Statistics
| Branch: | Revision:

ffmpeg / libavcodec / vp8.c @ c55e0d34

History | View | Annotate | Download (56.2 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 segment;
39
    uint8_t skip;
40
    // todo: make it possible to check for at least (i4x4 or split_mv)
41
    // in one op. are others needed?
42
    uint8_t mode;
43
    uint8_t ref_frame;
44
    uint8_t partitioning;
45
    VP56mv mv;
46
    VP56mv bmv[16];
47
} VP8Macroblock;
48

    
49
typedef struct {
50
    AVCodecContext *avctx;
51
    DSPContext dsp;
52
    VP8DSPContext vp8dsp;
53
    H264PredContext hpc;
54
    vp8_mc_func put_pixels_tab[3][3][3];
55
    AVFrame frames[4];
56
    AVFrame *framep[4];
57
    uint8_t *edge_emu_buffer;
58
    VP56RangeCoder c;   ///< header context, includes mb modes and motion vectors
59
    int profile;
60

    
61
    int mb_width;   /* number of horizontal MB */
62
    int mb_height;  /* number of vertical MB */
63
    int linesize;
64
    int uvlinesize;
65

    
66
    int keyframe;
67
    int invisible;
68
    int update_last;    ///< update VP56_FRAME_PREVIOUS with the current one
69
    int update_golden;  ///< VP56_FRAME_NONE if not updated, or which frame to copy if so
70
    int update_altref;
71
    int deblock_filter;
72

    
73
    /**
74
     * If this flag is not set, all the probability updates
75
     * are discarded after this frame is decoded.
76
     */
77
    int update_probabilities;
78

    
79
    /**
80
     * All coefficients are contained in separate arith coding contexts.
81
     * There can be 1, 2, 4, or 8 of these after the header context.
82
     */
83
    int num_coeff_partitions;
84
    VP56RangeCoder coeff_partition[8];
85

    
86
    VP8Macroblock *macroblocks;
87
    VP8Macroblock *macroblocks_base;
88
    VP8FilterStrength *filter_strength;
89
    int mb_stride;
90

    
91
    uint8_t *intra4x4_pred_mode;
92
    uint8_t *intra4x4_pred_mode_base;
93
    uint8_t *segmentation_map;
94
    int b4_stride;
95

    
96
    /**
97
     * Cache of the top row needed for intra prediction
98
     * 16 for luma, 8 for each chroma plane
99
     */
100
    uint8_t (*top_border)[16+8+8];
101

    
102
    /**
103
     * For coeff decode, we need to know whether the above block had non-zero
104
     * coefficients. This means for each macroblock, we need data for 4 luma
105
     * blocks, 2 u blocks, 2 v blocks, and the luma dc block, for a total of 9
106
     * per macroblock. We keep the last row in top_nnz.
107
     */
108
    uint8_t (*top_nnz)[9];
109
    DECLARE_ALIGNED(8, uint8_t, left_nnz)[9];
110

    
111
    /**
112
     * This is the index plus one of the last non-zero coeff
113
     * for each of the blocks in the current macroblock.
114
     * So, 0 -> no coeffs
115
     *     1 -> dc-only (special transform)
116
     *     2+-> full transform
117
     */
118
    DECLARE_ALIGNED(16, uint8_t, non_zero_count_cache)[6][4];
119
    DECLARE_ALIGNED(16, DCTELEM, block)[6][4][16];
120
    uint8_t intra4x4_pred_mode_mb[16];
121

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

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

    
127
    /**
128
     * Base parameters for segmentation, i.e. per-macroblock parameters.
129
     * These must be kept unchanged even if segmentation is not used for
130
     * a frame, since the values persist between interframes.
131
     */
132
    struct {
133
        int enabled;
134
        int absolute_vals;
135
        int update_map;
136
        int8_t base_quant[4];
137
        int8_t filter_level[4];     ///< base loop filter level
138
    } segmentation;
139

    
140
    /**
141
     * Macroblocks can have one of 4 different quants in a frame when
142
     * segmentation is enabled.
143
     * If segmentation is disabled, only the first segment's values are used.
144
     */
145
    struct {
146
        // [0] - DC qmul  [1] - AC qmul
147
        int16_t luma_qmul[2];
148
        int16_t luma_dc_qmul[2];    ///< luma dc-only block quant
149
        int16_t chroma_qmul[2];
150
    } qmat[4];
151

    
152
    struct {
153
        int simple;
154
        int level;
155
        int sharpness;
156
    } filter;
157

    
158
    struct {
159
        int enabled;    ///< whether each mb can have a different strength based on mode/ref
160

    
161
        /**
162
         * filter strength adjustment for the following macroblock modes:
163
         * [0] - i4x4
164
         * [1] - zero mv
165
         * [2] - inter modes except for zero or split mv
166
         * [3] - split mv
167
         *  i16x16 modes never have any adjustment
168
         */
169
        int8_t mode[4];
170

    
171
        /**
172
         * filter strength adjustment for macroblocks that reference:
173
         * [0] - intra / VP56_FRAME_CURRENT
174
         * [1] - VP56_FRAME_PREVIOUS
175
         * [2] - VP56_FRAME_GOLDEN
176
         * [3] - altref / VP56_FRAME_GOLDEN2
177
         */
178
        int8_t ref[4];
179
    } lf_delta;
180

    
181
    /**
182
     * These are all of the updatable probabilities for binary decisions.
183
     * They are only implictly reset on keyframes, making it quite likely
184
     * for an interframe to desync if a prior frame's header was corrupt
185
     * or missing outright!
186
     */
187
    struct {
188
        uint8_t segmentid[3];
189
        uint8_t mbskip;
190
        uint8_t intra;
191
        uint8_t last;
192
        uint8_t golden;
193
        uint8_t pred16x16[4];
194
        uint8_t pred8x8c[3];
195
        uint8_t token[4][8][3][NUM_DCT_TOKENS-1];
196
        uint8_t mvc[2][19];
197
    } prob[2];
198
} VP8Context;
199

    
200
#define RL24(p) (AV_RL16(p) + ((p)[2] << 16))
201

    
202
static void vp8_decode_flush(AVCodecContext *avctx)
203
{
204
    VP8Context *s = avctx->priv_data;
205
    int i;
206

    
207
    for (i = 0; i < 4; i++)
208
        if (s->frames[i].data[0])
209
            avctx->release_buffer(avctx, &s->frames[i]);
210
    memset(s->framep, 0, sizeof(s->framep));
211

    
212
    av_freep(&s->macroblocks_base);
213
    av_freep(&s->intra4x4_pred_mode_base);
214
    av_freep(&s->top_nnz);
215
    av_freep(&s->edge_emu_buffer);
216
    av_freep(&s->top_border);
217
    av_freep(&s->segmentation_map);
218

    
219
    s->macroblocks        = NULL;
220
    s->intra4x4_pred_mode = NULL;
221
}
222

    
223
static int update_dimensions(VP8Context *s, int width, int height)
224
{
225
    int i;
226

    
227
    if (avcodec_check_dimensions(s->avctx, width, height))
228
        return AVERROR_INVALIDDATA;
229

    
230
    vp8_decode_flush(s->avctx);
231

    
232
    avcodec_set_dimensions(s->avctx, width, height);
233

    
234
    s->mb_width  = (s->avctx->coded_width +15) / 16;
235
    s->mb_height = (s->avctx->coded_height+15) / 16;
236

    
237
    // we allocate a border around the top/left of intra4x4 modes
238
    // this is 4 blocks for intra4x4 to keep 4-byte alignment for fill_rectangle
239
    s->mb_stride = s->mb_width+1;
240
    s->b4_stride = 4*s->mb_stride;
241

    
242
    s->macroblocks_base        = av_mallocz((s->mb_stride+s->mb_height*2+2)*sizeof(*s->macroblocks));
243
    s->filter_strength         = av_mallocz(s->mb_stride*sizeof(*s->filter_strength));
244
    s->intra4x4_pred_mode_base = av_mallocz(s->b4_stride*(4*s->mb_height+1));
245
    s->top_nnz                 = av_mallocz(s->mb_width*sizeof(*s->top_nnz));
246
    s->top_border              = av_mallocz((s->mb_width+1)*sizeof(*s->top_border));
247
    s->segmentation_map        = av_mallocz(s->mb_stride*s->mb_height);
248

    
249
    if (!s->macroblocks_base || !s->filter_strength || !s->intra4x4_pred_mode_base ||
250
        !s->top_nnz || !s->top_border || !s->segmentation_map)
251
        return AVERROR(ENOMEM);
252

    
253
    s->macroblocks        = s->macroblocks_base + 1;
254
    s->intra4x4_pred_mode = s->intra4x4_pred_mode_base + 4 + s->b4_stride;
255

    
256
    memset(s->intra4x4_pred_mode_base, DC_PRED, s->b4_stride);
257
    for (i = 0; i < 4*s->mb_height; i++)
258
        s->intra4x4_pred_mode[i*s->b4_stride-1] = DC_PRED;
259

    
260
    return 0;
261
}
262

    
263
static void parse_segment_info(VP8Context *s)
264
{
265
    VP56RangeCoder *c = &s->c;
266
    int i;
267

    
268
    s->segmentation.update_map = vp8_rac_get(c);
269

    
270
    if (vp8_rac_get(c)) { // update segment feature data
271
        s->segmentation.absolute_vals = vp8_rac_get(c);
272

    
273
        for (i = 0; i < 4; i++)
274
            s->segmentation.base_quant[i]   = vp8_rac_get_sint(c, 7);
275

    
276
        for (i = 0; i < 4; i++)
277
            s->segmentation.filter_level[i] = vp8_rac_get_sint(c, 6);
278
    }
279
    if (s->segmentation.update_map)
280
        for (i = 0; i < 3; i++)
281
            s->prob->segmentid[i] = vp8_rac_get(c) ? vp8_rac_get_uint(c, 8) : 255;
282
}
283

    
284
static void update_lf_deltas(VP8Context *s)
285
{
286
    VP56RangeCoder *c = &s->c;
287
    int i;
288

    
289
    for (i = 0; i < 4; i++)
290
        s->lf_delta.ref[i]  = vp8_rac_get_sint(c, 6);
291

    
292
    for (i = 0; i < 4; i++)
293
        s->lf_delta.mode[i] = vp8_rac_get_sint(c, 6);
294
}
295

    
296
static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size)
297
{
298
    const uint8_t *sizes = buf;
299
    int i;
300

    
301
    s->num_coeff_partitions = 1 << vp8_rac_get_uint(&s->c, 2);
302

    
303
    buf      += 3*(s->num_coeff_partitions-1);
304
    buf_size -= 3*(s->num_coeff_partitions-1);
305
    if (buf_size < 0)
306
        return -1;
307

    
308
    for (i = 0; i < s->num_coeff_partitions-1; i++) {
309
        int size = RL24(sizes + 3*i);
310
        if (buf_size - size < 0)
311
            return -1;
312

    
313
        vp56_init_range_decoder(&s->coeff_partition[i], buf, size);
314
        buf      += size;
315
        buf_size -= size;
316
    }
317
    vp56_init_range_decoder(&s->coeff_partition[i], buf, buf_size);
318

    
319
    return 0;
320
}
321

    
322
static void get_quants(VP8Context *s)
323
{
324
    VP56RangeCoder *c = &s->c;
325
    int i, base_qi;
326

    
327
    int yac_qi     = vp8_rac_get_uint(c, 7);
328
    int ydc_delta  = vp8_rac_get_sint(c, 4);
329
    int y2dc_delta = vp8_rac_get_sint(c, 4);
330
    int y2ac_delta = vp8_rac_get_sint(c, 4);
331
    int uvdc_delta = vp8_rac_get_sint(c, 4);
332
    int uvac_delta = vp8_rac_get_sint(c, 4);
333

    
334
    for (i = 0; i < 4; i++) {
335
        if (s->segmentation.enabled) {
336
            base_qi = s->segmentation.base_quant[i];
337
            if (!s->segmentation.absolute_vals)
338
                base_qi += yac_qi;
339
        } else
340
            base_qi = yac_qi;
341

    
342
        s->qmat[i].luma_qmul[0]    =       vp8_dc_qlookup[av_clip(base_qi + ydc_delta , 0, 127)];
343
        s->qmat[i].luma_qmul[1]    =       vp8_ac_qlookup[av_clip(base_qi             , 0, 127)];
344
        s->qmat[i].luma_dc_qmul[0] =   2 * vp8_dc_qlookup[av_clip(base_qi + y2dc_delta, 0, 127)];
345
        s->qmat[i].luma_dc_qmul[1] = 155 * vp8_ac_qlookup[av_clip(base_qi + y2ac_delta, 0, 127)] / 100;
346
        s->qmat[i].chroma_qmul[0]  =       vp8_dc_qlookup[av_clip(base_qi + uvdc_delta, 0, 127)];
347
        s->qmat[i].chroma_qmul[1]  =       vp8_ac_qlookup[av_clip(base_qi + uvac_delta, 0, 127)];
348

    
349
        s->qmat[i].luma_dc_qmul[1] = FFMAX(s->qmat[i].luma_dc_qmul[1], 8);
350
        s->qmat[i].chroma_qmul[0]  = FFMIN(s->qmat[i].chroma_qmul[0], 132);
351
    }
352
}
353

    
354
/**
355
 * Determine which buffers golden and altref should be updated with after this frame.
356
 * The spec isn't clear here, so I'm going by my understanding of what libvpx does
357
 *
358
 * Intra frames update all 3 references
359
 * Inter frames update VP56_FRAME_PREVIOUS if the update_last flag is set
360
 * If the update (golden|altref) flag is set, it's updated with the current frame
361
 *      if update_last is set, and VP56_FRAME_PREVIOUS otherwise.
362
 * If the flag is not set, the number read means:
363
 *      0: no update
364
 *      1: VP56_FRAME_PREVIOUS
365
 *      2: update golden with altref, or update altref with golden
366
 */
367
static VP56Frame ref_to_update(VP8Context *s, int update, VP56Frame ref)
368
{
369
    VP56RangeCoder *c = &s->c;
370

    
371
    if (update)
372
        return VP56_FRAME_CURRENT;
373

    
374
    switch (vp8_rac_get_uint(c, 2)) {
375
    case 1:
376
        return VP56_FRAME_PREVIOUS;
377
    case 2:
378
        return (ref == VP56_FRAME_GOLDEN) ? VP56_FRAME_GOLDEN2 : VP56_FRAME_GOLDEN;
379
    }
380
    return VP56_FRAME_NONE;
381
}
382

    
383
static void update_refs(VP8Context *s)
384
{
385
    VP56RangeCoder *c = &s->c;
386

    
387
    int update_golden = vp8_rac_get(c);
388
    int update_altref = vp8_rac_get(c);
389

    
390
    s->update_golden = ref_to_update(s, update_golden, VP56_FRAME_GOLDEN);
391
    s->update_altref = ref_to_update(s, update_altref, VP56_FRAME_GOLDEN2);
392
}
393

    
394
static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)
395
{
396
    VP56RangeCoder *c = &s->c;
397
    int header_size, hscale, vscale, i, j, k, l, ret;
398
    int width  = s->avctx->width;
399
    int height = s->avctx->height;
400

    
401
    s->keyframe  = !(buf[0] & 1);
402
    s->profile   =  (buf[0]>>1) & 7;
403
    s->invisible = !(buf[0] & 0x10);
404
    header_size  = RL24(buf) >> 5;
405
    buf      += 3;
406
    buf_size -= 3;
407

    
408
    if (s->profile > 3)
409
        av_log(s->avctx, AV_LOG_WARNING, "Unknown profile %d\n", s->profile);
410

    
411
    if (!s->profile)
412
        memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_epel_pixels_tab, sizeof(s->put_pixels_tab));
413
    else    // profile 1-3 use bilinear, 4+ aren't defined so whatever
414
        memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_bilinear_pixels_tab, sizeof(s->put_pixels_tab));
415

    
416
    if (header_size > buf_size - 7*s->keyframe) {
417
        av_log(s->avctx, AV_LOG_ERROR, "Header size larger than data provided\n");
418
        return AVERROR_INVALIDDATA;
419
    }
420

    
421
    if (s->keyframe) {
422
        if (RL24(buf) != 0x2a019d) {
423
            av_log(s->avctx, AV_LOG_ERROR, "Invalid start code 0x%x\n", RL24(buf));
424
            return AVERROR_INVALIDDATA;
425
        }
426
        width  = AV_RL16(buf+3) & 0x3fff;
427
        height = AV_RL16(buf+5) & 0x3fff;
428
        hscale = buf[4] >> 6;
429
        vscale = buf[6] >> 6;
430
        buf      += 7;
431
        buf_size -= 7;
432

    
433
        if (hscale || vscale)
434
            av_log_missing_feature(s->avctx, "Upscaling", 1);
435

    
436
        s->update_golden = s->update_altref = VP56_FRAME_CURRENT;
437
        memcpy(s->prob->token    , vp8_token_default_probs , sizeof(s->prob->token));
438
        memcpy(s->prob->pred16x16, vp8_pred16x16_prob_inter, sizeof(s->prob->pred16x16));
439
        memcpy(s->prob->pred8x8c , vp8_pred8x8c_prob_inter , sizeof(s->prob->pred8x8c));
440
        memcpy(s->prob->mvc      , vp8_mv_default_prob     , sizeof(s->prob->mvc));
441
        memset(&s->segmentation, 0, sizeof(s->segmentation));
442
    }
443

    
444
    if (!s->macroblocks_base || /* first frame */
445
        width != s->avctx->width || height != s->avctx->height) {
446
        if ((ret = update_dimensions(s, width, height) < 0))
447
            return ret;
448
    }
449

    
450
    vp56_init_range_decoder(c, buf, header_size);
451
    buf      += header_size;
452
    buf_size -= header_size;
453

    
454
    if (s->keyframe) {
455
        if (vp8_rac_get(c))
456
            av_log(s->avctx, AV_LOG_WARNING, "Unspecified colorspace\n");
457
        vp8_rac_get(c); // whether we can skip clamping in dsp functions
458
    }
459

    
460
    if ((s->segmentation.enabled = vp8_rac_get(c)))
461
        parse_segment_info(s);
462
    else
463
        s->segmentation.update_map = 0; // FIXME: move this to some init function?
464

    
465
    s->filter.simple    = vp8_rac_get(c);
466
    s->filter.level     = vp8_rac_get_uint(c, 6);
467
    s->filter.sharpness = vp8_rac_get_uint(c, 3);
468

    
469
    if ((s->lf_delta.enabled = vp8_rac_get(c)))
470
        if (vp8_rac_get(c))
471
            update_lf_deltas(s);
472

    
473
    if (setup_partitions(s, buf, buf_size)) {
474
        av_log(s->avctx, AV_LOG_ERROR, "Invalid partitions\n");
475
        return AVERROR_INVALIDDATA;
476
    }
477

    
478
    get_quants(s);
479

    
480
    if (!s->keyframe) {
481
        update_refs(s);
482
        s->sign_bias[VP56_FRAME_GOLDEN]               = vp8_rac_get(c);
483
        s->sign_bias[VP56_FRAME_GOLDEN2 /* altref */] = vp8_rac_get(c);
484
    }
485

    
486
    // if we aren't saving this frame's probabilities for future frames,
487
    // make a copy of the current probabilities
488
    if (!(s->update_probabilities = vp8_rac_get(c)))
489
        s->prob[1] = s->prob[0];
490

    
491
    s->update_last = s->keyframe || vp8_rac_get(c);
492

    
493
    for (i = 0; i < 4; i++)
494
        for (j = 0; j < 8; j++)
495
            for (k = 0; k < 3; k++)
496
                for (l = 0; l < NUM_DCT_TOKENS-1; l++)
497
                    if (vp56_rac_get_prob(c, vp8_token_update_probs[i][j][k][l]))
498
                        s->prob->token[i][j][k][l] = vp8_rac_get_uint(c, 8);
499

    
500
    if ((s->mbskip_enabled = vp8_rac_get(c)))
501
        s->prob->mbskip = vp8_rac_get_uint(c, 8);
502

    
503
    if (!s->keyframe) {
504
        s->prob->intra  = vp8_rac_get_uint(c, 8);
505
        s->prob->last   = vp8_rac_get_uint(c, 8);
506
        s->prob->golden = vp8_rac_get_uint(c, 8);
507

    
508
        if (vp8_rac_get(c))
509
            for (i = 0; i < 4; i++)
510
                s->prob->pred16x16[i] = vp8_rac_get_uint(c, 8);
511
        if (vp8_rac_get(c))
512
            for (i = 0; i < 3; i++)
513
                s->prob->pred8x8c[i]  = vp8_rac_get_uint(c, 8);
514

    
515
        // 17.2 MV probability update
516
        for (i = 0; i < 2; i++)
517
            for (j = 0; j < 19; j++)
518
                if (vp56_rac_get_prob(c, vp8_mv_update_prob[i][j]))
519
                    s->prob->mvc[i][j] = vp8_rac_get_nn(c);
520
    }
521

    
522
    return 0;
523
}
524

    
525
static inline void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src,
526
                            int mb_x, int mb_y)
527
{
528
#define MARGIN (16 << 2)
529
    dst->x = av_clip(src->x, -((mb_x << 6) + MARGIN),
530
                     ((s->mb_width  - 1 - mb_x) << 6) + MARGIN);
531
    dst->y = av_clip(src->y, -((mb_y << 6) + MARGIN),
532
                     ((s->mb_height - 1 - mb_y) << 6) + MARGIN);
533
}
534

    
535
static void find_near_mvs(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
536
                          VP56mv near[2], VP56mv *best, uint8_t cnt[4])
537
{
538
    VP8Macroblock *mb_edge[3] = { mb + 2 /* top */,
539
                                  mb - 1 /* left */,
540
                                  mb + 1 /* top-left */ };
541
    enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT };
542
    VP56mv near_mv[4]  = {{ 0 }};
543
    enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV };
544
    int idx = CNT_ZERO;
545
    int best_idx = CNT_ZERO;
546
    int cur_sign_bias = s->sign_bias[mb->ref_frame];
547
    int *sign_bias = s->sign_bias;
548

    
549
    /* Process MB on top, left and top-left */
550
    #define MV_EDGE_CHECK(n)\
551
    {\
552
        VP8Macroblock *edge = mb_edge[n];\
553
        int edge_ref = edge->ref_frame;\
554
        if (edge_ref != VP56_FRAME_CURRENT) {\
555
            uint32_t mv = AV_RN32A(&edge->mv);\
556
            if (mv) {\
557
                if (cur_sign_bias != sign_bias[edge_ref]) {\
558
                    /* SWAR negate of the values in mv. */\
559
                    mv = ((mv&0x80008000) + 0x00010001) ^ (mv&0x7fff7fff);\
560
                }\
561
                if (!n || mv != AV_RN32A(&near_mv[idx]))\
562
                    AV_WN32A(&near_mv[++idx], mv);\
563
                cnt[idx]      += 1 + (n != 2);\
564
            } else\
565
                cnt[CNT_ZERO] += 1 + (n != 2);\
566
        }\
567
    }
568
    MV_EDGE_CHECK(0)
569
    MV_EDGE_CHECK(1)
570
    MV_EDGE_CHECK(2)
571

    
572
    /* If we have three distinct MVs, merge first and last if they're the same */
573
    if (cnt[CNT_SPLITMV] && AV_RN32A(&near_mv[1+EDGE_TOP]) == AV_RN32A(&near_mv[1+EDGE_TOPLEFT]))
574
        cnt[CNT_NEAREST] += 1;
575

    
576
    cnt[CNT_SPLITMV] = ((mb_edge[EDGE_LEFT]->mode   == VP8_MVMODE_SPLIT) +
577
                        (mb_edge[EDGE_TOP]->mode    == VP8_MVMODE_SPLIT)) * 2 +
578
                       (mb_edge[EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT);
579

    
580
    /* Swap near and nearest if necessary */
581
    if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) {
582
        FFSWAP(uint8_t,     cnt[CNT_NEAREST],     cnt[CNT_NEAR]);
583
        FFSWAP( VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]);
584
    }
585

    
586
    /* Choose the best mv out of 0,0 and the nearest mv */
587
    if (cnt[CNT_NEAREST] >= cnt[CNT_ZERO])
588
        best_idx = CNT_NEAREST;
589

    
590
    clamp_mv(s, best, &near_mv[best_idx], mb_x, mb_y);
591
    near[0] = near_mv[CNT_NEAREST];
592
    near[1] = near_mv[CNT_NEAR];
593
}
594

    
595
/**
596
 * Motion vector coding, 17.1.
597
 */
598
static int read_mv_component(VP56RangeCoder *c, const uint8_t *p)
599
{
600
    int x = 0;
601

    
602
    if (vp56_rac_get_prob(c, p[0])) {
603
        int i;
604

    
605
        for (i = 0; i < 3; i++)
606
            x += vp56_rac_get_prob(c, p[9 + i]) << i;
607
        for (i = 9; i > 3; i--)
608
            x += vp56_rac_get_prob(c, p[9 + i]) << i;
609
        if (!(x & 0xFFF0) || vp56_rac_get_prob(c, p[12]))
610
            x += 8;
611
    } else
612
        x = vp8_rac_get_tree(c, vp8_small_mvtree, &p[2]);
613

    
614
    return (x && vp56_rac_get_prob(c, p[1])) ? -x : x;
615
}
616

    
617
static const uint8_t *get_submv_prob(uint32_t left, uint32_t top)
618
{
619
    if (left == top)
620
        return vp8_submv_prob[4-!!left];
621
    if (!top)
622
        return vp8_submv_prob[2];
623
    return vp8_submv_prob[1-!!left];
624
}
625

    
626
/**
627
 * Split motion vector prediction, 16.4.
628
 * @returns the number of motion vectors parsed (2, 4 or 16)
629
 */
630
static int decode_splitmvs(VP8Context    *s,  VP56RangeCoder *c,
631
                            VP8Macroblock *mb, VP56mv         *base_mv)
632
{
633
    int part_idx = mb->partitioning =
634
        vp8_rac_get_tree(c, vp8_mbsplit_tree, vp8_mbsplit_prob);
635
    int n, num = vp8_mbsplit_count[part_idx];
636
    VP8Macroblock *top_mb  = &mb[2];
637
    VP8Macroblock *left_mb = &mb[-1];
638
    const uint8_t *mbsplits_left = vp8_mbsplits[left_mb->partitioning],
639
                  *mbsplits_top = vp8_mbsplits[top_mb->partitioning],
640
                  *mbsplits_cur = vp8_mbsplits[part_idx],
641
                  *firstidx = vp8_mbfirstidx[part_idx];
642
    VP56mv *top_mv  = top_mb->bmv;
643
    VP56mv *left_mv = left_mb->bmv;
644
    VP56mv *cur_mv  = mb->bmv;
645

    
646
    for (n = 0; n < num; n++) {
647
        int k = firstidx[n];
648
        uint32_t left, above;
649
        const uint8_t *submv_prob;
650

    
651
        if (!(k & 3))
652
            left = AV_RN32A(&left_mv[mbsplits_left[k + 3]]);
653
        else
654
            left  = AV_RN32A(&cur_mv[mbsplits_cur[k - 1]]);
655
        if (k <= 3)
656
            above = AV_RN32A(&top_mv[mbsplits_top[k + 12]]);
657
        else
658
            above = AV_RN32A(&cur_mv[mbsplits_cur[k - 4]]);
659

    
660
        submv_prob = get_submv_prob(left, above);
661

    
662
        switch (vp8_rac_get_tree(c, vp8_submv_ref_tree, submv_prob)) {
663
        case VP8_SUBMVMODE_NEW4X4:
664
            mb->bmv[n].y = base_mv->y + read_mv_component(c, s->prob->mvc[0]);
665
            mb->bmv[n].x = base_mv->x + read_mv_component(c, s->prob->mvc[1]);
666
            break;
667
        case VP8_SUBMVMODE_ZERO4X4:
668
            AV_WN32A(&mb->bmv[n], 0);
669
            break;
670
        case VP8_SUBMVMODE_LEFT4X4:
671
            AV_WN32A(&mb->bmv[n], left);
672
            break;
673
        case VP8_SUBMVMODE_TOP4X4:
674
            AV_WN32A(&mb->bmv[n], above);
675
            break;
676
        }
677
    }
678

    
679
    return num;
680
}
681

    
682
static inline void decode_intra4x4_modes(VP56RangeCoder *c, uint8_t *intra4x4,
683
                                         int stride, int keyframe)
684
{
685
    int x, y, t, l, i;
686

    
687
    if (keyframe) {
688
        const uint8_t *ctx;
689
        for (y = 0; y < 4; y++) {
690
            for (x = 0; x < 4; x++) {
691
                t = intra4x4[x - stride];
692
                l = intra4x4[x - 1];
693
                ctx = vp8_pred4x4_prob_intra[t][l];
694
                intra4x4[x] = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx);
695
            }
696
            intra4x4 += stride;
697
        }
698
    } else {
699
        for (i = 0; i < 16; i++)
700
            intra4x4[i] = vp8_rac_get_tree(c, vp8_pred4x4_tree, vp8_pred4x4_prob_inter);
701
    }
702
}
703

    
704
static void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
705
                           uint8_t *intra4x4, uint8_t *segment)
706
{
707
    VP56RangeCoder *c = &s->c;
708

    
709
    if (s->segmentation.update_map)
710
        *segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid);
711
    mb->segment = *segment;
712

    
713
    mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0;
714

    
715
    if (s->keyframe) {
716
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra);
717

    
718
        if (mb->mode == MODE_I4x4) {
719
            decode_intra4x4_modes(c, intra4x4, s->b4_stride, 1);
720
        } else
721
            fill_rectangle(intra4x4, 4, 4, s->b4_stride, vp8_pred4x4_mode[mb->mode], 1);
722

    
723
        s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra);
724
        mb->ref_frame = VP56_FRAME_CURRENT;
725
    } else if (vp56_rac_get_prob(c, s->prob->intra)) {
726
        VP56mv near[2], best;
727
        uint8_t cnt[4] = { 0 };
728
        uint8_t p[4];
729

    
730
        // inter MB, 16.2
731
        if (vp56_rac_get_prob(c, s->prob->last))
732
            mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ?
733
                VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN;
734
        else
735
            mb->ref_frame = VP56_FRAME_PREVIOUS;
736

    
737
        // motion vectors, 16.3
738
        find_near_mvs(s, mb, mb_x, mb_y, near, &best, cnt);
739
        p[0] = vp8_mode_contexts[cnt[0]][0];
740
        p[1] = vp8_mode_contexts[cnt[1]][1];
741
        p[2] = vp8_mode_contexts[cnt[2]][2];
742
        p[3] = vp8_mode_contexts[cnt[3]][3];
743
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_mvinter, p);
744
        switch (mb->mode) {
745
        case VP8_MVMODE_SPLIT:
746
            mb->mv = mb->bmv[decode_splitmvs(s, c, mb, &best) - 1];
747
            break;
748
        case VP8_MVMODE_ZERO:
749
            mb->mv.x = 0;
750
            mb->mv.y = 0;
751
            break;
752
        case VP8_MVMODE_NEAREST:
753
            clamp_mv(s, &mb->mv, &near[0], mb_x, mb_y);
754
            break;
755
        case VP8_MVMODE_NEAR:
756
            clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y);
757
            break;
758
        case VP8_MVMODE_NEW:
759
            mb->mv.y = best.y + read_mv_component(c, s->prob->mvc[0]);
760
            mb->mv.x = best.x + read_mv_component(c, s->prob->mvc[1]);
761
            break;
762
        }
763
        if (mb->mode != VP8_MVMODE_SPLIT) {
764
            mb->partitioning = VP8_SPLITMVMODE_NONE;
765
            mb->bmv[0] = mb->mv;
766
        }
767
    } else {
768
        // intra MB, 16.1
769
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16);
770

    
771
        if (mb->mode == MODE_I4x4)
772
            decode_intra4x4_modes(c, intra4x4, 4, 0);
773

    
774
        s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c);
775
        mb->ref_frame = VP56_FRAME_CURRENT;
776
    }
777
}
778

    
779
/**
780
 * @param c arithmetic bitstream reader context
781
 * @param block destination for block coefficients
782
 * @param probs probabilities to use when reading trees from the bitstream
783
 * @param i initial coeff index, 0 unless a separate DC block is coded
784
 * @param zero_nhood the initial prediction context for number of surrounding
785
 *                   all-zero blocks (only left/top, so 0-2)
786
 * @param qmul array holding the dc/ac dequant factor at position 0/1
787
 * @return 0 if no coeffs were decoded
788
 *         otherwise, the index of the last coeff decoded plus one
789
 */
790
static int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16],
791
                               uint8_t probs[8][3][NUM_DCT_TOKENS-1],
792
                               int i, int zero_nhood, int16_t qmul[2])
793
{
794
    int token, nonzero = 0;
795
    int offset = 0;
796

    
797
    for (; i < 16; i++) {
798
        token = vp8_rac_get_tree_with_offset(c, vp8_coeff_tree, probs[vp8_coeff_band[i]][zero_nhood], offset);
799

    
800
        if (token == DCT_EOB)
801
            break;
802
        else if (token >= DCT_CAT1) {
803
            int cat = token-DCT_CAT1;
804
            token = vp8_rac_get_coeff(c, vp8_dct_cat_prob[cat]);
805
            token += vp8_dct_cat_offset[cat];
806
        }
807

    
808
        // after the first token, the non-zero prediction context becomes
809
        // based on the last decoded coeff
810
        if (!token) {
811
            zero_nhood = 0;
812
            offset = 1;
813
            continue;
814
        } else if (token == 1)
815
            zero_nhood = 1;
816
        else
817
            zero_nhood = 2;
818

    
819
        // todo: full [16] qmat? load into register?
820
        block[zigzag_scan[i]] = (vp8_rac_get(c) ? -token : token) * qmul[!!i];
821
        nonzero = i+1;
822
        offset = 0;
823
    }
824
    return nonzero;
825
}
826

    
827
static void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,
828
                             uint8_t t_nnz[9], uint8_t l_nnz[9])
829
{
830
    LOCAL_ALIGNED_16(DCTELEM, dc,[16]);
831
    int i, x, y, luma_start = 0, luma_ctx = 3;
832
    int nnz_pred, nnz, nnz_total = 0;
833
    int segment = s->segmentation.enabled ? mb->segment : 0;
834

    
835
    s->dsp.clear_blocks((DCTELEM *)s->block);
836

    
837
    if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
838
        AV_ZERO128(dc);
839
        AV_ZERO128(dc+8);
840
        nnz_pred = t_nnz[8] + l_nnz[8];
841

    
842
        // decode DC values and do hadamard
843
        nnz = decode_block_coeffs(c, dc, s->prob->token[1], 0, nnz_pred,
844
                                  s->qmat[segment].luma_dc_qmul);
845
        l_nnz[8] = t_nnz[8] = !!nnz;
846
        nnz_total += nnz;
847
        s->vp8dsp.vp8_luma_dc_wht(s->block, dc);
848
        luma_start = 1;
849
        luma_ctx = 0;
850
    }
851

    
852
    // luma blocks
853
    for (y = 0; y < 4; y++)
854
        for (x = 0; x < 4; x++) {
855
            nnz_pred = l_nnz[y] + t_nnz[x];
856
            nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start,
857
                                      nnz_pred, s->qmat[segment].luma_qmul);
858
            // nnz+luma_start may be one more than the actual last index, but we don't care
859
            s->non_zero_count_cache[y][x] = nnz + luma_start;
860
            t_nnz[x] = l_nnz[y] = !!nnz;
861
            nnz_total += nnz;
862
        }
863

    
864
    // chroma blocks
865
    // TODO: what to do about dimensions? 2nd dim for luma is x,
866
    // but for chroma it's (y<<1)|x
867
    for (i = 4; i < 6; i++)
868
        for (y = 0; y < 2; y++)
869
            for (x = 0; x < 2; x++) {
870
                nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x];
871
                nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0,
872
                                          nnz_pred, s->qmat[segment].chroma_qmul);
873
                s->non_zero_count_cache[i][(y<<1)+x] = nnz;
874
                t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz;
875
                nnz_total += nnz;
876
            }
877

    
878
    // if there were no coded coeffs despite the macroblock not being marked skip,
879
    // we MUST not do the inner loop filter and should not do IDCT
880
    // Since skip isn't used for bitstream prediction, just manually set it.
881
    if (!nnz_total)
882
        mb->skip = 1;
883
}
884

    
885
static av_always_inline
886
void backup_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
887
                      int linesize, int uvlinesize, int simple)
888
{
889
    AV_COPY128(top_border, src_y + 15*linesize);
890
    if (!simple) {
891
        AV_COPY64(top_border+16, src_cb + 7*uvlinesize);
892
        AV_COPY64(top_border+24, src_cr + 7*uvlinesize);
893
    }
894
}
895

    
896
static av_always_inline
897
void xchg_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
898
                    int linesize, int uvlinesize, int mb_x, int mb_y, int mb_width,
899
                    int simple, int xchg)
900
{
901
    uint8_t *top_border_m1 = top_border-32;     // for TL prediction
902
    src_y  -=   linesize;
903
    src_cb -= uvlinesize;
904
    src_cr -= uvlinesize;
905

    
906
#define XCHG(a,b,xchg) do {                     \
907
        if (xchg) AV_SWAP64(b,a);               \
908
        else      AV_COPY64(b,a);               \
909
    } while (0)
910

    
911
    XCHG(top_border_m1+8, src_y-8, xchg);
912
    XCHG(top_border,      src_y,   xchg);
913
    XCHG(top_border+8,    src_y+8, 1);
914
    if (mb_x < mb_width-1)
915
        XCHG(top_border+32, src_y+16, 1);
916

    
917
    // only copy chroma for normal loop filter
918
    // or to initialize the top row to 127
919
    if (!simple || !mb_y) {
920
        XCHG(top_border_m1+16, src_cb-8, xchg);
921
        XCHG(top_border_m1+24, src_cr-8, xchg);
922
        XCHG(top_border+16,    src_cb, 1);
923
        XCHG(top_border+24,    src_cr, 1);
924
    }
925
}
926

    
927
static int check_intra_pred_mode(int mode, int mb_x, int mb_y)
928
{
929
    if (mode == DC_PRED8x8) {
930
        if (!(mb_x|mb_y))
931
            mode = DC_128_PRED8x8;
932
        else if (!mb_y)
933
            mode = LEFT_DC_PRED8x8;
934
        else if (!mb_x)
935
            mode = TOP_DC_PRED8x8;
936
    }
937
    return mode;
938
}
939

    
940
static void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
941
                          uint8_t *intra4x4, int mb_x, int mb_y)
942
{
943
    int x, y, mode, nnz, tr;
944

    
945
    // for the first row, we need to run xchg_mb_border to init the top edge to 127
946
    // otherwise, skip it if we aren't going to deblock
947
    if (s->deblock_filter || !mb_y)
948
        xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
949
                       s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
950
                       s->filter.simple, 1);
951

    
952
    if (mb->mode < MODE_I4x4) {
953
        mode = check_intra_pred_mode(mb->mode, mb_x, mb_y);
954
        s->hpc.pred16x16[mode](dst[0], s->linesize);
955
    } else {
956
        uint8_t *ptr = dst[0];
957
        int stride = s->keyframe ? s->b4_stride : 4;
958

    
959
        // all blocks on the right edge of the macroblock use bottom edge
960
        // the top macroblock for their topright edge
961
        uint8_t *tr_right = ptr - s->linesize + 16;
962

    
963
        // if we're on the right edge of the frame, said edge is extended
964
        // from the top macroblock
965
        if (mb_x == s->mb_width-1) {
966
            tr = tr_right[-1]*0x01010101;
967
            tr_right = (uint8_t *)&tr;
968
        }
969

    
970
        for (y = 0; y < 4; y++) {
971
            uint8_t *topright = ptr + 4 - s->linesize;
972
            for (x = 0; x < 4; x++) {
973
                if (x == 3)
974
                    topright = tr_right;
975

    
976
                s->hpc.pred4x4[intra4x4[x]](ptr+4*x, topright, s->linesize);
977

    
978
                nnz = s->non_zero_count_cache[y][x];
979
                if (nnz) {
980
                    if (nnz == 1)
981
                        s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize);
982
                    else
983
                        s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize);
984
                }
985
                topright += 4;
986
            }
987

    
988
            ptr   += 4*s->linesize;
989
            intra4x4 += stride;
990
        }
991
    }
992

    
993
    mode = check_intra_pred_mode(s->chroma_pred_mode, mb_x, mb_y);
994
    s->hpc.pred8x8[mode](dst[1], s->uvlinesize);
995
    s->hpc.pred8x8[mode](dst[2], s->uvlinesize);
996

    
997
    if (s->deblock_filter || !mb_y)
998
        xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
999
                       s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
1000
                       s->filter.simple, 0);
1001
}
1002

    
1003
/**
1004
 * Generic MC function.
1005
 *
1006
 * @param s VP8 decoding context
1007
 * @param luma 1 for luma (Y) planes, 0 for chroma (Cb/Cr) planes
1008
 * @param dst target buffer for block data at block position
1009
 * @param src reference picture buffer at origin (0, 0)
1010
 * @param mv motion vector (relative to block position) to get pixel data from
1011
 * @param x_off horizontal position of block from origin (0, 0)
1012
 * @param y_off vertical position of block from origin (0, 0)
1013
 * @param block_w width of block (16, 8 or 4)
1014
 * @param block_h height of block (always same as block_w)
1015
 * @param width width of src/dst plane data
1016
 * @param height height of src/dst plane data
1017
 * @param linesize size of a single line of plane data, including padding
1018
 * @param mc_func motion compensation function pointers (bilinear or sixtap MC)
1019
 */
1020
static inline void vp8_mc(VP8Context *s, int luma,
1021
                          uint8_t *dst, uint8_t *src, const VP56mv *mv,
1022
                          int x_off, int y_off, int block_w, int block_h,
1023
                          int width, int height, int linesize,
1024
                          vp8_mc_func mc_func[3][3])
1025
{
1026
    if (AV_RN32A(mv)) {
1027
        static const uint8_t idx[8] = { 0, 1, 2, 1, 2, 1, 2, 1 };
1028
        int mx = (mv->x << luma)&7, mx_idx = idx[mx];
1029
        int my = (mv->y << luma)&7, my_idx = idx[my];
1030

    
1031
        x_off += mv->x >> (3 - luma);
1032
        y_off += mv->y >> (3 - luma);
1033

    
1034
        // edge emulation
1035
        src += y_off * linesize + x_off;
1036
        if (x_off < 2 || x_off >= width  - block_w - 3 ||
1037
            y_off < 2 || y_off >= height - block_h - 3) {
1038
            ff_emulated_edge_mc(s->edge_emu_buffer, src - 2 * linesize - 2, linesize,
1039
                                block_w + 5, block_h + 5,
1040
                                x_off - 2, y_off - 2, width, height);
1041
            src = s->edge_emu_buffer + 2 + linesize * 2;
1042
        }
1043
        mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);
1044
    } else
1045
        mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0);
1046
}
1047

    
1048
static inline void vp8_mc_part(VP8Context *s, uint8_t *dst[3],
1049
                               AVFrame *ref_frame, int x_off, int y_off,
1050
                               int bx_off, int by_off,
1051
                               int block_w, int block_h,
1052
                               int width, int height, VP56mv *mv)
1053
{
1054
    VP56mv uvmv = *mv;
1055

    
1056
    /* Y */
1057
    vp8_mc(s, 1, dst[0] + by_off * s->linesize + bx_off,
1058
           ref_frame->data[0], mv, x_off + bx_off, y_off + by_off,
1059
           block_w, block_h, width, height, s->linesize,
1060
           s->put_pixels_tab[block_w == 8]);
1061

    
1062
    /* U/V */
1063
    if (s->profile == 3) {
1064
        uvmv.x &= ~7;
1065
        uvmv.y &= ~7;
1066
    }
1067
    x_off   >>= 1; y_off   >>= 1;
1068
    bx_off  >>= 1; by_off  >>= 1;
1069
    width   >>= 1; height  >>= 1;
1070
    block_w >>= 1; block_h >>= 1;
1071
    vp8_mc(s, 0, dst[1] + by_off * s->uvlinesize + bx_off,
1072
           ref_frame->data[1], &uvmv, x_off + bx_off, y_off + by_off,
1073
           block_w, block_h, width, height, s->uvlinesize,
1074
           s->put_pixels_tab[1 + (block_w == 4)]);
1075
    vp8_mc(s, 0, dst[2] + by_off * s->uvlinesize + bx_off,
1076
           ref_frame->data[2], &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
}
1080

    
1081
/* Fetch pixels for estimated mv 4 macroblocks ahead.
1082
 * Optimized for 64-byte cache lines.  Inspired by ffh264 prefetch_motion. */
1083
static inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int x_off, int y_off, int ref)
1084
{
1085
    if (mb->ref_frame != VP56_FRAME_CURRENT) {
1086
        int mx = mb->mv.x + x_off + 8;
1087
        int my = mb->mv.y + y_off;
1088
        uint8_t **src= s->framep[mb->ref_frame]->data;
1089
        int off= mx + (my + (mb_x&3)*4)*s->linesize + 64;
1090
        s->dsp.prefetch(src[0]+off, s->linesize, 4);
1091
        off= (mx>>1) + ((my>>1) + (mb_x&7))*s->uvlinesize + 64;
1092
        s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1093
    }
1094
}
1095

    
1096
/**
1097
 * Apply motion vectors to prediction buffer, chapter 18.
1098
 */
1099
static void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
1100
                          int mb_x, int mb_y)
1101
{
1102
    int x_off = mb_x << 4, y_off = mb_y << 4;
1103
    int width = 16*s->mb_width, height = 16*s->mb_height;
1104

    
1105
    prefetch_motion(s, mb, mb_x, mb_y, x_off, y_off, VP56_FRAME_PREVIOUS);
1106

    
1107
    if (mb->mode < VP8_MVMODE_SPLIT) {
1108
        vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1109
                    0, 0, 16, 16, width, height, &mb->mv);
1110
    } else switch (mb->partitioning) {
1111
    case VP8_SPLITMVMODE_4x4: {
1112
        int x, y;
1113
        VP56mv uvmv;
1114

    
1115
        /* Y */
1116
        for (y = 0; y < 4; y++) {
1117
            for (x = 0; x < 4; x++) {
1118
                vp8_mc(s, 1, dst[0] + 4*y*s->linesize + x*4,
1119
                       s->framep[mb->ref_frame]->data[0], &mb->bmv[4*y + x],
1120
                       4*x + x_off, 4*y + y_off, 4, 4,
1121
                       width, height, s->linesize,
1122
                       s->put_pixels_tab[2]);
1123
            }
1124
        }
1125

    
1126
        /* U/V */
1127
        x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1;
1128
        for (y = 0; y < 2; y++) {
1129
            for (x = 0; x < 2; x++) {
1130
                uvmv.x = mb->bmv[ 2*y    * 4 + 2*x  ].x +
1131
                         mb->bmv[ 2*y    * 4 + 2*x+1].x +
1132
                         mb->bmv[(2*y+1) * 4 + 2*x  ].x +
1133
                         mb->bmv[(2*y+1) * 4 + 2*x+1].x;
1134
                uvmv.y = mb->bmv[ 2*y    * 4 + 2*x  ].y +
1135
                         mb->bmv[ 2*y    * 4 + 2*x+1].y +
1136
                         mb->bmv[(2*y+1) * 4 + 2*x  ].y +
1137
                         mb->bmv[(2*y+1) * 4 + 2*x+1].y;
1138
                uvmv.x = (uvmv.x + 2 + (uvmv.x >> (INT_BIT-1))) >> 2;
1139
                uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2;
1140
                if (s->profile == 3) {
1141
                    uvmv.x &= ~7;
1142
                    uvmv.y &= ~7;
1143
                }
1144
                vp8_mc(s, 0, dst[1] + 4*y*s->uvlinesize + x*4,
1145
                       s->framep[mb->ref_frame]->data[1], &uvmv,
1146
                       4*x + x_off, 4*y + y_off, 4, 4,
1147
                       width, height, s->uvlinesize,
1148
                       s->put_pixels_tab[2]);
1149
                vp8_mc(s, 0, dst[2] + 4*y*s->uvlinesize + x*4,
1150
                       s->framep[mb->ref_frame]->data[2], &uvmv,
1151
                       4*x + x_off, 4*y + y_off, 4, 4,
1152
                       width, height, s->uvlinesize,
1153
                       s->put_pixels_tab[2]);
1154
            }
1155
        }
1156
        break;
1157
    }
1158
    case VP8_SPLITMVMODE_16x8:
1159
        vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1160
                    0, 0, 16, 8, width, height, &mb->bmv[0]);
1161
        vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1162
                    0, 8, 16, 8, width, height, &mb->bmv[1]);
1163
        break;
1164
    case VP8_SPLITMVMODE_8x16:
1165
        vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1166
                    0, 0, 8, 16, width, height, &mb->bmv[0]);
1167
        vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1168
                    8, 0, 8, 16, width, height, &mb->bmv[1]);
1169
        break;
1170
    case VP8_SPLITMVMODE_8x8:
1171
        vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1172
                    0, 0, 8, 8, width, height, &mb->bmv[0]);
1173
        vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1174
                    8, 0, 8, 8, width, height, &mb->bmv[1]);
1175
        vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1176
                    0, 8, 8, 8, width, height, &mb->bmv[2]);
1177
        vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1178
                    8, 8, 8, 8, width, height, &mb->bmv[3]);
1179
        break;
1180
    }
1181

    
1182
    prefetch_motion(s, mb, mb_x, mb_y, x_off, y_off, VP56_FRAME_GOLDEN);
1183
}
1184

    
1185
static void idct_mb(VP8Context *s, uint8_t *y_dst, uint8_t *u_dst, uint8_t *v_dst,
1186
                    VP8Macroblock *mb)
1187
{
1188
    int x, y, nnz;
1189

    
1190
    if (mb->mode != MODE_I4x4)
1191
        for (y = 0; y < 4; y++) {
1192
            for (x = 0; x < 4; x++) {
1193
                nnz = s->non_zero_count_cache[y][x];
1194
                if (nnz) {
1195
                    if (nnz == 1)
1196
                        s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize);
1197
                    else
1198
                        s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize);
1199
                }
1200
            }
1201
            y_dst += 4*s->linesize;
1202
        }
1203

    
1204
    for (y = 0; y < 2; y++) {
1205
        for (x = 0; x < 2; x++) {
1206
            nnz = s->non_zero_count_cache[4][(y<<1)+x];
1207
            if (nnz) {
1208
                if (nnz == 1)
1209
                    s->vp8dsp.vp8_idct_dc_add(u_dst+4*x, s->block[4][(y<<1)+x], s->uvlinesize);
1210
                else
1211
                    s->vp8dsp.vp8_idct_add(u_dst+4*x, s->block[4][(y<<1)+x], s->uvlinesize);
1212
            }
1213

    
1214
            nnz = s->non_zero_count_cache[5][(y<<1)+x];
1215
            if (nnz) {
1216
                if (nnz == 1)
1217
                    s->vp8dsp.vp8_idct_dc_add(v_dst+4*x, s->block[5][(y<<1)+x], s->uvlinesize);
1218
                else
1219
                    s->vp8dsp.vp8_idct_add(v_dst+4*x, s->block[5][(y<<1)+x], s->uvlinesize);
1220
            }
1221
        }
1222
        u_dst += 4*s->uvlinesize;
1223
        v_dst += 4*s->uvlinesize;
1224
    }
1225
}
1226

    
1227
static void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f )
1228
{
1229
    int interior_limit, filter_level;
1230

    
1231
    if (s->segmentation.enabled) {
1232
        filter_level = s->segmentation.filter_level[mb->segment];
1233
        if (!s->segmentation.absolute_vals)
1234
            filter_level += s->filter.level;
1235
    } else
1236
        filter_level = s->filter.level;
1237

    
1238
    if (s->lf_delta.enabled) {
1239
        filter_level += s->lf_delta.ref[mb->ref_frame];
1240

    
1241
        if (mb->ref_frame == VP56_FRAME_CURRENT) {
1242
            if (mb->mode == MODE_I4x4)
1243
                filter_level += s->lf_delta.mode[0];
1244
        } else {
1245
            if (mb->mode == VP8_MVMODE_ZERO)
1246
                filter_level += s->lf_delta.mode[1];
1247
            else if (mb->mode == VP8_MVMODE_SPLIT)
1248
                filter_level += s->lf_delta.mode[3];
1249
            else
1250
                filter_level += s->lf_delta.mode[2];
1251
        }
1252
    }
1253
    filter_level = av_clip(filter_level, 0, 63);
1254

    
1255
    interior_limit = filter_level;
1256
    if (s->filter.sharpness) {
1257
        interior_limit >>= s->filter.sharpness > 4 ? 2 : 1;
1258
        interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness);
1259
    }
1260
    interior_limit = FFMAX(interior_limit, 1);
1261

    
1262
    f->filter_level = filter_level;
1263
    f->inner_limit = interior_limit;
1264
    f->inner_filter = !mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT;
1265
}
1266

    
1267
static void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f, int mb_x, int mb_y)
1268
{
1269
    int mbedge_lim, bedge_lim, hev_thresh;
1270
    int filter_level = f->filter_level;
1271
    int inner_limit = f->inner_limit;
1272
    int inner_filter = f->inner_filter;
1273

    
1274
    if (!filter_level)
1275
        return;
1276

    
1277
    mbedge_lim = 2*(filter_level+2) + inner_limit;
1278
     bedge_lim = 2* filter_level    + inner_limit;
1279
    hev_thresh = filter_level >= 15;
1280

    
1281
    if (s->keyframe) {
1282
        if (filter_level >= 40)
1283
            hev_thresh = 2;
1284
    } else {
1285
        if (filter_level >= 40)
1286
            hev_thresh = 3;
1287
        else if (filter_level >= 20)
1288
            hev_thresh = 2;
1289
    }
1290

    
1291
    if (mb_x) {
1292
        s->vp8dsp.vp8_h_loop_filter16y(dst[0],     s->linesize,
1293
                                       mbedge_lim, inner_limit, hev_thresh);
1294
        s->vp8dsp.vp8_h_loop_filter8uv(dst[1],     dst[2],      s->uvlinesize,
1295
                                       mbedge_lim, inner_limit, hev_thresh);
1296
    }
1297

    
1298
    if (inner_filter) {
1299
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 4, s->linesize, bedge_lim,
1300
                                             inner_limit,   hev_thresh);
1301
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 8, s->linesize, bedge_lim,
1302
                                             inner_limit,   hev_thresh);
1303
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+12, s->linesize, bedge_lim,
1304
                                             inner_limit,   hev_thresh);
1305
        s->vp8dsp.vp8_h_loop_filter8uv_inner(dst[1] + 4,    dst[2] + 4,
1306
                                             s->uvlinesize, bedge_lim,
1307
                                             inner_limit,   hev_thresh);
1308
    }
1309

    
1310
    if (mb_y) {
1311
        s->vp8dsp.vp8_v_loop_filter16y(dst[0],     s->linesize,
1312
                                       mbedge_lim, inner_limit, hev_thresh);
1313
        s->vp8dsp.vp8_v_loop_filter8uv(dst[1],     dst[2],      s->uvlinesize,
1314
                                       mbedge_lim, inner_limit, hev_thresh);
1315
    }
1316

    
1317
    if (inner_filter) {
1318
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 4*s->linesize,
1319
                                             s->linesize,   bedge_lim,
1320
                                             inner_limit,   hev_thresh);
1321
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 8*s->linesize,
1322
                                             s->linesize,   bedge_lim,
1323
                                             inner_limit,   hev_thresh);
1324
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+12*s->linesize,
1325
                                             s->linesize,   bedge_lim,
1326
                                             inner_limit, hev_thresh);
1327
        s->vp8dsp.vp8_v_loop_filter8uv_inner(dst[1] + 4 * s->uvlinesize,
1328
                                             dst[2] + 4 * s->uvlinesize,
1329
                                             s->uvlinesize, bedge_lim,
1330
                                             inner_limit,   hev_thresh);
1331
    }
1332
}
1333

    
1334
static void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f, int mb_x, int mb_y)
1335
{
1336
    int mbedge_lim, bedge_lim;
1337
    int filter_level = f->filter_level;
1338
    int inner_limit = f->inner_limit;
1339
    int inner_filter = f->inner_filter;
1340

    
1341
    if (!filter_level)
1342
        return;
1343

    
1344
    mbedge_lim = 2*(filter_level+2) + inner_limit;
1345
     bedge_lim = 2* filter_level    + inner_limit;
1346

    
1347
    if (mb_x)
1348
        s->vp8dsp.vp8_h_loop_filter_simple(dst, s->linesize, mbedge_lim);
1349
    if (inner_filter) {
1350
        s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, s->linesize, bedge_lim);
1351
        s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, s->linesize, bedge_lim);
1352
        s->vp8dsp.vp8_h_loop_filter_simple(dst+12, s->linesize, bedge_lim);
1353
    }
1354

    
1355
    if (mb_y)
1356
        s->vp8dsp.vp8_v_loop_filter_simple(dst, s->linesize, mbedge_lim);
1357
    if (inner_filter) {
1358
        s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*s->linesize, s->linesize, bedge_lim);
1359
        s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*s->linesize, s->linesize, bedge_lim);
1360
        s->vp8dsp.vp8_v_loop_filter_simple(dst+12*s->linesize, s->linesize, bedge_lim);
1361
    }
1362
}
1363

    
1364
static void filter_mb_row(VP8Context *s, int mb_y)
1365
{
1366
    VP8FilterStrength *f = s->filter_strength;
1367
    uint8_t *dst[3] = {
1368
        s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize,
1369
        s->framep[VP56_FRAME_CURRENT]->data[1] +  8*mb_y*s->uvlinesize,
1370
        s->framep[VP56_FRAME_CURRENT]->data[2] +  8*mb_y*s->uvlinesize
1371
    };
1372
    int mb_x;
1373

    
1374
    for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1375
        backup_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], s->linesize, s->uvlinesize, 0);
1376
        filter_mb(s, dst, f++, mb_x, mb_y);
1377
        dst[0] += 16;
1378
        dst[1] += 8;
1379
        dst[2] += 8;
1380
    }
1381
}
1382

    
1383
static void filter_mb_row_simple(VP8Context *s, int mb_y)
1384
{
1385
    VP8FilterStrength *f = s->filter_strength;
1386
    uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize;
1387
    int mb_x;
1388

    
1389
    for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1390
        backup_mb_border(s->top_border[mb_x+1], dst, NULL, NULL, s->linesize, 0, 1);
1391
        filter_mb_simple(s, dst, f++, mb_x, mb_y);
1392
        dst += 16;
1393
    }
1394
}
1395

    
1396
static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
1397
                            AVPacket *avpkt)
1398
{
1399
    VP8Context *s = avctx->priv_data;
1400
    int ret, mb_x, mb_y, i, y, referenced;
1401
    enum AVDiscard skip_thresh;
1402
    AVFrame *curframe;
1403

    
1404
    if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0)
1405
        return ret;
1406

    
1407
    referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT
1408
                                || s->update_altref == VP56_FRAME_CURRENT;
1409

    
1410
    skip_thresh = !referenced ? AVDISCARD_NONREF :
1411
                    !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL;
1412

    
1413
    if (avctx->skip_frame >= skip_thresh) {
1414
        s->invisible = 1;
1415
        goto skip_decode;
1416
    }
1417
    s->deblock_filter = s->filter.level && avctx->skip_loop_filter < skip_thresh;
1418

    
1419
    for (i = 0; i < 4; i++)
1420
        if (&s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1421
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1422
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) {
1423
            curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i];
1424
            break;
1425
        }
1426
    if (curframe->data[0])
1427
        avctx->release_buffer(avctx, curframe);
1428

    
1429
    curframe->key_frame = s->keyframe;
1430
    curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE;
1431
    curframe->reference = referenced ? 3 : 0;
1432
    if ((ret = avctx->get_buffer(avctx, curframe))) {
1433
        av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n");
1434
        return ret;
1435
    }
1436

    
1437
    // Given that arithmetic probabilities are updated every frame, it's quite likely
1438
    // that the values we have on a random interframe are complete junk if we didn't
1439
    // start decode on a keyframe. So just don't display anything rather than junk.
1440
    if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] ||
1441
                         !s->framep[VP56_FRAME_GOLDEN] ||
1442
                         !s->framep[VP56_FRAME_GOLDEN2])) {
1443
        av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n");
1444
        return AVERROR_INVALIDDATA;
1445
    }
1446

    
1447
    s->linesize   = curframe->linesize[0];
1448
    s->uvlinesize = curframe->linesize[1];
1449

    
1450
    if (!s->edge_emu_buffer)
1451
        s->edge_emu_buffer = av_malloc(21*s->linesize);
1452

    
1453
    memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz));
1454

    
1455
    /* Zero macroblock structures for top/left prediction from outside the frame. */
1456
    memset(s->macroblocks, 0, (s->mb_width + s->mb_height*2)*sizeof(*s->macroblocks));
1457

    
1458
    // top edge of 127 for intra prediction
1459
    memset(s->top_border, 127, (s->mb_width+1)*sizeof(*s->top_border));
1460

    
1461
    for (mb_y = 0; mb_y < s->mb_height; mb_y++) {
1462
        VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)];
1463
        VP8Macroblock *mb = s->macroblocks + (s->mb_height - mb_y - 1)*2;
1464
        uint8_t *intra4x4 = s->intra4x4_pred_mode + 4*mb_y*s->b4_stride;
1465
        uint8_t *segment_map = s->segmentation_map + mb_y*s->mb_stride;
1466
        uint8_t *dst[3] = {
1467
            curframe->data[0] + 16*mb_y*s->linesize,
1468
            curframe->data[1] +  8*mb_y*s->uvlinesize,
1469
            curframe->data[2] +  8*mb_y*s->uvlinesize
1470
        };
1471

    
1472
        memset(s->left_nnz, 0, sizeof(s->left_nnz));
1473

    
1474
        // left edge of 129 for intra prediction
1475
        if (!(avctx->flags & CODEC_FLAG_EMU_EDGE))
1476
            for (i = 0; i < 3; i++)
1477
                for (y = 0; y < 16>>!!i; y++)
1478
                    dst[i][y*curframe->linesize[i]-1] = 129;
1479
        if (mb_y)
1480
            memset(s->top_border, 129, sizeof(*s->top_border));
1481

    
1482
        for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1483
            uint8_t *intra4x4_mb = s->keyframe ? intra4x4 + 4*mb_x : s->intra4x4_pred_mode_mb;
1484
            uint8_t *segment_mb = segment_map+mb_x;
1485

    
1486
            /* Prefetch the current frame, 4 MBs ahead */
1487
            s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4);
1488
            s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2);
1489

    
1490
            decode_mb_mode(s, mb, mb_x, mb_y, intra4x4_mb, segment_mb);
1491

    
1492
            if (!mb->skip)
1493
                decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);
1494
            else {
1495
                AV_ZERO128(s->non_zero_count_cache);    // luma
1496
                AV_ZERO64(s->non_zero_count_cache[4]);  // chroma
1497
            }
1498

    
1499
            if (mb->mode <= MODE_I4x4) {
1500
                intra_predict(s, dst, mb, intra4x4_mb, mb_x, mb_y);
1501
                memset(mb->bmv, 0, sizeof(mb->bmv));
1502
            } else {
1503
                inter_predict(s, dst, mb, mb_x, mb_y);
1504
            }
1505

    
1506
            if (!mb->skip) {
1507
                idct_mb(s, dst[0], dst[1], dst[2], mb);
1508
            } else {
1509
                AV_ZERO64(s->left_nnz);
1510
                AV_WN64(s->top_nnz[mb_x], 0);   // array of 9, so unaligned
1511

    
1512
                // Reset DC block predictors if they would exist if the mb had coefficients
1513
                if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
1514
                    s->left_nnz[8]      = 0;
1515
                    s->top_nnz[mb_x][8] = 0;
1516
                }
1517
            }
1518

    
1519
            if (s->deblock_filter)
1520
                filter_level_for_mb(s, mb, &s->filter_strength[mb_x]);
1521

    
1522
            dst[0] += 16;
1523
            dst[1] += 8;
1524
            dst[2] += 8;
1525
            mb++;
1526
        }
1527
        if (s->deblock_filter) {
1528
            if (s->filter.simple)
1529
                filter_mb_row_simple(s, mb_y);
1530
            else
1531
                filter_mb_row(s, mb_y);
1532
        }
1533
    }
1534

    
1535
skip_decode:
1536
    // if future frames don't use the updated probabilities,
1537
    // reset them to the values we saved
1538
    if (!s->update_probabilities)
1539
        s->prob[0] = s->prob[1];
1540

    
1541
    // check if golden and altref are swapped
1542
    if (s->update_altref == VP56_FRAME_GOLDEN &&
1543
        s->update_golden == VP56_FRAME_GOLDEN2)
1544
        FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN], s->framep[VP56_FRAME_GOLDEN2]);
1545
    else {
1546
        if (s->update_altref != VP56_FRAME_NONE)
1547
            s->framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref];
1548

    
1549
        if (s->update_golden != VP56_FRAME_NONE)
1550
            s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden];
1551
    }
1552

    
1553
    if (s->update_last) // move cur->prev
1554
        s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT];
1555

    
1556
    // release no longer referenced frames
1557
    for (i = 0; i < 4; i++)
1558
        if (s->frames[i].data[0] &&
1559
            &s->frames[i] != s->framep[VP56_FRAME_CURRENT] &&
1560
            &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1561
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1562
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2])
1563
            avctx->release_buffer(avctx, &s->frames[i]);
1564

    
1565
    if (!s->invisible) {
1566
        *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT];
1567
        *data_size = sizeof(AVFrame);
1568
    }
1569

    
1570
    return avpkt->size;
1571
}
1572

    
1573
static av_cold int vp8_decode_init(AVCodecContext *avctx)
1574
{
1575
    VP8Context *s = avctx->priv_data;
1576

    
1577
    s->avctx = avctx;
1578
    avctx->pix_fmt = PIX_FMT_YUV420P;
1579

    
1580
    dsputil_init(&s->dsp, avctx);
1581
    ff_h264_pred_init(&s->hpc, CODEC_ID_VP8);
1582
    ff_vp8dsp_init(&s->vp8dsp);
1583

    
1584
    // intra pred needs edge emulation among other things
1585
    if (avctx->flags&CODEC_FLAG_EMU_EDGE) {
1586
        av_log(avctx, AV_LOG_ERROR, "Edge emulation not supported\n");
1587
        return AVERROR_PATCHWELCOME;
1588
    }
1589

    
1590
    return 0;
1591
}
1592

    
1593
static av_cold int vp8_decode_free(AVCodecContext *avctx)
1594
{
1595
    vp8_decode_flush(avctx);
1596
    return 0;
1597
}
1598

    
1599
AVCodec vp8_decoder = {
1600
    "vp8",
1601
    AVMEDIA_TYPE_VIDEO,
1602
    CODEC_ID_VP8,
1603
    sizeof(VP8Context),
1604
    vp8_decode_init,
1605
    NULL,
1606
    vp8_decode_free,
1607
    vp8_decode_frame,
1608
    CODEC_CAP_DR1,
1609
    .flush = vp8_decode_flush,
1610
    .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"),
1611
};