mbrookhart commented on pull request #9767: URL: https://github.com/apache/tvm/pull/9767#issuecomment-1032931537
Ah, I think I understand now. I see a couple of options. The simplest might be to run this pass on the graph before doing pattern matching, that should remove most of the pads, but it would also require removing the pads from your patterns. https://github.com/apache/tvm/blob/824772489e514faf06025d7c09ce4dc13dcd7d08/python/tvm/relay/transform/transform.py#L1214-L1224 The next best option might be to align the rewrite callback and check APIs. For Rewrite, we have the callback take the pre-rewrite grap, the post-rewrite graph, and the map of pattern->node matches: https://github.com/apache/tvm/blob/824772489e514faf06025d7c09ce4dc13dcd7d08/src/relay/ir/dataflow_matcher.cc#L821 If we adjusted the `check` API to take the same arguments, that might solve the problem. Again, it would require adjusting definitions in other parts of the codebase. -- 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]
