mathematical rounding...

hi, What's the fastest algorythm to do rounding of decimal numbers... since rebol does not seem to have an explicit rounding function... to-integer only truncates... thanks! Max

For rational arithmetic on a computer look at www2.hursley.ibm.com and look at big decimal. bobh

Here are some ones I tried that seem to be real fast: ; This is fastest, but is limited to positive integers. round: func [n] [ to-integer n + .5 ] ; This is the fastest for working with large decimals, but only does positive numbers... round: func [n] [ (n + .5) - ((n + .5) // 1) ] ; This does positive and negatives fine... round: func [n] [ (n: either n < 0 [ n - .5 ] [ n + .5 ]) - (n // 1) ] ; If you need to certain places do this: to-thousandth: func [n] [ (n: either n < 0 [ n - .0005 ] [ n + .0005 ]) - (n // .001) ] You can test your own with the program I have been using to improve my ray casters performance: do http://www.sonic.net/~gaia/misc/benchmark.r --Ryan Media wrote:

> hi, > What's the fastest algorythm to do rounding of decimal numbers... since > [rebol-request--rebol--com] with "unsubscribe" in the > subject, without the quotes.

Hi Max, Here's is Ladislav's rounding.r module, though there may be others out there as well. Rebol [ Title: "Rounding" Purpose: {Rounding functions} Author: "Ladislav Mecir" Date: 2/8/2000 Email: [lmecir--geocities--com] File: %rounding.r Category: [Math] ] mod: func [ {compute a non-negative remainder} a [number!] b [number!] /local r ] [ either negative? r: a // b [ r + abs b ] [r] ] round: func ["Round a number" n [number!] /places p [integer!] {Decimal places - can be negative} /local factor r ] [ factor: either places [10 ** (- p)] [1] n: 0.5 * factor + n n - mod n factor ] floor: func [ n [number!] /places p [integer!] {Decimal places - can be negative} /local factor r ] [ factor: either places [10 ** (- p)] [1] n - mod n factor ] ceiling: func [ n [number!] /places p [integer!] {Decimal places - can be negative} /local factor r ] [ factor: either places [10 ** (- p)] [1] n + mod (- n) factor ] truncate: func [ n [number!] /places p [integer!] {Decimal places - can be negative} /local factor r ] [ factor: either places [10 ** (- p)] [1] n - (n // factor) ] --Gregg

Performance was pretty good for what you get, but not as fast as the functions I posted. One thing though...

>> round -5.5

> Hi Max, > Here's is Ladislav's rounding.r module, though there may be others out there > [rebol-request--rebol--com] with "unsubscribe" in the > subject, without the quotes.

If I recall, this is ok since: in negative, when you are at -5.5 you are actually closer to -5 (a larger value) than -6 (a smaller value) THANKS EVERYONE! -Maxim

HI, Following the rounding question I have done some tests... and Tought I'd send this little note to everyone... Although this seems to be normal IEEE sanctioned behaviour, I have noticed that producing a floating point calculation with more than 15 characters of precisions as in: 1.499999999999999 will affect the actuall value of the flating-point number. evaluating the above actually returns 1.5 ... try it, you'll see ! so the above rounded return 2 not 1 !!! adding 0.5 to the above will cause an overflow to occur and 1.999999999999999... actually becomes 2.0 ! thanks for that excellent reference go to Robert Hamilton... This behaviour is mentioned and somewhat explained... in some of the pages when you dig a little bit... it seems to have to do with binary translation of decimal values (or something like that, excuse my gross interpretation)... Just thought I'd point it out since it isn't such an obvious side-effect to debug when coding... And Like many if you I'm not a scientific maths expert... making it THAT much less obvious... Ciao! -MAx

Hi Ryan, << Performance was pretty good for what you get, but not as fast as the functions I posted. One thing though...

>> round -5.5

in the case of .5 there is a choice and different results have been in vouge at different times "round to the even number" and "random" are a couple of strategies I recall but I think these days "round towards zero" is expected (the point of these equally (in)valid strateges is to minimize bias) On Tue, 6 Nov 2001, Gregg Irwin wrote:

>> round -5.5

Hi Ryan, The standard rounding rule used in most numeric analysis packages (and also implemented as the default in IEEE Std 754 for binary floating point) is round nearest even . This means round to the nearest integer; in case of a tie make the least significant digit of the result an even number. Rounding even avoids bias in the result if the digits to be rounded off occur with equal frequency. Of course, rounding odd would work just as well, but the chosen convention is to round even. In base 10, rounding goes like this: 5.1 thru 5.4 rounds to 5 5.5 rounds to 6 5.6 thru 5.9 rounds to 6 -2.5 rounds to -2 -1.5 rounds to -2 -0.5 rounds to 0 0.5 rounds to 0 1.5 rounds to 2 2.5 rounds to 2 etc. Negatives are just the mirror image, so the correct answer is: -5.5 rounds to -6 Note that 5.5 is exact in both base 10 arithmetic and in base 2 arithmetic, so in this case there is no problem due to conversion between bases. However, in general, the base conversion will introduce further issues. For a set of functions which handle all the gory details for rounding doubles of any size in REBOL, see the scripts format.r and decimal.r on the Ecotope rebsite. www.nwlink.com/~ecotope1/reb/format.r www.nwlink.com/~ecotope1/reb/decimal.r The format function gives this result:

>> format -5.5 #0

Amazing.

> Hi Ryan, > << Performance was pretty good for what you get, but not as fast as the > This is another one of those times where a definitive set of documented > library functions would be helpful.

> --Gregg > > -- > To unsubscribe from this list, please send an email to > [rebol-request--rebol--com] with "unsubscribe" in the > subject, without the quotes. >

from: Gregg

> Hi Ryan, > << Performance was pretty good for what you get, but not as fast as the > This is another one of those times where a definitive set of documented > library functions would be helpful.

> --Gregg > > -- > To unsubscribe from this list, please send an email to > [rebol-request--rebol--com] with "unsubscribe" in the > subject, without the quotes. >

> There are several accepted methods of rounding. One version for rounding > from a 5 (i.e., 0.5, 0.05, etc.) says to round up if the number to the left > is even, and truncate if it is odd. The purpose (hope) is to randomize the > error so any statistics behave better.

Hi Alan, << There are several accepted methods of rounding. One version for rounding from 0.5 says to round up if the number to the left of the decimal is even, and truncate if it is odd. The purpose (hope) is to randomize the error so any statistics behave better. >> Right. That's what I meant by statistical rounding. --Gregg

--- Thomas E Conlin <[tomc--darkwing--uoregon--edu]> wrote:

> in the case of .5 there is a choice and different > results have been in > (the point of these equally (in)valid strateges is > to minimize bias)

--- Gregg Irwin <[greggirwin--mindspring--com]> wrote:

> Hi Alan, > << There are several accepted methods of rounding. > so any statistics behave better. >> > Right. That's what I meant by statistical rounding.

> --- Thomas E Conlin <[tomc--darkwing--uoregon--edu]> wrote: > > > a specific reason for using another method) round > *away* from zero at 0.5 as the standard.

> Chris Dicely > __________________________________________________ > [rebol-request--rebol--com] with "unsubscribe" in the > subject, without the quotes.