Re: Some failed tests in Octave 3.3.53 [more details]

[Ben Abbott]

On Nov 5, 2010, at 9:09 PM, Jordi Gutiérrez Hermoso <address@hidden> wrote:

> 2010/11/4 José Luis García Pallero <address@hidden>:
>> 2. Test failed in chol.cc (FAIL 1). This error appears in all
>> versions: 3.3.51, 3.3.53 and current hg source code. The log file
>> says:
>> 
>> !!!!! test failed
>> assert (norm (R1 - R, Inf) < 1e1 * eps) failed
>> shared variables {
>>  A =
>> 
>>     0.436997  -0.131721   0.124120  -0.061673
>>    -0.131721   0.738529   0.019851  -0.140295
>>     0.124120   0.019851   0.354879  -0.059472
>>    -0.061673  -0.140295  -0.059472   0.600939
>> 
>>  u =
>> 
>>     0.98950
>>     0.39844
>>     0.63484
>>     0.13351
>> 
>>  Ac =
>> 
>>   Columns 1 through 3:
>> 
>>     0.55855 + 0.00000i  -0.16621 - 0.03153i   0.01079 + 0.02364i
>>    -0.16621 + 0.03153i   0.67601 + 0.00000i   0.00115 - 0.04755i
>>     0.01079 - 0.02364i   0.00115 + 0.04755i   0.62631 + 0.00000i
>>    -0.02768 + 0.01861i   0.01460 - 0.02476i  -0.15858 + 0.07198i
>> 
>>   Column 4:
>> 
>>    -0.02768 - 0.01861i
>>     0.01460 + 0.02476i
>>    -0.15858 - 0.07198i
>>     0.60342 + 0.00000i
>> 
>>  uc =
>> 
>>     0.54267 + 0.91519i
>>     0.99647 + 0.43141i
>>     0.83760 + 0.68977i
>>     0.39160 + 0.90378i
>> 
>> If I do test chol the results are the same than for make check
> 
> What BLAS are you using? Most of those computations are handled by
> BLAS, so you could have found a bug in its implementation. My guess is
> you're probably using ATLAS. Can you please check?

I see this failure about 10% of the time with Apple's vecLib.

I suggest the test be run a few times to see if it is reproducible.

Ben