search

tenstorrent / tt-mlir

Tenstorrent MLIR compiler
https://tenstorrent.github.io/tt-mlir/
Apache License 2.0
78 stars 13 forks source link

Use ttnn::transpose instead of runtime decomposing transpose #1288

Closed dgolubovicTT closed 1 week ago

dgolubovicTT commented 1 week ago

Previously, transpose was decomposed in runtime to permute. This PR transfers directly to ttnn::transpose. I ran transpose tests on forge with this change and everything seems fine.

@vladimirjovanovicTT do you have some insight for why you added this decomposition? Was that a workaround due to some transpose case failing or transpose wan't implemented (less likely)?

vladimirjovanovicTT commented 1 week ago

Af first, ttnn::transpose wasn't avaiable, then after a metal uplift it became available, but simple use-cases didn't work. I'm happy to hear that it works well now!