Proposal to add stream/queue as an optional argument to few GPU dialect ops

[stellaraccident]

@mehdi_amini

That said it is perfectly reasonable to start with the types to model the concept (Stream) without having all the operations to manage it, in particular in a multi-device kind of settings. Operations can always be added later when there is more clarity on “what are the appropriate abstractions” there.

I wasn’t necessarily suggesting to punt completely but to isolate. Those aspects are far from standard in all of the programming models, and in MLIR, we usually model such differences with dialects vs lowest common denominator. There’s also substantial priors for these items just coming from a related runtime library with some form of handle interop.

@mehdi_amini

Actually, while “stream” exists in Cuda and this proposal is likely a good match, I’m more curious about how this would all map to the Vulkan model?

Not very well, I suspect, but these are already at two different levels of abstraction. Would need to check with the vulkan folks regarding how much of the GPU dialect really matches that (vs just happening to have some correspondence that can be used in simple cases).