[Mlir-commits] [mlir] [MLIR] [Vector] Linearization patterns for vector.load and vector.store (PR #145115)

[James Newling]

newling wrote:

> @newling following up on [#143420 (comment)](https://github.com/llvm/llvm-project/pull/143420#issuecomment-2967406606)
> 
> For 2, Is this what you meant? created a draft PR because that would be quicker

Yeah, this is what I meant 👍  

Ideally we'll eventually have something like flattening of transfer_read, done [here](https://github.com/llvm/llvm-project/blob/be7e4113c84317feda055cac68a855429e46e381/mlir/lib/Dialect/Vector/Transforms/VectorTransferOpTransforms.cpp#L593). i.e. linearize even when there is more than 1 dimension of size > 1, and it needn't be the inner-most dim. But I guess that can wait. FWIW IMO that transfer_read code should be in VectorLinearize too, I mentioned that at the bottom of [this](https://discourse.llvm.org/t/rfc-mlir-vector-modify-remove-bit-width-specific-flattening-linearization/86876/3?u=newling) comment. And the vector.load linearization code could probably then reuse some of it. Something for the future, maybe!

https://github.com/llvm/llvm-project/pull/145115