[Mlir-commits] [mlir] [mlir][gpu] Store the all-reduce result from the first lane only (PR #221532)

[Fabian Mora]

https://github.com/fabianmcg requested changes to this pull request.

Thank you for the PR! The comments in the lowering are wrong, the lowering uses `xor` shuffles which have the property of delivering the reduced value to all threads (see https://nvlabs.github.io/cuda-oxide/advanced/warp-level-programming.html), so it's not a coincidence it works. If you want you can amend the patch to fix the docs.

https://github.com/llvm/llvm-project/pull/221532