Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

sign special case implementations #136

Closed
mdhaber opened this issue May 9, 2024 · 14 comments · Fixed by #137
Closed

sign special case implementations #136

mdhaber opened this issue May 9, 2024 · 14 comments · Fixed by #137

Comments

@mdhaber
Copy link

mdhaber commented May 9, 2024

According to v2022.12 (and v2023.12) of the array API standard, the special cases of sign include:

For real-valued operands... If x_i is NaN, the result is NaN.

However,

from array_api_compat import numpy as np, cupy as cp, torch
np.sign(np.asarray(np.nan))  # nan
cp.sign(cp.asarray(cp.nan))  # array( 0.000e+00)
torch.sign(torch.asarray(torch.nan))  # tensor(0.)

There may be other special cases that are not yet implemented. I haven't done a complete review, but I noticed that torch gives an error when the input is complex.

torch.sign(torch.asarray(1+1j))
# RuntimeError: Unlike NumPy, torch.sign is not intended to support complex numbers. Please use torch.sgn instead.
@mdhaber mdhaber changed the title array_api_compat sign special case implementations sign special case implementations May 9, 2024
@asmeurer
Copy link
Member

I haven't tried to do any special-cases workarounds here yet. I guess if this is causing issues for you we can add a workaround.

@asmeurer
Copy link
Member

We definitely should add complex support to torch.sign. It sounds like that will just be an easy wrapper of torch.sgn.

@mdhaber
Copy link
Author

mdhaber commented May 10, 2024 •
edited
Loading

Yeah I actually did run into this special case while converting code from NumPy to array API. The purpose of the code is to produce a different status flag that depends on whether an element is positive, negative, zero, or NaN, so no wonder I ran into it. So it would be helpful to add the special case, but support for 2023.12 would be higher priority to me.

@asmeurer
Copy link
Member

We can add it. The main concern with adding special-case handling is it means adding a mask to the function, so it could be a minor hit to performance. I would at least make sure there are upstream issues about this to the libraries that don't implement it correctly.

@rgommers
Copy link
Member

The main concern with adding special-case handling is it means adding a mask to the function, so it could be a minor hit to performance.

I think that won't be a minor hit. For element-wise functions, adding isnan checks and masking may slow things down by 2x or more.

In general, these special cases have not been well validated yet, so I'd be quite reluctant to assume they are all correctly specified or support them in the compat layer.

In this case, it may be possible to fix in NumPy. There's a bunch of special-casing to make sign(nan) return nan, which looks like old code and disagrees with C99's signbit on purpose:

https://github.com/numpy/numpy/blob/2a9b9134270371b43223fc848b753fceab96b4a5/numpy/_core/src/umath/loops.c.src#L1341-L1350

https://github.com/numpy/numpy/blob/2a9b9134270371b43223fc848b753fceab96b4a5/numpy/_core/src/umath/loops.c.src#L2262-L2271

Changing it would be a minor backwards-compat break, but probably an improvement. Much larger changes were made to np.sign recently in gh-numpy/numpy#25441, however the sign(nan) = nan was not discussed there. It does seem like sign deviates from signbit for no good reason; it was probably an ad-hoc pre-C99 decision.

@asmeurer
Copy link
Member

NumPy is the one that agrees with what the array API says. If we think sign(nan) should not return NaN, then the array API should change.

@rgommers
Copy link
Member

If we think sign(nan) should not return NaN, then the array API should change.

I'm honestly not sure. These special cases are a pain, it requires a lot of time investment to figure out what was discussed before, and why different libraries end up with different return values. I suspect the following:

@seberg
Copy link

seberg commented May 10, 2024

I think one normal rule of NaNs is that they are propagated unless there is a clear reason why not.
sign is very different from signbit after all as that is defined to return a bool and additionally maps to an implementation detail of IEEE float representation.

Further, returning NaN preserves the full "partial order" (e.g. in C++20) of value <=> 0 (less, equal, greater, and unordered).

So IMO, NumPy does the right thing unless there is a good argument why typical use-cases would expect a 0 return for NaN and it sounds like the report here is a use-case where you want the full partial order to be preserved!

TBH, I would lean towards torch should fix this, but if there is a good reason why they don't want to (what is it?), then it has to stay undefined which seems unfortunate for the actual use-case above.

@asmeurer
Copy link
Member

I found this old discussion https://mail.python.org/archives/list/numpy-discussion@python.org/thread/A2JFHOZZOF634CNZ7E27THQEBU4EZFTS/#F3KS7QWVPIXSYB7CSSY37OXYM4JVZTZQ

sign and signbit are completely different things for complex numbers, as I pointed out at https://mail.python.org/archives/list/numpy-discussion@python.org/message/VBYOVSTN2GTBPEJ3OPDS2S5DLPQJFFX3/ It's probably not a coincidence that torch also doesn't define complex sign.

It does seem valuable to figure this out, since there's at least one real-world use-case. Maybe we should move this discussion to the array-api repo.

@asmeurer asmeurer mentioned this issue May 10, 2024
Merged
@asmeurer
Copy link
Member

I have a fix at #137, which we can at least use to check the performance implications.

@asmeurer
Copy link
Member

By the way, you can see the other special cases that are not being followed in the xfails files (we do not yet attempt to fix any of them, and most are on operators which can't be fixed anyways):

torch

array-api-compat/torch-xfails.txt

Lines 89 to 179 in 376038e

array_api_tests/test_special_cases.py::test_binary[__add__((x1_i is +0 or x1_i == -0) and isfinite(x2_i) and x2_i != 0) -> x2_i]
array_api_tests/test_special_cases.py::test_binary[__add__(isfinite(x1_i) and x1_i != 0 and (x2_i is +0 or x2_i == -0)) -> x1_i]
array_api_tests/test_special_cases.py::test_binary[__add__(isfinite(x1_i) and x1_i != 0 and x2_i == -x1_i) -> +0]
array_api_tests/test_special_cases.py::test_binary[__add__(isfinite(x1_i) and x2_i is +infinity) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[__add__(isfinite(x1_i) and x2_i is -infinity) -> -infinity]
array_api_tests/test_special_cases.py::test_binary[__add__(x1_i is +infinity and isfinite(x2_i)) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[__add__(x1_i is -infinity and isfinite(x2_i)) -> -infinity]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(isfinite(x1_i) and x1_i < 0 and x2_i is +infinity) -> -0]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(isfinite(x1_i) and x1_i < 0 and x2_i is -infinity) -> +0]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(isfinite(x1_i) and x1_i > 0 and x2_i is +infinity) -> +0]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(isfinite(x1_i) and x1_i > 0 and x2_i is -infinity) -> -0]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(x1_i < 0 and x2_i is +0) -> -infinity]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(x1_i < 0 and x2_i is -0) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(x1_i > 0 and x2_i is +0) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(x1_i > 0 and x2_i is -0) -> -infinity]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(x1_i is +0 and x2_i < 0) -> -0]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(x1_i is +0 and x2_i > 0) -> +0]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(x1_i is +infinity and isfinite(x2_i) and x2_i < 0) -> -infinity]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(x1_i is +infinity and isfinite(x2_i) and x2_i > 0) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(x1_i is -0 and x2_i < 0) -> +0]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(x1_i is -0 and x2_i > 0) -> -0]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(x1_i is -infinity and isfinite(x2_i) and x2_i < 0) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(x1_i is -infinity and isfinite(x2_i) and x2_i > 0) -> -infinity]
array_api_tests/test_special_cases.py::test_binary[__mod__(isfinite(x1_i) and x1_i < 0 and x2_i is +infinity) -> x2_i]
array_api_tests/test_special_cases.py::test_binary[__mod__(isfinite(x1_i) and x1_i < 0 and x2_i is -infinity) -> x1_i]
array_api_tests/test_special_cases.py::test_binary[__mod__(isfinite(x1_i) and x1_i > 0 and x2_i is +infinity) -> x1_i]
array_api_tests/test_special_cases.py::test_binary[__mod__(isfinite(x1_i) and x1_i > 0 and x2_i is -infinity) -> x2_i]
array_api_tests/test_special_cases.py::test_binary[__mod__(x1_i is +0 and x2_i < 0) -> -0]
array_api_tests/test_special_cases.py::test_binary[__mod__(x1_i is +0 and x2_i > 0) -> +0]
array_api_tests/test_special_cases.py::test_binary[__mod__(x1_i is -0 and x2_i < 0) -> -0]
array_api_tests/test_special_cases.py::test_binary[__mod__(x1_i is -0 and x2_i > 0) -> +0]
array_api_tests/test_special_cases.py::test_binary[__pow__(abs(x1_i) < 1 and x2_i is +infinity) -> +0]
array_api_tests/test_special_cases.py::test_binary[__pow__(abs(x1_i) < 1 and x2_i is -infinity) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[__pow__(abs(x1_i) > 1 and x2_i is +infinity) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[__pow__(abs(x1_i) > 1 and x2_i is -infinity) -> +0]
array_api_tests/test_special_cases.py::test_binary[__pow__(x1_i < 0 and isfinite(x1_i) and isfinite(x2_i) and not x2_i.is_integer()) -> NaN]
array_api_tests/test_special_cases.py::test_binary[__pow__(x1_i is +0 and x2_i < 0) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[__pow__(x1_i is +0 and x2_i > 0) -> +0]
array_api_tests/test_special_cases.py::test_binary[__pow__(x1_i is +infinity and x2_i < 0) -> +0]
array_api_tests/test_special_cases.py::test_binary[__pow__(x1_i is +infinity and x2_i > 0) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[__pow__(x1_i is -0 and x2_i < 0 and not (x2_i.is_integer() and x2_i % 2 == 1)) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[__pow__(x1_i is -0 and x2_i < 0 and x2_i.is_integer() and x2_i % 2 == 1) -> -infinity]
array_api_tests/test_special_cases.py::test_binary[__pow__(x1_i is -0 and x2_i > 0 and not (x2_i.is_integer() and x2_i % 2 == 1)) -> +0]
array_api_tests/test_special_cases.py::test_binary[__pow__(x1_i is -0 and x2_i > 0 and x2_i.is_integer() and x2_i % 2 == 1) -> -0]
array_api_tests/test_special_cases.py::test_binary[__pow__(x1_i is -infinity and x2_i < 0 and not (x2_i.is_integer() and x2_i % 2 == 1)) -> +0]
array_api_tests/test_special_cases.py::test_binary[__pow__(x1_i is -infinity and x2_i < 0 and x2_i.is_integer() and x2_i % 2 == 1) -> -0]
array_api_tests/test_special_cases.py::test_binary[__pow__(x1_i is -infinity and x2_i > 0 and not (x2_i.is_integer() and x2_i % 2 == 1)) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[__pow__(x1_i is -infinity and x2_i > 0 and x2_i.is_integer() and x2_i % 2 == 1) -> -infinity]
array_api_tests/test_special_cases.py::test_binary[__pow__(x1_i is NaN and not x2_i == 0) -> NaN]
array_api_tests/test_special_cases.py::test_binary[__truediv__(isfinite(x1_i) and x1_i < 0 and x2_i is +infinity) -> -0]
array_api_tests/test_special_cases.py::test_binary[__truediv__(isfinite(x1_i) and x1_i < 0 and x2_i is -infinity) -> +0]
array_api_tests/test_special_cases.py::test_binary[__truediv__(isfinite(x1_i) and x1_i > 0 and x2_i is +infinity) -> +0]
array_api_tests/test_special_cases.py::test_binary[__truediv__(isfinite(x1_i) and x1_i > 0 and x2_i is -infinity) -> -0]
array_api_tests/test_special_cases.py::test_binary[__truediv__(x1_i < 0 and x2_i is +0) -> -infinity]
array_api_tests/test_special_cases.py::test_binary[__truediv__(x1_i < 0 and x2_i is -0) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[__truediv__(x1_i > 0 and x2_i is +0) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[__truediv__(x1_i > 0 and x2_i is -0) -> -infinity]
array_api_tests/test_special_cases.py::test_binary[__truediv__(x1_i is +0 and x2_i < 0) -> -0]
array_api_tests/test_special_cases.py::test_binary[__truediv__(x1_i is +0 and x2_i > 0) -> +0]
array_api_tests/test_special_cases.py::test_binary[__truediv__(x1_i is +infinity and isfinite(x2_i) and x2_i < 0) -> -infinity]
array_api_tests/test_special_cases.py::test_binary[__truediv__(x1_i is +infinity and isfinite(x2_i) and x2_i > 0) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[__truediv__(x1_i is -0 and x2_i < 0) -> +0]
array_api_tests/test_special_cases.py::test_binary[__truediv__(x1_i is -0 and x2_i > 0) -> -0]
array_api_tests/test_special_cases.py::test_binary[__truediv__(x1_i is -infinity and isfinite(x2_i) and x2_i < 0) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[__truediv__(x1_i is -infinity and isfinite(x2_i) and x2_i > 0) -> -infinity]
array_api_tests/test_special_cases.py::test_binary[add(isfinite(x1_i) and x1_i != 0 and x2_i == -x1_i) -> +0]
array_api_tests/test_special_cases.py::test_binary[floor_divide(isfinite(x1_i) and x1_i < 0 and x2_i is +infinity) -> -0]
array_api_tests/test_special_cases.py::test_binary[floor_divide(isfinite(x1_i) and x1_i > 0 and x2_i is -infinity) -> -0]
array_api_tests/test_special_cases.py::test_binary[floor_divide(x1_i is +infinity and isfinite(x2_i) and x2_i < 0) -> -infinity]
array_api_tests/test_special_cases.py::test_binary[floor_divide(x1_i is +infinity and isfinite(x2_i) and x2_i > 0) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[floor_divide(x1_i is -infinity and isfinite(x2_i) and x2_i < 0) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[floor_divide(x1_i is -infinity and isfinite(x2_i) and x2_i > 0) -> -infinity]
array_api_tests/test_special_cases.py::test_binary[remainder(x1_i is +0 and x2_i < 0) -> -0]
array_api_tests/test_special_cases.py::test_binary[remainder(x1_i is -0 and x2_i > 0) -> +0]
array_api_tests/test_special_cases.py::test_iop[__iadd__(isfinite(x1_i) and x1_i != 0 and x2_i == -x1_i) -> +0]
array_api_tests/test_special_cases.py::test_iop[__ifloordiv__(isfinite(x1_i) and x1_i < 0 and x2_i is +infinity) -> -0]
array_api_tests/test_special_cases.py::test_iop[__ifloordiv__(isfinite(x1_i) and x1_i > 0 and x2_i is -infinity) -> -0]
array_api_tests/test_special_cases.py::test_iop[__ifloordiv__(x1_i is +infinity and isfinite(x2_i) and x2_i < 0) -> -infinity]
array_api_tests/test_special_cases.py::test_iop[__ifloordiv__(x1_i is +infinity and isfinite(x2_i) and x2_i > 0) -> +infinity]
array_api_tests/test_special_cases.py::test_iop[__ifloordiv__(x1_i is -infinity and isfinite(x2_i) and x2_i < 0) -> +infinity]
array_api_tests/test_special_cases.py::test_iop[__ifloordiv__(x1_i is -infinity and isfinite(x2_i) and x2_i > 0) -> -infinity]
array_api_tests/test_special_cases.py::test_iop[__imod__(x1_i is +0 and x2_i < 0) -> -0]
array_api_tests/test_special_cases.py::test_iop[__imod__(x1_i is -0 and x2_i > 0) -> +0]
array_api_tests/test_special_cases.py::test_unary[sign(x_i is NaN) -> NaN]
# Float correction is not supported by pytorch
# (https://github.com/data-apis/array-api-tests/issues/168)
array_api_tests/test_special_cases.py::test_empty_arrays[std]
array_api_tests/test_special_cases.py::test_empty_arrays[var]
array_api_tests/test_special_cases.py::test_nan_propagation[std]
array_api_tests/test_special_cases.py::test_nan_propagation[var]

cupy

array-api-compat/cupy-xfails.txt

Lines 61 to 169 in 376038e

array_api_tests/test_special_cases.py::test_binary[__add__(isfinite(x1_i) and x1_i != 0 and x2_i == -x1_i) -> +0]
array_api_tests/test_special_cases.py::test_binary[__add__(x1_i is +0 and x2_i is -0) -> +0]
array_api_tests/test_special_cases.py::test_binary[__add__(x1_i is -0 and x2_i is +0) -> +0]
array_api_tests/test_special_cases.py::test_binary[__add__(x1_i is -0 and x2_i is -0) -> -0]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(x1_i < 0 and x2_i is -0) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(x1_i > 0 and x2_i is -0) -> -infinity]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(x1_i is -0 and x2_i < 0) -> +0]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(x1_i is -0 and x2_i > 0) -> -0]
array_api_tests/test_special_cases.py::test_binary[__mod__(isfinite(x1_i) and x1_i < 0 and x2_i is +infinity) -> x2_i]
array_api_tests/test_special_cases.py::test_binary[__mod__(isfinite(x1_i) and x1_i < 0 and x2_i is -infinity) -> x1_i]
array_api_tests/test_special_cases.py::test_binary[__mod__(isfinite(x1_i) and x1_i > 0 and x2_i is +infinity) -> x1_i]
array_api_tests/test_special_cases.py::test_binary[__mod__(isfinite(x1_i) and x1_i > 0 and x2_i is -infinity) -> x2_i]
array_api_tests/test_special_cases.py::test_binary[__mod__(x1_i < 0 and x2_i is -0) -> NaN]
array_api_tests/test_special_cases.py::test_binary[__mod__(x1_i > 0 and x2_i is -0) -> NaN]
array_api_tests/test_special_cases.py::test_binary[__mod__(x1_i is +0 and x2_i < 0) -> -0]
array_api_tests/test_special_cases.py::test_binary[__mod__(x1_i is +0 and x2_i > 0) -> +0]
array_api_tests/test_special_cases.py::test_binary[__mod__(x1_i is -0 and x2_i < 0) -> -0]
array_api_tests/test_special_cases.py::test_binary[__mod__(x1_i is -0 and x2_i > 0) -> +0]
array_api_tests/test_special_cases.py::test_binary[__pow__(x1_i is -0 and x2_i < 0 and not (x2_i.is_integer() and x2_i % 2 == 1)) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[__pow__(x1_i is -0 and x2_i < 0 and x2_i.is_integer() and x2_i % 2 == 1) -> -infinity]
array_api_tests/test_special_cases.py::test_binary[__pow__(x1_i is -0 and x2_i > 0 and not (x2_i.is_integer() and x2_i % 2 == 1)) -> +0]
array_api_tests/test_special_cases.py::test_binary[__pow__(x1_i is -0 and x2_i > 0 and x2_i.is_integer() and x2_i % 2 == 1) -> -0]
array_api_tests/test_special_cases.py::test_binary[__pow__(x2_i is -0) -> 1]
array_api_tests/test_special_cases.py::test_binary[__truediv__(x1_i < 0 and x2_i is -0) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[__truediv__(x1_i > 0 and x2_i is -0) -> -infinity]
array_api_tests/test_special_cases.py::test_binary[__truediv__(x1_i is -0 and x2_i < 0) -> +0]
array_api_tests/test_special_cases.py::test_binary[__truediv__(x1_i is -0 and x2_i > 0) -> -0]
array_api_tests/test_special_cases.py::test_binary[add(isfinite(x1_i) and x1_i != 0 and x2_i == -x1_i) -> +0]
array_api_tests/test_special_cases.py::test_binary[add(x1_i is +0 and x2_i is -0) -> +0]
array_api_tests/test_special_cases.py::test_binary[add(x1_i is -0 and x2_i is +0) -> +0]
array_api_tests/test_special_cases.py::test_binary[add(x1_i is -0 and x2_i is -0) -> -0]
array_api_tests/test_special_cases.py::test_binary[atan2(x1_i < 0 and x2_i is -0) -> roughly -pi/2]
array_api_tests/test_special_cases.py::test_binary[atan2(x1_i > 0 and x2_i is -0) -> roughly +pi/2]
array_api_tests/test_special_cases.py::test_binary[atan2(x1_i is +0 and x2_i is -0) -> roughly +pi]
array_api_tests/test_special_cases.py::test_binary[atan2(x1_i is -0 and x2_i < 0) -> roughly -pi]
array_api_tests/test_special_cases.py::test_binary[atan2(x1_i is -0 and x2_i > 0) -> -0]
array_api_tests/test_special_cases.py::test_binary[atan2(x1_i is -0 and x2_i is +0) -> -0]
array_api_tests/test_special_cases.py::test_binary[atan2(x1_i is -0 and x2_i is -0) -> roughly -pi]
array_api_tests/test_special_cases.py::test_binary[divide(x1_i < 0 and x2_i is -0) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[divide(x1_i > 0 and x2_i is -0) -> -infinity]
array_api_tests/test_special_cases.py::test_binary[divide(x1_i is -0 and x2_i < 0) -> +0]
array_api_tests/test_special_cases.py::test_binary[divide(x1_i is -0 and x2_i > 0) -> -0]
array_api_tests/test_special_cases.py::test_binary[floor_divide(x1_i < 0 and x2_i is -0) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[floor_divide(x1_i > 0 and x2_i is -0) -> -infinity]
array_api_tests/test_special_cases.py::test_binary[floor_divide(x1_i is -0 and x2_i < 0) -> +0]
array_api_tests/test_special_cases.py::test_binary[floor_divide(x1_i is -0 and x2_i > 0) -> -0]
array_api_tests/test_special_cases.py::test_binary[pow(x1_i is -0 and x2_i < 0 and not (x2_i.is_integer() and x2_i % 2 == 1)) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[pow(x1_i is -0 and x2_i < 0 and x2_i.is_integer() and x2_i % 2 == 1) -> -infinity]
array_api_tests/test_special_cases.py::test_binary[pow(x1_i is -0 and x2_i > 0 and not (x2_i.is_integer() and x2_i % 2 == 1)) -> +0]
array_api_tests/test_special_cases.py::test_binary[pow(x1_i is -0 and x2_i > 0 and x2_i.is_integer() and x2_i % 2 == 1) -> -0]
array_api_tests/test_special_cases.py::test_binary[pow(x2_i is -0) -> 1]
array_api_tests/test_special_cases.py::test_binary[remainder(isfinite(x1_i) and x1_i < 0 and x2_i is +infinity) -> x2_i]
array_api_tests/test_special_cases.py::test_binary[remainder(isfinite(x1_i) and x1_i < 0 and x2_i is -infinity) -> x1_i]
array_api_tests/test_special_cases.py::test_binary[remainder(isfinite(x1_i) and x1_i > 0 and x2_i is +infinity) -> x1_i]
array_api_tests/test_special_cases.py::test_binary[remainder(isfinite(x1_i) and x1_i > 0 and x2_i is -infinity) -> x2_i]
array_api_tests/test_special_cases.py::test_binary[remainder(x1_i < 0 and x2_i is -0) -> NaN]
array_api_tests/test_special_cases.py::test_binary[remainder(x1_i > 0 and x2_i is -0) -> NaN]
array_api_tests/test_special_cases.py::test_binary[remainder(x1_i is +0 and x2_i < 0) -> -0]
array_api_tests/test_special_cases.py::test_binary[remainder(x1_i is +0 and x2_i > 0) -> +0]
array_api_tests/test_special_cases.py::test_binary[remainder(x1_i is -0 and x2_i < 0) -> -0]
array_api_tests/test_special_cases.py::test_binary[remainder(x1_i is -0 and x2_i > 0) -> +0]
array_api_tests/test_special_cases.py::test_iop[__iadd__(isfinite(x1_i) and x1_i != 0 and x2_i == -x1_i) -> +0]
array_api_tests/test_special_cases.py::test_iop[__iadd__(x1_i is +0 and x2_i is -0) -> +0]
array_api_tests/test_special_cases.py::test_iop[__iadd__(x1_i is -0 and x2_i is +0) -> +0]
array_api_tests/test_special_cases.py::test_iop[__iadd__(x1_i is -0 and x2_i is -0) -> -0]
array_api_tests/test_special_cases.py::test_iop[__ifloordiv__(x1_i < 0 and x2_i is -0) -> +infinity]
array_api_tests/test_special_cases.py::test_iop[__ifloordiv__(x1_i > 0 and x2_i is -0) -> -infinity]
array_api_tests/test_special_cases.py::test_iop[__ifloordiv__(x1_i is -0 and x2_i < 0) -> +0]
array_api_tests/test_special_cases.py::test_iop[__ifloordiv__(x1_i is -0 and x2_i > 0) -> -0]
array_api_tests/test_special_cases.py::test_iop[__imod__(isfinite(x1_i) and x1_i < 0 and x2_i is +infinity) -> x2_i]
array_api_tests/test_special_cases.py::test_iop[__imod__(isfinite(x1_i) and x1_i < 0 and x2_i is -infinity) -> x1_i]
array_api_tests/test_special_cases.py::test_iop[__imod__(isfinite(x1_i) and x1_i > 0 and x2_i is +infinity) -> x1_i]
array_api_tests/test_special_cases.py::test_iop[__imod__(isfinite(x1_i) and x1_i > 0 and x2_i is -infinity) -> x2_i]
array_api_tests/test_special_cases.py::test_iop[__imod__(x1_i < 0 and x2_i is -0) -> NaN]
array_api_tests/test_special_cases.py::test_iop[__imod__(x1_i > 0 and x2_i is -0) -> NaN]
array_api_tests/test_special_cases.py::test_iop[__imod__(x1_i is +0 and x2_i < 0) -> -0]
array_api_tests/test_special_cases.py::test_iop[__imod__(x1_i is +0 and x2_i > 0) -> +0]
array_api_tests/test_special_cases.py::test_iop[__imod__(x1_i is -0 and x2_i < 0) -> -0]
array_api_tests/test_special_cases.py::test_iop[__imod__(x1_i is -0 and x2_i > 0) -> +0]
array_api_tests/test_special_cases.py::test_iop[__ipow__(x1_i is -0 and x2_i < 0 and not (x2_i.is_integer() and x2_i % 2 == 1)) -> +infinity]
array_api_tests/test_special_cases.py::test_iop[__ipow__(x1_i is -0 and x2_i < 0 and x2_i.is_integer() and x2_i % 2 == 1) -> -infinity]
array_api_tests/test_special_cases.py::test_iop[__ipow__(x1_i is -0 and x2_i > 0 and not (x2_i.is_integer() and x2_i % 2 == 1)) -> +0]
array_api_tests/test_special_cases.py::test_iop[__ipow__(x1_i is -0 and x2_i > 0 and x2_i.is_integer() and x2_i % 2 == 1) -> -0]
array_api_tests/test_special_cases.py::test_iop[__ipow__(x2_i is -0) -> 1]
array_api_tests/test_special_cases.py::test_iop[__itruediv__(x1_i < 0 and x2_i is -0) -> +infinity]
array_api_tests/test_special_cases.py::test_iop[__itruediv__(x1_i > 0 and x2_i is -0) -> -infinity]
array_api_tests/test_special_cases.py::test_iop[__itruediv__(x1_i is -0 and x2_i < 0) -> +0]
array_api_tests/test_special_cases.py::test_iop[__itruediv__(x1_i is -0 and x2_i > 0) -> -0]
array_api_tests/test_special_cases.py::test_unary[__abs__(x_i is -0) -> +0]
array_api_tests/test_special_cases.py::test_unary[abs(x_i is -0) -> +0]
array_api_tests/test_special_cases.py::test_unary[asin(x_i is -0) -> -0]
array_api_tests/test_special_cases.py::test_unary[asinh(x_i is -0) -> -0]
array_api_tests/test_special_cases.py::test_unary[atan(x_i is -0) -> -0]
array_api_tests/test_special_cases.py::test_unary[atanh(x_i is -0) -> -0]
array_api_tests/test_special_cases.py::test_unary[ceil(x_i is -0) -> -0]
array_api_tests/test_special_cases.py::test_unary[cos(x_i is -0) -> 1]
array_api_tests/test_special_cases.py::test_unary[cosh(x_i is -0) -> 1]
array_api_tests/test_special_cases.py::test_unary[exp(x_i is -0) -> 1]
array_api_tests/test_special_cases.py::test_unary[expm1(x_i is -0) -> -0]
array_api_tests/test_special_cases.py::test_unary[floor(x_i is -0) -> -0]
array_api_tests/test_special_cases.py::test_unary[log1p(x_i is -0) -> -0]
array_api_tests/test_special_cases.py::test_unary[round(x_i is -0) -> -0]
array_api_tests/test_special_cases.py::test_unary[sign(x_i is NaN) -> NaN]
array_api_tests/test_special_cases.py::test_unary[sin(x_i is -0) -> -0]
array_api_tests/test_special_cases.py::test_unary[sinh(x_i is -0) -> -0]
array_api_tests/test_special_cases.py::test_unary[sqrt(x_i is -0) -> -0]
array_api_tests/test_special_cases.py::test_unary[tan(x_i is -0) -> -0]
array_api_tests/test_special_cases.py::test_unary[tanh(x_i is -0) -> -0]
array_api_tests/test_special_cases.py::test_unary[trunc(x_i is -0) -> -0]

numpy

array-api-compat/numpy-xfails.txt

Lines 18 to 41 in 376038e

array_api_tests/test_special_cases.py::test_binary[floor_divide(x1_i is +infinity and isfinite(x2_i) and x2_i > 0) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[floor_divide(x1_i is +infinity and isfinite(x2_i) and x2_i < 0) -> -infinity]
array_api_tests/test_special_cases.py::test_binary[floor_divide(x1_i is -infinity and isfinite(x2_i) and x2_i > 0) -> -infinity]
array_api_tests/test_special_cases.py::test_binary[floor_divide(x1_i is -infinity and isfinite(x2_i) and x2_i < 0) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[floor_divide(isfinite(x1_i) and x1_i > 0 and x2_i is -infinity) -> -0]
array_api_tests/test_special_cases.py::test_binary[floor_divide(isfinite(x1_i) and x1_i < 0 and x2_i is +infinity) -> -0]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(x1_i is +infinity and isfinite(x2_i) and x2_i > 0) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(x1_i is +infinity and isfinite(x2_i) and x2_i < 0) -> -infinity]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(x1_i is -infinity and isfinite(x2_i) and x2_i > 0) -> -infinity]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(x1_i is -infinity and isfinite(x2_i) and x2_i < 0) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(isfinite(x1_i) and x1_i > 0 and x2_i is -infinity) -> -0]
array_api_tests/test_special_cases.py::test_binary[__floordiv__(isfinite(x1_i) and x1_i < 0 and x2_i is +infinity) -> -0]
array_api_tests/test_special_cases.py::test_iop[__ifloordiv__(x1_i is +infinity and isfinite(x2_i) and x2_i > 0) -> +infinity]
array_api_tests/test_special_cases.py::test_iop[__ifloordiv__(x1_i is +infinity and isfinite(x2_i) and x2_i < 0) -> -infinity]
array_api_tests/test_special_cases.py::test_iop[__ifloordiv__(x1_i is -infinity and isfinite(x2_i) and x2_i > 0) -> -infinity]
array_api_tests/test_special_cases.py::test_iop[__ifloordiv__(x1_i is -infinity and isfinite(x2_i) and x2_i < 0) -> +infinity]
array_api_tests/test_special_cases.py::test_iop[__ifloordiv__(isfinite(x1_i) and x1_i > 0 and x2_i is -infinity) -> -0]
array_api_tests/test_special_cases.py::test_iop[__ifloordiv__(isfinite(x1_i) and x1_i < 0 and x2_i is +infinity) -> -0]
# https://github.com/numpy/numpy/issues/21213
array_api_tests/test_special_cases.py::test_binary[__pow__(x1_i is -infinity and x2_i > 0 and not (x2_i.is_integer() and x2_i % 2 == 1)) -> +infinity]
array_api_tests/test_special_cases.py::test_binary[__pow__(x1_i is -0 and x2_i > 0 and not (x2_i.is_integer() and x2_i % 2 == 1)) -> +0]
array_api_tests/test_special_cases.py::test_iop[__ipow__(x1_i is -infinity and x2_i > 0 and not (x2_i.is_integer() and x2_i % 2 == 1)) -> +infinity]
array_api_tests/test_special_cases.py::test_iop[__ipow__(x1_i is -0 and x2_i > 0 and not (x2_i.is_integer() and x2_i % 2 == 1)) -> +0]

I will say that even though there are quite a few of these, the sign special case seems to stand out. Almost all the rest seem to have to do with handling -0 correctly.

@mdhaber mdhaber mentioned this issue May 10, 2024
Merged
@kgryte
Copy link

kgryte commented May 11, 2024 •
edited
Loading

@rgommers The addition of the special case for NaN handling in sign comes from data-apis/array-api#556.

The specification is correct on this, as returning NaN follows naturally from the definition of the signum function where

$$\textrm{sgn\ } x = \frac{x}{|x|}$$

and where NaN/NaN follows the special cases for division, thus ensuring arithmetic consistency.

This was not a pre-C99 oversight and was intentional. We shouldn't expect the signum and signbit functions to return equivalent results, and NumPy is correct here. It would perhaps have been better if NumPy had chosen signum as the name rather than sign to make this delineation in behavior more clear.

@seberg
Copy link

seberg commented May 18, 2024

Since its maybe interesting. Turns out getting NaNs may make this faster on the GPU: cupy/cupy#8327

@mdhaber
Copy link
Author

mdhaber commented Aug 25, 2024

Just thought I'd mention that I ran into this again in a different context.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging a pull request may close this issue.

Fix sign() for torch and cupy asmeurer/array-api-compat
5 participants
@seberg @asmeurer @rgommers @kgryte @mdhaber