Skip to content

GitLab

Commit 9c95b942 authored by Martin Reinecke's avatar Martin Reinecke
Browse files

add pypocketfft

parent dde8c508
Hide whitespace changes
Inline Side-by-side
pypocketfft/pypocketfft.cc 0 → 100644
View file @ 9c95b942
/*
* This file is part of pocketfft.
* Licensed under a 3-clause BSD style license - see LICENSE.md
*/
/*
* Python interface.
*
* Copyright (C) 2019 Max-Planck-Society
* Copyright (C) 2019 Peter Bell
* \author Martin Reinecke
* \author Peter Bell
*/
#include <pybind11/pybind11.h>
#include <pybind11/numpy.h>
#include <pybind11/stl.h>
#include "mr_util/fft.h"
namespace {
using mr::shape_t;
using mr::stride_t;
using std::size_t;
using std::ptrdiff_t;
namespace py = pybind11;
// Only instantiate long double transforms if they offer more precision
using ldbl_t = typename std::conditional<
sizeof(long double)==sizeof(double), double, long double>::type;
using c64 = std::complex<float>;
using c128 = std::complex<double>;
using clong = std::complex<ldbl_t>;
using f32 = float;
using f64 = double;
using flong = ldbl_t;
auto None = py::none();
shape_t copy_shape(const py::array &arr)
{
shape_t res(size_t(arr.ndim()));
for (size_t i=0; i<res.size(); ++i)
res[i] = size_t(arr.shape(int(i)));
return res;
}
stride_t copy_strides(const py::array &arr)
{
stride_t res(size_t(arr.ndim()));
for (size_t i=0; i<res.size(); ++i)
res[i] = arr.strides(int(i));
return res;
}
shape_t makeaxes(const py::array &in, const py::object &axes)
{
if (axes.is_none())
{
shape_t res(size_t(in.ndim()));
for (size_t i=0; i<res.size(); ++i)
res[i]=i;
return res;
}
auto tmp=axes.cast<std::vector<ptrdiff_t>>();
auto ndim = in.ndim();
if ((tmp.size()>size_t(ndim)) || (tmp.size()==0))
throw std::runtime_error("bad axes argument");
for (auto& sz: tmp)
{
if (sz<0)
sz += ndim;
if ((sz>=ndim) || (sz<0))
throw std::invalid_argument("axes exceeds dimensionality of output");
}
return shape_t(tmp.begin(), tmp.end());
}
#define DISPATCH(arr, T1, T2, T3, func, args) \
{ \
if (py::isinstance<py::array_t<T1>>(arr)) return func<double> args; \
if (py::isinstance<py::array_t<T2>>(arr)) return func<float> args; \
if (py::isinstance<py::array_t<T3>>(arr)) return func<ldbl_t> args; \
throw std::runtime_error("unsupported data type"); \
}
template<typename T> T norm_fct(int inorm, size_t N)
{
if (inorm==0) return T(1);
if (inorm==2) return T(1/ldbl_t(N));
if (inorm==1) return T(1/sqrt(ldbl_t(N)));
throw std::invalid_argument("invalid value for inorm (must be 0, 1, or 2)");
}
template<typename T> T norm_fct(int inorm, const shape_t &shape,
const shape_t &axes, size_t fct=1, int delta=0)
{
if (inorm==0) return T(1);
size_t N(1);
for (auto a: axes)
N *= fct * size_t(int64_t(shape[a])+delta);
return norm_fct<T>(inorm, N);
}
template<typename T> py::array_t<T> prepare_output(py::object &out_,
shape_t &dims)
{
if (out_.is_none()) return py::array_t<T>(dims);
auto tmp = out_.cast<py::array_t<T>>();
if (!tmp.is(out_)) // a new object was created during casting
throw std::runtime_error("unexpected data type for output array");
return tmp;
}
template<typename T> py::array c2c_internal(const py::array &in,
const py::object &axes_, bool forward, int inorm, py::object &out_,
size_t nthreads)
{
auto axes = makeaxes(in, axes_);
auto dims(copy_shape(in));
auto res = prepare_output<std::complex<T>>(out_, dims);
auto s_in=copy_strides(in);
auto s_out=copy_strides(res);
auto d_in=reinterpret_cast<const std::complex<T> *>(in.data());
auto d_out=reinterpret_cast<std::complex<T> *>(res.mutable_data());
{
py::gil_scoped_release release;
T fct = norm_fct<T>(inorm, dims, axes);
mr::c2c(dims, s_in, s_out, axes, forward, d_in, d_out, fct, nthreads);
}
return move(res);
}
template<typename T> py::array c2c_sym_internal(const py::array &in,
const py::object &axes_, bool forward, int inorm, py::object &out_,
size_t nthreads)
{
auto axes = makeaxes(in, axes_);
auto dims(copy_shape(in));
auto res = prepare_output<std::complex<T>>(out_, dims);
auto s_in=copy_strides(in);
auto s_out=copy_strides(res);
auto d_in=reinterpret_cast<const T *>(in.data());
auto d_out=reinterpret_cast<std::complex<T> *>(res.mutable_data());
{
py::gil_scoped_release release;
T fct = norm_fct<T>(inorm, dims, axes);
mr::r2c(dims, s_in, s_out, axes, forward, d_in, d_out, fct, nthreads);
// now fill in second half
using namespace mr::detail_fft;
ndarr<std::complex<T>> ares(res.mutable_data(), dims, s_out);
rev_iter iter(ares, axes);
while(iter.remaining()>0)
{
auto v = ares[iter.ofs()];
ares[iter.rev_ofs()] = conj(v);
iter.advance();
}
}
return move(res);
}
py::array c2c(const py::array &a, const py::object &axes_, bool forward,
int inorm, py::object &out_, size_t nthreads)
{
if (a.dtype().kind() == 'c')
DISPATCH(a, c128, c64, clong, c2c_internal, (a, axes_, forward,
inorm, out_, nthreads))
DISPATCH(a, f64, f32, flong, c2c_sym_internal, (a, axes_, forward,
inorm, out_, nthreads))
}
template<typename T> py::array r2c_internal(const py::array &in,
const py::object &axes_, bool forward, int inorm, py::object &out_,
size_t nthreads)
{
auto axes = makeaxes(in, axes_);
auto dims_in(copy_shape(in)), dims_out(dims_in);
dims_out[axes.back()] = (dims_out[axes.back()]>>1)+1;
py::array res = prepare_output<std::complex<T>>(out_, dims_out);
auto s_in=copy_strides(in);
auto s_out=copy_strides(res);
auto d_in=reinterpret_cast<const T *>(in.data());
auto d_out=reinterpret_cast<std::complex<T> *>(res.mutable_data());
{
py::gil_scoped_release release;
T fct = norm_fct<T>(inorm, dims_in, axes);
mr::r2c(dims_in, s_in, s_out, axes, forward, d_in, d_out, fct,
nthreads);
}
return res;
}
py::array r2c(const py::array &in, const py::object &axes_, bool forward,
int inorm, py::object &out_, size_t nthreads)
{
DISPATCH(in, f64, f32, flong, r2c_internal, (in, axes_, forward, inorm, out_,
nthreads))
}
template<typename T> py::array r2r_fftpack_internal(const py::array &in,
const py::object &axes_, bool real2hermitian, bool forward, int inorm,
py::object &out_, size_t nthreads)
{
auto axes = makeaxes(in, axes_);
auto dims(copy_shape(in));
py::array res = prepare_output<T>(out_, dims);
auto s_in=copy_strides(in);
auto s_out=copy_strides(res);
auto d_in=reinterpret_cast<const T *>(in.data());
auto d_out=reinterpret_cast<T *>(res.mutable_data());
{
py::gil_scoped_release release;
T fct = norm_fct<T>(inorm, dims, axes);
mr::r2r_fftpack(dims, s_in, s_out, axes, real2hermitian, forward,
d_in, d_out, fct, nthreads);
}
return res;
}
py::array r2r_fftpack(const py::array &in, const py::object &axes_,
bool real2hermitian, bool forward, int inorm, py::object &out_,
size_t nthreads)
{
DISPATCH(in, f64, f32, flong, r2r_fftpack_internal, (in, axes_,
real2hermitian, forward, inorm, out_, nthreads))
}
template<typename T> py::array dct_internal(const py::array &in,
const py::object &axes_, int type, int inorm, py::object &out_,
size_t nthreads)
{
auto axes = makeaxes(in, axes_);
auto dims(copy_shape(in));
py::array res = prepare_output<T>(out_, dims);
auto s_in=copy_strides(in);
auto s_out=copy_strides(res);
auto d_in=reinterpret_cast<const T *>(in.data());
auto d_out=reinterpret_cast<T *>(res.mutable_data());
{
py::gil_scoped_release release;
T fct = (type==1) ? norm_fct<T>(inorm, dims, axes, 2, -1)
: norm_fct<T>(inorm, dims, axes, 2);
bool ortho = inorm == 1;
mr::dct(dims, s_in, s_out, axes, type, d_in, d_out, fct, ortho,
nthreads);
}
return res;
}
py::array dct(const py::array &in, int type, const py::object &axes_,
int inorm, py::object &out_, size_t nthreads)
{
if ((type<1) || (type>4)) throw std::invalid_argument("invalid DCT type");
DISPATCH(in, f64, f32, flong, dct_internal, (in, axes_, type, inorm, out_,
nthreads))
}
template<typename T> py::array dst_internal(const py::array &in,
const py::object &axes_, int type, int inorm, py::object &out_,
size_t nthreads)
{
auto axes = makeaxes(in, axes_);
auto dims(copy_shape(in));
py::array res = prepare_output<T>(out_, dims);
auto s_in=copy_strides(in);
auto s_out=copy_strides(res);
auto d_in=reinterpret_cast<const T *>(in.data());
auto d_out=reinterpret_cast<T *>(res.mutable_data());
{
py::gil_scoped_release release;
T fct = (type==1) ? norm_fct<T>(inorm, dims, axes, 2, 1)
: norm_fct<T>(inorm, dims, axes, 2);
bool ortho = inorm == 1;
mr::dst(dims, s_in, s_out, axes, type, d_in, d_out, fct, ortho,
nthreads);
}
return res;
}
py::array dst(const py::array &in, int type, const py::object &axes_,
int inorm, py::object &out_, size_t nthreads)
{
if ((type<1) || (type>4)) throw std::invalid_argument("invalid DST type");
DISPATCH(in, f64, f32, flong, dst_internal, (in, axes_, type, inorm,
out_, nthreads))
}
template<typename T> py::array c2r_internal(const py::array &in,
const py::object &axes_, size_t lastsize, bool forward, int inorm,
py::object &out_, size_t nthreads)
{
auto axes = makeaxes(in, axes_);
size_t axis = axes.back();
shape_t dims_in(copy_shape(in)), dims_out=dims_in;
if (lastsize==0) lastsize=2*dims_in[axis]-1;
if ((lastsize/2) + 1 != dims_in[axis])
throw std::invalid_argument("bad lastsize");
dims_out[axis] = lastsize;
py::array res = prepare_output<T>(out_, dims_out);
auto s_in=copy_strides(in);
auto s_out=copy_strides(res);
auto d_in=reinterpret_cast<const std::complex<T> *>(in.data());
auto d_out=reinterpret_cast<T *>(res.mutable_data());
{
py::gil_scoped_release release;
T fct = norm_fct<T>(inorm, dims_out, axes);
mr::c2r(dims_out, s_in, s_out, axes, forward, d_in, d_out, fct,
nthreads);
}
return res;
}
py::array c2r(const py::array &in, const py::object &axes_, size_t lastsize,
bool forward, int inorm, py::object &out_, size_t nthreads)
{
DISPATCH(in, c128, c64, clong, c2r_internal, (in, axes_, lastsize, forward,
inorm, out_, nthreads))
}
template<typename T> py::array separable_hartley_internal(const py::array &in,
const py::object &axes_, int inorm, py::object &out_, size_t nthreads)
{
auto dims(copy_shape(in));
py::array res = prepare_output<T>(out_, dims);
auto axes = makeaxes(in, axes_);
auto s_in=copy_strides(in);
auto s_out=copy_strides(res);
auto d_in=reinterpret_cast<const T *>(in.data());
auto d_out=reinterpret_cast<T *>(res.mutable_data());
{
py::gil_scoped_release release;
T fct = norm_fct<T>(inorm, dims, axes);
mr::r2r_separable_hartley(dims, s_in, s_out, axes, d_in, d_out, fct,
nthreads);
}
return res;
}
py::array separable_hartley(const py::array &in, const py::object &axes_,
int inorm, py::object &out_, size_t nthreads)
{
DISPATCH(in, f64, f32, flong, separable_hartley_internal, (in, axes_, inorm,
out_, nthreads))
}
template<typename T> py::array genuine_hartley_internal(const py::array &in,
const py::object &axes_, int inorm, py::object &out_, size_t nthreads)
{
auto dims(copy_shape(in));
py::array res = prepare_output<T>(out_, dims);
auto axes = makeaxes(in, axes_);
auto s_in=copy_strides(in);
auto s_out=copy_strides(res);
auto d_in=reinterpret_cast<const T *>(in.data());
auto d_out=reinterpret_cast<T *>(res.mutable_data());
{
py::gil_scoped_release release;
T fct = norm_fct<T>(inorm, dims, axes);
mr::r2r_genuine_hartley(dims, s_in, s_out, axes, d_in, d_out, fct,
nthreads);
}
return res;
}
py::array genuine_hartley(const py::array &in, const py::object &axes_,
int inorm, py::object &out_, size_t nthreads)
{
DISPATCH(in, f64, f32, flong, genuine_hartley_internal, (in, axes_, inorm,
out_, nthreads))
}
// Export good_size in raw C-API to reduce overhead (~4x faster)
PyObject * good_size(PyObject * /*self*/, PyObject * args)
{
Py_ssize_t n_ = -1;
int real = false;
if (!PyArg_ParseTuple(args, "n|p:good_size", &n_, &real))
return nullptr;
if (n_<0)
{
PyErr_SetString(PyExc_ValueError, "Target length must be positive");
return nullptr;
}
if ((n_-1) > static_cast<Py_ssize_t>(std::numeric_limits<size_t>::max() / 11))
{
PyErr_Format(PyExc_ValueError,
"Target length is too large to perform an FFT: %zi", n_);
return nullptr;
}
const auto n = static_cast<size_t>(n_);
using namespace mr::detail_fft;
return PyLong_FromSize_t(
real ? util::good_size_real(n) : util::good_size_cmplx(n));
}
const char *pypocketfft_DS = R"""(Fast Fourier and Hartley transforms.
This module supports
- single, double, and long double precision
- complex and real-valued transforms
- multi-dimensional transforms
For two- and higher-dimensional transforms the code will use SSE2 and AVX
vector instructions for faster execution if these are supported by the CPU and
were enabled during compilation.
)""";
const char *c2c_DS = R"""(Performs a complex FFT.
Parameters
----------
a : numpy.ndarray (any complex or real type)
The input data. If its type is real, a more efficient real-to-complex
transform will be used.
axes : list of integers
The axes along which the FFT is carried out.
If not set, all axes will be transformed.
forward : bool
If `True`, a negative sign is used in the exponent, else a positive one.
inorm : int
Normalization type
0 : no normalization
1 : divide by sqrt(N)
2 : divide by N
where N is the product of the lengths of the transformed axes.
out : numpy.ndarray (same shape as `a`, complex type with same accuracy as `a`)
May be identical to `a`, but if it isn't, it must not overlap with `a`.
If None, a new array is allocated to store the output.
nthreads : int
Number of threads to use. If 0, use the system default (typically governed
by the `OMP_NUM_THREADS` environment variable).
Returns
-------
numpy.ndarray (same shape as `a`, complex type with same accuracy as `a`)
The transformed data.
)""";
const char *r2c_DS = R"""(Performs an FFT whose input is strictly real.
Parameters
----------
a : numpy.ndarray (any real type)
The input data
axes : list of integers
The axes along which the FFT is carried out.
If not set, all axes will be transformed in ascending order.
forward : bool
If `True`, a negative sign is used in the exponent, else a positive one.
inorm : int
Normalization type
0 : no normalization
1 : divide by sqrt(N)
2 : divide by N
where N is the product of the lengths of the transformed input axes.
out : numpy.ndarray (complex type with same accuracy as `a`)
For the required shape, see the `Returns` section.
Must not overlap with `a`.
If None, a new array is allocated to store the output.
nthreads : int
Number of threads to use. If 0, use the system default (typically governed
by the `OMP_NUM_THREADS` environment variable).
Returns
-------
numpy.ndarray (complex type with same accuracy as `a`)
The transformed data. The shape is identical to that of the input array,
except for the axis that was transformed last. If the length of that axis
was n on input, it is n//2+1 on output.
)""";
const char *c2r_DS = R"""(Performs an FFT whose output is strictly real.
Parameters
----------
a : numpy.ndarray (any complex type)
The input data
axes : list of integers
The axes along which the FFT is carried out.
If not set, all axes will be transformed in ascending order.
lastsize : the output size of the last axis to be transformed.
If the corresponding input axis has size n, this can be 2*n-2 or 2*n-1.
forward : bool
If `True`, a negative sign is used in the exponent, else a positive one.
inorm : int
Normalization type
0 : no normalization
1 : divide by sqrt(N)
2 : divide by N
where N is the product of the lengths of the transformed output axes.
out : numpy.ndarray (real type with same accuracy as `a`)
For the required shape, see the `Returns` section.
Must not overlap with `a`.
If None, a new array is allocated to store the output.
nthreads : int
Number of threads to use. If 0, use the system default (typically governed
by the `OMP_NUM_THREADS` environment variable).
Returns
-------
numpy.ndarray (real type with same accuracy as `a`)
The transformed data. The shape is identical to that of the input array,
except for the axis that was transformed last, which has now `lastsize`
entries.
)""";
const char *r2r_fftpack_DS = R"""(Performs a real-valued FFT using the FFTPACK storage scheme.
Parameters
----------
a : numpy.ndarray (any real type)
The input data
axes : list of integers
The axes along which the FFT is carried out.
If not set, all axes will be transformed.
real2hermitian : bool
if True, the input is purely real and the output will have Hermitian
symmetry and be stored in FFTPACK's halfcomplex ordering, otherwise the
opposite.
forward : bool
If `True`, a negative sign is used in the exponent, else a positive one.
inorm : int
Normalization type
0 : no normalization
1 : divide by sqrt(N)
2 : divide by N
where N is the length of `axis`.
out : numpy.ndarray (same shape and data type as `a`)
May be identical to `a`, but if it isn't, it must not overlap with `a`.
If None, a new array is allocated to store the output.
nthreads : int
Number of threads to use. If 0, use the system default (typically governed
by the `OMP_NUM_THREADS` environment variable).
Returns
-------
numpy.ndarray (same shape and data type as `a`)
The transformed data. The shape is identical to that of the input array.
)""";
const char *separable_hartley_DS = R"""(Performs a separable Hartley transform.
For every requested axis, a 1D forward Fourier transform is carried out, and
the real and imaginary parts of the result are added before the next axis is
processed.
Parameters
----------
a : numpy.ndarray (any real type)
The input data
axes : list of integers
The axes along which the transform is carried out.
If not set, all axes will be transformed.
inorm : int
Normalization type
0 : no normalization
1 : divide by sqrt(N)
2 : divide by N
where N is the product of the lengths of the transformed axes.
out : numpy.ndarray (same shape and data type as `a`)
May be identical to `a`, but if it isn't, it must not overlap with `a`.
If None, a new array is allocated to store the output.
nthreads : int
Number of threads to use. If 0, use the system default (typically governed
by the `OMP_NUM_THREADS` environment variable).
Returns
-------
numpy.ndarray (same shape and data type as `a`)
The transformed data
)""";
const char *genuine_hartley_DS = R"""(Performs a full Hartley transform.
A full Fourier transform is carried out over the requested axes, and the
sum of real and imaginary parts of the result is stored in the output
array. For a single transformed axis, this is identical to `separable_hartley`,
but when transforming multiple axes, the results are different.
Parameters
----------
a : numpy.ndarray (any real type)
The input data
axes : list of integers
The axes along which the transform is carried out.
If not set, all axes will be transformed.
inorm : int
Normalization type
0 : no normalization
1 : divide by sqrt(N)
2 : divide by N
where N is the product of the lengths of the transformed axes.
out : numpy.ndarray (same shape and data type as `a`)
May be identical to `a`, but if it isn't, it must not overlap with `a`.
If None, a new array is allocated to store the output.
nthreads : int
Number of threads to use. If 0, use the system default (typically governed
by the `OMP_NUM_THREADS` environment variable).
Returns
-------
numpy.ndarray (same shape and data type as `a`)
The transformed data
)""";
const char *dct_DS = R"""(Performs a discrete cosine transform.
Parameters
----------
a : numpy.ndarray (any real type)
The input data
type : integer
the type of DCT. Must be in [1; 4].
axes : list of integers
The axes along which the transform is carried out.
If not set, all axes will be transformed.