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

[Benoit Jacob]

bjacob wrote:

Right, so the specification is under-defined when it comes to E8M0 conversions. I'm just suggesting what I believe is the more important invariant to preserve: "if a conversion produces a finite value then it should be a nearest value". Also, what Krzystof said about type conversions being monotonic functions is another important invariant to preserve. By contrast, dropping the sign bit seems arbitrary, except maybe if one thinks less in terms of the real numbers being encoded and more in terms of how they are encoded with or without a sign bit.

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