Statistics
| Branch: | Revision:

ffmpeg / libavcodec / vp8.c @ 10bf2eeb

History | View | Annotate | Download (58.2 KB)

1 3b636f21 David Conrad
/**
2
 * VP8 compatible video decoder
3
 *
4
 * Copyright (C) 2010 David Conrad
5
 * Copyright (C) 2010 Ronald S. Bultje
6 13a1304b Jason Garrett-Glaser
 * Copyright (C) 2010 Jason Garrett-Glaser
7 3b636f21 David Conrad
 *
8
 * This file is part of FFmpeg.
9
 *
10
 * FFmpeg is free software; you can redistribute it and/or
11
 * modify it under the terms of the GNU Lesser General Public
12
 * License as published by the Free Software Foundation; either
13
 * version 2.1 of the License, or (at your option) any later version.
14
 *
15
 * FFmpeg is distributed in the hope that it will be useful,
16
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18
 * Lesser General Public License for more details.
19
 *
20
 * You should have received a copy of the GNU Lesser General Public
21
 * License along with FFmpeg; if not, write to the Free Software
22
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23
 */
24
25
#include "avcodec.h"
26
#include "vp56.h"
27
#include "vp8data.h"
28
#include "vp8dsp.h"
29
#include "h264pred.h"
30
#include "rectangle.h"
31
32
typedef struct {
33 968570d6 Jason Garrett-Glaser
    uint8_t filter_level;
34
    uint8_t inner_limit;
35 c55e0d34 Jason Garrett-Glaser
    uint8_t inner_filter;
36 968570d6 Jason Garrett-Glaser
} VP8FilterStrength;
37
38
typedef struct {
39 3b636f21 David Conrad
    uint8_t skip;
40
    // todo: make it possible to check for at least (i4x4 or split_mv)
41
    // in one op. are others needed?
42
    uint8_t mode;
43
    uint8_t ref_frame;
44
    uint8_t partitioning;
45
    VP56mv mv;
46
    VP56mv bmv[16];
47
} VP8Macroblock;
48
49
typedef struct {
50
    AVCodecContext *avctx;
51
    DSPContext dsp;
52
    VP8DSPContext vp8dsp;
53
    H264PredContext hpc;
54 0ef1dbed David Conrad
    vp8_mc_func put_pixels_tab[3][3][3];
55 3b636f21 David Conrad
    AVFrame frames[4];
56
    AVFrame *framep[4];
57
    uint8_t *edge_emu_buffer;
58
    VP56RangeCoder c;   ///< header context, includes mb modes and motion vectors
59
    int profile;
60
61
    int mb_width;   /* number of horizontal MB */
62
    int mb_height;  /* number of vertical MB */
63
    int linesize;
64
    int uvlinesize;
65
66
    int keyframe;
67
    int invisible;
68
    int update_last;    ///< update VP56_FRAME_PREVIOUS with the current one
69
    int update_golden;  ///< VP56_FRAME_NONE if not updated, or which frame to copy if so
70
    int update_altref;
71 9ac831c2 David Conrad
    int deblock_filter;
72 3b636f21 David Conrad
73
    /**
74
     * If this flag is not set, all the probability updates
75
     * are discarded after this frame is decoded.
76
     */
77
    int update_probabilities;
78
79
    /**
80
     * All coefficients are contained in separate arith coding contexts.
81
     * There can be 1, 2, 4, or 8 of these after the header context.
82
     */
83
    int num_coeff_partitions;
84
    VP56RangeCoder coeff_partition[8];
85
86
    VP8Macroblock *macroblocks;
87
    VP8Macroblock *macroblocks_base;
88 968570d6 Jason Garrett-Glaser
    VP8FilterStrength *filter_strength;
89 3b636f21 David Conrad
    int mb_stride;
90
91
    uint8_t *intra4x4_pred_mode;
92
    uint8_t *intra4x4_pred_mode_base;
93 c55e0d34 Jason Garrett-Glaser
    uint8_t *segmentation_map;
94 3b636f21 David Conrad
    int b4_stride;
95
96
    /**
97 9ac831c2 David Conrad
     * Cache of the top row needed for intra prediction
98
     * 16 for luma, 8 for each chroma plane
99
     */
100
    uint8_t (*top_border)[16+8+8];
101
102
    /**
103 3b636f21 David Conrad
     * For coeff decode, we need to know whether the above block had non-zero
104
     * coefficients. This means for each macroblock, we need data for 4 luma
105
     * blocks, 2 u blocks, 2 v blocks, and the luma dc block, for a total of 9
106
     * per macroblock. We keep the last row in top_nnz.
107
     */
108
    uint8_t (*top_nnz)[9];
109
    DECLARE_ALIGNED(8, uint8_t, left_nnz)[9];
110
111
    /**
112
     * This is the index plus one of the last non-zero coeff
113
     * for each of the blocks in the current macroblock.
114
     * So, 0 -> no coeffs
115
     *     1 -> dc-only (special transform)
116
     *     2+-> full transform
117
     */
118
    DECLARE_ALIGNED(16, uint8_t, non_zero_count_cache)[6][4];
119
    DECLARE_ALIGNED(16, DCTELEM, block)[6][4][16];
120 d1c58fce Jason Garrett-Glaser
    uint8_t intra4x4_pred_mode_mb[16];
121 3b636f21 David Conrad
122
    int chroma_pred_mode;    ///< 8x8c pred mode of the current macroblock
123 b9a7186b Jason Garrett-Glaser
    int segment;             ///< segment of the current macroblock
124 3b636f21 David Conrad
125
    int mbskip_enabled;
126
    int sign_bias[4]; ///< one state [0, 1] per ref frame type
127 c4211046 Jason Garrett-Glaser
    int ref_count[3];
128 3b636f21 David Conrad
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 a8ab0ccc Pascal Massimino
        uint8_t mbskip;
192
        uint8_t intra;
193
        uint8_t last;
194
        uint8_t golden;
195 3b636f21 David Conrad
        uint8_t pred16x16[4];
196
        uint8_t pred8x8c[3];
197
        uint8_t token[4][8][3][NUM_DCT_TOKENS-1];
198
        uint8_t mvc[2][19];
199
    } prob[2];
200
} VP8Context;
201
202
static void vp8_decode_flush(AVCodecContext *avctx)
203
{
204
    VP8Context *s = avctx->priv_data;
205
    int i;
206
207
    for (i = 0; i < 4; i++)
208
        if (s->frames[i].data[0])
209
            avctx->release_buffer(avctx, &s->frames[i]);
210
    memset(s->framep, 0, sizeof(s->framep));
211
212
    av_freep(&s->macroblocks_base);
213 fca05ea8 Jason Garrett-Glaser
    av_freep(&s->filter_strength);
214 3b636f21 David Conrad
    av_freep(&s->intra4x4_pred_mode_base);
215
    av_freep(&s->top_nnz);
216
    av_freep(&s->edge_emu_buffer);
217 9ac831c2 David Conrad
    av_freep(&s->top_border);
218 c55e0d34 Jason Garrett-Glaser
    av_freep(&s->segmentation_map);
219 3b636f21 David Conrad
220
    s->macroblocks        = NULL;
221
    s->intra4x4_pred_mode = NULL;
222
}
223
224
static int update_dimensions(VP8Context *s, int width, int height)
225
{
226
    int i;
227
228
    if (avcodec_check_dimensions(s->avctx, width, height))
229
        return AVERROR_INVALIDDATA;
230
231
    vp8_decode_flush(s->avctx);
232
233
    avcodec_set_dimensions(s->avctx, width, height);
234
235
    s->mb_width  = (s->avctx->coded_width +15) / 16;
236
    s->mb_height = (s->avctx->coded_height+15) / 16;
237
238
    // we allocate a border around the top/left of intra4x4 modes
239
    // this is 4 blocks for intra4x4 to keep 4-byte alignment for fill_rectangle
240
    s->mb_stride = s->mb_width+1;
241
    s->b4_stride = 4*s->mb_stride;
242
243 c55e0d34 Jason Garrett-Glaser
    s->macroblocks_base        = av_mallocz((s->mb_stride+s->mb_height*2+2)*sizeof(*s->macroblocks));
244 968570d6 Jason Garrett-Glaser
    s->filter_strength         = av_mallocz(s->mb_stride*sizeof(*s->filter_strength));
245 3b636f21 David Conrad
    s->intra4x4_pred_mode_base = av_mallocz(s->b4_stride*(4*s->mb_height+1));
246
    s->top_nnz                 = av_mallocz(s->mb_width*sizeof(*s->top_nnz));
247 9ac831c2 David Conrad
    s->top_border              = av_mallocz((s->mb_width+1)*sizeof(*s->top_border));
248 c55e0d34 Jason Garrett-Glaser
    s->segmentation_map        = av_mallocz(s->mb_stride*s->mb_height);
249 3b636f21 David Conrad
250 c55e0d34 Jason Garrett-Glaser
    if (!s->macroblocks_base || !s->filter_strength || !s->intra4x4_pred_mode_base ||
251
        !s->top_nnz || !s->top_border || !s->segmentation_map)
252 b6c420ce David Conrad
        return AVERROR(ENOMEM);
253
254 c55e0d34 Jason Garrett-Glaser
    s->macroblocks        = s->macroblocks_base + 1;
255 3b636f21 David Conrad
    s->intra4x4_pred_mode = s->intra4x4_pred_mode_base + 4 + s->b4_stride;
256
257
    memset(s->intra4x4_pred_mode_base, DC_PRED, s->b4_stride);
258
    for (i = 0; i < 4*s->mb_height; i++)
259
        s->intra4x4_pred_mode[i*s->b4_stride-1] = DC_PRED;
260
261
    return 0;
262
}
263
264
static void parse_segment_info(VP8Context *s)
265
{
266
    VP56RangeCoder *c = &s->c;
267
    int i;
268
269
    s->segmentation.update_map = vp8_rac_get(c);
270
271
    if (vp8_rac_get(c)) { // update segment feature data
272
        s->segmentation.absolute_vals = vp8_rac_get(c);
273
274
        for (i = 0; i < 4; i++)
275
            s->segmentation.base_quant[i]   = vp8_rac_get_sint(c, 7);
276
277
        for (i = 0; i < 4; i++)
278
            s->segmentation.filter_level[i] = vp8_rac_get_sint(c, 6);
279
    }
280
    if (s->segmentation.update_map)
281
        for (i = 0; i < 3; i++)
282
            s->prob->segmentid[i] = vp8_rac_get(c) ? vp8_rac_get_uint(c, 8) : 255;
283
}
284
285
static void update_lf_deltas(VP8Context *s)
286
{
287
    VP56RangeCoder *c = &s->c;
288
    int i;
289
290
    for (i = 0; i < 4; i++)
291
        s->lf_delta.ref[i]  = vp8_rac_get_sint(c, 6);
292
293
    for (i = 0; i < 4; i++)
294
        s->lf_delta.mode[i] = vp8_rac_get_sint(c, 6);
295
}
296
297
static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size)
298
{
299
    const uint8_t *sizes = buf;
300
    int i;
301
302
    s->num_coeff_partitions = 1 << vp8_rac_get_uint(&s->c, 2);
303
304
    buf      += 3*(s->num_coeff_partitions-1);
305
    buf_size -= 3*(s->num_coeff_partitions-1);
306
    if (buf_size < 0)
307
        return -1;
308
309
    for (i = 0; i < s->num_coeff_partitions-1; i++) {
310 06d50ca8 Jason Garrett-Glaser
        int size = AV_RL24(sizes + 3*i);
311 3b636f21 David Conrad
        if (buf_size - size < 0)
312
            return -1;
313
314
        vp56_init_range_decoder(&s->coeff_partition[i], buf, size);
315
        buf      += size;
316
        buf_size -= size;
317
    }
318
    vp56_init_range_decoder(&s->coeff_partition[i], buf, buf_size);
319
320
    return 0;
321
}
322
323
static void get_quants(VP8Context *s)
324
{
325
    VP56RangeCoder *c = &s->c;
326
    int i, base_qi;
327
328
    int yac_qi     = vp8_rac_get_uint(c, 7);
329
    int ydc_delta  = vp8_rac_get_sint(c, 4);
330
    int y2dc_delta = vp8_rac_get_sint(c, 4);
331
    int y2ac_delta = vp8_rac_get_sint(c, 4);
332
    int uvdc_delta = vp8_rac_get_sint(c, 4);
333
    int uvac_delta = vp8_rac_get_sint(c, 4);
334
335
    for (i = 0; i < 4; i++) {
336
        if (s->segmentation.enabled) {
337
            base_qi = s->segmentation.base_quant[i];
338
            if (!s->segmentation.absolute_vals)
339
                base_qi += yac_qi;
340
        } else
341
            base_qi = yac_qi;
342
343
        s->qmat[i].luma_qmul[0]    =       vp8_dc_qlookup[av_clip(base_qi + ydc_delta , 0, 127)];
344
        s->qmat[i].luma_qmul[1]    =       vp8_ac_qlookup[av_clip(base_qi             , 0, 127)];
345
        s->qmat[i].luma_dc_qmul[0] =   2 * vp8_dc_qlookup[av_clip(base_qi + y2dc_delta, 0, 127)];
346 a8ab0ccc Pascal Massimino
        s->qmat[i].luma_dc_qmul[1] = 155 * vp8_ac_qlookup[av_clip(base_qi + y2ac_delta, 0, 127)] / 100;
347
        s->qmat[i].chroma_qmul[0]  =       vp8_dc_qlookup[av_clip(base_qi + uvdc_delta, 0, 127)];
348 3b636f21 David Conrad
        s->qmat[i].chroma_qmul[1]  =       vp8_ac_qlookup[av_clip(base_qi + uvac_delta, 0, 127)];
349 a8ab0ccc Pascal Massimino
350
        s->qmat[i].luma_dc_qmul[1] = FFMAX(s->qmat[i].luma_dc_qmul[1], 8);
351
        s->qmat[i].chroma_qmul[0]  = FFMIN(s->qmat[i].chroma_qmul[0], 132);
352 3b636f21 David Conrad
    }
353
}
354
355
/**
356
 * Determine which buffers golden and altref should be updated with after this frame.
357
 * The spec isn't clear here, so I'm going by my understanding of what libvpx does
358
 *
359
 * Intra frames update all 3 references
360
 * Inter frames update VP56_FRAME_PREVIOUS if the update_last flag is set
361
 * If the update (golden|altref) flag is set, it's updated with the current frame
362
 *      if update_last is set, and VP56_FRAME_PREVIOUS otherwise.
363
 * If the flag is not set, the number read means:
364
 *      0: no update
365
 *      1: VP56_FRAME_PREVIOUS
366
 *      2: update golden with altref, or update altref with golden
367
 */
368
static VP56Frame ref_to_update(VP8Context *s, int update, VP56Frame ref)
369
{
370
    VP56RangeCoder *c = &s->c;
371
372
    if (update)
373
        return VP56_FRAME_CURRENT;
374
375
    switch (vp8_rac_get_uint(c, 2)) {
376
    case 1:
377
        return VP56_FRAME_PREVIOUS;
378
    case 2:
379
        return (ref == VP56_FRAME_GOLDEN) ? VP56_FRAME_GOLDEN2 : VP56_FRAME_GOLDEN;
380
    }
381
    return VP56_FRAME_NONE;
382
}
383
384
static void update_refs(VP8Context *s)
385
{
386
    VP56RangeCoder *c = &s->c;
387
388
    int update_golden = vp8_rac_get(c);
389
    int update_altref = vp8_rac_get(c);
390
391
    s->update_golden = ref_to_update(s, update_golden, VP56_FRAME_GOLDEN);
392
    s->update_altref = ref_to_update(s, update_altref, VP56_FRAME_GOLDEN2);
393
}
394
395
static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)
396
{
397
    VP56RangeCoder *c = &s->c;
398
    int header_size, hscale, vscale, i, j, k, l, ret;
399
    int width  = s->avctx->width;
400
    int height = s->avctx->height;
401
402
    s->keyframe  = !(buf[0] & 1);
403
    s->profile   =  (buf[0]>>1) & 7;
404
    s->invisible = !(buf[0] & 0x10);
405 06d50ca8 Jason Garrett-Glaser
    header_size  = AV_RL24(buf) >> 5;
406 3b636f21 David Conrad
    buf      += 3;
407
    buf_size -= 3;
408
409 0ef1dbed David Conrad
    if (s->profile > 3)
410
        av_log(s->avctx, AV_LOG_WARNING, "Unknown profile %d\n", s->profile);
411
412
    if (!s->profile)
413
        memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_epel_pixels_tab, sizeof(s->put_pixels_tab));
414
    else    // profile 1-3 use bilinear, 4+ aren't defined so whatever
415
        memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_bilinear_pixels_tab, sizeof(s->put_pixels_tab));
416 3b636f21 David Conrad
417
    if (header_size > buf_size - 7*s->keyframe) {
418
        av_log(s->avctx, AV_LOG_ERROR, "Header size larger than data provided\n");
419
        return AVERROR_INVALIDDATA;
420
    }
421
422
    if (s->keyframe) {
423 06d50ca8 Jason Garrett-Glaser
        if (AV_RL24(buf) != 0x2a019d) {
424
            av_log(s->avctx, AV_LOG_ERROR, "Invalid start code 0x%x\n", AV_RL24(buf));
425 3b636f21 David Conrad
            return AVERROR_INVALIDDATA;
426
        }
427
        width  = AV_RL16(buf+3) & 0x3fff;
428
        height = AV_RL16(buf+5) & 0x3fff;
429
        hscale = buf[4] >> 6;
430
        vscale = buf[6] >> 6;
431
        buf      += 7;
432
        buf_size -= 7;
433
434 92a54426 Måns Rullgård
        if (hscale || vscale)
435
            av_log_missing_feature(s->avctx, "Upscaling", 1);
436
437 3b636f21 David Conrad
        s->update_golden = s->update_altref = VP56_FRAME_CURRENT;
438
        memcpy(s->prob->token    , vp8_token_default_probs , sizeof(s->prob->token));
439
        memcpy(s->prob->pred16x16, vp8_pred16x16_prob_inter, sizeof(s->prob->pred16x16));
440
        memcpy(s->prob->pred8x8c , vp8_pred8x8c_prob_inter , sizeof(s->prob->pred8x8c));
441
        memcpy(s->prob->mvc      , vp8_mv_default_prob     , sizeof(s->prob->mvc));
442
        memset(&s->segmentation, 0, sizeof(s->segmentation));
443
    }
444
445
    if (!s->macroblocks_base || /* first frame */
446
        width != s->avctx->width || height != s->avctx->height) {
447
        if ((ret = update_dimensions(s, width, height) < 0))
448
            return ret;
449
    }
450
451
    vp56_init_range_decoder(c, buf, header_size);
452
    buf      += header_size;
453
    buf_size -= header_size;
454
455
    if (s->keyframe) {
456
        if (vp8_rac_get(c))
457
            av_log(s->avctx, AV_LOG_WARNING, "Unspecified colorspace\n");
458
        vp8_rac_get(c); // whether we can skip clamping in dsp functions
459
    }
460
461
    if ((s->segmentation.enabled = vp8_rac_get(c)))
462
        parse_segment_info(s);
463
    else
464
        s->segmentation.update_map = 0; // FIXME: move this to some init function?
465
466
    s->filter.simple    = vp8_rac_get(c);
467
    s->filter.level     = vp8_rac_get_uint(c, 6);
468
    s->filter.sharpness = vp8_rac_get_uint(c, 3);
469
470
    if ((s->lf_delta.enabled = vp8_rac_get(c)))
471
        if (vp8_rac_get(c))
472
            update_lf_deltas(s);
473
474
    if (setup_partitions(s, buf, buf_size)) {
475
        av_log(s->avctx, AV_LOG_ERROR, "Invalid partitions\n");
476
        return AVERROR_INVALIDDATA;
477
    }
478
479
    get_quants(s);
480
481
    if (!s->keyframe) {
482
        update_refs(s);
483
        s->sign_bias[VP56_FRAME_GOLDEN]               = vp8_rac_get(c);
484
        s->sign_bias[VP56_FRAME_GOLDEN2 /* altref */] = vp8_rac_get(c);
485
    }
486
487
    // if we aren't saving this frame's probabilities for future frames,
488
    // make a copy of the current probabilities
489
    if (!(s->update_probabilities = vp8_rac_get(c)))
490
        s->prob[1] = s->prob[0];
491
492
    s->update_last = s->keyframe || vp8_rac_get(c);
493
494
    for (i = 0; i < 4; i++)
495
        for (j = 0; j < 8; j++)
496
            for (k = 0; k < 3; k++)
497
                for (l = 0; l < NUM_DCT_TOKENS-1; l++)
498 7697cdcf David Conrad
                    if (vp56_rac_get_prob_branchy(c, vp8_token_update_probs[i][j][k][l]))
499 3b636f21 David Conrad
                        s->prob->token[i][j][k][l] = vp8_rac_get_uint(c, 8);
500
501
    if ((s->mbskip_enabled = vp8_rac_get(c)))
502 a8ab0ccc Pascal Massimino
        s->prob->mbskip = vp8_rac_get_uint(c, 8);
503 3b636f21 David Conrad
504
    if (!s->keyframe) {
505 a8ab0ccc Pascal Massimino
        s->prob->intra  = vp8_rac_get_uint(c, 8);
506
        s->prob->last   = vp8_rac_get_uint(c, 8);
507
        s->prob->golden = vp8_rac_get_uint(c, 8);
508 3b636f21 David Conrad
509
        if (vp8_rac_get(c))
510
            for (i = 0; i < 4; i++)
511
                s->prob->pred16x16[i] = vp8_rac_get_uint(c, 8);
512
        if (vp8_rac_get(c))
513
            for (i = 0; i < 3; i++)
514
                s->prob->pred8x8c[i]  = vp8_rac_get_uint(c, 8);
515
516
        // 17.2 MV probability update
517
        for (i = 0; i < 2; i++)
518
            for (j = 0; j < 19; j++)
519 7697cdcf David Conrad
                if (vp56_rac_get_prob_branchy(c, vp8_mv_update_prob[i][j]))
520 3b636f21 David Conrad
                    s->prob->mvc[i][j] = vp8_rac_get_nn(c);
521
    }
522
523
    return 0;
524
}
525
526 414ac27d Jason Garrett-Glaser
static av_always_inline
527
void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src, int mb_x, int mb_y)
528 3b636f21 David Conrad
{
529
#define MARGIN (16 << 2)
530
    dst->x = av_clip(src->x, -((mb_x << 6) + MARGIN),
531
                     ((s->mb_width  - 1 - mb_x) << 6) + MARGIN);
532
    dst->y = av_clip(src->y, -((mb_y << 6) + MARGIN),
533
                     ((s->mb_height - 1 - mb_y) << 6) + MARGIN);
534
}
535
536 414ac27d Jason Garrett-Glaser
static av_always_inline
537
void find_near_mvs(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
538
                   VP56mv near[2], VP56mv *best, uint8_t cnt[4])
539 3b636f21 David Conrad
{
540 c55e0d34 Jason Garrett-Glaser
    VP8Macroblock *mb_edge[3] = { mb + 2 /* top */,
541
                                  mb - 1 /* left */,
542
                                  mb + 1 /* top-left */ };
543 3b636f21 David Conrad
    enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT };
544
    VP56mv near_mv[4]  = {{ 0 }};
545
    enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV };
546 702e8d33 Jason Garrett-Glaser
    int idx = CNT_ZERO;
547 3b636f21 David Conrad
    int best_idx = CNT_ZERO;
548 702e8d33 Jason Garrett-Glaser
    int cur_sign_bias = s->sign_bias[mb->ref_frame];
549
    int *sign_bias = s->sign_bias;
550 3b636f21 David Conrad
551
    /* Process MB on top, left and top-left */
552 702e8d33 Jason Garrett-Glaser
    #define MV_EDGE_CHECK(n)\
553
    {\
554
        VP8Macroblock *edge = mb_edge[n];\
555
        int edge_ref = edge->ref_frame;\
556
        if (edge_ref != VP56_FRAME_CURRENT) {\
557
            uint32_t mv = AV_RN32A(&edge->mv);\
558
            if (mv) {\
559
                if (cur_sign_bias != sign_bias[edge_ref]) {\
560
                    /* SWAR negate of the values in mv. */\
561 0087aa47 Jason Garrett-Glaser
                    mv = ~mv;\
562
                    mv = ((mv&0x7fff7fff) + 0x00010001) ^ (mv&0x80008000);\
563 702e8d33 Jason Garrett-Glaser
                }\
564
                if (!n || mv != AV_RN32A(&near_mv[idx]))\
565
                    AV_WN32A(&near_mv[++idx], mv);\
566
                cnt[idx]      += 1 + (n != 2);\
567
            } else\
568
                cnt[CNT_ZERO] += 1 + (n != 2);\
569
        }\
570 3b636f21 David Conrad
    }
571 702e8d33 Jason Garrett-Glaser
    MV_EDGE_CHECK(0)
572
    MV_EDGE_CHECK(1)
573
    MV_EDGE_CHECK(2)
574 3b636f21 David Conrad
575 702e8d33 Jason Garrett-Glaser
    /* If we have three distinct MVs, merge first and last if they're the same */
576
    if (cnt[CNT_SPLITMV] && AV_RN32A(&near_mv[1+EDGE_TOP]) == AV_RN32A(&near_mv[1+EDGE_TOPLEFT]))
577 3b636f21 David Conrad
        cnt[CNT_NEAREST] += 1;
578
579
    cnt[CNT_SPLITMV] = ((mb_edge[EDGE_LEFT]->mode   == VP8_MVMODE_SPLIT) +
580
                        (mb_edge[EDGE_TOP]->mode    == VP8_MVMODE_SPLIT)) * 2 +
581
                       (mb_edge[EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT);
582
583
    /* Swap near and nearest if necessary */
584
    if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) {
585 702e8d33 Jason Garrett-Glaser
        FFSWAP(uint8_t,     cnt[CNT_NEAREST],     cnt[CNT_NEAR]);
586
        FFSWAP( VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]);
587 3b636f21 David Conrad
    }
588
589
    /* Choose the best mv out of 0,0 and the nearest mv */
590
    if (cnt[CNT_NEAREST] >= cnt[CNT_ZERO])
591
        best_idx = CNT_NEAREST;
592
593 9fddd14a Jason Garrett-Glaser
    mb->mv  = near_mv[best_idx];
594 3b636f21 David Conrad
    near[0] = near_mv[CNT_NEAREST];
595
    near[1] = near_mv[CNT_NEAR];
596
}
597
598
/**
599
 * Motion vector coding, 17.1.
600
 */
601
static int read_mv_component(VP56RangeCoder *c, const uint8_t *p)
602
{
603 ca18a478 David Conrad
    int bit, x = 0;
604 3b636f21 David Conrad
605 7697cdcf David Conrad
    if (vp56_rac_get_prob_branchy(c, p[0])) {
606 3b636f21 David Conrad
        int i;
607
608
        for (i = 0; i < 3; i++)
609
            x += vp56_rac_get_prob(c, p[9 + i]) << i;
610
        for (i = 9; i > 3; i--)
611
            x += vp56_rac_get_prob(c, p[9 + i]) << i;
612
        if (!(x & 0xFFF0) || vp56_rac_get_prob(c, p[12]))
613
            x += 8;
614 ca18a478 David Conrad
    } else {
615
        // small_mvtree
616
        const uint8_t *ps = p+2;
617
        bit = vp56_rac_get_prob(c, *ps);
618
        ps += 1 + 3*bit;
619
        x  += 4*bit;
620
        bit = vp56_rac_get_prob(c, *ps);
621
        ps += 1 + bit;
622
        x  += 2*bit;
623
        x  += vp56_rac_get_prob(c, *ps);
624
    }
625 3b636f21 David Conrad
626
    return (x && vp56_rac_get_prob(c, p[1])) ? -x : x;
627
}
628
629 414ac27d Jason Garrett-Glaser
static av_always_inline
630
const uint8_t *get_submv_prob(uint32_t left, uint32_t top)
631 3b636f21 David Conrad
{
632 7bf254c4 Jason Garrett-Glaser
    if (left == top)
633
        return vp8_submv_prob[4-!!left];
634
    if (!top)
635 3b636f21 David Conrad
        return vp8_submv_prob[2];
636 7bf254c4 Jason Garrett-Glaser
    return vp8_submv_prob[1-!!left];
637 3b636f21 David Conrad
}
638
639
/**
640
 * Split motion vector prediction, 16.4.
641 7ed06b2b Ronald S. Bultje
 * @returns the number of motion vectors parsed (2, 4 or 16)
642 3b636f21 David Conrad
 */
643 414ac27d Jason Garrett-Glaser
static av_always_inline
644
int decode_splitmvs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb)
645 3b636f21 David Conrad
{
646
    int part_idx = mb->partitioning =
647
        vp8_rac_get_tree(c, vp8_mbsplit_tree, vp8_mbsplit_prob);
648
    int n, num = vp8_mbsplit_count[part_idx];
649 c55e0d34 Jason Garrett-Glaser
    VP8Macroblock *top_mb  = &mb[2];
650 7bf254c4 Jason Garrett-Glaser
    VP8Macroblock *left_mb = &mb[-1];
651
    const uint8_t *mbsplits_left = vp8_mbsplits[left_mb->partitioning],
652
                  *mbsplits_top = vp8_mbsplits[top_mb->partitioning],
653
                  *mbsplits_cur = vp8_mbsplits[part_idx],
654 7ed06b2b Ronald S. Bultje
                  *firstidx = vp8_mbfirstidx[part_idx];
655 c55e0d34 Jason Garrett-Glaser
    VP56mv *top_mv  = top_mb->bmv;
656
    VP56mv *left_mv = left_mb->bmv;
657
    VP56mv *cur_mv  = mb->bmv;
658 3b636f21 David Conrad
659
    for (n = 0; n < num; n++) {
660 7ed06b2b Ronald S. Bultje
        int k = firstidx[n];
661 7bf254c4 Jason Garrett-Glaser
        uint32_t left, above;
662 7ed06b2b Ronald S. Bultje
        const uint8_t *submv_prob;
663
664 7bf254c4 Jason Garrett-Glaser
        if (!(k & 3))
665
            left = AV_RN32A(&left_mv[mbsplits_left[k + 3]]);
666
        else
667
            left  = AV_RN32A(&cur_mv[mbsplits_cur[k - 1]]);
668
        if (k <= 3)
669
            above = AV_RN32A(&top_mv[mbsplits_top[k + 12]]);
670
        else
671
            above = AV_RN32A(&cur_mv[mbsplits_cur[k - 4]]);
672 7ed06b2b Ronald S. Bultje
673
        submv_prob = get_submv_prob(left, above);
674 3b636f21 David Conrad
675
        switch (vp8_rac_get_tree(c, vp8_submv_ref_tree, submv_prob)) {
676
        case VP8_SUBMVMODE_NEW4X4:
677 9fddd14a Jason Garrett-Glaser
            mb->bmv[n].y = mb->mv.y + read_mv_component(c, s->prob->mvc[0]);
678
            mb->bmv[n].x = mb->mv.x + read_mv_component(c, s->prob->mvc[1]);
679 3b636f21 David Conrad
            break;
680
        case VP8_SUBMVMODE_ZERO4X4:
681 14767f35 Jason Garrett-Glaser
            AV_ZERO32(&mb->bmv[n]);
682 3b636f21 David Conrad
            break;
683
        case VP8_SUBMVMODE_LEFT4X4:
684 7bf254c4 Jason Garrett-Glaser
            AV_WN32A(&mb->bmv[n], left);
685 3b636f21 David Conrad
            break;
686
        case VP8_SUBMVMODE_TOP4X4:
687 7bf254c4 Jason Garrett-Glaser
            AV_WN32A(&mb->bmv[n], above);
688 3b636f21 David Conrad
            break;
689
        }
690
    }
691 7ed06b2b Ronald S. Bultje
692
    return num;
693 3b636f21 David Conrad
}
694
695 414ac27d Jason Garrett-Glaser
static av_always_inline
696
void decode_intra4x4_modes(VP56RangeCoder *c, uint8_t *intra4x4,
697
                           int stride, int keyframe)
698 3b636f21 David Conrad
{
699 d1c58fce Jason Garrett-Glaser
    int x, y, t, l, i;
700 3b636f21 David Conrad
701 d1c58fce Jason Garrett-Glaser
    if (keyframe) {
702
        const uint8_t *ctx;
703
        for (y = 0; y < 4; y++) {
704
            for (x = 0; x < 4; x++) {
705 3b636f21 David Conrad
                t = intra4x4[x - stride];
706
                l = intra4x4[x - 1];
707
                ctx = vp8_pred4x4_prob_intra[t][l];
708 d1c58fce Jason Garrett-Glaser
                intra4x4[x] = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx);
709 3b636f21 David Conrad
            }
710 d1c58fce Jason Garrett-Glaser
            intra4x4 += stride;
711 3b636f21 David Conrad
        }
712 d1c58fce Jason Garrett-Glaser
    } else {
713
        for (i = 0; i < 16; i++)
714
            intra4x4[i] = vp8_rac_get_tree(c, vp8_pred4x4_tree, vp8_pred4x4_prob_inter);
715 3b636f21 David Conrad
    }
716
}
717
718 414ac27d Jason Garrett-Glaser
static av_always_inline
719
void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
720
                    uint8_t *intra4x4, uint8_t *segment)
721 3b636f21 David Conrad
{
722
    VP56RangeCoder *c = &s->c;
723
724
    if (s->segmentation.update_map)
725 c55e0d34 Jason Garrett-Glaser
        *segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid);
726 b9a7186b Jason Garrett-Glaser
    s->segment = *segment;
727 3b636f21 David Conrad
728 a8ab0ccc Pascal Massimino
    mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0;
729 3b636f21 David Conrad
730
    if (s->keyframe) {
731
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra);
732
733
        if (mb->mode == MODE_I4x4) {
734
            decode_intra4x4_modes(c, intra4x4, s->b4_stride, 1);
735
        } else
736
            fill_rectangle(intra4x4, 4, 4, s->b4_stride, vp8_pred4x4_mode[mb->mode], 1);
737
738
        s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra);
739
        mb->ref_frame = VP56_FRAME_CURRENT;
740 a8ab0ccc Pascal Massimino
    } else if (vp56_rac_get_prob_branchy(c, s->prob->intra)) {
741 3b636f21 David Conrad
        VP56mv near[2], best;
742 702e8d33 Jason Garrett-Glaser
        uint8_t cnt[4] = { 0 };
743 3b636f21 David Conrad
        uint8_t p[4];
744
745
        // inter MB, 16.2
746 a8ab0ccc Pascal Massimino
        if (vp56_rac_get_prob_branchy(c, s->prob->last))
747
            mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ?
748 3b636f21 David Conrad
                VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN;
749
        else
750
            mb->ref_frame = VP56_FRAME_PREVIOUS;
751 c4211046 Jason Garrett-Glaser
        s->ref_count[mb->ref_frame-1]++;
752 3b636f21 David Conrad
753
        // motion vectors, 16.3
754
        find_near_mvs(s, mb, mb_x, mb_y, near, &best, cnt);
755 702e8d33 Jason Garrett-Glaser
        p[0] = vp8_mode_contexts[cnt[0]][0];
756
        p[1] = vp8_mode_contexts[cnt[1]][1];
757
        p[2] = vp8_mode_contexts[cnt[2]][2];
758
        p[3] = vp8_mode_contexts[cnt[3]][3];
759 3b636f21 David Conrad
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_mvinter, p);
760
        switch (mb->mode) {
761
        case VP8_MVMODE_SPLIT:
762 9fddd14a Jason Garrett-Glaser
            clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y);
763
            mb->mv = mb->bmv[decode_splitmvs(s, c, mb) - 1];
764 3b636f21 David Conrad
            break;
765
        case VP8_MVMODE_ZERO:
766 14767f35 Jason Garrett-Glaser
            AV_ZERO32(&mb->mv);
767 3b636f21 David Conrad
            break;
768
        case VP8_MVMODE_NEAREST:
769
            clamp_mv(s, &mb->mv, &near[0], mb_x, mb_y);
770
            break;
771
        case VP8_MVMODE_NEAR:
772
            clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y);
773
            break;
774
        case VP8_MVMODE_NEW:
775 9fddd14a Jason Garrett-Glaser
            clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y);
776
            mb->mv.y += + read_mv_component(c, s->prob->mvc[0]);
777
            mb->mv.x += + read_mv_component(c, s->prob->mvc[1]);
778 3b636f21 David Conrad
            break;
779
        }
780
        if (mb->mode != VP8_MVMODE_SPLIT) {
781 7ed06b2b Ronald S. Bultje
            mb->partitioning = VP8_SPLITMVMODE_NONE;
782
            mb->bmv[0] = mb->mv;
783 3b636f21 David Conrad
        }
784
    } else {
785
        // intra MB, 16.1
786
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16);
787
788 158e062c Jason Garrett-Glaser
        if (mb->mode == MODE_I4x4)
789 d1c58fce Jason Garrett-Glaser
            decode_intra4x4_modes(c, intra4x4, 4, 0);
790 3b636f21 David Conrad
791
        s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c);
792
        mb->ref_frame = VP56_FRAME_CURRENT;
793 b946111f Jason Garrett-Glaser
        mb->partitioning = VP8_SPLITMVMODE_NONE;
794 14767f35 Jason Garrett-Glaser
        AV_ZERO32(&mb->bmv[0]);
795 3b636f21 David Conrad
    }
796
}
797
798
/**
799 e394953e Ronald S. Bultje
 * @param c arithmetic bitstream reader context
800
 * @param block destination for block coefficients
801
 * @param probs probabilities to use when reading trees from the bitstream
802 3b636f21 David Conrad
 * @param i initial coeff index, 0 unless a separate DC block is coded
803
 * @param zero_nhood the initial prediction context for number of surrounding
804
 *                   all-zero blocks (only left/top, so 0-2)
805 3fa76268 Diego Biurrun
 * @param qmul array holding the dc/ac dequant factor at position 0/1
806 3b636f21 David Conrad
 * @return 0 if no coeffs were decoded
807
 *         otherwise, the index of the last coeff decoded plus one
808
 */
809
static int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16],
810
                               uint8_t probs[8][3][NUM_DCT_TOKENS-1],
811
                               int i, int zero_nhood, int16_t qmul[2])
812
{
813 10bf2eeb Jason Garrett-Glaser
    uint8_t *token_prob = probs[vp8_coeff_band[i]][zero_nhood];
814 fe1b5d97 David Conrad
    int nonzero = 0;
815
    int coeff;
816 3b636f21 David Conrad
817 fe1b5d97 David Conrad
    do {
818
        if (!vp56_rac_get_prob_branchy(c, token_prob[0]))   // DCT_EOB
819
            return nonzero;
820 3b636f21 David Conrad
821 fe1b5d97 David Conrad
skip_eob:
822
        if (!vp56_rac_get_prob_branchy(c, token_prob[1])) { // DCT_0
823 c22b4468 Pascal Massimino
            if (++i == 16)
824
                return nonzero; // invalid input; blocks should end with EOB
825
            token_prob = probs[vp8_coeff_band[i]][0];
826
            goto skip_eob;
827 fe1b5d97 David Conrad
        }
828
829
        if (!vp56_rac_get_prob_branchy(c, token_prob[2])) { // DCT_1
830
            coeff = 1;
831 10bf2eeb Jason Garrett-Glaser
            token_prob = probs[vp8_coeff_band[i+1]][1];
832 fe1b5d97 David Conrad
        } else {
833
            if (!vp56_rac_get_prob_branchy(c, token_prob[3])) { // DCT 2,3,4
834
                coeff = vp56_rac_get_prob(c, token_prob[4]);
835
                if (coeff)
836
                    coeff += vp56_rac_get_prob(c, token_prob[5]);
837
                coeff += 2;
838
            } else {
839
                // DCT_CAT*
840
                if (!vp56_rac_get_prob_branchy(c, token_prob[6])) {
841
                    if (!vp56_rac_get_prob_branchy(c, token_prob[7])) { // DCT_CAT1
842
                        coeff  = 5 + vp56_rac_get_prob(c, vp8_dct_cat1_prob[0]);
843
                    } else {                                    // DCT_CAT2
844
                        coeff  = 7;
845
                        coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[0]) << 1;
846
                        coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[1]);
847
                    }
848
                } else {    // DCT_CAT3 and up
849
                    int a = vp56_rac_get_prob(c, token_prob[8]);
850
                    int b = vp56_rac_get_prob(c, token_prob[9+a]);
851
                    int cat = (a<<1) + b;
852
                    coeff  = 3 + (8<<cat);
853
                    coeff += vp8_rac_get_coeff(c, vp8_dct_cat_prob[cat]);
854
                }
855
            }
856 10bf2eeb Jason Garrett-Glaser
            token_prob = probs[vp8_coeff_band[i+1]][2];
857 fe1b5d97 David Conrad
        }
858
859 3b636f21 David Conrad
        // todo: full [16] qmat? load into register?
860 fe1b5d97 David Conrad
        block[zigzag_scan[i]] = (vp8_rac_get(c) ? -coeff : coeff) * qmul[!!i];
861
        nonzero = ++i;
862
    } while (i < 16);
863
864 3b636f21 David Conrad
    return nonzero;
865
}
866
867 414ac27d Jason Garrett-Glaser
static av_always_inline
868
void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,
869
                      uint8_t t_nnz[9], uint8_t l_nnz[9])
870 3b636f21 David Conrad
{
871
    LOCAL_ALIGNED_16(DCTELEM, dc,[16]);
872
    int i, x, y, luma_start = 0, luma_ctx = 3;
873
    int nnz_pred, nnz, nnz_total = 0;
874 b9a7186b Jason Garrett-Glaser
    int segment = s->segment;
875 3b636f21 David Conrad
876
    if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
877
        AV_ZERO128(dc);
878
        AV_ZERO128(dc+8);
879
        nnz_pred = t_nnz[8] + l_nnz[8];
880
881
        // decode DC values and do hadamard
882
        nnz = decode_block_coeffs(c, dc, s->prob->token[1], 0, nnz_pred,
883
                                  s->qmat[segment].luma_dc_qmul);
884
        l_nnz[8] = t_nnz[8] = !!nnz;
885
        nnz_total += nnz;
886
        s->vp8dsp.vp8_luma_dc_wht(s->block, dc);
887
        luma_start = 1;
888
        luma_ctx = 0;
889
    }
890
891
    // luma blocks
892
    for (y = 0; y < 4; y++)
893
        for (x = 0; x < 4; x++) {
894
            nnz_pred = l_nnz[y] + t_nnz[x];
895
            nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start,
896
                                      nnz_pred, s->qmat[segment].luma_qmul);
897
            // nnz+luma_start may be one more than the actual last index, but we don't care
898
            s->non_zero_count_cache[y][x] = nnz + luma_start;
899
            t_nnz[x] = l_nnz[y] = !!nnz;
900
            nnz_total += nnz;
901
        }
902
903
    // chroma blocks
904
    // TODO: what to do about dimensions? 2nd dim for luma is x,
905
    // but for chroma it's (y<<1)|x
906
    for (i = 4; i < 6; i++)
907
        for (y = 0; y < 2; y++)
908
            for (x = 0; x < 2; x++) {
909
                nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x];
910
                nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0,
911
                                          nnz_pred, s->qmat[segment].chroma_qmul);
912
                s->non_zero_count_cache[i][(y<<1)+x] = nnz;
913
                t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz;
914
                nnz_total += nnz;
915
            }
916
917
    // if there were no coded coeffs despite the macroblock not being marked skip,
918
    // we MUST not do the inner loop filter and should not do IDCT
919
    // Since skip isn't used for bitstream prediction, just manually set it.
920
    if (!nnz_total)
921
        mb->skip = 1;
922
}
923
924 9ac831c2 David Conrad
static av_always_inline
925
void backup_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
926
                      int linesize, int uvlinesize, int simple)
927
{
928
    AV_COPY128(top_border, src_y + 15*linesize);
929
    if (!simple) {
930
        AV_COPY64(top_border+16, src_cb + 7*uvlinesize);
931
        AV_COPY64(top_border+24, src_cr + 7*uvlinesize);
932
    }
933
}
934
935
static av_always_inline
936
void xchg_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
937
                    int linesize, int uvlinesize, int mb_x, int mb_y, int mb_width,
938
                    int simple, int xchg)
939
{
940
    uint8_t *top_border_m1 = top_border-32;     // for TL prediction
941
    src_y  -=   linesize;
942
    src_cb -= uvlinesize;
943
    src_cr -= uvlinesize;
944
945 096971e8 Måns Rullgård
#define XCHG(a,b,xchg) do {                     \
946
        if (xchg) AV_SWAP64(b,a);               \
947
        else      AV_COPY64(b,a);               \
948
    } while (0)
949 9ac831c2 David Conrad
950
    XCHG(top_border_m1+8, src_y-8, xchg);
951
    XCHG(top_border,      src_y,   xchg);
952
    XCHG(top_border+8,    src_y+8, 1);
953 070ce7ef Måns Rullgård
    if (mb_x < mb_width-1)
954 9ac831c2 David Conrad
        XCHG(top_border+32, src_y+16, 1);
955 070ce7ef Måns Rullgård
956 9ac831c2 David Conrad
    // only copy chroma for normal loop filter
957
    // or to initialize the top row to 127
958
    if (!simple || !mb_y) {
959
        XCHG(top_border_m1+16, src_cb-8, xchg);
960
        XCHG(top_border_m1+24, src_cr-8, xchg);
961
        XCHG(top_border+16,    src_cb, 1);
962
        XCHG(top_border+24,    src_cr, 1);
963
    }
964
}
965
966 414ac27d Jason Garrett-Glaser
static av_always_inline
967
int check_intra_pred_mode(int mode, int mb_x, int mb_y)
968 3b636f21 David Conrad
{
969
    if (mode == DC_PRED8x8) {
970 a71abb71 Jason Garrett-Glaser
        if (!mb_x) {
971
            mode = mb_y ? TOP_DC_PRED8x8 : DC_128_PRED8x8;
972
        } else if (!mb_y) {
973 09959ec4 Jason Garrett-Glaser
            mode = LEFT_DC_PRED8x8;
974 a71abb71 Jason Garrett-Glaser
        }
975 3b636f21 David Conrad
    }
976
    return mode;
977
}
978
979 414ac27d Jason Garrett-Glaser
static av_always_inline
980
void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
981
                   uint8_t *intra4x4, int mb_x, int mb_y)
982 3b636f21 David Conrad
{
983
    int x, y, mode, nnz, tr;
984
985 9ac831c2 David Conrad
    // for the first row, we need to run xchg_mb_border to init the top edge to 127
986
    // otherwise, skip it if we aren't going to deblock
987
    if (s->deblock_filter || !mb_y)
988
        xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
989
                       s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
990
                       s->filter.simple, 1);
991
992 3b636f21 David Conrad
    if (mb->mode < MODE_I4x4) {
993
        mode = check_intra_pred_mode(mb->mode, mb_x, mb_y);
994
        s->hpc.pred16x16[mode](dst[0], s->linesize);
995
    } else {
996
        uint8_t *ptr = dst[0];
997 d1c58fce Jason Garrett-Glaser
        int stride = s->keyframe ? s->b4_stride : 4;
998 3b636f21 David Conrad
999
        // all blocks on the right edge of the macroblock use bottom edge
1000
        // the top macroblock for their topright edge
1001
        uint8_t *tr_right = ptr - s->linesize + 16;
1002
1003
        // if we're on the right edge of the frame, said edge is extended
1004
        // from the top macroblock
1005
        if (mb_x == s->mb_width-1) {
1006
            tr = tr_right[-1]*0x01010101;
1007
            tr_right = (uint8_t *)&tr;
1008
        }
1009
1010 b74f70d6 Jason Garrett-Glaser
        if (mb->skip)
1011
            AV_ZERO128(s->non_zero_count_cache);
1012
1013 3b636f21 David Conrad
        for (y = 0; y < 4; y++) {
1014
            uint8_t *topright = ptr + 4 - s->linesize;
1015
            for (x = 0; x < 4; x++) {
1016
                if (x == 3)
1017
                    topright = tr_right;
1018
1019 d1c58fce Jason Garrett-Glaser
                s->hpc.pred4x4[intra4x4[x]](ptr+4*x, topright, s->linesize);
1020 3b636f21 David Conrad
1021
                nnz = s->non_zero_count_cache[y][x];
1022
                if (nnz) {
1023
                    if (nnz == 1)
1024
                        s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize);
1025
                    else
1026
                        s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize);
1027
                }
1028
                topright += 4;
1029
            }
1030
1031
            ptr   += 4*s->linesize;
1032 d1c58fce Jason Garrett-Glaser
            intra4x4 += stride;
1033 3b636f21 David Conrad
        }
1034
    }
1035
1036
    mode = check_intra_pred_mode(s->chroma_pred_mode, mb_x, mb_y);
1037
    s->hpc.pred8x8[mode](dst[1], s->uvlinesize);
1038
    s->hpc.pred8x8[mode](dst[2], s->uvlinesize);
1039 9ac831c2 David Conrad
1040
    if (s->deblock_filter || !mb_y)
1041
        xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
1042
                       s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
1043
                       s->filter.simple, 0);
1044 3b636f21 David Conrad
}
1045
1046
/**
1047
 * Generic MC function.
1048
 *
1049
 * @param s VP8 decoding context
1050
 * @param luma 1 for luma (Y) planes, 0 for chroma (Cb/Cr) planes
1051
 * @param dst target buffer for block data at block position
1052
 * @param src reference picture buffer at origin (0, 0)
1053
 * @param mv motion vector (relative to block position) to get pixel data from
1054
 * @param x_off horizontal position of block from origin (0, 0)
1055
 * @param y_off vertical position of block from origin (0, 0)
1056
 * @param block_w width of block (16, 8 or 4)
1057
 * @param block_h height of block (always same as block_w)
1058
 * @param width width of src/dst plane data
1059
 * @param height height of src/dst plane data
1060
 * @param linesize size of a single line of plane data, including padding
1061 e394953e Ronald S. Bultje
 * @param mc_func motion compensation function pointers (bilinear or sixtap MC)
1062 3b636f21 David Conrad
 */
1063 414ac27d Jason Garrett-Glaser
static av_always_inline
1064
void vp8_mc(VP8Context *s, int luma,
1065
            uint8_t *dst, uint8_t *src, const VP56mv *mv,
1066
            int x_off, int y_off, int block_w, int block_h,
1067
            int width, int height, int linesize,
1068
            vp8_mc_func mc_func[3][3])
1069 3b636f21 David Conrad
{
1070 c0498b30 Jason Garrett-Glaser
    if (AV_RN32A(mv)) {
1071
        static const uint8_t idx[8] = { 0, 1, 2, 1, 2, 1, 2, 1 };
1072
        int mx = (mv->x << luma)&7, mx_idx = idx[mx];
1073
        int my = (mv->y << luma)&7, my_idx = idx[my];
1074
1075
        x_off += mv->x >> (3 - luma);
1076
        y_off += mv->y >> (3 - luma);
1077
1078
        // edge emulation
1079
        src += y_off * linesize + x_off;
1080
        if (x_off < 2 || x_off >= width  - block_w - 3 ||
1081
            y_off < 2 || y_off >= height - block_h - 3) {
1082
            ff_emulated_edge_mc(s->edge_emu_buffer, src - 2 * linesize - 2, linesize,
1083
                                block_w + 5, block_h + 5,
1084
                                x_off - 2, y_off - 2, width, height);
1085
            src = s->edge_emu_buffer + 2 + linesize * 2;
1086
        }
1087
        mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);
1088
    } else
1089
        mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0);
1090 3b636f21 David Conrad
}
1091
1092 414ac27d Jason Garrett-Glaser
static av_always_inline
1093
void vp8_mc_part(VP8Context *s, uint8_t *dst[3],
1094
                 AVFrame *ref_frame, int x_off, int y_off,
1095
                 int bx_off, int by_off,
1096
                 int block_w, int block_h,
1097
                 int width, int height, VP56mv *mv)
1098 7c4dcf81 Ronald S. Bultje
{
1099
    VP56mv uvmv = *mv;
1100
1101
    /* Y */
1102
    vp8_mc(s, 1, dst[0] + by_off * s->linesize + bx_off,
1103
           ref_frame->data[0], mv, x_off + bx_off, y_off + by_off,
1104
           block_w, block_h, width, height, s->linesize,
1105
           s->put_pixels_tab[block_w == 8]);
1106
1107
    /* U/V */
1108
    if (s->profile == 3) {
1109
        uvmv.x &= ~7;
1110
        uvmv.y &= ~7;
1111
    }
1112
    x_off   >>= 1; y_off   >>= 1;
1113
    bx_off  >>= 1; by_off  >>= 1;
1114
    width   >>= 1; height  >>= 1;
1115
    block_w >>= 1; block_h >>= 1;
1116
    vp8_mc(s, 0, dst[1] + by_off * s->uvlinesize + bx_off,
1117
           ref_frame->data[1], &uvmv, x_off + bx_off, y_off + by_off,
1118
           block_w, block_h, width, height, s->uvlinesize,
1119
           s->put_pixels_tab[1 + (block_w == 4)]);
1120
    vp8_mc(s, 0, dst[2] + by_off * s->uvlinesize + bx_off,
1121
           ref_frame->data[2], &uvmv, x_off + bx_off, y_off + by_off,
1122
           block_w, block_h, width, height, s->uvlinesize,
1123
           s->put_pixels_tab[1 + (block_w == 4)]);
1124
}
1125
1126 d864dee8 Jason Garrett-Glaser
/* Fetch pixels for estimated mv 4 macroblocks ahead.
1127
 * Optimized for 64-byte cache lines.  Inspired by ffh264 prefetch_motion. */
1128 414ac27d Jason Garrett-Glaser
static av_always_inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int mb_xy, int ref)
1129 d864dee8 Jason Garrett-Glaser
{
1130 ef38842f Jason Garrett-Glaser
    /* Don't prefetch refs that haven't been used very often this frame. */
1131
    if (s->ref_count[ref-1] > (mb_xy >> 5)) {
1132 c4211046 Jason Garrett-Glaser
        int x_off = mb_x << 4, y_off = mb_y << 4;
1133
        int mx = mb->mv.x + x_off + 8;
1134
        int my = mb->mv.y + y_off;
1135
        uint8_t **src= s->framep[ref]->data;
1136
        int off= mx + (my + (mb_x&3)*4)*s->linesize + 64;
1137
        s->dsp.prefetch(src[0]+off, s->linesize, 4);
1138
        off= (mx>>1) + ((my>>1) + (mb_x&7))*s->uvlinesize + 64;
1139
        s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1140
    }
1141 d864dee8 Jason Garrett-Glaser
}
1142
1143 3b636f21 David Conrad
/**
1144
 * Apply motion vectors to prediction buffer, chapter 18.
1145
 */
1146 414ac27d Jason Garrett-Glaser
static av_always_inline
1147
void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
1148
                   int mb_x, int mb_y)
1149 3b636f21 David Conrad
{
1150
    int x_off = mb_x << 4, y_off = mb_y << 4;
1151
    int width = 16*s->mb_width, height = 16*s->mb_height;
1152 d292c345 Jason Garrett-Glaser
    AVFrame *ref = s->framep[mb->ref_frame];
1153
    VP56mv *bmv = mb->bmv;
1154 3b636f21 David Conrad
1155
    if (mb->mode < VP8_MVMODE_SPLIT) {
1156 d292c345 Jason Garrett-Glaser
        vp8_mc_part(s, dst, ref, x_off, y_off,
1157 7c4dcf81 Ronald S. Bultje
                    0, 0, 16, 16, width, height, &mb->mv);
1158
    } else switch (mb->partitioning) {
1159
    case VP8_SPLITMVMODE_4x4: {
1160 3b636f21 David Conrad
        int x, y;
1161 7c4dcf81 Ronald S. Bultje
        VP56mv uvmv;
1162 3b636f21 David Conrad
1163
        /* Y */
1164
        for (y = 0; y < 4; y++) {
1165
            for (x = 0; x < 4; x++) {
1166
                vp8_mc(s, 1, dst[0] + 4*y*s->linesize + x*4,
1167 d292c345 Jason Garrett-Glaser
                       ref->data[0], &bmv[4*y + x],
1168 3b636f21 David Conrad
                       4*x + x_off, 4*y + y_off, 4, 4,
1169
                       width, height, s->linesize,
1170 0ef1dbed David Conrad
                       s->put_pixels_tab[2]);
1171 3b636f21 David Conrad
            }
1172
        }
1173
1174
        /* U/V */
1175
        x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1;
1176
        for (y = 0; y < 2; y++) {
1177
            for (x = 0; x < 2; x++) {
1178
                uvmv.x = mb->bmv[ 2*y    * 4 + 2*x  ].x +
1179
                         mb->bmv[ 2*y    * 4 + 2*x+1].x +
1180
                         mb->bmv[(2*y+1) * 4 + 2*x  ].x +
1181
                         mb->bmv[(2*y+1) * 4 + 2*x+1].x;
1182
                uvmv.y = mb->bmv[ 2*y    * 4 + 2*x  ].y +
1183
                         mb->bmv[ 2*y    * 4 + 2*x+1].y +
1184
                         mb->bmv[(2*y+1) * 4 + 2*x  ].y +
1185
                         mb->bmv[(2*y+1) * 4 + 2*x+1].y;
1186 8f910a56 Stefan Gehrer
                uvmv.x = (uvmv.x + 2 + (uvmv.x >> (INT_BIT-1))) >> 2;
1187
                uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2;
1188 3b636f21 David Conrad
                if (s->profile == 3) {
1189
                    uvmv.x &= ~7;
1190
                    uvmv.y &= ~7;
1191
                }
1192
                vp8_mc(s, 0, dst[1] + 4*y*s->uvlinesize + x*4,
1193 d292c345 Jason Garrett-Glaser
                       ref->data[1], &uvmv,
1194 3b636f21 David Conrad
                       4*x + x_off, 4*y + y_off, 4, 4,
1195
                       width, height, s->uvlinesize,
1196 0ef1dbed David Conrad
                       s->put_pixels_tab[2]);
1197 3b636f21 David Conrad
                vp8_mc(s, 0, dst[2] + 4*y*s->uvlinesize + x*4,
1198 d292c345 Jason Garrett-Glaser
                       ref->data[2], &uvmv,
1199 3b636f21 David Conrad
                       4*x + x_off, 4*y + y_off, 4, 4,
1200
                       width, height, s->uvlinesize,
1201 0ef1dbed David Conrad
                       s->put_pixels_tab[2]);
1202 3b636f21 David Conrad
            }
1203
        }
1204 7c4dcf81 Ronald S. Bultje
        break;
1205
    }
1206
    case VP8_SPLITMVMODE_16x8:
1207 d292c345 Jason Garrett-Glaser
        vp8_mc_part(s, dst, ref, x_off, y_off,
1208
                    0, 0, 16, 8, width, height, &bmv[0]);
1209
        vp8_mc_part(s, dst, ref, x_off, y_off,
1210
                    0, 8, 16, 8, width, height, &bmv[1]);
1211 7c4dcf81 Ronald S. Bultje
        break;
1212
    case VP8_SPLITMVMODE_8x16:
1213 d292c345 Jason Garrett-Glaser
        vp8_mc_part(s, dst, ref, x_off, y_off,
1214
                    0, 0, 8, 16, width, height, &bmv[0]);
1215
        vp8_mc_part(s, dst, ref, x_off, y_off,
1216
                    8, 0, 8, 16, width, height, &bmv[1]);
1217 7c4dcf81 Ronald S. Bultje
        break;
1218
    case VP8_SPLITMVMODE_8x8:
1219 d292c345 Jason Garrett-Glaser
        vp8_mc_part(s, dst, ref, x_off, y_off,
1220
                    0, 0, 8, 8, width, height, &bmv[0]);
1221
        vp8_mc_part(s, dst, ref, x_off, y_off,
1222
                    8, 0, 8, 8, width, height, &bmv[1]);
1223
        vp8_mc_part(s, dst, ref, x_off, y_off,
1224
                    0, 8, 8, 8, width, height, &bmv[2]);
1225
        vp8_mc_part(s, dst, ref, x_off, y_off,
1226
                    8, 8, 8, 8, width, height, &bmv[3]);
1227 7c4dcf81 Ronald S. Bultje
        break;
1228 3b636f21 David Conrad
    }
1229
}
1230
1231 414ac27d Jason Garrett-Glaser
static av_always_inline void idct_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb)
1232 3b636f21 David Conrad
{
1233 3df56f41 Jason Garrett-Glaser
    int x, y, ch;
1234 3b636f21 David Conrad
1235 8a467b2d Jason Garrett-Glaser
    if (mb->mode != MODE_I4x4) {
1236
        uint8_t *y_dst = dst[0];
1237 3b636f21 David Conrad
        for (y = 0; y < 4; y++) {
1238 3df56f41 Jason Garrett-Glaser
            uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[y]);
1239
            if (nnz4) {
1240
                if (nnz4&~0x01010101) {
1241 8a467b2d Jason Garrett-Glaser
                    for (x = 0; x < 4; x++) {
1242 3df56f41 Jason Garrett-Glaser
                        int nnz = s->non_zero_count_cache[y][x];
1243 8a467b2d Jason Garrett-Glaser
                        if (nnz) {
1244
                            if (nnz == 1)
1245
                                s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize);
1246
                            else
1247
                                s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize);
1248
                        }
1249
                    }
1250
                } else {
1251 3ae079a3 Jason Garrett-Glaser
                    s->vp8dsp.vp8_idct_dc_add4y(y_dst, s->block[y], s->linesize);
1252 3b636f21 David Conrad
                }
1253
            }
1254
            y_dst += 4*s->linesize;
1255
        }
1256 8a467b2d Jason Garrett-Glaser
    }
1257 3b636f21 David Conrad
1258 8a467b2d Jason Garrett-Glaser
    for (ch = 0; ch < 2; ch++) {
1259 3ae079a3 Jason Garrett-Glaser
        uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[4+ch]);
1260
        if (nnz4) {
1261 8a467b2d Jason Garrett-Glaser
            uint8_t *ch_dst = dst[1+ch];
1262 3ae079a3 Jason Garrett-Glaser
            if (nnz4&~0x01010101) {
1263
                for (y = 0; y < 2; y++) {
1264
                    for (x = 0; x < 2; x++) {
1265
                        int nnz = s->non_zero_count_cache[4+ch][(y<<1)+x];
1266
                        if (nnz) {
1267
                            if (nnz == 1)
1268
                                s->vp8dsp.vp8_idct_dc_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);
1269
                            else
1270
                                s->vp8dsp.vp8_idct_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);
1271
                        }
1272 8a467b2d Jason Garrett-Glaser
                    }
1273 3ae079a3 Jason Garrett-Glaser
                    ch_dst += 4*s->uvlinesize;
1274 8a467b2d Jason Garrett-Glaser
                }
1275 3ae079a3 Jason Garrett-Glaser
            } else {
1276
                s->vp8dsp.vp8_idct_dc_add4uv(ch_dst, s->block[4+ch], s->uvlinesize);
1277 3b636f21 David Conrad
            }
1278
        }
1279
    }
1280
}
1281
1282 414ac27d Jason Garrett-Glaser
static av_always_inline void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f )
1283 3b636f21 David Conrad
{
1284
    int interior_limit, filter_level;
1285
1286
    if (s->segmentation.enabled) {
1287 b9a7186b Jason Garrett-Glaser
        filter_level = s->segmentation.filter_level[s->segment];
1288 3b636f21 David Conrad
        if (!s->segmentation.absolute_vals)
1289
            filter_level += s->filter.level;
1290
    } else
1291
        filter_level = s->filter.level;
1292
1293
    if (s->lf_delta.enabled) {
1294
        filter_level += s->lf_delta.ref[mb->ref_frame];
1295
1296
        if (mb->ref_frame == VP56_FRAME_CURRENT) {
1297
            if (mb->mode == MODE_I4x4)
1298
                filter_level += s->lf_delta.mode[0];
1299
        } else {
1300
            if (mb->mode == VP8_MVMODE_ZERO)
1301
                filter_level += s->lf_delta.mode[1];
1302
            else if (mb->mode == VP8_MVMODE_SPLIT)
1303
                filter_level += s->lf_delta.mode[3];
1304
            else
1305
                filter_level += s->lf_delta.mode[2];
1306
        }
1307
    }
1308
    filter_level = av_clip(filter_level, 0, 63);
1309
1310
    interior_limit = filter_level;
1311
    if (s->filter.sharpness) {
1312
        interior_limit >>= s->filter.sharpness > 4 ? 2 : 1;
1313
        interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness);
1314
    }
1315
    interior_limit = FFMAX(interior_limit, 1);
1316
1317 968570d6 Jason Garrett-Glaser
    f->filter_level = filter_level;
1318
    f->inner_limit = interior_limit;
1319 c55e0d34 Jason Garrett-Glaser
    f->inner_filter = !mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT;
1320 3b636f21 David Conrad
}
1321
1322 414ac27d Jason Garrett-Glaser
static av_always_inline void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f, int mb_x, int mb_y)
1323 3b636f21 David Conrad
{
1324 968570d6 Jason Garrett-Glaser
    int mbedge_lim, bedge_lim, hev_thresh;
1325
    int filter_level = f->filter_level;
1326
    int inner_limit = f->inner_limit;
1327 c55e0d34 Jason Garrett-Glaser
    int inner_filter = f->inner_filter;
1328 145d3186 Jason Garrett-Glaser
    int linesize = s->linesize;
1329
    int uvlinesize = s->uvlinesize;
1330 3b636f21 David Conrad
1331
    if (!filter_level)
1332
        return;
1333
1334 5245c04d David Conrad
    mbedge_lim = 2*(filter_level+2) + inner_limit;
1335
     bedge_lim = 2* filter_level    + inner_limit;
1336 968570d6 Jason Garrett-Glaser
    hev_thresh = filter_level >= 15;
1337
1338
    if (s->keyframe) {
1339
        if (filter_level >= 40)
1340
            hev_thresh = 2;
1341
    } else {
1342
        if (filter_level >= 40)
1343
            hev_thresh = 3;
1344
        else if (filter_level >= 20)
1345
            hev_thresh = 2;
1346
    }
1347 5245c04d David Conrad
1348 3b636f21 David Conrad
    if (mb_x) {
1349 145d3186 Jason Garrett-Glaser
        s->vp8dsp.vp8_h_loop_filter16y(dst[0],     linesize,
1350 3facfc99 Ronald S. Bultje
                                       mbedge_lim, inner_limit, hev_thresh);
1351 145d3186 Jason Garrett-Glaser
        s->vp8dsp.vp8_h_loop_filter8uv(dst[1],     dst[2],      uvlinesize,
1352 3facfc99 Ronald S. Bultje
                                       mbedge_lim, inner_limit, hev_thresh);
1353 3b636f21 David Conrad
    }
1354
1355 c55e0d34 Jason Garrett-Glaser
    if (inner_filter) {
1356 145d3186 Jason Garrett-Glaser
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 4, linesize, bedge_lim,
1357
                                             inner_limit, hev_thresh);
1358
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 8, linesize, bedge_lim,
1359
                                             inner_limit, hev_thresh);
1360
        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+12, linesize, bedge_lim,
1361
                                             inner_limit, hev_thresh);
1362
        s->vp8dsp.vp8_h_loop_filter8uv_inner(dst[1] + 4, dst[2] + 4,
1363
                                             uvlinesize,  bedge_lim,
1364
                                             inner_limit, hev_thresh);
1365 3b636f21 David Conrad
    }
1366
1367
    if (mb_y) {
1368 145d3186 Jason Garrett-Glaser
        s->vp8dsp.vp8_v_loop_filter16y(dst[0],     linesize,
1369 3facfc99 Ronald S. Bultje
                                       mbedge_lim, inner_limit, hev_thresh);
1370 145d3186 Jason Garrett-Glaser
        s->vp8dsp.vp8_v_loop_filter8uv(dst[1],     dst[2],      uvlinesize,
1371 3facfc99 Ronald S. Bultje
                                       mbedge_lim, inner_limit, hev_thresh);
1372 3b636f21 David Conrad
    }
1373
1374 c55e0d34 Jason Garrett-Glaser
    if (inner_filter) {
1375 145d3186 Jason Garrett-Glaser
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 4*linesize,
1376
                                             linesize,    bedge_lim,
1377
                                             inner_limit, hev_thresh);
1378
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 8*linesize,
1379
                                             linesize,    bedge_lim,
1380
                                             inner_limit, hev_thresh);
1381
        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+12*linesize,
1382
                                             linesize,    bedge_lim,
1383
                                             inner_limit, hev_thresh);
1384
        s->vp8dsp.vp8_v_loop_filter8uv_inner(dst[1] + 4 * uvlinesize,
1385
                                             dst[2] + 4 * uvlinesize,
1386
                                             uvlinesize,  bedge_lim,
1387 3facfc99 Ronald S. Bultje
                                             inner_limit, hev_thresh);
1388 3b636f21 David Conrad
    }
1389
}
1390
1391 414ac27d Jason Garrett-Glaser
static av_always_inline void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f, int mb_x, int mb_y)
1392 3b636f21 David Conrad
{
1393 968570d6 Jason Garrett-Glaser
    int mbedge_lim, bedge_lim;
1394
    int filter_level = f->filter_level;
1395
    int inner_limit = f->inner_limit;
1396 c55e0d34 Jason Garrett-Glaser
    int inner_filter = f->inner_filter;
1397 145d3186 Jason Garrett-Glaser
    int linesize = s->linesize;
1398 3b636f21 David Conrad
1399
    if (!filter_level)
1400
        return;
1401
1402
    mbedge_lim = 2*(filter_level+2) + inner_limit;
1403
     bedge_lim = 2* filter_level    + inner_limit;
1404
1405
    if (mb_x)
1406 145d3186 Jason Garrett-Glaser
        s->vp8dsp.vp8_h_loop_filter_simple(dst, linesize, mbedge_lim);
1407 c55e0d34 Jason Garrett-Glaser
    if (inner_filter) {
1408 145d3186 Jason Garrett-Glaser
        s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, linesize, bedge_lim);
1409
        s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, linesize, bedge_lim);
1410
        s->vp8dsp.vp8_h_loop_filter_simple(dst+12, linesize, bedge_lim);
1411 3b636f21 David Conrad
    }
1412
1413
    if (mb_y)
1414 145d3186 Jason Garrett-Glaser
        s->vp8dsp.vp8_v_loop_filter_simple(dst, linesize, mbedge_lim);
1415 c55e0d34 Jason Garrett-Glaser
    if (inner_filter) {
1416 145d3186 Jason Garrett-Glaser
        s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*linesize, linesize, bedge_lim);
1417
        s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*linesize, linesize, bedge_lim);
1418
        s->vp8dsp.vp8_v_loop_filter_simple(dst+12*linesize, linesize, bedge_lim);
1419 3b636f21 David Conrad
    }
1420
}
1421
1422
static void filter_mb_row(VP8Context *s, int mb_y)
1423
{
1424 968570d6 Jason Garrett-Glaser
    VP8FilterStrength *f = s->filter_strength;
1425 3b636f21 David Conrad
    uint8_t *dst[3] = {
1426
        s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize,
1427
        s->framep[VP56_FRAME_CURRENT]->data[1] +  8*mb_y*s->uvlinesize,
1428
        s->framep[VP56_FRAME_CURRENT]->data[2] +  8*mb_y*s->uvlinesize
1429
    };
1430
    int mb_x;
1431
1432
    for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1433 9ac831c2 David Conrad
        backup_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], s->linesize, s->uvlinesize, 0);
1434 c55e0d34 Jason Garrett-Glaser
        filter_mb(s, dst, f++, mb_x, mb_y);
1435 3b636f21 David Conrad
        dst[0] += 16;
1436
        dst[1] += 8;
1437
        dst[2] += 8;
1438
    }
1439
}
1440
1441
static void filter_mb_row_simple(VP8Context *s, int mb_y)
1442
{
1443 968570d6 Jason Garrett-Glaser
    VP8FilterStrength *f = s->filter_strength;
1444
    uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize;
1445 3b636f21 David Conrad
    int mb_x;
1446
1447
    for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
1448 9ac831c2 David Conrad
        backup_mb_border(s->top_border[mb_x+1], dst, NULL, NULL, s->linesize, 0, 1);
1449 c55e0d34 Jason Garrett-Glaser
        filter_mb_simple(s, dst, f++, mb_x, mb_y);
1450 3b636f21 David Conrad
        dst += 16;
1451
    }
1452
}
1453
1454
static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
1455
                            AVPacket *avpkt)
1456
{
1457
    VP8Context *s = avctx->priv_data;
1458
    int ret, mb_x, mb_y, i, y, referenced;
1459
    enum AVDiscard skip_thresh;
1460 28e241de Carl Eugen Hoyos
    AVFrame *av_uninit(curframe);
1461 3b636f21 David Conrad
1462
    if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0)
1463
        return ret;
1464
1465
    referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT
1466
                                || s->update_altref == VP56_FRAME_CURRENT;
1467
1468
    skip_thresh = !referenced ? AVDISCARD_NONREF :
1469
                    !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL;
1470
1471
    if (avctx->skip_frame >= skip_thresh) {
1472
        s->invisible = 1;
1473
        goto skip_decode;
1474
    }
1475 9ac831c2 David Conrad
    s->deblock_filter = s->filter.level && avctx->skip_loop_filter < skip_thresh;
1476 3b636f21 David Conrad
1477
    for (i = 0; i < 4; i++)
1478
        if (&s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1479
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1480
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) {
1481
            curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i];
1482
            break;
1483
        }
1484
    if (curframe->data[0])
1485
        avctx->release_buffer(avctx, curframe);
1486
1487
    curframe->key_frame = s->keyframe;
1488
    curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE;
1489
    curframe->reference = referenced ? 3 : 0;
1490
    if ((ret = avctx->get_buffer(avctx, curframe))) {
1491
        av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n");
1492
        return ret;
1493
    }
1494
1495
    // Given that arithmetic probabilities are updated every frame, it's quite likely
1496
    // that the values we have on a random interframe are complete junk if we didn't
1497
    // start decode on a keyframe. So just don't display anything rather than junk.
1498
    if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] ||
1499
                         !s->framep[VP56_FRAME_GOLDEN] ||
1500
                         !s->framep[VP56_FRAME_GOLDEN2])) {
1501
        av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n");
1502
        return AVERROR_INVALIDDATA;
1503
    }
1504
1505
    s->linesize   = curframe->linesize[0];
1506
    s->uvlinesize = curframe->linesize[1];
1507
1508
    if (!s->edge_emu_buffer)
1509
        s->edge_emu_buffer = av_malloc(21*s->linesize);
1510
1511
    memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz));
1512
1513 c55e0d34 Jason Garrett-Glaser
    /* Zero macroblock structures for top/left prediction from outside the frame. */
1514
    memset(s->macroblocks, 0, (s->mb_width + s->mb_height*2)*sizeof(*s->macroblocks));
1515
1516 3b636f21 David Conrad
    // top edge of 127 for intra prediction
1517 9ac831c2 David Conrad
    memset(s->top_border, 127, (s->mb_width+1)*sizeof(*s->top_border));
1518 c4211046 Jason Garrett-Glaser
    memset(s->ref_count, 0, sizeof(s->ref_count));
1519 3b636f21 David Conrad
1520
    for (mb_y = 0; mb_y < s->mb_height; mb_y++) {
1521
        VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)];
1522 c55e0d34 Jason Garrett-Glaser
        VP8Macroblock *mb = s->macroblocks + (s->mb_height - mb_y - 1)*2;
1523 3b636f21 David Conrad
        uint8_t *intra4x4 = s->intra4x4_pred_mode + 4*mb_y*s->b4_stride;
1524 c55e0d34 Jason Garrett-Glaser
        uint8_t *segment_map = s->segmentation_map + mb_y*s->mb_stride;
1525 ef38842f Jason Garrett-Glaser
        int mb_xy = mb_y * s->mb_stride;
1526 3b636f21 David Conrad
        uint8_t *dst[3] = {
1527
            curframe->data[0] + 16*mb_y*s->linesize,
1528
            curframe->data[1] +  8*mb_y*s->uvlinesize,
1529
            curframe->data[2] +  8*mb_y*s->uvlinesize
1530
        };
1531
1532
        memset(s->left_nnz, 0, sizeof(s->left_nnz));
1533
1534
        // left edge of 129 for intra prediction
1535
        if (!(avctx->flags & CODEC_FLAG_EMU_EDGE))
1536
            for (i = 0; i < 3; i++)
1537
                for (y = 0; y < 16>>!!i; y++)
1538
                    dst[i][y*curframe->linesize[i]-1] = 129;
1539 9ac831c2 David Conrad
        if (mb_y)
1540
            memset(s->top_border, 129, sizeof(*s->top_border));
1541 3b636f21 David Conrad
1542 ef38842f Jason Garrett-Glaser
        for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb_xy++, mb++) {
1543 d1c58fce Jason Garrett-Glaser
            uint8_t *intra4x4_mb = s->keyframe ? intra4x4 + 4*mb_x : s->intra4x4_pred_mode_mb;
1544 c55e0d34 Jason Garrett-Glaser
            uint8_t *segment_mb = segment_map+mb_x;
1545 d1c58fce Jason Garrett-Glaser
1546 d864dee8 Jason Garrett-Glaser
            /* Prefetch the current frame, 4 MBs ahead */
1547
            s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4);
1548
            s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2);
1549
1550 c55e0d34 Jason Garrett-Glaser
            decode_mb_mode(s, mb, mb_x, mb_y, intra4x4_mb, segment_mb);
1551 3b636f21 David Conrad
1552 ef38842f Jason Garrett-Glaser
            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_PREVIOUS);
1553 c4211046 Jason Garrett-Glaser
1554 3b636f21 David Conrad
            if (!mb->skip)
1555
                decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);
1556
1557 b946111f Jason Garrett-Glaser
            if (mb->mode <= MODE_I4x4)
1558 d1c58fce Jason Garrett-Glaser
                intra_predict(s, dst, mb, intra4x4_mb, mb_x, mb_y);
1559 b946111f Jason Garrett-Glaser
            else
1560 3b636f21 David Conrad
                inter_predict(s, dst, mb, mb_x, mb_y);
1561
1562 ef38842f Jason Garrett-Glaser
            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN);
1563 c4211046 Jason Garrett-Glaser
1564 3b636f21 David Conrad
            if (!mb->skip) {
1565 8a467b2d Jason Garrett-Glaser
                idct_mb(s, dst, mb);
1566 3b636f21 David Conrad
            } else {
1567
                AV_ZERO64(s->left_nnz);
1568
                AV_WN64(s->top_nnz[mb_x], 0);   // array of 9, so unaligned
1569
1570
                // Reset DC block predictors if they would exist if the mb had coefficients
1571
                if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
1572
                    s->left_nnz[8]      = 0;
1573
                    s->top_nnz[mb_x][8] = 0;
1574
                }
1575
            }
1576
1577 968570d6 Jason Garrett-Glaser
            if (s->deblock_filter)
1578
                filter_level_for_mb(s, mb, &s->filter_strength[mb_x]);
1579
1580 ef38842f Jason Garrett-Glaser
            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN2);
1581 c4211046 Jason Garrett-Glaser
1582 3b636f21 David Conrad
            dst[0] += 16;
1583
            dst[1] += 8;
1584
            dst[2] += 8;
1585
        }
1586 9ac831c2 David Conrad
        if (s->deblock_filter) {
1587 3b636f21 David Conrad
            if (s->filter.simple)
1588 9ac831c2 David Conrad
                filter_mb_row_simple(s, mb_y);
1589 3b636f21 David Conrad
            else
1590 9ac831c2 David Conrad
                filter_mb_row(s, mb_y);
1591 3b636f21 David Conrad
        }
1592
    }
1593
1594
skip_decode:
1595
    // if future frames don't use the updated probabilities,
1596
    // reset them to the values we saved
1597
    if (!s->update_probabilities)
1598
        s->prob[0] = s->prob[1];
1599
1600
    // check if golden and altref are swapped
1601
    if (s->update_altref == VP56_FRAME_GOLDEN &&
1602
        s->update_golden == VP56_FRAME_GOLDEN2)
1603
        FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN], s->framep[VP56_FRAME_GOLDEN2]);
1604
    else {
1605
        if (s->update_altref != VP56_FRAME_NONE)
1606
            s->framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref];
1607
1608
        if (s->update_golden != VP56_FRAME_NONE)
1609
            s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden];
1610
    }
1611
1612
    if (s->update_last) // move cur->prev
1613
        s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT];
1614
1615
    // release no longer referenced frames
1616
    for (i = 0; i < 4; i++)
1617
        if (s->frames[i].data[0] &&
1618
            &s->frames[i] != s->framep[VP56_FRAME_CURRENT] &&
1619
            &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1620
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1621
            &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2])
1622
            avctx->release_buffer(avctx, &s->frames[i]);
1623
1624
    if (!s->invisible) {
1625
        *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT];
1626
        *data_size = sizeof(AVFrame);
1627
    }
1628
1629
    return avpkt->size;
1630
}
1631
1632
static av_cold int vp8_decode_init(AVCodecContext *avctx)
1633
{
1634
    VP8Context *s = avctx->priv_data;
1635
1636
    s->avctx = avctx;
1637
    avctx->pix_fmt = PIX_FMT_YUV420P;
1638
1639
    dsputil_init(&s->dsp, avctx);
1640
    ff_h264_pred_init(&s->hpc, CODEC_ID_VP8);
1641
    ff_vp8dsp_init(&s->vp8dsp);
1642
1643
    // intra pred needs edge emulation among other things
1644
    if (avctx->flags&CODEC_FLAG_EMU_EDGE) {
1645 03ac56e7 Jason Garrett-Glaser
        av_log(avctx, AV_LOG_ERROR, "Edge emulation not supported\n");
1646 3b636f21 David Conrad
        return AVERROR_PATCHWELCOME;
1647
    }
1648
1649
    return 0;
1650
}
1651
1652
static av_cold int vp8_decode_free(AVCodecContext *avctx)
1653
{
1654
    vp8_decode_flush(avctx);
1655
    return 0;
1656
}
1657
1658
AVCodec vp8_decoder = {
1659
    "vp8",
1660
    AVMEDIA_TYPE_VIDEO,
1661
    CODEC_ID_VP8,
1662
    sizeof(VP8Context),
1663
    vp8_decode_init,
1664
    NULL,
1665
    vp8_decode_free,
1666
    vp8_decode_frame,
1667
    CODEC_CAP_DR1,
1668
    .flush = vp8_decode_flush,
1669
    .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"),
1670
};