masahi commented on pull request #9048: URL: https://github.com/apache/tvm/pull/9048#issuecomment-927116431
@comaniac There is one thing that's very odd about what `ReduceInferCorrectLayout` does. When the new layout is "NCHWc", rather than computing reduction over "NCHWc" layout directly, `ReduceInferCorrectLayout` inserts a conversion "NCHWc -> NCHW" before reduction, and reduction is always done in NCHW (see Animesh's comment in https://github.com/apache/tvm/pull/4059/files#r332704381). So when compiling efficientnet v2 for x86, I still get many "NCHWc -> NCHW" conversions. There is nothing that prevents reducing over `NCHWc` layout (transforming axes = [1] to axes = [1, 4], for example). What do you think? If this sounds good, I can follow up with another PR. -- 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]
