mbaret commented on issue #5277: [BYOC] Refine AnnotateTarget and MergeCompilerRegion Passes URL: https://github.com/apache/incubator-tvm/pull/5277#issuecomment-611084398 To explain the point about QNN a bit better, a concrete example would be our Ethos-N NPU codegen paired up with something like DNNL. The NPU directly supports quantised convolutions and so wants to mark qnn.conv2d as supported. DNNL doesn't supported quantised convolutions but does support normal convolutions, so it wants to see nn.conv2d. Now the TFLite frontend will initially produce a graph containing qnn.conv2d's from a qnn model. If we choose to annotate here, only the Ethos-N codegen will be able to target the graph. If, however, we choose to run QnnCanonicalize, the qnn.conv2d's are lowered to nn.conv2d's but now only DNNL can annotate the graph. To me the ideal solution here is to be able to run AnnotateTarget multiple times on different lowerings of the graph, because I agree that running the partitioning pipeline multiple times is not a good long term solution (although it probably is a good short term one).
---------------------------------------------------------------- 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: [email protected] With regards, Apache Git Services
