BUG: Incorrect type list for some ufunc loops (sin, cos, arctan2)

[seberg]

This is a follow up from gh-22376, where I forgot the sin/cos loops. CuPy noticed these in cupy/cupy#7243 and also found that arctan2 has the same issue.

For arctan2 this is actually a bug also leads to wrong results within NumPy (rather than only causing numba/cupy to be confused).

Just opening the issue in case 1.24.2 is around the corner, should fix it today in either case.

[kmaehashi]

Thanks for opening an issue @seberg! It seems sin and cos are also affected when int8 is involved because np.can_cast(np.int8, np.float16) is True:

NumPy 1.23:

>>> np.sin.types
['e->e', 'f->f', 'd->d', 'f->f', 'd->d', 'g->g', 'F->F', 'D->D', 'G->G', 'O->O']
>>> np.sin(np.array([0, 1, 2], dtype=np.int8))
array([0.    , 0.8413, 0.909 ], dtype=float16)
>>> np.cos(np.array([0, 1, 2], dtype=np.int8))
array([ 1.    ,  0.5405, -0.4163], dtype=float16)

NumPy 1.24:

>>> np.sin.types
['f->f', 'e->e', 'd->d', 'f->f', 'd->d', 'g->g', 'F->F', 'D->D', 'G->G', 'O->O']
>>> np.sin(np.array([0, 1, 2], dtype=np.int8))
array([0.       , 0.841471 , 0.9092974], dtype=float32)
>>> np.cos(np.array([0, 1, 2], dtype=np.int8))
array([ 1.       ,  0.5403023, -0.4161468], dtype=float32)

[seberg]

Ah, good point, although I really don't like how we guess at float16 for integer inputs...