Page Fault GPU Crash in Lumen Reflections (Hardware Raytracing)

Development Rendering
, ,
21

Thanks for the update!

[Attachment Removed]

22

Hello,

Even after fixing the Residency-related issues, we’re still collecting a small number of GPU Page Faults on the Lumen HWRT path. Aftermath reports that the resource involved — an Upload Heap — was FullyResident at the time, so it seems like the VA had gone stale. Our suspicion is that there are leftover/stale shader records in the SBT (dangling VAs).

Could you confirm whether this is the “other issue” you were referring to (the one to discuss in a separate case)?

Thanks!

[Image Removed]

[Attachment Removed]

23

We also suspected that some stale SBT records were accessed by RT geometry. We cannot locate the exact issue, so we added following protection code.

void ClearHitRecordData(uint32 WorkerIndex, uint32 RecordIndex)
{
......
	// Reset record to default state
	RecordData.Reset();
		
	// FIX BEGIN
	if (EnumHasAnyFlags(ShaderBindingMode, ERayTracingShaderBindingMode::RTPSO))
		SetHitGroupIdentifier(RecordIndex, FD3D12ShaderIdentifier::Null);
	// FIX END
}

Let me know if it helps your situation.

Thanks.

Yichen

[Attachment Removed]

24

Hi Tim,

Thanks! I’ve opened a separate ticket here and linked this one to it:

[Content removed]

I’ve attached the Aftermath crash artifacts there. Unfortunately we weren’t able to obtain the shader PDB for the Closest Hit Shader on our side — I’ve included the DXIL instead, and I’ll follow up on the new ticket if we manage to get the PDB.

Thanks,

Yanchen

[Attachment Removed]

25

Hi Yichen,

Thanks a lot for sharing this.

We’ll take a closer look and try it out on our side.

Thanks,

Yanchen

[Attachment Removed]