[Mlir-commits] [mlir] [mlir][linalg] Vectorize directly to a named contraction (PR #147296)

[Han-Chung Wang]

https://github.com/hanhanW commented:

The approach looks okay to me because I'm -1 on recovering information to achieve the same thing when we have enough information. It'd be good to see what is missing to make it default, as I think it is usually bad to have boolean options, especially when they are linalg specific. It makes generalization harder if we want vectorization interface in the future. I'd like to learn why it can't be default and what is missing.

I also have a question about "materializing broadcast" in my inline comment.

https://github.com/llvm/llvm-project/pull/147296