Steve Harris wrote:
> OK, Sounds reasonable, denormalise didn't quite seem like the right word
> to me either, but I didn't really understand it! To be honest I havn't
> done any real measurments to see how much it helps, but it did seem to
> stop the exceptional slowdowns when feeding in a stream of very low
> numbers.
It depends on how often you check for denormals, which I haven't tried
to optimize, but I get numbers like this for an iterated multiply:
naive FTZ() add DC
normalized 10.0 10.6 10.3 cycles/iteration
denormal 260 11.1 11.5
--
Eli Brandt | [EMAIL PROTECTED] | http://www.cs.cmu.edu/~eli/
