ENH: Add `__array_ufunc__` #8247

This reverts commit bac094c and enables __numpy_ufunc__ for development in the NumPy 1.13.0 development cycle.

The first commit in changing __numpy_ufunc__ by removing the index argument and making the out argument value always a tuple. These changes were proposed in numpygh-5986 and have been accepted. Renaming before further changes avoids triggering tests in scipy and astropy while keeping the numpy tests working.

Previously when __array_ufunc__ for one of the ufunc arguments was called, that arguments position was passed in the call. This PR removes that argument as proposed in numpygh-5986.

This is an ugly kluge, but until we merge multiarray.so and umath.so moving stuff into private/*.h serves as a reasonable workaround.

For some reason, np.generic.__pow__ felt the need to reimplement Python's full binop dispatch logic. The code has been there since the beginning of (git) history, isn't required to pass any tests, and AFAICT doesn't actually do anything interesting. I suspect it was just split apart from the (much more trivial) implementation of all the other np.generic binop methods because nb_power has a slightly different signature than all the others (it takes an extra argument, which we ignore, but still has to be accepted and passed as appropriate), and then got forgotten when the other implementations were simplified long ago. This commit throws away all that code and makes gentype_power a 2-liner, just like all the other gentype_@name@ methods.

As per the discussion at numpygh-5844, and in particular numpy#5844 (comment) this commit switches binop dispatch to mostly defer to ufuncs, except in some specific cases elaborated in a long comment in number.c. The basic strategy is to define a single piece of C code that knows how to handle forward binop overrides, and we put it into private/binop_override.h so that it can be accessed by both the array code in multiarray.so and the scalar code in umath.so.

In previous versions, one could force ndarray binops to defer by setting a high __array_priority__. With __array_ufunc__ this gets ignored, and this commit ensures it is still possible to avoid using the standard python language feature that setting something to None means it is not implemented. In consequence, inside a ufunc, if __array_ufunc__ is None, it will be treated as if it had returned NotImplemented (leading to a TypeError if no other object had a functioning __array_ufunc__ override).

This includes the use of super everywhere, and in the brief description of __array_ufunc__ in the reference section.

As part of this, split off the checking for overrides from the actual trying to execute them (in ufunc_override.c), so that just the check can be done in ndarray.__array_ufunc__.

Bring into compliance with current ndarray.__array_ufunc__ implementation and type casting hierarchy.

As part of this, it turns out the number of output arguments nout was needed, so that was passed in. Also, testing showed that arguments passed in both as positional and keyword arguments were not caught, so a duplication check was added.

@njsmith

The text had been changed following the suggestion of @njsmith, but somehow I had forgotten to delete the original.

In particular, use fact that we're guaranteed to have a PyUFuncObject in PyUFunc_CheckOverride.

And put it in src/umath instead of src/private. This leaves only the PyUFunc_WithOverride function in ufunc_override.c, which is all that is needed outside of umath (it is called by array_ufunc in multiarray/methods.c).

[ci skip]

Code inadvertantly removed in rebasing old PR on new work.

Also add a few more tests of the same example for good measure.

This mixin class provides an easy way to implement arithmetic operators that defer to __array_ufunc__ like numpy.ndarray in non-ndarray subclasses.

* ENH: Better error message for __array_ufunc__ not implemented New behavior: >>> import numpy as np >>> class Dummy: ... def __array_ufunc__(self, *args, **kwargs): ... return NotImplemented >>> np.negative(Dummy()) TypeError: operand type(s) do not implement __array_ufunc__( <ufunc 'negative'>, '__call__', <__main__.Dummy object at 0x1106df8d0>): 'Dummy' * check for null errmsg_formatter

* ENH: NDArrayOperatorsMixin calls ufuncs directly, like ndarray Per our discussion in numpy#8247 (comment) * add back in accidentally dropped __repr__

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

ENH: Add `__array_ufunc__` #8247

ENH: Add `__array_ufunc__` #8247

Commits on Apr 27, 2017

ENH: Add __array_ufunc__ #8247

ENH: Add __array_ufunc__ #8247

Commits on Apr 27, 2017

ENH: Add `__array_ufunc__` #8247

ENH: Add `__array_ufunc__` #8247