Question

so if i want do Unified Virtual Address (UVA) copy between two GPUs (using cudaMemcpyAsync with kind argument is cudaMemcopyDefault), whose stream should I use? stream of device of the source memory? or stream of dev of the destined memory?

thank you

Answer 1

Suggestion: use cudaMemcpyPeerAsync instead. See this question as an example.

I guess to answer your question, from here:

Stream and Event Behavior

A kernel launch or memory copy will fail if it is issued to a stream
that is not associated to the current device

So choose a stream that corresponds to the device that corresponds to the most recent cudaSetDevice() call that you made.