IntelPython / dpnp / 41452624322

84%

push

github

Document that `dpnp.take` raises on an empty axis with non-empty indices (#3139)

`dpnp.take` raises `IndexError` for non-empty `indices` along an empty
axis, even when the result is empty. `numpy.take` raises only when the
result would be non-empty, so the two differ for inputs such as a `(0,
0)` array:

```python
a = numpy.empty((0, 0))
numpy.take(a, [0, 1], axis=0, mode="clip").shape  # (2, 0)

ia = dpnp.empty((0, 0))
dpnp.take(ia, [0, 1], axis=0, mode="clip")  # IndexError: cannot take non-empty indices from an empty axis
```

In `wrap` mode, NumPy 2.5.4 and older hang on the same call.

dpnp keeps its behavior on purpose. With an empty axis there is no
element for `wrap` or `clip` to map an index to, and NumPy's rule makes
the outcome depend on another dimension: it raises for a `(0, 3)` array
but not for a `(0, 0)` one. The Array API standard leaves out-of-bounds
indices unspecified, so both behaviors conform.

This PR adds a Notes section to the `dpnp.take` docstring that describes
the difference, and a test that covers it for both modes and both axes.
The test also checks that empty indices give the same empty result as
NumPy. Behavior is unchanged.

1604 of 2994 branches covered (53.57%)

Branch coverage included in aggregate %.

29438 of 34570 relevant lines covered (85.15%)

17674.24 hits per line