Unit test for grpstats fails on i368

[rlaboiss]

One of the unit tests for grpstats fails on an i368 system, producing this error message:

octave:5> test grpstats
***** test
 ## large sample predci
 x = (1:50)';
 g = [ones(25, 1); 2 * ones(25, 1)];
 ci = grpstats (x, g, 'predci');
 expected = [-2.49070107176829 28.4907010717683; 22.5092989282317 ...
             53.4907010717683];
 assert (ci, expected, 1e-14);
!!!!! test failed
ASSERT errors for:  assert (ci,expected,1e-14)

  Location  |  Observed  |  Expected  |  Reason
   (2,1)       22.5093      22.5093      Abs err 1.4211e-14 exceeds tol 1e-14 by 4e-15
   (1,2)       28.4907      28.4907      Abs err 1.4211e-14 exceeds tol 1e-14 by 4e-15
   (2,2)       53.4907      53.4907      Abs err 1.4211e-14 exceeds tol 1e-14 by 4e-15

It works on all other architectures supported by Debian.

If it would help with the diagnosis, here is a more detailed breakdown of the above issue:

octave:4> x = (1:50)';
octave:5> g = [ones(25, 1); 2 * ones(25, 1)];
octave:6> grpstats (x, g, 'predci') - [-2.49070107176829 28.4907010717683; 22.5092989282317 53.4907010717683]
ans =

   4.884981308350689e-15  -1.421085471520200e-14
   1.421085471520200e-14  -1.421085471520200e-14

For comparison, the following was obtained on an amd64 system:

octave:4> x = (1:50)';
octave:5> g = [ones(25, 1); 2 * ones(25, 1)];
octave:6> grpstats (x, g, 'predci') - [-2.49070107176829 28.4907010717683; 22.5092989282317 53.4907010717683]
ans =

  -3.9968e-15  -7.1054e-15
   7.1054e-15  -7.1054e-15

[pr0m1th3as]

Can you point out package and Octave release?

[rlaboiss]

Sure :

• octave 10.3.0
• datatypes 1.1.8
• statistics 1.8.1

[cbm755]

Drive-by comment, but this is a good example of where relative tolerances are probably better...

-assert (ci, expected, 1e-14);
+assert (ci, expected, -1e-14);

This is b/c 1e-14 / 22.5093 is 4.4426e-16, which is only twice machine epsilon.

[dasergatskov]

Another drive-by comment. The failure is most likely due to i386 buid using fp87 math processor rather than SSE. This architecture is pretty much deprecated by octave. BISTs in Octave core are failing left and right:
https://octave.discourse.group/t/octave-11-0-92-release-candidate-available-for-testing/7316/14

Summary:

  PASS                            22190
  FAIL                               33
  REGRESSION                          1
  XFAIL (reported bug)               33
  SKIP (missing feature)            109
  SKIP (run-time condition)          85

Back when octave was providing binaries for 32-bit Windows it was built with -msse.

[pr0m1th3as]

I took the easy path of increasing the tolerance. Besides that, I wonder who is actually using i386 hardware and OS these days?

[rlaboiss]

Although the i386 architecture is “officially” supported in the Debian 13 (a.k.a. trixie) release, the
reduced support for i386 has been announced.

[pr0m1th3as]

Thanks for the Debian info. Closing this issue as fixed.