search

iree-org / iree

A retargetable MLIR-based machine learning compiler and runtime toolkit.
http://iree.dev/
Apache License 2.0
2.85k stars 621 forks source link

[Dispatch] Don't bubble extract slice ops through reduction ops #19259

Closed IanWood1 closed 3 days ago

IanWood1 commented 3 days ago

Fixes the compilation issue in https://github.com/iree-org/iree/issues/19230 but a more nuanced solution is needed since it can sometimes be beneficial to reduce the computation done by linalg.generics with reduction iterators.

saienduri commented 3 days ago

Actually, we don't need this. We are good on TOM vae decode. The compilation issue is with vae encode which we don't need but is part of the IR. I changed the export, so it doesn't worry about vae encode. This is also why it wasn't picked up by IREE CI. We can wait for Nirvedh's long term changes here as I think we were saying this is not the most ideal route. But, thanks for getting in a quick fix @IanWood1 :)

IanWood1 commented 3 days ago

Actually, we don't need this. We are good on TOM vae decode. The compilation issue is with vae encode which we don't need but is part of the IR. I changed the export, so it doesn't worry about vae encode. This is also why it wasn't picked up by IREE CI. We can wait for Nirvedh's long term changes here as I think we were saying this is not the most ideal route. But, thanks for getting in a quick fix @IanWood1 :)

Ahh that makes sense. Great, I'll close this PR then.