[Mlir-commits] [mlir] [MLIR] Add f8E8M0FNU type (PR #111028)

[Stella Laurenzo]

stellaraccident wrote:

> > Fly on the wall, but at what point, upon removing all features that traditionally make something a "floating point number" (mantissa, zero, denorms, infinities) does something no longer make any sense at all being part of a floating point hierarchy. It's just a bit-vector with a special error value.
> > I'm not blocking this in any way or even asking seriously. Just kind of balking at the cargo cult mentality that is going into bundling these things together like this.
> 
> I had the same gut reaction... We are now up to ~18 floating point types... That kind of points to a serious issue with the way things are scaling here, and I think we should really rethink what's being done (especially given that each one of these PRs are identical, add big chunks of code in the core library, etc).

Ok, I didn't want to unilaterally hit pause, but it seems like we've got a number of people with the same analysis. Should we at least discuss this a bit more? Or proceed with this patch? I agree that the path we're on is not very sustainable.

https://github.com/llvm/llvm-project/pull/111028