Nirvedh Meshram

Based on some initial investigation the issue comes down to the trucation instructions we make see [here](https://gist.github.com/nirvedhmeshram/85007881488508ea3752ae008f8f9675) `taf` does tuncation using ```   v_pk_mul_f32 v[0:1], v[20:21], v[0:1]   s_nop 0   v_cvt_f16_f32_e32 v12,...

I can confirm that the issue is with the zeroing semantics of `v_fma_mixlo_f16` which is basically acknowledged here https://github.com/llvm/llvm-project/blob/ac0f64f06d67a93817ccd9a3c529ad40920115c9/llvm/lib/Target/AMDGPU/SIInstructions.td#L2835-L2843 Since this is not stable we can disable the use of...

> Is this closed now? I think it should be left open. Using that feature flag worked on gfx9 but caused ISEL assert on gfx11 but we didn't have the...

@MaheshRavishankar and @IanWood1 do you guys have thoughts on how to solve this? I was thinking we could have an option to not sink collapseshape through pad and use it...

I thought #20108 had API dependency concerns by Ben so we decided to not go though with it? Edit : oh I see Ben has approved it.

Sounds good! Is it still useful to also have the option to not sink through the pad as we might not always have it as a edge op e.g in...

Confirmed that after the PR from Ian this issue isnt happening

> Do you only want to generalize it when it is a transposition? How about other cases? We have a [DecomposePackUnPackOpsPass](https://github.com/iree-org/iree/blob/main/compiler/src/iree/compiler/Codegen/Common/DecomposePackUnPackOps.cpp), which converts pack ops to pad + reshape +...

> okay, I just figured out what we want to do and a better way to approach it. I was confused that why there are no reshapes generated in the...

> Instead of a pass, cant this be made a pattern in > > https://github.com/iree-org/iree/blob/04144f65f5fee10cfffca851e52d64ac08aadb43/compiler/src/iree/compiler/Codegen/Transforms/Transforms.cpp#L777 > > ? Ya I could see that working well, with transpose + expandshape that...