[yujiaxu-ops]

The kernel used to draw white noise from a single PCG stream, so a draw's position in that stream depended on which branches the path took. It now uses shuffled, Owen-scrambled Sobol (Burley 2020, the same design as Cycles' Sobol-Burley), in shaders/sample.wgsl.

• Dimensions. The sequence has four Sobol dimensions; higher dimensions come from padding by dimension id. Each decision at a path vertex has its own id:

  • rng_dim builds the id from the path, the vertex and a named slot (D_*);
  • rng_sub builds the id for one iteration of a loop or for one light.

  The id seeds the scramble.

• Sample index. Offline, the index is the sample number; live, it is the frame. It is 16 bits wide.

• Offline: every draw comes from Sobol, except the steps of delta tracking. Those form a loop with no fixed length, so each step draws from a hash instead, as pbrt-v4 does.

• Live: only the camera path's first vertex uses Sobol; everything else uses a hash, as NVIDIA's RTXPT does. Using Sobol for every draw cost 2.1 ms of Bistro's 8.4 ms live trace.

• Sobol dimensions 1–3 are unrolled, with the direction numbers written inline.

• The loop guard now skips counted loops, i.e. for (var i = N; i < M; i++) where N and M are literals. Tint does the same. Without this, the sampler's own loops would use up the guard's iteration budget.

• sample.rs is the same sampler on the CPU. Its tests check:

  • the (0, m, 2)-net property;
  • the direction numbers against Cycles' table;
  • the WGSL against the Rust version on the GPU.

Checks

• Offline convergence. Raw HDR at 320×180 with no denoiser. Each build's error is measured against the other build's render at 4096 spp. The table gives RMSE of this branch divided by RMSE of main:

  spp	Cornell box	Light Room	Bistro
  1	0.99	0.97	1.00
  4	0.49	0.67	0.63
  16	0.33	0.37	0.44
  64	0.29	0.38	0.35
  256	0.32	0.46	0.32

  At 4096 spp the means of the two builds agree within 0.01 %.

• Live cost. Bistro's trace takes the same time as on main: 8.60 ms against 8.61 ms natively, 47.5 ms against 47.6 ms merged.

• Live quality. bench-window shows the same results on every shot:

  • still Bistro: dE 1.82 on both;
  • Kabukichō: dE 4.64 against 4.63;
  • slice and the three Tokyo shots: within 0.03 dE and 0.002 SSIM.

• Atomic tests. test-render passes 190/190 natively. Merged, the same 24 known software-traversal failures as main, with identical names.

• Unit tests. cargo test --release with the guard: 142 passed.

• Browser. web-hello passes in Chrome.

Fixed after the first review (6e670ad):

• The render seed (THREERT_SEED) had stopped affecting renders. It is mixed into each pixel's seed again.
• Delta tracking now draws from a hash. With Sobol, a dense grid's region mean varied 0.9 % between seeds at 4096 spp; with the hash it varies 0.3 %, the same as white noise.
• The parity test declares LIVE (Bugbot).
• volume/lamp-inside now expects 6.72, its converged value. Both samplers give 6.70–6.74 at 32768 spp. The old 6.607 came from the original repository and was never converged: the build from the day of the extraction also reads 6.697.

With these fixes, test-render passes 190/190 natively, and merged shows the same 24 failures as main.

The blue-noise pixel distribution (Cycles 4.2) comes in its own PR.

🤖 Generated with Claude Code

---

Note

Medium Risk
Wide-reaching changes to how every path-tracing random draw is produced, though live cost is unchanged and tests/parity checks cover the new sampler.

Overview
Replaces the path kernel’s PCG white-noise stream (draw order tied to branch paths) with shuffled, Owen-scrambled Sobol (Burley 2020 / Cycles-style) in new sample.wgsl, compiled into the kernel via KERNEL_FILES.

Each stochastic decision now uses a stable dimension id (rng_dim / rng_sub and D_* slots) so samples stratify by index offline (per spp) and live (per frame). Offline draws are Sobol everywhere; live uses Sobol only on the camera path’s first vertex and hashed RNG elsewhere (RTXPT-style) to avoid trace cost. pathtrace.wgsl and translucent.wgsl drop rnd()/rng_state and thread rd through lighting, BSDF, media, cache, and fog paths.

Adds sample.rs with matching CPU RNG plus (0,m,2)-net, table, GPU parity, and decorrelation tests. The loop guard skips provably finite literal for loops so Sobol’s own loops don’t burn the guard budget. Docs (RUNTIME.md, ROADMAP.md) document the scheme; volume/lamp-inside tightens its offline radiance expectation after better convergence.

^{Reviewed by Cursor Bugbot for commit 6e670ad. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Reviewed by Cursor Bugbot for commit a37eb18. Configure here.}

[cursor]

GPU sampler test omits LIVE

Low Severity

The standalone shader in the_shader_is_this includes sample.wgsl, which reads LIVE inside rng_white, but never declares that override. pathtrace.wgsl provides it in the real kernel; this test module does not. On a GPU the module fails to compile, so the CPU/GPU sampler check never runs; without a GPU the test skips and the mismatch stays hidden.

 

^{Reviewed by Cursor Bugbot for commit a37eb18. Configure here.}

[yujiaxu-ops]

Right: since rng_white the test's shader needed LIVE, and the test only runs where there is a GPU. Fixed in 6e670ad (override LIVE = false); it passes on this Mac's GPU again.