ROOM_PERFORMANCE.md

Live room-water performance pass

Baseline: e083618 (including the latest white-grid minification filter and rolling-safe caustic history). This is the room-wide liquid scene, not the old pit benchmark. Measurements and image checks are recorded in room-performance-validation.json.

Measured result

Pooled medians from 804 timed real frames per version/view:

View	Camera GPU ms, before → after	Renderer ms, before → after	Raw throughput gain	Renderer p95 ms, before → after
Stationary gameplay	8.011 → 7.805	13.563 → 13.546	effectively unchanged	16.456 → 15.278
Orbit	6.372 → 5.188	12.366 → 11.717	5.5%	15.201 → 14.338
Rolling / jumping	6.605 → 6.025	12.842 → 12.273	4.6%	15.627 → 14.792

Camera-pass reductions are 2.6%, 18.6% and 8.8%, respectively; total-frame improvements are smaller because simulation, reconstruction, RR, AS updates and submission remain. The stationary median-frame difference is within run noise, not a meaningful speedup. The p95 interval improves by approximately 5–7%.

Scope and retained change

The dominant measured pass is camera transport through the glass/water scene; the APIC simulation is about 2 ms, surface reconstruction about 1.0–1.2 ms, and spectral photons about 0.7 ms at these settings. No solver iteration counts, particle counts, field spacing, photon budgets or optical bounce limits are reduced.

Boolean visibility rays now use ACCEPT_FIRST_HIT_AND_END_SEARCH plus SKIP_CLOSEST_HIT_SHADER. They stop on any valid occluder and do not request surface attributes. A hit leaves a hit sentinel in the payload; the existing miss shader clears it. This follows the DXR ray-flag contract. Procedural water and whitewater intersections are not skipped. Masks, ray offsets, segment limits and the rule that photons own refracted illumination are unchanged. This covers direct-light visibility, photon launch blocker checks and RTXDI PT reconnection visibility across plain DXR, standard SER and NVAPI. Camera/continuation/photon transport rays still find the closest surface.

The unchanged scalar field, normal filter, spectral estimator, atlas histories, DLSS guides and RR evaluation remain in use. There is no new temporal denoiser, frozen liquid, reduced-quality water setting or reliance on generated frames.

Other experiments were discarded: shared normal taps were slower; wave-reduced diagnostic counters and an inline reflection-stack rewrite showed no reliable gain; primary-hit reuse caused a small raw-radiance mismatch in the deterministic optics comparison. None of these experiments is in the final shader.

Reproduce

Before changing the shader, save the six original Transport-*.dxil libraries to a separate directory. Keep the executable, non-Transport kernels and scene assets identical. Then compile the candidate with compile-shaders.ps1 and run:

.\profile-room-paired.ps1 -BaselineShaderDir C:\path\to\original-shaders -Tag visibility-final
.\test-render-performance.ps1 -BaselineShaderDir C:\path\to\original-shaders
.\test-water-temporal.ps1 -Name water-temporal-performance
.\test-water-temporal.ps1 -Name water-temporal-performance-foam -Whitewater

node engine/validate-room-performance.mjs /path/to/runtime visibility-final
node engine/validate-water-temporal.mjs /path/to/runtime water-temporal-performance water-temporal-final --preserve
node --test engine/*.test.mjs

The paired runner saves/restores the candidate libraries even if a test fails, refuses to overlap another lab/game process, and records executable/shader hashes and actual run order. The baseline is compiled original code, not a runtime reference branch that changes the shader's register allocation.

Settings: RTX 5090, output 1920×1080, DLSS Balanced (1114×626 internal), FG off, 100k initial particles / 250k capacity, live wall inlet and foam/bubbles, 65,536 spectral photons, 120 Hz APIC, 120 pressure and 60 density iterations, two simulation substeps per real frame. Each run has 300 frames, of which the first 32 are excluded. Three adjacent before/after pairs per view; order reverses in the second pair. There are no per-cell GPU validation probes in timed runs.

The reported renderer interval includes submission, present and the GPU fence, but excludes the outer gameplay update. Its reciprocal is not presented FG FPS. Column medians need not sum. Live APIC summation order and GPU scheduling vary between runs, so per-run measurements are included, not just a best case.

Static sphere, thin sheet and ordinary gameplay fixtures compare raw radiance, all DLSS guides, RR output and both caustic atlases on four backend/atomic paths. The raw-color aggregate error limit is 0.001%, RR 0.5%; geometry/material guides and irradiance also use a 0.001% limit. This is bounded numerical agreement, not a claim of bit-identical learned reconstruction. The camera event totals and procedural entry/exit/root probes are checked independently. The separate scripted orbit/top-down/rolling sequence checks temporal behavior with live and paused liquid; its primary-guide reprojection is not ground-truth refracted motion.

Windows Graphics Tools / GPU-based validation was unavailable on this setup; successful ordinary runs do not constitute a D3D12 debug-layer validation pass.

Validation result

All six transport libraries and the other kernels compile; 54 Node tests and five Windows CTests pass. The 12 deterministic before/after image comparisons pass: maximum raw relative L1 error is below 6e-9, all geometry/material/motion/ specular-distance guides match exactly, camera event/truncation counts match, and RR aggregate difference stays below 0.215%. Float photon sums differ only by accumulation-order rounding; fixed-point comparisons are exact. All tested procedural fluid roots and glass-shell entry/exit probes pass.

The 320-frame orbit/top-down/rolling sequence passes with and without whitewater, using --preserve against the pre-optimization corrected sequences. RR variation changes by less than 1%; mean luminance by less than 0.11%. Water history stays at ten resets while ordinary rolling invalidates the static atlas 58 times; global RR resets once. The previous flicker fix is retained, not redefined as zero residual flicker. Room before/after captures were also visually inspected.

Five additional live-room GPU cases pass with fixed atomics, standard SER, NVAPI SER, 2× FG and optional RTXDI PT (540 real frames). Captured guides are finite, photon energy is bounded, and fluid/whitewater probes report no bad hits. These probe-instrumented runs are correctness checks, not the timings above.