Aftermath SDK 32-bit seems not working & an TDR issue

Developer Tools Nsight Aftermath SDK
1

Hi, i’m tracing down an TDR issue only reproducible on 32-bit dx12 program. It seems i can’t properly initialize the Aftermath SDK 32-bit version, it just always fails the GFSDK_Aftermath_DX12_CreateContextHandle call, after a successful GFSDK_Aftermath_DX12_Initialize call. The return value is alwaysGFSDK_Aftermath_Result_FAIL_Unknown. I’ve tried the aftermath sdk sample code on github and they fail with the same error code too. I’m curious what could be the problem.

(By the way, the aftermath monitor can capture the original TDR issue, it says there’s 0000000 gpu memory addr page fault within internal shader compute01, i just can’t use aftermath sdk event markers to narrow down the possible issue point)

(I’ve also attached the TDR crash dump as attachment at this github issue report: DX12 Device Removal issue on nvidia device · Issue #670 · DiligentGraphics/DiligentCore · GitHub)

2

Hi crlf0710,

Thank you for using Aftermath SDK and providing your feedback. We are sorry for any inconvenience you have encountered.

It seems i can’t properly initialize the Aftermath SDK 32-bit version, it just always fails the GFSDK_Aftermath_DX12_CreateContextHandle call, after a successful GFSDK_Aftermath_DX12_Initialize call. The return value is alwaysGFSDK_Aftermath_Result_FAIL_Unknown . I’ve tried the aftermath sdk sample code on github and they fail with the same error code too. I’m curious what could be the problem.

Yes, we can reproduce the similar issue on Aftermath SDK samples on github (32bit specific and you need to change the build config). An internal case has been created to track this.

Thanks
An

4

@AYan Thanks! Beside this and not related to aftermath itself, could you also supply more information around the nv-gpudmp file we posted within the linked github issue?

It’s really hard to understand from outside why there could be a compute01 shader issue there from a seemingly innocent program, the only thing it does is to switch between two PSOs and generate a lot of draw calls continously. We cannot repro this issue with other chips, only a NVIDIA GeForce GTX 1050 Ti chip card with newest gpu driver installed. And this issue is 32bit DX12 only. And if we insert WaitForIdle() calls, the issue disappear (but of course we can’t do this in production).

5

Hi crlf0710,

According to the nv-gpudmp file, it’s some page fault during GPU Memory Write in internal shader. compute01 is just the name of internal shader.

Your crash doesn’t look like an Aftermath SDK issue, maybe you should post that on some other topic?

Thanks
An

1 Like