[Cosmo-torun] 200-bit precision floating point arithmetic and trigonometry?

[Boud Roukema]

hi cosmo-torun,

This follows from a conversation some of us had recently.
Debian stable (etch) includes GMP and MPFR - i ran a test program
which quite nicely does e.g. 200-bit floating point operations,
including trig functions, and gets reasonable answers. :)

Source + doc:
   http://gmplib.org/
   http://www.mpfr.org/

These are both under the LGPL instead of the GPL, so it turns out
that... the non-free program Mathematica uses the free program GMP !
So apparently non-free software libraries are not good enough for
Mathematica for this purpose. Or Mathematica is just "abusing the
commons".  [the commons: http://en.wikipedia.org/wiki/The_commons ]

pozdr
boud




On Wed, 14 Jan 2009, Bartosz Lew wrote:

>>> But within standard types long double is better than double etc.
>>
>> i assume this means you've experimented with this? Based on the
>> wikipedia entries, i suggested to Zbyszek that this could be
>> interesting, but i haven't tried it. My understanding is that it
>> pushes to the limit of the machine you're using, which can be better
>> than 64bit, but is nowhere near 128bit. :)
>
> mmm, don't remember which means that probably not much.
>
> But, I never experienced a need for such extremely super-duper high accuracy.
> As far as the calculation is actually correct within  the given precision it's cool.
> Eg. CMBFast works in 4byte float numbers and it's sufficient.
> AFAIK people still very often use 4byte floats.
> My opinion is that is the code returns wrong numbers its an indication that it should be rewritten
> in smarter way rather than going to long doubles or quadruple precisions.
>
>>
>> Tomorrow Thursday i've got linux w IF so i'm very unlikely to be at
>> KRA. But that's not an argument against both of you coming here and
>> doing stuff. :)
>
> ok.
>
> B.
>