def _promote_dtype(x):
if x.dtype.kind in 'bui':
# use float64 instead of promote_types to match SciPy's behavior
float_dtype = cupy.float64
else:
float_dtype = cupy.promote_types(x.dtype, cupy.float32)
return x.astype(float_dtype, copy=False)
def _convolve(
in1,
in2,
use_convolve,
swapped_inputs,
mode,
cp_stream,
autosync,
):
val = _valfrommode(mode)
# Promote inputs
promType = cp.promote_types(in1.dtype, in2.dtype)
in1 = in1.astype(promType)
in2 = in2.astype(promType)
# Create empty array to hold number of aout dimensions
out_dimens = np.empty(in1.ndim, np.int)
if val == VALID:
for i in range(in1.ndim):
out_dimens[i] = (max(in1.shape[i], in2.shape[i]) -
min(in1.shape[i], in2.shape[i]) + 1)
if out_dimens[i] < 0:
raise Exception(
"no part of the output is valid, use option 1 (same) or 2 \
(full) for third argument")
elif val == SAME:
for i in range(in1.ndim):
if not swapped_inputs:
out_dimens[i] = in1.shape[i] # Per scipy docs
else:
out_dimens[i] = min(in1.shape[i], in2.shape[i])
elif val == FULL:
for i in range(in1.ndim):
out_dimens[i] = in1.shape[i] + in2.shape[i] - 1
else:
raise Exception("mode must be 0 (valid), 1 (same), or 2 (full)")
# Create empty array out on GPU
out = cp.empty(out_dimens.tolist(), in1.dtype)
out = _convolve_gpu(
in1,
out,
in2,
val,
use_convolve,
swapped_inputs,
cp_stream,
)
if autosync is True:
cp_stream.synchronize()
return out
def _get_spline_output(input, output):
"""Create workspace array, temp, and the final dtype for the output.
Differs from SciPy by not always forcing the internal floating point dtype
to be double precision.
"""
complex_data = input.dtype.kind == 'c'
if complex_data:
min_float_dtype = cupy.complex64
else:
min_float_dtype = cupy.float32
if isinstance(output, cupy.ndarray):
if complex_data and output.dtype.kind != 'c':
raise ValueError(
'output must have complex dtype for complex inputs'
)
float_dtype = cupy.promote_types(output.dtype, min_float_dtype)
output_dtype = output.dtype
else:
if output is None:
output = output_dtype = input.dtype
else:
output_dtype = cupy.dtype(output)
float_dtype = cupy.promote_types(output, min_float_dtype)
if (isinstance(output, cupy.ndarray)
and output.dtype == float_dtype == output_dtype
and output.flags.c_contiguous):
if output is not input:
output[...] = input[...]
temp = output
else:
temp = input.astype(float_dtype, copy=False)
temp = cupy.ascontiguousarray(temp)
if cupy.shares_memory(temp, input, 'MAY_SHARE_BOUNDS'):
temp = temp.copy()
return temp, float_dtype, output_dtype
def channelize_poly(x, h, n_chans):
"""
Polyphase channelize signal into n channels
Parameters
----------
x : array_like
The input data to be channelized
h : array_like
The 1-D input filter; will be split into n
channels of int number of taps
n_chans : int
Number of channels for channelizer
Returns
----------
yy : channelized output matrix
Notes
----------
Currently only supports simple channelizer where channel
spacing is equivalent to the number of channels used (zero overlap).
Number of filter taps (len of filter / n_chans) must be <=32.
"""
dtype = cp.promote_types(x.dtype, h.dtype)
x = asarray(x, dtype=dtype)
h = asarray(h, dtype=dtype)
# number of taps in each h_n filter
n_taps = int(len(h) / n_chans)
if n_taps > 32:
raise NotImplementedError(
"The number of calculated taps ({}) in \
each filter is currently capped at 32. Please reduce filter \
length or number of channels".format(
n_taps
)
)
if n_taps > 32:
raise NotImplementedError(
"Number of taps ({}) must be less than (32).".format(n_taps)
)
# number of outputs
n_pts = int(len(x) / n_chans)
if x.dtype == cp.float32 or x.dtype == cp.complex64:
y = cp.empty((n_pts, n_chans), dtype=cp.complex64)
elif x.dtype == cp.float64 or x.dtype == cp.complex128:
y = cp.empty((n_pts, n_chans), dtype=cp.complex128)
_channelizer(x, h, y, n_chans, n_taps, n_pts)
# Remove with CuPy v8
if (x.dtype, n_pts, n_taps, n_chans) in _cupy_fft_cache:
plan = _cupy_fft_cache[(x.dtype, n_pts, n_taps, n_chans)]
else:
plan = _cupy_fft_cache[
(x.dtype, n_pts, n_taps, n_chans)
] = fftpack.get_fft_plan(y, axes=-1)
return cp.conj(fftpack.fft(y, overwrite_x=True, plan=plan)).T
def _dct_or_dst_type3(x,
n=None,
axis=-1,
norm=None,
forward=True,
dst=False,
overwrite_x=False):
"""Forward DCT/DST-III (or inverse DCT/DST-II) along a single axis.
Parameters
----------
x : cupy.ndarray
The data to transform.
n : int
The size of the transform. If None, ``x.shape[axis]`` is used.
axis : int
Axis along which the transform is applied.
forward : bool
Set true to indicate that this is a forward DCT-II as opposed to an
inverse DCT-III (The difference between the two is only in the
normalization factor).
norm : {None, 'ortho', 'forward', 'backward'}
The normalization convention to use.
dst : bool
If True, a discrete sine transform is computed rather than the discrete
cosine transform.
overwrite_x : bool
Indicates that it is okay to overwrite x. In practice, the current
implementation never performs the transform in-place.
Returns
-------
y: cupy.ndarray
The transformed array.
"""
if axis < -x.ndim or axis >= x.ndim:
raise numpy.AxisError('axis out of range')
if axis < 0:
axis += x.ndim
if n is not None and n < 1:
raise ValueError(f'invalid number of data points ({n}) specified')
x = _cook_shape(x, (n, ), (axis, ), 'R2R')
n = x.shape[axis]
# determine normalization factor
if norm == 'ortho':
sl0_scale = 0.5 * math.sqrt(2)
inorm = 'sqrt'
elif norm == 'forward':
sl0_scale = 0.5
inorm = 'full' if forward else 'none'
elif norm == 'backward' or norm is None:
sl0_scale = 0.5
inorm = 'none' if forward else 'full'
else:
raise ValueError(f'Invalid norm value "{norm}", should be "backward", '
'"ortho" or "forward"')
norm_factor = _get_dct_norm_factor(n, inorm=inorm, dct_type=3)
dtype = cupy.promote_types(x, cupy.complex64)
sl0 = [slice(None)] * x.ndim
sl0[axis] = slice(1)
if dst:
if norm == 'ortho':
float_dtype = cupy.promote_types(x.dtype, cupy.float32)
if x.dtype != float_dtype:
x = x.astype(float_dtype)
elif not overwrite_x:
x = x.copy()
x[tuple(sl0)] *= math.sqrt(2)
sl0_scale = 0.5
slrev = [slice(None)] * x.ndim
slrev[axis] = slice(None, None, -1)
x = x[tuple(slrev)]
# scale by exponentials and normalization factor
tmp = _exp_factor_dct3(x, n, axis, dtype, norm_factor)
x = x * tmp # broadcasting
x[tuple(sl0)] *= sl0_scale
# inverse fft
x = _fft.ifft(x, n=n, axis=axis, overwrite_x=True)
x = cupy.real(x)
# reorder entries
return _reshuffle_dct3(x, n, axis, dst)