Re: [Bug-gsl] 'make check' failes with Intel Compiler 10

[Stefan Gerlach]

Hi,

I'm using the Intel Compiler since they produce faster code in many cases.
Let me know if i should try to find out more informations, otherwise i'll live 
with it :-)

Thanks for your help.

-- 
Stefan Gerlach

> At Thu, 3 Apr 2008 11:33:00 +0200,
> Stefan Gerlach wrote:
> > compiling gsl-1.11 on openSUSE 10.2 (x86_64) with Intel C Compiler 10 i get 
> > the following problems during make check :
> > 
> > ...
> > FAIL: gsl_sf_angle_restrict_symm_err_e(-1e9, &r) [2769]
> >   expected: -5.7739542350138517e-01
> >   obtained: -5.7739543914794922e-01 +/- 4.4408920959364753e-07 
> > (rel=7.69124e-07)
> >   fracdiff: 1.3549262078941325e-08
> >  tolerance: 3.5527136788005009e-15
> >   value not within tolerance of expected value
> >    -5.773954391479492188e-01  4.440892095936475318e-07
> > FAIL:   gsl_angle_restrict_pos_e: theta = 50000.0 + 1.0/65536.0 [2775]
> > FAIL:   gsl_angle_restrict_pos_e: theta = 5000000.0 + 1.0/65536.0 [2776]
> > FAIL:   gsl_angle_restrict_symm_e: theta = 50000.0 + 1.0/65536.0 [2782]
> > FAIL: Trigonometric and Related Functions [2783]
> > FAIL: test
> > ===================
> > 1 of 1 tests failed
> > ===================
> > 
> > this is the last of a long list of failed tests in specfunc/
> 
> Thanks for your email.  
> 
> Most likely many of the failures are errors just above the test limit.
> 
> In test_sf.h you'll find two factors
> 
>    #define TEST_FACTOR 100.0  
>    #define TEST_SIGMA 1.5  
> 
> If you increase these you may find that some of the errors go away.
> 
> In the example above, it looks like the formula in
> gsl_sf_angle_restrict_symm_err_e has broken down and caused a loss of
> precision.  That could be a problem with the compiler.  If you used
> optimisation you could try compiling without optimisation.  As icc is
> proprietary software I don't run any tests with it myself, only gcc.
>

pgpWmUm9TEkbG.pgp
Description: PGP signature