[Home]FastMath/SquareRoot

Robo Home | FastMath | Changes | Preferences | AllPages

Difference (from prior major revision) (minor diff)

Changed: 19c19,21
I think a lot of these functions are implemented in hardware, so making interpreted software variants may not help. -- Kawigi
I think a lot of these functions are implemented in hardware, so making interpreted software variants may not help. -- Kawigi

Yeah, square roots should be handled in the math units in your processor. I recall reading somewhere they took about 60-70 cycles on a Pentium II era processor (don't have a source for this), which is slow for hardware, but still blowing away anything you'll ever do in software. This means picking a faster algorithm helps jack all if you're moving it from a one-step in pure hardware to a series of math calls in a software virtual machine :p If this were C or something then we could probably still talk, but in java I think this is just go-kart effect that traps some developers (oo.. I wrote a faster algorithm I can save cycles.. uh.. let's discount the fact that I'm probably spending the 70 cycles the original took in overhead for the while statement on it's own). -- Kuuran

From http://www-106.ibm.com/developerworks/library/wi-elite17.html?ca=dnt-435

    // very fast square root approximation
    public final static int sqrt( long val ) {
        long temp, g = 0, b = 0x8000, bshft = 15;
        do {
            if( val >= ( temp = ( ( ( g << 1 ) + b ) << bshft-- ) ) ) {
                g += b;
                val -= temp;
            }
        } while( ( b >>= 1 ) > 0 );
        return (int)g;
    }

Hang on... that casts the answer into an int. That's a little TOO approximate for my liking. -- Tango

It's also about 10 times slower than Math.sqrt() on my machine... Maybe this whole idea was bad. :-p --David Alves

I think a lot of these functions are implemented in hardware, so making interpreted software variants may not help. -- Kawigi

Yeah, square roots should be handled in the math units in your processor. I recall reading somewhere they took about 60-70 cycles on a Pentium II era processor (don't have a source for this), which is slow for hardware, but still blowing away anything you'll ever do in software. This means picking a faster algorithm helps jack all if you're moving it from a one-step in pure hardware to a series of math calls in a software virtual machine :p If this were C or something then we could probably still talk, but in java I think this is just go-kart effect that traps some developers (oo.. I wrote a faster algorithm I can save cycles.. uh.. let's discount the fact that I'm probably spending the 70 cycles the original took in overhead for the while statement on it's own). -- Kuuran


Robo Home | FastMath | Changes | Preferences | AllPages
Edit text of this page | View other revisions
Last edited February 20, 2004 10:54 EST by Kuuran (diff)
Search: