BUG: abs(np.int8(-128)) should arguably give an OverflowWarning

[hujie369]

Describe the issue:

The range of np.int8 is -128 to 127.
The ideal range of np.abs(np.int8) is -127 to 128.
But apparently 128 overflow.
So np.abs(-128, dtype=int8) = -128, while the rest of the cases are all correct.
This error is hard to detect and is also found in np.int16.

Reproduce the code example:

import numpy as np
a = np.array([-128, -127], dtype=np.int8)
print(a)
# [-128 -127]

print(np.abs(a))
# [-128, 127]

Error message:

No response

NumPy/Python version information:

numpy: 1.18.5
sys: 3.7.10 (default, Jun 4 2021, 14:48:32)
[GCC 7.5.0]

[charris]

This is known behavior. One option would be to return uint8, but that has its own problems as it would lead to type promotions when combined with int8. Best to just accept it and do your own promotion if needed.

In [1]: np.abs(np.array(-128, np.int8)).astype(np.uint8)
Out[1]: 128

[hujie369]

Thanks for your response!

[alonme]

Hey @charris,

This behavior is a bit confusing, what do you think about printing a warning in cases like these?
Or maybe adding a parameter to abs to do the type promotions if needed?

Otherwise - is there an easy way to check if an array contains the max/min of the dtype

[seberg]

Yeah, this is mainly a duplicate of e.g. gh-21237 (which is very similar). The basic thing, is that this issue is part of a long list of issues around the fact that it would be good if NumPy was able to give (maybe optional) integer overflow/rollover warnings.

However, we do give warnings for scalar operations. That may be slightly strange, but for scalar operations the check does not impose any significant performance cost.

In that sense, the following arguably should already give a warning:

abs(np.int8(-128))

but does not. I think this is a nice easy thing to fix, although would prefer it happening after gh-21188 is merged.

EDIT: Fixing this requires diving into a fairly complex file of C code (scalarmath.c.src), the relvant code in question will not be complicated though.