[Mne_analysis] test_filter.py value mismatch error

[Lee, KyuHwa]

Hello,

I installed MNE 0.15-dev from git with CUDA 8.0 and Anaconda 4.4 (Python
3.6), following this manual:
http://martinos.org/mne/dev/advanced_setup.html?highlight=dev%20install

When I run:
mne.utils.set_config('MNE_USE_CUDA', 'true'); mne.cuda.init_cuda()
I get the correct result:
Enabling CUDA with 761.0 MB available memory

When I run test_filter.py, I get the following error:

AssertionError:
Not equal to tolerance rtol=0.001, atol=0.001

(mismatch 0.9333333333333371%)
 x: array([[ -2.558717e-16,  -1.525432e-01,  -2.835838e-01, ...,
          4.634404e-01,   2.543413e-01,  -6.938894e-17],
       [ -2.532696e-16,   1.745948e-01,   3.111300e-01, ...,
          6.255135e-01,   3.478937e-01,  -2.775558e-17]])
 y: array([[  8.933826e-17,  -1.527561e-01,  -2.836241e-01, ...,
          4.634049e-01,   2.543312e-01,  -2.428613e-16],
       [  1.565588e-16,   1.746173e-01,   3.108850e-01, ...,
          6.256731e-01,   3.481433e-01,  -3.053113e-16]])

It was caused by line 332 of test_filter.py:
assert_allclose(hp, bp, rtol=1e-3, atol=1e-3)

I tried increasing the tolerence to 1e-2, and now it throws another error
caused by line 336 (FFT calculation):
AssertionError:
Not equal to tolerance rtol=1e-07, atol=0.02

(mismatch 100.0%)
 x: array(1.0259702184926076)
 y: array(1.0)

Seems like there's a serious numerical problem with the computation. What
should I do in this case?
Weird thing is, even if I set MNE_USE_CUDA to 'false', it still gives me
exactly the same error, which might suggest it's independent of CUDA
setting?

Best wishes,
Kyuhwa
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.nmr.mgh.harvard.edu/pipermail/mne_analysis/attachments/20170923/156f32d0/attachment.html