Andrea Faulds

I'm thinking of adding this part to the description of `gpu.subgroup_reduce`: > If a `cluster_size` is provided, the subgroup is divided into clusters of `cluster_size` lanes each, a reduction is...

I think I have a finished implementation of the MLIR part now: https://github.com/llvm/llvm-project/pull/104851

When I got to trying to update `GPUNestedLayoutDistributionPatterns.cpp`, me and @Groverkss realised that being able to specify cluster sizes isn't enough, we also need to be able to specify cluster...

I made a patch to add "cluster strides" to MLIR (https://github.com/llvm/llvm-project/pull/107142), and I've successfully prototyped using this in `GPUNestedLayoutDistributionPatterns.cpp` downstream; I was able to get the existing test to pass...

Second round of MLIR changes is now merged. Using this in IREE will require a few different commits (will update this list as I go along): - https://github.com/iree-org/iree/pull/18443 (replacement for...

Main PR is up at https://github.com/iree-org/iree/pull/18515 and I made an issue for tracking potential follow-up cleanup work (https://github.com/iree-org/iree/issues/18516); I realised I shouldn't block this issue on it.

Third round of MLIR changes, to fix a problem encountered in https://github.com/iree-org/iree/pull/18515 (difficult to use the upstream patterns in the pipeline optimally because they're too general): * https://github.com/llvm/llvm-project/pull/109015 * https://github.com/llvm/llvm-project/pull/109158

Okay, I'll try to find time to look at it today. FWIW I doubt this has any relation to my recent fptrunc-related changes, but I'm still happy to look at...

https://github.com/llvm/torch-mlir/issues/3527 is relevant