threers-codec 0.0.10

Pure-Rust media codecs: H.264/HEVC/VP9 encoders, H.264/HEVC/VP9/AV1 decoders, AAC/MP3/Opus/Vorbis/FLAC/WAV decoders, MP4/WebM/Matroska/Ogg, GIF/APNG, captions, and VideoToolbox acceleration on macOS
Documentation
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
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
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
217
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
299
300
301
302
303
304
305
306
307
308
309
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
345
346
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
452
453
454
455
456
457
458
459
460
461
462
463
464
465
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
516
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
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
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
679
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
714
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
//! Spatial and temporal motion-vector candidates (dav1d `src/refmvs.c`).
//!
//! dav1d keeps 35 rotating rows of 4x4 block info per superblock row; the
//! reads only ever reach the current superblock row and the odd rows -1,
//! -3, -5 above it, so a whole-frame array is equivalent and is used here.

use std::sync::Arc;

use super::env::{fix_mv_precision, get_gmv_2d, Mv};
use super::intra_edge::EDGE_I444_TOP_HAS_RIGHT;
use super::obu::{get_poc_diff, FrameHeader, SeqHeader, WM_TRANSLATION};
use super::tables::BLOCK_DIMENSIONS;

#[derive(Clone, Copy, Default, PartialEq, Eq, Debug)]
pub struct TemporalBlock {
    pub mv: Mv,
    pub r: i8,
}

#[derive(Clone, Copy, PartialEq, Eq, Debug)]
pub struct RefmvsBlock {
    pub mv: [Mv; 2],
    /// `[0]` = 0: intra; `[1]` = -1: not compound.
    pub r: [i8; 2],
    pub bs: u8,
    /// 1 = globalmv+affine, 2 = newmv.
    pub mf: u8,
}

impl Default for RefmvsBlock {
    fn default() -> Self {
        RefmvsBlock { mv: [Mv::ZERO; 2], r: [0, -1], bs: 0, mf: 0 }
    }
}

#[derive(Clone, Copy, Default, Debug)]
pub struct RefmvsCandidate {
    pub mv: [Mv; 2],
    pub weight: i32,
}

pub struct RefmvsFrame {
    pub iw4: i32,
    pub ih4: i32,
    pub iw8: i32,
    pub ih8: i32,
    pub use_ref_frame_mvs: bool,
    pub sign_bias: [u8; 7],
    pub mfmv_sign: [u8; 7],
    pub pocdiff: [i8; 7],
    pub mfmv_ref: [u8; 3],
    pub mfmv_ref2cur: [i32; 3],
    pub mfmv_ref2ref: [[i32; 7]; 3],
    pub n_mfmvs: usize,
    /// This frame's saved temporal MVs (8x8), `rp_stride` wide.
    pub rp: Vec<TemporalBlock>,
    pub rp_ref: [Option<Arc<Vec<TemporalBlock>>>; 7],
    /// Projected temporal MVs for the current frame, whole frame.
    pub rp_proj: Vec<TemporalBlock>,
    pub rp_stride: usize,
    /// 4x4 block info, whole frame (`r_stride` wide).
    pub r: Vec<RefmvsBlock>,
    pub r_stride: usize,
}

pub struct RefmvsTile {
    pub col_start: i32,
    pub col_end: i32,
    pub row_start: i32,
    pub row_end: i32,
}

impl RefmvsFrame {
    /// `dav1d_refmvs_init_frame`.
    pub fn new(seq: &SeqHeader, hdr: &FrameHeader, ref_poc: &[u32; 7], ref_ref_poc: &[[u32; 7]; 7], rp_ref: [Option<Arc<Vec<TemporalBlock>>>; 7]) -> Self {
        let rp_stride = (((hdr.width[0] + 127) & !127) >> 3) as usize;
        let iw8 = (hdr.width[0] + 7) >> 3;
        let ih8 = (hdr.height + 7) >> 3;
        let sb = 16 << seq.sb128 as i32;
        // Rows to the superblock-aligned height (blocks may overhang).
        let r_rows = (((ih8 * 2 + sb - 1) / sb) * sb) as usize;
        let mut rf = RefmvsFrame {
            iw4: iw8 << 1,
            ih4: ih8 << 1,
            iw8,
            ih8,
            use_ref_frame_mvs: false,
            sign_bias: [0; 7],
            mfmv_sign: [0; 7],
            pocdiff: [0; 7],
            mfmv_ref: [0; 3],
            mfmv_ref2cur: [0; 3],
            mfmv_ref2ref: [[0; 7]; 3],
            n_mfmvs: 0,
            rp: vec![TemporalBlock::default(); rp_stride * (r_rows / 2 + 1)],
            rp_ref,
            rp_proj: vec![TemporalBlock::default(); rp_stride * (r_rows / 2 + 1)],
            rp_stride,
            r: vec![RefmvsBlock::default(); rp_stride * 2 * r_rows],
            r_stride: rp_stride * 2,
        };
        let nb = seq.order_hint_n_bits;
        let poc = hdr.frame_offset as i32;
        for i in 0..7 {
            let d = get_poc_diff(nb, ref_poc[i] as i32, poc);
            rf.sign_bias[i] = (d > 0) as u8;
            rf.mfmv_sign[i] = (d < 0) as u8;
            rf.pocdiff[i] = get_poc_diff(nb, poc, ref_poc[i] as i32).clamp(-31, 31) as i8;
        }
        if hdr.use_ref_frame_mvs && nb != 0 {
            let mut total = 2;
            let rr = &rf.rp_ref;
            let mut refs = Vec::new();
            if rr[0].is_some() && ref_ref_poc[0][6] != ref_poc[3] {
                refs.push(0u8);
                total = 3;
            }
            if rr[4].is_some() && get_poc_diff(nb, ref_poc[4] as i32, poc) > 0 {
                refs.push(4);
            }
            if rr[5].is_some() && get_poc_diff(nb, ref_poc[5] as i32, poc) > 0 {
                refs.push(5);
            }
            if refs.len() < total && rr[6].is_some() && get_poc_diff(nb, ref_poc[6] as i32, poc) > 0 {
                refs.push(6);
            }
            if refs.len() < total && rr[1].is_some() {
                refs.push(1);
            }
            rf.n_mfmvs = refs.len();
            for (n, &m) in refs.iter().enumerate() {
                rf.mfmv_ref[n] = m;
                let rpoc = ref_poc[m as usize] as i32;
                let diff1 = get_poc_diff(nb, rpoc, poc);
                if diff1.abs() > 31 {
                    rf.mfmv_ref2cur[n] = i32::MIN;
                } else {
                    rf.mfmv_ref2cur[n] = if m < 4 { -diff1 } else { diff1 };
                    for k in 0..7 {
                        let rrpoc = ref_ref_poc[m as usize][k] as i32;
                        let diff2 = get_poc_diff(nb, rpoc, rrpoc);
                        rf.mfmv_ref2ref[n][k] = if diff2 as u32 > 31 { 0 } else { diff2 };
                    }
                }
            }
        }
        rf.use_ref_frame_mvs = rf.n_mfmvs > 0;
        rf
    }

    #[inline]
    pub fn blk(&self, y4: i32, x4: i32) -> &RefmvsBlock {
        &self.r[y4 as usize * self.r_stride + x4 as usize]
    }

    /// `splat_mv`.
    pub fn splat(&mut self, b: &RefmvsBlock, by4: i32, bx4: i32, bw4: i32, bh4: i32) {
        for y in 0..bh4 {
            let o = (by4 + y) as usize * self.r_stride + bx4 as usize;
            for v in &mut self.r[o..o + bw4 as usize] {
                *v = *b;
            }
        }
    }

    /// `load_tmvs_c` (rows of 8x8 blocks `row_start8..row_end8`).
    pub fn load_tmvs(&mut self, col_start8: i32, col_end8: i32, row_start8: i32, row_end8: i32) {
        let row_end8 = row_end8.min(self.ih8);
        let col_start8i = (col_start8 - 8).max(0);
        let col_end8i = (col_end8 + 8).min(self.iw8);
        let stride = self.rp_stride;
        for y in row_start8..row_end8 {
            for x in col_start8..col_end8 {
                self.rp_proj[y as usize * stride + x as usize].mv = Mv::INVALID;
            }
        }
        for n in 0..self.n_mfmvs {
            let ref2cur = self.mfmv_ref2cur[n];
            if ref2cur == i32::MIN {
                continue;
            }
            let r = self.mfmv_ref[n] as usize;
            let ref_sign = r as i32 - 4;
            let rp_ref = self.rp_ref[r].clone().unwrap();
            for y in row_start8..row_end8 {
                let y_sb_align = y & !7;
                let y_proj_start = y_sb_align.max(row_start8);
                let y_proj_end = (y_sb_align + 8).min(row_end8);
                let row = &rp_ref[y as usize * stride..];
                let mut x = col_start8i;
                while x < col_end8i {
                    let mut rb = x as usize;
                    let b_ref = row[rb].r;
                    if b_ref == 0 {
                        x += 1;
                        continue;
                    }
                    let ref2ref = self.mfmv_ref2ref[n][b_ref as usize - 1];
                    if ref2ref == 0 {
                        x += 1;
                        continue;
                    }
                    let b_mv = row[rb].mv;
                    let off = mv_projection(b_mv, ref2cur, ref2ref);
                    let sgn = |v: i32, s: i32| if s < 0 { -v } else { v };
                    let mut pos_x = x + sgn((off.x as i32).abs() >> 6, off.x as i32 ^ ref_sign);
                    let pos_y = y + sgn((off.y as i32).abs() >> 6, off.y as i32 ^ ref_sign);
                    if pos_y >= y_proj_start && pos_y < y_proj_end {
                        let pos = pos_y as usize * stride;
                        loop {
                            let x_sb_align = x & !7;
                            if pos_x >= (x_sb_align - 8).max(col_start8) && pos_x < (x_sb_align + 16).min(col_end8) {
                                let t = &mut self.rp_proj[pos + pos_x as usize];
                                t.mv = row[rb].mv;
                                t.r = ref2ref as i8;
                            }
                            x += 1;
                            if x >= col_end8i {
                                break;
                            }
                            rb += 1;
                            if row[rb].r != b_ref || row[rb].mv != b_mv {
                                break;
                            }
                            pos_x += 1;
                        }
                    } else {
                        loop {
                            x += 1;
                            if x >= col_end8i {
                                break;
                            }
                            rb += 1;
                            if row[rb].r != b_ref || row[rb].mv != b_mv {
                                break;
                            }
                        }
                    }
                }
            }
        }
    }

    /// `save_tmvs_c` for 8x8 rows `row_start8..row_end8`, clipped to the frame.
    pub fn save_tmvs(&mut self, col_start8: i32, col_end8: i32, row_start8: i32, row_end8: i32) {
        let row_end8 = row_end8.min(self.ih8);
        let col_end8 = col_end8.min(self.iw8);
        for y in row_start8..row_end8 {
            let b = (y as usize * 2 + 1) * self.r_stride;
            let rp = y as usize * self.rp_stride;
            let mut x = col_start8;
            while x < col_end8 {
                let c = self.r[b + x as usize * 2 + 1];
                let bw8 = (BLOCK_DIMENSIONS[c.bs as usize][0] as i32 + 1) >> 1;
                let ok = |k: usize| {
                    c.r[k] > 0 && self.mfmv_sign[c.r[k] as usize - 1] != 0 && ((c.mv[k].y as i32).abs() | (c.mv[k].x as i32).abs()) < 4096
                };
                let t = if ok(1) {
                    TemporalBlock { mv: c.mv[1], r: c.r[1] }
                } else if ok(0) {
                    TemporalBlock { mv: c.mv[0], r: c.r[0] }
                } else {
                    TemporalBlock { mv: Mv::ZERO, r: 0 }
                };
                for _ in 0..bw8 {
                    self.rp[rp + x as usize] = t;
                    x += 1;
                }
            }
        }
    }
}

fn mv_projection(mv: Mv, num: i32, den: i32) -> Mv {
    const DIV_MULT: [u16; 32] = [
        0, 16384, 8192, 5461, 4096, 3276, 2730, 2340, 2048, 1820, 1638, 1489, 1365, 1260, 1170, 1092, 1024, 963, 910, 862, 819, 780,
        744, 712, 682, 655, 630, 606, 585, 564, 546, 528,
    ];
    let frac = num * DIV_MULT[den as usize] as i32;
    let y = mv.y as i32 * frac;
    let x = mv.x as i32 * frac;
    Mv { y: ((y + 8192 + (y >> 31)) >> 14).clamp(-0x3fff, 0x3fff) as i16, x: ((x + 8192 + (x >> 31)) >> 14).clamp(-0x3fff, 0x3fff) as i16 }
}

struct Find<'a> {
    stack: [RefmvsCandidate; 9],
    cnt: usize,
    r: [i8; 2],
    gmv: [Mv; 2],
    have_newmv: bool,
    rf: &'a RefmvsFrame,
}

impl Find<'_> {
    fn add_spatial(&mut self, weight: i32, b: &RefmvsBlock, row_match: &mut bool, newmv: bool) {
        if b.mv[0] == Mv::INVALID {
            return;
        }
        let gmv = self.gmv;
        if self.r[1] == -1 {
            for n in 0..2 {
                if b.r[n] == self.r[0] {
                    let cand = if b.mf & 1 != 0 && gmv[0] != Mv::INVALID { gmv[0] } else { b.mv[n] };
                    *row_match = true;
                    if newmv {
                        self.have_newmv |= b.mf >> 1 != 0;
                    }
                    let last = self.cnt;
                    for m in 0..last {
                        if self.stack[m].mv[0] == cand {
                            self.stack[m].weight += weight;
                            return;
                        }
                    }
                    if last < 8 {
                        self.stack[last].mv[0] = cand;
                        self.stack[last].weight = weight;
                        self.cnt = last + 1;
                    }
                    return;
                }
            }
        } else if b.r == self.r {
            let cand = [
                if b.mf & 1 != 0 && gmv[0] != Mv::INVALID { gmv[0] } else { b.mv[0] },
                if b.mf & 1 != 0 && gmv[1] != Mv::INVALID { gmv[1] } else { b.mv[1] },
            ];
            *row_match = true;
            if newmv {
                self.have_newmv |= b.mf >> 1 != 0;
            }
            let last = self.cnt;
            for n in 0..last {
                if self.stack[n].mv == cand {
                    self.stack[n].weight += weight;
                    return;
                }
            }
            if last < 8 {
                self.stack[last].mv = cand;
                self.stack[last].weight = weight;
                self.cnt = last + 1;
            }
        }
    }

    #[allow(clippy::too_many_arguments)]
    fn scan_row(&mut self, y4: i32, x4: i32, bw4: i32, w4: i32, max_rows: i32, step: i32, m: &mut bool, newmv: bool) -> u32 {
        let rf = self.rf;
        let first = *rf.blk(y4, x4);
        let fdim = BLOCK_DIMENSIONS[first.bs as usize];
        let cand_bw4 = fdim[0] as i32;
        let mut len = step.max(bw4.min(cand_bw4));
        if bw4 <= cand_bw4 {
            let weight = if bw4 == 1 { 2 } else { 2.max((2 * max_rows).min(fdim[1] as i32)) };
            self.add_spatial(len * weight, &first, m, newmv);
            return (weight >> 1) as u32;
        }
        let mut x = 0;
        let mut cand = first;
        loop {
            self.add_spatial(len * 2, &cand, m, newmv);
            x += len;
            if x >= w4 {
                return 1;
            }
            cand = *rf.blk(y4, x4 + x);
            len = step.max(BLOCK_DIMENSIONS[cand.bs as usize][0] as i32);
        }
    }

    #[allow(clippy::too_many_arguments)]
    fn scan_col(&mut self, y4: i32, x4: i32, bh4: i32, h4: i32, max_cols: i32, step: i32, m: &mut bool, newmv: bool) -> u32 {
        let rf = self.rf;
        let first = *rf.blk(y4, x4);
        let fdim = BLOCK_DIMENSIONS[first.bs as usize];
        let cand_bh4 = fdim[1] as i32;
        let mut len = step.max(bh4.min(cand_bh4));
        if bh4 <= cand_bh4 {
            let weight = if bh4 == 1 { 2 } else { 2.max((2 * max_cols).min(fdim[0] as i32)) };
            self.add_spatial(len * weight, &first, m, newmv);
            return (weight >> 1) as u32;
        }
        let mut y = 0;
        let mut cand = first;
        loop {
            self.add_spatial(len * 2, &cand, m, newmv);
            y += len;
            if y >= h4 {
                return 1;
            }
            cand = *rf.blk(y4 + y, x4);
            len = step.max(BLOCK_DIMENSIONS[cand.bs as usize][1] as i32);
        }
    }

    fn add_temporal(&mut self, hdr: &FrameHeader, rb: &TemporalBlock, globalmv_ctx: Option<&mut i32>, tgmv: &[Mv; 2]) {
        if rb.mv == Mv::INVALID {
            return;
        }
        let rf = self.rf;
        let mut mv = mv_projection(rb.mv, rf.pocdiff[self.r[0] as usize - 1] as i32, rb.r as i32);
        fix_mv_precision(hdr, &mut mv);
        let last = self.cnt;
        if self.r[1] == -1 {
            if let Some(g) = globalmv_ctx {
                *g = (((mv.x as i32 - tgmv[0].x as i32).abs() | (mv.y as i32 - tgmv[0].y as i32).abs()) >= 16) as i32;
            }
            for n in 0..last {
                if self.stack[n].mv[0] == mv {
                    self.stack[n].weight += 2;
                    return;
                }
            }
            if last < 8 {
                self.stack[last].mv[0] = mv;
                self.stack[last].weight = 2;
                self.cnt = last + 1;
            }
        } else {
            let mut mv1 = mv_projection(rb.mv, rf.pocdiff[self.r[1] as usize - 1] as i32, rb.r as i32);
            fix_mv_precision(hdr, &mut mv1);
            let p = [mv, mv1];
            for n in 0..last {
                if self.stack[n].mv == p {
                    self.stack[n].weight += 2;
                    return;
                }
            }
            if last < 8 {
                self.stack[last].mv = p;
                self.stack[last].weight = 2;
                self.cnt = last + 1;
            }
        }
    }
}

fn neg(m: Mv) -> Mv {
    Mv { y: m.y.wrapping_neg(), x: m.x.wrapping_neg() }
}

/// `add_compound_extended_candidate`: `same` is stack[base..base+2],
/// `diff` stack[base+2..base+4].
fn add_compound_extended(st: &mut [RefmvsCandidate; 9], base: usize, cnt: &mut [usize; 4], c: &RefmvsBlock, s0: u8, s1: u8, r: [i8; 2], sign_bias: &[u8; 7]) {
    for n in 0..2 {
        let cr = c.r[n];
        if cr <= 0 {
            break;
        }
        let mut m = c.mv[n];
        let sb = sign_bias[cr as usize - 1];
        if cr == r[0] {
            if cnt[0] < 2 {
                st[base + cnt[0]].mv[0] = m;
                cnt[0] += 1;
            }
            if cnt[3] < 2 {
                if s1 ^ sb != 0 {
                    m = neg(m);
                }
                st[base + 2 + cnt[3]].mv[1] = m;
                cnt[3] += 1;
            }
        } else if cr == r[1] {
            if cnt[1] < 2 {
                st[base + cnt[1]].mv[1] = m;
                cnt[1] += 1;
            }
            if cnt[2] < 2 {
                if s0 ^ sb != 0 {
                    m = neg(m);
                }
                st[base + 2 + cnt[2]].mv[0] = m;
                cnt[2] += 1;
            }
        } else {
            let im = neg(m);
            if cnt[2] < 2 {
                st[base + 2 + cnt[2]].mv[0] = if s0 ^ sb != 0 { im } else { m };
                cnt[2] += 1;
            }
            if cnt[3] < 2 {
                st[base + 2 + cnt[3]].mv[1] = if s1 ^ sb != 0 { im } else { m };
                cnt[3] += 1;
            }
        }
    }
}

fn add_single_extended(st: &mut [RefmvsCandidate; 9], cnt: &mut usize, c: &RefmvsBlock, sign: u8, sign_bias: &[u8; 7]) {
    for n in 0..2 {
        let cr = c.r[n];
        if cr <= 0 {
            break;
        }
        let mut m = c.mv[n];
        if sign ^ sign_bias[cr as usize - 1] != 0 {
            m = neg(m);
        }
        let last = *cnt;
        if !(0..last).any(|k| st[k].mv[0] == m) {
            st[last].mv[0] = m;
            st[last].weight = 2;
            *cnt = last + 1;
        }
    }
}

/// `dav1d_refmvs_find`: (candidate stack, count, context).
#[allow(clippy::too_many_arguments)]
pub fn find(rf: &RefmvsFrame, hdr: &FrameHeader, rt: &RefmvsTile, r: [i8; 2], bs: usize, edge_flags: u8, by4: i32, bx4: i32) -> ([RefmvsCandidate; 8], usize, i32) {
    let b_dim = BLOCK_DIMENSIONS[bs];
    let bw4 = b_dim[0] as i32;
    let w4 = bw4.min(16).min(rt.col_end - bx4);
    let bh4 = b_dim[1] as i32;
    let h4 = bh4.min(16).min(rt.row_end - by4);
    let mut gmv = [Mv::INVALID; 2];
    let mut tgmv = [Mv::ZERO; 2];
    if r[0] > 0 {
        let g = &hdr.gmv[r[0] as usize - 1];
        tgmv[0] = get_gmv_2d(g, bx4, by4, bw4, bh4, hdr);
        gmv[0] = if g.typ > WM_TRANSLATION { tgmv[0] } else { Mv::INVALID };
    }
    if r[1] > 0 {
        let g = &hdr.gmv[r[1] as usize - 1];
        tgmv[1] = get_gmv_2d(g, bx4, by4, bw4, bh4, hdr);
        gmv[1] = if g.typ > WM_TRANSLATION { tgmv[1] } else { Mv::INVALID };
    }
    let mut f = Find { stack: [RefmvsCandidate::default(); 9], cnt: 0, r, gmv, have_newmv: false, rf };

    let (mut have_col_mvs, mut have_row_mvs) = (false, false);
    let mut max_rows = 0;
    let mut n_rows = !0u32;
    if by4 > rt.row_start {
        max_rows = ((by4 - rt.row_start + 1) >> 1).min(2 + (bh4 > 1) as i32);
        n_rows = f.scan_row(by4 - 1, bx4, bw4, w4, max_rows, if bw4 >= 16 { 4 } else { 1 }, &mut have_row_mvs, true);
    }
    let mut max_cols = 0;
    let mut n_cols = !0u32;
    if bx4 > rt.col_start {
        max_cols = ((bx4 - rt.col_start + 1) >> 1).min(2 + (bw4 > 1) as i32);
        n_cols = f.scan_col(by4, bx4 - 1, bh4, h4, max_cols, if bh4 >= 16 { 4 } else { 1 }, &mut have_col_mvs, true);
    }
    if n_rows != !0 && edge_flags & EDGE_I444_TOP_HAS_RIGHT != 0 && bw4.max(bh4) <= 16 && bw4 + bx4 < rt.col_end {
        let b = *rf.blk(by4 - 1, bx4 + bw4);
        f.add_spatial(4, &b, &mut have_row_mvs, true);
    }
    let nearest_match = have_col_mvs as i32 + have_row_mvs as i32;
    let nearest_cnt = f.cnt;
    for c in f.stack.iter_mut().take(nearest_cnt) {
        c.weight += 640;
    }

    // Temporal.
    let mut globalmv_ctx = hdr.use_ref_frame_mvs as i32;
    if rf.use_ref_frame_mvs {
        let stride = rf.rp_stride as i32;
        let (by8, bx8) = (by4 >> 1, bx4 >> 1);
        let rbi = by8 * stride + bx8;
        let step_h = if bw4 >= 16 { 2 } else { 1 };
        let step_v = if bh4 >= 16 { 2 } else { 1 };
        let w8 = ((w4 + 1) >> 1).min(8);
        let h8 = ((h4 + 1) >> 1).min(8);
        let mut y = 0;
        while y < h8 {
            let mut x = 0;
            while x < w8 {
                let rb = rf.rp_proj[(rbi + y * stride + x) as usize];
                if x == 0 && y == 0 {
                    f.add_temporal(hdr, &rb, Some(&mut globalmv_ctx), &tgmv);
                } else {
                    f.add_temporal(hdr, &rb, None, &tgmv);
                }
                x += step_h;
            }
            y += step_v;
        }
        if bw4.min(bh4) >= 2 && bw4.max(bh4) < 16 {
            let (bh8, bw8) = (bh4 >> 1, bw4 >> 1);
            let rb = rbi + bh8 * stride;
            let has_bottom = by8 + bh8 < (rt.row_end >> 1).min((by8 & !7) + 8);
            if has_bottom && bx8 > (rt.col_start >> 1).max(bx8 & !7) {
                let t = rf.rp_proj[(rb - 1) as usize];
                f.add_temporal(hdr, &t, None, &tgmv);
            }
            if bx8 + bw8 < (rt.col_end >> 1).min((bx8 & !7) + 8) {
                if has_bottom {
                    let t = rf.rp_proj[(rb + bw8) as usize];
                    f.add_temporal(hdr, &t, None, &tgmv);
                }
                if by8 + bh8 - 1 < (rt.row_end >> 1).min((by8 & !7) + 8) {
                    let t = rf.rp_proj[(rb + bw8 - stride) as usize];
                    f.add_temporal(hdr, &t, None, &tgmv);
                }
            }
        }
    }

    // Top/left.
    let mut dummy = false;
    if n_rows != !0 && n_cols != !0 {
        let b = *rf.blk(by4 - 1, bx4 - 1);
        f.add_spatial(4, &b, &mut have_row_mvs, false);
    }
    let _ = &mut dummy;
    // Secondary top and left edges (8x8 resolution).
    for n in 2..=3u32 {
        if n > n_rows && n as i32 <= max_rows {
            let y4 = by4 - (by4 & 31) + (((by4 & 31) - 2 * n as i32 + 1) | 1);
            n_rows = n_rows.wrapping_add(f.scan_row(y4, bx4 | 1, bw4, w4, 1 + max_rows - n as i32, if bw4 >= 16 { 4 } else { 2 }, &mut have_row_mvs, false));
        }
        if n > n_cols && n as i32 <= max_cols {
            let y4 = by4 - (by4 & 31) + ((by4 & 31) | 1);
            n_cols = n_cols.wrapping_add(f.scan_col(y4, (bx4 - n as i32 * 2 + 1) | 1, bh4, h4, 1 + max_cols - n as i32, if bh4 >= 16 { 4 } else { 2 }, &mut have_col_mvs, false));
        }
    }
    let ref_match_count = have_col_mvs as i32 + have_row_mvs as i32;
    let have_newmv = f.have_newmv as i32;
    let (refmv_ctx, newmv_ctx) = match nearest_match {
        0 => (ref_match_count.min(2), (ref_match_count > 0) as i32),
        1 => ((ref_match_count * 3).min(4), 3 - have_newmv),
        _ => (5, 5 - have_newmv),
    };

    // Sort (nearest, then secondary).
    let mut len = nearest_cnt;
    while len > 0 {
        let mut last = 0;
        for n in 1..len {
            if f.stack[n - 1].weight < f.stack[n].weight {
                f.stack.swap(n - 1, n);
                last = n;
            }
        }
        len = last;
    }
    len = f.cnt;
    while len > nearest_cnt {
        let mut last = nearest_cnt;
        for n in nearest_cnt + 1..len {
            if f.stack[n - 1].weight < f.stack[n].weight {
                f.stack.swap(n - 1, n);
                last = n;
            }
        }
        len = last;
    }

    let mut cnt = f.cnt;
    let mut st = f.stack;
    let clamp_bounds = (
        -(bx4 + bw4 + 4) * 4 * 8,
        (rf.iw4 - bx4 + 4) * 4 * 8,
        -(by4 + bh4 + 4) * 4 * 8,
        (rf.ih4 - by4 + 4) * 4 * 8,
    );
    let clampmv = |m: &mut Mv| {
        let (l, rr, t, b) = clamp_bounds;
        m.x = (m.x as i32).clamp(l, rr) as i16;
        m.y = (m.y as i32).clamp(t, b) as i16;
    };
    let ctx = if r[1] > 0 {
        if cnt < 2 {
            let s0 = rf.sign_bias[r[0] as usize - 1];
            let s1 = rf.sign_bias[r[1] as usize - 1];
            let sz4 = w4.min(h4);
            let base = cnt;
            let mut sc = [0usize; 4];
            if n_rows != !0 {
                let mut x = 0;
                while x < sz4 {
                    let c = *rf.blk(by4 - 1, bx4 + x);
                    add_compound_extended(&mut st, base, &mut sc, &c, s0, s1, r, &rf.sign_bias);
                    x += BLOCK_DIMENSIONS[c.bs as usize][0] as i32;
                }
            }
            if n_cols != !0 {
                let mut y = 0;
                while y < sz4 {
                    let c = *rf.blk(by4 + y, bx4 - 1);
                    add_compound_extended(&mut st, base, &mut sc, &c, s0, s1, r, &rf.sign_bias);
                    y += BLOCK_DIMENSIONS[c.bs as usize][1] as i32;
                }
            }
            for n in 0..2 {
                let mut m = sc[n];
                if m >= 2 {
                    continue;
                }
                let l = sc[2 + n];
                if l != 0 {
                    st[base + m].mv[n] = st[base + 2].mv[n];
                    m += 1;
                    if m == 2 {
                        continue;
                    }
                    if l == 2 {
                        st[base + 1].mv[n] = st[base + 3].mv[n];
                        continue;
                    }
                }
                loop {
                    st[base + m].mv[n] = tgmv[n];
                    m += 1;
                    if m >= 2 {
                        break;
                    }
                }
            }
            let mut n = cnt;
            if n == 1 && st[0].mv == st[base].mv {
                st[1].mv = st[2].mv;
            }
            loop {
                st[n].weight = 2;
                n += 1;
                if n >= 2 {
                    break;
                }
            }
            cnt = 2;
        }
        for c in st.iter_mut().take(cnt) {
            clampmv(&mut c.mv[0]);
            clampmv(&mut c.mv[1]);
        }
        match refmv_ctx >> 1 {
            0 => newmv_ctx.min(1),
            1 => 1 + newmv_ctx.min(3),
            _ => (3 + newmv_ctx).clamp(4, 7),
        }
    } else {
        if cnt < 2 && r[0] > 0 {
            let sign = rf.sign_bias[r[0] as usize - 1];
            let sz4 = w4.min(h4);
            if n_rows != !0 {
                let mut x = 0;
                while x < sz4 && cnt < 2 {
                    let c = *rf.blk(by4 - 1, bx4 + x);
                    add_single_extended(&mut st, &mut cnt, &c, sign, &rf.sign_bias);
                    x += BLOCK_DIMENSIONS[c.bs as usize][0] as i32;
                }
            }
            if n_cols != !0 {
                let mut y = 0;
                while y < sz4 && cnt < 2 {
                    let c = *rf.blk(by4 + y, bx4 - 1);
                    add_single_extended(&mut st, &mut cnt, &c, sign, &rf.sign_bias);
                    y += BLOCK_DIMENSIONS[c.bs as usize][1] as i32;
                }
            }
        }
        for c in st.iter_mut().take(cnt) {
            clampmv(&mut c.mv[0]);
        }
        for c in st.iter_mut().take(2).skip(cnt) {
            c.mv[0] = tgmv[0];
        }
        (refmv_ctx << 4) | (globalmv_ctx << 3) | newmv_ctx
    };
    let mut out = [RefmvsCandidate::default(); 8];
    out.copy_from_slice(&st[..8]);
    (out, cnt, ctx)
}