Thanks, do you think it might make sense to push it even further down to the integer_representation function? Since the "signed magnitude" representation is not just a view as far as I understand it, that might make sense?

not so sure about that, it is the max ULP diff function that doesn't want to differentiate between NAN's. The integer_representation function simply returns the binary representation of a floating point and I think that should stay true to its intent for NAN's as well.

x[np.argwhere(np.isnan(x))] = np.nan fails with E IndexError: too many indices for array for a single element array like x = np.array(1.2) Any suggestions on how to elegantly fix it without special casing it?

@r-devulap any reasons you are using np.argwhere at all, just remove it? Not sure I agree with the interpretation of integer_representation, but it also does not really matter much right now.

@r-devulap any reasons you are using np.argwhere at all, just remove it?

Thanks :)

@r-devulap looks good to me, would you be up for adding a test as well, since I think the ulp thing is arguably a bug fix in our testing utils, and also testing utils need testing :).

Sure, what do you have in mind? ensure ulp error between two different representations of NAN = 0?

@seberg let me know if that suffices.

Test would be nicer if you move it into numpy/testing/tests/test_utils.py. Otherwise looks good to me, I will let Matti sign off on the line removal changes and merge. Thanks!

I think there is more to do with the umath_accuracy test but this solves the current issues. xref gh-15605 for some possible improvements

Thanks @r-devulap for the quick response