prevent access to vp8_coeff_band[16]
[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
13a1304b 6 * Copyright (C) 2010 Jason Garrett-Glaser
3b636f21
DC
7 *
8 * This file is part of FFmpeg.
9 *
10 * FFmpeg is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public
12 * License as published by the Free Software Foundation; either
13 * version 2.1 of the License, or (at your option) any later version.
14 *
15 * FFmpeg is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
19 *
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with FFmpeg; if not, write to the Free Software
22 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 */
24
25#include "avcodec.h"
26#include "vp56.h"
27#include "vp8data.h"
28#include "vp8dsp.h"
29#include "h264pred.h"
30#include "rectangle.h"
31
32typedef struct {
968570d6
JGG
33 uint8_t filter_level;
34 uint8_t inner_limit;
c55e0d34 35 uint8_t inner_filter;
968570d6
JGG
36} VP8FilterStrength;
37
38typedef struct {
3b636f21
DC
39 uint8_t skip;
40 // todo: make it possible to check for at least (i4x4 or split_mv)
41 // in one op. are others needed?
42 uint8_t mode;
43 uint8_t ref_frame;
44 uint8_t partitioning;
45 VP56mv mv;
46 VP56mv bmv[16];
47} VP8Macroblock;
48
49typedef struct {
50 AVCodecContext *avctx;
51 DSPContext dsp;
52 VP8DSPContext vp8dsp;
53 H264PredContext hpc;
0ef1dbed 54 vp8_mc_func put_pixels_tab[3][3][3];
3b636f21
DC
55 AVFrame frames[4];
56 AVFrame *framep[4];
57 uint8_t *edge_emu_buffer;
58 VP56RangeCoder c; ///< header context, includes mb modes and motion vectors
59 int profile;
60
61 int mb_width; /* number of horizontal MB */
62 int mb_height; /* number of vertical MB */
63 int linesize;
64 int uvlinesize;
65
66 int keyframe;
67 int invisible;
68 int update_last; ///< update VP56_FRAME_PREVIOUS with the current one
69 int update_golden; ///< VP56_FRAME_NONE if not updated, or which frame to copy if so
70 int update_altref;
9ac831c2 71 int deblock_filter;
3b636f21
DC
72
73 /**
74 * If this flag is not set, all the probability updates
75 * are discarded after this frame is decoded.
76 */
77 int update_probabilities;
78
79 /**
80 * All coefficients are contained in separate arith coding contexts.
81 * There can be 1, 2, 4, or 8 of these after the header context.
82 */
83 int num_coeff_partitions;
84 VP56RangeCoder coeff_partition[8];
85
86 VP8Macroblock *macroblocks;
87 VP8Macroblock *macroblocks_base;
968570d6 88 VP8FilterStrength *filter_strength;
3b636f21
DC
89 int mb_stride;
90
91 uint8_t *intra4x4_pred_mode;
92 uint8_t *intra4x4_pred_mode_base;
c55e0d34 93 uint8_t *segmentation_map;
3b636f21
DC
94 int b4_stride;
95
96 /**
9ac831c2
DC
97 * Cache of the top row needed for intra prediction
98 * 16 for luma, 8 for each chroma plane
99 */
100 uint8_t (*top_border)[16+8+8];
101
102 /**
3b636f21
DC
103 * For coeff decode, we need to know whether the above block had non-zero
104 * coefficients. This means for each macroblock, we need data for 4 luma
105 * blocks, 2 u blocks, 2 v blocks, and the luma dc block, for a total of 9
106 * per macroblock. We keep the last row in top_nnz.
107 */
108 uint8_t (*top_nnz)[9];
109 DECLARE_ALIGNED(8, uint8_t, left_nnz)[9];
110
111 /**
112 * This is the index plus one of the last non-zero coeff
113 * for each of the blocks in the current macroblock.
114 * So, 0 -> no coeffs
115 * 1 -> dc-only (special transform)
116 * 2+-> full transform
117 */
118 DECLARE_ALIGNED(16, uint8_t, non_zero_count_cache)[6][4];
119 DECLARE_ALIGNED(16, DCTELEM, block)[6][4][16];
d1c58fce 120 uint8_t intra4x4_pred_mode_mb[16];
3b636f21
DC
121
122 int chroma_pred_mode; ///< 8x8c pred mode of the current macroblock
b9a7186b 123 int segment; ///< segment of the current macroblock
3b636f21
DC
124
125 int mbskip_enabled;
126 int sign_bias[4]; ///< one state [0, 1] per ref frame type
c4211046 127 int ref_count[3];
3b636f21
DC
128
129 /**
130 * Base parameters for segmentation, i.e. per-macroblock parameters.
131 * These must be kept unchanged even if segmentation is not used for
132 * a frame, since the values persist between interframes.
133 */
134 struct {
135 int enabled;
136 int absolute_vals;
137 int update_map;
138 int8_t base_quant[4];
139 int8_t filter_level[4]; ///< base loop filter level
140 } segmentation;
141
142 /**
143 * Macroblocks can have one of 4 different quants in a frame when
144 * segmentation is enabled.
145 * If segmentation is disabled, only the first segment's values are used.
146 */
147 struct {
148 // [0] - DC qmul [1] - AC qmul
149 int16_t luma_qmul[2];
150 int16_t luma_dc_qmul[2]; ///< luma dc-only block quant
151 int16_t chroma_qmul[2];
152 } qmat[4];
153
154 struct {
155 int simple;
156 int level;
157 int sharpness;
158 } filter;
159
160 struct {
161 int enabled; ///< whether each mb can have a different strength based on mode/ref
162
163 /**
164 * filter strength adjustment for the following macroblock modes:
165 * [0] - i4x4
166 * [1] - zero mv
167 * [2] - inter modes except for zero or split mv
168 * [3] - split mv
169 * i16x16 modes never have any adjustment
170 */
171 int8_t mode[4];
172
173 /**
174 * filter strength adjustment for macroblocks that reference:
175 * [0] - intra / VP56_FRAME_CURRENT
176 * [1] - VP56_FRAME_PREVIOUS
177 * [2] - VP56_FRAME_GOLDEN
178 * [3] - altref / VP56_FRAME_GOLDEN2
179 */
180 int8_t ref[4];
181 } lf_delta;
182
183 /**
184 * These are all of the updatable probabilities for binary decisions.
185 * They are only implictly reset on keyframes, making it quite likely
186 * for an interframe to desync if a prior frame's header was corrupt
187 * or missing outright!
188 */
189 struct {
190 uint8_t segmentid[3];
a8ab0ccc
PM
191 uint8_t mbskip;
192 uint8_t intra;
193 uint8_t last;
194 uint8_t golden;
3b636f21
DC
195 uint8_t pred16x16[4];
196 uint8_t pred8x8c[3];
197 uint8_t token[4][8][3][NUM_DCT_TOKENS-1];
198 uint8_t mvc[2][19];
199 } prob[2];
200} VP8Context;
201
3b636f21
DC
202static void vp8_decode_flush(AVCodecContext *avctx)
203{
204 VP8Context *s = avctx->priv_data;
205 int i;
206
207 for (i = 0; i < 4; i++)
208 if (s->frames[i].data[0])
209 avctx->release_buffer(avctx, &s->frames[i]);
210 memset(s->framep, 0, sizeof(s->framep));
211
212 av_freep(&s->macroblocks_base);
fca05ea8 213 av_freep(&s->filter_strength);
3b636f21
DC
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++) {
06d50ca8 310 int size = AV_RL24(sizes + 3*i);
3b636f21
DC
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)];
a8ab0ccc
PM
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)];
3b636f21 348 s->qmat[i].chroma_qmul[1] = vp8_ac_qlookup[av_clip(base_qi + uvac_delta, 0, 127)];
a8ab0ccc
PM
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);
3b636f21
DC
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);
06d50ca8 405 header_size = AV_RL24(buf) >> 5;
3b636f21
DC
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) {
06d50ca8
JGG
423 if (AV_RL24(buf) != 0x2a019d) {
424 av_log(s->avctx, AV_LOG_ERROR, "Invalid start code 0x%x\n", AV_RL24(buf));
3b636f21
DC
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++)
7697cdcf 498 if (vp56_rac_get_prob_branchy(c, vp8_token_update_probs[i][j][k][l]))
3b636f21
DC
499 s->prob->token[i][j][k][l] = vp8_rac_get_uint(c, 8);
500
501 if ((s->mbskip_enabled = vp8_rac_get(c)))
a8ab0ccc 502 s->prob->mbskip = vp8_rac_get_uint(c, 8);
3b636f21
DC
503
504 if (!s->keyframe) {
a8ab0ccc
PM
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);
3b636f21
DC
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++)
7697cdcf 519 if (vp56_rac_get_prob_branchy(c, vp8_mv_update_prob[i][j]))
3b636f21
DC
520 s->prob->mvc[i][j] = vp8_rac_get_nn(c);
521 }
522
523 return 0;
524}
525
414ac27d
JGG
526static av_always_inline
527void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src, int mb_x, int mb_y)
3b636f21
DC
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
414ac27d
JGG
536static av_always_inline
537void find_near_mvs(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
538 VP56mv near[2], VP56mv *best, uint8_t cnt[4])
3b636f21 539{
c55e0d34
JGG
540 VP8Macroblock *mb_edge[3] = { mb + 2 /* top */,
541 mb - 1 /* left */,
542 mb + 1 /* top-left */ };
3b636f21
DC
543 enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT };
544 VP56mv near_mv[4] = {{ 0 }};
545 enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV };
702e8d33 546 int idx = CNT_ZERO;
3b636f21 547 int best_idx = CNT_ZERO;
702e8d33
JGG
548 int cur_sign_bias = s->sign_bias[mb->ref_frame];
549 int *sign_bias = s->sign_bias;
3b636f21
DC
550
551 /* Process MB on top, left and top-left */
702e8d33
JGG
552 #define MV_EDGE_CHECK(n)\
553 {\
554 VP8Macroblock *edge = mb_edge[n];\
555 int edge_ref = edge->ref_frame;\
556 if (edge_ref != VP56_FRAME_CURRENT) {\
557 uint32_t mv = AV_RN32A(&edge->mv);\
558 if (mv) {\
559 if (cur_sign_bias != sign_bias[edge_ref]) {\
560 /* SWAR negate of the values in mv. */\
0087aa47
JGG
561 mv = ~mv;\
562 mv = ((mv&0x7fff7fff) + 0x00010001) ^ (mv&0x80008000);\
702e8d33
JGG
563 }\
564 if (!n || mv != AV_RN32A(&near_mv[idx]))\
565 AV_WN32A(&near_mv[++idx], mv);\
566 cnt[idx] += 1 + (n != 2);\
567 } else\
568 cnt[CNT_ZERO] += 1 + (n != 2);\
569 }\
3b636f21 570 }
702e8d33
JGG
571 MV_EDGE_CHECK(0)
572 MV_EDGE_CHECK(1)
573 MV_EDGE_CHECK(2)
3b636f21 574
702e8d33
JGG
575 /* If we have three distinct MVs, merge first and last if they're the same */
576 if (cnt[CNT_SPLITMV] && AV_RN32A(&near_mv[1+EDGE_TOP]) == AV_RN32A(&near_mv[1+EDGE_TOPLEFT]))
3b636f21
DC
577 cnt[CNT_NEAREST] += 1;
578
579 cnt[CNT_SPLITMV] = ((mb_edge[EDGE_LEFT]->mode == VP8_MVMODE_SPLIT) +
580 (mb_edge[EDGE_TOP]->mode == VP8_MVMODE_SPLIT)) * 2 +
581 (mb_edge[EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT);
582
583 /* Swap near and nearest if necessary */
584 if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) {
702e8d33
JGG
585 FFSWAP(uint8_t, cnt[CNT_NEAREST], cnt[CNT_NEAR]);
586 FFSWAP( VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]);
3b636f21
DC
587 }
588
589 /* Choose the best mv out of 0,0 and the nearest mv */
590 if (cnt[CNT_NEAREST] >= cnt[CNT_ZERO])
591 best_idx = CNT_NEAREST;
592
9fddd14a 593 mb->mv = near_mv[best_idx];
3b636f21
DC
594 near[0] = near_mv[CNT_NEAREST];
595 near[1] = near_mv[CNT_NEAR];
596}
597
598/**
599 * Motion vector coding, 17.1.
600 */
601static int read_mv_component(VP56RangeCoder *c, const uint8_t *p)
602{
ca18a478 603 int bit, x = 0;
3b636f21 604
7697cdcf 605 if (vp56_rac_get_prob_branchy(c, p[0])) {
3b636f21
DC
606 int i;
607
608 for (i = 0; i < 3; i++)
609 x += vp56_rac_get_prob(c, p[9 + i]) << i;
610 for (i = 9; i > 3; i--)
611 x += vp56_rac_get_prob(c, p[9 + i]) << i;
612 if (!(x & 0xFFF0) || vp56_rac_get_prob(c, p[12]))
613 x += 8;
ca18a478
DC
614 } else {
615 // small_mvtree
616 const uint8_t *ps = p+2;
617 bit = vp56_rac_get_prob(c, *ps);
618 ps += 1 + 3*bit;
619 x += 4*bit;
620 bit = vp56_rac_get_prob(c, *ps);
621 ps += 1 + bit;
622 x += 2*bit;
623 x += vp56_rac_get_prob(c, *ps);
624 }
3b636f21
DC
625
626 return (x && vp56_rac_get_prob(c, p[1])) ? -x : x;
627}
628
414ac27d
JGG
629static av_always_inline
630const uint8_t *get_submv_prob(uint32_t left, uint32_t top)
3b636f21 631{
7bf254c4
JGG
632 if (left == top)
633 return vp8_submv_prob[4-!!left];
634 if (!top)
3b636f21 635 return vp8_submv_prob[2];
7bf254c4 636 return vp8_submv_prob[1-!!left];
3b636f21
DC
637}
638
639/**
640 * Split motion vector prediction, 16.4.
7ed06b2b 641 * @returns the number of motion vectors parsed (2, 4 or 16)
3b636f21 642 */
414ac27d
JGG
643static av_always_inline
644int decode_splitmvs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb)
3b636f21
DC
645{
646 int part_idx = mb->partitioning =
647 vp8_rac_get_tree(c, vp8_mbsplit_tree, vp8_mbsplit_prob);
648 int n, num = vp8_mbsplit_count[part_idx];
c55e0d34 649 VP8Macroblock *top_mb = &mb[2];
7bf254c4
JGG
650 VP8Macroblock *left_mb = &mb[-1];
651 const uint8_t *mbsplits_left = vp8_mbsplits[left_mb->partitioning],
652 *mbsplits_top = vp8_mbsplits[top_mb->partitioning],
653 *mbsplits_cur = vp8_mbsplits[part_idx],
7ed06b2b 654 *firstidx = vp8_mbfirstidx[part_idx];
c55e0d34
JGG
655 VP56mv *top_mv = top_mb->bmv;
656 VP56mv *left_mv = left_mb->bmv;
657 VP56mv *cur_mv = mb->bmv;
3b636f21
DC
658
659 for (n = 0; n < num; n++) {
7ed06b2b 660 int k = firstidx[n];
7bf254c4 661 uint32_t left, above;
7ed06b2b
RB
662 const uint8_t *submv_prob;
663
7bf254c4
JGG
664 if (!(k & 3))
665 left = AV_RN32A(&left_mv[mbsplits_left[k + 3]]);
666 else
667 left = AV_RN32A(&cur_mv[mbsplits_cur[k - 1]]);
668 if (k <= 3)
669 above = AV_RN32A(&top_mv[mbsplits_top[k + 12]]);
670 else
671 above = AV_RN32A(&cur_mv[mbsplits_cur[k - 4]]);
7ed06b2b
RB
672
673 submv_prob = get_submv_prob(left, above);
3b636f21
DC
674
675 switch (vp8_rac_get_tree(c, vp8_submv_ref_tree, submv_prob)) {
676 case VP8_SUBMVMODE_NEW4X4:
9fddd14a
JGG
677 mb->bmv[n].y = mb->mv.y + read_mv_component(c, s->prob->mvc[0]);
678 mb->bmv[n].x = mb->mv.x + read_mv_component(c, s->prob->mvc[1]);
3b636f21
DC
679 break;
680 case VP8_SUBMVMODE_ZERO4X4:
14767f35 681 AV_ZERO32(&mb->bmv[n]);
3b636f21
DC
682 break;
683 case VP8_SUBMVMODE_LEFT4X4:
7bf254c4 684 AV_WN32A(&mb->bmv[n], left);
3b636f21
DC
685 break;
686 case VP8_SUBMVMODE_TOP4X4:
7bf254c4 687 AV_WN32A(&mb->bmv[n], above);
3b636f21
DC
688 break;
689 }
3b636f21 690 }
7ed06b2b
RB
691
692 return num;
3b636f21
DC
693}
694
414ac27d
JGG
695static av_always_inline
696void decode_intra4x4_modes(VP56RangeCoder *c, uint8_t *intra4x4,
697 int stride, int keyframe)
3b636f21 698{
d1c58fce 699 int x, y, t, l, i;
3b636f21 700
d1c58fce
JGG
701 if (keyframe) {
702 const uint8_t *ctx;
703 for (y = 0; y < 4; y++) {
704 for (x = 0; x < 4; x++) {
3b636f21
DC
705 t = intra4x4[x - stride];
706 l = intra4x4[x - 1];
707 ctx = vp8_pred4x4_prob_intra[t][l];
d1c58fce 708 intra4x4[x] = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx);
3b636f21 709 }
d1c58fce 710 intra4x4 += stride;
3b636f21 711 }
d1c58fce
JGG
712 } else {
713 for (i = 0; i < 16; i++)
714 intra4x4[i] = vp8_rac_get_tree(c, vp8_pred4x4_tree, vp8_pred4x4_prob_inter);
3b636f21
DC
715 }
716}
717
414ac27d
JGG
718static av_always_inline
719void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
720 uint8_t *intra4x4, uint8_t *segment)
3b636f21
DC
721{
722 VP56RangeCoder *c = &s->c;
3b636f21
DC
723
724 if (s->segmentation.update_map)
c55e0d34 725 *segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid);
b9a7186b 726 s->segment = *segment;
3b636f21 727
a8ab0ccc 728 mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0;
3b636f21
DC
729
730 if (s->keyframe) {
731 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra);
732
733 if (mb->mode == MODE_I4x4) {
734 decode_intra4x4_modes(c, intra4x4, s->b4_stride, 1);
735 } else
736 fill_rectangle(intra4x4, 4, 4, s->b4_stride, vp8_pred4x4_mode[mb->mode], 1);
737
738 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra);
739 mb->ref_frame = VP56_FRAME_CURRENT;
a8ab0ccc 740 } else if (vp56_rac_get_prob_branchy(c, s->prob->intra)) {
3b636f21 741 VP56mv near[2], best;
702e8d33 742 uint8_t cnt[4] = { 0 };
3b636f21
DC
743 uint8_t p[4];
744
745 // inter MB, 16.2
a8ab0ccc
PM
746 if (vp56_rac_get_prob_branchy(c, s->prob->last))
747 mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ?
3b636f21
DC
748 VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN;
749 else
750 mb->ref_frame = VP56_FRAME_PREVIOUS;
c4211046 751 s->ref_count[mb->ref_frame-1]++;
3b636f21
DC
752
753 // motion vectors, 16.3
754 find_near_mvs(s, mb, mb_x, mb_y, near, &best, cnt);
702e8d33
JGG
755 p[0] = vp8_mode_contexts[cnt[0]][0];
756 p[1] = vp8_mode_contexts[cnt[1]][1];
757 p[2] = vp8_mode_contexts[cnt[2]][2];
758 p[3] = vp8_mode_contexts[cnt[3]][3];
3b636f21
DC
759 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_mvinter, p);
760 switch (mb->mode) {
761 case VP8_MVMODE_SPLIT:
9fddd14a
JGG
762 clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y);
763 mb->mv = mb->bmv[decode_splitmvs(s, c, mb) - 1];
3b636f21
DC
764 break;
765 case VP8_MVMODE_ZERO:
14767f35 766 AV_ZERO32(&mb->mv);
3b636f21
DC
767 break;
768 case VP8_MVMODE_NEAREST:
769 clamp_mv(s, &mb->mv, &near[0], mb_x, mb_y);
770 break;
771 case VP8_MVMODE_NEAR:
772 clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y);
773 break;
774 case VP8_MVMODE_NEW:
9fddd14a
JGG
775 clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y);
776 mb->mv.y += + read_mv_component(c, s->prob->mvc[0]);
777 mb->mv.x += + read_mv_component(c, s->prob->mvc[1]);
3b636f21
DC
778 break;
779 }
780 if (mb->mode != VP8_MVMODE_SPLIT) {
7ed06b2b
RB
781 mb->partitioning = VP8_SPLITMVMODE_NONE;
782 mb->bmv[0] = mb->mv;
3b636f21
DC
783 }
784 } else {
785 // intra MB, 16.1
786 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16);
787
158e062c 788 if (mb->mode == MODE_I4x4)
d1c58fce 789 decode_intra4x4_modes(c, intra4x4, 4, 0);
3b636f21
DC
790
791 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c);
792 mb->ref_frame = VP56_FRAME_CURRENT;
b946111f 793 mb->partitioning = VP8_SPLITMVMODE_NONE;
14767f35 794 AV_ZERO32(&mb->bmv[0]);
3b636f21
DC
795 }
796}
797
798/**
e394953e
RB
799 * @param c arithmetic bitstream reader context
800 * @param block destination for block coefficients
801 * @param probs probabilities to use when reading trees from the bitstream
3b636f21
DC
802 * @param i initial coeff index, 0 unless a separate DC block is coded
803 * @param zero_nhood the initial prediction context for number of surrounding
804 * all-zero blocks (only left/top, so 0-2)
3fa76268 805 * @param qmul array holding the dc/ac dequant factor at position 0/1
3b636f21
DC
806 * @return 0 if no coeffs were decoded
807 * otherwise, the index of the last coeff decoded plus one
808 */
809static int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16],
810 uint8_t probs[8][3][NUM_DCT_TOKENS-1],
811 int i, int zero_nhood, int16_t qmul[2])
812{
fe1b5d97
DC
813 uint8_t *token_prob;
814 int nonzero = 0;
815 int coeff;
3b636f21 816
fe1b5d97
DC
817 do {
818 token_prob = probs[vp8_coeff_band[i]][zero_nhood];
3b636f21 819
fe1b5d97
DC
820 if (!vp56_rac_get_prob_branchy(c, token_prob[0])) // DCT_EOB
821 return nonzero;
3b636f21 822
fe1b5d97
DC
823skip_eob:
824 if (!vp56_rac_get_prob_branchy(c, token_prob[1])) { // DCT_0
c22b4468
PM
825 if (++i == 16)
826 return nonzero; // invalid input; blocks should end with EOB
3b636f21 827 zero_nhood = 0;
c22b4468
PM
828 token_prob = probs[vp8_coeff_band[i]][0];
829 goto skip_eob;
fe1b5d97
DC
830 }
831
832 if (!vp56_rac_get_prob_branchy(c, token_prob[2])) { // DCT_1
833 coeff = 1;
3b636f21 834 zero_nhood = 1;
fe1b5d97 835 } else {
3b636f21
DC
836 zero_nhood = 2;
837
fe1b5d97
DC
838 if (!vp56_rac_get_prob_branchy(c, token_prob[3])) { // DCT 2,3,4
839 coeff = vp56_rac_get_prob(c, token_prob[4]);
840 if (coeff)
841 coeff += vp56_rac_get_prob(c, token_prob[5]);
842 coeff += 2;
843 } else {
844 // DCT_CAT*
845 if (!vp56_rac_get_prob_branchy(c, token_prob[6])) {
846 if (!vp56_rac_get_prob_branchy(c, token_prob[7])) { // DCT_CAT1
847 coeff = 5 + vp56_rac_get_prob(c, vp8_dct_cat1_prob[0]);
848 } else { // DCT_CAT2
849 coeff = 7;
850 coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[0]) << 1;
851 coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[1]);
852 }
853 } else { // DCT_CAT3 and up
854 int a = vp56_rac_get_prob(c, token_prob[8]);
855 int b = vp56_rac_get_prob(c, token_prob[9+a]);
856 int cat = (a<<1) + b;
857 coeff = 3 + (8<<cat);
858 coeff += vp8_rac_get_coeff(c, vp8_dct_cat_prob[cat]);
859 }
860 }
861 }
862
3b636f21 863 // todo: full [16] qmat? load into register?
fe1b5d97
DC
864 block[zigzag_scan[i]] = (vp8_rac_get(c) ? -coeff : coeff) * qmul[!!i];
865 nonzero = ++i;
866 } while (i < 16);
867
3b636f21
DC
868 return nonzero;
869}
870
414ac27d
JGG
871static av_always_inline
872void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,
873 uint8_t t_nnz[9], uint8_t l_nnz[9])
3b636f21
DC
874{
875 LOCAL_ALIGNED_16(DCTELEM, dc,[16]);
876 int i, x, y, luma_start = 0, luma_ctx = 3;
877 int nnz_pred, nnz, nnz_total = 0;
b9a7186b 878 int segment = s->segment;
3b636f21 879
3b636f21
DC
880 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
881 AV_ZERO128(dc);
882 AV_ZERO128(dc+8);
883 nnz_pred = t_nnz[8] + l_nnz[8];
884
885 // decode DC values and do hadamard
886 nnz = decode_block_coeffs(c, dc, s->prob->token[1], 0, nnz_pred,
887 s->qmat[segment].luma_dc_qmul);
888 l_nnz[8] = t_nnz[8] = !!nnz;
889 nnz_total += nnz;
890 s->vp8dsp.vp8_luma_dc_wht(s->block, dc);
891 luma_start = 1;
892 luma_ctx = 0;
893 }
894
895 // luma blocks
896 for (y = 0; y < 4; y++)
897 for (x = 0; x < 4; x++) {
898 nnz_pred = l_nnz[y] + t_nnz[x];
899 nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start,
900 nnz_pred, s->qmat[segment].luma_qmul);
901 // nnz+luma_start may be one more than the actual last index, but we don't care
902 s->non_zero_count_cache[y][x] = nnz + luma_start;
903 t_nnz[x] = l_nnz[y] = !!nnz;
904 nnz_total += nnz;
905 }
906
907 // chroma blocks
908 // TODO: what to do about dimensions? 2nd dim for luma is x,
909 // but for chroma it's (y<<1)|x
910 for (i = 4; i < 6; i++)
911 for (y = 0; y < 2; y++)
912 for (x = 0; x < 2; x++) {
913 nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x];
914 nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0,
915 nnz_pred, s->qmat[segment].chroma_qmul);
916 s->non_zero_count_cache[i][(y<<1)+x] = nnz;
917 t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz;
918 nnz_total += nnz;
919 }
920
921 // if there were no coded coeffs despite the macroblock not being marked skip,
922 // we MUST not do the inner loop filter and should not do IDCT
923 // Since skip isn't used for bitstream prediction, just manually set it.
924 if (!nnz_total)
925 mb->skip = 1;
926}
927
9ac831c2
DC
928static av_always_inline
929void backup_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
930 int linesize, int uvlinesize, int simple)
931{
932 AV_COPY128(top_border, src_y + 15*linesize);
933 if (!simple) {
934 AV_COPY64(top_border+16, src_cb + 7*uvlinesize);
935 AV_COPY64(top_border+24, src_cr + 7*uvlinesize);
936 }
937}
938
939static av_always_inline
940void xchg_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
941 int linesize, int uvlinesize, int mb_x, int mb_y, int mb_width,
942 int simple, int xchg)
943{
944 uint8_t *top_border_m1 = top_border-32; // for TL prediction
945 src_y -= linesize;
946 src_cb -= uvlinesize;
947 src_cr -= uvlinesize;
948
096971e8
MR
949#define XCHG(a,b,xchg) do { \
950 if (xchg) AV_SWAP64(b,a); \
951 else AV_COPY64(b,a); \
952 } while (0)
9ac831c2
DC
953
954 XCHG(top_border_m1+8, src_y-8, xchg);
955 XCHG(top_border, src_y, xchg);
956 XCHG(top_border+8, src_y+8, 1);
070ce7ef 957 if (mb_x < mb_width-1)
9ac831c2 958 XCHG(top_border+32, src_y+16, 1);
070ce7ef 959
9ac831c2
DC
960 // only copy chroma for normal loop filter
961 // or to initialize the top row to 127
962 if (!simple || !mb_y) {
963 XCHG(top_border_m1+16, src_cb-8, xchg);
964 XCHG(top_border_m1+24, src_cr-8, xchg);
965 XCHG(top_border+16, src_cb, 1);
966 XCHG(top_border+24, src_cr, 1);
967 }
968}
969
414ac27d
JGG
970static av_always_inline
971int check_intra_pred_mode(int mode, int mb_x, int mb_y)
3b636f21
DC
972{
973 if (mode == DC_PRED8x8) {
a71abb71
JGG
974 if (!mb_x) {
975 mode = mb_y ? TOP_DC_PRED8x8 : DC_128_PRED8x8;
976 } else if (!mb_y) {
09959ec4 977 mode = LEFT_DC_PRED8x8;
a71abb71 978 }
3b636f21
DC
979 }
980 return mode;
981}
982
414ac27d
JGG
983static av_always_inline
984void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
985 uint8_t *intra4x4, int mb_x, int mb_y)
3b636f21
DC
986{
987 int x, y, mode, nnz, tr;
988
9ac831c2
DC
989 // for the first row, we need to run xchg_mb_border to init the top edge to 127
990 // otherwise, skip it if we aren't going to deblock
991 if (s->deblock_filter || !mb_y)
992 xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
993 s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
994 s->filter.simple, 1);
995
3b636f21
DC
996 if (mb->mode < MODE_I4x4) {
997 mode = check_intra_pred_mode(mb->mode, mb_x, mb_y);
998 s->hpc.pred16x16[mode](dst[0], s->linesize);
999 } else {
1000 uint8_t *ptr = dst[0];
d1c58fce 1001 int stride = s->keyframe ? s->b4_stride : 4;
3b636f21
DC
1002
1003 // all blocks on the right edge of the macroblock use bottom edge
1004 // the top macroblock for their topright edge
1005 uint8_t *tr_right = ptr - s->linesize + 16;
1006
1007 // if we're on the right edge of the frame, said edge is extended
1008 // from the top macroblock
1009 if (mb_x == s->mb_width-1) {
1010 tr = tr_right[-1]*0x01010101;
1011 tr_right = (uint8_t *)&tr;
1012 }
1013
b74f70d6
JGG
1014 if (mb->skip)
1015 AV_ZERO128(s->non_zero_count_cache);
1016
3b636f21
DC
1017 for (y = 0; y < 4; y++) {
1018 uint8_t *topright = ptr + 4 - s->linesize;
1019 for (x = 0; x < 4; x++) {
1020 if (x == 3)
1021 topright = tr_right;
1022
d1c58fce 1023 s->hpc.pred4x4[intra4x4[x]](ptr+4*x, topright, s->linesize);
3b636f21
DC
1024
1025 nnz = s->non_zero_count_cache[y][x];
1026 if (nnz) {
1027 if (nnz == 1)
1028 s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize);
1029 else
1030 s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize);
1031 }
1032 topright += 4;
1033 }
1034
1035 ptr += 4*s->linesize;
d1c58fce 1036 intra4x4 += stride;
3b636f21
DC
1037 }
1038 }
1039
1040 mode = check_intra_pred_mode(s->chroma_pred_mode, mb_x, mb_y);
1041 s->hpc.pred8x8[mode](dst[1], s->uvlinesize);
1042 s->hpc.pred8x8[mode](dst[2], s->uvlinesize);
9ac831c2
DC
1043
1044 if (s->deblock_filter || !mb_y)
1045 xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
1046 s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
1047 s->filter.simple, 0);
3b636f21
DC
1048}
1049
1050/**
1051 * Generic MC function.
1052 *
1053 * @param s VP8 decoding context
1054 * @param luma 1 for luma (Y) planes, 0 for chroma (Cb/Cr) planes
1055 * @param dst target buffer for block data at block position
1056 * @param src reference picture buffer at origin (0, 0)
1057 * @param mv motion vector (relative to block position) to get pixel data from
1058 * @param x_off horizontal position of block from origin (0, 0)
1059 * @param y_off vertical position of block from origin (0, 0)
1060 * @param block_w width of block (16, 8 or 4)
1061 * @param block_h height of block (always same as block_w)
1062 * @param width width of src/dst plane data
1063 * @param height height of src/dst plane data
1064 * @param linesize size of a single line of plane data, including padding
e394953e 1065 * @param mc_func motion compensation function pointers (bilinear or sixtap MC)
3b636f21 1066 */
414ac27d
JGG
1067static av_always_inline
1068void vp8_mc(VP8Context *s, int luma,
1069 uint8_t *dst, uint8_t *src, const VP56mv *mv,
1070 int x_off, int y_off, int block_w, int block_h,
1071 int width, int height, int linesize,
1072 vp8_mc_func mc_func[3][3])
3b636f21 1073{
c0498b30
JGG
1074 if (AV_RN32A(mv)) {
1075 static const uint8_t idx[8] = { 0, 1, 2, 1, 2, 1, 2, 1 };
1076 int mx = (mv->x << luma)&7, mx_idx = idx[mx];
1077 int my = (mv->y << luma)&7, my_idx = idx[my];
1078
1079 x_off += mv->x >> (3 - luma);
1080 y_off += mv->y >> (3 - luma);
1081
1082 // edge emulation
1083 src += y_off * linesize + x_off;
1084 if (x_off < 2 || x_off >= width - block_w - 3 ||
1085 y_off < 2 || y_off >= height - block_h - 3) {
1086 ff_emulated_edge_mc(s->edge_emu_buffer, src - 2 * linesize - 2, linesize,
1087 block_w + 5, block_h + 5,
1088 x_off - 2, y_off - 2, width, height);
1089 src = s->edge_emu_buffer + 2 + linesize * 2;
1090 }
1091 mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);
1092 } else
1093 mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0);
3b636f21
DC
1094}
1095
414ac27d
JGG
1096static av_always_inline
1097void vp8_mc_part(VP8Context *s, uint8_t *dst[3],
1098 AVFrame *ref_frame, int x_off, int y_off,
1099 int bx_off, int by_off,
1100 int block_w, int block_h,
1101 int width, int height, VP56mv *mv)
7c4dcf81
RB
1102{
1103 VP56mv uvmv = *mv;
1104
1105 /* Y */
1106 vp8_mc(s, 1, dst[0] + by_off * s->linesize + bx_off,
1107 ref_frame->data[0], mv, x_off + bx_off, y_off + by_off,
1108 block_w, block_h, width, height, s->linesize,
1109 s->put_pixels_tab[block_w == 8]);
1110
1111 /* U/V */
1112 if (s->profile == 3) {
1113 uvmv.x &= ~7;
1114 uvmv.y &= ~7;
1115 }
1116 x_off >>= 1; y_off >>= 1;
1117 bx_off >>= 1; by_off >>= 1;
1118 width >>= 1; height >>= 1;
1119 block_w >>= 1; block_h >>= 1;
1120 vp8_mc(s, 0, dst[1] + by_off * s->uvlinesize + bx_off,
1121 ref_frame->data[1], &uvmv, x_off + bx_off, y_off + by_off,
1122 block_w, block_h, width, height, s->uvlinesize,
1123 s->put_pixels_tab[1 + (block_w == 4)]);
1124 vp8_mc(s, 0, dst[2] + by_off * s->uvlinesize + bx_off,
1125 ref_frame->data[2], &uvmv, x_off + bx_off, y_off + by_off,
1126 block_w, block_h, width, height, s->uvlinesize,
1127 s->put_pixels_tab[1 + (block_w == 4)]);
1128}
1129
d864dee8
JGG
1130/* Fetch pixels for estimated mv 4 macroblocks ahead.
1131 * Optimized for 64-byte cache lines. Inspired by ffh264 prefetch_motion. */
414ac27d 1132static av_always_inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int mb_xy, int ref)
d864dee8 1133{
ef38842f
JGG
1134 /* Don't prefetch refs that haven't been used very often this frame. */
1135 if (s->ref_count[ref-1] > (mb_xy >> 5)) {
c4211046
JGG
1136 int x_off = mb_x << 4, y_off = mb_y << 4;
1137 int mx = mb->mv.x + x_off + 8;
1138 int my = mb->mv.y + y_off;
1139 uint8_t **src= s->framep[ref]->data;
1140 int off= mx + (my + (mb_x&3)*4)*s->linesize + 64;
1141 s->dsp.prefetch(src[0]+off, s->linesize, 4);
1142 off= (mx>>1) + ((my>>1) + (mb_x&7))*s->uvlinesize + 64;
1143 s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1144 }
d864dee8
JGG
1145}
1146
3b636f21
DC
1147/**
1148 * Apply motion vectors to prediction buffer, chapter 18.
1149 */
414ac27d
JGG
1150static av_always_inline
1151void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
1152 int mb_x, int mb_y)
3b636f21
DC
1153{
1154 int x_off = mb_x << 4, y_off = mb_y << 4;
1155 int width = 16*s->mb_width, height = 16*s->mb_height;
d292c345
JGG
1156 AVFrame *ref = s->framep[mb->ref_frame];
1157 VP56mv *bmv = mb->bmv;
3b636f21
DC
1158
1159 if (mb->mode < VP8_MVMODE_SPLIT) {
d292c345 1160 vp8_mc_part(s, dst, ref, x_off, y_off,
7c4dcf81
RB
1161 0, 0, 16, 16, width, height, &mb->mv);
1162 } else switch (mb->partitioning) {
1163 case VP8_SPLITMVMODE_4x4: {
3b636f21 1164 int x, y;
7c4dcf81 1165 VP56mv uvmv;
3b636f21
DC
1166
1167 /* Y */
1168 for (y = 0; y < 4; y++) {
1169 for (x = 0; x < 4; x++) {
1170 vp8_mc(s, 1, dst[0] + 4*y*s->linesize + x*4,
d292c345 1171 ref->data[0], &bmv[4*y + x],
3b636f21
DC
1172 4*x + x_off, 4*y + y_off, 4, 4,
1173 width, height, s->linesize,
0ef1dbed 1174 s->put_pixels_tab[2]);
3b636f21
DC
1175 }
1176 }
1177
1178 /* U/V */
1179 x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1;
1180 for (y = 0; y < 2; y++) {
1181 for (x = 0; x < 2; x++) {
1182 uvmv.x = mb->bmv[ 2*y * 4 + 2*x ].x +
1183 mb->bmv[ 2*y * 4 + 2*x+1].x +
1184 mb->bmv[(2*y+1) * 4 + 2*x ].x +
1185 mb->bmv[(2*y+1) * 4 + 2*x+1].x;
1186 uvmv.y = mb->bmv[ 2*y * 4 + 2*x ].y +
1187 mb->bmv[ 2*y * 4 + 2*x+1].y +
1188 mb->bmv[(2*y+1) * 4 + 2*x ].y +
1189 mb->bmv[(2*y+1) * 4 + 2*x+1].y;
8f910a56
SG
1190 uvmv.x = (uvmv.x + 2 + (uvmv.x >> (INT_BIT-1))) >> 2;
1191 uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2;
3b636f21
DC
1192 if (s->profile == 3) {
1193 uvmv.x &= ~7;
1194 uvmv.y &= ~7;
1195 }
1196 vp8_mc(s, 0, dst[1] + 4*y*s->uvlinesize + x*4,
d292c345 1197 ref->data[1], &uvmv,
3b636f21
DC
1198 4*x + x_off, 4*y + y_off, 4, 4,
1199 width, height, s->uvlinesize,
0ef1dbed 1200 s->put_pixels_tab[2]);
3b636f21 1201 vp8_mc(s, 0, dst[2] + 4*y*s->uvlinesize + x*4,
d292c345 1202 ref->data[2], &uvmv,
3b636f21
DC
1203 4*x + x_off, 4*y + y_off, 4, 4,
1204 width, height, s->uvlinesize,
0ef1dbed 1205 s->put_pixels_tab[2]);
3b636f21
DC
1206 }
1207 }
7c4dcf81
RB
1208 break;
1209 }
1210 case VP8_SPLITMVMODE_16x8:
d292c345
JGG
1211 vp8_mc_part(s, dst, ref, x_off, y_off,
1212 0, 0, 16, 8, width, height, &bmv[0]);
1213 vp8_mc_part(s, dst, ref, x_off, y_off,
1214 0, 8, 16, 8, width, height, &bmv[1]);
7c4dcf81
RB
1215 break;
1216 case VP8_SPLITMVMODE_8x16:
d292c345
JGG
1217 vp8_mc_part(s, dst, ref, x_off, y_off,
1218 0, 0, 8, 16, width, height, &bmv[0]);
1219 vp8_mc_part(s, dst, ref, x_off, y_off,
1220 8, 0, 8, 16, width, height, &bmv[1]);
7c4dcf81
RB
1221 break;
1222 case VP8_SPLITMVMODE_8x8:
d292c345
JGG
1223 vp8_mc_part(s, dst, ref, x_off, y_off,
1224 0, 0, 8, 8, width, height, &bmv[0]);
1225 vp8_mc_part(s, dst, ref, x_off, y_off,
1226 8, 0, 8, 8, width, height, &bmv[1]);
1227 vp8_mc_part(s, dst, ref, x_off, y_off,
1228 0, 8, 8, 8, width, height, &bmv[2]);
1229 vp8_mc_part(s, dst, ref, x_off, y_off,
1230 8, 8, 8, 8, width, height, &bmv[3]);
7c4dcf81 1231 break;
3b636f21
DC
1232 }
1233}
1234
414ac27d 1235static av_always_inline void idct_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb)
3b636f21 1236{
3df56f41 1237 int x, y, ch;
3b636f21 1238
8a467b2d
JGG
1239 if (mb->mode != MODE_I4x4) {
1240 uint8_t *y_dst = dst[0];
3b636f21 1241 for (y = 0; y < 4; y++) {
3df56f41
JGG
1242 uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[y]);
1243 if (nnz4) {
1244 if (nnz4&~0x01010101) {
8a467b2d 1245 for (x = 0; x < 4; x++) {
3df56f41 1246 int nnz = s->non_zero_count_cache[y][x];
8a467b2d
JGG
1247 if (nnz) {
1248 if (nnz == 1)
1249 s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize);
1250 else
1251 s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize);
1252 }
1253 }
1254 } else {
3ae079a3 1255 s->vp8dsp.vp8_idct_dc_add4y(y_dst, s->block[y], s->linesize);
3b636f21
DC
1256 }
1257 }
1258 y_dst += 4*s->linesize;
1259 }
8a467b2d 1260 }
3b636f21 1261
8a467b2d 1262 for (ch = 0; ch < 2; ch++) {
3ae079a3
JGG
1263 uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[4+ch]);
1264 if (nnz4) {
8a467b2d 1265 uint8_t *ch_dst = dst[1+ch];
3ae079a3
JGG
1266 if (nnz4&~0x01010101) {
1267 for (y = 0; y < 2; y++) {
1268 for (x = 0; x < 2; x++) {
1269 int nnz = s->non_zero_count_cache[4+ch][(y<<1)+x];
1270 if (nnz) {
1271 if (nnz == 1)
1272 s->vp8dsp.vp8_idct_dc_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);
1273 else
1274 s->vp8dsp.vp8_idct_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);
1275 }
8a467b2d 1276 }
3ae079a3 1277 ch_dst += 4*s->uvlinesize;
8a467b2d 1278 }
3ae079a3
JGG
1279 } else {
1280 s->vp8dsp.vp8_idct_dc_add4uv(ch_dst, s->block[4+ch], s->uvlinesize);
3b636f21
DC
1281 }
1282 }
3b636f21
DC
1283 }
1284}
1285
414ac27d 1286static av_always_inline void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f )
3b636f21
DC
1287{
1288 int interior_limit, filter_level;
1289
1290 if (s->segmentation.enabled) {
b9a7186b 1291 filter_level = s->segmentation.filter_level[s->segment];
3b636f21
DC
1292 if (!s->segmentation.absolute_vals)
1293 filter_level += s->filter.level;
1294 } else
1295 filter_level = s->filter.level;
1296
1297 if (s->lf_delta.enabled) {
1298 filter_level += s->lf_delta.ref[mb->ref_frame];
1299
1300 if (mb->ref_frame == VP56_FRAME_CURRENT) {
1301 if (mb->mode == MODE_I4x4)
1302 filter_level += s->lf_delta.mode[0];
1303 } else {
1304 if (mb->mode == VP8_MVMODE_ZERO)
1305 filter_level += s->lf_delta.mode[1];
1306 else if (mb->mode == VP8_MVMODE_SPLIT)
1307 filter_level += s->lf_delta.mode[3];
1308 else
1309 filter_level += s->lf_delta.mode[2];
1310 }
1311 }
1312 filter_level = av_clip(filter_level, 0, 63);
1313
1314 interior_limit = filter_level;
1315 if (s->filter.sharpness) {
1316 interior_limit >>= s->filter.sharpness > 4 ? 2 : 1;
1317 interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness);
1318 }
1319 interior_limit = FFMAX(interior_limit, 1);
1320
968570d6
JGG
1321 f->filter_level = filter_level;
1322 f->inner_limit = interior_limit;
c55e0d34 1323 f->inner_filter = !mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT;
3b636f21
DC
1324}
1325
414ac27d 1326static av_always_inline void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f, int mb_x, int mb_y)
3b636f21 1327{
968570d6
JGG
1328 int mbedge_lim, bedge_lim, hev_thresh;
1329 int filter_level = f->filter_level;
1330 int inner_limit = f->inner_limit;
c55e0d34 1331 int inner_filter = f->inner_filter;
145d3186
JGG
1332 int linesize = s->linesize;
1333 int uvlinesize = s->uvlinesize;
3b636f21 1334
3b636f21
DC
1335 if (!filter_level)
1336 return;
1337
5245c04d
DC
1338 mbedge_lim = 2*(filter_level+2) + inner_limit;
1339 bedge_lim = 2* filter_level + inner_limit;
968570d6
JGG
1340 hev_thresh = filter_level >= 15;
1341
1342 if (s->keyframe) {
1343 if (filter_level >= 40)
1344 hev_thresh = 2;
1345 } else {
1346 if (filter_level >= 40)
1347 hev_thresh = 3;
1348 else if (filter_level >= 20)
1349 hev_thresh = 2;
1350 }
5245c04d 1351
3b636f21 1352 if (mb_x) {
145d3186 1353 s->vp8dsp.vp8_h_loop_filter16y(dst[0], linesize,
3facfc99 1354 mbedge_lim, inner_limit, hev_thresh);
145d3186 1355 s->vp8dsp.vp8_h_loop_filter8uv(dst[1], dst[2], uvlinesize,
3facfc99 1356 mbedge_lim, inner_limit, hev_thresh);
3b636f21
DC
1357 }
1358
c55e0d34 1359 if (inner_filter) {
145d3186
JGG
1360 s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 4, linesize, bedge_lim,
1361 inner_limit, hev_thresh);
1362 s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 8, linesize, bedge_lim,
1363 inner_limit, hev_thresh);
1364 s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+12, linesize, bedge_lim,
1365 inner_limit, hev_thresh);
1366 s->vp8dsp.vp8_h_loop_filter8uv_inner(dst[1] + 4, dst[2] + 4,
1367 uvlinesize, bedge_lim,
1368 inner_limit, hev_thresh);
3b636f21
DC
1369 }
1370
1371 if (mb_y) {
145d3186 1372 s->vp8dsp.vp8_v_loop_filter16y(dst[0], linesize,
3facfc99 1373 mbedge_lim, inner_limit, hev_thresh);
145d3186 1374 s->vp8dsp.vp8_v_loop_filter8uv(dst[1], dst[2], uvlinesize,
3facfc99 1375 mbedge_lim, inner_limit, hev_thresh);
3b636f21
DC
1376 }
1377
c55e0d34 1378 if (inner_filter) {
145d3186
JGG
1379 s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 4*linesize,
1380 linesize, bedge_lim,
1381 inner_limit, hev_thresh);
1382 s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 8*linesize,
1383 linesize, bedge_lim,
1384 inner_limit, hev_thresh);
1385 s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+12*linesize,
1386 linesize, bedge_lim,
1387 inner_limit, hev_thresh);
1388 s->vp8dsp.vp8_v_loop_filter8uv_inner(dst[1] + 4 * uvlinesize,
1389 dst[2] + 4 * uvlinesize,
1390 uvlinesize, bedge_lim,
3facfc99 1391 inner_limit, hev_thresh);
3b636f21
DC
1392 }
1393}
1394
414ac27d 1395static av_always_inline void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f, int mb_x, int mb_y)
3b636f21 1396{
968570d6
JGG
1397 int mbedge_lim, bedge_lim;
1398 int filter_level = f->filter_level;
1399 int inner_limit = f->inner_limit;
c55e0d34 1400 int inner_filter = f->inner_filter;
145d3186 1401 int linesize = s->linesize;
3b636f21 1402
3b636f21
DC
1403 if (!filter_level)
1404 return;
1405
1406 mbedge_lim = 2*(filter_level+2) + inner_limit;
1407 bedge_lim = 2* filter_level + inner_limit;
1408
1409 if (mb_x)
145d3186 1410 s->vp8dsp.vp8_h_loop_filter_simple(dst, linesize, mbedge_lim);
c55e0d34 1411 if (inner_filter) {
145d3186
JGG
1412 s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, linesize, bedge_lim);
1413 s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, linesize, bedge_lim);
1414 s->vp8dsp.vp8_h_loop_filter_simple(dst+12, linesize, bedge_lim);
3b636f21
DC
1415 }
1416
1417 if (mb_y)
145d3186 1418 s->vp8dsp.vp8_v_loop_filter_simple(dst, linesize, mbedge_lim);
c55e0d34 1419 if (inner_filter) {
145d3186
JGG
1420 s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*linesize, linesize, bedge_lim);
1421 s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*linesize, linesize, bedge_lim);
1422 s->vp8dsp.vp8_v_loop_filter_simple(dst+12*linesize, linesize, bedge_lim);
3b636f21
DC
1423 }
1424}
1425
1426static void filter_mb_row(VP8Context *s, int mb_y)
1427{
968570d6 1428 VP8FilterStrength *f = s->filter_strength;
3b636f21
DC
1429 uint8_t *dst[3] = {
1430 s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize,
1431 s->framep[VP56_FRAME_CURRENT]->data[1] + 8*mb_y*s->uvlinesize,
1432 s->framep[VP56_FRAME_CURRENT]->data[2] + 8*mb_y*s->uvlinesize
1433 };
1434 int mb_x;
1435
1436 for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
9ac831c2 1437 backup_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], s->linesize, s->uvlinesize, 0);
c55e0d34 1438 filter_mb(s, dst, f++, mb_x, mb_y);
3b636f21
DC
1439 dst[0] += 16;
1440 dst[1] += 8;
1441 dst[2] += 8;
1442 }
1443}
1444
1445static void filter_mb_row_simple(VP8Context *s, int mb_y)
1446{
968570d6 1447 VP8FilterStrength *f = s->filter_strength;
968570d6 1448 uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize;
3b636f21
DC
1449 int mb_x;
1450
1451 for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
9ac831c2 1452 backup_mb_border(s->top_border[mb_x+1], dst, NULL, NULL, s->linesize, 0, 1);
c55e0d34 1453 filter_mb_simple(s, dst, f++, mb_x, mb_y);
3b636f21
DC
1454 dst += 16;
1455 }
1456}
1457
1458static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
1459 AVPacket *avpkt)
1460{
1461 VP8Context *s = avctx->priv_data;
1462 int ret, mb_x, mb_y, i, y, referenced;
1463 enum AVDiscard skip_thresh;
28e241de 1464 AVFrame *av_uninit(curframe);
3b636f21
DC
1465
1466 if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0)
1467 return ret;
1468
1469 referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT
1470 || s->update_altref == VP56_FRAME_CURRENT;
1471
1472 skip_thresh = !referenced ? AVDISCARD_NONREF :
1473 !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL;
1474
1475 if (avctx->skip_frame >= skip_thresh) {
1476 s->invisible = 1;
1477 goto skip_decode;
1478 }
9ac831c2 1479 s->deblock_filter = s->filter.level && avctx->skip_loop_filter < skip_thresh;
3b636f21
DC
1480
1481 for (i = 0; i < 4; i++)
1482 if (&s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1483 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1484 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) {
1485 curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i];
1486 break;
1487 }
1488 if (curframe->data[0])
1489 avctx->release_buffer(avctx, curframe);
1490
1491 curframe->key_frame = s->keyframe;
1492 curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE;
1493 curframe->reference = referenced ? 3 : 0;
1494 if ((ret = avctx->get_buffer(avctx, curframe))) {
1495 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n");
1496 return ret;
1497 }
1498
1499 // Given that arithmetic probabilities are updated every frame, it's quite likely
1500 // that the values we have on a random interframe are complete junk if we didn't
1501 // start decode on a keyframe. So just don't display anything rather than junk.
1502 if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] ||
1503 !s->framep[VP56_FRAME_GOLDEN] ||
1504 !s->framep[VP56_FRAME_GOLDEN2])) {
1505 av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n");
1506 return AVERROR_INVALIDDATA;
1507 }
1508
1509 s->linesize = curframe->linesize[0];
1510 s->uvlinesize = curframe->linesize[1];
1511
1512 if (!s->edge_emu_buffer)
1513 s->edge_emu_buffer = av_malloc(21*s->linesize);
1514
1515 memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz));
1516
c55e0d34
JGG
1517 /* Zero macroblock structures for top/left prediction from outside the frame. */
1518 memset(s->macroblocks, 0, (s->mb_width + s->mb_height*2)*sizeof(*s->macroblocks));
1519
3b636f21 1520 // top edge of 127 for intra prediction
9ac831c2 1521 memset(s->top_border, 127, (s->mb_width+1)*sizeof(*s->top_border));
c4211046 1522 memset(s->ref_count, 0, sizeof(s->ref_count));
3b636f21
DC
1523
1524 for (mb_y = 0; mb_y < s->mb_height; mb_y++) {
1525 VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)];
c55e0d34 1526 VP8Macroblock *mb = s->macroblocks + (s->mb_height - mb_y - 1)*2;
3b636f21 1527 uint8_t *intra4x4 = s->intra4x4_pred_mode + 4*mb_y*s->b4_stride;
c55e0d34 1528 uint8_t *segment_map = s->segmentation_map + mb_y*s->mb_stride;
ef38842f 1529 int mb_xy = mb_y * s->mb_stride;
3b636f21
DC
1530 uint8_t *dst[3] = {
1531 curframe->data[0] + 16*mb_y*s->linesize,
1532 curframe->data[1] + 8*mb_y*s->uvlinesize,
1533 curframe->data[2] + 8*mb_y*s->uvlinesize
1534 };
1535
1536 memset(s->left_nnz, 0, sizeof(s->left_nnz));
1537
1538 // left edge of 129 for intra prediction
1539 if (!(avctx->flags & CODEC_FLAG_EMU_EDGE))
1540 for (i = 0; i < 3; i++)
1541 for (y = 0; y < 16>>!!i; y++)
1542 dst[i][y*curframe->linesize[i]-1] = 129;
9ac831c2
DC
1543 if (mb_y)
1544 memset(s->top_border, 129, sizeof(*s->top_border));
3b636f21 1545
ef38842f 1546 for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb_xy++, mb++) {
d1c58fce 1547 uint8_t *intra4x4_mb = s->keyframe ? intra4x4 + 4*mb_x : s->intra4x4_pred_mode_mb;
c55e0d34 1548 uint8_t *segment_mb = segment_map+mb_x;
d1c58fce 1549
d864dee8
JGG
1550 /* Prefetch the current frame, 4 MBs ahead */
1551 s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4);
1552 s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2);
1553
c55e0d34 1554 decode_mb_mode(s, mb, mb_x, mb_y, intra4x4_mb, segment_mb);
3b636f21 1555
ef38842f 1556 prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_PREVIOUS);
c4211046 1557
3b636f21
DC
1558 if (!mb->skip)
1559 decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);
3b636f21 1560
b946111f 1561 if (mb->mode <= MODE_I4x4)
d1c58fce 1562 intra_predict(s, dst, mb, intra4x4_mb, mb_x, mb_y);
b946111f 1563 else
3b636f21 1564 inter_predict(s, dst, mb, mb_x, mb_y);
3b636f21 1565
ef38842f 1566 prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN);
c4211046 1567
3b636f21 1568 if (!mb->skip) {
8a467b2d 1569 idct_mb(s, dst, mb);
3b636f21
DC
1570 } else {
1571 AV_ZERO64(s->left_nnz);
1572 AV_WN64(s->top_nnz[mb_x], 0); // array of 9, so unaligned
1573
1574 // Reset DC block predictors if they would exist if the mb had coefficients
1575 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
1576 s->left_nnz[8] = 0;
1577 s->top_nnz[mb_x][8] = 0;
1578 }
1579 }
1580
968570d6
JGG
1581 if (s->deblock_filter)
1582 filter_level_for_mb(s, mb, &s->filter_strength[mb_x]);
1583
ef38842f 1584 prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN2);
c4211046 1585
3b636f21
DC
1586 dst[0] += 16;
1587 dst[1] += 8;
1588 dst[2] += 8;
3b636f21 1589 }
9ac831c2 1590 if (s->deblock_filter) {
3b636f21 1591 if (s->filter.simple)
9ac831c2 1592 filter_mb_row_simple(s, mb_y);
3b636f21 1593 else
9ac831c2 1594 filter_mb_row(s, mb_y);
3b636f21
DC
1595 }
1596 }
3b636f21
DC
1597
1598skip_decode:
1599 // if future frames don't use the updated probabilities,
1600 // reset them to the values we saved
1601 if (!s->update_probabilities)
1602 s->prob[0] = s->prob[1];
1603
1604 // check if golden and altref are swapped
1605 if (s->update_altref == VP56_FRAME_GOLDEN &&
1606 s->update_golden == VP56_FRAME_GOLDEN2)
1607 FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN], s->framep[VP56_FRAME_GOLDEN2]);
1608 else {
1609 if (s->update_altref != VP56_FRAME_NONE)
1610 s->framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref];
1611
1612 if (s->update_golden != VP56_FRAME_NONE)
1613 s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden];
1614 }
1615
1616 if (s->update_last) // move cur->prev
1617 s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT];
1618
1619 // release no longer referenced frames
1620 for (i = 0; i < 4; i++)
1621 if (s->frames[i].data[0] &&
1622 &s->frames[i] != s->framep[VP56_FRAME_CURRENT] &&
1623 &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1624 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1625 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2])
1626 avctx->release_buffer(avctx, &s->frames[i]);
1627
1628 if (!s->invisible) {
1629 *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT];
1630 *data_size = sizeof(AVFrame);
1631 }
1632
1633 return avpkt->size;
1634}
1635
1636static av_cold int vp8_decode_init(AVCodecContext *avctx)
1637{
1638 VP8Context *s = avctx->priv_data;
1639
1640 s->avctx = avctx;
1641 avctx->pix_fmt = PIX_FMT_YUV420P;
1642
1643 dsputil_init(&s->dsp, avctx);
1644 ff_h264_pred_init(&s->hpc, CODEC_ID_VP8);
1645 ff_vp8dsp_init(&s->vp8dsp);
1646
1647 // intra pred needs edge emulation among other things
1648 if (avctx->flags&CODEC_FLAG_EMU_EDGE) {
03ac56e7 1649 av_log(avctx, AV_LOG_ERROR, "Edge emulation not supported\n");
3b636f21
DC
1650 return AVERROR_PATCHWELCOME;
1651 }
1652
1653 return 0;
1654}
1655
1656static av_cold int vp8_decode_free(AVCodecContext *avctx)
1657{
1658 vp8_decode_flush(avctx);
1659 return 0;
1660}
1661
1662AVCodec vp8_decoder = {
1663 "vp8",
1664 AVMEDIA_TYPE_VIDEO,
1665 CODEC_ID_VP8,
1666 sizeof(VP8Context),
1667 vp8_decode_init,
1668 NULL,
1669 vp8_decode_free,
1670 vp8_decode_frame,
1671 CODEC_CAP_DR1,
1672 .flush = vp8_decode_flush,
1673 .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"),
1674};