[Mlir-commits] [mlir] [Linalg][Vectorization] Add support for linalg vectorization of a tensor.extract case (PR #107922)

[Andrzej Warzyński]

banach-space wrote:

> @banach-space #100582 breaks the logic here as one of the test cases is
> 
> ```
> %5 = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>], iterator_types = ["parallel", "parallel"]} outs(%arg1 : tensor<?x?xf32>) {
> ^bb0(%out: f32):
>   %6 = linalg.index 1 : index
>   %7 = arith.addi %6, %arg2 : index
>   %extracted = tensor.extract %arg0[%c79, %7] : tensor<?x?xf32>
>   linalg.yield %extracted : f32
> } -> tensor<?x?xf32>
> ```
> 
> where it will find two non unit dims and hit the assert

Apologies, I didn't expect that to be such a curve ball. But I am really glad that we are hitting these cases, it really helps to stress-test this logic. Hopefully my explanation makes sense. I've also deleted a couple of my comments - #100582 invalidated them 😅 

https://github.com/llvm/llvm-project/pull/107922