test_perfect_match assertions failing occasionally with conda-forge

[duncanmmacleod]

When building PyCBC for conda-forge, the test suite fails on the test_perfect_match tests in test_matchedfilter.py.

As far as I can tell, the failures only occur on zen2 machines (as reported by conda info under 'archspec'), while standard x86_64 works fine.

Examples failures:

• https://dev.azure.com/conda-forge/feedstock-builds/_build/results?buildId=1429426&view=logs&j=5eaeabf9-98bb-5ad4-ae9b-9c34ad8b134f&t=ef28f095-7a97-583f-1501-aab4442be48f
• https://dev.azure.com/conda-forge/feedstock-builds/_build/results?buildId=1434370&view=logs&j=d818e662-7181-5ce9-b265-50d03742b7af&t=367317f3-9b77-59ac-9e12-e56370c98acd
• https://dev.azure.com/conda-forge/feedstock-builds/_build/results?buildId=1434370&view=logs&j=d49dbcef-bf20-5a8a-a3c0-6b6309067d07&t=10fdae1f-e526-5485-4900-661e8ef7cddc

Pytest output

=================================== FAILURES ===================================
_____________________ TestMatchedFilter.test_perfect_match _____________________
[gw0] linux -- Python 3.11.14 $PREFIX/bin/python
self = <test_matchedfilter.TestMatchedFilter testMethod=test_perfect_match>
    def test_perfect_match(self):
        with self.context:
            o,i = match(self.filt,self.filt)
            self.assertAlmostEqual(1,o,places=4)
            self.assertEqual(0,i)
            o,i = match(self.filtD,self.filtD)
            self.assertAlmostEqual(1,o,places=4)
            self.assertEqual(0,i)
            o,i = match(self.filt,self.filt, subsample_interpolation=True)
            self.assertAlmostEqual(1,o,places=4)
>           self.assertAlmostEqual(0,i,places=1)
E           AssertionError: 0 != np.float64(-0.0714285746216774) within 1 places (np.float64(0.0714285746216774) difference)
test/test_matchedfilter.py:121: AssertionError
_________________ TestMatchedFilter.test_perfect_match_offset __________________
[gw0] linux -- Python 3.11.14 $PREFIX/bin/python
self = <test_matchedfilter.TestMatchedFilter testMethod=test_perfect_match_offset>
    def test_perfect_match_offset(self):
        with self.context:
            o,i = match(self.filt,self.filt_offset)
            self.assertAlmostEqual(1,o,places=4)
            self.assertEqual(4096*32,i)
    
            o,i = match(self.filtD,self.filt_offsetD)
            self.assertAlmostEqual(1,o,places=4)
            self.assertEqual(4096*32,i)
    
            o,i = match(self.filt, self.filt_offset,
                        subsample_interpolation=True)
            self.assertAlmostEqual(1, o, places=4)
>           self.assertAlmostEqual(4096*32, i, places=1)
E           AssertionError: 131072 != np.float64(131071.92857142538) within 1 places (np.float64(0.0714285746216774) difference)
test/test_matchedfilter.py:139: AssertionError
=============================== warnings summary ===============================
test/test_infmodel.py::TestModels::test_brute_pol_phase_marg
test/test_infmodel.py::TestWaveformErrors::test_brute_pol_phase_marg
test/test_infmodel.py::TestWaveformErrors::test_brute_pol_phase_marg
test/test_infmodel.py::TestMarginalizedPolModels::test_gaussian_models
  $PREFIX/lib/python3.11/site-packages/pycbc/inference/models/tools.py:136: DeprecationWarning: use marginalize_vector_samples rather than 'polarization_samples'
    warnings.warn("use marginalize_vector_samples rather "
-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
=========================== short test summary info ============================
FAILED test/test_matchedfilter.py::TestMatchedFilter::test_perfect_match - AssertionError: 0 != np.float64(-0.0714285746216774) within 1 places (np.float64(0.0714285746216774) difference)
FAILED test/test_matchedfilter.py::TestMatchedFilter::test_perfect_match_offset - AssertionError: 131072 != np.float64(131071.92857142538) within 1 places (np.float64(0.0714285746216774) difference)
============ 2 failed, 1488 passed, 4 warnings in 184.42s (0:03:04) ============

[spxiwh]

Interesting! Is there a zen2 machine accessible in IGWN that I could interactively test this on?

[duncanmmacleod]

Interesting! Is there a zen2 machine accessible in IGWN that I could interactively test this on?

I have no idea. I certainly don't have one...

[spxiwh]

It's a particularly bizarre failure ... The subsample interpolation feature that it's failing on isn't really being properly tested here, and it should just return the same as the not-subsample interpolation test, which does seem to work ???