Statistics
| Branch: | Revision:

ffmpeg / libavcodec / vp8.c @ 8cfae560

History | View | Annotate | Download (55.9 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

    
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
    s->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
            AV_WN32A(&mb->mv, 0);
750
            break;
751
        case VP8_MVMODE_NEAREST:
752
            clamp_mv(s, &mb->mv, &near[0], mb_x, mb_y);
753
            break;
754
        case VP8_MVMODE_NEAR:
755
            clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y);
756
            break;
757
        case VP8_MVMODE_NEW:
758
            mb->mv.y = best.y + read_mv_component(c, s->prob->mvc[0]);
759
            mb->mv.x = best.x + read_mv_component(c, s->prob->mvc[1]);
760
            break;
761
        }
762
        if (mb->mode != VP8_MVMODE_SPLIT) {
763
            mb->partitioning = VP8_SPLITMVMODE_NONE;
764
            mb->bmv[0] = mb->mv;
765
        }
766
    } else {
767
        // intra MB, 16.1
768
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16);
769

    
770
        if (mb->mode == MODE_I4x4)
771
            decode_intra4x4_modes(c, intra4x4, 4, 0);
772

    
773
        s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c);
774
        mb->ref_frame = VP56_FRAME_CURRENT;
775
        mb->partitioning = VP8_SPLITMVMODE_NONE;
776
        AV_WN32A(&mb->bmv[0], 0);
777
    }
778
}
779

    
780
/**
781
 * @param c arithmetic bitstream reader context
782
 * @param block destination for block coefficients
783
 * @param probs probabilities to use when reading trees from the bitstream
784
 * @param i initial coeff index, 0 unless a separate DC block is coded
785
 * @param zero_nhood the initial prediction context for number of surrounding
786
 *                   all-zero blocks (only left/top, so 0-2)
787
 * @param qmul array holding the dc/ac dequant factor at position 0/1
788
 * @return 0 if no coeffs were decoded
789
 *         otherwise, the index of the last coeff decoded plus one
790
 */
791
static int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16],
792
                               uint8_t probs[8][3][NUM_DCT_TOKENS-1],
793
                               int i, int zero_nhood, int16_t qmul[2])
794
{
795
    int token, nonzero = 0;
796
    int offset = 0;
797

    
798
    for (; i < 16; i++) {
799
        token = vp8_rac_get_tree_with_offset(c, vp8_coeff_tree, probs[vp8_coeff_band[i]][zero_nhood], offset);
800

    
801
        if (token == DCT_EOB)
802
            break;
803
        else if (token >= DCT_CAT1) {
804
            int cat = token-DCT_CAT1;
805
            token = vp8_rac_get_coeff(c, vp8_dct_cat_prob[cat]);
806
            token += 3 + (2<<cat);
807
        }
808

    
809
        // after the first token, the non-zero prediction context becomes
810
        // based on the last decoded coeff
811
        if (!token) {
812
            zero_nhood = 0;
813
            offset = 1;
814
            continue;
815
        } else if (token == 1)
816
            zero_nhood = 1;
817
        else
818
            zero_nhood = 2;
819

    
820
        // todo: full [16] qmat? load into register?
821
        block[zigzag_scan[i]] = (vp8_rac_get(c) ? -token : token) * qmul[!!i];
822
        nonzero = i+1;
823
        offset = 0;
824
    }
825
    return nonzero;
826
}
827

    
828
static void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,
829
                             uint8_t t_nnz[9], uint8_t l_nnz[9])
830
{
831
    LOCAL_ALIGNED_16(DCTELEM, dc,[16]);
832
    int i, x, y, luma_start = 0, luma_ctx = 3;
833
    int nnz_pred, nnz, nnz_total = 0;
834
    int segment = s->segment;
835

    
836
    s->dsp.clear_blocks((DCTELEM *)s->block);
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
        for (y = 0; y < 4; y++) {
972
            uint8_t *topright = ptr + 4 - s->linesize;
973
            for (x = 0; x < 4; x++) {
974
                if (x == 3)
975
                    topright = tr_right;
976

    
977
                s->hpc.pred4x4[intra4x4[x]](ptr+4*x, topright, s->linesize);
978

    
979
                nnz = s->non_zero_count_cache[y][x];
980
                if (nnz) {
981
                    if (nnz == 1)
982
                        s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize);
983
                    else
984
                        s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize);
985
                }
986
                topright += 4;
987
            }
988

    
989
            ptr   += 4*s->linesize;
990
            intra4x4 += stride;
991
        }
992
    }
993

    
994
    mode = check_intra_pred_mode(s->chroma_pred_mode, mb_x, mb_y);
995
    s->hpc.pred8x8[mode](dst[1], s->uvlinesize);
996
    s->hpc.pred8x8[mode](dst[2], s->uvlinesize);
997

    
998
    if (s->deblock_filter || !mb_y)
999
        xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
1000
                       s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
1001
                       s->filter.simple, 0);
1002
}
1003

    
1004
/**
1005
 * Generic MC function.
1006
 *
1007
 * @param s VP8 decoding context
1008
 * @param luma 1 for luma (Y) planes, 0 for chroma (Cb/Cr) planes
1009
 * @param dst target buffer for block data at block position
1010
 * @param src reference picture buffer at origin (0, 0)
1011
 * @param mv motion vector (relative to block position) to get pixel data from
1012
 * @param x_off horizontal position of block from origin (0, 0)
1013
 * @param y_off vertical position of block from origin (0, 0)
1014
 * @param block_w width of block (16, 8 or 4)
1015
 * @param block_h height of block (always same as block_w)
1016
 * @param width width of src/dst plane data
1017
 * @param height height of src/dst plane data
1018
 * @param linesize size of a single line of plane data, including padding
1019
 * @param mc_func motion compensation function pointers (bilinear or sixtap MC)
1020
 */
1021
static inline void vp8_mc(VP8Context *s, int luma,
1022
                          uint8_t *dst, uint8_t *src, const VP56mv *mv,
1023
                          int x_off, int y_off, int block_w, int block_h,
1024
                          int width, int height, int linesize,
1025
                          vp8_mc_func mc_func[3][3])
1026
{
1027
    if (AV_RN32A(mv)) {
1028
        static const uint8_t idx[8] = { 0, 1, 2, 1, 2, 1, 2, 1 };
1029
        int mx = (mv->x << luma)&7, mx_idx = idx[mx];
1030
        int my = (mv->y << luma)&7, my_idx = idx[my];
1031

    
1032
        x_off += mv->x >> (3 - luma);
1033
        y_off += mv->y >> (3 - luma);
1034

    
1035
        // edge emulation
1036
        src += y_off * linesize + x_off;
1037
        if (x_off < 2 || x_off >= width  - block_w - 3 ||
1038
            y_off < 2 || y_off >= height - block_h - 3) {
1039
            ff_emulated_edge_mc(s->edge_emu_buffer, src - 2 * linesize - 2, linesize,
1040
                                block_w + 5, block_h + 5,
1041
                                x_off - 2, y_off - 2, width, height);
1042
            src = s->edge_emu_buffer + 2 + linesize * 2;
1043
        }
1044
        mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);
1045
    } else
1046
        mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0);
1047
}
1048

    
1049
static inline void vp8_mc_part(VP8Context *s, uint8_t *dst[3],
1050
                               AVFrame *ref_frame, int x_off, int y_off,
1051
                               int bx_off, int by_off,
1052
                               int block_w, int block_h,
1053
                               int width, int height, VP56mv *mv)
1054
{
1055
    VP56mv uvmv = *mv;
1056

    
1057
    /* Y */
1058
    vp8_mc(s, 1, dst[0] + by_off * s->linesize + bx_off,
1059
           ref_frame->data[0], mv, x_off + bx_off, y_off + by_off,
1060
           block_w, block_h, width, height, s->linesize,
1061
           s->put_pixels_tab[block_w == 8]);
1062

    
1063
    /* U/V */
1064
    if (s->profile == 3) {
1065
        uvmv.x &= ~7;
1066
        uvmv.y &= ~7;
1067
    }
1068
    x_off   >>= 1; y_off   >>= 1;
1069
    bx_off  >>= 1; by_off  >>= 1;
1070
    width   >>= 1; height  >>= 1;
1071
    block_w >>= 1; block_h >>= 1;
1072
    vp8_mc(s, 0, dst[1] + by_off * s->uvlinesize + bx_off,
1073
           ref_frame->data[1], &uvmv, x_off + bx_off, y_off + by_off,
1074
           block_w, block_h, width, height, s->uvlinesize,
1075
           s->put_pixels_tab[1 + (block_w == 4)]);
1076
    vp8_mc(s, 0, dst[2] + by_off * s->uvlinesize + bx_off,
1077
           ref_frame->data[2], &uvmv, x_off + bx_off, y_off + by_off,
1078
           block_w, block_h, width, height, s->uvlinesize,
1079
           s->put_pixels_tab[1 + (block_w == 4)]);
1080
}
1081

    
1082
/* Fetch pixels for estimated mv 4 macroblocks ahead.
1083
 * Optimized for 64-byte cache lines.  Inspired by ffh264 prefetch_motion. */
1084
static inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int x_off, int y_off, int ref)
1085
{
1086
    int mx = mb->mv.x + x_off + 8;
1087
    int my = mb->mv.y + y_off;
1088
    uint8_t **src= s->framep[ref]->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
 * Apply motion vectors to prediction buffer, chapter 18.
1097
 */
1098
static void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
1099
                          int mb_x, int mb_y)
1100
{
1101
    int x_off = mb_x << 4, y_off = mb_y << 4;
1102
    int width = 16*s->mb_width, height = 16*s->mb_height;
1103
    AVFrame *ref = s->framep[mb->ref_frame];
1104
    VP56mv *bmv = mb->bmv;
1105

    
1106
    prefetch_motion(s, mb, mb_x, mb_y, x_off, y_off, VP56_FRAME_PREVIOUS);
1107

    
1108
    if (mb->mode < VP8_MVMODE_SPLIT) {
1109
        vp8_mc_part(s, dst, ref, x_off, y_off,
1110
                    0, 0, 16, 16, width, height, &mb->mv);
1111
    } else switch (mb->partitioning) {
1112
    case VP8_SPLITMVMODE_4x4: {
1113
        int x, y;
1114
        VP56mv uvmv;
1115

    
1116
        /* Y */
1117
        for (y = 0; y < 4; y++) {
1118
            for (x = 0; x < 4; x++) {
1119
                vp8_mc(s, 1, dst[0] + 4*y*s->linesize + x*4,
1120
                       ref->data[0], &bmv[4*y + x],
1121
                       4*x + x_off, 4*y + y_off, 4, 4,
1122
                       width, height, s->linesize,
1123
                       s->put_pixels_tab[2]);
1124
            }
1125
        }
1126

    
1127
        /* U/V */
1128
        x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1;
1129
        for (y = 0; y < 2; y++) {
1130
            for (x = 0; x < 2; x++) {
1131
                uvmv.x = mb->bmv[ 2*y    * 4 + 2*x  ].x +
1132
                         mb->bmv[ 2*y    * 4 + 2*x+1].x +
1133
                         mb->bmv[(2*y+1) * 4 + 2*x  ].x +
1134
                         mb->bmv[(2*y+1) * 4 + 2*x+1].x;
1135
                uvmv.y = mb->bmv[ 2*y    * 4 + 2*x  ].y +
1136
                         mb->bmv[ 2*y    * 4 + 2*x+1].y +
1137
                         mb->bmv[(2*y+1) * 4 + 2*x  ].y +
1138
                         mb->bmv[(2*y+1) * 4 + 2*x+1].y;
1139
                uvmv.x = (uvmv.x + 2 + (uvmv.x >> (INT_BIT-1))) >> 2;
1140
                uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2;
1141
                if (s->profile == 3) {
1142
                    uvmv.x &= ~7;
1143
                    uvmv.y &= ~7;
1144
                }
1145
                vp8_mc(s, 0, dst[1] + 4*y*s->uvlinesize + x*4,
1146
                       ref->data[1], &uvmv,
1147
                       4*x + x_off, 4*y + y_off, 4, 4,
1148
                       width, height, s->uvlinesize,
1149
                       s->put_pixels_tab[2]);
1150
                vp8_mc(s, 0, dst[2] + 4*y*s->uvlinesize + x*4,
1151
                       ref->data[2], &uvmv,
1152
                       4*x + x_off, 4*y + y_off, 4, 4,
1153
                       width, height, s->uvlinesize,
1154
                       s->put_pixels_tab[2]);
1155
            }
1156
        }
1157
        break;
1158
    }
1159
    case VP8_SPLITMVMODE_16x8:
1160
        vp8_mc_part(s, dst, ref, x_off, y_off,
1161
                    0, 0, 16, 8, width, height, &bmv[0]);
1162
        vp8_mc_part(s, dst, ref, x_off, y_off,
1163
                    0, 8, 16, 8, width, height, &bmv[1]);
1164
        break;
1165
    case VP8_SPLITMVMODE_8x16:
1166
        vp8_mc_part(s, dst, ref, x_off, y_off,
1167
                    0, 0, 8, 16, width, height, &bmv[0]);
1168
        vp8_mc_part(s, dst, ref, x_off, y_off,
1169
                    8, 0, 8, 16, width, height, &bmv[1]);
1170
        break;
1171
    case VP8_SPLITMVMODE_8x8:
1172
        vp8_mc_part(s, dst, ref, x_off, y_off,
1173
                    0, 0, 8, 8, width, height, &bmv[0]);
1174
        vp8_mc_part(s, dst, ref, x_off, y_off,
1175
                    8, 0, 8, 8, width, height, &bmv[1]);
1176
        vp8_mc_part(s, dst, ref, x_off, y_off,
1177
                    0, 8, 8, 8, width, height, &bmv[2]);
1178
        vp8_mc_part(s, dst, ref, x_off, y_off,
1179
                    8, 8, 8, 8, width, height, &bmv[3]);
1180
        break;
1181
    }
1182

    
1183
    prefetch_motion(s, mb, mb_x, mb_y, x_off, y_off, VP56_FRAME_GOLDEN);
1184
}
1185

    
1186
static void idct_mb(VP8Context *s, uint8_t *y_dst, uint8_t *u_dst, uint8_t *v_dst,
1187
                    VP8Macroblock *mb)
1188
{
1189
    int x, y, nnz;
1190

    
1191
    if (mb->mode != MODE_I4x4)
1192
        for (y = 0; y < 4; y++) {
1193
            for (x = 0; x < 4; x++) {
1194
                nnz = s->non_zero_count_cache[y][x];
1195
                if (nnz) {
1196
                    if (nnz == 1)
1197
                        s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize);
1198
                    else
1199
                        s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize);
1200
                }
1201
            }
1202
            y_dst += 4*s->linesize;
1203
        }
1204

    
1205
    for (y = 0; y < 2; y++) {
1206
        for (x = 0; x < 2; x++) {
1207
            nnz = s->non_zero_count_cache[4][(y<<1)+x];
1208
            if (nnz) {
1209
                if (nnz == 1)
1210
                    s->vp8dsp.vp8_idct_dc_add(u_dst+4*x, s->block[4][(y<<1)+x], s->uvlinesize);
1211
                else
1212
                    s->vp8dsp.vp8_idct_add(u_dst+4*x, s->block[4][(y<<1)+x], s->uvlinesize);
1213
            }
1214

    
1215
            nnz = s->non_zero_count_cache[5][(y<<1)+x];
1216
            if (nnz) {
1217
                if (nnz == 1)
1218
                    s->vp8dsp.vp8_idct_dc_add(v_dst+4*x, s->block[5][(y<<1)+x], s->uvlinesize);
1219
                else
1220
                    s->vp8dsp.vp8_idct_add(v_dst+4*x, s->block[5][(y<<1)+x], s->uvlinesize);
1221
            }
1222
        }
1223
        u_dst += 4*s->uvlinesize;
1224
        v_dst += 4*s->uvlinesize;
1225
    }
1226
}
1227

    
1228
static void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f )
1229
{
1230
    int interior_limit, filter_level;
1231

    
1232
    if (s->segmentation.enabled) {
1233
        filter_level = s->segmentation.filter_level[s->segment];
1234
        if (!s->segmentation.absolute_vals)
1235
            filter_level += s->filter.level;
1236
    } else
1237
        filter_level = s->filter.level;
1238

    
1239
    if (s->lf_delta.enabled) {
1240
        filter_level += s->lf_delta.ref[mb->ref_frame];
1241

    
1242
        if (mb->ref_frame == VP56_FRAME_CURRENT) {
1243
            if (mb->mode == MODE_I4x4)
1244
                filter_level += s->lf_delta.mode[0];
1245
        } else {
1246
            if (mb->mode == VP8_MVMODE_ZERO)
1247
                filter_level += s->lf_delta.mode[1];
1248
            else if (mb->mode == VP8_MVMODE_SPLIT)
1249
                filter_level += s->lf_delta.mode[3];
1250
            else
1251
                filter_level += s->lf_delta.mode[2];
1252
        }
1253
    }
1254
    filter_level = av_clip(filter_level, 0, 63);
1255

    
1256
    interior_limit = filter_level;
1257
    if (s->filter.sharpness) {
1258
        interior_limit >>= s->filter.sharpness > 4 ? 2 : 1;
1259
        interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness);
1260
    }
1261
    interior_limit = FFMAX(interior_limit, 1);
1262

    
1263
    f->filter_level = filter_level;
1264
    f->inner_limit = interior_limit;
1265
    f->inner_filter = !mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT;
1266
}
1267

    
1268
static void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f, int mb_x, int mb_y)
1269
{
1270
    int mbedge_lim, bedge_lim, hev_thresh;
1271
    int filter_level = f->filter_level;
1272
    int inner_limit = f->inner_limit;
1273
    int inner_filter = f->inner_filter;
1274

    
1275
    if (!filter_level)
1276
        return;
1277

    
1278
    mbedge_lim = 2*(filter_level+2) + inner_limit;
1279
     bedge_lim = 2* filter_level    + inner_limit;
1280
    hev_thresh = filter_level >= 15;
1281

    
1282
    if (s->keyframe) {
1283
        if (filter_level >= 40)
1284
            hev_thresh = 2;
1285
    } else {
1286
        if (filter_level >= 40)
1287
            hev_thresh = 3;
1288
        else if (filter_level >= 20)
1289
            hev_thresh = 2;
1290
    }
1291

    
1292
    if (mb_x) {
1293
        s->vp8dsp.vp8_h_loop_filter16y(dst[0],     s->linesize,
1294
                                       mbedge_lim, inner_limit, hev_thresh);
1295
        s->vp8dsp.vp8_h_loop_filter8uv(dst[1],     dst[2],      s->uvlinesize,
1296
                                       mbedge_lim, inner_limit, hev_thresh);
1297
    }
1298

    
1299
    if (inner_filter) {
1300
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 4, s->linesize, bedge_lim,
1301
                                             inner_limit,   hev_thresh);
1302
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 8, s->linesize, bedge_lim,
1303
                                             inner_limit,   hev_thresh);
1304
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+12, s->linesize, bedge_lim,
1305
                                             inner_limit,   hev_thresh);
1306
        s->vp8dsp.vp8_h_loop_filter8uv_inner(dst[1] + 4,    dst[2] + 4,
1307
                                             s->uvlinesize, bedge_lim,
1308
                                             inner_limit,   hev_thresh);
1309
    }
1310

    
1311
    if (mb_y) {
1312
        s->vp8dsp.vp8_v_loop_filter16y(dst[0],     s->linesize,
1313
                                       mbedge_lim, inner_limit, hev_thresh);
1314
        s->vp8dsp.vp8_v_loop_filter8uv(dst[1],     dst[2],      s->uvlinesize,
1315
                                       mbedge_lim, inner_limit, hev_thresh);
1316
    }
1317

    
1318
    if (inner_filter) {
1319
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 4*s->linesize,
1320
                                             s->linesize,   bedge_lim,
1321
                                             inner_limit,   hev_thresh);
1322
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 8*s->linesize,
1323
                                             s->linesize,   bedge_lim,
1324
                                             inner_limit,   hev_thresh);
1325
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+12*s->linesize,
1326
                                             s->linesize,   bedge_lim,
1327
                                             inner_limit, hev_thresh);
1328
        s->vp8dsp.vp8_v_loop_filter8uv_inner(dst[1] + 4 * s->uvlinesize,
1329
                                             dst[2] + 4 * s->uvlinesize,
1330
                                             s->uvlinesize, bedge_lim,
1331
                                             inner_limit,   hev_thresh);
1332
    }
1333
}
1334

    
1335
static void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f, int mb_x, int mb_y)
1336
{
1337
    int mbedge_lim, bedge_lim;
1338
    int filter_level = f->filter_level;
1339
    int inner_limit = f->inner_limit;
1340
    int inner_filter = f->inner_filter;
1341

    
1342
    if (!filter_level)
1343
        return;
1344

    
1345
    mbedge_lim = 2*(filter_level+2) + inner_limit;
1346
     bedge_lim = 2* filter_level    + inner_limit;
1347

    
1348
    if (mb_x)
1349
        s->vp8dsp.vp8_h_loop_filter_simple(dst, s->linesize, mbedge_lim);
1350
    if (inner_filter) {
1351
        s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, s->linesize, bedge_lim);
1352
        s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, s->linesize, bedge_lim);
1353
        s->vp8dsp.vp8_h_loop_filter_simple(dst+12, s->linesize, bedge_lim);
1354
    }
1355

    
1356
    if (mb_y)
1357
        s->vp8dsp.vp8_v_loop_filter_simple(dst, s->linesize, mbedge_lim);
1358
    if (inner_filter) {
1359
        s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*s->linesize, s->linesize, bedge_lim);
1360
        s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*s->linesize, s->linesize, bedge_lim);
1361
        s->vp8dsp.vp8_v_loop_filter_simple(dst+12*s->linesize, s->linesize, bedge_lim);
1362
    }
1363
}
1364

    
1365
static void filter_mb_row(VP8Context *s, int mb_y)
1366
{
1367
    VP8FilterStrength *f = s->filter_strength;
1368
    uint8_t *dst[3] = {
1369
        s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize,
1370
        s->framep[VP56_FRAME_CURRENT]->data[1] +  8*mb_y*s->uvlinesize,
1371
        s->framep[VP56_FRAME_CURRENT]->data[2] +  8*mb_y*s->uvlinesize
1372
    };
1373
    int mb_x;
1374

    
1375
    for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1376
        backup_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], s->linesize, s->uvlinesize, 0);
1377
        filter_mb(s, dst, f++, mb_x, mb_y);
1378
        dst[0] += 16;
1379
        dst[1] += 8;
1380
        dst[2] += 8;
1381
    }
1382
}
1383

    
1384
static void filter_mb_row_simple(VP8Context *s, int mb_y)
1385
{
1386
    VP8FilterStrength *f = s->filter_strength;
1387
    uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize;
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, NULL, NULL, s->linesize, 0, 1);
1392
        filter_mb_simple(s, dst, f++, mb_x, mb_y);
1393
        dst += 16;
1394
    }
1395
}
1396

    
1397
static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
1398
                            AVPacket *avpkt)
1399
{
1400
    VP8Context *s = avctx->priv_data;
1401
    int ret, mb_x, mb_y, i, y, referenced;
1402
    enum AVDiscard skip_thresh;
1403
    AVFrame *curframe;
1404

    
1405
    if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0)
1406
        return ret;
1407

    
1408
    referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT
1409
                                || s->update_altref == VP56_FRAME_CURRENT;
1410

    
1411
    skip_thresh = !referenced ? AVDISCARD_NONREF :
1412
                    !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL;
1413

    
1414
    if (avctx->skip_frame >= skip_thresh) {
1415
        s->invisible = 1;
1416
        goto skip_decode;
1417
    }
1418
    s->deblock_filter = s->filter.level && avctx->skip_loop_filter < skip_thresh;
1419

    
1420
    for (i = 0; i < 4; i++)
1421
        if (&s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1422
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1423
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) {
1424
            curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i];
1425
            break;
1426
        }
1427
    if (curframe->data[0])
1428
        avctx->release_buffer(avctx, curframe);
1429

    
1430
    curframe->key_frame = s->keyframe;
1431
    curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE;
1432
    curframe->reference = referenced ? 3 : 0;
1433
    if ((ret = avctx->get_buffer(avctx, curframe))) {
1434
        av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n");
1435
        return ret;
1436
    }
1437

    
1438
    // Given that arithmetic probabilities are updated every frame, it's quite likely
1439
    // that the values we have on a random interframe are complete junk if we didn't
1440
    // start decode on a keyframe. So just don't display anything rather than junk.
1441
    if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] ||
1442
                         !s->framep[VP56_FRAME_GOLDEN] ||
1443
                         !s->framep[VP56_FRAME_GOLDEN2])) {
1444
        av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n");
1445
        return AVERROR_INVALIDDATA;
1446
    }
1447

    
1448
    s->linesize   = curframe->linesize[0];
1449
    s->uvlinesize = curframe->linesize[1];
1450

    
1451
    if (!s->edge_emu_buffer)
1452
        s->edge_emu_buffer = av_malloc(21*s->linesize);
1453

    
1454
    memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz));
1455

    
1456
    /* Zero macroblock structures for top/left prediction from outside the frame. */
1457
    memset(s->macroblocks, 0, (s->mb_width + s->mb_height*2)*sizeof(*s->macroblocks));
1458

    
1459
    // top edge of 127 for intra prediction
1460
    memset(s->top_border, 127, (s->mb_width+1)*sizeof(*s->top_border));
1461

    
1462
    for (mb_y = 0; mb_y < s->mb_height; mb_y++) {
1463
        VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)];
1464
        VP8Macroblock *mb = s->macroblocks + (s->mb_height - mb_y - 1)*2;
1465
        uint8_t *intra4x4 = s->intra4x4_pred_mode + 4*mb_y*s->b4_stride;
1466
        uint8_t *segment_map = s->segmentation_map + mb_y*s->mb_stride;
1467
        uint8_t *dst[3] = {
1468
            curframe->data[0] + 16*mb_y*s->linesize,
1469
            curframe->data[1] +  8*mb_y*s->uvlinesize,
1470
            curframe->data[2] +  8*mb_y*s->uvlinesize
1471
        };
1472

    
1473
        memset(s->left_nnz, 0, sizeof(s->left_nnz));
1474

    
1475
        // left edge of 129 for intra prediction
1476
        if (!(avctx->flags & CODEC_FLAG_EMU_EDGE))
1477
            for (i = 0; i < 3; i++)
1478
                for (y = 0; y < 16>>!!i; y++)
1479
                    dst[i][y*curframe->linesize[i]-1] = 129;
1480
        if (mb_y)
1481
            memset(s->top_border, 129, sizeof(*s->top_border));
1482

    
1483
        for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1484
            uint8_t *intra4x4_mb = s->keyframe ? intra4x4 + 4*mb_x : s->intra4x4_pred_mode_mb;
1485
            uint8_t *segment_mb = segment_map+mb_x;
1486

    
1487
            /* Prefetch the current frame, 4 MBs ahead */
1488
            s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4);
1489
            s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2);
1490

    
1491
            decode_mb_mode(s, mb, mb_x, mb_y, intra4x4_mb, segment_mb);
1492

    
1493
            if (!mb->skip)
1494
                decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);
1495
            else {
1496
                AV_ZERO128(s->non_zero_count_cache);    // luma
1497
                AV_ZERO64(s->non_zero_count_cache[4]);  // chroma
1498
            }
1499

    
1500
            if (mb->mode <= MODE_I4x4)
1501
                intra_predict(s, dst, mb, intra4x4_mb, mb_x, mb_y);
1502
            else
1503
                inter_predict(s, dst, mb, mb_x, mb_y);
1504

    
1505
            if (!mb->skip) {
1506
                idct_mb(s, dst[0], dst[1], dst[2], mb);
1507
            } else {
1508
                AV_ZERO64(s->left_nnz);
1509
                AV_WN64(s->top_nnz[mb_x], 0);   // array of 9, so unaligned
1510

    
1511
                // Reset DC block predictors if they would exist if the mb had coefficients
1512
                if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
1513
                    s->left_nnz[8]      = 0;
1514
                    s->top_nnz[mb_x][8] = 0;
1515
                }
1516
            }
1517

    
1518
            if (s->deblock_filter)
1519
                filter_level_for_mb(s, mb, &s->filter_strength[mb_x]);
1520

    
1521
            dst[0] += 16;
1522
            dst[1] += 8;
1523
            dst[2] += 8;
1524
            mb++;
1525
        }
1526
        if (s->deblock_filter) {
1527
            if (s->filter.simple)
1528
                filter_mb_row_simple(s, mb_y);
1529
            else
1530
                filter_mb_row(s, mb_y);
1531
        }
1532
    }
1533

    
1534
skip_decode:
1535
    // if future frames don't use the updated probabilities,
1536
    // reset them to the values we saved
1537
    if (!s->update_probabilities)
1538
        s->prob[0] = s->prob[1];
1539

    
1540
    // check if golden and altref are swapped
1541
    if (s->update_altref == VP56_FRAME_GOLDEN &&
1542
        s->update_golden == VP56_FRAME_GOLDEN2)
1543
        FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN], s->framep[VP56_FRAME_GOLDEN2]);
1544
    else {
1545
        if (s->update_altref != VP56_FRAME_NONE)
1546
            s->framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref];
1547

    
1548
        if (s->update_golden != VP56_FRAME_NONE)
1549
            s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden];
1550
    }
1551

    
1552
    if (s->update_last) // move cur->prev
1553
        s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT];
1554

    
1555
    // release no longer referenced frames
1556
    for (i = 0; i < 4; i++)
1557
        if (s->frames[i].data[0] &&
1558
            &s->frames[i] != s->framep[VP56_FRAME_CURRENT] &&
1559
            &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1560
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1561
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2])
1562
            avctx->release_buffer(avctx, &s->frames[i]);
1563

    
1564
    if (!s->invisible) {
1565
        *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT];
1566
        *data_size = sizeof(AVFrame);
1567
    }
1568

    
1569
    return avpkt->size;
1570
}
1571

    
1572
static av_cold int vp8_decode_init(AVCodecContext *avctx)
1573
{
1574
    VP8Context *s = avctx->priv_data;
1575

    
1576
    s->avctx = avctx;
1577
    avctx->pix_fmt = PIX_FMT_YUV420P;
1578

    
1579
    dsputil_init(&s->dsp, avctx);
1580
    ff_h264_pred_init(&s->hpc, CODEC_ID_VP8);
1581
    ff_vp8dsp_init(&s->vp8dsp);
1582

    
1583
    // intra pred needs edge emulation among other things
1584
    if (avctx->flags&CODEC_FLAG_EMU_EDGE) {
1585
        av_log(avctx, AV_LOG_ERROR, "Edge emulation not supported\n");
1586
        return AVERROR_PATCHWELCOME;
1587
    }
1588

    
1589
    return 0;
1590
}
1591

    
1592
static av_cold int vp8_decode_free(AVCodecContext *avctx)
1593
{
1594
    vp8_decode_flush(avctx);
1595
    return 0;
1596
}
1597

    
1598
AVCodec vp8_decoder = {
1599
    "vp8",
1600
    AVMEDIA_TYPE_VIDEO,
1601
    CODEC_ID_VP8,
1602
    sizeof(VP8Context),
1603
    vp8_decode_init,
1604
    NULL,
1605
    vp8_decode_free,
1606
    vp8_decode_frame,
1607
    CODEC_CAP_DR1,
1608
    .flush = vp8_decode_flush,
1609
    .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"),
1610
};