indicatrix-worker 0.2.2

Library server for a indicatrix-backed design catalogue (`serve`, default build): serves indicatrix-net's read-only design-library protocol over mutual TLS. The optional `worker` feature adds render capacity on top -- accepting RenderRequests and replying with traced radiance -- and `render` a one-shot scene-to-PNG CLI.
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
//! The one function both subcommands use to actually run the tracer: [`trace_samples`].
//!
//! It traces an explicit `[first_sample, first_sample + samples)` range across a whole
//! frame and returns the summed (never averaged) per-pixel XYZ radiance -- exactly what
//! a caller accumulating contributions from multiple workers needs to add into its own
//! running total. See `indicatrix_net`'s crate docs for why sample-index partitioning
//! (not screen-space tiling) is what makes that additive.
//!
//! # Keeping the seed formula in sync with the viewer
//!
//! The per-sample RNG seed and pixel-jitter derivation below is copied verbatim from
//! `apps/indicatrix-cut/src/bridge/render_thread.rs`'s `render_frame_scanlines`. It is
//! not re-derived here: the formula is a function of `(global_pixel_idx, sample_num)`
//! alone, `sample_num` being the absolute sample index, never a batch-relative offset --
//! exactly what lets tracing a sample range here, remotely, produce numbers identical to
//! the viewer tracing that range itself. If this formula and the viewer's ever drift
//! apart, a remote worker's contribution silently composes into a wrong but
//! plausible-looking image (the wire-protocol analogue of the build-hash check that
//! catches drift on the physics side).
//!
//! # A process-wide permit pool bounds total CPU-tracer threads across connections
//!
//! This worker runs one thread per connection and one tracer thread per request; each
//! tracer's [`trace_into`] then spawns `effective_thread_count(threads)` scoped threads
//! of its own for one sub-batch. Each of those numbers is reasonable in isolation, but
//! nothing before this coordinated them across CONCURRENT requests: two clients
//! streaming at once each ask for (typically) every available core, oversubscribing the
//! CPU 2x (and N clients, Nx) -- every thread fighting the others for the same cores,
//! worse throughput for everyone rather than either request simply queueing its share.
//!
//! [`THREAD_PERMITS`] fixes this: a process-wide pool sized to
//! `effective_thread_count(0)` (this machine's `available_parallelism`, the same number
//! a solo request would ask for) that every [`trace_into`] call acquires `num_threads`
//! permits from before its `thread::scope` and releases right after. A request wanting
//! more threads than the pool's total capacity is clamped to that capacity (see
//! [`ThreadPermits::acquire`]) rather than blocking forever on permits that could never
//! all be free simultaneously. Permits are held only for the duration of one `trace_into`
//! call (one sub-batch, not a whole request), so a large request doesn't starve a second
//! client indefinitely -- it interleaves at sub-batch granularity instead.
//!
//! The GPU submitter thread (`hybrid::dispatch_concurrently`'s scoped thread that submits
//! and polls a GPU dispatch) takes no permits at all: it isn't a CPU tracer thread, and
//! the "one core reserved for the GPU submitter" accounting in
//! [`hybrid::cpu_threads_beside_gpu`] already shrinks the CPU side's OWN permit request
//! by one to leave that thread room -- the two mechanisms compose without double-counting
//! or deadlocking each other, since neither ever waits on the other's permits.

use glam::Vec3;
use indicatrix::optics::raytracer::{
    Camera, EnvironmentSource, FacetFinish, build_plane_soa, pixel_rotations, sample_draws,
    trace_spectral_ray_with_finish_soa,
};
use indicatrix_net::SceneState;
use std::{
    sync::{
        Condvar, LazyLock, Mutex,
        atomic::{AtomicBool, AtomicUsize, Ordering},
    },
    thread,
};

use indicatrix::renderer::gpu_backend::{GpuAccumulate, GpuBackend, GpuSceneRef};

pub(crate) mod hybrid;

/// Resolves the per-facet finish `scene.girdle_frosted` implies.
///
/// `indicatrix_net::SceneState` carries no `Vec<FacetFinish>` of its own -- the
/// wire-format encoding is the single `girdle_frosted` bool, and the worker re-derives
/// the same `Vec<FacetFinish>` the viewer would via `indicatrix::geometry::girdle_facet_finishes`,
/// a pure, deterministic function of `scene.planes` alone. `false` returns an empty
/// slice, equivalent to every facet using `FacetFinish::default() == Polished`.
#[must_use]
fn resolve_facet_finishes(scene: &SceneState) -> Vec<FacetFinish> {
    if scene.girdle_frosted {
        indicatrix::geometry::girdle_facet_finishes(&scene.planes)
    } else {
        Vec::new()
    }
}

/// Fixed FOV matching the viewer's own hard-coded value. `SceneState` carries no FOV
/// field of its own (see that struct's doc comment on what it deliberately does and
/// doesn't carry), so this must match `apps/indicatrix-cut/src/bridge/render_thread.rs`
/// and `export_thread.rs`'s `42.0` exactly for a remote worker's camera rays to line up
/// with the viewer's.
const VIEWER_FOV_DEG: f32 = 42.0;

/// Resolves a `--threads`-style argument (`0` meaning "let the OS decide") to an actual
/// thread count.
///
/// Shared by [`trace_samples`]'s own chunking and by `serve`'s `Welcome` message (which
/// reports the thread count it actually renders with, not the literal `0` sentinel).
#[must_use]
pub fn effective_thread_count(threads: usize) -> usize {
    if threads == 0 {
        thread::available_parallelism().map_or(8, std::num::NonZero::get)
    } else {
        threads
    }
}

/// A process-wide counting semaphore over CPU tracer threads -- see this module's own
/// doc comment ("A process-wide permit pool...") for why this exists.
struct ThreadPermits {
    /// Total capacity, fixed at construction (`effective_thread_count(0)`) -- never
    /// mutated afterward; only [`Self::available`] moves.
    capacity: usize,
    /// How many permits are currently unclaimed. Guarded by the same `Mutex` the
    /// [`Condvar`] waits against, per the standard condvar pattern.
    available: Mutex<usize>,
    /// Notified every time [`Self::release`] returns permits, so a blocked
    /// [`Self::acquire`] wakes up to re-check rather than polling.
    freed: Condvar,
}

impl ThreadPermits {
    fn new(capacity: usize) -> Self {
        Self {
            capacity: capacity.max(1),
            available: Mutex::new(capacity.max(1)),
            freed: Condvar::new(),
        }
    }

    /// Blocks until `wanted` permits are free, then claims them, returning how many
    /// were actually acquired (== `wanted.clamp(1, self.capacity)`).
    ///
    /// `wanted` is clamped to `self.capacity` first: a single request asking for more
    /// threads than exist on this machine (or than this pool was ever sized for) must
    /// still be satisfiable -- waiting for MORE permits than the pool's total capacity
    /// would otherwise block forever, since that many can never be simultaneously free.
    /// Clamped to at least 1 too, so a caller can't accidentally acquire zero permits
    /// and skip the accounting entirely.
    fn acquire(&self, wanted: usize) -> usize {
        let wanted = wanted.clamp(1, self.capacity);
        let mut available = self
            .available
            .lock()
            .unwrap_or_else(std::sync::PoisonError::into_inner);
        while *available < wanted {
            available = self
                .freed
                .wait(available)
                .unwrap_or_else(std::sync::PoisonError::into_inner);
        }
        *available -= wanted;
        wanted
    }

    /// Returns `held` permits (as returned by a matching [`Self::acquire`] call) to the
    /// pool and wakes every waiter to re-check -- `notify_all`, not `notify_one`: more
    /// than one blocked request can each now have enough newly-freed permits to
    /// proceed, and `notify_one` could wake the wrong one first and starve the others
    /// for another full cycle.
    fn release(&self, held: usize) {
        let mut available = self
            .available
            .lock()
            .unwrap_or_else(std::sync::PoisonError::into_inner);
        *available += held;
        drop(available);
        self.freed.notify_all();
    }
}

/// The process-wide pool itself, sized once at first use to `effective_thread_count(0)`
/// -- this machine's `available_parallelism`, matching what a single solo request would
/// ask for. Every [`trace_into`] call acquires from and releases back to this same
/// instance regardless of which connection or request it's tracing for.
static THREAD_PERMITS: LazyLock<ThreadPermits> =
    LazyLock::new(|| ThreadPermits::new(effective_thread_count(0)));

/// RAII guard releasing `n` permits back to [`THREAD_PERMITS`] on drop -- including on
/// an unwind, so a tracer thread panicking mid-`thread::scope` (already re-raised by
/// `hybrid::dispatch_concurrently`'s own `resume_unwind`, or a panic surfacing through
/// `run_tracer`'s `catch_unwind`) never leaks permits the pool would otherwise never see
/// again, slowly starving every future request on this worker.
struct PermitGuard(usize);

impl Drop for PermitGuard {
    fn drop(&mut self) {
        THREAD_PERMITS.release(self.0);
    }
}

/// Traces samples `[first_sample, first_sample + samples)` across the whole frame.
///
/// Parallel across `threads` CPU threads (`0` for "all available cores" -- see
/// [`effective_thread_count`]), over `scene.width x scene.height`.
///
/// Returns the summed (not averaged) per-pixel XYZ radiance, one [`Vec3`] per pixel in
/// row-major order. Callers that want a displayable image divide by their own total
/// sample count when tone-mapping; callers accumulating contributions from multiple
/// workers add these sums directly into their own running total -- this function itself
/// never divides.
///
/// Returns an all-[`Vec3::ZERO`] buffer (still correctly sized) if `samples`, `width`,
/// or `height` is zero -- callers are expected to have rejected those via
/// [`crate::validate`] if invalid for their use case; this function has no opinion.
#[must_use]
pub fn trace_samples(
    scene: &SceneState,
    first_sample: u32,
    samples: u32,
    threads: usize,
) -> Vec<Vec3> {
    let width = scene.width;
    let height = scene.height;
    let mut buffer = vec![Vec3::ZERO; width as usize * height as usize];
    if samples == 0 || width == 0 || height == 0 {
        return buffer;
    }

    let camera = Camera::new(scene.yaw, scene.pitch, scene.distance, VIEWER_FOV_DEG);
    let environment =
        scene
            .lighting_preset
            .studio(scene.exposure, scene.light_yaw, scene.light_pitch);

    trace_into(
        scene,
        first_sample,
        samples,
        threads,
        &camera,
        environment,
        &mut buffer,
    );
    buffer
}

/// Traces samples `[first_sample, first_sample + samples)` across the whole frame,
/// preferring `gpu`.
///
/// Falls back to the CPU tracer (via [`trace_into`], the same code [`trace_samples`]
/// runs) whenever `gpu` declines -- a decline is normal, not an error.
///
/// A drop-in replacement for [`trace_samples`] at every call site: same return contract
/// (summed, never averaged), same disjoint-and-additive sample-range property both
/// `render_cmd` and `serve` depend on -- both backends add into a
/// `Vec3::ZERO`-initialized buffer with identical semantics, keeping a GPU worker's
/// sample ranges mergeable with a CPU viewer's.
///
/// A thin wrapper over [`trace_samples_with_gpu_cancellable`] with a cancellation flag
/// that is never set -- kept for `render_cmd` (a one-shot CLI render with no
/// cancellation concept) and tests that predate cooperative cancellation.
///
/// # Panics
///
/// Never panics in practice: the `.expect` below can only fire if
/// [`trace_samples_with_gpu_cancellable`] returned `None`, which means it observed
/// `cancel` set -- impossible here, since `never_cancel` is never set.
#[must_use]
pub fn trace_samples_with_gpu(
    gpu: &GpuBackend,
    scene: &SceneState,
    first_sample: u32,
    samples: u32,
    threads: usize,
) -> Vec<Vec3> {
    let never_cancel = AtomicBool::new(false);
    trace_samples_with_gpu_cancellable(gpu, scene, first_sample, samples, threads, &never_cancel)
        .expect("cancel is never set, so this dispatch can never observe GpuAccumulate::Cancelled")
}

/// Cancellable counterpart of [`trace_samples_with_gpu`], used by
/// `stream_emit::tracer::run_tracer`'s single-engine sub-batch dispatch.
///
/// Used whenever a job hasn't calibrated a hybrid split, so a GPU sub-batch (which can
/// run up to `MAX_SUBBATCH` samples, far more than `TARGET_SUBBATCH`'s wall-clock target
/// on a fast adapter) can be interrupted mid-dispatch rather than only between whole
/// sub-batches, tightening worst-case cancellation latency.
///
/// Returns `None` if `cancel` was observed set before the GPU dispatch finished --
/// `buffer` is guaranteed untouched in that case, so the caller has nothing to fold in.
/// A `Declined` GPU still falls back to the CPU tracer for the full `samples` exactly as
/// [`trace_samples_with_gpu`] does; `cancel` plays no further role once declined.
#[must_use]
pub fn trace_samples_with_gpu_cancellable(
    gpu: &GpuBackend,
    scene: &SceneState,
    first_sample: u32,
    samples: u32,
    threads: usize,
    cancel: &AtomicBool,
) -> Option<Vec<Vec3>> {
    let width = scene.width;
    let height = scene.height;
    let mut buffer = vec![Vec3::ZERO; width as usize * height as usize];
    if samples == 0 || width == 0 || height == 0 {
        return Some(buffer);
    }

    let camera = Camera::new(scene.yaw, scene.pitch, scene.distance, VIEWER_FOV_DEG);
    let environment =
        scene
            .lighting_preset
            .studio(scene.exposure, scene.light_yaw, scene.light_pitch);

    // scene.girdle_frosted re-expanded into Vec<FacetFinish> (see
    // resolve_facet_finishes). The CPU fallback below resolves identical finishes from
    // the same field, so a GPU decline mid-trace can't silently switch polish state.
    let facet_finishes = resolve_facet_finishes(scene);
    let gpu_scene = GpuSceneRef {
        camera: &camera,
        width,
        height,
        planes: &scene.planes,
        facet_finishes: &facet_finishes,
        material: &scene.material,
        max_bounces: scene.max_bounces,
        environment,
    };
    match gpu.try_accumulate_cancellable(&gpu_scene, first_sample, samples, &mut buffer, cancel) {
        GpuAccumulate::Done => return Some(buffer),
        GpuAccumulate::Cancelled => return None,
        GpuAccumulate::Declined => {}
    }

    // The GPU declined (no adapter, ComputeMode::OnlyCpu, or an unsupported material).
    // `buffer` is guaranteed still all-zero -- a decline never partially writes -- so
    // falling back is exactly trace_samples' own zero-initialized start.
    trace_into(
        scene,
        first_sample,
        samples,
        threads,
        &camera,
        environment,
        &mut buffer,
    );
    Some(buffer)
}

/// The CPU tracing loop itself, adding into an already-sized `buffer` (`width x height`
/// entries, row-major). Shared by [`trace_samples`] and [`trace_samples_with_gpu`]'s CPU
/// fallback so both go through the identical seed formula and thread-chunking.
///
/// # Work distribution: a shared atomic row counter, not contiguous row bands
///
/// Rows through the stone cost far more than background rows, so this claims rows
/// dynamically through a shared `AtomicUsize` counter (`fetch_add` per row) rather than
/// splitting the frame into `num_threads` contiguous bands, mirroring the same fix in
/// `apps/indicatrix-cut`'s scanline/batch renderers. `buffer` is pre-split into per-row
/// slices behind a `Mutex<Vec<Option<&mut [Vec3]>>>`; a thread claims row `y`, locks
/// just long enough to take that row's slice, then traces it without the lock held.
/// Every row is claimed by exactly one thread, so per-pixel sums stay bit-identical to
/// the old contiguous split.
fn trace_into(
    scene: &SceneState,
    first_sample: u32,
    samples: u32,
    threads: usize,
    camera: &Camera,
    environment: EnvironmentSource<'_>,
    buffer: &mut [Vec3],
) {
    let width = scene.width;
    let height = scene.height;
    let width_usize = width as usize;

    let requested_threads = effective_thread_count(threads).max(1);
    // Acquire the process-wide CPU-tracer permits before spawning anything -- see this
    // module's own doc comment ("A process-wide permit pool...") and `THREAD_PERMITS`.
    // `acquire` blocks until enough are free (clamping to the pool's total capacity if
    // `requested_threads` exceeds it) and returns exactly how many were granted; that's
    // the number of scoped threads actually spawned below, so held permits and live
    // threads always agree. `_permits` releases them (even on panic) once this
    // `trace_into` call's `thread::scope` returns.
    let num_threads = THREAD_PERMITS.acquire(requested_threads);
    let _permits = PermitGuard(num_threads);

    let planes = &scene.planes;
    let material = &scene.material;
    let max_bounces = scene.max_bounces;
    // Frosted girdle -- see `resolve_facet_finishes`'s doc comment.
    let facet_finishes = resolve_facet_finishes(scene);
    let facet_finishes = facet_finishes.as_slice();
    // Built once for this whole `trace_into` call (one sub-batch, potentially many
    // samples x many rows) rather than once per traced sample -- see
    // `trace_spectral_ray_with_finish_soa`'s doc comment: rebuilding this arena per
    // sample was a heap allocation on every single traced sample, measured at ~10-20% of
    // per-sample cost on a 57-plane cut. `planes` is the exact slice this is built from,
    // as that function requires; results stay bit-identical to the rebuild-every-call
    // `trace_spectral_ray_with_finish` this replaces.
    let plane_soa = build_plane_soa(planes);

    let rows: Vec<Option<&mut [Vec3]>> = buffer.chunks_mut(width_usize).map(Some).collect();
    let rows = Mutex::new(rows);
    let next_row = AtomicUsize::new(0);

    thread::scope(|s| {
        for _ in 0..num_threads {
            let rows = &rows;
            let next_row = &next_row;
            let plane_soa = &plane_soa;

            s.spawn(move || {
                loop {
                    let y = next_row.fetch_add(1, Ordering::Relaxed);
                    if y >= height as usize {
                        break;
                    }

                    let row = rows
                        .lock()
                        .unwrap_or_else(std::sync::PoisonError::into_inner)[y]
                        .take()
                        .expect("each row index is claimed by exactly one thread via fetch_add");

                    for (x, pixel) in row.iter_mut().enumerate() {
                        let global_pixel_idx = (y * width_usize + x) as u32;
                        let mut sample_sum = Vec3::ZERO;

                        // Per-pixel Cranley-Patterson rotations for the stratified
                        // pixel-jitter/hero-wavelength draws below -- pure functions of
                        // `global_pixel_idx`, hoisted out of the sample loop. jx/jy/
                        // hero_rand each use a different prime base (2, 3, 5) rather
                        // than one base rotated three ways -- measurably better
                        // variance for the highest-variance pixels this targets.
                        let rot = pixel_rotations(global_pixel_idx);

                        for s_idx in 0..samples {
                            let sample_num = first_sample.wrapping_add(s_idx);
                            // Stratified, a pure function of the absolute sample index
                            // alone (never batch-relative) -- what lets disjoint worker
                            // sample ranges compose correctly.
                            let draws = sample_draws(global_pixel_idx, sample_num, &rot);

                            let ray = camera.generate_ray(
                                x as f32,
                                y as f32,
                                width as f32,
                                height as f32,
                                draws.jitter_x,
                                draws.jitter_y,
                            );

                            sample_sum += trace_spectral_ray_with_finish_soa(
                                ray,
                                planes,
                                plane_soa,
                                facet_finishes,
                                material,
                                max_bounces,
                                environment,
                                draws.seed,
                                draws.hero_rand,
                                None,
                            );
                        }

                        *pixel += sample_sum;
                    }
                }
            });
        }
    });
}

#[cfg(test)]
mod tests {
    use super::*;
    use indicatrix::{
        geometry::cuts::StandardGemCuts,
        optics::{materials::GemMaterial, raytracer::LightingPreset},
    };

    fn tiny_scene() -> SceneState {
        SceneState {
            width: 8,
            height: 8,
            yaw: 0.4,
            pitch: 0.3,
            distance: 3.0,
            light_yaw: 0.85,
            light_pitch: 0.95,
            exposure: 1.0,
            max_bounces: 4,
            lighting_preset: LightingPreset::Daylight,
            material: GemMaterial::diamond(),
            planes: StandardGemCuts::standard_round_brilliant(),
            girdle_frosted: false,
        }
    }

    #[test]
    #[ignore = "manual probe: reports whether this machine has a usable wgpu adapter"]
    fn probe_gpu_adapter() {
        let backend = GpuBackend::acquire();
        match backend.adapter_label() {
            Some(label) => eprintln!("GPU ADAPTER FOUND: {label}"),
            None => eprintln!("NO GPU ADAPTER (declined)"),
        }
    }

    #[test]
    fn returns_a_buffer_sized_to_the_scene() {
        let scene = tiny_scene();
        let buf = trace_samples(&scene, 0, 4, 1);
        assert_eq!(buf.len(), 64);
    }

    #[test]
    fn zero_samples_returns_an_all_zero_buffer_of_the_right_size() {
        let scene = tiny_scene();
        let buf = trace_samples(&scene, 0, 0, 1);
        assert_eq!(buf.len(), 64);
        assert!(buf.iter().all(|v| *v == Vec3::ZERO));
    }

    #[test]
    fn single_and_multi_threaded_traces_agree() {
        let scene = tiny_scene();
        let single = trace_samples(&scene, 0, 4, 1);
        let multi = trace_samples(&scene, 0, 4, 4);
        for (a, b) in single.iter().zip(multi.iter()) {
            // Thread count only changes chunking, never which (pixel, sample) pairs
            // are traced -- must be bit-exact.
            assert_eq!(a, b);
        }
    }

    #[test]
    fn splitting_a_sample_range_across_two_calls_sums_to_the_same_result_as_one_call() {
        // The additivity property this crate exists to preserve. Relative tolerance,
        // not bit-exact: float addition isn't associative, so a different grouping can
        // differ in the last bit or two of an f32 even when correct.
        let scene = tiny_scene();
        let whole = trace_samples(&scene, 0, 8, 2);
        let first_half = trace_samples(&scene, 0, 4, 2);
        let second_half = trace_samples(&scene, 4, 4, 2);

        for i in 0..whole.len() {
            let split_sum = first_half[i] + second_half[i];
            let diff = (whole[i] - split_sum).abs();
            let scale = whole[i].abs().max(split_sum.abs()).max(Vec3::splat(1e-6));
            let rel = diff / scale;
            assert!(
                rel.max_element() < 1e-3,
                "pixel {i}: whole={:?} split_sum={:?} rel={:?}",
                whole[i],
                split_sum,
                rel
            );
        }
    }
}