[Mlir-commits] [mlir] Add arith expansion of f8E8M0 type for extf/trunc ops (PR #140332)

[Benoit Jacob]

bjacob wrote:

Just want to make clear that I was only commenting from a perspective of trying to treat E8M0 as a real FP type.

Maybe that was pointless since that type is inherently only ever useful as a MX scale type.  So it may make more sense to treat it as the specialized type that it is, which means, do whatever you feel is the established practice in this field.

I personally think that E8M0 is a mistake *even* within its intended application domain for MX scales, because in that domain, since scales are shared across many matrix elements, their bit-width doesn't matter that much, and accordingly, it would have been better so simply have, say, bf16 scales.  I don't say this to thrown another wrench in this discussion, on the contrary: this means that in my mental model, E8M0 is probably a short-lived oddity so we may as well not think too hard about it.

https://github.com/llvm/llvm-project/pull/140332