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

[Pearu Peterson]

On Thu, 25 Nov 2004, Giovanni Samaey wrote:

> What is remarkable is that, the error is the same in the 3 cases.  Moreover, 
> if you add up the first two components of array1 you get the first 
> component of array2.

It would look like atlas bug but since the same results are obtained with 
Fortran blas/lapack, then I don't know what is happening. I presume that
you did
   rm -rf build
each time before rebuilding scipy with different blas/lapack libraries.

What happens when you call geev routine directly? Try for example:

In [1]: import scipy

In [2]: scipy.linalg.flapack.dgeev([[1,2,3],[1,2,3],[2,5,6]],0,0)
Out[2]:
(array([  9.32182538e+00,  -6.03426271e-16,  -3.21825380e-01]),
  array([ 0.,  0.,  0.]),
  array([       [ 0.,  0.,  0.]]),
  array([       [ 0.,  0.,  0.]]),
  0)

In [3]: scipy.linalg.flapack.sgeev([[1,2,3],[1,2,3],[2,5,6]],0,0)
Out[3]:
(array([  9.32182503e+00,  -2.82419336e-07,  -3.21825117e-01],'f'),
  array([ 0.,  0.,  0.],'f'),
  array([       [ 0.,  0.,  0.]],'f'),
  array([       [ 0.,  0.,  0.]],'f'),
  0)

In [4]: scipy.linalg.calc_lwork.geev('d',3)
Out[4]: (12, 102)

Is there any difference in your case?

Pearu