This is the mail archive of the
mailing list for the GCC project.
denormals/subnormals are heading for extinction
- To: gcc at gcc dot gnu dot org
- Subject: denormals/subnormals are heading for extinction
- From: "Thomas R. Truscott" <trt at cs dot duke dot edu>
- Date: Thu, 16 Aug 2001 12:16:34 -0400 (EDT)
Perhaps denormals/subnormals are off-topic for this list,
but I cannot resist responding to certain statements:
> However, most architectures do adequately support denormals.
What?!! Performance on denormals is abysmal on
typical implementations of the following architectures:
ALPHA, SPARC, PA-RISC, IA32, IA64
IBM's POWER and PPC implementations have good performance on denormals.
(Are there any others?)
> Flush-to-zero fails.
> For explicit examples, see James Demmel's
> "Underflow and the Reliability of Numerical Software."
What sort of non-theoretical evidence is there?
The cited paper is dated 1984, before most of the
architectures listed above even existed.
In the 1990s, when IBM's POWER architecture was announced,
the numerical theorists were quite upset with
its "fused multiply-add", and dire predictions
appeared in comp.risks and elsewhere.
Over time, a different reality emerged.
Theoretically speaking, I think denormals are just fine.
But there is a major disconnect between theory and practice.