lm_atm-bubble regression test fails with numpy 1.24
Closed this issue · 5 comments
First showed up in https://github.com/python-hydro/pyro2/actions/runs/4865084901.
I can reproduce the failure on the current main
branch with python 3.11, numpy 1.24.3, and numba 0.57.0, but it works fine if I downgrade to numpy 1.23.5, leaving everything else the same.
I also modified the test to run for only one step and got similar results. However, when I ran this version under valgrind, it passed fine with numpy 1.24.
Running with limiter = 0
seems to fix it too
should we pin the version to numpy < 1.24?
That sounds like a good idea while we track down what changed in numpy
I managed to track this down to numpy 1.24 now using SVML on Skylake processors for power and arctan: numpy/numpy#23289
Running with NPY_DISABLE_CPU_FEATURES="AVX512_SKX"
gives the exact same output as the benchmark, for both numpy versions.
for now, I'll disable this for the tests. But once 1.24 is more widely used, we should rerun the tests with this on.
We might also want to add a note to the docs