Statistics
| Branch: | Revision:

ffmpeg / libavcodec / vp8.c @ 23117d69

History | View | Annotate | Download (58.7 KB)

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

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

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

    
38
typedef struct {
39
    uint8_t skip;
40
    // todo: make it possible to check for at least (i4x4 or split_mv)
41
    // in one op. are others needed?
42
    uint8_t mode;
43
    uint8_t ref_frame;
44
    uint8_t partitioning;
45
    VP56mv mv;
46
    VP56mv bmv[16];
47
} VP8Macroblock;
48

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

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

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

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

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

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

    
91
    uint8_t *intra4x4_pred_mode_top;
92
    uint8_t intra4x4_pred_mode_left[4];
93
    uint8_t *segmentation_map;
94
    int b4_stride;
95

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

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

    
111
    /**
112
     * This is the index plus one of the last non-zero coeff
113
     * for each of the blocks in the current macroblock.
114
     * So, 0 -> no coeffs
115
     *     1 -> dc-only (special transform)
116
     *     2+-> full transform
117
     */
118
    DECLARE_ALIGNED(16, uint8_t, non_zero_count_cache)[6][4];
119
    DECLARE_ALIGNED(16, DCTELEM, block)[6][4][16];
120
    DECLARE_ALIGNED(16, DCTELEM, block_dc)[16];
121
    uint8_t intra4x4_pred_mode_mb[16];
122

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

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

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

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

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

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

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

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

    
184
    /**
185
     * These are all of the updatable probabilities for binary decisions.
186
     * They are only implictly reset on keyframes, making it quite likely
187
     * for an interframe to desync if a prior frame's header was corrupt
188
     * or missing outright!
189
     */
190
    struct {
191
        uint8_t segmentid[3];
192
        uint8_t mbskip;
193
        uint8_t intra;
194
        uint8_t last;
195
        uint8_t golden;
196
        uint8_t pred16x16[4];
197
        uint8_t pred8x8c[3];
198
        /* Padded to allow overreads */
199
        uint8_t token[4][17][3][NUM_DCT_TOKENS-1];
200
        uint8_t mvc[2][19];
201
    } prob[2];
202
} VP8Context;
203

    
204
static void vp8_decode_flush(AVCodecContext *avctx)
205
{
206
    VP8Context *s = avctx->priv_data;
207
    int i;
208

    
209
    for (i = 0; i < 4; i++)
210
        if (s->frames[i].data[0])
211
            avctx->release_buffer(avctx, &s->frames[i]);
212
    memset(s->framep, 0, sizeof(s->framep));
213

    
214
    av_freep(&s->macroblocks_base);
215
    av_freep(&s->filter_strength);
216
    av_freep(&s->intra4x4_pred_mode_top);
217
    av_freep(&s->top_nnz);
218
    av_freep(&s->edge_emu_buffer);
219
    av_freep(&s->top_border);
220
    av_freep(&s->segmentation_map);
221

    
222
    s->macroblocks        = NULL;
223
}
224

    
225
static int update_dimensions(VP8Context *s, int width, int height)
226
{
227
    if (avcodec_check_dimensions(s->avctx, width, height))
228
        return AVERROR_INVALIDDATA;
229

    
230
    vp8_decode_flush(s->avctx);
231

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

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

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

    
242
    s->macroblocks_base        = av_mallocz((s->mb_stride+s->mb_height*2+2)*sizeof(*s->macroblocks));
243
    s->filter_strength         = av_mallocz(s->mb_stride*sizeof(*s->filter_strength));
244
    s->intra4x4_pred_mode_top  = av_mallocz(s->b4_stride*4);
245
    s->top_nnz                 = av_mallocz(s->mb_width*sizeof(*s->top_nnz));
246
    s->top_border              = av_mallocz((s->mb_width+1)*sizeof(*s->top_border));
247
    s->segmentation_map        = av_mallocz(s->mb_stride*s->mb_height);
248

    
249
    if (!s->macroblocks_base || !s->filter_strength || !s->intra4x4_pred_mode_top ||
250
        !s->top_nnz || !s->top_border || !s->segmentation_map)
251
        return AVERROR(ENOMEM);
252

    
253
    s->macroblocks        = s->macroblocks_base + 1;
254

    
255
    return 0;
256
}
257

    
258
static void parse_segment_info(VP8Context *s)
259
{
260
    VP56RangeCoder *c = &s->c;
261
    int i;
262

    
263
    s->segmentation.update_map = vp8_rac_get(c);
264

    
265
    if (vp8_rac_get(c)) { // update segment feature data
266
        s->segmentation.absolute_vals = vp8_rac_get(c);
267

    
268
        for (i = 0; i < 4; i++)
269
            s->segmentation.base_quant[i]   = vp8_rac_get_sint(c, 7);
270

    
271
        for (i = 0; i < 4; i++)
272
            s->segmentation.filter_level[i] = vp8_rac_get_sint(c, 6);
273
    }
274
    if (s->segmentation.update_map)
275
        for (i = 0; i < 3; i++)
276
            s->prob->segmentid[i] = vp8_rac_get(c) ? vp8_rac_get_uint(c, 8) : 255;
277
}
278

    
279
static void update_lf_deltas(VP8Context *s)
280
{
281
    VP56RangeCoder *c = &s->c;
282
    int i;
283

    
284
    for (i = 0; i < 4; i++)
285
        s->lf_delta.ref[i]  = vp8_rac_get_sint(c, 6);
286

    
287
    for (i = 0; i < 4; i++)
288
        s->lf_delta.mode[i] = vp8_rac_get_sint(c, 6);
289
}
290

    
291
static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size)
292
{
293
    const uint8_t *sizes = buf;
294
    int i;
295

    
296
    s->num_coeff_partitions = 1 << vp8_rac_get_uint(&s->c, 2);
297

    
298
    buf      += 3*(s->num_coeff_partitions-1);
299
    buf_size -= 3*(s->num_coeff_partitions-1);
300
    if (buf_size < 0)
301
        return -1;
302

    
303
    for (i = 0; i < s->num_coeff_partitions-1; i++) {
304
        int size = AV_RL24(sizes + 3*i);
305
        if (buf_size - size < 0)
306
            return -1;
307

    
308
        vp56_init_range_decoder(&s->coeff_partition[i], buf, size);
309
        buf      += size;
310
        buf_size -= size;
311
    }
312
    vp56_init_range_decoder(&s->coeff_partition[i], buf, buf_size);
313

    
314
    return 0;
315
}
316

    
317
static void get_quants(VP8Context *s)
318
{
319
    VP56RangeCoder *c = &s->c;
320
    int i, base_qi;
321

    
322
    int yac_qi     = vp8_rac_get_uint(c, 7);
323
    int ydc_delta  = vp8_rac_get_sint(c, 4);
324
    int y2dc_delta = vp8_rac_get_sint(c, 4);
325
    int y2ac_delta = vp8_rac_get_sint(c, 4);
326
    int uvdc_delta = vp8_rac_get_sint(c, 4);
327
    int uvac_delta = vp8_rac_get_sint(c, 4);
328

    
329
    for (i = 0; i < 4; i++) {
330
        if (s->segmentation.enabled) {
331
            base_qi = s->segmentation.base_quant[i];
332
            if (!s->segmentation.absolute_vals)
333
                base_qi += yac_qi;
334
        } else
335
            base_qi = yac_qi;
336

    
337
        s->qmat[i].luma_qmul[0]    =       vp8_dc_qlookup[av_clip(base_qi + ydc_delta , 0, 127)];
338
        s->qmat[i].luma_qmul[1]    =       vp8_ac_qlookup[av_clip(base_qi             , 0, 127)];
339
        s->qmat[i].luma_dc_qmul[0] =   2 * vp8_dc_qlookup[av_clip(base_qi + y2dc_delta, 0, 127)];
340
        s->qmat[i].luma_dc_qmul[1] = 155 * vp8_ac_qlookup[av_clip(base_qi + y2ac_delta, 0, 127)] / 100;
341
        s->qmat[i].chroma_qmul[0]  =       vp8_dc_qlookup[av_clip(base_qi + uvdc_delta, 0, 127)];
342
        s->qmat[i].chroma_qmul[1]  =       vp8_ac_qlookup[av_clip(base_qi + uvac_delta, 0, 127)];
343

    
344
        s->qmat[i].luma_dc_qmul[1] = FFMAX(s->qmat[i].luma_dc_qmul[1], 8);
345
        s->qmat[i].chroma_qmul[0]  = FFMIN(s->qmat[i].chroma_qmul[0], 132);
346
    }
347
}
348

    
349
/**
350
 * Determine which buffers golden and altref should be updated with after this frame.
351
 * The spec isn't clear here, so I'm going by my understanding of what libvpx does
352
 *
353
 * Intra frames update all 3 references
354
 * Inter frames update VP56_FRAME_PREVIOUS if the update_last flag is set
355
 * If the update (golden|altref) flag is set, it's updated with the current frame
356
 *      if update_last is set, and VP56_FRAME_PREVIOUS otherwise.
357
 * If the flag is not set, the number read means:
358
 *      0: no update
359
 *      1: VP56_FRAME_PREVIOUS
360
 *      2: update golden with altref, or update altref with golden
361
 */
362
static VP56Frame ref_to_update(VP8Context *s, int update, VP56Frame ref)
363
{
364
    VP56RangeCoder *c = &s->c;
365

    
366
    if (update)
367
        return VP56_FRAME_CURRENT;
368

    
369
    switch (vp8_rac_get_uint(c, 2)) {
370
    case 1:
371
        return VP56_FRAME_PREVIOUS;
372
    case 2:
373
        return (ref == VP56_FRAME_GOLDEN) ? VP56_FRAME_GOLDEN2 : VP56_FRAME_GOLDEN;
374
    }
375
    return VP56_FRAME_NONE;
376
}
377

    
378
static void update_refs(VP8Context *s)
379
{
380
    VP56RangeCoder *c = &s->c;
381

    
382
    int update_golden = vp8_rac_get(c);
383
    int update_altref = vp8_rac_get(c);
384

    
385
    s->update_golden = ref_to_update(s, update_golden, VP56_FRAME_GOLDEN);
386
    s->update_altref = ref_to_update(s, update_altref, VP56_FRAME_GOLDEN2);
387
}
388

    
389
static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)
390
{
391
    VP56RangeCoder *c = &s->c;
392
    int header_size, hscale, vscale, i, j, k, l, m, ret;
393
    int width  = s->avctx->width;
394
    int height = s->avctx->height;
395

    
396
    s->keyframe  = !(buf[0] & 1);
397
    s->profile   =  (buf[0]>>1) & 7;
398
    s->invisible = !(buf[0] & 0x10);
399
    header_size  = AV_RL24(buf) >> 5;
400
    buf      += 3;
401
    buf_size -= 3;
402

    
403
    if (s->profile > 3)
404
        av_log(s->avctx, AV_LOG_WARNING, "Unknown profile %d\n", s->profile);
405

    
406
    if (!s->profile)
407
        memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_epel_pixels_tab, sizeof(s->put_pixels_tab));
408
    else    // profile 1-3 use bilinear, 4+ aren't defined so whatever
409
        memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_bilinear_pixels_tab, sizeof(s->put_pixels_tab));
410

    
411
    if (header_size > buf_size - 7*s->keyframe) {
412
        av_log(s->avctx, AV_LOG_ERROR, "Header size larger than data provided\n");
413
        return AVERROR_INVALIDDATA;
414
    }
415

    
416
    if (s->keyframe) {
417
        if (AV_RL24(buf) != 0x2a019d) {
418
            av_log(s->avctx, AV_LOG_ERROR, "Invalid start code 0x%x\n", AV_RL24(buf));
419
            return AVERROR_INVALIDDATA;
420
        }
421
        width  = AV_RL16(buf+3) & 0x3fff;
422
        height = AV_RL16(buf+5) & 0x3fff;
423
        hscale = buf[4] >> 6;
424
        vscale = buf[6] >> 6;
425
        buf      += 7;
426
        buf_size -= 7;
427

    
428
        if (hscale || vscale)
429
            av_log_missing_feature(s->avctx, "Upscaling", 1);
430

    
431
        s->update_golden = s->update_altref = VP56_FRAME_CURRENT;
432
        for (i = 0; i < 4; i++)
433
            for (j = 0; j < 16; j++)
434
                memcpy(s->prob->token[i][j], vp8_token_default_probs[i][vp8_coeff_band[j]],
435
                       sizeof(s->prob->token[i][j]));
436
        memcpy(s->prob->pred16x16, vp8_pred16x16_prob_inter, sizeof(s->prob->pred16x16));
437
        memcpy(s->prob->pred8x8c , vp8_pred8x8c_prob_inter , sizeof(s->prob->pred8x8c));
438
        memcpy(s->prob->mvc      , vp8_mv_default_prob     , sizeof(s->prob->mvc));
439
        memset(&s->segmentation, 0, sizeof(s->segmentation));
440
    }
441

    
442
    if (!s->macroblocks_base || /* first frame */
443
        width != s->avctx->width || height != s->avctx->height) {
444
        if ((ret = update_dimensions(s, width, height) < 0))
445
            return ret;
446
    }
447

    
448
    vp56_init_range_decoder(c, buf, header_size);
449
    buf      += header_size;
450
    buf_size -= header_size;
451

    
452
    if (s->keyframe) {
453
        if (vp8_rac_get(c))
454
            av_log(s->avctx, AV_LOG_WARNING, "Unspecified colorspace\n");
455
        vp8_rac_get(c); // whether we can skip clamping in dsp functions
456
    }
457

    
458
    if ((s->segmentation.enabled = vp8_rac_get(c)))
459
        parse_segment_info(s);
460
    else
461
        s->segmentation.update_map = 0; // FIXME: move this to some init function?
462

    
463
    s->filter.simple    = vp8_rac_get(c);
464
    s->filter.level     = vp8_rac_get_uint(c, 6);
465
    s->filter.sharpness = vp8_rac_get_uint(c, 3);
466

    
467
    if ((s->lf_delta.enabled = vp8_rac_get(c)))
468
        if (vp8_rac_get(c))
469
            update_lf_deltas(s);
470

    
471
    if (setup_partitions(s, buf, buf_size)) {
472
        av_log(s->avctx, AV_LOG_ERROR, "Invalid partitions\n");
473
        return AVERROR_INVALIDDATA;
474
    }
475

    
476
    get_quants(s);
477

    
478
    if (!s->keyframe) {
479
        update_refs(s);
480
        s->sign_bias[VP56_FRAME_GOLDEN]               = vp8_rac_get(c);
481
        s->sign_bias[VP56_FRAME_GOLDEN2 /* altref */] = vp8_rac_get(c);
482
    }
483

    
484
    // if we aren't saving this frame's probabilities for future frames,
485
    // make a copy of the current probabilities
486
    if (!(s->update_probabilities = vp8_rac_get(c)))
487
        s->prob[1] = s->prob[0];
488

    
489
    s->update_last = s->keyframe || vp8_rac_get(c);
490

    
491
    for (i = 0; i < 4; i++)
492
        for (j = 0; j < 8; j++)
493
            for (k = 0; k < 3; k++)
494
                for (l = 0; l < NUM_DCT_TOKENS-1; l++)
495
                    if (vp56_rac_get_prob_branchy(c, vp8_token_update_probs[i][j][k][l])) {
496
                        int prob = vp8_rac_get_uint(c, 8);
497
                        for (m = 0; m < 16; m++)
498
                            if (vp8_coeff_band[m] == j)
499
                                s->prob->token[i][m][k][l] = prob;
500
                    }
501

    
502
    if ((s->mbskip_enabled = vp8_rac_get(c)))
503
        s->prob->mbskip = vp8_rac_get_uint(c, 8);
504

    
505
    if (!s->keyframe) {
506
        s->prob->intra  = vp8_rac_get_uint(c, 8);
507
        s->prob->last   = vp8_rac_get_uint(c, 8);
508
        s->prob->golden = vp8_rac_get_uint(c, 8);
509

    
510
        if (vp8_rac_get(c))
511
            for (i = 0; i < 4; i++)
512
                s->prob->pred16x16[i] = vp8_rac_get_uint(c, 8);
513
        if (vp8_rac_get(c))
514
            for (i = 0; i < 3; i++)
515
                s->prob->pred8x8c[i]  = vp8_rac_get_uint(c, 8);
516

    
517
        // 17.2 MV probability update
518
        for (i = 0; i < 2; i++)
519
            for (j = 0; j < 19; j++)
520
                if (vp56_rac_get_prob_branchy(c, vp8_mv_update_prob[i][j]))
521
                    s->prob->mvc[i][j] = vp8_rac_get_nn(c);
522
    }
523

    
524
    return 0;
525
}
526

    
527
static av_always_inline
528
void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src, int mb_x, int mb_y)
529
{
530
#define MARGIN (16 << 2)
531
    dst->x = av_clip(src->x, -((mb_x << 6) + MARGIN),
532
                     ((s->mb_width  - 1 - mb_x) << 6) + MARGIN);
533
    dst->y = av_clip(src->y, -((mb_y << 6) + MARGIN),
534
                     ((s->mb_height - 1 - mb_y) << 6) + MARGIN);
535
}
536

    
537
static av_always_inline
538
void find_near_mvs(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
539
                   VP56mv near[2], VP56mv *best, uint8_t cnt[4])
540
{
541
    VP8Macroblock *mb_edge[3] = { mb + 2 /* top */,
542
                                  mb - 1 /* left */,
543
                                  mb + 1 /* top-left */ };
544
    enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT };
545
    VP56mv near_mv[4]  = {{ 0 }};
546
    enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV };
547
    int idx = CNT_ZERO;
548
    int best_idx = CNT_ZERO;
549
    int cur_sign_bias = s->sign_bias[mb->ref_frame];
550
    int *sign_bias = s->sign_bias;
551

    
552
    /* Process MB on top, left and top-left */
553
    #define MV_EDGE_CHECK(n)\
554
    {\
555
        VP8Macroblock *edge = mb_edge[n];\
556
        int edge_ref = edge->ref_frame;\
557
        if (edge_ref != VP56_FRAME_CURRENT) {\
558
            uint32_t mv = AV_RN32A(&edge->mv);\
559
            if (mv) {\
560
                if (cur_sign_bias != sign_bias[edge_ref]) {\
561
                    /* SWAR negate of the values in mv. */\
562
                    mv = ~mv;\
563
                    mv = ((mv&0x7fff7fff) + 0x00010001) ^ (mv&0x80008000);\
564
                }\
565
                if (!n || mv != AV_RN32A(&near_mv[idx]))\
566
                    AV_WN32A(&near_mv[++idx], mv);\
567
                cnt[idx]      += 1 + (n != 2);\
568
            } else\
569
                cnt[CNT_ZERO] += 1 + (n != 2);\
570
        }\
571
    }
572
    MV_EDGE_CHECK(0)
573
    MV_EDGE_CHECK(1)
574
    MV_EDGE_CHECK(2)
575

    
576
    /* If we have three distinct MVs, merge first and last if they're the same */
577
    if (cnt[CNT_SPLITMV] && AV_RN32A(&near_mv[1+EDGE_TOP]) == AV_RN32A(&near_mv[1+EDGE_TOPLEFT]))
578
        cnt[CNT_NEAREST] += 1;
579

    
580
    cnt[CNT_SPLITMV] = ((mb_edge[EDGE_LEFT]->mode   == VP8_MVMODE_SPLIT) +
581
                        (mb_edge[EDGE_TOP]->mode    == VP8_MVMODE_SPLIT)) * 2 +
582
                       (mb_edge[EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT);
583

    
584
    /* Swap near and nearest if necessary */
585
    if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) {
586
        FFSWAP(uint8_t,     cnt[CNT_NEAREST],     cnt[CNT_NEAR]);
587
        FFSWAP( VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]);
588
    }
589

    
590
    /* Choose the best mv out of 0,0 and the nearest mv */
591
    if (cnt[CNT_NEAREST] >= cnt[CNT_ZERO])
592
        best_idx = CNT_NEAREST;
593

    
594
    mb->mv  = near_mv[best_idx];
595
    near[0] = near_mv[CNT_NEAREST];
596
    near[1] = near_mv[CNT_NEAR];
597
}
598

    
599
/**
600
 * Motion vector coding, 17.1.
601
 */
602
static int read_mv_component(VP56RangeCoder *c, const uint8_t *p)
603
{
604
    int bit, x = 0;
605

    
606
    if (vp56_rac_get_prob_branchy(c, p[0])) {
607
        int i;
608

    
609
        for (i = 0; i < 3; i++)
610
            x += vp56_rac_get_prob(c, p[9 + i]) << i;
611
        for (i = 9; i > 3; i--)
612
            x += vp56_rac_get_prob(c, p[9 + i]) << i;
613
        if (!(x & 0xFFF0) || vp56_rac_get_prob(c, p[12]))
614
            x += 8;
615
    } else {
616
        // small_mvtree
617
        const uint8_t *ps = p+2;
618
        bit = vp56_rac_get_prob(c, *ps);
619
        ps += 1 + 3*bit;
620
        x  += 4*bit;
621
        bit = vp56_rac_get_prob(c, *ps);
622
        ps += 1 + bit;
623
        x  += 2*bit;
624
        x  += vp56_rac_get_prob(c, *ps);
625
    }
626

    
627
    return (x && vp56_rac_get_prob(c, p[1])) ? -x : x;
628
}
629

    
630
static av_always_inline
631
const uint8_t *get_submv_prob(uint32_t left, uint32_t top)
632
{
633
    if (left == top)
634
        return vp8_submv_prob[4-!!left];
635
    if (!top)
636
        return vp8_submv_prob[2];
637
    return vp8_submv_prob[1-!!left];
638
}
639

    
640
/**
641
 * Split motion vector prediction, 16.4.
642
 * @returns the number of motion vectors parsed (2, 4 or 16)
643
 */
644
static av_always_inline
645
int decode_splitmvs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb)
646
{
647
    int part_idx = mb->partitioning =
648
        vp8_rac_get_tree(c, vp8_mbsplit_tree, vp8_mbsplit_prob);
649
    int n, num = vp8_mbsplit_count[part_idx];
650
    VP8Macroblock *top_mb  = &mb[2];
651
    VP8Macroblock *left_mb = &mb[-1];
652
    const uint8_t *mbsplits_left = vp8_mbsplits[left_mb->partitioning],
653
                  *mbsplits_top = vp8_mbsplits[top_mb->partitioning],
654
                  *mbsplits_cur = vp8_mbsplits[part_idx],
655
                  *firstidx = vp8_mbfirstidx[part_idx];
656
    VP56mv *top_mv  = top_mb->bmv;
657
    VP56mv *left_mv = left_mb->bmv;
658
    VP56mv *cur_mv  = mb->bmv;
659

    
660
    for (n = 0; n < num; n++) {
661
        int k = firstidx[n];
662
        uint32_t left, above;
663
        const uint8_t *submv_prob;
664

    
665
        if (!(k & 3))
666
            left = AV_RN32A(&left_mv[mbsplits_left[k + 3]]);
667
        else
668
            left  = AV_RN32A(&cur_mv[mbsplits_cur[k - 1]]);
669
        if (k <= 3)
670
            above = AV_RN32A(&top_mv[mbsplits_top[k + 12]]);
671
        else
672
            above = AV_RN32A(&cur_mv[mbsplits_cur[k - 4]]);
673

    
674
        submv_prob = get_submv_prob(left, above);
675

    
676
        switch (vp8_rac_get_tree(c, vp8_submv_ref_tree, submv_prob)) {
677
        case VP8_SUBMVMODE_NEW4X4:
678
            mb->bmv[n].y = mb->mv.y + read_mv_component(c, s->prob->mvc[0]);
679
            mb->bmv[n].x = mb->mv.x + read_mv_component(c, s->prob->mvc[1]);
680
            break;
681
        case VP8_SUBMVMODE_ZERO4X4:
682
            AV_ZERO32(&mb->bmv[n]);
683
            break;
684
        case VP8_SUBMVMODE_LEFT4X4:
685
            AV_WN32A(&mb->bmv[n], left);
686
            break;
687
        case VP8_SUBMVMODE_TOP4X4:
688
            AV_WN32A(&mb->bmv[n], above);
689
            break;
690
        }
691
    }
692

    
693
    return num;
694
}
695

    
696
static av_always_inline
697
void decode_intra4x4_modes(VP8Context *s, VP56RangeCoder *c,
698
                           int mb_x, int keyframe)
699
{
700
    uint8_t *intra4x4 = s->intra4x4_pred_mode_mb;
701
    if (keyframe) {
702
        int x, y;
703
        uint8_t* const top = s->intra4x4_pred_mode_top + 4 * mb_x;
704
        uint8_t* const left = s->intra4x4_pred_mode_left;
705
        for (y = 0; y < 4; y++) {
706
            for (x = 0; x < 4; x++) {
707
                const uint8_t *ctx;
708
                ctx = vp8_pred4x4_prob_intra[top[x]][left[y]];
709
                *intra4x4 = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx);
710
                left[y] = top[x] = *intra4x4;
711
                intra4x4++;
712
            }
713
        }
714
    } else {
715
        int i;
716
        for (i = 0; i < 16; i++)
717
            intra4x4[i] = vp8_rac_get_tree(c, vp8_pred4x4_tree, vp8_pred4x4_prob_inter);
718
    }
719
}
720

    
721
static av_always_inline
722
void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, uint8_t *segment)
723
{
724
    VP56RangeCoder *c = &s->c;
725

    
726
    if (s->segmentation.update_map)
727
        *segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid);
728
    s->segment = *segment;
729

    
730
    mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0;
731

    
732
    if (s->keyframe) {
733
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra);
734

    
735
        if (mb->mode == MODE_I4x4) {
736
            decode_intra4x4_modes(s, c, mb_x, 1);
737
        } else {
738
            const uint32_t modes = vp8_pred4x4_mode[mb->mode] * 0x01010101u;
739
            AV_WN32A(s->intra4x4_pred_mode_top + 4 * mb_x, modes);
740
            AV_WN32A(s->intra4x4_pred_mode_left, modes);
741
        }
742

    
743
        s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra);
744
        mb->ref_frame = VP56_FRAME_CURRENT;
745
    } else if (vp56_rac_get_prob_branchy(c, s->prob->intra)) {
746
        VP56mv near[2], best;
747
        uint8_t cnt[4] = { 0 };
748

    
749
        // inter MB, 16.2
750
        if (vp56_rac_get_prob_branchy(c, s->prob->last))
751
            mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ?
752
                VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN;
753
        else
754
            mb->ref_frame = VP56_FRAME_PREVIOUS;
755
        s->ref_count[mb->ref_frame-1]++;
756

    
757
        // motion vectors, 16.3
758
        find_near_mvs(s, mb, mb_x, mb_y, near, &best, cnt);
759
        if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[0]][0])) {
760
            if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[1]][1])) {
761
                if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[2]][2])) {
762
                    if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[3]][3])) {
763
                        mb->mode = VP8_MVMODE_SPLIT;
764
                        clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y);
765
                        mb->mv = mb->bmv[decode_splitmvs(s, c, mb) - 1];
766
                    } else {
767
                        mb->mode = VP8_MVMODE_NEW;
768
                        clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y);
769
                        mb->mv.y += + read_mv_component(c, s->prob->mvc[0]);
770
                        mb->mv.x += + read_mv_component(c, s->prob->mvc[1]);
771
                    }
772
                } else {
773
                    mb->mode = VP8_MVMODE_NEAR;
774
                    clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y);
775
                }
776
            } else {
777
                mb->mode = VP8_MVMODE_NEAREST;
778
                clamp_mv(s, &mb->mv, &near[0], mb_x, mb_y);
779
            }
780
        } else {
781
            mb->mode = VP8_MVMODE_ZERO;
782
            AV_ZERO32(&mb->mv);
783
        }
784
        if (mb->mode != VP8_MVMODE_SPLIT) {
785
            mb->partitioning = VP8_SPLITMVMODE_NONE;
786
            mb->bmv[0] = mb->mv;
787
        }
788
    } else {
789
        // intra MB, 16.1
790
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16);
791

    
792
        if (mb->mode == MODE_I4x4)
793
            decode_intra4x4_modes(s, c, mb_x, 0);
794

    
795
        s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c);
796
        mb->ref_frame = VP56_FRAME_CURRENT;
797
        mb->partitioning = VP8_SPLITMVMODE_NONE;
798
        AV_ZERO32(&mb->bmv[0]);
799
    }
800
}
801

    
802
/**
803
 * @param c arithmetic bitstream reader context
804
 * @param block destination for block coefficients
805
 * @param probs probabilities to use when reading trees from the bitstream
806
 * @param i initial coeff index, 0 unless a separate DC block is coded
807
 * @param zero_nhood the initial prediction context for number of surrounding
808
 *                   all-zero blocks (only left/top, so 0-2)
809
 * @param qmul array holding the dc/ac dequant factor at position 0/1
810
 * @return 0 if no coeffs were decoded
811
 *         otherwise, the index of the last coeff decoded plus one
812
 */
813
static int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16],
814
                               uint8_t probs[8][3][NUM_DCT_TOKENS-1],
815
                               int i, int zero_nhood, int16_t qmul[2])
816
{
817
    uint8_t *token_prob = probs[i][zero_nhood];
818
    int nonzero = 0;
819
    int coeff;
820

    
821
    do {
822
        if (!vp56_rac_get_prob_branchy(c, token_prob[0]))   // DCT_EOB
823
            return nonzero;
824

    
825
skip_eob:
826
        if (!vp56_rac_get_prob_branchy(c, token_prob[1])) { // DCT_0
827
            if (++i == 16)
828
                return nonzero; // invalid input; blocks should end with EOB
829
            token_prob = probs[i][0];
830
            goto skip_eob;
831
        }
832

    
833
        if (!vp56_rac_get_prob_branchy(c, token_prob[2])) { // DCT_1
834
            coeff = 1;
835
            token_prob = probs[i+1][1];
836
        } else {
837
            if (!vp56_rac_get_prob_branchy(c, token_prob[3])) { // DCT 2,3,4
838
                coeff = vp56_rac_get_prob(c, token_prob[4]);
839
                if (coeff)
840
                    coeff += vp56_rac_get_prob(c, token_prob[5]);
841
                coeff += 2;
842
            } else {
843
                // DCT_CAT*
844
                if (!vp56_rac_get_prob_branchy(c, token_prob[6])) {
845
                    if (!vp56_rac_get_prob_branchy(c, token_prob[7])) { // DCT_CAT1
846
                        coeff  = 5 + vp56_rac_get_prob(c, vp8_dct_cat1_prob[0]);
847
                    } else {                                    // DCT_CAT2
848
                        coeff  = 7;
849
                        coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[0]) << 1;
850
                        coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[1]);
851
                    }
852
                } else {    // DCT_CAT3 and up
853
                    int a = vp56_rac_get_prob(c, token_prob[8]);
854
                    int b = vp56_rac_get_prob(c, token_prob[9+a]);
855
                    int cat = (a<<1) + b;
856
                    coeff  = 3 + (8<<cat);
857
                    coeff += vp8_rac_get_coeff(c, vp8_dct_cat_prob[cat]);
858
                }
859
            }
860
            token_prob = probs[i+1][2];
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
    int i, x, y, luma_start = 0, luma_ctx = 3;
876
    int nnz_pred, nnz, nnz_total = 0;
877
    int segment = s->segment;
878
    int block_dc = 0;
879

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

    
883
        // decode DC values and do hadamard
884
        nnz = decode_block_coeffs(c, s->block_dc, s->prob->token[1], 0, nnz_pred,
885
                                  s->qmat[segment].luma_dc_qmul);
886
        l_nnz[8] = t_nnz[8] = !!nnz;
887
        if (nnz) {
888
            nnz_total += nnz;
889
            block_dc = 1;
890
            if (nnz == 1)
891
                s->vp8dsp.vp8_luma_dc_wht_dc(s->block, s->block_dc);
892
            else
893
                s->vp8dsp.vp8_luma_dc_wht(s->block, s->block_dc);
894
        }
895
        luma_start = 1;
896
        luma_ctx = 0;
897
    }
898

    
899
    // luma blocks
900
    for (y = 0; y < 4; y++)
901
        for (x = 0; x < 4; x++) {
902
            nnz_pred = l_nnz[y] + t_nnz[x];
903
            nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start,
904
                                      nnz_pred, s->qmat[segment].luma_qmul);
905
            // nnz+block_dc may be one more than the actual last index, but we don't care
906
            s->non_zero_count_cache[y][x] = nnz + block_dc;
907
            t_nnz[x] = l_nnz[y] = !!nnz;
908
            nnz_total += nnz;
909
        }
910

    
911
    // chroma blocks
912
    // TODO: what to do about dimensions? 2nd dim for luma is x,
913
    // but for chroma it's (y<<1)|x
914
    for (i = 4; i < 6; i++)
915
        for (y = 0; y < 2; y++)
916
            for (x = 0; x < 2; x++) {
917
                nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x];
918
                nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0,
919
                                          nnz_pred, s->qmat[segment].chroma_qmul);
920
                s->non_zero_count_cache[i][(y<<1)+x] = nnz;
921
                t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz;
922
                nnz_total += nnz;
923
            }
924

    
925
    // if there were no coded coeffs despite the macroblock not being marked skip,
926
    // we MUST not do the inner loop filter and should not do IDCT
927
    // Since skip isn't used for bitstream prediction, just manually set it.
928
    if (!nnz_total)
929
        mb->skip = 1;
930
}
931

    
932
static av_always_inline
933
void backup_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
934
                      int linesize, int uvlinesize, int simple)
935
{
936
    AV_COPY128(top_border, src_y + 15*linesize);
937
    if (!simple) {
938
        AV_COPY64(top_border+16, src_cb + 7*uvlinesize);
939
        AV_COPY64(top_border+24, src_cr + 7*uvlinesize);
940
    }
941
}
942

    
943
static av_always_inline
944
void xchg_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
945
                    int linesize, int uvlinesize, int mb_x, int mb_y, int mb_width,
946
                    int simple, int xchg)
947
{
948
    uint8_t *top_border_m1 = top_border-32;     // for TL prediction
949
    src_y  -=   linesize;
950
    src_cb -= uvlinesize;
951
    src_cr -= uvlinesize;
952

    
953
#define XCHG(a,b,xchg) do {                     \
954
        if (xchg) AV_SWAP64(b,a);               \
955
        else      AV_COPY64(b,a);               \
956
    } while (0)
957

    
958
    XCHG(top_border_m1+8, src_y-8, xchg);
959
    XCHG(top_border,      src_y,   xchg);
960
    XCHG(top_border+8,    src_y+8, 1);
961
    if (mb_x < mb_width-1)
962
        XCHG(top_border+32, src_y+16, 1);
963

    
964
    // only copy chroma for normal loop filter
965
    // or to initialize the top row to 127
966
    if (!simple || !mb_y) {
967
        XCHG(top_border_m1+16, src_cb-8, xchg);
968
        XCHG(top_border_m1+24, src_cr-8, xchg);
969
        XCHG(top_border+16,    src_cb, 1);
970
        XCHG(top_border+24,    src_cr, 1);
971
    }
972
}
973

    
974
static av_always_inline
975
int check_intra_pred_mode(int mode, int mb_x, int mb_y)
976
{
977
    if (mode == DC_PRED8x8) {
978
        if (!mb_x) {
979
            mode = mb_y ? TOP_DC_PRED8x8 : DC_128_PRED8x8;
980
        } else if (!mb_y) {
981
            mode = LEFT_DC_PRED8x8;
982
        }
983
    }
984
    return mode;
985
}
986

    
987
static av_always_inline
988
void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
989
                   int mb_x, int mb_y)
990
{
991
    int x, y, mode, nnz, tr;
992

    
993
    // for the first row, we need to run xchg_mb_border to init the top edge to 127
994
    // otherwise, skip it if we aren't going to deblock
995
    if (s->deblock_filter || !mb_y)
996
        xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
997
                       s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
998
                       s->filter.simple, 1);
999

    
1000
    if (mb->mode < MODE_I4x4) {
1001
        mode = check_intra_pred_mode(mb->mode, mb_x, mb_y);
1002
        s->hpc.pred16x16[mode](dst[0], s->linesize);
1003
    } else {
1004
        uint8_t *ptr = dst[0];
1005
        uint8_t *intra4x4 = s->intra4x4_pred_mode_mb;
1006

    
1007
        // all blocks on the right edge of the macroblock use bottom edge
1008
        // the top macroblock for their topright edge
1009
        uint8_t *tr_right = ptr - s->linesize + 16;
1010

    
1011
        // if we're on the right edge of the frame, said edge is extended
1012
        // from the top macroblock
1013
        if (mb_x == s->mb_width-1) {
1014
            tr = tr_right[-1]*0x01010101;
1015
            tr_right = (uint8_t *)&tr;
1016
        }
1017

    
1018
        if (mb->skip)
1019
            AV_ZERO128(s->non_zero_count_cache);
1020

    
1021
        for (y = 0; y < 4; y++) {
1022
            uint8_t *topright = ptr + 4 - s->linesize;
1023
            for (x = 0; x < 4; x++) {
1024
                if (x == 3)
1025
                    topright = tr_right;
1026

    
1027
                s->hpc.pred4x4[intra4x4[x]](ptr+4*x, topright, s->linesize);
1028

    
1029
                nnz = s->non_zero_count_cache[y][x];
1030
                if (nnz) {
1031
                    if (nnz == 1)
1032
                        s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize);
1033
                    else
1034
                        s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize);
1035
                }
1036
                topright += 4;
1037
            }
1038

    
1039
            ptr   += 4*s->linesize;
1040
            intra4x4 += 4;
1041
        }
1042
    }
1043

    
1044
    mode = check_intra_pred_mode(s->chroma_pred_mode, mb_x, mb_y);
1045
    s->hpc.pred8x8[mode](dst[1], s->uvlinesize);
1046
    s->hpc.pred8x8[mode](dst[2], s->uvlinesize);
1047

    
1048
    if (s->deblock_filter || !mb_y)
1049
        xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
1050
                       s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
1051
                       s->filter.simple, 0);
1052
}
1053

    
1054
/**
1055
 * Generic MC function.
1056
 *
1057
 * @param s VP8 decoding context
1058
 * @param luma 1 for luma (Y) planes, 0 for chroma (Cb/Cr) planes
1059
 * @param dst target buffer for block data at block position
1060
 * @param src reference picture buffer at origin (0, 0)
1061
 * @param mv motion vector (relative to block position) to get pixel data from
1062
 * @param x_off horizontal position of block from origin (0, 0)
1063
 * @param y_off vertical position of block from origin (0, 0)
1064
 * @param block_w width of block (16, 8 or 4)
1065
 * @param block_h height of block (always same as block_w)
1066
 * @param width width of src/dst plane data
1067
 * @param height height of src/dst plane data
1068
 * @param linesize size of a single line of plane data, including padding
1069
 * @param mc_func motion compensation function pointers (bilinear or sixtap MC)
1070
 */
1071
static av_always_inline
1072
void vp8_mc(VP8Context *s, int luma,
1073
            uint8_t *dst, uint8_t *src, const VP56mv *mv,
1074
            int x_off, int y_off, int block_w, int block_h,
1075
            int width, int height, int linesize,
1076
            vp8_mc_func mc_func[3][3])
1077
{
1078
    if (AV_RN32A(mv)) {
1079
        static const uint8_t idx[8] = { 0, 1, 2, 1, 2, 1, 2, 1 };
1080
        int mx = (mv->x << luma)&7, mx_idx = idx[mx];
1081
        int my = (mv->y << luma)&7, my_idx = idx[my];
1082

    
1083
        x_off += mv->x >> (3 - luma);
1084
        y_off += mv->y >> (3 - luma);
1085

    
1086
        // edge emulation
1087
        src += y_off * linesize + x_off;
1088
        if (x_off < 2 || x_off >= width  - block_w - 3 ||
1089
            y_off < 2 || y_off >= height - block_h - 3) {
1090
            ff_emulated_edge_mc(s->edge_emu_buffer, src - 2 * linesize - 2, linesize,
1091
                                block_w + 5, block_h + 5,
1092
                                x_off - 2, y_off - 2, width, height);
1093
            src = s->edge_emu_buffer + 2 + linesize * 2;
1094
        }
1095
        mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);
1096
    } else
1097
        mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0);
1098
}
1099

    
1100
static av_always_inline
1101
void vp8_mc_part(VP8Context *s, uint8_t *dst[3],
1102
                 AVFrame *ref_frame, int x_off, int y_off,
1103
                 int bx_off, int by_off,
1104
                 int block_w, int block_h,
1105
                 int width, int height, VP56mv *mv)
1106
{
1107
    VP56mv uvmv = *mv;
1108

    
1109
    /* Y */
1110
    vp8_mc(s, 1, dst[0] + by_off * s->linesize + bx_off,
1111
           ref_frame->data[0], mv, x_off + bx_off, y_off + by_off,
1112
           block_w, block_h, width, height, s->linesize,
1113
           s->put_pixels_tab[block_w == 8]);
1114

    
1115
    /* U/V */
1116
    if (s->profile == 3) {
1117
        uvmv.x &= ~7;
1118
        uvmv.y &= ~7;
1119
    }
1120
    x_off   >>= 1; y_off   >>= 1;
1121
    bx_off  >>= 1; by_off  >>= 1;
1122
    width   >>= 1; height  >>= 1;
1123
    block_w >>= 1; block_h >>= 1;
1124
    vp8_mc(s, 0, dst[1] + by_off * s->uvlinesize + bx_off,
1125
           ref_frame->data[1], &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
    vp8_mc(s, 0, dst[2] + by_off * s->uvlinesize + bx_off,
1129
           ref_frame->data[2], &uvmv, x_off + bx_off, y_off + by_off,
1130
           block_w, block_h, width, height, s->uvlinesize,
1131
           s->put_pixels_tab[1 + (block_w == 4)]);
1132
}
1133

    
1134
/* Fetch pixels for estimated mv 4 macroblocks ahead.
1135
 * Optimized for 64-byte cache lines.  Inspired by ffh264 prefetch_motion. */
1136
static av_always_inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int mb_xy, int ref)
1137
{
1138
    /* Don't prefetch refs that haven't been used very often this frame. */
1139
    if (s->ref_count[ref-1] > (mb_xy >> 5)) {
1140
        int x_off = mb_x << 4, y_off = mb_y << 4;
1141
        int mx = mb->mv.x + x_off + 8;
1142
        int my = mb->mv.y + y_off;
1143
        uint8_t **src= s->framep[ref]->data;
1144
        int off= mx + (my + (mb_x&3)*4)*s->linesize + 64;
1145
        s->dsp.prefetch(src[0]+off, s->linesize, 4);
1146
        off= (mx>>1) + ((my>>1) + (mb_x&7))*s->uvlinesize + 64;
1147
        s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1148
    }
1149
}
1150

    
1151
/**
1152
 * Apply motion vectors to prediction buffer, chapter 18.
1153
 */
1154
static av_always_inline
1155
void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
1156
                   int mb_x, int mb_y)
1157
{
1158
    int x_off = mb_x << 4, y_off = mb_y << 4;
1159
    int width = 16*s->mb_width, height = 16*s->mb_height;
1160
    AVFrame *ref = s->framep[mb->ref_frame];
1161
    VP56mv *bmv = mb->bmv;
1162

    
1163
    if (mb->mode < VP8_MVMODE_SPLIT) {
1164
        vp8_mc_part(s, dst, ref, x_off, y_off,
1165
                    0, 0, 16, 16, width, height, &mb->mv);
1166
    } else switch (mb->partitioning) {
1167
    case VP8_SPLITMVMODE_4x4: {
1168
        int x, y;
1169
        VP56mv uvmv;
1170

    
1171
        /* Y */
1172
        for (y = 0; y < 4; y++) {
1173
            for (x = 0; x < 4; x++) {
1174
                vp8_mc(s, 1, dst[0] + 4*y*s->linesize + x*4,
1175
                       ref->data[0], &bmv[4*y + x],
1176
                       4*x + x_off, 4*y + y_off, 4, 4,
1177
                       width, height, s->linesize,
1178
                       s->put_pixels_tab[2]);
1179
            }
1180
        }
1181

    
1182
        /* U/V */
1183
        x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1;
1184
        for (y = 0; y < 2; y++) {
1185
            for (x = 0; x < 2; x++) {
1186
                uvmv.x = mb->bmv[ 2*y    * 4 + 2*x  ].x +
1187
                         mb->bmv[ 2*y    * 4 + 2*x+1].x +
1188
                         mb->bmv[(2*y+1) * 4 + 2*x  ].x +
1189
                         mb->bmv[(2*y+1) * 4 + 2*x+1].x;
1190
                uvmv.y = mb->bmv[ 2*y    * 4 + 2*x  ].y +
1191
                         mb->bmv[ 2*y    * 4 + 2*x+1].y +
1192
                         mb->bmv[(2*y+1) * 4 + 2*x  ].y +
1193
                         mb->bmv[(2*y+1) * 4 + 2*x+1].y;
1194
                uvmv.x = (uvmv.x + 2 + (uvmv.x >> (INT_BIT-1))) >> 2;
1195
                uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2;
1196
                if (s->profile == 3) {
1197
                    uvmv.x &= ~7;
1198
                    uvmv.y &= ~7;
1199
                }
1200
                vp8_mc(s, 0, dst[1] + 4*y*s->uvlinesize + x*4,
1201
                       ref->data[1], &uvmv,
1202
                       4*x + x_off, 4*y + y_off, 4, 4,
1203
                       width, height, s->uvlinesize,
1204
                       s->put_pixels_tab[2]);
1205
                vp8_mc(s, 0, dst[2] + 4*y*s->uvlinesize + x*4,
1206
                       ref->data[2], &uvmv,
1207
                       4*x + x_off, 4*y + y_off, 4, 4,
1208
                       width, height, s->uvlinesize,
1209
                       s->put_pixels_tab[2]);
1210
            }
1211
        }
1212
        break;
1213
    }
1214
    case VP8_SPLITMVMODE_16x8:
1215
        vp8_mc_part(s, dst, ref, x_off, y_off,
1216
                    0, 0, 16, 8, width, height, &bmv[0]);
1217
        vp8_mc_part(s, dst, ref, x_off, y_off,
1218
                    0, 8, 16, 8, width, height, &bmv[1]);
1219
        break;
1220
    case VP8_SPLITMVMODE_8x16:
1221
        vp8_mc_part(s, dst, ref, x_off, y_off,
1222
                    0, 0, 8, 16, width, height, &bmv[0]);
1223
        vp8_mc_part(s, dst, ref, x_off, y_off,
1224
                    8, 0, 8, 16, width, height, &bmv[1]);
1225
        break;
1226
    case VP8_SPLITMVMODE_8x8:
1227
        vp8_mc_part(s, dst, ref, x_off, y_off,
1228
                    0, 0, 8, 8, width, height, &bmv[0]);
1229
        vp8_mc_part(s, dst, ref, x_off, y_off,
1230
                    8, 0, 8, 8, width, height, &bmv[1]);
1231
        vp8_mc_part(s, dst, ref, x_off, y_off,
1232
                    0, 8, 8, 8, width, height, &bmv[2]);
1233
        vp8_mc_part(s, dst, ref, x_off, y_off,
1234
                    8, 8, 8, 8, width, height, &bmv[3]);
1235
        break;
1236
    }
1237
}
1238

    
1239
static av_always_inline void idct_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb)
1240
{
1241
    int x, y, ch;
1242

    
1243
    if (mb->mode != MODE_I4x4) {
1244
        uint8_t *y_dst = dst[0];
1245
        for (y = 0; y < 4; y++) {
1246
            uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[y]);
1247
            if (nnz4) {
1248
                if (nnz4&~0x01010101) {
1249
                    for (x = 0; x < 4; x++) {
1250
                        int nnz = s->non_zero_count_cache[y][x];
1251
                        if (nnz) {
1252
                            if (nnz == 1)
1253
                                s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize);
1254
                            else
1255
                                s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize);
1256
                        }
1257
                    }
1258
                } else {
1259
                    s->vp8dsp.vp8_idct_dc_add4y(y_dst, s->block[y], s->linesize);
1260
                }
1261
            }
1262
            y_dst += 4*s->linesize;
1263
        }
1264
    }
1265

    
1266
    for (ch = 0; ch < 2; ch++) {
1267
        uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[4+ch]);
1268
        if (nnz4) {
1269
            uint8_t *ch_dst = dst[1+ch];
1270
            if (nnz4&~0x01010101) {
1271
                for (y = 0; y < 2; y++) {
1272
                    for (x = 0; x < 2; x++) {
1273
                        int nnz = s->non_zero_count_cache[4+ch][(y<<1)+x];
1274
                        if (nnz) {
1275
                            if (nnz == 1)
1276
                                s->vp8dsp.vp8_idct_dc_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);
1277
                            else
1278
                                s->vp8dsp.vp8_idct_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);
1279
                        }
1280
                    }
1281
                    ch_dst += 4*s->uvlinesize;
1282
                }
1283
            } else {
1284
                s->vp8dsp.vp8_idct_dc_add4uv(ch_dst, s->block[4+ch], s->uvlinesize);
1285
            }
1286
        }
1287
    }
1288
}
1289

    
1290
static av_always_inline void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f )
1291
{
1292
    int interior_limit, filter_level;
1293

    
1294
    if (s->segmentation.enabled) {
1295
        filter_level = s->segmentation.filter_level[s->segment];
1296
        if (!s->segmentation.absolute_vals)
1297
            filter_level += s->filter.level;
1298
    } else
1299
        filter_level = s->filter.level;
1300

    
1301
    if (s->lf_delta.enabled) {
1302
        filter_level += s->lf_delta.ref[mb->ref_frame];
1303

    
1304
        if (mb->ref_frame == VP56_FRAME_CURRENT) {
1305
            if (mb->mode == MODE_I4x4)
1306
                filter_level += s->lf_delta.mode[0];
1307
        } else {
1308
            if (mb->mode == VP8_MVMODE_ZERO)
1309
                filter_level += s->lf_delta.mode[1];
1310
            else if (mb->mode == VP8_MVMODE_SPLIT)
1311
                filter_level += s->lf_delta.mode[3];
1312
            else
1313
                filter_level += s->lf_delta.mode[2];
1314
        }
1315
    }
1316
    filter_level = av_clip(filter_level, 0, 63);
1317

    
1318
    interior_limit = filter_level;
1319
    if (s->filter.sharpness) {
1320
        interior_limit >>= s->filter.sharpness > 4 ? 2 : 1;
1321
        interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness);
1322
    }
1323
    interior_limit = FFMAX(interior_limit, 1);
1324

    
1325
    f->filter_level = filter_level;
1326
    f->inner_limit = interior_limit;
1327
    f->inner_filter = !mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT;
1328
}
1329

    
1330
static av_always_inline void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f, int mb_x, int mb_y)
1331
{
1332
    int mbedge_lim, bedge_lim, hev_thresh;
1333
    int filter_level = f->filter_level;
1334
    int inner_limit = f->inner_limit;
1335
    int inner_filter = f->inner_filter;
1336
    int linesize = s->linesize;
1337
    int uvlinesize = s->uvlinesize;
1338

    
1339
    if (!filter_level)
1340
        return;
1341

    
1342
    mbedge_lim = 2*(filter_level+2) + inner_limit;
1343
     bedge_lim = 2* filter_level    + inner_limit;
1344
    hev_thresh = filter_level >= 15;
1345

    
1346
    if (s->keyframe) {
1347
        if (filter_level >= 40)
1348
            hev_thresh = 2;
1349
    } else {
1350
        if (filter_level >= 40)
1351
            hev_thresh = 3;
1352
        else if (filter_level >= 20)
1353
            hev_thresh = 2;
1354
    }
1355

    
1356
    if (mb_x) {
1357
        s->vp8dsp.vp8_h_loop_filter16y(dst[0],     linesize,
1358
                                       mbedge_lim, inner_limit, hev_thresh);
1359
        s->vp8dsp.vp8_h_loop_filter8uv(dst[1],     dst[2],      uvlinesize,
1360
                                       mbedge_lim, inner_limit, hev_thresh);
1361
    }
1362

    
1363
    if (inner_filter) {
1364
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 4, linesize, bedge_lim,
1365
                                             inner_limit, hev_thresh);
1366
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 8, linesize, bedge_lim,
1367
                                             inner_limit, hev_thresh);
1368
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+12, linesize, bedge_lim,
1369
                                             inner_limit, hev_thresh);
1370
        s->vp8dsp.vp8_h_loop_filter8uv_inner(dst[1] + 4, dst[2] + 4,
1371
                                             uvlinesize,  bedge_lim,
1372
                                             inner_limit, hev_thresh);
1373
    }
1374

    
1375
    if (mb_y) {
1376
        s->vp8dsp.vp8_v_loop_filter16y(dst[0],     linesize,
1377
                                       mbedge_lim, inner_limit, hev_thresh);
1378
        s->vp8dsp.vp8_v_loop_filter8uv(dst[1],     dst[2],      uvlinesize,
1379
                                       mbedge_lim, inner_limit, hev_thresh);
1380
    }
1381

    
1382
    if (inner_filter) {
1383
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 4*linesize,
1384
                                             linesize,    bedge_lim,
1385
                                             inner_limit, hev_thresh);
1386
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 8*linesize,
1387
                                             linesize,    bedge_lim,
1388
                                             inner_limit, hev_thresh);
1389
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+12*linesize,
1390
                                             linesize,    bedge_lim,
1391
                                             inner_limit, hev_thresh);
1392
        s->vp8dsp.vp8_v_loop_filter8uv_inner(dst[1] + 4 * uvlinesize,
1393
                                             dst[2] + 4 * uvlinesize,
1394
                                             uvlinesize,  bedge_lim,
1395
                                             inner_limit, hev_thresh);
1396
    }
1397
}
1398

    
1399
static av_always_inline void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f, int mb_x, int mb_y)
1400
{
1401
    int mbedge_lim, bedge_lim;
1402
    int filter_level = f->filter_level;
1403
    int inner_limit = f->inner_limit;
1404
    int inner_filter = f->inner_filter;
1405
    int linesize = s->linesize;
1406

    
1407
    if (!filter_level)
1408
        return;
1409

    
1410
    mbedge_lim = 2*(filter_level+2) + inner_limit;
1411
     bedge_lim = 2* filter_level    + inner_limit;
1412

    
1413
    if (mb_x)
1414
        s->vp8dsp.vp8_h_loop_filter_simple(dst, linesize, mbedge_lim);
1415
    if (inner_filter) {
1416
        s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, linesize, bedge_lim);
1417
        s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, linesize, bedge_lim);
1418
        s->vp8dsp.vp8_h_loop_filter_simple(dst+12, linesize, bedge_lim);
1419
    }
1420

    
1421
    if (mb_y)
1422
        s->vp8dsp.vp8_v_loop_filter_simple(dst, linesize, mbedge_lim);
1423
    if (inner_filter) {
1424
        s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*linesize, linesize, bedge_lim);
1425
        s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*linesize, linesize, bedge_lim);
1426
        s->vp8dsp.vp8_v_loop_filter_simple(dst+12*linesize, linesize, bedge_lim);
1427
    }
1428
}
1429

    
1430
static void filter_mb_row(VP8Context *s, int mb_y)
1431
{
1432
    VP8FilterStrength *f = s->filter_strength;
1433
    uint8_t *dst[3] = {
1434
        s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize,
1435
        s->framep[VP56_FRAME_CURRENT]->data[1] +  8*mb_y*s->uvlinesize,
1436
        s->framep[VP56_FRAME_CURRENT]->data[2] +  8*mb_y*s->uvlinesize
1437
    };
1438
    int mb_x;
1439

    
1440
    for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1441
        backup_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], s->linesize, s->uvlinesize, 0);
1442
        filter_mb(s, dst, f++, mb_x, mb_y);
1443
        dst[0] += 16;
1444
        dst[1] += 8;
1445
        dst[2] += 8;
1446
    }
1447
}
1448

    
1449
static void filter_mb_row_simple(VP8Context *s, int mb_y)
1450
{
1451
    VP8FilterStrength *f = s->filter_strength;
1452
    uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize;
1453
    int mb_x;
1454

    
1455
    for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1456
        backup_mb_border(s->top_border[mb_x+1], dst, NULL, NULL, s->linesize, 0, 1);
1457
        filter_mb_simple(s, dst, f++, mb_x, mb_y);
1458
        dst += 16;
1459
    }
1460
}
1461

    
1462
static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
1463
                            AVPacket *avpkt)
1464
{
1465
    VP8Context *s = avctx->priv_data;
1466
    int ret, mb_x, mb_y, i, y, referenced;
1467
    enum AVDiscard skip_thresh;
1468
    AVFrame *av_uninit(curframe);
1469

    
1470
    if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0)
1471
        return ret;
1472

    
1473
    referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT
1474
                                || s->update_altref == VP56_FRAME_CURRENT;
1475

    
1476
    skip_thresh = !referenced ? AVDISCARD_NONREF :
1477
                    !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL;
1478

    
1479
    if (avctx->skip_frame >= skip_thresh) {
1480
        s->invisible = 1;
1481
        goto skip_decode;
1482
    }
1483
    s->deblock_filter = s->filter.level && avctx->skip_loop_filter < skip_thresh;
1484

    
1485
    for (i = 0; i < 4; i++)
1486
        if (&s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1487
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1488
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) {
1489
            curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i];
1490
            break;
1491
        }
1492
    if (curframe->data[0])
1493
        avctx->release_buffer(avctx, curframe);
1494

    
1495
    curframe->key_frame = s->keyframe;
1496
    curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE;
1497
    curframe->reference = referenced ? 3 : 0;
1498
    if ((ret = avctx->get_buffer(avctx, curframe))) {
1499
        av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n");
1500
        return ret;
1501
    }
1502

    
1503
    // Given that arithmetic probabilities are updated every frame, it's quite likely
1504
    // that the values we have on a random interframe are complete junk if we didn't
1505
    // start decode on a keyframe. So just don't display anything rather than junk.
1506
    if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] ||
1507
                         !s->framep[VP56_FRAME_GOLDEN] ||
1508
                         !s->framep[VP56_FRAME_GOLDEN2])) {
1509
        av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n");
1510
        return AVERROR_INVALIDDATA;
1511
    }
1512

    
1513
    s->linesize   = curframe->linesize[0];
1514
    s->uvlinesize = curframe->linesize[1];
1515

    
1516
    if (!s->edge_emu_buffer)
1517
        s->edge_emu_buffer = av_malloc(21*s->linesize);
1518

    
1519
    memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz));
1520

    
1521
    /* Zero macroblock structures for top/left prediction from outside the frame. */
1522
    memset(s->macroblocks, 0, (s->mb_width + s->mb_height*2)*sizeof(*s->macroblocks));
1523

    
1524
    // top edge of 127 for intra prediction
1525
    memset(s->top_border, 127, (s->mb_width+1)*sizeof(*s->top_border));
1526
    memset(s->ref_count, 0, sizeof(s->ref_count));
1527
    if (s->keyframe)
1528
        memset(s->intra4x4_pred_mode_top, DC_PRED, s->b4_stride*4);
1529

    
1530
    for (mb_y = 0; mb_y < s->mb_height; mb_y++) {
1531
        VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)];
1532
        VP8Macroblock *mb = s->macroblocks + (s->mb_height - mb_y - 1)*2;
1533
        uint8_t *segment_map = s->segmentation_map + mb_y*s->mb_stride;
1534
        int mb_xy = mb_y * s->mb_stride;
1535
        uint8_t *dst[3] = {
1536
            curframe->data[0] + 16*mb_y*s->linesize,
1537
            curframe->data[1] +  8*mb_y*s->uvlinesize,
1538
            curframe->data[2] +  8*mb_y*s->uvlinesize
1539
        };
1540

    
1541
        memset(s->left_nnz, 0, sizeof(s->left_nnz));
1542
        AV_WN32A(s->intra4x4_pred_mode_left, DC_PRED*0x01010101);
1543

    
1544
        // left edge of 129 for intra prediction
1545
        if (!(avctx->flags & CODEC_FLAG_EMU_EDGE))
1546
            for (i = 0; i < 3; i++)
1547
                for (y = 0; y < 16>>!!i; y++)
1548
                    dst[i][y*curframe->linesize[i]-1] = 129;
1549
        if (mb_y)
1550
            memset(s->top_border, 129, sizeof(*s->top_border));
1551

    
1552
        for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb_xy++, mb++) {
1553
            uint8_t *segment_mb = segment_map+mb_x;
1554

    
1555
            /* Prefetch the current frame, 4 MBs ahead */
1556
            s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4);
1557
            s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2);
1558

    
1559
            decode_mb_mode(s, mb, mb_x, mb_y, segment_mb);
1560

    
1561
            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_PREVIOUS);
1562

    
1563
            if (!mb->skip)
1564
                decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);
1565

    
1566
            if (mb->mode <= MODE_I4x4)
1567
                intra_predict(s, dst, mb, mb_x, mb_y);
1568
            else
1569
                inter_predict(s, dst, mb, mb_x, mb_y);
1570

    
1571
            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN);
1572

    
1573
            if (!mb->skip) {
1574
                idct_mb(s, dst, mb);
1575
            } else {
1576
                AV_ZERO64(s->left_nnz);
1577
                AV_WN64(s->top_nnz[mb_x], 0);   // array of 9, so unaligned
1578

    
1579
                // Reset DC block predictors if they would exist if the mb had coefficients
1580
                if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
1581
                    s->left_nnz[8]      = 0;
1582
                    s->top_nnz[mb_x][8] = 0;
1583
                }
1584
            }
1585

    
1586
            if (s->deblock_filter)
1587
                filter_level_for_mb(s, mb, &s->filter_strength[mb_x]);
1588

    
1589
            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN2);
1590

    
1591
            dst[0] += 16;
1592
            dst[1] += 8;
1593
            dst[2] += 8;
1594
        }
1595
        if (s->deblock_filter) {
1596
            if (s->filter.simple)
1597
                filter_mb_row_simple(s, mb_y);
1598
            else
1599
                filter_mb_row(s, mb_y);
1600
        }
1601
    }
1602

    
1603
skip_decode:
1604
    // if future frames don't use the updated probabilities,
1605
    // reset them to the values we saved
1606
    if (!s->update_probabilities)
1607
        s->prob[0] = s->prob[1];
1608

    
1609
    // check if golden and altref are swapped
1610
    if (s->update_altref == VP56_FRAME_GOLDEN &&
1611
        s->update_golden == VP56_FRAME_GOLDEN2)
1612
        FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN], s->framep[VP56_FRAME_GOLDEN2]);
1613
    else {
1614
        if (s->update_altref != VP56_FRAME_NONE)
1615
            s->framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref];
1616

    
1617
        if (s->update_golden != VP56_FRAME_NONE)
1618
            s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden];
1619
    }
1620

    
1621
    if (s->update_last) // move cur->prev
1622
        s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT];
1623

    
1624
    // release no longer referenced frames
1625
    for (i = 0; i < 4; i++)
1626
        if (s->frames[i].data[0] &&
1627
            &s->frames[i] != s->framep[VP56_FRAME_CURRENT] &&
1628
            &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1629
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1630
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2])
1631
            avctx->release_buffer(avctx, &s->frames[i]);
1632

    
1633
    if (!s->invisible) {
1634
        *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT];
1635
        *data_size = sizeof(AVFrame);
1636
    }
1637

    
1638
    return avpkt->size;
1639
}
1640

    
1641
static av_cold int vp8_decode_init(AVCodecContext *avctx)
1642
{
1643
    VP8Context *s = avctx->priv_data;
1644

    
1645
    s->avctx = avctx;
1646
    avctx->pix_fmt = PIX_FMT_YUV420P;
1647

    
1648
    dsputil_init(&s->dsp, avctx);
1649
    ff_h264_pred_init(&s->hpc, CODEC_ID_VP8);
1650
    ff_vp8dsp_init(&s->vp8dsp);
1651

    
1652
    // intra pred needs edge emulation among other things
1653
    if (avctx->flags&CODEC_FLAG_EMU_EDGE) {
1654
        av_log(avctx, AV_LOG_ERROR, "Edge emulation not supported\n");
1655
        return AVERROR_PATCHWELCOME;
1656
    }
1657

    
1658
    return 0;
1659
}
1660

    
1661
static av_cold int vp8_decode_free(AVCodecContext *avctx)
1662
{
1663
    vp8_decode_flush(avctx);
1664
    return 0;
1665
}
1666

    
1667
AVCodec vp8_decoder = {
1668
    "vp8",
1669
    AVMEDIA_TYPE_VIDEO,
1670
    CODEC_ID_VP8,
1671
    sizeof(VP8Context),
1672
    vp8_decode_init,
1673
    NULL,
1674
    vp8_decode_free,
1675
    vp8_decode_frame,
1676
    CODEC_CAP_DR1,
1677
    .flush = vp8_decode_flush,
1678
    .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"),
1679
};