[Jazzlib-developers] Higher performance for Adler32

[Thomas M. <hyp...@gm...>]

Hi, 

I have made some speed improvements for your Adler32 implementation.
In my test, the old code takes 1731 ms, my code takes 1061 ms.

public void update(byte[] buffer, int off, int len) {
    int s1 = checksum & 0xffff;
    int s2 = checksum >>> 16;
    while (off < len) {
        int max = Math.min(off + DEFER, off + len);
        while (off < max) {
            s1 = (buffer[off++] & 0xff) + s1;
            s2 += s1;
        }
        s1 %= BASE;
        s2 %= BASE;
    }
    checksum = (s2 << 16) | s1;
}

The important changes are:

- only increment one value (off++) in the main loop
  instead of (off++) and (n--)

- s1=...+s1 is a lot faster than s1=s1+...
  for some unknown reason (maybe the processor pipeline?)

DEFER is a new constant:

private final static int DEFER = 3850;

I calculated this value like this:

static {
    int i=0, a=BASE, b=a;
    for(;b>0;i++)
        b += (a += 255);
    System.out.println("max defer = " + (i-1));
}

I hope this helps,
Thomas
(developer of the original Hypersonic SQL)

-- 
Handyrechnung zu hoch? Tipp: SMS und MMS mit GMX
Seien Sie so frei: Alle Infos unter http://www.gmx.net/de/go/freesms