Statistics
| Branch: | Revision:

ffmpeg / libavcodec / vp8.c @ d2840fa4

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
    uint8_t intra4x4_pred_mode_mb[16];
121

    
122
    int chroma_pred_mode;    ///< 8x8c pred mode of the current macroblock
123
    int segment;             ///< segment of the current macroblock
124

    
125
    int mbskip_enabled;
126
    int sign_bias[4]; ///< one state [0, 1] per ref frame type
127
    int ref_count[3];
128

    
129
    /**
130
     * Base parameters for segmentation, i.e. per-macroblock parameters.
131
     * These must be kept unchanged even if segmentation is not used for
132
     * a frame, since the values persist between interframes.
133
     */
134
    struct {
135
        int enabled;
136
        int absolute_vals;
137
        int update_map;
138
        int8_t base_quant[4];
139
        int8_t filter_level[4];     ///< base loop filter level
140
    } segmentation;
141

    
142
    /**
143
     * Macroblocks can have one of 4 different quants in a frame when
144
     * segmentation is enabled.
145
     * If segmentation is disabled, only the first segment's values are used.
146
     */
147
    struct {
148
        // [0] - DC qmul  [1] - AC qmul
149
        int16_t luma_qmul[2];
150
        int16_t luma_dc_qmul[2];    ///< luma dc-only block quant
151
        int16_t chroma_qmul[2];
152
    } qmat[4];
153

    
154
    struct {
155
        int simple;
156
        int level;
157
        int sharpness;
158
    } filter;
159

    
160
    struct {
161
        int enabled;    ///< whether each mb can have a different strength based on mode/ref
162

    
163
        /**
164
         * filter strength adjustment for the following macroblock modes:
165
         * [0] - i4x4
166
         * [1] - zero mv
167
         * [2] - inter modes except for zero or split mv
168
         * [3] - split mv
169
         *  i16x16 modes never have any adjustment
170
         */
171
        int8_t mode[4];
172

    
173
        /**
174
         * filter strength adjustment for macroblocks that reference:
175
         * [0] - intra / VP56_FRAME_CURRENT
176
         * [1] - VP56_FRAME_PREVIOUS
177
         * [2] - VP56_FRAME_GOLDEN
178
         * [3] - altref / VP56_FRAME_GOLDEN2
179
         */
180
        int8_t ref[4];
181
    } lf_delta;
182

    
183
    /**
184
     * These are all of the updatable probabilities for binary decisions.
185
     * They are only implictly reset on keyframes, making it quite likely
186
     * for an interframe to desync if a prior frame's header was corrupt
187
     * or missing outright!
188
     */
189
    struct {
190
        uint8_t segmentid[3];
191
        uint8_t mbskip;
192
        uint8_t intra;
193
        uint8_t last;
194
        uint8_t golden;
195
        uint8_t pred16x16[4];
196
        uint8_t pred8x8c[3];
197
        uint8_t token[4][8][3][NUM_DCT_TOKENS-1];
198
        uint8_t mvc[2][19];
199
    } prob[2];
200
} VP8Context;
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->filter_strength);
214
    av_freep(&s->intra4x4_pred_mode_top);
215
    av_freep(&s->top_nnz);
216
    av_freep(&s->edge_emu_buffer);
217
    av_freep(&s->top_border);
218
    av_freep(&s->segmentation_map);
219

    
220
    s->macroblocks        = NULL;
221
}
222

    
223
static int update_dimensions(VP8Context *s, int width, int height)
224
{
225
    if (avcodec_check_dimensions(s->avctx, width, height))
226
        return AVERROR_INVALIDDATA;
227

    
228
    vp8_decode_flush(s->avctx);
229

    
230
    avcodec_set_dimensions(s->avctx, width, height);
231

    
232
    s->mb_width  = (s->avctx->coded_width +15) / 16;
233
    s->mb_height = (s->avctx->coded_height+15) / 16;
234

    
235
    // we allocate a border around the top/left of intra4x4 modes
236
    // this is 4 blocks for intra4x4 to keep 4-byte alignment for fill_rectangle
237
    s->mb_stride = s->mb_width+1;
238
    s->b4_stride = 4*s->mb_stride;
239

    
240
    s->macroblocks_base        = av_mallocz((s->mb_stride+s->mb_height*2+2)*sizeof(*s->macroblocks));
241
    s->filter_strength         = av_mallocz(s->mb_stride*sizeof(*s->filter_strength));
242
    s->intra4x4_pred_mode_top  = av_mallocz(s->b4_stride*4);
243
    s->top_nnz                 = av_mallocz(s->mb_width*sizeof(*s->top_nnz));
244
    s->top_border              = av_mallocz((s->mb_width+1)*sizeof(*s->top_border));
245
    s->segmentation_map        = av_mallocz(s->mb_stride*s->mb_height);
246

    
247
    if (!s->macroblocks_base || !s->filter_strength || !s->intra4x4_pred_mode_top ||
248
        !s->top_nnz || !s->top_border || !s->segmentation_map)
249
        return AVERROR(ENOMEM);
250

    
251
    s->macroblocks        = s->macroblocks_base + 1;
252

    
253
    return 0;
254
}
255

    
256
static void parse_segment_info(VP8Context *s)
257
{
258
    VP56RangeCoder *c = &s->c;
259
    int i;
260

    
261
    s->segmentation.update_map = vp8_rac_get(c);
262

    
263
    if (vp8_rac_get(c)) { // update segment feature data
264
        s->segmentation.absolute_vals = vp8_rac_get(c);
265

    
266
        for (i = 0; i < 4; i++)
267
            s->segmentation.base_quant[i]   = vp8_rac_get_sint(c, 7);
268

    
269
        for (i = 0; i < 4; i++)
270
            s->segmentation.filter_level[i] = vp8_rac_get_sint(c, 6);
271
    }
272
    if (s->segmentation.update_map)
273
        for (i = 0; i < 3; i++)
274
            s->prob->segmentid[i] = vp8_rac_get(c) ? vp8_rac_get_uint(c, 8) : 255;
275
}
276

    
277
static void update_lf_deltas(VP8Context *s)
278
{
279
    VP56RangeCoder *c = &s->c;
280
    int i;
281

    
282
    for (i = 0; i < 4; i++)
283
        s->lf_delta.ref[i]  = vp8_rac_get_sint(c, 6);
284

    
285
    for (i = 0; i < 4; i++)
286
        s->lf_delta.mode[i] = vp8_rac_get_sint(c, 6);
287
}
288

    
289
static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size)
290
{
291
    const uint8_t *sizes = buf;
292
    int i;
293

    
294
    s->num_coeff_partitions = 1 << vp8_rac_get_uint(&s->c, 2);
295

    
296
    buf      += 3*(s->num_coeff_partitions-1);
297
    buf_size -= 3*(s->num_coeff_partitions-1);
298
    if (buf_size < 0)
299
        return -1;
300

    
301
    for (i = 0; i < s->num_coeff_partitions-1; i++) {
302
        int size = AV_RL24(sizes + 3*i);
303
        if (buf_size - size < 0)
304
            return -1;
305

    
306
        vp56_init_range_decoder(&s->coeff_partition[i], buf, size);
307
        buf      += size;
308
        buf_size -= size;
309
    }
310
    vp56_init_range_decoder(&s->coeff_partition[i], buf, buf_size);
311

    
312
    return 0;
313
}
314

    
315
static void get_quants(VP8Context *s)
316
{
317
    VP56RangeCoder *c = &s->c;
318
    int i, base_qi;
319

    
320
    int yac_qi     = vp8_rac_get_uint(c, 7);
321
    int ydc_delta  = vp8_rac_get_sint(c, 4);
322
    int y2dc_delta = vp8_rac_get_sint(c, 4);
323
    int y2ac_delta = vp8_rac_get_sint(c, 4);
324
    int uvdc_delta = vp8_rac_get_sint(c, 4);
325
    int uvac_delta = vp8_rac_get_sint(c, 4);
326

    
327
    for (i = 0; i < 4; i++) {
328
        if (s->segmentation.enabled) {
329
            base_qi = s->segmentation.base_quant[i];
330
            if (!s->segmentation.absolute_vals)
331
                base_qi += yac_qi;
332
        } else
333
            base_qi = yac_qi;
334

    
335
        s->qmat[i].luma_qmul[0]    =       vp8_dc_qlookup[av_clip(base_qi + ydc_delta , 0, 127)];
336
        s->qmat[i].luma_qmul[1]    =       vp8_ac_qlookup[av_clip(base_qi             , 0, 127)];
337
        s->qmat[i].luma_dc_qmul[0] =   2 * vp8_dc_qlookup[av_clip(base_qi + y2dc_delta, 0, 127)];
338
        s->qmat[i].luma_dc_qmul[1] = 155 * vp8_ac_qlookup[av_clip(base_qi + y2ac_delta, 0, 127)] / 100;
339
        s->qmat[i].chroma_qmul[0]  =       vp8_dc_qlookup[av_clip(base_qi + uvdc_delta, 0, 127)];
340
        s->qmat[i].chroma_qmul[1]  =       vp8_ac_qlookup[av_clip(base_qi + uvac_delta, 0, 127)];
341

    
342
        s->qmat[i].luma_dc_qmul[1] = FFMAX(s->qmat[i].luma_dc_qmul[1], 8);
343
        s->qmat[i].chroma_qmul[0]  = FFMIN(s->qmat[i].chroma_qmul[0], 132);
344
    }
345
}
346

    
347
/**
348
 * Determine which buffers golden and altref should be updated with after this frame.
349
 * The spec isn't clear here, so I'm going by my understanding of what libvpx does
350
 *
351
 * Intra frames update all 3 references
352
 * Inter frames update VP56_FRAME_PREVIOUS if the update_last flag is set
353
 * If the update (golden|altref) flag is set, it's updated with the current frame
354
 *      if update_last is set, and VP56_FRAME_PREVIOUS otherwise.
355
 * If the flag is not set, the number read means:
356
 *      0: no update
357
 *      1: VP56_FRAME_PREVIOUS
358
 *      2: update golden with altref, or update altref with golden
359
 */
360
static VP56Frame ref_to_update(VP8Context *s, int update, VP56Frame ref)
361
{
362
    VP56RangeCoder *c = &s->c;
363

    
364
    if (update)
365
        return VP56_FRAME_CURRENT;
366

    
367
    switch (vp8_rac_get_uint(c, 2)) {
368
    case 1:
369
        return VP56_FRAME_PREVIOUS;
370
    case 2:
371
        return (ref == VP56_FRAME_GOLDEN) ? VP56_FRAME_GOLDEN2 : VP56_FRAME_GOLDEN;
372
    }
373
    return VP56_FRAME_NONE;
374
}
375

    
376
static void update_refs(VP8Context *s)
377
{
378
    VP56RangeCoder *c = &s->c;
379

    
380
    int update_golden = vp8_rac_get(c);
381
    int update_altref = vp8_rac_get(c);
382

    
383
    s->update_golden = ref_to_update(s, update_golden, VP56_FRAME_GOLDEN);
384
    s->update_altref = ref_to_update(s, update_altref, VP56_FRAME_GOLDEN2);
385
}
386

    
387
static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)
388
{
389
    VP56RangeCoder *c = &s->c;
390
    int header_size, hscale, vscale, i, j, k, l, ret;
391
    int width  = s->avctx->width;
392
    int height = s->avctx->height;
393

    
394
    s->keyframe  = !(buf[0] & 1);
395
    s->profile   =  (buf[0]>>1) & 7;
396
    s->invisible = !(buf[0] & 0x10);
397
    header_size  = AV_RL24(buf) >> 5;
398
    buf      += 3;
399
    buf_size -= 3;
400

    
401
    if (s->profile > 3)
402
        av_log(s->avctx, AV_LOG_WARNING, "Unknown profile %d\n", s->profile);
403

    
404
    if (!s->profile)
405
        memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_epel_pixels_tab, sizeof(s->put_pixels_tab));
406
    else    // profile 1-3 use bilinear, 4+ aren't defined so whatever
407
        memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_bilinear_pixels_tab, sizeof(s->put_pixels_tab));
408

    
409
    if (header_size > buf_size - 7*s->keyframe) {
410
        av_log(s->avctx, AV_LOG_ERROR, "Header size larger than data provided\n");
411
        return AVERROR_INVALIDDATA;
412
    }
413

    
414
    if (s->keyframe) {
415
        if (AV_RL24(buf) != 0x2a019d) {
416
            av_log(s->avctx, AV_LOG_ERROR, "Invalid start code 0x%x\n", AV_RL24(buf));
417
            return AVERROR_INVALIDDATA;
418
        }
419
        width  = AV_RL16(buf+3) & 0x3fff;
420
        height = AV_RL16(buf+5) & 0x3fff;
421
        hscale = buf[4] >> 6;
422
        vscale = buf[6] >> 6;
423
        buf      += 7;
424
        buf_size -= 7;
425

    
426
        if (hscale || vscale)
427
            av_log_missing_feature(s->avctx, "Upscaling", 1);
428

    
429
        s->update_golden = s->update_altref = VP56_FRAME_CURRENT;
430
        memcpy(s->prob->token    , vp8_token_default_probs , sizeof(s->prob->token));
431
        memcpy(s->prob->pred16x16, vp8_pred16x16_prob_inter, sizeof(s->prob->pred16x16));
432
        memcpy(s->prob->pred8x8c , vp8_pred8x8c_prob_inter , sizeof(s->prob->pred8x8c));
433
        memcpy(s->prob->mvc      , vp8_mv_default_prob     , sizeof(s->prob->mvc));
434
        memset(&s->segmentation, 0, sizeof(s->segmentation));
435
    }
436

    
437
    if (!s->macroblocks_base || /* first frame */
438
        width != s->avctx->width || height != s->avctx->height) {
439
        if ((ret = update_dimensions(s, width, height) < 0))
440
            return ret;
441
    }
442

    
443
    vp56_init_range_decoder(c, buf, header_size);
444
    buf      += header_size;
445
    buf_size -= header_size;
446

    
447
    if (s->keyframe) {
448
        if (vp8_rac_get(c))
449
            av_log(s->avctx, AV_LOG_WARNING, "Unspecified colorspace\n");
450
        vp8_rac_get(c); // whether we can skip clamping in dsp functions
451
    }
452

    
453
    if ((s->segmentation.enabled = vp8_rac_get(c)))
454
        parse_segment_info(s);
455
    else
456
        s->segmentation.update_map = 0; // FIXME: move this to some init function?
457

    
458
    s->filter.simple    = vp8_rac_get(c);
459
    s->filter.level     = vp8_rac_get_uint(c, 6);
460
    s->filter.sharpness = vp8_rac_get_uint(c, 3);
461

    
462
    if ((s->lf_delta.enabled = vp8_rac_get(c)))
463
        if (vp8_rac_get(c))
464
            update_lf_deltas(s);
465

    
466
    if (setup_partitions(s, buf, buf_size)) {
467
        av_log(s->avctx, AV_LOG_ERROR, "Invalid partitions\n");
468
        return AVERROR_INVALIDDATA;
469
    }
470

    
471
    get_quants(s);
472

    
473
    if (!s->keyframe) {
474
        update_refs(s);
475
        s->sign_bias[VP56_FRAME_GOLDEN]               = vp8_rac_get(c);
476
        s->sign_bias[VP56_FRAME_GOLDEN2 /* altref */] = vp8_rac_get(c);
477
    }
478

    
479
    // if we aren't saving this frame's probabilities for future frames,
480
    // make a copy of the current probabilities
481
    if (!(s->update_probabilities = vp8_rac_get(c)))
482
        s->prob[1] = s->prob[0];
483

    
484
    s->update_last = s->keyframe || vp8_rac_get(c);
485

    
486
    for (i = 0; i < 4; i++)
487
        for (j = 0; j < 8; j++)
488
            for (k = 0; k < 3; k++)
489
                for (l = 0; l < NUM_DCT_TOKENS-1; l++)
490
                    if (vp56_rac_get_prob_branchy(c, vp8_token_update_probs[i][j][k][l]))
491
                        s->prob->token[i][j][k][l] = vp8_rac_get_uint(c, 8);
492

    
493
    if ((s->mbskip_enabled = vp8_rac_get(c)))
494
        s->prob->mbskip = vp8_rac_get_uint(c, 8);
495

    
496
    if (!s->keyframe) {
497
        s->prob->intra  = vp8_rac_get_uint(c, 8);
498
        s->prob->last   = vp8_rac_get_uint(c, 8);
499
        s->prob->golden = vp8_rac_get_uint(c, 8);
500

    
501
        if (vp8_rac_get(c))
502
            for (i = 0; i < 4; i++)
503
                s->prob->pred16x16[i] = vp8_rac_get_uint(c, 8);
504
        if (vp8_rac_get(c))
505
            for (i = 0; i < 3; i++)
506
                s->prob->pred8x8c[i]  = vp8_rac_get_uint(c, 8);
507

    
508
        // 17.2 MV probability update
509
        for (i = 0; i < 2; i++)
510
            for (j = 0; j < 19; j++)
511
                if (vp56_rac_get_prob_branchy(c, vp8_mv_update_prob[i][j]))
512
                    s->prob->mvc[i][j] = vp8_rac_get_nn(c);
513
    }
514

    
515
    return 0;
516
}
517

    
518
static av_always_inline
519
void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src, int mb_x, int mb_y)
520
{
521
#define MARGIN (16 << 2)
522
    dst->x = av_clip(src->x, -((mb_x << 6) + MARGIN),
523
                     ((s->mb_width  - 1 - mb_x) << 6) + MARGIN);
524
    dst->y = av_clip(src->y, -((mb_y << 6) + MARGIN),
525
                     ((s->mb_height - 1 - mb_y) << 6) + MARGIN);
526
}
527

    
528
static av_always_inline
529
void find_near_mvs(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
530
                   VP56mv near[2], VP56mv *best, uint8_t cnt[4])
531
{
532
    VP8Macroblock *mb_edge[3] = { mb + 2 /* top */,
533
                                  mb - 1 /* left */,
534
                                  mb + 1 /* top-left */ };
535
    enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT };
536
    VP56mv near_mv[4]  = {{ 0 }};
537
    enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV };
538
    int idx = CNT_ZERO;
539
    int best_idx = CNT_ZERO;
540
    int cur_sign_bias = s->sign_bias[mb->ref_frame];
541
    int *sign_bias = s->sign_bias;
542

    
543
    /* Process MB on top, left and top-left */
544
    #define MV_EDGE_CHECK(n)\
545
    {\
546
        VP8Macroblock *edge = mb_edge[n];\
547
        int edge_ref = edge->ref_frame;\
548
        if (edge_ref != VP56_FRAME_CURRENT) {\
549
            uint32_t mv = AV_RN32A(&edge->mv);\
550
            if (mv) {\
551
                if (cur_sign_bias != sign_bias[edge_ref]) {\
552
                    /* SWAR negate of the values in mv. */\
553
                    mv = ~mv;\
554
                    mv = ((mv&0x7fff7fff) + 0x00010001) ^ (mv&0x80008000);\
555
                }\
556
                if (!n || mv != AV_RN32A(&near_mv[idx]))\
557
                    AV_WN32A(&near_mv[++idx], mv);\
558
                cnt[idx]      += 1 + (n != 2);\
559
            } else\
560
                cnt[CNT_ZERO] += 1 + (n != 2);\
561
        }\
562
    }
563
    MV_EDGE_CHECK(0)
564
    MV_EDGE_CHECK(1)
565
    MV_EDGE_CHECK(2)
566

    
567
    /* If we have three distinct MVs, merge first and last if they're the same */
568
    if (cnt[CNT_SPLITMV] && AV_RN32A(&near_mv[1+EDGE_TOP]) == AV_RN32A(&near_mv[1+EDGE_TOPLEFT]))
569
        cnt[CNT_NEAREST] += 1;
570

    
571
    cnt[CNT_SPLITMV] = ((mb_edge[EDGE_LEFT]->mode   == VP8_MVMODE_SPLIT) +
572
                        (mb_edge[EDGE_TOP]->mode    == VP8_MVMODE_SPLIT)) * 2 +
573
                       (mb_edge[EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT);
574

    
575
    /* Swap near and nearest if necessary */
576
    if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) {
577
        FFSWAP(uint8_t,     cnt[CNT_NEAREST],     cnt[CNT_NEAR]);
578
        FFSWAP( VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]);
579
    }
580

    
581
    /* Choose the best mv out of 0,0 and the nearest mv */
582
    if (cnt[CNT_NEAREST] >= cnt[CNT_ZERO])
583
        best_idx = CNT_NEAREST;
584

    
585
    mb->mv  = near_mv[best_idx];
586
    near[0] = near_mv[CNT_NEAREST];
587
    near[1] = near_mv[CNT_NEAR];
588
}
589

    
590
/**
591
 * Motion vector coding, 17.1.
592
 */
593
static int read_mv_component(VP56RangeCoder *c, const uint8_t *p)
594
{
595
    int bit, x = 0;
596

    
597
    if (vp56_rac_get_prob_branchy(c, p[0])) {
598
        int i;
599

    
600
        for (i = 0; i < 3; i++)
601
            x += vp56_rac_get_prob(c, p[9 + i]) << i;
602
        for (i = 9; i > 3; i--)
603
            x += vp56_rac_get_prob(c, p[9 + i]) << i;
604
        if (!(x & 0xFFF0) || vp56_rac_get_prob(c, p[12]))
605
            x += 8;
606
    } else {
607
        // small_mvtree
608
        const uint8_t *ps = p+2;
609
        bit = vp56_rac_get_prob(c, *ps);
610
        ps += 1 + 3*bit;
611
        x  += 4*bit;
612
        bit = vp56_rac_get_prob(c, *ps);
613
        ps += 1 + bit;
614
        x  += 2*bit;
615
        x  += vp56_rac_get_prob(c, *ps);
616
    }
617

    
618
    return (x && vp56_rac_get_prob(c, p[1])) ? -x : x;
619
}
620

    
621
static av_always_inline
622
const uint8_t *get_submv_prob(uint32_t left, uint32_t top)
623
{
624
    if (left == top)
625
        return vp8_submv_prob[4-!!left];
626
    if (!top)
627
        return vp8_submv_prob[2];
628
    return vp8_submv_prob[1-!!left];
629
}
630

    
631
/**
632
 * Split motion vector prediction, 16.4.
633
 * @returns the number of motion vectors parsed (2, 4 or 16)
634
 */
635
static av_always_inline
636
int decode_splitmvs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb)
637
{
638
    int part_idx = mb->partitioning =
639
        vp8_rac_get_tree(c, vp8_mbsplit_tree, vp8_mbsplit_prob);
640
    int n, num = vp8_mbsplit_count[part_idx];
641
    VP8Macroblock *top_mb  = &mb[2];
642
    VP8Macroblock *left_mb = &mb[-1];
643
    const uint8_t *mbsplits_left = vp8_mbsplits[left_mb->partitioning],
644
                  *mbsplits_top = vp8_mbsplits[top_mb->partitioning],
645
                  *mbsplits_cur = vp8_mbsplits[part_idx],
646
                  *firstidx = vp8_mbfirstidx[part_idx];
647
    VP56mv *top_mv  = top_mb->bmv;
648
    VP56mv *left_mv = left_mb->bmv;
649
    VP56mv *cur_mv  = mb->bmv;
650

    
651
    for (n = 0; n < num; n++) {
652
        int k = firstidx[n];
653
        uint32_t left, above;
654
        const uint8_t *submv_prob;
655

    
656
        if (!(k & 3))
657
            left = AV_RN32A(&left_mv[mbsplits_left[k + 3]]);
658
        else
659
            left  = AV_RN32A(&cur_mv[mbsplits_cur[k - 1]]);
660
        if (k <= 3)
661
            above = AV_RN32A(&top_mv[mbsplits_top[k + 12]]);
662
        else
663
            above = AV_RN32A(&cur_mv[mbsplits_cur[k - 4]]);
664

    
665
        submv_prob = get_submv_prob(left, above);
666

    
667
        switch (vp8_rac_get_tree(c, vp8_submv_ref_tree, submv_prob)) {
668
        case VP8_SUBMVMODE_NEW4X4:
669
            mb->bmv[n].y = mb->mv.y + read_mv_component(c, s->prob->mvc[0]);
670
            mb->bmv[n].x = mb->mv.x + read_mv_component(c, s->prob->mvc[1]);
671
            break;
672
        case VP8_SUBMVMODE_ZERO4X4:
673
            AV_ZERO32(&mb->bmv[n]);
674
            break;
675
        case VP8_SUBMVMODE_LEFT4X4:
676
            AV_WN32A(&mb->bmv[n], left);
677
            break;
678
        case VP8_SUBMVMODE_TOP4X4:
679
            AV_WN32A(&mb->bmv[n], above);
680
            break;
681
        }
682
    }
683

    
684
    return num;
685
}
686

    
687
static av_always_inline
688
void decode_intra4x4_modes(VP8Context *s, VP56RangeCoder *c,
689
                           int mb_x, int keyframe)
690
{
691
    uint8_t *intra4x4 = s->intra4x4_pred_mode_mb;
692
    if (keyframe) {
693
        int x, y;
694
        uint8_t* const top = s->intra4x4_pred_mode_top + 4 * mb_x;
695
        uint8_t* const left = s->intra4x4_pred_mode_left;
696
        for (y = 0; y < 4; y++) {
697
            for (x = 0; x < 4; x++) {
698
                const uint8_t *ctx;
699
                ctx = vp8_pred4x4_prob_intra[top[x]][left[y]];
700
                *intra4x4 = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx);
701
                left[y] = top[x] = *intra4x4;
702
                intra4x4++;
703
            }
704
        }
705
    } else {
706
        int i;
707
        for (i = 0; i < 16; i++)
708
            intra4x4[i] = vp8_rac_get_tree(c, vp8_pred4x4_tree, vp8_pred4x4_prob_inter);
709
    }
710
}
711

    
712
static av_always_inline
713
void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, uint8_t *segment)
714
{
715
    VP56RangeCoder *c = &s->c;
716

    
717
    if (s->segmentation.update_map)
718
        *segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid);
719
    s->segment = *segment;
720

    
721
    mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0;
722

    
723
    if (s->keyframe) {
724
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra);
725

    
726
        if (mb->mode == MODE_I4x4) {
727
            decode_intra4x4_modes(s, c, mb_x, 1);
728
        } else {
729
            const uint32_t modes = vp8_pred4x4_mode[mb->mode] * 0x01010101u;
730
            AV_WN32A(s->intra4x4_pred_mode_top + 4 * mb_x, modes);
731
            AV_WN32A(s->intra4x4_pred_mode_left, modes);
732
        }
733

    
734
        s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra);
735
        mb->ref_frame = VP56_FRAME_CURRENT;
736
    } else if (vp56_rac_get_prob_branchy(c, s->prob->intra)) {
737
        VP56mv near[2], best;
738
        uint8_t cnt[4] = { 0 };
739
        uint8_t p[4];
740

    
741
        // inter MB, 16.2
742
        if (vp56_rac_get_prob_branchy(c, s->prob->last))
743
            mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ?
744
                VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN;
745
        else
746
            mb->ref_frame = VP56_FRAME_PREVIOUS;
747
        s->ref_count[mb->ref_frame-1]++;
748

    
749
        // motion vectors, 16.3
750
        find_near_mvs(s, mb, mb_x, mb_y, near, &best, cnt);
751
        p[0] = vp8_mode_contexts[cnt[0]][0];
752
        p[1] = vp8_mode_contexts[cnt[1]][1];
753
        p[2] = vp8_mode_contexts[cnt[2]][2];
754
        p[3] = vp8_mode_contexts[cnt[3]][3];
755
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_mvinter, p);
756
        switch (mb->mode) {
757
        case VP8_MVMODE_SPLIT:
758
            clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y);
759
            mb->mv = mb->bmv[decode_splitmvs(s, c, mb) - 1];
760
            break;
761
        case VP8_MVMODE_ZERO:
762
            AV_ZERO32(&mb->mv);
763
            break;
764
        case VP8_MVMODE_NEAREST:
765
            clamp_mv(s, &mb->mv, &near[0], mb_x, mb_y);
766
            break;
767
        case VP8_MVMODE_NEAR:
768
            clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y);
769
            break;
770
        case VP8_MVMODE_NEW:
771
            clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y);
772
            mb->mv.y += + read_mv_component(c, s->prob->mvc[0]);
773
            mb->mv.x += + read_mv_component(c, s->prob->mvc[1]);
774
            break;
775
        }
776
        if (mb->mode != VP8_MVMODE_SPLIT) {
777
            mb->partitioning = VP8_SPLITMVMODE_NONE;
778
            mb->bmv[0] = mb->mv;
779
        }
780
    } else {
781
        // intra MB, 16.1
782
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16);
783

    
784
        if (mb->mode == MODE_I4x4)
785
            decode_intra4x4_modes(s, c, mb_x, 0);
786

    
787
        s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c);
788
        mb->ref_frame = VP56_FRAME_CURRENT;
789
        mb->partitioning = VP8_SPLITMVMODE_NONE;
790
        AV_ZERO32(&mb->bmv[0]);
791
    }
792
}
793

    
794
/**
795
 * @param c arithmetic bitstream reader context
796
 * @param block destination for block coefficients
797
 * @param probs probabilities to use when reading trees from the bitstream
798
 * @param i initial coeff index, 0 unless a separate DC block is coded
799
 * @param zero_nhood the initial prediction context for number of surrounding
800
 *                   all-zero blocks (only left/top, so 0-2)
801
 * @param qmul array holding the dc/ac dequant factor at position 0/1
802
 * @return 0 if no coeffs were decoded
803
 *         otherwise, the index of the last coeff decoded plus one
804
 */
805
static int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16],
806
                               uint8_t probs[8][3][NUM_DCT_TOKENS-1],
807
                               int i, int zero_nhood, int16_t qmul[2])
808
{
809
    uint8_t *token_prob = probs[vp8_coeff_band[i]][zero_nhood];
810
    int nonzero = 0;
811
    int coeff;
812

    
813
    do {
814
        if (!vp56_rac_get_prob_branchy(c, token_prob[0]))   // DCT_EOB
815
            return nonzero;
816

    
817
skip_eob:
818
        if (!vp56_rac_get_prob_branchy(c, token_prob[1])) { // DCT_0
819
            if (++i == 16)
820
                return nonzero; // invalid input; blocks should end with EOB
821
            token_prob = probs[vp8_coeff_band[i]][0];
822
            goto skip_eob;
823
        }
824

    
825
        if (!vp56_rac_get_prob_branchy(c, token_prob[2])) { // DCT_1
826
            coeff = 1;
827
            token_prob = probs[vp8_coeff_band[i+1]][1];
828
        } else {
829
            if (!vp56_rac_get_prob_branchy(c, token_prob[3])) { // DCT 2,3,4
830
                coeff = vp56_rac_get_prob(c, token_prob[4]);
831
                if (coeff)
832
                    coeff += vp56_rac_get_prob(c, token_prob[5]);
833
                coeff += 2;
834
            } else {
835
                // DCT_CAT*
836
                if (!vp56_rac_get_prob_branchy(c, token_prob[6])) {
837
                    if (!vp56_rac_get_prob_branchy(c, token_prob[7])) { // DCT_CAT1
838
                        coeff  = 5 + vp56_rac_get_prob(c, vp8_dct_cat1_prob[0]);
839
                    } else {                                    // DCT_CAT2
840
                        coeff  = 7;
841
                        coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[0]) << 1;
842
                        coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[1]);
843
                    }
844
                } else {    // DCT_CAT3 and up
845
                    int a = vp56_rac_get_prob(c, token_prob[8]);
846
                    int b = vp56_rac_get_prob(c, token_prob[9+a]);
847
                    int cat = (a<<1) + b;
848
                    coeff  = 3 + (8<<cat);
849
                    coeff += vp8_rac_get_coeff(c, vp8_dct_cat_prob[cat]);
850
                }
851
            }
852
            token_prob = probs[vp8_coeff_band[i+1]][2];
853
        }
854

    
855
        // todo: full [16] qmat? load into register?
856
        block[zigzag_scan[i]] = (vp8_rac_get(c) ? -coeff : coeff) * qmul[!!i];
857
        nonzero = ++i;
858
    } while (i < 16);
859

    
860
    return nonzero;
861
}
862

    
863
static av_always_inline
864
void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,
865
                      uint8_t t_nnz[9], uint8_t l_nnz[9])
866
{
867
    LOCAL_ALIGNED_16(DCTELEM, dc,[16]);
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
        AV_ZERO128(dc);
874
        AV_ZERO128(dc+8);
875
        nnz_pred = t_nnz[8] + l_nnz[8];
876

    
877
        // decode DC values and do hadamard
878
        nnz = decode_block_coeffs(c, dc, s->prob->token[1], 0, nnz_pred,
879
                                  s->qmat[segment].luma_dc_qmul);
880
        l_nnz[8] = t_nnz[8] = !!nnz;
881
        nnz_total += nnz;
882
        s->vp8dsp.vp8_luma_dc_wht(s->block, dc);
883
        luma_start = 1;
884
        luma_ctx = 0;
885
    }
886

    
887
    // luma blocks
888
    for (y = 0; y < 4; y++)
889
        for (x = 0; x < 4; x++) {
890
            nnz_pred = l_nnz[y] + t_nnz[x];
891
            nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start,
892
                                      nnz_pred, s->qmat[segment].luma_qmul);
893
            // nnz+luma_start may be one more than the actual last index, but we don't care
894
            s->non_zero_count_cache[y][x] = nnz + luma_start;
895
            t_nnz[x] = l_nnz[y] = !!nnz;
896
            nnz_total += nnz;
897
        }
898

    
899
    // chroma blocks
900
    // TODO: what to do about dimensions? 2nd dim for luma is x,
901
    // but for chroma it's (y<<1)|x
902
    for (i = 4; i < 6; i++)
903
        for (y = 0; y < 2; y++)
904
            for (x = 0; x < 2; x++) {
905
                nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x];
906
                nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0,
907
                                          nnz_pred, s->qmat[segment].chroma_qmul);
908
                s->non_zero_count_cache[i][(y<<1)+x] = nnz;
909
                t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz;
910
                nnz_total += nnz;
911
            }
912

    
913
    // if there were no coded coeffs despite the macroblock not being marked skip,
914
    // we MUST not do the inner loop filter and should not do IDCT
915
    // Since skip isn't used for bitstream prediction, just manually set it.
916
    if (!nnz_total)
917
        mb->skip = 1;
918
}
919

    
920
static av_always_inline
921
void backup_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
922
                      int linesize, int uvlinesize, int simple)
923
{
924
    AV_COPY128(top_border, src_y + 15*linesize);
925
    if (!simple) {
926
        AV_COPY64(top_border+16, src_cb + 7*uvlinesize);
927
        AV_COPY64(top_border+24, src_cr + 7*uvlinesize);
928
    }
929
}
930

    
931
static av_always_inline
932
void xchg_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
933
                    int linesize, int uvlinesize, int mb_x, int mb_y, int mb_width,
934
                    int simple, int xchg)
935
{
936
    uint8_t *top_border_m1 = top_border-32;     // for TL prediction
937
    src_y  -=   linesize;
938
    src_cb -= uvlinesize;
939
    src_cr -= uvlinesize;
940

    
941
#define XCHG(a,b,xchg) do {                     \
942
        if (xchg) AV_SWAP64(b,a);               \
943
        else      AV_COPY64(b,a);               \
944
    } while (0)
945

    
946
    XCHG(top_border_m1+8, src_y-8, xchg);
947
    XCHG(top_border,      src_y,   xchg);
948
    XCHG(top_border+8,    src_y+8, 1);
949
    if (mb_x < mb_width-1)
950
        XCHG(top_border+32, src_y+16, 1);
951

    
952
    // only copy chroma for normal loop filter
953
    // or to initialize the top row to 127
954
    if (!simple || !mb_y) {
955
        XCHG(top_border_m1+16, src_cb-8, xchg);
956
        XCHG(top_border_m1+24, src_cr-8, xchg);
957
        XCHG(top_border+16,    src_cb, 1);
958
        XCHG(top_border+24,    src_cr, 1);
959
    }
960
}
961

    
962
static av_always_inline
963
int check_intra_pred_mode(int mode, int mb_x, int mb_y)
964
{
965
    if (mode == DC_PRED8x8) {
966
        if (!mb_x) {
967
            mode = mb_y ? TOP_DC_PRED8x8 : DC_128_PRED8x8;
968
        } else if (!mb_y) {
969
            mode = LEFT_DC_PRED8x8;
970
        }
971
    }
972
    return mode;
973
}
974

    
975
static av_always_inline
976
void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
977
                   int mb_x, int mb_y)
978
{
979
    int x, y, mode, nnz, tr;
980

    
981
    // for the first row, we need to run xchg_mb_border to init the top edge to 127
982
    // otherwise, skip it if we aren't going to deblock
983
    if (s->deblock_filter || !mb_y)
984
        xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
985
                       s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
986
                       s->filter.simple, 1);
987

    
988
    if (mb->mode < MODE_I4x4) {
989
        mode = check_intra_pred_mode(mb->mode, mb_x, mb_y);
990
        s->hpc.pred16x16[mode](dst[0], s->linesize);
991
    } else {
992
        uint8_t *ptr = dst[0];
993
        uint8_t *intra4x4 = s->intra4x4_pred_mode_mb;
994

    
995
        // all blocks on the right edge of the macroblock use bottom edge
996
        // the top macroblock for their topright edge
997
        uint8_t *tr_right = ptr - s->linesize + 16;
998

    
999
        // if we're on the right edge of the frame, said edge is extended
1000
        // from the top macroblock
1001
        if (mb_x == s->mb_width-1) {
1002
            tr = tr_right[-1]*0x01010101;
1003
            tr_right = (uint8_t *)&tr;
1004
        }
1005

    
1006
        if (mb->skip)
1007
            AV_ZERO128(s->non_zero_count_cache);
1008

    
1009
        for (y = 0; y < 4; y++) {
1010
            uint8_t *topright = ptr + 4 - s->linesize;
1011
            for (x = 0; x < 4; x++) {
1012
                if (x == 3)
1013
                    topright = tr_right;
1014

    
1015
                s->hpc.pred4x4[intra4x4[x]](ptr+4*x, topright, s->linesize);
1016

    
1017
                nnz = s->non_zero_count_cache[y][x];
1018
                if (nnz) {
1019
                    if (nnz == 1)
1020
                        s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize);
1021
                    else
1022
                        s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize);
1023
                }
1024
                topright += 4;
1025
            }
1026

    
1027
            ptr   += 4*s->linesize;
1028
            intra4x4 += 4;
1029
        }
1030
    }
1031

    
1032
    mode = check_intra_pred_mode(s->chroma_pred_mode, mb_x, mb_y);
1033
    s->hpc.pred8x8[mode](dst[1], s->uvlinesize);
1034
    s->hpc.pred8x8[mode](dst[2], s->uvlinesize);
1035

    
1036
    if (s->deblock_filter || !mb_y)
1037
        xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
1038
                       s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
1039
                       s->filter.simple, 0);
1040
}
1041

    
1042
/**
1043
 * Generic MC function.
1044
 *
1045
 * @param s VP8 decoding context
1046
 * @param luma 1 for luma (Y) planes, 0 for chroma (Cb/Cr) planes
1047
 * @param dst target buffer for block data at block position
1048
 * @param src reference picture buffer at origin (0, 0)
1049
 * @param mv motion vector (relative to block position) to get pixel data from
1050
 * @param x_off horizontal position of block from origin (0, 0)
1051
 * @param y_off vertical position of block from origin (0, 0)
1052
 * @param block_w width of block (16, 8 or 4)
1053
 * @param block_h height of block (always same as block_w)
1054
 * @param width width of src/dst plane data
1055
 * @param height height of src/dst plane data
1056
 * @param linesize size of a single line of plane data, including padding
1057
 * @param mc_func motion compensation function pointers (bilinear or sixtap MC)
1058
 */
1059
static av_always_inline
1060
void vp8_mc(VP8Context *s, int luma,
1061
            uint8_t *dst, uint8_t *src, const VP56mv *mv,
1062
            int x_off, int y_off, int block_w, int block_h,
1063
            int width, int height, int linesize,
1064
            vp8_mc_func mc_func[3][3])
1065
{
1066
    if (AV_RN32A(mv)) {
1067
        static const uint8_t idx[8] = { 0, 1, 2, 1, 2, 1, 2, 1 };
1068
        int mx = (mv->x << luma)&7, mx_idx = idx[mx];
1069
        int my = (mv->y << luma)&7, my_idx = idx[my];
1070

    
1071
        x_off += mv->x >> (3 - luma);
1072
        y_off += mv->y >> (3 - luma);
1073

    
1074
        // edge emulation
1075
        src += y_off * linesize + x_off;
1076
        if (x_off < 2 || x_off >= width  - block_w - 3 ||
1077
            y_off < 2 || y_off >= height - block_h - 3) {
1078
            ff_emulated_edge_mc(s->edge_emu_buffer, src - 2 * linesize - 2, linesize,
1079
                                block_w + 5, block_h + 5,
1080
                                x_off - 2, y_off - 2, width, height);
1081
            src = s->edge_emu_buffer + 2 + linesize * 2;
1082
        }
1083
        mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);
1084
    } else
1085
        mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0);
1086
}
1087

    
1088
static av_always_inline
1089
void vp8_mc_part(VP8Context *s, uint8_t *dst[3],
1090
                 AVFrame *ref_frame, int x_off, int y_off,
1091
                 int bx_off, int by_off,
1092
                 int block_w, int block_h,
1093
                 int width, int height, VP56mv *mv)
1094
{
1095
    VP56mv uvmv = *mv;
1096

    
1097
    /* Y */
1098
    vp8_mc(s, 1, dst[0] + by_off * s->linesize + bx_off,
1099
           ref_frame->data[0], mv, x_off + bx_off, y_off + by_off,
1100
           block_w, block_h, width, height, s->linesize,
1101
           s->put_pixels_tab[block_w == 8]);
1102

    
1103
    /* U/V */
1104
    if (s->profile == 3) {
1105
        uvmv.x &= ~7;
1106
        uvmv.y &= ~7;
1107
    }
1108
    x_off   >>= 1; y_off   >>= 1;
1109
    bx_off  >>= 1; by_off  >>= 1;
1110
    width   >>= 1; height  >>= 1;
1111
    block_w >>= 1; block_h >>= 1;
1112
    vp8_mc(s, 0, dst[1] + by_off * s->uvlinesize + bx_off,
1113
           ref_frame->data[1], &uvmv, x_off + bx_off, y_off + by_off,
1114
           block_w, block_h, width, height, s->uvlinesize,
1115
           s->put_pixels_tab[1 + (block_w == 4)]);
1116
    vp8_mc(s, 0, dst[2] + by_off * s->uvlinesize + bx_off,
1117
           ref_frame->data[2], &uvmv, x_off + bx_off, y_off + by_off,
1118
           block_w, block_h, width, height, s->uvlinesize,
1119
           s->put_pixels_tab[1 + (block_w == 4)]);
1120
}
1121

    
1122
/* Fetch pixels for estimated mv 4 macroblocks ahead.
1123
 * Optimized for 64-byte cache lines.  Inspired by ffh264 prefetch_motion. */
1124
static av_always_inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int mb_xy, int ref)
1125
{
1126
    /* Don't prefetch refs that haven't been used very often this frame. */
1127
    if (s->ref_count[ref-1] > (mb_xy >> 5)) {
1128
        int x_off = mb_x << 4, y_off = mb_y << 4;
1129
        int mx = mb->mv.x + x_off + 8;
1130
        int my = mb->mv.y + y_off;
1131
        uint8_t **src= s->framep[ref]->data;
1132
        int off= mx + (my + (mb_x&3)*4)*s->linesize + 64;
1133
        s->dsp.prefetch(src[0]+off, s->linesize, 4);
1134
        off= (mx>>1) + ((my>>1) + (mb_x&7))*s->uvlinesize + 64;
1135
        s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1136
    }
1137
}
1138

    
1139
/**
1140
 * Apply motion vectors to prediction buffer, chapter 18.
1141
 */
1142
static av_always_inline
1143
void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
1144
                   int mb_x, int mb_y)
1145
{
1146
    int x_off = mb_x << 4, y_off = mb_y << 4;
1147
    int width = 16*s->mb_width, height = 16*s->mb_height;
1148
    AVFrame *ref = s->framep[mb->ref_frame];
1149
    VP56mv *bmv = mb->bmv;
1150

    
1151
    if (mb->mode < VP8_MVMODE_SPLIT) {
1152
        vp8_mc_part(s, dst, ref, x_off, y_off,
1153
                    0, 0, 16, 16, width, height, &mb->mv);
1154
    } else switch (mb->partitioning) {
1155
    case VP8_SPLITMVMODE_4x4: {
1156
        int x, y;
1157
        VP56mv uvmv;
1158

    
1159
        /* Y */
1160
        for (y = 0; y < 4; y++) {
1161
            for (x = 0; x < 4; x++) {
1162
                vp8_mc(s, 1, dst[0] + 4*y*s->linesize + x*4,
1163
                       ref->data[0], &bmv[4*y + x],
1164
                       4*x + x_off, 4*y + y_off, 4, 4,
1165
                       width, height, s->linesize,
1166
                       s->put_pixels_tab[2]);
1167
            }
1168
        }
1169

    
1170
        /* U/V */
1171
        x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1;
1172
        for (y = 0; y < 2; y++) {
1173
            for (x = 0; x < 2; x++) {
1174
                uvmv.x = mb->bmv[ 2*y    * 4 + 2*x  ].x +
1175
                         mb->bmv[ 2*y    * 4 + 2*x+1].x +
1176
                         mb->bmv[(2*y+1) * 4 + 2*x  ].x +
1177
                         mb->bmv[(2*y+1) * 4 + 2*x+1].x;
1178
                uvmv.y = mb->bmv[ 2*y    * 4 + 2*x  ].y +
1179
                         mb->bmv[ 2*y    * 4 + 2*x+1].y +
1180
                         mb->bmv[(2*y+1) * 4 + 2*x  ].y +
1181
                         mb->bmv[(2*y+1) * 4 + 2*x+1].y;
1182
                uvmv.x = (uvmv.x + 2 + (uvmv.x >> (INT_BIT-1))) >> 2;
1183
                uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2;
1184
                if (s->profile == 3) {
1185
                    uvmv.x &= ~7;
1186
                    uvmv.y &= ~7;
1187
                }
1188
                vp8_mc(s, 0, dst[1] + 4*y*s->uvlinesize + x*4,
1189
                       ref->data[1], &uvmv,
1190
                       4*x + x_off, 4*y + y_off, 4, 4,
1191
                       width, height, s->uvlinesize,
1192
                       s->put_pixels_tab[2]);
1193
                vp8_mc(s, 0, dst[2] + 4*y*s->uvlinesize + x*4,
1194
                       ref->data[2], &uvmv,
1195
                       4*x + x_off, 4*y + y_off, 4, 4,
1196
                       width, height, s->uvlinesize,
1197
                       s->put_pixels_tab[2]);
1198
            }
1199
        }
1200
        break;
1201
    }
1202
    case VP8_SPLITMVMODE_16x8:
1203
        vp8_mc_part(s, dst, ref, x_off, y_off,
1204
                    0, 0, 16, 8, width, height, &bmv[0]);
1205
        vp8_mc_part(s, dst, ref, x_off, y_off,
1206
                    0, 8, 16, 8, width, height, &bmv[1]);
1207
        break;
1208
    case VP8_SPLITMVMODE_8x16:
1209
        vp8_mc_part(s, dst, ref, x_off, y_off,
1210
                    0, 0, 8, 16, width, height, &bmv[0]);
1211
        vp8_mc_part(s, dst, ref, x_off, y_off,
1212
                    8, 0, 8, 16, width, height, &bmv[1]);
1213
        break;
1214
    case VP8_SPLITMVMODE_8x8:
1215
        vp8_mc_part(s, dst, ref, x_off, y_off,
1216
                    0, 0, 8, 8, width, height, &bmv[0]);
1217
        vp8_mc_part(s, dst, ref, x_off, y_off,
1218
                    8, 0, 8, 8, width, height, &bmv[1]);
1219
        vp8_mc_part(s, dst, ref, x_off, y_off,
1220
                    0, 8, 8, 8, width, height, &bmv[2]);
1221
        vp8_mc_part(s, dst, ref, x_off, y_off,
1222
                    8, 8, 8, 8, width, height, &bmv[3]);
1223
        break;
1224
    }
1225
}
1226

    
1227
static av_always_inline void idct_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb)
1228
{
1229
    int x, y, ch;
1230

    
1231
    if (mb->mode != MODE_I4x4) {
1232
        uint8_t *y_dst = dst[0];
1233
        for (y = 0; y < 4; y++) {
1234
            uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[y]);
1235
            if (nnz4) {
1236
                if (nnz4&~0x01010101) {
1237
                    for (x = 0; x < 4; x++) {
1238
                        int nnz = s->non_zero_count_cache[y][x];
1239
                        if (nnz) {
1240
                            if (nnz == 1)
1241
                                s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize);
1242
                            else
1243
                                s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize);
1244
                        }
1245
                    }
1246
                } else {
1247
                    s->vp8dsp.vp8_idct_dc_add4y(y_dst, s->block[y], s->linesize);
1248
                }
1249
            }
1250
            y_dst += 4*s->linesize;
1251
        }
1252
    }
1253

    
1254
    for (ch = 0; ch < 2; ch++) {
1255
        uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[4+ch]);
1256
        if (nnz4) {
1257
            uint8_t *ch_dst = dst[1+ch];
1258
            if (nnz4&~0x01010101) {
1259
                for (y = 0; y < 2; y++) {
1260
                    for (x = 0; x < 2; x++) {
1261
                        int nnz = s->non_zero_count_cache[4+ch][(y<<1)+x];
1262
                        if (nnz) {
1263
                            if (nnz == 1)
1264
                                s->vp8dsp.vp8_idct_dc_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);
1265
                            else
1266
                                s->vp8dsp.vp8_idct_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);
1267
                        }
1268
                    }
1269
                    ch_dst += 4*s->uvlinesize;
1270
                }
1271
            } else {
1272
                s->vp8dsp.vp8_idct_dc_add4uv(ch_dst, s->block[4+ch], s->uvlinesize);
1273
            }
1274
        }
1275
    }
1276
}
1277

    
1278
static av_always_inline void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f )
1279
{
1280
    int interior_limit, filter_level;
1281

    
1282
    if (s->segmentation.enabled) {
1283
        filter_level = s->segmentation.filter_level[s->segment];
1284
        if (!s->segmentation.absolute_vals)
1285
            filter_level += s->filter.level;
1286
    } else
1287
        filter_level = s->filter.level;
1288

    
1289
    if (s->lf_delta.enabled) {
1290
        filter_level += s->lf_delta.ref[mb->ref_frame];
1291

    
1292
        if (mb->ref_frame == VP56_FRAME_CURRENT) {
1293
            if (mb->mode == MODE_I4x4)
1294
                filter_level += s->lf_delta.mode[0];
1295
        } else {
1296
            if (mb->mode == VP8_MVMODE_ZERO)
1297
                filter_level += s->lf_delta.mode[1];
1298
            else if (mb->mode == VP8_MVMODE_SPLIT)
1299
                filter_level += s->lf_delta.mode[3];
1300
            else
1301
                filter_level += s->lf_delta.mode[2];
1302
        }
1303
    }
1304
    filter_level = av_clip(filter_level, 0, 63);
1305

    
1306
    interior_limit = filter_level;
1307
    if (s->filter.sharpness) {
1308
        interior_limit >>= s->filter.sharpness > 4 ? 2 : 1;
1309
        interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness);
1310
    }
1311
    interior_limit = FFMAX(interior_limit, 1);
1312

    
1313
    f->filter_level = filter_level;
1314
    f->inner_limit = interior_limit;
1315
    f->inner_filter = !mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT;
1316
}
1317

    
1318
static av_always_inline void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f, int mb_x, int mb_y)
1319
{
1320
    int mbedge_lim, bedge_lim, hev_thresh;
1321
    int filter_level = f->filter_level;
1322
    int inner_limit = f->inner_limit;
1323
    int inner_filter = f->inner_filter;
1324
    int linesize = s->linesize;
1325
    int uvlinesize = s->uvlinesize;
1326

    
1327
    if (!filter_level)
1328
        return;
1329

    
1330
    mbedge_lim = 2*(filter_level+2) + inner_limit;
1331
     bedge_lim = 2* filter_level    + inner_limit;
1332
    hev_thresh = filter_level >= 15;
1333

    
1334
    if (s->keyframe) {
1335
        if (filter_level >= 40)
1336
            hev_thresh = 2;
1337
    } else {
1338
        if (filter_level >= 40)
1339
            hev_thresh = 3;
1340
        else if (filter_level >= 20)
1341
            hev_thresh = 2;
1342
    }
1343

    
1344
    if (mb_x) {
1345
        s->vp8dsp.vp8_h_loop_filter16y(dst[0],     linesize,
1346
                                       mbedge_lim, inner_limit, hev_thresh);
1347
        s->vp8dsp.vp8_h_loop_filter8uv(dst[1],     dst[2],      uvlinesize,
1348
                                       mbedge_lim, inner_limit, hev_thresh);
1349
    }
1350

    
1351
    if (inner_filter) {
1352
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 4, linesize, bedge_lim,
1353
                                             inner_limit, hev_thresh);
1354
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 8, linesize, bedge_lim,
1355
                                             inner_limit, hev_thresh);
1356
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+12, linesize, bedge_lim,
1357
                                             inner_limit, hev_thresh);
1358
        s->vp8dsp.vp8_h_loop_filter8uv_inner(dst[1] + 4, dst[2] + 4,
1359
                                             uvlinesize,  bedge_lim,
1360
                                             inner_limit, hev_thresh);
1361
    }
1362

    
1363
    if (mb_y) {
1364
        s->vp8dsp.vp8_v_loop_filter16y(dst[0],     linesize,
1365
                                       mbedge_lim, inner_limit, hev_thresh);
1366
        s->vp8dsp.vp8_v_loop_filter8uv(dst[1],     dst[2],      uvlinesize,
1367
                                       mbedge_lim, inner_limit, hev_thresh);
1368
    }
1369

    
1370
    if (inner_filter) {
1371
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 4*linesize,
1372
                                             linesize,    bedge_lim,
1373
                                             inner_limit, hev_thresh);
1374
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 8*linesize,
1375
                                             linesize,    bedge_lim,
1376
                                             inner_limit, hev_thresh);
1377
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+12*linesize,
1378
                                             linesize,    bedge_lim,
1379
                                             inner_limit, hev_thresh);
1380
        s->vp8dsp.vp8_v_loop_filter8uv_inner(dst[1] + 4 * uvlinesize,
1381
                                             dst[2] + 4 * uvlinesize,
1382
                                             uvlinesize,  bedge_lim,
1383
                                             inner_limit, hev_thresh);
1384
    }
1385
}
1386

    
1387
static av_always_inline void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f, int mb_x, int mb_y)
1388
{
1389
    int mbedge_lim, bedge_lim;
1390
    int filter_level = f->filter_level;
1391
    int inner_limit = f->inner_limit;
1392
    int inner_filter = f->inner_filter;
1393
    int linesize = s->linesize;
1394

    
1395
    if (!filter_level)
1396
        return;
1397

    
1398
    mbedge_lim = 2*(filter_level+2) + inner_limit;
1399
     bedge_lim = 2* filter_level    + inner_limit;
1400

    
1401
    if (mb_x)
1402
        s->vp8dsp.vp8_h_loop_filter_simple(dst, linesize, mbedge_lim);
1403
    if (inner_filter) {
1404
        s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, linesize, bedge_lim);
1405
        s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, linesize, bedge_lim);
1406
        s->vp8dsp.vp8_h_loop_filter_simple(dst+12, linesize, bedge_lim);
1407
    }
1408

    
1409
    if (mb_y)
1410
        s->vp8dsp.vp8_v_loop_filter_simple(dst, linesize, mbedge_lim);
1411
    if (inner_filter) {
1412
        s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*linesize, linesize, bedge_lim);
1413
        s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*linesize, linesize, bedge_lim);
1414
        s->vp8dsp.vp8_v_loop_filter_simple(dst+12*linesize, linesize, bedge_lim);
1415
    }
1416
}
1417

    
1418
static void filter_mb_row(VP8Context *s, int mb_y)
1419
{
1420
    VP8FilterStrength *f = s->filter_strength;
1421
    uint8_t *dst[3] = {
1422
        s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize,
1423
        s->framep[VP56_FRAME_CURRENT]->data[1] +  8*mb_y*s->uvlinesize,
1424
        s->framep[VP56_FRAME_CURRENT]->data[2] +  8*mb_y*s->uvlinesize
1425
    };
1426
    int mb_x;
1427

    
1428
    for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1429
        backup_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], s->linesize, s->uvlinesize, 0);
1430
        filter_mb(s, dst, f++, mb_x, mb_y);
1431
        dst[0] += 16;
1432
        dst[1] += 8;
1433
        dst[2] += 8;
1434
    }
1435
}
1436

    
1437
static void filter_mb_row_simple(VP8Context *s, int mb_y)
1438
{
1439
    VP8FilterStrength *f = s->filter_strength;
1440
    uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize;
1441
    int mb_x;
1442

    
1443
    for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1444
        backup_mb_border(s->top_border[mb_x+1], dst, NULL, NULL, s->linesize, 0, 1);
1445
        filter_mb_simple(s, dst, f++, mb_x, mb_y);
1446
        dst += 16;
1447
    }
1448
}
1449

    
1450
static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
1451
                            AVPacket *avpkt)
1452
{
1453
    VP8Context *s = avctx->priv_data;
1454
    int ret, mb_x, mb_y, i, y, referenced;
1455
    enum AVDiscard skip_thresh;
1456
    AVFrame *av_uninit(curframe);
1457

    
1458
    if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0)
1459
        return ret;
1460

    
1461
    referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT
1462
                                || s->update_altref == VP56_FRAME_CURRENT;
1463

    
1464
    skip_thresh = !referenced ? AVDISCARD_NONREF :
1465
                    !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL;
1466

    
1467
    if (avctx->skip_frame >= skip_thresh) {
1468
        s->invisible = 1;
1469
        goto skip_decode;
1470
    }
1471
    s->deblock_filter = s->filter.level && avctx->skip_loop_filter < skip_thresh;
1472

    
1473
    for (i = 0; i < 4; i++)
1474
        if (&s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1475
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1476
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) {
1477
            curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i];
1478
            break;
1479
        }
1480
    if (curframe->data[0])
1481
        avctx->release_buffer(avctx, curframe);
1482

    
1483
    curframe->key_frame = s->keyframe;
1484
    curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE;
1485
    curframe->reference = referenced ? 3 : 0;
1486
    if ((ret = avctx->get_buffer(avctx, curframe))) {
1487
        av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n");
1488
        return ret;
1489
    }
1490

    
1491
    // Given that arithmetic probabilities are updated every frame, it's quite likely
1492
    // that the values we have on a random interframe are complete junk if we didn't
1493
    // start decode on a keyframe. So just don't display anything rather than junk.
1494
    if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] ||
1495
                         !s->framep[VP56_FRAME_GOLDEN] ||
1496
                         !s->framep[VP56_FRAME_GOLDEN2])) {
1497
        av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n");
1498
        return AVERROR_INVALIDDATA;
1499
    }
1500

    
1501
    s->linesize   = curframe->linesize[0];
1502
    s->uvlinesize = curframe->linesize[1];
1503

    
1504
    if (!s->edge_emu_buffer)
1505
        s->edge_emu_buffer = av_malloc(21*s->linesize);
1506

    
1507
    memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz));
1508

    
1509
    /* Zero macroblock structures for top/left prediction from outside the frame. */
1510
    memset(s->macroblocks, 0, (s->mb_width + s->mb_height*2)*sizeof(*s->macroblocks));
1511

    
1512
    // top edge of 127 for intra prediction
1513
    memset(s->top_border, 127, (s->mb_width+1)*sizeof(*s->top_border));
1514
    memset(s->ref_count, 0, sizeof(s->ref_count));
1515
    if (s->keyframe)
1516
        memset(s->intra4x4_pred_mode_top, DC_PRED, s->b4_stride*4);
1517

    
1518
    for (mb_y = 0; mb_y < s->mb_height; mb_y++) {
1519
        VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)];
1520
        VP8Macroblock *mb = s->macroblocks + (s->mb_height - mb_y - 1)*2;
1521
        uint8_t *segment_map = s->segmentation_map + mb_y*s->mb_stride;
1522
        int mb_xy = mb_y * s->mb_stride;
1523
        uint8_t *dst[3] = {
1524
            curframe->data[0] + 16*mb_y*s->linesize,
1525
            curframe->data[1] +  8*mb_y*s->uvlinesize,
1526
            curframe->data[2] +  8*mb_y*s->uvlinesize
1527
        };
1528

    
1529
        memset(s->left_nnz, 0, sizeof(s->left_nnz));
1530
        AV_WN32A(s->intra4x4_pred_mode_left, DC_PRED*0x01010101);
1531

    
1532
        // left edge of 129 for intra prediction
1533
        if (!(avctx->flags & CODEC_FLAG_EMU_EDGE))
1534
            for (i = 0; i < 3; i++)
1535
                for (y = 0; y < 16>>!!i; y++)
1536
                    dst[i][y*curframe->linesize[i]-1] = 129;
1537
        if (mb_y)
1538
            memset(s->top_border, 129, sizeof(*s->top_border));
1539

    
1540
        for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb_xy++, mb++) {
1541
            uint8_t *segment_mb = segment_map+mb_x;
1542

    
1543
            /* Prefetch the current frame, 4 MBs ahead */
1544
            s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4);
1545
            s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2);
1546

    
1547
            decode_mb_mode(s, mb, mb_x, mb_y, segment_mb);
1548

    
1549
            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_PREVIOUS);
1550

    
1551
            if (!mb->skip)
1552
                decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);
1553

    
1554
            if (mb->mode <= MODE_I4x4)
1555
                intra_predict(s, dst, mb, mb_x, mb_y);
1556
            else
1557
                inter_predict(s, dst, mb, mb_x, mb_y);
1558

    
1559
            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN);
1560

    
1561
            if (!mb->skip) {
1562
                idct_mb(s, dst, mb);
1563
            } else {
1564
                AV_ZERO64(s->left_nnz);
1565
                AV_WN64(s->top_nnz[mb_x], 0);   // array of 9, so unaligned
1566

    
1567
                // Reset DC block predictors if they would exist if the mb had coefficients
1568
                if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
1569
                    s->left_nnz[8]      = 0;
1570
                    s->top_nnz[mb_x][8] = 0;
1571
                }
1572
            }
1573

    
1574
            if (s->deblock_filter)
1575
                filter_level_for_mb(s, mb, &s->filter_strength[mb_x]);
1576

    
1577
            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN2);
1578

    
1579
            dst[0] += 16;
1580
            dst[1] += 8;
1581
            dst[2] += 8;
1582
        }
1583
        if (s->deblock_filter) {
1584
            if (s->filter.simple)
1585
                filter_mb_row_simple(s, mb_y);
1586
            else
1587
                filter_mb_row(s, mb_y);
1588
        }
1589
    }
1590

    
1591
skip_decode:
1592
    // if future frames don't use the updated probabilities,
1593
    // reset them to the values we saved
1594
    if (!s->update_probabilities)
1595
        s->prob[0] = s->prob[1];
1596

    
1597
    // check if golden and altref are swapped
1598
    if (s->update_altref == VP56_FRAME_GOLDEN &&
1599
        s->update_golden == VP56_FRAME_GOLDEN2)
1600
        FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN], s->framep[VP56_FRAME_GOLDEN2]);
1601
    else {
1602
        if (s->update_altref != VP56_FRAME_NONE)
1603
            s->framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref];
1604

    
1605
        if (s->update_golden != VP56_FRAME_NONE)
1606
            s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden];
1607
    }
1608

    
1609
    if (s->update_last) // move cur->prev
1610
        s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT];
1611

    
1612
    // release no longer referenced frames
1613
    for (i = 0; i < 4; i++)
1614
        if (s->frames[i].data[0] &&
1615
            &s->frames[i] != s->framep[VP56_FRAME_CURRENT] &&
1616
            &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1617
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1618
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2])
1619
            avctx->release_buffer(avctx, &s->frames[i]);
1620

    
1621
    if (!s->invisible) {
1622
        *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT];
1623
        *data_size = sizeof(AVFrame);
1624
    }
1625

    
1626
    return avpkt->size;
1627
}
1628

    
1629
static av_cold int vp8_decode_init(AVCodecContext *avctx)
1630
{
1631
    VP8Context *s = avctx->priv_data;
1632

    
1633
    s->avctx = avctx;
1634
    avctx->pix_fmt = PIX_FMT_YUV420P;
1635

    
1636
    dsputil_init(&s->dsp, avctx);
1637
    ff_h264_pred_init(&s->hpc, CODEC_ID_VP8);
1638
    ff_vp8dsp_init(&s->vp8dsp);
1639

    
1640
    // intra pred needs edge emulation among other things
1641
    if (avctx->flags&CODEC_FLAG_EMU_EDGE) {
1642
        av_log(avctx, AV_LOG_ERROR, "Edge emulation not supported\n");
1643
        return AVERROR_PATCHWELCOME;
1644
    }
1645

    
1646
    return 0;
1647
}
1648

    
1649
static av_cold int vp8_decode_free(AVCodecContext *avctx)
1650
{
1651
    vp8_decode_flush(avctx);
1652
    return 0;
1653
}
1654

    
1655
AVCodec vp8_decoder = {
1656
    "vp8",
1657
    AVMEDIA_TYPE_VIDEO,
1658
    CODEC_ID_VP8,
1659
    sizeof(VP8Context),
1660
    vp8_decode_init,
1661
    NULL,
1662
    vp8_decode_free,
1663
    vp8_decode_frame,
1664
    CODEC_CAP_DR1,
1665
    .flush = vp8_decode_flush,
1666
    .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"),
1667
};