Open andfau-amd opened 2 months ago
The upstream MLIR gpu.subgroup_reduce operation is quite powerful, especially following the work in https://github.com/iree-org/iree/issues/18142. There are still places in IREE that manually build a series of shuffles that could be replaced with gpu.subgroup_reduce. One example is emitGPUReduction (WIP patch to clean it up: https://github.com/iree-org/iree/compare/main...andfau-amd:18142-use-subgroup_reduce-for-emitGPUGroupReduction).
gpu.subgroup_reduce
emitGPUReduction
is this solved by https://github.com/andfau-amd/iree/commit/4b984f1d56ead80cd7921b01937bfbb1640c01df
@dan-garvey I think the point of this ticket was to do the same kind of improvement in more places than was done in that commit.
The upstream MLIR
gpu.subgroup_reduce
operation is quite powerful, especially following the work in https://github.com/iree-org/iree/issues/18142. There are still places in IREE that manually build a series of shuffles that could be replaced withgpu.subgroup_reduce
. One example isemitGPUReduction
(WIP patch to clean it up: https://github.com/iree-org/iree/compare/main...andfau-amd:18142-use-subgroup_reduce-for-emitGPUGroupReduction).