Statistics
| Branch: | Revision:

ffmpeg / libavcodec / vp8.c @ 8cde1b79

History | View | Annotate | Download (65.2 KB)

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

    
25
#include "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
         * [0-3] - i16x16 (always zero)
166
         * [4]   - i4x4
167
         * [5]   - zero mv
168
         * [6]   - inter modes except for zero or split mv
169
         * [7]   - split mv
170
         *  i16x16 modes never have any adjustment
171
         */
172
        int8_t mode[VP8_MVMODE_SPLIT+1];
173

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

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

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

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

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

    
222
    s->macroblocks        = NULL;
223
}
224

    
225
static int update_dimensions(VP8Context *s, int width, int height)
226
{
227
    if (av_image_check_size(width, height, 0, s->avctx))
228
        return AVERROR_INVALIDDATA;
229

    
230
    vp8_decode_flush(s->avctx);
231

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

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

    
237
    s->macroblocks_base        = av_mallocz((s->mb_width+s->mb_height*2+1)*sizeof(*s->macroblocks));
238
    s->filter_strength         = av_mallocz(s->mb_width*sizeof(*s->filter_strength));
239
    s->intra4x4_pred_mode_top  = av_mallocz(s->mb_width*4);
240
    s->top_nnz                 = av_mallocz(s->mb_width*sizeof(*s->top_nnz));
241
    s->top_border              = av_mallocz((s->mb_width+1)*sizeof(*s->top_border));
242
    s->segmentation_map        = av_mallocz(s->mb_width*s->mb_height);
243

    
244
    if (!s->macroblocks_base || !s->filter_strength || !s->intra4x4_pred_mode_top ||
245
        !s->top_nnz || !s->top_border || !s->segmentation_map)
246
        return AVERROR(ENOMEM);
247

    
248
    s->macroblocks        = s->macroblocks_base + 1;
249

    
250
    return 0;
251
}
252

    
253
static void parse_segment_info(VP8Context *s)
254
{
255
    VP56RangeCoder *c = &s->c;
256
    int i;
257

    
258
    s->segmentation.update_map = vp8_rac_get(c);
259

    
260
    if (vp8_rac_get(c)) { // update segment feature data
261
        s->segmentation.absolute_vals = vp8_rac_get(c);
262

    
263
        for (i = 0; i < 4; i++)
264
            s->segmentation.base_quant[i]   = vp8_rac_get_sint(c, 7);
265

    
266
        for (i = 0; i < 4; i++)
267
            s->segmentation.filter_level[i] = vp8_rac_get_sint(c, 6);
268
    }
269
    if (s->segmentation.update_map)
270
        for (i = 0; i < 3; i++)
271
            s->prob->segmentid[i] = vp8_rac_get(c) ? vp8_rac_get_uint(c, 8) : 255;
272
}
273

    
274
static void update_lf_deltas(VP8Context *s)
275
{
276
    VP56RangeCoder *c = &s->c;
277
    int i;
278

    
279
    for (i = 0; i < 4; i++)
280
        s->lf_delta.ref[i]  = vp8_rac_get_sint(c, 6);
281

    
282
    for (i = MODE_I4x4; i <= VP8_MVMODE_SPLIT; i++)
283
        s->lf_delta.mode[i] = vp8_rac_get_sint(c, 6);
284
}
285

    
286
static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size)
287
{
288
    const uint8_t *sizes = buf;
289
    int i;
290

    
291
    s->num_coeff_partitions = 1 << vp8_rac_get_uint(&s->c, 2);
292

    
293
    buf      += 3*(s->num_coeff_partitions-1);
294
    buf_size -= 3*(s->num_coeff_partitions-1);
295
    if (buf_size < 0)
296
        return -1;
297

    
298
    for (i = 0; i < s->num_coeff_partitions-1; i++) {
299
        int size = AV_RL24(sizes + 3*i);
300
        if (buf_size - size < 0)
301
            return -1;
302

    
303
        ff_vp56_init_range_decoder(&s->coeff_partition[i], buf, size);
304
        buf      += size;
305
        buf_size -= size;
306
    }
307
    ff_vp56_init_range_decoder(&s->coeff_partition[i], buf, buf_size);
308

    
309
    return 0;
310
}
311

    
312
static void get_quants(VP8Context *s)
313
{
314
    VP56RangeCoder *c = &s->c;
315
    int i, base_qi;
316

    
317
    int yac_qi     = vp8_rac_get_uint(c, 7);
318
    int ydc_delta  = vp8_rac_get_sint(c, 4);
319
    int y2dc_delta = vp8_rac_get_sint(c, 4);
320
    int y2ac_delta = vp8_rac_get_sint(c, 4);
321
    int uvdc_delta = vp8_rac_get_sint(c, 4);
322
    int uvac_delta = vp8_rac_get_sint(c, 4);
323

    
324
    for (i = 0; i < 4; i++) {
325
        if (s->segmentation.enabled) {
326
            base_qi = s->segmentation.base_quant[i];
327
            if (!s->segmentation.absolute_vals)
328
                base_qi += yac_qi;
329
        } else
330
            base_qi = yac_qi;
331

    
332
        s->qmat[i].luma_qmul[0]    =       vp8_dc_qlookup[av_clip(base_qi + ydc_delta , 0, 127)];
333
        s->qmat[i].luma_qmul[1]    =       vp8_ac_qlookup[av_clip(base_qi             , 0, 127)];
334
        s->qmat[i].luma_dc_qmul[0] =   2 * vp8_dc_qlookup[av_clip(base_qi + y2dc_delta, 0, 127)];
335
        s->qmat[i].luma_dc_qmul[1] = 155 * vp8_ac_qlookup[av_clip(base_qi + y2ac_delta, 0, 127)] / 100;
336
        s->qmat[i].chroma_qmul[0]  =       vp8_dc_qlookup[av_clip(base_qi + uvdc_delta, 0, 127)];
337
        s->qmat[i].chroma_qmul[1]  =       vp8_ac_qlookup[av_clip(base_qi + uvac_delta, 0, 127)];
338

    
339
        s->qmat[i].luma_dc_qmul[1] = FFMAX(s->qmat[i].luma_dc_qmul[1], 8);
340
        s->qmat[i].chroma_qmul[0]  = FFMIN(s->qmat[i].chroma_qmul[0], 132);
341
    }
342
}
343

    
344
/**
345
 * Determine which buffers golden and altref should be updated with after this frame.
346
 * The spec isn't clear here, so I'm going by my understanding of what libvpx does
347
 *
348
 * Intra frames update all 3 references
349
 * Inter frames update VP56_FRAME_PREVIOUS if the update_last flag is set
350
 * If the update (golden|altref) flag is set, it's updated with the current frame
351
 *      if update_last is set, and VP56_FRAME_PREVIOUS otherwise.
352
 * If the flag is not set, the number read means:
353
 *      0: no update
354
 *      1: VP56_FRAME_PREVIOUS
355
 *      2: update golden with altref, or update altref with golden
356
 */
357
static VP56Frame ref_to_update(VP8Context *s, int update, VP56Frame ref)
358
{
359
    VP56RangeCoder *c = &s->c;
360

    
361
    if (update)
362
        return VP56_FRAME_CURRENT;
363

    
364
    switch (vp8_rac_get_uint(c, 2)) {
365
    case 1:
366
        return VP56_FRAME_PREVIOUS;
367
    case 2:
368
        return (ref == VP56_FRAME_GOLDEN) ? VP56_FRAME_GOLDEN2 : VP56_FRAME_GOLDEN;
369
    }
370
    return VP56_FRAME_NONE;
371
}
372

    
373
static void update_refs(VP8Context *s)
374
{
375
    VP56RangeCoder *c = &s->c;
376

    
377
    int update_golden = vp8_rac_get(c);
378
    int update_altref = vp8_rac_get(c);
379

    
380
    s->update_golden = ref_to_update(s, update_golden, VP56_FRAME_GOLDEN);
381
    s->update_altref = ref_to_update(s, update_altref, VP56_FRAME_GOLDEN2);
382
}
383

    
384
static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)
385
{
386
    VP56RangeCoder *c = &s->c;
387
    int header_size, hscale, vscale, i, j, k, l, m, ret;
388
    int width  = s->avctx->width;
389
    int height = s->avctx->height;
390

    
391
    s->keyframe  = !(buf[0] & 1);
392
    s->profile   =  (buf[0]>>1) & 7;
393
    s->invisible = !(buf[0] & 0x10);
394
    header_size  = AV_RL24(buf) >> 5;
395
    buf      += 3;
396
    buf_size -= 3;
397

    
398
    if (s->profile > 3)
399
        av_log(s->avctx, AV_LOG_WARNING, "Unknown profile %d\n", s->profile);
400

    
401
    if (!s->profile)
402
        memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_epel_pixels_tab, sizeof(s->put_pixels_tab));
403
    else    // profile 1-3 use bilinear, 4+ aren't defined so whatever
404
        memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_bilinear_pixels_tab, sizeof(s->put_pixels_tab));
405

    
406
    if (header_size > buf_size - 7*s->keyframe) {
407
        av_log(s->avctx, AV_LOG_ERROR, "Header size larger than data provided\n");
408
        return AVERROR_INVALIDDATA;
409
    }
410

    
411
    if (s->keyframe) {
412
        if (AV_RL24(buf) != 0x2a019d) {
413
            av_log(s->avctx, AV_LOG_ERROR, "Invalid start code 0x%x\n", AV_RL24(buf));
414
            return AVERROR_INVALIDDATA;
415
        }
416
        width  = AV_RL16(buf+3) & 0x3fff;
417
        height = AV_RL16(buf+5) & 0x3fff;
418
        hscale = buf[4] >> 6;
419
        vscale = buf[6] >> 6;
420
        buf      += 7;
421
        buf_size -= 7;
422

    
423
        if (hscale || vscale)
424
            av_log_missing_feature(s->avctx, "Upscaling", 1);
425

    
426
        s->update_golden = s->update_altref = VP56_FRAME_CURRENT;
427
        for (i = 0; i < 4; i++)
428
            for (j = 0; j < 16; j++)
429
                memcpy(s->prob->token[i][j], vp8_token_default_probs[i][vp8_coeff_band[j]],
430
                       sizeof(s->prob->token[i][j]));
431
        memcpy(s->prob->pred16x16, vp8_pred16x16_prob_inter, sizeof(s->prob->pred16x16));
432
        memcpy(s->prob->pred8x8c , vp8_pred8x8c_prob_inter , sizeof(s->prob->pred8x8c));
433
        memcpy(s->prob->mvc      , vp8_mv_default_prob     , sizeof(s->prob->mvc));
434
        memset(&s->segmentation, 0, sizeof(s->segmentation));
435
    }
436

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

    
443
    ff_vp56_init_range_decoder(c, buf, header_size);
444
    buf      += header_size;
445
    buf_size -= header_size;
446

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

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

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

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

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

    
471
    get_quants(s);
472

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

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

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

    
486
    for (i = 0; i < 4; i++)
487
        for (j = 0; j < 8; j++)
488
            for (k = 0; k < 3; k++)
489
                for (l = 0; l < NUM_DCT_TOKENS-1; l++)
490
                    if (vp56_rac_get_prob_branchy(c, vp8_token_update_probs[i][j][k][l])) {
491
                        int prob = vp8_rac_get_uint(c, 8);
492
                        for (m = 0; vp8_coeff_band_indexes[j][m] >= 0; m++)
493
                            s->prob->token[i][vp8_coeff_band_indexes[j][m]][k][l] = prob;
494
                    }
495

    
496
    if ((s->mbskip_enabled = vp8_rac_get(c)))
497
        s->prob->mbskip = vp8_rac_get_uint(c, 8);
498

    
499
    if (!s->keyframe) {
500
        s->prob->intra  = vp8_rac_get_uint(c, 8);
501
        s->prob->last   = vp8_rac_get_uint(c, 8);
502
        s->prob->golden = vp8_rac_get_uint(c, 8);
503

    
504
        if (vp8_rac_get(c))
505
            for (i = 0; i < 4; i++)
506
                s->prob->pred16x16[i] = vp8_rac_get_uint(c, 8);
507
        if (vp8_rac_get(c))
508
            for (i = 0; i < 3; i++)
509
                s->prob->pred8x8c[i]  = vp8_rac_get_uint(c, 8);
510

    
511
        // 17.2 MV probability update
512
        for (i = 0; i < 2; i++)
513
            for (j = 0; j < 19; j++)
514
                if (vp56_rac_get_prob_branchy(c, vp8_mv_update_prob[i][j]))
515
                    s->prob->mvc[i][j] = vp8_rac_get_nn(c);
516
    }
517

    
518
    return 0;
519
}
520

    
521
static av_always_inline
522
void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src, int mb_x, int mb_y)
523
{
524
#define MARGIN (16 << 2)
525
    dst->x = av_clip(src->x, -((mb_x << 6) + MARGIN),
526
                     ((s->mb_width  - 1 - mb_x) << 6) + MARGIN);
527
    dst->y = av_clip(src->y, -((mb_y << 6) + MARGIN),
528
                     ((s->mb_height - 1 - mb_y) << 6) + MARGIN);
529
}
530

    
531
static av_always_inline
532
void find_near_mvs(VP8Context *s, VP8Macroblock *mb,
533
                   VP56mv near[2], VP56mv *best, uint8_t cnt[4])
534
{
535
    VP8Macroblock *mb_edge[3] = { mb + 2 /* top */,
536
                                  mb - 1 /* left */,
537
                                  mb + 1 /* top-left */ };
538
    enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT };
539
    VP56mv near_mv[4]  = {{ 0 }};
540
    enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV };
541
    int idx = CNT_ZERO;
542
    int best_idx = CNT_ZERO;
543
    int cur_sign_bias = s->sign_bias[mb->ref_frame];
544
    int *sign_bias = s->sign_bias;
545

    
546
    /* Process MB on top, left and top-left */
547
    #define MV_EDGE_CHECK(n)\
548
    {\
549
        VP8Macroblock *edge = mb_edge[n];\
550
        int edge_ref = edge->ref_frame;\
551
        if (edge_ref != VP56_FRAME_CURRENT) {\
552
            uint32_t mv = AV_RN32A(&edge->mv);\
553
            if (mv) {\
554
                if (cur_sign_bias != sign_bias[edge_ref]) {\
555
                    /* SWAR negate of the values in mv. */\
556
                    mv = ~mv;\
557
                    mv = ((mv&0x7fff7fff) + 0x00010001) ^ (mv&0x80008000);\
558
                }\
559
                if (!n || mv != AV_RN32A(&near_mv[idx]))\
560
                    AV_WN32A(&near_mv[++idx], mv);\
561
                cnt[idx]      += 1 + (n != 2);\
562
            } else\
563
                cnt[CNT_ZERO] += 1 + (n != 2);\
564
        }\
565
    }
566
    MV_EDGE_CHECK(0)
567
    MV_EDGE_CHECK(1)
568
    MV_EDGE_CHECK(2)
569

    
570
    /* If we have three distinct MVs, merge first and last if they're the same */
571
    if (cnt[CNT_SPLITMV] && AV_RN32A(&near_mv[1+EDGE_TOP]) == AV_RN32A(&near_mv[1+EDGE_TOPLEFT]))
572
        cnt[CNT_NEAREST] += 1;
573

    
574
    cnt[CNT_SPLITMV] = ((mb_edge[EDGE_LEFT]->mode   == VP8_MVMODE_SPLIT) +
575
                        (mb_edge[EDGE_TOP]->mode    == VP8_MVMODE_SPLIT)) * 2 +
576
                       (mb_edge[EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT);
577

    
578
    /* Swap near and nearest if necessary */
579
    if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) {
580
        FFSWAP(uint8_t,     cnt[CNT_NEAREST],     cnt[CNT_NEAR]);
581
        FFSWAP( VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]);
582
    }
583

    
584
    /* Choose the best mv out of 0,0 and the nearest mv */
585
    if (cnt[CNT_NEAREST] >= cnt[CNT_ZERO])
586
        best_idx = CNT_NEAREST;
587

    
588
    mb->mv  = near_mv[best_idx];
589
    near[0] = near_mv[CNT_NEAREST];
590
    near[1] = near_mv[CNT_NEAR];
591
}
592

    
593
/**
594
 * Motion vector coding, 17.1.
595
 */
596
static int read_mv_component(VP56RangeCoder *c, const uint8_t *p)
597
{
598
    int bit, x = 0;
599

    
600
    if (vp56_rac_get_prob_branchy(c, p[0])) {
601
        int i;
602

    
603
        for (i = 0; i < 3; i++)
604
            x += vp56_rac_get_prob(c, p[9 + i]) << i;
605
        for (i = 9; i > 3; i--)
606
            x += vp56_rac_get_prob(c, p[9 + i]) << i;
607
        if (!(x & 0xFFF0) || vp56_rac_get_prob(c, p[12]))
608
            x += 8;
609
    } else {
610
        // small_mvtree
611
        const uint8_t *ps = p+2;
612
        bit = vp56_rac_get_prob(c, *ps);
613
        ps += 1 + 3*bit;
614
        x  += 4*bit;
615
        bit = vp56_rac_get_prob(c, *ps);
616
        ps += 1 + bit;
617
        x  += 2*bit;
618
        x  += vp56_rac_get_prob(c, *ps);
619
    }
620

    
621
    return (x && vp56_rac_get_prob(c, p[1])) ? -x : x;
622
}
623

    
624
static av_always_inline
625
const uint8_t *get_submv_prob(uint32_t left, uint32_t top)
626
{
627
    if (left == top)
628
        return vp8_submv_prob[4-!!left];
629
    if (!top)
630
        return vp8_submv_prob[2];
631
    return vp8_submv_prob[1-!!left];
632
}
633

    
634
/**
635
 * Split motion vector prediction, 16.4.
636
 * @returns the number of motion vectors parsed (2, 4 or 16)
637
 */
638
static av_always_inline
639
int decode_splitmvs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb)
640
{
641
    int part_idx;
642
    int n, num;
643
    VP8Macroblock *top_mb  = &mb[2];
644
    VP8Macroblock *left_mb = &mb[-1];
645
    const uint8_t *mbsplits_left = vp8_mbsplits[left_mb->partitioning],
646
                  *mbsplits_top = vp8_mbsplits[top_mb->partitioning],
647
                  *mbsplits_cur, *firstidx;
648
    VP56mv *top_mv  = top_mb->bmv;
649
    VP56mv *left_mv = left_mb->bmv;
650
    VP56mv *cur_mv  = mb->bmv;
651

    
652
    if (vp56_rac_get_prob_branchy(c, vp8_mbsplit_prob[0])) {
653
        if (vp56_rac_get_prob_branchy(c, vp8_mbsplit_prob[1])) {
654
            part_idx = VP8_SPLITMVMODE_16x8 + vp56_rac_get_prob(c, vp8_mbsplit_prob[2]);
655
        } else {
656
            part_idx = VP8_SPLITMVMODE_8x8;
657
        }
658
    } else {
659
        part_idx = VP8_SPLITMVMODE_4x4;
660
    }
661

    
662
    num = vp8_mbsplit_count[part_idx];
663
    mbsplits_cur = vp8_mbsplits[part_idx],
664
    firstidx = vp8_mbfirstidx[part_idx];
665
    mb->partitioning = part_idx;
666

    
667
    for (n = 0; n < num; n++) {
668
        int k = firstidx[n];
669
        uint32_t left, above;
670
        const uint8_t *submv_prob;
671

    
672
        if (!(k & 3))
673
            left = AV_RN32A(&left_mv[mbsplits_left[k + 3]]);
674
        else
675
            left  = AV_RN32A(&cur_mv[mbsplits_cur[k - 1]]);
676
        if (k <= 3)
677
            above = AV_RN32A(&top_mv[mbsplits_top[k + 12]]);
678
        else
679
            above = AV_RN32A(&cur_mv[mbsplits_cur[k - 4]]);
680

    
681
        submv_prob = get_submv_prob(left, above);
682

    
683
        if (vp56_rac_get_prob_branchy(c, submv_prob[0])) {
684
            if (vp56_rac_get_prob_branchy(c, submv_prob[1])) {
685
                if (vp56_rac_get_prob_branchy(c, submv_prob[2])) {
686
                    mb->bmv[n].y = mb->mv.y + read_mv_component(c, s->prob->mvc[0]);
687
                    mb->bmv[n].x = mb->mv.x + read_mv_component(c, s->prob->mvc[1]);
688
                } else {
689
                    AV_ZERO32(&mb->bmv[n]);
690
                }
691
            } else {
692
                AV_WN32A(&mb->bmv[n], above);
693
            }
694
        } else {
695
            AV_WN32A(&mb->bmv[n], left);
696
        }
697
    }
698

    
699
    return num;
700
}
701

    
702
static av_always_inline
703
void decode_intra4x4_modes(VP8Context *s, VP56RangeCoder *c,
704
                           int mb_x, int keyframe)
705
{
706
    uint8_t *intra4x4 = s->intra4x4_pred_mode_mb;
707
    if (keyframe) {
708
        int x, y;
709
        uint8_t* const top = s->intra4x4_pred_mode_top + 4 * mb_x;
710
        uint8_t* const left = s->intra4x4_pred_mode_left;
711
        for (y = 0; y < 4; y++) {
712
            for (x = 0; x < 4; x++) {
713
                const uint8_t *ctx;
714
                ctx = vp8_pred4x4_prob_intra[top[x]][left[y]];
715
                *intra4x4 = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx);
716
                left[y] = top[x] = *intra4x4;
717
                intra4x4++;
718
            }
719
        }
720
    } else {
721
        int i;
722
        for (i = 0; i < 16; i++)
723
            intra4x4[i] = vp8_rac_get_tree(c, vp8_pred4x4_tree, vp8_pred4x4_prob_inter);
724
    }
725
}
726

    
727
static av_always_inline
728
void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, uint8_t *segment)
729
{
730
    VP56RangeCoder *c = &s->c;
731

    
732
    if (s->segmentation.update_map)
733
        *segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid);
734
    s->segment = *segment;
735

    
736
    mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0;
737

    
738
    if (s->keyframe) {
739
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra);
740

    
741
        if (mb->mode == MODE_I4x4) {
742
            decode_intra4x4_modes(s, c, mb_x, 1);
743
        } else {
744
            const uint32_t modes = vp8_pred4x4_mode[mb->mode] * 0x01010101u;
745
            AV_WN32A(s->intra4x4_pred_mode_top + 4 * mb_x, modes);
746
            AV_WN32A(s->intra4x4_pred_mode_left, modes);
747
        }
748

    
749
        s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra);
750
        mb->ref_frame = VP56_FRAME_CURRENT;
751
    } else if (vp56_rac_get_prob_branchy(c, s->prob->intra)) {
752
        VP56mv near[2], best;
753
        uint8_t cnt[4] = { 0 };
754

    
755
        // inter MB, 16.2
756
        if (vp56_rac_get_prob_branchy(c, s->prob->last))
757
            mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ?
758
                VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN;
759
        else
760
            mb->ref_frame = VP56_FRAME_PREVIOUS;
761
        s->ref_count[mb->ref_frame-1]++;
762

    
763
        // motion vectors, 16.3
764
        find_near_mvs(s, mb, near, &best, cnt);
765
        if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[0]][0])) {
766
            mb->mode = VP8_MVMODE_MV;
767
            if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[1]][1])) {
768
                if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[2]][2])) {
769
                    if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[3]][3])) {
770
                        mb->mode = VP8_MVMODE_SPLIT;
771
                        clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y);
772
                        mb->mv = mb->bmv[decode_splitmvs(s, c, mb) - 1];
773
                    } else {
774
                        clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y);
775
                        mb->mv.y += read_mv_component(c, s->prob->mvc[0]);
776
                        mb->mv.x += read_mv_component(c, s->prob->mvc[1]);
777
                    }
778
                } else
779
                    clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y);
780
            } else
781
                clamp_mv(s, &mb->mv, &near[0], mb_x, mb_y);
782
        } else {
783
            mb->mode = VP8_MVMODE_ZERO;
784
            AV_ZERO32(&mb->mv);
785
        }
786
        if (mb->mode != VP8_MVMODE_SPLIT) {
787
            mb->partitioning = VP8_SPLITMVMODE_NONE;
788
            mb->bmv[0] = mb->mv;
789
        }
790
    } else {
791
        // intra MB, 16.1
792
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16);
793

    
794
        if (mb->mode == MODE_I4x4)
795
            decode_intra4x4_modes(s, c, mb_x, 0);
796

    
797
        s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c);
798
        mb->ref_frame = VP56_FRAME_CURRENT;
799
        mb->partitioning = VP8_SPLITMVMODE_NONE;
800
        AV_ZERO32(&mb->bmv[0]);
801
    }
802
}
803

    
804
/**
805
 * @param c arithmetic bitstream reader context
806
 * @param block destination for block coefficients
807
 * @param probs probabilities to use when reading trees from the bitstream
808
 * @param i initial coeff index, 0 unless a separate DC block is coded
809
 * @param zero_nhood the initial prediction context for number of surrounding
810
 *                   all-zero blocks (only left/top, so 0-2)
811
 * @param qmul array holding the dc/ac dequant factor at position 0/1
812
 * @return 0 if no coeffs were decoded
813
 *         otherwise, the index of the last coeff decoded plus one
814
 */
815
static int decode_block_coeffs_internal(VP56RangeCoder *c, DCTELEM block[16],
816
                                        uint8_t probs[8][3][NUM_DCT_TOKENS-1],
817
                                        int i, uint8_t *token_prob, int16_t qmul[2])
818
{
819
    goto skip_eob;
820
    do {
821
        int coeff;
822
        if (!vp56_rac_get_prob_branchy(c, token_prob[0]))   // DCT_EOB
823
            return i;
824

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

    
833
        if (!vp56_rac_get_prob_branchy(c, token_prob[2])) { // DCT_1
834
            coeff = 1;
835
            token_prob = probs[i+1][1];
836
        } else {
837
            if (!vp56_rac_get_prob_branchy(c, token_prob[3])) { // DCT 2,3,4
838
                coeff = vp56_rac_get_prob_branchy(c, token_prob[4]);
839
                if (coeff)
840
                    coeff += vp56_rac_get_prob(c, token_prob[5]);
841
                coeff += 2;
842
            } else {
843
                // DCT_CAT*
844
                if (!vp56_rac_get_prob_branchy(c, token_prob[6])) {
845
                    if (!vp56_rac_get_prob_branchy(c, token_prob[7])) { // DCT_CAT1
846
                        coeff  = 5 + vp56_rac_get_prob(c, vp8_dct_cat1_prob[0]);
847
                    } else {                                    // DCT_CAT2
848
                        coeff  = 7;
849
                        coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[0]) << 1;
850
                        coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[1]);
851
                    }
852
                } else {    // DCT_CAT3 and up
853
                    int a = vp56_rac_get_prob(c, token_prob[8]);
854
                    int b = vp56_rac_get_prob(c, token_prob[9+a]);
855
                    int cat = (a<<1) + b;
856
                    coeff  = 3 + (8<<cat);
857
                    coeff += vp8_rac_get_coeff(c, vp8_dct_cat_prob[cat]);
858
                }
859
            }
860
            token_prob = probs[i+1][2];
861
        }
862
        block[zigzag_scan[i]] = (vp8_rac_get(c) ? -coeff : coeff) * qmul[!!i];
863
    } while (++i < 16);
864

    
865
    return i;
866
}
867

    
868
static av_always_inline
869
int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16],
870
                        uint8_t probs[8][3][NUM_DCT_TOKENS-1],
871
                        int i, int zero_nhood, int16_t qmul[2])
872
{
873
    uint8_t *token_prob = probs[i][zero_nhood];
874
    if (!vp56_rac_get_prob_branchy(c, token_prob[0]))   // DCT_EOB
875
        return 0;
876
    return decode_block_coeffs_internal(c, block, probs, i, token_prob, qmul);
877
}
878

    
879
static av_always_inline
880
void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,
881
                      uint8_t t_nnz[9], uint8_t l_nnz[9])
882
{
883
    int i, x, y, luma_start = 0, luma_ctx = 3;
884
    int nnz_pred, nnz, nnz_total = 0;
885
    int segment = s->segment;
886
    int block_dc = 0;
887

    
888
    if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
889
        nnz_pred = t_nnz[8] + l_nnz[8];
890

    
891
        // decode DC values and do hadamard
892
        nnz = decode_block_coeffs(c, s->block_dc, s->prob->token[1], 0, nnz_pred,
893
                                  s->qmat[segment].luma_dc_qmul);
894
        l_nnz[8] = t_nnz[8] = !!nnz;
895
        if (nnz) {
896
            nnz_total += nnz;
897
            block_dc = 1;
898
            if (nnz == 1)
899
                s->vp8dsp.vp8_luma_dc_wht_dc(s->block, s->block_dc);
900
            else
901
                s->vp8dsp.vp8_luma_dc_wht(s->block, s->block_dc);
902
        }
903
        luma_start = 1;
904
        luma_ctx = 0;
905
    }
906

    
907
    // luma blocks
908
    for (y = 0; y < 4; y++)
909
        for (x = 0; x < 4; x++) {
910
            nnz_pred = l_nnz[y] + t_nnz[x];
911
            nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start,
912
                                      nnz_pred, s->qmat[segment].luma_qmul);
913
            // nnz+block_dc may be one more than the actual last index, but we don't care
914
            s->non_zero_count_cache[y][x] = nnz + block_dc;
915
            t_nnz[x] = l_nnz[y] = !!nnz;
916
            nnz_total += nnz;
917
        }
918

    
919
    // chroma blocks
920
    // TODO: what to do about dimensions? 2nd dim for luma is x,
921
    // but for chroma it's (y<<1)|x
922
    for (i = 4; i < 6; i++)
923
        for (y = 0; y < 2; y++)
924
            for (x = 0; x < 2; x++) {
925
                nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x];
926
                nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0,
927
                                          nnz_pred, s->qmat[segment].chroma_qmul);
928
                s->non_zero_count_cache[i][(y<<1)+x] = nnz;
929
                t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz;
930
                nnz_total += nnz;
931
            }
932

    
933
    // if there were no coded coeffs despite the macroblock not being marked skip,
934
    // we MUST not do the inner loop filter and should not do IDCT
935
    // Since skip isn't used for bitstream prediction, just manually set it.
936
    if (!nnz_total)
937
        mb->skip = 1;
938
}
939

    
940
static av_always_inline
941
void backup_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
942
                      int linesize, int uvlinesize, int simple)
943
{
944
    AV_COPY128(top_border, src_y + 15*linesize);
945
    if (!simple) {
946
        AV_COPY64(top_border+16, src_cb + 7*uvlinesize);
947
        AV_COPY64(top_border+24, src_cr + 7*uvlinesize);
948
    }
949
}
950

    
951
static av_always_inline
952
void xchg_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
953
                    int linesize, int uvlinesize, int mb_x, int mb_y, int mb_width,
954
                    int simple, int xchg)
955
{
956
    uint8_t *top_border_m1 = top_border-32;     // for TL prediction
957
    src_y  -=   linesize;
958
    src_cb -= uvlinesize;
959
    src_cr -= uvlinesize;
960

    
961
#define XCHG(a,b,xchg) do {                     \
962
        if (xchg) AV_SWAP64(b,a);               \
963
        else      AV_COPY64(b,a);               \
964
    } while (0)
965

    
966
    XCHG(top_border_m1+8, src_y-8, xchg);
967
    XCHG(top_border,      src_y,   xchg);
968
    XCHG(top_border+8,    src_y+8, 1);
969
    if (mb_x < mb_width-1)
970
        XCHG(top_border+32, src_y+16, 1);
971

    
972
    // only copy chroma for normal loop filter
973
    // or to initialize the top row to 127
974
    if (!simple || !mb_y) {
975
        XCHG(top_border_m1+16, src_cb-8, xchg);
976
        XCHG(top_border_m1+24, src_cr-8, xchg);
977
        XCHG(top_border+16,    src_cb, 1);
978
        XCHG(top_border+24,    src_cr, 1);
979
    }
980
}
981

    
982
static av_always_inline
983
int check_dc_pred8x8_mode(int mode, int mb_x, int mb_y)
984
{
985
    if (!mb_x) {
986
        return mb_y ? TOP_DC_PRED8x8 : DC_128_PRED8x8;
987
    } else {
988
        return mb_y ? mode : LEFT_DC_PRED8x8;
989
    }
990
}
991

    
992
static av_always_inline
993
int check_tm_pred8x8_mode(int mode, int mb_x, int mb_y)
994
{
995
    if (!mb_x) {
996
        return mb_y ? VERT_PRED8x8 : DC_129_PRED8x8;
997
    } else {
998
        return mb_y ? mode : HOR_PRED8x8;
999
    }
1000
}
1001

    
1002
static av_always_inline
1003
int check_intra_pred8x8_mode(int mode, int mb_x, int mb_y)
1004
{
1005
    if (mode == DC_PRED8x8) {
1006
        return check_dc_pred8x8_mode(mode, mb_x, mb_y);
1007
    } else {
1008
        return mode;
1009
    }
1010
}
1011

    
1012
static av_always_inline
1013
int check_intra_pred8x8_mode_emuedge(int mode, int mb_x, int mb_y)
1014
{
1015
    switch (mode) {
1016
    case DC_PRED8x8:
1017
        return check_dc_pred8x8_mode(mode, mb_x, mb_y);
1018
    case VERT_PRED8x8:
1019
        return !mb_y ? DC_127_PRED8x8 : mode;
1020
    case HOR_PRED8x8:
1021
        return !mb_x ? DC_129_PRED8x8 : mode;
1022
    case PLANE_PRED8x8 /*TM*/:
1023
        return check_tm_pred8x8_mode(mode, mb_x, mb_y);
1024
    }
1025
    return mode;
1026
}
1027

    
1028
static av_always_inline
1029
int check_tm_pred4x4_mode(int mode, int mb_x, int mb_y)
1030
{
1031
    if (!mb_x) {
1032
        return mb_y ? VERT_VP8_PRED : DC_129_PRED;
1033
    } else {
1034
        return mb_y ? mode : HOR_VP8_PRED;
1035
    }
1036
}
1037

    
1038
static av_always_inline
1039
int check_intra_pred4x4_mode_emuedge(int mode, int mb_x, int mb_y, int *copy_buf)
1040
{
1041
    switch (mode) {
1042
    case VERT_PRED:
1043
        if (!mb_x && mb_y) {
1044
            *copy_buf = 1;
1045
            return mode;
1046
        }
1047
        /* fall-through */
1048
    case DIAG_DOWN_LEFT_PRED:
1049
    case VERT_LEFT_PRED:
1050
        return !mb_y ? DC_127_PRED : mode;
1051
    case HOR_PRED:
1052
        if (!mb_y) {
1053
            *copy_buf = 1;
1054
            return mode;
1055
        }
1056
        /* fall-through */
1057
    case HOR_UP_PRED:
1058
        return !mb_x ? DC_129_PRED : mode;
1059
    case TM_VP8_PRED:
1060
        return check_tm_pred4x4_mode(mode, mb_x, mb_y);
1061
    case DC_PRED: // 4x4 DC doesn't use the same "H.264-style" exceptions as 16x16/8x8 DC
1062
    case DIAG_DOWN_RIGHT_PRED:
1063
    case VERT_RIGHT_PRED:
1064
    case HOR_DOWN_PRED:
1065
        if (!mb_y || !mb_x)
1066
            *copy_buf = 1;
1067
        return mode;
1068
    }
1069
    return mode;
1070
}
1071

    
1072
static av_always_inline
1073
void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
1074
                   int mb_x, int mb_y)
1075
{
1076
    AVCodecContext *avctx = s->avctx;
1077
    int x, y, mode, nnz, tr;
1078

    
1079
    // for the first row, we need to run xchg_mb_border to init the top edge to 127
1080
    // otherwise, skip it if we aren't going to deblock
1081
    if (!(avctx->flags & CODEC_FLAG_EMU_EDGE && !mb_y) && (s->deblock_filter || !mb_y))
1082
        xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
1083
                       s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
1084
                       s->filter.simple, 1);
1085

    
1086
    if (mb->mode < MODE_I4x4) {
1087
        if (avctx->flags & CODEC_FLAG_EMU_EDGE) { // tested
1088
            mode = check_intra_pred8x8_mode_emuedge(mb->mode, mb_x, mb_y);
1089
        } else {
1090
            mode = check_intra_pred8x8_mode(mb->mode, mb_x, mb_y);
1091
        }
1092
        s->hpc.pred16x16[mode](dst[0], s->linesize);
1093
    } else {
1094
        uint8_t *ptr = dst[0];
1095
        uint8_t *intra4x4 = s->intra4x4_pred_mode_mb;
1096
        uint8_t tr_top[4] = { 127, 127, 127, 127 };
1097

    
1098
        // all blocks on the right edge of the macroblock use bottom edge
1099
        // the top macroblock for their topright edge
1100
        uint8_t *tr_right = ptr - s->linesize + 16;
1101

    
1102
        // if we're on the right edge of the frame, said edge is extended
1103
        // from the top macroblock
1104
        if (!(!mb_y && avctx->flags & CODEC_FLAG_EMU_EDGE) &&
1105
            mb_x == s->mb_width-1) {
1106
            tr = tr_right[-1]*0x01010101;
1107
            tr_right = (uint8_t *)&tr;
1108
        }
1109

    
1110
        if (mb->skip)
1111
            AV_ZERO128(s->non_zero_count_cache);
1112

    
1113
        for (y = 0; y < 4; y++) {
1114
            uint8_t *topright = ptr + 4 - s->linesize;
1115
            for (x = 0; x < 4; x++) {
1116
                int copy = 0, linesize = s->linesize;
1117
                uint8_t *dst = ptr+4*x;
1118
                DECLARE_ALIGNED(4, uint8_t, copy_dst)[5*8];
1119

    
1120
                if ((y == 0 || x == 3) && mb_y == 0 && avctx->flags & CODEC_FLAG_EMU_EDGE) {
1121
                    topright = tr_top;
1122
                } else if (x == 3)
1123
                    topright = tr_right;
1124

    
1125
                if (avctx->flags & CODEC_FLAG_EMU_EDGE) { // mb_x+x or mb_y+y is a hack but works
1126
                    mode = check_intra_pred4x4_mode_emuedge(intra4x4[x], mb_x + x, mb_y + y, &copy);
1127
                    if (copy) {
1128
                        dst = copy_dst + 12;
1129
                        linesize = 8;
1130
                        if (!(mb_y + y)) {
1131
                            copy_dst[3] = 127U;
1132
                            AV_WN32A(copy_dst+4, 127U * 0x01010101U);
1133
                        } else {
1134
                            AV_COPY32(copy_dst+4, ptr+4*x-s->linesize);
1135
                            if (!(mb_x + x)) {
1136
                                copy_dst[3] = 129U;
1137
                            } else {
1138
                                copy_dst[3] = ptr[4*x-s->linesize-1];
1139
                            }
1140
                        }
1141
                        if (!(mb_x + x)) {
1142
                            copy_dst[11] =
1143
                            copy_dst[19] =
1144
                            copy_dst[27] =
1145
                            copy_dst[35] = 129U;
1146
                        } else {
1147
                            copy_dst[11] = ptr[4*x              -1];
1148
                            copy_dst[19] = ptr[4*x+s->linesize  -1];
1149
                            copy_dst[27] = ptr[4*x+s->linesize*2-1];
1150
                            copy_dst[35] = ptr[4*x+s->linesize*3-1];
1151
                        }
1152
                    }
1153
                } else {
1154
                    mode = intra4x4[x];
1155
                }
1156
                s->hpc.pred4x4[mode](dst, topright, linesize);
1157
                if (copy) {
1158
                    AV_COPY32(ptr+4*x              , copy_dst+12);
1159
                    AV_COPY32(ptr+4*x+s->linesize  , copy_dst+20);
1160
                    AV_COPY32(ptr+4*x+s->linesize*2, copy_dst+28);
1161
                    AV_COPY32(ptr+4*x+s->linesize*3, copy_dst+36);
1162
                }
1163

    
1164
                nnz = s->non_zero_count_cache[y][x];
1165
                if (nnz) {
1166
                    if (nnz == 1)
1167
                        s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize);
1168
                    else
1169
                        s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize);
1170
                }
1171
                topright += 4;
1172
            }
1173

    
1174
            ptr   += 4*s->linesize;
1175
            intra4x4 += 4;
1176
        }
1177
    }
1178

    
1179
    if (avctx->flags & CODEC_FLAG_EMU_EDGE) {
1180
        mode = check_intra_pred8x8_mode_emuedge(s->chroma_pred_mode, mb_x, mb_y);
1181
    } else {
1182
        mode = check_intra_pred8x8_mode(s->chroma_pred_mode, mb_x, mb_y);
1183
    }
1184
    s->hpc.pred8x8[mode](dst[1], s->uvlinesize);
1185
    s->hpc.pred8x8[mode](dst[2], s->uvlinesize);
1186

    
1187
    if (!(avctx->flags & CODEC_FLAG_EMU_EDGE && !mb_y) && (s->deblock_filter || !mb_y))
1188
        xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
1189
                       s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
1190
                       s->filter.simple, 0);
1191
}
1192

    
1193
static const uint8_t subpel_idx[3][8] = {
1194
    { 0, 1, 2, 1, 2, 1, 2, 1 }, // nr. of left extra pixels,
1195
                                // also function pointer index
1196
    { 0, 3, 5, 3, 5, 3, 5, 3 }, // nr. of extra pixels required
1197
    { 0, 2, 3, 2, 3, 2, 3, 2 }, // nr. of right extra pixels
1198
};
1199

    
1200
/**
1201
 * Generic MC function.
1202
 *
1203
 * @param s VP8 decoding context
1204
 * @param luma 1 for luma (Y) planes, 0 for chroma (Cb/Cr) planes
1205
 * @param dst target buffer for block data at block position
1206
 * @param src reference picture buffer at origin (0, 0)
1207
 * @param mv motion vector (relative to block position) to get pixel data from
1208
 * @param x_off horizontal position of block from origin (0, 0)
1209
 * @param y_off vertical position of block from origin (0, 0)
1210
 * @param block_w width of block (16, 8 or 4)
1211
 * @param block_h height of block (always same as block_w)
1212
 * @param width width of src/dst plane data
1213
 * @param height height of src/dst plane data
1214
 * @param linesize size of a single line of plane data, including padding
1215
 * @param mc_func motion compensation function pointers (bilinear or sixtap MC)
1216
 */
1217
static av_always_inline
1218
void vp8_mc_luma(VP8Context *s, uint8_t *dst, uint8_t *src, const VP56mv *mv,
1219
                 int x_off, int y_off, int block_w, int block_h,
1220
                 int width, int height, int linesize,
1221
                 vp8_mc_func mc_func[3][3])
1222
{
1223
    if (AV_RN32A(mv)) {
1224

    
1225
        int mx = (mv->x << 1)&7, mx_idx = subpel_idx[0][mx];
1226
        int my = (mv->y << 1)&7, my_idx = subpel_idx[0][my];
1227

    
1228
        x_off += mv->x >> 2;
1229
        y_off += mv->y >> 2;
1230

    
1231
        // edge emulation
1232
        src += y_off * linesize + x_off;
1233
        if (x_off < mx_idx || x_off >= width  - block_w - subpel_idx[2][mx] ||
1234
            y_off < my_idx || y_off >= height - block_h - subpel_idx[2][my]) {
1235
            s->dsp.emulated_edge_mc(s->edge_emu_buffer, src - my_idx * linesize - mx_idx, linesize,
1236
                                    block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],
1237
                                    x_off - mx_idx, y_off - my_idx, width, height);
1238
            src = s->edge_emu_buffer + mx_idx + linesize * my_idx;
1239
        }
1240
        mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);
1241
    } else
1242
        mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0);
1243
}
1244

    
1245
static av_always_inline
1246
void vp8_mc_chroma(VP8Context *s, uint8_t *dst1, uint8_t *dst2, uint8_t *src1,
1247
                   uint8_t *src2, const VP56mv *mv, int x_off, int y_off,
1248
                   int block_w, int block_h, int width, int height, int linesize,
1249
                   vp8_mc_func mc_func[3][3])
1250
{
1251
    if (AV_RN32A(mv)) {
1252
        int mx = mv->x&7, mx_idx = subpel_idx[0][mx];
1253
        int my = mv->y&7, my_idx = subpel_idx[0][my];
1254

    
1255
        x_off += mv->x >> 3;
1256
        y_off += mv->y >> 3;
1257

    
1258
        // edge emulation
1259
        src1 += y_off * linesize + x_off;
1260
        src2 += y_off * linesize + x_off;
1261
        if (x_off < mx_idx || x_off >= width  - block_w - subpel_idx[2][mx] ||
1262
            y_off < my_idx || y_off >= height - block_h - subpel_idx[2][my]) {
1263
            s->dsp.emulated_edge_mc(s->edge_emu_buffer, src1 - my_idx * linesize - mx_idx, linesize,
1264
                                    block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],
1265
                                    x_off - mx_idx, y_off - my_idx, width, height);
1266
            src1 = s->edge_emu_buffer + mx_idx + linesize * my_idx;
1267
            mc_func[my_idx][mx_idx](dst1, linesize, src1, linesize, block_h, mx, my);
1268

    
1269
            s->dsp.emulated_edge_mc(s->edge_emu_buffer, src2 - my_idx * linesize - mx_idx, linesize,
1270
                                    block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],
1271
                                    x_off - mx_idx, y_off - my_idx, width, height);
1272
            src2 = s->edge_emu_buffer + mx_idx + linesize * my_idx;
1273
            mc_func[my_idx][mx_idx](dst2, linesize, src2, linesize, block_h, mx, my);
1274
        } else {
1275
            mc_func[my_idx][mx_idx](dst1, linesize, src1, linesize, block_h, mx, my);
1276
            mc_func[my_idx][mx_idx](dst2, linesize, src2, linesize, block_h, mx, my);
1277
        }
1278
    } else {
1279
        mc_func[0][0](dst1, linesize, src1 + y_off * linesize + x_off, linesize, block_h, 0, 0);
1280
        mc_func[0][0](dst2, linesize, src2 + y_off * linesize + x_off, linesize, block_h, 0, 0);
1281
    }
1282
}
1283

    
1284
static av_always_inline
1285
void vp8_mc_part(VP8Context *s, uint8_t *dst[3],
1286
                 AVFrame *ref_frame, int x_off, int y_off,
1287
                 int bx_off, int by_off,
1288
                 int block_w, int block_h,
1289
                 int width, int height, VP56mv *mv)
1290
{
1291
    VP56mv uvmv = *mv;
1292

    
1293
    /* Y */
1294
    vp8_mc_luma(s, dst[0] + by_off * s->linesize + bx_off,
1295
                ref_frame->data[0], mv, x_off + bx_off, y_off + by_off,
1296
                block_w, block_h, width, height, s->linesize,
1297
                s->put_pixels_tab[block_w == 8]);
1298

    
1299
    /* U/V */
1300
    if (s->profile == 3) {
1301
        uvmv.x &= ~7;
1302
        uvmv.y &= ~7;
1303
    }
1304
    x_off   >>= 1; y_off   >>= 1;
1305
    bx_off  >>= 1; by_off  >>= 1;
1306
    width   >>= 1; height  >>= 1;
1307
    block_w >>= 1; block_h >>= 1;
1308
    vp8_mc_chroma(s, dst[1] + by_off * s->uvlinesize + bx_off,
1309
                  dst[2] + by_off * s->uvlinesize + bx_off, ref_frame->data[1],
1310
                  ref_frame->data[2], &uvmv, x_off + bx_off, y_off + by_off,
1311
                  block_w, block_h, width, height, s->uvlinesize,
1312
                  s->put_pixels_tab[1 + (block_w == 4)]);
1313
}
1314

    
1315
/* Fetch pixels for estimated mv 4 macroblocks ahead.
1316
 * Optimized for 64-byte cache lines.  Inspired by ffh264 prefetch_motion. */
1317
static av_always_inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int mb_xy, int ref)
1318
{
1319
    /* Don't prefetch refs that haven't been used very often this frame. */
1320
    if (s->ref_count[ref-1] > (mb_xy >> 5)) {
1321
        int x_off = mb_x << 4, y_off = mb_y << 4;
1322
        int mx = (mb->mv.x>>2) + x_off + 8;
1323
        int my = (mb->mv.y>>2) + y_off;
1324
        uint8_t **src= s->framep[ref]->data;
1325
        int off= mx + (my + (mb_x&3)*4)*s->linesize + 64;
1326
        s->dsp.prefetch(src[0]+off, s->linesize, 4);
1327
        off= (mx>>1) + ((my>>1) + (mb_x&7))*s->uvlinesize + 64;
1328
        s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1329
    }
1330
}
1331

    
1332
/**
1333
 * Apply motion vectors to prediction buffer, chapter 18.
1334
 */
1335
static av_always_inline
1336
void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
1337
                   int mb_x, int mb_y)
1338
{
1339
    int x_off = mb_x << 4, y_off = mb_y << 4;
1340
    int width = 16*s->mb_width, height = 16*s->mb_height;
1341
    AVFrame *ref = s->framep[mb->ref_frame];
1342
    VP56mv *bmv = mb->bmv;
1343

    
1344
    switch (mb->partitioning) {
1345
    case VP8_SPLITMVMODE_NONE:
1346
        vp8_mc_part(s, dst, ref, x_off, y_off,
1347
                    0, 0, 16, 16, width, height, &mb->mv);
1348
        break;
1349
    case VP8_SPLITMVMODE_4x4: {
1350
        int x, y;
1351
        VP56mv uvmv;
1352

    
1353
        /* Y */
1354
        for (y = 0; y < 4; y++) {
1355
            for (x = 0; x < 4; x++) {
1356
                vp8_mc_luma(s, dst[0] + 4*y*s->linesize + x*4,
1357
                            ref->data[0], &bmv[4*y + x],
1358
                            4*x + x_off, 4*y + y_off, 4, 4,
1359
                            width, height, s->linesize,
1360
                            s->put_pixels_tab[2]);
1361
            }
1362
        }
1363

    
1364
        /* U/V */
1365
        x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1;
1366
        for (y = 0; y < 2; y++) {
1367
            for (x = 0; x < 2; x++) {
1368
                uvmv.x = mb->bmv[ 2*y    * 4 + 2*x  ].x +
1369
                         mb->bmv[ 2*y    * 4 + 2*x+1].x +
1370
                         mb->bmv[(2*y+1) * 4 + 2*x  ].x +
1371
                         mb->bmv[(2*y+1) * 4 + 2*x+1].x;
1372
                uvmv.y = mb->bmv[ 2*y    * 4 + 2*x  ].y +
1373
                         mb->bmv[ 2*y    * 4 + 2*x+1].y +
1374
                         mb->bmv[(2*y+1) * 4 + 2*x  ].y +
1375
                         mb->bmv[(2*y+1) * 4 + 2*x+1].y;
1376
                uvmv.x = (uvmv.x + 2 + (uvmv.x >> (INT_BIT-1))) >> 2;
1377
                uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2;
1378
                if (s->profile == 3) {
1379
                    uvmv.x &= ~7;
1380
                    uvmv.y &= ~7;
1381
                }
1382
                vp8_mc_chroma(s, dst[1] + 4*y*s->uvlinesize + x*4,
1383
                              dst[2] + 4*y*s->uvlinesize + x*4,
1384
                              ref->data[1], ref->data[2], &uvmv,
1385
                              4*x + x_off, 4*y + y_off, 4, 4,
1386
                              width, height, s->uvlinesize,
1387
                              s->put_pixels_tab[2]);
1388
            }
1389
        }
1390
        break;
1391
    }
1392
    case VP8_SPLITMVMODE_16x8:
1393
        vp8_mc_part(s, dst, ref, x_off, y_off,
1394
                    0, 0, 16, 8, width, height, &bmv[0]);
1395
        vp8_mc_part(s, dst, ref, x_off, y_off,
1396
                    0, 8, 16, 8, width, height, &bmv[1]);
1397
        break;
1398
    case VP8_SPLITMVMODE_8x16:
1399
        vp8_mc_part(s, dst, ref, x_off, y_off,
1400
                    0, 0, 8, 16, width, height, &bmv[0]);
1401
        vp8_mc_part(s, dst, ref, x_off, y_off,
1402
                    8, 0, 8, 16, width, height, &bmv[1]);
1403
        break;
1404
    case VP8_SPLITMVMODE_8x8:
1405
        vp8_mc_part(s, dst, ref, x_off, y_off,
1406
                    0, 0, 8, 8, width, height, &bmv[0]);
1407
        vp8_mc_part(s, dst, ref, x_off, y_off,
1408
                    8, 0, 8, 8, width, height, &bmv[1]);
1409
        vp8_mc_part(s, dst, ref, x_off, y_off,
1410
                    0, 8, 8, 8, width, height, &bmv[2]);
1411
        vp8_mc_part(s, dst, ref, x_off, y_off,
1412
                    8, 8, 8, 8, width, height, &bmv[3]);
1413
        break;
1414
    }
1415
}
1416

    
1417
static av_always_inline void idct_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb)
1418
{
1419
    int x, y, ch;
1420

    
1421
    if (mb->mode != MODE_I4x4) {
1422
        uint8_t *y_dst = dst[0];
1423
        for (y = 0; y < 4; y++) {
1424
            uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[y]);
1425
            if (nnz4) {
1426
                if (nnz4&~0x01010101) {
1427
                    for (x = 0; x < 4; x++) {
1428
                        int nnz = s->non_zero_count_cache[y][x];
1429
                        if (nnz) {
1430
                            if (nnz == 1)
1431
                                s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize);
1432
                            else
1433
                                s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize);
1434
                        }
1435
                    }
1436
                } else {
1437
                    s->vp8dsp.vp8_idct_dc_add4y(y_dst, s->block[y], s->linesize);
1438
                }
1439
            }
1440
            y_dst += 4*s->linesize;
1441
        }
1442
    }
1443

    
1444
    for (ch = 0; ch < 2; ch++) {
1445
        uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[4+ch]);
1446
        if (nnz4) {
1447
            uint8_t *ch_dst = dst[1+ch];
1448
            if (nnz4&~0x01010101) {
1449
                for (y = 0; y < 2; y++) {
1450
                    for (x = 0; x < 2; x++) {
1451
                        int nnz = s->non_zero_count_cache[4+ch][(y<<1)+x];
1452
                        if (nnz) {
1453
                            if (nnz == 1)
1454
                                s->vp8dsp.vp8_idct_dc_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);
1455
                            else
1456
                                s->vp8dsp.vp8_idct_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);
1457
                        }
1458
                    }
1459
                    ch_dst += 4*s->uvlinesize;
1460
                }
1461
            } else {
1462
                s->vp8dsp.vp8_idct_dc_add4uv(ch_dst, s->block[4+ch], s->uvlinesize);
1463
            }
1464
        }
1465
    }
1466
}
1467

    
1468
static av_always_inline void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f )
1469
{
1470
    int interior_limit, filter_level;
1471

    
1472
    if (s->segmentation.enabled) {
1473
        filter_level = s->segmentation.filter_level[s->segment];
1474
        if (!s->segmentation.absolute_vals)
1475
            filter_level += s->filter.level;
1476
    } else
1477
        filter_level = s->filter.level;
1478

    
1479
    if (s->lf_delta.enabled) {
1480
        filter_level += s->lf_delta.ref[mb->ref_frame];
1481
        filter_level += s->lf_delta.mode[mb->mode];
1482
    }
1483
    filter_level = av_clip(filter_level, 0, 63);
1484

    
1485
    interior_limit = filter_level;
1486
    if (s->filter.sharpness) {
1487
        interior_limit >>= s->filter.sharpness > 4 ? 2 : 1;
1488
        interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness);
1489
    }
1490
    interior_limit = FFMAX(interior_limit, 1);
1491

    
1492
    f->filter_level = filter_level;
1493
    f->inner_limit = interior_limit;
1494
    f->inner_filter = !mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT;
1495
}
1496

    
1497
static av_always_inline void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f, int mb_x, int mb_y)
1498
{
1499
    int mbedge_lim, bedge_lim, hev_thresh;
1500
    int filter_level = f->filter_level;
1501
    int inner_limit = f->inner_limit;
1502
    int inner_filter = f->inner_filter;
1503
    int linesize = s->linesize;
1504
    int uvlinesize = s->uvlinesize;
1505

    
1506
    if (!filter_level)
1507
        return;
1508

    
1509
    mbedge_lim = 2*(filter_level+2) + inner_limit;
1510
     bedge_lim = 2* filter_level    + inner_limit;
1511
    hev_thresh = filter_level >= 15;
1512

    
1513
    if (s->keyframe) {
1514
        if (filter_level >= 40)
1515
            hev_thresh = 2;
1516
    } else {
1517
        if (filter_level >= 40)
1518
            hev_thresh = 3;
1519
        else if (filter_level >= 20)
1520
            hev_thresh = 2;
1521
    }
1522

    
1523
    if (mb_x) {
1524
        s->vp8dsp.vp8_h_loop_filter16y(dst[0],     linesize,
1525
                                       mbedge_lim, inner_limit, hev_thresh);
1526
        s->vp8dsp.vp8_h_loop_filter8uv(dst[1],     dst[2],      uvlinesize,
1527
                                       mbedge_lim, inner_limit, hev_thresh);
1528
    }
1529

    
1530
    if (inner_filter) {
1531
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 4, linesize, bedge_lim,
1532
                                             inner_limit, hev_thresh);
1533
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 8, linesize, bedge_lim,
1534
                                             inner_limit, hev_thresh);
1535
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+12, linesize, bedge_lim,
1536
                                             inner_limit, hev_thresh);
1537
        s->vp8dsp.vp8_h_loop_filter8uv_inner(dst[1] + 4, dst[2] + 4,
1538
                                             uvlinesize,  bedge_lim,
1539
                                             inner_limit, hev_thresh);
1540
    }
1541

    
1542
    if (mb_y) {
1543
        s->vp8dsp.vp8_v_loop_filter16y(dst[0],     linesize,
1544
                                       mbedge_lim, inner_limit, hev_thresh);
1545
        s->vp8dsp.vp8_v_loop_filter8uv(dst[1],     dst[2],      uvlinesize,
1546
                                       mbedge_lim, inner_limit, hev_thresh);
1547
    }
1548

    
1549
    if (inner_filter) {
1550
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 4*linesize,
1551
                                             linesize,    bedge_lim,
1552
                                             inner_limit, hev_thresh);
1553
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 8*linesize,
1554
                                             linesize,    bedge_lim,
1555
                                             inner_limit, hev_thresh);
1556
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+12*linesize,
1557
                                             linesize,    bedge_lim,
1558
                                             inner_limit, hev_thresh);
1559
        s->vp8dsp.vp8_v_loop_filter8uv_inner(dst[1] + 4 * uvlinesize,
1560
                                             dst[2] + 4 * uvlinesize,
1561
                                             uvlinesize,  bedge_lim,
1562
                                             inner_limit, hev_thresh);
1563
    }
1564
}
1565

    
1566
static av_always_inline void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f, int mb_x, int mb_y)
1567
{
1568
    int mbedge_lim, bedge_lim;
1569
    int filter_level = f->filter_level;
1570
    int inner_limit = f->inner_limit;
1571
    int inner_filter = f->inner_filter;
1572
    int linesize = s->linesize;
1573

    
1574
    if (!filter_level)
1575
        return;
1576

    
1577
    mbedge_lim = 2*(filter_level+2) + inner_limit;
1578
     bedge_lim = 2* filter_level    + inner_limit;
1579

    
1580
    if (mb_x)
1581
        s->vp8dsp.vp8_h_loop_filter_simple(dst, linesize, mbedge_lim);
1582
    if (inner_filter) {
1583
        s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, linesize, bedge_lim);
1584
        s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, linesize, bedge_lim);
1585
        s->vp8dsp.vp8_h_loop_filter_simple(dst+12, linesize, bedge_lim);
1586
    }
1587

    
1588
    if (mb_y)
1589
        s->vp8dsp.vp8_v_loop_filter_simple(dst, linesize, mbedge_lim);
1590
    if (inner_filter) {
1591
        s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*linesize, linesize, bedge_lim);
1592
        s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*linesize, linesize, bedge_lim);
1593
        s->vp8dsp.vp8_v_loop_filter_simple(dst+12*linesize, linesize, bedge_lim);
1594
    }
1595
}
1596

    
1597
static void filter_mb_row(VP8Context *s, int mb_y)
1598
{
1599
    VP8FilterStrength *f = s->filter_strength;
1600
    uint8_t *dst[3] = {
1601
        s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize,
1602
        s->framep[VP56_FRAME_CURRENT]->data[1] +  8*mb_y*s->uvlinesize,
1603
        s->framep[VP56_FRAME_CURRENT]->data[2] +  8*mb_y*s->uvlinesize
1604
    };
1605
    int mb_x;
1606

    
1607
    for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1608
        backup_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], s->linesize, s->uvlinesize, 0);
1609
        filter_mb(s, dst, f++, mb_x, mb_y);
1610
        dst[0] += 16;
1611
        dst[1] += 8;
1612
        dst[2] += 8;
1613
    }
1614
}
1615

    
1616
static void filter_mb_row_simple(VP8Context *s, int mb_y)
1617
{
1618
    VP8FilterStrength *f = s->filter_strength;
1619
    uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize;
1620
    int mb_x;
1621

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

    
1629
static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
1630
                            AVPacket *avpkt)
1631
{
1632
    VP8Context *s = avctx->priv_data;
1633
    int ret, mb_x, mb_y, i, y, referenced;
1634
    enum AVDiscard skip_thresh;
1635
    AVFrame *av_uninit(curframe);
1636

    
1637
    if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0)
1638
        return ret;
1639

    
1640
    referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT
1641
                                || s->update_altref == VP56_FRAME_CURRENT;
1642

    
1643
    skip_thresh = !referenced ? AVDISCARD_NONREF :
1644
                    !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL;
1645

    
1646
    if (avctx->skip_frame >= skip_thresh) {
1647
        s->invisible = 1;
1648
        goto skip_decode;
1649
    }
1650
    s->deblock_filter = s->filter.level && avctx->skip_loop_filter < skip_thresh;
1651

    
1652
    for (i = 0; i < 4; i++)
1653
        if (&s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1654
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1655
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) {
1656
            curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i];
1657
            break;
1658
        }
1659
    if (curframe->data[0])
1660
        avctx->release_buffer(avctx, curframe);
1661

    
1662
    curframe->key_frame = s->keyframe;
1663
    curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE;
1664
    curframe->reference = referenced ? 3 : 0;
1665
    if ((ret = avctx->get_buffer(avctx, curframe))) {
1666
        av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n");
1667
        return ret;
1668
    }
1669

    
1670
    // Given that arithmetic probabilities are updated every frame, it's quite likely
1671
    // that the values we have on a random interframe are complete junk if we didn't
1672
    // start decode on a keyframe. So just don't display anything rather than junk.
1673
    if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] ||
1674
                         !s->framep[VP56_FRAME_GOLDEN] ||
1675
                         !s->framep[VP56_FRAME_GOLDEN2])) {
1676
        av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n");
1677
        return AVERROR_INVALIDDATA;
1678
    }
1679

    
1680
    s->linesize   = curframe->linesize[0];
1681
    s->uvlinesize = curframe->linesize[1];
1682

    
1683
    if (!s->edge_emu_buffer)
1684
        s->edge_emu_buffer = av_malloc(21*s->linesize);
1685

    
1686
    memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz));
1687

    
1688
    /* Zero macroblock structures for top/top-left prediction from outside the frame. */
1689
    memset(s->macroblocks + s->mb_height*2 - 1, 0, (s->mb_width+1)*sizeof(*s->macroblocks));
1690

    
1691
    // top edge of 127 for intra prediction
1692
    if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) {
1693
        s->top_border[0][15] = s->top_border[0][23] = 127;
1694
        memset(s->top_border[1]-1, 127, s->mb_width*sizeof(*s->top_border)+1);
1695
    }
1696
    memset(s->ref_count, 0, sizeof(s->ref_count));
1697
    if (s->keyframe)
1698
        memset(s->intra4x4_pred_mode_top, DC_PRED, s->mb_width*4);
1699

    
1700
    for (mb_y = 0; mb_y < s->mb_height; mb_y++) {
1701
        VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)];
1702
        VP8Macroblock *mb = s->macroblocks + (s->mb_height - mb_y - 1)*2;
1703
        int mb_xy = mb_y*s->mb_width;
1704
        uint8_t *dst[3] = {
1705
            curframe->data[0] + 16*mb_y*s->linesize,
1706
            curframe->data[1] +  8*mb_y*s->uvlinesize,
1707
            curframe->data[2] +  8*mb_y*s->uvlinesize
1708
        };
1709

    
1710
        memset(mb - 1, 0, sizeof(*mb));   // zero left macroblock
1711
        memset(s->left_nnz, 0, sizeof(s->left_nnz));
1712
        AV_WN32A(s->intra4x4_pred_mode_left, DC_PRED*0x01010101);
1713

    
1714
        // left edge of 129 for intra prediction
1715
        if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) {
1716
            for (i = 0; i < 3; i++)
1717
                for (y = 0; y < 16>>!!i; y++)
1718
                    dst[i][y*curframe->linesize[i]-1] = 129;
1719
            if (mb_y == 1) // top left edge is also 129
1720
                s->top_border[0][15] = s->top_border[0][23] = s->top_border[0][31] = 129;
1721
        }
1722

    
1723
        for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb_xy++, mb++) {
1724
            /* Prefetch the current frame, 4 MBs ahead */
1725
            s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4);
1726
            s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2);
1727

    
1728
            decode_mb_mode(s, mb, mb_x, mb_y, s->segmentation_map + mb_xy);
1729

    
1730
            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_PREVIOUS);
1731

    
1732
            if (!mb->skip)
1733
                decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);
1734

    
1735
            if (mb->mode <= MODE_I4x4)
1736
                intra_predict(s, dst, mb, mb_x, mb_y);
1737
            else
1738
                inter_predict(s, dst, mb, mb_x, mb_y);
1739

    
1740
            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN);
1741

    
1742
            if (!mb->skip) {
1743
                idct_mb(s, dst, mb);
1744
            } else {
1745
                AV_ZERO64(s->left_nnz);
1746
                AV_WN64(s->top_nnz[mb_x], 0);   // array of 9, so unaligned
1747

    
1748
                // Reset DC block predictors if they would exist if the mb had coefficients
1749
                if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
1750
                    s->left_nnz[8]      = 0;
1751
                    s->top_nnz[mb_x][8] = 0;
1752
                }
1753
            }
1754

    
1755
            if (s->deblock_filter)
1756
                filter_level_for_mb(s, mb, &s->filter_strength[mb_x]);
1757

    
1758
            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN2);
1759

    
1760
            dst[0] += 16;
1761
            dst[1] += 8;
1762
            dst[2] += 8;
1763
        }
1764
        if (s->deblock_filter) {
1765
            if (s->filter.simple)
1766
                filter_mb_row_simple(s, mb_y);
1767
            else
1768
                filter_mb_row(s, mb_y);
1769
        }
1770
    }
1771

    
1772
skip_decode:
1773
    // if future frames don't use the updated probabilities,
1774
    // reset them to the values we saved
1775
    if (!s->update_probabilities)
1776
        s->prob[0] = s->prob[1];
1777

    
1778
    // check if golden and altref are swapped
1779
    if (s->update_altref == VP56_FRAME_GOLDEN &&
1780
        s->update_golden == VP56_FRAME_GOLDEN2)
1781
        FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN], s->framep[VP56_FRAME_GOLDEN2]);
1782
    else {
1783
        if (s->update_altref != VP56_FRAME_NONE)
1784
            s->framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref];
1785

    
1786
        if (s->update_golden != VP56_FRAME_NONE)
1787
            s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden];
1788
    }
1789

    
1790
    if (s->update_last) // move cur->prev
1791
        s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT];
1792

    
1793
    // release no longer referenced frames
1794
    for (i = 0; i < 4; i++)
1795
        if (s->frames[i].data[0] &&
1796
            &s->frames[i] != s->framep[VP56_FRAME_CURRENT] &&
1797
            &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1798
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1799
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2])
1800
            avctx->release_buffer(avctx, &s->frames[i]);
1801

    
1802
    if (!s->invisible) {
1803
        *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT];
1804
        *data_size = sizeof(AVFrame);
1805
    }
1806

    
1807
    return avpkt->size;
1808
}
1809

    
1810
static av_cold int vp8_decode_init(AVCodecContext *avctx)
1811
{
1812
    VP8Context *s = avctx->priv_data;
1813

    
1814
    s->avctx = avctx;
1815
    avctx->pix_fmt = PIX_FMT_YUV420P;
1816

    
1817
    dsputil_init(&s->dsp, avctx);
1818
    ff_h264_pred_init(&s->hpc, CODEC_ID_VP8);
1819
    ff_vp8dsp_init(&s->vp8dsp);
1820

    
1821
    return 0;
1822
}
1823

    
1824
static av_cold int vp8_decode_free(AVCodecContext *avctx)
1825
{
1826
    vp8_decode_flush(avctx);
1827
    return 0;
1828
}
1829

    
1830
AVCodec ff_vp8_decoder = {
1831
    "vp8",
1832
    AVMEDIA_TYPE_VIDEO,
1833
    CODEC_ID_VP8,
1834
    sizeof(VP8Context),
1835
    vp8_decode_init,
1836
    NULL,
1837
    vp8_decode_free,
1838
    vp8_decode_frame,
1839
    CODEC_CAP_DR1,
1840
    .flush = vp8_decode_flush,
1841
    .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"),
1842
};