96, 128 bit floating point C++ library?

[KevinHall]

Anyone know of an open-source 96 or 128 bit floating point C++ library? (Or even better, a specifiable but fixed bit library?)

I have seen and used apfloat, but the one thing I don't like about it is that apfloat numbers extend its precision with certain operations -- meaning that I have to continually go back an tell it I only need n significant digits.

What I'm trying to say is that I'm looking for something like:

class Float96;
class Float128;

or:

template<size_t bits>
class NBitFloat
{
//...
};

Thanks guys!

- Kevin

[galathaea]

Have you tried doubledouble from the oonumerics site? They also have EXTNUM, which changes only the exponent precision, and I believe LiDiA and some of the other packages use earlier forms of doubledouble.

[KevinHall]

No, I hadn't seen it yet. It certainly looks easy to use.... I'll give it a whirl right now! Thanks!!!

-Kevin

[Paul McKenzie]

You can try the 96bit library on PJ Naughter's site:

http://www.naughter.com/int96.html

Even though it says "MFC", very litte of it is MFC and won't take long to convert it or just plain rip it out.

Regards,

Paul McKenzie

[KevinHall]

Thanks guys for the help. I really needed to test a routine for singular value decomposition of matrices -- there is one matrix that I'm having trouble with docomposing accurately. I wanted to make sure it was a precision thing (and that I wasn't going insane). It turns out it is!

I think that the doubledouble library is pretty simple for doing tests like this one. Thanks again!

- Kevin