Yes. The reason is that there is no appropriate group op at tensor level, but there is a fused_brgemm op in XSMM. We’re also discussing group ops in Linalg to implement this (and other things).
It’s better to split refactoring from code changes in separate PRs.
This isn’t a single RFC, but involves everything from lowering from ingress, target description and cost modelling, pattern recognition, pattern matchers and the multiple topics discussed at the round table (link above).
All those discussions are happening upstream in Github or here.