[Mlir-commits] [mlir] [MLIR][XeGPU] Add intel_reqd_sub_group_size attribute to gpu.func to facilitate gpu.shuffle lowering (PR #182434)

[Artem Kroviakov]

akroviakov wrote:

Oh, I thought it already was part of the target attachment pass. In any case, I believe it should not be limited to sg distribution only, since, at least, 2d block ops behavior is sensitive to it (e.g., the infra assumes 16, but the default size at execution is 32), and 2d ops are also present in lane-level code.

https://github.com/llvm/llvm-project/pull/182434