Lunderberg opened a new pull request, #14982: URL: https://github.com/apache/tvm/pull/14982
This resolves an issue introduced by the combination of https://github.com/apache/tvm/pull/14918 and https://github.com/apache/tvm/pull/14945. The bug occurred for targets that do not require device-side codegen, but do require a `device_type` other than `kDLCPU`. It wasn't caught by CI, as the issue only occurred with the combination of both PRs. 1. #14918 updated `SplitHostDevice` to only modify the `"target"` attribute when a device-side function has been extracted. 2. For VTA, there is no device-side function, as everything is done through host-side API calls. 3. From (1) and (2), the VTA examples kept the target `T.target("ext_dev", host="llvm")` after the `SplitHostDevice` pass, instead of being updated to `T.target("llvm")`. 4. #14945 restricted CombineContextCall to only apply to host-side passes. 5. From (4) and (5), the `CombineContextCall` pass was no longer applied to the VTA context calls. This PR fixes `SplitHostDevice`, updating the target from `T.target("ext_dev", host="llvm")` to `T.target("llvm")`, even if no device sections have been extracted from the function. -- 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]
