GI re-shade: add a significance test and find why ~25 lights count as changed every frame #980

New issue

Closed

opened 2026-10-23 18:07:41 +00:00 by sky_line · 1 comment

sky_line commented

2026-10-23 18:07:41 +00:00

Owner

0.235 re-shades per GI ray (about 1-2 ms). Each re-shade is a full closest-hit ray plus hit shading. giChangedReaches has no significance test: any changed light in range with attenuation above 0 triggers it, and about 25 of 256 lights count as changed every frame at steady state.

Plan

Log which lights are flagged changed every frame (moving vs radiance-changed; RenderVKReSTIRFlickerEps too tight? lights on avatars?).
Add a contribution threshold: re-shade only when the changed light's unshadowed contribution at the stored point is a meaningful share of the stored radiance.

Estimate

1-2 ms, small effort, risk: a slightly stale tint in rare cases.

From the 2026-10-23 GI performance pass. Baseline: RTX 5090, 4K, DLSS Ultra Performance (1280x720 render), GI at full render resolution, GI 3 + cache, ReSTIR DI 3, emissive on: GPU 30.9 ms/frame (GPU-bound), async deferred lighting 20.1 ms, `c_gi_trace` 12.7 ms (883k rays/frame). 0.235 re-shades per GI ray (about 1-2 ms). Each re-shade is a full closest-hit ray plus hit shading. `giChangedReaches` has no significance test: any changed light in range with attenuation above 0 triggers it, and about 25 of 256 lights count as changed every frame at steady state. ## Plan 1. Log which lights are flagged changed every frame (moving vs radiance-changed; `RenderVKReSTIRFlickerEps` too tight? lights on avatars?). 2. Add a contribution threshold: re-shade only when the changed light's unshadowed contribution at the stored point is a meaningful share of the stored radiance. ## Estimate 1-2 ms, small effort, risk: a slightly stale tint in rare cases.

sky_line added the

labels

2026-10-23 18:07:41 +00:00

sky_line added this to the Slipstream project

2026-10-23 18:07:41 +00:00

sky_line referenced this issue from a commit

2026-10-23 18:44:08 +00:00

#980: GI re-shade -- GI's own changed-light rule and a significance test