[RFC] Should the OCP microscaling float scalars be added to APFloat and FloatType?

[ThomasRaoux]

sorry for joining this thread late. Few comments/questions.

I do like the proposal to add getScalarBitwidth() and (we may have discussed it before) having an interface for it would allow this to work with any type that may live downstream which would be a big improvement. I’m happy to help if we have a consensus about this.

About allowing fp4/fp6 formats in tensors, I still think this would be useful for high level representations. We already support sub-bytes type in tensor for integers so it doesn’t seem like a new problem.

@stellaraccident

As for the basic scalar types themselves, I think that is an option we should keep an open mind about. The hint that we should define them in APFloat is if we end up starting to use them as scalars or basic element types in constants. I could see such a situation arising naturally as part of the rest of the design, and the advantage to defining them properly is that you get basic parse/print/typing and arithmetic emulation suitable for folding and such. If we find ourselves there, it makes a lot of sense to add them as real types, and I’ve found that a much better option than trying to define such things another way.

+1, and this seems orthogonal to the block representation.
Note that there is a recent PR that goes in that direction (I’m not the author nor I have discussed much with them but I think this would be useful):

[APFloat] Add APFloat support for FP6 data types