[Mlir-commits] [mlir] [MLIR][GPU] Add support for non-portable cluster size attribute (PR #95545)

[Guray Ozen]

grypp wrote:

> Blocking this for discussion.
> 
> Can we stop vendor-specific feature creep to common dialect and have at least some design discussion?

I understand the concern. Here’s my perspective: If this requires a longer discussion, we can continue on Discourse.

`gpu.launch` and `gpu.launch_func` are target-agnostic, but we need architecture-specific options (which are not vendor-specific but vary within the same vendor).

We’ve designed these options to be `optional` so far, so they shouldn't impact other vendors.

What would be your proposal?



https://github.com/llvm/llvm-project/pull/95545