Floating-point rounding

In floating-point arithmetic, rounding aims to about-face a accustomed amount x into a amount z with a defined amount of cogent digits. In added words, z should be a assorted of a amount m that depends on the consequence of z. The amount m is a ability of the abject (usually 2 or 10) of the floating-point representation.

Apart from this detail, all the variants of rounding discussed aloft administer to the rounding of floating-point numbers as well. The algorithm for such rounding is presented in the Scaled rounding area above, but with a connected ascent agency s=1, and an accumulation abject b>1.

For after-effects area the angled aftereffect would overflow the aftereffect for a directed rounding is either the adapted active infinity, or the accomplished representable absolute bound amount (or the everyman representable abrogating bound amount if x is negative), depending on the administration of rounding. The aftereffect of an overflow for the accepted case of annular to even is consistently the adapted infinity.

In addition, if the angled aftereffect would underflow, i.e. if the backer would beat the everyman representable accumulation value, the able aftereffect may be either aught (possibly active if the representation can advance a acumen of signs for zeroes), or the aboriginal representable absolute bound amount (or the accomplished representable abrogating bound amount if x is negative), possibly a denormal absolute or abrogating amount (if the mantissa is autumn all its cogent digits, in which case the a lot of cogent chiffre may still be stored in a lower position by ambience the accomplished stored digits to zero, and this stored mantissa does not bead the a lot of cogent digit, something that is accessible if abject b=2 because the a lot of cogent chiffre is consistently 1 in that base), depending on the administration of rounding. The aftereffect of an underflow for the accepted case of annular to even is consistently the adapted zero.