S.W.Harris at ecs.soton.ac.uk
Tue Apr 27 08:22:16 PDT 2004
On Tue, Apr 27, 2004 at 10:41:31 -0400, Ross Vandegrift wrote:
> On Tue, Apr 27, 2004 at 11:16:58AM +0300, Sampo Savolainen wrote:
> > I've a few had problems which seem like denormalisation issues. In case
> > someone reading this doesn't know what it is, you should read
> > http://phonophunk.phreakin.com/p4denormal.php . In short, it's got to do
> > with processors handling very very very small floating point numbers very
> > slowly.
> This problem should be easily fixable in software. According to IEEE
> 754, denormalizing results in loss of significance and starts a gradual
> underflow. (there's a version at:
> http://www.cs.berkeley.edu/~wkahan/ieee754status/IEEE754.PDF - I'm
> looking at pages 2 and 15)
> If you're just going to set a register that sets denormals to zero, why
> not do it right and detect underflows? When an underflow happens, just
> stop banging the processor against the wall.
Yeah, unfortunatly that means killing potential denormals everywhere they
might occur, and that's a lot of places.
More information about the Ardour-Dev