Multiple async memcpy


I was wondering how multiple async memcpy from host to device were handled by nvidia. If the first memcpy isn’t finished when I fire a second one, is it automatically queued ?


If they are in the same stream, yes queuing is guaranteed. If they are in separate streams, you are telling CUDA that it is OK to run the two copies at the same time. Whether it does so or not depends on your system’s capabilities.