@mattip @seiko2plus Do you know where I should add the tests for this?

Do you know where I should add the tests for this?

maybe within [numpy/core/tests/test_cpu_dispatcher.py] (https://github.com/numpy/numpy/blob/main/numpy/core/tests/test_cpu_dispatcher.py) but it will require loading the module in a separate process.

@seiko2plus Is there a preference in NumPy for warnings vs errors? My question here is partially because of the fact that enabling a feature that doesn't exist is not equivalent to disabling a feature that doesn't exist, and so may require different ways of handling the error--a warning for the latter seems appropriate, but in the case of the former, it may be better to throw an error and exit. Thoughts?

@seiko2plus Is there a preference in NumPy for warnings vs errors? My question here is partially because of the fact that enabling a feature that doesn't exist is not equivalent to disabling a feature that doesn't exist, and so may require different ways of handling the error--a warning for the latter seems appropriate, but in the case of the former, it may be better to throw an error and exit. Thoughts?

For reviewers and interested parties, the current PR's behavior regarding errors/warnings:

@magsol,

As long as the error does not affect the workflow of the following procedures, then choosing to raise runtime-warning would make sense. according to Python's warning-doc RuntimeWarning is reserved for warnings about dubious runtime features. maybe ImportWarning would be a better choice since it's ignored by default.
however, the end-user has the upper hand here, since this behavior can be changed during the runtime through module warning, for example, warnings can be treated as errors via:

import os, warnings
os.environ['NPY_DISABLE_CPU_FEATURES'] = 'XXXX'
warnings.simplefilter("error", RuntimeWarning)
import numpy as np

@seiko2plus That's perfect. Thank you!

Tests are missing. For an example of tests that change the environment variables and then run python in a subprocess, see env manipulation in numpy/core/tests/test_cpu_feature.py

I can't reproduce these errors locally. Even when setting up a machine with --cpu-baseline=native correctly, all the tests pass on my end. Wondering if you may have any insights @mattip @seiko2plus

@Micky774, to reproduce this error, you have to disable the dispatch-able features similar to what the smoke test does, for example:

python setup.py build --cpu-baseline=native --cpu-dispatch=none install --user
python runtests.py -n

or just

python runtests.py --cpu-dispatch=none

In general, you have to escape the dispatch tests if __cpu_dispatch__ is equal to None or if it's an empty list.

@Micky774, please let me know when it's ready for review.

@Micky774, please let me know when it's ready for review.

@seiko2plus Should be good now! Thanks for the help :)

@seiko2plus could you take a look? I think the aarch64 failure is systematic and not connected to this PR.

@seiko2plus Should be ready when you get a chance to review.

@seiko2plus Just a gentle reminder ping. Let me know if you have any other concerns to address for this PR :)

@seiko2plus @seberg Just wanted to ping again regarding this PR.

Coverage thinks tests are missing for some of the corner cases.

rebuild CI

@Micky774 will you be continuing this?

Admittedly this had fallen off my radar, but yes I'll resume work on it over the next few days! Thank you for the reminder :)

@mattip Do you know if the failing TestMatmul.test_vector_matrix_values is related to this PR? And what should be done about the code coverage?

if the failing TestMatmul.test_vector_matrix_values is related to this PR

No. It is one of those failures that happens now and then on Windows for unknown reasons. It is probably compiler related, you can ignore it.

I think this needs to be rebased off main rather than having main merged in for the coverage report to be correct? Otherwise I don't understand why coverage is claiming so much of the code is not tested.

Squashed and rebased off main

Coverage is still claiming the new code is not tested. When I run them locally on my AMD machine, they are skipped since I do not have any features beyond the baseline.

I have verified the coverage report is bogus and the new tests run.

Thanks @Micky774