Re: effectiveness of gnu.math

[Per Bothner <per at bothner dot com>]

Yaroslav Kavenchuk wrote:
> Is BigInteger slower gnu.math?

I'd like to think so, but when I tried it I got that gnu.math
was slower.  gnu.math should be faster when working with small
integer (which is the common case).  However, I don't understand
why gnu.math is slower for the larger numbers, since it uses
basically the same algorithm, as far as I can tell.  I also don't
understand why you got different results.  (However, I tried 20000
rather than 50000 presumably because I'm using a slower machine.)

I'd love to find out what is going on.  Would you mind trying to
run some profiling tool (e.g. http://profiler.netbeans.org/)
to see where the <integer> version is spending its time?
I used the code at the end of this message, and compiled it with -C,
and then tried:

$ time /opt/jdk1.5/bin/java kawa.repl -e '(require <foo>)' -e '(il 20000)'

vs

$ time /opt/jdk1.5/bin/java kawa.repl -e '(require <foo>)' -e '(il 
(java.math.BigInteger:valueOf 20000))'

(For some reason my code would automatically convert 20000 to a
BigInteger, so I did it manually.)

> Is untyped code slower typed code very slightly?

Sometimes it can be a lot slower.  However, in this case most
of the time is spent doing (I would guess) multiplications,
so that the difference made by type declaration would be minor.

(define (bil x :: <java.math.BigInteger>) :: <java.math.BigInteger>
 (do (((i :: <java.math.BigInteger>) x (i:subtract 
java.math.BigInteger:ONE))
      ((k :: <java.math.BigInteger>) java.math.BigInteger:ONE 
(k:multiply i)))
;     ((= i 1) k)))
     ((i:equals java.math.BigInteger:ONE) k)))
(define (il x :: <integer>) :: <integer>
 (do (((i :: <integer>) x (- i 1))
      ((k :: <integer>) 1 (* k i)))
     ((i:equals 1) k)))

--
	--Per Bothner
per@bothner.com   http://per.bothner.com/