Tests generate large number of numpy warnings about scico

[bwohlberg]

Confirmed using the test recommended in the linked analysis that the dynamic libraries included with the bm3d package do mess with the CPU floating point flags, as also described there.

[bwohlberg]

An additional update on this issue. The dynamic libraries distributed with bm3d and bm4d were tested using a minor variation of the test script recommended in the link above:

import sys

from ctypes import CDLL

CDLL(sys.argv[1])

import numpy as np

print("%e" % np.finfo(np.float32).smallest_subnormal)
print("Smallest subnormal is zero: ", np.finfo(np.float32).smallest_subnormal == 0.0)

which was run as in

python fptest.py /[full path redacted]/lib/python3.9/site-packages/bm4d/libbm4d.so

All of the dynamic libraries distributed with bm3d and bm4d test positive for the undesirable floating point behaviour, but strangely, when testing by actually importing the two packages, only bm4d seems to result in the undesirable handling of subnormal numbers.

[bwohlberg]

This issue is confirmed under Linux, but does not seem to be present in the MacOS version of bm4d.

[bwohlberg]

Resolved with the recent bm3d and bm4d updates.