`linalg.cholesky` not tested for complex dtypes

[lucascolley]

As reported in data-apis/array-api-compat#61 (comment), linalg.cholesky is only tested on real arrays, hence some bugs slipped through into both array-api-compat and numpy.array_api.

[lucascolley]

relevant lines:

array-api-tests/array_api_tests/test_linalg.py

Lines 126 to 130 in d295a0a

	@given(
	x=positive_definite_matrices(),
	kw=kwargs(upper=booleans())
	)
	def test_cholesky(x, kw):

array-api-tests/array_api_tests/hypothesis_helpers.py

Line 314 in d295a0a

def positive_definite_matrices(draw, dtypes=floating_dtypes):

[ev-br]

Are complex dtypes really skipped for new enough spec levels?

floating_dtypes are defined as floating_dtypes = sampled_from(dh.all_float_dtypes) at
https://github.com/data-apis/array-api-tests/blob/master/array_api_tests/hypothesis_helpers.py#L187

and all_float_dtypes include complex ones if the standard is above 2022.12:
https://github.com/data-apis/array-api-tests/blob/master/array_api_tests/hypothesis_helpers.py#L195

This me is eyeballing the code though, not sure how to check it.

[ev-br]

This me is eyeballing the code though, not sure how to check it.

Actually,

$ git diff
diff --git a/array_api_tests/test_linalg.py b/array_api_tests/test_linalg.py
index 40fe035..5f39da0 100644
--- a/array_api_tests/test_linalg.py
+++ b/array_api_tests/test_linalg.py
@@ -130,6 +130,8 @@ def _test_namedtuple(res, fields, func_name):
 def test_cholesky(x, kw):
     res = linalg.cholesky(x, **kw)
 
+    print(f"{x.dtype = }")
+
     ph.assert_dtype("cholesky", in_dtype=x.dtype, out_dtype=res.dtype)
     ph.assert_result_shape("cholesky", in_shapes=[x.shape],
                            out_shape=res.shape, expected=x.shape)

prints (on several reruns)

$ pytest array_api_tests/test_linalg.py::test_cholesky -v -s
============================================ test session starts ============================================
platform linux -- Python 3.11.0, pytest-8.3.3, pluggy-1.5.0 -- /home/br/miniforge3/envs/array-api-tests/bin/python3.11
cachedir: .pytest_cache
hypothesis profile 'array-api-tests' -> deadline=timedelta(milliseconds=800), database=DirectoryBasedExampleDatabase(PosixPath('/home/br/repos/array-api-tests/.hypothesis/examples'))
metadata: {'Python': '3.11.0', 'Platform': 'Linux-5.15.0-76-generic-x86_64-with-glibc2.31', 'Packages': {'pytest': '8.3.3', 'pluggy': '1.5.0'}, 'Plugins': {'json-report': '1.5.0', 'hypothesis': '6.118.8', 'metadata': '3.1.1'}, 'array_api_tests_module': 'array_api_tests._array_module', 'array_api_tests_version': '2022.09.30+406.g9db56ec.dirty'}
Array API Tests Module: numpy (2.1.3). API Version: 2023.12. Enabled Extensions: fft, linalg
rootdir: /home/br/repos/array-api-tests
configfile: pytest.ini
plugins: json-report-1.5.0, hypothesis-6.118.8, metadata-3.1.1
collected 1 item                                                                                            

array_api_tests/test_linalg.py::test_cholesky x.dtype = dtype('float32')
x.dtype = dtype('float32')
x.dtype = dtype('float32')
x.dtype = dtype('float32')
x.dtype = dtype('complex128')
x.dtype = dtype('float32')
x.dtype = dtype('float64')
x.dtype = dtype('float32')
x.dtype = dtype('complex64')
x.dtype = dtype('float32')
PASSED

[lucascolley]

okay cool, so it seems like Ralf's initial diagnosis was incorrect. I don't know why the linked bug was able to slip through, then.

[asmeurer]

I think this is something that I fixed a few months ago.