Can't instrument relocatable device code

[AKKamath]

I have a program written in main.cu with a kernel that is executed calling some device functions declared in header.cuh. In header.cu, I've given the definitions.

When I compile this using -rdc flag, and try instrumenting the code, none of the instrumented device code gets executed.

When I use verbose mode during the instrumentation process, the instructions inside header.cu show up, meaning NVBit is aware that they exist. However, the instrumented device function never gets executed if the code is inside header.cu.

[ovilla]

We never seen anything like that, but we did not test specifically for it. Could you please attach your simplified example application so we can try reproduce on our side? Meaning a simplified version of these files:

Makefile main.cu header.cuh header.cu

Thanks!

[AKKamath]

Thanks for the quick response!

NVBit_Example.zip

I've attached the files. You can try running the mem_trace tool on it.

My observation is the code within the device function doesn't get instrumented, but if I replace the function call with the code inside the function it gets instrumented properly. Moving the device function to the main.cu file also allows it to get instrumented.

[ovilla]

You are right, I see the same issue. I am surprised we never discovered this before. Thanks a lot for reporting this, we will look into it as soon as possible.

I can't estimate how long it will take to solve it, but for sure it is something we have to address or at least document.

[AKKamath]

Alright. Thanks again for the response.

[x-y-z]

It is fixed in NVBit 1.4.