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

[Andrzej Warzyński]

banach-space wrote:

Thanks for elaborating, Adam! Now I see what was the source of confusion for me. Basically, there is no ambiguity at the `Linalg` level, but rather in how we materlialise broadcasts at the `Vector` level.

> I lean toward forcing users to first decompose unsupported affine maps to not have to guess or compromise on an arbitrary default decomposition for these edge cases.

+1

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