isomesh-gpu 0.0.15

GPU isosurface extraction for isomesh: wgpu buffers and pipelines, no engine types.
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
//! How far GPU arithmetic drifts from `libm`'s, per field.
//!
//! This is the deliverable of GPU-011a, not a side check. The whole reason to
//! measure it is that moving field evaluation onto the GPU **gives up the
//! sample-identity** that lets M-142 compare CPU and GPU meshes at all, and the
//! size of what replaces it decides whether the general mechanism is worth
//! building.
#![allow(clippy::float_cmp)]

use super::{FieldSampler, GpuField};
use crate::headless::Gpu;
use crate::{GridParams, read_buffer};

fn gpu() -> &'static Gpu {
    crate::headless::shared()
}

/// How one field's GPU evaluation compares with `isomesh`'s.
///
/// `worst_ulps` is reported but **not** asserted on except for `box_exact`, and
/// the reason is that these fields are zero on their own surface: at a value of
/// `1e-8` an absolute difference of `1e-8` is astronomically many ULPs while
/// being physically nothing. ULP distance is only meaningful between numbers of
/// comparable magnitude, and a signed distance field crosses zero by
/// construction. `worst_abs` is the number that decides whether a crossing
/// moves.
struct Deviation {
    exact: usize,
    total: usize,
    worst_abs: f32,
    worst_ulps: u32,
    at: [f32; 3],
}

fn measure(gpu: &Gpu, sampler: &FieldSampler, field: GpuField, grid: GridParams) -> Deviation {
    let buffer = sampler
        .sample(gpu.device(), gpu.queue(), grid, field)
        .expect("gpu sample");
    let got = read_buffer(
        gpu.device(),
        gpu.queue(),
        buffer.buffer(),
        grid.field_buffer_size(),
    )
    .expect("read back");

    let [sx, sy, sz] = grid.samples();
    let mut deviation = Deviation {
        exact: 0,
        total: 0,
        worst_abs: 0.0,
        worst_ulps: 0,
        at: [0.0; 3],
    };
    let mut i = 0usize;
    for z in 0..sz {
        for y in 0..sy {
            for x in 0..sx {
                let p = grid.sample_position([x, y, z]);
                let cpu = field.sample_on_cpu(p);
                let gpu_value = got[i];
                deviation.total += 1;
                if gpu_value.to_bits() == cpu.to_bits() {
                    deviation.exact += 1;
                } else {
                    let abs = (gpu_value - cpu).abs();
                    // ULPs between two finite floats of the same sign is the
                    // distance between their bit patterns.
                    let ulps = gpu_value.to_bits().abs_diff(cpu.to_bits());
                    if abs > deviation.worst_abs {
                        deviation.worst_abs = abs;
                        deviation.at = p;
                    }
                    deviation.worst_ulps = deviation.worst_ulps.max(ulps);
                }
                i += 1;
            }
        }
    }
    deviation
}

/// The deviation table, printed and asserted.
///
/// Asserted on the **absolute** deviation, which is the quantity that decides
/// whether a crossing moves. The bit-exact counts and ULP figures are printed
/// rather than gated: they are properties of this driver's contraction choices
/// and would make a brittle test, where "no field is wrong by enough to move a
/// surface" is the claim that has to hold on any adapter.
#[test]
fn the_gpu_field_matches_the_cpu_within_a_measured_bound() {
    let gpu = gpu();
    let sampler = FieldSampler::new(gpu.device()).expect("pipeline");
    let grid = GridParams::new([33; 3], [-2.0; 3], 0.125).expect("grid");

    println!(
        "{:<12} {:>10} {:>8} {:>12} {:>8}",
        "field", "bit-exact", "of", "worst abs", "ulps"
    );
    for field in GpuField::ALL {
        let d = measure(gpu, &sampler, field, grid);
        println!(
            "{:<12} {:>10} {:>8} {:>12.3e} {:>8}   worst at {:?}",
            field.name(),
            d.exact,
            d.total,
            d.worst_abs,
            d.worst_ulps,
            d.at
        );

        // A cell is 0.125 world units. Anything within a millionth of that
        // cannot move a crossing anywhere a mesh would notice.
        assert!(
            d.worst_abs < 1e-6,
            "{}: worst deviation {:e} is too large to be rounding",
            field.name(),
            d.worst_abs
        );
    }
}

/// **No field agrees bit-for-bit — not even the ones with no transcendentals.**
///
/// That was the hypothesis and it is wrong. The grid here uses `h = 0.125` from
/// origin `-2.0`, both powers of two, so `origin + h·i` is *exact* and the two
/// sides evaluate at identical positions. The divergence is therefore inside the
/// field expression itself: a GPU is free to contract `x*x + y*y + z*z` into
/// fused multiply-adds, which rounds once where `libm` rounds twice. Same cause
/// as M-142, one layer further up.
///
/// Asserted as an inequality so it cannot rot into a belief: if a driver starts
/// matching exactly, this fails and the table gets re-read.
#[test]
fn no_field_agrees_with_libm_bit_for_bit() {
    let gpu = gpu();
    let sampler = FieldSampler::new(gpu.device()).expect("pipeline");
    let grid = GridParams::new([33; 3], [-2.0; 3], 0.125).expect("grid");

    for field in GpuField::ALL {
        let d = measure(gpu, &sampler, field, grid);
        assert!(
            d.exact < d.total,
            "{} now agrees with libm on all {} samples -- the GPU/CPU divergence \
             this ticket measured has gone away for it, and everything resting \
             on the deviation table needs re-reading",
            field.name(),
            d.total
        );
    }
}

/// `box_exact` is far closer than the others, and the reason is structural.
///
/// Its expression is `abs`, `max` and `min` — all exact in IEEE — over a `sqrt`
/// whose argument is **identically zero everywhere inside the box**, where
/// `sqrt(0)` is exact too. Only the exterior shell has anything to round. So it
/// lands within a single ULP while `sphere` and `torus`, which square and sum
/// three coordinates on every sample, do not.
///
/// This is the observation GPU-011b needs: the drift is a property of the
/// *expression*, not of the GPU, so a field built from exact operations crosses
/// unchanged and one built from products does not.
#[test]
fn the_field_of_exact_operations_is_the_closest() {
    let gpu = gpu();
    let sampler = FieldSampler::new(gpu.device()).expect("pipeline");
    let grid = GridParams::new([33; 3], [-2.0; 3], 0.125).expect("grid");

    let boxed = measure(gpu, &sampler, GpuField::BoxExact, grid);
    assert_eq!(
        boxed.worst_ulps, 1,
        "box_exact's worst deviation is {} ULPs, not 1 -- its arithmetic is \
         supposed to be exact except on the exterior shell",
        boxed.worst_ulps
    );

    for other in [GpuField::Sphere, GpuField::Torus, GpuField::Gyroid] {
        let d = measure(gpu, &sampler, other, grid);
        assert!(
            d.exact < boxed.exact,
            "{} matches on {} samples, box_exact on {} -- box_exact should be \
             the closest",
            other.name(),
            d.exact,
            boxed.exact
        );
    }
}

/// The samples land where `grid_sample_at` says, not transposed.
///
/// A transposed field is sampled correctly and stored rotated, which looks like
/// a meshing bug rather than an indexing one — so the inverse index mapping gets
/// its own asymmetric fixture.
#[test]
fn the_flat_index_is_not_transposed() {
    let gpu = gpu();
    let sampler = FieldSampler::new(gpu.device()).expect("pipeline");
    // Deliberately unequal extents: a transposition on a cube is invisible.
    let grid = GridParams::new([5, 7, 3], [-1.0, -2.0, -0.5], 0.25).expect("grid");
    let buffer = sampler
        .sample(gpu.device(), gpu.queue(), grid, GpuField::BoxExact)
        .expect("sample");
    let got = read_buffer(
        gpu.device(),
        gpu.queue(),
        buffer.buffer(),
        grid.field_buffer_size(),
    )
    .expect("read back");

    let [sx, sy, _] = grid.samples();
    for (index, value) in got.iter().enumerate() {
        let i = index as u32;
        let at = [i % sx, (i / sx) % sy, i / (sx * sy)];
        let expected = GpuField::BoxExact.sample_on_cpu(grid.sample_position(at));
        assert_eq!(
            value.to_bits(),
            expected.to_bits(),
            "sample {index} holds the value for a different position"
        );
    }
}

/// **Does the deviation change the mesh?** — the question the table only
/// implies.
///
/// A sample within `7e-7` of zero can land on the other side of the sign test,
/// and a flipped sign changes a cell's case, its triangle count and its
/// topology. Over two million samples "the deviation is tiny" is an argument;
/// this is the measurement.
#[test]
fn a_gpu_evaluated_field_extracts_the_same_mesh() {
    use crate::{FieldBuffer, MarchingCubesGpu};

    let gpu = gpu();
    let sampler = FieldSampler::new(gpu.device()).expect("pipeline");
    let mc = MarchingCubesGpu::new(gpu.device(), gpu.queue()).expect("pipeline");

    for (field, samples, cell) in [
        (GpuField::Sphere, 65u32, 0.0625f32),
        (GpuField::Torus, 65, 0.0625),
        (GpuField::BoxExact, 65, 0.0625),
        // The one with the largest deviation, at a spacing that is not a power
        // of two so the positions differ as well as the values.
        (GpuField::Gyroid, 49, 4.0 / 48.0),
    ] {
        let grid = GridParams::new([samples; 3], [-2.0; 3], cell).expect("grid");

        let uploaded = FieldBuffer::sampled(gpu.device(), gpu.queue(), grid, &FieldOf(field))
            .expect("cpu sample");
        let produced = sampler
            .sample(gpu.device(), gpu.queue(), grid, field)
            .expect("gpu sample");

        let from_cpu = mc
            .extract_buffers(gpu.device(), gpu.queue(), &uploaded)
            .expect("extract");
        let from_gpu = mc
            .extract_buffers(gpu.device(), gpu.queue(), &produced)
            .expect("extract");

        assert_eq!(
            from_cpu.triangles,
            from_gpu.triangles,
            "{} at {samples}^3: a GPU-evaluated field changed the triangle count \
             ({} against {}), so a sample crossed the sign test",
            field.name(),
            from_gpu.triangles,
            from_cpu.triangles
        );
    }
}

/// `GpuField` as an `Sdf`, so the CPU side of the comparison goes through the
/// same upload path a consumer uses.
struct FieldOf(GpuField);

impl isomesh::Sdf for FieldOf {
    type Scalar = f32;
    fn sample(&self, p: [f32; 3]) -> f32 {
        self.0.sample_on_cpu(p)
    }
}

// -- GPU-011b: the edit log --------------------------------------------------

use super::{GpuBrush, GpuOp, GpuShape};

/// A base field plus an edit log, as `isomesh` folds it.
fn cpu_stack(base: GpuField, brushes: &[GpuBrush], p: [f32; 3]) -> f32 {
    use isomesh::Sdf;
    use isomesh::brush::{Brush, BrushStack};

    struct Base(GpuField);
    impl Sdf for Base {
        type Scalar = f32;
        fn sample(&self, p: [f32; 3]) -> f32 {
            self.0.sample_on_cpu(p)
        }
    }

    let ops: Vec<Brush<GpuShape>> = brushes.iter().map(|b| b.to_cpu()).collect();
    BrushStack {
        base: Base(base),
        brushes: &ops,
    }
    .sample(p)
}

/// Compare a GPU-folded log against `BrushStack`, sample for sample.
fn check_stack(
    gpu: &Gpu,
    sampler: &FieldSampler,
    base: GpuField,
    brushes: &[GpuBrush],
    what: &str,
) {
    let grid = GridParams::new([25; 3], [-2.0; 3], 0.125).expect("grid");
    let buffer = sampler
        .sample_stack(gpu.device(), gpu.queue(), grid, base, brushes)
        .expect("gpu stack");
    let got = read_buffer(
        gpu.device(),
        gpu.queue(),
        buffer.buffer(),
        grid.field_buffer_size(),
    )
    .expect("read back");

    let [sx, sy, sz] = grid.samples();
    let mut worst = 0.0f32;
    let mut i = 0usize;
    for z in 0..sz {
        for y in 0..sy {
            for x in 0..sx {
                let p = grid.sample_position([x, y, z]);
                worst = worst.max((got[i] - cpu_stack(base, brushes, p)).abs());
                i += 1;
            }
        }
    }
    assert!(
        worst < 1e-5,
        "{what}: worst deviation {worst:e} from BrushStack -- too large to be rounding"
    );
    println!("{what:<44} worst {worst:e}");
}

/// Each op, alone, against `brush::apply`.
#[test]
fn every_brush_op_matches_the_cpu_fold() {
    let gpu = gpu();
    let sampler = FieldSampler::new(gpu.device()).expect("pipeline");
    let shape = GpuShape::Sphere {
        center: [0.4, 0.1, -0.2],
        radius: 0.7,
    };

    for (op, name) in [
        (GpuOp::Add, "add"),
        (GpuOp::Subtract, "subtract"),
        (GpuOp::SmoothAdd { k: 0.25 }, "smooth_add k=0.25"),
        // k <= 0 degenerates to a plain min on both sides rather than dividing
        // by zero -- the degenerate case has a right answer and gets it.
        (GpuOp::SmoothAdd { k: 0.0 }, "smooth_add k=0"),
    ] {
        check_stack(
            gpu,
            &sampler,
            GpuField::BoxExact,
            &[GpuBrush { shape, op }],
            name,
        );
    }
}

/// Each shape, against its `isomesh` definition.
#[test]
fn every_brush_shape_matches_the_cpu() {
    let gpu = gpu();
    let sampler = FieldSampler::new(gpu.device()).expect("pipeline");

    for (shape, name) in [
        (
            GpuShape::Sphere {
                center: [0.3, -0.2, 0.1],
                radius: 0.6,
            },
            "sphere brush",
        ),
        (
            GpuShape::BoxExact {
                center: [-0.2, 0.3, 0.0],
                half_extents: [0.5, 0.3, 0.7],
            },
            "box brush",
        ),
        (
            GpuShape::Capsule {
                a: [-0.6, 0.0, 0.0],
                b: [0.6, 0.4, -0.3],
                radius: 0.35,
            },
            "capsule brush",
        ),
        // A zero-length capsule is a sphere, which `brush.rs` calls the right
        // answer rather than a case to reject. Both sides must agree on that.
        (
            GpuShape::Capsule {
                a: [0.1, 0.1, 0.1],
                b: [0.1, 0.1, 0.1],
                radius: 0.4,
            },
            "degenerate capsule",
        ),
    ] {
        check_stack(
            gpu,
            &sampler,
            GpuField::Sphere,
            &[GpuBrush {
                shape,
                op: GpuOp::Subtract,
            }],
            name,
        );
    }
}

/// A long mixed log, which is where an interpreter that reorders would show.
///
/// Mixed adds and subtracts do not commute and a smooth add is not associative
/// (M-36..M-38), so folding first-to-last is part of the answer rather than an
/// implementation detail. A shader that batched or reordered for parallelism
/// would pass every single-op test above and fail this one.
#[test]
fn a_mixed_log_matches_the_cpu_in_order() {
    let gpu = gpu();
    let sampler = FieldSampler::new(gpu.device()).expect("pipeline");

    let mut log = Vec::new();
    for i in 0..12u32 {
        let t = i as f32;
        let shape = match i % 3 {
            0 => GpuShape::Sphere {
                center: [0.9 * (t * 0.7).sin(), 0.6 * (t * 1.1).cos(), 0.4 * t.sin()],
                radius: 0.3 + 0.05 * (t % 3.0),
            },
            1 => GpuShape::BoxExact {
                center: [0.5 * (t * 0.9).cos(), 0.2 * t.sin(), 0.7 * (t * 0.3).sin()],
                half_extents: [0.25, 0.35, 0.2],
            },
            _ => GpuShape::Capsule {
                a: [-0.5, 0.2 * t.sin(), 0.0],
                b: [0.5, 0.1 * t.cos(), 0.2],
                radius: 0.2,
            },
        };
        let op = match i % 4 {
            0 => GpuOp::Add,
            1 => GpuOp::Subtract,
            2 => GpuOp::SmoothAdd { k: 0.2 },
            _ => GpuOp::Subtract,
        };
        log.push(GpuBrush { shape, op });
    }

    check_stack(gpu, &sampler, GpuField::Gyroid, &log, "12-brush mixed log");
}

/// An empty log is the base field, unchanged.
#[test]
fn an_empty_log_is_the_base_field() {
    let gpu = gpu();
    let sampler = FieldSampler::new(gpu.device()).expect("pipeline");
    let grid = GridParams::new([17; 3], [-2.0; 3], 0.25).expect("grid");

    for field in GpuField::ALL {
        let plain = sampler
            .sample(gpu.device(), gpu.queue(), grid, field)
            .expect("sample");
        let empty = sampler
            .sample_stack(gpu.device(), gpu.queue(), grid, field, &[])
            .expect("sample_stack");

        let a = read_buffer(
            gpu.device(),
            gpu.queue(),
            plain.buffer(),
            grid.field_buffer_size(),
        )
        .expect("read");
        let b = read_buffer(
            gpu.device(),
            gpu.queue(),
            empty.buffer(),
            grid.field_buffer_size(),
        )
        .expect("read");
        assert_eq!(
            a.iter().map(|v| v.to_bits()).collect::<Vec<_>>(),
            b.iter().map(|v| v.to_bits()).collect::<Vec<_>>(),
            "{}: an empty log changed the field",
            field.name()
        );
    }
}

/// And the mesh is unchanged, which is the acceptance rather than the deviation.
#[test]
fn a_gpu_folded_log_extracts_the_same_mesh() {
    use crate::{FieldBuffer, MarchingCubesGpu};
    use isomesh::Sdf;
    use isomesh::brush::{Brush, BrushStack};

    let gpu = gpu();
    let sampler = FieldSampler::new(gpu.device()).expect("pipeline");
    let mc = MarchingCubesGpu::new(gpu.device(), gpu.queue()).expect("pipeline");
    let grid = GridParams::new([49; 3], [-2.0; 3], 4.0 / 48.0).expect("grid");

    let log = [
        GpuBrush {
            shape: GpuShape::Sphere {
                center: [0.6, 0.2, 0.0],
                radius: 0.55,
            },
            op: GpuOp::Subtract,
        },
        GpuBrush {
            shape: GpuShape::Capsule {
                a: [-0.8, -0.3, 0.0],
                b: [0.4, 0.5, 0.2],
                radius: 0.22,
            },
            op: GpuOp::Add,
        },
        GpuBrush {
            shape: GpuShape::BoxExact {
                center: [-0.3, 0.4, 0.1],
                half_extents: [0.3, 0.2, 0.4],
            },
            op: GpuOp::SmoothAdd { k: 0.15 },
        },
    ];

    struct Base;
    impl Sdf for Base {
        type Scalar = f32;
        fn sample(&self, p: [f32; 3]) -> f32 {
            GpuField::Sphere.sample_on_cpu(p)
        }
    }
    let ops: Vec<Brush<GpuShape>> = log.iter().map(|b| b.to_cpu()).collect();
    let stack = BrushStack {
        base: Base,
        brushes: &ops,
    };

    let uploaded =
        FieldBuffer::sampled(gpu.device(), gpu.queue(), grid, &stack).expect("cpu sample");
    let folded = sampler
        .sample_stack(gpu.device(), gpu.queue(), grid, GpuField::Sphere, &log)
        .expect("gpu fold");

    let from_cpu = mc
        .extract_buffers(gpu.device(), gpu.queue(), &uploaded)
        .expect("extract");
    let from_gpu = mc
        .extract_buffers(gpu.device(), gpu.queue(), &folded)
        .expect("extract");

    assert!(from_cpu.triangles > 0, "the fixture extracted nothing");
    assert_eq!(
        from_cpu.triangles, from_gpu.triangles,
        "a GPU-folded edit log changed the triangle count: {} against {}",
        from_gpu.triangles, from_cpu.triangles
    );
}

// -- GPU-014: a fold over a base the shader cannot evaluate -------------------

use crate::FieldBuffer;

/// A base `field.wgsl` has no name for: `game_dig`'s own terrain.
///
/// Deliberately not one of [`GpuField::ALL`] — the whole claim of
/// [`FieldSampler::fold_into`] is that the base never enters the shader's
/// vocabulary, so a fixture the shader could evaluate would prove nothing.
struct WavyGround;

impl isomesh::Sdf for WavyGround {
    type Scalar = f32;

    fn sample(&self, p: [f32; 3]) -> f32 {
        let height = 0.35 * (p[0] * 0.9).sin() * (p[2] * 0.7).cos() + 0.15 * (p[0] * 2.1).sin();
        p[1] - height
    }
}

/// The grid the fold tests share: `h = 0.125` from `-2.0`, both powers of two,
/// so `origin + h·i` is exact and the sample positions are bit-identical on the
/// two sides.
fn fold_grid() -> GridParams {
    GridParams::new([25; 3], [-2.0; 3], 0.125).expect("grid")
}

/// [`WavyGround`] on the CPU, in the buffer's own `x`-fastest order.
fn ground_samples(grid: GridParams) -> Vec<f32> {
    use isomesh::Sdf;

    let [sx, sy, sz] = grid.samples();
    let mut out = Vec::with_capacity(grid.sample_count() as usize);
    for z in 0..sz {
        for y in 0..sy {
            for x in 0..sx {
                out.push(WavyGround.sample(grid.sample_position([x, y, z])));
            }
        }
    }
    out
}

/// The device fold over an uploaded base equals `BrushStack` over the same one.
///
/// **The base is bit-identical on the two sides here**, because the CPU produced
/// it and the upload copies it — so unlike every test above, what this measures
/// is the log fold alone, with none of `GpuField`'s own evaluation drift in it.
///
/// `1e-6` rather than bit-exactness: M-154 measures the drift as a property of
/// the *expression* rather than the platform — a GPU may contract a
/// sum-of-products into fused multiply-adds — and M-157 records edit-log
/// agreement at 8.4e-7.
#[test]
fn a_fold_over_an_uploaded_base_matches_the_cpu() {
    use isomesh::Sdf;
    use isomesh::brush::{Brush, BrushStack};

    let gpu = gpu();
    let sampler = FieldSampler::new(gpu.device()).expect("pipeline");
    let grid = fold_grid();
    let base = FieldBuffer::uploaded(gpu.device(), gpu.queue(), grid, &ground_samples(grid))
        .expect("base upload");
    let out = FieldBuffer::new(gpu.device(), grid);

    // One add and one subtract at least: the two do not commute, so a fold that
    // reordered them would pass a single-op fixture and fail this.
    let log = [
        GpuBrush {
            shape: GpuShape::Sphere {
                center: [0.3, -0.1, 0.2],
                radius: 0.6,
            },
            op: GpuOp::Add,
        },
        GpuBrush {
            shape: GpuShape::BoxExact {
                center: [-0.4, 0.2, 0.1],
                half_extents: [0.5, 0.3, 0.4],
            },
            op: GpuOp::Subtract,
        },
        GpuBrush {
            shape: GpuShape::Capsule {
                a: [-0.5, 0.1, -0.2],
                b: [0.5, 0.3, 0.2],
                radius: 0.18,
            },
            op: GpuOp::SmoothAdd { k: 0.2 },
        },
    ];
    sampler
        .fold_into(gpu.device(), gpu.queue(), &base, &out, &log)
        .expect("fold");

    let got = read_buffer(
        gpu.device(),
        gpu.queue(),
        out.buffer(),
        grid.field_buffer_size(),
    )
    .expect("read back");

    let ops: Vec<Brush<GpuShape>> = log.iter().map(|b| b.to_cpu()).collect();
    let stack = BrushStack {
        base: WavyGround,
        brushes: &ops,
    };
    let [sx, sy, sz] = grid.samples();
    let mut worst = 0.0f32;
    let mut i = 0usize;
    for z in 0..sz {
        for y in 0..sy {
            for x in 0..sx {
                worst = worst.max((got[i] - stack.sample(grid.sample_position([x, y, z]))).abs());
                i += 1;
            }
        }
    }
    assert!(
        worst <= 1e-6,
        "worst deviation {worst:e} from BrushStack over an uploaded base -- \
         too large to be rounding"
    );
    println!("fold over an uploaded base                   worst {worst:e}");
}

/// An empty log copies the base **bit for bit**.
///
/// The only thing `fold_into` changes about the kernel is where the first value
/// comes from, so with nothing to fold the output must be the base itself rather
/// than the base to within rounding. Any arithmetic on the way through -- a
/// gathered neighbour, a position recomputed and re-evaluated -- shows here and
/// nowhere else.
#[test]
fn an_empty_fold_copies_the_base_bit_for_bit() {
    let gpu = gpu();
    let sampler = FieldSampler::new(gpu.device()).expect("pipeline");
    let grid = fold_grid();
    let samples = ground_samples(grid);
    let base =
        FieldBuffer::uploaded(gpu.device(), gpu.queue(), grid, &samples).expect("base upload");
    let out = FieldBuffer::new(gpu.device(), grid);
    sampler
        .fold_into(gpu.device(), gpu.queue(), &base, &out, &[])
        .expect("fold");

    let got = read_buffer(
        gpu.device(),
        gpu.queue(),
        out.buffer(),
        grid.field_buffer_size(),
    )
    .expect("read back");
    assert_eq!(
        got.iter().map(|v| v.to_bits()).collect::<Vec<_>>(),
        samples.iter().map(|v| v.to_bits()).collect::<Vec<_>>(),
        "an empty fold changed the base it was handed"
    );
}

/// A base and an output describing different sample counts are refused.
///
/// The kernel indexes both by the same flat id, so a fold across two grids reads
/// the wrong position for every sample past the first row and writes a buffer
/// that still looks meshable. That is the failure that cannot be allowed to be
/// silent, and it is the caller's grid rather than the kernel's to get right.
#[test]
fn a_base_of_a_different_size_is_refused() {
    let gpu = gpu();
    let sampler = FieldSampler::new(gpu.device()).expect("pipeline");
    let small = GridParams::new([9; 3], [-2.0; 3], 0.125).expect("grid");
    let large = GridParams::new([17; 3], [-2.0; 3], 0.125).expect("grid");
    let base = FieldBuffer::new(gpu.device(), small);
    let out = FieldBuffer::new(gpu.device(), large);

    assert_eq!(
        sampler.fold_into(gpu.device(), gpu.queue(), &base, &out, &[]),
        Err(crate::Error::SampleCountMismatch {
            expected: small.sample_count(),
            got: large.sample_count(),
        }),
        "a fold across two grids was accepted"
    );
}