Noise audit: half-res GI / shadows under DLSS -- findings to look at #1613

Open
opened 2026-11-18 15:32:46 +00:00 by TheSlowly · 0 comments
Collaborator

A read-through of the GI, shadow denoiser and DLSS Ray Reconstruction paths, each checked against the code. None of these are fixed or tested in-world yet; they're candidates for the "no noise, more FPS" work. (The GI ray-sequence bug split out as #1611.)

DLSS Ray Reconstruction

  • At default settings RR gets no raw signal: RenderVKDlssRawSignals 26 lacks bit 32 and GI / shadow mask scale is 0.5, and GI is raw only at DLAA. Auto still picks RR, so every signal is denoised by us and then again by RR. Measured 2026-11-18 (DLSS Quality, 2560x1440, busy club): Auto 38.5 ms GPU/frame (p_dlssrr 3.34 ms) vs RR off 36.2 ms (p_dlsssr 0.96 ms); our GI denoiser cost the same in both. Idea: Auto picks Super Resolution when no signal would reach RR raw. Image quality not compared.
  • RR is created with preset 0; the responsivity mask may only work on preset F (unverified).

GI

  • GI temporal reprojection ignores last frame's jitter (clean_dn_temporal.comp ~1197: GI doesn't set sampleGridHistory), and so does the lag fetch (clean_deferred_light.comp ~3962, comment at ~3838 says so). History is read up to half a pixel off: blur + edge shimmer, at full res too.

Shadows

  • RenderVKReSTIRRadianceMoveReset (described as a transient A/B) defaults on: wherever a moving / flickering light is 5 %+ of the light, shadow history resets each frame (clean_dn_temporal.comp ~1293). Try off in a club.
  • ReGIR population 1 (the ReSTIR set) still draws one hashed xi per cell per frame (clean_regir_build.comp ~138): cells reached by more than 16 lights flicker as a whole; population 0 already has the stratified xi.
  • Half-res RT mask read inverts the jitter linearly, not the floor'd sample pixel (clean_shadow_mask_read.glsl ~91): up to ~0.5 px wobble (merge off / mover mask only).

Not noise, but FPS: g_p2_post_static was 10.8 ms of 38 ms in the same capture.

A read-through of the GI, shadow denoiser and DLSS Ray Reconstruction paths, each checked against the code. None of these are fixed or tested in-world yet; they're candidates for the "no noise, more FPS" work. (The GI ray-sequence bug split out as #1611.) **DLSS Ray Reconstruction** - [ ] At default settings RR gets no raw signal: RenderVKDlssRawSignals 26 lacks bit 32 and GI / shadow mask scale is 0.5, and GI is raw only at DLAA. Auto still picks RR, so every signal is denoised by us and then again by RR. Measured 2026-11-18 (DLSS Quality, 2560x1440, busy club): Auto 38.5 ms GPU/frame (p_dlssrr 3.34 ms) vs RR off 36.2 ms (p_dlsssr 0.96 ms); our GI denoiser cost the same in both. Idea: Auto picks Super Resolution when no signal would reach RR raw. Image quality not compared. - [ ] RR is created with preset 0; the responsivity mask may only work on preset F (unverified). **GI** - [ ] GI temporal reprojection ignores last frame's jitter (clean_dn_temporal.comp ~1197: GI doesn't set sampleGridHistory), and so does the lag fetch (clean_deferred_light.comp ~3962, comment at ~3838 says so). History is read up to half a pixel off: blur + edge shimmer, at full res too. **Shadows** - [ ] RenderVKReSTIRRadianceMoveReset (described as a transient A/B) defaults on: wherever a moving / flickering light is 5 %+ of the light, shadow history resets each frame (clean_dn_temporal.comp ~1293). Try off in a club. - [ ] ReGIR population 1 (the ReSTIR set) still draws one hashed xi per cell per frame (clean_regir_build.comp ~138): cells reached by more than 16 lights flicker as a whole; population 0 already has the stratified xi. - [ ] Half-res RT mask read inverts the jitter linearly, not the floor'd sample pixel (clean_shadow_mask_read.glsl ~91): up to ~0.5 px wobble (merge off / mover mask only). **Not noise, but FPS:** g_p2_post_static was 10.8 ms of 38 ms in the same capture.
TheSlowly added this to the Slipstream project 2026-11-18 15:32:46 +00:00
Sign in to join this conversation.
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
sky_line/Slipstream#1613
No description provided.