[Mlir-commits] [mlir] [mlir][linalg] Fix for bias handling for Winograd (PR #110331)
Georgios Pinitas
llvmlistbot at llvm.org
Mon Sep 30 04:07:53 PDT 2024
GeorgeARM wrote:
> > I am happy to be corrected but I am not sure adding bias to an op like this is warranted.
>
> When lowering tosa.conv2d to linalg, the output tensor will be filled with bias values. The implementation here assumes the values inside output tensor is meaningful. If we do not added it back inside output transform, we will have another round to add the bias values back to the transformed results after output transform op.
Sure but this itself is not per se coupled with bias. It is just using and operating over the initial values that were injected to the operator (unless I am missing something). So I would suggest dropping anything that has to do with bias here.
https://github.com/llvm/llvm-project/pull/110331
More information about the Mlir-commits
mailing list