The main rationale behind a single conv2d op is to be able to support grouped conv2d, where depthwise and conv2d are special case of the grouped conv2d.
Having three instances(conv2d, grouped conv2d and depthwise) or two won't remove the additional complexities in handling the conv2d case, since the groups parameters is expected to be supported. So I still think it is useful to keep the conv2d as it is. It is OK to have different compute to support the possibilities of different conv2d setups. Perhaps the main reason behind the pain was not clearly specifying the group conv2d semantics. It might be useful to discuss that case, and streamline every part to support group conv2d well. --- [Visit Topic](https://discuss.tvm.ai/t/separate-relay-depthwise-conv-operator/6114/2) to respond. You are receiving this because you enabled mailing list mode. To unsubscribe from these emails, [click here](https://discuss.tvm.ai/email/unsubscribe/aebc082cc33ca8ccd313602a6a9d58cc93ef48a5a0d35b1619a6fa12a7717c83).
