[BUG] Invalid stochastic rounding with infinite

[Thukisdo]

Hi,

We've spotted an issue in Float32_stochastic_round() where rounding an Infinite will sometime return NaN, where we may expect Infinite, see output below.

julia> Float32(Float32_stochastic_round(Inf))
Inf32

julia> Float32(Float32_stochastic_round(Inf))
NaN32

The only difference between NaN and infinite is an empty mantissa, hence adding a random noise will result in NaN.

[milankl]

Sorry @Thukisdo, missed that issue. Yes, I observed that earlier, but couldn't be bothered resolve this, as I didn't think this would be of practical importance. But maybe that's also me thinking that NaR (= one bitpattern for NaN, +-Inf) from posits is a good idea. I guess this could be resolved by another early exit, something like isfinite(x) || return BFloat16sr(x) such that for x=NaN32 it returns NaNB16sr and for x=Inf32 it returns Inf16sr ?

[Thukisdo]

No problem :)

Indeed, i also solved it using an early exit in my code.

[milankl]

Happy for you to create a PR with your solution?

[Thukisdo]

I applied this fix in a C++ adaptation of your rounding method here, so i unfortunately do not have a patched Julia version.

[milankl]

Thanks @Thukisdo, this is addressed in upcoming v0.6.1 as I've replaced the stochastic rounding via signed integer arithmetic with unsigned integers. This avoids any stochastic rounding between Inf and NaN and doesn't require a check for special cases like isfinite(x) || return ... or isnan(x) || return .... Check the new Float32_stochastic_rounding function definition if you are interested!