masahi commented on PR #14097: URL: https://github.com/apache/tvm/pull/14097#issuecomment-1446090660
Thanks for the discussion, I think we shouldn't push all responsibilities to codegen and call it "a more general solution". At least for the cases where there is a one-to-one mapping between a composite function and an op, clearly there is one "right" function signature. So each backend shouldn't need to worry about `add` or other trivial functions possibly having multiple signatures. For more complicated, "truly fused" composite functions that are meant to be consumed by library-based BYOC (dnnl, cutlass), I think it is reasonable to ask codegen to handle the signature problem. I liked @yelite suggestion of leaving the current op fusion impl as is, and instead adding a standalone pass that would identify such trivial composite functions (the one that merely wraps an op, e.g. `tensorrt.add`) and fix up their signature if necessary. Later we may extend that pass to support the `fused_permute_matmul` case - convert the signature into more "natural" one based on the anchor op call site. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
