Adding optional CUDA support to an application

Accelerated Computing CUDA CUDA Programming and Performance
1

I am considering adding optional CUDA support to an application and I believe the only way to do this is via the Driver API?

The application must be able to start regardless of whether CUDA is installed or if there is a CUDA capable GPU etc.

I can’t seem to step through debug any of the CUDA driver API examples in NSight - is this a limitation of using the Driver API?

So I am a bit concerned my debugging experience won’t be ideal if I do attempt to use the Driver API? I haven’t come across this specific issue in the nsight or CUDA docs?

2

Oh for some reason I can now step through one of the DriverAPI examples in Nsight - quite possibly I did something wrong before.

And I have now found the part in the guide regards deploying applications(!):
http://docs.nvidia.com/cuda/cuda-c-best-practices-guide/index.html#deploying-cuda-applications

“If cudaGetDeviceCount() reports an error, the application should fall back to an alternative code path.”

I was nervous that if there was no CUDA installed my program would fail to run (so I could choose an alternative non-CUDA path). But it appears that is not the case.

Sorry for missing that in the docs.

3

There’s no particular reason that adding CUDA support to an existing application needs to be done with the driver API. It could be done with the runtime API also.