MasterJH5574 opened a new pull request, #15687: URL: https://github.com/apache/tvm/pull/15687
This PR enhances the fallback dlight GPU rule so that it can support more non-trivial spatial workloads. Particularly, to this end, this PR makes the following changes: * for function normalization: when a block of a TIR PrimFunc cannot be applied by "TransformLayout", it will now give up layout transformation instead of throwing the error in TransformLayout. * for the fallback rule, if the return value of the normalization is None, it will directly return, instead of proceeding the inline attempt. * for the fallback rule, when iterating the blocks in the PrimFunc, if the loops of a block are already bound to GPU threads, the block will be skipped so the loops will not be processed again. -- 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]
