test_linalg_matrix_power fails due to tolerance issues when built against lapack 3.12.0

[badshah400]

Reporting a bug

• I have tried using the latest released version of Numba (most recent is
  visible in the release notes
  (https://numba.readthedocs.io/en/stable/release-notes-overview.html).
• I have included a self contained code sample to reproduce the problem.
  i.e. it's possible to run as 'python bug.py'.

Description

This came up when trying to update packages for lapack on openSUSE to its latest version 3.12.0. I understand numba depends on Lapack indirectly via numpy, and we find that when running the test-suite:

~> python3.11 -m numba.runtests -v -b --exclude-tags=long_running -m 4 -- numba.tests

test_linalg_matrix_power fails with the following log. Previously we were building against lapack 3.9.0, with otherwise identical versions of packages/libraries,¹ and the tests all succeeded.

[ 2481s] FAIL: test_linalg_matrix_power (numba.tests.test_linalg.TestLinalgMatrixPower.test_linalg_matrix_power)
[ 2481s] ----------------------------------------------------------------------
[ 2481s] Traceback (most recent call last):
[ 2481s]   File "/usr/lib64/python3.12/site-packages/numba/tests/test_linalg.py", line 2440, in test_linalg_matrix_power
[ 2481s]     check(a, pwr)
[ 2481s]   File "/usr/lib64/python3.12/site-packages/numba/tests/test_linalg.py", line 2428, in check
[ 2481s]     np.testing.assert_allclose(got, expected, rtol=res, atol=res)
[ 2481s]   File "/usr/lib64/python3.12/site-packages/numpy/testing/_private/utils.py", line 1504, in assert_allclose
[ 2481s]     assert_array_compare(compare, actual, desired, err_msg=str(err_msg),
[ 2481s]   File "/usr/lib64/python3.12/contextlib.py", line 81, in inner
[ 2481s]     return func(*args, **kwds)
[ 2481s]            ^^^^^^^^^^^^^^^^^^^
[ 2481s]   File "/usr/lib64/python3.12/site-packages/numpy/testing/_private/utils.py", line 797, in assert_array_compare
[ 2481s]     raise AssertionError(msg)
[ 2481s] AssertionError: 
[ 2481s] Not equal to tolerance rtol=5e-15, atol=5e-15
[ 2481s] 
[ 2481s] Mismatched elements: 1 / 25 (4%)
[ 2481s] Max absolute difference: 1.12132525e-14
[ 2481s] Max relative difference: 3.46189033e-14
[ 2481s]  x: array([[-0.644914,  0.433068,  0.244191,  0.299553,  0.497169],
[ 2481s]        [ 0.360293, -0.456952,  0.328252, -0.061924,  0.741485],
[ 2481s]        [ 0.241062, -0.141202,  0.180883,  0.9199  , -0.207403],...
[ 2481s]  y: array([[-0.644914,  0.433068,  0.244191,  0.299553,  0.497169],
[ 2481s]        [ 0.360293, -0.456952,  0.328252, -0.061924,  0.741485],
[ 2481s]        [ 0.241062, -0.141202,  0.180883,  0.9199  , -0.207403],...
[ 2481s] 
[ 2481s] Stderr:
[ 2481s] /usr/lib64/python3.12/site-packages/numba/tests/test_linalg.py:399: NumbaPerformanceWarning: np.dot() is faster on contiguous arrays, called on (Array(float64, 2, 'A', False, aligned=True), Array(float64, 2, 'A', False, aligned=True))
[ 2481s]   return np.linalg.matrix_power(a, n)
[ 2481s] /usr/lib64/python3.12/site-packages/numba/tests/test_linalg.py:399: NumbaPerformanceWarning: np.dot() is faster on contiguous arrays, called on (Array(float64, 2, 'C', False, aligned=True), Array(float64, 2, 'A', False, aligned=True))
[ 2481s]   return np.linalg.matrix_power(a, n)

Thanks in advance for any suggestions and/or fixes.

Footnotes

1. Package versions used:

   • GCC 13.2.1
   • Lapack 3.12.0
   • Numba: 0.59.1
   • Numpy 1.26.4
   • Python 3.10 / 3.11 / 3.12
   • Scipy 1.13.0
   ↩

[stuartarchibald]

Thanks for the report, and apologies for the delay in responding. It looks like there's some minor numerical difference the results being checked in the matrix_power tests, the Numba buildfarm has also picked up on it: #9602 (comment). Perhaps subscribe to that issue for updates as it will be fixed as part of the upcoming Numba 0.60.0 release. Thanks again!