Statistics
| Branch: | Revision:

ffmpeg / libavcodec / vp8.c @ 5245c04d

History | View | Annotate | Download (51 KB)

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

    
24
#include "avcodec.h"
25
#include "vp56.h"
26
#include "vp8data.h"
27
#include "vp8dsp.h"
28
#include "h264pred.h"
29
#include "rectangle.h"
30

    
31
typedef struct {
32
    uint8_t segment;
33
    uint8_t skip;
34
    // todo: make it possible to check for at least (i4x4 or split_mv)
35
    // in one op. are others needed?
36
    uint8_t mode;
37
    uint8_t ref_frame;
38
    uint8_t partitioning;
39
    VP56mv mv;
40
    VP56mv bmv[16];
41
} VP8Macroblock;
42

    
43
typedef struct {
44
    AVCodecContext *avctx;
45
    DSPContext dsp;
46
    VP8DSPContext vp8dsp;
47
    H264PredContext hpc;
48
    vp8_mc_func put_pixels_tab[3][3][3];
49
    AVFrame frames[4];
50
    AVFrame *framep[4];
51
    uint8_t *edge_emu_buffer;
52
    VP56RangeCoder c;   ///< header context, includes mb modes and motion vectors
53
    int profile;
54

    
55
    int mb_width;   /* number of horizontal MB */
56
    int mb_height;  /* number of vertical MB */
57
    int linesize;
58
    int uvlinesize;
59

    
60
    int keyframe;
61
    int invisible;
62
    int update_last;    ///< update VP56_FRAME_PREVIOUS with the current one
63
    int update_golden;  ///< VP56_FRAME_NONE if not updated, or which frame to copy if so
64
    int update_altref;
65

    
66
    /**
67
     * If this flag is not set, all the probability updates
68
     * are discarded after this frame is decoded.
69
     */
70
    int update_probabilities;
71

    
72
    /**
73
     * All coefficients are contained in separate arith coding contexts.
74
     * There can be 1, 2, 4, or 8 of these after the header context.
75
     */
76
    int num_coeff_partitions;
77
    VP56RangeCoder coeff_partition[8];
78

    
79
    VP8Macroblock *macroblocks;
80
    VP8Macroblock *macroblocks_base;
81
    int mb_stride;
82

    
83
    uint8_t *intra4x4_pred_mode;
84
    uint8_t *intra4x4_pred_mode_base;
85
    int b4_stride;
86

    
87
    /**
88
     * For coeff decode, we need to know whether the above block had non-zero
89
     * coefficients. This means for each macroblock, we need data for 4 luma
90
     * blocks, 2 u blocks, 2 v blocks, and the luma dc block, for a total of 9
91
     * per macroblock. We keep the last row in top_nnz.
92
     */
93
    uint8_t (*top_nnz)[9];
94
    DECLARE_ALIGNED(8, uint8_t, left_nnz)[9];
95

    
96
    /**
97
     * This is the index plus one of the last non-zero coeff
98
     * for each of the blocks in the current macroblock.
99
     * So, 0 -> no coeffs
100
     *     1 -> dc-only (special transform)
101
     *     2+-> full transform
102
     */
103
    DECLARE_ALIGNED(16, uint8_t, non_zero_count_cache)[6][4];
104
    DECLARE_ALIGNED(16, DCTELEM, block)[6][4][16];
105

    
106
    int chroma_pred_mode;    ///< 8x8c pred mode of the current macroblock
107

    
108
    int mbskip_enabled;
109
    int sign_bias[4]; ///< one state [0, 1] per ref frame type
110

    
111
    /**
112
     * Base parameters for segmentation, i.e. per-macroblock parameters.
113
     * These must be kept unchanged even if segmentation is not used for
114
     * a frame, since the values persist between interframes.
115
     */
116
    struct {
117
        int enabled;
118
        int absolute_vals;
119
        int update_map;
120
        int8_t base_quant[4];
121
        int8_t filter_level[4];     ///< base loop filter level
122
    } segmentation;
123

    
124
    /**
125
     * Macroblocks can have one of 4 different quants in a frame when
126
     * segmentation is enabled.
127
     * If segmentation is disabled, only the first segment's values are used.
128
     */
129
    struct {
130
        // [0] - DC qmul  [1] - AC qmul
131
        int16_t luma_qmul[2];
132
        int16_t luma_dc_qmul[2];    ///< luma dc-only block quant
133
        int16_t chroma_qmul[2];
134
    } qmat[4];
135

    
136
    struct {
137
        int simple;
138
        int level;
139
        int sharpness;
140
    } filter;
141

    
142
    struct {
143
        int enabled;    ///< whether each mb can have a different strength based on mode/ref
144

    
145
        /**
146
         * filter strength adjustment for the following macroblock modes:
147
         * [0] - i4x4
148
         * [1] - zero mv
149
         * [2] - inter modes except for zero or split mv
150
         * [3] - split mv
151
         *  i16x16 modes never have any adjustment
152
         */
153
        int8_t mode[4];
154

    
155
        /**
156
         * filter strength adjustment for macroblocks that reference:
157
         * [0] - intra / VP56_FRAME_CURRENT
158
         * [1] - VP56_FRAME_PREVIOUS
159
         * [2] - VP56_FRAME_GOLDEN
160
         * [3] - altref / VP56_FRAME_GOLDEN2
161
         */
162
        int8_t ref[4];
163
    } lf_delta;
164

    
165
    /**
166
     * These are all of the updatable probabilities for binary decisions.
167
     * They are only implictly reset on keyframes, making it quite likely
168
     * for an interframe to desync if a prior frame's header was corrupt
169
     * or missing outright!
170
     */
171
    struct {
172
        uint8_t segmentid[3];
173
        uint8_t mbskip;
174
        uint8_t intra;
175
        uint8_t last;
176
        uint8_t golden;
177
        uint8_t pred16x16[4];
178
        uint8_t pred8x8c[3];
179
        uint8_t token[4][8][3][NUM_DCT_TOKENS-1];
180
        uint8_t mvc[2][19];
181
    } prob[2];
182
} VP8Context;
183

    
184
#define RL24(p) (AV_RL16(p) + ((p)[2] << 16))
185

    
186
static void vp8_decode_flush(AVCodecContext *avctx)
187
{
188
    VP8Context *s = avctx->priv_data;
189
    int i;
190

    
191
    for (i = 0; i < 4; i++)
192
        if (s->frames[i].data[0])
193
            avctx->release_buffer(avctx, &s->frames[i]);
194
    memset(s->framep, 0, sizeof(s->framep));
195

    
196
    av_freep(&s->macroblocks_base);
197
    av_freep(&s->intra4x4_pred_mode_base);
198
    av_freep(&s->top_nnz);
199
    av_freep(&s->edge_emu_buffer);
200

    
201
    s->macroblocks        = NULL;
202
    s->intra4x4_pred_mode = NULL;
203
}
204

    
205
static int update_dimensions(VP8Context *s, int width, int height)
206
{
207
    int i;
208

    
209
    if (avcodec_check_dimensions(s->avctx, width, height))
210
        return AVERROR_INVALIDDATA;
211

    
212
    vp8_decode_flush(s->avctx);
213

    
214
    avcodec_set_dimensions(s->avctx, width, height);
215

    
216
    s->mb_width  = (s->avctx->coded_width +15) / 16;
217
    s->mb_height = (s->avctx->coded_height+15) / 16;
218

    
219
    // we allocate a border around the top/left of intra4x4 modes
220
    // this is 4 blocks for intra4x4 to keep 4-byte alignment for fill_rectangle
221
    s->mb_stride = s->mb_width+1;
222
    s->b4_stride = 4*s->mb_stride;
223

    
224
    s->macroblocks_base        = av_mallocz(s->mb_stride*(s->mb_height+1)*sizeof(*s->macroblocks));
225
    s->intra4x4_pred_mode_base = av_mallocz(s->b4_stride*(4*s->mb_height+1));
226
    s->top_nnz                 = av_mallocz(s->mb_width*sizeof(*s->top_nnz));
227

    
228
    s->macroblocks        = s->macroblocks_base        + 1 + s->mb_stride;
229
    s->intra4x4_pred_mode = s->intra4x4_pred_mode_base + 4 + s->b4_stride;
230

    
231
    memset(s->intra4x4_pred_mode_base, DC_PRED, s->b4_stride);
232
    for (i = 0; i < 4*s->mb_height; i++)
233
        s->intra4x4_pred_mode[i*s->b4_stride-1] = DC_PRED;
234

    
235
    return 0;
236
}
237

    
238
static void parse_segment_info(VP8Context *s)
239
{
240
    VP56RangeCoder *c = &s->c;
241
    int i;
242

    
243
    s->segmentation.update_map = vp8_rac_get(c);
244

    
245
    if (vp8_rac_get(c)) { // update segment feature data
246
        s->segmentation.absolute_vals = vp8_rac_get(c);
247

    
248
        for (i = 0; i < 4; i++)
249
            s->segmentation.base_quant[i]   = vp8_rac_get_sint(c, 7);
250

    
251
        for (i = 0; i < 4; i++)
252
            s->segmentation.filter_level[i] = vp8_rac_get_sint(c, 6);
253
    }
254
    if (s->segmentation.update_map)
255
        for (i = 0; i < 3; i++)
256
            s->prob->segmentid[i] = vp8_rac_get(c) ? vp8_rac_get_uint(c, 8) : 255;
257
}
258

    
259
static void update_lf_deltas(VP8Context *s)
260
{
261
    VP56RangeCoder *c = &s->c;
262
    int i;
263

    
264
    for (i = 0; i < 4; i++)
265
        s->lf_delta.ref[i]  = vp8_rac_get_sint(c, 6);
266

    
267
    for (i = 0; i < 4; i++)
268
        s->lf_delta.mode[i] = vp8_rac_get_sint(c, 6);
269
}
270

    
271
static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size)
272
{
273
    const uint8_t *sizes = buf;
274
    int i;
275

    
276
    s->num_coeff_partitions = 1 << vp8_rac_get_uint(&s->c, 2);
277

    
278
    buf      += 3*(s->num_coeff_partitions-1);
279
    buf_size -= 3*(s->num_coeff_partitions-1);
280
    if (buf_size < 0)
281
        return -1;
282

    
283
    for (i = 0; i < s->num_coeff_partitions-1; i++) {
284
        int size = RL24(sizes + 3*i);
285
        if (buf_size - size < 0)
286
            return -1;
287

    
288
        vp56_init_range_decoder(&s->coeff_partition[i], buf, size);
289
        buf      += size;
290
        buf_size -= size;
291
    }
292
    vp56_init_range_decoder(&s->coeff_partition[i], buf, buf_size);
293

    
294
    return 0;
295
}
296

    
297
static void get_quants(VP8Context *s)
298
{
299
    VP56RangeCoder *c = &s->c;
300
    int i, base_qi;
301

    
302
    int yac_qi     = vp8_rac_get_uint(c, 7);
303
    int ydc_delta  = vp8_rac_get_sint(c, 4);
304
    int y2dc_delta = vp8_rac_get_sint(c, 4);
305
    int y2ac_delta = vp8_rac_get_sint(c, 4);
306
    int uvdc_delta = vp8_rac_get_sint(c, 4);
307
    int uvac_delta = vp8_rac_get_sint(c, 4);
308

    
309
    for (i = 0; i < 4; i++) {
310
        if (s->segmentation.enabled) {
311
            base_qi = s->segmentation.base_quant[i];
312
            if (!s->segmentation.absolute_vals)
313
                base_qi += yac_qi;
314
        } else
315
            base_qi = yac_qi;
316

    
317
        s->qmat[i].luma_qmul[0]    =       vp8_dc_qlookup[av_clip(base_qi + ydc_delta , 0, 127)];
318
        s->qmat[i].luma_qmul[1]    =       vp8_ac_qlookup[av_clip(base_qi             , 0, 127)];
319
        s->qmat[i].luma_dc_qmul[0] =   2 * vp8_dc_qlookup[av_clip(base_qi + y2dc_delta, 0, 127)];
320
        s->qmat[i].luma_dc_qmul[1] = 155 * vp8_ac_qlookup[av_clip(base_qi + y2ac_delta, 0, 127)] / 100;
321
        s->qmat[i].chroma_qmul[0]  =       vp8_dc_qlookup[av_clip(base_qi + uvdc_delta, 0, 127)];
322
        s->qmat[i].chroma_qmul[1]  =       vp8_ac_qlookup[av_clip(base_qi + uvac_delta, 0, 127)];
323

    
324
        s->qmat[i].luma_dc_qmul[1] = FFMAX(s->qmat[i].luma_dc_qmul[1], 8);
325
        s->qmat[i].chroma_qmul[0]  = FFMIN(s->qmat[i].chroma_qmul[0], 132);
326
    }
327
}
328

    
329
/**
330
 * Determine which buffers golden and altref should be updated with after this frame.
331
 * The spec isn't clear here, so I'm going by my understanding of what libvpx does
332
 *
333
 * Intra frames update all 3 references
334
 * Inter frames update VP56_FRAME_PREVIOUS if the update_last flag is set
335
 * If the update (golden|altref) flag is set, it's updated with the current frame
336
 *      if update_last is set, and VP56_FRAME_PREVIOUS otherwise.
337
 * If the flag is not set, the number read means:
338
 *      0: no update
339
 *      1: VP56_FRAME_PREVIOUS
340
 *      2: update golden with altref, or update altref with golden
341
 */
342
static VP56Frame ref_to_update(VP8Context *s, int update, VP56Frame ref)
343
{
344
    VP56RangeCoder *c = &s->c;
345

    
346
    if (update)
347
        return VP56_FRAME_CURRENT;
348

    
349
    switch (vp8_rac_get_uint(c, 2)) {
350
    case 1:
351
        return VP56_FRAME_PREVIOUS;
352
    case 2:
353
        return (ref == VP56_FRAME_GOLDEN) ? VP56_FRAME_GOLDEN2 : VP56_FRAME_GOLDEN;
354
    }
355
    return VP56_FRAME_NONE;
356
}
357

    
358
static void update_refs(VP8Context *s)
359
{
360
    VP56RangeCoder *c = &s->c;
361

    
362
    int update_golden = vp8_rac_get(c);
363
    int update_altref = vp8_rac_get(c);
364

    
365
    s->update_golden = ref_to_update(s, update_golden, VP56_FRAME_GOLDEN);
366
    s->update_altref = ref_to_update(s, update_altref, VP56_FRAME_GOLDEN2);
367
}
368

    
369
static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)
370
{
371
    VP56RangeCoder *c = &s->c;
372
    int header_size, hscale, vscale, i, j, k, l, ret;
373
    int width  = s->avctx->width;
374
    int height = s->avctx->height;
375

    
376
    s->keyframe  = !(buf[0] & 1);
377
    s->profile   =  (buf[0]>>1) & 7;
378
    s->invisible = !(buf[0] & 0x10);
379
    header_size  = RL24(buf) >> 5;
380
    buf      += 3;
381
    buf_size -= 3;
382

    
383
    if (s->profile > 3)
384
        av_log(s->avctx, AV_LOG_WARNING, "Unknown profile %d\n", s->profile);
385

    
386
    if (!s->profile)
387
        memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_epel_pixels_tab, sizeof(s->put_pixels_tab));
388
    else    // profile 1-3 use bilinear, 4+ aren't defined so whatever
389
        memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_bilinear_pixels_tab, sizeof(s->put_pixels_tab));
390

    
391
    if (header_size > buf_size - 7*s->keyframe) {
392
        av_log(s->avctx, AV_LOG_ERROR, "Header size larger than data provided\n");
393
        return AVERROR_INVALIDDATA;
394
    }
395

    
396
    if (s->keyframe) {
397
        if (RL24(buf) != 0x2a019d) {
398
            av_log(s->avctx, AV_LOG_ERROR, "Invalid start code 0x%x\n", RL24(buf));
399
            return AVERROR_INVALIDDATA;
400
        }
401
        width  = AV_RL16(buf+3) & 0x3fff;
402
        height = AV_RL16(buf+5) & 0x3fff;
403
        hscale = buf[4] >> 6;
404
        vscale = buf[6] >> 6;
405
        buf      += 7;
406
        buf_size -= 7;
407

    
408
        if (hscale || vscale)
409
            av_log_missing_feature(s->avctx, "Upscaling", 1);
410

    
411
        s->update_golden = s->update_altref = VP56_FRAME_CURRENT;
412
        memcpy(s->prob->token    , vp8_token_default_probs , sizeof(s->prob->token));
413
        memcpy(s->prob->pred16x16, vp8_pred16x16_prob_inter, sizeof(s->prob->pred16x16));
414
        memcpy(s->prob->pred8x8c , vp8_pred8x8c_prob_inter , sizeof(s->prob->pred8x8c));
415
        memcpy(s->prob->mvc      , vp8_mv_default_prob     , sizeof(s->prob->mvc));
416
        memset(&s->segmentation, 0, sizeof(s->segmentation));
417
    }
418

    
419
    if (!s->macroblocks_base || /* first frame */
420
        width != s->avctx->width || height != s->avctx->height) {
421
        if ((ret = update_dimensions(s, width, height) < 0))
422
            return ret;
423
    }
424

    
425
    vp56_init_range_decoder(c, buf, header_size);
426
    buf      += header_size;
427
    buf_size -= header_size;
428

    
429
    if (s->keyframe) {
430
        if (vp8_rac_get(c))
431
            av_log(s->avctx, AV_LOG_WARNING, "Unspecified colorspace\n");
432
        vp8_rac_get(c); // whether we can skip clamping in dsp functions
433
    }
434

    
435
    if ((s->segmentation.enabled = vp8_rac_get(c)))
436
        parse_segment_info(s);
437
    else
438
        s->segmentation.update_map = 0; // FIXME: move this to some init function?
439

    
440
    s->filter.simple    = vp8_rac_get(c);
441
    s->filter.level     = vp8_rac_get_uint(c, 6);
442
    s->filter.sharpness = vp8_rac_get_uint(c, 3);
443

    
444
    if ((s->lf_delta.enabled = vp8_rac_get(c)))
445
        if (vp8_rac_get(c))
446
            update_lf_deltas(s);
447

    
448
    if (setup_partitions(s, buf, buf_size)) {
449
        av_log(s->avctx, AV_LOG_ERROR, "Invalid partitions\n");
450
        return AVERROR_INVALIDDATA;
451
    }
452

    
453
    get_quants(s);
454

    
455
    if (!s->keyframe) {
456
        update_refs(s);
457
        s->sign_bias[VP56_FRAME_GOLDEN]               = vp8_rac_get(c);
458
        s->sign_bias[VP56_FRAME_GOLDEN2 /* altref */] = vp8_rac_get(c);
459
    }
460

    
461
    // if we aren't saving this frame's probabilities for future frames,
462
    // make a copy of the current probabilities
463
    if (!(s->update_probabilities = vp8_rac_get(c)))
464
        s->prob[1] = s->prob[0];
465

    
466
    s->update_last = s->keyframe || vp8_rac_get(c);
467

    
468
    for (i = 0; i < 4; i++)
469
        for (j = 0; j < 8; j++)
470
            for (k = 0; k < 3; k++)
471
                for (l = 0; l < NUM_DCT_TOKENS-1; l++)
472
                    if (vp56_rac_get_prob(c, vp8_token_update_probs[i][j][k][l]))
473
                        s->prob->token[i][j][k][l] = vp8_rac_get_uint(c, 8);
474

    
475
    if ((s->mbskip_enabled = vp8_rac_get(c)))
476
        s->prob->mbskip = vp8_rac_get_uint(c, 8);
477

    
478
    if (!s->keyframe) {
479
        s->prob->intra  = vp8_rac_get_uint(c, 8);
480
        s->prob->last   = vp8_rac_get_uint(c, 8);
481
        s->prob->golden = vp8_rac_get_uint(c, 8);
482

    
483
        if (vp8_rac_get(c))
484
            for (i = 0; i < 4; i++)
485
                s->prob->pred16x16[i] = vp8_rac_get_uint(c, 8);
486
        if (vp8_rac_get(c))
487
            for (i = 0; i < 3; i++)
488
                s->prob->pred8x8c[i]  = vp8_rac_get_uint(c, 8);
489

    
490
        // 17.2 MV probability update
491
        for (i = 0; i < 2; i++)
492
            for (j = 0; j < 19; j++)
493
                if (vp56_rac_get_prob(c, vp8_mv_update_prob[i][j]))
494
                    s->prob->mvc[i][j] = vp8_rac_get_nn(c);
495
    }
496

    
497
    return 0;
498
}
499

    
500
static inline void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src,
501
                            int mb_x, int mb_y)
502
{
503
#define MARGIN (16 << 2)
504
    dst->x = av_clip(src->x, -((mb_x << 6) + MARGIN),
505
                     ((s->mb_width  - 1 - mb_x) << 6) + MARGIN);
506
    dst->y = av_clip(src->y, -((mb_y << 6) + MARGIN),
507
                     ((s->mb_height - 1 - mb_y) << 6) + MARGIN);
508
}
509

    
510
static void find_near_mvs(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
511
                          VP56mv near[2], VP56mv *best, int cnt[4])
512
{
513
    VP8Macroblock *mb_edge[3] = { mb - s->mb_stride     /* top */,
514
                                  mb - 1                /* left */,
515
                                  mb - s->mb_stride - 1 /* top-left */ };
516
    enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT };
517
    VP56mv near_mv[4]  = {{ 0 }};
518
    enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV };
519
    int idx = CNT_ZERO, n;
520
    int best_idx = CNT_ZERO;
521

    
522
    /* Process MB on top, left and top-left */
523
    for (n = 0; n < 3; n++) {
524
        VP8Macroblock *edge = mb_edge[n];
525
        if (edge->ref_frame != VP56_FRAME_CURRENT) {
526
            if (edge->mv.x | edge->mv.y) {
527
                VP56mv tmp = edge->mv;
528
                if (s->sign_bias[mb->ref_frame] != s->sign_bias[edge->ref_frame]) {
529
                    tmp.x *= -1;
530
                    tmp.y *= -1;
531
                }
532
                if ((tmp.x ^ near_mv[idx].x) | (tmp.y ^ near_mv[idx].y))
533
                    near_mv[++idx] = tmp;
534
                cnt[idx]       += 1 + (n != 2);
535
            } else
536
                cnt[CNT_ZERO] += 1 + (n != 2);
537
        }
538
    }
539

    
540
    /* If we have three distinct MV's, merge first and last if they're the same */
541
    if (cnt[CNT_SPLITMV] &&
542
        !((near_mv[1+EDGE_TOP].x ^ near_mv[1+EDGE_TOPLEFT].x) |
543
          (near_mv[1+EDGE_TOP].y ^ near_mv[1+EDGE_TOPLEFT].y)))
544
        cnt[CNT_NEAREST] += 1;
545

    
546
    cnt[CNT_SPLITMV] = ((mb_edge[EDGE_LEFT]->mode   == VP8_MVMODE_SPLIT) +
547
                        (mb_edge[EDGE_TOP]->mode    == VP8_MVMODE_SPLIT)) * 2 +
548
                       (mb_edge[EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT);
549

    
550
    /* Swap near and nearest if necessary */
551
    if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) {
552
        FFSWAP(int,    cnt[CNT_NEAREST],     cnt[CNT_NEAR]);
553
        FFSWAP(VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]);
554
    }
555

    
556
    /* Choose the best mv out of 0,0 and the nearest mv */
557
    if (cnt[CNT_NEAREST] >= cnt[CNT_ZERO])
558
        best_idx = CNT_NEAREST;
559

    
560
    clamp_mv(s, best, &near_mv[best_idx], mb_x, mb_y);
561
    near[0] = near_mv[CNT_NEAREST];
562
    near[1] = near_mv[CNT_NEAR];
563
}
564

    
565
/**
566
 * Motion vector coding, 17.1.
567
 */
568
static int read_mv_component(VP56RangeCoder *c, const uint8_t *p)
569
{
570
    int x = 0;
571

    
572
    if (vp56_rac_get_prob(c, p[0])) {
573
        int i;
574

    
575
        for (i = 0; i < 3; i++)
576
            x += vp56_rac_get_prob(c, p[9 + i]) << i;
577
        for (i = 9; i > 3; i--)
578
            x += vp56_rac_get_prob(c, p[9 + i]) << i;
579
        if (!(x & 0xFFF0) || vp56_rac_get_prob(c, p[12]))
580
            x += 8;
581
    } else
582
        x = vp8_rac_get_tree(c, vp8_small_mvtree, &p[2]);
583

    
584
    return (x && vp56_rac_get_prob(c, p[1])) ? -x : x;
585
}
586

    
587
static const uint8_t *get_submv_prob(const VP56mv *left, const VP56mv *top)
588
{
589
    int l_is_zero = !(left->x | left->y);
590
    int t_is_zero = !(top->x  | top->y);
591
    int equal = !((left->x ^ top->x) | (left->y ^ top->y));
592

    
593
    if (equal)
594
        return l_is_zero ? vp8_submv_prob[4] : vp8_submv_prob[3];
595
    if (t_is_zero)
596
        return vp8_submv_prob[2];
597
    return l_is_zero ? vp8_submv_prob[1] : vp8_submv_prob[0];
598
}
599

    
600
/**
601
 * Split motion vector prediction, 16.4.
602
 * @returns the number of motion vectors parsed (2, 4 or 16)
603
 */
604
static int decode_splitmvs(VP8Context    *s,  VP56RangeCoder *c,
605
                            VP8Macroblock *mb, VP56mv         *base_mv)
606
{
607
    int part_idx = mb->partitioning =
608
        vp8_rac_get_tree(c, vp8_mbsplit_tree, vp8_mbsplit_prob);
609
    int n, num = vp8_mbsplit_count[part_idx];
610
    const uint8_t *mbsplits = vp8_mbsplits[part_idx],
611
                  *firstidx = vp8_mbfirstidx[part_idx];
612

    
613
    for (n = 0; n < num; n++) {
614
        int k = firstidx[n];
615
        const VP56mv *left, *above;
616
        const uint8_t *submv_prob;
617

    
618
        if (!(k & 3)) {
619
            VP8Macroblock *left_mb = &mb[-1];
620
            left = &left_mb->bmv[vp8_mbsplits[left_mb->partitioning][k + 3]];
621
        } else
622
            left  = &mb->bmv[mbsplits[k - 1]];
623
        if (k <= 3) {
624
            VP8Macroblock *above_mb = &mb[-s->mb_stride];
625
            above = &above_mb->bmv[vp8_mbsplits[above_mb->partitioning][k + 12]];
626
        } else
627
            above = &mb->bmv[mbsplits[k - 4]];
628

    
629
        submv_prob = get_submv_prob(left, above);
630

    
631
        switch (vp8_rac_get_tree(c, vp8_submv_ref_tree, submv_prob)) {
632
        case VP8_SUBMVMODE_NEW4X4:
633
            mb->bmv[n].y = base_mv->y + read_mv_component(c, s->prob->mvc[0]);
634
            mb->bmv[n].x = base_mv->x + read_mv_component(c, s->prob->mvc[1]);
635
            break;
636
        case VP8_SUBMVMODE_ZERO4X4:
637
            mb->bmv[n].x = 0;
638
            mb->bmv[n].y = 0;
639
            break;
640
        case VP8_SUBMVMODE_LEFT4X4:
641
            mb->bmv[n] = *left;
642
            break;
643
        case VP8_SUBMVMODE_TOP4X4:
644
            mb->bmv[n] = *above;
645
            break;
646
        }
647
    }
648

    
649
    return num;
650
}
651

    
652
static inline void decode_intra4x4_modes(VP56RangeCoder *c, uint8_t *intra4x4,
653
                                         int stride, int keyframe)
654
{
655
    int x, y, t, l;
656
    const uint8_t *ctx = vp8_pred4x4_prob_inter;
657

    
658
    for (y = 0; y < 4; y++) {
659
        for (x = 0; x < 4; x++) {
660
            if (keyframe) {
661
                t = intra4x4[x - stride];
662
                l = intra4x4[x - 1];
663
                ctx = vp8_pred4x4_prob_intra[t][l];
664
            }
665
            intra4x4[x] = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx);
666
        }
667
        intra4x4 += stride;
668
    }
669
}
670

    
671
static void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
672
                           uint8_t *intra4x4)
673
{
674
    VP56RangeCoder *c = &s->c;
675
    int n;
676

    
677
    if (s->segmentation.update_map)
678
        mb->segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid);
679

    
680
    mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0;
681

    
682
    if (s->keyframe) {
683
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra);
684

    
685
        if (mb->mode == MODE_I4x4) {
686
            decode_intra4x4_modes(c, intra4x4, s->b4_stride, 1);
687
        } else
688
            fill_rectangle(intra4x4, 4, 4, s->b4_stride, vp8_pred4x4_mode[mb->mode], 1);
689

    
690
        s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra);
691
        mb->ref_frame = VP56_FRAME_CURRENT;
692
    } else if (vp56_rac_get_prob(c, s->prob->intra)) {
693
        VP56mv near[2], best;
694
        int cnt[4] = { 0 };
695
        uint8_t p[4];
696

    
697
        // inter MB, 16.2
698
        if (vp56_rac_get_prob(c, s->prob->last))
699
            mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ?
700
                VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN;
701
        else
702
            mb->ref_frame = VP56_FRAME_PREVIOUS;
703

    
704
        // motion vectors, 16.3
705
        find_near_mvs(s, mb, mb_x, mb_y, near, &best, cnt);
706
        for (n = 0; n < 4; n++)
707
            p[n] = vp8_mode_contexts[cnt[n]][n];
708
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_mvinter, p);
709
        switch (mb->mode) {
710
        case VP8_MVMODE_SPLIT:
711
            mb->mv = mb->bmv[decode_splitmvs(s, c, mb, &best) - 1];
712
            break;
713
        case VP8_MVMODE_ZERO:
714
            mb->mv.x = 0;
715
            mb->mv.y = 0;
716
            break;
717
        case VP8_MVMODE_NEAREST:
718
            clamp_mv(s, &mb->mv, &near[0], mb_x, mb_y);
719
            break;
720
        case VP8_MVMODE_NEAR:
721
            clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y);
722
            break;
723
        case VP8_MVMODE_NEW:
724
            mb->mv.y = best.y + read_mv_component(c, s->prob->mvc[0]);
725
            mb->mv.x = best.x + read_mv_component(c, s->prob->mvc[1]);
726
            break;
727
        }
728
        if (mb->mode != VP8_MVMODE_SPLIT) {
729
            mb->partitioning = VP8_SPLITMVMODE_NONE;
730
            mb->bmv[0] = mb->mv;
731
        }
732
    } else {
733
        // intra MB, 16.1
734
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16);
735

    
736
        if (mb->mode == MODE_I4x4) {
737
            decode_intra4x4_modes(c, intra4x4, s->b4_stride, 0);
738
        } else
739
            fill_rectangle(intra4x4, 4, 4, s->b4_stride, vp8_pred4x4_mode[mb->mode], 1);
740

    
741
        s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c);
742
        mb->ref_frame = VP56_FRAME_CURRENT;
743
    }
744
}
745

    
746
/**
747
 * @param i initial coeff index, 0 unless a separate DC block is coded
748
 * @param zero_nhood the initial prediction context for number of surrounding
749
 *                   all-zero blocks (only left/top, so 0-2)
750
 * @param qmul array holding the dc/ac dequant factor at position 0/1
751
 * @return 0 if no coeffs were decoded
752
 *         otherwise, the index of the last coeff decoded plus one
753
 */
754
static int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16],
755
                               uint8_t probs[8][3][NUM_DCT_TOKENS-1],
756
                               int i, int zero_nhood, int16_t qmul[2])
757
{
758
    int token, nonzero = 0;
759
    int offset = 0;
760

    
761
    for (; i < 16; i++) {
762
        token = vp8_rac_get_tree_with_offset(c, vp8_coeff_tree, probs[vp8_coeff_band[i]][zero_nhood], offset);
763

    
764
        if (token == DCT_EOB)
765
            break;
766
        else if (token >= DCT_CAT1) {
767
            int cat = token-DCT_CAT1;
768
            token = vp8_rac_get_coeff(c, vp8_dct_cat_prob[cat]);
769
            token += vp8_dct_cat_offset[cat];
770
        }
771

    
772
        // after the first token, the non-zero prediction context becomes
773
        // based on the last decoded coeff
774
        if (!token) {
775
            zero_nhood = 0;
776
            offset = 1;
777
            continue;
778
        } else if (token == 1)
779
            zero_nhood = 1;
780
        else
781
            zero_nhood = 2;
782

    
783
        // todo: full [16] qmat? load into register?
784
        block[zigzag_scan[i]] = (vp8_rac_get(c) ? -token : token) * qmul[!!i];
785
        nonzero = i+1;
786
        offset = 0;
787
    }
788
    return nonzero;
789
}
790

    
791
static void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,
792
                             uint8_t t_nnz[9], uint8_t l_nnz[9])
793
{
794
    LOCAL_ALIGNED_16(DCTELEM, dc,[16]);
795
    int i, x, y, luma_start = 0, luma_ctx = 3;
796
    int nnz_pred, nnz, nnz_total = 0;
797
    int segment = s->segmentation.enabled ? mb->segment : 0;
798

    
799
    s->dsp.clear_blocks((DCTELEM *)s->block);
800

    
801
    if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
802
        AV_ZERO128(dc);
803
        AV_ZERO128(dc+8);
804
        nnz_pred = t_nnz[8] + l_nnz[8];
805

    
806
        // decode DC values and do hadamard
807
        nnz = decode_block_coeffs(c, dc, s->prob->token[1], 0, nnz_pred,
808
                                  s->qmat[segment].luma_dc_qmul);
809
        l_nnz[8] = t_nnz[8] = !!nnz;
810
        nnz_total += nnz;
811
        s->vp8dsp.vp8_luma_dc_wht(s->block, dc);
812
        luma_start = 1;
813
        luma_ctx = 0;
814
    }
815

    
816
    // luma blocks
817
    for (y = 0; y < 4; y++)
818
        for (x = 0; x < 4; x++) {
819
            nnz_pred = l_nnz[y] + t_nnz[x];
820
            nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start,
821
                                      nnz_pred, s->qmat[segment].luma_qmul);
822
            // nnz+luma_start may be one more than the actual last index, but we don't care
823
            s->non_zero_count_cache[y][x] = nnz + luma_start;
824
            t_nnz[x] = l_nnz[y] = !!nnz;
825
            nnz_total += nnz;
826
        }
827

    
828
    // chroma blocks
829
    // TODO: what to do about dimensions? 2nd dim for luma is x,
830
    // but for chroma it's (y<<1)|x
831
    for (i = 4; i < 6; i++)
832
        for (y = 0; y < 2; y++)
833
            for (x = 0; x < 2; x++) {
834
                nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x];
835
                nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0,
836
                                          nnz_pred, s->qmat[segment].chroma_qmul);
837
                s->non_zero_count_cache[i][(y<<1)+x] = nnz;
838
                t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz;
839
                nnz_total += nnz;
840
            }
841

    
842
    // if there were no coded coeffs despite the macroblock not being marked skip,
843
    // we MUST not do the inner loop filter and should not do IDCT
844
    // Since skip isn't used for bitstream prediction, just manually set it.
845
    if (!nnz_total)
846
        mb->skip = 1;
847
}
848

    
849
static int check_intra_pred_mode(int mode, int mb_x, int mb_y)
850
{
851
    if (mode == DC_PRED8x8) {
852
        if (!(mb_x|mb_y))
853
            mode = DC_128_PRED8x8;
854
        else if (!mb_y)
855
            mode = LEFT_DC_PRED8x8;
856
        else if (!mb_x)
857
            mode = TOP_DC_PRED8x8;
858
    }
859
    return mode;
860
}
861

    
862
static void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
863
                          uint8_t *bmode, int mb_x, int mb_y)
864
{
865
    int x, y, mode, nnz, tr;
866

    
867
    if (mb->mode < MODE_I4x4) {
868
        mode = check_intra_pred_mode(mb->mode, mb_x, mb_y);
869
        s->hpc.pred16x16[mode](dst[0], s->linesize);
870
    } else {
871
        uint8_t *ptr = dst[0];
872

    
873
        // all blocks on the right edge of the macroblock use bottom edge
874
        // the top macroblock for their topright edge
875
        uint8_t *tr_right = ptr - s->linesize + 16;
876

    
877
        // if we're on the right edge of the frame, said edge is extended
878
        // from the top macroblock
879
        if (mb_x == s->mb_width-1) {
880
            tr = tr_right[-1]*0x01010101;
881
            tr_right = (uint8_t *)&tr;
882
        }
883

    
884
        for (y = 0; y < 4; y++) {
885
            uint8_t *topright = ptr + 4 - s->linesize;
886
            for (x = 0; x < 4; x++) {
887
                if (x == 3)
888
                    topright = tr_right;
889

    
890
                s->hpc.pred4x4[bmode[x]](ptr+4*x, topright, s->linesize);
891

    
892
                nnz = s->non_zero_count_cache[y][x];
893
                if (nnz) {
894
                    if (nnz == 1)
895
                        s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize);
896
                    else
897
                        s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize);
898
                }
899
                topright += 4;
900
            }
901

    
902
            ptr   += 4*s->linesize;
903
            bmode += s->b4_stride;
904
        }
905
    }
906

    
907
    mode = check_intra_pred_mode(s->chroma_pred_mode, mb_x, mb_y);
908
    s->hpc.pred8x8[mode](dst[1], s->uvlinesize);
909
    s->hpc.pred8x8[mode](dst[2], s->uvlinesize);
910
}
911

    
912
/**
913
 * Generic MC function.
914
 *
915
 * @param s VP8 decoding context
916
 * @param luma 1 for luma (Y) planes, 0 for chroma (Cb/Cr) planes
917
 * @param dst target buffer for block data at block position
918
 * @param src reference picture buffer at origin (0, 0)
919
 * @param mv motion vector (relative to block position) to get pixel data from
920
 * @param x_off horizontal position of block from origin (0, 0)
921
 * @param y_off vertical position of block from origin (0, 0)
922
 * @param block_w width of block (16, 8 or 4)
923
 * @param block_h height of block (always same as block_w)
924
 * @param width width of src/dst plane data
925
 * @param height height of src/dst plane data
926
 * @param linesize size of a single line of plane data, including padding
927
 */
928
static inline void vp8_mc(VP8Context *s, int luma,
929
                          uint8_t *dst, uint8_t *src, const VP56mv *mv,
930
                          int x_off, int y_off, int block_w, int block_h,
931
                          int width, int height, int linesize,
932
                          vp8_mc_func mc_func[3][3])
933
{
934
    static const uint8_t idx[8] = { 0, 1, 2, 1, 2, 1, 2, 1 };
935
    int mx = (mv->x << luma)&7, mx_idx = idx[mx];
936
    int my = (mv->y << luma)&7, my_idx = idx[my];
937

    
938
    x_off += mv->x >> (3 - luma);
939
    y_off += mv->y >> (3 - luma);
940

    
941
    // edge emulation
942
    src += y_off * linesize + x_off;
943
    if (x_off < 2 || x_off >= width  - block_w - 3 ||
944
        y_off < 2 || y_off >= height - block_h - 3) {
945
        ff_emulated_edge_mc(s->edge_emu_buffer, src - 2 * linesize - 2, linesize,
946
                            block_w + 5, block_h + 5,
947
                            x_off - 2, y_off - 2, width, height);
948
        src = s->edge_emu_buffer + 2 + linesize * 2;
949
    }
950

    
951
    mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);
952
}
953

    
954
static inline void vp8_mc_part(VP8Context *s, uint8_t *dst[3],
955
                               AVFrame *ref_frame, int x_off, int y_off,
956
                               int bx_off, int by_off,
957
                               int block_w, int block_h,
958
                               int width, int height, VP56mv *mv)
959
{
960
    VP56mv uvmv = *mv;
961

    
962
    /* Y */
963
    vp8_mc(s, 1, dst[0] + by_off * s->linesize + bx_off,
964
           ref_frame->data[0], mv, x_off + bx_off, y_off + by_off,
965
           block_w, block_h, width, height, s->linesize,
966
           s->put_pixels_tab[block_w == 8]);
967

    
968
    /* U/V */
969
    if (s->profile == 3) {
970
        uvmv.x &= ~7;
971
        uvmv.y &= ~7;
972
    }
973
    x_off   >>= 1; y_off   >>= 1;
974
    bx_off  >>= 1; by_off  >>= 1;
975
    width   >>= 1; height  >>= 1;
976
    block_w >>= 1; block_h >>= 1;
977
    vp8_mc(s, 0, dst[1] + by_off * s->uvlinesize + bx_off,
978
           ref_frame->data[1], &uvmv, x_off + bx_off, y_off + by_off,
979
           block_w, block_h, width, height, s->uvlinesize,
980
           s->put_pixels_tab[1 + (block_w == 4)]);
981
    vp8_mc(s, 0, dst[2] + by_off * s->uvlinesize + bx_off,
982
           ref_frame->data[2], &uvmv, x_off + bx_off, y_off + by_off,
983
           block_w, block_h, width, height, s->uvlinesize,
984
           s->put_pixels_tab[1 + (block_w == 4)]);
985
}
986

    
987
/**
988
 * Apply motion vectors to prediction buffer, chapter 18.
989
 */
990
static void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
991
                          int mb_x, int mb_y)
992
{
993
    int x_off = mb_x << 4, y_off = mb_y << 4;
994
    int width = 16*s->mb_width, height = 16*s->mb_height;
995

    
996
    if (mb->mode < VP8_MVMODE_SPLIT) {
997
        vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
998
                    0, 0, 16, 16, width, height, &mb->mv);
999
    } else switch (mb->partitioning) {
1000
    case VP8_SPLITMVMODE_4x4: {
1001
        int x, y;
1002
        VP56mv uvmv;
1003

    
1004
        /* Y */
1005
        for (y = 0; y < 4; y++) {
1006
            for (x = 0; x < 4; x++) {
1007
                vp8_mc(s, 1, dst[0] + 4*y*s->linesize + x*4,
1008
                       s->framep[mb->ref_frame]->data[0], &mb->bmv[4*y + x],
1009
                       4*x + x_off, 4*y + y_off, 4, 4,
1010
                       width, height, s->linesize,
1011
                       s->put_pixels_tab[2]);
1012
            }
1013
        }
1014

    
1015
        /* U/V */
1016
        x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1;
1017
        for (y = 0; y < 2; y++) {
1018
            for (x = 0; x < 2; x++) {
1019
                uvmv.x = mb->bmv[ 2*y    * 4 + 2*x  ].x +
1020
                         mb->bmv[ 2*y    * 4 + 2*x+1].x +
1021
                         mb->bmv[(2*y+1) * 4 + 2*x  ].x +
1022
                         mb->bmv[(2*y+1) * 4 + 2*x+1].x;
1023
                uvmv.y = mb->bmv[ 2*y    * 4 + 2*x  ].y +
1024
                         mb->bmv[ 2*y    * 4 + 2*x+1].y +
1025
                         mb->bmv[(2*y+1) * 4 + 2*x  ].y +
1026
                         mb->bmv[(2*y+1) * 4 + 2*x+1].y;
1027
                uvmv.x = (uvmv.x + 2 + (uvmv.x >> (INT_BIT-1))) >> 2;
1028
                uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2;
1029
                if (s->profile == 3) {
1030
                    uvmv.x &= ~7;
1031
                    uvmv.y &= ~7;
1032
                }
1033
                vp8_mc(s, 0, dst[1] + 4*y*s->uvlinesize + x*4,
1034
                       s->framep[mb->ref_frame]->data[1], &uvmv,
1035
                       4*x + x_off, 4*y + y_off, 4, 4,
1036
                       width, height, s->uvlinesize,
1037
                       s->put_pixels_tab[2]);
1038
                vp8_mc(s, 0, dst[2] + 4*y*s->uvlinesize + x*4,
1039
                       s->framep[mb->ref_frame]->data[2], &uvmv,
1040
                       4*x + x_off, 4*y + y_off, 4, 4,
1041
                       width, height, s->uvlinesize,
1042
                       s->put_pixels_tab[2]);
1043
            }
1044
        }
1045
        break;
1046
    }
1047
    case VP8_SPLITMVMODE_16x8:
1048
        vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1049
                    0, 0, 16, 8, width, height, &mb->bmv[0]);
1050
        vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1051
                    0, 8, 16, 8, width, height, &mb->bmv[1]);
1052
        break;
1053
    case VP8_SPLITMVMODE_8x16:
1054
        vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1055
                    0, 0, 8, 16, width, height, &mb->bmv[0]);
1056
        vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1057
                    8, 0, 8, 16, width, height, &mb->bmv[1]);
1058
        break;
1059
    case VP8_SPLITMVMODE_8x8:
1060
        vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1061
                    0, 0, 8, 8, width, height, &mb->bmv[0]);
1062
        vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1063
                    8, 0, 8, 8, width, height, &mb->bmv[1]);
1064
        vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1065
                    0, 8, 8, 8, width, height, &mb->bmv[2]);
1066
        vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off,
1067
                    8, 8, 8, 8, width, height, &mb->bmv[3]);
1068
        break;
1069
    }
1070
}
1071

    
1072
static void idct_mb(VP8Context *s, uint8_t *y_dst, uint8_t *u_dst, uint8_t *v_dst,
1073
                    VP8Macroblock *mb)
1074
{
1075
    int x, y, nnz;
1076

    
1077
    if (mb->mode != MODE_I4x4)
1078
        for (y = 0; y < 4; y++) {
1079
            for (x = 0; x < 4; x++) {
1080
                nnz = s->non_zero_count_cache[y][x];
1081
                if (nnz) {
1082
                    if (nnz == 1)
1083
                        s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize);
1084
                    else
1085
                        s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize);
1086
                }
1087
            }
1088
            y_dst += 4*s->linesize;
1089
        }
1090

    
1091
    for (y = 0; y < 2; y++) {
1092
        for (x = 0; x < 2; x++) {
1093
            nnz = s->non_zero_count_cache[4][(y<<1)+x];
1094
            if (nnz) {
1095
                if (nnz == 1)
1096
                    s->vp8dsp.vp8_idct_dc_add(u_dst+4*x, s->block[4][(y<<1)+x], s->uvlinesize);
1097
                else
1098
                    s->vp8dsp.vp8_idct_add(u_dst+4*x, s->block[4][(y<<1)+x], s->uvlinesize);
1099
            }
1100

    
1101
            nnz = s->non_zero_count_cache[5][(y<<1)+x];
1102
            if (nnz) {
1103
                if (nnz == 1)
1104
                    s->vp8dsp.vp8_idct_dc_add(v_dst+4*x, s->block[5][(y<<1)+x], s->uvlinesize);
1105
                else
1106
                    s->vp8dsp.vp8_idct_add(v_dst+4*x, s->block[5][(y<<1)+x], s->uvlinesize);
1107
            }
1108
        }
1109
        u_dst += 4*s->uvlinesize;
1110
        v_dst += 4*s->uvlinesize;
1111
    }
1112
}
1113

    
1114
static void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, int *level, int *inner, int *hev_thresh)
1115
{
1116
    int interior_limit, filter_level;
1117

    
1118
    if (s->segmentation.enabled) {
1119
        filter_level = s->segmentation.filter_level[mb->segment];
1120
        if (!s->segmentation.absolute_vals)
1121
            filter_level += s->filter.level;
1122
    } else
1123
        filter_level = s->filter.level;
1124

    
1125
    if (s->lf_delta.enabled) {
1126
        filter_level += s->lf_delta.ref[mb->ref_frame];
1127

    
1128
        if (mb->ref_frame == VP56_FRAME_CURRENT) {
1129
            if (mb->mode == MODE_I4x4)
1130
                filter_level += s->lf_delta.mode[0];
1131
        } else {
1132
            if (mb->mode == VP8_MVMODE_ZERO)
1133
                filter_level += s->lf_delta.mode[1];
1134
            else if (mb->mode == VP8_MVMODE_SPLIT)
1135
                filter_level += s->lf_delta.mode[3];
1136
            else
1137
                filter_level += s->lf_delta.mode[2];
1138
        }
1139
    }
1140
    filter_level = av_clip(filter_level, 0, 63);
1141

    
1142
    interior_limit = filter_level;
1143
    if (s->filter.sharpness) {
1144
        interior_limit >>= s->filter.sharpness > 4 ? 2 : 1;
1145
        interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness);
1146
    }
1147
    interior_limit = FFMAX(interior_limit, 1);
1148

    
1149
    *level = filter_level;
1150
    *inner = interior_limit;
1151

    
1152
    if (hev_thresh) {
1153
        *hev_thresh = filter_level >= 15;
1154

    
1155
        if (s->keyframe) {
1156
            if (filter_level >= 40)
1157
                *hev_thresh = 2;
1158
        } else {
1159
            if (filter_level >= 40)
1160
                *hev_thresh = 3;
1161
            else if (filter_level >= 20)
1162
                *hev_thresh = 2;
1163
        }
1164
    }
1165
}
1166

    
1167
// TODO: look at backup_mb_border / xchg_mb_border in h264.c
1168
static void filter_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb, int mb_x, int mb_y)
1169
{
1170
    int filter_level, inner_limit, hev_thresh, mbedge_lim, bedge_lim;
1171

    
1172
    filter_level_for_mb(s, mb, &filter_level, &inner_limit, &hev_thresh);
1173
    if (!filter_level)
1174
        return;
1175

    
1176
    mbedge_lim = 2*(filter_level+2) + inner_limit;
1177
     bedge_lim = 2* filter_level    + inner_limit;
1178

    
1179
    if (mb_x) {
1180
        s->vp8dsp.vp8_h_loop_filter16(dst[0], s->linesize,   mbedge_lim, inner_limit, hev_thresh);
1181
        s->vp8dsp.vp8_h_loop_filter8 (dst[1], s->uvlinesize, mbedge_lim, inner_limit, hev_thresh);
1182
        s->vp8dsp.vp8_h_loop_filter8 (dst[2], s->uvlinesize, mbedge_lim, inner_limit, hev_thresh);
1183
    }
1184

    
1185
    if (!mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT) {
1186
        s->vp8dsp.vp8_h_loop_filter16_inner(dst[0]+ 4, s->linesize,   bedge_lim, inner_limit, hev_thresh);
1187
        s->vp8dsp.vp8_h_loop_filter16_inner(dst[0]+ 8, s->linesize,   bedge_lim, inner_limit, hev_thresh);
1188
        s->vp8dsp.vp8_h_loop_filter16_inner(dst[0]+12, s->linesize,   bedge_lim, inner_limit, hev_thresh);
1189
        s->vp8dsp.vp8_h_loop_filter8_inner (dst[1]+ 4, s->uvlinesize, bedge_lim, inner_limit, hev_thresh);
1190
        s->vp8dsp.vp8_h_loop_filter8_inner (dst[2]+ 4, s->uvlinesize, bedge_lim, inner_limit, hev_thresh);
1191
    }
1192

    
1193
    if (mb_y) {
1194
        s->vp8dsp.vp8_v_loop_filter16(dst[0], s->linesize,   mbedge_lim, inner_limit, hev_thresh);
1195
        s->vp8dsp.vp8_v_loop_filter8 (dst[1], s->uvlinesize, mbedge_lim, inner_limit, hev_thresh);
1196
        s->vp8dsp.vp8_v_loop_filter8 (dst[2], s->uvlinesize, mbedge_lim, inner_limit, hev_thresh);
1197
    }
1198

    
1199
    if (!mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT) {
1200
        s->vp8dsp.vp8_v_loop_filter16_inner(dst[0]+ 4*s->linesize,   s->linesize,   bedge_lim, inner_limit, hev_thresh);
1201
        s->vp8dsp.vp8_v_loop_filter16_inner(dst[0]+ 8*s->linesize,   s->linesize,   bedge_lim, inner_limit, hev_thresh);
1202
        s->vp8dsp.vp8_v_loop_filter16_inner(dst[0]+12*s->linesize,   s->linesize,   bedge_lim, inner_limit, hev_thresh);
1203
        s->vp8dsp.vp8_v_loop_filter8_inner (dst[1]+ 4*s->uvlinesize, s->uvlinesize, bedge_lim, inner_limit, hev_thresh);
1204
        s->vp8dsp.vp8_v_loop_filter8_inner (dst[2]+ 4*s->uvlinesize, s->uvlinesize, bedge_lim, inner_limit, hev_thresh);
1205
    }
1206
}
1207

    
1208
static void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8Macroblock *mb, int mb_x, int mb_y)
1209
{
1210
    int filter_level, inner_limit, mbedge_lim, bedge_lim;
1211

    
1212
    filter_level_for_mb(s, mb, &filter_level, &inner_limit, NULL);
1213
    if (!filter_level)
1214
        return;
1215

    
1216
    mbedge_lim = 2*(filter_level+2) + inner_limit;
1217
     bedge_lim = 2* filter_level    + inner_limit;
1218

    
1219
    if (mb_x)
1220
        s->vp8dsp.vp8_h_loop_filter_simple(dst, s->linesize, mbedge_lim);
1221
    if (!mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT) {
1222
        s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, s->linesize, bedge_lim);
1223
        s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, s->linesize, bedge_lim);
1224
        s->vp8dsp.vp8_h_loop_filter_simple(dst+12, s->linesize, bedge_lim);
1225
    }
1226

    
1227
    if (mb_y)
1228
        s->vp8dsp.vp8_v_loop_filter_simple(dst, s->linesize, mbedge_lim);
1229
    if (!mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT) {
1230
        s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*s->linesize, s->linesize, bedge_lim);
1231
        s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*s->linesize, s->linesize, bedge_lim);
1232
        s->vp8dsp.vp8_v_loop_filter_simple(dst+12*s->linesize, s->linesize, bedge_lim);
1233
    }
1234
}
1235

    
1236
static void filter_mb_row(VP8Context *s, int mb_y)
1237
{
1238
    VP8Macroblock *mb = s->macroblocks + mb_y*s->mb_stride;
1239
    uint8_t *dst[3] = {
1240
        s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize,
1241
        s->framep[VP56_FRAME_CURRENT]->data[1] +  8*mb_y*s->uvlinesize,
1242
        s->framep[VP56_FRAME_CURRENT]->data[2] +  8*mb_y*s->uvlinesize
1243
    };
1244
    int mb_x;
1245

    
1246
    for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1247
        filter_mb(s, dst, mb++, mb_x, mb_y);
1248
        dst[0] += 16;
1249
        dst[1] += 8;
1250
        dst[2] += 8;
1251
    }
1252
}
1253

    
1254
static void filter_mb_row_simple(VP8Context *s, int mb_y)
1255
{
1256
    uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize;
1257
    VP8Macroblock *mb = s->macroblocks + mb_y*s->mb_stride;
1258
    int mb_x;
1259

    
1260
    for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1261
        filter_mb_simple(s, dst, mb++, mb_x, mb_y);
1262
        dst += 16;
1263
    }
1264
}
1265

    
1266
static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
1267
                            AVPacket *avpkt)
1268
{
1269
    VP8Context *s = avctx->priv_data;
1270
    int ret, mb_x, mb_y, i, y, referenced;
1271
    enum AVDiscard skip_thresh;
1272
    AVFrame *curframe;
1273

    
1274
    if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0)
1275
        return ret;
1276

    
1277
    referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT
1278
                                || s->update_altref == VP56_FRAME_CURRENT;
1279

    
1280
    skip_thresh = !referenced ? AVDISCARD_NONREF :
1281
                    !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL;
1282

    
1283
    if (avctx->skip_frame >= skip_thresh) {
1284
        s->invisible = 1;
1285
        goto skip_decode;
1286
    }
1287

    
1288
    for (i = 0; i < 4; i++)
1289
        if (&s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1290
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1291
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) {
1292
            curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i];
1293
            break;
1294
        }
1295
    if (curframe->data[0])
1296
        avctx->release_buffer(avctx, curframe);
1297

    
1298
    curframe->key_frame = s->keyframe;
1299
    curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE;
1300
    curframe->reference = referenced ? 3 : 0;
1301
    if ((ret = avctx->get_buffer(avctx, curframe))) {
1302
        av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n");
1303
        return ret;
1304
    }
1305

    
1306
    // Given that arithmetic probabilities are updated every frame, it's quite likely
1307
    // that the values we have on a random interframe are complete junk if we didn't
1308
    // start decode on a keyframe. So just don't display anything rather than junk.
1309
    if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] ||
1310
                         !s->framep[VP56_FRAME_GOLDEN] ||
1311
                         !s->framep[VP56_FRAME_GOLDEN2])) {
1312
        av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n");
1313
        return AVERROR_INVALIDDATA;
1314
    }
1315

    
1316
    s->linesize   = curframe->linesize[0];
1317
    s->uvlinesize = curframe->linesize[1];
1318

    
1319
    if (!s->edge_emu_buffer)
1320
        s->edge_emu_buffer = av_malloc(21*s->linesize);
1321

    
1322
    memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz));
1323

    
1324
    // top edge of 127 for intra prediction
1325
    if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) {
1326
        memset(curframe->data[0] - s->linesize  -1, 127, s->linesize  +1);
1327
        memset(curframe->data[1] - s->uvlinesize-1, 127, s->uvlinesize+1);
1328
        memset(curframe->data[2] - s->uvlinesize-1, 127, s->uvlinesize+1);
1329
    }
1330

    
1331
    for (mb_y = 0; mb_y < s->mb_height; mb_y++) {
1332
        VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)];
1333
        VP8Macroblock *mb = s->macroblocks + mb_y*s->mb_stride;
1334
        uint8_t *intra4x4 = s->intra4x4_pred_mode + 4*mb_y*s->b4_stride;
1335
        uint8_t *dst[3] = {
1336
            curframe->data[0] + 16*mb_y*s->linesize,
1337
            curframe->data[1] +  8*mb_y*s->uvlinesize,
1338
            curframe->data[2] +  8*mb_y*s->uvlinesize
1339
        };
1340

    
1341
        memset(s->left_nnz, 0, sizeof(s->left_nnz));
1342

    
1343
        // left edge of 129 for intra prediction
1344
        if (!(avctx->flags & CODEC_FLAG_EMU_EDGE))
1345
            for (i = 0; i < 3; i++)
1346
                for (y = 0; y < 16>>!!i; y++)
1347
                    dst[i][y*curframe->linesize[i]-1] = 129;
1348

    
1349
        for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1350
            decode_mb_mode(s, mb, mb_x, mb_y, intra4x4 + 4*mb_x);
1351

    
1352
            if (!mb->skip)
1353
                decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);
1354
            else {
1355
                AV_ZERO128(s->non_zero_count_cache);    // luma
1356
                AV_ZERO64(s->non_zero_count_cache[4]);  // chroma
1357
            }
1358

    
1359
            if (mb->mode <= MODE_I4x4) {
1360
                intra_predict(s, dst, mb, intra4x4 + 4*mb_x, mb_x, mb_y);
1361
                memset(mb->bmv, 0, sizeof(mb->bmv));
1362
            } else {
1363
                inter_predict(s, dst, mb, mb_x, mb_y);
1364
            }
1365

    
1366
            if (!mb->skip) {
1367
                idct_mb(s, dst[0], dst[1], dst[2], mb);
1368
            } else {
1369
                AV_ZERO64(s->left_nnz);
1370
                AV_WN64(s->top_nnz[mb_x], 0);   // array of 9, so unaligned
1371

    
1372
                // Reset DC block predictors if they would exist if the mb had coefficients
1373
                if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
1374
                    s->left_nnz[8]      = 0;
1375
                    s->top_nnz[mb_x][8] = 0;
1376
                }
1377
            }
1378

    
1379
            dst[0] += 16;
1380
            dst[1] += 8;
1381
            dst[2] += 8;
1382
            mb++;
1383
        }
1384
        if (mb_y && s->filter.level && avctx->skip_loop_filter < skip_thresh) {
1385
            if (s->filter.simple)
1386
                filter_mb_row_simple(s, mb_y-1);
1387
            else
1388
                filter_mb_row(s, mb_y-1);
1389
        }
1390
    }
1391
    if (s->filter.level && avctx->skip_loop_filter < skip_thresh) {
1392
        if (s->filter.simple)
1393
            filter_mb_row_simple(s, mb_y-1);
1394
        else
1395
            filter_mb_row(s, mb_y-1);
1396
    }
1397

    
1398
skip_decode:
1399
    // if future frames don't use the updated probabilities,
1400
    // reset them to the values we saved
1401
    if (!s->update_probabilities)
1402
        s->prob[0] = s->prob[1];
1403

    
1404
    // check if golden and altref are swapped
1405
    if (s->update_altref == VP56_FRAME_GOLDEN &&
1406
        s->update_golden == VP56_FRAME_GOLDEN2)
1407
        FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN], s->framep[VP56_FRAME_GOLDEN2]);
1408
    else {
1409
        if (s->update_altref != VP56_FRAME_NONE)
1410
            s->framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref];
1411

    
1412
        if (s->update_golden != VP56_FRAME_NONE)
1413
            s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden];
1414
    }
1415

    
1416
    if (s->update_last) // move cur->prev
1417
        s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT];
1418

    
1419
    // release no longer referenced frames
1420
    for (i = 0; i < 4; i++)
1421
        if (s->frames[i].data[0] &&
1422
            &s->frames[i] != s->framep[VP56_FRAME_CURRENT] &&
1423
            &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1424
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1425
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2])
1426
            avctx->release_buffer(avctx, &s->frames[i]);
1427

    
1428
    if (!s->invisible) {
1429
        *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT];
1430
        *data_size = sizeof(AVFrame);
1431
    }
1432

    
1433
    return avpkt->size;
1434
}
1435

    
1436
static av_cold int vp8_decode_init(AVCodecContext *avctx)
1437
{
1438
    VP8Context *s = avctx->priv_data;
1439

    
1440
    s->avctx = avctx;
1441
    avctx->pix_fmt = PIX_FMT_YUV420P;
1442

    
1443
    dsputil_init(&s->dsp, avctx);
1444
    ff_h264_pred_init(&s->hpc, CODEC_ID_VP8);
1445
    ff_vp8dsp_init(&s->vp8dsp);
1446

    
1447
    // intra pred needs edge emulation among other things
1448
    if (avctx->flags&CODEC_FLAG_EMU_EDGE) {
1449
        av_log(avctx, AV_LOG_ERROR, "Edge emulation not supported\n");
1450
        return AVERROR_PATCHWELCOME;
1451
    }
1452

    
1453
    return 0;
1454
}
1455

    
1456
static av_cold int vp8_decode_free(AVCodecContext *avctx)
1457
{
1458
    vp8_decode_flush(avctx);
1459
    return 0;
1460
}
1461

    
1462
AVCodec vp8_decoder = {
1463
    "vp8",
1464
    AVMEDIA_TYPE_VIDEO,
1465
    CODEC_ID_VP8,
1466
    sizeof(VP8Context),
1467
    vp8_decode_init,
1468
    NULL,
1469
    vp8_decode_free,
1470
    vp8_decode_frame,
1471
    CODEC_CAP_DR1,
1472
    .flush = vp8_decode_flush,
1473
    .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"),
1474
};