v1.23.0: np.unique returns incorrect result for float16 array containing NaNs

This only seems to affect float16 arrays (float32 and float64 are fine), and only if the array contains NaN values.

Reproduction:

import numpy as np
x = np.array([0, 1, np.nan], dtype='float16')
print(np.__version__)
print(x)
print(np.unique(x))

Output in numpy 1.22.4:

1.22.4
[ 0.  1. nan]
[ 0.  1. nan]

Output in numpy 1.23.0

1.23.0
[ 0.  1. nan]
[0.]

[postmalloc]

@seberg I added some test cases specifically for float16 and float32; includes the test case you shared earlier.

The underlying issue seems that np.searchsorted is broken with NaN for float16, I am not quite sure why… Maybe it happened as part of the C++ conversions?\

EDIT:

x = np.array([0, 1, np.nan], dtype='float16')
np.searchsorted(x, x[-1])
# Should return 2 but returns 0

[postmalloc]

I’ve taken a quick look. It seems the type_num for np.float16 is 23 according to the NPY_TYPES enum sequence. However, the taglist in binarysearch has npy::half_tag at a different position (at index 11). The indices don’t match, and it never enters HALF_LT as you observed.