2 * VP8 compatible video decoder
4 * Copyright (C) 2010 David Conrad
5 * Copyright (C) 2010 Ronald S. Bultje
7 * This file is part of FFmpeg.
9 * FFmpeg is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
14 * FFmpeg is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with FFmpeg; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
29 #include "rectangle.h"
34 // todo: make it possible to check for at least (i4x4 or split_mv)
35 // in one op. are others needed?
44 AVCodecContext *avctx;
48 vp8_mc_func put_pixels_tab[3][3][3];
51 uint8_t *edge_emu_buffer;
52 VP56RangeCoder c; ///< header context, includes mb modes and motion vectors
55 int mb_width; /* number of horizontal MB */
56 int mb_height; /* number of vertical MB */
62 int update_last; ///< update VP56_FRAME_PREVIOUS with the current one
63 int update_golden; ///< VP56_FRAME_NONE if not updated, or which frame to copy if so
67 * If this flag is not set, all the probability updates
68 * are discarded after this frame is decoded.
70 int update_probabilities;
73 * All coefficients are contained in separate arith coding contexts.
74 * There can be 1, 2, 4, or 8 of these after the header context.
76 int num_coeff_partitions;
77 VP56RangeCoder coeff_partition[8];
79 VP8Macroblock *macroblocks;
80 VP8Macroblock *macroblocks_base;
83 uint8_t *intra4x4_pred_mode;
84 uint8_t *intra4x4_pred_mode_base;
88 * For coeff decode, we need to know whether the above block had non-zero
89 * coefficients. This means for each macroblock, we need data for 4 luma
90 * blocks, 2 u blocks, 2 v blocks, and the luma dc block, for a total of 9
91 * per macroblock. We keep the last row in top_nnz.
93 uint8_t (*top_nnz)[9];
94 DECLARE_ALIGNED(8, uint8_t, left_nnz)[9];
97 * This is the index plus one of the last non-zero coeff
98 * for each of the blocks in the current macroblock.
100 * 1 -> dc-only (special transform)
101 * 2+-> full transform
103 DECLARE_ALIGNED(16, uint8_t, non_zero_count_cache)[6][4];
104 DECLARE_ALIGNED(16, DCTELEM, block)[6][4][16];
106 int chroma_pred_mode; ///< 8x8c pred mode of the current macroblock
109 int sign_bias[4]; ///< one state [0, 1] per ref frame type
112 * Base parameters for segmentation, i.e. per-macroblock parameters.
113 * These must be kept unchanged even if segmentation is not used for
114 * a frame, since the values persist between interframes.
120 int8_t base_quant[4];
121 int8_t filter_level[4]; ///< base loop filter level
125 * Macroblocks can have one of 4 different quants in a frame when
126 * segmentation is enabled.
127 * If segmentation is disabled, only the first segment's values are used.
130 // [0] - DC qmul [1] - AC qmul
131 int16_t luma_qmul[2];
132 int16_t luma_dc_qmul[2]; ///< luma dc-only block quant
133 int16_t chroma_qmul[2];
143 int enabled; ///< whether each mb can have a different strength based on mode/ref
146 * filter strength adjustment for the following macroblock modes:
149 * [2] - inter modes except for zero or split mv
151 * i16x16 modes never have any adjustment
156 * filter strength adjustment for macroblocks that reference:
157 * [0] - intra / VP56_FRAME_CURRENT
158 * [1] - VP56_FRAME_PREVIOUS
159 * [2] - VP56_FRAME_GOLDEN
160 * [3] - altref / VP56_FRAME_GOLDEN2
166 * These are all of the updatable probabilities for binary decisions.
167 * They are only implictly reset on keyframes, making it quite likely
168 * for an interframe to desync if a prior frame's header was corrupt
169 * or missing outright!
172 uint8_t segmentid[3];
177 uint8_t pred16x16[4];
179 uint8_t token[4][8][3][NUM_DCT_TOKENS-1];
184 #define RL24(p) (AV_RL16(p) + ((p)[2] << 16))
186 static void vp8_decode_flush(AVCodecContext *avctx)
188 VP8Context *s = avctx->priv_data;
191 for (i = 0; i < 4; i++)
192 if (s->frames[i].data[0])
193 avctx->release_buffer(avctx, &s->frames[i]);
194 memset(s->framep, 0, sizeof(s->framep));
196 av_freep(&s->macroblocks_base);
197 av_freep(&s->intra4x4_pred_mode_base);
198 av_freep(&s->top_nnz);
199 av_freep(&s->edge_emu_buffer);
201 s->macroblocks = NULL;
202 s->intra4x4_pred_mode = NULL;
205 static int update_dimensions(VP8Context *s, int width, int height)
209 if (avcodec_check_dimensions(s->avctx, width, height))
210 return AVERROR_INVALIDDATA;
212 vp8_decode_flush(s->avctx);
214 avcodec_set_dimensions(s->avctx, width, height);
216 s->mb_width = (s->avctx->coded_width +15) / 16;
217 s->mb_height = (s->avctx->coded_height+15) / 16;
219 // we allocate a border around the top/left of intra4x4 modes
220 // this is 4 blocks for intra4x4 to keep 4-byte alignment for fill_rectangle
221 s->mb_stride = s->mb_width+1;
222 s->b4_stride = 4*s->mb_stride;
224 s->macroblocks_base = av_mallocz(s->mb_stride*(s->mb_height+1)*sizeof(*s->macroblocks));
225 s->intra4x4_pred_mode_base = av_mallocz(s->b4_stride*(4*s->mb_height+1));
226 s->top_nnz = av_mallocz(s->mb_width*sizeof(*s->top_nnz));
228 s->macroblocks = s->macroblocks_base + 1 + s->mb_stride;
229 s->intra4x4_pred_mode = s->intra4x4_pred_mode_base + 4 + s->b4_stride;
231 memset(s->intra4x4_pred_mode_base, DC_PRED, s->b4_stride);
232 for (i = 0; i < 4*s->mb_height; i++)
233 s->intra4x4_pred_mode[i*s->b4_stride-1] = DC_PRED;
238 static void parse_segment_info(VP8Context *s)
240 VP56RangeCoder *c = &s->c;
243 s->segmentation.update_map = vp8_rac_get(c);
245 if (vp8_rac_get(c)) { // update segment feature data
246 s->segmentation.absolute_vals = vp8_rac_get(c);
248 for (i = 0; i < 4; i++)
249 s->segmentation.base_quant[i] = vp8_rac_get_sint(c, 7);
251 for (i = 0; i < 4; i++)
252 s->segmentation.filter_level[i] = vp8_rac_get_sint(c, 6);
254 if (s->segmentation.update_map)
255 for (i = 0; i < 3; i++)
256 s->prob->segmentid[i] = vp8_rac_get(c) ? vp8_rac_get_uint(c, 8) : 255;
259 static void update_lf_deltas(VP8Context *s)
261 VP56RangeCoder *c = &s->c;
264 for (i = 0; i < 4; i++)
265 s->lf_delta.ref[i] = vp8_rac_get_sint(c, 6);
267 for (i = 0; i < 4; i++)
268 s->lf_delta.mode[i] = vp8_rac_get_sint(c, 6);
271 static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size)
273 const uint8_t *sizes = buf;
276 s->num_coeff_partitions = 1 << vp8_rac_get_uint(&s->c, 2);
278 buf += 3*(s->num_coeff_partitions-1);
279 buf_size -= 3*(s->num_coeff_partitions-1);
283 for (i = 0; i < s->num_coeff_partitions-1; i++) {
284 int size = RL24(sizes + 3*i);
285 if (buf_size - size < 0)
288 vp56_init_range_decoder(&s->coeff_partition[i], buf, size);
292 vp56_init_range_decoder(&s->coeff_partition[i], buf, buf_size);
297 static void get_quants(VP8Context *s)
299 VP56RangeCoder *c = &s->c;
302 int yac_qi = vp8_rac_get_uint(c, 7);
303 int ydc_delta = vp8_rac_get_sint(c, 4);
304 int y2dc_delta = vp8_rac_get_sint(c, 4);
305 int y2ac_delta = vp8_rac_get_sint(c, 4);
306 int uvdc_delta = vp8_rac_get_sint(c, 4);
307 int uvac_delta = vp8_rac_get_sint(c, 4);
309 for (i = 0; i < 4; i++) {
310 if (s->segmentation.enabled) {
311 base_qi = s->segmentation.base_quant[i];
312 if (!s->segmentation.absolute_vals)
317 s->qmat[i].luma_qmul[0] = vp8_dc_qlookup[av_clip(base_qi + ydc_delta , 0, 127)];
318 s->qmat[i].luma_qmul[1] = vp8_ac_qlookup[av_clip(base_qi , 0, 127)];
319 s->qmat[i].luma_dc_qmul[0] = 2 * vp8_dc_qlookup[av_clip(base_qi + y2dc_delta, 0, 127)];
320 s->qmat[i].luma_dc_qmul[1] = 155 * vp8_ac_qlookup[av_clip(base_qi + y2ac_delta, 0, 127)] / 100;
321 s->qmat[i].chroma_qmul[0] = vp8_dc_qlookup[av_clip(base_qi + uvdc_delta, 0, 127)];
322 s->qmat[i].chroma_qmul[1] = vp8_ac_qlookup[av_clip(base_qi + uvac_delta, 0, 127)];
324 s->qmat[i].luma_dc_qmul[1] = FFMAX(s->qmat[i].luma_dc_qmul[1], 8);
325 s->qmat[i].chroma_qmul[0] = FFMIN(s->qmat[i].chroma_qmul[0], 132);
330 * Determine which buffers golden and altref should be updated with after this frame.
331 * The spec isn't clear here, so I'm going by my understanding of what libvpx does
333 * Intra frames update all 3 references
334 * Inter frames update VP56_FRAME_PREVIOUS if the update_last flag is set
335 * If the update (golden|altref) flag is set, it's updated with the current frame
336 * if update_last is set, and VP56_FRAME_PREVIOUS otherwise.
337 * If the flag is not set, the number read means:
339 * 1: VP56_FRAME_PREVIOUS
340 * 2: update golden with altref, or update altref with golden
342 static VP56Frame ref_to_update(VP8Context *s, int update, VP56Frame ref)
344 VP56RangeCoder *c = &s->c;
347 return VP56_FRAME_CURRENT;
349 switch (vp8_rac_get_uint(c, 2)) {
351 return VP56_FRAME_PREVIOUS;
353 return (ref == VP56_FRAME_GOLDEN) ? VP56_FRAME_GOLDEN2 : VP56_FRAME_GOLDEN;
355 return VP56_FRAME_NONE;
358 static void update_refs(VP8Context *s)
360 VP56RangeCoder *c = &s->c;
362 int update_golden = vp8_rac_get(c);
363 int update_altref = vp8_rac_get(c);
365 s->update_golden = ref_to_update(s, update_golden, VP56_FRAME_GOLDEN);
366 s->update_altref = ref_to_update(s, update_altref, VP56_FRAME_GOLDEN2);
369 static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)
371 VP56RangeCoder *c = &s->c;
372 int header_size, hscale, vscale, i, j, k, l, ret;
373 int width = s->avctx->width;
374 int height = s->avctx->height;
376 s->keyframe = !(buf[0] & 1);
377 s->profile = (buf[0]>>1) & 7;
378 s->invisible = !(buf[0] & 0x10);
379 header_size = RL24(buf) >> 5;
384 av_log(s->avctx, AV_LOG_WARNING, "Unknown profile %d\n", s->profile);
387 memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_epel_pixels_tab, sizeof(s->put_pixels_tab));
388 else // profile 1-3 use bilinear, 4+ aren't defined so whatever
389 memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_bilinear_pixels_tab, sizeof(s->put_pixels_tab));
391 if (header_size > buf_size - 7*s->keyframe) {
392 av_log(s->avctx, AV_LOG_ERROR, "Header size larger than data provided\n");
393 return AVERROR_INVALIDDATA;
397 if (RL24(buf) != 0x2a019d) {
398 av_log(s->avctx, AV_LOG_ERROR, "Invalid start code 0x%x\n", RL24(buf));
399 return AVERROR_INVALIDDATA;
401 width = AV_RL16(buf+3) & 0x3fff;
402 height = AV_RL16(buf+5) & 0x3fff;
403 hscale = buf[4] >> 6;
404 vscale = buf[6] >> 6;
408 if (hscale || vscale)
409 av_log_missing_feature(s->avctx, "Upscaling", 1);
411 s->update_golden = s->update_altref = VP56_FRAME_CURRENT;
412 memcpy(s->prob->token , vp8_token_default_probs , sizeof(s->prob->token));
413 memcpy(s->prob->pred16x16, vp8_pred16x16_prob_inter, sizeof(s->prob->pred16x16));
414 memcpy(s->prob->pred8x8c , vp8_pred8x8c_prob_inter , sizeof(s->prob->pred8x8c));
415 memcpy(s->prob->mvc , vp8_mv_default_prob , sizeof(s->prob->mvc));
416 memset(&s->segmentation, 0, sizeof(s->segmentation));
419 if (!s->macroblocks_base || /* first frame */
420 width != s->avctx->width || height != s->avctx->height) {
421 if ((ret = update_dimensions(s, width, height) < 0))
425 vp56_init_range_decoder(c, buf, header_size);
427 buf_size -= header_size;
431 av_log(s->avctx, AV_LOG_WARNING, "Unspecified colorspace\n");
432 vp8_rac_get(c); // whether we can skip clamping in dsp functions
435 if ((s->segmentation.enabled = vp8_rac_get(c)))
436 parse_segment_info(s);
438 s->segmentation.update_map = 0; // FIXME: move this to some init function?
440 s->filter.simple = vp8_rac_get(c);
441 s->filter.level = vp8_rac_get_uint(c, 6);
442 s->filter.sharpness = vp8_rac_get_uint(c, 3);
444 if ((s->lf_delta.enabled = vp8_rac_get(c)))
448 if (setup_partitions(s, buf, buf_size)) {
449 av_log(s->avctx, AV_LOG_ERROR, "Invalid partitions\n");
450 return AVERROR_INVALIDDATA;
457 s->sign_bias[VP56_FRAME_GOLDEN] = vp8_rac_get(c);
458 s->sign_bias[VP56_FRAME_GOLDEN2 /* altref */] = vp8_rac_get(c);
461 // if we aren't saving this frame's probabilities for future frames,
462 // make a copy of the current probabilities
463 if (!(s->update_probabilities = vp8_rac_get(c)))
464 s->prob[1] = s->prob[0];
466 s->update_last = s->keyframe || vp8_rac_get(c);
468 for (i = 0; i < 4; i++)
469 for (j = 0; j < 8; j++)
470 for (k = 0; k < 3; k++)
471 for (l = 0; l < NUM_DCT_TOKENS-1; l++)
472 if (vp56_rac_get_prob(c, vp8_token_update_probs[i][j][k][l]))
473 s->prob->token[i][j][k][l] = vp8_rac_get_uint(c, 8);
475 if ((s->mbskip_enabled = vp8_rac_get(c)))
476 s->prob->mbskip = vp8_rac_get_uint(c, 8);
479 s->prob->intra = vp8_rac_get_uint(c, 8);
480 s->prob->last = vp8_rac_get_uint(c, 8);
481 s->prob->golden = vp8_rac_get_uint(c, 8);
484 for (i = 0; i < 4; i++)
485 s->prob->pred16x16[i] = vp8_rac_get_uint(c, 8);
487 for (i = 0; i < 3; i++)
488 s->prob->pred8x8c[i] = vp8_rac_get_uint(c, 8);
490 // 17.2 MV probability update
491 for (i = 0; i < 2; i++)
492 for (j = 0; j < 19; j++)
493 if (vp56_rac_get_prob(c, vp8_mv_update_prob[i][j]))
494 s->prob->mvc[i][j] = vp8_rac_get_nn(c);
500 static inline void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src,
503 #define MARGIN (16 << 2)
504 dst->x = av_clip(src->x, -((mb_x << 6) + MARGIN),
505 ((s->mb_width - 1 - mb_x) << 6) + MARGIN);
506 dst->y = av_clip(src->y, -((mb_y << 6) + MARGIN),
507 ((s->mb_height - 1 - mb_y) << 6) + MARGIN);
510 static void find_near_mvs(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
511 VP56mv near[2], VP56mv *best, int cnt[4])
513 VP8Macroblock *mb_edge[3] = { mb - s->mb_stride /* top */,
515 mb - s->mb_stride - 1 /* top-left */ };
516 enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT };
517 VP56mv near_mv[4] = {{ 0 }};
518 enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV };
519 int idx = CNT_ZERO, n;
520 int best_idx = CNT_ZERO;
522 /* Process MB on top, left and top-left */
523 for (n = 0; n < 3; n++) {
524 VP8Macroblock *edge = mb_edge[n];
525 if (edge->ref_frame != VP56_FRAME_CURRENT) {
526 if (edge->mv.x | edge->mv.y) {
527 VP56mv tmp = edge->mv;
528 if (s->sign_bias[mb->ref_frame] != s->sign_bias[edge->ref_frame]) {
532 if ((tmp.x ^ near_mv[idx].x) | (tmp.y ^ near_mv[idx].y))
533 near_mv[++idx] = tmp;
534 cnt[idx] += 1 + (n != 2);
536 cnt[CNT_ZERO] += 1 + (n != 2);
540 /* If we have three distinct MV's, merge first and last if they're the same */
541 if (cnt[CNT_SPLITMV] &&
542 !((near_mv[1+EDGE_TOP].x ^ near_mv[1+EDGE_TOPLEFT].x) |
543 (near_mv[1+EDGE_TOP].y ^ near_mv[1+EDGE_TOPLEFT].y)))
544 cnt[CNT_NEAREST] += 1;
546 cnt[CNT_SPLITMV] = ((mb_edge[EDGE_LEFT]->mode == VP8_MVMODE_SPLIT) +
547 (mb_edge[EDGE_TOP]->mode == VP8_MVMODE_SPLIT)) * 2 +
548 (mb_edge[EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT);
550 /* Swap near and nearest if necessary */
551 if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) {
552 FFSWAP(int, cnt[CNT_NEAREST], cnt[CNT_NEAR]);
553 FFSWAP(VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]);
556 /* Choose the best mv out of 0,0 and the nearest mv */
557 if (cnt[CNT_NEAREST] >= cnt[CNT_ZERO])
558 best_idx = CNT_NEAREST;
560 clamp_mv(s, best, &near_mv[best_idx], mb_x, mb_y);
561 near[0] = near_mv[CNT_NEAREST];
562 near[1] = near_mv[CNT_NEAR];
566 * Motion vector coding, 17.1.
568 static int read_mv_component(VP56RangeCoder *c, const uint8_t *p)
572 if (vp56_rac_get_prob(c, p[0])) {
575 for (i = 0; i < 3; i++)
576 x += vp56_rac_get_prob(c, p[9 + i]) << i;
577 for (i = 9; i > 3; i--)
578 x += vp56_rac_get_prob(c, p[9 + i]) << i;
579 if (!(x & 0xFFF0) || vp56_rac_get_prob(c, p[12]))
582 x = vp8_rac_get_tree(c, vp8_small_mvtree, &p[2]);
584 return (x && vp56_rac_get_prob(c, p[1])) ? -x : x;
587 static const uint8_t *get_submv_prob(const VP56mv *left, const VP56mv *top)
589 int l_is_zero = !(left->x | left->y);
590 int t_is_zero = !(top->x | top->y);
591 int equal = !((left->x ^ top->x) | (left->y ^ top->y));
594 return l_is_zero ? vp8_submv_prob[4] : vp8_submv_prob[3];
596 return vp8_submv_prob[2];
597 return l_is_zero ? vp8_submv_prob[1] : vp8_submv_prob[0];
601 * Split motion vector prediction, 16.4.
602 * @returns the number of motion vectors parsed (2, 4 or 16)
604 static int decode_splitmvs(VP8Context *s, VP56RangeCoder *c,
605 VP8Macroblock *mb, VP56mv *base_mv)
607 int part_idx = mb->partitioning =
608 vp8_rac_get_tree(c, vp8_mbsplit_tree, vp8_mbsplit_prob);
609 int n, num = vp8_mbsplit_count[part_idx];
610 const uint8_t *mbsplits = vp8_mbsplits[part_idx],
611 *firstidx = vp8_mbfirstidx[part_idx];
613 for (n = 0; n < num; n++) {
615 const VP56mv *left, *above;
616 const uint8_t *submv_prob;
619 VP8Macroblock *left_mb = &mb[-1];
620 left = &left_mb->bmv[vp8_mbsplits[left_mb->partitioning][k + 3]];
622 left = &mb->bmv[mbsplits[k - 1]];
624 VP8Macroblock *above_mb = &mb[-s->mb_stride];
625 above = &above_mb->bmv[vp8_mbsplits[above_mb->partitioning][k + 12]];
627 above = &mb->bmv[mbsplits[k - 4]];
629 submv_prob = get_submv_prob(left, above);
631 switch (vp8_rac_get_tree(c, vp8_submv_ref_tree, submv_prob)) {
632 case VP8_SUBMVMODE_NEW4X4:
633 mb->bmv[n].y = base_mv->y + read_mv_component(c, s->prob->mvc[0]);
634 mb->bmv[n].x = base_mv->x + read_mv_component(c, s->prob->mvc[1]);
636 case VP8_SUBMVMODE_ZERO4X4:
640 case VP8_SUBMVMODE_LEFT4X4:
643 case VP8_SUBMVMODE_TOP4X4:
652 static inline void decode_intra4x4_modes(VP56RangeCoder *c, uint8_t *intra4x4,
653 int stride, int keyframe)
656 const uint8_t *ctx = vp8_pred4x4_prob_inter;
658 for (y = 0; y < 4; y++) {
659 for (x = 0; x < 4; x++) {
661 t = intra4x4[x - stride];
663 ctx = vp8_pred4x4_prob_intra[t][l];
665 intra4x4[x] = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx);
671 static void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
674 VP56RangeCoder *c = &s->c;
677 if (s->segmentation.update_map)
678 mb->segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid);
680 mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0;
683 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra);
685 if (mb->mode == MODE_I4x4) {
686 decode_intra4x4_modes(c, intra4x4, s->b4_stride, 1);
688 fill_rectangle(intra4x4, 4, 4, s->b4_stride, vp8_pred4x4_mode[mb->mode], 1);
690 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra);
691 mb->ref_frame = VP56_FRAME_CURRENT;
692 } else if (vp56_rac_get_prob(c, s->prob->intra)) {
693 VP56mv near[2], best;
698 if (vp56_rac_get_prob(c, s->prob->last))
699 mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ?
700 VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN;
702 mb->ref_frame = VP56_FRAME_PREVIOUS;
704 // motion vectors, 16.3
705 find_near_mvs(s, mb, mb_x, mb_y, near, &best, cnt);
706 for (n = 0; n < 4; n++)
707 p[n] = vp8_mode_contexts[cnt[n]][n];
708 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_mvinter, p);
710 case VP8_MVMODE_SPLIT:
711 mb->mv = mb->bmv[decode_splitmvs(s, c, mb, &best) - 1];
713 case VP8_MVMODE_ZERO:
717 case VP8_MVMODE_NEAREST:
718 clamp_mv(s, &mb->mv, &near[0], mb_x, mb_y);
720 case VP8_MVMODE_NEAR:
721 clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y);
724 mb->mv.y = best.y + read_mv_component(c, s->prob->mvc[0]);
725 mb->mv.x = best.x + read_mv_component(c, s->prob->mvc[1]);
728 if (mb->mode != VP8_MVMODE_SPLIT) {
729 mb->partitioning = VP8_SPLITMVMODE_NONE;
734 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16);
736 if (mb->mode == MODE_I4x4) {
737 decode_intra4x4_modes(c, intra4x4, s->b4_stride, 0);
739 fill_rectangle(intra4x4, 4, 4, s->b4_stride, vp8_pred4x4_mode[mb->mode], 1);
741 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c);
742 mb->ref_frame = VP56_FRAME_CURRENT;
747 * @param c arithmetic bitstream reader context
748 * @param block destination for block coefficients
749 * @param probs probabilities to use when reading trees from the bitstream
750 * @param i initial coeff index, 0 unless a separate DC block is coded
751 * @param zero_nhood the initial prediction context for number of surrounding
752 * all-zero blocks (only left/top, so 0-2)
753 * @param qmul array holding the dc/ac dequant factor at position 0/1
754 * @return 0 if no coeffs were decoded
755 * otherwise, the index of the last coeff decoded plus one
757 static int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16],
758 uint8_t probs[8][3][NUM_DCT_TOKENS-1],
759 int i, int zero_nhood, int16_t qmul[2])
761 int token, nonzero = 0;
764 for (; i < 16; i++) {
765 token = vp8_rac_get_tree_with_offset(c, vp8_coeff_tree, probs[vp8_coeff_band[i]][zero_nhood], offset);
767 if (token == DCT_EOB)
769 else if (token >= DCT_CAT1) {
770 int cat = token-DCT_CAT1;
771 token = vp8_rac_get_coeff(c, vp8_dct_cat_prob[cat]);
772 token += vp8_dct_cat_offset[cat];
775 // after the first token, the non-zero prediction context becomes
776 // based on the last decoded coeff
781 } else if (token == 1)
786 // todo: full [16] qmat? load into register?
787 block[zigzag_scan[i]] = (vp8_rac_get(c) ? -token : token) * qmul[!!i];
794 static void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,
795 uint8_t t_nnz[9], uint8_t l_nnz[9])
797 LOCAL_ALIGNED_16(DCTELEM, dc,[16]);
798 int i, x, y, luma_start = 0, luma_ctx = 3;
799 int nnz_pred, nnz, nnz_total = 0;
800 int segment = s->segmentation.enabled ? mb->segment : 0;
802 s->dsp.clear_blocks((DCTELEM *)s->block);
804 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
807 nnz_pred = t_nnz[8] + l_nnz[8];
809 // decode DC values and do hadamard
810 nnz = decode_block_coeffs(c, dc, s->prob->token[1], 0, nnz_pred,
811 s->qmat[segment].luma_dc_qmul);
812 l_nnz[8] = t_nnz[8] = !!nnz;
814 s->vp8dsp.vp8_luma_dc_wht(s->block, dc);
820 for (y = 0; y < 4; y++)
821 for (x = 0; x < 4; x++) {
822 nnz_pred = l_nnz[y] + t_nnz[x];
823 nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start,
824 nnz_pred, s->qmat[segment].luma_qmul);
825 // nnz+luma_start may be one more than the actual last index, but we don't care
826 s->non_zero_count_cache[y][x] = nnz + luma_start;
827 t_nnz[x] = l_nnz[y] = !!nnz;
832 // TODO: what to do about dimensions? 2nd dim for luma is x,
833 // but for chroma it's (y<<1)|x
834 for (i = 4; i < 6; i++)
835 for (y = 0; y < 2; y++)
836 for (x = 0; x < 2; x++) {
837 nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x];
838 nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0,
839 nnz_pred, s->qmat[segment].chroma_qmul);
840 s->non_zero_count_cache[i][(y<<1)+x] = nnz;
841 t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz;
845 // if there were no coded coeffs despite the macroblock not being marked skip,
846 // we MUST not do the inner loop filter and should not do IDCT
847 // Since skip isn't used for bitstream prediction, just manually set it.
852 static int check_intra_pred_mode(int mode, int mb_x, int mb_y)
854 if (mode == DC_PRED8x8) {
856 mode = DC_128_PRED8x8;
858 mode = LEFT_DC_PRED8x8;
860 mode = TOP_DC_PRED8x8;
865 static void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
866 uint8_t *bmode, int mb_x, int mb_y)
868 int x, y, mode, nnz, tr;
870 if (mb->mode < MODE_I4x4) {
871 mode = check_intra_pred_mode(mb->mode, mb_x, mb_y);
872 s->hpc.pred16x16[mode](dst[0], s->linesize);
874 uint8_t *ptr = dst[0];
876 // all blocks on the right edge of the macroblock use bottom edge
877 // the top macroblock for their topright edge
878 uint8_t *tr_right = ptr - s->linesize + 16;
880 // if we're on the right edge of the frame, said edge is extended
881 // from the top macroblock
882 if (mb_x == s->mb_width-1) {
883 tr = tr_right[-1]*0x01010101;
884 tr_right = (uint8_t *)&tr;
887 for (y = 0; y < 4; y++) {
888 uint8_t *topright = ptr + 4 - s->linesize;
889 for (x = 0; x < 4; x++) {
893 s->hpc.pred4x4[bmode[x]](ptr+4*x, topright, s->linesize);
895 nnz = s->non_zero_count_cache[y][x];
898 s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize);
900 s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize);
905 ptr += 4*s->linesize;
906 bmode += s->b4_stride;
910 mode = check_intra_pred_mode(s->chroma_pred_mode, mb_x, mb_y);
911 s->hpc.pred8x8[mode](dst[1], s->uvlinesize);
912 s->hpc.pred8x8[mode](dst[2], s->uvlinesize);
916 * Generic MC function.
918 * @param s VP8 decoding context
919 * @param luma 1 for luma (Y) planes, 0 for chroma (Cb/Cr) planes
920 * @param dst target buffer for block data at block position
921 * @param src reference picture buffer at origin (0, 0)
922 * @param mv motion vector (relative to block position) to get pixel data from
923 * @param x_off horizontal position of block from origin (0, 0)
924 * @param y_off vertical position of block from origin (0, 0)
925 * @param block_w width of block (16, 8 or 4)
926 * @param block_h height of block (always same as block_w)
927 * @param width width of src/dst plane data
928 * @param height height of src/dst plane data
929 * @param linesize size of a single line of plane data, including padding
930 * @param mc_func motion compensation function pointers (bilinear or sixtap MC)
932 static inline void vp8_mc(VP8Context *s, int luma,
933 uint8_t *dst, uint8_t *src, const VP56mv *mv,
934 int x_off, int y_off, int block_w, int block_h,
935 int width, int height, int linesize,
936 vp8_mc_func mc_func[3][3])
938 static const uint8_t idx[8] = { 0, 1, 2, 1, 2, 1, 2, 1 };
939 int mx = (mv->x << luma)&7, mx_idx = idx[mx];
940 int my = (mv->y << luma)&7, my_idx = idx[my];
942 x_off += mv->x >> (3 - luma);
943 y_off += mv->y >> (3 - luma);
946 src += y_off * linesize + x_off;
947 if (x_off < 2 || x_off >= width - block_w - 3 ||
948 y_off < 2 || y_off >= height - block_h - 3) {
949 ff_emulated_edge_mc(s->edge_emu_buffer, src - 2 * linesize - 2, linesize,
950 block_w + 5, block_h + 5,
951 x_off - 2, y_off - 2, width, height);
952 src = s->edge_emu_buffer + 2 + linesize * 2;
955 mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);
958 static inline void vp8_mc_part(VP8Context *s, uint8_t *dst[3],
959 AVFrame *ref_frame, int x_off, int y_off,
960 int bx_off, int by_off,
961 int block_w, int block_h,
962 int width, int height, VP56mv *mv)
967 vp8_mc(s, 1, dst[0] + by_off * s->linesize + bx_off,
968 ref_frame->data[0], mv, x_off + bx_off, y_off + by_off,
969 block_w, block_h, width, height, s->linesize,
970 s->put_pixels_tab[block_w == 8]);
973 if (s->profile == 3) {
977 x_off >>= 1; y_off >>= 1;
978 bx_off >>= 1; by_off >>= 1;
979 width >>= 1; height >>= 1;
980 block_w >>= 1; block_h >>= 1;
981 vp8_mc(s, 0, dst[1] + by_off * s->uvlinesize + bx_off,
982 ref_frame->data[1], &uvmv, x_off + bx_off, y_off + by_off,
983 block_w, block_h, width, height, s->uvlinesize,
984 s->put_pixels_tab[1 + (block_w == 4)]);
985 vp8_mc(s, 0, dst[2] + by_off * s->uvlinesize + bx_off,
986 ref_frame->data[2], &uvmv, x_off + bx_off, y_off + by_off,
987 block_w, block_h, width, height, s->uvlinesize,
988 s->put_pixels_tab[1 + (block_w == 4)]);
992 * Apply motion vectors to prediction buffer, chapter 18.
994 static void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
997 int x_off = mb_x << 4, y_off = mb_y << 4;
998 int width = 16*s->mb_width, height = 16*s->mb_height;
1000 if (mb->mode < VP8_MVMODE_SPLIT) {
1001 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1002 0, 0, 16, 16, width, height, &mb->mv);
1003 } else switch (mb->partitioning) {
1004 case VP8_SPLITMVMODE_4x4: {
1009 for (y = 0; y < 4; y++) {
1010 for (x = 0; x < 4; x++) {
1011 vp8_mc(s, 1, dst[0] + 4*y*s->linesize + x*4,
1012 s->framep[mb->ref_frame]->data[0], &mb->bmv[4*y + x],
1013 4*x + x_off, 4*y + y_off, 4, 4,
1014 width, height, s->linesize,
1015 s->put_pixels_tab[2]);
1020 x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1;
1021 for (y = 0; y < 2; y++) {
1022 for (x = 0; x < 2; x++) {
1023 uvmv.x = mb->bmv[ 2*y * 4 + 2*x ].x +
1024 mb->bmv[ 2*y * 4 + 2*x+1].x +
1025 mb->bmv[(2*y+1) * 4 + 2*x ].x +
1026 mb->bmv[(2*y+1) * 4 + 2*x+1].x;
1027 uvmv.y = mb->bmv[ 2*y * 4 + 2*x ].y +
1028 mb->bmv[ 2*y * 4 + 2*x+1].y +
1029 mb->bmv[(2*y+1) * 4 + 2*x ].y +
1030 mb->bmv[(2*y+1) * 4 + 2*x+1].y;
1031 uvmv.x = (uvmv.x + 2 + (uvmv.x >> (INT_BIT-1))) >> 2;
1032 uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2;
1033 if (s->profile == 3) {
1037 vp8_mc(s, 0, dst[1] + 4*y*s->uvlinesize + x*4,
1038 s->framep[mb->ref_frame]->data[1], &uvmv,
1039 4*x + x_off, 4*y + y_off, 4, 4,
1040 width, height, s->uvlinesize,
1041 s->put_pixels_tab[2]);
1042 vp8_mc(s, 0, dst[2] + 4*y*s->uvlinesize + x*4,
1043 s->framep[mb->ref_frame]->data[2], &uvmv,
1044 4*x + x_off, 4*y + y_off, 4, 4,
1045 width, height, s->uvlinesize,
1046 s->put_pixels_tab[2]);
1051 case VP8_SPLITMVMODE_16x8:
1052 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1053 0, 0, 16, 8, width, height, &mb->bmv[0]);
1054 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1055 0, 8, 16, 8, width, height, &mb->bmv[1]);
1057 case VP8_SPLITMVMODE_8x16:
1058 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1059 0, 0, 8, 16, width, height, &mb->bmv[0]);
1060 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1061 8, 0, 8, 16, width, height, &mb->bmv[1]);
1063 case VP8_SPLITMVMODE_8x8:
1064 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1065 0, 0, 8, 8, width, height, &mb->bmv[0]);
1066 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1067 8, 0, 8, 8, width, height, &mb->bmv[1]);
1068 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1069 0, 8, 8, 8, width, height, &mb->bmv[2]);
1070 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1071 8, 8, 8, 8, width, height, &mb->bmv[3]);
1076 static void idct_mb(VP8Context *s, uint8_t *y_dst, uint8_t *u_dst, uint8_t *v_dst,
1081 if (mb->mode != MODE_I4x4)
1082 for (y = 0; y < 4; y++) {
1083 for (x = 0; x < 4; x++) {
1084 nnz = s->non_zero_count_cache[y][x];
1087 s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize);
1089 s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize);
1092 y_dst += 4*s->linesize;
1095 for (y = 0; y < 2; y++) {
1096 for (x = 0; x < 2; x++) {
1097 nnz = s->non_zero_count_cache[4][(y<<1)+x];
1100 s->vp8dsp.vp8_idct_dc_add(u_dst+4*x, s->block[4][(y<<1)+x], s->uvlinesize);
1102 s->vp8dsp.vp8_idct_add(u_dst+4*x, s->block[4][(y<<1)+x], s->uvlinesize);
1105 nnz = s->non_zero_count_cache[5][(y<<1)+x];
1108 s->vp8dsp.vp8_idct_dc_add(v_dst+4*x, s->block[5][(y<<1)+x], s->uvlinesize);
1110 s->vp8dsp.vp8_idct_add(v_dst+4*x, s->block[5][(y<<1)+x], s->uvlinesize);
1113 u_dst += 4*s->uvlinesize;
1114 v_dst += 4*s->uvlinesize;
1118 static void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, int *level, int *inner, int *hev_thresh)
1120 int interior_limit, filter_level;
1122 if (s->segmentation.enabled) {
1123 filter_level = s->segmentation.filter_level[mb->segment];
1124 if (!s->segmentation.absolute_vals)
1125 filter_level += s->filter.level;
1127 filter_level = s->filter.level;
1129 if (s->lf_delta.enabled) {
1130 filter_level += s->lf_delta.ref[mb->ref_frame];
1132 if (mb->ref_frame == VP56_FRAME_CURRENT) {
1133 if (mb->mode == MODE_I4x4)
1134 filter_level += s->lf_delta.mode[0];
1136 if (mb->mode == VP8_MVMODE_ZERO)
1137 filter_level += s->lf_delta.mode[1];
1138 else if (mb->mode == VP8_MVMODE_SPLIT)
1139 filter_level += s->lf_delta.mode[3];
1141 filter_level += s->lf_delta.mode[2];
1144 filter_level = av_clip(filter_level, 0, 63);
1146 interior_limit = filter_level;
1147 if (s->filter.sharpness) {
1148 interior_limit >>= s->filter.sharpness > 4 ? 2 : 1;
1149 interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness);
1151 interior_limit = FFMAX(interior_limit, 1);
1153 *level = filter_level;
1154 *inner = interior_limit;
1157 *hev_thresh = filter_level >= 15;
1160 if (filter_level >= 40)
1163 if (filter_level >= 40)
1165 else if (filter_level >= 20)
1171 // TODO: look at backup_mb_border / xchg_mb_border in h264.c
1172 static void filter_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb, int mb_x, int mb_y)
1174 int filter_level, inner_limit, hev_thresh, mbedge_lim, bedge_lim;
1176 filter_level_for_mb(s, mb, &filter_level, &inner_limit, &hev_thresh);
1180 mbedge_lim = 2*(filter_level+2) + inner_limit;
1181 bedge_lim = 2* filter_level + inner_limit;
1184 s->vp8dsp.vp8_h_loop_filter16(dst[0], s->linesize, mbedge_lim, inner_limit, hev_thresh);
1185 s->vp8dsp.vp8_h_loop_filter8 (dst[1], s->uvlinesize, mbedge_lim, inner_limit, hev_thresh);
1186 s->vp8dsp.vp8_h_loop_filter8 (dst[2], s->uvlinesize, mbedge_lim, inner_limit, hev_thresh);
1189 if (!mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT) {
1190 s->vp8dsp.vp8_h_loop_filter16_inner(dst[0]+ 4, s->linesize, bedge_lim, inner_limit, hev_thresh);
1191 s->vp8dsp.vp8_h_loop_filter16_inner(dst[0]+ 8, s->linesize, bedge_lim, inner_limit, hev_thresh);
1192 s->vp8dsp.vp8_h_loop_filter16_inner(dst[0]+12, s->linesize, bedge_lim, inner_limit, hev_thresh);
1193 s->vp8dsp.vp8_h_loop_filter8_inner (dst[1]+ 4, s->uvlinesize, bedge_lim, inner_limit, hev_thresh);
1194 s->vp8dsp.vp8_h_loop_filter8_inner (dst[2]+ 4, s->uvlinesize, bedge_lim, inner_limit, hev_thresh);
1198 s->vp8dsp.vp8_v_loop_filter16(dst[0], s->linesize, mbedge_lim, inner_limit, hev_thresh);
1199 s->vp8dsp.vp8_v_loop_filter8 (dst[1], s->uvlinesize, mbedge_lim, inner_limit, hev_thresh);
1200 s->vp8dsp.vp8_v_loop_filter8 (dst[2], s->uvlinesize, mbedge_lim, inner_limit, hev_thresh);
1203 if (!mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT) {
1204 s->vp8dsp.vp8_v_loop_filter16_inner(dst[0]+ 4*s->linesize, s->linesize, bedge_lim, inner_limit, hev_thresh);
1205 s->vp8dsp.vp8_v_loop_filter16_inner(dst[0]+ 8*s->linesize, s->linesize, bedge_lim, inner_limit, hev_thresh);
1206 s->vp8dsp.vp8_v_loop_filter16_inner(dst[0]+12*s->linesize, s->linesize, bedge_lim, inner_limit, hev_thresh);
1207 s->vp8dsp.vp8_v_loop_filter8_inner (dst[1]+ 4*s->uvlinesize, s->uvlinesize, bedge_lim, inner_limit, hev_thresh);
1208 s->vp8dsp.vp8_v_loop_filter8_inner (dst[2]+ 4*s->uvlinesize, s->uvlinesize, bedge_lim, inner_limit, hev_thresh);
1212 static void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8Macroblock *mb, int mb_x, int mb_y)
1214 int filter_level, inner_limit, mbedge_lim, bedge_lim;
1216 filter_level_for_mb(s, mb, &filter_level, &inner_limit, NULL);
1220 mbedge_lim = 2*(filter_level+2) + inner_limit;
1221 bedge_lim = 2* filter_level + inner_limit;
1224 s->vp8dsp.vp8_h_loop_filter_simple(dst, s->linesize, mbedge_lim);
1225 if (!mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT) {
1226 s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, s->linesize, bedge_lim);
1227 s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, s->linesize, bedge_lim);
1228 s->vp8dsp.vp8_h_loop_filter_simple(dst+12, s->linesize, bedge_lim);
1232 s->vp8dsp.vp8_v_loop_filter_simple(dst, s->linesize, mbedge_lim);
1233 if (!mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT) {
1234 s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*s->linesize, s->linesize, bedge_lim);
1235 s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*s->linesize, s->linesize, bedge_lim);
1236 s->vp8dsp.vp8_v_loop_filter_simple(dst+12*s->linesize, s->linesize, bedge_lim);
1240 static void filter_mb_row(VP8Context *s, int mb_y)
1242 VP8Macroblock *mb = s->macroblocks + mb_y*s->mb_stride;
1244 s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize,
1245 s->framep[VP56_FRAME_CURRENT]->data[1] + 8*mb_y*s->uvlinesize,
1246 s->framep[VP56_FRAME_CURRENT]->data[2] + 8*mb_y*s->uvlinesize
1250 for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1251 filter_mb(s, dst, mb++, mb_x, mb_y);
1258 static void filter_mb_row_simple(VP8Context *s, int mb_y)
1260 uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize;
1261 VP8Macroblock *mb = s->macroblocks + mb_y*s->mb_stride;
1264 for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1265 filter_mb_simple(s, dst, mb++, mb_x, mb_y);
1270 static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
1273 VP8Context *s = avctx->priv_data;
1274 int ret, mb_x, mb_y, i, y, referenced;
1275 enum AVDiscard skip_thresh;
1278 if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0)
1281 referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT
1282 || s->update_altref == VP56_FRAME_CURRENT;
1284 skip_thresh = !referenced ? AVDISCARD_NONREF :
1285 !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL;
1287 if (avctx->skip_frame >= skip_thresh) {
1292 for (i = 0; i < 4; i++)
1293 if (&s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1294 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1295 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) {
1296 curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i];
1299 if (curframe->data[0])
1300 avctx->release_buffer(avctx, curframe);
1302 curframe->key_frame = s->keyframe;
1303 curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE;
1304 curframe->reference = referenced ? 3 : 0;
1305 if ((ret = avctx->get_buffer(avctx, curframe))) {
1306 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n");
1310 // Given that arithmetic probabilities are updated every frame, it's quite likely
1311 // that the values we have on a random interframe are complete junk if we didn't
1312 // start decode on a keyframe. So just don't display anything rather than junk.
1313 if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] ||
1314 !s->framep[VP56_FRAME_GOLDEN] ||
1315 !s->framep[VP56_FRAME_GOLDEN2])) {
1316 av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n");
1317 return AVERROR_INVALIDDATA;
1320 s->linesize = curframe->linesize[0];
1321 s->uvlinesize = curframe->linesize[1];
1323 if (!s->edge_emu_buffer)
1324 s->edge_emu_buffer = av_malloc(21*s->linesize);
1326 memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz));
1328 // top edge of 127 for intra prediction
1329 if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) {
1330 memset(curframe->data[0] - s->linesize -1, 127, s->linesize +1);
1331 memset(curframe->data[1] - s->uvlinesize-1, 127, s->uvlinesize+1);
1332 memset(curframe->data[2] - s->uvlinesize-1, 127, s->uvlinesize+1);
1335 for (mb_y = 0; mb_y < s->mb_height; mb_y++) {
1336 VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)];
1337 VP8Macroblock *mb = s->macroblocks + mb_y*s->mb_stride;
1338 uint8_t *intra4x4 = s->intra4x4_pred_mode + 4*mb_y*s->b4_stride;
1340 curframe->data[0] + 16*mb_y*s->linesize,
1341 curframe->data[1] + 8*mb_y*s->uvlinesize,
1342 curframe->data[2] + 8*mb_y*s->uvlinesize
1345 memset(s->left_nnz, 0, sizeof(s->left_nnz));
1347 // left edge of 129 for intra prediction
1348 if (!(avctx->flags & CODEC_FLAG_EMU_EDGE))
1349 for (i = 0; i < 3; i++)
1350 for (y = 0; y < 16>>!!i; y++)
1351 dst[i][y*curframe->linesize[i]-1] = 129;
1353 for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1354 decode_mb_mode(s, mb, mb_x, mb_y, intra4x4 + 4*mb_x);
1357 decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);
1359 AV_ZERO128(s->non_zero_count_cache); // luma
1360 AV_ZERO64(s->non_zero_count_cache[4]); // chroma
1363 if (mb->mode <= MODE_I4x4) {
1364 intra_predict(s, dst, mb, intra4x4 + 4*mb_x, mb_x, mb_y);
1365 memset(mb->bmv, 0, sizeof(mb->bmv));
1367 inter_predict(s, dst, mb, mb_x, mb_y);
1371 idct_mb(s, dst[0], dst[1], dst[2], mb);
1373 AV_ZERO64(s->left_nnz);
1374 AV_WN64(s->top_nnz[mb_x], 0); // array of 9, so unaligned
1376 // Reset DC block predictors if they would exist if the mb had coefficients
1377 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
1379 s->top_nnz[mb_x][8] = 0;
1388 if (mb_y && s->filter.level && avctx->skip_loop_filter < skip_thresh) {
1389 if (s->filter.simple)
1390 filter_mb_row_simple(s, mb_y-1);
1392 filter_mb_row(s, mb_y-1);
1395 if (s->filter.level && avctx->skip_loop_filter < skip_thresh) {
1396 if (s->filter.simple)
1397 filter_mb_row_simple(s, mb_y-1);
1399 filter_mb_row(s, mb_y-1);
1403 // if future frames don't use the updated probabilities,
1404 // reset them to the values we saved
1405 if (!s->update_probabilities)
1406 s->prob[0] = s->prob[1];
1408 // check if golden and altref are swapped
1409 if (s->update_altref == VP56_FRAME_GOLDEN &&
1410 s->update_golden == VP56_FRAME_GOLDEN2)
1411 FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN], s->framep[VP56_FRAME_GOLDEN2]);
1413 if (s->update_altref != VP56_FRAME_NONE)
1414 s->framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref];
1416 if (s->update_golden != VP56_FRAME_NONE)
1417 s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden];
1420 if (s->update_last) // move cur->prev
1421 s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT];
1423 // release no longer referenced frames
1424 for (i = 0; i < 4; i++)
1425 if (s->frames[i].data[0] &&
1426 &s->frames[i] != s->framep[VP56_FRAME_CURRENT] &&
1427 &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1428 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1429 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2])
1430 avctx->release_buffer(avctx, &s->frames[i]);
1432 if (!s->invisible) {
1433 *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT];
1434 *data_size = sizeof(AVFrame);
1440 static av_cold int vp8_decode_init(AVCodecContext *avctx)
1442 VP8Context *s = avctx->priv_data;
1445 avctx->pix_fmt = PIX_FMT_YUV420P;
1447 dsputil_init(&s->dsp, avctx);
1448 ff_h264_pred_init(&s->hpc, CODEC_ID_VP8);
1449 ff_vp8dsp_init(&s->vp8dsp);
1451 // intra pred needs edge emulation among other things
1452 if (avctx->flags&CODEC_FLAG_EMU_EDGE) {
1453 av_log(avctx, AV_LOG_ERROR, "Edge emulation not supported\n");
1454 return AVERROR_PATCHWELCOME;
1460 static av_cold int vp8_decode_free(AVCodecContext *avctx)
1462 vp8_decode_flush(avctx);
1466 AVCodec vp8_decoder = {
1476 .flush = vp8_decode_flush,
1477 .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"),