masahi opened a new pull request, #13631: URL: https://github.com/apache/tvm/pull/13631
Since https://github.com/apache/tvm/pull/12515, const folding on fp16, when both lhs and rhs are constant, is disabled. This is the case even when one of the arguments are one or zero. This leads to a real problem described in https://github.com/apache/tvm/pull/13532#issuecomment-1353715208. So in practice, compiling winograd fp16 on GPU is currently broken. This PR enables trivial algebraic simplification for fp16 const folding when (1) both arguments are constant and (2) one or both of the inputs are one or zero. Since we already does such simplification when only one of the inputs are constant (for example https://github.com/apache/tvm/blob/3a639a41a830ece914df10a98f807123e46cd750/src/arith/const_fold.h#L209-L216), I just removed all early exit `return NullOpt;` for fp16 cases. -- 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]
