vp8: Check for malloc failure
[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
b6c420ce
DC
228 if (!s->macroblocks_base || !s->intra4x4_pred_mode_base || !s->top_nnz)
229 return AVERROR(ENOMEM);
230
3b636f21
DC
231 s->macroblocks = s->macroblocks_base + 1 + s->mb_stride;
232 s->intra4x4_pred_mode = s->intra4x4_pred_mode_base + 4 + s->b4_stride;
233
234 memset(s->intra4x4_pred_mode_base, DC_PRED, s->b4_stride);
235 for (i = 0; i < 4*s->mb_height; i++)
236 s->intra4x4_pred_mode[i*s->b4_stride-1] = DC_PRED;
237
238 return 0;
239}
240
241static void parse_segment_info(VP8Context *s)
242{
243 VP56RangeCoder *c = &s->c;
244 int i;
245
246 s->segmentation.update_map = vp8_rac_get(c);
247
248 if (vp8_rac_get(c)) { // update segment feature data
249 s->segmentation.absolute_vals = vp8_rac_get(c);
250
251 for (i = 0; i < 4; i++)
252 s->segmentation.base_quant[i] = vp8_rac_get_sint(c, 7);
253
254 for (i = 0; i < 4; i++)
255 s->segmentation.filter_level[i] = vp8_rac_get_sint(c, 6);
256 }
257 if (s->segmentation.update_map)
258 for (i = 0; i < 3; i++)
259 s->prob->segmentid[i] = vp8_rac_get(c) ? vp8_rac_get_uint(c, 8) : 255;
260}
261
262static void update_lf_deltas(VP8Context *s)
263{
264 VP56RangeCoder *c = &s->c;
265 int i;
266
267 for (i = 0; i < 4; i++)
268 s->lf_delta.ref[i] = vp8_rac_get_sint(c, 6);
269
270 for (i = 0; i < 4; i++)
271 s->lf_delta.mode[i] = vp8_rac_get_sint(c, 6);
272}
273
274static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size)
275{
276 const uint8_t *sizes = buf;
277 int i;
278
279 s->num_coeff_partitions = 1 << vp8_rac_get_uint(&s->c, 2);
280
281 buf += 3*(s->num_coeff_partitions-1);
282 buf_size -= 3*(s->num_coeff_partitions-1);
283 if (buf_size < 0)
284 return -1;
285
286 for (i = 0; i < s->num_coeff_partitions-1; i++) {
287 int size = RL24(sizes + 3*i);
288 if (buf_size - size < 0)
289 return -1;
290
291 vp56_init_range_decoder(&s->coeff_partition[i], buf, size);
292 buf += size;
293 buf_size -= size;
294 }
295 vp56_init_range_decoder(&s->coeff_partition[i], buf, buf_size);
296
297 return 0;
298}
299
300static void get_quants(VP8Context *s)
301{
302 VP56RangeCoder *c = &s->c;
303 int i, base_qi;
304
305 int yac_qi = vp8_rac_get_uint(c, 7);
306 int ydc_delta = vp8_rac_get_sint(c, 4);
307 int y2dc_delta = vp8_rac_get_sint(c, 4);
308 int y2ac_delta = vp8_rac_get_sint(c, 4);
309 int uvdc_delta = vp8_rac_get_sint(c, 4);
310 int uvac_delta = vp8_rac_get_sint(c, 4);
311
312 for (i = 0; i < 4; i++) {
313 if (s->segmentation.enabled) {
314 base_qi = s->segmentation.base_quant[i];
315 if (!s->segmentation.absolute_vals)
316 base_qi += yac_qi;
317 } else
318 base_qi = yac_qi;
319
320 s->qmat[i].luma_qmul[0] = vp8_dc_qlookup[av_clip(base_qi + ydc_delta , 0, 127)];
321 s->qmat[i].luma_qmul[1] = vp8_ac_qlookup[av_clip(base_qi , 0, 127)];
322 s->qmat[i].luma_dc_qmul[0] = 2 * vp8_dc_qlookup[av_clip(base_qi + y2dc_delta, 0, 127)];
323 s->qmat[i].luma_dc_qmul[1] = 155 * vp8_ac_qlookup[av_clip(base_qi + y2ac_delta, 0, 127)] / 100;
324 s->qmat[i].chroma_qmul[0] = vp8_dc_qlookup[av_clip(base_qi + uvdc_delta, 0, 127)];
325 s->qmat[i].chroma_qmul[1] = vp8_ac_qlookup[av_clip(base_qi + uvac_delta, 0, 127)];
326
327 s->qmat[i].luma_dc_qmul[1] = FFMAX(s->qmat[i].luma_dc_qmul[1], 8);
328 s->qmat[i].chroma_qmul[0] = FFMIN(s->qmat[i].chroma_qmul[0], 132);
329 }
330}
331
332/**
333 * Determine which buffers golden and altref should be updated with after this frame.
334 * The spec isn't clear here, so I'm going by my understanding of what libvpx does
335 *
336 * Intra frames update all 3 references
337 * Inter frames update VP56_FRAME_PREVIOUS if the update_last flag is set
338 * If the update (golden|altref) flag is set, it's updated with the current frame
339 * if update_last is set, and VP56_FRAME_PREVIOUS otherwise.
340 * If the flag is not set, the number read means:
341 * 0: no update
342 * 1: VP56_FRAME_PREVIOUS
343 * 2: update golden with altref, or update altref with golden
344 */
345static VP56Frame ref_to_update(VP8Context *s, int update, VP56Frame ref)
346{
347 VP56RangeCoder *c = &s->c;
348
349 if (update)
350 return VP56_FRAME_CURRENT;
351
352 switch (vp8_rac_get_uint(c, 2)) {
353 case 1:
354 return VP56_FRAME_PREVIOUS;
355 case 2:
356 return (ref == VP56_FRAME_GOLDEN) ? VP56_FRAME_GOLDEN2 : VP56_FRAME_GOLDEN;
357 }
358 return VP56_FRAME_NONE;
359}
360
361static void update_refs(VP8Context *s)
362{
363 VP56RangeCoder *c = &s->c;
364
365 int update_golden = vp8_rac_get(c);
366 int update_altref = vp8_rac_get(c);
367
368 s->update_golden = ref_to_update(s, update_golden, VP56_FRAME_GOLDEN);
369 s->update_altref = ref_to_update(s, update_altref, VP56_FRAME_GOLDEN2);
370}
371
372static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)
373{
374 VP56RangeCoder *c = &s->c;
375 int header_size, hscale, vscale, i, j, k, l, ret;
376 int width = s->avctx->width;
377 int height = s->avctx->height;
378
379 s->keyframe = !(buf[0] & 1);
380 s->profile = (buf[0]>>1) & 7;
381 s->invisible = !(buf[0] & 0x10);
382 header_size = RL24(buf) >> 5;
383 buf += 3;
384 buf_size -= 3;
385
0ef1dbed
DC
386 if (s->profile > 3)
387 av_log(s->avctx, AV_LOG_WARNING, "Unknown profile %d\n", s->profile);
388
389 if (!s->profile)
390 memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_epel_pixels_tab, sizeof(s->put_pixels_tab));
391 else // profile 1-3 use bilinear, 4+ aren't defined so whatever
392 memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_bilinear_pixels_tab, sizeof(s->put_pixels_tab));
3b636f21
DC
393
394 if (header_size > buf_size - 7*s->keyframe) {
395 av_log(s->avctx, AV_LOG_ERROR, "Header size larger than data provided\n");
396 return AVERROR_INVALIDDATA;
397 }
398
399 if (s->keyframe) {
400 if (RL24(buf) != 0x2a019d) {
401 av_log(s->avctx, AV_LOG_ERROR, "Invalid start code 0x%x\n", RL24(buf));
402 return AVERROR_INVALIDDATA;
403 }
404 width = AV_RL16(buf+3) & 0x3fff;
405 height = AV_RL16(buf+5) & 0x3fff;
406 hscale = buf[4] >> 6;
407 vscale = buf[6] >> 6;
408 buf += 7;
409 buf_size -= 7;
410
92a54426
MR
411 if (hscale || vscale)
412 av_log_missing_feature(s->avctx, "Upscaling", 1);
413
3b636f21
DC
414 s->update_golden = s->update_altref = VP56_FRAME_CURRENT;
415 memcpy(s->prob->token , vp8_token_default_probs , sizeof(s->prob->token));
416 memcpy(s->prob->pred16x16, vp8_pred16x16_prob_inter, sizeof(s->prob->pred16x16));
417 memcpy(s->prob->pred8x8c , vp8_pred8x8c_prob_inter , sizeof(s->prob->pred8x8c));
418 memcpy(s->prob->mvc , vp8_mv_default_prob , sizeof(s->prob->mvc));
419 memset(&s->segmentation, 0, sizeof(s->segmentation));
420 }
421
422 if (!s->macroblocks_base || /* first frame */
423 width != s->avctx->width || height != s->avctx->height) {
424 if ((ret = update_dimensions(s, width, height) < 0))
425 return ret;
426 }
427
428 vp56_init_range_decoder(c, buf, header_size);
429 buf += header_size;
430 buf_size -= header_size;
431
432 if (s->keyframe) {
433 if (vp8_rac_get(c))
434 av_log(s->avctx, AV_LOG_WARNING, "Unspecified colorspace\n");
435 vp8_rac_get(c); // whether we can skip clamping in dsp functions
436 }
437
438 if ((s->segmentation.enabled = vp8_rac_get(c)))
439 parse_segment_info(s);
440 else
441 s->segmentation.update_map = 0; // FIXME: move this to some init function?
442
443 s->filter.simple = vp8_rac_get(c);
444 s->filter.level = vp8_rac_get_uint(c, 6);
445 s->filter.sharpness = vp8_rac_get_uint(c, 3);
446
447 if ((s->lf_delta.enabled = vp8_rac_get(c)))
448 if (vp8_rac_get(c))
449 update_lf_deltas(s);
450
451 if (setup_partitions(s, buf, buf_size)) {
452 av_log(s->avctx, AV_LOG_ERROR, "Invalid partitions\n");
453 return AVERROR_INVALIDDATA;
454 }
455
456 get_quants(s);
457
458 if (!s->keyframe) {
459 update_refs(s);
460 s->sign_bias[VP56_FRAME_GOLDEN] = vp8_rac_get(c);
461 s->sign_bias[VP56_FRAME_GOLDEN2 /* altref */] = vp8_rac_get(c);
462 }
463
464 // if we aren't saving this frame's probabilities for future frames,
465 // make a copy of the current probabilities
466 if (!(s->update_probabilities = vp8_rac_get(c)))
467 s->prob[1] = s->prob[0];
468
469 s->update_last = s->keyframe || vp8_rac_get(c);
470
471 for (i = 0; i < 4; i++)
472 for (j = 0; j < 8; j++)
473 for (k = 0; k < 3; k++)
474 for (l = 0; l < NUM_DCT_TOKENS-1; l++)
475 if (vp56_rac_get_prob(c, vp8_token_update_probs[i][j][k][l]))
476 s->prob->token[i][j][k][l] = vp8_rac_get_uint(c, 8);
477
478 if ((s->mbskip_enabled = vp8_rac_get(c)))
479 s->prob->mbskip = vp8_rac_get_uint(c, 8);
480
481 if (!s->keyframe) {
482 s->prob->intra = vp8_rac_get_uint(c, 8);
483 s->prob->last = vp8_rac_get_uint(c, 8);
484 s->prob->golden = vp8_rac_get_uint(c, 8);
485
486 if (vp8_rac_get(c))
487 for (i = 0; i < 4; i++)
488 s->prob->pred16x16[i] = vp8_rac_get_uint(c, 8);
489 if (vp8_rac_get(c))
490 for (i = 0; i < 3; i++)
491 s->prob->pred8x8c[i] = vp8_rac_get_uint(c, 8);
492
493 // 17.2 MV probability update
494 for (i = 0; i < 2; i++)
495 for (j = 0; j < 19; j++)
496 if (vp56_rac_get_prob(c, vp8_mv_update_prob[i][j]))
497 s->prob->mvc[i][j] = vp8_rac_get_nn(c);
498 }
499
500 return 0;
501}
502
503static inline void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src,
504 int mb_x, int mb_y)
505{
506#define MARGIN (16 << 2)
507 dst->x = av_clip(src->x, -((mb_x << 6) + MARGIN),
508 ((s->mb_width - 1 - mb_x) << 6) + MARGIN);
509 dst->y = av_clip(src->y, -((mb_y << 6) + MARGIN),
510 ((s->mb_height - 1 - mb_y) << 6) + MARGIN);
511}
512
513static void find_near_mvs(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
514 VP56mv near[2], VP56mv *best, int cnt[4])
515{
516 VP8Macroblock *mb_edge[3] = { mb - s->mb_stride /* top */,
517 mb - 1 /* left */,
518 mb - s->mb_stride - 1 /* top-left */ };
519 enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT };
520 VP56mv near_mv[4] = {{ 0 }};
521 enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV };
522 int idx = CNT_ZERO, n;
523 int best_idx = CNT_ZERO;
524
525 /* Process MB on top, left and top-left */
526 for (n = 0; n < 3; n++) {
527 VP8Macroblock *edge = mb_edge[n];
528 if (edge->ref_frame != VP56_FRAME_CURRENT) {
529 if (edge->mv.x | edge->mv.y) {
530 VP56mv tmp = edge->mv;
531 if (s->sign_bias[mb->ref_frame] != s->sign_bias[edge->ref_frame]) {
532 tmp.x *= -1;
533 tmp.y *= -1;
534 }
535 if ((tmp.x ^ near_mv[idx].x) | (tmp.y ^ near_mv[idx].y))
536 near_mv[++idx] = tmp;
537 cnt[idx] += 1 + (n != 2);
538 } else
539 cnt[CNT_ZERO] += 1 + (n != 2);
540 }
541 }
542
543 /* If we have three distinct MV's, merge first and last if they're the same */
544 if (cnt[CNT_SPLITMV] &&
545 !((near_mv[1+EDGE_TOP].x ^ near_mv[1+EDGE_TOPLEFT].x) |
546 (near_mv[1+EDGE_TOP].y ^ near_mv[1+EDGE_TOPLEFT].y)))
547 cnt[CNT_NEAREST] += 1;
548
549 cnt[CNT_SPLITMV] = ((mb_edge[EDGE_LEFT]->mode == VP8_MVMODE_SPLIT) +
550 (mb_edge[EDGE_TOP]->mode == VP8_MVMODE_SPLIT)) * 2 +
551 (mb_edge[EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT);
552
553 /* Swap near and nearest if necessary */
554 if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) {
555 FFSWAP(int, cnt[CNT_NEAREST], cnt[CNT_NEAR]);
556 FFSWAP(VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]);
557 }
558
559 /* Choose the best mv out of 0,0 and the nearest mv */
560 if (cnt[CNT_NEAREST] >= cnt[CNT_ZERO])
561 best_idx = CNT_NEAREST;
562
563 clamp_mv(s, best, &near_mv[best_idx], mb_x, mb_y);
564 near[0] = near_mv[CNT_NEAREST];
565 near[1] = near_mv[CNT_NEAR];
566}
567
568/**
569 * Motion vector coding, 17.1.
570 */
571static int read_mv_component(VP56RangeCoder *c, const uint8_t *p)
572{
573 int x = 0;
574
575 if (vp56_rac_get_prob(c, p[0])) {
576 int i;
577
578 for (i = 0; i < 3; i++)
579 x += vp56_rac_get_prob(c, p[9 + i]) << i;
580 for (i = 9; i > 3; i--)
581 x += vp56_rac_get_prob(c, p[9 + i]) << i;
582 if (!(x & 0xFFF0) || vp56_rac_get_prob(c, p[12]))
583 x += 8;
584 } else
585 x = vp8_rac_get_tree(c, vp8_small_mvtree, &p[2]);
586
587 return (x && vp56_rac_get_prob(c, p[1])) ? -x : x;
588}
589
590static const uint8_t *get_submv_prob(const VP56mv *left, const VP56mv *top)
591{
592 int l_is_zero = !(left->x | left->y);
593 int t_is_zero = !(top->x | top->y);
594 int equal = !((left->x ^ top->x) | (left->y ^ top->y));
595
596 if (equal)
597 return l_is_zero ? vp8_submv_prob[4] : vp8_submv_prob[3];
598 if (t_is_zero)
599 return vp8_submv_prob[2];
600 return l_is_zero ? vp8_submv_prob[1] : vp8_submv_prob[0];
601}
602
603/**
604 * Split motion vector prediction, 16.4.
7ed06b2b 605 * @returns the number of motion vectors parsed (2, 4 or 16)
3b636f21 606 */
7ed06b2b 607static int decode_splitmvs(VP8Context *s, VP56RangeCoder *c,
3b636f21
DC
608 VP8Macroblock *mb, VP56mv *base_mv)
609{
610 int part_idx = mb->partitioning =
611 vp8_rac_get_tree(c, vp8_mbsplit_tree, vp8_mbsplit_prob);
612 int n, num = vp8_mbsplit_count[part_idx];
7ed06b2b
RB
613 const uint8_t *mbsplits = vp8_mbsplits[part_idx],
614 *firstidx = vp8_mbfirstidx[part_idx];
3b636f21
DC
615
616 for (n = 0; n < num; n++) {
7ed06b2b
RB
617 int k = firstidx[n];
618 const VP56mv *left, *above;
619 const uint8_t *submv_prob;
620
621 if (!(k & 3)) {
622 VP8Macroblock *left_mb = &mb[-1];
623 left = &left_mb->bmv[vp8_mbsplits[left_mb->partitioning][k + 3]];
624 } else
625 left = &mb->bmv[mbsplits[k - 1]];
626 if (k <= 3) {
627 VP8Macroblock *above_mb = &mb[-s->mb_stride];
628 above = &above_mb->bmv[vp8_mbsplits[above_mb->partitioning][k + 12]];
629 } else
630 above = &mb->bmv[mbsplits[k - 4]];
631
632 submv_prob = get_submv_prob(left, above);
3b636f21
DC
633
634 switch (vp8_rac_get_tree(c, vp8_submv_ref_tree, submv_prob)) {
635 case VP8_SUBMVMODE_NEW4X4:
7ed06b2b
RB
636 mb->bmv[n].y = base_mv->y + read_mv_component(c, s->prob->mvc[0]);
637 mb->bmv[n].x = base_mv->x + read_mv_component(c, s->prob->mvc[1]);
3b636f21
DC
638 break;
639 case VP8_SUBMVMODE_ZERO4X4:
7ed06b2b
RB
640 mb->bmv[n].x = 0;
641 mb->bmv[n].y = 0;
3b636f21
DC
642 break;
643 case VP8_SUBMVMODE_LEFT4X4:
7ed06b2b 644 mb->bmv[n] = *left;
3b636f21
DC
645 break;
646 case VP8_SUBMVMODE_TOP4X4:
7ed06b2b 647 mb->bmv[n] = *above;
3b636f21
DC
648 break;
649 }
3b636f21 650 }
7ed06b2b
RB
651
652 return num;
3b636f21
DC
653}
654
655static inline void decode_intra4x4_modes(VP56RangeCoder *c, uint8_t *intra4x4,
656 int stride, int keyframe)
657{
658 int x, y, t, l;
659 const uint8_t *ctx = vp8_pred4x4_prob_inter;
660
661 for (y = 0; y < 4; y++) {
662 for (x = 0; x < 4; x++) {
663 if (keyframe) {
664 t = intra4x4[x - stride];
665 l = intra4x4[x - 1];
666 ctx = vp8_pred4x4_prob_intra[t][l];
667 }
668 intra4x4[x] = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx);
669 }
670 intra4x4 += stride;
671 }
672}
673
674static void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
675 uint8_t *intra4x4)
676{
677 VP56RangeCoder *c = &s->c;
678 int n;
679
680 if (s->segmentation.update_map)
681 mb->segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid);
682
683 mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0;
684
685 if (s->keyframe) {
686 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra);
687
688 if (mb->mode == MODE_I4x4) {
689 decode_intra4x4_modes(c, intra4x4, s->b4_stride, 1);
690 } else
691 fill_rectangle(intra4x4, 4, 4, s->b4_stride, vp8_pred4x4_mode[mb->mode], 1);
692
693 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra);
694 mb->ref_frame = VP56_FRAME_CURRENT;
695 } else if (vp56_rac_get_prob(c, s->prob->intra)) {
696 VP56mv near[2], best;
697 int cnt[4] = { 0 };
698 uint8_t p[4];
699
700 // inter MB, 16.2
701 if (vp56_rac_get_prob(c, s->prob->last))
702 mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ?
703 VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN;
704 else
705 mb->ref_frame = VP56_FRAME_PREVIOUS;
706
707 // motion vectors, 16.3
708 find_near_mvs(s, mb, mb_x, mb_y, near, &best, cnt);
709 for (n = 0; n < 4; n++)
710 p[n] = vp8_mode_contexts[cnt[n]][n];
711 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_mvinter, p);
712 switch (mb->mode) {
713 case VP8_MVMODE_SPLIT:
7ed06b2b 714 mb->mv = mb->bmv[decode_splitmvs(s, c, mb, &best) - 1];
3b636f21
DC
715 break;
716 case VP8_MVMODE_ZERO:
717 mb->mv.x = 0;
718 mb->mv.y = 0;
719 break;
720 case VP8_MVMODE_NEAREST:
721 clamp_mv(s, &mb->mv, &near[0], mb_x, mb_y);
722 break;
723 case VP8_MVMODE_NEAR:
724 clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y);
725 break;
726 case VP8_MVMODE_NEW:
727 mb->mv.y = best.y + read_mv_component(c, s->prob->mvc[0]);
728 mb->mv.x = best.x + read_mv_component(c, s->prob->mvc[1]);
729 break;
730 }
731 if (mb->mode != VP8_MVMODE_SPLIT) {
7ed06b2b
RB
732 mb->partitioning = VP8_SPLITMVMODE_NONE;
733 mb->bmv[0] = mb->mv;
3b636f21
DC
734 }
735 } else {
736 // intra MB, 16.1
737 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16);
738
739 if (mb->mode == MODE_I4x4) {
740 decode_intra4x4_modes(c, intra4x4, s->b4_stride, 0);
741 } else
742 fill_rectangle(intra4x4, 4, 4, s->b4_stride, vp8_pred4x4_mode[mb->mode], 1);
743
744 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c);
745 mb->ref_frame = VP56_FRAME_CURRENT;
746 }
747}
748
749/**
e394953e
RB
750 * @param c arithmetic bitstream reader context
751 * @param block destination for block coefficients
752 * @param probs probabilities to use when reading trees from the bitstream
3b636f21
DC
753 * @param i initial coeff index, 0 unless a separate DC block is coded
754 * @param zero_nhood the initial prediction context for number of surrounding
755 * all-zero blocks (only left/top, so 0-2)
3fa76268 756 * @param qmul array holding the dc/ac dequant factor at position 0/1
3b636f21
DC
757 * @return 0 if no coeffs were decoded
758 * otherwise, the index of the last coeff decoded plus one
759 */
760static int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16],
761 uint8_t probs[8][3][NUM_DCT_TOKENS-1],
762 int i, int zero_nhood, int16_t qmul[2])
763{
764 int token, nonzero = 0;
765 int offset = 0;
766
767 for (; i < 16; i++) {
768 token = vp8_rac_get_tree_with_offset(c, vp8_coeff_tree, probs[vp8_coeff_band[i]][zero_nhood], offset);
769
770 if (token == DCT_EOB)
771 break;
772 else if (token >= DCT_CAT1) {
773 int cat = token-DCT_CAT1;
774 token = vp8_rac_get_coeff(c, vp8_dct_cat_prob[cat]);
775 token += vp8_dct_cat_offset[cat];
776 }
777
778 // after the first token, the non-zero prediction context becomes
779 // based on the last decoded coeff
780 if (!token) {
781 zero_nhood = 0;
782 offset = 1;
783 continue;
784 } else if (token == 1)
785 zero_nhood = 1;
786 else
787 zero_nhood = 2;
788
789 // todo: full [16] qmat? load into register?
790 block[zigzag_scan[i]] = (vp8_rac_get(c) ? -token : token) * qmul[!!i];
791 nonzero = i+1;
792 offset = 0;
793 }
794 return nonzero;
795}
796
797static void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,
798 uint8_t t_nnz[9], uint8_t l_nnz[9])
799{
800 LOCAL_ALIGNED_16(DCTELEM, dc,[16]);
801 int i, x, y, luma_start = 0, luma_ctx = 3;
802 int nnz_pred, nnz, nnz_total = 0;
803 int segment = s->segmentation.enabled ? mb->segment : 0;
804
805 s->dsp.clear_blocks((DCTELEM *)s->block);
806
807 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
808 AV_ZERO128(dc);
809 AV_ZERO128(dc+8);
810 nnz_pred = t_nnz[8] + l_nnz[8];
811
812 // decode DC values and do hadamard
813 nnz = decode_block_coeffs(c, dc, s->prob->token[1], 0, nnz_pred,
814 s->qmat[segment].luma_dc_qmul);
815 l_nnz[8] = t_nnz[8] = !!nnz;
816 nnz_total += nnz;
817 s->vp8dsp.vp8_luma_dc_wht(s->block, dc);
818 luma_start = 1;
819 luma_ctx = 0;
820 }
821
822 // luma blocks
823 for (y = 0; y < 4; y++)
824 for (x = 0; x < 4; x++) {
825 nnz_pred = l_nnz[y] + t_nnz[x];
826 nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start,
827 nnz_pred, s->qmat[segment].luma_qmul);
828 // nnz+luma_start may be one more than the actual last index, but we don't care
829 s->non_zero_count_cache[y][x] = nnz + luma_start;
830 t_nnz[x] = l_nnz[y] = !!nnz;
831 nnz_total += nnz;
832 }
833
834 // chroma blocks
835 // TODO: what to do about dimensions? 2nd dim for luma is x,
836 // but for chroma it's (y<<1)|x
837 for (i = 4; i < 6; i++)
838 for (y = 0; y < 2; y++)
839 for (x = 0; x < 2; x++) {
840 nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x];
841 nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0,
842 nnz_pred, s->qmat[segment].chroma_qmul);
843 s->non_zero_count_cache[i][(y<<1)+x] = nnz;
844 t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz;
845 nnz_total += nnz;
846 }
847
848 // if there were no coded coeffs despite the macroblock not being marked skip,
849 // we MUST not do the inner loop filter and should not do IDCT
850 // Since skip isn't used for bitstream prediction, just manually set it.
851 if (!nnz_total)
852 mb->skip = 1;
853}
854
855static int check_intra_pred_mode(int mode, int mb_x, int mb_y)
856{
857 if (mode == DC_PRED8x8) {
858 if (!(mb_x|mb_y))
859 mode = DC_128_PRED8x8;
860 else if (!mb_y)
861 mode = LEFT_DC_PRED8x8;
862 else if (!mb_x)
863 mode = TOP_DC_PRED8x8;
864 }
865 return mode;
866}
867
868static void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
869 uint8_t *bmode, int mb_x, int mb_y)
870{
871 int x, y, mode, nnz, tr;
872
873 if (mb->mode < MODE_I4x4) {
874 mode = check_intra_pred_mode(mb->mode, mb_x, mb_y);
875 s->hpc.pred16x16[mode](dst[0], s->linesize);
876 } else {
877 uint8_t *ptr = dst[0];
878
879 // all blocks on the right edge of the macroblock use bottom edge
880 // the top macroblock for their topright edge
881 uint8_t *tr_right = ptr - s->linesize + 16;
882
883 // if we're on the right edge of the frame, said edge is extended
884 // from the top macroblock
885 if (mb_x == s->mb_width-1) {
886 tr = tr_right[-1]*0x01010101;
887 tr_right = (uint8_t *)&tr;
888 }
889
890 for (y = 0; y < 4; y++) {
891 uint8_t *topright = ptr + 4 - s->linesize;
892 for (x = 0; x < 4; x++) {
893 if (x == 3)
894 topright = tr_right;
895
896 s->hpc.pred4x4[bmode[x]](ptr+4*x, topright, s->linesize);
897
898 nnz = s->non_zero_count_cache[y][x];
899 if (nnz) {
900 if (nnz == 1)
901 s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize);
902 else
903 s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize);
904 }
905 topright += 4;
906 }
907
908 ptr += 4*s->linesize;
909 bmode += s->b4_stride;
910 }
911 }
912
913 mode = check_intra_pred_mode(s->chroma_pred_mode, mb_x, mb_y);
914 s->hpc.pred8x8[mode](dst[1], s->uvlinesize);
915 s->hpc.pred8x8[mode](dst[2], s->uvlinesize);
916}
917
918/**
919 * Generic MC function.
920 *
921 * @param s VP8 decoding context
922 * @param luma 1 for luma (Y) planes, 0 for chroma (Cb/Cr) planes
923 * @param dst target buffer for block data at block position
924 * @param src reference picture buffer at origin (0, 0)
925 * @param mv motion vector (relative to block position) to get pixel data from
926 * @param x_off horizontal position of block from origin (0, 0)
927 * @param y_off vertical position of block from origin (0, 0)
928 * @param block_w width of block (16, 8 or 4)
929 * @param block_h height of block (always same as block_w)
930 * @param width width of src/dst plane data
931 * @param height height of src/dst plane data
932 * @param linesize size of a single line of plane data, including padding
e394953e 933 * @param mc_func motion compensation function pointers (bilinear or sixtap MC)
3b636f21
DC
934 */
935static inline void vp8_mc(VP8Context *s, int luma,
936 uint8_t *dst, uint8_t *src, const VP56mv *mv,
937 int x_off, int y_off, int block_w, int block_h,
938 int width, int height, int linesize,
d6f8476b 939 vp8_mc_func mc_func[3][3])
3b636f21
DC
940{
941 static const uint8_t idx[8] = { 0, 1, 2, 1, 2, 1, 2, 1 };
942 int mx = (mv->x << luma)&7, mx_idx = idx[mx];
943 int my = (mv->y << luma)&7, my_idx = idx[my];
944
945 x_off += mv->x >> (3 - luma);
946 y_off += mv->y >> (3 - luma);
947
948 // edge emulation
949 src += y_off * linesize + x_off;
950 if (x_off < 2 || x_off >= width - block_w - 3 ||
951 y_off < 2 || y_off >= height - block_h - 3) {
952 ff_emulated_edge_mc(s->edge_emu_buffer, src - 2 * linesize - 2, linesize,
953 block_w + 5, block_h + 5,
954 x_off - 2, y_off - 2, width, height);
955 src = s->edge_emu_buffer + 2 + linesize * 2;
956 }
957
d6f8476b 958 mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);
3b636f21
DC
959}
960
7c4dcf81
RB
961static inline void vp8_mc_part(VP8Context *s, uint8_t *dst[3],
962 AVFrame *ref_frame, int x_off, int y_off,
963 int bx_off, int by_off,
964 int block_w, int block_h,
965 int width, int height, VP56mv *mv)
966{
967 VP56mv uvmv = *mv;
968
969 /* Y */
970 vp8_mc(s, 1, dst[0] + by_off * s->linesize + bx_off,
971 ref_frame->data[0], mv, x_off + bx_off, y_off + by_off,
972 block_w, block_h, width, height, s->linesize,
973 s->put_pixels_tab[block_w == 8]);
974
975 /* U/V */
976 if (s->profile == 3) {
977 uvmv.x &= ~7;
978 uvmv.y &= ~7;
979 }
980 x_off >>= 1; y_off >>= 1;
981 bx_off >>= 1; by_off >>= 1;
982 width >>= 1; height >>= 1;
983 block_w >>= 1; block_h >>= 1;
984 vp8_mc(s, 0, dst[1] + by_off * s->uvlinesize + bx_off,
985 ref_frame->data[1], &uvmv, x_off + bx_off, y_off + by_off,
986 block_w, block_h, width, height, s->uvlinesize,
987 s->put_pixels_tab[1 + (block_w == 4)]);
988 vp8_mc(s, 0, dst[2] + by_off * s->uvlinesize + bx_off,
989 ref_frame->data[2], &uvmv, x_off + bx_off, y_off + by_off,
990 block_w, block_h, width, height, s->uvlinesize,
991 s->put_pixels_tab[1 + (block_w == 4)]);
992}
993
3b636f21
DC
994/**
995 * Apply motion vectors to prediction buffer, chapter 18.
996 */
997static void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
998 int mb_x, int mb_y)
999{
1000 int x_off = mb_x << 4, y_off = mb_y << 4;
1001 int width = 16*s->mb_width, height = 16*s->mb_height;
3b636f21
DC
1002
1003 if (mb->mode < VP8_MVMODE_SPLIT) {
7c4dcf81
RB
1004 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1005 0, 0, 16, 16, width, height, &mb->mv);
1006 } else switch (mb->partitioning) {
1007 case VP8_SPLITMVMODE_4x4: {
3b636f21 1008 int x, y;
7c4dcf81 1009 VP56mv uvmv;
3b636f21
DC
1010
1011 /* Y */
1012 for (y = 0; y < 4; y++) {
1013 for (x = 0; x < 4; x++) {
1014 vp8_mc(s, 1, dst[0] + 4*y*s->linesize + x*4,
1015 s->framep[mb->ref_frame]->data[0], &mb->bmv[4*y + x],
1016 4*x + x_off, 4*y + y_off, 4, 4,
1017 width, height, s->linesize,
0ef1dbed 1018 s->put_pixels_tab[2]);
3b636f21
DC
1019 }
1020 }
1021
1022 /* U/V */
1023 x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1;
1024 for (y = 0; y < 2; y++) {
1025 for (x = 0; x < 2; x++) {
1026 uvmv.x = mb->bmv[ 2*y * 4 + 2*x ].x +
1027 mb->bmv[ 2*y * 4 + 2*x+1].x +
1028 mb->bmv[(2*y+1) * 4 + 2*x ].x +
1029 mb->bmv[(2*y+1) * 4 + 2*x+1].x;
1030 uvmv.y = mb->bmv[ 2*y * 4 + 2*x ].y +
1031 mb->bmv[ 2*y * 4 + 2*x+1].y +
1032 mb->bmv[(2*y+1) * 4 + 2*x ].y +
1033 mb->bmv[(2*y+1) * 4 + 2*x+1].y;
8f910a56
SG
1034 uvmv.x = (uvmv.x + 2 + (uvmv.x >> (INT_BIT-1))) >> 2;
1035 uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2;
3b636f21
DC
1036 if (s->profile == 3) {
1037 uvmv.x &= ~7;
1038 uvmv.y &= ~7;
1039 }
1040 vp8_mc(s, 0, dst[1] + 4*y*s->uvlinesize + x*4,
1041 s->framep[mb->ref_frame]->data[1], &uvmv,
1042 4*x + x_off, 4*y + y_off, 4, 4,
1043 width, height, s->uvlinesize,
0ef1dbed 1044 s->put_pixels_tab[2]);
3b636f21
DC
1045 vp8_mc(s, 0, dst[2] + 4*y*s->uvlinesize + x*4,
1046 s->framep[mb->ref_frame]->data[2], &uvmv,
1047 4*x + x_off, 4*y + y_off, 4, 4,
1048 width, height, s->uvlinesize,
0ef1dbed 1049 s->put_pixels_tab[2]);
3b636f21
DC
1050 }
1051 }
7c4dcf81
RB
1052 break;
1053 }
1054 case VP8_SPLITMVMODE_16x8:
1055 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1056 0, 0, 16, 8, width, height, &mb->bmv[0]);
1057 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
7ed06b2b 1058 0, 8, 16, 8, width, height, &mb->bmv[1]);
7c4dcf81
RB
1059 break;
1060 case VP8_SPLITMVMODE_8x16:
1061 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1062 0, 0, 8, 16, width, height, &mb->bmv[0]);
1063 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
7ed06b2b 1064 8, 0, 8, 16, width, height, &mb->bmv[1]);
7c4dcf81
RB
1065 break;
1066 case VP8_SPLITMVMODE_8x8:
1067 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1068 0, 0, 8, 8, width, height, &mb->bmv[0]);
1069 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
7ed06b2b 1070 8, 0, 8, 8, width, height, &mb->bmv[1]);
7c4dcf81 1071 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
7ed06b2b 1072 0, 8, 8, 8, width, height, &mb->bmv[2]);
7c4dcf81 1073 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
7ed06b2b 1074 8, 8, 8, 8, width, height, &mb->bmv[3]);
7c4dcf81 1075 break;
3b636f21
DC
1076 }
1077}
1078
1079static void idct_mb(VP8Context *s, uint8_t *y_dst, uint8_t *u_dst, uint8_t *v_dst,
1080 VP8Macroblock *mb)
1081{
1082 int x, y, nnz;
1083
1084 if (mb->mode != MODE_I4x4)
1085 for (y = 0; y < 4; y++) {
1086 for (x = 0; x < 4; x++) {
1087 nnz = s->non_zero_count_cache[y][x];
1088 if (nnz) {
1089 if (nnz == 1)
1090 s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize);
1091 else
1092 s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize);
1093 }
1094 }
1095 y_dst += 4*s->linesize;
1096 }
1097
1098 for (y = 0; y < 2; y++) {
1099 for (x = 0; x < 2; x++) {
1100 nnz = s->non_zero_count_cache[4][(y<<1)+x];
1101 if (nnz) {
1102 if (nnz == 1)
1103 s->vp8dsp.vp8_idct_dc_add(u_dst+4*x, s->block[4][(y<<1)+x], s->uvlinesize);
1104 else
1105 s->vp8dsp.vp8_idct_add(u_dst+4*x, s->block[4][(y<<1)+x], s->uvlinesize);
1106 }
1107
1108 nnz = s->non_zero_count_cache[5][(y<<1)+x];
1109 if (nnz) {
1110 if (nnz == 1)
1111 s->vp8dsp.vp8_idct_dc_add(v_dst+4*x, s->block[5][(y<<1)+x], s->uvlinesize);
1112 else
1113 s->vp8dsp.vp8_idct_add(v_dst+4*x, s->block[5][(y<<1)+x], s->uvlinesize);
1114 }
1115 }
1116 u_dst += 4*s->uvlinesize;
1117 v_dst += 4*s->uvlinesize;
1118 }
1119}
1120
1121static void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, int *level, int *inner, int *hev_thresh)
1122{
1123 int interior_limit, filter_level;
1124
1125 if (s->segmentation.enabled) {
1126 filter_level = s->segmentation.filter_level[mb->segment];
1127 if (!s->segmentation.absolute_vals)
1128 filter_level += s->filter.level;
1129 } else
1130 filter_level = s->filter.level;
1131
1132 if (s->lf_delta.enabled) {
1133 filter_level += s->lf_delta.ref[mb->ref_frame];
1134
1135 if (mb->ref_frame == VP56_FRAME_CURRENT) {
1136 if (mb->mode == MODE_I4x4)
1137 filter_level += s->lf_delta.mode[0];
1138 } else {
1139 if (mb->mode == VP8_MVMODE_ZERO)
1140 filter_level += s->lf_delta.mode[1];
1141 else if (mb->mode == VP8_MVMODE_SPLIT)
1142 filter_level += s->lf_delta.mode[3];
1143 else
1144 filter_level += s->lf_delta.mode[2];
1145 }
1146 }
1147 filter_level = av_clip(filter_level, 0, 63);
1148
1149 interior_limit = filter_level;
1150 if (s->filter.sharpness) {
1151 interior_limit >>= s->filter.sharpness > 4 ? 2 : 1;
1152 interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness);
1153 }
1154 interior_limit = FFMAX(interior_limit, 1);
1155
1156 *level = filter_level;
1157 *inner = interior_limit;
1158
1159 if (hev_thresh) {
1160 *hev_thresh = filter_level >= 15;
1161
1162 if (s->keyframe) {
1163 if (filter_level >= 40)
1164 *hev_thresh = 2;
1165 } else {
1166 if (filter_level >= 40)
1167 *hev_thresh = 3;
1168 else if (filter_level >= 20)
1169 *hev_thresh = 2;
1170 }
1171 }
1172}
1173
1174// TODO: look at backup_mb_border / xchg_mb_border in h264.c
1175static void filter_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb, int mb_x, int mb_y)
1176{
5245c04d 1177 int filter_level, inner_limit, hev_thresh, mbedge_lim, bedge_lim;
3b636f21
DC
1178
1179 filter_level_for_mb(s, mb, &filter_level, &inner_limit, &hev_thresh);
1180 if (!filter_level)
1181 return;
1182
5245c04d
DC
1183 mbedge_lim = 2*(filter_level+2) + inner_limit;
1184 bedge_lim = 2* filter_level + inner_limit;
1185
3b636f21 1186 if (mb_x) {
5245c04d
DC
1187 s->vp8dsp.vp8_h_loop_filter16(dst[0], s->linesize, mbedge_lim, inner_limit, hev_thresh);
1188 s->vp8dsp.vp8_h_loop_filter8 (dst[1], s->uvlinesize, mbedge_lim, inner_limit, hev_thresh);
1189 s->vp8dsp.vp8_h_loop_filter8 (dst[2], s->uvlinesize, mbedge_lim, inner_limit, hev_thresh);
3b636f21
DC
1190 }
1191
1192 if (!mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT) {
5245c04d
DC
1193 s->vp8dsp.vp8_h_loop_filter16_inner(dst[0]+ 4, s->linesize, bedge_lim, inner_limit, hev_thresh);
1194 s->vp8dsp.vp8_h_loop_filter16_inner(dst[0]+ 8, s->linesize, bedge_lim, inner_limit, hev_thresh);
1195 s->vp8dsp.vp8_h_loop_filter16_inner(dst[0]+12, s->linesize, bedge_lim, inner_limit, hev_thresh);
1196 s->vp8dsp.vp8_h_loop_filter8_inner (dst[1]+ 4, s->uvlinesize, bedge_lim, inner_limit, hev_thresh);
1197 s->vp8dsp.vp8_h_loop_filter8_inner (dst[2]+ 4, s->uvlinesize, bedge_lim, inner_limit, hev_thresh);
3b636f21
DC
1198 }
1199
1200 if (mb_y) {
5245c04d
DC
1201 s->vp8dsp.vp8_v_loop_filter16(dst[0], s->linesize, mbedge_lim, inner_limit, hev_thresh);
1202 s->vp8dsp.vp8_v_loop_filter8 (dst[1], s->uvlinesize, mbedge_lim, inner_limit, hev_thresh);
1203 s->vp8dsp.vp8_v_loop_filter8 (dst[2], s->uvlinesize, mbedge_lim, inner_limit, hev_thresh);
3b636f21
DC
1204 }
1205
1206 if (!mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT) {
5245c04d
DC
1207 s->vp8dsp.vp8_v_loop_filter16_inner(dst[0]+ 4*s->linesize, s->linesize, bedge_lim, inner_limit, hev_thresh);
1208 s->vp8dsp.vp8_v_loop_filter16_inner(dst[0]+ 8*s->linesize, s->linesize, bedge_lim, inner_limit, hev_thresh);
1209 s->vp8dsp.vp8_v_loop_filter16_inner(dst[0]+12*s->linesize, s->linesize, bedge_lim, inner_limit, hev_thresh);
1210 s->vp8dsp.vp8_v_loop_filter8_inner (dst[1]+ 4*s->uvlinesize, s->uvlinesize, bedge_lim, inner_limit, hev_thresh);
1211 s->vp8dsp.vp8_v_loop_filter8_inner (dst[2]+ 4*s->uvlinesize, s->uvlinesize, bedge_lim, inner_limit, hev_thresh);
3b636f21
DC
1212 }
1213}
1214
1215static void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8Macroblock *mb, int mb_x, int mb_y)
1216{
1217 int filter_level, inner_limit, mbedge_lim, bedge_lim;
1218
1219 filter_level_for_mb(s, mb, &filter_level, &inner_limit, NULL);
1220 if (!filter_level)
1221 return;
1222
1223 mbedge_lim = 2*(filter_level+2) + inner_limit;
1224 bedge_lim = 2* filter_level + inner_limit;
1225
1226 if (mb_x)
1227 s->vp8dsp.vp8_h_loop_filter_simple(dst, s->linesize, mbedge_lim);
1228 if (!mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT) {
1229 s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, s->linesize, bedge_lim);
1230 s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, s->linesize, bedge_lim);
1231 s->vp8dsp.vp8_h_loop_filter_simple(dst+12, s->linesize, bedge_lim);
1232 }
1233
1234 if (mb_y)
1235 s->vp8dsp.vp8_v_loop_filter_simple(dst, s->linesize, mbedge_lim);
1236 if (!mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT) {
1237 s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*s->linesize, s->linesize, bedge_lim);
1238 s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*s->linesize, s->linesize, bedge_lim);
1239 s->vp8dsp.vp8_v_loop_filter_simple(dst+12*s->linesize, s->linesize, bedge_lim);
1240 }
1241}
1242
1243static void filter_mb_row(VP8Context *s, int mb_y)
1244{
1245 VP8Macroblock *mb = s->macroblocks + mb_y*s->mb_stride;
1246 uint8_t *dst[3] = {
1247 s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize,
1248 s->framep[VP56_FRAME_CURRENT]->data[1] + 8*mb_y*s->uvlinesize,
1249 s->framep[VP56_FRAME_CURRENT]->data[2] + 8*mb_y*s->uvlinesize
1250 };
1251 int mb_x;
1252
1253 for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1254 filter_mb(s, dst, mb++, mb_x, mb_y);
1255 dst[0] += 16;
1256 dst[1] += 8;
1257 dst[2] += 8;
1258 }
1259}
1260
1261static void filter_mb_row_simple(VP8Context *s, int mb_y)
1262{
1263 uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize;
1264 VP8Macroblock *mb = s->macroblocks + mb_y*s->mb_stride;
1265 int mb_x;
1266
1267 for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1268 filter_mb_simple(s, dst, mb++, mb_x, mb_y);
1269 dst += 16;
1270 }
1271}
1272
1273static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
1274 AVPacket *avpkt)
1275{
1276 VP8Context *s = avctx->priv_data;
1277 int ret, mb_x, mb_y, i, y, referenced;
1278 enum AVDiscard skip_thresh;
1279 AVFrame *curframe;
1280
1281 if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0)
1282 return ret;
1283
1284 referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT
1285 || s->update_altref == VP56_FRAME_CURRENT;
1286
1287 skip_thresh = !referenced ? AVDISCARD_NONREF :
1288 !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL;
1289
1290 if (avctx->skip_frame >= skip_thresh) {
1291 s->invisible = 1;
1292 goto skip_decode;
1293 }
1294
1295 for (i = 0; i < 4; i++)
1296 if (&s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1297 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1298 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) {
1299 curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i];
1300 break;
1301 }
1302 if (curframe->data[0])
1303 avctx->release_buffer(avctx, curframe);
1304
1305 curframe->key_frame = s->keyframe;
1306 curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE;
1307 curframe->reference = referenced ? 3 : 0;
1308 if ((ret = avctx->get_buffer(avctx, curframe))) {
1309 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n");
1310 return ret;
1311 }
1312
1313 // Given that arithmetic probabilities are updated every frame, it's quite likely
1314 // that the values we have on a random interframe are complete junk if we didn't
1315 // start decode on a keyframe. So just don't display anything rather than junk.
1316 if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] ||
1317 !s->framep[VP56_FRAME_GOLDEN] ||
1318 !s->framep[VP56_FRAME_GOLDEN2])) {
1319 av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n");
1320 return AVERROR_INVALIDDATA;
1321 }
1322
1323 s->linesize = curframe->linesize[0];
1324 s->uvlinesize = curframe->linesize[1];
1325
1326 if (!s->edge_emu_buffer)
1327 s->edge_emu_buffer = av_malloc(21*s->linesize);
1328
1329 memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz));
1330
1331 // top edge of 127 for intra prediction
1332 if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) {
1333 memset(curframe->data[0] - s->linesize -1, 127, s->linesize +1);
1334 memset(curframe->data[1] - s->uvlinesize-1, 127, s->uvlinesize+1);
1335 memset(curframe->data[2] - s->uvlinesize-1, 127, s->uvlinesize+1);
1336 }
1337
1338 for (mb_y = 0; mb_y < s->mb_height; mb_y++) {
1339 VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)];
1340 VP8Macroblock *mb = s->macroblocks + mb_y*s->mb_stride;
1341 uint8_t *intra4x4 = s->intra4x4_pred_mode + 4*mb_y*s->b4_stride;
1342 uint8_t *dst[3] = {
1343 curframe->data[0] + 16*mb_y*s->linesize,
1344 curframe->data[1] + 8*mb_y*s->uvlinesize,
1345 curframe->data[2] + 8*mb_y*s->uvlinesize
1346 };
1347
1348 memset(s->left_nnz, 0, sizeof(s->left_nnz));
1349
1350 // left edge of 129 for intra prediction
1351 if (!(avctx->flags & CODEC_FLAG_EMU_EDGE))
1352 for (i = 0; i < 3; i++)
1353 for (y = 0; y < 16>>!!i; y++)
1354 dst[i][y*curframe->linesize[i]-1] = 129;
1355
1356 for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1357 decode_mb_mode(s, mb, mb_x, mb_y, intra4x4 + 4*mb_x);
1358
1359 if (!mb->skip)
1360 decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);
1361 else {
1362 AV_ZERO128(s->non_zero_count_cache); // luma
1363 AV_ZERO64(s->non_zero_count_cache[4]); // chroma
1364 }
1365
1366 if (mb->mode <= MODE_I4x4) {
1367 intra_predict(s, dst, mb, intra4x4 + 4*mb_x, mb_x, mb_y);
1368 memset(mb->bmv, 0, sizeof(mb->bmv));
1369 } else {
1370 inter_predict(s, dst, mb, mb_x, mb_y);
1371 }
1372
1373 if (!mb->skip) {
1374 idct_mb(s, dst[0], dst[1], dst[2], mb);
1375 } else {
1376 AV_ZERO64(s->left_nnz);
1377 AV_WN64(s->top_nnz[mb_x], 0); // array of 9, so unaligned
1378
1379 // Reset DC block predictors if they would exist if the mb had coefficients
1380 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
1381 s->left_nnz[8] = 0;
1382 s->top_nnz[mb_x][8] = 0;
1383 }
1384 }
1385
1386 dst[0] += 16;
1387 dst[1] += 8;
1388 dst[2] += 8;
1389 mb++;
1390 }
1391 if (mb_y && s->filter.level && avctx->skip_loop_filter < skip_thresh) {
1392 if (s->filter.simple)
1393 filter_mb_row_simple(s, mb_y-1);
1394 else
1395 filter_mb_row(s, mb_y-1);
1396 }
1397 }
1398 if (s->filter.level && avctx->skip_loop_filter < skip_thresh) {
1399 if (s->filter.simple)
1400 filter_mb_row_simple(s, mb_y-1);
1401 else
1402 filter_mb_row(s, mb_y-1);
1403 }
1404
1405skip_decode:
1406 // if future frames don't use the updated probabilities,
1407 // reset them to the values we saved
1408 if (!s->update_probabilities)
1409 s->prob[0] = s->prob[1];
1410
1411 // check if golden and altref are swapped
1412 if (s->update_altref == VP56_FRAME_GOLDEN &&
1413 s->update_golden == VP56_FRAME_GOLDEN2)
1414 FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN], s->framep[VP56_FRAME_GOLDEN2]);
1415 else {
1416 if (s->update_altref != VP56_FRAME_NONE)
1417 s->framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref];
1418
1419 if (s->update_golden != VP56_FRAME_NONE)
1420 s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden];
1421 }
1422
1423 if (s->update_last) // move cur->prev
1424 s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT];
1425
1426 // release no longer referenced frames
1427 for (i = 0; i < 4; i++)
1428 if (s->frames[i].data[0] &&
1429 &s->frames[i] != s->framep[VP56_FRAME_CURRENT] &&
1430 &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1431 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1432 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2])
1433 avctx->release_buffer(avctx, &s->frames[i]);
1434
1435 if (!s->invisible) {
1436 *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT];
1437 *data_size = sizeof(AVFrame);
1438 }
1439
1440 return avpkt->size;
1441}
1442
1443static av_cold int vp8_decode_init(AVCodecContext *avctx)
1444{
1445 VP8Context *s = avctx->priv_data;
1446
1447 s->avctx = avctx;
1448 avctx->pix_fmt = PIX_FMT_YUV420P;
1449
1450 dsputil_init(&s->dsp, avctx);
1451 ff_h264_pred_init(&s->hpc, CODEC_ID_VP8);
1452 ff_vp8dsp_init(&s->vp8dsp);
1453
1454 // intra pred needs edge emulation among other things
1455 if (avctx->flags&CODEC_FLAG_EMU_EDGE) {
03ac56e7 1456 av_log(avctx, AV_LOG_ERROR, "Edge emulation not supported\n");
3b636f21
DC
1457 return AVERROR_PATCHWELCOME;
1458 }
1459
1460 return 0;
1461}
1462
1463static av_cold int vp8_decode_free(AVCodecContext *avctx)
1464{
1465 vp8_decode_flush(avctx);
1466 return 0;
1467}
1468
1469AVCodec vp8_decoder = {
1470 "vp8",
1471 AVMEDIA_TYPE_VIDEO,
1472 CODEC_ID_VP8,
1473 sizeof(VP8Context),
1474 vp8_decode_init,
1475 NULL,
1476 vp8_decode_free,
1477 vp8_decode_frame,
1478 CODEC_CAP_DR1,
1479 .flush = vp8_decode_flush,
1480 .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"),
1481};