Statistics
| Branch: | Revision:

ffmpeg / libavcodec / vp8.c @ 827d43bb

History | View | Annotate | Download (58 KB)

1
/**
2
 * VP8 compatible video decoder
3
 *
4
 * Copyright (C) 2010 David Conrad
5
 * Copyright (C) 2010 Ronald S. Bultje
6
 * Copyright (C) 2010 Jason Garrett-Glaser
7
 *
8
 * This file is part of FFmpeg.
9
 *
10
 * FFmpeg is free software; you can redistribute it and/or
11
 * modify it under the terms of the GNU Lesser General Public
12
 * License as published by the Free Software Foundation; either
13
 * version 2.1 of the License, or (at your option) any later version.
14
 *
15
 * FFmpeg is distributed in the hope that it will be useful,
16
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18
 * Lesser General Public License for more details.
19
 *
20
 * You should have received a copy of the GNU Lesser General Public
21
 * License along with FFmpeg; if not, write to the Free Software
22
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23
 */
24

    
25
#include "avcodec.h"
26
#include "vp56.h"
27
#include "vp8data.h"
28
#include "vp8dsp.h"
29
#include "h264pred.h"
30
#include "rectangle.h"
31

    
32
typedef struct {
33
    uint8_t filter_level;
34
    uint8_t inner_limit;
35
    uint8_t inner_filter;
36
} VP8FilterStrength;
37

    
38
typedef struct {
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_top;
92
    uint8_t intra4x4_pred_mode_left[4];
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
    DECLARE_ALIGNED(16, DCTELEM, block_dc)[16];
121
    uint8_t intra4x4_pred_mode_mb[16];
122

    
123
    int chroma_pred_mode;    ///< 8x8c pred mode of the current macroblock
124
    int segment;             ///< segment of the current macroblock
125

    
126
    int mbskip_enabled;
127
    int sign_bias[4]; ///< one state [0, 1] per ref frame type
128
    int ref_count[3];
129

    
130
    /**
131
     * Base parameters for segmentation, i.e. per-macroblock parameters.
132
     * These must be kept unchanged even if segmentation is not used for
133
     * a frame, since the values persist between interframes.
134
     */
135
    struct {
136
        int enabled;
137
        int absolute_vals;
138
        int update_map;
139
        int8_t base_quant[4];
140
        int8_t filter_level[4];     ///< base loop filter level
141
    } segmentation;
142

    
143
    /**
144
     * Macroblocks can have one of 4 different quants in a frame when
145
     * segmentation is enabled.
146
     * If segmentation is disabled, only the first segment's values are used.
147
     */
148
    struct {
149
        // [0] - DC qmul  [1] - AC qmul
150
        int16_t luma_qmul[2];
151
        int16_t luma_dc_qmul[2];    ///< luma dc-only block quant
152
        int16_t chroma_qmul[2];
153
    } qmat[4];
154

    
155
    struct {
156
        int simple;
157
        int level;
158
        int sharpness;
159
    } filter;
160

    
161
    struct {
162
        int enabled;    ///< whether each mb can have a different strength based on mode/ref
163

    
164
        /**
165
         * filter strength adjustment for the following macroblock modes:
166
         * [0] - i4x4
167
         * [1] - zero mv
168
         * [2] - inter modes except for zero or split mv
169
         * [3] - split mv
170
         *  i16x16 modes never have any adjustment
171
         */
172
        int8_t mode[4];
173

    
174
        /**
175
         * filter strength adjustment for macroblocks that reference:
176
         * [0] - intra / VP56_FRAME_CURRENT
177
         * [1] - VP56_FRAME_PREVIOUS
178
         * [2] - VP56_FRAME_GOLDEN
179
         * [3] - altref / VP56_FRAME_GOLDEN2
180
         */
181
        int8_t ref[4];
182
    } lf_delta;
183

    
184
    /**
185
     * These are all of the updatable probabilities for binary decisions.
186
     * They are only implictly reset on keyframes, making it quite likely
187
     * for an interframe to desync if a prior frame's header was corrupt
188
     * or missing outright!
189
     */
190
    struct {
191
        uint8_t segmentid[3];
192
        uint8_t mbskip;
193
        uint8_t intra;
194
        uint8_t last;
195
        uint8_t golden;
196
        uint8_t pred16x16[4];
197
        uint8_t pred8x8c[3];
198
        uint8_t token[4][8][3][NUM_DCT_TOKENS-1];
199
        uint8_t mvc[2][19];
200
    } prob[2];
201
} VP8Context;
202

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

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

    
213
    av_freep(&s->macroblocks_base);
214
    av_freep(&s->filter_strength);
215
    av_freep(&s->intra4x4_pred_mode_top);
216
    av_freep(&s->top_nnz);
217
    av_freep(&s->edge_emu_buffer);
218
    av_freep(&s->top_border);
219
    av_freep(&s->segmentation_map);
220

    
221
    s->macroblocks        = NULL;
222
}
223

    
224
static int update_dimensions(VP8Context *s, int width, int height)
225
{
226
    if (avcodec_check_dimensions(s->avctx, width, height))
227
        return AVERROR_INVALIDDATA;
228

    
229
    vp8_decode_flush(s->avctx);
230

    
231
    avcodec_set_dimensions(s->avctx, width, height);
232

    
233
    s->mb_width  = (s->avctx->coded_width +15) / 16;
234
    s->mb_height = (s->avctx->coded_height+15) / 16;
235

    
236
    // we allocate a border around the top/left of intra4x4 modes
237
    // this is 4 blocks for intra4x4 to keep 4-byte alignment for fill_rectangle
238
    s->mb_stride = s->mb_width+1;
239
    s->b4_stride = 4*s->mb_stride;
240

    
241
    s->macroblocks_base        = av_mallocz((s->mb_stride+s->mb_height*2+2)*sizeof(*s->macroblocks));
242
    s->filter_strength         = av_mallocz(s->mb_stride*sizeof(*s->filter_strength));
243
    s->intra4x4_pred_mode_top  = av_mallocz(s->b4_stride*4);
244
    s->top_nnz                 = av_mallocz(s->mb_width*sizeof(*s->top_nnz));
245
    s->top_border              = av_mallocz((s->mb_width+1)*sizeof(*s->top_border));
246
    s->segmentation_map        = av_mallocz(s->mb_stride*s->mb_height);
247

    
248
    if (!s->macroblocks_base || !s->filter_strength || !s->intra4x4_pred_mode_top ||
249
        !s->top_nnz || !s->top_border || !s->segmentation_map)
250
        return AVERROR(ENOMEM);
251

    
252
    s->macroblocks        = s->macroblocks_base + 1;
253

    
254
    return 0;
255
}
256

    
257
static void parse_segment_info(VP8Context *s)
258
{
259
    VP56RangeCoder *c = &s->c;
260
    int i;
261

    
262
    s->segmentation.update_map = vp8_rac_get(c);
263

    
264
    if (vp8_rac_get(c)) { // update segment feature data
265
        s->segmentation.absolute_vals = vp8_rac_get(c);
266

    
267
        for (i = 0; i < 4; i++)
268
            s->segmentation.base_quant[i]   = vp8_rac_get_sint(c, 7);
269

    
270
        for (i = 0; i < 4; i++)
271
            s->segmentation.filter_level[i] = vp8_rac_get_sint(c, 6);
272
    }
273
    if (s->segmentation.update_map)
274
        for (i = 0; i < 3; i++)
275
            s->prob->segmentid[i] = vp8_rac_get(c) ? vp8_rac_get_uint(c, 8) : 255;
276
}
277

    
278
static void update_lf_deltas(VP8Context *s)
279
{
280
    VP56RangeCoder *c = &s->c;
281
    int i;
282

    
283
    for (i = 0; i < 4; i++)
284
        s->lf_delta.ref[i]  = vp8_rac_get_sint(c, 6);
285

    
286
    for (i = 0; i < 4; i++)
287
        s->lf_delta.mode[i] = vp8_rac_get_sint(c, 6);
288
}
289

    
290
static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size)
291
{
292
    const uint8_t *sizes = buf;
293
    int i;
294

    
295
    s->num_coeff_partitions = 1 << vp8_rac_get_uint(&s->c, 2);
296

    
297
    buf      += 3*(s->num_coeff_partitions-1);
298
    buf_size -= 3*(s->num_coeff_partitions-1);
299
    if (buf_size < 0)
300
        return -1;
301

    
302
    for (i = 0; i < s->num_coeff_partitions-1; i++) {
303
        int size = AV_RL24(sizes + 3*i);
304
        if (buf_size - size < 0)
305
            return -1;
306

    
307
        vp56_init_range_decoder(&s->coeff_partition[i], buf, size);
308
        buf      += size;
309
        buf_size -= size;
310
    }
311
    vp56_init_range_decoder(&s->coeff_partition[i], buf, buf_size);
312

    
313
    return 0;
314
}
315

    
316
static void get_quants(VP8Context *s)
317
{
318
    VP56RangeCoder *c = &s->c;
319
    int i, base_qi;
320

    
321
    int yac_qi     = vp8_rac_get_uint(c, 7);
322
    int ydc_delta  = vp8_rac_get_sint(c, 4);
323
    int y2dc_delta = vp8_rac_get_sint(c, 4);
324
    int y2ac_delta = vp8_rac_get_sint(c, 4);
325
    int uvdc_delta = vp8_rac_get_sint(c, 4);
326
    int uvac_delta = vp8_rac_get_sint(c, 4);
327

    
328
    for (i = 0; i < 4; i++) {
329
        if (s->segmentation.enabled) {
330
            base_qi = s->segmentation.base_quant[i];
331
            if (!s->segmentation.absolute_vals)
332
                base_qi += yac_qi;
333
        } else
334
            base_qi = yac_qi;
335

    
336
        s->qmat[i].luma_qmul[0]    =       vp8_dc_qlookup[av_clip(base_qi + ydc_delta , 0, 127)];
337
        s->qmat[i].luma_qmul[1]    =       vp8_ac_qlookup[av_clip(base_qi             , 0, 127)];
338
        s->qmat[i].luma_dc_qmul[0] =   2 * vp8_dc_qlookup[av_clip(base_qi + y2dc_delta, 0, 127)];
339
        s->qmat[i].luma_dc_qmul[1] = 155 * vp8_ac_qlookup[av_clip(base_qi + y2ac_delta, 0, 127)] / 100;
340
        s->qmat[i].chroma_qmul[0]  =       vp8_dc_qlookup[av_clip(base_qi + uvdc_delta, 0, 127)];
341
        s->qmat[i].chroma_qmul[1]  =       vp8_ac_qlookup[av_clip(base_qi + uvac_delta, 0, 127)];
342

    
343
        s->qmat[i].luma_dc_qmul[1] = FFMAX(s->qmat[i].luma_dc_qmul[1], 8);
344
        s->qmat[i].chroma_qmul[0]  = FFMIN(s->qmat[i].chroma_qmul[0], 132);
345
    }
346
}
347

    
348
/**
349
 * Determine which buffers golden and altref should be updated with after this frame.
350
 * The spec isn't clear here, so I'm going by my understanding of what libvpx does
351
 *
352
 * Intra frames update all 3 references
353
 * Inter frames update VP56_FRAME_PREVIOUS if the update_last flag is set
354
 * If the update (golden|altref) flag is set, it's updated with the current frame
355
 *      if update_last is set, and VP56_FRAME_PREVIOUS otherwise.
356
 * If the flag is not set, the number read means:
357
 *      0: no update
358
 *      1: VP56_FRAME_PREVIOUS
359
 *      2: update golden with altref, or update altref with golden
360
 */
361
static VP56Frame ref_to_update(VP8Context *s, int update, VP56Frame ref)
362
{
363
    VP56RangeCoder *c = &s->c;
364

    
365
    if (update)
366
        return VP56_FRAME_CURRENT;
367

    
368
    switch (vp8_rac_get_uint(c, 2)) {
369
    case 1:
370
        return VP56_FRAME_PREVIOUS;
371
    case 2:
372
        return (ref == VP56_FRAME_GOLDEN) ? VP56_FRAME_GOLDEN2 : VP56_FRAME_GOLDEN;
373
    }
374
    return VP56_FRAME_NONE;
375
}
376

    
377
static void update_refs(VP8Context *s)
378
{
379
    VP56RangeCoder *c = &s->c;
380

    
381
    int update_golden = vp8_rac_get(c);
382
    int update_altref = vp8_rac_get(c);
383

    
384
    s->update_golden = ref_to_update(s, update_golden, VP56_FRAME_GOLDEN);
385
    s->update_altref = ref_to_update(s, update_altref, VP56_FRAME_GOLDEN2);
386
}
387

    
388
static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)
389
{
390
    VP56RangeCoder *c = &s->c;
391
    int header_size, hscale, vscale, i, j, k, l, ret;
392
    int width  = s->avctx->width;
393
    int height = s->avctx->height;
394

    
395
    s->keyframe  = !(buf[0] & 1);
396
    s->profile   =  (buf[0]>>1) & 7;
397
    s->invisible = !(buf[0] & 0x10);
398
    header_size  = AV_RL24(buf) >> 5;
399
    buf      += 3;
400
    buf_size -= 3;
401

    
402
    if (s->profile > 3)
403
        av_log(s->avctx, AV_LOG_WARNING, "Unknown profile %d\n", s->profile);
404

    
405
    if (!s->profile)
406
        memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_epel_pixels_tab, sizeof(s->put_pixels_tab));
407
    else    // profile 1-3 use bilinear, 4+ aren't defined so whatever
408
        memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_bilinear_pixels_tab, sizeof(s->put_pixels_tab));
409

    
410
    if (header_size > buf_size - 7*s->keyframe) {
411
        av_log(s->avctx, AV_LOG_ERROR, "Header size larger than data provided\n");
412
        return AVERROR_INVALIDDATA;
413
    }
414

    
415
    if (s->keyframe) {
416
        if (AV_RL24(buf) != 0x2a019d) {
417
            av_log(s->avctx, AV_LOG_ERROR, "Invalid start code 0x%x\n", AV_RL24(buf));
418
            return AVERROR_INVALIDDATA;
419
        }
420
        width  = AV_RL16(buf+3) & 0x3fff;
421
        height = AV_RL16(buf+5) & 0x3fff;
422
        hscale = buf[4] >> 6;
423
        vscale = buf[6] >> 6;
424
        buf      += 7;
425
        buf_size -= 7;
426

    
427
        if (hscale || vscale)
428
            av_log_missing_feature(s->avctx, "Upscaling", 1);
429

    
430
        s->update_golden = s->update_altref = VP56_FRAME_CURRENT;
431
        memcpy(s->prob->token    , vp8_token_default_probs , sizeof(s->prob->token));
432
        memcpy(s->prob->pred16x16, vp8_pred16x16_prob_inter, sizeof(s->prob->pred16x16));
433
        memcpy(s->prob->pred8x8c , vp8_pred8x8c_prob_inter , sizeof(s->prob->pred8x8c));
434
        memcpy(s->prob->mvc      , vp8_mv_default_prob     , sizeof(s->prob->mvc));
435
        memset(&s->segmentation, 0, sizeof(s->segmentation));
436
    }
437

    
438
    if (!s->macroblocks_base || /* first frame */
439
        width != s->avctx->width || height != s->avctx->height) {
440
        if ((ret = update_dimensions(s, width, height) < 0))
441
            return ret;
442
    }
443

    
444
    vp56_init_range_decoder(c, buf, header_size);
445
    buf      += header_size;
446
    buf_size -= header_size;
447

    
448
    if (s->keyframe) {
449
        if (vp8_rac_get(c))
450
            av_log(s->avctx, AV_LOG_WARNING, "Unspecified colorspace\n");
451
        vp8_rac_get(c); // whether we can skip clamping in dsp functions
452
    }
453

    
454
    if ((s->segmentation.enabled = vp8_rac_get(c)))
455
        parse_segment_info(s);
456
    else
457
        s->segmentation.update_map = 0; // FIXME: move this to some init function?
458

    
459
    s->filter.simple    = vp8_rac_get(c);
460
    s->filter.level     = vp8_rac_get_uint(c, 6);
461
    s->filter.sharpness = vp8_rac_get_uint(c, 3);
462

    
463
    if ((s->lf_delta.enabled = vp8_rac_get(c)))
464
        if (vp8_rac_get(c))
465
            update_lf_deltas(s);
466

    
467
    if (setup_partitions(s, buf, buf_size)) {
468
        av_log(s->avctx, AV_LOG_ERROR, "Invalid partitions\n");
469
        return AVERROR_INVALIDDATA;
470
    }
471

    
472
    get_quants(s);
473

    
474
    if (!s->keyframe) {
475
        update_refs(s);
476
        s->sign_bias[VP56_FRAME_GOLDEN]               = vp8_rac_get(c);
477
        s->sign_bias[VP56_FRAME_GOLDEN2 /* altref */] = vp8_rac_get(c);
478
    }
479

    
480
    // if we aren't saving this frame's probabilities for future frames,
481
    // make a copy of the current probabilities
482
    if (!(s->update_probabilities = vp8_rac_get(c)))
483
        s->prob[1] = s->prob[0];
484

    
485
    s->update_last = s->keyframe || vp8_rac_get(c);
486

    
487
    for (i = 0; i < 4; i++)
488
        for (j = 0; j < 8; j++)
489
            for (k = 0; k < 3; k++)
490
                for (l = 0; l < NUM_DCT_TOKENS-1; l++)
491
                    if (vp56_rac_get_prob_branchy(c, vp8_token_update_probs[i][j][k][l]))
492
                        s->prob->token[i][j][k][l] = vp8_rac_get_uint(c, 8);
493

    
494
    if ((s->mbskip_enabled = vp8_rac_get(c)))
495
        s->prob->mbskip = vp8_rac_get_uint(c, 8);
496

    
497
    if (!s->keyframe) {
498
        s->prob->intra  = vp8_rac_get_uint(c, 8);
499
        s->prob->last   = vp8_rac_get_uint(c, 8);
500
        s->prob->golden = vp8_rac_get_uint(c, 8);
501

    
502
        if (vp8_rac_get(c))
503
            for (i = 0; i < 4; i++)
504
                s->prob->pred16x16[i] = vp8_rac_get_uint(c, 8);
505
        if (vp8_rac_get(c))
506
            for (i = 0; i < 3; i++)
507
                s->prob->pred8x8c[i]  = vp8_rac_get_uint(c, 8);
508

    
509
        // 17.2 MV probability update
510
        for (i = 0; i < 2; i++)
511
            for (j = 0; j < 19; j++)
512
                if (vp56_rac_get_prob_branchy(c, vp8_mv_update_prob[i][j]))
513
                    s->prob->mvc[i][j] = vp8_rac_get_nn(c);
514
    }
515

    
516
    return 0;
517
}
518

    
519
static av_always_inline
520
void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src, int mb_x, int mb_y)
521
{
522
#define MARGIN (16 << 2)
523
    dst->x = av_clip(src->x, -((mb_x << 6) + MARGIN),
524
                     ((s->mb_width  - 1 - mb_x) << 6) + MARGIN);
525
    dst->y = av_clip(src->y, -((mb_y << 6) + MARGIN),
526
                     ((s->mb_height - 1 - mb_y) << 6) + MARGIN);
527
}
528

    
529
static av_always_inline
530
void find_near_mvs(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
531
                   VP56mv near[2], VP56mv *best, uint8_t cnt[4])
532
{
533
    VP8Macroblock *mb_edge[3] = { mb + 2 /* top */,
534
                                  mb - 1 /* left */,
535
                                  mb + 1 /* top-left */ };
536
    enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT };
537
    VP56mv near_mv[4]  = {{ 0 }};
538
    enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV };
539
    int idx = CNT_ZERO;
540
    int best_idx = CNT_ZERO;
541
    int cur_sign_bias = s->sign_bias[mb->ref_frame];
542
    int *sign_bias = s->sign_bias;
543

    
544
    /* Process MB on top, left and top-left */
545
    #define MV_EDGE_CHECK(n)\
546
    {\
547
        VP8Macroblock *edge = mb_edge[n];\
548
        int edge_ref = edge->ref_frame;\
549
        if (edge_ref != VP56_FRAME_CURRENT) {\
550
            uint32_t mv = AV_RN32A(&edge->mv);\
551
            if (mv) {\
552
                if (cur_sign_bias != sign_bias[edge_ref]) {\
553
                    /* SWAR negate of the values in mv. */\
554
                    mv = ~mv;\
555
                    mv = ((mv&0x7fff7fff) + 0x00010001) ^ (mv&0x80008000);\
556
                }\
557
                if (!n || mv != AV_RN32A(&near_mv[idx]))\
558
                    AV_WN32A(&near_mv[++idx], mv);\
559
                cnt[idx]      += 1 + (n != 2);\
560
            } else\
561
                cnt[CNT_ZERO] += 1 + (n != 2);\
562
        }\
563
    }
564
    MV_EDGE_CHECK(0)
565
    MV_EDGE_CHECK(1)
566
    MV_EDGE_CHECK(2)
567

    
568
    /* If we have three distinct MVs, merge first and last if they're the same */
569
    if (cnt[CNT_SPLITMV] && AV_RN32A(&near_mv[1+EDGE_TOP]) == AV_RN32A(&near_mv[1+EDGE_TOPLEFT]))
570
        cnt[CNT_NEAREST] += 1;
571

    
572
    cnt[CNT_SPLITMV] = ((mb_edge[EDGE_LEFT]->mode   == VP8_MVMODE_SPLIT) +
573
                        (mb_edge[EDGE_TOP]->mode    == VP8_MVMODE_SPLIT)) * 2 +
574
                       (mb_edge[EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT);
575

    
576
    /* Swap near and nearest if necessary */
577
    if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) {
578
        FFSWAP(uint8_t,     cnt[CNT_NEAREST],     cnt[CNT_NEAR]);
579
        FFSWAP( VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]);
580
    }
581

    
582
    /* Choose the best mv out of 0,0 and the nearest mv */
583
    if (cnt[CNT_NEAREST] >= cnt[CNT_ZERO])
584
        best_idx = CNT_NEAREST;
585

    
586
    mb->mv  = near_mv[best_idx];
587
    near[0] = near_mv[CNT_NEAREST];
588
    near[1] = near_mv[CNT_NEAR];
589
}
590

    
591
/**
592
 * Motion vector coding, 17.1.
593
 */
594
static int read_mv_component(VP56RangeCoder *c, const uint8_t *p)
595
{
596
    int bit, x = 0;
597

    
598
    if (vp56_rac_get_prob_branchy(c, p[0])) {
599
        int i;
600

    
601
        for (i = 0; i < 3; i++)
602
            x += vp56_rac_get_prob(c, p[9 + i]) << i;
603
        for (i = 9; i > 3; i--)
604
            x += vp56_rac_get_prob(c, p[9 + i]) << i;
605
        if (!(x & 0xFFF0) || vp56_rac_get_prob(c, p[12]))
606
            x += 8;
607
    } else {
608
        // small_mvtree
609
        const uint8_t *ps = p+2;
610
        bit = vp56_rac_get_prob(c, *ps);
611
        ps += 1 + 3*bit;
612
        x  += 4*bit;
613
        bit = vp56_rac_get_prob(c, *ps);
614
        ps += 1 + bit;
615
        x  += 2*bit;
616
        x  += vp56_rac_get_prob(c, *ps);
617
    }
618

    
619
    return (x && vp56_rac_get_prob(c, p[1])) ? -x : x;
620
}
621

    
622
static av_always_inline
623
const uint8_t *get_submv_prob(uint32_t left, uint32_t top)
624
{
625
    if (left == top)
626
        return vp8_submv_prob[4-!!left];
627
    if (!top)
628
        return vp8_submv_prob[2];
629
    return vp8_submv_prob[1-!!left];
630
}
631

    
632
/**
633
 * Split motion vector prediction, 16.4.
634
 * @returns the number of motion vectors parsed (2, 4 or 16)
635
 */
636
static av_always_inline
637
int decode_splitmvs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb)
638
{
639
    int part_idx = mb->partitioning =
640
        vp8_rac_get_tree(c, vp8_mbsplit_tree, vp8_mbsplit_prob);
641
    int n, num = vp8_mbsplit_count[part_idx];
642
    VP8Macroblock *top_mb  = &mb[2];
643
    VP8Macroblock *left_mb = &mb[-1];
644
    const uint8_t *mbsplits_left = vp8_mbsplits[left_mb->partitioning],
645
                  *mbsplits_top = vp8_mbsplits[top_mb->partitioning],
646
                  *mbsplits_cur = vp8_mbsplits[part_idx],
647
                  *firstidx = vp8_mbfirstidx[part_idx];
648
    VP56mv *top_mv  = top_mb->bmv;
649
    VP56mv *left_mv = left_mb->bmv;
650
    VP56mv *cur_mv  = mb->bmv;
651

    
652
    for (n = 0; n < num; n++) {
653
        int k = firstidx[n];
654
        uint32_t left, above;
655
        const uint8_t *submv_prob;
656

    
657
        if (!(k & 3))
658
            left = AV_RN32A(&left_mv[mbsplits_left[k + 3]]);
659
        else
660
            left  = AV_RN32A(&cur_mv[mbsplits_cur[k - 1]]);
661
        if (k <= 3)
662
            above = AV_RN32A(&top_mv[mbsplits_top[k + 12]]);
663
        else
664
            above = AV_RN32A(&cur_mv[mbsplits_cur[k - 4]]);
665

    
666
        submv_prob = get_submv_prob(left, above);
667

    
668
        switch (vp8_rac_get_tree(c, vp8_submv_ref_tree, submv_prob)) {
669
        case VP8_SUBMVMODE_NEW4X4:
670
            mb->bmv[n].y = mb->mv.y + read_mv_component(c, s->prob->mvc[0]);
671
            mb->bmv[n].x = mb->mv.x + read_mv_component(c, s->prob->mvc[1]);
672
            break;
673
        case VP8_SUBMVMODE_ZERO4X4:
674
            AV_ZERO32(&mb->bmv[n]);
675
            break;
676
        case VP8_SUBMVMODE_LEFT4X4:
677
            AV_WN32A(&mb->bmv[n], left);
678
            break;
679
        case VP8_SUBMVMODE_TOP4X4:
680
            AV_WN32A(&mb->bmv[n], above);
681
            break;
682
        }
683
    }
684

    
685
    return num;
686
}
687

    
688
static av_always_inline
689
void decode_intra4x4_modes(VP8Context *s, VP56RangeCoder *c,
690
                           int mb_x, int keyframe)
691
{
692
    uint8_t *intra4x4 = s->intra4x4_pred_mode_mb;
693
    if (keyframe) {
694
        int x, y;
695
        uint8_t* const top = s->intra4x4_pred_mode_top + 4 * mb_x;
696
        uint8_t* const left = s->intra4x4_pred_mode_left;
697
        for (y = 0; y < 4; y++) {
698
            for (x = 0; x < 4; x++) {
699
                const uint8_t *ctx;
700
                ctx = vp8_pred4x4_prob_intra[top[x]][left[y]];
701
                *intra4x4 = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx);
702
                left[y] = top[x] = *intra4x4;
703
                intra4x4++;
704
            }
705
        }
706
    } else {
707
        int i;
708
        for (i = 0; i < 16; i++)
709
            intra4x4[i] = vp8_rac_get_tree(c, vp8_pred4x4_tree, vp8_pred4x4_prob_inter);
710
    }
711
}
712

    
713
static av_always_inline
714
void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, uint8_t *segment)
715
{
716
    VP56RangeCoder *c = &s->c;
717

    
718
    if (s->segmentation.update_map)
719
        *segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid);
720
    s->segment = *segment;
721

    
722
    mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0;
723

    
724
    if (s->keyframe) {
725
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra);
726

    
727
        if (mb->mode == MODE_I4x4) {
728
            decode_intra4x4_modes(s, c, mb_x, 1);
729
        } else {
730
            const uint32_t modes = vp8_pred4x4_mode[mb->mode] * 0x01010101u;
731
            AV_WN32A(s->intra4x4_pred_mode_top + 4 * mb_x, modes);
732
            AV_WN32A(s->intra4x4_pred_mode_left, modes);
733
        }
734

    
735
        s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra);
736
        mb->ref_frame = VP56_FRAME_CURRENT;
737
    } else if (vp56_rac_get_prob_branchy(c, s->prob->intra)) {
738
        VP56mv near[2], best;
739
        uint8_t cnt[4] = { 0 };
740
        uint8_t p[4];
741

    
742
        // inter MB, 16.2
743
        if (vp56_rac_get_prob_branchy(c, s->prob->last))
744
            mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ?
745
                VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN;
746
        else
747
            mb->ref_frame = VP56_FRAME_PREVIOUS;
748
        s->ref_count[mb->ref_frame-1]++;
749

    
750
        // motion vectors, 16.3
751
        find_near_mvs(s, mb, mb_x, mb_y, near, &best, cnt);
752
        p[0] = vp8_mode_contexts[cnt[0]][0];
753
        p[1] = vp8_mode_contexts[cnt[1]][1];
754
        p[2] = vp8_mode_contexts[cnt[2]][2];
755
        p[3] = vp8_mode_contexts[cnt[3]][3];
756
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_mvinter, p);
757
        switch (mb->mode) {
758
        case VP8_MVMODE_SPLIT:
759
            clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y);
760
            mb->mv = mb->bmv[decode_splitmvs(s, c, mb) - 1];
761
            break;
762
        case VP8_MVMODE_ZERO:
763
            AV_ZERO32(&mb->mv);
764
            break;
765
        case VP8_MVMODE_NEAREST:
766
            clamp_mv(s, &mb->mv, &near[0], mb_x, mb_y);
767
            break;
768
        case VP8_MVMODE_NEAR:
769
            clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y);
770
            break;
771
        case VP8_MVMODE_NEW:
772
            clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y);
773
            mb->mv.y += + read_mv_component(c, s->prob->mvc[0]);
774
            mb->mv.x += + read_mv_component(c, s->prob->mvc[1]);
775
            break;
776
        }
777
        if (mb->mode != VP8_MVMODE_SPLIT) {
778
            mb->partitioning = VP8_SPLITMVMODE_NONE;
779
            mb->bmv[0] = mb->mv;
780
        }
781
    } else {
782
        // intra MB, 16.1
783
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16);
784

    
785
        if (mb->mode == MODE_I4x4)
786
            decode_intra4x4_modes(s, c, mb_x, 0);
787

    
788
        s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c);
789
        mb->ref_frame = VP56_FRAME_CURRENT;
790
        mb->partitioning = VP8_SPLITMVMODE_NONE;
791
        AV_ZERO32(&mb->bmv[0]);
792
    }
793
}
794

    
795
/**
796
 * @param c arithmetic bitstream reader context
797
 * @param block destination for block coefficients
798
 * @param probs probabilities to use when reading trees from the bitstream
799
 * @param i initial coeff index, 0 unless a separate DC block is coded
800
 * @param zero_nhood the initial prediction context for number of surrounding
801
 *                   all-zero blocks (only left/top, so 0-2)
802
 * @param qmul array holding the dc/ac dequant factor at position 0/1
803
 * @return 0 if no coeffs were decoded
804
 *         otherwise, the index of the last coeff decoded plus one
805
 */
806
static int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16],
807
                               uint8_t probs[8][3][NUM_DCT_TOKENS-1],
808
                               int i, int zero_nhood, int16_t qmul[2])
809
{
810
    uint8_t *token_prob = probs[vp8_coeff_band[i]][zero_nhood];
811
    int nonzero = 0;
812
    int coeff;
813

    
814
    do {
815
        if (!vp56_rac_get_prob_branchy(c, token_prob[0]))   // DCT_EOB
816
            return nonzero;
817

    
818
skip_eob:
819
        if (!vp56_rac_get_prob_branchy(c, token_prob[1])) { // DCT_0
820
            if (++i == 16)
821
                return nonzero; // invalid input; blocks should end with EOB
822
            token_prob = probs[vp8_coeff_band[i]][0];
823
            goto skip_eob;
824
        }
825

    
826
        if (!vp56_rac_get_prob_branchy(c, token_prob[2])) { // DCT_1
827
            coeff = 1;
828
            token_prob = probs[vp8_coeff_band[i+1]][1];
829
        } else {
830
            if (!vp56_rac_get_prob_branchy(c, token_prob[3])) { // DCT 2,3,4
831
                coeff = vp56_rac_get_prob(c, token_prob[4]);
832
                if (coeff)
833
                    coeff += vp56_rac_get_prob(c, token_prob[5]);
834
                coeff += 2;
835
            } else {
836
                // DCT_CAT*
837
                if (!vp56_rac_get_prob_branchy(c, token_prob[6])) {
838
                    if (!vp56_rac_get_prob_branchy(c, token_prob[7])) { // DCT_CAT1
839
                        coeff  = 5 + vp56_rac_get_prob(c, vp8_dct_cat1_prob[0]);
840
                    } else {                                    // DCT_CAT2
841
                        coeff  = 7;
842
                        coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[0]) << 1;
843
                        coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[1]);
844
                    }
845
                } else {    // DCT_CAT3 and up
846
                    int a = vp56_rac_get_prob(c, token_prob[8]);
847
                    int b = vp56_rac_get_prob(c, token_prob[9+a]);
848
                    int cat = (a<<1) + b;
849
                    coeff  = 3 + (8<<cat);
850
                    coeff += vp8_rac_get_coeff(c, vp8_dct_cat_prob[cat]);
851
                }
852
            }
853
            token_prob = probs[vp8_coeff_band[i+1]][2];
854
        }
855

    
856
        // todo: full [16] qmat? load into register?
857
        block[zigzag_scan[i]] = (vp8_rac_get(c) ? -coeff : coeff) * qmul[!!i];
858
        nonzero = ++i;
859
    } while (i < 16);
860

    
861
    return nonzero;
862
}
863

    
864
static av_always_inline
865
void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,
866
                      uint8_t t_nnz[9], uint8_t l_nnz[9])
867
{
868
    int i, x, y, luma_start = 0, luma_ctx = 3;
869
    int nnz_pred, nnz, nnz_total = 0;
870
    int segment = s->segment;
871

    
872
    if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
873
        nnz_pred = t_nnz[8] + l_nnz[8];
874

    
875
        // decode DC values and do hadamard
876
        nnz = decode_block_coeffs(c, s->block_dc, s->prob->token[1], 0, nnz_pred,
877
                                  s->qmat[segment].luma_dc_qmul);
878
        l_nnz[8] = t_nnz[8] = !!nnz;
879
        nnz_total += nnz;
880
        s->vp8dsp.vp8_luma_dc_wht(s->block, s->block_dc);
881
        luma_start = 1;
882
        luma_ctx = 0;
883
    }
884

    
885
    // luma blocks
886
    for (y = 0; y < 4; y++)
887
        for (x = 0; x < 4; x++) {
888
            nnz_pred = l_nnz[y] + t_nnz[x];
889
            nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start,
890
                                      nnz_pred, s->qmat[segment].luma_qmul);
891
            // nnz+luma_start may be one more than the actual last index, but we don't care
892
            s->non_zero_count_cache[y][x] = nnz + luma_start;
893
            t_nnz[x] = l_nnz[y] = !!nnz;
894
            nnz_total += nnz;
895
        }
896

    
897
    // chroma blocks
898
    // TODO: what to do about dimensions? 2nd dim for luma is x,
899
    // but for chroma it's (y<<1)|x
900
    for (i = 4; i < 6; i++)
901
        for (y = 0; y < 2; y++)
902
            for (x = 0; x < 2; x++) {
903
                nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x];
904
                nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0,
905
                                          nnz_pred, s->qmat[segment].chroma_qmul);
906
                s->non_zero_count_cache[i][(y<<1)+x] = nnz;
907
                t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz;
908
                nnz_total += nnz;
909
            }
910

    
911
    // if there were no coded coeffs despite the macroblock not being marked skip,
912
    // we MUST not do the inner loop filter and should not do IDCT
913
    // Since skip isn't used for bitstream prediction, just manually set it.
914
    if (!nnz_total)
915
        mb->skip = 1;
916
}
917

    
918
static av_always_inline
919
void backup_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
920
                      int linesize, int uvlinesize, int simple)
921
{
922
    AV_COPY128(top_border, src_y + 15*linesize);
923
    if (!simple) {
924
        AV_COPY64(top_border+16, src_cb + 7*uvlinesize);
925
        AV_COPY64(top_border+24, src_cr + 7*uvlinesize);
926
    }
927
}
928

    
929
static av_always_inline
930
void xchg_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
931
                    int linesize, int uvlinesize, int mb_x, int mb_y, int mb_width,
932
                    int simple, int xchg)
933
{
934
    uint8_t *top_border_m1 = top_border-32;     // for TL prediction
935
    src_y  -=   linesize;
936
    src_cb -= uvlinesize;
937
    src_cr -= uvlinesize;
938

    
939
#define XCHG(a,b,xchg) do {                     \
940
        if (xchg) AV_SWAP64(b,a);               \
941
        else      AV_COPY64(b,a);               \
942
    } while (0)
943

    
944
    XCHG(top_border_m1+8, src_y-8, xchg);
945
    XCHG(top_border,      src_y,   xchg);
946
    XCHG(top_border+8,    src_y+8, 1);
947
    if (mb_x < mb_width-1)
948
        XCHG(top_border+32, src_y+16, 1);
949

    
950
    // only copy chroma for normal loop filter
951
    // or to initialize the top row to 127
952
    if (!simple || !mb_y) {
953
        XCHG(top_border_m1+16, src_cb-8, xchg);
954
        XCHG(top_border_m1+24, src_cr-8, xchg);
955
        XCHG(top_border+16,    src_cb, 1);
956
        XCHG(top_border+24,    src_cr, 1);
957
    }
958
}
959

    
960
static av_always_inline
961
int check_intra_pred_mode(int mode, int mb_x, int mb_y)
962
{
963
    if (mode == DC_PRED8x8) {
964
        if (!mb_x) {
965
            mode = mb_y ? TOP_DC_PRED8x8 : DC_128_PRED8x8;
966
        } else if (!mb_y) {
967
            mode = LEFT_DC_PRED8x8;
968
        }
969
    }
970
    return mode;
971
}
972

    
973
static av_always_inline
974
void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
975
                   int mb_x, int mb_y)
976
{
977
    int x, y, mode, nnz, tr;
978

    
979
    // for the first row, we need to run xchg_mb_border to init the top edge to 127
980
    // otherwise, skip it if we aren't going to deblock
981
    if (s->deblock_filter || !mb_y)
982
        xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
983
                       s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
984
                       s->filter.simple, 1);
985

    
986
    if (mb->mode < MODE_I4x4) {
987
        mode = check_intra_pred_mode(mb->mode, mb_x, mb_y);
988
        s->hpc.pred16x16[mode](dst[0], s->linesize);
989
    } else {
990
        uint8_t *ptr = dst[0];
991
        uint8_t *intra4x4 = s->intra4x4_pred_mode_mb;
992

    
993
        // all blocks on the right edge of the macroblock use bottom edge
994
        // the top macroblock for their topright edge
995
        uint8_t *tr_right = ptr - s->linesize + 16;
996

    
997
        // if we're on the right edge of the frame, said edge is extended
998
        // from the top macroblock
999
        if (mb_x == s->mb_width-1) {
1000
            tr = tr_right[-1]*0x01010101;
1001
            tr_right = (uint8_t *)&tr;
1002
        }
1003

    
1004
        if (mb->skip)
1005
            AV_ZERO128(s->non_zero_count_cache);
1006

    
1007
        for (y = 0; y < 4; y++) {
1008
            uint8_t *topright = ptr + 4 - s->linesize;
1009
            for (x = 0; x < 4; x++) {
1010
                if (x == 3)
1011
                    topright = tr_right;
1012

    
1013
                s->hpc.pred4x4[intra4x4[x]](ptr+4*x, topright, s->linesize);
1014

    
1015
                nnz = s->non_zero_count_cache[y][x];
1016
                if (nnz) {
1017
                    if (nnz == 1)
1018
                        s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize);
1019
                    else
1020
                        s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize);
1021
                }
1022
                topright += 4;
1023
            }
1024

    
1025
            ptr   += 4*s->linesize;
1026
            intra4x4 += 4;
1027
        }
1028
    }
1029

    
1030
    mode = check_intra_pred_mode(s->chroma_pred_mode, mb_x, mb_y);
1031
    s->hpc.pred8x8[mode](dst[1], s->uvlinesize);
1032
    s->hpc.pred8x8[mode](dst[2], s->uvlinesize);
1033

    
1034
    if (s->deblock_filter || !mb_y)
1035
        xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
1036
                       s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
1037
                       s->filter.simple, 0);
1038
}
1039

    
1040
/**
1041
 * Generic MC function.
1042
 *
1043
 * @param s VP8 decoding context
1044
 * @param luma 1 for luma (Y) planes, 0 for chroma (Cb/Cr) planes
1045
 * @param dst target buffer for block data at block position
1046
 * @param src reference picture buffer at origin (0, 0)
1047
 * @param mv motion vector (relative to block position) to get pixel data from
1048
 * @param x_off horizontal position of block from origin (0, 0)
1049
 * @param y_off vertical position of block from origin (0, 0)
1050
 * @param block_w width of block (16, 8 or 4)
1051
 * @param block_h height of block (always same as block_w)
1052
 * @param width width of src/dst plane data
1053
 * @param height height of src/dst plane data
1054
 * @param linesize size of a single line of plane data, including padding
1055
 * @param mc_func motion compensation function pointers (bilinear or sixtap MC)
1056
 */
1057
static av_always_inline
1058
void vp8_mc(VP8Context *s, int luma,
1059
            uint8_t *dst, uint8_t *src, const VP56mv *mv,
1060
            int x_off, int y_off, int block_w, int block_h,
1061
            int width, int height, int linesize,
1062
            vp8_mc_func mc_func[3][3])
1063
{
1064
    if (AV_RN32A(mv)) {
1065
        static const uint8_t idx[8] = { 0, 1, 2, 1, 2, 1, 2, 1 };
1066
        int mx = (mv->x << luma)&7, mx_idx = idx[mx];
1067
        int my = (mv->y << luma)&7, my_idx = idx[my];
1068

    
1069
        x_off += mv->x >> (3 - luma);
1070
        y_off += mv->y >> (3 - luma);
1071

    
1072
        // edge emulation
1073
        src += y_off * linesize + x_off;
1074
        if (x_off < 2 || x_off >= width  - block_w - 3 ||
1075
            y_off < 2 || y_off >= height - block_h - 3) {
1076
            ff_emulated_edge_mc(s->edge_emu_buffer, src - 2 * linesize - 2, linesize,
1077
                                block_w + 5, block_h + 5,
1078
                                x_off - 2, y_off - 2, width, height);
1079
            src = s->edge_emu_buffer + 2 + linesize * 2;
1080
        }
1081
        mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);
1082
    } else
1083
        mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0);
1084
}
1085

    
1086
static av_always_inline
1087
void vp8_mc_part(VP8Context *s, uint8_t *dst[3],
1088
                 AVFrame *ref_frame, int x_off, int y_off,
1089
                 int bx_off, int by_off,
1090
                 int block_w, int block_h,
1091
                 int width, int height, VP56mv *mv)
1092
{
1093
    VP56mv uvmv = *mv;
1094

    
1095
    /* Y */
1096
    vp8_mc(s, 1, dst[0] + by_off * s->linesize + bx_off,
1097
           ref_frame->data[0], mv, x_off + bx_off, y_off + by_off,
1098
           block_w, block_h, width, height, s->linesize,
1099
           s->put_pixels_tab[block_w == 8]);
1100

    
1101
    /* U/V */
1102
    if (s->profile == 3) {
1103
        uvmv.x &= ~7;
1104
        uvmv.y &= ~7;
1105
    }
1106
    x_off   >>= 1; y_off   >>= 1;
1107
    bx_off  >>= 1; by_off  >>= 1;
1108
    width   >>= 1; height  >>= 1;
1109
    block_w >>= 1; block_h >>= 1;
1110
    vp8_mc(s, 0, dst[1] + by_off * s->uvlinesize + bx_off,
1111
           ref_frame->data[1], &uvmv, x_off + bx_off, y_off + by_off,
1112
           block_w, block_h, width, height, s->uvlinesize,
1113
           s->put_pixels_tab[1 + (block_w == 4)]);
1114
    vp8_mc(s, 0, dst[2] + by_off * s->uvlinesize + bx_off,
1115
           ref_frame->data[2], &uvmv, x_off + bx_off, y_off + by_off,
1116
           block_w, block_h, width, height, s->uvlinesize,
1117
           s->put_pixels_tab[1 + (block_w == 4)]);
1118
}
1119

    
1120
/* Fetch pixels for estimated mv 4 macroblocks ahead.
1121
 * Optimized for 64-byte cache lines.  Inspired by ffh264 prefetch_motion. */
1122
static av_always_inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int mb_xy, int ref)
1123
{
1124
    /* Don't prefetch refs that haven't been used very often this frame. */
1125
    if (s->ref_count[ref-1] > (mb_xy >> 5)) {
1126
        int x_off = mb_x << 4, y_off = mb_y << 4;
1127
        int mx = mb->mv.x + x_off + 8;
1128
        int my = mb->mv.y + y_off;
1129
        uint8_t **src= s->framep[ref]->data;
1130
        int off= mx + (my + (mb_x&3)*4)*s->linesize + 64;
1131
        s->dsp.prefetch(src[0]+off, s->linesize, 4);
1132
        off= (mx>>1) + ((my>>1) + (mb_x&7))*s->uvlinesize + 64;
1133
        s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1134
    }
1135
}
1136

    
1137
/**
1138
 * Apply motion vectors to prediction buffer, chapter 18.
1139
 */
1140
static av_always_inline
1141
void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
1142
                   int mb_x, int mb_y)
1143
{
1144
    int x_off = mb_x << 4, y_off = mb_y << 4;
1145
    int width = 16*s->mb_width, height = 16*s->mb_height;
1146
    AVFrame *ref = s->framep[mb->ref_frame];
1147
    VP56mv *bmv = mb->bmv;
1148

    
1149
    if (mb->mode < VP8_MVMODE_SPLIT) {
1150
        vp8_mc_part(s, dst, ref, x_off, y_off,
1151
                    0, 0, 16, 16, width, height, &mb->mv);
1152
    } else switch (mb->partitioning) {
1153
    case VP8_SPLITMVMODE_4x4: {
1154
        int x, y;
1155
        VP56mv uvmv;
1156

    
1157
        /* Y */
1158
        for (y = 0; y < 4; y++) {
1159
            for (x = 0; x < 4; x++) {
1160
                vp8_mc(s, 1, dst[0] + 4*y*s->linesize + x*4,
1161
                       ref->data[0], &bmv[4*y + x],
1162
                       4*x + x_off, 4*y + y_off, 4, 4,
1163
                       width, height, s->linesize,
1164
                       s->put_pixels_tab[2]);
1165
            }
1166
        }
1167

    
1168
        /* U/V */
1169
        x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1;
1170
        for (y = 0; y < 2; y++) {
1171
            for (x = 0; x < 2; x++) {
1172
                uvmv.x = mb->bmv[ 2*y    * 4 + 2*x  ].x +
1173
                         mb->bmv[ 2*y    * 4 + 2*x+1].x +
1174
                         mb->bmv[(2*y+1) * 4 + 2*x  ].x +
1175
                         mb->bmv[(2*y+1) * 4 + 2*x+1].x;
1176
                uvmv.y = mb->bmv[ 2*y    * 4 + 2*x  ].y +
1177
                         mb->bmv[ 2*y    * 4 + 2*x+1].y +
1178
                         mb->bmv[(2*y+1) * 4 + 2*x  ].y +
1179
                         mb->bmv[(2*y+1) * 4 + 2*x+1].y;
1180
                uvmv.x = (uvmv.x + 2 + (uvmv.x >> (INT_BIT-1))) >> 2;
1181
                uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2;
1182
                if (s->profile == 3) {
1183
                    uvmv.x &= ~7;
1184
                    uvmv.y &= ~7;
1185
                }
1186
                vp8_mc(s, 0, dst[1] + 4*y*s->uvlinesize + x*4,
1187
                       ref->data[1], &uvmv,
1188
                       4*x + x_off, 4*y + y_off, 4, 4,
1189
                       width, height, s->uvlinesize,
1190
                       s->put_pixels_tab[2]);
1191
                vp8_mc(s, 0, dst[2] + 4*y*s->uvlinesize + x*4,
1192
                       ref->data[2], &uvmv,
1193
                       4*x + x_off, 4*y + y_off, 4, 4,
1194
                       width, height, s->uvlinesize,
1195
                       s->put_pixels_tab[2]);
1196
            }
1197
        }
1198
        break;
1199
    }
1200
    case VP8_SPLITMVMODE_16x8:
1201
        vp8_mc_part(s, dst, ref, x_off, y_off,
1202
                    0, 0, 16, 8, width, height, &bmv[0]);
1203
        vp8_mc_part(s, dst, ref, x_off, y_off,
1204
                    0, 8, 16, 8, width, height, &bmv[1]);
1205
        break;
1206
    case VP8_SPLITMVMODE_8x16:
1207
        vp8_mc_part(s, dst, ref, x_off, y_off,
1208
                    0, 0, 8, 16, width, height, &bmv[0]);
1209
        vp8_mc_part(s, dst, ref, x_off, y_off,
1210
                    8, 0, 8, 16, width, height, &bmv[1]);
1211
        break;
1212
    case VP8_SPLITMVMODE_8x8:
1213
        vp8_mc_part(s, dst, ref, x_off, y_off,
1214
                    0, 0, 8, 8, width, height, &bmv[0]);
1215
        vp8_mc_part(s, dst, ref, x_off, y_off,
1216
                    8, 0, 8, 8, width, height, &bmv[1]);
1217
        vp8_mc_part(s, dst, ref, x_off, y_off,
1218
                    0, 8, 8, 8, width, height, &bmv[2]);
1219
        vp8_mc_part(s, dst, ref, x_off, y_off,
1220
                    8, 8, 8, 8, width, height, &bmv[3]);
1221
        break;
1222
    }
1223
}
1224

    
1225
static av_always_inline void idct_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb)
1226
{
1227
    int x, y, ch;
1228

    
1229
    if (mb->mode != MODE_I4x4) {
1230
        uint8_t *y_dst = dst[0];
1231
        for (y = 0; y < 4; y++) {
1232
            uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[y]);
1233
            if (nnz4) {
1234
                if (nnz4&~0x01010101) {
1235
                    for (x = 0; x < 4; x++) {
1236
                        int nnz = s->non_zero_count_cache[y][x];
1237
                        if (nnz) {
1238
                            if (nnz == 1)
1239
                                s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize);
1240
                            else
1241
                                s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize);
1242
                        }
1243
                    }
1244
                } else {
1245
                    s->vp8dsp.vp8_idct_dc_add4y(y_dst, s->block[y], s->linesize);
1246
                }
1247
            }
1248
            y_dst += 4*s->linesize;
1249
        }
1250
    }
1251

    
1252
    for (ch = 0; ch < 2; ch++) {
1253
        uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[4+ch]);
1254
        if (nnz4) {
1255
            uint8_t *ch_dst = dst[1+ch];
1256
            if (nnz4&~0x01010101) {
1257
                for (y = 0; y < 2; y++) {
1258
                    for (x = 0; x < 2; x++) {
1259
                        int nnz = s->non_zero_count_cache[4+ch][(y<<1)+x];
1260
                        if (nnz) {
1261
                            if (nnz == 1)
1262
                                s->vp8dsp.vp8_idct_dc_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);
1263
                            else
1264
                                s->vp8dsp.vp8_idct_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);
1265
                        }
1266
                    }
1267
                    ch_dst += 4*s->uvlinesize;
1268
                }
1269
            } else {
1270
                s->vp8dsp.vp8_idct_dc_add4uv(ch_dst, s->block[4+ch], s->uvlinesize);
1271
            }
1272
        }
1273
    }
1274
}
1275

    
1276
static av_always_inline void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f )
1277
{
1278
    int interior_limit, filter_level;
1279

    
1280
    if (s->segmentation.enabled) {
1281
        filter_level = s->segmentation.filter_level[s->segment];
1282
        if (!s->segmentation.absolute_vals)
1283
            filter_level += s->filter.level;
1284
    } else
1285
        filter_level = s->filter.level;
1286

    
1287
    if (s->lf_delta.enabled) {
1288
        filter_level += s->lf_delta.ref[mb->ref_frame];
1289

    
1290
        if (mb->ref_frame == VP56_FRAME_CURRENT) {
1291
            if (mb->mode == MODE_I4x4)
1292
                filter_level += s->lf_delta.mode[0];
1293
        } else {
1294
            if (mb->mode == VP8_MVMODE_ZERO)
1295
                filter_level += s->lf_delta.mode[1];
1296
            else if (mb->mode == VP8_MVMODE_SPLIT)
1297
                filter_level += s->lf_delta.mode[3];
1298
            else
1299
                filter_level += s->lf_delta.mode[2];
1300
        }
1301
    }
1302
    filter_level = av_clip(filter_level, 0, 63);
1303

    
1304
    interior_limit = filter_level;
1305
    if (s->filter.sharpness) {
1306
        interior_limit >>= s->filter.sharpness > 4 ? 2 : 1;
1307
        interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness);
1308
    }
1309
    interior_limit = FFMAX(interior_limit, 1);
1310

    
1311
    f->filter_level = filter_level;
1312
    f->inner_limit = interior_limit;
1313
    f->inner_filter = !mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT;
1314
}
1315

    
1316
static av_always_inline void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f, int mb_x, int mb_y)
1317
{
1318
    int mbedge_lim, bedge_lim, hev_thresh;
1319
    int filter_level = f->filter_level;
1320
    int inner_limit = f->inner_limit;
1321
    int inner_filter = f->inner_filter;
1322
    int linesize = s->linesize;
1323
    int uvlinesize = s->uvlinesize;
1324

    
1325
    if (!filter_level)
1326
        return;
1327

    
1328
    mbedge_lim = 2*(filter_level+2) + inner_limit;
1329
     bedge_lim = 2* filter_level    + inner_limit;
1330
    hev_thresh = filter_level >= 15;
1331

    
1332
    if (s->keyframe) {
1333
        if (filter_level >= 40)
1334
            hev_thresh = 2;
1335
    } else {
1336
        if (filter_level >= 40)
1337
            hev_thresh = 3;
1338
        else if (filter_level >= 20)
1339
            hev_thresh = 2;
1340
    }
1341

    
1342
    if (mb_x) {
1343
        s->vp8dsp.vp8_h_loop_filter16y(dst[0],     linesize,
1344
                                       mbedge_lim, inner_limit, hev_thresh);
1345
        s->vp8dsp.vp8_h_loop_filter8uv(dst[1],     dst[2],      uvlinesize,
1346
                                       mbedge_lim, inner_limit, hev_thresh);
1347
    }
1348

    
1349
    if (inner_filter) {
1350
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 4, linesize, bedge_lim,
1351
                                             inner_limit, hev_thresh);
1352
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 8, linesize, bedge_lim,
1353
                                             inner_limit, hev_thresh);
1354
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+12, linesize, bedge_lim,
1355
                                             inner_limit, hev_thresh);
1356
        s->vp8dsp.vp8_h_loop_filter8uv_inner(dst[1] + 4, dst[2] + 4,
1357
                                             uvlinesize,  bedge_lim,
1358
                                             inner_limit, hev_thresh);
1359
    }
1360

    
1361
    if (mb_y) {
1362
        s->vp8dsp.vp8_v_loop_filter16y(dst[0],     linesize,
1363
                                       mbedge_lim, inner_limit, hev_thresh);
1364
        s->vp8dsp.vp8_v_loop_filter8uv(dst[1],     dst[2],      uvlinesize,
1365
                                       mbedge_lim, inner_limit, hev_thresh);
1366
    }
1367

    
1368
    if (inner_filter) {
1369
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 4*linesize,
1370
                                             linesize,    bedge_lim,
1371
                                             inner_limit, hev_thresh);
1372
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 8*linesize,
1373
                                             linesize,    bedge_lim,
1374
                                             inner_limit, hev_thresh);
1375
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+12*linesize,
1376
                                             linesize,    bedge_lim,
1377
                                             inner_limit, hev_thresh);
1378
        s->vp8dsp.vp8_v_loop_filter8uv_inner(dst[1] + 4 * uvlinesize,
1379
                                             dst[2] + 4 * uvlinesize,
1380
                                             uvlinesize,  bedge_lim,
1381
                                             inner_limit, hev_thresh);
1382
    }
1383
}
1384

    
1385
static av_always_inline void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f, int mb_x, int mb_y)
1386
{
1387
    int mbedge_lim, bedge_lim;
1388
    int filter_level = f->filter_level;
1389
    int inner_limit = f->inner_limit;
1390
    int inner_filter = f->inner_filter;
1391
    int linesize = s->linesize;
1392

    
1393
    if (!filter_level)
1394
        return;
1395

    
1396
    mbedge_lim = 2*(filter_level+2) + inner_limit;
1397
     bedge_lim = 2* filter_level    + inner_limit;
1398

    
1399
    if (mb_x)
1400
        s->vp8dsp.vp8_h_loop_filter_simple(dst, linesize, mbedge_lim);
1401
    if (inner_filter) {
1402
        s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, linesize, bedge_lim);
1403
        s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, linesize, bedge_lim);
1404
        s->vp8dsp.vp8_h_loop_filter_simple(dst+12, linesize, bedge_lim);
1405
    }
1406

    
1407
    if (mb_y)
1408
        s->vp8dsp.vp8_v_loop_filter_simple(dst, linesize, mbedge_lim);
1409
    if (inner_filter) {
1410
        s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*linesize, linesize, bedge_lim);
1411
        s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*linesize, linesize, bedge_lim);
1412
        s->vp8dsp.vp8_v_loop_filter_simple(dst+12*linesize, linesize, bedge_lim);
1413
    }
1414
}
1415

    
1416
static void filter_mb_row(VP8Context *s, int mb_y)
1417
{
1418
    VP8FilterStrength *f = s->filter_strength;
1419
    uint8_t *dst[3] = {
1420
        s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize,
1421
        s->framep[VP56_FRAME_CURRENT]->data[1] +  8*mb_y*s->uvlinesize,
1422
        s->framep[VP56_FRAME_CURRENT]->data[2] +  8*mb_y*s->uvlinesize
1423
    };
1424
    int mb_x;
1425

    
1426
    for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1427
        backup_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], s->linesize, s->uvlinesize, 0);
1428
        filter_mb(s, dst, f++, mb_x, mb_y);
1429
        dst[0] += 16;
1430
        dst[1] += 8;
1431
        dst[2] += 8;
1432
    }
1433
}
1434

    
1435
static void filter_mb_row_simple(VP8Context *s, int mb_y)
1436
{
1437
    VP8FilterStrength *f = s->filter_strength;
1438
    uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize;
1439
    int mb_x;
1440

    
1441
    for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1442
        backup_mb_border(s->top_border[mb_x+1], dst, NULL, NULL, s->linesize, 0, 1);
1443
        filter_mb_simple(s, dst, f++, mb_x, mb_y);
1444
        dst += 16;
1445
    }
1446
}
1447

    
1448
static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
1449
                            AVPacket *avpkt)
1450
{
1451
    VP8Context *s = avctx->priv_data;
1452
    int ret, mb_x, mb_y, i, y, referenced;
1453
    enum AVDiscard skip_thresh;
1454
    AVFrame *av_uninit(curframe);
1455

    
1456
    if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0)
1457
        return ret;
1458

    
1459
    referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT
1460
                                || s->update_altref == VP56_FRAME_CURRENT;
1461

    
1462
    skip_thresh = !referenced ? AVDISCARD_NONREF :
1463
                    !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL;
1464

    
1465
    if (avctx->skip_frame >= skip_thresh) {
1466
        s->invisible = 1;
1467
        goto skip_decode;
1468
    }
1469
    s->deblock_filter = s->filter.level && avctx->skip_loop_filter < skip_thresh;
1470

    
1471
    for (i = 0; i < 4; i++)
1472
        if (&s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1473
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1474
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) {
1475
            curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i];
1476
            break;
1477
        }
1478
    if (curframe->data[0])
1479
        avctx->release_buffer(avctx, curframe);
1480

    
1481
    curframe->key_frame = s->keyframe;
1482
    curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE;
1483
    curframe->reference = referenced ? 3 : 0;
1484
    if ((ret = avctx->get_buffer(avctx, curframe))) {
1485
        av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n");
1486
        return ret;
1487
    }
1488

    
1489
    // Given that arithmetic probabilities are updated every frame, it's quite likely
1490
    // that the values we have on a random interframe are complete junk if we didn't
1491
    // start decode on a keyframe. So just don't display anything rather than junk.
1492
    if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] ||
1493
                         !s->framep[VP56_FRAME_GOLDEN] ||
1494
                         !s->framep[VP56_FRAME_GOLDEN2])) {
1495
        av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n");
1496
        return AVERROR_INVALIDDATA;
1497
    }
1498

    
1499
    s->linesize   = curframe->linesize[0];
1500
    s->uvlinesize = curframe->linesize[1];
1501

    
1502
    if (!s->edge_emu_buffer)
1503
        s->edge_emu_buffer = av_malloc(21*s->linesize);
1504

    
1505
    memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz));
1506

    
1507
    /* Zero macroblock structures for top/left prediction from outside the frame. */
1508
    memset(s->macroblocks, 0, (s->mb_width + s->mb_height*2)*sizeof(*s->macroblocks));
1509

    
1510
    // top edge of 127 for intra prediction
1511
    memset(s->top_border, 127, (s->mb_width+1)*sizeof(*s->top_border));
1512
    memset(s->ref_count, 0, sizeof(s->ref_count));
1513
    if (s->keyframe)
1514
        memset(s->intra4x4_pred_mode_top, DC_PRED, s->b4_stride*4);
1515

    
1516
    for (mb_y = 0; mb_y < s->mb_height; mb_y++) {
1517
        VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)];
1518
        VP8Macroblock *mb = s->macroblocks + (s->mb_height - mb_y - 1)*2;
1519
        uint8_t *segment_map = s->segmentation_map + mb_y*s->mb_stride;
1520
        int mb_xy = mb_y * s->mb_stride;
1521
        uint8_t *dst[3] = {
1522
            curframe->data[0] + 16*mb_y*s->linesize,
1523
            curframe->data[1] +  8*mb_y*s->uvlinesize,
1524
            curframe->data[2] +  8*mb_y*s->uvlinesize
1525
        };
1526

    
1527
        memset(s->left_nnz, 0, sizeof(s->left_nnz));
1528
        AV_WN32A(s->intra4x4_pred_mode_left, DC_PRED*0x01010101);
1529

    
1530
        // left edge of 129 for intra prediction
1531
        if (!(avctx->flags & CODEC_FLAG_EMU_EDGE))
1532
            for (i = 0; i < 3; i++)
1533
                for (y = 0; y < 16>>!!i; y++)
1534
                    dst[i][y*curframe->linesize[i]-1] = 129;
1535
        if (mb_y)
1536
            memset(s->top_border, 129, sizeof(*s->top_border));
1537

    
1538
        for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb_xy++, mb++) {
1539
            uint8_t *segment_mb = segment_map+mb_x;
1540

    
1541
            /* Prefetch the current frame, 4 MBs ahead */
1542
            s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4);
1543
            s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2);
1544

    
1545
            decode_mb_mode(s, mb, mb_x, mb_y, segment_mb);
1546

    
1547
            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_PREVIOUS);
1548

    
1549
            if (!mb->skip)
1550
                decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);
1551

    
1552
            if (mb->mode <= MODE_I4x4)
1553
                intra_predict(s, dst, mb, mb_x, mb_y);
1554
            else
1555
                inter_predict(s, dst, mb, mb_x, mb_y);
1556

    
1557
            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN);
1558

    
1559
            if (!mb->skip) {
1560
                idct_mb(s, dst, mb);
1561
            } else {
1562
                AV_ZERO64(s->left_nnz);
1563
                AV_WN64(s->top_nnz[mb_x], 0);   // array of 9, so unaligned
1564

    
1565
                // Reset DC block predictors if they would exist if the mb had coefficients
1566
                if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
1567
                    s->left_nnz[8]      = 0;
1568
                    s->top_nnz[mb_x][8] = 0;
1569
                }
1570
            }
1571

    
1572
            if (s->deblock_filter)
1573
                filter_level_for_mb(s, mb, &s->filter_strength[mb_x]);
1574

    
1575
            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN2);
1576

    
1577
            dst[0] += 16;
1578
            dst[1] += 8;
1579
            dst[2] += 8;
1580
        }
1581
        if (s->deblock_filter) {
1582
            if (s->filter.simple)
1583
                filter_mb_row_simple(s, mb_y);
1584
            else
1585
                filter_mb_row(s, mb_y);
1586
        }
1587
    }
1588

    
1589
skip_decode:
1590
    // if future frames don't use the updated probabilities,
1591
    // reset them to the values we saved
1592
    if (!s->update_probabilities)
1593
        s->prob[0] = s->prob[1];
1594

    
1595
    // check if golden and altref are swapped
1596
    if (s->update_altref == VP56_FRAME_GOLDEN &&
1597
        s->update_golden == VP56_FRAME_GOLDEN2)
1598
        FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN], s->framep[VP56_FRAME_GOLDEN2]);
1599
    else {
1600
        if (s->update_altref != VP56_FRAME_NONE)
1601
            s->framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref];
1602

    
1603
        if (s->update_golden != VP56_FRAME_NONE)
1604
            s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden];
1605
    }
1606

    
1607
    if (s->update_last) // move cur->prev
1608
        s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT];
1609

    
1610
    // release no longer referenced frames
1611
    for (i = 0; i < 4; i++)
1612
        if (s->frames[i].data[0] &&
1613
            &s->frames[i] != s->framep[VP56_FRAME_CURRENT] &&
1614
            &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1615
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1616
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2])
1617
            avctx->release_buffer(avctx, &s->frames[i]);
1618

    
1619
    if (!s->invisible) {
1620
        *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT];
1621
        *data_size = sizeof(AVFrame);
1622
    }
1623

    
1624
    return avpkt->size;
1625
}
1626

    
1627
static av_cold int vp8_decode_init(AVCodecContext *avctx)
1628
{
1629
    VP8Context *s = avctx->priv_data;
1630

    
1631
    s->avctx = avctx;
1632
    avctx->pix_fmt = PIX_FMT_YUV420P;
1633

    
1634
    dsputil_init(&s->dsp, avctx);
1635
    ff_h264_pred_init(&s->hpc, CODEC_ID_VP8);
1636
    ff_vp8dsp_init(&s->vp8dsp);
1637

    
1638
    // intra pred needs edge emulation among other things
1639
    if (avctx->flags&CODEC_FLAG_EMU_EDGE) {
1640
        av_log(avctx, AV_LOG_ERROR, "Edge emulation not supported\n");
1641
        return AVERROR_PATCHWELCOME;
1642
    }
1643

    
1644
    return 0;
1645
}
1646

    
1647
static av_cold int vp8_decode_free(AVCodecContext *avctx)
1648
{
1649
    vp8_decode_flush(avctx);
1650
    return 0;
1651
}
1652

    
1653
AVCodec vp8_decoder = {
1654
    "vp8",
1655
    AVMEDIA_TYPE_VIDEO,
1656
    CODEC_ID_VP8,
1657
    sizeof(VP8Context),
1658
    vp8_decode_init,
1659
    NULL,
1660
    vp8_decode_free,
1661
    vp8_decode_frame,
1662
    CODEC_CAP_DR1,
1663
    .flush = vp8_decode_flush,
1664
    .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"),
1665
};