Re: [Gambas-user] rounding a half

[Benoît Minisini <benoit.minisini@xxxxxxxxxxxxxxxx>]

Le 11/12/2024 à 16:05, Benoît Minisini a écrit :
> Le 09/12/2024 à 21:27, Jussi Lahtinen a écrit :
> > Hmmm... no this seems like a fundamental problem in floating points.
> > You could do something like this:
> >
> > RETURN->_float.value = floor((PARAM->_float.value / power + 0.5) * 
> > t) / t * power;
> >
> > Where t is double with a large value. This would hide the problem into 
> > smaller decimal places.
> > Not sure it's worth it. Probably implementing arbitrary precision 
> > datatype would be the right way. Benoit?
> >
> >
> > Jussi
>
> The reason is that Round() rounds in binary, not in decimal.
>
> 1.015 and 1.055 do not have an exact representation in binary.
>
> 1.015 = 1.000000(11110101110000101000)*
> 1.055 = 1.0000(11100001010001111010)*
>
> Moreover, Round() sometimes uses a multiplication (when its second 
> argument is positive) or a division (when its second argument is 
> negative) to round.
>
> So, when the decimal digit where we want to round is "5", sometimes the 
> number is rounding down, sometimes up.
>
> Regards,

I just introduce a little change in Round(): it now always use positive 
powers of ten for rounding. It leads to less errors, and return the 
expected result in the previous specific case.

But for sure we will find other numbers where Round() won't round the 
way we think.

Regards,

--
Benoît Minisini.