Convert vp56_mv to 16-bit.
[libav.git] / libavcodec / vp8.c
CommitLineData
3b636f21
DC
1/**
2 * VP8 compatible video decoder
3 *
4 * Copyright (C) 2010 David Conrad
5 * Copyright (C) 2010 Ronald S. Bultje
6 *
7 * This file is part of FFmpeg.
8 *
9 * FFmpeg is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
13 *
14 * FFmpeg is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
18 *
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with FFmpeg; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22 */
23
24#include "avcodec.h"
25#include "vp56.h"
26#include "vp8data.h"
27#include "vp8dsp.h"
28#include "h264pred.h"
29#include "rectangle.h"
30
31typedef struct {
32 uint8_t segment;
33 uint8_t skip;
34 // todo: make it possible to check for at least (i4x4 or split_mv)
35 // in one op. are others needed?
36 uint8_t mode;
37 uint8_t ref_frame;
38 uint8_t partitioning;
39 VP56mv mv;
40 VP56mv bmv[16];
41} VP8Macroblock;
42
43typedef struct {
44 AVCodecContext *avctx;
45 DSPContext dsp;
46 VP8DSPContext vp8dsp;
47 H264PredContext hpc;
0ef1dbed 48 vp8_mc_func put_pixels_tab[3][3][3];
3b636f21
DC
49 AVFrame frames[4];
50 AVFrame *framep[4];
51 uint8_t *edge_emu_buffer;
52 VP56RangeCoder c; ///< header context, includes mb modes and motion vectors
53 int profile;
54
55 int mb_width; /* number of horizontal MB */
56 int mb_height; /* number of vertical MB */
57 int linesize;
58 int uvlinesize;
59
60 int keyframe;
61 int invisible;
62 int update_last; ///< update VP56_FRAME_PREVIOUS with the current one
63 int update_golden; ///< VP56_FRAME_NONE if not updated, or which frame to copy if so
64 int update_altref;
9ac831c2 65 int deblock_filter;
3b636f21
DC
66
67 /**
68 * If this flag is not set, all the probability updates
69 * are discarded after this frame is decoded.
70 */
71 int update_probabilities;
72
73 /**
74 * All coefficients are contained in separate arith coding contexts.
75 * There can be 1, 2, 4, or 8 of these after the header context.
76 */
77 int num_coeff_partitions;
78 VP56RangeCoder coeff_partition[8];
79
80 VP8Macroblock *macroblocks;
81 VP8Macroblock *macroblocks_base;
82 int mb_stride;
83
84 uint8_t *intra4x4_pred_mode;
85 uint8_t *intra4x4_pred_mode_base;
86 int b4_stride;
87
88 /**
9ac831c2
DC
89 * Cache of the top row needed for intra prediction
90 * 16 for luma, 8 for each chroma plane
91 */
92 uint8_t (*top_border)[16+8+8];
93
94 /**
3b636f21
DC
95 * For coeff decode, we need to know whether the above block had non-zero
96 * coefficients. This means for each macroblock, we need data for 4 luma
97 * blocks, 2 u blocks, 2 v blocks, and the luma dc block, for a total of 9
98 * per macroblock. We keep the last row in top_nnz.
99 */
100 uint8_t (*top_nnz)[9];
101 DECLARE_ALIGNED(8, uint8_t, left_nnz)[9];
102
103 /**
104 * This is the index plus one of the last non-zero coeff
105 * for each of the blocks in the current macroblock.
106 * So, 0 -> no coeffs
107 * 1 -> dc-only (special transform)
108 * 2+-> full transform
109 */
110 DECLARE_ALIGNED(16, uint8_t, non_zero_count_cache)[6][4];
111 DECLARE_ALIGNED(16, DCTELEM, block)[6][4][16];
112
113 int chroma_pred_mode; ///< 8x8c pred mode of the current macroblock
114
115 int mbskip_enabled;
116 int sign_bias[4]; ///< one state [0, 1] per ref frame type
117
118 /**
119 * Base parameters for segmentation, i.e. per-macroblock parameters.
120 * These must be kept unchanged even if segmentation is not used for
121 * a frame, since the values persist between interframes.
122 */
123 struct {
124 int enabled;
125 int absolute_vals;
126 int update_map;
127 int8_t base_quant[4];
128 int8_t filter_level[4]; ///< base loop filter level
129 } segmentation;
130
131 /**
132 * Macroblocks can have one of 4 different quants in a frame when
133 * segmentation is enabled.
134 * If segmentation is disabled, only the first segment's values are used.
135 */
136 struct {
137 // [0] - DC qmul [1] - AC qmul
138 int16_t luma_qmul[2];
139 int16_t luma_dc_qmul[2]; ///< luma dc-only block quant
140 int16_t chroma_qmul[2];
141 } qmat[4];
142
143 struct {
144 int simple;
145 int level;
146 int sharpness;
147 } filter;
148
149 struct {
150 int enabled; ///< whether each mb can have a different strength based on mode/ref
151
152 /**
153 * filter strength adjustment for the following macroblock modes:
154 * [0] - i4x4
155 * [1] - zero mv
156 * [2] - inter modes except for zero or split mv
157 * [3] - split mv
158 * i16x16 modes never have any adjustment
159 */
160 int8_t mode[4];
161
162 /**
163 * filter strength adjustment for macroblocks that reference:
164 * [0] - intra / VP56_FRAME_CURRENT
165 * [1] - VP56_FRAME_PREVIOUS
166 * [2] - VP56_FRAME_GOLDEN
167 * [3] - altref / VP56_FRAME_GOLDEN2
168 */
169 int8_t ref[4];
170 } lf_delta;
171
172 /**
173 * These are all of the updatable probabilities for binary decisions.
174 * They are only implictly reset on keyframes, making it quite likely
175 * for an interframe to desync if a prior frame's header was corrupt
176 * or missing outright!
177 */
178 struct {
179 uint8_t segmentid[3];
180 uint8_t mbskip;
181 uint8_t intra;
182 uint8_t last;
183 uint8_t golden;
184 uint8_t pred16x16[4];
185 uint8_t pred8x8c[3];
186 uint8_t token[4][8][3][NUM_DCT_TOKENS-1];
187 uint8_t mvc[2][19];
188 } prob[2];
189} VP8Context;
190
191#define RL24(p) (AV_RL16(p) + ((p)[2] << 16))
192
193static void vp8_decode_flush(AVCodecContext *avctx)
194{
195 VP8Context *s = avctx->priv_data;
196 int i;
197
198 for (i = 0; i < 4; i++)
199 if (s->frames[i].data[0])
200 avctx->release_buffer(avctx, &s->frames[i]);
201 memset(s->framep, 0, sizeof(s->framep));
202
203 av_freep(&s->macroblocks_base);
204 av_freep(&s->intra4x4_pred_mode_base);
205 av_freep(&s->top_nnz);
206 av_freep(&s->edge_emu_buffer);
9ac831c2 207 av_freep(&s->top_border);
3b636f21
DC
208
209 s->macroblocks = NULL;
210 s->intra4x4_pred_mode = NULL;
211}
212
213static int update_dimensions(VP8Context *s, int width, int height)
214{
215 int i;
216
217 if (avcodec_check_dimensions(s->avctx, width, height))
218 return AVERROR_INVALIDDATA;
219
220 vp8_decode_flush(s->avctx);
221
222 avcodec_set_dimensions(s->avctx, width, height);
223
224 s->mb_width = (s->avctx->coded_width +15) / 16;
225 s->mb_height = (s->avctx->coded_height+15) / 16;
226
227 // we allocate a border around the top/left of intra4x4 modes
228 // this is 4 blocks for intra4x4 to keep 4-byte alignment for fill_rectangle
229 s->mb_stride = s->mb_width+1;
230 s->b4_stride = 4*s->mb_stride;
231
232 s->macroblocks_base = av_mallocz(s->mb_stride*(s->mb_height+1)*sizeof(*s->macroblocks));
233 s->intra4x4_pred_mode_base = av_mallocz(s->b4_stride*(4*s->mb_height+1));
234 s->top_nnz = av_mallocz(s->mb_width*sizeof(*s->top_nnz));
9ac831c2 235 s->top_border = av_mallocz((s->mb_width+1)*sizeof(*s->top_border));
3b636f21 236
9ac831c2 237 if (!s->macroblocks_base || !s->intra4x4_pred_mode_base || !s->top_nnz || !s->top_border)
b6c420ce
DC
238 return AVERROR(ENOMEM);
239
3b636f21
DC
240 s->macroblocks = s->macroblocks_base + 1 + s->mb_stride;
241 s->intra4x4_pred_mode = s->intra4x4_pred_mode_base + 4 + s->b4_stride;
242
243 memset(s->intra4x4_pred_mode_base, DC_PRED, s->b4_stride);
244 for (i = 0; i < 4*s->mb_height; i++)
245 s->intra4x4_pred_mode[i*s->b4_stride-1] = DC_PRED;
246
247 return 0;
248}
249
250static void parse_segment_info(VP8Context *s)
251{
252 VP56RangeCoder *c = &s->c;
253 int i;
254
255 s->segmentation.update_map = vp8_rac_get(c);
256
257 if (vp8_rac_get(c)) { // update segment feature data
258 s->segmentation.absolute_vals = vp8_rac_get(c);
259
260 for (i = 0; i < 4; i++)
261 s->segmentation.base_quant[i] = vp8_rac_get_sint(c, 7);
262
263 for (i = 0; i < 4; i++)
264 s->segmentation.filter_level[i] = vp8_rac_get_sint(c, 6);
265 }
266 if (s->segmentation.update_map)
267 for (i = 0; i < 3; i++)
268 s->prob->segmentid[i] = vp8_rac_get(c) ? vp8_rac_get_uint(c, 8) : 255;
269}
270
271static void update_lf_deltas(VP8Context *s)
272{
273 VP56RangeCoder *c = &s->c;
274 int i;
275
276 for (i = 0; i < 4; i++)
277 s->lf_delta.ref[i] = vp8_rac_get_sint(c, 6);
278
279 for (i = 0; i < 4; i++)
280 s->lf_delta.mode[i] = vp8_rac_get_sint(c, 6);
281}
282
283static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size)
284{
285 const uint8_t *sizes = buf;
286 int i;
287
288 s->num_coeff_partitions = 1 << vp8_rac_get_uint(&s->c, 2);
289
290 buf += 3*(s->num_coeff_partitions-1);
291 buf_size -= 3*(s->num_coeff_partitions-1);
292 if (buf_size < 0)
293 return -1;
294
295 for (i = 0; i < s->num_coeff_partitions-1; i++) {
296 int size = RL24(sizes + 3*i);
297 if (buf_size - size < 0)
298 return -1;
299
300 vp56_init_range_decoder(&s->coeff_partition[i], buf, size);
301 buf += size;
302 buf_size -= size;
303 }
304 vp56_init_range_decoder(&s->coeff_partition[i], buf, buf_size);
305
306 return 0;
307}
308
309static void get_quants(VP8Context *s)
310{
311 VP56RangeCoder *c = &s->c;
312 int i, base_qi;
313
314 int yac_qi = vp8_rac_get_uint(c, 7);
315 int ydc_delta = vp8_rac_get_sint(c, 4);
316 int y2dc_delta = vp8_rac_get_sint(c, 4);
317 int y2ac_delta = vp8_rac_get_sint(c, 4);
318 int uvdc_delta = vp8_rac_get_sint(c, 4);
319 int uvac_delta = vp8_rac_get_sint(c, 4);
320
321 for (i = 0; i < 4; i++) {
322 if (s->segmentation.enabled) {
323 base_qi = s->segmentation.base_quant[i];
324 if (!s->segmentation.absolute_vals)
325 base_qi += yac_qi;
326 } else
327 base_qi = yac_qi;
328
329 s->qmat[i].luma_qmul[0] = vp8_dc_qlookup[av_clip(base_qi + ydc_delta , 0, 127)];
330 s->qmat[i].luma_qmul[1] = vp8_ac_qlookup[av_clip(base_qi , 0, 127)];
331 s->qmat[i].luma_dc_qmul[0] = 2 * vp8_dc_qlookup[av_clip(base_qi + y2dc_delta, 0, 127)];
332 s->qmat[i].luma_dc_qmul[1] = 155 * vp8_ac_qlookup[av_clip(base_qi + y2ac_delta, 0, 127)] / 100;
333 s->qmat[i].chroma_qmul[0] = vp8_dc_qlookup[av_clip(base_qi + uvdc_delta, 0, 127)];
334 s->qmat[i].chroma_qmul[1] = vp8_ac_qlookup[av_clip(base_qi + uvac_delta, 0, 127)];
335
336 s->qmat[i].luma_dc_qmul[1] = FFMAX(s->qmat[i].luma_dc_qmul[1], 8);
337 s->qmat[i].chroma_qmul[0] = FFMIN(s->qmat[i].chroma_qmul[0], 132);
338 }
339}
340
341/**
342 * Determine which buffers golden and altref should be updated with after this frame.
343 * The spec isn't clear here, so I'm going by my understanding of what libvpx does
344 *
345 * Intra frames update all 3 references
346 * Inter frames update VP56_FRAME_PREVIOUS if the update_last flag is set
347 * If the update (golden|altref) flag is set, it's updated with the current frame
348 * if update_last is set, and VP56_FRAME_PREVIOUS otherwise.
349 * If the flag is not set, the number read means:
350 * 0: no update
351 * 1: VP56_FRAME_PREVIOUS
352 * 2: update golden with altref, or update altref with golden
353 */
354static VP56Frame ref_to_update(VP8Context *s, int update, VP56Frame ref)
355{
356 VP56RangeCoder *c = &s->c;
357
358 if (update)
359 return VP56_FRAME_CURRENT;
360
361 switch (vp8_rac_get_uint(c, 2)) {
362 case 1:
363 return VP56_FRAME_PREVIOUS;
364 case 2:
365 return (ref == VP56_FRAME_GOLDEN) ? VP56_FRAME_GOLDEN2 : VP56_FRAME_GOLDEN;
366 }
367 return VP56_FRAME_NONE;
368}
369
370static void update_refs(VP8Context *s)
371{
372 VP56RangeCoder *c = &s->c;
373
374 int update_golden = vp8_rac_get(c);
375 int update_altref = vp8_rac_get(c);
376
377 s->update_golden = ref_to_update(s, update_golden, VP56_FRAME_GOLDEN);
378 s->update_altref = ref_to_update(s, update_altref, VP56_FRAME_GOLDEN2);
379}
380
381static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)
382{
383 VP56RangeCoder *c = &s->c;
384 int header_size, hscale, vscale, i, j, k, l, ret;
385 int width = s->avctx->width;
386 int height = s->avctx->height;
387
388 s->keyframe = !(buf[0] & 1);
389 s->profile = (buf[0]>>1) & 7;
390 s->invisible = !(buf[0] & 0x10);
391 header_size = RL24(buf) >> 5;
392 buf += 3;
393 buf_size -= 3;
394
0ef1dbed
DC
395 if (s->profile > 3)
396 av_log(s->avctx, AV_LOG_WARNING, "Unknown profile %d\n", s->profile);
397
398 if (!s->profile)
399 memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_epel_pixels_tab, sizeof(s->put_pixels_tab));
400 else // profile 1-3 use bilinear, 4+ aren't defined so whatever
401 memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_bilinear_pixels_tab, sizeof(s->put_pixels_tab));
3b636f21
DC
402
403 if (header_size > buf_size - 7*s->keyframe) {
404 av_log(s->avctx, AV_LOG_ERROR, "Header size larger than data provided\n");
405 return AVERROR_INVALIDDATA;
406 }
407
408 if (s->keyframe) {
409 if (RL24(buf) != 0x2a019d) {
410 av_log(s->avctx, AV_LOG_ERROR, "Invalid start code 0x%x\n", RL24(buf));
411 return AVERROR_INVALIDDATA;
412 }
413 width = AV_RL16(buf+3) & 0x3fff;
414 height = AV_RL16(buf+5) & 0x3fff;
415 hscale = buf[4] >> 6;
416 vscale = buf[6] >> 6;
417 buf += 7;
418 buf_size -= 7;
419
92a54426
MR
420 if (hscale || vscale)
421 av_log_missing_feature(s->avctx, "Upscaling", 1);
422
3b636f21
DC
423 s->update_golden = s->update_altref = VP56_FRAME_CURRENT;
424 memcpy(s->prob->token , vp8_token_default_probs , sizeof(s->prob->token));
425 memcpy(s->prob->pred16x16, vp8_pred16x16_prob_inter, sizeof(s->prob->pred16x16));
426 memcpy(s->prob->pred8x8c , vp8_pred8x8c_prob_inter , sizeof(s->prob->pred8x8c));
427 memcpy(s->prob->mvc , vp8_mv_default_prob , sizeof(s->prob->mvc));
428 memset(&s->segmentation, 0, sizeof(s->segmentation));
429 }
430
431 if (!s->macroblocks_base || /* first frame */
432 width != s->avctx->width || height != s->avctx->height) {
433 if ((ret = update_dimensions(s, width, height) < 0))
434 return ret;
435 }
436
437 vp56_init_range_decoder(c, buf, header_size);
438 buf += header_size;
439 buf_size -= header_size;
440
441 if (s->keyframe) {
442 if (vp8_rac_get(c))
443 av_log(s->avctx, AV_LOG_WARNING, "Unspecified colorspace\n");
444 vp8_rac_get(c); // whether we can skip clamping in dsp functions
445 }
446
447 if ((s->segmentation.enabled = vp8_rac_get(c)))
448 parse_segment_info(s);
449 else
450 s->segmentation.update_map = 0; // FIXME: move this to some init function?
451
452 s->filter.simple = vp8_rac_get(c);
453 s->filter.level = vp8_rac_get_uint(c, 6);
454 s->filter.sharpness = vp8_rac_get_uint(c, 3);
455
456 if ((s->lf_delta.enabled = vp8_rac_get(c)))
457 if (vp8_rac_get(c))
458 update_lf_deltas(s);
459
460 if (setup_partitions(s, buf, buf_size)) {
461 av_log(s->avctx, AV_LOG_ERROR, "Invalid partitions\n");
462 return AVERROR_INVALIDDATA;
463 }
464
465 get_quants(s);
466
467 if (!s->keyframe) {
468 update_refs(s);
469 s->sign_bias[VP56_FRAME_GOLDEN] = vp8_rac_get(c);
470 s->sign_bias[VP56_FRAME_GOLDEN2 /* altref */] = vp8_rac_get(c);
471 }
472
473 // if we aren't saving this frame's probabilities for future frames,
474 // make a copy of the current probabilities
475 if (!(s->update_probabilities = vp8_rac_get(c)))
476 s->prob[1] = s->prob[0];
477
478 s->update_last = s->keyframe || vp8_rac_get(c);
479
480 for (i = 0; i < 4; i++)
481 for (j = 0; j < 8; j++)
482 for (k = 0; k < 3; k++)
483 for (l = 0; l < NUM_DCT_TOKENS-1; l++)
484 if (vp56_rac_get_prob(c, vp8_token_update_probs[i][j][k][l]))
485 s->prob->token[i][j][k][l] = vp8_rac_get_uint(c, 8);
486
487 if ((s->mbskip_enabled = vp8_rac_get(c)))
488 s->prob->mbskip = vp8_rac_get_uint(c, 8);
489
490 if (!s->keyframe) {
491 s->prob->intra = vp8_rac_get_uint(c, 8);
492 s->prob->last = vp8_rac_get_uint(c, 8);
493 s->prob->golden = vp8_rac_get_uint(c, 8);
494
495 if (vp8_rac_get(c))
496 for (i = 0; i < 4; i++)
497 s->prob->pred16x16[i] = vp8_rac_get_uint(c, 8);
498 if (vp8_rac_get(c))
499 for (i = 0; i < 3; i++)
500 s->prob->pred8x8c[i] = vp8_rac_get_uint(c, 8);
501
502 // 17.2 MV probability update
503 for (i = 0; i < 2; i++)
504 for (j = 0; j < 19; j++)
505 if (vp56_rac_get_prob(c, vp8_mv_update_prob[i][j]))
506 s->prob->mvc[i][j] = vp8_rac_get_nn(c);
507 }
508
509 return 0;
510}
511
512static inline void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src,
513 int mb_x, int mb_y)
514{
515#define MARGIN (16 << 2)
516 dst->x = av_clip(src->x, -((mb_x << 6) + MARGIN),
517 ((s->mb_width - 1 - mb_x) << 6) + MARGIN);
518 dst->y = av_clip(src->y, -((mb_y << 6) + MARGIN),
519 ((s->mb_height - 1 - mb_y) << 6) + MARGIN);
520}
521
522static void find_near_mvs(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
523 VP56mv near[2], VP56mv *best, int cnt[4])
524{
525 VP8Macroblock *mb_edge[3] = { mb - s->mb_stride /* top */,
526 mb - 1 /* left */,
527 mb - s->mb_stride - 1 /* top-left */ };
528 enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT };
529 VP56mv near_mv[4] = {{ 0 }};
530 enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV };
531 int idx = CNT_ZERO, n;
532 int best_idx = CNT_ZERO;
533
534 /* Process MB on top, left and top-left */
535 for (n = 0; n < 3; n++) {
536 VP8Macroblock *edge = mb_edge[n];
537 if (edge->ref_frame != VP56_FRAME_CURRENT) {
538 if (edge->mv.x | edge->mv.y) {
539 VP56mv tmp = edge->mv;
540 if (s->sign_bias[mb->ref_frame] != s->sign_bias[edge->ref_frame]) {
541 tmp.x *= -1;
542 tmp.y *= -1;
543 }
544 if ((tmp.x ^ near_mv[idx].x) | (tmp.y ^ near_mv[idx].y))
545 near_mv[++idx] = tmp;
546 cnt[idx] += 1 + (n != 2);
547 } else
548 cnt[CNT_ZERO] += 1 + (n != 2);
549 }
550 }
551
552 /* If we have three distinct MV's, merge first and last if they're the same */
553 if (cnt[CNT_SPLITMV] &&
554 !((near_mv[1+EDGE_TOP].x ^ near_mv[1+EDGE_TOPLEFT].x) |
555 (near_mv[1+EDGE_TOP].y ^ near_mv[1+EDGE_TOPLEFT].y)))
556 cnt[CNT_NEAREST] += 1;
557
558 cnt[CNT_SPLITMV] = ((mb_edge[EDGE_LEFT]->mode == VP8_MVMODE_SPLIT) +
559 (mb_edge[EDGE_TOP]->mode == VP8_MVMODE_SPLIT)) * 2 +
560 (mb_edge[EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT);
561
562 /* Swap near and nearest if necessary */
563 if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) {
564 FFSWAP(int, cnt[CNT_NEAREST], cnt[CNT_NEAR]);
565 FFSWAP(VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]);
566 }
567
568 /* Choose the best mv out of 0,0 and the nearest mv */
569 if (cnt[CNT_NEAREST] >= cnt[CNT_ZERO])
570 best_idx = CNT_NEAREST;
571
572 clamp_mv(s, best, &near_mv[best_idx], mb_x, mb_y);
573 near[0] = near_mv[CNT_NEAREST];
574 near[1] = near_mv[CNT_NEAR];
575}
576
577/**
578 * Motion vector coding, 17.1.
579 */
580static int read_mv_component(VP56RangeCoder *c, const uint8_t *p)
581{
582 int x = 0;
583
584 if (vp56_rac_get_prob(c, p[0])) {
585 int i;
586
587 for (i = 0; i < 3; i++)
588 x += vp56_rac_get_prob(c, p[9 + i]) << i;
589 for (i = 9; i > 3; i--)
590 x += vp56_rac_get_prob(c, p[9 + i]) << i;
591 if (!(x & 0xFFF0) || vp56_rac_get_prob(c, p[12]))
592 x += 8;
593 } else
594 x = vp8_rac_get_tree(c, vp8_small_mvtree, &p[2]);
595
596 return (x && vp56_rac_get_prob(c, p[1])) ? -x : x;
597}
598
599static const uint8_t *get_submv_prob(const VP56mv *left, const VP56mv *top)
600{
601 int l_is_zero = !(left->x | left->y);
602 int t_is_zero = !(top->x | top->y);
603 int equal = !((left->x ^ top->x) | (left->y ^ top->y));
604
605 if (equal)
606 return l_is_zero ? vp8_submv_prob[4] : vp8_submv_prob[3];
607 if (t_is_zero)
608 return vp8_submv_prob[2];
609 return l_is_zero ? vp8_submv_prob[1] : vp8_submv_prob[0];
610}
611
612/**
613 * Split motion vector prediction, 16.4.
7ed06b2b 614 * @returns the number of motion vectors parsed (2, 4 or 16)
3b636f21 615 */
7ed06b2b 616static int decode_splitmvs(VP8Context *s, VP56RangeCoder *c,
3b636f21
DC
617 VP8Macroblock *mb, VP56mv *base_mv)
618{
619 int part_idx = mb->partitioning =
620 vp8_rac_get_tree(c, vp8_mbsplit_tree, vp8_mbsplit_prob);
621 int n, num = vp8_mbsplit_count[part_idx];
7ed06b2b
RB
622 const uint8_t *mbsplits = vp8_mbsplits[part_idx],
623 *firstidx = vp8_mbfirstidx[part_idx];
3b636f21
DC
624
625 for (n = 0; n < num; n++) {
7ed06b2b
RB
626 int k = firstidx[n];
627 const VP56mv *left, *above;
628 const uint8_t *submv_prob;
629
630 if (!(k & 3)) {
631 VP8Macroblock *left_mb = &mb[-1];
632 left = &left_mb->bmv[vp8_mbsplits[left_mb->partitioning][k + 3]];
633 } else
634 left = &mb->bmv[mbsplits[k - 1]];
635 if (k <= 3) {
636 VP8Macroblock *above_mb = &mb[-s->mb_stride];
637 above = &above_mb->bmv[vp8_mbsplits[above_mb->partitioning][k + 12]];
638 } else
639 above = &mb->bmv[mbsplits[k - 4]];
640
641 submv_prob = get_submv_prob(left, above);
3b636f21
DC
642
643 switch (vp8_rac_get_tree(c, vp8_submv_ref_tree, submv_prob)) {
644 case VP8_SUBMVMODE_NEW4X4:
7ed06b2b
RB
645 mb->bmv[n].y = base_mv->y + read_mv_component(c, s->prob->mvc[0]);
646 mb->bmv[n].x = base_mv->x + read_mv_component(c, s->prob->mvc[1]);
3b636f21
DC
647 break;
648 case VP8_SUBMVMODE_ZERO4X4:
7ed06b2b
RB
649 mb->bmv[n].x = 0;
650 mb->bmv[n].y = 0;
3b636f21
DC
651 break;
652 case VP8_SUBMVMODE_LEFT4X4:
7ed06b2b 653 mb->bmv[n] = *left;
3b636f21
DC
654 break;
655 case VP8_SUBMVMODE_TOP4X4:
7ed06b2b 656 mb->bmv[n] = *above;
3b636f21
DC
657 break;
658 }
3b636f21 659 }
7ed06b2b
RB
660
661 return num;
3b636f21
DC
662}
663
664static inline void decode_intra4x4_modes(VP56RangeCoder *c, uint8_t *intra4x4,
665 int stride, int keyframe)
666{
667 int x, y, t, l;
668 const uint8_t *ctx = vp8_pred4x4_prob_inter;
669
670 for (y = 0; y < 4; y++) {
671 for (x = 0; x < 4; x++) {
672 if (keyframe) {
673 t = intra4x4[x - stride];
674 l = intra4x4[x - 1];
675 ctx = vp8_pred4x4_prob_intra[t][l];
676 }
677 intra4x4[x] = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx);
678 }
679 intra4x4 += stride;
680 }
681}
682
683static void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
684 uint8_t *intra4x4)
685{
686 VP56RangeCoder *c = &s->c;
687 int n;
688
689 if (s->segmentation.update_map)
690 mb->segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid);
691
692 mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0;
693
694 if (s->keyframe) {
695 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra);
696
697 if (mb->mode == MODE_I4x4) {
698 decode_intra4x4_modes(c, intra4x4, s->b4_stride, 1);
699 } else
700 fill_rectangle(intra4x4, 4, 4, s->b4_stride, vp8_pred4x4_mode[mb->mode], 1);
701
702 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra);
703 mb->ref_frame = VP56_FRAME_CURRENT;
704 } else if (vp56_rac_get_prob(c, s->prob->intra)) {
705 VP56mv near[2], best;
706 int cnt[4] = { 0 };
707 uint8_t p[4];
708
709 // inter MB, 16.2
710 if (vp56_rac_get_prob(c, s->prob->last))
711 mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ?
712 VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN;
713 else
714 mb->ref_frame = VP56_FRAME_PREVIOUS;
715
716 // motion vectors, 16.3
717 find_near_mvs(s, mb, mb_x, mb_y, near, &best, cnt);
718 for (n = 0; n < 4; n++)
719 p[n] = vp8_mode_contexts[cnt[n]][n];
720 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_mvinter, p);
721 switch (mb->mode) {
722 case VP8_MVMODE_SPLIT:
7ed06b2b 723 mb->mv = mb->bmv[decode_splitmvs(s, c, mb, &best) - 1];
3b636f21
DC
724 break;
725 case VP8_MVMODE_ZERO:
726 mb->mv.x = 0;
727 mb->mv.y = 0;
728 break;
729 case VP8_MVMODE_NEAREST:
730 clamp_mv(s, &mb->mv, &near[0], mb_x, mb_y);
731 break;
732 case VP8_MVMODE_NEAR:
733 clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y);
734 break;
735 case VP8_MVMODE_NEW:
736 mb->mv.y = best.y + read_mv_component(c, s->prob->mvc[0]);
737 mb->mv.x = best.x + read_mv_component(c, s->prob->mvc[1]);
738 break;
739 }
740 if (mb->mode != VP8_MVMODE_SPLIT) {
7ed06b2b
RB
741 mb->partitioning = VP8_SPLITMVMODE_NONE;
742 mb->bmv[0] = mb->mv;
3b636f21
DC
743 }
744 } else {
745 // intra MB, 16.1
746 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16);
747
748 if (mb->mode == MODE_I4x4) {
749 decode_intra4x4_modes(c, intra4x4, s->b4_stride, 0);
750 } else
751 fill_rectangle(intra4x4, 4, 4, s->b4_stride, vp8_pred4x4_mode[mb->mode], 1);
752
753 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c);
754 mb->ref_frame = VP56_FRAME_CURRENT;
755 }
756}
757
758/**
e394953e
RB
759 * @param c arithmetic bitstream reader context
760 * @param block destination for block coefficients
761 * @param probs probabilities to use when reading trees from the bitstream
3b636f21
DC
762 * @param i initial coeff index, 0 unless a separate DC block is coded
763 * @param zero_nhood the initial prediction context for number of surrounding
764 * all-zero blocks (only left/top, so 0-2)
3fa76268 765 * @param qmul array holding the dc/ac dequant factor at position 0/1
3b636f21
DC
766 * @return 0 if no coeffs were decoded
767 * otherwise, the index of the last coeff decoded plus one
768 */
769static int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16],
770 uint8_t probs[8][3][NUM_DCT_TOKENS-1],
771 int i, int zero_nhood, int16_t qmul[2])
772{
773 int token, nonzero = 0;
774 int offset = 0;
775
776 for (; i < 16; i++) {
777 token = vp8_rac_get_tree_with_offset(c, vp8_coeff_tree, probs[vp8_coeff_band[i]][zero_nhood], offset);
778
779 if (token == DCT_EOB)
780 break;
781 else if (token >= DCT_CAT1) {
782 int cat = token-DCT_CAT1;
783 token = vp8_rac_get_coeff(c, vp8_dct_cat_prob[cat]);
784 token += vp8_dct_cat_offset[cat];
785 }
786
787 // after the first token, the non-zero prediction context becomes
788 // based on the last decoded coeff
789 if (!token) {
790 zero_nhood = 0;
791 offset = 1;
792 continue;
793 } else if (token == 1)
794 zero_nhood = 1;
795 else
796 zero_nhood = 2;
797
798 // todo: full [16] qmat? load into register?
799 block[zigzag_scan[i]] = (vp8_rac_get(c) ? -token : token) * qmul[!!i];
800 nonzero = i+1;
801 offset = 0;
802 }
803 return nonzero;
804}
805
806static void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,
807 uint8_t t_nnz[9], uint8_t l_nnz[9])
808{
809 LOCAL_ALIGNED_16(DCTELEM, dc,[16]);
810 int i, x, y, luma_start = 0, luma_ctx = 3;
811 int nnz_pred, nnz, nnz_total = 0;
812 int segment = s->segmentation.enabled ? mb->segment : 0;
813
814 s->dsp.clear_blocks((DCTELEM *)s->block);
815
816 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
817 AV_ZERO128(dc);
818 AV_ZERO128(dc+8);
819 nnz_pred = t_nnz[8] + l_nnz[8];
820
821 // decode DC values and do hadamard
822 nnz = decode_block_coeffs(c, dc, s->prob->token[1], 0, nnz_pred,
823 s->qmat[segment].luma_dc_qmul);
824 l_nnz[8] = t_nnz[8] = !!nnz;
825 nnz_total += nnz;
826 s->vp8dsp.vp8_luma_dc_wht(s->block, dc);
827 luma_start = 1;
828 luma_ctx = 0;
829 }
830
831 // luma blocks
832 for (y = 0; y < 4; y++)
833 for (x = 0; x < 4; x++) {
834 nnz_pred = l_nnz[y] + t_nnz[x];
835 nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start,
836 nnz_pred, s->qmat[segment].luma_qmul);
837 // nnz+luma_start may be one more than the actual last index, but we don't care
838 s->non_zero_count_cache[y][x] = nnz + luma_start;
839 t_nnz[x] = l_nnz[y] = !!nnz;
840 nnz_total += nnz;
841 }
842
843 // chroma blocks
844 // TODO: what to do about dimensions? 2nd dim for luma is x,
845 // but for chroma it's (y<<1)|x
846 for (i = 4; i < 6; i++)
847 for (y = 0; y < 2; y++)
848 for (x = 0; x < 2; x++) {
849 nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x];
850 nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0,
851 nnz_pred, s->qmat[segment].chroma_qmul);
852 s->non_zero_count_cache[i][(y<<1)+x] = nnz;
853 t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz;
854 nnz_total += nnz;
855 }
856
857 // if there were no coded coeffs despite the macroblock not being marked skip,
858 // we MUST not do the inner loop filter and should not do IDCT
859 // Since skip isn't used for bitstream prediction, just manually set it.
860 if (!nnz_total)
861 mb->skip = 1;
862}
863
9ac831c2
DC
864static av_always_inline
865void backup_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
866 int linesize, int uvlinesize, int simple)
867{
868 AV_COPY128(top_border, src_y + 15*linesize);
869 if (!simple) {
870 AV_COPY64(top_border+16, src_cb + 7*uvlinesize);
871 AV_COPY64(top_border+24, src_cr + 7*uvlinesize);
872 }
873}
874
875static av_always_inline
876void xchg_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
877 int linesize, int uvlinesize, int mb_x, int mb_y, int mb_width,
878 int simple, int xchg)
879{
880 uint8_t *top_border_m1 = top_border-32; // for TL prediction
881 src_y -= linesize;
882 src_cb -= uvlinesize;
883 src_cr -= uvlinesize;
884
096971e8
MR
885#define XCHG(a,b,xchg) do { \
886 if (xchg) AV_SWAP64(b,a); \
887 else AV_COPY64(b,a); \
888 } while (0)
9ac831c2
DC
889
890 XCHG(top_border_m1+8, src_y-8, xchg);
891 XCHG(top_border, src_y, xchg);
892 XCHG(top_border+8, src_y+8, 1);
070ce7ef 893 if (mb_x < mb_width-1)
9ac831c2 894 XCHG(top_border+32, src_y+16, 1);
070ce7ef 895
9ac831c2
DC
896 // only copy chroma for normal loop filter
897 // or to initialize the top row to 127
898 if (!simple || !mb_y) {
899 XCHG(top_border_m1+16, src_cb-8, xchg);
900 XCHG(top_border_m1+24, src_cr-8, xchg);
901 XCHG(top_border+16, src_cb, 1);
902 XCHG(top_border+24, src_cr, 1);
903 }
904}
905
3b636f21
DC
906static int check_intra_pred_mode(int mode, int mb_x, int mb_y)
907{
908 if (mode == DC_PRED8x8) {
909 if (!(mb_x|mb_y))
910 mode = DC_128_PRED8x8;
911 else if (!mb_y)
912 mode = LEFT_DC_PRED8x8;
913 else if (!mb_x)
914 mode = TOP_DC_PRED8x8;
915 }
916 return mode;
917}
918
919static void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
920 uint8_t *bmode, int mb_x, int mb_y)
921{
922 int x, y, mode, nnz, tr;
923
9ac831c2
DC
924 // for the first row, we need to run xchg_mb_border to init the top edge to 127
925 // otherwise, skip it if we aren't going to deblock
926 if (s->deblock_filter || !mb_y)
927 xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
928 s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
929 s->filter.simple, 1);
930
3b636f21
DC
931 if (mb->mode < MODE_I4x4) {
932 mode = check_intra_pred_mode(mb->mode, mb_x, mb_y);
933 s->hpc.pred16x16[mode](dst[0], s->linesize);
934 } else {
935 uint8_t *ptr = dst[0];
936
937 // all blocks on the right edge of the macroblock use bottom edge
938 // the top macroblock for their topright edge
939 uint8_t *tr_right = ptr - s->linesize + 16;
940
941 // if we're on the right edge of the frame, said edge is extended
942 // from the top macroblock
943 if (mb_x == s->mb_width-1) {
944 tr = tr_right[-1]*0x01010101;
945 tr_right = (uint8_t *)&tr;
946 }
947
948 for (y = 0; y < 4; y++) {
949 uint8_t *topright = ptr + 4 - s->linesize;
950 for (x = 0; x < 4; x++) {
951 if (x == 3)
952 topright = tr_right;
953
954 s->hpc.pred4x4[bmode[x]](ptr+4*x, topright, s->linesize);
955
956 nnz = s->non_zero_count_cache[y][x];
957 if (nnz) {
958 if (nnz == 1)
959 s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize);
960 else
961 s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize);
962 }
963 topright += 4;
964 }
965
966 ptr += 4*s->linesize;
967 bmode += s->b4_stride;
968 }
969 }
970
971 mode = check_intra_pred_mode(s->chroma_pred_mode, mb_x, mb_y);
972 s->hpc.pred8x8[mode](dst[1], s->uvlinesize);
973 s->hpc.pred8x8[mode](dst[2], s->uvlinesize);
9ac831c2
DC
974
975 if (s->deblock_filter || !mb_y)
976 xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
977 s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
978 s->filter.simple, 0);
3b636f21
DC
979}
980
981/**
982 * Generic MC function.
983 *
984 * @param s VP8 decoding context
985 * @param luma 1 for luma (Y) planes, 0 for chroma (Cb/Cr) planes
986 * @param dst target buffer for block data at block position
987 * @param src reference picture buffer at origin (0, 0)
988 * @param mv motion vector (relative to block position) to get pixel data from
989 * @param x_off horizontal position of block from origin (0, 0)
990 * @param y_off vertical position of block from origin (0, 0)
991 * @param block_w width of block (16, 8 or 4)
992 * @param block_h height of block (always same as block_w)
993 * @param width width of src/dst plane data
994 * @param height height of src/dst plane data
995 * @param linesize size of a single line of plane data, including padding
e394953e 996 * @param mc_func motion compensation function pointers (bilinear or sixtap MC)
3b636f21
DC
997 */
998static inline void vp8_mc(VP8Context *s, int luma,
999 uint8_t *dst, uint8_t *src, const VP56mv *mv,
1000 int x_off, int y_off, int block_w, int block_h,
1001 int width, int height, int linesize,
d6f8476b 1002 vp8_mc_func mc_func[3][3])
3b636f21
DC
1003{
1004 static const uint8_t idx[8] = { 0, 1, 2, 1, 2, 1, 2, 1 };
1005 int mx = (mv->x << luma)&7, mx_idx = idx[mx];
1006 int my = (mv->y << luma)&7, my_idx = idx[my];
1007
1008 x_off += mv->x >> (3 - luma);
1009 y_off += mv->y >> (3 - luma);
1010
1011 // edge emulation
1012 src += y_off * linesize + x_off;
1013 if (x_off < 2 || x_off >= width - block_w - 3 ||
1014 y_off < 2 || y_off >= height - block_h - 3) {
1015 ff_emulated_edge_mc(s->edge_emu_buffer, src - 2 * linesize - 2, linesize,
1016 block_w + 5, block_h + 5,
1017 x_off - 2, y_off - 2, width, height);
1018 src = s->edge_emu_buffer + 2 + linesize * 2;
1019 }
1020
d6f8476b 1021 mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);
3b636f21
DC
1022}
1023
7c4dcf81
RB
1024static inline void vp8_mc_part(VP8Context *s, uint8_t *dst[3],
1025 AVFrame *ref_frame, int x_off, int y_off,
1026 int bx_off, int by_off,
1027 int block_w, int block_h,
1028 int width, int height, VP56mv *mv)
1029{
1030 VP56mv uvmv = *mv;
1031
1032 /* Y */
1033 vp8_mc(s, 1, dst[0] + by_off * s->linesize + bx_off,
1034 ref_frame->data[0], mv, x_off + bx_off, y_off + by_off,
1035 block_w, block_h, width, height, s->linesize,
1036 s->put_pixels_tab[block_w == 8]);
1037
1038 /* U/V */
1039 if (s->profile == 3) {
1040 uvmv.x &= ~7;
1041 uvmv.y &= ~7;
1042 }
1043 x_off >>= 1; y_off >>= 1;
1044 bx_off >>= 1; by_off >>= 1;
1045 width >>= 1; height >>= 1;
1046 block_w >>= 1; block_h >>= 1;
1047 vp8_mc(s, 0, dst[1] + by_off * s->uvlinesize + bx_off,
1048 ref_frame->data[1], &uvmv, x_off + bx_off, y_off + by_off,
1049 block_w, block_h, width, height, s->uvlinesize,
1050 s->put_pixels_tab[1 + (block_w == 4)]);
1051 vp8_mc(s, 0, dst[2] + by_off * s->uvlinesize + bx_off,
1052 ref_frame->data[2], &uvmv, x_off + bx_off, y_off + by_off,
1053 block_w, block_h, width, height, s->uvlinesize,
1054 s->put_pixels_tab[1 + (block_w == 4)]);
1055}
1056
d864dee8
JGG
1057/* Fetch pixels for estimated mv 4 macroblocks ahead.
1058 * Optimized for 64-byte cache lines. Inspired by ffh264 prefetch_motion. */
1059static inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int x_off, int y_off, int ref)
1060{
1061 if (mb->ref_frame != VP56_FRAME_CURRENT) {
1062 int mx = mb->mv.x + x_off + 8;
1063 int my = mb->mv.y + y_off;
1064 uint8_t **src= s->framep[mb->ref_frame]->data;
1065 int off= mx + (my + (mb_x&3)*4)*s->linesize + 64;
1066 s->dsp.prefetch(src[0]+off, s->linesize, 4);
1067 off= (mx>>1) + ((my>>1) + (mb_x&7))*s->uvlinesize + 64;
1068 s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1069 }
1070}
1071
3b636f21
DC
1072/**
1073 * Apply motion vectors to prediction buffer, chapter 18.
1074 */
1075static void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
1076 int mb_x, int mb_y)
1077{
1078 int x_off = mb_x << 4, y_off = mb_y << 4;
1079 int width = 16*s->mb_width, height = 16*s->mb_height;
3b636f21 1080
d864dee8
JGG
1081 prefetch_motion(s, mb, mb_x, mb_y, x_off, y_off, VP56_FRAME_PREVIOUS);
1082
3b636f21 1083 if (mb->mode < VP8_MVMODE_SPLIT) {
7c4dcf81
RB
1084 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1085 0, 0, 16, 16, width, height, &mb->mv);
1086 } else switch (mb->partitioning) {
1087 case VP8_SPLITMVMODE_4x4: {
3b636f21 1088 int x, y;
7c4dcf81 1089 VP56mv uvmv;
3b636f21
DC
1090
1091 /* Y */
1092 for (y = 0; y < 4; y++) {
1093 for (x = 0; x < 4; x++) {
1094 vp8_mc(s, 1, dst[0] + 4*y*s->linesize + x*4,
1095 s->framep[mb->ref_frame]->data[0], &mb->bmv[4*y + x],
1096 4*x + x_off, 4*y + y_off, 4, 4,
1097 width, height, s->linesize,
0ef1dbed 1098 s->put_pixels_tab[2]);
3b636f21
DC
1099 }
1100 }
1101
1102 /* U/V */
1103 x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1;
1104 for (y = 0; y < 2; y++) {
1105 for (x = 0; x < 2; x++) {
1106 uvmv.x = mb->bmv[ 2*y * 4 + 2*x ].x +
1107 mb->bmv[ 2*y * 4 + 2*x+1].x +
1108 mb->bmv[(2*y+1) * 4 + 2*x ].x +
1109 mb->bmv[(2*y+1) * 4 + 2*x+1].x;
1110 uvmv.y = mb->bmv[ 2*y * 4 + 2*x ].y +
1111 mb->bmv[ 2*y * 4 + 2*x+1].y +
1112 mb->bmv[(2*y+1) * 4 + 2*x ].y +
1113 mb->bmv[(2*y+1) * 4 + 2*x+1].y;
8f910a56
SG
1114 uvmv.x = (uvmv.x + 2 + (uvmv.x >> (INT_BIT-1))) >> 2;
1115 uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2;
3b636f21
DC
1116 if (s->profile == 3) {
1117 uvmv.x &= ~7;
1118 uvmv.y &= ~7;
1119 }
1120 vp8_mc(s, 0, dst[1] + 4*y*s->uvlinesize + x*4,
1121 s->framep[mb->ref_frame]->data[1], &uvmv,
1122 4*x + x_off, 4*y + y_off, 4, 4,
1123 width, height, s->uvlinesize,
0ef1dbed 1124 s->put_pixels_tab[2]);
3b636f21
DC
1125 vp8_mc(s, 0, dst[2] + 4*y*s->uvlinesize + x*4,
1126 s->framep[mb->ref_frame]->data[2], &uvmv,
1127 4*x + x_off, 4*y + y_off, 4, 4,
1128 width, height, s->uvlinesize,
0ef1dbed 1129 s->put_pixels_tab[2]);
3b636f21
DC
1130 }
1131 }
7c4dcf81
RB
1132 break;
1133 }
1134 case VP8_SPLITMVMODE_16x8:
1135 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1136 0, 0, 16, 8, width, height, &mb->bmv[0]);
1137 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
7ed06b2b 1138 0, 8, 16, 8, width, height, &mb->bmv[1]);
7c4dcf81
RB
1139 break;
1140 case VP8_SPLITMVMODE_8x16:
1141 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1142 0, 0, 8, 16, width, height, &mb->bmv[0]);
1143 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
7ed06b2b 1144 8, 0, 8, 16, width, height, &mb->bmv[1]);
7c4dcf81
RB
1145 break;
1146 case VP8_SPLITMVMODE_8x8:
1147 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1148 0, 0, 8, 8, width, height, &mb->bmv[0]);
1149 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
7ed06b2b 1150 8, 0, 8, 8, width, height, &mb->bmv[1]);
7c4dcf81 1151 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
7ed06b2b 1152 0, 8, 8, 8, width, height, &mb->bmv[2]);
7c4dcf81 1153 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
7ed06b2b 1154 8, 8, 8, 8, width, height, &mb->bmv[3]);
7c4dcf81 1155 break;
3b636f21 1156 }
d864dee8
JGG
1157
1158 prefetch_motion(s, mb, mb_x, mb_y, x_off, y_off, VP56_FRAME_GOLDEN);
3b636f21
DC
1159}
1160
1161static void idct_mb(VP8Context *s, uint8_t *y_dst, uint8_t *u_dst, uint8_t *v_dst,
1162 VP8Macroblock *mb)
1163{
1164 int x, y, nnz;
1165
1166 if (mb->mode != MODE_I4x4)
1167 for (y = 0; y < 4; y++) {
1168 for (x = 0; x < 4; x++) {
1169 nnz = s->non_zero_count_cache[y][x];
1170 if (nnz) {
1171 if (nnz == 1)
1172 s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize);
1173 else
1174 s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize);
1175 }
1176 }
1177 y_dst += 4*s->linesize;
1178 }
1179
1180 for (y = 0; y < 2; y++) {
1181 for (x = 0; x < 2; x++) {
1182 nnz = s->non_zero_count_cache[4][(y<<1)+x];
1183 if (nnz) {
1184 if (nnz == 1)
1185 s->vp8dsp.vp8_idct_dc_add(u_dst+4*x, s->block[4][(y<<1)+x], s->uvlinesize);
1186 else
1187 s->vp8dsp.vp8_idct_add(u_dst+4*x, s->block[4][(y<<1)+x], s->uvlinesize);
1188 }
1189
1190 nnz = s->non_zero_count_cache[5][(y<<1)+x];
1191 if (nnz) {
1192 if (nnz == 1)
1193 s->vp8dsp.vp8_idct_dc_add(v_dst+4*x, s->block[5][(y<<1)+x], s->uvlinesize);
1194 else
1195 s->vp8dsp.vp8_idct_add(v_dst+4*x, s->block[5][(y<<1)+x], s->uvlinesize);
1196 }
1197 }
1198 u_dst += 4*s->uvlinesize;
1199 v_dst += 4*s->uvlinesize;
1200 }
1201}
1202
1203static void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, int *level, int *inner, int *hev_thresh)
1204{
1205 int interior_limit, filter_level;
1206
1207 if (s->segmentation.enabled) {
1208 filter_level = s->segmentation.filter_level[mb->segment];
1209 if (!s->segmentation.absolute_vals)
1210 filter_level += s->filter.level;
1211 } else
1212 filter_level = s->filter.level;
1213
1214 if (s->lf_delta.enabled) {
1215 filter_level += s->lf_delta.ref[mb->ref_frame];
1216
1217 if (mb->ref_frame == VP56_FRAME_CURRENT) {
1218 if (mb->mode == MODE_I4x4)
1219 filter_level += s->lf_delta.mode[0];
1220 } else {
1221 if (mb->mode == VP8_MVMODE_ZERO)
1222 filter_level += s->lf_delta.mode[1];
1223 else if (mb->mode == VP8_MVMODE_SPLIT)
1224 filter_level += s->lf_delta.mode[3];
1225 else
1226 filter_level += s->lf_delta.mode[2];
1227 }
1228 }
1229 filter_level = av_clip(filter_level, 0, 63);
1230
1231 interior_limit = filter_level;
1232 if (s->filter.sharpness) {
1233 interior_limit >>= s->filter.sharpness > 4 ? 2 : 1;
1234 interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness);
1235 }
1236 interior_limit = FFMAX(interior_limit, 1);
1237
1238 *level = filter_level;
1239 *inner = interior_limit;
1240
1241 if (hev_thresh) {
1242 *hev_thresh = filter_level >= 15;
1243
1244 if (s->keyframe) {
1245 if (filter_level >= 40)
1246 *hev_thresh = 2;
1247 } else {
1248 if (filter_level >= 40)
1249 *hev_thresh = 3;
1250 else if (filter_level >= 20)
1251 *hev_thresh = 2;
1252 }
1253 }
1254}
1255
3b636f21
DC
1256static void filter_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb, int mb_x, int mb_y)
1257{
5245c04d 1258 int filter_level, inner_limit, hev_thresh, mbedge_lim, bedge_lim;
3b636f21
DC
1259
1260 filter_level_for_mb(s, mb, &filter_level, &inner_limit, &hev_thresh);
1261 if (!filter_level)
1262 return;
1263
5245c04d
DC
1264 mbedge_lim = 2*(filter_level+2) + inner_limit;
1265 bedge_lim = 2* filter_level + inner_limit;
1266
3b636f21 1267 if (mb_x) {
3facfc99
RB
1268 s->vp8dsp.vp8_h_loop_filter16y(dst[0], s->linesize,
1269 mbedge_lim, inner_limit, hev_thresh);
1270 s->vp8dsp.vp8_h_loop_filter8uv(dst[1], dst[2], s->uvlinesize,
1271 mbedge_lim, inner_limit, hev_thresh);
3b636f21
DC
1272 }
1273
1274 if (!mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT) {
3facfc99
RB
1275 s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 4, s->linesize, bedge_lim,
1276 inner_limit, hev_thresh);
1277 s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 8, s->linesize, bedge_lim,
1278 inner_limit, hev_thresh);
1279 s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+12, s->linesize, bedge_lim,
1280 inner_limit, hev_thresh);
1281 s->vp8dsp.vp8_h_loop_filter8uv_inner(dst[1] + 4, dst[2] + 4,
1282 s->uvlinesize, bedge_lim,
1283 inner_limit, hev_thresh);
3b636f21
DC
1284 }
1285
1286 if (mb_y) {
3facfc99
RB
1287 s->vp8dsp.vp8_v_loop_filter16y(dst[0], s->linesize,
1288 mbedge_lim, inner_limit, hev_thresh);
1289 s->vp8dsp.vp8_v_loop_filter8uv(dst[1], dst[2], s->uvlinesize,
1290 mbedge_lim, inner_limit, hev_thresh);
3b636f21
DC
1291 }
1292
1293 if (!mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT) {
3facfc99
RB
1294 s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 4*s->linesize,
1295 s->linesize, bedge_lim,
1296 inner_limit, hev_thresh);
1297 s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 8*s->linesize,
1298 s->linesize, bedge_lim,
1299 inner_limit, hev_thresh);
1300 s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+12*s->linesize,
1301 s->linesize, bedge_lim,
1302 inner_limit, hev_thresh);
1303 s->vp8dsp.vp8_v_loop_filter8uv_inner(dst[1] + 4 * s->uvlinesize,
1304 dst[2] + 4 * s->uvlinesize,
1305 s->uvlinesize, bedge_lim,
1306 inner_limit, hev_thresh);
3b636f21
DC
1307 }
1308}
1309
1310static void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8Macroblock *mb, int mb_x, int mb_y)
1311{
1312 int filter_level, inner_limit, mbedge_lim, bedge_lim;
1313
1314 filter_level_for_mb(s, mb, &filter_level, &inner_limit, NULL);
1315 if (!filter_level)
1316 return;
1317
1318 mbedge_lim = 2*(filter_level+2) + inner_limit;
1319 bedge_lim = 2* filter_level + inner_limit;
1320
1321 if (mb_x)
1322 s->vp8dsp.vp8_h_loop_filter_simple(dst, s->linesize, mbedge_lim);
1323 if (!mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT) {
1324 s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, s->linesize, bedge_lim);
1325 s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, s->linesize, bedge_lim);
1326 s->vp8dsp.vp8_h_loop_filter_simple(dst+12, s->linesize, bedge_lim);
1327 }
1328
1329 if (mb_y)
1330 s->vp8dsp.vp8_v_loop_filter_simple(dst, s->linesize, mbedge_lim);
1331 if (!mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT) {
1332 s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*s->linesize, s->linesize, bedge_lim);
1333 s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*s->linesize, s->linesize, bedge_lim);
1334 s->vp8dsp.vp8_v_loop_filter_simple(dst+12*s->linesize, s->linesize, bedge_lim);
1335 }
1336}
1337
1338static void filter_mb_row(VP8Context *s, int mb_y)
1339{
1340 VP8Macroblock *mb = s->macroblocks + mb_y*s->mb_stride;
1341 uint8_t *dst[3] = {
1342 s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize,
1343 s->framep[VP56_FRAME_CURRENT]->data[1] + 8*mb_y*s->uvlinesize,
1344 s->framep[VP56_FRAME_CURRENT]->data[2] + 8*mb_y*s->uvlinesize
1345 };
1346 int mb_x;
1347
1348 for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
9ac831c2 1349 backup_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], s->linesize, s->uvlinesize, 0);
3b636f21
DC
1350 filter_mb(s, dst, mb++, mb_x, mb_y);
1351 dst[0] += 16;
1352 dst[1] += 8;
1353 dst[2] += 8;
1354 }
1355}
1356
1357static void filter_mb_row_simple(VP8Context *s, int mb_y)
1358{
1359 uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize;
1360 VP8Macroblock *mb = s->macroblocks + mb_y*s->mb_stride;
1361 int mb_x;
1362
1363 for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
9ac831c2 1364 backup_mb_border(s->top_border[mb_x+1], dst, NULL, NULL, s->linesize, 0, 1);
3b636f21
DC
1365 filter_mb_simple(s, dst, mb++, mb_x, mb_y);
1366 dst += 16;
1367 }
1368}
1369
1370static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
1371 AVPacket *avpkt)
1372{
1373 VP8Context *s = avctx->priv_data;
1374 int ret, mb_x, mb_y, i, y, referenced;
1375 enum AVDiscard skip_thresh;
1376 AVFrame *curframe;
1377
1378 if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0)
1379 return ret;
1380
1381 referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT
1382 || s->update_altref == VP56_FRAME_CURRENT;
1383
1384 skip_thresh = !referenced ? AVDISCARD_NONREF :
1385 !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL;
1386
1387 if (avctx->skip_frame >= skip_thresh) {
1388 s->invisible = 1;
1389 goto skip_decode;
1390 }
9ac831c2 1391 s->deblock_filter = s->filter.level && avctx->skip_loop_filter < skip_thresh;
3b636f21
DC
1392
1393 for (i = 0; i < 4; i++)
1394 if (&s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1395 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1396 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) {
1397 curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i];
1398 break;
1399 }
1400 if (curframe->data[0])
1401 avctx->release_buffer(avctx, curframe);
1402
1403 curframe->key_frame = s->keyframe;
1404 curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE;
1405 curframe->reference = referenced ? 3 : 0;
1406 if ((ret = avctx->get_buffer(avctx, curframe))) {
1407 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n");
1408 return ret;
1409 }
1410
1411 // Given that arithmetic probabilities are updated every frame, it's quite likely
1412 // that the values we have on a random interframe are complete junk if we didn't
1413 // start decode on a keyframe. So just don't display anything rather than junk.
1414 if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] ||
1415 !s->framep[VP56_FRAME_GOLDEN] ||
1416 !s->framep[VP56_FRAME_GOLDEN2])) {
1417 av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n");
1418 return AVERROR_INVALIDDATA;
1419 }
1420
1421 s->linesize = curframe->linesize[0];
1422 s->uvlinesize = curframe->linesize[1];
1423
1424 if (!s->edge_emu_buffer)
1425 s->edge_emu_buffer = av_malloc(21*s->linesize);
1426
1427 memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz));
1428
1429 // top edge of 127 for intra prediction
9ac831c2 1430 memset(s->top_border, 127, (s->mb_width+1)*sizeof(*s->top_border));
3b636f21
DC
1431
1432 for (mb_y = 0; mb_y < s->mb_height; mb_y++) {
1433 VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)];
1434 VP8Macroblock *mb = s->macroblocks + mb_y*s->mb_stride;
1435 uint8_t *intra4x4 = s->intra4x4_pred_mode + 4*mb_y*s->b4_stride;
1436 uint8_t *dst[3] = {
1437 curframe->data[0] + 16*mb_y*s->linesize,
1438 curframe->data[1] + 8*mb_y*s->uvlinesize,
1439 curframe->data[2] + 8*mb_y*s->uvlinesize
1440 };
1441
1442 memset(s->left_nnz, 0, sizeof(s->left_nnz));
1443
1444 // left edge of 129 for intra prediction
1445 if (!(avctx->flags & CODEC_FLAG_EMU_EDGE))
1446 for (i = 0; i < 3; i++)
1447 for (y = 0; y < 16>>!!i; y++)
1448 dst[i][y*curframe->linesize[i]-1] = 129;
9ac831c2
DC
1449 if (mb_y)
1450 memset(s->top_border, 129, sizeof(*s->top_border));
3b636f21
DC
1451
1452 for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
d864dee8
JGG
1453 /* Prefetch the current frame, 4 MBs ahead */
1454 s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4);
1455 s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2);
1456
3b636f21
DC
1457 decode_mb_mode(s, mb, mb_x, mb_y, intra4x4 + 4*mb_x);
1458
1459 if (!mb->skip)
1460 decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);
1461 else {
1462 AV_ZERO128(s->non_zero_count_cache); // luma
1463 AV_ZERO64(s->non_zero_count_cache[4]); // chroma
1464 }
1465
1466 if (mb->mode <= MODE_I4x4) {
1467 intra_predict(s, dst, mb, intra4x4 + 4*mb_x, mb_x, mb_y);
1468 memset(mb->bmv, 0, sizeof(mb->bmv));
1469 } else {
1470 inter_predict(s, dst, mb, mb_x, mb_y);
1471 }
1472
1473 if (!mb->skip) {
1474 idct_mb(s, dst[0], dst[1], dst[2], mb);
1475 } else {
1476 AV_ZERO64(s->left_nnz);
1477 AV_WN64(s->top_nnz[mb_x], 0); // array of 9, so unaligned
1478
1479 // Reset DC block predictors if they would exist if the mb had coefficients
1480 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
1481 s->left_nnz[8] = 0;
1482 s->top_nnz[mb_x][8] = 0;
1483 }
1484 }
1485
1486 dst[0] += 16;
1487 dst[1] += 8;
1488 dst[2] += 8;
1489 mb++;
1490 }
9ac831c2 1491 if (s->deblock_filter) {
3b636f21 1492 if (s->filter.simple)
9ac831c2 1493 filter_mb_row_simple(s, mb_y);
3b636f21 1494 else
9ac831c2 1495 filter_mb_row(s, mb_y);
3b636f21
DC
1496 }
1497 }
3b636f21
DC
1498
1499skip_decode:
1500 // if future frames don't use the updated probabilities,
1501 // reset them to the values we saved
1502 if (!s->update_probabilities)
1503 s->prob[0] = s->prob[1];
1504
1505 // check if golden and altref are swapped
1506 if (s->update_altref == VP56_FRAME_GOLDEN &&
1507 s->update_golden == VP56_FRAME_GOLDEN2)
1508 FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN], s->framep[VP56_FRAME_GOLDEN2]);
1509 else {
1510 if (s->update_altref != VP56_FRAME_NONE)
1511 s->framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref];
1512
1513 if (s->update_golden != VP56_FRAME_NONE)
1514 s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden];
1515 }
1516
1517 if (s->update_last) // move cur->prev
1518 s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT];
1519
1520 // release no longer referenced frames
1521 for (i = 0; i < 4; i++)
1522 if (s->frames[i].data[0] &&
1523 &s->frames[i] != s->framep[VP56_FRAME_CURRENT] &&
1524 &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1525 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1526 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2])
1527 avctx->release_buffer(avctx, &s->frames[i]);
1528
1529 if (!s->invisible) {
1530 *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT];
1531 *data_size = sizeof(AVFrame);
1532 }
1533
1534 return avpkt->size;
1535}
1536
1537static av_cold int vp8_decode_init(AVCodecContext *avctx)
1538{
1539 VP8Context *s = avctx->priv_data;
1540
1541 s->avctx = avctx;
1542 avctx->pix_fmt = PIX_FMT_YUV420P;
1543
1544 dsputil_init(&s->dsp, avctx);
1545 ff_h264_pred_init(&s->hpc, CODEC_ID_VP8);
1546 ff_vp8dsp_init(&s->vp8dsp);
1547
1548 // intra pred needs edge emulation among other things
1549 if (avctx->flags&CODEC_FLAG_EMU_EDGE) {
03ac56e7 1550 av_log(avctx, AV_LOG_ERROR, "Edge emulation not supported\n");
3b636f21
DC
1551 return AVERROR_PATCHWELCOME;
1552 }
1553
1554 return 0;
1555}
1556
1557static av_cold int vp8_decode_free(AVCodecContext *avctx)
1558{
1559 vp8_decode_flush(avctx);
1560 return 0;
1561}
1562
1563AVCodec vp8_decoder = {
1564 "vp8",
1565 AVMEDIA_TYPE_VIDEO,
1566 CODEC_ID_VP8,
1567 sizeof(VP8Context),
1568 vp8_decode_init,
1569 NULL,
1570 vp8_decode_free,
1571 vp8_decode_frame,
1572 CODEC_CAP_DR1,
1573 .flush = vp8_decode_flush,
1574 .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"),
1575};