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"
39 // todo: make it possible to check for at least (i4x4 or split_mv)
40 // in one op. are others needed?
49 AVCodecContext *avctx;
53 vp8_mc_func put_pixels_tab[3][3][3];
56 uint8_t *edge_emu_buffer;
57 VP56RangeCoder c; ///< header context, includes mb modes and motion vectors
60 int mb_width; /* number of horizontal MB */
61 int mb_height; /* number of vertical MB */
67 int update_last; ///< update VP56_FRAME_PREVIOUS with the current one
68 int update_golden; ///< VP56_FRAME_NONE if not updated, or which frame to copy if so
73 * If this flag is not set, all the probability updates
74 * are discarded after this frame is decoded.
76 int update_probabilities;
79 * All coefficients are contained in separate arith coding contexts.
80 * There can be 1, 2, 4, or 8 of these after the header context.
82 int num_coeff_partitions;
83 VP56RangeCoder coeff_partition[8];
85 VP8Macroblock *macroblocks;
86 VP8Macroblock *macroblocks_base;
87 VP8FilterStrength *filter_strength;
90 uint8_t *intra4x4_pred_mode;
91 uint8_t *intra4x4_pred_mode_base;
92 uint8_t *segmentation_map;
96 * Cache of the top row needed for intra prediction
97 * 16 for luma, 8 for each chroma plane
99 uint8_t (*top_border)[16+8+8];
102 * For coeff decode, we need to know whether the above block had non-zero
103 * coefficients. This means for each macroblock, we need data for 4 luma
104 * blocks, 2 u blocks, 2 v blocks, and the luma dc block, for a total of 9
105 * per macroblock. We keep the last row in top_nnz.
107 uint8_t (*top_nnz)[9];
108 DECLARE_ALIGNED(8, uint8_t, left_nnz)[9];
111 * This is the index plus one of the last non-zero coeff
112 * for each of the blocks in the current macroblock.
114 * 1 -> dc-only (special transform)
115 * 2+-> full transform
117 DECLARE_ALIGNED(16, uint8_t, non_zero_count_cache)[6][4];
118 DECLARE_ALIGNED(16, DCTELEM, block)[6][4][16];
119 uint8_t intra4x4_pred_mode_mb[16];
121 int chroma_pred_mode; ///< 8x8c pred mode of the current macroblock
122 int segment; ///< segment of the current macroblock
125 int sign_bias[4]; ///< one state [0, 1] per ref frame type
129 * Base parameters for segmentation, i.e. per-macroblock parameters.
130 * These must be kept unchanged even if segmentation is not used for
131 * a frame, since the values persist between interframes.
137 int8_t base_quant[4];
138 int8_t filter_level[4]; ///< base loop filter level
142 * Macroblocks can have one of 4 different quants in a frame when
143 * segmentation is enabled.
144 * If segmentation is disabled, only the first segment's values are used.
147 // [0] - DC qmul [1] - AC qmul
148 int16_t luma_qmul[2];
149 int16_t luma_dc_qmul[2]; ///< luma dc-only block quant
150 int16_t chroma_qmul[2];
160 int enabled; ///< whether each mb can have a different strength based on mode/ref
163 * filter strength adjustment for the following macroblock modes:
166 * [2] - inter modes except for zero or split mv
168 * i16x16 modes never have any adjustment
173 * filter strength adjustment for macroblocks that reference:
174 * [0] - intra / VP56_FRAME_CURRENT
175 * [1] - VP56_FRAME_PREVIOUS
176 * [2] - VP56_FRAME_GOLDEN
177 * [3] - altref / VP56_FRAME_GOLDEN2
183 * These are all of the updatable probabilities for binary decisions.
184 * They are only implictly reset on keyframes, making it quite likely
185 * for an interframe to desync if a prior frame's header was corrupt
186 * or missing outright!
189 uint8_t segmentid[3];
194 uint8_t pred16x16[4];
196 uint8_t token[4][8][3][NUM_DCT_TOKENS-1];
201 static void vp8_decode_flush(AVCodecContext *avctx)
203 VP8Context *s = avctx->priv_data;
206 for (i = 0; i < 4; i++)
207 if (s->frames[i].data[0])
208 avctx->release_buffer(avctx, &s->frames[i]);
209 memset(s->framep, 0, sizeof(s->framep));
211 av_freep(&s->macroblocks_base);
212 av_freep(&s->intra4x4_pred_mode_base);
213 av_freep(&s->top_nnz);
214 av_freep(&s->edge_emu_buffer);
215 av_freep(&s->top_border);
216 av_freep(&s->segmentation_map);
218 s->macroblocks = NULL;
219 s->intra4x4_pred_mode = NULL;
222 static int update_dimensions(VP8Context *s, int width, int height)
226 if (avcodec_check_dimensions(s->avctx, width, height))
227 return AVERROR_INVALIDDATA;
229 vp8_decode_flush(s->avctx);
231 avcodec_set_dimensions(s->avctx, width, height);
233 s->mb_width = (s->avctx->coded_width +15) / 16;
234 s->mb_height = (s->avctx->coded_height+15) / 16;
236 // we allocate a border around the top/left of intra4x4 modes
237 // this is 4 blocks for intra4x4 to keep 4-byte alignment for fill_rectangle
238 s->mb_stride = s->mb_width+1;
239 s->b4_stride = 4*s->mb_stride;
241 s->macroblocks_base = av_mallocz((s->mb_stride+s->mb_height*2+2)*sizeof(*s->macroblocks));
242 s->filter_strength = av_mallocz(s->mb_stride*sizeof(*s->filter_strength));
243 s->intra4x4_pred_mode_base = av_mallocz(s->b4_stride*(4*s->mb_height+1));
244 s->top_nnz = av_mallocz(s->mb_width*sizeof(*s->top_nnz));
245 s->top_border = av_mallocz((s->mb_width+1)*sizeof(*s->top_border));
246 s->segmentation_map = av_mallocz(s->mb_stride*s->mb_height);
248 if (!s->macroblocks_base || !s->filter_strength || !s->intra4x4_pred_mode_base ||
249 !s->top_nnz || !s->top_border || !s->segmentation_map)
250 return AVERROR(ENOMEM);
252 s->macroblocks = s->macroblocks_base + 1;
253 s->intra4x4_pred_mode = s->intra4x4_pred_mode_base + 4 + s->b4_stride;
255 memset(s->intra4x4_pred_mode_base, DC_PRED, s->b4_stride);
256 for (i = 0; i < 4*s->mb_height; i++)
257 s->intra4x4_pred_mode[i*s->b4_stride-1] = DC_PRED;
262 static void parse_segment_info(VP8Context *s)
264 VP56RangeCoder *c = &s->c;
267 s->segmentation.update_map = vp8_rac_get(c);
269 if (vp8_rac_get(c)) { // update segment feature data
270 s->segmentation.absolute_vals = vp8_rac_get(c);
272 for (i = 0; i < 4; i++)
273 s->segmentation.base_quant[i] = vp8_rac_get_sint(c, 7);
275 for (i = 0; i < 4; i++)
276 s->segmentation.filter_level[i] = vp8_rac_get_sint(c, 6);
278 if (s->segmentation.update_map)
279 for (i = 0; i < 3; i++)
280 s->prob->segmentid[i] = vp8_rac_get(c) ? vp8_rac_get_uint(c, 8) : 255;
283 static void update_lf_deltas(VP8Context *s)
285 VP56RangeCoder *c = &s->c;
288 for (i = 0; i < 4; i++)
289 s->lf_delta.ref[i] = vp8_rac_get_sint(c, 6);
291 for (i = 0; i < 4; i++)
292 s->lf_delta.mode[i] = vp8_rac_get_sint(c, 6);
295 static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size)
297 const uint8_t *sizes = buf;
300 s->num_coeff_partitions = 1 << vp8_rac_get_uint(&s->c, 2);
302 buf += 3*(s->num_coeff_partitions-1);
303 buf_size -= 3*(s->num_coeff_partitions-1);
307 for (i = 0; i < s->num_coeff_partitions-1; i++) {
308 int size = AV_RL24(sizes + 3*i);
309 if (buf_size - size < 0)
312 vp56_init_range_decoder(&s->coeff_partition[i], buf, size);
316 vp56_init_range_decoder(&s->coeff_partition[i], buf, buf_size);
321 static void get_quants(VP8Context *s)
323 VP56RangeCoder *c = &s->c;
326 int yac_qi = vp8_rac_get_uint(c, 7);
327 int ydc_delta = vp8_rac_get_sint(c, 4);
328 int y2dc_delta = vp8_rac_get_sint(c, 4);
329 int y2ac_delta = vp8_rac_get_sint(c, 4);
330 int uvdc_delta = vp8_rac_get_sint(c, 4);
331 int uvac_delta = vp8_rac_get_sint(c, 4);
333 for (i = 0; i < 4; i++) {
334 if (s->segmentation.enabled) {
335 base_qi = s->segmentation.base_quant[i];
336 if (!s->segmentation.absolute_vals)
341 s->qmat[i].luma_qmul[0] = vp8_dc_qlookup[av_clip(base_qi + ydc_delta , 0, 127)];
342 s->qmat[i].luma_qmul[1] = vp8_ac_qlookup[av_clip(base_qi , 0, 127)];
343 s->qmat[i].luma_dc_qmul[0] = 2 * vp8_dc_qlookup[av_clip(base_qi + y2dc_delta, 0, 127)];
344 s->qmat[i].luma_dc_qmul[1] = 155 * vp8_ac_qlookup[av_clip(base_qi + y2ac_delta, 0, 127)] / 100;
345 s->qmat[i].chroma_qmul[0] = vp8_dc_qlookup[av_clip(base_qi + uvdc_delta, 0, 127)];
346 s->qmat[i].chroma_qmul[1] = vp8_ac_qlookup[av_clip(base_qi + uvac_delta, 0, 127)];
348 s->qmat[i].luma_dc_qmul[1] = FFMAX(s->qmat[i].luma_dc_qmul[1], 8);
349 s->qmat[i].chroma_qmul[0] = FFMIN(s->qmat[i].chroma_qmul[0], 132);
354 * Determine which buffers golden and altref should be updated with after this frame.
355 * The spec isn't clear here, so I'm going by my understanding of what libvpx does
357 * Intra frames update all 3 references
358 * Inter frames update VP56_FRAME_PREVIOUS if the update_last flag is set
359 * If the update (golden|altref) flag is set, it's updated with the current frame
360 * if update_last is set, and VP56_FRAME_PREVIOUS otherwise.
361 * If the flag is not set, the number read means:
363 * 1: VP56_FRAME_PREVIOUS
364 * 2: update golden with altref, or update altref with golden
366 static VP56Frame ref_to_update(VP8Context *s, int update, VP56Frame ref)
368 VP56RangeCoder *c = &s->c;
371 return VP56_FRAME_CURRENT;
373 switch (vp8_rac_get_uint(c, 2)) {
375 return VP56_FRAME_PREVIOUS;
377 return (ref == VP56_FRAME_GOLDEN) ? VP56_FRAME_GOLDEN2 : VP56_FRAME_GOLDEN;
379 return VP56_FRAME_NONE;
382 static void update_refs(VP8Context *s)
384 VP56RangeCoder *c = &s->c;
386 int update_golden = vp8_rac_get(c);
387 int update_altref = vp8_rac_get(c);
389 s->update_golden = ref_to_update(s, update_golden, VP56_FRAME_GOLDEN);
390 s->update_altref = ref_to_update(s, update_altref, VP56_FRAME_GOLDEN2);
393 static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)
395 VP56RangeCoder *c = &s->c;
396 int header_size, hscale, vscale, i, j, k, l, ret;
397 int width = s->avctx->width;
398 int height = s->avctx->height;
400 s->keyframe = !(buf[0] & 1);
401 s->profile = (buf[0]>>1) & 7;
402 s->invisible = !(buf[0] & 0x10);
403 header_size = AV_RL24(buf) >> 5;
408 av_log(s->avctx, AV_LOG_WARNING, "Unknown profile %d\n", s->profile);
411 memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_epel_pixels_tab, sizeof(s->put_pixels_tab));
412 else // profile 1-3 use bilinear, 4+ aren't defined so whatever
413 memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_bilinear_pixels_tab, sizeof(s->put_pixels_tab));
415 if (header_size > buf_size - 7*s->keyframe) {
416 av_log(s->avctx, AV_LOG_ERROR, "Header size larger than data provided\n");
417 return AVERROR_INVALIDDATA;
421 if (AV_RL24(buf) != 0x2a019d) {
422 av_log(s->avctx, AV_LOG_ERROR, "Invalid start code 0x%x\n", AV_RL24(buf));
423 return AVERROR_INVALIDDATA;
425 width = AV_RL16(buf+3) & 0x3fff;
426 height = AV_RL16(buf+5) & 0x3fff;
427 hscale = buf[4] >> 6;
428 vscale = buf[6] >> 6;
432 if (hscale || vscale)
433 av_log_missing_feature(s->avctx, "Upscaling", 1);
435 s->update_golden = s->update_altref = VP56_FRAME_CURRENT;
436 memcpy(s->prob->token , vp8_token_default_probs , sizeof(s->prob->token));
437 memcpy(s->prob->pred16x16, vp8_pred16x16_prob_inter, sizeof(s->prob->pred16x16));
438 memcpy(s->prob->pred8x8c , vp8_pred8x8c_prob_inter , sizeof(s->prob->pred8x8c));
439 memcpy(s->prob->mvc , vp8_mv_default_prob , sizeof(s->prob->mvc));
440 memset(&s->segmentation, 0, sizeof(s->segmentation));
443 if (!s->macroblocks_base || /* first frame */
444 width != s->avctx->width || height != s->avctx->height) {
445 if ((ret = update_dimensions(s, width, height) < 0))
449 vp56_init_range_decoder(c, buf, header_size);
451 buf_size -= header_size;
455 av_log(s->avctx, AV_LOG_WARNING, "Unspecified colorspace\n");
456 vp8_rac_get(c); // whether we can skip clamping in dsp functions
459 if ((s->segmentation.enabled = vp8_rac_get(c)))
460 parse_segment_info(s);
462 s->segmentation.update_map = 0; // FIXME: move this to some init function?
464 s->filter.simple = vp8_rac_get(c);
465 s->filter.level = vp8_rac_get_uint(c, 6);
466 s->filter.sharpness = vp8_rac_get_uint(c, 3);
468 if ((s->lf_delta.enabled = vp8_rac_get(c)))
472 if (setup_partitions(s, buf, buf_size)) {
473 av_log(s->avctx, AV_LOG_ERROR, "Invalid partitions\n");
474 return AVERROR_INVALIDDATA;
481 s->sign_bias[VP56_FRAME_GOLDEN] = vp8_rac_get(c);
482 s->sign_bias[VP56_FRAME_GOLDEN2 /* altref */] = vp8_rac_get(c);
485 // if we aren't saving this frame's probabilities for future frames,
486 // make a copy of the current probabilities
487 if (!(s->update_probabilities = vp8_rac_get(c)))
488 s->prob[1] = s->prob[0];
490 s->update_last = s->keyframe || vp8_rac_get(c);
492 for (i = 0; i < 4; i++)
493 for (j = 0; j < 8; j++)
494 for (k = 0; k < 3; k++)
495 for (l = 0; l < NUM_DCT_TOKENS-1; l++)
496 if (vp56_rac_get_prob(c, vp8_token_update_probs[i][j][k][l]))
497 s->prob->token[i][j][k][l] = vp8_rac_get_uint(c, 8);
499 if ((s->mbskip_enabled = vp8_rac_get(c)))
500 s->prob->mbskip = vp8_rac_get_uint(c, 8);
503 s->prob->intra = vp8_rac_get_uint(c, 8);
504 s->prob->last = vp8_rac_get_uint(c, 8);
505 s->prob->golden = vp8_rac_get_uint(c, 8);
508 for (i = 0; i < 4; i++)
509 s->prob->pred16x16[i] = vp8_rac_get_uint(c, 8);
511 for (i = 0; i < 3; i++)
512 s->prob->pred8x8c[i] = vp8_rac_get_uint(c, 8);
514 // 17.2 MV probability update
515 for (i = 0; i < 2; i++)
516 for (j = 0; j < 19; j++)
517 if (vp56_rac_get_prob(c, vp8_mv_update_prob[i][j]))
518 s->prob->mvc[i][j] = vp8_rac_get_nn(c);
524 static inline void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src,
527 #define MARGIN (16 << 2)
528 dst->x = av_clip(src->x, -((mb_x << 6) + MARGIN),
529 ((s->mb_width - 1 - mb_x) << 6) + MARGIN);
530 dst->y = av_clip(src->y, -((mb_y << 6) + MARGIN),
531 ((s->mb_height - 1 - mb_y) << 6) + MARGIN);
534 static void find_near_mvs(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
535 VP56mv near[2], VP56mv *best, uint8_t cnt[4])
537 VP8Macroblock *mb_edge[3] = { mb + 2 /* top */,
539 mb + 1 /* top-left */ };
540 enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT };
541 VP56mv near_mv[4] = {{ 0 }};
542 enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV };
544 int best_idx = CNT_ZERO;
545 int cur_sign_bias = s->sign_bias[mb->ref_frame];
546 int *sign_bias = s->sign_bias;
548 /* Process MB on top, left and top-left */
549 #define MV_EDGE_CHECK(n)\
551 VP8Macroblock *edge = mb_edge[n];\
552 int edge_ref = edge->ref_frame;\
553 if (edge_ref != VP56_FRAME_CURRENT) {\
554 uint32_t mv = AV_RN32A(&edge->mv);\
556 if (cur_sign_bias != sign_bias[edge_ref]) {\
557 /* SWAR negate of the values in mv. */\
559 mv = ((mv&0x7fff7fff) + 0x00010001) ^ (mv&0x80008000);\
561 if (!n || mv != AV_RN32A(&near_mv[idx]))\
562 AV_WN32A(&near_mv[++idx], mv);\
563 cnt[idx] += 1 + (n != 2);\
565 cnt[CNT_ZERO] += 1 + (n != 2);\
572 /* If we have three distinct MVs, merge first and last if they're the same */
573 if (cnt[CNT_SPLITMV] && AV_RN32A(&near_mv[1+EDGE_TOP]) == AV_RN32A(&near_mv[1+EDGE_TOPLEFT]))
574 cnt[CNT_NEAREST] += 1;
576 cnt[CNT_SPLITMV] = ((mb_edge[EDGE_LEFT]->mode == VP8_MVMODE_SPLIT) +
577 (mb_edge[EDGE_TOP]->mode == VP8_MVMODE_SPLIT)) * 2 +
578 (mb_edge[EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT);
580 /* Swap near and nearest if necessary */
581 if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) {
582 FFSWAP(uint8_t, cnt[CNT_NEAREST], cnt[CNT_NEAR]);
583 FFSWAP( VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]);
586 /* Choose the best mv out of 0,0 and the nearest mv */
587 if (cnt[CNT_NEAREST] >= cnt[CNT_ZERO])
588 best_idx = CNT_NEAREST;
590 mb->mv = near_mv[best_idx];
591 near[0] = near_mv[CNT_NEAREST];
592 near[1] = near_mv[CNT_NEAR];
596 * Motion vector coding, 17.1.
598 static int read_mv_component(VP56RangeCoder *c, const uint8_t *p)
602 if (vp56_rac_get_prob(c, p[0])) {
605 for (i = 0; i < 3; i++)
606 x += vp56_rac_get_prob(c, p[9 + i]) << i;
607 for (i = 9; i > 3; i--)
608 x += vp56_rac_get_prob(c, p[9 + i]) << i;
609 if (!(x & 0xFFF0) || vp56_rac_get_prob(c, p[12]))
612 x = vp8_rac_get_tree(c, vp8_small_mvtree, &p[2]);
614 return (x && vp56_rac_get_prob(c, p[1])) ? -x : x;
617 static const uint8_t *get_submv_prob(uint32_t left, uint32_t top)
620 return vp8_submv_prob[4-!!left];
622 return vp8_submv_prob[2];
623 return vp8_submv_prob[1-!!left];
627 * Split motion vector prediction, 16.4.
628 * @returns the number of motion vectors parsed (2, 4 or 16)
630 static int decode_splitmvs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb)
632 int part_idx = mb->partitioning =
633 vp8_rac_get_tree(c, vp8_mbsplit_tree, vp8_mbsplit_prob);
634 int n, num = vp8_mbsplit_count[part_idx];
635 VP8Macroblock *top_mb = &mb[2];
636 VP8Macroblock *left_mb = &mb[-1];
637 const uint8_t *mbsplits_left = vp8_mbsplits[left_mb->partitioning],
638 *mbsplits_top = vp8_mbsplits[top_mb->partitioning],
639 *mbsplits_cur = vp8_mbsplits[part_idx],
640 *firstidx = vp8_mbfirstidx[part_idx];
641 VP56mv *top_mv = top_mb->bmv;
642 VP56mv *left_mv = left_mb->bmv;
643 VP56mv *cur_mv = mb->bmv;
645 for (n = 0; n < num; n++) {
647 uint32_t left, above;
648 const uint8_t *submv_prob;
651 left = AV_RN32A(&left_mv[mbsplits_left[k + 3]]);
653 left = AV_RN32A(&cur_mv[mbsplits_cur[k - 1]]);
655 above = AV_RN32A(&top_mv[mbsplits_top[k + 12]]);
657 above = AV_RN32A(&cur_mv[mbsplits_cur[k - 4]]);
659 submv_prob = get_submv_prob(left, above);
661 switch (vp8_rac_get_tree(c, vp8_submv_ref_tree, submv_prob)) {
662 case VP8_SUBMVMODE_NEW4X4:
663 mb->bmv[n].y = mb->mv.y + read_mv_component(c, s->prob->mvc[0]);
664 mb->bmv[n].x = mb->mv.x + read_mv_component(c, s->prob->mvc[1]);
666 case VP8_SUBMVMODE_ZERO4X4:
667 AV_ZERO32(&mb->bmv[n]);
669 case VP8_SUBMVMODE_LEFT4X4:
670 AV_WN32A(&mb->bmv[n], left);
672 case VP8_SUBMVMODE_TOP4X4:
673 AV_WN32A(&mb->bmv[n], above);
681 static inline void decode_intra4x4_modes(VP56RangeCoder *c, uint8_t *intra4x4,
682 int stride, int keyframe)
688 for (y = 0; y < 4; y++) {
689 for (x = 0; x < 4; x++) {
690 t = intra4x4[x - stride];
692 ctx = vp8_pred4x4_prob_intra[t][l];
693 intra4x4[x] = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx);
698 for (i = 0; i < 16; i++)
699 intra4x4[i] = vp8_rac_get_tree(c, vp8_pred4x4_tree, vp8_pred4x4_prob_inter);
703 static void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
704 uint8_t *intra4x4, uint8_t *segment)
706 VP56RangeCoder *c = &s->c;
708 if (s->segmentation.update_map)
709 *segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid);
710 s->segment = *segment;
712 mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0;
715 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra);
717 if (mb->mode == MODE_I4x4) {
718 decode_intra4x4_modes(c, intra4x4, s->b4_stride, 1);
720 fill_rectangle(intra4x4, 4, 4, s->b4_stride, vp8_pred4x4_mode[mb->mode], 1);
722 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra);
723 mb->ref_frame = VP56_FRAME_CURRENT;
724 } else if (vp56_rac_get_prob(c, s->prob->intra)) {
725 VP56mv near[2], best;
726 uint8_t cnt[4] = { 0 };
730 if (vp56_rac_get_prob(c, s->prob->last))
731 mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ?
732 VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN;
734 mb->ref_frame = VP56_FRAME_PREVIOUS;
735 s->ref_count[mb->ref_frame-1]++;
737 // motion vectors, 16.3
738 find_near_mvs(s, mb, mb_x, mb_y, near, &best, cnt);
739 p[0] = vp8_mode_contexts[cnt[0]][0];
740 p[1] = vp8_mode_contexts[cnt[1]][1];
741 p[2] = vp8_mode_contexts[cnt[2]][2];
742 p[3] = vp8_mode_contexts[cnt[3]][3];
743 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_mvinter, p);
745 case VP8_MVMODE_SPLIT:
746 clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y);
747 mb->mv = mb->bmv[decode_splitmvs(s, c, mb) - 1];
749 case VP8_MVMODE_ZERO:
752 case VP8_MVMODE_NEAREST:
753 clamp_mv(s, &mb->mv, &near[0], mb_x, mb_y);
755 case VP8_MVMODE_NEAR:
756 clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y);
759 clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y);
760 mb->mv.y += + read_mv_component(c, s->prob->mvc[0]);
761 mb->mv.x += + read_mv_component(c, s->prob->mvc[1]);
764 if (mb->mode != VP8_MVMODE_SPLIT) {
765 mb->partitioning = VP8_SPLITMVMODE_NONE;
770 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16);
772 if (mb->mode == MODE_I4x4)
773 decode_intra4x4_modes(c, intra4x4, 4, 0);
775 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c);
776 mb->ref_frame = VP56_FRAME_CURRENT;
777 mb->partitioning = VP8_SPLITMVMODE_NONE;
778 AV_ZERO32(&mb->bmv[0]);
783 * @param c arithmetic bitstream reader context
784 * @param block destination for block coefficients
785 * @param probs probabilities to use when reading trees from the bitstream
786 * @param i initial coeff index, 0 unless a separate DC block is coded
787 * @param zero_nhood the initial prediction context for number of surrounding
788 * all-zero blocks (only left/top, so 0-2)
789 * @param qmul array holding the dc/ac dequant factor at position 0/1
790 * @return 0 if no coeffs were decoded
791 * otherwise, the index of the last coeff decoded plus one
793 static int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16],
794 uint8_t probs[8][3][NUM_DCT_TOKENS-1],
795 int i, int zero_nhood, int16_t qmul[2])
797 int token, nonzero = 0;
800 for (; i < 16; i++) {
801 token = vp8_rac_get_tree_with_offset(c, vp8_coeff_tree, probs[vp8_coeff_band[i]][zero_nhood], offset);
803 if (token == DCT_EOB)
805 else if (token >= DCT_CAT1) {
806 int cat = token-DCT_CAT1;
807 token = vp8_rac_get_coeff(c, vp8_dct_cat_prob[cat]);
808 token += 3 + (2<<cat);
811 // after the first token, the non-zero prediction context becomes
812 // based on the last decoded coeff
817 } else if (token == 1)
822 // todo: full [16] qmat? load into register?
823 block[zigzag_scan[i]] = (vp8_rac_get(c) ? -token : token) * qmul[!!i];
830 static void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,
831 uint8_t t_nnz[9], uint8_t l_nnz[9])
833 LOCAL_ALIGNED_16(DCTELEM, dc,[16]);
834 int i, x, y, luma_start = 0, luma_ctx = 3;
835 int nnz_pred, nnz, nnz_total = 0;
836 int segment = s->segment;
838 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
841 nnz_pred = t_nnz[8] + l_nnz[8];
843 // decode DC values and do hadamard
844 nnz = decode_block_coeffs(c, dc, s->prob->token[1], 0, nnz_pred,
845 s->qmat[segment].luma_dc_qmul);
846 l_nnz[8] = t_nnz[8] = !!nnz;
848 s->vp8dsp.vp8_luma_dc_wht(s->block, dc);
854 for (y = 0; y < 4; y++)
855 for (x = 0; x < 4; x++) {
856 nnz_pred = l_nnz[y] + t_nnz[x];
857 nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start,
858 nnz_pred, s->qmat[segment].luma_qmul);
859 // nnz+luma_start may be one more than the actual last index, but we don't care
860 s->non_zero_count_cache[y][x] = nnz + luma_start;
861 t_nnz[x] = l_nnz[y] = !!nnz;
866 // TODO: what to do about dimensions? 2nd dim for luma is x,
867 // but for chroma it's (y<<1)|x
868 for (i = 4; i < 6; i++)
869 for (y = 0; y < 2; y++)
870 for (x = 0; x < 2; x++) {
871 nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x];
872 nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0,
873 nnz_pred, s->qmat[segment].chroma_qmul);
874 s->non_zero_count_cache[i][(y<<1)+x] = nnz;
875 t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz;
879 // if there were no coded coeffs despite the macroblock not being marked skip,
880 // we MUST not do the inner loop filter and should not do IDCT
881 // Since skip isn't used for bitstream prediction, just manually set it.
886 static av_always_inline
887 void backup_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
888 int linesize, int uvlinesize, int simple)
890 AV_COPY128(top_border, src_y + 15*linesize);
892 AV_COPY64(top_border+16, src_cb + 7*uvlinesize);
893 AV_COPY64(top_border+24, src_cr + 7*uvlinesize);
897 static av_always_inline
898 void xchg_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
899 int linesize, int uvlinesize, int mb_x, int mb_y, int mb_width,
900 int simple, int xchg)
902 uint8_t *top_border_m1 = top_border-32; // for TL prediction
904 src_cb -= uvlinesize;
905 src_cr -= uvlinesize;
907 #define XCHG(a,b,xchg) do { \
908 if (xchg) AV_SWAP64(b,a); \
909 else AV_COPY64(b,a); \
912 XCHG(top_border_m1+8, src_y-8, xchg);
913 XCHG(top_border, src_y, xchg);
914 XCHG(top_border+8, src_y+8, 1);
915 if (mb_x < mb_width-1)
916 XCHG(top_border+32, src_y+16, 1);
918 // only copy chroma for normal loop filter
919 // or to initialize the top row to 127
920 if (!simple || !mb_y) {
921 XCHG(top_border_m1+16, src_cb-8, xchg);
922 XCHG(top_border_m1+24, src_cr-8, xchg);
923 XCHG(top_border+16, src_cb, 1);
924 XCHG(top_border+24, src_cr, 1);
928 static int check_intra_pred_mode(int mode, int mb_x, int mb_y)
930 if (mode == DC_PRED8x8) {
932 mode = mb_y ? TOP_DC_PRED8x8 : DC_128_PRED8x8;
934 mode = LEFT_DC_PRED8x8;
940 static void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
941 uint8_t *intra4x4, int mb_x, int mb_y)
943 int x, y, mode, nnz, tr;
945 // for the first row, we need to run xchg_mb_border to init the top edge to 127
946 // otherwise, skip it if we aren't going to deblock
947 if (s->deblock_filter || !mb_y)
948 xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
949 s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
950 s->filter.simple, 1);
952 if (mb->mode < MODE_I4x4) {
953 mode = check_intra_pred_mode(mb->mode, mb_x, mb_y);
954 s->hpc.pred16x16[mode](dst[0], s->linesize);
956 uint8_t *ptr = dst[0];
957 int stride = s->keyframe ? s->b4_stride : 4;
959 // all blocks on the right edge of the macroblock use bottom edge
960 // the top macroblock for their topright edge
961 uint8_t *tr_right = ptr - s->linesize + 16;
963 // if we're on the right edge of the frame, said edge is extended
964 // from the top macroblock
965 if (mb_x == s->mb_width-1) {
966 tr = tr_right[-1]*0x01010101;
967 tr_right = (uint8_t *)&tr;
971 AV_ZERO128(s->non_zero_count_cache);
973 for (y = 0; y < 4; y++) {
974 uint8_t *topright = ptr + 4 - s->linesize;
975 for (x = 0; x < 4; x++) {
979 s->hpc.pred4x4[intra4x4[x]](ptr+4*x, topright, s->linesize);
981 nnz = s->non_zero_count_cache[y][x];
984 s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize);
986 s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize);
991 ptr += 4*s->linesize;
996 mode = check_intra_pred_mode(s->chroma_pred_mode, mb_x, mb_y);
997 s->hpc.pred8x8[mode](dst[1], s->uvlinesize);
998 s->hpc.pred8x8[mode](dst[2], s->uvlinesize);
1000 if (s->deblock_filter || !mb_y)
1001 xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
1002 s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
1003 s->filter.simple, 0);
1007 * Generic MC function.
1009 * @param s VP8 decoding context
1010 * @param luma 1 for luma (Y) planes, 0 for chroma (Cb/Cr) planes
1011 * @param dst target buffer for block data at block position
1012 * @param src reference picture buffer at origin (0, 0)
1013 * @param mv motion vector (relative to block position) to get pixel data from
1014 * @param x_off horizontal position of block from origin (0, 0)
1015 * @param y_off vertical position of block from origin (0, 0)
1016 * @param block_w width of block (16, 8 or 4)
1017 * @param block_h height of block (always same as block_w)
1018 * @param width width of src/dst plane data
1019 * @param height height of src/dst plane data
1020 * @param linesize size of a single line of plane data, including padding
1021 * @param mc_func motion compensation function pointers (bilinear or sixtap MC)
1023 static inline void vp8_mc(VP8Context *s, int luma,
1024 uint8_t *dst, uint8_t *src, const VP56mv *mv,
1025 int x_off, int y_off, int block_w, int block_h,
1026 int width, int height, int linesize,
1027 vp8_mc_func mc_func[3][3])
1030 static const uint8_t idx[8] = { 0, 1, 2, 1, 2, 1, 2, 1 };
1031 int mx = (mv->x << luma)&7, mx_idx = idx[mx];
1032 int my = (mv->y << luma)&7, my_idx = idx[my];
1034 x_off += mv->x >> (3 - luma);
1035 y_off += mv->y >> (3 - luma);
1038 src += y_off * linesize + x_off;
1039 if (x_off < 2 || x_off >= width - block_w - 3 ||
1040 y_off < 2 || y_off >= height - block_h - 3) {
1041 ff_emulated_edge_mc(s->edge_emu_buffer, src - 2 * linesize - 2, linesize,
1042 block_w + 5, block_h + 5,
1043 x_off - 2, y_off - 2, width, height);
1044 src = s->edge_emu_buffer + 2 + linesize * 2;
1046 mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);
1048 mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0);
1051 static inline void vp8_mc_part(VP8Context *s, uint8_t *dst[3],
1052 AVFrame *ref_frame, int x_off, int y_off,
1053 int bx_off, int by_off,
1054 int block_w, int block_h,
1055 int width, int height, VP56mv *mv)
1060 vp8_mc(s, 1, dst[0] + by_off * s->linesize + bx_off,
1061 ref_frame->data[0], mv, x_off + bx_off, y_off + by_off,
1062 block_w, block_h, width, height, s->linesize,
1063 s->put_pixels_tab[block_w == 8]);
1066 if (s->profile == 3) {
1070 x_off >>= 1; y_off >>= 1;
1071 bx_off >>= 1; by_off >>= 1;
1072 width >>= 1; height >>= 1;
1073 block_w >>= 1; block_h >>= 1;
1074 vp8_mc(s, 0, dst[1] + by_off * s->uvlinesize + bx_off,
1075 ref_frame->data[1], &uvmv, x_off + bx_off, y_off + by_off,
1076 block_w, block_h, width, height, s->uvlinesize,
1077 s->put_pixels_tab[1 + (block_w == 4)]);
1078 vp8_mc(s, 0, dst[2] + by_off * s->uvlinesize + bx_off,
1079 ref_frame->data[2], &uvmv, x_off + bx_off, y_off + by_off,
1080 block_w, block_h, width, height, s->uvlinesize,
1081 s->put_pixels_tab[1 + (block_w == 4)]);
1084 /* Fetch pixels for estimated mv 4 macroblocks ahead.
1085 * Optimized for 64-byte cache lines. Inspired by ffh264 prefetch_motion. */
1086 static inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int mb_xy, int ref)
1088 /* Don't prefetch refs that haven't been used very often this frame. */
1089 if (s->ref_count[ref-1] > (mb_xy >> 5)) {
1090 int x_off = mb_x << 4, y_off = mb_y << 4;
1091 int mx = mb->mv.x + x_off + 8;
1092 int my = mb->mv.y + y_off;
1093 uint8_t **src= s->framep[ref]->data;
1094 int off= mx + (my + (mb_x&3)*4)*s->linesize + 64;
1095 s->dsp.prefetch(src[0]+off, s->linesize, 4);
1096 off= (mx>>1) + ((my>>1) + (mb_x&7))*s->uvlinesize + 64;
1097 s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1102 * Apply motion vectors to prediction buffer, chapter 18.
1104 static void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
1107 int x_off = mb_x << 4, y_off = mb_y << 4;
1108 int width = 16*s->mb_width, height = 16*s->mb_height;
1109 AVFrame *ref = s->framep[mb->ref_frame];
1110 VP56mv *bmv = mb->bmv;
1112 if (mb->mode < VP8_MVMODE_SPLIT) {
1113 vp8_mc_part(s, dst, ref, x_off, y_off,
1114 0, 0, 16, 16, width, height, &mb->mv);
1115 } else switch (mb->partitioning) {
1116 case VP8_SPLITMVMODE_4x4: {
1121 for (y = 0; y < 4; y++) {
1122 for (x = 0; x < 4; x++) {
1123 vp8_mc(s, 1, dst[0] + 4*y*s->linesize + x*4,
1124 ref->data[0], &bmv[4*y + x],
1125 4*x + x_off, 4*y + y_off, 4, 4,
1126 width, height, s->linesize,
1127 s->put_pixels_tab[2]);
1132 x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1;
1133 for (y = 0; y < 2; y++) {
1134 for (x = 0; x < 2; x++) {
1135 uvmv.x = mb->bmv[ 2*y * 4 + 2*x ].x +
1136 mb->bmv[ 2*y * 4 + 2*x+1].x +
1137 mb->bmv[(2*y+1) * 4 + 2*x ].x +
1138 mb->bmv[(2*y+1) * 4 + 2*x+1].x;
1139 uvmv.y = mb->bmv[ 2*y * 4 + 2*x ].y +
1140 mb->bmv[ 2*y * 4 + 2*x+1].y +
1141 mb->bmv[(2*y+1) * 4 + 2*x ].y +
1142 mb->bmv[(2*y+1) * 4 + 2*x+1].y;
1143 uvmv.x = (uvmv.x + 2 + (uvmv.x >> (INT_BIT-1))) >> 2;
1144 uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2;
1145 if (s->profile == 3) {
1149 vp8_mc(s, 0, dst[1] + 4*y*s->uvlinesize + x*4,
1150 ref->data[1], &uvmv,
1151 4*x + x_off, 4*y + y_off, 4, 4,
1152 width, height, s->uvlinesize,
1153 s->put_pixels_tab[2]);
1154 vp8_mc(s, 0, dst[2] + 4*y*s->uvlinesize + x*4,
1155 ref->data[2], &uvmv,
1156 4*x + x_off, 4*y + y_off, 4, 4,
1157 width, height, s->uvlinesize,
1158 s->put_pixels_tab[2]);
1163 case VP8_SPLITMVMODE_16x8:
1164 vp8_mc_part(s, dst, ref, x_off, y_off,
1165 0, 0, 16, 8, width, height, &bmv[0]);
1166 vp8_mc_part(s, dst, ref, x_off, y_off,
1167 0, 8, 16, 8, width, height, &bmv[1]);
1169 case VP8_SPLITMVMODE_8x16:
1170 vp8_mc_part(s, dst, ref, x_off, y_off,
1171 0, 0, 8, 16, width, height, &bmv[0]);
1172 vp8_mc_part(s, dst, ref, x_off, y_off,
1173 8, 0, 8, 16, width, height, &bmv[1]);
1175 case VP8_SPLITMVMODE_8x8:
1176 vp8_mc_part(s, dst, ref, x_off, y_off,
1177 0, 0, 8, 8, width, height, &bmv[0]);
1178 vp8_mc_part(s, dst, ref, x_off, y_off,
1179 8, 0, 8, 8, width, height, &bmv[1]);
1180 vp8_mc_part(s, dst, ref, x_off, y_off,
1181 0, 8, 8, 8, width, height, &bmv[2]);
1182 vp8_mc_part(s, dst, ref, x_off, y_off,
1183 8, 8, 8, 8, width, height, &bmv[3]);
1188 static void idct_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb)
1192 if (mb->mode != MODE_I4x4) {
1193 uint8_t *y_dst = dst[0];
1194 for (y = 0; y < 4; y++) {
1195 uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[y]);
1197 if (nnz4&~0x01010101) {
1198 for (x = 0; x < 4; x++) {
1199 int nnz = s->non_zero_count_cache[y][x];
1202 s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize);
1204 s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize);
1208 s->vp8dsp.vp8_idct_dc_add4y(y_dst, s->block[y], s->linesize);
1211 y_dst += 4*s->linesize;
1215 for (ch = 0; ch < 2; ch++) {
1216 uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[4+ch]);
1218 uint8_t *ch_dst = dst[1+ch];
1219 if (nnz4&~0x01010101) {
1220 for (y = 0; y < 2; y++) {
1221 for (x = 0; x < 2; x++) {
1222 int nnz = s->non_zero_count_cache[4+ch][(y<<1)+x];
1225 s->vp8dsp.vp8_idct_dc_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);
1227 s->vp8dsp.vp8_idct_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);
1230 ch_dst += 4*s->uvlinesize;
1233 s->vp8dsp.vp8_idct_dc_add4uv(ch_dst, s->block[4+ch], s->uvlinesize);
1239 static void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f )
1241 int interior_limit, filter_level;
1243 if (s->segmentation.enabled) {
1244 filter_level = s->segmentation.filter_level[s->segment];
1245 if (!s->segmentation.absolute_vals)
1246 filter_level += s->filter.level;
1248 filter_level = s->filter.level;
1250 if (s->lf_delta.enabled) {
1251 filter_level += s->lf_delta.ref[mb->ref_frame];
1253 if (mb->ref_frame == VP56_FRAME_CURRENT) {
1254 if (mb->mode == MODE_I4x4)
1255 filter_level += s->lf_delta.mode[0];
1257 if (mb->mode == VP8_MVMODE_ZERO)
1258 filter_level += s->lf_delta.mode[1];
1259 else if (mb->mode == VP8_MVMODE_SPLIT)
1260 filter_level += s->lf_delta.mode[3];
1262 filter_level += s->lf_delta.mode[2];
1265 filter_level = av_clip(filter_level, 0, 63);
1267 interior_limit = filter_level;
1268 if (s->filter.sharpness) {
1269 interior_limit >>= s->filter.sharpness > 4 ? 2 : 1;
1270 interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness);
1272 interior_limit = FFMAX(interior_limit, 1);
1274 f->filter_level = filter_level;
1275 f->inner_limit = interior_limit;
1276 f->inner_filter = !mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT;
1279 static void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f, int mb_x, int mb_y)
1281 int mbedge_lim, bedge_lim, hev_thresh;
1282 int filter_level = f->filter_level;
1283 int inner_limit = f->inner_limit;
1284 int inner_filter = f->inner_filter;
1285 int linesize = s->linesize;
1286 int uvlinesize = s->uvlinesize;
1291 mbedge_lim = 2*(filter_level+2) + inner_limit;
1292 bedge_lim = 2* filter_level + inner_limit;
1293 hev_thresh = filter_level >= 15;
1296 if (filter_level >= 40)
1299 if (filter_level >= 40)
1301 else if (filter_level >= 20)
1306 s->vp8dsp.vp8_h_loop_filter16y(dst[0], linesize,
1307 mbedge_lim, inner_limit, hev_thresh);
1308 s->vp8dsp.vp8_h_loop_filter8uv(dst[1], dst[2], uvlinesize,
1309 mbedge_lim, inner_limit, hev_thresh);
1313 s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 4, linesize, bedge_lim,
1314 inner_limit, hev_thresh);
1315 s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 8, linesize, bedge_lim,
1316 inner_limit, hev_thresh);
1317 s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+12, linesize, bedge_lim,
1318 inner_limit, hev_thresh);
1319 s->vp8dsp.vp8_h_loop_filter8uv_inner(dst[1] + 4, dst[2] + 4,
1320 uvlinesize, bedge_lim,
1321 inner_limit, hev_thresh);
1325 s->vp8dsp.vp8_v_loop_filter16y(dst[0], linesize,
1326 mbedge_lim, inner_limit, hev_thresh);
1327 s->vp8dsp.vp8_v_loop_filter8uv(dst[1], dst[2], uvlinesize,
1328 mbedge_lim, inner_limit, hev_thresh);
1332 s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 4*linesize,
1333 linesize, bedge_lim,
1334 inner_limit, hev_thresh);
1335 s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 8*linesize,
1336 linesize, bedge_lim,
1337 inner_limit, hev_thresh);
1338 s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+12*linesize,
1339 linesize, bedge_lim,
1340 inner_limit, hev_thresh);
1341 s->vp8dsp.vp8_v_loop_filter8uv_inner(dst[1] + 4 * uvlinesize,
1342 dst[2] + 4 * uvlinesize,
1343 uvlinesize, bedge_lim,
1344 inner_limit, hev_thresh);
1348 static void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f, int mb_x, int mb_y)
1350 int mbedge_lim, bedge_lim;
1351 int filter_level = f->filter_level;
1352 int inner_limit = f->inner_limit;
1353 int inner_filter = f->inner_filter;
1354 int linesize = s->linesize;
1359 mbedge_lim = 2*(filter_level+2) + inner_limit;
1360 bedge_lim = 2* filter_level + inner_limit;
1363 s->vp8dsp.vp8_h_loop_filter_simple(dst, linesize, mbedge_lim);
1365 s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, linesize, bedge_lim);
1366 s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, linesize, bedge_lim);
1367 s->vp8dsp.vp8_h_loop_filter_simple(dst+12, linesize, bedge_lim);
1371 s->vp8dsp.vp8_v_loop_filter_simple(dst, linesize, mbedge_lim);
1373 s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*linesize, linesize, bedge_lim);
1374 s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*linesize, linesize, bedge_lim);
1375 s->vp8dsp.vp8_v_loop_filter_simple(dst+12*linesize, linesize, bedge_lim);
1379 static void filter_mb_row(VP8Context *s, int mb_y)
1381 VP8FilterStrength *f = s->filter_strength;
1383 s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize,
1384 s->framep[VP56_FRAME_CURRENT]->data[1] + 8*mb_y*s->uvlinesize,
1385 s->framep[VP56_FRAME_CURRENT]->data[2] + 8*mb_y*s->uvlinesize
1389 for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1390 backup_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], s->linesize, s->uvlinesize, 0);
1391 filter_mb(s, dst, f++, mb_x, mb_y);
1398 static void filter_mb_row_simple(VP8Context *s, int mb_y)
1400 VP8FilterStrength *f = s->filter_strength;
1401 uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize;
1404 for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1405 backup_mb_border(s->top_border[mb_x+1], dst, NULL, NULL, s->linesize, 0, 1);
1406 filter_mb_simple(s, dst, f++, mb_x, mb_y);
1411 static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
1414 VP8Context *s = avctx->priv_data;
1415 int ret, mb_x, mb_y, i, y, referenced;
1416 enum AVDiscard skip_thresh;
1417 AVFrame *curframe = NULL;
1419 if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0)
1422 referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT
1423 || s->update_altref == VP56_FRAME_CURRENT;
1425 skip_thresh = !referenced ? AVDISCARD_NONREF :
1426 !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL;
1428 if (avctx->skip_frame >= skip_thresh) {
1432 s->deblock_filter = s->filter.level && avctx->skip_loop_filter < skip_thresh;
1434 for (i = 0; i < 4; i++)
1435 if (&s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1436 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1437 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) {
1438 curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i];
1441 if (curframe->data[0])
1442 avctx->release_buffer(avctx, curframe);
1444 curframe->key_frame = s->keyframe;
1445 curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE;
1446 curframe->reference = referenced ? 3 : 0;
1447 if ((ret = avctx->get_buffer(avctx, curframe))) {
1448 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n");
1452 // Given that arithmetic probabilities are updated every frame, it's quite likely
1453 // that the values we have on a random interframe are complete junk if we didn't
1454 // start decode on a keyframe. So just don't display anything rather than junk.
1455 if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] ||
1456 !s->framep[VP56_FRAME_GOLDEN] ||
1457 !s->framep[VP56_FRAME_GOLDEN2])) {
1458 av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n");
1459 return AVERROR_INVALIDDATA;
1462 s->linesize = curframe->linesize[0];
1463 s->uvlinesize = curframe->linesize[1];
1465 if (!s->edge_emu_buffer)
1466 s->edge_emu_buffer = av_malloc(21*s->linesize);
1468 memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz));
1470 /* Zero macroblock structures for top/left prediction from outside the frame. */
1471 memset(s->macroblocks, 0, (s->mb_width + s->mb_height*2)*sizeof(*s->macroblocks));
1473 // top edge of 127 for intra prediction
1474 memset(s->top_border, 127, (s->mb_width+1)*sizeof(*s->top_border));
1475 memset(s->ref_count, 0, sizeof(s->ref_count));
1477 for (mb_y = 0; mb_y < s->mb_height; mb_y++) {
1478 VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)];
1479 VP8Macroblock *mb = s->macroblocks + (s->mb_height - mb_y - 1)*2;
1480 uint8_t *intra4x4 = s->intra4x4_pred_mode + 4*mb_y*s->b4_stride;
1481 uint8_t *segment_map = s->segmentation_map + mb_y*s->mb_stride;
1482 int mb_xy = mb_y * s->mb_stride;
1484 curframe->data[0] + 16*mb_y*s->linesize,
1485 curframe->data[1] + 8*mb_y*s->uvlinesize,
1486 curframe->data[2] + 8*mb_y*s->uvlinesize
1489 memset(s->left_nnz, 0, sizeof(s->left_nnz));
1491 // left edge of 129 for intra prediction
1492 if (!(avctx->flags & CODEC_FLAG_EMU_EDGE))
1493 for (i = 0; i < 3; i++)
1494 for (y = 0; y < 16>>!!i; y++)
1495 dst[i][y*curframe->linesize[i]-1] = 129;
1497 memset(s->top_border, 129, sizeof(*s->top_border));
1499 for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb_xy++, mb++) {
1500 uint8_t *intra4x4_mb = s->keyframe ? intra4x4 + 4*mb_x : s->intra4x4_pred_mode_mb;
1501 uint8_t *segment_mb = segment_map+mb_x;
1503 /* Prefetch the current frame, 4 MBs ahead */
1504 s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4);
1505 s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2);
1507 decode_mb_mode(s, mb, mb_x, mb_y, intra4x4_mb, segment_mb);
1509 prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_PREVIOUS);
1512 decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);
1514 if (mb->mode <= MODE_I4x4)
1515 intra_predict(s, dst, mb, intra4x4_mb, mb_x, mb_y);
1517 inter_predict(s, dst, mb, mb_x, mb_y);
1519 prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN);
1522 idct_mb(s, dst, mb);
1524 AV_ZERO64(s->left_nnz);
1525 AV_WN64(s->top_nnz[mb_x], 0); // array of 9, so unaligned
1527 // Reset DC block predictors if they would exist if the mb had coefficients
1528 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
1530 s->top_nnz[mb_x][8] = 0;
1534 if (s->deblock_filter)
1535 filter_level_for_mb(s, mb, &s->filter_strength[mb_x]);
1537 prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN2);
1543 if (s->deblock_filter) {
1544 if (s->filter.simple)
1545 filter_mb_row_simple(s, mb_y);
1547 filter_mb_row(s, mb_y);
1552 // if future frames don't use the updated probabilities,
1553 // reset them to the values we saved
1554 if (!s->update_probabilities)
1555 s->prob[0] = s->prob[1];
1557 // check if golden and altref are swapped
1558 if (s->update_altref == VP56_FRAME_GOLDEN &&
1559 s->update_golden == VP56_FRAME_GOLDEN2)
1560 FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN], s->framep[VP56_FRAME_GOLDEN2]);
1562 if (s->update_altref != VP56_FRAME_NONE)
1563 s->framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref];
1565 if (s->update_golden != VP56_FRAME_NONE)
1566 s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden];
1569 if (s->update_last) // move cur->prev
1570 s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT];
1572 // release no longer referenced frames
1573 for (i = 0; i < 4; i++)
1574 if (s->frames[i].data[0] &&
1575 &s->frames[i] != s->framep[VP56_FRAME_CURRENT] &&
1576 &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1577 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1578 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2])
1579 avctx->release_buffer(avctx, &s->frames[i]);
1581 if (!s->invisible) {
1582 *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT];
1583 *data_size = sizeof(AVFrame);
1589 static av_cold int vp8_decode_init(AVCodecContext *avctx)
1591 VP8Context *s = avctx->priv_data;
1594 avctx->pix_fmt = PIX_FMT_YUV420P;
1596 dsputil_init(&s->dsp, avctx);
1597 ff_h264_pred_init(&s->hpc, CODEC_ID_VP8);
1598 ff_vp8dsp_init(&s->vp8dsp);
1600 // intra pred needs edge emulation among other things
1601 if (avctx->flags&CODEC_FLAG_EMU_EDGE) {
1602 av_log(avctx, AV_LOG_ERROR, "Edge emulation not supported\n");
1603 return AVERROR_PATCHWELCOME;
1609 static av_cold int vp8_decode_free(AVCodecContext *avctx)
1611 vp8_decode_flush(avctx);
1615 AVCodec vp8_decoder = {
1625 .flush = vp8_decode_flush,
1626 .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"),