iree-org / iree

A retargetable MLIR-based machine learning compiler and runtime toolkit.
http://iree.dev/
Apache License 2.0
2.47k stars 551 forks source link

Merge `round_dims_to` and `matmul_narrow_{M,N}` into a single `max_padding` array attribute #17729

Open bjacob opened 6 days ago

bjacob commented 6 days ago

This is a take 2 on https://github.com/iree-org/iree/issues/17553. There, I had suggested renaming round_dims_to to max_padding and making it a scalar, while retaining matmul_narrow_{M,N} overriding it for narrow dimensions and clarifying the whole semantics.

Last week we discussed in-person that it would be even nicer to resolve this the other way around: keep the renaming of round_dims_to to max_padding but keep that an array attribute and instead drop the separate matmul_narrow_{M,N}, folding their semantics into that.