underflow / saturation of unsigned list not working

[AndrzejKowalski9917]

Description

When subtracting a larger unsigned number from a smaller one the result is correctly 0.0. If both values are part of a list, the result is wrong, if an underflow occurs (seems to be set to the upper limit, instead of the lower one).

Minimal example:

from fxpmath import Fxp

# single unsigned value does work
a = Fxp(10, False, 14, 3)
b = Fxp(15, False, 14, 3)
c = a - b
print(c)  # 0.0 --> correct
c.info(verbose=3)

# unsigned list does not work
d = Fxp([10, 21], False, 14, 3)
e = Fxp([15, 15], False, 14, 3)
f = d - e
print(f)  # [4095.875 6.0] --> 4095.875 is the upper limit
f.info(verbose=3)

c.info(verbose=3) delivers the following output:

dtype = fxp-u15/3
Value = 0.0
underflow = True
inaccuracy = True
Signed = False
Word bits = 15
Fract bits = 3
Int bits = 12
Val data type = None
Upper = 4095.875
Lower = 0.0
Precision = 0.125
Overflow = saturate
Rounding = trunc
Shifting = expand

f.info(verbose=3) this output:

dtype = fxp-u15/3
Value = [4095.875 6.0]
overflow = True
inaccuracy = True
Signed = False
Word bits = 15
Fract bits = 3
Int bits = 12
Val data type = None
Upper = 4095.875
Lower = 0.0
Precision = 0.125
Overflow = saturate
Rounding = trunc
Shifting = expand

Version infos

• OS: Windows 10 21H2 (19044.2130)
• python: 3.7.6 (miniconda), 3.10.8 (mingw64)
• fxpmath: 0.4.8

[francof2a]

Hello @AndrzejKowalski9917

I tested your example in a fresh conda environment with python 3.7.6 (conda), fxpmath 0.4.8 (from anaconda) and I couldn't reproduce de error.
I also tested in a conda env with python 3.8.13 and fxpmath installed from Pip, and it's also working well.
Both runs were performed in linux, so I can not tell about relationship with Windows OS yet.

What version of numpy are you using?
Could you run again your example in a fresh conda env?

[AndrzejKowalski9917]

Hi @francof2a ,

thank you for your answer. I tried now in Linux (Arch) with

• python 3.10.8
• numpy 1.23.4
• fxpmath 0.4.8

and also were not able to reproduce the error. In Windows I also tried python 3.10.0, installed via the installer from python.org, fxpmath and numpy installed via pip (same versions as on Linux). Here the error occurs.
I can try a fresh conda env in Windows tomorrow at work.

[AndrzejKowalski9917]

Hello @francof2a,

I just tried at work again (Windows 10) with a clean virtualenv (conda env does not work here for reasons unknown).
Python is still version 3.7.6 and only the following packages are installed:

(test_env) D:\tmp\py>pip list
Package Version
---------- -------
fxpmath 0.4.8
numpy 1.21.6
pip 22.3.1
setuptools 65.5.1
wheel 0.38.4

The error persists. So it seems to be an error specific to the Windows OS.

EDIT:
I just noticed, while c.info(verbose=3) prints underflow = True, that on Windows f.info(verbose=3) prints overflow = True, which is also wrong.