tqchen edited a comment on pull request #8472: URL: https://github.com/apache/tvm/pull/8472#issuecomment-880203373
Thanks @manupa-arm . Changing the IR definition can create a bigger impact than attaching optional attrs etc. Since the new node introduce semantics that existing passes may not taken into account. So it would be indeed great to put some thoughts into this. We need to put extra care into this to maintain architecture consistency across all the sub areas of interest. My current read is that the function attrs approach can also be used to run the same optimization mentioned(pooling the constant). Except that rewrite needs to be done which takes the function attrs and intrinsics. I would still recommend us take that route first. Given the common need of embedding artifacts, we might want to find a way to express global constant artifact in the IRModule, perhaps as `Map<GlobalVar, DataSegment>`, and referring them in the IR. This way could be cleaner than inlining the constant directly. -- 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: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
