);
  283. if (ret < 0)
  284. return ret;
  285. }
  286. for (i = 0; i < FF_ARRAY_ELEMS(h->ps.pps_list); i++) {
  287. ret = av_buffer_replace(&h->ps.pps_list[i], h1->ps.pps_list[i]);
  288. if (ret < 0)
  289. return ret;
  290. }
  291. ret = av_buffer_replace(&h->ps.pps_ref, h1->ps.pps_ref);
  292. if (ret < 0)
  293. return ret;
  294. h->ps.pps = NULL;
  295. h->ps.sps = NULL;
  296. if (h1->ps.pps_ref) {
  297. h->ps.pps = (const PPS*)h->ps.pps_ref->data;
  298. h->ps.sps = h->ps.pps->sps;
  299. }
  300. if (need_reinit || !inited) {
  301. h->width = h1->width;
  302. h->height = h1->height;
  303. h->mb_height = h1->mb_height;
  304. h->mb_width = h1->mb_width;
  305. h->mb_num = h1->mb_num;
  306. h->mb_stride = h1->mb_stride;
  307. h->b_stride = h1->b_stride;
  308. h->x264_build = h1->x264_build;
  309. if (h->context_initialized || h1->context_initialized) {
  310. if ((err = h264_slice_header_init(h)) < 0) {
  311. av_log(h->avctx, AV_LOG_ERROR, "h264_slice_header_init() failed");
  312. return err;
  313. }
  314. }
  315. /* copy block_offset since frame_start may not be called */
  316. memcpy(h->block_offset, h1->block_offset, sizeof(h->block_offset));
  317. }
  318. h->avctx->coded_height = h1->avctx->coded_height;
  319. h->avctx->coded_width = h1->avctx->coded_width;
  320. h->avctx->width = h1->avctx->width;
  321. h->avctx->height = h1->avctx->height;
  322. h->width_from_caller = h1->width_from_caller;
  323. h->height_from_caller = h1->height_from_caller;
  324. h->coded_picture_number = h1->coded_picture_number;
  325. h->first_field = h1->first_field;
  326. h->picture_structure = h1->picture_structure;
  327. h->mb_aff_frame = h1->mb_aff_frame;
  328. h->droppable = h1->droppable;
  329. for (i = 0; i < H264_MAX_PICTURE_COUNT; i++) {
  330. ff_h264_unref_picture(h, &h->DPB[i]);
  331. if (h1->DPB[i].f->buf[0] &&
  332. (ret = ff_h264_ref_picture(h, &h->DPB[i], &h1->DPB[i])) < 0)
  333. return ret;
  334. }
  335. h->cur_pic_ptr = REBASE_PICTURE(h1->cur_pic_ptr, h, h1);
  336. ff_h264_unref_picture(h, &h->cur_pic);
  337. if (h1->cur_pic.f->buf[0]) {
  338. ret = ff_h264_ref_picture(h, &h->cur_pic, &h1->cur_pic);
  339. if (ret < 0)
  340. return ret;
  341. }
  342. h->enable_er = h1->enable_er;
  343. h->workaround_bugs = h1->workaround_bugs;
  344. h->droppable = h1->droppable;
  345. // extradata/NAL handling
  346. h->is_avc = h1->is_avc;
  347. h->nal_length_size = h1->nal_length_size;
  348. memcpy(&h->poc, &h1->poc, sizeof(h->poc));
  349. memcpy(h->short_ref, h1->short_ref, sizeof(h->short_ref));
  350. memcpy(h->long_ref, h1->long_ref, sizeof(h->long_ref));
  351. memcpy(h->delayed_pic, h1->delayed_pic, sizeof(h->delayed_pic));
  352. memcpy(h->last_pocs, h1->last_pocs, sizeof(h->last_pocs));
  353. h->next_output_pic = h1->next_output_pic;
  354. h->next_outputed_poc = h1->next_outputed_poc;
  355. memcpy(h->mmco, h1->mmco, sizeof(h->mmco));
  356. h->nb_mmco = h1->nb_mmco;
  357. h->mmco_reset = h1->mmco_reset;
  358. h->explicit_ref_marking = h1->explicit_ref_marking;
  359. h->long_ref_count = h1->long_ref_count;
  360. h->short_ref_count = h1->short_ref_count;
  361. copy_picture_range(h->short_ref, h1->short_ref, 32, h, h1);
  362. copy_picture_range(h->long_ref, h1->long_ref, 32, h, h1);
  363. copy_picture_range(h->delayed_pic, h1->delayed_pic,
  364. MAX_DELAYED_PIC_COUNT + 2, h, h1);
  365. h->frame_recovered = h1->frame_recovered;
  366. ret = av_buffer_replace(&h->sei.a53_caption.buf_ref, h1->sei.a53_caption.buf_ref);
  367. if (ret < 0)
  368. return ret;
  369. for (i = 0; i < h->sei.unregistered.nb_buf_ref; i++)
  370. av_buffer_unref(&h->sei.unregistered.buf_ref[i]);
  371. h->sei.unregistered.nb_buf_ref = 0;
  372. if (h1->sei.unregistered.nb_buf_ref) {
  373. ret = av_reallocp_array(&h->sei.unregistered.buf_ref,
  374. h1->sei.unregistered.nb_buf_ref,
  375. sizeof(*h->sei.unregistered.buf_ref));
  376. if (ret < 0)
  377. return ret;
  378. for (i = 0; i < h1->sei.unregistered.nb_buf_ref; i++) {
  379. h->sei.unregistered.buf_ref[i] = av_buffer_ref(h1->sei.unregistered.buf_ref[i]);
  380. if (!h->sei.unregistered.buf_ref[i])
  381. return AVERROR(ENOMEM);
  382. h->sei.unregistered.nb_buf_ref++;
  383. }
  384. }
  385. h->sei.unregistered.x264_build = h1->sei.unregistered.x264_build;
  386. if (!h->cur_pic_ptr)
  387. return 0;
  388. if (!h->droppable) {
  389. err = ff_h264_execute_ref_pic_marking(h);
  390. h->poc.prev_poc_msb = h->poc.poc_msb;
  391. h->poc.prev_poc_lsb = h->poc.poc_lsb;
  392. }
  393. h->poc.prev_frame_num_offset = h->poc.frame_num_offset;
  394. h->poc.prev_frame_num = h->poc.frame_num;
  395. h->recovery_frame = h1->recovery_frame;
  396. return err;
  397. }
  398. static int h264_frame_start(H264Context *h)
  399. {
  400. H264Picture *pic;
  401. int i, ret;
  402. const int pixel_shift = h->pixel_shift;
  403. if (!ff_thread_can_start_frame(h->avctx)) {
  404. av_log(h->avctx, AV_LOG_ERROR, "Attempt to start a frame outside SETUP state\n");
  405. return -1;
  406. }
  407. release_unused_pictures(h, 1);
  408. h->cur_pic_ptr = NULL;
  409. i = find_unused_picture(h);
  410. if (i < 0) {
  411. av_log(h->avctx, AV_LOG_ERROR, "no frame buffer available\n");
  412. return i;
  413. }
  414. pic = &h->DPB[i];
  415. pic->reference = h->droppable ? 0 : h->picture_structure;
  416. pic->f->coded_picture_number = h->coded_picture_number++;
  417. pic->field_picture = h->picture_structure != PICT_FRAME;
  418. pic->frame_num = h->poc.frame_num;
  419. /*
  420. * Zero key_frame here; IDR markings per slice in frame or fields are ORed
  421. * in later.
  422. * See decode_nal_units().
  423. */
  424. pic->f->key_frame = 0;
  425. pic->mmco_reset = 0;
  426. pic->recovered = 0;
  427. pic->invalid_gap = 0;
  428. pic->sei_recovery_frame_cnt = h->sei.recovery_point.recovery_frame_cnt;
  429. pic->f->pict_type = h->slice_ctx[0].slice_type;
  430. pic->f->crop_left = h->crop_left;
  431. pic->f->crop_right = h->crop_right;
  432. pic->f->crop_top = h->crop_top;
  433. pic->f->crop_bottom = h->crop_bottom;
  434. if ((ret = alloc_picture(h, pic)) < 0)
  435. return ret;
  436. h->cur_pic_ptr = pic;
  437. ff_h264_unref_picture(h, &h->cur_pic);
  438. if (CONFIG_ERROR_RESILIENCE) {
  439. ff_h264_set_erpic(&h->slice_ctx[0].er.cur_pic, NULL);
  440. }
  441. if ((ret = ff_h264_ref_picture(h, &h->cur_pic, h->cur_pic_ptr)) < 0)
  442. return ret;
  443. for (i = 0; i < h->nb_slice_ctx; i++) {
  444. h->slice_ctx[i].linesize = h->cur_pic_ptr->f->linesize[0];
  445. h->slice_ctx[i].uvlinesize = h->cur_pic_ptr->f->linesize[1];
  446. }
  447. if (CONFIG_ERROR_RESILIENCE && h->enable_er) {
  448. ff_er_frame_start(&h->slice_ctx[0].er);
  449. ff_h264_set_erpic(&h->slice_ctx[0].er.last_pic, NULL);
  450. ff_h264_set_erpic(&h->slice_ctx[0].er.next_pic, NULL);
  451. }
  452. for (i = 0; i < 16; i++) {
  453. h->block_offset[i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * pic->f->linesize[0] * ((scan8[i] - scan8[0]) >> 3);
  454. h->block_offset[48 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * pic->f->linesize[0] * ((scan8[i] - scan8[0]) >> 3);
  455. }
  456. for (i = 0; i < 16; i++) {
  457. h->block_offset[16 + i] =
  458. h->block_offset[32 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * pic->f->linesize[1] * ((scan8[i] - scan8[0]) >> 3);
  459. h->block_offset[48 + 16 + i] =
  460. h->block_offset[48 + 32 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * pic->f->linesize[1] * ((scan8[i] - scan8[0]) >> 3);
  461. }
  462. /* We mark the current picture as non-reference after allocating it, so
  463. * that if we break out due to an error it can be released automatically
  464. * in the next ff_mpv_frame_start().
  465. */
  466. h->cur_pic_ptr->reference = 0;
  467. h->cur_pic_ptr->field_poc[0] = h->cur_pic_ptr->field_poc[1] = INT_MAX;
  468. h->next_output_pic = NULL;
  469. h->postpone_filter = 0;
  470. h->mb_aff_frame = h->ps.sps->mb_aff && (h->picture_structure == PICT_FRAME);
  471. if (h->sei.unregistered.x264_build >= 0)
  472. h->x264_build = h->sei.unregistered.x264_build;
  473. assert(h->cur_pic_ptr->long_ref == 0);
  474. return 0;
  475. }
  476. static av_always_inline void backup_mb_border(const H264Context *h, H264SliceContext *sl,
  477. uint8_t *src_y,
  478. uint8_t *src_cb, uint8_t *src_cr,
  479. int linesize, int uvlinesize,
  480. int simple)
  481. {
  482. uint8_t *top_border;
  483. int top_idx = 1;
  484. const int pixel_shift = h->pixel_shift;
  485. int chroma444 = CHROMA444(h);
  486. int chroma422 = CHROMA422(h);
  487. src_y -= linesize;
  488. src_cb -= uvlinesize;
  489. src_cr -= uvlinesize;
  490. if (!simple && FRAME_MBAFF(h)) {
  491. if (sl->mb_y & 1) {
  492. if (!MB_MBAFF(sl)) {
  493. top_border = sl->top_borders[0][sl->mb_x];
  494. AV_COPY128(top_border, src_y + 15 * linesize);
  495. if (pixel_shift)
  496. AV_COPY128(top_border + 16, src_y + 15 * linesize + 16);
  497. if (simple || !CONFIG_GRAY || !(h->flags & AV_CODEC_FLAG_GRAY)) {
  498. if (chroma444) {
  499. if (pixel_shift) {
  500. AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
  501. AV_COPY128(top_border + 48, src_cb + 15 * uvlinesize + 16);
  502. AV_COPY128(top_border + 64, src_cr + 15 * uvlinesize);
  503. AV_COPY128(top_border + 80, src_cr + 15 * uvlinesize + 16);
  504. } else {
  505. AV_COPY128(top_border + 16, src_cb + 15 * uvlinesize);
  506. AV_COPY128(top_border + 32, src_cr + 15 * uvlinesize);
  507. }
  508. } else if (chroma422) {
  509. if (pixel_shift) {
  510. AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
  511. AV_COPY128(top_border + 48, src_cr + 15 * uvlinesize);
  512. } else {
  513. AV_COPY64(top_border + 16, src_cb + 15 * uvlinesize);
  514. AV_COPY64(top_border + 24, src_cr + 15 * uvlinesize);
  515. }
  516. } else {
  517. if (pixel_shift) {
  518. AV_COPY128(top_border + 32, src_cb + 7 * uvlinesize);
  519. AV_COPY128(top_border + 48, src_cr + 7 * uvlinesize);
  520. } else {
  521. AV_COPY64(top_border + 16, src_cb + 7 * uvlinesize);
  522. AV_COPY64(top_border + 24, src_cr + 7 * uvlinesize);
  523. }
  524. }
  525. }
  526. }
  527. } else if (MB_MBAFF(sl)) {
  528. top_idx = 0;
  529. } else
  530. return;
  531. }
  532. top_border = sl->top_borders[top_idx][sl->mb_x];
  533. /* There are two lines saved, the line above the top macroblock
  534. * of a pair, and the line above the bottom macroblock. */
  535. AV_COPY128(top_border, src_y + 16 * linesize);
  536. if (pixel_shift)
  537. AV_COPY128(top_border + 16, src_y + 16 * linesize + 16);
  538. if (simple || !CONFIG_GRAY || !(h->flags & AV_CODEC_FLAG_GRAY)) {
  539. if (chroma444) {
  540. if (pixel_shift) {
  541. AV_COPY128(top_border + 32, src_cb + 16 * linesize);
  542. AV_COPY128(top_border + 48, src_cb + 16 * linesize + 16);
  543. AV_COPY128(top_border + 64, src_cr + 16 * linesize);
  544. AV_COPY128(top_border + 80, src_cr + 16 * linesize + 16);
  545. } else {
  546. AV_COPY128(top_border + 16, src_cb + 16 * linesize);
  547. AV_COPY128(top_border + 32, src_cr + 16 * linesize);
  548. }
  549. } else if (chroma422) {
  550. if (pixel_shift) {
  551. AV_COPY128(top_border + 32, src_cb + 16 * uvlinesize);
  552. AV_COPY128(top_border + 48, src_cr + 16 * uvlinesize);
  553. } else {
  554. AV_COPY64(top_border + 16, src_cb + 16 * uvlinesize);
  555. AV_COPY64(top_border + 24, src_cr + 16 * uvlinesize);
  556. }
  557. } else {
  558. if (pixel_shift) {
  559. AV_COPY128(top_border + 32, src_cb + 8 * uvlinesize);
  560. AV_COPY128(top_border + 48, src_cr + 8 * uvlinesize);
  561. } else {
  562. AV_COPY64(top_border + 16, src_cb + 8 * uvlinesize);
  563. AV_COPY64(top_border + 24, src_cr + 8 * uvlinesize);
  564. }
  565. }
  566. }
  567. }
  568. /**
  569. * Initialize implicit_weight table.
  570. * @param field 0/1 initialize the weight for interlaced MBAFF
  571. * -1 initializes the rest
  572. */
  573. static void implicit_weight_table(const H264Context *h, H264SliceContext *sl, int field)
  574. {
  575. int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
  576. for (i = 0; i < 2; i++) {
  577. sl->pwt.luma_weight_flag[i] = 0;
  578. sl->pwt.chroma_weight_flag[i] = 0;
  579. }
  580. if (field < 0) {
  581. if (h->picture_structure == PICT_FRAME) {
  582. cur_poc = h->cur_pic_ptr->poc;
  583. } else {
  584. cur_poc = h->cur_pic_ptr->field_poc[h->picture_structure - 1];
  585. }
  586. if (sl->ref_count[0] == 1 && sl->ref_count[1] == 1 && !FRAME_MBAFF(h) &&
  587. sl->ref_list[0][0].poc + (int64_t)sl->ref_list[1][0].poc == 2LL * cur_poc) {
  588. sl->pwt.use_weight = 0;
  589. sl->pwt.use_weight_chroma = 0;
  590. return;
  591. }
  592. ref_start = 0;
  593. ref_count0 = sl->ref_count[0];
  594. ref_count1 = sl->ref_count[1];
  595. } else {
  596. cur_poc = h->cur_pic_ptr->field_poc[field];
  597. ref_start = 16;
  598. ref_count0 = 16 + 2 * sl->ref_count[0];
  599. ref_count1 = 16 + 2 * sl->ref_count[1];
  600. }
  601. sl->pwt.use_weight = 2;
  602. sl->pwt.use_weight_chroma = 2;
  603. sl->pwt.luma_log2_weight_denom = 5;
  604. sl->pwt.chroma_log2_weight_denom = 5;
  605. for (ref0 = ref_start; ref0 < ref_count0; ref0++) {
  606. int64_t poc0 = sl->ref_list[0][ref0].poc;
  607. for (ref1 = ref_start; ref1 < ref_count1; ref1++) {
  608. int w = 32;
  609. if (!sl->ref_list[0][ref0].parent->long_ref && !sl->ref_list[1][ref1].parent->long_ref) {
  610. int poc1 = sl->ref_list[1][ref1].poc;
  611. int td = av_clip_int8(poc1 - poc0);
  612. if (td) {
  613. int tb = av_clip_int8(cur_poc - poc0);
  614. int tx = (16384 + (FFABS(td) >> 1)) / td;
  615. int dist_scale_factor = (tb * tx + 32) >> 8;
  616. if (dist_scale_factor >= -64 && dist_scale_factor <= 128)
  617. w = 64 - dist_scale_factor;
  618. }
  619. }
  620. if (field < 0) {
  621. sl->pwt.implicit_weight[ref0][ref1][0] =
  622. sl->pwt.implicit_weight[ref0][ref1][1] = w;
  623. } else {
  624. sl->pwt.implicit_weight[ref0][ref1][field] = w;
  625. }
  626. }
  627. }
  628. }
  629. /**
  630. * initialize scan tables
  631. */
  632. static void init_scan_tables(H264Context *h)
  633. {
  634. int i;
  635. for (i = 0; i < 16; i++) {
  636. #define TRANSPOSE(x) ((x) >> 2) | (((x) << 2) & 0xF)
  637. h->zigzag_scan[i] = TRANSPOSE(ff_zigzag_scan[i]);
  638. h->field_scan[i] = TRANSPOSE(field_scan[i]);
  639. #undef TRANSPOSE
  640. }
  641. for (i = 0; i < 64; i++) {
  642. #define TRANSPOSE(x) ((x) >> 3) | (((x) & 7) << 3)
  643. h->zigzag_scan8x8[i] = TRANSPOSE(ff_zigzag_direct[i]);
  644. h->zigzag_scan8x8_cavlc[i] = TRANSPOSE(zigzag_scan8x8_cavlc[i]);
  645. h->field_scan8x8[i] = TRANSPOSE(field_scan8x8[i]);
  646. h->field_scan8x8_cavlc[i] = TRANSPOSE(field_scan8x8_cavlc[i]);
  647. #undef TRANSPOSE
  648. }
  649. if (h->ps.sps->transform_bypass) { // FIXME same ugly
  650. memcpy(h->zigzag_scan_q0 , ff_zigzag_scan , sizeof(h->zigzag_scan_q0 ));
  651. memcpy(h->zigzag_scan8x8_q0 , ff_zigzag_direct , sizeof(h->zigzag_scan8x8_q0 ));
  652. memcpy(h->zigzag_scan8x8_cavlc_q0 , zigzag_scan8x8_cavlc , sizeof(h->zigzag_scan8x8_cavlc_q0));
  653. memcpy(h->field_scan_q0 , field_scan , sizeof(h->field_scan_q0 ));
  654. memcpy(h->field_scan8x8_q0 , field_scan8x8 , sizeof(h->field_scan8x8_q0 ));
  655. memcpy(h->field_scan8x8_cavlc_q0 , field_scan8x8_cavlc , sizeof(h->field_scan8x8_cavlc_q0 ));
  656. } else {
  657. memcpy(h->zigzag_scan_q0 , h->zigzag_scan , sizeof(h->zigzag_scan_q0 ));
  658. memcpy(h->zigzag_scan8x8_q0 , h->zigzag_scan8x8 , sizeof(h->zigzag_scan8x8_q0 ));
  659. memcpy(h->zigzag_scan8x8_cavlc_q0 , h->zigzag_scan8x8_cavlc , sizeof(h->zigzag_scan8x8_cavlc_q0));
  660. memcpy(h->field_scan_q0 , h->field_scan , sizeof(h->field_scan_q0 ));
  661. memcpy(h->field_scan8x8_q0 , h->field_scan8x8 , sizeof(h->field_scan8x8_q0 ));
  662. memcpy(h->field_scan8x8_cavlc_q0 , h->field_scan8x8_cavlc , sizeof(h->field_scan8x8_cavlc_q0 ));
  663. }
  664. }
  665. static enum AVPixelFormat get_pixel_format(H264Context *h, int force_callback)
  666. {
  667. #define HWACCEL_MAX (CONFIG_H264_DXVA2_HWACCEL + \
  668. (CONFIG_H264_D3D11VA_HWACCEL * 2) + \
  669. CONFIG_H264_NVDEC_HWACCEL + \
  670. CONFIG_H264_VAAPI_HWACCEL + \
  671. CONFIG_H264_VIDEOTOOLBOX_HWACCEL + \
  672. CONFIG_H264_VDPAU_HWACCEL)
  673. enum AVPixelFormat pix_fmts[HWACCEL_MAX + 2], *fmt = pix_fmts;
  674. const enum AVPixelFormat *choices = pix_fmts;
  675. int i;
  676. switch (h->ps.sps->bit_depth_luma) {
  677. case 9:
  678. if (CHROMA444(h)) {
  679. if (h->avctx->colorspace == AVCOL_SPC_RGB) {
  680. *fmt++ = AV_PIX_FMT_GBRP9;
  681. } else
  682. *fmt++ = AV_PIX_FMT_YUV444P9;
  683. } else if (CHROMA422(h))
  684. *fmt++ = AV_PIX_FMT_YUV422P9;
  685. else
  686. *fmt++ = AV_PIX_FMT_YUV420P9;
  687. break;
  688. case 10:
  689. if (CHROMA444(h)) {
  690. if (h->avctx->colorspace == AVCOL_SPC_RGB) {
  691. *fmt++ = AV_PIX_FMT_GBRP10;
  692. } else
  693. *fmt++ = AV_PIX_FMT_YUV444P10;
  694. } else if (CHROMA422(h))
  695. *fmt++ = AV_PIX_FMT_YUV422P10;
  696. else
  697. *fmt++ = AV_PIX_FMT_YUV420P10;
  698. break;
  699. case 12:
  700. if (CHROMA444(h)) {
  701. if (h->avctx->colorspace == AVCOL_SPC_RGB) {
  702. *fmt++ = AV_PIX_FMT_GBRP12;
  703. } else
  704. *fmt++ = AV_PIX_FMT_YUV444P12;
  705. } else if (CHROMA422(h))
  706. *fmt++ = AV_PIX_FMT_YUV422P12;
  707. else
  708. *fmt++ = AV_PIX_FMT_YUV420P12;
  709. break;
  710. case 14:
  711. if (CHROMA444(h)) {
  712. if (h->avctx->colorspace == AVCOL_SPC_RGB) {
  713. *fmt++ = AV_PIX_FMT_GBRP14;
  714. } else
  715. *fmt++ = AV_PIX_FMT_YUV444P14;
  716. } else if (CHROMA422(h))
  717. *fmt++ = AV_PIX_FMT_YUV422P14;
  718. else
  719. *fmt++ = AV_PIX_FMT_YUV420P14;
  720. break;
  721. case 8:
  722. #if CONFIG_H264_VDPAU_HWACCEL
  723. *fmt++ = AV_PIX_FMT_VDPAU;
  724. #endif
  725. #if CONFIG_H264_NVDEC_HWACCEL
  726. *fmt++ = AV_PIX_FMT_CUDA;
  727. #endif
  728. if (CHROMA444(h)) {
  729. if (h->avctx->colorspace == AVCOL_SPC_RGB)
  730. *fmt++ = AV_PIX_FMT_GBRP;
  731. else if (h->avctx->color_range == AVCOL_RANGE_JPEG)
  732. *fmt++ = AV_PIX_FMT_YUVJ444P;
  733. else
  734. *fmt++ = AV_PIX_FMT_YUV444P;
  735. } else if (CHROMA422(h)) {
  736. if (h->avctx->color_range == AVCOL_RANGE_JPEG)
  737. *fmt++ = AV_PIX_FMT_YUVJ422P;
  738. else
  739. *fmt++ = AV_PIX_FMT_YUV422P;
  740. } else {
  741. #if CONFIG_H264_DXVA2_HWACCEL
  742. *fmt++ = AV_PIX_FMT_DXVA2_VLD;
  743. #endif
  744. #if CONFIG_H264_D3D11VA_HWACCEL
  745. *fmt++ = AV_PIX_FMT_D3D11VA_VLD;
  746. *fmt++ = AV_PIX_FMT_D3D11;
  747. #endif
  748. #if CONFIG_H264_VAAPI_HWACCEL
  749. *fmt++ = AV_PIX_FMT_VAAPI;
  750. #endif
  751. #if CONFIG_H264_VIDEOTOOLBOX_HWACCEL
  752. *fmt++ = AV_PIX_FMT_VIDEOTOOLBOX;
  753. #endif
  754. if (h->avctx->codec->pix_fmts)
  755. choices = h->avctx->codec->pix_fmts;
  756. else if (h->avctx->color_range == AVCOL_RANGE_JPEG)
  757. *fmt++ = AV_PIX_FMT_YUVJ420P;
  758. else
  759. *fmt++ = AV_PIX_FMT_YUV420P;
  760. }
  761. break;
  762. default:
  763. av_log(h->avctx, AV_LOG_ERROR,
  764. "Unsupported bit depth %d\n", h->ps.sps->bit_depth_luma);
  765. return AVERROR_INVALIDDATA;
  766. }
  767. *fmt = AV_PIX_FMT_NONE;
  768. for (i=0; choices[i] != AV_PIX_FMT_NONE; i++)
  769. if (choices[i] == h->avctx->pix_fmt && !force_callback)
  770. return choices[i];
  771. return ff_thread_get_format(h->avctx, choices);
  772. }
  773. /* export coded and cropped frame dimensions to AVCodecContext */
  774. static void init_dimensions(H264Context *h)
  775. {
  776. const SPS *sps = (const SPS*)h->ps.sps;
  777. int cr = sps->crop_right;
  778. int cl = sps->crop_left;
  779. int ct = sps->crop_top;
  780. int cb = sps->crop_bottom;
  781. int width = h->width - (cr + cl);
  782. int height = h->height - (ct + cb);
  783. av_assert0(sps->crop_right + sps->crop_left < (unsigned)h->width);
  784. av_assert0(sps->crop_top + sps->crop_bottom < (unsigned)h->height);
  785. /* handle container cropping */
  786. if (h->width_from_caller > 0 && h->height_from_caller > 0 &&
  787. !sps->crop_top && !sps->crop_left &&
  788. FFALIGN(h->width_from_caller, 16) == FFALIGN(width, 16) &&
  789. FFALIGN(h->height_from_caller, 16) == FFALIGN(height, 16) &&
  790. h->width_from_caller <= width &&
  791. h->height_from_caller <= height) {
  792. width = h->width_from_caller;
  793. height = h->height_from_caller;
  794. cl = 0;
  795. ct = 0;
  796. cr = h->width - width;
  797. cb = h->height - height;
  798. } else {
  799. h->width_from_caller = 0;
  800. h->height_from_caller = 0;
  801. }
  802. h->avctx->coded_width = h->width;
  803. h->avctx->coded_height = h->height;
  804. h->avctx->width = width;
  805. h->avctx->height = height;
  806. h->crop_right = cr;
  807. h->crop_left = cl;
  808. h->crop_top = ct;
  809. h->crop_bottom = cb;
  810. }
  811. static int h264_slice_header_init(H264Context *h)
  812. {
  813. const SPS *sps = h->ps.sps;
  814. int i, ret;
  815. if (!sps) {
  816. ret = AVERROR_INVALIDDATA;
  817. goto fail;
  818. }
  819. ff_set_sar(h->avctx, sps->sar);
  820. av_pix_fmt_get_chroma_sub_sample(h->avctx->pix_fmt,
  821. &h->chroma_x_shift, &h->chroma_y_shift);
  822. if (sps->timing_info_present_flag) {
  823. int64_t den = sps->time_scale;
  824. if (h->x264_build < 44U)
  825. den *= 2;
  826. av_reduce(&h->avctx->framerate.den, &h->avctx->framerate.num,
  827. sps->num_units_in_tick * h->avctx->ticks_per_frame, den, 1 << 30);
  828. }
  829. ff_h264_free_tables(h);
  830. h->first_field = 0;
  831. h->prev_interlaced_frame = 1;
  832. init_scan_tables(h);
  833. ret = ff_h264_alloc_tables(h);
  834. if (ret < 0) {
  835. av_log(h->avctx, AV_LOG_ERROR, "Could not allocate memory\n");
  836. goto fail;
  837. }
  838. if (sps->bit_depth_luma < 8 || sps->bit_depth_luma > 14 ||
  839. sps->bit_depth_luma == 11 || sps->bit_depth_luma == 13
  840. ) {
  841. av_log(h->avctx, AV_LOG_ERROR, "Unsupported bit depth %d\n",
  842. sps->bit_depth_luma);
  843. ret = AVERROR_INVALIDDATA;
  844. goto fail;
  845. }
  846. h->cur_bit_depth_luma =
  847. h->avctx->bits_per_raw_sample = sps->bit_depth_luma;
  848. h->cur_chroma_format_idc = sps->chroma_format_idc;
  849. h->pixel_shift = sps->bit_depth_luma > 8;
  850. h->chroma_format_idc = sps->chroma_format_idc;
  851. h->bit_depth_luma = sps->bit_depth_luma;
  852. ff_h264dsp_init(&h->h264dsp, sps->bit_depth_luma,
  853. sps->chroma_format_idc);
  854. ff_h264chroma_init(&h->h264chroma, sps->bit_depth_chroma);
  855. ff_h264qpel_init(&h->h264qpel, sps->bit_depth_luma);
  856. ff_h264_pred_init(&h->hpc, h->avctx->codec_id, sps->bit_depth_luma,
  857. sps->chroma_format_idc);
  858. ff_videodsp_init(&h->vdsp, sps->bit_depth_luma);
  859. if (!HAVE_THREADS || !(h->avctx->active_thread_type & FF_THREAD_SLICE)) {
  860. ret = ff_h264_slice_context_init(h, &h->slice_ctx[0]);
  861. if (ret < 0) {
  862. av_log(h->avctx, AV_LOG_ERROR, "context_init() failed.\n");
  863. goto fail;
  864. }
  865. } else {
  866. for (i = 0; i < h->nb_slice_ctx; i++) {
  867. H264SliceContext *sl = &h->slice_ctx[i];
  868. sl->h264 = h;
  869. sl->intra4x4_pred_mode = h->intra4x4_pred_mode + i * 8 * 2 * h->mb_stride;
  870. sl->mvd_table[0] = h->mvd_table[0] + i * 8 * 2 * h->mb_stride;
  871. sl->mvd_table[1] = h->mvd_table[1] + i * 8 * 2 * h->mb_stride;
  872. if ((ret = ff_h264_slice_context_init(h, sl)) < 0) {
  873. av_log(h->avctx, AV_LOG_ERROR, "context_init() failed.\n");
  874. goto fail;
  875. }
  876. }
  877. }
  878. h->context_initialized = 1;
  879. return 0;
  880. fail:
  881. ff_h264_free_tables(h);
  882. h->context_initialized = 0;
  883. return ret;
  884. }
  885. static enum AVPixelFormat non_j_pixfmt(enum AVPixelFormat a)
  886. {
  887. switch (a) {
  888. case AV_PIX_FMT_YUVJ420P: return AV_PIX_FMT_YUV420P;
  889. case AV_PIX_FMT_YUVJ422P: return AV_PIX_FMT_YUV422P;
  890. case AV_PIX_FMT_YUVJ444P: return AV_PIX_FMT_YUV444P;
  891. default:
  892. return a;
  893. }
  894. }
  895. static int h264_init_ps(H264Context *h, const H264SliceContext *sl, int first_slice)
  896. {
  897. const SPS *sps;
  898. int needs_reinit = 0, must_reinit, ret;
  899. if (first_slice) {
  900. av_buffer_unref(&h->ps.pps_ref);
  901. h->ps.pps = NULL;
  902. h->ps.pps_ref = av_buffer_ref(h->ps.pps_list[sl->pps_id]);
  903. if (!h->ps.pps_ref)
  904. return AVERROR(ENOMEM);
  905. h->ps.pps = (const PPS*)h->ps.pps_ref->data;
  906. }
  907. if (h->ps.sps != h->ps.pps->sps) {
  908. h->ps.sps = (const SPS*)h->ps.pps->sps;
  909. if (h->mb_width != h->ps.sps->mb_width ||
  910. h->mb_height != h->ps.sps->mb_height ||
  911. h->cur_bit_depth_luma != h->ps.sps->bit_depth_luma ||
  912. h->cur_chroma_format_idc != h->ps.sps->chroma_format_idc
  913. )
  914. needs_reinit = 1;
  915. if (h->bit_depth_luma != h->ps.sps->bit_depth_luma ||
  916. h->chroma_format_idc != h->ps.sps->chroma_format_idc)
  917. needs_reinit = 1;
  918. }
  919. sps = h->ps.sps;
  920. must_reinit = (h->context_initialized &&
  921. ( 16*sps->mb_width != h->avctx->coded_width
  922. || 16*sps->mb_height != h->avctx->coded_height
  923. || h->cur_bit_depth_luma != sps->bit_depth_luma
  924. || h->cur_chroma_format_idc != sps->chroma_format_idc
  925. || h->mb_width != sps->mb_width
  926. || h->mb_height != sps->mb_height
  927. ));
  928. if (h->avctx->pix_fmt == AV_PIX_FMT_NONE
  929. || (non_j_pixfmt(h->avctx->pix_fmt) != non_j_pixfmt(get_pixel_format(h, 0))))
  930. must_reinit = 1;
  931. if (first_slice && av_cmp_q(sps->sar, h->avctx->sample_aspect_ratio))
  932. must_reinit = 1;
  933. if (!h->setup_finished) {
  934. h->avctx->profile = ff_h264_get_profile(sps);
  935. h->avctx->level = sps->level_idc;
  936. h->avctx->refs = sps->ref_frame_count;
  937. h->mb_width = sps->mb_width;
  938. h->mb_height = sps->mb_height;
  939. h->mb_num = h->mb_width * h->mb_height;
  940. h->mb_stride = h->mb_width + 1;
  941. h->b_stride = h->mb_width * 4;
  942. h->chroma_y_shift = sps->chroma_format_idc <= 1; // 400 uses yuv420p
  943. h->width = 16 * h->mb_width;
  944. h->height = 16 * h->mb_height;
  945. init_dimensions(h);
  946. if (sps->video_signal_type_present_flag) {
  947. h->avctx->color_range = sps->full_range > 0 ? AVCOL_RANGE_JPEG
  948. : AVCOL_RANGE_MPEG;
  949. if (sps->colour_description_present_flag) {
  950. if (h->avctx->colorspace != sps->colorspace)
  951. needs_reinit = 1;
  952. h->avctx->color_primaries = sps->color_primaries;
  953. h->avctx->color_trc = sps->color_trc;
  954. h->avctx->colorspace = sps->colorspace;
  955. }
  956. }
  957. if (h->sei.alternative_transfer.present &&
  958. av_color_transfer_name(h->sei.alternative_transfer.preferred_transfer_characteristics) &&
  959. h->sei.alternative_transfer.preferred_transfer_characteristics != AVCOL_TRC_UNSPECIFIED) {
  960. h->avctx->color_trc = h->sei.alternative_transfer.preferred_transfer_characteristics;
  961. }
  962. }
  963. h->avctx->chroma_sample_location = sps->chroma_location;
  964. if (!h->context_initialized || must_reinit || needs_reinit) {
  965. int flush_changes = h->context_initialized;
  966. h->context_initialized = 0;
  967. if (sl != h->slice_ctx) {
  968. av_log(h->avctx, AV_LOG_ERROR,
  969. "changing width %d -> %d / height %d -> %d on "
  970. "slice %d\n",
  971. h->width, h->avctx->coded_width,
  972. h->height, h->avctx->coded_height,
  973. h->current_slice + 1);
  974. return AVERROR_INVALIDDATA;
  975. }
  976. av_assert1(first_slice);
  977. if (flush_changes)
  978. ff_h264_flush_change(h);
  979. if ((ret = get_pixel_format(h, 1)) < 0)
  980. return ret;
  981. h->avctx->pix_fmt = ret;
  982. av_log(h->avctx, AV_LOG_VERBOSE, "Reinit context to %dx%d, "
  983. "pix_fmt: %s\n", h->width, h->height, av_get_pix_fmt_name(h->avctx->pix_fmt));
  984. if ((ret = h264_slice_header_init(h)) < 0) {
  985. av_log(h->avctx, AV_LOG_ERROR,
  986. "h264_slice_header_init() failed\n");
  987. return ret;
  988. }
  989. }
  990. return 0;
  991. }
  992. static int h264_export_frame_props(H264Context *h)
  993. {
  994. const SPS *sps = h->ps.sps;
  995. H264Picture *cur = h->cur_pic_ptr;
  996. AVFrame *out = cur->f;
  997. out->interlaced_frame = 0;
  998. out->repeat_pict = 0;
  999. /* Signal interlacing information externally. */
  1000. /* Prioritize picture timing SEI information over used
  1001. * decoding process if it exists. */
  1002. if (h->sei.picture_timing.present) {
  1003. int ret = ff_h264_sei_process_picture_timing(&h->sei.picture_timing, sps,
  1004. h->avctx);
  1005. if (ret < 0) {
  1006. av_log(h->avctx, AV_LOG_ERROR, "Error processing a picture timing SEI\n");
  1007. if (h->avctx->err_recognition & AV_EF_EXPLODE)
  1008. return ret;
  1009. h->sei.picture_timing.present = 0;
  1010. }
  1011. }
  1012. if (sps->pic_struct_present_flag && h->sei.picture_timing.present) {
  1013. H264SEIPictureTiming *pt = &h->sei.picture_timing;
  1014. switch (pt->pic_struct) {
  1015. case H264_SEI_PIC_STRUCT_FRAME:
  1016. break;
  1017. case H264_SEI_PIC_STRUCT_TOP_FIELD:
  1018. case H264_SEI_PIC_STRUCT_BOTTOM_FIELD:
  1019. out->interlaced_frame = 1;
  1020. break;
  1021. case H264_SEI_PIC_STRUCT_TOP_BOTTOM:
  1022. case H264_SEI_PIC_STRUCT_BOTTOM_TOP:
  1023. if (FIELD_OR_MBAFF_PICTURE(h))
  1024. out->interlaced_frame = 1;
  1025. else
  1026. // try to flag soft telecine progressive
  1027. out->interlaced_frame = h->prev_interlaced_frame;
  1028. break;
  1029. case H264_SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
  1030. case H264_SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
  1031. /* Signal the possibility of telecined film externally
  1032. * (pic_struct 5,6). From these hints, let the applications
  1033. * decide if they apply deinterlacing. */
  1034. out->repeat_pict = 1;
  1035. break;
  1036. case H264_SEI_PIC_STRUCT_FRAME_DOUBLING:
  1037. out->repeat_pict = 2;
  1038. break;
  1039. case H264_SEI_PIC_STRUCT_FRAME_TRIPLING:
  1040. out->repeat_pict = 4;
  1041. break;
  1042. }
  1043. if ((pt->ct_type & 3) &&
  1044. pt->pic_struct <= H264_SEI_PIC_STRUCT_BOTTOM_TOP)
  1045. out->interlaced_frame = (pt->ct_type & (1 << 1)) != 0;
  1046. } else {
  1047. /* Derive interlacing flag from used decoding process. */
  1048. out->interlaced_frame = FIELD_OR_MBAFF_PICTURE(h);
  1049. }
  1050. h->prev_interlaced_frame = out->interlaced_frame;
  1051. if (cur->field_poc[0] != cur->field_poc[1]) {
  1052. /* Derive top_field_first from field pocs. */
  1053. out->top_field_first = cur->field_poc[0] < cur->field_poc[1];
  1054. } else {
  1055. if (sps->pic_struct_present_flag && h->sei.picture_timing.present) {
  1056. /* Use picture timing SEI information. Even if it is a
  1057. * information of a past frame, better than nothing. */
  1058. if (h->sei.picture_timing.pic_struct == H264_SEI_PIC_STRUCT_TOP_BOTTOM ||
  1059. h->sei.picture_timing.pic_struct == H264_SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
  1060. out->top_field_first = 1;
  1061. else
  1062. out->top_field_first = 0;
  1063. } else if (out->interlaced_frame) {
  1064. /* Default to top field first when pic_struct_present_flag
  1065. * is not set but interlaced frame detected */
  1066. out->top_field_first = 1;
  1067. } else {
  1068. /* Most likely progressive */
  1069. out->top_field_first = 0;
  1070. }
  1071. }
  1072. if (h->sei.frame_packing.present &&
  1073. h->sei.frame_packing.arrangement_type <= 6 &&
  1074. h->sei.frame_packing.content_interpretation_type > 0 &&
  1075. h->sei.frame_packing.content_interpretation_type < 3) {
  1076. H264SEIFramePacking *fp = &h->sei.frame_packing;
  1077. AVStereo3D *stereo = av_stereo3d_create_side_data(out);
  1078. if (stereo) {
  1079. switch (fp->arrangement_type) {
  1080. case H264_SEI_FPA_TYPE_CHECKERBOARD:
  1081. stereo->type = AV_STEREO3D_CHECKERBOARD;
  1082. break;
  1083. case H264_SEI_FPA_TYPE_INTERLEAVE_COLUMN:
  1084. stereo->type = AV_STEREO3D_COLUMNS;
  1085. break;
  1086. case H264_SEI_FPA_TYPE_INTERLEAVE_ROW:
  1087. stereo->type = AV_STEREO3D_LINES;
  1088. break;
  1089. case H264_SEI_FPA_TYPE_SIDE_BY_SIDE:
  1090. if (fp->quincunx_sampling_flag)
  1091. stereo->type = AV_STEREO3D_SIDEBYSIDE_QUINCUNX;
  1092. else
  1093. stereo->type = AV_STEREO3D_SIDEBYSIDE;
  1094. break;
  1095. case H264_SEI_FPA_TYPE_TOP_BOTTOM:
  1096. stereo->type = AV_STEREO3D_TOPBOTTOM;
  1097. break;
  1098. case H264_SEI_FPA_TYPE_INTERLEAVE_TEMPORAL:
  1099. stereo->type = AV_STEREO3D_FRAMESEQUENCE;
  1100. break;
  1101. case H264_SEI_FPA_TYPE_2D:
  1102. stereo->type = AV_STEREO3D_2D;
  1103. break;
  1104. }
  1105. if (fp->content_interpretation_type == 2)
  1106. stereo->flags = AV_STEREO3D_FLAG_INVERT;
  1107. if (fp->arrangement_type == H264_SEI_FPA_TYPE_INTERLEAVE_TEMPORAL) {
  1108. if (fp->current_frame_is_frame0_flag)
  1109. stereo->view = AV_STEREO3D_VIEW_LEFT;
  1110. else
  1111. stereo->view = AV_STEREO3D_VIEW_RIGHT;
  1112. }
  1113. }
  1114. }
  1115. if (h->sei.display_orientation.present &&
  1116. (h->sei.display_orientation.anticlockwise_rotation ||
  1117. h->sei.display_orientation.hflip ||
  1118. h->sei.display_orientation.vflip)) {
  1119. H264SEIDisplayOrientation *o = &h->sei.display_orientation;
  1120. double angle = o->anticlockwise_rotation * 360 / (double) (1 << 16);
  1121. AVFrameSideData *rotation = av_frame_new_side_data(out,
  1122. AV_FRAME_DATA_DISPLAYMATRIX,
  1123. sizeof(int32_t) * 9);
  1124. if (rotation) {
  1125. av_display_rotation_set((int32_t *)rotation->data, angle);
  1126. av_display_matrix_flip((int32_t *)rotation->data,
  1127. o->hflip, o->vflip);
  1128. }
  1129. }
  1130. if (h->sei.afd.present) {
  1131. AVFrameSideData *sd = av_frame_new_side_data(out, AV_FRAME_DATA_AFD,
  1132. sizeof(uint8_t));
  1133. if (sd) {
  1134. *sd->data = h->sei.afd.active_format_description;
  1135. h->sei.afd.present = 0;
  1136. }
  1137. }
  1138. if (h->sei.a53_caption.buf_ref) {
  1139. H264SEIA53Caption *a53 = &h->sei.a53_caption;
  1140. AVFrameSideData *sd = av_frame_new_side_data_from_buf(out, AV_FRAME_DATA_A53_CC, a53->buf_ref);
  1141. if (!sd)
  1142. av_buffer_unref(&a53->buf_ref);
  1143. a53->buf_ref = NULL;
  1144. h->avctx->properties |= FF_CODEC_PROPERTY_CLOSED_CAPTIONS;
  1145. }
  1146. for (int i = 0; i < h->sei.unregistered.nb_buf_ref; i++) {
  1147. H264SEIUnregistered *unreg = &h->sei.unregistered;
  1148. if (unreg->buf_ref[i]) {
  1149. AVFrameSideData *sd = av_frame_new_side_data_from_buf(out,
  1150. AV_FRAME_DATA_SEI_UNREGISTERED,
  1151. unreg->buf_ref[i]);
  1152. if (!sd)
  1153. av_buffer_unref(&unreg->buf_ref[i]);
  1154. unreg->buf_ref[i] = NULL;
  1155. }
  1156. }
  1157. h->sei.unregistered.nb_buf_ref = 0;
  1158. if (h->sei.picture_timing.timecode_cnt > 0) {
  1159. uint32_t *tc_sd;
  1160. char tcbuf[AV_TIMECODE_STR_SIZE];
  1161. AVFrameSideData *tcside = av_frame_new_side_data(out,
  1162. AV_FRAME_DATA_S12M_TIMECODE,
  1163. sizeof(uint32_t)*4);
  1164. if (!tcside)
  1165. return AVERROR(ENOMEM);
  1166. tc_sd = (uint32_t*)tcside->data;
  1167. tc_sd[0] = h->sei.picture_timing.timecode_cnt;
  1168. for (int i = 0; i < tc_sd[0]; i++) {
  1169. int drop = h->sei.picture_timing.timecode[i].dropframe;
  1170. int hh = h->sei.picture_timing.timecode[i].hours;
  1171. int mm = h->sei.picture_timing.timecode[i].minutes;
  1172. int ss = h->sei.picture_timing.timecode[i].seconds;
  1173. int ff = h->sei.picture_timing.timecode[i].frame;
  1174. tc_sd[i + 1] = av_timecode_get_smpte(h->avctx->framerate, drop, hh, mm, ss, ff);
  1175. av_timecode_make_smpte_tc_string2(tcbuf, h->avctx->framerate, tc_sd[i + 1], 0, 0);
  1176. av_dict_set(&out->metadata, "timecode", tcbuf, 0);
  1177. }
  1178. h->sei.picture_timing.timecode_cnt = 0;
  1179. }
  1180. return 0;
  1181. }
  1182. static int h264_select_output_frame(H264Context *h)
  1183. {
  1184. const SPS *sps = h->ps.sps;
  1185. H264Picture *out = h->cur_pic_ptr;
  1186. H264Picture *cur = h->cur_pic_ptr;
  1187. int i, pics, out_of_order, out_idx;
  1188. cur->mmco_reset = h->mmco_reset;
  1189. h->mmco_reset = 0;
  1190. if (sps->bitstream_restriction_flag ||
  1191. h->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT) {
  1192. h->avctx->has_b_frames = FFMAX(h->avctx->has_b_frames, sps->num_reorder_frames);
  1193. }
  1194. for (i = 0; 1; i++) {
  1195. if(i == MAX_DELAYED_PIC_COUNT || cur->poc < h->last_pocs[i]){
  1196. if(i)
  1197. h->last_pocs[i-1] = cur->poc;
  1198. break;
  1199. } else if(i) {
  1200. h->last_pocs[i-1]= h->last_pocs[i];
  1201. }
  1202. }
  1203. out_of_order = MAX_DELAYED_PIC_COUNT - i;
  1204. if( cur->f->pict_type == AV_PICTURE_TYPE_B
  1205. || (h->last_pocs[MAX_DELAYED_PIC_COUNT-2] > INT_MIN && h->last_pocs[MAX_DELAYED_PIC_COUNT-1] - (int64_t)h->last_pocs[MAX_DELAYED_PIC_COUNT-2] > 2))
  1206. out_of_order = FFMAX(out_of_order, 1);
  1207. if (out_of_order == MAX_DELAYED_PIC_COUNT) {
  1208. av_log(h->avctx, AV_LOG_VERBOSE, "Invalid POC %d<%d\n", cur->poc, h->last_pocs[0]);
  1209. for (i = 1; i < MAX_DELAYED_PIC_COUNT; i++)
  1210. h->last_pocs[i] = INT_MIN;
  1211. h->last_pocs[0] = cur->poc;
  1212. cur->mmco_reset = 1;
  1213. } else if(h->avctx->has_b_frames < out_of_order && !sps->bitstream_restriction_flag){
  1214. int loglevel = h->avctx->frame_number > 1 ? AV_LOG_WARNING : AV_LOG_VERBOSE;
  1215. av_log(h->avctx, loglevel, "Increasing reorder buffer to %d\n", out_of_order);
  1216. h->avctx->has_b_frames = out_of_order;
  1217. }
  1218. pics = 0;
  1219. while (h->delayed_pic[pics])
  1220. pics++;
  1221. av_assert0(pics <= MAX_DELAYED_PIC_COUNT);
  1222. h->delayed_pic[pics++] = cur;
  1223. if (cur->reference == 0)
  1224. cur->reference = DELAYED_PIC_REF;
  1225. out = h->delayed_pic[0];
  1226. out_idx = 0;
  1227. for (i = 1; h->delayed_pic[i] &&
  1228. !h->delayed_pic[i]->f->key_frame &&
  1229. !h->delayed_pic[i]->mmco_reset;
  1230. i++)
  1231. if (h->delayed_pic[i]->poc < out->poc) {
  1232. out = h->delayed_pic[i];
  1233. out_idx = i;
  1234. }
  1235. if (h->avctx->has_b_frames == 0 &&
  1236. (h->delayed_pic[0]->f->key_frame || h->delayed_pic[0]->mmco_reset))
  1237. h->next_outputed_poc = INT_MIN;
  1238. out_of_order = out->poc < h->next_outputed_poc;
  1239. if (out_of_order || pics > h->avctx->has_b_frames) {
  1240. out->reference &= ~DELAYED_PIC_REF;
  1241. for (i = out_idx; h->delayed_pic[i]; i++)
  1242. h->delayed_pic[i] = h->delayed_pic[i + 1];
  1243. }
  1244. if (!out_of_order && pics > h->avctx->has_b_frames) {
  1245. h->next_output_pic = out;
  1246. if (out_idx == 0 && h->delayed_pic[0] && (h->delayed_pic[0]->f->key_frame || h->delayed_pic[0]->mmco_reset)) {
  1247. h->next_outputed_poc = INT_MIN;
  1248. } else
  1249. h->next_outputed_poc = out->poc;
  1250. if (out->recovered) {
  1251. // We have reached an recovery point and all frames after it in
  1252. // display order are "recovered".
  1253. h->frame_recovered |= FRAME_RECOVERED_SEI;
  1254. }
  1255. out->recovered |= !!(h->frame_recovered & FRAME_RECOVERED_SEI);
  1256. if (!out->recovered) {
  1257. if (!(h->avctx->flags & AV_CODEC_FLAG_OUTPUT_CORRUPT) &&
  1258. !(h->avctx->flags2 & AV_CODEC_FLAG2_SHOW_ALL)) {
  1259. h->next_output_pic = NULL;
  1260. } else {
  1261. out->f->flags |= AV_FRAME_FLAG_CORRUPT;
  1262. }
  1263. }
  1264. } else {
  1265. av_log(h->avctx, AV_LOG_DEBUG, "no picture %s\n", out_of_order ? "ooo" : "");
  1266. }
  1267. return 0;
  1268. }
  1269. /* This function is called right after decoding the slice header for a first
  1270. * slice in a field (or a frame). It decides whether we are decoding a new frame
  1271. * or a second field in a pair and does the necessary setup.
  1272. */
  1273. static int h264_field_start(H264Context *h, const H264SliceContext *sl,
  1274. const H2645NAL *nal, int first_slice)
  1275. {
  1276. int i;
  1277. const SPS *sps;
  1278. int last_pic_structure, last_pic_droppable, ret;
  1279. ret = h264_init_ps(h, sl, first_slice);
  1280. if (ret < 0)
  1281. return ret;
  1282. sps = h->ps.sps;
  1283. if (sps && sps->bitstream_restriction_flag &&
  1284. h->avctx->has_b_frames < sps->num_reorder_frames) {
  1285. h->avctx->has_b_frames = sps->num_reorder_frames;
  1286. }
  1287. last_pic_droppable = h->droppable;
  1288. last_pic_structure = h->picture_structure;
  1289. h->droppable = (nal->ref_idc == 0);
  1290. h->picture_structure = sl->picture_structure;
  1291. h->poc.frame_num = sl->frame_num;
  1292. h->poc.poc_lsb = sl->poc_lsb;
  1293. h->poc.delta_poc_bottom = sl->delta_poc_bottom;
  1294. h->poc.delta_poc[0] = sl->delta_poc[0];
  1295. h->poc.delta_poc[1] = sl->delta_poc[1];
  1296. /* Shorten frame num gaps so we don't have to allocate reference
  1297. * frames just to throw them away */
  1298. if (h->poc.frame_num != h->poc.prev_frame_num) {
  1299. int unwrap_prev_frame_num = h->poc.prev_frame_num;
  1300. int max_frame_num = 1 << sps->log2_max_frame_num;
  1301. if (unwrap_prev_frame_num > h->poc.frame_num)
  1302. unwrap_prev_frame_num -= max_frame_num;
  1303. if ((h->poc.frame_num - unwrap_prev_frame_num) > sps->ref_frame_count) {
  1304. unwrap_prev_frame_num = (h->poc.frame_num - sps->ref_frame_count) - 1;
  1305. if (unwrap_prev_frame_num < 0)
  1306. unwrap_prev_frame_num += max_frame_num;
  1307. h->poc.prev_frame_num = unwrap_prev_frame_num;
  1308. }
  1309. }
  1310. /* See if we have a decoded first field looking for a pair...
  1311. * Here, we're using that to see if we should mark previously
  1312. * decode frames as "finished".
  1313. * We have to do that before the "dummy" in-between frame allocation,
  1314. * since that can modify h->cur_pic_ptr. */
  1315. if (h->first_field) {
  1316. int last_field = last_pic_structure == PICT_BOTTOM_FIELD;
  1317. av_assert0(h->cur_pic_ptr);
  1318. av_assert0(h->cur_pic_ptr->f->buf[0]);
  1319. assert(h->cur_pic_ptr->reference != DELAYED_PIC_REF);
  1320. /* Mark old field/frame as completed */
  1321. if (h->cur_pic_ptr->tf.owner[last_field] == h->avctx) {
  1322. ff_thread_report_progress(&h->cur_pic_ptr->tf, INT_MAX, last_field);
  1323. }
  1324. /* figure out if we have a complementary field pair */
  1325. if (!FIELD_PICTURE(h) || h->picture_structure == last_pic_structure) {
  1326. /* Previous field is unmatched. Don't display it, but let it
  1327. * remain for reference if marked as such. */
  1328. if (last_pic_structure != PICT_FRAME) {
  1329. ff_thread_report_progress(&h->cur_pic_ptr->tf, INT_MAX,
  1330. last_pic_structure == PICT_TOP_FIELD);
  1331. }
  1332. } else {
  1333. if (h->cur_pic_ptr->frame_num != h->poc.frame_num) {
  1334. /* This and previous field were reference, but had
  1335. * different frame_nums. Consider this field first in
  1336. * pair. Throw away previous field except for reference
  1337. * purposes. */
  1338. if (last_pic_structure != PICT_FRAME) {
  1339. ff_thread_report_progress(&h->cur_pic_ptr->tf, INT_MAX,
  1340. last_pic_structure == PICT_TOP_FIELD);
  1341. }
  1342. } else {
  1343. /* Second field in complementary pair */
  1344. if (!((last_pic_structure == PICT_TOP_FIELD &&
  1345. h->picture_structure == PICT_BOTTOM_FIELD) ||
  1346. (last_pic_structure == PICT_BOTTOM_FIELD &&
  1347. h->picture_structure == PICT_TOP_FIELD))) {
  1348. av_log(h->avctx, AV_LOG_ERROR,
  1349. "Invalid field mode combination %d/%d\n",
  1350. last_pic_structure, h->picture_structure);
  1351. h->picture_structure = last_pic_structure;
  1352. h->droppable = last_pic_droppable;
  1353. return AVERROR_INVALIDDATA;
  1354. } else if (last_pic_droppable != h->droppable) {
  1355. avpriv_request_sample(h->avctx,
  1356. "Found reference and non-reference fields in the same frame, which");
  1357. h->picture_structure = last_pic_structure;
  1358. h->droppable = last_pic_droppable;
  1359. return AVERROR_PATCHWELCOME;
  1360. }
  1361. }
  1362. }
  1363. }
  1364. while (h->poc.frame_num != h->poc.prev_frame_num && !h->first_field &&
  1365. h->poc.frame_num != (h->poc.prev_frame_num + 1) % (1 << sps->log2_max_frame_num)) {
  1366. H264Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
  1367. av_log(h->avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n",
  1368. h->poc.frame_num, h->poc.prev_frame_num);
  1369. if (!sps->gaps_in_frame_num_allowed_flag)
  1370. for(i=0; i<FF_ARRAY_ELEMS(h->last_pocs); i++)
  1371. h->last_pocs[i] = INT_MIN;
  1372. ret = h264_frame_start(h);
  1373. if (ret < 0) {
  1374. h->first_field = 0;
  1375. return ret;
  1376. }
  1377. h->poc.prev_frame_num++;
  1378. h->poc.prev_frame_num %= 1 << sps->log2_max_frame_num;
  1379. h->cur_pic_ptr->frame_num = h->poc.prev_frame_num;
  1380. h->cur_pic_ptr->invalid_gap = !sps->gaps_in_frame_num_allowed_flag;
  1381. ff_thread_report_progress(&h->cur_pic_ptr->tf, INT_MAX, 0);
  1382. ff_thread_report_progress(&h->cur_pic_ptr->tf, INT_MAX, 1);
  1383. h->explicit_ref_marking = 0;
  1384. ret = ff_h264_execute_ref_pic_marking(h);
  1385. if (ret < 0 && (h->avctx->err_recognition & AV_EF_EXPLODE))
  1386. return ret;
  1387. /* Error concealment: If a ref is missing, copy the previous ref
  1388. * in its place.
  1389. * FIXME: Avoiding a memcpy would be nice, but ref handling makes
  1390. * many assumptions about there being no actual duplicates.
  1391. * FIXME: This does not copy padding for out-of-frame motion
  1392. * vectors. Given we are concealing a lost frame, this probably
  1393. * is not noticeable by comparison, but it should be fixed. */
  1394. if (h->short_ref_count) {
  1395. int c[4] = {
  1396. 1<<(h->ps.sps->bit_depth_luma-1),
  1397. 1<<(h->ps.sps->bit_depth_chroma-1),
  1398. 1<<(h->ps.sps->bit_depth_chroma-1),
  1399. -1
  1400. };
  1401. if (prev &&
  1402. h->short_ref[0]->f->width == prev->f->width &&
  1403. h->short_ref[0]->f->height == prev->f->height &&
  1404. h->short_ref[0]->f->format == prev->f->format) {
  1405. ff_thread_await_progress(&prev->tf, INT_MAX, 0);
  1406. if (prev->field_picture)
  1407. ff_thread_await_progress(&prev->tf, INT_MAX, 1);
  1408. ff_thread_release_buffer(h->avctx, &h->short_ref[0]->tf);
  1409. h->short_ref[0]->tf.f = h->short_ref[0]->f;
  1410. ret = ff_thread_ref_frame(&h->short_ref[0]->tf, &prev->tf);
  1411. if (ret < 0)
  1412. return ret;
  1413. h->short_ref[0]->poc = prev->poc + 2U;
  1414. ff_thread_report_progress(&h->short_ref[0]->tf, INT_MAX, 0);
  1415. if (h->short_ref[0]->field_picture)
  1416. ff_thread_report_progress(&h->short_ref[0]->tf, INT_MAX, 1);
  1417. } else if (!h->frame_recovered && !h->avctx->hwaccel)
  1418. ff_color_frame(h->short_ref[0]->f, c);
  1419. h->short_ref[0]->frame_num = h->poc.prev_frame_num;
  1420. }
  1421. }
  1422. /* See if we have a decoded first field looking for a pair...
  1423. * We're using that to see whether to continue decoding in that
  1424. * frame, or to allocate a new one. */
  1425. if (h->first_field) {
  1426. av_assert0(h->cur_pic_ptr);
  1427. av_assert0(h->cur_pic_ptr->f->buf[0]);
  1428. assert(h->cur_pic_ptr->reference != DELAYED_PIC_REF);
  1429. /* figure out if we have a complementary field pair */
  1430. if (!FIELD_PICTURE(h) || h->picture_structure == last_pic_structure) {
  1431. /* Previous field is unmatched. Don't display it, but let it
  1432. * remain for reference if marked as such. */
  1433. h->missing_fields ++;
  1434. h->cur_pic_ptr = NULL;
  1435. h->first_field = FIELD_PICTURE(h);
  1436. } else {
  1437. h->missing_fields = 0;
  1438. if (h->cur_pic_ptr->frame_num != h->poc.frame_num) {
  1439. ff_thread_report_progress(&h->cur_pic_ptr->tf, INT_MAX,
  1440. h->picture_structure==PICT_BOTTOM_FIELD);
  1441. /* This and the previous field had different frame_nums.
  1442. * Consider this field first in pair. Throw away previous
  1443. * one except for reference purposes. */
  1444. h->first_field = 1;
  1445. h->cur_pic_ptr = NULL;
  1446. } else if (h->cur_pic_ptr->reference & DELAYED_PIC_REF) {
  1447. /* This frame was already output, we cannot draw into it
  1448. * anymore.
  1449. */
  1450. h->first_field = 1;
  1451. h->cur_pic_ptr = NULL;
  1452. } else {
  1453. /* Second field in complementary pair */
  1454. h->first_field = 0;
  1455. }
  1456. }
  1457. } else {
  1458. /* Frame or first field in a potentially complementary pair */
  1459. h->first_field = FIELD_PICTURE(h);
  1460. }
  1461. if (!FIELD_PICTURE(h) || h->first_field) {
  1462. if (h264_frame_start(h) < 0) {
  1463. h->first_field = 0;
  1464. return AVERROR_INVALIDDATA;
  1465. }
  1466. } else {
  1467. int field = h->picture_structure == PICT_BOTTOM_FIELD;
  1468. release_unused_pictures(h, 0);
  1469. h->cur_pic_ptr->tf.owner[field] = h->avctx;
  1470. }
  1471. /* Some macroblocks can be accessed before they're available in case
  1472. * of lost slices, MBAFF or threading. */
  1473. if (FIELD_PICTURE(h)) {
  1474. for(i = (h->picture_structure == PICT_BOTTOM_FIELD); i<h->mb_height; i++)
  1475. memset(h->slice_table + i*h->mb_stride, -1, (h->mb_stride - (i+1==h->mb_height)) * sizeof(*h->slice_table));
  1476. } else {
  1477. memset(h->slice_table, -1,
  1478. (h->mb_height * h->mb_stride - 1) * sizeof(*h->slice_table));
  1479. }
  1480. ret = ff_h264_init_poc(h->cur_pic_ptr->field_poc, &h->cur_pic_ptr->poc,
  1481. h->ps.sps, &h->poc, h->picture_structure, nal->ref_idc);
  1482. if (ret < 0)
  1483. return ret;
  1484. memcpy(h->mmco, sl->mmco, sl->nb_mmco * sizeof(*h->mmco));
  1485. h->nb_mmco = sl->nb_mmco;
  1486. h->explicit_ref_marking = sl->explicit_ref_marking;
  1487. h->picture_idr = nal->type == H264_NAL_IDR_SLICE;
  1488. if (h->sei.recovery_point.recovery_frame_cnt >= 0) {
  1489. const int sei_recovery_frame_cnt = h->sei.recovery_point.recovery_frame_cnt;
  1490. if (h->poc.frame_num != sei_recovery_frame_cnt || sl->slice_type_nos != AV_PICTURE_TYPE_I)
  1491. h->valid_recovery_point = 1;
  1492. if ( h->recovery_frame < 0
  1493. || av_mod_uintp2(h->recovery_frame - h->poc.frame_num, h->ps.sps->log2_max_frame_num) > sei_recovery_frame_cnt) {
  1494. h->recovery_frame = av_mod_uintp2(h->poc.frame_num + sei_recovery_frame_cnt, h->ps.sps->log2_max_frame_num);
  1495. if (!h->valid_recovery_point)
  1496. h->recovery_frame = h->poc.frame_num;
  1497. }
  1498. }
  1499. h->cur_pic_ptr->f->key_frame |= (nal->type == H264_NAL_IDR_SLICE);
  1500. if (nal->type == H264_NAL_IDR_SLICE ||
  1501. (h->recovery_frame == h->poc.frame_num && nal->ref_idc)) {
  1502. h->recovery_frame = -1;
  1503. h->cur_pic_ptr->recovered = 1;
  1504. }
  1505. // If we have an IDR, all frames after it in decoded order are
  1506. // "recovered".
  1507. if (nal->type == H264_NAL_IDR_SLICE)
  1508. h->frame_recovered |= FRAME_RECOVERED_IDR;
  1509. #if 1
  1510. h->cur_pic_ptr->recovered |= h->frame_recovered;
  1511. #else
  1512. h->cur_pic_ptr->recovered |= !!(h->frame_recovered & FRAME_RECOVERED_IDR);
  1513. #endif
  1514. /* Set the frame properties/side data. Only done for the second field in
  1515. * field coded frames, since some SEI information is present for each field
  1516. * and is merged by the SEI parsing code. */
  1517. if (!FIELD_PICTURE(h) || !h->first_field || h->missing_fields > 1) {
  1518. ret = h264_export_frame_props(h);
  1519. if (ret < 0)
  1520. return ret;
  1521. ret = h264_select_output_frame(h);
  1522. if (ret < 0)
  1523. return ret;
  1524. }
  1525. return 0;
  1526. }
  1527. static int h264_slice_header_parse(const H264Context *h, H264SliceContext *sl,
  1528. const H2645NAL *nal)
  1529. {
  1530. const SPS *sps;
  1531. const PPS *pps;
  1532. int ret;
  1533. unsigned int slice_type, tmp, i;
  1534. int field_pic_flag, bottom_field_flag;
  1535. int first_slice = sl == h->slice_ctx && !h->current_slice;
  1536. int picture_structure;
  1537. if (first_slice)
  1538. av_assert0(!h->setup_finished);
  1539. sl->first_mb_addr = get_ue_golomb_long(&sl->gb);
  1540. slice_type = get_ue_golomb_31(&sl->gb);
  1541. if (slice_type > 9) {
  1542. av_log(h->avctx, AV_LOG_ERROR,
  1543. "slice type %d too large at %d\n",
  1544. slice_type, sl->first_mb_addr);
  1545. return AVERROR_INVALIDDATA;
  1546. }
  1547. if (slice_type > 4) {
  1548. slice_type -= 5;
  1549. sl->slice_type_fixed = 1;
  1550. } else
  1551. sl->slice_type_fixed = 0;
  1552. slice_type = ff_h264_golomb_to_pict_type[slice_type];
  1553. sl->slice_type = slice_type;
  1554. sl->slice_type_nos = slice_type & 3;
  1555. if (nal->type == H264_NAL_IDR_SLICE &&
  1556. sl->slice_type_nos != AV_PICTURE_TYPE_I) {
  1557. av_log(h->avctx, AV_LOG_ERROR, "A non-intra slice in an IDR NAL unit.\n");
  1558. return AVERROR_INVALIDDATA;
  1559. }
  1560. sl->pps_id = get_ue_golomb(&sl->gb);
  1561. if (sl->pps_id >= MAX_PPS_COUNT) {
  1562. av_log(h->avctx, AV_LOG_ERROR, "pps_id %u out of range\n", sl->pps_id);
  1563. return AVERROR_INVALIDDATA;
  1564. }
  1565. if (!h->ps.pps_list[sl->pps_id]) {
  1566. av_log(h->avctx, AV_LOG_ERROR,
  1567. "non-existing PPS %u referenced\n",
  1568. sl->pps_id);
  1569. return AVERROR_INVALIDDATA;
  1570. }
  1571. pps = (const PPS*)h->ps.pps_list[sl->pps_id]->data;
  1572. sps = pps->sps;
  1573. sl->frame_num = get_bits(&sl->gb, sps->log2_max_frame_num);
  1574. if (!first_slice) {
  1575. if (h->poc.frame_num != sl->frame_num) {
  1576. av_log(h->avctx, AV_LOG_ERROR, "Frame num change from %d to %d\n",
  1577. h->poc.frame_num, sl->frame_num);
  1578. return AVERROR_INVALIDDATA;
  1579. }
  1580. }
  1581. sl->mb_mbaff = 0;
  1582. if (sps->frame_mbs_only_flag) {
  1583. picture_structure = PICT_FRAME;
  1584. } else {
  1585. if (!sps->direct_8x8_inference_flag && slice_type == AV_PICTURE_TYPE_B) {
  1586. av_log(h->avctx, AV_LOG_ERROR, "This stream was generated by a broken encoder, invalid 8x8 inference\n");
  1587. return -1;
  1588. }
  1589. field_pic_flag = get_bits1(&sl->gb);
  1590. if (field_pic_flag) {
  1591. bottom_field_flag = get_bits1(&sl->gb);
  1592. picture_structure = PICT_TOP_FIELD + bottom_field_flag;
  1593. } else {
  1594. picture_structure = PICT_FRAME;
  1595. }
  1596. }
  1597. sl->picture_structure = picture_structure;
  1598. sl->mb_field_decoding_flag = picture_structure != PICT_FRAME;
  1599. if (picture_structure == PICT_FRAME) {
  1600. sl->curr_pic_num = sl->frame_num;
  1601. sl->max_pic_num = 1 << sps->log2_max_frame_num;
  1602. } else {
  1603. sl->curr_pic_num = 2 * sl->frame_num + 1;
  1604. sl->max_pic_num = 1 << (sps->log2_max_frame_num + 1);
  1605. }
  1606. if (nal->type == H264_NAL_IDR_SLICE)
  1607. get_ue_golomb_long(&sl->gb); /* idr_pic_id */
  1608. if (sps->poc_type == 0) {
  1609. sl->poc_lsb = get_bits(&sl->gb, sps->log2_max_poc_lsb);
  1610. if (pps->pic_order_present == 1 && picture_structure == PICT_FRAME)
  1611. sl->delta_poc_bottom = get_se_golomb(&sl->gb);
  1612. }
  1613. if (sps->poc_type == 1 && !sps->delta_pic_order_always_zero_flag) {
  1614. sl->delta_poc[0] = get_se_golomb(&sl->gb);
  1615. if (pps->pic_order_present == 1 && picture_structure == PICT_FRAME)
  1616. sl->delta_poc[1] = get_se_golomb(&sl->gb);
  1617. }
  1618. sl->redundant_pic_count = 0;
  1619. if (pps->redundant_pic_cnt_present)
  1620. sl->redundant_pic_count = get_ue_golomb(&sl->gb);
  1621. if (sl->slice_type_nos == AV_PICTURE_TYPE_B)
  1622. sl->direct_spatial_mv_pred = get_bits1(&sl->gb);
  1623. ret = ff_h264_parse_ref_count(&sl->list_count, sl->ref_count,
  1624. &sl->gb, pps, sl->slice_type_nos,
  1625. picture_structure, h->avctx);
  1626. if (ret < 0)
  1627. return ret;
  1628. if (sl->slice_type_nos != AV_PICTURE_TYPE_I) {
  1629. ret = ff_h264_decode_ref_pic_list_reordering(sl, h->avctx);
  1630. if (ret < 0) {
  1631. sl->ref_count[1] = sl->ref_count[0] = 0;
  1632. return ret;
  1633. }
  1634. }
  1635. sl->pwt.use_weight = 0;
  1636. for (i = 0; i < 2; i++) {
  1637. sl->pwt.luma_weight_flag[i] = 0;
  1638. sl->pwt.chroma_weight_flag[i] = 0;
  1639. }
  1640. if ((pps->weighted_pred && sl->slice_type_nos == AV_PICTURE_TYPE_P) ||
  1641. (pps->weighted_bipred_idc == 1 &&
  1642. sl->slice_type_nos == AV_PICTURE_TYPE_B)) {
  1643. ret = ff_h264_pred_weight_table(&sl->gb, sps, sl->ref_count,
  1644. sl->slice_type_nos, &sl->pwt,
  1645. picture_structure, h->avctx);
  1646. if (ret < 0)
  1647. return ret;
  1648. }
  1649. sl->explicit_ref_marking = 0;
  1650. if (nal->ref_idc) {
  1651. ret = ff_h264_decode_ref_pic_marking(sl, &sl->gb, nal, h->avctx);
  1652. if (ret < 0 && (h->avctx->err_recognition & AV_EF_EXPLODE))
  1653. return AVERROR_INVALIDDATA;
  1654. }
  1655. if (sl->slice_type_nos != AV_PICTURE_TYPE_I && pps->cabac) {
  1656. tmp = get_ue_golomb_31(&sl->gb);
  1657. if (tmp > 2) {
  1658. av_log(h->avctx, AV_LOG_ERROR, "cabac_init_idc %u overflow\n", tmp);
  1659. return AVERROR_INVALIDDATA;
  1660. }
  1661. sl->cabac_init_idc = tmp;
  1662. }
  1663. sl->last_qscale_diff = 0;
  1664. tmp = pps->init_qp + (unsigned)get_se_golomb(&sl->gb);
  1665. if (tmp > 51 + 6 * (sps->bit_depth_luma - 8)) {
  1666. av_log(h->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
  1667. return AVERROR_INVALIDDATA;
  1668. }
  1669. sl->qscale = tmp;
  1670. sl->chroma_qp[0] = get_chroma_qp(pps, 0, sl->qscale);
  1671. sl->chroma_qp[1] = get_chroma_qp(pps, 1, sl->qscale);
  1672. // FIXME qscale / qp ... stuff
  1673. if (sl->slice_type == AV_PICTURE_TYPE_SP)
  1674. get_bits1(&sl->gb); /* sp_for_switch_flag */
  1675. if (sl->slice_type == AV_PICTURE_TYPE_SP ||
  1676. sl->slice_type == AV_PICTURE_TYPE_SI)
  1677. get_se_golomb(&sl->gb); /* slice_qs_delta */
  1678. sl->deblocking_filter = 1;
  1679. sl->slice_alpha_c0_offset = 0;
  1680. sl->slice_beta_offset = 0;
  1681. if (pps->deblocking_filter_parameters_present) {
  1682. tmp = get_ue_golomb_31(&sl->gb);
  1683. if (tmp > 2) {
  1684. av_log(h->avctx, AV_LOG_ERROR,
  1685. "deblocking_filter_idc %u out of range\n", tmp);
  1686. return AVERROR_INVALIDDATA;
  1687. }
  1688. sl->deblocking_filter = tmp;
  1689. if (sl->deblocking_filter < 2)
  1690. sl->deblocking_filter ^= 1; // 1<->0
  1691. if (sl->deblocking_filter) {
  1692. int slice_alpha_c0_offset_div2 = get_se_golomb(&sl->gb);
  1693. int slice_beta_offset_div2 = get_se_golomb(&sl->gb);
  1694. if (slice_alpha_c0_offset_div2 > 6 ||
  1695. slice_alpha_c0_offset_div2 < -6 ||
  1696. slice_beta_offset_div2 > 6 ||
  1697. slice_beta_offset_div2 < -6) {
  1698. av_log(h->avctx, AV_LOG_ERROR,
  1699. "deblocking filter parameters %d %d out of range\n",
  1700. slice_alpha_c0_offset_div2, slice_beta_offset_div2);
  1701. return AVERROR_INVALIDDATA;
  1702. }
  1703. sl->slice_alpha_c0_offset = slice_alpha_c0_offset_div2 * 2;
  1704. sl->slice_beta_offset = slice_beta_offset_div2 * 2;
  1705. }
  1706. }
  1707. return 0;
  1708. }
  1709. /* do all the per-slice initialization needed before we can start decoding the
  1710. * actual MBs */
  1711. static int h264_slice_init(H264Context *h, H264SliceContext *sl,
  1712. const H2645NAL *nal)
  1713. {
  1714. int i, j, ret = 0;
  1715. if (h->picture_idr && nal->type != H264_NAL_IDR_SLICE) {
  1716. av_log(h->avctx, AV_LOG_ERROR, "Invalid mix of IDR and non-IDR slices\n");
  1717. return AVERROR_INVALIDDATA;
  1718. }
  1719. av_assert1(h->mb_num == h->mb_width * h->mb_height);
  1720. if (sl->first_mb_addr << FIELD_OR_MBAFF_PICTURE(h) >= h->mb_num ||
  1721. sl->first_mb_addr >= h->mb_num) {
  1722. av_log(h->avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
  1723. return AVERROR_INVALIDDATA;
  1724. }
  1725. sl->resync_mb_x = sl->mb_x = sl->first_mb_addr % h->mb_width;
  1726. sl->resync_mb_y = sl->mb_y = (sl->first_mb_addr / h->mb_width) <<
  1727. FIELD_OR_MBAFF_PICTURE(h);
  1728. if (h->picture_structure == PICT_BOTTOM_FIELD)
  1729. sl->resync_mb_y = sl->mb_y = sl->mb_y + 1;
  1730. av_assert1(sl->mb_y < h->mb_height);
  1731. ret = ff_h264_build_ref_list(h, sl);
  1732. if (ret < 0)
  1733. return ret;
  1734. if (h->ps.pps->weighted_bipred_idc == 2 &&
  1735. sl->slice_type_nos == AV_PICTURE_TYPE_B) {
  1736. implicit_weight_table(h, sl, -1);
  1737. if (FRAME_MBAFF(h)) {
  1738. implicit_weight_table(h, sl, 0);
  1739. implicit_weight_table(h, sl, 1);
  1740. }
  1741. }
  1742. if (sl->slice_type_nos == AV_PICTURE_TYPE_B && !sl->direct_spatial_mv_pred)
  1743. ff_h264_direct_dist_scale_factor(h, sl);
  1744. if (!h->setup_finished)
  1745. ff_h264_direct_ref_list_init(h, sl);
  1746. if (h->avctx->skip_loop_filter >= AVDISCARD_ALL ||
  1747. (h->avctx->skip_loop_filter >= AVDISCARD_NONKEY &&
  1748. h->nal_unit_type != H264_NAL_IDR_SLICE) ||
  1749. (h->avctx->skip_loop_filter >= AVDISCARD_NONINTRA &&
  1750. sl->slice_type_nos != AV_PICTURE_TYPE_I) ||
  1751. (h->avctx->skip_loop_filter >= AVDISCARD_BIDIR &&
  1752. sl->slice_type_nos == AV_PICTURE_TYPE_B) ||
  1753. (h->avctx->skip_loop_filter >= AVDISCARD_NONREF &&
  1754. nal->ref_idc == 0))
  1755. sl->deblocking_filter = 0;
  1756. if (sl->deblocking_filter == 1 && h->nb_slice_ctx > 1) {
  1757. if (h->avctx->flags2 & AV_CODEC_FLAG2_FAST) {
  1758. /* Cheat slightly for speed:
  1759. * Do not bother to deblock across slices. */
  1760. sl->deblocking_filter = 2;
  1761. } else {
  1762. h->postpone_filter = 1;
  1763. }
  1764. }
  1765. sl->qp_thresh = 15 -
  1766. FFMIN(sl->slice_alpha_c0_offset, sl->slice_beta_offset) -
  1767. FFMAX3(0,
  1768. h->ps.pps->chroma_qp_index_offset[0],
  1769. h->ps.pps->chroma_qp_index_offset[1]) +
  1770. 6 * (h->ps.sps->bit_depth_luma - 8);
  1771. sl->slice_num = ++h->current_slice;
  1772. if (sl->slice_num)
  1773. h->slice_row[(sl->slice_num-1)&(MAX_SLICES-1)]= sl->resync_mb_y;
  1774. if ( h->slice_row[sl->slice_num&(MAX_SLICES-1)] + 3 >= sl->resync_mb_y
  1775. && h->slice_row[sl->slice_num&(MAX_SLICES-1)] <= sl->resync_mb_y
  1776. && sl->slice_num >= MAX_SLICES) {
  1777. //in case of ASO this check needs to be updated depending on how we decide to assign slice numbers in this case
  1778. av_log(h->avctx, AV_LOG_WARNING, "Possibly too many slices (%d >= %d), increase MAX_SLICES and recompile if there are artifacts\n", sl->slice_num, MAX_SLICES);
  1779. }
  1780. for (j = 0; j < 2; j++) {
  1781. int id_list[16];
  1782. int *ref2frm = h->ref2frm[sl->slice_num & (MAX_SLICES - 1)][j];
  1783. for (i = 0; i < 16; i++) {
  1784. id_list[i] = 60;
  1785. if (j < sl->list_count && i < sl->ref_count[j] &&
  1786. sl->ref_list[j][i].parent->f->buf[0]) {
  1787. int k;
  1788. AVBuffer *buf = sl->ref_list[j][i].parent->f->buf[0]->buffer;
  1789. for (k = 0; k < h->short_ref_count; k++)
  1790. if (h->short_ref[k]->f->buf[0]->buffer == buf) {
  1791. id_list[i] = k;
  1792. break;
  1793. }
  1794. for (k = 0; k < h->long_ref_count; k++)
  1795. if (h->long_ref[k] && h->long_ref[k]->f->buf[0]->buffer == buf) {
  1796. id_list[i] = h->short_ref_count + k;
  1797. break;
  1798. }
  1799. }
  1800. }
  1801. ref2frm[0] =
  1802. ref2frm[1] = -1;
  1803. for (i = 0; i < 16; i++)
  1804. ref2frm[i + 2] = 4 * id_list[i] + (sl->ref_list[j][i].reference & 3);
  1805. ref2frm[18 + 0] =
  1806. ref2frm[18 + 1] = -1;
  1807. for (i = 16; i < 48; i++)
  1808. ref2frm[i + 4] = 4 * id_list[(i - 16) >> 1] +
  1809. (sl->ref_list[j][i].reference & 3);
  1810. }
  1811. if (h->avctx->debug & FF_DEBUG_PICT_INFO) {
  1812. av_log(h->avctx, AV_LOG_DEBUG,
  1813. "slice:%d %s mb:%d %c%s%s frame:%d poc:%d/%d ref:%d/%d qp:%d loop:%d:%d:%d weight:%d%s %s\n",
  1814. sl->slice_num,
  1815. (h->picture_structure == PICT_FRAME ? "F" : h->picture_structure == PICT_TOP_FIELD ? "T" : "B"),
  1816. sl->mb_y * h->mb_width + sl->mb_x,
  1817. av_get_picture_type_char(sl->slice_type),
  1818. sl->slice_type_fixed ? " fix" : "",
  1819. nal->type == H264_NAL_IDR_SLICE ? " IDR" : "",
  1820. h->poc.frame_num,
  1821. h->cur_pic_ptr->field_poc[0],
  1822. h->cur_pic_ptr->field_poc[1],
  1823. sl->ref_count[0], sl->ref_count[1],
  1824. sl->qscale,
  1825. sl->deblocking_filter,
  1826. sl->slice_alpha_c0_offset, sl->slice_beta_offset,
  1827. sl->pwt.use_weight,
  1828. sl->pwt.use_weight == 1 && sl->pwt.use_weight_chroma ? "c" : "",
  1829. sl->slice_type == AV_PICTURE_TYPE_B ? (sl->direct_spatial_mv_pred ? "SPAT" : "TEMP") : "");
  1830. }
  1831. return 0;
  1832. }
  1833. int ff_h264_queue_decode_slice(H264Context *h, const H2645NAL *nal)
  1834. {
  1835. H264SliceContext *sl = h->slice_ctx + h->nb_slice_ctx_queued;
  1836. int first_slice = sl == h->slice_ctx && !h->current_slice;
  1837. int ret;
  1838. sl->gb = nal->gb;
  1839. ret = h264_slice_header_parse(h, sl, nal);
  1840. if (ret < 0)
  1841. return ret;
  1842. // discard redundant pictures
  1843. if (sl->redundant_pic_count > 0) {
  1844. sl->ref_count[0] = sl->ref_count[1] = 0;
  1845. return 0;
  1846. }
  1847. if (sl->first_mb_addr == 0 || !h->current_slice) {
  1848. if (h->setup_finished) {
  1849. av_log(h->avctx, AV_LOG_ERROR, "Too many fields\n");
  1850. return AVERROR_INVALIDDATA;
  1851. }
  1852. }
  1853. if (sl->first_mb_addr == 0) { // FIXME better field boundary detection
  1854. if (h->current_slice) {
  1855. // this slice starts a new field
  1856. // first decode any pending queued slices
  1857. if (h->nb_slice_ctx_queued) {
  1858. H264SliceContext tmp_ctx;
  1859. ret = ff_h264_execute_decode_slices(h);
  1860. if (ret < 0 && (h->avctx->err_recognition & AV_EF_EXPLODE))
  1861. return ret;
  1862. memcpy(&tmp_ctx, h->slice_ctx, sizeof(tmp_ctx));
  1863. memcpy(h->slice_ctx, sl, sizeof(tmp_ctx));
  1864. memcpy(sl, &tmp_ctx, sizeof(tmp_ctx));
  1865. sl = h->slice_ctx;
  1866. }
  1867. if (h->cur_pic_ptr && FIELD_PICTURE(h) && h->first_field) {
  1868. ret = ff_h264_field_end(h, h->slice_ctx, 1);
  1869. if (ret < 0)
  1870. return ret;
  1871. } else if (h->cur_pic_ptr && !FIELD_PICTURE(h) && !h->first_field && h->nal_unit_type == H264_NAL_IDR_SLICE) {
  1872. av_log(h, AV_LOG_WARNING, "Broken frame packetizing\n");
  1873. ret = ff_h264_field_end(h, h->slice_ctx, 1);
  1874. ff_thread_report_progress(&h->cur_pic_ptr->tf, INT_MAX, 0);
  1875. ff_thread_report_progress(&h->cur_pic_ptr->tf, INT_MAX, 1);
  1876. h->cur_pic_ptr = NULL;
  1877. if (ret < 0)
  1878. return ret;
  1879. } else
  1880. return AVERROR_INVALIDDATA;
  1881. }
  1882. if (!h->first_field) {
  1883. if (h->cur_pic_ptr && !h->droppable) {
  1884. ff_thread_report_progress(&h->cur_pic_ptr->tf, INT_MAX,
  1885. h->picture_structure == PICT_BOTTOM_FIELD);
  1886. }
  1887. h->cur_pic_ptr = NULL;
  1888. }
  1889. }
  1890. if (!h->current_slice)
  1891. av_assert0(sl == h->slice_ctx);
  1892. if (h->current_slice == 0 && !h->first_field) {
  1893. if (
  1894. (h->avctx->skip_frame >= AVDISCARD_NONREF && !h->nal_ref_idc) ||
  1895. (h->avctx->skip_frame >= AVDISCARD_BIDIR && sl->slice_type_nos == AV_PICTURE_TYPE_B) ||
  1896. (h->avctx->skip_frame >= AVDISCARD_NONINTRA && sl->slice_type_nos != AV_PICTURE_TYPE_I) ||
  1897. (h->avctx->skip_frame >= AVDISCARD_NONKEY && h->nal_unit_type != H264_NAL_IDR_SLICE && h->sei.recovery_point.recovery_frame_cnt < 0) ||
  1898. h->avctx->skip_frame >= AVDISCARD_ALL) {
  1899. return 0;
  1900. }
  1901. }
  1902. if (!first_slice) {
  1903. const PPS *pps = (const PPS*)h->ps.pps_list[sl->pps_id]->data;
  1904. if (h->ps.pps->sps_id != pps->sps_id ||
  1905. h->ps.pps->transform_8x8_mode != pps->transform_8x8_mode /*||
  1906. (h->setup_finished && h->ps.pps != pps)*/) {
  1907. av_log(h->avctx, AV_LOG_ERROR, "PPS changed between slices\n");
  1908. return AVERROR_INVALIDDATA;
  1909. }
  1910. if (h->ps.sps != pps->sps) {
  1911. av_log(h->avctx, AV_LOG_ERROR,
  1912. "SPS changed in the middle of the frame\n");
  1913. return AVERROR_INVALIDDATA;
  1914. }
  1915. }
  1916. if (h->current_slice == 0) {
  1917. ret = h264_field_start(h, sl, nal, first_slice);
  1918. if (ret < 0)
  1919. return ret;
  1920. } else {
  1921. if (h->picture_structure != sl->picture_structure ||
  1922. h->droppable != (nal->ref_idc == 0)) {
  1923. av_log(h->avctx, AV_LOG_ERROR,
  1924. "Changing field mode (%d -> %d) between slices is not allowed\n",
  1925. h->picture_structure, sl->picture_structure);
  1926. return AVERROR_INVALIDDATA;
  1927. } else if (!h->cur_pic_ptr) {
  1928. av_log(h->avctx, AV_LOG_ERROR,
  1929. "unset cur_pic_ptr on slice %d\n",
  1930. h->current_slice + 1);
  1931. return AVERROR_INVALIDDATA;
  1932. }
  1933. }
  1934. ret = h264_slice_init(h, sl, nal);
  1935. if (ret < 0)
  1936. return ret;
  1937. h->nb_slice_ctx_queued++;
  1938. return 0;
  1939. }
  1940. int ff_h264_get_slice_type(const H264SliceContext *sl)
  1941. {
  1942. switch (sl->slice_type) {
  1943. case AV_PICTURE_TYPE_P:
  1944. return 0;
  1945. case AV_PICTURE_TYPE_B:
  1946. return 1;
  1947. case AV_PICTURE_TYPE_I:
  1948. return 2;
  1949. case AV_PICTURE_TYPE_SP:
  1950. return 3;
  1951. case AV_PICTURE_TYPE_SI:
  1952. return 4;
  1953. default:
  1954. return AVERROR_INVALIDDATA;
  1955. }
  1956. }
  1957. static av_always_inline void fill_filter_caches_inter(const H264Context *h,
  1958. H264SliceContext *sl,
  1959. int mb_type, int top_xy,
  1960. int left_xy[LEFT_MBS],
  1961. int top_type,
  1962. int left_type[LEFT_MBS],
  1963. int mb_xy, int list)
  1964. {
  1965. int b_stride = h->b_stride;
  1966. int16_t(*mv_dst)[2] = &sl->mv_cache[list][scan8[0]];
  1967. int8_t *ref_cache = &sl->ref_cache[list][scan8[0]];
  1968. if (IS_INTER(mb_type) || IS_DIRECT(mb_type)) {
  1969. if (USES_LIST(top_type, list)) {
  1970. const int b_xy = h->mb2b_xy[top_xy] + 3 * b_stride;
  1971. const int b8_xy = 4 * top_xy + 2;
  1972. const int *ref2frm = &h->ref2frm[h->slice_table[top_xy] & (MAX_SLICES - 1)][list][(MB_MBAFF(sl) ? 20 : 2)];
  1973. AV_COPY128(mv_dst - 1 * 8, h->cur_pic.motion_val[list][b_xy + 0]);
  1974. ref_cache[0 - 1 * 8] =
  1975. ref_cache[1 - 1 * 8] = ref2frm[h->cur_pic.ref_index[list][b8_xy + 0]];
  1976. ref_cache[2 - 1 * 8] =
  1977. ref_cache[3 - 1 * 8] = ref2frm[h->cur_pic.ref_index[list][b8_xy + 1]];
  1978. } else {
  1979. AV_ZERO128(mv_dst - 1 * 8);
  1980. AV_WN32A(&ref_cache[0 - 1 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
  1981. }
  1982. if (!IS_INTERLACED(mb_type ^ left_type[LTOP])) {
  1983. if (USES_LIST(left_type[LTOP], list)) {
  1984. const int b_xy = h->mb2b_xy[left_xy[LTOP]] + 3;
  1985. const int b8_xy = 4 * left_xy[LTOP] + 1;
  1986. const int *ref2frm = &h->ref2frm[h->slice_table[left_xy[LTOP]] & (MAX_SLICES - 1)][list][(MB_MBAFF(sl) ? 20 : 2)];
  1987. AV_COPY32(mv_dst - 1 + 0, h->cur_pic.motion_val[list][b_xy + b_stride * 0]);
  1988. AV_COPY32(mv_dst - 1 + 8, h->cur_pic.motion_val[list][b_xy + b_stride * 1]);
  1989. AV_COPY32(mv_dst - 1 + 16, h->cur_pic.motion_val[list][b_xy + b_stride * 2]);
  1990. AV_COPY32(mv_dst - 1 + 24, h->cur_pic.motion_val[list][b_xy + b_stride * 3]);
  1991. ref_cache[-1 + 0] =
  1992. ref_cache[-1 + 8] = ref2frm[h->cur_pic.ref_index[list][b8_xy + 2 * 0]];
  1993. ref_cache[-1 + 16] =
  1994. ref_cache[-1 + 24] = ref2frm[h->cur_pic.ref_index[list][b8_xy + 2 * 1]];
  1995. } else {
  1996. AV_ZERO32(mv_dst - 1 + 0);
  1997. AV_ZERO32(mv_dst - 1 + 8);
  1998. AV_ZERO32(mv_dst - 1 + 16);
  1999. AV_ZERO32(mv_dst - 1 + 24);
  2000. ref_cache[-1 + 0] =
  2001. ref_cache[-1 + 8] =
  2002. ref_cache[-1 + 16] =
  2003. ref_cache[-1 + 24] = LIST_NOT_USED;
  2004. }
  2005. }
  2006. }
  2007. if (!USES_LIST(mb_type, list)) {
  2008. fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0, 0), 4);
  2009. AV_WN32A(&ref_cache[0 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
  2010. AV_WN32A(&ref_cache[1 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
  2011. AV_WN32A(&ref_cache[2 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
  2012. AV_WN32A(&ref_cache[3 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
  2013. return;
  2014. }
  2015. {
  2016. int8_t *ref = &h->cur_pic.ref_index[list][4 * mb_xy];
  2017. const int *ref2frm = &h->ref2frm[sl->slice_num & (MAX_SLICES - 1)][list][(MB_MBAFF(sl) ? 20 : 2)];
  2018. uint32_t ref01 = (pack16to32(ref2frm[ref[0]], ref2frm[ref[1]]) & 0x00FF00FF) * 0x0101;
  2019. uint32_t ref23 = (pack16to32(ref2frm[ref[2]], ref2frm[ref[3]]) & 0x00FF00FF) * 0x0101;
  2020. AV_WN32A(&ref_cache[0 * 8], ref01);
  2021. AV_WN32A(&ref_cache[1 * 8], ref01);
  2022. AV_WN32A(&ref_cache[2 * 8], ref23);
  2023. AV_WN32A(&ref_cache[3 * 8], ref23);
  2024. }
  2025. {
  2026. int16_t(*mv_src)[2] = &h->cur_pic.motion_val[list][4 * sl->mb_x + 4 * sl->mb_y * b_stride];
  2027. AV_COPY128(mv_dst + 8 * 0, mv_src + 0 * b_stride);
  2028. AV_COPY128(mv_dst + 8 * 1, mv_src + 1 * b_stride);
  2029. AV_COPY128(mv_dst + 8 * 2, mv_src + 2 * b_stride);
  2030. AV_COPY128(mv_dst + 8 * 3, mv_src + 3 * b_stride);
  2031. }
  2032. }
  2033. /**
  2034. * @return non zero if the loop filter can be skipped
  2035. */
  2036. static int fill_filter_caches(const H264Context *h, H264SliceContext *sl, int mb_type)
  2037. {
  2038. const int mb_xy = sl->mb_xy;
  2039. int top_xy, left_xy[LEFT_MBS];
  2040. int top_type, left_type[LEFT_MBS];
  2041. uint8_t *nnz;
  2042. uint8_t *nnz_cache;
  2043. top_xy = mb_xy - (h->mb_stride << MB_FIELD(sl));
  2044. left_xy[LBOT] = left_xy[LTOP] = mb_xy - 1;
  2045. if (FRAME_MBAFF(h)) {
  2046. const int left_mb_field_flag = IS_INTERLACED(h->cur_pic.mb_type[mb_xy - 1]);
  2047. const int curr_mb_field_flag = IS_INTERLACED(mb_type);
  2048. if (sl->mb_y & 1) {
  2049. if (left_mb_field_flag != curr_mb_field_flag)
  2050. left_xy[LTOP] -= h->mb_stride;
  2051. } else {
  2052. if (curr_mb_field_flag)
  2053. top_xy += h->mb_stride &
  2054. (((h->cur_pic.mb_type[top_xy] >> 7) & 1) - 1);
  2055. if (left_mb_field_flag != curr_mb_field_flag)
  2056. left_xy[LBOT] += h->mb_stride;
  2057. }
  2058. }
  2059. sl->top_mb_xy = top_xy;
  2060. sl->left_mb_xy[LTOP] = left_xy[LTOP];
  2061. sl->left_mb_xy[LBOT] = left_xy[LBOT];
  2062. {
  2063. /* For sufficiently low qp, filtering wouldn't do anything.
  2064. * This is a conservative estimate: could also check beta_offset
  2065. * and more accurate chroma_qp. */
  2066. int qp_thresh = sl->qp_thresh; // FIXME strictly we should store qp_thresh for each mb of a slice
  2067. int qp = h->cur_pic.qscale_table[mb_xy];
  2068. if (qp <= qp_thresh &&
  2069. (left_xy[LTOP] < 0 ||
  2070. ((qp + h->cur_pic.qscale_table[left_xy[LTOP]] + 1) >> 1) <= qp_thresh) &&
  2071. (top_xy < 0 ||
  2072. ((qp + h->cur_pic.qscale_table[top_xy] + 1) >> 1) <= qp_thresh)) {
  2073. if (!FRAME_MBAFF(h))
  2074. return 1;
  2075. if ((left_xy[LTOP] < 0 ||
  2076. ((qp + h->cur_pic.qscale_table[left_xy[LBOT]] + 1) >> 1) <= qp_thresh) &&
  2077. (top_xy < h->mb_stride ||
  2078. ((qp + h->cur_pic.qscale_table[top_xy - h->mb_stride] + 1) >> 1) <= qp_thresh))
  2079. return 1;
  2080. }
  2081. }
  2082. top_type = h->cur_pic.mb_type[top_xy];
  2083. left_type[LTOP] = h->cur_pic.mb_type[left_xy[LTOP]];
  2084. left_type[LBOT] = h->cur_pic.mb_type[left_xy[LBOT]];
  2085. if (sl->deblocking_filter == 2) {
  2086. if (h->slice_table[top_xy] != sl->slice_num)
  2087. top_type = 0;
  2088. if (h->slice_table[left_xy[LBOT]] != sl->slice_num)
  2089. left_type[LTOP] = left_type[LBOT] = 0;
  2090. } else {
  2091. if (h->slice_table[top_xy] == 0xFFFF)
  2092. top_type = 0;
  2093. if (h->slice_table[left_xy[LBOT]] == 0xFFFF)
  2094. left_type[LTOP] = left_type[LBOT] = 0;
  2095. }
  2096. sl->top_type = top_type;
  2097. sl->left_type[LTOP] = left_type[LTOP];
  2098. sl->left_type[LBOT] = left_type[LBOT];
  2099. if (IS_INTRA(mb_type))
  2100. return 0;
  2101. fill_filter_caches_inter(h, sl, mb_type, top_xy, left_xy,
  2102. top_type, left_type, mb_xy, 0);
  2103. if (sl->list_count == 2)
  2104. fill_filter_caches_inter(h, sl, mb_type, top_xy, left_xy,
  2105. top_type, left_type, mb_xy, 1);
  2106. nnz = h->non_zero_count[mb_xy];
  2107. nnz_cache = sl->non_zero_count_cache;
  2108. AV_COPY32(&nnz_cache[4 + 8 * 1], &nnz[0]);
  2109. AV_COPY32(&nnz_cache[4 + 8 * 2], &nnz[4]);
  2110. AV_COPY32(&nnz_cache[4 + 8 * 3], &nnz[8]);
  2111. AV_COPY32(&nnz_cache[4 + 8 * 4], &nnz[12]);
  2112. sl->cbp = h->cbp_table[mb_xy];
  2113. if (top_type) {
  2114. nnz = h->non_zero_count[top_xy];
  2115. AV_COPY32(&nnz_cache[4 + 8 * 0], &nnz[3 * 4]);
  2116. }
  2117. if (left_type[LTOP]) {
  2118. nnz = h->non_zero_count[left_xy[LTOP]];
  2119. nnz_cache[3 + 8 * 1] = nnz[3 + 0 * 4];
  2120. nnz_cache[3 + 8 * 2] = nnz[3 + 1 * 4];
  2121. nnz_cache[3 + 8 * 3] = nnz[3 + 2 * 4];
  2122. nnz_cache[3 + 8 * 4] = nnz[3 + 3 * 4];
  2123. }
  2124. /* CAVLC 8x8dct requires NNZ values for residual decoding that differ
  2125. * from what the loop filter needs */
  2126. if (!CABAC(h) && h->ps.pps->transform_8x8_mode) {
  2127. if (IS_8x8DCT(top_type)) {
  2128. nnz_cache[4 + 8 * 0] =
  2129. nnz_cache[5 + 8 * 0] = (h->cbp_table[top_xy] & 0x4000) >> 12;
  2130. nnz_cache[6 + 8 * 0] =
  2131. nnz_cache[7 + 8 * 0] = (h->cbp_table[top_xy] & 0x8000) >> 12;
  2132. }
  2133. if (IS_8x8DCT(left_type[LTOP])) {
  2134. nnz_cache[3 + 8 * 1] =
  2135. nnz_cache[3 + 8 * 2] = (h->cbp_table[left_xy[LTOP]] & 0x2000) >> 12; // FIXME check MBAFF
  2136. }
  2137. if (IS_8x8DCT(left_type[LBOT])) {
  2138. nnz_cache[3 + 8 * 3] =
  2139. nnz_cache[3 + 8 * 4] = (h->cbp_table[left_xy[LBOT]] & 0x8000) >> 12; // FIXME check MBAFF
  2140. }
  2141. if (IS_8x8DCT(mb_type)) {
  2142. nnz_cache[scan8[0]] =
  2143. nnz_cache[scan8[1]] =
  2144. nnz_cache[scan8[2]] =
  2145. nnz_cache[scan8[3]] = (sl->cbp & 0x1000) >> 12;
  2146. nnz_cache[scan8[0 + 4]] =
  2147. nnz_cache[scan8[1 + 4]] =
  2148. nnz_cache[scan8[2 + 4]] =
  2149. nnz_cache[scan8[3 + 4]] = (sl->cbp & 0x2000) >> 12;
  2150. nnz_cache[scan8[0 + 8]] =
  2151. nnz_cache[scan8[1 + 8]] =
  2152. nnz_cache[scan8[2 + 8]] =
  2153. nnz_cache[scan8[3 + 8]] = (sl->cbp & 0x4000) >> 12;
  2154. nnz_cache[scan8[0 + 12]] =
  2155. nnz_cache[scan8[1 + 12]] =
  2156. nnz_cache[scan8[2 + 12]] =
  2157. nnz_cache[scan8[3 + 12]] = (sl->cbp & 0x8000) >> 12;
  2158. }
  2159. }
  2160. return 0;
  2161. }
  2162. static void loop_filter(const H264Context *h, H264SliceContext *sl, int start_x, int end_x)
  2163. {
  2164. uint8_t *dest_y, *dest_cb, *dest_cr;
  2165. int linesize, uvlinesize, mb_x, mb_y;
  2166. const int end_mb_y = sl->mb_y + FRAME_MBAFF(h);
  2167. const int old_slice_type = sl->slice_type;
  2168. const int pixel_shift = h->pixel_shift;
  2169. const int block_h = 16 >> h->chroma_y_shift;
  2170. if (h->postpone_filter)
  2171. return;
  2172. if (sl->deblocking_filter) {
  2173. for (mb_x = start_x; mb_x < end_x; mb_x++)
  2174. for (mb_y = end_mb_y - FRAME_MBAFF(h); mb_y <= end_mb_y; mb_y++) {
  2175. int mb_xy, mb_type;
  2176. mb_xy = sl->mb_xy = mb_x + mb_y * h->mb_stride;
  2177. mb_type = h->cur_pic.mb_type[mb_xy];
  2178. if (FRAME_MBAFF(h))
  2179. sl->mb_mbaff =
  2180. sl->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
  2181. sl->mb_x = mb_x;
  2182. sl->mb_y = mb_y;
  2183. dest_y = h->cur_pic.f->data[0] +
  2184. ((mb_x << pixel_shift) + mb_y * sl->linesize) * 16;
  2185. dest_cb = h->cur_pic.f->data[1] +
  2186. (mb_x << pixel_shift) * (8 << CHROMA444(h)) +
  2187. mb_y * sl->uvlinesize * block_h;
  2188. dest_cr = h->cur_pic.f->data[2] +
  2189. (mb_x << pixel_shift) * (8 << CHROMA444(h)) +
  2190. mb_y * sl->uvlinesize * block_h;
  2191. // FIXME simplify above
  2192. if (MB_FIELD(sl)) {
  2193. linesize = sl->mb_linesize = sl->linesize * 2;
  2194. uvlinesize = sl->mb_uvlinesize = sl->uvlinesize * 2;
  2195. if (mb_y & 1) { // FIXME move out of this function?
  2196. dest_y -= sl->linesize * 15;
  2197. dest_cb -= sl->uvlinesize * (block_h - 1);
  2198. dest_cr -= sl->uvlinesize * (block_h - 1);
  2199. }
  2200. } else {
  2201. linesize = sl->mb_linesize = sl->linesize;
  2202. uvlinesize = sl->mb_uvlinesize = sl->uvlinesize;
  2203. }
  2204. backup_mb_border(h, sl, dest_y, dest_cb, dest_cr, linesize,
  2205. uvlinesize, 0);
  2206. if (fill_filter_caches(h, sl, mb_type))
  2207. continue;
  2208. sl->chroma_qp[0] = get_chroma_qp(h->ps.pps, 0, h->cur_pic.qscale_table[mb_xy]);
  2209. sl->chroma_qp[1] = get_chroma_qp(h->ps.pps, 1, h->cur_pic.qscale_table[mb_xy]);
  2210. if (FRAME_MBAFF(h)) {
  2211. ff_h264_filter_mb(h, sl, mb_x, mb_y, dest_y, dest_cb, dest_cr,
  2212. linesize, uvlinesize);
  2213. } else {
  2214. ff_h264_filter_mb_fast(h, sl, mb_x, mb_y, dest_y, dest_cb,
  2215. dest_cr, linesize, uvlinesize);
  2216. }
  2217. }
  2218. }
  2219. sl->slice_type = old_slice_type;
  2220. sl->mb_x = end_x;
  2221. sl->mb_y = end_mb_y - FRAME_MBAFF(h);
  2222. sl->chroma_qp[0] = get_chroma_qp(h->ps.pps, 0, sl->qscale);
  2223. sl->chroma_qp[1] = get_chroma_qp(h->ps.pps, 1, sl->qscale);
  2224. }
  2225. static void predict_field_decoding_flag(const H264Context *h, H264SliceContext *sl)
  2226. {
  2227. const int mb_xy = sl->mb_x + sl->mb_y * h->mb_stride;
  2228. int mb_type = (h->slice_table[mb_xy - 1] == sl->slice_num) ?
  2229. h->cur_pic.mb_type[mb_xy - 1] :
  2230. (h->slice_table[mb_xy - h->mb_stride] == sl->slice_num) ?
  2231. h->cur_pic.mb_type[mb_xy - h->mb_stride] : 0;
  2232. sl->mb_mbaff = sl->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
  2233. }
  2234. /**
  2235. * Draw edges and report progress for the last MB row.
  2236. */
  2237. static void decode_finish_row(const H264Context *h, H264SliceContext *sl)
  2238. {
  2239. int top = 16 * (sl->mb_y >> FIELD_PICTURE(h));
  2240. int pic_height = 16 * h->mb_height >> FIELD_PICTURE(h);
  2241. int height = 16 << FRAME_MBAFF(h);
  2242. int deblock_border = (16 + 4) << FRAME_MBAFF(h);
  2243. if (sl->deblocking_filter) {
  2244. if ((top + height) >= pic_height)
  2245. height += deblock_border;
  2246. top -= deblock_border;
  2247. }
  2248. if (top >= pic_height || (top + height) < 0)
  2249. return;
  2250. height = FFMIN(height, pic_height - top);
  2251. if (top < 0) {
  2252. height = top + height;
  2253. top = 0;
  2254. }
  2255. ff_h264_draw_horiz_band(h, sl, top, height);
  2256. if (h->droppable || sl->h264->slice_ctx[0].er.error_occurred)
  2257. return;
  2258. ff_thread_report_progress(&h->cur_pic_ptr->tf, top + height - 1,
  2259. h->picture_structure == PICT_BOTTOM_FIELD);
  2260. }
  2261. static void er_add_slice(H264SliceContext *sl,
  2262. int startx, int starty,
  2263. int endx, int endy, int status)
  2264. {
  2265. if (!sl->h264->enable_er)
  2266. return;
  2267. if (CONFIG_ERROR_RESILIENCE) {
  2268. ERContext *er = &sl->h264->slice_ctx[0].er;
  2269. ff_er_add_slice(er, startx, starty, endx, endy, status);
  2270. }
  2271. }
  2272. static int decode_slice(struct AVCodecContext *avctx, void *arg)
  2273. {
  2274. H264SliceContext *sl = arg;
  2275. const H264Context *h = sl->h264;
  2276. int lf_x_start = sl->mb_x;
  2277. int orig_deblock = sl->deblocking_filter;
  2278. int ret;
  2279. sl->linesize = h->cur_pic_ptr->f->linesize[0];
  2280. sl->uvlinesize = h->cur_pic_ptr->f->linesize[1];
  2281. ret = alloc_scratch_buffers(sl, sl->linesize);
  2282. if (ret < 0)
  2283. return ret;
  2284. sl->mb_skip_run = -1;
  2285. av_assert0(h->block_offset[15] == (4 * ((scan8[15] - scan8[0]) & 7) << h->pixel_shift) + 4 * sl->linesize * ((scan8[15] - scan8[0]) >> 3));
  2286. if (h->postpone_filter)
  2287. sl->deblocking_filter = 0;
  2288. sl->is_complex = FRAME_MBAFF(h) || h->picture_structure != PICT_FRAME ||
  2289. (CONFIG_GRAY && (h->flags & AV_CODEC_FLAG_GRAY));
  2290. if (!(h->avctx->active_thread_type & FF_THREAD_SLICE) && h->picture_structure == PICT_FRAME && h->slice_ctx[0].er.error_status_table) {
  2291. const int start_i = av_clip(sl->resync_mb_x + sl->resync_mb_y * h->mb_width, 0, h->mb_num - 1);
  2292. if (start_i) {
  2293. int prev_status = h->slice_ctx[0].er.error_status_table[h->slice_ctx[0].er.mb_index2xy[start_i - 1]];
  2294. prev_status &= ~ VP_START;
  2295. if (prev_status != (ER_MV_END | ER_DC_END | ER_AC_END))
  2296. h->slice_ctx[0].er.error_occurred = 1;
  2297. }
  2298. }
  2299. if (h->ps.pps->cabac) {
  2300. /* realign */
  2301. align_get_bits(&sl->gb);
  2302. /* init cabac */
  2303. ret = ff_init_cabac_decoder(&sl->cabac,
  2304. sl->gb.buffer + get_bits_count(&sl->gb) / 8,
  2305. (get_bits_left(&sl->gb) + 7) / 8);
  2306. if (ret < 0)
  2307. return ret;
  2308. ff_h264_init_cabac_states(h, sl);
  2309. for (;;) {
  2310. int ret, eos;
  2311. if (sl->mb_x + sl->mb_y * h->mb_width >= sl->next_slice_idx) {
  2312. av_log(h->avctx, AV_LOG_ERROR, "Slice overlaps with next at %d\n",
  2313. sl->next_slice_idx);
  2314. er_add_slice(sl, sl->resync_mb_x, sl->resync_mb_y, sl->mb_x,
  2315. sl->mb_y, ER_MB_ERROR);
  2316. return AVERROR_INVALIDDATA;
  2317. }
  2318. ret = ff_h264_decode_mb_cabac(h, sl);
  2319. if (ret >= 0)
  2320. ff_h264_hl_decode_mb(h, sl);
  2321. // FIXME optimal? or let mb_decode decode 16x32 ?
  2322. if (ret >= 0 && FRAME_MBAFF(h)) {
  2323. sl->mb_y++;
  2324. ret = ff_h264_decode_mb_cabac(h, sl);
  2325. if (ret >= 0)
  2326. ff_h264_hl_decode_mb(h, sl);
  2327. sl->mb_y--;
  2328. }
  2329. eos = get_cabac_terminate(&sl->cabac);
  2330. if ((h->workaround_bugs & FF_BUG_TRUNCATED) &&
  2331. sl->cabac.bytestream > sl->cabac.bytestream_end + 2) {
  2332. er_add_slice(sl, sl->resync_mb_x, sl->resync_mb_y, sl->mb_x - 1,
  2333. sl->mb_y, ER_MB_END);
  2334. if (sl->mb_x >= lf_x_start)
  2335. loop_filter(h, sl, lf_x_start, sl->mb_x + 1);
  2336. goto finish;
  2337. }
  2338. if (sl->cabac.bytestream > sl->cabac.bytestream_end + 2 )
  2339. av_log(h->avctx, AV_LOG_DEBUG, "bytestream overread %"PTRDIFF_SPECIFIER"\n", sl->cabac.bytestream_end - sl->cabac.bytestream);
  2340. if (ret < 0 || sl->cabac.bytestream > sl->cabac.bytestream_end + 4) {
  2341. av_log(h->avctx, AV_LOG_ERROR,
  2342. "error while decoding MB %d %d, bytestream %"PTRDIFF_SPECIFIER"\n",
  2343. sl->mb_x, sl->mb_y,
  2344. sl->cabac.bytestream_end - sl->cabac.bytestream);
  2345. er_add_slice(sl, sl->resync_mb_x, sl->resync_mb_y, sl->mb_x,
  2346. sl->mb_y, ER_MB_ERROR);
  2347. return AVERROR_INVALIDDATA;
  2348. }
  2349. if (++sl->mb_x >= h->mb_width) {
  2350. loop_filter(h, sl, lf_x_start, sl->mb_x);
  2351. sl->mb_x = lf_x_start = 0;
  2352. decode_finish_row(h, sl);
  2353. ++sl->mb_y;
  2354. if (FIELD_OR_MBAFF_PICTURE(h)) {
  2355. ++sl->mb_y;
  2356. if (FRAME_MBAFF(h) && sl->mb_y < h->mb_height)
  2357. predict_field_decoding_flag(h, sl);
  2358. }
  2359. }
  2360. if (eos || sl->mb_y >= h->mb_height) {
  2361. ff_tlog(h->avctx, "slice end %d %d\n",
  2362. get_bits_count(&sl->gb), sl->gb.size_in_bits);
  2363. er_add_slice(sl, sl->resync_mb_x, sl->resync_mb_y, sl->mb_x - 1,
  2364. sl->mb_y, ER_MB_END);
  2365. if (sl->mb_x > lf_x_start)
  2366. loop_filter(h, sl, lf_x_start, sl->mb_x);
  2367. goto finish;
  2368. }
  2369. }
  2370. } else {
  2371. for (;;) {
  2372. int ret;
  2373. if (sl->mb_x + sl->mb_y * h->mb_width >= sl->next_slice_idx) {
  2374. av_log(h->avctx, AV_LOG_ERROR, "Slice overlaps with next at %d\n",
  2375. sl->next_slice_idx);
  2376. er_add_slice(sl, sl->resync_mb_x, sl->resync_mb_y, sl->mb_x,
  2377. sl->mb_y, ER_MB_ERROR);
  2378. return AVERROR_INVALIDDATA;
  2379. }
  2380. ret = ff_h264_decode_mb_cavlc(h, sl);
  2381. if (ret >= 0)
  2382. ff_h264_hl_decode_mb(h, sl);
  2383. // FIXME optimal? or let mb_decode decode 16x32 ?
  2384. if (ret >= 0 && FRAME_MBAFF(h)) {
  2385. sl->mb_y++;
  2386. ret = ff_h264_decode_mb_cavlc(h, sl);
  2387. if (ret >= 0)
  2388. ff_h264_hl_decode_mb(h, sl);
  2389. sl->mb_y--;
  2390. }
  2391. if (ret < 0) {
  2392. av_log(h->avctx, AV_LOG_ERROR,
  2393. "error while decoding MB %d %d\n", sl->mb_x, sl->mb_y);
  2394. er_add_slice(sl, sl->resync_mb_x, sl->resync_mb_y, sl->mb_x,
  2395. sl->mb_y, ER_MB_ERROR);
  2396. return ret;
  2397. }
  2398. if (++sl->mb_x >= h->mb_width) {
  2399. loop_filter(h, sl, lf_x_start, sl->mb_x);
  2400. sl->mb_x = lf_x_start = 0;
  2401. decode_finish_row(h, sl);
  2402. ++sl->mb_y;
  2403. if (FIELD_OR_MBAFF_PICTURE(h)) {
  2404. ++sl->mb_y;
  2405. if (FRAME_MBAFF(h) && sl->mb_y < h->mb_height)
  2406. predict_field_decoding_flag(h, sl);
  2407. }
  2408. if (sl->mb_y >= h->mb_height) {
  2409. ff_tlog(h->avctx, "slice end %d %d\n",
  2410. get_bits_count(&sl->gb), sl->gb.size_in_bits);
  2411. if ( get_bits_left(&sl->gb) == 0
  2412. || get_bits_left(&sl->gb) > 0 && !(h->avctx->err_recognition & AV_EF_AGGRESSIVE)) {
  2413. er_add_slice(sl, sl->resync_mb_x, sl->resync_mb_y,
  2414. sl->mb_x - 1, sl->mb_y, ER_MB_END);
  2415. goto finish;
  2416. } else {
  2417. er_add_slice(sl, sl->resync_mb_x, sl->resync_mb_y,
  2418. sl->mb_x, sl->mb_y, ER_MB_END);
  2419. return AVERROR_INVALIDDATA;
  2420. }
  2421. }
  2422. }
  2423. if (get_bits_left(&sl->gb) <= 0 && sl->mb_skip_run <= 0) {
  2424. ff_tlog(h->avctx, "slice end %d %d\n",
  2425. get_bits_count(&sl->gb), sl->gb.size_in_bits);
  2426. if (get_bits_left(&sl->gb) == 0) {
  2427. er_add_slice(sl, sl->resync_mb_x, sl->resync_mb_y,
  2428. sl->mb_x - 1, sl->mb_y, ER_MB_END);
  2429. if (sl->mb_x > lf_x_start)
  2430. loop_filter(h, sl, lf_x_start, sl->mb_x);
  2431. goto finish;
  2432. } else {
  2433. er_add_slice(sl, sl->resync_mb_x, sl->resync_mb_y, sl->mb_x,
  2434. sl->mb_y, ER_MB_ERROR);
  2435. return AVERROR_INVALIDDATA;
  2436. }
  2437. }
  2438. }
  2439. }
  2440. finish:
  2441. sl->deblocking_filter = orig_deblock;
  2442. return 0;
  2443. }
  2444. /**
  2445. * Call decode_slice() for each context.
  2446. *
  2447. * @param h h264 master context
  2448. */
  2449. int ff_h264_execute_decode_slices(H264Context *h)
  2450. {
  2451. AVCodecContext *const avctx = h->avctx;
  2452. H264SliceContext *sl;
  2453. int context_count = h->nb_slice_ctx_queued;
  2454. int ret = 0;
  2455. int i, j;
  2456. h->slice_ctx[0].next_slice_idx = INT_MAX;
  2457. if (h->avctx->hwaccel || context_count < 1)
  2458. return 0;
  2459. av_assert0(context_count && h->slice_ctx[context_count - 1].mb_y < h->mb_height);
  2460. if (context_count == 1) {
  2461. h->slice_ctx[0].next_slice_idx = h->mb_width * h->mb_height;
  2462. h->postpone_filter = 0;
  2463. ret = decode_slice(avctx, &h->slice_ctx[0]);
  2464. h->mb_y = h->slice_ctx[0].mb_y;
  2465. if (ret < 0)
  2466. goto finish;
  2467. } else {
  2468. av_assert0(context_count > 0);
  2469. for (i = 0; i < context_count; i++) {
  2470. int next_slice_idx = h->mb_width * h->mb_height;
  2471. int slice_idx;
  2472. sl = &h->slice_ctx[i];
  2473. if (CONFIG_ERROR_RESILIENCE) {
  2474. sl->er.error_count = 0;
  2475. }
  2476. /* make sure none of those slices overlap */
  2477. slice_idx = sl->mb_y * h->mb_width + sl->mb_x;
  2478. for (j = 0; j < context_count; j++) {
  2479. H264SliceContext *sl2 = &h->slice_ctx[j];
  2480. int slice_idx2 = sl2->mb_y * h->mb_width + sl2->mb_x;
  2481. if (i == j || slice_idx2 < slice_idx)
  2482. continue;
  2483. next_slice_idx = FFMIN(next_slice_idx, slice_idx2);
  2484. }
  2485. sl->next_slice_idx = next_slice_idx;
  2486. }
  2487. avctx->execute(avctx, decode_slice, h->slice_ctx,
  2488. NULL, context_count, sizeof(h->slice_ctx[0]));
  2489. /* pull back stuff from slices to master context */
  2490. sl = &h->slice_ctx[context_count - 1];
  2491. h->mb_y = sl->mb_y;
  2492. if (CONFIG_ERROR_RESILIENCE) {
  2493. for (i = 1; i < context_count; i++)
  2494. h->slice_ctx[0].er.error_count += h->slice_ctx[i].er.error_count;
  2495. }
  2496. if (h->postpone_filter) {
  2497. h->postpone_filter = 0;
  2498. for (i = 0; i < context_count; i++) {
  2499. int y_end, x_end;
  2500. sl = &h->slice_ctx[i];
  2501. y_end = FFMIN(sl->mb_y + 1, h->mb_height);
  2502. x_end = (sl->mb_y >= h->mb_height) ? h->mb_width : sl->mb_x;
  2503. for (j = sl->resync_mb_y; j < y_end; j += 1 + FIELD_OR_MBAFF_PICTURE(h)) {
  2504. sl->mb_y = j;
  2505. loop_filter(h, sl, j > sl->resync_mb_y ? 0 : sl->resync_mb_x,
  2506. j == y_end - 1 ? x_end : h->mb_width);
  2507. }
  2508. }
  2509. }
  2510. }
  2511. finish:
  2512. h->nb_slice_ctx_queued = 0;
  2513. return ret;
  2514. }