Reindent
[libav.git] / libavcodec / vp8.c
CommitLineData
3b636f21
DC
1/**
2 * VP8 compatible video decoder
3 *
4 * Copyright (C) 2010 David Conrad
5 * Copyright (C) 2010 Ronald S. Bultje
6 *
7 * This file is part of FFmpeg.
8 *
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.
13 *
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.
18 *
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
22 */
23
24#include "avcodec.h"
25#include "vp56.h"
26#include "vp8data.h"
27#include "vp8dsp.h"
28#include "h264pred.h"
29#include "rectangle.h"
30
31typedef struct {
32 uint8_t segment;
33 uint8_t skip;
34 // todo: make it possible to check for at least (i4x4 or split_mv)
35 // in one op. are others needed?
36 uint8_t mode;
37 uint8_t ref_frame;
38 uint8_t partitioning;
39 VP56mv mv;
40 VP56mv bmv[16];
41} VP8Macroblock;
42
43typedef struct {
44 AVCodecContext *avctx;
45 DSPContext dsp;
46 VP8DSPContext vp8dsp;
47 H264PredContext hpc;
0ef1dbed 48 vp8_mc_func put_pixels_tab[3][3][3];
3b636f21
DC
49 AVFrame frames[4];
50 AVFrame *framep[4];
51 uint8_t *edge_emu_buffer;
52 VP56RangeCoder c; ///< header context, includes mb modes and motion vectors
53 int profile;
54
55 int mb_width; /* number of horizontal MB */
56 int mb_height; /* number of vertical MB */
57 int linesize;
58 int uvlinesize;
59
60 int keyframe;
61 int invisible;
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
64 int update_altref;
65
66 /**
67 * If this flag is not set, all the probability updates
68 * are discarded after this frame is decoded.
69 */
70 int update_probabilities;
71
72 /**
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.
75 */
76 int num_coeff_partitions;
77 VP56RangeCoder coeff_partition[8];
78
79 VP8Macroblock *macroblocks;
80 VP8Macroblock *macroblocks_base;
81 int mb_stride;
82
83 uint8_t *intra4x4_pred_mode;
84 uint8_t *intra4x4_pred_mode_base;
85 int b4_stride;
86
87 /**
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.
92 */
93 uint8_t (*top_nnz)[9];
94 DECLARE_ALIGNED(8, uint8_t, left_nnz)[9];
95
96 /**
97 * This is the index plus one of the last non-zero coeff
98 * for each of the blocks in the current macroblock.
99 * So, 0 -> no coeffs
100 * 1 -> dc-only (special transform)
101 * 2+-> full transform
102 */
103 DECLARE_ALIGNED(16, uint8_t, non_zero_count_cache)[6][4];
104 DECLARE_ALIGNED(16, DCTELEM, block)[6][4][16];
105
106 int chroma_pred_mode; ///< 8x8c pred mode of the current macroblock
107
108 int mbskip_enabled;
109 int sign_bias[4]; ///< one state [0, 1] per ref frame type
110
111 /**
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.
115 */
116 struct {
117 int enabled;
118 int absolute_vals;
119 int update_map;
120 int8_t base_quant[4];
121 int8_t filter_level[4]; ///< base loop filter level
122 } segmentation;
123
124 /**
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.
128 */
129 struct {
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];
134 } qmat[4];
135
136 struct {
137 int simple;
138 int level;
139 int sharpness;
140 } filter;
141
142 struct {
143 int enabled; ///< whether each mb can have a different strength based on mode/ref
144
145 /**
146 * filter strength adjustment for the following macroblock modes:
147 * [0] - i4x4
148 * [1] - zero mv
149 * [2] - inter modes except for zero or split mv
150 * [3] - split mv
151 * i16x16 modes never have any adjustment
152 */
153 int8_t mode[4];
154
155 /**
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
161 */
162 int8_t ref[4];
163 } lf_delta;
164
165 /**
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!
170 */
171 struct {
172 uint8_t segmentid[3];
173 uint8_t mbskip;
174 uint8_t intra;
175 uint8_t last;
176 uint8_t golden;
177 uint8_t pred16x16[4];
178 uint8_t pred8x8c[3];
179 uint8_t token[4][8][3][NUM_DCT_TOKENS-1];
180 uint8_t mvc[2][19];
181 } prob[2];
182} VP8Context;
183
184#define RL24(p) (AV_RL16(p) + ((p)[2] << 16))
185
186static void vp8_decode_flush(AVCodecContext *avctx)
187{
188 VP8Context *s = avctx->priv_data;
189 int i;
190
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));
195
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);
200
201 s->macroblocks = NULL;
202 s->intra4x4_pred_mode = NULL;
203}
204
205static int update_dimensions(VP8Context *s, int width, int height)
206{
207 int i;
208
209 if (avcodec_check_dimensions(s->avctx, width, height))
210 return AVERROR_INVALIDDATA;
211
212 vp8_decode_flush(s->avctx);
213
214 avcodec_set_dimensions(s->avctx, width, height);
215
216 s->mb_width = (s->avctx->coded_width +15) / 16;
217 s->mb_height = (s->avctx->coded_height+15) / 16;
218
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;
223
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));
227
228 s->macroblocks = s->macroblocks_base + 1 + s->mb_stride;
229 s->intra4x4_pred_mode = s->intra4x4_pred_mode_base + 4 + s->b4_stride;
230
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;
234
235 return 0;
236}
237
238static void parse_segment_info(VP8Context *s)
239{
240 VP56RangeCoder *c = &s->c;
241 int i;
242
243 s->segmentation.update_map = vp8_rac_get(c);
244
245 if (vp8_rac_get(c)) { // update segment feature data
246 s->segmentation.absolute_vals = vp8_rac_get(c);
247
248 for (i = 0; i < 4; i++)
249 s->segmentation.base_quant[i] = vp8_rac_get_sint(c, 7);
250
251 for (i = 0; i < 4; i++)
252 s->segmentation.filter_level[i] = vp8_rac_get_sint(c, 6);
253 }
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;
257}
258
259static void update_lf_deltas(VP8Context *s)
260{
261 VP56RangeCoder *c = &s->c;
262 int i;
263
264 for (i = 0; i < 4; i++)
265 s->lf_delta.ref[i] = vp8_rac_get_sint(c, 6);
266
267 for (i = 0; i < 4; i++)
268 s->lf_delta.mode[i] = vp8_rac_get_sint(c, 6);
269}
270
271static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size)
272{
273 const uint8_t *sizes = buf;
274 int i;
275
276 s->num_coeff_partitions = 1 << vp8_rac_get_uint(&s->c, 2);
277
278 buf += 3*(s->num_coeff_partitions-1);
279 buf_size -= 3*(s->num_coeff_partitions-1);
280 if (buf_size < 0)
281 return -1;
282
283 for (i = 0; i < s->num_coeff_partitions-1; i++) {
284 int size = RL24(sizes + 3*i);
285 if (buf_size - size < 0)
286 return -1;
287
288 vp56_init_range_decoder(&s->coeff_partition[i], buf, size);
289 buf += size;
290 buf_size -= size;
291 }
292 vp56_init_range_decoder(&s->coeff_partition[i], buf, buf_size);
293
294 return 0;
295}
296
297static void get_quants(VP8Context *s)
298{
299 VP56RangeCoder *c = &s->c;
300 int i, base_qi;
301
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);
308
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)
313 base_qi += yac_qi;
314 } else
315 base_qi = yac_qi;
316
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)];
323
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);
326 }
327}
328
329/**
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
332 *
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:
338 * 0: no update
339 * 1: VP56_FRAME_PREVIOUS
340 * 2: update golden with altref, or update altref with golden
341 */
342static VP56Frame ref_to_update(VP8Context *s, int update, VP56Frame ref)
343{
344 VP56RangeCoder *c = &s->c;
345
346 if (update)
347 return VP56_FRAME_CURRENT;
348
349 switch (vp8_rac_get_uint(c, 2)) {
350 case 1:
351 return VP56_FRAME_PREVIOUS;
352 case 2:
353 return (ref == VP56_FRAME_GOLDEN) ? VP56_FRAME_GOLDEN2 : VP56_FRAME_GOLDEN;
354 }
355 return VP56_FRAME_NONE;
356}
357
358static void update_refs(VP8Context *s)
359{
360 VP56RangeCoder *c = &s->c;
361
362 int update_golden = vp8_rac_get(c);
363 int update_altref = vp8_rac_get(c);
364
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);
367}
368
369static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)
370{
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;
375
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;
380 buf += 3;
381 buf_size -= 3;
382
0ef1dbed
DC
383 if (s->profile > 3)
384 av_log(s->avctx, AV_LOG_WARNING, "Unknown profile %d\n", s->profile);
385
386 if (!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));
3b636f21
DC
390
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;
394 }
395
396 if (s->keyframe) {
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;
400 }
401 width = AV_RL16(buf+3) & 0x3fff;
402 height = AV_RL16(buf+5) & 0x3fff;
403 hscale = buf[4] >> 6;
404 vscale = buf[6] >> 6;
405 buf += 7;
406 buf_size -= 7;
407
92a54426
MR
408 if (hscale || vscale)
409 av_log_missing_feature(s->avctx, "Upscaling", 1);
410
3b636f21
DC
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));
417 }
418
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))
422 return ret;
423 }
424
425 vp56_init_range_decoder(c, buf, header_size);
426 buf += header_size;
427 buf_size -= header_size;
428
429 if (s->keyframe) {
430 if (vp8_rac_get(c))
431 av_log(s->avctx, AV_LOG_WARNING, "Unspecified colorspace\n");
432 vp8_rac_get(c); // whether we can skip clamping in dsp functions
433 }
434
435 if ((s->segmentation.enabled = vp8_rac_get(c)))
436 parse_segment_info(s);
437 else
438 s->segmentation.update_map = 0; // FIXME: move this to some init function?
439
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);
443
444 if ((s->lf_delta.enabled = vp8_rac_get(c)))
445 if (vp8_rac_get(c))
446 update_lf_deltas(s);
447
448 if (setup_partitions(s, buf, buf_size)) {
449 av_log(s->avctx, AV_LOG_ERROR, "Invalid partitions\n");
450 return AVERROR_INVALIDDATA;
451 }
452
453 get_quants(s);
454
455 if (!s->keyframe) {
456 update_refs(s);
457 s->sign_bias[VP56_FRAME_GOLDEN] = vp8_rac_get(c);
458 s->sign_bias[VP56_FRAME_GOLDEN2 /* altref */] = vp8_rac_get(c);
459 }
460
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];
465
466 s->update_last = s->keyframe || vp8_rac_get(c);
467
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);
474
475 if ((s->mbskip_enabled = vp8_rac_get(c)))
476 s->prob->mbskip = vp8_rac_get_uint(c, 8);
477
478 if (!s->keyframe) {
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);
482
483 if (vp8_rac_get(c))
484 for (i = 0; i < 4; i++)
485 s->prob->pred16x16[i] = vp8_rac_get_uint(c, 8);
486 if (vp8_rac_get(c))
487 for (i = 0; i < 3; i++)
488 s->prob->pred8x8c[i] = vp8_rac_get_uint(c, 8);
489
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);
495 }
496
497 return 0;
498}
499
500static inline void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src,
501 int mb_x, int mb_y)
502{
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);
508}
509
510static void find_near_mvs(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
511 VP56mv near[2], VP56mv *best, int cnt[4])
512{
513 VP8Macroblock *mb_edge[3] = { mb - s->mb_stride /* top */,
514 mb - 1 /* left */,
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;
521
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]) {
529 tmp.x *= -1;
530 tmp.y *= -1;
531 }
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);
535 } else
536 cnt[CNT_ZERO] += 1 + (n != 2);
537 }
538 }
539
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;
545
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);
549
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]);
554 }
555
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;
559
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];
563}
564
565/**
566 * Motion vector coding, 17.1.
567 */
568static int read_mv_component(VP56RangeCoder *c, const uint8_t *p)
569{
570 int x = 0;
571
572 if (vp56_rac_get_prob(c, p[0])) {
573 int i;
574
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]))
580 x += 8;
581 } else
582 x = vp8_rac_get_tree(c, vp8_small_mvtree, &p[2]);
583
584 return (x && vp56_rac_get_prob(c, p[1])) ? -x : x;
585}
586
587static const uint8_t *get_submv_prob(const VP56mv *left, const VP56mv *top)
588{
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));
592
593 if (equal)
594 return l_is_zero ? vp8_submv_prob[4] : vp8_submv_prob[3];
595 if (t_is_zero)
596 return vp8_submv_prob[2];
597 return l_is_zero ? vp8_submv_prob[1] : vp8_submv_prob[0];
598}
599
600/**
601 * Split motion vector prediction, 16.4.
602 */
603static void decode_splitmvs(VP8Context *s, VP56RangeCoder *c,
604 VP8Macroblock *mb, VP56mv *base_mv)
605{
606 int part_idx = mb->partitioning =
607 vp8_rac_get_tree(c, vp8_mbsplit_tree, vp8_mbsplit_prob);
608 int n, num = vp8_mbsplit_count[part_idx];
609 VP56mv part_mv[16];
610
611 for (n = 0; n < num; n++) {
612 int k = vp8_mbfirstidx[part_idx][n];
613 const VP56mv *left = (k & 3) ? &mb->bmv[k - 1] : &mb[-1].bmv[k + 3],
614 *above = (k > 3) ? &mb->bmv[k - 4] : &mb[-s->mb_stride].bmv[k + 12];
615 const uint8_t *submv_prob = get_submv_prob(left, above);
616
617 switch (vp8_rac_get_tree(c, vp8_submv_ref_tree, submv_prob)) {
618 case VP8_SUBMVMODE_NEW4X4:
619 part_mv[n].y = base_mv->y + read_mv_component(c, s->prob->mvc[0]);
620 part_mv[n].x = base_mv->x + read_mv_component(c, s->prob->mvc[1]);
621 break;
622 case VP8_SUBMVMODE_ZERO4X4:
623 part_mv[n].x = 0;
624 part_mv[n].y = 0;
625 break;
626 case VP8_SUBMVMODE_LEFT4X4:
627 part_mv[n] = *left;
628 break;
629 case VP8_SUBMVMODE_TOP4X4:
630 part_mv[n] = *above;
631 break;
632 }
633
634 /* fill out over the 4x4 blocks in MB */
635 for (k = 0; k < 16; k++)
636 if (vp8_mbsplits[part_idx][k] == n) {
637 mb->bmv[k] = part_mv[n];
638 }
639 }
640}
641
642static inline void decode_intra4x4_modes(VP56RangeCoder *c, uint8_t *intra4x4,
643 int stride, int keyframe)
644{
645 int x, y, t, l;
646 const uint8_t *ctx = vp8_pred4x4_prob_inter;
647
648 for (y = 0; y < 4; y++) {
649 for (x = 0; x < 4; x++) {
650 if (keyframe) {
651 t = intra4x4[x - stride];
652 l = intra4x4[x - 1];
653 ctx = vp8_pred4x4_prob_intra[t][l];
654 }
655 intra4x4[x] = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx);
656 }
657 intra4x4 += stride;
658 }
659}
660
661static void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
662 uint8_t *intra4x4)
663{
664 VP56RangeCoder *c = &s->c;
665 int n;
666
667 if (s->segmentation.update_map)
668 mb->segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid);
669
670 mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0;
671
672 if (s->keyframe) {
673 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra);
674
675 if (mb->mode == MODE_I4x4) {
676 decode_intra4x4_modes(c, intra4x4, s->b4_stride, 1);
677 } else
678 fill_rectangle(intra4x4, 4, 4, s->b4_stride, vp8_pred4x4_mode[mb->mode], 1);
679
680 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra);
681 mb->ref_frame = VP56_FRAME_CURRENT;
682 } else if (vp56_rac_get_prob(c, s->prob->intra)) {
683 VP56mv near[2], best;
684 int cnt[4] = { 0 };
685 uint8_t p[4];
686
687 // inter MB, 16.2
688 if (vp56_rac_get_prob(c, s->prob->last))
689 mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ?
690 VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN;
691 else
692 mb->ref_frame = VP56_FRAME_PREVIOUS;
693
694 // motion vectors, 16.3
695 find_near_mvs(s, mb, mb_x, mb_y, near, &best, cnt);
696 for (n = 0; n < 4; n++)
697 p[n] = vp8_mode_contexts[cnt[n]][n];
698 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_mvinter, p);
699 switch (mb->mode) {
700 case VP8_MVMODE_SPLIT:
701 decode_splitmvs(s, c, mb, &best);
702 mb->mv = mb->bmv[15];
703 break;
704 case VP8_MVMODE_ZERO:
705 mb->mv.x = 0;
706 mb->mv.y = 0;
707 break;
708 case VP8_MVMODE_NEAREST:
709 clamp_mv(s, &mb->mv, &near[0], mb_x, mb_y);
710 break;
711 case VP8_MVMODE_NEAR:
712 clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y);
713 break;
714 case VP8_MVMODE_NEW:
715 mb->mv.y = best.y + read_mv_component(c, s->prob->mvc[0]);
716 mb->mv.x = best.x + read_mv_component(c, s->prob->mvc[1]);
717 break;
718 }
719 if (mb->mode != VP8_MVMODE_SPLIT) {
720 for (n = 0; n < 16; n++)
721 mb->bmv[n] = mb->mv;
722 }
723 } else {
724 // intra MB, 16.1
725 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16);
726
727 if (mb->mode == MODE_I4x4) {
728 decode_intra4x4_modes(c, intra4x4, s->b4_stride, 0);
729 } else
730 fill_rectangle(intra4x4, 4, 4, s->b4_stride, vp8_pred4x4_mode[mb->mode], 1);
731
732 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c);
733 mb->ref_frame = VP56_FRAME_CURRENT;
734 }
735}
736
737/**
738 * @param i initial coeff index, 0 unless a separate DC block is coded
739 * @param zero_nhood the initial prediction context for number of surrounding
740 * all-zero blocks (only left/top, so 0-2)
741 * @param qmul[0] dc dequant factor
742 * @param qmul[1] ac dequant factor
743 * @return 0 if no coeffs were decoded
744 * otherwise, the index of the last coeff decoded plus one
745 */
746static int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16],
747 uint8_t probs[8][3][NUM_DCT_TOKENS-1],
748 int i, int zero_nhood, int16_t qmul[2])
749{
750 int token, nonzero = 0;
751 int offset = 0;
752
753 for (; i < 16; i++) {
754 token = vp8_rac_get_tree_with_offset(c, vp8_coeff_tree, probs[vp8_coeff_band[i]][zero_nhood], offset);
755
756 if (token == DCT_EOB)
757 break;
758 else if (token >= DCT_CAT1) {
759 int cat = token-DCT_CAT1;
760 token = vp8_rac_get_coeff(c, vp8_dct_cat_prob[cat]);
761 token += vp8_dct_cat_offset[cat];
762 }
763
764 // after the first token, the non-zero prediction context becomes
765 // based on the last decoded coeff
766 if (!token) {
767 zero_nhood = 0;
768 offset = 1;
769 continue;
770 } else if (token == 1)
771 zero_nhood = 1;
772 else
773 zero_nhood = 2;
774
775 // todo: full [16] qmat? load into register?
776 block[zigzag_scan[i]] = (vp8_rac_get(c) ? -token : token) * qmul[!!i];
777 nonzero = i+1;
778 offset = 0;
779 }
780 return nonzero;
781}
782
783static void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,
784 uint8_t t_nnz[9], uint8_t l_nnz[9])
785{
786 LOCAL_ALIGNED_16(DCTELEM, dc,[16]);
787 int i, x, y, luma_start = 0, luma_ctx = 3;
788 int nnz_pred, nnz, nnz_total = 0;
789 int segment = s->segmentation.enabled ? mb->segment : 0;
790
791 s->dsp.clear_blocks((DCTELEM *)s->block);
792
793 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
794 AV_ZERO128(dc);
795 AV_ZERO128(dc+8);
796 nnz_pred = t_nnz[8] + l_nnz[8];
797
798 // decode DC values and do hadamard
799 nnz = decode_block_coeffs(c, dc, s->prob->token[1], 0, nnz_pred,
800 s->qmat[segment].luma_dc_qmul);
801 l_nnz[8] = t_nnz[8] = !!nnz;
802 nnz_total += nnz;
803 s->vp8dsp.vp8_luma_dc_wht(s->block, dc);
804 luma_start = 1;
805 luma_ctx = 0;
806 }
807
808 // luma blocks
809 for (y = 0; y < 4; y++)
810 for (x = 0; x < 4; x++) {
811 nnz_pred = l_nnz[y] + t_nnz[x];
812 nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start,
813 nnz_pred, s->qmat[segment].luma_qmul);
814 // nnz+luma_start may be one more than the actual last index, but we don't care
815 s->non_zero_count_cache[y][x] = nnz + luma_start;
816 t_nnz[x] = l_nnz[y] = !!nnz;
817 nnz_total += nnz;
818 }
819
820 // chroma blocks
821 // TODO: what to do about dimensions? 2nd dim for luma is x,
822 // but for chroma it's (y<<1)|x
823 for (i = 4; i < 6; i++)
824 for (y = 0; y < 2; y++)
825 for (x = 0; x < 2; x++) {
826 nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x];
827 nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0,
828 nnz_pred, s->qmat[segment].chroma_qmul);
829 s->non_zero_count_cache[i][(y<<1)+x] = nnz;
830 t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz;
831 nnz_total += nnz;
832 }
833
834 // if there were no coded coeffs despite the macroblock not being marked skip,
835 // we MUST not do the inner loop filter and should not do IDCT
836 // Since skip isn't used for bitstream prediction, just manually set it.
837 if (!nnz_total)
838 mb->skip = 1;
839}
840
841static int check_intra_pred_mode(int mode, int mb_x, int mb_y)
842{
843 if (mode == DC_PRED8x8) {
844 if (!(mb_x|mb_y))
845 mode = DC_128_PRED8x8;
846 else if (!mb_y)
847 mode = LEFT_DC_PRED8x8;
848 else if (!mb_x)
849 mode = TOP_DC_PRED8x8;
850 }
851 return mode;
852}
853
854static void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
855 uint8_t *bmode, int mb_x, int mb_y)
856{
857 int x, y, mode, nnz, tr;
858
859 if (mb->mode < MODE_I4x4) {
860 mode = check_intra_pred_mode(mb->mode, mb_x, mb_y);
861 s->hpc.pred16x16[mode](dst[0], s->linesize);
862 } else {
863 uint8_t *ptr = dst[0];
864
865 // all blocks on the right edge of the macroblock use bottom edge
866 // the top macroblock for their topright edge
867 uint8_t *tr_right = ptr - s->linesize + 16;
868
869 // if we're on the right edge of the frame, said edge is extended
870 // from the top macroblock
871 if (mb_x == s->mb_width-1) {
872 tr = tr_right[-1]*0x01010101;
873 tr_right = (uint8_t *)&tr;
874 }
875
876 for (y = 0; y < 4; y++) {
877 uint8_t *topright = ptr + 4 - s->linesize;
878 for (x = 0; x < 4; x++) {
879 if (x == 3)
880 topright = tr_right;
881
882 s->hpc.pred4x4[bmode[x]](ptr+4*x, topright, s->linesize);
883
884 nnz = s->non_zero_count_cache[y][x];
885 if (nnz) {
886 if (nnz == 1)
887 s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize);
888 else
889 s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize);
890 }
891 topright += 4;
892 }
893
894 ptr += 4*s->linesize;
895 bmode += s->b4_stride;
896 }
897 }
898
899 mode = check_intra_pred_mode(s->chroma_pred_mode, mb_x, mb_y);
900 s->hpc.pred8x8[mode](dst[1], s->uvlinesize);
901 s->hpc.pred8x8[mode](dst[2], s->uvlinesize);
902}
903
904/**
905 * Generic MC function.
906 *
907 * @param s VP8 decoding context
908 * @param luma 1 for luma (Y) planes, 0 for chroma (Cb/Cr) planes
909 * @param dst target buffer for block data at block position
910 * @param src reference picture buffer at origin (0, 0)
911 * @param mv motion vector (relative to block position) to get pixel data from
912 * @param x_off horizontal position of block from origin (0, 0)
913 * @param y_off vertical position of block from origin (0, 0)
914 * @param block_w width of block (16, 8 or 4)
915 * @param block_h height of block (always same as block_w)
916 * @param width width of src/dst plane data
917 * @param height height of src/dst plane data
918 * @param linesize size of a single line of plane data, including padding
919 */
920static inline void vp8_mc(VP8Context *s, int luma,
921 uint8_t *dst, uint8_t *src, const VP56mv *mv,
922 int x_off, int y_off, int block_w, int block_h,
923 int width, int height, int linesize,
d6f8476b 924 vp8_mc_func mc_func[3][3])
3b636f21
DC
925{
926 static const uint8_t idx[8] = { 0, 1, 2, 1, 2, 1, 2, 1 };
927 int mx = (mv->x << luma)&7, mx_idx = idx[mx];
928 int my = (mv->y << luma)&7, my_idx = idx[my];
929
930 x_off += mv->x >> (3 - luma);
931 y_off += mv->y >> (3 - luma);
932
933 // edge emulation
934 src += y_off * linesize + x_off;
935 if (x_off < 2 || x_off >= width - block_w - 3 ||
936 y_off < 2 || y_off >= height - block_h - 3) {
937 ff_emulated_edge_mc(s->edge_emu_buffer, src - 2 * linesize - 2, linesize,
938 block_w + 5, block_h + 5,
939 x_off - 2, y_off - 2, width, height);
940 src = s->edge_emu_buffer + 2 + linesize * 2;
941 }
942
d6f8476b 943 mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);
3b636f21
DC
944}
945
946/**
947 * Apply motion vectors to prediction buffer, chapter 18.
948 */
949static void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
950 int mb_x, int mb_y)
951{
952 int x_off = mb_x << 4, y_off = mb_y << 4;
953 int width = 16*s->mb_width, height = 16*s->mb_height;
954 VP56mv uvmv;
955
956 if (mb->mode < VP8_MVMODE_SPLIT) {
957 /* Y */
958 vp8_mc(s, 1, dst[0], s->framep[mb->ref_frame]->data[0], &mb->mv,
959 x_off, y_off, 16, 16, width, height, s->linesize,
0ef1dbed 960 s->put_pixels_tab[0]);
3b636f21
DC
961
962 /* U/V */
963 uvmv = mb->mv;
964 if (s->profile == 3) {
965 uvmv.x &= ~7;
966 uvmv.y &= ~7;
967 }
968 x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1;
969 vp8_mc(s, 0, dst[1], s->framep[mb->ref_frame]->data[1], &uvmv,
970 x_off, y_off, 8, 8, width, height, s->uvlinesize,
0ef1dbed 971 s->put_pixels_tab[1]);
3b636f21
DC
972 vp8_mc(s, 0, dst[2], s->framep[mb->ref_frame]->data[2], &uvmv,
973 x_off, y_off, 8, 8, width, height, s->uvlinesize,
0ef1dbed 974 s->put_pixels_tab[1]);
3b636f21
DC
975 } else {
976 int x, y;
977
978 /* Y */
979 for (y = 0; y < 4; y++) {
980 for (x = 0; x < 4; x++) {
981 vp8_mc(s, 1, dst[0] + 4*y*s->linesize + x*4,
982 s->framep[mb->ref_frame]->data[0], &mb->bmv[4*y + x],
983 4*x + x_off, 4*y + y_off, 4, 4,
984 width, height, s->linesize,
0ef1dbed 985 s->put_pixels_tab[2]);
3b636f21
DC
986 }
987 }
988
989 /* U/V */
990 x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1;
991 for (y = 0; y < 2; y++) {
992 for (x = 0; x < 2; x++) {
993 uvmv.x = mb->bmv[ 2*y * 4 + 2*x ].x +
994 mb->bmv[ 2*y * 4 + 2*x+1].x +
995 mb->bmv[(2*y+1) * 4 + 2*x ].x +
996 mb->bmv[(2*y+1) * 4 + 2*x+1].x;
997 uvmv.y = mb->bmv[ 2*y * 4 + 2*x ].y +
998 mb->bmv[ 2*y * 4 + 2*x+1].y +
999 mb->bmv[(2*y+1) * 4 + 2*x ].y +
1000 mb->bmv[(2*y+1) * 4 + 2*x+1].y;
8f910a56
SG
1001 uvmv.x = (uvmv.x + 2 + (uvmv.x >> (INT_BIT-1))) >> 2;
1002 uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2;
3b636f21
DC
1003 if (s->profile == 3) {
1004 uvmv.x &= ~7;
1005 uvmv.y &= ~7;
1006 }
1007 vp8_mc(s, 0, dst[1] + 4*y*s->uvlinesize + x*4,
1008 s->framep[mb->ref_frame]->data[1], &uvmv,
1009 4*x + x_off, 4*y + y_off, 4, 4,
1010 width, height, s->uvlinesize,
0ef1dbed 1011 s->put_pixels_tab[2]);
3b636f21
DC
1012 vp8_mc(s, 0, dst[2] + 4*y*s->uvlinesize + x*4,
1013 s->framep[mb->ref_frame]->data[2], &uvmv,
1014 4*x + x_off, 4*y + y_off, 4, 4,
1015 width, height, s->uvlinesize,
0ef1dbed 1016 s->put_pixels_tab[2]);
3b636f21
DC
1017 }
1018 }
1019 }
1020}
1021
1022static void idct_mb(VP8Context *s, uint8_t *y_dst, uint8_t *u_dst, uint8_t *v_dst,
1023 VP8Macroblock *mb)
1024{
1025 int x, y, nnz;
1026
1027 if (mb->mode != MODE_I4x4)
1028 for (y = 0; y < 4; y++) {
1029 for (x = 0; x < 4; x++) {
1030 nnz = s->non_zero_count_cache[y][x];
1031 if (nnz) {
1032 if (nnz == 1)
1033 s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize);
1034 else
1035 s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize);
1036 }
1037 }
1038 y_dst += 4*s->linesize;
1039 }
1040
1041 for (y = 0; y < 2; y++) {
1042 for (x = 0; x < 2; x++) {
1043 nnz = s->non_zero_count_cache[4][(y<<1)+x];
1044 if (nnz) {
1045 if (nnz == 1)
1046 s->vp8dsp.vp8_idct_dc_add(u_dst+4*x, s->block[4][(y<<1)+x], s->uvlinesize);
1047 else
1048 s->vp8dsp.vp8_idct_add(u_dst+4*x, s->block[4][(y<<1)+x], s->uvlinesize);
1049 }
1050
1051 nnz = s->non_zero_count_cache[5][(y<<1)+x];
1052 if (nnz) {
1053 if (nnz == 1)
1054 s->vp8dsp.vp8_idct_dc_add(v_dst+4*x, s->block[5][(y<<1)+x], s->uvlinesize);
1055 else
1056 s->vp8dsp.vp8_idct_add(v_dst+4*x, s->block[5][(y<<1)+x], s->uvlinesize);
1057 }
1058 }
1059 u_dst += 4*s->uvlinesize;
1060 v_dst += 4*s->uvlinesize;
1061 }
1062}
1063
1064static void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, int *level, int *inner, int *hev_thresh)
1065{
1066 int interior_limit, filter_level;
1067
1068 if (s->segmentation.enabled) {
1069 filter_level = s->segmentation.filter_level[mb->segment];
1070 if (!s->segmentation.absolute_vals)
1071 filter_level += s->filter.level;
1072 } else
1073 filter_level = s->filter.level;
1074
1075 if (s->lf_delta.enabled) {
1076 filter_level += s->lf_delta.ref[mb->ref_frame];
1077
1078 if (mb->ref_frame == VP56_FRAME_CURRENT) {
1079 if (mb->mode == MODE_I4x4)
1080 filter_level += s->lf_delta.mode[0];
1081 } else {
1082 if (mb->mode == VP8_MVMODE_ZERO)
1083 filter_level += s->lf_delta.mode[1];
1084 else if (mb->mode == VP8_MVMODE_SPLIT)
1085 filter_level += s->lf_delta.mode[3];
1086 else
1087 filter_level += s->lf_delta.mode[2];
1088 }
1089 }
1090 filter_level = av_clip(filter_level, 0, 63);
1091
1092 interior_limit = filter_level;
1093 if (s->filter.sharpness) {
1094 interior_limit >>= s->filter.sharpness > 4 ? 2 : 1;
1095 interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness);
1096 }
1097 interior_limit = FFMAX(interior_limit, 1);
1098
1099 *level = filter_level;
1100 *inner = interior_limit;
1101
1102 if (hev_thresh) {
1103 *hev_thresh = filter_level >= 15;
1104
1105 if (s->keyframe) {
1106 if (filter_level >= 40)
1107 *hev_thresh = 2;
1108 } else {
1109 if (filter_level >= 40)
1110 *hev_thresh = 3;
1111 else if (filter_level >= 20)
1112 *hev_thresh = 2;
1113 }
1114 }
1115}
1116
1117// TODO: look at backup_mb_border / xchg_mb_border in h264.c
1118static void filter_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb, int mb_x, int mb_y)
1119{
1120 int filter_level, inner_limit, hev_thresh;
1121
1122 filter_level_for_mb(s, mb, &filter_level, &inner_limit, &hev_thresh);
1123 if (!filter_level)
1124 return;
1125
1126 if (mb_x) {
1127 s->vp8dsp.vp8_h_loop_filter16(dst[0], s->linesize, filter_level+2, inner_limit, hev_thresh);
1128 s->vp8dsp.vp8_h_loop_filter8 (dst[1], s->uvlinesize, filter_level+2, inner_limit, hev_thresh);
1129 s->vp8dsp.vp8_h_loop_filter8 (dst[2], s->uvlinesize, filter_level+2, inner_limit, hev_thresh);
1130 }
1131
1132 if (!mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT) {
1133 s->vp8dsp.vp8_h_loop_filter16_inner(dst[0]+ 4, s->linesize, filter_level, inner_limit, hev_thresh);
1134 s->vp8dsp.vp8_h_loop_filter16_inner(dst[0]+ 8, s->linesize, filter_level, inner_limit, hev_thresh);
1135 s->vp8dsp.vp8_h_loop_filter16_inner(dst[0]+12, s->linesize, filter_level, inner_limit, hev_thresh);
1136 s->vp8dsp.vp8_h_loop_filter8_inner (dst[1]+ 4, s->uvlinesize, filter_level, inner_limit, hev_thresh);
1137 s->vp8dsp.vp8_h_loop_filter8_inner (dst[2]+ 4, s->uvlinesize, filter_level, inner_limit, hev_thresh);
1138 }
1139
1140 if (mb_y) {
1141 s->vp8dsp.vp8_v_loop_filter16(dst[0], s->linesize, filter_level+2, inner_limit, hev_thresh);
1142 s->vp8dsp.vp8_v_loop_filter8 (dst[1], s->uvlinesize, filter_level+2, inner_limit, hev_thresh);
1143 s->vp8dsp.vp8_v_loop_filter8 (dst[2], s->uvlinesize, filter_level+2, inner_limit, hev_thresh);
1144 }
1145
1146 if (!mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT) {
1147 s->vp8dsp.vp8_v_loop_filter16_inner(dst[0]+ 4*s->linesize, s->linesize, filter_level, inner_limit, hev_thresh);
1148 s->vp8dsp.vp8_v_loop_filter16_inner(dst[0]+ 8*s->linesize, s->linesize, filter_level, inner_limit, hev_thresh);
1149 s->vp8dsp.vp8_v_loop_filter16_inner(dst[0]+12*s->linesize, s->linesize, filter_level, inner_limit, hev_thresh);
1150 s->vp8dsp.vp8_v_loop_filter8_inner (dst[1]+ 4*s->uvlinesize, s->uvlinesize, filter_level, inner_limit, hev_thresh);
1151 s->vp8dsp.vp8_v_loop_filter8_inner (dst[2]+ 4*s->uvlinesize, s->uvlinesize, filter_level, inner_limit, hev_thresh);
1152 }
1153}
1154
1155static void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8Macroblock *mb, int mb_x, int mb_y)
1156{
1157 int filter_level, inner_limit, mbedge_lim, bedge_lim;
1158
1159 filter_level_for_mb(s, mb, &filter_level, &inner_limit, NULL);
1160 if (!filter_level)
1161 return;
1162
1163 mbedge_lim = 2*(filter_level+2) + inner_limit;
1164 bedge_lim = 2* filter_level + inner_limit;
1165
1166 if (mb_x)
1167 s->vp8dsp.vp8_h_loop_filter_simple(dst, s->linesize, mbedge_lim);
1168 if (!mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT) {
1169 s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, s->linesize, bedge_lim);
1170 s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, s->linesize, bedge_lim);
1171 s->vp8dsp.vp8_h_loop_filter_simple(dst+12, s->linesize, bedge_lim);
1172 }
1173
1174 if (mb_y)
1175 s->vp8dsp.vp8_v_loop_filter_simple(dst, s->linesize, mbedge_lim);
1176 if (!mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT) {
1177 s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*s->linesize, s->linesize, bedge_lim);
1178 s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*s->linesize, s->linesize, bedge_lim);
1179 s->vp8dsp.vp8_v_loop_filter_simple(dst+12*s->linesize, s->linesize, bedge_lim);
1180 }
1181}
1182
1183static void filter_mb_row(VP8Context *s, int mb_y)
1184{
1185 VP8Macroblock *mb = s->macroblocks + mb_y*s->mb_stride;
1186 uint8_t *dst[3] = {
1187 s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize,
1188 s->framep[VP56_FRAME_CURRENT]->data[1] + 8*mb_y*s->uvlinesize,
1189 s->framep[VP56_FRAME_CURRENT]->data[2] + 8*mb_y*s->uvlinesize
1190 };
1191 int mb_x;
1192
1193 for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1194 filter_mb(s, dst, mb++, mb_x, mb_y);
1195 dst[0] += 16;
1196 dst[1] += 8;
1197 dst[2] += 8;
1198 }
1199}
1200
1201static void filter_mb_row_simple(VP8Context *s, int mb_y)
1202{
1203 uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize;
1204 VP8Macroblock *mb = s->macroblocks + mb_y*s->mb_stride;
1205 int mb_x;
1206
1207 for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1208 filter_mb_simple(s, dst, mb++, mb_x, mb_y);
1209 dst += 16;
1210 }
1211}
1212
1213static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
1214 AVPacket *avpkt)
1215{
1216 VP8Context *s = avctx->priv_data;
1217 int ret, mb_x, mb_y, i, y, referenced;
1218 enum AVDiscard skip_thresh;
1219 AVFrame *curframe;
1220
1221 if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0)
1222 return ret;
1223
1224 referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT
1225 || s->update_altref == VP56_FRAME_CURRENT;
1226
1227 skip_thresh = !referenced ? AVDISCARD_NONREF :
1228 !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL;
1229
1230 if (avctx->skip_frame >= skip_thresh) {
1231 s->invisible = 1;
1232 goto skip_decode;
1233 }
1234
1235 for (i = 0; i < 4; i++)
1236 if (&s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1237 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1238 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) {
1239 curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i];
1240 break;
1241 }
1242 if (curframe->data[0])
1243 avctx->release_buffer(avctx, curframe);
1244
1245 curframe->key_frame = s->keyframe;
1246 curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE;
1247 curframe->reference = referenced ? 3 : 0;
1248 if ((ret = avctx->get_buffer(avctx, curframe))) {
1249 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n");
1250 return ret;
1251 }
1252
1253 // Given that arithmetic probabilities are updated every frame, it's quite likely
1254 // that the values we have on a random interframe are complete junk if we didn't
1255 // start decode on a keyframe. So just don't display anything rather than junk.
1256 if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] ||
1257 !s->framep[VP56_FRAME_GOLDEN] ||
1258 !s->framep[VP56_FRAME_GOLDEN2])) {
1259 av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n");
1260 return AVERROR_INVALIDDATA;
1261 }
1262
1263 s->linesize = curframe->linesize[0];
1264 s->uvlinesize = curframe->linesize[1];
1265
1266 if (!s->edge_emu_buffer)
1267 s->edge_emu_buffer = av_malloc(21*s->linesize);
1268
1269 memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz));
1270
1271 // top edge of 127 for intra prediction
1272 if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) {
1273 memset(curframe->data[0] - s->linesize -1, 127, s->linesize +1);
1274 memset(curframe->data[1] - s->uvlinesize-1, 127, s->uvlinesize+1);
1275 memset(curframe->data[2] - s->uvlinesize-1, 127, s->uvlinesize+1);
1276 }
1277
1278 for (mb_y = 0; mb_y < s->mb_height; mb_y++) {
1279 VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)];
1280 VP8Macroblock *mb = s->macroblocks + mb_y*s->mb_stride;
1281 uint8_t *intra4x4 = s->intra4x4_pred_mode + 4*mb_y*s->b4_stride;
1282 uint8_t *dst[3] = {
1283 curframe->data[0] + 16*mb_y*s->linesize,
1284 curframe->data[1] + 8*mb_y*s->uvlinesize,
1285 curframe->data[2] + 8*mb_y*s->uvlinesize
1286 };
1287
1288 memset(s->left_nnz, 0, sizeof(s->left_nnz));
1289
1290 // left edge of 129 for intra prediction
1291 if (!(avctx->flags & CODEC_FLAG_EMU_EDGE))
1292 for (i = 0; i < 3; i++)
1293 for (y = 0; y < 16>>!!i; y++)
1294 dst[i][y*curframe->linesize[i]-1] = 129;
1295
1296 for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1297 decode_mb_mode(s, mb, mb_x, mb_y, intra4x4 + 4*mb_x);
1298
1299 if (!mb->skip)
1300 decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);
1301 else {
1302 AV_ZERO128(s->non_zero_count_cache); // luma
1303 AV_ZERO64(s->non_zero_count_cache[4]); // chroma
1304 }
1305
1306 if (mb->mode <= MODE_I4x4) {
1307 intra_predict(s, dst, mb, intra4x4 + 4*mb_x, mb_x, mb_y);
1308 memset(mb->bmv, 0, sizeof(mb->bmv));
1309 } else {
1310 inter_predict(s, dst, mb, mb_x, mb_y);
1311 }
1312
1313 if (!mb->skip) {
1314 idct_mb(s, dst[0], dst[1], dst[2], mb);
1315 } else {
1316 AV_ZERO64(s->left_nnz);
1317 AV_WN64(s->top_nnz[mb_x], 0); // array of 9, so unaligned
1318
1319 // Reset DC block predictors if they would exist if the mb had coefficients
1320 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
1321 s->left_nnz[8] = 0;
1322 s->top_nnz[mb_x][8] = 0;
1323 }
1324 }
1325
1326 dst[0] += 16;
1327 dst[1] += 8;
1328 dst[2] += 8;
1329 mb++;
1330 }
1331 if (mb_y && s->filter.level && avctx->skip_loop_filter < skip_thresh) {
1332 if (s->filter.simple)
1333 filter_mb_row_simple(s, mb_y-1);
1334 else
1335 filter_mb_row(s, mb_y-1);
1336 }
1337 }
1338 if (s->filter.level && avctx->skip_loop_filter < skip_thresh) {
1339 if (s->filter.simple)
1340 filter_mb_row_simple(s, mb_y-1);
1341 else
1342 filter_mb_row(s, mb_y-1);
1343 }
1344
1345skip_decode:
1346 // if future frames don't use the updated probabilities,
1347 // reset them to the values we saved
1348 if (!s->update_probabilities)
1349 s->prob[0] = s->prob[1];
1350
1351 // check if golden and altref are swapped
1352 if (s->update_altref == VP56_FRAME_GOLDEN &&
1353 s->update_golden == VP56_FRAME_GOLDEN2)
1354 FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN], s->framep[VP56_FRAME_GOLDEN2]);
1355 else {
1356 if (s->update_altref != VP56_FRAME_NONE)
1357 s->framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref];
1358
1359 if (s->update_golden != VP56_FRAME_NONE)
1360 s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden];
1361 }
1362
1363 if (s->update_last) // move cur->prev
1364 s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT];
1365
1366 // release no longer referenced frames
1367 for (i = 0; i < 4; i++)
1368 if (s->frames[i].data[0] &&
1369 &s->frames[i] != s->framep[VP56_FRAME_CURRENT] &&
1370 &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1371 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1372 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2])
1373 avctx->release_buffer(avctx, &s->frames[i]);
1374
1375 if (!s->invisible) {
1376 *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT];
1377 *data_size = sizeof(AVFrame);
1378 }
1379
1380 return avpkt->size;
1381}
1382
1383static av_cold int vp8_decode_init(AVCodecContext *avctx)
1384{
1385 VP8Context *s = avctx->priv_data;
1386
1387 s->avctx = avctx;
1388 avctx->pix_fmt = PIX_FMT_YUV420P;
1389
1390 dsputil_init(&s->dsp, avctx);
1391 ff_h264_pred_init(&s->hpc, CODEC_ID_VP8);
1392 ff_vp8dsp_init(&s->vp8dsp);
1393
1394 // intra pred needs edge emulation among other things
1395 if (avctx->flags&CODEC_FLAG_EMU_EDGE) {
03ac56e7 1396 av_log(avctx, AV_LOG_ERROR, "Edge emulation not supported\n");
3b636f21
DC
1397 return AVERROR_PATCHWELCOME;
1398 }
1399
1400 return 0;
1401}
1402
1403static av_cold int vp8_decode_free(AVCodecContext *avctx)
1404{
1405 vp8_decode_flush(avctx);
1406 return 0;
1407}
1408
1409AVCodec vp8_decoder = {
1410 "vp8",
1411 AVMEDIA_TYPE_VIDEO,
1412 CODEC_ID_VP8,
1413 sizeof(VP8Context),
1414 vp8_decode_init,
1415 NULL,
1416 vp8_decode_free,
1417 vp8_decode_frame,
1418 CODEC_CAP_DR1,
1419 .flush = vp8_decode_flush,
1420 .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"),
1421};