[Mlir-commits] [mlir] [mlir][linalg/scf/transform] scalable tiling and fusion for pack/unpack ops (PR #204007)

[Andrzej Warzyński]

https://github.com/banach-space approved this pull request.

I am posting comments for commits 3 and 4, which concludes my review for now 😅 

Btw, it would be great if we could avoid numeric values in `{inner_tile_alignments = array<i64: 2, 2>}` and instead could use some symbolic names. That's a nice-to have to and can be left as an improvement for later.

All in all, while it's a large change, most of the newly added code are tests and otherwise it's just plumbing things together. It's actually pretty clear and the way you split this into 4 commits really helped me review it, so massive thank you!

One big downside of this approach is the fact that extra logic is added specifically for `linalg.pack` and `linalg.unpack`. However, as i said earlier, those two Ops are special and IMHO require some compromise to fully support for dynamic shapes. This PR is a very healthy compromise and really neatly executed. All in all LG!

https://github.com/llvm/llvm-project/pull/204007