Help: Numerical precision issues?

[KimB]

Alternatively, if you have access to a Solaris SPARC machine, they have software libraries for IEEE quad precision arithmetic (113 mantissa bits), although I'm not sure what the state of their special function implementations are.

Regards,
Serge

Actually, I do have access to Seaborg at NERSC which makes use of 106-bit mantissa long doubles (on IBM Power3 processors). I had been having troubles with them, though. But I just now just went back to check on the issues, based on Arjan's post:

One note: the C exp() and log() functions expect a 'double' argument, so your long-doubles will be downconverted to ordinary doubles. Check if your C compiler supports the functions expl() and logl() (long double versions of the exp and log functions).

It looks like IBM's xlC compiler does not overload the log, exp functions for long doubles, and has a bug when using them as arguments, so I had some issues when I attempted to run my code there. It works now, though, and gives 10^-31 precision, so I'm going to give my code a go there and see if I notice the same creep with change in lmax.

Thanks for the ideas, guys!