[Mlir-commits] [mlir] [MLIR][XeVM] Add truncf and mma_mx op. (PR #180055)

[Sang Ik Lee]

silee2 wrote:

> Ok, so, just because I forget which dialect you're operating on - is this the MLIR wrappers dialect (analogous to AMDGPU) or the LLVM intrinsic wrappers dialect (analogous to ROCDL)?
> 
> If it's the MLIR-side dialect, what LLVM supports doesn't matter - if it's a first-class type in MLIR, you shouldn't redundantly be encoding the type information.
> 
> If it's the LLVM-side dialect, then having type attributes and complex translation logic is an anti-pattern - the mapping from 8-bit float types to ... an integer or picking a different intrinsic or however it works for you ... should be handled by the translation layer from the MLIR wrappers.
> 
> (I'm willing to be overruled by dialect maintainers here, but I don't want this sort of layering issue to become common)

XeVM is analogous to ROCDL. 
Agree that types should be kept as true type and not the wrapper type if possible.
8bit float types can handled during translate MLIR to LLVMIR stage but sub-byte type, fp4 (e2m1) can be tricky.
At the same time, I believe translation layer should stay simple.
Thanks for the feedback. Will consider all factors and make a decision.


https://github.com/llvm/llvm-project/pull/180055