Try to fix nan value mismatches in filtfilt tests

[andfoy]

After #7496, some failures started to appear in the cuda102 CIs regarding nan and zeros mismatches. Since these errors do not appear (and cannot be reproduced) in CIs that run newer CUDA versions, they are most probably related to unstable input coefficients which cause lfilter_zi to be nan, which then when they are applied to a signal that is zero, it will output nan as opposed to zero.

[takagi]

We can only merge this to v12 as CUDA 10.2 is planned to drop in v13.

[takagi]

Would you move this change on top of v12 branch?

[kmaehashi]

CUDA 10.2 drop is under RFC in #7557 although no one complained so far. How about skipping these tests for CUDA 10.2 (cupy.cuda.runtime.runtimeGetVersion() == 10020) like we do in

cupy/tests/cupyx_tests/scipy_tests/sparse_tests/test_csc.py

Lines 1112 to 1114 in 6f135c4

	@pytest.mark.skipif(
	not runtime.is_hip and cupy.cuda.runtime.runtimeGetVersion() < 8000,
	reason='CUDA <8 cannot keep number of non-zero entries ')

, and merge this one to both main/v12 branch?

[takagi]

/test full

[takagi]

Thanks, @andfoy!

[chainer-ci]

@takagi Failed to backport automatically.

---

Not cleanly cherry-picked

[takagi]

No need to make a backport.