masahi edited a comment on pull request #9334:
URL: https://github.com/apache/tvm/pull/9334#issuecomment-948921743
> For removing FInferCorrectLayout, it looks like most of attrs are set to
ElemwiseArbitraryLayout, but some of the QNN ops are set to custom layout
methods
Oh yes, `InferLayout` functions for QNN ops are used when a user calls
`ConvertLayout` on a quantized model after the model is imported. So this is
very important. `QnnConcatenateLayout` is indeed used and there are other QNN
ops that have a specialized `FInferCorrectLayout`, such as
`RequantizeInferCorrectLayout`.
On the other hand, the ops you removed `FTVMCompute` from in this PR are
not "user-facing" in the same sense that QNN ops are, they are only used by VM
during compilation and `ConvertLayout` pass will never see them. So it should
be safe to remove `FInferCorrectLayout` from them.
I didn't realize that QNN ops are marked `TNonComputational`, but they are
kind of distinct from other `TNonComputational` ops in this sense.
--
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: commits-unsubscr...@tvm.apache.org
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
