[Mlir-commits] [mlir] [mlir][ArmNeon] Add linalg.matmul e2e tests (PR #212809)

[Federico Bruzzone]

================

----------------
FedericoBruzzone wrote:

Switched to `transform.structured.transpose_matmul %matmul <rhs>` on a vanilla `linalg.matmul`, as suggested. A much cleaner approach, thanks! :D

---

One snag: with the RHS cast to dynamic `tensor<?x?xi8>`, `structured.vectorize {create_named_contraction}` on the resulting `linalg.matmul_transpose_b` produced a `vector.mask` wrapping several ops instead of the single op it requires, currently MLIR rejected it with: _`vector.mask` op expects only one operation to mask_. Switching to static shapes avoided the mask being generated at all. 
I don't think this is intentional. However, I'll look into it further and open an issue if necessary. What do you think? @banach-space 

https://github.com/llvm/llvm-project/pull/212809