Seems not quite that simple. The tests fail. It seems that this change as is breaks compatibility for array subclasses (i.e. matrix and masked arrays)? I don't have time to look into it, my best guess right now is that the reflected op, i.e.__radd__ also callsnp.add and then fails again, because whatever mechanism made sure that doesn't happen is somehow not active? Unfortunatly we have to be careful about what might be changed, there are probably no tests defining the current behaviour well.

Ok, second attempt. This stays completely out of theufunc machinery and only applies to binary operators (as it should). Also see the lengthy comment in the code. Tests pass for me (I only rannumpy.test() the first time around)

Can you explain to me why the array is converted currently? As far as I
understand the ufunc machinery expects that such objects as yours got
converted to an 0-d object array (or such). Where and why does this
happen? If we cannot fix it, can we hook into there to make your type
become a 0-d object array instead of a doing a full array conversion
process? Just thinking, but hooking your fix in further up doesn't
really sound right to me right now, though I will not look into it
deeply now.

First, to answer your question:get_ufunc_arguments is called from these two sites:

• https://github.com/numpy/numpy/blob/master/numpy/core/src/umath/ufunc_object.c#L1787
• https://github.com/numpy/numpy/blob/master/numpy/core/src/umath/ufunc_object.c#L2311

That routine callsPyArray_FromAny on the ufunc arguments, which convertsMyThing into a 3x3 (IOW: large) object array of "scalar"MyThings. That's the core of the issue, because it is O(N) in the number of array elements. Nothing produces a 0-D array, and I'm not sure that would be correct or would fix anything.

Second, in my opinion, having__array_priority__ logic inumath is just broken. It should be ripped out. A tell-tale sign of this is that the logic is activated only for two-argument ufuncs with one output. What's so special about them? Why should they behave differently from three-argument ufuncs?

Another thing that's broken in the current code and would have to be fixed along with the issue above is that lots of places (such as1,2) usePyArray_Type->tp_as_number->nb_multiply to call themultiply ufunc (etc.). Because of__array_priority__, the semantics ofnb_multiply are no longer the same as the ufunc. Specifically,nb_multiply is required to to returnNotImplemented in cases of priority inversion, whereas the ufunc has a job to do and just needs to do it, without asking questions about priorities.

My patch above is a valid band aid for#3375. It does not claim to fix the issues I've mentioned. I do claim that it doesn't break anything that was previously working. Moreover, having priority logic "further up" (as you called it) is actually less risky, because itonly touches dispatch for binary operators. In other words,fewer code paths go through there than through the ufunc.

The issues I mentioned above should probably be filed as bugs of their own.

Good points, it is true that it doesn't even make sense that ufuncs can possibly return NotImplemented, etc. I agree that this whole thing isn't ideal and likely it is the only reasonable thing to fix it where you fixed it. It is basically only because of how python handles subclassing with operators (i.e. Python knows that the subclass should be asked to handle the operator) that this code works at all. To be honest, I am not quite certain that array priority is necessary at all :).

I think the whole thing needs more thinking. I would digg a bit into checking other options before deciding that this perfect for myself, but I don't have the time right now. This creates one change I think, that I don't really mind too much, but should maybe be deprecated? If YourThing does not implement the__rop__ it would seem that things break?

Array priority itself is a reasonable idea--it makes my example possible at all. Without it,numpy would not know that it's not supposed to try and convertMyThing into an array. (which is possible, but a terrible idea)

If some object declares an__array_priority__, but does not implement__rop__, an error will result as you point out. To me, that's fine. I would argue that that is correct behavior on the part ofnumpy, and a mistake on the part of the priority-declaring object, who claimed that it wanted to handle reverse arithmetic but then didn't. Onecould argue thatnumpy should check whether the corresponding__rop__ is implemented before returningNotImplemented, but no other part of Python works like that, and it inserts more expensive (and possibly brittle) checks into what is a performance-relevant code path. That's why I'm hesitant.

@seberg Are you happy with this?

It seems like it is the right spot for such a check. I am still a little worried that breaking the case of people defining array priority but not supporting all the operators actually might affect code out there. I don't mind that as such, so maybe just try and see... What I am thinking of is mostly objects defining__array__ and for some reason also__array_priority__, i.e. objects in the general direction of PIL images or so? Though those shouldn't need array priority defined, so probably its fine.

Btw, I think I've addressed@charris's concerns in the commit I pushed earlier.

OK, one more thing. Because this changes behavior it should be mentioned indoc\release\1.8.0-notes.rst, probably under theChanges and maybe underCompatibility notes also if you think this might affect current code, as in combining__array_priority__ and__array__.

For next time note the standard prefixes for commit messages indoc/source/dev/gitwash/development_workflow.rst.

Revised, with change comments and better commit message, as requested.

Thanks for your help getting this merged!