Junius-Wynn commented on PR #20057: URL: https://github.com/apache/tvm/pull/20057#issuecomment-5427577733
@tlopex Sorry, my previous fix was not correct. I focused too much on mixed spatial/reduction accesses, but that was not the actual cause. You were also right that catching `ScheduleError` globally in `_apply_rules` is not a good idea, since it could hide real scheduling bugs. I removed that change. I updated `Reduction` to check the actual fused spatial geometry instead. The check now follows the normalized dominant-read order, so the reordered-access case is handled correctly. Unsupported non-affine write-back shapes return `None` from `Reduction`, while unexpected `ScheduleError`s are still raised normally. I added tests for the original conv2d case, plain `relax.op.sum`, reordered access, affine shapes, and unexpected errors. The tests pass with `25 passed`, and the original reproduction now builds successfully. Sorry again for the earlier incorrect fix, and thanks for pointing it out. -- 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] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
