Diego Caballero

I'm still getting this error with the to-be-landed integrate (https://github.com/iree-org/iree/pull/10781): ``` /usr/local/google/home/diegocaballero/sandbox/models/deeplabv3/deeplabv3.mlir:229:12: error: 'tosa.resize' op requires attribute 'border' %226 = "tosa.resize"(%225) {mode = "BILINEAR", offset = [0, 0], offset_fp =...

I'm still seeing the `tosa.resize` in deeplabv3 after: ``` commit bba48dfe4ac4be72d74bb33e646d0b3984eeded1 Author: Rob Suderman Date: Wed Oct 5 16:28:19 2022 -0700 [mlir][tosa] tosa.resize canonicalizer for trivial noop If the scaling...

@pzread, I think this could be related to the vector gather work that you are doing. Would you mind taking a look?

@MaheshRavishankar, would it make sense to prevent fusion of non-vectorizable gathers for now? We have quite a few ops here that are not being vectorized so it might be an...

> AFAICS, this is an issue with RISC-V only. This is not a RISC-V specific issue. This is a limitation in the vectorizer that impacts all the backends using it....

> To start with you can try to make the change to see the difference (at least to evaluate the impact on all backends). No gaurantees though if we can...

Actually, most of the ops in the dispatch are part of the `tosa.resize` lowering so this is not a fusion problem. > One of the low-hanging fruit is adding the...

@rsuderman is looking into the canonicalization problem here: #10412

Hi @banach-space! Let me reply to you in #9198 since this issue is already fixed by just canonicalizing away `tosa.resize` ops that are actually no-ops.

After talking to @rsuderman, the main `tosa.resize` and hottest dispatch in DeepLabV3 can't be optimized away, as we initially thought. Therefore, the only option to optimize it is improving gather...