Work around AMD's broken Linux drivers when it comes to pinned memory and base_vertex usage. It seems that using pinned memory with base_vertex disabled is quicker than the other way around.
