Statistics
| Branch: | Revision:

ffmpeg / libavcodec / vp8.c @ 64233e70

History | View | Annotate | Download (65.6 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 "libavcore/imgutils.h"
26
#include "avcodec.h"
27
#include "vp56.h"
28
#include "vp8data.h"
29
#include "vp8dsp.h"
30
#include "h264pred.h"
31
#include "rectangle.h"
32

    
33
typedef struct {
34
    uint8_t filter_level;
35
    uint8_t inner_limit;
36
    uint8_t inner_filter;
37
} VP8FilterStrength;
38

    
39
typedef struct {
40
    uint8_t skip;
41
    // todo: make it possible to check for at least (i4x4 or split_mv)
42
    // in one op. are others needed?
43
    uint8_t mode;
44
    uint8_t ref_frame;
45
    uint8_t partitioning;
46
    VP56mv mv;
47
    VP56mv bmv[16];
48
} VP8Macroblock;
49

    
50
typedef struct {
51
    AVCodecContext *avctx;
52
    DSPContext dsp;
53
    VP8DSPContext vp8dsp;
54
    H264PredContext hpc;
55
    vp8_mc_func put_pixels_tab[3][3][3];
56
    AVFrame frames[4];
57
    AVFrame *framep[4];
58
    uint8_t *edge_emu_buffer;
59
    VP56RangeCoder c;   ///< header context, includes mb modes and motion vectors
60
    int profile;
61

    
62
    int mb_width;   /* number of horizontal MB */
63
    int mb_height;  /* number of vertical MB */
64
    int linesize;
65
    int uvlinesize;
66

    
67
    int keyframe;
68
    int invisible;
69
    int update_last;    ///< update VP56_FRAME_PREVIOUS with the current one
70
    int update_golden;  ///< VP56_FRAME_NONE if not updated, or which frame to copy if so
71
    int update_altref;
72
    int deblock_filter;
73

    
74
    /**
75
     * If this flag is not set, all the probability updates
76
     * are discarded after this frame is decoded.
77
     */
78
    int update_probabilities;
79

    
80
    /**
81
     * All coefficients are contained in separate arith coding contexts.
82
     * There can be 1, 2, 4, or 8 of these after the header context.
83
     */
84
    int num_coeff_partitions;
85
    VP56RangeCoder coeff_partition[8];
86

    
87
    VP8Macroblock *macroblocks;
88
    VP8Macroblock *macroblocks_base;
89
    VP8FilterStrength *filter_strength;
90

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

    
95
    /**
96
     * Cache of the top row needed for intra prediction
97
     * 16 for luma, 8 for each chroma plane
98
     */
99
    uint8_t (*top_border)[16+8+8];
100

    
101
    /**
102
     * For coeff decode, we need to know whether the above block had non-zero
103
     * coefficients. This means for each macroblock, we need data for 4 luma
104
     * blocks, 2 u blocks, 2 v blocks, and the luma dc block, for a total of 9
105
     * per macroblock. We keep the last row in top_nnz.
106
     */
107
    uint8_t (*top_nnz)[9];
108
    DECLARE_ALIGNED(8, uint8_t, left_nnz)[9];
109

    
110
    /**
111
     * This is the index plus one of the last non-zero coeff
112
     * for each of the blocks in the current macroblock.
113
     * So, 0 -> no coeffs
114
     *     1 -> dc-only (special transform)
115
     *     2+-> full transform
116
     */
117
    DECLARE_ALIGNED(16, uint8_t, non_zero_count_cache)[6][4];
118
    DECLARE_ALIGNED(16, DCTELEM, block)[6][4][16];
119
    DECLARE_ALIGNED(16, DCTELEM, block_dc)[16];
120
    uint8_t intra4x4_pred_mode_mb[16];
121

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

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

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

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

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

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

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

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

    
183
    /**
184
     * These are all of the updatable probabilities for binary decisions.
185
     * They are only implictly reset on keyframes, making it quite likely
186
     * for an interframe to desync if a prior frame's header was corrupt
187
     * or missing outright!
188
     */
189
    struct {
190
        uint8_t segmentid[3];
191
        uint8_t mbskip;
192
        uint8_t intra;
193
        uint8_t last;
194
        uint8_t golden;
195
        uint8_t pred16x16[4];
196
        uint8_t pred8x8c[3];
197
        /* Padded to allow overreads */
198
        uint8_t token[4][17][3][NUM_DCT_TOKENS-1];
199
        uint8_t mvc[2][19];
200
    } prob[2];
201
} VP8Context;
202

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

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

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

    
221
    s->macroblocks        = NULL;
222
}
223

    
224
static int update_dimensions(VP8Context *s, int width, int height)
225
{
226
    if (av_image_check_size(width, height, 0, s->avctx))
227
        return AVERROR_INVALIDDATA;
228

    
229
    vp8_decode_flush(s->avctx);
230

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

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

    
236
    s->macroblocks_base        = av_mallocz((s->mb_width+s->mb_height*2+1)*sizeof(*s->macroblocks));
237
    s->filter_strength         = av_mallocz(s->mb_width*sizeof(*s->filter_strength));
238
    s->intra4x4_pred_mode_top  = av_mallocz(s->mb_width*4);
239
    s->top_nnz                 = av_mallocz(s->mb_width*sizeof(*s->top_nnz));
240
    s->top_border              = av_mallocz((s->mb_width+1)*sizeof(*s->top_border));
241
    s->segmentation_map        = av_mallocz(s->mb_width*s->mb_height);
242

    
243
    if (!s->macroblocks_base || !s->filter_strength || !s->intra4x4_pred_mode_top ||
244
        !s->top_nnz || !s->top_border || !s->segmentation_map)
245
        return AVERROR(ENOMEM);
246

    
247
    s->macroblocks        = s->macroblocks_base + 1;
248

    
249
    return 0;
250
}
251

    
252
static void parse_segment_info(VP8Context *s)
253
{
254
    VP56RangeCoder *c = &s->c;
255
    int i;
256

    
257
    s->segmentation.update_map = vp8_rac_get(c);
258

    
259
    if (vp8_rac_get(c)) { // update segment feature data
260
        s->segmentation.absolute_vals = vp8_rac_get(c);
261

    
262
        for (i = 0; i < 4; i++)
263
            s->segmentation.base_quant[i]   = vp8_rac_get_sint(c, 7);
264

    
265
        for (i = 0; i < 4; i++)
266
            s->segmentation.filter_level[i] = vp8_rac_get_sint(c, 6);
267
    }
268
    if (s->segmentation.update_map)
269
        for (i = 0; i < 3; i++)
270
            s->prob->segmentid[i] = vp8_rac_get(c) ? vp8_rac_get_uint(c, 8) : 255;
271
}
272

    
273
static void update_lf_deltas(VP8Context *s)
274
{
275
    VP56RangeCoder *c = &s->c;
276
    int i;
277

    
278
    for (i = 0; i < 4; i++)
279
        s->lf_delta.ref[i]  = vp8_rac_get_sint(c, 6);
280

    
281
    for (i = 0; i < 4; i++)
282
        s->lf_delta.mode[i] = vp8_rac_get_sint(c, 6);
283
}
284

    
285
static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size)
286
{
287
    const uint8_t *sizes = buf;
288
    int i;
289

    
290
    s->num_coeff_partitions = 1 << vp8_rac_get_uint(&s->c, 2);
291

    
292
    buf      += 3*(s->num_coeff_partitions-1);
293
    buf_size -= 3*(s->num_coeff_partitions-1);
294
    if (buf_size < 0)
295
        return -1;
296

    
297
    for (i = 0; i < s->num_coeff_partitions-1; i++) {
298
        int size = AV_RL24(sizes + 3*i);
299
        if (buf_size - size < 0)
300
            return -1;
301

    
302
        ff_vp56_init_range_decoder(&s->coeff_partition[i], buf, size);
303
        buf      += size;
304
        buf_size -= size;
305
    }
306
    ff_vp56_init_range_decoder(&s->coeff_partition[i], buf, buf_size);
307

    
308
    return 0;
309
}
310

    
311
static void get_quants(VP8Context *s)
312
{
313
    VP56RangeCoder *c = &s->c;
314
    int i, base_qi;
315

    
316
    int yac_qi     = vp8_rac_get_uint(c, 7);
317
    int ydc_delta  = vp8_rac_get_sint(c, 4);
318
    int y2dc_delta = vp8_rac_get_sint(c, 4);
319
    int y2ac_delta = vp8_rac_get_sint(c, 4);
320
    int uvdc_delta = vp8_rac_get_sint(c, 4);
321
    int uvac_delta = vp8_rac_get_sint(c, 4);
322

    
323
    for (i = 0; i < 4; i++) {
324
        if (s->segmentation.enabled) {
325
            base_qi = s->segmentation.base_quant[i];
326
            if (!s->segmentation.absolute_vals)
327
                base_qi += yac_qi;
328
        } else
329
            base_qi = yac_qi;
330

    
331
        s->qmat[i].luma_qmul[0]    =       vp8_dc_qlookup[av_clip(base_qi + ydc_delta , 0, 127)];
332
        s->qmat[i].luma_qmul[1]    =       vp8_ac_qlookup[av_clip(base_qi             , 0, 127)];
333
        s->qmat[i].luma_dc_qmul[0] =   2 * vp8_dc_qlookup[av_clip(base_qi + y2dc_delta, 0, 127)];
334
        s->qmat[i].luma_dc_qmul[1] = 155 * vp8_ac_qlookup[av_clip(base_qi + y2ac_delta, 0, 127)] / 100;
335
        s->qmat[i].chroma_qmul[0]  =       vp8_dc_qlookup[av_clip(base_qi + uvdc_delta, 0, 127)];
336
        s->qmat[i].chroma_qmul[1]  =       vp8_ac_qlookup[av_clip(base_qi + uvac_delta, 0, 127)];
337

    
338
        s->qmat[i].luma_dc_qmul[1] = FFMAX(s->qmat[i].luma_dc_qmul[1], 8);
339
        s->qmat[i].chroma_qmul[0]  = FFMIN(s->qmat[i].chroma_qmul[0], 132);
340
    }
341
}
342

    
343
/**
344
 * Determine which buffers golden and altref should be updated with after this frame.
345
 * The spec isn't clear here, so I'm going by my understanding of what libvpx does
346
 *
347
 * Intra frames update all 3 references
348
 * Inter frames update VP56_FRAME_PREVIOUS if the update_last flag is set
349
 * If the update (golden|altref) flag is set, it's updated with the current frame
350
 *      if update_last is set, and VP56_FRAME_PREVIOUS otherwise.
351
 * If the flag is not set, the number read means:
352
 *      0: no update
353
 *      1: VP56_FRAME_PREVIOUS
354
 *      2: update golden with altref, or update altref with golden
355
 */
356
static VP56Frame ref_to_update(VP8Context *s, int update, VP56Frame ref)
357
{
358
    VP56RangeCoder *c = &s->c;
359

    
360
    if (update)
361
        return VP56_FRAME_CURRENT;
362

    
363
    switch (vp8_rac_get_uint(c, 2)) {
364
    case 1:
365
        return VP56_FRAME_PREVIOUS;
366
    case 2:
367
        return (ref == VP56_FRAME_GOLDEN) ? VP56_FRAME_GOLDEN2 : VP56_FRAME_GOLDEN;
368
    }
369
    return VP56_FRAME_NONE;
370
}
371

    
372
static void update_refs(VP8Context *s)
373
{
374
    VP56RangeCoder *c = &s->c;
375

    
376
    int update_golden = vp8_rac_get(c);
377
    int update_altref = vp8_rac_get(c);
378

    
379
    s->update_golden = ref_to_update(s, update_golden, VP56_FRAME_GOLDEN);
380
    s->update_altref = ref_to_update(s, update_altref, VP56_FRAME_GOLDEN2);
381
}
382

    
383
static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)
384
{
385
    VP56RangeCoder *c = &s->c;
386
    int header_size, hscale, vscale, i, j, k, l, m, ret;
387
    int width  = s->avctx->width;
388
    int height = s->avctx->height;
389

    
390
    s->keyframe  = !(buf[0] & 1);
391
    s->profile   =  (buf[0]>>1) & 7;
392
    s->invisible = !(buf[0] & 0x10);
393
    header_size  = AV_RL24(buf) >> 5;
394
    buf      += 3;
395
    buf_size -= 3;
396

    
397
    if (s->profile > 3)
398
        av_log(s->avctx, AV_LOG_WARNING, "Unknown profile %d\n", s->profile);
399

    
400
    if (!s->profile)
401
        memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_epel_pixels_tab, sizeof(s->put_pixels_tab));
402
    else    // profile 1-3 use bilinear, 4+ aren't defined so whatever
403
        memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_bilinear_pixels_tab, sizeof(s->put_pixels_tab));
404

    
405
    if (header_size > buf_size - 7*s->keyframe) {
406
        av_log(s->avctx, AV_LOG_ERROR, "Header size larger than data provided\n");
407
        return AVERROR_INVALIDDATA;
408
    }
409

    
410
    if (s->keyframe) {
411
        if (AV_RL24(buf) != 0x2a019d) {
412
            av_log(s->avctx, AV_LOG_ERROR, "Invalid start code 0x%x\n", AV_RL24(buf));
413
            return AVERROR_INVALIDDATA;
414
        }
415
        width  = AV_RL16(buf+3) & 0x3fff;
416
        height = AV_RL16(buf+5) & 0x3fff;
417
        hscale = buf[4] >> 6;
418
        vscale = buf[6] >> 6;
419
        buf      += 7;
420
        buf_size -= 7;
421

    
422
        if (hscale || vscale)
423
            av_log_missing_feature(s->avctx, "Upscaling", 1);
424

    
425
        s->update_golden = s->update_altref = VP56_FRAME_CURRENT;
426
        for (i = 0; i < 4; i++)
427
            for (j = 0; j < 16; j++)
428
                memcpy(s->prob->token[i][j], vp8_token_default_probs[i][vp8_coeff_band[j]],
429
                       sizeof(s->prob->token[i][j]));
430
        memcpy(s->prob->pred16x16, vp8_pred16x16_prob_inter, sizeof(s->prob->pred16x16));
431
        memcpy(s->prob->pred8x8c , vp8_pred8x8c_prob_inter , sizeof(s->prob->pred8x8c));
432
        memcpy(s->prob->mvc      , vp8_mv_default_prob     , sizeof(s->prob->mvc));
433
        memset(&s->segmentation, 0, sizeof(s->segmentation));
434
    }
435

    
436
    if (!s->macroblocks_base || /* first frame */
437
        width != s->avctx->width || height != s->avctx->height) {
438
        if ((ret = update_dimensions(s, width, height) < 0))
439
            return ret;
440
    }
441

    
442
    ff_vp56_init_range_decoder(c, buf, header_size);
443
    buf      += header_size;
444
    buf_size -= header_size;
445

    
446
    if (s->keyframe) {
447
        if (vp8_rac_get(c))
448
            av_log(s->avctx, AV_LOG_WARNING, "Unspecified colorspace\n");
449
        vp8_rac_get(c); // whether we can skip clamping in dsp functions
450
    }
451

    
452
    if ((s->segmentation.enabled = vp8_rac_get(c)))
453
        parse_segment_info(s);
454
    else
455
        s->segmentation.update_map = 0; // FIXME: move this to some init function?
456

    
457
    s->filter.simple    = vp8_rac_get(c);
458
    s->filter.level     = vp8_rac_get_uint(c, 6);
459
    s->filter.sharpness = vp8_rac_get_uint(c, 3);
460

    
461
    if ((s->lf_delta.enabled = vp8_rac_get(c)))
462
        if (vp8_rac_get(c))
463
            update_lf_deltas(s);
464

    
465
    if (setup_partitions(s, buf, buf_size)) {
466
        av_log(s->avctx, AV_LOG_ERROR, "Invalid partitions\n");
467
        return AVERROR_INVALIDDATA;
468
    }
469

    
470
    get_quants(s);
471

    
472
    if (!s->keyframe) {
473
        update_refs(s);
474
        s->sign_bias[VP56_FRAME_GOLDEN]               = vp8_rac_get(c);
475
        s->sign_bias[VP56_FRAME_GOLDEN2 /* altref */] = vp8_rac_get(c);
476
    }
477

    
478
    // if we aren't saving this frame's probabilities for future frames,
479
    // make a copy of the current probabilities
480
    if (!(s->update_probabilities = vp8_rac_get(c)))
481
        s->prob[1] = s->prob[0];
482

    
483
    s->update_last = s->keyframe || vp8_rac_get(c);
484

    
485
    for (i = 0; i < 4; i++)
486
        for (j = 0; j < 8; j++)
487
            for (k = 0; k < 3; k++)
488
                for (l = 0; l < NUM_DCT_TOKENS-1; l++)
489
                    if (vp56_rac_get_prob_branchy(c, vp8_token_update_probs[i][j][k][l])) {
490
                        int prob = vp8_rac_get_uint(c, 8);
491
                        for (m = 0; vp8_coeff_band_indexes[j][m] >= 0; m++)
492
                            s->prob->token[i][vp8_coeff_band_indexes[j][m]][k][l] = prob;
493
                    }
494

    
495
    if ((s->mbskip_enabled = vp8_rac_get(c)))
496
        s->prob->mbskip = vp8_rac_get_uint(c, 8);
497

    
498
    if (!s->keyframe) {
499
        s->prob->intra  = vp8_rac_get_uint(c, 8);
500
        s->prob->last   = vp8_rac_get_uint(c, 8);
501
        s->prob->golden = vp8_rac_get_uint(c, 8);
502

    
503
        if (vp8_rac_get(c))
504
            for (i = 0; i < 4; i++)
505
                s->prob->pred16x16[i] = vp8_rac_get_uint(c, 8);
506
        if (vp8_rac_get(c))
507
            for (i = 0; i < 3; i++)
508
                s->prob->pred8x8c[i]  = vp8_rac_get_uint(c, 8);
509

    
510
        // 17.2 MV probability update
511
        for (i = 0; i < 2; i++)
512
            for (j = 0; j < 19; j++)
513
                if (vp56_rac_get_prob_branchy(c, vp8_mv_update_prob[i][j]))
514
                    s->prob->mvc[i][j] = vp8_rac_get_nn(c);
515
    }
516

    
517
    return 0;
518
}
519

    
520
static av_always_inline
521
void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src, int mb_x, int mb_y)
522
{
523
#define MARGIN (16 << 2)
524
    dst->x = av_clip(src->x, -((mb_x << 6) + MARGIN),
525
                     ((s->mb_width  - 1 - mb_x) << 6) + MARGIN);
526
    dst->y = av_clip(src->y, -((mb_y << 6) + MARGIN),
527
                     ((s->mb_height - 1 - mb_y) << 6) + MARGIN);
528
}
529

    
530
static av_always_inline
531
void find_near_mvs(VP8Context *s, VP8Macroblock *mb,
532
                   VP56mv near[2], VP56mv *best, uint8_t cnt[4])
533
{
534
    VP8Macroblock *mb_edge[3] = { mb + 2 /* top */,
535
                                  mb - 1 /* left */,
536
                                  mb + 1 /* top-left */ };
537
    enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT };
538
    VP56mv near_mv[4]  = {{ 0 }};
539
    enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV };
540
    int idx = CNT_ZERO;
541
    int best_idx = CNT_ZERO;
542
    int cur_sign_bias = s->sign_bias[mb->ref_frame];
543
    int *sign_bias = s->sign_bias;
544

    
545
    /* Process MB on top, left and top-left */
546
    #define MV_EDGE_CHECK(n)\
547
    {\
548
        VP8Macroblock *edge = mb_edge[n];\
549
        int edge_ref = edge->ref_frame;\
550
        if (edge_ref != VP56_FRAME_CURRENT) {\
551
            uint32_t mv = AV_RN32A(&edge->mv);\
552
            if (mv) {\
553
                if (cur_sign_bias != sign_bias[edge_ref]) {\
554
                    /* SWAR negate of the values in mv. */\
555
                    mv = ~mv;\
556
                    mv = ((mv&0x7fff7fff) + 0x00010001) ^ (mv&0x80008000);\
557
                }\
558
                if (!n || mv != AV_RN32A(&near_mv[idx]))\
559
                    AV_WN32A(&near_mv[++idx], mv);\
560
                cnt[idx]      += 1 + (n != 2);\
561
            } else\
562
                cnt[CNT_ZERO] += 1 + (n != 2);\
563
        }\
564
    }
565
    MV_EDGE_CHECK(0)
566
    MV_EDGE_CHECK(1)
567
    MV_EDGE_CHECK(2)
568

    
569
    /* If we have three distinct MVs, merge first and last if they're the same */
570
    if (cnt[CNT_SPLITMV] && AV_RN32A(&near_mv[1+EDGE_TOP]) == AV_RN32A(&near_mv[1+EDGE_TOPLEFT]))
571
        cnt[CNT_NEAREST] += 1;
572

    
573
    cnt[CNT_SPLITMV] = ((mb_edge[EDGE_LEFT]->mode   == VP8_MVMODE_SPLIT) +
574
                        (mb_edge[EDGE_TOP]->mode    == VP8_MVMODE_SPLIT)) * 2 +
575
                       (mb_edge[EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT);
576

    
577
    /* Swap near and nearest if necessary */
578
    if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) {
579
        FFSWAP(uint8_t,     cnt[CNT_NEAREST],     cnt[CNT_NEAR]);
580
        FFSWAP( VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]);
581
    }
582

    
583
    /* Choose the best mv out of 0,0 and the nearest mv */
584
    if (cnt[CNT_NEAREST] >= cnt[CNT_ZERO])
585
        best_idx = CNT_NEAREST;
586

    
587
    mb->mv  = near_mv[best_idx];
588
    near[0] = near_mv[CNT_NEAREST];
589
    near[1] = near_mv[CNT_NEAR];
590
}
591

    
592
/**
593
 * Motion vector coding, 17.1.
594
 */
595
static int read_mv_component(VP56RangeCoder *c, const uint8_t *p)
596
{
597
    int bit, x = 0;
598

    
599
    if (vp56_rac_get_prob_branchy(c, p[0])) {
600
        int i;
601

    
602
        for (i = 0; i < 3; i++)
603
            x += vp56_rac_get_prob(c, p[9 + i]) << i;
604
        for (i = 9; i > 3; i--)
605
            x += vp56_rac_get_prob(c, p[9 + i]) << i;
606
        if (!(x & 0xFFF0) || vp56_rac_get_prob(c, p[12]))
607
            x += 8;
608
    } else {
609
        // small_mvtree
610
        const uint8_t *ps = p+2;
611
        bit = vp56_rac_get_prob(c, *ps);
612
        ps += 1 + 3*bit;
613
        x  += 4*bit;
614
        bit = vp56_rac_get_prob(c, *ps);
615
        ps += 1 + bit;
616
        x  += 2*bit;
617
        x  += vp56_rac_get_prob(c, *ps);
618
    }
619

    
620
    return (x && vp56_rac_get_prob(c, p[1])) ? -x : x;
621
}
622

    
623
static av_always_inline
624
const uint8_t *get_submv_prob(uint32_t left, uint32_t top)
625
{
626
    if (left == top)
627
        return vp8_submv_prob[4-!!left];
628
    if (!top)
629
        return vp8_submv_prob[2];
630
    return vp8_submv_prob[1-!!left];
631
}
632

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

    
651
    if (vp56_rac_get_prob_branchy(c, vp8_mbsplit_prob[0])) {
652
        if (vp56_rac_get_prob_branchy(c, vp8_mbsplit_prob[1])) {
653
            part_idx = VP8_SPLITMVMODE_16x8 + vp56_rac_get_prob(c, vp8_mbsplit_prob[2]);
654
        } else {
655
            part_idx = VP8_SPLITMVMODE_8x8;
656
        }
657
    } else {
658
        part_idx = VP8_SPLITMVMODE_4x4;
659
    }
660

    
661
    num = vp8_mbsplit_count[part_idx];
662
    mbsplits_cur = vp8_mbsplits[part_idx],
663
    firstidx = vp8_mbfirstidx[part_idx];
664
    mb->partitioning = part_idx;
665

    
666
    for (n = 0; n < num; n++) {
667
        int k = firstidx[n];
668
        uint32_t left, above;
669
        const uint8_t *submv_prob;
670

    
671
        if (!(k & 3))
672
            left = AV_RN32A(&left_mv[mbsplits_left[k + 3]]);
673
        else
674
            left  = AV_RN32A(&cur_mv[mbsplits_cur[k - 1]]);
675
        if (k <= 3)
676
            above = AV_RN32A(&top_mv[mbsplits_top[k + 12]]);
677
        else
678
            above = AV_RN32A(&cur_mv[mbsplits_cur[k - 4]]);
679

    
680
        submv_prob = get_submv_prob(left, above);
681

    
682
        if (vp56_rac_get_prob_branchy(c, submv_prob[0])) {
683
            if (vp56_rac_get_prob_branchy(c, submv_prob[1])) {
684
                if (vp56_rac_get_prob_branchy(c, submv_prob[2])) {
685
                    mb->bmv[n].y = mb->mv.y + read_mv_component(c, s->prob->mvc[0]);
686
                    mb->bmv[n].x = mb->mv.x + read_mv_component(c, s->prob->mvc[1]);
687
                } else {
688
                    AV_ZERO32(&mb->bmv[n]);
689
                }
690
            } else {
691
                AV_WN32A(&mb->bmv[n], above);
692
            }
693
        } else {
694
            AV_WN32A(&mb->bmv[n], left);
695
        }
696
    }
697

    
698
    return num;
699
}
700

    
701
static av_always_inline
702
void decode_intra4x4_modes(VP8Context *s, VP56RangeCoder *c,
703
                           int mb_x, int keyframe)
704
{
705
    uint8_t *intra4x4 = s->intra4x4_pred_mode_mb;
706
    if (keyframe) {
707
        int x, y;
708
        uint8_t* const top = s->intra4x4_pred_mode_top + 4 * mb_x;
709
        uint8_t* const left = s->intra4x4_pred_mode_left;
710
        for (y = 0; y < 4; y++) {
711
            for (x = 0; x < 4; x++) {
712
                const uint8_t *ctx;
713
                ctx = vp8_pred4x4_prob_intra[top[x]][left[y]];
714
                *intra4x4 = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx);
715
                left[y] = top[x] = *intra4x4;
716
                intra4x4++;
717
            }
718
        }
719
    } else {
720
        int i;
721
        for (i = 0; i < 16; i++)
722
            intra4x4[i] = vp8_rac_get_tree(c, vp8_pred4x4_tree, vp8_pred4x4_prob_inter);
723
    }
724
}
725

    
726
static av_always_inline
727
void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, uint8_t *segment)
728
{
729
    VP56RangeCoder *c = &s->c;
730

    
731
    if (s->segmentation.update_map)
732
        *segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid);
733
    s->segment = *segment;
734

    
735
    mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0;
736

    
737
    if (s->keyframe) {
738
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra);
739

    
740
        if (mb->mode == MODE_I4x4) {
741
            decode_intra4x4_modes(s, c, mb_x, 1);
742
        } else {
743
            const uint32_t modes = vp8_pred4x4_mode[mb->mode] * 0x01010101u;
744
            AV_WN32A(s->intra4x4_pred_mode_top + 4 * mb_x, modes);
745
            AV_WN32A(s->intra4x4_pred_mode_left, modes);
746
        }
747

    
748
        s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra);
749
        mb->ref_frame = VP56_FRAME_CURRENT;
750
    } else if (vp56_rac_get_prob_branchy(c, s->prob->intra)) {
751
        VP56mv near[2], best;
752
        uint8_t cnt[4] = { 0 };
753

    
754
        // inter MB, 16.2
755
        if (vp56_rac_get_prob_branchy(c, s->prob->last))
756
            mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ?
757
                VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN;
758
        else
759
            mb->ref_frame = VP56_FRAME_PREVIOUS;
760
        s->ref_count[mb->ref_frame-1]++;
761

    
762
        // motion vectors, 16.3
763
        find_near_mvs(s, mb, near, &best, cnt);
764
        if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[0]][0])) {
765
            if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[1]][1])) {
766
                if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[2]][2])) {
767
                    if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[3]][3])) {
768
                        mb->mode = VP8_MVMODE_SPLIT;
769
                        clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y);
770
                        mb->mv = mb->bmv[decode_splitmvs(s, c, mb) - 1];
771
                    } else {
772
                        mb->mode = VP8_MVMODE_NEW;
773
                        clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y);
774
                        mb->mv.y += read_mv_component(c, s->prob->mvc[0]);
775
                        mb->mv.x += read_mv_component(c, s->prob->mvc[1]);
776
                    }
777
                } else {
778
                    mb->mode = VP8_MVMODE_NEAR;
779
                    clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y);
780
                }
781
            } else {
782
                mb->mode = VP8_MVMODE_NEAREST;
783
                clamp_mv(s, &mb->mv, &near[0], mb_x, mb_y);
784
            }
785
        } else {
786
            mb->mode = VP8_MVMODE_ZERO;
787
            AV_ZERO32(&mb->mv);
788
        }
789
        if (mb->mode != VP8_MVMODE_SPLIT) {
790
            mb->partitioning = VP8_SPLITMVMODE_NONE;
791
            mb->bmv[0] = mb->mv;
792
        }
793
    } else {
794
        // intra MB, 16.1
795
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16);
796

    
797
        if (mb->mode == MODE_I4x4)
798
            decode_intra4x4_modes(s, c, mb_x, 0);
799

    
800
        s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c);
801
        mb->ref_frame = VP56_FRAME_CURRENT;
802
        mb->partitioning = VP8_SPLITMVMODE_NONE;
803
        AV_ZERO32(&mb->bmv[0]);
804
    }
805
}
806

    
807
/**
808
 * @param c arithmetic bitstream reader context
809
 * @param block destination for block coefficients
810
 * @param probs probabilities to use when reading trees from the bitstream
811
 * @param i initial coeff index, 0 unless a separate DC block is coded
812
 * @param zero_nhood the initial prediction context for number of surrounding
813
 *                   all-zero blocks (only left/top, so 0-2)
814
 * @param qmul array holding the dc/ac dequant factor at position 0/1
815
 * @return 0 if no coeffs were decoded
816
 *         otherwise, the index of the last coeff decoded plus one
817
 */
818
static int decode_block_coeffs_internal(VP56RangeCoder *c, DCTELEM block[16],
819
                                        uint8_t probs[8][3][NUM_DCT_TOKENS-1],
820
                                        int i, uint8_t *token_prob, int16_t qmul[2])
821
{
822
    goto skip_eob;
823
    do {
824
        int coeff;
825
        if (!vp56_rac_get_prob_branchy(c, token_prob[0]))   // DCT_EOB
826
            return i;
827

    
828
skip_eob:
829
        if (!vp56_rac_get_prob_branchy(c, token_prob[1])) { // DCT_0
830
            if (++i == 16)
831
                return i; // invalid input; blocks should end with EOB
832
            token_prob = probs[i][0];
833
            goto skip_eob;
834
        }
835

    
836
        if (!vp56_rac_get_prob_branchy(c, token_prob[2])) { // DCT_1
837
            coeff = 1;
838
            token_prob = probs[i+1][1];
839
        } else {
840
            if (!vp56_rac_get_prob_branchy(c, token_prob[3])) { // DCT 2,3,4
841
                coeff = vp56_rac_get_prob_branchy(c, token_prob[4]);
842
                if (coeff)
843
                    coeff += vp56_rac_get_prob(c, token_prob[5]);
844
                coeff += 2;
845
            } else {
846
                // DCT_CAT*
847
                if (!vp56_rac_get_prob_branchy(c, token_prob[6])) {
848
                    if (!vp56_rac_get_prob_branchy(c, token_prob[7])) { // DCT_CAT1
849
                        coeff  = 5 + vp56_rac_get_prob(c, vp8_dct_cat1_prob[0]);
850
                    } else {                                    // DCT_CAT2
851
                        coeff  = 7;
852
                        coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[0]) << 1;
853
                        coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[1]);
854
                    }
855
                } else {    // DCT_CAT3 and up
856
                    int a = vp56_rac_get_prob(c, token_prob[8]);
857
                    int b = vp56_rac_get_prob(c, token_prob[9+a]);
858
                    int cat = (a<<1) + b;
859
                    coeff  = 3 + (8<<cat);
860
                    coeff += vp8_rac_get_coeff(c, vp8_dct_cat_prob[cat]);
861
                }
862
            }
863
            token_prob = probs[i+1][2];
864
        }
865
        block[zigzag_scan[i]] = (vp8_rac_get(c) ? -coeff : coeff) * qmul[!!i];
866
    } while (++i < 16);
867

    
868
    return i;
869
}
870

    
871
static av_always_inline
872
int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16],
873
                        uint8_t probs[8][3][NUM_DCT_TOKENS-1],
874
                        int i, int zero_nhood, int16_t qmul[2])
875
{
876
    uint8_t *token_prob = probs[i][zero_nhood];
877
    if (!vp56_rac_get_prob_branchy(c, token_prob[0]))   // DCT_EOB
878
        return 0;
879
    return decode_block_coeffs_internal(c, block, probs, i, token_prob, qmul);
880
}
881

    
882
static av_always_inline
883
void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,
884
                      uint8_t t_nnz[9], uint8_t l_nnz[9])
885
{
886
    int i, x, y, luma_start = 0, luma_ctx = 3;
887
    int nnz_pred, nnz, nnz_total = 0;
888
    int segment = s->segment;
889
    int block_dc = 0;
890

    
891
    if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
892
        nnz_pred = t_nnz[8] + l_nnz[8];
893

    
894
        // decode DC values and do hadamard
895
        nnz = decode_block_coeffs(c, s->block_dc, s->prob->token[1], 0, nnz_pred,
896
                                  s->qmat[segment].luma_dc_qmul);
897
        l_nnz[8] = t_nnz[8] = !!nnz;
898
        if (nnz) {
899
            nnz_total += nnz;
900
            block_dc = 1;
901
            if (nnz == 1)
902
                s->vp8dsp.vp8_luma_dc_wht_dc(s->block, s->block_dc);
903
            else
904
                s->vp8dsp.vp8_luma_dc_wht(s->block, s->block_dc);
905
        }
906
        luma_start = 1;
907
        luma_ctx = 0;
908
    }
909

    
910
    // luma blocks
911
    for (y = 0; y < 4; y++)
912
        for (x = 0; x < 4; x++) {
913
            nnz_pred = l_nnz[y] + t_nnz[x];
914
            nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start,
915
                                      nnz_pred, s->qmat[segment].luma_qmul);
916
            // nnz+block_dc may be one more than the actual last index, but we don't care
917
            s->non_zero_count_cache[y][x] = nnz + block_dc;
918
            t_nnz[x] = l_nnz[y] = !!nnz;
919
            nnz_total += nnz;
920
        }
921

    
922
    // chroma blocks
923
    // TODO: what to do about dimensions? 2nd dim for luma is x,
924
    // but for chroma it's (y<<1)|x
925
    for (i = 4; i < 6; i++)
926
        for (y = 0; y < 2; y++)
927
            for (x = 0; x < 2; x++) {
928
                nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x];
929
                nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0,
930
                                          nnz_pred, s->qmat[segment].chroma_qmul);
931
                s->non_zero_count_cache[i][(y<<1)+x] = nnz;
932
                t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz;
933
                nnz_total += nnz;
934
            }
935

    
936
    // if there were no coded coeffs despite the macroblock not being marked skip,
937
    // we MUST not do the inner loop filter and should not do IDCT
938
    // Since skip isn't used for bitstream prediction, just manually set it.
939
    if (!nnz_total)
940
        mb->skip = 1;
941
}
942

    
943
static av_always_inline
944
void backup_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
945
                      int linesize, int uvlinesize, int simple)
946
{
947
    AV_COPY128(top_border, src_y + 15*linesize);
948
    if (!simple) {
949
        AV_COPY64(top_border+16, src_cb + 7*uvlinesize);
950
        AV_COPY64(top_border+24, src_cr + 7*uvlinesize);
951
    }
952
}
953

    
954
static av_always_inline
955
void xchg_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
956
                    int linesize, int uvlinesize, int mb_x, int mb_y, int mb_width,
957
                    int simple, int xchg)
958
{
959
    uint8_t *top_border_m1 = top_border-32;     // for TL prediction
960
    src_y  -=   linesize;
961
    src_cb -= uvlinesize;
962
    src_cr -= uvlinesize;
963

    
964
#define XCHG(a,b,xchg) do {                     \
965
        if (xchg) AV_SWAP64(b,a);               \
966
        else      AV_COPY64(b,a);               \
967
    } while (0)
968

    
969
    XCHG(top_border_m1+8, src_y-8, xchg);
970
    XCHG(top_border,      src_y,   xchg);
971
    XCHG(top_border+8,    src_y+8, 1);
972
    if (mb_x < mb_width-1)
973
        XCHG(top_border+32, src_y+16, 1);
974

    
975
    // only copy chroma for normal loop filter
976
    // or to initialize the top row to 127
977
    if (!simple || !mb_y) {
978
        XCHG(top_border_m1+16, src_cb-8, xchg);
979
        XCHG(top_border_m1+24, src_cr-8, xchg);
980
        XCHG(top_border+16,    src_cb, 1);
981
        XCHG(top_border+24,    src_cr, 1);
982
    }
983
}
984

    
985
static av_always_inline
986
int check_dc_pred8x8_mode(int mode, int mb_x, int mb_y)
987
{
988
    if (!mb_x) {
989
        return mb_y ? TOP_DC_PRED8x8 : DC_128_PRED8x8;
990
    } else {
991
        return mb_y ? mode : LEFT_DC_PRED8x8;
992
    }
993
}
994

    
995
static av_always_inline
996
int check_tm_pred8x8_mode(int mode, int mb_x, int mb_y)
997
{
998
    if (!mb_x) {
999
        return mb_y ? VERT_PRED8x8 : DC_129_PRED8x8;
1000
    } else {
1001
        return mb_y ? mode : HOR_PRED8x8;
1002
    }
1003
}
1004

    
1005
static av_always_inline
1006
int check_intra_pred8x8_mode(int mode, int mb_x, int mb_y)
1007
{
1008
    if (mode == DC_PRED8x8) {
1009
        return check_dc_pred8x8_mode(mode, mb_x, mb_y);
1010
    } else {
1011
        return mode;
1012
    }
1013
}
1014

    
1015
static av_always_inline
1016
int check_intra_pred8x8_mode_emuedge(int mode, int mb_x, int mb_y)
1017
{
1018
    switch (mode) {
1019
    case DC_PRED8x8:
1020
        return check_dc_pred8x8_mode(mode, mb_x, mb_y);
1021
    case VERT_PRED8x8:
1022
        return !mb_y ? DC_127_PRED8x8 : mode;
1023
    case HOR_PRED8x8:
1024
        return !mb_x ? DC_129_PRED8x8 : mode;
1025
    case PLANE_PRED8x8 /*TM*/:
1026
        return check_tm_pred8x8_mode(mode, mb_x, mb_y);
1027
    }
1028
    return mode;
1029
}
1030

    
1031
static av_always_inline
1032
int check_tm_pred4x4_mode(int mode, int mb_x, int mb_y)
1033
{
1034
    if (!mb_x) {
1035
        return mb_y ? VERT_VP8_PRED : DC_129_PRED;
1036
    } else {
1037
        return mb_y ? mode : HOR_VP8_PRED;
1038
    }
1039
}
1040

    
1041
static av_always_inline
1042
int check_intra_pred4x4_mode_emuedge(int mode, int mb_x, int mb_y, int *copy_buf)
1043
{
1044
    switch (mode) {
1045
    case VERT_PRED:
1046
        if (!mb_x && mb_y) {
1047
            *copy_buf = 1;
1048
            return mode;
1049
        }
1050
        /* fall-through */
1051
    case DIAG_DOWN_LEFT_PRED:
1052
    case VERT_LEFT_PRED:
1053
        return !mb_y ? DC_127_PRED : mode;
1054
    case HOR_PRED:
1055
        if (!mb_y) {
1056
            *copy_buf = 1;
1057
            return mode;
1058
        }
1059
        /* fall-through */
1060
    case HOR_UP_PRED:
1061
        return !mb_x ? DC_129_PRED : mode;
1062
    case TM_VP8_PRED:
1063
        return check_tm_pred4x4_mode(mode, mb_x, mb_y);
1064
    case DC_PRED: // 4x4 DC doesn't use the same "H.264-style" exceptions as 16x16/8x8 DC
1065
    case DIAG_DOWN_RIGHT_PRED:
1066
    case VERT_RIGHT_PRED:
1067
    case HOR_DOWN_PRED:
1068
        if (!mb_y || !mb_x)
1069
            *copy_buf = 1;
1070
        return mode;
1071
    }
1072
    return mode;
1073
}
1074

    
1075
static av_always_inline
1076
void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
1077
                   int mb_x, int mb_y)
1078
{
1079
    AVCodecContext *avctx = s->avctx;
1080
    int x, y, mode, nnz, tr;
1081

    
1082
    // for the first row, we need to run xchg_mb_border to init the top edge to 127
1083
    // otherwise, skip it if we aren't going to deblock
1084
    if (!(avctx->flags & CODEC_FLAG_EMU_EDGE && !mb_y) && (s->deblock_filter || !mb_y))
1085
        xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
1086
                       s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
1087
                       s->filter.simple, 1);
1088

    
1089
    if (mb->mode < MODE_I4x4) {
1090
        if (avctx->flags & CODEC_FLAG_EMU_EDGE) { // tested
1091
            mode = check_intra_pred8x8_mode_emuedge(mb->mode, mb_x, mb_y);
1092
        } else {
1093
            mode = check_intra_pred8x8_mode(mb->mode, mb_x, mb_y);
1094
        }
1095
        s->hpc.pred16x16[mode](dst[0], s->linesize);
1096
    } else {
1097
        uint8_t *ptr = dst[0];
1098
        uint8_t *intra4x4 = s->intra4x4_pred_mode_mb;
1099
        uint8_t tr_top[4] = { 127, 127, 127, 127 };
1100

    
1101
        // all blocks on the right edge of the macroblock use bottom edge
1102
        // the top macroblock for their topright edge
1103
        uint8_t *tr_right = ptr - s->linesize + 16;
1104

    
1105
        // if we're on the right edge of the frame, said edge is extended
1106
        // from the top macroblock
1107
        if (!(!mb_y && avctx->flags & CODEC_FLAG_EMU_EDGE) &&
1108
            mb_x == s->mb_width-1) {
1109
            tr = tr_right[-1]*0x01010101;
1110
            tr_right = (uint8_t *)&tr;
1111
        }
1112

    
1113
        if (mb->skip)
1114
            AV_ZERO128(s->non_zero_count_cache);
1115

    
1116
        for (y = 0; y < 4; y++) {
1117
            uint8_t *topright = ptr + 4 - s->linesize;
1118
            for (x = 0; x < 4; x++) {
1119
                int copy = 0, linesize = s->linesize;
1120
                uint8_t *dst = ptr+4*x;
1121
                DECLARE_ALIGNED(4, uint8_t, copy_dst)[5*8];
1122

    
1123
                if ((y == 0 || x == 3) && mb_y == 0 && avctx->flags & CODEC_FLAG_EMU_EDGE) {
1124
                    topright = tr_top;
1125
                } else if (x == 3)
1126
                    topright = tr_right;
1127

    
1128
                if (avctx->flags & CODEC_FLAG_EMU_EDGE) { // mb_x+x or mb_y+y is a hack but works
1129
                    mode = check_intra_pred4x4_mode_emuedge(intra4x4[x], mb_x + x, mb_y + y, &copy);
1130
                    if (copy) {
1131
                        dst = copy_dst + 12;
1132
                        linesize = 8;
1133
                        if (!(mb_y + y)) {
1134
                            copy_dst[3] = 127U;
1135
                            AV_WN32A(copy_dst+4, 127U * 0x01010101U);
1136
                        } else {
1137
                            AV_COPY32(copy_dst+4, ptr+4*x-s->linesize);
1138
                            if (!(mb_x + x)) {
1139
                                copy_dst[3] = 129U;
1140
                            } else {
1141
                                copy_dst[3] = ptr[4*x-s->linesize-1];
1142
                            }
1143
                        }
1144
                        if (!(mb_x + x)) {
1145
                            copy_dst[11] =
1146
                            copy_dst[19] =
1147
                            copy_dst[27] =
1148
                            copy_dst[35] = 129U;
1149
                        } else {
1150
                            copy_dst[11] = ptr[4*x              -1];
1151
                            copy_dst[19] = ptr[4*x+s->linesize  -1];
1152
                            copy_dst[27] = ptr[4*x+s->linesize*2-1];
1153
                            copy_dst[35] = ptr[4*x+s->linesize*3-1];
1154
                        }
1155
                    }
1156
                } else {
1157
                    mode = intra4x4[x];
1158
                }
1159
                s->hpc.pred4x4[mode](dst, topright, linesize);
1160
                if (copy) {
1161
                    AV_COPY32(ptr+4*x              , copy_dst+12);
1162
                    AV_COPY32(ptr+4*x+s->linesize  , copy_dst+20);
1163
                    AV_COPY32(ptr+4*x+s->linesize*2, copy_dst+28);
1164
                    AV_COPY32(ptr+4*x+s->linesize*3, copy_dst+36);
1165
                }
1166

    
1167
                nnz = s->non_zero_count_cache[y][x];
1168
                if (nnz) {
1169
                    if (nnz == 1)
1170
                        s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize);
1171
                    else
1172
                        s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize);
1173
                }
1174
                topright += 4;
1175
            }
1176

    
1177
            ptr   += 4*s->linesize;
1178
            intra4x4 += 4;
1179
        }
1180
    }
1181

    
1182
    if (avctx->flags & CODEC_FLAG_EMU_EDGE) {
1183
        mode = check_intra_pred8x8_mode_emuedge(s->chroma_pred_mode, mb_x, mb_y);
1184
    } else {
1185
        mode = check_intra_pred8x8_mode(s->chroma_pred_mode, mb_x, mb_y);
1186
    }
1187
    s->hpc.pred8x8[mode](dst[1], s->uvlinesize);
1188
    s->hpc.pred8x8[mode](dst[2], s->uvlinesize);
1189

    
1190
    if (!(avctx->flags & CODEC_FLAG_EMU_EDGE && !mb_y) && (s->deblock_filter || !mb_y))
1191
        xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
1192
                       s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
1193
                       s->filter.simple, 0);
1194
}
1195

    
1196
static const uint8_t subpel_idx[3][8] = {
1197
    { 0, 1, 2, 1, 2, 1, 2, 1 }, // nr. of left extra pixels,
1198
                                // also function pointer index
1199
    { 0, 3, 5, 3, 5, 3, 5, 3 }, // nr. of extra pixels required
1200
    { 0, 2, 3, 2, 3, 2, 3, 2 }, // nr. of right extra pixels
1201
};
1202

    
1203
/**
1204
 * Generic MC function.
1205
 *
1206
 * @param s VP8 decoding context
1207
 * @param luma 1 for luma (Y) planes, 0 for chroma (Cb/Cr) planes
1208
 * @param dst target buffer for block data at block position
1209
 * @param src reference picture buffer at origin (0, 0)
1210
 * @param mv motion vector (relative to block position) to get pixel data from
1211
 * @param x_off horizontal position of block from origin (0, 0)
1212
 * @param y_off vertical position of block from origin (0, 0)
1213
 * @param block_w width of block (16, 8 or 4)
1214
 * @param block_h height of block (always same as block_w)
1215
 * @param width width of src/dst plane data
1216
 * @param height height of src/dst plane data
1217
 * @param linesize size of a single line of plane data, including padding
1218
 * @param mc_func motion compensation function pointers (bilinear or sixtap MC)
1219
 */
1220
static av_always_inline
1221
void vp8_mc_luma(VP8Context *s, uint8_t *dst, uint8_t *src, const VP56mv *mv,
1222
                 int x_off, int y_off, int block_w, int block_h,
1223
                 int width, int height, int linesize,
1224
                 vp8_mc_func mc_func[3][3])
1225
{
1226
    if (AV_RN32A(mv)) {
1227

    
1228
        int mx = (mv->x << 1)&7, mx_idx = subpel_idx[0][mx];
1229
        int my = (mv->y << 1)&7, my_idx = subpel_idx[0][my];
1230

    
1231
        x_off += mv->x >> 2;
1232
        y_off += mv->y >> 2;
1233

    
1234
        // edge emulation
1235
        src += y_off * linesize + x_off;
1236
        if (x_off < mx_idx || x_off >= width  - block_w - subpel_idx[2][mx] ||
1237
            y_off < my_idx || y_off >= height - block_h - subpel_idx[2][my]) {
1238
            s->dsp.emulated_edge_mc(s->edge_emu_buffer, src - my_idx * linesize - mx_idx, linesize,
1239
                                    block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],
1240
                                    x_off - mx_idx, y_off - my_idx, width, height);
1241
            src = s->edge_emu_buffer + mx_idx + linesize * my_idx;
1242
        }
1243
        mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);
1244
    } else
1245
        mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0);
1246
}
1247

    
1248
static av_always_inline
1249
void vp8_mc_chroma(VP8Context *s, uint8_t *dst1, uint8_t *dst2, uint8_t *src1,
1250
                   uint8_t *src2, const VP56mv *mv, int x_off, int y_off,
1251
                   int block_w, int block_h, int width, int height, int linesize,
1252
                   vp8_mc_func mc_func[3][3])
1253
{
1254
    if (AV_RN32A(mv)) {
1255
        int mx = mv->x&7, mx_idx = subpel_idx[0][mx];
1256
        int my = mv->y&7, my_idx = subpel_idx[0][my];
1257

    
1258
        x_off += mv->x >> 3;
1259
        y_off += mv->y >> 3;
1260

    
1261
        // edge emulation
1262
        src1 += y_off * linesize + x_off;
1263
        src2 += y_off * linesize + x_off;
1264
        if (x_off < mx_idx || x_off >= width  - block_w - subpel_idx[2][mx] ||
1265
            y_off < my_idx || y_off >= height - block_h - subpel_idx[2][my]) {
1266
            s->dsp.emulated_edge_mc(s->edge_emu_buffer, src1 - my_idx * linesize - mx_idx, linesize,
1267
                                    block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],
1268
                                    x_off - mx_idx, y_off - my_idx, width, height);
1269
            src1 = s->edge_emu_buffer + mx_idx + linesize * my_idx;
1270
            mc_func[my_idx][mx_idx](dst1, linesize, src1, linesize, block_h, mx, my);
1271

    
1272
            s->dsp.emulated_edge_mc(s->edge_emu_buffer, src2 - my_idx * linesize - mx_idx, linesize,
1273
                                    block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],
1274
                                    x_off - mx_idx, y_off - my_idx, width, height);
1275
            src2 = s->edge_emu_buffer + mx_idx + linesize * my_idx;
1276
            mc_func[my_idx][mx_idx](dst2, linesize, src2, linesize, block_h, mx, my);
1277
        } else {
1278
            mc_func[my_idx][mx_idx](dst1, linesize, src1, linesize, block_h, mx, my);
1279
            mc_func[my_idx][mx_idx](dst2, linesize, src2, linesize, block_h, mx, my);
1280
        }
1281
    } else {
1282
        mc_func[0][0](dst1, linesize, src1 + y_off * linesize + x_off, linesize, block_h, 0, 0);
1283
        mc_func[0][0](dst2, linesize, src2 + y_off * linesize + x_off, linesize, block_h, 0, 0);
1284
    }
1285
}
1286

    
1287
static av_always_inline
1288
void vp8_mc_part(VP8Context *s, uint8_t *dst[3],
1289
                 AVFrame *ref_frame, int x_off, int y_off,
1290
                 int bx_off, int by_off,
1291
                 int block_w, int block_h,
1292
                 int width, int height, VP56mv *mv)
1293
{
1294
    VP56mv uvmv = *mv;
1295

    
1296
    /* Y */
1297
    vp8_mc_luma(s, dst[0] + by_off * s->linesize + bx_off,
1298
                ref_frame->data[0], mv, x_off + bx_off, y_off + by_off,
1299
                block_w, block_h, width, height, s->linesize,
1300
                s->put_pixels_tab[block_w == 8]);
1301

    
1302
    /* U/V */
1303
    if (s->profile == 3) {
1304
        uvmv.x &= ~7;
1305
        uvmv.y &= ~7;
1306
    }
1307
    x_off   >>= 1; y_off   >>= 1;
1308
    bx_off  >>= 1; by_off  >>= 1;
1309
    width   >>= 1; height  >>= 1;
1310
    block_w >>= 1; block_h >>= 1;
1311
    vp8_mc_chroma(s, dst[1] + by_off * s->uvlinesize + bx_off,
1312
                  dst[2] + by_off * s->uvlinesize + bx_off, ref_frame->data[1],
1313
                  ref_frame->data[2], &uvmv, x_off + bx_off, y_off + by_off,
1314
                  block_w, block_h, width, height, s->uvlinesize,
1315
                  s->put_pixels_tab[1 + (block_w == 4)]);
1316
}
1317

    
1318
/* Fetch pixels for estimated mv 4 macroblocks ahead.
1319
 * Optimized for 64-byte cache lines.  Inspired by ffh264 prefetch_motion. */
1320
static av_always_inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int mb_xy, int ref)
1321
{
1322
    /* Don't prefetch refs that haven't been used very often this frame. */
1323
    if (s->ref_count[ref-1] > (mb_xy >> 5)) {
1324
        int x_off = mb_x << 4, y_off = mb_y << 4;
1325
        int mx = (mb->mv.x>>2) + x_off + 8;
1326
        int my = (mb->mv.y>>2) + y_off;
1327
        uint8_t **src= s->framep[ref]->data;
1328
        int off= mx + (my + (mb_x&3)*4)*s->linesize + 64;
1329
        s->dsp.prefetch(src[0]+off, s->linesize, 4);
1330
        off= (mx>>1) + ((my>>1) + (mb_x&7))*s->uvlinesize + 64;
1331
        s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1332
    }
1333
}
1334

    
1335
/**
1336
 * Apply motion vectors to prediction buffer, chapter 18.
1337
 */
1338
static av_always_inline
1339
void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
1340
                   int mb_x, int mb_y)
1341
{
1342
    int x_off = mb_x << 4, y_off = mb_y << 4;
1343
    int width = 16*s->mb_width, height = 16*s->mb_height;
1344
    AVFrame *ref = s->framep[mb->ref_frame];
1345
    VP56mv *bmv = mb->bmv;
1346

    
1347
    switch (mb->partitioning) {
1348
    case VP8_SPLITMVMODE_NONE:
1349
        vp8_mc_part(s, dst, ref, x_off, y_off,
1350
                    0, 0, 16, 16, width, height, &mb->mv);
1351
        break;
1352
    case VP8_SPLITMVMODE_4x4: {
1353
        int x, y;
1354
        VP56mv uvmv;
1355

    
1356
        /* Y */
1357
        for (y = 0; y < 4; y++) {
1358
            for (x = 0; x < 4; x++) {
1359
                vp8_mc_luma(s, dst[0] + 4*y*s->linesize + x*4,
1360
                            ref->data[0], &bmv[4*y + x],
1361
                            4*x + x_off, 4*y + y_off, 4, 4,
1362
                            width, height, s->linesize,
1363
                            s->put_pixels_tab[2]);
1364
            }
1365
        }
1366

    
1367
        /* U/V */
1368
        x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1;
1369
        for (y = 0; y < 2; y++) {
1370
            for (x = 0; x < 2; x++) {
1371
                uvmv.x = mb->bmv[ 2*y    * 4 + 2*x  ].x +
1372
                         mb->bmv[ 2*y    * 4 + 2*x+1].x +
1373
                         mb->bmv[(2*y+1) * 4 + 2*x  ].x +
1374
                         mb->bmv[(2*y+1) * 4 + 2*x+1].x;
1375
                uvmv.y = mb->bmv[ 2*y    * 4 + 2*x  ].y +
1376
                         mb->bmv[ 2*y    * 4 + 2*x+1].y +
1377
                         mb->bmv[(2*y+1) * 4 + 2*x  ].y +
1378
                         mb->bmv[(2*y+1) * 4 + 2*x+1].y;
1379
                uvmv.x = (uvmv.x + 2 + (uvmv.x >> (INT_BIT-1))) >> 2;
1380
                uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2;
1381
                if (s->profile == 3) {
1382
                    uvmv.x &= ~7;
1383
                    uvmv.y &= ~7;
1384
                }
1385
                vp8_mc_chroma(s, dst[1] + 4*y*s->uvlinesize + x*4,
1386
                              dst[2] + 4*y*s->uvlinesize + x*4,
1387
                              ref->data[1], ref->data[2], &uvmv,
1388
                              4*x + x_off, 4*y + y_off, 4, 4,
1389
                              width, height, s->uvlinesize,
1390
                              s->put_pixels_tab[2]);
1391
            }
1392
        }
1393
        break;
1394
    }
1395
    case VP8_SPLITMVMODE_16x8:
1396
        vp8_mc_part(s, dst, ref, x_off, y_off,
1397
                    0, 0, 16, 8, width, height, &bmv[0]);
1398
        vp8_mc_part(s, dst, ref, x_off, y_off,
1399
                    0, 8, 16, 8, width, height, &bmv[1]);
1400
        break;
1401
    case VP8_SPLITMVMODE_8x16:
1402
        vp8_mc_part(s, dst, ref, x_off, y_off,
1403
                    0, 0, 8, 16, width, height, &bmv[0]);
1404
        vp8_mc_part(s, dst, ref, x_off, y_off,
1405
                    8, 0, 8, 16, width, height, &bmv[1]);
1406
        break;
1407
    case VP8_SPLITMVMODE_8x8:
1408
        vp8_mc_part(s, dst, ref, x_off, y_off,
1409
                    0, 0, 8, 8, width, height, &bmv[0]);
1410
        vp8_mc_part(s, dst, ref, x_off, y_off,
1411
                    8, 0, 8, 8, width, height, &bmv[1]);
1412
        vp8_mc_part(s, dst, ref, x_off, y_off,
1413
                    0, 8, 8, 8, width, height, &bmv[2]);
1414
        vp8_mc_part(s, dst, ref, x_off, y_off,
1415
                    8, 8, 8, 8, width, height, &bmv[3]);
1416
        break;
1417
    }
1418
}
1419

    
1420
static av_always_inline void idct_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb)
1421
{
1422
    int x, y, ch;
1423

    
1424
    if (mb->mode != MODE_I4x4) {
1425
        uint8_t *y_dst = dst[0];
1426
        for (y = 0; y < 4; y++) {
1427
            uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[y]);
1428
            if (nnz4) {
1429
                if (nnz4&~0x01010101) {
1430
                    for (x = 0; x < 4; x++) {
1431
                        int nnz = s->non_zero_count_cache[y][x];
1432
                        if (nnz) {
1433
                            if (nnz == 1)
1434
                                s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize);
1435
                            else
1436
                                s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize);
1437
                        }
1438
                    }
1439
                } else {
1440
                    s->vp8dsp.vp8_idct_dc_add4y(y_dst, s->block[y], s->linesize);
1441
                }
1442
            }
1443
            y_dst += 4*s->linesize;
1444
        }
1445
    }
1446

    
1447
    for (ch = 0; ch < 2; ch++) {
1448
        uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[4+ch]);
1449
        if (nnz4) {
1450
            uint8_t *ch_dst = dst[1+ch];
1451
            if (nnz4&~0x01010101) {
1452
                for (y = 0; y < 2; y++) {
1453
                    for (x = 0; x < 2; x++) {
1454
                        int nnz = s->non_zero_count_cache[4+ch][(y<<1)+x];
1455
                        if (nnz) {
1456
                            if (nnz == 1)
1457
                                s->vp8dsp.vp8_idct_dc_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);
1458
                            else
1459
                                s->vp8dsp.vp8_idct_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);
1460
                        }
1461
                    }
1462
                    ch_dst += 4*s->uvlinesize;
1463
                }
1464
            } else {
1465
                s->vp8dsp.vp8_idct_dc_add4uv(ch_dst, s->block[4+ch], s->uvlinesize);
1466
            }
1467
        }
1468
    }
1469
}
1470

    
1471
static av_always_inline void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f )
1472
{
1473
    int interior_limit, filter_level;
1474

    
1475
    if (s->segmentation.enabled) {
1476
        filter_level = s->segmentation.filter_level[s->segment];
1477
        if (!s->segmentation.absolute_vals)
1478
            filter_level += s->filter.level;
1479
    } else
1480
        filter_level = s->filter.level;
1481

    
1482
    if (s->lf_delta.enabled) {
1483
        filter_level += s->lf_delta.ref[mb->ref_frame];
1484

    
1485
        if (mb->ref_frame == VP56_FRAME_CURRENT) {
1486
            if (mb->mode == MODE_I4x4)
1487
                filter_level += s->lf_delta.mode[0];
1488
        } else {
1489
            if (mb->mode == VP8_MVMODE_ZERO)
1490
                filter_level += s->lf_delta.mode[1];
1491
            else if (mb->mode == VP8_MVMODE_SPLIT)
1492
                filter_level += s->lf_delta.mode[3];
1493
            else
1494
                filter_level += s->lf_delta.mode[2];
1495
        }
1496
    }
1497
    filter_level = av_clip(filter_level, 0, 63);
1498

    
1499
    interior_limit = filter_level;
1500
    if (s->filter.sharpness) {
1501
        interior_limit >>= s->filter.sharpness > 4 ? 2 : 1;
1502
        interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness);
1503
    }
1504
    interior_limit = FFMAX(interior_limit, 1);
1505

    
1506
    f->filter_level = filter_level;
1507
    f->inner_limit = interior_limit;
1508
    f->inner_filter = !mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT;
1509
}
1510

    
1511
static av_always_inline void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f, int mb_x, int mb_y)
1512
{
1513
    int mbedge_lim, bedge_lim, hev_thresh;
1514
    int filter_level = f->filter_level;
1515
    int inner_limit = f->inner_limit;
1516
    int inner_filter = f->inner_filter;
1517
    int linesize = s->linesize;
1518
    int uvlinesize = s->uvlinesize;
1519

    
1520
    if (!filter_level)
1521
        return;
1522

    
1523
    mbedge_lim = 2*(filter_level+2) + inner_limit;
1524
     bedge_lim = 2* filter_level    + inner_limit;
1525
    hev_thresh = filter_level >= 15;
1526

    
1527
    if (s->keyframe) {
1528
        if (filter_level >= 40)
1529
            hev_thresh = 2;
1530
    } else {
1531
        if (filter_level >= 40)
1532
            hev_thresh = 3;
1533
        else if (filter_level >= 20)
1534
            hev_thresh = 2;
1535
    }
1536

    
1537
    if (mb_x) {
1538
        s->vp8dsp.vp8_h_loop_filter16y(dst[0],     linesize,
1539
                                       mbedge_lim, inner_limit, hev_thresh);
1540
        s->vp8dsp.vp8_h_loop_filter8uv(dst[1],     dst[2],      uvlinesize,
1541
                                       mbedge_lim, inner_limit, hev_thresh);
1542
    }
1543

    
1544
    if (inner_filter) {
1545
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 4, linesize, bedge_lim,
1546
                                             inner_limit, hev_thresh);
1547
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 8, linesize, bedge_lim,
1548
                                             inner_limit, hev_thresh);
1549
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+12, linesize, bedge_lim,
1550
                                             inner_limit, hev_thresh);
1551
        s->vp8dsp.vp8_h_loop_filter8uv_inner(dst[1] + 4, dst[2] + 4,
1552
                                             uvlinesize,  bedge_lim,
1553
                                             inner_limit, hev_thresh);
1554
    }
1555

    
1556
    if (mb_y) {
1557
        s->vp8dsp.vp8_v_loop_filter16y(dst[0],     linesize,
1558
                                       mbedge_lim, inner_limit, hev_thresh);
1559
        s->vp8dsp.vp8_v_loop_filter8uv(dst[1],     dst[2],      uvlinesize,
1560
                                       mbedge_lim, inner_limit, hev_thresh);
1561
    }
1562

    
1563
    if (inner_filter) {
1564
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 4*linesize,
1565
                                             linesize,    bedge_lim,
1566
                                             inner_limit, hev_thresh);
1567
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 8*linesize,
1568
                                             linesize,    bedge_lim,
1569
                                             inner_limit, hev_thresh);
1570
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+12*linesize,
1571
                                             linesize,    bedge_lim,
1572
                                             inner_limit, hev_thresh);
1573
        s->vp8dsp.vp8_v_loop_filter8uv_inner(dst[1] + 4 * uvlinesize,
1574
                                             dst[2] + 4 * uvlinesize,
1575
                                             uvlinesize,  bedge_lim,
1576
                                             inner_limit, hev_thresh);
1577
    }
1578
}
1579

    
1580
static av_always_inline void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f, int mb_x, int mb_y)
1581
{
1582
    int mbedge_lim, bedge_lim;
1583
    int filter_level = f->filter_level;
1584
    int inner_limit = f->inner_limit;
1585
    int inner_filter = f->inner_filter;
1586
    int linesize = s->linesize;
1587

    
1588
    if (!filter_level)
1589
        return;
1590

    
1591
    mbedge_lim = 2*(filter_level+2) + inner_limit;
1592
     bedge_lim = 2* filter_level    + inner_limit;
1593

    
1594
    if (mb_x)
1595
        s->vp8dsp.vp8_h_loop_filter_simple(dst, linesize, mbedge_lim);
1596
    if (inner_filter) {
1597
        s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, linesize, bedge_lim);
1598
        s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, linesize, bedge_lim);
1599
        s->vp8dsp.vp8_h_loop_filter_simple(dst+12, linesize, bedge_lim);
1600
    }
1601

    
1602
    if (mb_y)
1603
        s->vp8dsp.vp8_v_loop_filter_simple(dst, linesize, mbedge_lim);
1604
    if (inner_filter) {
1605
        s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*linesize, linesize, bedge_lim);
1606
        s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*linesize, linesize, bedge_lim);
1607
        s->vp8dsp.vp8_v_loop_filter_simple(dst+12*linesize, linesize, bedge_lim);
1608
    }
1609
}
1610

    
1611
static void filter_mb_row(VP8Context *s, int mb_y)
1612
{
1613
    VP8FilterStrength *f = s->filter_strength;
1614
    uint8_t *dst[3] = {
1615
        s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize,
1616
        s->framep[VP56_FRAME_CURRENT]->data[1] +  8*mb_y*s->uvlinesize,
1617
        s->framep[VP56_FRAME_CURRENT]->data[2] +  8*mb_y*s->uvlinesize
1618
    };
1619
    int mb_x;
1620

    
1621
    for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1622
        backup_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], s->linesize, s->uvlinesize, 0);
1623
        filter_mb(s, dst, f++, mb_x, mb_y);
1624
        dst[0] += 16;
1625
        dst[1] += 8;
1626
        dst[2] += 8;
1627
    }
1628
}
1629

    
1630
static void filter_mb_row_simple(VP8Context *s, int mb_y)
1631
{
1632
    VP8FilterStrength *f = s->filter_strength;
1633
    uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize;
1634
    int mb_x;
1635

    
1636
    for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1637
        backup_mb_border(s->top_border[mb_x+1], dst, NULL, NULL, s->linesize, 0, 1);
1638
        filter_mb_simple(s, dst, f++, mb_x, mb_y);
1639
        dst += 16;
1640
    }
1641
}
1642

    
1643
static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
1644
                            AVPacket *avpkt)
1645
{
1646
    VP8Context *s = avctx->priv_data;
1647
    int ret, mb_x, mb_y, i, y, referenced;
1648
    enum AVDiscard skip_thresh;
1649
    AVFrame *av_uninit(curframe);
1650

    
1651
    if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0)
1652
        return ret;
1653

    
1654
    referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT
1655
                                || s->update_altref == VP56_FRAME_CURRENT;
1656

    
1657
    skip_thresh = !referenced ? AVDISCARD_NONREF :
1658
                    !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL;
1659

    
1660
    if (avctx->skip_frame >= skip_thresh) {
1661
        s->invisible = 1;
1662
        goto skip_decode;
1663
    }
1664
    s->deblock_filter = s->filter.level && avctx->skip_loop_filter < skip_thresh;
1665

    
1666
    for (i = 0; i < 4; i++)
1667
        if (&s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1668
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1669
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) {
1670
            curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i];
1671
            break;
1672
        }
1673
    if (curframe->data[0])
1674
        avctx->release_buffer(avctx, curframe);
1675

    
1676
    curframe->key_frame = s->keyframe;
1677
    curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE;
1678
    curframe->reference = referenced ? 3 : 0;
1679
    if ((ret = avctx->get_buffer(avctx, curframe))) {
1680
        av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n");
1681
        return ret;
1682
    }
1683

    
1684
    // Given that arithmetic probabilities are updated every frame, it's quite likely
1685
    // that the values we have on a random interframe are complete junk if we didn't
1686
    // start decode on a keyframe. So just don't display anything rather than junk.
1687
    if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] ||
1688
                         !s->framep[VP56_FRAME_GOLDEN] ||
1689
                         !s->framep[VP56_FRAME_GOLDEN2])) {
1690
        av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n");
1691
        return AVERROR_INVALIDDATA;
1692
    }
1693

    
1694
    s->linesize   = curframe->linesize[0];
1695
    s->uvlinesize = curframe->linesize[1];
1696

    
1697
    if (!s->edge_emu_buffer)
1698
        s->edge_emu_buffer = av_malloc(21*s->linesize);
1699

    
1700
    memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz));
1701

    
1702
    /* Zero macroblock structures for top/top-left prediction from outside the frame. */
1703
    memset(s->macroblocks + s->mb_height*2 - 1, 0, (s->mb_width+1)*sizeof(*s->macroblocks));
1704

    
1705
    // top edge of 127 for intra prediction
1706
    if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) {
1707
        s->top_border[0][15] = s->top_border[0][23] = 127;
1708
        memset(s->top_border[1]-1, 127, s->mb_width*sizeof(*s->top_border)+1);
1709
    }
1710
    memset(s->ref_count, 0, sizeof(s->ref_count));
1711
    if (s->keyframe)
1712
        memset(s->intra4x4_pred_mode_top, DC_PRED, s->mb_width*4);
1713

    
1714
    for (mb_y = 0; mb_y < s->mb_height; mb_y++) {
1715
        VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)];
1716
        VP8Macroblock *mb = s->macroblocks + (s->mb_height - mb_y - 1)*2;
1717
        int mb_xy = mb_y*s->mb_width;
1718
        uint8_t *dst[3] = {
1719
            curframe->data[0] + 16*mb_y*s->linesize,
1720
            curframe->data[1] +  8*mb_y*s->uvlinesize,
1721
            curframe->data[2] +  8*mb_y*s->uvlinesize
1722
        };
1723

    
1724
        memset(mb - 1, 0, sizeof(*mb));   // zero left macroblock
1725
        memset(s->left_nnz, 0, sizeof(s->left_nnz));
1726
        AV_WN32A(s->intra4x4_pred_mode_left, DC_PRED*0x01010101);
1727

    
1728
        // left edge of 129 for intra prediction
1729
        if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) {
1730
            for (i = 0; i < 3; i++)
1731
                for (y = 0; y < 16>>!!i; y++)
1732
                    dst[i][y*curframe->linesize[i]-1] = 129;
1733
            if (mb_y == 1) // top left edge is also 129
1734
                s->top_border[0][15] = s->top_border[0][23] = s->top_border[0][31] = 129;
1735
        }
1736

    
1737
        for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb_xy++, mb++) {
1738
            /* Prefetch the current frame, 4 MBs ahead */
1739
            s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4);
1740
            s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2);
1741

    
1742
            decode_mb_mode(s, mb, mb_x, mb_y, s->segmentation_map + mb_xy);
1743

    
1744
            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_PREVIOUS);
1745

    
1746
            if (!mb->skip)
1747
                decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);
1748

    
1749
            if (mb->mode <= MODE_I4x4)
1750
                intra_predict(s, dst, mb, mb_x, mb_y);
1751
            else
1752
                inter_predict(s, dst, mb, mb_x, mb_y);
1753

    
1754
            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN);
1755

    
1756
            if (!mb->skip) {
1757
                idct_mb(s, dst, mb);
1758
            } else {
1759
                AV_ZERO64(s->left_nnz);
1760
                AV_WN64(s->top_nnz[mb_x], 0);   // array of 9, so unaligned
1761

    
1762
                // Reset DC block predictors if they would exist if the mb had coefficients
1763
                if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
1764
                    s->left_nnz[8]      = 0;
1765
                    s->top_nnz[mb_x][8] = 0;
1766
                }
1767
            }
1768

    
1769
            if (s->deblock_filter)
1770
                filter_level_for_mb(s, mb, &s->filter_strength[mb_x]);
1771

    
1772
            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN2);
1773

    
1774
            dst[0] += 16;
1775
            dst[1] += 8;
1776
            dst[2] += 8;
1777
        }
1778
        if (s->deblock_filter) {
1779
            if (s->filter.simple)
1780
                filter_mb_row_simple(s, mb_y);
1781
            else
1782
                filter_mb_row(s, mb_y);
1783
        }
1784
    }
1785

    
1786
skip_decode:
1787
    // if future frames don't use the updated probabilities,
1788
    // reset them to the values we saved
1789
    if (!s->update_probabilities)
1790
        s->prob[0] = s->prob[1];
1791

    
1792
    // check if golden and altref are swapped
1793
    if (s->update_altref == VP56_FRAME_GOLDEN &&
1794
        s->update_golden == VP56_FRAME_GOLDEN2)
1795
        FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN], s->framep[VP56_FRAME_GOLDEN2]);
1796
    else {
1797
        if (s->update_altref != VP56_FRAME_NONE)
1798
            s->framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref];
1799

    
1800
        if (s->update_golden != VP56_FRAME_NONE)
1801
            s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden];
1802
    }
1803

    
1804
    if (s->update_last) // move cur->prev
1805
        s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT];
1806

    
1807
    // release no longer referenced frames
1808
    for (i = 0; i < 4; i++)
1809
        if (s->frames[i].data[0] &&
1810
            &s->frames[i] != s->framep[VP56_FRAME_CURRENT] &&
1811
            &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1812
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1813
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2])
1814
            avctx->release_buffer(avctx, &s->frames[i]);
1815

    
1816
    if (!s->invisible) {
1817
        *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT];
1818
        *data_size = sizeof(AVFrame);
1819
    }
1820

    
1821
    return avpkt->size;
1822
}
1823

    
1824
static av_cold int vp8_decode_init(AVCodecContext *avctx)
1825
{
1826
    VP8Context *s = avctx->priv_data;
1827

    
1828
    s->avctx = avctx;
1829
    avctx->pix_fmt = PIX_FMT_YUV420P;
1830

    
1831
    dsputil_init(&s->dsp, avctx);
1832
    ff_h264_pred_init(&s->hpc, CODEC_ID_VP8);
1833
    ff_vp8dsp_init(&s->vp8dsp);
1834

    
1835
    return 0;
1836
}
1837

    
1838
static av_cold int vp8_decode_free(AVCodecContext *avctx)
1839
{
1840
    vp8_decode_flush(avctx);
1841
    return 0;
1842
}
1843

    
1844
AVCodec ff_vp8_decoder = {
1845
    "vp8",
1846
    AVMEDIA_TYPE_VIDEO,
1847
    CODEC_ID_VP8,
1848
    sizeof(VP8Context),
1849
    vp8_decode_init,
1850
    NULL,
1851
    vp8_decode_free,
1852
    vp8_decode_frame,
1853
    CODEC_CAP_DR1,
1854
    .flush = vp8_decode_flush,
1855
    .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"),
1856
};