[Mlir-commits] [mlir] [mlir][gpu] Improve `gpu-lower-to-nvvm-pipeline` Documentation (PR #77062)

[Guray Ozen]

grypp wrote:

> > The pipeline requires explicitly parallel IR and doesn't do parallelism to any of the operations especially for linalg. Does this explanation align with our goals?
> 
> I would think that "gpu" is already parallel code (the pipeline is "gpu-to-nvvm" here): so it seems to me that handling of linalg should be done outside of this pipeline and this pipeline invoked on the results.

I see yor point, it makes sense. I removed the lowering of linalg. In integration tests I added `convert-linalg-to-loops` before `gpu-to-nvvm-pipeline`.

https://github.com/llvm/llvm-project/pull/77062