Lunderberg commented on issue #14100: URL: https://github.com/apache/tvm/issues/14100#issuecomment-1445372847
Good catch, and thank you on the detailed analysis. The use of `extent *= new_arg->extent` does assume that the iterators are independent, and would produce incorrect output when the iterated aren't independent. I'm somewhat surprised that this expression didn't raise an error earlier when attempting to detect the transformation as an itermap. `Schedule::Blockize` internally calls `DetectIterMap` using `IterMapLevel::kSurjective` prior to calling `SubspaceDivider`, and I would have expected the non-injective transform `h_w_fused // 54 + kh` to raise an error at that point. I could see either (A) treating the incorrect extent for non-injective mappings as the bug, or (B) treating the failure to raise an error for the non-injective mapping as the bug. To help decide which path to take without restricting use cases, how did the `v0 = T.axis.spatial(56, h_w_fused // 54 + kh)` expression come about? The `h_w_fused // 54 + kh` doesn't look like a valid expression for `T.axis.spatial`, which expects the expression to be usable in data-parallel contexts. This looks more like two separate axes, one spatial and one reduction (i.e. `v1 = T.axis.spatial(54, h_w_fused // 54)` and `v2 = T.axis.reduction(3, kh)`). -- 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]
