Re: [Bug-gsl] multifit test failures for Solaris x86 with libgsl 2.1

[Rich Burridge]

On 12/09/2015 12:13 PM, Orion Poplawski wrote:
> On 12/08/2015 05:12 PM, Rich Burridge wrote:
> > Hi,
> >
> > I'm finally getting around to integrating libgsl 2.1 into Solaris 12.
> > One thing I've noticed is that I'm now seeing multifit test failures
> > on the x86 platform but only for 32-bit (64-bit is fine):
> >
> > ...
> > Making check in multifit
> > /usr/gnu/bin/make  test
> > /usr/gnu/bin/make  check-TESTS
> > FAIL: lmsder/scale=1/wnlin_internal_weights i=0 (5.17378552263401126 observed
> > vs 5.17378551196259195 expected) [190002]
> > FAIL: lmsder/scale=1/wnlin_internal_weights i=1 (0.111041758485330988 observed
> > vs 0.111041758006851149 expected) [190003]
> > FAIL: lmsder/scale=1/wnlin_internal_weights i=2 (1.05282724326308363 observed
> > vs 1.05282724070446099 expected) [190004]
> > FAIL: lmsder/scale=1/weights/wnlin_internal_weights i=0 (5.17378552263401126
> > observed vs 5.17378551196259195 expected) [190007]
> > FAIL: lmsder/scale=1/weights/wnlin_internal_weights i=1 (0.111041758485330988
> > observed vs 0.111041758006851149 expected) [190008]
> > FAIL: lmsder/scale=1/weights/wnlin_internal_weights i=2 (1.05282724326308363
> > observed vs 1.05282724070446099 expected) [190009]
> > FAIL: ridge/lmsder/wnlin_internal_weights i=0 (5.17378552263401126 observed vs
> > 5.17378551196259195 expected) [190022]
> > FAIL: ridge/lmsder/wnlin_internal_weights i=1 (0.111041758485330988 observed
> > vs 0.111041758006851149 expected) [190023]
> > FAIL: ridge/lmsder/wnlin_internal_weights i=2 (1.05282724326308363 observed vs
> > 1.05282724070446099 expected) [190024]
> > FAIL: lmsder/scale=1/wnlin_external_weights i=0 (5.17378552263401126 observed
> > vs 5.17378551196259195 expected) [190042]
> > FAIL: lmsder/scale=1/wnlin_external_weights i=1 (0.111041758485330988 observed
> > vs 0.111041758006851149 expected) [190043]
> > FAIL: lmsder/scale=1/wnlin_external_weights i=2 (1.05282724326308363 observed
> > vs 1.05282724070446099 expected) [190044]
> > FAIL: ridge/lmsder/wnlin_external_weights i=0 (5.17378552263401126 observed vs
> > 5.17378551196259195 expected) [190052]
> > FAIL: ridge/lmsder/wnlin_external_weights i=1 (0.111041758485330988 observed
> > vs 0.111041758006851149 expected) [190053]
> > FAIL: ridge/lmsder/wnlin_external_weights i=2 (1.05282724326308363 observed vs
> > 1.05282724070446099 expected) [190054]
> > FAIL: test
> > ==================
> > 1 of 1 test failed
> > ==================
> > make[3]: [check-TESTS] Error 1 (ignored)
> > ...
> >
> > Is this another case where the tolerances just need to be tweaked,
> > similar to what we recently did with the linalg tests?
>
> Looks like:
>
> http://savannah.gnu.org/bugs/?46593
> http://lists.gnu.org/archive/html/bug-gsl/2015-11/msg00011.html

Finally getting back to this.

Based on what Patrick suggested at:

  http://lists.gnu.org/archive/html/bug-gsl/2015-12/msg00018.html

I found that the following change works nicely for me:

--- gsl-2.1/multifit/test_wnlin.c.orig  2016-01-12 08:26:06.987977367 -0800
+++ gsl-2.1/multifit/test_wnlin.c       2016-01-12 12:29:34.092460907 -0800
@@ -4,7 +4,7 @@
 #define wnlin_NTRIES    1

 static double wnlin_x0[wnlin_P] = { 1.0, 0.0, 0.0 };
-static double wnlin_epsrel = 1.0e-9;
+static double wnlin_epsrel = 1.0e-8;

 static int wnlin_internal_weight = 1;

and all GSL multifit tests now PASS on x86 32-bit.

Thanks!