Junius-Wynn commented on PR #20057: URL: https://github.com/apache/tvm/pull/20057#issuecomment-5439484393
@tlopex Thanks, you're right. The guard was still using the original iterator domain, so it could disagree with `_normalize` when `lower_factor > 1`. I updated the spatial-domain analysis to use `split_expr.extent` for split spatial iterators, and the same normalized effective extent is now used to select the thread extent. For the `(2, 3, 8)` case with `A[n, y, k, x // 2]`, both paths now use extent 4 and select `tx=4`. I also added this case as a regression test, including a check for the generated `threadIdx.x` extent. The related DLight tests pass with `26 passed`, and pre-commit passes as well. -- 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]
