[GPU] Prefer TileAndFuse pipeline over SIMT pipeline

[nirvedhmeshram]

TileandFuse is the modernized pipeline that we would want to use over the older SIMT pipeline when possible.

[kuhar]

Can we observe the difference using the gemm shapes in https://github.com/nod-ai/iree-kernel-benchmark ?

[nirvedhmeshram]

Can we observe the difference using the gemm shapes in https://github.com/nod-ai/iree-kernel-benchmark ?

As I am running it I am realizing that all of but three shapes in the benchmark are multiples of 32 so dont go down these default path, but lets see what I get for those 3.

[kuhar]

@nirvedhmeshram We can add a few more. Do you have some specific shapes you are interested in?

[nirvedhmeshram]

@nirvedhmeshram We can add a few more. Do you have some specific shapes you are interested in?

@kuhar actually on close inspection I do see some shapes that go down SIMT for e.g

        %2 = linalg.matmul_transpose_a ins(%arg0, %arg1 : tensor<8192x14336xbf16>, tensor<8192x4xbf16>)
                                       outs(%1 : tensor<14336x4xf32>)

However, I am realizing this PR is most likely just an NFC, because TileandFuse just bails on such shapes as it wont find a schedule
https://github.com/iree-org/iree/blob/main/compiler/src/iree/compiler/Codegen/Dialect/GPU/TargetUtils/ConfigUtils.cpp#L210-L213

However I will change that in a follow up, next
#18791

[MaheshRavishankar]

Lets fold this into the subseqent PR? Hard to see what this flag is for.

[nirvedhmeshram]

Lets fold this into the subseqent PR? Hard to see what this flag is for.

Ya that is fine by me, Quinn suggested not to do too much at once, but since this is doing nothing, folding makes sense

[kuhar]

Ah, interesting, could you add this shape to iree-kernel-benchmark? We can tag it as 'corner_case'.

[nirvedhmeshram]

Ah, interesting, could you add this shape to iree-kernel-benchmark? We can tag it as 'corner_case'.

@jakub that shape is already generated by the benchmark so I was saying no need for adding more cases as the sizes pulled from the models seem to have enough diversity.