def test_backward_fft(self, dtype):
t = dtype
idtype = odtype = edtype = cupy.dtype(t)
shape = self.shape
length = cupy.core.internal.prod(shape[1:])
a = testing.shaped_random(shape, cupy, dtype)
out = cupy.empty_like(a)
plan = cufft.XtPlanNd(shape[1:],
shape[1:],
1,
length,
idtype,
shape[1:],
1,
length,
odtype,
shape[0],
edtype,
order='C',
last_axis=-1,
last_size=None)
plan.fft(a, out, cufft.CUFFT_INVERSE)
if len(shape) <= 2:
out_cp = cupy.fft.ifft(a)
else:
out_cp = cupy.fft.ifftn(a, axes=(-1, -2))
testing.assert_allclose(out / length, out_cp)
def test_backward_fft_complex32(self):
t = 'E'
idtype = odtype = edtype = t
old_shape = self.shape
shape = list(self.shape)
shape[-1] = 2 * shape[-1]
shape = tuple(shape)
a = testing.shaped_random(shape, cupy, cupy.float16)
out = cupy.empty_like(a)
shape = old_shape
length = cupy.core.internal.prod(shape[1:])
plan = cufft.XtPlanNd(shape[1:],
shape[1:],
1,
length,
idtype,
shape[1:],
1,
length,
odtype,
shape[0],
edtype,
order='C',
last_axis=-1,
last_size=None)
plan.fft(a, out, cufft.CUFFT_INVERSE)
a_cp = a.astype(cupy.float32) # upcast
a_cp = a_cp.view(cupy.complex64)
if len(shape) <= 2:
out_cp = cupy.fft.ifft(a_cp)
else:
out_cp = cupy.fft.ifftn(a_cp, axes=(-1, -2))
out_cp = out_cp.view(cupy.float32)
out_cp = out_cp.astype(cupy.float16) # downcast
# We shouldn't expect getting high accuracy, as both computations have
# precision losses...
testing.assert_allclose(out / length, out_cp, rtol=1E-1, atol=1E-1)