Re: [Help-gsl] eigen test fails on 32-bit (x86) Linux

[Patrick Alken]

Victor, sorry to take so long to apply your patch - it is now applied on
the git repository. Could you let me know if its working for you now?

Patrick

On 08/11/2016 02:40 PM, Victor Zverovich wrote:
> Hi Patrick and others,
>
> While testing GSL 2.1 I've noticed that poly_test has been fixed but
> eigen_test is still failing when compiled with -O3 on some platforms:
>
> FAIL: gen, direct eigenvalue(4) imag, random (-0.481216772353650846
> observed vs 0.481216772353650846 expected) [877968]
> FAIL: gen, direct eigenvalue(5) imag, random (0.481216772353650901
> observed vs -0.481216772353650846 expected) [877970]
> FAIL: gen, direct eigenvalue(15) imag, random (6.85872455924790447
> observed vs -6.85872455924790536 expected) [877990]
> FAIL: gen, direct eigenvalue(16) imag, random (-6.85872455924790536
> observed vs 6.85872455924790625 expected) [877992]
>
> Is it possible to fix eigen_test too e.g. by merging the patch
> I provided earlier?
>
> Thanks,
> Victor
>
> On Mon, Aug 18, 2014 at 4:11 PM address@hidden
> <mailto:address@hidden> <address@hidden
> <mailto:address@hidden>> wrote:
>
>     Indeed, using the version of poly/test.c from the repository fixes
>     the issue.
>
>     Thanks,
>     Victor
>
>
>     On Mon, Aug 18, 2014 at 3:06 PM, Patrick Alken
>     <address@hidden <mailto:address@hidden>>
>     wrote:
>
>         This should be fixed on the git repository (see bug #39055),
>         though it may be that the tolerances need adjusting for your
>         system.
>
>
>         On 08/18/2014 04:02 PM, address@hidden
>         <mailto:address@hidden> wrote:
>
>             There is another test failure of a similar kind, but now
>             without any -O
>             flags:
>
>             $ ./configure CFLAGS=
>             $ make
>             $ poly/test
>             FAIL: z1.real, 15th-order polynomial
>             (-0.999999999999994227 observed vs 1
>             expected) [159]
>             FAIL: z2.real, 15th-order polynomial (1.00000005616458409
>             observed vs -1
>             expected) [161]
>             FAIL: z13.real, 15th-order polynomial (1.99999988977862886
>             observed vs 2
>             expected) [183]
>             FAIL: z14.real, 15th-order polynomial (2.00000011022135338
>             observed vs 2
>             expected) [185]
>
>             Victor
>
>
>             On Mon, Aug 18, 2014 at 2:48 PM,
>             address@hidden
>             <mailto:address@hidden> <
>             address@hidden
>             <mailto:address@hidden>> wrote:
>
>                 Thanks for quick responses guys.
>
>                 Patrick, please find attached the patch that
>                 implements your suggestion
>                 which does fix the test failures. I've
>                 reused GSL_EIGEN_SORT_VAL_{ASC,DESC}, that were
>                 previously unused for
>                 complex numbers, for the new type of comparison.
>
>                 Victor
>
>
>
>                 On Mon, Aug 18, 2014 at 12:21 PM, Patrick Alken <
>                 address@hidden
>                 <mailto:address@hidden>> wrote:
>
>                     GSL is already sorting the eigenvalues based on
>                     magnitude, so the problem
>                     seems to be that with the extra -O3 optimization,
>                     the gsl_eigen_gen and
>                     gsl_eigen_genv routines are producing slightly
>                     different results, leading
>                     to a different sorting order. One way to fix this
>                     would be to sort by real
>                     part first and then by imaginary part, but it may
>                     be a little while until I
>                     can look into this further.
>
>                     Patrick
>
>
>                     On 08/18/2014 12:59 PM, Martin Jansche wrote:
>
>                         I'm not sure if GSL makes any guarantees about
>                         the direction of
>                         eigenvectors. It looks like the test might
>                         pass if one only looks at the
>                         magnitude of the values. Perhaps the test
>                         could be made robust against
>                         sign
>                         changes by looking at absolute values.
>
>
>
>                         On Mon, Aug 18, 2014 at 1:53 PM,
>                         address@hidden
>                         <mailto:address@hidden> <
>                         address@hidden
>                         <mailto:address@hidden>> wrote:
>
>                           Hello,
>
>                             I noticed that eigen test fails on 32-bit
>                             Linux (x86) when compiled with
>                             -O3 flag:
>
>                             $ ./configure CFLAGS="-O3"
>                             $ make
>                             $ eigen/test
>                             FAIL: gen, direct eigenvalue(4) imag,
>                             random (-0.481216772353650846
>                             observed vs 0.481216772353650846 expected)
>                             [877968]
>                             FAIL: gen, direct eigenvalue(5) imag,
>                             random (0.481216772353650901
>                             observed
>                             vs -0.481216772353650846 expected) [877970]
>                             FAIL: gen, direct eigenvalue(15) imag,
>                             random (6.85872455924790447
>                             observed
>                             vs -6.85872455924790536 expected) [877990]
>                             FAIL: gen, direct eigenvalue(16) imag,
>                             random (-6.85872455924790536
>                             observed vs 6.85872455924790625 expected)
>                             [877992]
>
>                             I'm using GSL version 1.16, 32-bit Ubuntu
>                             10.04 and GCC 4.4.3.
>
>                             Best regards,
>                             Victor
>
>
>
>
>
>