TEST: 1.24.x + blas variants

[h-vetinari]

Continuing the analysis from #273, #252, #237, #227 & #196. Should not be merged for the same reasons as #227.

[conda-forge-webservices]

Hi! This is the friendly automated conda-forge-linting service.

I just wanted to let you know that I linted all conda-recipes in your PR (recipe) and found it was in an excellent condition.

[h-vetinari]

Update for numpy 1.24.1

After being all green for 1.23.5, we're still all green for 1.24.1 🥳 (with some small caveats below)

Notable

• It looks like the new QEMU 7.2.0 might have finally solved our emulation problems with PPC (see Switch PPC to cross-compilation & testing in emulation #274), and thus the flakiness of Travis. 🤩 🥳 🚀
  Xref also this issue.
• I'm planning to get Switch PPC to cross-compilation & testing in emulation #274 merged (which should also clean up some test skips), and then redo this here.
• Since we're not able to run tests for osx-arm, we might miss issues such as linalg.eigh incorrect results for complex data on osx-arm64 #287 (though this one looks more like a channel-mixing problem)

Restrictions

• ppc64le and pypy-on-aarch only run tests with label='fast', not the full test suite
• the following test skips:

# https://github.com/numpy/numpy/issues/15243
{% set tests_to_skip = tests_to_skip + " or test_loss_of_precision[complex256]" %}  # [ppc64le]
# Sometimes crashes on travis (needs too much memory); since 1.22.4 it fails for pypy
{% set tests_to_skip = tests_to_skip + " or test_large_archive" %}                  # [(CI == "travis") or ((python_impl == "pypy") and linux)]
# there are some tests that cannot really work in emulation, see e.g. numpy/numpy#20445
{% set tests_to_skip = tests_to_skip + " or Test_ARM_Features" %}                   # [build_platform != target_platform]
# test_new_policy reruns part of test suite; including a CPU feature test that fails in emulation
{% set tests_to_skip = tests_to_skip + " or test_new_policy" %}                     # [build_platform != target_platform]
# flaky refcount-based test; already skipped upstream for win+py39
{% set tests_to_skip = tests_to_skip + " or test_partial_iteration_cleanup" %}      # [osx]
# non-critical polynomial repr divergence, see https://github.com/numpy/numpy/issues/22825
{% set tests_to_skip = tests_to_skip + " or (TestPrintOptions and (test_str or test_switch_to_exp or test_non_finite))" %}  # [osx and python_impl == "pypy"]

Details

variant	before	after
linux + ppc (on azure)	test failures due to emulation problems	looks like #274 is now working 🥳
linux + ppc (on travis)	fine (with 2 test skips)	Travis flakiness causing all builds to lose their agent 😑

---

lib	before	after	updated version	updated build
numpy	1.23.4	1.24.1	X
libblas	3.9.0-16	3.9.0-16
blis	0.9.0-0	0.9.0-0
openblas	0.3.21-pthreads-3	0.3.21-pthreads-3
mkl	2022.1.0-915	2022.1.0-915
netlib	3.9.0-5	3.9.0-5
pypy	7.3.9-7	7.3.9-7
qemu-user-static	7.2.0-1	7.2.0-1

---

variant	accelerate	blis	mkl	netlib	openblas	sum*
linux / x86	➖	✔️	✔️	✔️	✔️	-
linux / aarch	➖	➖	➖	✔️	✔️	-
linux / ppc64le	➖	➖	➖	✔️	✔️	-
osx / arm	✔️	➖	➖	✔️	✔️	-
osx / x86	✔️	✔️	✔️	✔️	✔️	-
win / x86	➖	✔️	✔️	✔️	✔️	-
sum*	-	-	-	-	-	-

* sum of Failures (out of a total of 114 CI combinations being tested)

Build logs:
Azure
Travis

[h-vetinari]

Update for numpy 1.24.3

After being all green for 1.24.1, we now have 2 failures out of 114 jobs for 1.24.3.

Notable

• The recent QEMU 7.2.0 finally solved our emulation problems with PPC (see Switch PPC to cross-compilation & testing in emulation #274), and thus the flakiness of Travis. 🤩 🥳 🚀
• The pypy failures don't seem to have a pattern, could be random OOM, but only appears in quite specific combination (win+openblas+pypy), so perhaps not that random.

Restrictions

• pypy-on-aarch/ppc64le only runs tests with label='fast', not the full test suite

Details

variant	before	after
linux + ppc	test failures due to emulation problems	solved by #274 🥳
win + openblas + pypy	-	two separate test crashes

---

lib	before	after	updated version	updated build
numpy	1.24.1	1.24.3	X
libblas	3.9.0-16	3.9.0-16
blis	0.9.0-0	0.9.0-0
openblas	0.3.21-pthreads-3	0.3.21-pthreads-3
mkl	2022.1.0-915	2022.2.1-16997	X
netlib	3.9.0-5	3.9.0-5
pypy	7.3.9-7	7.3.11-0	X
qemu-user-static	7.2.0-1	7.2.0-1

---

variant	accelerate	blis	mkl	netlib	openblas	sum*
linux / x86	➖	✔️	✔️	✔️	✔️	-
linux / aarch	➖	➖	➖	✔️	✔️	-
linux / ppc64le	➖	➖	➖	✔️	✔️	-
osx / arm	✔️	➖	➖	✔️	✔️	-
osx / x86	✔️	✔️	✔️	✔️	✔️	-
win / x86	➖	✔️	✔️	✔️	✔️ (CPython) ❌ (PyPy)	2F
sum*	-	-	-	-	2F	2F

* sum of Failures (out of a total of 114 CI combinations being tested)

Build logs:
Azure

win + openblas + pypy3.8: 1 failure

================================== FAILURES ===================================
___________________________ tests/test_warnings.py ____________________________
[gw1] win32 -- Python 3.8.16 %PREFIX%\python.exe
worker 'gw1' crashed while running 'tests/test_warnings.py::test_warning_calls'

win + openblas + pypy3.9: 1 failure

================================== FAILURES ===================================
____________________ matrixlib/tests/test_matrix_linalg.py ____________________
[gw1] win32 -- Python 3.9.16 %PREFIX%\python.exe
worker 'gw1' crashed while running 'matrixlib/tests/test_matrix_linalg.py::TestQRMatrix::test_stacked_inputs[float32-outer_size0-size1]'

[h-vetinari]

Closing this now that we have 1.26