[Mlir-commits] [mlir] [mlir][vector-to-gpu] Do not convert ops masked by vector.mask (PR #226548)

[Fabian Mora]

fabianmcg wrote:

In general LGTM, the limitation is real and cannot be easily addressed for the GPU (nvgpu==false) path, however, the `nvgpu` path relies on `vector.store` to do the store back which does allows masking, thus masking could be made to work on that path (I don't think it would be worth it though). Please document the asymmetry in the commit description.


https://github.com/llvm/llvm-project/pull/226548