Statistics
| Branch: | Revision:

ffmpeg / libavcodec / vp8.c @ e7e81959

History | View | Annotate | Download (58.2 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;
92
    uint8_t *intra4x4_pred_mode_base;
93
    uint8_t *segmentation_map;
94
    int b4_stride;
95

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

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

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

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

    
125
    int mbskip_enabled;
126
    uint8_t mbskip_proba;
127
    uint8_t intra_proba;
128
    uint8_t last_proba;
129
    uint8_t golden_proba;
130
    int sign_bias[4]; ///< one state [0, 1] per ref frame type
131
    int ref_count[3];
132

    
133
    /**
134
     * Base parameters for segmentation, i.e. per-macroblock parameters.
135
     * These must be kept unchanged even if segmentation is not used for
136
     * a frame, since the values persist between interframes.
137
     */
138
    struct {
139
        int enabled;
140
        int absolute_vals;
141
        int update_map;
142
        int8_t base_quant[4];
143
        int8_t filter_level[4];     ///< base loop filter level
144
    } segmentation;
145

    
146
    /**
147
     * Macroblocks can have one of 4 different quants in a frame when
148
     * segmentation is enabled.
149
     * If segmentation is disabled, only the first segment's values are used.
150
     */
151
    struct {
152
        // [0] - DC qmul  [1] - AC qmul
153
        int16_t luma_qmul[2];
154
        int16_t luma_dc_qmul[2];    ///< luma dc-only block quant
155
        int16_t chroma_qmul[2];
156
    } qmat[4];
157

    
158
    struct {
159
        int simple;
160
        int level;
161
        int sharpness;
162
    } filter;
163

    
164
    struct {
165
        int enabled;    ///< whether each mb can have a different strength based on mode/ref
166

    
167
        /**
168
         * filter strength adjustment for the following macroblock modes:
169
         * [0] - i4x4
170
         * [1] - zero mv
171
         * [2] - inter modes except for zero or split mv
172
         * [3] - split mv
173
         *  i16x16 modes never have any adjustment
174
         */
175
        int8_t mode[4];
176

    
177
        /**
178
         * filter strength adjustment for macroblocks that reference:
179
         * [0] - intra / VP56_FRAME_CURRENT
180
         * [1] - VP56_FRAME_PREVIOUS
181
         * [2] - VP56_FRAME_GOLDEN
182
         * [3] - altref / VP56_FRAME_GOLDEN2
183
         */
184
        int8_t ref[4];
185
    } lf_delta;
186

    
187
    /**
188
     * These are all of the updatable probabilities for binary decisions.
189
     * They are only implictly reset on keyframes, making it quite likely
190
     * for an interframe to desync if a prior frame's header was corrupt
191
     * or missing outright!
192
     */
193
    struct {
194
        uint8_t segmentid[3];
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_base);
215
    av_freep(&s->top_nnz);
216
    av_freep(&s->edge_emu_buffer);
217
    av_freep(&s->top_border);
218
    av_freep(&s->segmentation_map);
219

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

    
224
static int update_dimensions(VP8Context *s, int width, int height)
225
{
226
    int i;
227

    
228
    if (avcodec_check_dimensions(s->avctx, width, height))
229
        return AVERROR_INVALIDDATA;
230

    
231
    vp8_decode_flush(s->avctx);
232

    
233
    avcodec_set_dimensions(s->avctx, width, height);
234

    
235
    s->mb_width  = (s->avctx->coded_width +15) / 16;
236
    s->mb_height = (s->avctx->coded_height+15) / 16;
237

    
238
    // we allocate a border around the top/left of intra4x4 modes
239
    // this is 4 blocks for intra4x4 to keep 4-byte alignment for fill_rectangle
240
    s->mb_stride = s->mb_width+1;
241
    s->b4_stride = 4*s->mb_stride;
242

    
243
    s->macroblocks_base        = av_mallocz((s->mb_stride+s->mb_height*2+2)*sizeof(*s->macroblocks));
244
    s->filter_strength         = av_mallocz(s->mb_stride*sizeof(*s->filter_strength));
245
    s->intra4x4_pred_mode_base = av_mallocz(s->b4_stride*(4*s->mb_height+1));
246
    s->top_nnz                 = av_mallocz(s->mb_width*sizeof(*s->top_nnz));
247
    s->top_border              = av_mallocz((s->mb_width+1)*sizeof(*s->top_border));
248
    s->segmentation_map        = av_mallocz(s->mb_stride*s->mb_height);
249

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

    
254
    s->macroblocks        = s->macroblocks_base + 1;
255
    s->intra4x4_pred_mode = s->intra4x4_pred_mode_base + 4 + s->b4_stride;
256

    
257
    memset(s->intra4x4_pred_mode_base, DC_PRED, s->b4_stride);
258
    for (i = 0; i < 4*s->mb_height; i++)
259
        s->intra4x4_pred_mode[i*s->b4_stride-1] = DC_PRED;
260

    
261
    return 0;
262
}
263

    
264
static void parse_segment_info(VP8Context *s)
265
{
266
    VP56RangeCoder *c = &s->c;
267
    int i;
268

    
269
    s->segmentation.update_map = vp8_rac_get(c);
270

    
271
    if (vp8_rac_get(c)) { // update segment feature data
272
        s->segmentation.absolute_vals = vp8_rac_get(c);
273

    
274
        for (i = 0; i < 4; i++)
275
            s->segmentation.base_quant[i]   = vp8_rac_get_sint(c, 7);
276

    
277
        for (i = 0; i < 4; i++)
278
            s->segmentation.filter_level[i] = vp8_rac_get_sint(c, 6);
279
    }
280
    if (s->segmentation.update_map)
281
        for (i = 0; i < 3; i++)
282
            s->prob->segmentid[i] = vp8_rac_get(c) ? vp8_rac_get_uint(c, 8) : 255;
283
}
284

    
285
static void update_lf_deltas(VP8Context *s)
286
{
287
    VP56RangeCoder *c = &s->c;
288
    int i;
289

    
290
    for (i = 0; i < 4; i++)
291
        s->lf_delta.ref[i]  = vp8_rac_get_sint(c, 6);
292

    
293
    for (i = 0; i < 4; i++)
294
        s->lf_delta.mode[i] = vp8_rac_get_sint(c, 6);
295
}
296

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

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

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

    
309
    for (i = 0; i < s->num_coeff_partitions-1; i++) {
310
        int size = AV_RL24(sizes + 3*i);
311
        if (buf_size - size < 0)
312
            return -1;
313

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

    
320
    return 0;
321
}
322

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

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

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

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

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

    
355
/**
356
 * Determine which buffers golden and altref should be updated with after this frame.
357
 * The spec isn't clear here, so I'm going by my understanding of what libvpx does
358
 *
359
 * Intra frames update all 3 references
360
 * Inter frames update VP56_FRAME_PREVIOUS if the update_last flag is set
361
 * If the update (golden|altref) flag is set, it's updated with the current frame
362
 *      if update_last is set, and VP56_FRAME_PREVIOUS otherwise.
363
 * If the flag is not set, the number read means:
364
 *      0: no update
365
 *      1: VP56_FRAME_PREVIOUS
366
 *      2: update golden with altref, or update altref with golden
367
 */
368
static VP56Frame ref_to_update(VP8Context *s, int update, VP56Frame ref)
369
{
370
    VP56RangeCoder *c = &s->c;
371

    
372
    if (update)
373
        return VP56_FRAME_CURRENT;
374

    
375
    switch (vp8_rac_get_uint(c, 2)) {
376
    case 1:
377
        return VP56_FRAME_PREVIOUS;
378
    case 2:
379
        return (ref == VP56_FRAME_GOLDEN) ? VP56_FRAME_GOLDEN2 : VP56_FRAME_GOLDEN;
380
    }
381
    return VP56_FRAME_NONE;
382
}
383

    
384
static void update_refs(VP8Context *s)
385
{
386
    VP56RangeCoder *c = &s->c;
387

    
388
    int update_golden = vp8_rac_get(c);
389
    int update_altref = vp8_rac_get(c);
390

    
391
    s->update_golden = ref_to_update(s, update_golden, VP56_FRAME_GOLDEN);
392
    s->update_altref = ref_to_update(s, update_altref, VP56_FRAME_GOLDEN2);
393
}
394

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

    
402
    s->keyframe  = !(buf[0] & 1);
403
    s->profile   =  (buf[0]>>1) & 7;
404
    s->invisible = !(buf[0] & 0x10);
405
    header_size  = AV_RL24(buf) >> 5;
406
    buf      += 3;
407
    buf_size -= 3;
408

    
409
    if (s->profile > 3)
410
        av_log(s->avctx, AV_LOG_WARNING, "Unknown profile %d\n", s->profile);
411

    
412
    if (!s->profile)
413
        memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_epel_pixels_tab, sizeof(s->put_pixels_tab));
414
    else    // profile 1-3 use bilinear, 4+ aren't defined so whatever
415
        memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_bilinear_pixels_tab, sizeof(s->put_pixels_tab));
416

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

    
422
    if (s->keyframe) {
423
        if (AV_RL24(buf) != 0x2a019d) {
424
            av_log(s->avctx, AV_LOG_ERROR, "Invalid start code 0x%x\n", AV_RL24(buf));
425
            return AVERROR_INVALIDDATA;
426
        }
427
        width  = AV_RL16(buf+3) & 0x3fff;
428
        height = AV_RL16(buf+5) & 0x3fff;
429
        hscale = buf[4] >> 6;
430
        vscale = buf[6] >> 6;
431
        buf      += 7;
432
        buf_size -= 7;
433

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

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

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

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

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

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

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

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

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

    
479
    get_quants(s);
480

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

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

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

    
494
    for (i = 0; i < 4; i++)
495
        for (j = 0; j < 8; j++)
496
            for (k = 0; k < 3; k++)
497
                for (l = 0; l < NUM_DCT_TOKENS-1; l++)
498
                    if (vp56_rac_get_prob_branchy(c, vp8_token_update_probs[i][j][k][l]))
499
                        s->prob->token[i][j][k][l] = vp8_rac_get_uint(c, 8);
500

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

    
504
    if (!s->keyframe) {
505
        s->intra_proba  = vp8_rac_get_uint(c, 8);
506
        s->last_proba   = vp8_rac_get_uint(c, 8);
507
        s->golden_proba = vp8_rac_get_uint(c, 8);
508

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

    
516
        // 17.2 MV probability update
517
        for (i = 0; i < 2; i++)
518
            for (j = 0; j < 19; j++)
519
                if (vp56_rac_get_prob_branchy(c, vp8_mv_update_prob[i][j]))
520
                    s->prob->mvc[i][j] = vp8_rac_get_nn(c);
521
    }
522

    
523
    return 0;
524
}
525

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

    
536
static av_always_inline
537
void find_near_mvs(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
538
                   VP56mv near[2], VP56mv *best, uint8_t cnt[4])
539
{
540
    VP8Macroblock *mb_edge[3] = { mb + 2 /* top */,
541
                                  mb - 1 /* left */,
542
                                  mb + 1 /* top-left */ };
543
    enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT };
544
    VP56mv near_mv[4]  = {{ 0 }};
545
    enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV };
546
    int idx = CNT_ZERO;
547
    int best_idx = CNT_ZERO;
548
    int cur_sign_bias = s->sign_bias[mb->ref_frame];
549
    int *sign_bias = s->sign_bias;
550

    
551
    /* Process MB on top, left and top-left */
552
    #define MV_EDGE_CHECK(n)\
553
    {\
554
        VP8Macroblock *edge = mb_edge[n];\
555
        int edge_ref = edge->ref_frame;\
556
        if (edge_ref != VP56_FRAME_CURRENT) {\
557
            uint32_t mv = AV_RN32A(&edge->mv);\
558
            if (mv) {\
559
                if (cur_sign_bias != sign_bias[edge_ref]) {\
560
                    /* SWAR negate of the values in mv. */\
561
                    mv = ~mv;\
562
                    mv = ((mv&0x7fff7fff) + 0x00010001) ^ (mv&0x80008000);\
563
                }\
564
                if (!n || mv != AV_RN32A(&near_mv[idx]))\
565
                    AV_WN32A(&near_mv[++idx], mv);\
566
                cnt[idx]      += 1 + (n != 2);\
567
            } else\
568
                cnt[CNT_ZERO] += 1 + (n != 2);\
569
        }\
570
    }
571
    MV_EDGE_CHECK(0)
572
    MV_EDGE_CHECK(1)
573
    MV_EDGE_CHECK(2)
574

    
575
    /* If we have three distinct MVs, merge first and last if they're the same */
576
    if (cnt[CNT_SPLITMV] && AV_RN32A(&near_mv[1+EDGE_TOP]) == AV_RN32A(&near_mv[1+EDGE_TOPLEFT]))
577
        cnt[CNT_NEAREST] += 1;
578

    
579
    cnt[CNT_SPLITMV] = ((mb_edge[EDGE_LEFT]->mode   == VP8_MVMODE_SPLIT) +
580
                        (mb_edge[EDGE_TOP]->mode    == VP8_MVMODE_SPLIT)) * 2 +
581
                       (mb_edge[EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT);
582

    
583
    /* Swap near and nearest if necessary */
584
    if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) {
585
        FFSWAP(uint8_t,     cnt[CNT_NEAREST],     cnt[CNT_NEAR]);
586
        FFSWAP( VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]);
587
    }
588

    
589
    /* Choose the best mv out of 0,0 and the nearest mv */
590
    if (cnt[CNT_NEAREST] >= cnt[CNT_ZERO])
591
        best_idx = CNT_NEAREST;
592

    
593
    mb->mv  = near_mv[best_idx];
594
    near[0] = near_mv[CNT_NEAREST];
595
    near[1] = near_mv[CNT_NEAR];
596
}
597

    
598
/**
599
 * Motion vector coding, 17.1.
600
 */
601
static int read_mv_component(VP56RangeCoder *c, const uint8_t *p)
602
{
603
    int bit, x = 0;
604

    
605
    if (vp56_rac_get_prob_branchy(c, p[0])) {
606
        int i;
607

    
608
        for (i = 0; i < 3; i++)
609
            x += vp56_rac_get_prob(c, p[9 + i]) << i;
610
        for (i = 9; i > 3; i--)
611
            x += vp56_rac_get_prob(c, p[9 + i]) << i;
612
        if (!(x & 0xFFF0) || vp56_rac_get_prob(c, p[12]))
613
            x += 8;
614
    } else {
615
        // small_mvtree
616
        const uint8_t *ps = p+2;
617
        bit = vp56_rac_get_prob(c, *ps);
618
        ps += 1 + 3*bit;
619
        x  += 4*bit;
620
        bit = vp56_rac_get_prob(c, *ps);
621
        ps += 1 + bit;
622
        x  += 2*bit;
623
        x  += vp56_rac_get_prob(c, *ps);
624
    }
625

    
626
    return (x && vp56_rac_get_prob(c, p[1])) ? -x : x;
627
}
628

    
629
static av_always_inline
630
const uint8_t *get_submv_prob(uint32_t left, uint32_t top)
631
{
632
    if (left == top)
633
        return vp8_submv_prob[4-!!left];
634
    if (!top)
635
        return vp8_submv_prob[2];
636
    return vp8_submv_prob[1-!!left];
637
}
638

    
639
/**
640
 * Split motion vector prediction, 16.4.
641
 * @returns the number of motion vectors parsed (2, 4 or 16)
642
 */
643
static av_always_inline
644
int decode_splitmvs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb)
645
{
646
    int part_idx = mb->partitioning =
647
        vp8_rac_get_tree(c, vp8_mbsplit_tree, vp8_mbsplit_prob);
648
    int n, num = vp8_mbsplit_count[part_idx];
649
    VP8Macroblock *top_mb  = &mb[2];
650
    VP8Macroblock *left_mb = &mb[-1];
651
    const uint8_t *mbsplits_left = vp8_mbsplits[left_mb->partitioning],
652
                  *mbsplits_top = vp8_mbsplits[top_mb->partitioning],
653
                  *mbsplits_cur = vp8_mbsplits[part_idx],
654
                  *firstidx = vp8_mbfirstidx[part_idx];
655
    VP56mv *top_mv  = top_mb->bmv;
656
    VP56mv *left_mv = left_mb->bmv;
657
    VP56mv *cur_mv  = mb->bmv;
658

    
659
    for (n = 0; n < num; n++) {
660
        int k = firstidx[n];
661
        uint32_t left, above;
662
        const uint8_t *submv_prob;
663

    
664
        if (!(k & 3))
665
            left = AV_RN32A(&left_mv[mbsplits_left[k + 3]]);
666
        else
667
            left  = AV_RN32A(&cur_mv[mbsplits_cur[k - 1]]);
668
        if (k <= 3)
669
            above = AV_RN32A(&top_mv[mbsplits_top[k + 12]]);
670
        else
671
            above = AV_RN32A(&cur_mv[mbsplits_cur[k - 4]]);
672

    
673
        submv_prob = get_submv_prob(left, above);
674

    
675
        switch (vp8_rac_get_tree(c, vp8_submv_ref_tree, submv_prob)) {
676
        case VP8_SUBMVMODE_NEW4X4:
677
            mb->bmv[n].y = mb->mv.y + read_mv_component(c, s->prob->mvc[0]);
678
            mb->bmv[n].x = mb->mv.x + read_mv_component(c, s->prob->mvc[1]);
679
            break;
680
        case VP8_SUBMVMODE_ZERO4X4:
681
            AV_ZERO32(&mb->bmv[n]);
682
            break;
683
        case VP8_SUBMVMODE_LEFT4X4:
684
            AV_WN32A(&mb->bmv[n], left);
685
            break;
686
        case VP8_SUBMVMODE_TOP4X4:
687
            AV_WN32A(&mb->bmv[n], above);
688
            break;
689
        }
690
    }
691

    
692
    return num;
693
}
694

    
695
static av_always_inline
696
void decode_intra4x4_modes(VP56RangeCoder *c, uint8_t *intra4x4,
697
                           int stride, int keyframe)
698
{
699
    int x, y, t, l, i;
700

    
701
    if (keyframe) {
702
        const uint8_t *ctx;
703
        for (y = 0; y < 4; y++) {
704
            for (x = 0; x < 4; x++) {
705
                t = intra4x4[x - stride];
706
                l = intra4x4[x - 1];
707
                ctx = vp8_pred4x4_prob_intra[t][l];
708
                intra4x4[x] = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx);
709
            }
710
            intra4x4 += stride;
711
        }
712
    } else {
713
        for (i = 0; i < 16; i++)
714
            intra4x4[i] = vp8_rac_get_tree(c, vp8_pred4x4_tree, vp8_pred4x4_prob_inter);
715
    }
716
}
717

    
718
static av_always_inline
719
void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
720
                    uint8_t *intra4x4, uint8_t *segment)
721
{
722
    VP56RangeCoder *c = &s->c;
723

    
724
    if (s->segmentation.update_map)
725
        *segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid);
726
    s->segment = *segment;
727

    
728
    mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->mbskip_proba) : 0;
729

    
730
    if (s->keyframe) {
731
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra);
732

    
733
        if (mb->mode == MODE_I4x4) {
734
            decode_intra4x4_modes(c, intra4x4, s->b4_stride, 1);
735
        } else
736
            fill_rectangle(intra4x4, 4, 4, s->b4_stride, vp8_pred4x4_mode[mb->mode], 1);
737

    
738
        s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra);
739
        mb->ref_frame = VP56_FRAME_CURRENT;
740
    } else if (vp56_rac_get_prob_branchy(c, s->intra_proba)) {
741
        VP56mv near[2], best;
742
        uint8_t cnt[4] = { 0 };
743
        uint8_t p[4];
744

    
745
        // inter MB, 16.2
746
        if (vp56_rac_get_prob_branchy(c, s->last_proba))
747
            mb->ref_frame = vp56_rac_get_prob(c, s->golden_proba) ?
748
                VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN;
749
        else
750
            mb->ref_frame = VP56_FRAME_PREVIOUS;
751
        s->ref_count[mb->ref_frame-1]++;
752

    
753
        // motion vectors, 16.3
754
        find_near_mvs(s, mb, mb_x, mb_y, near, &best, cnt);
755
        p[0] = vp8_mode_contexts[cnt[0]][0];
756
        p[1] = vp8_mode_contexts[cnt[1]][1];
757
        p[2] = vp8_mode_contexts[cnt[2]][2];
758
        p[3] = vp8_mode_contexts[cnt[3]][3];
759
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_mvinter, p);
760
        switch (mb->mode) {
761
        case VP8_MVMODE_SPLIT:
762
            clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y);
763
            mb->mv = mb->bmv[decode_splitmvs(s, c, mb) - 1];
764
            break;
765
        case VP8_MVMODE_ZERO:
766
            AV_ZERO32(&mb->mv);
767
            break;
768
        case VP8_MVMODE_NEAREST:
769
            clamp_mv(s, &mb->mv, &near[0], mb_x, mb_y);
770
            break;
771
        case VP8_MVMODE_NEAR:
772
            clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y);
773
            break;
774
        case VP8_MVMODE_NEW:
775
            clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y);
776
            mb->mv.y += + read_mv_component(c, s->prob->mvc[0]);
777
            mb->mv.x += + read_mv_component(c, s->prob->mvc[1]);
778
            break;
779
        }
780
        if (mb->mode != VP8_MVMODE_SPLIT) {
781
            mb->partitioning = VP8_SPLITMVMODE_NONE;
782
            mb->bmv[0] = mb->mv;
783
        }
784
    } else {
785
        // intra MB, 16.1
786
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16);
787

    
788
        if (mb->mode == MODE_I4x4)
789
            decode_intra4x4_modes(c, intra4x4, 4, 0);
790

    
791
        s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c);
792
        mb->ref_frame = VP56_FRAME_CURRENT;
793
        mb->partitioning = VP8_SPLITMVMODE_NONE;
794
        AV_ZERO32(&mb->bmv[0]);
795
    }
796
}
797

    
798
/**
799
 * @param c arithmetic bitstream reader context
800
 * @param block destination for block coefficients
801
 * @param probs probabilities to use when reading trees from the bitstream
802
 * @param i initial coeff index, 0 unless a separate DC block is coded
803
 * @param zero_nhood the initial prediction context for number of surrounding
804
 *                   all-zero blocks (only left/top, so 0-2)
805
 * @param qmul array holding the dc/ac dequant factor at position 0/1
806
 * @return 0 if no coeffs were decoded
807
 *         otherwise, the index of the last coeff decoded plus one
808
 */
809
static int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16],
810
                               uint8_t probs[8][3][NUM_DCT_TOKENS-1],
811
                               int i, int zero_nhood, int16_t qmul[2])
812
{
813
    uint8_t *token_prob;
814
    int nonzero = 0;
815
    int coeff;
816

    
817
    do {
818
        token_prob = probs[vp8_coeff_band[i]][zero_nhood];
819

    
820
        if (!vp56_rac_get_prob_branchy(c, token_prob[0]))   // DCT_EOB
821
            return nonzero;
822

    
823
skip_eob:
824
        if (!vp56_rac_get_prob_branchy(c, token_prob[1])) { // DCT_0
825
            zero_nhood = 0;
826
            token_prob = probs[vp8_coeff_band[++i]][0];
827
            if (i < 16)
828
                goto skip_eob;
829
            return nonzero; // invalid input; blocks should end with EOB
830
        }
831

    
832
        if (!vp56_rac_get_prob_branchy(c, token_prob[2])) { // DCT_1
833
            coeff = 1;
834
            zero_nhood = 1;
835
        } else {
836
            zero_nhood = 2;
837

    
838
            if (!vp56_rac_get_prob_branchy(c, token_prob[3])) { // DCT 2,3,4
839
                coeff = vp56_rac_get_prob(c, token_prob[4]);
840
                if (coeff)
841
                    coeff += vp56_rac_get_prob(c, token_prob[5]);
842
                coeff += 2;
843
            } else {
844
                // DCT_CAT*
845
                if (!vp56_rac_get_prob_branchy(c, token_prob[6])) {
846
                    if (!vp56_rac_get_prob_branchy(c, token_prob[7])) { // DCT_CAT1
847
                        coeff  = 5 + vp56_rac_get_prob(c, vp8_dct_cat1_prob[0]);
848
                    } else {                                    // DCT_CAT2
849
                        coeff  = 7;
850
                        coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[0]) << 1;
851
                        coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[1]);
852
                    }
853
                } else {    // DCT_CAT3 and up
854
                    int a = vp56_rac_get_prob(c, token_prob[8]);
855
                    int b = vp56_rac_get_prob(c, token_prob[9+a]);
856
                    int cat = (a<<1) + b;
857
                    coeff  = 3 + (8<<cat);
858
                    coeff += vp8_rac_get_coeff(c, vp8_dct_cat_prob[cat]);
859
                }
860
            }
861
        }
862

    
863
        // todo: full [16] qmat? load into register?
864
        block[zigzag_scan[i]] = (vp8_rac_get(c) ? -coeff : coeff) * qmul[!!i];
865
        nonzero = ++i;
866
    } while (i < 16);
867

    
868
    return nonzero;
869
}
870

    
871
static av_always_inline
872
void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,
873
                      uint8_t t_nnz[9], uint8_t l_nnz[9])
874
{
875
    LOCAL_ALIGNED_16(DCTELEM, dc,[16]);
876
    int i, x, y, luma_start = 0, luma_ctx = 3;
877
    int nnz_pred, nnz, nnz_total = 0;
878
    int segment = s->segment;
879

    
880
    if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
881
        AV_ZERO128(dc);
882
        AV_ZERO128(dc+8);
883
        nnz_pred = t_nnz[8] + l_nnz[8];
884

    
885
        // decode DC values and do hadamard
886
        nnz = decode_block_coeffs(c, dc, s->prob->token[1], 0, nnz_pred,
887
                                  s->qmat[segment].luma_dc_qmul);
888
        l_nnz[8] = t_nnz[8] = !!nnz;
889
        nnz_total += nnz;
890
        s->vp8dsp.vp8_luma_dc_wht(s->block, dc);
891
        luma_start = 1;
892
        luma_ctx = 0;
893
    }
894

    
895
    // luma blocks
896
    for (y = 0; y < 4; y++)
897
        for (x = 0; x < 4; x++) {
898
            nnz_pred = l_nnz[y] + t_nnz[x];
899
            nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start,
900
                                      nnz_pred, s->qmat[segment].luma_qmul);
901
            // nnz+luma_start may be one more than the actual last index, but we don't care
902
            s->non_zero_count_cache[y][x] = nnz + luma_start;
903
            t_nnz[x] = l_nnz[y] = !!nnz;
904
            nnz_total += nnz;
905
        }
906

    
907
    // chroma blocks
908
    // TODO: what to do about dimensions? 2nd dim for luma is x,
909
    // but for chroma it's (y<<1)|x
910
    for (i = 4; i < 6; i++)
911
        for (y = 0; y < 2; y++)
912
            for (x = 0; x < 2; x++) {
913
                nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x];
914
                nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0,
915
                                          nnz_pred, s->qmat[segment].chroma_qmul);
916
                s->non_zero_count_cache[i][(y<<1)+x] = nnz;
917
                t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz;
918
                nnz_total += nnz;
919
            }
920

    
921
    // if there were no coded coeffs despite the macroblock not being marked skip,
922
    // we MUST not do the inner loop filter and should not do IDCT
923
    // Since skip isn't used for bitstream prediction, just manually set it.
924
    if (!nnz_total)
925
        mb->skip = 1;
926
}
927

    
928
static av_always_inline
929
void backup_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
930
                      int linesize, int uvlinesize, int simple)
931
{
932
    AV_COPY128(top_border, src_y + 15*linesize);
933
    if (!simple) {
934
        AV_COPY64(top_border+16, src_cb + 7*uvlinesize);
935
        AV_COPY64(top_border+24, src_cr + 7*uvlinesize);
936
    }
937
}
938

    
939
static av_always_inline
940
void xchg_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
941
                    int linesize, int uvlinesize, int mb_x, int mb_y, int mb_width,
942
                    int simple, int xchg)
943
{
944
    uint8_t *top_border_m1 = top_border-32;     // for TL prediction
945
    src_y  -=   linesize;
946
    src_cb -= uvlinesize;
947
    src_cr -= uvlinesize;
948

    
949
#define XCHG(a,b,xchg) do {                     \
950
        if (xchg) AV_SWAP64(b,a);               \
951
        else      AV_COPY64(b,a);               \
952
    } while (0)
953

    
954
    XCHG(top_border_m1+8, src_y-8, xchg);
955
    XCHG(top_border,      src_y,   xchg);
956
    XCHG(top_border+8,    src_y+8, 1);
957
    if (mb_x < mb_width-1)
958
        XCHG(top_border+32, src_y+16, 1);
959

    
960
    // only copy chroma for normal loop filter
961
    // or to initialize the top row to 127
962
    if (!simple || !mb_y) {
963
        XCHG(top_border_m1+16, src_cb-8, xchg);
964
        XCHG(top_border_m1+24, src_cr-8, xchg);
965
        XCHG(top_border+16,    src_cb, 1);
966
        XCHG(top_border+24,    src_cr, 1);
967
    }
968
}
969

    
970
static av_always_inline
971
int check_intra_pred_mode(int mode, int mb_x, int mb_y)
972
{
973
    if (mode == DC_PRED8x8) {
974
        if (!mb_x) {
975
            mode = mb_y ? TOP_DC_PRED8x8 : DC_128_PRED8x8;
976
        } else if (!mb_y) {
977
            mode = LEFT_DC_PRED8x8;
978
        }
979
    }
980
    return mode;
981
}
982

    
983
static av_always_inline
984
void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
985
                   uint8_t *intra4x4, int mb_x, int mb_y)
986
{
987
    int x, y, mode, nnz, tr;
988

    
989
    // for the first row, we need to run xchg_mb_border to init the top edge to 127
990
    // otherwise, skip it if we aren't going to deblock
991
    if (s->deblock_filter || !mb_y)
992
        xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
993
                       s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
994
                       s->filter.simple, 1);
995

    
996
    if (mb->mode < MODE_I4x4) {
997
        mode = check_intra_pred_mode(mb->mode, mb_x, mb_y);
998
        s->hpc.pred16x16[mode](dst[0], s->linesize);
999
    } else {
1000
        uint8_t *ptr = dst[0];
1001
        int stride = s->keyframe ? s->b4_stride : 4;
1002

    
1003
        // all blocks on the right edge of the macroblock use bottom edge
1004
        // the top macroblock for their topright edge
1005
        uint8_t *tr_right = ptr - s->linesize + 16;
1006

    
1007
        // if we're on the right edge of the frame, said edge is extended
1008
        // from the top macroblock
1009
        if (mb_x == s->mb_width-1) {
1010
            tr = tr_right[-1]*0x01010101;
1011
            tr_right = (uint8_t *)&tr;
1012
        }
1013

    
1014
        if (mb->skip)
1015
            AV_ZERO128(s->non_zero_count_cache);
1016

    
1017
        for (y = 0; y < 4; y++) {
1018
            uint8_t *topright = ptr + 4 - s->linesize;
1019
            for (x = 0; x < 4; x++) {
1020
                if (x == 3)
1021
                    topright = tr_right;
1022

    
1023
                s->hpc.pred4x4[intra4x4[x]](ptr+4*x, topright, s->linesize);
1024

    
1025
                nnz = s->non_zero_count_cache[y][x];
1026
                if (nnz) {
1027
                    if (nnz == 1)
1028
                        s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize);
1029
                    else
1030
                        s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize);
1031
                }
1032
                topright += 4;
1033
            }
1034

    
1035
            ptr   += 4*s->linesize;
1036
            intra4x4 += stride;
1037
        }
1038
    }
1039

    
1040
    mode = check_intra_pred_mode(s->chroma_pred_mode, mb_x, mb_y);
1041
    s->hpc.pred8x8[mode](dst[1], s->uvlinesize);
1042
    s->hpc.pred8x8[mode](dst[2], s->uvlinesize);
1043

    
1044
    if (s->deblock_filter || !mb_y)
1045
        xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
1046
                       s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
1047
                       s->filter.simple, 0);
1048
}
1049

    
1050
/**
1051
 * Generic MC function.
1052
 *
1053
 * @param s VP8 decoding context
1054
 * @param luma 1 for luma (Y) planes, 0 for chroma (Cb/Cr) planes
1055
 * @param dst target buffer for block data at block position
1056
 * @param src reference picture buffer at origin (0, 0)
1057
 * @param mv motion vector (relative to block position) to get pixel data from
1058
 * @param x_off horizontal position of block from origin (0, 0)
1059
 * @param y_off vertical position of block from origin (0, 0)
1060
 * @param block_w width of block (16, 8 or 4)
1061
 * @param block_h height of block (always same as block_w)
1062
 * @param width width of src/dst plane data
1063
 * @param height height of src/dst plane data
1064
 * @param linesize size of a single line of plane data, including padding
1065
 * @param mc_func motion compensation function pointers (bilinear or sixtap MC)
1066
 */
1067
static av_always_inline
1068
void vp8_mc(VP8Context *s, int luma,
1069
            uint8_t *dst, uint8_t *src, const VP56mv *mv,
1070
            int x_off, int y_off, int block_w, int block_h,
1071
            int width, int height, int linesize,
1072
            vp8_mc_func mc_func[3][3])
1073
{
1074
    if (AV_RN32A(mv)) {
1075
        static const uint8_t idx[8] = { 0, 1, 2, 1, 2, 1, 2, 1 };
1076
        int mx = (mv->x << luma)&7, mx_idx = idx[mx];
1077
        int my = (mv->y << luma)&7, my_idx = idx[my];
1078

    
1079
        x_off += mv->x >> (3 - luma);
1080
        y_off += mv->y >> (3 - luma);
1081

    
1082
        // edge emulation
1083
        src += y_off * linesize + x_off;
1084
        if (x_off < 2 || x_off >= width  - block_w - 3 ||
1085
            y_off < 2 || y_off >= height - block_h - 3) {
1086
            ff_emulated_edge_mc(s->edge_emu_buffer, src - 2 * linesize - 2, linesize,
1087
                                block_w + 5, block_h + 5,
1088
                                x_off - 2, y_off - 2, width, height);
1089
            src = s->edge_emu_buffer + 2 + linesize * 2;
1090
        }
1091
        mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);
1092
    } else
1093
        mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0);
1094
}
1095

    
1096
static av_always_inline
1097
void vp8_mc_part(VP8Context *s, uint8_t *dst[3],
1098
                 AVFrame *ref_frame, int x_off, int y_off,
1099
                 int bx_off, int by_off,
1100
                 int block_w, int block_h,
1101
                 int width, int height, VP56mv *mv)
1102
{
1103
    VP56mv uvmv = *mv;
1104

    
1105
    /* Y */
1106
    vp8_mc(s, 1, dst[0] + by_off * s->linesize + bx_off,
1107
           ref_frame->data[0], mv, x_off + bx_off, y_off + by_off,
1108
           block_w, block_h, width, height, s->linesize,
1109
           s->put_pixels_tab[block_w == 8]);
1110

    
1111
    /* U/V */
1112
    if (s->profile == 3) {
1113
        uvmv.x &= ~7;
1114
        uvmv.y &= ~7;
1115
    }
1116
    x_off   >>= 1; y_off   >>= 1;
1117
    bx_off  >>= 1; by_off  >>= 1;
1118
    width   >>= 1; height  >>= 1;
1119
    block_w >>= 1; block_h >>= 1;
1120
    vp8_mc(s, 0, dst[1] + by_off * s->uvlinesize + bx_off,
1121
           ref_frame->data[1], &uvmv, x_off + bx_off, y_off + by_off,
1122
           block_w, block_h, width, height, s->uvlinesize,
1123
           s->put_pixels_tab[1 + (block_w == 4)]);
1124
    vp8_mc(s, 0, dst[2] + by_off * s->uvlinesize + bx_off,
1125
           ref_frame->data[2], &uvmv, x_off + bx_off, y_off + by_off,
1126
           block_w, block_h, width, height, s->uvlinesize,
1127
           s->put_pixels_tab[1 + (block_w == 4)]);
1128
}
1129

    
1130
/* Fetch pixels for estimated mv 4 macroblocks ahead.
1131
 * Optimized for 64-byte cache lines.  Inspired by ffh264 prefetch_motion. */
1132
static av_always_inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int mb_xy, int ref)
1133
{
1134
    /* Don't prefetch refs that haven't been used very often this frame. */
1135
    if (s->ref_count[ref-1] > (mb_xy >> 5)) {
1136
        int x_off = mb_x << 4, y_off = mb_y << 4;
1137
        int mx = mb->mv.x + x_off + 8;
1138
        int my = mb->mv.y + y_off;
1139
        uint8_t **src= s->framep[ref]->data;
1140
        int off= mx + (my + (mb_x&3)*4)*s->linesize + 64;
1141
        s->dsp.prefetch(src[0]+off, s->linesize, 4);
1142
        off= (mx>>1) + ((my>>1) + (mb_x&7))*s->uvlinesize + 64;
1143
        s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1144
    }
1145
}
1146

    
1147
/**
1148
 * Apply motion vectors to prediction buffer, chapter 18.
1149
 */
1150
static av_always_inline
1151
void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
1152
                   int mb_x, int mb_y)
1153
{
1154
    int x_off = mb_x << 4, y_off = mb_y << 4;
1155
    int width = 16*s->mb_width, height = 16*s->mb_height;
1156
    AVFrame *ref = s->framep[mb->ref_frame];
1157
    VP56mv *bmv = mb->bmv;
1158

    
1159
    if (mb->mode < VP8_MVMODE_SPLIT) {
1160
        vp8_mc_part(s, dst, ref, x_off, y_off,
1161
                    0, 0, 16, 16, width, height, &mb->mv);
1162
    } else switch (mb->partitioning) {
1163
    case VP8_SPLITMVMODE_4x4: {
1164
        int x, y;
1165
        VP56mv uvmv;
1166

    
1167
        /* Y */
1168
        for (y = 0; y < 4; y++) {
1169
            for (x = 0; x < 4; x++) {
1170
                vp8_mc(s, 1, dst[0] + 4*y*s->linesize + x*4,
1171
                       ref->data[0], &bmv[4*y + x],
1172
                       4*x + x_off, 4*y + y_off, 4, 4,
1173
                       width, height, s->linesize,
1174
                       s->put_pixels_tab[2]);
1175
            }
1176
        }
1177

    
1178
        /* U/V */
1179
        x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1;
1180
        for (y = 0; y < 2; y++) {
1181
            for (x = 0; x < 2; x++) {
1182
                uvmv.x = mb->bmv[ 2*y    * 4 + 2*x  ].x +
1183
                         mb->bmv[ 2*y    * 4 + 2*x+1].x +
1184
                         mb->bmv[(2*y+1) * 4 + 2*x  ].x +
1185
                         mb->bmv[(2*y+1) * 4 + 2*x+1].x;
1186
                uvmv.y = mb->bmv[ 2*y    * 4 + 2*x  ].y +
1187
                         mb->bmv[ 2*y    * 4 + 2*x+1].y +
1188
                         mb->bmv[(2*y+1) * 4 + 2*x  ].y +
1189
                         mb->bmv[(2*y+1) * 4 + 2*x+1].y;
1190
                uvmv.x = (uvmv.x + 2 + (uvmv.x >> (INT_BIT-1))) >> 2;
1191
                uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2;
1192
                if (s->profile == 3) {
1193
                    uvmv.x &= ~7;
1194
                    uvmv.y &= ~7;
1195
                }
1196
                vp8_mc(s, 0, dst[1] + 4*y*s->uvlinesize + x*4,
1197
                       ref->data[1], &uvmv,
1198
                       4*x + x_off, 4*y + y_off, 4, 4,
1199
                       width, height, s->uvlinesize,
1200
                       s->put_pixels_tab[2]);
1201
                vp8_mc(s, 0, dst[2] + 4*y*s->uvlinesize + x*4,
1202
                       ref->data[2], &uvmv,
1203
                       4*x + x_off, 4*y + y_off, 4, 4,
1204
                       width, height, s->uvlinesize,
1205
                       s->put_pixels_tab[2]);
1206
            }
1207
        }
1208
        break;
1209
    }
1210
    case VP8_SPLITMVMODE_16x8:
1211
        vp8_mc_part(s, dst, ref, x_off, y_off,
1212
                    0, 0, 16, 8, width, height, &bmv[0]);
1213
        vp8_mc_part(s, dst, ref, x_off, y_off,
1214
                    0, 8, 16, 8, width, height, &bmv[1]);
1215
        break;
1216
    case VP8_SPLITMVMODE_8x16:
1217
        vp8_mc_part(s, dst, ref, x_off, y_off,
1218
                    0, 0, 8, 16, width, height, &bmv[0]);
1219
        vp8_mc_part(s, dst, ref, x_off, y_off,
1220
                    8, 0, 8, 16, width, height, &bmv[1]);
1221
        break;
1222
    case VP8_SPLITMVMODE_8x8:
1223
        vp8_mc_part(s, dst, ref, x_off, y_off,
1224
                    0, 0, 8, 8, width, height, &bmv[0]);
1225
        vp8_mc_part(s, dst, ref, x_off, y_off,
1226
                    8, 0, 8, 8, width, height, &bmv[1]);
1227
        vp8_mc_part(s, dst, ref, x_off, y_off,
1228
                    0, 8, 8, 8, width, height, &bmv[2]);
1229
        vp8_mc_part(s, dst, ref, x_off, y_off,
1230
                    8, 8, 8, 8, width, height, &bmv[3]);
1231
        break;
1232
    }
1233
}
1234

    
1235
static av_always_inline void idct_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb)
1236
{
1237
    int x, y, ch;
1238

    
1239
    if (mb->mode != MODE_I4x4) {
1240
        uint8_t *y_dst = dst[0];
1241
        for (y = 0; y < 4; y++) {
1242
            uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[y]);
1243
            if (nnz4) {
1244
                if (nnz4&~0x01010101) {
1245
                    for (x = 0; x < 4; x++) {
1246
                        int nnz = s->non_zero_count_cache[y][x];
1247
                        if (nnz) {
1248
                            if (nnz == 1)
1249
                                s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize);
1250
                            else
1251
                                s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize);
1252
                        }
1253
                    }
1254
                } else {
1255
                    s->vp8dsp.vp8_idct_dc_add4y(y_dst, s->block[y], s->linesize);
1256
                }
1257
            }
1258
            y_dst += 4*s->linesize;
1259
        }
1260
    }
1261

    
1262
    for (ch = 0; ch < 2; ch++) {
1263
        uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[4+ch]);
1264
        if (nnz4) {
1265
            uint8_t *ch_dst = dst[1+ch];
1266
            if (nnz4&~0x01010101) {
1267
                for (y = 0; y < 2; y++) {
1268
                    for (x = 0; x < 2; x++) {
1269
                        int nnz = s->non_zero_count_cache[4+ch][(y<<1)+x];
1270
                        if (nnz) {
1271
                            if (nnz == 1)
1272
                                s->vp8dsp.vp8_idct_dc_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);
1273
                            else
1274
                                s->vp8dsp.vp8_idct_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);
1275
                        }
1276
                    }
1277
                    ch_dst += 4*s->uvlinesize;
1278
                }
1279
            } else {
1280
                s->vp8dsp.vp8_idct_dc_add4uv(ch_dst, s->block[4+ch], s->uvlinesize);
1281
            }
1282
        }
1283
    }
1284
}
1285

    
1286
static av_always_inline void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f )
1287
{
1288
    int interior_limit, filter_level;
1289

    
1290
    if (s->segmentation.enabled) {
1291
        filter_level = s->segmentation.filter_level[s->segment];
1292
        if (!s->segmentation.absolute_vals)
1293
            filter_level += s->filter.level;
1294
    } else
1295
        filter_level = s->filter.level;
1296

    
1297
    if (s->lf_delta.enabled) {
1298
        filter_level += s->lf_delta.ref[mb->ref_frame];
1299

    
1300
        if (mb->ref_frame == VP56_FRAME_CURRENT) {
1301
            if (mb->mode == MODE_I4x4)
1302
                filter_level += s->lf_delta.mode[0];
1303
        } else {
1304
            if (mb->mode == VP8_MVMODE_ZERO)
1305
                filter_level += s->lf_delta.mode[1];
1306
            else if (mb->mode == VP8_MVMODE_SPLIT)
1307
                filter_level += s->lf_delta.mode[3];
1308
            else
1309
                filter_level += s->lf_delta.mode[2];
1310
        }
1311
    }
1312
    filter_level = av_clip(filter_level, 0, 63);
1313

    
1314
    interior_limit = filter_level;
1315
    if (s->filter.sharpness) {
1316
        interior_limit >>= s->filter.sharpness > 4 ? 2 : 1;
1317
        interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness);
1318
    }
1319
    interior_limit = FFMAX(interior_limit, 1);
1320

    
1321
    f->filter_level = filter_level;
1322
    f->inner_limit = interior_limit;
1323
    f->inner_filter = !mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT;
1324
}
1325

    
1326
static av_always_inline void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f, int mb_x, int mb_y)
1327
{
1328
    int mbedge_lim, bedge_lim, hev_thresh;
1329
    int filter_level = f->filter_level;
1330
    int inner_limit = f->inner_limit;
1331
    int inner_filter = f->inner_filter;
1332
    int linesize = s->linesize;
1333
    int uvlinesize = s->uvlinesize;
1334

    
1335
    if (!filter_level)
1336
        return;
1337

    
1338
    mbedge_lim = 2*(filter_level+2) + inner_limit;
1339
     bedge_lim = 2* filter_level    + inner_limit;
1340
    hev_thresh = filter_level >= 15;
1341

    
1342
    if (s->keyframe) {
1343
        if (filter_level >= 40)
1344
            hev_thresh = 2;
1345
    } else {
1346
        if (filter_level >= 40)
1347
            hev_thresh = 3;
1348
        else if (filter_level >= 20)
1349
            hev_thresh = 2;
1350
    }
1351

    
1352
    if (mb_x) {
1353
        s->vp8dsp.vp8_h_loop_filter16y(dst[0],     linesize,
1354
                                       mbedge_lim, inner_limit, hev_thresh);
1355
        s->vp8dsp.vp8_h_loop_filter8uv(dst[1],     dst[2],      uvlinesize,
1356
                                       mbedge_lim, inner_limit, hev_thresh);
1357
    }
1358

    
1359
    if (inner_filter) {
1360
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 4, linesize, bedge_lim,
1361
                                             inner_limit, hev_thresh);
1362
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 8, linesize, bedge_lim,
1363
                                             inner_limit, hev_thresh);
1364
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+12, linesize, bedge_lim,
1365
                                             inner_limit, hev_thresh);
1366
        s->vp8dsp.vp8_h_loop_filter8uv_inner(dst[1] + 4, dst[2] + 4,
1367
                                             uvlinesize,  bedge_lim,
1368
                                             inner_limit, hev_thresh);
1369
    }
1370

    
1371
    if (mb_y) {
1372
        s->vp8dsp.vp8_v_loop_filter16y(dst[0],     linesize,
1373
                                       mbedge_lim, inner_limit, hev_thresh);
1374
        s->vp8dsp.vp8_v_loop_filter8uv(dst[1],     dst[2],      uvlinesize,
1375
                                       mbedge_lim, inner_limit, hev_thresh);
1376
    }
1377

    
1378
    if (inner_filter) {
1379
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 4*linesize,
1380
                                             linesize,    bedge_lim,
1381
                                             inner_limit, hev_thresh);
1382
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 8*linesize,
1383
                                             linesize,    bedge_lim,
1384
                                             inner_limit, hev_thresh);
1385
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+12*linesize,
1386
                                             linesize,    bedge_lim,
1387
                                             inner_limit, hev_thresh);
1388
        s->vp8dsp.vp8_v_loop_filter8uv_inner(dst[1] + 4 * uvlinesize,
1389
                                             dst[2] + 4 * uvlinesize,
1390
                                             uvlinesize,  bedge_lim,
1391
                                             inner_limit, hev_thresh);
1392
    }
1393
}
1394

    
1395
static av_always_inline void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f, int mb_x, int mb_y)
1396
{
1397
    int mbedge_lim, bedge_lim;
1398
    int filter_level = f->filter_level;
1399
    int inner_limit = f->inner_limit;
1400
    int inner_filter = f->inner_filter;
1401
    int linesize = s->linesize;
1402

    
1403
    if (!filter_level)
1404
        return;
1405

    
1406
    mbedge_lim = 2*(filter_level+2) + inner_limit;
1407
     bedge_lim = 2* filter_level    + inner_limit;
1408

    
1409
    if (mb_x)
1410
        s->vp8dsp.vp8_h_loop_filter_simple(dst, linesize, mbedge_lim);
1411
    if (inner_filter) {
1412
        s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, linesize, bedge_lim);
1413
        s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, linesize, bedge_lim);
1414
        s->vp8dsp.vp8_h_loop_filter_simple(dst+12, linesize, bedge_lim);
1415
    }
1416

    
1417
    if (mb_y)
1418
        s->vp8dsp.vp8_v_loop_filter_simple(dst, linesize, mbedge_lim);
1419
    if (inner_filter) {
1420
        s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*linesize, linesize, bedge_lim);
1421
        s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*linesize, linesize, bedge_lim);
1422
        s->vp8dsp.vp8_v_loop_filter_simple(dst+12*linesize, linesize, bedge_lim);
1423
    }
1424
}
1425

    
1426
static void filter_mb_row(VP8Context *s, int mb_y)
1427
{
1428
    VP8FilterStrength *f = s->filter_strength;
1429
    uint8_t *dst[3] = {
1430
        s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize,
1431
        s->framep[VP56_FRAME_CURRENT]->data[1] +  8*mb_y*s->uvlinesize,
1432
        s->framep[VP56_FRAME_CURRENT]->data[2] +  8*mb_y*s->uvlinesize
1433
    };
1434
    int mb_x;
1435

    
1436
    for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1437
        backup_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], s->linesize, s->uvlinesize, 0);
1438
        filter_mb(s, dst, f++, mb_x, mb_y);
1439
        dst[0] += 16;
1440
        dst[1] += 8;
1441
        dst[2] += 8;
1442
    }
1443
}
1444

    
1445
static void filter_mb_row_simple(VP8Context *s, int mb_y)
1446
{
1447
    VP8FilterStrength *f = s->filter_strength;
1448
    uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize;
1449
    int mb_x;
1450

    
1451
    for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1452
        backup_mb_border(s->top_border[mb_x+1], dst, NULL, NULL, s->linesize, 0, 1);
1453
        filter_mb_simple(s, dst, f++, mb_x, mb_y);
1454
        dst += 16;
1455
    }
1456
}
1457

    
1458
static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
1459
                            AVPacket *avpkt)
1460
{
1461
    VP8Context *s = avctx->priv_data;
1462
    int ret, mb_x, mb_y, i, y, referenced;
1463
    enum AVDiscard skip_thresh;
1464
    AVFrame *av_uninit(curframe);
1465

    
1466
    if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0)
1467
        return ret;
1468

    
1469
    referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT
1470
                                || s->update_altref == VP56_FRAME_CURRENT;
1471

    
1472
    skip_thresh = !referenced ? AVDISCARD_NONREF :
1473
                    !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL;
1474

    
1475
    if (avctx->skip_frame >= skip_thresh) {
1476
        s->invisible = 1;
1477
        goto skip_decode;
1478
    }
1479
    s->deblock_filter = s->filter.level && avctx->skip_loop_filter < skip_thresh;
1480

    
1481
    for (i = 0; i < 4; i++)
1482
        if (&s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1483
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1484
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) {
1485
            curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i];
1486
            break;
1487
        }
1488
    if (curframe->data[0])
1489
        avctx->release_buffer(avctx, curframe);
1490

    
1491
    curframe->key_frame = s->keyframe;
1492
    curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE;
1493
    curframe->reference = referenced ? 3 : 0;
1494
    if ((ret = avctx->get_buffer(avctx, curframe))) {
1495
        av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n");
1496
        return ret;
1497
    }
1498

    
1499
    // Given that arithmetic probabilities are updated every frame, it's quite likely
1500
    // that the values we have on a random interframe are complete junk if we didn't
1501
    // start decode on a keyframe. So just don't display anything rather than junk.
1502
    if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] ||
1503
                         !s->framep[VP56_FRAME_GOLDEN] ||
1504
                         !s->framep[VP56_FRAME_GOLDEN2])) {
1505
        av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n");
1506
        return AVERROR_INVALIDDATA;
1507
    }
1508

    
1509
    s->linesize   = curframe->linesize[0];
1510
    s->uvlinesize = curframe->linesize[1];
1511

    
1512
    if (!s->edge_emu_buffer)
1513
        s->edge_emu_buffer = av_malloc(21*s->linesize);
1514

    
1515
    memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz));
1516

    
1517
    /* Zero macroblock structures for top/left prediction from outside the frame. */
1518
    memset(s->macroblocks, 0, (s->mb_width + s->mb_height*2)*sizeof(*s->macroblocks));
1519

    
1520
    // top edge of 127 for intra prediction
1521
    memset(s->top_border, 127, (s->mb_width+1)*sizeof(*s->top_border));
1522
    memset(s->ref_count, 0, sizeof(s->ref_count));
1523

    
1524
    for (mb_y = 0; mb_y < s->mb_height; mb_y++) {
1525
        VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)];
1526
        VP8Macroblock *mb = s->macroblocks + (s->mb_height - mb_y - 1)*2;
1527
        uint8_t *intra4x4 = s->intra4x4_pred_mode + 4*mb_y*s->b4_stride;
1528
        uint8_t *segment_map = s->segmentation_map + mb_y*s->mb_stride;
1529
        int mb_xy = mb_y * s->mb_stride;
1530
        uint8_t *dst[3] = {
1531
            curframe->data[0] + 16*mb_y*s->linesize,
1532
            curframe->data[1] +  8*mb_y*s->uvlinesize,
1533
            curframe->data[2] +  8*mb_y*s->uvlinesize
1534
        };
1535

    
1536
        memset(s->left_nnz, 0, sizeof(s->left_nnz));
1537

    
1538
        // left edge of 129 for intra prediction
1539
        if (!(avctx->flags & CODEC_FLAG_EMU_EDGE))
1540
            for (i = 0; i < 3; i++)
1541
                for (y = 0; y < 16>>!!i; y++)
1542
                    dst[i][y*curframe->linesize[i]-1] = 129;
1543
        if (mb_y)
1544
            memset(s->top_border, 129, sizeof(*s->top_border));
1545

    
1546
        for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb_xy++, mb++) {
1547
            uint8_t *intra4x4_mb = s->keyframe ? intra4x4 + 4*mb_x : s->intra4x4_pred_mode_mb;
1548
            uint8_t *segment_mb = segment_map+mb_x;
1549

    
1550
            /* Prefetch the current frame, 4 MBs ahead */
1551
            s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4);
1552
            s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2);
1553

    
1554
            decode_mb_mode(s, mb, mb_x, mb_y, intra4x4_mb, segment_mb);
1555

    
1556
            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_PREVIOUS);
1557

    
1558
            if (!mb->skip)
1559
                decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);
1560

    
1561
            if (mb->mode <= MODE_I4x4)
1562
                intra_predict(s, dst, mb, intra4x4_mb, mb_x, mb_y);
1563
            else
1564
                inter_predict(s, dst, mb, mb_x, mb_y);
1565

    
1566
            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN);
1567

    
1568
            if (!mb->skip) {
1569
                idct_mb(s, dst, mb);
1570
            } else {
1571
                AV_ZERO64(s->left_nnz);
1572
                AV_WN64(s->top_nnz[mb_x], 0);   // array of 9, so unaligned
1573

    
1574
                // Reset DC block predictors if they would exist if the mb had coefficients
1575
                if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
1576
                    s->left_nnz[8]      = 0;
1577
                    s->top_nnz[mb_x][8] = 0;
1578
                }
1579
            }
1580

    
1581
            if (s->deblock_filter)
1582
                filter_level_for_mb(s, mb, &s->filter_strength[mb_x]);
1583

    
1584
            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN2);
1585

    
1586
            dst[0] += 16;
1587
            dst[1] += 8;
1588
            dst[2] += 8;
1589
        }
1590
        if (s->deblock_filter) {
1591
            if (s->filter.simple)
1592
                filter_mb_row_simple(s, mb_y);
1593
            else
1594
                filter_mb_row(s, mb_y);
1595
        }
1596
    }
1597

    
1598
skip_decode:
1599
    // if future frames don't use the updated probabilities,
1600
    // reset them to the values we saved
1601
    if (!s->update_probabilities)
1602
        s->prob[0] = s->prob[1];
1603

    
1604
    // check if golden and altref are swapped
1605
    if (s->update_altref == VP56_FRAME_GOLDEN &&
1606
        s->update_golden == VP56_FRAME_GOLDEN2)
1607
        FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN], s->framep[VP56_FRAME_GOLDEN2]);
1608
    else {
1609
        if (s->update_altref != VP56_FRAME_NONE)
1610
            s->framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref];
1611

    
1612
        if (s->update_golden != VP56_FRAME_NONE)
1613
            s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden];
1614
    }
1615

    
1616
    if (s->update_last) // move cur->prev
1617
        s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT];
1618

    
1619
    // release no longer referenced frames
1620
    for (i = 0; i < 4; i++)
1621
        if (s->frames[i].data[0] &&
1622
            &s->frames[i] != s->framep[VP56_FRAME_CURRENT] &&
1623
            &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1624
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1625
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2])
1626
            avctx->release_buffer(avctx, &s->frames[i]);
1627

    
1628
    if (!s->invisible) {
1629
        *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT];
1630
        *data_size = sizeof(AVFrame);
1631
    }
1632

    
1633
    return avpkt->size;
1634
}
1635

    
1636
static av_cold int vp8_decode_init(AVCodecContext *avctx)
1637
{
1638
    VP8Context *s = avctx->priv_data;
1639

    
1640
    s->avctx = avctx;
1641
    avctx->pix_fmt = PIX_FMT_YUV420P;
1642

    
1643
    dsputil_init(&s->dsp, avctx);
1644
    ff_h264_pred_init(&s->hpc, CODEC_ID_VP8);
1645
    ff_vp8dsp_init(&s->vp8dsp);
1646

    
1647
    // intra pred needs edge emulation among other things
1648
    if (avctx->flags&CODEC_FLAG_EMU_EDGE) {
1649
        av_log(avctx, AV_LOG_ERROR, "Edge emulation not supported\n");
1650
        return AVERROR_PATCHWELCOME;
1651
    }
1652

    
1653
    return 0;
1654
}
1655

    
1656
static av_cold int vp8_decode_free(AVCodecContext *avctx)
1657
{
1658
    vp8_decode_flush(avctx);
1659
    return 0;
1660
}
1661

    
1662
AVCodec vp8_decoder = {
1663
    "vp8",
1664
    AVMEDIA_TYPE_VIDEO,
1665
    CODEC_ID_VP8,
1666
    sizeof(VP8Context),
1667
    vp8_decode_init,
1668
    NULL,
1669
    vp8_decode_free,
1670
    vp8_decode_frame,
1671
    CODEC_CAP_DR1,
1672
    .flush = vp8_decode_flush,
1673
    .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"),
1674
};