[GitHub] [tvm] NicolaLancellotti commented on pull request #10959: [microNPU] Add a pass to reorder copy and compute nodes

[GitBox <gi...@apache.org>]

NicolaLancellotti commented on PR #10959:
URL: https://github.com/apache/tvm/pull/10959#issuecomment-1096855691

   > I'd be interested in seeing how that pass interacts with graphs that have mixture of operators with and without weights, e.g. it seems to me that when we have a graph that looks like
   > `pooling -> copy -> copy -> conv2d -> copy -> copy -> depthwise2d`
   > it will end up after this pass as
   > `copy -> copy -> pooling -> copy -> copy -> conv2d ... `
   > I suppose that is intentional, that we start copying the conv2d weights in while the MAC engine is crunching the pooling? Maybe it's worth adding a test that exercises that kind of mixture of ops?
   
   Yes, it is intentional, and the reordering is just what you said. I have added a test too.


-- 
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-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org