On 11/11/2016 02:29 PM, Andreas Tille wrote:
>         Start 197: AmberFF_test
> 197/334 Test #197: AmberFF_test .................................***Failed    
> 0.19 sec
> (...)
>     (line 511 TEST_REAL_EQUAL(r4_r1 - r4_i, r1_r4 - r1_i): got -145.471, 
> expected -103.957)  - 
>     (line 513 TEST_REAL_EQUAL(r1_r4 - r1_i + r1_tpl + r4_tpl + tpl_i, 
> total_energy): got 1680.36, expected 1638.84)  - 
>     FAILED
> FAILED

> Considering Steffen's answer it might be worth trying the said
> snapshot if this might be fixed there.

Hello Andreas, hello Steffen,

this looks like upstream bug [584][], and appears to be intermittent,
but from what I can tell, no code has been added to address this in
upstream master.  So upgrading to a snapshot would not necessarily help.

However, I can *not* reproduce the AmberFF_test failure.  Neither in my
Debian testing environment, nor in an unstable chroot freshly created
with git-pbuilder.

I literally ran the AmberFF_test one thousand times and it did not fail
once.  Of course, that doesn't mean it won't fail on the 1001st time,
but I have no idea how to effectively approach this without deep
knowledge of the program and its algorithms.

Since we're talking floating-point arithmetic, and the results are not
completely off (same sign, and at least within the same order of
magnitude), this may be a hardware-specific corner-case resulting in
greatly reduced precision.  For reference, I'm running on an Intel Core2
6600.

Trying the same code again and expecting a different result might not be
very professional (or even sane), but we could upload to a buildd
(experimental?) and see if the failure is reproducible there.

Of course, it is always an option to exclude this test to restore
buildability.  The quality would certainly not be worse than before,
when the testsuite was skipped entirely.

[584]: https://github.com/BALL-Project/ball/issues/584

Bottom line:  As much as I'd like to help out; without access to a box
where the problem is reproducible, I am not able to.

Best regards,

        Danny

Attachment: signature.asc
Description: OpenPGP digital signature

Reply via email to