I have a few questions regarding UVA. If there’s a document with the answers, feel free to point me at it instead of copying the text.
If GPU 2 is executing a kernel that is accessing memory on GPU 1, can GPU 1 execute its own kernel that doesn’t access the same memory buffer? Can the kernel on GPU 2 be run asynchonously?
If GPU 2 is executing a kernel that is accessing memory on GPU 1, can an asynchronous transfer be performed between GPU 2 and the host? Between GPU 1 and the host?
The quest arises from the synchronous processing that gets performed when a peer to peer copy is performed. I suspect it has nothing to do with UVA, but I’d rather know for sure.