meshoptimizer-rs 0.2.20

Rust port of meshoptimizer 1.3 with byte-identical output. Covers mesh simplification, meshlets, cluster LOD, vertex/index compression analysis, remeshing and optional parallel batch APIs and audited SIMD codecs.
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
766
767
768
769
770
771
772
773
774
775
776
777
778
// meshoptimizer 1.3 scalar vertex encoder (vertexcodec.cpp), MIT, Arseny Kapoulkine.
//
// Every capacity check follows the reference in the same order, so a caller
// buffer is rejected exactly where meshopt_encodeVertexBufferLevel returns 0.
// Deltas for the four bytes of one channel group are computed together; the
// reference computes the same bytes one component at a time, in the same order.
use crate::Error;

const GROUP: usize = 16;
const DECODE_LIMIT: usize = 24;
const BLOCK_MAX: usize = 256;
const HEADER: u8 = 0xa0;
const TAIL_V0: usize = 32;
const TAIL_V1: usize = 24;

#[inline]
pub(super) fn block_size(stride: usize) -> usize {
    ((8192 / stride) & !(GROUP - 1)).min(BLOCK_MAX)
}

/// Upstream meshopt_encodeVertexBufferBound, with checked arithmetic.
#[inline]
pub(super) fn bound(count: usize, stride: usize) -> Result<usize, Error> {
    let block = block_size(stride);
    let blocks = count / block + usize::from(!count.is_multiple_of(block));
    let header = (block / GROUP).div_ceil(4);
    let tail = (stride + stride / 4).max(TAIL_V0.max(TAIL_V1));
    // stride <= 256, so the per-block size cannot overflow.
    let per = stride * (header + block) + stride / 4;
    blocks
        .checked_mul(per)
        .and_then(|n| n.checked_add(1 + tail))
        .ok_or(Error::SizeOverflow)
}

struct Writer<'a> {
    out: &'a mut [u8],
    pos: usize,
}
impl Writer<'_> {
    #[inline]
    fn remaining(&self) -> usize {
        self.out.len() - self.pos
    }
}

/// Encoded group sizes for every bit width, from one pass over the group.
#[derive(Clone, Copy, Default)]
struct Sizes {
    zero: bool,
    ge1: u8,
    ge3: u8,
    ge15: u8,
}
impl Sizes {
    #[inline]
    fn new(group: &[u8; GROUP]) -> Self {
        // Byte counters (at most 16) keep the comparisons in byte lanes.
        let (mut ge1, mut ge3, mut ge15) = (0u8, 0u8, 0u8);
        for &b in group {
            ge1 += u8::from(b >= 1);
            ge3 += u8::from(b >= 3);
            ge15 += u8::from(b >= 15);
        }
        Self {
            zero: ge1 == 0,
            ge1,
            ge3,
            ge15,
        }
    }
    // encodeBytesGroupMeasure; size_t(-1) marks an impossible zero group.
    #[inline]
    fn of(self, bits: u32) -> usize {
        match bits {
            0 => {
                if self.zero {
                    0
                } else {
                    usize::MAX
                }
            }
            1 => 2 + usize::from(self.ge1),
            2 => 4 + usize::from(self.ge3),
            4 => 8 + usize::from(self.ge15),
            _ => GROUP,
        }
    }
}

/// Smallest of the 1/2/4/8-bit group sizes (estimateChannel's measure),
/// computed in byte arithmetic: every candidate is at most 24.
#[inline(always)]
fn best_size_1248(group: &[u8; GROUP]) -> u8 {
    let (mut ge1, mut ge3, mut ge15) = (0u8, 0u8, 0u8);
    for &b in group {
        ge1 += u8::from(b >= 1);
        ge3 += u8::from(b >= 3);
        ge15 += u8::from(b >= 15);
    }
    (2 + ge1).min(4 + ge3).min(8 + ge15).min(GROUP as u8)
}

#[inline]
fn encode_group(dst: &mut [u8; DECODE_LIMIT], group: &[u8; GROUP], bits: u32) -> usize {
    match bits {
        0 => 0,
        1 => packed_group::<1>(dst, group),
        2 => packed_group::<2>(dst, group),
        4 => packed_group::<4>(dst, group),
        _ => {
            dst[..GROUP].copy_from_slice(group);
            GROUP
        }
    }
}

#[inline(always)]
fn packed_group<const BITS: u32>(dst: &mut [u8; DECODE_LIMIT], group: &[u8; GROUP]) -> usize {
    let per = (8 / BITS) as usize;
    let sentinel = (1u8 << BITS) - 1;
    let packed = GROUP / per;
    for (i, slot) in dst[..packed].iter_mut().enumerate() {
        let mut byte = 0u8;
        for &v in &group[i * per..(i + 1) * per] {
            byte = (byte << BITS) | v.min(sentinel);
        }
        // One-bit groups are stored in reverse bit order.
        *slot = if BITS == 1 { byte.reverse_bits() } else { byte };
    }
    let mut n = packed;
    for &v in group {
        // Branchless append of out-of-range values, as upstream: n <= 23
        // here, and a byte written past the final literal is overwritten by
        // later output (every stream ends with its tail).
        dst[n] = v;
        n += usize::from(v >= sentinel);
    }
    n
}

fn encode_bytes(
    w: &mut Writer<'_>,
    buffer: &[u8],
    measured: &[Sizes],
    bits: &[u32],
) -> Result<(), Error> {
    let header_size = (buffer.len() / GROUP).div_ceil(4);
    if w.remaining() < header_size {
        return Err(Error::BufferTooSmall);
    }
    let header = w.pos;
    w.out[header..header + header_size].fill(0);
    w.pos += header_size;
    let mut last_bits = u32::MAX;
    for (i, (group, &sizes)) in buffer
        .as_chunks::<GROUP>()
        .0
        .iter()
        .zip(measured)
        .enumerate()
    {
        if w.remaining() < DECODE_LIMIT {
            return Err(Error::BufferTooSmall);
        }
        let mut best_k = 3;
        let mut best_size = sizes.of(bits[3]);
        for (k, &b) in bits[..3].iter().enumerate() {
            let size = sizes.of(b);
            // Favor consistent selection across groups, but never replace literals.
            if size < best_size || (size == best_size && b == last_bits && bits[best_k] != 8) {
                best_k = k;
                best_size = size;
            }
        }
        w.out[header + i / 4] |= (best_k as u8) << ((i % 4) * 2);
        let dst: &mut [u8; DECODE_LIMIT] = (&mut w.out[w.pos..w.pos + DECODE_LIMIT])
            .try_into()
            .map_err(|_| Error::BufferTooSmall)?;
        w.pos += encode_group(dst, group, bits[best_k]);
        last_bits = bits[best_k];
    }
    Ok(())
}

#[inline]
fn word(data: &[u8], at: usize) -> u32 {
    u32::from_le_bytes(
        *data[at..]
            .first_chunk::<4>()
            .expect("validated vertex word"),
    )
}

/// encodeDeltas for the four byte lanes k4..k4+4, which share one channel.
fn deltas(
    lanes: &mut [[u8; BLOCK_MAX]; 4],
    data: &[u8],
    count: usize,
    stride: usize,
    last: &[u8],
    k4: usize,
    channel: u8,
) {
    match stride {
        4 => deltas_fixed::<4>(lanes, data, count, stride, last, k4, channel),
        12 => deltas_fixed::<12>(lanes, data, count, stride, last, k4, channel),
        32 => deltas_fixed::<32>(lanes, data, count, stride, last, k4, channel),
        _ => deltas_fixed::<0>(lanes, data, count, stride, last, k4, channel),
    }
}

fn deltas_fixed<const STRIDE: usize>(
    lanes: &mut [[u8; BLOCK_MAX]; 4],
    data: &[u8],
    count: usize,
    stride: usize,
    last: &[u8],
    k4: usize,
    channel: u8,
) {
    let stride = if STRIDE == 0 { stride } else { STRIDE };
    let k4 = if STRIDE == 4 { 0 } else { k4 };
    let mut p = word(last, k4);
    let [l0, l1, l2, l3] = lanes;
    let rows = data.chunks_exact(stride).take(count);
    let lanes = l0[..count]
        .iter_mut()
        .zip(&mut l1[..count])
        .zip(&mut l2[..count])
        .zip(&mut l3[..count]);
    match channel & 3 {
        0 => {
            for (row, (((a, b), c), d)) in rows.zip(lanes) {
                let v = word(row, k4);
                let x = v.to_le_bytes();
                let y = p.to_le_bytes();
                let z = |j: usize| {
                    let t = x[j].wrapping_sub(y[j]);
                    (t << 1) ^ 0u8.wrapping_sub(t >> 7)
                };
                (*a, *b, *c, *d) = (z(0), z(1), z(2), z(3));
                p = v;
            }
        }
        1 => {
            for (row, (((a, b), c), d)) in rows.zip(lanes) {
                let v = word(row, k4);
                let z = |shift: u32| {
                    let t = ((v >> shift) as u16).wrapping_sub((p >> shift) as u16);
                    (t << 1) ^ 0u16.wrapping_sub(t >> 15)
                };
                let (lo, hi) = (z(0), z(16));
                (*a, *b, *c, *d) = (lo as u8, (lo >> 8) as u8, hi as u8, (hi >> 8) as u8);
                p = v;
            }
        }
        _ => {
            // Channel values are produced internally: 0, 1 or 2 | rotation << 4.
            let rot = u32::from(channel >> 4);
            for (row, (((a, b), c), d)) in rows.zip(lanes) {
                let v = word(row, k4);
                let [x0, x1, x2, x3] = (v ^ p).rotate_left(rot).to_le_bytes();
                (*a, *b, *c, *d) = (x0, x1, x2, x3);
                p = v;
            }
        }
    }
}

fn estimate_bits(v: u8) -> usize {
    if v <= 15 {
        if v <= 3 {
            if v == 0 {
                0
            } else {
                2
            }
        } else {
            4
        }
    } else {
        8
    }
}

fn estimate_rotate(vertices: &[u8], count: usize, stride: usize, k: usize) -> u8 {
    pick_rotation(&rotate_sizes(
        vertices,
        count,
        stride,
        k,
        0..count.div_ceil(GROUP),
    ))
}

/// Encoded-size estimate of each rotation over the sixteen-vertex groups in
/// `groups`. Group sums are integers, so ranges may be summed in any order.
fn rotate_sizes(
    vertices: &[u8],
    count: usize,
    stride: usize,
    k: usize,
    groups: core::ops::Range<usize>,
) -> [usize; 8] {
    let mut sizes = [0usize; 8];
    // The first delta of the buffer is zero; later ranges start from the
    // vertex before their first group.
    let before = (groups.start * GROUP).saturating_sub(1);
    let mut last = word(&vertices[before * stride..], k);
    let end = (groups.end * GROUP).min(count);
    for group in vertices[groups.start * GROUP * stride..end * stride].chunks(stride * GROUP) {
        let mut bitg = 0u32;
        for row in group.chunks_exact(stride) {
            let v = word(row, k);
            bitg |= v ^ last;
            last = v;
        }
        for (j, size) in sizes.iter_mut().enumerate() {
            let [a, b, c, d] = bitg.rotate_left(j as u32).to_le_bytes();
            *size += estimate_bits(a) + estimate_bits(b) + estimate_bits(c) + estimate_bits(d);
        }
    }
    sizes
}

fn pick_rotation(sizes: &[usize; 8]) -> u8 {
    let mut best = 0;
    for rot in 1..8 {
        if sizes[rot] < sizes[best] {
            best = rot;
        }
    }
    best as u8
}

#[allow(clippy::too_many_arguments)]
fn estimate_channel(
    lanes: &mut [[u8; BLOCK_MAX]; 4],
    vertices: &[u8],
    count: usize,
    stride: usize,
    k: usize,
    block_size: usize,
    skip: usize,
    max_channel: u8,
    rot: u8,
) -> u8 {
    let sizes = channel_sizes(
        lanes,
        vertices,
        count,
        stride,
        k,
        block_size,
        skip,
        max_channel,
        rot,
        0..sample_count(count, block_size, skip),
    );
    pick_channel(&sizes, max_channel, rot)
}

/// Number of sampled blocks: every `skip`-th block, starting with the first.
fn sample_count(count: usize, block_size: usize, skip: usize) -> usize {
    count.div_ceil(block_size * skip)
}

/// Estimated encoded size of each delta channel over the sampled blocks
/// numbered `samples`. Block sums are integers, so ranges may be summed in
/// any order.
#[allow(clippy::too_many_arguments)]
fn channel_sizes(
    lanes: &mut [[u8; BLOCK_MAX]; 4],
    vertices: &[u8],
    count: usize,
    stride: usize,
    k: usize,
    block_size: usize,
    skip: usize,
    max_channel: u8,
    rot: u8,
    samples: core::ops::Range<usize>,
) -> [usize; 3] {
    let mut sizes = [0usize; 3];
    for sample in samples {
        let i = sample * block_size * skip;
        let block = block_size.min(count - i);
        let aligned = block.div_ceil(GROUP) * GROUP;
        let start = if i == 0 { 0 } else { i - 1 };
        // Only bytes k..k+4 of the previous vertex are read.
        let last = &vertices[start * stride..(start + 1) * stride];
        for lane in lanes.iter_mut() {
            lane[block..aligned].fill(0);
        }
        let data = &vertices[i * stride..(i + block) * stride];
        for channel in 0..max_channel {
            deltas(lanes, data, block, stride, last, k, channel | (rot << 4));
            for lane in lanes.iter() {
                let mut total = 0u32;
                for group in lane[..aligned].as_chunks::<GROUP>().0 {
                    total += u32::from(best_size_1248(group));
                }
                sizes[channel as usize] += total as usize;
            }
        }
    }
    sizes
}

fn pick_channel(sizes: &[usize; 3], max_channel: u8, rot: u8) -> u8 {
    let mut best = 0;
    for channel in 1..max_channel as usize {
        if sizes[channel] < sizes[best] {
            best = channel;
        }
    }
    if best == 2 {
        2 | (rot << 4)
    } else {
        best as u8
    }
}

fn estimate_control(measured: &[Sizes], count: usize, level: u8) -> u8 {
    if measured.iter().all(|s| s.zero) {
        return 2;
    }
    if level == 0 {
        return 1;
    }
    let header = measured.len().div_ceil(4);
    let (mut est0, mut est1) = (header, header);
    for &s in measured {
        let s124 = s.of(1).min(s.of(2)).min(s.of(4));
        est0 += s124.min(s.of(0));
        est1 += s124.min(s.of(8));
    }
    if est0 < count || est1 < count {
        if est0 < est1 {
            0
        } else {
            1
        }
    } else {
        3
    }
}

#[allow(clippy::too_many_arguments)]
fn encode_block(
    w: &mut Writer<'_>,
    lanes: &mut [[u8; BLOCK_MAX]; 4],
    data: &[u8],
    count: usize,
    stride: usize,
    last: &mut [u8; 256],
    channels: &[u8; 64],
    version: u8,
    level: u8,
) -> Result<(), Error> {
    let aligned = count.div_ceil(GROUP) * GROUP;
    // The reference zeroes its block buffer; only the first count bytes of a
    // component are written, so the aligned tail of every component encodes as zero.
    for lane in lanes.iter_mut() {
        lane[count..aligned].fill(0);
    }
    let control_size = if version == 0 { 0 } else { stride / 4 };
    if w.remaining() < control_size {
        return Err(Error::BufferTooSmall);
    }
    let control = w.pos;
    w.out[control..control + control_size].fill(0);
    w.pos += control_size;
    for k4 in (0..stride).step_by(4) {
        let channel = if version == 0 { 0 } else { channels[k4 / 4] };
        deltas(lanes, data, count, stride, &last[..], k4, channel);
        for (j, lane) in lanes.iter().enumerate() {
            let buffer = &lane[..aligned];
            // Measure each group once for control selection and encoding.
            let mut measured = [Sizes::default(); BLOCK_MAX / GROUP];
            let measured = &mut measured[..aligned / GROUP];
            for (s, group) in measured.iter_mut().zip(buffer.as_chunks::<GROUP>().0) {
                *s = Sizes::new(group);
            }
            let mut ctrl = 0;
            if version != 0 {
                ctrl = estimate_control(measured, count, level);
                w.out[control + k4 / 4] |= ctrl << (j * 2);
            }
            if ctrl == 3 {
                if w.remaining() < count {
                    return Err(Error::BufferTooSmall);
                }
                w.out[w.pos..w.pos + count].copy_from_slice(&buffer[..count]);
                w.pos += count;
            } else if ctrl != 2 {
                let bits: &[u32] = match (version, ctrl) {
                    (0, _) => &[0, 2, 4, 8],
                    (_, 0) => &[0, 1, 2, 4],
                    _ => &[1, 2, 4, 8],
                };
                encode_bytes(w, buffer, measured, bits)?;
            }
        }
    }
    last[..stride].copy_from_slice(&data[(count - 1) * stride..count * stride]);
    Ok(())
}

/// meshopt_encodeVertexBufferLevel. Parameters are validated by the caller:
/// stride is 4..=256 and a multiple of four, vertices holds count * stride
/// bytes, version is 0 or 1 and level is 0..=9.
pub(super) fn encode(
    out: &mut [u8],
    vertices: &[u8],
    count: usize,
    stride: usize,
    version: u8,
    level: u8,
) -> Result<usize, Error> {
    let mut w = Writer { out, pos: 0 };
    if w.remaining() < 1 {
        return Err(Error::BufferTooSmall);
    }
    w.out[0] = HEADER | version;
    w.pos = 1;
    let mut first = [0u8; 256];
    if count > 0 {
        first[..stride].copy_from_slice(&vertices[..stride]);
    }
    let mut last = first;
    let block = block_size(stride);
    let mut channels = [0u8; 64];
    let mut lanes = [[0u8; BLOCK_MAX]; 4];
    if version != 0 && level > 1 && count > 1 {
        for k in (0..stride).step_by(4) {
            let rot = if level >= 3 {
                estimate_rotate(vertices, count, stride, k)
            } else {
                0
            };
            let max_channel = if level >= 3 { 3 } else { 2 };
            channels[k / 4] = estimate_channel(
                &mut lanes,
                vertices,
                count,
                stride,
                k,
                block,
                3,
                max_channel,
                rot,
            );
        }
    }
    let mut offset = 0;
    while offset < count {
        let size = block.min(count - offset);
        encode_block(
            &mut w,
            &mut lanes,
            &vertices[offset * stride..(offset + size) * stride],
            size,
            stride,
            &mut last,
            &channels,
            version,
            level,
        )?;
        offset += size;
    }
    let tail = stride + if version == 0 { 0 } else { stride / 4 };
    let padded = tail.max(if version == 0 { TAIL_V0 } else { TAIL_V1 });
    if w.remaining() < padded {
        return Err(Error::BufferTooSmall);
    }
    let mut pos = w.pos;
    w.out[pos..pos + padded - tail].fill(0);
    pos += padded - tail;
    w.out[pos..pos + stride].copy_from_slice(&first[..stride]);
    pos += stride;
    if version != 0 {
        w.out[pos..pos + stride / 4].copy_from_slice(&channels[..stride / 4]);
        pos += stride / 4;
    }
    Ok(pos)
}

/// Fewest blocks in a parallel work unit; smaller buffers encode serially.
#[cfg(feature = "parallel")]
pub(super) const MIN_CHUNK_BLOCKS: usize = 4;
#[cfg(feature = "parallel")]
const ROTATE_TASK_GROUPS: usize = 2048;
#[cfg(feature = "parallel")]
const CHANNEL_TASK_SAMPLES: usize = 32;

#[cfg(feature = "parallel")]
fn ranges(total: usize, per: usize) -> impl Iterator<Item = core::ops::Range<usize>> {
    (0..total)
        .step_by(per)
        .map(move |start| start..(start + per).min(total))
}

#[cfg(feature = "parallel")]
fn zeroed(len: usize) -> Result<alloc::vec::Vec<u8>, Error> {
    let mut v = alloc::vec::Vec::new();
    v.try_reserve_exact(len)
        .map_err(|_| Error::AllocationFailed)?;
    v.resize(len, 0);
    Ok(v)
}

/// Per-channel-group control bytes (version 1, level above 1), computed with
/// the estimators split into integer-summed ranges that run concurrently.
#[cfg(feature = "parallel")]
fn estimate_channels_parallel(
    vertices: &[u8],
    count: usize,
    stride: usize,
    level: u8,
    block: usize,
) -> [u8; 64] {
    use rayon::prelude::*;
    let groups4 = stride / 4;
    let mut rotations = [0u8; 64];
    if level >= 3 {
        let tasks: alloc::vec::Vec<_> = (0..groups4)
            .flat_map(|g| {
                ranges(count.div_ceil(GROUP), ROTATE_TASK_GROUPS).map(move |range| (g, range))
            })
            .collect();
        let sums: alloc::vec::Vec<[usize; 8]> = tasks
            .par_iter()
            .map(|(g, range)| rotate_sizes(vertices, count, stride, g * 4, range.clone()))
            .collect();
        let mut totals = [[0usize; 8]; 64];
        for ((g, _), sum) in tasks.iter().zip(&sums) {
            for (total, value) in totals[*g].iter_mut().zip(sum) {
                *total += value;
            }
        }
        for g in 0..groups4 {
            rotations[g] = pick_rotation(&totals[g]);
        }
    }
    let max_channel = if level >= 3 { 3 } else { 2 };
    let tasks: alloc::vec::Vec<_> = (0..groups4)
        .flat_map(|g| {
            ranges(sample_count(count, block, 3), CHANNEL_TASK_SAMPLES).map(move |range| (g, range))
        })
        .collect();
    let sums: alloc::vec::Vec<[usize; 3]> = tasks
        .par_iter()
        .map(|(g, range)| {
            let mut lanes = [[0u8; BLOCK_MAX]; 4];
            channel_sizes(
                &mut lanes,
                vertices,
                count,
                stride,
                g * 4,
                block,
                3,
                max_channel,
                rotations[*g],
                range.clone(),
            )
        })
        .collect();
    let mut totals = [[0usize; 3]; 64];
    for ((g, _), sum) in tasks.iter().zip(&sums) {
        for (total, value) in totals[*g].iter_mut().zip(sum) {
            *total += value;
        }
    }
    let mut channels = [0u8; 64];
    for g in 0..groups4 {
        channels[g] = pick_channel(&totals[g], max_channel, rotations[g]);
    }
    channels
}

/// Parallel [`encode`] into an exactly sized vector, with the same bytes.
///
/// Block bodies depend only on the input (the baseline of a block is the last
/// input vertex of the previous block) and on the per-channel estimates, so
/// runs of blocks encode independently and concatenate in order. The
/// parameters are validated by the caller as for [`encode`].
#[cfg(feature = "parallel")]
pub(super) fn encode_parallel(
    vertices: &[u8],
    count: usize,
    stride: usize,
    version: u8,
    level: u8,
) -> Result<alloc::vec::Vec<u8>, Error> {
    use rayon::prelude::*;
    let block = block_size(stride);
    let channels = if version != 0 && level > 1 && count > 1 {
        estimate_channels_parallel(vertices, count, stride, level, block)
    } else {
        [0u8; 64]
    };
    let blocks = count.div_ceil(block);
    let per_block = stride * ((block / GROUP).div_ceil(4) + block) + stride / 4;
    let chunk_blocks = blocks
        .div_ceil(rayon::current_num_threads() * 4)
        .max(MIN_CHUNK_BLOCKS);
    let pieces: alloc::vec::Vec<Result<alloc::vec::Vec<u8>, Error>> = ranges(blocks, chunk_blocks)
        .collect::<alloc::vec::Vec<_>>()
        .into_par_iter()
        .map(|run| {
            // Slack beyond the worst case keeps every capacity check passing.
            let mut buffer = zeroed(run.len() * per_block + 64)?;
            let mut w = Writer {
                out: &mut buffer,
                pos: 0,
            };
            let mut lanes = [[0u8; BLOCK_MAX]; 4];
            let mut offset = run.start * block;
            let end = (run.end * block).min(count);
            let mut last = [0u8; 256];
            let previous = offset.saturating_sub(1);
            last[..stride].copy_from_slice(&vertices[previous * stride..(previous + 1) * stride]);
            while offset < end {
                let size = block.min(count - offset);
                encode_block(
                    &mut w,
                    &mut lanes,
                    &vertices[offset * stride..(offset + size) * stride],
                    size,
                    stride,
                    &mut last,
                    &channels,
                    version,
                    level,
                )?;
                offset += size;
            }
            let used = w.pos;
            buffer.truncate(used);
            buffer.shrink_to_fit();
            Ok(buffer)
        })
        .collect();
    let tail = stride + if version == 0 { 0 } else { stride / 4 };
    let padded = tail.max(if version == 0 { TAIL_V0 } else { TAIL_V1 });
    let mut total = 1 + padded;
    for piece in &pieces {
        match piece {
            Ok(piece) => total += piece.len(),
            Err(error) => return Err(*error),
        }
    }
    let mut out = super::vertex_parallel::zeroed(total)?;
    out[0] = HEADER | version;
    let (body, rest) = out[1..].split_at_mut(total - 1 - padded);
    let mut targets = alloc::vec::Vec::new();
    let mut remaining = body;
    for piece in pieces.iter().flatten() {
        let (target, after) = remaining.split_at_mut(piece.len());
        targets.push((target, piece));
        remaining = after;
    }
    targets
        .par_iter_mut()
        .for_each(|(target, piece)| target.copy_from_slice(piece));
    // The padding stays zero; the first vertex and the channel bytes follow it.
    let trailer = &mut rest[padded - tail..];
    if count > 0 {
        trailer[..stride].copy_from_slice(&vertices[..stride]);
    }
    if version != 0 {
        trailer[stride..].copy_from_slice(&channels[..stride / 4]);
    }
    Ok(out)
}