[SciPy-user] run=2339 errors=0 failures=3 on ibex

[Xavier Gnata]

David Cournapeau wrote:
> Xavier Gnata wrote:
>   
>> Ok so there is a simple fix which is fully correct:
>>
>> -assert_array_almost_equal(x, y, decimal=6, err_msg='', verbose=True)
>>
>> +assert_array_almost_equal(x, y, decimal=5, err_msg='', verbose=True)
>>   
>>     
>
> The obvious question is does that fix the issue or does it only hide the
> problem ? 5 decimals is relatively poor, but float32 can only be
> expected to have 7 decimals anyway. And assuming the failure is caused
> by different BLAS/LAPACK implementations, it is not rare to see
> relatively significant differences between two given BLAS/LAPACK (even
> same version but different compilers; I have never seen a BLAS/LAPACK
> passing all the netlib LAPACK tests, for example: g77 and gfortran break
> them at different places, for example).
>
>   
Ok my fault. It was not a good idea to "fix" it that way.
So first I can try to implement the same computation in C/whatever using 
float32 and see how many correct digits we get.
If it "breaks" on some systems, IMHO it means that the check is too 
stringent.

> Did you use g77 before intrepid (intrepid finally use gfortran as the
> default ABI for fortran) ?
>
>   
Full gfortran. No g77.

So what should be the way forward?  Should we involve gfortran/lapack 
guys asking for how many digits we should get?
Should we move from an error to a warning (because the result may only 
be not as accurate as expected).

As you know, I really would like to see scipy compiling and running the 
test suite smoothly :)
It is very important to "sell" scipy to people (people I know at least ;)).

Cheers,
Xavier


> cheers,
>
> David
> _______________________________________________
> SciPy-user mailing list
> SciPy-user at scipy.org
> http://projects.scipy.org/mailman/listinfo/scipy-user
>