raikonenfnu wrote: This looks very similar to the IREE VectorReduction to shuffle lowering? Is the plan to lower vector.multi_reduce to gpu.subgroup_reduce and use this lowering instead? https://github.com/llvm/llvm-project/pull/76530