Stuck code reviews (MLIR GPU interfacing)

I’d like to ping folks about ⚙ D112668 [MLIR][GPU] Update SerializeToHsaco to match downstream and ⚙ D110448 [MLIR][GPU] Define gpu.printf op and its lowerings, two pull requests of mine that have acquired a nice coat of dust. (In the case of the latter, there is probably a bit of discussion to be had about whether gpu is the right place for that one and whether maybe we shouldn’t have something like std.printf, but I’d still like the code itself looked at.)

2 Likes

Sorry about this! Seems it fells through the cracks…
I didn’t see the first one, overall it seems fine but likely could be split in a few independent changes.
The second one is mostly about where / how to define a gpu.printf op: is this something that really should be exposed in the GPU dialect?
Do we need instead are more extensible “platform specific intrinsic” mechanism (or namespacing)?

[RFC] Add a printf op is the RFC asking that question