Intermittent white / pink / noisy flashes for a few frames (probe-like corruption) #777

New issue

Open

opened 2026-10-15 03:01:53 +00:00 by sky_line · 0 comments

sky_line commented

2026-10-15 03:01:53 +00:00

Owner

History / what's been ruled in or out

Light-cluster overflow (white tile blocks with many shadowed projectors) was fixed in f9c171b048 / 11b7762103 -- the flashes continued after that, so it's not (only) light clusters.
Probe blend seam fix 0037c20d12 (no hard seam at automatic probe edges / top-N swaps) -- owner not convinced it's fully fixed and believes it can still be broken.
FSR mode change DEVICE_LOSTs were a stale Hi-Z descriptor (fixed 3c0f1cbbc2 / 273b17321d) -- different symptom.
The #VKWhite# resolveTexture warnings in the log are startup UI placeholders, unrelated.
No log line marks the event today: there's no telemetry for it.

Next steps

Add an always-on, cheap detector: e.g. a GPU-side NaN/Inf / out-of-range luminance counter on the HDR target, the probe atlas (sample a few texels per probe face after filter/encode) and the lit output, logged with the frame number and the passes active (probe job, realtime probe, mirror, env reset) when it fires; optional frame-dump hotkey/auto-capture on detection.
Suspects: probe capture/filter/BC6H encode writing NaN or garbage into a probe slot that shading samples for a few frames (recapture-on-change work #772 changed scheduling today), probe slot reuse/eviction races, uninitialised probe texels on slot (re)allocation, RT-traced probe faces (detail 3), descriptor/resource reuse across frames in flight.
Correlate with probe scheduler events (new VKProbe line) and the RT cache / static carry state.

Intermittent: for a handful of frames, objects or the whole scene flash white / pink / noisy-colourful, then return to normal. Seen several times on 2026-10-13/14 dev builds (RTX 5090, RT on), most recently 2026-10-15 ~03:00 UTC on e1b51ff8cf after a camera-still period. Looks similar to the earlier reflection-probe corruption. ## History / what's been ruled in or out - Light-cluster overflow (white tile blocks with many shadowed projectors) was fixed in f9c171b048 / 11b7762103 -- the flashes continued after that, so it's not (only) light clusters. - Probe blend seam fix 0037c20d12 (no hard seam at automatic probe edges / top-N swaps) -- owner not convinced it's fully fixed and believes it can still be broken. - FSR mode change DEVICE_LOSTs were a stale Hi-Z descriptor (fixed 3c0f1cbbc2 / 273b17321d) -- different symptom. - The `#VKWhite#` resolveTexture warnings in the log are startup UI placeholders, unrelated. - No log line marks the event today: there's no telemetry for it. ## Next steps - Add an always-on, cheap detector: e.g. a GPU-side NaN/Inf / out-of-range luminance counter on the HDR target, the probe atlas (sample a few texels per probe face after filter/encode) and the lit output, logged with the frame number and the passes active (probe job, realtime probe, mirror, env reset) when it fires; optional frame-dump hotkey/auto-capture on detection. - Suspects: probe capture/filter/BC6H encode writing NaN or garbage into a probe slot that shading samples for a few frames (recapture-on-change work #772 changed scheduling today), probe slot reuse/eviction races, uninitialised probe texels on slot (re)allocation, RT-traced probe faces (detail 3), descriptor/resource reuse across frames in flight. - Correlate with probe scheduler events (new VKProbe line) and the RT cache / static carry state.

sky_line added the

labels

2026-10-15 03:01:53 +00:00

sky_line added this to the Slipstream project

2026-10-15 03:01:53 +00:00