[BUG] numpy types cannot be used for ufunc arguments

[maxnoe]

Describe the bug

Using numpy's C types for ufunc declarations raises a very long error.

Compiling /home/mnoethe/test/cython/test_numpy_ufunc.pyx because it changed.
[1/1] Cythonizing /home/mnoethe/test/cython/test_numpy_ufunc.pyx

Error compiling Cython file:
------------------------------------------------------------
...
from stdlib.stdint import int64_t

np.import_array()

@cython.ufunc
cdef np.int64_t add_one(np.int64_t x):
     ^
------------------------------------------------------------

test_numpy_ufunc.pyx:9:5: Type 'int64_t' cannot be used as a ufunc argument


...

Code to reproduce the behaviour:

#cython: language_level=3
cimport cython
cimport numpy as np

np.import_array()

@cython.ufunc
cdef np.int64_t add_one(np.int64_t x):
    return x + 1

Expected behaviour

Numpy ufuncs work with numpy data types.

OS

Linux, Ubuntu 22.04

Python version

3.11

Cython version

cython 3.0.10 py311hb755f60_0 conda-forge

Additional context

No response

[da-woods]

Thanks for the report. It's good to know what people are trying this and finding the obvious deficiencies :-)

It's slightly harder than it looks. Cython doesn't actually know a lot about integer typedefs - mostly just that it's "some integer" and whether it's signed or not. However, it needs to tell Numpy exactly what type is actually being used. So it either needs some special-casing for these Numpy typedefs, or it needs a clever runtime mechanism to work out what to tell Numpy.