Port the opportunistic synchronization from CUDA.jl

Open christiangnrd opened this issue 1 year ago • 1 comments

It would probably be good to port the opportunistic synchronization from CUDA.jl

Originally posted by @maleadt in https://github.com/JuliaGPU/Metal.jl/pull/314#discussion_r1521559352

EDIT by maleadt: Specifically, the tracking of which stream has recently modified a buffer, and whether to synchronize or not.

Mar 12 '24 15:03 christiangnrd

In addition, copyto! should be asynchronous when possible, e.g., when copying between MTLArrays. As suggested by @tgymnich, this could probably be done by always using the MTLBlitCommandEncoder.

Mar 13 '24 11:03 maleadt

https://github.com/JuliaGPU/Metal.jl/issues/443

Oct 02 '24 08:10 maleadt