[Mlir-commits] [mlir] [mlir][Vector] Support mixed mode vector.contract lowering (PR #117753)

[Kunwar Grover]

Groverkss wrote:




> Hey, thanks a lot for the contribution. As you may know, this has been a controversial topic (actually, I'm not sure if we discussed it in the context of `vector.contract` or linalg at the time). The main problem is this:
> 
> > For integer types, only signless integer types are supported, and the promotion happens via sign extension.
> 
> If we want to walk this path, we should find a consistent way to represent signed and unsigned extensions per operand (and any potential flags associated with each operand extension, including FP ones). Otherwise, we would need different matches/handling depending on whether the specific extension each operand is going through.
> 
> Any ideas to address this issue?

Just to be clear, this patch is implementing mixed mode vector.contract lowering for lowering variants other than outer product. Outer product lowering already implemented mixed mode lowering, this PR simply adds it for other variants as well, in the same way outer product lowering did it. If something is controversial here, I'm only following what is already implemented.

On the point of signed/unsigned integer extensions per operand, the line that you quoted regarding using sign extension is not something I decided, but is from vector.contract documentation (https://mlir.llvm.org/docs/Dialects/Vector/#vectorcontract-vectorcontractionop):

> If operands and the result have types of different bitwidths, operands are promoted to have the same bitwidth as the result before performing the contraction. For integer types, only signless integer types are supported, and the promotion happens via sign extension.

The documentation already mentions that we use signed extension, which is what this pr implements.

Regarding FP types, if an operand extension requires fast math flags, I would guess that this should be annotated on vector.contract. But I'm not aware of any other lowerings looking at fast math flags or any fp flags. Do you have an example or a documentation link I could follow for what is the intended behavior there?

https://github.com/llvm/llvm-project/pull/117753