Tests currently fail on M1, but not on Intel MacOS

[paddyroddy]

I have managed to get this to build on M1 (see here flimlib/flimlib#60 (comment)) but it only works if I force the tests to be skipped i.e. mvn clean install -DskipTests

The error message is here:

From googling around it looks like Failed to load class "org.slf4j.impl.StaticLoggerBinder" could be fixed by e.g. https://stackoverflow.com/a/9919375/7359333

My colleague @drmattews doesn't need to force the tests to fail on his Intel Mac (it appears to work out of the box) so the issue seems to be somewhat M1/Apple Silicon related

[imagesc-bot]

This issue has been mentioned on Image.sc Forum. There might be relevant details there:

https://forum.image.sc/t/cmake-fails-to-build-on-m1-mac-apple-silicon/61964/8

[hinerm]

As of flimlib 2.1.1 I can confirm tests are still failing on arm64 macs. It looks like the numerical computation differs (slight rounding in most cases). My current failures are:

[ERROR] Failures: 
[ERROR]   FitTest.testBayesFitImg:151->assertSampleEquals:222 arrays first differed at element [2]; expected:<0.20058449> but was:<1164.4563>
[ERROR]   FitTest.testBinning:125->assertSampleEquals:222 arrays first differed at element [0]; expected:<15.917448> but was:<15.917488>
[ERROR]   FitTest.testGlobalFitImg:188->assertSampleEquals:222 arrays first differed at element [3]; expected:<4.498526> but was:<4.498537>
[ERROR]   FitTest.testRLDFitImg:115->assertSampleEquals:222 arrays first differed at element [3]; expected:<4.498526> but was:<4.498537>

[hinerm]

@paulbarber just wanted to bring you into this conversation. I assume this is an issue in core flimlib but it's obvious/easy to reproduce here in flimj-ops. We have not run these tests on a non-arm64 Mac yet, but they do pass on Windows.

[ctrueden]

As a next step, I'd like to see whether the tests fail also on Intel Macs, or only Apple Silicon machines. Either way though, we should try to minify this example to something that can be reproduced in flimlib alone.

[ctrueden]

I finally got around to testing on both macOS x64 and macOS arm64. Confirming: on x64, all tests pass, whereas on arm64, four tests fail exactly as @hinerm describes above. Next steps:

• Run one of the failing tests through a debugger to see exactly which flimlib calculations are happening, in parallel on both x64 and arm64, noting as soon as a difference pops up.
• Create a minimal example exposing the issue we can add to the flimlib library level.
• Fix the flimliib code to behave identically if possible, or at least document the nature of the difference if not.