v1.0.2: HessTest2 fails on some more exotic architectures

[junghans]

The HessTest2 fails on aarch64, ppc64le, s390x.

Details
build_aarch64.txt
build_ppc64le.txt
build_s390x.txt

i686, x86_64 snd armv7hl works perfectly.
(from https://koji.fedoraproject.org/koji/taskinfo?taskID=52711488)

[junghans]

on openSUSE that test fails on i586 as well:
opensuse_build_i586.txt

[robashaw]

Looking at the logs you've provided (thanks for another detailed report), the issue is purely numerical, in that the derivatives are disagreeing on the order of 10^-5.

This could be absolutely fine, I don't personally have enough experience of how stable Hessians are with respect to the numerics of integrals - in which case, the test could be fixed by changing the numerical tolerance. If this level of discrepancy is a problem, however, then it is an architecture thing that I have absolutely no expertise to deal with.

Actions to resolve:

• I will consult someone who knows more than me and maybe run some frequency calculation stability tests
• I will also add a note to the documentation with regards to this specific test

Given the above, this issue likely won't be resolved for a while.

[junghans]

On Gentoo (https://bugs.gentoo.org/746440) same test fails on amd64 (x86_64) as well.
gentoo_build_amd64.txt
gentoo_test_log.xt.txt

[robashaw]

I have adjusted the precision of this test to now be at a threshold where calculated frequencies are unaffected (PR #23). This should prevent numerical issues on different architectures - please open a new issue if this turns out to not be the case.