Test fails when using numpy with MKL backend

[gwtnb]

% pytest -m 'not slow' -x tests/cupy_tests/fft_tests/test_fft.py                                                                                                                                                                                              
==================================================================================================================== test session starts =====================================================================================================================
platform linux -- Python 3.6.9, pytest-4.6.2, py-1.8.0, pluggy-0.12.0
rootdir: /path/to/cupy, inifile: setup.cfg
plugins: xdist-1.28.0, forked-1.0.2
collected 389 items / 1 deselected / 388 selected                                                                                                                                                                                                            

tests/cupy_tests/fft_tests/test_fft.py ......................................................................................F

========================================================================================================================== FAILURES ==========================================================================================================================
_________________________________________________________________________________________________________________ TestFft2_param_1.test_fft2 _________________________________________________________________________________________________________________

self = <cupy.testing.parameterized.TestFft2_param_1 testMethod=test_fft2>, args = (), kw = {'enable_nd': True}, planning_state = True, nd_planning = True

    @functools.wraps(impl)
    def test_func(self, *args, **kw):
        # get original global planning state
        planning_state = config.enable_nd_planning
        try:
            for nd_planning in states:
                try:
                    # enable or disable nd planning
                    config.enable_nd_planning = nd_planning
    
                    kw[name] = nd_planning
>                   impl(self, *args, **kw)

tests/cupy_tests/fft_tests/test_fft.py:41: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
cupy/testing/helper.py:627: in test_func
    impl(self, *args, **kw)
cupy/testing/helper.py:156: in test_func
    accept_error=accept_error)
cupy/testing/helper.py:92: in _check_cupy_numpy_error
    self.fail('Only numpy raises error\n\n' + numpy_tb)
E   AssertionError: Parameterized test failed.
E   
E   Base test method: TestFft2.test_fft2
E   Test parameters:
E     shape: (3, 4)
E     s: (1, None)
E     axes: None
E     norm: None
E   
E   AssertionError: Only numpy raises error
E   
E   Traceback (most recent call last):
E     File "/path/to/cupy/cupy/testing/helper.py", line 28, in _call_func
E       result = impl(self, *args, **kw)
E     File "/path/to/cupy/tests/cupy_tests/fft_tests/test_fft.py", line 207, in test_fft2
E       out = xp.fft.fft2(a, s=self.s, norm=self.norm)
E     File "/some/path/lib/python3.6/site-packages/mkl_fft/_numpy_fft.py", line 875, in fft2
E       return fftn(x, s=s, axes=axes, norm=norm)
E     File "/some/path/lib/python3.6/site-packages/mkl_fft/_numpy_fft.py", line 683, in fftn
E       output = mkl_fft.fftn(x, s, axes)
E     File "mkl_fft/_pydfti.pyx", line 860, in mkl_fft._pydfti.fftn
E     File "mkl_fft/_pydfti.pyx", line 846, in mkl_fft._pydfti._fftnd_impl
E     File "mkl_fft/_pydfti.pyx", line 724, in mkl_fft._pydfti._iter_fftnd
E     File "mkl_fft/_pydfti.pyx", line 684, in mkl_fft._pydfti._init_nd_shape_and_axes
E   ValueError: when given, shape values must be integers
-------------------------------------------------------------------------------------------------------------------- Captured stdout call --------------------------------------------------------------------------------------------------------------------
dtype is <class 'numpy.float64'>
enable_nd is True
===================================================================================================== 1 failed, 86 passed, 1 deselected in 3.68 seconds ======================================================================================================

[gwtnb]

Note that numpy with MKL backend is default on conda.

[leofang]

It'd be nice to get this fixed. It's indeed annoying when I was testing the FFT module. But I am not sure if we are able to detect the NumPy's backend in tests?

[grlee77]

But I am not sure if we are able to detect the NumPy's backend in tests?

Even when MKL is present, non-MKL versions of the functions should still be available. For NumPy <1.17, these are numpy.fft.fftpack.*, etc. (for numpy >=1.17 numpy.fft.pocketfft.*)`.

[leofang]

@grlee77 wow is numpy.fft.fftpack an undocumented feature of numpy.fft?

[gwtnb]

chainer/chainer-test#177

[kmaehashi]

We discussed and using numpy.fft.fftpack in unit tests sounds reasonable.