I am writing wrappers for all CUDA functions including (__cudaRegisterFatBinary(…), __cudaUnregisterFatBinary()…).
And I am planning to disable the __cudaUnregisterFatBinary() call. Inside my own wrapper function, I will just exit before calling the real Cuda function. I am doing this because I am manipulating the process memory in a way that makes the real call to __cudaUnregisterFatBinary() fail.
And since the __cudaUnregisterFatBinary() is called at the very end of the program execution, I am thinking this won’t be harmful.
Would that be fine to do?
Thank you’ all for your answers