Thanks for the discussion. To provide more context, the A0 approach we discussed is TIR-Relax layout rewriting https://github.com/tlc-pack/relax/issues/162 (the general idea is to lift such transformation in TIR scheduling into the graph, and then cancels out redundant intermediate transformations by either proving fusing the pair of post-compute and pre-compute transformations produces an identity TIR function, or use high-level operator semantic). I think this is very similar to the [graph-level solution](https://discuss.tvm.apache.org/t/introducing-ty-nnp-backend-with-end2end-tensorir-integration/11807/4) mentioned by @wrongtest In general, both A0 and A1 are valid approaches. It is mainly about how we would like to handle the complexity in simplifications.
-- Reply to this email directly or view it on GitHub: https://github.com/apache/tvm-rfcs/pull/77#issuecomment-1152992143 You are receiving this because you are subscribed to this thread. Message ID: <apache/tvm-rfcs/pull/77/c1152992...@github.com>
