tqchen commented on issue #9022: URL: https://github.com/apache/tvm/issues/9022#issuecomment-921945722
Thanks @manupa-arm. I agree that putting TVMBAW as a peer to heap is not right(that was meant as an example to demonstrate the viewpoint. I do not necessary want to enforce heap as a peer to TVMBAW). For the case of AOT, however. I think it would still be useful to introduce a specialized scope eventually. So we can place the constraint that the memory is accessible by npu explicitly in the IR. While there is no micro specific target, we can introduce micro specific tags that set these properties. See https://github.com/apache/tvm/blob/main/src/target/tag.cc#L84 as an example of cuda tags I think all three attributes you proposed could work. A1 might be the simplest for now and it won't preclude A2/A3 options in the future. -- 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]
