[LAD] Denormals / subnormals (again)

[Fons Adriaensen]

On Mon, Jan 02, 2012 at 09:30:56PM -0500, Paul Davis wrote:
> On Mon, Jan 2, 2012 at 9:17 PM, Fons Adriaensen <fons at linuxaudio.org> wrote:
> 
> > I don't agree with blaming the Intel architecture. Denormals are
> > just one tiny aspect of a much wider issue which is numerical
> > precision and stability, and that affects all architectures.
> > Anyone writing numerical or DSP code should be aware of this
> > and be able to analyse the algorithm. You don't produce good
> > numerical or DSP code by blindly implementing textbook equations.
> 
> absolutely.
> 
> but there are two aspects to this issue:
> 
>     (1) good numerical or DSP code
>     (2) not slowing the processor down by a factor of 10 or more
> 
> non-Intel architectures took care of (2), while leaving (1) for the
> programmer to tackle.

Again I don't agree these are two different issues.

'Taking care of (2)' may avoid the slowdown, but it may also
hide the real problem, which is that you are trying to do a
computation that is beyond the limits of what the FPU can do.
And that problem is *not* solved by ensuring there will be
no slowdown. In some cases replacing small values (of which
denormals are just one form) by zero or adding a small bias
may help. In other cases it doesn't and it just produces new
problems. Such a method is no substitute for analysis, not any
more than blindly changing floats into doubles and hoping for
the best.


Ciao,

-- 
FA

Vor uns liegt ein weites Tal, die Sonne scheint - ein Glitzerstrahl.