IntelPython / dpnp / 39792505835

83%

push

github

Fix `dpnp.repeat` with `uint64` repeats above the `int64` maximum (#3115)

`dpnp.tensor.repeat`, and so `dpnp.repeat`, accepts a `uint64` repeats
array, but it checked the array for negative values before casting it to
`int64`. A `uint64` value is never negative, so values above the `int64`
maximum got past the check:

- with more than one element, the value wrapped to a negative number
only when copied to `int64`, and the result was silently wrong
- with a single element, it was turned into a Python `int` too large for
the output shape, which raised an `OverflowError`

`uint64` repeats are now cast to `int64` first, so such values become
negative and raise the same `ValueError` as any other negative repeat:
```python
import dpnp as np

x = np.arange(3)
np.repeat(x, np.array([1, 2**64 - 1, 1], dtype=np.uint64))  # was array([2]), now ValueError
np.repeat(x, np.array([2**64 - 1], dtype=np.uint64))        # was OverflowError, now ValueError
```

`test_repeat_arg_validation` now covers both cases, and both checks fail
without the fix.

1604 of 2994 branches covered (53.57%)

Branch coverage included in aggregate %.

6 of 6 new or added lines in 1 file covered. (100.0%)

29403 of 34537 relevant lines covered (85.13%)

19634.71 hits per line