[Mlir-commits] [mlir] [mlir][vector] Fold transpose(broadcast(shape_cast)) to broadcast (PR #215940)

[Jianhui Li]

Jianhui-Li wrote:

@banach-space  @dcaballe  I've further narrowed this PR's scope: the transpose may only reorder the dimensions the broadcast prepends; the dimensions inherited from the broadcast's input must keep their relative order. Stretching a size-1 dimension and then transposing it is therefore no longer folded.

I believe this already meet the transpose rule, although I didn't get clarification in my last post.  I would like to proceed. Please clearly indicate whether this scope of fusion is accepted or not. If not, I will have to move them to XeGPU. 

https://github.com/llvm/llvm-project/pull/215940