Support for IEEE 754 octuple precision

ginette.machin 3 years ago updated by Pavel Holoborodko 3 years ago 5

Advanpix already allows the declaration of quadruple precision as in the  IEEE 754 standard. 

It would be great to do the same with octuple precision since it is defined by this standard. 

My point here is not speed, but to test algorithms with this peculiar precision in a quite portable way.

Under review

Dear Ginette,

First of all, I am very happy to here from you! 

Hope you are doing fine and everything is good.

As for octuple precision, if we setup mp.Digits(71), toolbox will provide computations with ~99% compatibility with the standard. The only major difference would be exponent size  - IEEE requires 19 bits, whereas toolbox uses 32 bits (for better compatibility with CPU registers). 

Basically, if numbers stay inside IEEE exponent range (=most common case) - than computations with toolbox will be fully compatible with the standard. And also reproducible on different platforms.

In general, toolbox follows IEEE guidelines and requirements in implementation of arbitrary precision computations.


Maybe I am missing something here, do you have some particular compatibility test in mind?

Dear Pavel,

Thanks for your quick reply.

You did not miss something. 

I was just looking for a way 

to declare octuple precision, 

so that anyone can reproduce 

my results. The features of 

mp.Digits(71) seems sufficient 

for me.

If you use octuple precision most of the times, you can put the mp.Digits(71) into mpstartup.m.

This will make it the default precision for every MATLAB/toolbox session.

By the way, please update toolbox to the latest version (if you use older one). 

Now it has EIGS and many other improvements....