decode.c 125 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
/*
 * Copyright © 2018, VideoLAN and dav1d authors
 * Copyright © 2018, Two Orioles, LLC
 * All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions are met:
 *
 * 1. Redistributions of source code must retain the above copyright notice, this
 *    list of conditions and the following disclaimer.
 *
 * 2. Redistributions in binary form must reproduce the above copyright notice,
 *    this list of conditions and the following disclaimer in the documentation
 *    and/or other materials provided with the distribution.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
 * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */

#include "config.h"

#include <errno.h>
#include <limits.h>
#include <string.h>
#include <stdio.h>
#include <inttypes.h>

#include "dav1d/data.h"

#include "common/intops.h"
#include "common/mem.h"

#include "src/decode.h"
#include "src/dequant_tables.h"
#include "src/env.h"
#include "src/qm.h"
#include "src/recon.h"
#include "src/ref.h"
#include "src/tables.h"
#include "src/thread_task.h"
#include "src/warpmv.h"

static void init_quant_tables(const Av1SequenceHeader *const seq_hdr,
                              const Av1FrameHeader *const frame_hdr,
                              const int qidx, uint16_t (*dq)[3][2])
{
    for (int i = 0; i < (frame_hdr->segmentation.enabled ? 8 : 1); i++) {
        const int yac = frame_hdr->segmentation.enabled ?
            iclip_u8(qidx + frame_hdr->segmentation.seg_data.d[i].delta_q) : qidx;
        const int ydc = iclip_u8(yac + frame_hdr->quant.ydc_delta);
        const int uac = iclip_u8(yac + frame_hdr->quant.uac_delta);
        const int udc = iclip_u8(yac + frame_hdr->quant.udc_delta);
        const int vac = iclip_u8(yac + frame_hdr->quant.vac_delta);
        const int vdc = iclip_u8(yac + frame_hdr->quant.vdc_delta);

        dq[i][0][0] = dav1d_dq_tbl[seq_hdr->bpc > 8][ydc][0];
        dq[i][0][1] = dav1d_dq_tbl[seq_hdr->bpc > 8][yac][1];
        dq[i][1][0] = dav1d_dq_tbl[seq_hdr->bpc > 8][udc][0];
        dq[i][1][1] = dav1d_dq_tbl[seq_hdr->bpc > 8][uac][1];
        dq[i][2][0] = dav1d_dq_tbl[seq_hdr->bpc > 8][vdc][0];
        dq[i][2][1] = dav1d_dq_tbl[seq_hdr->bpc > 8][vac][1];
    }
}

static int read_mv_component_diff(Dav1dTileContext *const t,
                                  CdfMvComponent *const mv_comp,
                                  const int have_fp)
{
    Dav1dTileState *const ts = t->ts;
    const Dav1dFrameContext *const f = t->f;
    const int have_hp = f->frame_hdr.hp;
    const int sign = msac_decode_bool_adapt(&ts->msac, mv_comp->sign);
    const int cl = msac_decode_symbol_adapt(&ts->msac, mv_comp->classes, 11);
    int up, fp, hp;

    if (!cl) {
        up = msac_decode_bool_adapt(&ts->msac, mv_comp->class0);
        if (have_fp) {
            fp = msac_decode_symbol_adapt(&ts->msac, mv_comp->class0_fp[up], 4);
            hp = have_hp ? msac_decode_bool_adapt(&ts->msac, mv_comp->class0_hp) : 1;
        } else {
            fp = 3;
            hp = 1;
        }
    } else {
        up = 1 << cl;
        for (int n = 0; n < cl; n++)
            up |= msac_decode_bool_adapt(&ts->msac, mv_comp->classN[n]) << n;
        if (have_fp) {
            fp = msac_decode_symbol_adapt(&ts->msac, mv_comp->classN_fp, 4);
            hp = have_hp ? msac_decode_bool_adapt(&ts->msac, mv_comp->classN_hp) : 1;
        } else {
            fp = 3;
            hp = 1;
        }
    }

    const int diff = ((up << 3) | (fp << 1) | hp) + 1;

    return sign ? -diff : diff;
}

static void read_mv_residual(Dav1dTileContext *const t, mv *const ref_mv,
                             CdfMvContext *const mv_cdf, const int have_fp)
{
    switch (msac_decode_symbol_adapt(&t->ts->msac, t->ts->cdf.mv.joint, N_MV_JOINTS)) {
    case MV_JOINT_HV:
        ref_mv->y += read_mv_component_diff(t, &mv_cdf->comp[0], have_fp);
        ref_mv->x += read_mv_component_diff(t, &mv_cdf->comp[1], have_fp);
        break;
    case MV_JOINT_H:
        ref_mv->x += read_mv_component_diff(t, &mv_cdf->comp[1], have_fp);
        break;
    case MV_JOINT_V:
        ref_mv->y += read_mv_component_diff(t, &mv_cdf->comp[0], have_fp);
        break;
    default:
        break;
    }
}

static void read_tx_tree(Dav1dTileContext *const t,
                         const enum RectTxfmSize from,
                         const int depth, uint16_t *const masks,
                         const int x_off, const int y_off)
{
    const Dav1dFrameContext *const f = t->f;
    const int bx4 = t->bx & 31, by4 = t->by & 31;
137
    const TxfmInfo *const t_dim = &dav1d_txfm_dimensions[from];
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
    const int txw = t_dim->lw, txh = t_dim->lh;
    int is_split;

    if (depth < 2 && from > (int) TX_4X4) {
        const int cat = 2 * (TX_64X64 - t_dim->max) - depth;
        const int a = t->a->tx[bx4] < txw;
        const int l = t->l.tx[by4] < txh;

        is_split = msac_decode_bool_adapt(&t->ts->msac, t->ts->cdf.m.txpart[cat][a + l]);
        if (is_split)
            masks[depth] |= 1 << (y_off * 4 + x_off);
    } else {
        is_split = 0;
    }

    if (is_split && t_dim->max > TX_8X8) {
        const enum RectTxfmSize sub = t_dim->sub;
155
        const TxfmInfo *const sub_t_dim = &dav1d_txfm_dimensions[sub];
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
        const int txsw = sub_t_dim->w, txsh = sub_t_dim->h;

        read_tx_tree(t, sub, depth + 1, masks, x_off * 2 + 0, y_off * 2 + 0);
        t->bx += txsw;
        if (txw >= txh && t->bx < f->bw)
            read_tx_tree(t, sub, depth + 1, masks, x_off * 2 + 1, y_off * 2 + 0);
        t->bx -= txsw;
        t->by += txsh;
        if (txh >= txw && t->by < f->bh) {
            read_tx_tree(t, sub, depth + 1, masks, x_off * 2 + 0, y_off * 2 + 1);
            t->bx += txsw;
            if (txw >= txh && t->bx < f->bw)
                read_tx_tree(t, sub, depth + 1, masks,
                             x_off * 2 + 1, y_off * 2 + 1);
            t->bx -= txsw;
        }
        t->by -= txsh;
    } else {
        memset(&t->a->tx[bx4], is_split ? TX_4X4 : txw, t_dim->w);
        memset(&t->l.tx[by4], is_split ? TX_4X4 : txh, t_dim->h);
    }
}

179
static int neg_deinterleave(int diff, int ref, int max) {
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
    if (!ref) return diff;
    if (ref >= (max - 1)) return max - diff - 1;
    if (2 * ref < max) {
        if (diff <= 2 * ref) {
            if (diff & 1)
                return ref + ((diff + 1) >> 1);
            else
                return ref - (diff >> 1);
        }
        return diff;
    } else {
        if (diff <= 2 * (max - ref - 1)) {
            if (diff & 1)
                return ref + ((diff + 1) >> 1);
            else
                return ref - (diff >> 1);
        }
        return max - (diff + 1);
    }
}

static void find_matching_ref(const Dav1dTileContext *const t,
                              const enum EdgeFlags intra_edge_flags,
                              const int bw4, const int bh4,
                              const int w4, const int h4,
                              const int have_left, const int have_top,
                              const int ref, uint64_t masks[2])
{
    const Dav1dFrameContext *const f = t->f;
    const ptrdiff_t b4_stride = f->b4_stride;
    const refmvs *const r = &f->mvs[t->by * b4_stride + t->bx];
    int count = 0;
    int have_topleft = have_top && have_left;
    int have_topright = imax(bw4, bh4) < 32 &&
                        have_top && t->bx + bw4 < t->ts->tiling.col_end &&
                        (intra_edge_flags & EDGE_I444_TOP_HAS_RIGHT);

217
#define bs(rp) dav1d_block_dimensions[sbtype_to_bs[(rp)->sb_type]]
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
#define matches(rp) ((rp)->ref[0] == ref + 1 && (rp)->ref[1] == -1)

    if (have_top) {
        const refmvs *r2 = &r[-b4_stride];
        if (matches(r2)) {
            masks[0] |= 1;
            count = 1;
        }
        int aw4 = bs(r2)[0];
        if (aw4 >= bw4) {
            const int off = t->bx & (aw4 - 1);
            if (off) have_topleft = 0;
            if (aw4 - off > bw4) have_topright = 0;
        } else {
            unsigned mask = 1 << aw4;
            for (int x = aw4; x < w4; x += aw4) {
                r2 += aw4;
                if (matches(r2)) {
                    masks[0] |= mask;
                    if (++count >= 8) return;
                }
                aw4 = bs(r2)[0];
                mask <<= aw4;
            }
        }
    }
    if (have_left) {
        const refmvs *r2 = &r[-1];
        if (matches(r2)) {
            masks[1] |= 1;
            if (++count >= 8) return;
        }
        int lh4 = bs(r2)[1];
        if (lh4 >= bh4) {
            if (t->by & (lh4 - 1)) have_topleft = 0;
        } else {
            unsigned mask = 1 << lh4;
            for (int y = lh4; y < h4; y += lh4) {
                r2 += lh4 * b4_stride;
                if (matches(r2)) {
                    masks[1] |= mask;
                    if (++count >= 8) return;
                }
                lh4 = bs(r2)[1];
                mask <<= lh4;
            }
        }
    }
    if (have_topleft && matches(&r[-(1 + b4_stride)])) {
        masks[1] |= 1ULL << 32;
        if (++count >= 8) return;
    }
    if (have_topright && matches(&r[bw4 - b4_stride])) {
        masks[0] |= 1ULL << 32;
    }
#undef matches
}

static void derive_warpmv(const Dav1dTileContext *const t,
                          const int bw4, const int bh4,
                          const uint64_t masks[2], const struct mv mv,
                          WarpedMotionParams *const wmp)
{
    int pts[8][2 /* in, out */][2 /* x, y */], np = 0;
    const Dav1dFrameContext *const f = t->f;
    const ptrdiff_t b4_stride = f->b4_stride;
    const refmvs *const r = &f->mvs[t->by * b4_stride + t->bx];

#define add_sample(dx, dy, sx, sy, rp) do { \
    pts[np][0][0] = 16 * (2 * dx + sx * bs(rp)[0]) - 8; \
    pts[np][0][1] = 16 * (2 * dy + sy * bs(rp)[1]) - 8; \
    pts[np][1][0] = pts[np][0][0] + (rp)->mv[0].x; \
    pts[np][1][1] = pts[np][0][1] + (rp)->mv[0].y; \
    np++; \
} while (0)

    // use masks[] to find the projectable motion vectors in the edges
    if ((unsigned) masks[0] == 1 && !(masks[1] >> 32)) {
        const int off = t->bx & (bs(&r[-b4_stride])[0] - 1);
        add_sample(-off, 0, 1, -1, &r[-b4_stride]);
    } else for (unsigned off = 0, xmask = masks[0]; np < 8 && xmask;) { // top
299
        const int tz = ctz(xmask);
300
301
302
303
304
305
306
307
308
        off += tz;
        add_sample(off, 0, 1, -1, &r[off - b4_stride]);
        xmask >>= tz + 1;
        off += 1;
    }
    if (np < 8 && masks[1] == 1) {
        const int off = t->by & (bs(&r[-1])[1] - 1);
        add_sample(0, -off, -1, 1, &r[-1 - off * b4_stride]);
    } else for (unsigned off = 0, ymask = masks[1]; np < 8 && ymask;) { // left
309
        const int tz = ctz(ymask);
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
        off += tz;
        add_sample(0, off, -1, 1, &r[off * b4_stride - 1]);
        ymask >>= tz + 1;
        off += 1;
    }
    if (np < 8 && masks[1] >> 32) // top/left
        add_sample(0, 0, -1, -1, &r[-(1 + b4_stride)]);
    if (np < 8 && masks[0] >> 32) // top/right
        add_sample(bw4, 0, 1, -1, &r[bw4 - b4_stride]);
    assert(np > 0 && np <= 8);
#undef bs

    // select according to motion vector difference against a threshold
    int mvd[8], ret = 0;
    const int thresh = 4 * iclip(imax(bw4, bh4), 4, 28);
    for (int i = 0; i < np; i++) {
        mvd[i] = labs(pts[i][1][0] - pts[i][0][0] - mv.x) +
                 labs(pts[i][1][1] - pts[i][0][1] - mv.y);
        if (mvd[i] > thresh)
            mvd[i] = -1;
        else
            ret++;
    }
    if (!ret) {
        ret = 1;
    } else for (int i = 0, j = np - 1, k = 0; k < np - ret; k++, i++, j--) {
        while (mvd[i] != -1) i++;
        while (mvd[j] == -1) j--;
        assert(i != j);
        if (i > j) break;
        // replace the discarded samples;
        mvd[i] = mvd[j];
        memcpy(pts[i], pts[j], sizeof(*pts));
    }

345
346
    if (!dav1d_find_affine_int(pts, ret, bw4, bh4, mv, wmp, t->bx, t->by) &&
        !dav1d_get_shear_params(wmp))
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
    {
        wmp->type = WM_TYPE_AFFINE;
    } else
        wmp->type = WM_TYPE_IDENTITY;
}

static inline int findoddzero(const uint8_t *buf, int len) {
    for (int n = 0; n < len; n++)
        if (!buf[n * 2]) return 1;
    return 0;
}

static void read_pal_plane(Dav1dTileContext *const t, Av1Block *const b,
                           const int pl, const int sz_ctx,
                           const int bx4, const int by4)
{
    Dav1dTileState *const ts = t->ts;
    const Dav1dFrameContext *const f = t->f;
    const int pal_sz = b->pal_sz[pl] = 2 + msac_decode_symbol_adapt(&ts->msac,
                                                 ts->cdf.m.pal_sz[pl][sz_ctx], 7);
    uint16_t cache[16], used_cache[8];
    int l_cache = pl ? t->pal_sz_uv[1][by4] : t->l.pal_sz[by4];
    int n_cache = 0;
    // don't reuse above palette outside SB64 boundaries
    int a_cache = by4 & 15 ? pl ? t->pal_sz_uv[0][bx4] : t->a->pal_sz[bx4] : 0;
    const uint16_t *l = t->al_pal[1][by4][pl], *a = t->al_pal[0][bx4][pl];

    // fill/sort cache
    while (l_cache && a_cache) {
        if (*l < *a) {
            if (!n_cache || cache[n_cache - 1] != *l)
                cache[n_cache++] = *l;
            l++;
            l_cache--;
        } else {
            if (*a == *l) {
                l++;
                l_cache--;
            }
            if (!n_cache || cache[n_cache - 1] != *a)
                cache[n_cache++] = *a;
            a++;
            a_cache--;
        }
    }
    if (l_cache) {
        do {
            if (!n_cache || cache[n_cache - 1] != *l)
                cache[n_cache++] = *l;
            l++;
        } while (--l_cache > 0);
    } else if (a_cache) {
        do {
            if (!n_cache || cache[n_cache - 1] != *a)
                cache[n_cache++] = *a;
            a++;
        } while (--a_cache > 0);
    }

    // find reused cache entries
    int i = 0;
    for (int n = 0; n < n_cache && i < pal_sz; n++)
        if (msac_decode_bool(&ts->msac, 128 << 7))
            used_cache[i++] = cache[n];
    const int n_used_cache = i;

    // parse new entries
    uint16_t *const pal = f->frame_thread.pass ?
        f->frame_thread.pal[((t->by >> 1) + (t->bx & 1)) * (f->b4_stride >> 1) +
                            ((t->bx >> 1) + (t->by & 1))][pl] : t->pal[pl];
    if (i < pal_sz) {
        int prev = pal[i++] = msac_decode_bools(&ts->msac, f->cur.p.p.bpc);

        if (i < pal_sz) {
            int bits = f->cur.p.p.bpc - 3 + msac_decode_bools(&ts->msac, 2);
            const int max = (1 << f->cur.p.p.bpc) - 1;

            do {
                const int delta = msac_decode_bools(&ts->msac, bits);
                prev = pal[i++] = imin(prev + delta + !pl, max);
                if (prev + !pl >= max) {
                    for (; i < pal_sz; i++)
                        pal[i] = pal[i - 1];
                    break;
                }
                bits = imin(bits, 1 + ulog2(max - prev - !pl));
            } while (i < pal_sz);
        }

        // merge cache+new entries
        int n = 0, m = n_used_cache;
        for (i = 0; i < pal_sz; i++) {
            if (n < n_used_cache && (m >= pal_sz || used_cache[n] <= pal[m])) {
                pal[i] = used_cache[n++];
            } else {
                assert(m < pal_sz);
                pal[i] = pal[m++];
            }
        }
    } else {
        memcpy(pal, used_cache, n_used_cache * sizeof(*used_cache));
    }

    if (DEBUG_BLOCK_INFO) {
        printf("Post-pal[pl=%d,sz=%d,cache_size=%d,used_cache=%d]: r=%d, cache=",
452
               pl, pal_sz, n_cache, n_used_cache, ts->msac.rng);
453
454
455
        for (int n = 0; n < n_cache; n++)
            printf("%c%02x", n ? ' ' : '[', cache[n]);
        printf("%s, pal=", n_cache ? "]" : "[]");
456
        for (int n = 0; n < pal_sz; n++)
457
458
459
460
461
462
            printf("%c%02x", n ? ' ' : '[', pal[n]);
        printf("]\n");
    }
}

static void read_pal_uv(Dav1dTileContext *const t, Av1Block *const b,
463
                        const int sz_ctx, const int bx4, const int by4)
464
{
465
    read_pal_plane(t, b, 1, sz_ctx, bx4, by4);
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515

    // V pal coding
    Dav1dTileState *const ts = t->ts;
    const Dav1dFrameContext *const f = t->f;
    uint16_t *const pal = f->frame_thread.pass ?
        f->frame_thread.pal[((t->by >> 1) + (t->bx & 1)) * (f->b4_stride >> 1) +
                            ((t->bx >> 1) + (t->by & 1))][2] : t->pal[2];
    if (msac_decode_bool(&ts->msac, 128 << 7)) {
        const int bits = f->cur.p.p.bpc - 4 + msac_decode_bools(&ts->msac, 2);
        int prev = pal[0] = msac_decode_bools(&ts->msac, f->cur.p.p.bpc);
        const int max = (1 << f->cur.p.p.bpc) - 1;
        for (int i = 1; i < b->pal_sz[1]; i++) {
            int delta = msac_decode_bools(&ts->msac, bits);
            if (delta && msac_decode_bool(&ts->msac, 128 << 7)) delta = -delta;
            prev = pal[i] = (prev + delta) & max;
        }
    } else {
        for (int i = 0; i < b->pal_sz[1]; i++)
            pal[i] = msac_decode_bools(&ts->msac, f->cur.p.p.bpc);
    }
    if (DEBUG_BLOCK_INFO) {
        printf("Post-pal[pl=2]: r=%d ", ts->msac.rng);
        for (int n = 0; n < b->pal_sz[1]; n++)
            printf("%c%02x", n ? ' ' : '[', pal[n]);
        printf("]\n");
    }
}

// meant to be SIMD'able, so that theoretical complexity of this function
// times block size goes from w4*h4 to w4+h4-1
// a and b are previous two lines containing (a) top/left entries or (b)
// top/left entries, with a[0] being either the first top or first left entry,
// depending on top_offset being 1 or 0, and b being the first top/left entry
// for whichever has one. left_offset indicates whether the (len-1)th entry
// has a left neighbour.
// output is order[] and ctx for each member of this diagonal.
static void order_palette(const uint8_t *pal_idx, const ptrdiff_t stride,
                          const int i, const int first, const int last,
                          uint8_t (*const order)[8], uint8_t *const ctx)
{
    int have_top = i > first;

    pal_idx += first + (i - first) * stride;
    for (int j = first, n = 0; j >= last; have_top = 1, j--, n++, pal_idx += stride - 1) {
        const int have_left = j > 0;

        assert(have_left || have_top);

#define add(v_in) do { \
        const int v = v_in; \
516
        assert((unsigned)v < 8U); \
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
        order[n][o_idx++] = v; \
        mask |= 1 << v; \
    } while (0)

        unsigned mask = 0;
        int o_idx = 0;
        if (!have_left) {
            ctx[n] = 0;
            add(pal_idx[-stride]);
        } else if (!have_top) {
            ctx[n] = 0;
            add(pal_idx[-1]);
        } else {
            const int l = pal_idx[-1], t = pal_idx[-stride], tl = pal_idx[-(stride + 1)];
            const int same_t_l = t == l;
            const int same_t_tl = t == tl;
            const int same_l_tl = l == tl;
            const int same_all = same_t_l & same_t_tl & same_l_tl;

            if (same_all) {
                ctx[n] = 4;
                add(t);
            } else if (same_t_l) {
                ctx[n] = 3;
                add(t);
                add(tl);
            } else if (same_t_tl | same_l_tl) {
                ctx[n] = 2;
                add(tl);
                add(same_t_tl ? l : t);
            } else {
                ctx[n] = 1;
                add(imin(t, l));
                add(imax(t, l));
                add(tl);
            }
        }
        for (unsigned m = 1, bit = 0; m < 0x100; m <<= 1, bit++)
            if (!(mask & m))
                order[n][o_idx++] = bit;
        assert(o_idx == 8);
#undef add
    }
}

static void read_pal_indices(Dav1dTileContext *const t,
                             uint8_t *const pal_idx,
                             const Av1Block *const b, const int pl,
                             const int w4, const int h4,
                             const int bw4, const int bh4)
{
    Dav1dTileState *const ts = t->ts;
    const ptrdiff_t stride = bw4 * 4;
    pal_idx[0] = msac_decode_uniform(&ts->msac, b->pal_sz[pl]);
    uint16_t (*const color_map_cdf)[8 + 1] =
        ts->cdf.m.color_map[pl][b->pal_sz[pl] - 2];
    for (int i = 1; i < 4 * (w4 + h4) - 1; i++) {
        // top/left-to-bottom/right diagonals ("wave-front")
        uint8_t order[64][8], ctx[64];
        const int first = imin(i, w4 * 4 - 1);
        const int last = imax(0, i - h4 * 4 + 1);
        order_palette(pal_idx, stride, i, first, last, order, ctx);
        for (int j = first, m = 0; j >= last; j--, m++) {
            const int color_idx =
                msac_decode_symbol_adapt(&ts->msac, color_map_cdf[ctx[m]],
                                         b->pal_sz[pl]);
            pal_idx[(i - j) * stride + j] = order[m][color_idx];
        }
    }
    // fill invisible edges
    if (bw4 > w4)
        for (int y = 0; y < 4 * h4; y++)
            memset(&pal_idx[y * stride + 4 * w4],
                   pal_idx[y * stride + 4 * w4 - 1], 4 * (bw4 - w4));
    if (h4 < bh4) {
        const uint8_t *const src = &pal_idx[stride * (4 * h4 - 1)];
        for (int y = h4 * 4; y < bh4 * 4; y++)
            memcpy(&pal_idx[y * stride], src, bw4 * 4);
    }
}

static void read_vartx_tree(Dav1dTileContext *const t,
                            Av1Block *const b, const enum BlockSize bs,
                            const int bx4, const int by4)
{
    const Dav1dFrameContext *const f = t->f;
603
    const uint8_t *const b_dim = dav1d_block_dimensions[bs];
604
605
606
607
    const int bw4 = b_dim[0], bh4 = b_dim[1];

    // var-tx tree coding
    b->tx_split[0] = b->tx_split[1] = 0;
608
    b->max_ytx = dav1d_max_txfm_size_for_bs[bs][0];
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
    if (f->frame_hdr.segmentation.lossless[b->seg_id] ||
        b->max_ytx == TX_4X4)
    {
        b->max_ytx = b->uvtx = TX_4X4;
        if (f->frame_hdr.txfm_mode == TX_SWITCHABLE) {
            memset(&t->a->tx[bx4], TX_4X4, bw4);
            memset(&t->l.tx[by4], TX_4X4, bh4);
        }
    } else if (f->frame_hdr.txfm_mode != TX_SWITCHABLE || b->skip) {
        if (f->frame_hdr.txfm_mode == TX_SWITCHABLE) {
            memset(&t->a->tx[bx4], b_dim[2], bw4);
            memset(&t->l.tx[by4], b_dim[3], bh4);
        } else {
            assert(f->frame_hdr.txfm_mode == TX_LARGEST);
        }
624
        b->uvtx = dav1d_max_txfm_size_for_bs[bs][f->cur.p.p.layout];
625
626
627
    } else {
        assert(imin(bw4, bh4) <= 16 || b->max_ytx == TX_64X64);
        int y, x, y_off, x_off;
628
        const TxfmInfo *const ytx = &dav1d_txfm_dimensions[b->max_ytx];
629
630
631
632
633
634
635
636
637
638
639
640
641
        for (y = 0, y_off = 0; y < bh4; y += ytx->h, y_off++) {
            for (x = 0, x_off = 0; x < bw4; x += ytx->w, x_off++) {
                read_tx_tree(t, b->max_ytx, 0, b->tx_split, x_off, y_off);
                // contexts are updated inside read_tx_tree()
                t->bx += ytx->w;
            }
            t->bx -= x;
            t->by += ytx->h;
        }
        t->by -= y;
        if (DEBUG_BLOCK_INFO)
            printf("Post-vartxtree[%x/%x]: r=%d\n",
                   b->tx_split[0], b->tx_split[1], t->ts->msac.rng);
642
        b->uvtx = dav1d_max_txfm_size_for_bs[bs][f->cur.p.p.layout];
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
    }
}

static inline unsigned get_prev_frame_segid(const Dav1dFrameContext *const f,
                                            const int by, const int bx,
                                            const int w4, int h4,
                                            const uint8_t *ref_seg_map,
                                            const ptrdiff_t stride)
{
    unsigned seg_id = 8;

    assert(f->frame_hdr.primary_ref_frame != PRIMARY_REF_NONE);
    dav1d_thread_picture_wait(&f->refp[f->frame_hdr.primary_ref_frame],
                              (by + h4) * 4, PLANE_TYPE_BLOCK);

    ref_seg_map += by * stride + bx;
    do {
        for (int x = 0; x < w4; x++)
            seg_id = imin(seg_id, ref_seg_map[x]);
        ref_seg_map += stride;
    } while (--h4 > 0);
    assert(seg_id < 8);

    return seg_id;
}

static void decode_b(Dav1dTileContext *const t,
                     const enum BlockLevel bl,
                     const enum BlockSize bs,
                     const enum BlockPartition bp,
                     const enum EdgeFlags intra_edge_flags)
{
    Dav1dTileState *const ts = t->ts;
    const Dav1dFrameContext *const f = t->f;
    Av1Block b_mem, *const b = f->frame_thread.pass ?
        &f->frame_thread.b[t->by * f->b4_stride + t->bx] : &b_mem;
679
    const uint8_t *const b_dim = dav1d_block_dimensions[bs];
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
    const int bx4 = t->bx & 31, by4 = t->by & 31;
    const int ss_ver = f->cur.p.p.layout == DAV1D_PIXEL_LAYOUT_I420;
    const int ss_hor = f->cur.p.p.layout != DAV1D_PIXEL_LAYOUT_I444;
    const int cbx4 = bx4 >> ss_hor, cby4 = by4 >> ss_ver;
    const int bw4 = b_dim[0], bh4 = b_dim[1];
    const int w4 = imin(bw4, f->bw - t->bx), h4 = imin(bh4, f->bh - t->by);
    const int cbw4 = (bw4 + ss_hor) >> ss_hor, cbh4 = (bh4 + ss_ver) >> ss_ver;
    const int have_left = t->bx > ts->tiling.col_start;
    const int have_top = t->by > ts->tiling.row_start;
    const int has_chroma = f->seq_hdr.layout != DAV1D_PIXEL_LAYOUT_I400 &&
                           (bw4 > ss_hor || t->bx & 1) &&
                           (bh4 > ss_ver || t->by & 1);

    if (f->frame_thread.pass == 2) {
        if (b->intra) {
            f->bd_fn.recon_b_intra(t, bs, intra_edge_flags, b);

            if (has_chroma) {
                memset(&t->l.uvmode[cby4], b->uv_mode, cbh4);
                memset(&t->a->uvmode[cbx4], b->uv_mode, cbw4);
            }
            const enum IntraPredMode y_mode_nofilt =
                b->y_mode == FILTER_PRED ? DC_PRED : b->y_mode;
            memset(&t->l.mode[by4], y_mode_nofilt, bh4);
            memset(&t->a->mode[bx4], y_mode_nofilt, bw4);
        } else {
            if (b->comp_type == COMP_INTER_NONE && b->motion_mode == MM_WARP) {
                uint64_t mask[2] = { 0, 0 };
                find_matching_ref(t, intra_edge_flags, bw4, bh4, w4, h4,
                                  have_left, have_top, b->ref[0], mask);
                derive_warpmv(t, bw4, bh4, mask, b->mv[0], &t->warpmv);
            }
            f->bd_fn.recon_b_inter(t, bs, b);

714
            const uint8_t *const filter = dav1d_filter_dir[b->filter2d];
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
            memset(&t->l.filter[0][by4], filter[0], bh4);
            memset(&t->a->filter[0][bx4], filter[0], bw4);
            memset(&t->l.filter[1][by4], filter[1], bh4);
            memset(&t->a->filter[1][bx4], filter[1], bw4);
            if (has_chroma) {
                memset(&t->l.uvmode[cby4], DC_PRED, cbh4);
                memset(&t->a->uvmode[cbx4], DC_PRED, cbw4);
            }
        }
        memset(&t->l.intra[by4], b->intra, bh4);
        memset(&t->a->intra[bx4], b->intra, bw4);
        return;
    }

    const int cw4 = (w4 + ss_hor) >> ss_hor, ch4 = (h4 + ss_ver) >> ss_ver;

    b->bl = bl;
    b->bp = bp;
    b->bs = bs;

    // skip_mode
    if (f->frame_hdr.skip_mode_enabled && imin(bw4, bh4) > 1) {
        const int smctx = t->a->skip_mode[bx4] + t->l.skip_mode[by4];
        b->skip_mode = msac_decode_bool_adapt(&ts->msac,
                                              ts->cdf.m.skip_mode[smctx]);
        if (DEBUG_BLOCK_INFO)
            printf("Post-skipmode[%d]: r=%d\n", b->skip_mode, ts->msac.rng);
    } else {
        b->skip_mode = 0;
    }

    // segment_id (if seg_feature for skip/ref/gmv is enabled)
    int seg_pred = 0;
    if (f->frame_hdr.segmentation.enabled) {
        if (!f->frame_hdr.segmentation.update_map) {
            b->seg_id = f->prev_segmap ?
                        get_prev_frame_segid(f, t->by, t->bx, w4, h4,
                                             f->prev_segmap, f->b4_stride) : 0;
        } else if (f->frame_hdr.segmentation.seg_data.preskip) {
            if (f->frame_hdr.segmentation.temporal &&
                (seg_pred = msac_decode_bool_adapt(&ts->msac,
                                       ts->cdf.m.seg_pred[t->a->seg_pred[bx4] +
                                                          t->l.seg_pred[by4]])))
            {
                // temporal predicted seg_id
                b->seg_id = f->prev_segmap ?
                            get_prev_frame_segid(f, t->by, t->bx, w4, h4,
                                                 f->prev_segmap, f->b4_stride) : 0;
            } else {
                int seg_ctx;
                const unsigned pred_seg_id =
                    get_cur_frame_segid(t->by, t->bx, have_top, have_left,
                                        &seg_ctx, f->cur_segmap, f->b4_stride);
                const unsigned diff = msac_decode_symbol_adapt(&ts->msac,
                                                   ts->cdf.m.seg_id[seg_ctx],
                                                   NUM_SEGMENTS);
                const unsigned last_active_seg_id =
                    f->frame_hdr.segmentation.seg_data.last_active_segid;
773
774
                b->seg_id = neg_deinterleave(diff, pred_seg_id,
                                             last_active_seg_id + 1);
775
                if (b->seg_id > last_active_seg_id) b->seg_id = 0; // error?
776
                if (b->seg_id >= NUM_SEGMENTS) b->seg_id = 0; // error?
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
            }

            if (DEBUG_BLOCK_INFO)
                printf("Post-segid[preskip;%d]: r=%d\n",
                       b->seg_id, ts->msac.rng);
        }
    } else {
        b->seg_id = 0;
    }

    // skip
    const int sctx = t->a->skip[bx4] + t->l.skip[by4];
    b->skip = b->skip_mode ? 1 :
              msac_decode_bool_adapt(&ts->msac, ts->cdf.m.skip[sctx]);
    if (DEBUG_BLOCK_INFO)
        printf("Post-skip[%d]: r=%d\n", b->skip, ts->msac.rng);

    // segment_id
    if (f->frame_hdr.segmentation.enabled &&
        f->frame_hdr.segmentation.update_map &&
        !f->frame_hdr.segmentation.seg_data.preskip)
    {
        if (!b->skip && f->frame_hdr.segmentation.temporal &&
            (seg_pred = msac_decode_bool_adapt(&ts->msac,
                                   ts->cdf.m.seg_pred[t->a->seg_pred[bx4] +
                                                      t->l.seg_pred[by4]])))
        {
            // temporal predicted seg_id
            b->seg_id = f->prev_segmap ?
                        get_prev_frame_segid(f, t->by, t->bx, w4, h4,
                                             f->prev_segmap, f->b4_stride) : 0;
        } else {
            int seg_ctx;
            const unsigned pred_seg_id =
                get_cur_frame_segid(t->by, t->bx, have_top, have_left,
                                    &seg_ctx, f->cur_segmap, f->b4_stride);
            if (b->skip) {
                b->seg_id = pred_seg_id;
            } else {
                const unsigned diff = msac_decode_symbol_adapt(&ts->msac,
                                                   ts->cdf.m.seg_id[seg_ctx],
                                                   NUM_SEGMENTS);
                const unsigned last_active_seg_id =
                    f->frame_hdr.segmentation.seg_data.last_active_segid;
821
822
                b->seg_id = neg_deinterleave(diff, pred_seg_id,
                                             last_active_seg_id + 1);
823
824
                if (b->seg_id > last_active_seg_id) b->seg_id = 0; // error?
            }
825
            if (b->seg_id >= NUM_SEGMENTS) b->seg_id = 0; // error?
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
        }

        if (DEBUG_BLOCK_INFO)
            printf("Post-segid[postskip;%d]: r=%d\n",
                   b->seg_id, ts->msac.rng);
    }

    // cdef index
    if (!b->skip) {
        const int idx = f->seq_hdr.sb128 ? ((t->bx & 16) >> 4) +
                                           ((t->by & 16) >> 3) : 0;
        if (t->cur_sb_cdef_idx_ptr[idx] == -1) {
            const int v = msac_decode_bools(&ts->msac, f->frame_hdr.cdef.n_bits);
            t->cur_sb_cdef_idx_ptr[idx] = v;
            if (bw4 > 16) t->cur_sb_cdef_idx_ptr[idx + 1] = v;
            if (bh4 > 16) t->cur_sb_cdef_idx_ptr[idx + 2] = v;
            if (bw4 == 32 && bh4 == 32) t->cur_sb_cdef_idx_ptr[idx + 3] = v;

            if (DEBUG_BLOCK_INFO)
                printf("Post-cdef_idx[%d]: r=%d\n",
                        *t->cur_sb_cdef_idx_ptr, ts->msac.rng);
        }
    }

    // delta-q/lf
    if (!(t->bx & (31 >> !f->seq_hdr.sb128)) &&
        !(t->by & (31 >> !f->seq_hdr.sb128)))
    {
        const int prev_qidx = ts->last_qidx;
855
        const int have_delta_q = f->frame_hdr.delta.q.present &&
856
            (bs != (f->seq_hdr.sb128 ? BS_128x128 : BS_64x64) || !b->skip);
857
858
859
860

        int8_t prev_delta_lf[4];
        memcpy(prev_delta_lf, ts->last_delta_lf, 4);

861
862
863
864
865
866
867
868
        if (have_delta_q) {
            int delta_q = msac_decode_symbol_adapt(&ts->msac, ts->cdf.m.delta_q, 4);
            if (delta_q == 3) {
                const int n_bits = 1 + msac_decode_bools(&ts->msac, 3);
                delta_q = msac_decode_bools(&ts->msac, n_bits) + 1 + (1 << n_bits);
            }
            if (delta_q) {
                if (msac_decode_bool(&ts->msac, 128 << 7)) delta_q = -delta_q;
869
                delta_q *= 1 << f->frame_hdr.delta.q.res_log2;
870
871
872
873
874
            }
            ts->last_qidx = iclip(ts->last_qidx + delta_q, 1, 255);
            if (have_delta_q && DEBUG_BLOCK_INFO)
                printf("Post-delta_q[%d->%d]: r=%d\n",
                       delta_q, ts->last_qidx, ts->msac.rng);
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900

            if (f->frame_hdr.delta.lf.present) {
                const int n_lfs = f->frame_hdr.delta.lf.multi ?
                    f->seq_hdr.layout != DAV1D_PIXEL_LAYOUT_I400 ? 4 : 2 : 1;

                for (int i = 0; i < n_lfs; i++) {
                    int delta_lf =
                        msac_decode_symbol_adapt(&ts->msac,
                        ts->cdf.m.delta_lf[i + f->frame_hdr.delta.lf.multi], 4);
                    if (delta_lf == 3) {
                        const int n_bits = 1 + msac_decode_bools(&ts->msac, 3);
                        delta_lf = msac_decode_bools(&ts->msac, n_bits) +
                                   1 + (1 << n_bits);
                    }
                    if (delta_lf) {
                        if (msac_decode_bool(&ts->msac, 128 << 7))
                            delta_lf = -delta_lf;
                        delta_lf *= 1 << f->frame_hdr.delta.lf.res_log2;
                    }
                    ts->last_delta_lf[i] =
                        iclip(ts->last_delta_lf[i] + delta_lf, -63, 63);
                    if (have_delta_q && DEBUG_BLOCK_INFO)
                        printf("Post-delta_lf[%d:%d]: r=%d\n", i, delta_lf,
                               ts->msac.rng);
                }
            }
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
        }
        if (ts->last_qidx == f->frame_hdr.quant.yac) {
            // assign frame-wide q values to this sb
            ts->dq = f->dq;
        } else if (ts->last_qidx != prev_qidx) {
            // find sb-specific quant parameters
            init_quant_tables(&f->seq_hdr, &f->frame_hdr, ts->last_qidx, ts->dqmem);
            ts->dq = ts->dqmem;
        }
        if (!memcmp(ts->last_delta_lf, (int8_t[4]) { 0, 0, 0, 0 }, 4)) {
            // assign frame-wide lf values to this sb
            ts->lflvl = f->lf.lvl;
        } else if (memcmp(ts->last_delta_lf, prev_delta_lf, 4)) {
            // find sb-specific lf lvl parameters
            dav1d_calc_lf_values(ts->lflvlmem, &f->frame_hdr, ts->last_delta_lf);
            ts->lflvl = ts->lflvlmem;
        }
    }

    if (b->skip_mode) {
        b->intra = 0;
    } else if (f->frame_hdr.frame_type & 1) {
        const int ictx = get_intra_ctx(t->a, &t->l, by4, bx4,
                                       have_top, have_left);
        b->intra = !msac_decode_bool_adapt(&ts->msac, ts->cdf.m.intra[ictx]);
        if (DEBUG_BLOCK_INFO)
            printf("Post-intra[%d]: r=%d\n", b->intra, ts->msac.rng);
    } else if (f->frame_hdr.allow_intrabc) {
        b->intra = !msac_decode_bool_adapt(&ts->msac, ts->cdf.m.intrabc);
        if (DEBUG_BLOCK_INFO)
            printf("Post-intrabcflag[%d]: r=%d\n", b->intra, ts->msac.rng);
    } else {
        b->intra = 1;
    }

    // intra/inter-specific stuff
    if (b->intra) {
        uint16_t *const ymode_cdf = f->frame_hdr.frame_type & 1 ?
939
940
941
            ts->cdf.m.y_mode[dav1d_ymode_size_context[bs]] :
            ts->cdf.kfym[dav1d_intra_mode_context[t->a->mode[bx4]]]
                        [dav1d_intra_mode_context[t->l.mode[by4]]];
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
        b->y_mode = msac_decode_symbol_adapt(&ts->msac, ymode_cdf,
                                              N_INTRA_PRED_MODES);
        if (DEBUG_BLOCK_INFO)
            printf("Post-ymode[%d]: r=%d\n", b->y_mode, ts->msac.rng);

        // angle delta
        if (b_dim[2] + b_dim[3] >= 2 && b->y_mode >= VERT_PRED &&
            b->y_mode <= VERT_LEFT_PRED)
        {
            uint16_t *const acdf = ts->cdf.m.angle_delta[b->y_mode - VERT_PRED];
            const int angle = msac_decode_symbol_adapt(&ts->msac, acdf, 7);
            b->y_angle = angle - 3;
        } else {
            b->y_angle = 0;
        }

        if (has_chroma) {
959
960
            const int cfl_allowed = f->frame_hdr.segmentation.lossless[b->seg_id] ?
                cbw4 == 1 && cbh4 == 1 : !!(cfl_allowed_mask & (1 << bs));
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
            uint16_t *const uvmode_cdf = ts->cdf.m.uv_mode[cfl_allowed][b->y_mode];
            b->uv_mode = msac_decode_symbol_adapt(&ts->msac, uvmode_cdf,
                                         N_UV_INTRA_PRED_MODES - !cfl_allowed);
            if (DEBUG_BLOCK_INFO)
                printf("Post-uvmode[%d]: r=%d\n", b->uv_mode, ts->msac.rng);

            if (b->uv_mode == CFL_PRED) {
#define SIGN(a) (!!(a) + ((a) > 0))
                const int sign =
                    msac_decode_symbol_adapt(&ts->msac, ts->cdf.m.cfl_sign, 8) + 1;
                const int sign_u = sign * 0x56 >> 8, sign_v = sign - sign_u * 3;
                assert(sign_u == sign / 3);
                if (sign_u) {
                    const int ctx = (sign_u == 2) * 3 + sign_v;
                    b->cfl_alpha[0] = msac_decode_symbol_adapt(&ts->msac,
                                            ts->cdf.m.cfl_alpha[ctx], 16) + 1;
                    if (sign_u == 1) b->cfl_alpha[0] = -b->cfl_alpha[0];
                } else {
                    b->cfl_alpha[0] = 0;
                }
                if (sign_v) {
                    const int ctx = (sign_v == 2) * 3 + sign_u;
                    b->cfl_alpha[1] = msac_decode_symbol_adapt(&ts->msac,
                                            ts->cdf.m.cfl_alpha[ctx], 16) + 1;
                    if (sign_v == 1) b->cfl_alpha[1] = -b->cfl_alpha[1];
                } else {
                    b->cfl_alpha[1] = 0;
                }
#undef SIGN
                if (DEBUG_BLOCK_INFO)
                    printf("Post-uvalphas[%d/%d]: r=%d\n",
                           b->cfl_alpha[0], b->cfl_alpha[1], ts->msac.rng);
            } else if (b_dim[2] + b_dim[3] >= 2 && b->uv_mode >= VERT_PRED &&
                       b->uv_mode <= VERT_LEFT_PRED)
            {
                uint16_t *const acdf = ts->cdf.m.angle_delta[b->uv_mode - VERT_PRED];
                const int angle = msac_decode_symbol_adapt(&ts->msac, acdf, 7);
                b->uv_angle = angle - 3;
            } else {
                b->uv_angle = 0;