[Mlir-commits] [clang] [llvm] [mlir] [openmp] [OpenMP][offload] Cross-team reductions with variable number of teams (PR #195102)

[Dhruva Chakrabarti]

dhruvachak wrote:

Removing the chunked teams approach may have benefits only when the number of teams launched is > 1024. I think that in many cases, the number of teams used is <= 1024. However, the difference in code complexity may lead to a different number of register usage. Regardless, it would be better to bring in this change as a separate PR, it would definitely help understand the performance impact of this change separately.

https://github.com/llvm/llvm-project/pull/195102