I am wondering if it is possible somehow to specify the stream for a code block with array computations that rely on CuArray’s broadcasting capability as e.g.
A .= A .+ B in this snippet:
A = CUDA.zeros(2,3)
B = CUDA.ones(2,3)
A .= A .+ B
There currently isn’t. We need to set-up some global but task-local state to set/get the stream, and make all functions (like
broadcast use that). For now though, we’ve switched to using the implicit per-thread stream, so if you perform those computations on a separate thread you should get the same effect.
Thanks @maleadt for the reply. However, we would like to run these computations on a high priority stream, whereas implicit per-thread streams are normal priority, I assume.