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

Pearu Peterson pearu at scipy.org
Wed Nov 24 14:14:01 CST 2004



On Wed, 24 Nov 2004, Giovanni Samaey wrote:

> Pearu Peterson wrote:
>
>> 
>> 
>> On Wed, 24 Nov 2004, Giovanni Samaey wrote:
>> 
>>> Since I don't use linalg myself, I do not have any problems using SciPy 
>>> like this. But I would appreciate this being solved before giving access 
>>> to others. Since we plan on encouraging new students to use this 
>>> environment instead of matlab, any help would be welcome.
>> 
>> 
>> On a 64-bit platform you'll need F2PY version 2.43.239_1831, at least.
>> Also upgrading Numeric would be recommended (make sure that new
>> Numeric header files are copied to install directory as well).
>
> OK -- I succeeded in recompiling atlas version 3.7.8, Numeric 23.6 (I also 
> tried Numeric 23.5 ), and the latest
> f2py (2.43.239_1844).  Python version is 2.3.4. However, I use gcc3.3.2.
> Now the failed scipy tests reduce to only 3 (and 3 times the same computation 
> that fails).

> The lapack and blas were compiled by adding -fPIC -m64 to their options, to 
> make the compiler options compatible with those of atlas.
>
> ======================================================================
> FAIL: check_simple (scipy.linalg.decomp.test_decomp.test_eig)
> ----------------------------------------------------------------------
> Arrays are not almost equal (mismatch 66.6666666667%):
>       Array 1: [ 9.4371906+0.j -0.1153653+0.j -0.3218254+0.j]
>       Array 2: [ 9.3218254  0.        -0.3218254]
>
>
> ======================================================================
> FAIL: check_simple (scipy.linalg.decomp.test_decomp.test_eigvals)
> ----------------------------------------------------------------------
> Arrays are not almost equal (mismatch 66.6666666667%):
>       Array 1: [ 9.4371906+0.j -0.1153653+0.j -0.3218254+0.j]
>       Array 2: [ 9.3218254  0.        -0.3218254]
>
>
> ======================================================================
> FAIL: check_simple_tr (scipy.linalg.decomp.test_decomp.test_eigvals)
> ----------------------------------------------------------------------
> Arrays are not almost equal (mismatch 66.6666666667%):
>       Array 1: [ 9.4371906+0.j -0.1153653+0.j -0.3218254+0.j]
>       Array 2: [ 9.3218254  0.        -0.3218254]

It is interesting that only simple eig tests fail while complex and random
tests pass. And in these simple tests eigenvalues are only slightly 
incorrect.

Could you try building scipy against Fortran BLAS/LAPACK libriares to
see if your ATLAS libraries are ok? If ATLAS is OK, then you could try 
f2py from cvs though I can't remember that it would include any 64-bit 
specifix fixes.

Pearu



More information about the SciPy-user mailing list