DickJC123 edited a comment on issue #13598: More fine-grained operator implementation dispatch & memory planning flow URL: https://github.com/apache/incubator-mxnet/issues/13598#issuecomment-447533189 I looked this over and concluded it's a complicated issue, so I'm not ready to take a strong stand. Some thoughts though: One desire of mine would be to have an op implementation be a slight tweek of the existing "base op" without copying the entire op description. Also, whatever is decided on here should play well with the existing op registration mechanism and the subgraph api. I haven't studied subgraph frankly, but I'd hope one could make implementation decisions (and graph alterations) based on all the information normally provided the forward/backward op calls (so context, shapes, dtypes, stypes, etc.). So during subgraph graph passes, would there be a lightweight way to swap in a tweeked operator? So, for discussion, what about: ``` NNVM_REGISTER_OP(Convolution_CUDNN_Impl) .clone_of(Convolution) // clones with a new name the Convolution op .override_attr<FInplaceOption>("FInplaceOption", [] FInplaceOption(const NodeAttrs& attrs) { return {0,0}; }) ``` Then use the subgraph API to swap in a Convolution_CUDNN_Impl node for a Convolution node if the parameters, gpu arch, etc. supported it? Alternatively, in a subgraph API graph pass, could one keep the node, but attach an attribute that preempts the default FInplaceOption (or overwrites the FInplaceOption function pointer used by the node directly)?
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
