Statistics
| Branch: | Revision:

ffmpeg / libavcodec / vp8.c @ 3ae079a3

History | View | Annotate | Download (56.6 KB)

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

    
31
typedef struct {
32
    uint8_t filter_level;
33
    uint8_t inner_limit;
34
    uint8_t inner_filter;
35
} VP8FilterStrength;
36

    
37
typedef struct {
38
    uint8_t skip;
39
    // todo: make it possible to check for at least (i4x4 or split_mv)
40
    // in one op. are others needed?
41
    uint8_t mode;
42
    uint8_t ref_frame;
43
    uint8_t partitioning;
44
    VP56mv mv;
45
    VP56mv bmv[16];
46
} VP8Macroblock;
47

    
48
typedef struct {
49
    AVCodecContext *avctx;
50
    DSPContext dsp;
51
    VP8DSPContext vp8dsp;
52
    H264PredContext hpc;
53
    vp8_mc_func put_pixels_tab[3][3][3];
54
    AVFrame frames[4];
55
    AVFrame *framep[4];
56
    uint8_t *edge_emu_buffer;
57
    VP56RangeCoder c;   ///< header context, includes mb modes and motion vectors
58
    int profile;
59

    
60
    int mb_width;   /* number of horizontal MB */
61
    int mb_height;  /* number of vertical MB */
62
    int linesize;
63
    int uvlinesize;
64

    
65
    int keyframe;
66
    int invisible;
67
    int update_last;    ///< update VP56_FRAME_PREVIOUS with the current one
68
    int update_golden;  ///< VP56_FRAME_NONE if not updated, or which frame to copy if so
69
    int update_altref;
70
    int deblock_filter;
71

    
72
    /**
73
     * If this flag is not set, all the probability updates
74
     * are discarded after this frame is decoded.
75
     */
76
    int update_probabilities;
77

    
78
    /**
79
     * All coefficients are contained in separate arith coding contexts.
80
     * There can be 1, 2, 4, or 8 of these after the header context.
81
     */
82
    int num_coeff_partitions;
83
    VP56RangeCoder coeff_partition[8];
84

    
85
    VP8Macroblock *macroblocks;
86
    VP8Macroblock *macroblocks_base;
87
    VP8FilterStrength *filter_strength;
88
    int mb_stride;
89

    
90
    uint8_t *intra4x4_pred_mode;
91
    uint8_t *intra4x4_pred_mode_base;
92
    uint8_t *segmentation_map;
93
    int b4_stride;
94

    
95
    /**
96
     * Cache of the top row needed for intra prediction
97
     * 16 for luma, 8 for each chroma plane
98
     */
99
    uint8_t (*top_border)[16+8+8];
100

    
101
    /**
102
     * For coeff decode, we need to know whether the above block had non-zero
103
     * coefficients. This means for each macroblock, we need data for 4 luma
104
     * blocks, 2 u blocks, 2 v blocks, and the luma dc block, for a total of 9
105
     * per macroblock. We keep the last row in top_nnz.
106
     */
107
    uint8_t (*top_nnz)[9];
108
    DECLARE_ALIGNED(8, uint8_t, left_nnz)[9];
109

    
110
    /**
111
     * This is the index plus one of the last non-zero coeff
112
     * for each of the blocks in the current macroblock.
113
     * So, 0 -> no coeffs
114
     *     1 -> dc-only (special transform)
115
     *     2+-> full transform
116
     */
117
    DECLARE_ALIGNED(16, uint8_t, non_zero_count_cache)[6][4];
118
    DECLARE_ALIGNED(16, DCTELEM, block)[6][4][16];
119
    uint8_t intra4x4_pred_mode_mb[16];
120

    
121
    int chroma_pred_mode;    ///< 8x8c pred mode of the current macroblock
122
    int segment;             ///< segment of the current macroblock
123

    
124
    int mbskip_enabled;
125
    int sign_bias[4]; ///< one state [0, 1] per ref frame type
126
    int ref_count[3];
127

    
128
    /**
129
     * Base parameters for segmentation, i.e. per-macroblock parameters.
130
     * These must be kept unchanged even if segmentation is not used for
131
     * a frame, since the values persist between interframes.
132
     */
133
    struct {
134
        int enabled;
135
        int absolute_vals;
136
        int update_map;
137
        int8_t base_quant[4];
138
        int8_t filter_level[4];     ///< base loop filter level
139
    } segmentation;
140

    
141
    /**
142
     * Macroblocks can have one of 4 different quants in a frame when
143
     * segmentation is enabled.
144
     * If segmentation is disabled, only the first segment's values are used.
145
     */
146
    struct {
147
        // [0] - DC qmul  [1] - AC qmul
148
        int16_t luma_qmul[2];
149
        int16_t luma_dc_qmul[2];    ///< luma dc-only block quant
150
        int16_t chroma_qmul[2];
151
    } qmat[4];
152

    
153
    struct {
154
        int simple;
155
        int level;
156
        int sharpness;
157
    } filter;
158

    
159
    struct {
160
        int enabled;    ///< whether each mb can have a different strength based on mode/ref
161

    
162
        /**
163
         * filter strength adjustment for the following macroblock modes:
164
         * [0] - i4x4
165
         * [1] - zero mv
166
         * [2] - inter modes except for zero or split mv
167
         * [3] - split mv
168
         *  i16x16 modes never have any adjustment
169
         */
170
        int8_t mode[4];
171

    
172
        /**
173
         * filter strength adjustment for macroblocks that reference:
174
         * [0] - intra / VP56_FRAME_CURRENT
175
         * [1] - VP56_FRAME_PREVIOUS
176
         * [2] - VP56_FRAME_GOLDEN
177
         * [3] - altref / VP56_FRAME_GOLDEN2
178
         */
179
        int8_t ref[4];
180
    } lf_delta;
181

    
182
    /**
183
     * These are all of the updatable probabilities for binary decisions.
184
     * They are only implictly reset on keyframes, making it quite likely
185
     * for an interframe to desync if a prior frame's header was corrupt
186
     * or missing outright!
187
     */
188
    struct {
189
        uint8_t segmentid[3];
190
        uint8_t mbskip;
191
        uint8_t intra;
192
        uint8_t last;
193
        uint8_t golden;
194
        uint8_t pred16x16[4];
195
        uint8_t pred8x8c[3];
196
        uint8_t token[4][8][3][NUM_DCT_TOKENS-1];
197
        uint8_t mvc[2][19];
198
    } prob[2];
199
} VP8Context;
200

    
201
#define RL24(p) (AV_RL16(p) + ((p)[2] << 16))
202

    
203
static void vp8_decode_flush(AVCodecContext *avctx)
204
{
205
    VP8Context *s = avctx->priv_data;
206
    int i;
207

    
208
    for (i = 0; i < 4; i++)
209
        if (s->frames[i].data[0])
210
            avctx->release_buffer(avctx, &s->frames[i]);
211
    memset(s->framep, 0, sizeof(s->framep));
212

    
213
    av_freep(&s->macroblocks_base);
214
    av_freep(&s->intra4x4_pred_mode_base);
215
    av_freep(&s->top_nnz);
216
    av_freep(&s->edge_emu_buffer);
217
    av_freep(&s->top_border);
218
    av_freep(&s->segmentation_map);
219

    
220
    s->macroblocks        = NULL;
221
    s->intra4x4_pred_mode = NULL;
222
}
223

    
224
static 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

    
243
    s->macroblocks_base        = av_mallocz((s->mb_stride+s->mb_height*2+2)*sizeof(*s->macroblocks));
244
    s->filter_strength         = av_mallocz(s->mb_stride*sizeof(*s->filter_strength));
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));
247
    s->top_border              = av_mallocz((s->mb_width+1)*sizeof(*s->top_border));
248
    s->segmentation_map        = av_mallocz(s->mb_stride*s->mb_height);
249

    
250
    if (!s->macroblocks_base || !s->filter_strength || !s->intra4x4_pred_mode_base ||
251
        !s->top_nnz || !s->top_border || !s->segmentation_map)
252
        return AVERROR(ENOMEM);
253

    
254
    s->macroblocks        = s->macroblocks_base + 1;
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

    
264
static 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

    
285
static 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

    
297
static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size)
298
{
299
    const uint8_t *sizes = buf;
300
    int i;
301

    
302
    s->num_coeff_partitions = 1 << vp8_rac_get_uint(&s->c, 2);
303

    
304
    buf      += 3*(s->num_coeff_partitions-1);
305
    buf_size -= 3*(s->num_coeff_partitions-1);
306
    if (buf_size < 0)
307
        return -1;
308

    
309
    for (i = 0; i < s->num_coeff_partitions-1; i++) {
310
        int size = RL24(sizes + 3*i);
311
        if (buf_size - size < 0)
312
            return -1;
313

    
314
        vp56_init_range_decoder(&s->coeff_partition[i], buf, size);
315
        buf      += size;
316
        buf_size -= size;
317
    }
318
    vp56_init_range_decoder(&s->coeff_partition[i], buf, buf_size);
319

    
320
    return 0;
321
}
322

    
323
static void get_quants(VP8Context *s)
324
{
325
    VP56RangeCoder *c = &s->c;
326
    int i, base_qi;
327

    
328
    int yac_qi     = vp8_rac_get_uint(c, 7);
329
    int ydc_delta  = vp8_rac_get_sint(c, 4);
330
    int y2dc_delta = vp8_rac_get_sint(c, 4);
331
    int y2ac_delta = vp8_rac_get_sint(c, 4);
332
    int uvdc_delta = vp8_rac_get_sint(c, 4);
333
    int uvac_delta = vp8_rac_get_sint(c, 4);
334

    
335
    for (i = 0; i < 4; i++) {
336
        if (s->segmentation.enabled) {
337
            base_qi = s->segmentation.base_quant[i];
338
            if (!s->segmentation.absolute_vals)
339
                base_qi += yac_qi;
340
        } else
341
            base_qi = yac_qi;
342

    
343
        s->qmat[i].luma_qmul[0]    =       vp8_dc_qlookup[av_clip(base_qi + ydc_delta , 0, 127)];
344
        s->qmat[i].luma_qmul[1]    =       vp8_ac_qlookup[av_clip(base_qi             , 0, 127)];
345
        s->qmat[i].luma_dc_qmul[0] =   2 * vp8_dc_qlookup[av_clip(base_qi + y2dc_delta, 0, 127)];
346
        s->qmat[i].luma_dc_qmul[1] = 155 * vp8_ac_qlookup[av_clip(base_qi + y2ac_delta, 0, 127)] / 100;
347
        s->qmat[i].chroma_qmul[0]  =       vp8_dc_qlookup[av_clip(base_qi + uvdc_delta, 0, 127)];
348
        s->qmat[i].chroma_qmul[1]  =       vp8_ac_qlookup[av_clip(base_qi + uvac_delta, 0, 127)];
349

    
350
        s->qmat[i].luma_dc_qmul[1] = FFMAX(s->qmat[i].luma_dc_qmul[1], 8);
351
        s->qmat[i].chroma_qmul[0]  = FFMIN(s->qmat[i].chroma_qmul[0], 132);
352
    }
353
}
354

    
355
/**
356
 * Determine which buffers golden and altref should be updated with after this frame.
357
 * The spec isn't clear here, so I'm going by my understanding of what libvpx does
358
 *
359
 * Intra frames update all 3 references
360
 * Inter frames update VP56_FRAME_PREVIOUS if the update_last flag is set
361
 * If the update (golden|altref) flag is set, it's updated with the current frame
362
 *      if update_last is set, and VP56_FRAME_PREVIOUS otherwise.
363
 * If the flag is not set, the number read means:
364
 *      0: no update
365
 *      1: VP56_FRAME_PREVIOUS
366
 *      2: update golden with altref, or update altref with golden
367
 */
368
static 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

    
384
static 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

    
395
static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)
396
{
397
    VP56RangeCoder *c = &s->c;
398
    int header_size, hscale, vscale, i, j, k, l, ret;
399
    int width  = s->avctx->width;
400
    int height = s->avctx->height;
401

    
402
    s->keyframe  = !(buf[0] & 1);
403
    s->profile   =  (buf[0]>>1) & 7;
404
    s->invisible = !(buf[0] & 0x10);
405
    header_size  = RL24(buf) >> 5;
406
    buf      += 3;
407
    buf_size -= 3;
408

    
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));
416

    
417
    if (header_size > buf_size - 7*s->keyframe) {
418
        av_log(s->avctx, AV_LOG_ERROR, "Header size larger than data provided\n");
419
        return AVERROR_INVALIDDATA;
420
    }
421

    
422
    if (s->keyframe) {
423
        if (RL24(buf) != 0x2a019d) {
424
            av_log(s->avctx, AV_LOG_ERROR, "Invalid start code 0x%x\n", RL24(buf));
425
            return AVERROR_INVALIDDATA;
426
        }
427
        width  = AV_RL16(buf+3) & 0x3fff;
428
        height = AV_RL16(buf+5) & 0x3fff;
429
        hscale = buf[4] >> 6;
430
        vscale = buf[6] >> 6;
431
        buf      += 7;
432
        buf_size -= 7;
433

    
434
        if (hscale || vscale)
435
            av_log_missing_feature(s->avctx, "Upscaling", 1);
436

    
437
        s->update_golden = s->update_altref = VP56_FRAME_CURRENT;
438
        memcpy(s->prob->token    , vp8_token_default_probs , sizeof(s->prob->token));
439
        memcpy(s->prob->pred16x16, vp8_pred16x16_prob_inter, sizeof(s->prob->pred16x16));
440
        memcpy(s->prob->pred8x8c , vp8_pred8x8c_prob_inter , sizeof(s->prob->pred8x8c));
441
        memcpy(s->prob->mvc      , vp8_mv_default_prob     , sizeof(s->prob->mvc));
442
        memset(&s->segmentation, 0, sizeof(s->segmentation));
443
    }
444

    
445
    if (!s->macroblocks_base || /* first frame */
446
        width != s->avctx->width || height != s->avctx->height) {
447
        if ((ret = update_dimensions(s, width, height) < 0))
448
            return ret;
449
    }
450

    
451
    vp56_init_range_decoder(c, buf, header_size);
452
    buf      += header_size;
453
    buf_size -= header_size;
454

    
455
    if (s->keyframe) {
456
        if (vp8_rac_get(c))
457
            av_log(s->avctx, AV_LOG_WARNING, "Unspecified colorspace\n");
458
        vp8_rac_get(c); // whether we can skip clamping in dsp functions
459
    }
460

    
461
    if ((s->segmentation.enabled = vp8_rac_get(c)))
462
        parse_segment_info(s);
463
    else
464
        s->segmentation.update_map = 0; // FIXME: move this to some init function?
465

    
466
    s->filter.simple    = vp8_rac_get(c);
467
    s->filter.level     = vp8_rac_get_uint(c, 6);
468
    s->filter.sharpness = vp8_rac_get_uint(c, 3);
469

    
470
    if ((s->lf_delta.enabled = vp8_rac_get(c)))
471
        if (vp8_rac_get(c))
472
            update_lf_deltas(s);
473

    
474
    if (setup_partitions(s, buf, buf_size)) {
475
        av_log(s->avctx, AV_LOG_ERROR, "Invalid partitions\n");
476
        return AVERROR_INVALIDDATA;
477
    }
478

    
479
    get_quants(s);
480

    
481
    if (!s->keyframe) {
482
        update_refs(s);
483
        s->sign_bias[VP56_FRAME_GOLDEN]               = vp8_rac_get(c);
484
        s->sign_bias[VP56_FRAME_GOLDEN2 /* altref */] = vp8_rac_get(c);
485
    }
486

    
487
    // if we aren't saving this frame's probabilities for future frames,
488
    // make a copy of the current probabilities
489
    if (!(s->update_probabilities = vp8_rac_get(c)))
490
        s->prob[1] = s->prob[0];
491

    
492
    s->update_last = s->keyframe || vp8_rac_get(c);
493

    
494
    for (i = 0; i < 4; i++)
495
        for (j = 0; j < 8; j++)
496
            for (k = 0; k < 3; k++)
497
                for (l = 0; l < NUM_DCT_TOKENS-1; l++)
498
                    if (vp56_rac_get_prob(c, vp8_token_update_probs[i][j][k][l]))
499
                        s->prob->token[i][j][k][l] = vp8_rac_get_uint(c, 8);
500

    
501
    if ((s->mbskip_enabled = vp8_rac_get(c)))
502
        s->prob->mbskip = vp8_rac_get_uint(c, 8);
503

    
504
    if (!s->keyframe) {
505
        s->prob->intra  = vp8_rac_get_uint(c, 8);
506
        s->prob->last   = vp8_rac_get_uint(c, 8);
507
        s->prob->golden = vp8_rac_get_uint(c, 8);
508

    
509
        if (vp8_rac_get(c))
510
            for (i = 0; i < 4; i++)
511
                s->prob->pred16x16[i] = vp8_rac_get_uint(c, 8);
512
        if (vp8_rac_get(c))
513
            for (i = 0; i < 3; i++)
514
                s->prob->pred8x8c[i]  = vp8_rac_get_uint(c, 8);
515

    
516
        // 17.2 MV probability update
517
        for (i = 0; i < 2; i++)
518
            for (j = 0; j < 19; j++)
519
                if (vp56_rac_get_prob(c, vp8_mv_update_prob[i][j]))
520
                    s->prob->mvc[i][j] = vp8_rac_get_nn(c);
521
    }
522

    
523
    return 0;
524
}
525

    
526
static inline void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src,
527
                            int mb_x, int mb_y)
528
{
529
#define MARGIN (16 << 2)
530
    dst->x = av_clip(src->x, -((mb_x << 6) + MARGIN),
531
                     ((s->mb_width  - 1 - mb_x) << 6) + MARGIN);
532
    dst->y = av_clip(src->y, -((mb_y << 6) + MARGIN),
533
                     ((s->mb_height - 1 - mb_y) << 6) + MARGIN);
534
}
535

    
536
static void find_near_mvs(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
537
                          VP56mv near[2], VP56mv *best, uint8_t cnt[4])
538
{
539
    VP8Macroblock *mb_edge[3] = { mb + 2 /* top */,
540
                                  mb - 1 /* left */,
541
                                  mb + 1 /* top-left */ };
542
    enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT };
543
    VP56mv near_mv[4]  = {{ 0 }};
544
    enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV };
545
    int idx = CNT_ZERO;
546
    int best_idx = CNT_ZERO;
547
    int cur_sign_bias = s->sign_bias[mb->ref_frame];
548
    int *sign_bias = s->sign_bias;
549

    
550
    /* Process MB on top, left and top-left */
551
    #define MV_EDGE_CHECK(n)\
552
    {\
553
        VP8Macroblock *edge = mb_edge[n];\
554
        int edge_ref = edge->ref_frame;\
555
        if (edge_ref != VP56_FRAME_CURRENT) {\
556
            uint32_t mv = AV_RN32A(&edge->mv);\
557
            if (mv) {\
558
                if (cur_sign_bias != sign_bias[edge_ref]) {\
559
                    /* SWAR negate of the values in mv. */\
560
                    mv = ((mv&0x80008000) + 0x00010001) ^ (mv&0x7fff7fff);\
561
                }\
562
                if (!n || mv != AV_RN32A(&near_mv[idx]))\
563
                    AV_WN32A(&near_mv[++idx], mv);\
564
                cnt[idx]      += 1 + (n != 2);\
565
            } else\
566
                cnt[CNT_ZERO] += 1 + (n != 2);\
567
        }\
568
    }
569
    MV_EDGE_CHECK(0)
570
    MV_EDGE_CHECK(1)
571
    MV_EDGE_CHECK(2)
572

    
573
    /* If we have three distinct MVs, merge first and last if they're the same */
574
    if (cnt[CNT_SPLITMV] && AV_RN32A(&near_mv[1+EDGE_TOP]) == AV_RN32A(&near_mv[1+EDGE_TOPLEFT]))
575
        cnt[CNT_NEAREST] += 1;
576

    
577
    cnt[CNT_SPLITMV] = ((mb_edge[EDGE_LEFT]->mode   == VP8_MVMODE_SPLIT) +
578
                        (mb_edge[EDGE_TOP]->mode    == VP8_MVMODE_SPLIT)) * 2 +
579
                       (mb_edge[EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT);
580

    
581
    /* Swap near and nearest if necessary */
582
    if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) {
583
        FFSWAP(uint8_t,     cnt[CNT_NEAREST],     cnt[CNT_NEAR]);
584
        FFSWAP( VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]);
585
    }
586

    
587
    /* Choose the best mv out of 0,0 and the nearest mv */
588
    if (cnt[CNT_NEAREST] >= cnt[CNT_ZERO])
589
        best_idx = CNT_NEAREST;
590

    
591
    clamp_mv(s, best, &near_mv[best_idx], mb_x, mb_y);
592
    near[0] = near_mv[CNT_NEAREST];
593
    near[1] = near_mv[CNT_NEAR];
594
}
595

    
596
/**
597
 * Motion vector coding, 17.1.
598
 */
599
static int read_mv_component(VP56RangeCoder *c, const uint8_t *p)
600
{
601
    int x = 0;
602

    
603
    if (vp56_rac_get_prob(c, p[0])) {
604
        int i;
605

    
606
        for (i = 0; i < 3; i++)
607
            x += vp56_rac_get_prob(c, p[9 + i]) << i;
608
        for (i = 9; i > 3; i--)
609
            x += vp56_rac_get_prob(c, p[9 + i]) << i;
610
        if (!(x & 0xFFF0) || vp56_rac_get_prob(c, p[12]))
611
            x += 8;
612
    } else
613
        x = vp8_rac_get_tree(c, vp8_small_mvtree, &p[2]);
614

    
615
    return (x && vp56_rac_get_prob(c, p[1])) ? -x : x;
616
}
617

    
618
static const uint8_t *get_submv_prob(uint32_t left, uint32_t top)
619
{
620
    if (left == top)
621
        return vp8_submv_prob[4-!!left];
622
    if (!top)
623
        return vp8_submv_prob[2];
624
    return vp8_submv_prob[1-!!left];
625
}
626

    
627
/**
628
 * Split motion vector prediction, 16.4.
629
 * @returns the number of motion vectors parsed (2, 4 or 16)
630
 */
631
static int decode_splitmvs(VP8Context    *s,  VP56RangeCoder *c,
632
                            VP8Macroblock *mb, VP56mv         *base_mv)
633
{
634
    int part_idx = mb->partitioning =
635
        vp8_rac_get_tree(c, vp8_mbsplit_tree, vp8_mbsplit_prob);
636
    int n, num = vp8_mbsplit_count[part_idx];
637
    VP8Macroblock *top_mb  = &mb[2];
638
    VP8Macroblock *left_mb = &mb[-1];
639
    const uint8_t *mbsplits_left = vp8_mbsplits[left_mb->partitioning],
640
                  *mbsplits_top = vp8_mbsplits[top_mb->partitioning],
641
                  *mbsplits_cur = vp8_mbsplits[part_idx],
642
                  *firstidx = vp8_mbfirstidx[part_idx];
643
    VP56mv *top_mv  = top_mb->bmv;
644
    VP56mv *left_mv = left_mb->bmv;
645
    VP56mv *cur_mv  = mb->bmv;
646

    
647
    for (n = 0; n < num; n++) {
648
        int k = firstidx[n];
649
        uint32_t left, above;
650
        const uint8_t *submv_prob;
651

    
652
        if (!(k & 3))
653
            left = AV_RN32A(&left_mv[mbsplits_left[k + 3]]);
654
        else
655
            left  = AV_RN32A(&cur_mv[mbsplits_cur[k - 1]]);
656
        if (k <= 3)
657
            above = AV_RN32A(&top_mv[mbsplits_top[k + 12]]);
658
        else
659
            above = AV_RN32A(&cur_mv[mbsplits_cur[k - 4]]);
660

    
661
        submv_prob = get_submv_prob(left, above);
662

    
663
        switch (vp8_rac_get_tree(c, vp8_submv_ref_tree, submv_prob)) {
664
        case VP8_SUBMVMODE_NEW4X4:
665
            mb->bmv[n].y = base_mv->y + read_mv_component(c, s->prob->mvc[0]);
666
            mb->bmv[n].x = base_mv->x + read_mv_component(c, s->prob->mvc[1]);
667
            break;
668
        case VP8_SUBMVMODE_ZERO4X4:
669
            AV_WN32A(&mb->bmv[n], 0);
670
            break;
671
        case VP8_SUBMVMODE_LEFT4X4:
672
            AV_WN32A(&mb->bmv[n], left);
673
            break;
674
        case VP8_SUBMVMODE_TOP4X4:
675
            AV_WN32A(&mb->bmv[n], above);
676
            break;
677
        }
678
    }
679

    
680
    return num;
681
}
682

    
683
static inline void decode_intra4x4_modes(VP56RangeCoder *c, uint8_t *intra4x4,
684
                                         int stride, int keyframe)
685
{
686
    int x, y, t, l, i;
687

    
688
    if (keyframe) {
689
        const uint8_t *ctx;
690
        for (y = 0; y < 4; y++) {
691
            for (x = 0; x < 4; x++) {
692
                t = intra4x4[x - stride];
693
                l = intra4x4[x - 1];
694
                ctx = vp8_pred4x4_prob_intra[t][l];
695
                intra4x4[x] = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx);
696
            }
697
            intra4x4 += stride;
698
        }
699
    } else {
700
        for (i = 0; i < 16; i++)
701
            intra4x4[i] = vp8_rac_get_tree(c, vp8_pred4x4_tree, vp8_pred4x4_prob_inter);
702
    }
703
}
704

    
705
static void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
706
                           uint8_t *intra4x4, uint8_t *segment)
707
{
708
    VP56RangeCoder *c = &s->c;
709

    
710
    if (s->segmentation.update_map)
711
        *segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid);
712
    s->segment = *segment;
713

    
714
    mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0;
715

    
716
    if (s->keyframe) {
717
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra);
718

    
719
        if (mb->mode == MODE_I4x4) {
720
            decode_intra4x4_modes(c, intra4x4, s->b4_stride, 1);
721
        } else
722
            fill_rectangle(intra4x4, 4, 4, s->b4_stride, vp8_pred4x4_mode[mb->mode], 1);
723

    
724
        s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra);
725
        mb->ref_frame = VP56_FRAME_CURRENT;
726
    } else if (vp56_rac_get_prob(c, s->prob->intra)) {
727
        VP56mv near[2], best;
728
        uint8_t cnt[4] = { 0 };
729
        uint8_t p[4];
730

    
731
        // inter MB, 16.2
732
        if (vp56_rac_get_prob(c, s->prob->last))
733
            mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ?
734
                VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN;
735
        else
736
            mb->ref_frame = VP56_FRAME_PREVIOUS;
737
        s->ref_count[mb->ref_frame-1]++;
738

    
739
        // motion vectors, 16.3
740
        find_near_mvs(s, mb, mb_x, mb_y, near, &best, cnt);
741
        p[0] = vp8_mode_contexts[cnt[0]][0];
742
        p[1] = vp8_mode_contexts[cnt[1]][1];
743
        p[2] = vp8_mode_contexts[cnt[2]][2];
744
        p[3] = vp8_mode_contexts[cnt[3]][3];
745
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_mvinter, p);
746
        switch (mb->mode) {
747
        case VP8_MVMODE_SPLIT:
748
            mb->mv = mb->bmv[decode_splitmvs(s, c, mb, &best) - 1];
749
            break;
750
        case VP8_MVMODE_ZERO:
751
            AV_WN32A(&mb->mv, 0);
752
            break;
753
        case VP8_MVMODE_NEAREST:
754
            clamp_mv(s, &mb->mv, &near[0], mb_x, mb_y);
755
            break;
756
        case VP8_MVMODE_NEAR:
757
            clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y);
758
            break;
759
        case VP8_MVMODE_NEW:
760
            mb->mv.y = best.y + read_mv_component(c, s->prob->mvc[0]);
761
            mb->mv.x = best.x + read_mv_component(c, s->prob->mvc[1]);
762
            break;
763
        }
764
        if (mb->mode != VP8_MVMODE_SPLIT) {
765
            mb->partitioning = VP8_SPLITMVMODE_NONE;
766
            mb->bmv[0] = mb->mv;
767
        }
768
    } else {
769
        // intra MB, 16.1
770
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16);
771

    
772
        if (mb->mode == MODE_I4x4)
773
            decode_intra4x4_modes(c, intra4x4, 4, 0);
774

    
775
        s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c);
776
        mb->ref_frame = VP56_FRAME_CURRENT;
777
        mb->partitioning = VP8_SPLITMVMODE_NONE;
778
        AV_WN32A(&mb->bmv[0], 0);
779
    }
780
}
781

    
782
/**
783
 * @param c arithmetic bitstream reader context
784
 * @param block destination for block coefficients
785
 * @param probs probabilities to use when reading trees from the bitstream
786
 * @param i initial coeff index, 0 unless a separate DC block is coded
787
 * @param zero_nhood the initial prediction context for number of surrounding
788
 *                   all-zero blocks (only left/top, so 0-2)
789
 * @param qmul array holding the dc/ac dequant factor at position 0/1
790
 * @return 0 if no coeffs were decoded
791
 *         otherwise, the index of the last coeff decoded plus one
792
 */
793
static int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16],
794
                               uint8_t probs[8][3][NUM_DCT_TOKENS-1],
795
                               int i, int zero_nhood, int16_t qmul[2])
796
{
797
    int token, nonzero = 0;
798
    int offset = 0;
799

    
800
    for (; i < 16; i++) {
801
        token = vp8_rac_get_tree_with_offset(c, vp8_coeff_tree, probs[vp8_coeff_band[i]][zero_nhood], offset);
802

    
803
        if (token == DCT_EOB)
804
            break;
805
        else if (token >= DCT_CAT1) {
806
            int cat = token-DCT_CAT1;
807
            token = vp8_rac_get_coeff(c, vp8_dct_cat_prob[cat]);
808
            token += 3 + (2<<cat);
809
        }
810

    
811
        // after the first token, the non-zero prediction context becomes
812
        // based on the last decoded coeff
813
        if (!token) {
814
            zero_nhood = 0;
815
            offset = 1;
816
            continue;
817
        } else if (token == 1)
818
            zero_nhood = 1;
819
        else
820
            zero_nhood = 2;
821

    
822
        // todo: full [16] qmat? load into register?
823
        block[zigzag_scan[i]] = (vp8_rac_get(c) ? -token : token) * qmul[!!i];
824
        nonzero = i+1;
825
        offset = 0;
826
    }
827
    return nonzero;
828
}
829

    
830
static void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,
831
                             uint8_t t_nnz[9], uint8_t l_nnz[9])
832
{
833
    LOCAL_ALIGNED_16(DCTELEM, dc,[16]);
834
    int i, x, y, luma_start = 0, luma_ctx = 3;
835
    int nnz_pred, nnz, nnz_total = 0;
836
    int segment = s->segment;
837

    
838
    if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
839
        AV_ZERO128(dc);
840
        AV_ZERO128(dc+8);
841
        nnz_pred = t_nnz[8] + l_nnz[8];
842

    
843
        // decode DC values and do hadamard
844
        nnz = decode_block_coeffs(c, dc, s->prob->token[1], 0, nnz_pred,
845
                                  s->qmat[segment].luma_dc_qmul);
846
        l_nnz[8] = t_nnz[8] = !!nnz;
847
        nnz_total += nnz;
848
        s->vp8dsp.vp8_luma_dc_wht(s->block, dc);
849
        luma_start = 1;
850
        luma_ctx = 0;
851
    }
852

    
853
    // luma blocks
854
    for (y = 0; y < 4; y++)
855
        for (x = 0; x < 4; x++) {
856
            nnz_pred = l_nnz[y] + t_nnz[x];
857
            nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start,
858
                                      nnz_pred, s->qmat[segment].luma_qmul);
859
            // nnz+luma_start may be one more than the actual last index, but we don't care
860
            s->non_zero_count_cache[y][x] = nnz + luma_start;
861
            t_nnz[x] = l_nnz[y] = !!nnz;
862
            nnz_total += nnz;
863
        }
864

    
865
    // chroma blocks
866
    // TODO: what to do about dimensions? 2nd dim for luma is x,
867
    // but for chroma it's (y<<1)|x
868
    for (i = 4; i < 6; i++)
869
        for (y = 0; y < 2; y++)
870
            for (x = 0; x < 2; x++) {
871
                nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x];
872
                nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0,
873
                                          nnz_pred, s->qmat[segment].chroma_qmul);
874
                s->non_zero_count_cache[i][(y<<1)+x] = nnz;
875
                t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz;
876
                nnz_total += nnz;
877
            }
878

    
879
    // if there were no coded coeffs despite the macroblock not being marked skip,
880
    // we MUST not do the inner loop filter and should not do IDCT
881
    // Since skip isn't used for bitstream prediction, just manually set it.
882
    if (!nnz_total)
883
        mb->skip = 1;
884
}
885

    
886
static av_always_inline
887
void backup_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
888
                      int linesize, int uvlinesize, int simple)
889
{
890
    AV_COPY128(top_border, src_y + 15*linesize);
891
    if (!simple) {
892
        AV_COPY64(top_border+16, src_cb + 7*uvlinesize);
893
        AV_COPY64(top_border+24, src_cr + 7*uvlinesize);
894
    }
895
}
896

    
897
static av_always_inline
898
void xchg_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
899
                    int linesize, int uvlinesize, int mb_x, int mb_y, int mb_width,
900
                    int simple, int xchg)
901
{
902
    uint8_t *top_border_m1 = top_border-32;     // for TL prediction
903
    src_y  -=   linesize;
904
    src_cb -= uvlinesize;
905
    src_cr -= uvlinesize;
906

    
907
#define XCHG(a,b,xchg) do {                     \
908
        if (xchg) AV_SWAP64(b,a);               \
909
        else      AV_COPY64(b,a);               \
910
    } while (0)
911

    
912
    XCHG(top_border_m1+8, src_y-8, xchg);
913
    XCHG(top_border,      src_y,   xchg);
914
    XCHG(top_border+8,    src_y+8, 1);
915
    if (mb_x < mb_width-1)
916
        XCHG(top_border+32, src_y+16, 1);
917

    
918
    // only copy chroma for normal loop filter
919
    // or to initialize the top row to 127
920
    if (!simple || !mb_y) {
921
        XCHG(top_border_m1+16, src_cb-8, xchg);
922
        XCHG(top_border_m1+24, src_cr-8, xchg);
923
        XCHG(top_border+16,    src_cb, 1);
924
        XCHG(top_border+24,    src_cr, 1);
925
    }
926
}
927

    
928
static int check_intra_pred_mode(int mode, int mb_x, int mb_y)
929
{
930
    if (mode == DC_PRED8x8) {
931
        if (!(mb_x|mb_y))
932
            mode = DC_128_PRED8x8;
933
        else if (!mb_y)
934
            mode = LEFT_DC_PRED8x8;
935
        else if (!mb_x)
936
            mode = TOP_DC_PRED8x8;
937
    }
938
    return mode;
939
}
940

    
941
static void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
942
                          uint8_t *intra4x4, int mb_x, int mb_y)
943
{
944
    int x, y, mode, nnz, tr;
945

    
946
    // for the first row, we need to run xchg_mb_border to init the top edge to 127
947
    // otherwise, skip it if we aren't going to deblock
948
    if (s->deblock_filter || !mb_y)
949
        xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
950
                       s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
951
                       s->filter.simple, 1);
952

    
953
    if (mb->mode < MODE_I4x4) {
954
        mode = check_intra_pred_mode(mb->mode, mb_x, mb_y);
955
        s->hpc.pred16x16[mode](dst[0], s->linesize);
956
    } else {
957
        uint8_t *ptr = dst[0];
958
        int stride = s->keyframe ? s->b4_stride : 4;
959

    
960
        // all blocks on the right edge of the macroblock use bottom edge
961
        // the top macroblock for their topright edge
962
        uint8_t *tr_right = ptr - s->linesize + 16;
963

    
964
        // if we're on the right edge of the frame, said edge is extended
965
        // from the top macroblock
966
        if (mb_x == s->mb_width-1) {
967
            tr = tr_right[-1]*0x01010101;
968
            tr_right = (uint8_t *)&tr;
969
        }
970

    
971
        if (mb->skip)
972
            AV_ZERO128(s->non_zero_count_cache);
973

    
974
        for (y = 0; y < 4; y++) {
975
            uint8_t *topright = ptr + 4 - s->linesize;
976
            for (x = 0; x < 4; x++) {
977
                if (x == 3)
978
                    topright = tr_right;
979

    
980
                s->hpc.pred4x4[intra4x4[x]](ptr+4*x, topright, s->linesize);
981

    
982
                nnz = s->non_zero_count_cache[y][x];
983
                if (nnz) {
984
                    if (nnz == 1)
985
                        s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize);
986
                    else
987
                        s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize);
988
                }
989
                topright += 4;
990
            }
991

    
992
            ptr   += 4*s->linesize;
993
            intra4x4 += stride;
994
        }
995
    }
996

    
997
    mode = check_intra_pred_mode(s->chroma_pred_mode, mb_x, mb_y);
998
    s->hpc.pred8x8[mode](dst[1], s->uvlinesize);
999
    s->hpc.pred8x8[mode](dst[2], s->uvlinesize);
1000

    
1001
    if (s->deblock_filter || !mb_y)
1002
        xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
1003
                       s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
1004
                       s->filter.simple, 0);
1005
}
1006

    
1007
/**
1008
 * Generic MC function.
1009
 *
1010
 * @param s VP8 decoding context
1011
 * @param luma 1 for luma (Y) planes, 0 for chroma (Cb/Cr) planes
1012
 * @param dst target buffer for block data at block position
1013
 * @param src reference picture buffer at origin (0, 0)
1014
 * @param mv motion vector (relative to block position) to get pixel data from
1015
 * @param x_off horizontal position of block from origin (0, 0)
1016
 * @param y_off vertical position of block from origin (0, 0)
1017
 * @param block_w width of block (16, 8 or 4)
1018
 * @param block_h height of block (always same as block_w)
1019
 * @param width width of src/dst plane data
1020
 * @param height height of src/dst plane data
1021
 * @param linesize size of a single line of plane data, including padding
1022
 * @param mc_func motion compensation function pointers (bilinear or sixtap MC)
1023
 */
1024
static inline void vp8_mc(VP8Context *s, int luma,
1025
                          uint8_t *dst, uint8_t *src, const VP56mv *mv,
1026
                          int x_off, int y_off, int block_w, int block_h,
1027
                          int width, int height, int linesize,
1028
                          vp8_mc_func mc_func[3][3])
1029
{
1030
    if (AV_RN32A(mv)) {
1031
        static const uint8_t idx[8] = { 0, 1, 2, 1, 2, 1, 2, 1 };
1032
        int mx = (mv->x << luma)&7, mx_idx = idx[mx];
1033
        int my = (mv->y << luma)&7, my_idx = idx[my];
1034

    
1035
        x_off += mv->x >> (3 - luma);
1036
        y_off += mv->y >> (3 - luma);
1037

    
1038
        // edge emulation
1039
        src += y_off * linesize + x_off;
1040
        if (x_off < 2 || x_off >= width  - block_w - 3 ||
1041
            y_off < 2 || y_off >= height - block_h - 3) {
1042
            ff_emulated_edge_mc(s->edge_emu_buffer, src - 2 * linesize - 2, linesize,
1043
                                block_w + 5, block_h + 5,
1044
                                x_off - 2, y_off - 2, width, height);
1045
            src = s->edge_emu_buffer + 2 + linesize * 2;
1046
        }
1047
        mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);
1048
    } else
1049
        mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0);
1050
}
1051

    
1052
static inline void vp8_mc_part(VP8Context *s, uint8_t *dst[3],
1053
                               AVFrame *ref_frame, int x_off, int y_off,
1054
                               int bx_off, int by_off,
1055
                               int block_w, int block_h,
1056
                               int width, int height, VP56mv *mv)
1057
{
1058
    VP56mv uvmv = *mv;
1059

    
1060
    /* Y */
1061
    vp8_mc(s, 1, dst[0] + by_off * s->linesize + bx_off,
1062
           ref_frame->data[0], mv, x_off + bx_off, y_off + by_off,
1063
           block_w, block_h, width, height, s->linesize,
1064
           s->put_pixels_tab[block_w == 8]);
1065

    
1066
    /* U/V */
1067
    if (s->profile == 3) {
1068
        uvmv.x &= ~7;
1069
        uvmv.y &= ~7;
1070
    }
1071
    x_off   >>= 1; y_off   >>= 1;
1072
    bx_off  >>= 1; by_off  >>= 1;
1073
    width   >>= 1; height  >>= 1;
1074
    block_w >>= 1; block_h >>= 1;
1075
    vp8_mc(s, 0, dst[1] + by_off * s->uvlinesize + bx_off,
1076
           ref_frame->data[1], &uvmv, x_off + bx_off, y_off + by_off,
1077
           block_w, block_h, width, height, s->uvlinesize,
1078
           s->put_pixels_tab[1 + (block_w == 4)]);
1079
    vp8_mc(s, 0, dst[2] + by_off * s->uvlinesize + bx_off,
1080
           ref_frame->data[2], &uvmv, x_off + bx_off, y_off + by_off,
1081
           block_w, block_h, width, height, s->uvlinesize,
1082
           s->put_pixels_tab[1 + (block_w == 4)]);
1083
}
1084

    
1085
/* Fetch pixels for estimated mv 4 macroblocks ahead.
1086
 * Optimized for 64-byte cache lines.  Inspired by ffh264 prefetch_motion. */
1087
static inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int mb_xy, int ref)
1088
{
1089
    /* Don't prefetch refs that haven't been used very often this frame. */
1090
    if (s->ref_count[ref-1] > (mb_xy >> 5)) {
1091
        int x_off = mb_x << 4, y_off = mb_y << 4;
1092
        int mx = mb->mv.x + x_off + 8;
1093
        int my = mb->mv.y + y_off;
1094
        uint8_t **src= s->framep[ref]->data;
1095
        int off= mx + (my + (mb_x&3)*4)*s->linesize + 64;
1096
        s->dsp.prefetch(src[0]+off, s->linesize, 4);
1097
        off= (mx>>1) + ((my>>1) + (mb_x&7))*s->uvlinesize + 64;
1098
        s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1099
    }
1100
}
1101

    
1102
/**
1103
 * Apply motion vectors to prediction buffer, chapter 18.
1104
 */
1105
static void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
1106
                          int mb_x, int mb_y)
1107
{
1108
    int x_off = mb_x << 4, y_off = mb_y << 4;
1109
    int width = 16*s->mb_width, height = 16*s->mb_height;
1110
    AVFrame *ref = s->framep[mb->ref_frame];
1111
    VP56mv *bmv = mb->bmv;
1112

    
1113
    if (mb->mode < VP8_MVMODE_SPLIT) {
1114
        vp8_mc_part(s, dst, ref, x_off, y_off,
1115
                    0, 0, 16, 16, width, height, &mb->mv);
1116
    } else switch (mb->partitioning) {
1117
    case VP8_SPLITMVMODE_4x4: {
1118
        int x, y;
1119
        VP56mv uvmv;
1120

    
1121
        /* Y */
1122
        for (y = 0; y < 4; y++) {
1123
            for (x = 0; x < 4; x++) {
1124
                vp8_mc(s, 1, dst[0] + 4*y*s->linesize + x*4,
1125
                       ref->data[0], &bmv[4*y + x],
1126
                       4*x + x_off, 4*y + y_off, 4, 4,
1127
                       width, height, s->linesize,
1128
                       s->put_pixels_tab[2]);
1129
            }
1130
        }
1131

    
1132
        /* U/V */
1133
        x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1;
1134
        for (y = 0; y < 2; y++) {
1135
            for (x = 0; x < 2; x++) {
1136
                uvmv.x = mb->bmv[ 2*y    * 4 + 2*x  ].x +
1137
                         mb->bmv[ 2*y    * 4 + 2*x+1].x +
1138
                         mb->bmv[(2*y+1) * 4 + 2*x  ].x +
1139
                         mb->bmv[(2*y+1) * 4 + 2*x+1].x;
1140
                uvmv.y = mb->bmv[ 2*y    * 4 + 2*x  ].y +
1141
                         mb->bmv[ 2*y    * 4 + 2*x+1].y +
1142
                         mb->bmv[(2*y+1) * 4 + 2*x  ].y +
1143
                         mb->bmv[(2*y+1) * 4 + 2*x+1].y;
1144
                uvmv.x = (uvmv.x + 2 + (uvmv.x >> (INT_BIT-1))) >> 2;
1145
                uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2;
1146
                if (s->profile == 3) {
1147
                    uvmv.x &= ~7;
1148
                    uvmv.y &= ~7;
1149
                }
1150
                vp8_mc(s, 0, dst[1] + 4*y*s->uvlinesize + x*4,
1151
                       ref->data[1], &uvmv,
1152
                       4*x + x_off, 4*y + y_off, 4, 4,
1153
                       width, height, s->uvlinesize,
1154
                       s->put_pixels_tab[2]);
1155
                vp8_mc(s, 0, dst[2] + 4*y*s->uvlinesize + x*4,
1156
                       ref->data[2], &uvmv,
1157
                       4*x + x_off, 4*y + y_off, 4, 4,
1158
                       width, height, s->uvlinesize,
1159
                       s->put_pixels_tab[2]);
1160
            }
1161
        }
1162
        break;
1163
    }
1164
    case VP8_SPLITMVMODE_16x8:
1165
        vp8_mc_part(s, dst, ref, x_off, y_off,
1166
                    0, 0, 16, 8, width, height, &bmv[0]);
1167
        vp8_mc_part(s, dst, ref, x_off, y_off,
1168
                    0, 8, 16, 8, width, height, &bmv[1]);
1169
        break;
1170
    case VP8_SPLITMVMODE_8x16:
1171
        vp8_mc_part(s, dst, ref, x_off, y_off,
1172
                    0, 0, 8, 16, width, height, &bmv[0]);
1173
        vp8_mc_part(s, dst, ref, x_off, y_off,
1174
                    8, 0, 8, 16, width, height, &bmv[1]);
1175
        break;
1176
    case VP8_SPLITMVMODE_8x8:
1177
        vp8_mc_part(s, dst, ref, x_off, y_off,
1178
                    0, 0, 8, 8, width, height, &bmv[0]);
1179
        vp8_mc_part(s, dst, ref, x_off, y_off,
1180
                    8, 0, 8, 8, width, height, &bmv[1]);
1181
        vp8_mc_part(s, dst, ref, x_off, y_off,
1182
                    0, 8, 8, 8, width, height, &bmv[2]);
1183
        vp8_mc_part(s, dst, ref, x_off, y_off,
1184
                    8, 8, 8, 8, width, height, &bmv[3]);
1185
        break;
1186
    }
1187
}
1188

    
1189
static void idct_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb)
1190
{
1191
    int x, y, ch;
1192

    
1193
    if (mb->mode != MODE_I4x4) {
1194
        uint8_t *y_dst = dst[0];
1195
        for (y = 0; y < 4; y++) {
1196
            uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[y]);
1197
            if (nnz4) {
1198
                if (nnz4&~0x01010101) {
1199
                    for (x = 0; x < 4; x++) {
1200
                        int nnz = s->non_zero_count_cache[y][x];
1201
                        if (nnz) {
1202
                            if (nnz == 1)
1203
                                s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize);
1204
                            else
1205
                                s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize);
1206
                        }
1207
                    }
1208
                } else {
1209
                    s->vp8dsp.vp8_idct_dc_add4y(y_dst, s->block[y], s->linesize);
1210
                }
1211
            }
1212
            y_dst += 4*s->linesize;
1213
        }
1214
    }
1215

    
1216
    for (ch = 0; ch < 2; ch++) {
1217
        uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[4+ch]);
1218
        if (nnz4) {
1219
            uint8_t *ch_dst = dst[1+ch];
1220
            if (nnz4&~0x01010101) {
1221
                for (y = 0; y < 2; y++) {
1222
                    for (x = 0; x < 2; x++) {
1223
                        int nnz = s->non_zero_count_cache[4+ch][(y<<1)+x];
1224
                        if (nnz) {
1225
                            if (nnz == 1)
1226
                                s->vp8dsp.vp8_idct_dc_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);
1227
                            else
1228
                                s->vp8dsp.vp8_idct_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);
1229
                        }
1230
                    }
1231
                    ch_dst += 4*s->uvlinesize;
1232
                }
1233
            } else {
1234
                s->vp8dsp.vp8_idct_dc_add4uv(ch_dst, s->block[4+ch], s->uvlinesize);
1235
            }
1236
        }
1237
    }
1238
}
1239

    
1240
static void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f )
1241
{
1242
    int interior_limit, filter_level;
1243

    
1244
    if (s->segmentation.enabled) {
1245
        filter_level = s->segmentation.filter_level[s->segment];
1246
        if (!s->segmentation.absolute_vals)
1247
            filter_level += s->filter.level;
1248
    } else
1249
        filter_level = s->filter.level;
1250

    
1251
    if (s->lf_delta.enabled) {
1252
        filter_level += s->lf_delta.ref[mb->ref_frame];
1253

    
1254
        if (mb->ref_frame == VP56_FRAME_CURRENT) {
1255
            if (mb->mode == MODE_I4x4)
1256
                filter_level += s->lf_delta.mode[0];
1257
        } else {
1258
            if (mb->mode == VP8_MVMODE_ZERO)
1259
                filter_level += s->lf_delta.mode[1];
1260
            else if (mb->mode == VP8_MVMODE_SPLIT)
1261
                filter_level += s->lf_delta.mode[3];
1262
            else
1263
                filter_level += s->lf_delta.mode[2];
1264
        }
1265
    }
1266
    filter_level = av_clip(filter_level, 0, 63);
1267

    
1268
    interior_limit = filter_level;
1269
    if (s->filter.sharpness) {
1270
        interior_limit >>= s->filter.sharpness > 4 ? 2 : 1;
1271
        interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness);
1272
    }
1273
    interior_limit = FFMAX(interior_limit, 1);
1274

    
1275
    f->filter_level = filter_level;
1276
    f->inner_limit = interior_limit;
1277
    f->inner_filter = !mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT;
1278
}
1279

    
1280
static void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f, int mb_x, int mb_y)
1281
{
1282
    int mbedge_lim, bedge_lim, hev_thresh;
1283
    int filter_level = f->filter_level;
1284
    int inner_limit = f->inner_limit;
1285
    int inner_filter = f->inner_filter;
1286
    int linesize = s->linesize;
1287
    int uvlinesize = s->uvlinesize;
1288

    
1289
    if (!filter_level)
1290
        return;
1291

    
1292
    mbedge_lim = 2*(filter_level+2) + inner_limit;
1293
     bedge_lim = 2* filter_level    + inner_limit;
1294
    hev_thresh = filter_level >= 15;
1295

    
1296
    if (s->keyframe) {
1297
        if (filter_level >= 40)
1298
            hev_thresh = 2;
1299
    } else {
1300
        if (filter_level >= 40)
1301
            hev_thresh = 3;
1302
        else if (filter_level >= 20)
1303
            hev_thresh = 2;
1304
    }
1305

    
1306
    if (mb_x) {
1307
        s->vp8dsp.vp8_h_loop_filter16y(dst[0],     linesize,
1308
                                       mbedge_lim, inner_limit, hev_thresh);
1309
        s->vp8dsp.vp8_h_loop_filter8uv(dst[1],     dst[2],      uvlinesize,
1310
                                       mbedge_lim, inner_limit, hev_thresh);
1311
    }
1312

    
1313
    if (inner_filter) {
1314
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 4, linesize, bedge_lim,
1315
                                             inner_limit, hev_thresh);
1316
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 8, linesize, bedge_lim,
1317
                                             inner_limit, hev_thresh);
1318
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+12, linesize, bedge_lim,
1319
                                             inner_limit, hev_thresh);
1320
        s->vp8dsp.vp8_h_loop_filter8uv_inner(dst[1] + 4, dst[2] + 4,
1321
                                             uvlinesize,  bedge_lim,
1322
                                             inner_limit, hev_thresh);
1323
    }
1324

    
1325
    if (mb_y) {
1326
        s->vp8dsp.vp8_v_loop_filter16y(dst[0],     linesize,
1327
                                       mbedge_lim, inner_limit, hev_thresh);
1328
        s->vp8dsp.vp8_v_loop_filter8uv(dst[1],     dst[2],      uvlinesize,
1329
                                       mbedge_lim, inner_limit, hev_thresh);
1330
    }
1331

    
1332
    if (inner_filter) {
1333
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 4*linesize,
1334
                                             linesize,    bedge_lim,
1335
                                             inner_limit, hev_thresh);
1336
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 8*linesize,
1337
                                             linesize,    bedge_lim,
1338
                                             inner_limit, hev_thresh);
1339
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+12*linesize,
1340
                                             linesize,    bedge_lim,
1341
                                             inner_limit, hev_thresh);
1342
        s->vp8dsp.vp8_v_loop_filter8uv_inner(dst[1] + 4 * uvlinesize,
1343
                                             dst[2] + 4 * uvlinesize,
1344
                                             uvlinesize,  bedge_lim,
1345
                                             inner_limit, hev_thresh);
1346
    }
1347
}
1348

    
1349
static void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f, int mb_x, int mb_y)
1350
{
1351
    int mbedge_lim, bedge_lim;
1352
    int filter_level = f->filter_level;
1353
    int inner_limit = f->inner_limit;
1354
    int inner_filter = f->inner_filter;
1355
    int linesize = s->linesize;
1356

    
1357
    if (!filter_level)
1358
        return;
1359

    
1360
    mbedge_lim = 2*(filter_level+2) + inner_limit;
1361
     bedge_lim = 2* filter_level    + inner_limit;
1362

    
1363
    if (mb_x)
1364
        s->vp8dsp.vp8_h_loop_filter_simple(dst, linesize, mbedge_lim);
1365
    if (inner_filter) {
1366
        s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, linesize, bedge_lim);
1367
        s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, linesize, bedge_lim);
1368
        s->vp8dsp.vp8_h_loop_filter_simple(dst+12, linesize, bedge_lim);
1369
    }
1370

    
1371
    if (mb_y)
1372
        s->vp8dsp.vp8_v_loop_filter_simple(dst, linesize, mbedge_lim);
1373
    if (inner_filter) {
1374
        s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*linesize, linesize, bedge_lim);
1375
        s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*linesize, linesize, bedge_lim);
1376
        s->vp8dsp.vp8_v_loop_filter_simple(dst+12*linesize, linesize, bedge_lim);
1377
    }
1378
}
1379

    
1380
static void filter_mb_row(VP8Context *s, int mb_y)
1381
{
1382
    VP8FilterStrength *f = s->filter_strength;
1383
    uint8_t *dst[3] = {
1384
        s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize,
1385
        s->framep[VP56_FRAME_CURRENT]->data[1] +  8*mb_y*s->uvlinesize,
1386
        s->framep[VP56_FRAME_CURRENT]->data[2] +  8*mb_y*s->uvlinesize
1387
    };
1388
    int mb_x;
1389

    
1390
    for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1391
        backup_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], s->linesize, s->uvlinesize, 0);
1392
        filter_mb(s, dst, f++, mb_x, mb_y);
1393
        dst[0] += 16;
1394
        dst[1] += 8;
1395
        dst[2] += 8;
1396
    }
1397
}
1398

    
1399
static void filter_mb_row_simple(VP8Context *s, int mb_y)
1400
{
1401
    VP8FilterStrength *f = s->filter_strength;
1402
    uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize;
1403
    int mb_x;
1404

    
1405
    for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1406
        backup_mb_border(s->top_border[mb_x+1], dst, NULL, NULL, s->linesize, 0, 1);
1407
        filter_mb_simple(s, dst, f++, mb_x, mb_y);
1408
        dst += 16;
1409
    }
1410
}
1411

    
1412
static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
1413
                            AVPacket *avpkt)
1414
{
1415
    VP8Context *s = avctx->priv_data;
1416
    int ret, mb_x, mb_y, i, y, referenced;
1417
    enum AVDiscard skip_thresh;
1418
    AVFrame *curframe = NULL;
1419

    
1420
    if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0)
1421
        return ret;
1422

    
1423
    referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT
1424
                                || s->update_altref == VP56_FRAME_CURRENT;
1425

    
1426
    skip_thresh = !referenced ? AVDISCARD_NONREF :
1427
                    !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL;
1428

    
1429
    if (avctx->skip_frame >= skip_thresh) {
1430
        s->invisible = 1;
1431
        goto skip_decode;
1432
    }
1433
    s->deblock_filter = s->filter.level && avctx->skip_loop_filter < skip_thresh;
1434

    
1435
    for (i = 0; i < 4; i++)
1436
        if (&s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1437
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1438
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) {
1439
            curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i];
1440
            break;
1441
        }
1442
    if (curframe->data[0])
1443
        avctx->release_buffer(avctx, curframe);
1444

    
1445
    curframe->key_frame = s->keyframe;
1446
    curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE;
1447
    curframe->reference = referenced ? 3 : 0;
1448
    if ((ret = avctx->get_buffer(avctx, curframe))) {
1449
        av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n");
1450
        return ret;
1451
    }
1452

    
1453
    // Given that arithmetic probabilities are updated every frame, it's quite likely
1454
    // that the values we have on a random interframe are complete junk if we didn't
1455
    // start decode on a keyframe. So just don't display anything rather than junk.
1456
    if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] ||
1457
                         !s->framep[VP56_FRAME_GOLDEN] ||
1458
                         !s->framep[VP56_FRAME_GOLDEN2])) {
1459
        av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n");
1460
        return AVERROR_INVALIDDATA;
1461
    }
1462

    
1463
    s->linesize   = curframe->linesize[0];
1464
    s->uvlinesize = curframe->linesize[1];
1465

    
1466
    if (!s->edge_emu_buffer)
1467
        s->edge_emu_buffer = av_malloc(21*s->linesize);
1468

    
1469
    memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz));
1470

    
1471
    /* Zero macroblock structures for top/left prediction from outside the frame. */
1472
    memset(s->macroblocks, 0, (s->mb_width + s->mb_height*2)*sizeof(*s->macroblocks));
1473

    
1474
    // top edge of 127 for intra prediction
1475
    memset(s->top_border, 127, (s->mb_width+1)*sizeof(*s->top_border));
1476
    memset(s->ref_count, 0, sizeof(s->ref_count));
1477

    
1478
    for (mb_y = 0; mb_y < s->mb_height; mb_y++) {
1479
        VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)];
1480
        VP8Macroblock *mb = s->macroblocks + (s->mb_height - mb_y - 1)*2;
1481
        uint8_t *intra4x4 = s->intra4x4_pred_mode + 4*mb_y*s->b4_stride;
1482
        uint8_t *segment_map = s->segmentation_map + mb_y*s->mb_stride;
1483
        int mb_xy = mb_y * s->mb_stride;
1484
        uint8_t *dst[3] = {
1485
            curframe->data[0] + 16*mb_y*s->linesize,
1486
            curframe->data[1] +  8*mb_y*s->uvlinesize,
1487
            curframe->data[2] +  8*mb_y*s->uvlinesize
1488
        };
1489

    
1490
        memset(s->left_nnz, 0, sizeof(s->left_nnz));
1491

    
1492
        // left edge of 129 for intra prediction
1493
        if (!(avctx->flags & CODEC_FLAG_EMU_EDGE))
1494
            for (i = 0; i < 3; i++)
1495
                for (y = 0; y < 16>>!!i; y++)
1496
                    dst[i][y*curframe->linesize[i]-1] = 129;
1497
        if (mb_y)
1498
            memset(s->top_border, 129, sizeof(*s->top_border));
1499

    
1500
        for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb_xy++, mb++) {
1501
            uint8_t *intra4x4_mb = s->keyframe ? intra4x4 + 4*mb_x : s->intra4x4_pred_mode_mb;
1502
            uint8_t *segment_mb = segment_map+mb_x;
1503

    
1504
            /* Prefetch the current frame, 4 MBs ahead */
1505
            s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4);
1506
            s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2);
1507

    
1508
            decode_mb_mode(s, mb, mb_x, mb_y, intra4x4_mb, segment_mb);
1509

    
1510
            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_PREVIOUS);
1511

    
1512
            if (!mb->skip)
1513
                decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);
1514

    
1515
            if (mb->mode <= MODE_I4x4)
1516
                intra_predict(s, dst, mb, intra4x4_mb, mb_x, mb_y);
1517
            else
1518
                inter_predict(s, dst, mb, mb_x, mb_y);
1519

    
1520
            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN);
1521

    
1522
            if (!mb->skip) {
1523
                idct_mb(s, dst, mb);
1524
            } else {
1525
                AV_ZERO64(s->left_nnz);
1526
                AV_WN64(s->top_nnz[mb_x], 0);   // array of 9, so unaligned
1527

    
1528
                // Reset DC block predictors if they would exist if the mb had coefficients
1529
                if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
1530
                    s->left_nnz[8]      = 0;
1531
                    s->top_nnz[mb_x][8] = 0;
1532
                }
1533
            }
1534

    
1535
            if (s->deblock_filter)
1536
                filter_level_for_mb(s, mb, &s->filter_strength[mb_x]);
1537

    
1538
            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN2);
1539

    
1540
            dst[0] += 16;
1541
            dst[1] += 8;
1542
            dst[2] += 8;
1543
        }
1544
        if (s->deblock_filter) {
1545
            if (s->filter.simple)
1546
                filter_mb_row_simple(s, mb_y);
1547
            else
1548
                filter_mb_row(s, mb_y);
1549
        }
1550
    }
1551

    
1552
skip_decode:
1553
    // if future frames don't use the updated probabilities,
1554
    // reset them to the values we saved
1555
    if (!s->update_probabilities)
1556
        s->prob[0] = s->prob[1];
1557

    
1558
    // check if golden and altref are swapped
1559
    if (s->update_altref == VP56_FRAME_GOLDEN &&
1560
        s->update_golden == VP56_FRAME_GOLDEN2)
1561
        FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN], s->framep[VP56_FRAME_GOLDEN2]);
1562
    else {
1563
        if (s->update_altref != VP56_FRAME_NONE)
1564
            s->framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref];
1565

    
1566
        if (s->update_golden != VP56_FRAME_NONE)
1567
            s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden];
1568
    }
1569

    
1570
    if (s->update_last) // move cur->prev
1571
        s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT];
1572

    
1573
    // release no longer referenced frames
1574
    for (i = 0; i < 4; i++)
1575
        if (s->frames[i].data[0] &&
1576
            &s->frames[i] != s->framep[VP56_FRAME_CURRENT] &&
1577
            &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1578
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1579
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2])
1580
            avctx->release_buffer(avctx, &s->frames[i]);
1581

    
1582
    if (!s->invisible) {
1583
        *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT];
1584
        *data_size = sizeof(AVFrame);
1585
    }
1586

    
1587
    return avpkt->size;
1588
}
1589

    
1590
static av_cold int vp8_decode_init(AVCodecContext *avctx)
1591
{
1592
    VP8Context *s = avctx->priv_data;
1593

    
1594
    s->avctx = avctx;
1595
    avctx->pix_fmt = PIX_FMT_YUV420P;
1596

    
1597
    dsputil_init(&s->dsp, avctx);
1598
    ff_h264_pred_init(&s->hpc, CODEC_ID_VP8);
1599
    ff_vp8dsp_init(&s->vp8dsp);
1600

    
1601
    // intra pred needs edge emulation among other things
1602
    if (avctx->flags&CODEC_FLAG_EMU_EDGE) {
1603
        av_log(avctx, AV_LOG_ERROR, "Edge emulation not supported\n");
1604
        return AVERROR_PATCHWELCOME;
1605
    }
1606

    
1607
    return 0;
1608
}
1609

    
1610
static av_cold int vp8_decode_free(AVCodecContext *avctx)
1611
{
1612
    vp8_decode_flush(avctx);
1613
    return 0;
1614
}
1615

    
1616
AVCodec vp8_decoder = {
1617
    "vp8",
1618
    AVMEDIA_TYPE_VIDEO,
1619
    CODEC_ID_VP8,
1620
    sizeof(VP8Context),
1621
    vp8_decode_init,
1622
    NULL,
1623
    vp8_decode_free,
1624
    vp8_decode_frame,
1625
    CODEC_CAP_DR1,
1626
    .flush = vp8_decode_flush,
1627
    .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"),
1628
};