cupy.ndarray cannot be compared with ellipsis

[leofang]

Not sure if it's due to #4198. The __richcmp__ method does not handle trivial comparison with Python objects correctly:

>>> np.empty(3) == ...
array([False, False, False])
>>>
>>> cp.empty(3) == ...
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "cupy/_core/core.pyx", line 1094, in cupy._core.core.ndarray.__richcmp__
    return numpy.equal(self, other)
TypeError: operand type(s) all returned NotImplemented from __array_ufunc__(<ufunc 'equal'>, '__call__', array([0., 0., 0.]), Ellipsis): 'ndarray', 'ellipsis'

It seems NumPy internally calls PyObject_RichCompare to handle this. Perhaps before delegating to numpy.equal() (and friends) we should handle this edge case first?

[leofang]

Note: this relates to the test_indexing.py failure seen in #5698 (comment).

[toslunar]

#1989 discusses comparison with None.

[asmeurer]

Can you post the traceback from the test failure? We should probably make the array API tests not use == for comparing arrays and use is to compare against Ellipsis.

[github-actions]

cc/ @asmeurer @leofang

[takagi]

This line brings the error, where key is a tuple that contains array indices and they are compared to an ellipsis. When an index is a ndarray, the error above occurs. I think we should count using is here.

cupy/cupy/array_api/_array_object.py

Line 330 in 0955ad9

n_ellipsis = key.count(...)

[toslunar]

n_ellipsis = key.count(...)

To be precise, each array index in key must have _.size <= 1, or key.count(...) will raise ValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all(). The line works fine because indexing with integer arrays is unspecified in Array API. https://data-apis.org/array-api/latest/API_specification/indexing.html

cupy/cupy/array_api/_array_object.py

Lines 346 to 351 in 0955ad9

	elif isinstance(key, Array):
	if key.dtype in _integer_dtypes:
	if key.ndim != 0:
	raise IndexError(
	"Non-zero dimensional integer array indices are not allowed in the array API namespace"
	)

[asmeurer]

Do you know if this code leads to a problem in the NumPY implementation? I think something like this would be more robust.

diff --git a/numpy/array_api/_array_object.py b/numpy/array_api/_array_object.py
index ead061882d..e7ddcf033b 100644
--- a/numpy/array_api/_array_object.py
+++ b/numpy/array_api/_array_object.py
@@ -325,10 +325,13 @@ def _validate_index(key, shape):

             if shape is None:
                 return key
-            n_ellipsis = key.count(...)
+            # Use 'is' to compare against ellipsis to avoid performing == on
+            # an array.
+            ellipses = [i for i, k in enumerate(key) if k is ...]
+            n_ellipsis = len(ellipses)
             if n_ellipsis > 1:
                 return key
-            ellipsis_i = key.index(...) if n_ellipsis else len(key)
+            ellipsis_i = ellipses[0] if ellipses else len(key)

             for idx, size in list(zip(key[:ellipsis_i], shape)) + list(
                 zip(key[:ellipsis_i:-1], shape[:ellipsis_i:-1])

[takagi]

As @toslunar mentioned in #5715 (comment), I think NumPy's current code works without any problem. But, yes, your diff would be exactly what we want to do there.