[SciPy-user] problems with scipy.test() on Opteron 64-bit

Pearu Peterson pearu at scipy.org
Thu Nov 25 09:14:23 CST 2004



On Thu, 25 Nov 2004, Giovanni Samaey wrote:

>> If those fail as well then try a C program that uses dgeev to solve the 
>> same eigenvalue problem:
>> 
>> $ gcc main.c -llapack
>> $ ./a.out
>> wr=9.32183,-6.20979e-16,-0.321825,wi=0,0,0
>
> I get (after adding some -lblas -lm and paths etc...)
> ./a.out
> wr=9.43719,-0.115365,-0.321825,wi=0,0,0
>
> which probably shows that the error comes from lapack.

Yes.

> (This is with the lapack I compiled myself in an intermediate step to build 
> atlas for scipy.)
> Just to repeat how I compiled lapack: I took the Makefile and added -fPIC and 
> -m64 there to give it the same compiler options as that atlas makefile.

Now try building Fortran lapack without optimization flags and see if you 
still get incorrect results.

I think it is reasonable to check also if the incorrect results are 
related to gcc 3.2 that you are using. I might be wrong but gcc 3.2
is probably older than Opterons..

Pearu



More information about the SciPy-user mailing list