[Mlir-commits] [mlir] [MLIR][Vector] Relax shape_cast unrolling to per-reassociation-group contiguity (PR #205684)

[Andrzej Warzyński]

banach-space wrote:

> @banach-space does my explanation make sense to you? For HW target which has the issue (extract_strided_slice non-contiguous high-d vector lowering to complex ops like shuffling), it should choose proper tile size avoiding it. But vector unrolling should support the case which can be properly use vector.extract_strided_slice and lower to NOP, and this PR exactly does so.

Yes, sorry for the delay replying!

Your reply addresses my concerns. It would be good to add comments in code to re-enforce these assumptions/guidelines, i.e. this unrolling should not be used by targets for which `tensor.extract_strided_slice` is a concern and `vector.shape_cast` should be treated as NOP.

Otherwise LGTM

https://github.com/llvm/llvm-project/pull/205684