denormal floats (was Re: [linux-audio-dev] XAP spec - early scribbles)

[Steve Harris]

On Thu, Feb 06, 2003 at 03:59:31 +0100, Dave Griffiths wrote:
> On Thu, 6 Feb 2003 14:07:06 +0000, Steve Harris wrote
> > On Thu, Feb 06, 2003 at 12:18:47 +0100, Dave Griffiths wrote:
> > > I had a vague attempt at doing something like this (after noticing that
> > > filters filtering silence uses up a lot of cpu). Each sample buffer object
> > 
> > OT: Thats probably because the zeros weren't 0.0, they were probably
> > denormal numbers. If you squash them the filter will run at the same 
> > cost as when its processing any other data.
> 
> ahah, I was hoping for an explanation :) any ideas on how to combat this, what
> the squashing threshold should be?

I use:

#define FLUSH_TO_ZERO(fv) (((*(unsigned int*)&(fv))&0x7f800000)==0)?0.0f:(fv)
I think it came from the music-dsp list.

Its uses pointer casting, which is mildly evil, but C doesn't have a
non-interprting cast like C++ has, I cant remeber the C++ syntax, but it
should be easy to convert.

It looks for the denormal pattern in the IEEE bitpattern and zeros the
float if its there IIRC. You probably can't do the comparison with < as
that would cause the denormal to be evaluated.

- Steve