[Mlir-commits] [mlir] [mlir][gpu] Productize `test-lower-to-nvvm` as `gpu-lower-to-nvvm` (PR #75775)

[Guray Ozen]

grypp wrote:

I had two key goals: firstly, to use this pipeline via Python, and secondly, having a universal pipeline to compile `nvgpu, memref, vector, scf, vector->nvvm->cubin/ptx`.

Regarding the `createConvertLinalgToLoopsPass`, it appears to generate loops without parallelization. Considering our integration tests, it might be needed. I can try to exclude it and possibly rename the pipeline if it is your concern.

https://github.com/llvm/llvm-project/pull/75775