Avoid in-place multiplication of a large value to an array with small integer dtype

[Illviljan]

Upstream numpy has become a bit more particular with which types you can use for inplace operations. This PR fixes

_______________ TestImshow.test_imshow_rgb_values_in_valid_range _______________
self = <xarray.tests.test_plot.TestImshow object at 0x7f88320c2780>
def test_imshow_rgb_values_in_valid_range(self) -> None:
da = DataArray(np.arange(75, dtype="uint8").reshape((5, 5, 3)))
_, ax = plt.subplots()
> out = da.plot.imshow(ax=ax).get_array()
/home/runner/work/xarray/xarray/xarray/tests/test_plot.py:2034: _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ /home/runner/work/xarray/xarray/xarray/plot/accessor.py:421: in imshow
return dataarray_plot.imshow(self._da, *args, **kwargs)
/home/runner/work/xarray/xarray/xarray/plot/dataarray_plot.py:1601: in newplotfunc
primitive = plotfunc(
/home/runner/work/xarray/xarray/xarray/plot/dataarray_plot.py:1853: in imshow
alpha *= 255
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ self = masked_array(
data=[[[1],
[1],
[1],
[1],
[1]],
[[1],
[1],
...,
[1],
[1],
[1],
[1]]],
mask=False,
fill_value=np.int64(999999),
dtype=uint8)
other = 255
def __imul__(self, other):
"""
Multiply self by other in-place.
"""
m = getmask(other)
if self._mask is nomask:
if m is not nomask and m.any():
self._mask = make_mask_none(self.shape, self.dtype)
self._mask += m
elif m is not nomask:
self._mask += m
other_data = getdata(other)
other_data = np.where(self._mask, other_data.dtype.type(1), other_data)
> self._data.__imul__(other_data)
E numpy._core._exceptions._UFuncOutputCastingError: Cannot cast ufunc 'multiply' output from dtype('int64') to dtype('uint8') with casting rule 'same_kind'
/home/runner/micromamba/envs/xarray-tests/lib/python3.12/site-packages/numpy/ma/core.py:4415: UFuncTypeError

Some curious behaviors seen while debugging:

alpha=np.array([1], dtype=np.int8)
alpha*=255repr(alpha) # 'array([-1], dtype=int8)'alpha=np.array([1], dtype=np.int16)
alpha*=255repr(alpha) # 'array([255], dtype=int16)'

xref: #8844

[dopplershift]

Why is in-place multiplication a problem here?

[Illviljan]

Quick response for a draft PR! My idea 5 minutes ago is that

alpha = np.array(1, dtype=np.int8) alpha *= 255

will create a number larger than int8, which is strange. So I think we should create a new array instead with whatever numpy thinks is the best dtype.

[Illviljan]

The dtype should allow at least 0 and 255. But it will be converted in np.ma.concatenate at some point anyway so I thought it's best to just figure it out early and initialize alpha correctly.

[dcherian]

Thanks for figuring this out!