Official cuBLAS documentation contains some information about pointer mode: http://docs.nvidia.com/cuda/cublas/#scalar-parameters
But as far as I understand this section refers only to scalar parameters. And what about vector parameters?
I’ve tried to set pointer mode to CUBLAS_POINTER_MODE_HOST and pass a vector dynamically allocated in device memory to some functions.
Everything works fine. I’m aware of the fact that some function calls are no longer asynchronous.
So my question is: are there any other pitfalls? For example, I don’t want my whole vector to copied to some other place.
Any help would be greatly appreciated.