floating point inconsistency

Christoph Groth cwg@falma.de
Wed Feb 17 16:09:00 GMT 2010


Vincent Lefevre <vincent+gcc@vinc17.org> writes:

> On 2010-02-16 18:24:02 +0100, Christoph Groth wrote:
>> It's not broken but I would like the runs to be reproducible, i.e. to
>> get exactly (bit by bit) the same results when starting with the same
>> parameters and the same random seed on all the platforms I use.
>
> Then you should probably use CRlibm[1] (fast, for double precision)
> or MPFR[2] (much slower, because not targeted at a fixed precision).
>
> [1] https://lipforge.ens-lyon.fr/projects/crlibm/
> [2] http://www.mpfr.org/
>
> CRlibm is used by the LHC@home project at CERN precisely for this
> reason.

Thanks for the suggestions -- I stumbled across exactly these two
libraries yesterday and found that crlibm solved my problem.  Works like
a charm!  Compared to standard libm it does not noticeably slow down my
calculations.

There seem to be many French guys in this business of verifiable
computation.

Cheers

Christoph



More information about the Gcc-help mailing list