__array_ufunc__(np.matmul, ...) fails

[crusaderky]

These work:

• sparse.COO @ np.ndarray
• sparse.matmul(sparse.COO, np.ndarray)
• sparse.matmul(np.ndarray, sparse.COO)

These crash:

• np.ndarray @ sparse.COO
• np.matmul(sparse.COO, np.ndarray)
• np.matmul(np.ndarray, sparse.COO)

Using:

• Python 3.7
• numpy 1.16
• sparse 0.7.0
• NUMPY_EXPERIMENTAL_ARRAY_FUNCTION=1

>>> import numpy
>>> import sparse
>>> a = numpy.zeros((2, 2))
>>> sa = sparse.COO(a)
>>> sa @ a
array([[0., 0.],
       [0., 0.]])
>>> a @ sa
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-9-85bcf6384716> in <module>
----> 1 a @ sa

/nfs/crusaderky/prg/github/sparse/sparse/coo/core.py in __array_ufunc__(self, ufunc, method, *inputs, **kwargs)
   1608 
   1609         if method == '__call__':
-> 1610             result = elemwise(ufunc, *inputs, **kwargs)
   1611         elif method == 'reduce':
   1612             result = COO._reduce(ufunc, *inputs, **kwargs)

/nfs/crusaderky/prg/github/sparse/sparse/coo/umath.py in elemwise(func, *args, **kwargs)
     46     """
     47 
---> 48     return _Elemwise(func, *args, **kwargs).get_result()
     49 
     50 

/nfs/crusaderky/prg/github/sparse/sparse/coo/umath.py in __init__(self, func, *args, **kwargs)
    427         self.cache = {}
    428 
--> 429         self._get_fill_value()
    430         self._check_broadcast()
    431 

/nfs/crusaderky/prg/github/sparse/sparse/coo/umath.py in _get_fill_value(self)
    476         # Some elemwise functions require a dtype argument, some abhorr it.
    477         try:
--> 478             fill_value_array = self.func(*zero_args, dtype=self.dtype, **self.kwargs)
    479             self.dtype = None
    480         except TypeError:

ValueError: matmul: Input operand 1 does not have enough dimensions (has 0, gufunc core with signature (n?,k),(k,m?)->(n?,m?) requires 1)

>>> numpy.matmul(a, sa)
(same ValueError as above)

[hameerabbasi]

For matmul, we'd need to special-case anything that's a gufunc but not a regular ufunc in __array_ufunc__. Are you willing to submit a patch for this?

[hameerabbasi]

(We'd need to route it through __array_function__).