Re: [AD] [ alleg-Bugs-1650580 ] Problem in fixed point library.

[Elias Pschernig <elias@xxxxxxxxxx>]

On Mon, 2007-02-05 at 07:45 -0700, Thomas Fjellstrom wrote:
> On Monday 05 February 2007 7:43 am, Elias Pschernig wrote:
> > On Tue, 2007-02-06 at 00:42 +1100, Peter Wang wrote:
> > > I retried the benchmark in that thread, and fixdivl now beats fixdivf on
> > > my machine by ~10%, where as it was slower by ~15% one and a half years
> > > ago.  With a couple of other silly benchmarks fixdivl wins by an even
> > > greater margin.  Probably the only relevant difference to my setup since
> > > then would be a newer version of gcc (3.4.6 vs 3.3.4).
> > >
> > > Anyone up for some benchmarking?
> >
> > Results here are:
> > with -O3 -march=pentium4: fixdivf is ~60% faster
> > with -O3 -ffast-math: fixdivf is ~10% faster
> > with -O3: fixdivf is ~10% faster
> > with -O2: fixdivf is ~10% faster
> > with -O1: fixdivf is ~10% faster
> > with -O0: fixdivl is ~250% faster (irrelevant of course)
> 
> I'd like to see -Os as well. :)
> 

In fact, -Os didn't behave as I thought: fixdivl is ~80% faster
So -Os is much closer to -O0 than -O1 (and I always thought it's about
like -O2..)

I also should have noted, I did not use test.c, but my standard
benchmarking setup. In the fastest case above, my P4 3.2 MHz runs 50
million fixfivf in a second, and the most fixdivl per second were 30
million.

-- 
Elias Pschernig