Testsuite failures on several architectures

As part of the automatic building of bornagain on several architectures by the Debian project, several testsuite failures have been discovered.

As of 2024.11.25, the current Git repository fails its testsuite on arm64, i386, ppc64el, riscv64 and s390x. The same test (Example.specular.GaussianBeams.persist) fails on all of these architecture, but three others (Example.scatter2d.DodecahedraSAS.persist, Suite.Persist.DepthprobeAveragedLayer and Suite.Persist.DepthprobeSimpleLayer) also fail in addition to the first one on i386. amd64, armel and armhf are fine.

The "Build-Attempted" links at https://buildd.debian.org/status/package.php?p=bornagain point at the full logs.

(This looks similar to #952 (closed), with a different set of failing tests)

@lolando I looked specifically into the log [1] from arm64, where the persistence test fails for the Python script GaussianBeams.py. To investigate further, we would need to see the files mentioned in these log lines:

359: To overwrite references:
359: cp -f /build/reproducible-path/bornagain-22~git20241218175952.966c34a+ds3/obj-aarch64-linux-gnu/test_output/MiniExamples/specular/GaussianBeams.0.int /build/reproducible-path/bornagain-22~git20241218175952.966c34a+ds3/Tests/ReferenceData/MiniExamples/specular/GaussianBeams.0.int
359: cp -f /build/reproducible-path/bornagain-22~git20241218175952.966c34a+ds3/obj-aarch64-linux-gnu/test_output/MiniExamples/specular/GaussianBeams.1.int /build/reproducible-path/bornagain-22~git20241218175952.966c34a+ds3/Tests/ReferenceData/MiniExamples/specular/GaussianBeams.1.int
359: cp -f /build/reproducible-path/bornagain-22~git20241218175952.966c34a+ds3/obj-aarch64-linux-gnu/test_output/MiniExamples/specular/GaussianBeams.2.int /build/reproducible-path/bornagain-22~git20241218175952.966c34a+ds3/Tests/ReferenceData/MiniExamples/specular/GaussianBeams.2.int
359: cp -f /build/reproducible-path/bornagain-22~git20241218175952.966c34a+ds3/obj-aarch64-linux-gnu/test_output/MiniExamples/specular/GaussianBeams.3.int /build/reproducible-path/bornagain-22~git20241218175952.966c34a+ds3/Tests/ReferenceData/MiniExamples/specular/GaussianBeams.3.int

Furthermore, to resolve issues like these, we would need a way to trigger builds on the respective target platforms, so that we can print debug messages and thereby trace divergent outcomes to specific code lines.

[1] https://buildd.debian.org/status/fetch.php?pkg=bornagain&arch=arm64&ver=22~git20241218175952.966c34a%2Bds3-1&stamp=1735932020&raw=0

added OS:Linux bug labels

I've uploaded a new version of the package, based on the current state of your master branch, with an extra patch that displays the output files when they differ from the reference files. I don't expect the build to succeed due to unrelated problems (lmfit-py vs. numpy2), but I think the testsuite should at least reach the point where the expected/obtained files differ. Hope this helps :-)

A new snapshot as of March 12 still fails on at least i386, arm64 and ppc64el.

changed milestone to %v22

assigned to @j.wuttke

Failures

 Store simulation as GUI/XML project file
 unknown file: Failure
 C++ exception with description "BUG: Assertion axis->isEquiScan() failed in ./GUI/ Model/Beam/SourceItems.cpp, line 219.

are due to different arithmetic outcomes for

• factory functions EquiScan or EquiDivision called when creating XML file on amd64 (memory model [1] LP64)
• functions Scale::isEquiScan or Scale::isEquiDivision called when reading the XML file on i386 (memory model ILP32).

I replaced size_t by uint32_t in the above cited four functions. I expected some tests to fail, but this did not happen. Therefore I stashed this change, as it is unlikely to resolve the problem on i386.

Next, we should replace the exact replication of factory computations by standard floating-point tests allowing for some tolerance.

[1] https://en.cppreference.com/w/cpp/language/types

mentioned in merge request !2922 (merged)

mentioned in commit bb4164ed

@lolando I relaxed the numerical test that caused most (if not all) errors in the i386 build. How to trigger a new build?