Failed code 7052 with optixDeviceContextCreate

[sway]

Hi there,
I came across with a weird error after upgrading the driver from 495.46 to 515.76 on RTX 3060. The error code says OPTIX_ERROR_CUDA_NOT_INITIALIZED, but it confuses me cause’ I have a CUDA context already. Is there something wrong with my following code? Thanks.

Sway

cudaSetDevice(0);//using device 0 for cuda context
CHECK_CUDA_ERROR()     
CUresult  cuRes = cuCtxGetCurrent(&cudaContext);
if( cuRes != CUDA_SUCCESS ) {
  printf("failed to get CUDA context!\n");
}
OptixDeviceContextOptions options = {};
options.logCallbackFunction       = &context_log_cb;
options.logCallbackLevel          = 1;
OPTIX_CHECK( optixDeviceContextCreate( cudaContext, &options, &optixContext ) );

Other info:
Ubuntu 20.04, CUDA 11.5, OptiX 7.3~7.5

[dhart]

Hi @sway, welcome!

Does this code work as-is with the older driver? (You can re-install the old one if you need.) You’re mixing the CUDA runtime API and the driver API here, it might be worth double-checking that it breaks if you switch to only runtime API CUDA calls. In our SDK samples, the CUDA initialization is typically done via a call to cudaFree(0) before using OptiX- does that style of CUDA initialization work for you?

Do you have multiple GPU devices in your system? (Looking at the cudaSetDevice(0) call, I’m curious if the hard-coded 0 is guaranteed to be correct on your system, or if the device ordering could have changed.)

Are you able to run a CUDA kernel before trying to initialize OptiX?

–
David.

[sway]

Hi David,

Thanks for your suggestions. Finally, I found the issue. I wanted to ship my executable with the shared libs, so I put libcuda.so.515.76 and its symbolic link libcuda.so.1 alongside my executable. Now, I am sure the error is related to the module loading. But I didn’t find any major difference when I tried to ‘strace’ the lib loading.

The error can be reproduced by just copying libcuda.so.xxx and its symbolic link to the lib directory of SDK samples.

best,
Sway