def test_input_and_filter_sizes(nh, nx):
dtype_data = dtype_filter = np.float32
x = cp.arange(nx, dtype=dtype_data)
x_cpu = x.get()
h_cpu = np.arange(1, nh + 1, dtype=dtype_filter)
h = cp.asarray(h_cpu)
# up=1 kernel case
cp.testing.assert_allclose(signal.upfirdn(h_cpu, x_cpu, up=1, down=2),
upfirdn(h, x, up=1, down=2))
# down=1 kernel case
cp.testing.assert_allclose(signal.upfirdn(h_cpu, x_cpu, up=2, down=1),
upfirdn(h, x, up=2, down=1))
def test_dtype_combos(dtype_data, dtype_filter):
shape = (64, 64)
size = int(np.prod(shape))
x = cp.arange(size, dtype=dtype_data).reshape(shape)
x_cpu = x.get()
h_cpu = np.arange(5, dtype=dtype_filter)
h = cp.asarray(h_cpu)
# up=1 kernel case
cp.testing.assert_allclose(signal.upfirdn(h_cpu, x_cpu, up=1, down=2),
upfirdn(h, x, up=1, down=2))
# down=1 kernel case
cp.testing.assert_allclose(signal.upfirdn(h_cpu, x_cpu, up=2, down=1),
upfirdn(h, x, up=2, down=1))
def test_general_up_and_down(up, down, nx, nh):
dtype_data = dtype_filter = np.float32
x = cp.arange(nx, dtype=dtype_data)
x_cpu = x.get()
h_cpu = np.arange(1, nh + 1, dtype=dtype_filter)
h = cp.asarray(h_cpu)
cp.testing.assert_allclose(
signal.upfirdn(h_cpu, x_cpu, up=up, down=down),
upfirdn(h, x, up=up, down=down),
)
def test_modes(self, size, h_len, mode, dtype):
random_state = cp.random.RandomState(5)
x = random_state.randn(size).astype(dtype)
if dtype in (cp.complex64, cp.complex128):
x += 1j * random_state.randn(size)
h = cp.arange(1, 1 + h_len, dtype=x.real.dtype)
y = upfirdn(h, x, up=1, down=1, mode=mode)
# expected result: pad the input, filter with zero padding, then crop
npad = h_len - 1
if mode in ["antisymmetric", "antireflect", "smooth", "line"]:
# use _pad_test test function for modes not supported by cp.pad.
xpad = _pad_test(x, npre=npad, npost=npad, mode=mode)
else:
xpad = cp.pad(x, npad, mode=mode)
ypad = upfirdn(h, xpad, up=1, down=1, mode="constant")
y_expected = ypad[npad:-npad]
atol = rtol = cp.finfo(dtype).eps * 1e2
assert_allclose(y, y_expected, atol=atol, rtol=rtol)
def test_up(up):
dtype_data = dtype_filter = np.float32
nx = 16
nh = 4
x = cp.arange(nx, dtype=dtype_data)
x_cpu = x.get()
h_cpu = np.arange(1, nh + 1, dtype=dtype_filter)
h = cp.asarray(h_cpu)
# up=1 kernel case
cp.testing.assert_allclose(
signal.upfirdn(h_cpu, x_cpu, up=up, down=1),
upfirdn(h, x, up=up, down=1),
)
def test_axis_and_order(shape, axis, order):
dtype_data = dtype_filter = np.float32
size = int(np.prod(shape))
x_cpu = np.arange(size, dtype=dtype_data).reshape(shape, order=order)
h_cpu = np.arange(3, dtype=dtype_filter)
x = cp.asarray(x_cpu, order=order)
h = cp.asarray(h_cpu)
ndim = len(shape)
if axis >= -ndim and axis < ndim:
# up=1 case
cp.testing.assert_allclose(
signal.upfirdn(h_cpu, x_cpu, up=1, down=2, axis=axis),
upfirdn(h, x, up=1, down=2, axis=axis),
)
# down=1 case
cp.testing.assert_allclose(
signal.upfirdn(h_cpu, x_cpu, up=2, down=1, axis=axis),
upfirdn(h, x, up=2, down=1, axis=axis),
)
else:
with pytest.raises(ValueError):
upfirdn(h, x, up=2, down=1, axis=axis)
def scrub(self, x, axis=-1):
yr = np.apply_along_axis(upfirdn_naive, axis, x.get(), self.h.get(),
self.up, self.down)
yr = cp.asarray(yr)
y = upfirdn(self.h, x, self.up, self.down, axis=axis)
dtypes = (self.h.dtype, x.dtype)
if all(d == cp.complex64 for d in dtypes):
assert_equal(y.dtype, cp.complex64)
elif cp.complex64 in dtypes and cp.float32 in dtypes:
assert_equal(y.dtype, cp.complex64)
elif all(d == cp.float32 for d in dtypes):
assert_equal(y.dtype, cp.float32)
elif cp.complex128 in dtypes or cp.complex64 in dtypes:
assert_equal(y.dtype, cp.complex128)
else:
assert_equal(y.dtype, cp.float64)
assert_allclose(yr, y)
def test_vs_lfilter(self):
# Check that up=1.0 gives same answer as lfilter + slicing
random_state = cp.random.RandomState(17)
try_types = (int, cp.float32, cp.complex64, float, complex)
size = 10000
down_factors = [2, 11, 79]
for dtype in try_types:
x = random_state.randn(size).astype(dtype)
if dtype in (cp.complex64, cp.complex128):
x += 1j * random_state.randn(size)
tol = cp.finfo(cp.float32).eps * 100
for down in down_factors:
h = firwin(31, 1.0 / down, window="hamming")
yl = cp.asarray(lfilter(h, 1.0, x.get())[::down])
h = cp.asarray(h)
y = upfirdn(h, x, up=1, down=down)
assert_allclose(yl, y[: yl.size], atol=tol, rtol=tol)
def test_convolve1d(dtype_x, dtype_h, len_x, mode):
x_cpu = np.arange(1, 1 + len_x, dtype=dtype_x)
for len_h in range(1, len_x):
h_cpu = np.arange(1, 1 + len_h, dtype=dtype_h)
min_origin = -(len_h // 2)
max_origin = (len_h - 1) // 2
for origin in range(min_origin, max_origin + 1):
y = ndi.convolve1d(x_cpu, h_cpu, mode=mode, cval=0, origin=origin)
# test via convolve1d
y3 = convolve1d(
cp.asarray(x_cpu),
cp.asarray(h_cpu),
mode=mode,
cval=0,
origin=origin,
)
cp.testing.assert_allclose(y, y3)
# test using upfirdn directly
offset = len(h_cpu) // 2 + origin
mode_kwargs = _get_ndimage_mode_kwargs(mode, cval=0)
y2 = upfirdn(
cp.asarray(h_cpu),
cp.asarray(x_cpu),
offset=offset,
**mode_kwargs,
)[:len_x]
cp.testing.assert_allclose(y, y2)
for origin in [min_origin - 1, max_origin + 1]:
with pytest.raises(ValueError):
convolve1d(
cp.asarray(x_cpu),
cp.asarray(h_cpu),
mode=mode,
cval=0,
origin=origin,
)
def _pad_test(x, npre, npost, mode):
# test array extension by convolving with an impulse padded with zeros
h = cp.zeros((npre + npost + 1))
h[npre] = 1
return upfirdn(h, x, up=1, down=1, mode=mode)