FrozenGene commented on issue #4803: [WIP][Frontend] Asymmetric padding of convolution support URL: https://github.com/apache/incubator-tvm/pull/4803#issuecomment-581488095 Initial Investigation: For QNN, before asymmetric padding support, we will handle it in tflite frontend inserting `nn.pad` so that the `attr['padding']` is always 0 even it is `same` padding. However, the things become different when we have asymmetric padding support. If we pass 4D padding (top, left, bottom, right) directly to QNN, if we are in the target of CPU, we will call ```python def helper_no_fast_int8_hw_legalization ``` if we don't have int8 acceleration. there isn't pad handling here. The logic of QNN is always do like `Conv2DPadInput`. This is not perfect way because we will always have `pad` operator before QNN. We should have `pad_const` and let the tvm backend to handle pad value. I prefer this PR make FP32 be asymmetric padding and leave QNN padding support as previous way (i.e. inserting `nn.pad` in tflite frontend). However, I wish we could add one attribute `pad_const` to `Conv2DAttr` and related api to support asymmetric padding computation of QNN. How about you? @anijain2305
---------------------------------------------------------------- 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. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
