ispass-2009-LIB with record_reg_vals tool crashes in SM_120

[rodhuega]

Hello,

I am trying the new 1.7.5 release. Despite it fixes the problem of decoding instructions reported in #148 . It crashes during the execution of the "record_reg_vals" tool included with the release. Furthermore, it crashes with a similar tool such as the accel-sim tracer, but as it crashes with the "record_reg_vals", I guess that is an internal issue of NVBIT instead of an implementation problem of the tracer.

The problem just happens with the tool "record_reg_vals"; the other tools included with the release work perfectly with that benchmark.

Here I attach the benchmark (ISPASS_LIB.tar.gz).

I have tested the benchmark in SM_86 and SM_120, it only crashes with SM_120.

[x-y-z]

I cannot reproduce it on CUDA 12.9 and r575.57.08 on a 5080. I compiled your program with CUDA 12.9's nvcc and ran it with record_reg_vals. Can you share your error message?

[rodhuega]

Sorry for the delay, I had that machine busy. My driver version is 570.169 desktop open and with CUDA 12.8. The error message is

ispass-2009-LIB: record_reg_vals.cu:260: void nvbit_at_cuda_event(CUcontext, int, nvbit_api_cuda_t, const char*, void*, CUresult*): Assertion `0' failed.
Abortado (`core' generado)

The tracer tool of Accel-sim reports a similar message with CUDA 12.8.

However, after your message, I have tried with CUDA 12.9, and the behavior is different. The record_reg_vals tool works, but the Accel-sim tracer fails as it does in CUDA 12.8. So, I am still thinking that there is an internal bug inside NVBIT, because it is strange that record_reg_vals works with CUDA 12.9 but not with CUDA 12.8.

[x-y-z]

The error message is truncated, since the most important part Kernel launch error: ... is not there. There is no way of figuring out the issue.

If record_reg_vals works on CUDA 12.9 but Accel-sim tracer fails, the problem is in Accel-sim tracer.

[x-y-z]

OK, I tried record_reg_vals on CUDA 12.8 and r570.169 with a 5080 and saw no issue from my side.