Rounding Algorithms

dtmos writes "Clive Maxfield has an interesting article up on PL DesignLine cataloging most (all?) of the known rounding algorithms used in computer math. As he states, "...the mind soon boggles at the variety and intricacies of the rounding algorithms that may be used for different applications ... round-up, round-down, round-toward-nearest, arithmetic rounding, round-half-up, round-half-down, round-half-even, round-half-odd, round-toward-zero, round-away-from-zero, round-ceiling, round-floor, truncation (chopping), round-alternate, and round-random (stochastic rounding), to name but a few." It's a good read, especially if you *think* you know what your programs are doing."

Comments

I read the first half of the article...

[under_score]

...where it discusses the various rounding methods. I had actually thought of/used most of them. The one that was new to me was the round-half-even (banker's rounding). Very cool idea, and I had no idea it was commonly used.

This is a great reference article! If you are programmer working with numerical algorithms, keep this article handy.

Slide Rules and precision

[goombah99]

These days kids are not taught to round. Instead you just do the compuations at absurdly large precision then on the last step round off. This way you don't accumulate systematic round-off error. It's good as long as you have the luxury of doing that. It used to be that C-programmers had a cavalier attitude of always writing the double-precision libraries first. Which is why Scientific programmers were initially slow to migrate from fortran.

These days it's not so true any more. First there's lots of good scientific C programmers now so the problem of parcimonius computation is well appreciated. Moreover the creation of math co-processing, vector calcualtions, and math co-processors often makes it counter-intuitive what to do.

  For example it's quite likely that brute forcing a stiff calculation is double precision using a numeric co-processor will beat doing it in single precision with a few extra steps added to keep the precision in range. So being clever is not always helpful. people used to create math libraries that even cheated on using the full precision of the avialable floating point word size (sub-single precision accuracy) since it was fast (e.g. the radius libs for macintoshes) Pipelining adds more confusion, since the processor can be doing other stuff during those wait states for the higher precision. Vector code reverse this: if you are clever maybe shaving precision willlet you double the number of simultanoeus calcualtions.

In any case, what was once intuitive: minimal precision and clever rounding to avoid systematic errors means faster computation is no longer true.

Of course in the old days people learned to round early in life: no one wanted to use a 5 digit precision slide rule if you could use a 2 digit precision slide rule.