Karim Belabas on Wed, 24 Sep 2014 12:52:29 +0200


[Date Prev] [Date Next] [Thread Prev] [Thread Next] [Date Index] [Thread Index]

Re: 32-bit test-real fails


* Jeroen Demeyer [2014-09-24 12:34]:
> Dear pari-dev,
> 
> the recent changes to src/test/in/real affect 32-bit test-all results.
> The reason is simple: in the 64-bit branch you change the realprecision to
> 2, but not in the 32-bit branch. Doing that in both branches would fix the
> problem.

I actually fixed it this morning :-) (By removing the offending test,
I'll make a more comprehensive one later; what it tests was unrelated to
the source code changes in the commit that introduced it)

> However, I don't know if you consider the following difference a bug or not:
> 
> 
> i686 running linux (ix86/GMP-5.0.2 kernel) 32-bit version
> ? \p38
>    realprecision = 38 significant digits
> ? 1+0e-1
> %1 = 1.000000000
> 
> 
> amd64 running linux (x86-64/GMP-5.0.2 kernel) 64-bit version
> gp> \p38
> gp> 1+0e-1
> %1 = 1.0000000000000000000

No. In the current floating point model, precision must vary by 64
(resp. 32) full bits depending on the architecture. In that case, we
should have 4 bits of accuracy, "rounded up" to 64 (resp. 32) bits.

Cheers,

    K.B.
--
Karim Belabas, IMB (UMR 5251)  Tel: (+33) (0)5 40 00 26 17
Universite de Bordeaux         Fax: (+33) (0)5 40 00 69 50
351, cours de la Liberation    http://www.math.u-bordeaux1.fr/~kbelabas/
F-33405 Talence (France)       http://pari.math.u-bordeaux1.fr/  [PARI/GP]
`