Shut up an uninitialized variable GCC warning in VP8.
[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 {
968570d6
JGG
32 uint8_t filter_level;
33 uint8_t inner_limit;
c55e0d34 34 uint8_t inner_filter;
968570d6
JGG
35} VP8FilterStrength;
36
37typedef struct {
3b636f21
DC
38 uint8_t skip;
39 // todo: make it possible to check for at least (i4x4 or split_mv)
40 // in one op. are others needed?
41 uint8_t mode;
42 uint8_t ref_frame;
43 uint8_t partitioning;
44 VP56mv mv;
45 VP56mv bmv[16];
46} VP8Macroblock;
47
48typedef struct {
49 AVCodecContext *avctx;
50 DSPContext dsp;
51 VP8DSPContext vp8dsp;
52 H264PredContext hpc;
0ef1dbed 53 vp8_mc_func put_pixels_tab[3][3][3];
3b636f21
DC
54 AVFrame frames[4];
55 AVFrame *framep[4];
56 uint8_t *edge_emu_buffer;
57 VP56RangeCoder c; ///< header context, includes mb modes and motion vectors
58 int profile;
59
60 int mb_width; /* number of horizontal MB */
61 int mb_height; /* number of vertical MB */
62 int linesize;
63 int uvlinesize;
64
65 int keyframe;
66 int invisible;
67 int update_last; ///< update VP56_FRAME_PREVIOUS with the current one
68 int update_golden; ///< VP56_FRAME_NONE if not updated, or which frame to copy if so
69 int update_altref;
9ac831c2 70 int deblock_filter;
3b636f21
DC
71
72 /**
73 * If this flag is not set, all the probability updates
74 * are discarded after this frame is decoded.
75 */
76 int update_probabilities;
77
78 /**
79 * All coefficients are contained in separate arith coding contexts.
80 * There can be 1, 2, 4, or 8 of these after the header context.
81 */
82 int num_coeff_partitions;
83 VP56RangeCoder coeff_partition[8];
84
85 VP8Macroblock *macroblocks;
86 VP8Macroblock *macroblocks_base;
968570d6 87 VP8FilterStrength *filter_strength;
3b636f21
DC
88 int mb_stride;
89
90 uint8_t *intra4x4_pred_mode;
91 uint8_t *intra4x4_pred_mode_base;
c55e0d34 92 uint8_t *segmentation_map;
3b636f21
DC
93 int b4_stride;
94
95 /**
9ac831c2
DC
96 * Cache of the top row needed for intra prediction
97 * 16 for luma, 8 for each chroma plane
98 */
99 uint8_t (*top_border)[16+8+8];
100
101 /**
3b636f21
DC
102 * For coeff decode, we need to know whether the above block had non-zero
103 * coefficients. This means for each macroblock, we need data for 4 luma
104 * blocks, 2 u blocks, 2 v blocks, and the luma dc block, for a total of 9
105 * per macroblock. We keep the last row in top_nnz.
106 */
107 uint8_t (*top_nnz)[9];
108 DECLARE_ALIGNED(8, uint8_t, left_nnz)[9];
109
110 /**
111 * This is the index plus one of the last non-zero coeff
112 * for each of the blocks in the current macroblock.
113 * So, 0 -> no coeffs
114 * 1 -> dc-only (special transform)
115 * 2+-> full transform
116 */
117 DECLARE_ALIGNED(16, uint8_t, non_zero_count_cache)[6][4];
118 DECLARE_ALIGNED(16, DCTELEM, block)[6][4][16];
d1c58fce 119 uint8_t intra4x4_pred_mode_mb[16];
3b636f21
DC
120
121 int chroma_pred_mode; ///< 8x8c pred mode of the current macroblock
b9a7186b 122 int segment; ///< segment of the current macroblock
3b636f21
DC
123
124 int mbskip_enabled;
125 int sign_bias[4]; ///< one state [0, 1] per ref frame type
c4211046 126 int ref_count[3];
3b636f21
DC
127
128 /**
129 * Base parameters for segmentation, i.e. per-macroblock parameters.
130 * These must be kept unchanged even if segmentation is not used for
131 * a frame, since the values persist between interframes.
132 */
133 struct {
134 int enabled;
135 int absolute_vals;
136 int update_map;
137 int8_t base_quant[4];
138 int8_t filter_level[4]; ///< base loop filter level
139 } segmentation;
140
141 /**
142 * Macroblocks can have one of 4 different quants in a frame when
143 * segmentation is enabled.
144 * If segmentation is disabled, only the first segment's values are used.
145 */
146 struct {
147 // [0] - DC qmul [1] - AC qmul
148 int16_t luma_qmul[2];
149 int16_t luma_dc_qmul[2]; ///< luma dc-only block quant
150 int16_t chroma_qmul[2];
151 } qmat[4];
152
153 struct {
154 int simple;
155 int level;
156 int sharpness;
157 } filter;
158
159 struct {
160 int enabled; ///< whether each mb can have a different strength based on mode/ref
161
162 /**
163 * filter strength adjustment for the following macroblock modes:
164 * [0] - i4x4
165 * [1] - zero mv
166 * [2] - inter modes except for zero or split mv
167 * [3] - split mv
168 * i16x16 modes never have any adjustment
169 */
170 int8_t mode[4];
171
172 /**
173 * filter strength adjustment for macroblocks that reference:
174 * [0] - intra / VP56_FRAME_CURRENT
175 * [1] - VP56_FRAME_PREVIOUS
176 * [2] - VP56_FRAME_GOLDEN
177 * [3] - altref / VP56_FRAME_GOLDEN2
178 */
179 int8_t ref[4];
180 } lf_delta;
181
182 /**
183 * These are all of the updatable probabilities for binary decisions.
184 * They are only implictly reset on keyframes, making it quite likely
185 * for an interframe to desync if a prior frame's header was corrupt
186 * or missing outright!
187 */
188 struct {
189 uint8_t segmentid[3];
190 uint8_t mbskip;
191 uint8_t intra;
192 uint8_t last;
193 uint8_t golden;
194 uint8_t pred16x16[4];
195 uint8_t pred8x8c[3];
196 uint8_t token[4][8][3][NUM_DCT_TOKENS-1];
197 uint8_t mvc[2][19];
198 } prob[2];
199} VP8Context;
200
201#define RL24(p) (AV_RL16(p) + ((p)[2] << 16))
202
203static void vp8_decode_flush(AVCodecContext *avctx)
204{
205 VP8Context *s = avctx->priv_data;
206 int i;
207
208 for (i = 0; i < 4; i++)
209 if (s->frames[i].data[0])
210 avctx->release_buffer(avctx, &s->frames[i]);
211 memset(s->framep, 0, sizeof(s->framep));
212
213 av_freep(&s->macroblocks_base);
214 av_freep(&s->intra4x4_pred_mode_base);
215 av_freep(&s->top_nnz);
216 av_freep(&s->edge_emu_buffer);
9ac831c2 217 av_freep(&s->top_border);
c55e0d34 218 av_freep(&s->segmentation_map);
3b636f21
DC
219
220 s->macroblocks = NULL;
221 s->intra4x4_pred_mode = NULL;
222}
223
224static int update_dimensions(VP8Context *s, int width, int height)
225{
226 int i;
227
228 if (avcodec_check_dimensions(s->avctx, width, height))
229 return AVERROR_INVALIDDATA;
230
231 vp8_decode_flush(s->avctx);
232
233 avcodec_set_dimensions(s->avctx, width, height);
234
235 s->mb_width = (s->avctx->coded_width +15) / 16;
236 s->mb_height = (s->avctx->coded_height+15) / 16;
237
238 // we allocate a border around the top/left of intra4x4 modes
239 // this is 4 blocks for intra4x4 to keep 4-byte alignment for fill_rectangle
240 s->mb_stride = s->mb_width+1;
241 s->b4_stride = 4*s->mb_stride;
242
c55e0d34 243 s->macroblocks_base = av_mallocz((s->mb_stride+s->mb_height*2+2)*sizeof(*s->macroblocks));
968570d6 244 s->filter_strength = av_mallocz(s->mb_stride*sizeof(*s->filter_strength));
3b636f21
DC
245 s->intra4x4_pred_mode_base = av_mallocz(s->b4_stride*(4*s->mb_height+1));
246 s->top_nnz = av_mallocz(s->mb_width*sizeof(*s->top_nnz));
9ac831c2 247 s->top_border = av_mallocz((s->mb_width+1)*sizeof(*s->top_border));
c55e0d34 248 s->segmentation_map = av_mallocz(s->mb_stride*s->mb_height);
3b636f21 249
c55e0d34
JGG
250 if (!s->macroblocks_base || !s->filter_strength || !s->intra4x4_pred_mode_base ||
251 !s->top_nnz || !s->top_border || !s->segmentation_map)
b6c420ce
DC
252 return AVERROR(ENOMEM);
253
c55e0d34 254 s->macroblocks = s->macroblocks_base + 1;
3b636f21
DC
255 s->intra4x4_pred_mode = s->intra4x4_pred_mode_base + 4 + s->b4_stride;
256
257 memset(s->intra4x4_pred_mode_base, DC_PRED, s->b4_stride);
258 for (i = 0; i < 4*s->mb_height; i++)
259 s->intra4x4_pred_mode[i*s->b4_stride-1] = DC_PRED;
260
261 return 0;
262}
263
264static void parse_segment_info(VP8Context *s)
265{
266 VP56RangeCoder *c = &s->c;
267 int i;
268
269 s->segmentation.update_map = vp8_rac_get(c);
270
271 if (vp8_rac_get(c)) { // update segment feature data
272 s->segmentation.absolute_vals = vp8_rac_get(c);
273
274 for (i = 0; i < 4; i++)
275 s->segmentation.base_quant[i] = vp8_rac_get_sint(c, 7);
276
277 for (i = 0; i < 4; i++)
278 s->segmentation.filter_level[i] = vp8_rac_get_sint(c, 6);
279 }
280 if (s->segmentation.update_map)
281 for (i = 0; i < 3; i++)
282 s->prob->segmentid[i] = vp8_rac_get(c) ? vp8_rac_get_uint(c, 8) : 255;
283}
284
285static void update_lf_deltas(VP8Context *s)
286{
287 VP56RangeCoder *c = &s->c;
288 int i;
289
290 for (i = 0; i < 4; i++)
291 s->lf_delta.ref[i] = vp8_rac_get_sint(c, 6);
292
293 for (i = 0; i < 4; i++)
294 s->lf_delta.mode[i] = vp8_rac_get_sint(c, 6);
295}
296
297static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size)
298{
299 const uint8_t *sizes = buf;
300 int i;
301
302 s->num_coeff_partitions = 1 << vp8_rac_get_uint(&s->c, 2);
303
304 buf += 3*(s->num_coeff_partitions-1);
305 buf_size -= 3*(s->num_coeff_partitions-1);
306 if (buf_size < 0)
307 return -1;
308
309 for (i = 0; i < s->num_coeff_partitions-1; i++) {
310 int size = RL24(sizes + 3*i);
311 if (buf_size - size < 0)
312 return -1;
313
314 vp56_init_range_decoder(&s->coeff_partition[i], buf, size);
315 buf += size;
316 buf_size -= size;
317 }
318 vp56_init_range_decoder(&s->coeff_partition[i], buf, buf_size);
319
320 return 0;
321}
322
323static void get_quants(VP8Context *s)
324{
325 VP56RangeCoder *c = &s->c;
326 int i, base_qi;
327
328 int yac_qi = vp8_rac_get_uint(c, 7);
329 int ydc_delta = vp8_rac_get_sint(c, 4);
330 int y2dc_delta = vp8_rac_get_sint(c, 4);
331 int y2ac_delta = vp8_rac_get_sint(c, 4);
332 int uvdc_delta = vp8_rac_get_sint(c, 4);
333 int uvac_delta = vp8_rac_get_sint(c, 4);
334
335 for (i = 0; i < 4; i++) {
336 if (s->segmentation.enabled) {
337 base_qi = s->segmentation.base_quant[i];
338 if (!s->segmentation.absolute_vals)
339 base_qi += yac_qi;
340 } else
341 base_qi = yac_qi;
342
343 s->qmat[i].luma_qmul[0] = vp8_dc_qlookup[av_clip(base_qi + ydc_delta , 0, 127)];
344 s->qmat[i].luma_qmul[1] = vp8_ac_qlookup[av_clip(base_qi , 0, 127)];
345 s->qmat[i].luma_dc_qmul[0] = 2 * vp8_dc_qlookup[av_clip(base_qi + y2dc_delta, 0, 127)];
346 s->qmat[i].luma_dc_qmul[1] = 155 * vp8_ac_qlookup[av_clip(base_qi + y2ac_delta, 0, 127)] / 100;
347 s->qmat[i].chroma_qmul[0] = vp8_dc_qlookup[av_clip(base_qi + uvdc_delta, 0, 127)];
348 s->qmat[i].chroma_qmul[1] = vp8_ac_qlookup[av_clip(base_qi + uvac_delta, 0, 127)];
349
350 s->qmat[i].luma_dc_qmul[1] = FFMAX(s->qmat[i].luma_dc_qmul[1], 8);
351 s->qmat[i].chroma_qmul[0] = FFMIN(s->qmat[i].chroma_qmul[0], 132);
352 }
353}
354
355/**
356 * Determine which buffers golden and altref should be updated with after this frame.
357 * The spec isn't clear here, so I'm going by my understanding of what libvpx does
358 *
359 * Intra frames update all 3 references
360 * Inter frames update VP56_FRAME_PREVIOUS if the update_last flag is set
361 * If the update (golden|altref) flag is set, it's updated with the current frame
362 * if update_last is set, and VP56_FRAME_PREVIOUS otherwise.
363 * If the flag is not set, the number read means:
364 * 0: no update
365 * 1: VP56_FRAME_PREVIOUS
366 * 2: update golden with altref, or update altref with golden
367 */
368static VP56Frame ref_to_update(VP8Context *s, int update, VP56Frame ref)
369{
370 VP56RangeCoder *c = &s->c;
371
372 if (update)
373 return VP56_FRAME_CURRENT;
374
375 switch (vp8_rac_get_uint(c, 2)) {
376 case 1:
377 return VP56_FRAME_PREVIOUS;
378 case 2:
379 return (ref == VP56_FRAME_GOLDEN) ? VP56_FRAME_GOLDEN2 : VP56_FRAME_GOLDEN;
380 }
381 return VP56_FRAME_NONE;
382}
383
384static void update_refs(VP8Context *s)
385{
386 VP56RangeCoder *c = &s->c;
387
388 int update_golden = vp8_rac_get(c);
389 int update_altref = vp8_rac_get(c);
390
391 s->update_golden = ref_to_update(s, update_golden, VP56_FRAME_GOLDEN);
392 s->update_altref = ref_to_update(s, update_altref, VP56_FRAME_GOLDEN2);
393}
394
395static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)
396{
397 VP56RangeCoder *c = &s->c;
398 int header_size, hscale, vscale, i, j, k, l, ret;
399 int width = s->avctx->width;
400 int height = s->avctx->height;
401
402 s->keyframe = !(buf[0] & 1);
403 s->profile = (buf[0]>>1) & 7;
404 s->invisible = !(buf[0] & 0x10);
405 header_size = RL24(buf) >> 5;
406 buf += 3;
407 buf_size -= 3;
408
0ef1dbed
DC
409 if (s->profile > 3)
410 av_log(s->avctx, AV_LOG_WARNING, "Unknown profile %d\n", s->profile);
411
412 if (!s->profile)
413 memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_epel_pixels_tab, sizeof(s->put_pixels_tab));
414 else // profile 1-3 use bilinear, 4+ aren't defined so whatever
415 memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_bilinear_pixels_tab, sizeof(s->put_pixels_tab));
3b636f21
DC
416
417 if (header_size > buf_size - 7*s->keyframe) {
418 av_log(s->avctx, AV_LOG_ERROR, "Header size larger than data provided\n");
419 return AVERROR_INVALIDDATA;
420 }
421
422 if (s->keyframe) {
423 if (RL24(buf) != 0x2a019d) {
424 av_log(s->avctx, AV_LOG_ERROR, "Invalid start code 0x%x\n", RL24(buf));
425 return AVERROR_INVALIDDATA;
426 }
427 width = AV_RL16(buf+3) & 0x3fff;
428 height = AV_RL16(buf+5) & 0x3fff;
429 hscale = buf[4] >> 6;
430 vscale = buf[6] >> 6;
431 buf += 7;
432 buf_size -= 7;
433
92a54426
MR
434 if (hscale || vscale)
435 av_log_missing_feature(s->avctx, "Upscaling", 1);
436
3b636f21
DC
437 s->update_golden = s->update_altref = VP56_FRAME_CURRENT;
438 memcpy(s->prob->token , vp8_token_default_probs , sizeof(s->prob->token));
439 memcpy(s->prob->pred16x16, vp8_pred16x16_prob_inter, sizeof(s->prob->pred16x16));
440 memcpy(s->prob->pred8x8c , vp8_pred8x8c_prob_inter , sizeof(s->prob->pred8x8c));
441 memcpy(s->prob->mvc , vp8_mv_default_prob , sizeof(s->prob->mvc));
442 memset(&s->segmentation, 0, sizeof(s->segmentation));
443 }
444
445 if (!s->macroblocks_base || /* first frame */
446 width != s->avctx->width || height != s->avctx->height) {
447 if ((ret = update_dimensions(s, width, height) < 0))
448 return ret;
449 }
450
451 vp56_init_range_decoder(c, buf, header_size);
452 buf += header_size;
453 buf_size -= header_size;
454
455 if (s->keyframe) {
456 if (vp8_rac_get(c))
457 av_log(s->avctx, AV_LOG_WARNING, "Unspecified colorspace\n");
458 vp8_rac_get(c); // whether we can skip clamping in dsp functions
459 }
460
461 if ((s->segmentation.enabled = vp8_rac_get(c)))
462 parse_segment_info(s);
463 else
464 s->segmentation.update_map = 0; // FIXME: move this to some init function?
465
466 s->filter.simple = vp8_rac_get(c);
467 s->filter.level = vp8_rac_get_uint(c, 6);
468 s->filter.sharpness = vp8_rac_get_uint(c, 3);
469
470 if ((s->lf_delta.enabled = vp8_rac_get(c)))
471 if (vp8_rac_get(c))
472 update_lf_deltas(s);
473
474 if (setup_partitions(s, buf, buf_size)) {
475 av_log(s->avctx, AV_LOG_ERROR, "Invalid partitions\n");
476 return AVERROR_INVALIDDATA;
477 }
478
479 get_quants(s);
480
481 if (!s->keyframe) {
482 update_refs(s);
483 s->sign_bias[VP56_FRAME_GOLDEN] = vp8_rac_get(c);
484 s->sign_bias[VP56_FRAME_GOLDEN2 /* altref */] = vp8_rac_get(c);
485 }
486
487 // if we aren't saving this frame's probabilities for future frames,
488 // make a copy of the current probabilities
489 if (!(s->update_probabilities = vp8_rac_get(c)))
490 s->prob[1] = s->prob[0];
491
492 s->update_last = s->keyframe || vp8_rac_get(c);
493
494 for (i = 0; i < 4; i++)
495 for (j = 0; j < 8; j++)
496 for (k = 0; k < 3; k++)
497 for (l = 0; l < NUM_DCT_TOKENS-1; l++)
498 if (vp56_rac_get_prob(c, vp8_token_update_probs[i][j][k][l]))
499 s->prob->token[i][j][k][l] = vp8_rac_get_uint(c, 8);
500
501 if ((s->mbskip_enabled = vp8_rac_get(c)))
502 s->prob->mbskip = vp8_rac_get_uint(c, 8);
503
504 if (!s->keyframe) {
505 s->prob->intra = vp8_rac_get_uint(c, 8);
506 s->prob->last = vp8_rac_get_uint(c, 8);
507 s->prob->golden = vp8_rac_get_uint(c, 8);
508
509 if (vp8_rac_get(c))
510 for (i = 0; i < 4; i++)
511 s->prob->pred16x16[i] = vp8_rac_get_uint(c, 8);
512 if (vp8_rac_get(c))
513 for (i = 0; i < 3; i++)
514 s->prob->pred8x8c[i] = vp8_rac_get_uint(c, 8);
515
516 // 17.2 MV probability update
517 for (i = 0; i < 2; i++)
518 for (j = 0; j < 19; j++)
519 if (vp56_rac_get_prob(c, vp8_mv_update_prob[i][j]))
520 s->prob->mvc[i][j] = vp8_rac_get_nn(c);
521 }
522
523 return 0;
524}
525
526static inline void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src,
527 int mb_x, int mb_y)
528{
529#define MARGIN (16 << 2)
530 dst->x = av_clip(src->x, -((mb_x << 6) + MARGIN),
531 ((s->mb_width - 1 - mb_x) << 6) + MARGIN);
532 dst->y = av_clip(src->y, -((mb_y << 6) + MARGIN),
533 ((s->mb_height - 1 - mb_y) << 6) + MARGIN);
534}
535
536static void find_near_mvs(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
702e8d33 537 VP56mv near[2], VP56mv *best, uint8_t cnt[4])
3b636f21 538{
c55e0d34
JGG
539 VP8Macroblock *mb_edge[3] = { mb + 2 /* top */,
540 mb - 1 /* left */,
541 mb + 1 /* top-left */ };
3b636f21
DC
542 enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT };
543 VP56mv near_mv[4] = {{ 0 }};
544 enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV };
702e8d33 545 int idx = CNT_ZERO;
3b636f21 546 int best_idx = CNT_ZERO;
702e8d33
JGG
547 int cur_sign_bias = s->sign_bias[mb->ref_frame];
548 int *sign_bias = s->sign_bias;
3b636f21
DC
549
550 /* Process MB on top, left and top-left */
702e8d33
JGG
551 #define MV_EDGE_CHECK(n)\
552 {\
553 VP8Macroblock *edge = mb_edge[n];\
554 int edge_ref = edge->ref_frame;\
555 if (edge_ref != VP56_FRAME_CURRENT) {\
556 uint32_t mv = AV_RN32A(&edge->mv);\
557 if (mv) {\
558 if (cur_sign_bias != sign_bias[edge_ref]) {\
559 /* SWAR negate of the values in mv. */\
560 mv = ((mv&0x80008000) + 0x00010001) ^ (mv&0x7fff7fff);\
561 }\
562 if (!n || mv != AV_RN32A(&near_mv[idx]))\
563 AV_WN32A(&near_mv[++idx], mv);\
564 cnt[idx] += 1 + (n != 2);\
565 } else\
566 cnt[CNT_ZERO] += 1 + (n != 2);\
567 }\
3b636f21 568 }
702e8d33
JGG
569 MV_EDGE_CHECK(0)
570 MV_EDGE_CHECK(1)
571 MV_EDGE_CHECK(2)
3b636f21 572
702e8d33
JGG
573 /* If we have three distinct MVs, merge first and last if they're the same */
574 if (cnt[CNT_SPLITMV] && AV_RN32A(&near_mv[1+EDGE_TOP]) == AV_RN32A(&near_mv[1+EDGE_TOPLEFT]))
3b636f21
DC
575 cnt[CNT_NEAREST] += 1;
576
577 cnt[CNT_SPLITMV] = ((mb_edge[EDGE_LEFT]->mode == VP8_MVMODE_SPLIT) +
578 (mb_edge[EDGE_TOP]->mode == VP8_MVMODE_SPLIT)) * 2 +
579 (mb_edge[EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT);
580
581 /* Swap near and nearest if necessary */
582 if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) {
702e8d33
JGG
583 FFSWAP(uint8_t, cnt[CNT_NEAREST], cnt[CNT_NEAR]);
584 FFSWAP( VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]);
3b636f21
DC
585 }
586
587 /* Choose the best mv out of 0,0 and the nearest mv */
588 if (cnt[CNT_NEAREST] >= cnt[CNT_ZERO])
589 best_idx = CNT_NEAREST;
590
591 clamp_mv(s, best, &near_mv[best_idx], mb_x, mb_y);
592 near[0] = near_mv[CNT_NEAREST];
593 near[1] = near_mv[CNT_NEAR];
594}
595
596/**
597 * Motion vector coding, 17.1.
598 */
599static int read_mv_component(VP56RangeCoder *c, const uint8_t *p)
600{
601 int x = 0;
602
603 if (vp56_rac_get_prob(c, p[0])) {
604 int i;
605
606 for (i = 0; i < 3; i++)
607 x += vp56_rac_get_prob(c, p[9 + i]) << i;
608 for (i = 9; i > 3; i--)
609 x += vp56_rac_get_prob(c, p[9 + i]) << i;
610 if (!(x & 0xFFF0) || vp56_rac_get_prob(c, p[12]))
611 x += 8;
612 } else
613 x = vp8_rac_get_tree(c, vp8_small_mvtree, &p[2]);
614
615 return (x && vp56_rac_get_prob(c, p[1])) ? -x : x;
616}
617
7bf254c4 618static const uint8_t *get_submv_prob(uint32_t left, uint32_t top)
3b636f21 619{
7bf254c4
JGG
620 if (left == top)
621 return vp8_submv_prob[4-!!left];
622 if (!top)
3b636f21 623 return vp8_submv_prob[2];
7bf254c4 624 return vp8_submv_prob[1-!!left];
3b636f21
DC
625}
626
627/**
628 * Split motion vector prediction, 16.4.
7ed06b2b 629 * @returns the number of motion vectors parsed (2, 4 or 16)
3b636f21 630 */
7ed06b2b 631static int decode_splitmvs(VP8Context *s, VP56RangeCoder *c,
3b636f21
DC
632 VP8Macroblock *mb, VP56mv *base_mv)
633{
634 int part_idx = mb->partitioning =
635 vp8_rac_get_tree(c, vp8_mbsplit_tree, vp8_mbsplit_prob);
636 int n, num = vp8_mbsplit_count[part_idx];
c55e0d34 637 VP8Macroblock *top_mb = &mb[2];
7bf254c4
JGG
638 VP8Macroblock *left_mb = &mb[-1];
639 const uint8_t *mbsplits_left = vp8_mbsplits[left_mb->partitioning],
640 *mbsplits_top = vp8_mbsplits[top_mb->partitioning],
641 *mbsplits_cur = vp8_mbsplits[part_idx],
7ed06b2b 642 *firstidx = vp8_mbfirstidx[part_idx];
c55e0d34
JGG
643 VP56mv *top_mv = top_mb->bmv;
644 VP56mv *left_mv = left_mb->bmv;
645 VP56mv *cur_mv = mb->bmv;
3b636f21
DC
646
647 for (n = 0; n < num; n++) {
7ed06b2b 648 int k = firstidx[n];
7bf254c4 649 uint32_t left, above;
7ed06b2b
RB
650 const uint8_t *submv_prob;
651
7bf254c4
JGG
652 if (!(k & 3))
653 left = AV_RN32A(&left_mv[mbsplits_left[k + 3]]);
654 else
655 left = AV_RN32A(&cur_mv[mbsplits_cur[k - 1]]);
656 if (k <= 3)
657 above = AV_RN32A(&top_mv[mbsplits_top[k + 12]]);
658 else
659 above = AV_RN32A(&cur_mv[mbsplits_cur[k - 4]]);
7ed06b2b
RB
660
661 submv_prob = get_submv_prob(left, above);
3b636f21
DC
662
663 switch (vp8_rac_get_tree(c, vp8_submv_ref_tree, submv_prob)) {
664 case VP8_SUBMVMODE_NEW4X4:
7ed06b2b
RB
665 mb->bmv[n].y = base_mv->y + read_mv_component(c, s->prob->mvc[0]);
666 mb->bmv[n].x = base_mv->x + read_mv_component(c, s->prob->mvc[1]);
3b636f21
DC
667 break;
668 case VP8_SUBMVMODE_ZERO4X4:
7bf254c4 669 AV_WN32A(&mb->bmv[n], 0);
3b636f21
DC
670 break;
671 case VP8_SUBMVMODE_LEFT4X4:
7bf254c4 672 AV_WN32A(&mb->bmv[n], left);
3b636f21
DC
673 break;
674 case VP8_SUBMVMODE_TOP4X4:
7bf254c4 675 AV_WN32A(&mb->bmv[n], above);
3b636f21
DC
676 break;
677 }
3b636f21 678 }
7ed06b2b
RB
679
680 return num;
3b636f21
DC
681}
682
683static inline void decode_intra4x4_modes(VP56RangeCoder *c, uint8_t *intra4x4,
684 int stride, int keyframe)
685{
d1c58fce 686 int x, y, t, l, i;
3b636f21 687
d1c58fce
JGG
688 if (keyframe) {
689 const uint8_t *ctx;
690 for (y = 0; y < 4; y++) {
691 for (x = 0; x < 4; x++) {
3b636f21
DC
692 t = intra4x4[x - stride];
693 l = intra4x4[x - 1];
694 ctx = vp8_pred4x4_prob_intra[t][l];
d1c58fce 695 intra4x4[x] = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx);
3b636f21 696 }
d1c58fce 697 intra4x4 += stride;
3b636f21 698 }
d1c58fce
JGG
699 } else {
700 for (i = 0; i < 16; i++)
701 intra4x4[i] = vp8_rac_get_tree(c, vp8_pred4x4_tree, vp8_pred4x4_prob_inter);
3b636f21
DC
702 }
703}
704
705static void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
c55e0d34 706 uint8_t *intra4x4, uint8_t *segment)
3b636f21
DC
707{
708 VP56RangeCoder *c = &s->c;
3b636f21
DC
709
710 if (s->segmentation.update_map)
c55e0d34 711 *segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid);
b9a7186b 712 s->segment = *segment;
3b636f21
DC
713
714 mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0;
715
716 if (s->keyframe) {
717 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra);
718
719 if (mb->mode == MODE_I4x4) {
720 decode_intra4x4_modes(c, intra4x4, s->b4_stride, 1);
721 } else
722 fill_rectangle(intra4x4, 4, 4, s->b4_stride, vp8_pred4x4_mode[mb->mode], 1);
723
724 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra);
725 mb->ref_frame = VP56_FRAME_CURRENT;
726 } else if (vp56_rac_get_prob(c, s->prob->intra)) {
727 VP56mv near[2], best;
702e8d33 728 uint8_t cnt[4] = { 0 };
3b636f21
DC
729 uint8_t p[4];
730
731 // inter MB, 16.2
732 if (vp56_rac_get_prob(c, s->prob->last))
733 mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ?
734 VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN;
735 else
736 mb->ref_frame = VP56_FRAME_PREVIOUS;
c4211046 737 s->ref_count[mb->ref_frame-1]++;
3b636f21
DC
738
739 // motion vectors, 16.3
740 find_near_mvs(s, mb, mb_x, mb_y, near, &best, cnt);
702e8d33
JGG
741 p[0] = vp8_mode_contexts[cnt[0]][0];
742 p[1] = vp8_mode_contexts[cnt[1]][1];
743 p[2] = vp8_mode_contexts[cnt[2]][2];
744 p[3] = vp8_mode_contexts[cnt[3]][3];
3b636f21
DC
745 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_mvinter, p);
746 switch (mb->mode) {
747 case VP8_MVMODE_SPLIT:
7ed06b2b 748 mb->mv = mb->bmv[decode_splitmvs(s, c, mb, &best) - 1];
3b636f21
DC
749 break;
750 case VP8_MVMODE_ZERO:
b946111f 751 AV_WN32A(&mb->mv, 0);
3b636f21
DC
752 break;
753 case VP8_MVMODE_NEAREST:
754 clamp_mv(s, &mb->mv, &near[0], mb_x, mb_y);
755 break;
756 case VP8_MVMODE_NEAR:
757 clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y);
758 break;
759 case VP8_MVMODE_NEW:
760 mb->mv.y = best.y + read_mv_component(c, s->prob->mvc[0]);
761 mb->mv.x = best.x + read_mv_component(c, s->prob->mvc[1]);
762 break;
763 }
764 if (mb->mode != VP8_MVMODE_SPLIT) {
7ed06b2b
RB
765 mb->partitioning = VP8_SPLITMVMODE_NONE;
766 mb->bmv[0] = mb->mv;
3b636f21
DC
767 }
768 } else {
769 // intra MB, 16.1
770 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16);
771
158e062c 772 if (mb->mode == MODE_I4x4)
d1c58fce 773 decode_intra4x4_modes(c, intra4x4, 4, 0);
3b636f21
DC
774
775 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c);
776 mb->ref_frame = VP56_FRAME_CURRENT;
b946111f
JGG
777 mb->partitioning = VP8_SPLITMVMODE_NONE;
778 AV_WN32A(&mb->bmv[0], 0);
3b636f21
DC
779 }
780}
781
782/**
e394953e
RB
783 * @param c arithmetic bitstream reader context
784 * @param block destination for block coefficients
785 * @param probs probabilities to use when reading trees from the bitstream
3b636f21
DC
786 * @param i initial coeff index, 0 unless a separate DC block is coded
787 * @param zero_nhood the initial prediction context for number of surrounding
788 * all-zero blocks (only left/top, so 0-2)
3fa76268 789 * @param qmul array holding the dc/ac dequant factor at position 0/1
3b636f21
DC
790 * @return 0 if no coeffs were decoded
791 * otherwise, the index of the last coeff decoded plus one
792 */
793static int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16],
794 uint8_t probs[8][3][NUM_DCT_TOKENS-1],
795 int i, int zero_nhood, int16_t qmul[2])
796{
797 int token, nonzero = 0;
798 int offset = 0;
799
800 for (; i < 16; i++) {
801 token = vp8_rac_get_tree_with_offset(c, vp8_coeff_tree, probs[vp8_coeff_band[i]][zero_nhood], offset);
802
803 if (token == DCT_EOB)
804 break;
805 else if (token >= DCT_CAT1) {
806 int cat = token-DCT_CAT1;
807 token = vp8_rac_get_coeff(c, vp8_dct_cat_prob[cat]);
2a38c2e9 808 token += 3 + (2<<cat);
3b636f21
DC
809 }
810
811 // after the first token, the non-zero prediction context becomes
812 // based on the last decoded coeff
813 if (!token) {
814 zero_nhood = 0;
815 offset = 1;
816 continue;
817 } else if (token == 1)
818 zero_nhood = 1;
819 else
820 zero_nhood = 2;
821
822 // todo: full [16] qmat? load into register?
823 block[zigzag_scan[i]] = (vp8_rac_get(c) ? -token : token) * qmul[!!i];
824 nonzero = i+1;
825 offset = 0;
826 }
827 return nonzero;
828}
829
830static void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,
831 uint8_t t_nnz[9], uint8_t l_nnz[9])
832{
833 LOCAL_ALIGNED_16(DCTELEM, dc,[16]);
834 int i, x, y, luma_start = 0, luma_ctx = 3;
835 int nnz_pred, nnz, nnz_total = 0;
b9a7186b 836 int segment = s->segment;
3b636f21
DC
837
838 s->dsp.clear_blocks((DCTELEM *)s->block);
839
840 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
841 AV_ZERO128(dc);
842 AV_ZERO128(dc+8);
843 nnz_pred = t_nnz[8] + l_nnz[8];
844
845 // decode DC values and do hadamard
846 nnz = decode_block_coeffs(c, dc, s->prob->token[1], 0, nnz_pred,
847 s->qmat[segment].luma_dc_qmul);
848 l_nnz[8] = t_nnz[8] = !!nnz;
849 nnz_total += nnz;
850 s->vp8dsp.vp8_luma_dc_wht(s->block, dc);
851 luma_start = 1;
852 luma_ctx = 0;
853 }
854
855 // luma blocks
856 for (y = 0; y < 4; y++)
857 for (x = 0; x < 4; x++) {
858 nnz_pred = l_nnz[y] + t_nnz[x];
859 nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start,
860 nnz_pred, s->qmat[segment].luma_qmul);
861 // nnz+luma_start may be one more than the actual last index, but we don't care
862 s->non_zero_count_cache[y][x] = nnz + luma_start;
863 t_nnz[x] = l_nnz[y] = !!nnz;
864 nnz_total += nnz;
865 }
866
867 // chroma blocks
868 // TODO: what to do about dimensions? 2nd dim for luma is x,
869 // but for chroma it's (y<<1)|x
870 for (i = 4; i < 6; i++)
871 for (y = 0; y < 2; y++)
872 for (x = 0; x < 2; x++) {
873 nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x];
874 nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0,
875 nnz_pred, s->qmat[segment].chroma_qmul);
876 s->non_zero_count_cache[i][(y<<1)+x] = nnz;
877 t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz;
878 nnz_total += nnz;
879 }
880
881 // if there were no coded coeffs despite the macroblock not being marked skip,
882 // we MUST not do the inner loop filter and should not do IDCT
883 // Since skip isn't used for bitstream prediction, just manually set it.
884 if (!nnz_total)
885 mb->skip = 1;
886}
887
9ac831c2
DC
888static av_always_inline
889void backup_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
890 int linesize, int uvlinesize, int simple)
891{
892 AV_COPY128(top_border, src_y + 15*linesize);
893 if (!simple) {
894 AV_COPY64(top_border+16, src_cb + 7*uvlinesize);
895 AV_COPY64(top_border+24, src_cr + 7*uvlinesize);
896 }
897}
898
899static av_always_inline
900void xchg_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
901 int linesize, int uvlinesize, int mb_x, int mb_y, int mb_width,
902 int simple, int xchg)
903{
904 uint8_t *top_border_m1 = top_border-32; // for TL prediction
905 src_y -= linesize;
906 src_cb -= uvlinesize;
907 src_cr -= uvlinesize;
908
096971e8
MR
909#define XCHG(a,b,xchg) do { \
910 if (xchg) AV_SWAP64(b,a); \
911 else AV_COPY64(b,a); \
912 } while (0)
9ac831c2
DC
913
914 XCHG(top_border_m1+8, src_y-8, xchg);
915 XCHG(top_border, src_y, xchg);
916 XCHG(top_border+8, src_y+8, 1);
070ce7ef 917 if (mb_x < mb_width-1)
9ac831c2 918 XCHG(top_border+32, src_y+16, 1);
070ce7ef 919
9ac831c2
DC
920 // only copy chroma for normal loop filter
921 // or to initialize the top row to 127
922 if (!simple || !mb_y) {
923 XCHG(top_border_m1+16, src_cb-8, xchg);
924 XCHG(top_border_m1+24, src_cr-8, xchg);
925 XCHG(top_border+16, src_cb, 1);
926 XCHG(top_border+24, src_cr, 1);
927 }
928}
929
3b636f21
DC
930static int check_intra_pred_mode(int mode, int mb_x, int mb_y)
931{
932 if (mode == DC_PRED8x8) {
933 if (!(mb_x|mb_y))
934 mode = DC_128_PRED8x8;
935 else if (!mb_y)
936 mode = LEFT_DC_PRED8x8;
937 else if (!mb_x)
938 mode = TOP_DC_PRED8x8;
939 }
940 return mode;
941}
942
943static void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
d1c58fce 944 uint8_t *intra4x4, int mb_x, int mb_y)
3b636f21
DC
945{
946 int x, y, mode, nnz, tr;
947
9ac831c2
DC
948 // for the first row, we need to run xchg_mb_border to init the top edge to 127
949 // otherwise, skip it if we aren't going to deblock
950 if (s->deblock_filter || !mb_y)
951 xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
952 s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
953 s->filter.simple, 1);
954
3b636f21
DC
955 if (mb->mode < MODE_I4x4) {
956 mode = check_intra_pred_mode(mb->mode, mb_x, mb_y);
957 s->hpc.pred16x16[mode](dst[0], s->linesize);
958 } else {
959 uint8_t *ptr = dst[0];
d1c58fce 960 int stride = s->keyframe ? s->b4_stride : 4;
3b636f21
DC
961
962 // all blocks on the right edge of the macroblock use bottom edge
963 // the top macroblock for their topright edge
964 uint8_t *tr_right = ptr - s->linesize + 16;
965
966 // if we're on the right edge of the frame, said edge is extended
967 // from the top macroblock
968 if (mb_x == s->mb_width-1) {
969 tr = tr_right[-1]*0x01010101;
970 tr_right = (uint8_t *)&tr;
971 }
972
973 for (y = 0; y < 4; y++) {
974 uint8_t *topright = ptr + 4 - s->linesize;
975 for (x = 0; x < 4; x++) {
976 if (x == 3)
977 topright = tr_right;
978
d1c58fce 979 s->hpc.pred4x4[intra4x4[x]](ptr+4*x, topright, s->linesize);
3b636f21
DC
980
981 nnz = s->non_zero_count_cache[y][x];
982 if (nnz) {
983 if (nnz == 1)
984 s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize);
985 else
986 s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize);
987 }
988 topright += 4;
989 }
990
991 ptr += 4*s->linesize;
d1c58fce 992 intra4x4 += stride;
3b636f21
DC
993 }
994 }
995
996 mode = check_intra_pred_mode(s->chroma_pred_mode, mb_x, mb_y);
997 s->hpc.pred8x8[mode](dst[1], s->uvlinesize);
998 s->hpc.pred8x8[mode](dst[2], s->uvlinesize);
9ac831c2
DC
999
1000 if (s->deblock_filter || !mb_y)
1001 xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
1002 s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
1003 s->filter.simple, 0);
3b636f21
DC
1004}
1005
1006/**
1007 * Generic MC function.
1008 *
1009 * @param s VP8 decoding context
1010 * @param luma 1 for luma (Y) planes, 0 for chroma (Cb/Cr) planes
1011 * @param dst target buffer for block data at block position
1012 * @param src reference picture buffer at origin (0, 0)
1013 * @param mv motion vector (relative to block position) to get pixel data from
1014 * @param x_off horizontal position of block from origin (0, 0)
1015 * @param y_off vertical position of block from origin (0, 0)
1016 * @param block_w width of block (16, 8 or 4)
1017 * @param block_h height of block (always same as block_w)
1018 * @param width width of src/dst plane data
1019 * @param height height of src/dst plane data
1020 * @param linesize size of a single line of plane data, including padding
e394953e 1021 * @param mc_func motion compensation function pointers (bilinear or sixtap MC)
3b636f21
DC
1022 */
1023static inline void vp8_mc(VP8Context *s, int luma,
1024 uint8_t *dst, uint8_t *src, const VP56mv *mv,
1025 int x_off, int y_off, int block_w, int block_h,
1026 int width, int height, int linesize,
d6f8476b 1027 vp8_mc_func mc_func[3][3])
3b636f21 1028{
c0498b30
JGG
1029 if (AV_RN32A(mv)) {
1030 static const uint8_t idx[8] = { 0, 1, 2, 1, 2, 1, 2, 1 };
1031 int mx = (mv->x << luma)&7, mx_idx = idx[mx];
1032 int my = (mv->y << luma)&7, my_idx = idx[my];
1033
1034 x_off += mv->x >> (3 - luma);
1035 y_off += mv->y >> (3 - luma);
1036
1037 // edge emulation
1038 src += y_off * linesize + x_off;
1039 if (x_off < 2 || x_off >= width - block_w - 3 ||
1040 y_off < 2 || y_off >= height - block_h - 3) {
1041 ff_emulated_edge_mc(s->edge_emu_buffer, src - 2 * linesize - 2, linesize,
1042 block_w + 5, block_h + 5,
1043 x_off - 2, y_off - 2, width, height);
1044 src = s->edge_emu_buffer + 2 + linesize * 2;
1045 }
1046 mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);
1047 } else
1048 mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0);
3b636f21
DC
1049}
1050
7c4dcf81
RB
1051static inline void vp8_mc_part(VP8Context *s, uint8_t *dst[3],
1052 AVFrame *ref_frame, int x_off, int y_off,
1053 int bx_off, int by_off,
1054 int block_w, int block_h,
1055 int width, int height, VP56mv *mv)
1056{
1057 VP56mv uvmv = *mv;
1058
1059 /* Y */
1060 vp8_mc(s, 1, dst[0] + by_off * s->linesize + bx_off,
1061 ref_frame->data[0], mv, x_off + bx_off, y_off + by_off,
1062 block_w, block_h, width, height, s->linesize,
1063 s->put_pixels_tab[block_w == 8]);
1064
1065 /* U/V */
1066 if (s->profile == 3) {
1067 uvmv.x &= ~7;
1068 uvmv.y &= ~7;
1069 }
1070 x_off >>= 1; y_off >>= 1;
1071 bx_off >>= 1; by_off >>= 1;
1072 width >>= 1; height >>= 1;
1073 block_w >>= 1; block_h >>= 1;
1074 vp8_mc(s, 0, dst[1] + by_off * s->uvlinesize + bx_off,
1075 ref_frame->data[1], &uvmv, x_off + bx_off, y_off + by_off,
1076 block_w, block_h, width, height, s->uvlinesize,
1077 s->put_pixels_tab[1 + (block_w == 4)]);
1078 vp8_mc(s, 0, dst[2] + by_off * s->uvlinesize + bx_off,
1079 ref_frame->data[2], &uvmv, x_off + bx_off, y_off + by_off,
1080 block_w, block_h, width, height, s->uvlinesize,
1081 s->put_pixels_tab[1 + (block_w == 4)]);
1082}
1083
d864dee8
JGG
1084/* Fetch pixels for estimated mv 4 macroblocks ahead.
1085 * Optimized for 64-byte cache lines. Inspired by ffh264 prefetch_motion. */
c4211046 1086static inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int ref)
d864dee8 1087{
c4211046
JGG
1088 /* Don't prefetch refs that haven't been used yet this frame. */
1089 if (s->ref_count[ref-1]) {
1090 int x_off = mb_x << 4, y_off = mb_y << 4;
1091 int mx = mb->mv.x + x_off + 8;
1092 int my = mb->mv.y + y_off;
1093 uint8_t **src= s->framep[ref]->data;
1094 int off= mx + (my + (mb_x&3)*4)*s->linesize + 64;
1095 s->dsp.prefetch(src[0]+off, s->linesize, 4);
1096 off= (mx>>1) + ((my>>1) + (mb_x&7))*s->uvlinesize + 64;
1097 s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1098 }
d864dee8
JGG
1099}
1100
3b636f21
DC
1101/**
1102 * Apply motion vectors to prediction buffer, chapter 18.
1103 */
1104static void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
1105 int mb_x, int mb_y)
1106{
1107 int x_off = mb_x << 4, y_off = mb_y << 4;
1108 int width = 16*s->mb_width, height = 16*s->mb_height;
d292c345
JGG
1109 AVFrame *ref = s->framep[mb->ref_frame];
1110 VP56mv *bmv = mb->bmv;
3b636f21
DC
1111
1112 if (mb->mode < VP8_MVMODE_SPLIT) {
d292c345 1113 vp8_mc_part(s, dst, ref, x_off, y_off,
7c4dcf81
RB
1114 0, 0, 16, 16, width, height, &mb->mv);
1115 } else switch (mb->partitioning) {
1116 case VP8_SPLITMVMODE_4x4: {
3b636f21 1117 int x, y;
7c4dcf81 1118 VP56mv uvmv;
3b636f21
DC
1119
1120 /* Y */
1121 for (y = 0; y < 4; y++) {
1122 for (x = 0; x < 4; x++) {
1123 vp8_mc(s, 1, dst[0] + 4*y*s->linesize + x*4,
d292c345 1124 ref->data[0], &bmv[4*y + x],
3b636f21
DC
1125 4*x + x_off, 4*y + y_off, 4, 4,
1126 width, height, s->linesize,
0ef1dbed 1127 s->put_pixels_tab[2]);
3b636f21
DC
1128 }
1129 }
1130
1131 /* U/V */
1132 x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1;
1133 for (y = 0; y < 2; y++) {
1134 for (x = 0; x < 2; x++) {
1135 uvmv.x = mb->bmv[ 2*y * 4 + 2*x ].x +
1136 mb->bmv[ 2*y * 4 + 2*x+1].x +
1137 mb->bmv[(2*y+1) * 4 + 2*x ].x +
1138 mb->bmv[(2*y+1) * 4 + 2*x+1].x;
1139 uvmv.y = mb->bmv[ 2*y * 4 + 2*x ].y +
1140 mb->bmv[ 2*y * 4 + 2*x+1].y +
1141 mb->bmv[(2*y+1) * 4 + 2*x ].y +
1142 mb->bmv[(2*y+1) * 4 + 2*x+1].y;
8f910a56
SG
1143 uvmv.x = (uvmv.x + 2 + (uvmv.x >> (INT_BIT-1))) >> 2;
1144 uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2;
3b636f21
DC
1145 if (s->profile == 3) {
1146 uvmv.x &= ~7;
1147 uvmv.y &= ~7;
1148 }
1149 vp8_mc(s, 0, dst[1] + 4*y*s->uvlinesize + x*4,
d292c345 1150 ref->data[1], &uvmv,
3b636f21
DC
1151 4*x + x_off, 4*y + y_off, 4, 4,
1152 width, height, s->uvlinesize,
0ef1dbed 1153 s->put_pixels_tab[2]);
3b636f21 1154 vp8_mc(s, 0, dst[2] + 4*y*s->uvlinesize + x*4,
d292c345 1155 ref->data[2], &uvmv,
3b636f21
DC
1156 4*x + x_off, 4*y + y_off, 4, 4,
1157 width, height, s->uvlinesize,
0ef1dbed 1158 s->put_pixels_tab[2]);
3b636f21
DC
1159 }
1160 }
7c4dcf81
RB
1161 break;
1162 }
1163 case VP8_SPLITMVMODE_16x8:
d292c345
JGG
1164 vp8_mc_part(s, dst, ref, x_off, y_off,
1165 0, 0, 16, 8, width, height, &bmv[0]);
1166 vp8_mc_part(s, dst, ref, x_off, y_off,
1167 0, 8, 16, 8, width, height, &bmv[1]);
7c4dcf81
RB
1168 break;
1169 case VP8_SPLITMVMODE_8x16:
d292c345
JGG
1170 vp8_mc_part(s, dst, ref, x_off, y_off,
1171 0, 0, 8, 16, width, height, &bmv[0]);
1172 vp8_mc_part(s, dst, ref, x_off, y_off,
1173 8, 0, 8, 16, width, height, &bmv[1]);
7c4dcf81
RB
1174 break;
1175 case VP8_SPLITMVMODE_8x8:
d292c345
JGG
1176 vp8_mc_part(s, dst, ref, x_off, y_off,
1177 0, 0, 8, 8, width, height, &bmv[0]);
1178 vp8_mc_part(s, dst, ref, x_off, y_off,
1179 8, 0, 8, 8, width, height, &bmv[1]);
1180 vp8_mc_part(s, dst, ref, x_off, y_off,
1181 0, 8, 8, 8, width, height, &bmv[2]);
1182 vp8_mc_part(s, dst, ref, x_off, y_off,
1183 8, 8, 8, 8, width, height, &bmv[3]);
7c4dcf81 1184 break;
3b636f21
DC
1185 }
1186}
1187
1188static void idct_mb(VP8Context *s, uint8_t *y_dst, uint8_t *u_dst, uint8_t *v_dst,
1189 VP8Macroblock *mb)
1190{
1191 int x, y, nnz;
1192
1193 if (mb->mode != MODE_I4x4)
1194 for (y = 0; y < 4; y++) {
1195 for (x = 0; x < 4; x++) {
1196 nnz = s->non_zero_count_cache[y][x];
1197 if (nnz) {
1198 if (nnz == 1)
1199 s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize);
1200 else
1201 s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize);
1202 }
1203 }
1204 y_dst += 4*s->linesize;
1205 }
1206
1207 for (y = 0; y < 2; y++) {
1208 for (x = 0; x < 2; x++) {
1209 nnz = s->non_zero_count_cache[4][(y<<1)+x];
1210 if (nnz) {
1211 if (nnz == 1)
1212 s->vp8dsp.vp8_idct_dc_add(u_dst+4*x, s->block[4][(y<<1)+x], s->uvlinesize);
1213 else
1214 s->vp8dsp.vp8_idct_add(u_dst+4*x, s->block[4][(y<<1)+x], s->uvlinesize);
1215 }
1216
1217 nnz = s->non_zero_count_cache[5][(y<<1)+x];
1218 if (nnz) {
1219 if (nnz == 1)
1220 s->vp8dsp.vp8_idct_dc_add(v_dst+4*x, s->block[5][(y<<1)+x], s->uvlinesize);
1221 else
1222 s->vp8dsp.vp8_idct_add(v_dst+4*x, s->block[5][(y<<1)+x], s->uvlinesize);
1223 }
1224 }
1225 u_dst += 4*s->uvlinesize;
1226 v_dst += 4*s->uvlinesize;
1227 }
1228}
1229
968570d6 1230static void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f )
3b636f21
DC
1231{
1232 int interior_limit, filter_level;
1233
1234 if (s->segmentation.enabled) {
b9a7186b 1235 filter_level = s->segmentation.filter_level[s->segment];
3b636f21
DC
1236 if (!s->segmentation.absolute_vals)
1237 filter_level += s->filter.level;
1238 } else
1239 filter_level = s->filter.level;
1240
1241 if (s->lf_delta.enabled) {
1242 filter_level += s->lf_delta.ref[mb->ref_frame];
1243
1244 if (mb->ref_frame == VP56_FRAME_CURRENT) {
1245 if (mb->mode == MODE_I4x4)
1246 filter_level += s->lf_delta.mode[0];
1247 } else {
1248 if (mb->mode == VP8_MVMODE_ZERO)
1249 filter_level += s->lf_delta.mode[1];
1250 else if (mb->mode == VP8_MVMODE_SPLIT)
1251 filter_level += s->lf_delta.mode[3];
1252 else
1253 filter_level += s->lf_delta.mode[2];
1254 }
1255 }
1256 filter_level = av_clip(filter_level, 0, 63);
1257
1258 interior_limit = filter_level;
1259 if (s->filter.sharpness) {
1260 interior_limit >>= s->filter.sharpness > 4 ? 2 : 1;
1261 interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness);
1262 }
1263 interior_limit = FFMAX(interior_limit, 1);
1264
968570d6
JGG
1265 f->filter_level = filter_level;
1266 f->inner_limit = interior_limit;
c55e0d34 1267 f->inner_filter = !mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT;
3b636f21
DC
1268}
1269
c55e0d34 1270static void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f, int mb_x, int mb_y)
3b636f21 1271{
968570d6
JGG
1272 int mbedge_lim, bedge_lim, hev_thresh;
1273 int filter_level = f->filter_level;
1274 int inner_limit = f->inner_limit;
c55e0d34 1275 int inner_filter = f->inner_filter;
3b636f21 1276
3b636f21
DC
1277 if (!filter_level)
1278 return;
1279
5245c04d
DC
1280 mbedge_lim = 2*(filter_level+2) + inner_limit;
1281 bedge_lim = 2* filter_level + inner_limit;
968570d6
JGG
1282 hev_thresh = filter_level >= 15;
1283
1284 if (s->keyframe) {
1285 if (filter_level >= 40)
1286 hev_thresh = 2;
1287 } else {
1288 if (filter_level >= 40)
1289 hev_thresh = 3;
1290 else if (filter_level >= 20)
1291 hev_thresh = 2;
1292 }
5245c04d 1293
3b636f21 1294 if (mb_x) {
3facfc99
RB
1295 s->vp8dsp.vp8_h_loop_filter16y(dst[0], s->linesize,
1296 mbedge_lim, inner_limit, hev_thresh);
1297 s->vp8dsp.vp8_h_loop_filter8uv(dst[1], dst[2], s->uvlinesize,
1298 mbedge_lim, inner_limit, hev_thresh);
3b636f21
DC
1299 }
1300
c55e0d34 1301 if (inner_filter) {
3facfc99
RB
1302 s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 4, s->linesize, bedge_lim,
1303 inner_limit, hev_thresh);
1304 s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 8, s->linesize, bedge_lim,
1305 inner_limit, hev_thresh);
1306 s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+12, s->linesize, bedge_lim,
1307 inner_limit, hev_thresh);
1308 s->vp8dsp.vp8_h_loop_filter8uv_inner(dst[1] + 4, dst[2] + 4,
1309 s->uvlinesize, bedge_lim,
1310 inner_limit, hev_thresh);
3b636f21
DC
1311 }
1312
1313 if (mb_y) {
3facfc99
RB
1314 s->vp8dsp.vp8_v_loop_filter16y(dst[0], s->linesize,
1315 mbedge_lim, inner_limit, hev_thresh);
1316 s->vp8dsp.vp8_v_loop_filter8uv(dst[1], dst[2], s->uvlinesize,
1317 mbedge_lim, inner_limit, hev_thresh);
3b636f21
DC
1318 }
1319
c55e0d34 1320 if (inner_filter) {
3facfc99
RB
1321 s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 4*s->linesize,
1322 s->linesize, bedge_lim,
1323 inner_limit, hev_thresh);
1324 s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 8*s->linesize,
1325 s->linesize, bedge_lim,
1326 inner_limit, hev_thresh);
1327 s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+12*s->linesize,
1328 s->linesize, bedge_lim,
1329 inner_limit, hev_thresh);
1330 s->vp8dsp.vp8_v_loop_filter8uv_inner(dst[1] + 4 * s->uvlinesize,
1331 dst[2] + 4 * s->uvlinesize,
1332 s->uvlinesize, bedge_lim,
1333 inner_limit, hev_thresh);
3b636f21
DC
1334 }
1335}
1336
c55e0d34 1337static void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f, int mb_x, int mb_y)
3b636f21 1338{
968570d6
JGG
1339 int mbedge_lim, bedge_lim;
1340 int filter_level = f->filter_level;
1341 int inner_limit = f->inner_limit;
c55e0d34 1342 int inner_filter = f->inner_filter;
3b636f21 1343
3b636f21
DC
1344 if (!filter_level)
1345 return;
1346
1347 mbedge_lim = 2*(filter_level+2) + inner_limit;
1348 bedge_lim = 2* filter_level + inner_limit;
1349
1350 if (mb_x)
1351 s->vp8dsp.vp8_h_loop_filter_simple(dst, s->linesize, mbedge_lim);
c55e0d34 1352 if (inner_filter) {
3b636f21
DC
1353 s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, s->linesize, bedge_lim);
1354 s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, s->linesize, bedge_lim);
1355 s->vp8dsp.vp8_h_loop_filter_simple(dst+12, s->linesize, bedge_lim);
1356 }
1357
1358 if (mb_y)
1359 s->vp8dsp.vp8_v_loop_filter_simple(dst, s->linesize, mbedge_lim);
c55e0d34 1360 if (inner_filter) {
3b636f21
DC
1361 s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*s->linesize, s->linesize, bedge_lim);
1362 s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*s->linesize, s->linesize, bedge_lim);
1363 s->vp8dsp.vp8_v_loop_filter_simple(dst+12*s->linesize, s->linesize, bedge_lim);
1364 }
1365}
1366
1367static void filter_mb_row(VP8Context *s, int mb_y)
1368{
968570d6 1369 VP8FilterStrength *f = s->filter_strength;
3b636f21
DC
1370 uint8_t *dst[3] = {
1371 s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize,
1372 s->framep[VP56_FRAME_CURRENT]->data[1] + 8*mb_y*s->uvlinesize,
1373 s->framep[VP56_FRAME_CURRENT]->data[2] + 8*mb_y*s->uvlinesize
1374 };
1375 int mb_x;
1376
1377 for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
9ac831c2 1378 backup_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], s->linesize, s->uvlinesize, 0);
c55e0d34 1379 filter_mb(s, dst, f++, mb_x, mb_y);
3b636f21
DC
1380 dst[0] += 16;
1381 dst[1] += 8;
1382 dst[2] += 8;
1383 }
1384}
1385
1386static void filter_mb_row_simple(VP8Context *s, int mb_y)
1387{
968570d6 1388 VP8FilterStrength *f = s->filter_strength;
968570d6 1389 uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize;
3b636f21
DC
1390 int mb_x;
1391
1392 for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
9ac831c2 1393 backup_mb_border(s->top_border[mb_x+1], dst, NULL, NULL, s->linesize, 0, 1);
c55e0d34 1394 filter_mb_simple(s, dst, f++, mb_x, mb_y);
3b636f21
DC
1395 dst += 16;
1396 }
1397}
1398
1399static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
1400 AVPacket *avpkt)
1401{
1402 VP8Context *s = avctx->priv_data;
1403 int ret, mb_x, mb_y, i, y, referenced;
1404 enum AVDiscard skip_thresh;
86721533 1405 AVFrame *curframe = NULL;
3b636f21
DC
1406
1407 if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0)
1408 return ret;
1409
1410 referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT
1411 || s->update_altref == VP56_FRAME_CURRENT;
1412
1413 skip_thresh = !referenced ? AVDISCARD_NONREF :
1414 !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL;
1415
1416 if (avctx->skip_frame >= skip_thresh) {
1417 s->invisible = 1;
1418 goto skip_decode;
1419 }
9ac831c2 1420 s->deblock_filter = s->filter.level && avctx->skip_loop_filter < skip_thresh;
3b636f21
DC
1421
1422 for (i = 0; i < 4; i++)
1423 if (&s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1424 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1425 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) {
1426 curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i];
1427 break;
1428 }
1429 if (curframe->data[0])
1430 avctx->release_buffer(avctx, curframe);
1431
1432 curframe->key_frame = s->keyframe;
1433 curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE;
1434 curframe->reference = referenced ? 3 : 0;
1435 if ((ret = avctx->get_buffer(avctx, curframe))) {
1436 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n");
1437 return ret;
1438 }
1439
1440 // Given that arithmetic probabilities are updated every frame, it's quite likely
1441 // that the values we have on a random interframe are complete junk if we didn't
1442 // start decode on a keyframe. So just don't display anything rather than junk.
1443 if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] ||
1444 !s->framep[VP56_FRAME_GOLDEN] ||
1445 !s->framep[VP56_FRAME_GOLDEN2])) {
1446 av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n");
1447 return AVERROR_INVALIDDATA;
1448 }
1449
1450 s->linesize = curframe->linesize[0];
1451 s->uvlinesize = curframe->linesize[1];
1452
1453 if (!s->edge_emu_buffer)
1454 s->edge_emu_buffer = av_malloc(21*s->linesize);
1455
1456 memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz));
1457
c55e0d34
JGG
1458 /* Zero macroblock structures for top/left prediction from outside the frame. */
1459 memset(s->macroblocks, 0, (s->mb_width + s->mb_height*2)*sizeof(*s->macroblocks));
1460
3b636f21 1461 // top edge of 127 for intra prediction
9ac831c2 1462 memset(s->top_border, 127, (s->mb_width+1)*sizeof(*s->top_border));
c4211046 1463 memset(s->ref_count, 0, sizeof(s->ref_count));
3b636f21
DC
1464
1465 for (mb_y = 0; mb_y < s->mb_height; mb_y++) {
1466 VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)];
c55e0d34 1467 VP8Macroblock *mb = s->macroblocks + (s->mb_height - mb_y - 1)*2;
3b636f21 1468 uint8_t *intra4x4 = s->intra4x4_pred_mode + 4*mb_y*s->b4_stride;
c55e0d34 1469 uint8_t *segment_map = s->segmentation_map + mb_y*s->mb_stride;
3b636f21
DC
1470 uint8_t *dst[3] = {
1471 curframe->data[0] + 16*mb_y*s->linesize,
1472 curframe->data[1] + 8*mb_y*s->uvlinesize,
1473 curframe->data[2] + 8*mb_y*s->uvlinesize
1474 };
1475
1476 memset(s->left_nnz, 0, sizeof(s->left_nnz));
1477
1478 // left edge of 129 for intra prediction
1479 if (!(avctx->flags & CODEC_FLAG_EMU_EDGE))
1480 for (i = 0; i < 3; i++)
1481 for (y = 0; y < 16>>!!i; y++)
1482 dst[i][y*curframe->linesize[i]-1] = 129;
9ac831c2
DC
1483 if (mb_y)
1484 memset(s->top_border, 129, sizeof(*s->top_border));
3b636f21
DC
1485
1486 for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
d1c58fce 1487 uint8_t *intra4x4_mb = s->keyframe ? intra4x4 + 4*mb_x : s->intra4x4_pred_mode_mb;
c55e0d34 1488 uint8_t *segment_mb = segment_map+mb_x;
d1c58fce 1489
d864dee8
JGG
1490 /* Prefetch the current frame, 4 MBs ahead */
1491 s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4);
1492 s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2);
1493
c55e0d34 1494 decode_mb_mode(s, mb, mb_x, mb_y, intra4x4_mb, segment_mb);
3b636f21 1495
c4211046
JGG
1496 prefetch_motion(s, mb, mb_x, mb_y, VP56_FRAME_PREVIOUS);
1497
3b636f21
DC
1498 if (!mb->skip)
1499 decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);
1500 else {
1501 AV_ZERO128(s->non_zero_count_cache); // luma
1502 AV_ZERO64(s->non_zero_count_cache[4]); // chroma
1503 }
1504
b946111f 1505 if (mb->mode <= MODE_I4x4)
d1c58fce 1506 intra_predict(s, dst, mb, intra4x4_mb, mb_x, mb_y);
b946111f 1507 else
3b636f21 1508 inter_predict(s, dst, mb, mb_x, mb_y);
3b636f21 1509
c4211046
JGG
1510 prefetch_motion(s, mb, mb_x, mb_y, VP56_FRAME_GOLDEN);
1511
3b636f21
DC
1512 if (!mb->skip) {
1513 idct_mb(s, dst[0], dst[1], dst[2], mb);
1514 } else {
1515 AV_ZERO64(s->left_nnz);
1516 AV_WN64(s->top_nnz[mb_x], 0); // array of 9, so unaligned
1517
1518 // Reset DC block predictors if they would exist if the mb had coefficients
1519 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
1520 s->left_nnz[8] = 0;
1521 s->top_nnz[mb_x][8] = 0;
1522 }
1523 }
1524
968570d6
JGG
1525 if (s->deblock_filter)
1526 filter_level_for_mb(s, mb, &s->filter_strength[mb_x]);
1527
c4211046
JGG
1528 prefetch_motion(s, mb, mb_x, mb_y, VP56_FRAME_GOLDEN2);
1529
3b636f21
DC
1530 dst[0] += 16;
1531 dst[1] += 8;
1532 dst[2] += 8;
1533 mb++;
1534 }
9ac831c2 1535 if (s->deblock_filter) {
3b636f21 1536 if (s->filter.simple)
9ac831c2 1537 filter_mb_row_simple(s, mb_y);
3b636f21 1538 else
9ac831c2 1539 filter_mb_row(s, mb_y);
3b636f21
DC
1540 }
1541 }
3b636f21
DC
1542
1543skip_decode:
1544 // if future frames don't use the updated probabilities,
1545 // reset them to the values we saved
1546 if (!s->update_probabilities)
1547 s->prob[0] = s->prob[1];
1548
1549 // check if golden and altref are swapped
1550 if (s->update_altref == VP56_FRAME_GOLDEN &&
1551 s->update_golden == VP56_FRAME_GOLDEN2)
1552 FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN], s->framep[VP56_FRAME_GOLDEN2]);
1553 else {
1554 if (s->update_altref != VP56_FRAME_NONE)
1555 s->framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref];
1556
1557 if (s->update_golden != VP56_FRAME_NONE)
1558 s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden];
1559 }
1560
1561 if (s->update_last) // move cur->prev
1562 s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT];
1563
1564 // release no longer referenced frames
1565 for (i = 0; i < 4; i++)
1566 if (s->frames[i].data[0] &&
1567 &s->frames[i] != s->framep[VP56_FRAME_CURRENT] &&
1568 &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1569 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1570 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2])
1571 avctx->release_buffer(avctx, &s->frames[i]);
1572
1573 if (!s->invisible) {
1574 *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT];
1575 *data_size = sizeof(AVFrame);
1576 }
1577
1578 return avpkt->size;
1579}
1580
1581static av_cold int vp8_decode_init(AVCodecContext *avctx)
1582{
1583 VP8Context *s = avctx->priv_data;
1584
1585 s->avctx = avctx;
1586 avctx->pix_fmt = PIX_FMT_YUV420P;
1587
1588 dsputil_init(&s->dsp, avctx);
1589 ff_h264_pred_init(&s->hpc, CODEC_ID_VP8);
1590 ff_vp8dsp_init(&s->vp8dsp);
1591
1592 // intra pred needs edge emulation among other things
1593 if (avctx->flags&CODEC_FLAG_EMU_EDGE) {
03ac56e7 1594 av_log(avctx, AV_LOG_ERROR, "Edge emulation not supported\n");
3b636f21
DC
1595 return AVERROR_PATCHWELCOME;
1596 }
1597
1598 return 0;
1599}
1600
1601static av_cold int vp8_decode_free(AVCodecContext *avctx)
1602{
1603 vp8_decode_flush(avctx);
1604 return 0;
1605}
1606
1607AVCodec vp8_decoder = {
1608 "vp8",
1609 AVMEDIA_TYPE_VIDEO,
1610 CODEC_ID_VP8,
1611 sizeof(VP8Context),
1612 vp8_decode_init,
1613 NULL,
1614 vp8_decode_free,
1615 vp8_decode_frame,
1616 CODEC_CAP_DR1,
1617 .flush = vp8_decode_flush,
1618 .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"),
1619};