search

NVlabs / NVBit

198 stars 18 forks source link

Segmentation fault on accessing cuda kernel launch's kernel parameters with different CUDA versions #91

Open William-An opened 2 years ago

William-An commented 2 years ago

Hello all,

I have written a tracer tool to collect the CUDA API call parameters within the function nvbit_at_cuda_event(CUcontext ctx, int is_exit, nvbit_api_cuda_t cbid, const char *name, void *params, CUresult *pStatus).

When I compiled both the tool and my test apps with CUDA 11.0 and GCC 7.5.0, it worked perfectly. However, with a different setting for the apps (CUDA 9.1 with GCC 5.5.0), NVBIT throws segmentation faults which are due to the kernelParams member in the cuLaunchKernel_params being null.

Some faulty apps can be found here:

  1. l1_bw_128
  2. l1_lat
ovilla commented 2 years ago

Hi, it could easily be due to the older version of GCC but I am not sure. We currently don't have that exact configuration around, but we will keep an eye on it for future releases. Thanks for reporting this.