[Mlir-commits] [mlir] [mlir][nvgpu] update commit group and wait async ops (PR #130482)

[Guray Ozen]

grypp wrote:

Here is the RFC and code review discussion:  
https://discourse.llvm.org/t/modeling-gpu-async-copy-ampere-feature/4924  
https://reviews.llvm.org/D119191  

It seems this PR's approach (without token) was previously discussed there. However, since it's close to low-level PTX, the design decision was to avoid this approach in favor of using SSA and tokens.

In my personal opinion, there's not much practical difference with or without `token`(I've learned that reordering isn't an issue). Having ssa&tokens is slightly better as it makes dependencies clearer.

Overall, there doesn't seem to be a clear advantage to adopting this PR. Let me know what you think. 



https://github.com/llvm/llvm-project/pull/130482