[Mlir-commits] [mlir] [mlir][gpu] Deprecate gpu::Serialization* passes. (PR #65857)

[Guray Ozen]

grypp wrote:

> Deprecate the `gpu-to-cubin` & `gpu-to-hsaco` passes in favor of the `TargetAttr` workflow. 

Seems reasonable to me. Thank you for doing this.
I no longer use the `gpu-to-cubin` pass since you've implemented the new workflow.

> 2. The [SM_90 integration tests](https://github.com/llvm/llvm-project/tree/main/mlir/test/Integration/GPU/CUDA/sm90) still need to be ported into the new workflow, so this patch is dependent on that porting.

Let me know if you need help here. It seems like removing the 'convert-gpu-to-nvvm' addind `TargetAttr` would be enough. We are already using "TargetAttr" in one test.

FWIW, I'm about to land for 4 more tests. I will make sure to use the new workflow. 

https://github.com/llvm/llvm-project/pull/65857